Abstract
We discuss the convergence rate of the QR algorithm with Wilkinson’s shift for tridiagonal symmetric eigenvalue problems. It is well known that the convergence rate is theoretically at least quadratic, and practically better than cubic for most matrices. In an effort to derive the convergence rate, the limiting patterns of some lower right submatrices have been intensively investigated. In this paper, we first describe a new limiting pattern of the lower right 3-by-3 submatrix with a concrete example, and then prove that the convergence rate of this new pattern is strictly cubic. In addition, we stress that our analysis identifies three classes of the limiting patterns of the tridiagonal QR algorithm with Wilkinson’s shift.
Similar content being viewed by others
References
Hoffmann, W., Parlett, B.N.: A new proof of global convergence for the tridiagonal QL algorithm. SIAM J. Numer. Anal. 15, 929–937 (1978)
Huang, H., Tam, T.: On the QR iterations of real matrices. Linear Algebra Appl. 408, 161–176 (2005)
Jiang, E., Zhang, Z.: A new shift of the QL algorithm for irreducible symmetric matrices. Linear Algebra Appl. 65, 261–272 (1985)
Leitte, R.S., Saldanha, N.C., Tomei, C.: The asymptotics of Wilkinson’s shift: loss of cubic convergence. Found. Comput. Math. 10, 15–36 (2010)
Parlett, B.N.: Canonical decomposition of Hessenberg matrices. Math. Comput. 21, 223–227 (1967)
Parlett, B.N.: Global convergence of the basic QR algorithm on Hessenberg matrices. Math. Comput. 22, 803–817 (1968)
Parlett, B.N.: The symmetric eigenvalue problem. Prentice-Hall,Englewood Cliffs, New Jersey, 1980; SIAM, Philadelphia (1998)
Wang, T.: Convergence of the tridiagonal QR algorithm. Linear Algebra Appl. 322, 1–17 (2001)
Wilkinson, J.H.: The algebraic eigenvalue problem. Clarendon Press, Oxford (1965)
Wilkinson, J.H.: Global convergence of tridiagonal QR algorithm with origin shifts. Linear Algebra Appl. 1, 409–420 (1968)
Zhang, G.: On the convergence rate of the QL algorithm with Wilkinson’s shift. Linear Algebra Appl. 113, 131–137 (1989)
Acknowledgments
The author is grateful to Professor Takayasu Matsuo and Professor Beresford Parlett for their valuable comments and suggestions. The author also thanks the anonymous reviewer for the helpful comments.
Author information
Authors and Affiliations
Corresponding author
Additional information
The author is supported by JSPS Grant-in-Aid for Young Scientists (Grant Number 25790096).
Appendices
Appendix A: Proof of convergence
We prove here that all the tridiagonal elements of \(T^{(n)}\) converge for any initial matrix \(T^{(0)}\). To this end, we consider a general shift \(s^{(n)}\) satisfying the following conditions:
-
(i)
The shift \(s^{(n)}\) converges to a certain eigenvalue \(s^{(\infty )}=\lambda _{l}\);
-
(ii)
\(|s^{(n)}-\lambda _{l}|=\mathrm{o}(c^{n})\) for a positive constant \(c<1\).
Note that Wilkinson’s shift satisfies the two conditions: (i) has been proved by [8]; then the convergence rate by the shifts \(s^{(n)}\rightarrow \lambda _{l}\) is at least quadratic because \(|s^{(n)}-\lambda _{l}|\le |s^{(n)}-\alpha _{m}^{(n)}|+|\alpha _{m}^{(n)}-\lambda _{l}| \le 2|\beta _{m-1}^{(n)}|\) by the definition of Wilkinson’s shift and Gershgorin’s circle theorem, and then we have \(|\beta _{m-1}^{(n+1)}|=\mathrm{O}(|\beta _{m-1}^{(n)}|^2)\) by Wilkinson’s proof, which implies (ii).
