3 * Copyright (C) 1996,1997,1998,1999,2000,2001,2002,2003 Gerard Jungman
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License as published by
7 * the Free Software Foundation; either version 3 of the License, or (at
8 * your option) any later version.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License for more details.
15 * You should have received a copy of the GNU General Public License
16 * along with this program; if not, write to the Free Software
17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301, USA.
20 /* Author: G. Jungman */
21 /* Miscellaneous support functions for Bessel function evaluations.
24 #include <gsl/gsl_math.h>
25 #include <gsl/gsl_errno.h>
26 #include <gsl/gsl_sf_airy.h>
27 #include <gsl/gsl_sf_elementary.h>
28 #include <gsl/gsl_sf_exp.h>
29 #include <gsl/gsl_sf_gamma.h>
30 #include <gsl/gsl_sf_trig.h>
34 #include "bessel_amp_phase.h"
35 #include "bessel_temme.h"
38 #define CubeRoot2_ 1.25992104989487316476721060728
42 /* Debye functions [Abramowitz+Stegun, 9.3.9-10] */
45 debye_u1(const double * tpow)
47 return (3.0*tpow[1] - 5.0*tpow[3])/24.0;
51 debye_u2(const double * tpow)
53 return (81.0*tpow[2] - 462.0*tpow[4] + 385.0*tpow[6])/1152.0;
57 static double debye_u3(const double * tpow)
59 return (30375.0*tpow[3] - 369603.0*tpow[5] + 765765.0*tpow[7] - 425425.0*tpow[9])/414720.0;
63 static double debye_u4(const double * tpow)
65 return (4465125.0*tpow[4] - 94121676.0*tpow[6] + 349922430.0*tpow[8] -
66 446185740.0*tpow[10] + 185910725.0*tpow[12])/39813120.0;
70 static double debye_u5(const double * tpow)
72 return (1519035525.0*tpow[5] - 49286948607.0*tpow[7] +
73 284499769554.0*tpow[9] - 614135872350.0*tpow[11] +
74 566098157625.0*tpow[13] - 188699385875.0*tpow[15])/6688604160.0;
79 static double debye_u6(const double * tpow)
81 return (2757049477875.0*tpow[6] - 127577298354750.0*tpow[8] +
82 1050760774457901.0*tpow[10] - 3369032068261860.0*tpow[12] +
83 5104696716244125.0*tpow[14] - 3685299006138750.0*tpow[16] +
84 1023694168371875.0*tpow[18])/4815794995200.0;
89 /*-*-*-*-*-*-*-*-*-*-*-* Functions with Error Codes *-*-*-*-*-*-*-*-*-*-*-*/
92 gsl_sf_bessel_IJ_taylor_e(const double nu, const double x,
95 const double threshold,
96 gsl_sf_result * result
99 /* CHECK_POINTER(result) */
101 if(nu < 0.0 || x < 0.0) {
102 DOMAIN_ERROR(result);
116 gsl_sf_result prefactor; /* (x/2)^nu / Gamma(nu+1) */
126 stat_pre = GSL_SUCCESS;
128 else if(nu < INT_MAX-1) {
129 /* Separate the integer part and use
130 * y^nu / Gamma(nu+1) = y^N /N! y^f / (N+1)_f,
131 * to control the error.
