the Unitary Polar Factor æ Ren-Cang Li P.O. Box 2008, Bldg 6012

Similar documents
the Unitary Polar Factor æ Ren-Cang Li Department of Mathematics University of California at Berkeley Berkeley, California 94720

deviation of D and D from similarity (Theorem 6.). The bound is tight when the perturbation is a similarity transformation D = D? or when ^ = 0. From

Key words. multiplicative perturbation, relative perturbation theory, relative distance, eigenvalue, singular value, graded matrix

A Note on Eigenvalues of Perturbed Hermitian Matrices

Department of Mathematics Technical Report May 2000 ABSTRACT. for any matrix norm that is reduced by a pinching. In addition to known

S.F. Xu (Department of Mathematics, Peking University, Beijing)

Linear Algebra Review. Vectors

On Computing Accurate Singular. Values and Eigenvalues of. Matrices with Acyclic Graphs. James W. Demmel. William Gragg.

The skew-symmetric orthogonal solutions of the matrix equation AX = B

UMIACS-TR July CS-TR 2721 Revised March Perturbation Theory for. Rectangular Matrix Pencils. G. W. Stewart.

Yimin Wei a,b,,1, Xiezhang Li c,2, Fanbin Bu d, Fuzhen Zhang e. Abstract

arxiv: v1 [math.na] 1 Sep 2018

linearly indepedent eigenvectors as the multiplicity of the root, but in general there may be no more than one. For further discussion, assume matrice

MULTIPLICATIVE PERTURBATION ANALYSIS FOR QR FACTORIZATIONS. Xiao-Wen Chang. Ren-Cang Li. (Communicated by Wenyu Sun)

Multiplicative Perturbation Analysis for QR Factorizations

Chapter 3 Transformations

arxiv: v3 [math.ra] 22 Aug 2014

Total least squares. Gérard MEURANT. October, 2008

On Orthogonal Block Elimination. Christian Bischof and Xiaobai Sun. Argonne, IL Argonne Preprint MCS-P

We first repeat some well known facts about condition numbers for normwise and componentwise perturbations. Consider the matrix

Computational math: Assignment 1

2 and bound the error in the ith eigenvector in terms of the relative gap, min j6=i j i? jj j i j j 1=2 : In general, this theory usually restricts H

1 Quasi-definite matrix

A fast randomized algorithm for overdetermined linear least-squares regression

Bounding the Spectrum of Large Hermitian Matrices

Index. book 2009/5/27 page 121. (Page numbers set in bold type indicate the definition of an entry.)

For δa E, this motivates the definition of the Bauer-Skeel condition number ([2], [3], [14], [15])

Lecture notes: Applied linear algebra Part 1. Version 2

OPTIMAL SCALING FOR P -NORMS AND COMPONENTWISE DISTANCE TO SINGULARITY

ON WEIGHTED PARTIAL ORDERINGS ON THE SET OF RECTANGULAR COMPLEX MATRICES

A DIVIDE-AND-CONQUER METHOD FOR THE TAKAGI FACTORIZATION

Institute for Advanced Computer Studies. Department of Computer Science. On the Perturbation of. LU and Cholesky Factors. G. W.

Two Results About The Matrix Exponential

Preliminary/Qualifying Exam in Numerical Analysis (Math 502a) Spring 2012

Interlacing Inequalities for Totally Nonnegative Matrices

Designing Information Devices and Systems II

Majorization for Changes in Ritz Values and Canonical Angles Between Subspaces (Part I and Part II)

Positive Denite Matrix. Ya Yan Lu 1. Department of Mathematics. City University of Hong Kong. Kowloon, Hong Kong. Abstract

Quantum Computing Lecture 2. Review of Linear Algebra

The Hermitian R-symmetric Solutions of the Matrix Equation AXA = B

Linear Algebra in Actuarial Science: Slides to the lecture

Computation of eigenvalues and singular values Recall that your solutions to these questions will not be collected or evaluated.

