14 typedef long long BLASLONG;
15 typedef unsigned long long BLASULONG;
17 typedef long BLASLONG;
18 typedef unsigned long BLASULONG;
22 typedef BLASLONG blasint;
24 #define blasabs(x) llabs(x)
26 #define blasabs(x) labs(x)
30 #define blasabs(x) abs(x)
33 typedef blasint integer;
35 typedef unsigned int uinteger;
36 typedef char *address;
37 typedef short int shortint;
39 typedef double doublereal;
40 typedef struct { real r, i; } complex;
41 typedef struct { doublereal r, i; } doublecomplex;
43 static inline _Fcomplex Cf(complex *z) {_Fcomplex zz={z->r , z->i}; return zz;}
44 static inline _Dcomplex Cd(doublecomplex *z) {_Dcomplex zz={z->r , z->i};return zz;}
45 static inline _Fcomplex * _pCf(complex *z) {return (_Fcomplex*)z;}
46 static inline _Dcomplex * _pCd(doublecomplex *z) {return (_Dcomplex*)z;}
48 static inline _Complex float Cf(complex *z) {return z->r + z->i*_Complex_I;}
49 static inline _Complex double Cd(doublecomplex *z) {return z->r + z->i*_Complex_I;}
50 static inline _Complex float * _pCf(complex *z) {return (_Complex float*)z;}
51 static inline _Complex double * _pCd(doublecomplex *z) {return (_Complex double*)z;}
53 #define pCf(z) (*_pCf(z))
54 #define pCd(z) (*_pCd(z))
56 typedef short int shortlogical;
57 typedef char logical1;
58 typedef char integer1;
63 /* Extern is for use with -E */
74 /*external read, write*/
83 /*internal read, write*/
113 /*rewind, backspace, endfile*/
125 ftnint *inex; /*parameters in standard's order*/
151 union Multitype { /* for multiple entry points */
162 typedef union Multitype Multitype;
164 struct Vardesc { /* for Namelist */
170 typedef struct Vardesc Vardesc;
177 typedef struct Namelist Namelist;
179 #define abs(x) ((x) >= 0 ? (x) : -(x))
180 #define dabs(x) (fabs(x))
181 #define f2cmin(a,b) ((a) <= (b) ? (a) : (b))
182 #define f2cmax(a,b) ((a) >= (b) ? (a) : (b))
183 #define dmin(a,b) (f2cmin(a,b))
184 #define dmax(a,b) (f2cmax(a,b))
185 #define bit_test(a,b) ((a) >> (b) & 1)
186 #define bit_clear(a,b) ((a) & ~((uinteger)1 << (b)))
187 #define bit_set(a,b) ((a) | ((uinteger)1 << (b)))
189 #define abort_() { sig_die("Fortran abort routine called", 1); }
190 #define c_abs(z) (cabsf(Cf(z)))
191 #define c_cos(R,Z) { pCf(R)=ccos(Cf(Z)); }
193 #define c_div(c, a, b) {Cf(c)._Val[0] = (Cf(a)._Val[0]/Cf(b)._Val[0]); Cf(c)._Val[1]=(Cf(a)._Val[1]/Cf(b)._Val[1]);}
194 #define z_div(c, a, b) {Cd(c)._Val[0] = (Cd(a)._Val[0]/Cd(b)._Val[0]); Cd(c)._Val[1]=(Cd(a)._Val[1]/df(b)._Val[1]);}
196 #define c_div(c, a, b) {pCf(c) = Cf(a)/Cf(b);}
197 #define z_div(c, a, b) {pCd(c) = Cd(a)/Cd(b);}
199 #define c_exp(R, Z) {pCf(R) = cexpf(Cf(Z));}
200 #define c_log(R, Z) {pCf(R) = clogf(Cf(Z));}
201 #define c_sin(R, Z) {pCf(R) = csinf(Cf(Z));}
202 //#define c_sqrt(R, Z) {*(R) = csqrtf(Cf(Z));}
203 #define c_sqrt(R, Z) {pCf(R) = csqrtf(Cf(Z));}
204 #define d_abs(x) (fabs(*(x)))
205 #define d_acos(x) (acos(*(x)))
206 #define d_asin(x) (asin(*(x)))
207 #define d_atan(x) (atan(*(x)))
208 #define d_atn2(x, y) (atan2(*(x),*(y)))
209 #define d_cnjg(R, Z) { pCd(R) = conj(Cd(Z)); }
210 #define r_cnjg(R, Z) { pCf(R) = conjf(Cf(Z)); }
211 #define d_cos(x) (cos(*(x)))
212 #define d_cosh(x) (cosh(*(x)))
213 #define d_dim(__a, __b) ( *(__a) > *(__b) ? *(__a) - *(__b) : 0.0 )
214 #define d_exp(x) (exp(*(x)))
215 #define d_imag(z) (cimag(Cd(z)))
216 #define r_imag(z) (cimagf(Cf(z)))
217 #define d_int(__x) (*(__x)>0 ? floor(*(__x)) : -floor(- *(__x)))
218 #define r_int(__x) (*(__x)>0 ? floor(*(__x)) : -floor(- *(__x)))
219 #define d_lg10(x) ( 0.43429448190325182765 * log(*(x)) )
220 #define r_lg10(x) ( 0.43429448190325182765 * log(*(x)) )
221 #define d_log(x) (log(*(x)))
222 #define d_mod(x, y) (fmod(*(x), *(y)))
223 #define u_nint(__x) ((__x)>=0 ? floor((__x) + .5) : -floor(.5 - (__x)))
224 #define d_nint(x) u_nint(*(x))
225 #define u_sign(__a,__b) ((__b) >= 0 ? ((__a) >= 0 ? (__a) : -(__a)) : -((__a) >= 0 ? (__a) : -(__a)))
226 #define d_sign(a,b) u_sign(*(a),*(b))
227 #define r_sign(a,b) u_sign(*(a),*(b))
228 #define d_sin(x) (sin(*(x)))
