Matrix Analysis Matrix Analysis for Scientists for Scientists & & Engineers Engineers This page intentionally intentionally left left blank blank This page Matrix Matrix Analysis Analysis for Scientists Engineers for Scientists & & Engineers Alan J. J. Laub Alan Laub University of California Davis, California slam. Copyright © 2005 by the the Society Society for Industrial and and Applied Mathematics. Copyright 2005 by for Industrial Applied Mathematics. 10987654321 10987654321 All America. No this book All rights rights reserved. reserved. Printed Printed in in the the United United States States of of America. No part part of of this book may be be reproduced, reproduced, stored, stored, or or transmitted transmitted in in any any manner manner without the written may without the written permission permission of the publisher. For For information, information, write to the the Society Society for Industrial and Applied of the publisher. write to for Industrial and Applied Mathematics, Mathematics, 3600 3600 University University City City Science Science Center, Center, Philadelphia, Philadelphia, PA PA 19104-2688. 19104-2688. MATLAB® is is a a registered registered trademark trademark of The MathWorks, MathWorks, Inc. Inc. For For MATLAB MATLAB product product information, information, MATLAB® of The please contact The Apple Hill 01760-2098 USA, USA, please contact The MathWorks, MathWorks, Inc., Inc., 3 3 Apple Hill Drive, Drive, Natick, Natick, MA MA 01760-2098 508-647-7000, Fax: Fax: 508-647-7101, 508-647-7101, info@mathworks.com, www.mathworks.com 508-647-7000, info@mathworks.com, wwwmathworks.com Mathematica is is a a registered registered trademark trademark of of Wolfram Wolfram Research, Research, Inc. Mathematica Inc. Mathcad is is a a registered registered trademark of Mathsoft Mathsoft Engineering Engineering & & Education, Education, Inc. Mathcad trademark of Inc. Library of of Congress Congress Cataloging-in-Publication Cataloging-in-Publication Data Data Library Laub, Alan J., 19481948Laub, Alan J., Matrix analysis scientists and and engineers engineers // Alan Matrix analysis for for scientists Alan J. J. Laub. Laub. p. cm. cm. p. Includes bibliographical bibliographical references references and and index. Includes index. ISBN 0-89871-576-8 0-89871-576-8 (pbk.) (pbk.) ISBN 1. Matrices. Matrices. 2. 2. Mathematical Mathematical analysis. analysis. I.I. Title. Title. 1. QA188138 2005 QA 188.L38 2005 512.9'434—dc22 512.9'434-dc22 2004059962 2004059962 About the cover: cover: The The original original artwork artwork featured on the cover was created by by freelance About the featured on the cover was created freelance permission . artist Aaron Tallon artist Aaron Tallon of of Philadelphia, Philadelphia, PA. PA. Used Used by by permission. • slam 5.lam... is a a registered registered trademark. is trademark. To To my my wife, wife, Beverley Beverley (who captivated captivated me in the UBC UBC math library nearly forty years ago) nearly forty This page intentionally intentionally left left blank blank This page Contents Contents Preface Preface xi xi 11 Introduction Introduction and and Review Review 1.1 Notation and 1.1 Some Some Notation and Terminology Terminology 1.2 Matrix Matrix Arithmetic 1.2 Arithmetic . . . . . . . . 1.3 Inner Inner Products and Orthogonality 1.3 Products and Orthogonality . 1.4 Determinants 1.4 Determinants 11 11 33 4 44 2 2 Vector Vector Spaces Spaces 2.1 Definitions Examples . 2.1 Definitions and and Examples 2.2 Subspaces......... 2.2 Subspaces 2.3 2.3 Linear Linear Independence Independence . . . 2.4 Sums and Intersections Intersections of 2.4 Sums and of Subspaces Subspaces 77 77 99 10 10 13 13 33 Linear Linear Transformations Transformations 3.1 Definition Definition and Examples . . . . . . . . . . . . . 3.1 and Examples 3.2 Matrix Representation of Linear 3.2 Matrix Representation of Linear Transformations Transformations 3.3 Composition Transformations . . 3.3 Composition of of Transformations 3.4 Structure of Linear Linear Transformations Transformations 3.4 Structure of 3.5 3.5 Four Four Fundamental Fundamental Subspaces Subspaces . . . . 17 17 17 17 18 18 19 19 20 20 22 22 4 4 Introduction Introduction to to the the Moore-Penrose Moore-Penrose Pseudoinverse Pseudoinverse 4.1 Definitions and Characterizations Characterizations. 4.1 Definitions and 4.2 Examples.......... 4.2 Examples 4.3 Properties and and Applications Applications . . . . 4.3 Properties 29 29 30 30 31 31 55 Introduction Introduction to to the the Singular Singular Value Value Decomposition Decomposition 5.1 5.1 The The Fundamental Fundamental Theorem Theorem . . . 5.2 Some Basic Properties Properties . . . . . 