11.6 The QR Algorithm For Real Hessenberg Matrices: Eigensystems
11.6 The QR Algorithm For Real Hessenberg Matrices: Eigensystems
Chapter 11.
Eigensystems
} } if (i != m) { Interchange rows and columns. for (j=m-1;j<=n;j++) SWAP(a[i][j],a[m][j]) for (j=1;j<=n;j++) SWAP(a[j][i],a[j][m]) } if (x) { Carry out the elimination. for (i=m+1;i<=n;i++) { if ((y=a[i][m-1]) != 0.0) { y /= x; a[i][m-1]=y; for (j=m;j<=n;j++) a[i][j] -= y*a[m][j]; for (j=1;j<=n;j++) a[j][m] += y*a[j][i]; } } } } }
Sample page from NUMERICAL RECIPES IN C: THE ART OF SCIENTIFIC COMPUTING (ISBN 0-521-43108-5) Copyright (C) 1988-1992 by Cambridge University Press. Programs Copyright (C) 1988-1992 by Numerical Recipes Software. Permission is granted for internet users to make one paper copy for their own personal use. Further reproduction, or any copying of machinereadable files (including this one) to any server computer, is strictly prohibited. To order Numerical Recipes books or CDROMs, visit website http://www.nr.com or call 1-800-872-7423 (North America only), or send email to [email protected] (outside North America).
CITED REFERENCES AND FURTHER READING: Wilkinson, J.H., and Reinsch, C. 1971, Linear Algebra, vol. II of Handbook for Automatic Computation (New York: Springer-Verlag). [1] Smith, B.T., et al. 1976, Matrix Eigensystem Routines EISPACK Guide, 2nd ed., vol. 6 of Lecture Notes in Computer Science (New York: Springer-Verlag). [2] Stoer, J., and Bulirsch, R. 1980, Introduction to Numerical Analysis (New York: Springer-Verlag), 6.5.4. [3]
The QR transformation preserves the upper Hessenberg form of the original matrix A A1 , and the workload on such a matrix is O(n 2 ) per iteration as opposed to O(n3 ) on a general matrix. As s , A s converges to a form where the eigenvalues are either isolated on the diagonal or are eigenvalues of a 2 2 submatrix on the diagonal. As we pointed out in 11.3, shifting is essential for rapid convergence. A key difference here is that a nonsymmetric real matrix can have complex eigenvalues. This
487
means that good choices for the shifts k s may be complex, apparently necessitating complex arithmetic. Complex arithmetic can be avoided, however, by a clever trick. The trick depends on a result analogous to the lemma we used for implicit shifts in 11.3. The lemma we need here states that if B is a nonsingular matrix such that BQ= QH (11.6.3)
Sample page from NUMERICAL RECIPES IN C: THE ART OF SCIENTIFIC COMPUTING (ISBN 0-521-43108-5) Copyright (C) 1988-1992 by Cambridge University Press. Programs Copyright (C) 1988-1992 by Numerical Recipes Software. Permission is granted for internet users to make one paper copy for their own personal use. Further reproduction, or any copying of machinereadable files (including this one) to any server computer, is strictly prohibited. To order Numerical Recipes books or CDROMs, visit website http://www.nr.com or call 1-800-872-7423 (North America only), or send email to [email protected] (outside North America).
where Q is orthogonal and H is upper Hessenberg, then Q and H are fully determined by the rst column of Q. (The determination is unique if H has positive subdiagonal elements.) The lemma can be proved by induction analogously to the proof given for tridiagonal matrices in 11.3. The lemma is used in practice by taking two steps of the QR algorithm, either with two real shifts ks and ks+1 , or with complex conjugate values k s and ks+1 = ks *. This gives a real matrix A s+2 , where
T As+2 = Qs+1 Qs As QT s Qs+1
(11.6.4)
The Qs are determined by As+1 = Qs As QT s R As+1 ks+1 1 = QT s +1 s+1 Using (11.6.6), equation (11.6.7) can be rewritten
T As ks+1 1 = QT s Qs+1 Rs+1 Qs
As ks 1 = QT s Rs
(11.6.8)
Hence, if we dene M = (As ks+1 1) (As ks 1) equations (11.6.5) and (11.6.8) give R=QM where Q = Qs+1 Qs R = Rs+1 Rs Equation (11.6.4) can be rewritten As QT = QT As+2 (11.6.13) (11.6.11) (11.6.12) (11.6.10) (11.6.9)
(11.6.14)
where Q is orthogonal. If Q has the same rst column as QT (i.e., Q has the same rst row as Q), then Q = Q and A s+2 = H.