229 #define d_sinh(x) (sinh(*(x)))
230 #define d_sqrt(x) (sqrt(*(x)))
231 #define d_tan(x) (tan(*(x)))
232 #define d_tanh(x) (tanh(*(x)))
233 #define i_abs(x) abs(*(x))
234 #define i_dnnt(x) ((integer)u_nint(*(x)))
235 #define i_len(s, n) (n)
236 #define i_nint(x) ((integer)u_nint(*(x)))
237 #define i_sign(a,b) ((integer)u_sign((integer)*(a),(integer)*(b)))
238 #define pow_dd(ap, bp) ( pow(*(ap), *(bp)))
239 #define pow_si(B,E) spow_ui(*(B),*(E))
240 #define pow_ri(B,E) spow_ui(*(B),*(E))
241 #define pow_di(B,E) dpow_ui(*(B),*(E))
242 #define pow_zi(p, a, b) {pCd(p) = zpow_ui(Cd(a), *(b));}
243 #define pow_ci(p, a, b) {pCf(p) = cpow_ui(Cf(a), *(b));}
244 #define pow_zz(R,A,B) {pCd(R) = cpow(Cd(A),*(B));}
245 #define s_cat(lpp, rpp, rnp, np, llp) { ftnlen i, nc, ll; char *f__rp, *lp; ll = (llp); lp = (lpp); for(i=0; i < (int)*(np); ++i) { nc = ll; if((rnp)[i] < nc) nc = (rnp)[i]; ll -= nc; f__rp = (rpp)[i]; while(--nc >= 0) *lp++ = *(f__rp)++; } while(--ll >= 0) *lp++ = ' '; }
246 #define s_cmp(a,b,c,d) ((integer)strncmp((a),(b),f2cmin((c),(d))))
247 #define s_copy(A,B,C,D) { int __i,__m; for (__i=0, __m=f2cmin((C),(D)); __i<__m && (B)[__i] != 0; ++__i) (A)[__i] = (B)[__i]; }
248 #define sig_die(s, kill) { exit(1); }
249 #define s_stop(s, n) {exit(0);}
250 static char junk[] = "\n@(#)LIBF77 VERSION 19990503\n";
251 #define z_abs(z) (cabs(Cd(z)))
252 #define z_exp(R, Z) {pCd(R) = cexp(Cd(Z));}
253 #define z_sqrt(R, Z) {pCd(R) = csqrt(Cd(Z));}
254 #define myexit_() break;
255 #define mycycle() continue;
256 #define myceiling(w) {ceil(w)}
257 #define myhuge(w) {HUGE_VAL}
258 //#define mymaxloc_(w,s,e,n) {if (sizeof(*(w)) == sizeof(double)) dmaxloc_((w),*(s),*(e),n); else dmaxloc_((w),*(s),*(e),n);}
259 #define mymaxloc(w,s,e,n) {dmaxloc_(w,*(s),*(e),n)}
261 /* procedure parameter types for -A and -C++ */
263 #define F2C_proc_par_types 1
265 typedef logical (*L_fp)(...);
267 typedef logical (*L_fp)();
270 static float spow_ui(float x, integer n) {
271 float pow=1.0; unsigned long int u;
273 if(n < 0) n = -n, x = 1/x;
282 static double dpow_ui(double x, integer n) {
283 double pow=1.0; unsigned long int u;
285 if(n < 0) n = -n, x = 1/x;
295 static _Fcomplex cpow_ui(complex x, integer n) {
296 complex pow={1.0,0.0}; unsigned long int u;
298 if(n < 0) n = -n, x.r = 1/x.r, x.i=1/x.i;
300 if(u & 01) pow.r *= x.r, pow.i *= x.i;
301 if(u >>= 1) x.r *= x.r, x.i *= x.i;
305 _Fcomplex p={pow.r, pow.i};
309 static _Complex float cpow_ui(_Complex float x, integer n) {
310 _Complex float pow=1.0; unsigned long int u;
312 if(n < 0) n = -n, x = 1/x;
323 static _Dcomplex zpow_ui(_Dcomplex x, integer n) {
324 _Dcomplex pow={1.0,0.0}; unsigned long int u;
326 if(n < 0) n = -n, x._Val[0] = 1/x._Val[0], x._Val[1] =1/x._Val[1];
328 if(u & 01) pow._Val[0] *= x._Val[0], pow._Val[1] *= x._Val[1];
329 if(u >>= 1) x._Val[0] *= x._Val[0], x._Val[1] *= x._Val[1];
333 _Dcomplex p = {pow._Val[0], pow._Val[1]};
337 static _Complex double zpow_ui(_Complex double x, integer n) {
338 _Complex double pow=1.0; unsigned long int u;
340 if(n < 0) n = -n, x = 1/x;
350 static integer pow_ii(integer x, integer n) {
351 integer pow; unsigned long int u;
353 if (n == 0 || x == 1) pow = 1;
354 else if (x != -1) pow = x == 0 ? 1/x : 0;
357 if ((n > 0) || !(n == 0 || x == 1 || x != -1)) {
367 static integer dmaxloc_(double *w, integer s, integer e, integer *n)
369 double m; integer i, mi;
370 for(m=w[s-1], mi=s, i=s+1; i<=e; i++)
371 if (w[i-1]>m) mi=i ,m=w[i-1];
374 static integer smaxloc_(float *w, integer s, integer e, integer *n)
376 float m; integer i, mi;
377 for(m=w[s-1], mi=s, i=s+1; i<=e; i++)
378 if (w[i-1]>m) mi=i ,m=w[i-1];
381 static inline void cdotc_(complex *z, integer *n_, complex *x, integer *incx_, complex *y, integer *incy_) {
382 integer n = *n_, incx = *incx_, incy = *incy_, i;
384 _Fcomplex zdotc = {0.0, 0.0};
385 if (incx == 1 && incy == 1) {
386 for (i=0;i<n;i++) { /* zdotc = zdotc + dconjg(x(i))* y(i) */
387 zdotc._Val[0] += conjf(Cf(&x[i]))._Val[0] * Cf(&y[i])._Val[0];