5.2 Some Basic 5.3 Row and Column Compressions 5.3 Rowand Column Compressions 35 35 35 35 38 40 6 6 Linear Linear Equations Equations 6.1 Vector Vector Linear Linear Equations Equations . . . . . . . . . 6.1 6.2 Matrix Linear Equations Equations . . . . . . . . 6.2 Matrix Linear 6.3 6.3 A A More More General General Matrix Matrix Linear Linear Equation Equation 6.4 Some Useful and and Interesting Inverses. 6.4 Some Useful Interesting Inverses 43 43 43 43 vii 44 47 47 47 47 viii viii Contents Contents 7 Projections, Inner Product Spaces, and Norms 7.1 Projections . . . . . . . . . . . . . . . . . . . . . . 7.1 Projections 7.1.1 The fundamental orthogonal orthogonal projections projections 7.1.1 The four four fundamental 7.2 Inner Product Product Spaces Spaces 7.2 Inner 7.3 7.3 Vector Vector Norms Norms 7.4 Matrix Norms Norms . . . . 7.4 Matrix 51 51 51 51 52 52 54 54 57 57 59 59 8 Linear Least Squares Problems 8.1 Linear Least Least Squares Problem . . . . . . . . . . . . . . 8.1 The The Linear Squares Problem 8.2 8.2 Geometric Geometric Solution Solution . . . . . . . . . . . . . . . . . . . . . . 8.3 Linear Regression Regression and and Other 8.3 Linear Other Linear Linear Least Least Squares Squares Problems Problems 8.3.1 Linear regression 8.3.1 Example: Example: Linear regression . . . . . . . 8.3.2 problems . . . . . . . 8.3.2 Other Other least least squares squares problems 8.4 Least Squares 8.4 Least Squares and and Singular Singular Value Value Decomposition Decomposition 8.5 Least Squares and QR Factorization Factorization . . . . . . . 8.5 Least Squares and QR 65 65 65 65 67 67 67 67 67 67 69 70 70 71 71 9 Eigenvalues and Eigenvectors 9.1 Fundamental Definitions Definitions and Properties 9.1 Fundamental and Properties 9.2 Jordan Jordan Canonical Canonical Form Form . . . . . 9.2 the JCF 9.3 Determination of 9.3 Determination of the JCF . . . . . 9.3.1 Theoretical computation . 9.3.1 Theoretical computation l's in in JCF blocks 9.3.2 On the + 9.3.2 On the +1's JCF blocks 9.4 Geometric Aspects of JCF of the the JCF 9.4 Geometric Aspects 9.5 The The Matrix Sign Function Function. 9.5 Matrix Sign 75 75 75 82 82 85 85 86 86 88 88 89 89 91 91 10 Canonical Forms 10.1 Basic Canonical 10.1 Some Some Basic Canonical Forms Forms . 10.2 Definite 10.2 Definite Matrices Matrices . . . . . . . 10.3 Equivalence Transformations Transformations and 10.3 Equivalence and Congruence Congruence 10.3.1 matrices and 10.3.1 Block Block matrices and definiteness definiteness 10.4 Rational Canonical 10.4 Rational Canonical Form Form . . . . . . . . . 95 95 95 95 99 102 102 104 104 104 104 11 Linear Differential and and Difference Difference Equations Equations 11 Linear Differential 11.1 Differential ILl Differential Equations Equations . . . . . . . . . . . . . . . . 11.1.1 matrix exponential 11.1.1 Properties Properties ofthe of the matrix exponential . . . . 11.1.2 11.1.2 Homogeneous Homogeneous linear linear differential differential equations equations 11.1.3 11.1.3 Inhomogeneous Inhomogeneous linear linear differential differential equations equations 11.1.4 Linear matrix differential equations 11.1.4 Linear matrix differential equations . . 11.1.5 decompositions . . . . . . . . . 11.1.5 Modal Modal decompositions matrix exponential 11.1.6 11.1.6 Computation Computation of of the the matrix exponential 11.2 Difference Equations . . . . . . . . . . . . . . 11.2 Difference Equations 11.2.1 linear difference difference equations 11.2.1 Homogeneous Homogeneous linear equations 11.2.2 Inhomogeneous linear difference equations 11.2.2 Inhomogeneous linear difference equations 11.2.3 powers . 11.2.3 Computation Computation of of matrix matrix powers Equations. . . . . . . . . . . . . . . 11.3 Higher-Order Equations 11.3 Higher-Order 109 109 109 109 109 109 112 112 112 112 113 113 114 114 114 114 118 118 118 118 118 118 119 119 120 120 Contents Contents ix ix 12 Generalized Eigenvalue Eigenvalue Problems Problems 12 Generalized 12.1 The Generalized EigenvaluelEigenvector 12.1 The Generalized Eigenvalue/Eigenvector Problem Problem 12.2 Forms . . . . . . . . . . . . . . . . . 12.2 Canonical Canonical Forms 12.3 Application to to the the Computation of System Zeros . 12.3 Application Computation of System Zeros 12.4 Generalized Eigenvalue Eigenvalue Problems 12.4 Symmetric Symmetric Generalized Problems . 12.5 Simultaneous Simultaneous Diagonalization 12.5 Diagonalization . . . . . . . . . 12.5.1 Simultaneous Simultaneous diagonalization 12.5.1 diagonalization via via SVD SVD 12.6 Higher-Order Higher-Order Eigenvalue Problems .. 12.6 Eigenvalue Problems 12.6.1 Conversion Conversion to first-order form form 12.6.1 to first-order 125 125 125 127 127 130 131 131 133 133 133 135 135 135 13 Kronecker 13 Kronecker Products Products 13.1 Definition and Examples Examples . . . . . . . . . . . . . 