We use the overhead arrow to denote a column vector, i.e., a number with a direction. For example, in three-space, we write

Multiplicative Perturbation Bounds of the Group Inverse and Oblique Projection

Derivation of the Kalman Filter

POSITIVE MAP AS DIFFERENCE OF TWO COMPLETELY POSITIVE OR SUPER-POSITIVE MAPS

Singular Value Decompsition

RITZ VALUE BOUNDS THAT EXPLOIT QUASI-SPARSITY

Algorithms and Perturbation Theory for Matrix Eigenvalue Problems and the SVD

~ g-inverses are indeed an integral part of linear algebra and should be treated as such even at an elementary level.

Institute for Advanced Computer Studies. Department of Computer Science. On the Adjoint Matrix. G. W. Stewart y ABSTRACT

EIGENVALUES AND SINGULAR VALUE DECOMPOSITION

MATH36001 Generalized Inverses and the SVD 2015

MASSACHUSETTS INSTITUTE OF TECHNOLOGY Chemistry 5.76 Revised February, 1982 NOTES ON MATRIX METHODS

1 Multiply Eq. E i by λ 0: (λe i ) (E i ) 2 Multiply Eq. E j by λ and add to Eq. E i : (E i + λe j ) (E i )

ELA THE OPTIMAL PERTURBATION BOUNDS FOR THE WEIGHTED MOORE-PENROSE INVERSE. 1. Introduction. Let C m n be the set of complex m n matrices and C m n

Accuracy and Stability in Numerical Linear Algebra

PERTURBATION ANALYSIS OF THE EIGENVECTOR MATRIX AND SINGULAR VECTOR MATRICES. Xiao Shan Chen, Wen Li and Wei Wei Xu 1. INTRODUCTION A = UΛU H,

The Solvability Conditions for the Inverse Eigenvalue Problem of Hermitian and Generalized Skew-Hamiltonian Matrices and Its Approximation

Block Bidiagonal Decomposition and Least Squares Problems

Block Lanczos Tridiagonalization of Complex Symmetric Matrices

Conceptual Questions for Review

Linear Algebra Massoud Malek

Bare minimum on matrix algebra. Psychology 588: Covariance structure and factor models

Numerical Methods in Matrix Computations

EE/ACM Applications of Convex Optimization in Signal Processing and Communications Lecture 2

ON WEIGHTED PARTIAL ORDERINGS ON THE SET OF RECTANGULAR COMPLEX MATRICES

Mathematics. EC / EE / IN / ME / CE. for

MATH 304 Linear Algebra Lecture 34: Review for Test 2.

Chapter 7. Canonical Forms. 7.1 Eigenvalues and Eigenvectors

The Best Circulant Preconditioners for Hermitian Toeplitz Systems II: The Multiple-Zero Case Raymond H. Chan Michael K. Ng y Andy M. Yip z Abstract In

Some inequalities for sum and product of positive semide nite matrices

Multivariate Statistical Analysis

The Singular Value Decomposition

On the eigenvalues of specially low-rank perturbed matrices

Elementary linear algebra

Distribution for the Standard Eigenvalues of Quaternion Matrices

1 Linear Algebra Problems

Perturbation results for nearly uncoupled Markov. chains with applications to iterative methods. Jesse L. Barlow. December 9, 1992.

APPLIED NUMERICAL LINEAR ALGEBRA

Notes on Eigenvalues, Singular Values and QR

Foundations of Matrix Analysis

Ir O D = D = ( ) Section 2.6 Example 1. (Bottom of page 119) dim(v ) = dim(l(v, W )) = dim(v ) dim(f ) = dim(v )

Linear Algebra: Matrix Eigenvalue Problems

Multiple eigenvalues

On condition numbers for the canonical generalized polar decompostion of real matrices

We will discuss matrix diagonalization algorithms in Numerical Recipes in the context of the eigenvalue problem in quantum mechanics, m A n = λ m

onto the column space = I P Zj. Embed PZ j

Is there a Small Skew Cayley Transform with Zero Diagonal?