388 zdotc._Val[1] += conjf(Cf(&x[i]))._Val[1] * Cf(&y[i])._Val[1];
391 for (i=0;i<n;i++) { /* zdotc = zdotc + dconjg(x(i))* y(i) */
392 zdotc._Val[0] += conjf(Cf(&x[i*incx]))._Val[0] * Cf(&y[i*incy])._Val[0];
393 zdotc._Val[1] += conjf(Cf(&x[i*incx]))._Val[1] * Cf(&y[i*incy])._Val[1];
399 _Complex float zdotc = 0.0;
400 if (incx == 1 && incy == 1) {
401 for (i=0;i<n;i++) { /* zdotc = zdotc + dconjg(x(i))* y(i) */
402 zdotc += conjf(Cf(&x[i])) * Cf(&y[i]);
405 for (i=0;i<n;i++) { /* zdotc = zdotc + dconjg(x(i))* y(i) */
406 zdotc += conjf(Cf(&x[i*incx])) * Cf(&y[i*incy]);
412 static inline void zdotc_(doublecomplex *z, integer *n_, doublecomplex *x, integer *incx_, doublecomplex *y, integer *incy_) {
413 integer n = *n_, incx = *incx_, incy = *incy_, i;
415 _Dcomplex zdotc = {0.0, 0.0};
416 if (incx == 1 && incy == 1) {
417 for (i=0;i<n;i++) { /* zdotc = zdotc + dconjg(x(i))* y(i) */
418 zdotc._Val[0] += conj(Cd(&x[i]))._Val[0] * Cd(&y[i])._Val[0];
419 zdotc._Val[1] += conj(Cd(&x[i]))._Val[1] * Cd(&y[i])._Val[1];
422 for (i=0;i<n;i++) { /* zdotc = zdotc + dconjg(x(i))* y(i) */
423 zdotc._Val[0] += conj(Cd(&x[i*incx]))._Val[0] * Cd(&y[i*incy])._Val[0];
424 zdotc._Val[1] += conj(Cd(&x[i*incx]))._Val[1] * Cd(&y[i*incy])._Val[1];
430 _Complex double zdotc = 0.0;
431 if (incx == 1 && incy == 1) {
432 for (i=0;i<n;i++) { /* zdotc = zdotc + dconjg(x(i))* y(i) */
433 zdotc += conj(Cd(&x[i])) * Cd(&y[i]);
436 for (i=0;i<n;i++) { /* zdotc = zdotc + dconjg(x(i))* y(i) */
437 zdotc += conj(Cd(&x[i*incx])) * Cd(&y[i*incy]);
443 static inline void cdotu_(complex *z, integer *n_, complex *x, integer *incx_, complex *y, integer *incy_) {
444 integer n = *n_, incx = *incx_, incy = *incy_, i;
446 _Fcomplex zdotc = {0.0, 0.0};
447 if (incx == 1 && incy == 1) {
448 for (i=0;i<n;i++) { /* zdotc = zdotc + dconjg(x(i))* y(i) */
449 zdotc._Val[0] += Cf(&x[i])._Val[0] * Cf(&y[i])._Val[0];
450 zdotc._Val[1] += Cf(&x[i])._Val[1] * Cf(&y[i])._Val[1];
453 for (i=0;i<n;i++) { /* zdotc = zdotc + dconjg(x(i))* y(i) */
454 zdotc._Val[0] += Cf(&x[i*incx])._Val[0] * Cf(&y[i*incy])._Val[0];
455 zdotc._Val[1] += Cf(&x[i*incx])._Val[1] * Cf(&y[i*incy])._Val[1];
461 _Complex float zdotc = 0.0;
462 if (incx == 1 && incy == 1) {
463 for (i=0;i<n;i++) { /* zdotc = zdotc + dconjg(x(i))* y(i) */
464 zdotc += Cf(&x[i]) * Cf(&y[i]);
467 for (i=0;i<n;i++) { /* zdotc = zdotc + dconjg(x(i))* y(i) */
468 zdotc += Cf(&x[i*incx]) * Cf(&y[i*incy]);
474 static inline void zdotu_(doublecomplex *z, integer *n_, doublecomplex *x, integer *incx_, doublecomplex *y, integer *incy_) {
475 integer n = *n_, incx = *incx_, incy = *incy_, i;
477 _Dcomplex zdotc = {0.0, 0.0};
478 if (incx == 1 && incy == 1) {
479 for (i=0;i<n;i++) { /* zdotc = zdotc + dconjg(x(i))* y(i) */
480 zdotc._Val[0] += Cd(&x[i])._Val[0] * Cd(&y[i])._Val[0];
481 zdotc._Val[1] += Cd(&x[i])._Val[1] * Cd(&y[i])._Val[1];
484 for (i=0;i<n;i++) { /* zdotc = zdotc + dconjg(x(i))* y(i) */
485 zdotc._Val[0] += Cd(&x[i*incx])._Val[0] * Cd(&y[i*incy])._Val[0];
486 zdotc._Val[1] += Cd(&x[i*incx])._Val[1] * Cd(&y[i*incy])._Val[1];
492 _Complex double zdotc = 0.0;
493 if (incx == 1 && incy == 1) {
494 for (i=0;i<n;i++) { /* zdotc = zdotc + dconjg(x(i))* y(i) */
495 zdotc += Cd(&x[i]) * Cd(&y[i]);
498 for (i=0;i<n;i++) { /* zdotc = zdotc + dconjg(x(i))* y(i) */
499 zdotc += Cd(&x[i*incx]) * Cd(&y[i*incy]);
505 /* -- translated by f2c (version 20000121).
506 You must link the resulting object file with the libraries:
507 -lf2c -lm (in that order)
513 /* > \brief \b SLASD4 computes the square root of the i-th updated eigenvalue of a positive symmetric rank-one
514 modification to a positive diagonal matrix. Used by sbdsdc. */
516 /* =========== DOCUMENTATION =========== */
518 /* Online html documentation available at */
519 /* http://www.netlib.org/lapack/explore-html/ */
522 /* > Download SLASD4 + dependencies */
523 /* > <a href="http://www.netlib.org/cgi-bin/netlibfiles.tgz?format=tgz&filename=/lapack/lapack_routine/slasd4.
