Abstract
We provide sufficient conditions for the general sequential block Jacobi-type method to converge to the diagonal form for cyclic pivot strategies which are weakly equivalent to the column-cyclic strategy. Given a block-matrix partition \((A_{ij})\) of a square matrix \(\mathbf {A}\), the paper analyzes the iterative process of the form \(\mathbf {A}^{(k+1)} = [\mathbf {P}^{(k)}]^*\,\mathbf {A}^{(k)}\,\mathbf {Q}^{(k)}\), \(k\ge 0\), \(\mathbf {A}^{(0)}=\mathbf {A}\), where \(\mathbf {P}^{(k)}\) and \(\mathbf {Q}^{(k)}\) are elementary block matrices which differ from the identity matrix in four blocks, two diagonal and the two corresponding off-diagonal blocks. In our analysis of convergence a promising new tool is used, namely, the theory of block Jacobi operators. Typical applications lie in proving the global convergence of block Jacobi-type methods for solving standard and generalized eigenvalue and singular value problems.
Similar content being viewed by others
References
Bujanović, Z., Drmač, Z.: A contribution to the theory and practise of the block Kogbetliantz method for computing the SVD. BIT Numer. Math. 52(4), 827–849 (2012)
Chandrasekaran, S., Ipsen, I.: On rank-revealing QR factorizations. SIAM J. Matrix Anal. Appl. 15, 592–622 (1994)
Charlier, J.P., Vanbegin, M.P., Van Dooren, P.: On efficient implementation of Kogbetliantz’s algorithm for computing the singular value decomposition. Numer. Math. 52, 279–300 (1988)
Demmel, J.W.: Applied numerical linear algebra. SIAM, Philadelpia (1997)
Demmel, J.W., Veselić, K.: Jacobi’s method is more accurate than QR. SIAM J. Math. Anal. Appl. 13, 1204–1245 (1992)
Demmel, W.J., Gu, M., Eisenstat, S., Slapničar, I., Veselić, K., Drmač, Z.: Computing the singular value decomposition with high relative accuracy. Linear Algebra Appl. 299, 21–80 (1999)
de Rijk, P.P.M.: A one-sided Jacobi algorithm for computing the singular-value decomposition on a vector computer. SIAM J. Sci. Stat. Comput. 10(2), 359–371 (1989)
Dopico, F.M., Molera, J.M., Moro, J.: An orthogonal high relative accuracy algorithm for the symmetric eigenproblem. SIAM J. Math. Anal. Appl. 25(2), 301–351 (2003)
Dopico, F.M., Koev, P., Molera, J.M.: Implicit standard Jacobi gives high relative accuracy. Numer. Math. 113(2), 519–553 (2009)
Drmač, Z., Veselić, K.: New fast and accurate Jacobi SVD algorithm I. SIAM J. Math. Anal. Appl. 29(4), 1322–1342 (2008)
Drmač, Z., Veselić, K.: New fast and accurate Jacobi SVD algorithm II. SIAM J. Math. Anal. Appl. 29(4), 1343–1362 (2008)
Drmač, Z.: A global convergence proof of cyclic Jacobi methods with block rotations. SIAM J. Math. Anal. Appl. 31(3), 1329–1350 (2009)
Eberlein, P.J.: A Jacobi-like method for the automatic computation of eigenvalues and eigenvectors of an arbitrary matrix. J. Soc. Ind. Appl. Math. 10(1), 74–88 (1962)
Eberlein, P.J., Boothroyd, J.: Solution to the eigenproblem by a norm-reducing Jacobi-type method. Numer. Math. 11, 1–12 (1968)
Eberlein, P.J., Park, H.: Efficient implementation of Jacobi algorithms and Jacobi sets on distributed memory architectures. Special issue on algorithms for hypercube computers. J. Parallel Distrib. Comput. 8, 358–366 (1990)
Falk, S., Langemeyer, P.: Das Jacobische Rotations-Verfahren für reelsymmetrische Matrizen-Paare I, II. Elektronische Datenverarbeitung 30–43 (1960)
Fernando, K.V.: Linear convergence of the row cyclic Jacobi and Kogbetliantz methods. Numer. Math. 56, 73–94 (1989)
Forsythe, G.E., Henrici, P.: The cyclic Jacobi method for computing the principal values of a complex matrix. Trans. Am. Math. Soc. 94, 1–23 (1960)