13.1 Definition and 13.2 Properties Properties of of the the Kronecker Kronecker Product Product . . . . . . . 13.2 13.3 Application to to Sylvester and Lyapunov Lyapunov Equations Equations 13.3 Application Sylvester and 139 139 139 139 140 144 144 Bibliography Bibliography 151 Index Index 153 This page intentionally intentionally left left blank blank This page Preface Preface This intended to for beginning (or even even senior-level) This book book is is intended to be be used used as as aa text text for beginning graduate-level graduate-level (or senior-level) students in the sciences, sciences, mathematics, computer science, science, or students in engineering, engineering, the mathematics, computer or computational computational science who wish to be familar with enough prepared to science enough matrix analysis analysis that they they are are prepared to use its tools and ideas comfortably in aa variety variety of applications. By By matrix matrix analysis analysis II mean mean linear tools and ideas comfortably in of applications. linear algebra and and matrix application to algebra matrix theory theory together together with with their their intrinsic intrinsic interaction interaction with with and and application to linear linear differential text linear dynamical dynamical systems systems (systems (systems of of linear differential or or difference difference equations). equations). The The text can be used used in one-quarter or or one-semester one-semester course course to to provide provide aa compact compact overview of can be in aa one-quarter overview of much important and and useful useful mathematics mathematics that, that, in many cases, cases, students meant to to learn learn much of of the the important in many students meant thoroughly somehow didn't manage to topics thoroughly as as undergraduates, undergraduates, but but somehow didn't quite quite manage to do. do. Certain Certain topics that may may have have been been treated treated cursorily cursorily in in undergraduate undergraduate courses courses are treated in more depth that are treated in more depth and more more advanced is introduced. only the and advanced material material is introduced. II have have tried tried throughout throughout to to emphasize emphasize only the more important and "useful" tools, methods, and mathematical structures. Instructors are encouraged to supplement the book book with with specific specific application from their their own own encouraged to supplement the application examples examples from particular area. particular subject subject area. The choice of algebra and and matrix matrix theory theory is is motivated motivated both both by by The choice of topics topics covered covered in in linear linear algebra applications and computational utility relevance. The The concept of matrix applications and by by computational utility and and relevance. concept of matrix factorization factorization is is emphasized emphasized throughout throughout to to provide provide aa foundation foundation for for aa later later course course in in numerical numerical linear linear algebra. are stressed than abstract vector spaces, spaces, although although Chapters and 3 3 algebra. Matrices Matrices are stressed more more than abstract vector Chapters 22 and do cover cover some geometric (i.e., subspace) aspects aspects of fundamental do some geometric (i.e., basis-free basis-free or or subspace) of many many of of the the fundamental notions. The books by Meyer [18], Noble and Daniel [20], Ortega Ortega [21], and Strang [24] are excellent companion companion texts for this book. Upon course based based on on this this are excellent texts for this book. Upon completion completion of of aa course text, the student is then then well-equipped to pursue, pursue, either via formal formal courses through selftext, the student is well-equipped to either via courses or or through selfstudy, follow-on topics on the computational side (at the level of [7], [II], [11], [23], or [25], for example) or or on on the side (at level of [12], [13], [13], or [16], for example). of [12], or [16], for example). example) the theoretical theoretical side (at the the level essentially just an understanding Prerequisites for for using this this text are quite modest: essentially understanding of and definitely some previous previous exposure to matrices matrices and linear algebra. Basic of calculus calculus and definitely some exposure to and linear algebra. Basic concepts such such as determinants, singularity singularity of eigenvalues and concepts as determinants, of matrices, matrices, eigenvalues and eigenvectors, eigenvectors, and and positive definite matrices matrices should have been covered at least least