526 /* > <a href="http://www.netlib.org/cgi-bin/netlibfiles.zip?format=zip&filename=/lapack/lapack_routine/slasd4.
529 /* > <a href="http://www.netlib.org/cgi-bin/netlibfiles.txt?format=txt&filename=/lapack/lapack_routine/slasd4.
537 /* SUBROUTINE SLASD4( N, I, D, Z, DELTA, RHO, SIGMA, WORK, INFO ) */
539 /* INTEGER I, INFO, N */
540 /* REAL RHO, SIGMA */
541 /* REAL D( * ), DELTA( * ), WORK( * ), Z( * ) */
544 /* > \par Purpose: */
549 /* > This subroutine computes the square root of the I-th updated */
550 /* > eigenvalue of a positive symmetric rank-one modification to */
551 /* > a positive diagonal matrix whose entries are given as the squares */
552 /* > of the corresponding entries in the array d, and that */
554 /* > 0 <= D(i) < D(j) for i < j */
556 /* > and that RHO > 0. This is arranged by the calling routine, and is */
557 /* > no loss in generality. The rank-one modified system is thus */
559 /* > diag( D ) * diag( D ) + RHO * Z * Z_transpose. */
561 /* > where we assume the Euclidean norm of Z is 1. */
563 /* > The method consists of approximating the rational functions in the */
564 /* > secular equation by simpler interpolating rational functions. */
573 /* > The length of all arrays. */
579 /* > The index of the eigenvalue to be computed. 1 <= I <= N. */
584 /* > D is REAL array, dimension ( N ) */
585 /* > The original eigenvalues. It is assumed that they are in */
586 /* > order, 0 <= D(I) < D(J) for I < J. */
591 /* > Z is REAL array, dimension ( N ) */
592 /* > The components of the updating vector. */
595 /* > \param[out] DELTA */
597 /* > DELTA is REAL array, dimension ( N ) */
598 /* > If N .ne. 1, DELTA contains (D(j) - sigma_I) in its j-th */
599 /* > component. If N = 1, then DELTA(1) = 1. The vector DELTA */
600 /* > contains the information necessary to construct the */
601 /* > (singular) eigenvectors. */
604 /* > \param[in] RHO */
607 /* > The scalar in the symmetric updating formula. */
610 /* > \param[out] SIGMA */
612 /* > SIGMA is REAL */
613 /* > The computed sigma_I, the I-th updated eigenvalue. */
616 /* > \param[out] WORK */
618 /* > WORK is REAL array, dimension ( N ) */
619 /* > If N .ne. 1, WORK contains (D(j) + sigma_I) in its j-th */
620 /* > component. If N = 1, then WORK( 1 ) = 1. */
623 /* > \param[out] INFO */
625 /* > INFO is INTEGER */
626 /* > = 0: successful exit */
627 /* > > 0: if INFO = 1, the updating process failed. */
630 /* > \par Internal Parameters: */
631 /* ========================= */
634 /* > Logical variable ORGATI (origin-at-i?) is used for distinguishing */
635 /* > whether D(i) or D(i+1) is treated as the origin. */
637 /* > ORGATI = .true. origin at i */
638 /* > ORGATI = .false. origin at i+1 */
640 /* > Logical variable SWTCH3 (switch-for-3-poles?) is for noting */
641 /* > if we are working with THREE poles! */
643 /* > MAXIT is the maximum number of iterations allowed for each */
650 /* > \author Univ. of Tennessee */
651 /* > \author Univ. of California Berkeley */
652 /* > \author Univ. of Colorado Denver */
653 /* > \author NAG Ltd. */
655 /* > \date December 2016 */
657 /* > \ingroup OTHERauxiliary */
659 /* > \par Contributors: */
660 /* ================== */
662 /* > Ren-Cang Li, Computer Science Division, University of California */
663 /* > at Berkeley, USA */
665 /* ===================================================================== */
666 /* Subroutine */ int slasd4_(integer *n, integer *i__, real *d__, real *z__,
667 real *delta, real *rho, real *sigma, real *work, integer *info)
669 /* System generated locals */
673 /* Local variables */
674 real dphi, sglb, dpsi, sgub;
676 real temp, prew, temp1, temp2, a, b, c__;
678 real w, dtiim, delsq, dtiip;
683 extern /* Subroutine */ int slaed6_(integer *, logical *, real *, real *,
684 real *, real *, real *, integer *);
686 extern /* Subroutine */ int slasd5_(integer *, real *, real *, real *,
687 real *, real *, real *);
692 extern real slamch_(char *);
695 real erretm, dtipsq, rhoinv;
697 real sq2, eta, phi, eps, tau, psi;
703 /* -- LAPACK auxiliary routine (version 3.7.0) -- */
704 /* -- LAPACK is a software package provided by Univ. of Tennessee, -- */