Gu, M., Eisenstat, S.C.: Efficient algorithms for computing a strong rank-revealing QR factorizations. SIAM J. Sci. Comput. 17(4), 848–869 (1996)
Hansen, E.R.: On Jacobi methods and block Jacobi methods for computing matrix Eigenvalues. Ph.D. thesis, Stanford University (1960)
Hansen, E.R.: On cyclic Jacobi methods. SIAM J. Appl. Math. 11, 449–459 (1963)
Hari, V.: On the global convergence of Eberlein method for real matrices. Numer. Math. 39, 361–369 (1982)
Hari V.: On cyclic Jacobi methods for the positive definite generalized Eigenvalue problem. Ph.D. thesis, University of Hagen (1984)
Hari, V.: On the convergence of certain cyclic Jacobi-like norm-reducing processes for real matrices. Radovi Matem. 1, 121–147 (1985)
Hari, V.: On the convergence of cyclic Jacobi-like processes. Linear Algebra Appl. 81, 105–127 (1986)
Hari, V., Veselić, K.: On Jacobi methods for singular value decompositions. SIAM J. Sci. Stat. Comput. 8(5), 741–754 (1987)
Hari, V.: On sharp quadratic convergence bounds for the serial Jacobi methods. Numer. Math. 60, 375–406 (1991)
Hari, V.: Accelerating the SVD block-Jacobi method. Computing 75, 27–53 (2005)
Hari, V.: Convergence of a block-oriented quasi-cyclic Jacobi method. SIAM J. Math. Anal. Appl. 29(2), 349–369 (2007)
Hari, V., Matejaš, J.: Accuracy of two SVD algorithms for \(2\times 2\) triangular matrices. Appl. Math. Comput. 210(1), 232–257 (2009)
Hari, V., Zadelj-Martić, V.: Parallelizing the Kogbetliantz method: a first attempt. JNAIAM 2(1–2), 49–66 (2007)
Hari, V.: On block Jacobi annihilators. Proc. Algoritm. 2009, 1–13 (2009)
Hari, V., Singer, S., Singer, S.: Block-oriented J-Jacobi methods for Hermitian matrices. Linear Algebra Appl. 433(8–10), 1491–1512 (2010)
Hari, V., Singer, S., Singer, S.: Full block \(J\)-Jacobi method for Hermitian matrices. Linear Algebra Appl. 444, 1–27 (2014)
Henrici, P., Zimmermann, K.: An estimate for the norms of certain cyclic Jacobi operators. Linear Algebra Appl. 1, 289–501 (1968)
Hestenes, M.R.: Inversion of Matrices by Biorthonalization and Related Results. J. SIAM 6(1), 51–90 (1958)
Kogbetliantz, E.: Diagonalization of general complex matrices as a new method for solution of linear equations. Proc. Int. Congr. Math. Amsterdam 2, 356–357 (1954)
Kogbetliantz, E.: Solutions of linear equations by diagonalization of coefficient matrices. Q. Appl. Math. 13, 123–132 (1955)
Luk, F., Park, H.: On the equivalence and convergence of parallel Jacobi SVD algorithms. IEEE Trans. Comput. 38(6), 806–811 (1989)
Luk, F., Park, H.: On parallel Jacobi orderings. SIAM J. Sci. Stat. Comput. 10, 18–26 (1989)
Mascarenhas, W.F.: On the convergence of the Jacobi method for arbitrary orderings. SIAM J. Matrix Anal. Appl. 16, 1197–1209 (1995)
Matejaš, J.: Accuracy of the Jacobi method on scaled diagonally dominant symmetric matrices. SIAM. J. Matrix Anal. Appl. 31, 133–153 (2009)
Matejaš, J., Hari, V.: Accuracy of the Kogbetliantz method for scaled diagonally dominant triangular matrices. Appl. Math. Comput. 217, 3726–3746 (2010)
Mathias, R.: Accurate eigensystem computations by Jacobi methods. SIAM J. Math. Anal. Appl. 16(3), 977–1003 (1995)
Nazareth, L.: On the convergence of the cyclic Jacobi method. Linear Algebra Appl. 12(2), 151–164 (1975)
Novaković, V., Singer, S.: A GPU-based hyperbolic SVD algorithm. BIT Numer. Math. 51(4), 1009–1030 (2011)
Pupovci, D., Hari, V.: On the convergence of parallelized Eberlein methods. Radovi Matem. 8, 249–267 (1999)
Rhee, H.N., Hari, V.: On the global and cubic convergence of a quasi-cyclic Jacobi method. Numer. Math. 66, 97–122 (1993)
Rutishauser, H.: Une mthode pour le calcul des valeurs propres des matrices non symtriques. Comptes Rendus 259, 2758–2758 (1964)