once, even though their recollection may occasionally occasionally be be "hazy." However, requiring requiring such material as as prerequisite prerequisite permits tion may "hazy." However, such material permits the early "out-of-order" by standards) introduction of topics the early (but (but "out-of-order" by conventional conventional standards) introduction of topics such such as as pseupseudoinverses and and the singular value decomposition (SVD). tools doinverses the singular value decomposition (SVD). These These powerful powerful and and versatile versatile tools can can then be exploited exploited to to provide a unifying foundation foundation upon which to base subsequent subsequent toptopics. Because tools tools such the SVD are not not generally generally amenable to "hand "hand computation," computation," this this ics. Because such as as the SVD are amenable to approach necessarily availability of of appropriate mathematical software software on appropriate mathematical on approach necessarily presupposes presupposes the the availability aa digital digital computer. computer. For For this, this, II highly highly recommend recommend MAlLAB® MATLAB® although although other other software software such such as as xi xi xii xii Preface Preface Mathcad® is also excellent. Since this text is not intended for a course in Mathematica® or Mathcad® numerical linear algebra per per se, se, the details of most of the numerical aspects of linear algebra are deferred to are deferred to such such aa course. course. The presentation of the material in this book is is strongly influenced influenced by by computacomputational issues for two principal reasons. First, "real-life" "real-life" problems seldom yield to simple closed-form closed-form formulas or solutions. They must generally be solved computationally and it is important to know which types of algorithms can be relied upon and which cannot. Some of of the numerical linear linear algebra, form the Some the key key algorithms algorithms of of numerical algebra, in in particular, particular, form the foundation foundation virtually all of modern modem scientific and engineering computation. A second upon which rests virtually motivation for a computational emphasis is that it provides many of the essential tools for what I call "qualitative mathematics." mathematics." For example, in an elementary linear algebra course, a set of vectors is either linearly independent or it is not. This is an absolutely fundamental fundamental concept. But in most engineering or scientific contexts we want to know more than that. If linearly independent, independent, how "nearly dependent" are the vectors? If If a set of vectors is linearly If they are linearly dependent, are there "best" linearly independent subsets? These tum turn out to be more difficult difficult problems frequently involve involve research-level research-level questions questions when be much much more problems and and frequently when set set in the context of of the finite-precision, finite-range floating-point arithmetic environment of of most modem modern computing platforms. Some of of the the applications applications of of matrix matrix analysis analysis mentioned mentioned briefly briefly in in this this book book derive modem state-space from the modern state-space approach to dynamical systems. State-space State-space methods are modem engineering where, for example, control systems with now standard standard in much of modern large numbers numbers of interacting inputs, outputs, and states often give rise to models models of very high order that must be analyzed, simulated, and evaluated. The "language" in which such described involves vectors and matrices. It is thus crucial to acquire models are conveniently described knowledge of the vocabulary vocabulary and grammar of this language. The tools of matrix a working knowledge analysis are also applied applied on a daily basis to problems in biology, chemistry, econometrics, physics, statistics, and a wide variety of other fields, and thus the text can serve a rather diverse audience. audience. Mastery of the material in this text should enable the student to read and diverse understand the modern modem language of matrices used throughout mathematics, science, and engineering. prerequisites for this text are modest, and while most material is developed developed from While prerequisites basic ideas in the book, the student does require a certain amount of what is conventionally referred to as "mathematical maturity." Proofs Proofs are given for many theorems. When they are referred not given explicitly, obvious or or easily easily found found in literature. This This is is ideal ideal not given explicitly, they they are are