Lecture 9: Numerical Linear Algebra Primer (February 11st)

1. The Polar Decomposition

Linear Algebra and its Applications

Eigenvalues and eigenvectors

or H = UU = nx i=1 i u i u i ; where H is a non-singular Hermitian matrix of order n, = diag( i ) is a diagonal matrix whose diagonal elements are the

Ann. Funct. Anal. 5 (2014), no. 2, A nnals of F unctional A nalysis ISSN: (electronic) URL:

Singular Value Decomposition and Principal Component Analysis (PCA) I

A note on eigenvalue computation for a tridiagonal matrix with real eigenvalues Akiko Fukuda

Applied Numerical Linear Algebra. Lecture 8

Distance Geometry-Matrix Completion

Eigenvalue and Eigenvector Problems

Transcription:

Relative Perturbation Bounds for the Unitary Polar actor Ren-Cang Li Mathematical Science Section Oak Ridge National Laboratory P.O. Box 2008, Bldg 602 Oak Ridge, TN 3783-6367 èli@msr.epm.ornl.govè LAPACK working notes è 83 èrst published July, 994, revised January, 996è Abstract Let B be an m n èm ç nè complex matrix. It is known that there is a unique polar decomposition B = QH, where Q Q = I, the n n identity matrix, and H is positive denite, provided B has full column rank. Existing perturbation bounds for complex matrices suggest that in the worst case, the change in Q be proportional to the reciprocal of the smallest singular value of B. However, there are situations where this unitary polar factor is much more accurately determined by the data than the existing perturbation bounds would indicate. In this paper the following question is addressed: how much mayqchange if B is perturbed to e B = D BD 2? Here D and D 2 are nonsingular and close to the identity matrices of suitable dimensions. It will be proved that for such kinds of perturbations, the change in Q is bounded only by the distances from D and D 2 to identity matrices, and thus independent of the singular values of B. This material is based in part upon work supported, during January, 992íAugust, 995, by Argonne National Laboratory under grant No. 20552402 and by the University of Tennessee through the Advanced Research Projects Agency under contract No. DAAL03-9-C-0047, by the National Science oundation under grant No. ASC-9005933, and by the National Science Infrastructure grants No. CDA-8722788 and CDA-94056, and supported, since August, 995, by a Householder ellowship in Scientic Computing at Oak Ridge National Laboratory, supported by the Applied Mathematical Sciences Research Program, Oce of Energy Research, United States Department of Energy contract DE-AC05-84OR2400 with Lockheed Martin Energy System, Inc. Part of this work was done during summer of 994 while the author was at Department of Mathematics, University of California at Berkeley.

Ren-Cang Li: Relative Perturbation Bounds for the Unitary Polar actor 2 Let B be an m n èm ç nè complex matrix. It is known that there are Q with orthonormal column vectors, i.e., Q Q = I, and a unique positive semidenite H such that B = QH: èè Hereafter I denotes an identity matrix with appropriate dimensions which should be clear from the context or specied. The decomposition èè is called the polar decomposition of B. If, in addition, B has full column rank, then Q is uniquely determined also. In fact, H =èb Bè =2 ; Q=BèB Bè,=2 ; è2è where superscript ë " denotes conjugate transpose. The decomposition èè can also be computed from the singular value decomposition èsvdè B = UV by H = V V ; Q = U V ; è3è è! where U =èu ;U 2 è and V are unitary, U is mn,= and 0 = diag èç ;:::;ç n è is nonnegative. There are several published bounds stating how much the two factor matrices Q and H may change if entries of B are perturbed in arbitrary manner ë2, 3,6,8,9,, 2, 3, 4ë. In these papers, no assumption was made on how B was perturbed to B e except possibly an assumption on the smallness of kb e, Bk for some matrix norm kk. Roughly speaking, bounds in these published papers suggest that in the worst case, the change in Q be proportional to the reciprocal of the smallest singular value of B. In this paper, on the other hand, we study the change in Q, assuming B is complex and perturbed to B e = D BD 2, where D and D 2 are nonsingular and close to the identity matrices of suitable dimensions. Such perturbations covers, for example, component-wise relative perturbations to entries of symmetric tridiagonal matrices with zero diagonal ë5, 7ë, entries of bidiagonal and biacyclic matrices ë, 4, 5ë. Our results indicate that under such perturbations, the change in Q is independent of the smallest singular value of B. Assume that B has full column rank and so does B e = D BD 2. Let be the polar decompositions of B and e B respectively, and let B = QH; e B = e Q e H è4è B = UV ; e B = e U e e V è5è be the SVDs of B and e B, respectively, where e U =è e U ; e U 2 è, e U is m n, and e = and e = diag èeç ;:::;eç n è. Assume as usual that è e 0 ç ç ç ç n é0; and eç ç ç eç n é0: è6è It follows from è2è and è5è that Q = U V and e Q = e U e V.!