133 const int N = (int)floor(nu + 0.5);
134 const double f = nu - N;
135 gsl_sf_result poch_factor;
136 gsl_sf_result tc_factor;
137 const int stat_poch = gsl_sf_poch_e(N+1.0, f, &poch_factor);
138 const int stat_tc = gsl_sf_taylorcoeff_e(N, 0.5*x, &tc_factor);
139 const double p = pow(0.5*x,f);
140 prefactor.val = tc_factor.val * p / poch_factor.val;
141 prefactor.err = tc_factor.err * p / poch_factor.val;
142 prefactor.err += fabs(prefactor.val) / poch_factor.val * poch_factor.err;
143 prefactor.err += 2.0 * GSL_DBL_EPSILON * fabs(prefactor.val);
144 stat_pre = GSL_ERROR_SELECT_2(stat_tc, stat_poch);
148 const int stat_lg = gsl_sf_lngamma_e(nu+1.0, &lg);
149 const double term1 = nu*log(0.5*x);
150 const double term2 = lg.val;
151 const double ln_pre = term1 - term2;
152 const double ln_pre_err = GSL_DBL_EPSILON * (fabs(term1)+fabs(term2)) + lg.err;
153 const int stat_ex = gsl_sf_exp_err_e(ln_pre, ln_pre_err, &prefactor);
154 stat_pre = GSL_ERROR_SELECT_2(stat_ex, stat_lg);
158 * [Abramowitz+Stegun, 9.1.10]
159 * [Abramowitz+Stegun, 9.6.7]
162 const double y = sign * 0.25 * x*x;
167 for(k=1; k<=kmax; k++) {
168 term *= y/((nu+k)*k);
170 if(fabs(term/sumk) < threshold) break;
174 sum.err = threshold * fabs(sumk);
176 stat_sum = ( k >= kmax ? GSL_EMAXITER : GSL_SUCCESS );
179 stat_mul = gsl_sf_multiply_err_e(prefactor.val, prefactor.err,
183 return GSL_ERROR_SELECT_3(stat_mul, stat_pre, stat_sum);
188 /* Hankel's Asymptotic Expansion - A&S 9.2.5
191 * error ~ O( ((nu*nu+1)/x)^4 )
193 * empirical error analysis:
194 * choose GSL_ROOT4_MACH_EPS * x > (nu*nu + 1)
196 * This is not especially useful. When the argument gets
197 * large enough for this to apply, the cos() and sin()
198 * start loosing digits. However, this seems inevitable
199 * for this particular method.
201 * Wed Jun 25 14:39:38 MDT 2003 [GJ]
202 * This function was inconsistent since the Q term did not
203 * go to relative order eps^2. That's why the error estimate
204 * originally given was screwy (it didn't make sense that the
205 * "empirical" error was coming out O(eps^3)).
206 * With Q to proper order, the error is O(eps^4).
208 * Sat Mar 15 05:16:18 GMT 2008 [BG]
209 * Extended to use additional terms in the series to gain
215 gsl_sf_bessel_Jnu_asympx_e(const double nu, const double x, gsl_sf_result * result)
217 double mu = 4.0*nu*nu;
218 double chi = x - (0.5*nu + 0.25)*M_PI;
228 t *= (k == 0) ? 1 : -(mu - (2*k-1)*(2*k-1)) / (k * (8 * x));
229 convP = fabs(t) < GSL_DBL_EPSILON * fabs(P);
234 t *= (mu - (2*k-1)*(2*k-1)) / (k * (8 * x));
235 convQ = fabs(t) < GSL_DBL_EPSILON * fabs(Q);
238 /* To preserve the consistency of the series we need to exit
239 when P and Q have the same number of terms */
241 if (convP && convQ && k > (nu / 2))
249 double pre = sqrt(2.0/(M_PI*x));
253 result->val = pre * (c*P - s*Q);
254 result->err = pre * GSL_DBL_EPSILON * (fabs(c*P) + fabs(s*Q) + fabs(t)) * (1 + fabs(x));
255 /* NB: final term accounts for phase error with large x */
265 gsl_sf_bessel_Ynu_asympx_e(const double nu, const double x, gsl_sf_result * result)
270 double beta = -0.5*nu*M_PI;
271 int stat_a = gsl_sf_bessel_asymp_Mnu_e(nu, x, &l);
272 int stat_t = gsl_sf_bessel_asymp_thetanu_corr_e(nu, x, &theta);
273 double sin_alpha = sin(alpha);
274 double cos_alpha = cos(alpha);
275 double sin_chi = sin(beta + theta);
276 double cos_chi = cos(beta + theta);
277 double sin_term = sin_alpha * cos_chi + sin_chi * cos_alpha;
278 double sin_term_mag = fabs(sin_alpha * cos_chi) + fabs(sin_chi * cos_alpha);
279 result->val = ampl * sin_term;
280 result->err = fabs(ampl) * GSL_DBL_EPSILON * sin_term_mag;
281 result->err += fabs(result->val) * 2.0 * GSL_DBL_EPSILON;
283 if(fabs(alpha) > 1.0/GSL_DBL_EPSILON) {
284 result->err *= 0.5 * fabs(alpha);
286 else if(fabs(alpha) > 1.0/GSL_SQRT_DBL_EPSILON) {
287 result->err *= 256.0 * fabs(alpha) * GSL_SQRT_DBL_EPSILON;
290 return GSL_ERROR_SELECT_2(stat_t, stat_a);
297 gsl_sf_bessel_Inu_scaled_asympx_e(const double nu, const double x, gsl_sf_result * result)
299 double mu = 4.0*nu*nu;
300 double mum1 = mu-1.0;
301 double mum9 = mu-9.0;
302 double pre = 1.0/sqrt(2.0*M_PI*x);
304 result->val = pre * (1.0 - mum1/(8.0*x) + mum1*mum9/(128.0*x*x));
305 result->err = 2.0 * GSL_DBL_EPSILON * fabs(result->val) + pre * fabs(0.1*r*r*r);
312 gsl_sf_bessel_Knu_scaled_asympx_e(const double nu, const double x, gsl_sf_result * result)
314 double mu = 4.0*nu*nu;
315 double mum1 = mu-1.0;
316 double mum9 = mu-9.0;
317 double pre = sqrt(M_PI/(2.0*x));
319 result->val = pre * (1.0 + mum1/(8.0*x) + mum1*mum9/(128.0*x*x));
320 result->err = 2.0 * GSL_DBL_EPSILON * fabs(result->val) + pre * fabs(0.1*r*r*r);
325 /* nu -> Inf; uniform in x > 0 [Abramowitz+Stegun, 9.7.7]
328 * The error has the form u_N(t)/nu^N where 0 <= t <= 1.