either either obvious in the the literature. material from which to learn a bit about mathematical proofs and the mathematical maturity and insight gained thereby. It is my firm conviction conviction that such maturity is neither neither encouraged nor nurtured by relegating the mathematical aspects of applications (for example, linear algebra for elementary state-space theory) to introducing it "on-the-f1y" "on-the-fly" when algebra to an appendix or introducing foundation upon necessary. Rather, Rather, one must must lay lay a firm firm foundation upon which which subsequent applications and and perspectives can be built in a logical, consistent, and coherent fashion. perspectives I have taught this material for many years, many times at UCSB and twice at UC Davis, course has successful at enabling students students from from Davis, and and the the course has proven proven to to be be remarkably remarkably successful at enabling disparate backgrounds to acquire a quite acceptable acceptable level of mathematical maturity and graduate studies in a variety of disciplines. Indeed, many students who rigor for subsequent graduate completed the course, especially especially the first few times it was offered, offered, remarked afterward that completed if only they had had this course before they took linear systems, or signal processing. processing, if Preface Preface xiii XIII or estimation theory, etc., they would have been able to concentrate on the new ideas deficiencies in their they wanted to learn, rather than having to spend time making up for deficiencies background in matrices and linear algebra. My fellow instructors, too, realized that by background requiring this course as a prerequisite, they no longer had to provide as much time for "review" and could focus instead on the subject at hand. The concept seems to work. -AJL, — AJL, June 2004 This page intentionally intentionally left left blank blank This page Chapter 1 Chapter 1 Introduction and and Review Introduction Review 1.1 1.1 Some Notation Notation and and Terminology Terminology Some We begin with with aa brief brief introduction notation and used We begin introduction to to some some standard standard notation and terminology terminology to to be be used throughout the text. This This is review of of some some basic notions in throughout the text. is followed followed by by aa review basic notions in matrix matrix analysis analysis and linear linear algebra. algebra. and The The following following sets sets appear appear frequently frequently throughout throughout subsequent subsequent chapters: chapters: 1. Rnn== the the set set of of n-tuples n-tuples of of real real numbers as column column vectors. vectors. Thus, Thus, xx Ee Rn I. IR numbers represented represented as IR n means means where Xi xi Ee R for ii Ee !!. n. IR for where Henceforth, the notation!! notation n denotes denotes the the set set {I, {1, ... ..., , nn}. Henceforth, the }. Note: Vectors Vectors are vectors. A vector is where Note: are always always column column vectors. A row row vector is denoted denoted by by y~ yT, where yy G E Rn IR n and and the the superscript superscript T T is is the the transpose transpose operation. operation. That That aa vector vector is is always always aa column vector vector rather rather than row vector vector is entirely arbitrary, arbitrary, but this convention convention makes makes column than aa row is entirely but this it text that, x TTyy is while it easy easy to to recognize recognize immediately immediately throughout throughout the the text that, e.g., e.g., X is aa scalar scalar while T xy is an an nn xx nn matrix. xyT is matrix. en 2. Cn = the the set set of of n-tuples n-tuples of of complex complex numbers numbers represented represented as as column column vectors. vectors. 2. 3. IR xn = Rrnmxn = the the set set of of real real (or (or real-valued) real-valued) m m xx nn matrices. matrices. 4. 1R;n xn Rmxnr = xn denotes = the set set of of real real m x n matrices of of rank rank r. Thus, Thus, IR~ Rnxnn denotes the the set set of of real real nonsingular matrices. nonsingular n n xx nn matrices. e mxn 5. = 5. Crnxn = the the set set of of complex complex (or (or complex-valued) complex-valued) m xx nn matrices. matrices. 6. e;n xn Cmxn = n matrices = the the set set of of complex complex m m xx n matrices of of rank rank r. r. 1 Chapter 1. 