705 /* -- Univ. of California Berkeley, Univ. of Colorado Denver and NAG Ltd..-- */
709 /* ===================================================================== */
712 /* Since this routine is called in an inner loop, we do no argument */
715 /* Quick return for N=1 and 2. */
717 /* Parameter adjustments */
727 /* Presumably, I=1 upon entry */
729 *sigma = sqrt(d__[1] * d__[1] + *rho * z__[1] * z__[1]);
735 slasd5_(i__, &d__[1], &z__[1], &delta[1], rho, sigma, &work[1]);
739 /* Compute machine epsilon */
741 eps = slamch_("Epsilon");
749 /* Initialize some basic variables */
754 /* Calculate initial guess */
758 /* If ||Z||_2 is not one, then TEMP should be set to */
759 /* RHO * ||Z||_2^2 / TWO */
761 temp1 = temp / (d__[*n] + sqrt(d__[*n] * d__[*n] + temp));
763 for (j = 1; j <= i__1; ++j) {
764 work[j] = d__[j] + d__[*n] + temp1;
765 delta[j] = d__[j] - d__[*n] - temp1;
771 for (j = 1; j <= i__1; ++j) {
772 psi += z__[j] * z__[j] / (delta[j] * work[j]);
777 w = c__ + z__[ii] * z__[ii] / (delta[ii] * work[ii]) + z__[*n] * z__[*
778 n] / (delta[*n] * work[*n]);
781 temp1 = sqrt(d__[*n] * d__[*n] + *rho);
782 temp = z__[*n - 1] * z__[*n - 1] / ((d__[*n - 1] + temp1) * (d__[*
783 n] - d__[*n - 1] + *rho / (d__[*n] + temp1))) + z__[*n] *
786 /* The following TAU2 is to approximate */
787 /* SIGMA_n^2 - D( N )*D( N ) */
792 delsq = (d__[*n] - d__[*n - 1]) * (d__[*n] + d__[*n - 1]);
793 a = -c__ * delsq + z__[*n - 1] * z__[*n - 1] + z__[*n] * z__[*
795 b = z__[*n] * z__[*n] * delsq;
797 tau2 = b * 2.f / (sqrt(a * a + b * 4.f * c__) - a);
799 tau2 = (a + sqrt(a * a + b * 4.f * c__)) / (c__ * 2.f);
801 tau = tau2 / (d__[*n] + sqrt(d__[*n] * d__[*n] + tau2));
804 /* It can be proved that */
805 /* D(N)^2+RHO/2 <= SIGMA_n^2 < D(N)^2+TAU2 <= D(N)^2+RHO */
808 delsq = (d__[*n] - d__[*n - 1]) * (d__[*n] + d__[*n - 1]);
809 a = -c__ * delsq + z__[*n - 1] * z__[*n - 1] + z__[*n] * z__[*n];
810 b = z__[*n] * z__[*n] * delsq;
812 /* The following TAU2 is to approximate */
813 /* SIGMA_n^2 - D( N )*D( N ) */
816 tau2 = b * 2.f / (sqrt(a * a + b * 4.f * c__) - a);
818 tau2 = (a + sqrt(a * a + b * 4.f * c__)) / (c__ * 2.f);
820 tau = tau2 / (d__[*n] + sqrt(d__[*n] * d__[*n] + tau2));
822 /* It can be proved that */
823 /* D(N)^2 < D(N)^2+TAU2 < SIGMA(N)^2 < D(N)^2+RHO/2 */
827 /* The following TAU is to approximate SIGMA_n - D( N ) */
829 /* TAU = TAU2 / ( D( N )+SQRT( D( N )*D( N )+TAU2 ) ) */
831 *sigma = d__[*n] + tau;
833 for (j = 1; j <= i__1; ++j) {
834 delta[j] = d__[j] - d__[*n] - tau;
835 work[j] = d__[j] + d__[*n] + tau;
839 /* Evaluate PSI and the derivative DPSI */
845 for (j = 1; j <= i__1; ++j) {
846 temp = z__[j] / (delta[j] * work[j]);
847 psi += z__[j] * temp;
852 erretm = abs(erretm);
854 /* Evaluate PHI and the derivative DPHI */
856 temp = z__[*n] / (delta[*n] * work[*n]);
857 phi = z__[*n] * temp;
859 erretm = (-phi - psi) * 8.f + erretm - phi + rhoinv;
860 /* $ + ABS( TAU2 )*( DPSI+DPHI ) */
862 w = rhoinv + phi + psi;
864 /* Test for convergence */
866 if (abs(w) <= eps * erretm) {
870 /* Calculate the new step */
873 dtnsq1 = work[*n - 1] * delta[*n - 1];
874 dtnsq = work[*n] * delta[*n];
875 c__ = w - dtnsq1 * dpsi - dtnsq * dphi;
876 a = (dtnsq + dtnsq1) * w - dtnsq * dtnsq1 * (dpsi + dphi);
877 b = dtnsq * dtnsq1 * w;
882 eta = *rho - *sigma * *sigma;
883 } else if (a >= 0.f) {
884 eta = (a + sqrt((r__1 = a * a - b * 4.f * c__, abs(r__1)))) / (
887 eta = b * 2.f / (a - sqrt((r__1 = a * a - b * 4.f * c__, abs(r__1)
891 /* Note, eta should be positive if w is negative, and */
892 /* eta should be negative otherwise. However, */
893 /* if for some reason caused by roundoff, eta*w > 0, */
894 /* we simply use one Newton step instead. This way */
895 /* will guarantee eta*w < 0. */
898 eta = -w / (dpsi + dphi);
905 eta /= *sigma + sqrt(eta + *sigma * *sigma);
910 for (j = 1; j <= i__1; ++j) {
916 /* Evaluate PSI and the derivative DPSI */
922 for (j = 1; j <= i__1; ++j) {
923 temp = z__[j] / (work[j] * delta[j]);
924 psi += z__[j] * temp;
929 erretm = abs(erretm);
931 /* Evaluate PHI and the derivative DPHI */
933 tau2 = work[*n] * delta[*n];
934 temp = z__[*n] / tau2;
935 phi = z__[*n] * temp;
937 erretm = (-phi - psi) * 8.f + erretm - phi + rhoinv;