329 * It is not hard to show that |u_N(t)| is small for such t.
330 * We have N=6 here, and |u_6(t)| < 0.025, so the error is clearly
331 * bounded by 0.025/nu^6. This gives the asymptotic bound on nu
332 * seen below as nu ~ 100. For general MACH_EPS it will be
333 * nu > 0.5 / MACH_EPS^(1/6)
334 * When t is small, the bound is even better because |u_N(t)| vanishes
335 * as t->0. In fact u_N(t) ~ C t^N as t->0, with C ~= 0.1.
337 * err_N <= min(0.025, C(1/(1+(x/nu)^2))^3) / nu^6
339 * min(0.29/nu^2, 0.5/(nu^2+x^2)) < MACH_EPS^{1/3}
340 * and this is the general form.
342 * empirical error analysis, assuming 14 digit requirement:
343 * choose x > 50.000 nu ==> nu > 3
344 * choose x > 10.000 nu ==> nu > 15
345 * choose x > 2.000 nu ==> nu > 50
346 * choose x > 1.000 nu ==> nu > 75
347 * choose x > 0.500 nu ==> nu > 80
348 * choose x > 0.100 nu ==> nu > 83
350 * This makes sense. For x << nu, the error will be of the form u_N(1)/nu^N,
351 * since the polynomial term will be evaluated near t=1, so the bound
352 * on nu will become constant for small x. Furthermore, increasing x with
353 * nu fixed will decrease the error.
356 gsl_sf_bessel_Inu_scaled_asymp_unif_e(const double nu, const double x, gsl_sf_result * result)
360 double root_term = hypot(1.0,z);
361 double pre = 1.0/sqrt(2.0*M_PI*nu * root_term);
362 double eta = root_term + log(z/(1.0+root_term));
363 double ex_arg = ( z < 1.0/GSL_ROOT3_DBL_EPSILON ? nu*(-z + eta) : -0.5*nu/z*(1.0 - 1.0/(12.0*z*z)) );
364 gsl_sf_result ex_result;
365 int stat_ex = gsl_sf_exp_e(ex_arg, &ex_result);
366 if(stat_ex == GSL_SUCCESS) {
367 double t = 1.0/root_term;
371 for(i=1; i<16; i++) tpow[i] = t * tpow[i-1];
372 sum = 1.0 + debye_u1(tpow)/nu + debye_u2(tpow)/(nu*nu) + debye_u3(tpow)/(nu*nu*nu)
373 + debye_u4(tpow)/(nu*nu*nu*nu) + debye_u5(tpow)/(nu*nu*nu*nu*nu);
374 result->val = pre * ex_result.val * sum;
375 result->err = pre * ex_result.val / (nu*nu*nu*nu*nu*nu);
376 result->err += pre * ex_result.err * fabs(sum);
377 result->err += 2.0 * GSL_DBL_EPSILON * fabs(result->val);
388 /* nu -> Inf; uniform in x > 0 [Abramowitz+Stegun, 9.7.8]
391 * identical to that above for Inu_scaled
394 gsl_sf_bessel_Knu_scaled_asymp_unif_e(const double nu, const double x, gsl_sf_result * result)
398 double root_term = hypot(1.0,z);
399 double pre = sqrt(M_PI/(2.0*nu*root_term));
400 double eta = root_term + log(z/(1.0+root_term));
401 double ex_arg = ( z < 1.0/GSL_ROOT3_DBL_EPSILON ? nu*(z - eta) : 0.5*nu/z*(1.0 + 1.0/(12.0*z*z)) );
