#include "clapack.h" /* Subroutine */ int slaed4_(integer *n, integer *i__, real *d__, real *z__, real *delta, real *rho, real *dlam, integer *info) { /* System generated locals */ integer i__1; real r__1; /* Builtin functions */ double sqrt(doublereal); /* Local variables */ real a, b, c__; integer j; real w; integer ii; real dw, zz[3]; integer ip1; real del, eta, phi, eps, tau, psi; integer iim1, iip1; real dphi, dpsi; integer iter; real temp, prew, temp1, dltlb, dltub, midpt; integer niter; logical swtch; extern /* Subroutine */ int slaed5_(integer *, real *, real *, real *, real *, real *), slaed6_(integer *, logical *, real *, real *, real *, real *, real *, integer *); logical swtch3; extern doublereal slamch_(char *); logical orgati; real erretm, rhoinv; /* -- LAPACK routine (version 3.1) -- */ /* Univ. of Tennessee, Univ. of California Berkeley and NAG Ltd.. */ /* November 2006 */ /* .. Scalar Arguments .. */ /* .. */ /* .. Array Arguments .. */ /* .. */ /* Purpose */ /* ======= */ /* This subroutine computes the I-th updated eigenvalue of a symmetric */ /* rank-one modification to a diagonal matrix whose elements are */ /* given in the array d, and that */ /* D(i) < D(j) for i < j */ /* and that RHO > 0. This is arranged by the calling routine, and is */ /* no loss in generality. The rank-one modified system is thus */ /* diag( D ) + RHO * Z * Z_transpose. */ /* where we assume the Euclidean norm of Z is 1. */ /* The method consists of approximating the rational functions in the */ /* secular equation by simpler interpolating rational functions. */ /* Arguments */ /* ========= */ /* N (input) INTEGER */ /* The length of all arrays. */ /* I (input) INTEGER */ /* The index of the eigenvalue to be computed. 1 <= I <= N. */ /* D (input) REAL array, dimension (N) */ /* The original eigenvalues. It is assumed that they are in */ /* order, D(I) < D(J) for I < J. */ /* Z (input) REAL array, dimension (N) */ /* The components of the updating vector. */ /* DELTA (output) REAL array, dimension (N) */ /* If N .GT. 2, DELTA contains (D(j) - lambda_I) in its j-th */ /* component. If N = 1, then DELTA(1) = 1. If N = 2, see SLAED5 */ /* for detail. The vector DELTA contains the information necessary */ /* to construct the eigenvectors by SLAED3 and SLAED9. */ /* RHO (input) REAL */ /* The scalar in the symmetric updating formula. */ /* DLAM (output) REAL */ /* The computed lambda_I, the I-th updated eigenvalue. */ /* INFO (output) INTEGER */ /* = 0: successful exit */ /* > 0: if INFO = 1, the updating process failed. */ /* Internal Parameters */ /* =================== */ /* Logical variable ORGATI (origin-at-i?) is used for distinguishing */ /* whether D(i) or D(i+1) is treated as the origin. */ /* ORGATI = .true. origin at i */ /* ORGATI = .false. origin at i+1 */ /* Logical variable SWTCH3 (switch-for-3-poles?) is for noting */ /* if we are working with THREE poles! */ /* MAXIT is the maximum number of iterations allowed for each */ /* eigenvalue. */ /* Further Details */ /* =============== */ /* Based on contributions by */ /* Ren-Cang Li, Computer Science