Ren-Cang Li: Relative Perturbation Bounds for the Unitary Polar actor 3 In what follows, kxk denotes the robenius norm which is the square root of the trace of X X. Notice that and eu è e B, BèV = e e V V, e U U; eu è e B, BèV = e U èd BD 2, D B + D B, BèV = e U h ebèi, D, 2 è+èd,ièb iv = e e V èi,d, 2 èv + e U èd,ièu; U è e B, Bè e V = U e U e, V e V; U èb,bè e V e = U èd BD 2, BD 2 + BD 2, BèV e h = U èi, D, è e B + BèD 2, Iè i V e to obtain two perturbation equations: = U èi, D, è e U e +V èd 2,Iè e V: e V e V, U e U = e V e èi, D, 2 èv + U e èd, IèU; è7è U U e e, V V e = U èi, D, U+V e èd 2,IèV: e è8è The rst n rows of equation è7è yields e e V V, e U U = e e V èi, D, 2 èv + e U èd, IèU : è9è The rst n rows of equation è8è yields U e U e, V V e = U èi, D, è e U e + V èd 2,IèV; e on taking conjugate transpose of which, one has e e U U, e V V = e e U èi, D, èu + e V èd 2, IèV : è0è Now subtracting è0è from è9è leads to e è e U U, e V V è+è e U U, e V Vè = e h eu èi, D, èu, e V èi, D, 2 èv i + To continue, we need a lemma from Li ë0ë. h ev èd 2, IèV, U e i èd, IèU : Lemma Let 2 C ss and, 2 C tt be two Hermitian matrices, and let E; 2 C st. If çèè T çè,è = ;, where çè è is the spectrum of a matrix, then q matrix equation X, X, =E+,has a unique solution, and moreover kxk ç kek 2 +. k k2 ç, where ç def = min!2çèè;2çè,è j!,j pj!j 2 +jj 2. èè

Ren-Cang Li: Relative Perturbation Bounds for the Unitary Polar actor 4 Apply this lemma to equation èè with = e,,=,, and X = e U U, e V V = e V è e V e U U V, IèV = e V è e Q Q, IèV; min E = U e èi, D, èu, V e èi, D, 2 èv; ee = V e èd 2, IèV, U e èd, IèU q kek 2 + k Ek e 2 =ç, where ç = to get kxk = k e Q Q, Ik ç q eç i +ç j çi; jçn eç 2 i +ç2 j Theorem Let B and B e = D BD 2 be two m n èm ç nè complex matrices having full column rank and with polar decompositions è4è. Then r ç kq, Qk e 2 ç ki, D, k + ki, D, 2 ç k +èkd2,ik +kd,ik è 2 ; è2è p q ç 2 ki, D, k2 + ki, D, 2 k2 + kd 2, Ik 2 + kd, Ik 2 : è3è Proof: Inequality è3è follows from è2è by the fact that è+è 2 ç 2è 2 + 2 è for ; ç 0. Now we prove è2è. When m = n, both Q and e Q are unitary. Thus k e Q Q, Ik = k e Q èq, e Qèk = kq, e Qk. Inequality è2è is a consequence of kek çki,d, k +ki,d, 2 k and k e Ek çkd 2,Ik +kd,ik : è4è or the case mén, the last m, n rows of equation è8è produce that U 2 e U e =U 2èI,D, è e U e : Since e B has full column rank, e is nonsingular. We haveu 2 e U =U 2 èi,d, è e U : Thus ç. ku 2 e U k çku 2 èi,d, èk = kèi, D, èu 2k : è5è Notice that èu V ;U 2 è=èq; U 2 è is unitary. Hence U 2 Q = 0, and è kq, Qk e = kèq; U 2 è èq, Qèk e I, Q = e! Q,U e 2 Q = qki, Q Qk e 2 + e q k,u2 U V e k 2 ç kek 2 +ke Ek 2 e +ku 2U k 2: è6è By è5è, we get kek 2 + ku e ç ç 2 2 U k 2 ç kèi, D, èu k + ki, D, 2 k + kèi, D, èu 2k 2 = kèi, D, èu k 2 +2kèI,D, èu k ki,d, 2 k +ki,d, 2 k2 +kèi,d, èu 2k 2 ç kèi,d, èu k 2 +kèi,d, èu 2k 2 +2kI, D, k ki, D, 2 k + ki, D, 2 k2 = ki, D, k2 +2kI,D, k ki,d, 2 k +ki,d, 2 k2 = ç 2: ki,d, k +ki,d, 2 ç k

