for Scientists & Engineers
Matrix Analysis
for Scientists & Engineers
This page intentionally left blank This page intentionally left blank
Matrix Analysis
for Scientists & Engineers
Alan J. Laub
University of California
Davis, California
slam.
Matrix Analysis
for Scientists & Engineers
Alan J. Laub
University of California
Davis, California
Copyright © 2005 by the Society for Industrial and Applied Mathematics.
1 0 9 8 7 6 5 4 3 2 1
All rights reserved. Printed in the United States of America. No part of this book
may be reproduced, stored, or transmitted in any manner without the written permission
of the publisher. For information, write to the Society for Industrial and Applied
Mathematics, 3600 University City Science Center, Philadelphia, PA 191042688.
MATLAB® is a registered trademark of The MathWorks, Inc. For MATLAB product information,
please contact The MathWorks, Inc., 3 Apple Hill Drive, Natick, MA 017602098 USA,
5086477000, Fax: 5086477101, info@mathworks.com, www.mathworks.com
Mathematica is a registered trademark of Wolfram Research, Inc.
Mathcad is a registered trademark of Mathsoft Engineering & Education, Inc.
Library of Congress CataloginginPublication Data
Laub, Alan J., 1948
Matrix analysis for scientists and engineers / Alan J. Laub.
p. cm.
Includes bibliographical references and index.
ISBN 0898715768 (pbk.)
1. Matrices. 2. Mathematical analysis. I. Title.
QA188138 2005
512.9'434—dc22
2004059962
About the cover: The original artwork featured on the cover was created by freelance
artist Aaron Tallon of Philadelphia, PA. Used by permission.
slam is a registered trademark.
Copyright © 2005 by the Society for Industrial and Applied Mathematics.
10987654321
All rights reserved. Printed in the United States of America. No part of this book
may be reproduced, stored, or transmitted in any manner without the written permission
of the publisher. For information, write to the Society for Industrial and Applied
Mathematics, 3600 University City Science Center, Philadelphia, PA 191042688.
MATLAB® is a registered trademark of The MathWorks, Inc. For MATLAB product information,
please contact The MathWorks, Inc., 3 Apple Hill Drive, Natick, MA 017602098 USA,
5086477000, Fax: 5086477101, info@mathworks.com, wwwmathworks.com
Mathematica is a registered trademark of Wolfram Research, Inc.
Mathcad is a registered trademark of Mathsoft Engineering & Education, Inc.
Library of Congress CataloginginPublication Data
Laub, Alan J., 1948
Matrix analysis for scientists and engineers / Alan J. Laub.
p. cm.
Includes bibliographical references and index.
ISBN 0898715768 (pbk.)
1. Matrices. 2. Mathematical analysis. I. Title.
QA 188.L38 2005
512.9'434dc22
2004059962
About the cover: The original artwork featured on the cover was created by freelance
artist Aaron Tallon of Philadelphia, PA. Used by permission .
•
5.lam... is a registered trademark.
To my wife, Beverley
(who captivated me in the UBC math library
nearly forty years ago)
To my wife, Beverley
(who captivated me in the UBC math library
nearly forty years ago)
This page intentionally left blank This page intentionally left blank
Contents
Preface xi
1 Introduction and Review 1
1.1 Some Notation and Terminology 1
1.2 Matrix Arithmetic 3
1.3 Inner Products and Orthogonality 4
1.4 Determinants 4
2 Vector Spaces 7
2.1 Definitions and Examples 7
2.2 Subspaces 9
2.3 Linear Independence 10
2.4 Sums and Intersections of Subspaces 13
3 Linear Transformations 17
3.1 Definition and Examples 17
3.2 Matrix Representation of Linear Transformations 18
3.3 Composition of Transformations 19
3.4 Structure of Linear Transformations 20
3.5 Four Fundamental Subspaces 22
4 Introduction to the MoorePenrose Pseudoinverse 29
4.1 Definitions and Characterizations 29
4.2 Examples 30
4.3 Properties and Applications 31
5 Introduction to the Singular Value Decomposition 35
5.1 The Fundamental Theorem 35
5.2 Some Basic Properties 38
5.3 Row and Column Compressions 40
6 Linear Equations 43
6.1 Vector Linear Equations 43
6.2 Matrix Linear Equations 44
6.3 A More General Matrix Linear Equation 47
6.4 Some Useful and Interesting Inverses 47
vii
Contents
Preface
1 Introduction and Review
1.1 Some Notation and Terminology
1.2 Matrix Arithmetic . . . . . . . .
1.3 Inner Products and Orthogonality .
1.4 Determinants
2 Vector Spaces
2.1 Definitions and Examples .
2.2 Subspaces.........
2.3 Linear Independence . . .
2.4 Sums and Intersections of Subspaces
3 Linear Transformations
3.1 Definition and Examples . . . . . . . . . . . . .
3.2 Matrix Representation of Linear Transformations
3.3 Composition of Transformations . .
3.4 Structure of Linear Transformations
3.5 Four Fundamental Subspaces . . . .
4 Introduction to the MoorePenrose Pseudoinverse
4.1 Definitions and Characterizations.
4.2 Examples..........
4.3 Properties and Applications . . . .
5 Introduction to the Singular Value Decomposition
5.1 The Fundamental Theorem . . .
5.2 Some Basic Properties .....
5.3 Rowand Column Compressions
6 Linear Equations
6.1 Vector Linear Equations . . . . . . . . .
6.2 Matrix Linear Equations ....... .
6.3 A More General Matrix Linear Equation
6.4 Some Useful and Interesting Inverses.
vii
xi
1
1
3
4
4
7
7
9
10
13
17
17
18
19
20
22
29
29
30
31
35
35
38
40
43
43
44
47
47
viii Contents
7 Projections, Inner Product Spaces, and Norms 51
7.1 Projections 51
7.1.1 The four fundamental orthogonal projections 52
7.2 Inner Product Spaces 54
7.3 Vector Norms 57
7.4 Matrix Norms 59
8 Linear Least Squares Problems 65
8.1 The Linear Least Squares Problem 65
8.2 Geometric Solution 67
8.3 Linear Regression and Other Linear Least Squares Problems 67
8.3.1 Example: Linear regression 67
8.3.2 Other least squares problems 69
8.4 Least Squares and Singular Value Decomposition 70
8.5 Least Squares and QR Factorization 71
9 Eigenvalues and Eigenvectors 75
9.1 Fundamental Definitions and Properties 75
9.2 Jordan Canonical Form 82
9.3 Determination of the JCF 85
9.3.1 Theoretical computation 86
9.3.2 On the +1's in JCF blocks 88
9.4 Geometric Aspects of the JCF 89
9.5 The Matrix Sign Function 91
10 Canonical Forms 95
10.1 Some Basic Canonical Forms 95
10.2 Definite Matrices 99
10.3 Equivalence Transformations and Congruence 102
10.3.1 Block matrices and definiteness 104
10.4 Rational Canonical Form 104
11 Linear Differential and Difference Equations 109
11.1 Differential Equations 109
11.1.1 Properties of the matrix exponential 109
11.1.2 Homogeneous linear differential equations 112
11.1.3 Inhomogeneous linear differential equations 112
11.1.4 Linear matrix differential equations 113
11.1.5 Modal decompositions 114
11.1.6 Computation of the matrix exponential 114
11.2 Difference Equations 118
11.2.1 Homogeneous linear difference equations 118
11.2.2 Inhomogeneous linear difference equations 118
11.2.3 Computation of matrix powers 119
11.3 HigherOrder Equations 120
viii
7 Projections, Inner Product Spaces, and Norms
7.1 Projections ..................... .
7.1.1 The four fundamental orthogonal projections
7.2 Inner Product Spaces
7.3 Vector Norms
7.4 Matrix Norms ....
8 Linear Least Squares Problems
8.1 The Linear Least Squares Problem . . . . . . . . . . . . . .
8.2 Geometric Solution . . . . . . . . . . . . . . . . . . . . . .
8.3 Linear Regression and Other Linear Least Squares Problems
8.3.1 Example: Linear regression ...... .
8.3.2 Other least squares problems ...... .
8.4 Least Squares and Singular Value Decomposition
8.5 Least Squares and QR Factorization . . . . . . .
9 Eigenvalues and Eigenvectors
9.1 Fundamental Definitions and Properties
9.2 Jordan Canonical Form .... .
9.3 Determination of the JCF .... .
9.3.1 Theoretical computation .
9.3.2 On the + l's in JCF blocks
9.4 Geometric Aspects of the JCF
9.5 The Matrix Sign Function.
10 Canonical Forms
10.1 Some Basic Canonical Forms .
10.2 Definite Matrices . . . . . . .
10.3 Equivalence Transformations and Congruence
10.3.1 Block matrices and definiteness
10.4 Rational Canonical Form . . . . . . . . .
11 Linear Differential and Difference Equations
ILl Differential Equations . . . . . . . . . . . . . . . .
11.1.1 Properties ofthe matrix exponential . . . .
11.1.2 Homogeneous linear differential equations
11.1.3 Inhomogeneous linear differential equations
11.1.4 Linear matrix differential equations . .
11.1.5 Modal decompositions . . . . . . . . .
11.1.6 Computation of the matrix exponential
11.2 Difference Equations . . . . . . . . . . . . . .
11.2.1 Homogeneous linear difference equations
11.2.2 Inhomogeneous linear difference equations
11.2.3 Computation of matrix powers .
11.3 HigherOrder Equations. . . . . . . . . . . . . . .
Contents
51
51
52
54
57
59
65
65
67
67
67
69
70
71
75
75
82
85
86
88
89
91
95
95
99
102
104
104
109
109
109
112
112
113
114
114
118
118
118
119
120
Contents ix
12 Generalized Eigenvalue Problems 125
12.1 The Generalized Eigenvalue/Eigenvector Problem 125
12.2 Canonical Forms 127
12.3 Application to the Computation of System Zeros 130
12.4 Symmetric Generalized Eigenvalue Problems 131
12.5 Simultaneous Diagonalization 133
12.5.1 Simultaneous diagonalization via SVD 133
12.6 HigherOrder Eigenvalue Problems 135
12.6.1 Conversion to firstorder form 135
13 Kronecker Products 139
13.1 Definition and Examples 139
13.2 Properties of the Kronecker Product 140
13.3 Application to Sylvester and Lyapunov Equations 144
Bibliography 151
Index 153
Contents
12 Generalized Eigenvalue Problems
12.1 The Generalized EigenvaluelEigenvector Problem
12.2 Canonical Forms ................ .
12.3 Application to the Computation of System Zeros .
12.4 Symmetric Generalized Eigenvalue Problems .
12.5 Simultaneous Diagonalization ........ .
12.5.1 Simultaneous diagonalization via SVD
12.6 HigherOrder Eigenvalue Problems ..
12.6.1 Conversion to firstorder form
13 Kronecker Products
13.1 Definition and Examples ............ .
13.2 Properties of the Kronecker Product ...... .
13.3 Application to Sylvester and Lyapunov Equations
Bibliography
Index
ix
125
125
127
130
131
133
133
135
135
139
139
140
144
151
153
This page intentionally left blank This page intentionally left blank
Preface
This book is intended to be used as a text for beginning graduatelevel (or even seniorlevel)
students in engineering, the sciences, mathematics, computer science, or computational
science who wish to be familar with enough matrix analysis that they are prepared to use its
tools and ideas comfortably in a variety of applications. By matrix analysis I mean linear
algebra and matrix theory together with their intrinsic interaction with and application to
linear dynamical systems (systems of linear differential or difference equations). The text
can be used in a onequarter or onesemester course to provide a compact overview of
much of the important and useful mathematics that, in many cases, students meant to learn
thoroughly as undergraduates, but somehow didn't quite manage to do. Certain topics
that may have been treated cursorily in undergraduate courses are treated in more depth
and more advanced material is introduced. I have tried throughout to emphasize only the
more important and "useful" tools, methods, and mathematical structures. Instructors are
encouraged to supplement the book with specific application examples from their own
particular subject area.
The choice of topics covered in linear algebra and matrix theory is motivated both by
applications and by computational utility and relevance. The concept of matrix factorization
is emphasized throughout to provide a foundation for a later course in numerical linear
algebra. Matrices are stressed more than abstract vector spaces, although Chapters 2 and 3
do cover some geometric (i.e., basisfree or subspace) aspects of many of the fundamental
notions. The books by Meyer [18], Noble and Daniel [20], Ortega [21], and Strang [24]
are excellent companion texts for this book. Upon completion of a course based on this
text, the student is then wellequipped to pursue, either via formal courses or through self
study, followon topics on the computational side (at the level of [7], [11], [23], or [25], for
example) or on the theoretical side (at the level of [12], [13], or [16], for example).
Prerequisites for using this text are quite modest: essentially just an understanding
of calculus and definitely some previous exposure to matrices and linear algebra. Basic
concepts such as determinants, singularity of matrices, eigenvalues and eigenvectors, and
positive definite matrices should have been covered at least once, even though their recollec
tion may occasionally be "hazy." However, requiring such material as prerequisite permits
the early (but "outoforder" by conventional standards) introduction of topics such as pseu
doinverses and the singular value decomposition (SVD). These powerful and versatile tools
can then be exploited to provide a unifying foundation upon which to base subsequent top
ics. Because tools such as the SVD are not generally amenable to "hand computation," this
approach necessarily presupposes the availability of appropriate mathematical software on
a digital computer. For this, I highly recommend MA TL A B® although other software such as
xi
Preface
This book is intended to be used as a text for beginning graduatelevel (or even seniorlevel)
students in engineering, the sciences, mathematics, computer science, or computational
science who wish to be familar with enough matrix analysis that they are prepared to use its
tools and ideas comfortably in a variety of applications. By matrix analysis I mean linear
algebra and matrix theory together with their intrinsic interaction with and application to
linear dynamical systems (systems of linear differential or difference equations). The text
can be used in a onequarter or onesemester course to provide a compact overview of
much of the important and useful mathematics that, in many cases, students meant to learn
thoroughly as undergraduates, but somehow didn't quite manage to do. Certain topics
that may have been treated cursorily in undergraduate courses are treated in more depth
and more advanced material is introduced. I have tried throughout to emphasize only the
more important and "useful" tools, methods, and mathematical structures. Instructors are
encouraged to supplement the book with specific application examples from their own
particular subject area.
The choice of topics covered in linear algebra and matrix theory is motivated both by
applications and by computational utility and relevance. The concept of matrix factorization
is emphasized throughout to provide a foundation for a later course in numerical linear
algebra. Matrices are stressed more than abstract vector spaces, although Chapters 2 and 3
do cover some geometric (i.e., basisfree or subspace) aspects of many of the fundamental
notions. The books by Meyer [18], Noble and Daniel [20], Ortega [21], and Strang [24]
are excellent companion texts for this book. Upon completion of a course based on this
text, the student is then wellequipped to pursue, either via formal courses or through self
study, followon topics on the computational side (at the level of [7], [II], [23], or [25], for
example) or on the theoretical side (at the level of [12], [13], or [16], for example).
Prerequisites for using this text are quite modest: essentially just an understanding
of calculus and definitely some previous exposure to matrices and linear algebra. Basic
concepts such as determinants, singularity of matrices, eigenvalues and eigenvectors, and
positive definite matrices should have been covered at least once, even though their recollec
tion may occasionally be "hazy." However, requiring such material as prerequisite permits
the early (but "outoforder" by conventional standards) introduction of topics such as pseu
doinverses and the singular value decomposition (SVD). These powerful and versatile tools
can then be exploited to provide a unifying foundation upon which to base subsequent top
ics. Because tools such as the SVD are not generally amenable to "hand computation," this
approach necessarily presupposes the availability of appropriate mathematical software on
a digital computer. For this, I highly recommend MAlLAB® although other software such as
xi
xii Preface
Mathematica® or Mathcad® is also excellent. Since this text is not intended for a course in
numerical linear algebra per se, the details of most of the numerical aspects of linear algebra
are deferred to such a course.
The presentation of the material in this book is strongly influenced by computa
tional issues for two principal reasons. First, "reallife" problems seldom yield to simple
closedform 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 the key algorithms of numerical linear algebra, in particular, form the foundation
upon which rests virtually all of modern scientific and engineering computation. A second
motivation for a computational emphasis is that it provides many of the essential tools for
what I call "qualitative 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
concept. But in most engineering or scientific contexts we want to know more than that.
If a set of vectors is linearly independent, how "nearly dependent" are the vectors? If they
are linearly dependent, are there "best" linearly independent subsets? These turn out to
be much more difficult problems and frequently involve researchlevel questions when set
in the context of the finiteprecision, finiterange floatingpoint arithmetic environment of
most modern computing platforms.
Some of the applications of matrix analysis mentioned briefly in this book derive
from the modern statespace approach to dynamical systems. Statespace methods are
now standard in much of modern engineering where, for example, control systems with
large numbers of interacting inputs, outputs, and states often give rise to models of very
high order that must be analyzed, simulated, and evaluated. The "language" in which such
models are conveniently described involves vectors and matrices. It is thus crucial to acquire
a working knowledge of the vocabulary and grammar of this language. The tools of matrix
analysis are also 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. Mastery of the material in this text should enable the student to read and
understand the modern language of matrices used throughout mathematics, science, and
engineering.
While prerequisites for this text are modest, and while most material is developed from
basic ideas in the book, the student does require a certain amount of what is conventionally
referred to as "mathematical maturity." Proofs are given for many theorems. When they are
not given explicitly, they are either obvious or easily found in the literature. This is ideal
material from which to learn a bit about mathematical proofs and the mathematical maturity
and insight gained thereby. It is my firm conviction that such maturity is neither encouraged
nor nurtured by relegating the mathematical aspects of applications (for example, linear
algebra for elementary statespace theory) to an appendix or introducing it "onthefly" when
necessary. Rather, one must lay a firm foundation upon which subsequent applications and
perspectives can be built in a logical, consistent, and coherent fashion.
I have taught this material for many years, many times at UCSB and twice at UC
Davis, and the course has proven to be remarkably successful at enabling students from
disparate backgrounds to acquire a quite acceptable level of mathematical maturity and
rigor for subsequent graduate studies in a variety of disciplines. Indeed, many students who
completed the course, especially the first few times it was offered, remarked afterward that
if only they had had this course before they took linear systems, or signal processing,
xii Preface
Mathematica® or Mathcad® is also excellent. Since this text is not intended for a course in
numerical linear algebra per se, the details of most of the numerical aspects of linear algebra
are deferred to such a course.
The presentation of the material in this book is strongly influenced by computa
tional issues for two principal reasons. First, "reallife" problems seldom yield to simple
closedform 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 the key algorithms of numerical linear algebra, in particular, form the foundation
upon which rests virtually all of modem scientific and engineering computation. A second
motivation for a computational emphasis is that it provides many of the essential tools for
what I call "qualitative 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
concept. But in most engineering or scientific contexts we want to know more than that.
If a set of vectors is linearly independent, how "nearly dependent" are the vectors? If they
are linearly dependent, are there "best" linearly independent subsets? These tum out to
be much more difficult problems and frequently involve researchlevel questions when set
in the context of the finiteprecision, finiterange floatingpoint arithmetic environment of
most modem computing platforms.
Some of the applications of matrix analysis mentioned briefly in this book derive
from the modem statespace approach to dynamical systems. Statespace methods are
now standard in much of modem engineering where, for example, control systems with
large numbers of interacting inputs, outputs, and states often give rise to models of very
high order that must be analyzed, simulated, and evaluated. The "language" in which such
models are conveniently described involves vectors and matrices. It is thus crucial to acquire
a working knowledge of the vocabulary and grammar of this language. The tools of matrix
analysis are also 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. Mastery of the material in this text should enable the student to read and
understand the modem language of matrices used throughout mathematics, science, and
engineering.
While prerequisites for this text are modest, and while most material is developed from
basic ideas in the book, the student does require a certain amount of what is conventionally
referred to as "mathematical maturity." Proofs are given for many theorems. When they are
not given explicitly, they are either obvious or easily found in the literature. This is ideal
material from which to learn a bit about mathematical proofs and the mathematical maturity
and insight gained thereby. It is my firm conviction that such maturity is neither encouraged
nor nurtured by relegating the mathematical aspects of applications (for example, linear
algebra for elementary statespace theory) to an appendix or introducing it "onthef1y" when
necessary. Rather, one must lay a firm foundation upon which subsequent applications and
perspectives can be built in a logical, consistent, and coherent fashion.
I have taught this material for many years, many times at UCSB and twice at UC
Davis, and the course has proven to be remarkably successful at enabling students from
disparate backgrounds to acquire a quite acceptable level of mathematical maturity and
rigor for subsequent graduate studies in a variety of disciplines. Indeed, many students who
completed the course, especially the first few times it was offered, remarked afterward that
if only they had had this course before they took linear systems, or signal processing.
Preface xiii
or estimation theory, etc., they would have been able to concentrate on the new ideas
they wanted to learn, rather than having to spend time making up for deficiencies in their
background in matrices and linear algebra. My fellow instructors, too, realized that by
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, June 2004
Preface XIII
or estimation theory, etc., they would have been able to concentrate on the new ideas
they wanted to learn, rather than having to spend time making up for deficiencies in their
background in matrices and linear algebra. My fellow instructors, too, realized that by
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, June 2004
This page intentionally left blank This page intentionally left blank
Chapter 1
Introduction and Review
1.1 Some Notation and Terminology
We begin with a brief introduction to some standard notation and terminology to be used
throughout the text. This is followed by a review of some basic notions in matrix analysis
and linear algebra.
The following sets appear frequently throughout subsequent chapters:
1. R
n
= the set of ntuples of real numbers represented as column vectors. Thus, x e Rn
means
where xi e R for i e n.
Henceforth, the notation n denotes the set {1, . . . , n}.
Note: Vectors are always column vectors. A row vector is denoted by y
T
, where
y G Rn and the superscript T is the transpose operation. That a vector is always a
column vector rather than a row vector is entirely arbitrary, but this convention makes
it easy to recognize immediately throughout the text that, e.g., X
T
y is a scalar while
xy
T
is an n x n matrix.
2. Cn = the set of ntuples of complex numbers represented as column vectors.
3. R
mxn
= the set of real (or realvalued) m x n matrices.
4. R
mxnr
= the set of real m x n matrices of rank r. Thus, R
nxnn
denotes the set of real
nonsingular n x n matrices.
5. C
mxn
= the set of complex (or complexvalued) m x n matrices.
6. C
mxn
= the set of complex m x n matrices of rank r.
1
Chapter 1
Introduction and Review
1.1 Some Notation and Terminology
We begin with a brief introduction to some standard notation and terminology to be used
throughout the text. This is followed by a review of some basic notions in matrix analysis
and linear algebra.
The following sets appear frequently throughout subsequent chapters:
I. IR
n
= the set of ntuples of real numbers represented as column vectors. Thus, x E IR
n
means
where Xi E IR for i E !!.
Henceforth, the notation!! denotes the set {I, ... , n }.
Note: Vectors are always column vectors. A row vector is denoted by y ~ where
y E IR
n
and the superscript T is the transpose operation. That a vector is always a
column vector rather than a row vector is entirely arbitrary, but this convention makes
it easy to recognize immediately throughout the text that, e.g., x
T
y is a scalar while
xyT is an n x n matrix.
2. en = the set of ntuples of complex numbers represented as column vectors.
3. IR
rn
xn = the set of real (or realvalued) m x n matrices.
4. 1R;n xn = the set of real m x n matrices of rank r. Thus, I R ~ xn denotes the set of real
nonsingular n x n matrices.
5. e
rnxn
= the set of complex (or complexvalued) m x n matrices.
6. e;n xn = the set of complex m x n matrices of rank r.
Chapter 1. Introduction and Review
Each of the above also has a "block" analogue obtained by replacing scalar components in
the respective definitions by block submatrices. For example, if A e R
nxn
, B e R
mx n
, and
C e R
mxm
, then the (m+ n) x (m+ n) matrix [ A0 Bc ] is block upper triangular.
The transpose of a matrix A is denoted by A
T
and is the matrix whose (i, j)th entry
is the (7, Oth entry of A, that is, (A
7
),, = a,,. Note that if A e R
mx
", then A
7
" e E"
xm
.
If A e C
mx
", then its Hermitian transpose (or conjugate transpose) is denoted by A
H
(or
sometimes A*) and its (i, j)\h entry is (A
H
),
7
= («77), where the bar indicates complex
conjugation; i.e., if z = a + jf$ (j = i = v^T), then z = a — jfi. A matrix A is symmetric
if A = A
T
and Hermitian if A = A
H
. We henceforth adopt the convention that, unless
otherwise noted, an equation like A = A
T
implies that A is realvalued while a statement
like A = A
H
implies that A is complexvalued.
Remark 1.1. While \/—\ is most commonly denoted by i in mathematics texts, j is
the more common notation in electrical engineering and system theory. There is some
advantage to being conversant with both notations. The notation j is used throughout the
text but reminders are placed at strategic locations.
Example 1.2.
Transposes of block matrices can be defined in an obvious way. For example, it is
easy to see that if A,, are appropriately dimensioned subblocks, then
is symmetric (and Hermitian).
is complexvalued symmetric but not Hermitian.
is Hermitian (but not symmetric).
2
We now classify some of the more familiar "shaped" matrices. A matrix A e
(or A eC"
x
")i s
• diagonal if a,
7
= 0 for i ^ j.
• upper triangular if a,
;
= 0 for i > j.
• lower triangular if a,
7
= 0 for / < j.
• tridiagonal if a
(y
= 0 for z — j\ > 1.
• pentadiagonal if a
i;
= 0 for / — j\ > 2.
• upper Hessenberg if a
f
j = 0 for i — j > 1.
• lower Hessenberg if a,
;
= 0 for j — i > 1.
2 Chapter 1. Introduction and Review
We now classify some of the more familiar "shaped" matrices. A matrix A E IR
n
xn
(or A E e
nxn
) is
• diagonal if aij = 0 for i i= }.
• upper triangular if aij = 0 for i > }.
• lower triangular if aij = 0 for i < }.
• tridiagonal if aij = 0 for Ii  JI > 1.
• pentadiagonal if aij = 0 for Ii  J I > 2.
• upper Hessenberg if aij = 0 for i  j > 1.
• lower Hessenberg if aij = 0 for }  i > 1.
Each of the above also has a "block" analogue obtained by replacing scalar components in
the respective definitions by block submatrices. For example, if A E IR
nxn
, B E IR
nxm
, and
C E jRmxm, then the (m + n) x (m + n) matrix [ ~ ~ ] is block upper triangular.
The transpose of a matrix A is denoted by AT and is the matrix whose (i, j)th entry
is the (j, i)th entry of A, that is, (AT)ij = aji. Note that if A E jRmxn, then AT E jRnxm.
If A E em xn, then its Hermitian transpose (or conjugate transpose) is denoted by A H (or
sometimes A*) and its (i, j)th entry is (AH)ij = (aji), where the bar indicates complex
conjugation; i.e., if z = IX + jfJ (j = i = R), then z = IX  jfJ. A matrix A is symmetric
if A = A T and Hermitian if A = A H. We henceforth adopt the convention that, unless
otherwise noted, an equation like A = A T implies that A is realvalued while a statement
like A = AH implies that A is complexvalued.
Remark 1.1. While R is most commonly denoted by i in mathematics texts, } is
the more common notation in electrical engineering and system theory. There is some
advantage to being conversant with both notations. The notation j is used throughout the
text but reminders are placed at strategic locations.
Example 1.2.
1. A = [
; ~ ] is symmetric (and Hermitian).
2. A = [
5
7+}
7 + j ]
2 is complexvalued symmetric but not Hermitian.
[
5 7+} ]
3 A  2 is Hermitian (but not symmetric).
·  7  j
Transposes of block matrices can be defined in an obvious way. For example, it is
easy to see that if Aij are appropriately dimensioned subblocks, then
r = [
1.2. Matrix Arithmetic
1.2 Matrix 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 occurs when the second matrix is a column
vector x, i.e., the matrixvector product Ax. A very important way to view this product is
to interpret it as a weighted sum (linear combination) of the columns of A. That is, suppose
The importance of this interpretation cannot be overemphasized. As a numerical example,
take A = [96 85 74]x = 2 . Then we can quickly calculate dot products of the rows of A
with the column x to find Ax =[50 32]' but this matrixvector product can also be computed
v1a
For large arrays of numbers, there can be important computerarchitecturerelated advan
tages to preferring the latter calculation method.
For matrix multiplication, suppose A e R
mxn
and B = [bi,...,b
p
] e R
nxp
with
bi e W
1
. Then the matrix product A 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 is deceptively simple and its full understanding is well rewarded.
Theorem 1.3. Let U = [M I , . . . , u
n
] e R
mxn
with u
t
e R
m
and V = [v
{
,..., v
n
] e R
pxn
with v
t
e R
p
. Then
If matrices C and D are compatible for multiplication, recall that (CD)
T
= D
T
C
T
(or (CD}
H
— D
H
C
H
). This gives a dual to the matrixvector result above. Namely, if
C eR
mxn
has row vectors cj e E
lx
", and is premultiplied by a row vector y
T
e R
l xm
,
then the product can be written as a weighted linear sum of the rows of C as follows:
3
Theorem 1.3 can then also be generalized to its "row dual." The details are left to the readei
Then
1.2. Matrix Arithmetic 3
1 .2 Matrix 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 occurs when the second matrix is a column
vector x, i.e., the matrixvector product Ax. A very important way to view this product is
to interpret it as a weighted sum (linear combination) of the columns of A. That is, suppose
I ]
A = la' ....• a"1 E JR
m
" with a, E JRm and x = l
Then
Ax = Xjal + ... + Xnan E jRm.
The importance of this interpretation cannot be overemphasized. As a numerical example,
take A = ! x = Then we can quickly calculate dot products of the rows of A
with the column x to find Ax = but this matrixvector product can also be computed
via
3.[ J+2.[ J+l.[ l
For large arrays of numbers, there can be important computerarchitecturerelated advan
tages to preferring the latter calculation method.
For matrix multiplication, suppose A E jRmxn and B = [hI,.'" h
p
] E jRnxp with
hi E jRn. Then the matrix product AB 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 is deceptively simple and its full understanding is well rewarded.
Theorem 1.3. Let U = [Uj, ... , un] E jRmxn with Ui E jRm and V = [VI, .•. , Vn] E lR
Pxn
with Vi E jRP. Then
n
UV
T
= LUiVr E jRmxp.
i=I
If matrices C and D are compatible for multiplication, recall that (C D)T = DT C
T
(or (C D)H = DH C
H
). This gives a dual to the matrixvector result above. Namely, if
C E jRmxn has row vectors cJ E jRlxn, and is premultiplied by a row vector yT E jRlxm,
then the product can be written as a weighted linear sum of the rows of C as follows:
yTC=YICf EjRlxn.
Theorem 1.3 can then also be generalized to its "row dual." The details are left to the reader.
Chapter 1. Introduction and Review
1.3 Inner Products and Orthogonality
For vectors x, y e R", the Euclidean inner product (or inner product, for short) of x and
y is given by
Note that the inner product is a scalar.
If x, y e C", we define their complex Euclidean inner product (or inner product,
for short) by
and we see that, indeed, (x, y)
c
= (y, x)
c
.
Note that x
T
x = 0 if and only if x = 0 when x e Rn but that this is not true if x e Cn.
What is true in the complex case is that X
H
x = 0 if and only if x = 0. To illustrate, consider
the nonzero vector x above. Then X
T
X = 0 but X
H
X = 2.
Two nonzero vectors x, y e R are said to be orthogonal if their inner product is
zero, i.e., x
T
y = 0. Nonzero complex vectors are orthogonal if X
H
y = 0. If x and y are
orthogonal and X
T
X = 1 and y
T
y = 1, then we say that x and y are orthonormal. A
matrix A e R
nxn
is an orthogonal matrix if A
T
A = AA
T
= /, where / is the n x n
identity matrix. The notation /„ is sometimes used to denote the identity matrix in R
nx
"
(orC"
x
"). Similarly, a matrix A e C
nxn
is said to be unitary if A
H
A = AA
H
= I. Clearly
an orthogonal or unitary matrix has orthonormal rows and orthonormal columns. There is
no special name attached to a nonsquare matrix A e R
mxn
(or € C
mxn
) with orthonormal
rows or columns.
1.4 Determinants
It is assumed that the reader is familiar with the basic theory of determinants. For A e R
nxn
(or A 6 C
nxn
) we use the notation det A for the determinant of A. We list below some of
Note that (x, y)
c
= (y, x)
c
, i.e., the order in which x and y appear in the complex inner
product is important. The more conventional definition of the complex inner product is
( x , y )
c
= y
H
x = Eni=1 xiyi but throughout the text we prefer the symmetry with the real
case.
Example 1.4. Let x = [ 1j ] and y = [ 1/ 2 ]. Then
while
44 Chapter 1. Introduction and Review
1.3 Inner Products and Orthogonality
For vectors x, y E IRn, the Euclidean inner product (or inner product, for short) of x and
y is given by
n
(x, y) := x
T
y = Lx;y;.
;=1
Note that the inner product is a scalar.
If x, y E <en, we define their complex Euclidean inner product (or inner product,
for short) by
n
(x'Y}c :=xHy = Lx;y;.
;=1
Note that (x, y)c = (y, x}c, i.e., the order in which x and y appear in the complex inner
product is important. The more conventional definition of the complex inner product is
(x, y)c = yH x = L:7=1 x;y; but throughout the text we prefer the symmetry with the real
case.
Example 1.4. Let x = [} ] and y = [ ~ ] . Then
(x, Y}c = [ } JH [ ~ ] = [I  j] [ ~ ] = 1  2j
while
and we see that, indeed, (x, Y}c = {y, x)c'
Note that x
T
x = 0 if and only if x = 0 when x E IR
n
but that this is not true if x E en.
What is true in the complex case is that x
H
x = 0 if and only if x = O. To illustrate, consider
the nonzero vector x above. Then x
T
x = 0 but x
H
X = 2.
Two nonzero vectors x, y E IR
n
are said to be orthogonal if their inner product is
zero, i.e., x
T
y = O. Nonzero complex vectors are orthogonal if x
H
y = O. If x and y are
orthogonal and x
T
x = 1 and yT y = 1, then we say that x and y are orthonormal. A
matrix A E IR
nxn
is an orthogonal matrix if AT A = AAT = I, where I is the n x n
identity matrix. The notation In is sometimes used to denote the identity matrix in IR
nxn
(or en xn). Similarly, a matrix A E en xn is said to be unitary if A H A = AA H = I. Clearly
an orthogonal or unitary matrix has orthonormal rows and orthonormal columns. There is
no special name attached to a nonsquare matrix A E ]Rrn"n (or E e
mxn
) with orthonormal
rows or columns.
1.4 Determinants
It is assumed that the reader is familiar with the basic theory of determinants. For A E IR
n
xn
(or A E en xn) we use the notation det A for the determinant of A. We list below some of
1.4. Determinants
the more useful properties of determinants. Note that this is not a minimal set, i.e., several
properties are consequences of one or more of the others.
1. If A has a zero row or if any two rows of A are equal, then det A = 0.
2. If A has a zero column or if any two columns of A are equal, then det A = 0.
3. Interchanging two rows of A changes only the sign of the determinant.
4. Interchanging two columns of A changes only the sign of the determinant.
5. Multiplying a row of A by a scalar a results in a new matrix whose determinant is
a det A.
6. Multiplying a column of A by a scalar a results in a new matrix whose determinant
is a det A.
7. Multiplying a row of A by a scalar and then adding it to another row does not change
the determinant.
8. Multiplying a column of A by a scalar and then adding it to another column does not
change the determinant.
9. det A
T
= det A (det A
H
= det A if A e C
nxn
).
10. If A is diagonal, then det A = a11a22 • • • a
nn
, i.e., det A is the product of its diagonal
elements.
11. If A is upper triangular, then det A = a11a22 • • • a
nn
.
12. If A is lower triangular, then det A = a11a22 • • • a
nn
.
13. If A is block diagonal (or block upper triangular or block lower triangular), with
square diagonal blocks A11, A22, • • •, A
nn
(of possibly different sizes), then det A =
det A11 det A22 • • • det A
nn
.
14. If A, B eR
nxn
,thendet(AB) = det A det 5.
15. If A € R
nxn
, then det(A
1
) = 1det A.
16. If A e R
nxn
and D e R
mxm
, then det [Ac
B
D
] = del A det ( D – CA–
l
B).
Proof: This follows easily from the block LU factorization
17. If A eR
nxn
and D e RM
mxm
, then det [Ac
B
D
] = det D det(A – BD–
1
C) .
Proof: This follows easily from the block UL factorization
5 1.4. Determinants 5
the more useful properties of determinants. Note that this is not a minimal set, i.e., several
properties are consequences of one or more of the others.
1. If A has a zero row or if any two rows of A are equal, then det A = o.
2. If A has a zero column or if any two columns of A are equal, then det A = O.
3. Interchanging two rows of A changes only the sign of the determinant.
4. Interchanging two columns of A changes only the sign of the determinant.
5. Multiplying a row of A by a scalar ex results in a new matrix whose determinant is
exdetA.
6. Multiplying a column of A by a scalar ex results in a new matrix whose determinant
is ex det A.
7. Multiplying a row of A by a scalar and then adding it to another row does not change
the determinant.
8. Multiplying a column of A by a scalar and then adding it to another column does not
change the determinant.
9. detAT = detA (detA
H
= detA if A E C"X").
10. If A is diagonal, then det A = alla22 ... ann, i.e., det A is the product of its diagonal
elements.
11. If A is upper triangular, then det A = all a22 ... a"n.
12. If A is lower triangUlar, then det A = alla22 ... ann.
13. If A is block diagonal (or block upper triangular or block lower triangular), with
square diagonal blocks A 11, A
22
, ... , An" (of possibly different sizes), then det A =
det A 11 det A22 ... det Ann.
14. If A, B E IR
nxn
, then det(AB) = det A det B.
15. If A E then det(A
1
) = de: A .
16. If A E and DE IR
mxm
, then det = detA det(D  CA
1
B).
Proof" This follows easily from the block LU factorization
] [
17. If A E IR
nxn
and D E then det = det D det(A  B D
1
C).
Proof" This follows easily from the block UL factorization
BD
1
I
] [
Chapter 1. Introduction and Review
Remark 1.5. The factorization of a matrix A into the product of a unit lower triangular
matrix L (i.e., lower triangular with all 1's on the diagonal) and an upper triangular matrix
U is called an LU factorization; see, for example, [24]. Another such factorization is UL
where U is unit upper triangular and L is lower triangular. The factorizations used above
are block analogues of these.
Remark 1.6. The matrix D — CA–
1
B is called the Schur complement of A in [AC BD].
Similarly, A – BD–
l
C is the Schur complement of D in [AC
B
D
].
EXERCISES
1. If A e R
nxn
and or is a scalar, what is det(aA)? What is det(–A)?
2. If A is orthogonal, what is det A? If A is unitary, what is det A?
3. Let x, y e Rn. Show that det(I – xy
T
) = 1 – y
T
x.
4. Let U1, U
2
, . . ., Uk € R
nxn
be orthogonal matrices. Show that the product U =
U1 U2 • • • Uk is an orthogonal matrix.
5. Let A e R
n x n
. The trace of A, denoted TrA, is defined as the sum of its diagonal
elements, i.e., TrA = Eni=1
aii.
(a) Show that the trace is a linear function; i.e., if A, B e R
nxn
and a, ft e R, then
Tr(aA + fiB)= aTrA + fiTrB.
(b) Show that Tr(Afl) = Tr(£A), even though in general AB ^ B A.
(c) Let S € R
nxn
be skewsymmetric, i.e., S
T
= S. Show that TrS = 0. Then
either prove the converse or provide a counterexample.
6. A matrix A e W
x
" is said to be idempotent if A
2
= A.
/ x ™ . , • , ! T 2cos
2
<9 sin 20 1 . . _ ,
(a) Show that the matrix A =  . _ .. _ .
2rt
is idempotent for all #.
2 _ sin 2^ 2sm
z
# J
r
(b) Suppose A e IR"
X
" is idempotent and A ^ I. Show that A must be singular.
66 Chapter 1. Introduction and Review
Remark 1.5. The factorization of a matrix A into the product of a unit lower triangular
matrix L (i.e., lower triangular with all l's on the diagonal) and an upper triangular matrix
V is called an LV factorization; see, for example, [24]. Another such factorization is VL
where V is unit upper triangular and L is lower triangular. The factorizations used above
are block analogues of these.
Remark 1.6. The matrix D  e A I B is called the Schur complement of A in [ ~ ~ ].
Similarly, A  BDIe is the Schur complement of Din [ ~ ~ l
EXERCISES
1. If A E jRnxn and a is a scalar, what is det(aA)? What is det(A)?
2. If A is orthogonal, what is det A? If A is unitary, what is det A?
3. Letx,y E jRn. Showthatdet(lxyT) = 1 yTx.
4. Let VI, V2, ... ,Vk E jRn xn be orthogonal matrices. Show that the product V =
VI V2 ... V
k
is an orthogonal matrix.
5. Let A E jRNxn. The trace of A, denoted Tr A, is defined as the sum of its diagonal
elements, i.e., TrA = L ~ = I au·
(a) Show that the trace is a linear function; i.e., if A, B E JRn xn and a, f3 E JR, then
Tr(aA + f3B) = aTrA + f3TrB.
(b) Show that Tr(AB) = Tr(BA), even though in general AB i= BA.
(c) Let S E jRnxn be skewsymmetric, i.e., ST = So Show that TrS = O. Then
either prove the converse or provide a counterexample.
6. A matrix A E jRnxn is said to be idempotent if A2 = A.
I [ 2cos
2
0
(a) Show that the matrix A =  . 2f)
2 sm 0
sin 20 J. . d .. II II
2sin
2
0 IS I empotent lor a o.
(b) Suppose A E jRn xn is idempotent and A i= I. Show that A must be singular.
Chapter 2
Vector Spaces
In this chapter we give a brief review of some of the basic concepts of vector spaces. The
emphasis is on finitedimensional vector spaces, including spaces formed by special classes
of matrices, but some infinitedimensional examples are also cited. An excellent reference
for this and the next chapter is [10], where some of the proofs that are not given here may
be found.
2.1 Definitions and Examples
Definition 2.1. A field is a set F together with two operations +, • : F x F — > F such that
Axioms (A1)(A3) state that (F, +) is a group and an abelian group if (A4) also holds.
Axioms (M1)(M4) state that (F \ {0}, •) is an abelian group.
Generally speaking, when no confusion can arise, the multiplication operator "•" is
not written explicitly.
7
(Al) a + (P + y ) = (a + p ) + y f o r all a, f t, y € F.
(A2) there exists an element 0 e F such that a + 0 = a. for all a e F.
(A3 ) for all a e F, there exists an element (—a) e F such that a + (— a) = 0.
(A4 ) a + p = ft + afar all a, ft e F.
(M l) a  ( p  y ) = ( a  p )  y f o r al l a, p, y e F.
(M 2) there exists an element 1 e F such that a • I = a for all a e F.
(M 3 ) for all a e ¥, a ^0, there exists an element a"
1
€ F such that a • a~
l
= 1.
(M 4 ) a • p = P • a for all a, p e F.
(D) a  ( p + y)=ci p+a y f or alia, p,ye¥.
Chapter 2
Vector Spaces
In this chapter we give a brief review of some of the basic concepts of vector spaces. The
emphasis is on finitedimensional vector spaces, including spaces formed by special classes
of matrices, but some infinitedimensional examples are also cited. An excellent reference
for this and the next chapter is [10], where some of the proofs that are not given here may
be found.
2.1 Definitions and Examples
Definition 2.1. A field is a set IF together with two operations +, . : IF x IF ~ IF such that
(Al) a + (,8 + y) = (a +,8) + y for all a,,8, y Elf.
(A2) there exists an element 0 E IF such that a + 0 = a for all a E IF.
(A3) for all a E IF, there exists an element (a) E IF such that a + (a) = O.
(A4) a + ,8 = ,8 + a for all a, ,8 Elf.
(Ml) a· (,8, y) = (a·,8)· y for all a,,8, y Elf.
(M2) there exists an element I E IF such that a . I = a for all a E IF.
(M3) for all a E IF, a f. 0, there exists an element aI E IF such that a . aI = 1.
(M4) a·,8 =,8 . afar all a, ,8 E IF.
(D) a· (,8 + y) = a·,8 +a· y for all a, ,8, y Elf.
Axioms (Al)(A3) state that (IF, +) is a group and an abelian group if (A4) also holds.
Axioms (MI)(M4) state that (IF \ to), .) is an abelian group.
Generally speaking, when no confusion can arise, the multiplication operator "." is
not written explicitly.
7
Chapter 2. Vector Spaces
Example 2.2.
1. R with ordinary addition and multiplication is a field.
2. C with ordinary complex addition and multiplication is a field.
3. Raf. r] = the field of rational functions in the indeterminate x
8
where Z+ = {0,1,2, . . . }, is a field.
4. RMr
mxn
= { m x n matrices of rank r with real coefficients) is clearly not a field since,
for example, (Ml) does not hold unless m = n. Moreover, R"
x
" is not a field either
since (M4) does not hold in general (although the other 8 axioms hold).
Definition 2.3. A vector space over a field F is a set V together with two operations
+ :V x V ^V and : F xV »• V such that
A vector space is denoted by (V, F) or, when there is no possibility of confusion as to the
underlying fie Id, simply by V.
Remark 2.4. Note that + and • in Definition 2.3 are different from the + and • in Definition
2.1 in the sense of operating on different objects in different sets. In practice, this causes
no confusion and the • operator is usually not even written explicitly.
Example 2.5.
1. (R", R) with addition defined by
and scalar multiplication defined by
is a vector space. Similar definitions hold for (C", C).
(VI) (V, +) is an abelian group.
(V2) ( a  p )  v = a  ( P ' V ) f o r all a, p e F and for all v e V.
(V3) (a + ft) • v = a • v + p • v for all a, p € F and for all v e V.
(V4) a(v + w)=av + a w for all a e F and for all v, w e V.
(V5) 1 • v = v for all v e V (1 e F).
8 Chapter 2. Vector Spaces
Example 2.2.
I. IR with ordinary addition and multiplication is a field.
2. e with ordinary complex addition and multiplication is a field.
3. Ra[x] = the field of rational functions in the indeterminate x
= {a
o
+ atX + ... + apxP +}
:aj,f3i EIR ;P,qEZ ,
f30 + f3t
X
+ ... + f3qX
q
where Z+ = {O,l,2, ... }, is a field.
4. I R ~ xn = { m x n matrices of rank r with real coefficients} is clearly not a field since,
for example, (MI) does not hold unless m = n. Moreover, l R ~ x n is not a field either
since (M4) does not hold in general (although the other 8 axioms hold).
Definition 2.3. A vector space over a field IF is a set V together with two operations
+ : V x V + V and· : IF x V + V such that
(VI) (V, +) is an abelian group.
(V2) (a· f3) . v = a . (f3 . v) for all a, f3 E IF andfor all v E V.
(V3) (a + f3). v = a· v + f3. v for all a, f3 Elf andforall v E V.
(V4) a· (v + w) = a . v + a . w for all a ElF andfor all v, w E V.
(V5) I· v = v for all v E V (1 Elf).
A vector space is denoted by (V, IF) or, when there is no possibility of confusion as to the
underlying field, simply by V.
Remark 2.4. Note that + and· in Definition 2.3 are different from the + and . in Definition
2.1 in the sense of operating on different objects in different sets. In practice, this causes
no confusion and the· operator is usually not even written explicitly.
Example 2.5.
I. (IRn, IR) with addition defined by
and scalar multiplication defined by
is a vector space. Similar definitions hold for (en, e).
2.2. Subspaces
3. Let (V, F) be an arbitrary vector space and V be an arbitrary set. Let O (X > , V) be the
set of functions / mapping D to V. Then O (D, V) is a vector space with addition
defined by
2.2 Subspaces
Definition 2.6. Let (V, F) be a vector space and let W c V, W = 0. Then (W, F) is a
subspace of (V, F) i f and only i f (W, F) is i tself a vector space or, equi valently, i f and only
i f ( a w 1 + ß W 2 ) e W for all a, ß e ¥ and for all w 1 , w
2
e W.
Remark 2.7. The latter characterization of a subspace is often the easiest way to check
or prove that something is indeed a subspace (or vector space); i.e., verify that the set in
question is closed under addition and scalar multiplication. Note, too, that since 0 e F, this
implies that the zero vector must be in any subspace.
Notation: When the underlying field is understood, we write W c V, and the symbol c,
when used with vector spaces, is henceforth understood to mean "is a subspace of." The
less restrictive meaning "is a subset of" is specifically flagged as such.
9
2. (E
mxn
, E) is a vector space with addition defined by
and scalar multiplication defined by
and scalar multiplication defined by
Special Cases:
(a) V = [to, t \ ] , (V, F) = (IR", E), and the functions are piecewise continuous
=: (PC[f
0
, t\ ] )
n
or continuous =: (C[?
0
, h] )
n
.
4. Let A € R"
x
". Then (x(t) : x ( t ) = Ax(t}} is a vector space (of dimension n) .
2.2. Subspaces 9
2.
(JRmxn, JR) is a vector space with addition defined by
[ ." + P"
al2 + fJI2 aln + fJln
l
a21 + fJ2I a22 + fJ22 a2n + fJ2n
A+B= .
amI + fJml am2 + fJm2 amn + fJmn
and scalar multiplication defined by
[ ya"
y
a
l2
ya," l
y
a
21 y
a
22 ya2n
yA = . . .
yaml ya
m
2
ya
mn
3. Let (V, IF) be an arbitrary vector space and '0 be an arbitrary set. Let cf>('O, V) be the
set of functions f mapping '0 to V. Then cf>('O, V) is a vector space with addition
defined by
(f + g)(d) = fed) + g(d) for all d E '0 and for all f, g E cf>
and scalar multiplication defined by
(af)(d) = af(d) for all a E IF, for all d ED, and for all f E cf>.
Special Cases:
(a) '0 = [to, td, (V, IF) = (JR
n
, JR), and the functions are piecewise continuous
=: (PC[to, td)n or continuous =: (C[to, td)n.
(b) '0 = [to, +00), (V, IF) = (JRn, JR), etc.
4. Let A E JR(nxn. Then {x(t) : x(t) = Ax(t)} is a vector space (of dimension n).
2.2 Subspaces
Definition 2.6. Let (V, IF) be a vector space and let W ~ V, W f= 0. Then (W, IF) is a
subspace of (V, IF) if and only if (W, IF) is itself a vector space or, equivalently, if and only
if(awl + fJw2) E W foral! a, fJ E IF andforall WI, W2 E W.
Remark 2.7. The latter characterization of a subspace is often the easiest way to check
or prove that something is indeed a subspace (or vector space); i.e., verify that the set in
question is closed under addition and scalar multiplication. Note, too, that since 0 E IF, this
implies that the zero vector must be in any subspace.
Notation: When the underlying field is understood, we write W ~ V, and the symbol ~ ,
when used with vector spaces, is henceforth understood to mean "is a subspace of." The
less restrictive meaning "is a subset of' is specifically flagged as such.
Then W
a
,ß is a subspace of V if and only if ß = 0. As an interesting exercise, sketch
W2,1, W2,o, W1/2,1, and W1/2,
0
. Note, too, that the vertical line through the origin (i.e.,
a = oo) is also a subspace.
All lines through the origin are subspaces. Shifted subspaces W
a
,ß with ß = 0 are
called linear varieties.
Henceforth, we drop the explicit dependence of a vector space on an underlying field.
Thus, V usually denotes a vector space with the underlying field generally being R unless
explicitly stated otherwise.
Definition 2.9. If 12, and S are vector spaces (or subspaces), then R = S if and only if
R C S and S C R.
Note: To prove two vector spaces are equal, one usually proves the two inclusions separately:
An arbitrary r e R is shown to be an element of S and then an arbitrary 5 € S is shown to
be an element of R.
2.3 Linear Independence
Let X = { v1 , v2, • • •} be a nonempty collection of vectors u, in some vector space V.
Definition 2.10. X is a linearly dependent set of vectors if and only if there exist k distinct
elements v1, . . . , vk e X and scalars a1, . . . , ak not all zero such that
10 Chapter 2. Vector Spaces
Example 2.8.
1. Consider (V, F) = (R"
X
",R) and let W = [A e R"
x
" : A is symmetric}. Then
We V.
Proof: Suppose A\, A
2
are symmetric. Then it is easily shown that ctA\ + fiAi is
symmetric for all a, ft e R.
2. Let W = { A € R"
x
" : A is orthogonal}. Then W is /wf a subspace of R"
x
".
3. Consider (V, F) = (R
2
, R) and for each v € R
2
of the form v = [v1v2 ] identify v1 with
the jccoordinate in the plane and u
2
with the ycoordinate. For a, ß e R, define
X is a linearly independent set of vectors if and only if for any collection of k distinct
elements v1, . . . ,Vk of X and for any scalars a1, . . . , ak,
10 Chapter 2. Vector Spaces
Example 2.S.
1. Consider (V,lF) = (JR.nxn,JR.) and let W = {A E JR.nxn : A is symmetric}. Then
Proof' Suppose AI, A2 are symmetric. Then it is easily shown that aAI + f3A2 is
symmetric for all a, f3 E R
2. Let W = {A E ]Rnxn : A is orthogonal}. Then W is not a subspace of JR.nxn.
3. Consider (V, IF) = (]R2, JR.) and for each v E ]R2 of the form v = ] identify VI with
the xcoordinate in the plane and V2 with the ycoordinate. For a, f3 E R define
W",/l = {V : v = [ ac f3 ] ; c E JR.} .
Then W",/l is a subspace of V if and only if f3 = O. As an interesting exercise, sketch
W2.I, W2,O, Wi,I' and Wi,o, Note, too, that the vertical line through the origin (i.e.,
a = 00) is also a subspace.
All lines through the origin are subspaces. Shifted subspaces W",/l with f3 =1= 0 are
called linear varieties.
Henceforth, we drop the explicit dependence of a vector space on an underlying field.
Thus, V usually denotes a vector space with the underlying field generally being JR. unless
explicitly stated otherwise.
Definition 2.9. ffR and S are vector spaces (or subspaces), then R = S if and only if
R R.
Note: To prove two vector spaces are equal, one usually proves the two inclusions separately:
An arbitrary r E R is shown to be an element of S and then an arbitrary s E S is shown to
be an element of R.
2.3 Linear Independence
Let X = {VI, V2, •.• } be a nonempty collection of vectors Vi in some vector space V.
Definition 2.10. X is a linearly dependent set of vectors if and only if there exist k distinct
elements VI, ... , Vk E X and scalars aI, ..• , (Xk not all zero such that
X is a linearly independent set of vectors if and only if for any collection of k distinct
elements VI, ... , Vk of X and for any scalars aI, ••• , ak,
al VI + ... + (XkVk = 0 implies al = 0, ... , ak = O.
2.3. Linear Independence 11
(since 2v\ — v
2
+ v3 = 0).
2. Let A e R
xn
and 5 e R"
xm
. Then consider the rows of e
tA
B as vectors in C
m
[t
0
, t1]
(recall that e
fA
denotes the matrix exponential, which is discussed in more detail in
Chapter 11). Independence of these vectors turns out to be equivalent to a concept
called controllability, to be studied further in what follows.
Let v
f
e R", i e k, and consider the matrix V = [ v1 , ... ,Vk] e R
nxk
. The linear
dependence of this set of vectors is equivalent to the existence of a nonzero vector a e R
k
such that Va = 0. An equivalent condition for linear dependence is that the k x k matrix
V
T
V is singular. If the set of vectors is independent, and there exists a e R* such that
Va = 0, then a = 0. An equivalent condition for linear independence is that the matrix
V
T
V is nonsingular.
Definition 2.12. Let X = [ v1 , v2, . . . } be a collection of vectors vi. e V. Then the span of
X is defined as
Example 2.13. Let V = R
n
and define
Then Sp{e1, e
2
, ...,e
n
} = Rn.
Definition 2.14. A set of vectors X is a basis for V if and only ij
1. X is a linearly independent set (of basis vectors), and
2. Sp(X) = V.
Example 2.11.
is a linearly independent set. Why?
s a linearly dependent set However,
1. LetV = R
3
. Then
where N = {1, 2, ...}.
2.3. Linear Independence 11
Example 2.11.
I. 1£t V = Then {[ H i Hi] } i" independent.. Why?
Howe,."I [ i 1 [ i 1 [ l ] } is a Iin=ly
(since 2vI  V2 + V3 = 0).
2. Let A E ]Rnxn and B E ]Rnxm. Then consider the rows of etA B as vectors in em [to, tIl
(recall that etA denotes the matrix exponential, which is discussed in more detail in
Chapter 11). Independence of these vectors turns out to be equivalent to a concept
called controllability, to be studied further in what follows.
Let Vi E ]Rn, i E If, and consider the matrix V = [VI, ... , Vk] E ]Rnxk. The linear
dependence of this set of vectors is equivalent to the existence of a nonzero vector a E ]Rk
such that Va = O. An equivalent condition for linear dependence is that the k x k matrix
VT V is singular. If the set of vectors is independent, and there exists a E ]Rk such that
Va = 0, then a = O. An equivalent condition for linear independence is that the matrix
V T V is nonsingular.
Definition 2.12. Let X = {VI, V2, ..• } be a collection of vectors Vi E V. Then the span of
X is defined as
Sp(X) = Sp{VI, V2, ... }
= {v : V = (Xl VI + ... + (XkVk ; (Xi ElF, Vi EX, kEN},
where N = {I, 2, ... }.
Example 2.13. Let V = ]Rn and define
0 0
0 1 0
el =
0
, e2 =
0
,'" ,en =
0
o o
Then SpIel, e2, ... , en} = ]Rn.
Definition 2.14. A set of vectors X is a basis for V if and only if
1. X is a linearly independent set (of basis vectors), and
2. Sp(X) = V.
12 Chapter 2. Vector Spaces
Example 2.15. [e\,..., e
n
} is a basis for IR" (sometimes called the natural basis).
Now let b1, ..., b
n
be a basis (with a specific order associated with the basis vectors)
for V. Then for all v e V there exists a unique ntuple {E1 , . . . , E n} such that
Definition 2.16. The scalars {Ei} are called the components (or sometimes the coordinates)
of v with respect to the basis (b1, ..., b
n
] and are unique. We say that the vector x of
components represents the vector v with respect to the basis B.
Example 2.17. In Rn,
we have
To see this, write
Then
Theorem 2.18. The number of elements in a basis of a vector space is independent of the
particular basis considered.
Definition 2.19. If a basis X for a vector space V= 0) has n elements, V is said to
be ndimensional or have dimension n and we write dim(V) = n or dim V — n. For
We can also determine components of v with respect to another basis. For example, while
with respect to the basis
where
12 Chapter 2. Vector Spaces
Example 2.15. {el, ... , en} is a basis for]Rn (sometimes called the natural basis).
Now let b
l
, ... , b
n
be a basis (with a specific order associated with the basis vectors)
for V. Then for all v E V there exists a unique ntuple ... , such that
v = + ... + = Bx,
where
B [b".,b.l. x D J
Definition 2.16. The scalars } are called the components (or sometimes the coordinates)
of v with respect to the basis {b
l
, ... , b
n
} and are unique. We say that the vector x of
components represents the vector v with respect to the basis B.
Example 2.17. In]Rn,
VI ]
: = vlel + V2e2 + ... + vne
n
·
Vn
We can also determine components of v with respect to another basis. For example, while
with respect to the basis
we have
To see this, write
Then
[ ] = I . el + 2 . e2,
[ ] = 3 . [ ] + 4· [ l
[ ] = XI • [  ] + X2 • [ _! ]
= [  ! ] [ l
[ ] = [ ; 1 r I [ ; ] = [ l
Theorem 2.18. The number of elements in a basis of a vector space is independent of the
particular basis considered.
Definition 2.19. If a basis X for a vector space V(Jf 0) has n elements, V is said to
be n.dimensional or have dimension n and we write dim (V) = n or dim V = n. For
2.4 Sums and Intersections of Subspaces
Definition 2.21. Let (V, F) be a vector space and let 71, S c V. The sum and intersection
of R, and S are defined respectively by:
The subspaces R, and S are said to be complements of each other in T.
Remark 2.23. The union of two subspaces, R C S, is not necessarily a subspace.
Definition 2.24. T = R 0 S is the direct sum of R and S if
Theorem 2.22.
2.4. Sums and Intersections of Subspaces 13
consistency, and because the 0 vector is in any vector space, we define dim(O) = 0. A
vector space V is finitedimensional if there exists a basis X with n < +00 elements;
otherwise, V is infinitedimensional.
Thus, Theorem 2.18 says that dim(V) = the number of elements in a basis.
Example 2.20.
1. d i m(Rn)=n.
2. dim(R
mxn
) = mn.
Note: Check that a basis for R
mxn
is given by the mn matrices Eij; i e m, j e n,
where E
f
j is a matrix all of whose elements are 0 except for a 1 in the (i, j)th location.
The collection of Eij matrices can be called the "natural basis matrices."
3. dim(C[to, t1])  +00.
4. dim{A € R
nxn
: A = A
T
} = {1/2(n + 1).
1
2
(To see why, determine 1/ 2n( n + 1) symmetric basis matrices.)
5. dim{A e R
nxn
: A is upper (lower) triangular} = 1/ 2n( n + 1).
1. n + S = {r + s : r e U, s e 5}.
2. ft H 5 = {v : v e 7^ and v e 5}.
K
1. K + S C V (in general, U\  \ h 7^ =: ]T ft/ C V, for finite k).
1=1
2. 72. D 5 C V (in general, f] * R,
a
C V/ or an arbitrary index set A).
a e A
1. n n S = 0, and
2. U + S = T (in general ft; n (^ ft,) = 0 am/ ]Pft, = T).
y>f «
2.4. Sums and Intersections of Subspaces 13
consistency, and because the 0 vector is in any vector space, we define dim(O) = O. A
vector space V is finitedimensional if there exists a basis X with n < +00 elements;
otherwise, V is infinitedimensional.
Thus, Theorem 2.18 says that dim (V) = the number of elements in a basis.
Example 2.20.
1. = n.
2. = mn.
Note: Check that a basis for is given by the mn matrices Eij; i E m, j E
where Eij is a matrix all of whose elements are 0 except for a 1 in the (i, J)th location.
The collection of E;j matrices can be called the "natural basis matrices."
3. dim(C[to, tJJ) = +00.
4. dim{A E : A = AT} = !n(n + 1).
(To see why, determine !n(n + 1) symmetric basis matrices.)
5. dim{A E : A is upper (lower) triangular} = !n(n + 1).
2.4 Sums and Intersections of Subspaces
Definition 2.21. Let (V, JF') be a vector space and let R, S S; V. The sum and intersection
ofR and S are defined respectively by:
1. R + S = {r + s : r E R, s E S}.
2. R n S = {v : v E R and v E S}.
Theorem 2.22.
k
1. R + S S; V (in general, RI + ... + Rk =: L R; S; V, for finite k).
;=1
2. R n S S; V (in general, n Ra S; V for an arbitrary index set A).
CiEA
Remark 2.23. The union of two subspaces, R U S, is not necessarily a subspace.
Definition 2.24. T = REB S is the direct sum ofR and S if
1. R n S = 0, and
2. R + S = T (in general, R; n (L R
j
) = 0 and L Ri = T).
H;
The subspaces Rand S are said to be complements of each other in T.
14 Chapter 2. Vector Spaces
Remark 2.25. The complement of ft (or S) is not unique. For example, consider V = R
2
and let ft be any line through the origin. Then any other distinct line through the origin is
a complement of ft. Among all the complements there is a unique one orthogonal to ft.
We discuss more about orthogonal complements elsewhere in the text.
Theorem 2.26. Suppose T =R O S. Then
1. every t € T can be written uniquely in the form t = r + s with r e R and s e S.
2. dim(T) = dim(ft) + dim(S).
Proof: To prove the first part, suppose an arbitrary vector t e T can be written in two ways
as t = r1 + s1 = r2 + S2, where r1, r2 e R. and s1, S2 e S. Then r1 — r2 = s2— s\. But
r1 –r2 £ ft and 52 — si e S. Since ft fl S = 0, we must have r\ = ri and s\ = si from
which uniqueness follows.
The statement of the second part is a special case of the next theorem. D
Theorem 2.27. For arbitrary subspaces ft, S of a vector space V,
EXERCISES
1. Suppose {vi,..., Vk} is a linearly dependent set. Then show that one of the vectors
must be a linear combination of the others.
2. Let x\, *2, . . . , x/c E R" be nonzero mutually orthogonal vectors. Show that [x\,...,
X k} must be a linearly independent set.
3. Let v\,... ,v
n
be orthonormal vectors in R". Show that Av\,..., Av
n
are also or
thonormal if and only if Ae R"
x
" is orthogonal.
4. Consider the vectors v\ — [2 l]
r
and 1*2 = [3 l]
r
. Prove that vi and V2 form a basis
for R
2
. Find the components of the vector v = [4 l]
r
with respect to this basis.
Example 2.28. Let U be the subspace of upper triangular matrices in E"
x
" and let £ be the
subspace of lower triangular matrices in R
nxn
. Then it may be checked that U + L = R
nxn
while U n £ is the set of diagonal matrices in R
nxn
. Using the fact that dim (diagonal
matrices} = n, together with Examples 2.20.2 and 2.20.5, one can easily verify the validity
of the formula given in Theorem 2.27.
Example 2.29. Let (V, F) = (R
nxn
, R), let ft be the set of skewsymmetric matrices in
R"
x
", and let S be the set of symmetric matrices in R"
x
". Then V = U 0 S.
Proof: This follows easily from the fact that any Ae R"
x
" can be written in the form
The first matrix on the righthand side above is in S while the second is in ft.
14 Chapter 2. Vector Spaces
Remark 2.25. The complement of R (or S) is not unique. For example, consider V = jR2
and let R be any line through the origin. Then any other distinct line through the origin is
a complement of R. Among all the complements there is a unique one orthogonal to R.
We discuss more about orthogonal complements elsewhere in the text.
Theorem 2.26. Suppose T = R EB S. Then
1. every t E T can be written uniquely in the form t = r + s with r E Rand s E S.
2. dim(T) = dim(R) + dim(S).
Proof: To prove the first part, suppose an arbitrary vector t E T can be written in two ways
as t = rl + Sl = r2 + S2, where rl, r2 E Rand SI, S2 E S. Then r,  r2 = S2  SI. But
rl  r2 E Rand S2  SI E S. Since R n S = 0, we must have rl = r2 and SI = S2 from
which uniqueness follows.
The statement of the second part is a special case of the next theorem. 0
Theorem 2.27. For arbitrary subspaces R, S of a vector space V,
dim(R + S) = dim(R) + dim(S)  dim(R n S).
Example 2.28. Let U be the subspace of upper triangular matrices in jRn xn and let .c be the
subspace of lower triangUlar matrices in jRn xn. Then it may be checked that U + .c = jRn xn
while un.c is the set of diagonal matrices in jRnxn. Using the fact that dim {diagonal
matrices} = n, together with Examples 2.20.2 and 2.20.5, one can easily verify the validity
of the formula given in Theorem 2.27.
Example 2.29. Let (V, IF) = (jRnxn, jR), let R be the set of skewsymmetric matrices in
jRnxn, and let S be the set of symmetric matrices in jRnxn. Then V = n $ S.
Proof: This follows easily from the fact that any A E jRnxn can be written in the form
1 TIT
A=2:(A+A )+2:(AA).
The first matrix on the righthand side above is in S while the second is in R.
EXERCISES
1. Suppose {VI, ... , vd is a linearly dependent set. Then show that one of the vectors
must be a linear combination of the others.
2. Let XI, X2, ... , Xk E jRn be nonzero mutually orthogonal vectors. Show that {XI, ... ,
Xk} must be a linearly independent set.
3. Let VI, ... , Vn be orthonormal vectors in jRn. Show that Av" •.. , AV
n
are also or
thonormal if and only if A E jRnxn is orthogonal.
4. Consider the vectors VI = [2 1 f and V2 = [3 1 f. Prove that VI and V2 form a basis
for jR2. Find the components of the vector v = [4 If with respect to this basis.
Exercises 15
5. Let P denote the set of polynomials of degree less than or equal to two of the form
Po + p\x + pix
2
, where po, p\, p2 e R. Show that P is a vector space over E. Show
that the polynomials 1, *, and 2x
2
— 1 are a basis for P. Find the components of the
polynomial 2 + 3x + 4x
2
with respect to this basis.
6. Prove Theorem 2.22 (for the case of two subspaces R and S only).
7. Let P
n
denote the vector space of polynomials of degree less than or equal to n, and of
the form p( x) = po + p\x + • • • + p
n
x
n
, where the coefficients /?, are all real. Let PE
denote the subspace of all even polynomials in P
n
, i.e., those that satisfy the property
p(—x} = p(x). Similarly, let PQ denote the subspace of all odd polynomials, i.e.,
those satisfying p(—x} = – p( x) . Show that P
n
= P
E
© PO
8. Repeat Example 2.28 using instead the two subspaces 7" of tridiagonal matrices and
U of upper triangular matrices.
Exercises 15
5. Let P denote the set of polynomials of degree less than or equal to two of the form
Po + PI X + P2x2, where Po, PI, P2 E R Show that P is a vector space over R Show
that the polynomials 1, x, and 2x2  1 are a basis for P. Find the components of the
polynomial 2 + 3x + 4x
2
with respect to this basis.
6. Prove Theorem 2.22 (for the case of two subspaces Rand S only).
7. Let P
n
denote the vector space of polynomials of degree less than or equal to n, and of
the form p(x) = Po + PIX + ... + Pnxn, where the coefficients Pi are all real. Let PE
denote the subspace of all even polynomials in P
n
, i.e., those that satisfy the property
p( x) = p(x). Similarly, let Po denote the subspace of all odd polynomials, i.e.,
those satisfying p(x) = p(x). Show that P
n
= PE EB Po·
8. Repeat Example 2.28 using instead the two subspaces T of tridiagonal matrices and
U of upper triangular matrices.
This page intentionally left blank This page intentionally left blank
Chapter 3
Linear Transformations
3.1 Definition and Examples
We begin with the basic definition of a linear transformation (or linear map, linear function,
or linear operator) between two vector spaces.
Definition 3.1. Let (V, F) and (W, F) be vector spaces. Then C : V > W is a linear
transformation if and only if
£(avi + pv
2
) = aCv\ + fi£v
2
far all a, £ e F and far all v
}
,v
2
e V.
The vector space V is called the domain of the transformation C while VV, the space into
which it maps, is called the codomain.
Example 3.2.
1. Let F = R and take V = W = PC[f
0
, +00).
Define £ : PC[t
0
, +00) > PC[t
0
, +00) by
2. Let F = R and take V = W = R
mx
". Fix M e R
mxm
.
Define £ : R
mx
" > M
mxn
by
3. Let F = R and take V = P" = (p(x) = a
0
+ ct
}
x H h a
n
x" : a, E R} and
w = p
n