938 /* $ + ABS( TAU2 )*( DPSI+DPHI ) */
940 w = rhoinv + phi + psi;
942 /* Main loop to update the values of the array DELTA */
946 for (niter = iter; niter <= 400; ++niter) {
948 /* Test for convergence */
950 if (abs(w) <= eps * erretm) {
954 /* Calculate the new step */
956 dtnsq1 = work[*n - 1] * delta[*n - 1];
957 dtnsq = work[*n] * delta[*n];
958 c__ = w - dtnsq1 * dpsi - dtnsq * dphi;
959 a = (dtnsq + dtnsq1) * w - dtnsq1 * dtnsq * (dpsi + dphi);
960 b = dtnsq1 * dtnsq * w;
962 eta = (a + sqrt((r__1 = a * a - b * 4.f * c__, abs(r__1)))) /
965 eta = b * 2.f / (a - sqrt((r__1 = a * a - b * 4.f * c__, abs(
969 /* Note, eta should be positive if w is negative, and */
970 /* eta should be negative otherwise. However, */
971 /* if for some reason caused by roundoff, eta*w > 0, */
972 /* we simply use one Newton step instead. This way */
973 /* will guarantee eta*w < 0. */
976 eta = -w / (dpsi + dphi);
983 eta /= *sigma + sqrt(eta + *sigma * *sigma);
988 for (j = 1; j <= i__1; ++j) {
994 /* Evaluate PSI and the derivative DPSI */
1000 for (j = 1; j <= i__1; ++j) {
1001 temp = z__[j] / (work[j] * delta[j]);
1002 psi += z__[j] * temp;
1003 dpsi += temp * temp;
1007 erretm = abs(erretm);
1009 /* Evaluate PHI and the derivative DPHI */
1011 tau2 = work[*n] * delta[*n];
1012 temp = z__[*n] / tau2;
1013 phi = z__[*n] * temp;
1015 erretm = (-phi - psi) * 8.f + erretm - phi + rhoinv;
1016 /* $ + ABS( TAU2 )*( DPSI+DPHI ) */
1018 w = rhoinv + phi + psi;
1022 /* Return with INFO = 1, NITER = MAXIT and not converged */
1027 /* End for the case I = N */
1031 /* The case for I < N */
1036 /* Calculate initial guess */
1038 delsq = (d__[ip1] - d__[*i__]) * (d__[ip1] + d__[*i__]);
1039 delsq2 = delsq / 2.f;
1040 sq2 = sqrt((d__[*i__] * d__[*i__] + d__[ip1] * d__[ip1]) / 2.f);
1041 temp = delsq2 / (d__[*i__] + sq2);
1043 for (j = 1; j <= i__1; ++j) {
1044 work[j] = d__[j] + d__[*i__] + temp;
1045 delta[j] = d__[j] - d__[*i__] - temp;
1051 for (j = 1; j <= i__1; ++j) {
1052 psi += z__[j] * z__[j] / (work[j] * delta[j]);
1058 for (j = *n; j >= i__1; --j) {
1059 phi += z__[j] * z__[j] / (work[j] * delta[j]);
1062 c__ = rhoinv + psi + phi;
1063 w = c__ + z__[*i__] * z__[*i__] / (work[*i__] * delta[*i__]) + z__[
1064 ip1] * z__[ip1] / (work[ip1] * delta[ip1]);
1069 /* d(i)^2 < the ith sigma^2 < (d(i)^2+d(i+1)^2)/2 */
1071 /* We choose d(i) as origin. */
1076 sgub = delsq2 / (d__[*i__] + sq2);
1077 a = c__ * delsq + z__[*i__] * z__[*i__] + z__[ip1] * z__[ip1];
1078 b = z__[*i__] * z__[*i__] * delsq;
1080 tau2 = b * 2.f / (a + sqrt((r__1 = a * a - b * 4.f * c__, abs(
1083 tau2 = (a - sqrt((r__1 = a * a - b * 4.f * c__, abs(r__1)))) /
1087 /* TAU2 now is an estimation of SIGMA^2 - D( I )^2. The */
1088 /* following, however, is the corresponding estimation of */
1089 /* SIGMA - D( I ). */
1091 tau = tau2 / (d__[*i__] + sqrt(d__[*i__] * d__[*i__] + tau2));
1093 if (d__[*i__] <= temp * d__[ip1] && (r__1 = z__[*i__], abs(r__1))
1094 <= temp && d__[*i__] > 0.f) {
1096 r__1 = d__[*i__] * 10.f;
1097 tau = f2cmin(r__1,sgub);
1102 /* (d(i)^2+d(i+1)^2)/2 <= the ith sigma^2 < d(i+1)^2/2 */
1104 /* We choose d(i+1) as origin. */
1108 sglb = -delsq2 / (d__[ii] + sq2);
1110 a = c__ * delsq - z__[*i__] * z__[*i__] - z__[ip1] * z__[ip1];
1111 b = z__[ip1] * z__[ip1] * delsq;
1113 tau2 = b * 2.f / (a - sqrt((r__1 = a * a + b * 4.f * c__, abs(
1116 tau2 = -(a + sqrt((r__1 = a * a + b * 4.f * c__, abs(r__1))))
1120 /* TAU2 now is an estimation of SIGMA^2 - D( IP1 )^2. The */
1121 /* following, however, is the corresponding estimation of */
1122 /* SIGMA - D( IP1 ). */
1124 tau = tau2 / (d__[ip1] + sqrt((r__1 = d__[ip1] * d__[ip1] + tau2,
1128 *sigma = d__[ii] + tau;
1130 for (j = 1; j <= i__1; ++j) {
1131 work[j] = d__[j] + d__[ii] + tau;
1132 delta[j] = d__[j] - d__[ii] - tau;
1138 /* Evaluate PSI and the derivative DPSI */
1144 for (j = 1; j <= i__1; ++j) {
1145 temp = z__[j] / (work[j] * delta[j]);
1146 psi += z__[j] * temp;
1147 dpsi += temp * temp;
1151 erretm = abs(erretm);
1153 /* Evaluate PHI and the derivative DPHI */
1158 for (j = *n; j >= i__1; --j) {
1159 temp = z__[j] / (work[j] * delta[j]);
1160 phi += z__[j] * temp;
1161 dphi += temp * temp;