488
Chapter 11.
Eigensystems
The rst row of Q is found as follows. Equation (11.6.10) shows that Q is the orthogonal matrix that triangularizes the real matrix M. Any real matrix can be triangularized by premultiplying it by a sequence of Householder matrices P 1 (acting on the rst column), P 2 (acting on the second column), . . . , P n1 . Thus Q = Pn1 P2 P1 , and the rst row of Q is the rst row of P 1 since Pi is an (i 1) (i 1) identity matrix in the top left-hand corner. We now must nd Q satisfying (11.6.14) whose rst row is that of P 1 . The Householder matrix P 1 is determined by the rst column of M. Since A s is upper Hessenberg, equation (11.6.9) shows that the rst column of M has the form [p1 , q1 , r1 , 0, ..., 0]T , where p 1 = a2 11 a11 (ks + ks+1 ) + ks ks+1 + a12 a21 q1 = a21 (a11 + a22 ks ks+1 ) r1 = a21 a32 Hence P 1 = 1 2 w 1 wT 1 (11.6.16) (11.6.15)
Sample page from NUMERICAL RECIPES IN C: THE ART OF SCIENTIFIC COMPUTING (ISBN 0-521-43108-5) Copyright (C) 1988-1992 by Cambridge University Press. Programs Copyright (C) 1988-1992 by Numerical Recipes Software. Permission is granted for internet users to make one paper copy for their own personal use. Further reproduction, or any copying of machinereadable files (including this one) to any server computer, is strictly prohibited. To order Numerical Recipes books or CDROMs, visit website http://www.nr.com or call 1-800-872-7423 (North America only), or send email to [email protected] (outside North America).
where w1 has only its rst 3 elements nonzero (cf. equation 11.2.5). The matrix P 1 As P T 1 is therefore upper Hessenberg with 3 extra elements: x T P 1 A1 P 1 = x x
(11.6.17)
This matrix can be restored to upper Hessenberg form without affecting the rst row by a sequence of Householder similarity transformations. The rst such Householder matrix, P2 , acts on elements 2, 3, and 4 in the rst column, annihilating elements 3 and 4. This produces a matrix of the same form as (11.6.17), with the 3 extra elements appearing one column over: x x x
(11.6.18)
Proceeding in this way up to P n1 , we see that at each stage the Householder matrix Pr has a vector wr that is nonzero only in elements r, r + 1, and r + 2. These elements are determined by the elements r, r + 1, and r + 2 in the (r 1)st
489
column of the current matrix. Note that the preliminary matrix P 1 has the same structure as P2 , . . . , Pn1 . The result is that
T T Pn1 P2 P1 As PT 1 P2 Pn1 = H
(11.6.19)
Sample page from NUMERICAL RECIPES IN C: THE ART OF SCIENTIFIC COMPUTING (ISBN 0-521-43108-5) Copyright (C) 1988-1992 by Cambridge University Press. Programs Copyright (C) 1988-1992 by Numerical Recipes Software. Permission is granted for internet users to make one paper copy for their own personal use. Further reproduction, or any copying of machinereadable files (including this one) to any server computer, is strictly prohibited. To order Numerical Recipes books or CDROMs, visit website http://www.nr.com or call 1-800-872-7423 (North America only), or send email to [email protected] (outside North America).