1
.
Define C.: V —> W by Lp — p', where' denotes differentiation with respect to x.
17
Chapter 3
Linear Transformations
3.1 Definition and Examples
We begin with the basic definition of a linear transformation (or linear map, linear function,
or linear operator) between two vector spaces.
Definition 3.1. Let (V, IF) and (W, IF) be vector spaces. Then I: : V + W is a linear
transformation if and only if
I:(avi + {3V2) = al:vi + {3I:V2 for all a, {3 ElF and for all VI, V2 E V.
The vector space V is called the domain of the transformation I: while W, the space into
which it maps, is called the codomain.
Example 3.2.
1. Let IF = JR and take V = W = PC[to, +00).
Define I: : PC[to, +00) + PC[to, +00) by
vet) f+ wet) = (I:v)(t) = 11 e(tr)v(r) dr.
to
2. Let IF = JR and take V = W = JRmxn. Fix ME JRmxm.
Define I: : JRmxn + JRmxn by
X f+ Y = I:X = MX.
3. Let IF = JR and take V = pn = {p(x) = ao + alx + ... + anx
n
: ai E JR} and
W = pnl.
Define I: : V + W by I: p = p', where I denotes differentiation with respect to x.
17
18 Chapters. Li near Transformations
3.2 Matrix Representation of Linear Transformations
Linear transformations between vector spaces with specific bases can be represented con
veniently in matrix form. Specifically, suppose £ : (V, F) — > • (W, F) is linear and further
suppose that {u,, i e n} and {Wj, j e m] are bases for V and W, respectively. Then the
ith column of A = Mat £ (the matrix representation of £ with respect to the given bases
for V and W) is the representation of £i> , with respect to {w
}
•, j e raj. In other words,
represents £ since
where W = [w\,..., w
m
] and
is the z'th column of A. Note that A = Mat £ depends on the particular bases for V and W.
This could be reflected by subscripts, say, in the notation, but this is usually not done.
The action of £ on an arbitrary vector v e V is uniquely determined (by linearity)
by its action on a basis. Thus, if v = E1v1 + • • • + E
n
v
n
= Vx (where u, and hence jc, is
arbitrary), then
Thinking of A both as a matrix and as a linear transformation from Rn to R
m
usually causes no
confusion. Change of basis then corresponds naturally to appropriate matrix multiplication.
Thus, £V = WA since x was arbitrary.
When V = R", W = R
m
and [vi , i e n}, [wj , j e m} are the usual (natural) bases
the equation £V = WA becomes simply £ = A. We thus commonly identify A as a linea
transformation with its matrix representation, i.e.,
18 Chapter 3. Linear Transformations
3.2 Matrix Representation of Linear Transformations
Linear transformations between vector spaces with specific bases can be represented con
veniently in matrix form. Specifically, suppose L : (V, IF) (W, IF) is linear and further
suppose that {Vi, i E and {w j, j E !!!.} are bases for V and W, respectively. Then the
ith column of A = Mat L (the matrix representation of L with respect to the given bases
for V and W) is the representation of LVi with respect to {w j, j E m}. In other words,
represents L since
A=
al
n
]
: E JR.mxn
a
mn
LVi = aliwl + ... + amiWm
=Wai,
where W = [WI, ... , w
m
] and
is the ith column of A. Note that A = Mat L depends on the particular bases for V and W.
This could be reflected by subscripts, say, in the notation, but this is usually not done.
The action of L on an arbitrary vector V E V is uniquely determined (by linearity)
by its action on a basis. Thus, if V = VI + ... + Vn = V x (where v, and hence x, is
arbitrary), then
LVx = Lv = + ... +
= WAx.
Thus, LV = W A since x was arbitrary.
When V = JR.n, W = lR.
m
and {Vi, i E {W j' j E !!!.} are the usual (natural) bases,
the equation LV = W A becomes simply L = A. We thus commonly identify A as a linear
transformation with its matrix representation, i.e.,
Thinking of A both as a matrix and as a linear transformation from JR." to lR.
m
usually causes no
confusion. Change of basis then corresponds naturally to appropriate matrix multiplication.
3.3. Composition of Transformations 19
3.3 Composition of Transformations
Consider three vector spaces U, V, and W and transformations B from U to V and A from
V to W. Then we can define a new transformation C as follows:
formula
Two Special Cases:
Inner Product: Let x, y e Rn. Then their inner product is the scalar
Outer Product: Let x e R
m
, y e Rn. Then their outer product is the m x n
matrix
Note that any rankone matrix A e R
mxn
can be written in the form A = xy
T
above (or xy
H
if A e C
mxn
). A rankone symmetric matrix can be written in
the form XX
T
(or XX
H
).
The above diagram illustrates the composition of transformations C = AB. Note that in
most texts, the arrows above are reversed as follows:
However, it might be useful to prefer the former since the transformations A and B appear
in the same order in both the diagram and the equation. If dimZ// = p, dimV = n,
and dim W = m, and if we associate matrices with the transformations in the usual way,
then composition of transformations corresponds to standard matrix multiplication. That is,
we have C — A B . The above is sometimes expressed componentwise by the
3.3. Composition ofTransformations 19
3.3 Composition of Transformations
Consider three vector spaces U, V, and Wand transformations B from U to V and A from
V to W. Then we can define a new transformation C as follows:
C
The above diagram illustrates the composition of transformations C = AB. Note that in
most texts, the arrows above are reversed as follows:
C
However, it might be useful to prefer the former since the transformations A and B appear
in the same order in both the diagram and the equation. If dimU = p, dim V = n,
and dim W = m, and if we associate matrices with the transformations in the usual way,
then composition of transformations corresponds to standard matrix mUltiplication. That is,
we have CAB . The above is sometimes expressed componentwise by the
mxp
formula
Two Special Cases:
nxp
n
cij = L aikbkj.
k=1
Inner Product: Let x, y E ~ n . Then their inner product is the scalar
n
xTy = Lx;y;.
;=1
Outer Product: Let x E ~ m , y E ~ n . Then their outer product is the m x n
matrix
Note that any rankone matrix A E ~ m x n can be written in the form A = xyT
above (or xyH if A E c
mxn
). A rankone symmetric matrix can be written in
the form xx
T
(or xx
H
).
20 Chapter 3. Li near Transformations
3.4 Structure of Linear Transformations
Let A : V —> W be a linear transformation.
Definition 3.3. The range of A, denotedlZ( A), is the set {w e W : w = Av for some v e V}.
Equivalently, R(A) — {Av : v e V}. The range of A is also known as the image of A and
denoted Im(A).
The nullspace of A, denoted N(A), is the set {v e V : Av = 0}. The nullspace of
A is also known as the kernel of A and denoted Ker (A).
Theorem 3.4. Let A : V — >• W be a linear transformation. Then
1. R( A) C W.
2. N(A) c V.
Note that N(A) and R(A) are, in general, subspaces of different spaces.
Theorem 3.5. Let A e R
mxn
. If A is written in terms of its columns as A = [a\,... ,a
n
],
then
Proof: The proof of this theorem is easy, essentially following immediately from the defi
nition. D
Remark 3.6. Note that in Theorem 3.5 and throughout the text, the same symbol (A) is
used to denote both a linear transformation and its matrix representation with respect to the
usual (natural) bases. See also the last paragraph of Section 3.2.
Definition 3.7. Let {v1 , . . . , vk] be a set of nonzero vectors u, e Rn. The set is said to
be orthogonal if' vjvj = 0 for i ^ j and orthonormal if vf vj = 8
ij
, where 8
t
j is the
Kronecker delta defined by
Example 3.8.
is an orthogonal set.
is an orthonormal set.
3. If { t > i , . . . , Vk} with u, € M." is an orthogonal set, then I — /==,  ., — /===  is an
I ^/v, vi ^/v'
k
v
k
]
orthonormal set.
then
20 Chapter 3. LinearTransformations
3.4 Structure of Linear Transformations
Let A : V + W be a linear transformation.
Definition3.3. The range of A, denotedR(A), is the set {w E W : w = Av for some v E V}.
Equivalently, R(A) = {Av : v E V}. The range of A is also known as the image of A and
denoted Im(A).
The nullspace of A, denoted N(A), is the set {v E V : Av = O}. The nullspace of
A is also known as the kernel of A and denoted Ker (A).
Theorem 3.4. Let A : V + W be a linear transformation. Then
1. R(A) S; W.
2. N(A) S; V.
Note that N(A) and R(A) are, in general, subspaces of different spaces.
Theorem 3.5. Let A E If A is written in terms of its columns as A = [ai, ... , an],
then
R(A) = Sp{al, ... , an} .
Proof: The proof of this theorem is easy, essentially following immediately from the defi
nition. 0
Remark 3.6. Note that in Theorem 3.5 and throughout the text, the same symbol (A) is
used to denote both a linear transformation and its matrix representation with respect to the
usual (natural) bases. See also the last paragraph of Section 3.2.
Definition 3.7. Let {VI, ... , vd be a set of nonzero vectors Vi E The set is said to
be orthogonal if vr v j = 0 for i f= j and orthonormal if vr v j = 8ij' where 8ij is the
Kronecker delta defined by
8 {I ifi=j,
ij = 0 if i f= j.
Example 3.8.
1. {[ J. [ : J} is an orthogonal set.
2. {[ ] ,[ J} is an orthonormal set.
3 If { }
. h 1Tlln • h I th { .
. VI, •.• ,Vk Wit Vi E.IN,. IS an ort ogona set, en ... , IS an
VI
orthonormal set.
3.4. Structure of Linear Transformations 21
Definition 3.9. Let S c Rn. Then the orthogonal complement of S is defined as the set
S
1
 = {v e Rn : V
T
S = 0 for all s e S}.
Example 3.10. Let
Then it can be shown that
Working from the definition, the computation involved is simply to find all nontrivial (i.e.,
nonzero) solutions of the system of equations
Note that there is nothing special about the two vectors in the basis defining S being or
thogonal. Any set of vectors will do, including dependent spanning vectors (which would,
of course, then give rise to redundant equations).
Theorem 311 Let R S C R
n
The
Proof: We prove and discuss only item 2 here. The proofs of the other results are left as
exercises. Let { v1 , ..., v
k
} be an orthonormal basis for S and let x e Rn be an arbitrary
vector. Set
3.4. Structure of Li near Transformations 21
Definition 3.9. Let S <; ]Rn. Then the orthogonal complement of S is defined as the set
vTs=OforallsES}.
Example 3.10. Let
Then it can be shown that
Working from the definition, the computation involved is simply to find all nontrivial (i.e.,
nonzero) solutions of the system of equations
3xI + 5X2 + 7X3 = 0,
4xI + X2 + X3 = 0.
Note that there is nothing special about the two vectors in the basis defining S being or
thogonal. Any set of vectors will do, including dependent spanning vectors (which would,
of course, then give rise to redundant equations).
Theorem 3.11. Let n, S <; ]Rn. Then
2. S \B = ]Rn.
3. = S.
4. n <; S if and only if <;
5. (n + = nl. n
6. (n n = +
Proof: We prove and discuss only item 2 here. The proofs of the other results are left as
exercises. Let {VI, ... , Vk} be an orthonormal basis for S and let x E ]Rn be an arbitrary
vector. Set
k
XI = L (xT Vi)Vi,
;=1
X2 = X XI.
we see that x2 is orthogonal to v1, ..., Vk and hence to any linear combination of these
vectors. In other words, X2 is orthogonal to any vector in S. We have thus shown that
S + S
1
= Rn. We also have that S U S
1
=0 since the only vector s e S orthogonal to
everything in S (i.e., including itself) is 0.
It is also easy to see directly that, when we have such direct sum decompositions, we
can write vectors in a unique way with respect to the corresponding subspaces. Suppose,
for example, that x = x1 + x2. = x'1+ x'
2
, where x\, x 1 E S and x2, x'
2
e S
1
. Then
(x'1 — x1)
T
( x'
2
— x2) = 0 by definition of ST . But then (x'1 — x1)
T
( x' 1 – x1) = 0 since
x
2
— X2 = — (x'1 — x1) (which follows by rearranging the equation x1+x2 = x'1 + x'
2
) . Thus,
x1 — x'1 and x2 = x
2
. D
Theorem 3.12. Let A : Rn —> R
m
. Then
1. N(A)
1
" = 7£(A
r
). (Note: This holds only for finitedimensional vector spaces.)
2. 'R,(A)
1
~ — J\f(A
T
). (Note: This also holds for infinitedimensional vector spaces.)
Proof: To prove the first part, take an arbitrary x e A/ "(A). Then Ax = 0 and this is
equivalent to y
T
Ax = 0 for all v. But y
T
Ax = ( A
T
y ) x. Thus, Ax = 0 if and only if x
is orthogonal to all vectors of the form A
T
v, i.e., x e R(A
r
) . Since x was arbitrary, we
have established that N(A)
1
= U(A
T
}.
The proof of the second part is similar and is left as an exercise. D
Definition 3.13. Let A : R
n
> R
m
. Then {v e R" : Av = 0} is sometimes called the
right nullspace of A. Similarly, (w e R
m
: W
T
A = 0} is called the left nullspace of A.
Clearly, the right nullspace is A/"(A) while the left nullspace is J\f(A
T
).
Theorem 3.12 and part 2 of Theorem 3.11 can be combined to give two very fun
damental and useful decompositions of vectors in the domain and codomain of a linear
transformation A. See also Theorem 2.26.
Theorem 3.14 (Decomposition Theorem). Let A : R" > R
m
. Then
7. every vector v in the domain space R" can be written in a unique way as v = x + y,
where x € M(A) and y € J\f(A)
±
= ft(A
r
) (i.e., R" = M(A) 0 ft(A
r
)).
2. every vector w in the codomain space R
m
can be written in a unique way asw = x+y,
where x e U(A) and y e ft(A)
1
 = Af(A
T
) (i.e., R
m
= 7l(A) 0 M(A
T
)).
This key theorem becomes very easy to remember by carefully studying and under
standing Figure 3.1 in the next section.
3.5 Four Fundamental Subspaces
Consider a general matrix A € E^
x
". When thought of as a linear transformation from E"
to R
m
, many properties of A can be developed in terms of the four fundamental subspaces
22 Chapters. L i near Transformations
Then x\ e < S and, since
22 Chapter 3. Linear Transformations
Then XI E S and, since
T T T
x
2
V j = X V j  X I V j
=XTVjXTVj=O,
we see that X2 is orthogonal to VI, .•. , Vk and hence to any linear combination of these
vectors. In other words, X2 is orthogonal to any vector in S. We have thus shown that
S + S.l = IRn. We also have that S n S.l = 0 since the only vector s E S orthogonal to
everything in S (i.e., including itself) is O.
It is also easy to see directly that, when we have such direct sum decompositions, we
can write vectors in a unique way with respect to the corresponding subspaces. Suppose,
for example, that x = XI + X2 = x; + x ~ , where XI, x; E Sand X2, x ~ E S.l. Then
(x;  XI/ ( x ~  X2) = 0 by definition of S.l. But then (x;  XI)T (x;  xd = 0 since
x ~ X2 = (x; XI) (which follows by rearranging the equation XI +X2 = x; + x ~ ) . Thus,
XI = x; andx2 = x ~ . 0
Theorem 3.12. Let A : IR
n
+ IRm. Then
1. N(A).l = R(A
T
). (Note: This holds only for finitedimensional vector spaces.)
2. R(A).l = N(A
T
). (Note: This also holds for infinitedimensional vector spaces.)
Proof: To prove the first part, take an arbitrary x E N(A). Then Ax = 0 and this is
equivalent to yT Ax = 0 for all y. But yT Ax = (AT y{ x. Thus, Ax = 0 if and only if x
is orthogonal to all vectors of the form AT y, i.e., x E R(AT).l. Since x was arbitrary, we
have established thatN(A).l = R(A
T
).
The proof of the second part is similar and is left as an exercise. 0
Definition 3.13. Let A : IR
n
+ IRm. Then {v E IR
n
: A v = O} is sometimes called the
right nullspace of A. Similarly, {w E IR
m
: w
T
A = O} is called the left nullspace of A.
Clearly, the right nullspace is N(A) while the left nullspace is N(A
T
).
Theorem 3.12 and part 2 of Theorem 3.11 can be combined to give two very fun
damental and useful decompositions of vectors in the domain and codomain of a linear
transformation A. See also Theorem 2.26.
Theorem 3.14 (Decomposition Theorem). Let A : IR
n
+ IRm. Then
1. every vector v in the domain space IR
n
can be written in a unique way as v = x + y,
where x E N(A) and y E N(A).l = R(AT) (i.e., IR
n
= N(A) EB R(A
T
».
2. every vector w in the codomain space IR
m
can be written ina unique way as w = x+y,
where x E R(A) and y E R(A).l = N(A
T
) (i.e., IR
m
= R(A) EBN(A
T
».
This key theorem becomes very easy to remember by carefully studying and under
standing Figure 3.1 in the next section.
3.5 Four Fundamental Subspaces
Consider a general matrix A E lR;,xn. When thought of as a linear transformation from IR
n
to IRm, many properties of A can be developed in terms of the four fundamental subspaces
3.5. Four Fundamental Subspaces 23
Figure 3.1. Four fundamental subspaces.
7£(A), 'R.(A)^, Af ( A) , and N(A)T. Figure 3.1 makes many key properties seem almost
obvious and we return to this figure frequently both in the context of linear transformations
and in illustrating concepts such as controllability and observability.
Definition 3.15. Let V and W be vector spaces and let A : V
motion.
1. A is onto (also called epic or surjective) ifR,(A) = W.
W be a linear transfor
2. A is onetoone or 11 (also called monic or infective) ifJ\f(A) = 0. Two equivalent
characterizations of A being 11 that are often easier to verify in practice are the
following:
Definition 3.16. Let A : E" > R
m
. Then rank(A) = dimftCA). This is sometimes called
the column rank of A (maximum number of independent columns). The row rank of A is
3.5. Four Fundamental Subspaces 23
A
r
N(A)1
r
EB {OJ
X {O}Gl
nr m r
Figure 3.1. Four fundamental subspaces.
R(A), R(A)1, N(A), and N(A)1. Figure 3.1 makes many key properties seem almost
obvious and we return to this figure frequently both in the context of linear transformations
and in illustrating concepts such as controllability and observability.
Definition 3.15. Let V and W be vector spaces and let A : V + W be a linear transfor
mation.
1. A is onto (also called epic or surjective) ifR(A) = W.
2. A is onetoone or 11 (also called monic or injective) if N(A) = O. Two equivalent
characterizations of A being 11 that are often easier to verify in practice are the
following:
(a) AVI = AV2 ===} VI = V2 .
(b) VI t= V2 ===} AVI t= AV2 .
Definition 3.16. Let A : IR
n
+ IRm. Then rank(A) = dim R(A). This is sometimes called
the column rank of A (maximum number of independent columns). The row rank of A is
24 Chapter3. Linear Transformations
dim 7£(A
r
) (maximum number of independent rows). The dual notion to rank is the nullity
of A, sometimes denoted nullity(A) or corank(A), and is defined as dim A/"(A).
Theorem 3.17. Let A : R
n
> R
m
. Then dim K(A) = dimA/ '(A)
±
. (Note: Since
A/^A)
1
" = 7l(A
T
), this theorem is sometimes colloquially stated "row rank of A = column
rank of A.")
Proof: Define a linear transformation T : J\f(A)~
L
— >• 7£(A) by
Clearly T is 11 (since A/"(T) = 0). To see that T is also onto, take any w e 7£(A). Then
by definition there is a vector x e R" such that Ax — w. Write x = x\ + X2, where
x\ e A/^A)
1
 and jc
2
e A/"(A). Then Ajti = u; = r*i since *i e A/^A)
1
. The last equality
shows that T is onto. We thus have that dim7?.(A) = dimA/^A^ since it is easily shown
that if { ui , . . . , iv} is abasis forA/'CA)
1
, then {Tv\, . . . , Tv
r
] is abasis for 7?.(A). Finally, if
we apply this and several previous results, the following string of equalities follows easily:
"column rank of A" = rank(A) = dim7e(A) = dim A/^A)
1
= dim7l(A
T
) = rank(A
r
) =
"row rank of A." D
The following corollary is immediate. Like the theorem, it is a statement about equality
of dimensions; the subspaces themselves are not necessarily in the same vector space.
Corollary 3.18. Let A : R" > R
m
. Then dimA/"(A) + dimft(A) = n, where n is the
dimension of the domain of A.
Proof: From Theorems 3.11 and 3.17 we see immediately that
For completeness, we include here a few miscellaneous results about ranks of sums
and products of matrices.
Theorem 3.19. Let A, B e R"
xn
. Then
Part 4 of Theorem 3.19 suggests looking at the general problem of the four fundamental
subspaces of matrix products. The basic results are contained in the following easily proved
theorem.
24 Chapter 3. LinearTransformations
dim R(AT) (maximum number of independent rows). The dual notion to rank is the nullity
of A, sometimes denoted nullity(A) or corank(A), and is defined as dimN(A).
Theorem 3.17. Let A : ]Rn ~ ]Rm. Then dim R(A) = dimNCA)L. (Note: Since
N(A)L = R(A
T
), this theorem is sometimes colloquially stated "row rank of A = column
rank of A.")
Proof: Define a linear transformation T : N(A)L ~ R(A) by
Tv = Av for all v E N(A)L.
Clearly T is 11 (since N(T) = 0). To see that T is also onto, take any W E R(A). Then
by definition there is a vector x E ]Rn such that Ax = w. Write x = Xl + X2, where
Xl E N(A)L andx2 E N(A). Then AXI = W = TXI since Xl E N(A)L. The last equality
shows that T is onto. We thus have that dim R(A) = dimN(A)L since it is easily shown
that if {VI, ... , v
r
} is a basis for N(A)L, then {TVI, ... , Tv
r
} is a basis for R(A). Finally, if
we apply this and several previous results, the following string of equalities follows easily:
"column rank of A" = rank(A) = dim R(A) = dimN(A)L = dim R(AT) = rank(AT) =
"row rank of A." 0
The following corollary is immediate. Like the theorem, it is a statement about equality
of dimensions; the subspaces themselves are not necessarily in the same vector space.
Corollary 3.18. Let A : ]Rn ~ ]Rm. Then dimN(A) + dim R(A) = n, where n is the
dimension of the domain of A.
Proof: From Theorems 3.11 and 3.17 we see immediately that
n = dimN(A) + dimN(A)L
= dimN(A) + dim R(A) . 0
For completeness, we include here a few miscellaneous results about ranks of sums
and products of matrices.
Theorem 3.19. Let A, B E ]Rnxn. Then
1. O:s rank(A + B) :s rank(A) + rank(B).
2. rank(A) + rank(B)  n :s rank(AB) :s min{rank(A), rank(B)}.
3. nullity(B) :s nullity(AB) :s nullity(A) + nullity(B).
4. if B is nonsingular, rank(AB) = rank(BA) = rank(A) and N(BA) = N(A).
Part 4 of Theorem 3.19 suggests looking atthe general problem of the four fundamental
subspaces of matrix products. The basic results are contained in the following easily proved
theorem.
3.5. Four F u n d a me n t a l Subspaces 25
Theorem 3.20. Let A e R
mxn
, B e R
nxp
. Then
The next theorem is closely related to Theorem 3.20 and is also easily proved. It
is extremely useful in text that follows, especially when dealing with pseudoinverses and
linear least squares problems.
Theorem 3.21. Let A e R
mxn
. Then
We now characterize 11 and onto transformations and provide characterizations in
terms of rank and invertibility.
Theorem 3.22. Let A : R
n
» R
m
. Then
1. A is onto if and only //"rank(A) — m (A has linearly independent rows or is said to
have full row rank; equivalently, AA
T
is nonsingular).
2. A is 11 if and only z/r a nk(A) = n (A has linearly independent columns or is said
to have full column rank; equivalently, A
T
A is nonsingular).
Proof: Proof of part 1: If A is onto, dim7?,(A) — m — rank (A). Conversely, let y e R
m
be arbitrary. Let jc = A
T
(AA
T
)~
]
y e R
n
. Then y = Ax, i.e., y e 7?.(A), so A is onto.
Proof of part 2: If A is 11, then A/"(A) = 0, which implies that dim A/^A)
1
—n —
dim 7£(A
r
), and hence dim 7£(A) = n by Theorem 3.17. Conversely, suppose Ax\ = Ax^.
Then A
r
A;t i = A
T
Ax2, which implies x\ = x^. since A
r
A is invertible. Thus, A is
11. D
Definition 3.23. A : V —» W is invertible (or bijective) if and only if it is 11 and onto.
Note that if A is invertible, then dim V — dim W. Also, A : W
1
»• E" is invertible or
nonsingular if and only z/r ank(A) = n.
Note that in the special case when A € R"
x
", the transformations A, A
r
, and A"
1
are all 11 and onto between the two spaces M(A)
±
and 7£(A). The transformations A
T
and A~
!
have the same domain and range but are in general different maps unless A is
orthogonal. Similar remarks apply to A and A~
T
.
3.5. Four Fundamental Subspaces 25
Theorem 3.20. Let A E IRmxn, B E IRnxp. Then
1. RCAB) S; RCA).
2. N(AB) ;2 N(B).
3. R«AB)T) S; R(B
T
).
4. N«AB)T) ;2 N(A
T
).
The next theorem is closely related to Theorem 3.20 and is also easily proved. It
is extremely useful in text that follows, especially when dealing with pseudoinverses and
linear least squares problems.
Theorem 3.21. Let A E IRmxn. Then
1. R(A) = R(AA
T
).
2. R(AT) = R(A
T
A).
3. N(A) = N(A
T
A).
4. N(A
T
) = N(AA
T
).
We now characterize II and onto transformations and provide characterizations in
terms of rank and invertibility.
Theorem 3.22. Let A : IR
n
+ IRm. Then
1. A is onto if and only if rank (A) = m (A has linearly independent rows or is said to
have full row rank; equivalently, AA T is nonsingular).
2. A is 11 if and only ifrank(A) = n (A has linearly independent columns or is said
to have full column rank; equivalently, AT A is nonsingular).
Proof' Proof of part 1: If A is onto, dim R(A) = m = rank(A). Conversely, let y E IRm
be arbitrary. Let x = AT (AAT)I Y E IRn. Then y = Ax, i.e., y E R(A), so A is onto.
Proof of part 2: If A is 11, then N(A) = 0, which implies that dimN(A)1 = n =
dim R(A
T
), and hence dim R(A) = n by Theorem 3.17. Conversely, suppose AXI = AX2.
Then AT AXI = AT AX2, which implies XI = X2 since AT A is invertible. Thus, A is
11. D
Definition 3.23. A : V + W is invertible (or bijective) if and only if it is 11 and onto.
Note that if A is invertible, then dim V = dim W. Also, A : IRn + IR
n
is invertible or
nonsingular ifand only ifrank(A) = n.
Note that in the special case when A E I R ~ x n , the transformations A, AT, and AI
are all 11 and onto between the two spaces N(A)1 and R(A). The transformations AT
and A I have the same domain and range but are in general different maps unless A is
orthogonal. Similar remarks apply to A and A T.
26 Chapters. Li near Transformations
If a linear transformation is not invertible, it may still be right or left invertible. Defi
nitions of these concepts are followed by a theorem characterizing left and right invertible
transformations.
Definition 3.24. Let A : V > W. Then
1. A is said to be right invertible if there exists a right inverse transformation A~
R
:
W —> V such that AA~
R
= I
w
, where I
w
denotes the identity transformation on W.
2. A is said to be left invertible if there exists a left inverse transformation A~
L
: W —>
V such that A~
L
A = I
v
, where I
v
denotes the identity transformation on V.
Theorem 3.25. Let A : V > W. Then
1. A is right invertible if and only if it is onto.
2. A is left invertible if and only if it is 11.
Moreover, A is invertible if and only if it is both right and left invertible, i.e., both 11 and
onto, in which case A~
l
= A~
R
= A~
L
.
Note: From Theorem 3.22 we see that if A : E" >• E
m
is onto, then a right inverse
is given by A~
R
= A
T
(AA
T
) . Similarly, if A is 11, then a left inverse is given by
A~
L
= (A
T
A)~
1
A
T
.
Theorem 3.26. Let A : V » V.
1. If there exists a unique right inverse A~
R
such that AA~
R
= I, then A is invertible.
2. If there exists a unique left inverse A~
L
such that A~
L
A = I, then A is invertible.
Proof: We prove the first part and leave the proof of the second to the reader. Notice the
following:
Thus, (A
R
+ A
R
A — /) must be a right inverse and, therefore, by uniqueness it must be
the case that A~
R
+ A~
R
A — I = A~
R
. But this implies that A~
R
A = /, i.e., that A~
R
is
a left inverse. It then follows from Theorem 3.25 that A is invertible. D
Example 3.27.
1. Let A = [1 2] : E
2
»• E
1
. Then A is onto. (Proof: Take any a € E
1
; then one
can always find v e E
2
such that [1 2][^] = a). Obviously A has full row rank
(=1) and A~
R
= [ _j j is a right inverse. Also, it is clear that there are infinitely many
right inverses for A. In Chapter 6 we characterize all right inverses of a matrix by
characterizing all solutions of the linear matrix equation AR = I.
26 Chapter 3. linear Transformations
If a linear transformation is not invertible, it may still be right or left invertible. Defi
nitions of these concepts are followed by a theorem characterizing left and right invertible
transformations.
Definition 3.24. Let A : V + W. Then
1. A is said to be right invertible if there exists a right inverse transformation A
R
:
W + V such that AA R = I
w
, where Iw denotes the identity transfonnation on W.
2. A is said to be left invertible if there exists a left inverse transformation A L : W +
V such that A L A = Iv, where Iv denotes the identity transfonnation on V.
Theorem 3.25. Let A : V + W. Then
1. A is right invertible if and only if it is onto.
2. A is left invertible if and only ifit is 11.
Moreover, A is invertible if and only if it is both right and left invertible, i.e., both 11 and
onto, in which case A I = A R = A L.
Note: From Theorem 3.22 we see that if A : ]Rn + ]Rm is onto, then a right inverse
is given by A R = AT (AAT) I. Similarly, if A is 11, then a left inverse is given by
A
L
= (AT A)I AT.
Theorem 3.26. Let A : V + V.
1. If there exists a unique right inverse A  R such that A A  R = I, then A is invertible.
2. If there exists a unique left inverse A L such that A L A = I, then A is invertible.
Proof: We prove the first part and leave the proof of the second to the reader. Notice the
following:
A(A
R
+ ARA I) = AA
R
+ AARA  A
= I + I A  A since AA R = I
= I.
Thus, (A R + A R A  I) must be a right inverse and, therefore, by uniqueness it must be
the case that A R + A R A  I = A R. But this implies that A R A = I, i.e., that A R is
a left inverse. It then follows from Theorem 3.25 that A is invertible. 0
Example 3.27.
1. Let A = [1 2]:]R2 + ]R I. Then A is onto. (Proo!' Take any a E ]R I; then one
can always find v E ]R2 such that [1 2][ ~ ~ ] = a). Obviously A has full row rank
(= 1) and A  R = [ _ ~ ] is a right inverse. Also, it is clear that there are infinitely many
right inverses for A. In Chapter 6 we characterize all right inverses of a matrix by
characterizing all solutions of the linear matrix equation A R = I.
Exercises 27
2. Let A = [J] : E
1
> E
2
. ThenAis 11. (Proof: The only solution to 0 = Av = [
I
2
]v
is v = 0, whence A/"(A) = 0 so A is 11). It is now obvious that A has full column
rank (=1) and A~
L
= [3 — 1] is a left inverse. Again, it is clear that there are
infinitely many left inverses for A. In Chapter 6 we characterize all left inverses of a
matrix by characterizing all solutions of the linear matrix equation LA = I.
3. The matrix
when considered as a linear transformation on IE
below bases for its four fundamental subspaces.
\ is neither 11 nor onto. We give
EXERCISES
3 4
1. Let A = [
8 5
J and consider A as a linear transformation mapping E
3
to E
2
.
Find the matrix representation of A with respect to the bases
2. Consider the vector space R
nx
" over E, let S denote the subspace of symmetric
matrices, and let 7£ denote the subspace of skewsymmetric matrices. For matrices
X, Y e E
nx
" define their inner product by (X, Y) = Tr( X
r
F) . Show that, with
respect to this inner product, 'R, — S^.
3. Consider the differentiation operator C defined in Example 3.2.3. Is £ 11? Is£
onto?
4. Prove Theorem 3.4.
of R
3
and
of E
2
.
Exercises 27
2. LetA = [i]:]Rl ~ ]R2. Then A is 11. (Proof The only solution toO = Av = [i]v
is v = 0, whence N(A) = 0 so A is 11). It is now obvious that A has full column
rank (=1) and A L = [3  1] is a left inverse. Again, it is clear that there are
infinitely many left inverses for A. In Chapter 6 we characterize all left inverses of a
matrix by characterizing all solutions of the linear matrix equation LA = I.
3. The matrix
[
1 1
A = 2 1
3 1
when considered as a linear transformation on ]R3, is neither 11 nor onto. We give
below bases for its four fundamental subspaces.
EXERCISES
1. Let A = [ ~ ; i) and consider A as a linear transformation mapping ]R3 to ]R2.
Find the matrix representation of A with respect to the bases
{[lHHU]}
{ [ i l [ ~ J }
2. Consider the vector space ]Rnxn over ]R, let S denote the subspace of symmetric
matrices, and let R denote the subspace of skewsymmetric matrices. For matrices
X, Y E ]Rnxn define their inner product by (X, y) = Tr(X
T
Y). Show that, with
respect to this inner product, R = S J. .
3. Consider the differentiation operator £, defined in Example 3.2.3. Is £, II? Is £,
onto?
4. Prove Theorem 3.4.
28 Chapters. Linear Transformations
5. Prove Theorem 3.11.4.
6. Prove Theorem 3.12.2.
7. Determine bases for the four fundamental subspaces of the matrix
8. Suppose A e R
mxn
has a left inverse. Show that A
T
has a right inverse.
9. Let A = [ J o]. Determine A/"(A) and 7£(A). Are they equal? Is this true in general?
If this is true in general, prove it; if not, provide a counterexample.
10. Suppose A € Mg
9x48
. How many linearly independent solutions can be found to the
homogeneous linear system Ax = 0?
11. Modify Figure 3.1 to illustrate the four fundamental subspaces associated with A
T
e
R
nxm
thought of as a transformation from R
m
to R".
28 Chapter 3. Linear Transformations
5. Prove Theorem 3.Il.4.
6. Prove Theorem 3.12.2.
7. Detennine bases for the four fundamental subspaces of the matrix
2 5 5 3
8. Suppose A E IR
m
xn has a left inverse. Show that A T has a right inverse.
9. Let A = n DetennineN(A) and R(A). Are they equal? Is this true in general?
If this is true in general, prove it; if not, provide a counterexample.
10. Suppose A E How many linearly independent solutions can be found to the
homogeneous linear system Ax = O?
11. Modify Figure 3.1 to illustrate the four fundamental subspaces associated with ATE
IR
nxm
thought of as a transformation from IR
m
to IRn.
Chapter 4
Introduction to the
MoorePen rose
Pseudoinverse
In this chapter we give a brief introduction to the MoorePenrose pseudoinverse, a gener
alization of the inverse of a matrix. The MoorePenrose pseudoinverse is defined for any
matrix and, as is shown in the following text, brings great notational and conceptual clarity
to the study of solutions to arbitrary systems of linear equations and linear least squares
problems.
4.1 Definitions and Characterizations
Consider a linear transformation A : X —>• y, where X and y are arbitrary finite
dimensional vector spaces. Define a transformation T : Af(A)
1
 —>• Tl(A) by
Then, as noted in the proof of Theorem 3.17, T is bijective (11 and onto), and hence we
can define a unique inverse transformation T~
l
: 7£(A) —>• J\f(A}~
L
. This transformation
can be used to give our first definition of A
+
, the MoorePenrose pseudoinverse of A.
Unfortunately, the definition neither provides nor suggests a good computational strategy
for determining A
+
.
Definition 4.1. With A and T as defined above, define a transformation A
+
: y —» • X by
where y = y\ + j2 with y\ e 7£(A) and yi e Tl(A}
L
. Then A
+
is the MoorePenrose
pseudoinverse of A.
Although X and y were arbitrary vector spaces above, let us henceforth consider the
case X = W
1
and y = R
m
. We have thus defined A+ for all A e IR ™
X
" . A purely algebraic
characterization of A
+
is given in the next theorem, which was proved by Penrose in 1955;
see [22].
29
Chapter 4
Introduction to the
MoorePenrose
Pseudoinverse
In this chapter we give a brief introduction to the MoorePenrose pseudoinverse, a gener
alization of the inverse of a matrix. The MoorePenrose pseudoinverse is defined for any
matrix and, as is shown in the following text, brings great notational and conceptual clarity
to the study of solutions to arbitrary systems of linear equations and linear least squares
problems.
4.1 Definitions and Characterizations
Consider a linear transformation A : X + y, where X and Y are arbitrary finite
dimensional vector spaces. Define a transformation T : N(A).l + R(A) by
Tx = Ax for all x E NCA).l.
Then, as noted in the proof of Theorem 3.17, T is bijective Cll and onto), and hence we
can define a unique inverse transformation T
1
: RCA) + NCA).l. This transformation
can be used to give our first definition of A +, the MoorePenrose pseudoinverse of A.
Unfortunately, the definition neither provides nor suggests a good computational strategy
for determining A + .
Definition 4.1. With A and T as defined above, define a transformation A + : Y + X by
where Y = YI + Yz with Yl E RCA) and Yz E RCA).l. Then A+ is the MoorePenrose
pseudoinverse of A.
Although X and Y were arbitrary vector spaces above, let us henceforth consider the
case X = ~ n and Y = lP1.
m
. We have thus defined A + for all A E lP1.;" xn. A purely algebraic
characterization of A + is given in the next theorem, which was proved by Penrose in 1955;
see [22].
29
30 Chapter 4. Introduction to the MoorePenrose Pseudoinverse
Theorem 4.2. Let A e R?
xn
. Then G = A
+
i f and only i f
(PI) AGA = A.
(P2) GAG = G.
(P3) (AGf = AG.
(P4) (GA)
T
= GA.
Furthermore, A
+
always exi sts and i s uni que.
Note that the inverse of a nonsingular matrix satisfies all four Penrose properties. Also,
a right or left inverse satisfies no fewer than three of the four properties. Unfortunately, as
with Definition 4.1, neither the statement of Theorem 4.2 nor its proof suggests a computa
tional algorithm. However, the Penrose properties do offer the great virtue of providing a
checkable criterion in the following sense. Given a matrix G that is a candidate for being
the pseudoinverse of A, one need simply verify the four Penrose conditions (P1)(P4). If G
satisfies all four, then by uniqueness, it must be A
+
. Such a verification is often relatively
straightforward.
Example 4.3. Consider A = [' ]. Verify directly that A
+
= [ f ] satisfies (P1)(P4).
Note that other left inverses (for example, A~
L
= [3 — 1]) satisfy properties (PI), (P2),
and (P4) but not (P3).
Still another characterization of A
+
is given in the following theorem, whose proof
can be found in [1, p. 19]. While not generally suitable for computer implementation, this
characterization can be useful for hand calculation of small examples.
Theorem 4.4. Let A e R™
xn
. Then
4.2 Examples
Each of the following can be derived or verified by using the above definitions or charac
terizations.
Example 4.5. A
+
= A
T
(AA
T
)~ if A is onto (independent rows) (A is right invertible).
Example 4.6. A
+
= (A
T
A)~ A
T
i f A is 11 (independent columns) (A is left invertible).
Example 4.7. For any scalar a,
30 Chapter 4. Introduction to the MoorePenrose Pseudoinverse
Theorem 4.2. Let A E lR;" xn. Then G = A + if and only if
(Pl) AGA = A.
(P2) GAG = G.
(P3) (AG)T = AG.
(P4) (GA)T = GA.
Furthermore, A + always exists and is unique.
Note that the inverse of a nonsingular matrix satisfies all four Penrose properties. Also,
a right or left inverse satisfies no fewer than three of the four properties. Unfortunately, as
with Definition 4.1, neither the statement of Theorem 4.2 nor its proof suggests a computa
tional algorithm. However, the Penrose properties do offer the great virtue of providing a
checkable criterion in the following sense. Given a matrix G that is a candidate for being
the pseudoinverse of A, one need simply verify the four Penrose conditions (P1)(P4). If G
satisfies all four, then by uniqueness, it must be A +. Such a verification is often relatively
straightforward.
Example 4.3. Consider A = [a Verify directly that A+ = [! ~ ] satisfies (PI)(P4).
Note that other left inverses (for example, A L = [3  1]) satisfy properties (PI), (P2),
and (P4) but not (P3).
Still another characterization of A + is given in the following theorem, whose proof
can be found in [1, p. 19]. While not generally suitable for computer implementation, this
characterization can be useful for hand calculation of small examples.
Theorem 4.4. Let A E lR;" xn. Then
4.2 Examples
A + = lim (AT A + 8
2
1) I AT
6+0
= limAT(AAT +8
2
1)1.
6+0
(4.1)
(4.2)
Each of the following can be derived or verified by using the above definitions or charac
terizations.
Example 4.5. X
t
= AT (AA T) I if A is onto (independent rows) (A is right invertible).
Example 4.6. A+ = (AT A)I AT if A is 11 (independent columns) (A is left invertible).
Example 4.7. For any scalar a,
if a t= 0,
if a =0.
4.3. Properties and Appl ications 31
Example 4.8. For any vector v e M",
Example 4.9.
Example 4.10.
4.3 Properties and Applications
This section presents some miscellaneous useful results on pseudoinverses. Many of these
are used in the text that follows.
Theorem 4.11. Let A e R
mx
" and suppose U e R
mxm
, V e R
nx
" are orthogonal (M is
orthogonal if M
T
= M
1
). Then
Proof: For the proof, simply verify that the expression above does indeed satisfy each c
the four Penrose conditions. D
Theorem 4.12. Let S e R
nxn
be symmetric with U
T
SU = D, where U is orthogonal an
D is diagonal. Then S
+
= UD
+
U
T
, where D
+
is again a diagonal matrix whose diagonc
elements are determined according to Example 4.7.
Theorem 4.13. For all A e R
mxn
,
Proof: Both results can be proved using the limit characterization of Theorem 4.4. The
proof of the first result is not particularly easy and does not even have the virtue of being
especially illuminating. The interested reader can consult the proof in [1, p. 27]. The
proof of the second result (which can also be proved easily by verifying the four Penrose
conditions) is as follows:
4.3. Properties and Applications
Example 4.8. For any vector v E jRn,
Example 4.9.
[ ~ ~ r = [
0
~ l
[ ~ ~ r = [
I I
1
Example 4.10.
4 4
I I
4 4
4.3 Properties and Applications
if v i= 0,
if v = O.
31
This section presents some miscellaneous useful results on pseudoinverses. Many of these
are used in the text that follows.
Theorem 4.11. Let A E jRmxn and suppose U E jRmxm, V E jRnxn are orthogonal (M is
orthogonal if MT = M
1
). Then
Proof: For the proof, simply verify that the expression above does indeed satisfy each of
the four Penrose conditions. 0
Theorem 4.12. Let S E jRnxn be symmetric with U
T
SU = D, where U is orthogonal and
D is diagonal. Then S+ = U D+U
T
, where D+ is again a diagonal matrix whose diagonal
elements are determined according to Example 4.7.
Theorem 4.13. For all A E jRmxn,
1. A+ = (AT A)+ AT = AT (AA
T
)+.
2. (A
T
)+ = (A+{.
Proof: Both results can be proved using the limit characterization of Theorem 4.4. The
proof of the first result is not particularly easy and does not even have the virtue of being
especially illuminating. The interested reader can consult the proof in [1, p. 27]. The
proof of the second result (which can also be proved easily by verifying the four Penrose
conditions) is as follows:
(A
T
)+ = lim (AA
T
+ 8
2
l)IA
~   + O
= lim [AT(AAT + 8
2
l)1{
~   + O
= [limAT(AAT + 8
2
l)1{
~   + O
= (A+{. 0
32 Chapter 4. Introduction to the MoorePenrose Pseudoinverse
Note that by combining Theorems 4.12 and 4.13 we can, in theory at least, compute
the MoorePenrose pseudoinverse of any matrix (since A A
T
and A
T
A are symmetric). This
turns out to be a poor approach in finiteprecision arithmetic, however (see, e.g., [7], [11],
[23]), and better methods are suggested in text that follows.
Theorem 4.11 is suggestive of a "reverseorder" property for pseudoinverses of prod
nets of matrices such as exists for inverses of nroducts TTnfortnnatelv. in peneraK
As an example consider A = [0 1J and B = LI. Then
while
However, necessary and sufficient conditions under which the reverseorder property does
hold are known and we quote a couple of moderately useful results for reference.
Theorem 4.14. (AB)
+
= B
+
A
+
if and only if
Proof: For the proof, see [9]. D
Theorem 4.15. (AB)
+
= B?A+, where BI = A+AB and A) = AB\B+.
Proof: For the proof, see [5]. D
Theorem 4.16. If A e R
n
r
xr
, B e R
r
r
xm
, then (AB)
+
= B+A+.
Proof: Since A e R
n
r
xr
, then A
+
= (A
T
A)~
l
A
T
, whence A
+
A = I
r
. Similarly, since
B e W
r
xm
, we have B
+
= B
T
(BB
T
)~\ whence BB
+
= I
r
. The result then follows by
taking BI = B, A\ = A in Theorem 4.15. D
The following theorem gives some additional useful properties of pseudoinverses.
Theorem 4.17. For all A e R
mxn
,
32 Chapter 4. Introduction to the MoorePenrose Pseudo inverse
Note that by combining Theorems 4.12 and 4.13 we can, in theory at least, compute
the MoorePenrose pseudoinverse of any matrix (since AAT and AT A are symmetric). This
turns out to be a poor approach in finiteprecision arithmetic, however (see, e.g., [7], [II],
[23]), and better methods are suggested in text that follows.
Theorem 4.11 is suggestive of a "reverseorder" property for pseudoinverses of prod
ucts of matrices such as exists for inverses of products. Unfortunately, in general,
As an example consider A = [0 I] and B = [ : J. Then
(AB)+ = 1+ = I
while
B+ A+ = [ ] =
However, necessary and sufficient conditions under which the reverseorder property does
hold are known and we quote a couple of moderately useful results for reference.
Theorem 4.14. (AB)+ = B+ A + if and only if
1. n(BB
T
AT) n(AT)
and
2. n(A T AB) nCB) .
Proof: For the proof, see [9]. 0
Theorem 4.15. (AB)+ = B{ Ai, where BI = A+ AB and AI = ABIB{.
Proof: For the proof, see [5]. 0
Theorem 4.16. If A E B E then (AB)+ = B+ A+.
Proof' Since A E then A+ = (AT A)I AT, whence A+ A = f
r
• Similarly, since
B E lR;xm, we have B+ = BT(BBT)I, whence BB+ = f
r
. The result then follows by
takingB
t
= B,At = A in Theorem 4.15. 0
The following theorem gives some additional useful properties of pseudoinverses.
Theorem 4.17. For all A E lR
mxn
,
1. (A+)+ = A.
2. (AT A)+ = A+(A
T
)+, (AA
T
)+ = (A
T
)+ A+.
3. n(A+) = n(A
T
) = n(A+ A) = n(A
T
A).
4. N(A+) = N(AA+) = N«AA
T
)+) = N(AA
T
) = N(A
T
).
5. If A is normal, then AkA+ = A+ Ak and (Ak)+ = (A+)kforall integers k > O.
Exercises 33
Note: Recall that A e R"
xn
is normal if AA
T
= A
T
A. For example, if A is symmetric,
skewsymmetric, or orthogonal, then it is normal. However, a matrix can be none of the
preceding but still be normal, such as
for scalars a, b e E.
The next theorem is fundamental to facilitating a compact and unifying approach
to studying the existence of solutions of (matrix) linear equations and linear least squares
problems.
Theorem 4.18. Suppose A e R
nxp
, B e E
MX m
. Then K(B) c U(A) if and only if
AA+B = B.
Proof: Suppose K(B) c U(A) and take arbitrary jc e R
m
. Then Bx e H(B) c H(A), so
there exists a vector y e R
p
such that Ay = Bx. Then we have
where one of the Penrose properties is used above. Since x was arbitrary, we have shown
that B = AA+B.
To prove the converse, assume that AA
+
B = B and take arbitrary y e K(B). Then
there exists a vector x e R
m
such that Bx = y, whereupon
EXERCISES
1. Use Theorem 4.4 to compute the pseudoinverse of \
2 2
1 •
2. If jc, y e R", show that (xy
T
)
+
= (x
T
x)
+
(y
T
y)
+
yx
T
.
3. For A e R
mxn
, prove that 7£(A) = 7£(AA
r
) using only definitions and elementary
properties of the MoorePenrose pseudoinverse.
4. For A e R
mxn
, prove that ft(A+) = ft(A
r
).
5. For A e R
pxn
and 5 € R
mx
", show that JV(A) C A/"(S) if and only if fiA+A = B.
6. Let A G M"
xn
, 5 e E
nxm
, and D € E
mxm
and suppose further that D is nonsingular.
(a) Prove or disprove that
(b) Prove or disprove that
Exercises 33
Note: Recall that A E IRn xn is normal if A A T = A T A. For example, if A is symmetric,
skewsymmetric, or orthogonal, then it is normal. However, a matrix can be none of the
preceding but still be normal, such as
A=[ a b]
b a
for scalars a, b E R
The next theorem is fundamental to facilitating a compact and unifying approach
to studying the existence of solutions of (matrix) linear equations and linear least squares
problems.
Theorem 4.18. Suppose A E IRnxp, B E IRnxm. Then R(B) S; R(A) if and only if
AA+B = B.
Proof: Suppose R(B) S; R(A) and take arbitrary x E IRm. Then Bx E R(B) S; RCA), so
there exists a vector y E IRP such that Ay = Bx. Then we have
Bx = Ay = AA + Ay = AA + Bx,
where one of the Penrose properties is used above. Since x was arbitrary, we have shown
that B = AA+ B.
To prove the converse, assume that AA + B = B and take arbitrary y E R(B). Then
there exists a vector x E IR
m
such that Bx = y, whereupon
y = Bx = AA+Bx E R(A). 0
EXERCISES
1. Use Theorem 4.4 to compute the pseudoinverse of U ;].
2. If x, Y E IRn, show that (xyT)+ = (x
T
x)+(yT y)+ yx
T
.
3. For A E IRmxn, prove that RCA) = R(AAT) using only definitions and elementary
properties of the MoorePenrose pseudoinverse.
4. For A E IRmxn, prove that R(A+) = R(A
T
).
5. For A E IRPxn and BE IRmxn, show thatN(A) S; N(B) if and only if BA+ A = B.
6. Let A E IRn xn, B E JRn xm , and D E IRm xm and suppose further that D is nonsingular.
(a) Prove or disprove that
[ ~
AB
r = [
A+ A+ABD
i
].
D 0
D
i
(b) Prove or disprove that
[ ~
B
r =[
A+ A+BD
1
l
D 0
D
i
This page intentionally left blank This page intentionally left blank
Chapter 5
Introduction to the Singular
Value Decomposition
In this chapter we give a brief introduction to the singular value decomposition (SVD). We
show that every matrix has an SVD and describe some useful properties and applications
of this important matrix factorization. The SVD plays a key conceptual and computational
role throughout (numerical) linear algebra and its applications.
5.1 The Fundamental Theorem
Theorem 5.1. Let A e R™
xn
. Then there exist orthogonal matrices U e R
mxm
and
V € R
nxn
such that
where S = [J °
0
], S = diagfcri, ... , o>) e R
rxr
, and a\ > • • • > o
r
> 0. More
specifically, we have
The submatrix sizes are all determined by r (which must be < min{m, «}), i.e., U\ e W
nxr
,
U
2 e
^x(mr)
; Vi e R
«xr
j
y
2 €
Rnxfor^
and the
0
JM
^/ocJb in E are compatibly
dimensioned.
Proof: Since A
r
A> 0 ( A
r
Ai s symmetric and nonnegative definite; recall, for example,
[24, Ch. 6]), its eigenvalues are all real and nonnegative. (Note: The rest of the proof follows
analogously if we start with the observation that A A
T
> 0 and the details are left to the reader
as an exercise.) Denote the set of eigenvalues of A
T
A by {of , / e n} with a\ > • • • > a
r
>
0 = o>
+
i = • • • = a
n
. Let {u, , i e n} be a set of corresponding orthonormal eigenvectors
and let V\ = [v\, ..., v
r
] , Vi = [v
r+
\, . . . , v
n
]. Letting S — diag(cri, . . . , cf
r
), we can
write A
r
AVi = ViS
2
. Premultiplying by Vf gives Vf A
T
AVi = VfV^S
2
= S
2
, the latter
equality following from the orthonormality of the r;, vectors. Pre and postmultiplying by
S~
l
eives the emotion
35
Chapter 5
Introduction to the Singular
Value Decomposition
In this chapter we give a brief introduction to the singular value decomposition (SVD). We
show that every matrix has an SVD and describe some useful properties and applications
of this important matrix factorization. The SVD plays a key conceptual and computational
role throughout (numerical) linear algebra and its applications.
5.1 The Fundamental Theorem
Theorem 5.1. Let A E Then there exist orthogonal matrices U E IRmxm and
V E IR
nxn
such that
A =
(5.1)
where =
n
S diag(ul, ... , u
r
) E
IRrxr, and UI
> > U
r
> O. More
specifically, we have
U2) [
0
] [
V
T
]
A = [U
I
I
(5.2)
0
VT
2
= Ulsvt·
(5.3)
The submatrix sizes are all determined by r (which must be S min{m, n}), i.e., UI E IRmxr,
U2 E IRrnx(mrl, VI E IRnxr, V
2
E IRnx(nr), and the Osubblocks in are compatibly
dimensioned.
Proof: Since AT A ::::: 0 (AT A is symmetric and nonnegative definite; recall, for example,
[24, Ch. 6]), its eigenvalues are all real and nonnegative. (Note: The rest of the proof follows
analogously if we start with the observation that AAT ::::: 0 and the details are left to the reader
as an exercise.) Denote the set of eigenvalues of AT A by {U?, i E !!.} with UI ::::: ... ::::: U
r
>
0= Ur+1 = ... = Un. Let {Vi, i E !!.} be a set of corresponding orthonormal eigenvectors
and let VI = [VI, ... ,V
r
), V2 = [Vr+I, ... ,V
n
]. LettingS = diag(uI, ... ,u
r
), we can
write A T A VI = VI S2. Premultiplying by vt gives vt A T A VI = vt VI S2 = S2, the latter
equality following from the orthonormality of the Vi vectors. Pre and postmultiplying by
SI gives the equation
(5.4)
35
36 Chapter 5. Introduction to the Singular Value Decomposition
Turning now to the eigenvalue equations corresponding to the eigenvalues o
r+
\, . . . , a
n
we
have that A
T
AV
2
= V
2
0 = 0, whence Vf A
T
AV
2
= 0. Thus, AV
2
= 0. Now define the
matrix Ui e M
mx/
" by U\ = AViS~
l
. Then from (5.4) we see that UfU\ = /; i.e., the
columns of U\ are orthonormal. Choose any matrix U
2
£ ^
7 7 I X(
™~
r)
such that [U\ U
2
] is
orthogonal. Then
since A V
2
=0. Referring to the equation U\ = A V\ S
l
defining U\, we see that U{ AV\ =
S and 1/2 AVi = U^UiS = 0. The latter equality follows from the orthogonality of the
columns of U\ andU
2
. Thus, we see that, in fact, U
T
AV = [Q Q], and defining this matrix
to be S completes the proof. D
Definition 5.2. Let A = t/E V
T
be an SVD of A as in Theorem 5.1.
1. The set [a\,..., a
r
} is called the set of (nonzero) singular values of the matrix A and
i
is denoted £(A). From the proof of Theorem 5.1 we see that cr,(A) = A
(
2
(A
T
A) =
A.? (AA
T
). Note that there are alsomin{m, n] — r zero singular values.
2. The columns ofU are called the left singular vectors of A (and are the orthonormal
eigenvectors of AA
T
).
3. The columns of V are called the right singular vectors of A (and are the orthonormal
eigenvectors of A
1
A).
Remark 5.3. The analogous complex case in which A e C™
x
" is quite straightforward.
The decomposition is A = t/E V
H
, where U and V are unitary and the proof is essentially
identical, except for Hermitian transposes replacing transposes.
Remark 5.4. Note that U and V can be interpreted as changes of basis in both the domain
and codomain spaces with respect to which A then has a diagonal matrix representation.
Specifically, let C, denote A thought of as a linear transformation mapping W to W. Then
rewriting A = U^V
T
as AV = U E we see that Mat £ is S with respect to the bases
[v\,..., v
n
} for R" and {u\,..., u
m
] for R
m
(see the discussion in Section 3.2). See also
Remark 5.16.
Remark 5.5. The singular value decomposition is not unique. For example, an examination
of the proof of Theorem 5.1 reveals that
• any orthonormal basis for jV(A) can be used for V
2
.
there may be nonuniqueness associated with the columns of V\ (and hence U\) cor
responding to multiple cr/' s.
36 Chapter 5. Introduction to the Singular Value Decomposition
Turning now to the eigenvalue equations corresponding to the eigenvalues ar+l, ... , an we
have that A T A V
z
= VzO = 0, whence Vi A T A V
2
= O. Thus, A V
2
= O. Now define the
matrix VI E IRmxr by VI = AViSI. Then from (5.4) we see that VrVI = /; i.e., the
columns of VI are orthonormal. Choose any matrix V2 E IRmx(mr) such that [VI V2] is
orthogonal. Then
V
T
AV = [
VrAV
I
VIAV
I
=[
VrAV
I
vIA VI
Vr AV
z
]
vI AV
z
]
since A V
2
= O. Referring to the equation V I = A VI SI defining VI, we see that V r A VI =
S and vI A VI = vI VI S = O. The latter equality follows from the orthogonality of the
columns of VI and V
2
. Thus, we see that, in fact, VT A V = and defining this matrix
to be completes the proof. 0
Definition 5.2. Let A = V"i:. VT be an SVD of A as in Theorem 5.1.
1. The set {ai, ... , a
r
} is called the set of (nonzero) singular values of the matrix A and
I
is denoted From the proof of Theorem 5.1 we see that ai(A) = A;' (AT A) =
I
At (AA
T
). Note that there are also min{m, n}  r zero singular values.
2. The columns of V are called the left singular vectors of A (and are the orthonormal
eigenvectors of AA
T
).
3. The columns of V are called the right singular vectors of A (and are the orthonormal
eigenvectors of AT A).
Remark 5.3. The analogous complex case in which A E xn is quite straightforward.
The decomposition is A = V"i:. V H, where V and V are unitary and the proof is essentially
identical, except for Hermitian transposes replacing transposes.
Remark 5.4. Note that V and V can be interpreted as changes of basis in both the domain
and codomain spaces with respect to which A then has a diagonal matrix representation.
Specifically, let C denote A thought of as a linear transformation mapping IR
n
to IRm. Then
rewriting A = V"i:. VT as A V = V"i:. we see that Mat C is "i:. with respect to the bases
{VI, ... , v
n
} for IR
n
and {u I, •.. , u
m
} for IR
m
(see the discussion in Section 3.2). See also
Remark 5.16.
Remark 5.5. The !:ingular value decomposition is not unique. For example, an examination
of the proof of Theorem 5.1 reveals that
• £lny orthonormal basis for N(A) can be used for V2.
• there may be nonuniqueness associated with the columns of VI (and hence VI) cor
responding to multiple O'i'S.
5.1. The Fundamental Theorem 37
• any C/
2
can be used so long as [U\ Ui] is orthogonal.
• columns of U and V can be changed (in tandem) by sign (or multiplier of the form
e
je
in the complex case).
What is unique, however, is the matrix E and the span of the columns of U\, f/2, Vi, and
¥2 (see Theorem 5.11). Note, too, that a "full SVD" (5.2) can always be constructed from
a "compact SVD" (5.3).
Remark 5.6. Computing an SVD by working directly with the eigenproblem for A
T
A or
AA
T
is numerically poor in finiteprecision arithmetic. Better algorithms exist that work
directly on A via a sequence of orthogonal transformations; see, e.g., [7], [11], [25].
F/vamnlp 5.7.
Example 5.10. Let A e R
MX
" be symmetric and positive definite. Let V be an orthogonal
matrix of eigenvectors that diagonalizes A, i.e., V
T
AV = A > 0. Then A = VAV
T
is an
SVD of A.
A factorization t/SV
r
o f a n m x n matrix A qualifies as an SVD if U and V are
orthogonal and £ is an m x n "diagonal" matrix whose diagonal elements in the upper
left corner are positive (and ordered). For example, if A = f/E V
T
is an SVD of A, then
VS
r
C/
r
i sanSVDof A
T
.
where U is an arbitrary 2x2 orthogonal matrix, is an SVD.
Example 5.8.
where 0 is arbitrary, is an SVD.
Example 5.9.
is an SVD.
5.1. The Fundamental Theorem 37
• any U2 can be used so long as [U
I
U2] is orthogonal.
• columns of U and V can be changed (in tandem) by sign (or multiplier of the form
e
j8
in the complex case).
What is unique, however, is the matrix I: and the span of the columns of UI, U2, VI, and
V
2
(see Theorem 5.11). Note, too, that a "full SVD" (5.2) can always be constructed from
a "compact SVD" (5.3).
Remark 5.6. Computing an SVD by working directly with the eigenproblem for A T A or
AA T is numerically poor in finiteprecision arithmetic. Better algorithms exist that work
directly on A via a sequence of orthogonal transformations; see, e.g., [7], [11], [25],
Example 5.7.
A  [1 0 ]  U I U
T
 01 ,
where U is an arbitrary 2 x 2 orthogonal matrix, is an SVD.
Example 5.8.
A _ [ 1
 0  ~ ] = [
where e is arbitrary, is an SVD.
Example 5.9.
cose
 sine
sin e
cose J [ ~ ~ J [
cose
sine
A=U n=[
I 2y'5
2 ~ ][ 3 ~ 0][
3
5
2
y'5
4y'5 0 0
3 S 15
2
0
_y'5 0 0
3
3
[
I
]
3
3J2 [ ~
~ ]
=
2
3
2
3
is an SVD.
Sine]
cose '
v'2 v'2
]
T T
v'2 v'2
T
2
Example 5.10. Let A E IR
nxn
be symmetric and positive definite. Let V be an orthogonal
matrix of eigenvectors that diagonalizes A, i.e., VT A V = A > O. Then A = V A V
T
is an
SVDof A.
A factorization UI: VT of an m x n matrix A qualifies as an SVD if U and V are
orthogonal and I: is an m x n "diagonal" matrix whose diagonal elements in the upper
left comer are positive (and ordered). For example, if A = UI:V
T
is an SVD of A, then
VI:TU
T
is an SVD of AT.
38 Chapter 5. Introduction to the Singular Value Decomposition
5.2 Some Basic Properties
Theorem 5.11. Let A e R
mxn
have a singular value decomposition A = VLV
T
. Using
the notation of Theorem 5.1, the following properties hold:
1. rank(A) = r = the number of nonzero singular values of A.
2. Let U =. [ H I , ..., u
m
] and V = [v\, ..., v
n
]. Then A has the dyadic (or outer
product) expansion
Remark 5.12. Part 4 of the above theorem provides a numerically superior method for
finding (orthonormal) bases for the four fundamental subspaces compared to methods based
on, for example, reduction to row or column echelon form. Note that each subspace requires
knowledge of the rank r. The relationship to the four fundamental subspaces is summarized
nicely in Figure 5.1.
Remark 5.13. The elegance of the dyadic decomposition (5.5) as a sum of outer products
and the key vector relations (5.6) and (5.7) explain why it is conventional to write the SVD
as A = UZV
T
rather than, say, A = UZV.
Theorem 5.14. Let A e E
mx
" have a singular value decomposition A = UHV
T
as in
TheoremS.]. Then
where
3. The singular vectors satisfy the relations
38 Chapter 5. Introduction to the Singular Value Decomposition
5.2 Some Basic Properties
Theorem 5.11. Let A E jRrnxn have a singular value decomposition A = U'£ V
T
. Using
the notation of Theorem 5.1, the following properties hold:
1. rank(A) = r = the number of nonzero singular values of A.
2. Let V = [UI, ... , urn] and V = [VI, ... , v
n
]. Then A has the dyadic (or outer
product) expansion
r
A = Laiuiv;.
i=1
3. The singular vectors satisfy the relations
for i E r.
AVi = ajui,
AT Uj = aivi
(5.5)
(5.6)
(5.7)
4. LetUI = [UI, ... , u
r
], U2 = [Ur+I, ... , urn], VI = [VI, ... , v
r
], andV2 = [Vr+I, ... , V
n
].
Then
(a) R(VI) = R(A) = N(A
T
/.
(b) R(U
2
) = R(A)1 = N(A
T
).
(c) R(VI) = N(A)1 = R(A
T
).
(d) R(V2) = N(A) = R(AT)1.
Remark 5.12. Part 4 of the above theorem provides a numerically superior method for
finding (orthonormal) bases for the four fundamental subspaces compared to methods based
on, for example, reduction to row or column echelon form. Note that each subspace requires
knowledge of the rank r. The relationship to the four fundamental subspaces is summarized
nicely in Figure 5.1.
Remark 5.13. The elegance of the dyadic decomposition (5.5) as a sum of outer products
and the key vector relations (5.6) and (5.7) explain why it is conventional to write the SVD
as A = U'£V
T
rather than, say, A = U,£V.
Theorem 5.14. Let A E jRmxn have a singular value decomposition A = U,£V
T
as in
Theorem 5.1. Then
(5.8)
where
5.2. Some Basic Properties 39
Figure 5.1. SVD and the four fundamental subspaces.
with the Qsubblocks appropriately sized. Furthermore, if we let the columns of U and V
be as defined in Theorem 5.11, then
Proof: The proof follows easily by verifying the four Penrose conditions. D
Remark 5.15. Note that none of the expressions above quite qualifies as an SVD of A
+
if we insist that the singular values be ordered from largest to smallest. However, a simple
reordering accomplishes the task:
This can also be written in matrix terms by using the socalled reverseorder identity matrix
(or exchange matrix) P = \e
r
,e
r
^\, ..., e^, e\\, which is clearly orthogonal and symmetric.
5.2. Some Basic Properties 39
A
r r
E9 {O}
/ {O)<!l
nr mr
Figure 5.1. SVD and the four fundamental subspaces.
with the Osubblocks appropriately sized. Furthermore, if we let the columns of U and V
be as defined in Theorem 5.11, then
r 1
= L v;u;, (5.10)
;=1 U;
Proof' The proof follows easily by verifying the four Penrose conditions. 0
Remark 5.15. Note that none of the expressions above quite qualifies as an SVD of A +
if we insist that the singular values be ordered from largest to smallest. However, a simple
reordering accomplishes the task:
(5.11)
This can also be written in matrix terms by using the socalled reverseorder identity matrix
(or exchange matrix) P = [e
r
, erI, ... , e2, ed, which is clearly orthogonal and symmetric.
is the matrix version of (5.11). A "full SVD" can be similarly constructed.
Remark 5.16. Recall the linear transformation T used in the proof of Theorem 3.17 and
in Definition 4.1. Since T is determined by its action on a basis, and since ( v \ , . . . , v
r
} is a
basis forJ\f(A)
±
, then T can be defined by TV; = cr, w, , / e r. Similarly, since [u\, ... ,u
r
}
isabasisfor7£(.4), then T~
l
can be defined by T^' M, = ^u, , / e r. From Section 3.2, the
matrix representation for T with respect to the bases { v \ , ..., v
r
} and { MI , . . . , u
r
] is clearly
S, while the matrix representation for the inverse linear transformation T~
l
with respect to
the same bases is 5""
1
.
5.3 Row and Column Compressions
Row compression
Let A E R
mxn
have an SVD given by (5.1). Then
Notice that M(A)  M(U
T
A) = A/"(SV,
r
) and the matrix SVf e R
r x
" has full row
rank. I n other words, premultiplication of A by U
T
is an orthogonal transformation that
"compresses" A by row transformations. Such a row compression can also be accomplished
D _
by orthogonal row transformations performed directly on A to reduce it to the form
0
,
where R is upper triangular. Both compressions are analogous to the socalled rowreduced
echelon form which, when derived by a Gaussian elimination algorithm implemented in
finiteprecision arithmetic, is not generally as reliable a procedure.
Column compression
Again, let A e R
mxn
have an SVD given by (5.1). Then
This time, notice that H(A) = K(AV) = K(UiS) and the matrix UiS e R
mxr
has full
column rank. I n other words, postmultiplication of A by V is an orthogonal transformation
that "compresses" A by column transformations. Such a compression is analogous to the
40 Chapters. Introduction to the Singular Value Decomposition
Then
40 Chapter 5. Introduction to the Singular Value Decomposition
Then
A+ = (VI p)(PS1 p)(PVr)
is the matrix version of (5.11). A "full SVD" can be similarly constructed.
Remark 5.16. Recall the linear transformation T used in the proof of Theorem 3.17 and
in Definition 4.1. Since T is determined by its action on a basis, and since {VI, ... , v
r
} is a
basisforN(A).l, then T can be defined by TVj = OjUj , i E ~ . Similarly, since {UI, ... , u
r
}
is a basis forR(A), then T
I
canbedefinedbyTIu; = tv; ,i E ~ . From Section 3.2, the
matrix representation for T with respect to the bases {VI, ... , v
r
} and {u I, ... , u
r
} is clearly
S, while the matrix representation for the inverse linear transformation T
I
with respect to
the same bases is SI.
5.3 Rowand Column Compressions
Row compression
Let A E lR.
mxn
have an SVD given by (5.1). Then
VT A = :EVT
= [ ~ ~ ] [ ~ i ]
 [ SVr ] lR.
mxn
 0 E .
Notice that N(A) = N(V
T
A) = N(svr> and the matrix SVr E lR.
rxll
has full row
rank. In other words, premultiplication of A by VT is an orthogonal transformation that
"compresses" A by row transformations. Such a row compression can also be accomplished
by orthogonal row transformations performed directly on A to reduce it to the form [ ~ ] ,
where R is upper triangular. Both compressions are analogous to the socalled rowreduced
echelon form which, when derived by a Gaussian elimination algorithm implemented in
finiteprecision arithmetic, is not generally as reliable a procedure.
Column compression
Again, let A E lR.
mxn
have an SVD given by (5.1). Then
AV = V:E
= [VI U2] [ ~ ~ ]
=[VIS 0] ElR.mxn.
This time, notice that R(A) = R(A V) = R(UI S) and the matrix VI S E lR.
m
xr has full
column rank. In other words, postmultiplication of A by V is an orthogonal transformation
that "compresses" A by I;olumn transformations. Such a compression is analogous to the
Exercises 41
socalled columnreduced echelon form, which is not generally a reliable procedure when
performed by Gauss transformations in finiteprecision arithmetic. For details, see, for
example, [7], [11], [23], [25].
EXERCISES
1. Let X € M
mx
". If X
T
X = 0, show that X = 0.
2. Prove Theorem 5.1 starting from the observation that AA
T
> 0.
3. Let A e E"
xn
be symmetric but indefinite. Determine an SVD of A.
4. Let x e R
m
, y e R
n
be nonzero vectors. Determine an SVD of the matrix A e R™
defined by A = xy
T
.
6. Let A e R
mxn
and suppose W eR
mxm
and 7 e R
nxn
are orthogonal.
(a) Show that A and W A F have the same singular values (and hence the same rank).
(b) Suppose that W and Y are nonsingular but not necessarily orthogonal. Do A
and WAY have the same singular values? Do they have the same rank?
7. Let A € R"
XM
. Use the SVD to determine a polar factorization of A, i.e., A = QP
where Q is orthogonal and P = P
T
> 0. Note: this is analogous to the polar form
z = re
l&
ofa complex scalar z (where i = j = V^T).
5. Determine SVDs of the matrices
Exercises 41
socalled columnreduced echelon form, which is not generally a reliable procedure when
performed by Gauss transformations in finiteprecision arithmetic. For details, see, for
example, [7], [11], [23], [25].
EXERCISES
1. Let X E IRmxn. If XT X = 0, show that X = o.
2. Prove Theorem 5.1 starting from the observation that AAT ~ O.
3. Let A E IR
nxn
be symmetric but indefinite. Determine an SVD of A.
4. Let x E IRm, y E ~ n be nonzero vectors. Determine an SVD of the matrix A E ~ ~ xn
defined by A = xyT.
5. Determine SVDs of the matrices
(a)
[
1
]
0 1
(b)
[
~ l
6. Let A E ~ m x n and suppose W E IR
mxm
and Y E ~ n x n are orthogonal.
(a) Show that A and WAY have the same singular values (and hence the same rank).
(b) Suppose that Wand Yare nonsingular but not necessarily orthogonal. Do A
and WAY have the same singular values? Do they have the same rank?
7. Let A E ~ ~ x n . Use the SVD to determine a polar factorization of A, i.e., A = Q P
where Q is orthogonal and P = p
T
> O. Note: this is analogous to the polar form
z = re
iO
of a complex scalar z (where i = j = J=I).
This page intentionally left blank This page intentionally left blank
Chapter 6
Li near Equations
In this chapter we examine existence and uniqueness of solutions of systems of linear
equations. General linear systems of the form
are studied and include, as a special case, the familiar vector system
6.1 Vector Li near Equations
We begin with a review of some of the principal results associated with vector linear systems.
Theorem 6.1. Consider the system of linear equations
1. There exists a solution to (6.3) if and only ifbeH(A).
2. There exists a solution to (6.3} for all b e R
m
if and only ifU(A) = W", i.e., A is
onto; equivalently, there exists a solution if and only j/"rank([A, b]) = rank(A), and
this is possible only ifm < n (since m = dimT^(A) = rank(A) < min{m, n}).
3. A solution to (6.3) is unique if and only ifJ\f(A) = 0, i.e., A is 11.
4. There exists a unique solution to (6.3) for all b e W" if and only if A is nonsingular;
equivalently, A G M
mxm
and A has neither a 0 singular value nor a 0 eigenvalue.
5. There exists at most one solution to (6.3) for all b e W
1
if and only if the columns of
A are linearly independent, i.e., A/"(A) = 0, and this is possible only ifm > n.
6. There exists a nontrivial solution to the homogeneous system Ax = 0 if and only if
rank(A) < n.
43
Chapter 6
Linear Equations
In this chapter we examine existence and uniqueness of solutions of systems of linear
equations. General linear systems of the form
(6.1)
are studied and include, as a special case, the familiar vector system
Ax = b; A E ]Rn xn, b E ]Rn.
(6.2)
6.1 Vector Linear Equations
We begin with a review of some of the principal results associated with vector linear systems.
Theorem 6.1. Consider the system of linear equations
Ax = b; A E lR
m
xn, b E lRm.
(6.3)
1. There exists a solution to (6.3) if and only if b E R(A).
2. There exists a solution to (6.3) for all b E lR
m
if and only ifR(A) = lR
m
, i.e., A is
onto; equivalently, there exists a solution if and only ifrank([A, b]) = rank(A), and
this is possible only ifm :::: n (since m = dim R(A) = rank(A) :::: min{m, n n.
3. A solution to (6.3) is unique if and only if N(A) = 0, i.e., A is 11.
4. There exists a unique solution to (6.3) for all b E ]Rm if and only if A is nonsingular;
equivalently, A E lR
mxm
and A has neither a 0 singular value nor a 0 eigenvalue.
5. There exists at most one solution to (6.3) for all b E lR
m
if and only if the columns of
A are linearly independent, i.e., N(A) = 0, and this is possible only ifm ::: n.
6. There exists a nontrivial solution to the homogeneous system Ax = 0 if and only if
rank(A) < n.
43
44 Chapter 6. Linear Equations
Proof: The proofs are straightforward and can be consulted in standard texts on linear
algebra. Note that some parts of the theorem follow directly from others. For example, to
prove part 6, note that x = 0 is always a solution to the homogeneous system. Therefore, we
must have the case of a nonunique solution, i.e., A is not 11, which implies rank(A) < n
by part 3. D
6.2 Matrix Linear Equations
In this section we present some of the principal results concerning existence and uniqueness
of solutions to the general matrix linear system (6.1). Note that the results of Theorem
6.1 follow from those below for the special case k = 1, while results for (6.2) follow by
specializing even further to the case m = n.
Theorem 6.2 (Existence). The matrix linear equation
and this is clearly of the form (6.5).
has a solution if and only ifl^(B) C 7£(A); equivalently, a solution exists if and only if
AA
+
B = B.
Proof: The subspace inclusion criterion follows essentially from the definition of the range
of a matrix. The matrix criterion is Theorem 4.18.
Theorem 6.3. Let A e R
mxn
, B eR
mxk
and suppose that AA
+
B = B. Then any matrix
of the form
is a solution of
Furthermore, all solutions of (6.6) are of this form.
Proof: To verify that (6.5) is a solution, premultiply by A:
That all solutions arc of this form can be seen as follows. Let Z be an arbitrary solution of
(6.6), i.e., AZ — B. Then we can write
44 Chapter 6. Linear Equations
Proof: The proofs are straightforward and can be consulted in standard texts on linear
algebra. Note that some parts of the theorem follow directly from others. For example, to
prove part 6, note that x = 0 is always a solution to the homogeneous system. Therefore, we
must have the case of a nonunique solution, i.e., A is not II, which implies rank(A) < n
by part 3. 0
6.2 Matrix Linear Equations
In this section we present some of the principal results concerning existence and uniqueness
of solutions to the general matrix linear system (6.1). Note that the results of Theorem
6.1 follow from those below for the special case k = 1, while results for (6.2) follow by
specializing even further to the case m = n.
Theorem 6.2 (Existence). The matrix linear equation
AX = B; A E JR.
mxn
, BE JR.mxk, (6.4)
has a solution if and only ifR(B) S; R(A); equivalently, a solution exists if and only if
AA+B = B.
Proof: The subspace inclusion criterion follows essentially from the definition of the range
of a matrix. The matrix criterion is Theorem 4.18. 0
Theorem 6.3. Let A E JR.mxn, B E JR.mxk and suppose that AA + B = B. Then any matrix
of the form
X = A+ B + (/  A+ A)Y, where Y E JR.nxk is arbitrary, (6.5)
is a solution of
AX=B. (6.6)
Furthermore, all solutions of (6.6) are of this form.
Proof: To verify that (6.5) is a solution, premultiply by A:
AX = AA+ B + A(I  A+ A)Y
= B + (A  AA+ A)Y by hypothesis
= B since AA + A = A by the first Penrose condition.
That all solutions are of this form can be seen as follows. Let Z be an arbitrary solution of
(6.6). i.e .. AZ :::: B. Then we can write
Z=A+AZ+(IA+A)Z
=A+B+(IA+A)Z
and this is clearly of the form (6.5). 0
6.2. Matrix Linear Equations 45
Remark 6.4. When A is square and nonsingular, A
+
= A"
1
and so (/ — A
+
A) = 0. Thus,
there is no "arbitrary" component, leaving only the unique solution X• = A~
1
B.
Remark 6.5. It can be shown that the particular solution X = A
+
B is the solution of (6.6)
that minimizes TrX
7
X. (Tr() denotes the trace of a matrix; recall that TrX
r
X = £\ • jcj.)
Theorem 6.6 (Uniqueness). A solution of the matrix linear equation
is unique if and only if A
+
A = /; equivalently, (6.7) has a unique solution if and only if
M(A) = 0.
Proof: The first equivalence is immediate from Theorem 6.3. The second follows by noting
that A
+
A = / can occur only if r — n, where r = rank(A) (recall r < h). But rank(A) = n
if and only if A is 11 or _ /V(A) = 0. D
Example 6.7. Suppose A e E"
x
". Find all solutions of the homogeneous system Ax — 0.
Solution:
where y e R" is arbitrary. Hence, there exists a nonzero solution if and only if A
+
A /= I.
This is equivalent to either rank (A) = r < n or A being singular. Clearly, if there exists a
nonzero solution, it is not unique.
Computation: Since y is arbitrary, it is easy to see that all solutions are generated
from a basis for 7£(7 — A
+
A). But if A has an SVD given by A = f/E V
T
, then it is easily
checked that /  A+A = V
2
V
2
r
and U(V
2
V^) = K(V
2
) = N(A).
Example 6.8. Characterize all right inverses of a matrix A e ]R
mx
"; equivalently, find all
solutions R of the equation AR = I
m
. Here, we write I
m
to emphasize the m x m identity
matrix.
Solution: There exists a right inverse if and only if 7£(/
m
) c 7£(A) and this is
equivalent to AA
+
I
m
= I
m
. Clearly, this can occur if and only if rank(A) = r = m (since
r < m) and this is equivalent to A being onto (A
+
is then a right inverse). All right inverses
of A are then of the form
where Y e E"
xm
is arbitrary. There is a unique right inverse if and only if A
+
A = /
(AA(A) = 0), in which case A must be invertible and R = A"
1
.
Example 6.9. Consider the system of linear firstorder difference equations
6.2. Matrix Linear Equations 45
Remark 6.4. When A is square and nonsingular, A + = AI and so (I  A + A) = O. Thus,
there is no "arbitrary" component, leaving only the unique solution X = AI B.
Remark 6.5. It can be shown that the particular solution X = A + B is the solution of (6.6)
that minimizes TrXT X. (TrO denotes the trace of a matrix; recall that TrXT X = Li,j xlj.)
Theorem 6.6 (Uniqueness). A solution of the matrix linear equation
AX = B; A E lR,mxn, BE lR,mxk
(6,7)
is unique if and only if A + A = I; equivalently, (6.7) has a unique solution if and only if
N(A) = O.
Proof: The first equivalence is immediate from Theorem 6.3. The second follows by noting
thatA+ A = I can occur only ifr = n, wherer = rank(A) (recallr ::: n), Butrank(A) = n
if and only if A is Ilor N(A) = O. 0
Example 6.7. Suppose A E lR,nxn. Find all solutions of the homogeneous system Ax = 0,
Solution:
x=A+O+(IA+A)y
= (IA+A)y,
where y E lR,n is arbitrary. Hence, there exists a nonzero solution if and only if A + A t= I,
This is equivalent to either rank(A) = r < n or A being singular. Clearly, if there exists a
nonzero solution, it is not unique,
Computation: Since y is arbitrary, it is easy to see that all solutions are generated
from a basis for R(I  A + A). But if A has an SVD given by A = U h VT, then it is easily
checked that 1 A+ A = Vz V[ and R(Vz vD = R(Vz) = N(A),
Example 6.S. Characterize all right inverses of a matrix A E lR,mxn; equivalently, find all
solutions R of the equation AR = 1
m
, Here, we write 1m to emphasize the m x m identity
matrix,
Solution: There exists a right inverse if and only if R(Im) S; R(A) and this is
equivalent to AA + 1m = 1m. Clearly, this can occur if and only if rank(A) = r = m (since
r ::: m) and this is equivalent to A being onto (A + is then a right inverse). All right inverses
of A are then of the form
R = A+ 1m + (In  A+ A)Y
=A++(IA+A)Y,
where Y E lR,nxm is arbitrary, There is a unique right inverse if and only if A+ A I
(N(A) = 0), in which case A must be invertible and R = AI.
Example 6.9. Consider the system of linear firstorder difference equations
(6,8)
46 Chapter 6. Linear Equations
with A e R"
xn
and fieR"
xm
(rc>l,ra>l). The vector Jt* in linear system theory is
known as the state vector at time k while Uk is the input (control) vector. The general
solution of (6.8) is given by
for k > 1. We might now ask the question: Given X Q = 0, does there exist an input sequence
{uj } y~ Q such that x^ takes an arbitrary va
of reachability. Since m > 1, from the
see that (6.8) is reachable if and only if
[ Uj }
k
jj^ such that X k takes an arbitrary value in W ? In linear system theory, this is a question
of reachability. Since m > 1, from the fundamental Existence Theorem, Theorem 6.2, we
or, equivalently, if and only if
A related question is the following: Given an arbitrary initial vector X Q , does there ex
ist an input sequence {"y} "~ o such that x
n
= 0? In linear system theory, this is called
controllability. Again from Theorem 6.2, we see that (6.8) is controllable if and only if
Clearly, reachability always implies controllability and, if A is nonsingular, control
lability and reachability are equivalent. The matrices A = [ °
1
Q
1 and 5 = f ^ 1 provide an
example of a system that is controllable but not reachable.
The above are standard conditions with analogues for continuoustime models (i.e.,
linear differential equations). There are many other algebraically equivalent conditions.
Example 6.10. We now introduce an output vector y
k
to the system (6.8) of Example 6.9
by appending the equation
with C e R
pxn
and D € R
pxm
(p > 1). We can then pose some new questions about the
overall system that are dual in the systemtheoretic sense to reachability and controllability.
The answers are cast in terms that are dual in the linear algebra sense as well. The condition
dual to reachability is called observability: When does knowledge of {"
7
}"!Q and {y_ / } "~ o
suffice to determine (uniquely) Jt
0
? As a dual to controllability, we have the notion of
reconstructibility: When does knowledge of {w
y
} "~ Q and {;y/ } "Io suffice to determine
(uniquely) x
n
l The fundamental duality result from linear system theory is the following:
(A, B) is reachable [ controllable] if and only if (A
T
, B
T
] is observable [ reconstructive].
46 Chapter 6. Linear Equations
with A E IR
nx
" and B E IR
nxm
(n I, m I). The vector Xk in linear system theory is
known as the state vector at time k while Uk is the input (control) vector. The general
solution of (6.8) is given by
kJ
Xk = Akxo + LAkJj BUj
j=O
k kJ Uk2
[
UkJ ]
•...• A B]
(6.9)
(6.10)
for k 1. We might now ask the question: Given Xo = 0, does there exist an input sequence
{u j 1 such that Xk takes an arbitrary value in 1R"? In linear system theory, this is a question
of reacbability. Since m I, from the fundamental Existence Theorem, Theorem 6.2, we
see that (6.8) is reachable if and only if
R([ B, AB, ... , A
n

J
B]) = 1R"
or, equivalently, if and only if
rank [B, AB, ... , A
n

J
B] = n.
A related question is the following: Given an arbitrary initial vector Xo, does there ex
ist an input sequence {u j l'/:b such that Xn = O? In linear system theory, this is called
controllability. Again from Theorem 6.2, we see that (6.8) is controllable if and only if
Clearly, reachability always implies controllability and, if A is nonsingular, control
lability and reachability are equivalent. The matrices A = and B = provide an
example of a system that is controllable but not reachable.
The above are standard conditions with analogues for continuoustime models (i.e.,
linear differential equations). There are many other algebraically equivalent conditions.
Example 6.10. We now introduce an output vector Yk to the system (6.8) of Example 6.9
by appending the equation
(6.11)
with C E IR
Pxn
and D E IR
Pxm
(p 1). We can then pose some new questions about the
overall system that are dual in the systemtheoretic sense to reachability and controllability.
The answers are cast in terms that are dual in the linear algebra sense as well. The condition
dual to reachability is called observability: When does knowledge of {u j r/:b and {Yj l';:b
suffice to determine (uniquely) xo? As a dual to controllability, we have the notion of
reconstructibility: When does knowledge of {u j r/:b and {YJ lj:b suffice to determine
(uniquely) xn? The fundamental duality result from linear system theory is the following:
(A. B) iJ reachable [controllablcl if and only if (A T. B T) is observable [reconsrrucrible]
6.4 Some Us ef u l and I nt er es t i ng Inverses 47
To derive a condition for observability, notice that
Thus,
Let v denote the (known) vector on the lefthand side of (6.13) and let R denote the matrix on
the righthand side. Then, by definition, v e Tl(R), so a solution exists. By the fundamental
Uniqueness Theorem, Theorem 6.6, the solution is then unique if and only if N(R) = 0,
or, equivalently, if and only if
6.3 A More General Matrix Linear Equation
Theorem 6.11. Let A e R
mxn
, B e R
mxq
, and C e R
pxti
. Then the equation
has a solution if and only if AA
+
BC
+
C = B, in which case the general solution is of the
where Y € R
n
*
p
is arbitrary.
A compact matrix criterion for uniqueness of solutions to (6.14) requires the notion
of the Kronecker product of matrices for its statement. Such a criterion (CC
+
< g) A
+
A — I)
is stated and proved in Theorem 13.27.
6.4 Some Useful and Interesting Inverses
In many applications, the coefficient matrices of interest are square and nonsingular. Listed
below is a small collection of useful matrix identities, particularly for block matrices, as
sociated with matrix inverses. In these identities, A e R
nxn
, B E R
nxm
, C e R
mxn
,
and D € E
mxm
. Invertibility is assumed for any component or subblock whose inverse is
indicated. Verification of each identity is recommended as an exercise for the reader.
6.4 Some Useful and Interesting Inverses
Thus,
To derive a condition for observability, notice that
kl
Yk = CAkxo + L CAk1j BUj + DUk.
j=O
r
Yo  Duo
Yl  CBuo  Du]
Yn]  L j : ~ CA
n

2
j BUj  DUnl
47
(6.12)
(6.13)
Let v denote the (known) vector on the lefthand side of (6.13) and let R denote the matrix on
the righthand side. Then, by definition, v E R(R), so a solution exists. By the fundamental
Uniqueness Theorem, Theorem 6.6, the solution is then unique if and only if N(R) = 0,
or, equivalently, if and only if
6.3 A More General Matrix Linear Equation
Theorem 6.11. Let A E jRmxn, B E jRmx
q
, and C E jRpxq. Then the equation
AXC=B (6.14)
has a solution if and only if AA + BC+C = B, in which case the general solution is of the
form
(6.15)
where Y E jRnxp is arbitrary.
A compact matrix criterion for uniqueness of solutions to (6.14) requires the notion
of the Kronecker product of matrices for its statement. Such a criterion (C C+ ® A + A = I)
is stated and proved in Theorem 13.27.
6.4 Some Useful and Interesting Inverses
In many applications, the coefficient matrices of interest are square and nonsingular. Listed
below is a small collection of useful matrix identities, particularly for block matrices, as
sociated with matrix inverses. In these identities, A E jRnxn, B E jRnxm, C E jRmxn,
and D E jRm xm. Invertibility is assumed for any component or subblock whose inverse is
indicated. Verification of each identity is recommended as an exercise for the reader.
48 Chapter 6. Linear Equations
1. (A + BDCr
1
= A~
l
 A~
l
B(D~
l
+ CA~
l
B)~
[
CA~
l
.
This result is known as the ShermanMorrisonWoodbury formula. It has many
applications (and is frequently "rediscovered") including, for example, formulas for
the inverse of a sum of matrices such as (A + D)"
1
or (A"
1
+ D"
1
) . It also
yields very efficient "updating" or "downdating" formulas in expressions such as
T — 1
(A + JUT ) (with symmetric A e R"
x
" and ;c e E") that arise in optimization
theory.
EXERCISES
1. As in Example 6.8, characterize all left inverses of a matrix A e M
mx
".
2. Let A € E
mx
", B e R
mxk
and suppose A has an SVD as in Theorem 5.1. Assuming
7Z(B) c 7£(A), characterize all solutions of the matrix linear equation
Both of these matrices satisfy the matrix equation X^ = I from which it is obvious
that X~
l
= X. Note that the positions of the / and — / blocks may be exchanged.
where E = (D — CA B) (E is the inverse of the Schur complement of A). This
result follows easily from the block LU factorization in property 16 of Section 1.4.
where F = (A — ED C) . This result follows easily from the block UL factor
ization in property 17 of Section 1.4.
in terms of the SVD of A
48 Chapter 6. Linear Equations
1. (A + BDC)I = AI  AIB(D
I
+ CAIB)ICAI.
This result is known as the ShermanMorrisonWoodbury formula. It has many
applications (and is frequently "rediscovered") including, for example, formulas for
the inverse of a sum of matrices such as (A + D)lor (AI + DI)I. It also
yields very efficient "updating" or "downdating" formulas in expressions such as
(A + xx
T
) I (with symmetric A E lR
nxn
and x E lRn) that arise in optimization
theory.
2. r
l
= [
3. !/ r
l
= l r
l
= 1
Both of these matrices satisfy the matrix equation X2 = / from which it is obvious
that XI = X. Note that the positions of the / and  / blocks may be exchanged.
4. r
l
= [
AI BD
I
]
D I .
5. r
l
= 1
6. [ / +c
BC
r
l
= [!C / 1
7. r
l
= [ AI l
where E = (D  CA
I
B)I (E is the inverse of the Schur complement of A). This
result follows easily from the block LU factorization in property 16 of Section 1.4.
8. r
l
= D
I
l
where F = (A  B D
I
C) I. This result follows easily from the block UL factor
ization in property 17 of Section 1.4.
EXERCISES
1. As in Example 6.8, characterize all left inverses of a matrix A E lR
m
xn .
2. Let A E lRmxn, B E lR
fflxk
and suppose A has an SVD as in Theorem 5.1. Assuming
R(B) R(A), characterize all solutions of the matrix linear equation
AX=B
in terms of the SVD of A.
Exercises 49
3. Let jc, y e E" and suppose further that X
T
y ^ 1. Show that
4. Let x, y € E" and suppose further that X
T
y ^ 1. Show that
where c = 1/(1 — x
T
y).
5. Let A e R"
x
" and let A"
1
have columns c\, ..., c
n
and individual elements y
;y
.
Assume that x/
(
7^ 0 for some / and j. Show that the matrix B — A —
l
—ei e
T
: (i.e.,
A with — subtracted from its (zy)th element) is singular.
Hint: Show that c
t
< = M(B).
6. As in Example 6.10, check directly that the condition for reconstructibility takes the
form
Exercises 49
3. Let x, y E IR
n
and suppose further that x T y i= 1. Show that
T 1 1 T
(/  xy) = I  xy .
xTy 1
4. Let x, y E IR
n
and suppose further that x T y i= 1. Show that
cxJ
C '
where C = 1/(1  x
T
y).
5. Let A E 1 R ~ xn and let A 1 have columns Cl, ... ,C
n
and individual elements Yij.
Assume that Yji i= 0 for some i and j. Show that the matrix B = A  ~ i e;e; (i.e.,
A with yl subtracted from its (ij)th element) is singular.
l'
Hint: Show that Ci E N(B).
6. As in Example 6.10, check directly that the condition for reconstructibility takes the
form
N[
fA J ~ N(A
n
).
CA
n

1
This page intentionally left blank This page intentionally left blank
Chapter 7
Projections, Inner Product
Spaces, and Norms
7.1 Projections
Definition 7.1. Let V be a vector space with V = X 0 y. By Theorem 2.26, every v e V
has a unique decomposition v = x + y with x e X and y e y. Define PX y • V — > • X c V
by
Figure 7.1. Oblique projections.
Theorem 7.2. Px,y is linear and P# y — Px,y
Theorem 7.3. A linear transformation P is a projection if and only if it is idempotent, i.e.,
P
2
= P. Also, P is a projection if and only if I —P is a projection. Infact, Py,x — I — Px,y
Proof: Suppose P is a projection, say on X along y (using the notation of Definition 7.1).
51
Px,y is called the (oblique) projection on X along 3^.
Figure 7.1 displays the projection of v on both X and 3^ in the case V =
Chapter 7
Projections, Inner Product
Spaces, and Norms
7.1 Projections
Definition 7.1. Let V be a vector space with V = X EEl Y. By Theorem 2.26, every v E V
has a unique decomposition v = x + y with x E X and y E y. Define pX,y : V + X <; V
by
PX,yV = x for all v E V.
PX,y is called the (oblique) projection on X along y.
Figure 7.1 displays the projection of von both X and Y in the case V = ]R2.
y
x
Figure 7.1. Oblique projections.
Theorem 7.2. px.y is linear and pl.
y
= px.y.
Theorem 7.3. A linear transformation P is a projection if and only if it is idempotent, i.e.,
p2 = P. Also, P isaprojectionifandonlyifl P isaprojection. Infact, Py.x = I px.y.
Proof: Suppose P is a projection, say on X along Y (using the notation of Definition 7.1).
51
52 Chapter 7. Projections, Inner Product Spaces, and Norms
Let u e V be arbitrary. Then Pv = P(x + y) = Px = x. Moreover, P
2
v = PPv —
Px = x = Pv. Thus, P
2
= P. Conversely, suppose P
2
= P. Let X = {v e V : Pv = v}
and y = {v € V : Pv = 0}. It is easy to check that X and 3^ are subspaces. We now prove
that V = X 0 y. First note that tfveX, then Pv = v. If v e y, then Pv = 0. Hence
i f v € X n y, then v = 0. Now let u e V be arbitrary. Then v = Pv + (I  P)v. Let
x = Pv, y = (I  P)v. Then Px = P
2
v = Pv = x so x e X, while Py = P(I  P}v =
Pv  P
2
v = 0 so y e y. Thus, V = X 0 y and the projection on X along y is P.
Essentially the same argument shows that / — P is the projection on y along X. D
Definition 7.4. In the speci al case where y = X^, PX.X
L
*
s
called an orthogonal projec
tion and we then use the notati on PX = PX,X
L

Theorem 7.5. P e E"
xn
i s the matri x of an orthogonal projecti on (onto K(P)} i f and only
i fP
2
= p = P
T
.
Proof: Let P be an orthogonal projection (on X, say, along X
L
} and let jc, y e R" be
arbitrary. Note that (/  P)x = (I  PX,X^X = P
x
±,
x
x by Theorem 7.3. Thus,
(/  P)x e X
L
. Since Py e X, we have ( P y f ( I  P)x = y
T
P
T
(I  P)x = 0.
Since x and y were arbitrary, we must have P
T
(I — P) = 0. Hence P
T
= P
T
P = P,
with the second equality following since P
T
P is symmetric. Conversely, suppose P is a
symmetric projection matrix and let x be arbitrary. Write x = Px + (I — P)x. Then
x
T
P
T
(I  P)x = x
T
P(I  P}x = 0. Thus, since Px e U(P), then (/  P)x 6 ft(P)
1
and P must be an orthogonal projection. D
7.1.1 The four fundamental orthogonal projections
Using the notation of Theorems 5.1 and 5.11, let A 6 R
mxn
with SVD A = UT,V
T
=
UtSVf. Then
are easily checked to be (unique) orthogonal projections onto the respective four funda
mental subspaces,
52 Chapter 7. Projections, Inner Product Spaces, and Norms
Let v E V be arbitrary. Then Pv = P(x + y) = Px = x. Moreover, p
2
v = P Pv =
Px = x = Pv. Thus, p2 = P. Conversely, suppose p2 = P. Let X = {v E V : Pv = v}
and Y = {v E V : Pv = OJ. It is easy to check that X and Y are subspaces. We now prove
that V = X $ y. First note that if v E X, then Pv = v. If v E Y, then Pv = O. Hence
if v E X ny, then v = O. Now let v E V be arbitrary. Then v = Pv + (I  P)v. Let
x = Pv, y = (I  P)v. Then Px = p
2
v = Pv = x so x E X, while Py = P(l  P)v =
Pv  p
2
v = 0 so Y E y. Thus, V = X $ Y and the projection on X along Y is P.
Essentially the same argument shows that I  P is the projection on Y along X. 0
Definition 7.4. In the special case where Y = X1, px.xl. is called an orthogonal projec
tion and we then use the notation P
x
= PX.XL
Theorem 7.5. P E jRnxn is the matrix of an orthogonal projection (onto R(P)) if and only
if p2 = P = pT.
Proof: Let P be an orthogonal projection (on X, say, along X 1) and let x, y E jR" be
arbitrary. Note that (I  P)x = (I  px.xJ.)x = PXJ..xx by Theorem 7.3. Thus,
(I  P)x E X1. Since Py E X, we have (py)T (I  P)x = yT pT (I  P)x = O.
Since x and y were arbitrary, we must have pT (I  P) = O. Hence pT = pT P = P,
with the second equality following since pT P is symmetric. Conversely, suppose P is a
symmetric projection matrix and let x be arbitrary. Write x = P x + (I  P)x. Then
x
T
pT (I  P)x = x
T
P(l  P)x = O. Thus, since Px E R(P), then (I  P)x E R(P)1
and P must be an orthogonal projection. 0
7.1 .1 The four fundamental orthogonal projections
Using the notation of Theorems 5.1 and 5.11, let A E jRmxII with SVD A = U!:V
T
U\SVr Then
r
PR(A)
AA+
U\U[
Lu;uT,
;=1
m
PR(A).L
1 AA+
U2
U
! LUiUT,
i=r+l
11
PN(A)
1 A+A
V2V{
L ViVf,
i=r+l
r
PN(A)J.
A+A
VIV{
LViVT
i=l
are easily checked to be (unique) orthogonal projections onto the respective four funda
mental subspaces.
7.1. Projections 53
Example 7.6. Determine the orthogonal projection of a vector v e M" on another nonzero
vector w e R
n
.
Solution: Think of the vector w as an element of the onedimensional subspace IZ(w).
Then the desired projection is simply
(using Example 4.8)
Moreover, the vector z that is orthogonal to w and such that v = Pv + z is given by
z = PK(
W
)±V = (/ — PK(W))V = v — (^^ j w. See Figure 7.2. A direct calculation shows
that z and u; are, in fact, orthogonal:
Figure 7.2. Orthogonal projection on a "line."
Example 7.7. Recall the proof of Theorem 3.11. There, {v\ , ..., Vk} was an orthornormal
basis for a subset S of W
1
. An arbitrary vector x e R" was chosen and a formula for x\
appeared rather mysteriously. The expression for x\ is simply the orthogonal projection of
x on S. Specifically,
Example 7.8. Recall the diagram of the four fundamental subspaces. The indicated direct
sum decompositions of the domain E" and codomain R
m
are given easily as follows.
Let x e W
1
be an arbitrary vector. Then
7.1. Projections 53
Example 7.6. Determine the orthogonal projection of a vector v E IR
n
on another nonzero
vector w E IRn.
Solution: Think of the vector w as an element of the onedimensional subspace R( w).
Then the desired projection is simply
Pn(w)v = ww+v
wwTv
(using Example 4.8)
= (WTV)
T W.
W W
Moreover, the vector z that is orthogonal to wand such that v = P v + z is given by
z = Pn(w)"' v = (l  Pn(w»v = v  ( : ; ~ ) w. See Figure 7.2. A direct calculation shows
that z and ware, in fact, orthogonal:
v
z
Pv w
Figure 7.2. Orthogonal projection on a "line."
Example 7.7. Recall the proof of Theorem 3.11. There, {VI, ... , Vk} was an orthomormal
basis for a subset S of IRn. An arbitrary vector x E IR
n
was chosen and a formula for XI
appeared rather mysteriously. The expression for XI is simply the orthogonal projection of
X on S. Specifically,
Example 7.8. Recall the diagram of the four fundamental subspaces. The indicated direct
sum decompositions of the domain IR
n
and codomain IR
m
are given easily as follows.
Let X E IR
n
be an arbitrary vector. Then
X = PN(A)u + PN(A)X
= A+ Ax + (I  A+ A)x
= VI vt x + V
2
Vi x (recall VVT = I).
54 Chapter 7. Projections, Inner Product Spaces, and Norms
Similarly, let y e ]R
m
be an arbitrary vector. Then
Example 7.9. Let
Then
and we can decompose the vector [2 3 4]
r
uniquely into the sum of a vector in A/' CA)
1
and a vector in J\f(A), respectively, as follows:
7.2 Inner Product Spaces
Definition 7.10. Let V be a vector space over R. Then { • , • ) : V x V
product if
is a real inner
1. (x, x) > Qfor all x 6V and ( x , x } =0 if and only ifx = 0.
2. (x, y) = (y,x)forallx,y e V.
3. { *, cryi + ^2) = a(x, y\) + / 3( j t, y^} for all jc, yi, j2 ^ V and for alia, ft e R.
Example 7.11. Let V = R". Then { ^, y} = X
T
y is the "usual" Euclidean inner product or
dot product.
Example 7.12. Let V = E". Then ( j c, y)
Q
= X
T
Qy, where Q = Q
T
> 0 is an arbitrary
n x n positive definite matrix, defines a "weighted" inner product.
Definition 7.13. If A e R
mx
", then A
T
e R
nxm
is the unique linear transformation or map
such that (x, Ay)  (A
T
x, y) for all x € R
m
and for all y e R".
54 Chapter 7. Projections, Inner Product Spaces, and Norms
Similarly, let Y E IR
m
be an arbitrary vector. Then
Y = PR(A)Y +
= AA+y + (l AA+)y
= U1Ur y + U2U[ Y (recall UU
T
= I).
Example 7.9. Let
Then
1/4
1/4
o
1/4 ]
1/4
o
and we can decompose the vector [2 3 4V uniquely into the sum of a vector in N(A)L
and a vector in N(A), respectively, as follows:
[ ! ] A' Ax + (l  A' A)x
[
1/2 1/2 0] [ 2] [
= ! +
[
5/2] [1/2]
= + .
7.2 Inner Product Spaces
1/2
1/2
o
1/2
1/2
o
Definition 7.10. Let V be a vector space over IR. Then (', .) : V x V + IR is a real inner
product if
1. (x, x) ::: Of or aU x E V and (x, x) = 0 if and only ifx = O.
2. (x, y) = (y, x) for all x, y E V.
3. (x, aYI + PY2) = a(x, Yl) + f3(x, Y2) for all x, Yl, Y2 E V and/or all a, f3 E IR.
Example 7.11. Let V = IRn. Then (x, y) = x
T
Y is the "usual" Euclidean inner product or
dot product.
Example 7.12. Let V = IRn. Then (x, y) Q = X T Qy, where Q = Q T > 0 is an arbitrary
n x n positive definite matrix, defines a "weighted" inner product.
Definition 7.13. If A E IR
m
xn, then ATE IR
n
xm is the unique linear transformation or map
such that {x, Ay) = {AT x, y) for all x E IR
m
andfor all y e IRn.
7.2. Inner Product Spaces 55
It is easy to check that, with this more "abstract" definition of transpose, and if the
(/, y)th element of A is a
(;
, then the (i, y)t h element of A
T
is a/ , . It can also be checked
that all the usual properties of the transpose hold, such as (Afl) = B
T
A
T
. However, the
definition above allows us to extend the concept of transpose to the case of weighted inner
products in the following way. Suppose A e R
mxn
and let {, }g and (•, }
R
, with Q and
R positive definite, be weighted inner products on R
m
and W, respectively. Then we can
define the "weighted transpose" A
#
as the unique map that satisfies
(x, Ay)
Q
= (A
#
x, y)
R
for all x e R
m
and for all y e W
1
.
By Example 7.12 above, we must then have X
T
QAy = x
T
(A
#
) Ry for all x, y. Hence we
must have QA = (A
#
) R. Taking transposes (of the usual variety) gives A
T
Q = RA
#
.
Since R is nonsingular, we find
A* = /r'A' Q.
We can also generalize the notion of orthogonality (x
T
y = 0) to Q orthogonality (Q is
a positive definite matrix). Two vectors x, y e W are <2orthogonal (or conjugate with
respect to Q) if ( x, y}
Q
= X
T
Qy = 0. Q orthogonality is an important tool used in
studying conjugate direction methods in optimization theory.
Definition 7.14. Let V be a vector space over C. Then {, •} : V x V > C is a complex
inner product if
1. ( x, x) > Qfor all x e V and ( x, x) =0 if and only ifx = 0.
2. (x, y) = (y, x) for all x, y e V.
3. (x,ayi + fiy
2
) = a(x, y\) + fi(x, y
2
}forallx, y\, y
2
e V and for alia, ft 6 C.
Remark 7.15. We could use the notation {•, }
c
to denote a complex inner product, but
if the vectors involved are complexvalued, the complex inner product is to be understood.
Note, too, from part 2 of the definition, that ( x, x) must be real for all x.
Remark 7.16. Note from parts 2 and 3 of Definition 7.14 that we have
(ax\ + fix
2
, y) = a(x\, y) + P(x
2
, y}.
Remark 7.17. The Euclidean inner product of x, y e C" is given by
The conventional definition of the complex Euclidean inner product is (x, y} = y
H
x but we
use its complex conjugate x
H
y here for symmetry with the real case.
Remark 7.18. A weighted inner product can be defined as in the real case by (x, y}
Q
—
X
H
Qy, for arbitrary Q = Q
H
> 0. The notion of Q orthogonality can be similarly
generalized to the complex case.
7.2. Inner product Spaces 55
It is easy to check that, with this more "abstract" definition of transpose, and if the
(i, j)th element of A is aij, then the (i, j)th element of AT is ap. It can also be checked
that all the usual properties of the transpose hold, such as (AB) = BT AT. However, the
definition above allows us to extend the concept of transpose to the case of weighted inner
products in the following way. Suppose A E ]Rm xn and let (., .) Q and (., .) R, with Q and
R positive definite, be weighted inner products on IR
m
and IRn, respectively. Then we can
define the "weighted transpose" A # as the unique map that satisfies
(x, AY)Q = (A#x, Y)R for all x E IRm and for all Y E IRn.
By Example 7.l2 above, we must then have x
T
QAy = x
T
(A#{ Ry for all x, y. Hence we
must have QA = (A#{ R. Taking transposes (of the usual variety) gives AT Q = RA#.
Since R is nonsingular, we find
A# = R1A
T
Q.
We can also generalize the notion of orthogonality (x
T
y = 0) to Qorthogonality (Q is
a positive definite matrix). Two vectors x, y E IRn are Qorthogonal (or conjugate with
respect to Q) if (x, y) Q = X T Qy = O. Qorthogonality is an important tool used in
studying conjugate direction methods in optimization theory.
Definition 7.14. Let V be a vector space over <C. Then (., .) : V x V + C is a complex
inner product if
1. (x, x) :::: 0 for all x E V and (x, x) = 0 if and only if x = O.
2. (x, y) = (y, x) for all x, y E V.
3. (x, aYI + f3Y2) = a(x, yll + f3(x, Y2) for all x, YI, Y2 E V andfor all a, f3 E c.
Remark 7.15. We could use the notation (., ·)e to denote a complex inner product, but
if the vectors involved are complexvalued, the complex inner product is to be understood.
Note, too, from part 2 of the definition, that (x, x) must be real for all x.
Remark 7.16. Note from parts 2 and 3 of Definition 7.14 that we have
Remark 7.17. The Euclidean inner product of x, y E C
n
is given by
n
(x, y) = LXiYi = xHy.
i=1
The conventional definition of the complex Euclidean inner product is (x, y) = yH x but we
use its complex conjugate x
H
y here for symmetry with the real case.
Remark 7.1S. A weighted inner product can be defined as in the real case by (x, y)Q =
x
H
Qy, for arbitrary Q = QH > o. The notion of Qorthogonality can be similarly
generalized to the complex case.
56 Chapter 7. Projections, Inner Product Spaces, and Norms
Definition 7.19. A vector space (V, F) endowed with a specific inner product is called an
inner product space. If F = C, we call V a complex inner product space. If F = R, we
call V a real inner product space.
Example 7.20.
1. Check that V = R"
x
" with the inner product (A, B) = Tr A
T
B is a real inner product
space. Note that other choices are possible since by properties of the trace function,
TrA
T
B = TrB
T
A = TrAB
T
= TrBA
T
.
2. Check that V = C
nx
" with the inner product (A, B) = Tr A
H
B is a complex inner
product space. Again, other choices are possible.
Definition 7.21. Let V be an inner product space. For v e V, we define the norm (or
length) ofv by \\v\\ = */(v, v). This is called the norm induced by (  ,  ) .
Example 7.22.
1. If V = E." with the usual inner product, the induced norm is given by   i>   =
xV—*« 9\ 7
( E , =i < Y )
2

2. If V = C" with the usual inner product, the induced norm is given by \\v\\ =
(£ ?
=
, l » ,  l
2
)* .
Theorem 7.23. Let P be an orthogonal projection on an inner product space V. Then
\\Pv\\ < \\v\\forallv e V.
Proof: Since P is an orthogonal projection, P
2
= P = P
#
. (Here, the notation P
#
denotes
the unique linear transformation that satisfies ( P u , v } = (u, P
#
v) for all u, v e V. If this
seems a little too abstract, consider V = R" (or C"), where P
#
is simply the usual P
T
(or
P
H
)). Hence ( P v , v) = (P
2
v, v) = (Pv, P
#
v) = ( P v , Pv) = \\Pv\\
2
> 0. Now /  P is
also a projection, so the above result applies and we get
from which the theorem follows.
Definition 7.24. The norm induced on an inner product space by the "usual" inner product
is called the natural norm.
In case V = C" or V = R", the natural norm is also called the Euclidean norm. In
the next section, other norms on these vector spaces are defined. A converse to the above
procedure is also available. That is, given a norm defined by \\x\\ — • > /(• * > x), an inner
product can be defined via the following.
56 Chapter 7. Projections, Inner Product Spaces, and Norms
Definition 7.19. A vector space (V, IF) endowed with a specific inner product is called an
inner product space. If IF = e, we call V a complex inner product space. If IF = R we
call V a real inner product space.
Example 7.20.
1. Check that V = IR
n
xn with the inner product (A, B) = Tr AT B is a real inner product
space. Note that other choices are possible since by properties of the trace function,
Tr AT B = Tr B T A = Tr A B T = Tr BAT.
2. Check that V = e
nxn
with the inner product (A, B) = Tr AH B is a complex inner
product space. Again, other choices are possible.
Definition 7.21. Let V be an inner product space. For v E V, we define the norm (or
length) ofv by IIvll = J(V,V). This is called the norm induced by (', .).
Example 7.22.
1. If V = IR
n
with the usual inner product, the induced norm is given by II v II
n 2 1
(Li=l V
i
)2.
2. If V = en with the usual inner product, the induced norm is given by II v II =
"n 2 !
(L...i=l IVi I ) .
Theorem 7.23. Let P be an orthogonal projection on an inner product space V. Then
IIPvll ::::: Ilvll for all v E V.
Proof: Since P is an orthogonal projection, p2 = P = pH. (Here, the notation p# denotes
the unique linear transformation that satisfies (Pu, v) = (u, p#v) for all u, v E V. If this
seems a little too abstract, consider V = IR
n
(or en), where p# is simply the usual pT (or
pH)). Hence (Pv, v) = (P
2
v, v) = (Pv, p#v) = (Pv, Pv) = IIPvll
2
::: O. Now / P is
also a projection, so the above result applies and we get
0::::: ((I  P)v. v) = (v. v)  (Pv, v)
= IIvll2  IIPvll
2
from which the theorem follows. 0
Definition 7.24. The norm induced on an inner product space by the "usual" inner product
is called the natural norm.
In case V = en or V = IR
n
, the natural norm is also called the Euclidean norm. In
the next section, other norms on these vector spaces are defined. A converse to the above
procedure is also available. That is, given a norm defined by IIx II = .j(X,X}, an inner
product can be defined via the following.
7.3. Vector Norms 57
Theorem 7.25 (Polarization Identity).
1. For x, y € R", an inner product is defined by
7.3 Vector Norms
Definition 7.26. Let (V, F) be a vector space. Then \ \  \ \ : V >• R is a vector norm if it
satisfies the following three properties:
2. For x, y e C", an inner product is defined by
where j = i = \/—T.
(This is called the triangle inequality, as seen readily from the usual diagram illus
trating the sum of two vectors in R
2
.)
Remark 7.27. It is convenient in the remainder of this section to state results for complex
valued vectors. The specialization to the real case is obvious.
Definition 7.28. A vector space (V, F) is said to be a normed linear space if and only if
there exists a vector norm  •  : V > R satisfying the three conditions of Definition 7.26.
Example 7.29.
1. For x e C", the Holder norms, or pnorms, are defined by
Special cases:
(The second equality is a theorem that requires proof.)
7.3. Vector Norms
Theorem 7.25 (Polarization Identity).
1. For x, y E an inner product is defined by
(x,y)=xTy=
2. For x, y E en, an inner product is defined by
where j = i = .J=I.
7.3 Vector Norms
IIx + yll2 _ IIxll2 _ lIyll2
2
57
Definition 7.26. Let (V, IF) be a vector space. Then II . II : V + IR is a vector norm ifit
satisfies the following three properties:
1. Ilxll::: Of or all x E V and IIxll = 0 ifand only ifx = O.
2. Ilaxll = lalllxllforallx E Vandforalla E IF.
3. IIx + yll :::: IIxll + IIYliforall x, y E V.
(This is called the triangle inequality, as seen readily from the usual diagram illus
trating the sum of two vectors in ]R2 .)
Remark 7.27. It is convenient in the remainder of this section to state results for complex
valued vectors. The specialization to the real case is obvious.
Definition 7.28. A vector space (V, IF) is said to be a normed linear space if and only if
there exists a vector norm II . II : V + ]R satisfying the three conditions of Definition 7.26.
Example 7.29.
1. For x E en, the HOlder norms, or pnorms, are defined by
Special cases:
(a) Ilx III = L:7=1 IXi I (the "Manhattan" norm).
1 1
(b) Ilxllz = (L:7=1Ix;l2)2 = (X
H
X)2 (the Euclidean norm).
(c) Ilxlioo = maxlx;l = lim IIxllp
IE!! p++oo
(The second equality is a theorem that requires proof.)
58 Chapter 7. Projections, Inner Product Spaces, and Norms
2. Some weighted pnorms:
(a)   JC  , .
D
= E^rf/l*/!, where 4 > 0.
(b) I k llz . g — (x
h
Q
X
Y > where Q = Q
H
> 0 (this norm is more commonly
denoted  • 
c
).
3. On the vector space (C[to, t \ ] , R), define the vector norm
On the vector space ((C[to, t\])
n
, R), define the vector norm
Fhcorem 7.30 (Holder Inequality). Let x, y e C". Ther,
A particular case of the Holder inequality is of special interest.
Theorem 7.31 (CauchyBunyakovskySchwarz Inequality). Let x, y e C". Then
with equality if and only if x and y are linearly dependent.
Proof: Consider the matrix [x y] e C"
x2
. Since
is a nonnegative definite matrix, its determinant must be nonnegative. In other words,
0 < ( x
H
x ) ( y
H
y ) — ( x
H
y ) ( y
H
x ) . Since y
H
x = x
H
y, we see immediately that \X
H
y\ <
\\X\\2\\y\\2
D
Note: This is not the classical algebraic proof of the CauchyBunyakovskySchwarz
(CBS) inequality (see, e.g., [20, p. 217]). However, it is particularly easy to remember.
Remark 7.32. The angle 0 between two nonzero vectors x, y e C" may be defined by
cos# = I, „ .^ , 0 < 0 < 5. The CBS inequality is thus equivalent to the statement
IlMmlylb — ^
COS 0 <1.
Remark 7.33. Theorem 7.31 and Remark 7.32 are true for general inner product spaces.
Remark 7.34. The norm  • 
2
is unitarily invariant, i.e., if U € C"
x
" is unitary, then
\\Ux\\
2
= \\x\\
2
(Proof. \\Ux\\l = x
H
U
H
Ux = X
H
X = \\x\\\). However,   , and   1^
58 Chapter 7. Projections, Inner Product Spaces, and Norms
2. Some weighted pnorms:
(a) IIxll1.D = whered; > O.
1
(b) IIx IIz.Q = (x
H
Qx) 2, where Q = QH > 0 (this norm is more commonly
denoted II . IIQ)'
3. On the vector space (C[to, ttl, 1Ft), define the vector norm
11111 = max 1/(t)I·
On the vector space «e[to, ttlr, 1Ft), define the vector norm
1111100 = max II/(t) 11
00
,
Theorem 7.30 (HOlder Inequality). Let x, y E en. Then
I I
+=1.
p q
A particular case of the HOlder inequality is of special interest.
Theorem 7.31 (CauchyBunyakovskySchwarz Inequality). Let x, y E en. Then
with equality if and only if x and yare linearly dependent.
Proof' Consider the matrix [x y] E en
x2
. Since
is a nonnegative definite matrix, its determinant must be nonnegative. In other words,
o (x
H
x)(yH y)  (x
H
y)(yH x). Since yH x = x
H
y, we see immediately that IXH yl
IIxll2l1yllz. 0
Note: This is not the classical algebraic proof of the CauchyBunyakovskySchwarz
(CBS) inequality (see, e.g., [20, p. 217]). However, it is particularly easy to remember.
Remark 7.32. The angle e between two nonzero vectors x, y E en may be defined by
cos e = 0 e I' The CBS inequality is thus equivalent to the statement
1 cose 1 1.
Remark 7.33. Theorem 7.31 and Remark 7.32 are true for general inner product spaces.
Remark 7.34. The norm II . 112 is unitarily invariant, i.e., if U E e
nxn
is unitary, then
IIUxll2 = IIxll2 (Proof IIUxili = XHUHUx = xHx = IIxlli)· However, 11·111 and 1I·IIClO
7.4. Matrix Norms 59
are not unitarily invariant. Similar remarks apply to the unitary invariance of norms of real
vectors under orthogonal transformation.
Remark 7.35. If x, y € C" are orthogonal, then we have the Pythagorean Identity
7.4 Matrix Norms
In this section we introduce the concept of matrix norm. As with vectors, the motivation for
using matrix norms is to have a notion of either the size of or the nearness of matrices. The
former notion is useful for perturbation analysis, while the latter is needed to make sense of
"convergence" of matrices. Attention is confined to the vector space (W
nxn
, R) since that is
what arises in the majority of applications. Extension to the complex case is straightforward
and essentially obvious.
Definition 7.39.  •  : R
mx
" > E is a matrix norm if it satisfies the following three
properties:
2 _ _/ / .
the proof of which follows easily from z2 = z z.
Theorem 7.36. All norms on C" are equivalent; i.e., there exist constants c\, ci (possibly
depending onn) such that
Example 7.37. For x G C", the following inequalities are all tight bounds; i.e., there exist
vectors x for which equality holds:
Finally, we conclude this section with a theorem about convergence of vectors. Con
vergence of a sequence of vectors to some limit vector can be converted into a statement
about convergence of real numbers, i.e., convergence in terms of vector norms.
Theorem 7.38. Let \\ • \\ be a vector norm and suppose v, i»
( 1 )
, v
(2
\ ... e C". Then
7.4. Matrix Norms 59
are not unitarily invariant. Similar remarks apply to the unitary invariance of norms of real
vectors under orthogonal transformation.
Remark 7.35. If x, y E en are orthogonal, then we have the Pythagorean Identity
Ilx ± = +
the proof of which follows easily from liz = ZH z.
Theorem 7.36. All norms on en are equivalent; i.e., there exist constants CI, C2 (possibly
depending on n) such that
Example 7.37. For x E en, the following inequalities are all tight bounds; i.e., there exist
vectors x for which equality holds:
Ilxlll :::: Jn Ilxlb
Ilxll2:::: IIxll»
IIxlloo :::: IIxll»
Ilxlll :::: n IIxlloo;
IIxl12 :::: Jn Ilxll
oo
;
IIxlioo :::: IIxllz.
Finally, we conclude this section with a theorem about convergence of vectors. Con
vergence of a sequence of vectors to some limit vector can be converted into a statement
about convergence of real numbers, i.e., convergence in terms of vector norms.
Theorem 7.38. Let II· II be a vector norm and suppose v, v(l), v(2), ... E en. Then
lim V(k) = v if and only if lim II v(k)  v II = O.
k4+00
7.4 Matrix Norms
In this section we introduce the concept of matrix norm. As with vectors, the motivation for
using matrix norms is to have a notion of either the size of or the nearness of matrices. The
former notion is useful for perturbation analysis, while the latter is needed to make sense of
"convergence" of matrices. Attention is confined to the vector space (IRm xn , IR) since that is
what arises in the majority of applications. Extension to the complex case is straightforward
and essentially obvious.
Definition 7.39. II· II : IR
mxn
IR is a matrix norm if it satisfies the following three
properties:
1. IIAII Of or all A E IR
mxn
and IIAII = 0 if and only if A = O.
2. lIaAl1 = lalliAliforall A E IR
mxn
andfor all a E IR.
3. IIA + BII :::: IIAII + IIBII for all A, BE IRmxn.
(As with vectors, this is called the triangle inequality.)
60 Chapter 7. Projections, Inner Product Spaces, and Norms
Example 7.40. Let A e R
mx
". Then the Frobenius norm (or matrix Euclidean norm) is
defined by
^wncic r = laiiK^/i;;.
Example 7.41. Let A e R
mxn
. Then the matrix pnorms are defined by
The following three special cases are important because they are "computable." Each is a
theorem and requires a proof.
1. The "maximum column sum" norm is
2. The "maximum row sum" norm is
3. The spectral norm is
Example 7.42. Let A E R
mxn
. The Schatten/7norms are defined by
Some special cases of Schatten /?norms are equal to norms defined previously. For example,
 . 
5 2
=  . \\
F
and  • 
5i00
=  • 
2
. The norm  • 
5>1
is often called the trace norm.
Example 7.43. Let A e K
mx
". Then "mixed" norms can also be defined by
Example 7.44. The "matrix analogue of the vector 1norm,"  A\\
s
= ^ j \a
i}
; , is a norm.
The concept of a matrix norm alone is not altogether useful since it does not allow us
to estimate the size of a matrix product A B in terms of the sizes of A and B individually.
60 Chapter 7. Projections, Inner Product Spaces, and Norms
Example 7.40. Let A E lR,mxn. Then the Frobenius norm (or matrix Euclidean norm) is
defined by
IIAIIF ~ (t. t ai;) I ~ (t. altA)) 1 ~ (T, (A' A)) 1 ~ (T, (AA '));
(where r = rank(A)).
Example 7.41. Let A E lR,mxn. Then the matrix pnorms are defined by
IIAxll
IIAII = max _P = max IIAxll .
P Ilxllp;60 Ilxli
p
IIxllp=1 p
The following three special cases are important because they are "computable." Each is a
theorem and requires a proof.
I. The "maximum column sum" norm is
2. The "maximum row sum" norm is
IIAlioo = max
rE!!l. (
t laUI ).
J=1
3. The spectral norm is
tTL T
IIAII2 = Amax(A A) = A ~ a x ( A A ) = a1(A).
Note: IIA+llz = l/ar(A), where r = rank(A).
Example 7.42. Let A E lR,mxn. The Schattenpnorms are defined by
I
IIAlls.p = (at' + ... + a!)"".
Some special cases of Schatten pnorms are equal to norms defined previously. For example,
11·115.2 = II . IIF and 11'115,00 = II . 112' The norm II . 115.1 is often called the trace norm.
Example 7.43. Let A E lR,mxn _ Then "mixed" norms can also be defined by
IIAII = max IIAxil
p
p,q 11.<110#0 IIxllq
Example 7.44. The "matrix analogue of the vector Inorm," IIAlis = Li.j laij I, is a norm.
The concept of a matrix norm alone is not altogether useful since it does not allow us
to estimate the size of a matrix product AB in terms of the sizes of A and B individually.
7.4. Matrix Norms 61
Notice that this difficulty did not arise for vectors, although there are analogues for, e.g.,
inner products or outer products of vectors. We thus need the following definition.
Definition 7.45. Let A e R
mxn
, B e R
nxk
. Then the norms \\ • \\
a
, \\ • \\
p
, and \\ • \\
y
are
mutually consistent if \\ A B \\
a
< \\A\\p\\B\\
y
. A matrix norm\\ • \\ is said to be consistent
if \\AB\\ <  A   fi whenever the matrix product is defined.
Example 7.46.
1.  • /7 and  • 
p
for all p are consistent matrix norms.
2. The "mixed" norm
is a matrix norm but it is not consistent. For example, take A = B = \ \ J1. Then
  Af l  
l i 00
= 2whil e  A 
l i 00
  B 
1 >00
= l.
The p norms are examples of matrix norms that are subordinate to (or induced by)
a vector norm, i.e.,
11^ 4^ 11
(or, more generally, A = max^o ., . .
p
) . For such subordinate norms, also called oper
ator norms, we clearly have Aj c < A1jt. Since   Af ij c  <   A    f l j c  < Aflj t,
it follows that all subordinate norms are consistent.
Theorem 7.47. There exists a vector x* such that Ajt* = A jc* if the matrix normis
subordinate to the vector norm.
Theorem 7.48. If \\ • \\
m
is a consistent matrix norm, there exists a vector norm \\ • \\
v
consistent with it, i.e., H Aj c JI ^ < \\A\\
m
\\x\\
v
.
Not every consistent matrix norm is subordinate to a vector norm. For example,
consider  • \\
F
. Then  A^ 
2
< A
F
j c
2
, so  • 
2
is consistent with  • 
F
, but there does
not exist a vector norm  •  such that A
F
is given by max^o \^ •
Useful Results
The following miscellaneous results about matrix norms are collected for future reference.
The interested reader is invited to prove each of them as an exercise.
2. For A e R"
x
", the following inequalities are all tight, i.e., there exist matrices A for
which equality holds:
7.4. Matrix Norms 61
Notice that this difficulty did not arise for vectors, although there are analogues for, e.g.,
inner products or outer products of vectors. We thus need the following definition.
Definition 7.45. Let A E ]Rmxn, B E ]Rnxk. Then the norms II . II", II· Ilfl' and II . lIy are
mutuallyconsistentifIlABII,,::S IIAllfllIBlly. A matrix norm 11·11 is said to be consistent
if II A B II ::s II A 1111 B II whenever the matrix product is defined.
Example 7.46.
1. II· II F and II . II p for all p are consistent matrix norms.
2. The "mixed" norm
IIAxll1
II· 11
100
= max = max laijl
, x;60 Ilx 1100 i,j
is a matrix norm but it is not consistent. For example, take A = B = [: :]. Then
IIABIII,oo = 2 while IIAIII,ooIlBIII,oo = 1.
The pnorms are examples of matrix norms that are subordinate to (or induced by)
a vector norm, i.e.,
IIAxl1
IIAII = max  = max IIAxl1
x;60 IIx II Ilxll=1
IIAxll .
(or, more generally, IIAllp,q = maxx;60 IIxll
q
P
), For such subordmate norms, also caUedoper
atornorms, wec1earlyhave IIAxll ::s IIAllllxll· Since IIABxl1 ::s IIAlIllBxll ::s IIAIIIIBllllxll,
it follows that all subordinate norms are consistent.
Theorem 7.47. There exists a vector x* such that IIAx*11 = IIAllllx*11 if the matrix norm is
subordinate to the vector norm.
Theorem 7.48. If II . 11m is a consistent matrix norm, there exists a vector norm II . IIv
consistent with it, i.e., IIAxliv ::s IIAlim Ilxli
v
'
Not every consistent matrix norm is subordinate to a vector norm. For example,
consider II . II F' Then II Ax 112 ::s II A II Filx 112, so II . 112 is consistent with II . II F, but there does
not exist a vector norm II . II such that IIAIIF is given by max
x
;60 " , ~ ~ i ' .
Useful Results
The following miscellaneous results about matrix norms are collected for future reference.
The interested reader is invited to prove each of them as an exercise.
1. II In II p = 1 for all p, while IIIn II F = .jii.
2. For A E ]Rnxn, the following inequalities are all tight, i.e., there exist matrices A for
which equality holds:
IIAIII ::s .jii IIAlb
IIAII2 ::s.jii IIAII
I
,
II A 1100 ::s n IIAII
I
,
IIAIIF ::s.jii IIAII
I
,
IIAIII ::s n IIAlloo,
IIAII2 ::s .jii IIAlloo,
IIAlioo ::s .jii IIAII2,
IIAIIF ::s .jii IIAlb
IIAIII ::s .jii II
A
IIF;
IIAII2::S IIAIIF;
IIAlioo ::s .jii IIAIIF;
IIAIIF ::s .jii IIAlioo'
62 Chapter 7. Projections, Inner Product Spaces, and Norms
3. For A eR
mxa
.
4. The norms  • \\
F
and  • 
2
(as well as all the Schatten /?norms, but not necessarily
other pnorms) are unitarily invariant; i.e., for all A e R
mx
" and for all orthogonal
matrices Q zR
mxm
and Z e M"
x
", (MZ
a
=   A 
a
fora = 2 or F.
Convergence
The following theorem uses matrix norms to convert a statement about convergence of a
sequence of matrices into a statement about the convergence of an associated sequence of
scalars.
Theorem 7.49. Let \\ \\bea matrix normand suppose A, A
( 1)
, A
(2)
, ... e R
mx
". Then
EXERCISES
1. If P is an orthogonal projection, prove that P
+
= P.
2. Suppose P and Q are orthogonal projections and P + Q = I. Prove that P — Q
must be an orthogonal matrix.
3. Prove that / — A
+
A is an orthogonal projection. Also, prove directly that V
2
V/ is an
orthogonal projection, where ¥2 is defined as in Theorem 5.1.
4. Suppose that a matrix A e W
nxn
has linearly independent columns. Prove that the
orthogonal projection onto the space spanned by these column vectors is given by the
matrix P = A(A
T
A)~
}
A
T
.
5. Find the (orthogonal) projection of the vector [2 3 4]
r
onto the subspace of R
3
spanned by the plane 3;c — v + 2z = 0.
6. Prove that E"
x
" with the inner product (A, B) = Tr A
T
B is a real inner product
space.
7. Show that the matrix norms  • 
2
and  • \\
F
are unitarily invariant.
8. Definition: Let A e R
nxn
and denote its set of eigenvalues (not necessarily distinct)
by { Ai , . . . , > . „ } . The spectral radius of A is the scalar
62 Chapter 7. Projections, Inner Product Spaces, and Norms
3. For A E IR
mxn
,
max laijl :::: IIAII2 :::: ~ max laijl.
l.] l.]
4. The norms II . IIF and II . 112 (as well as all the Schatten pnorms, but not necessarily
other pnorms) are unitarily invariant; i.e., for all A E IR
mxn
and for all orthogonal
matrices Q E IR
mxm
and Z E IR
nxn
, IIQAZlia = IIAlla fora = 2 or F.
Convergence
The following theorem uses matrix norms to convert a statement about convergence of a
sequence of matrices into a statement about the convergence of an associated sequence of
scalars.
Theorem 7.49. Let II ·11 be a matrix norm and suppose A, A(I), A(2), ... E IRmxn. Then
lim A (k) = A if and only if lim IIA (k)  A II = o.
k ~ + o o k ~ + o o
EXERCISES
1. If P is an orthogonal projection, prove that p+ = P.
2. Suppose P and Q are orthogonal projections and P + Q = I. Prove that P  Q
must be an orthogonal matrix.
3. Prove that I  A + A is an orthogonal projection. Also, prove directly that V
2
Vl is an
orthogonal projection, where V2 is defined as in Theorem 5.1.
4. Suppose that a matrix A E IR
mxn
has linearly independent columns. Prove that the
orthogonal projection onto the space spanned by these column vectors is given by the
matrix P = A(AT A) 1 AT.
5. Find the (orthogonal) projection of the vector [2 3 4f onto the subspace of 1R
3
spanned by the plane 3x  y + 2z = O.
6. Prove that IR
n
xn with the inner product (A, B) = Tr AT B is a real inner product
space.
7. Show that the matrix norms II . 112 and II . IIF are unitarily invariant.
8. Definition: Let A E IR
nxn
and denote its set of eigenvalues (not necessarily distinct)
by P.l, ... , An}. The spectral radius of A is the scalar
p(A) = max IA;I.
i
Exercises 63
Determine A
F
, H AI d , A
2
, H AH ^ , and p(A). (An n x n matrix, all of whose
columns and rows as well as main d iagonal and antid iagonal sum to s = n(n
2
+ l)/2,
is called a "magic square" matrix. I f M is a magic square matrix, it can be proved
that  M U p = s for all/?.)
10. Let A = xy
T
, where both x, y e R" are nonzero. Determine A
F
, Aj, A
2
,
and Aoo in terms of \\x\\
a
and /or \\y\\p, where a and ft take the value 1, 2, or oo as
appropriate.
Let
9. Let
Determine A
F
, \\A\\
lt
A
2
, H A^ , and p(A).
Exercises 63
Let
A = [ ~ 0 ~ ] .
14 12 5
Determine IIAIIF' IIAII
I
, IIAlb IIAlloo, and peA).
9. Let
A = [ ~ ~ ~ ] .
492
Determine IIAIIF' IIAII
I
, IIAlb IIAlloo, and peA). (An n x n matrix, all of whose
columns and rows as well as main diagonal and antidiagonal sum to s = n (n
2
+ 1) /2,
is called a "magic square" matrix. If M is a magic square matrix, it can be proved
that IIMllp = s for all p.)
10. Let A = xyT, where both x, y E IR
n
are nonzero. Determine IIAIIF' IIAIII> IIAlb
and II A 1100 in terms of IIxlla and/or IlylljJ, where ex and {3 take the value 1,2, or (Xl as
appropriate.
This page intentionally left blank This page intentionally left blank
Chapter 8
Li near Least Squares
Problems
8.1 The Li near Least Squares Problem
Problem: Suppose A e R
mx
" with m > n and b <= R
m
is a given vector. The linear least
squares problem consists of finding an element of the set
Solution: The set X has a number of easily verified properties:
1. A vector x e X if and only if A
T
r = 0, where r = b — Ax is the residual associated
with x. The equations A
T
r — 0 can be rewritten in the form A
T
Ax = A
T
b and the
latter form is commonly known as the normal equations, i.e., x e X if and only if
x is a solution of the normal equations. For further details, see Section 8.2.
2. A vector x E X if and onlv if x is of the form
To see why this must be so, write the residual r in the form
Now, (Pn(A)b — AJ C ) is clearly in 7£(A) , while
so these two vectors are orthogonal. Hence,
from the Pythagorean identity (Remark 7.35). Thus, A.x — b\\\ (and hence p ( x ) =
\\Ax —b\\2) assumes its minimum value if and only if
65
Chapter 8
Linear Least Squares
Problems
8.1 The Linear Least Squares Problem
Problem: Suppose A E jRmxn with m 2: nand b E jRm is a given vector. The linear least
squares problem consists of finding an element of the set
x = {x E jRn : p(x) = IIAx  bll
2
is minimized}.
Solution: The set X has a number of easily verified properties:
1. A vector x E X if and only if AT r = 0, where r = b  Ax is the residual associated
with x. The equations AT r = 0 can be rewritten in the form A T Ax = AT b and the
latter form is commonly known as the normal equations, i.e., x E X if and only if
x is a solution of the normal equations. For further details, see Section 8.2.
2. A vector x E X if and only if x is of the form
x=A+b+(IA+A)y, whereyEjRnisarbitrary. (8.1)
To see why this must be so, write the residual r in the form
r = (b  PR(A)b) + (PR(A)b  Ax).
Now, (PR(A)b  Ax) is clearly in 'R(A), while
(b  PR(A)b) = (I  PR(A))b
= PR(A),,b E 'R(A)L
so these two vectors are orthogonal. Hence,
= lib 
= lib  + IIPR(A)b 
from the Pythagorean identity (Remark 7.35). Thus, IIAx  (and hence p(x) =
II Ax  b 112) assumes its minimum value if and only if
(8.2)
65
66 Chapter 8. Linear Least Squares Problems
and this equation always has a solution since AA
+
b e 7£(A). By Theorem 6.3, all
solutions of (8.2) are of the form
where y e W is arbitrary. The minimum value of p ( x ) is then clearly equal to
the last inequality following by Theorem 7.23.
3. X is convex. To see why, consider two arbitrary vectors jci = A
+
b + (I — A+A)y
and *2 = A+b + (I — A+A)z in X. Let 6 e [0, 1]. Then the convex combination
0*i + (1  #)*
2
= A+b + (I  A
+
A)(Oy + (1  0)z) is clearly in X.
4. X has a unique element x* of minimal 2norm. In fact, x* = A
+
b is the unique vector
that solves this "double minimization" problem, i.e., x * minimizes the residual p ( x )
and is the vector of minimum 2norm that does so. This follows immediately from
convexity or directly from the fact that all x e X are of the form (8.1) and
which follows since the two vectors are orthogonal.
5. There is a unique solution to the least squares problem, i.e., X = {x*} = {A+b}, if
and only if A
+
A = I or, equivalently, if and only if rank (A) = n.
Just as for the solution of linear equations, we can generalize the linear least squares
problem to the matrix case.
Theorem 8.1. Let A e E
mx
" and B € R
mxk
. The general solution to
is of the form
where Y € R"
xfc
is arbitrary. The unique solution of minimum 2norm or Fnorm is
X = A+B.
Remark 8.2. Notice that solutions of the linear least squares problem look exactly the
same as solutions of the linear system AX = B. The only difference is that in the case
of linear least squares solutions, there is no "existence condition" such as K(B) c 7£(A).
If the existence condition happens to be satisfied, then equality holds and the least squares
66 Chapter 8. Linear Least Squares Problems
and this equation always has a solution since AA+b E R(A). By Theorem 6.3, all
solutions of (8.2) are of the form
x = A+ AA+b + (I  A+ A)y
=A+b+(IA+A)y,
where y E ]R.n is arbitrary. The minimum value of p (x) is then clearly equal to
lib  PR(A)bll
z
= 11(1  AA+)bI1
2
~ Ilbll z,
the last inequality following by Theorem 7.23.
3. X is convex. To see why, consider two arbitrary vectors Xl = A + b + (I  A + A) y
and Xz = A+b + (I  A+ A)z in X. Let 8 E [0,1]. Then the convex combination
8x, + (1  8)xz = A+b + (I  A+ A)(8y + (1  8)z) is clearly in X.
4. X has a unique element x" of minimal2norm. In fact, x" = A + b is the unique vector
that solves this "double minimization" problem, i.e., x* minimizes the residual p(x)
and is the vector of minimum 2norm that does so. This follows immediately from
convexity or directly from the fact that all x E X are of the form (8.1) and
which follows since the two vectors are orthogonal.
5. There is a unique solution to the least squares problem, i.e., X = {x"} = {A+b}, if
and only if A + A = lor, equivalently, if and only if rank(A) = n.
Just as for the solution of linear equations, we can generalize the linear least squares
problem to the matrix case.
Theorem 8.1. Let A E ]R.mxn and BE ]R.mxk. The general solution to
min IIAX  Bib
XElR
Plxk
is of the form
X=A+B+(IA+A)Y,
where Y E ]R.nxk is arbitrary. The unique solution of minimum 2norm or Fnorm is
X = A+B.
Remark 8.2. Notice that solutions of the linear least squares problem look exactly the
same as solutions of the linear system AX = B. The only difference is that in the case
of linear least squares solutions, there is no "existence condition" such as R(B) S; R(A).
If the existence condition happens to be satisfied. then equality holds and the least squares
8.3 Linear Regression and Other Linear Least Squares Problems 67
residual is 0. Of all solutions that give a residual of 0, the unique solution X = A
+
B has
minimum 2norm or Fnorm.
Remark 8.3. If we take B = I
m
in Theorem 8.1, then X = A
+
can be interpreted as
saying that the MoorePenrose pseudoinverse of A is the best (in the matrix 2norm sense)
matrix such that AX approximates the identity.
Remark 8.4. Many other interesting and useful approximation results are available for the
matrix 2norm (and Fnorm). One such is the following. Let A e M™
x
" with SVD
8.2 Geometric Solution
Looking at the schematic provided in Figure 8.1, it is apparent that minimizing  Ax — b\\
2
is equivalent to finding the vector x e W
1
for which p — Ax is closest to b (in the Euclidean
norm sense). Clearly, r = b — Ax must be orthogonal to 7£(A). Thus, if Ay is an arbitrary
vector in 7£(A) (i.e., y is arbitrary), we must have
Then a best rank k approximation to A for l <f c <r , i . e . , a solution to
is given by
The special case in which m = n and k = n — 1 gives a nearest singular matrix to A e
Since y is arbitrary, we must have A
T
b — A
T
Ax = 0 or A
r
A;c = A
T
b.
Special case: If A is full (column) rank, then x = (A
T
A) A
T
b.
8.3 Linear Regression and Other Linear Least Squares
Problems
8.3.1 Example: Linear regression
Suppose we have m measurements (t\,y\), . . . , (t
m
,y
m
) for which we hypothesize a linear
(affine) relationship
8.3 Linear Regression and Other Linear Least Squares Problems 67
residual is O. Of all solutions that give a residual of 0, the unique solution X = A + B has
minimum 2norm or F norm.
Remark 8.3. If we take B = 1m in Theorem 8.1, then X = A+ can be interpreted as
saying that the MoorePenrose pseudoinverse of A is the best (in the matrix 2norm sense)
matrix such that AX approximates the identity.
Remark 8.4. Many other interesting and useful approximation results are available for the
matrix 2norm (and F norm). One such is the following. Let A E with SVD
A = = LOiUiV!.
i=l
Then a best rank k approximation to A for 1 :s k :s r, i.e., a solution to
min IIA  MIi2,
MEJRZ'xn
is given by
k
Mk = LOiUiV!.
i=1
The special case in which m = nand k = n  1 gives a nearest singular matrix to A E x n .
8.2 Geometric Solution
Looking at the schematic provided in Figure 8.1, it is apparent that minimizing IIAx  bll
2
is equivalent to finding the vector x E lR
n
for which p = Ax is closest to b (in the Euclidean
norm sense). Clearly, r = b  Ax must be orthogonal to R(A). Thus, if Ay is an arbitrary
vector in R(A) (i.e., y is arbitrary), we must have
0= (Ay)T (b  Ax)
=yTAT(bAx)
= yT (ATb _ AT Ax).
Since y is arbitrary, we must have AT b  AT Ax = 0 or AT Ax = AT b.
Special case: If A is full (column) rank, then x = (AT A)l ATb.
8.3 Linear Regression and Other Linear Least Squares
Problems
8.3.1 Example: Linear regression
Suppose we have m measurements (ll, YI), ... , (trn, Ym) for which we hypothesize a linear
(affine) relationship
y = at + f3
(8.3)
68 Chapter 8. Linear Least Squares Problems
Figure 8.1. Projection of b on K(A).
for certain constants a. and ft. One way to solve this problem is to find the line that best fits
the data in the least squares sense; i.e., with the model (8.3), we have
where &\,..., 8
m
are "errors" and we wish to minimize 8\ + • • • + 8^ Geometrically, we
are trying to find the best line that minimizes the (sum of squares of the) distances from the
given data points. See, for example, Figure 8.2.
Figure 8.2. Simple linear regression.
Note that distances are measured in the vertical sense from the points to the line (as
indicated, for example, for the point (t\, y\}}. However, other criteria arc possible. For ex
ample, one could measure the distances in the horizontal sense, or the perpendicular distance
from the points to the line could be used. The latter is called total least squares. Instead
of 2norms, one could also use 1norms or oonorms. The latter two are computationally
68 Chapter 8. Linear Least Squares Problems
b
r
p=Ax Ay E R(A)
Figure S.l. Projection of b on R(A).
for certain constants a and {3. One way to solve this problem is to find the line that best fits
the data in the least squares sense; i.e., with the model (8.3), we have
YI = all + {3 + 81,
Y2 = al2 + {3 + 82
where 8
1
, ... , 8
m
are "errors" and we wish to minimize 8? + ... + 8;. Geometrically, we
are trying to find the best line that minimizes the (sum of squares of the) distances from the
given data points. See, for example, Figure 8.2.
y
Figure 8.2. Simple linear regression.
Note that distances are measured in the venical sense from the point!; to [he line (a!;
indicated. for example. for the point (tl. YIn. However. other criteria nrc For cx
ample, one could measure the distances in the horizontal sense, or the perpendiculnr distance
from the points to the line could be used. The latter is called total least squares. Instead
of 2norms, one could also use Inorms or oonorms. The latter two are computationally
8.3. Linear Regression and Other Linear Least Squares Problems 69
much more difficult to handle, and thus we present only the more tractable 2norm case in
text that follows.
The ra "error equations" can be written in matrix form as
where
We then want to solve the problem
or, equivalently,
Solution: x — [^1 is a solution of the normal equations A
T
Ax = A
T
y where, for the
special form of the matrices above, we have
and
8.3.2 Other least squares problems
Suppose the hypothesized model is not the linear equation (8.3) but rather is of the form
y = f ( t ) =
Cl
0!(0+ • • • 4 c
n
<t>
n
(t). (8.5)
In (8.5) the < / > ,(0 are given (basis) functions and the c
;
are constants to be determined to
minimize the least squares error. The matrix problem is still (8.4), where we now have
An important special case of (8.5) is least squares polynomial approximation, which
corresponds to choosing 0,• (?) = t'~
l
, i
;
e n, although this choice can lead to computational
The solution for the parameters a and ft can then be written
8.3. Linear Regression and Other Linear Least Squares Problems 69
much more difficult to handle, and thus we present only the more tractable 2norm case in
text that follows.
The m "error equations" can be written in matrix form as
Y = Ax +0,
where
We then want to solve the problem
minoT 0 = min (Ax  y)T (Ax  y)
x
or, equivalently,
min = min II Ax 
x
Solution: x = is a solution of the normal equations AT Ax
special form of the matrices above, we have
and
AT Y = [ Li ti Yi J.
LiYi
The solution for the parameters a and f3 can then be written
8.3.2 Other least squares problems
(8.4)
AT y where, for the
Suppose the hypothesized model is not the linear equation (S.3) but rather is of the form
(8.5)
In (8.5) the ¢i(t) are given (basis) functions and the Ci are constants to be determined to
minimize the least squares error. The matrix problem is still (S.4), where we now have
An important special case of (8.5) is least squares polynomial approximation, which
corresponds to choosing ¢i (t) = t
i