402 gsl_sf_result ex_result;
403 int stat_ex = gsl_sf_exp_e(ex_arg, &ex_result);
404 if(stat_ex == GSL_SUCCESS) {
405 double t = 1.0/root_term;
409 for(i=1; i<16; i++) tpow[i] = t * tpow[i-1];
410 sum = 1.0 - debye_u1(tpow)/nu + debye_u2(tpow)/(nu*nu) - debye_u3(tpow)/(nu*nu*nu)
411 + debye_u4(tpow)/(nu*nu*nu*nu) - debye_u5(tpow)/(nu*nu*nu*nu*nu);
412 result->val = pre * ex_result.val * sum;
413 result->err = pre * ex_result.err * fabs(sum);
414 result->err += pre * ex_result.val / (nu*nu*nu*nu*nu*nu);
415 result->err += 2.0 * GSL_DBL_EPSILON * fabs(result->val);
426 /* Evaluate J_mu(x),J_{mu+1}(x) and Y_mu(x),Y_{mu+1}(x) for |mu| < 1/2
429 gsl_sf_bessel_JY_mu_restricted(const double mu, const double x,
430 gsl_sf_result * Jmu, gsl_sf_result * Jmup1,
431 gsl_sf_result * Ymu, gsl_sf_result * Ymup1)
433 /* CHECK_POINTER(Jmu) */
434 /* CHECK_POINTER(Jmup1) */
435 /* CHECK_POINTER(Ymu) */
436 /* CHECK_POINTER(Ymup1) */
438 if(x < 0.0 || fabs(mu) > 0.5) {
447 GSL_ERROR ("error", GSL_EDOM);
464 GSL_ERROR ("error", GSL_EDOM);
471 /* Use Taylor series for J and the Temme series for Y.
472 * The Taylor series for J requires nu > 0, so we shift
473 * up one and use the recursion relation to get Jmu, in
477 int stat_J1 = gsl_sf_bessel_IJ_taylor_e(mu+1.0, x, -1, 100, GSL_DBL_EPSILON, Jmup1);
478 int stat_J2 = gsl_sf_bessel_IJ_taylor_e(mu+2.0, x, -1, 100, GSL_DBL_EPSILON, &Jmup2);
479 double c = 2.0*(mu+1.0)/x;
480 Jmu->val = c * Jmup1->val - Jmup2.val;
481 Jmu->err = c * Jmup1->err + Jmup2.err;
482 Jmu->err += 2.0 * GSL_DBL_EPSILON * fabs(Jmu->val);
483 stat_J = GSL_ERROR_SELECT_2(stat_J1, stat_J2);
484 stat_Y = gsl_sf_bessel_Y_temme(mu, x, Ymu, Ymup1);
485 return GSL_ERROR_SELECT_2(stat_J, stat_Y);
487 else if(x < 1000.0) {
491 const int stat_CF1 = gsl_sf_bessel_J_CF1(mu, x, &J_ratio, &J_sgn);
492 const int stat_CF2 = gsl_sf_bessel_JY_steed_CF2(mu, x, &P, &Q);
493 double Jprime_J_ratio = mu/x - J_ratio;
494 double gamma = (P - Jprime_J_ratio)/Q;
495 Jmu->val = J_sgn * sqrt(2.0/(M_PI*x) / (Q + gamma*(P-Jprime_J_ratio)));
496 Jmu->err = 4.0 * GSL_DBL_EPSILON * fabs(Jmu->val);
497 Jmup1->val = J_ratio * Jmu->val;
498 Jmup1->err = fabs(J_ratio) * Jmu->err;
499 Ymu->val = gamma * Jmu->val;
500 Ymu->err = fabs(gamma) * Jmu->err;
501 Ymup1->val = Ymu->val * (mu/x - P - Q/gamma);
502 Ymup1->err = Ymu->err * fabs(mu/x - P - Q/gamma) + 4.0*GSL_DBL_EPSILON*fabs(Ymup1->val);
503 return GSL_ERROR_SELECT_2(stat_CF1, stat_CF2);
506 /* Use asymptotics for large argument.