where H is upper Hessenberg. Thus Q = Q = Pn1 P2 P1 and As+2 = H (11.6.21) The shifts of origin at each stage are taken to be the eigenvalues of the 2 2 matrix in the bottom right-hand corner of the current A s . This gives ks + ks+2 = an1,n1 + ann ks ks+1 = an1,n1 ann an1,n an,n1 Substituting (11.6.22) in (11.6.15), we get p1 = a21 {[(ann a11 )(an1,n1 a11 ) an1,n an,n1 ]/a21 + a12 } q1 = a21 [a22 a11 (ann a11 ) (an1,n1 a11 )] r1 = a21 a32 (11.6.23) (11.6.22) (11.6.20)
We have judiciously grouped terms to reduce possible roundoff when there are small off-diagonal elements. Since only the ratios of elements are relevant for a Householder transformation, we can omit the factor a 21 from (11.6.23). In summary, to carry out a double QR step we construct the Householder matrices Pr , r = 1, . . . , n 1. For P1 we use p1 , q1 , and r1 given by (11.6.23). For the remaining matrices, p r , qr , and rr are determined by the (r, r 1), (r + 1, r 1), and (r + 2, r 1) elements of the current matrix. The number of arithmetic operations can be reduced by writing the nonzero elements of the 2w w T part of the Householder matrix in the form (p s)/(s) 2w wT = q/(s) [ 1 q/(p s) r/(p s) ] (11.6.24) r/(s) where s2 = p 2 + q 2 + r 2 (11.6.25)
(We have simply divided each element by a piece of the normalizing factor; cf. the equations in 11.2.) If we proceed in this way, convergence is usually very fast. There are two possible ways of terminating the iteration for an eigenvalue. First, if a n,n1 becomes negligible, then a nn is an eigenvalue. We can then delete the nth row and column
490
Chapter 11.
Eigensystems
of the matrix and look for the next eigenvalue. Alternatively, a n1,n2 may become negligible. In this case the eigenvalues of the 2 2 matrix in the lower right-hand corner may be taken to be eigenvalues. We delete the nth and (n 1)st rows and columns of the matrix and continue. The test for convergence to an eigenvalue is combined with a test for negligible subdiagonal elements that allows splitting of the matrix into submatrices. We nd the largest i such that ai,i1 is negligible. If i = n, we have found a single eigenvalue. If i = n 1, we have found two eigenvalues. Otherwise we continue the iteration on the submatrix in rows i to n (i being set to unity if there is no small subdiagonal element). After determining i, the submatrix in rows i to n is examined to see if the product of any two consecutive subdiagonal elements is small enough that we can work with an even smaller submatrix, starting say in row m. We start with m = n 2 and decrement it down to i + 1, computing p, q , and r according to equations (11.6.23) with 1 replaced by m and 2 by m + 1. If these were indeed the elements of the special rst Householder matrix in a double QR step, then applying the Householder matrix would lead to nonzero elements in positions (m + 1, m 1), (m + 2, m 1), and (m + 2, m). We require that the rst two of these elements be small compared with the local diagonal elements a m1,m1 , amm and am+1,m+1 . A satisfactory approximate criterion is |am,m1 |(|q | + |r|) |p|(|am+1,m+1 | + |amm | + |am1,m1 |) (11.6.26)
Sample page from NUMERICAL RECIPES IN C: THE ART OF SCIENTIFIC COMPUTING (ISBN 0-521-43108-5) Copyright (C) 1988-1992 by Cambridge University Press. Programs Copyright (C) 1988-1992 by Numerical Recipes Software. Permission is granted for internet users to make one paper copy for their own personal use. Further reproduction, or any copying of machinereadable files (including this one) to any server computer, is strictly prohibited. To order Numerical Recipes books or CDROMs, visit website http://www.nr.com or call 1-800-872-7423 (North America only), or send email to [email protected] (outside North America).
Very rarely, the procedure described so far will fail to converge. On such matrices, experience shows that if one double step is performed with any shifts that are of order the norm of the matrix, convergence is subsequently very rapid. Accordingly, if ten iterations occur without determining an eigenvalue, the usual shifts are replaced for the next iteration by shifts dened by ks + ks+1 = 1.5 (|an,n1 | + |an1,n2 |) ks ks+1 = (|an,n1 | + |an1,n2 |)2 (11.6.27)
The factor 1.5 was arbitrarily chosen to lessen the likelihood of an unfortunate choice of shifts. This strategy is repeated after 20 unsuccessful iterations. After 30 unsuccessful iterations, the routine reports failure. The operation count for the QR algorithm described here is 5k 2 per iteration, where k is the current size of the matrix. The typical average number of iterations per eigenvalue is 1.8, so the total operation count for all the eigenvalues is 3n 3 . This estimate neglects any possible efciency due to splitting or sparseness of the matrix. The following routine hqr is based algorithmically on the above description, in turn following the implementations in [1,2].