1
, i E !!, although this choice can lead to computational
70 Chapter 8. Linear Least Squares Problems
difficulties because of numerical ill conditioning for large n. Numerically better approaches
are based on orthogonal polynomials, piecewise polynomial functions, splines, etc.
The key feature in (8.5) is that the coefficients c, appear linearly. The basis functions
< / > , can be arbitrarily nonlinear. Sometimes a problem in which the c, 's appear nonlinearly
can be converted into a linear problem. For example, if the fitting function is of the form
y = f ( t ) = c\e
C2i
, then taking logarithms yields the equation logy = logci + cjt. Then
defining y — logy, c\ = logci, and GI = cj_ results in a standard linear least squares
problem.
8.4 Least Squares and Singular Value Decomposition
In the numerical linear algebra literature (e.g., [4], [7], [11], [23]), it is shown that solution
of linear least squares problems via the normal equations can be a very poor numerical
method in finite precision arithmetic. Since the standard Kalman filter essentially amounts
to sequential updating of normal equations, it can be expected to exhibit such poor numerical
behavior in practice (and it does). Better numerical methods are based on algorithms that
work directly and solely on A itself rather than A
T
A. Two basic classes of algorithms are
based on S VD and QR (orthogonal upper triangular) factorization, respectively. The former
is much more expensive but is generally more reliable and offers considerable theoretical
insight.
In this section we investigate solution of the linear least squares problem
The last equality follows from the fact that if v = [£ ], then u^ =   i> i \\\ + \\vi\\\ (note
that orthogonality is not what is used here; the subvectors can have different lengths). This
explains why it is convenient to work above with the square of the norm rather than the
norm. As far as the minimization is concerned, the two are equivalent. In fact, the last
quantity above is clearly minimized by taking z\ = S~
l
c\. The subvector z
2
is arbitrary,
while the minimum value of \\Ax — b\\^ is l ^l l r
via the SVD. Specifically, we assume that A has an SVD given by A = UT, V
T
= U\SVf
as in Theorem 5.1. We now note that
70 Chapter 8. Linear Least Squares Problems
difficulties because of numerical ill conditioning for large n. Numerically better approaches
are based on orthogonal polynomials, piecewise polynomial functions, splines, etc.
The key feature in (8.5) is that the coefficients Ci appear linearly. The basis functions
¢i can be arbitrarily nonlinear. Sometimes a problem in which the Ci'S appear nonlinearly
can be converted into a linear problem. For example, if the fitting function is of the form
Y = f (t) = c, e
C2
/ , then taking logarithms yields the equation log y = log c, + c2f. Then
defining y = log y, c, = log c" and C2 = C2 results in a standard linear least squares
problem.
8.4 Least Squares and Singular Value Decomposition
In the numerical linear algebra literature (e.g., [4], [7], [11], [23]), it is shown that solution
of linear least squares problems via the normal equations can be a very poor numerical
method in finiteprecision arithmetic. Since the standard Kalman filter essentially amounts
to sequential updating of normal equations, it can be expected to exhibit such poor numerical
behavior in practice (and it does). Better numerical methods are based on algorithms that
work directly and solely on A itself rather than AT A. Two basic classes of algorithms are
based on SVD and QR (orthogonalupper triangular) factorization, respectively. The former
is much more expensive but is generally more reliable and offers considerable theoretical
insight.
In this section we investigate solution of the linear least squares problem
min II Ax  b11
2
, A E IR
mxn
, bE IR
m
, (8.6)
x
via the SVD. Specifically, we assume that A has an SVD given by A = = U,SVr
as in Theorem 5.1. We now note that
IIAx  = x 
= II VT X  U
T
bll; since II . Ib is unitarily invariant
wherez=VTx,c=UTb
= II [ ]  [ ] II:
= II [ c, ] II:
The last equality follows from the fact that if v = then II v II = II viii + II v211 (note
that orthogonality is not what is used here; the subvectors can have different lengths). This
explains why it is convenient to work above with the square of the norm rather than the
norm. As far as the minimization is concerned. the two are equivalent. In fact. the last
quantity above is clearly minimized by taking z, = S'c,. The subvector Z2 is arbitrary,
while the minimum value of II Ax  b II is II czll
8.5. Least Squares and QR Factorization 71
Now transform back to the original coordinates:
The last equality follows from
Note that since 12 is arbitrary, V
2
z
2
is an arbitrary vector in 7Z(V
2
) = A/"(A). Thus, x has
been written in the form x = A
+
b + (/ — A
+
A ) _ y, where y e R
m
is arbitrary. This agrees,
of course, with (8.1).
The minimum value of the least squares residual is
and we clearly have that
minimum least squares residual is 0 4=> b is orthogonal to all vectors in U
2
•<=^ b is orthogonal to all vectors in 7l(A}
L
Another expression for the minimum residual is  (/ — AA
+
) b 
2
. This follows easily since
(7  AA+)b\\
2
2
 \\U2Ufb\\l = b
T
U
2
U^U
2
UJb = b
T
U
2
U*b = \\U?b\\
2
2
.
Finally, an important special case of the linear least squares problem is the
socalled fullrank problem, i.e., A e 1R™
X
" . In this case the SVD of A is given by
A = UZV
T
= [U
{
t/ 2][o]^i
r
> and there is thus "no V
2
part" to the solution.
8.5 Least Squares and QR Factorization
In this section, we again look at the solution of the linear least squares problem (8.6) but this
time in terms of the QR factorization. This matrix factorization is much cheaper to compute
than an SVD and, with appropriate numerical enhancements, can be quite reliable.
To simplify the exposition, we add the simplifying assumption that A has full column
rank, i.e., A e R™
X M
. It is then possible, via a sequence of socalled Householder or Givens
transformations, to reduce A in the following way. A finite sequence of simple orthogonal
row transformations (of Householder or Givens type) can be performed on A to reduce it
to triangular form. If we label the product of such orthogonal row transformations as the
orthogonal matrix Q
T
€ R
mxm
, we have
B.S. Least Squares and QR Factorization 71
Now transform back to the original coordinates:
x = Vz
= [VI V
2
1 [ ]
= VIZI + V2Z2
= VISici + V2Z2
= vlsIufb + V
2
Z
2
.
The last equality follows from
c = U T b = [ f: ] = [ l
Note that since Z2 is arbitrary, V
2
Z
2
is an arbitrary vector in R(V
2
) = N(A). Thus, x has
been written in the form x = A + b + (I  A + A) y, where y E ffi.m is arbitrary. This agrees,
of course, with (8.1).
The minimum value of the least squares residual is
and we clearly have that
minimum least squares residual is 0 {::=:} b is orthogonal to all vectors in U2
{::=:} b is orthogonal to all vectors in R(A)l.
{::=:} b E R(A).
Another expression for the minimum residual is II (I  AA +)bllz. This follows easily since
11(1 = = b
T
U
Z
V!V
2
V!b = bTVZV!b =
Finally, an important special case of the linear least squares problem is the
socalled fullrank problem, i.e., A E In this case the SVD of A is given by
A = V:EV
T
= [VI Vzl[g]Vr, and there is thus "no V
2
part" to the solution.
8.5 Least Squares and QR Factorization
In this section, we again look at the solution of the linear least squares problem (8.6) but this
time in terms of the QR factorization. This matrix factorization is much cheaper to compute
than an SVD and, with appropriate numerical enhancements, can be quite reliable.
To simplify the exposition, we add the simplifying assumption that A has full column
rank, i.e., A E It is then possible, via a sequence of socalled Householder or Givens
transformations, to reduce A in the following way. A finite sequence of simple orthogonal
row transformations (of Householder or Givens type) can be performed on A to reduce it
to triangular form. If we label the product of such orthogonal row transformations as the
orthogonal matrix QT E ffi.mxm, we have
(8.7)
72 Chapter 8. Linear Least Squares Problems
where R e M£
x
" is upper triangular. Now write Q = [Q\ Q
2
], where Q\ e R
mx
" and
Q
2
€ K"
IX(m
~"
)
. Both Q\ and <2
2
have orthonormal columns. Multiplying through by Q
in (8.7), we see that
Any of (8.7), (8.8), or (8.9) are variously referred to as QR factorizations of A. Note that
(8.9) is essentially what is accomplished by the GramSchmidt process, i.e., by writing
AR~
l
= Q\ we see that a "triangular" linear combination (given by the coefficients of
R~
l
) of the columns of A yields the orthonormal columns of Q\.
Now note that
The last quantity above is clearly minimized by taking x = R
l
c\ and the minimum residual
is \\C 2\\2 Equivalently, we have x = R~
l
Q\b = A
+
b and the minimum residual is IIC?^!^
EXERCISES
1. For A € W
xn
, b e E
m
, and any y e R", check directly that (I  A
+
A)y and A
+
b
are orthogonal vectors.
2. Consider the following set of measurements (*,, y
t
):
(a) Find the best (in the 2norm sense) line of the form y = ax + ft that fits this
data.
(b) Find the best (in the 2norm sense) line of the form jc = ay + (3 that fits this
data.
3. Suppose qi and q
2
are two orthonormal vectors and b is a fixed vector, all in R".
(a) Find the optimal linear combination aq^ + fiq
2
that is closest to b (in the 2norm
sense).
(b) Let r denote the "error vector" b — ctq\ — flq
2
 Show that r is orthogonal to
both^i and q
2
.
72 Chapter 8. Linear Least Squares Problems
where R E is upper triangular. Now write Q = [QI Qz], where QI E ffi.mxn and
Qz E ffi.m x (mn). Both Q I and Qz have orthonormal columns. Multiplying through by Q
in (8.7), we see that
(8.8)
= [QI Qz] [ ]
= QIR.
(8.9)
Any of (8.7), (8.8), or (8.9) are variously referred to as QR factorizations of A. Note that
(8.9) is essentially what is accomplished by the GramSchmidt process, i.e., by writing
AR
1
= QI we see that a "triangular" linear combination (given by the coefficients of
R
I
) of the columns of A yields the orthonormal columns of Q I.
Now note that
IIAx  = IIQ
T
Ax  since II . 112 is unitarily invariant
= II [ ] x  [ ] If:,
The last quantity above is clearly minimized by taking x = R
I
Cl and the minimum residual
is Ilczllz. Equivalently, we have x = R
1
Qf b = A +b and the minimum residual is II Qr bllz'
EXERCISES
1. For A E ffi.
mxn
, b E ffi.
m
, and any y E ffi.
n
, check directly that (I  A + A)y and A +b
are orthogonal vectors.
2. Consider the following set of measurements (Xi, Yi):
(1,2), (2,1), (3,3).
(a) Find the best (in the 2norm sense) line of the form y = ax + fJ that fits this
data.
(b) Find the best (in the 2norm sense) line of the form x = ay + fJ that fits this
data.
3. Suppose q, and qz are two orthonormal vectors and b is a fixed vector, all in ffi.
n
•
(a) Find the optimallinear combination aql + (3q2 that is closest to b (in the 2norm
sense).
(b) Let r denote the "error vector" b  aql  {3qz. Show that r is orthogonal to
both ql and q2.
Exercises 73
4. Find all solutions of the linear least squares problem
5. Consider the problem of finding the minimum 2norm solution of the linear least
«rmarp« nrr»h1<=>m
(a) Consider a perturbation E\ = [
0
pi of A, where 8 is a small positive number.
Solve the perturbed version of the above problem,
where AI = A + E\. What happens to jt* — y 
2
as 8 approaches 0?
(b) Now consider the perturbation EI = \
0 s
~\ of A, where again 8 is a small
positive number. Solve the perturbed problem
where A
2
— A + E
2
. What happens to \\x* — z
2
as 8 approaches 0?
6. Use the four Penrose conditions and the fact that Q\ has orthonormal columns to
verify that if A e R™
x
" can be factored in the form (8.9), then A+ = R~
l
Q\.
1. Let A e R"
x
", not necessarily nonsingular, and suppose A = QR, where Q is
orthogonal. Prove that A
+
= R
+
Q
T
.
Exercises 73
4. Find all solutions of the linear least squares problem
min II Ax  bll
2
x
when A = [
5. Consider the problem of finding the minimum 2norm solution of the linear least
squares problem
min II Ax  bl1
2
x
when A = ] and b = [ ! 1 The solution is
(a) Consider a perturbation EI = of A, where 8 is a small positive number.
Solve the perturbed version of the above problem,
where AI = A + E
I
. What happens to IIx*  yII2 as 8 approaches O?
(b) Now consider the perturbation E2 = n of A, where again 8 is a small
positive number. Solve the perturbed problem
min II A
2
z  bib
z
where A2 = A + E
2
• What happens to IIx*  zll2 as 8 approaches O?
6. Use the four Penrose conditions and the fact that QI has orthonormal columns to
verify that if A E can be factored in the form (8.9), then A+ = R
I
Qf.
7. Let A E not necessarily nonsingular, and suppose A = QR, where Q is
orthogonal. Prove that A + = R+ QT .
This page intentionally left blank This page intentionally left blank
Chapter 9
Eigenvalues and
Eigenvectors
9.1 Fundamental Definitions and Properties
Definition 9.1. A nonzero vector x e C" is a right eigenvector of A e C
nxn
if there exists
a scalar A. e C, called an eigenvalue, such that
Similarly, a nonzero vector y e C" is a left eigenvector corresponding to an eigenvalue
a if
By taking Hermitian transposes in (9.1), we see immediately that X
H
is a left eigen
vector of A
H
associated with A . Note that if x [y] is a right [left] eigenvector of A, then
so is ax [ay] for any nonzero scalar a E C. One oftenused scaling for an eigenvector is
a — \j';t so that the scaled eigenvector has norm 1. The 2norm is the most common
norm used for such scaling.
Definition 9.2. The polynomial n (A.) = det(A —A ,/ ) is called the characteristic polynomial
of A. (Note that the characteristic polynomial can also be defined as det(A . / — A ). This
results in at most a change of sign and, as a matter of convenience, we use both forms
throughout the text.}
The following classical theorem can be very useful in hand calculation. It can be
proved easily from the Jordan canonical form to be discussed in the text to follow (see, for
example, [21]) or directly using elementary properties of inverses and determinants (see,
for example, [3]).
Theorem 9.3 (CayleyHamilton). For any A e C
nxn
, n(A) = 0.
Example 9.4. Let A = [~g ~g] . Then n(k) = X
2
+ 2A , — 3. It is an easy exercise to
verify that n(A) = A
2
+ 2A  31 = 0.
It can be proved from elementary properties of determinants that if A e C"
x
", then
7 t (X) is a polynomial of degree n. Thus, the Fundamental Theorem of A lgebra says that
75
Chapter 9
Eigenvalues and
Eigenvectors
9.1 Fundamental Definitions and Properties
Definition 9.1. A nonzero vector x E en is a right eigenvector of A E e
nxn
if there exists
a scalar A E e, called an eigenvalue, such that
Ax = AX. (9.1)
Similarly, a nonzero vector y E en is a left eigenvector corresponding to an eigenvalue
Mif
(9.2)
By taking Hennitian transposes in (9.1), we see immediately that x
H
is a left eigen
vector of A H associated with I. Note that if x [y] is a right [left] eigenvector of A, then
so is ax [ay] for any nonzero scalar a E C. One oftenused scaling for an eigenvector is
a = 1/ IIx II so that the scaled eigenvector has nonn 1. The 2nonn is the most common
nonn used for such scaling.
Definition 9.2. The polynomialn (A) = det (A  A l) is called the characteristic polynomial
of A. (Note that the characteristic polynomial can also be defined as det(Al  A). This
results in at most a change of sign and, as a matter of convenience, we use both forms
throughout the text.)
The following classical theorem can be very useful in hand calculation. It can be
proved easily from the Jordan canonical fonn to be discussed in the text to follow (see, for
example, [21D or directly using elementary properties of inverses and determinants (see,
for example, [3]).
Theorem 9.3 (CayleyHamilton). For any A E e
nxn
, n(A) = O.
Example 9.4. Let A = [  ~  ~ ] . Then n(A) = A2 + 2A  3. It is an easy exercise to
verify that n(A) = A2 + 2A  31 = O.
It can be proved from elementary properties of detenninants that if A E e
nxn
, then
n(A) is a polynomial of degree n. Thus, the Fundamental Theorem of Algebra says that
75
and set X = 0 in this identity, we get the interesting fact that del (A) = AI • A.2 • • • A
M
(see
also Theorem 9.25).
If A e W
xn
, then n(X) has real coefficients. Hence the roots of 7 r( A) , i.e., the
eigenvalues of A, must occur in complex conjugate pairs.
Example 9.6. Let a, ft e R and let A = [ _^ £ ]. Then jr( A. ) = A.
2
 2aA + a
2
+ ft
2
and
A has eigenvalues a ± fij (where j = i = •>/—!)•
If A € R"
x
", then there is an easily checked relationship between the left and right
eigenvectors of A and A
T
(take Hermitian transposes of both sides of (9.2)). Specifically, if
y is a left eigenvector of A corresponding to A e A( A) , then y is a right eigenvector of A
T
corresponding to A. € A ( A) . Note, too, that by elementary properties of the determinant,
we always have A ( A ) = A ( A
r
) , but that A ( A ) = A ( A ) only if A e R"
x
".
Definition 9.7. IfX is a root of multiplicity m ofjr(X), we say that X is an eigenvalue of A
of algebraic multiplicity m. The geometric multiplicity ofX is the number of associated
independent eigenvectors = n — rank( A — A/) = dim J \ f(A — XI).
If A € A ( A ) has algebraic multiplicity m, then 1 < di mA/ "(A — A/) < m. Thus, if
we denote the geometric multiplicity of A by g, then we must have 1 < g < m.
Definition 9.8. A matrix A e W
x
" is said to be defective if it has an eigenvalue whose
geometric multiplicity is not equal to (i.e., less than) its algebraic multiplicity. Equivalently,
A is said to be defective if it does not have n linearly independent (right or left) eigenvectors.
From the CayleyHamilton Theorem, we know that n(A) = 0. However, it is pos
sible for A to satisfy a lowerorder polynomial. For example, if A = \
1
Q
®], then A sat
isfies (1 — I)
2
= 0. But it also clearly satisfies the smaller degree polynomial equation
a  n = o.
Definition 5.5. The minimal polynomial of A G K""" is the polynomial o/ (X) of least
degree such that a (A) =0.
It can be shown that or(l) is essentially unique (unique if we force the coefficient
of the highest power of A to be +1, say; such a polynomial is said to be monic and we
generally write et (A) as a monic polynomial throughout the text). Moreover, it can also be
7 6 Chapt er 9. Ei g e n va l ue s and Ei genvect ors
7 r( A) has n roots, possibly repeated. These roots, as solutions of the determinant equation
are the eigenvalues of A and imply the singularity of the matrix A — XI, and hence further
guarantee the existence of corresponding nonzero eigenvectors.
Definition 9.5. The spectrum of A e C"
x
" is the set of all eigenvalues of A, i.e., the set of
all roots of its characteristic polynomial n(X). The spectrum of A is denoted A ( A) .
Let the eigenvalues of A e C"
x
" be denoted X\ ,..., X
n
. Then if we write (9.3) in the
form
76 Chapter 9. Eigenvalues and Eigenvectors
n(A) has n roots, possibly repeated. These roots, as solutions of the determinant equation
n(A) = det(A  AI) = 0, (9.3)
are the eigenvalues of A and imply the singularity of the matrix A  AI, and hence further
guarantee the existence of corresponding nonzero eigenvectors.
Definition 9.5. The spectrum of A E c
nxn
is the set of all eigenvalues of A, i.e., the set of
all roots of its characteristic polynomialn(A). The spectrum of A is denoted A(A).
Let the eigenvalues of A E en xn be denoted A], ... , An. Then if we write (9.3) in the
form
n(A) = det(A  AI) = (A]  A) ... (An  A) (9.4)
and set A = 0 in this identity, we get the interesting fact that det(A) = A] . A2 ... An (see
also Theorem 9.25).
If A E 1Ftnxn, then n(A) has real coefficients. Hence the roots of n(A), i.e., the
eigenvalues of A, must occur in complex conjugate pairs.
Example 9.6. Let a, f3 E 1Ft and let A = [ ~ f 3 !]. Then n(A) = A
2
 2aA + a
2
+ f32 and
A has eigenvalues a ± f3j (where j = i = R).
If A E 1Ftnxn, then there is an easily checked relationship between the left and right
eigenvectors of A and AT (take Hermitian transposes of both sides of (9.2». Specifically, if
y is a left eigenvector of A corresponding to A E A(A), then y is a right eigenvector of AT
corresponding to I E A(A). Note, too, that by elementary properties of the determinant,
we always have A(A) = A(AT), but that A(A) = A(A) only if A E 1Ftnxn.
Definition 9.7. If A is a root of multiplicity m of n(A), we say that A is an eigenvalue of A
of algebraic multiplicity m. The geometric multiplicity of A is the number of associated
independent eigenvectors = n  rank(A  AI) = dimN(A  AI).
If A E A(A) has algebraic multiplicity m, then I :::: dimN(A  AI) :::: m. Thus, if
we denote the geometric multiplicity of A by g, then we must have I :::: g :::: m.
Definition 9.8. A matrix A E 1Ft
nxn
is said to be defective if it has an eigenvalue whose
geometric multiplicity is not equal to (i.e., less than) its algebraic multiplicity. Equivalently,
A is said to be defective if it does not have n linearly independent (right or left) eigenvectors.
From the CayleyHamilton Theorem, we know that n(A) = O. However, it is pos
sible for A to satisfy a lowerorder polynomial. For example, if A = [ ~ ~ ] , then A sat
isfies (Je  1)2 = O. But it also clearly satisfies the smaller degree polynomial equation
(it.  1) ;;;:; 0
neftnhion ~ . ~ . Thll minimal polynomial Of A l::: l!if.nxn ix (hI' polynomilll a(A) oJ IPll.ft
degree such that a(A) ~ O.
It can be shown that a(Je) is essentially unique (unique if we force the coefficient
of the highest power of A to be + 1. say; such a polynomial is said to be monic and we
generally write a(A) as a monic polynomial throughout the text). Moreover, it can also be
9.1. Fundamental Definitions and Properties 77
shown that a (A.) divides every nonzero polynomial fi(k} for which ft (A) = 0. In particular,
a(X) divides n(X).
There is an algorithm to determine or ( A . ) directly ( without knowing eigenvalues and as
sociated eigenvector structure). Unfortunately, this algorithm, called the Bezout algorithm,
is numerically unstable.
Example 9.10. The above definitions are illustrated below for a series of matrices, each
of which has an eigenvalue 2 of algebraic multiplicity 4, i. e. , 7r( A ) = ( A — 2)
4
. We denote
the geometric multiplicity by g.
A t this point, one might speculate that g plus the degree of a must always be five.
Unfortunately, such is not the case. The matrix
Theorem 9.11. Let A e C«
x
"
ana
[
e
t A ., be an eigenvalue of A with corresponding right
eigenvector j c,. Furthermore, let yj be a left eigenvector corresponding to any A
;
e A ( A )
such that Xj =£ A . ,. Then yfx{ = 0.
Proof: Since Ax
t
= A ,*,,
9.1. Fundamental Definitions and Properties 77
shown that a(A) divides every nonzero polynomial f3(A) for which f3(A) = O. In particular,
a(A) divides n(A).
There is an algorithm to determine a(A) directly (without knowing eigenvalues and as
sociated eigenvector structure). Unfortunately, this algorithm, called the Bezout algorithm,
is numerically unstable.
Example 9.10. The above definitions are illustrated below for a series of matrices, each
of which has an eigenvalue 2 of algebraic multiplicity 4, i.e., n(A) = (A  2)4. We denote
the geometric multiplicity by g.
A  [ ~
0
! ] ha,"(A) ~ (A  2)' ""d g ~ 1.
2 I
 0
0 2
0 0 0
A ~ [ ~
0
~ ] ha< a(A) ~ (A  2)' ""d g ~ 2.
2
0 2
0 0
A ~ U
I 0
~ ] h'" a(A) ~ (A  2)2 ""d g ~ 3.
2 0
0 2
0 0
A ~ U
0 0
~ ] ha<a(A) ~ (A  2) andg ~ 4.
2 0
0 2
0 0
At this point, one might speculate that g plus the degree of a must always be five.
Unfortunately, such is not the case. The matrix
A ~ U
I 0
!]
2 0
0 2
0 0
has a(A) = (A  2)2 and g = 2.
Theorem 9.11. Let A E cc
nxn
and let Ai be an eigenvalue of A with corresponding right
eigenvector Xi. Furthermore, let Yj be a left eigenvector corresponding to any Aj E l\(A)
such that Aj 1= Ai. Then YY Xi = O.
Proof' Since AXi = AiXi,
(9.5)
78 Chapter 9. Eigenvalues and Eigenvectors
Similarly, since y" A = Xjyf,
Subtracting (9.6) from (9.5), we find 0 = (A., — A
y
)j ^j c, . Since A,, — A.
7
 ^ 0, we must have
yfxt =0.
The proof of Theorem 9.11 is very similar to two other fundamental and important
results.
Theorem 9.12. Let A e C"
x
" be Hermitian, i.e., A = A
H
. Then all eigenvalues of A must
be real.
Proof: Suppose (A ., x) is an arbitrary eigenvalue/eigenvector pair such that Ax = A .J C. Then
Taking Hermitian transposes in (9.7) yields
Using the fact that A is Hermitian, we have that Xx
H
x = Xx
H
x. However, since x is an
eigenvector, we have X
H
X /= 0, from which we conclude A . = A , i.e., A . is real. D
Theorem 9.13. Let A e C"
x
" be Hermitian and suppose X and / J L are distinct eigenvalues
of A with corresponding right eigenvectors x and z, respectively. Then x and z must be
orthogonal.
Proof: Premultiply the equation Ax = A.J C by Z
H
to get Z
H
Ax = X z
H
x . Take the Hermitian
transpose of this equation and use the facts that A is Hermitian and A . is real to get X
H
Az =
Xx
H
z. Premultiply the equation Az = i^z by X
H
to get X
H
Az = / ^X
H
Z = Xx
H
z. Since
A, ^ /z, we must have that X
H
z = 0, i.e., the two vectors must be orthogonal. D
Let us now return to the general case.
Theorem 9.14. Let A €. C
nxn
have distinct eigenvalues A ,
1 ?
. . . , A .
n
with corresponding
right eigenvectors x\,... ,x
n
. Then [x\,..., x
n
} is a linearly independent set. The same
result holds for the corresponding left eigenvectors.
Proof: For the proof see, for example, [21, p. 118].
If A e C
nx
" has distinct eigenvalues, and if A ., e A (A ), then by Theorem 9.11, jc, is
orthogonal to all yj's for which j ^ i. However, it cannot be the case that yf*x
t
= 0 as
well, or else x
f
would be orthogonal to n linearly independent vectors (by Theorem 9.14)
and would thus have to be 0, contradicting the fact that it is an eigenvector. Since yf*Xi ^ 0
for each i, we can choose the normalization of the *, 's, or the y, 's, or both, so that y
t
H
x; = 1
f or / € n.
78 Chapter 9. Eigenvalues and Eigenvectors
Similarly, since YY A = A j yy,
(9.6)
Subtracting (9.6) from (9.5), we find 0 = (Ai  Aj)YY xi. Since Ai  Aj =1= 0, we must have
YyXi = O. 0
The proof of Theorem 9.11 is very similar to two other fundamental and important
results.
Theorem 9.12. Let A E c
nxn
be Hermitian, i.e., A = AH. Then all eigenvalues of A must
be real.
Proof: Suppose (A, x) is an arbitrary eigenvalue/eigenvector pair such that Ax = AX. Then
(9.7)
Taking Hermitian transposes in (9.7) yields
Using the fact that A is Hermitian, we have that IXH x = AXH x. However, since x is an
eigenvector, we have xH x =1= 0, from which we conclude I = A, i.e., A is real. 0
Theorem 9.13. Let A E c
nxn
be Hermitian and suppose A and iJ are distinct eigenvalues
of A with corresponding right eigenvectors x and z, respectively. Then x and z must be
orthogonal.
Proof: Premultiply the equation Ax = AX by ZH to get ZH Ax = AZ
H
x. Take the Hermitian
transpose of this equation and use the facts that A is Hermitian and A is real to get x H Az =
AxH z. Premultiply the equation Az = iJZ by x
H
to get x
H
Az = iJXH Z = AXH z. Since
A =1= iJ, we must have that x
H
z = 0, i.e., the two vectors must be orthogonal. 0
Let us now return to the general case.
Theorem 9.14. Let A E c
nxn
have distinct eigenvalues AI, ... , An with corresponding
right eigenvectors XI, ... , x
n
• Then {XI, ... , x
n
} is a linearly independent set. The same
result holds for the corresponding left eigenvectors.
Proof: For the proof see, for example, [21, p. 118]. 0
If A E c
nxn
has distinct eigenvalues, and if Ai E A(A), then by Theorem 9.11, Xi is
orthogonal to all y/s for which j =1= i. However, it cannot be the case that Yi
H
Xi = 0 as
well, or else Xi would be orthogonal to n linearly independent vectors (by Theorem 9.14)
and would thus have to be 0, contradicting the fact that it is an eigenvector. Since yr Xi =1= 0
for each i, we can choose the normalization of the Xi'S, or the Yi 's, or both, so that Yi
H
Xi = 1
for i E !1.
9.1. Fundament al Def i ni t i o ns and Properties 79
Theorem 9.15. Let A e C"
x
" have distinct eigenvalues A .I , ..., A .
n
and let the correspond
ing right eigenvectors form a matrix X = [x\, ..., x
n
]. Similarly, let Y — [y\, ..., y
n
]
be the matrix of corresponding left eigenvectors. Furthermore, suppose that the left and
right eigenvectors have been normalized so that yf
1
Xi = 1, / en. Finally, let A =
di ag ( A ,j , . . . , X
n
) e W
txn
. Then A J C, = A ., * /, / e n, can be written in matrix form as
Example 9.16. Let
Then n(X) = det( A  A ./) =  (A .
3
+ 4A .
2
+ 9 A . + 10) =  (A . + 2 )(A .
2
+ 2 A , + 5), from
which we find A ( A ) = { — 2 , — 1 ± 2 j } . We can now find the right and left eigenvectors
corresponding to these eigenvalues.
For A  i = — 2 , solve the 3 x 3 linear system (A — (—2 } I)x\ = 0 to get
while y^Xj = 5,
;
, / en, y' e n, is expressed by the equation
These matrix equations can be combined to yield the following matrix factorizations:
and
Note that one component of ;ci can be set arbitrarily, and this then determines the other two
(since di mA /XA — ( — 2 )7) = 1). To get the corresponding left eigenvector y\, solve the
linear system y\(A + 2 1) = 0 to get
This time we have chosen the arbitrary scale factor for y\ so that y f x \ = 1.
For A
2
= — 1 + 2 j , solve the linear system (A — (— 1 + 2 j )I)x
2
= 0 to get
9.1. Fundamental Definitions and Properties 79
Theorem 9.15. Let A E en xn have distinct eigenvalues A I, ... , An and let the correspond
ing right eigenvectors form a matrix X = [XI, ... , xn]. Similarly, let Y = [YI,"" Yn]
be the matrix of corresponding left eigenvectors. Furthermore, suppose that the left and
right eigenvectors have been normalized so that YiH Xi = 1, i E !!:: Finally, let A =
diag(AJ, ... , An) E ]Rnxn. Then AXi = AiXi, i E !!, can be written in matrixform as
AX=XA (9.8)
while YiH X j = oij, i E!!, j E !!, is expressed by the equation
yHX = I.
(9.9)
These matrix equations can be combined to yield the following matrix factorizations:
and
Example 9.16. Let
XlAX = A = yRAX
n
A = XAX
I
= XAyH = LAixiyr
2
5
3
3
2
i=1
~ ] .
4
(9.10)
(9.11)
Then rr(A) = det(A  AI) = (A
3
+ 4A2 + 9)" + 10) = ()" + 2)(),,2 + 2)" + 5), from
which we find A(A) = {2, 1 ± 2j}. We can now find the right and left eigenvectors
corresponding to these eigenvalues.
For Al = 2, solve the 3 x 3 linear system (A  (2)l)xI = 0 to get
Note that one component of XI can be set arbitrarily, and this then determines the other two
(since dimN(A  (2)1) = 1). To get the corresponding left eigenvector YI, solve the
linear system yi (A + 21) = 0 to get
This time we have chosen the arbitrary scale factor for YJ so that yi XI = 1.
For A2 = 1 + 2j, solve the linear system (A  (1 + 2j) I)x2 = 0 to get
[
3+ j ]
X2 = 3 ~ / .
80 Chapter 9. Eigenvalues and Eigenvectors
Solve the linear system y" (A — (1 + 27')/) = 0 and normalize y>
2
so that y"x
2
= 1 to get
For X T , = — 1 — 2 j, we could proceed to solve linear systems as for A.
2
. However, we
can also note that x$ =x
2
' and yi = jj. To see this, use the fact that A, 3 = A.2 and simply
conjugate the equation A;c
2
— ^.2 *2 to get Ax^ = ^2 X 2  A similar argument yields the result
for left eigenvectors.
Now define the matrix X of right eigenvectors:
It is then easy to verify that
Other results in Theorem 9.15 can also be verified. For example,
Finally, note that we could have solved directly only for *i and x
2
(and X T , = x
2
). Then,
instead of determining the j,'s directly, we could have found them instead by computing
X ~
l
and reading off its rows.
Example 9.17. Let
Then 7r(A.) = det(A  A./) = (A
3
+ 8A
2
+ 19X + 12) = (A. + 1)(A. + 3)(A, + 4),
from which we find A (A) = { —1, —3, —4}. Proceeding as in the previous example, it is
straightforward to compute
and
80 Chapter 9. Eigenvalues and Eigenvectors
Solve the linear system yf (A  ( I + 2 j) I) = 0 and nonnalize Y2 so that yf X2 = 1 to get
For A3 = I  2j, we could proceed to solve linear systems as for A2. However, we
can also note that X3 = X2 and Y3 = Y2. To see this, use the fact that A3 = A2 and simply
conjugate the equation AX2 = A2X2 to get AX2 = A2X2. A similar argument yields the result
for left eigenvectors.
Now define the matrix X of right eigenvectors:
3+j 3
j
]
3j 3+j .
2 2
It is then easy to verify that
.!.=.L
4
l+j
4
!.±1
4
.!.=.L
4
Other results in Theorem 9.15 can also be verified. For example,
[
2 0
XIAX=A= 0 1+2j
o 0
Finally, note that we could have solved directly only for XI and X2 (and X3 = X2). Then,
instead of detennining the Yi'S directly, we could have found them instead by computing
XI and reading off its rows.
Example 9.17. Let
A = .
o 3
Then Jl"(A) = det(A  AI) = _(A
3
+ 8A
2
+ 19A + 12) = (A + I)(A + 3)(A + 4),
from which we find A(A) = {I, 3, 4}. Proceeding as in the previous example, it is
gtruightforw!U"d to
I
i ]
0
I
and
1 2 1
] y'
3 0 3
2 2 2
9.1. Fundamental Definitions and Properties 81
We also have X~
l
AX = A = di ag( —1, —3, —4 ) , which is equivalent to the dyadic expan
sion
Theorem 9.18. Eigenvalues (but not eigenvectors) are invariant under a similarity trans
formation T.
Proof: Suppose (A, jc) is an eigenvalue/eigenvector pair such that Ax = Xx. Then, since T
is nonsingular, we have the equivalent statement (T~
l
AT)(T~
l
x) = X ( T ~
l
x ) , from which
the theorem statement follows. For left eigenvectors we have a similar statement, namely
y
H
A = Xy
H
ifandon\yif(T
H
y)
H
(T~
1
AT) =X(T
H
yf. D
Remark 9.19. If / is an analytic function (e.g., f ( x ) is a polynomial, or e
x
, or sin*,
or, in general, representable by a power series X^^o
fl
n*
n
)> then it is easy to show that
the eigenvalues of /( A) (defined as X^o^A") are /( A) , but /( A) does not necessarily
have all the same eigenvectors (unless, say, A is diagonalizable). For example, A = T
0 O
j
has only one right eigenvector corresponding to the eigenvalue 0, but A
2
= f
0 0
1 has two
independent right eigenvectors associated with the eigenvalue 0. What is true is that the
eigenvalue/eigenvector pair (A, x) maps to ( /( A) , x) but not conversely.
The following theorem is useful when solving systems of linear differential equations.
Details of how the matrix exponential e'
A
is used to solve the system x = Ax are the subject
of Chapter 11.
Theorem 9.20. Let A e R"
xn
and suppose X~~
1
AX — A, where A is diagonal. Then
9.1. Fundamental Definitions and Properties 81
We also have XI AX = A = diag( 1, 3, 4), which is equivalent to the dyadic expan
sion
3
A = LAixiyr
i=1
W j 0
+(4) [ ; ]
1
 
3
(I) [
I I I
J + (3) [
I
0
I
] + (4) [
I I I
l
(;
3
(;
2
2 3
3
3
I 2 I
0 0 0
I I I
3 3 3
3
3
3
I I I
I
0
I
I I I
(;
3
(;
2
2
3
3
3
Theorem 9.18. Eigenvalues (but not eigenvectors) are invariant under a similarity trans
formation T.
Proof: Suppose (A, X) is an eigenvalue/eigenvector pair such that Ax = AX. Then, since T
is nonsingular, we have the equivalent statement (T
I
AT)(T
I
x) = A(T
I
x), from which
the theorem statement follows. For left eigenvectors we have a similar statement, namely
yH A = AyH if and only if (T
H
y)H (T
1
AT) = A(T
H
y)H. D
Remark 9.19. If f is an analytic function (e.g., f(x) is a polynomial, or eX, or sinx,
or, in general, representable by a power series anxn), then it is easy to show that
the eigenvalues of f(A) (defined as are f(A), but f(A) does not necessarily
have all the same eigenvectors (unless, say, A is diagonalizable). For example, A = 6 ]
has only one right eigenvector corresponding to the eigenvalue 0, but A
2
= ] has two
independent right eigenvectors associated with the eigenvalue o. What is true is that the
eigenvalue/eigenvector pair (A, x) maps to (f(A), x) but not conversely.
The following theorem is useful when solving systems of linear differential equations.
Details of how the matrix exponential etA is used to solve the system i = Ax are the subject
of Chapter 11.
Theorem 9.20. Let A E jRnxn and suppose XI AX = A, where A is diagonal. Then
n
= LeA,txiYiH.
i=1
82 Chapter 9. Eigenvalues and Eigenvectors
Proof: Starting from the definition, we have
The following corollary is immediate from the theorem upon setting t = I.
Corollary 9.21. If A e R
nx
" is diagonalizable with eigenvalues A ., , /' en, and right
eigenvectors x
t
•, / € n_, then e
A
has eigenvalues e
X i
, i € n_, and the same eigenvectors.
There are extensions to Theorem 9.20 and Corollary 9.21 for any function that is
analytic on the spectrum of A , i.e., f ( A ) = X f ( A ) X ~
l
= Xdi ag ( / ( A . i ) , . . . , f ( X
t t
) ) X ~
l
.
It is desirable, of course, to have a version of Theorem 9.20 and its corollary in which
A is not necessarily diagonalizable. It is necessary first to consider the notion of Jordan
canonical form, from which such a result is then available and presented later in this chapter.
9.2 Jordan Canonical Form
Theorem 9.22.
1. Jordan Canonical Form (/CF): For all A e C"
x
" with eigenvalues X\,..., k
n
e C
(not necessarily distinct), there exists X € C^
x
" such that
where each of the Jordan block matrices / i , . . . , J
q
is of the form
82 Chapter 9. Eigenvalues and Eigenvectors
Proof' Starting from the definition, we have
n
= LeA;IXiYiH. 0
i=1
The following corollary is immediate from the theorem upon setting t = I.
Corollary 9.21. If A E ]Rn xn is diagonalizable with eigenvalues Ai, i E ~ , and right
eigenvectors Xi, i E ~ , then e
A
has eigenvalues e
A
" i E ~ , and the same eigenvectors.
There are extensions to Theorem 9.20 and Corollary 9.21 for any function that is
analytic on the spectrum of A, i.e., f(A) = Xf(A)X
I
= Xdiag(J(AI), ... , f(An))X
I
.
It is desirable, of course, to have a version of Theorem 9.20 and its corollary in which
A is not necessarily diagonalizable. It is necessary first to consider the notion of Jordan
canonical form, from which such a result is then available and presented later in this chapter.
9.2 Jordan Canonical Form
Theorem 9.22.
I. lordan Canonical Form (JCF): For all A E c
nxn
with eigenvalues AI, ... , An E C
(not necessarily distinct), there exists X E c ~ x n such that
XI AX = 1 = diag(ll, ... , 1q), (9.12)
where each of the lordan block matrices 1
1
, ••• , 1q is of the form
Ai
0 o
0
Ai
0
1i =
Ai
(9.13)
o
Ai
o o Ai
9.2. Jordan Canonical Form 83
2. Real Jordan Canonical Form: For all A € R
nx
" with eigenvalues Xi, . . . , X
n
(not
necessarily distinct), there exists X € R"
xn
such that
where each of the Jordan block matrices J\, ..., J
q
is of the form
in the case of real eigenvalues A., e A (A), and
where M
;
= [ _»' ^ 1 and I
2
= \
0
A in the case of complex conjugate eigenvalues
a
i
±jp
i
eA(A
>
).
Proof: For the proof see, for example, [21, pp. 120124]. D
Transformations like T = I " _, "•{ "] allow us to go back and forth between a real JCF
and its complex counterpart:
For nontrivial Jordan blocks, the situation is only a bit more complicated. With
9.2. Jordan Canonical Form 83
and L;=1 ki = n.
2. Real Jordan Canonical Form: For all A E jRnxn with eigenvalues AI, ... , An (not
necessarily distinct), there exists X E such that
(9.14)
where each of the Jordan block matrices 11, ... , 1q is of the form
where Mi = [ ] and h = [6 in the case of complex conjugate eigenvalues
(Xi ± jfJi E A(A).
Proof: For the proof see, for example, [21, pp. 120124]. 0
Transformations like T = [ _  { ] allow us to go back and forth between a real JCF
and its complex counterpart:
TI [ (X + jfJ O. ] T = [ (X fJ ] = M.
o (X  JfJ fJ (X
For nontrivial Jordan blocks, the situation is only a bit more complicated. With
1
o
j
o
j
o
1 o 0 '
o j 1
84 Chapter 9. Ei genval ues and Eigenvectors
it is easily checked that
Definition 9.23. The characteristic polynomials of the Jordan blocks defined in Theorem
9 . 2 2 are called the elementary divisors or invariant factors of A.
Theorem 9.24. The characteristic polynomial of a matrix is the product of its elementary
divisors. The minimal polynomial of a matrix is the product of the elementary divisors of
highest degree corresponding to distinct eigenvalues.
Theorem 9.25. Let A e C"
x
" with eigenvalues AI, . . . , X
n
. Then
Proof:
1. From Theorem 9.22 we have that A = X J X ~
l
. Thus,
det(A) = det(XJX
1
) = det(7) = ] ~ [ "
=l
A,  .
2. Again, from Theorem 9.22 we have that A = X J X ~
l
. Thus,
Tr(A) = Tr(XJX~
l
) = TrC/X"
1
*) = Tr(/) = £"
=1
A., . D
Example 9.26. Suppose A e E
7x7
is known to have 7r(A) = (A.  1)
4
(A  2)
3
and
a (A.) = (A. — I)
2
(A. — 2)
2
. Then A has two possible JCFs (not counting reorderings of the
diagonal blocks):
Note that 7
(1)
has elementary divisors (A  I )
2
, (A.  1), (A.  1), (A,  2)
2
, and (A  2),
while /(
2)
has elementary divisors (A  I )
2
, (A  I )
2
, (A  2)
2
, and (A  2).
84 Chapter 9. Eigenvalues and Eigenvectors
it is easily checked that
[ "+ jfi
0 0
] T ~ [ ~
T
I
0
et + jf3 0 0
h
l
0 0 et  jf3 M
0 0 0 et  jf3
Definition 9.23. The characteristic polynomials of the Jordan blocks defined in Theorem
9.22 are called the elementary divisors or invariant factors of A.
Theorem 9.24. The characteristic polynomial of a matrix is the product of its elementary
divisors. The minimal polynomial of a matrix is the product of the elementary divisors of
highest degree corresponding to distinct eigenvalues.
Theorem 9.25. Let A E c
nxn
with eigenvalues AI, .. " An. Then
n
1. det(A) = nAi.
i=1
n
2. Tr(A) = 2,)i.
i=1
Proof:
1. From Theorem 9.22 we have that A = X J XI. Thus,
det(A) = det(X J XI) = det(J) = n7=1 Ai.
2. Again, from Theorem 9.22 we have that A = X J XI. Thus,
Tr(A) = Tr(X J XI) = Tr(JX
1
X) = Tr(J) = L7=1 Ai. 0
Example 9.26. Suppose A E lR.
7x7
is known to have :rr(A) = (A  1)4(A  2)3 and
et(A) = (A  1)2(A  2)2. Then A has two possible JCFs (not counting reorderings of the
diagonal blocks):
1 0 0 0 0 0 1 0 0 0 0 0
0 1 0 0 0 0 0 0 0 0 0 0 0
0 0 1 0 0 0 0 0 0 1 I 0 0 0
J(l) =
0 0 0 1 0 0 0
and f2) =
0 0 0 1 0 0 0
0 0 0 0 2 1 0 0 0 0 0 2 1 0
0 0 0 0 0 2 0 0 0 0 0 0 2 0
0 0 0 0 0 0 2
0 0 0 0 0 0 2
Note that J(l) has elementary divisors (A  1)z, (A  1), (A  1), (A  2)2, and (A  2),
while J(2) has elementary divisors (A  1)2, (A  1)2, (A  2)2, and (A  2).
9.3. Determination of the JCF &5
Example 9.27. Knowing T T (A.), a ( A ) , and rank (A — A,,7) for distinct A., is not sufficient to
determine the JCF of A uniquely. T he matrices
both have 7r( A . ) = (A. — a) , a( A . ) = (A. — a) , and rank( A — al) = 4, i.e., three eigen
vectors.
9.3 Determination of the JCF
T he first critical item of information in determining the JCF of a matrix A e W
lxn
is its
number of eigenvectors. For each distinct eigenvalue A , , , the associated number of linearly
independent right (or left) eigenvectors is given by dim A^(A — A.,7) = n — rank( A — A.
(
7).
T he straightforward case is, of course, when X, is simple, i.e., of algebraic multiplicity 1; it
then has precisely one eigenvector. T he more interesting (and difficult) case occurs when
A, is of algebraic multiplicity greater than one. For example, suppose
T hen
has rank 1, so the eigenvalue 3 has two eigenvectors associated with it. If we let [^i £2 &]
T
denote a solution to the linear system (A — 3/) £ = 0, we find that 2£
2
+ £3=0. T hus, both
are eigenvectors (and are independent). T o get a third vector JC3 such that X = [x\ KJ_ XT,]
reduces A to JCF, we need the notion of principal vector.
Definition 9.28. Let A e C"
xn
(or R"
x
") . Then x is a right principal vector of degree k
associated with X e A (A) if and only if (A  XI)
k
x = 0 and (A  U}
k
~
l
x ^ 0.
Remark 9.29.
1. An analogous definition holds for a left principal vector of degree k.
9.3. Determination of the JCF 85
Example 9.27. Knowing rr(A), a(A), and rank(A  Ai l) for distinct Ai is not sufficient to
determine the JCF of A uniquely. The matrices
a 0 0 0 0 0 a 0 0 0 0 0
0 a 0 0 0 0 0 a 0 0 0 0
0 0 a 0 0 0 0 0 0 a 0 0 0 0
Al=
0 0 0 a 0 0
A2 =
0 0 0 a 0 0
0 0 0 0 a 0 0 0 0 0 0 a 0
0 0 0 0 0 a 1 0 0 0 0 0 a 0
0 0 0 0 0 0 a 0 0 0 0 0 0 a
both have rr(A) = (A  a)7, a(A) = (A  a)\ and rank(A  al) = 4, i.e., three eigen
vectors.
9.3 Determination of the JCF
The first critical item of information in determining the JCF of a matrix A E ]R.nxn is its
number of eigenvectors. For each distinct eigenvalue Ai, the associated number of linearly
independent right (or left) eigenvectors is given by dimN(A  A;l) = n  rank(A  A;l).
The straightforward case is, of course, when Ai is simple, i.e., of algebraic multiplicity 1; it
then has precisely one eigenvector. The more interesting (and difficult) case occurs when
Ai is of algebraic multiplicity greater than one. For example, suppose
[3 2
n
A = 0 3
o 0
Then
A3I= U
2 I]
o 0
o 0
has rank 1, so the eigenvalue 3 has two eigenvectors associated with it. If we let
denote a solution to the linear system (A  = 0, we find that + = O. Thus, both
are eigenvectors (and are independent). To get a third vector X3 such that X = [Xl X2 X3]
reduces A to JCF, we need the notion of principal vector.
Definition 9.28. Let A E c
nxn
(or ]R.nxn). Then X is a right principal vector of degree k
associated with A E A(A) ifand only if(A  ulx = 0 and (A  AI)kl x i= o.
Remark 9.29.
1. An analogous definition holds for a left principal vector of degree k.
86 Chapter 9. Eigenvalues and Eigenvectors
2. The phrase "of grade k" is often used synonymously with "of degree k."
3. Principal vectors are sometimes also called generalized eigenvectors, but the latter
term will be assigned a much different meaning in Chapter 12.
4. The case k = 1 corresponds to the "usual" eigenvector.
5. A right (or left) principal vector of degree k is associated with a Jordan block ji of
dimension k or larger.
9.3.1 Theoretical computation
To motivate the development of a procedure for determining principal vectors, consider a
2x2 Jordan block {h
0
h1. Denote by x
(1)
and x
(2)
the two columns of a matrix X e R
2
,x
2
that reduces a matrix A to this JCF. Then the equation AX = XJ can be written
The first column yields the equation Ax
(1)
= hx
(1)
which simply says that x
(1)
is a right
eigenvector. The second column yields the following equation for x
(2)
, the principal vector
of degree 2:
If we premultiply (9.17) by (A  XI), we find (A  XI )
z
x
( 2 )
= (A  XI)x
w
= 0. Thus,
the definition of principal vector is satisfied.
This suggests a "general" procedure. First, determine all eigenvalues of A e R"
x
"
(or C
nxn
). Then for each distinct X e A (A) perform the following:
1. Solve
This step finds all the eigenvectors (i.e., principal vectors of degree 1) associated with
X. The number of eigenvectors depends on the rank of A — XI. For example, if
rank(A — XI) = n — 1, there is only one eigenvector. If the algebraic multiplicity of
X is greater than its geometric multiplicity, principal vectors still need to be computed
from succeeding steps.
2. For each independent jc
(1)
, solve
The number of linearly independent solutions at this step depends on the rank of
(A — XI )
2
. If, for example, this rank is n — 2 , there are two linearly independent
solutions to the homogeneous equation (A — XI)
2
x^ = 0. One of these solutions
is, of course, x
(l)
(^ 0), since (A  XI )
2
x
( l )
= (A  XI)0 = 0. The other solution
is the desired principal vector of degree 2. (It may be necessary to take a linear
combination of jc
(1)
vectors to get a righthand side that is in 7£(A — XI). See, for
example, Exercise 7.)
86 Chapter 9. Eigenvalues and Eigenvectors
2. The phrase "of grade k" is often used synonymously with "of degree k."
3. Principal vectors are sometimes also called generalized eigenvectors, but the latter
term will be assigned a much different meaning in Chapter 12.
4. The case k = 1 corresponds to the "usual" eigenvector.
S. A right (or left) principal vector of degree k is associated with a Jordan block J; of
dimension k or larger.
9.3.1 Theoretical computation
To motivate the development of a procedure for determining principal vectors, consider a
2 x 2 Jordan block [ ~ i]. Denote by x(l) and x(2) the two columns of a matrix X E l R ~ X 2
that reduces a matrix A to this JCF. Then the equation AX = X J can be written
A [x(l) x(2)] = [x(l) X(2)] [ ~ ~ J.
The first column yields the equation Ax(!) = AX(!), which simply says that x(!) is a right
eigenvector. The second column yields the following equation for x(2), the principal vector
of degree 2:
(A  A/)x(2) = x(l). (9.17)
If we premultiply (9.17) by (A  AI), we find (A  A1)2 x(2) = (A  A1)X(l) = O. Thus,
the definition of principal vector is satisfied.
This suggests a "general" procedure. First, determine all eigenvalues of A E lR
nxn
(or c
nxn
). Then for each distinct A E A(A) perform the following:
1. Solve
(A  A1)X(l) = O.
This step finds all the eigenvectors (i.e., principal vectors of degree I) associated with
A. The number of eigenvectors depends on the rank of A  AI. For example, if
rank(A  A/) = n  1, there is only one eigenvector. If the algebraic multiplicity of
A is greater than its geometric multiplicity, principal vectors still need to be computed
from succeeding steps.
2. For each independent x(l), solve
(A  A1)x(2) = x(l).
The number of linearly independent solutions at this step depends on the rank of
(A  uf. If, for example, this rank is n  2, there are two linearly independent
solutions to the homogeneous equation (A  AI)2x (2) = o. One of these solutions
is, of course, x(l) (1= 0), since (A  'A1)
2
x(l) = (A  AI)O = o. The other solution
is the desired principal vector of degree 2. (It may be necessary to take a linear
combination of x(l) vectors to get a righthand side that is in R(A  AI). See, for
example, Exercise 7.)
9. 3. Determination of the JCF 87
4. Continue in this way until the total number of independent eigenvectors and principal
vectors is equal to the algebraic multiplicity of A.
Unfortunately, this naturallooking procedure can fail to find all Jordan vectors. For
more extensive treatments, see, for example, [20] and [21]. Determination of eigenvectors
and principal vectors is obviously very tedious for anything beyond simple problems (n = 2
or 3, say). Attempts to do such calculations in finiteprecision floatingpoint arithmetic
generally prove unreliable. There are significant numerical difficulties inherent in attempting
to compute a JCF, and the interested student is strongly urged to consult the classical and very
readable [8] to learn why. Notice that highquality mathematical software such as MATLAB
does not offer a jcf command, although a jordan command is available in MATLAB'S
Symbolic Toolbox.
Theorem 9.30. Suppose A e C
kxk
has an eigenvalue A, of algebraic multiplicity k and
suppose further that rank(A — AI) = k — 1. Let X = [ x
( l )
, . . . , x
(k)
], where the chain of
vectors x(i) is constructed as above. Then
Theorem 9.31. (x
( 1)
, . . . , x
(k)
} is a linearly independent set.
Theorem 9.32. Principal vectors associated with different Jordan blocks are linearly inde
pendent.
Example 9.33. Let
The eigenvalues of A are A1 = 1, h2 = 1, and h
3
= 2. First, find the eigenvectors associated
with the distinct eigenvalues 1 and 2.
,(1)
(A  2/)x3(1) = 0 yields
3. For each independent x
(2)
from step 2, solve
9.3. Determination of the JCF 87
3. For each independent X(2) from step 2, solve
(A  AI)x(3) = x(2).
4. Continue in this way until the total number of independent eigenvectors and principal
vectors is equal to the algebraic multiplicity of A.
Unfortunately, this naturallooking procedure can fail to find all Jordan vectors. For
more extensive treatments, see, for example, [20] and [21]. Determination of eigenvectors
and principal vectors is obviously very tedious for anything beyond simple problems (n = 2
or 3, say). Attempts to do such calculations in finiteprecision floatingpoint arithmetic
generally prove unreliable. There are significant numerical difficulties inherent in attempting
to compute a JCF, and the interested student is strongly urged to consult the classical and very
readable [8] to learn why. Notice that highquality mathematical software such as MATLAB
does not offer a j cf command, although a j ardan command is available in MATLAB's
Symbolic Toolbox.
Theorem 9.30. Suppose A E C
kxk
has an eigenvalue A of algebraic multiplicity k and
suppose further that rank(A  AI) = k  1. Let X = [x(l), ... , X(k)], where the chain of
vectors x(i) is constructed as above. Then
Theorem 9.31. {x(l), ... , X(k)} is a linearly independent set.
Theorem 9.32. Principal vectors associated with different Jordan blocks are linearly inde
pendent.
Example 9.33. Let
1 ;].
002
The eigenvalues of A are AI = I, A2 = 1, and A3 = 2. First, find the eigenvectors associated
with the distinct eigenvalues 1 and 2.
(A  = 0 yields
88 Chapter 9. Eigenvalues and Eigenvectors
(A l/)x,
(1)
=0 yields
Then it is easy to check that
9.3.2 On the +1 's in JCF blocks
In this subsection we show that the nonzero superdiagonal elements of a JCF need not be
1 's but can be arbitrary — so long as they are nonzero. For the sake of defmiteness, we
consider below the case of a single Jordan block, but the result clearly holds for any JCF.
Supposed A € R
nxn
and
Let D = d i a g ( d 1 , . . . , d
n
) be a nonsingular "scaling" matrix. Then
To find a principal vector of degree 2 associated with the multiple eigenvalue 1, solve
( A – l/)x,
(2)
= x,
(1)
toeet
Now let
88 Chapter 9. Eigenvalues and Eigenvectors
(A  11)x?J = 0 yields
To find a principal vector of degree 2 associated with the multiple eigenvalue 1, solve
(A  1I)xl
2
) = xiI) to get
[ 0 ]
(2)
x, = ~ .
Now let
xl" xl"] ~ [ ~
0 5
l
X = [xiI) 1 3
0
Then it is easy to check that
X  ' ~ U
0
5 ] [ I
n
1
i and XlAX = ~
1
0 0
9.3.2 On the +1 's in JCF blocks
In this subsection we show that the nonzero superdiagonal elements of a JCF need not be
1 's but can be arbitrary  so long as they are nonzero. For the sake of definiteness, we
consider below the case of a single Jordan block, but the result clearly holds for any JCF.
Suppose A E jRnxn and
Let D = diag(d" ... , d
n
) be a nonsingular "scaling" matrix. Then
A
4l.
0 0
d,
0
)...
!b.
0
d,
D'(X' AX)D = D' J D = j =
A
d
n

I
0
d
n

2
A
d
n
d
n

I
0 0
)...
9.4. Geometric Aspects of the JCF 89
Appropriate choice of the di 's then yields any desired nonzero superdiagonal elements.
This result can also be interpreted in terms of the matrix X = [x\,..., x
n
] of eigenvectors
and principal vectors that reduces A to its JCF. Specifically, J is obtained from A via the
similarity transformation XD = \d\x\,..., d
n
x
n
}.
In a similar fashion, the reverseorder identity matrix (or exchange matrix)
9.4 Geometric Aspects of the JCF
Note that di mM( A — A.,/ )
w
= «,.
Definition 9.35. Let V be a vector space over F and suppose A : V —>• V is a linear
transformation. A subspace S c V is Ainvariant if AS c S, where AS is defined as the
set {As : s e S}.
can be used to put the superdiagonal elements in the subdiagonal instead if that is desired:
The matrix X that reduces a matrix A e IR"
X
" (or C
nxn
) to a JCF provides a change of basis
with respect to which the matrix is diagonal or block diagonal. It is thus natural to expect an
associated direct sum decomposition of R. Such a decomposition is given in the following
theorem.
Theorem 9.34. Suppose A e R"
x
" has characteristic polynomial
and minimal polynomial
with A i , . . . , A.
m
distinct. Then
9.4. Geometric Aspects of the JCF 89
Appropriate choice of the di's then yields any desired nonzero superdiagonal elements.
This result can also be interpreted in terms of the matrix X = [x[, ... ,x
n
] of eigenvectors
and principal vectors that reduces A to its lCF. Specifically, j is obtained from A via the
similarity transformation XD = [d[x[, ... , dnxn].
In a similar fashion, the reverseorder identity matrix (or exchange matrix)
0 0 I
0
p = pT = p[ =
(9.18)
0 1
I 0 0
can be used to put the superdiagonal elements in the subdiagonal instead if that is desired:
A I 0 0 A 0 0
0 A 0 A 0
p[
A
p=
0 1 A
0
A I A 0
0 0 A 0 0 A
9.4 Geometric Aspects of the JCF
The matrix X that reduces a matrix A E jH.nxn (or c
nxn
) to a lCF provides a change of basis
with respect to which the matrix is diagonal or block diagonal. It is thus natural to expect an
associated direct sum decomposition of jH.n. Such a decomposition is given in the following
theorem.
Theorem 9.34. Suppose A E jH.nxn has characteristic polynomial
n(A) = (A  A[)n) ... (A  Amt
m
and minimal polynomial
a(A) = (A  A[)V) '" (A  Am)V
m
with A I, ... , Am distinct. Then
jH.n = N(A  AlIt) E6 ... E6 N(A  AmItm
= N (A  A 1 I) v) E6 ... E6 N (A  Am I) Vm .
Note that dimN(A  AJ)Vi = ni.
Definition 9.35. Let V be a vector space over IF and suppose A : V + V is a linear
transformation. A subspace S ~ V is A invariant if AS ~ S, where AS is defined as the
set {As: s E S}.
90 Chapter 9. Eigenvalues and Eigenvectors
If V is taken to be R" over R, and S e R"
x
* is a matrix whose columns s\,..., s/t
span a /^dimensional subspace <S, i.e., K(S) = <S, then <S is Ainvariant if and only if there
exists M eR
kxk
such that
This follows easily by comparing the /th columns of each side of (9.19):
Example 9.36. The equation Ax = A* = xA defining a right eigenvector x of an eigenvalue
X says that * spans an Ainvariant subspace (of dimension one).
Example 9.37. Suppose X block diagonalizes A, i.e.,
Rewriting in the form
we have that A A, = A", /,, / = 1, 2, so the columns of A, span an Amvanant subspace.
Theorem 9.38. Suppose A e E"
x
".
7. Let p(A) = «o/ + o?i A + • • • + <x
q
A
q
be a polynomial in A. Then N(p(A)) and
7£(p(A)) are Ainvariant.
2. S isAinvariant if and only ifS
1
 is A
T
invariant.
Theorem 9.39. If V isa vector space over F such that V = N\ ® • • • 0 N
m
, where each
A// isAinvariant, then a basisfor V can be chosen with respect to which A hasa block
diagonal representation.
The Jordan canonical form is a special case of the above theorem. If A has distinct
eigenvalues A,, as in Theorem 9.34, we could choose bases for N(A — A.,/)"' by SVD, for
example (note that the power n, could be replaced by v,). We would then get a block diagonal
representation for A with full blocks rather than the highly structured Jordan blocks. Other
such "canonical" forms are discussed in text that follows.
Suppose A" = [ X i , . . . , X
m
] e R"
n
xn
is such that X ^AX = diag(7i,. . . , J
m
), where
each Ji = diag(/,i,..., //*,.) and each /,* is a Jordan block corresponding to A, e A(A).
We could also use other block diagonal decompositions (e.g., via SVD), but we restrict our
attention here to only the Jordan block case. Note that A A", = A*, /,, so by (9.19) the columns
of A", (i.e., the eigenvectors and principal vectors associated with A.,) span an Ainvariant
subspace of W.
Finally, we return to the problem of developing a formula for e'
A
in the case that A
is not necessarily diagonalizable. Let 7, € C"
x
"' be a Jordan basis for N(A
T
— A.,/)"' .
Equivalently, partition
90 Chapter 9. Eigenvalues and Eigenvectors
If V is taken to be ]Rn over Rand S E ]Rn xk is a matrix whose columns SI, ... , Sk
span a kdimensional subspace S, i.e., R(S) = S, then S is Ainvariant if and only if there
exists M E ]Rkxk such that
AS = SM. (9.19)
This follows easily by comparing the ith columns of each side of (9.19):
Example 9.36. The equation Ax = AX = x A defining a right eigenvector x of an eigenvalue
A says that x spans an Ainvariant subspace (of dimension one).
Example 9.37. Suppose X block diagonalizes A, i.e.,
XI AX = [ ~ J
2
].
Rewriting in the form
~ J,
we have that AX
i
= X;li, i = 1,2, so the columns of Xi span an Ainvariant subspace.
Theorem 9.38. Suppose A E ]Rnxn.
1. Let peA) = CloI + ClIA + '" + ClqAq be a polynomial in A. Then N(p(A)) and
R(p(A)) are Ainvariant.
2. S is A invariant if and only if S 1. is A T invariant.
Theorem 9.39. If V is a vector space over IF such that V = NI EB ... EB N
m
, where each
N; is Ainvariant, then a basis for V can be chosen with respect to which A has a block
diagonal representation.
The Jordan canonical form is a special case of the above theorem. If A has distinct
eigenvalues Ai as in Theorem 9.34, we could choose bases for N(A  Ai/)n, by SVD, for
example (note that the power ni could be replaced by Vi). We would then get a block diagonal
representation for A with full blocks rather than the highly structured Jordan blocks. Other
such "canonical" forms are discussed in text that follows.
Suppose X = [Xl ..... Xm] E ] R ~ x n is such that XI AX = diag(J1, ... , J
m
), where
each J
i
= diag(JiI,"" Jik,) and each Jik is a Jordan block corresponding to Ai E A(A).
We could also use other block diagonal decompositions (e.g., via SVD), but we restrict our
attention here to only the Jordan block case. Note that AXi = Xi J
i
, so by (9.19) the columns
of Xi (i.e., the eigenvectors and principal vectors associated with Ai) span an Ainvariant
subspace of]Rn.
Finally, we return to the problem of developing a formula for e
l
A in the case that A
is not necessarily diagonalizable. Let Yi E <e
nxn
, be a Jordan basis for N (AT  A;lt.
Equivalently, partition
9.5. The Matrix Sign Function 91
compatibly. Then
In a similar fashion we can compute
which is a useful formula when used in conjunction with the result
for a k x k Jordan block 7, associated with an eigenvalue A. = A.,.
9.5 The Matrix Sign Function
In this section we give a very brief introduction to an interesting and useful matrix function
called the matrix sign function. It is a generalization of the sign (or signum) of a scalar. A
survey of the matrix sign function and some of its applications can be found in [15].
Definition 9.40. Let z E C with Re(z) ^ 0. Then the sign of z is defined by
Definition 9.41. Suppose A e C"
x
" has no eigenvalues on the imaginary axis, and let
be a Jordan canonical form for A, with N containing all Jordan blocks corresponding to the
eigenvalues of A in the left halfplane and P containing all Jordan blocks corresponding to
eigenvalues in the right halfplane. Then the sign of A, denoted sgn(A), is given by
9.S. The Matrix Sign Function
compatibly. Then
A = XJX
I
= XJy
H
= [XI, ... , Xm] diag(JI, ... , J
m
) [Y
I
, ••• , Ym]H
m
= LX;JiYi
H
.
i=1
In a similar fashion we can compute
m
etA = LXietJ;YiH,
i=1
which is a useful formula when used in conjunction with the result
A 0 0
eAt teAt
.lt
2
e
At
2!
0 A
0
eAt teAt
exp t
A 0
0 0
eAt
1
0 0 A
0 0
for a k x k Jordan block J
i
associated with an eigenvalue A = Ai.
9.5 The Matrix Sign Function
91
In this section we give a very brief introduction to an interesting and useful matrix function
called the matrix sign function. It is a generalization of the sign (or signum) of a scalar. A
survey of the matrix sign function and some of its applications can be found in [15].
Definition 9.40. Let z E C with Re(z) f= O. Then the sign of z is defined by
Re(z) {+1
sgn(z) = IRe(z) I = 1
ifRe(z) > 0,
ifRe(z) < O.
Definition 9.41. Suppose A E cnxn has no eigenvalues on the imaginary axis, and let
be a Jordan canonical form for A, with N containing all Jordan blocks corresponding to the
eigenvalues of A in the left halfplane and P containing all Jordan blocks corresponding to
eigenvalues in the right halfplane. Then the sign of A, denoted sgn(A), is given by
[
/ 0] I
sgn(A) = X 0 / X ,
92 Chapter 9. Eigenvalues and Eigenvectors
where the negative and positive identity matrices are of the same dimensions as N and P,
respectively.
There are other equivalent definitions of the matrix sign function, but the one given
here is especially useful in deriving many of its key properties. The JCF definition of the
matrix sign function does not generally lend itself to reliable computation on a finiteword
length digital computer. In fact, its reliable numerical calculation is an interesting topic in
its own right.
We state some of the more useful properties of the matrix sign function as theorems.
Their straightforward proofs are left to the exercises.
Theorem 9.42. Suppose A e C"
x
" has no eigenvalues on the imaginary axis, and let
S = sgn(A). Then the following hold:
1. S is diagonalizable with eigenvalues equal to del.
2. S
2
= I.
3. AS = SA.
4. sgn(A") = (sgn(A))".
5. sgn(T
l
AT) = T
l
sgn(A)TforallnonsingularT e C"
x
".
6. sgn(cA) = sgn(c) sgn(A)/or all nonzero real scalars c.
Theorem 9.43. Suppose A e C"
x
" has no eigenvalues on the imaginary axis, and let
S — sgn(A). Then the following hold:
1. 7l(S — /) is an Ainvariant subspace corresponding to the left halfplane eigenvalues
of A (the negative invariant subspace).
2. R(S+/) is an Ainvariant subspace corresponding to the right halfplane eigenvalues
of A (the positive invariant subspace).
3. negA = (/ — S)/2 is a projection onto the negative invariant subspace of A.
4. posA = (/ + S)/2 is a projection onto the positive invariant subspace of A.
EXERCISES
1. Let A e C
nxn
have distinct eigenvalues AI, ..., X
n
with corresponding right eigen
vectors Xi, ... ,x
n
and left eigenvectors y\, ..., y
n
, respectively. Let v e C" be an
arbitrary vector. Show that v can be expressed (uniquely) as a linear combination
of the right eigenvectors. Find the appropriate expression for v as a linear combination
of the left eigenvectors as well.
92 Chapter 9. Eigenvalues and Eigenvectors
where the negative and positive identity matrices are of the same dimensions as N and p,
respectively.
There are other equivalent definitions of the matrix sign function, but the one given
here is especially useful in deriving many of its key properties. The JCF definition of the
matrix sign function does not generally lend itself to reliable computation on a finiteword
length digital computer. In fact, its reliable numerical calculation is an interesting topic in
its own right.
We state some of the more useful properties of the matrix sign function as theorems.
Their straightforward proofs are left to the exercises.
Theorem 9.42. Suppose A E e
nxn
has no eigenvalues on the imaginary axis, and let
S = sgn(A). Then the following hold:
1. S is diagonalizable with eigenvalues equal to ± 1.
2. S2 = I.
3. AS = SA.
4. sgn(AH) = (sgn(A»H.
5. sgn(T1AT) = T1sgn(A)T foralinonsingularT E e
nxn
.
6. sgn(cA) = sgn(c) sgn(A) for all nonzero real scalars c.
Theorem 9.43. Suppose A E e
nxn
has no eigenvalues on the imaginary axis, and let
S = sgn(A). Then the following hold:
I. R(S l) is an Ainvariant subspace corresponding to the left halfplane eigenvalues
of A (the negative invariant subspace).
2. R(S + l) is an A invariant subspace corresponding to the right halfplane eigenvalues
of A (the positive invariant subspace).
3. negA == (l  S) /2 is a projection onto the negative invariant subspace of A.
4. posA == (l + S)/2 is a projection onto the positive invariant subspace of A.
EXERCISES
1. Let A E e
nxn
have distinct eigenvalues ),.1> ••• , ),.n with corresponding right eigen
vectors Xl, ... , Xn and left eigenvectors Yl, ••. , Yn, respectively. Let v E en be an
arbitrary vector. Show that v can be expressed (uniquely) as a linear combination
of the right eigenvectors. Find the appropriate expression for v as a linear combination
of the left eigenvectors as well.
Exercises 93
2. Suppose A € C"
x
" is skewHermitian, i.e., A
H
= —A. Prove that all eigenvalues of
a skewHermitian matrix must be pure imaginary.
3. Suppose A e C"
x
" is Hermitian. Let A be an eigenvalue of A with corresponding
right eigenvector x. Show that x is also a left eigenvector for A. Prove the same result
if A is skewHermitian.
5. Determine the eigenvalues, right eigenvectors and right principal vectors if necessary,
and (real) JCFs of the following matrices:
6. Determine the JCFs of the following matrices:
Find a nonsingular matrix X such that X
1
AX = J, where J is the JCF
Hint: Use[ — 1 1 — l]
r
as an eigenvector. The vectors [0 1 — l]
r
and[ l 0 0]
r
are both eigenvectors, but then the equation (A — /)jc
(2)
= x
(1)
can't be solved.
8. Show that all right eigenvectors of the Jordan block matrix in Theorem 9.30 must be
multiples of e\ e R*. Characterize all left eigenvectors.
9. Let A e R"
x
" be of the form A = xy
T
, where x, y e R" are nonzero vectors with
x
T
y = 0. Determine the JCF of A.
10. Let A e R"
xn
be of the form A = / + xy
T
, where x, y e R" are nonzero vectors
with x
T
y = 0. Determine the JCF of A.
11. Suppose a matrix A e R
16x 16
has 16 eigenvalues at 0 and its JCF consists of a single
Jordan block of the form specified in Theorem 9.22. Suppose the small number 10~
16
is added to the (16,1) element of J. What are the eigenvalues of this slightly perturbed
matrix?
4. Suppose a matrix A € R
5x5
has eigenvalues {2, 2, 2, 2, 3}. Determine all possible
JCFs for A.
7. Let
Exercises 93
2. Suppose A E rc
nxn
is skewHermitian, i.e., AH = A. Prove that all eigenvalues of
a skewHermitian matrix must be pure imaginary.
3. Suppose A E rc
nxn
is Hermitian. Let A be an eigenvalue of A with corresponding
right eigenvector x. Show that x is also a left eigenvector for A. Prove the same result
if A is skewHermitian.
4. Suppose a matrix A E lR.
5x5
has eigenvalues {2, 2, 2, 2, 3}. Determine all possible
JCFs for A.
5. Determine the eigenvalues, right eigenvectors and right principal vectors if necessary,
and (real) JCFs of the following matrices:
[
2 1 ]
(a) 1 0 '
6. Determine the JCFs of the following matrices:
<a) U j n
2
1
2
=n
7. Let
A = [H 1]·
2 2"
Find a nonsingular matrix X such that XI AX = J, where J is the JCF
J = [ ~ ~ ~ ] .
001
Hint: Use[1 1  I]T as an eigenvector. The vectors [0 If and[1 0 of
are both eigenvectors, but then the equation (A  I)x(2) = x(1) can't be solved.
8. Show that all right eigenvectors of the Jordan block matrix in Theorem 9.30 must be
multiples of el E lR.
k
. Characterize all left eigenvectors.
9. Let A E lR.
nxn
be of the form A = xyT, where x, y E lR.
n
are nonzero vectors with
x
T
y = O. Determine the JCF of A.
10. Let A E lR.
nxn
be of the form A = 1+ xyT, where x, y E lR.
n
are nonzero vectors
with x
T
y = O. Determine the JCF of A.
11. Suppose a matrix A E lR.
16x
16 has 16 eigenvalues at 0 and its JCF consists of a single
Jordan block of the form specified in Theorem 9.22. Suppose the small number 10
16
is added to the (16,1) element of J. What are the eigenvalues of this slightly perturbed
matrix?
94 Chapter 9. Eigenvalues and Eigenvectors
12. Show that every matrix A e R"
x
" can be factored in the form A = Si$2, where Si
and £2 are real symmetric matrices and one of them, say Si, is nonsingular.
Hint: Suppose A = X J X ~
l
is a reduction of A to JCF and suppose we can construct
the "symmetric factorization" of J . Then A = ( X S i X
T
) ( X ~
T
S
2
X ~
l
) would be the
required symmetric factorization of A. Thus, it suffices to prove the result for the
JCF. The transformation P in (9.18) is useful.
13. Prove that every matrix A e W
x
" is similar to its transpose and determine a similarity
transformation explicitly.
Hint: Use the factorization in the previous exercise.
14. Consider the block upper triangular matrix
where A e M"
xn
and A
n
e R
kxk
with 1 < k < n. Suppose A
u
^ 0 and that we
want to block diagonalize A via the similarity transformation
where X e R*
x
<«  *), i.e.,
Find a matrix equation that X must satisfy for this to be possible. If n = 2 and k = 1,
what can you say further, in terms of AU and A 22, about when the equation for X is
solvable?
15. Prove Theorem 9.42.
16. Prove Theorem 9.43.
17. Suppose A e C"
xn
has all its eigenvalues in the left half plane. Prove that
sgn(A) =  /.
94 Chapter 9. Eigenvalues and Eigenvectors
12. Show that every matrix A E jRnxn can be factored in the form A = SIS2, where SI
and S2 are real symmetric matrices and one of them, say S1, is nonsingular.
Hint: Suppose A = Xl XI is a reduction of A to JCF and suppose we can construct
the "symmetric factorization" of 1. Then A = (X SIXT)(X
T
S2XI) would be the
required symmetric factorization of A. Thus, it suffices to prove the result for the
JCF. The transformation P in (9.18) is useful.
13. Prove that every matrix A E jRn xn is similar to its transpose and determine a similarity
transformation explicitly.
Hint: Use the factorization in the previous exercise.
14. Consider the block upper triangular matrix
A _ [ All
 0
Al2 ]
A22 '
where A E jRnxn and All E jRkxk with 1 ::s: k ::s: n. Suppose Al2 =1= 0 and that we
want to block diagonalize A via the similarity transformation
where X E IRkx(nk), i.e.,
TIAT = [A 011 0 ]
A22 .
Find a matrix equation that X must satisfy for this to be possible. If n = 2 and k = 1,
what can you say further, in terms of All and A22, about when the equation for X is
solvable?
15. Prove Theorem 9.42.
16. Prove Theorem 9.43.
17. Suppose A E en xn has all its eigenvalues in the left halfplane. Prove that
sgn(A) = 1.
Chapter 10
Canonical Forms
10.1 Some Basic Canonical Forms
Problem: Let V and W be vector spaces and suppose A : V — > • W is a linear transformation.
Find bases in V and W with respect to which Mat A has a "simple form" or "canonical
form." In matrix terms, if A e R
mxn
, find P e R™
xm
and Q e R
n
n
xn
such that PAQ has a
"canonical form." The transformation A M» PAQ is called an equivalence; it is called an
orthogonal equivalence if P and Q are orthogonal matrices.
Remark 10.1. We can also consider the case A e C
m xn
and unitary equivalence if P and
< 2 are unitary.
Two special cases are of interest:
1. If W = V and < 2 = P"
1
, the transformation A H> PAP"
1
is called a similarity.
2 . If W = V and if Q = P
T
is orthogonal, the transformation A i» PAP
T
is called
an orthogonal similarity (or unitary similarity in the complex case).
The following results are typical of what can be achieved under a unitary similarity. If
A = A
H
6 C"
x
" has eigenvalues AI, . . . , A
n
, then there exists a unitary matrix £7 such that
U
H
AU — D, where D = di ag( A. j , . . . , A.
n
). This is proved in Theorem 10.2 . What other
matrices are "diagonalizable" under unitary similarity? The answer is given in Theorem
10.9, where it is proved that a general matrix A e C"
x
" is unitarily similar to a diagonal
matrix if and only if it is normal (i.e., AA
H
= A
H
A). Normal matrices include Hermitian,
skewHermitian, and unitary matrices (and their "real" counterparts: symmetric, skew
symmetric, and orthogonal, respectively), as well as other matrices that merely satisfy the
definition, such as A = [ _
a
b
^1 for real scalars a and b. If a matrix A is not normal, the
most "diagonal" we can get is the JCF described in Chapter 9.
Theorem 10.2. Let A = A
H
e C"
x
" have (real) eigenvalues A. I, . . . , X
n
. Then there
exists a unitary matrix X such that X
H
AX = D = diag(A.j , . . . , X
n
) (the columns ofX are
orthonormal eigenvectors for A).
95
Chapter 10
Canonical Forms
10.1 Some Basic Canonical Forms
Problem: Let V and W be vector spaces and suppose A : V + W is a linear transformation.
Find bases in V and W with respect to which Mat A has a "simple form" or "canonical
form." In matrix terms, if A E IR
mxn
, find P E lR;;:xm and Q E l R ~ x n such that P AQ has a
"canonical form." The transformation A f+ P AQ is called an equivalence; it is called an
orthogonal equivalence if P and Q are orthogonal matrices.
Remark 10.1. We can also consider the case A E e
mxn
and unitary equivalence if P and
Q are unitary.
Two special cases are of interest:
1. If W = V and Q = p
1
, the transformation A f+ PAPI is called a similarity.
2. If W = V and if Q = pT is orthogonal, the transformation A f+ P ApT is called
an orthogonal similarity (or unitary similarity in the complex case).
The following results are typical of what can be achieved under a unitary similarity. If
A = A H E en xn has eigenvalues AI, ... , An, then there exists a unitary matrix U such that
U
H
AU = D, where D = diag(AJ, ... , An). This is proved in Theorem 10.2. What other
matrices are "diagonalizable" under unitary similarity? The answer is given in Theorem
10.9, where it is proved that a general matrix A E e
nxn
is unitarily similar to a diagonal
matrix if and only if it is normal (i.e., AA H = AHA). Normal matrices include Hermitian,
skewHermitian, and unitary matrices (and their "real" counterparts: symmetric, skew
symmetric, and orthogonal, respectively), as well as other matrices that merely satisfy the
definition, such as A = [ _ ~ !] for real scalars a and h. If a matrix A is not normal, the
most "diagonal" we can get is the JCF described in Chapter 9.
Theorem 10.2. Let A = A H E en xn have (real) eigenvalues AI, ... ,An. Then there
exists a unitary matrix X such that X
H
AX = D = diag(Al, ... , An) (the columns of X are
orthonormal eigenvectors for A).
95
96 Chapter 10. Canonical Forms
Proof: Let x\ be a right eigenvector corresponding to X\, and normalize it such that xf*x\ =
1. Then there exist n — 1 additional vectors x
2
, ..., x
n
such that X = [x\,..., x
n
] =
[x\ X
2
] is unitary. Now
Then x^U
2
= 0 (/ € k) means that x
f
is orthogonal to each of the n — k columns of U
2
.
But the latter are orthonormal since they are the last n — k rows of the unitary matrix U.
Thus, [Xi f/2] is unitary. D
The construction called for in Theorem 10.2 is then a special case of Theorem 10.3
for k = 1. We illustrate the construction of the necessary Householder matrix for k — 1.
For simplicity, we consider the real case. Let the unit vector x\ be denoted by [£i, ..., %
n
]
T
.
In (10.1) we have used the fact that Ax\ = k\x\. When combined with the fact that
x"xi = 1, we get Ai remaining in the (l,l)block. We also get 0 in the (2,l)block by
noting that x\ is orthogonal to all vectors in X
2
. In (10.2), we get 0 in the (l,2)block by
noting that X
H
AX is Hermitian. The proof is completed easily by induction upon noting
that the (2,2)block must have eigenvalues X
2
,..., A.
n
. D
Given a unit vector x\ e E", the construction of X
2
e ]R"
X
("
1
) such that X —
[x\ X
2
] is orthogonal is frequently required. The construction can actually be performed
quite easily by means of Householder (or Givens) transformations as in the proof of the
following general result.
Theorem 10.3. Let X\ e C
nxk
have orthonormal columns and suppose U is a unitary
matrix such that UX\ = \
0
1, where R € C
kxk
is upper triangular. Write U
H
= [U\ U
2
]
with Ui € C
nxk
. Then [Xi U
2
] is unitary.
Proof: Let X\ = [x\,..., Xk]. Construct a sequence of Householder matrices (also known
as elementary reflectors) H\,..., H
k
in the usual way (see below) such that
where R is upper triangular (and nonsingular since x\, ..., Xk are orthonormal). Let U =
H
k
...H
v
. Then U
H
= / / , • • H
k
and
96 Chapter 10. Canonical Forms
Proof' Let XI be a right eigenvector corresponding to AI, and normalize it such that XI =
1. Then there exist n  1 additional vectors X2, ... , Xn such that X = (XI, ... , xn] =
[XI X
2
] is unitary. Now
XHAX = [
xH
] A [XI X2] = [
]
I
XH
XfAxl XfAX
2
2
=[
Al
]
(10.1)
0 XfAX
2
=[
Al 0
l
(10.2)
0
XfAX
z
In (l0.1) we have used the fact that AXI = AIXI. When combined with the fact that
XI = 1, we get Al remaining in the (l,I)block. We also get 0 in the (2, I)block by
noting that XI is orthogonal to all vectors in Xz. In (10.2), we get 0 in the (l,2)block by
noting that XH AX is Hermitian. The proof is completed easily by induction upon noting
that the (2,2)block must have eigenvalues A2, ... , An. 0
Given a unit vector XI E JRn, the construction of X
z
E JRnx(nl) such that X =
[XI X
2
] is orthogonal is frequently required. The construction can actually be performed
quite easily by means of Householder (or Givens) transformations as in the proof of the
following general result.
Theorem 10.3. Let XI E C
nxk
have orthonormal columns and suppose V is a unitary
matrix such that V X I = [ where R E C
kxk
is upper triangular. Write V H = [VI Vz]
with VI E C
nxk
. Then [XI V
2
] is unitary.
Proof: Let X I = [XI, ... ,xd. Construct a sequence of Householder matrices (also known
as elementary reflectors) HI, ... , Hk in the usual way (see below) such that
Hk ... HdxI, ... , xd = [ l
where R is upper triangular (and nonsingular since XI, ... , Xk are orthonormal). Let V =
Hk'" HI. Then VH = HI'" Hk and
Then X
i
H
U2 = 0 (i E means that Xi is orthogonal to each of the n  k columns of V2.
But the latter are orthonormal since they are the last n  k rows of the unitary matrix U.
Thus. [XI U2] is unitary. 0
The construction called for in Theorem 10.2 is then a special case of Theorem 10.3
for k = 1. We illustrate the construction of the necessary Householder matrix for k = 1.
For simplicity, we consider the real case. Let the unit vector XI be denoted by .. , ,
10.1. Some Basic Canonical Forms 97
Then the necessary Householder matrix needed for the construction of X^ is given by
U = I — 2uu
+
= I — ^UU
T
, where u = [t\ ± 1, £2, • • •» £«]
r
 It can easily be checked
that U is symmetric and U
T
U = U
2
= I, so U is orthogonal. To see that U effects the
necessary compression of j ci, it is easily verified that U
T
U = 2 ± 2£i and U
T
X\ = 1 ± £1.
Thus,
Further details on Householder matrices, including the choice of sign and the complex case,
can be consulted in standard numerical linear algebra texts such as [7], [11], [23], [25].
The real version of Theorem 10.2 is worth stating separately since it is applied fre
quently in applications.
Theorem 10.4. Let A = A
T
e E
nxn
have eigenvalues k\, ... ,X
n
. Then there exists an
orthogonal matrix X e W
lxn
(whose columns are orthonormal eigenvectors of A) such that
X
T
AX = D = diag(Xi, . . . , X
n
).
Note that Theorem 10.4 implies that a symmetric matrix A (with the obvious analogue
from Theorem 10.2 for Hermitian matrices) can be written
which is often called the spectral representation of A. In fact, A in (10.3) is actually a
weighted sum of orthogonal projections P, (onto the onedimensional eigenspaces corre
sponding to the A., 's), i.e.,
where P, = PUM —
x
i
x
f =
x
i
x
j since xj xi — 1.
The following pair of theorems form the theoretical foundation of the doubleFrancis
QR algorithm used to compute matrix eigenvalues in a numerically stable and reliable way.
10.1. Some Basic Canonical Forms 97
Then the necessary Householder matrix needed for the construction of X
2
is given by
U = I  2uu+ = I  +uu
T
, where u = [';1 ± 1, ';2, ... , ';nf. It can easily be checked
u u
that U is symmetric and U
T
U = U
2
= I, so U is orthogonal. To see that U effects the
necessary compression of Xl, it is easily verified that u
T
u = 2 ± 2';1 and u
T
Xl = 1 ± ';1.
Thus,
Further details on Householder matrices, including the choice of sign and the complex case,
can be consulted in standard numerical linear algebra texts such as [7], [11], [23], [25].
The real version of Theorem 10.2 is worth stating separately since it is applied fre
quently in applications.
Theorem 10.4. Let A = AT E jRnxn have eigenvalues AI, ... , An. Then there exists an
orthogonal matrix X E jRn xn (whose columns are orthonormal eigenvectors of A) such that
XT AX = D = diag(Al, ... , An).
Note that Theorem 10.4 implies that a symmetric matrix A (with the obvious analogue
from Theorem 10.2 for Hermitian matrices) can be written
n
A = XDX
T
= LAiXiXT,
(10.3)
i=1
which is often called the spectral representation of A. In fact, A in (10.3) is actually a
weighted sum of orthogonal projections Pi (onto the onedimensional eigenspaces corre
sponding to the Ai'S), i.e.,
n
A = LAiPi,
i=l
where Pi = PR(x;) = xiXt = xixT since xT Xi = 1.
The following pair of theorems form the theoretical foundation of the doubleFrancis
QR algorithm used to compute matrix eigenvalues in a numerically stable and reliable way.
98 Chapter 10. Canonical Forms
Theorem 10.5 (Schur). Let A e C"
x
". Then there exists a unitary matrix U such that
U
H
AU = T, where T is upper triangular.
Proof: The proof of this theorem is essentially the same as that of Theorem 10.2 except that
in this case (using the notation U rather than X) the (l,2)block wf AU2 is not 0. D
In the case of A e R"
x
", it is thus unitarily similar to an upper triangular matrix, but
if A has a complex conjugate pair of eigenvalues, then complex arithmetic is clearly needed
to place such eigenvalues on the diagonal of T. However, the next theorem shows that every
A e W
xn
is also orthogonally similar (i.e., real arithmetic) to a quasiuppertriangular
matrix. A quasiuppertriangular matrix is block upper triangular with 1 x 1 diagonal
blocks corresponding to its real eigenvalues and 2x2 diagonal blocks corresponding to its
complex conjugate pairs of eigenvalues.
Theorem 10.6 (MurnaghanWintner). Let A e R"
x
". Then there exists an orthogonal
matrix U such that U
T
AU = S, where S is quasiuppertriangular.
Definition 10.7. The triangular matrix T in Theorem 10.5 is called a Schur canonical
form or Schur form. The quasiuppertriangular matrix S in Theorem 10.6 is called a real
Schur canonical form or real Schur form (RSF). The columns of a unitary [orthogonal]
matrix U that reduces a matrix to [real] Schur form are called Schur vectors.
Example 10.8. The matrix
is in RSF. Its real JCF is
Note that only the first Schur vector (and then only if the corresponding first eigenvalue
is real if U is orthogonal) is an eigenvector. However, what is true, and sufficient for virtually
all applications (see, for example, [17]), is that the first k Schur vectors span the same A
invariant subspace as the eigenvectors corresponding to the first k eigenvalues along the
diagonal of T (or S).
While every matrix can be reduced to Schur form (or RSF), it is of interest to know
when we can go further and reduce a matrix via unitary similarity to diagonal form. The
following theorem answers this question.
Theorem 10.9. A matrix A e C"
x
" is unitarily similar to a diagonal matrix if and only if
A is normal (i.e., A
H
A = AA
H
).
Proof: Suppose U is a unitary matrix such that U
H
AU = D, where D is diagonal. Then
so A is normal.
98 Chapter 10. Canonical Forms
Theorem 10.5 (Schur). Let A E c
nxn
. Then there exists a unitary matrix U such that
U
H
AU = T, where T is upper triangular.
Proof: The proof of this theorem is essentially the same as that of Theorem lO.2 except that
in this case (using the notation U rather than X) the (l,2)block ur AU
2
is not O. 0
In the case of A E IR
n
xn , it is thus unitarily similar to an upper triangular matrix, but
if A has a complex conjugate pair of eigenvalues, then complex arithmetic is clearly needed
to place such eigenValues on the diagonal of T. However, the next theorem shows that every
A E IR
nxn
is also orthogonally similar (i.e., real arithmetic) to a quasiuppertriangular
matrix. A quasiuppertriangular matrix is block upper triangular with 1 x 1 diagonal
blocks corresponding to its real eigenvalues and 2 x 2 diagonal blocks corresponding to its
complex conjugate pairs of eigenvalues.
Theorem 10.6 (MurnaghanWintner). Let A E IR
n
xn. Then there exists an orthogonal
matrix U such that U
T
AU = S, where S is quasiuppertriangular.
Definition 10.7. The triangular matrix T in Theorem 10.5 is called a Schur canonical
form or Schur fonn. The quasiuppertriangular matrix S in Theorem 10.6 is called a real
Schur canonical form or real Schur fonn (RSF). The columns of a unitary [orthogonal}
matrix U that reduces a matrix to [real} Schur fonn are called Schur vectors.
Example 10.8. The matrix
s ~ [
2 5
n
2 4
0 0
is in RSF. Its real JCF is
h[
1
n
1 1
0 0
Note that only the first Schur vector (and then only if the corresponding first eigenvalue
is real if U is orthogonal) is an eigenvector. However, what is true, and sufficient for virtually
all applications (see, for example, [17]), is that the first k Schur vectors span the same A
invariant subspace as the eigenvectors corresponding to the first k eigenvalues along the
diagonal of T (or S).
While every matrix can be reduced to Schur form (or RSF), it is of interest to know
when we can go further and reduce a matrix via unitary similarity to diagonal form. The
following theorem answers this question.
Theorem 10.9. A matrix A E c
nxn
is unitarily similar to a diagonal matrix if and only if
A is normal (i.e., AH A = AA
H
).
Proof: Suppose U is a unitary matrix such that U
H
AU = D, where D is diagonal. Then
AAH = U VUHU VHU
H
= U DDHU
H
== U DH DU
H
== AH A
so A is normal.
10.2. Definite Matrices 99
Conversely, suppose A is normal and let U be a unitary matrix such that U
H
AU = T,
where T is an upper triangular matrix (Theorem 10.5). Then
It is then a routine exercise to show that T must, in fact, be diagonal. D
10.2 Definite Matrices
Definition 10.10. A symmetric matrix A e W
xn
is
1. positive definite if and only ifx
T
Ax > Qfor all nonzero x G W
1
. We write A > 0.
2. nonnegative definite (or positive semidefinite) if and only if X
T
Ax > 0 for all
nonzero x e W. We write A > 0.
3. negative definite if—A is positive definite. We write A < 0.
4. nonpositive definite (or negative semidefinite) if— A is nonnegative definite. We
write A < 0.
Also, if A and B are symmetric matrices, we write A > B if and only if A — B > 0 or
B — A < 0. Similarly, we write A > B if and only ifA — B>QorB — A < 0.
Remark 10.11. If A e C"
x
" is Hermitian, all the above definitions hold except that
superscript //s replace Ts. Indeed, this is generally true for all results in the remainder of
this section that may be stated in the real case for simplicity.
Remark 10.12. If a matrix is neither definite nor semidefinite, it is said to be indefinite.
Theorem 10.13. Let A = A
H
e C
nxn
with eigenvalues X
{
> A
2
> • • • > A
n
. Then for all
x eC",
Proof: Let U be a unitary matrix that diagonalizes A as in Theorem 10.2. Furthermore,
let v = U
H
x, where x is an arbitrary vector in C
M
, and denote the components of y by
j]i, i € n. Then
But clearly
10.2. Definite Matrices 99
Conversely, suppose A is normal and let U be a unitary matrix such that U H A U = T,
where T is an upper triangular matrix (Theorem 10.5). Then
It is then a routine exercise to show that T must, in fact, be diagonal. 0
10.2 Definite Matrices
Definition 10.10. A symmetric matrix A E lR.
nxn
is
1. positive definite if and only if x T Ax > 0 for all nonzero x E lR.
n
. We write A > O.
2. nonnegative definite (or positive semidefinite) if and only if x
T
Ax :::: 0 for all
nonzero x E lR.
n
• We write A :::: O.
3. negative definite if  A is positive definite. We write A < O.
4. nonpositive definite (or negative semidefinite) if A is nonnegative definite. We
write A ~ O.
Also, if A and B are symmetric matrices, we write A > B if and only if A  B > 0 or
B  A < O. Similarly, we write A :::: B if and only if A  B :::: 0 or B  A ~ O.
Remark 10.11. If A E e
nxn
is Hermitian, all the above definitions hold except that
superscript H s replace T s. Indeed, this is generally true for all results in the remainder of
this section that may be stated in the real case for simplicity.
Remark 10.12. If a matrix is neither definite nor semidefinite, it is said to be indefinite.
Theorem 10.13. Let A = AH E e
nxn
with eigenvalues AI :::: A2 :::: ... :::: An. Thenfor all
x E en,
Proof: Let U be a unitary matrix that diagonalizes A as in Theorem 10.2. Furthermore,
let y = U H x, where x is an arbitrary vector in en, and denote the components of y by
11;, i En. Then
But clearly
n
x
H
Ax = (U
H
X)H U
H
AU(U
H
x) = yH Dy = LA; 111;12.
n
LA; 11'/;12 ~ AlyH Y = AIX
H
X
;=1
;=1
100 Chapter 10. Canonical Forms
and
from which the theorem follows. D
Remark 10.14. The ratio ^^ for A = A
H
< = C
nxn
and nonzero jc e C" is called the
Rayleigh quotient of jc. Theorem 10.13 provides upper (AO and lower (A.
w
) bounds for
the Rayleigh quotient. If A = A
H
e C"
x
" is positive definite, X
H
Ax > 0 for all nonzero
x E C", soO < X
n
< • • • < A. I.
Corollary 10.15. Let A e C"
x
". Then \\A\\
2
= ^
m
(A
H
A}.
Proof: For all x € C" we have
Let jc be an eigenvector corresponding to X
max
(A
H
A). Then ^pjp
2
= ^^(A" A) , whence
Definition 10.16. A principal submatrix of an nxn matrix A is the (n — k)x(n — k) matrix
that remains by deleting k rows and the corresponding k columns. A leading principal
submatrix of order n — k is obtained by deleting the last k rows and columns.
Theorem 10.17. A symmetric matrix A e E"
x
" is positive definite if and only if any of the
following three equivalent conditions hold:
1. The determinants of all leading principal submatrices of A are positive.
2. All eigenvalues of A are positive.
3. A can be written in the formM
T
M, where M e R"
x
" is nonsingular.
Theorem 10.18. A symmetric matrix A € R"
x
" is nonnegative definite if and only if any
of the following three equivalent conditions hold:
1. The determinants of all principal submatrices of A are nonnegative.
2. All eigenvalues of A are nonnegative.
3. A can be written in the formM
T
M, where M 6 R
ix
" and k > rank(A) — rank(M) .
Remark 10.19. Note that the determinants of all principal eubmatrioes muet bQ nonnogativo
in Theorem 10.18.1, not just those of the leading principal submatrices. For example,
consider the matrix A — [
0
_
l
1. The determinant of the 1x1 leading submatrix is 0 and
the determinant of the 2x2 leading submatrix is also 0 (cf. Theorem 10.17). However, the
100 Chapter 10. Canonical Forms
and
n
LAillJilZ::: AnyHy = An
xHx
,
i=l
from which the theorem follows. 0
Remark 10.14. The ratio XHHAx for A = AH E e
nxn
and nonzero x E en is called the
x x
Rayleigh quotient of x. Theorem 1O.l3 provides upper (A 1) and lower (An) bounds for
the Rayleigh quotient. If A = AH E e
nxn
is positive definite, x
H
Ax > 0 for all nonzero
x E en, so 0 < An ::::: ... ::::: A I.
I
Corollary 10.15. Let A E e
nxn
. Then IIAII2 = Ar1ax(AH A).
Proof: For all x E en we have
I
Let x be an eigenvector corresponding to Amax (A H A). Then = Ar1ax (A H A), whence
IIAxll2 ! H
IIAliz = max = Amax{A A). 0
xfO IIxll2
Definition 10.16. A principal submatrixofan n x n matrix A is the (n k) x (n k) matrix
that remains by deleting k rows and the corresponding k columns. A leading principal
submatrix of order n  k is obtained by deleting the last k rows and columns.
Theorem 10.17. A symmetric matrix A E is positive definite if and only if any of the
following three equivalent conditions hold:
1. The determinants of all leading principal submatrices of A are positive.
2. All eigenvalues of A are positive.
3. A can be written in the form MT M, where M E xn is nonsingular.
Theorem 10.18. A symmetric matrix A E xn is nonnegative definite if and only if any
of the following three equivalent conditions hold:
1. The determinants of all principal submatrices of A are nonnegative.
2. All eigenvalues of A are nonnegaTive.
3. A can be wrirren in [he/orm MT M, where M E IRb<n and k ranlc(A) "" ranlc(M).
R.@mllrk 10.19. Not@th!ltthl!dl!termin!lntl:ofnllprincip!ll "ubm!ltriC[!!l mu"t bB nonnBgmivB
in Theorem 10.18.1, not just those of the leading principal submatrices. For example,
consider the matrix A = _ The determinant of the I x 1 leading submatrix is 0 and
the determinant of the 2 x 2 leading submatrix is also 0 (cf. Theorem 10.17). However, the
10.2. Definite Matrices 101
principal submatrix consisting of the (2,2) element is, in fact, negative and A is nonpositive
definite.
Remark 10.20. The factor M in Theorem 10.18.3 is not unique. For example, if
Recall that A > B if the matrix A — B is nonnegative definite. The following
theorem is useful in "comparing" symmetric matrices. Its proof is straightforward from
basic definitions.
Theorem 10.21. Let A, B e R
nxn
be symmetric.
1. If A >BandMe R
nxm
, then M
T
AM > M
T
BM.
2. If A >B and M e R
nxm
, then M
T
AM > M.
T
BM.
j m
The following standard theorem is stated without proof (see, for example, [16, p.
181]). It concerns the notion of the "square root" of a matrix. That is, if A € E"
xn
, we say
that S e R
nx
" is a square root of A if S
2
— A. In general, matrices (both symmetric and
nonsymmetric) have infinitely many square roots. For example, if A = /2, any matrix S of
the form [
c
s
°*
e
e
_
c
s
™
9
e
] is a square root.
Theorem 10.22. Let A e R"
x
" be nonnegative definite. Then A has a unique nonnegative
definite square root S. Moreover, SA = AS and rankS = rank A (and hence S is positive
definite if A is positive definite).
A stronger form of the third characterization in Theorem 10.17 is available and is
known as the Cholesky factorization. It is stated and proved below for the more general
Hermitian case.
Theorem 10.23. Let A e <C
nxn
be Hermitian and positive definite. Then there exists a
unique nonsingular lower triangular matrix L with positive diagonal elements such that
A = LL
H
.
Proof: The proof is by induction. The case n = 1 is trivially true. Write the matrix A in
the form
By our induction hypothesis, assume the result is true for matrices of order n — 1 so that B
may be written as B = L\L^, where L\ e C
1
""
1
^""^ is nonsingular and lower triangular
then M can be
10.2. Definite Matrices 101
principal submatrix consisting of the (2,2) element is, in fact, negative and A is nonpositive
definite.
Remark 10.20. The factor M in Theorem 10.18.3 is not unique. For example, if
then M can be
[1 0], [ fz
ti
o [ ~ 0]
o l ~ 0 , ...
v'3 0
Recall that A :::: B if the matrix A  B is nonnegative definite. The following
theorem is useful in "comparing" symmetric matrices. Its proof is straightforward from
basic definitions.
Theorem 10.21. Let A, B E jRnxn be symmetric.
1. 1f A :::: Band M E jRnxm, then MT AM :::: MT BM.
2. If A> Band M E j R ~ x m , then MT AM> MT BM.
The following standard theorem is stated without proof (see, for example, [16, p.
181]). It concerns the notion of the "square root" of a matrix. That is, if A E lR.
nxn
, we say
that S E jRn xn is a square root of A if S2 = A. In general, matrices (both symmetric and
nonsymmetric) have infinitely many square roots. For example, if A = lz, any matrix S of
h
" [COSO Sino] .
t e 10rm sinO _ cosO IS a square root.
Theorem 10.22. Let A E lR.
nxn
be nonnegative definite. Then A has a unique nonnegative
definite square root S. Moreover, SA = AS and rankS = rankA (and hence S is positive
definite if A is positive definite).
A stronger form of the third characterization in Theorem 10.17 is available and is
known as the Cholesky factorization. It is stated and proved below for the more general
Hermitian case.
Theorem 10.23. Let A E c
nxn
be Hermitian and positive definite. Then there exists a
unique nonsingular lower triangular matrix L with positive diagonal elements such that
A = LLH.
Proof: The proof is by induction. The case n = 1 is trivially true. Write the matrix A in
the form
By our induction hypothesis, assume the result is true for matrices of order n  1 so that B
may be written as B = L1Lf, where Ll E c(nl)x(nl) is nonsingular and lower triangular
102 Chapt er 10. Ca n o n i c a l Forms
with positive diagonal elements. It remains to prove that we can write the n x n matrix A
in the form
10.3 Equivalence Transformations and Congruence
Theorem 10.24. Let A € C™*
7 1
. Then there exist matrices P e C ™
xm
and Q e C"
n
x
" such
that
Proof: A classical proof can be consulted in, for example, [21, p. 131]. Alternatively,
suppose A has an SVD of the form (5.2) in its complex version. Then
Note that the greater freedom afforded by the equivalence transformation of Theorem
10.24, as opposed to the more restrictive situation of a similarity transformation, yields a
far "simpler" canonical form (10.4). However, numerical procedures for computing such
an equivalence directly via, say, Gaussian or elementary row and column operations, are
generally unreliable. The numerically preferred equivalence is, of course, the unitary equiv
alence known as the SVD. However, the SVD is relatively expensive to compute and other
canonical forms exist that are intermediate between (10.4) and the SVD; see, for example
[7, Ch. 5], [4, Ch. 2]. Two such forms are stated here. They are more stably computable
than (10.4) and more efficiently computable than a ful l SVD. Many similar results are also
available.
where a is positive. Performing the indicated matrix multiplication and equating the cor
responding submatrices, we see that we must have L\c = b and a
nn
= C
H
C + a
2
. Clearly
c is given simply by c = L^b. Substituting in the expression involving a, we find
a
2
= a
nn
— b
H
L\
H
L\
l
b = a
nn
— b
H
B~
l
b (= the Schur complement of B in A). But we
know that
Since det (fi ) > 0, we must have a
nn
—b
H
B
l
b > 0. Choosing a to be the positive square
root of «„„ — b
H
B~
l
b completes the proof. D
102 Chapter 10. Canonical Forms
with positive diagonal elements. It remains to prove that we can write the n x n matrix A
in the form
b ] = [L
J
0 ] [Lf c J,
ann c a 0 a
where a is positive. Performing the indicated matrix multiplication and equating the cor
responding submatrices, we see that we must have L I C = b and ann = c
H
c + a
2
• Clearly
c is given simply by c = C,lb. Substituting in the expression involving a, we find
a
2
= ann  b
H
LIH L11b = ann  b
H
B1b (= the Schur complement of B in A). But we
know that
o < det(A) = det [
b ] = det(B) det(a
nn
_ b
H
B1b).
ann
Since det(B) > 0, we must have ann  b
H
B1b > O. Choosing a to be the positive square
root of ann  b
H
B1b completes the proof. 0
10.3 Equivalence Transformations and Congruence
Theorem 10.24. Let A E c;,xn. Then there exist matrices P E C:
xm
and Q E such
that
(l0.4)
Proof: A classical proof can be consulted in, for example, [21, p. 131]. Alternatively,
suppose A has an SVD of the form (5.2) in its complex version. Then
[
Sl 0 ] [ U
H
] [I 0 ]
o I Uf AV = 0 0 .
Take P = [ 'f [I ] and Q = V to complete the proof. 0
Note that the greater freedom afforded by the equivalence transformation of Theorem
10.24, as opposed to the more restrictive situation of a similarity transformation, yields a
far "simpler" canonical form (10.4). However, numerical procedures for computing such
an equivalence directly via, say, Gaussian or elementary row and column operations, are
generally unreliable. The numerically preferred equivalence is, of course, the unitary equiv
alence known as the SVD. However, the SVD is relatively expensive to compute and other
canonical forms exist that are intermediate between (l0.4) and the SVD; see, for example
[7, Ch. 5], [4, Ch. 2]. Two such forms are stated here. They are more stably computable
than (lOA) and more efficiently computable than a full SVD. Many similar results are also
available.
10.3. Equivalence Transformations and Congruence 103
Theorem 10.25 (Complete Orthogonal Decomposition). Let A e C™
x
". Then there exist
unitary matrices U e C
mxm
and V e C
nxn
such that
where R e €,
r
r
xr
is upper (or lower) triangular with positive diagonal elements.
Proof: For the proof, see [4]. D
Theorem 10.26. Let A e C™
x
". Then there exists a unitary matrix Q e C
mxm
and a
permutation matrix Fl e C"
x
" such that
where R E C
r
r
xr
is upper triangular and S e C
r x(
"
r)
is arbitrary but in general nonzero.
Proof: For the proof, see [4]. D
Remark 10.27. When A has full column rank but is "near" a rank deficient matrix,
various rank revealing QR decompositions are available that can sometimes detect such
phenomena at a cost considerably less than a full SVD. Again, see [4] for details.
Definition 10.28. Let A e C
nxn
and X e C
n
n
xn
. The transformation A i> X
H
AX is called
a congruence. Note that a congruence is a similarity if and only ifX is unitary.
Note that congruence preserves the property of being Hermitian; i.e., if A is Hermitian,
then X
H
AX is also Hermitian. It is of interest to ask what other properties of a matrix are
preserved under congruence. It turns out that the principal property so preserved is the sign
of each eigenvalue.
Definition 10.29. Let A = A
H
e C"
x
" and let 7t, v, and £ denote the numbers of positive,
negative, and zero eigenvalues, respectively, of A. Then the inertia of A is the triple of
numbers In(A) = (n, v, £). The signature of A is given by sig(A) = n — v.
Example 10.30.
2. If A = A" eC
n x
" , t h enA > 0 if and only if In (A) = (n, 0, 0).
3. If In(A) = (TT, v, £), then rank(A) = n + v.
Theorem 10.31 (Sylvester's Law of Inertia). Let A = A
H
e C
nxn
and X e C
n
n
xn
. Then
In(A) = ln(X
H
AX).
Proof: For the proof, see, for example, [21, p. 134]. D
Theorem 10.31 guarantees that rank and signature of a matrix are preserved under
congruence. We then have the following.
10.3. Equivalence Transformations and Congruence 103
Theorem 10.25 (Complete Orthogonal Decomposition). Let A E e ~ x n . Then there exist
unitary matrices U E e
mxm
and V E e
nxn
such that
where R E e;xr is upper (or lower) triangular with positive diagonal elements.
Proof: For the proof, see [4]. 0
(10.5)
Theorem 10.26. Let A E e ~ x n . Then there exists a unitary matrix Q E e
mxm
and a
permutation matrix IT E en xn such that
QAIT = [ ~ ~ l
(10.6)
where R E e;xr is upper triangular and S E erx(nr) is arbitrary but in general nonzero.
Proof: For the proof, see [4]. 0
Remark 10.27. When A has full column rank but is "near" a rank deficient matrix,
various rank revealing QR decompositions are available that can sometimes detect such
phenomena at a cost considerably less than a full SVD. Again, see [4] for details.
Definition 10.28. Let A E e
nxn
and X E e ~ x n . The transformation A H XH AX is called
a congruence. Note that a congruence is a similarity if and only if X is unitary.
Note that congruence preserves the property of being Hermitian; i.e., if A is Hermitian,
then X H AX is also Hermitian. It is of interest to ask what other properties of a matrix are
preserved under congruence. It turns out that the principal property so preserved is the sign
of each eigenvalue.
Definition 10.29. Let A = AH E e
nxn
and let rr, v, and ~ denote the numbers of positive,
negative, and zero eigenvalues, respectively, of A. Then the inertia of A is the triple of
numbers In(A) = (rr, v, n The signature of A is given by sig(A) = rr  v.
Example 10.30.
l.In[! 1
o 0
00] o 0
10 =(2,1,1).
o 0
2. If A = AH E e
nxn
, then A> 0 if and only if In(A) = (n, 0, 0).
3. If In(A) = (rr, v, n, then rank(A) = rr + v.
Theorem 10.31 (Sylvester's Law of Inertia). Let A = A HE en xn and X E e ~ xn. Then
In(A) = In(X
H
AX).
Proof: For the proof, see, for example, [21, p. 134]. D
Theorem 10.31 guarantees that rank and signature of a matrix are preserved under
congruence. We then have the following.
104 Chapter 10. Canonical Forms
Theorem 10.32. Let A = A
H
e C"
xn
with In(A) = (jt, v, £). Then there exists a matrix
X e C"
n
xn
such that X
H
AX = diag(l, . . . , 1, 1,..., 1, 0, . . . , 0), where the number of
1 's is 7i, the number of — l's is v, and the number 0/0 's is (,.
Proof: Let AI , . . . , X
w
denote the eigenvalues of A and order them such that the first TT are
positive, the next v are negative, and the final £ are 0. By Theorem 10.2 there exists a unitary
matrix U such that U
H
AU = diag(Ai, . . . , A
w
). Define the n x n matrix
Then it is easy to check that X = U W yields the desired result. D
10.3.1 Block matrices and definiteness
Theorem 10.33. Suppose A = A
T
and D = D
T
. Then
if and only if either A > 0 and D  B
T
A~
l
B > 0, or D > 0 and A  BD^B
T
> 0.
Proof: The proof follows by considering, for example, the congruence
The details are straightforward and are left to the reader. D
Remark 10.34. Note the symmetric Schur complements of A (or D) in the theorem.
Theorem 10.35. Suppose A = A
T
and D = D
T
. Then
if and only ifA>0, AA
+
B = B, and D  B
T
A
+
B > 0.
Proof: Consider the congruence with
and proceed as in the proof of Theorem 10.33. D
10.4 Rational Canonical Form
One final canonical form to be mentioned is the rational canonical form.
104 Chapter 10. Canonical Forms
Theorem 10.32. Let A = AH E c
nxn
with In(A) = (Jr, v, O. Then there exists a matrix
X E such that XH AX = diag(1, ... , I, I, ... , 1,0, ... ,0), where the number of
1 's is Jr, the number of I 's is v, and the numberofO's
Proof: Let A I, ... , An denote the eigenvalues of A and order them such that the first Jr are
positive, the next v are negative, and the final are O. By Theorem 10.2 there exists a unitary
matrix V such that VH AV = diag(AI, ... , An). Define the n x n matrix
vv = ... , 1/.fArr+I' ... , I/.fArr+v, I, ... ,1).
Then it is easy to check that X = V VV yields the desired result. 0
10.3.1 Block matrices and definiteness
Theorem 10.33. Suppose A = AT and D = DT. Then
ifand only ifeither A> ° and D  BT AI B > 0, or D > 0 and A  BD
I
BT > O.
Proof: The proof follows by considering, for example, the congruence
B ] [I _AI B JT [ A
D 0 I BT
] [
The details are straightforward and are left to the reader. 0
Remark 10.34. Note the symmetric Schur complements of A (or D) in the theorem.
Theorem 10.35. Suppose A = AT and D = DT. Then
B ] > °
D 
if and only if A:::: 0, AA+B = B. and D  BT A+B:::: o.
Proof: Consider the congruence with
and proceed as in the proof of Theorem 10.33. 0
10.4 Rational Canonical Form
One final canonical form to be mentioned is the rational canonical form.
10.4. Rational Canonical Form 105
Definition 10.36. A matrix A e M"
x
" is said to be nonderogatory if its minimal polynomial
and characteristic polynomial are the same or, equivalently, if its Jordan canonical f orm
has only one block associated with each distinct eigenvalue.
Suppose A E W
xn
is a nonderogatory matrix and suppose its characteristic polyno
mial is 7 r( A ) = A " — ( a
0
+ «A +
is similar to a matrix of the form
+ a
n
_ i A
n
~ ' )  Then it can be shown (see [12]) that A
Definition 10.37. A matrix A e E
nx
" of the f orm (10.7) is called a companion matrix or
is said to be in companion form.
Companion matrices also appear in the literature in several equivalent forms. To
illustrate, consider the companion matrix
This matrix is a special case of a matrix in lower Hessenberg form. Using the reverseorder
identity similarity P given by (9.18), A is easily seen to be similar to the following matrix
in upper Hessenberg form:
Moreover, since a matrix is similar to its transpose (see exercise 13 in Chapter 9), the
following are also companion matrices similar to the above:
Notice that in all cases a companion matrix is nonsingular if and only if aO /= 0.
In fact, the inverse of a nonsingular companion matrix is again in companion form. For
£*Yamr\1j=»
10.4. Rational Canonical Form 105
Definition 10.36. A matrix A E lR
n
Xn is said to be nonderogatory ifits minimal polynomial
and characteristic polynomial are the same or; equivalently, if its Jordan canonical form
has only one block associated with each distinct eigenvalue.
Suppose A E lR
nxn
is a nonderogatory matrix and suppose its characteristic polyno
mial is n(A) = An  (ao + alA + ... + an_IAnI). Then it can be shown (see [12]) that A
is similar to a matrix of the form
o o o
o 0
o
(10.7)
o o
Definition 10.37. A matrix A E lR
nxn
of the form (10.7) is called a cornpanion rnatrix or
is said to be in cornpanion forrn.
Companion matrices also appear in the literature in several equivalent forms. To
illustrate, consider the companion matrix
(l0.8)
This matrix is a special case of a matrix in lower Hessenberg form. Using the reverseorder
identity similarity P given by (9.18), A is easily seen to be similar to the following matrix
in upper Hessenberg form:
a2 al
o 0
1 0
o 1
6]
o .
o
(10.9)
Moreover, since a matrix is similar to its transpose (see exercise 13 in Chapter 9), the
following are also companion matrices similar to the above:
l
:: ~ ! ~ 0 1 ] .
ao 0 0
(10.10)
Notice that in all cases a companion matrix is nonsingular if and only if ao i= O.
In fact, the inverse of a nonsingular companion matrix is again in companion form. For
example,
o
1
o
 ~
ao
1
o
o
 ~
ao
o
o
_!!l
o
o
(10.11)
106 Chapter 10. Canonical Forms
with a similar result for companion matrices of the form (10.10).
If a companion matrix of the form (10.7) is singular, i.e., if ao = 0, then its pseudo
inverse can still be computed. Let a e M""
1
denote the vector \a\, 02,..., a
n
i] and let
c =
l+
l
a
r
a
. Then it is easily verified that
Note that / — caa
T
= (I + aa
T
) , and hence the pseudoinverse of a singular companion
matrix is not a companion matrix unless a = 0.
Companion matrices have many other interesting properties, among which, and per
haps surprisingly, is the fact that their singular values can be found in closed form; see
[14].
Theorem 10.38. Let a\ > GI > • • • > a
n
be the singular values of the companion matrix
A in (10.7). Let a = a\ + a\ + • • • +a%_
{
and y = 1 + «.Q + a. Then
Remark 10.39. Explicit formulas for all the associated right and left singular vectors can
also be derived easily.
If A € R
nx
" is derogatory, i.e., has more than one Jordan block associated with
at least one eigenvalue, then it is not similar to a companion matrix of the form (10.7).
However, it can be shown that a derogatory matrix is similar to a block diagonal matrix,
each of whose diagonal blocks is a companion matrix. Such matrices are said to be in
rational canonical form (or Frobenius canonical form). For details, see, for example, [12].
Companion matrices appear frequently in the control and signal processing literature
but unfortunately they are often very difficult to work with numerically. Algorithms to reduce
an arbitrary matrix to companion form are numerically unstable. Moreover, companion
matrices are known to possess many undesirable numerical properties. For example, in
general and especially as n increases, their eigenstructure is extremely ill conditioned,
nonsingular ones are nearly singular, stable ones are nearly unstable, and so forth [14].
Ifao ^ 0, the largest and smallest singular values can also be written in the equivalent form
106 Chapter 10. Canonical Forms
with a similar result for companion matrices of the form (10.10).
If a companion matrix of the form (10.7) is singular, i.e., if ao = 0, then its pseudo
inverse can still be computed. Let a E JRn1 denote the vector [ai, a2, ... , anIf and let
c = I + ~ T a' Then it is easily verified that
o
o
o
o
o o
o
o
+
o
1 caa
T
o J.
ca
Note that I  caa T = (I + aa T) I , and hence the pseudoinverse of a singular companion
matrix is not a companion matrix unless a = O.
Companion matrices have many other interesting properties, among which, and per
haps surprisingly, is the fact that their singular values can be found in closed form; see
[14].
Theorem 10.38. Let al ~ a2 ~ ... ~ an be the singular values of the companion matrix
A in (10.7). Leta = ar + ai + ... + a;_1 and y = 1 + aJ + a. Then
2 _ 1 ( J 2 2)
a
l
 2 y + y  4a
o
'
a? = 1 for i = 2, 3, ... , n  1,
a; = ~ (y  J y2  4a
J
) .
If ao =1= 0, the largest and smallest singular values can also be written in the equivalent form
Remark 10.39. Explicit formulas for all the associated right and left singular vectors can
also be derived easily.
If A E JRnxn is derogatory, i.e., has more than one Jordan block associated with
at least one eigenvalue, then it is not similar to a companion matrix of the form (10.7).
However, it can be shown that a derogatory matrix is similar to a block diagonal matrix,
each of whose diagonal blocks is a companion matrix. Such matrices are said to be in
rational canonical form (or Frobenius canonical form). For details, see, for example, [12].
Companion matrices appear frequently in the control and signal processing literature
but unfortunately they are often very difficult to work with numerically. Algorithms to reduce
an arbitrary matrix to companion form are numerically unstable. Moreover, companion
matrices are known to possess many undesirable numerical properties. For example, in
general and especially as n increases, their eigenstructure is extremely ill conditioned,
nonsingular ones are nearly singular, stable ones are nearly unstable, and so forth [14].
Exercises 107
Companion matrices and rational canonical forms are generally to be avoided in floating
point computation.
Remark 10.40. Theorem 10.38 yields some understanding of why difficult numerical
behavior might be expected for companion matrices. For example, when solving linear
systems of equations of the form (6.2), one measure of numerical sensitivity is K
P
(A) =
I I ^ I I
p
I I A~
l
I I
p
>
m
e socalled condition number of A with respect to inversion and with respect
to the matrix Pnorm. I f this number is large, say 0(10*), one may lose up to k digits of
precision. I n the 2norm, this condition number is the ratio of largest to smallest singular
values which, by the theorem, can be determined explicitly as
I t is easy to show that y/2/ao < k2(A) < £,, and when GO is small or y is large (or both),
then K2(A) ^ T~I. I t is not unusual for y to be large for large n. Note that explicit formulas
for K\ (A) and K oo(A) can also be determined easily by using (10.11).
EXERCISES
1. Show that if a triangular matrix is normal, then it must be diagonal.
2. Prove that if A e M"
x
" is normal, then Af(A) = A/"(A
r
).
3. Let A G C
nx
" and define p(A) = maxx
€
A(A) I ' M Then p(A) is called the spectral
radius of A. Show that if A is normal, then p(A) = A
2
. Show that the converse
is true if n = 2.
4. Let A € C
nxn
be normal with eigenvalues y1 , ..., y
n
and singular values a\ > a
2
>
• • • > o
n
> 0. Show that a, (A) = A.,(A) for i e n.
5. Use the reverseorder identity matrix P introduced in (9.18) and the matrix U in
Theorem 10.5 to find a unitary matrix Q that reduces A e C"
x
" to lower triangular
form.
6. Let A = I J MeC
2x2
. Find a unitary matrix U such that
7. I f A e W
xn
is positive definite, show that A
[
must also be positive definite.
3. Suppose A e E"
x
" is positive definite. I s [ ^ /i 1 > 0?
}. Let R, S 6 E
nxn
be symmetric. Show that [ * J 1 > 0 if and only if S > 0 and
R> S
Exercises 107
Companion matrices and rational canonical forms are generally to be avoided in fioating
point computation.
Remark 10.40. Theorem 10.38 yields some understanding of why difficult numerical
behavior might be expected for companion matrices. For example, when solving linear
systems of equations of the form (6.2), one measure of numerical sensitivity is Kp(A) =
II A II p II A ] II p' the socalled condition number of A with respect to inversion and with respect
to the matrix pnorm. If this number is large, say O(lO
k
), one may lose up to k digits of
precision. In the 2norm, this condition number is the ratio of largest to smallest singular
values which, by the theorem, can be determined explicitly as
y+J
y
2 4
a5
21
a
ol
It is easy to show that :::: K2(A) :::: 1:01' and when ao is small or y is large (or both),
then K2(A) It is not unusualfor y to be large forlarge n. Note that explicit formulas
for K] (A) and Koo(A) can also be determined easily by using (l0.11).
EXERCISES
1. Show that if a triangular matrix is normal, then it must be diagonal.
2. Prove that if A E jRnxn is normal, then N(A) = N(A
T
).
3. Let A E cc
nxn
and define peA) = max)..EA(A) IAI. Then peA) is called the spectral
radius of A. Show that if A is normal, then peA) = IIAII2' Show that the converse
is true if n = 2.
4. Let A E en xn be normal with eigenvalues A], ... , An and singular values 0'1 0'2
... an O. Show that a; (A) = IA;(A)I for i E!l.
5. Use the reverseorder identity matrix P introduced in (9.18) and the matrix U in
Theorem 10.5 to find a unitary matrix Q that reduces A E cc
nxn
to lower triangular
form.
6. Let A = :] E CC
2x2
. Find a unitary matrix U such that
7. If A E jRn xn is positive definite, show that A I must also be positive definite.
8. Suppose A E jRnxn is positive definite. Is [1 O?
9. Let R, S E jRnxn be symmetric. Show that > 0 if and only if S > 0 and
R > SI.
108 Chapter 10. Canonical Forms
10. Find the inertia of the following matrices:
108
10. Find the inertia of the following matrices:
(a) [ ~ ~ l (b) [
(d) [1 1 + j ]
1  j 1 .
Chapter 10. Canonical Forms
2 1 + j ]
1  j 2 '
Chapter 11
Linear Differential and
Difference Equations
11.1 Differential Equations
In this section we study solutions of the linear homogeneous system of differential equations
for t > IQ. This is known as an initialvalue problem. We restrict our attention in this
chapter only to the socalled timeinvariant case, where the matrix A e R
nxn
is constant
and does not depend on t. The solution of (11.1) is then known always to exist and be
unique. It can be described conveniently in terms of the matrix exponential.
Definition 11.1. For all A e R
nxn
, the matrix exponential e
A
e R
nxn
is defined by the
power series
The series (11.2) can be shown to converge for all A (has radius of convergence equal
to +00). The solution of (11.1) involves the matrix
which thus also converges for all A and uniformly in t.
11.1.1 Properties of the matrix exponential
1. e° = I.
Proof: This follows immediately from Definition 11.1 by setting A = 0.
2. For all A G R"
XM
, (e
A
f  e^.
Proof: This follows immediately from Definition 11.1 and linearity of the transpose.
109
Chapter 11
Linear Differential and
Difference Equations
11.1 Differential Equations
In this section we study solutions of the linear homogeneous system of differential equations
x(t) = Ax(t); x(to) = Xo E JR.n (11.1)
for t 2: to. This is known as an initialvalue problem. We restrict our attention in this
chapter only to the socalled timeinvariant case, where the matrix A E JR.nxn is constant
and does not depend on t. The solution of (11.1) is then known always to exist and be
unique. It can be described conveniently in terms of the matrix exponential.
Definition 11.1. For all A E JR.nxn, the matrix exponential e
A
E JR.nxn is defined by the
power series
+00 1
e
A
= L ,Ak.
k=O k.
(11.2)
The series (11.2) can be shown to converge for all A (has radius of convergence equal
to +(0). The solution of (11.1) involves the matrix
(11.3)
which thus also converges for all A and uniformly in t.
11.1.1 Properties of the matrix exponential
1. eO = I.
Proof This follows immediately from Definition 11.1 by setting A = O.
T T
2. For all A E JR.nxn, (e
A
) = e
A
•
Proof This follows immediately from Definition 11.1 and linearity of the transpose.
109
110 Chapter 11. Linear Differential and Difference Equations
3. For all A e R"
x
" and for all t, r e R, e
(t
+
T)A
= e'
A
e
rA
= e
lA
e'
A
.
Proof: Note that
Compare like powers of A in the above two equations and use the binomial theorem
on (t + T)*.
4. For all A, B e R"
xn
and for all t e R, e
t(A+B)
=^e'
A
e'
B
= e'
B
e'
A
if and only if A
and B commute, i.e., AB = B A.
Proof: Note that
and
and
while
Compare like powers of t in the first equation and the second or third and use the
binomial theorem on (A + B)
k
and the commutativity of A and B.
5. For all A e R"
x
" and for all t e R, (e'
A
)~
l
= e~'
A
.
Proof: Simply take T = — t in property 3.
6. Let £ denote the Laplace transform and £~
!
the inverse Laplace transform. Then for
all A € R"
x
" and for all t € R,
(a) C{e
tA
} = (sIAr
l
.
(b) £
1
{(j/A)
1
} = «
M
.
Proof: We prove only (a). Part (b) follows similarly.
110 Chapter 11. Linear Differential and Difference Equations
3. For all A E JRnxn and for all t, T E JR, e(t+r)A = etA erA = erAe
tA
.
Proof" Note that
(t + T)2 2
e(t+r)A = I + (t + T)A + A + ...
2!
and
tA rA t 2 T 2
(
2 )( 2 )
e e = I + t A + 2! A +... I + T A + 2! A +... .
Compare like powers of A in the above two equations and use the binomial theorem
on(t+T)k.
4. For all A, B E JRnxn and for all t E JR, et(A+B) =etAe
tB
= etBe
tA
if and only if A
and B commute, i.e., AB = BA.
Proof' Note that
and
while
t
2
et(A+B) = I + teA + B) + (A + B)2 + ...
2!
tB tA t 2 t 2
(
2 )( 2 )
e e = 1+ tB + 2iB +... 1+ tA + 2!A +... .
Compare like powers of t in the first equation and the second or third and use the
binomial theorem on (A + B/ and the commutativity of A and B.
5. ForaH A E JRnxn and for all t E JR, (etA)1 = e
tA
.
Proof" Simply take T = t in property 3.
6. Let £ denote the Laplace transform and £1 the inverse Laplace transform. Then for
all A E JRnxn and for all t E lR,
(a) .l{e
tA
} = (sI  A)I.
(b) .lI{(sl A)I} = erA.
Proof" We prove only (a). Part (b) follows similarly.
{+oo
= io et(sl)e
tA
dt
(+oo
= io ef(Asl) dt since A and (sf) commute
11.1. Differential Equations 111
= (sl A)
1
.
The matrix (s I — A) ~' is called the resolvent of A and is defined for all s not in A (A).
Notice in the proof that we have assumed, for convenience, that A is diagonalizable.
If this is not the case, the scalar dyadic decomposition can be replaced by
using the JCF. All succeeding steps in the proof then follow in a straightforward way.
7. For all A e R"
x
" and for all t e R, £(e'
A
) = Ae
tA
= e'
A
A.
Proof: Since the series (11.3) is uniformly convergent, it can be differentiated termby
term from which the result follows immediately. Alternatively, the formal definition
can be employed as follows. For any consistent matrix norm,
11.1. Differential Equations 111
= {+oo t e(AiS)t x;y;H dt assuming A is diagonalizable
10 ;=1
= e(AiS)t dt]x;y;H
n 1
= '"' Xi y;H assuming Re s > Re Ai for i E !!
L..... s  A"
i=1 I
= (sI  A)I.
The matrix (s I  A) I is called the resolvent of A and is defined for all s not in A (A).
Notice in the proof that we have assumed, for convenience, that A is diagonalizable.
If this is not the case, the scalar dyadic decomposition can be replaced by
m
et(Asl) = L Xiet(Jisl)y;H
;=1
using the JCF. All succeeding steps in the proof then follow in a straightforward way.
7. For all A E JRnxn and for all t E JR, 1h(e
tA
) = Ae
tA
= etA A.
Proof: Since the series (11.3) is uniformly convergent, it can be differentiated termby
term from which the result follows immediately. Alternatively, the formal definition
d e(t+M)A _ etA
_(/A) = lim
d t L'lt
can be employed as follows. For any consistent matrix norm,
II
etA II III II
u.  Ae
tA
= L'lt  /A)  Ae
tA
= II  etA)  Ae
tA
II
= II  l)e
tA
 Ae
tA
II
II
I ( (M)2 2 ) tA tAil
= L'lt M A + A +... e  Ae
= II ( Ae
tA
+ A
2
e
tA
+ ... )  Ae
tA
II
= II ( A2 + A
3
+ .. , ) etA II
< MIIA21111e
tA
II _ + IIAII + IIAI12 + ...
(
1 L'lt (L'lt)2 )
 2! 3! 4!
< L'lt1lA21111e
tA
Il (1 + L'ltiIAIl + IIAII2 + ... )
= L'lt IIA 21111e
tA
112 Chapter 11. Linear Differential and Difference Equations
For fixed t, the righthand side above clearly goes to 0 as At goes to 0. Thus, the
limit exists and equals Ae'
A
. A similar proof yields the limit e'
A
A, or one can use the
fact that A commutes with any polynomial of A of finite degree and hence with e'
A
.
11.1.2 Homogeneous linear differential equations
Theorem 11.2. Let A e R
nxn
. The solution of the linear homogeneous initialvalue problem
Proof: Differentiate (11.5) and use property 7 of the matrix exponential to get x( t ) =
Ae
( t
~
to) A
xo = Ax( t) . Also, x( t
0
) — e
( fo
~
t
° '
) A
X Q — X Q so, by the fundamental existence and
uniqueness theorem for ordinary differential equations, (11.5) is the solution of (11.4). D
11.1.3 Inhomogeneous linear differential equations
Theorem 11.3. Let A e R
nxn
, B e W
xm
and let the vectorvalued function u be given
and, say, continuous. Then the solution of the linear inhomogeneous initialvalue problem
for t > IQ is given by the variation of parameters formula
Proof: Differentiate (11.7) and again use property 7 of the matrix exponential. The general
formula
is used to get x( t ) = Ae
{
'
to) A
x
0
+ f'
o
Ae
(
'
s) A
Bu( s) ds + Bu( t) = Ax( t) + Bu( t) . Also,
*('o)
=
< ?
(f
° ~
fo)/ 1
.¥ o + 0 = X Q so, by the fundamental existence and uniqueness theorem for
ordinary differential equations, (11.7) is the solution of (11.6). D
Remark 11.4. The proof above simply verifies the variation of parameters formula by
direct differentiation. The formula can be derived by means of an integrating factor "trick"
as follows. Premultiply the equation x — Ax = Bu by e~
tA
to get
112 Chapter 11. Linear Differential and Difference Equations
For fixed t, the righthand side above clearly goes to 0 as t:.t goes to O. Thus, the
limit exists and equals Ae
t
A • A similar proof yields the limit e
t
A A, or one can use the
fact that A commutes with any polynomial of A of finite degree and hence with etA.
11.1.2 Homogeneous linear differential equations
Theorem 11.2. Let A E IR
n
xn. The solution of the linear homogeneous initialvalue problem
x(t) = Ax(l); x(to) = Xo E IR
n
(11.4)
for t ::: to is given by
(11.5)
Proof: Differentiate (11.5) and use property 7 of the matrix exponential to get x (t) =
Ae(tto)A
xo
= Ax(t). Also, x(to) = e(toto)A Xo = Xo so, by the fundamental existence and
uniqueness theorem for ordinary differential equations, (11.5) is the solution of (11.4). 0
11.1.3 Inhomogeneous linear differential equations
Theorem 11.3. Let A E IR
nxn
, B E IR
nxm
and let the vectorvalued function u be given
and, say, continuous. Then the solution of the linear inhomogeneous initialvalue problem
x(t) = Ax(t) + Bu(t); x(to) = Xo E IR
n
for t ::: to is given by the variation of parameters formula
x(t) = e(tto)A
xo
+ t e(ts)A Bu(s) ds.
l t o
(11.6)
(11.7)
Proof: Differentiate (11.7) and again use property 7 of the matrix exponential. The general
formula
d l
q
(t) l
q
(t) af(x t) dq(t) dp(t)
 f(x, t) dx = ' dx + f(q(t), t)  f(p(t), t)
dt pet) pet) at dt dt
is used to get x(t) = Ae(tto)A Xo + Ir: Ae(ts)A Bu(s) ds + Bu(t) = Ax(t) + Bu(t). Also,
x(t
o
} = e(totolA Xo + 0 = Xo so, by the fundilm()ntill nnd uniqu()Oc:s:s theorem for
ordinary differential equations, (11.7) is the solution of (1l.6). 0
Remark 11.4. The proof above simply verifies the variation of parameters formula by
direct differentiation. The formula can be derived by means of an integrating factor "trick"
as follows. Premultiply the equation x  Ax = Bu by e
tA
to get
(11.8)
11.1. Differential Equations 113
Now integrate (11.8) over the interval [to, t]:
Thus,
and hence
11.1.4 Linear matrix differential equations
Matrixvalued initialvalue problems also occur frequently. The first is an obvious general
ization of Theorem 11.2, and the proof is essentially the same.
Theorem 11.5. Let A e W
lxn
. The solution of the matrix linear homogeneous initialvalue
nrohlcm
for t > to is given by
In the matrix case, we can have coefficient matrices on both the right and left. For
convenience, the following theorem is stated with initial time to = 0.
Theorem 11.6. Let A e Rn
xn
, B e R
mxm
, and C e Rn
xm
. Then the matrix initialvalue
problem
—
a
tA
ra
tB
has the solutionX ( t ) = e Ce
Proof: Differentiate e
tA
Ce
tB
with respect to t and use property 7 of the matrix exponential.
The fact that X ( t ) satisfies the initial condition is trivial. D
Corollary 11.7. Let A, C e IR"
X
". Then the matrix initialvalue problem
has the solution X(t} = e
tA
Ce
tAT
.
When C is symmetric in (11.12), X ( t ) is symmetric and (11.12) is known as a Lya
punov differential equation. The initialvalue problem (11.11) is known as a Sylvester
differential equation.
11.1. Differential Equations
Now integrate (11.8) over the interval [to, t]:
Thus,
and hence
esAx(s) ds = eSABu(s) ds.
1
t d 1t
to ds to
etAx(t)  etoAx(to) = t e
sA
Bu(s) ds
lto
x(t) = e(tt
olA
xo
+ t e(ts)A Bu(s) ds.
lto
11.1.4 Linear matrix differential equations
113
Matrixvalued initialvalue problems also occur frequently. The first is an obvious general
ization of Theorem 11.2, and the proof is essentially the same.
Theorem 11.5. Let A E jRnxn. The solution of the matrix linear homogeneous initialvalue
problem
X(t) = AX(t); X(to) = C E jRnxn (11.9)
for t ::: to is given by
X(t) = e(tto)Ac.
(11.10)
In the matrix case, we can have coefficient matrices on both the right and left. For
convenience, the following theorem is stated with initial time to = O.
Theorem 11.6. Let A E jRnxn, B E jRmxm, and C E ]R.nxm. Then the matrix initialvalue
problem
X(t) = AX(t) + X(t)B; X(O) = C (11.11)
has the solution X(t) = etACe
tB
.
Proof: Differentiate etACe
tB
with respect to t and use property 7 of the matrix exponential.
The fact that X (t) satisfies the initial condition is trivial. 0
Corollary 11.7. Let A, C E ]R.nxn. Then the matrix initialvalue problem
X(t) = AX(t) + X(t)AT; X(O) = C (11.12)
has the solution X(t) = etACetAT.
When C is symmetric in (11.12), X (t) is symmetric and (11.12) is known as a Lya
punov differential equation. The initialvalue problem (11.11) is known as a Sylvester
differential equation.
114 Chapter 11. Linear Differential and Difference Equations
11.1.5 Modal decompositions
Let A E W
xn
and suppose, for convenience, that it is diagonalizable (if A is not diagonaliz
able, the rest of this subsection is easily generalized by using the JCF and the decomposition
A — ^ X f Ji Y
t
H
as discussed in Chapter 9). Then the solution x(t) of (11.4) can be written
The ki s are called the modal velocities and the right eigenvectors *, are called the modal
directions. The decomposition above expresses the solution x(t) as a weighted sum of its
modal velocities and directions.
This modal decomposition can be expressed in a different looking but identical form
if we write the initial condition X Q as a weighted sum of the right eigenvectors
Then
In the last equality we have used the fact that y f * X j = S f j .
Similarly, in the inhomogeneous case we can write
11.1.6 Computation of the matrix exponential
JCF method
Let A e R"
x
" and suppose X e Rn
xn
is such that X"
1
AX = J, where J is a JCF for A.
Then
114 Chapter 11. Linear Differential and Difference Equations
11.1 .5 Modal decompositions
Let A E jRnxn and suppose, for convenience, that it is diagonalizable (if A is not diagonaliz
able, the rest of this subsection is easily generalized by using the JCF and the decomposition
A = L X;li y
i
H
as discussed in Chapter 9). Then the solution x(t) of (11.4) can be written
x(t) = e(tto)A Xo
= (ti.iUtO)Xiyr) Xo
1=1
n
= L(Yi
H
xoeAi(ttO»Xi.
i=1
The Ai s are called the modal velocities and the right eigenvectors Xi are called the modal
directions. The decomposition above expresses the solution x (t) as a weighted sum of its
modal velocities and directions.
This modal decomposition can be expressed in a different looking but identical form
n
if we write the initial condition Xo as a weighted sum of the right eigenvectors Xo = L ai Xi.
Then
n
= L(aieAiUtO»Xi.
i=1
In the last equality we have used the fact that Yi
H
X j = flij.
Similarly, in the inhomogeneous case we can write
i
t e(ts)A Bu(s) ds = t (it eAiUS)YiH Bu(s) dS) Xi.
~ i=1 ~
11.1.6 Computation of the matrix exponential
JCF method
i=1
Let A E jRnxn and suppose X E j R ~ x n is such that XI AX = J, where J is a JCF for A.
Then
etA = etXJX1
= XetJX
1
I
n
Le
A
•
,
X'Yi
H
if A is diagonalizable
1=1
~ t,x;e'J,y;H in geneml.
11.1. Differential Equations 115
If A is diagonalizable, it is then easy to compute e
tA
via the formula e
tA
= Xe
tJ
X '
since e
tj
is simply a diagonal matrix.
In the more general case, the problem clearly reduces simply to the computation of
the exponential of a Jordan block. To be specific, let .7, e <C
kxk
be a Jordan block of the form
Clearly A/ and N commute. Thus, e
tJi
= e'
u
e
tN
by property 4 of the matrix exponential.
The diagonal part is easy: e
tu
= diag(e
x
',..., e
xt
}. But e
tN
is almost as easy since N is
nilpotent of degree k.
Definition 11.8. A matrix M e M
nx
"
M
p
= 0, while M
p
~
l
^ 0.
is nilpotent of degree (or index, or grade) p if
For the matrix N defined above, it is easy to check that while N has 1's along only
its first superdiagonal (and O's elsewhere), N
2
has 1's along only its second superdiagonal,
and so forth. Finally, N
k
~
l
has a 1 in its (1, k) element and has O's everywhere else, and
N
k
= 0. Thus, the series expansion of e'
N
is finite, i.e.,
Thus,
In the case when A. is complex, a real version of the above can be worked out.
11.1. Differential Equations 115
If A is diagonalizable, it is then easy to compute etA via the formula etA = Xe
tl
XI
since e
t
I is simply a diagonal matrix.
In the more general case, the problem clearly reduces simply to the computation of
the exponential of a Jordan block. To be specific, let J
i
E C
kxk
be a Jordan block of the form
J
i
=
A 1
o A
o
o o
o =U+N.
o A
Clearly AI and N commute. Thus, e
t
I, = eO.! e
l
N by property 4 of the matrix exponential.
The diagonal part is easy: e
lH
= diag(e
At
, ••• ,eAt). But e
lN
is almost as easy since N is
nilpotent of degree k.
Definition 11.8. A matrix M E jRnxn is nilpotent of degree (or index, or grade) p if
MP = 0, while MPI t= O.
For the matrix N defined above, it is easy to check that while N has l's along only
its first superdiagonal (and O's elsewhere), N
2
has l's along only its second superdiagonal,
and so forth. Finally, N
k