In what follows, we prove that all the tridiagonal elements \(T^{(n)}\) always converge for such general shifts. The following convergence proof might have been noticed by the experts in this research field because its proof is almost the same as those by [2, 6, 9] for the unshifted QR algorithm. However, to the best of the authors’ knowledge, the proof for the shifted algorithm is not explicitly stated in any reference. For the readers’ convenience, we prove it as follows.
Theorem 5
Suppose that the QR algorithm with a general shift satisfying the above conditions (i) and (ii) is applied to an irreducible tridiagonal matrix T. Then, \(T^{(n)}\) converges to a block diagonal matrix whose block size is at most 2. Further, the lower right diagonal element \(T^{(n)}(m,m)\) converges to \(\lambda _{l}\) and the lower right off-diagonal element \(T^{(n)}(m,m-1)\) converges to 0.
Proof
First of all, we show several important facts for the convergence analysis. Similarly to the discussion in [6] and [9, Chapter8, §28], let \(\tilde{Q}^{(n)},\ \tilde{R}^{(n)}\) be
By the orthogonal matrix \(\tilde{Q}^{(n)}\), \(T^{(n)}\) is described as
in view of \(T^{(n+1)}=R^{(n)}Q^{(n)}+s^{(n)}I=(Q^{(n)})^\mathrm{T}(Q^{(n)}R^{(n)}+s^{(n)}I)Q^{(n)}=(Q^{(n)})^\mathrm{T}T^{(n)}Q^{(n)}\). Using (38), we see
Let p(l) denote a permutation of the indices l \((l = 1, \ldots , m)\). Then in view of the condition (i) we can place the shifted eigenvalues in a descending order as
The last inequality follows from (i) (note that all the eigenvalues are distinct since T is irreducible).
Next, we focus on the eigendecomposition
where X is the orthogonal matrix consisting of the eigenvectors and \(\varLambda \) is the diagonal matrix with the eigenvalues: \(\mathrm{diag}(\lambda _{p(1)},\ldots ,\lambda _{p(m)})\). Then we see
where
From (38) and (41), we have \(T^{(n)}=(\tilde{Q}^{(n)})^\mathrm{T}X\varLambda X^\mathrm{T}\tilde{Q}^{(n)}\). If all the inequalities are strict in (40), \(T^{(n)}\) converges to \(\varLambda \). Actually, \(T^{(n)}=(\tilde{Q}^{(n)})^\mathrm{T}X\varLambda X^\mathrm{T}\tilde{Q}^{(n)}\) always converges to a block diagonal matrix whose block size is at most 2. In order to prove it, we firstly apply the LU factorization \(X^\mathrm{T}=LU\). Note that \(X^\mathrm{T}\) constructed by the normalized eigenvectors of an irreducible tridiagonal matrix is always LU factorizable [2, 5]. Hence, we see
Combining it with (39) we have
To reveal the relationship between \(\tilde{Q}^{(n)}\) and X, we discuss the behavior of the QR factorization of \(\varLambda ^{(n)} L(\varLambda ^{(n)})^{-1}\) as follows. Let \(D_{\varLambda ^{(n)}}\) be an orthogonal matrix
It is easy to see that
In the right-hand side of (45), we have
and by applying the QR factorization we see
where \(P^{(n)}\) is an orthogonal matrix, \(\varGamma ^{(n)}\) is an upper triangular matrix whose diagonal elements are positive. Let \(D_{U}\) be an orthogonal matrix \(D_{U}=\mathrm{diag}(|u_{11}|/u_{11}, \ldots ,|u_{mm}|/u_{mm})\). Then we see
from (45), (48), and (49). Therefore, we have
Since our aim is to prove the convergence of all the elements of \(T^{(n)}\), let us discuss the behavior of the orthogonal matrix \(P^{(n)}\) as \(n\rightarrow \infty \). To this end, we focus on (49). The lower left elements are