1166 w = rhoinv + phi + psi;
1168 /* W is the value of the secular function with */
1169 /* its ii-th element removed. */
1181 if (ii == 1 || ii == *n) {
1185 temp = z__[ii] / (work[ii] * delta[ii]);
1186 dw = dpsi + dphi + temp * temp;
1187 temp = z__[ii] * temp;
1189 erretm = (phi - psi) * 8.f + erretm + rhoinv * 2.f + abs(temp) * 3.f;
1190 /* $ + ABS( TAU2 )*DW */
1192 /* Test for convergence */
1194 if (abs(w) <= eps * erretm) {
1199 sglb = f2cmax(sglb,tau);
1201 sgub = f2cmin(sgub,tau);
1204 /* Calculate the new step */
1208 dtipsq = work[ip1] * delta[ip1];
1209 dtisq = work[*i__] * delta[*i__];
1211 /* Computing 2nd power */
1212 r__1 = z__[*i__] / dtisq;
1213 c__ = w - dtipsq * dw + delsq * (r__1 * r__1);
1215 /* Computing 2nd power */
1216 r__1 = z__[ip1] / dtipsq;
1217 c__ = w - dtisq * dw - delsq * (r__1 * r__1);
1219 a = (dtipsq + dtisq) * w - dtipsq * dtisq * dw;
1220 b = dtipsq * dtisq * w;
1224 a = z__[*i__] * z__[*i__] + dtipsq * dtipsq * (dpsi +
1227 a = z__[ip1] * z__[ip1] + dtisq * dtisq * (dpsi +
1232 } else if (a <= 0.f) {
1233 eta = (a - sqrt((r__1 = a * a - b * 4.f * c__, abs(r__1)))) /
1236 eta = b * 2.f / (a + sqrt((r__1 = a * a - b * 4.f * c__, abs(
1241 /* Interpolation using THREE most relevant poles */
1243 dtiim = work[iim1] * delta[iim1];
1244 dtiip = work[iip1] * delta[iip1];
1245 temp = rhoinv + psi + phi;
1247 temp1 = z__[iim1] / dtiim;
1249 c__ = temp - dtiip * (dpsi + dphi) - (d__[iim1] - d__[iip1]) *
1250 (d__[iim1] + d__[iip1]) * temp1;
1251 zz[0] = z__[iim1] * z__[iim1];
1253 zz[2] = dtiip * dtiip * dphi;
1255 zz[2] = dtiip * dtiip * (dpsi - temp1 + dphi);
1258 temp1 = z__[iip1] / dtiip;
1260 c__ = temp - dtiim * (dpsi + dphi) - (d__[iip1] - d__[iim1]) *
1261 (d__[iim1] + d__[iip1]) * temp1;
1263 zz[0] = dtiim * dtiim * dpsi;
1265 zz[0] = dtiim * dtiim * (dpsi + (dphi - temp1));
1267 zz[2] = z__[iip1] * z__[iip1];
1269 zz[1] = z__[ii] * z__[ii];
1271 dd[1] = delta[ii] * work[ii];
1273 slaed6_(&niter, &orgati, &c__, dd, zz, &w, &eta, info);
1277 /* If INFO is not 0, i.e., SLAED6 failed, switch back */
1278 /* to 2 pole interpolation. */
1282 dtipsq = work[ip1] * delta[ip1];
1283 dtisq = work[*i__] * delta[*i__];
1285 /* Computing 2nd power */
1286 r__1 = z__[*i__] / dtisq;
1287 c__ = w - dtipsq * dw + delsq * (r__1 * r__1);
1289 /* Computing 2nd power */
1290 r__1 = z__[ip1] / dtipsq;
1291 c__ = w - dtisq * dw - delsq * (r__1 * r__1);
1293 a = (dtipsq + dtisq) * w - dtipsq * dtisq * dw;
1294 b = dtipsq * dtisq * w;
1298 a = z__[*i__] * z__[*i__] + dtipsq * dtipsq * (
1301 a = z__[ip1] * z__[ip1] + dtisq * dtisq * (dpsi +
1306 } else if (a <= 0.f) {
1307 eta = (a - sqrt((r__1 = a * a - b * 4.f * c__, abs(r__1)))
1310 eta = b * 2.f / (a + sqrt((r__1 = a * a - b * 4.f * c__,
1316 /* Note, eta should be positive if w is negative, and */
1317 /* eta should be negative otherwise. However, */
1318 /* if for some reason caused by roundoff, eta*w > 0, */
1319 /* we simply use one Newton step instead. This way */
1320 /* will guarantee eta*w < 0. */
1322 if (w * eta >= 0.f) {
1326 eta /= *sigma + sqrt(*sigma * *sigma + eta);
1328 if (temp > sgub || temp < sglb) {
1330 eta = (sgub - tau) / 2.f;
1332 eta = (sglb - tau) / 2.f;
1337 eta = sqrt(sgub * tau) - tau;
1341 eta = sqrt(sglb * tau) - tau;
1353 for (j = 1; j <= i__1; ++j) {
1359 /* Evaluate PSI and the derivative DPSI */
1365 for (j = 1; j <= i__1; ++j) {
1366 temp = z__[j] / (work[j] * delta[j]);
1367 psi += z__[j] * temp;
1368 dpsi += temp * temp;
1372 erretm = abs(erretm);
1374 /* Evaluate PHI and the derivative DPHI */
1379 for (j = *n; j >= i__1; --j) {
1380 temp = z__[j] / (work[j] * delta[j]);
1381 phi += z__[j] * temp;
1382 dphi += temp * temp;
1387 tau2 = work[ii] * delta[ii];
1388 temp = z__[ii] / tau2;
1389 dw = dpsi + dphi + temp * temp;
1390 temp = z__[ii] * temp;
1391 w = rhoinv + phi + psi + temp;
1392 erretm = (phi - psi) * 8.f + erretm + rhoinv * 2.f + abs(temp) * 3.f;
1393 /* $ + ABS( TAU2 )*DW */
1397 if (-w > abs(prew) / 10.f) {
1401 if (w > abs(prew) / 10.f) {
1406 /* Main loop to update the values of the array DELTA and WORK */
1410 for (niter = iter; niter <= 400; ++niter) {
1412 /* Test for convergence */
1414 if (abs(w) <= eps * erretm) {