I
has a 1 in its (1, k) element and has O's everywhere else, and
N
k
= O. Thus, the series expansion of e
lN
is finite, i.e.,
Thus,
t
2
t
k

I
e
IN
=I+tN+N
2
+ ... + N
k

I
2! (k  I)!
o
o o
eAt
teAt
12 At
2I
e
0
eAt teAl
ell; =
0 0
eAt
0 0
t
1
IkI At
(kI)! e
12 At
2I
e
teAl
eAt
In the case when A is complex, a real version of the above can be worked out.
116 Chapter 11. Linear Differential and Difference Equations
Example 11.9. Let A = [ ~ _ \ J]. Then A (A) = {2, 2} and
Interpolation method
This method is numerically unstable in finiteprecision arithmetic but is quite effective for
hand calculation in smallorder problems. The method is stated and illustrated for the
exponential function but applies equally well to other functions.
Given A € E.
nxn
and /(A) = e
tx
, compute f(A) = e'
A
, where t is a fixed scalar.
Suppose the characteristic polynomial of A can be written as n ( X ) = Yi?=i (^ ~~ ^ i)" ' »
where the A.,  s are distinct. Define
where O TQ , . . . , a
n
i are n constants that are to be determined. They are, in fact, the unique
solution of the n equations:
Here, the superscript (&) denotes the fcth derivative with respect to X. With the a, s then
known, the function g is known and /(A) = g(A). The motivation for this method is
the CayleyHamilton Theorem, Theorem 9.3, which says that all powers of A greater than
n — 1 can be expressed as linear combinations of A
k
for k = 0, 1, . . . , n — 1. Thus, all the
terms of order greater than n — 1 in the power series for e'
A
can be written in terms of these
lowerorder powers as well. The polynomial g gives the appropriate linear combination.
Example 11.10. Let
and /(A) = e
tK
. Then j r(A.) = (A. + I)
3
, so m = 1 and n
{
= 3.
Let g(X) — UQ + a\X + o^A.
2
. Then the three equations for the a, s are given by
116 Chapter 11. Linear Differential and Difference Equations
Example 11.9.
Let A = [=i
a Then A(A) = {2, 2} and
etA = Xe
tJ
xI
=[
2 1
] exp t [
2
 ~ ] [
1
]
0
1 2
=[
2
] [ e ~ 2 t
te
2t
] [
1
]
1
e
2t
1 2
Interpolation method
This method is numerically unstable in finiteprecision arithmetic but is quite effective for
hand calculation in smallorder problems. The method is stated and illustrated for the
exponential function but applies equally well to other functions.
Given A E jRnxn and f(A) = etA, compute f(A) = etA, where t is a fixed scalar.
Suppose the characteristic polynomial of A can be written as n(A) = nr=1 (A  Ai t',
where the Ai s are distinct. Define
where ao, ... , anl are n constants that are to be determined. They are, in fact, the unique
solution of the n equations:
g(k)(Ai) = f(k)(Ai); k = 0, I, ... , ni  I, i Em.
Here, the superscript (k) denotes the kth derivative with respect to A. With the aiS then
known, the function g is known and f(A) = g(A). The motivation for this method is
the CayleyHamilton Theorem, Theorem 9.3, which says that all powers of A greater than
n  1 can be expressed as linear combinations of A k for k = 0, I, ... , n  1. Thus, all the
terms of order greater than n  1 in the power series for e
t
A can be written in terms of these
lowerorder powers as well. The polynomial g gives the appropriate linear combination.
Example 11.10. Let
A = [  ~  ~ ~ ]
o 01
and f(A) = etA. Then n(A) = (A + 1)3, so m = 1 and nl = 3.
Let g(A) = ao + alA + a2A2. Then the three equations for the aiS are given by
g(I) = f(1) ==> ao al +a2 = e
t
,
g'(1) = f'(1) ==> at  2a2 = te
t
,
g"(I) = 1"(1) ==> 2a2 = t
2
e
t
•
11.1. Differential Equations 117
Solving for the a, s, we find
Thus,
~4 4i t f f > \ t k TU^^ _/"i\ f \ i o\ 2
Example 11.11. Let A = [ _* J] and /(A) = e
a
. Then 7 r(X ) = (A + 2)
2
so m = 1 and
«i = 2.
Let g(A.) = «o + ofiA.. Then the defining equations for the a,s are given by
Solving for the a,s, we find
Thus,
Other methods
1. Use e
tA
= £~
l
{(sl — A)^
1
} and techniques for inverse Laplace transforms. This
is quite effective for smallorder problems, but general nonsymbolic computational
techniques are numerically unstable since the problem is theoretically equivalent to
knowing precisely a JCF.
2. Use Pade approximation. There is an extensive literature on approximating cer
tain nonlinear functions by rational functions. The matrix analogue yields e
A
=
11 .1. Differential Equations
117
Solving for the ai s, we find
Thus,
Example 11.11. Let A = [ : : : : ~ 6] and f(A) = eO. Then rr(A) = (A + 2)2 so m = 1 and
nL = 2.
Let g(A) = ao + aLA. Then the defining equations for the aiS are given by
g(2) = f(2) ==> ao  2al = e
2t
,
g'(2) = f'(2) ==> al = te
2t
.
Solving for the aiS, we find
Thus,
ao = e
2t
+ 2te
2t
,
aL = te
2t
.
f(A) = etA = g(A) = aoI + al A
= (e
2t
+ 2te
2t
) [ ~
_ [ e
2t
_ 2te
2t
 te
2t
Other methods
o ] + te
2t
[4 4 ]
I I 0
1. Use etA = .cI{(sI  A)I} and techniques for inverse Laplace transforms. This
is quite effective for smallorder problems, but general nonsymbolic computational
techniques are numerically unstable since the problem is theoretically equivalent to
knowing precisely a JCE
2. Use Pade approximation. There is an extensive literature on approximating cer
tain nonlinear functions by rational functions. The matrix analogue yields e
A
~
118 Chapter 11. Linear Differential and Difference Equations
D~
l
(A)N(A), where D(A) = 8
0
I + Si A H h S
P
A
P
and N(A) = v
0
I + v
l
A +
• • • + v
q
A
q
. Explicit formulas are known for the coefficients of the numerator and
denominator polynomials of various orders. Unfortunately, a Fade approximation for
the exponential is accurate only in a neighborhood of the origin; in the matrix case
this means when  A is sufficiently small. This can be arranged by scaling A, say, by
/ * \
2
*
multiplying it by 1/2* for sufficiently large k and using the fact that e
A
= ( e
{ ] / 2 )A
j .
Numerical loss of accuracy can occur in this procedure from the successive squarings.
3. Reduce A to (real) Schur form S via the unitary similarity U and use e
A
= Ue
s
U
H
and successive recursions up the superdiagonals of the (quasi) upper triangular matrix
e
s
.
4. Many methods are outlined in, for example, [19]. Reliable and efficient computation
of matrix functions such as e
A
and log(A) remains a fertile area for research.
11.2 Difference Equations
In this section we outline solutions of discretetime analogues of the linear differential
equations of the previous section. Linear discretetime systems, modeled by systems of
difference equations, exhibit many parallels to the continuoustime differential equation
case, and this observation is exploited frequently.
11.2.1 Homogeneous linear difference equations
Theorem 11.12. Let A e Rn
xn
. The solution of the linear homogeneous system of difference
equations
11.2.2 Inhomogeneous linear difference equations
Theorem 11.14. Let A e R
nxn
, B e R
nxm
and suppose {«*}£§ « a given sequence of
mvectors. Then the solution of the inhomogeneous initialvalue problem
for k > 0 is given by
Proof: The proof is almost immediate upon substitution of (11.14) into (11.13). D
Remark 11.13. Again, we restrict our attention only to the socalled timeinvariant
case, where the matrix A in (11.13) is constant and does not depend on k. We could also
consider an arbitrary "initial time" ko, but since the system is timeinvariant, and since we
want to keep the formulas "clean" (i.e., no double subscripts), we have chosen ko = 0 for
convenience.
118 Chapter 11. Linear Differential and Difference Equations
DI(A)N(A), where D(A) = 001 + olA + ... + opAP and N(A) = vol + vIA +
... + Vq A q. Explicit formulas are known for the coefficients of the numerator and
denominator polynomials of various orders. Unfortunately, a Pad6 approximation for
the exponential is accurate only in a neighborhood of the origin; in the matrix case
this means when IIAII is sufficiently small. This can be arranged by scaling A, say, by
2'
multiplying it by 1/2k for sufficiently large k and using the fact that e
A
= (e( I /2')A )
Numerical loss of accuracy can occur in this procedure from the successive squarings.
3. Reduce A to (real) Schur form S via the unitary similarity U and use e
A
= U e
S
U H
and successive recursions up the superdiagonals of the (quasi) upper triangular matrix
e
S
.
4. Many methods are outlined in, for example, [19]. Reliable and efficient computation
of matrix functions such as e
A
and 10g(A) remains a fertile area for research.
11.2 Difference Equations
In this section we outline solutions of discretetime analogues of the linear differential
equations of the previous section. Linear discretetime systems, modeled by systems of
difference equations, exhibit many parallels to the continuoustime differential equation
case, and this observation is exploited frequently.
11.2.1 Homogeneous linear difference equations
Theorem 11.12. Let A E jRn xn. The solution of the linear homogeneous system of difference
equations
(11.13)
for k 2:: 0 is given by
Proof: The proof is almost immediate upon substitution of (11.14) into (11.13). 0
Remark 11.13. Again, we restrict our attention only to the socalled timeinvariant
case, where the matrix A in (11.13) is constant and does not depend on k. We could also
consider an arbitrary "initial time" ko, but since the system is timeinvariant, and since we
want to keep the formulas "clean" (i.e., no double subscripts), we have chosen ko = 0 for
convenience.
11.2.2 Inhomogeneous linear difference equations
Theorem 11.14. Let A E jRnxn, B E jRnxm and suppose { u d t ~ is a given sequence of
mvectors. Then the solution of the inhomogeneous initialvalue problem
(11.15)
11.2. Difference Equations 119
is given by
11.2.3 Computation of matrix powers
It is clear that solution of linear systems of difference equations involves computation of
A
k
. One solution method, which is numerically unstable but sometimes useful for hand
calculation, is to use ztransforms, by analogy with the use of Laplace transforms to compute
a matrix exponential. One definition of the ztransform of a sequence {gk} is
Assuming z > max A, the ztransform of the sequence {A
k
} is then given by
X€A(A)
Proof: The proof is again almost immediate upon substitution of (11.16) into (11.15). D
Methods based on the JCF are sometimes useful, again mostly for smallorder prob
lems. Assume that A e M"
xn
and let X e R^
n
be such that X~
1
AX = /, where J is a
JCF for A. Then
If A is diagonalizable, it is then easy to compute A
k
via the formula A
k
— XJ
k
X
l
since /* is simply a diagonal matrix.
11.2. Difference Equations
is given by
kI
xk=AkXO+LAkjIBUj, k:::.O.
j=O
119
(11.16)
Proof: The proof is again almost immediate upon substitution of (11.16) into (11.15). 0
11.2.3 Computation of matrix powers
It is clear that solution of linear systems of difference equations involves computation of
A k. One solution method, which is numerically unstable but sometimes useful for hand
calculation, is to use ztransforms, by analogy with the use of Laplace transforms to compute
a matrix exponential. One definition of the ztransform of a sequence {gk} is
+00
= LgkZ
k
.
k=O
Assuming Izl > max IAI, the ztransform of the sequence {Ak} is then given by
AEA(A)
+00
k "'kk 1 12
Z({A})=L...zA =I+A+"2A + ...
k=O z z
= (lzIA)I
= z(zI  A)I.
Methods based on the JCF are sometimes useful, again mostly for smallorder prob
lems. Assume that A E jRnxn and let X E be such that XI AX = J, where J is a
JCF for A. Then
Ak = (XJXI)k
= XJkX
1
_I
 m
LXi Jty
i
H
;=1
if A is diagonalizable,
in general.
If A is diagonalizable, it is then easy to compute Ak via the formula Ak = X Jk XI
since Jk is simply a diagonal matrix.
120 Chapter 11. Linear Differential and Difference Equations
In the general case, the problem again reduces to the computation of the power of a
Jordan block. To be specific, let 7, e C
pxp
be a Jordan block of the form
Writing /,• = XI + N and noting that XI and the nilpotent matrix N commute, it is
then straightforward to apply the binomial theorem to (XI + N)
k
and verify that
The symbol ( ) has the usual definition of ,
(
^ ., and is to be interpreted as 0 if k < q.
In the case when A. is complex, a real version of the above can be worked out.
4
Example 11.15. Let A = [_J J]. Then
Basic analogues of other methods such as those mentioned in Section 11.1.6 can also
be derived for the computation of matrix powers, but again no universally "best" method
exists. For an erudite discussion of the state of the art, see [11, Ch. 18].
11.3 HigherOrder Equations
It is well known that a higherorder (scalar) linear differential equation can be converted to
a firstorder linear system. Consider, for example, the initialvalue problem
with 4 > (t } a given function and n initial conditions
120 Chapter 11. Linear Differential and Difference Equations
In the general case, the problem again reduces to the computation of the power of a
Jordan block. To be specific, let J
i
E Cpxp be a Jordan block of the form
o ... 0 A
Writing J
i
= AI + N and noting that AI and the nilpotent matrix N commute, it is
then straightforward to apply the binomial theorem to (AI + N)k and verify that
Ak
kA kI
(;)A
k

2
(
k ) AkP+I
pl
0
Ak kA
k

1
J/ =
0 0
Ak
( ; ) A
k

2
kA
k

1
0 0
Ak
The symbol (: ) has the usual definition of q ! ( k k ~ q ) ! and is to be interpreted as 0 if k < q.
In the case when A is complex, a real version of the above can be worked out.
Example 11.15. Let A = [=i a Then
Ak = XJkX1 = [2 1 ] [(_2)k k(2)kk
1
] [ 1 2
1
]
1 1 0 (2) 1
_ [ (_2/
1
(2  2k) k( 2l+
1
]
 k( _2)k1 (2l
1
(2k  2) .
Basic analogues of other methods such as those mentioned in Section 11.1.6 can also
be derived for the computation of matrix powers, but again no universally "best" method
exists. For an erudite discussion of the state of the art, see [11, Ch. 18].
11.3 HigherOrder Equations
It is well known that a higherorder (scalar) linear differential equation can be converted to
a firstorder linear system. Consider, for example, the initialvalue problem
(11.17)
with ¢J(t) a given function and n initial conditions
y(O) = Co, y(O) = CI, ... , inI)(O) = CnI' (1l.l8)
Exercises 121
Here, v
(m)
denotes the mth derivative of y with respect to t. Define a vector x (?) e R" with
components *i(0 = y ( t ) , x
2
( t) = y ( t ) , . . . , x
n
( t) = y
{ n
~
l )
( t ) . Then
These equations can then be rewritten as the firstorder linear system
The initial conditions take the form ^(0) = c = [ C Q , c\, ..., C
M
_ I ] .
Note that det(X7 — A) = A." + a
n
\X
n
~
l
H h a\X + ao. However, the companion
matrix A in (11.19) possesses many nasty numerical properties for even moderately sized n
and, as mentioned before, is often well worth avoiding, at least for computational purposes.
A similar procedure holds for the conversion of a higherorder difference equation
EXERCISES
1. Let P € R
nxn
be a projection. Show that e
p
% / + 1.718P.
2. Suppose x, y € R" and let A = xy
T
. Further, let a = x
T
y. Show that e'
A
I + g ( t , a) xy
T
, where
3. Let
with n initial conditions, into a linear firstorder difference equation with (vector) initial
condition.
Exercises 121
Here, y(m) denotes the mth derivative of y with respect to t. Define a vector x (t) E ]Rn with
components Xl (t) = yet), X2(t) = yet), ... , Xn(t) = Inl)(t). Then
Xl (I) = X2(t) = y(t),
X2(t) = X3(t) = yet),
Xnl (t) = Xn(t) = y(nl)(t),
Xn(t) = y(n)(t) = aoy(t)  aly(t)  ...  an_llnl)(t) + ¢(t)
= aOx\ (t)  a\X2(t)  ...  anlXn(t) + ¢(t).
These equations can then be rewritten as the firstorder linear system
0 0 0
0 0 1
x(t)+ [ n ~ ( t )
x(t) =
0
0 0 1
ao a\ a
n
\
The initial conditions take the form X (0) = C = [co, Cl, •.. , C
n
\ r.
(11.19)
Note that det(A!  A) = An + an_1A
n