1. Introduction Introduction and and Review Review Chapter 22 We now classify some of the more familiar "shaped" matrices. A matrix A Ee IRn xn x (or A A E enxn ) is eC" ")is diagonal if if aij a,7 == 00 for forii i= ^ }.j. •• diagonal upper triangular triangular if if aij a,; == 00 for forii >> }.j. •• upper lower triangular triangular if if aij a,7 == 00 for for i/ << }.j. •• lower tridiagonal if if aij a(y = = 00 for for Ii|z -—JI j\ > > 1. •• tridiagonal 1. pentadiagonal if if aij ai; = = 00 for for Ii|/ -—J j\I >> 2. •• pentadiagonal 2. upper Hessenberg Hessenberg if if aij afj == 00 for for ii -— jj >> 1. •• upper 1. lower Hessenberg Hessenberg if if aij a,; == 00 for for }j -—ii >> 1. •• lower 1. Each of the above also has a "block" analogue obtained by replacing scalar components in nxn mxn the respective definitions definitions by block block submatrices. submatrices. For For example, example, if if A Ee IR Rnxn , , B Ee IR R nxm ,, and C Ee jRmxm, Rmxm, then then the the (m (m + n) n) xx (m (m + n) n) matrix matrix [~ [A0Bc block upper upper triangular. triangular. ~]] isisblock C T A is AT and is the matrix whose entry The transpose of The of aa matrix matrix A is denoted denoted by by A and is the matrix whose (i, j)th j)th entry 7 mx A, that is, (AT)ij A E jRmxn, AT7" e E jRnxm. is the (j, (7, i)th Oth entry of A, (A ),, = aji. a,,. Note that if A e R ", then A E" xm . If A Ee em If A C mxxn, ", then its Hermitian Hermitian transpose (or conjugate conjugate transpose) is denoted by AHH (or H sometimes A*) and j)th entry is (AH)ij the bar bar indicates sometimes A*) and its its (i, j)\h entry is (A ), 7 = = (aji), («77), where where the indicates complex complex = a IX + jf$ jfJ (j = ii = jfJ. A A is conjugation; i.e., i.e., if z = (j = = R), v^T), then z = = IX a -— jfi. A matrix A is symmetric T H if A = A T and Hermitian A = A H. We henceforth if A = A Hermitian if A = A . We henceforth adopt the convention that, that, unless otherwise noted, an equation equation like = A ATT implies implies that that A is real-valued real-valued while while aa statement A = A is statement otherwise noted, an like A H like A A = AH implies that A A is complex-valued. = A complex-valued. z Remark While \/—\ most commonly commonly denoted denoted by in mathematics mathematics texts, Remark 1.1. While R isis most by ii in texts, }j is is the common notation notation in in electrical and system system theory. is some some the more more common electrical engineering engineering and theory. There There is advantage to being conversant with both notations. The notation j is used throughout the text but but reminders reminders are text are placed placed at at strategic strategic locations. locations. Example 1.2. 1.2. Example ~ 1. A = [ ; 2. A 5 = [ 7+} 3 · A -- [ 7 -5 j is symmetric symmetric (and Hermitian). ] is (and Hermitian). 7+ is complex-valued symmetric but Hermitian. 2 j ] is complex-valued symmetric but not not Hermitian. 7+} is Hermitian Hermitian (but symmetric). 2 ] is (but not not symmetric). Transposes block matrices be defined defined in obvious way. is Transposes of of block matrices can can be in an an obvious way. For For example, example, it it is easy to to see see that that if if A,, are appropriately appropriately dimensioned dimensioned subblocks, subblocks, then easy Aij are then r = [ 1.2. Matrix Arithmetic 3 11.2 .2 Matrix Arithmetic Arithmetic It is assumed that the reader is familiar with the fundamental notions of matrix addition, multiplication of a matrix by a scalar, and multiplication of matrices. A special case of matrix multiplication multiplication occurs when the second second matrix is a column i.e., the matrix-vector product Ax. Ax. A very important way to view this product is vector x, i.e., interpret it as a weighted weighted sum (linear combination) of the columns of A. That is, suppose to interpret (linear combination) suppose A = la' ....• a"1 E m JR " with a, Then Ax = Xjal E JRm and x = + ... + Xnan Il ;xn~ ] E jRm. The importance importance of this interpretation interpretation cannot be overemphasized. As a numerical example, take = [96 take A A = [~ 85 74]x ~], x == ! 2 . Then can quickly quickly calculate dot products rows of [~]. Then we we can calculate dot products of of the the rows of A A column x to find Ax Ax = = [50[;~], matrix-vector product product can also be computed with the column 32]' but this matrix-vector computed via v1a 3.[ ~ J+2.[ ~ J+l.[ ~ l For large arrays of numbers, there can be important computer-architecture-related computer-architecture-related advantages to preferring the latter calculation method. mxn nxp multiplication, suppose A e E R jRmxn and and B = [bi,...,b [hI,.'" hpp]] e E R jRnxp with For matrix multiplication, suppose A 1 hi E jRn.. Then the matrix product A AB bi e W B can be thought of as above, applied p times: There is also an alternative, but equivalent, formulation of matrix multiplication that appears frequently in the text and is presented below as a theorem. Again, its importance cannot be overemphasized. It It is deceptively simple and its full understanding is well rewarded. pxn Theorem 1.3. [Uj, .... Theorem 1.3. Let U U = [MI, . . ,, un] un]Ee jRmxn Rmxn with withUiut Ee jRm Rm and andVV == [VI, [v{.