Ren-Cang Li: Relative Perturbation Bounds for the Unitary Polar actor 5 Inequality è2è is a consequence of è6è, è4è, and the above inequalities. Now we are in the position to apply Theorem to perturbations for one-side scaling. Here we consider two n n nonsingular matrices B = S G and B e = S G, e where S is a scaling matrix and usually diagonal. èbut this is not necessary to the theorem below.è The elements of S can vary wildly. G is nonsingular and usually better conditioned than B itself. Set G def = G e, G: eg is guaranteed nonsingular if kgk 2 kg, k 2 é which will be assumed henceforth. Notice that eb = S e G = S èg +Gè=S GèI+G, ègèè = BèI + G, ègèè: So applying Theorem with D = 0 and D 2 = I + G, ègè leads to Theorem 2 Let B = S G and e B = S e G be two n n nonsingular matrices with the polar decompositions è4è. If kgk 2 kg, k 2 é then kq, e Qk ç ç r kg, ègèk 2 + I, èi + G, ègèè, 2 ; s + è,kg, k 2 kgk 2 è 2 kg, k 2 kgk : One can deal with one-side scaling from the right in the same way. Acknowledgement: I thank Professor W. Kahan for his supervision and Professor J. Demmel for valuable discussions.

Ren-Cang Li: Relative Perturbation Bounds for the Unitary Polar actor 6 References ëë J. Barlow and J. Demmel. Computing accurate eigensystems of scaled diagonally dominant matrices. SIAM Journal on Numerical Analysis, 27:762í79, 990. ë2ë A. Barrlund. Perturbation bounds on the polar decomposition. 990. BIT, 30:0í3, ë3ë C.-H. Chen and J.-G. Sun. Math., 7:397í40, 989. Perturbation bounds for the polar factors. J. Comp. ë4ë J. Demmel and W. Gragg. On computing accurate singular values and eigenvalues of matrices with acyclic graphs. Linear Algebra and its Application, 85:203í27, 993. ë5ë J. Demmel and W. Kahan. Accurate singular values of bidiagonal matrices. SIAM Journal on Scientic and Statistical Computing, :873í92, 990. ë6ë N. J. Higham. Computing the polar decompositioníwith applications. SIAM Journal on Scientic and Statistical Computing, 7:60í74, 986. ë7ë W. Kahan. Accurate eigenvalues of a symmetric tridiagonal matrix. Technical Report CS4, Computer Science Department, Stanford University, Stanford, CA, 966. èrevised June 968è. ë8ë C. Kenney and A. J. Laub. Polar decompostion and matrix sign function condition estimates. SIAM Journal on Scientic and Statistical Computing, 2:488í504, 99. ë9ë R.-C. Li. A perturbation bound for the generalized polar decomposition. BIT, 33:304í 308, 993. ë0ë R.-C. Li. Relative perturbation theory: èiiè eigenspace and singular subspace variations. Technical Report UCBèèCSD-94-856, Computer Science Division, Department of EECS, University of California at Berkeley, 994. èrevised January 996è. ëë R.-C. Li. New perturbation bounds for the unitary polar factor. SIAM J. Matrix Anal. Appl., 6:327í332, 995. ë2ë J.-Q. Mao. The perturbation analysis of the product of singular vector matrices UV H. J. Comp. Math., 4:245í248, 986. ë3ë R. Mathias. Perturbation bounds for the polar decomposition. SIAM J. Matrix Anal. Appl., 4:588í597, 993. ë4ë J.-G. Sun and C.-H. Chen. Generalized polar decomposition. Math. Numer. Sinica, :262í273, 989. In Chinese.