1
+ ... + alA + ao. However, the companion
matrix A in (11.19) possesses many nasty numerical properties for even moderately sized n
and, as mentioned before, is often well worth avoiding, at least for computational purposes.
A similar procedure holds for the conversion of a higherorder difference equation
with n initial conditions, into a linear firstorder difference equation with (vector) initial
condition.
EXERCISES
1. Let P E lR
nxn
be a projection. Show that e
P
~ ! + 1.718P.
2. Suppose x, y E lR
n
and let A = xyT. Further, let a = XT y. Show that etA
1+ get, a)xyT, where
{
!(eat  I)
g(t,a)= a t
3. Let
if a 1= 0,
if a = O.
122 Chapter 11. L i n ear Di f f eren ti al and Di f f erence Equati on s
where X e M'
nx
" is arbitrary. Show that
4. Let K denote the skewsymmetric matrix
where /„ denotes the n x n identity matrix. A matrix A e R
2n x2n
is said to be
Hamiltonian if K~
1
A
T
K = A and to be symplectic if K~
l
A
T
K  A
1
.
(a) Suppose E is Hamiltonian and let A, be an eigenvalue of H. Show that — A, must
also be an eigenvalue of H.
(b) Suppose S is symplectic and let A. be an eigenvalue of S. Show that 1 /A, must
also be an eigenvalue of S.
(c) Suppose that H is Hamiltonian and S is symplectic. Show that S~
1
HS must be
Hamiltonian.
(d) Suppose H is Hamiltonian. Show that e
H
must be symplectic.
5. Let a, ft € R and
Then show that
6. Find a general expression for
7. Find e
M
when A =
5. Let
(a) Solve the differential equation
122 Chapter 11. Linear Differential and Difference Equations
where X E jRmxn is arbitrary. Show that
e
A = [eo I sinh 1 X ]
~ I .
4. Let K denote the skewsymmetric matrix
[
0 In ]
In 0 '
where In denotes the n x n identity matrix. A matrix A E jR2nx2n is said to be
Hamiltonian if K I AT K =  A and to be symplectic if K I AT K = A I.
(a) Suppose H is Hamiltonian and let).. be an eigenvalue of H. Show that ).. must
also be an eigenvalue of H.
(b) Suppose S is symplectic and let).. be an eigenvalue of S. Show that 1/).. must
also be an eigenValue of S.
(c) Suppose that H is Hamiltonian and S is symplectic. Show that SI H S must be
Hamiltonian.
(d) Suppose H is Hamiltonian. Show that e
H
must be symplectic.
5. Let a, f3 E lR and
Then show that
6. Find a general expression for
7. Find etA when A =
8. Let
ectt cos f3t
_eut sin f3t
ectctrt sin ~ t J.
e cos/A
(a) Solve the differential equation
i = Ax ; x(O) = [ ~ J.
Exercises 123
Show that the eigenvalues of the solution X ( t ) of this problem are the same as those
of Cf or all?.
11. The year is 2004 and there are three large "free trade zones" in the world: Asia (A),
Europe (E), and the Americas (R). Suppose certain multinational companies have
total assets of $40 trillion of which $20 trillion is in E and $20 trillion is in R. Each
year half of the Americas' money stays home, a quarter goes to Europe, and a quarter
goes to Asia. For Europe and Asia, half stays home and half goes to the Americas.
(a) Find the matrix M that gives
(b) Find the eigenvalues and right eigenvectors of M.
(c) Find the distribution of the companies' assets at year k.
(d) Find the limiting distribution of the $40 trillion as the universe ends, i.e., as
k — » • +00 (i.e., around the time the Cubs win a World Series).
(Exercise adapted from Problem 5.3.11 in [24].)
(b) Solve the differential equation
9. Consider the initialvalue problem
for t > 0. Suppose that A e E"
x
" is skewsymmetric and let a = \\XQ\\
2
. Show that
*(OII
2
= af or al l f > 0.
10. Consider the n x n matrix initialvalue problem
12. (a) Find the solution of the initialvalue problem
(b) Consider the difference equation
If £
0
= 1 and z\ = 2, what is the value of Z IQ OO? What is the value of Zk in
general?
Exercises 123
(b) Solve the differential equation
i = Ax + b; x(O) = [ ~ l
9. Consider the initialvalue problem
i(t) = Ax(t); x(O) = Xo
for t ~ O. Suppose that A E ~ n x n is skewsymmetric and let ex = Ilxol12. Show that
I/X(t)1/2 = ex for all t > O.
10. Consider the n x n matrix initialvalue problem
X(t) = AX(t)  X(t)A; X(O) = c.
Show that the eigenvalues of the solution X (t) of this problem are the same as those
of C for all t.
11. The year is 2004 and there are three large "free trade zones" in the world: Asia (A),
Europe (E), and the Americas (R). Suppose certain multinational companies have
total assets of $40 trillion of which $20 trillion is in E and $20 trillion is in R. Each
year half of the Americas' money stays home, a quarter goes to Europe, and a quarter
goes to Asia. For Europe and Asia, half stays home and half goes to the Americas.
(a) Find the matrix M that gives
[
A] [A]
E =M E
R year k+1 R year k
(b) Find the eigenvalues and right eigenvectors of M.
(c) Find the distribution of the companies' assets at year k.
(d) Find the limiting distribution of the $40 trillion as the universe ends, i.e., as
k * +00 (i.e., around the time the Cubs win a World Series).
(Exercise adapted from Problem 5.3.11 in [24].)
12. (a) Find the solution of the initialvalue problem
.Yet) + 2y(t) + yet) = 0; yeO) = 1, .YeO) = O.
(b) Consider the difference equation
Zk+2 + 2Zk+1 + Zk = O.
If Zo = 1 and ZI = 2, what is the value of ZIOOO? What is the value of Zk in
general?
This page intentionally left blank This page intentionally left blank
Chapter 12
Generalized Eigenvalue
Problems
12.1 The Generalized Eigenvalue/Eigenvector Problem
In this chapter we consider the generalized eigenvalue problem
125
where A, B e C"
xn
. The standard eigenvalue problem considered in Chapter 9 obviously
corresponds to the special case that B = I.
Definition 12.1. A nonzero vector x e C" is a right generalized eigenvector of the pair
(A, B) with A, B e C
MX
" if there exists a scalar A. e C, called a generalized eigenvalue,
such that
Similarly, a nonzero vector y e C" is a left generalized eigenvector corresponding to an
eigenvalue X if
When the context is such that no confusion can arise, the adjective "generalized"
is usually dropped. As with the standard eigenvalue problem, if x [y] is a right [left]
eigenvector, then so is ax [ay] for any nonzero scalar a. e C.
Definition 12.2. The matrix A — X B is called a matrix pencil (or pencil of the matrices A
and B).
As with the standard eigenvalue problem, eigenvalues for the generalized eigenvalue
problem occur where the matrix pencil A — X B is singular.
Definition 12.3. The polynomial 7 r(A.) = det(A — A.5) is called the characteristic poly
nomial of the matrix pair (A, B) . The roots ofn(X .) are the eigenvalues of the associated
generalized eigenvalue problem.
Remark 12.4. When A, B e E"
xn
, the characteristic polynomial is obviously real, and
hence nonreal eigenvalues must occur in complex conjugate pairs.
Chapter 12
Generalized Eigenvalue
Problems
12.1 The Generalized Eigenvalue/Eigenvector Problem
In this chapter we consider the generalized eigenvalue problem
Ax = 'ABx,
where A, B E e
nxn
. The standard eigenvalue problem considered in Chapter 9 obviously
corresponds to the special case that B = I.
Definition 12.1. A nonzero vector x E en is a right generalized eigenvector of the pair
(A, B) with A, B E e
nxn
if there exists a scalar 'A E e, called a generalized eigenvalue,
such that
Ax = 'ABx. (12.1)
Similarly, a nonzero vector y E en is a left generalized eigenvector corresponding to an
eigenvalue 'A if
(12.2)
When the context is such that no confusion can arise, the adjective "generalized"
is usually dropped. As with the standard eigenvalue problem, if x [y] is a right [left]
eigenvector, then so is ax [ay] for any nonzero scalar a E <C.
Definition 12.2. The matrix A  'AB is called a matrix pencil (or pencil of the matrices A
and B).
As with the standard eigenvalue problem, eigenvalues for the generalized eigenvalue
problem occur where the matrix pencil A  'AB is singular.
Definition 12.3. The polynomial n('A) = det(A  'AB) is called the characteristic poly
nomial of the matrix pair (A, B). The roots ofn('A) are the eigenvalues of the associated
generalized eigenvalue problem.
Remark 12.4. When A, B E jRnxn, the characteristic polynomial is obviously real, and
hence nonreal eigenvalues must occur in complex conjugate pairs.
125
and there are again four cases to consider.
Case 1: a ^ 0, ft ^ 0. There are two eigenvalues, 1 and ^.
Case 2: a = 0, ft ^ 0. There is only one eigenvalue, 1 (of multiplicity 1).
Case 3: a ^ 0, f3 = 0. There are two eigenvalues, 1 and 0.
Case 4: a = 0, (3 = 0. All A 6 C are eigenvalues since det(B — uA) = 0.
At least for the case of regular pencils, it is apparent where the "missing" eigenvalues have
gone in Cases 2 and 3. That is to say, there is a second eigenvalue "at infinity" for Case 3 of
A — A.B, with its reciprocal eigenvalue being 0 in Case 3 of the reciprocal pencil B — nA.
A similar reciprocal symmetry holds for Case 2.
While there are applications in system theory and control where singular pencils
appear, only the case of regular pencils is considered in the remainder of this chapter. Note
that A and/or B may still be singular. If B is singular, the pencil A — KB always has
126 Chapter 12. Generalized Eigenvalue Problems
Remark 12.5. If B = I (or in general when B is nonsingular), then n ( X ) is a polynomial
of degree n, and hence there are n eigenvalues associated with the pencil A — X B. However,
when B = I, in particular, when B is singular, there may be 0, k e n, or infinitely many
eigenvalues associated with the pencil A — X B. For example, suppose
where a and ft are scalars. Then the characteristic polynomial is
and there are several cases to consider.
Case 1: a ^ 0, ft ^ 0. There are two eigenvalues, 1 and .
Case 2: a = 0, f3 / 0. There are two eigenvalues, 1 and 0.
Case 3: a = 0, f3 = 0. There is only one eigenvalue, 1 (of multiplicity 1).
Case 4: a = 0, f3 = 0. All A e C are eigenvalues since det(A — A. B ) =0.
Definition 12.6. If del (A — X B) is not identically zero, the pencil A — X B is said to be
regular; otherwise, it is said to be singular.
Note that if AA(A) n J\f(B) ^ 0, the associated matrix pencil is singular (as in Case
4 above).
Associated with any matrix pencil A — X B is a reciprocal pencil B — n,A and cor
responding generalized eigenvalue problem. Clearly the reciprocal pencil has eigenvalues
(JL = £. It is instructive to consider the reciprocal pencil associated with the example in
Remark 12.5. With A and B as in (12.3), the characteristic polynomial is
126 Chapter 12. Generalized Eigenvalue Problems
Remark 12.5. If B = I (or in general when B is nonsingular), then rr(A) is a polynomial
of degree n, and hence there are n eigenvalues associated with the pencil A  AB. However,
when B =I I, in particular, when B is singular, there may be 0, k E !!, or infinitely many
eigenvalues associated with the pencil A  AB. For example, suppose
where a and (3 are scalars. Then the characteristic polynomial is
det(A  AB) = (I  AHa  (3A)
and there are several cases to consider.
Case 1: a =I 0, {3 =I O. There are two eigenvalues, I and ~ .
Case 2: a = 0, {3 =I O. There are two eigenvalues, I and O.
Case 3: a =I 0, {3 = O. There is only one eigenvalue, I (of multiplicity 1).
Case 4: a = 0, (3 = O. All A E C are eigenvalues since det(A  AB) == O.
(12.3)
Definition 12.6. If det(A  AB) is not identically zero, the pencil A  AB is said to be
regular; otherwise, it is said to be singular.
Note that if N(A) n N(B) =I 0, the associated matrix pencil is singular (as in Case
4 above).
Associated with any matrix pencil A  AB is a reciprocal pencil B  /.LA and cor
responding generalized eigenvalue problem. Clearly the reciprocal pencil has eigenvalues
/.L = ±. It is instructive to consider the reciprocal pencil associated with the example in
Remark 12.5. With A and B as in (12.3), the characteristic polynomial is
det(B  /.LA) = (1  /.L)({3  a/.L)
and there are again four cases to consider.
Case 1: a =I 0, {3 =I O. There are two eigenvalues, I and ~ .
Case 2: a = 0, {3 =I O. There is only one eigenvalue, I (of multiplicity I).
Case 3: a =I 0, {3 = O. There are two eigenvalues, 1 and O.
Case 4: a = 0, (3 = O. All A E C are eigenvalues since det(B  /.LA) == O.
At least for the case of regular pencils, it is apparent where the "missing" eigenvalues have
gone in Cases 2 and 3. That is to say, there is a second eigenvalue "at infinity" for Case 3 of
A  AB, with its reciprocal eigenvalue being 0 in Case 3 of the reciprocal pencil B  /.LA.
A similar reciprocal symmetry holds for Case 2.
While there are applications in system theory and control where singular pencils
appear, only the case of regular pencils is considered in the remainder of this chapter. Note
that A and/or B may still be singular. If B is singular, the pencil A  AB always has
12. 2. Canonical Forms 127
fewer than n eigenvalues. If B is nonsingular, the pencil A A. f i always has precisely n
eigenvalues, since the generalized eigenvalue problem is then easily seen to be equivalent
to the standard eigenvalue problem B~
l
Ax = Xx (or AB~
l
w = Xw). However, this turns
out to be a very poor numerical procedure for handling the generalized eigenvalue problem
if B is even moderately ill conditioned with respect to inversion. Numerical methods that
work directly on A and B are discussed in standard textbooks on numerical linear algebra;
see, for example, [7, Sec. 7.7] or [25, Sec. 6.7].
12.2 Canonical Forms
Just as for the standard eigenvalue problem, canonical forms are available for the generalized
eigenvalue problem. Since the latter involves a pair of matrices, we now deal with equiva
lencies rather than similarities, and the first theorem deals with what happens to eigenvalues
and eigenvectors under equivalence.
Theorem 12.7. Let A, fl, Q, Z e C
nxn
with Q and Z nonsingular. Then
1. the eigenvalues of the problems A — XB and QAZ — XQBZ are the same (the two
problems are said to be equivalent).
2. ifx is a right eigenvector of A—XB, then Z~
l
x is a right eigenvector of QAZ—XQ B Z.
3. ify is a left eigenvector of A —KB, then Q~
H
y isa left eigenvector ofQAZ — XQBZ.
Proof:
1. det(QAZXQBZ) = det[0(A  XB)Z] = det gdet Zdet(A  XB). Since det 0
and det Z are nonzero, the result follows.
2. The result follows by noting that (A – yB)x  Oif andonly if Q(AXB)Z(Z~
l
x) =
0.
3. Again, the result follows easily by noting that y
H
(A — XB) — 0 if and only if
( Q~
H
y )
H
Q( A– XB ) Z = Q. D
where T
a
and Tp are upper triangular.
By Theorem 12.7, the eigenvalues of the pencil A — XB are then the ratios of the diag
onal elements of T
a
to the corresponding diagonal elements of Tp, with the understanding
that a zero diagonal element of Tp corresponds to an infinite generalized eigenvalue.
There is also an analogue of the MurnaghanWintner Theorem for real matrices.
The first canonical form is an analogue of Schur's Theorem and forms, in fact, the
theoretical foundation for the QZ algorithm, which is the generally preferred method for
solving the generalized eigenvalue problem; see, for example, [7, Sec. 7.7] or [25, Sec. 6.7].
Theorem 12.8. Let A, B e Cn
xn
. Then there exist unitary matrices Q, Z e Cn
xn
such that
12.2. Canonical Forms 127
fewer than n eigenvalues. If B is nonsingular, the pencil A  AB always has precisely n
eigenvalues, since the generalized eigenvalue problem is then easily seen to be equivalent
to the standard eigenvalue problem B
1
Ax = Ax (or AB
1
W = AW). However, this turns
out to be a very poor numerical procedure for handling the generalized eigenvalue problem
if B is even moderately ill conditioned with respect to inversion. Numerical methods that
work directly on A and B are discussed in standard textbooks on numerical linear algebra;
see, for example, [7, Sec. 7.7] or [25, Sec. 6.7].
12.2 Canonical Forms
Just as for the standard eigenvalue problem, canonical forms are available for the generalized
eigenvalue problem. Since the latter involves a pair of matrices, we now deal with equiva
lencies rather than similarities, and the first theorem deals with what happens to eigenvalues
and eigenvectors under equivalence.
Theorem 12.7. Let A, B, Q, Z E c
nxn
with Q and Z nonsingular. Then
1. the eigenvalues of the problems A  AB and QAZ  AQBZ are the same (the two
problems are said to be equivalent).
2. ifx isa right eigenvector of AAB, then Zl x isa righteigenvectorofQAZAQB Z.
3. ify isa left eigenvector of A AB, then QH y isa lefteigenvectorofQAZ AQBZ.
Proof:
1. det(QAZ  AQBZ) = det[Q(A  AB)Z] = det Q det Z det(A  AB). Since det Q
and det Z are nonzero, the result follows.
2. The result follows by noting that (A AB)x = 0 if and only if Q(A AB)Z(Zl x) =
o.
3. Again, the result follows easily by noting that yH (A  AB) o if and only if
(QH y)H Q(A _ AB)Z = O. 0
The first canonical form is an analogue of Schur's Theorem and forms, in fact, the
theoretical foundation for the QZ algorithm, which is the generally preferred method for
solving the generalized eigenvalue problem; see, for example, [7, Sec. 7.7] or [25, Sec. 6.7].
Theorem 12.8. Let A, B E c
nxn
. Then there exist unitary matrices Q, Z E c
nxn
such that
QAZ = T
a
, QBZ = T
fJ
,
where Ta and TfJ are upper triangular.
By Theorem 12.7, the eigenvalues ofthe pencil A  AB are then the ratios of the diag
onal elements of Ta to the corresponding diagonal elements of T
fJ
, with the understanding
that a zero diagonal element of TfJ corresponds to an infinite generalized eigenvalue.
There is also an analogue of the MurnaghanWintner Theorem for real matrices.
128 Chapter 12. Generalized Eigenvalue Problems
Theorem 12.9. Let A, B e R
nxn
. Then there exist orthogonal matrices Q, Z e R"
xn
such
thnt
where T is upper triangular and S is quasiuppertriangular.
When S has a 2 x 2 diagonal block, the 2 x 2 subpencil formed with the corresponding
2x2 diagonal subblock of T has a pair of complex conjugate eigenvalues. Otherwise, real
eigenvalues are given as above by the ratios of diagonal elements of S to corresponding
elements of T.
There is also an analogue of the Jordan canonical form called the Kronecker canonical
form (KCF). A full description of the KCF, including analogues of principal vectors and
so forth, is beyond the scope of this book. In this chapter, we present only statements of
the basic theorems and some examples. The first theorem pertains only to "square" regular
pencils, while the full KCF in all its generality applies also to "rectangular" and singular
pencils.
Theorem 12.10. Let A, B e C
nxn
and suppose the pencil A — XB is regular. Then there
exist nonsingular matrices P, Q € C"
x
" such that
where J is a Jordan canonical form corresponding to the finite eigenvalues of A A.fi and
N is a nilpotent matrix of Jordan blocks associated with 0 and corresponding to the infinite
eigenvalues of A — XB.
Example 12.11. The matrix pencil
with characteristic polynomial (X — 2)
2
has a finite eigenvalue 2 of multiplicty 2 and three
infinite eigenvalues.
Theorem 12.12 (Kronecker Canonical Form). Let A, B e C
mxn
. Then there exist
nonsingular matrices P e C
mxm
and Q e C
nxn
such that
128 Chapter 12. Generalized Eigenvalue Problems
Theorem 12.9. Let A, B E jRnxn. Then there exist orthogonal matrices Q, Z E jRnxn such
that
QAZ = S, QBZ = T,
where T is upper triangular and S is quasiuppertriangular.
When S has a 2 x 2 diagonal block, the 2 x 2 subpencil fonned with the corresponding
2 x 2 diagonal subblock of T has a pair of complex conjugate eigenvalues. Otherwise, real
eigenvalues are given as above by the ratios of diagonal elements of S to corresponding
elements of T.
There is also an analogue of the Jordan canonical fonn called the Kronecker canonical
form (KeF). A full description of the KeF, including analogues of principal vectors and
so forth, is beyond the scope of this book. In this chapter, we present only statements of
the basic theorems and some examples. The first theorem pertains only to "square" regular
pencils, while the full KeF in all its generality applies also to "rectangular" and singular
pencils.
Theorem 12.10. Let A, B E c
nxn
and suppose the pencil A  AB is regular. Then there
exist nonsingular matrices P, Q E c
nxn
such that
peA  AB)Q = [ ~ ~ ]  A [ ~ ~ l
where J is a Jordan canonical form corresponding to the finite eigenvalues of A  AB and
N is a nilpotent matrix of Jordan blocks associated with 0 and corresponding to the infinite
eigenvalues of A  AB.
Example 12.11. The matrix pencil
[2 I
0 0
~ ]> [ ~
0 0
o 0] o 2 0 0 I 0 o 0
o 0 1 0 0 0 I 0
o 0 0 1 0 0 o 0
o 0 0 0 0 0 0 0
with characteristic polynomial (A  2)2 has a finite eigenvalue 2 of multiplicty 2 and three
infinite eigenvalues.
Theorem 12.12 (Kronecker Canonical Form). Let A, B E c
mxn
• Then there exist
nonsingular matrices P E c
mxm
and Q E c
nxn
such that
peA  AB)Q = diag(LII' ... , L
l
" L ~ , ...• L;'. J  A.I, I  )"N),
12.2. Canonical Forms 129
where N is nilpotent, both N and J are in Jordan canonical form, and L^ is the (k + 1) x k
bidiagonal pencil
The /( are called the left minimal indices while the r, are called the right minimal indices.
Left or right minimal indices can take the value 0.
Such a matrix is in KCF. The first block of zeros actually corresponds to LQ, LQ, LQ, LQ ,
LQ, where each LQ has "zero columns" and one row, while each LQ has "zero rows" and
one column. The second block is L\ while the third block is L\. The next two blocks
correspond to
Just as sets of eigenvectors span Ainvariant subspaces in the case of the standard
eigenproblem (recall Definition 9.35), there is an analogous geometric concept for the
generalized eigenproblem.
Definition 12.14. Let A, B e W
lxn
and suppose the pencil A — XB is regular. Then V is a
deflating subspace if
Just as in the standard eigenvalue case, there is a matrix characterization of deflating
subspace. Specifically, suppose S e R n*
xk
is a matrix whose columns span a ^dimensional
subspace S of R
n
, i.e., R ( S) = <S. Then S is a deflating subspace for the pencil A — XB if
and only if there exists M e R
kxk
such that
while the nilpotent matrix N in this example is
12.2. Canonical Forms 129
where N is nilpotent, both Nand J are in Jordan canonical form, and Lk is the (k + I) x k
bidiagonal pencil
A 0 0
A
Lk =
0 0
A
0 0 I
The Ii are called the left minimal indices while the ri are called the right minimal indices.
Left or right minimal indices can take the value O.
Example 12.13. Consider a 13 x 12 block diagonal matrix whose diagonal blocks are
A 0]
I A .
o I
Such a matrix is in KCF. The first block of zeros actually corresponds to Lo, Lo, Lo, L6,
L6, where each Lo has "zero columns" and one row, while each L6 has "zero rows" and
one column. The second block is L\ while the third block is LI The next two blocks
correspond to
[
21
J = 0 2
o 0
while the nilpotent matrix N in this example is
000
Just as sets of eigenvectors span Ainvariant subspaces in the case of the standard
eigenproblem (recall Definition 9.35), there is an analogous geometric concept for the
generalized eigenproblem.
Definition 12.14. Let A, B E and suppose the pencil A  AB is regular. Then V is a
deflating subspace if
dim(AV + BV) = dimV. (12.4)
Just as in the standard eigenvalue case, there is a matrix characterization of deflating
subspace. Specifically, suppose S E is a matrix whose columns span a kdimensional
subspace S of i.e., n(S) = S. Then S is a deflating subspace for the pencil A  AB if
and only if there exists M E such that
AS = BSM. (12.5)
130 Chapter 12. Generalized Eigenvalue Problems
If B = /, then (12.4) becomes dim(AV + V) = dimV, which is clearly equivalent to
AV c V. Similarly, (12.5) becomes AS = SM as before. If the pencil is not regular, there
is a concept analogous to deflating subspace called a reducing subspace.
12.3 Application to the Computation of System Zeros
Consider the linear svstem
which has a root at —2.8 .
The method of finding system zeros via a generalized eigenvalue problem also works
well for general multiinput, multioutput systems. Numerically, however, one must be
careful first to "deflate out" the infinite zeros (infinite eigenvalues of (12.6)). This is accom
plished by computing a certain unitary equivalence on the system pencil that then yields a
smaller generalized eigenvalue problem with only finite generalized eigenvalues (the finite
zeros).
The connection between system zeros and the corresponding system pencil is non
trivial. However, we offer some insight below into the special case of a singleinput,
with A € M
n x n
, B € R"
x m
, C e R
pxn
, and D € R
pxm
. This linear timeinvariant state
space model is often used in multivariable control theory, where x(= x(t)) is called the state
vector, u is the vector of inputs or controls, and y is the vector of outputs or observables.
For details, see, for example, [26].
In general, the (finite) zeros of this system are given by the (finite) complex numbers
z, where the "system pencil"
drops rank. In the special case p = m, these values are the generalized eigenvalues of the
(n + m) x (n + m) pencil.
Example 12.15. Let
Then the transfer matrix (see [26]) of this system is
which clearly has a zero at —2.8 . Checking the finite eigenvalues of the pencil (12.6), we
find the characteristic polynomial to be
130 Chapter 12. Generalized Eigenvalue Problems
If B = I, then (12.4) becomes dim (A V + V) = dim V, which is clearly equivalent to
AV ~ V. Similarly, (12.5) becomes AS = SM as before. lEthe pencil is not regular, there
is a concept analogous to deflating subspace called a reducing subspace.
12.3 Application to the Computation of System Zeros
Consider the linear system
i = Ax + Bu,
y = Cx + Du
with A E jRnxn, B E jRnxm, C E jRPxn, and D E jRPxm. This linear timeinvariant state
space model is often used in multivariable control theory, where x(= x(t)) is called the state
vector, u is the vector of inputs or controls, and y is the vector of outputs or observables.
For details, see, for example, [26].
In general, the (finite) zeros of this system are given by the (finite) complex numbers
z, where the "system pencil"
(12.6)
drops rank. In the special case p = m, these values are the generalized eigenvalues of the
(n + m) x (n + m) pencil.
Example 12.15. Let
A=[
4
2
Then the transfer matrix (see [26)) of this system is
C = [I 2],
55 + 14
g(5)=C(sIA)'B+D= 2 '
5 + 3s + 2
D=O.
which clearly has a zero at 2.8. Checking the finite eigenvalues of the pencil (12.6), we
find the characteristic polynomial to be
det
[
A c
M
B]
D "'" 5A + 14,
which has a root at 2.8.
The method of finding system zeros via a generalized eigenvalue problem also works
well for general mUltiinput, multioutput systems. Numerically, however, one must be
careful first to "deflate out" the infinite zeros (infinite eigenvalues of (12.6». This is accom
plished by computing a certain unitary equivalence on the system pencil that then yields a
smaller generalized eigenvalue problem with only finite generalized eigenvalues (the finite
zeros).
The connection between system zeros and the corresponding system pencil is non
trivial. However, we offer some insight below into the special case of a singleinput.
12.4. Symmetric Generalized Eigenvalue Problems 131
singleoutput system. Specifically, let B = b e Rn, C = c
1
e R
l xn
, and D = d e R.
Furthermore, let g(.s) = c
r
(s7 — A )~
!
Z ? + d denote the system transfer function (matrix),
and assume that g ( s ) can be written in the form
where T T (S ) is the characteristic polynomial of A, and v(s) and T T (S ) are relatively prime
(i.e., there are no "pole/zero cancellations").
Suppose z € C is such that
is singular. Then there exists a nonzero solution to
or
Assuming z is not an eigenvalue of A (i.e., no pole/zero cancellations), then from (12.7) we
get
Substituting this in (12.8), we have
or g ( z ) y = 0 by the definition of g . Now _ y ^ 0 (else x = 0 from (12.9)). Hence g(z) = 0,
i.e., z is a zero of g.
12.4 Symmetric Generalized Eigenvalue Problems
A very important special case of the generalized eigenvalue problem
for A, B e R
nxn
arises when A = A and B = B
1
> 0. For example, the secondorder
system of differential equations
where M is a symmetric positive definite "mass matrix" and K is a symmetric "stiffness
matrix," is a frequently employed model of structures or vibrating systems and yields a
generalized eigenvalue problem of the form (12.10).
Since B is positive definite it is nonsingular. Thus, the problem (12.10) is equivalent
to the standard eigenvalue problem B~
l
Ax = A J C. However, B~
1
A is not necessarily
symmetric.
12.4. Symmetric Generalized Eigenvalue Problems 131
singleoutput system. Specifically, let B = b E ffi.n, C = c
T
E ffi.l xn, and D = d E R
Furthermore, let g(s) = c
T
(s I  A) 1 b + d denote the system transfer function (matrix),
and assume that g(s) can be written in the form
v(s)
g(s) = n(s)'
where n(s) is the characteristic polynomial of A, and v(s) and n(s) are relatively prime
(i.e., there are no "pole/zero cancellations").
Suppose Z E C is such that
[
A  zI b ]
c
T
d
is singular. Then there exists a nonzero solution to
or
(A  zl)x + by = 0,
c
T
x +dy = O.
(12.7)
(12.8)
Assuming z is not an eigenvalue of A (i.e., no pole/zero cancellations), then from (12.7) we
get
x = (A  zl)lby.
(12.9)
Substituting this in (12.8), we have
_c
T
(A  zl)lby + dy = 0,
or g(z)y = 0 by the definition of g. Now y 1= 0 (else x = 0 from (12.9». Hence g(z) = 0,
i.e., z is a zero of g.
12.4 Symmetric Generalized Eigenvalue Problems
A very important special case of the generalized eigenvalue problem
Ax = ABx (12.10)
for A, B E ffi.nxn arises when A = AT and B = BT > O. For example, the secondorder
system of differential equations
Mx+Kx=O,
where M is a symmetric positive definite "mass matrix" and K is a symmetric "stiffness
matrix," is a frequently employed model of structures or vibrating systems and yields a
generalized eigenvalue problem ofthe form (12.10).
Since B is positive definite it is nonsingular. Thus, the problem (12.10) is equivalent
to the standard eigenvalue problem B
1
Ax = AX. However, B
1
A is not necessarily
symmetric.
Nevertheless, the eigenvalues of B
l
A are always real (and are approximately 2.1926
and 3.1926 in Example 12.16).
Theorem 12.17. Let A, B e R
nxn
with A = A
T
and B = B
T
> 0. Then the generalized
eigenvalue problem
whose eigenvalues are approximately 2.1926 and —3.1926 as expected.
The material of this section can, of course, be generalized easily to the case where A
and B are Hermitian, but since realvalued matrices are commonly used in most applications,
we have restricted our attention to that case only.
has n real eigenvalues, and the n corresponding right eigenvectors can be chosen to be
orthogonal with respect to the inner product (x, y)
B
= X
T
By. Moreover, if A > 0, then
the eigenvalues are also all positive.
Proof: Since B > 0, it has a Cholesky factorization B = LL
T
, where L is nonsingular
(Theorem 10.23). Then the eigenvalue problem
can be rewritten as the equivalent problem
Letting C = L
1
AL
J
and z = L
1
x, (12.11) can then be rewritten as
Since C = C
T
, the eigenproblem (12.12) has n real eigenvalues, with corresponding eigen
vectors zi,..., z
n
satisfying
Then x, = L
T
zi, i € n, are eigenvectors of the original generalized eigenvalue problem
and satisfy
Finally, if A = A
T
> 0, then C = C
T
> 0, so the eigenvalues are positive. D
Example 12.18. The Cholesky factor for the matrix B in Example 12.16 is
Then it is easily checked thai
132 Chapter 12. Generalized Eigenvalue Problems
Example 12.16. Let A ThenB~
l
A
132 Chapter 12. Generalized Eigenvalue Problems
Example 12.16. Let A = ; l B = [i J Then A = J
Nevertheless, the eigenvalues of A are always real (and are approximately 2.1926
and 3.1926 in Example 12.16).
Theorem 12.17. Let A, B E jRnxn with A = AT and B = BT > O. Then the generalized
eigenvalue problem
Ax = ABx
has n real eigenvalues, and the n corresponding right eigenvectors can be chosen to be
orthogonal with respect to the inner product (x, y) B = x
T
By. Moreover, if A > 0, then
the eigenvalues are also all positive.
Proof: Since B > 0, it has a Cholesky factorization B = LL T, where L is nonsingular
(Theorem 10.23). Then the eigenvalue problem
Ax = ABx = ALL T x
can be rewritten as the equivalent problem
(12.11)
Letting C = L AL and Z = LT x, (12.11) can then be rewritten as
Cz = AZ. (12.12)
Since C = C
T
, the eigenproblem (12.12) has n real eigenvalues, with corresponding eigen
vectors Z I, •.. , Zn satisfying
zi Zj = Dij.
Then Xi = L Zi, i E !!., are eigenvectors of the original generalized eigenvalue problem
and satisfy
(Xi, Xj)B = xr BXj = (zi L Zj) = Dij.
Finally, if A = AT> 0, then C = C
T
> 0, so the eigenvalues are positive. 0
Example 12.18. The Cholesky factor for the matrix B in Example 12.16 is
1] .
.,fi .,fi
Then it is easily checked that
c = = [ 0 . .5
2.5
2 . .5 ]
1.5 '
whose eigenvalues are approximately 2.1926 and 3.1926 as expected.
The material of this section can, of course, be generalized easily to the case where A
and B are Hermitian, but since realvalued matrices are commonly used in most applications,
we have restricted our attention to that case only.
12.5. Simultaneous Diagonalization 133
12.5 Simultaneous Diagonalization
Recall that many matrices can be diagonalized by a similarity. In particular, normal ma
trices can be diagonalized by a unitary similarity. It turns out that in some cases a pair of
matrices (A, B) can be simultaneously diagonalized by the same matrix. There are many
such results and we present only a representative (but important and useful) theorem here.
Again, we restrict our attention only to the real case, with the complex case following in a
straightforward way.
Theorem 12.19 (Simultaneous Reduction to Diagonal Form). Let A, B e E"
x
" with
A = A
T
and B = B
T
> 0. Then there exists a nonsingular matrix Q such that
\ 2.5.1 Simultaneous diagonalization via SVD
There are situations in which forming C = L~
1
AL~
T
as in the proof of Theorem 12.19 is
numerically problematic, e.g., when L is highly ill conditioned with respect to inversion. In
such cases, simultaneous reduction can also be accomplished via an SVD. To illustrate, let
where D is diagonal. In fact, the diagonal elements of D are the eigenvalues of B
1
A.
Proof: Let B = LL
T
be the Cholesky factorization of B and setC = L~
1
AL~
T
. Since
C is symmetric, there exists an orthogonal matrix P such that P
T
CP = D, where D is
diagonal. Let Q = L~
T
P. Then
and
Finally, since QDQ~
l
= QQ
T
AQQ~
l
= L
T
PP
T
L~
1
A = L~
T
L~
1
A = B~
1
A, we
haveA(D) = A(B~
1
A). D
Note that Q is not in general orthogonal, so it does not preserve eigenvalues of A and B
individually. However, it does preserve the eigenvalues of A — XB. This can be seen directly.
LetA = Q
T
AQandB = Q
T
BQ. Then/HA = Q~
l
B~
l
Q~
T
Q
T
AQ = Q~
1
B~
1
AQ.
Theorem 12.19 is very useful for reducing many statements about pairs of symmetric
matrices to "the diagonal case." The following is typical.
Theorem 12.20. Let A, B e M"
xn
be positive definite. Then A > B if and only if B~
l
>
A
1
.
Proof: By Theorem 12.19, there exists Q e E"
x
" such that Q
T
AQ = D and Q
T
BQ = I,
where D is diagonal. Now D > 0 by Theorem 10.31. Also, since A > B, by Theorem
10.21 we have that Q
T
AQ > Q
T
BQ, i.e., D > I. But then D"
1
< / (this is trivially true
since the two matrices are diagonal). Thus, QD~
l
Q
T
< QQ
T
, i.e., A~
l
< B~
l
. D
12.5. Simultaneous Diagonalization 133
12.5 Simultaneous Diagonalization
Recall that many matrices can be diagonalized by a similarity. In particular, normal ma
trices can be diagonalized by a unitary similarity. It turns out that in some cases a pair of
matrices (A, B) can be simultaneously diagonalized by the same matrix. There are many
such results and we present only a representative (but important and useful) theorem here.
Again, we restrict our attention only to the real case, with the complex case following in a
straightforward way.
Theorem 12.19 (Simultaneous Reduction to Diagonal Form). Let A, B E ] [ ~ n x n with
A = AT and B = BT > O. Then there exists a nonsingular matrix Q such that
where D is diagonal. Infact, the diagonal elements of D are the eigenvalues of B
1
A.
Proof: Let B = LLT be the Cholesky factorization of B and set C = L I AL T. Since
C is symmetric, there exists an orthogonal matrix P such that pTe p = D, where D is
diagonal. Let Q = L  T P. Then
and
QT BQ = pT L I(LLT)L T P = pT P = [.
Finally, since QDQI = QQT AQQI = L T P pT L I A = L T L I A
B
1
A, we
have A(D) = A(B
1
A). 0
Note that Q is not in general orthogonal, so it does not preserve eigenvalues of A and B
individually. However, it does preserve the eigenvalues of A  'AB. This can be seen directly.
Let A = QT AQ and B = QT BQ. Then B
1
A = Q1 B
1
QT QT AQ = Q1 B
1
AQ.
Theorem 12.19 is very useful for reducing many statements about pairs of symmetric
matrices to "the diagonal case." The following is typical.
Theorem 12.20. Let A, B E lR
nxn
be positive definite. Then A 2: B if and only if B
1
2:
AI.
Proof: By Theorem 12.19, there exists Q E l R ~ x n such that QT AQ = D and QT BQ = [,
where D is diagonal. Now D > 0 by Theorem 10.31. Also, since A 2: B, by Theorem
10.21 we have that QT AQ 2: QT BQ, i.e., D 2: [. But then D
I
:::: [(this is trivially true
since the two matrices are diagonal). Thus, Q D
I
QT :::: Q QT, i.e., A I :::: B
1
. 0
12.5.1 Simultaneous diagonalization via SVD
There are situations in which forming C = L I AL T as in the proof of Theorem 12.19 is
numerically problematic, e.g., when L is highly iII conditioned with respect to inversion. In
such cases, simultaneous reduction can also be accomplished via an SVD. To illustrate. let
The problem (12.15) is called a generalized singular value problem and algorithms exist to
solve it (and hence equivalently (12.13)) via arithmetic operations performed only on LA
and LB separately, i.e., without forming the products L
A
L
T
A
or L
B
L
T
B
explicitly; see, for
example, [7, Sec. 8.7.3]. This is analogous to finding the singular values of a matrix M by
operations performed directly on M rather than by forming the matrix M
T
M and solving
the eigenproblem M
T
MX = Xx.
Remark 12.22. Various generalizations of the results in Remark 12.21 are possible, for
example, when A = A
T
> 0. The case when A is symmetric but indefinite is not so
straightforward, at least in real arithmetic. For example, A can be written as A = PDP
T
,
~ ~ ~ ~ T
where D is diagonal and P is orthogonal, but in writing A — PDDP = PD(PD) with
D diagonal, D may have pure imaginary elements.
134 Chapter 12. Generalized Eigenvalue Problems
us assume that both A and B are positive definite. Further, let A = L
A
L
T
A
and B — LsL
T
B
be Cholesky factorizations of A and B, respectively. Compute the SVD
where E e R£
x
" is diagonal. Then the matrix Q = L
B
T
U performs the simultaneous
diagonalization. To check this, note that
while
Remark 12.21. The SVD in (12.13) can be computed without explicitly forming the
indicated matrix product or the inverse by using the socalled generalized singular value
decomposition (GSVD). Note that
and thus the singular values of L
B
L
A
can be found from the eigenvalue problem
Letting x = L
B
z we see that (12.14) can be rewritten in the form L
A
L
A
x = XL
B
z =
ALgL^Lg
7
z, which is thus equivalent to the generalized eigenvalue problem
134 Chapter 12. Generalized Eigenvalue Problems
us assume that both A and B are positive definite. Further, let A = and B =
be Cholesky factorizations of A and B, respectively. Compute the SVD
(12.13)
where L E xn is diagonal. Then the matrix Q = L i/ u performs the simultaneous
diagonalization. To check this, note that
while
QT AQ = U
T
= UTULVTVLTUTU
= L2
QT BQ = U
T
= UTU
= I.
Remark 12.21. The SVD in (12.13) can be computed without explicitly forming the
indicated matrix product or the inverse by using the socalled generalized singular value
decomposition (GSVD). Note that
and thus the singular values of L B 1 L A can be found from the eigenvalue problem
02.14)
Letting x = LBT Z we see that 02.14) can be rewritten in the form = ALBz =
z, which is thus equivalent to the generalized eigenvalue problem
02.15)
The problem (12.15) is called a generalized singular value problem and algorithms exist to
solve it (and hence equivalently (12.13» via arithmetic operations performed only on LA
and L B separately, i.e., without forming the products LA L or L B L explicitly; see, for
example, [7, Sec. 8.7.3]. This is analogous to finding the singular values of a matrix M by
operations performed directly on M rather than by forming the matrix MT M and solving
the eigenproblem MT M x = AX.
Remark 12.22. Various generalizations of the results in Remark 12.21 are possible, for
example, when A = AT::: O. The case when A is symmetric but indefinite is not so
straightforward, at least in real arithmetic. For example, A can be written as A = PDP T,
where Disdiagonaland P is orthogonal,butin writing A = PDDp
T
= PD(PD{ with
D diagonal, b may have pure imaginary elements.
12.6. HigherOrder Eigenvalue Problems 135
12.6 HigherOrder Eigenvalue Problems
Consider the secondorder system of differential equations
where q(t} e W
1
and M, C, K e Rn
xn
. Assume for simplicity that M is nonsingular.
Suppose, by analogy with the firstorder case, that we try to find a solution of (12.16) of the
form q(t) = e
xt
p, where the nvector p and scalar A. are to be determined. Substituting in
(12.16) we get
To get a nonzero solution /?, we thus seek values of A. for which the matrix A.
2
M + A.C + K
is singular. Since the determinantal equation
yields a polynomial of degree 2rc, there are 2n eigenvalues for the secondorder (or
quadratic) eigenvalue problem A.
2
M + A.C + K.
A special case of (12.16) arises frequently in applications: M = I, C = 0, and
K = K
T
. Suppose K has eigenvalues
Let a > k =  f j i k 1
2
• Then the 2n eigenvalues of the secondorder eigenvalue problem A.
2
/ + K
are
If r = n (i.e., K = K
T
> 0), then all solutions of q + Kq = 0 are oscillatory.
12.6.1 Conversion to firstorder form
Let x\ = q and \i = q. Then (12.16) can be written as a firstorder system (with block
companion matrix)
where x(t) €. E
2
". If M is singular, or if it is desired to avoid the calculation of M
l
because
M is too ill conditioned with respect to inversion, the secondorder problem (12.16) can still
be converted to the firstorder generalized linear system
or, since
12.6. HigherOrder Eigenvalue Problems 135
12.6 HigherOrder Eigenvalue Problems
Consider the secondorder system of differential equations
Mq+Cq+Kq=O, (12.16)
where q(t) E ~ n and M, C, K E ~ n x n . Assume for simplicity that M is nonsingular.
Suppose, by analogy with the firstorder case, that we try to find a solution of (12.16) of the
form q(t) = eAt p, where the nvector p and scalar A are to be determined. Substituting in
(12.16) we get
or, since eAt :F 0,
(A
2
M + AC + K) p = O.
To get a nonzero solution p, we thus seek values of A for which the matrix A
2
M + AC + K
is singular. Since the determinantal equation
o = det(A
2
M + AC + K) = A 2n + ...
yields a polynomial of degree 2n, there are 2n eigenvalues for the secondorder (or
quadratic) eigenvalue problem A
2
M + AC + K.
A special case of (12.16) arises frequently in applications: M = I, C = 0, and
K = KT. Suppose K has eigenvalues
IL I ::: ... ::: ILr ::: 0 > ILr+ I ::: ... ::: ILn·
Let Wk = I ILk I !. Then the 2n eigenvalues of the secondorder eigenvalue problem A
2
I + K
are
± jWk; k = 1, ... , r,
± Wk; k = r + 1, ... , n.
If r = n (i.e., K = KT ::: 0), then all solutions of q + K q = 0 are oscillatory.
12.6.1 Conversion to firstorder form
Let XI = q and X2 = q. Then (12.16) can be written as a firstorder system (with block
companion matrix)
. [ 0
X = M1K
where x (t) E ~ 2 n . If M is singular, or if it is desired to avoid the calculation of M
I
because
M is too ill conditioned with respect to inversion, the secondorder problem (12.16) can still
be converted to the firstorder generalized linear system
[
I OJ' [0 I J
o M x = K C x.
136 Chapter 12. Generalized Eigenvalue Problems
Many other firstorder realizations are possible. Some can be useful when M, C, and/or K
have special symmetry or skewsymmetry properties that can exploited.
Higherorder analogues of (12.16) involving, say, the kth derivative of q, lead naturally
to higherorder eigenvalue problems that can be converted to firstorder form using aknxkn
block companion matrix analogue of (11.19). Similar procedures hold for the general k\h
order difference equation
EXERCISES
are the eigenvalues of the matrix A — BD
1
C.
2. Let F, G € C
MX
". Show that the nonzero eigenvalues of FG and GF are the same.
Hint: An easy "trick proof is to verify that the matrices
are similar via the similarity transformation
are identical for all F 6 E"
1
*" and all G G R"
xm
.
Hint: Consider the equivalence
(A similar result is also true for "nonsquare" pencils. In the parlance of control theory,
such results show that zeros are invariant under state feedback or output injection.)
which can be converted to various firstorder systems of dimension kn.
1. Suppose A e R
nx
" and D e R™
xm
. Show that the finite generalized eigenvalues of
the pencil
3. Let F e C
nxm
, G e C
mx
". Are the nonzero singular values of FG and GF the
same?
4. Suppose A € R
nxn
, B e R
n
*
m
, and C e E
wx
". Show that the generalized eigenval
ues of the pencils
and
136 Chapter 12. Generalized Eigenvalue Problems
Many other firstorder realizations are possible. Some can be useful when M, C, andlor K
have special symmetry or skewsymmetry properties that can exploited.
Higherorder analogues of (12.16) involving, say, the kth derivative of q, lead naturally
to higherorder eigenvalue problems that can be converted to firstorder form using a kn x kn
block companion matrix analogue of (11.19). Similar procedures hold for the general kth
order difference equation
which can be converted to various firstorder systems of dimension kn.
EXERCISES
1. Suppose A E lR
n
xn and D E lR::! xm. Show that the finite generalized eigenvalues of
the pencil
[ ~ ~ J  A [ ~ ~ J
are the eigenvalues of the matrix A  B D
1
C.
2. Let F, G E e
nxn
• Show that the nonzero eigenvalues of FG and G F are the same.
Hint: An easy "trick proof' is to verify that the matrices
[Fg ~ ] and [ ~ GOF ]
are similar via the similarity transformation
3. Let F E e
nxm
, G E e
mxn
• Are the nonzero singular values of FG and GF the
same?
4. Suppose A E ]Rnxn, B E lR
nxm
, and C E lRmxn. Show that the generalized eigenval
ues of the pencils
[ ~ ~ J  A [ ~ ~ J
and
[ A + B ~ + GC ~ ] _ A [ ~ ~ ]
are identical for all F E Rm xn and all G E R" xm .
Hint: Consider the equivalence
[
I G][AU B][I 0]
01 CO Fl'
(A similar result is also true for "nonsquare" pencils. In the parlance of control theory,
such results show that zeros are invariant under state feedback or output injection.)
Exercises 137
5. Another family of simultaneous diagonalization problems arises when it is desired
that the simultaneous diagonalizing transformation Q operates on matrices A, B e
]R
nx
" in such a way that Q~
l
AQ~
T
and Q
T
BQ are simultaneously diagonal. Such
a transformation is called contragredient. Consider the case where both A and
B are positive definite with Cholesky factorizations A = L&L
T
A
and B = L#Lg,
respectively, and let UW
T
be an SVD of L
T
B
L
A
.
(a) Show that Q = LA V£~
5
is a contragredient transformation that reduces both
A and B to the same diagonal matrix.
(b) Show that Q~
l
= ^~^U
T
L
T
B
.
(c) Show that the eigenvalues of A B are the same as those of E
2
and hence are
positive.
Exercises 137
5. Another family of simultaneous diagonalization problems arises when it is desired
that the simultaneous diagonalizing transformation Q operates on matrices A, B E
jRnxn in such a way that Ql AQT and QT BQ are simultaneously diagonal. Such
a transformation is called contragredient. Consider the case where both A and
B are positive definite with Cholesky factorizations A = LA L and B = L B L
respectively, and let be an SVD of
(a) Show that Q = LA is a contragredient transformation that reduces both
A and B to the same diagonal matrix.
(b) Show that Ql =
(c) Show that the eigenvalues of AB are the same as those of 1;2 and hence are
positive.
This page intentionally left blank This page intentionally left blank
Chapter 13
Kronecker Products
13.1 Definition and Examples
Definition 13.1. Let A e R
mx
", B e R
pxq
. Then the Kronecker product (or tensor
product) of A and B is defined as the matrix
Obviously, the same definition holds if A and B are complexvalued matrices. We
restrict our attention in this chapter primarily to realvalued matrices, pointing out the
extension to the complex case only where it is not obvious.
Example 13.2.
Note that B < g> A / A < g> B.
2. Foranyfl e!F
X(
7, /
2
< 8 > f l = [o l\
Replacing I
2
by /„ yields a block diagonal matrix with n copies of B along the
diagonal.
3. Let B be an arbitrary 2x2 matrix. Then
139
Chapter 13
Kronecker Products
13.1 Definition and Examples
Definition 13.1. Let A E lR
mxn
, B E lR
pxq
. Then the Kronecker product (or tensor
product) of A and B is defined as the matrix
[
allB
A@B= :
amlB
alnB ]
: E lRmpxnq.
amnB
(13.1)
Obviously, the same definition holds if A and B are complexvalued matrices. We
restrict our attention in this chapter primarily to realvalued matrices, pointing out the
extension to the complex case only where it is not obvious.
Example 13.2.
1. Let A =
2
nand B = [; Then
2
4 2 6
n
A@B = [
2B 3 4 6 6
2B 3 4 2 2
9 4 6 2
Note that B @ A i A @ B.
2. Forany B E lR
pxq
, /z @ B = J.
Replacing 12 by In yields a block diagonal matrix with n copies of B along the
diagonal.
3. Let B be an arbitrary 2 x 2 matrix. Then
l b"
0
b12
0
l
B @/z =
b
ll
0 b12
0
b
2
2
0
b
21
0 b
22
139
140 Chapter 13. Kronecker Products
The extension to arbitrary B and /„ is obvious.
4. Let Jt € R
m
, y e R". Then
5. Let* eR
m
, y eR". Then
13.2 Properties of the Kronecker Product
Theorem 13.3. Let A e R
mx
", 5 e R
rxi
, C e R"
x
^ and D e R
sxt
. Then
Proof: Simply verify that
Theorem 13.4. For all A and B,
Proof: For the proof, simply verify using the definitions of transpose and Kronecker
product. D
Corollary 13.5. If A e R"
xn
and B e R
mxm
are symmetric, then A® B is symmetric.
Theorem 13.6. If A and B are nonsingular,
Proof: Using Theorem 13.3, simply note that
140 Chapter 13. Kronecker Products
The extension to arbitrary B and In is obvious.
4. Let x E y E !R.n. Then
[
T T]T
X ® Y = XIY , ... , XmY
= [XIYJ, ... , XIYn, X2Yl, ... , xmYnf E !R.
mn
.
13.2 Properties of the Kronecker Product
Theorem 13.3. Let A E B E C E and D E Then
(A 0 B)(C 0 D) = AC 0 BD (E
Proof; Simply verify that
=AC0BD. 0
Theorem 13.4. Foral! A and B, (A ® Bl = AT ® BT.
al;kCkPBD ]
amkckpBD
(13.2)
Proof' For the proof, simply verify using the definitions of transpose and Kronecker
product. 0
Corollary 13.5. If A E ]Rn xn and B E !R.
m
xm are symmetric, then A ® B is symmetric.
Theorem 13.6. If A and Bare nonsingular, (A ® B)I = AI ® B
1
.
Proof: Using Theorem 13.3, simply note that (A ® B)(A 1 ® B
1
) = 1 ® 1 = I. 0
Corollary 13.8. If A € E"
xn
is orthogonal and B e M
mxm
15 orthogonal, then A < g > B is
orthogonal.
13.2. Properties of the Kronecker Product 141
Theorem 13.7. If A e IR"
xn
am/ B eR
mxm
are normal, then A® B is normal.
Proof:
yields a singular value decomposition of A < 8 > B (after a simple reordering of the diagonal
elements O/£A < 8 > £5 and the corresponding right and left singular vectors).
Corollary 13.11. Let A e R™
x
" have singular values a\ > • • • > a
r
> 0 and let B e
have singular values T\ > • • • > T
S
> 0. Then A < g ) B (or B < 8 > A) has rs singular values
^iT\ > • • • > ff
r
T
s
> Qand
Theorem 13.12. Let A e R
nx
" have eigenvalues A.,  , / e n, and let B e R
mxw
/zave
eigenvalues jJij, 7 € m. TTzen ?/ze mn eigenvalues of A® B are
Moreover, if x\, ..., x
p
are linearly independent right eigenvectors of A corresponding
to A  i , . . . , A.
p
(p < n), and zi, • • •, z
q
are linearly independent right eigenvectors of B
corresponding to JJL\ , ..., \Ju
q
(q < m), then ;c, < 8 > Zj € ffi.
m
" are linearly independent right
eigenvectors of A® B corresponding to A., /u ,
7
, i e /?, 7 e q.
Proof: The basic idea of the proof is as follows:
If A and B are diag onalizable in Theorem 13.12, we can take p = n and q —mand
thu s g et the complete eig enstru ctu re of A < 8 > B. In g eneral, if A and fi have Jordan form
Example 13.9. Let A and B  Then it is easily seen that
A i s orthog onal wi th eig envalu es e
±j9
and B i s orthog onal wi th eig envalu es e
±j(i>
. T he 4x4
matrix A ® 5 is then also orthog onal with eig envalu es e^'^+'W and e
±
^
( 6>
~^
>
\
Theorem 13.10. Lg f A G E
mx
" have a singular value decomposition l/^E^Vj an^ /ef
fi e ^
pxq
have a singular value decomposition UB^B^B Then
13.2. Properties of the Kronecker Product
Theorem 13.7. If A E IR
nxn
and B E IR
mxm
are normal, then A 0 B is normal.
Proof:
(A 0 B{ (A 0 B) = (AT 0 BT)(A 0 B) by Theorem 13.4
= AT A 0 BT B by Theorem 13.3
= AAT 0 B BT since A and B are normal
= (A 0 B)(A 0 B)T by Theorem 13.3. 0
141
Corollary 13.8. If A E IR
nxn
is orthogonal and B E IR
mxm
is orthogonal, then A 0 B is
orthogonal.
E I 139 L A
[
eose Sine] dB [Cos</> Sin</>] Th ., '1 h
xamp e .• et = _ sin e cose an = _ sin</> cos</>O en It IS easl y seen t at
A is orthogonal with eigenvalues e±jO and B is orthogonal with eigenvalues e±j</J. The 4 x 4
matrix A 0 B is then also orthogonal with eigenvalues e±jeH</» and e±je
fJ
</».
Theorem 13.10. Let A E IR
mxn
have a singular value decomposition VA ~ A vI and let
B E IR
pxq
have a singular value decomposition V B ~ B VI. Then
yields a singular value decomposition of A 0 B (after a simple reordering of the diagonal
elements of ~ A 0 ~ B and the corresponding right and left singular vectors).
Corollary 13.11. Let A E lR;"xn have singular values UI :::: ... :::: U
r
> 0 and let B E IRfx
q
have singular values <I :::: ... :::: <s > O. Then A 0 B (or B 0 A) has rs singular values
U, <I :::: ... :::: U
r
<s > 0 and
rank(A 0 B) = (rankA)(rankB) = rank(B 0 A) .
Theorem 13.12. Let A E IR
n
xn have eigenvalues Ai, i E !!, and let B E IR
m
xm have
eigenvalues JL j, j E m. Then the mn eigenvalues of A 0 Bare
Moreover, if Xl, ••. , xp are linearly independent right eigenvectors of A corresponding
to AI, ... , A p (p ::::: n), and Z I, ... ,Zq are linearly independent right eigenvectors of B
corresponding to JLI, ... ,JLq (q ::::: m), then Xi 0 Zj E IR
mn
are linearly independent right
eigenvectors of A 0 B corresponding to Ai JL j, i E l!! j E 1·
Proof: The basic idea of the proof is as follows:
(A 0 B)(x 0 z) = Ax 0 Bz
= AX 0 JLZ
= AJL(X 0 z). 0
If A and Bare diagonalizable in Theorem 13.12, we can take p = nand q = m and
thus get the complete eigenstructure of A 0 B. In general, if A and B have Jordan form
142 Chapter 13. Kronecker Products
decompositions given by P~
l
AP = JA and Q~
]
BQ = JB, respectively, then we get the
following Jordanlike structure:
Note that JA® JB, while upper triangular, is generally not quite in Jordan form and needs
further reduction (to an ultimate Jordan form that also depends on whether or not certain
eigenvalues are zero or nonzero).
A Schur form for A ® B can be derived similarly. For example, suppose P and
Q are unitary matrices that reduce A and 5, respectively, to Schur (triangular) form, i.e.,
P
H
AP = T
A
and Q
H
BQ = T
B
(and similarly if P and Q are orthogonal similarities
reducing A and B to real Schur form). Then
Corollary 13.13. Let A e R
nxn
and B e R
mxm
. Then
Definition 13.14. Let A e R
nxn
and B e R
mxm
. Then the Kronecker sum (or tensor sum)
of A and B, denoted A © B, is the mn x mn matrix (I
m
< g> A) + (B ® /„). Note that, in
general, A ® B ^ B © A.
Example 13.15.
Then
The reader is invited to compute B 0 A = (/3 ® B) + (A < g> /2) and note the difference
with A © B.
1. Let
142 Chapter 1 3. Kronecker Products
decompositions given by p
I
AP = J
A
and Ql BQ = J
B
, respectively, then we get the
following Jordanlike structure:
(P ® Q)I(A ® B)(P ® Q) = (P
I
® Ql)(A ® B)(P ® Q)
= (P
1
AP) ® (Ql BQ)
= J
A
® J
B ·
Note that h ® JR, while upper triangular, is generally not quite in Jordan form and needs
further reduction (to an ultimate Jordan form that also depends on whether or not certain
eigenvalues are zero or nonzero).
A Schur form for A ® B can be derived similarly. For example, suppose P and
Q are unitary matrices that reduce A and B, respectively, to Schur (triangular) form, i.e.,
pH AP = TA and QH BQ = TB (and similarly if P and Q are orthogonal similarities
reducing A and B to real Schur form). Then
(P ® Q)H (A ® B)(P ® Q) = (pH ® QH)(A ® B)(P ® Q)
= (pH AP) ® (QH BQ)
= TA ® T
R
.
Corollary 13.13. Let A E IR
n
xn and B E IR
rn
xm. Then
1. Tr(A ® B) = (TrA)(TrB) = Tr(B ® A).
2. det(A ® B) = (det A)m(det Bt = det(B ® A).
Definition 13.14. Let A E IR
n
Xn and B E IR
m
xrn. Then the Kronecker sum (or tensor sum)
of A and B, denoted A EEl B, is the mn x mn matrix Urn ® A) + (B ® In). Note that, in
general, A EEl B i= B EEl A.
Example 13.15.
1. Let
2
;
2
Then
2 3 0 0 0 2 0 0 0 0
3 2 1 0 0 0 0 2 0 0 1 0
AfflB = (h®A)+(B®h) =
1 1 4 0 0 0 0 0 2 0 0
0 0 0 2 3
+
2 0 0 3 0 0
0 0 0 3 2 0 2 0 0 3 0
0 0 0 4 0 0 2 0 0 3
The reader is invited to compute B EEl A = (h ® B) + (A 0 h) and note the difference
with A EEl B.
13.2. Properties of the Kronecker Product 143
If A and B are diagonalizable in Theorem 13.16, we can take p = n and q = m and
thus get the complete eigenstructure of A 0 B. In general, if A and B have Jordan form
decompositions given by P~
1
AP = JA and Q"
1
BQ = JB, respectively, then
is a Jordanlike structure for A © B.
Then J can be written in the very compact form J = (4 < 8 > M) + (E^®l2) = M 0 Ek.
Theorem 13.16. Let A e E"
x
" have eigenvalues A,  , i e n, and let B e R
mx
'" have
eigenvalues /z
;
, 7 e ra. TTzen r/ze Kronecker sum A® B = (I
m
(g> A) + (B < g> /„) /za^ ran
e/genva/wes
Moreover, if x\,... ,x
p
are linearly independent right eigenvectors of A corresponding
to AI, . . . , X
p
(p < n), and z\, ..., z
q
are linearly independent right eigenvectors of B
corresponding to f j i \ , . . . , f^
q
(q < ra), then Zj < 8 > Xi € W
1
" are linearly independent right
eigenvectors of A® B corresponding to A., + [ij , i € p, j e q.
Proof: The basic idea of the proof is as follows:
2. Recall the real JCF
where M =
13.2. Properties of the Kronecker Product
2. Recall the real JCF
1=
where M = [
a
f3
M I 0 0
f3
a
o M I 0
M
0
J. Define
0 0
0 0
Ek =
0
I 0
M I
o M
o
o
o
143
E jR2kx2k,
Then 1 can be written in the very compact form 1 = (I} ® M) + (Ek ® h) = M $ E
k
.
Theorem 13.16. Let A E jRnxn have eigenvalues Ai, i E !!. and let B E jRmxm have
eigenvalues fJj, j E I!!. Then the Kronecker sum A $ B = (1m ® A) + (B ® In) has mn
eigenvalues
Al + fJt, ... , AI + fJm, A2 + fJt,···, A2 + fJm, ... , An + fJm'
Moreover, if XI, .•• , xp are linearly independent right eigenvectors of A corresponding
to AI, ... , Ap (p ::s: n), and ZI, ... , Zq are linearly independent right eigenvectors of B
corresponding to fJt, ... , fJq (q ::s: m), then Z j ® Xi E jRmn are linearly independent right
eigenvectors of A $ B corresponding to Ai + fJj' i E E, j E fl·
Proof: The basic idea of the proof is as follows:
[(1m ® A) + (B ® In)](Z ® X) = (Z ® Ax) + (Bz ® X)
= (Z ® Ax) + (fJZ ® X)
= (A + fJ)(Z ® X). 0
If A and Bare diagonalizable in Theorem 13.16, we can take p = nand q = m and
thus get the complete eigenstructure of A $ B. In general, if A and B have Jordan form
decompositions given by pI AP = lA and Qt BQ = l
B
, respectively, then
[(Q ® In)(lm ® p)rt[(lm ® A) + (B ® In)][CQ ® In)(lm ® P)]
= [(1m ® p)I(Q ® In)I][(lm ® A) + (B ® In)][(Q ® In)(/m ® P)]
= [(1m ® pI)(QI ® In)][(lm ® A) + (B ® In)][CQ ® In)(/m <:9 P)]
= (1m ® lA) + (JB ® In)
is a Jordanlike structure for A $ B.
144 Chapter 13. Kronecker Products
A Schur form for A © B can be derived similarly. Again, suppose P and Q are unitary
matrices that reduce A and B, respectively, to Schur (triangular) form, i.e., P
H
AP = T
A
and Q
H
BQ = T
B
(and similarly if P and Q are orthogonal similarities reducing A and B
to real Schur form). Then
((Q ® /„)(/« ® P)]"[(/m < 8 > A) + (B ® /
B
)][(e (g) /„)(/„, ® P)] = (/
m
< 8 > r
A
) + (7* (g) /„),
where [(Q < 8 > /„)(/« ® P)] = (< 2 ® P) is unitary by Theorem 13.3 and Corollary 13.8 .
13.3 Application to Sylvester and Lyapunov Equations
In this section we study the linear matrix equation
A special case of (13.3) is the symmetric equation
obtained by taking B = A
T
. When C is symmetric, the solution X e W
x
" is easily shown
also to be symmetric and (13.4) is known as a Lyapunov equation. Lyapunov equations
arise naturally in stability theory.
The first important question to ask regarding (13.3) is, When does a solution exist?
By writing the matrices in (13.3) in terms of their columns, it is easily seen by equating the
z'th columns that
The coefficient matrix in (13.5) clearly can be written as the Kronecker sum (I
m
* A) +
(B
T
® /„). The following definition is very helpful in completing the writing of (13.5) as
an "ordinary" linear system.
where A e R"
x
", B e R
mxm
, and C e M"
xm
. This equation is now often called a Sylvester
equation in honor of J.J. Sylvester who studied general linear matrix equations of the form
These equations can then be rewritten as the mn x mn linear system
144 Chapter 13. Kronecker Products
A Schur fonn for A EB B can be derived similarly. Again, suppose P and Q are unitary
matrices that reduce A and B, respectively, to Schur (triangular) fonn, i.e., pH AP = TA
and QH BQ = TB (and similarly if P and Q are orthogonal similarities reducing A and B
to real Schur fonn). Then
where [(Q ® In)(lm ® P)] = (Q ® P) is unitary by Theorem 13.3 and Corollary 13.8.
13.3 Application to Sylvester and Lyapunov Equations
In this section we study the linear matrix equation
AX+XB=C, (13.3)
where A E IR
nxn
, B E IR
mxm
, and C E IRnxm. This equation is now often called a Sylvester
equation in honor of 1.1. Sylvester who studied general linear matrix equations of the fonn
k
LA;XB; =C.
;=1
A special case of (13.3) is the symmetric equation
AX +XAT = C (13.4)
obtained by taking B = AT. When C is symmetric, the solution X E IR
n
xn is easily shown
also to be symmetric and (13.4) is known as a Lyapunov equation. Lyapunovequations
arise naturally in stability theory.
The first important question to ask regarding (13.3) is, When does a solution exist?
By writing the matrices in (13.3) in tenns of their columns, it is easily seen by equating the
ith columns that
m
AXi + Xb; = C; = AXi +
j=1
These equations can then be rewritten as the mn x mn linear system
[
A+blll b
21
1
bl21 A + b
2Z
1
blml b2ml
(13.5)
The coefficient matrix in (13.5) clearly can be written as the Kronecker sum (1m 0 A) +
(B
T
0 In). The following definition is very helpful in completing the writing of (13.5) as
an "ordinary" linear system.
13.3. Application to Sylvester and Lyapunov Equations 145
Definition 13.17. Let c
(
€ E.
n
denote the columns ofC e R
nxm
so that C = [n,..., c
m
}.
Then vec(C) is defined to be the mnvector formed by stacking the columns ofC on top of
one another, i.e., vec(C) =
Using Definition 13.17, the linear system (13.5) can be rewritten in the form
There exists a unique solution to (13.6) if and only if [(I
m
® A) + (B
T
® /„)] is nonsingular.
But [(I
m
< 8 > A) + (B
T
(g) /„)] is nonsingular if and only if it has no zero eigenvalues.
From Theorem 13.16, the eigenvalues of [(/
m
<g> A) + (B
T
<8> /„)] are A., + IJ LJ , where
A,, e A (A), i e n_, and ^j e A(fi), j e m. We thus have the following theorem.
Theorem 13.18. Let A e R
nxn
, B G R
mxm
, and C e R"
xm
. 77ie/i the Sylvester equation
has a unique solution if and only if A and —B have no eigenvalues in common.
Sylvester equations of the form (13.3) (or symmetric Lyapunov equations of the form
(13.4)) are generally not solved using the mn x mn "vec" formulation (13.6). The most
commonly preferred numerical algorithm is described in [2]. First A and B are reduced to
(real) Schur form. An equivalent linear system is then solved in which the triangular form
of the reduced A and B can be exploited to solve successively for the columns of a suitably
transformed solution matrix X. Assuming that, say, n > m, this algorithm takes only O(n
3
)
operations rather than the O(n
6
) that would be required by solving (13.6) directly with
Gaussian elimination. A further enhancement to this algorithm is available in [6] whereby
the larger of A or B is initially reduced only to upper Hessenberg rather than triangular
Schur form.
The next few theorems are classical. They culminate in Theorem 13.24, one of many
elegant connections between matrix theory and stability theory for differential equations.
Theorem 13.19. Let A e Rn
xn
, B e R
mxm
, and C e R
nxm
. Suppose further that A and B
are asymptotically stable (a matrix is asymptotically stable if all its eigenvalues have real
parts in the open left halfplane). Then the (unique) solution of the Sylvester equation
can be written as
Proof: Since A and B are stable, A., (A) + A
;
 (B) ^ 0 for all i, j so there exists a unique
solution to(13.8 )by Theorem 13.18. Now integrate the differential equation X = AX + XB
(with X(0) = C) on [0, +00):
13.3. Application to Sylvester and Lyapunov Equations 145
Definition 13.17. Let Ci E jRn denote the columns ofC E jRnxm so that C = [CI, ... , C
m
].
: : ~ ~ : : ~ : : : d ~ ~ : : : O :[]::::fonned by "ocking the colunuu of C on top of
Using Definition 13.17, the linear system (13.5) can be rewritten in the form
[(1m ® A) + (B
T
® In)]vec(X) = vec(C). (13.6)
There exists a unique solution to (13.6) if and only if [(1m ® A) + (B
T
® In)] is nonsingular.
But [(1m ® A) + (B
T
® In)] is nonsingular if and only if it has no zero eigenvalues.
From Theorem 13.16, the eigenvalues of [(1m ® A) + (BT ® In)] are Ai + Mj, where
Ai E A(A), i E!!, and Mj E A(B), j E!!!.. We thus have the following theorem.
Theorem 13.1S. Let A E lR
nxn
, B E jRmxm, and C E jRnxm. Then the Sylvester equation
AX+XB=C
(13.7)
has a unique solution if and only if A and  B have no eigenvalues in common.
Sylvester equations of the form (13.3) (or symmetric Lyapunov equations of the form
(13.4» are generally not solved using the mn x mn "vee" formulation (13.6). The most
commonly preferred numerical algorithm is described in [2]. First A and B are reduced to
(real) Schur form. An equivalent linear system is then solved in which the triangular form
of the reduced A and B can be exploited to solve successively for the columns of a suitably
transformed solution matrix X. Assuming that, say, n :::: m, this algorithm takes only 0 (n
3
)
operations rather than the O(n
6
) that would be required by solving (13.6) directly with
Gaussian elimination. A further enhancement to this algorithm is available in [6] whereby
the larger of A or B is initially reduced only to upper Hessenberg rather than triangular
Schur form.
The next few theorems are classical. They culminate in Theorem 13.24, one of many
elegant connections between matrix theory and stability theory for differential equations.
Theorem 13.19. Let A E jRnxn, B E jRmxm, and C E jRnxm. Suppose further that A and B
are asymptotically stable (a matrix is asymptotically stable if all its eigenvalues have real
parts in the open left halfplane). Then the (unique) solution of the Sylvester equation
AX+XB=C (13.8)
can be written as
(13.9)
Proof: Since A and B are stable, Aj(A) + Aj(B) =I 0 for all i, j so there exists a unique
solution to (13.8) by Theorem 13.18. Now integrate the differential equation X = AX + X B
(with X(O) = C) on [0, +00):
lim XU)  X(O) = A roo X(t)dt + ([+00 X(t)dt) B.
IHoo 10 10
(13.10)
146 Chapter 13. Kronecker Products
Using the results of Section 11.1.6, it can be shown easily that lim e = lim e = 0.
r—> + oo t—v+oo
Hence, using the solution X ( t ) = e
tA
Ce
tB
from Theorem 11.6, we have that lim X ( t ) — 0.
/—<+3C
Substituting in (13.10) we have
Remark 13.20. An equivalent condition for the existence of a unique solution to AX +
XB = C is that [ J _
c
fi
] be similar to [ J _°
B
] (via the similarity [ J _* ]).
Theorem 13.21. Lef A, C e R"
x
". TTzen r/z e Lyapunov equation
has a unique solution if and only if A and —A
T
have no eigenvalues in common. If C is
symmetric and (13.11) has a unique solution, then that solution is symmetric.
Remark 13.22. If the matrix A e W
xn
has eigenvalues A.I ,...,!„, then — A
T
has eigen
values —A.], . . . , —k
n
. Thus, a sufficient condition that guarantees that A and — A
T
have
no common eigenvalues is that A be asymptotically stable. Many useful results exist con
cerning the relationship between stability and Lyapunov equations. Two basic results due
to Lyapunov are the following, the first of which follows immediately from Theorem 13.19.
Theorem 13.23. Let A,C e R"
x
" and suppose further that A is asymptotically stable.
Then the (unique) solution of the Lyapunov equation
can be written as
Theorem 13.24. A matrix A e R"
x
" is asymptotically stable if and only if there exists a
positive definite solution to the Lyapunov equation
Proof: Suppose A is asymptotically stable. By Theorems 13.21 and 13.23 a solution to
(13.13) exists and takes the form (13.12). Now let v be an arbitrary nonz ero vector in E".
Then
and so X
where C 
146 Chapter 13. Kronecker Products
Using the results of Section 11.1.6, it can be shown easily that lim e
lA
= lim e
lB
= O.
1>+00 1 .... +00
Hence, using the solution X (t) = elACe
lB
from Theorem 11.6, we have that lim X (t) = O.
t ~ + x
Substituting in (13.10) we have
C = A (1+
00
elACe
lB
dt) + (1+
00
elACe
lB
dt) B
{+oo
and so X = 1o elACe
lB
dt satisfies (13.8). o
Remark 13.20. An equivalent condition for the existence of a unique solution to AX +
X B = C is that [ ~ _C
B
] be similar to [ ~ _OB] (via the similarity [ ~ _ ~ ]).
Theorem 13.21. Let A, C E jRnxn. Then the Lyapunov equation
AX+XAT = C (13.11)
has a unique solution if and only if A and  A T have no eigenvalues in common. If C is
symmetric and ( 13.11) has a unique solution, then that solution is symmetric.
Remark 13.22. If the matrix A E jRn xn has eigenvalues )"" ... , An, then  AT has eigen
values AI, ... ,  An. Thus, a sufficient condition that guarantees that A and  A T have
no common eigenvalues is that A be asymptotically stable. Many useful results exist con
cerning the relationship between stability and Lyapunov equations. Two basic results due
to Lyapunov are the following, the first of which follows immediately from Theorem 13.19.
Theorem 13.23. Let A, C E jRnxn and suppose further that A is asymptotically stable.
Then the (unique) solution o/the Lyapunov equation
AX+XAT=C
can be written as
(13.12)
Theorem 13.24. A matrix A E jRnxn is asymptotically stable if and only if there exists a
positive definite solution to the Lyapunov equation
AX +XAT = C, (13.13)
where C = C
T
< O.
Proof: Suppose A is asymptotically stable. By Theorems l3.21 and l3.23 a solution to
(13.13) exists and takes the form (13.12). Now let v be an arbitrary nonzero vector in jRn.
Then
13.3. Application to Sylvester and Lyapunov Equations 147
Since — C > 0 and e
tA
is nonsingular for all t, the integrand above is positive. Hence
v
T
Xv > 0 and thus X is positive definite.
Conversely, suppose X = X
T
> 0 and let A. e A (A) with corresponding left eigen
vector y. Then
Since y
H
Xy > 0, we must have A + A = 2 Re A < 0 . Since A was arbitrary, A must be
asymptotically stable. D
Remark 13.25. The Lyapunov equation AX + XA
T
= C can also be written using the
vec notation in the equivalent form
A subtle point arises when dealing with the "dual" Lyapunov equation A
T
X + XA = C.
The equivalent "vec form" of this equation is
However, the complexvalued equation A
H
X + XA = C is equivalent to
The vec operator has many useful properties, most of which derive from one key
result.
Theorem 13.26. For any three matrices A, B, and C for which the matrix product ABC is
defined,
Proof: The proof follows in a fairly straightforward fashion either directly from the defini
tions or from the fact that vec(;t;y
r
) = y <8 > x. D
An immediate application is to the derivation of existence and uniqueness conditions
for the solution of the simple Sylvesterlike equation introduced in Theorem 6.11.
Theorem 13.27. Let A e R
mxn
, B e R
px(}
, and C e R
mxq
. Then the equation
has a solution X e R.
nxp
if and only ifAA
+
CB
+
B = C, in which case the general solution
is of the form
where Y e R
nxp
is arbitrary. The solution of (13.14) is unique if BB
+
® A
+
A = I.
Proof: Write (13.14) as
13.3. Application to Sylvester and Lyapunov Equations 147
Since C > 0 and etA is nonsingular for all t, the integrand above is positive. Hence
v
T
Xv > 0 and thus X is positive definite.
Conversely, suppose X = XT > 0 and let A E A(A) with corresponding left eigen
vector y. Then
0> yHCy = yH AXy + yHXAT Y
= (A + I)yH Xy.
Since yH Xy > 0, we must have A + I = 2 Re A < O. Since A was arbitrary, A must be
asymptotically stable. D
Remark 13.25. The Lyapunov equation AX + X A T = C can also be written using the
vec notation in the equivalent form
[(/ ® A) + (A ® l)]vec(X) = vec(C).
A subtle point arises when dealing with the "dual" Lyapunov equation A T X + X A = C.
The equivalent "vec form" of this equation is
[(/ ® AT) + (AT ® l)]vec(X) = vec(C).
However, the complexvalued equation A H X + X A = C is equivalent to
[(/ ® AH) + (AT ® l)]vec(X) = vec(C).
The vec operator has many useful properties, most of which derive from one key
result.
Theorem 13.26. For any three matrices A, B, and C for which the matrix product ABC is
defined,
vec(ABC) = (C
T
® A)vec(B).
Proof: The proof follows in a fairly straightforward fashion either directly from the defini
tions or from the fact that vec(xyT) = y ® x. D
An immediate application is to the derivation of existence and uniqueness conditions
for the solution of the simple Sylvesterlike equation introduced in Theorem 6.11.
Theorem 13.27. Let A E jRrnxn, B E jRPxq, and C E jRrnxq. Then the equation
AXB =C (13.14)
has a solution X E jRn x p if and only if A A + C B+ B = C, in which case the general solution
is of the form
(13.15)
where Y E jRnxp is arbitrary. The solution of (13. 14) is unique if BB+ ® A+ A = [.
Proof: Write (13.14) as
(B
T
® A)vec(X) = vec(C) (13.16)
148 Chapter 13. Kronecker Products
by Theorem 13.26. This "vector equation" has a solution if and only if
It is a straightforward exercise to show that (M ® N)
+
= M
+
< 8> N
+
. Thus, (13.16) has a
solution if and only if
and hence if and only if AA
+
CB
+
B = C.
The general solution of (13.16) is then given by
where Y is arbitrary. This equation can then be rewritten in the form
or, using Theorem 13.26,
The solution is clearly unique if BB
+
< 8> A
+
A = I. D
EXERCISES
1. For any two matrices A and B for which the indicated matrix product is defined,
show that (vec(A))
r
(vec(fl)) = Tr(A
r
£). In particular, if B e Rn
xn
, then Tr(fl) =
vec(/J
r
vec(fl).
2. Prove that for all matrices A and B, (A ® B)
+
= A
+
® B
+
.
3. Show that the equation AX B = C has a solution for all C if A has full row rank and
B has full column rank. Also, show that a solution, if it exists, is unique if A has full
column rank and B has full row rank. What is the solution in this case?
4. Show that the general linear equation
can be written in the form
148 Chapter 1 3. Kronecker Products
by Theorem 13.26. This "vector equation" has a solution if and only if
(B
T
® A)(B
T
® A)+ vec(C) = vec(C).
It is a straightforward exercise to show that (M ® N) + = M+ ® N+. Thus, (13.16) has a
solution if and only if
vec(C) = (B
T
® A)«B+{ ® A+)vec(C)
= [(B+ B{ ® AA+]vec(C)
= vec(AA +C B+ B)
and hence if and only if AA + C B+ B = C.
The general solution of (13 .16) is then given by
vec(X) = (B
T
® A) + vec(C) + [I  (B
T
® A) + (B
T
® A)]vec(Y),
where Y is arbitrary. This equation can then be rewritten in the form
vec(X) = «B+{ ® A+)vec(C) + [I  (BB+{ ® A+ A]vec(y)
or, using Theorem 13.26,
The solution is clearly unique if B B+ ® A + A = I. 0
EXERCISES
I. For any two matrices A and B for which the indicated matrix product is defined,
show that (vec(A»T (vec(B» = Tr(A
T
B). In particular, if B E lR
nxn
, then Tr(B) =
vec(Inl vec(B).
2. Prove that for all matrices A and B, (A ® B)+ = A+ ® B+.
3. Show that the equation AX B = C has a solution for all C if A has full row rank and
B has full column rank. Also, show that a solution, if it exists, is unique if A has full
column rank and B has full row rank. What is the solution in this case?
4. Show that the general linear equation
k
LAiXB
i
=C
i=1
can be written in the form
[BT ® AI + ... + B[ ® Ak]vec(X) = vec(C).
Exercises 149
5. Let x € M
m
and y e E". Show that *
r
< 8 > y = yx
T
.
6. Let A e R"
xn
and £ e M
mxm
.
(a) Show that A < 8 > B
2
= A
2
£
2
.
(b) What is A ® B\\
F
in terms of the Frobenius norms of A and B? Justify your
answer carefully.
(c) What is the spectral radius of A < 8 > B in terms of the spectral radii of A and B?
Justify your answer carefully.
7. Let A, 5 eR"
x
".
(a) Show that (/ ® A)* = / < 8 > A* and (fl < g > /)* = B
fc
® / for all integ ers &.
(b) Show that e
l
®
A
= I < g ) e
A
and e
5
®
7
= e
B
(g ) /.
(c) Show that the matrices / (8 ) A and B ® / commute.
(d) Show that
(Note: This result would look a little "nicer" had we defined our Kronecker
sum the other way around. However, Definition 13.14 is conventional in the
literature.)
8 . Consider the Lyapunov matrix equation (13.11) with
and C the symmetric matrix
Clearly
is a symmetric solution of the equation. Verify that
is also a solution and is nonsymmetric. Explain in lig ht of Theorem 13.21.
9. Block Triangularization: Let
where A e Rn
xn
and D e R
mxm
. It is desired to find a similarity transformation
of the form
such that T
l
ST is block upper triang ular.
Exercises 149
5. Let x E ]Rm and y E ]Rn. Show that x T ® y = y X T •
(a) Show that IIA ® BII2 = IIAII2I1Blb.
(b) What is II A ® B II F in terms of the Frobenius norms of A and B? Justify your
answer carefully.
(c) What is the spectral radius of A ® B in terms of the spectral radii of A and B?
Justify your answer carefully.
7. Let A, B E ]Rnxn.
(a) Show that (l ® A)k = I ® Ak and (B ® Il = Bk ® I for all integers k.
(b) Show that el®A = I ® e
A
and eB®1 = e
B
® I.
(c) Show that the matrices I ® A and B ® I commute.
(d) Show that
e
AEIlB
= eU®A)+(B®l) = e
B
® e
A
.
(Note: This result would look a little "nicer" had we defined our Kronecker
sum the other way around. However, Definition 13.14 is conventional in the
literature.)
8. Consider the Lyapunov matrix equation (13.11) with
A = [ ~ _ ~ ]
and C the symmetric matrix
[ ~
Clearly
Xs = [ ~ ~ ]
is a symmetric solution of the equation. Verify that
Xns = [ _ ~ ~ ]
is also a solution and is nonsymmetric. Explain in light of Theorem 13.21.
9. Block Triangularization: Let
where A E ]Rn xn and D E ]Rm xm. It is desired to find a similarity transformation
of the form
T = [ ~ ~ J
such that T
1
ST is block upper triangular.
150 Chapter 13. Kronecker Products
(a) Show that S is similar to
if X satisfies the socalled matrix Riccati equation
(b) Formulate a similar result for block lower triangularization of S.
10. Block Diagonalization: Let
where A e Rn
xn
and D E R
mxm
. It is desired to find a similarity transformation of
the form
such that T
l
ST is block diagonal,
(a) Show that S is similar to
if Y satisfies the Sylvester equation
(b) Formulate a similar result for block diagonalization of
150 Chapter 13. Kronecker Products
(a) Show that S is similar to
[
A +OBX B ]
DXB
if X satisfies the socalled matrix Riccati equation
CXA+DXXBX=O.
(b) Fonnulate a similar result for block lower triangularization of S.
to. Block Diagonalization: Let
S= [ ~ ~ l
where A E jRnxn and D E jRmxm. It is desired to find a similarity transfonnation of
the fonn
T = [ ~ ~ ]
such that T
1
ST is block diagonal.
(a) Show that S is similar to
if Y satisfies the Sylvester equation
AY  YD = B.
(b) Fonnulate a similar result for block diagonalization of
Bibliography
[1] Albert, A., Regression and the MoorePenrose Pseudoinverse, Academic Press, New
York, NY, 1972.
[2] Bartels, R.H., and G.W. Stewart, "Algorithm 432. Solution of the Matrix Equation
AX + XB = C," Cornm. ACM, 15(1972), 820826.
[3] Bellman, R., Introduction to Matrix Analysis, Second Edition, McGrawHill, New
York, NY, 1970.
[4] Bjorck, A., Numerical Methods for Least Squares Problems, SIAM, Philadelphia, PA,
1996.
[5] Cline, R.E., "Note on the Generalized Inverse of the Product of Matrices," SIAM Rev.,
6(1964), 57–58.
[6] Golub, G.H., S. Nash, and C. Van Loan, "A HessenbergSchur Method for the Problem
AX + XB = C," IEEE Trans. Autom. Control, AC24(1979), 909913.
[7] Golub, G.H., and C.F. Van Loan, Matrix Computations, Third Edition, Johns Hopkins
Univ. Press, Baltimore, MD, 1996.
[8] Golub, G.H., and J.H. Wilkinson, "IllConditioned Eigensystems and the Computation
of the Jordan Canonical Form," SIAM Rev., 18(1976), 578619.
[9] Greville, T.N.E., "Note on the Generalized Inverse of a Matrix Product," SIAM Rev.,
8(1966), 518–521 [Erratum, SIAM Rev., 9(1967), 249].
[10] Halmos, PR., FiniteDimensional Vector Spaces, Second Edition, Van Nostrand,
Princeton, NJ, 1958.
[11] Higham, N.J., Accuracy and Stability of'Numerical Algorithms, Second Edition, SIAM,
Philadelphia, PA, 2002.
[12] Horn, R.A., and C.R. Johnson, Matrix Analysis, Cambridge Univ. Press, Cambridge,
UK, 1985.
[13] Horn, R.A., and C.R. Johnson, Topics in Matrix Analysis, Cambridge Univ. Press,
Cambridge, UK, 1991.
151
Bibliography
[1] Albert, A., Regression and the MoorePenrose Pseudoinverse, Academic Press, New
York, NY, 1972.
[2] Bartels, RH., and G.w. Stewart, "Algorithm 432. Solution of the Matrix Equation
AX + X B = C," Comm. ACM, 15(1972),820826.
[3] Bellman, R, Introduction to Matrix Analysis, Second Edition, McGrawHill, New
York, NY, 1970.
[4] Bjorck, A., Numerical Methodsfor Least Squares Problems, SIAM, Philadelphia, PA,
1996.
[5] Cline, R.E., "Note on the Generalized Inverse of the Product of Matrices," SIAM Rev.,
6(1964),5758.
[6] Golub, G.H., S. Nash, and C. Van Loan, "A HessenbergSchur Method for the Problem
AX + X B = C," IEEE Trans. Autom. Control, AC24(1979), 909913.
[7] Golub, G.H., and c.F. Van Loan, Matrix Computations, Third Edition, Johns Hopkins
Univ. Press, Baltimore, MD, 1996.
[8] Golub, G.H., and lH. Wilkinson, "IllConditioned Eigensystems and the Computation
ofthe Jordan Canonical Form," SIAM Rev., 18(1976),578619.
[9] Greville, T.N.E., "Note on the Generalized Inverse of a Matrix Product," SIAM Rev.,
8(1966),518521 [Erratum, SIAM Rev., 9(1967), 249].
[10] Halmos, P.R, FiniteDimensional Vector Spaces, Second Edition, Van Nostrand,
Princeton, NJ, 1958.
[11] Higham, N.1., Accuracy and Stability of Numerical Algorithms, Second Edition, SIAM,
Philadelphia, PA, 2002.
[12] Hom, RA., and C.R. Johnson, Matrix Analysis, Cambridge Univ. Press, Cambridge,
UK, 1985.
[13] Hom, RA., and C.R. Johnson, Topics in Matrix Analysis, Cambridge Univ. Press,
Cambridge, UK, 1991.
151
152 Bibliography
[14] Kenney, C, and A.J. Laub, "Controllability and Stability Radii for Companion Form
Systems," Math, of Control, Signals, and Systems, 1(1988), 361390.
[15] Kenney, C.S., and A.J. Laub, "The Matrix Sign Function," IEEE Trans. Autom. Control,
40(1995), 1330–1348.
[16] Lancaster, P., and M. Tismenetsky, The Theory of Matrices, Second Edition with
Applications, Academic Press, Orlando, FL, 1985.
[17] Laub, A.J., "A Schur Method for Solving Algebraic Riccati Equations," IEEE Trans..
Autom. Control, AC24( 1979), 913–921.
[18] Meyer, C.D., Matrix Analysis and Applied Linear Algebra, SIAM, Philadelphia, PA,
2000.
[19] Moler, C.B., and C.F. Van Loan, "Nineteen Dubious Ways to Compute the Exponential
of a Matrix," SIAM Rev., 20(1978), 801836.
[20] Noble, B., and J.W. Daniel, Applied Linear Algebra, Third Edition, PrenticeHall,
Englewood Cliffs, NJ, 1988.
[21] Ortega, J., Matrix Theory. A Second Course, Plenum, New York, NY, 1987.
[22] Penrose, R., "A Generalized Inverse for Matrices," Proc. Cambridge Philos. Soc.,
51(1955), 406–413.
[23] Stewart, G. W., Introduction to Matrix Computations, Academic Press, New York, NY,
1973.
[24] Strang, G., Linear Algebra and Its Applications, Third Edition, Harcourt Brace
Jovanovich, San Diego, CA, 1988.
[25] Watkins, D.S., Fundamentals of Matrix Computations, Second Edition, Wiley
Interscience, New York, 2002.
[26] Wonham, W.M., Linear Multivariable Control. A Geometric Approach, Third Edition,
SpringerVerlag, New York, NY, 1985.
152 Bibliography
[14] Kenney, C., and AJ. Laub, "Controllability and Stability Radii for Companion Fonn
Systems," Math. of Control, Signals, and Systems, 1(1988),361390.
[15] Kenney, C.S., andAJ. Laub, "The Matrix Sign Function," IEEE Trans. Autom. Control,
40(1995),13301348.
[16] Lancaster, P., and M. Tismenetsky, The Theory of Matrices, Second Edition with
Applications, Academic Press, Orlando, FL, 1985.
[17] Laub, AJ., "A Schur Method for Solving Algebraic Riccati Equations," IEEE Trans ..
Autom. Control, AC24( 1979), 913921.
[18] Meyer, C.D., Matrix Analysis and Applied Linear Algebra, SIAM, Philadelphia, PA,
2000.
[19] Moler, c.B., and c.P. Van Loan, "Nineteen Dubious Ways to Compute the Exponential
of a Matrix," SIAM Rev., 20(1978),801836.
[20] Noble, B., and J.w. Daniel, Applied Linear Algebra, Third Edition, PrenticeHall,
Englewood Cliffs, NJ, 1988.
[21] Ortega, J., Matrix Theory. A Second Course, Plenum, New York, NY, 1987.
[22] Pemose, R., "A Generalized Inverse for Matrices," Proc. Cambridge Philos. Soc.,
51(1955),406413.
[23] Stewart, G.W., Introduction to Matrix Computations, Academic Press, New York, NY,
1973.
[24] Strang, G., Linear Algebra and Its Applications, Third Edition, Harcourt Brace
Jovanovich, San Diego, CA, 1988.
[25] Watkins, D.S., Fundamentals of Matrix Computations, Second Edition, Wiley
Interscience, New York, 2002.
[26] Wonham, W.M., Linear Multivariable Control. A Geometric Approach, Third Edition,
SpringerVerlag, New York, NY, 1985.
Index
A–invariant subspace, 89
matrix characterization of, 90
algebraic multiplicity, 76
angle between vectors, 58
basis, 11
natural, 12
block matrix, 2
definiteness of, 104
diagonalization, 150
inverse of, 48
LU factorization, 5
triangularization, 149
C", 1
(pmxn i
(p/nxn 1
Cauchy–Bunyakovsky–Schwarz Inequal
ity, 58
Cayley–Hamilton Theorem, 75
chain
of eigenvectors, 87
characteristic polynomial
of a matrix, 75
of a matrix pencil, 125
Cholesky factorization, 101
co–domain, 17
column
rank, 23
vector, 1
companion matrix
inverse of, 105
pseudoinverse of, 106
singular values of, 106
singular vectors of, 106
complement
of a subspace, 13
orthogonal, 21
congruence, 103
conjugate transpose, 2
contragredient transformation, 137
controllability, 46
defective, 76
degree
of a principal vector, 85
determinant, 4
of a block matrix, 5
properties of, 4–6
dimension, 12
direct sum
of subspaces, 13
domain, 17
eigenvalue, 75
invariance under similarity transfor
mation, 81
elementary divisors, 84
equivalence transformation, 95
orthogonal, 95
unitary, 95
equivalent generalized eigenvalue prob
lems, 127
equivalent matrix pencils, 127
exchange matrix, 39, 89
exponential of a Jordan block, 91, 115
exponential of a matrix, 81, 109
computation of, 114–118
inverse of, 110
properties of, 109–112
field, 7
four fundamental subspaces, 23
function of a matrix, 81
generalized eigenvalue, 125
generalized real Schur form, 128
153
Index
Ainvariant subspace, 89
matrix characterization of, 90
algebraic multiplicity, 76
angle between vectors, 58
basis, 11
natural, 12
block matrix, 2
definiteness of, 104
diagonalization, 150
inverse of, 48
LV factorization, 5
triangularization, 149
en, 1
e
mxn
, 1
e ~ x n , 1
CauchyBunyakovskySchwarz Inequal
ity,58
CayleyHamilton Theorem, 75
chain
of eigenvectors, 87
characteristic polynomial
of a matrix, 75
of a matrix pencil, 125
Cholesky factorization, 101
codomain, 17
column
rank, 23
vector, 1
companion matrix
inverse of, 105
pseudoinverse of, 106
singular values of, 106
singular vectors of, 106
complement
of a subspace, 13
orthogonal, 21
153
congruence, 103
conjugate transpose, 2
contragredient transformation, 137
controllability, 46
defective, 76
degree
of a principal vector, 85
determinant, 4
of a block matrix, 5
properties of, 46
dimension, 12
direct sum
of subspaces, 13
domain, 17
eigenvalue, 75
invariance under similarity transfor
mation,81
elementary divisors, 84
equivalence transformation, 95
orthogonal, 95
unitary, 95
equivalent generalized eigenvalue prob
lems, 127
equivalent matrix pencils, 127
exchange matrix, 39, 89
exponential of a Jordan block, 91, 115
exponential of a matrix, 81, 109
computation of, 114118
inverse of, 110
properties of, 109112
field, 7
four fundamental subspaces, 23
function of a matrix, 81
generalized eigenvalue, 125
generalized real Schur form, 128
154 Index
generalized Schur form, 127
generalized singular value decomposition,
134
geometric multiplicity, 76
Holder Inequality, 58
Hermitian transpose, 2
higher–order difference equations
conversion to first–order form, 121
higher–order differential equations
conversion to first–order form, 120
higher–order eigenvalue problems
conversion to first–order form, 136
i, 2
idempotent, 6, 51
identity matrix, 4
inertia, 103
initial–value problem, 109
for higher–order equations, 120
for homogeneous linear difference
equations, 118
for homogeneous linear differential
equations, 112
for inhomogeneous linear difference
equations, 119
for inhomogeneous linear differen
tial equations, 112
inner product
complex, 55
complex Euclidean, 4
Euclidean, 4, 54
real, 54
usual, 54
weighted, 54
invariant factors, 84
inverses
of block matrices, 47
7, 2
Jordan block, 82
Jordan canonical form (JCF), 82
Kronecker canonical form (KCF), 129
Kronecker delta, 20
Kronecker product, 139
determinant of, 142
eigenvalues of, 141
eigenvectors of, 141
products of, 140
pseudoinverse of, 148
singular values of, 141
trace of, 142
transpose of, 140
Kronecker sum, 142
eigenvalues of, 143
eigenvectors of, 143
exponential of, 149
leading principal submatrix, 100
left eigenvector, 75
left generalized eigenvector, 125
left invertible, 26
left nullspace, 22
left principal vector, 85
linear dependence, 10
linear equations
characterization of all solutions, 44
existence of solutions, 44
uniqueness of solutions, 45
linear independence, 10
linear least squares problem, 65
general solution of, 66
geometric solution of, 67
residual of, 65
solution via QR factorization, 71
solution via singular value decom
position, 70
statement of, 65
uniqueness of solution, 66
linear regression, 67
linear transformation, 17
co–domain of, 17
composition of, 19
domain of, 17
invertible, 25
left invertible, 26
matrix representation of, 18
nonsingular, 25
nullspace of, 20
154
generalized Schur form, 127
generalized singular value decomposition,
134
geometric multiplicity, 76
Holder Inequality, 58
Hermitian transpose, 2
higherorder difference equations
conversion to firstorder form, 121
higherorder differential equations
conversion to firstorder form, 120
higherorder eigenvalue problems
conversion to firstorder form, 136
i,2
idempotent, 6, 51
identity matrix, 4
inertia, 103
initialvalue problem, 109
for higherorder equations, 120
for homogeneous linear difference
equations, 118
for homogeneous linear differential
equations, 112
for inhomogeneous linear difference
equations, 119
for inhomogeneous linear differen
tial equations, 112
inner product
complex, 55
complex Euclidean, 4
Euclidean, 4, 54
real, 54
usual, 54
weighted, 54
invariant factors, 84
inverses
of block matrices, 47
j,2
Jordan block, 82
Jordan canonical form (JCF), 82
Kronecker canonical form (KCF), 129
Kronecker delta, 20
Kronecker product, 139
determinant of, 142
eigenvalues of, 141
eigenvectors of, 141
products of, 140
pseudoinverse of, 148
singUlar values of, 141
trace of, 142
transpose of, 140
Kronecker sum, 142
eigenvalues of, 143
eigenvectors of, 143
exponential of, 149
leading principal submatrix, 100
left eigenvector, 75
left generalized eigenvector, 125
left invertible. 26
left nullspace, 22
left principal vector, 85
linear dependence, 10
linear equations
Index
characterization of all solutions, 44
existence of solutions, 44
uniqueness of solutions, 45
linear independence, 10
linear least squares problem, 65
general solution of, 66
geometric solution of, 67
residual of, 65
solution via QR factorization, 71
solution via singular value decom
position, 70
statement of, 65
uniqueness of solution, 66
linear regression, 67
linear transformation, 17
codomain of, 17
composition of, 19
domain of, 17
invertible, 25
left invertible. 26
matrix representation of, 18
nonsingular, 25
nulls pace of, 20
Index 155
range of, 20
right invertible, 26
LU factorization, 6
block, 5
Lyapunov differential equation, 113
Lyapunov equation, 144
and asymptotic stability, 146
integral form of solution, 146
symmetry of solution, 146
uniqueness of solution, 146
matrix
asymptotically stable, 145
best rank k approximation to, 67
companion, 105
defective, 76
definite, 99
derogatory, 106
diagonal, 2
exponential, 109
Hamiltonian, 122
Hermitian, 2
Householder, 97
indefinite, 99
lower Hessenberg, 2
lower triangular, 2
nearest singular matrix to, 67
nilpotent, 115
nonderogatory, 105
normal, 33, 95
orthogonal, 4
pentadiagonal, 2
quasi–upper–triangular, 98
sign of a, 91
square root of a, 101
symmetric, 2
symplectic, 122
tridiagonal, 2
unitary, 4
upper Hessenberg, 2
upper triangular, 2
matrix exponential, 81, 91, 109
matrix norm, 59
1–.60
2–, 60
oo–, 60
/?–, 60
consistent, 61
Frobenius, 60
induced by a vector norm, 61
mixed, 60
mutually consistent, 61
relations among, 61
Schatten, 60
spectral, 60
subordinate to a vector norm, 61
unitarily invariant, 62
matrix pencil, 125
equivalent, 127
reciprocal, 126
regular, 126
singular, 126
matrix sign function, 91
minimal polynomial, 76
monic polynomial, 76
Moore–Penrose pseudoinverse, 29
multiplication
matrix–matrix, 3
matrix–vector, 3
Murnaghan–Wintner Theorem, 98
negative definite, 99
negative invariant subspace, 92
nonnegative definite, 99
criteria for, 100
nonpositive definite, 99
norm
induced, 56
natural, 56
normal equations, 65
normed linear space, 57
nullity, 24
nullspace, 20
left, 22
right, 22
observability, 46
one–to–one (1–1), 23
conditions for, 25
onto, 23
conditions for, 25
Index
range of, 20
right invertible, 26
LV factorization, 6
block,5
Lyapunov differential equation, 113
Lyapunov equation, 144
and asymptotic stability, 146
integral form of solution, 146
symmetry of solution, 146
uniqueness of solution, 146
matrix
asymptotically stable, 145
best rank k approximation to, 67
companion, 105
defective, 76
definite, 99
derogatory, 106
diagonal,2
exponential, 109
Hamiltonian, 122
Hermitian, 2
Householder, 97
indefinite, 99
lower Hessenberg, 2
lower triangular, 2
nearest singular matrix to, 67
nilpotent, 115
nonderogatory, 105
normal, 33, 95
orthogonal, 4
pentadiagonal, 2
quasiuppertriangular, 98
sign of a, 91
square root of a, 10 1
symmetric, 2
symplectic, 122
tridiagonal, 2
unitary, 4
upper Hessenberg, 2
upper triangular, 2
matrix exponential, 81, 91, 109
matrix norm, 59
1,60
2,60
00,60
p,60
consistent, 61
Frobenius, 60
induced by a vector norm, 61
mixed,60
mutually consistent, 61
relations among, 61
Schatten,60
spectral, 60
155
subordinate to a vector norm, 61
unitarily invariant, 62
matrix pencil, 125
equivalent, 127
reciprocal, 126
regular, 126
singUlar, 126
matrix sign function, 91
minimal polynomial, 76
monic polynomial, 76
MoorePenrose pseudoinverse, 29
multiplication
matrixmatrix, 3
matrixvector, 3
MumaghanWintner Theorem, 98
negative definite, 99
negative invariant subspace, 92
nonnegative definite, 99
criteria for, 100
nonpositive definite, 99
norm
induced,56
natural,56
normal equations, 65
normed linear space, 57
nullity, 24
nullspace,20
left, 22
right, 22
observability, 46
onetoone (11), 23
conditions for, 25
onto, 23
conditions for, 25
156 Index
orthogonal
complement, 21
matrix, 4
projection, 52
subspaces, 14
vectors, 4, 20
orthonormal
vectors, 4, 20
outer product, 19
and Kronecker product, 140
exponential of, 121
pseudoinverse of, 33
singular value decomposition of, 41
various matrix norms of, 63
pencil
equivalent, 127
of matrices, 125
reciprocal, 126
regular, 126
singular, 126
Penrose theorem, 30
polar factorization, 41
polarization identity, 57
positive definite, 99
criteria for, 100
positive invariant subspace, 92
power (Kth) of a Jordan block, 120
powers of a matrix
computation of, 119–120
principal submatrix, 100
projection
oblique, 51
on four fundamental subspaces, 52
orthogonal, 52
pseudoinverse, 29
four Penrose conditions for, 30
of a full–column–rank matrix, 30
of a full–row–rank matrix, 30
of a matrix product, 32
of a scalar, 31
of a vector, 31
uniqueness, 30
via singular value decomposition, 38
Pythagorean Identity, 59
Q –orthogonality, 55
QR factorization, 72
T O " 1
IK , 1
M
mxn i
, 1
M
mxn 1
r '
M nxn 1
n ' '
range, 20
range inclusion
characterized by pseudoinverses, 33
rank, 23
column, 23
row, 23
rank–one matrix, 19
rational canonical form, 104
Rayleigh quotient, 100
reachability, 46
real Schur canonical form, 98
real Schur form, 98
reciprocal matrix pencil, 126
reconstructibility, 46
regular matrix pencil, 126
residual, 65
resolvent, 111
reverse–order identity matrix, 39, 89
right eigenvector, 75
right generalized eigenvector, 125
right invertible, 26
right nullspace, 22
right principal vector, 85
row
rank, 23
vector, 1
Schur canonical form, 98
generalized, 127
Schur complement, 6, 48, 102, 104
Schur T heorem, 98
Schur vectors, 98
second–order eigenvalue problem, 135
conversion to first–order form, 135
Sherman–Morrison–Woodbury formula,
48
signature, 103
similarity transformation, 95
and invariance of eigenvalues, h
156
orthogonal
complement, 21
matrix, 4
projection, 52
subspaces, 14
vectors, 4, 20
orthonormal
vectors, 4, 20
outer product, 19
and Kronecker product, 140
exponential of, 121
pseudoinverse of, 33
singular value decomposition of, 41
various matrix norms of, 63
pencil
equivalent, 127
of matrices, 125
reciprocal, 126
regular, 126
singular, 126
Penrose theorem, 30
polar factorization, 41
polarization identity, 57
positive definite, 99
criteria for, 100
positive invariant subspace, 92
power (kth) of a Jordan block, 120
powers of a matrix
computation of, 119120
principal submatrix, 100
projection
oblique, 51
on four fundamental subspaces, 52
orthogonal, 52
pseudoinverse, 29
four Penrose conditions for, 30
of a fullcolumnrank matrix, 30
of a fullrowrank matrix, 30
of a matrix product, 32
of a scalar, 31
of a vector, 31
uniqueness, 30
via singular value decomposition, 38
Pythagorean Identity, 59
Qorthogonality, 55
QR factorization, 72
JR.n, I
JR.mxn,1
1
I
range, 20
range inclusion
Index
characterized by pseudoinverses, 33
rank, 23
column, 23
row, 23
rankone matrix, 19
rational canonical form, 104
Rayleigh quotient, 100
reachability, 46
real Schur canonical form, 98
real Schur form, 98
reciprocal matrix pencil, 126
reconstructibility, 46
regular matrix pencil, 126
residual, 65
resolvent, III
reverseorder identity matrix, 39, 89
right eigenvector, 75
right generalized eigenvector, 125
right invertible, 26
right nullspace, 22
right principal vector, 85
row
rank, 23
vector, I
Schur canonical form, 98
generalized, 127
Schur complement, 6, 48, 102, 104
Schur Theorem, 98
Schur vectors, 98
secondorder eigenvalue problem, 135
conversion to firstorder form, 135
ShermanMorrisonWoodbury formula,
48
signature, 103
similarity transformation, 95
and invariance of eigenvalues, 81
Index 157
orthogonal, 95
unitary, 95
simple eigenvalue, 85
simultaneous diagonalization, 133
via singular value decomposition, 134
singular matrix pencil, 126
singular value decomposition (SVD), 35
and bases for four fundamental
subspaces, 38
and pseudoinverse, 38
and rank, 38
characterization of a matrix factor
ization as, 37
dyadic expansion, 38
examples, 37
full vs. compact, 37
fundamental theorem, 35
nonuniqueness, 36
singular values, 36
singular vectors
left, 36
right, 36
span, 11
spectral radius, 62, 107
spectral representation, 97
spectrum, 76
subordinate norm, 61
subspace, 9
A–invariant, 89
deflating, 129
reducing, 130
subspaces
complements of, 13
direct sum of, 13
equality of, 10
four fundamental, 23
intersection of, 13
orthogonal, 14
sum of, 13
Sylvester differential equation, 113
Sylvester equation, 144
integral form of solution, 145
uniqueness of solution, 145
Sylvester's Law of Inertia, 103
symmetric generalized eigenvalue prob
lem, 131
total least squares, 68
trace, 6
transpose, 2
characterization by inner product, 54
of a block matrix, 2
triangle inequality
for matrix norms, 59
for vector norms, 57
unitarily invariant
matrix norm, 62
vector norm, 58
variation of parameters, 112
vec
of a matrix, 145
of a matrix product, 147
vector norm, 57
l–, 57
2–, 57
oo–, 57
P–, 51
equivalent, 59
Euclidean, 57
Manhattan, 57
relations among, 59
unitarily invariant, 58
weighted, 58
weighted p–, 58
vector space, 8
dimension of, 12
vectors, 1
column, 1
linearly dependent, 10
linearly independent, 10
orthogonal, 4, 20
orthonormal, 4, 20
row, 1
span of a set of, 11
zeros
of a linear dynamical system, 130
Index
orthogonal, 95
unitary, 95
simple eigenvalue, 85
simultaneous diagonalization, 133
via singular value decomposition, 134
singular matrix pencil, 126
singular value decomposition (SVD), 35
and bases for four fundamental
subspaces, 38
and pseudoinverse, 38
and rank, 38
characterization of a matrix factor
ization as, 37
dyadic expansion, 38
examples, 37
full vs. compact, 37
fundamental theorem, 35
nonuniqueness, 36
singular values, 36
singular vectors
left, 36
right, 36
span, 11
spectral radius, 62, 107
spectral representation, 97
spectrum, 76
subordinate norm, 61
subspace, 9
Ainvariant, 89
deflating, 129
reducing, 130
subspaces
complements of, 13
direct sum of, 13
equality of, 10
four fundamental, 23
intersection of, 13
orthogonal, 14
sum of, 13
Sylvester differential equation, 113
Sylvester equation, 144
integral form of solution, 145
uniqueness of solution, 145
157
Sylvester's Law of Inertia, 103
symmetric generalized eigenvalue prob
lem,131
total least squares, 68
trace, 6
transpose, 2
characterization by inner product, 54
of a block matrix, 2
triangle inequality
for matrix norms, 59
for vector norms, 57
unitarily invariant
matrix norm, 62
vector norm, 58
variation of parameters, 112
vec
of a matrix, 145
of a matrix product, 147
vector norm, 57
1,57
2,57
00,57
p,57
equivalent, 59
Euclidean, 57
Manhattan, 57
relations among, 59
unitarily invariant, 58
weighted, 58
weighted p, 58
vector space, 8
dimension of, 12
vectors, 1
column, 1
linearly dependent, 10
linearly independent, 10
orthogonal, 4, 20
orthonormal, 4, 20
row, 1
span of a set of, 11
zeros
of a linear dynamical system, 130
Matrix Analysis Matrix Analysis
for Scientists & Engineers for Scientists & Engineers
This page intentionally left blank This page intentionally left blank
California slam. . Laub Alan J.Matrix Analysis Matrix Analysis for Scientists & Engineers for Scientists & Engineers Alan J. Laub University of California Davis.
Printed in the United States of America. Alan J. No part of this book may be reproduced. Mathematical analysis.. Mathematica is a registered trademark of Wolfram Research.lam. Natick. Mathcad is a registered trademark of Mathsoft Engineering & Education.. artist Aaron Tallon of Philadelphia. info@mathworks. p. Library of Congress CataloginginPublication Data Library of Congress CataloginginPublication Data Laub. For information. 2. write to the Society for Industrial and Applied Mathematics. Philadelphia. Mathematical analysis.. Alan J.. Inc. Used by permission. Mathcad is a registered trademark of Mathsoft Engineering & Education.. • slam is a registered trademark.com.Copyright Copyright © 2005 by the Society for Industrial and Applied Mathematics. cm. 5086477000. 3600 University City Science Center. Inc. MATLAB® is a registered trademark of The MathWorks. Matrix analysis for scientists and engineers / Alan J. Title. No part of this book All rights reserved.9'434dc22 2004059962 2004059962 About the cover: The original artwork featured on the cover was created by freelance About the cover: The original artwork featured on the cover was created by freelance artist Aaron Tallon of Philadelphia. I.com 5086477000.L38 2005 512. 1948Matrix analysis for scientists and engineers / Alan J. p. Inc.com. PA 191042688. Includes bibliographical references and index. I. 3 Apple Hill Drive. For MATLAB product information. Title. Inc. Matrices. MATLAB® is a registered trademark of The MathWorks. For MATLAB product information. Fax: 5086477101. or transmitted in any manner without the written permission may be reproduced. cm. 1.com Mathematica is a registered trademark of Wolfram Research. Inc. 10987654321 10987654321 All rights reserved. For information. . Philadelphia. 2. PA. Matrices. MA 017602098 USA. is a registered trademark. Printed in the United States of America. wwwmathworks. Inc. PA 191042688.) ISBN 0898715768 (pbk. 3600 University City Science Center. ISBN 0898715768 (pbk. or transmitted in any manner without the written permission of the publisher. 2005 by the Society for Industrial and Applied Mathematics. QA188138 2005 QA 188. please contact The MathWorks. Inc. stored. www. please contact The MathWorks.9'434—dc22 512. Laub. Laub.. Inc. 3 Apple Hill Drive. Mathematics. info@mathworks.mathworks.) 1.. Fax: 5086477101. write to the Society for Industrial and Applied of the publisher. 1948Laub. PA. MA 017602098 USA. Includes bibliographical references and index. Natick. stored. Used by permission 5.
Beverley (who captivated me in the UBC math library captivated UBC nearly forty years ago) nearly forty .To my wife. Beverley To my wife.
This page intentionally left blank This page intentionally left blank .
..2 Subspaces. . ..1 Definitions and Examples . 2. .. .. . 1.. . 3. . . .2 Subspaces 2. 6.2 Matrix Linear Equations . . ... . 5.3 A More General Matrix Linear Equation 6. .. 3.1 Definition and Examples 3.3 Composition of Transformations 3. . 4. .1 The Fundamental Theorem 5. .3 Inner Products and Orthogonality 1.2 Some Basic Properties 5. . 2. 6. .3 Linear Independence 2. 4. .1 Some Notation and Terminology 1.4 Some Useful and Interesting Inverses.2 Matrix Arithmetic .5 Four Fundamental Subspaces ..1 Vector Linear Equations . . .. .2 Matrix Representation of Linear Transformations 3. 6. .1 Some Notation and Terminology 1. .3 Properties and Applications .4 Determinants 1.. ....1 2. . . .4 Sums and Intersections of Subspaces Linear Transformations Linear Transformations 3.1 Definition and Examples . . . .1 4. ..3 Properties and Applications Introduction to the Singular Value Decomposition Introduction to the Singular Value Decomposition 5.2 Matrix Linear Equations 6. 3.3 Row and Column Compressions Linear Equations Linear Equations 6. ..2 Examples.4 Sums and Intersections of Subspaces 2.3 Composition of Transformations .2 Examples 4. .4 Structure of Linear Transformations 3. . .. . .2 Matrix Representation of Linear Transformations 3. . .4 Some Useful and Interesting Inverses vii xi xi 1 1 1 1 3 3 4 4 4 7 7 7 7 9 9 10 10 13 13 17 17 17 17 18 18 19 19 20 20 22 22 2 2 3 3 4 4 29 29 30 30 31 31 35 35 35 35 38 40 5 5 6 6 43 43 43 43 44 47 47 47 47 . .1 Vector Linear Equations 6. .1 Definitions and Characterizations Definitions and Characterizations. 4. 5.5 Four Fundamental Subspaces Introduction to the MoorePenrose Pseudoinverse Introduction to the MoorePenrose Pseudoinverse 4.2 Matrix Arithmetic 1.1 The Fundamental Theorem .3 Linear Independence . . .3 Rowand Column Compressions 5. 1.4 Determinants Vector Spaces Vector Spaces 2. . .4 Structure of Linear Transformations 3.3 Inner Products and Orthogonality . . .3 A More General Matrix Linear Equation 6.2 Some Basic Properties .Contents Contents Preface Preface 1 1 Introduction and Review Introduction and Review 1. Definitions and Examples 2.
. . . . . . .2 Inhomogeneous linear difference equations 11. Inner Product Spaces. .3 HigherOrder Equations . . .1 9. .1 Block matrices and definiteness 10.3 Vector Norms 7.3 Linear Regression and Other Linear Least Squares Problems 8. .2. .3 Determination of the JCF . .1 Fundamental Definitions and Properties 9.3. . . .1.1 Theoretical computation . . . .2 Definite Matrices . Eigenvalues and Eigenvectors 9.2 On the +1's in JCF blocks 9.2 Other least squares problems . . . . . .1 Homogeneous linear difference equations 11. . 8. . . .1 The Linear Least Squares Problem 8.5 Modal decompositions . . . . . 11. .5 The Matrix Sign Function. 11.1 7. .2 Jordan Canonical Form . . . . . . . . . 7. .6 Computation of the matrix exponential 11. . .1 Some Basic Canonical Forms 10.1 8.5 Least Squares and QR Factorization 8. Example: Linear regression 8. . 9.2 Difference Equations . . 9. 8.3 Determination of the JCF 9. . . . 7. .1.3. . .3. . . .5 Least Squares and QR Factorization .1. . . . and Norms 7. . .3. .1 Projections 7.3 Computation of matrix powers 11.2 Difference Equations 11.2 Definite Matrices 10. 9.2.2.2 Homogeneous linear differential equations 11.3.3 Computation of matrix powers . . .1.1 Fundamental Definitions and Properties 9. .viii viii Contents Contents 7 Projections.2. . . .4 Least Squares and Singular Value Decomposition 8. . .1 Properties ofthe matrix exponential . . . .1. . 11.1 The Linear Least Squares Problem . . . .3 Equivalence Transformations and Congruence 10.3. .1 Block matrices and definiteness 10. . . . . .4 Rational Canonical Form 11 Linear Differential and Difference Equations 11 Linear Differential and Difference Equations 11. . . . . 95 95 99 102 102 104 104 104 104 109 109 109 109 109 109 112 112 112 112 113 113 114 114 114 114 118 118 118 118 118 118 119 119 120 120 .4 Linear matrix differential equations 11. . . .3.1 Homogeneous linear difference equations 11.3 Inhomogeneous linear differential equations 11.3 HigherOrder Equations.1. . .4 Least Squares and Singular Value Decomposition 8. . .1. .3 Inhomogeneous linear differential equations 11.2 Other least squares problems 8. .1 Some Basic Canonical Forms .1. . . .2 Inner Product Spaces 7. . 11. .1. 11. .1 The four fundamental orthogonal projections The four fundamental orthogonal projections 7. . .1 Example: Linear regression . . . . . . .3 Linear Regression and Other Linear Least Squares Problems 8. 11. . . .4 Matrix Norms .2 Inhomogeneous linear difference equations 11. 10.1.2. . . .4 Linear matrix differential equations .3 Equivalence Transformations and Congruence 10.3. .1. . 8.4 Matrix Norms Linear Least Squares Problems 8.1.1 Properties of the matrix exponential 11. . .3 Vector Norms 7.3. . .2 Geometric Solution .2. .2 Jordan Canonical Form 9. . . .1 Differential Equations ILl Differential Equations . . 10. .2 Homogeneous linear differential equations 11.2 Geometric Solution 8. . .2 On the + l's in JCF blocks 9. . .6 Computation of the matrix exponential 11. .1.1 Projections .4 Rational Canonical Form . . . .2 Inner Product Spaces 7. . . .3.1.4 Geometric Aspects of the JCF 9. . . .5 The Matrix Sign Function 51 51 51 51 52 52 54 54 57 57 59 59 8 65 65 65 65 67 67 67 67 67 67 69 70 70 71 71 9 75 75 75 82 82 85 85 86 86 88 88 89 89 91 91 95 95 10 Canonical Forms 10. 11. . . . . . . Theoretical computation 9.4 Geometric Aspects of the JCF 9.5 Modal decompositions 11. . 10. .
1 Definition and Examples 13.5 Simultaneous Diagonalization .1 Definition and Examples . 12. . . . . .1 The Generalized Eigenvalue/Eigenvector Problem 12. .1 Simultaneous diagonalization via SVD 12. . . 12. . . . .1 Conversion to firstorder form 125 125 125 127 127 130 131 131 133 133 133 135 135 135 139 139 139 139 140 144 144 151 153 13 Kronecker Products 13 Kronecker Products 13. 12. .2 Canonical Forms 12. .2 Properties of the Kronecker Product . . . 13.6 HigherOrder Eigenvalue Problems 12. . .2 Properties of the Kronecker Product 13.1 Conversion to firstorder form 12. . .5. .6 HigherOrder Eigenvalue Problems . 13. . .Contents Contents ix ix 12 Generalized Eigenvalue Problems 12 Generalized Eigenvalue Problems 12.. . . . .3 Application to Sylvester and Lyapunov Equations 13. .1 The Generalized EigenvaluelEigenvector Problem 12. .3 Application to the Computation of System Zeros . 12.4 Symmetric Generalized Eigenvalue Problems .4 Symmetric Generalized Eigenvalue Problems 12.6. . . .2 Canonical Forms . .1 Simultaneous diagonalization via SVD 12.3 Application to the Computation of System Zeros 12. 12. . . . .3 Application to Sylvester and Lyapunov Equations Bibliography Bibliography Index Index . . . .6.5.5 Simultaneous Diagonalization 12. . .
This page intentionally left blank This page intentionally left blank .
Matrices are stressed more than abstract vector spaces. eigenvalues and eigenvectors. The books by Meyer [18]. For this. followon topics on the computational side (at the level of [7]. By matrix analysis I mean linear algebra and matrix theory together with their intrinsic interaction with and application to algebra and matrix theory together with their intrinsic interaction with and application to linear dynamical systems (systems of linear differential or difference equations). or computational science science who wish to be familar with enough matrix analysis that they are prepared to use its enough analysis they are prepared to tools and ideas comfortably in a variety of applications." However. The text can be used in a onequarter or onesemester course to provide a compact overview of can be used in a onequarter or onesemester course to provide a compact overview of much of the important and useful mathematics that. methods. For this. but somehow didn't quite manage to do. or [16]. requiring such material as prerequisite permits tion may occasionally be "hazy. By matrix analysis I mean linear tools and ideas comfortably in a variety of applications. the student is then wellequipped to pursue. either via formal courses or through selfstudy. the sciences. Instructors are encouraged to supplement the book with specific application examples from their own encouraged to supplement the book with specific application examples from their own particular subject area. example) or on the theoretical side (at the level of [12]. [13].Preface Preface This book is intended to be used as a text for beginning graduatelevel (or even seniorlevel) This book is intended to be used as a text for beginning graduatelevel (or even seniorlevel) students in engineering. or [25]. Certain topics that may have been treated cursorily in undergraduate courses are treated in more depth that may have been treated cursorily in undergraduate courses are treated in more depth and more advanced material is introduced. essentially Prerequisites for using this text are quite modest: essentially just an understanding for this understanding of calculus and definitely some previous exposure to matrices and linear algebra. [23]. Ortega [21].e. and Strang [24] Ortega are excellent companion texts for this book. Upon completion of a course based on this are excellent companion texts for this book. I have tried throughout to emphasize only the and more advanced material is introduced. requiring such material as prerequisite permits the early (but "outoforder" by conventional standards) introduction of topics such as pseuthe early (but "outoforder" by conventional standards) introduction of topics such as pseudoinverses and the singular value decomposition (SVD). Noble and Daniel [20]. and mathematical structures. Basic concepts such as determinants. I have tried throughout to emphasize only the more important and "useful" tools." However. singularity of matrices. computer science. the sciences.. in many cases. These powerful and versatile tools can then be exploited to provide a unifying foundation upon which to base subsequent topcan exploited to foundation subsequent topics. Matrices are stressed more than abstract vector spaces. I highly recommend MATLAB® although other software such as xi xi . example) or on the theoretical side (at the level of [12]. basisfree or subspace) aspects of many of the fundamental notions. and positive definite matrices should have been covered at least once. Upon completion of a course based on this text. Because tools such as the SVD are not generally amenable to "hand computation. students meant to learn much of the important and useful mathematics that. mathematics. students meant to learn thoroughly as undergraduates.e. Because tools such as the SVD are not generally amenable to "hand computation. for example). particular subject area. for [11]. and concepts such as determinants. [II]. eigenvalues and eigenvectors. The choice of topics covered in linear algebra and matrix theory is motivated both by The choice of topics covered in linear algebra and matrix theory is motivated both by applications and by computational utility and relevance. The concept of matrix factorization applications and by computational utility and relevance." this ics. The text linear dynamical systems (systems of linear differential or difference equations). although Chapters 2 and 3 do cover some geometric (i. for example). in many cases. Certain topics thoroughly as undergraduates. mathematics. either via formal courses or through selftext.. but somehow didn't quite manage to do. singularity of matrices. the student is then wellequipped to pursue. The concept of matrix factorization is emphasized throughout to provide a foundation for a later course in numerical linear is emphasized throughout to provide a foundation for a later course in numerical linear algebra." this approach necessarily presupposes the availability of appropriate mathematical software on approach necessarily presupposes the availability of appropriate mathematical software on a digital computer. computer science. These powerful and versatile tools doinverses and the singular value decomposition (SVD). [13]. basisfree or subspace) aspects of many of the fundamental do cover some geometric (i. even though their recollecmatrices least tion may occasionally be "hazy. although Chapters 2 and 3 algebra. I highly recommend MAlLAB® although other software such as a digital computer. Basic of calculus and definitely some previous exposure to matrices and linear algebra. or computational students in engineering. or [16].
in particular. chemistry. simulated. Some of the key algorithms of numerical linear algebra. applied physics. the student does require a certain amount of what is conventionally referred Proofs referred to as "mathematical maturity. If If are linearly dependent." a set of vectors is either linearly independent or it is not. "reallife" problems seldom yield to simple "reallife" closedform formulas or solutions. When they are not given explicitly. This is ideal material from which to learn a bit about mathematical proofs and the mathematical maturity and insight gained thereby. Mastery of the material in this text should enable the student to read and understand the modern language of matrices used throughout mathematics. for example. completed the course. and the course has proven to be remarkably successful at enabling students from Davis. how "nearly dependent" are the vectors? If they linearly independent. the details of most of the numerical aspects of linear algebra per se. are there "best" linearly independent subsets? These tum out to turn be much more difficult problems and frequently involve researchlevel questions when set be much more difficult problems and frequently involve researchlevel questions when set in the context of the finiteprecision. The presentation of the material in this book is strongly influenced by computais influenced by computational issues for two principal reasons. econometrics. A second motivation for a computational emphasis is that it provides many of the essential tools for what I call "qualitative mathematics. they are either obvious or easily found in the literature. they are either obvious or easily found in the literature. modern Some of the applications of matrix analysis mentioned briefly in this book derive of the applications of matrix analysis mentioned briefly in this book modem statespace from the modern statespace approach to dynamical systems." Proofs are given for many theorems. . and coherent fashion. First. This is an absolutely fundamental fundamental concept. finiterange floatingpoint arithmetic environment of of of most modem computing platforms. diverse audience. It is my firm conviction that such maturity is neither encouraged conviction neither nor nurtured by relegating the mathematical aspects of applications (for example. The "language" in which such described models are conveniently described involves vectors and matrices. many students who completed especially offered. form the foundation Some of the key algorithms of numerical linear algebra. many times at UCSB and twice at UC Davis. Indeed. form the foundation virtually modem upon which rests virtually all of modern scientific and engineering computation. and states often give rise to models of very numbers models high order that must be analyzed. in an elementary linear algebra course. But in most engineering or scientific contexts we want to know more than that. or signal processing. and thus the text can serve a rather diverse audience. This is ideal not given explicitly. statistics. consistent. and while most material is developed from basic ideas in the book. linear algebra introducing "onthefly" algebra for elementary statespace theory) to an appendix or introducing it "onthef1y" when to necessary. outputs. one must lay a firm foundation upon which subsequent applications and Rather. mathematics. and modem engineering. science. Rather. prerequisites developed While prerequisites for this text are modest." For example. I have taught this material for many years. It is thus crucial to acquire knowledge vocabulary a working knowledge of the vocabulary and grammar of this language. If a set of vectors is linearly independent. and evaluated. control systems with standard large numbers of interacting inputs. remarked afterward that if processing. are deferred to such a course. The tools of matrix analysis are also applied on a daily basis to problems in biology. They must generally be solved computationally and closedform it is important to know which types of algorithms can be relied upon and which cannot. are deferred to such a course. and a wide variety of other fields. if only they had had this course before they took linear systems. Statespace methods are Statespace modem now standard in much of modern engineering where. and the course has proven to be remarkably successful at enabling students from disparate backgrounds to acquire a quite acceptable level of mathematical maturity and acceptable graduate rigor for subsequent graduate studies in a variety of disciplines. in particular. must lay firm foundation upon which and perspectives perspectives can be built in a logical. especially the first few times it was offered.xii xii Preface Preface Mathcad® Mathematica® or Mathcad® is also excellent. Since this text is not intended for a course in numerical linear algebra per se.
My fellow instructors. they would have been able to concentrate on the new ideas deficiencies they wanted to learn. they no longer had to provide as much time for "review" and could focus instead on the subject at hand. etc. June 2004 — AJL. AJL. realized that by requiring this course as a prerequisite. The concept seems to work. too..Preface Preface xiii XIII or estimation theory. . rather than having to spend time making up for deficiencies in their background background in matrices and linear algebra.
This page intentionally left blank This page intentionally left blank .
xyT is an n x n matrix. IR rn xn = the set of real (or realvalued) m x n matrices. the notation!! denotes the set {I. Thus.n xn Rmxnr = the set of real m x n matrices of rank r. That a vector is always a y E IR n and the superscript T is the transpose operation. The following sets appear frequently throughout subsequent chapters: The following sets appear frequently throughout subsequent chapters: 1. mxn = the set of complex (or complexvalued) x n matrices. 3. This is followed by a review of some basic notions in matrix analysis and linear algebra. This is followed by a review of some basic notions in matrix analysis throughout the text.. nonsingular n x n matrices.g. 1 . n }. Henceforth. x e IR n means means where xi e IR for e n. That a vector is always a column vector rather than a row vector is entirely arbitrary. the set of ntuples of complex numbers represented as column vectors.1 Some Notation and Terminology Some Notation and Terminology We begin with a brief introduction to some standard notation and terminology to be used We begin with a brief introduction to some standard notation and terminology to be used throughout the text. e. but this convention makes it easy to recognize immediately throughout the text that. Rn = the set of ntuples of real numbers represented as column vectors. and linear algebra.. x E Rn I. IR n = the set of ntuples of real numbers represented as column vectors. XTy is a scalar while it easy to recognize immediately throughout the text that.. e. 5. R mxn = the set of real (or realvalued) m x n matrices. Henceforth. Thus. 2. IR~ xn denotes the set of real = set of real of rank Thus. A row vector is denoted by y~.n xn Cmxn = the set of complex m x n matrices of rank r. 2. n}.. Crnxn = the set of complex (or complexvalued) m x n matrices. but this convention makes column vector rather than a row vector is entirely arbitrary.Chapter 1 Chapter 1 Introduction and Review Introduction and Review 1. e 6. where Xi E R for ii E !!..g. . e. en 4. = the set of complex m x n matrices of rank r.. x T y is a scalar while xyT is an n x n matrix. .1 1. . Note: Vectors are always column vectors. Cn = the set of ntuples of complex numbers represented as column vectors. 1R. Thus.. where y G Rn and the superscript T is the transpose operation. the notation n denotes the set {1. 5. A row vector is denoted by yT where Note: Vectors are always column vectors. Rnxnn denotes the set of real nonsingular n x n matrices.