508 const int stat_J0 = gsl_sf_bessel_Jnu_asympx_e(mu, x, Jmu);
509 const int stat_J1 = gsl_sf_bessel_Jnu_asympx_e(mu+1.0, x, Jmup1);
510 const int stat_Y0 = gsl_sf_bessel_Ynu_asympx_e(mu, x, Ymu);
511 const int stat_Y1 = gsl_sf_bessel_Ynu_asympx_e(mu+1.0, x, Ymup1);
512 stat_J = GSL_ERROR_SELECT_2(stat_J0, stat_J1);
513 stat_Y = GSL_ERROR_SELECT_2(stat_Y0, stat_Y1);
514 return GSL_ERROR_SELECT_2(stat_J, stat_Y);
521 gsl_sf_bessel_J_CF1(const double nu, const double x,
522 double * ratio, double * sgn)
524 const double RECUR_BIG = GSL_SQRT_DBL_MAX;
525 const double RECUR_SMALL = GSL_SQRT_DBL_MIN;
526 const int maxiter = 10000;
532 double a1 = x/(2.0*(nu+1.0));
533 double An = Anm1 + a1*Anm2;
534 double Bn = Bnm1 + a1*Bnm2;
548 an = -x*x/(4.0*(nu+n-1.0)*(nu+n));
552 if(fabs(An) > RECUR_BIG || fabs(Bn) > RECUR_BIG) {
558 } else if(fabs(An) < RECUR_SMALL || fabs(Bn) < RECUR_SMALL) {
571 dn = 1.0 / (2.0*(nu+n)/x - dn);
574 if(fabs(del - 1.0) < 2.0*GSL_DBL_EPSILON) break;
577 /* FIXME: we should return an error term here as well, because the
578 error from this recurrence affects the overall error estimate. */
584 GSL_ERROR ("error", GSL_EMAXITER);
591 /* Evaluate the continued fraction CF1 for J_{nu+1}/J_nu
592 * using Gautschi (Euler) equivalent series.
593 * This exhibits an annoying problem because the
594 * a_k are not positive definite (in fact they are all negative).
595 * There are cases when rho_k blows up. Example: nu=1,x=4.
599 gsl_sf_bessel_J_CF1_ser(const double nu, const double x,
600 double * ratio, double * sgn)
602 const int maxk = 20000;
610 for(k=1; k<maxk; k++) {
611 double ak = -0.25 * (x/(nu+k)) * x/(nu+k+1.0);
612 rhok = -ak*(1.0 + rhok)/(1.0 + ak*(1.0 + rhok));
616 dk = 1.0 / (2.0/x - (nu+k-1.0)/(nu+k) * dk);
619 if(fabs(tk/sum) < GSL_DBL_EPSILON) break;
622 *ratio = x/(2.0*(nu+1.0)) * sum;
626 GSL_ERROR ("error", GSL_EMAXITER);
633 /* Evaluate the continued fraction CF1 for I_{nu+1}/I_nu
634 * using Gautschi (Euler) equivalent series.
637 gsl_sf_bessel_I_CF1_ser(const double nu, const double x, double * ratio)
639 const int maxk = 20000;
645 for(k=1; k<maxk; k++) {
646 double ak = 0.25 * (x/(nu+k)) * x/(nu+k+1.0);
647 rhok = -ak*(1.0 + rhok)/(1.0 + ak*(1.0 + rhok));
650 if(fabs(tk/sum) < GSL_DBL_EPSILON) break;
653 *ratio = x/(2.0*(nu+1.0)) * sum;
656 GSL_ERROR ("error", GSL_EMAXITER);
663 gsl_sf_bessel_JY_steed_CF2(const double nu, const double x,
664 double * P, double * Q)
666 const int max_iter = 10000;
667 const double SMALL = 1.0e-100;
671 double x_inv = 1.0/x;
672 double a = 0.25 - nu*nu;
673 double p = -0.5*x_inv;
677 double fact = a*x_inv/(p*p + q*q);
678 double cr = br + q*fact;
679 double ci = bi + p*fact;
680 double den = br*br + bi*bi;
683 double dlr = cr*dr - ci*di;
684 double dli = cr*di + ci*dr;
685 double temp = p*dlr - q*dli;
688 for (i=2; i<=max_iter; i++) {
693 if(fabs(dr)+fabs(di) < SMALL) dr = SMALL;
694 fact = a/(cr*cr+ci*ci);
697 if(fabs(cr)+fabs(ci) < SMALL) cr = SMALL;
703 temp = p*dlr - q*dli;
706 if(fabs(dlr-1.0)+fabs(dli) < GSL_DBL_EPSILON) break;
713 GSL_ERROR ("error", GSL_EMAXITER);
719 /* Evaluate continued fraction CF2, using Thompson-Barnett-Temme method,
720 * to obtain values of exp(x)*K_nu and exp(x)*K_{nu+1}.
