xref: /illumos-gate/usr/src/lib/libm/common/complex/casin.c (revision 4b5c8e93cab28d3c65ba9d407fd8f46e3be1db1c)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 
22 /*
23  * Copyright 2011 Nexenta Systems, Inc.  All rights reserved.
24  */
25 /*
26  * Copyright 2006 Sun Microsystems, Inc.  All rights reserved.
27  * Use is subject to license terms.
28  */
29 
30 #pragma weak __casin = casin
31 
32 /* INDENT OFF */
33 /*
34  * dcomplex casin(dcomplex z);
35  *
36  * Alogrithm
37  * (based on T.E.Hull, Thomas F. Fairgrieve and Ping Tak Peter Tang's
38  * paper "Implementing the Complex Arcsine and Arccosine Functins Using
39  * Exception Handling", ACM TOMS, Vol 23, pp 299-335)
40  *
41  * The principal value of complex inverse sine function casin(z),
42  * where z = x+iy, can be defined by
43  *
44  * 	casin(z) = asin(B) + i sign(y) log (A + sqrt(A*A-1)),
45  *
46  * where the log function is the natural log, and
47  *             ____________           ____________
48  *       1    /     2    2      1    /     2    2
49  *  A = ---  / (x+1)  + y   +  ---  / (x-1)  + y
50  *       2 \/                   2 \/
51  *             ____________           ____________
52  *       1    /     2    2      1    /     2    2
53  *  B = ---  / (x+1)  + y   -  ---  / (x-1)  + y   .
54  *       2 \/                   2 \/
55  *
56  * The Branch cuts are on the real line from -inf to -1 and from 1 to inf.
57  * The real and imaginary parts are based on Abramowitz and Stegun
58  * [Handbook of Mathematic Functions, 1972].  The sign of the imaginary
59  * part is chosen to be the generally considered the principal value of
60  * this function.
61  *
62  * Notes:1. A is the average of the distances from z to the points (1,0)
63  *          and (-1,0) in the complex z-plane, and in particular A>=1.
64  *       2. B is in [-1,1], and A*B = x.
65  *
66  * Special notes: if casin( x, y) = ( u, v), then
67  *		    casin(-x, y) = (-u, v),
68  *		    casin( x,-y) = ( u,-v),
69  *    in general, we have casin(conj(z))     =  conj(casin(z))
70  *                       casin(-z)          = -casin(z)
71  *			 casin(z)           =  pi/2 - cacos(z)
72  *
73  * EXCEPTION CASES (conform to ISO/IEC 9899:1999(E)):
74  *    casin( 0 + i 0   ) =  0    + i 0
75  *    casin( 0 + i NaN ) =  0    + i NaN
76  *    casin( x + i inf ) =  0    + i inf for finite x
77  *    casin( x + i NaN ) =  NaN  + i NaN with invalid for finite x != 0
78  *    casin(inf + iy   ) =  pi/2 + i inf finite y
79  *    casin(inf + i inf) =  pi/4 + i inf
80  *    casin(inf + i NaN) =  NaN  + i inf
81  *    casin(NaN + i y  ) =  NaN  + i NaN for finite y
82  *    casin(NaN + i inf) =  NaN  + i inf
83  *    casin(NaN + i NaN) =  NaN  + i NaN
84  *
85  * Special Regions (better formula for accuracy and for avoiding spurious
86  * overflow or underflow) (all x and y are assumed nonnegative):
87  *  case 1: y = 0
88  *  case 2: tiny y relative to x-1: y <= ulp(0.5)*|x-1|
89  *  case 3: tiny y: y < 4 sqrt(u), where u = minimum normal number
90  *  case 4: huge y relative to x+1: y >= (1+x)/ulp(0.5)
91  *  case 5: huge x and y: x and y >= sqrt(M)/8, where M = maximum normal number
92  *  case 6: tiny x: x < 4 sqrt(u)
93  *  --------
94  *  case	1 & 2. y=0 or y/|x-1| is tiny. We have
95  *             ____________              _____________
96  *            /      2    2             /       y    2
97  *           / (x+-1)  + y   =  |x+-1| / 1 + (------)
98  *         \/                        \/       |x+-1|
99  *
100  *                                            1      y   2
101  *                           ~  |x+-1| ( 1 + --- (------)  )
102  *                                            2   |x+-1|
103  *
104  *                                           2
105  *                                          y
106  *                           =  |x+-1| + --------.