Each of the above also has a "block" analogue obtained by replacing scalar components in the respective definitions by block submatrices. For example.. We henceforth adopt the convention that.j  7+} ] is Hermitian (but not symmetric). then the (m + n) x (m + n) matrix [~ Bc] is block upper triangular. (7. 7 + j ] is complexvalued symmetric but not Hermitian. = a jfJ. j)th entry of a matrix A is denoted by AT and is the matrix whose j)th entry A. where the bar indicates complex j)th entry is (A H ). There is some advantage to being conversant with both notations. it is easy to see that if Aij are appropriately dimensioned subblocks. Introduction and Review Chapter 1. that is. A = [ .e.e. • tridiagonal if aij = 0 for Ii .. • upper triangular if aij. and definitions block submatrices. • upper Hessenberg if aij = 0 for ii .2.[ 7 . are appropriately dimensioned subblocks. = 0 for i > j. For example. 2 Transposes of block matrices can be defined in an obvious way. If A E em xn. A matrix A is symmetric i. text but reminders are placed at strategic locations. • lower Hessenberg if a. an equation like A = A T implies that A is realvalued while a statement like A = AH implies that A is complexvalued. 2 2. if e Rnxn e Rmxn C e Rmxm then the (m n) x (m n) matrix [A0 ~] is block upper triangular.ii > 1. unless if A = A T Hermitian A = A H.2 2 Chapter 1. A if A = AT and Hermitian if A = AH... C E jRmxm. The The transpose of a matrix A is denoted by AT and is the matrix whose (i. Example 1. = 0 for i > }. Example 1. a. z Remark 1. • pentadiagonal if ai. } is While R is most commonly denoted by i in mathematics texts. . otherwise noted. • lower Hessenberg if aij = 0 for } . A is conjugation. • lower triangular if aij7 = 0 for i/ < }.. it is Transposes of block matrices can be defined in an obvious way.. • pentadiagonal if aij = 0 for Ii . Hermitian conjugate sometimes A*) and its (i. then z = IX — jfi. i. B E IR nxm .1. • upper triangular if a. is complexvalued symmetric but not Hermitian. where the bar indicates complex sometimes A*) and its = IX jfJ (j = = v^T). For example. 1. = 0 for / — j\ > 2. an equation like A = A T implies that A is realvalued while a statement otherwise noted. = 0 for j — > 1.J I > 2. if A E IRnxn. Introduction and Review We now classify some of the more familiar "shaped" matrices. • upper Hessenberg if afj = 0 for — > 1. • tridiagonal if a(y = 0 for z — j\ > 1. A matrix A E IRn xn e (or A E enxn ) is A eC" x ")is • diagonal if a. i)th entry of A.. AT E E" xm is the (j. A = AH A complexvalued... = 0 for i ^ j. j)\h entry is (AH)ij = (aji). then easy to see that if A. then its Hermitian transpose (or conjugate transpose) is denoted by AH (or H If A e C mx ". then A7" e jRnxm. The notation j is used throughout the text but reminders are placed at strategic locations. Note that if A E R mx ". Oth (A 7 ). For example. • diagonal if aij7 = 0 for i i= }. 7 = («77). • lower triangular if a.jj > 1. While \/—\ is most commonly denoted by i in mathematics texts.JI > 1. A e jRmxn. ~ 5 is symmetric (and Hermitian). is Hermitian (but not symmetric). We henceforth that. There is some the more common notation in electrical engineering and system theory. j is Remark the more common notation in electrical engineering and system theory. then r = [ .2. ] is symmetric (and Hermitian). if z = a + jf$ (j = ii = R). (AT)ij = aji. A = [ 7+} 5 3· A .. = 0 for < j.
[ ~ J+l. suppose (linear combination) suppose A = la' .1.. Then the matrix product A B can be thought of as above. .3. Then v E jRP. {. i=I If matrices C and D are compatible for multiplication.. eRmxn has row cj e E l x ". suppose A e Rmxn and B = [bi. and multiplication of matrices. and is premultiplied by a row vector yTe jRlxm.[ ~ J+2. if (C D)H — D C ). the matrixvector product Ax.. vector x.. Ax..bhp ] e Rnxp with hi e jRn....2 Arithmetic It is assumed that the reader is familiar with the fundamental notions of matrix addition. A very important way to view this product is interpret weighted to interpret it as a weighted sum (linear combination) of the columns of A. That is. suppose A E jRmxn and [hI. Again. Let U = [MI.. . Theorem reader.. Namely. .. + Xnan E jRm. importance interpretation take A = [96 85 74]x = take A = [~ ~]. formulation of matrix multiplication that appears frequently in the text and is presented below as a theorem.. It Theorem 1. but equivalent. multiplication of a matrix by a scalar. its importance cannot be overemphasized.3 can then also be generalized to its "row dual.xn~ ] Then Ax = Xjal + . matrixvector if C E jRmxn has row vectors cJ E jRlxn. i. Vn] ]Ee lR Pxn U [Uj. As a numerical example..• a"1 E m JR " with a. vn Rpxn p with Vit e R .2.." The details are left to the readei "row left . Theorem 1.•.~].. A special case of matrix multiplication occurs when the second matrix is a column multiplication second i. there can be important computerarchitecturerelated advancomputerarchitecturerelated tages to preferring the latter calculation method.. n UV T = LUiVr E jRmxp.. applied p times: There is also an alternative.2 Matrix Arithmetic 1. AB bi E W1. Theorem 1. un Rmxn with u Rm and V = [v .. . matrixvector product with the column x to find Ax = [50 32]' but this matrixvector product can also be computed computed via v1a 3. It is deceptively simple and its full understanding is well rewarded.. This gives a dual to the matrixvector result above. E JRm and x = l I.e. x = ! 2 Then we can quickly calculate dot products of the rows of A [~]. Matrix Arithmetic 3 1. The importance of this interpretation cannot be overemphasized.3.. recall that (CD)T = DT C T (C D)T = DT T If H H H (or (CD} = DHC H ). multiplication.'" p] E jRnxp For matrix multiplication.[ ~ l For large arrays of numbers. and is premultiplied by a row yT E R l x m then the product can be written as a weighted linear sum of the rows of C as follows: follows: yTC=YICf +"'+Ymc~ EjRlxn. Then we can quickly calculate dot products of the rows of A column Ax = [. un]]Ee jRmxn with Ui t Ee jRm and V = [VI.e. .. .
4.=1 Note that the inner product is a scalar..e. the Euclidean inner product (or inner product. we define their complex Euclidean inner product (or inner product. then we say that x and y are orthonormal. Introduction and Review Chapter 1. where / is the n x n identity matrix. Note that x Tx = 0 if and only if x = 0 when x e Rn but that this is not true if x e Cn. E R are said to be orthogonal if their inner product is Two nonzero vectors x.. There is an orthogonal or unitary matrix has orthonormal rows and orthonormal columns. the nonzero vector x above. case. Note that x T x = 0 if and only if x = 0 when x E IRn but that this is not true if x E en. (or A E Cnxn) we use the notation det A for the determinant of A. where I is the n x n matrix A e IRnxn is an orthogonal matrix if ATA = AAT = /. the Euclidean inner product inner for short) y is given by y is given by n T (x.4. i.3 1. Y}c = {y. a matrix A e en xn is said to be unitary if A H A = AA H = I. Note that the inner product is a scalar. We list below some of (or A 6 en xn) we use the notation det A for the determinant of A. To illustrate. xTyy = 0.e. Let x = [} ]] and y = [~].4 1. A EC = (orC" xn). y e IRn are said to be orthogonal if their inner product is zero. The more conventional definition of the complex inner product is H ( x . . The notation /„ is sometimes used to denote the identity matrix in IR nxn in Rnx" x nxn H H (or en "). 1. A orthogonal and XTX = 1 and yTyy = 1.y.2j while while and we see that. There is no special name attached to a nonsquare matrix A E R mxn (or E Cmxn with orthonormal no special name attached to a nonsquare matrix A e ]Rrn"n (or € e mxn ))with orthonormal rows or columns. x)c'. What is true in the complex case is that XH x = 0 if and only if x = 0. consider the nonzero vector x above. (x. for short) by for short) by n (x'Y}c :=xHy = Lx. Then XTX = 0 but XHX = 2. Nonzero complex vectors are orthogonal if XHy = O. Similarly. for short) of x and For vectors y e IRn.3 Inner Products and Orthogonality Inner Products and Orthogonality For vectors x.y. A nxn matrix E R is an orthogonal matrix if AT A = AAT = I. If x.e. y)c = (y. Then (x. y)c = y = Eni=1 xiyi but throughout the text we prefer the symmetry with the real case. i. y E R".j] [ ~ ] = 1 . order in which x product is important. consider What is true in the complex case is that x H 0 if and only if O. For A E R nnxn A e IR xn It assumed of determinants. Introduction and Review 1. Two nonzero vectors x. If e C". Then x T x = 0 but x H X = 2. x T = O.4 4 Chapter 1. Example 1. . the order in which x and y appear in the complex inner (x. y)c = (y. Then Example 1..4 Determinants Determinants It is assumed that the reader is familiar with the basic theory of determinants. The more conventional definition of the complex inner product is product is important. indeed. x}c. rows or columns. y E <en. we define their complex Euclidean inner product (or inner product. To illustrate. then we say that x and y are orthonormal... x and y are orthogonal and x Tx = 1 and yT = 1.e. y)c = (y.y. x)c. Y}c = [ } JH [ ~ ] = [I . but throughout the text we prefer the symmetry with the real (x. Let x = [1j and y = [1/2]. If x and y are zero. Nonzero complex vectors are orthogonal if x H y = 0. y) := x y = Lx. In sometimes used denote identity matrix. x)c and we see that. Similarly. indeed.. i. (x. i. We list below some of . Clearly said = an orthogonal or unitary matrix has orthonormal rows and orthonormal columns.=1 y appear in Note that (x. y ) c = yHxx = L:7=1 x.
If A A A = o. B E IRnxn . 11. = alla22 • • ann i. Multiplying a row of A by a scalar and then adding it to another row does not change 7. If A is lower triangUlar. i. If A € lR~xn.thendet(AB) = det A det 5. Multiplying a row of A by a scalar and then adding it to another row does not change the determinant. Proof" Proof: This follows easily from the block UL factorization BD. If A. 11. 2. 16. then det A = alla22 • • ann 12.C).1 ) = de: A.. 16. If A is block diagonal (or block upper triangular or block lower triangular).• ann. Determinants 5 properties the more useful properties of determinants.. is a det A.A22. Note that this is not a minimal set. then det A = different det A11 det A22 .• ann. change the determinant.4.. If A e IRnxn and D E lR~xm. If A. A 22 .1 I ][ . exdetA. If A E lR~xn and DE IR mxm det [Ac ~] detA det(D . Interchanging two rows of A changes only the sign of the determinant. Multiplying a column of A by a scalar and then adding it to another column does not a column of scalar column does change the determinant.. 10. then det(AB) = det A det B. If A is upper triangular.. 14. 17.1..e. If elements. then det [Ac ~] det D det(A B D. 8. 3. 15. If A has a zero column or if any two columns of A are equal.4.e. 3. the determinant. then det A = all a22 .. If A is lower triangular. 15. then det [~ BD] = det D det(A – B D – 11C ) .• det Ann. If A e R n x n and D e R m x m .. i. 13. Multiplying A 6. B eR n x n .. properties 1. 4. det AT = det A (detA H = detA if A e C nxn ).. Multiplying a column of A by a scalar ex results in a new matrix whose determinant scalar a determinant is ex det A.. with A block diagonal (or block 13. then det [~ BD] = del A det(D – CA– l 1 B).B). then det A = O.• a"n. detAT = detA (det A H = det A A E C"X"). Determinants 1. then det A = a11a22 . If A E R n x n and D e RMmxm. det A 11 det A22 • • det Ann 14. several more is a properties are consequences of one or more of the others. If A has a zero row or if any two rows of A are equal. of 5.CA. Multiplying a row of A by a scalar a results in a new matrix whose determinant is 5. 9. then det A = a11a22 . If A is diagonal. Multiplying a row of A by a scalar ex results in a new matrix whose determinant is a det A. • • An" (of A = square diagonal blocks A11. If A E Rnxn. Interchanging two columns of A changes only the sign of the determinant. If A has a zero column or if any two columns of A are equal. det A is the product of its diagonal 10. 7. det A is the product of its diagonal diagonal..•... Proof: This follows easily from the block LU factorization Proof" This follows easily from the block LU factorization [~ ~J=[ ~ ][ ~ 17. . then det(A. If det = a11a22 • • ann 12. are consequences of one or more of the others. then det(A1) = 1detA .. A 11. Ann (of possibly different sizes). 8... then det A = 0.. then det A = 0.. Interchanging two rows of A changes only the sign of the determinant.e.
lower triangular with all l's on the diagonal) and an upper triangular matrix L 1's an V is called an LV factorization.e. lor z r 2sm2rt # J. U1 U2 • • Uk is an 5. U2 . Show that det(I – xyT) = 1. Tr(aA + f3B) = aTrA + fiTrB. example. aII o. If A e jRnxn and or is a scalar. (c) Let S € Rnxn be skewsymmetric. denoted TrA. Let A E jRNxn.e A – 1 B is called the Schur complement of A in[ACBD]. The factorization of a matrix A into the product of a unit lower triangular Remark 1. what is det(aA)? What is det(–A)? E R a det(A)? A? If A unitary. Another such factorization is UL where V is unit upper triangular and L is lower triangular. Then E jRnxn skewsymmetric. ... even though in general AB i= B A. see.y E jRn. 6. _. • . 2sin20 J is idempotent for all #.6. of Din [AC ~ l EXERCISES EXERCISES nxn 1.. what is det A? If A is unitary. B e JRn xn and a. A matrix A e Wx" is said to be idempotent if A2 = A.. A . ST = S. are block analogues of these. Suppose A E jRn xn is idempotent and A ^ I.6 6 Chapter 1. A? 2. Another such factorization is VL U is an LU factorization. The trace of A. (b) Suppose A e IR" X "is idempotent and A i= I. Showthatdet(lxyT) 1 – yTx.5. Introduction and Review Chapter 1.• V k is an orthogonal matrix. A E jRnxn A2 / x™ . y e Rn.. Remark — C I B – BDIe Similarly.. what is det A? If A 3. if A. The factorizations used above U triangular. . ft e R. is defined as the sum of its diagonal A e Rnxn.. 2 0 IS I dempotent . TrS O. i.e.5.e. II _ .. TrA = Eni=1 aii. [24].yTx. elements.e.. If A is orthogonal. . Uk E Rnxn U = VI V2 .. [~ ~ ].. . V2.. i. Letx. Introduction and Review Remark 1. Show that the product V = VI. The factorization of a matrix A into the product of a unit lower triangular matrix L (i.Vk € jRn xn be orthogonal matrices. Let U1.. i. see. either prove the converse or provide a counterexample. .. Remark 1. Tr(Afl) = Tr(£A). (a) Show that the trace is a linear function.e. Let x. 4. Show that A must be singular. ! [ 2cos2<9 I T 2cos2 0 (a) Show that the matrix A = _.B D – l C is the Schur complement of D in [~ BD ]. AB ^ BA. [24]. for example. then Tr(aA fiB)= aTrA + f3TrB. of denoted Tr A.. (b) Show that Tr(AB) = Tr(BA). nxn linear E R f3 E JR. . TrA = L~=I au· elements. Show that A must be singular. ST = So Show that TrS = 0. 2f) 2 _ sm 2^ sin 0 sin sin 20 1 . i. A =. The matrix D .
8) + y ffor all a.. (A3) for all a e F. (A4) a + . (M2) 1 e IF • I = for a e IF.p ) . aI = 1.8)· yyf for all a. (A2) there exists an element 0 e IF such that a 0 = a. (M2) there exists an element I E F such that a . including spaces formed by special classes emphasis is on finitedimensional vector spaces. (D) (D) a· p a . A field is a set IF together with two operations +.. . •) is an abelian group.ye¥. Axioms (A1)(A3) state that (F.8. (M3) e IF. .8 = ft + afar all a.8 . +) is a group and an abelian group if (A4) also holds. y Elf.1 Definitions and Examples Definition 2. when no confusion can arise. Axioms (MI)(M4) state that (IF \ {0}. ^ 0. (Ml) a . . ft Elf. Axioms (Al)(A3) state that (IF.((.8. (M4) a·.8 E IF.8 + y) = a·. I = a for all a E F. y)=cip+a.8.8. 2. where some of the proofs that are not given here may for this and the next chapter is [10].1. p.1. y Elf. (A2) there exists an element 0 E F such that a + 0 = a for all a E F. A field is a set F together with two operations +. Generally speaking. afar all a. there exists an element (a) e F such that a + (a) = 0. The emphasis is on finitedimensional vector spaces. (A4) a + p = . +) is a group and an abelian group if (A4) also holds. (M3) for all a E ¥. there exists an element aI E F such that a . the multiplication operator ". An excellent reference of matrices. An excellent reference for this and the next chapter is [10]. The In this chapter we give a brief review of some of the basic concepts of vector spaces. a"1 € IF • a~l = 1.8 +a· y for all a. not written explicitly. but some infinitedimensional examples are also cited. 7 . . ft.) ( a . : IF x F —> IF such that Definition 2.8 e F. y Elf. y € F. including spaces formed by special classes of matrices. when no confusion can arise. . the multiplication operator "•" is Generally speaking. y e F. but some infinitedimensional examples are also cited. (A3) for all a E IF. be found.Chapter 2 Vector Spaces Vector Spaces In this chapter we give a brief review of some of the basic concepts of vector spaces. where some of the proofs that are not given here may be found.y for alia.8 = P • a for all a. p.) is an abelian group." is not written explicitly. (M4) a • p =. (Ml) a· p . 0. (Al) a + (.8 a for all a.8 + y) = (a +. Axioms (M1)(M4) state that (F \ to). there exists an element (—a) E IF such that a (—a) O. yy) = (a·.((. a f.o r all a. • F x IF ~ F such that (Al) a (P y ) = (a + p ) y o r all a. p e F. for all a e IF.
A vector space is denoted by (V.3.2. IR with ordinary addition and multiplication is a field. 1.2. 4. ) f for all a. RMrmxn= {m x n matrices of rank r with real coefficients} is clearly not a field since. in Definition Remark 2. f3 e F and for all v E V.qEZ +} . Similar definitions hold for (C". (V2) ( a f3) v = a P . (V5) 1 v = v for all v e V (1 Elf). is a field. +) is an abelian group. when there is no possibility of confusion as to the underlying fie Id. w for all a e F and for all v..5. since (M4) does not hold in general (although the other 8 axioms hold). (V3) (a (V4) a(v w)=av a w for all a ElF andfor all v. ft) v = a v + f3.3 are different from the + and . R) with addition defined by I. e with ordinary complex addition and multiplication is a field. is a vector space. is a field. for example. IR~ xn = m x n matrices of rank r with real coefficients) is clearly not a field since.. +) is an abelian group.. In practice. (R". Note that + and· in Definition 2. f3 Elf andforall v E V. (V5) I·• v = v for all v E V (1 e F).1 in the sense of operating on different objects in different sets. v = a . Ra[x] = the field of rational functions in the indeterminate x 3. simply by V. p € F and for all v e V. 1. 4.5..2.1. 2. (V4) a· (v + w) = a . Moreover. no confusion and the operator is usually not even written explicitly. when there is no possibility of confusion as to the A vector space is denoted by (V.. }.. where Z+ = {O.3. 3. + f3qXq :aj.4. Example 2.r] = the field of rational functions in the indeterminate x = {ao + f30 + atX f3t X + . IF) or. this causes no confusion and the·• operator is usually not even written explicitly.• v = a·• v + p • v for all a. F) or.( (f3' V v) o r all a.3 are different from the + and • in Definition 2. p E IF andfor all v e V.. . Example 2.1 in the sense of operating on different objects in different sets. IR) with addition defined by and scalar multiplication defined by and scalar multiplication defined by is a vector space.p ) .8 Chapter 2. . Raf. (V3) (a + f3).P. w E V. v + a. + apxP + .f3i EIR . Definition 2. v for all a. Vector Spaces Example 2. this causes 2. C with ordinary complex addition and multiplication is a field.l. . C). A vector space over a field IF is a set V together with two operations + ::V x V + V and· :: IF x V + V such that V x V ^V and.4.}. lR~xn is not a field either since (M4) does not hold in general (although the other 8 axioms hold). Remark 2. Moreover. I. underlying field. A vector space over a field F is a set V together with two operations Definition 2. simply by V. where Z+ = {0. e). (V2) (a·.2. Note that + and • in Definition 2.F xV »• V such that (VI) (V. Example 2. (IRn. R"x" is not a field either for example. (VI) (V. 2. (Ml) does not hold unless m = n.. In practice. Similar definitions hold for (en. (MI) does not hold unless m = n.. w e V. R with ordinary addition and multiplication is a field.
F) be a vector space and let W ~ V. + fJmn l yaml yamn 3." + fJ2I a21 + P" .. Then (x(t) : x ( t ) = Ax(t)} is a vector space (of dimension n). too. and the functions are piecewise continuous (a) '0 = [to. F) be an arbitrary vector space and V be an arbitrary set. is henceforth understood to mean "is a subspace of. (JRmxn.6.7. td)n. IF) = (JR n . 4. td)n or continuous =: (C[?0. and for all f E cf>.. F) is itself a vector space or. E). that since 0 e F. IF) = (JRn. we write W ~ V." The when used with vector spaces. F) is a Definition 2.7. if and only subspace of (V.2 Subspaces Subspaces Definition 2. IF) be an arbitrary vector space and '0 be an arbitrary set. when used with vector spaces. Let cf>('O. Special Cases: Special Cases: (a) V = [to. Note. Let A € R"x"." ya2n . Then O(D. t\])n continuous =: (C[to. verify that the set in question is closed under addition and scalar multiplication. this question is closed under addition and scalar multiplication.2.e. and the functions are piecewise continuous =: (PC[to. and the symbol c. The latter characterization of a subspace is often the easiest way to check Remark 2. too. that since 0 E IF. amI al2 a22 + fJI2 + fJ22 aln + fJln a2n + fJ2n a mn + fJml am2 + fJm2 and scalar multiplication defined by and scalar multiplication defined by [ ya" y a 21 yA = .. implies that the zero vector must be in any subspace. 2. foral! a. V) be the 3.e. Then (W. E) is a vector space with addition defined by 9 9 A+B= [ . Then (W. (V. Notation: When the underlying field is understood. Let (V. and the symbol ~. less restrictive meaning "is a subset of' is specifically flagged as such. IF) is a subspace of (V. equivalently. IF) if and only if (W. Let A E JR(nxn. fJ e IF andforall WI. IF) is itself a vector space or. for all d ED. Then cf>('O.6. l . +00). (E mxn JR) is a vector space with addition defined by 2. Let (V. =: (PC[f0. (V. i. 4. if and only if(aw1 ßW2) E if(awl + fJw2) e W for all a. this implies that the zero vector must be in any subspace.2. V) is a vector space with addition set of functions f mapping '0 to V. verify that the set in or prove that something is indeed a subspace (or vector space). Subspaces 2.2 2. h])n (b) '0 = [to. we write W c V. IF) be a vector space and let W c V. The latter characterization of a subspace is often the easiest way to check or prove that something is indeed a subspace (or vector space). etc. JR). i. ß E ¥ and for all w1. is henceforth understood to mean "is a subspace of. F) if and only if (W. . y a l2 y a 22 yam 2 ya. w2 e Remark 2. equivalently. t\]. W f= 0. Let (V. g E cf> and scalar multiplication defined by and scalar multiplication defined by (af)(d) = af(d) for all a E IF. V) is a vector space with addition defined by defined by (f + g)(d) = fed) + g(d) for all d E '0 and for all f. Subspaces 2. F) = (IR". Notation: When the underlying field is understood.. Let (V. JR). W2 E W." The less restrictive meaning "is a subset of" is specifically flagged as such. V) be the set of functions / mapping D to V. W = 0. Note. td. Then {x(t) : x(t) = Ax(t}} is a vector space (of dimension n). (V. Let O(X>.2.
) and for each v E ]R2 of the form v = [~~ ] identify VI with the xcoordinate in the plane and u2 with the ycoordinate. Consider (V. 2. Consider (V. Vector Spaces Example 2. then R = S if and only if Definition 2.I.. Shifted subspaces W".nxn. V usually denotes a vector space with the underlying field generally being R unless Thus. unless explicitly stated otherwise. W2. vk E X and scalars a1. and S are vector spaces (or subspaces).S. W1/2. . + (XkVk = 0 implies al = 0.3 Linear Independence Linear Independence Let X = {v1. ft E R symmetric for all a. E ]Rnxn : not 2. that the vertical line through the origin (i. V usually denotes a vector space with the underlying field generally being JR. As an interesting exercise.e.10. Then W is /wf a subspace of JR. . Wi.• } be a nonempty collection of vectors Vi in some vector space V. one usually proves the two inclusions separately: An arbitrary r e R is shown to be an element of S and then an arbitrary s E S is shown to is shown to be an element of and then an arbitrary 5 € is shown to An arbitrary r E be an element of R. Then W". .e. v2.9.. Note. To prove two vector spaces are equal. ..../l with f3 = 0 are All lines through the origin are subspaces. al VI + .. As an interesting exercise../l = {V : v = [ ac ~ f3 ] .10.JR. . . . = {A E JR.lF) = (R" X ". W~V. ak = O. ffR and S are vector spaces (or subspaces). . we drop the explicit dependence of a vector space on an underlying field. f3 e R. explicitly stated otherwise. define W". c E JR.. V2. W2. too. Consider (V.} . sketch Then Wa. . Then it is easily shown that aAI + fiAi is symmetric for all a. Henceforth. . If 12.. that the vertical line through the origin (i. Then (V. For ß E R define the jccoordinate in the plane and V2 with the ycoordinate. R) and for each v € R2 of the form v = [v1v2 ] identify v1 with 3.) and let W = [A e R"x" : A is symmetric}.•. ak.nxn : A We V.9. Let X {VI.10 Chapter 2. All lines through the origin are subspaces. A2 are symmetric. . in some vector space V. . Proof: Suppose A\. one usually proves the two inclusions separately: Note: To prove two vector spaces are equal. X is a linearly dependent set of vectors ifand only if there exist k distinct if and only if exist distinct elements v1.nxn. W2. X linearly set of Definition 2.0. sketch W2. ak not all zero such that elements VI.1. Then it is easily shown that ctA\ + f3A2 is Proof' Suppose AI. f3 e R.R) and 1. Shifted subspaces Wa.Vk of X and for any scalars a1. Vk of X and for any scalars aI. ~SandS ~ R. JR. 1. too. Vk e X and scalars aI.I' and Wi.8. ak. . For a. elements VI. • • •} be a nonempty collection of vectors u. •••. we drop the explicit dependence of a vector space on an underlying field. Example 2.. IF) = (]R2. a = 00) is also a subspace. Henceforth. F) = (JR.. Note.O. . called linear varieties. . R"x".o. (Xk not all zero such that X is a linearly independent set of vectors if and only if for any collection of k distinct X is a linearly independent set of vectors if and only if for any collection of k distinct elements v1. 3. be an element of R. F) = (R2..o.. a = oo) is also a subspace.and W1/2. Definition 2. Let W = {A € R"x" : A is orthogonal}. . Definition 2. then R = S if and only if R C S and S C R. A2 are symmetric. ...ß is a subspace of V if and only if ß = O.1. Thus. •. ..ß with ß =1= 0 are called linear varieties.3 2. ./l is a subspace of V if and only if f3 = 0.
If the set of vectors is independent. (Xi ElF.. Linear Independence Example 2. Example 2.. and there exists a e R* such that Va = 0. The linear v E ]Rn. + (XkVk . Then consider the rows of etA B as vectors in Cm [t0.en} = ]Rn. Vi EX. o Definition 2. e2.. then = O. .. 1... Then consider the rows of etA B as vectors in em [to.. Why? However. The dependence of this set of vectors is equivalent to the existence of a nonzero vector E Rk dependence of this set of vectors is equivalent to the existence of a nonzero vector a e ]Rk O. = {v : where N = {I. = [ v 1 .3. Definition 2. LetV = 11 11 ~. An equivalent condition for linear dependence is that the k x k matrix condition VT V is singular. Let V = Rn and define = ]Rn and el = 0 0 . e2 = 0 1 0 ..en = 0 0 0 o SpIel. kEN}. Sp(X) = V. A set of vectors X is a basis for V if and only if 1. Vk] E Rnxk. {1. Vi e span of Definition 2. 2.. (since 2vI .13. . to be studied further in what follows. Then Sp{e1.. t1] (recall that etA denotes the matrix exponential. Sp(X) = V. Let A E ]Rnxn and 5 e R"xm.. X is a linearly independent set (of basis vectors). } = (Xl VI + .12."I [ i1i1l ]} [[ s a linearly is a Iin=ly dependent set de~ndent ~t (since 2v\ — V2 + v3 = 0). 2. 1£t V = R3. to be studied further in what follows. A set of vectors X is a basis for V if and only ij Definition 2. e k. Let X = {VI. An equivalent condition for linear independence is that the matrix Va = 0. linear dependence x such that Va = 0. which is discussed in more detail in efA Chapter 11).3. and consider the matrix V = [VI.Vk] e ]Rnxk. and X (of and 2.. . V2. If the set of vectors is independent. .14.11. }... Example 2. An equivalent condition for linear independence is that the matrix V TV is nonsingular. . Why? independent.v2 + V3 = 0). and there exists a E ]Rk such that VT V is singular. Independence of these vectors turns out to be equivalent to a concept called controllability.. e2 .14. E V.13. Linear Independence 2.. ii E If. . .. called consider Let Vif e R".12. X = [v1 v2 . . . tIl 2..11.. Then {[ Then I.. Howe. Then the span of of X is defined as X is defined as Sp(X) = Sp{VI. 2.'" . A e R xn B E ]Rnxm. then a = 0. V V2. ~ HHi] } Ime~ly i is a i" linearly independent set. .•}} be a collection of vectors vi. T V is nonsingular.. . Independence of these vectors turns out to be equivalent to a concept Chapter 11).}.. en} = Rn...2. .
with respect to the basis with respect to the basis {[~l[!J} we have we have [ ~ ~ ] = 3.E~n} such that v= where ~Ibl + . .. .. ] l = = Theorem 2... while [ ~ ] = I . + vne n · Vn We can also determine components of v with respect to another basis. Example 2. bn]} and are unique.. Definition 2. B ~ [b". ... r I [ . For example.. en} natural Now let b l . n for Then for all e there exists a unique ntuple {E1 ...18.. For be n. . .. el + 2 . particular basis considered. write [ ] = XI • [ ~ + ] X2 • [ _! ] =[ ~ = [ ~ Then Then ! ][ ~~ l 1 [ ~~ ] = [ ..l. For example. en} is a basis for IR" (sometimes called the natural basis). Vector Spaces Example 2. + ~nbn = Bx. If V= 0) V is Definition 2..17.. We say that the vector x of of of (b1. .dimensional or have dimension n and we write dim (V) = n or dim V — n. .. V is said to X for be ndimensional or have dimension n and we write dim(V) n or dim V n..16.19.19.15.12 12 Chapter 2. while We can also determine components of v with respect to another basis. .. . VI ] : = vlel + V2e2 + . The number of elements in a basis of a vector space is independent of the Theorem 2. If a basis X for a vector space V(Jf 0) has n elements. The number of elements in a basis of a vector space is independent of the particular basis considered. Then for all v E V there exists a unique ntuple {~I'.16. In Rn. For . .. components represents the vector v with respect to the basis B.. n unique... . e2.b.[ ~  ] + 4· [ ~ l To see this.. {el. The scalars {Ei}are called the components (or sometimes the coordinates) components coordinates) Definition 2. x ~ D J Definition 2.. In]Rn.18. . bn be a basis (with a specific order associated with the basis vectors) b1. n } such that for V. We represents B. {~i } of v with respect to the basis {b l . for]Rn [e\.
U\ + 1. where Eij is a matrix all of whose elements are 0 except for a 1 in the (i. V.. dim{A E ~nxn :: A = AT} = !n(n + 1). and because the 0 vector is in any vector space.) = 0 and ]P ft. s e S}. 2. determine 1/2n(n + 1) symmetric basis matrices. The union of two subspaces. Ra C V/or an arbitrary index set A). we define dim(O) = O.23. dim{A € Rnxn A AT} = {1/2(n 1 (To see why. Theorem 2. Thus. V is infinitedimensional. n 5 S. Example 2. R H S = {v : v e R and v e S}.a S. is not necessarily a subspace. Definition 2. 72. n S {r s : r E U. dim{A e Rnxn A is upper (lower) triangular} = !n(n 1)..20. t1]) . R. The sum and intersection Definition 2. The union of two subspaces. R = 0. 1.) (To see why. H. 5. Note: Check that a basis for ~mxn is given by the mn matrices Eij. Remark 2. ft n 5 = {v : v E 7^ and v E 5}. V for an arbitrary index set A). j E ~. dim(R mxn ) mn. for finite k).18 says that dim(V) = the number of elements in a basis. R C S. The collection of Eij matrices can be called the "natural basis matrices. and because the 0 vector is in any vector space. n n S = 0. y>f (L L . J)th location. 2. V is infinitedimensional. and 2. T = R 0 S is the direct sum of R and S if = REB S is the direct sum ofR and S if Definition 2. we define dim(O) = 0.=1 K k 1=1 2. dim(Rn)=n. dim(~n) = n.23. + 7^ =: L R. where Efj is a matrix all of whose elements are 0 except for a 1 in the (i. Sums and Intersections of Subspaces 2. for finite k). 2. and S are defined respectively by: 1. V (in general.21. tJJ) = +00. Remark 2.) 2 5. F) be a vector space and let R. U S.4. Thus. Let (V. Theorem 2. dim(C[to. 1. S.24. Theorem 2.2. a eA CiEA f] n *R. 2. j e n. 4. R S C V (in general. RI \ h Rk =: ]T ft/ C V. vector space V is finitedimensional if there exists a basis X with n < +00 elements. determine !n(n 1) symmetric basis matrices.18 says that dim (V) the number of elements in a basis.24.4 Sums and Intersections of Subspaces Subspaces Definition 2. R S = (in general.j matrices can be called the "natural basis matrices. dim(~mXn) = mn. R D S C V (in general. The sum and intersection ofR and S are defined respectively by: of R.21.4.= T). The subspaces Rand S are said to be complements of each other in T. s E 5}. U + S = T (in general ft. Example 2.20. R + S = {r + s : r e R. j)th location. . 2. 1. otherwise. 1.. . i E m. n (^ ft. i e m. V. S S. 2. A consistency. « The subspaces R. Theorem 2. otherwise. K + S S. is not necessarily a subspace. dim{A E ~nxn :: A is upper (lower) triangular} = 1/2n(n+ 1). Let (V." 3.22. S c V. V (in general. R j ) = 0 am/ Ri = T). A vector space V is finitedimensional if there exists a basis X with n < +00 elements. JF') be a vector space and let 71.22." The collection of E. and 1. and S are said to be complements of each other in T.4 2. Note: Check that a basis for Rmxn is given by the mn matrices Eij. Sums and Intersections of Subspaces 13 13 consistency.+00.
.25.27. v = [4 l]r jR2. 3. . . Then show that one of the vectors 1. we must have rl = r2 and s\ = si from S2 from which uniqueness follows. mutually [x\. . we must have r\ ri and SI rl . Let x\. . dim(R + S) = dim(R) + dim(S)  dim(R n S). jRnxn. R). Example 2.s\. every t € can be written uniquely in the form r s with r e R and s e S. S2 e rl Sl r2 Then r.c the Example 2.. Let U be the subspace of upper triangular matrices in E" x" and let £ be the subspace of lower triangUlar matrices in Rnxn. For arbitrary subspaces R. Since R n S = 0.r2 S2 . D Theorem 2. 1 TIT The first matrix on the righthand side above is in S while the second is in R. *2. 0 S. .. . Then it may be checked that U + . e jRnxn 4. Consider the vectors VI — [2 1f and V2 = [3 1f.14 14 Chapter 2. n Proof: A e jRnxn written Proof: This follows easily from the fact that any A E R"x" can be written in the form A=2:(A+A )+2:(AA). ft be the set of symmetric matrices in R" x ".26.29. For example. dim(T) = dim(R) + dim(S). ft S) = jR2 and let ft be any line through the origin. Since ft fl 0. where r1. where rl. . Vector Spaces Remark 2.2 and 2. ft. 0 The statement of the second part is a special case of the next theorem. Then any other distinct line through the origin is a complement of R..27. Suppose =R EB Then 1.si e S.20. while U n £ is the set of diagonal matrices in Rnxn... Suppose {VI. vd must be a linear combination of the others.. Xk} must be a linearly independent set. .. 2. But r1 –r2 E ft and S2 — SI E S. Vn thonormal if and only if A E R"x" is orthogonal. ft.... consider V = R2 unique. {vi. ..28..27.. validity of the formula given in Theorem 2. Theorem 2. r2 E Rand s1. triangular + L = R xn un..25. Let VI. S of a vector space V. XI. Find the components of the vector v = [4 If with respect to this basis.c jRnxn.vn be orthonormal vectors in R". . x/c E R" be nonzero mutually orthogonal vectors. Using the fact that dim {diagonal (diagonal matrices} = n... . . 2. Av" •. the set in jRnxn. which uniqueness follows. = dim(ft) + Proof: To Proof: To prove the first part. F) (R n x n .26.. Then any other distinct line through the origin is and let R be any line through the origin. Vector Spaces Chapter 2. one can easily verify the validity = n.r2 £ Rand 52 . Example 2. We discuss more about orthogonal complements elsewhere in the text.. The complement of R (or S) is not unique. EXERCISES EXERCISES 1.27. and let R"x".28. Vk} is a linearly dependent set. X2. But as t = r1 + s1 = r2 + S2. Then V = U $ S..... and let S Let (V. IF) = (jRnxn. r2 e R. let R be the set of skewsymmetric matrices in (V.29. .. . Show that {XI. S2 E S. AVn are orv\.20. Avn are also orjRn. Suppose T = R O S.r . Example 2. Among all the complements there is a unique one orthogonal to R. together with Examples 2. unique ft. every t E T can be written uniquely in the form tt = r + s with r E Rand s E S.5. Xk E jRn 2. Show that Av\. of the formula given in Theorem 2.c = jRnnxn jRn xn. 2.. Then Theorem 2. Then r1 — r2 = s2— SI. For arbitrary subspaces ft. Prove that viand V2 form a a basis Consider v\ = [2 l]r 1*2 = [3 l] Prove that VI and V2 form basis 2 for R . and SI. S of a vector space V. jRn xn . Xk} must be a linearly independent set. jR). Theorem 2.. suppose an arbitrary vector t E T can be written in two ways t e as t S2.
Exercises Exercises
15
5. Let denote the set of polynomials of degree less than or equal to two of the form 5. Let P denote the set of polynomials of degree less than or equal to two of the form Po + PI X + pix2, where Po, PI, p2 E R. Show that P is a vector space over R Show Po p\x P2x2, where po, p\, P2 e R Show that is a vector space over E. Show Find the components of the that the polynomials 1, *, and 2x2 — 1 are a basis for P. Find the components of the that the polynomials 1, x, and 2x2  1 are a basis for 2 2 with respect to this basis. polynomial 2 + 3x 4x polynomial 2 + 3x + 4x with respect to this basis.
6. Prove Theorem 2.22 (for the case of two subspaces Rand S only). 6. Prove Theorem 2.22 (for the case of two subspaces R and only).
7. Let n denote the vector space of polynomials of degree less than or equal to n, and of 7. Let Pn denote the vector space of polynomials of degree less than or equal to n, and of the form p ( x ) = Po + PIX + ...•+ Pnxn,, where the coefficients Pi are all real. Let PE po + p\x + • • + pnxn where the coefficients /?, are all real. Let PE the form p(x) denote the subspace of all even polynomials in Pn,, i.e., those that satisfy the property denote the subspace of all even polynomials in n i.e., those that satisfy the property p(—x} = p(x). Similarly, let PQ denote the subspace of all odd polynomials, i.e., p( x) = p(x). Similarly, let Po denote the subspace of all odd polynomials, i.e., those satisfying p(—x} = p(x). Show that Pn = PE EB Po· those satisfying p(x) = – p ( x ) . Show that n = PE © PO8. Repeat Example 2.28 using instead the two subspaces 7" of tridiagonal matrices and 8. Repeat Example 2.28 using instead the two subspaces T of tridiagonal matrices and U of upper triangular matrices. U of upper triangular matrices.
This page intentionally left blank This page intentionally left blank
Chapter 3 Chapter 3
Linear Transformations Linear Transformations
3.1 3.1
Definition and Examples Definition and Examples
definition of a linear (or function, We begin with the basic definition of a linear transformation (or linear map, linear function, or linear operator) between two vector spaces. or linear operator) between two vector spaces.
Let IF) and (W, IF) be vector spaces. Then I: : > a Definition 3.1. Let (V, F) and (W, F) be vector spaces. Then C : V + W is a linear transformation if and only if transformation if and only if I:(avi £(avi + {3V2) = aCv\ + {3I:V2 far all a, {3 e F andfor all v},v2e V. pv2) = al:vi fi£v2 for all a, £ ElF and far all VI, V2 E V. The vector space V is called the domain of the transformation C while VV, the space into called the of the transformation I: while W, the space into The vector space which it maps, is called the which it maps, is called the codomain.
Example 3.2. Example 3.2.
1. Let F = R and take V = W = PC[f0, +00). 1. Let IF JR and take V W PC[to, +00). Define I: : PC[to, +00) > PC[to, +00) by Define £ : PC[t0, +00) + PC[t0, +00) by
vet)
f+
wet) = (I:v)(t) =
11
to
e(tr)v(r) dr.
2. Let F = R and take V = W = JRmxn. Fix M e R m x m . Let IF JR and V W R mx ". Fix ME JRmxm. Define £ : JRmxn + M mxn by I: : R mx " > JRmxn by
X
f+
Y
= I:X = MX.
3. Let F = R and take V = P" = {p(x) = a0 + ct}x H ... + anx"n : a, E R} and ao alx + ai E JR} and 3. Let IF = JR and take V = pn (p(x) h anx W = pnl. w = pn1. I: : —> Define C.: V + W by I: p = p', where'I denotes differentiation with respect to x. Lp — p', where denotes differentiation x.
17
E ~} e m] V ith column of A = Mat £ (the matrix representation of £ with respect to the given bases = L L for V and W) is the representation of LVi with respect to {w j. then LVx = Lv = ~ILvI + . Note that A = Mat £ depends on the particular bases for V and W. When V = JR. Thus. j E raj. F) —>• (W. w m] and where W = [WI. Li near Transformations Chapters. in the notation. i e ~}.2 3. transformation with its matrix representation.. In other words. is by its action on a basis. + ~nLvn =~IWal+"'+~nWan = WAx. n A= al : ] E JR. but this is usually not done. Thinking of both as a matrix and as a linear transformation from JR.. ... IF) veniently in matrix form. Linear Transformations 3. suppose £ : (V. is arbitrary).e.. i.. j E !!!...n. usually L The action of £ on an arbitrary vector V e V is uniquely determined (by linearity) v E V uniquely determined by its action on a basis. w ] and L is the ith column of A. if V = E1v1 + . and hence jc. W = lR.2 Matrix Representation of Linear Transformations Matrix Representation of Linear Transformations Linear transformations between vector spaces with specific bases can be represented conLinear transformations between vector spaces with specific bases can be represented conSpecifically...• + E nVnn = V x (where u.. Thus. with respect to {w }•..m and {Vi.. Then the {w j. Specifically. When V = R".. We identify A the equation £V = W A becomes simply £ = A. {u.} are the usual (natural) bases. £V = W A since x was arbitrary.m usually causes no naturally confusion. Thus. z'th V This could be reflected by subscripts. i e n} and {Wj. [ w . i. for V and W) is the representation of £i>. j E m} are the usual (natural) bases WA linea LV L = A. W = R m and [ v i . LV WA since x was arbitrary.e. {W jj' j e !!!.." to Rm usually causes no Thinking of A both as a matrix and as a linear transformation from Rn to lR. if v = ~I VI + • • + ~n v = Vx (where v. L IF) ~ (W. respectively. Change of basis then corresponds naturally to appropriate matrix multiplication.} are bases for V and W.mxn a mn represents L since represents £ since LVi = aliwl =Wai. We thus commonly identify A as a linear transformation with its matrix representation. . Thus.. say. i E n}. In other words. then arbitrary). j e m}. and hence x. F) is linear and further suppose that {Vi.18 Chapter 3. + amiWm where W = [w\. + .
Two Special Cases: Two Special Inner Product: Let x. If dimU = p. That is. . Then we can define a new transformation C as follows: to W. If dimZ// = p. and dim W m. and W and transformations B from U to V and A from Wand V to W. in the same order in both the diagram and the equation.y. The above is sometimes expressed componentwise by the C — A B . we have C A B . . y e ~n. xx T XX ).=1 Outer Product: Let x e Rm. V. dimV = n. then composition of transformations corresponds to standard matrix multiplication. the form XXT (or xx HH). Composition of Transformations 19 19 3. then composition of transformations corresponds to standard matrix mUltiplication. That is. A rankone symmetric matrix can be written in above (or xy if A E C xyH e c ). dim V = n. and dim W = m.3. Inner Product: n xTy = Lx. Note that in most texts. Then their outer product is the m x n E ~m. Note that in The above diagram illustrates the composition of transformations C = AB. and if we associate matrices with the transformations in the usual way.3.3. Then we can define a new transformation C as follows: C The above diagram illustrates the composition of transformations C = AB.3 Composition of Transformations Composition Consider three vector spaces U. Outer Product: matrix matrix mxn E R Note that any rankone matrix A e ~mxn can be written in the form A = xyT = xyT H mxn mxn). and if we associate matrices with the transformations in the usual way. Then their inner product is the scalar E ~n. y E Rn. the arrows above are reversed as follows: C However. Composition ofTransformations 3. it might be useful to prefer the former since the transformations A and B appear in the same order in both the diagram and the equation. y e Rn.. expressed mxp nxp formula cij = L k=1 n aikbkj.
where 8ij is the Kronecker delta defined by Kronecker delta defined by 8 = {I0 ij ifi=j. is the set {w E W : w = Av for some v E V}. Equivalently.. Then Let A : V —>• be a linear transformation.2. an M. . Definition 3. . essentially following immediately from the defiProof: The proof of this theorem is easy. 1. Theorem 3.4 Structure of Linear Transformations Structure of Linear Transformations Let A : V —> W be a linear transformation. in general. . If A is written in terms of its columns as A = [ai.. V. is an orthonormal set. {[ ~~i ].." • orthonormal set. . Definition3. 3. be orthogonal if' vjvj 0 for i ^ j and orthonormal if vf vj 8ij' where 8tj is the be orthogonal if vr v j = 0 for i f= j and orthonormal if vr v j = 8ij. the same symbol (A) is Note that in Theorem and throughout the text.4.. Definition 3. denotedR(A). Li near Transformations 3. . D Remark 3. denoted N(A). ~}  ISisan 3.3. vk] be a set of nonzero vectors Vi E ~n.. Let A E Rmxn. See also the of Section 3. ~ 3 . is an orthogonal set. The nullspace of A. . usual (natural) bases. Then 1. {[ ~ J. .4 3. The set is said to 3.5 and throughout the text. See also the last paragraph of Section 3. 2. orthonormal set. then {I —/==. denotedlZ( A). The range of A. The nullspace of kernel of and A is also known as the kernel of A and denoted Ker (A). ~. subspaces of different spaces.. Theorem 3.. Let A : V + W be a linear transformation. . is the set {v E V : Av = O}... 2. Note that N(A) and R(A) are. [: J} is an orthogonal set. then Proof: The proof of this theorem is easy. of of denoted Im(A). N(A) c V. in general. R ( A ) S. I ~VI VI ^/v.i •.2. If {VI. 0 nition.. then ~ . Note that in Theorem 3. . Example 3. The range of A is also known as the image of A and — {Av e V}.•.. R(A) C W. 1. essentially following immediately from the definition. {v1. an]..7.8. If in of = [a\.. is the {v e V Av = 0}. LinearTransformations Chapter 3.5.. subspaces of different spaces. is an orthonormal set. e Rn. Note that N(A) and R(A) are. The nullspace of The of denoted N(A). IS an orthogonaI set. e ~mxn. € 1Tlln is an orthogonal set. Vk} with u. —/=== .8. Let {VI.. ..Vk } With Vi E. W..3.7.6.. (A). the same symbol (A) is used to denote both a linear transformation and its matrix representation with respect to the used to denote both a linear transformation and its matrix representation with respect to the usual (natural) bases.an]. then then R(A) = Sp{al. The range of A. { t > . vi ^/v'k vk ~~~ ] . R(A) = {Av : v E V}.[ :~~ J} . an} . 2.. vd of u. N(A) S. if i f= j. is the set {w e w = Av for some v e V}. Let A : V + be a linear transformation.IN.20 20 Chapter 3.. denoted Im(A). .
Structure of Li near Transformations 3. S \B S~ = ]Rn. S 5. Any set of vectors will do. . (n n S)~ = n~ + S~. . Example 3. = S.9. S1. n <. 3. Then the orthogonal complement of S is defined as the set c ]Rn. including dependent spanning vectors (which would. k =X . Set vector. 2. the computation involved is simply to find all nontrivial (i.3. Let {VI. Then n.e.. + X2 + X3 = 0..11.. The proofs of the other results are left as Proof: left exercises.. . Theorem 311 Let Theorem 3. Rn. Set XI X2 = L (xT Vi)Vi. . Let 3.4. then give rise to redundant equations). (n + S)~ = nl. nonzero) solutions of the system of equations 3xI 4xI + 5X2 + 7X3 = 0. Let S <. if and only if S~ <. vk} e ]Rn vector.= {v e Rn : vTs=OforallsES}.. Then it can be shown that Working from the definition. Vk} be an orthonormal basis for S and let x E Rn be an arbitrary {v1. Structure of Linear Transformations 21 21 Definition 3. ]Rn.10. Then the of defined T S~={VE]Rn: V S = 0 for all s e S}. n~. Proof: We prove and discuss only item 2 here. n S~. 6. of course. Note that there is nothing special about the two vectors in the basis defining S being orthogonal. 4. (S~)l.10. Let R S C Rn The S <..=1 XI.4.
l N(A T (i.13. +x~).5 Four Fundamental Subspaces Four Fundamental Subspaces Consider a general matrix A E lR.) 2. where XI. . . Clearly. Suppose.e. But then (x'1 —XI)TT (x. R" N(A) 0 ft(Ar ».e. x. Then {v E R" : Av = 0} is sometimes called the right nullspace of A.e. Vk and hence to any linear combination of these we see that X2 is orthogonal to VI. Ax = 0 if and only if x equivalent to yT Ax = 0 for all y... including itself) is 0. X2 is orthogonal to any vector in S. and x2 = x~. that x = XI for example. We S n S1 =0 the e orthogonal everything in (i. (Note: This also holds for infinitedimensional vector spaces.. the right nullspace is A/"(A) while the left nullspace is N(A T ).. In other words.l = 7£(AT).= Af(AT) ) (i. Thus. We also have that S U S. Linear Transformations Then x\ E <S and. Then R(A r ). Then X2 = x.e.e. But then (x. Similarly.. .1 in the next section. When thought of as a linear transformation from IR n to Rm. E S and X2..l = Rn. It is also easy to see directly that.l where x € M(A) and y € J\f(A)± = R(AT) (i. . X2 is orthogonal to any vector in S. i. x~ X2 = (x.. everything in S (i. Theorem 3. R(A). (Note: This also holds for infinitedimensional vector spaces. Then {v e IRn : A v = O} is sometimes called the Definition 3.XITVj =XTVjXTVj=O. XI) (which follows by rearranging the equation XI +X2 = x. But yT Ax = ( A T ) x. Theorem 3. including itself) is O.22 22 Chapter 3. we see that x2 is orthogonal to v1. the right nullspace is N(A) while the left nullspace is J\f(AT). But yT Ax = (ATyy{ x. since T x 2 Vj = XTVj . It can write vectors in a unique way with respect to the corresponding subspaces. Since x was arbitrary. x 1 E Sand x2.26. = x'1+ x'2. Suppose. standing Figure 3. D Definition 3..26.e.xn.) N(A)1" spaces. Clearly. we form AT v.x1) (x'1 xd x2 — X2 = — (x'1 — x1) (which follows by rearranging the equation x1+x2 = x'1 + x'2). 'R. Then T (x. (w e Rm : w T A = 0} is called the left nullspace of A. where x\. Ax = 0 if and only if x orthogonal is orthogonal to all vectors of the form AT y.12.l. for example. In other words.. x E R(A r ) Since x 1 have established thatN(A).13. Similarly.. We have thus shown that S + S. 2. (Note: This for finitedimensional 1. y.l = N(A ).l = 0 since the only vector s E S orthogonal to S1 = IRn. every vector in the codomain space R m can be written ina unique way asw = x+y. IRm = R(A) EBN(A T». D Theorem 3. where x e U(A) and y e ft(A)1.•. Thus. When thought of as a linear transformation from E" Consider a general matrix A € E^ x ". transformation A. Li near Transformations Chapters. N(A). . E N(A) and E N(A). established that N(A) U(AT ). Thus. We have thus shown that vectors. every vector w in the codomain space IRm can be written in a unique way as w = x+y. every vector v in the domain space R" can be written in a unique way as v = x + y. Let A : IRn —> Rm. Ax = Proof: To prove the first part. Vk and hence to any linear combination of these vectors. since Then XI e S and.12 and part 2 of Theorem 3. E R(A) and E R(A).. (Note: This holds only for finitedimensional vector spaces. {w E IR m : WT A = O} is called the left nullspace of A.12. 0 The proof of the second part is similar and is left as an exercise. Let A : Rn + IRm. See also Theorem 2.l.XI/ (x~ .14 (Decomposition Theorem). The proof of the second part is similar and is left as an exercise.11 can be combined to give two very fundamental decompositions damental and useful decompositions of vectors in the domain and codomain of a linear and transformation A. many properties of A can be developed in terms of the four fundamental subspaces . Rm = 7l(A) 0 M(AT)). that x = x1 + x2. x e R(AT). every vector v in the domain space IRn can be written in a unique way as v = x 7. can write vectors in a unique way with respect to the corresponding subspaces. Let A : R" + IRm.12 and part 2 of Theorem 3. Then Ax = 0 and this is an and equivalent to yT Ax = 0 for all v. XI = x. 0 x1 — x'1 andx2 = x2. IRn = M(A) EB R(A T)).X2) 0 since (x'1 — x1) (x' 2 — x2) = 0 by definition of ST. 3.e. Then Theorem 3. x'2 E S1. – x1) = 0 since 0 by definition of S. we decompositions. x~ e S.(A)1~ — J\f(ATT ).l. Then Theorem 3. take an arbitrary x e A/"(A). Let A : Rn + Rm.1 in the next section. Thus. + x~.) Proof: To x E N(A).14 (Decomposition Theorem). when we have such direct sum decompositions.l = R(A T}. right nullspace of A. Then 1.11 can be combined to give two very funTheorem 3. i.) 2. Let A : IRn > Rm. ft(Ar) (i.e. many properties of A can be developed in terms of the four fundamental subspaces to IRm.5 3. See also Theorem 2.. Let A : IRn > IRm.. This key theorem becomes very easy to remember by carefully studying and underThis key theorem becomes very easy to remember by carefully studying and understanding Figure 3.
1 makes many key properties seem almost N(A)T. The row rank of A is column rank of of independent row rank of . A is onetoone or 11 (also called monic or infective) ifJ\f(A) = 0. Let V and W be vector spaces and let A : V + W be a linear transforDefinition 3. and in illustrating concepts such as controllability and observability.5.16. t= V2 ===} AVI t= AV2 . mation. Four Fundamental Subspaces 3. rank(A) dimftCA).1 obvious and we return to this figure frequently both in the context of linear transformations obvious and we return to this figure frequently both in the context of linear transformations and in illustrating concepts such as controllability and observability. Figure 3.16. R(A)1. 1. Four fundamental subspaces. Let A : E" + Rm. This is sometimes called 3. Two equivalent characterizations of A being 11 that are often easier to verify in practice are the characterizations of A being 11 that are often easier to verify in practice are the following: following: (a) AVI = AV2 (b) VI ===} VI = V2 .15. A is onto (also called epic or surjective) ifR(A) = W.1.15. the column rank of A (maximum number of independent columns). Four Fundamental Subspaces 23 23 A r N(A)1 r EB {OJ X {O}Gl nr m r Figure 3.(A)^.(A) = W. 2. Definition 3. Figure 3. properties 7£(A). 3.1. A f ( A ) . Then rank(A) = dim R(A).3. 1. A is onto (also called epic or surjective) ifR. and N(A)1. N(A). Let and W be vector spaces and let A : motion. 'R. A is onetoone or 11 (also called monic or injective) if N(A) = O. R(A).5. Two equivalent 2. IR n > IRm. fundamental subspaces. be a linear transforDefinition 3.
Linear Transformations dim 7£(A r ) (maximum number of independent rows). . rank(A) + B) :s rank(A) + rank(B).. v r } is a basis for N(A)L. the subspaces themselves are not necessarily in the same vector space. if B is nonsingular. rank(AB) = rank(BA) = rank(A) and N(BA) = N(A). Proof: From Theorems 3. Let A : R" ~ Rm.") of A. where n is the ]Rn > ]Rm. Tv abasis 7?. (Note: Since 3. Finally. dimension of the domain of A.. + rank(B)  n :s rank(AB) :s min{rank(A). we include here a few miscellaneous results about ranks of sums completeness.17.andx22 e A/"(A). of A. Let A : Rn > Rm. O:s rank(A 2.11 and 3. The dual notion to rank is the nullity R(AT) of independent rows).19 suggests looking atthe general problem of the four fundamental subspaces of matrix products. this theorem is sometimes colloquially stated "row rank of A = column N(A)L = R(A A/^A) " = 7l(A ). . Then dimN(A) + dim R(A) = n." 0 of D The following corollary is immediate. + nullity(B).19. of A.. denoted nullity(A) or corank(A). and is defined as dim A/"(A). 3.17 we see immediately that Proof: From Theorems 3.(A). LinearTransformations Chapter3. . take any W e R(A).17. Like the theorem. . . e ]Rnxn. if {ui. 1 1 Xl E A/^A) . 3. sometimes denoted nullity(A) or corank(A).18. 1. Let A.. The last equality AXI x\ e N(A)L and jc E N(A). Theorem 3. shows that T is onto.17 we see immediately that n = dimN(A) = dimN(A) + dimN(A)L + dim R(A) . Then N(T) = To w E 7£(A). then {TVI. Tvrr]} is a basis for R(A). Theorem 3. {Tv\. r*i *i E N(A)L. rank(B)}. To see that T is also onto. dimA/"(A) + dimft(A) = dimension of the domain of A. Write x = Xl + X2. Then Ajti = W = TXI since Xl e A/^A).11 and 3. colloquially of = rank of A. . Then dim K(A) = dimNCA)L. it is a statement about equality of dimensions. .19 suggests looking at the general problem of the four fundamental Part 4 of Theorem 3. the following string of equalities follows easily: "column rank of A" = rank(A) = dim R(A) = dimN(A)L1 = dim R(AT) = rank(AT)) = A" rank(A) = dim7e(A) = dim A/^A) = dim7l(AT) = rank(A r = "column "row rank of A. dimA/'(A) ± (Note: 1 T T ).") Proof: Proof: Define a linear transformation T : N(A)L ~ R(A) by J\f(A)~L —>• 7£(A) by Tv = Av for all v E N(A)L. of Corollary 3. the subspaces themselves are not necessarily in the same vector space. nullity(B) :s nullity(AB) :s nullity(A) 4. where Ax — w. B E R" xn .. and is defined as dimN(A). by definition there is a vector x E ]Rn such that Ax = w. Clearly T is 11 (since A/"(T) = 0)..24 24 Chapter 3. 0 For completeness. R(A) : ]Rn ~ ]Rm. u. following follows we apply this and several previous results. Then 3. Part 4 of Theorem 3. ..19.(A) = dimA/^A^ 1 if that if {VI. . . The basic results are contained in the following easily proved following theorem.18.. iv} abasis forA/'CA) . x x e R" x\ X2. dimensions. and products of matrices. We thus have that dim R(A) = dimN(A)L since it is easily shown T dim7?. .
It is extremely useful in text that follows. Conversely. which implies x\ = x^. to have full column rank. e IRmxn. have full row rank. dim R(A) = m = rank(A). e 7?. N(A) = N(A T A). i. y E R(A). AA is nonsingular). A is onto if and only if rank (A) = m (A has linearly independent rows or is said to have full row rank. The transformations AT and A I have the same domain and range but are in general different maps unless A is and A~! have the same domain and range but are in general different maps unless A is orthogonal. The transformations AT are all 11 and onto between the two spaces N(A)1.17.—n = dim 7£(A r ). 4.5. Similar remarks apply to A and A~T. R(B T ).20. A is 11 if and only z/rank(A) = n (A has linearly independent columns or is said to have full column rank. Thus. Conversely. RCAB) S. linear least squares problems.2 N(A T ). It The next theorem is closely related to Theorem 3. N(AB) . Then y = Ax.20 and is also easily proved.22.e. since ArA is invertible. Let e IRmxn. especially when dealing with pseudoinverses and linear least squares problems. Let A E Rmxn. dim7?. 2. R«AB)T) S. Theorem 3.and R(A).17. Then 3. Also. A € IR~xn. R(AT) 3. A : IRn »• IR n is invertible or Note that if A is invertible. 2.22. XI = X2 AT A A 11. . AT A nonsingular). Proof' Proof of part 1: If A is onto. especially when dealing with pseudoinverses and is extremely useful in text that follows.20. 1. Let A : IRn + IRm. which implies that dimN(A)11 = n — Proof of part 2: If A is 11. suppose AXI = Ax^. A Proof of part 2: If A is 11. B E Rnxp. AATT is nonsingular). 1.(A). 4. Note that in the special case when A E R"x". and hence dim R(A) n by Theorem 3. N(A T ) = N(AA T ). suppose Ax\ dim R(A T). A : W1 + E" is invertible or nonsingular if and only z/rank(A) = n. which implies that dim A/^A). 3. A is 11 if and only ifrank(A) = n (A has linearly independent columns or is said 2. equivalently. then A/"(A) = 0. then dim V — dim W. 1. equivalently. 2. Then 3. x AT (AAT)I e IRn. We now characterize 11 and onto transformations and provide characterizations in We now characterize II and onto transformations and provide characterizations in terms of rank and invertibility. : R n » Rm.. AT. equivalently. let y E IRm be arbitrary. A is onto if and only //"rank(A) — m (A has linearly independent rows or is said to 1. Definition 3.20 and is also easily proved. the transformations A. Then A r A. Four Fundamental Subspaces Theorem 3. then N(A) = 0. and hence dim 7£(A) = n by Theorem 3. Let jc = AT(AAT)~]y Y E Rn. A : V + W is invertible (or bijective) if and only if it is 11 and onto. Conversely. D D 11.21. nonsingular ifand only ifrank(A) = n. then dim V = dim W. let y e Rm Proof: Proof of part 1: If A is onto. N«AB)T) . A"1 ± are all 11 and onto between the two spaces M(A) and 7£(A). AX2. Ar.(A) — m — rank (A). A A T. AT A is nonsingular). Conversely. and AI A.3. terms of rank and invertibility. A : V —» W is invertible (or bijective) if and only if it is 11 and onto.21. RCA). so A is onto.23.2 N(B). equivalently. Theorem 3. R(A) = R(AA T ). Definition 3. Also.ti = AT Ax2. Let A E Rmxn. A is AT AXI AT AX2. Four Fundamental Subspaces 3. Then Theorem 3. e IRnxp. = R(A T A). Note that if A is invertible.23.5. 25 25 The next theorem is closely related to Theorem 3.
It then follows from Theorem 3. Theorem 3. Similarly. (A R + A R A — /) must be a right inverse and.. by uniqueness it must be Thus. 2.R + AARA = I A +IA  A since AA R = I = I. In Chapter 6 we characterize all right inverses of a matrix by Chapter characterize characterizing all solutions of the linear matrix equation A R = I. Then Definition 3.R = If left A L A L A = 2. by uniqueness it must be A R + A R A — = A R.26.24. 0 a left inverse. where Iv denotes the identity transfonnation on V. Let + V.both 11 and is if and if onto.e. Let A : V > W. (Proof: Take any a E E1I.26 Chapter 3.: AA R = w Iw W + V such that AA~R = Iw. is left invertible if and if it and left invertible.e. A right invertible if and only if it onto. Theorem 3.24.25.L = (ATTA)I1AT.I) must be a right inverse and. Then A is onto. i. (A R + A RA .R = _~] (=1) and A~R = [ _j j is a right inverse. are infinitely A. Let A : V + Then 1. Let A : V + W. Obviously A has full row rank (= 1) and A . If there exists a unique left inverse A~L such that A~LA = I. Defileft If linear concepts left nitions of these concepts are followed by a theorem characterizing left and right invertible transformations. i.27. If there exists a unique right inverse A~R such that AA~R = I. i. then a right inverse is given by A~R = AT(AAT) I. right inverses for A. linear Transformations If a linear transformation is not invertible.22 we see that if A : E" + Em is onto.. 1. Let Theorem 3. where Iw denotes the identity transfonnation on W.. in A I = A R = A L.22 ]Rn >• ]Rm Note: From Theorem 3. it is clear that there are infinitely many right inverse. then A is invertible. i. A is said to be left invertible if there exists a left inverse transformation A~L : W —> to transformation A L : V such that A L A = Iv. . Let A : V » V. Notice the and leave the following: following: A(A.e. 1. D Example 3.R + ARA I) = AA.25 that A is invertible. Definition 3. if A is 11. A R the case that A~R + A~RA . If Proof: proof of second Proof: We prove the first part and leave the proof of the second to the reader. Let A = [1 2] : E2 »• E1I. A is right invertible if and only if it is onto. Also.. A is left invertible if and only ifit is 11.R AA. A is invertible if and only if it is both right and left invertible. Then > 1. A. A R A = I. can always find v e E2 such that [1 2][^] = a)..25 that A is invertible. therefore.26. characterizing all solutions of the linear matrix equation AR = I.I = A~R. it may still be right or left invertible. such that A~LA = Iv where Iv denotes the identity transformation on V. therefore. € ]R . 3. A is said to be right invertible if there exists a right inverse transformation A~RR : if A. that A~R is a left inverse.. then a left inverse is given by A R = AT (AAT) left T L = A. But this implies that A~RA = /. —> transformation if left + 2. It then follows from Theorem 3. both 11 and Moreover. Li near Transformations Chapters. then A is invertible. in which case A~l = A~R = A~L. Obviously A has full row rank can always find v E ]R2 such that [1 2][ ~~] = a). A~ (A A)~ A . 1.e. then one (Proo!' = [1 2]:]R2 + ]R ..
Consider the vector space R nx " over E.2. Y) = Tr(X Tr F). 2. let denote the subspace of symmetric 2. and let 7£ denote the subspace of skewsymmetric matrices. Prove Theorem 3. We give when considered as a linear on ]R3. LetA [J] : E1 ~ E2. II? Is £. y) = Tr(X Y). In Chapter 6 we characterize all left inverses of a matrix by characterizing all solutions of the linear matrix equation LA = I. respect to this inner product. J E2. It is now obvious that A has full column is v 0. Find the matrix representation of A with respect to the bases Find the matrix representation of A to bases {[lHHU]} of R3 and {[il[~J} of E . Y E Enx" define their inner product by (X.4. Then A is 11. consider A linear transformation ]R3 1. The matrix A = 1 1 2 1 [ 3 1 when considered as a linear transformation onIE \ is neither 11 nor onto. whence N(A) = 0 so A is 11). it is clear that there are A L = [3 — 1] infinitely many left inverses for A. For matrices X. Show that. Prove Theorem 3. In Chapter 6 we characterize all left inverses of a infinitely many left inverses for A. 3. Consider the vector space ]Rnxn over ]R. with Y e ]Rnxn (X. We give below bases for its four fundamental subspaces. Consider differentiation £ 11? Is£ onto? onto? 4. Let A = [i]:]Rl > ]R2. For matrices matrices. The matrix 3. 4. (Proof: The only solution to 0 = Av = [I2]v is v = 0. is neither 11 nor onto.3. matrix characterizing all solutions of the linear matrix equation LA = I. EXERCISES EXERCISES 3 4 1. below bases for its four fundamental subspaces. whence A/"(A) = 0 so A is 11). . Let A = [8 5 i) and consider A as a linear transformation mapping E3 to ]R2. 3. Again. 2 . — S^.4. £. Let A = [~ . It is now obvious that A has full column rank (=1) and A~L = [3 . Consider the differentiation operator C defined in Example 3. let S denote the subspace of symmetric matrices. 'R. R = S J. respect to this inner product.1] is a left inverse. and let R denote the subspace of skewsymmetric matrices. Is £. (Proof Theonly solution toO = Av = [i]v 2.Exercises 27 2. ThenAis 11.
Let A = [ J o]. Prove Theorem 3. Let = [~ 9. if not. Show that AT has a right inverse.4.11.12. left T Suppose e Rmxn 9. Modify Figure 3. Are they equal? Is this true in general? If this is true in general. Theorem 6. linearly independent solutions 10.Il. ~ ~ 3 8. Suppose A € Mg 9x48 . Rnxm thought of as a transformation from Rm to IRn.28 5. homogeneous linear system Ax = 0? homogeneous linear system Ax = O? n 3. Are they equal? Is this true in general? DetennineN(A) and R(A). Determine bases for the four fundamental subspaces of the matrix Detennine fundamental A=[~2 5 5 ~]. Linear Transformations 7. If E 1R~9X48. provide a counterexample.2.1 to illustrate the four fundamental subspaces associated with AT e associated ATE nxm IR from IR m R". prove it. Suppose A E IR m xn has a left inverse. 3. Prove Theorem 3. Chapter 3. Linear Transformations Chapters.1 11.4. Determine A/"(A) and 7£(A). How many linearly independent solutions can be found to the 10. .
1. Definition 4. as noted in the proof of Theorem 3. T is bijective (11 and onto). see [22].17.m We A+ A e lP1.Chapter 4 Chapter 4 Introduction to the Introduction to the MoorePenrose MoorePen rose Pseudoinverse Pseudoinverse In this chapter we give a brief introduction to the MoorePenrose pseudoinverse. define a transformation A+ : Y —»• X by Definition 4.1. Define a transformation T : Af(A)1. the MoorePenrose pseudoinverse of A. which was proved by Penrose in 1955. which was proved by Penrose in 1955. case X = W1 and Y = Rm. define a transformation A + y + X by where Y = YI + Yz with Yl e 7£(A) and Yz e Tl(A}L. where X Xand Y y are arbitrary finitedimensional vector spaces. This transformation T~ + can be used to give our first definition of A+. as noted in the proof of Theorem 3. for determining A+ . pseudoinverse of A.l.. We have thus defined A+ for all A E IR™xn.1 4. and hence we can RCA) —>• J\f(A}~L This transformation can define a unique inverse transformation Tl 1 :: 7£(A) + NCA). With A and T as defined above. brings great notational and conceptual clarity to the study of solutions to arbitrary systems of linear equations and linear least squares to the study of solutions to arbitrary systems of linear equations and linear least squares problems. Then A+ is the MoorePenrose j2 with y\ E RCA) and yi E RCA). 4." X ". Then. let us henceforth consider the X ~n lP1.17.+ R(A) by dimensional Define transformation T : N(A). see [22]. as is shown in the following text.l —>• Tl(A) by Tx = Ax for all x E NCA).1 Definitions and Characterizations Definitions and Characterizations Consider a linear transformation A : X —>• y. Then A+ is the MoorePenrose where y = y\ pseudoinverse of A. The MoorePenrose pseudoinverse is defined for any matrix and. as is shown in the following text. With A and T as defined above. brings great notational and conceptual clarity matrix and.. A purely algebraic y + characterization of A+ is given in the next theorem.l. the definition neither provides nor suggests a good computational strategy good computational strategy for determining A +. can be used to give our first definition of A . Although X and y were arbitrary vector spaces above. the MoorePenrose pseudoinverse of A.. problems. let us henceforth consider the Although X and Y were arbitrary vector spaces above.l. 29 . neither provides Unfortunately. a generalization of the inverse of a matrix. a generIn this chapter we give a brief introduction to the MoorePenrose pseudoinverse. characterization of A is given in the next theorem. and hence we Then. T is bijective Cll and onto). where and are arbitrary finiteConsider a linear transformation A : X + y.
Example 4.4.6. A + always exists and is unique. 19]. then by uniqueness. this can be found in [1.. If G the pseudoinverse of A. and (P4) but not (P3). Consider A = f ] satisfies (P1)(P4). (P2). as with Definition 4. it must be A +. if a t= 0. = Furthermore. (P2). Let A e R™xn. one need simply verify the four Penrose conditions (P1)(P4). Example 4.2 Examples Examples Each of the following can be derived or verified by using the above definitions or characEach of the following can be derived or verified by using the above definitions or characterizations.1) = limAT(AAT +8 2 1)1. While not generally suitable for computer implementation. Example 4. Introduction to the MoorePenrose Pseudoinverse Theorem 4. Let A E lR. as a right or left inverse satisfies no fewer than three of the four properties. one need simply verify the four Penrose conditions (P1)(P4). Let A E lR. then by uniqueness. A t = AT (AA )~ if A is onto (independent rows) (A is Example 4. the Penrose properties do offer the great virtue of providing a tional algorithm.2. While not generally suitable for computer implementation. Such a verification is often relatively straightforward. Unfortunately. Theorem 4.5.3. p. . the Penrose properties do offer the great virtue of providing a checkable criterion in the following sense." xn. this characterization can be useful for hand calculation of small examples. it must be A+. (P3) (AGf (P3) (AG)T = AG. characterization can be useful for hand calculation of small examples. Consider A = [']. L Note that other left inverses (for example.30 Chapter 4. If G satisfies all four. A+ = (AT A)~ AT if A is 11 (independent columns) (A is left invertible).1]) satisfy properties (PI). a right or left inverse satisfies no fewer than three of the four properties.4. (4. Example 4. Example 4.5. Then Theorem 4. Then G = A+ if and only if Theorem 4. Note that other left inverses (for example. whose proof Still another characterization of A + is given in the following theorem. A~ = [3 . Introduction to the MoorePenrose Pseudoinverse Chapter 4. X+ = AT(AATT) I if A is onto (independent rows) (A is right invertible). Still another characterization of A+ is given in the following theorem. 19]. Example 4.1. Given a matrix G that is a candidate for being checkable criterion in the following sense. whose proof can be found in [1.7.1." xn. AG. However. p. neither the statement of Theorem 4. straightforward. neither the statement of Theorem 4. Unfortunately. Verify directly that A+ = Example 4. Furthermore. However. Let A e R?xn Then G = A + if and only if (Pl) AGA = A. terizations.7. Given a matrix G that is a candidate for being the pseudoinverse of A. Also.2) 4.2 4.2. (P4) (GA)T = GA. Then A+ [a [! = lim (AT A + 82 1) I AT 6+0 6+0 (4. (P2) GAG = G. (P4) (GA)T = GA. For any scalar a.3.2 nor its proof suggests a computational algorithm. if a =0. (PI) AGA = A. Note that the inverse of a nonsingular matrix satisfies all four Penrose properties. (P2) GAG G.2 nor its proof suggests a computawith Definition 4.6. A+ always exists and is unique. Also. For any scalar a. Such a verification is often relatively satisfies all four. Verify directly that A+ = [ ~] satisfies (PI)(P4). Note that the inverse of a nonsingular matrix satisfies all four Penrose properties. A+ = (AT A)I AT if A is 11 (independent columns) (A is left invertible). and (P4) but not (P3). A L = [3 — 1]) satisfy properties (PI).
Theorem 4.3. Many of these are used in the text that follows. Many of these This section presents some miscellaneous useful results on pseudoinverses. where U is orthogonal an Theorem 4.9. e jRmxn and suppose Rmxm R n are orthogonal (M is T 1 1 orthogonal if M M ). Then S+ UD+U T where D+ is again a diagonal matrix whose diagonal elements are determined according to Example 4.13. .. Proof: Both results can be proved using the limit characterization of Theorem 4. where U is orthogonal and D is diagonal. 27].9. p.4.10.3. 27]. are used in the text that follows. (A T )+ = (A+{. 31 31 Example 4. For any vector v E M". Then S+ = U D+UT. D Theorem 4.7. simply verify that the expression above does indeed satisfy each of Proof: For the proof.4.8.7. The proof of the second result (which can also be proved easily by verifying the four Penrose proof of the second result (which can also be proved easily by verifying the four Penrose conditions) is as follows: conditions) is as follows: (A T )+ = lim (AA T ~+O + 82 l)IA = lim [AT(AAT ~+O + 82 l)1{ + 82 l)1{ 0 = [limAT(AAT ~+O = (A+{. The especially illuminating. Then orthogonal if MT = M. Then Proof: For the proof. The interested reader can consult the proof in [1.). Example 4. For A e Rmxn 1. if v i= 0.3 Properties and Applications Properties and Applications This section presents some miscellaneous useful results on pseudoinverses.12. Let S E Rnxn be symmetric with U TSU = D.4. Let A E R m x "and suppose UUEejRmxm. 2.8. Let S e jRnxn be symmetric with UT SU = D. where D+ is again a diagonal matrix whose diagonc D is diagonal. if v = O.3 4.10. 0 the four Penrose conditions. The proof of the first result is not particularly easy and does not even have the virtue of being proof of the first result is not particularly easy and does not even have the virtue of being especially illuminating. p. For all A E jRmxn. Properties and Applications 4. Example 4. Properties and Applications Example 4. 4. . elements are determined according to Example 4.13.11.11. Example 4.12. The interested reader can consult the proof in [1. [~ ~ r ~ =[ 0 Example 4. . [~ r 1 =[ 4 4 I I ~l 4 I I 4 4. Theorem 4. A+ = (AT A)+ AT = AT (AA T)+. The Proof: Both results can be proved using the limit characterization of Theorem 4. For any vector e jRn.VVEejRnxnx " are orthogonal (M is 4. simply verify that the expression above does indeed satisfy each c the four Penrose conditions.
13 we can.xm. see [9]. [7]. Theorem 4. 1.12 Note that by combining Theorems 4. [II]. B E Rrrxm. (A+)+ = A. (AB)+ = B?A+. For all A E lR mxn . we have B = B (BBT)~\ whence BB+ = Ir. whence A+A = f r. Then As an example consider A = [0 I] and B = LI. 3..g..15. in theory at least. since e lR~xr.At = A in Theorem 4. As an example consider A = [0 1J and B = [ : J.13 can.16. where BI = A+ AB and AI = AB\B+. Proof: Proof: For the proof.11 is suggestive of a "reverseorder" property for pseudoinverses of prodTheorem 4. 0 D E lR~xr. (AT A)+ = A+(A T)+. 0 The following theorem gives some additional useful properties of pseudoinverses. then (AB)+ = B+ A+.12 and 4. Theorem 4. Introduction to the MoorePenrose Pseudo inverse 4. necessary and sufficient conditions under which the reverseorder property does hold are known and we quote a couple of moderately useful results for reference.15. we B+ BT(BBT)I. 2. in general. Introduction to the MoorePenrose Pseudoinverse Chapter 4. Proof' A+ A Proof: Since A E Rnrxr. e. Proof: Proof: For the proof. and better methods are suggested in text that follows. If A is normal. If A e Rnrxr.16. (AB)+ = B+ A + if and only if 1. n(A T AB) ~ nCB) . A\ = A in Theorem 4. poor (see. where BI = A+AB and A) = ABIB{.17. D takingB t = B. Ir Similarly.. A+ = (AT A)I AT. however (see.15. The result then follows by E lR. [9]. then A+ = (ATA)~lAT. see [5]. (AA T )+ = (A T)+ A+. xm + T B e Wr . [5]. (AB)+ = B+A+ if and only if 4. e. = N(AA+) = N«AA T)+) = N(AA T) = N(A T). If e lR~xm. 0 D Theorem 4.17. n(A+) 4. N(A+) 5. = n(A T) = n(A+ A) = n(A TA). then AkA+ = A+ Ak and (Ak)+ = (A+)kforall integers k > O. in peneraK ucts of matrices such as exists for inverses of products.• Similarly. properties Theorem 4.15.11 nets of matrices such as exists for inverses of nroducts Unfortunately. 4. 4.14. TTnfortnnatelv. Then (AB)+ = 1+ = I while while B+ A+ = [~ ~J ~ = ~.g. . 4. [23]). n(BB T AT) ~ n(AT) and 2.32 Chapter 4. [] sufficient reverseorder However. (AB)+ = B+A+. [11]. For e Rmxn . BB+ f r The by taking BI = B. [7]. This A AT AT turns out to be a poor approach in finiteprecision arithmetic. the MoorePenrose pseudoinverse of any matrix (since AAT and AT A are symmetric). compute 4. (AB)+ = B{ Ai. Theorem 4.14.
18. Suppose A E Rnxp. (a) Prove or disprove that Prove or disprove that [~ (b) Prove or disprove that (b) Prove or disprove that AB D [~ B D r r=[ =[ A+ 0 A+ABD. ft(A+) ft(Ar 5. The next theorem is fundamental to facilitating a compact and unifying approach The next theorem is fundamental to facilitating a compact and unifying approach to studying the existence of solutions of (matrix) linear equations and linear least squares to studying the existence of solutions of (matrix) linear equations and linear least squares problems. prove that R(A+) = R(A T). or orthogonal. we have shown where one of the Penrose properties is used above. prove that 7£(A) = 7£(AA r ) using only definitions and elementary properties of the MoorePenrose pseudoinverse. 5 e JRn x m . prove that RCA) = R(AAT) using only definitions and elementary 3. For A E Rpxn and BE R mx ". Use Theorem 4. such as preceding but still be normal. whereupon there exists a vector x e IR m such that Bx = y. where one of the Penrose properties is used above. show that JV(A) C A/"(S) if and only if BA+ A = B. For A e Rmxn. € IRm xm D 6. Let A G M"xn.• 1 2 x. Y e R". A e IRPxn thatN(A) S.i l . b e R for scalars a. For example. show that (xyT)+ = (x Tx)+(yT y)++yxT. if A is symmetric.i D.4 to compute the pseudoinverse of U .i ]. Then Bx E H(B) S. = B. If jc.. Since x was arbitrary. skewsymmetric. then it is normal. RCA). whereupon y = Bx = AA+Bx E R(A). For A e R m x n . However. and D E E mxm and suppose further that D is nonsingular. assume that AA+B = B and take arbitrary y e K(B). assume that AA + B To prove the converse. y E IRn. such as A=[ b a a b] for scalars a.4 to compute the pseudoinverse of \ 2 1. R(A) and take arbitrary x e IRm. Since x was arbitrary. For example. U(A) if and only if AA+B = B. A E IRmxn. a matrix can be none of the skewsymmetric. 0 EXERCISES EXERCISES 1. we have shown that B = AA+B. (xyT)+ = (xTx)+(yTy) yx T 3. 4. Then we have there exists a vector y e IRP such that Ay = Bx. that B = AA+ B. N(B) and 5 € IRmxn. Use Theorem 4. For A E IRmxn. b E E. properties of the MoorePenrose pseudoinverse.1 D. A E IRn xn B E E n xm 6. Then R(B) c R(A) if and only if Suppose e IRnxp. A+ 0 A+BD. However. 2. Then B and take arbitrary y E R(B). Proof: Suppose K(B) S.]. then it is normal. To prove the converse. e IRnxm. Then we have Bx = Ay = AA + Ay = AA + Bx. problems. Theorem 4. Then there exists a vector x E Rm such that Bx = y. or orthogonal. if A is symmetric. Then Bx e R(B) c H(A).Exercises 33 Note: Recall that A E IRn xn is normal if A A = A T A. so Proof: Suppose R(B) c U(A) and take arbitrary jc E Rm. fiA+A B. Note: Recall that A e R" xn is normal if AATT = AT A. so there exists a vector y E Rp such that Ay = Bx. a matrix can be none of the preceding but still be normal. Then K(B) S. B E E M X m .
This page intentionally left blank This page intentionally left blank .
The SVD plays a key conceptual and computational of this important matrix factorization.. < min{m. ... Ch.Chapter 5 Chapter 5 Introduction to the Singular Introduction to the Singular Value Decomposition Value Decomposition In this chapter we give a brief introduction to the singular value decomposition (SVD).e. . ... [24.2) (5.and postmultiplying by equality following from the orthonormality of the r. y22 €E Rnxfor^ and the 0JM^/ocJb in E~are compatibly IRnx(nr). More S > o r > O. Let {Vi. The SVD plays a key conceptual and computational role throughout (numerical) linear algebra and its applications. Ch. .1) rxr A = [U I U2) [ ~ 0 0 ][ ] 2 T VI VT (5.LettingSS = diag(uI. the latter VfV^S2 = S2. e IRmxm and V E IR nxn such that V € Rnxn such that UI > diag(ul. Proof: Since A r A > 00 A r A i is symmetric and nonnegative definite...] . 5. 6]). vn]. recall. its eigenvalues are all real and nonnegative.Vn ].u )... Proof: Since AT A ::::: ( (AT A s symmetric and nonnegative definite. 6]). . . V2 = [Vr+I.4) 35 .1 The Fundamental Theorem Theorem A Theorem 5. (Note: The rest of the proof follows analogously if we start with the observation that AAT > 0 and the details are left to the reader analogously if we start with the observation that A A T ::::: 0 and the details are left to the reader as an exercise. rcfr)..) Denote the set of eigenvalues of AT A by {U?. Let {u.. recall.o>) E IRrxr. its eigenvalues are all real and nonnegative. we have n = U~VT. ii E !!. (Note: The rest of the proof follows [24. for example.... r write A r A VI = ViS2.. ..we can Vi = [vr+ . We show that every matrix has an SVD and describe some useful properties and applications show that every matrix has an SVD and describe some useful properties and applications of this important matrix factorization. Premultiplying by Vf gives Vf A T A VI write ATAVi = VI S2. Vi eE RIRnxr.. . . . . Then there exist orthogonal matrices U E Rmxm and E IR~xn.. the latter equality following from the orthonormality of the Vi vectors.e. i. role throughout (numerical) linear algebra and its applications..• ::::: Urr > as an exercise. VI «xr j V U2 e ^x(mr) . i e !!. S = diagfcri..1. specifically. . dimensioned.. «}).. U\ E IRmxr. More where ~ = [~ specifically. and a\ > • • • > Ur > 0. i.• = Un. . Theorem 5. We In this chapter we give a brief introduction to the singular value decomposition (SVD).Vv r).. n}). e n} be a set of corresponding orthonormal eigenvectors 0= Ur+1 = . for example. vectors. Pre.) Denote the set of eigenvalues of AT A by {of / E n} with a\ > • • > a > 0 = o>+i = • • an.} be a set of corresponding orthonormal eigenvectors and let VI = [v\. (5. Let A e R™ x n . where S = [J °0]. .3) = Ulsvt· The submatrix sizes are all determined by r (which must be S min{m. Premultiplying by vt gives vt ATAVi = vt VI S2 = S2.} with UI ::::: . UI e Wnxr. U2 E IRrnx(mrl.\. .1.. we can and let V\ [VI. . u r ) e R . and the Osubblocks in are compatibly dimensioned. Preand postmultiplying by SI gives the emotion S~l eives the equation (5. Letting — diag(cri.
. Remark 5. The !:ingular value decomposition is not unique. . ...1 we see that ai(A) = A(2 (ATA) = £(A).(AATT). Remark 5.2. Introduction to the Singular Value Decomposition Chapter 5. .5.e. responding to multiple cr/'s. . i... } for R" {VI. we see that U{ AV\ = S and 1/2 A VI = U^UiS = 0.. whence Vi ATAV22 = O.)..36 36 Chapter 5. A = t/E VT SVD of A 5. there may be nonuniqueness associated with the columns of V\ (and hence U\) cor• there may be nonuniqueness associated with the columns of VI (and hence VI) corresponding to multiple O'i'S.3. Remark 5. we see that Mat C is "i:..r zero singular values.? (AA I min{m.4) see UfU\ = columns of U\ are orthonormal. The decomposition is A = V"i:. Definition 5. except for Hermitian transposes replacing transposes.16. Remark 5. 0 Definition 5. From the proof of Theorem 5. Then T rewriting A = V"i:. VT as AV = V"i:..1. matrix VI E M mx/ " by U\ = AViS~l.. to be ~ completes the proof. u ]} for Rm (see the discussion in Section 3. we see that V r A VI = since A V2 = O. Specifically. vn } for IR and {u\. Ui e (5. . Now V20 Vf A T A V 0. an examination decomposition Remark 5. Thus.(AT A) = is denoted ~(A). For example. The columns of V are called the right singular vectors of A (and are the orthonormal right singular vectors of of called orthonormal eigenvectors of AT1A).2). . AV2 = 0. .1 reveals that proof Theorem any orthonormal basis for jV(A) can be used for V2 • £lny orthonormal basis for N(A) can be used for V2. The columns of V are called the left singular vectors of A (and are the orthonormal called orthonormal columns ofU left singular vectors of eigenvectors of AA ). •. VTAV = [~ Q].denote A thought of as aalinear transformation mapping IR n to IRm. and codomain spaces with respect to which A then has a diagonal matrix representation. Then Specifically. Referring to the equation V I = A VI SI defining U\.4. The set {ai. VT be an SVD of A as in Theorem 5. Remark 5. . let C. an we Turning now to the eigenvalue equations corresponding to the eigenvalues ar+l.16. eigenvectors of AA TT). cr.. See also m Remark 5. of A A). an we have that ATAV2z = VzO = 0. analogous complex e IC~ xn straightforward. n] — Note that there are also min{m.. with respect to the bases A = U^V as A V Mat £ is S the U E we see respect n and {u I. Note that V and V can be interpreted as changes of basis in both the domain Remark 5.4. . m for IR (see the discussion in Section 3. Introduction to the Singular Value Decomposition Turning now to the eigenvalue equations corresponding to the eigenvalues or+\. Choose U2 £ IRmx(mr) [U\ U2] orthogonal.'. V H. Then T V AV =[ =[ VrAVI VIAVI VrAVI vIA VI Vr AVz vI AVz ] ~] since A V2 =0.4) we see that VrVI = /.. Then from (5. and defining this matrix U\ andU UT A V [Q ~]. ar}} is called the set of (nonzero) singular values of the matrix A and called [a\.2. Choose any matrix V2 E ^ 77IX( ™~ r) such that [VI V2] is VI columns orthonormal.. .. n} . the IRmxr VI AViSI. D to be S completes the proof. singular unique. Now define the ATA V AV O... The latter equality follows from the orthogonality of the columns of VI and V 2. in fact. 3. . The analogous complex case in which A E C™ x " is quite straightforward. where V and V are unitary and the proof is essentially proof decomposition A = t/E V H. The latter equality follows from the orthogonality of the S and vI AVi = vI VI S = O.3. C denote A thought of as linear transformation mapping W to W. of values of i I proof of A.. U be interpreted changes domain and codomain spaces with respect to which A then has a diagonal matrix representation.. Let A = V"i:..(A) At.. 1. Thus.2). Referring to the equation U\ = A V\ S l defining VI. of the proof of Theorem 5. U V identical.1. A.5. singular 2. we see that... See also m [v\.
is the matrix I: and the span of the columns of UI. i.6. SVD of A. Computing AA AATT is numerically poor in finiteprecision arithmetic. orthogonal transformations..sine sin e cose J[~ ~J[ cose sine Sine] cose ' where e is arbitrary. and E U\. too. Example 5. A _ [ 1  0 ~ ] cose = [ ..3). Let A e IRnxn" be symmetric and positive definite.. The Fundamental Theorem 37 37 • any U22can be used so long as [U I U2] is orthogonal.2) can always be constructed from a "compact SVD" (5. V2 (see Theorem 5. AT A Remark 5. then A A. SVD" (5. is an SVD.[1 0 ] .3).1. that aa"full SVD" (5. f/2. VT A V = A = VAV eigenvectors > 0.U I U T .. however.2) can always be constructed from ¥2 Theorem too.e.1. [7]. that "full SVD" (5. The Fundamental Theorem 5. is an SVD. U arbitrary 2 x orthogonal 5. see. if A = UI:VT is an SVD of A. VI:TU/ s n SVD of A VS C .5.11). VI.. Let V be an orthogonal 5.8. Note. corner f/E V T r r T Ti isaanS V D o f AT.8. Computing an SVD by working directly with the eigenproblem for AT A or 5. U2.9. A factorization UI: VT of a n m x n matrix A qualifies as an SVD if U and V are A t/SV r o f an m n U orthogonal and I: is an m x n "diagonal" matrix whose diagonal elements in the upper £ left comer are positive (and ordered). e. VT AV = A > O. Example A . Example 5.10. 01  where U is an arbitrary 2x2 2 orthogonal matrix.10. [11]. i. [25]. [11]. What is unique. Then A = V A VTT is an A. F/vamnlp 5. [25]. e j8 the case).g. For example.e.7. symmetric V orthogonal matrix of eigenvectors that diagonalizes A.9. A=U is an SVD. 0 Example 5. Vi. n=[ [] 3 2 I 3 2 I 5 2y'5 y'5 S 4y'5 15 2~ ][ 3~ 0 0 0][ 0 0 3 0 _y'5 3 v'2 T v'2 T v'2 T v'2 2 ] 3 2 2 = 3 3 3J2 [~ ~] A E R MX Example 5. see. C/ [U\ Ui] orthogonal. Better algorithms exist that work directly on A via a sequence of orthogonal transformations.6.g. [7]. e. U V form • columns of U and V can be changed (in tandem) by sign (or multiplier of the form e je in the complex case). SVDof A.
14.6) and (5. Part 4 of the above theorem provides a numerically superior method for finding (orthonormal) bases for the four fundamental subspaces compared to methods based finding (orthonormal) bases for the four fundamental subspaces compared to methods based on.. [HI.5) as a sum of outer products Remark 5...13. A = UZV. nicely in Figure 5.7) explain why it is conventional to write the SVD and the key vector relations (5. .1.11.2 Some Basic Properties Some Basic Properties Theorem 5. the following properties hold: the notation of Theorem 5.. u r ]. Remark 5.£V.. (c) R(VI) = N(A)1. Part 4 of the above theorem provides a numerically superior method for Remark 5.12. Introduction to the Singular Value Decomposition Chapter 5. reduction to row or column echelon form. (d) R(V2) = N(A) = R(AT)1.. vr ]. . . The elegance of the dyadic decomposition (5. i=1 (5.£VTT as in Let A e jRmxn have a singular value decomposition A = UHV as in Theorem Theorem 5. vn]. Then A has the dyadic (or outer 2. . . . Note that each subspace requires knowledge of the rank r. Let V = [UI. 2. the following properties hold: 1. . The singular vectors satisfy the relations 3. say. . for example.. as A = UZV rather than. . VI = [VI. Let A e jRrnxn have a singular value decomposition A = VLV T Using Theorem 5.1.. The relationship to the four fundamental subspaces is summarized knowledge of the rank r.8) where where . . rank(A) = r = the number of nonzero singular values of A.5) as a sum of outer products and the key vector relations (5. . Let A E E mx " have a singular value decomposition A = U.13. 1. Vn].1.. for example. U2 = [Ur+I. say. The elegance of the dyadic decomposition (5. Using the notation of Theorem 5.12. A = U.]. (5.1. urn].. ..6) (5. Then TheoremS. reduction to row or column echelon form. (b) R(U2) = R(A)1. 4.6) and (5.. The relationship to the four fundamental subspaces is summarized nicely in Figure 5. The singular vectors satisfy the relations AVi = ajui. . Then (5. .. Theorem 5. . LetUI = [UI.7) explain why it is conventional to write the SVD as A = U'£VTT rather than. Then A has the dyadic (or outer product) expansion product) expansion r A = Laiuiv. Let U =.11.38 38 Chapter 5..2 5..7) AT Uj = aivi for i E r. vn]. . Note that each subspace requires on. andV2 = [Vr+I. Let A E Rmxn have a singular value decomposition A = U'£ VT. rank(A) = r = the number of nonzero singular values of A. urn] and V = [v\.. Introduction to the Singular Value Decomposition 5. Then (a) R(VI) = R(A) = N(A T / . um] and V = [VI.= R(A T ). .. Remark 5.5) 3.1.= N(A T ).
Remark 5.. a simple if we insist that the singular values be ordered from largest to smallest.=1 U. e\\.u.15. then = L r 1 v.2. SVD and the four fundamental subspaces. Figure 5. Furthermore.11. SVD and the four fundamental subspaces. if we let the columns of U and V with the Qsubblocks appropriately sized.. .15.5. a simple reordering accomplishes the task: reordering accomplishes the task: (5. 0 D (5. which is clearly orthogonal and symmetric. Proof: The proof follows easily by verifying the four Penrose conditions..2.11. e^. Furthermore.1.10) . which is clearly orthogonal and symmetric.11) This can also be written in matrix terms by using the socalled reverseorder identity matrix This can also be written in matrix terms by using the socalled reverseorder identity matrix (or exchange matrix) P = \err. However. ed. However. Some Basic Properties 5. with the Osubblocks appropriately sized. Note that none of the expressions above quite qualifies as an SVD of A + Remark 5.. then be as defined in Theorem 5.er^\. . Proof' The proof follows easily by verifying the four Penrose conditions. (or exchange matrix) P = [e erI.. Note that none of the expressions above quite qualifies as an SVD of A+ if we insist that the singular values be ordered from largest to smallest. Some Basic Properties 39 39 A r r E9 {O} / {O)<!l nr mr Figure 5. if we let the columns of U and V be as defined in Theorem 5. . e2... .1.
/ E~. Then Let A E lR...olumn transformations. Then VT A = :EVT = [~ ~ ] [ ~i ] D _ . since {UI.. . postmultiplication of A by V is an orthogonal transformation column rank.mxn. In other words. since [u\. Both compressions are analogous to the socalled rowreduced echelon form which. premultiplication of A by UT is an orthogonal transformation that rank. Recall the linear transformation T used in the proof of Theorem 3. then T~ canbedefinedbyTIu. . postmultiplication of A by V is an orthogonal transformation that "compresses" A by column transformations. . . the matrix representation for T with respect to the bases {VI. . From Section 3.. is not generally as reliable a procedure. From Section 3.. r x has full row rank..16. Remark 5.1). and since ( v \ .. Then Again. Similarly.3 5. then T can be defined by TVj = OjUj . and since {VI.. the isabasisfor7£(. u is clearly matrix representation for T with respect to the bases { v \ . Since T is determined by its action on a basis. premultiplication of A by VT is an orthogonal transformation that "compresses" A by row transformations. while the matrix representation for the inverse linear transformation T~ with respect to the same bases is SI.vvr}}is aa is r basisforN(A). then T can be defined by TV.M(UT Notice that M(A) = N(V T A) = N(svr> and the matrix SVf E Rrxll" has full row A/"(SV. finiteprecision arithmetic.16..1. In other words. Both compressions are analogous to the socalled rowreduced where R is upper triangular. Such a compression is analogous to the "compresses" A by I.17 and in Definition 4. mxr has full column rank. Such a row compression can also be accomplished "compresses" A by row transformations. basis forJ\f(A)±..17 and Remark 5. vrr} and {u I.. Then AV = V:E = [VI U2] [~ ~ ] =[VIS 0] ElR. Notice that N(A) . then TlI can be defined by T^'M. Similarly. v } and {MI .1.2. . A "full SVD" can be similarly constructed. have an SVD given by (5. . . Introduction to the Singular Value Decomposition Chapter 5.urr}} e r.40 40 Then Then Chapters. = tv. = cr. u is a basis forR(A). Column compression Column compression Again. .1). . let A E lR.4). Since T is determined by its action on a basis. when derived by a Gaussian elimination algorithm implemented in finiteprecision arithmetic. = ^u. by orthogonal row transformations performed directly on A to reduce it to the form [~]. notice that R(A) R(A V) This time.i / E~. while the matrix representation for the inverse linear transformation TlI with respect to S. the same bases is 5""1.11).l.1). . is not generally as reliable a procedure. This time. ..2.. 5. notice that H(A) = K(AV) = R(UI S) and the matrix UiS e Rm xr has full K(UiS) and the matrix VI S E lR. .[ SVr ] 0 mxn E lR. w. Such a compression is analogous to the .3 Rowand Column Compressions Row and Column Compressions Row compression Let A E R have an SVD given by (5. when derived by a Gaussian elimination algorithm implemented in echelon form which..1).r) and the matrix SVr e lR. .. . urr]} is clearly S.11). .. let A e Rmxn have an SVD given by (5. A "full SVD" can be similarly constructed.. Introduction to the Singular Value Decomposition A+ = (VI p)(PS1 p)(PVr) is the matrix version of (5. is the matrix version of (5.i e r. in Definition 4. Such a row compression can also be accomplished by orthogonal row transformations performed directly on A to reduce it to the form 0 . .. In other words. In other words. where R is upper triangular.mxn have an SVD given by (5.. Recall the linear transformation T used in the proof of Theorem 3.
y E ~n Determine A e ~~ 4. Prove Theorem 5. Determine an SVD of A. for example. If XTX = 0. [11].1 starting from the observation that AAT ~ O. Note: this is analogous to the polar form where Q is orthogonal and P = PT > 0.e. of defined by A defined by A = xyT. [25]. y e Rn be nonzero vectors. xyT 5. For details. [7]. € IRmxn. EXERCISES EXERCISES 1. which is not generally a reliable procedure when performed by Gauss transformations in finiteprecision arithmetic. Let A E ~mxn and W E IR mxm and 7 E ~nxn are (a) Show that A and WAY have the same singular values (and hence the same rank). see. [25]. Determine an SVD of the matrix A E R™ xn E IRm. Prove Theorem 5. A E IRnxn indefinite. which is not generally a reliable procedure when socalled columnreduced echelon form..e. Note: this is analogous to the polar form iO z = rel&ofaa complex scalar z (where i = j = V^T). an SVD A. If XT X = 2. [23].. Determine SVDs of the matrices 5. z of complex scalar z (where i j J=I). For details. Use the SVD to determine a where Q is orthogonal and P p T > O. Do A Wand Yare A and WAY have the same singular values? Do they have the same rank? and WAY have the same singular values? Do they have the same rank? factorization of i. Let A e E"xn be symmetric but indefinite. Let X E M mx ".1 starting from the observation that AAT > 0. Let E ~~xn. Use the SVD to determine a polar factorization of A. A = Q P 7.Exercises Exercises 41 41 socalled columnreduced echelon form. see.[11]. (b) Suppose that W and Y are nonsingular but not necessarily orthogonal. Determine SVDs of the matrices (a) (b) [ ] [ ~l 1 0 1 6. Let A e Rmxn and suppose W eRmxm and Y e Rnxn are orthogonal. show that X = 0. for performed by Gauss transformations in finiteprecision arithmetic. A = QP 7. 3. 4. Let x e Rm. [23]. . = o. 2. [7]. i. (a) Show that and W A F have the same singular values (and hence the same rank). Let A € R" X M ..
This page intentionally left blank This page intentionally left blank .
43 .3) if and only b E R(A).3) for all b E lRm if and only ifR(A) = lRm. and this is possible only ifm > n. There exists at most one solution to (6. as a special case. n}). General linear systems of the form (6.e.1 6. 4. Consider the system of linear equations Theorem 6.e. equivalently.e. (6. (6. b]) = rank(A). i. A/"(A) = 0.1. There exists a unique to (6.2) 6. 4. (6.1. 6. onto. A are linearly independent. 5.3) for all b e W1 if and only if the columns of A are linearly independent. A G M m x m and A has neither a singular value nor a eigenvalue. there exists a solution if and only ifrank([A. n. i. equivalently.3) for all b e W" if and only if is nonsingular.. 2. There exists a unique solution to (6. i. and onto. 1. A solution to (6.3) if and only ififbeH(A). Consider the system of linear equations Ax = b.3} for e R m if only ifU(A) = W". and this is possible only ifm ::: n. b E lRm. i. N(A) 0.3) for all b E ]Rm if and only if A is nonsingular.3) is unique if and only if N(A) = 0.e.1) are studied and include. b E ]Rn. A is 11.. n this is possible only ifm < n (since m dimT^(A) = rank(A) < min{m. only if rank(A) < n.3) for all b E lRm if and only if the columns of 5. i.3) 1. There exists a solution to (6.3) is unique if and only ifJ\f(A) = 0. 3. A E lRmxn.. 3. and this is possible only ifm :::: n (since m = dim R(A) = rank(A) :::: min{m. There exists a solution to (6. A solution to (6.. b]) = rank(A). A is 11. equivalently. i. there exists a solution if and only j/"rank([A. A E ]Rn xn. A E lR mxm and A has neither a 0 singular value nor a 0 eigenvalue. There exists a solution to (6.1 Vector Linear Equations Vector Linear Equations We begin with a review of some of the principal results associated with vector linear systems. There exists a nontrivial solution to the homogeneous system Ax = 0 if and only if Ax = 0 if 6.e. Theorem 6. equivalently. rank(A) < n.e. There exists at most one solution to (6. A is 2. General linear systems of the form equations. We begin with a review of some of the principal results associated with vector linear systems. the familiar vector system Ax = b. as a special case.Chapter 6 Chapter 6 Linear Equations Linear Equations In this chapter we examine existence and uniqueness of solutions of systems of linear In this chapter we examine existence and uniqueness of solutions of systems of linear equations... the familiar vector system are studied and include.
(6.mxn. i. A is not 11. AA+B B.nxk is arbitrary. all solutions of (6. Linear Equations Proof: The proofs are straightforward and can be consulted in standard texts on linear Proof: The proofs are straightforward and can be consulted in standard texts on linear algebra. A E JR.6) Furthermore. while results for (6.44 Chapter 6. to algebra.mxk.4) has a solution if and only ifl^(B) C 7£(A).3.A+ A)Y. mxn .6) are of this form.e.mxk and suppose that AA+B = B. A is not II. Let Z be an arbitrary solution of That all solutions arc of this form can be seen as follows. AZ — B. Proof: To verify that (6.e .6). we prove part 6. AZ :::: B.1 follow from those below for the special case = 1.2 6. Note that some parts of the theorem follow directly from others. we must have the case of a nonunique solution. i.e.5) is a solution. Linear Equations Chapter 6.5) is a solution of is a solution of AX=B. i. 0 Theorem 6. note that x 0 is always a solution to the homogeneous system. Therefore. Proof: The subspace inclusion criterion follows essentially from the definition of the range Proof: The subspace inclusion criterion follows essentially from the definition of the range of a matrix. to prove part 6.3.. premultiply by A: AX = AA+ B + A(I = B A+ A)Y + (A  AA+ A)Y by hypothesis = B since AA + A = A by the first Penrose condition. which implies rank(A) < n must have the case of a nonunique solution. (6..2 (Existence). Theorem 6.2)follow by 6. Then we can write Z=A+AZ+(IA+A)Z =A+B+(IA+A)Z and this is clearly of the form (6. Furthermore. equivalently. a solution exists if and only if AA+B = B. and this is clearly of the form (6. For example. Let Z be an arbitrary solution of (6.2 (Existence).5) is a solution.6). all solutions of (6. Note that the results of Theorem 6. Note that some parts of the theorem follow directly from others. The matrix criterion is Theorem 4. The matrix linear equation Theorem 6. E JR. while results for (6. Note that the results of Theorem of solutions to the general matrix linear system (6.e. where Y E JR. a solution exists if and only if has a solution if and only ifR(B) S.2) follow by specializing even further to the case m = n. premultiply by A: Proof: To verify that (6. which implies rank(A) < n by part 0 by part 3.18. equivalently. The matrix linear equation AX = B. of a matrix.6) are of this form...1). specializing even further to the case m = n. Then any matrix eRmxk of the form of the form X = A+ B + (/ . i. B E JR.5). R(A). D 6. note that x = 0 is always a solution to the homogeneous system.1). Let A e Rmxn. Therefore. BE JR. Then we can write (6. For example. (6.5). That all solutions are of this form can be seen as follows.1 follow from those below for the special case k = 1. +B = Theorem 6. The matrix criterion is Theorem 4.2 Matrix Linear Equations In this section we present some of the principal results concerning existence and uniqueness In this section we present some of the principal results concerning existence and uniqueness of solutions to the general matrix linear system (6.18. 0 .
A+A) = O. Example 6.9.6) that minimizes TrX7 (Tr() denotes the trace of a matrix. wherer = rank(A) (recallr ::: h). equivalently.7) has a unique solution if and only if M(A) = 0. Computation: Since y is arbitrary. equivalently. Example 6. Clearly.7) has a unique solution if and only if unique if and only if A + A = I.5. A A+ AI Remark (/ — A + A) 0..6) +B Remark 6. But rank(A) = n that A+ A = / if r — n.5. this can occur if and only if rank(A) = r = m (since r ::: m) and this is equivalent to A being onto (A + is then a right inverse).4. Solution: x=A+O+(IA+A)y = (IA+A)y. BE lR.7.6 (Uniqueness). Proof: Proof: The first equivalence is immediate from Theorem 6. recall that TrXT X = £\ •xlj.A+ A)Y =A++(IA+A)Y. A+ = A"1 and so (I .2. But if A has an SVD given by A = U h VT.mxn. A solution of the matrix linear equation Theorem 6. it is not unique. Butrank(A) = n if and only if A is 11 or N(A) = 0. in which case A must be invertible and R = AI. Matrix Linear Equations 45 Remark 6.7) is unique if and only if A+A = /. equivalently.S. Matrix Linear Equations 6. Here. nonzero solution.mxn. X• = A~ B. leaving only the unique solution X = AI1B.) that minimizes TrXT X. All right inverses r < m) A (A+ of A are then of the form of A R = A+ 1m + (In . there is no "arbitrary" component. Clearly.7. N(A) = O. When A is square and nonsingular. if there exists a unique. this can occur if and only if rank(A) = r m (since equivalent to AA+Im = 1m. y E R" A + A t= I.nxm is arbitrary. A solution of the matrix linear equation AX = B. where y e lR. (TrO denotes the trace of a matrix. then it is easily R(I — A + A). (6. Consider the system of linear firstorder difference equations (6. A A = f/E VT. where r rank(A) (recall r < n). there exists a nonzero solution if and only if A+A /= I. Thus.8) .9. Characterize all right inverses of a matrix A E lR. recall that TrX r = Li. Clearly.) Theorem 6. vD Example 6. A e E"x".mxk (6. D 0 Example 6. we write 1m to emphasize the m x m identity Im matrix.A+A). There is a unique right inverse if and only if A+A = I/ e E"xm arbitrary. rank(A) = < A This is equivalent to either rank (A) = r < n or A being singular. r checked that 1. find all A e ]Rmx". The second follows by noting thatA+A = I can occur only ifr = n.n is arbitrary. 7£(A) and this is 7£(/m) c R(A) equivalent to AA + 1m Im. Example 6. Solution: There exists a right inverse if and only if R(Im) S. Find all solutions of the homogeneous system Ax = 0.6. A E lR. Hence.A+A = Vz V2 and R(Vz2V^) = R(Vz) = N(A). Ax — 0. A+ A where Y E lR. Suppose A E lR. A R (AA(A) = A"1.A+ A V2 V[ and U(V = K(V2) = N(A). it is easy to see that all solutions are generated y from a basis for 7£(7 . Remark 6.6 (Uniqueness).j jcj. (6. matrix. Consider Example 6. It particular (6. It can be shown that the particular solution X = A+B is the solution of (6. and (N(A) = 0).nxn. Characterize AR = Im solutions R of the equation AR = 1m.8. if and only if A is Ilor _/V(A) = O. / .3.2.
Again from Theorem 6.2. AB. We now introduce an output vector yk to the system (6. B) is if(AT . we of reachability.. this is a question va [Uj }k~:b such that x^ takes an arbitrary value in W ? In linear system theory. from the fundamental Existence Theorem. (A. equivalently. this is a question {u }y~Q Xk of reacbability. The matrices A = [~ ~]1and B5 == [~] 1 providean example of a system that is controllable but not reachable.:b dual to reachability is called observability: When does knowledge of {" j }"!Q and {y_/}"~o suffice to determine (uniquely) Jt0? As a dual to controllability. from the fundamental Existence Theorem. if and only if rank [B.46 46 Equations Chapter 6. The answers are cast in terms that are dual in the linear algebra sense as well.J B]) = 1R" or. m known as the state vector at time while Uk is the input (control) vector. equivalently.J B] = n.8) of Example 6. The vector Jt* in linear system theory is e IR nx " fieR" (n ~ I. we have the notion of reconstructibility: When does knowledge of {u jy }"~Q and {.T. linear differential equations). if and only if or. this is called controllability. The general known as the state vector at time k while Uk is the input (control) vector. There are many other algebraically equivalent conditions. The condition The answers are cast in terms that are dual in the linear algebra sense as well.9) ~Axo+[B.8) is given by kJ Xk = Akxo + LAkJj BUj j=O UkJ ] Uk2 (6.e. Since m ~ I.. A n . does there exist an input sequence {u j an input sequence {"y}"~o such that xn = O? In linear system theory. We might now ask the question: Given Xo 0.10. Since > 1.10) for k ~ 1.ra>l).8) of Example 6. The above are standard conditions with analogues for continuoustime models (i. We can then pose some new questions about the overall system that are dual in the systemtheoretic sense to reachability and controllability.8) is given by solution of (6. B T] is observable [reconsrrucrible] [controllablcl if and T) observable [reconstructive]. reachability always implies controllability and. controlA 1 lability and reachability are equivalent. We now introduce an output vector Yk to the system (6. We might now ask the question: Given XQ = 0.. The linear differential equations).. AB.y/}"Io suffice to determine reconstructibility: When does knowledge of {w r/:b and {YJ lj:b suffice to determine (uniquely) xn? The fundamental duality result from linear system theory is the following: (uniquely) xnl The fundamental duality result from linear system theory is the following: E RPxn e IR pxn E RPxm € IR pxm (A. The condition dual to reachability is called observability: When does knowledge of {u 7 r/:b and {Yj l'. . B) iJ reachable [controllable] ifand only if (A . The general solution of (6. see that (6..~ I). A n .• A k kJ B] [ ~o (6. Example 6. There are many other algebraically equivalent conditions. overall system that are dual in the systemtheoretic sense to reachability and controllability.2.8) is controllable if and only if if controllability.AB •. . this is called such that Xn = 0? linear system theory. . we see that (6. we have the notion of suffice to determine (uniquely) xo? As a dual to controllability. example of a system that is controllable but not reachable. Theorem 6. We can then pose some new questions about the with C and (p > 1)..10. does there exist an input sequence {ujj 1jj^ such that Xk takes an arbitrary value in 1R"? In linear system theory.9 Example 6. The matrices A = [ ° Q and f ^ provide an lability and reachability are equivalent.11) with and D (p ~ 1).. standard conditions with analogues for continuoustime models (i. . if A is nonsingular. A related question is the following: Given an arbitrary initial vector Xo. Linear Equations Xk with A E R"xn and B E IR nxmxm(rc>l.8) is reachable if and only if if R([ B.. does there exA related question is the following: Given an arbitrary initial vector XQ.9 by appending the equation by appending the equation (6. does there exist an input sequence for k > 1. Theorem l'/:b Clearly.e.
v E R(R)..14) requires the notion A compact matrix criterion for uniqueness of solutions to (6. if and only if or. Let A E Rmxn. Listed In many applications.15) E jRnxp where Y € Rn*p is arbitrary. by definition. arbitrary. Theorem 6. B E Rmxq. the solution is then unique if and only if N(R) ==0.6. Theorem 6. Verification of each identity is recommended as an exercise for the reader. by definition. Invertibility is assumed for any component or subblock whose inverse is indicated. the solution is then unique if and only if N(R) Uniqueness Theorem. equivalently.13) Let v denote the (known) vector on the lefthand side of (6.Du] (6. in which case the general solution is of the form (6. By the fundamental Uniqueness Theorem. C E jRmxn.27.13) and let R denote the matrix on Let denote the (known) vector on the lefthand side of (6. Thus. 0. so a solution exists. particularly for block matrices. mxm and D E jRm Invertibility is assumed for any component or subblock whose inverse is and D € E xm.11. or. A E Rnxn.CBuo .14) Theorem 6. asbelow is a small collection of useful matrix identities. Then the equation e jRmxn.6. the coefficient matrices of interest are square and nonsingular. the has a solution if and only if AA+BC+C = B. e Tl(R). particularly for block matrices. e Rmxn. 6.4 6.27.14) requires the notion of the Kronecker product of matrices for its statement. Such a criterion (CC+ <g) A+ A — I) is stated and proved in Theorem 13. if and only if r Yn]  Lj:~ CA n . Such a criterion (C C+ ® A +A = I) of the Kronecker product of matrices for its statement. and C E Rpxti. sociated e jRnxn. E jRnxm. +L kl CAk1j BUj + DUk.6.12) j=O Yo . . notice that Yk = CAkxo Thus.Duo Yl .4 Some Useful and Interesting Inverses 47 To derive a condition for observability.2 j BUj . equivalently. associated with matrix inverses. Then.3 A More General Matrix Linear Equation A More General Matrix Linear Equation AXC=B (6. Verification of each identity is recommended as an exercise for the reader. B E Rnxm. so a solution exists. (6. Then.DUnl 6. In these identities.3 6. the coefficient matrices of interest are square and nonsingular. indicated.4 Some Useful and Interesting Inverses Some Useful and Interesting Inverses In many applications. By the fundamental the righthand side.13) and let denote the matrix on the righthand side. in which case the general solution is of the has a solution if and only if AA + BC+C = B. notice that To derive a condition for observability. B e jRmx q . and C e jRpxq.4 Some Useful and Interesting Inverses 6. is stated and proved in Theorem 13. A compact matrix criterion for uniqueness of solutions to (6. Listed below is a small collection of useful matrix identities.
l = l = [!C / [~ ~ l = [ AI +_~~!~CAI A~BE = D.CA.c E E") that arise in optimization (A + xx T ) — (with symmetric A e lRnxn and x e lRn) that arise in optimization theory. 1. Assuming 2. This result follows easily from the block UL factorwhere F = (A — ED C) This result follows easily from the block UL factorization in property 17 of Section 1. Linear Equations 1. for example. that X~l [~ !/ [~ ~ r [~ ~ l [~ ~/ r [~ ~ 1 l l l = [ ~ 4. This result follows easily from the block LU factorization in property 16 of Section 1.BDI l = [ AI BD. (A BDCr1 = AI . Let A E lRmxn. result follows easily from the block LU factorization in property 16 of Section 1. 5. [~ ~ r l 3.4. characterize all left inverses of a matrix A E Mm xn .I EXERCISES EXERCISES 1.1. (A + BDC)I = A~l . 1. mx . BC 6.B D.I B)I (E is the inverse of the Schur complement of A).I C) I.I ..A~lB(D~ CA~lB)~[CA~l This result is known as the ShermanMorrisonWoodbury formula. It also the inverse of a sum of matrices such as (A + D)"1 or (A" + D"1) It also yields very efficient "updating" or "downdating" formulas in expressions such as yields very efficient "updating" or "downdating" formulas in expressions such as T (A + JUT ) I1 (with symmetric A E R"x" and . BB EelR fflxk and suppose AAhas an SVD as in Theorem 5.I ] D. As in Example 6. Note that the positions of the / and . formulas for the inverse of a sum of matrices such as (A + D)lor (AI1 + DI)I.8. for example. Linear Equations Chapter 6.. 2.AIB(DlI + CAIB)ICAI. As in Example 6. [ / +c 7. formulas for applications (and is frequently "rediscovered") including./ blocks may be exchanged.8. X. characterize all solutions of the matrix linear equation AX=B in terms of the SVD of A in terms of the SVD of A. Let A € E mx ". theory. 2. where F = (A .. It has many This result is known as the ShermanMorrisonWoodbury formula. Note that the positions of the / and — / blocks may be exchanged. Assuming R(B) ~ R(A). r A~I [~ ~ r [D~I~AI D~I 1 ~r ~~B 1 r l [~ ~ r [D~CF +~~I~. This where E = (D — CA B) (E is the inverse of the Schur complement of A). Rmxk and suppose has an SVD as in Theorem 5. It has many applications (and is frequently "rediscovered") including. characterize all left inverses of a matrix A e lR ".1.4.4.4.48 Chapter 6. ization in property 17 of Section 1. l 8. = = Both of these matrices satisfy the matrix equation X2 = / from which it is obvious these matrices satisfy the matrix equation X^ = I from which it is obvious Both of that XI = X. characterize all solutions of the matrix linear equation 7Z(B) c 7£(A). where E = (D .
in Example 6.. Let x. T 4. y e IRn and suppose further that x T y ^ 1..Cn and individual elements Yij.10. 6. . Show that the matrix B = A .. Show that the matrix B — A — —eie T : (i. 5. Let jc. check directly condition for reconstructibility the form form N[ fA J CA n 1 ~ N(A n ). € IRn and suppose that x T y i= 1. A with — subtracted from its (zy)th element) is singular. c and individual elements y.. Let A e R"xxn and let A"1 have columns c\. l' Hint: Show that Ci E N(B). Assume that x/( 7^ 0 for some and j. where C = 1/(1 .Exercises Exercises 3..10.. As in Example 6.y Assume that Yji i= 0 for some i/ and j. . Show that 3. A with yl subtracted from its (ij)th element) is singular. . y E E" and suppose further that XTy i= 1.xy) T 1 49 = I  1 xTy 1 xy .l ~i e.e. Let A E 1R~ " and let A 1 have columns Cl. y E E" and suppose further that XTy ^ 1. check directly that the condition for reconstructibility takes the 6..e.x xTy). Show that (/ . (i. Show that 4.e.. Let x. . Hint: Show that ct <= M(B). Show that cxJ C ' where c 1/(1 — T y).
This page intentionally left blank This page intentionally left blank .
2. Figure 7. By Theorem 2.2. say on X along Y (using the notation of Definition 7. Define pX.26. PX. Infact.x = I px.yV = x for all v E V.1 displays the projection of von both and 3^ in the case = Figure 7. Let V be a vector space with V X 0 y.y.e. Inner Product Projections.1 displays the projection of v on both X and Y in the case V = ]R2. Theorem 7. P isaprojectionifandonlyifl P isaprojection.1.yp2 = P. y x Figure 7. Oblique projections.3..1 7. every v E V Definition 7.1). Let V be a vector space with V = X EEl Y. Theorem 7. Infact. Oblique projections.y.1.y is called the (oblique) projection on X along 3^.y • V —>• c V has a unique decomposition v = x + y with x e X and y E y.1. By Theorem 2.x — I — Px. Define PX y : V + X <.26.y is called the (oblique) projection on X along y. Figure 7. Also.1.y is linear and P# y — px. A linear transformation P is a projection if and only if it is idempotent. px. every v e V has a unique decomposition v x y with x E and y e y. i. y = Px. and Norms Spaces. Inner Product Spaces. P2 = P.. say on X along y (using the notation of Definition 7.1 Projections Definition 7. Proof: Suppose P is a projection. Theorem 7. Py. Proof: Suppose P is a projection. P is a projection if and only if I —P is a projection. and Norms 7.1). V by by PX.e.y Theorem 7. 51 51 . Px. Px. Also. Py.3. i.y is linear and pl.Chapter 7 Chapter 7 Projections. A linear transformation P is a projection if and only if it is idempotent.
.P)x = xTP(I . we have ( P y f I (/ .P)x E ft(P)1 xTPT(I . i=r+l PN(A) 1. Thus. Py e X. Moreover. Then x T pT (I .=1 m PR(A).3.P is the projection on Y along X.. First note that v E X. then Pv = 0.P}x = 0. P Proof: Let P be an orthogonal projection (on X. Now let v E V be arbitrary.A+A V2V{ L i=r+l i=l 11 ViVf.XL iss called an orthogonal projection and we then use the notation PX = PX. PN(A)J. with the second equality following since PTP is symmetric. Write x = Px (I — P)x. In the special case where y X1.P)v. Since x and y were arbitrary.1 and 5.P2v = 0 so Y E y. and Norms Let v e V be arbitrary.52 52 Chapter 7.P)v. and P must be an orthogonal projection. We now prove and Y = {v E V : Pv = OJ. Inner Product Spaces. * called an orthogonal projecDefinition 7. then v = O. p 2v = PPv = Let u E V be arbitrary. Moreover. while Py = P(I P}v = x Pv . Then v if v € Pv (I .P)x = O.P)v. Conversely. Now let u e V be arbitrary. along XXL} and let x. Then v = Pv + (I . Thus.XL Theorem 7. Px E R(P). then Pv v.xJ. say. First note that iftfveX.11. y = (I .X^X = Px±.P)x = yTTpT (I ..11. Conversely. .xl. Let x = Pv. we must have P (I — P) = O. along 1) and let jc. We now prove that V = X $ y. Thus. since Px e U(P). Then Px = p 2v = Pv = x so x e X. y = (I . D 0 7.5.P) = 0.xx by Theorem 7.1 .P)x 6 R(P)1and P must be an orthogonal projection.P)x e X1. Hence PT = PTP = P. Hence if v E X ny. Hence that V X 0 y. suppose p2 = P. let A E Rmxn with SVD A = U!:VTT = A = UT. Let X = {v E V : Pv = v} Px = x = Pv. then (/ . and Norms Chapter 7. we must have pT (I . Thus.L 1. mental subspaces. Then symmetric projection matrix and let x be arbitrary.V Theorems 5. P e jRnxn is the matrix of an orthogonal projection (onto R(P)) if and only ifP2 PT if p2 = p = pT.. It is easy to check that X and Y are subspaces. Since Py E X. Inner Product Spaces.uT. suppose P = P. It is easy to check that X and 3^ are subspaces. px. Hence pT = pT P = P. while Py = P(l . P = P. say. . Then Px = P2v = Pv = x so x E X. R" be Proof: Let P be an orthogonal projection (on X. PX.P)x = (I .P)x E XL. A 6 jRmxII UtSVf.1 The four fundamental orthogonal projections The four fundamental orthogonal projections Using the notation of Theorems 5. then v = 0. p2 = P. suppose symmetric projection matrix and let x be arbitrary.P)x. Let X n y. Note that (I . If v e y. (I . T Since x and y were arbitrary.1 7. V = X $ Y and the projection on X along Y is P. V Pv .XLtion and we then use the notation P x = PX.P)x = (I .5. Then Pv = P(x + y) = Px = x.P)x = x T P(l . Then Pv = P(x + y) = Px = x. P E E"xn is the matrix of an orthogonal projection (onto K(P)} if and only 7.p 2 v 0 so y e Thus. P)x = y PT(I P)x = 0. arbitrary.px. D Essentially the same argument shows that I — P is the projection on y along X. 0 Definition 7. suppose P is a is a with the second equality following since pT P is symmetric. Projections. Note that (/ .P)v.AA+ U2 U ! LUiUT. Write x = P x + (I . A+A VIV{ r LViVT are easily checked to be (unique) orthogonal projections onto the respective four fundaare easily checked to be (unique) orthogonal projections onto the respective four fundamental subspaces.1 5. Thus. yy Ee jR" be arbitrary. we have (py)T ((I . Let X = {v e V : Pv = v} and y {v € V : Pv 0}.xx by Theorem 7. Projections.)x = PXJ.4. (I .1.PX. Then U\SVr Then r PR(A) AA+ U\U[ Lu. If v E Y. Conversely. then Pv = O. Conversely. In the special case where Y = X^.4. P2v = P Pv — 2 2 Px = x = Pv. Thus. then Pv = v.P)x = O.P)v = = Pv. X 0 y and the projection on X along y is P.3. Essentially the same argument shows that / .
Solution: Think of the vector w as an element of the onedimensional subspace R( w). The indicated direct sum decompositions of the domain E" and codomain IRm are given easily as follows. Vk} was an orthomormal basis for a subset S of W1. There.. .1. Then Let x e W be an arbitrary vector. An arbitrary vector x e IRn was chosen and a formula for x\ appeared rather mysteriously.6.8) = (WTV) W. Recall the proof of Theorem 3. Projections 7.. Example 7. orthogonal: v z Pv w Figure 7." Figure 7. the vector z that is orthogonal to w and such that Pv Moreover. Recall the diagram of the four fundamental subspaces.8. . Determine the orthogonal projection of a vector v E IR n on another nonzero vector w E IRn. T W W Moreover.7. . See Figure 7. . Orthogonal projection on a "line. X on Specifically. Specifically.2. Orthogonal projection on a "line. The expression for x\ is simply the orthogonal projection of XI projection of rather x on S. in fact.2. orthogonal: that z and u. are.11.11..A+ A)x 2 = A+ Ax + (I = VI vt x + V Vi x (recall VVT = I). A direct calculation shows z = Pn(w)"' = (l .1. Then the desired projection is simply Then the desired projection is simply Pn(w)v = ww+v wwTv (using Example 4. IR n Rm 1 n Let X E IR be an arbitrary vector. the vector z that is orthogonal to wand such that v = P v + z is given by z is given by z = PK(W)±Vv = (/ — PK(W))V = v — (^^ j w.7. Projections 53 Example 7. {VI.Pn(w»v = v . Then X = PN(A)u + PN(A)X . There. The indicated direct Example 7. Recall the proof of Theorem 3.(:... An arbitrary vector x E R" was chosen and a formula for XI basis for a subset of IRn. in fact. Recall the diagram of the four fundamental subspaces. e Rn Solution: Think of the vector w as an element of the onedimensional subspace IZ(w)." Example 7. Determine the orthogonal projection of a vector e M" on another nonzero Example 7.6. Vk} was an orthornormal Example 7.2. { v \ .~) w. A direct calculation shows that and ware.7.8) (using Example 4. See Figure 7..8.2.