722 * This is unstable for small x; x > 2 is a good cutoff.
723 * Also requires |nu| < 1/2.
726 gsl_sf_bessel_K_scaled_steed_temme_CF2(const double nu, const double x,
727 double * K_nu, double * K_nup1,
730 const int maxiter = 10000;
733 double bi = 2.0*(1.0 + x);
741 double ai = -(0.25 - nu*nu);
746 double s = 1.0 + Qi*delhi;
748 for(i=2; i<=maxiter; i++) {
753 tmp = (qi - bi*qip1)/ai;
758 di = 1.0/(bi + ai*di);
759 delhi = (bi*di - 1.0) * delhi;
763 if(fabs(dels/s) < GSL_DBL_EPSILON) break;
768 *K_nu = sqrt(M_PI/(2.0*x)) / s;
769 *K_nup1 = *K_nu * (nu + x + 0.5 - hi)/x;
770 *Kp_nu = - *K_nup1 + nu/x * *K_nu;
772 GSL_ERROR ("error", GSL_EMAXITER);
778 int gsl_sf_bessel_cos_pi4_e(double y, double eps, gsl_sf_result * result)
780 const double sy = sin(y);
781 const double cy = cos(y);
782 const double s = sy + cy;
783 const double d = sy - cy;
784 const double abs_sum = fabs(cy) + fabs(sy);
787 if(fabs(eps) < GSL_ROOT5_DBL_EPSILON) {
788 const double e2 = eps*eps;
789 seps = eps * (1.0 - e2/6.0 * (1.0 - e2/20.0));
790 ceps = 1.0 - e2/2.0 * (1.0 - e2/12.0);
796 result->val = (ceps * s - seps * d)/ M_SQRT2;
797 result->err = 2.0 * GSL_DBL_EPSILON * (fabs(ceps) + fabs(seps)) * abs_sum / M_SQRT2;
799 /* Try to account for error in evaluation of sin(y), cos(y).
800 * This is a little sticky because we don't really know
801 * how the library routines are doing their argument reduction.
802 * However, we will make a reasonable guess.
805 if(y > 1.0/GSL_DBL_EPSILON) {
806 result->err *= 0.5 * y;
808 else if(y > 1.0/GSL_SQRT_DBL_EPSILON) {
809 result->err *= 256.0 * y * GSL_SQRT_DBL_EPSILON;
816 int gsl_sf_bessel_sin_pi4_e(double y, double eps, gsl_sf_result * result)
818 const double sy = sin(y);
819 const double cy = cos(y);
820 const double s = sy + cy;
821 const double d = sy - cy;
822 const double abs_sum = fabs(cy) + fabs(sy);
825 if(fabs(eps) < GSL_ROOT5_DBL_EPSILON) {
826 const double e2 = eps*eps;
827 seps = eps * (1.0 - e2/6.0 * (1.0 - e2/20.0));
828 ceps = 1.0 - e2/2.0 * (1.0 - e2/12.0);
834 result->val = (ceps * d + seps * s)/ M_SQRT2;
835 result->err = 2.0 * GSL_DBL_EPSILON * (fabs(ceps) + fabs(seps)) * abs_sum / M_SQRT2;
837 /* Try to account for error in evaluation of sin(y), cos(y).
841 if(y > 1.0/GSL_DBL_EPSILON) {
842 result->err *= 0.5 * y;
844 else if(y > 1.0/GSL_SQRT_DBL_EPSILON) {
845 result->err *= 256.0 * y * GSL_SQRT_DBL_EPSILON;
852 /************************************************************************
854 Asymptotic approximations 8.11.5, 8.12.5, and 8.42.7 from
855 G.N.Watson, A Treatise on the Theory of Bessel Functions,
856 2nd Edition (Cambridge University Press, 1944).