•. ,...,, Vn] vn]Ee lRRPxn p jRP. with Vi vt eE R . Then n UV T = LUiVr E jRmxp. i=I If (C D)TT = If matrices C and D are compatible for multiplication, recall that (CD) = DT DT C TT H H H (or (CD} (C D)H =— DH C H).). This gives a dual to the matrix-vector matrix-vector result above. Namely, if if D C mxn jRmxn has C EeR has row row vectors cJ cj Ee jRlxn, E l x ", and and is is premultiplied premultiplied by by aa row row vector yT yTeE jRlxm, Rlxm, then the product can be written as a weighted linear sum of the rows of C as follows: follows: yTC=YICf +"'+Ymc~ EjRlxn. Theorem 1.3 can then also be generalized to its "row reader. Theorem "row dual." The details are left left to the readei 4 4 1.3 1.3 Chapter Review Chapter 1. 1. Introduction Introduction and and Review Inner Inner Products Products and and Orthogonality Orthogonality For IRn, the Euclidean inner inner product For vectors vectors x, yy E e R", the Euclidean product (or inner inner product, for for short) short) of x and is given given by by yy is n T (x, y) := x y = Lx;y;. ;=1 Note that that the inner product product is is aa scalar. Note the inner scalar. If we define complex Euclidean inner product product (or (or inner inner product, product, If x, y Ee <en, C", we define their their complex Euclidean inner for short) short) by for by n (x'Y}c :=xHy = Lx;y;. ;=1 y)c x}c, Note that (x, (x, y) = (y, (y, x) i.e., the order order in in which which xx and yy appear appear in in the complex inner c = c, i.e., product is is important. important. The The more more conventional conventional definition definition of of the the complex inner product product is is product complex inner ((x, x , yy)c )c = yHxx = Eni=1 x;y; xiyi but the text text we with the = yH = L:7=1 but throughout throughout the we prefer prefer the the symmetry symmetry with the real real case. case. Example 1.4. Let [1j]] and and yy == [~]. [1/2]. Then Then Example 1.4. Let xx = = [} (x, Y}c = [ } JH [ ~ ] = [I - j] [ ~ ] = 1 - 2j while while and we see that, indeed, (x, (x, Y}c y)c = = {y, (y, x)c' x)c. and we see that, indeed, Note that that xx TTxx = = 0 0 if if and and only only if if xx = = 00 when when xx eE Rn IRn but but that that this this is is not not true true if ifxx eE Cn. en. Note HH What is true complex case and only if x = 0. illustrate, consider consider What is true in in the the complex case is is that that X x x = 00 if if and only if O. To To illustrate, T H the nonzero vector =0 the nonzero vector xx above. above. Then Then X x TXx = 0 but but X x HXX = = 2.2. n Two nonzero nonzero vectors vectors x, x, y eE IR to be be orthogonal if their their inner product is is Two R are are said said to orthogonal if inner product H zero, i.e., xxTTyy = = 0. if X 0. If xx and zero, i.e., O. Nonzero Nonzero complex complex vectors vectors are are orthogonal orthogonal if x Hyy = = O. and yy are are T T orthogonal and and X x TXx = and yyT = 1,1, then then we we say say that that xx and are orthonormal. orthonormal. A A orthogonal = 11 and yy = and yy are nxn T T nxn matrix A eE IR is an orthogonal matrix matrix if if A AT AAT = I, where where /I is is the the n n x x nn matrix R is an orthogonal AA = = AA = /, nx identity matrix. matrix. The notation /„ In is sometimes identity sometimes used used to denote denote the identity matrix in in IRRnxn " x nxn H H (or en xn). A eE en = I. Clearly (orC" "). Similarly, Similarly, a matrix A C xn is said said to be unitary if A H A = = AA H = an orthogonal orthogonal or or unitary unitary matrix rows and is an matrix has has orthonormal orthonormal rows and orthonormal orthonormal columns. columns. There There is mxn no special name attached attached to to aa nonsquare nonsquare matrix matrix A A e E ]Rrn"n (or € E e ))with no special name R mxn (or Cmxn with orthonormal orthonormal rows columns. rows or or columns. 1.4 1.4 Determinants Determinants It A E IRnnxn xn It is assumed assumed that the reader is familiar with the basic theory of of determinants. determinants. For A eR nxn (or A A 6 E en we use use the the notation det A A for determinant of of A. A. We We list list below below some some of of (or C xn) ) we notation det for the the determinant 1.4. Determinants 1.4. Determinants 5 properties of determinants. Note that this is the more more useful properties is not aa minimal set, i.e., several of one or more of the others. properties are consequences properties are consequences of one or more of the others. 1. If If A A has a zero row or if any two rows of A A are equal, then det A A = = 0.o. = 0. 