Division, University of California */ /* at Berkeley, USA */ /* ===================================================================== */ /* .. Parameters .. */ /* .. */ /* .. Local Scalars .. */ /* .. */ /* .. Local Arrays .. */ /* .. */ /* .. External Functions .. */ /* .. */ /* .. External Subroutines .. */ /* .. */ /* .. Intrinsic Functions .. */ /* .. */ /* .. Executable Statements .. */ /* Since this routine is called in an inner loop, we do no argument */ /* checking. */ /* Quick return for N=1 and 2. */ /* Parameter adjustments */ --delta; --z__; --d__; /* Function Body */ *info = 0; if (*n == 1) { /* Presumably, I=1 upon entry */ *dlam = d__[1] + *rho * z__[1] * z__[1]; delta[1] = 1.f; return 0; } if (*n == 2) { slaed5_(i__, &d__[1], &z__[1], &delta[1], rho, dlam); return 0; } /* Compute machine epsilon */ eps = slamch_("Epsilon"); rhoinv = 1.f / *rho; /* The case I = N */ if (*i__ == *n) { /* Initialize some basic variables */ ii = *n - 1; niter = 1; /* Calculate initial guess */ midpt = *rho / 2.f; /* If ||Z||_2 is not one, then TEMP should be set to */ /* RHO * ||Z||_2^2 / TWO */ i__1 = *n; for (j = 1; j <= i__1; ++j) { delta[j] = d__[j] - d__[*i__] - midpt; /* L10: */ } psi = 0.f; i__1 = *n - 2; for (j = 1; j <= i__1; ++j) { psi += z__[j] * z__[j] / delta[j]; /* L20: */ } c__ = rhoinv + psi; w = c__ + z__[ii] * z__[ii] / delta[ii] + z__[*n] * z__[*n] / delta[* n]; if (w <= 0.f) { temp = z__[*n - 1] * z__[*n - 1] / (d__[*n] - d__[*n - 1] + *rho) + z__[*n] * z__[*n] / *rho; if (c__ <= temp) { tau = *rho; } else { del = d__[*n] - d__[*n - 1]; a = -c__ * del + z__[*n - 1] * z__[*n - 1] + z__[*n] * z__[*n] ; b = z__[*n] * z__[*n] * del; if (a < 0.f) { tau = b * 2.f / (sqrt(a * a + b * 4.f * c__) - a); } else { tau = (a + sqrt(a * a + b * 4.f * c__)) / (c__ * 2.f); } } /* It can be proved that */ /* D(N)+RHO/2 <= LAMBDA(N) < D(N)+TAU <= D(N)+RHO */ dltlb = midpt; dltub = *rho; } else { del = d__[*n] - d__[*n - 1]; a = -c__ * del + z__[*n - 1] * z__[*n - 1] + z__[*n] * z__[*n]; b = z__[*n] * z__[*n] * del; if (a < 0.f) { tau = b * 2.f / (sqrt(a * a + b * 4.f * c__) - a); } else { tau = (a + sqrt(a * a + b * 4.f * c__)) / (c__ * 2.f); } /* It can be proved that */ /* D(N) < D(N)+TAU < LAMBDA(N) < D(N)+RHO/2 */ dltlb = 0.f; dltub = midpt; } i__1 = *n; for (j = 1; j <= i__1; ++j) { delta[j] = d__[j] - d__[*i__] - tau; /* L30: */ } /* Evaluate PSI and the derivative DPSI */ dpsi = 0.f; psi = 0.f; erretm = 0.f; i__1 = ii; for (j = 1; j <= i__1; ++j) { temp = z__[j] / delta[j]; psi += z__[j] * temp; dpsi += temp * temp; erretm += psi; /* L40: */ } erretm = dabs(erretm); /* Evaluate PHI and the derivative DPHI */ temp = z__[*n] / delta[*n]; phi = z__[*n] * temp; dphi = temp * temp; erretm = (-phi - psi) * 8.f + erretm - phi + rhoinv + dabs(tau) * ( dpsi + dphi); w = rhoinv + phi + psi; /* Test for convergence */ if (dabs(w) <= eps * erretm) { *dlam = d__[*i__] + tau; goto L250; } if (w <= 0.f) { dltlb = dmax(dltlb,tau); } else { dltub = dmin(dltub,tau); } /* Calculate the new step */ ++niter; c__ = w - delta[*n - 1] * dpsi - delta[*n] * dphi; a = (delta[*n - 