107  *                                       2|x+-1|
108  *
109  *	Consequently, it is not difficult to see that
110  *                                 2
111  *                                y
112  *                    [ 1 + ------------ ,  if x < 1,
113  *                    [      2(1+x)(1-x)
114  *                    [
115  *                    [
116  *                    [ x,                 if x = 1 (y = 0),
117  *                    [
118  *		A ~=  [             2
119  *                    [        x * y
120  *                    [ x + ------------ ,  if x > 1
121  *                    [      2(1+x)(x-1)
122  *
123  *	and hence
124  *                      ______                                 2
125  *                     / 2                    y               y
126  *               A + \/ A  - 1  ~  1 + ---------------- + -----------, if x < 1,
127  *                                     sqrt((x+1)(1-x))   2(x+1)(1-x)
128  *
129  *
130  *			       ~  x + sqrt((x-1)*(x+1)),              if x >= 1.
131  *
132  *                                         2
133  *                                        y
134  *                          [ x(1 - ------------), if x < 1,
135  *                          [       2(1+x)(1-x)
136  *		B = x/A  ~  [
137  *                          [ 1,                  if x = 1,
138  *			    [
139  *                          [           2
140  *                          [          y
141  *                          [ 1 - ------------ ,   if x > 1,
142  *                          [      2(1+x)(1-x)
143  *	Thus
144  *                            [ asin(x) + i y/sqrt((x-1)*(x+1)), if x <  1
145  *		casin(x+i*y)=[
146  *                            [ pi/2    + i log(x+sqrt(x*x-1)),  if x >= 1
147  *
148  *  case 3. y < 4 sqrt(u), where u = minimum normal x.
149  *	After case 1 and 2, this will only occurs when x=1. When x=1, we have
150  *	   A = (sqrt(4+y*y)+y)/2 ~ 1 + y/2 + y^2/8 + ...
151  *	and
152  *	   B = 1/A = 1 - y/2 + y^2/8 + ...
153  * 	Since
154  *	   asin(x) = pi/2-2*asin(sqrt((1-x)/2))
155  *	   asin(x) = x + x^3/6 + x^5*3/40 + x^7*15/336 + ...
156  *	we have, for the real part asin(B),
157  *	   asin(1-y/2) ~ pi/2 - 2 asin(sqrt(y/4))
158  *	               ~ pi/2 - sqrt(y)
159  *	For the imaginary part,
160  *	   log(A+sqrt(A*A-1)) ~ log(1+y/2+sqrt(2*y/2))
161  *	                      = log(1+y/2+sqrt(y))
162  *	                      = (y/2+sqrt(y)) - (y/2+sqrt(y))^2/2 + ...
163  *	                      ~ sqrt(y) - y*(sqrt(y)+y/2)/2
164  *	                      ~ sqrt(y)
165  *
166  *  case 4. y >= (x+1)ulp(0.5). In this case, A ~ y and B ~ x/y. Thus
167  *	   real part = asin(B) ~ x/y (be careful, x/y may underflow)
168  * 	and
169  *	   imag part = log(y+sqrt(y*y-one))
170  *
171  *
172  *  case 5. Both x and y are large: x and y > sqrt(M)/8, where M = maximum x
173  *	In this case,
174  *	   A ~ sqrt(x*x+y*y)
175  *	   B ~ x/sqrt(x*x+y*y).
176  *	Thus
177  *	   real part = asin(B) = atan(x/y),
178  *	   imag part = log(A+sqrt(A*A-1)) ~ log(2A)
179  *	             = log(2) + 0.5*log(x*x+y*y)
180  *	             = log(2) + log(y) + 0.5*log(1+(x/y)^2)
181  *
182  *  case 6. x < 4 sqrt(u). In this case, we have
183  *	    A ~ sqrt(1+y*y), B = x/sqrt(1+y*y).