Sameh, A.H.: On Jacobi and Jacobi-like algorithms for parallel computer. Math. Comput. 25, 579–590 (1971)
Shroff, G., Schreiber, R.: On the convergence of the cyclic Jacobi method for parallel block orderings. SIAM J. Matrix Anal. Appl. 10, 326–346 (1989)
Singer, S., Singer, S., Novaković, V., Davidović, D., Bokulić, K., Ušćumlić, A.: Three-level parallel J-Jacobi algorithms for Hermitian matrices. Appl. Math. Comput. 218(9), 5704–5725 (2012)
Singer, S., Singer, S., Novaković, V., Ušćumlić, A., Dunjko, V.: Novel modifications of parallel Jacobi algorithms. Numer. Algorithm 59(1), 1–27 (2012)
Slapničar, I.: Accurate symmetric Eigenreduction by a Jacobi method. Ph.D. thesis, University of Hagen (1992)
Slapničar, I.: Highly accurate symmetric eigenvalue decomposition and hyperbolic SVD. Linear Algebra Appl. 358, 387–424 (2002)
Slapničar, I., Hari, V.: On the quadratic convergence of the Falk–Langemeyer method for definite matrix pairs. SIAM J. Math. Anal. Appl. 12(1), 84–114 (1991)
Veselić, K., Hari, V.: A note on a one-sided Jacobi algorithm. Numer. Math. 56, 627–633 (1989)
Veselić, K.: A Jacobi Eigenreduction Algorithm for Definite Matrix Pairs. Numer. Math. 64(1), 241–269 (1993)
Voevodin, V.V.: Cislennye metody linejnoj algebry. Nauka, Moscow (1966)
Zimmermann, K.: On the convergence of the Jacobi process for ordinary and generalized eigenvalue problems. Ph.D. Thesis, Dissertation No. 4305 ETH, Zürich (1965)
Acknowledgments
The author is thankful to D. Kressner, S. B. Coban and N. Strabić for reading and improving the paper. He is especially thankful to I. Ipsen and the anonymous referees for really excellent remarks which have made the paper much better and shorter.
Author information
Authors and Affiliations
Corresponding author
Additional information
This paper is dedicated to Peter Henrici.
This work was supported by the Croatian Ministry of Science, Education and Sports, grant 037-0372783-3042.
Appendix
Appendix
Figure 2 (Fig. 3) shows behavior of the zero pattern of \(\mathbf {F}^{(k)}\) (\(\mathbf {G}^{(k)}\)) for the first cycle. In the second cycle the roles of \(\mathbf {F}^{(k)}\) and \(\mathbf {G}^{(k)}\) are interchanged since \(\mathbf {F}^{(M)}\) (\(\mathbf {G}^{(M)}\)) is strictly block lower- (upper-) triangular. Hence, these figures give insight into the zero patterns of \(\mathbf {F}^{(k)}\) and \(\mathbf {G}^{(k)}\) for all \(k\ge 0\).
1.1 Proof of Lemma 4.8
As mentioned earlier, we can assume that every second equivalence in the inserted sequence is \(\mathop {\sim }\limits ^{s}\) and the others are \(\sim \). Let \(r+1\) be the length of the chain linking \(\fancyscript{O}=\fancyscript{O}_0\) and \(\fancyscript{O}'= \fancyscript{O}_r\). Then we have two possibilities:
(i) The proof uses induction with respect to \(d\). For \(d=0\), we have just one possibility: \(r=1\) and \(\fancyscript{O}=\fancyscript{O}_0\sim \fancyscript{O}_1=\fancyscript{O}'\). The assertion assumes the form \(\Vert { \large J}\!_{{\fancyscript{O}}'}^{\,[1]} { \large J}\!_{{\fancyscript{O}}'}^{\,[2]}\Vert \le c\). Lemma 4.4(i) implies \({ \large J}\!_{{\fancyscript{O}}'}^{\,[1]}={ \large J}\!_{{\fancyscript{O}}}^{\,[1]}\) and \({ \large J}\!_{{\fancyscript{O}}'}^{\,[2]}={ \large J}\!_{{\fancyscript{O}}}^{\,[2]}\), where \({ \large J}\!_{{\fancyscript{O}}}^{\,[1]}\) denotes the Jacobi operator composed by the same Jacobi annihilators as \({ \large J}\!_{{\fancyscript{O}}'}^{\,[1]}\), and similar holds for \({ \large J}\!_{{\fancyscript{O}}}^{\,[2]}\). Hence, the assertion is implied by the assumption \(\Vert { \large J}\!_{{\fancyscript{O}}}{ }\large {J}_{{\fancyscript{O}}}\Vert _2 \le c\).