1] + delta[*n]) * w - delta[*n - 1] * delta[*n] * ( dpsi + dphi); b = delta[*n - 1] * delta[*n] * w; if (c__ < 0.f) { c__ = dabs(c__); } if (c__ == 0.f) { /* ETA = B/A */ /* ETA = RHO - TAU */ eta = dltub - tau; } else if (a >= 0.f) { eta = (a + sqrt((r__1 = a * a - b * 4.f * c__, dabs(r__1)))) / ( c__ * 2.f); } else { eta = b * 2.f / (a - sqrt((r__1 = a * a - b * 4.f * c__, dabs( r__1)))); } /* Note, eta should be positive if w is negative, and */ /* eta should be negative otherwise. However, */ /* if for some reason caused by roundoff, eta*w > 0, */ /* we simply use one Newton step instead. This way */ /* will guarantee eta*w < 0. */ if (w * eta > 0.f) { eta = -w / (dpsi + dphi); } temp = tau + eta; if (temp > dltub || temp < dltlb) { if (w < 0.f) { eta = (dltub - tau) / 2.f; } else { eta = (dltlb - tau) / 2.f; } } i__1 = *n; for (j = 1; j <= i__1; ++j) { delta[j] -= eta; /* L50: */ } tau += eta; /* Evaluate PSI and the derivative DPSI */ dpsi = 0.f; psi = 0.f; erretm = 0.f; i__1 = ii; for (j = 1; j <= i__1; ++j) { temp = z__[j] / delta[j]; psi += z__[j] * temp; dpsi += temp * temp; erretm += psi; /* L60: */ } erretm = dabs(erretm); /* Evaluate PHI and the derivative DPHI */ temp = z__[*n] / delta[*n]; phi = z__[*n] * temp; dphi = temp * temp; erretm = (-phi - psi) * 8.f + erretm - phi + rhoinv + dabs(tau) * ( dpsi + dphi); w = rhoinv + phi + psi; /* Main loop to update the values of the array DELTA */ iter = niter + 1; for (niter = iter; niter <= 30; ++niter) { /* Test for convergence */ if (dabs(w) <= eps * erretm) { *dlam = d__[*i__] + tau; goto L250; } if (w <= 0.f) { dltlb = dmax(dltlb,tau); } else { dltub = dmin(dltub,tau); } /* Calculate the new step */ c__ = w - delta[*n - 1] * dpsi - delta[*n] * dphi; a = (delta[*n - 1] + delta[*n]) * w - delta[*n - 1] * delta[*n] * (dpsi + dphi); b = delta[*n - 1] * delta[*n] * w; if (a >= 0.f) { eta = (a + sqrt((r__1 = a * a - b * 4.f * c__, dabs(r__1)))) / (c__ * 2.f); } else { eta = b * 2.f / (a - sqrt((r__1 = a * a - b * 4.f * c__, dabs( r__1)))); } /* Note, eta should be positive if w is negative, and */ /* eta should be negative otherwise. However, */ /* if for some reason caused by roundoff, eta*w > 0, */ /* we simply use one Newton step instead. This way */ /* will guarantee eta*w < 0. */ if (w * eta > 0.f) { eta = -w / (dpsi + dphi); } temp = tau + eta; if (temp > dltub || temp < dltlb) { if (w < 0.f) { eta = (dltub - tau) / 2.f; } else { eta = (dltlb - tau) / 2.f; } } i__1 = *n; for (j = 1; j <= i__1; ++j) { delta[j] -= eta; /* L70: */ } tau += eta; /* Evaluate PSI and the derivative DPSI */ dpsi = 0.f; psi = 0.f; erretm = 0.f; i__1 = ii; for (j = 1; j <= i__1; ++j) { temp = z__[j] / delta[j]; psi += z__[j] * temp; dpsi += temp * temp; erretm += psi; /* L80: */ } erretm = dabs(erretm); /* Evaluate PHI and the derivative DPHI */ temp = z__[*n] / delta[*n]; phi = z__[*n] * temp; dphi = temp * temp; erretm = (-phi - psi) * 8.f + erretm - phi + rhoinv + dabs(tau) * (dpsi + dphi); w = rhoinv + phi + psi; /* L90: */ } /* Return with INFO = 1, NITER = MAXIT and not converged */ *info = 1; *dlam = d__[*i__] + tau; goto