857 Higher terms in expansion for x near l given by
858 Airey in Phil. Mag. 31, 520 (1916).
860 This approximation is accurate to near 0.1% at the boundaries
861 between the asymptotic regions; well away from the boundaries
862 the accuracy is better than 10^{-5}.
864 ************************************************************************/
866 double besselJ_meissel(double nu, double x)
868 double beta = pow(nu, 0.325);
871 /* Fitted matching points. */
872 double llimit = 1.1 * beta;
873 double ulimit = 1.3 * beta;
875 double nu2 = nu * nu;
877 if (nu < 5. && x < 1.)
879 /* Small argument and order. Use a Taylor expansion. */
881 double xo2 = 0.5 * x;
882 double gamfactor = pow(nu,nu) * exp(-nu) * sqrt(nu * 2. * M_PI)
883 * (1. + 1./(12.*nu) + 1./(288.*nu*nu));
884 double prefactor = pow(xo2, nu) / gamfactor;
888 C[1] = -C[0] / (nu+1.);
889 C[2] = -C[1] / (2.*(nu+2.));
890 C[3] = -C[2] / (3.*(nu+3.));
891 C[4] = -C[3] / (4.*(nu+4.));
895 result += C[k] * pow(xo2, 2.*k);
899 else if(x < nu - llimit)
901 /* Small x region: x << l. */
904 double rtomz2 = sqrt(1.-z2);
905 double omz2_2 = (1.-z2)*(1.-z2);
907 /* Calculate Meissel exponent. */
908 double term1 = 1./(24.*nu) * ((2.+3.*z2)/((1.-z2)*rtomz2) -2.);
909 double term2 = - z2*(4. + z2)/(16.*nu2*(1.-z2)*omz2_2);
910 double V_nu = term1 + term2;
912 /* Calculate the harmless prefactor. */
913 double sterlingsum = 1. + 1./(12.*nu) + 1./(288*nu2);
914 double harmless = 1. / (sqrt(rtomz2*2.*M_PI*nu) * sterlingsum);
916 /* Calculate the logarithm of the nu dependent prefactor. */
917 double ln_nupre = rtomz2 + log(z) - log(1. + rtomz2);
919 result = harmless * exp(nu*ln_nupre - V_nu);
921 else if(x < nu + ulimit)
923 /* Intermediate region 1: x near nu. */
924 double eps = 1.-nu/x;
925 double eps_x = eps * x;
926 double eps_x_2 = eps_x * eps_x;
929 static double gam[6] = {2.67894, 1.35412, 1., 0.89298, 0.902745, 1.};
930 static double sf[6] = {0.866025, 0.866025, 0., -0.866025, -0.866025, 0.};
932 /* Some terms are identically zero, because sf[] can be zero.
933 * Some terms do not appear in the result.
937 /* B[2] = 0.5 * eps_x_2 - 1./20.; */
938 B[3] = eps_x * (eps_x_2/6. - 1./15.);
939 B[4] = eps_x_2 * (eps_x_2 - 1.)/24. + 1./280.;
940 /* B[5] = eps_x * (eps_x_2*(0.5*eps_x_2 - 1.)/60. + 43./8400.); */
942 result = B[0] * gam[0] * sf[0] / pow(xo6, 1./3.);
943 result += B[1] * gam[1] * sf[1] / pow(xo6, 2./3.);
944 result += B[3] * gam[3] * sf[3] / pow(xo6, 4./3.);
945 result += B[4] * gam[4] * sf[4] / pow(xo6, 5./3.);
951 /* Region of very large argument. Use expansion
952 * for x>>l, and we need not be very exacting.
955 double sec2b= secb*secb;
957 double cotb = 1./sqrt(sec2b-1.); /* cotb=cot(beta) */
959 double beta = acos(nu/x);
960 double trigarg = nu/cotb - nu*beta - 0.25 * M_PI;
962 double cot3b = cotb * cotb * cotb;
963 double cot6b = cot3b * cot3b;
965 double sum1, sum2, expterm, prefactor, trigcos;
967 sum1 = 2.0 + 3.0 * sec2b;
968 trigarg -= sum1 * cot3b / (24.0 * nu);
970 trigcos = cos(trigarg);
973 expterm = sum2 * sec2b * cot6b / (16.0 * nu2);
975 expterm = exp(-expterm);
976 prefactor = sqrt(2. * cotb / (nu * M_PI));
978 result = prefactor * expterm * trigcos;