For \(d=1\), we have four possibilities: \(\fancyscript{O}\mathop {\sim }\limits ^{s}\fancyscript{O}'\), \(\fancyscript{O}\sim \fancyscript{O}_1\mathop {\sim }\limits ^{s}\fancyscript{O}'\), \(\fancyscript{O}\mathop {\sim }\limits ^{s}\fancyscript{O}_1\sim \fancyscript{O}'\) and \(\fancyscript{O}\sim \fancyscript{O}_1\mathop {\sim }\limits ^{s}\fancyscript{O}_2\sim \fancyscript{O}'\). For all these cases, the assertion assumes the form \(\chi \equiv \Vert { \large J}\!_{{\fancyscript{O}}'}^{\,[1]} { \large J}\!_{{\fancyscript{O}}'}^{\,[2]}{ \large J}\!_{{\fancyscript{O}}'}^{\,[3]}\Vert _2 \le c\).
In the first case, \(\fancyscript{O}\mathop {\sim }\limits ^{s} \fancyscript{O}'\), so we can write \( { \large J}\!_{{\fancyscript{O}}'}^{\,[t]}=\varTheta _2^{[t]}\varTheta _1^{[t]},\ { \large J}\!_{{\fancyscript{O}}}^{\,[t]}=\varTheta _1^{[t]}\varTheta _2^{[t]},\ t=1,2,3\) where \(\varTheta _1^{[t]}\) and \(\varTheta _2^{[t]}\) are made of the Jacobi annihilators which build \({ \large J}\!_{{\fancyscript{O}}'}^{\,[t]}\). Using Proposition 4.2, we have \(\chi =\Vert \varTheta _2^{[1]}\varTheta _1^{[1]}\varTheta _2^{[2]}\varTheta _1^{[2]}\varTheta _2^{[3]}\varTheta _1^{[3]}\Vert _2 \le \Vert (\varTheta _1^{[1]}\varTheta _2^{[2]})(\varTheta _1^{[2]}\varTheta _2^{[3]})\Vert _2 \). Because each parenthesis represents one Jacobi operator made of the Jacobi annihilators from \(\Upsilon _{\pi }\) and defined by the ordering \(\fancyscript{O}\), the assumption \(\Vert { \large J}\!_{{\fancyscript{O}}}{ }\large {J}_{{\fancyscript{O}}}\Vert _2 \le c\) implies \(\chi \le c^2<c\).
In the second case, we have \({ \large J}\!_{{\fancyscript{O}}_1}^{\,[t]}={ \large J}\!_{{\fancyscript{O}}}^{\,[t]}\) for \(t=1,2,3\), where \({ \large J}\!_{{\fancyscript{O}}}^{\,[t]}\) denotes the Jacobi operator composed by the same Jacobi annihilators as \({ \large J}\!_{{\fancyscript{O}}_1}^{\,[t]}\). Hence, \(\Vert { \large J}\!_{{\fancyscript{O}}_1}^{\;[t_1]}{ \large J}\!_{{\fancyscript{O}}_1}^{\,[t_2]}\Vert _2 \le c,\ t_1,t_2\in \{1,2,3\}\). The next equivalence relation is \(\mathop {\sim }\limits ^{s}\), so we can write \({ \large J}\!_{{\fancyscript{O}}'}^{\,[t]}=\varTheta _2^{[t]}\varTheta _1^{[t]}\), \({ \large J}\!_{{\fancyscript{O}}_1}^{\,[t]}=\varTheta _1^{[t]}\varTheta _2^{[t]}\) for \(t=1,2,3\). Here \(\varTheta _1^{[t]}\) and \(\varTheta _2^{[t]}\) are composed of the Jacobi annihilators from \({ \large J}\!_{{\fancyscript{O}}'}^{\,[t]}\) and are not necessarily the same as earlier. The rest of the proof is the same as earlier.