1415 /* $ .OR. (SGUB-SGLB).LE.EIGHT*ABS(SGUB+SGLB) ) THEN */
1420 sglb = f2cmax(sglb,tau);
1422 sgub = f2cmin(sgub,tau);
1425 /* Calculate the new step */
1428 dtipsq = work[ip1] * delta[ip1];
1429 dtisq = work[*i__] * delta[*i__];
1432 /* Computing 2nd power */
1433 r__1 = z__[*i__] / dtisq;
1434 c__ = w - dtipsq * dw + delsq * (r__1 * r__1);
1436 /* Computing 2nd power */
1437 r__1 = z__[ip1] / dtipsq;
1438 c__ = w - dtisq * dw - delsq * (r__1 * r__1);
1441 temp = z__[ii] / (work[ii] * delta[ii]);
1443 dpsi += temp * temp;
1445 dphi += temp * temp;
1447 c__ = w - dtisq * dpsi - dtipsq * dphi;
1449 a = (dtipsq + dtisq) * w - dtipsq * dtisq * dw;
1450 b = dtipsq * dtisq * w;
1455 a = z__[*i__] * z__[*i__] + dtipsq * dtipsq *
1458 a = z__[ip1] * z__[ip1] + dtisq * dtisq * (
1462 a = dtisq * dtisq * dpsi + dtipsq * dtipsq * dphi;
1466 } else if (a <= 0.f) {
1467 eta = (a - sqrt((r__1 = a * a - b * 4.f * c__, abs(r__1)))
1470 eta = b * 2.f / (a + sqrt((r__1 = a * a - b * 4.f * c__,
1475 /* Interpolation using THREE most relevant poles */
1477 dtiim = work[iim1] * delta[iim1];
1478 dtiip = work[iip1] * delta[iip1];
1479 temp = rhoinv + psi + phi;
1481 c__ = temp - dtiim * dpsi - dtiip * dphi;
1482 zz[0] = dtiim * dtiim * dpsi;
1483 zz[2] = dtiip * dtiip * dphi;
1486 temp1 = z__[iim1] / dtiim;
1488 temp2 = (d__[iim1] - d__[iip1]) * (d__[iim1] + d__[
1490 c__ = temp - dtiip * (dpsi + dphi) - temp2;
1491 zz[0] = z__[iim1] * z__[iim1];
1493 zz[2] = dtiip * dtiip * dphi;
1495 zz[2] = dtiip * dtiip * (dpsi - temp1 + dphi);
1498 temp1 = z__[iip1] / dtiip;
1500 temp2 = (d__[iip1] - d__[iim1]) * (d__[iim1] + d__[
1502 c__ = temp - dtiim * (dpsi + dphi) - temp2;
1504 zz[0] = dtiim * dtiim * dpsi;
1506 zz[0] = dtiim * dtiim * (dpsi + (dphi - temp1));
1508 zz[2] = z__[iip1] * z__[iip1];
1512 dd[1] = delta[ii] * work[ii];
1514 slaed6_(&niter, &orgati, &c__, dd, zz, &w, &eta, info);
1518 /* If INFO is not 0, i.e., SLAED6 failed, switch */
1519 /* back to two pole interpolation */
1523 dtipsq = work[ip1] * delta[ip1];
1524 dtisq = work[*i__] * delta[*i__];
1527 /* Computing 2nd power */
1528 r__1 = z__[*i__] / dtisq;
1529 c__ = w - dtipsq * dw + delsq * (r__1 * r__1);
1531 /* Computing 2nd power */
1532 r__1 = z__[ip1] / dtipsq;
1533 c__ = w - dtisq * dw - delsq * (r__1 * r__1);
1536 temp = z__[ii] / (work[ii] * delta[ii]);
1538 dpsi += temp * temp;
1540 dphi += temp * temp;
1542 c__ = w - dtisq * dpsi - dtipsq * dphi;
1544 a = (dtipsq + dtisq) * w - dtipsq * dtisq * dw;
1545 b = dtipsq * dtisq * w;
1550 a = z__[*i__] * z__[*i__] + dtipsq *
1551 dtipsq * (dpsi + dphi);
1553 a = z__[ip1] * z__[ip1] + dtisq * dtisq *
1557 a = dtisq * dtisq * dpsi + dtipsq * dtipsq *
1562 } else if (a <= 0.f) {
1563 eta = (a - sqrt((r__1 = a * a - b * 4.f * c__, abs(
1564 r__1)))) / (c__ * 2.f);
1566 eta = b * 2.f / (a + sqrt((r__1 = a * a - b * 4.f *
1572 /* Note, eta should be positive if w is negative, and */
1573 /* eta should be negative otherwise. However, */
1574 /* if for some reason caused by roundoff, eta*w > 0, */
1575 /* we simply use one Newton step instead. This way */
1576 /* will guarantee eta*w < 0. */
1578 if (w * eta >= 0.f) {
1582 eta /= *sigma + sqrt(*sigma * *sigma + eta);
1584 if (temp > sgub || temp < sglb) {
1586 eta = (sgub - tau) / 2.f;
1588 eta = (sglb - tau) / 2.f;
1593 eta = sqrt(sgub * tau) - tau;
1597 eta = sqrt(sglb * tau) - tau;
1609 for (j = 1; j <= i__1; ++j) {
1615 /* Evaluate PSI and the derivative DPSI */
1621 for (j = 1; j <= i__1; ++j) {
1622 temp = z__[j] / (work[j] * delta[j]);
1623 psi += z__[j] * temp;
1624 dpsi += temp * temp;
1628 erretm = abs(erretm);
1630 /* Evaluate PHI and the derivative DPHI */
1635 for (j = *n; j >= i__1; --j) {
1636 temp = z__[j] / (work[j] * delta[j]);
1637 phi += z__[j] * temp;
1638 dphi += temp * temp;
1643 tau2 = work[ii] * delta[ii];
1644 temp = z__[ii] / tau2;
1645 dw = dpsi + dphi + temp * temp;
1646 temp = z__[ii] * temp;
1647 w = rhoinv + phi + psi + temp;
1648 erretm = (phi - psi) * 8.f + erretm + rhoinv * 2.f + abs(temp) *
1650 /* $ + ABS( TAU2 )*DW */
1652 if (w * prew > 0.f && abs(w) > abs(prew) / 10.f) {
1659 /* Return with INFO = 1, NITER = MAXIT and not converged */