(A . definite defines Definition 7. Projections. as follows: and a vector in J\f(A). then AT e Rn xm is the unique linear transformation or map T E IRm andfor IRn.11. ATE IR nxm transformation Definition 7. Then (x. 3. (x.12. Then {^. only ifx = O. Then ('.AA+)y = U1Ur y + U2U[ Y (recall UU T = I). (x. {*. . Ay) = {AT x. Then Y = PR(A)Y + PR(A)~Y = AA+y + ( l .10. y) Q = X T Qy. Then { • • V x V if product if 1. Let V be a vector space over R. let y e IR m be an arbitrary vector. defines a "weighted" inner product. Then (x. y) x T Y is the "usual" Euclidean inner product or Example 7. y\) + /3(jt. y)Q = XT Qy. Let Example 7. e R. V = IRn.9. (x. Let V be a vector space over IR. Example 7. Let V = IRn. and Norms Chapter 7. > Ofor all E V ( x x) =0 if 2. Yl. Y2) for all x. y} = XTy is the "usual" Euclidean inner product or dot product.12. [ 5~2 + 7. Y2 E V and/or all a.11. and Norms Similarly. Projections.10. . y) = (y. respectively. f3ftE IR. let Y E ]Rm be an arbitrary vector.54 Chapter 7. If e IR mx ". 3. 2.. cryi + ^2) = a(x. y) for all x € Rm and for all y e R". j2 ^ V and for alia.2 Inner Product Inner Product Spaces Definition 7. (x. such that {x.13. y e V.9. yi.) ) :: V x V + IR is a real inner is a real inner Definition 7. where Q = QT > 0 is an arbitrary Q = Q T > is an Example 7. x) ::: Qfor aU x 6V and (x. Inner Product Spaces. y) = (y. . If A E Rm xn. Let V = R". Then Similarly.13. Yl) + f3(x. y^} for all jc. Let V = E". respectively. (x. n x n positive definite matrix. x } = 0 if and only ifx = 0. as follows: o o 4] uniquely into the sum of a vector in N(A)L 4V uniquely into the sum of a vector in A/'CA)1 r 1/4 1/4 ] 1/4 1/4 [!]~ = = A' Ax + (l  A' A)x 1/2 1/2 1/2 1/2 0] [ 2] [ 1/2 1/2 + [ 1~2 1~2 ~ o o ! 5/2] [1/2] 1~2 . Inner Product Spaces.x)forallx. Example 7.y E V. Let Then Then and we can decompose the vector [2 3 and we can decompose the vector [2 3 and a vector in N(A). aYI + PY2) = a(x. (jc. x) for all x. Example 7.
7.2. Inner product Spaces 7.2. Inner Product Spaces
55 55
It is easy to check that, with this more "abstract" definition of transpose, and if the It is easy to check that, with this more "abstract" definition of transpose, and if the (i, j)th element of A is aij, then the (i, j)th element of AT is ap. It can also be checked (/, y)th element of A is a(;, then the (i, y)th element of AT is a/,. It can also be checked that all the usual properties of the transpose hold, such as (Afl) = BT AT. However, the that all the usual properties of the transpose hold, such as (AB) = BT AT. However, the
definition above allows us to extend the concept of transpose to the case of weighted inner definition above allows us to extend the concept of transpose to the case of weighted inner products in the following way. Suppose A e Rmxn and let (., .) Q and (•, .) R, , with Q and A E ]Rm xn (., }R with Q and {, }g R positive definite, be weighted inner products on Rm and W, respectively. Then we can positive definite, be weighted inner products on IR m and IRn, respectively. Then we can define the "weighted transpose" A # as the unique map that satisfies define the "weighted transpose" A# as the unique map that satisfies
(x, AY)Q = (A#x, y)R all x e IRm (x, Ay)Q = (A#x, Y)R for all x E Rm and for all Y E W1. y e IRn.
By Example 7.l2 above, we must then have x T QAy x T (A#{ Ry for all x, y. Hence we By Example 7.12 above, we must then have XT QAy = xT(A#) Ry for all x, y. Hence we transposes (of AT Q = RA#. must have QA = (A#{ R. Taking transposes (of the usual variety) gives AT Q = RA#. QA = (A#) R. Since R is nonsingular, we find Since R is nonsingular, we find
A# = R1A Q. A* = /r'A' TQ.
We can also generalize the notion of orthogonality (x T = 0) to Q orthogonality (Q is We can also generalize the notion of orthogonality (xTyy = 0) to Qorthogonality (Q is a positive definite matrix). Two vectors x, y E IRn are Qorthogonal (or conjugate with a positive definite matrix). Two vectors x, y e W are <2orthogonal (or conjugate with T X Qy O. Qorthogonality is an important tool used in respect to Q) if ( x y) Q respect to Q) if (x,, y } Q = XT Qy = 0. Q orthogonality is an important tool used in studying conjugate direction methods in optimization theory. studying conjugate direction methods in optimization theory. Definition 7.14. Let V be a vector space over C. Then (., •} : V V > Definition 7.14. Let V be a vector space over <C. Then {, .) : V x V + C is a complex is a complex inner product if inner product if
1. (x,, x ) :::: Qfor all x e V and ( x , x ) = 0 if and only if x = 0. 1. ( x x) > 0 for all x E V and (x, x) =0 if and only ifx = O.
2. (x, y) = (y, x) for all x, y E V. (y, x) for all x, y e V. 2. (x, y)
3. (x, aYI + fiy2) = a(x, y\) + fi(x, Y2) for all x, YI, y2 E V and for alia, f3 6 C. 3. (x,ayi f3Y2) = a(x, yll f3(x, y2}forallx, y\, Y2 e V andfor all a, ft E c. Remark 7.15. We could use the notation (., ·)e to denote a complex inner product, but Remark 7.15. We could use the notation {•, }c to denote a complex inner product, but if the vectors involved are complexvalued, the complex inner product is to be understood. if the vectors involved are complexvalued, the complex inner product is to be understood. Note, too, from part 2 of the definition, that (x, x) must be real for all x. Note, too, from part 2 of the definition, that ( x , x ) must be real for all x.
Remark 7.16. Note from parts 2 and 3 of Definition 7.14 that we have Remark 7.16. Note from parts 2 and 3 of Definition 7.14 that we have
(ax\ + fix2, y) = a(x\, y) + P(x2, y}.
Remark 7.17. The Euclidean inner product of x, e C" is given by Remark 7.17. The Euclidean inner product of x, y E C n is given by
n
(x, y)
= LXiYi = xHy.
i=1
The conventional definition of the complex Euclidean inner product is (x, y) yH but we The conventional definition of the complex Euclidean inner product is (x, y} = yHxx but we use its complex conjugate H here for symmetry with the real case. use its complex conjugate xHyy here for symmetry with the real case.
Remark 7.18. A weighted inner product can be defined as in the real case by (x, y)Q = Remark 7.1S. A weighted inner product can be defined as in the real case by (x, y}Q — x H Qy, arbitrary Q QH > o. notion Qorthogonality can be similarly XH Qy, for arbitrary Q = QH > 0. The notion of Q orthogonality can be similarly generalized to the complex case. generalized to the complex case.
56 56
Chapter 7. Projections, Inner Product Spaces, and Norms Chapter 7. Projections, Inner Product Spaces, and Norms
Definition 7.19. A vector space (V, F) endowed with a specific inner product is called an Definition 7.19. A vector space (V, IF) endowed with a specific inner product is called an inner product space. If F = C, we call V a complex inner product space. If F = R, we inner product space. If IF = e, we call V a complex inner product space. If IF = R we call V a real inner product space. call V a real inner product space.
Example 7.20. Example 7.20. 1. Check that V = IRn x" with the inner product (A, B) = Tr AT B is a real inner product 1. Check that = R" xn with the inner product (A, B) = Tr AT B is a real inner product space. Note that other choices are possible since by properties of the trace function, space. Note that other choices are possible since by properties of the trace function, Tr AT B = TrB TA = Tr A B = TrBAT TrATB = Tr BTA = TrABTT = Tr BAT..
2. Check that V = e nxn with the inner product (A, B) = Tr AHB is a complex inner Tr AH B is a complex inner 2. Check that V = Cnx" with the inner product (A, B) product space. Again, other choices are possible. product space. Again, other choices are possible. Definition 7.21. Let V be an inner product space. For v e V, we define the norm (or Definition 7.21. Let V be an inner product space. For v E V, we define the norm (or length) ofv by IIvll = */(v, v). This is called the norm induced by (',, .).. length) ofv by \\v\\ = J(V,V). This is called the norm induced by (  ) Example 7.22. Example 7.22. 1. If V = E." with the usual inner product, the induced norm is given by i> 1. If V = IRn with the usual inner product, the induced norm is given by II v II = n 2 2 1
(Li=l V i (E,=i<Y))2.xV—*« 9\ 7
2. If V = en with the usual inner product, the induced norm is given by II v II = 2. If V = C" with the usual inner product, the induced norm is given by \\v\\ "n (L...i=l IVi ) ! (£? = ,l»,lI22)*.. Theorem 7.23. Let P be an orthogonal projection on an inner product space V. Then Then Theorem 7.23. Let P be an orthogonal projection on an inner product space \\Pv\\ ::::: Ilvll for all v e V. IIPvll < \\v\\forallv E V.
Proof: Since P is an orthogonal projection, p2 = P = pH. (Here, the notation p# denotes Proof: Since P is an orthogonal projection, P2 = P = P#. (Here, the notation P# denotes the unique linear transformation that satisfies ( P u , } = (u, p#v) for all u, v E If this the unique linear transformation that satisfies (Pu, vv) = (u, P#v) for all u, v e V. If this seems a little too abstract, consider V = R" (or en), where P# is simply the usual PT (or seems a little too abstract, consider = IRn (or C"), where p# is simply the usual pT (or pH)). Hence (Pv, v) = (P 2v, v) = (Pv, p#v) = (Pv, Pv) = IIPvll 2 > O. Now /  P is PH)). Hence ( P v , v) = (P2v, v) = (Pv, P#v) = ( P v , Pv) = \\Pv\\2 ::: 0. Now /  P is also a projection, so the above result applies and we get also a projection, so the above result applies and we get
0::::: ((I  P)v. v) = (v. v)  (Pv, v)
=
IIvll2  IIPvll 2
from which the theorem follows. from which the theorem follows.
0
Definition 7.24. The norm induced on an inner product space by the "usual" inner product Definition 7.24. The norm induced on an inner product space by the "usual" inner product is called the natural norm. is called the natural norm.
In case V = C" or V = R",, the natural norm is also called the Euclidean norm. In In case = en or = IR n the natural norm is also called the Euclidean norm. In the next section, other norms on these vector spaces are defined. A converse to the above the next section, other norms on these vector spaces are defined. A converse to the above procedure is also available. That is, given a norm defined by IIx II = •>/(•*> x), an inner procedure is also available. That is, given a norm defined by \\x\\ — .j(X,X}, an inner product can be defined via the following. product can be defined via the following.
7.3. Vector Norms 7.3. Vector Norms Theorem 7.25 (Polarization Identity). Theorem 7.25 (Polarization Identity).
1. For x, y E m~n, an inner product is defined by 1. For x, y € R", an inner product is defined by (x,y)=xTy=
57 57
IIx+YIl2~IIX_YI12_
IIx + yll2 _ IIxll2 _ lIyll2 2
2. For x, y E en, an inner product is defined by 2. For x, y e C", an inner product is defined by
where j = i = \/—T. where j = i = .J=I.
7.3 7.3
Vector Norms Vector Norms
Definition 7.26. Let (V, F) be a vector space. Then \ Definition 7.26. Let (V, IF) be a vector space. Then II \ . \ II\ : V + R is a vector norm ifit V >• IR is a vector norm if it satisfies the following three properties: satisfies the following three properties:
1. Ilxll::: Ofor all x E V and IIxll = 0 ifand only ifx
= O.
2. Ilaxll = lalllxllforallx
E
Vandforalla
E
IF.
3. IIx + yll :::: IIxll + IIYliforall x, y E V. (This is called the triangle inequality, as seen readily from the usual diagram illus (This is called the triangle inequality, as seen readily from the usual diagram illustrating the sum of two vectors in ]R2 .) trating the sum of two vectors in R2 .) Remark 7.27. It is convenient in the remainder of this section to state results for complexRemark 7.27. It is convenient in the remainder of this section to state results for complexvalued vectors. The specialization to the real case is obvious. valued vectors. The specialization to the real case is obvious. Definition 7.28. A vector space (V, F) is said to be a normed linear space if and only if Definition 7.28. A vector space (V, IF) is said to be a normed linear space if and only if there exists a vector norm  •  : V > R satisfying the three conditions of Definition 7.26. there exists a vector norm II . II : V + ]R satisfying the three conditions of Definition 7.26. Example 7.29. Example 7.29.
1. For x E en, the Holder norms, or pnorms, are defined by 1. For e C", the HOlder norms, or pnorms, are defined by
Special cases: Special cases: (a) Ilx III = L:7=1
IXi
I (the "Manhattan" norm).
1
(b) Ilxllz = (L:7=1Ix;l2)2 = (c) Ilxlioo
(X
H
1
X)2
(the Euclidean norm).
= maxlx;l
IE!!
=
(The second equality is a theorem that requires proof.) (The second equality is a theorem that requires proof.)
p++oo
lim IIxllp
(x Hyy)(yH x). Since is a nonnegative definite matrix. define the vector norm 3. The norm II . it is particularly easy to remember.1^ IIUxll2 = IIxll2 (Proof IIUxili = x U Ux = xHx = IIxlli)· However.  . is a nonnegative definite matrix. Proof' Consider the matrix [x y] E C" x2 . 1cose 1~ 1. tO~t:5. Theorem 7.e. y E C".58 58 Chapter 7. On the vector space (C[to. if U E enxn is unitary.. Some weighted pnorms: 2. p. On the vector space (C[to. „ . Since yHxx = x Hy. However. A particular case of the Holder inequality is of special interest. y E en may be defined by cos# = 1I. its determinant must be nonnegative.31 and Remark 7. define the vector norm 11111 = max 1/(t)I· to:::.D = E^rf/l*/!.33. Theorem 7. Since yH = x H y.30 (HOlder Inequality). if U € C"x" is unitary.. R). \\Ux\\l XHX = \\x\\\). Inner Product Spaces. Some weighted pnorms: (a) IIxll1.> where Q = QH > 0 (this norm is more commonly = QH > Ikllz. Remark 7. D 0 \\X\\2\\y\\2Note: This is not the classical algebraic proof of the CauchyBunyakovskySchwarz Note: This is not the classical algebraic proof of the CauchyBunyakovskySchwarz (CBS) inequality (see.33.lx. Since Proof: Consider the matrix [x y] e en x2 . o ~ (x Hxx)(yH y) . Let x. (b) IIx IIz. then Remark 7. 112 is unitarily invariant. i. Then Fhcorem 7.32 are true for general inner product spaces. Then Theorem 7. Then with equality if and only if x and yare linearly dependent. 1Ft). y E en. > 0. t \ ] R). define the vector norm 1111100 = max II/(t) 11 00 .tl Theorem 7. i. The angle e between two nonzero vectors x. whered. The angle 0 between two nonzero vectors x. Let x. 0 < e — 5. 217]). [20. it is particularly easy to remember. e. p..~~1~1112'. Remark 7. The norm  • 2 is unitarily invariant. IIQ)' 1 3. denoted II . Projections.g. JC. y e en.31 (CauchyBunyakovskySchwarz Inequality). Remark 7.D = L~=ld. +=1.34. The CBS inequality is thus equivalent to the statement I. and Norms Chapter 7. Projections.Q — (xhH Qx) 2. Theorem 7.e. Ther.. and Norms 2. ttlr.32. where 4 > O. However. Let x.30 (Holder Inequality).^ cos e = IlMmlylb 0 ~ 0 < I' The CBS inequality is thus equivalent to the statement ~ ^  COS 0 < 1. y e C" may be defined by Remark 7.g. we see immediately that IXH y\ ~ 0 < ( x H ) ( y H y ) — ( x H ) ( y H x ) . Let x.. t\])n. we see immediately that \XH yl < IIxll2l1yllz. e. In other words.32 are true for general inner product spaces. ttl. Inner Product Spaces. 217])..t~JI On the vector space ((C[to. However.l. In other words. 1Ft).. (CBS) inequality (see.34. define the vector norm On the vector space «e[to.31 (CauchyBunyakovskySchwarz Inequality).g = (x QXY denoted  • c). p q I I A particular case of the HOlder inequality is of special interest.31 and Remark 7.32. 11·111 and 1I·IIClO XHUHUx . with equality if and only if x and y are linearly dependent. [20. y e C". its determinant must be nonnegative. then H H \\Ux\\2 \\x\\2 (Proof. and  . Remark 7.
.e. ConFinally. this is called the triangle inequality. vectors under orthogonal transformation.e. lIaAl1 = lalliAliforall A E mxn andfor all a E IR.   R mx " ~ E is a matrix norm if it satisfies the following three Definition 7. IIAII ~ Ofor all A E IR mxn and IR IIAII = 0 if and only if A = O. the following inequalities are all tight bounds. there exist constants c\. IR) since that is "convergence" of matrices. there exist Example 7. C2 (possibly 7.. .e. If y € C" are orthogonal. IIxl12 :::: Jn Ilxll oo . we conclude this section with a theorem about convergence of vectors. 3. Remark 7.36. then we have the Pythagorean Identity Ilx ± YII~ = IIxll~ + IIYII~. Then lim k4+00 V(k) = v if and only if lim k~+oo II v(k)  v II = O. 2. Let \\ \\ be a vector norm and suppose v. Then 7. II·• II : IR mxn > IR is a matrix norm if it satisfies the following three properties: properties: 1. Attention is confined to the vector space (IRmnxn. ci (possibly depending onn) such that depending on n) such that Example 7. we conclude this section with a theorem about convergence of vectors. while the latter is needed to make sense of former notion is useful for perturbation analysis. i.e. i. All norms on en are equivalent.) .. 7.35.35. Definition 7. Theorem 7. the motivation for using matrix norms is to have a notion of either the size of or the nearness of matrices. Matrix Norms 7.. The former notion is useful for perturbation analysis. Extension to the complex case is straightforward what arises in the majority of applications. y E en are orthogonal.4 7. i. As with vectors.4. the following inequalities are all tight bounds.e. Extension to the complex case is straightforward and essentially obvious.37.37. Theorem 7.. while the latter is needed to make sense of "convergence" of matrices.e.. convergence in terms of vector norms. there exist vectors x for which equality holds: vectors x for which equality holds: Ilxlll :::: Jn Ilxlb Ilxll2:::: IIxll» IIxlloo :::: IIxll» Ilxlll :::: n IIxlloo. then we have the Pythagorean Identity Remark 7.. there exist constants CI. e C".. and essentially obvious. Attention is confined to the vector space (W xn R) since that is what arises in the majority of applications. Let II·• II be a vector norm and suppose v.. the motivation for In this section we introduce the concept of matrix norm. Matrix Norms 59 59 are not unitarily invariant.38.. As with vectors. For x E en. IIA + BII :::: IIAII + IIBII for all A.4 Matrix Norms Matrix Norms In this section we introduce the concept of matrix norm. Similar remarks apply to the unitary invariance of norms of real vectors under orthogonal transformation. v(2).4. BE IRmxn. Finally. i» (1) v(2\ . about convergence of real numbers. i. All norms on C" are equivalent. 2 the proof of which follows easily from z2 _ z_//. Convergence of a sequence of vectors to some limit vector can be converted into a statement vergence of a sequence of vectors to some limit vector can be converted into a statement about convergence of real numbers. (As with vectors. the proof of which follows easily from liz II~ = ZH z.7. convergence in terms of vector norms.. E en.39. v(l). i. IIxlioo :::: IIxllz. For x G C". If x.39. i.36. The using matrix norms is to have a notion of either the size of or the nearness of matrices..38. Similar remarks apply to the unitary invariance of norms of real are not unitarily invariant.
The "matrix analogue of the vector 1norm. is a norm._ Then "mixed" norms can also be defined by e lR..q = max IIAxil p 11.<110#0 IIxllq Example 7. Let A E K m x ". Example 7. IIAlioo = max rE!!l. J=1 3.42.41. Example 7." Each is a "computable.00 =  • 2.40. Schatten/7norms IIAlls.. The spectral norm is 3. (t laUI). Inner Product Spaces. Example 7.40. ^wncic = rank(A)). e R mx ". IIAII P t altA)) 1 ~ (T.42. (A' A)) 1 ~ (T. (where r = laiiK^/i. Inner Product Spaces. Then the Frobenius norm (or matrix Euclidean norm) is 7. and Norms Example 7.jj laij. Example 7. + a!)"". 1. \\F and 11'115.mxn IIAII p. The concept of a matrix norm alone is not altogether useful since it does not allow us to estimate the size of a matrix product AB in terms of the sizes of A and B individually. I Some special cases of Schatten /?norms are equal to norms defined previously. Then the matrix pnorms are defined by A e Rmxn. 5>1 is often called the trace norm.mxn. IIAII2 = Amax(A A) = A~ax(AA ) = a1(A). I. Let A E lR.44.44. The norm  . matrix = Ilxllp. Example 7. Example 7." theorem and requires a proof. defined by IIAIIF ~ (t. to estimate the size of a matrix product A B in terms of the sizes of A and B individually.  . pnorms previously. and Norms Chapter 7. is a norm.  5 2 = II IIF and  • 5i00 = II ..2 =  . Projections. Let A E R . The following three special cases are important because they are "computable.43.) I ~ (t. 11·115. I. The "matrix analogue of the vector Inorm.p = (at' + . (AA ')). For example. tTL T Note: IIA+llz = l/ar(A).43. Projections. 112' The norm II • 115. The Schattenpnorms are defined by E lR.60 max _P IIAxll = max Ilxli p IIxllp=1 IIAxll p . Let A E lR. where r mxn = rank(A). ai. .60 Chapter 7. The "maximum column sum" norm is 2."  A\\ = ^ \ai} .1 is often called the trace norm." IIAliss = Li.mxn.. The "maximum row sum" norm is 2.mxn.
IIAII2 ::s . i. Matrix Norms 61 61 Notice that this difficulty did not arise for vectors. . B e Rnxk.jii IIAlioo' .jii IIAII2. •II p for all p are consistent matrix norms.. IIAIIF ::s. a vector norm.q = maxx. Example 7. If II .jii IIAII I. we clearly have Ajc ::s A1jt.jii IIAlloo.e.1100 = max laijl x. IIAxl1 IIAII = max .4. more generally. although there are analogues for. IIAlioo ::s . Example 7. 11^4^11 P (or. Theorem 7. which equality holds: which equality holds: IIAIII ::s . \\v consistent with it. 1. p for all p are consistent matrix norms. there exists a vector norm II • IIv Theorem 7.~~i'.j is a matrix norm but it is not consistent. although there are analogues for. IIAlioo ::s .e. Then the norms II • \\a II· Ilfl' and . 2.48. Then The p norms are examples of matrix norms that are subordinate to (or induced by) The pnorms are examples of matrix norms that are subordinate to (or induced by) a vector norm. IIABIII.jii II A IIF. II such that AF is given by max^o ".4.47.. e.e. IIAIIF ::s . not exist a vector norm II •  such that IIAIIF is given by max x . IIAxliv < IIAlim \\x\\v' Not every consistent matrix norm is subordinate to a vector norm.46... consider  • \\F. We thus need the following definition. We thus need the following definition.ooIlBIII.45. \\ • \\p. inner products or outer products of vectors. Matrix Norms 7.jii IIAlb IIAIII ::s n IIAlloo. it follows that all subordinate norms are consistent. \\m is a consistent matrix norm.  • /7and II .::S \\A\\p\\B\\y A matrix norm \\ • is said to be consistent if \\AB\\ ::s  A   B II whenever the matrix product is defined.60 .oo 2 while IIAIII.47. if II A B II < II A 1111 fi whenever the matrix product is defined.e. II· II F and 1. IIAllp. For such subordinate norms. i. Theorem 7. the IIIn II F = .60 \^ • Useful Results The following miscellaneous results about matrix norms are collected for future reference. A A 2.= max IIAxl1 x..oo J1.46. more generally.so II . 1. IIAxll1 = max . e R" x ". For example. but there does consider II .7. there exist matrices A for i. subordinate to the vector norm. i. B E ]Rnxk.jii. take A = B = [: is a matrix norm but it is not consistent. 2. also caUedoperator norms. There exists a vector x* such that Ajt* = A jc* if the matrix norm is Theorem 7. there exists a vector norm \\ . i. A matrix norm 11·11\\is said to be consistent mutuallyconsistentifIlABII. wec1earlyhave IIAxll < IIAllllxll· Since Afijc < IIAlIllBxll < Afljt. II A 1100 ::s n IIAII I .g. l...Then II Ax 1122 ::s II AFjc2.. HAjcJI^ ::s \\A\\m Ilxli v.60 IIx II Ilxll=1 IIAxll p .jii IIAlb IIAIIF ::s . Then :]. also called oper(or. Then the norms \\ . Let A e Rmxn. while E ]Rnxn. For example. For example. IIAII2 ::s.jii IIAII I . atornorms. A = max^o IIxll. For A following inequalities are all tight. The "mixed" norm "mixed" norm II· 11 100 . IIAIII ::s . inner products or outer products of vectors.and II \\ •lIy y are mutually consistent if \\ A B \\ a < IIAllfllIBlly. reader The interested reader is invited to prove each of them as an exercise..60 Ilx i. take A = B = \ \ Afl li00 = 2whileA li00 B 1>00 = 1. e. q For such subordmate norms. Notice that this difficulty did not arise for vectors.e. Not every consistent matrix norm is subordinate to a vector norm.jii IIAIIF. •II F. II". IIAII2::S IIAIIF. 2. )). II In II p = 1 for all p.. it follows that all subordinate norms are consistent. consistent with it. If \\ • 11m is a consistent matrix norm. Since IIABxl1 ::s Afljc ::s IIAIIIIBllllxll.but there does  is consistent with F..45. i. There exists a vector x* such that IIAx*11 = IIAllllx*11 if the matrix norm is subordinate to the vector norm. II F' ThenA^ < A II Filx 112. so not exist a vector norm  . \\ are Definition 7. •1122is consistent with II . For example. exercise. . The following miscellaneous results about matrix norms are collected for future reference.e. Let A E ]Rmxn.. Definition 7..48.g.
Suppose P and Q are orthogonal projections and P + Q = I. \\ \\bea Rmx". Let II ·11 be a matrix norm and suppose A. For A E IR mxn . but not necessarily other pnorms) are unitarily invariant. Projections.. Definition: Let A e Rnxn and denote its set of eigenvalues (not necessarily distinct) 8. spanned by the plane 3x . 6.c — v + = 0.] 4. 3. and Norms Chapter 7. The spectral radius of A is the scalar by {Ai . scalars. 2. A(A A) 1 AT 5.l.A+A is an orthogonal projection. EeIRmxn.] l.1. matrices Q zR and e M" "...49. orthogonal projection. prove directly that V22Vl is an I — +A V V/ is an orthogonal projection. 3.Q — Q must be an orthogonal matrix. i. EXERCISES EXERCISES 1. where V2 is defined as in Theorem 5.1. space.. IIQAZlia = A fora = 2 or F. . prove that P+ = P. (MZa or F.An}. Then 7..e. A (1) . . Suppose that a matrix A E IR mxn has linearly independent columns. Prove that the A e Wnxn orthogonal projection onto the space spanned by these column vectors is given by the P matrix P = A(ATTA)~}AT. 4. [2 3 4]r R3 spanned by the plane 3. Theorem 7.. Show that the matrix norms II . A(2)... i. 112 (as well as all the Schatten /?norms. Projections. A (2) . Definition: Let A E IRnxn and denote its set of eigenvalues (not necessarily distinct) by P. Also.. must be an orthogonal matrix. 7. Inner Product Spaces. The norms II .49. IIF and II . Prove that E"xn with the inner product (A. Find the (orthogonal) projection of the vector [2 3 4f onto the subspace of 1R3 5.. B) = space. l. Then k~+oo lim A (k) = A if and only if k~+oo lim IIA (k)  A II = o. Prove that P . .. If P projection. 112 and II . „ } The spectral radius of A is the scalar p(A) = max IA. If P is an orthogonal projection. and Norms max laijl :::: IIAII2 :::: ~ max laijl. Prove that / . 1... IIF are unitarily invariant. but not necessarily The norms  • \\F and  • 2 (as well as all the Schatten pnorms. . e Rmx" mxm x mxm and Z E IRnxn . p+ = P. Inner Product Spaces. B) = Tr ATB is a real inner product IR n x" AT B (A. > . Chapter 7. for all A E IRmxn and for all orthogonal unitarily invariant.y + 2z = O. For A eRmxa . where ¥2 is defined as in Theorem 5. A(I). i .  • 2 and  • \\F 8. IIAllaa fora matrices Q E IR Convergence Convergence The following theorem uses matrix norms to convert a statement about convergence of a sequence of matrices into a statement about the convergence of an associated sequence of of scalars.I. .62 62 3.e.
) that  M Up = for all/?.) T 10. y E IR n are nonzero. all of whose Determine AF. 10. Determine AF. (An n x n matrix. where both x. \\A\\ A2. and p(A). IIAlb IIAlloo. H A I I A2. it can be proved that IIMllp = ss for all p. If M is a magic square matrix. y e R" are nonzero. and p(A). where ex and {3 take the value 1. Let A = xyT. . IIAlb IIAlloo.Exercises Exercises 63 63 Let Let A=[~ 14 0 12 5 ~]. appropriate. 2.. H A H ^ and peA). where both x. Determine IIAIIF' IIAII Ilt. A2. is called a "magic square" matrix.2. and peA). where a and ft take the value 1.. Let 9. Determine AF. Determine IIAIIF' IIAII d . or (Xl as appropriate. Determine IIAIIF' IIAIII> IIAlb and Aoo in terms of \\x\\a and/or \\y\\p. 9. Aj. all of whose columns and rows as well as main diagonal and antidiagonal sum to s = n(n2 + 1) /2. Let A = xy . (An n x n matrix. or oo as and II A 1100 in terms of IIxlla and/or IlylljJ. HA^. columns and rows as well as main diagonal and antidiagonal sum to s = n (n 2 l)/2. Let A=[~4 9 2 ~ ~].
This page intentionally left blank This page intentionally left blank .
so these two vectors are orthogonal.x — b\\\ (and hence p ( x ) = from the Pythagorean identity (Remark 7. i. x e X if and only if latter form is commonly known as the normal equations. write the residual r in the form To see why this must be so.. A.Ax is the residual associated 1. The linear least Problem: Suppose A e jRmxn with > n and b <= Rm is given vector. Hence. The equations ATrr = 0 can be rewritten in the form A TAx = ATb and the x.b 112) assumes its minimum value if and only if II Ax —b\\2) assumes its minimum value if and only if (8.Ax) is clearly in 'R(A). x E X if and only if x is a solution of the normal equations.PR(A))b = PR(A). 2..e.PR(A)b) + (PR(A)b  Ax). (PR(A)b .Axll~ = lib .bll 2 is minimized}. Thus. where r = b .1) To see why this must be so. IIAx .PR(A)b) = (I . For further details. Thus.PR(A)bll~ + IIPR(A)b  Axll~ from the Pythagorean identity (Remark 7. A vector x X if and onlv if x is of the x=A+b+(IA+A)y. is a solution of the normal equations. For further details. (8. whereyEjRnisarbitrary.. A vector x E X if and only if ATrr = 0. see Section 8. The linear least squares problem consists of finding an element of the set squares problem consists of finding an element of the set x = {x E jRn : p(x) = IIAx .2) 65 . vector x e X if and only if AT where b — Ax is the residual associated with x.1 The Linear Least Squares Problem The Linear Least Squares Problem Problem: Suppose A E Rmx" with m 2: nand b E jRm is aagiven vector. Now. Hence. AT — A T Ax = AT b latter form is commonly known as the normal equations.35).bll~ (and hence p(x) = \\Ax .b E 'R(A)L so these two vectors are orthogonal. (Pn(A)b — AJC) is clearly in 7£(A).1 8.35).e. i. A vector x E X if and only if x is of the form 2. IIrll~ = lib . see Section 8. Solution: The set X has a number of easily verified properties: The set X has a number of easily verified properties: 1. write the residual in the form r = (b . while Now.2.2. while (b .Chapter 8 Chapter 8 Linear Least Squares Linear Least Squares Problems Problems 8.
x" = + b is the unique vector 4. equivalently.. There is a unique solution to the least squares problem.3. x* = A+b is the unique vector that solves this "double minimization" problem. This follows immediately from convexity or directly from the fact that all x e X are of the form (8.e. X = A+B. all and this equation always has a solution since AA+b E R(A). 5. The unique solution of minimum 2norm or Fnorm is where Y € ]R. By Theorem 6. The only difference is that in the case same as solutions of the linear system AX = B.e.2) are of the form x = A+ AA+b + (I  A+ A)y =A+b+(IA+A)y..2. Let 6 E [0.mxn XElR Plxk min IIAX  Bib is of the form is of the form X=A+B+(IA+A)Y.1. + (1 .e.A+A)(Oy (1 . Let A E E mx " and B € Rmxk. consider two arbitrary vectors Xl = A + b 3. The only difference is that in the case of linear least squares solutions. if and only if rank (A) = n.nxk is arbitrary. where y e W is arbitrary.PR(A)bll z = ~ 11(1 Ilbll z. Let 8 e [0. then equality holds and the least squares .A+ A)z in X. equivalently. then equality holds and the least squares If the existence condition happens to be satisfied. 1]. Linear Least Squares Problems Chapter 8. To see why. There is a unique solution to the least squares problem. i.A+A)y and *2 = A+b + (I — A+A)z in X. Notice that solutions of the linear least squares problem look exactly the Remark 8.2) are of the form solutions of (8. By Theorem 6. x* minimizes the residual p(x) that solves this "double minimization" problem.3. problem to the matrix case. there is no "existence condition" such as R(B) S.66 Chapter 8.. we can generalize the linear least squares problem to the matrix case. AA+)bI1 2 the last inequality following by Theorem 7. The unique solution of minimum 2norm or Fnorm is X = A+B.. i.8)xz2 = A+b ++ (I A+ A)(8y ++ (1 8)z) is clearly in X. To see why. If the existence condition happens to be satisfied. has a unique element x" of minimal2norm.23. The Theorem 8. if and only if rank(A) n.. X = {x"} = {A+b}. Remark 8. Linear Least Squares Problems and this equation always has a solution since AA+b e 7£(A). 7£(A). i. In fact. and only if A+A = I or.1]. BE ]R. x* minimizes the residual p ( x ) and is the vector of minimum 2norm that does so. i. Then the convex combination and Xz = A+b (I . X has a unique element x* of minimal 2norm. Just as for the solution of linear equations. X = {x*} = {A+b}. the last inequality following by Theorem 7. of linear least squares solutions. This follows immediately from and is the vector of minimum 2norm that does so.23. X. where Y E R" xfc is arbitrary.0)z) is clearly in 4. The general solution to e ]R. The minimum value of p x ) is then clearly equal to lib .e. which follows since the two vectors are orthogonal. The minimum value of p ((x) is then clearly equal to where y E ]R.mxk. we can generalize the linear least squares Just as for the solution of linear equations. Notice that solutions of the linear least squares problem look exactly the same as solutions of the linear system AX = B. X is convex. if 5. Then the convex combination 8x. In fact.. X is convex. 0*i (1 #)* = A+b (I .1) and which follows since the two vectors are orthogonal. 3.1) and convexity or directly from the fact that all x E X are of the form (8. there is no "existence condition" such as K(B) c R(A). consider two arbitrary vectors jci = A+b + (I — A + A) y (I .2.n is arbitrary. if and only if A + A lor. all solutions of (8.
8.3 Linear Regression and Other Linear Least Squares Problems 8.3 Linear Regression and Other Linear Least Squares Problems
67
O. X = +B residual is 0. Of all solutions that give a residual of 0, the unique solution X = A+B has minimum 2norm or F norm. Fnorm. Remark 8.3. If we take B = 1m in Theorem 8.1, then X = A+ can be interpreted as Im in Theorem 8.1, then Remark 8.3. If we take B A+ can be interpreted as saying that the MoorePenrose pseudoinverse of A is the best (in the matrix 2norm sense) A AX matrix such that AX approximates the identity. Remark 8.4. Many other interesting and useful approximation results are available for the F norm). matrix 2norm (and Fnorm). One such is the following. Let A E M™ x " with SVD following. e lR~xn
A
= U~VT = LOiUiV!.
i=l
Then a best rank k approximation to A for 1< f c < r r,i . e . , a solution to A k l :s k :s , i.e., a
MEJRZ'xn
min IIA  MIi2,
is given by is given by
k
Mk =
LOiUiV!.
i=1
The special case in which m = n and k = n  1 gives a nearest singular matrix to A E A e = nand = —
lR~ xn .
8.2 8.2
Geometric Solution Geometric Solution
Looking at the schematic provided in Figure 8.1, it is apparent that minimizing IIAx —bll 2 2  Ax b\\ x e W1 p — Ax is equivalent to finding the vector x E lRn for which p = Ax is closest to b (in the Euclidean b Ay norm sense). Clearly, r = b  Ax must be orthogonal to R(A). Thus, if Ay is an arbitrary r b — Ax 7£(A). R(A) vector in 7£(A) (i.e., y is arbitrary), we must have y
0= (Ay)T (b  Ax) =yTAT(bAx) = yT (ATb _ AT Ax).
Since y is arbitrary, we must have ATb — ATAx = 0 or A r A;c = AT b. AT b  AT Ax AT Ax = ATb. T Special case: If A is full (column) rank, then x = (AT A) ATb. A = (A A)l ATb.
8.3 8.3
8.3.1 8.3.1
Linear Regression and Other Linear Least Squares Linear Regression and Other Linear Least Squares Problems Problems
Example: Linear regression
Suppose we have m measurements (ll, YI), ... ,, (trn,,ym) for which we hypothesize a linear (t\,y\), . . . (tm Ym) (affine) relationship (8.3) y = at + f3
68
Chapter 8. Linear Least Squares Problems Chapter 8. Linear Least Squares Problems
b
r
p=Ax
Ay E R(A)
Figure S.l. Projection of b on K(A). Figure 8.1. Projection of b on R(A).
for certain constants a. and {3. One way to solve this problem is to find the line that best fits for certain constants a and ft. One way to solve this problem is to find the line that best fits the data in the least squares sense; i.e., with the model (8.3), we have the data in the least squares sense; i.e., with the model (8.3), we have
YI
Y2
= all + {3 + 81 ,
= al2 + {3 + 82
where &\,..., 8m are "errors" and we wish to minimize 8\ + • • 8;. Geometrically, we where 81 , ... , 8m are "errors" and we wish to minimize 8? + ...• + 8^ Geometrically, we are trying to find the best line that minimizes the (sum of squares of the) distances from the are trying to find the best line that minimizes the (sum of squares of the) distances from the given data points. See, for example, Figure 8.2. given data points. See, for example, Figure 8.2.
y
Figure 8.2. Simple linear regression. Figure 8.2. Simple linear regression.
Note that distances are measured in the vertical sense from the points to [he line (as Note that distances are measured in the venical sense from the point!; to the line (a!; indicated, for example, for the point (t\, y\}}. However, other criteria arc possible. For exindicated. for example. for the point (tl. YIn. However. other criteria nrc po~~iblc. For cxample, one could measure the distances in the horizontal sense, or the perpendicular distance ample, one could measure the distances in the horizontal sense, or the perpendiculnr distance from the points to the line could be used. The latter is called from the points to the line could be used. The latter is called total least squares. Instead squares. Instead of 2norms, one could also use 1norms or oonorms. The latter two are computationally of 2norms, one could also use Inorms or oonorms. The latter two are computationally
8.3. Linear Regression and Other Linear Least Squares Problems 8.3. Linear Regression and Other Linear Least Squares Problems
69
much more difficult to handle, and thus we present only the more tractable 2norm case in difficult text that follows. follows. The m "error equations" can be written in matrix form as ra
Y = Ax +0,
where
We then want to solve the problem
minoT 0 = min (Ax  y)T (Ax  y)
x
or, equivalently, min lIoll~ = min II Ax  YII~.
x
(8.4)
AT Solution: x = [~] is a solution of the normal equations AT Ax Solution: x — [^1 is a solution of the normal equations ATAx = ATyy where, for the special form of the matrices above, we have special form of the matrices above, we have
and and
AT Y = [ Li ti Yi
LiYi
J.
The solution for the parameters a and f3 can then be written ft
8.3.2
Other least squares problems
y = f(t) =
Cl0!(0
(8.3) of the form Suppose the hypothesized model is not the linear equation (S.3) but rather is of the form + • • • 4 cn<t>n(t). (8.5) (8.5)
In (8.5) the ¢i(t) are given (basis) functions and the Ci; are constants to be determined to </>,(0 functions c minimize the least squares error. The matrix problem is still (S.4), where we now have minimize the least squares error. The matrix problem is still (8.4), where we now have
An important special case of (8.5) is least squares polynomial approximation, which corresponds to choosing ¢i (t) = t t'~1,, i i;Ee!!, although this choice can lead to computational 0,• (?) = i  l n, although this choice can lead to computational
VT = U. ] II: = II [ The last equality follows from the fact that if v [~~]. Since the standard Kalman filter essentially amounts method in finiteprecision arithmetic. Since the standard Kalman filter essentially amounts to sequential updating of normal equations. since II . etc.b\\^ is II czll ~. the subvectors can have different lengths). Linear Least Squares Problems difficulties because of numerical ill conditioning for large n. the last equivalent. we assume that A has an SVD given by A = UT. insight. piecewise polynomial functions.c=UTb = II [~ ~] [ ~~ ] . as in Theorem 5. A E IRmxn . bE IR m . The former is much more expensive but is generally more reliable and offers considerable theoretical offers insight. the two are equivalent. (8. We now note that IIAx  bll~ = IIU~VT x =  bll~ II ~ VT X  U T bll. [7]. 8. Better numerical methods are based on algorithms that behavior in practice (and it does). The key feature in (8. [11]. Then c. Then GI defining y = logy. then II v II ~ = II viii ~ + II v211 ~ (note that orthogonality is not what is used here. In this section we investigate solution of the linear least squares problem min II Ax x b11 2 . Ib is unitarily invariant =11~zcll~ wherez=VTx. splines.g. In fact. appear functions </>. if the fitting function is of the form y t) Y = ff( (t) = c\eC2i.1. norm. quantity above is clearly minimized by taking z\ = S'c. etc.. This that orthogonality is not what is used here. .[ ~~ ] II: sz~~ c. This explains why it is convenient to work above with the square of the norm rather than the concerned. Linear Least Squares Problems Chapter 8.4 8.can be arbitrarily nonlinear. Z2 while the minimum value of \\Ax — b II ~ is l^llr while the minimum value of II Ax .. For example. c. then u^ = i>i \\\ \\vi\\\ (note The last equality follows from the fact that if v = [£ ]. if the fitting function is of the form can be converted into a linear problem. = log c" and C2 = cj_ results in a standard linear least squares y — log y. we assume that A has an SVD given by A U\SVf via the SVD. Sometimes a problem in which the Ci'S appear nonlinearly nonlinearly can be converted into a linear problem. of linear least squares problems via the normal equations can be a very poor numerical method in finiteprecision arithmetic. For example. [7]. are based on orthogonal polynomials.. + c2f.70 70 Chapter 8.1. problem. As far as the minimization is concerned. fact. The former based on SVD and QR (orthogonalupper triangular) factorization. Specifically. it is shown that solution [4].6) via the SVD. S~lc\. [23]). Specifically. C2 problem. e C2 / then taking logarithms yields the equation log y = log c. c. then taking logarithms yields the equation logy = logci + cjt.5) is that the coefficients Ci appear linearly. c\ logci. the subvectors can have different lengths). respectively.SVr U~VT Theorem 5. Numerically better approaches ill difficulties n.4 Least Squares and Singular Value Decomposition Least Squares and Singular Value Decomposition In the numerical linear algebra literature (e. 's ¢i. it can be expected to exhibit such poor numerical behavior in practice (and it does). Better numerical methods are based on algorithms that AT work directly and solely on A itself rather than AT A. z. The subvector z2 is arbitrary. respectively. [4]. Two basic classes of algorithms are A itself S VD and QR (orthogonalupper triangular) factorization. The basis functions coefficients c. arbitrary.
This follows easily since Another expression for the minimum residual is II (I . x has been written in the form x = A+b + (I . To simplify the exposition. Thus. of course.5 Least Squares and QR Factorization Least Squares and QR Factorization In this section. Z VISici The last equality follows from The last equality follows from c = UTb = [ ~ f: ]= [ ~~ l Note that since Z2 is arbitrary. (8. UZV = [U t/2][o]^i r > and there is thus "no V2 part" to the solution. A e R™ X M . Least Squares and QR Factorization Now transform back to the original coordinates: Now transform back to the original coordinates: x = Vz 71 71 = [VI V2 1[ ~~ ] = VIZ I + V2Z2 = = + V2Z2 vlsIufb + V2 2. we again look at the solution of the linear least squares problem (8.. i.e.7) . to reduce A in the following way. This agrees. i. we again look at the solution of the linear least squares problem (8. can be quite reliable.. with appropriate numerical enhancements. with appropriate numerical enhancements. to reduce A in the following way.e. we have QT € ffi. It is then possible. It is then possible. A finite sequence of simple orthogonal row transformations (of Householder or Givens type) can be performed on A to reduce it row transformations (of Householder or Givens type) can be performed on A to reduce it to triangular form. If we label the product of such orthogonal row transformations as the orthogonal matrix QT E R m x m . The minimum value of the least squares residual is The minimum value of the least squares residual is and we clearly have that and we clearly have that minimum least squares residual is 0 4=> b is orthogonal to all vectors in U2 minimum least squares residual is 0 {::=:} b is orthogonal to all vectors in U2 {::=:} •<=^ {::=:} b is orthogonal to all vectors in R(A)l.1). This agrees. A e 1R™ X ".e. i. i. is orthogonal to all vectors in 7l(A}L b E R(A).. where y e ffi. we add the simplifying assumption that A has full column rank. = +b + (/ — A + A) y. Another expression for the minimum residual is  (/ — AA + )b 2 . can be quite reliable. V2 Z 2 is an arbitrary vector in 7Z(V2)) = A/"(A).~xn. In this case the SVD of A is given by socalled fullrank problem.A + A)_y. an important special case of the linear least squares problem is the Finally. Least Squares and QR Factorization B.~xn. 11(1. A E ffi.11U2U!b"~ = bTUZV!V UJb = bTVZV!b = IIV!bll~. This follows easily since (7 . than an SVD and.6) but this In this section.. of course. Thus. V2z is an arbitrary vector in R(V2 = N(A). In this case the SVD of A is given by A A = V:EVTT = [VI{ Vzl[g]Vr. x has Note that since 12 is arbitrary.mxm. A E ffi.S. where y E Rm is arbitrary.8. with (8. via a sequence of socalled Householder or Givens transformations. This matrix factorization is much cheaper to compute than an SVD and.AA+)bll~ . and there is thus "no V2 part" to the solution. an important special case of the linear least squares problem is the socalled fullrank problem. If we label the product of such orthogonal row transformations as the to triangular form. with (8.AA+)b\\22 = \\U2Ufb\\l = bTU2U^U22V!b = bTU2U*b = \\U?b\\22. Finally. A finite sequence of simple orthogonal transformations. This matrix factorization is much cheaper to compute time in terms of the QR factorization.m is arbitrary.5 8.6) but this time in terms of the QR factorization.5.AA+)bllz. 8. via a sequence of socalled Householder or Givens rank. we add the simplifying assumption that A has full column To simplify the exposition.e.1).
by writing (8. 112 is unitarily invariant ~ ] x .7). and any y E R". all in R". (b) Let r denote the "error vector" b .m IX(m ~" ) .9) is essentially what is accomplished by the GramSchmidt process. data. 3.. Now write Q = [Q\ Qz].Show that r is orthogonal to both^i and q2.mxn and where R e M£ x " is upper triangular.flq2 Show that r is orthogonal to (b) Let r denote the "error vector" b — ctq\ — {3qz.1Q\b = A+b and the minimum residual is II Qr bllz' is \\C2\\2.I of the columns of yields the orthonormal columns of Q\.8). Linear Least Squares Problems Chapter 8.8). check directly that (I . are orthogonal vectors..7).. (b) Find the best (in the 2norm sense) line of the form jc = ay + (3 that fits this (b) Find the best (in the 2norm sense) line of the form x = ay fJ that fits this data.9) are variously referred to as QR factorizations of A. and qz are two orthonormal vectors and b is a fixed vector. Linear Least Squares Problems where E ffi. data.. b e ffi. by writing AR~l1 = Q\ we see that a "triangular" linear combination (given by the coefficients of ARQI we see that a "triangular" linear combination (given by the coefficients of R. all in ffi. we have x = R. where Q\ e R mx " and Qz € K" x(mn). 3.e. i.aql .7). Note that Any of (8.3).1). sense). where QI E ffi. Now write Q = [QI Q2].e. Both Q I and Qz2 have orthonormal columns. Now note that Now note that IIAx  bll~ = IIQ T Ax = II [ QTbll~ since II . xn. The last quantity above is clearly minimized by taking x = R.9) Any of (8. i. Consider the following set of measurements (*. or (8. For A E Wmxn . Equivalently. we have = R~l Qf b = +b and the minimum residual is IIC?^!^ EXERCISES EXERCISES 1.7). m and any e ffi. For € ffi. Note that (8. n • (a) Find the optimal linear combination aq^ + (3q2 that is closest to b (in the 2norm (a) Find the optimallinear combination aql + fiq2 that is closest to b (in the 2norm sense).~xn is upper triangular. (2.72 Chapter 8. we see that A=Q[~J = [QI = QIR..+ A)y and A+b are orthogonal vectors. Suppose q. b E Em. Qz] [ (8. Yi): 2. Consider the following set of measurements (Xi. yt): (1.[ ~~ ] If:.2). (3.Cl and the minimum residual The last quantity above is clearly minimized by taking x = R lIc\ and the minimum residual is Ilczllz.Equivalently. Multiplying through by Q in (8.9) is essentially what is accomplished by the GramSchmidt process..8) ~ ] (8. (a) Find the best (in the 2norm sense) line of the form y = ax + fJ that fits this (a) Find the best (in the 2norm sense) line of the form y = ax + ft that fits this data. n . both ql and q2 . Multiplying through by Q Q2 E ffi. we see that in (8. Both Q\ and <2 have orthonormal columns. Suppose qi and q2 are two orthonormal vectors and b is a fixed vector. 2.9) are variously referred to as QR factorizations of A. (8. or (8. check directly that (I . R~l) ) of the columns of A yields the orthonormal columns of QI. (8.A+A)y and A +b 1.
bl1 2 when A = [~ ~ ] and b = [ !1 x The solution is (a) Consider a perturbation EI = [~ pi of A. where Q is orthogonal. Solve the perturbed problem positive number. not necessarily nonsingular. Solve the perturbed version of the above problem. A+ R+QT . Prove that A+ = R+ QT. verify that if A E ~. where 8 is a small positive number. then A+ R~ Q\. Let A E ~nxn. not necessarily nonsingular. Find all solutions of the linear least squares problem min II Ax .yII2 as 8 approaches O? (b) Now consider the perturbation E2 = [~ (b) Now consider the perturbation EI = \0 s~\ of A. What happens to jt* . then A+ == R.9).bll 2 x when A = [ ~ 5.:. 7.xn can be factored in the form (8. and suppose A where is 1. What happens to IIx* — y 2 as 8 approaches 0? where AI = A + E\. (a) Consider a perturbation E\ = [0 ~] of A. of 2norm solution of least «rmarp« problem squares nrr»h1<=>m min II Ax .bib z n where A2 = A + E2. Use the four Penrose conditions and the fact that Q\ has orthonormal columns to 6.9).IlQf. Let A e R"x". where again 8 is a small positive number. where AI = A + E I . where again 8 is a small of A.Exercises Exercises 73 4. Find all solutions of the linear least squares problem 4. Use the four Penrose conditions and the fact that QI has orthonormal columns to verify that if A e R™ x "can be factored in the form (8. where 8 is a small positive number. Consider the problem of finding the minimum 2norm solution of the linear least 5. and suppose A = QR. Solve the perturbed version of the above problem.• What happens to IIx* .z2 as 8 approaches O? where A2 — A E 2 What happens to \\x* — zll2 as 8 approaches 0? 6. Solve the perturbed problem min II A 2 z .
This page intentionally left blank This page intentionally left blank .
(Note that the characteristic polynomial can also be defined as det(Al .4. A nonzero vector x e C" is a right eigenvector of A e Cnxn if there exists a scalar A. Theorem 9. A nonzero vector x E en is a right eigenvector of A E e nxn if there exists Definition 9.— 3.2. as a matter of convenience. n(A) = 0. example.31 O. It is an easy exercise to Example 9. norm used for such scaling. Definition 9. Let A = [~g ~g]. For any A E e nxn . n(A) = O. It can be proved from elementary properties of determinants that if A E C" ". [3]).2) By taking Hennitian transposes in (9. Example 9. then so is ax [ay] for any nonzero scalar a E C. for example. [21D or directly using elementary properties of inverses and determinants (see.2. It can be proved easily from the Jordan canonical form to be discussed in the text to follow (see. verify that n(A) = A2 2A . This of A. called an eigenvalue.) throughout the text. the Fundamental Theorem of Algebra says that x 75 .1). One oftenused scaling for an eigenvector is One oftenused scaling for an eigenvector is so is ax [ay] for any nonzero scalar a E a = 1/ IIx II so that the scaled eigenvector has nonn 1. e C. a nonzero vector y E en is a left eigenvector corresponding to an eigenvalue Similarly./) is called the characteristic polynomial of A. It can be The following classical theorem can be very useful in hand calculation. then vector of AH associated with I. we use both forms throughout the text. Then n(A) A2 + 2A 3. Theorem 9.Al) is called the characteristic polynomial Definition 9. the Fundamental Theorem of Algebra says that 7t (X) is a polynomial of degree n. The 2norm is the most common nonn used for such scaling.1.) = det (A .4. we see immediately that XH is a left eigenvector of A H associated with A. we use both forms results in at most a change of sign and. such that a scalar A E e. for example. for proved easily from the Jordan canonical fonn to be discussed in the text to follow (see. The polynomial n (A) = det(A—A.1.1 9. Note that if x [y] is a right [left] eigenvector of A. such that Ax = AX. It is an easy exercise to 2 verify that n(A) = A + 2A . The 2nonn is the most common a — \j'. then n(A) is a polynomial of degree n. (9. This results in at most a change of sign and. Thus. Then n(k) = X2 + 2A. we see immediately that x H is a left eigenBy taking Hermitian transposes in (9../ — A).1) Similarly.A). Note that if x [y] is a right [left] eigenvector of A. Let A [~ ~]. [3]).Chapter 9 Chapter 9 Eigenvalues and Eigenvalues and Eigenvectors Eigenvectors 9. called an eigenvalue. a nonzero vector y e C" is a left eigenvector corresponding to an eigenvalue a if Mif (9.1).t so that the scaled eigenvector has norm 1.} The following classical theorem can be very useful in hand calculation.3 (CayleyHamilton). [21]) or directly using elementary properties of inverses and determinants (see. For any A e Cnxn . (Note that the characteristic polynomial can also be defined as det(A. as a matter of convenience. Thus.1 Fundamental Definitions and Properties Fundamental Definitions and Properties Definition 9. for example. The polynomialn (A.3 (CayleyHamilton).31 = 0. then It can be proved from elementary properties of detenninants that if A e enxn .
. If A E A(A) has algebraic multiplicity m. the set of Definition 9. It can be shown that or(l) is essentially unique (unique if we force the coefficient It can be shown that a(Je) is essentially unique (unique if we force the coefficient of the highest power of A to be + 1. A.5. If is a root of multiplicity m of n(A). Theorem If A E 1Ftnxn.AI) = (A] . n(A). 2aA + 2 + ft and Example 9..8..7. From the CayleyHamilton Theorem. we always have A(A) = A(AT). A. The spectrum of A E C"x" is the set of all eigenvalues of A. (9.. Let the eigenvalues of A E en xxn be denoted X\.AI) :::: m.AI). we know that n(A) = 0.:. a . degree such that a (A) =0. if left of A A E A(A). if eigenvectors of A and AT (take Hermitian transposes of both sides of (9.1)2 = O. But it also clearly satisfies the smaller degree polynomial equation (it. However.. possibly repeated. Eigenvalues and Eigenvectors n(A) has n roots. Note.e. But it also clearly satisfies the smaller degree polynomial equation isfies (1 .e. Let a.rank(A . then I < dimA/"(A — A/) < m. The minimal polynomial Of A l::: K""" ix (hI' polynomilll a(A) of least degree such that a(A) ~ O.2... then we must have 1 :::: g :::: m. Then if we write (9. but that A(A) = A(A) only if A E 1Ftnxn. Specifically. such a polynomial is said to be monic and we generally write et (A) as a monic polynomial throughout the text)..25). then A satisfies (Je — I)2 = 0. f3 e R and let A = [~f3 £ ]. Let a. A is said to be defective if it does not have n linearly independent (right or left) eigenvectors. we say that A is an eigenvalue of A Definition 9. Thus. if A = \1Q ®]. then 1 :::: dimN(A . checked eigenvectors of A and AT (take Hermitian transposes of both sides of (9. then we must have I < g < m. . i. then n(X) has real coefficients. The spectrum of A is denoted A(A). such a polynomial is said to be monic and we of the highest power of A to be +1.5. Moreover.5. ft E 1Ft and let A = [ _^ !]. we say that X is an eigenvalue of A of algebraic multiplicity m. the set of all roots of its characteristic polynomial n(X).. less than) its algebraic multiplicity. E A(A). must occur in complex conjugate pairs. eigenvalues of A. then there is an easily checked relationship between the left and right If A € R"x". Xn.7. neftnhion ~. (An . A matrix A e 1Ft x" is said to be defective if it has an eigenvalue whose geometric multiplicity is not equal to (i. For example. Then jr(A.) A.ft Definition 5. Thll minimal polynomial of A G l!if. it can also be . i. then A satsible for A to satisfy a lowerorder polynomial.nxn is the polynomial o/(X) oJ IPll. These roots. i... Thus. and hence further guarantee the existence of corresponding nonzero eigenvectors..3) are the eigenvalues of A and imply the singularity of the matrix A .n =0o.. eigenvalues of A. i •>/—!)• If A E 1Ftnxn. Definition 9. it is posn(A) = O.~.XI. say.8. that by elementary properties of the determinant. geometric multiplicity is not equal to (i.e.. it can also be generally write a(A) as a monic polynomial throughout the text).4) and set A = 0 in this identity.. Equivalently. A is said to be defective if it does not have n linearly independent (right or left) eigenvectors. The spectrum of A e nxn is the set of all eigenvalues of A. if we denote the geometric multiplicity of A by g. For example. sible for A to satisfy a lowerorder polynomial. Equivalently. Example 9. However. An.2aA + aa2+ f322 and A has eigenvalues a f3j (where A has eigenvalues a ± fij (where j = i = R). A matrix A E Wnxn is said to be defective if it has an eigenvalue whose Definition 9. and hence further are the eigenvalues of A and imply the singularity of the matrix A — AI. if A = [~ ~].e. Eigenvalues and Eigenvectors Chapter 9. .A) (9. independent eigenvectors = n . Definition 9.76 Chapter 9. Definition 9. say. we get the interesting fact that det(A) = AI .6.3) in the n(A) = det(A . too.6.2». IfXA is a root of multiplicity m ofjr(X). of we always have A(A) = A(A r ).A) . The geometric multiplicity of A is the number of associated independent eigenvectors = n — rank(A — A/) = dim J\f(A — XI). Hence the roots of 7r(A). c form form A e C" " A]. must occur in complex conjugate pairs.. the n(A) coefficients. Specifically... guarantee the existence of corresponding nonzero eigenvectors. less than) its algebraic multiplicity. as solutions of the determinant equation n(A) = det(A  AI) = 0.AI) = dimN(A . The geometric multiplicity ofX is the number of associated of algebraic multiplicity m.1) .. but that A(A) = A(A) only if A e R"x".e. The spectrum of A is denoted A (A). we get the interesting fact that del (A) = A] • A2 • • An (see also Theorem 9. possibly repeated. y of AT y is a left eigenvector of A corresponding to A e A(A). we denote the geometric multiplicity of A by g. Moreover. then y is a right eigenvector of AT corresponding to I € A (A). if If A € A(A) has algebraic multiplicity m. all roots of its characteristic polynomialn(A). as solutions of the determinant equation 7r(A) has n roots.2)). These roots. Then n(A) = A22. If e Wxn. • AM(see and set X = 0 in this identity.
let Yj be a left eigenvector corresponding to any A. In particular. Let A e C« x " ana [et A. YY Proof: Since Axt = A. of which has an eigenvalue 2 of algebraic multiplicity 4. Example 9. A[~  2 0 I 2 0 0 0 0 0 0 !] ~ ~ ~ ha. each 4.*. such that Aj =£ Ai. eigenvector is numerically unstable.2)4.) directly (without knowing eigenvalues and as Unfortunately. Fundamental Definitions and Properties 9. Then yfx{ = O. every nonzero polynomial f3(A) particular."(A) ~ ~ ~ ~ (A . Bezout algorithm..5) = (A  2)2 and g = 2. The above definitions are illustrated below for a series of matrices. this algorithm. Then Xi = 0.2)2 ""d g 3. 0 0 0 2 A~U 0 0 ] ha<a(A) (A .1. We denote 7r(A) (A .11.. the geometric multiplicity by g. such is not the case. i. a(A) directly (without knowing eigenvalues and asThere is an algorithm to determine or (A. A~[~ A~U 2 0 0 I 2 2 ] ha< a(A) (A . a(A) divides n(A).11. called the Bezout algorithm. Fundamental Definitions and Properties 77 77 a(A) f3(A) O..9. Furthermore. Example 9.10. one might speculate that g plus the degree of a must always be five.2)' ""d g ~ ~ ~ 1. n(A) = (A — 2)4.2) andg ~ 4. The matrix A~U has a(A) I 2 0 0 2 0 0 0 !] (9. Proof' Since AXi AiXi.. . Let A E cc nxn and let Ai be an eigenvalue of A with corresponding right eigenvector jc. Unfortunately. 0 g At this point.2)' ""d g 2.) divides every nonzero polynomial fi(k} for which ft (A) = 0.. shown that a (A.e. g.1. Unfortunately. be an eigenvalue of A with corresponding right Theorem 9. algorithm. 0 0 0 2 0 0 0 2 ] h'" a(A) (A . sociated eigenvector structure). a(X) n(X). Theorem 9. left Aj E A (A) such that Xj 1= A.10. e l\(A) yj Xi..
Let A E C"x" be Hermitian. Then all eigenvalues of A must be real. so that YitH x. However.— A. is If A e C nxn has distinct eigenvalues..11.. c Proof: Suppose (A.e. we find 0 = (Ai . Since equation Az i^z XH get X H Az = iJXH A.7) Taking Hermitian transposes in (9. . AXH z. YyXi =0.. 1 ?.13.7) yields Taking Hermitian transposes in (9.14. for [21. 118]. results. c Proof: Premultiply the equation Ax = A. 0 D Theorem 9. i. Since Xi ^ 0 and would thus have to be 0. from which conclude A. jc.— A y )j^jc..Aj)YY xi. from which we conclude I = A. Take the Hermitian A transpose of this equation and use the facts that A is Hermitian and A is real to get xXHAz == of equation facts A.12. Cnxn have distinct eigenvalues AI.. A.is real.11 is very similar to two other fundamental and important results. A = AH. x) is an arbitrary eigenvalue/eigenvector pair such that Ax = AX. .. since y" A = Similarly. 0 The proof of Theorem 9. or else Xi would be orthogonal to n linearly independent vectors (by Theorem 9... and if Ai E A(A). Since Ai . i. i.5). However. we have that XxH = AXH x.. .e. we must have that XHzz = 0..6) from (9. 's. Then and z must be orthogonal. we can choose the normalization of the Xi'S... . since x is an eigenvector. 's. Take the Hermitian Proof: Premultiply the equation Ax = AX by ZH to get ZH Ax = AZ H x. for i € !1.. since YY A = AjXjyf. yr . so that y H Xi = 1 for each i. Eigenvalues and Eigenvectors yy.13. contradicting the fact that it is an eigenvector. Xi is orthogonal to all y/s for which j =1= i. e A(A). Let A e C"x" be Hermitian and suppose A and /JL are distinct eigenvalues Theorem 9.14) and would thus have to be 0. = A.7. An with corresponding Theorem 9. A.JC by ZH to get ZH Ax = X z Hx . x is a linearly independent set. XH x /= 0. A is real. = 1 for/ E n. Proof: Proof: For the proof see. since x is an Using the fact that A is Hermitian.e.n with corresponding right eigenvectors x\. orthogonal.14. Theorem 9.e. [21..78 Similarly. x) is an arbitrary eigenvalue/eigenvector pair such that Ax = A.are distinct eigenvalues of A with corresponding right eigenvectors x and z.. Since XxHz. the proof see. Let us now return to the general case. for example. 0 If A E cnx " has distinct eigenvalues. Eigenvalues and Eigenvectors Chapter 9..Aj ^ 0.. Premultiply the equation Az = iJZ by x H to get x HAz = /^XHZZ = XxHz. However. xn}} is a linearly independent set. eigenvector. we can choose the normalization of the *. or both. 0 Let us now return to the general case.e.7) yields Using the fact that A is Hermitian. The same result holds for the corresponding left eigenvectors. ^ /z. it cannot be the case that YiH Xi = 0 as well. we find 0 = (A.. (9.5). result holds for the corresponding left eigenvectors. . i. we have that IXHxx = XxHx. the two vectors must be orthogonal. However.=1= 0. p. i.. p. The same right eigenvectors XI. Chapter 9. we have xHX =1= 0. Let A e nxn be Hermitian. we must have Subtracting (9. the two vectors must be orthogonal. x . i. Then [x\.6) Subtracting (9. Let A E cnxn have distinct eigenvalues A . be real..11. Let A E nxn be Hermitian and suppose X and iJ.. is real to get H Az AxH z.e. and if A. or both. D A =1= iJ..xnn • Then {XI. . then by Theorem 9. or the Yi 's. we must have yfxt = O.6) from (9.11 is very similar to two other fundamental and important The proof of Theorem 9.. Then all eigenvalues of A must Theorem 9. .. or the y. Then x and z must be of A with corresponding right eigenvectors and respectively.14) well. Then (9.12. we must have that x H = 0. it cannot be the case that yf*xt = 0 as orthogonal to all yj's for which j ^ i.. A = AH. or else xf would be orthogonal to n linearly independent vectors (by Theorem 9.. contradicting the fact that it is an eigenvector.. Theorem 9. Since A. Let A €. Since yf*Xi =1= 0 for each i. Then Proof: Suppose (A..JC. then by Theorem 9.. 118]. respectively.
let Y — [y\. Similarly. Fundamental Definitions and Properties 9. is expressed by the equation while YiH Xj = oij. . / en. i E!!. Yn] ing right eigenvectors form a matrix X = [XI. 2)(A. solve the linear system (A 21) = 0 to get linear system y\(A + 21) = 0 to get yi This time we have chosen the arbitrary scale factor for YJ so that y f xXI = 1. suppose that the left and be the matrix of corresponding left eigenvectors. solve the linear system (A — (—1 + 2j)I)x2 = 0 to get yi X2 =[ 3+ j ] 3 ~/ . = A. xn].15.. and this then determines the other two Note that one component of XI can be set arbitrarily. y' E !!... .15. Furthermore.8) while y^Xj = 5. xn]. suppose that the left and right eigenvectors have been normalized so that yf1 Xi = 1.11) Example 9.(2)l)xI = 0 to get Note that one component of . solve the 3 x 3 linear system (A — (—2}I)x\ = 0 to get For Ai = 2. Xn) E Wtxn. from which we find A(A) = {—2.*/...9) =A = XAyH yRAX n (9. Furthermore.9.. and this then determines the other two (since dimA/XA — (—2)7) = 1)..3 4A2 9 A.ci can be set arbitrarily. ..22 + 2)" + 5). .nand let the correspondTheorem 9..16.(2)1) = 1). A.10) = XAX.2 + 9)" + 10) = ()" + 2)().. can be written in matrixform as diag(A.. 10) (A. We can now find the right and left eigenvectors corresponding to these eigenvalues. . . Fundamental Definitions and 79 Theorem 9..(1 + 2j) I)x2 = 0 to get For A2 — 1 + 2j.."" yn] be the matrix of corresponding left eigenvectors. corresponding to these eigenvalues.. / e n. This time we have chosen the arbitrary scale factor for y\ so that \ = 1. Finally. j e n.. To get the corresponding left eigenvector y\. An and let the corresponding right eigenvectors form a matrix X [x\. . i E !!. . 1 ± 2j}. Then AJC. Let A e en xn have distinct eigenvalues AI. Then rr(A) = det(A .1. is expressed by the equation yHX = I. . An) e ]Rnxn. i E !!:: Finally. — 1 2 j } ..j.AI) (A.I . 2A. Then AXi = AiXi.. / en.A. For Al = —2. To get the corresponding left eigenvector YI.I = = LAixiyr i=1 (9. . We can now find the right and left eigenvectors which we find A (A) = {2. solve the (since dimN(A . . let A = diag(AJ. from Then n(X) det(A ./) = (A 3 + 4A. Similarly. can be written in matrix form as AX=XA (9. These matrix equations can be combined to yield the following matrix factorizations: These matrix equations can be combined to yield the following matrix factorizations: XlAX and and A (9. solve the linear system (A .1. For A2 = 1 + 2j. .. . solve the 3 x 3 linear system (A .16. Let A E C"x" have distinct eigenvalues A. Let Example 9. let Y = [YI. + 5). let A = right eigenvectors have been normalized so that YiH Xi = 1. Let 2 5 3 3 2 4 ~ ] ...
17. we could have found them instead by computing instead of detennining the Yi'S directly. 4}.2*2 to get Ax^ = ^2X2.=. itit is gtruightforw!U"d to compute straightforward to comput~ X~[~ and and I 0 I i ] 1 x.L 4 !. Proceeding as in the previous example. + 4). XIAX=A= [ 2 0 0 1+2j o 0 Finally.A. A. 3.=. Finally. + 1)(A.±1 4 4 4 l+j .c2 = ^. = I .2 However. Then. we could proceed to solve linear systems as for A2. note that we could have solved directly only for XI and X2 (and X3 = X2). To see this. To see this.!. Now define the matrix X of right eigenvectors: Now define the matrix of right eigenvectors: 3+j 3j 3.17. o 3 Then 7r(A. we could have found them instead by computing XI and reading off its rows. X~l Example 9. instead of determining the j. Then.2 and simply can also note that x$ = X2 and Y3 = Y2./) _(A + 8A from which we find A(A) = {—1. = x2). use the fact that A. Let A = [~ ~ ~] .15 can also be verified. we can also note that X3 =x2' and yi jj.A similar argument yields the result conjugate the equation AX2 — A2X2 to get AX2 A2X2. Eigenvalues and Eigenvectors Chapter 9.!. Proceeding as in the previous example. —4}. we could proceed to solve linear systems as for A.) 19X + 12) = (A. similar argument yields the result for left eigenvectors.( I + 2 j) I) = 0 and nonnalize Y22 so that y"xX2 = 1 to get For A3 = — 1 — 2j.j ] 3+j .15 can also be verified. Let Example 9. for left eigenvectors. —3.L Other results in Theorem 9. + 3)(A. Other results in Theorem 9. use the fact that A33 = A2 and simply conjugate the equation A.AI) = (A33 + 8A 22+ 19A + 12) = (A + I)(A + 3)(A + 4).'s directly. 2 It is then easy to verify that It is then easy to verify that 2 .~q 1 3 2 2 0 2 3 ] 2 ~ y' . Eigenvalues and Eigenvectors Solve the linear system yf (A — (1 + 27')/) = 0 and normalize y> so that yf 2 1 to get Solve the linear system y" (A . For example. we For XT.. However.80 Chapter 9. is from which we find A (A) = {I. Then Jl"(A) = det(A .2j. note that we could have solved directly only for *i and x2 (and XT. For example. det(A .
where A is diagonal. Proof: Suppose (A. namely yH A AyH if and only if Hy)H(T~1AT) = A(T Hy)H. A is diagonalizable). 3. since T is nonsingular. namely the theorem statement follows. ff(x) is a polynomial. Remark 9. say. 4). Theorem 9. I 3 I (. Then suppose XI AX = A. or eX.1. For left eigenvectors we have a similar statement. from the theorem statement follows. A = T0 6 2 has only one right eigenvector corresponding to the eigenvalue 0. or.g. but A = [~ has two independent right eigenvectors associated with the eigenvalue o.txiYiH. D D yHA = XyH ifandon\yif(T(T Hy)H (T.but f(A) does not necessarily have all the same eigenvectors (unless. For example.18. but /(A) does not necessarily the eigenvalues of /(A) (defined as X^o^A") are /(A). The following theorem is useful when solving systems of linear differential equations. x) but not conversely. x) maps to (/(A).g. Then. say. Then. Remark 9. A Theorem 9.I = A(T.9.. For left eigenvectors we have a similar statement. If / is an analytic function (e. etA Ax are Details of how the matrix exponential e'A is used to solve the system x = Ax are the subject solve system i of Chapter 11. Fundamental Definitions and Properties 81 81 We also have XI AX = A = diag(—1. or ex.20.. we have the equivalent statement (T. e jRnxn n = LeA. or sin*. If f is an analytic function (e. x) maps to (f(A). ] [~ ~ (I) [ I (. 2 3 I (. What is true is that the independent right eigenvectors associated with the eigenvalue 0. i=1 . —3. eigenvalue/eigenvector pair (A.18. A is diagonalizable). of Chapter 11.20. Let A E R" xn and suppose X~~1AX — A. which is equivalent to the dyadic expansion sion 3 A = LAixiyr i=1 ~(I)[ ~ W~l+(3)[ j ][~ ~ 1 . I I ~J I 2 0 0 0 3 3 3 I I (. Eigenvalues (but not eigenvectors) are invariant under a similarity transformation T. For example. from which equivalent statement (T~ AT)(T.lIAT)(T~lx)x) = X ( T ~ lIxx). then easy to show that the eigenvalues of f(A) (defined as L~:OanAn) are f(A). jc) is an eigenvalue/eigenvector pair such that Ax = Xx. Eigenvalues (but not eigenvectors) are invariant under a similarity transTheorem 9. Fundamental Definitions and Properties 9. —4).1. A = [~ Oj ] have all the same eigenvectors (unless. formation T. The following theorem is useful when solving systems of linear differential equations. but A2 = f0 0~1]has two has only one right eigenvector corresponding to the eigenvalue 0.) . representable by a power series X^^o fln*n)> then it is easy to show that representable L~:O anxn). J+ (3) [ 2 0 2 I I I 2 I ]+ 3 3 I I 3 I I I 3 3 I (4) [ 3 3 I I 0 3 3 l Theorem 9. or sinx.3 0 ~l +(4) [ . x) but not conversely. which is equivalent to the dyadic expanWe also have X~l AX = A = diag( 1. What is true is that the eigenvalue/eigenvector pair (A.19. in general. X) is an eigenvalue/eigenvector pair such that Ax = AX. since T Proof: Suppose (A. ( x ) is a polynomial.19.1 AT) =X(THyf.
.. we have Chapter 9.. ff(A) = X f ( A ) X ~ l I = Xdiag(J(AI). of course. then e A has eigenvalues e A There are extensions to Theorem 9.. An e C 1.21. If A E Rnx" is diagonalizable with eigenvalues A. /' en.... . I.22. . eigenvectors Xi. Theorem 9. and the same eigenvectors.2 9.Il .82 Proof: Starting from the definition.2 Jordan Canonical Form Jordan Canonical Form Theorem 9.= Xdiag(/(A... and the same eigenvectors. i E ~. . . / € n_. canonical form. from which such a result is then available and presented later in this chapter. . kn E C (not necessarily distinct). and right eigenvectors xt•.20 and Corollary 9.12) where each of the lordan block matrices 1i . .IXiYiH..20 and Corollary 9. It is necessary first to consider the notion of Jordan canonical form. Eigenvalues and Eigenvectors Chapter 9. .. i=1 0 The following corollary is immediate from the theorem upon setting t = I.13) 1i = o o Ai o Ai .21. there exists X E C^x" such that XI AX = 1 = diag(ll. i. there exists X € c~xn such that (not necessarily distinct). 0 (9. . Corollary 9..21 for any function that isis There are extensions to Theorem 9. The following corollary is immediate from the theorem upon setting t = I. 1q is of the form where each of the Jordan block matrices / 1 ••• Jq is of the form Ai 0 Ai Ai o 0 (9.i). 9... i E ~.20 and its corollary in which It is desirable.21 for any function that analytic on the spectrum of A.. € n_.. If A e ]Rn xn is diagonalizable with eigenvalues Ai. to have a version of Theorem 9. ( A ) = Xf(A)X..e.. f(An))X. analytic on the spectrum of A. It is necessary first to consider the notion of Jordan A is not necessarily diagonalizable.. lordan Canonical Form (JCF): For all A E C"x" with eigenvalues X\. from which such a result is then available and presented later in this chapter..22.. i.. we have Proof' Starting from the definition. then eA has eigenvalues e X"i .. and right Corollary 9. Eigenvalues and Eigenvectors n = LeA. to have a version of Theorem 9.20 and its corollary in which A is not necessarily diagonalizable. f ( X t t ) ) X ~ It is desirable. ii E ~. Jordan Canonical Form (/CF): For all A e c nxn with eigenvalues AI. 1q). of course.e.
83 83 Form: 2. 120124]. e A (A). the situation is only a bit more complicated. ] T (X . = [ _»' ^ 1 and h2 = [6 ~] in the case of complex conjugate eigenvalues where Mi = [ _~... and where M. Proof: proof D 0 Transformations like T = [ _~ "•{"]allow us to go back and forth between aareal JCF Transformations like T = I" _.JfJ =[ (X fJ fJ ] (X = M. . pp.. Jordan Canonical Form 9.. 1q is of form where each of the Jordan block matrices 11. Proof: For the proof see.9. . . Real Jordan Canonical Form: For all A E R n x " with eigenvalues AI. (Xii±jpieA(A>)..=1 ki = n. X (not necessarily X E lR. Jq is of the form of in the case of real eigenvalues A. ~: ] and I = \0 A in the case of complex conjugate eigenvalues a ± jfJi E A(A)..... for example. [21. there exists X € R" xn such that (9.An n (not € jRnxn Xi.2. . .2. Jordan Canonical Form and L. complicated..~xn necessarily distinct). For nontrivial Jordan blocks. { ] allow us to go back and forth between real JCF and its complex counterpart: TI [ (X + jfJ o O. . With 1 j o j o 1 o o o j ~ ~] 0 1 ' ..14) J\.
1). D 0 Example 9.1)2.23.. .24. i=1 n 2.7x7 is known to have 7r(A) = (A . Let A e C" " with eigenvalues AI. The characteristic polynomials of the Jordan blocks defined in Theorem 9.(A 1)... . 1)4(A 2) and 2 2 et(A) = (A . An.jf3 0 ]T~[~ l h M Definition 9. Then has two possible JCFs (not counting reorderings of the diagonal blocks): diagonal blocks): 1 J(l) = 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 2 0 0 0 0 0 0 0 0 0 0 0 1 2 0 0 0 0 1 0 0 0 and f2) 0 0 0 0 0 1 0 0 0 0 0 2 = 0 0 0 0 0 0 I 1 0 0 2 0 0 0 0 0 0 0 0 0 0 0 0 0 0 2 0 0 0 0 0 1 0 2 0 J(l) has elementary divisors (A Note that 7(1) haselementary divisors (A .25. The minimal polynomial of a matrix is the product of the elementary divisors of highest degree corresponding to distinct eigenvalues.. Theorem 9. The characteristic polynomial of a matrix is the product of its elementary divisors.25. Eigenvalues and Eigenvectors T. The characteristic polynomials of the Jordan blocks defined in Theorem Definition 9.. from Theorem 9.— I) (A. 2. from Theorem 9. 1. . 2 .2)2. highest degree corresponding to distinct eigenvalues.(A.2)2. J(2) has elementary divisors (A while /( 2) haselementary divisors (A .24.2)2. From Theorem 9.. Thus. .2). . det(A) = det(XJX.22 are called the elementary divisors or invariant factors of A. Then AAhas two possible JCFs (not counting reorderings of the a (A.) = (A. and (A . Tr(A) = Tr(XJX~ ) = TrC/X"1 X) = Tr(/) = £"=1 A. Then c n 1. 1).23. (A1). x Theorem 9. Then Theorem 9.22 we have that A = X JJX ~ l . Suppose A e lR. Tr(A) = 2. 2(A(A. and (A .26.. Thus. The minimal polynomial of a matrix is the product of the elementary divisors of divisors... 1)2(A .)i. Thus. 1 det(A) = det(X J XI) det(J) = n7=1 Ai. Eigenvalues and Eigenvectors Chapter 9.26.2)3 3and is known to have :rr(A) Example 9. The characteristic polynomial of a matrix is the product of its elementary Theorem 9. X XI. I) .1)2. det(A) = nAi.84 it is easily checked that it is easily checked that Chapter 9. 2)2. — 2) . Again.jf3 0 0 et .1)4(A ..22 l Tr(A) = Tr(X J XI) Tr(JX.1*) = Tr(J) = L7=1 Ai.I [ "+ jfi 0 0 0 et + jf3 0 0 0 0 et . 9.2). .1)z. ..22 we have that A X J XI. From Theorem 9.2).(A. Suppose A E E (A.2(A(A . and 2)..22 we have that A = X J X ~ .I)(A (A2)2.) = det(7) = ]~["=l A. I) .22 are called the elementary divisors or invariant factors of A. . Let A E nxn with eigenvalues AI. i=1 l Proof: Proof: 1. " Xn. and (A (A .
If we let [~l ~2 ~3]T associated If [^i £2 &]T denote a solution to the linear system (A — 3l)~ = 0.a(A) = (A .— al) == 4.. determine the JCF of A uniquely.27. X principal Definition 9. For example. a (A). both are eigenvectors (and are independent).ulx = 0 and (A ..AI)klx i= o.e.) = (A.(7). a)\ . Determination of the JCF 9.7) for distinct A.). c .7) = n . it The straightforward case is. we find that 2~2 + £ 3= 0 .3 Determination of the JCF Determination of the JCF The first critical item of information in determining the JCF of a matrix A E Wlxn is its A e ]R.is simple. a(A. i. so the eigenvalue 3 has two eigenvectors associated with it. left k. For each distinct eigenvalue Ai.) = (A.9.3. i. of algebraic multiplicity 1.e. An analogous definition holds for a left principal vector of degree k.rank(A . three eigenboth have rr(A) = (A .A. we find that 2£2 + ~3 = O. Determination of the JCF 85 &5 Example 9. a(A). is not sufficient to Example 9.l) independent right — — A. determine a 0 0 0 0 0 0 0 a 0 0 0 0 0 a 0 0 0 0 Al= 0 0 0 a 0 0 0 0 0 0 a 0 0 0 0 0 0 0 a 0 0 0 0 0 0 1 a A2 = a 0 0 0 0 0 0 0 a 0 0 0 0 0 a 0 0 0 0 0 0 0 a 0 0 0 0 0 0 a 0 0 0 0 0 0 a 0 0 0 0 0 0 0 a 4.— a) and rank(A .28. Then x is a right principal vector of degree k degree associated with A E A (A) ifand only if(A . of course. A e nxn ]R. i. and rank(A al) vectors. i. The straightforward case is. Thus.. Knowing TT (A.29.e. of course.27. when Ai is simple. eigenvectors dimN(A — A. Remark 9.. is of algebraic multiplicity greater than one. and rank(A —Ai l) for distinct Ai is not sufficient to rr(A). the associated number of linearly A. both denote a solution to the linear system (A .A.] are eigenvectors (and are independent). Definition 9. Remark 9. associated independent right (or left) eigenvectors is given by dim A^(A .nxn). Let A E C"xn (or R"x").29.. when X. Thus. suppose suppose A = [3 2 0 o Then Then 3 0 A3I= U2 I] o o 0 0 n has rank 1. To get a third vector JC3 such that X [x\ X2 XT. and rank (A A. it then has precisely one eigenvector.— a) .l). 1. To get a third vector X3 such that X = [Xl KJ_ X3] reduces A to JCF.3/)£ = 0. of algebraic multiplicity 1. The matrices A uniquely.3. 9. a)7. The more interesting (and difficult) case occurs when Ai multiplicity A.. X e A(A) if and only if (A XI)kx = 0 and (A U}k~l x ^ 0.. three eigen7r(A... we need the notion of principal vector.nxn number of eigenvectors.28..e.
Principal vectors are sometimes also called generalized eigenvectors. of — AI.17) by (A . of k 5. k = eigenvector. which simply says that x(!) is a right Ax(1) = hx(1) x (1) (2) x(2).86 Chapter 9. combination of jc(1) vectors to get a righthand side that is in 7£(A — XI). Solve (A .AI).XI. for get righthand example. solve (A .XI)0 = 0. The other solution necessary is the desired principal vector of degree 2. A right (or left) principal vector of degree k is associated with a Jordan block J. The phrase "of grade k" is often used synonymously with "of degree k. For example. but the latter generalized eigenvectors. of course. Eigenvalues and Eigenvectors synonymously "of 2. The case k = 1 corresponds to the "usual" eigenvector. Thus. the definition of principal vector is satisfied.A1)2 x(2) = (A . Then for each distinct X e A (A) perform the following: z (2) w c 1. is. Denote by x(l) and x(2) the two columns of a matrix X E lR~X2 2x2 2 Jordan block{h0 h1. the principal vector second of degree 2: of degree (A .) . 9. there are two linearly independent n — o.1 9.A/)x(2) = x(l).A/) = — multiplicity of rank(A — XI) = n .X I ) . I) associated This step finds all the eigenvectors (i.AI). principal vectors still need to be computed from succeeding steps.e.17) The first column yields the equation Ax(!) = AX(!).X I ) ( l ) = (A AI)O o.XI)2x^ = 0. E lR nxn This suggests a "general" procedure. x(l). One of these solutions (A — AI)2 x (2) x(l) (1= 0).. consider a determining 2 x Jordan [~ i]. If we premultiply (9. since (A . solutions solutions to the homogeneous equation (A . A E A(A) following: (or C ). for example.1 Theoretical computation Theoretical computation To motivate the development of a procedure for determining principal vectors. Then the equation AX = XJ can be written that reduces a matrix A to this JCF. If the algebraic multiplicity of If A principal need X is greater than its geometric multiplicity.3.A1)X(l) = O. Exercise 7. The second column yields the following equation for x . See.3.'A1)22xx(l) = (A . (It may be necessary to take a linear of x(l) R(A . by (A .x2 A JCF. 2. See.2." "of often 3. Thus. principal vectors of degree 1) associated with A. Denote by x(1) and x(2) the two columns of a matrix X e R2. 4. different term will be assigned a much different meaning in Chapter 12. determine all eigenvalues of A e R" x " nxn ). of The number of linearly independent solutions at this step depends on the rank of 2 (A . if of .1. there is only one eigenvector. For each independent jc (1) . ji of dimension k or larger. (9. (A — uf. S.A1)X(l) = O. eigenvector. First. The number of eigenvectors depends on the rank of A . If. this rank is n .A1)x(2) = x(l). Theother solution (A . we find (A If we premultiply XI) x = (A XI)x = 0. Eigenvalues and Eigenvectors Chapter 9. wefind(A. Then the equation AX = X J can be written A [x(l) x(2)] = [x(l) X(2)] [~ ~ J. x(l) (^ 0). if X.XI).
{x(l). this naturallooking procedure can fail to find all Jordan vectors. for example.AI) = k . .. (x (1) . Suppose A e Ckxk has an eigenvalue A. Let X = x ( l ) . x(k)]. say). pendent. [20] and [21]. .. and the interested student is strongly urged to consult the classical and very readable [8] to learn why. Theorem 9. 4. Continue in this way until the total number of independent eigenvectors and principal vectors is equal to the algebraic multiplicity of A. Symbolic Toolbox. Let X = [[x(l). [20] and [21]. Continue in this way until the total number of independent eigenvectors and principal 4. Then Theorem 9. Determination of eigenvectors and principal vectors is obviously very tedious for anything beyond simple problems (n = 2 and principal vectors is obviously very tedious for anything beyond simple problems (n = 2 or 3. First. Suppose A E C kxk has an eigenvalue A of algebraic multiplicity kkand suppose further that rank(A — AI) = k — 1. although a j ardan command is available in MATLAB'S does not offer a jcf command. . Theorem 9. where the chain of vectors x(i) is constructed as above. solve 3.. Unfortunately. x (k) } is a linearly independent set. see.31.9. Let Example 9. .. where the chain of suppose further that rank(A . see.3.3. Notice that highquality mathematical software such as MATLAB readable [8] to learn why. A2 = 1. . For each independent X(2) from step 2. . Example 9.. and A3 = 2. solve (A AI)x(3) 87 = x(2).32.1. . 0 The eigenvalues of A are AI = 1. Theorem 9. this naturallooking procedure can fail to find all Jordan vectors.2I)x~1) = 0 yields . of algebraic multiplicity and Theorem 9. For each independent x(2) from step 2. say). First. Attempts to do such calculations in finiteprecision floatingpoint arithmetic generally prove unreliable. although a jordan command is available in MATLAB's Symbolic Toolbox. . Determination of the JCF 3. X(k)} is a linearly independent set.(1) (A . X(k)]. Attempts to do such calculations in finiteprecision floatingpoint arithmetic or 3. . Determination of eigenvectors more extensive treatments. and h3 = 2. with the distinct eigenvalues 1 and 2. Principal vectors associated with different Jordan blocks are linearly indeTheorem 9.. find the eigenvectors associated The eigenvalues of A are A1 = I.33. vectors is equal to the algebraic multiplicity of A.32.2/)x3(1)= 0 yields (A . and the interested student is strongly urged to consult the classical and very to compute a JCF. Notice that highquality mathematical software such as MATLAB does not offer a j cf command.30. Determination of the JCF 9.33. There are significant numerical difficulties inherent in attempting generally prove unreliable. For Unfortunately.31. There are significant numerical difficulties inherent in attempting to compute a JCF.30.. For more extensive treatments. h2 = 1. 1 . for example. Let A=[~ 0 2 ] . Then vectors x(i) is constructed as above. find the eigenvectors associated with the distinct eigenvalues 1 and 2. . Principal vectors associated with different Jordan blocks are linearly independent.
. d n)) be a nonsingular "scaling" matrix. we 1 's but can be arbitrary — so long as they are nonzero.. consider below the case of a single Jordan block.(2) = x. . For the sake of defmiteness. 0 1 = [xiI) 0 xl" xl"] ~ [ ~ 5 ] and XlAX 5 3 0 Then it is easy to check that Then it is easy to check that l 1 X'~U i 1 =[ ~ I 0 0 n 9.1I)xl ) = xiI) to get (A – l/)x. (1) toeet x. 0 !b.2 On the +1 's in JCF blocks 's JCF In this subsection we show that the nonzero superdiagonal elements of a JCF need not be In this subsection we show that the nonzero superdiagonal elements of a JCF need not be 1's but can be arbitrary .. . For the sake of definiteness. Then Let D = diag(d" . solve 2 (A . but the result clearly holds for any JCF. =0 yields (1) Chapter 9. dn be a nonsingular "scaling" matrix. d.3.2 9. Now let Now let X (2) =[ 0 ] ~ . we consider below the case of a single Jordan block. 0 = 0 A dn dn I 2 0 dn dn I A 0 ).11)x?J = 0 yields (A.... d. .l/)x. Suppose A € Rnxn and SupposedA E jRnxn and Let D diag(d1.so long as they are nonzero. Then A 4l.3. 0 0 D'(X' AX)D = D' J D = j ). solve To find a principal vector of degree 2 associated with the multiple eigenvalue 1. but the result clearly holds for any JCF. Eigenvalues and Eigenvectors To find a principal vector of degree 2 associated with the multiple eigenvalue 1..88 (A .. .
dimN(A — AJ)Vi = ni.n = N(A = N (A .. Then jH. Such a decomposition is given in the following theorem. Such a decomposition is given in the following associated direct sum decomposition of jH. Let V be a vector space over F and suppose A : V —>• V is a linear Definition 9. (A . Note that dimM(A .A1I) v) E6 .Am)Vm with Ai...Amtm c and minimal polynomial a(A) = (A .n. where AS is defined as the set {As:: s e S}.. Theorem 9.AlIt) E6 .34... Suppose A E R"x" has characteristic polynomial 9.. Then AI. . Specifically.A[)V) '" (A .AmItm .. .4.. set {As s E S}. Let IF and suppose + transformation. In a similar fashion.. Suppose e jH.. Geometric Aspects of the JCF 89 di's Appropriate choice of the di 's then yields any desired nonzero superdiagonal elements.. . Geometric Aspects of the JCF 9..4 Geometric Aspects of the JCF Geometric Aspects of the JCF The matrix X that reduces a matrix A E IR"X"(or C nxn)) to aalCF provides aachange of basis X e jH..18) 0 I 1 0 0 can be used to put the superdiagonal elements in the subdiagonal instead if that is desired: to superdiagonal elements in instead desired: A I 0 0 A 0 A 0 A 0 0 A 0 p[ A p= 0 1 0 0 A 0 I A A 0 0 0 A 9. dnxn}..Am I) Vm .4. interpreted This result can also be interpreted in terms of the matrix X = [x\.nxn n(A) = (A .A[)n) . Am distinct.35. mdistinct.A. the reverseorder identity matrix (or exchange matrix) 0 p = pT = p[ = 0 I 0 (9. .. ../) w = «.9. the reverseorder identity matrix (or exchange matrix) In a similar fashion. J is obtained from A via the similarity transformation XD = \d\x\... A subspace S c V is A invariant if AS c S. similarity transformation XD [d[x[. Specifically. . x n eigenvectors and principal vectors that reduces A to its lCF. j is obtained from A via the and principal vectors that reduces A to its JCF..4 9. A subspace S ~ V is Ainvariant if AS ~ S.nxn (or nxn to JCF provides change of basis with respect to which the matrix is diagonal or block diagonal. It is thus natural to expect an with respect to which the matrix is diagonal or block diagonal.xn]] of eigenvectors = [x[. It is thus natural to expect an associated direct sum decomposition of R. . E6 N (A . Definition 9. dnxn]. E6 N(A .. where AS is defined as the transformation.. ...34. A..35.
. € C"x"' be a Jordan basis for N(AT — A. 9. We would then get a block diagonal example (note that the power ni could be replaced by Vi).38.19) the columns attention here to only the Jordan block case. S is Ainvariant if and only if S . but we restrict our attention here to only the Jordan block case. R(S) == S..A. so the columns of Xi span an Amvanant subspace. . The equation Ax Example 9.34. i. so by (9.34.g.. /. Equivalently. Then N(p(A)) and R(p(A)) 7£(p(A)) are Ainvariant.• EB Nm. Eigenvalues and Eigenvectors Chapter 9. = X. A".) span an Ainvariant of A".19): /th Example 9. 7.li.is A T invariant.37. 9. If A has distinct eigenvalues A. i. Rewriting in the form ~ J. Eigenvalues and Eigenvectors If V is taken to be ]Rn over Rand S E ]Rn x* is a matrix whose columns SI. Let peA) = «o/ + o?i A + • + <xqA be a polynomial in A..e."" Jik.). i /= 1. diagonal representation. //*.e. XI AX = [~ J 2 ]......39. K(S) <S..e.* is a Jordan block corresponding to Ai E A(A). be a Jordan basis for N (AT . each Ji = diag(JiI.. is Ainvariant. . then S <S is Ainvariant if and only if there exists M E ]Rkxk such that eRkxk (9. Then N(p(A)) and 1. This follows easily by comparing the ith columns of each side of (9. We could also use other block diagonal decompositions (e. Jm). s/t span a /^dimensional subspace <S. = Xi. Let Yi E <enxn .19) AS = SM. Suppose A E ]Rnxn. partition .. eigenvalues Ai 9. Suppose A"== [Xl .... Other representation for A with full blocks rather than the highly structured Jordan blocks. Jm).. . where X [ X i ..Ji .39. If V is a vector space over IF such that V = N\ EB . Suppose X block diagonalizes A.. then a basis for V can be chosen with respect to which A has a block N. We would then get a block diagonal representation for A with full blocks rather than the highly structured Jordan blocks. Ainvariant... where each Theorem 9.. e A(A).. we have that A A. = 1.) and each Jik is a Jordan block corresponding to A. could be replaced by v. where each Ji = diag(/. /.. Xm R"n such that XI AX diag(7i. so by (9. e E"x"..) and each /.38. Let p(A) = CloI + ClIA + '"• •+ ClqAqq be a polynomial in A..as in Theorem 9. AT Theorem 9.../)"'.. partition Equivalently.. . i. the eigenvectors and principal vectors associated with Ai) span an Ainvariant subspace of]Rn. so the columns of A.. The equation Ax = A* = x A defining a right eigenvector x of an eigenvalue AX = x A defining a right eigenvector x of an eigenvalue A x X says that * spans an Ainvariant subspace (of dimension one).. Note that AXi = A*. (i. the eigenvectors and principal vectors associated with A.Xm] ] Ee]R~xnxnisis such that X^AX ==diag(J1.lt.. then a basis for V can be chosen with respect to which A has a block diagonal representation. and S e R" xk s\. we could choose bases for N(A — A.span an Ainvariant subspace... then is Ainvariant if and only if there span a kdimensional subspace S.90 Chapter 9. Example 9. A invariant if only ifS1 1. If F = NI ® • • 0 m A// is Ainvariant.Ai/)n. . Sk If R" R..36../)"' by SVD.19) the columns of Xi (i. If A has distinct The Jordan canonical form is a special case of the above theorem. we have that AXi Theorem 9. . we return to the problem of developing a formula for e l A in the case that A A formula e' A is not necessarily diagonalizable.e..2. of W. Finally. via SVD). example (note that the power n. Other such "canonical" forms are discussed in text that follows. for N(A . Note that A A".2. Let 7.i. .e. 2.37. The Jordan canonical form is a special case of the above theorem.36. . such "canonical" forms are discussed in text that follows. is not necessarily diagonalizable.
40. ifRe(z) < O. Definition 9. Then the sign of z is defined by Re(z) {+1 sgn(z) = IRe(z) I = 1 ifRe(z) > 0.YiH. m ••• . ... Then the sign of A. with N containing all Jordan blocks corresponding to the be a Jordan canonical form for with N containing all Jordan blocks corresponding to the eigenvalues of in the left halfplane and P containing all Jordan blocks corresponding to eigenvalues of A in the left halfplane and P containing all Jordan blocks corresponding to eigenvalues in the right halfplane. i=1 H In a similar fashion we can compute m etA = LXietJ. Then A = XJX.5 9. The Matrix Sign Function 91 91 compatibly. Ym]H = LX.. and let e cnxn be a Jordan canonical form for A. denoted sgn(A). Jm) [YI . is given by sgn(A) = X [ / 0] 0 / X I .41.. .5. Let z E C with Re(z) ^ O. 9. Suppose A E C"x" has no eigenvalues on the imaginary axis. It is a generalization of the sign (or signum) of a scalar. . Definition 9.S. The Matrix Sign Function 9.= Ai. Then the sign of A. . i=1 which is a useful formula when used in conjunction with the result which is a useful formula when used in conjunction with the result A 0 A A 0 eAt teAt eAt . A called the matrix sign function.JiYi .9. Definition 9.lt 2 e At 2! 0 exp t 0 0 0 1 A teAt eAt 0 0 0 0 0 block Ji associated A = A. Xm] diag(JI.40. of defined Definition 9.. A survey of the matrix sign function and some of its applications can be found in [15]. is given by eigenvalues in the right halfplane.5 The Matrix Sign Function The Matrix Sign Function section brief interesting useful In this section we give a very brief introduction to an interesting and useful matrix function function called the matrix sign function. Then compatibly. .I = XJy H = [XI. denoted sgn(A).41. E f= 0. It is a generalization of the sign (or signum) of a scalar.. associated with an eigenvalue A. for a k x k Jordan block 7.
5. Let e C" be an arbitrary vector. sgn(AH) = (sgn(A))".42. Xn and left eigenvectors y\. negA == (l . distinct right eigenvectors Xi. 2. sgn(cA) = sgn(c) sgn(A) for all nonzero real scalars c. The JCF definition of the matrix sign function does not generally lend itself to reliable computation on a finitewordgenerally itself length digital computer. Their straightforward proofs are left to the exercises. Let A E Cnxn have distinct eigenvalues AI. sgn(T1AT) Tlsgn(A)TforallnonsingularT e enxn 6. respectively..S) /2 is a projection onto the negative invariant subspace of A. 5. respectively. its reliable numerical calculation is an interesting topic in calculation its own right. . Then the following hold: following e 1.. Find the appropriate expression for v as a linear combination expression of the left eigenvectors as well. Show that v can be expressed (uniquely) as a linear combination arbitrary vector. 3. S2 = I. projection subspace of 4. and let — sgn(A).n 1. . e nxn Theorem 9. Eigenvalues and Eigenvectors where the negative and positive identity matrices are of the same dimensions as N and p.. of A (the negative invariant subspace)... S = sgn(A). 6. 4.43. Show that v can be expressed (uniquely) as a linear combination e of the right eigenvectors. S2 = I. respectively. sgn(A") = (sgn(A»H. In fact. 7l(S l) is an Ainvariant subspace corresponding to the left halfplane eigenvalues left halfplane I. AS = SA. 3...92 92 Chapter 9.. R(S+/) is an Ainvariant subspace corresponding to the right halfplane eigenvalues R(S + l) A invariant halfplane of (the positive invariant of A (the positive invariant subspace). Theorem 9. Yn. ••.. sgn(TlAT) = T1sgn(A)T foralinonsingularT E C"x". sgn(cA) = sgn(c) sgn(A)/or c. AS = SA. but the one given here is especially useful in deriving many of its key properties. Their left exercises..xn and left eigenvectors Yl. e C"x" Theorem 9. ••• . 2. Theorem 9. . There are other equivalent definitions of the matrix sign function.42. posA == (l + S)/2 is a projection onto the positive invariant subspace of A. ). and let = sgn(A). . S = sgn(A). S is diagonalizable with eigenvalues equal to del.. positive = (/ + of A. 2. positive of P. 3. Suppose A E enxn has no eigenvalues on the imaginary axis. We state some of the more useful properties of the matrix sign function as theorems. Xn with corresponding right eigenA e nxn ). yn.. The JCF definition of the here is especially useful in deriving many of its key properties.1> .. Let v E en be an vectors Xl. EXERCISES EXERCISES 1. R(S — /) Ainvariant of (the negative invariant subspace). Eigenvalues and Eigenvectors Chapter 9. 4.. Suppose A E C"x" has no eigenvalues on the imaginary axis.43. but the one given There are other equivalent definitions of the matrix sign function. Then the following hold: following 1. negA = (/ — S)/2 3. . ± 1. . We state some of the more useful properties of the matrix sign function as theorems.
Determine the JCF of A. Let 7. What are the eigenvalues of this slightly perturbed matrix? matrix? . y e R" are nonzero vectors with A E lR.l]r as an eigenvector.22. eigenvalues. Suppose A e rc nxn is Hermitian. 3}. Determine the JCFs of the following matrices: 6.e. Let A E lR.30 must be 8. i. where x. Suppose the small number 10. n are nonzero vectors with with xTTyy = 0. Let A = [H 1]· 2 2" Find a nonsingular matrix X such that XI AX = J. y e R" are nonzero vectors 10. Suppose A E C"x" is skewHermitian. if A is skewHermitian. 10. Determine all possible € R 5x5 {2. Prove the same result right eigenvector x. nxn be of the form A = / + xyT. Show that x is also a left eigenvector for A.1) element of J. What are the eigenvalues of this slightly perturbed is added to the (16.22. Suppose 10~16 is added to the (16. AH = —A. Suppose a matrix A E lR. k .. (A — I)x(2) x(1) 8. where x. Let A be an eigenvalue of A with corresponding 3.1) element of J. Suppose A € rc nxn is skewHermitian. 2. Show that all right eigenvectors of the Jordan block matrix in Theorem 9. The vectors [0 1 Ifand[l 0 of [0 — l] r and[1 0]r (2) (1) are both eigenvectors. 5. Determine the JCF of A. Let A e R" xn be of the form A = 1+ xyT. x O. Let A be an eigenvalue of A with corresponding right eigenvector x. = O. JCFs for A.Exercises 93 93 2. Prove that all eigenvalues of a skewHermitian matrix must be pure imaginary. i.30 must be multiples of el e R*. Determine the eigenvalues. y E lR. Prove that all eigenvalues of 2. Determine the JCF of A. Prove the same result if A is skewHermitian. Show that x is also a left eigenvector for A. where J is the JCF Find a nonsingular matrix X such that X AX = J. Determine the JCFs of the following matrices: <a) Uj n 2 1 2 =n 7. eigenvectors and if and (real) JCFs of the following matrices: (a) 2 1 ] 0 ' [ 1 6.5x5 has eigenvalues {2.nxn A = xyT. Determine the JCF of A. 4. Suppose A E C"x" is Hermitian. 5.16 Jordan form specified 9. 3. AH = A. Show that all right eigenvectors of the Jordan block matrix in Theorem 9. multiples of e\ E lR. 16x 16 has eigenvalues at 0 its JCF consists of single Jordan block of the form specified in Theorem 9./)jc = x can't be solved. y E lR. right eigenvectors and right principal vectors if necessary. Characterize all left eigenvectors.. where x. 11. where J is the JCF 1 J=[~ 0 1~]. Suppose a matrix A E R 16x 16 has 16 eigenvalues at 0 and its JCF consists of a single A e lR. 9. ~ 0 Hint: Use[1 1 — I]T an Hint: Use[— 1 1 .e. x.n T T x yy = 0. a skewHermitian matrix must be pure imaginary. JCFs for A. 2. Let A e R"x" be of the form A = xyT. Characterize all left eigenvectors. but then the equation (A . 3}. 2.
Find a matrix equation that X must satisfy for this to be possible. Prove that every matrix e jRn xn is similar to its transpose and determine a similarity transformation explicitly. The transformation P in (9.42. Thus. Prove Theorem 9. Then = ( X SIXT)(X. Suppose Au =1= 0 and that we we e jRnxn and All e jRkxk 1 < ::s: n. Prove that 17. where SI and £2 are real symmetric matrices and one of them. Show that every matrix A e jRnxn can be factored in the form A Si$2. Show that every matrix A E R"x" can be factored in the form A = SIS2.S2X~l) would required symmetric factorization of A. Hint: Use the factorization in the previous exercise.42. and S2 are real symmetric matrices and one of them. Suppose Al2 ^ and want to block diagonalize A via the similarity transformation want to block diagonalize A via the similarity transformation where X E IRkx(nk). it suffices to prove the result for the required symmetric factorization of A. in terms of All and A22. Prove that every matrix A E W x" is similar to its transpose and determine a similarity 13. Then A = (XS i X T ) ( X ~ T T S2XI) would be the the "symmetric factorization" of J. If n = 2 and k = 1. sgn(A) = /. 14. in terms of AU and A 22. Prove that 17. it suffices to prove the result for the JCF. 15..18) is useful. about when the equation for X is what can you say further. Hint: Suppose A = Xl XI is a reduction of A to JCF and suppose we can construct Hint: Suppose A = X J X ~ l is a reduction of A to JCF and suppose we can construct the "symmetric factorization" of 1. xn has all its eigenvalues in the left halfplane. Prove Theorem 9. is nonsingular. Hint: Use the factorization in the previous exercise. X e R*x <«*). about when the equation for is solvable? solvable? 15. en . Find a matrix equation that X must satisfy for this to be possible. If n = 2 and k = 1. Prove Theorem 9. i. The transformation P in (9. Eigenvalues and Eigenvectors 12.e. Eigenvalues and Eigenvectors Chapter 9. TIAT = [A011 A22 0 ] . Prove Theorem 9.43. what can you say further. 13. say S1. 16. Suppose A e sgn(A) = 1. transformation explicitly. Consider the block upper triangular matrix 14. Thus. Consider the block upper triangular matrix A _ [ All  0 Al2 ] A22 ' where A E M"xn and An E Rkxk with 1 ::s: k < n.18) is useful. 16. is nonsingular.94 Chapter 9. where Si 12. Suppose A E C"xn has all its eigenvalues in the left halfplane. JCF.43. say Si.
95 95 . . ..1. the transformation A f+ PAPI is called aasimilarity.1 Some Basic Canonical Forms Some Basic Canonical Forms Problem: Let V and W be vector spaces and suppose A : V + W is a linear transformation. respectively). the transformation A f+ P ApT is called 2.. !] Theorem 10. What other U HAU = D.. = H E en xn exists a unitary matrix X such that X H AX = D = diag(Al. the transformation A H> PAP" 1 is called similarity. If P"1 . the for real scalars a and h... such as A = [ _ab ^1 for real scalars a and b. AAHH = AH A).. We can also consider the case A E emxn and unitary equivalence if P and <2 are unitary. . if A e Rmxn ..j. the definition. If W = V and <2== p. most "diagonal" we can get is the JCF described in Chapter 9..j. and unitary matrices (and their "real" counterparts: symmetric. This is proved in Theorem 10. .Chapter 10 Chapter 10 Canonical Forms Canonical Forms 10. it is called an orthogonal equivalence if P and Q are orthogonal matrices. Normal matrices include Hermitian." The transformation A f+ P AQ is called an equivalence. where D = diag(A.e.2. AA = AHA). Then there AI.. . Two special cases are of interest: Two special cases are of interest: 1. Q are unitary.. and unitary matrices (and their "real" counterparts: symmetric.. orthogonal equivalence if P and are orthogonal matrices. Remark 10... This is proved in Theorem 10. What other matrices are "diagonalizable" under unitary similarity? The answer is given in Theorem matrices are "diagonalizable" under unitary similarity? The answer is given in Theorem 10. . such as A = [_~ most "diagonal" we can get is the JCF described in Chapter 9." The transformation A M» PAQ is called an equivalence. skewsymmetric. Xn) (the columns ofX are orthonormal eigenvectors for A). If W = V and if Q = PT is orthogonal. then there exists a unitary matrix £7 such that A = AH E en xxn has eigenvalues AI. . Normal matrices include Hermitian. Xn Theorem 10.2.I. . where it is proved that a general matrix A e enxn is unitarily similar to a diagonal matrix if and only if it is normal (i. An. find P e lR. if A E IR mxn find E R™ and Q E lR~xn such that P AQ has a form...An. . where it is proved that a general matrix A E C"x" is unitarily similar to a diagonal 10. .2. matrix if and only if it is normal (i. skewskewHermitian. An).:xm and Q e Rnnxn such that PAQ has a "canonical form. . The following results are typical of what can be achieved under a unitary similarity. Let A = AH e C"x" have (real) eigenvalues A.9. If A = A H 6 C" " has eigenvalues AI.1. .. the transformation A i» PAPT is called If an orthogonal similarity (or unitary similarity in the complex case). and orthogonal. A.1 10. n ). V and Q 1. an orthogonal similarity (or unitary similarity in the complex case)." In matrix terms. An) (the columns of X are exists a unitary matrix X such that XHAX = D = diag(A. and orthogonal. If The following results are typical of what can be achieved under a unitary similarity. . If a matrix A is not normal.e. Problem: Let V and W be vector spaces and suppose A : V —>• W is a linear transformation.. . If a matrix A is not normal. . We can also consider the case A e Cm xn and unitary equivalence if P and Remark 10.2. then there exists a unitary matrix U such that UH AU — D.. respectively). it is called an "canonical form.. orthonormal eigenvectors for A). as well as other matrices that merely satisfy the symmetric. skewHermitian. = V and if pT is orthogonal. An.. . Find bases in V and W with respect to which Mat A has a "simple form" or "canonical Find bases in V and W with respect to which Mat A has a "simple form" or "canonical xm form. as well as other matrices that merely satisfy the definition. where D = diag(AJ." In matrix terms.9. .1.
.. Xk H Hk.96 96 Chapter 10... xn] = [x\ ] [XI X22] is unitary.3..k rows of the unitary matrix U... [£i. . k = For simplicity.. . An.. A. VI € Cnxk [Xi U ] Proof: Let X\I = [x\. The construction can actually be performed orthogonal frequently [x\ 2 quite easily by means of Householder (or Givens) transformations as in the proof of the Householder transformations proof following general result. HdxI. We also get 0 in the (2. xn] = 1.. XH AX induction noting that XH AX is Hermitian. Write V H matrix such that V X I = [ ~].. .l)block by Al (2.. When combined with the fact that In (l0.. Let X\ e Cnxk have orthonormal columns and suppose U is a unitary Theorem 10. where R E kxk is upper triangular. .. D The construction called for in Theorem 10. I)block x"xi = 1... k 1. Construct a sequence of Householder matrices (also known Proof: Let X [XI. we get 0 in the (l. [Xi f/2] unitary. %n] XI .2)block by XI Xz. n . ..I)block. ~nf. [XI U2] is unitary. . We illustrate the construction of the necessary Householder matrix for k — 1. Hk in the usual way (see below) such that Hk . . . D 0 (2.. When combined with the fact that x~ XI = 1. xf*x\ = Proof' Let x\ be a right eigenvector corresponding to X\. Canonical Forms Chapter 10. .... . Then VH = / / .. . X = Given a unit vector x\ E JRn. Xk are orthonormal).. orthogonal (l.) [XI X2]] is orthogonal is frequently required. we consider the real case. we get Ai remaining in the (l. . . (/ € k) U2 X i U2 = Xi . . .1) we have used the fact that AXI = k\x\... The proof is completed easily by induction upon noting proof that the (2. Hk as elementary reflectors) H\.. the construction of X2 E JRnx(nl) such that X — z e ]R" ("... Let the unit vector x\ be denoted by [~I.2) Al X~AX2 XfAX 2 0 Al ] 0 XfAX z 0 l In (10.3. Construct a sequence of Householder matrices (also known HI. xd = [ ~ l U = where R is upper triangular (and nonsingular since x\. .k But the latter are orthonormal since they are the last n . Now XHAX =[ xH I XH ] A [XI 2 X 2] =[ =[ =[ x~Axl XfAxl X~AX2 XfAX 2 ] (10.2 is then a special case of Theorem 10. Write UH = [U\ U ] [VI Vz] 0 2 with Ui E Cnxk .• • Hk and Hk'" HI.. .2)block must have eigenvalues A2. —k U. Let XI E Cnxk have orthonormal columns and suppose V is a unitary matrix such that UX\ = \ 1. . Let V = XI. .xd.. 10.1 additional vectors x2. (l..1) (10. simplicity. X 1 XI e E".2)block noting that x\ is orthogonal to all vectors in X2. 0 Thus..Hv.3 called Theorem 10. Xn such that [x\. Canonical Forms Proof: eigenvector corresponding AI...2).3 for k = 1.. Xk]... Then there exist n — 1 additional vectors X2. Then [XI V 2] is unitary. where R € Ckxk is upper triangular.1) we have used the fact that Ax\ = AIXI.. Thus..l)block.2)block X2 . .. . Then there exist n . xn such that X = (XI.T.. following general result.. and normalize it such that x~ XI = XI 1. .. In (10. . Then U = HI'" Hk and H Then x^U2 = 0 (i E ~) means that xf is orthogonal to each of the n — k columns of V2.
it is easily verified that UT U = 2 ± 2'. To see that U effects the U symmetric U U = U = I..i.) — xiXt = i j since xT Xi = 1. Theorem 10.. where u ^UU [t\ 1.It can easily be checked — 2uu+ — u u T . n A = LAiPi. [7].2 is worth stating separately since it is applied frequently in applications. i=l theoretical The following pair of theorems form the theoretical foundation of the doubleFrancisdoubleFrancisQR algorithm used to compute matrix eigenvalues in a numerically stable and reliable way.2 for Hermitian matrices) can be written from Theorem 10.1 ± 1. An).4..•» '.2. it is easily verified that u T u = ± 2£i and u T Xl = 1 ± '.Xn.2uu+ = I . . including the choice of sign and the complex case. so U is orthogonal. . An. Let A = AT e E nxn have eigenvalues k\.. 's). U effects necessary compression of jci. Thus.4. [25].nf. ... .4 implies that a symmetric matrix A (with the obvious analogue from Theorem 10.1. sponding to the Ai'S). [11].2 is worth stating separately since it is applied fre10. . quently in applications. [23]. . consulted standard numerical linear algebra can be consulted in standard numerical linear algebra texts such as [7]. In fact. (onto the onedimensional eigenspaces correPi onedimensional eigenspaces sponding to the A. • • . i=1 (10.2 for Hermitian matrices) can be written n A = XDX T = LAiXiXT.e.1. . U orthogonal. Then there exists an AT E jRnxn have eigenvalues AI.10. £2. necessary compression of Xl.3) is actually a often weighted sum of orthogonal projections P. Some Basic Canonical Forms 97 Then the necessary Householder matrix needed for the construction of X 2 is given by Then the necessary Householder matrix needed for the construction of X^ is given by U = I . [11].e.. . Further details on Householder matrices.. = PUM = xixf = xxixT since xj xi — 1. £«] r It can checked T 2 that U is symmetric and U TU = U 2 = I. i. [23]. .+uu T ... . The real version of Theorem 10. where Pi = PR(x. Then there exists an 10.1 and UT X\ = 1 ± £1. [25]. x where P. '. Some Basic Canonical Forms 10.1. where u = ['. X n ). Let A E jRn xn (whose orthogonal matrix X e Wlxn (whose columns are orthonormal eigenvectors of A) such that of XT AX = D = diag(Al. XTAX = D = diag(Xi. A in (10.3) spectral which is often called the spectral representation of A... . A Note that Theorem 10.
The following theorem answers this question. where D is diagonal. AHA = AA H). and sufficient for virtually is real if U is orthogonal) is an eigenvector. The columns of a unitary [orthogonal] matrix U that reduces a matrix to [real} Schur fonn are called Schur vectors.98 98 Chapter 10. but In the case of A E R"xxn . Then there exists an orthogonal 10. [17]). The matrix 10. . diagonal of T (or S). While every matrix can be reduced to Schur form (or RSF).9. Example 10.8.5 is called a Schur canonical form or Schur form. then complex arithmetic is clearly needed if A has a complex conjugate pair of eigenvalues.. A is normal (i.. where T is upper triangular. Theorem 10. UH AU = T.6 (MurnaghanWintner). Let A e C"x". what is true. where T is upper triangular. so A is normal. The quasiuppertriangular matrix S in Theorem 10. The triangular matrix T in Theorem 10. Definition 10. for example. following theorem answers this question..e.9. Let A E R"xxn.7. real arithmetic) to a quasiuppertriangular matrix. A matrix A E C"x" is unitarily similar to a diagonal matrix if and only if Theorem 10. Theorem 10.2 except that in this case (using the notation U rather than X) the (l. However. where S is quasiuppertriangular. D ur In the case of A e IRn ". for example. However. However. The matrix s~ [ 2 0 2 5 4 0 is in RSF. Canonical Forms Chapter 10. Proof: The proof of this theorem is essentially the same as that of Theorem lO.6 T T matrix U such that U AU = S. it is thus unitarily similar to an upper triangular matrix. matrix U such that U AU = S. what is true..2)block wf AU2 is not 0. Canonical Forms Theorem 10. Its real JCF is h[ 1 1 1 0 0 n n Note that only the first Schur vector (and then only if the corresponding first eigenvalue Note that only the first Schur vector (and then only if the corresponding first eigenvalue is real if U is orthogonal) is an eigenvector. then complex arithmetic is clearly needed to place such eigenValues on the diagonal of T.8. Then there exists an orthogonal Let A e IR n ". is that the first Schur vectors span the same all applications (see. The triangular matrix T in Theorem 10.e. where D is diagonal. Proof: Suppose U is a unitary matrix such that U H AU = D. and sufficient for virtually all applications (see.5 (Schur). However. the next theorem shows that every A E IR xn is also orthogonally similar (i. the next theorem shows that every to place such eigenvalues on the diagonal of T. The quasiuppertriangular matrix S in Theorem 10. it is of interest to know when we can go further and reduce a matrix via unitary similarity to diagonal form.5 is called a Schur canonical Definition 10. is that the first k Schur vectors span the same Ainvariant subspace as the eigenvectors corresponding to the first eigenvalues along the invariant subspace as the eigenvectors corresponding to the first k eigenvalues along the diagonal of T (or S).2)block AU2 is not O. it is of interest to know While every matrix can be reduced to Schur form (or RSF).6 is called a real Schur canonical form or real Schur form (RSF).6 is called a real form or Schur fonn. it is thus unitarily similar to an upper triangular matrix. Then AAH = U VUHU VHU H = U DDHU H == U DH DU H == AH A so A is normal. A quasiuppertriangular matrix is block upper triangular with 1 x 1 diagonal blocks corresponding to its real eigenvalues and 2x2 2 diagonal blocks corresponding to its blocks corresponding to its real eigenvalues and 2 x diagonal blocks corresponding to its complex conjugate pairs of eigenvalues. Then Proof: Suppose U is a unitary matrix such that U H AU = D. real arithmetic) to a quasiuppertriangular A e Wnxn is also orthogonally similar (i. AH A = AAH ).e. The when we can go further and reduce a matrix via unitary similarity to diagonal form. complex conjugate pairs of eigenvalues. The columns of a unitary [orthogonal} Schur canonical form or real Schur fonn (RSF). [17]). Its real JCF is is in RSF. but if A has a complex conjugate pair of eigenvalues.2 except that Proof: The proof of this theorem is essentially the same as that of Theorem 10. where S is quasiuppertriangular. A matrix A e c nxn is unitarily similar to a diagonal matrix if and only if A is normal (i. matrix U that reduces a matrix to [real] Schur form are called Schur vectors. Let A E cnxn Then there exists a unitary matrix U such that Theorem 10. 0 in this case (using the notation U rather than X) the (l. .7. Then there exists a unitary matrix U such that U H AU = T.e.5 (Schur). A quasiuppertriangular matrix is block upper triangular with 1 x 1 diagonal matrix.
111. . We write A < O. negative definite if . We (or negative if— A nonnegative definite. where x is an arbitrary vector in en. nonpositive definite (or negative semidefinite) if A is nonnegative definite. in fact. We write A > 0.2. we write A :::: B if and only ifA — B>QorB — A < 0.10. it is said to be indefinite. indefinite.10. we write A > B if and only if A — B > B .5). 3.2 Definite Matrices Definite Matrices Definition 10. CM j]i. suppose A is normal and let U be a unitary matrix such that U H AU = T. we write A > B if and only if A . all the above definitions hold except that A e nxn Remark 10. Indeed..B > 0 or or Also. B — A < 0. Remark 10. Similarly.. let y = U H x. we write A > B if and only if A . We write A < 0.=1 But clearly n LA. nonnegative definite (or positive semidefinite) if and only if XT Ax :::: 0 for all (or positive if and only if x T Ax > for all nonzero x e W. Then for all E en. Proof: Proof: Let U be a unitary matrix that diagonalizes A as in Theorem 10. Then n x HAx = (U HX)H U H AU(U Hx) = yH Dy = LA.. this is generally true for all results in the remainder of of superscript //s Ts.11. if A and B are symmetric matrices. be diagonal.11. We write A ~ 0. where T is an upper triangular matrix (Theorem 10. this section that may be stated in the real case for simplicity. if—A 4.A < O.B :::: 0 or B .=1 . e Theorem 10. positive definite if and only if xTT Ax > 0 for all nonzero x G lR. If neither semidefinite. i En. if A and B are symmetric matrices. nonzero x E lR. Definite Matrices 99 Conversely. It T 0 D 10.2. Remark 10. Definite Matrices 10. Thenfor all Let A = AH E Cnxn with eigenvalues AI > A2 > • • > An.10.13.2.A is positive definite. A U U HA U T.12 ~ AlyH Y = AIX HX .2. i € n.2 10. Furthermore. 11'/. x eC". We write A > O. 2.• :::: An. Furthermore. U diagonalizes A 10. negative positive definite.. Then 11. superscript H s replace T s.n. Also. and denote the components of y by v UHx.A ~ O.12. Similarly. Then T (Theorem It is then a routine exercise to show that T must. write A < O. If A E C"x" is Hermitian. Let A = AH e enxn with eigenvalues X{ :::: A2 :::: .n • We write A :::: 0. If a matrix is neither definite nor semidefinite. A symmetric matrix A E lR.. positive definite if and only ifx Ax > Qfor all nonzero x E W1 We write A > O.12. A symmetric matrix A e Wxn 1.nxn is Definition 10.
The determinant of the 1x1 1 leading submatrix is 0 and 1. The determinant of the I x leading submatrix is 0 and consider the matrix A = [~ 2x 0 (cf. where M e R"x" is nonsingular.w) x HAx > the Rayleigh quotient. A leading principal submatrix of order n — k is obtained by deleting the last k rows and columns. not just those of the leading principal submatrices.18. For example. consider the matrix A — [0 _l~].1. ..17). 3. form MT E ~n xn E ~n xn definite if and only if Theorem 10. Let A E enxn Then \\A\\2 = Ar1ax(AH A). However.100 100 Chapter 10. A principal submatrix of an nxn n matrix A is the (n — k)x(n(n — k) matrix that remains by deleting k rows and the corresponding k columns. of all principal submatrices of 2. 3.= Amax{A A). Remark 10. x E C". so 0 An ::::: .14. 2.13 provides (A 1) Rayleigh quotient of jc. Corollary Corollary 10.l3 provides upper (AO and lower (An) bounds for (A. A symmetric matrix A € R"x" is nonnegative definite if and only if any of following equivalent of the following three equivalent conditions hold: 1.I. Then IIAII2 = ^m(AH A}. If A = AH e C"x" is positive definite.@mllrk 10.soO < X n < ••• < A. Theorem 1O.17. whence Ar1ax (A A). A symmetric matrix A e E" x" is positive definite if and only if any of the following equivalent following three equivalent conditions hold: determinants of principal 1. All eigenvalues of A are positive. Let A e C"x". Canonical Forms and and n LAillJilZ::: i=l AnyHy = An xHx .1. Theorem 10. Then ^pjp2 = ^^(A" HA). All eigenvalues of A are nonnegative. Canonical Forms Chapter 10.18.. Theorem 10.16.19. xfO IIxll2 I 0 Definition submatrixofan n x k) x k) Definition 10. whence IIAxll2 ! H IIAliz = max . Note that the determinants of all principal "ubm!ltriC[!!l mu"t bB nonnBgmivB R. A can be written in the form MT M. The ratio ^^ x for A = AH <=enxn and nonzero x jc een isis calledthe = AH E Cnxn and nonzero E C" called the x of x. A can be written in the form MT M. Theorem 10. where M 6 R ix " and k > rank(A) "" rank(M). A can be wrirren in [he/orm MT M. 2. I Proof: E C" Proof: For all x € en we have Let x be an eigenvector corresponding to Amax (A HA). of obtained and E ~nxn positive definite if and only if any of the Theorem 10. Not@th!ltthl!dl!termin!lntl:ofnllprincip!ll eubmatrioes muet bQ nonnogativo in Theorem 10.18. of positive.19. Then 111~~1~2 Let jc be an eigenvector corresponding to Xmax(AHA). XHAx > 0 for all nonzero = AH E enxn E en. the . The determinants of all principal submatrices of A are nonnegative.17). ::::: AI. 3. from which the theorem follows. Theorem 10.. All eigenvalues of A are positive.15. 0 D Remark XHHAx Remark 10. where M E IRb<n and k ~ ranlc(A) — ranlc(M). All eigenvalues of A are nonnegaTive. determinant the determinant of the 2x2 2 leading submatrix is also 0 (cf. The determinants of all leading principal submatrices of A are positive.
negative and A is nonpositive principal submatrix consisting of the (2.10. Then there exists a positive definite. The factor M in Theorem 10. Let A E lR. assume the result is true for matrices of order — 1 so that B may be written as B = L\L^. p. matrices (both symmetric and nonsymmetric) have infinitely many square roots. It is stated and proved below for the more general Hermitian case. A e R"x be nonnegative definite. The factor M in Theorem 10. For example. definite if A is positive definite). In general. For example. Hermitian case. 1f A :::: Band M E Rnxm. in fact. Then A has unique nonnegative Theorem 10. negative and is nonpositive definite. if = /2. Its proof is straightforward from basic definitions.22. for example.. B e Rnxn be symmetric. 2. If >BandMe jRnxm. Then A has aaunique nonnegative definite square root S. for example. any matrix of nonsymmetric) have infinitely many square roots. Write the matrix A in Proof: The proof is by induction. A stronger form of the third characterization in Theorem 10.17 is available and is known as the Cholesky factorization. In general. Definite Matrices 10. and positive definite.18.2) element is. standard theorem stated 181]). E <C Theorem 10.nxn"be nonnegative definite. Theorem 10. If A > B and M e Rm . if € E" xn we say that e jRn x that S E R nxn"isisa asquare root of AA ifS2 2 =— A. Ll E C1""1^""^ and . 1. 10rm [COSO _ Sino] . rankS = rankA definite definite if positive definite).20. E jRnxn MT AM > M BM. That is.2. with positive diagonal elements such that positive Proof: The proof is by induction.17 is available and is A stronger form of the third characterization in Theorem 10. assume the result is true for matrices of order n . = LLH.20. if A E lR. Theorem 10. BM. 0 Recall that A :::: B if the matrix A . Let A. That is. Its proof is straightforward from theorem is useful in "comparing" symmetric matrices.23. If A> Band E jR~xm. .22.18. For example.. [ fz ti o o l [~ 0] ~ 0 v'3 . The case = is trivially true. Remark 10. if A = lz.B is nonnegative definite. matrices (both symmetric and square root of if S A.23.2) element is. Definite Matrices 101 101 principal submatrix consisting of the (2. For example.3 is not unique.nxn . nxm 2. Write the matrix A in the form the form By our induction hypothesis. nxn Theorem 10. basic definitions. The following Recall that A > B if the matrix A — B is nonnegative definite. The following standard theorem is stated without proof (see. p. It is stated and proved below for the more general known as the Cholesky factorization. Let A e c nxn be Hermitian unique nonsingular lower triangular matrix L nonsingular A = LLH. any matrix S of c e s 9 the " °* ™ the form [ ssinOe _ ccosOe ] IS a square root. Moreover.21. [16.. then MT AM > MT TBM.2. where L\ e c(nl)x(nl) is nonsingular and lower triangular as = L1Lf.3 is not unique. j proof (see. then MT AM :::: MTTBM. It concerns the notion of the "square root" of a matrix. SA = AS and rankS = rank A (and hence S is positive = AS S S.1 so that B By our induction hypothesis. [16. if Remark 10. in fact.is a square root. The case n = 1 is trivially true. if then M can be then M can be [1 0]. concerns the notion of the "square root" of a matrix. MT AM> M.we say 181]). The following theorem is useful in "comparing" symmetric matrices.