from (47). Obviously \(\lim _{n\rightarrow \infty }(D_{\varLambda ^{(n)}}\varLambda ^{(n)} L(D_{\varLambda ^{(n)}}\varLambda ^{(n)})^{-1})_{ij}=0\), when \(|\lambda _{p(i)}-s^{(\infty )}| > |\lambda _{p(j)}-s^{(\infty )}|\). Otherwise, from (40) and the condition (ii), we have
Since a sequence of the size \(\mathrm{o}(c^{l})\) with \(0<c<1\) absolutely converges, \((D_{\varLambda ^{(n)}}\varLambda ^{(n)} L(D_{\varLambda ^{(n)}}\varLambda ^{(n)})^{-1})_{ij}\) represented by the infinite product (53) is convergent:
The resulting matrix \(\tilde{L}\) is not only unit lower triangular, but also block diagonal with the block sizes at most 2, because if the equality \(|\lambda _{p(k)}-s^{(\infty )}|=|\lambda _{p(k+1)}-s^{(\infty )}|\) in (40) holds, then both inequalities \(|\lambda _{p(k-1)}-s^{(\infty )}|>|\lambda _{p(k)}-s^{(\infty )}|\) and \(|\lambda _{p(k+1)}-s^{(\infty )}|>|\lambda _{p(k+2)}-s^{(\infty )}|\) are satisfied thanks to the fact that the eigenvalues are all distinct. Hence, the orthogonal matrix \(P^{(n)}\) given by the QR factorization of \(D_{\varLambda ^{(n)}}\varLambda ^{(n)} L(D_{\varLambda ^{(n)}}\varLambda ^{(n)})^{-1}\) is convergent:
where \(\tilde{P}\) is a block diagonal matrix whose block size is at most 2. It then follows that
from (52). Therefore, \(T^{(n)}\) converges to a block diagonal matrix whose block size is at most 2. \(\square \)
Appendix B: Proofs of Lemmas 2 and 3
We prove here Lemmas 2 and 3 in turn.
By the discussion in Appendix A, \(\lambda _{p(m-1)},\ \lambda _{p(m-2)}\) are the eigenvalues of the \(2\times 2\) submatrix
In view of \(\beta _{m-2}^{(\infty )}\not = 0\), we see \(|\lambda _{p(m-2)}-\lambda _{p(m)}|=|\lambda _{p(m-1)}-\lambda _{p(m)}|\). The eigenvalues are real and distinct, which implies \((\lambda _{p(m-2)}-\lambda _{p(m)})+(\lambda _{p(m-1)}-\lambda _{p(m)})=0\). Since the sum of the eigenvalues of the matrix (58) is equal to the trace of that, we have \((\alpha _{m-2}^{(\infty )}-s^{(\infty )}) +(\alpha _{m-1}^{(\infty )}-s^{(\infty )})=0\). In other words,
holds for constants C and D. Obviously, the eigenvalues of the matrix (59) are \(\pm \sqrt{C^2+D^2}\). It then follows that
This completes the proof of Lemma 2.
Next, we we prove Lemma 3 based on Lemma 1. Actually, in Lemma 1,
holds. We prove (61) below. If \(\lim _{n\rightarrow \infty }\beta _{m-2}^{(n)}= 0\), then \(\lim _{n\rightarrow \infty }\alpha _{m-1}^{(n)}= \lambda _{p(m-1)}\) holds. Therefore, we obtain
from (17) and (19). Next, we consider the situation \(\lim _{n\rightarrow \infty }\beta _{m-2}^{(n)} \not =0\). It is easy to see that, if \(\lim _{n\rightarrow \infty }\beta _{m-2}^{(n)} \not =0\), then \(\lim _{n\rightarrow \infty }\beta _{m-3}^{(n)}= 0\) because the block size of \(T^{(\infty )}\) is at most 2. Combining it with (17), we have
In addition, we see
in (19). Noting \(\beta _{m-3}^{(\infty )}=0\) and (17), we see that
holds. Therefore, we obtain (61) from (59) and (60).
Obviously,
holds. Moreover, if \(D\not =0\), then
in the same way as (22).
From (18), (61), (65), and (66), we have
Noting \(|\rho _{1}|\) is defined as (25), we have
from (67).