184  *	Since B is tiny, we have
185  *	    real part = asin(B) ~ B = x/sqrt(1+y*y)
186  *	    imag part = log(A+sqrt(A*A-1)) = log (A+sqrt(y*y))
187  *	              = log(y+sqrt(1+y*y))
188  *	              = 0.5*log(y^2+2ysqrt(1+y^2)+1+y^2)
189  *	              = 0.5*log(1+2y(y+sqrt(1+y^2)));
190  *	              = 0.5*log1p(2y(y+A));
191  *
192  * 	casin(z) = asin(B) + i sign(y) log (A + sqrt(A*A-1)),
193  */
194 /* INDENT ON */
195 
196 #include "libm.h"		/* asin/atan/fabs/log/log1p/sqrt */
197 #include "complex_wrapper.h"
198 
199 /* INDENT OFF */
200 static const double
201 	zero = 0.0,
202 	one = 1.0,
203 	E = 1.11022302462515654042e-16,			/* 2**-53 */
204 	ln2 = 6.93147180559945286227e-01,
205 	pi_2 = 1.570796326794896558e+00,
206 	pi_2_l = 6.123233995736765886e-17,
207 	pi_4 = 7.85398163397448278999e-01,
208 	Foursqrtu = 5.96667258496016539463e-154,	/* 2**(-509) */
209 	Acrossover = 1.5,
210 	Bcrossover = 0.6417,
211 	half = 0.5;
212 /* INDENT ON */
213 
214 dcomplex
215 casin(dcomplex z) {
216 	double x, y, t, R, S, A, Am1, B, y2, xm1, xp1, Apx;
217 	int ix, iy, hx, hy;
218 	unsigned lx, ly;
219 	dcomplex ans;
220 
221 	x = D_RE(z);
222 	y = D_IM(z);
223 	hx = HI_WORD(x);
224 	lx = LO_WORD(x);
225 	hy = HI_WORD(y);
226 	ly = LO_WORD(y);
227 	ix = hx & 0x7fffffff;
228 	iy = hy & 0x7fffffff;
229 	x = fabs(x);
230 	y = fabs(y);
231 
232 	/* special cases */
233 
234 	/* x is inf or NaN */
235 	if (ix >= 0x7ff00000) {	/* x is inf or NaN */
236 		if (ISINF(ix, lx)) {	/* x is INF */
237 			D_IM(ans) = x;
238 			if (iy >= 0x7ff00000) {
239 				if (ISINF(iy, ly))
240 					/* casin(inf + i inf) = pi/4 + i inf */
241 					D_RE(ans) = pi_4;
242 				else	/* casin(inf + i NaN) = NaN  + i inf  */
243 					D_RE(ans) = y + y;
244 			} else	/* casin(inf + iy) = pi/2 + i inf */
245 				D_RE(ans) = pi_2;
246 		} else {		/* x is NaN */
247 			if (iy >= 0x7ff00000) {
248 				/* INDENT OFF */
249 				/*
250 				 * casin(NaN + i inf) = NaN + i inf
251 				 * casin(NaN + i NaN) = NaN + i NaN
252 				 */
253 				/* INDENT ON */
254 				D_IM(ans) = y + y;
255 				D_RE(ans) = x + x;
256 			} else {
257 				/* casin(NaN + i y ) = NaN  + i NaN */
258 				D_IM(ans) = D_RE(ans) = x + y;
259 			}
260 		}
261 		if (hx < 0)
262 			D_RE(ans) = -D_RE(ans);
263 		if (hy < 0)
264 			D_IM(ans) = -D_IM(ans);
265 		return (ans);
266 	}
267 
268 	/* casin(+0 + i 0  ) =  0   + i 0. */
269 	if ((ix | lx | iy | ly) == 0)
270 		return (z);
271 
272 	if (iy >= 0x7ff00000) {	/* y is inf or NaN */
273 		if (ISINF(iy, ly)) {	/* casin(x + i inf) =  0   + i inf */
274 			D_IM(ans) = y;
275 			D_RE(ans) = zero;
276 		} else {		/* casin(x + i NaN) = NaN  + i NaN */
277 			D_IM(ans) = x + y;
278 			if ((ix | lx) == 0)
279 				D_RE(ans) = x;
280 			else
281 				D_RE(ans) = y;
282 		}
283 		if (hx < 0)
284 			D_RE(ans) = -D_RE(ans);
285 		if (hy < 0)
286 			D_IM(ans) = -D_IM(ans);
287 		return (ans);