2. If If A A has has aa zero zero column column or or if if any any two two columns columns of of A A are are equal, equal, then then det det A A = O. 3. Interchanging of A sign of 3. Interchanging two two rows rows of A changes changes only only the the sign of the the determinant. determinant. 4. Interchanging two columns of A changes only the sign of of the determinant. 5. scalar a 5. Multiplying Multiplying aa row row of of A A by by aa scalar ex results results in in aa new new matrix matrix whose whose determinant determinant is is a det A. exdetA. Multiplying a column of A A by a scalar 6. Multiplying scalar ex a results in a new matrix whose determinant determinant is a det is ex det A. A. 7. Multiplying of A scalar and and then then adding adding it it to 7. Multiplying aa row row of A by by aa scalar to another another row row does does not not change change the the determinant. determinant. 8. Multiplying aa column 8. column of of A by a scalar scalar and then adding it to another column column does does not change the the determinant. change determinant. nxn 9. det detAT = det detA = detA A eE C C"X"). AT = A (detA (det AHH = det A if A ). 10. If A is diagonal, diagonal, then det A = =a11a22 alla22 ... 10. If • • • ann, ann, i.e., i.e., det det AA isis the the product product of of its its diagonal diagonal elements. a22 ... 11. 11. If If A is upper triangular, then det det A = = all a11a22 • • • a"n. ann. 12. If triangular, then = a11a22 • • • ann. ann. 12. If A A is is lower lower triangUlar, then det det A A= alla22 ... 13. A is block block diagonal block upper triangular or block lower triangular), with 13. If A diagonal (or (or block A 11, A22, A 22 , ... An" A == square diagonal blocks A11, • • •,, A (of possibly different different sizes), then det A nn (of det A 11 det det A22 A22 ... det Ann. det A11 • • • det Ann. xn 14. If eRIRnnxn ,thendet(AB) = det 5. 14. If A, A, B B E , then det(AB) = det A A det det B. 1 15. If If A Rnxn, then =1det 15. A € E lR~xn, then det(Adet(A- 1)) = de: AA. . nxn xm mxm 16. A eE R lR~xn and D DE IR m detA det(D –- CA– CA-l 1 B). B). 16. If If A and eR ,, then det det [~ [Ac B~] A det(D D] = del Proof: from the LU factorization Proof" This This follows follows easily easily from the block block LU factorization [~ ~J=[ ~ ][ ~ xn mxm 17. If If A and D D eE RM , then then det det [~ [Ac B~] BD – 11C ). 17. A Ee R IRnnxn and lR~xm, det D D det(A det(A -– B DC). D] = det Proof" This follows easily from the block UL factorization Proof: BD- 1 I ][ Chapter 1. 1. Introduction Introduction and and Review Chapter Review 6 6 Remark 1.5. The factorization of of aa matrix into the of aa unit lower triangular Remark 1.5. The factorization matrix A A into the product product of unit lower triangular matrix L L (i.e., lower triangular with all l's 1's on the diagonal) and an an upper triangular matrix V U is is called an an LV LU factorization; factorization; see, see, for example, example, [24]. [24]. Another Another such such factorization factorization is is VL UL where V U is unit upper triangular and L is lower triangular. triangular. The factorizations used above are block analogues of these. Remark [~ BD]. ~ ]. Remark 1.6. The matrix D -— e C A –-I1 BB is called the Schur complement of A in[AC l D – l C is the Schur complement of in [~ [AC B~D ]. Similarly, A -– B BD-Ie of D Din EXERCISES EXERCISES 1. If A eE jRnxn a is a scalar, what is det(aA)? What is det(–A)? det(-A)? Rnxn and or A is orthogonal, what is det A? A? If A is unitary, unitary, what is det A? A? 2. If If A If A 3. Let Letx,y jRn. Show Showthatdet(l-xyT) x, y eE Rn. that det(I – xyT) = 11 – yTx. yTx. 4. Let U1, VI, V2, E jRn xn be orthogonal matrices. Show that the product V U2, ... . . .,,Vk Uk € Rnxn U = = VI U1 V2 U2 ... • • •V Ukk is is an an orthogonal matrix. 5. Let A A E of A, denoted denoted TrA, Tr A, is defined as the sum of its diagonal e jRNxn. R n x n . The trace of aii. elements, Eni=1 au· elements, i.e., i.e., TrA TrA = = L~=I linear function; i.e., if A, B eE JRn xn and a, ft f3 eE R, JR, then (a) Show that the trace is a linear Rnxn Tr(aA + f3B) fiB)= + fiTrB. Tr(aA = aTrA aTrA + f3TrB. (b) Show that Tr(AB) = Tr(BA), AB i= BA. Tr(Afl) = Tr(£A), even though in general AB ^ B A. nxn (c) Let S € E R jRnxn be skew-symmetric, skew-symmetric, i.e., S STT = = -So TrS = 0. O. Then -S. Show that TrS either prove the converse or provide a counterexample. x 6. A matrix A A eE W jRnxn A22 = A. " is said to be idempotent if A 22 / x™ . , • , 2cos<9 0 (a) Show that the matrix A _.. A = = --2!I [T|_ 2cos . 2f) sin 2^ sm 0 J. . sin 20 1 . .d_, ..lor all II _sin. 20 is idempotent for 2sin aII #. o. r 2z 0 2sm2rt # J IS I empotent X (b) Suppose A eE IR" jRn xn"isisidempotent Suppose A idempotentand andAAi=^ I.I. Show Showthat thatAAmust mustbe besingular. singular.