In the third case, we first note that \({ \large J}\!_{{\fancyscript{O}}_1}^{\,[t]}={ \large J}\!_{{\fancyscript{O}}'}^{\,[t]}\) for \(t=1,2,3\). For each \(t\), we start from \({ \large J}\!_{{\fancyscript{O}}'}^{\,[t]}\), rearrange the Jacobi annihilators, and obtain \({ \large J}\!_{{\fancyscript{O}}_1}^{\,[t]}\). Then we continue as in the first case. We conclude that \(\chi = \Vert { \large J}\!_{{\fancyscript{O}}_1}^{\,[1]} { \large J}\!_{{\fancyscript{O}}_1}^{\,[2]}{ \large J}\!_{{\fancyscript{O}}_1}^{\,[3]}\Vert _2 \le c\).
Finally, in the fourth case, we start making conclusions as in the third case, and then we complete the analysis as in the second case.
The Induction Hypothesis. We assume that the assertion of the lemma holds provided the chain contains \(d-1\) shift-equivalences, where \(d\ge 1\).
The Induction Step. Let the chain contain \(d\) shift-equivalences. We divide it into two parts: the head and the tail, where the tail is the right-most part of the chain. The head contains the remaining part of the chain. The form of the tail can be either \(\sim \fancyscript{O}_{r-1}\mathop {\sim }\limits ^{s}\fancyscript{O}'\) or \(\mathop {\sim }\limits ^{s}\fancyscript{O}_{r-1}\sim \fancyscript{O}'\).
Let us consider the first case. By the induction hypothesis, the relation
holds if the Jacobi operators \({ \large J}\!_{{\fancyscript{O}}_{r-1}}^{\,[t]}\) are made of the Jacobi annihilators from \(\Upsilon _{\pi }\) and are defined by \(\fancyscript{O}_{r-1}\). Let us denote the left-hand side of the inequality (4.13) by \(\chi \). We have to show that \(\chi \le c\).
Since \(\fancyscript{O}_{r-1}\mathop {\sim }\limits ^{s}\fancyscript{O}'\), we have \({ \large J}\!_{{\fancyscript{O}}'}^{\,[t]}=\varTheta _2^{[t]}\varTheta _1^{[t]}\), \({ \large J}\!_{{\fancyscript{O}}_{r-1}}^{\,[t]}=\varTheta _1^{[t]}\varTheta _2^{[t]}\), \(t=1,\ \ldots \ ,d+2\), for some \(\varTheta _1^{[t]}\) and \(\varTheta _2^{[t]}\) which are made of the Jacobi annihilators from \({ \large J}\!_{{\fancyscript{O}}'}^{\,[t]}\). We have
Here each \({ }\large {J}_{{\fancyscript{O}}_{r-1}}^{\,[t]}=\varTheta _1^{[t]}\varTheta _2^{[t+1]}\), \(1\le t\le d+1\) is the Jacobi operator built of the Jacobi annihilators appearing in \({ \large J}\!_{{\fancyscript{O}}'}^{\,[t]}\), \(1\le t\le d+2\) and defined by \(\fancyscript{O}_{r-1}\). Since the Jacobi annihilators are from \(\Upsilon _{\pi }\), we can use the induction hypothesis (7.1) to conclude that \(\chi \le c\).
If the tail has the form \(\mathop {\sim }\limits ^{s}\fancyscript{O}_{r-1}\sim \fancyscript{O}'\), the proof is quite similar. One first notes that \(\chi = \Vert { \large J}\!_{{\fancyscript{O}}_{r-1}}^{\,[1]} { \large J}\!_{{\fancyscript{O}}_{r-1}}^{\,[2]}\cdots { \large J}\!_{{\fancyscript{O}}_{r-1}}^{\,[d+2]}\Vert _2\), where for each \(t\), \({ \large J}\!_{{\fancyscript{O}}_{r-1}}^{\,[t]}\) has the Jacobi annihilators from \({ \large J}\!_{{\fancyscript{O}}'}^{\,[t]}\), only rearranged according to \(\fancyscript{O}_{r-1}\). The rest of the proof is as in the first case, but the roles of \({ \large J}\!_{{\fancyscript{O}}'}^{\,[t]}\) and \({ \large J}\!_{{\fancyscript{O}}_{r-1}}^{\,[t]}\) are replaced by \({ \large J}\!_{{\fancyscript{O}}_{r-1}}^{\,[t]}\) and \({ \large J}\!_{{\fancyscript{O}}_{r-2}}^{\,[t]}\), respectively.
(ii) The proof is very similar to the proof of the assertion (i).
Rights and permissions
About this article
Cite this article
Hari, V. Convergence to diagonal form of block Jacobi-type methods. Numer. Math. 129, 449–481 (2015). https://doi.org/10.1007/s00211-014-0647-8
Received:
Revised:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s00211-014-0647-8