288 	}
289 
290 	if ((iy | ly) == 0) {	/* region 1: y=0 */
291 		if (ix < 0x3ff00000) {	/* |x| < 1 */
292 			D_RE(ans) = asin(x);
293 			D_IM(ans) = zero;
294 		} else {
295 			D_RE(ans) = pi_2;
296 			if (ix >= 0x43500000)	/* |x| >= 2**54 */
297 				D_IM(ans) = ln2 + log(x);
298 			else if (ix >= 0x3ff80000)	/* x > Acrossover */
299 				D_IM(ans) = log(x + sqrt((x - one) * (x +
300 					one)));
301 			else {
302 				xm1 = x - one;
303 				D_IM(ans) = log1p(xm1 + sqrt(xm1 * (x + one)));
304 			}
305 		}
306 	} else if (y <= E * fabs(x - one)) {	/* region 2: y < tiny*|x-1| */
307 		if (ix < 0x3ff00000) {	/* x < 1 */
308 			D_RE(ans) = asin(x);
309 			D_IM(ans) = y / sqrt((one + x) * (one - x));
310 		} else {
311 			D_RE(ans) = pi_2;
312 			if (ix >= 0x43500000) {	/* |x| >= 2**54 */
313 				D_IM(ans) = ln2 + log(x);
314 			} else if (ix >= 0x3ff80000)	/* x > Acrossover */
315 				D_IM(ans) = log(x + sqrt((x - one) * (x +
316 					one)));
317 			else
318 				D_IM(ans) = log1p((x - one) + sqrt((x - one) *
319 					(x + one)));
320 		}
321 	} else if (y < Foursqrtu) {	/* region 3 */
322 		t = sqrt(y);
323 		D_RE(ans) = pi_2 - (t - pi_2_l);
324 		D_IM(ans) = t;
325 	} else if (E * y - one >= x) {	/* region 4 */
326 		D_RE(ans) = x / y;	/* need to fix underflow cases */
327 		D_IM(ans) = ln2 + log(y);
328 	} else if (ix >= 0x5fc00000 || iy >= 0x5fc00000) {	/* x,y>2**509 */
329 		/* region 5: x+1 or y is very large (>= sqrt(max)/8) */
330 		t = x / y;
331 		D_RE(ans) = atan(t);
332 		D_IM(ans) = ln2 + log(y) + half * log1p(t * t);
333 	} else if (x < Foursqrtu) {
334 		/* region 6: x is very small, < 4sqrt(min) */
335 		A = sqrt(one + y * y);
336 		D_RE(ans) = x / A;	/* may underflow */
337 		if (iy >= 0x3ff80000)	/* if y > Acrossover */
338 			D_IM(ans) = log(y + A);
339 		else
340 			D_IM(ans) = half * log1p((y + y) * (y + A));
341 	} else {	/* safe region */
342 		y2 = y * y;
343 		xp1 = x + one;
344 		xm1 = x - one;
345 		R = sqrt(xp1 * xp1 + y2);
346 		S = sqrt(xm1 * xm1 + y2);
347 		A = half * (R + S);
348 		B = x / A;
349 
350 		if (B <= Bcrossover)
351 			D_RE(ans) = asin(B);
352 		else {		/* use atan and an accurate approx to a-x */
353 			Apx = A + x;
354 			if (x <= one)
355 				D_RE(ans) = atan(x / sqrt(half * Apx * (y2 /
356 					(R + xp1) + (S - xm1))));
357 			else
358 				D_RE(ans) = atan(x / (y * sqrt(half * (Apx /
359 					(R + xp1) + Apx / (S + xm1)))));
360 		}
361 		if (A <= Acrossover) {
362 			/* use log1p and an accurate approx to A-1 */
363 			if (x < one)
364 				Am1 = half * (y2 / (R + xp1) + y2 / (S - xm1));
365 			else
366 				Am1 = half * (y2 / (R + xp1) + (S + xm1));
367 			D_IM(ans) = log1p(Am1 + sqrt(Am1 * (A + one)));
368 		} else {
369 			D_IM(ans) = log(A + sqrt(A * A - one));
370 		}
371 	}
372 
373 	if (hx < 0)
374 		D_RE(ans) = -D_RE(ans);
375 	if (hy < 0)
376 		D_IM(ans) = -D_IM(ans);
377 
378 	return (ans);
379 }
380