We investigate the behavior of \(|d_{m-3}^{(n)}|\). First of all, we assume \(\beta _{m-3}^{(\infty )}\not =0\). Then \(\beta _{m-4}^{(\infty )}=0\) and \(\beta _{m-2}^{(\infty )}=0\). Similarly to (59) and (60),
holds. Noting \(\beta _{m-4}^{(\infty )}=0\), we have
in the same way as (59) and (60). Moreover, noting \(\left| \lambda _{p(m-3)}-\lambda _{p(m)}\right| =\left| \lambda _{p(m-2)}\right. \left. -\lambda _{p(m)}\right| \), we obtain
Actually, it is easy to see that (70) covers the case \(\beta _{m-3}^{(\infty )}=0\). Therefore, we obtain
where the first equality is due to (68) and (70), the second equality is due to the definition of \(|\rho _{2}|\) in (26). This completes the proof of (24).
The final task is to derive (28) in the case \(\alpha _{m-1}^{(\infty )}=\alpha _{m}^{(\infty )}\). In (21), \(|\beta _{m-1}^{(n)}|/|\alpha _{m-1}^{(n)}-s^{(n)}|\le 1\) in view of the definition of Wilkinson’s shift. Hence,
from (21). Also noting \(\left| \beta _{m-2}^{(\infty )}\right| =\left| \lambda _{p(m-1)} -\lambda _{p(m)}\right| =\left| \lambda _{p(m-2)}-\lambda _{p(m)}\right| \), \(\beta _{m-3}^{(\infty )}=0\), we have
Therefore, we obtain (28) from (18), (61), and (65).
Appendix C: Convergence analysis based on perturbation theory
In this section, using perturbation theory we prove Corollary 1. In other words, we prove the following facts:
-
if \(\lim _{n\rightarrow \infty }(\alpha _{m-1}^{(n)}-\alpha _{m}^{(n)})=0\), then \(|\beta _{m-1}^{(n+1)}|=\mathrm{O}(|\beta _{m-1}^{(n)}|^2)\);
-
if \(\lim _{n\rightarrow \infty }(\alpha _{m-1}^{(n)}-\alpha _{m}^{(n)})=D\not =0\), then \(|\beta _{m-1}^{(n+1)}|=\mathrm{O}(|\beta _{m-2}^{(n)}|^2|\beta _{m-1}^{(n)}|^3)\).
In what follows, we use the so-called gap theorem [7, Theorem 11.7.1].
Lemma 4
([7]) Let y be a unit vector, A be a symmetric matrix, and \(\lambda _{l}\) be the eigenvalue of A closest to \(y^\mathrm{T}Ay\). Then
holds.
First of all, we note
as \(n\rightarrow \infty \) for the general shifts satisfying \(\beta _{m-1}^{(\infty )}=0\) and the condition (i) in Appendix A. Although this fact might be noticed by the experts, the authors do not know the literature where its proof is explicitly written. Hence, we prove (72) below. Recall that \(T_{k}^{(n)}\) for \(k=1,\ldots ,m\) are the \(k \times k\) leading principal submatrix of \(T^{(n)}\) and \(d_{k}^{(n)}=\det (T_{k}^{(n)}-s^{(n)}I)\) for \(k=1,\ldots ,m\) defined in (17). Then, (65) holds for the general shifts in view of the condition (i) and \(\lim _{n\rightarrow \infty }\beta _{m-1}^{(n)}= 0\). Obviously, \(d_{m}^{(n)}=\prod _{1\le i \le m}(\lambda _{p(i)}-s^{(n)})\). Furthermore, noting that \(\gamma ^{(n)}\) in (19) is bounded and \(\lim _{n\rightarrow \infty }\beta _{m-1}^{(n)}= 0\), in (18) we have
as \(n\rightarrow \infty \). From (61), we have (72).
For the discussion below, we describe the relation (72) more precisely. For any \(\epsilon _{1} >0\),
holds for all sufficiently large n.