#include <math.h> #include "nrutil.h" void hqr(float **a, int n, float wr[], float wi[]) Finds all eigenvalues of an upper Hessenberg matrix a[1..n][1..n] . On input a can be exactly as output from elmhes 11.5; on output it is destroyed. The real and imaginary parts of the eigenvalues are returned in wr[1..n] and wi[1..n], respectively. {
491
int nn,m,l,k,j,its,i,mmin; float z,y,x,w,v,u,t,s,r,q,p,anorm; anorm=0.0; Compute matrix norm for possible use in lofor (i=1;i<=n;i++) cating single small subdiagonal element. for (j=IMAX(i-1,1);j<=n;j++) anorm += fabs(a[i][j]); nn=n; t=0.0; Gets changed only by an exceptional shift. while (nn >= 1) { Begin search for next eigenvalue. its=0; do { for (l=nn;l>=2;l--) { Begin iteration: look for single small subdis=fabs(a[l-1][l-1])+fabs(a[l][l]); agonal element. if (s == 0.0) s=anorm; if ((float)(fabs(a[l][l-1]) + s) == s) { a[l][l-1]=0.0; break; } } x=a[nn][nn]; if (l == nn) { One root found. wr[nn]=x+t; wi[nn--]=0.0; } else { y=a[nn-1][nn-1]; w=a[nn][nn-1]*a[nn-1][nn]; if (l == (nn-1)) { Two roots found... p=0.5*(y-x); q=p*p+w; z=sqrt(fabs(q)); x += t; if (q >= 0.0) { ...a real pair. z=p+SIGN(z,p); wr[nn-1]=wr[nn]=x+z; if (z) wr[nn]=x-w/z; wi[nn-1]=wi[nn]=0.0; } else { ...a complex pair. wr[nn-1]=wr[nn]=x+p; wi[nn-1]= -(wi[nn]=z); } nn -= 2; } else { No roots found. Continue iteration. if (its == 30) nrerror("Too many iterations in hqr"); if (its == 10 || its == 20) { Form exceptional shift. t += x; for (i=1;i<=nn;i++) a[i][i] -= x; s=fabs(a[nn][nn-1])+fabs(a[nn-1][nn-2]); y=x=0.75*s; w = -0.4375*s*s; } ++its; for (m=(nn-2);m>=l;m--) { Form shift and then look for z=a[m][m]; 2 consecutive small subr=x-z; diagonal elements. s=y-z; p=(r*s-w)/a[m+1][m]+a[m][m+1]; Equation (11.6.23). q=a[m+1][m+1]-z-r-s; r=a[m+2][m+1]; s=fabs(p)+fabs(q)+fabs(r); Scale to prevent overow or p /= s; underow. q /= s; r /= s; if (m == l) break;
Sample page from NUMERICAL RECIPES IN C: THE ART OF SCIENTIFIC COMPUTING (ISBN 0-521-43108-5) Copyright (C) 1988-1992 by Cambridge University Press. Programs Copyright (C) 1988-1992 by Numerical Recipes Software. Permission is granted for internet users to make one paper copy for their own personal use. Further reproduction, or any copying of machinereadable files (including this one) to any server computer, is strictly prohibited. To order Numerical Recipes books or CDROMs, visit website http://www.nr.com or call 1-800-872-7423 (North America only), or send email to [email protected] (outside North America).
492
Chapter 11.