L250; /* End for the case I = N */ } else { /* The case for I < N */ niter = 1; ip1 = *i__ + 1; /* Calculate initial guess */ del = d__[ip1] - d__[*i__]; midpt = del / 2.f; i__1 = *n; for (j = 1; j <= i__1; ++j) { delta[j] = d__[j] - d__[*i__] - midpt; /* L100: */ } psi = 0.f; i__1 = *i__ - 1; for (j = 1; j <= i__1; ++j) { psi += z__[j] * z__[j] / delta[j]; /* L110: */ } phi = 0.f; i__1 = *i__ + 2; for (j = *n; j >= i__1; --j) { phi += z__[j] * z__[j] / delta[j]; /* L120: */ } c__ = rhoinv + psi + phi; w = c__ + z__[*i__] * z__[*i__] / delta[*i__] + z__[ip1] * z__[ip1] / delta[ip1]; if (w > 0.f) { /* d(i)< the ith eigenvalue < (d(i)+d(i+1))/2 */ /* We choose d(i) as origin. */ orgati = TRUE_; a = c__ * del + z__[*i__] * z__[*i__] + z__[ip1] * z__[ip1]; b = z__[*i__] * z__[*i__] * del; if (a > 0.f) { tau = b * 2.f / (a + sqrt((r__1 = a * a - b * 4.f * c__, dabs( r__1)))); } else { tau = (a - sqrt((r__1 = a * a - b * 4.f * c__, dabs(r__1)))) / (c__ * 2.f); } dltlb = 0.f; dltub = midpt; } else { /* (d(i)+d(i+1))/2 <= the ith eigenvalue < d(i+1) */ /* We choose d(i+1) as origin. */ orgati = FALSE_; a = c__ * del - z__[*i__] * z__[*i__] - z__[ip1] * z__[ip1]; b = z__[ip1] * z__[ip1] * del; if (a < 0.f) { tau = b * 2.f / (a - sqrt((r__1 = a * a + b * 4.f * c__, dabs( r__1)))); } else { tau = -(a + sqrt((r__1 = a * a + b * 4.f * c__, dabs(r__1)))) / (c__ * 2.f); } dltlb = -midpt; dltub = 0.f; } if (orgati) { i__1 = *n; for (j = 1; j <= i__1; ++j) { delta[j] = d__[j] - d__[*i__] - tau; /* L130: */ } } else { i__1 = *n; for (j = 1; j <= i__1; ++j) { delta[j] = d__[j] - d__[ip1] - tau; /* L140: */ } } if (orgati) { ii = *i__; } else { ii = *i__ + 1; } iim1 = ii - 1; iip1 = ii + 1; /* Evaluate PSI and the derivative DPSI */ dpsi = 0.f; psi = 0.f; erretm = 0.f; i__1 = iim1; for (j = 1; j <= i__1; ++j) { temp = z__[j] / delta[j]; psi += z__[j] * temp; dpsi += temp * temp; erretm += psi; /* L150: */ } erretm = dabs(erretm); /* Evaluate PHI and the derivative DPHI */ dphi = 0.f; phi = 0.f; i__1 = iip1; for (j = *n; j >= i__1; --j) { temp = z__[j] / delta[j]; phi += z__[j] * temp; dphi += temp * temp; erretm += phi; /* L160: */ } w = rhoinv + phi + psi; /* W is the value of the secular function with */ /* its ii-th element removed. */ swtch3 = FALSE_; if (orgati) { if (w < 0.f) { swtch3 = TRUE_; } } else { if (w > 0.f) { swtch3 = TRUE_; } } if (ii == 1 || ii == *n) { swtch3 = FALSE_; } temp = z__[ii] / delta[ii]; dw = dpsi + dphi + temp * temp; temp = z__[ii] * temp; w += temp; erretm = (phi - psi) * 8.f + erretm + rhoinv * 2.f + dabs(temp) * 3.f + dabs(tau) * dw; /* Test for convergence */ if (dabs(w) <= eps * erretm) { if (orgati) { *dlam = d__[*i__] + tau; } else { *dlam = d__[ip1] + tau; } goto L250; } if (w <= 0.f) { dltlb = dmax(dltlb,tau); } else { dltub = dmin(dltub,tau); } /* Calculate the new step */ ++niter; if (! swtch3) { if (orgati) { /* Computing 2nd power */ r__1 = z__[*i__] / delta[*i__]; c__ = w - delta[ip1] * dw - (d__[*i__] - d__[ip1]) * (r__1 * r__1); } else { /* Computing 2nd power */ r__1 = z__[ip1] / delta[ip1]; c__ = w - delta[*i__] * dw - (d__[ip1] - d__[*i__]) * (r__1 * r__1); } a = (delta[*i__] + delta[ip1]) * w - delta[*i__] * delta[ip1] * dw; b = delta[*i__] * delta[ip1] * w; if (c__ == 0.f) { if (a == 0.f) { if (orgati) { a = z__[*i__] * z__[*i__] + delta[ip1] * delta[ip1] * (dpsi + dphi); } else { a = z__[ip1] * z__[ip1] + delta[*i__] * delta[*i__] * (dpsi + dphi); } } eta = b / a; } else if (a <= 0.f) { eta = (a - sqrt((r__1 = a * a - b * 4.f * c__, dabs(r__1)))) / (c__ * 2.f); } else { eta = b * 2.f / (a + sqrt((r__1 = a * a - b * 4.f * c__, dabs( r__1)))); } } else { /* Interpolation using THREE most relevant poles */ temp = rhoinv + psi + phi; if (orgati) { temp1 = z__[iim1] / delta[iim1]; temp1 *= temp1; c__ = temp - delta[iip1] * (dpsi + dphi) - (d__[iim1] - d__[ iip1]) * temp1; zz[0] = z__[iim1] * z__[iim1]; zz[2] = delta[iip1] * delta[iip1] * (dpsi - temp1 + dphi); } else { temp1 = z__[iip1] / delta[iip1]; temp1 *= temp1; c__ = temp - delta[iim1] * (dpsi + dphi) - (d__[iip1] - d__[ iim1]) * temp1; zz[0] = delta[iim1] * delta[iim1] * (dpsi + (dphi - temp1)); zz[2] = z__[iip1] * z__[iip1]; } zz[1] = z__[ii] * z__[ii]; slaed6_(&niter, &orgati, &c__, &delta[iim1], zz, &w, &eta, info); if (*info != 0) { goto L250; } } /* Note, eta should be positive if w is negative, and */ /* eta should be negative otherwise. However, */ /* if for some reason caused by roundoff, eta*w > 0, */ /* we simply use one Newton step instead. This way */ /* will guarantee eta*w < 0. */ if (w * eta >= 0.f) { eta = -w / dw; } temp = tau + eta; if (temp > dltub || temp < dltlb) { if (w < 0.f) { eta = (dltub - tau) / 2.f; } else { eta = (dltlb - tau) / 2.f; } } prew = w; i__1 = *n; for (j = 1; j <= i__1; ++j) { delta[j] -= eta; /* L180: */ } /* Evaluate PSI and the derivative DPSI */ dpsi = 0.f; psi = 0.f; erretm = 0.f; i__1 = iim1; for (j = 1; j <= i__1; ++j) { temp = z__[j] / delta[j]; psi += z__[j] * temp; dpsi += temp * temp; erretm += psi; /* L190: */ } erretm = dabs(erretm); /* Evaluate PHI and the derivative DPHI */ dphi = 0.f; phi = 0.f; i__1 = iip1; for (j = *n; j >= i__1; --j) { temp = z__[j] / delta[j]; phi += z__[j] * temp; dphi += temp * temp; erretm += phi; /* L200: */ } temp = z__[ii] / delta[ii]; dw = dpsi + dphi + temp * temp; temp = z__[ii] * temp; w = rhoinv + phi + psi + temp; erretm = (phi - psi) * 8.f + erretm + rhoinv * 2.f + dabs(temp) * 3.f + (r__1 = tau + eta, dabs(r__1)) * dw; swtch = FALSE_; if (orgati) { if (-w > dabs(prew) / 10.f) { swtch = TRUE_; } } else { if (w > dabs(prew) / 10.f) { swtch = TRUE_; } } tau += eta; /* Main loop to update the values of the array DELTA */ iter = niter + 1; for (niter = iter; niter <= 30; ++niter) { /* Test for convergence */ if (dabs(w) <= eps * erretm) { if (orgati) { *dlam = d__[*i__] + tau; } else { *dlam = d__[ip1] + tau; } goto L250; } if (w <= 0.f) { dltlb = dmax(dltlb,tau); } else { dltub = dmin(dltub,tau); } /* Calculate the new step */ if (! swtch3) { if (! swtch) { if (orgati) { /* Computing 2nd power */ r__1 = z__[*i__] / delta[*i__]; c__ = w - delta[ip1] * dw - (d__[*i__] - d__[ip1]) * ( r__1 * r__1); } else { /* Computing 2nd power */ r__1 = z__[ip1] / delta[ip1]; c__ = w - delta[*i__] * dw - (d__[ip1] - d__[*i__]) * (r__1 * r__1); } } else { temp = z__[ii] / delta[ii]; if (orgati) { dpsi += temp * temp; } else { dphi += temp * temp; } c__ = w - delta[*i__] * dpsi - delta[ip1] * dphi; } a = (delta[*i__] + delta[ip1]) * w - delta[*i__] * delta[ip1] * dw; b = delta[*i__] * delta[ip1] * w; if (c__ == 0.f) { if (a == 0.f) { if (! swtch) { if (orgati) { a = z__[*i__] * z__[*i__] + delta[ip1] * delta[ip1] * (dpsi + dphi); } else { a = z__[ip1] * z__[ip1] + delta[*i__] * delta[ *i__] * (dpsi + dphi); } } else { a = delta[*i__] * delta[*i__] * dpsi + delta[ip1] * delta[ip1] * dphi; } } eta = b / a; } else if (a <= 0.f) { eta = (a - sqrt((r__1 = a * a - b * 4.f * c__, dabs(r__1)) )) / (c__ * 2.f); } else { eta = b * 2.f / (a + sqrt((r__1 = a * a - b * 4.f * c__, dabs(r__1)))); } } else { /* Interpolation using THREE most relevant poles */ temp = rhoinv + psi + phi; if (swtch) { c__ = temp - delta[iim1] * dpsi - delta[iip1] * dphi; zz[0] = delta[iim1] * delta[iim1] * dpsi; zz[2] = delta[iip1] * delta[iip1] * dphi; } else { if (orgati) { temp1 = z__[iim1] / delta[iim1]; temp1 *= temp1; c__ = temp - delta[iip1] * (dpsi + dphi) - (d__[iim1] - d__[iip1]) * temp1; zz[0] = z__[iim1] * z__[iim1]; zz[2] = delta[iip1] * delta[iip1] * (dpsi - temp1 + dphi); } else { temp1 = z__[iip1] / delta[iip1]; temp1 *= temp1; c__ = temp - delta[iim1] * (dpsi + dphi) - (d__[iip1] - d__[iim1]) * temp1; zz[0] = delta[iim1] * delta[iim1] * (dpsi + (dphi - temp1)); zz[2] = z__[iip1] * z__[iip1]; } } slaed6_(&niter, &orgati, &c__, &delta[iim1], zz, &w, &eta, info); if (*info != 0) { goto L250; } } /* Note, eta should be positive if w is negative, and */ /* eta should be negative otherwise. However, */ /* if for some reason caused by roundoff, eta*w > 0, */ /* we simply use one Newton step instead. This way */ /* will guarantee eta*w < 0. */ if (w * eta >= 0.f) { eta = -w / dw; } temp = tau + eta; if (temp > dltub || temp < dltlb) { if (w < 0.f) { eta = (dltub - tau) / 2.f; } else { eta = (dltlb - tau) / 2.f; } } i__1 = *n; for (j = 1; j <= i__1; ++j) { delta[j] -= eta; /* L210: */ } tau += eta; prew = w; /* Evaluate PSI and the derivative DPSI */ dpsi = 0.f; psi = 0.f; erretm = 0.f; i__1 = iim1; for (j = 1; j <= i__1; ++j) { temp = z__[j] / delta[j]; psi += z__[j] * temp; dpsi += temp * temp; erretm += psi; /* L220: */ } erretm = dabs(erretm); /* Evaluate PHI and the derivative DPHI */ dphi = 0.f; phi = 0.f; i__1 = iip1; for (j = *n; j >= i__1; --j) { temp = z__[j] / delta[j]; phi += z__[j] * temp; dphi += temp * temp; erretm += phi; /* L230: */ } temp = z__[ii] / delta[ii]; dw = dpsi + dphi + temp * temp; temp = z__[ii] * temp; w = rhoinv + phi + psi + temp; erretm = (phi - psi) * 8.f + erretm + rhoinv * 2.f + dabs(temp) * 3.f + dabs(tau) * dw; if (w * prew > 0.f && dabs(w) > dabs(prew) / 10.f) { swtch = ! swtch; } /* L240: */ } /* Return with INFO = 1, NITER = MAXIT and not converged */ *info = 1; if (orgati) { *dlam = d__[*i__] + tau; } else { *dlam = d__[ip1] + tau; } } L250: return 0; /* End of SLAED4 */ } /* slaed4_ */