In order to reveal the convergence rate, let us estimate \(|\lambda _{p(m)}-s^{(n)}|\). To this end, suppose we apply one step of the Jacobi method for the lower right 2-by-2 submatrix of \(T^{(n)}\). Then we see from [7, Chapter9] that the angle \(\theta ^{(n)}\) of Givens rotation for annihilating \(\beta _{m-1}^{(n)}\) satisfies
where \(\theta ^{(n)}\) is chosen in the interval \([-\pi /4,\ \pi /4]\). It means that the transformed matrix can be described as
where
Note that \(s^{(\infty )}=\lambda _{p(m)}\). Let
For any \(\epsilon _{2} >0\), noting Lemma 4 with \(y:=(0,0,\ldots ,1)^\mathrm{T}\), we have
for all sufficiently large n. In addition, it is easy to see that, for any \(\epsilon _{3} > 0\),
for all sufficiently large n.
Now we consider the case of \(|\alpha _{m-1}^{(n)} - \alpha _m^{(n)}| \rightarrow D \ne 0\). We obtain
for all sufficiently large n by using (74), (79), (78), (76), (75) in turn. We see \(\epsilon _{1}, \epsilon _{2}, \epsilon _{3} \rightarrow 0\) and \(|\alpha _{m-1}^{(n)}-\alpha _{m}^{(n)}| \rightarrow |D| > 0\) as \(n\rightarrow \infty \). Therefore, \(|\beta _{m-1}^{(n+1)}|=\mathrm{O}(|\beta _{m-2}^{(n)}|^{2}|\beta _{m-1}^{(n)}|^{3})\).
Finally, we prove the quadratic convergence in the case \(|\alpha _{m-1}^{(n)}-\alpha _{m}^{(n)}|\rightarrow 0\). Since the estimate (80) by the Jacobi transformation cannot derive the quadratic convergence in the case \(|\alpha _{m-1}^{(n)}-\alpha _{m}^{(n)}|\rightarrow 0\), we give another estimate based on Lemma 4. For any \(\epsilon _{4} >0\), we see
for all sufficiently large n from Lemma 4 with \(y:=(0,0,\ldots ,1)^\mathrm{T}\), \(A:=T^{(n)}\). From the definition of Wilkinson’s shift, \(|\alpha _{m}^{(n)}-s^{(n)}|\le |\beta _{m-1}^{(n)}|\) holds. Hence,
for all sufficiently large n. Thus, we have
for all sufficiently large n from (74), (79), (82). We see \(\epsilon _{3}, \epsilon _{4} \rightarrow 0\) and \(\beta _{m-1}^{(n)} \rightarrow 0\) as \(n\rightarrow \infty \). Therefore, \(|\beta _{m-1}^{(n+1)}|=\mathrm{O}(|\beta _{m-1}^{(n)}|^{2})\). This completes the proof.
Although the convergence analysis above is readily accessible to the readers in the research fields of the numerical linear algebra, we also note that the right-hand side of (80) in our proof is an overestimate in the case of \(|\lambda _{p(m-1)}-\lambda _{p(m)}|< |\lambda _{p(m-2)}-\lambda _{p(m)}|\) because we have
where the first equality is due to the assumption \(\alpha _{m-1}^{(\infty )}-\alpha _{m}^{(\infty )}=D\) and the definition of \(\delta \) in (77), the next inequality is due to (27) and the above condition \(|\lambda _{p(m-1)}-\lambda _{p(m)}|< |\lambda _{p(m-2)}-\lambda _{p(m)}|\), the next equality is due to (25), the next inequality is due to (26), and the last equality is due to (24). Also note that, in the case of \(\beta _{m-2}^{(\infty )}=C\not =0\), the above relation (83) holds because we have the strict inequality in (83) from (60). Therefore, the right-hand side of (80) is an overestimate.
About this article
Cite this article
Aishima, K. A note on the convergence theorem of the tridiagonal QR algorithm with Wilkinson’s shift. Japan J. Indust. Appl. Math. 32, 465–487 (2015). https://doi.org/10.1007/s13160-015-0171-y
Received:
Revised:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s13160-015-0171-y
Keywords
- Numerical linear algebra
- Eigensolver
- QR algorithm
- Symmetric tridiagonal matrices
- Wilkinson’s shift
- Convergence rate