Eigensystems
u=fabs(a[m][m-1])*(fabs(q)+fabs(r)); v=fabs(p)*(fabs(a[m-1][m-1])+fabs(z)+fabs(a[m+1][m+1])); if ((float)(u+v) == v) break; Equation (11.6.26). } for (i=m+2;i<=nn;i++) { a[i][i-2]=0.0; if (i != (m+2)) a[i][i-3]=0.0; } for (k=m;k<=nn-1;k++) { Double QR step on rows l to nn and columns m to nn. if (k != m) { p=a[k][k-1]; Begin setup of Householder q=a[k+1][k-1]; vector. r=0.0; if (k != (nn-1)) r=a[k+2][k-1]; if ((x=fabs(p)+fabs(q)+fabs(r)) != 0.0) { p /= x; Scale to prevent overow or q /= x; underow. r /= x; } } if ((s=SIGN(sqrt(p*p+q*q+r*r),p)) != 0.0) { if (k == m) { if (l != m) a[k][k-1] = -a[k][k-1]; } else a[k][k-1] = -s*x; p += s; Equations (11.6.24). x=p/s; y=q/s; z=r/s; q /= p; r /= p; for (j=k;j<=nn;j++) { Row modication. p=a[k][j]+q*a[k+1][j]; if (k != (nn-1)) { p += r*a[k+2][j]; a[k+2][j] -= p*z; } a[k+1][j] -= p*y; a[k][j] -= p*x; } mmin = nn<k+3 ? nn : k+3; for (i=l;i<=mmin;i++) { Column modication. p=x*a[i][k]+y*a[i][k+1]; if (k != (nn-1)) { p += z*a[i][k+2]; a[i][k+2] -= p*r; } a[i][k+1] -= p*q; a[i][k] -= p; } } } } } } while (l < nn-1); } }
Sample page from NUMERICAL RECIPES IN C: THE ART OF SCIENTIFIC COMPUTING (ISBN 0-521-43108-5) Copyright (C) 1988-1992 by Cambridge University Press. Programs Copyright (C) 1988-1992 by Numerical Recipes Software. Permission is granted for internet users to make one paper copy for their own personal use. Further reproduction, or any copying of machinereadable files (including this one) to any server computer, is strictly prohibited. To order Numerical Recipes books or CDROMs, visit website http://www.nr.com or call 1-800-872-7423 (North America only), or send email to [email protected] (outside North America).
493
CITED REFERENCES AND FURTHER READING: Wilkinson, J.H., and Reinsch, C. 1971, Linear Algebra, vol. II of Handbook for Automatic Computation (New York: Springer-Verlag). [1] Golub, G.H., and Van Loan, C.F. 1989, Matrix Computations, 2nd ed. (Baltimore: Johns Hopkins University Press), 7.5. Smith, B.T., et al. 1976, Matrix Eigensystem Routines EISPACK Guide, 2nd ed., vol. 6 of Lecture Notes in Computer Science (New York: Springer-Verlag). [2]
Sample page from NUMERICAL RECIPES IN C: THE ART OF SCIENTIFIC COMPUTING (ISBN 0-521-43108-5) Copyright (C) 1988-1992 by Cambridge University Press. Programs Copyright (C) 1988-1992 by Numerical Recipes Software. Permission is granted for internet users to make one paper copy for their own personal use. Further reproduction, or any copying of machinereadable files (including this one) to any server computer, is strictly prohibited. To order Numerical Recipes books or CDROMs, visit website http://www.nr.com or call 1-800-872-7423 (North America only), or send email to [email protected] (outside North America).
where b is a random vector and is close to some eigenvalue of A. Then the solution y will be close to the eigenvector corresponding to . The procedure can be iterated: Replace b by y and solve for a new y, which will be even closer to the true eigenvector. We can see why this works by expanding both y and b as linear combinations of the eigenvectors x j of A: y=
j
j xj
b=
j
j x j
(11.7.2)
j x j
(11.7.3)
so that j = and y=
j
j j j xj j
(11.7.4)
(11.7.5)
If is close to n , say, then provided n is not accidentally too small, y will be approximately x n , up to a normalization. Moreover, each iteration of this procedure gives another power of j in the denominator of (11.7.5). Thus the convergence is rapid for well-separated eigenvalues. Suppose at the k th stage of iteration we are solving the equation (A k 1) y = bk (11.7.6)