3 10.24. Ch.24. 5].4) efficiently available. 131].4) [7. [21. available.4). as opposed to the more restrictive situation of a similarity transformation.• Clearly we see we L I C = b and ann = c HC a 2 c is given simply by c = C. Choosing a to be the positive square ann . . [21. Canonical Forms with positive diagonal elements. Then [ Sl o 0 ] [ I Uf U H ] AV = [I 0 0 ] 0 . The numerically preferred equivalence is. numerical procedures for computing such procedures an equivalence directly via. the unitary equivunitary alence known as the SVD.131]. Many similar results are also (10. yields a far "simpler" canonical form (10. Alternatively. Alternatively. Substituting in the involving we find 2 a2 = ann . suppose A has an SVD of the form (5. Then E c~xn such exist e C™ x m that that PAQ=[~ ~l (l0. They are more stably computable than (lOA) and more efficiently computable than a full SVD.102 102 Chapter 10.3 Equivalence Transformations and Congruence Equivalence Transformations and Congruence Theorem 10. But know that o < det(A) = det [ ~ b ] = det(B) det(a nn _ b H B1b). see. Take P =[ S~ 'f [I ] and Q = V to complete the proof. Choosing a be det(fi) > HB~lb completes the proof.b HL\H L11b = ann . the SVD is relatively expensive to compute and other canonical forms exist that are intermediate between (l0. [4. However. Ch. p. we see that we must have L\c = b and ann = CHc + a 2. multiplication where a is positive. for example.b H B1b > O. Let A € C™*71. are generally unreliable. However. we find by L^b.lb. we must have ann —bHB lb > 0.p. say. Canonical Forms Chapter 10. Two such forms are stated here.b B1b completes D 10. of course. Gaussian or elementary row and column operations. Then there exist matrices P E C: xm and Q e C"nx" such E c. 0 Note that the greater freedom afforded by the equivalence transformation of Theorem afforded 10.4) Proof: proof Proof: A classical proof can be consulted in. It remains to prove that we can write the n x n matrix A It in the form in the form ann b ] = [LJ c a 0 ] [Lf 0 c a J.2) in its complex version. 2]. of ann — b 0 root of «„„ .b H B1b (= the Schur complement of B in A). But we = ann — b LIH L\lb = ann — bH B~lb B A).xn.. Performing the indicated matrix multiplication and equating the corresponding submatrices. Substituting in the expression involving a.4) and the SVD. for example (10. ann Since det(B) > 0.
5) where R E e. 0 D x Theorem 10.29. If A AH e e x " then A> 0 if and only if In(A) = (n. £). Then there exists a unitary matrix Q E Cmxm and a permutation permutation matrix IT e en xn" such that Fl E C"x QAIT = [~ ~ l (10. It turns out that the principal property so preserved is the sign of each eigenvalue. see [4] for details.rrxr is upper (or lower) triangular with positive diagonal elements. When A has full column rank but is "near" a rank deficient matrix.28. D Proof: For the proof. Then the inertia of A is the triple of inertia of of negative. [21. sig(A) = rr — v.t h e n A > 0 if and only if In (A) = (n. of each eigenvalue.28.26.31 (Sylvester's Law of Inertia). In(A) = ln(X Proof: For the proof. numbers In(A) (n. Note that a congruence is a similarity if and only ifX is unitary. . In(A) 3. Theorem 10. upper Proof: For the proof.31 guarantees that rank and signature of matrix are preserved under congruence. v.26. .31 guarantees that rank and signature of a a matrixare preserved under Theorem 10. 0.25 (Complete Orthogonal Decomposition). The H.25 (Complete Orthogonal Decomposition). It turns out that the principal property so preserved is the sign preserved under congruence. Then H HAX). Then there exist Theorem 10. (TT.0. It is of interest to ask what other properties of a matrix are preserved under congruence.. v. respectively. It is of interest to ask what other properties of a matrix are then X H AX is also Hermitian. Proof: For the proof. Proof: For the proof. Again. Then there exists a unitary matrix Q e e mxm and a Theorem 10. If A = A" E C nnxn. see [4]. Equivalence Transformations and Congruence 103 103 Theorem 10. Theorem 10. Note that a congruence is a similarity if and only if X is unitary. Note that congruence preserves the property of being Hermitian. If In(A) = (rr. The signature of is Example 10. Let A = AH e C"x" and let 7t. and £ denote the numbers of positive. HE C xn E e~ xn. Let A = AH E e nxn and let rr. a congruence. Let A e e~xn. p. [21. where R E Crrxr is upper triangular and S e C rx( " r) is arbitrary but in general nonzero. v.29. 0 2. see [4]..27.e. When A has full column rank but is "near" a rank deficient matrix. Proof: For the proof. see [4]. various rank revealing QR decompositions are available that can sometimes detect such various rank revealing QR decompositions are available that can sometimes detect such phenomena at a cost considerably less than a full SVD. D 0 Remark 10. Then is the numbers In(A) = (rr.1). Let A e C™ ". Definition 10.v.27. Example 10. n. i. and eigenvalues.30. 134].XH AX Definition 10. i. Again. l. Definition 10. 134].e. We then have the following. negative. and zero eigenvalues.In[! 1o o o 0 0 00] 10 =(2. where R e €. for example. We then have the following. of A. then rank(A) = n + v. phenomena at a cost considerably less than a full SVD. v.31 (Sylvester's Law of Inertia).10. In(A) = In(X AX). see [4]. then XH AX is also Hermitian. 0). Let A = A He ennxn and X e Cnnxn.3. The transformation A i> XH AX is called a congruence. see [4] for details. nxn E e X E e~xn. v.3.30. see.1. for example. congruence. D Theorem 10. Then there exist unitary matrices U e Cmxm and V E Cnxn such that unitary matrices U E e mxm and V e e nxn such that (10. if A is Hermitian.xr is upper (or lower) triangular with positive diagonal elements.6) E e. 2.xr E erx(nr) arbitrary general nonzero. and ~ denote the numbers of positive. of A. n The signature of A is given by sig(A) = n . see. Equivalence Transformations and Congruence 10. then rank(A) rr v. Remark 10. Let A E C™ x ". if A is Hermitian. Let A E e~xn. Definition 10. respectively. Let A e Cnxn and X e Cnnxn. Note that congruence preserves the property of being Hermitian. £). p.0). v.
. . . . . £). ...0. if ifA>0. . 1.fArr+v. the congruence B ] [I D ~ 0 _AI B I ° JT [ A BT ~ ][ ~ 0 D The details are straightforward and are left to the reader. Canonical Forms Chapter 10. the next v are negative. Let A = AHeE cnxn with In(A) = (jt.0). Suppose A = AT and D = DT. 1/.BD. X e C"nxn such that XHAX = diag(l... if and if either A> and D . the number of Il's is v. . .. Proof: proof Proof: The proof follows by considering. Then = AT D = DT. I. . A w ). Define the nn x n matrix U UH AV = diag(Ai. . I..BT AI > 0.35.. and the numberofO's is~..BT A+B:::: o.34.3. . v. . Define the x n matrix vv = diag(I/~.. Note the symmetric Schur complements of A (or D) in the theorem. I/~..I BT > O.104 104 Chapter 10.. Theorem positive.BD^BT > 0. An). Proof: Consider the congruence with Proof: Consider proof Theorem and proceed as in the proof of Theorem 10.. .. . Then Remark Remark 10. An of Jr Proof: Let AI . and D . .2 there exists a unitary matrix V such that VHAU = diag(AI. AA+B = B. the number of — 's is v. .33. and the final £ are 0. and D . B D ] >  ° if and only if A:::: 0. Suppose A = AT and D = DT. .4 Rational Canonical Form Rational Canonical Form rational One final canonical form to be mentioned is the rational canonical form. Theorem 10. 1. where the number of E c~xn XH AX = diag(1. v. 1. Then there exists a matrix AH C"xn In(A) = (Jr.. ifand only ifeither A > 0 and D . 0. where the number of X 1's is Jr. Xw denote the eigenvalues of A and order them such that the first TTare ~ O.. left AT D DT.3. By Theorem 10.4 10. Proof: AI. D > and .. . . I/. the number 0/0 's is (.fArr+I' . . Canonical Forms Theorem 10. ..BT A+B > 0.1). .1 10..32. 0). I. or D > 0 and A . .1 Block matrices and definiteness Theorem 10. 0 D 10.. .33.. 's is 7i.BT A~l B > 0. 0 D Then it is easy to check that X = V VV yields the desired result.. X UW desired 10. . AA+B = B.. O. . .. 1.. for example..
A matrix A E lRnxn of the form (10. Using the reverseorder This matrix is a special case of a matrix in lower Hessenberg form.Then it can be shown (see [12]) that A mial is 7r(A) = A" . consider the companion matrix illustrate.18).8) This matrix is a special case of a matrix in lower Hessenberg form. Companion matrices also appear in the literature in several equivalent forms. Rational Canonical Form 105 105 Definition A matrix A e M"x" is said to be Definition 10. Notice that in all cases a companion matrix is nonsingular if and only if aO i= O. A is easily seen to be similar to the following matrix in upper Hessenberg form: in upper Hessenberg form: a2 al o 0 0 1 o 1 6] ao o . the inverse of a nonsingular companion matrix is again in companion form.7) is called a cornpanion rnatrix or Definition 10. o (10.4. since a matrix is similar to its transpose (see exercise 13 in Chapter 9). For £*Yamr\1j=» example.7) Definition 10. A is easily seen to be similar to the following matrix identity similarity P given by (9.(ao + «A + . Suppose A E lRnxn is a nonderogatory matrix and suppose its characteristic polynoSuppose A E Wxn is a nonderogatory matrix and suppose its characteristic polynon(A) An — (a0 + alA + a n _iA n ~'). equivalently. For In fact. : ~ ! ~01].11) . A matrix A E lRn Xn is said to be nonderogatory ifits minimal polynomial if its minimal polynomial and characteristic polynomial are the same or. is said to be in cornpanion form. To illustrate. the following are also companion matrices similar to the above: following are also companion matrices similar to the above: Notice that in all cases a companion matrix is nonsingular if and only if ao /= 0. Rational Canonical Form 10.36. l 0 0 ~ ao ~ ao _!!l (10.9) Moreover. + an_IAnI). the Moreover. if its Jordan canonical form and characteristic polynomial are the same or.37. equivalently. A matrix A e E nx " of the form (10.. if its Jordan canonical form has only one block associated with each distinct eigenvalue.4. since a matrix is similar to its transpose (see exercise 13 in Chapter 9).18).. In fact. To Companion matrices also appear in the literature in several equivalent forms.10. the inverse of a nonsingular companion matrix is again in companion form. Using the reverseorder identity similarity P given by (9.7) is called a companion matrix or is said to be in companion forrn.37. consider the companion matrix (l0.10) o 1 o 1 o o o o o o (10. has only one block associated with each distinct eigenvalue. Then it can be shown (see [12]) that A is similar to a matrix of the form is similar to a matrix of the form o o o o 0 o o o (10.
stable ones are nearly unstable.. if ao = 1 inverse can still be computed.Jy2 . For example. the largest and smallest singular values can also be written in the equivalent form Remark 10.10).. a. see. among which. Let a = a\ + a\ + • • • + a%_{ and y = 1 + «. see [14].4aJ) .e. Algorithms to reduce an arbitrary matrix to companion form are numerically unstable. and so forth [14]. matrix is not a companion matrix unless a = O. is the fact that their singular values can be found in closed form. Moreover. If A E R nx " is derogatory. nonsingular ones are nearly singular.38. i.4ao ' 1 2)  a? = 1 for i = 2. Explicit formulas for all the associated right and left singular vectors can also be derived easily. Ifao ^ 0. = ~ (y . form). it can be shown that a derogatory matrix is similar to a block diagonal matrix.39. . For example.38. companion an arbitrary matrix to companion form are numerically unstable. 02. then its pseudoIf singular. with a similar result for companion matrices of the form (10.. .e. especially nonsingular ones are nearly singular.7). then it is not similar to a companion matrix of the form (10.. Canonical Forms with a similar result for companion matrices of the form (10.10).. for example. a2.7) is singular. among which. I — T = T) Note that / . and perCompanion matrices have many other interesting properties.7). Then it is easily verified that c = l+ ara' Then it is easily verified that o o o + o o o o o o 1.caa T ca o J. Companion matrices have many other interesting properties.. anIf and let e M"" \a\. Companion matrices appear frequently in the control and signal processing literature Companion matrices appear frequently in the control and signal processing literature but unfortunately they are often very difficult to work with numerically. + a. at least one eigenvalue.. i. Such matrices are said to be in each of whose diagonal blocks is a companion matrix.7). Moreover.. see haps surprisingly..Q + a. [12]._1 and y = 1 + + a.39. companion matrices are known to possess many undesirable numerical properties. the largest and smallest singular values can also be written in the equivalent form If ao =1= 0. Explicit formulas for all the associated right and left singular vectors can Remark 10. then it is not similar to a companion matrix of the form (10. and hence the pseudoinverse of a singular companion + matrix is not a companion matrix unless a = 0. and so forth [14]. For details.1.. 3.106 Chapter 10. is the fact that their singular values can be found in closed form.7). n . has more than one Jordan block associated with If A € JRnxn derogatory. Let al ~ GI ~ • • ~ an be the singular values of the companion matrix Theorem 10. associated at least one eigenvalue. stable ones are nearly unstable.. Let a E JRn1 denote the vector [ai. in matrices are known to possess many undesirable numerical properties. in n general and especially as n increases. However. their eigenstructure is extremely ill conditioned. . If a companion matrix of the form (10.caa T = (I + aaT) I .. a n i] and l c I+~T a.. Leta = ar aJ al 2_ 2 ( y + Jy 2. Theorem 10. if ao = 0. each of whose diagonal blocks is a companion matrix.. Such matrices are said to be in rational canonical form Frobenius rational canonical form (or Frobenius canonical form).• > an be the singular values of the companion matrix A in (10. Then A in (10. Canonical Forms Chapter 10. . Then + ai + .. Let a\ > a2 > . also be derived easily. and perhaps surprisingly. . Algorithms to reduce but unfortunately they are often very difficult to work with numerically.
one may lose up to k digits of precision.38 yields some understanding of why difficult numerical Remark 10. A E jRnxn N(A) = A/"(A ). this condition number is the ratio of largest to smallest singular precision. Prove that if A e M"x" is normal. In the 2norm.40.Exercises Exercises 107 Companion matrices and rational canonical forms are generally to be avoided in floatingCompanion matrices and rational canonical forms are generally to be avoided in fioatingpoint computation. . then it must be diagonal. A [ must also be positive 7. one measure of numerical sensitivity is KP(A) = A A ] > the socalled condition number of A with respect to inversion and with respect II ^ IIpp II A~l IIpp'me socalled condition number of A with respect to inversion and with respect to the matrix pnorm. Let R.. A E cc nxn peA) = max).. . In the 2norm. Suppose A e E"x" is positive definite.18) and the matrix U in identity in (9..(A) for e n. then Af(A) = N(A Tr ). Show that if A is normal. 5. It is not unusual for y to be large for large n. Show that [~ R > SI. say O(lO k ). For example.4a5 21 a ol It is easy to show that 21~01 :::: k2(A) :::: 1:01' and when ao is small or y is large (or both).5 to find a unitary matrix Q that reduces A e C"x" to lower triangular form. Use the reverseorder identity matrix P introduced in (9. Let R. For example.. Let A 7. Let A € C n xn be normal with eigenvalues y1 . is true if n = 2. then K2(A) ^ T~I.18) U A E cc nxn Theorem 10. • • > on > 0..(A)I for ii E!l. yn and singular values a\ ~ a2 > . 9. and when GO is small or y is large (or both).. R> S [1 A~I] ~ O? /i 1 > 0? ~] > 0 if and only if > 0 and J 1 > 0 if and only if S > 0 and .2). E jRnxn be symmetric. then peA) = A2. Show that a. when solving linear equations numerical sensitivity Kp(A) = systems of equations of the form (6. Let A = I J : ]eEC 22x2. Theorem 10. A E en x n eigenvalues A]. one may lose up to k digits of to the matrix Pnorm. An and singular 0'1 > 0'2 ~ 4. (A) A. 6. Show that the converse radius of A.. Show that a.40. If this number is large. Theorem 10. 2.. K\ (A) (10. Note that explicit formulas then K2(A) ~ I~I' It is not unusualfor y to be large forlarge Note that explicit formulas Koo(A) for K] (A) and Koo(A) can also be determined easily by using (l0. EXERCISES EXERCISES 1. by the theorem.EA(A) I'MpeA) 3. Remark 10. can be determined explicitly as determined explicitly y+J y 2 .. say 0(10*). show that AI must also be positive definite. then p(A) = IIAII2' Show that the converse is true if n = 2.38 yields some understanding of why difficult numerical behavior might be expected for companion matrices. 1. Is [ ^ A E jRnxn is definite. this condition number is the ratio of largest to smallest singular values which. If A e jRn xn 8. Let A G Cnx" and define p(A) = maxx€A(A) IAI.11). (A) = IA. S 6 E nxn be symmetric..11).• ~ an ~ O. Then p(A) is called the spectral radius of A. 3. If this number is large.. Show that if A is normal. . Show that if a triangular matrix is normal. If A E Wxn is positive definite. Find a unitary matrix U such that [~ M CC x 2 Find a unitary matrix U such that 6. Show that [ * }. It is easy to show that y/2/ao < K2(A) < £. when solving linear behavior might be expected for companion matrices.
j 1+ j ] 1 .j 1+ j ] 2 ' (d) [ . Find the inertia of the following matrices: following 10.108 108 10. (a) Chapter 10. . Canonical Forms Chapter 10. Canonical Forms [~ ~ l (b) [ 2 1.1 1.
which thus also converges for all A and uniformly in t.1 Differential Equations Differential Equations = Ax(t). the matrix exponential e A E Rnxn is defined by the power series power series e = A L +00 1 .1) involves the matrix to +(0). = Xo In this section we study solutions of the linear homogeneous system of differential equations In this section we study solutions of the linear homogeneous system of differential equations x(t) x(to) E JR. For all A e Rnxn.1 and linearity of the transpose.2) can be shown to converge for all A (has radius of convergence equal to +00).1 Properties of the matrix exponential Properties of the matrix exponential 1.1 11.1) for t 2: to. e° = I.n (11.nxn is constant and does not depend on t. Proof This follows immediately from Definition 11. The solution of (11. It can be described conveniently in terms of the matrix exponential.Ak. eO = I. The solution of (11. Proof: This follows immediately from Definition 11.1.1 by setting = 0.1. Proof This follows immediately from Definition 11. (e(eAf = e A e^. k. The solution of (11.1 and linearity of the transpose. Forall A EG R" XM .1) involves the matrix (11.1) is then known always to exist and be and does not depend on t.1. For all A E JR. unique. We restrict our attention in this chapter only to the socalled timeinvariant case. (11.1 11. the matrix exponential e A e JR. Proof: This follows immediately from Definition 11.nxn is defined by Definition 11.nxn.1. A) • 2.3) which thus also converges for all A and uniformly in t.2) can be shown to converge for all A (has radius of convergence equal The series (11. The solution of (11. where the matrix A E Rnxn is constant chapter only to the socalled timeinvariant case.1) is then known always to exist and be unique.nxn. Definition 11. 11. We restrict our attention in this for t > IQ. For all A JR. It can be described conveniently in terms of the matrix exponential.Chapter 11 Chapter 11 Linear Differential and Linear Differential and Difference Equations Difference Equations 11. where the matrix A e JR.1 by setting AA =O.2) k=O The series (11. T T 109 109 . This is known as an initialvalue problem. This is known as an initialvalue problem.
and B commute. ) ( 1+ tA + t2!A 2 +. Let £ denote the Laplace transform and £1 the inverse Laplace transform. Proof: We prove only (a). i..A)I..1 {(j/A). (b) .. (a) C{etA = (sIArl. 6.110 110 Chapter 11. Proof" Simply take T = — t in property 3. 4. r e R. Part (b) follows similarly. For all e JRnxn and for all E R. Then for 6.A)I} = «M. Then for E JRnxn t E R..tA . ) ( I + T A + T2!2 A 2 +. et(A+B) =etAe tB = etBe tA if and only if A all e JRnxn and all e R. ) .. = e'A erA = elAe'A . et(A+B) =^e'Ae'B = e'Be'A and and B commute. 5. ) . 2 2! and and while while e e tB tA = ( 1+ tB t2 2 + 2iB 2 +. on (t + T)*. (e'A)~l e~'A. Part (b) follows similarly. AB = B A. (etA)1 = e. T E JR. Proof' Note that Proof: Note that et(A+B) = I t + teA + B) + (A + B)2 + . (a) . i.. 2! and and e e tA rA 2 = ( I + t A + t2! A 2 +. Compare like powers of A in the above two equations and use the binomial theorem Compare like powers of A in the above two equations and use the binomial theorem on(t+T)k..e.e. = I + (t + T)A + (t + T)2 A 2 + . Linear Differential and Difference Equations e(t+r)A e(t+T)A 3... AB = BA. Proof: Simply take T = t in property 3.. ForaH A E R" x " and for all t e JR. For all A E JRnxn and for all t. Compare like powers of t in the first equation and the second or third and use the Compare like powers of t in the first equation and the second or third and use the k binomial theorem on (A + B/ and the commutativity of A and B. binomial theorem on (A B) and the commutativity of A and B. Proof" We prove only (a).l{e tA}} = (sI .lI{(sl.. Proof" Note that Proof: Note that e(t+r)A = etA erA = erAe tA . (b) £. Let denote the Laplace transform and £~! the inverse Laplace transform. Linear Differential and Difference Equations Chapter 11. For all e R"x" and for all t.... For all A. B E R" xn and for all t E JR. {+oo = io et(sl)e tA dt since A and (sf) commute =io (+oo ef(Asl) dt .1 } = erA. all A € R"x" and for all t € lR.
using the JCF. ) = L'lt IIA 21111e tA IIe~tIIAII..AetAil Ae tA I ~t (e~tAetA I (M A I ~t (e~tA . Alternatively. £(e'A) 7.=1 m Xiet(Jisl)y.Ae tA I = III (etAe~tA L'lt = = /A) . Notice in the proof that we have assumed. If this is not the case. . All succeeding steps in the proof then follow in straightforward way..y.. ... s .H = '"' assuming Re s > Re Ai for i E !! = (sI .1. it can be differentiated termbyProof: Since the series (11. The matrix (s I .A)I....H using the JCF.. = (sl A). the scalar dyadic decomposition can be replaced by et(Asl) =L .3) is uniformly convergent.H L.3) is uniformly convergent.Ae tA tA tA I I e tA . that A is diagonalizable..X i y.11..=1 = ~[fo+oo e(AiS)t dt]x.All succeeding steps in the proof then follow in aastraightforward way. employed I e(t+~t)AAt.etA .A) ~' is called the resolvent of A and is defined for all s not in A (A).Ae tA etA) . For all A e R"x" and for all t e R.. . ) = I ( Ae + = tA ~... it can be differentiated termbyterm from which the result follows immediately.. ) etA I < MIIA21111e  L'lt (L'lt)2 + IIAII + IIAI12 + . the formal definition d dt _(/A) = lim ~t+O e(t+M)A _ etA L'lt can be employed as follows.l)e .y. for convenience... that A is diagonalizable.u . the scalar dyadic decomposition can be replaced by If this is not the case.Ae II = I L'lt (M)2 + ~ A 2 +. For any consistent matrix norm.H dt assuming A is diagonalizable ... A 2etA + . e'A Proof: Since the series (11.. Differential Equations 111 111 = {+oo 10 n t 1 e(AiS)t x.1. Notice in the proof that we have assumed. For all A E JRnxn and for all E JR.A"I i=1 .1 The matrix (s I — A) I is called the resolvent of A and is defined for all s not in A (A). for convenience. A2 + (~~)2 A tA II tA Il 1 (_ 2! + . ) 3 II I ( ~. 1h(e tA ) = AetA = etA A. ) 3! 4! L'ltiIAIl < L'lt1lA21111e (1 + + (~t IIAII2 + . Differential Equations 11.
6).2 Homogeneous linear differential equations Homogeneous equations x(t) Theorem 11. (11. The general Proof: Differentiate (11. The proof above simply verifies the variation of parameters formula by direct differentiation. D Ir: Remark 11.f(p(t). 11. Also.5) is the solution of (11..6) for t ::: to is given by the variation of parameters formula for t > IQ is given by the variation of parameters formula x(t) = e(tto)A xo + t e(ts)A Bu(s) ds. (11. (11. Also. x(to) = Xo E IR n (11. or one can use the fact that A commutes with any polynomial of A of finite degree and hence with e'A. 0 ordinary differential equations. by the fundamental existence and x(t0) — e(fo~t°')AXQ — Xo uniqueness theorem for ordinary differential equations. Linear Differential and Difference Equations For fixed t. The solution of the linear homogeneous initialvalue problem = Ax(l). The proof above simply verifies the variation of parameters formula by Remark 11. Then the solution of the linear inhomogeneous initialvalue problem and.dt dt is used to get x ( t ) = Ae(tto)Ax0 + f'o Ae('s)ABu(s) ds + Bu(t) = Ax(t) + Bu(t).¥o + 0 = XQ so. the For fixed t. the limit exists and equals Ae'A •. say.t goes to O.5) Proof: Differentiate (11.4. Premultiply the equation x .1. The formula can be derived by means of an integrating factor "trick" as follows. The solution ofthe linear homogeneous initialvalue problem Let A e Rnxn. B e IR xm and let the vectorvalued function u be given Theorem and.i~t()Oc() nnd uniqu()Oc:s:s theorem for *('o)} = <?(f°~fo)/1. The general formula formula d dt l q (t) pet) f(x. continuous. t ) . the righthand side above clearly goes to 0 as t:. lo t (11.8) . Then the solution of the linear inhomogeneous initialvalue problem x(t) = Ax(t) + Bu(t).4).5) is the solution of (11..Ax = Bu by e. fact that A commutes with any polynomial of A of finite degree and hence with etA.7) is the solution of (1l.5) and use property 7 of the matrix exponential to get x t ) = Ae(tto)A xo fundamental Ae(t~to)Axo = Ax(t). Linear Differential and Difference Equations Chapter 11. x(to) = Xo E IRn (11.112 112 Chapter 11.4.6). D 11. The formula can be derived by means of an integrating factor "trick" direct differentiation.3.2.tA to get as follows.3 Inhomogeneous linear differential equations Inhomogeneous equations Theorem 11. Thus.7) is the solution of (11. B E Wnxm and let the vectorvalued function u be given Let A e IR nxn . say. Let A E IR n xn.7) and again use property 7 of the matrix exponential. Premultiply the equation x — Ax = Bu by e~ to get (11. x(to) = e(toto)A Xo = XQ so. (11. t) dx = l af(x t) ' dx pet) at (t) q + dq(t) dp(t) f(q(t).7) and again use property 7 of the matrix exponential. Let A E Rnxn .4). Thus. the righthand side above clearly goes to 0 as At goes to 0. Ae(ts)A Bu(s) to get x(t) = Ae{'to)A Xo + Bu(t) = Ax(t) = x(to e(totolA Xo + = Xo fundilm()ntill ()lI.7) Proof: Differentiate (11. 0 uniqueness theorem for ordinary differential equations.5) and use property 7 of the matrix exponential to get x ((t) = Proof: Differentiate (11.1. by the fundamental existence and uniqueness theorem for ordinary differential equations. continuous.4) for t ::: to is given by (11. or one can use the limit exists and equals Ae t A A similar proof yields the limit et A A. A similar proof yields the limit e'A A. t ) .
2. and the proof is essentially the same. Then the matrix initialvalue E jRmxm.1. X(O) =C (11. Let A E Wlxn. differential equation. Theorem 11.1. punov differential equation.6.2. X t) X 0 D Corollary 11. Differential Equations [to. we can have coefficient matrices on both the right and left. and C e Rnxm. t exponential.5. X((t) is symmetric and (11.4 Linear matrix differential equations Linear matrix differential equations Matrixvalued initialvalue problems also occur frequently.11) is known as a Sylvester Sylvester differential equation. C e IR" ". Corollary 11. t]: Now integrate (11. e jRnxn. (11. The initialvalue problem (11. Then the matrix initialvalue problem X(t) = AX(t) + X(t)AT. The fact that X((t) satisfies the initial condition is trivial.. X(to) =C E jRnxn (11.etoAx(to) = lto t e. B e R m x m . For convenience. the following theorem is stated with initial time to = 0. t]: 113 1 Thus.nxm.1.11.6.9) for t ::: to is given by for t > to is given by X(t) = e(tto)Ac.7.4 11. E ]R. X(O) = C (11. and the proof is essentially the same. The of nrohlcm problem X(t) = AX(t).sA Bu(s) ds x(t) = e(ttolA xo + lto t e(ts)A Bu(s) ds. E ]R. Theorem 11.7.11) X(t) = etACe = e ratB has the solution X ( t ) — atACe tB . Theorem 11. the Theorem 11. problem problem X(t) = AX(t) + X(t)B. the When C is symmetric in (11.10) coefficient In the matrix case.1. Let A E Rnxn. . 11. and hence t d esAx(s) ds = to ds 1t to eSABu(s) ds.12). Differential Equations 11. The first is an obvious generalization of Theorem 11. etAx(t) .12) X(t) = etACetAT has the solution X(t} = etACetAT.8) over the interval [to. The solution of the matrix linear homogeneous initialvalue e jRnxn. following to = O. the Proof: Differentiate etACe tB property Proof: Differentiate etACetB with respect to t and use property 7 of the matrix exponential.nxn. Let A.12) is known as a LyaX t) punov differential equation.
Similarly. ~ 11.114 114 Chapter 11. the rest of this subsection is easily generalized by using the JCF and the decomposition able. the rest of this subsection is easily generalized by using the JCF and the decomposition H A — ^ Xf Ji YiH as discussed in Chapter 9). Then the solution x(t) of (11. Let A E jRnxn and suppose X e jR~xn is such that XI AX = J. where J is a JCF for A. Linear Differential and Difference Equations Chapter 11. in the inhomogeneous case we can write Similarly. i=1 The ki s are called the modal velocities and the right eigenvectors Xi are called the modal The Ai s are called the modal velocities and the right eigenvectors *. In the last equality we have used the fact that YiHXj = flij.x.1 . Then Then i=1 n = L(aieAiUtO»Xi.1. Then the solution x(t) of (11. where J is a JCF for A. modal velocities and directions.6 Computation of the matrix exponential Computation exponential JCF method JCF method Let A e R"x" and suppose X E Rnxn is such that X"1 AX = J.4) can be written x(t) = e(tto)A Xo E jRnxn E Wxn = (ti. are called the modal directions. i=1 In the last equality we have used the fact that yf*Xj = Sfj. for convenience. Then Then etA = etXJX1 = XetJX. The decomposition above expresses the solution x (t) as a weighted sum of its directions. The decomposition above expresses the solution x(t) as a weighted sum of its modal velocities and directions. that it is diagonalizable (if A is not diagonalizLet A and suppose.1 n Le A• X'YiH . if A is diagonalizable in geneml. for convenience.H . in the inhomogeneous case we can write t e(ts)A Bu(s) ds i~ = t i=1 (it eAiUS)YiH Bu(s) dS) Xi.e'J.4) can be written A = L X.5 11. that it is diagonalizable (if A is not diagonalizable. ~ 1=1 I t.li y t as discussed in Chapter 9).iUtO)Xiyr) Xo 1=1 n = L(YiHxoeAi(ttO»Xi. Linear Differential and Difference Equations 11. This modal decomposition can be expressed in a different looking but identical form This modal decomposition can be expressed in a different looking but identical form n if we write the initial condition Xo as a weighted sum of the right eigenvectors if we write the initial condition XQ as a weighted sum of the right eigenvectors Xo = L ai Xi.5 Modal decompositions Let A and suppose.y.1.
let .EeCkxk be aaJordan block of the form Ji <Ckxk be Jordan block of the form A Ji = 1 o o o =U+N. O. Differential Equations 115 If A is diagonalizable.. Thus. real version of the above can be worked out. or grade) p if if matrix M e jRnxn is nilpotent of degree (or index. its first superdiagonal (and O's elsewhere). eAt teAt eAt o 2I e 12 At IkI At e (kI)! 0 ell.11. e ttJi = eO. nilpotent Definition 11. k) element and has O's everywhere else. degree k. To be specific. + N k2! (k .7. while MPI t=. . and N kforth.1. and so forth.I)! I o t 1 o Thus. or grade) MP = 0. e lN finite.. Differential Equations 11. the problem clearly reduces simply to the computation of problem clearly reduces the exponential of a Jordan block. elN is is nilpotent of degree k. k) O's k k N = 0. Mp~l ^ O.!etN by property 4 of the matrix exponential. ext}. the series expansion of e'N is finite. A.e.8... i. Finally.. ••• . N22 has l's along only its second superdiagonal.. t2 t k. e'u e l N tu x lH = diag(e At . AI e I.0. N has 1's along only its second superdiagonal. it is then easy to compute etA via the formula etA = XetJ XI' Xe tl X If is etA etA tj since et I is simply a diagonal matrix.. Thus. Mp = 0.1. Nk~lI has a 1 in its (1. But e tN is almost as easy since N The diagonal part is easy: e e = diag(e '. (1. A matrix M E M nx " is nilpotent of degree (or index. of In the more general case.I e IN =I+tN+N 2 + . teAl eAt = 0 0 0 2I e 12 At teAl 0 eAt In the case when A is complex.eAt). it is easy to check that while N has 1's along only its first superdiagonal (and O's elsewhere). aareal version of the above can be worked out. is complex. l's For the matrix N defined above. o A o o A Clearly A/ and N commute.
Let A = [ ~_\ J].. Theorem 9. The motivation for this method is known.9.2t e.. the function g is known and /(A) = g(A). Let Example 11. which says that all powers of A greater than A n . ani solution of the n equations: g(k)(Ai) = f(k)(Ai). I. where t is a fixed scalar. Then jr(A. f(A) n(A) etK.s known. k = 0.116 Chapter 11. .9. . Then the three equations for the a. functions.t • g'(1) = f'(1) g"(I) = 1"(1) .. . Linear Differential and Difference Equations Example 11. (A. The method is stated and illustrated for the exponential function but applies equally well to other functions. Let A = [~ o ~01~ ] t . anl are n constants that are to be determined. .2.1 in the power series for et A can be written in terms of these greater n— e' A lowerorder powers as well. lowerorder g Example 11. 2} and etA Xe tJ =[=i a = xI =[ =[ 2 1 ] exp t ] [ [ 2 0 ~ ] [ 1 1 2 1 2 ] 2 1 e~2t te. compute f(A) = etA. . .t .2t ][ 1 ] Interpolation method Interpolation method This method is numerically unstable in finiteprecision arithmetic but is quite effective for effective hand calculation in smallorder problems. The method is stated and illustrated for the hand calculation in smallorder problems.I.2a2 = te.. . Let A Then A (A) = {2.1 can be expressed as linear combinations of Ak for k = 0..) = (A + 1)3. .10. all the Ak — expressed k 1. a. Suppose the characteristic polynomial of A can be written as n(A)) = Yi?=i (A . i Em. t fixed Given A € E. .. ni . . The polynomial g gives the appropriate linear combination. The motivation for this method is the CayleyHamilton Theorem. Given A E jRnxn and f(A) = etA. + I) 3 .. compute f(A) = e'A. I. characteristic of n(X (^ ~~ ^i)"'» where the A. ==> 2a2 = t 2 e.. Here. and /(A) = etA.1.s are given by g(A) — ao aiS a\X o^A..nxn and /(A) = etx. so m = 1 and nl Let g(X) = UQ + alA + a2A2. 2} and Example 11. the function g is known and f(A) = g(A).10. . Then A(A) = {2. Linear Differential and Difference Equations Chapter 11. They are.s are distinct.a l +a2 = e==> at . in fact.. n . the unique OTQ. the superscript (k) denotes the fcth derivative with respect to A. — 1. . g(I) = f(1) ==> ao . Thus.3. With the aiS then kth superscript (&) X. terms of order greater than n . Define the Ai nr=1 n where ao. so m = 1 and n{ = 3.Ai t'.
Let A _* Example 11. Then 7r(X) = (A+ 2)22 so m = 11and [::::~ 4i and f(A) = ea. Thus.2t _ Other methods Other methods 1.2t .A)^ 1 } and techniques for inverse Laplace transforms.2t [ ~ o ] + te. we find 117 Thus. 1. There is an extensive literature on approximating certain nonlinear functions by rational functions. g'(2) = f'(2) = te Solving for the a. The matrix analogue yields e A ~ functions rational eA = .11. Use etA = £~l{(sl . 2.2t . 2.11. Use Pade approximation. Then rr(A) = f\ + o\2 so m = and (A i 2) «i nL = 2. Let A = [ ~4 J] and /(A) = eO.) = «o + ofiA. we find Solving for the aiS.2t I [4 4] I 0 _ [  e. This etA = . we find ao = e. Then the defining equations for the a. but general nonsymbolic computational effective smallorder techniques numerically problem equivalent techniques are numerically unstable since the problem is theoretically equivalent to knowing precisely a JCE JCF.2t te. Let g(A.2t aL = + 2te.1. Differential Equations 11 . t ff>\ tk TU^^ _/"i\ Example 11.2t ) 2te. f(A) = etA = g(A) = aoI + al A = (e.s. Differential Equations Solving for the ai s. s.. te. Then the defining equations for the aiS are given by 6] g(2) = f(2) ==> ao ==> al 2al = e.11.2t + 2te.2t . we find Solving for the a. 2t .cI{(sI — A)I} is quite effective for smallorder problems.1.s are given by Let g(A) ao + aLA.
Many methods are outlined in.13. where the matrix A in (11.2 Inhomogeneous linear difference equations Inhomogeneous linear difference equations E jRnxn..15) .e. a Fade approximation for polynomials of various orders. 0 D Remark 11. eS .118 118 l Chapter 11.1 Homogeneous linear difference equations Homogeneous linear difference equations Theorem 11. no double subscripts). Then the solution of the inhomogeneous initialvalue problem mvectors. •• vq A . Many methods are outlined in. but since the system is timeinvariant. Unfortunately.13).13. Let A e Rnxn. The solution of the linear homogeneous system ofdifference Let A e jRn xn. in the matrix case the exponential is accurate only in a neighborhood of the origin..13). by 22' 2* )A A multiplying it by 1/2k for sufficiently large k and using the fact that A = / { ]I //2')A )\ * . Reliable and efficient computation of matrix functions such as e A and log(A) remains a fertile area for research. we have chosen ko = 0 for convenience. Linear Differential and Difference Equations Chapter 11. and since we want to keep the formulas "clean" (i. we have chosen ko = 0 for want to keep the formulas "clean" (i. [19]. We could also consider an arbitrary "initial time" ko. where the matrix A in (11.12.14.. say. exhibit many parallels to the continuoustime differential equation difference equations. modeled by systems of equations of the previous section. + opAP and N(A) = vol + vIA + D~ (A)N(A)..e. we restrict our attention only to the socalled timeinvariant Remark 11. but since the system is timeinvariant. exhibit many parallels to the continuoustime differential equation case. Again.2. Numerical loss of accuracy can occur in this procedure from the successive squarings. 11. 11. 11.14) into (11. Again. of matrix functions such as e A and 10g(A) remains a fertile area for research. and this observation is exploited frequently. = P = multiplying it by 1/2* for sufficiently large k and using the fact that e = ( e j . 4. Numerical loss of accuracy can occur in this procedure from the successive squarings.13) is constant and does not depend on k. by this means when IIAII is sufficiently small. case.13) is constant and does not depend on k. say. a Pad6 approximation for denominator the exponential is accurate only in a neighborhood of the origin.13) for k > 0 is given by for k 2:: 0 is given by Proof: The proof is almost immediate upon substitution of (11. Then the solution of the inhomogeneous initialvalue problem (11. Linear discretetime systems. modeled by systems of difference equations.2 11. Reduce A to (real) Schur form S via the unitary similarity U and use eA = U e SsUH Ue U H and successive recursions up the superdiagonals of the (quasi) upper triangular matrix and successive recursions up the superdiagonals of the (quasi) upper triangular matrix e s. for example.. Linear discretetime systems.• + Vq A q.2.2 Difference Equations Difference Equations In this section we outline solutions of discretetime analogues of the linear differential In this section we outline solutions of discretetime analogues of the linear differential equations of the previous section. Reliable and efficient computation 4. [19]. Proof: The proof is almost immediate upon substitution of (11. Unfortunately. e (e( 3.1 11. Linear Differential and Difference Equations DI(A)N(A). We could also case. for example. Let A E Rnxn. and since we consider an arbitrary "initial time" ko.2. no double subscripts). we restrict our attention only to the socalled timeinvariant case. The solution ofthe linear homogeneous system of difference equations equations (11. This can be arranged by scaling A. and this observation is exploited frequently. E jRnxm {udt~ is of Theorem 11. This can be arranged by scaling A. Reduce A to (real) Schur form S via the unitary similarity U and use e A 3. where D(A) 80I Si A H h SPA and N(A) v0I + vlA + q Explicit formulas are known for the coefficients of the numerator and Explicit formulas are known for the coefficients of the numerator and denominator polynomials of various orders. convenience.14) into (11. in the matrix case this means when  A is sufficiently small.. B e Rnxm and suppose {«*}£§ « a given sequence of mvectors. where D(A) = 001 + olA + .
15).1 _I tA~X. in general. LXi Jtyi . based Methods based on the JCF are sometimes useful. Difference Equations 11. again mostly for smallorder probsmallorder lems. j=O (11. substitution of (11.O. One solution method.2. Then Ak = (XJXI)k = XJkX.3 Computation of matrix powers Computation of matrix powers It is clear that solution of linear systems of difference equations involves computation of It is clear that solution of linear systems of difference equations involves computation of k.zA =I+A+"2 A + . the ztransform of the sequence {Ak is then given by Assuming z > max IAI. X~1 AX JCF for A. is to use ztransforms. where J is a E jRnxn and X E R^n J.A)I.y.3 11. k:::.11. since /* is simply a diagonal matrix. it is then easy to compute Ak via the formula Ak = XJkXXI Ak Ak — X Jk If diagonalizable.H m if A is diagonalizable. +00 k=O z z = (lzIA)I = z(zI .15).2.. Difference Equations 119 119 is given by kI xk=AkXO+LAkjIBUj. a matrix exponential.16) into (11.=1 H l If A is diagonalizable.16) into (11. Then JCF for A.16) Proof: The proof is again almost immediate Proof: The proof is again almost immediate upon substitution of (11. which is numerically unstable but sometimes useful for hand calculation.. One definition of the ztransform of a sequence {gk} is a matrix exponential. Assume that A e M" xn and let X e jR~xn be such that XI AX = /. sometimes useful Ak.. Jk . by analogy with the use of Laplace transforms to compute ztransforms. 0 D 11. the ztransform of the sequence {Ak}} is then given by AEA(A) X€A(A) k "'kk 1 12 Z({A})=L..2.2. k=O Assuming Izl > max A. One definition of the ztransform of a sequence is +00 Z({gk}t~) = LgkZk.
The symbol ( ) has the usual definition of . but again no universally "best" method be derived for the computation of matrix powers..15. let 7. the initialvalue problem initialvalue (11..1 Ak ( . A is complex. inI)(O) = CnI' (1l.15.• = AI and noting that AI and the nilpotent matrix Writing Ji = XI + N and noting that XI and the nilpotent matrix N commute. it is then straightforward to apply the binomial theorem to (AI + N)k and verify that straightforward N)k (XI verify Ak kA kI Ak k 2 (. 0 A Writing /.2 0 0 0 0 kA k . see [11. ) Ak. . y(O) = CI.2k) k( _2)k1 ] k( 2l+ (2l. 18]. Then Then 1 ] [(_2)k 1 0 k(2)kk(2) 1 ] [ _ [ (_2/.3 HigherOrder Equations HigherOrder Equations differential It is well known that a higherorder (scalar) linear differential equation can be converted to higherorder a firstorder linear system. real version of the above can be worked out.6 can also methods 11. Consider. aareal version of the above can be worked out.1(2k .1 Ak The symbol (: ) has the usual definition of q!(kk~q)! and is to be interpreted as 0 if k < q.6 be derived for the computation of matrix powers. but again no universally "best" method exists.. In the case when A. Let A = [_J Example 11. For an erudite discussion of the state of the art.3 11.. .l8) .is complex. it is commute. Ch.)A  ( k ) AkP+I pl 0 J/ = kA k. Linear Differential and Difference Equations Chapter 11. 11..1. and is to be interpreted as 0 if k < q.(^ . [11. for example.120 Chapter 11. Example 11. the problem again reduces to the computation of the power of a In the general case.17) with ¢J(t) a given function and n initial conditions 4>(t} y(O) = Co. 1 1 1 2 1 ] Basic analogues of other methods such as those mentioned in Section 11. Linear Differential and Difference Equations In the general case. To be specific. e Cpxp be a Jordan block of the form o .1 (2 . Let A Ak = XJkX1 = [=i 4 a [2 1 J]. the problem again reduces to the computation of the power of a To Ji E Cpxp Jordan block.2) .1.
as mentioned before. Suppose x.an_llnl)(t) Xnl (t) Xn(t) = y(n)(t) = aoy(t)  + ¢(t) = aOx\ (t) .. the companion Note that det(A! .718P. ... Further.19) possesses many nasty numerical properties for even moderately sized n and. Define a vector x (t) E ]Rn with Here. where !(eat . A similar procedure holds for the conversion of a higherorder difference equation A similar procedure holds for the conversion of a higherorder difference equation with n initial conditions.19) The initial conditions take the form ^(0) = c [CQ. at least for computational purposes. •. Note that det(X7 — A) = An + an\Xn 1l H alA + ao. and. the companion matrix A in (11. where I + get..a)= { a t nxn p if a if a 1= 0. Let P € R 1. Show that e'A 1+ g ( t . Then components Xl (t) yet). = Xn(t) = y(nl)(t). Cn \ . at least for computational purposes. be a projection. y E lRn and let A = xyT. Let . .A) = A.. = X3(t) = yet).19) possesses many nasty numerical properties for even moderately sized n matrix A in (11. . a)xyT. C M _I] The initial conditions take the form X (0) = C = [co.. Cl. Let P E lR nxn be a projection.a\X2(t) . into a linear firstorder difference equation with (vector) initial condition. let a = XT y.Exercises 121 121 Here. xn(t) = Inl)(t). into a linear firstorder difference equation with (vector) initial with n initial conditions. is often well worth avoiding. is often well worth avoiding. v (m) denotes the mth derivative of y with respect to t.. Let 3. y(m) denotes the mth derivative of y with respect to t. Suppose x. Show that etA 2. Further. 2. Xn(t) y { n ~ l ) ( t ) . x2(t) = y ( t ) . c\. . EXERCISES EXERCISES 1... +h a\X+ ao. However. Then Xl (I) X2(t) = X2(t) = y(t). y € R" and let A = xyT.... . . Define a vector x (?) e R" with components *i(0 = y ( t ) .I) g(t. condition.. (11. let a = xTy."+ an_1A n~ + . . X2(t) yet). aly(t) . . These equations can then be rewritten as the firstorder linear system These equations can then be rewritten as the firstorder linear system 0 0 x(t) = 0 0 1 0 0 0 ao a\ x(t)+ [ 0 1 a n\ n ~(t) r. as mentioned before.anlXn(t) + ¢(t). Show that e P ~ ! + 1.718P. However. 3.. a)xyT. Show that e % / + 1.. = O..
ft € lR and Let a. Let 5. x(O) =[ ~ J. be an eigenvalue of S. be an eigenvalue of H. (a) Suppose H is Hamiltonian and let).A 1 . 4. also eigenvalue of (c) Suppose that H is Hamiltonian and S is symplectic. . A matrix A e R 2nx2n is said to be K I AT K = . Show that eH must be symplectic.122 122 Chapter 11. 6. also be an eigenvalue of H. Let (a) Solve the differential equation (a) Solve the differential equation i = Ax .A and to be symplectic K~l AT K . Hamiltonian. must also be an eigenvalue of H. Find a general expression for Find a general expression for 7. Linear Differential and Difference where X e M'nx" is arbitrary. f3 E R and Then show that Then show that ectt _eut cos f3t sin f3t ectctrt e sin ~t cos/A J. Linear Differential and Difference Equations Chapter 11. Let K denote the skewsymmetric matrix 0 [ In In ] 0 ' In A E jR2nx2n where /„ denotes the n x n identity matrix.. Show that 1/).. (d) Suppose 5.. H (d) Suppose H is Hamiltonian. Show that E jRmxn e = [eoI A sinh 1 X ] ~I .. Find eM when A = Find etA = 8. Let a. Show S~1 H S Hamiltonian. Show that —A.be an eigenvalue of H. must also be an eigenValue of S. (b) Suppose S is symplectic and let A. Show that ). must (b) Suppose S is symplectic and let). Hamiltonian if K~1ATK = A and to be symplectic if K I ATK = A I. must (a) Suppose E is Hamiltonian and let A.be an eigenvalue of S. Let denote the skewsymmetric matrix 4. Show that SI HS must be Suppose and symplectic. Show that 1 /A.
and a quarter goes to Asia.Yet) + 2y(t) + yet) = 0. The year is 2004 and there are three large "free trade zones" in the world: Asia (A). of Cf or all t. Show that for t > 0.3. (b) Consider the difference equation (b) Consider the difference equation Zk+2 + 2Zk+1 + Zk = O. half stays home and half goes to the Americas.) (Exercise adapted from Problem 5. (a) Find the matrix M that gives (a) Find the matrix M that gives [ A] E R =M year k+1 [A] E R year k (b) Find the eigenvalues and right eigenvectors of M. x(O) = Xo for t ~ O. (d) Find the limiting distribution of the $40 trillion as the universe ends. as (d) Find the limiting distribution of the $40 trillion as the universe ends. Suppose certain multinational companies have Europe (E).YeO) = O. around the time the Cubs win a World Series). Suppose that A E ~nxn is skewsymmetric and let ex = Ilxol12. For Europe and Asia. what is the value of ZIQOO? What is the value of Zk in general? general? .11 in [24]. yeO) = 1. goes to Asia. 12. Each total assets of $40 trillion of which $20 trillion is in E and $20 trillion is in R.e. Each year half of the Americas' money stays home.e. i. The year is 2004 and there are three large "free trade zones" in the world: Asia (A). around the time the Cubs win a World Series). Consider the n x n matrix initialvalue problem 10.11 in [24].. and the Americas (R).) 12. .X(t)A. and the Americas (R). If £0 = 1 and z\ If Zo = 1 and ZI = 2. Suppose that e E"x" is skewsymmetric and let a = \\XQ\\2. Show that the eigenvalues of the solution X t ) of this problem are the same as those Show that the eigenvalues of the solution X ((t) of this problem are the same as those of C for all?.. 11. X(O) = c. Consider the initialvalue problem i(t) = Ax(t). (Exercise adapted from Problem 5. a quarter goes to Europe. and a quarter year half of the Americas' money stays home. as k —»• +00 (i. Consider the n x n matrix initialvalue problem X(t) = AX(t) . (b) Find the eigenvalues and right eigenvectors of M. Show that *(OII2 = aforallf > O. 11. Europe (E). what is the value of ZIOOO? What is the value of Zk in 2. a quarter goes to Europe. i. half stays home and half goes to the Americas. Suppose certain multinational companies have total assets of $40 trillion of which $20 trillion is in E and $20 trillion is in R.3. (a) Find the solution of the initialvalue problem (a) Find the solution of the initialvalue problem .e.e. k * +00 (i.Exercises Exercises (b) Solve the differential equation (b) Solve the differential equation i 123 = Ax + b. 10. Consider the initialvalue problem 9. I/X(t)1/2 = ex for all t > 0. (c) Find the distribution of the companies' assets at year k. For Europe and Asia.. (c) Find the distribution of the companies' assets at year k.. x(O) =[ ~ l 9.
This page intentionally left blank This page intentionally left blank .
then so is ax [ay] for any nonzero scalar a E <C.2. B E E" xn . The standard eigenvalue problem considered in Chapter 9 obviously where A. As with the standard eigenvalue problem. Definition 12. B E C MX if there exists a scalar A E C.1. such that that (12. the characteristic polynomial is obviously real.Chapter 12 Chapter 12 Generalized Eigenvalue Generalized Eigenvalue Problems 12. Definition 12. B).XB is called a matrix pencil (or pencil of the matrices A and B). and Remark 12. characteristic hence nonreal eigenvalues must occur in complex conjugate pairs. eigenvalues for the generalized eigenvalue problem occur pencil — XB problem occur where the matrix pencil A .2) When the context is such that no confusion can arise.4. Definition 12. A E en Definition 12. . In this chapter we consider the generalized eigenvalue problem In we the generalized eigenvalue problem where A.1) Ax = 'ABx.3. B) with A. The roots ofn('A) are the eigenvalues of the associated nomial of the matrix pair (A. e e. The polynomial 7r(A.2. B) with A. if x [y] is a right [left] ax [ay] for any eigenvector.4.) are the eigenvalues of the associated generalized eigenvalue problem. corresponds to the special case that B = I.1 The Generalized Eigenvalue/Eigenvector Problem The Generalized Eigenvalue/Eigenvector Problem Ax = 'ABx.3.1 12.1. Similarly. The roots ofn(X. B). As with the standard eigenvalue problem. called a generalized eigenvalue. hence nonreal eigenvalues must occur in complex conjugate pairs. e C. the adjective "generalized" "generalized" standard eigenvalue [y] is usually dropped. a. When A. The matrix A . and A.5) is called the characteristic polynomial of the matrix pair (A. 125 125 . B e jRnxn. The matrix A — 'AB is called a matrix pencil (or pencil of the matrices A Definition 12. a nonzero vector y e C" is a left generalized eigenvector corresponding to an E en generalized eigenvector eigenvalue 'X if eigenvalue A if (12.) = det(A . A nonzero vector x e C" is a right generalized eigenvector of the pair generalized eigenvector of (A. B E enxn. B). The polynomial n('A) = det(A — A. eigenvector. Remark 12.'AB is singular. (A. B e C" xn The standard eigenvalue problem considered in Chapter 9 obviously corresponds to the special case that B = I. B e enxn" if there exists a scalar 'A. generalized eigenvalue problem.'AB) is called the characteristic polyDefinition 12. called a generalized eigenvalue.
All A E C are eigenvalues since det(B . k E !!. Case 2: a = 0. Case 1: a =I.LA) == 0. there is a second eigenvalue "at infinity" for Case 3 of of . If = of degree n. {3 = 0. {3 =I.0.(3A) ±. when B is singular. There is only one eigenvalue. it is apparent where the "missing" eigenvalues have "missing" gone in Cases 2 and 3. Case 3: Case 4: = 0. Case 1: a ^ 0. n(X) Remark 12. There are two eigenvalues. f3 = O.X B is a reciprocal pencil B — n. Generalized Eigenvalue Problems Chapter 12./. suppose associated — AB. is singular. zero.L = (JL = £. =I. (3 = 0.KB always has pencil — AB .LA. There are two eigenvalues. f3 = 0.AB Definition 12. there may be 0.6. 1 and ~.5. A — A.O.I. There are two eigenvalues. If det(A — AB) not regular. I). Case 4: a = 0. Note that if AA(A) n J\f(B) ^ 0.LA and corresponding generalized eigenvalue problem. (3 = O. the characteristic polynomial is = (I . ^ 0. 1 and O. Note appear.LA) = (1 ./. With A and B as in (12.0. the pencil A — XB is said to be 12. Case 2: a = 0. (12. {3 =I. Associated with any matrix pencil A . (3 = O.5. Then the characteristic polynomial is ft det(A . Case 1: a =I. ft =I. There are two eigenvalues. Case 2: = 0.O. While While there are applications in system theory and control where singular pencils appear. There are two eigenvalues.L)({3 . 1 (of multiplicity 1). If B is singular.a/./. in particular. with its reciprocal eigenvalue being 0 in Case 3 of the reciprocal pencil B — /. All A E C are eigenvalues since det(A . I (of multiplicity 1). then rr(A) is a polynomial nonsingular).126 126 Chapter 12. B k e n.3).0. All A 6 C are eigenvalues since det(B — uA) = O. Case 4: a = 0. det(B . {3 =I. There are two eigenvalues.AHa . Case 1: ^ 0. Generalized Eigenvalue Problems Remark 12. For example.3) where a and (3 are scalars. pencil . eigenvalues associated with the pencil A . 1 Case 3: a =I. only the case of regular pencils is considered in the remainder of this chapter. f3 / 0. I1 and .6. Case 4: a = 0. A similar reciprocal symmetry holds for Case 2. regular.0./. I and O. otherwise. If B = I (or in general when B is nonsingular).5. However. and hence there are n eigenvalues associated with the pencil A . Case Case 3: a = 0.B. A similar reciprocal symmetry holds for Case 2. and ~. Case = ft ^ 0. However.{3 = 0. Clearly the reciprocal pencil has eigenvalues responding generalized /. I multiplicity 1). eigenvalues — AB.A. There are two eigenvalues. There is only one eigenvalue. All A e C are eigenvalues since det(A — AB) =0. it is said to be singular.AB) and there are several cases to consider. ft ^ O. Note that A and/or B may still be singular. {3 ^ 0. If del (A .nA.L) and there are again four cases to consider. the pencil A .O. I and ^.XB. 1 and 0. the associated matrix pencil is singular (as in Case N(A) n N(B) =Isingular 4 above). or infinitely many B = I. There are two eigenvalues. It is instructive to consider the reciprocal pencil associated with the example in It reciprocal Remark 12.AB. That is to say. when B =I. 1 and 0.XB) is not identically zero.XB.B) == O. At least for the case of regular pencils. f3 = O. only the case of regular pencils is considered in the remainder of this chapter. reciprocal Case of reciprocal .A and corAssociated with any matrix pencil — AB is a reciprocal pencil .
XB)Z] = detQ det Z det(A . 0 ( Q ~ H y)H Q(A X AB)Z = O. 6. Q. the pencil A fewer than eigenvalues. Numerical methods that if B is even moderately ill conditioned with respect to inversion. ifx isa right eigenvector of A—XB. the pencil A AAB always has precisely n . If B is nonsingular. which is the generally preferred method for solving the generalized eigenvalue problem. [7.2.7] or [25. and the first theorem deals with what happens to eigenvalues lencies rather than similarities. 3. see. for example. Sec. the eigenvalues of the pencil A . 6. Since det 0 and det Z are nonzero.7.AB)Z] = det gdet Zdet(A 1. det(QAZ . 2.AQBZ) = det[Q(A . lencies rather than similarities. see. and det Z are nonzero.8. 2.. det(QAZXQBZ) = det[0(A .7]. Again. then Z~lx isa right eigenvector of QAZ—XQ B Z. Let A.AB). Proof: Proof: 1.7. Sec. with the understanding that a zero diagonal element of Tp corresponds to an infinite generalized eigenvalue. There is also an analogue of the MurnaghanWintner Theorem for real matrices. 7.AB) o if and only if (QH y ) H Q ( A –_ B ) Z = Q. Sec. the result follows easily by noting that yH(A — XB) — 0 if and only if yH (A .2. for example.7]. of AAB. B. for example. we now deal with equivaa matrices. Numerical methods that work directly on A and are discussed in standard textbooks on numerical linear algebra.7. However. the result follows. c 3. D The first canonical form is an analogue of Schur's Theorem and forms. the theoretical foundation for the QZ algorithm. in fact. Let A.7] or [25. Canonical Forms 127 B is nonsingular.8. The result follows by noting that (A AB)x = 0 if and only if Q(A AB)Z(Zl x) = The result follows by noting that (A –yB)x . B e cnxn . ify isa left eigenvector of A —KB. in fact. ify is a left of AB. Canonical Forms 12. Sec. However. work directly on A and B are discussed in standard textbooks on numerical linear algebra. fewer than n eigenvalues. and the first theorem deals with what happens to eigenvalues and eigenvectors under equivalence. where Ta and TfJ are upper triangular. B E Cnxn Then there exist unitary matrices Q.Oif andonly if Q(AXB)Z(Z~lx) = 0. the eigenvalues of the problems A — XB and QAZ — XQBZ are the same (the two 1.l W AW). and eigenvectors under equivalence.7] or [25. Then 1. Q. [7.7]. then QHy isa left eigenvector ofQAZ AQBZ. 7. the The first canonical form is an analogue of Schur's Theorem and forms.AQBZ are the same (the two problems problems are said to be equivalent). [7. Sec. Sec.12. canonical forms are available for the generalized eigenvalue problem. . 6. By Theorem 12. the eigenvalues of the problems A . Since the latter involves a pair of matrices. E nxn with Q and nonsingular. Sec. the result follows. for example. QBZ = TfJ . Z e Cnxn with Q and Z nonsingular. the eigenvalues ofthe pencil A — XB are then the ratios of the diagonal elements of Ta to the corresponding diagonal elements of TfJ . 7. canonical forms are available for the generalized Just as for the standard eigenvalue problem. since the generalized eigenvalue problem is then easily seen to be equivalent to the standard eigenvalue problem B. E c nxn such that QAZ = Ta . fl. Let A.7. since the generalized eigenvalue problem is then easily seen to be equivalent eigenvalues. this turns out to be a very poor numerical procedure for handling the generalized eigenvalue problem out to be a very poor numerical procedure for handling the generalized eigenvalue problem if is even moderately ill conditioned with respect to inversion. see. Let A. Theorem 12. There is also an analogue of the MurnaghanWintner Theorem for real matrices. which is the generally preferred method for theoretical foundation for the QZ algorithm. with the understanding onal elements of Ta to the corresponding diagonal elements of Tp. Then there exist unitary matrices Q. Then 12. 6. Theorem 12. [7.7].2 Canonical Forms Canonical Forms Just as for the standard eigenvalue problem. o. solving the generalized eigenvalue problem. f i always has precisely eigenvalues.7] [25. that a zero diagonal element of TfJ corresponds to an infinite generalized eigenvalue. Z e Cnxn such that 12. where Ta and Tp are upper triangular. 7. see. Since det Q XB).l Ax Ax (or AB.AB and QAZ . 12. this turns to the standard eigenvalue problem B~1Ax = Xx (or AB~1w = Xw). to ifx is a Zl x is a righteigenvectorofQAZAQB Z.AB are then the ratios of the diagBy Theorem 12.2 12. Q~H y isa lefteigenvectorofQAZ — XQBZ.
In this chapter. B e c mxn .9. Generalized Eigenvalue Problems Theorem 12. L l" L~.'. real eigenvalues. . thnt that QAZ = S. Then there exist orthogonal matrices Q. QBZ = T. while the full KeF in all its generality applies also to "rectangular" and singular KCF "rectangular" pencils. There is also an analogue of the Jordan canonical form called the Kronecker canonical fonn Kronecker form (KeF). T.11. including analogues of principal vectors and description of of so forth.A [~ ~ l of . B e Rnxn.AB)Q = [~ ~ ] .9.AB where J is a Jordan canonical form corresponding to the finite eigenvalues of A A.XB is regular. The first theorem pertains only to "square" regular pencils. Let A.12 (Kronecker Canonical Form). we present only statements of the basic theorems and some examples.128 Chapter 12.fi and canonical form nilpotent matrix of associated and N is a nilpotent matrix of Jordan blocks associated with 0 and corresponding to the infinite infinite eigenvalues of A .. Otherwise.11. form (KCF)..AB.)"N). Z e R"xn such B E jRnxn. Example 12. is beyond the scope of this book.. E jRnxn 12. B E c nxn pencil — AB Theorem 12. .• L. Let A. A full description of the KeF.I. I .AB)Q = diag(LII' . The matrix pencil 12. Then there exist 12.12 mxm nxn mxm nxn E C nonsingular nonsingular matrices P e c and Q e c QE C such that peA . J . [2o I o o o 0 0 0 0 0 2 0 0 1 0 0 1 0 0 ~ ]> [~ 0 I 0 0 0 0 0 0 0 0 o o 0 I 0] 0 0 0 0 (X . .A. where T is upper triangular and S is quasiuppertriangular. of — XB. quasiuppertriangular. B e Cnxn and suppose the pencil A .10. Q € c nxn"such that nonsingular E C" such that peA . Then there x exist nonsingular matrices P. mxn E C • Theorem 12. Let A.. of eigenvalues are given as above by the ratios of diagonal elements of S to corresponding elements of T. Generalized Eigenvalue Problems Chapter 12.2)2 with characteristic polynomial (A — 2)2 has a finite eigenvalue 2 of multiplicty 2 and three 2 2 infinite eigenvalues. KCF. the 2 x 2 subpencil formed with the corresponding fonned 2 x diagonal subblock 2x2 2 diagonal subblock of T has a pair of complex conjugate eigenvalues. When S has a 2 x 2 diagonal block.
suppose S e Rn* xk is a matrix whose columns span a kdimensional E ~nxk ^dimensional subspace S of ~n.— XBif S Rn. i. Canonical Forms 12. Then V is a E ~nxn suppose pencil — AB deflating subspace if deflating subspace if dim(AV + BV) = dimV. B e Wlxn and suppose the pencil A . L6. 000 Just as sets of eigenvectors span Ainvariant subspaces in the case of the standard eigenvectors eigenproblem (recall Definition 9.2. LQ . n(S)) = S. (12. both Nand J are in Jordan canonical form. Lo. Lo L6 one column. Then is deflating subspace for the pencil A AB if and only if there exists M E Rkxk such that e ~kxk AS = BSM. Lo. (12. Specifically.. and Lk is the (k + 1) x k bidiagonal pencil bidiagonal pencil A 0 0 A Lk = 0 0 0 0 A I The Ii are called the left minimal indices while the ri are called the right minimal indices. Definition 12.The next two blocks second block L\ one the block is L\. both N and J are in Jordan canonical form. Example 12. LQ.e. generalized eigenproblem. LQ.14. are called the right minimal indices. Such a matrix is in KCF. The /( are called the left minimal indices while the r. Canonical Forms 129 where N is nilpotent. there is a matrix characterization of deflating subspace. The first block of zeros actually corresponds to LQ..XB is regular. and L^ is the (k + I) x k where N is nilpotent. there is an analogous geometric concept for the eigenproblem generalized eigenproblem.35). Consider a 13 x 12 block diagonal matrix whose diagonal blocks are A 0] I o A I .12. 0. next two correspond to correspond J = 21 0 2 [ o 0 while the nilpotent matrix N in this example is N [ ~6~].4) eigenvalue characterization Just as in the standard eigenvalue case. Lo. Let A. R ( S <S. i.5) .e. corresponds LQ. where each LQ has "zero columns" and one row.2.13. while each LQ has "zero rows" and L6. Left Left or right minimal indices can take the value O. The second block is L\ while the third block is LI. Then SS is aadeflating subspace for the pencil A .
3 12.130 Chapter 12. and y is the vector of outputs or observables. one must be well for general mUltiinput. where x(= x(t)) is called the state space model is often used in multivariable control theory. we offer some insight below into the special case of a singleinput. and D € Rpxm. Similarly. for example. If the pencil is not regular. This linear with A € M n x n . Similarly. Ac M D "'" 5A + 14.4) becomes dim (A V + V) = dim V. The method of finding system zeros via a generalized eigenvalue problem also works The method of finding system zeros via a generalized eigenvalue problem also works well for general multiinput. [26].4) becomes dim(AV + V) = dimV. for example. B] . Then the transfer matrix (see [26]) of this system is Then the transfer matrix (see [26)) of this system is g(5)=C(sIA)'B+D= 5 55 2 + 14 ' + 3s + 2 which clearly has a zero at —2. For details. Numerically. multioutput systems. (n + m) x (n + m) pencil. is a concept analogous to deflating subspace called a reducing subspace. This is accomcareful first to "deflate out" the infinite zeros (infinite eigenvalues of (12. In general. and E jRPxm. Generalized Eigenvalue Problems If B = /. then (12. which is clearly equivalent to AV c V. there AV ~ V. However. and y is the vector of outputs or observables. This linear timeinvariant statespace model is often used in multivariable control theory. the (finite) zeros of this system are given by the (finite) complex numbers In general.8. (12. trivial. Let Example 12. where x(= x(t)) is called the state vector. For details. these values are the generalized eigenvalues of the (n + m) x (n m) pencil. which has a root at —2. Numerically. Checking the finite eigenvalues of the pencil (12. we find the characteristic polynomial to be find the characteristic polynomial to be det [ which has a root at 2.15.6) drops rank. we offer some insight below into the special case of a singleinput.15. vector.6). However. u is the vector of inputs or controls.6». however. u is the vector of inputs or controls. lEthe pencil is not regular.8.3 Application to the Computation of System Zeros Application to the Computation of System Zeros i y Consider the linear system Consider the linear svstem = Ax + Bu. Checking the finite eigenvalues of the pencil (12. which is clearly equivalent to If B = I.6)). Let A=[ 4 2 C = [I 2]. In the special case p = m. In the special case p = m. Example 12. zeros). see. (12. we which clearly has a zero at 2. C e Rpxn. 12.6).8. where the "system pencil" (12. one must be careful first to "deflate out" the infinite zeros (infinite eigenvalues of (12. [26]. multioutput systems. these values are the generalized eigenvalues of the drops rank. B € R" xm . This is accomplished by computing a certain unitary equivalence on the system pencil that then yields a plished by computing a certain unitary equivalence on the system pencil that then yields a smaller generalized eigenvalue problem with only finite generalized eigenvalues (the finite smaller generalized eigenvalue problem with only finite generalized eigenvalues (the finite zeros). however. then (12. The connection between system zeros and the corresponding system pencil is nonThe connection between system zeros and the corresponding system pencil is nontrivial.5) becomes AS = SM as before. E jRPxn. D=O. = Cx + Du E jRnxn. the (finite) zeros of this system are given by the (finite) complex numbers where the "system pencil" z. there is a concept analogous to deflating subspace called a reducing subspace. see. E jRnxm.8.5) becomes AS = SM as before.
10) is equivalent Since B is positive definite it is nonsingular. Suppose z € is such that Suppose Z E C is such that [ A .10) for A.9».8) c T x +dy = O. "pole/zero cancellations").e. z is a zero of g.nxn A AT and B the B1 0. Symmetric Generalized Eigenvalue Problems 12. and D e R r T(s7 . e ffi. g(s) Furthermore. (12.zI cT b ] d is singular..zl)lby + dy = 0.9) Substituting this in (12. system of differential equations differential Mx+Kx=O.A)~ ! Z? + d denote the system transfer function (matrix).10) is equivalent B. However.8). B E Rnxn arises when A = A and B = BT > O. no pole/zero cancellations).7) we get get x = (A . Then there exists a nonzero solution to or or (A . Specifically..7) (12. C = c T E R l x n . Hence g(z) 0. and v(s) and n(s) are relatively prime TT(S) v(s) TT(S) (i. or g(z)y 0 by the definition of g. g.4.e. the secondorder A. . Thus.4 Symmetric Generalized Eigenvalue Problems Symmetric Generalized Eigenvalue Problems Ax = ABx A very important special case of the generalized eigenvalue problem (12.4 12.12. let g(.s) = c (s I — A) 1 b + d c function and assume that g(s) can be written in the form and assume that g ( s ) can be written in the form v(s) g(s) = n(s)' polynomial A.zl)x + by = 0. Now y ^ 0 (else x z i.8). Now _y 1= 0 (else x = 0 from (12. A pole/zero Assuming z is not an eigenvalue of A (i.e. B e ffi. Thus. let B = b E Rn.9)). of the Since B is positive definite it is nonsingular. 0 from (12. then from (12. we have Substituting this in (12. relatively where n(s) is the characteristic polynomial of A. For example. symmetric." is a frequently employed model of structures or vibrating systems and yields a frequently generalized eigenvalue problem ofthe form (12.10). Hence g(z) = 0. b e ffi. and D = d E R.A to the standard eigenvalue problem Bl1Ax = AJC.zl)lby.l xn. Symmetric Generalized Eigenvalue Problems 131 131 1 singleoutput system. M K where M is a symmetric positive definite "mass matrix" and K is a symmetric "stiffness definite "stiffness matrix. (12..n. we have _c T (A . B~11A is not necessarily B~ Ax = AX. or g ( z ) y = 0 by the definition of g. the problem (12. there are no "pole/zero cancellations").4. 12. the problem (12.
23).. be generalized easily to the case where A material of can. but since realvalued matrices are commonly used in most applications. B E Rnxn with A = AT and B = BT > 0. Zn Zj = Dij. Generalized Eigenvalue Problems Chapter 12. Let A = [~ .1926 and 3. The material of this section can. but since realvalued matrices are commonly used in most applications. where L is nonsingular (Theorem 10.12) Since C = C T the eigenproblem (12.fi Then it is easily checked that Then it is easily checked thai c = L~lAL~T = [ 0. l = [i ~ J B ThenB~ A Then A B~Il = [~ ~ J B~I A approximately Nevertheless.. Then the eigenvalue problem (Theorem 10.5 2. Finally. are eigenvectors of the original generalized eigenvalue problem Xi Zi.132 132 Chapter 12. zn satisfying vectors Z I. Let A Example 12. zi Then x.11) can then be rewritten as AL J and Z = LT x. Example 12. . Xj)B T T = xr BXj = (zi L ~l)(LLT)(L ~T Zj) = Dij..16 is D 0 L=[~ .. then C = C T > 0. so the eigenvalues are positive. with corresponding eigenvectors zi.12) has n real eigenvalues.1926 as expected. the eigenvalues of B l A are always real (and are approximately 2. Then the generalized A. Generalized Eigenvalue Problems Example 12. if A = A > 0. the eigenproblem (12.16 is Example 12. it has a Cholesky factorization B = LLT. B e jRnxn A AT and B BT > O. Finally. Theorem 12.1926 in Example 12.5 2. = L ~Tzi. then = C T > 0.11) can then be rewritten as = Cz = AZ. and the n corresponding right eigenvectors can be chosen to be orthogonal with respect to the inner product (x. ii € n.1926 and —3. .12) has n real eigenvalues. positive. we have restricted our attention to that case only.16). it has a Cholesky factorization B = LL T. Let A. y)BB = XT By. and are Hermitian. (12. the eigenvalue problem eigenvalue problem Ax = ABx has n real eigenvalues. Proof: Since B > 0. if A > 0. we have restricted our attention to that case only.18.. E !!. then product y) x T By.1926 whose eigenvalues are approximately 2.11) can be rewritten as the equivalent problem 1 Letting C = L ~I AL ~T and z = L1 x. The Cholesky factor for the matrix B in Example 12. . The Cholesky factor for the matrix B in Example 12.5 ] 1.16. Then the eigenvalue problem Ax = ABx = ALL Tx (12....5 ' 3. •.23).18.17. (12.. of course.16.. (12.. with corresponding eigenSince C = C T. are eigenvectors of the original generalized eigenvalue problem and satisfy and satisfy (Xi. if orthogonal > 0. the eigenvalues are also all positive. so the eigenvalues are positive. where L is nonsingular Proof: Since B > 0.fi 1] . Moreover. if A = AT> 0. generalized case A and B are Hermitian. Moreover.
Then and and QT BQ Finally. since QDQI Finally. B E lRnxn be positive definite. since A 2: B. simultaneous reduction can also be accomplished via an SVD. when L is highly ill conditioned with respect to inversion. it does preserve the eigenvalues of A .20. Now D > 0 by Theorem 10.20.1 Simultaneous diagonalization via SVD Simultaneous diagonalization via SVD There are situations in which forming C L I AL T as in the proof of Theorem 12.19.19 e ][~nxn A AT and B BT > O. it does preserve the eigenvalues of A — XB..5. = QQT AQQ~l = LTPPTL~IA = L~TL~1A L T P pT L 1 A L T L I A QQT AQQI 0 D = B1A. Simultaneous Diagonalization 12. we Note that Q is not in general orthogonal. In particular.19. Proof: Let B = LLT be the Cholesky factorization of B and set C = L~1AL T. normal maRecall that many matrices can be diagonalized by a similarity. Theorem 12. It turns out that in some cases a pair of trices can be diagonalized by a unitary similarity. Then there exists a nonsingular matrix Q such that where D is diagonal." The following is typical.< / (this is trivially true 0 since the two matrices are diagonal).19 is There are situations in which forming C = L~1AL~T as in the proof of Theorem 12. Also. by Theorem 10.1A..31. where D is diagonal. e. Let A.e. we restrict our attention only to the real case. Let Q = L~T P. Proof: By Theorem 12. the diagonal elements of D are the eigenvalues of B.1 12. Thus. Since Proof: Let T C is symmetric. there exists an orthogonal matrix P such that P CP = D. let . LetA QT AQ and B QT Then/HA Q~ B.21 we have that QT AQ 2: QT BQ. e. To illustrate. D > I.5. we restrict our attention only to the real case. In fact. This can be seen directly.1A = Q1l B~1QT QT AQ = Q11B. However. where D is C is symmetric. i. when L is highly iII conditioned with respect to inversion. To illustrate. Now D > 0 by Theorem 10. But then D. so it does not preserve eigenvalues of and B Note that Q is not in general orthogonal. In particular. QD~ QT < QQT. Let Q = L . there exists Q E lR~xn such that QT AQ = D and QT BQ = [.. matrices to "the diagonal case. such results and we present only a representative (but important and useful) theorem here. Then A 2: B if and only if B~ 2: AI. Then B. \ 2.31. Also. Q D. normal matrices can be diagonalized by a unitary similarity. Since LLT be the Cholesky factorization of and setC L I AL~T. In numerically problematic. In such cases. This can be seen directly.1AQ. straightforward way. But then D"1I :::: [(this is trivially true 10. Simultaneous Diagonalization 133 12. There are many such results and we present only a representative (but important and useful) theorem here. the diagonal elements of D are the eigenvalues of B 1A. A I < B~ . B E E"x" with 12. Thus. Proof: By Theorem 12. we B~ 1 A. where D is diagonal. Infact. Then diagonal.19 is very useful for reducing many statements about pairs of symmetric Theorem 12.e. B) can be simultaneously diagonalized by the same matrix. where D is diagonal. = pT L I(LLT)L T P = pT P = [.19 is numerically problematic.1A). with the complex case following in a straightforward way. individually.12.e. There are many matrices (A. Theorem 12.g.19 (Simultaneous Reduction to Diagonal Form). Let A = QT AQandB = QT BQ.5 Simultaneous Diagonalization Simultaneous Diagonalization Recall that many matrices can be diagonalized by a similarity.T P.. However." The following is typical. Let A.'AB. by Theorem where D is diagonal. Again. i.19 is very useful for reducing many statements about pairs of symmetric matrices to "the diagonal case. i. Then there exists a nonsingular matrix Q such that A = AT and B = BT > 0.5 12. since QDQ~l have A(D) = A(B~1A). Then A > B if and only if Bl1 > Theorem 12. i.. Theorem 12. B) can be simultaneously diagonalized by the same matrix. B e M" xn be positive definite. A1.lI QT :::: Q QT. Let A. D 2: [. simultaneous reduction can also be accomplished via an SVD. let such cases.e.5. there exists Q e E"x" such that QT AQ = D and QT BQ = I.. since A > B. It turns out that in some cases a pair of matrices (A.21 we have that QT AQ > QT BQ.g.5. D since the two matrices are diagonal). haveA(D) = A(B. with the complex case following in a Again. so it does not preserve eigenvalues of A and B individually. there exists an orthogonal matrix P such that pTe p = D. A~l :::: Bl1.l Q~T QT Q~ B~ AQ.
respectively.14) Letting x = LB z we see that (12. let A = LAL~ and B = LsLTB be Cholesky factorizations of A and B. which is thus equivalent to the generalized eigenvalue problem ALBL~LBT z.7. (12. let A = LALTA and B — LBL~ us assume that both A and B are positive definite. respectively. A can be written as A = PDP T.15) is called a generalized singular value problem and algorithms exist to problem generalized solve it (and hence equivalently (12. Remark 12. For example.21 example. but in writing = PDDp D diagonal. A straightforward. without forming the products LALTA or LBLTB explicitly. Note that LB A and thus the singular values of L B 1 LA can be found from the eigenvalue problem 02. Further.13) where E E R£ x " isisdiagonal. i.13) can be computed without explicitly forming the without Remark product indicated matrix product or the inverse by using the socalled generalized singular value decomposition (GSVD). To check this.. Various generalizations of the results in Remark 12. Generalized Eigenvalue Problems us assume that both A and B are positive definite.15) The problem (12. D may have pure imaginary elements. at least in real arithmetic. when A = AT > 0. for LB i. eigenproblem MT M x Xx.13)) and LB separately.13» via arithmetic operations performed only on LA LA (12. for generalizations results 12. note that T QT AQ = U Li/(LAL~)Li/U = UTULVTVLTUTU i/ = while L2 QT BQ = U T LB1(LBL~)Li/U = UTU = I.e. 8.21 are possible. D b .e.14) can be rewritten in the form LALAx = XLBz = Letting x = LBT Z we see 02.. Then the matrix Q == LLBTu performs the simultaneous diagonal.3].. PDPT ~ ~ ~ ~ T PD(PD{ with where Disdiagonaland P is orthogonal.134 134 Chapter 12. Compute the SVD Cholesky factorizations A B. operations performed directly on M rather than by forming the matrix MT M and solving performed MT forming the eigenproblem MT MX = AX. Further. Sec. Then the matrix Q U performs the simultaneous L e 1R~ xn diagonalization. Remark 12. The SVD in (12. Generalized Eigenvalue Problems Chapter 12. see.14) rewritten the LAL~x = ALBz = A L g L ^ L g 7 z . which is thus to the generalized eigenvalue problem 02. example. The case when A is symmetric but indefinite is not so A = AT::: O. [7. This is analogous to finding the singular values of a matrix M by Sec.21.22. products LA L ~ LBL~ see.butin writing A — PDDP T = PD(PD) with D is diagonal and P orthogonal.
(12. k = 1. quadratic) eigenvalue problem A.16) arises frequently in applications: 0. that we try to find a solution of (12. yields a polynomial of degree 2rc. K e Rnxn. Assume for simplicity that M is nonsingular. If r = n (i. ::: ILn· Let a>k = IILk I!. where q(t} e W1 and M.2 K are are ± jWk.. HigherOrder Eigenvalue Problems 12.12. polynomial 2n. seek A A2 M + AC + To get a nonzero solution /?.e.• Then the 2n eigenvalues of the secondorder eigenvalue problem A2 I /+ K Let Wk =  fjik 12 Then the 2n eigenvalues of the secondorder eigenvalue problem A. If r n (i. k = r + 1.. there are 2n eigenvalues for the secondorder (or A2 M + AC + K. KT > 0).16) can still M secondorder generalized linear be converted to the firstorder generalized linear system converted I [ o M OJ'x = [0 K I C Jx. the secondorder problem (12.16) can be written as a firstorder system (with block companion matrix) X .e.16) of the p A are to be determined. r. A special case of (12..1 Conversion to firstorder form Conversion to firstorder form Let x\ = q and \i = q. by analogy with the firstorder case. (A 2 M + AC + K) p = O. M Mwhere x(t) €. Suppose. Substituting in form q(t) = ext p. = [ M1K 0 x (t) E ~2n. or if it is desired to avoid the calculation of M lI because M is too ill conditioned with respect to inversion. . 12.16) we get (12. p. K = KT ::: 0). then all solutions of q K q 0 are oscillatory. C...16) can be written as a firstorder system (with block Let XI q and X2 Then (12.C + K is singular.16) or. then all solutions of q + Kq = 0 are oscillatory. .16) arises frequently in applications: M = I. since eAt :F 0. where the nvector p and scalar A..1 12.. Suppose K = KT.6.. ± Wk. .6.. and = = KT. Since the determinantal equation o = det(A 2 M + AC + K) = A2n + .6 HigherOrder Eigenvalue Problems HigherOrder Eigenvalue Problems Mq+Cq+Kq=O.16) Consider the secondorder system of differential equations Consider the secondorder system of differential equations q(t) E ~n E ~nxn. .2M + A. HigherOrder Eigenvalue Problems 135 12. . and A special case of (12. Suppose K has eigenvalues eigenvalues IL I ::: . n.2M + A... for which the matrix A. E2". are to be determined. (12. Then (12.6 12..6. ::: ILr ::: 0 > ILr+ I ::: . If M is singular. Since the determinantal equation is singular. Substituting in q(t) = eAt p. C = 0.C + K. we thus seek values of A.6.
Similar procedures hold for the general kthblock companion matrix analogue of (11.19). say. E Rnxm and E E 4. Suppose A € Rnxn. F 6 Rm *" G R" x . Show that the generalized eigenval". Let F e Cnxm .) . Show that the finite generalized eigenvalues of E lR " finite eigenvalues of e R™ x m the pencil [~ ~JA[~ ~J are the eigenvalues of the matrix A — BD 1 C.. Suppose A e Rnxn and D E lR::! xm.16) involving. verify Hint: An easy "trick proof is to verify that the matrices "trick proof' [Fg ~] and [~ GOF ] are similar via the similarity transformation are similar via the similarity transformation Let F E nxm G E mx ". C. . andlor K Many other firstorder realizations are possible. In the parlance of control theory. In the parlance of control theory. properties Higherorder analogues of (12. G e Cmxn • Are the nonzero singular values of FG and GF the same? same? wx E ]Rnxn. and/or K have special symmetry or skewsymmetry properties that can exploited. B e lRn*m. Generalized Eigenvalue Problems Many other firstorder realizations are possible. derivative q. the kth derivative of q. lead naturally naturally involving. Show that the generalized eigenvalues of the pencils ues of the pencils e e [~ ~JA[~ ~J and and [ A + B~ + GC ~] _ A [~ ~] are identical for all F E E"1xn and all G E R" xmm .19). Generalized Eigenvalue Problems Chapter 12. Let € C M X • Show that the nonzero eigenvalues of and G F are the same. Some can be useful when M. G E enxn".1 2.. Hint: Consider the equivalence I G][AUO F0]' B][I l [01 C (A similar result is also true for "nonsquare" pencils. Let F. which can be converted to various firstorder systems of dimension kn. such results show that zeros are invariant under state feedback or output injection. Similar procedures hold for the general k\horder difference equation order difference equation which can be converted to various firstorder systems of dimension kn.136 136 Chapter 12. EXERCISES EXERCISES nx 1. Show that the nonzero eigenvalues of FG and GF are the same. and C e lRmxn. (A similar result is also true for "nonsquare" pencils. Some can be useful when M.B D. C. to higherorder eigenvalue problems that can be converted to firstorder form using a kn x kn to higherorder eigenvalue problems that can be converted to firstorder form using aknxkn block companion matrix analogue of (11. Are the FG and GF the 3.
2 and hence are AB E2 positive. Consider the case where both A and transformation contragredient. B E e jRnxn Ql AQT ]Rnx" in such a way that Q~l AQ~T and QT BQ are simultaneously diagonal. positive. A B B are positive definite with Cholesky factorizations A = L<A and B = L#Lg. and let U~VT be an SVD of LTBLA (a) Show that Q = LA V £ ~ 5 is a contragredient transformation that reduces both contragredient = LA V~! A and B to the same diagonal matrix. (c) Show that the eigenvalues of A B are the same as those of 1.Exercises Exercises 137 137 desired 5. and let UWT be an SVD of L~LA'. A and B to the same diagonal matrix. . Such QT BQ a transformation is called contragredient. Another family of simultaneous diagonalization problems arises when it is desired Another simultaneous diagonalization problems operates that the simultaneous diagonalizing transformation Q operates on matrices A. positive Cholesky = LA L ~ = L B L ~. (b) Show that Q~l = ^~^UT LTB. respectively. respectively. Ql = ~!UTL~.
This page intentionally left blank This page intentionally left blank .
1) amnB Obviously. n 2. 4 3 4 3 4 9 4 2 6 2 6 6 6 2 2 Note that B @ A i.2. Example 13. Let B be an arbitrary 2 x 2 matrix. (13. Then the Kronecker product (or tensor Then the Kronecker product (or tensor product) of A and B is defined as the matrix product) of A and B is defined as the matrix allB A@B= [ : amlB alnB ] : E lRmpxnq.. Let A e R mx ". We restrict our attention in this chapter primarily to realvalued matrices.1. / 2 <8>fl = [o ~ l\ 2.Chapter 13 Chapter 13 Kronecker Products Kronecker Products 13. extension to the complex case only where it is not obvious. pointing out the restrict our attention in this chapter primarily to realvalued matrices. the same definition holds if A and B are complexvalued matrices. Let A E lRmxn B E R Definition 13. pointing out the extension to the complex case only where it is not obvious.2. Then A@B =[ 3~ ~]~U J.1 Definition and Examples Definition and Examples Definition 13. Foranyfl E lRX(7. Let A = [~ 2 2 nand B = [.. the same definition holds if A and B are complexvalued matrices.. Example 13.1. Then 0 b ll b12 B @/z = l b" b~l 139 0 b2 2 0 b21 0 0 b12 0 b 22 l . 1.A @ B. B e lR pxq. Forany B e!F pxq /z @ B = [~ In Replacing 12 by /„ yields a block diagonal matrix with n copies of B along the I2 diagonal with n copies of along the diagonal.1 13. We Obviously. Let B be an arbitrary 2x2 matrix. 2B 2B ~J. Note that B <g> A / A <g> B. Then 3.
simply verify using the definitions of transpose and Kronecker verify transpose Kronecker 0 product.140 Chapter 13. Let Jt € Rm. y eR". B e ~rxs.3. . simply note that (A ® B)(A 1 ® B.2) Proof: Simply verify that Proof.kCkPBD L~=1 amkckpBD ] 0 Theorem 13. C e ~nxp. . Then 13. Let E ~mxn. E R". If E ]Rn xn e Rmxm are Theorem 13. (A ® Bl = AT ® BT.5. A® 13. Then X ® Y = [ XIY T . If A e R"xn and B E !R.. C E R" x ^ and D E ~sxt. .3. L~=l al. D Corollary 13. Let* eR m . Then 13. and D e Rsxt. Theorem 13. For all A and B. y e !R.. XmY T]T = [XIYJ.1 ) Theorem 13. B In x E ~m.5. Simply verify that ~[ =AC0BD. then A® B is symmetric.m xm are symmetric. Proof: Proof: Using Theorem 13. 5.3... = 1 ® 1 = I.. Foral! Proof' Proof: For the proof.2 13. 4. XIYn. . X2Yl. Let A e R mx ". If A and B are nonsingular. 0 . .6. (A ® B)I = Bare 13.n. If AI ® B.4. Kronecker Products Kronecker Products The extension to arbitrary B and /„ is obvious. xmYnf E !R.1.. (13.6.2 Properties of the Kronecker Product Properties of the Kronecker Product (A 0 B)(C 0 D) = AC 0 BD (E ~mrxpt). . mn . 5 E R r x i .3.
• :::: TS > O.c.. xp are linearly independent right eigenvectors of A corresponding Moreover. Then A <g)B (or B 0<8> A) has rs singular values U.8. = (A 0 B)(A 0 B)T 0 Corollary 13. Then the mn eigenvalues of A® B are eigenvalues JL j..12..and let BB E e IRR mxwhave e IR nxn have eigenvalues A.3.• sin e = _ sin</> Sin</>] Then it is easily seen that A is orthogonal with eigenvalues e±jO and B is orthogonal with eigenvalues e±j</J. ./u. Lgf A E E mxn have a singular value decomposition VA ~A Theorem 13..m are linearly independent right corresponding to JJL\ . xp are linearly independent right eigenvectors of A corresponding AI.. Properties of the Kronecker Product 13.... Sine] and B . mxm /zave Theorem 13. Let A E R nx "have eigenvalues Ai..7.. then A® B is normal. Ap (p ::::: and ZI. j e q. if A and fi have Jordan form . then A 0 B is normal.. A0 B e±jeH</» e±jefJ </».[Cos</> cos</>O Then It IS easl'1y seen that . q Corollary 13. In general. If Corollary 13. • • zq independent of to A . <I :::: .. .•. . \Ju (q ::::: m). i E l!! 7 E 1· Proof: proof Proof: The basic idea of the proof is as follows: follows: (A 0 B)(x 0 z) = Ax 0 Bz =AX 0 JLZ = AJL(X 0 z). Then vI yields a singular value decomposition of A <8>B (after aasimple reordering of the diagonal yields a singular value decomposition of A 0 B (after simple reordering of the diagonal elements O/£A <8> £5 and the corresponding right and left singular vectors)... 0 Zj E€ IR mn "are linearly independent right eigenvectors of A 0 B corresponding to Ai JL 7 i e /?..p (p < n).."xn have singular values UI :::: . then Xi <8> Zj ffi.7.3 since A and B are normal by Theorem 13. if A and B have Jordan form thus get the complete eigenstructure of A <8> B. If A e IR nxn am/ B E IR mxm are normal. we can take p = nand q = m and n and q —m and If A and B are diagonalizable in Theorem 13. we can take p thus get the complete eigenstructure of A 0 B. L et A E xamp Ie 139 Let A = [ _eose cose andB .i .4 by Theorem 13...12.. Let A E lR. . and let eigenvalues jJij. A. The 4 x 4 orthogonal e±j9 orthogonal eigenvalues e±j(i>.. 141 141 Proof: Proof: (A 0 B{ (A 0 B) = (AT 0 BT)(A 0 B) = AT A 0 BT B = AAT 0 B BT by Theorem 13.n. If A E IR"xn and B eRmxm are normal. then A <g> B is € IR nxn orthogonal and e IR m x m 15 then 0 is orthogonal. and zi. matrix A ® 5 is then also orthogonal with eigenvalues e^'^+'W and e ± ^ (6> ~^ > \ Theorem 13. if Xl.j. .13. :::: U rTs > 0 and ^iT\ > • • • > ffr <s Qand rank(A 0 B) = (rankA)(rankB) = rank(B 0 A) .2.2. Example 13. i / E e!!. . if x\.12. Let A G IR mx " have a singular value decomposition l/^E^Vj an^ let and /ef singular decomposition UB^B^BB e IR pxq fi E ^pxq have a singular value decomposition V B ~B VI.. .• :::: U rr > 0 and let B E IRfx Corollary e R™x" singular a\ > • • > a > e have singular values T\ > • • > <s > 0. TTzen ?/ze mn eigenvalues of A 0 Bare Moreover.. . eigenvectors of A® B corresponding to A.. 0 If A and Bare diagonalizable in Theorem 13. ..10. . 7 E m. j € m. In general.Zq are linearly independent right eigenvectors of B corresponding to JLI. Properties of the Kronecker Product Theorem 13.8. then ...11. ••.9. Then A 0 B (or B A) has rs singular values have singular values <I :::: .JLqq (q < m). elements of ~A 0 ~B and the corresponding right and left singular vectors). Theorem 13. If A E E"xn is orthogonal and B E Mmxm is orthogonal.10.
is generally not quite in Jordan form and needs Note that JA® JB. Tr(A ® B) = (TrA)(TrB) = Tr(B ® A). Then (P ® Q)H (A ® B)(P ® Q) = (pH ® QH)(A ® B)(P ® Q) = (pH AP) ® (QH BQ) = TA ® TR . i. ~l 2 2 1 3 AfflB = (h®A)+(B®h) = 1 3 0 1 0 4 0 3 0 0 0 0 0 0 0 0 0 0 0 0 2 2 0 0 0 3 4 2 0 0 2 0 0 2 0 0 2 0 0 0 1 0 0 + 0 2 0 0 2 0 0 0 0 3 0 0 0 3 0 0 0 3 The reader is invited to compute B 0 A = (/3 ® B) + (A 0 h) and note the difference The reader is invited to compute B EEl A = (h ® B) (A <g> /2) and note the difference with A © B.e. 1. while upper triangular. in of A and B. while upper triangular.14. respectively. is the mn x mn matrix Urn <g> A) + (B ® In). det(A ® B) = (det A)m(det Bt = det(B ® A). denoted A © B. suppose P and Schur form for A ® B can be derived similarly.14.13. 2. is the mn mn matrix (Im ® A) + (B ® /„). Then 13. respectively. 1. A EEl B ^ B EEl A. i.. Corollary 13. pH AP = TA and QH BQ = TB (and similarly if and are orthogonal similarities PHAP = TA and QHBQ = TB (and similarly if P and Q are orthogonal similarities reducing A and B to real Schur form). For example. with A EEl B. Let A e Rn xn and B e Rrn xm.AP J B . suppose P and Q are unitary matrices that reduce A and B.15. Let 1. Let A~U Then Then 2 2 !]andB~[ . nxn mxm Definition 13. respectively. then we get the JA and Q~] BQ following Jordanlike structure: following Jordanlike structure: (P ® Q)I(A ® B)(P ® Q) = (P. .15. A ® B i= B © A. E IR nxn E IR mxm. to Schur (triangular) form. in general. eigenvalues are zero or nonzero). general.1 AP) ® (Ql BQ) = JA ® JB · Note that h ® JR. to Schur (triangular) form. Example 13.13. are unitary matrices that reduce A and 5.. Note that. Let A e Rn Xn and B e Rm xrn.e. denoted A EEl B.142 142 Chapter 13. respectively.I ® Ql)(A ® B)(P ® Q) = (P. Note that. of A and B. Then the Kronecker sum (or tensor sum) . Kronecker Products decompositions given by p. Kronecker Products Chapter 13. Then reducing A and B to real Schur form). Example 13. For example. E IR E IR Kronecker Definition 13. is generally not quite in Jordan form and needs further reduction (to an ultimate Jordan form that also depends on whether or not certain further reduction (to an ultimate Jordan form that also depends on whether or not certain eigenvalues are zero or nonzero). A Schur form for A ® B can be derived similarly. then we get the decompositions given by P~lI AP = J A and Ql BQ = JB.
is a Jordanlike structure for A © B... Properties of the Kronecker Product 2. 0 If A and Bare diagonalizable in Theorem 13. TTzen r/ze Kronecker sum A $ B eigenvalues e/genva/wes Al + fJt.16. .2.. j e q. xp are linearly independent right eigenvectors of A corresponding to AI.16. Ap (p < and ZI. i E !!. Properties of the Kronecker Product 13.. . ..•• . 7 e I!!. and let B E Rmx'" have e jRnxn eigenvalues A. and z\. . fJq (q < m). . then decompositions given by P~1AP = lA and Q"1 BQ = JB.. A2 + fJm. we can take p nand q and If A and B are diagonalizable in Theorem 13. if A and have Jordan form thus get the complete eigenstructure of A 0 B. .2.. then Zj ® Xi E€ jRmn" are linearly independent right Zj <8> Xi W1 are linearly independent right corresponding f j i . zq are linearly independent eigenvectors of corresponding to fJt. . In general. .. AI + fJm. .. e jRmxm eigenvalues /z. .. respectively.\ . eigenvectors of A® B corresponding to Ai + [ij.16.. . Xp (p ::s: n). . ... Then the Kronecker sum A® B = (1m (g>A) + (B ® In) has mn (Im ® A) + (B <g> /„) /za^ ran eigenvalues fJj. f^q (q ::s: ra)... if x\.13. 0 I M 0 where M = [ where M = o M a f3 f3 a J... . Then J can be written in the very compact form J Theorem 13. . if XI. if A and B have Jordan form pI l B . . Define 0 0 0 0 o o Ek = 0 o Then 1 can be written in the very compact form 1 = (4 <8>M) + (Ek ® h) = M $ E k . + fJj' € p. then decompositions given JA and Qt BQ [(Q ® In)(lm ® p)rt[(lm ® A) = [(1m ® p)I(Q ® In)I][(lm ® A) = (1m ® lA) + (B ® In)][CQ ® In)(lm ® P)] + (B ® In)][(Q ® In)(/m ® + (B ® P)] = [(1m ® pI)(QI ® In)][(lm ® A) In)][CQ ® In)(/m <:9 P)] + (JB ® In) is a Jordanlike structure for A $ B.···. . Let A E E"x" have eigenvalues Ai.xp are linearly independent right eigenvectors of A corresponding Moreover. ii E E. An + fJm' Moreover.. Proof: The basic idea of the proof is as follows: Proof: The basic idea of the proof is as follows: [(1m ® A) + (B ® In)](Z ® X) = (Z ® Ax) = (Z + (Bz ® X) ® Ax) + (fJZ ® X) = (A + fJ)(Z ® X). A2 + fJt. j E ra. we can take p = n and q = m and thus get the complete eigenstructure of A $ In general. Zq are linearly independent right eigenvectors of B AI.. (I} ® M) + (E^®l2) = M 0 Ek.. j E fl· eigenvectors of A $ B corresponding to A. Recall the real JCF M I M 143 143 0 I M I 0 o 1= 0 E jR2kx2k.. . respectively. .i e n. Recall the real JCF 2.
When symmetric.3 Application to Sylvester and Lyapunov Equations Application to Sylvester and Lyapunov Equations In this section we study the linear matrix equation In this section we study the linear matrix equation AX+XB=C.144 Chapter 13. Then to real Schur fonn). When does a solution exist? The first important question to ask regarding (13. =C.. the solution X E Wnx" is easily shown taking B = AT. j=1 These equations can then be rewritten as the These equations can then be rewritten as the mn x mn linear system x linear system A+blll bl21 A + b 2Z 1 b2ml b 21 1 (13.3 13. Kronecker Products A Schur fonn for A EB B can be derived similarly..=1 A special case of (13.e. i. [(Q ® /„)(/« ® P)] = (<2 ® P) is unitary by Theorem 13. Sylvester who studied general linear matrix equations of the form equation in honor of J.5) as (B T 0 /„). Sylvester where A e R"x". Then ((Q ® /„)(/« ® P)]"[(/m <8> A) + (B ® /B)][(e (g) /„)(/„. pH AP = TA matrices that reduce A and B.3) in terms of their easily seen z'th columns that ith columns that m AXi + Xb. i. 13. Again.5) [ blml The coefficient matrix in (13.8..3 and Corollary 13. and C e M" xm .3) is.4) obtained by taking B = AT.e. arise naturally in stability theory.3) in tenns of their columns.XB.4) is known as a Lyapunov equation. to Schur (triangular) form. This equation is now often called a Sylvester equation is now often equation in honor of 1. . PHAP = TA that reduce to Schur and QH BQ = TB (and similarly if P and Q are orthogonal similarities reducing A and B and QHBQ = TB (and similarly if P and Q are orthogonal similarities reducing A and B to real Schur form). Lyapunovequations arise naturally in stability theory.3 and Corollary 13. = C. When C is symmetric. an "ordinary" linear system.1. .3) is. where [(Q <8>In)(lm ® P)] = (Q ® P) is unitary by Theorem 13. Kronecker Products Chapter 13. = AXi + l:~>j. (13. Sylvester who studied general linear matrix equations of the fonn k LA.5) clearly can be written as the Kronecker sum (Im * A) + (BT ® In). The following definition is very helpful in completing the writing of (13. solution e IR xn also to be symmetric and (13. B e Rmxm . respectively. Again. The first important question to ask regarding (13.4) is known as a Lyapunov equation.5) as an "ordinary" linear system. ® P)] = (/m <8> rA) + (7* (g) /„). . it is easily seen by equating the writing (13. Lyapunov equations also to be symmetric and (13.J.8. The following definition is very helpful in completing the writing of (13. When does a solution exist? By writing the matrices in (13.5) clearly can be written as the Kronecker sum (1m 0 A) + The coefficient matrix in (13. suppose P and are unitary A Schur form for A © B can be derived similarly.Xj.3) is the symmetric equation AX +XAT = C (13.3) mxm E IRnxn E IR E IRnxm. suppose P and Q are unitary fonn.
Assuming that. +00): IHoo lim XU) . Application to Sylvester and Lyapunov Equations 145 145 Definition 13.18.8)by Theorem 13.6). First A and B are reduced to (real) Schur form. has a unique solution if and only if A and —B have no eigenvalues in common. (13. say. Theorem C E jRnxm. AX+XB=C (13. They culminate in Theorem 13. First A and B are reduced to commonly preferred numerical algorithm is described in [2]. n :::: m. A further enhancement to this algorithm is available in [6] whereby Gaussian elimination.16. We thus have the following theorem. Then the (unique) solution of the Sylvester equation AX+XB=C (13.1S.e.6) There exists a unique solution to (13.3. where From Theorem 13. But [(1m ® A) + (B (g) /„)] nonsingular and only has no zero eigenvalues.19. Assuming that. where A. c E jRn the Then vec(C) is defined to be the mnvector formed by stacking the columns ofC on top of by C ::~~::~: ::d~~:::O:[]::::fonned "ocking the colunuu of on top of one another. and C e R" xm . E!!. so there exists unique Proof: Since A and B are stable.9) Proof: Since A and B are stable. From Theorem 13.B have no eigenvalues in common.(B) ^ solution to(13.6).3) (or symmetric Lyapunov equations of the form Sylvester equations of the form (13.4» are generally not solved using the mn x mn "vec" formulation (13. The most commonly preferred numerical algorithm is described in [2].5) can be rewritten in the form [(1m ® A) + (B T ® In)]vec(X) = vec(C). Cm}.8) can be written as can be written as (13. the linear system (13..18. Aj(A) + A. and ^j Theorem 13. A further enhancement to this algorithm is available in [6] whereby the larger of A or B is initially reduced only to upper Hessenberg rather than triangular the larger of A or B is initially reduced only to upper Hessenberg rather than triangular Schur form..13. (A)+ Aj(B) =I 00 for all i.18.6) if and only if [(Im ® A) + (BT ® /„)] is nonsingular. the eigenvalues of [(/m <g> A) + (BT ® In)] are Ai A. E jRmxm. 77ie/i Theorem 13.. Let A e jRnxn. A. The next few theorems are classical.X(O) = A 10 roo X(t)dt + ([+00 X(t)dt) 10 B. Now integrate the differential equation X AX XB (with X(O) C) on [0.6) if and only if [(1m ® A) + (B T ® In)] is nonsingular.3.24. say.10) . ofC e jRnxm [CI. and C e Rnxm. Ai E A(A). B e Rmxm. xn Theorem 13. j j so there exists aaunique for all i.4)) are generally not solved using the mn x mn "vee" formulation (13.17.. Sylvester equations of the form (13. Let A e lRnxn. e m. c ]. one of many elegant connections between matrix theory and stability theory for differential equations. They culminate in Theorem 13. .. Suppose further that A and B E Rn . this algorithm takes only O(n3 ) operations rather than the O(n6)) that would be required by solving (13.e A (A). . Suppose further are asymptotically stable (a matrix is asymptotically stable if all its eigenvalues have real are asymptotically stable (a matrix is asymptotically stable if all its eigenvalues have real parts in the open left halfplane).8) by Theorem 13. Now integrate the differential equation X = AX + X B solution to (13. The most (13.17. the eigenvalues of [(1m ® A) + (BT <8> /„)] are + Mj. Definition 13. An equivalent linear system is then solved in which the triangular form equivalent linear system is then solved in which the triangular form of the reduced and can be exploited to solve successively for the columns of a suitably of the reduced A and B can be exploited to solve successively for the columns of a suitably transformed solution matrix X.. A(fi). But [(Im <8>A) + (B TT ® In)] isisnonsingular ififand only ififitithas no zero eigenvalues.6) directly with Gaussian elimination. ii e n_.17. There exists a unique solution to (13.and Mj Ee A(B). Then the Sylvester equation G jRmxm. .. Application to Sylvester and Lyapunov Equations 13.16.24. n > m. Let Ci( € E.17. +00): (with X(0) = C) on [0.n denote the columns ofC E Rnxm so that C = [ n . (real) Schur form. .7) has a unique solution if and only if A and . (13.3) (or symmetric Lyapunov equations of the form (13. + IJLJ. j j E!!!. vec(C) = Using Definition 13.6) directly with operations rather than the O(n 6 that would be required by solving (13. i. B E Rmxm.. We thus have the following theorem. E R E jRnxm. Schur form. one of many The next few theorems are classical.. elegant connections between matrix theory and stability theory for differential equations.. the linear system (13. this algorithm takes only 0 (n 3) transformed solution matrix X. .5) can be rewritten in the form Using Definition 13. Then the (unique) solution of the Sylvester equation parts in the open left halfplane).
C E R"x" and suppose further that A is asymptotically stable. Remark 13.. By Theorems 13. —kn. X B = is that [ J _Cfi ] be similar to [~ _OB] (via the similarity [ Let Theorem 13..8). Theorem 13. Remark 13. the first of which follows immediately from Theorem 13. Now let v be an arbitrary nonzero vector in jRn. An.23. Theorem 13.. v E".146 146 Chapter 13. Lef A. Hence.1. C E R"x". Many useful results exist concerning the relationship between stability and Lyapunov equations.21 and 13... If matrix A e jRn xn eigenvalues )"" .13) where C = C T < O. sufficient —A common eigenvalues A asymptotically no common eigenvalues is that A be asymptotically stable.I ..12). using the solution X ((t) = elACe tB from Theorem 11.21 l3. it can be shown easily that lim elA = lim elB = O. A.23 a solution to (13.]. An equivalent condition for the existence of a unique solution to AX + AX + Remark XB = C is that [~ _cB ] be similar to [ J _°B ](via the similarity [~J _~ ]).24..6.13) exists and takes the form (13. symmetric and ( 13. If symmetric and (13. Kronecker Products Chapter 13. results = 0. Kronecker Products Using the results of Section 11.6. .12) Theorem 13. 1>+00 1 .ATT have A —A.A T have no eigenvalues in common. Let A.. a sufficient condition that guarantees that A and . Theorem Substituting in (13.AT has eigen— AT eigenvalues AI. .10) we have C t~+x /—<+3C = A (1+ 00 elACe lB dt) + (1+ o 00 elACe lB dt) B and so X and so X = 1o {+oo elACe lB dt satisfies (13. _* ]). If C is has unique if and only if and —A T eigenvalues in common.11) has a unique solution if and only if A and .. A matrix A E R"x" is asymptotically stable if and only if there exists a only if e jRnxn asymptotically if positive definite solution to the Lyapunov equation positive definite solution to the Lyapunov equation AX +XAT = C.. . TTzen r/ze AX+XAT =C (13.20. If the matrix A E Wxn has eigenvalues A.21.22. then . . where C Proof: asymptotically l3.23 solution Proof: Suppose A is asymptotically stable.11) has a unique solution. Then the Lyapunov equation e jRnxn.19.. Thus.An. . then that solution is symmetric. then that solution is symmetric. Two basic results due to Lyapunov are the following.. . we have that lim X ((t) = 0. C e jRnxn further asymptotically stable. Then the (unique) solution o/the Lyapunov equation of the AX+XAT=C can be written as can be written as (13. (13.!„. .11) has a unique solution. +00 r—>+oo t—v+oo X t ) = etACelB X t ) — O.19. Then Then .
C. The Lyapunov equation AX + XATT = C can also be written using the vec notation in the equivalent form vec notation in the equivalent form [(/ ® A) + (A ® l)]vec(X) = vec(C).16) .yr) = <8> x. A must be asymptotically stable.t. we must have A + A = 2 R e A < O.3. v TXv > 0 and thus X is positive definite. the complexvalued equation AHX + XA = C is equivalent to [(/ ® AH) vec(C).14) is unique if BB+ ® A+ A = I. The solution of (13. for the solution of the simple Sylvesterlike equation introduced in Theorem 6. Since yH Xy > 0. the integrand above is positive. B. D Remark 13. Then the equation 13. in which the solution is of the form is of the form (13. we must have A + I = 2 Re A < 0 . and C for which the matrix product ABC is Theorem 13.25. B e jRPxq. suppose X = XT > 0 and let A.13. Then vector y. B E Rpx(}. result.3. Hence Since C > 0 and etA is nonsingular for all t. Then 0> yHCy = yH AXy + yHXAT Y = (A + I)yH Xy.26.26. The Proof: Write (13. The vec operator has many useful properties. The equivalent "vec form" of this equation is The equivalent "vec form" of this equation is [(/ ® AT) + (AT ® l)]vec(X) = + (AT ® l)]vec(X) = vec(C). and C for which the matrix product ABC is defined.15) of (13. D tions or from the fact that vec(xyT) = y ® x.11.14) as Proof: Write (13. Theorem 13. vec(ABC) = (C T ® A)vec(B). and C E Rmxq. For any three matrices A. defined. However. most of which derive from one key The vec operator has many useful properties. The Lyapunov equation AX X A = C can also be written using the Remark 13. the complexvalued equation H X X A = C is equivalent to However. Proof: The proof follows in a fairly straightforward fashion either directly from the definiProof: The proof follows in a fairly straightforward fashion either directly from the definitions or from the fact that vec(. 14) is unique if BB+ ® A+A = [. Theorem 13. the AXB =C (13. nx p if and only if A A+CB+BB = C. For any three matrices A. where Y e Rnxp is arbitrary. in which case the general solution has a if only ifAA + C B+ C. B.25. D asymptotically stable. suppose X = XT > 0 and let A E A (A) with corresponding left eigenConversely. Let A E Rmxn. A must be Since yHXy > 0.11. e jRrnxq. A subtle point arises when dealing with the "dual" Lyapunov equation A T X X A A subtle point arises when dealing with the "dual" Lyapunov equation ATX + XA = C. e A(A) with corresponding left eigenvector y.27. Since A was arbitrary. e jRrnxn. where Y E jRnxp is arbitrary. Hence vT Xv > 0 and thus X is positive definite. Application to Sylvester and Lyapunov Equations 13. Since A was arbitrary.14) xp E jRn has a solution X e R.14) as (B T ® A)vec(X) = vec(C) (13. Application to Sylvester and Lyapunov Equations 147 147 Since — C > 0 and etA is nonsingular for all the integrand above is positive. D An immediate application is to the derivation of existence and uniqueness conditions An immediate application is to the derivation of existence and uniqueness conditions for the solution of the simple Sylvesterlike equation introduced in Theorem 6. Conversely.27. most of which derive from one key result.
148 148
Chapter 1 3. Kronecker Products Chapter 13. Kronecker Products
by Theorem 13.26. This "vector equation" has a solution if and only if by Theorem 13.26. This "vector equation" has a solution if and only if
(B T ® A)(B T ® A)+ vec(C)
+
= vec(C).
+ +
It is a straightforward exercise to show that (M ® N) + = M+ ® N+.. Thus, (13.16) has aa It is a straightforward exercise to show that (M ® N) = M <8> N Thus, (13.16) has
solution if and only if solution if and only if vec(C)
=
(B T ® A)«B+{ ® A+)vec(C)
= [(B+ B{ ® AA+]vec(C)
= vec(AA +C B+ B)
and hence if and only if AA +CB+B = C. and hence if and only if AA+ C B+ B C. The general solution of (13 .16) is then given by The general solution of (13.16) is then given by vec(X) = (B T ® A) + vec(C)
+ [I 
(B T ® A) + (B T ® A)]vec(Y),
where Y is arbitrary. This equation can then be rewritten in the form where Y is arbitrary. This equation can then be rewritten in the form vec(X)
= «B+{
® A+)vec(C)
+ [I
 (BB+{ ® A+ A]vec(y)
or, using Theorem 13.26, or, using Theorem 13.26,
The solution is clearly unique if B B+ ® A + A ==I. The solution is clearly unique if BB+ <8> A+A I.
0 D
EXERCISES EXERCISES
I. For any two matrices A and B for which the indicated matrix product is defined, 1. For any two matrices A and B for which the indicated matrix product is defined, show that (vec(A»T(vec(fl)) = Tr(A T B). In particular, if B E Rn x n ,, then Tr(B) = show that (vec(A)) r (vec(B» = Tr(A r £). In particular, if B e lR nxn then Tr(fl) = vec(/J r vec(fl). vec(Inl vec(B). 2. Prove that for all matrices A and B, (A ® B)+ = A+ ® B+.. 2. Prove that for all matrices A and B, (A ® B)+ = A+ ® B+
3. Show that the equation AX B = C has a solution for all C if A has full row rank and 3. Show that the equation AX B = C has a solution for all C if A has full row rank and B has full column rank. Also, show that a solution, if it exists, is unique if A has full B has full column rank. Also, show that a solution, if it exists, is unique if A has full column rank and B has full row rank. What is the solution in this case? column rank and B has full row rank. What is the solution in this case? 4. Show that the general linear equation 4. Show that the general linear equation
k
LAiXBi =C
i=1
can be written in the form can be written in the form
[BT ® AI
+ ... + B[ ® Ak]vec(X) =
vec(C).
Exercises Exercises
149 149
5. Let x E ]Rm and y E E". Show that *rT ® yy==y X T T. x <8> € Mm e ]Rn. yx •
6. Let A e R" xn and £ e M m x m . (a) Show that IIA ® BII22 = IIAII2I1Blb. (a) Show that A <8> B = A2£2. (b) What is II A ® B II F in terms of the Frobenius norms of A and B? Justify your (b) What is A ® B\\F in terms of the Frobenius norms of A and B? Justify your answer carefully. answer carefully.
(c) What is the spectral radius of A ® B in terms of the spectral radii of A and B? of A <8> B in terms of the spectral radii of A and B? Justify your answer carefully. Justify your answer carefully. 7. Let A, 5 eR" x ". 7. Let A, B E ]Rnxn.
A)k = / <8> A* and (fl <g> l = B® I for all integers k. (a) Show that (l ® A)* = I ® Ak and (B ® I /)* =Bk fc ® / for all integers &. (/ l A (b) Show that el®A = I ® eeA and eB®1 7= eeB ® I./. e® <g) A and e5® = B (g)
(c) Show that the matrices I ® and (c) Show that the matrices / (8)AA andBB® I /commute. ® commute. (d) Show that (d) Show that
e AEIlB
= eU®A)+(B®l) = e B ® e A .
(Note: This result would look a little "nicer" had we defined our Kronecker (Note: This result would look a little "nicer" had we defined our Kronecker sum the other way around. However, Definition 13.14 is conventional in the 13.14 literature.)
8. Consider the Lyapunov matrix equation (13.11) with
A =
and C the symmetric matrix and C the symmetric matrix
[~ _~ ]
[~
Xs
Clearly Clearly
=
[~ ~ ]
[_~ ~
]
is a symmetric solution of the equation. Verify that is a symmetric solution of the equation. Verify that
Xns =
is also a solution and is nonsymmetric. Explain in light of Theorem 13.21. is also a solution and is nonsymmetric. Explain in light of Theorem 13.21. 9. Block Triangularization: Let 9. Block Triangularization: Let
A E ]Rn xn find similarity where A e Rnxn and D E ]Rm xm. It is desired to find a similarity transformation e Rmxm. of the form of the form
T=[~ ~J
such that T l1ST is block upper triangular. such that T ST is block upper triangular.
150 150 (a) Show that S is similar to
Chapter 13. Kronecker Products Chapter 13. Kronecker Products
[
A +OBX
B ] DXB
if X satisfies the socalled matrix Riccati equation if X satisfies the socalled matrix Riccati equation
CXA+DXXBX=O.
(b) Fonnulate a similar result for block lower triangularization of S. Formulate S.
to. Block 10. Block Diagonalization: Let
S=
[~ ~
l
where A E Rnxn and D E R m x m . It is desired to find a similarity transfonnation of e jRnxn E jRmxm. transformation of the fonn form
T=[~ ~]
such that T l1ST is block diagonal, T ST block diagonal. (a) Show that S is similar to
if Y satisfies the Sylvester equation Y
AY  YD = B.
(b) Formulate a similar result for block diagonalization of Fonnulate of
. 1970. Philadelphia.W. 1996. Cambridge." SIAM Rev. T. 1970.E.H. R.H. AC24(1979). N.R. [10] Halmos. 2002. Univ. Introduction to Matrix Analysis. NY. NJ.. Baltimore. and lH. R.R. "IllConditioned Eigensystems and the Computation [8] Golub. SIAM Rev. and C. A. Johnson. "A HessenbergSchur Method for the Problem [6] Golub. 9(1967). Matrix Analysis. New York.E. G. SIAM 9(1967). Cambridge Univ. and c. Numerical Methodsfor Least Squares Problems.Bibliography [1] Albert. NJ.F. Cambridge. [3] Bellman. Cambridge. Wilkinson. 6(1964). Wilkinson... 18(1976). of the SIAM 18(1976). Cambridge Univ. G. "Note on the Generalized Inverse of the Product of Matrices. Johns Hopkins [7] Golub. PA. [10] Halmos. and C..R. Third Edition. SIAM. 249]. [5] Cline. Stewart. and G.H. S. UK. 1996. Philadelphia. Numerical Methods for Least Squares Problems. MD. 1996. York. 1991.. Cambridge Univ.. McGrawHill. N.A.H.518521 [Erratum. G. S.H. Press.5758. New [3] Bellman. "Algorithm 432. RH." SIAM Rev. New [1] York. FiniteDimensional Vector Spaces. [9] Greville. Academic Press.. 1991.. Topics in Matrix Analysis. [4] Bjorck. Accuracy of'Numerical Algorithms.. R.820826. Press. 8(1966). "Note on the Generalized Inverse of the Product of Matrices.. ACM.R. 8(1966). [13] Hom.." IEEE [7] Golub. 909913.E. "A HessenbergSchur Method for the Problem AX + XB = C. Philadelphia.. Van Loan.1. T." SIAM Rev. Van Loan.H.. SIAM.A.. G. McGrawHill. 151 151 . 15(1972). SIAM. 1996. NY.578619.." IEEE Trans. 6(1964). Princeton.. 1972. Stewart. "Note on the Generalized Inverse of a Matrix Product. R. Second Edition. 1972. Nash.. [9] Greville. 1958.J. [8] Golub. 1958.. Johns Hopkins Univ.w. Matrix Analysis. Horn. [6] Golub. [11] Higham. UK. Princeton. Van Nostrand. [2] Bartels. 1985. Baltimore.. FiniteDimensional Vector Spaces. P. Analysis. Press. A." SIAM Rev. G. R. "Note on the Generalized Inverse of a Matrix Product.. 57–58. Regression and the MoorePenrose Pseudoinverse. Van Loan. RA. Second Edition. 2002. [11] Higham. SIAM. Johnson. Van Loan.N. Autom.. Second Edition. UK. Johnson." Comm." SIAM Rev. AX + XB = C. "IllConditioned Eigensystems and Computation ofthe Jordan Canonical Form. Matrix Computations. Accuracy and Stability ofNumerical Algorithms. Univ. R. and C. and J.. Second Edition.. 15(1972). Van Nostrand." Cornm. PA.. Press. R. AX X B = C. [4] Bjorck.. NY.R. G. and G. Solution of the Matrix Equation [2] Bartels. Solution Equation AX + X B = C. [5] Cline. to Second Edition. and C.H. Cambridge. UK. 249].H. [12] Hom.. NY..F. "Algorithm 432. 820826. and C. York. Control. Horn. C.N. Nash. RA. MD.E. Second Edition. PR. PA. Philadelphia. Third Edition. 518–521 [Erratum. and C. 578619. Matrix Computations. AC24(1979). 1985. Press.
San Diego. Analysis Applied Linear [18] Meyer. FL.M." SIAM Rev. Academic Press. 1985." IEEE Trans . "A Generalized Inverse for Matrices. NJ. Orlando. . 1(1988). Fundamentals of Matrix Computations. PA. and c. J. Daniel. 801836. Control. 2000. Philadelphia. W. and C. [15] Kenney. R. CA.J. Tismenetsky. Control. NY." IEEE Trans.F. [19] Moler. NY. Laub. Harcourt Brace [24] Strang. P. Matrix Analysis and Applied Linear Algebra. W. San Diego. [22] Penrose.J.. A. [22] Pemose. Linear Algebra and Its Applications. 2000. Van Loan. "The Matrix Sign Function. 1988.. P. 1973. The Theory of Matrices. A Second Course. New York. to Academic Press.. [18] Meyer. Autom. Third Edition.. and J..152 152 Bibliography Bibliography [14] Kenney. NJ. Linear Multivariable Control. 20(1978). [16] Lancaster. Laub. "Nineteen Dubious Ways to Compute the Exponential [19] Moler. C. Autom. "Controllability and Stability Radii for Companion Form Systems." SIAM 20(1978).361390.. "Nineteen Dubious Ways to Compute the Exponential of a Matrix. C. Academic Press. 361390.. Laub. [15] Kenney. Third Edition. Daniel. Soc. D. Laub. Control. Signals. 1(1988).. [20] Noble. 1979). Harcourt Brace Jovanovich. 1987. G. 1973.W. Academic Press. of a Matrix. G. 913921. G. and AJ. FL. 1985. SIAM. 40(1995). Second Edition.. SpringerVerlag.. SpringerVerlag. c. C. and A. G.B. andAJ.. Applied Linear Algebra.B.. New York. "The Matrix Sign Function. Third Edition. Second Edition." IEEE Trans. New York. [17] Laub.J.M.S.. "A Schur Method for Solving Algebraic Riccati Equations.w. and M. "A Schur Method for Solving Algebraic Riccati Equations. NY. Third Edition. New York. [26] Wonham.." Proc. 1985. C. "Controllability and Stability Radii for Companion Fonn [14] Kenney. PrenticeHall. R. Control. Interscience. Matrix Theory.W. C.P.. New York. and A. [24] Strang...D. [17] Laub.. Interscience. New York. Linear Algebra and Its Applications.. New York.S. Applied Linear Englewood Cliffs..S. AJ. of Control. and J. 2002. Systems. B. [23] Stewart.801836. Tismenetsky. Autom. 51(1955). Orlando. 406–413. 1987. [21] Ortega.. 1988. C.. Soc. 1985.13301348.. Applications. and Systems. A Geometric Approach. 913–921.." Proc. NY." Math.. Third Edition. [23] Stewart. PA. [21] Ortega. 1330–1348. Philadelphia. [20] Noble. CA. of [25] Watkins. SIAM. and M. Jovanovich. 1988.406413." IEEE Trans. Wiley[25] D. Plenum.. W. Cambridge Philos. PrenticeHall. "A Generalized Inverse for Matrices.. NY. 40(1995).. Englewood Cliffs. Theory of Second Edition with Applications. AC24( 1979). Second Edition with [16] Lancaster. 1988. Third Edition. 51(1955). Van Loan. of Control." Math.S. Introduction to Matrix Computations. [26] Wonham.D. NY. 2002. Plenum. C.
21 orthogonal. 58 basis. 101 codomain. 106 pseudoinverse of. 149 congruence. 109 computation of. 90 matrix characterization of. 23 vector. 95 orthogonal. 12 dimension. 106 singular vectors of. 81 generalized eigenvalue. 11 natural. 5 triangularization. 11 basis. 109 exponential of a matrix. 13 domain. 84 equivalence transformation. 91. 1 vector. 48 LV factorization. 125 generalized real Schur form. 87 of eigenvectors. 12 direct sum direct sum of subspaces. 12 block matrix. 1 CauchyBunyakovskySchwarz InequalCauchy–Bunyakovsky–Schwarz Inequality. 115 exponential of a Jordan block. 109112 field. 137 controllability. 23 function of a matrix. 106 singular values of. 7 field. 115 exponential of a matrix. 106 complement complement of a subspace. 125 generalized eigenvalue. 1 . 101 Cholesky factorization. 105 pseudoinverse of. 85 of a principal vector. four fundamental subspaces. 76 defective. 17 column column rank. 95 unitary. 150 diagonalization. 2 contragredient transformation. 110 properties of. 81 function of a matrix. 21 153 . 127 equivalent matrix pencils. 4 determinant. 104 diagonalization. 5 LU factorization. 12 natural. 75 chain chain of eigenvectors. 23 rank. 127 exchange matrix. 128 e (pmxn mxn en. 128 generalized real Schur form. 1 companion matrix companion matrix inverse of. 5 properties of. 105 inverse of. 95 unitary. 137 contragredient transformation. 127 lems. 95 equivalence transformation. 81 mation. 58 CayleyHamilton Theorem. 2 conjugate transpose. i 1 (p/nxn 1 e~xn. 104 definiteness of. 81. 109–112 properties of. 75 Cayley–Hamilton Theorem. 23 four fundamental subspaces. 114118 inverse of. 89 matrix characterization of. 17 co–domain. 75 invariance under similarity transforinvariance under similarity transformation. 85 determinant. 75 of a matrix pencil. 13 of subspaces. 75 eigenvalue. 95 equivalent generalized eigenvalue probequivalent generalized eigenvalue problems. 17 eigenvalue. 13 of a subspace. 89 exponential of a Jordan block. 46 defective. 17 domain. 75 of a matrix. 39. 103 conjugate transpose. 13 orthogonal. 4 of a block matrix. 90 algebraic multiplicity. 76 algebraic multiplicity. 76 degree degree of a principal vector. 89 exchange matrix. 87 characteristic polynomial characteristic polynomial of a matrix. 76 angle between vectors. 4–6 dimension. 150 inverse of. 103 congruence. 58 angle between vectors. 106 singular values of. 114–118 computation of. 46 properties of. 2 block matrix. 84 elementary divisors. C". 125 Cholesky factorization. 125 of a matrix pencil. 39. 46 controllability. 110 inverse of. 81.58 ity. 89 A–invariant subspace. 48 inverse of. 2 definiteness of. 91.Index Index Ainvariant subspace. 5 of a block matrix. 106 singular vectors of. equivalent matrix pencils.81 elementary divisors.
54 weighted. 2 transpose. 2 Jordan block. 134 134 geometric geometric multiplicity. 54 usual. 82 Kronecker canonical form (KCF). 140 Kronecker sum. 4 inertia. 10 linear equations linear equations characterization of all solutions. 55 complex Euclidean. 2 idempotent. 118 for homogeneous linear differential homogeneous equations.2 i. 51 identity matrix. 141 singular trace of. 85 linear dependence. left nullspace. 125 left invertible. 140 pseudoinverse of. 54 real. 17 composition of. 119 for inhomogeneous linear differeninhomogeneous differential equations. 66 geometric solution of. 26 invertible. 75 left generalized eigenvector. 51 idempotent. 129 Kronecker canonical Kronecker delta. 17 invertible. 44 existence of solutions. 44 characterization of all solutions. 17 co–domain codomain of. 143 exponential of. 121 conversion firstorder higher–order differential equations higherorder conversion to first–order form. 142 transpose of. 65 solution via QR factorization. 65 uniqueness of solution. 67 residual of. 148 singUlar values of. 20 Kronecker . 112 equations.2 7. 120 higher–order for homogeneous linear difference homogeneous equations. 65 problem. 71 71 decomsolution via singular value decomposition. 103 initialvalue problem. generalized decomposition. 112 for inhomogeneous linear difference for inhomogeneous linear difference equations. 66 linear regression. 25 left invertible. 20 nullspace j. 22 left left principal vector. 84 inverses of block matrices.154 generalized Schur form. 25 nulls pace of. 100 left eigenvector. 26 invertible. 4. 44 existence of solutions. 67 linear transformation. 143 eigenvectors of. 58 Hermitian transpose. 45 linear independence. 142 eigenvalues of. 45 uniqueness of solutions. 10 linear independence. 19 domain of. 44 uniqueness of solutions. 82 Jordan canonical form (JCF). 47 Index Index Kronecker product. 141 eigenvectors of. matrix representation of. 142 determinant eigenvalues of. 136 firstorder i. 6. 141 eigenvectors products of. general solution of. 70 statement of. 120 firstorder higherorder eigenvalue problems higher–order problems conversion to first–order form. 18 nonsingular. 143 eigenvectors of. 149 leading principal submatrix. 10 linear least squares problem. inner product inner product complex. 76 Holder Inequality. 54 invariant factors. 139 determinant of. 127 generalized generalized singular value decomposition. higherorder higher–order difference equations conversion to first–order form. 109 initial–value for higherorder equations. 4 Euclidean. 4 complex Euclidean.
3 MumaghanWintner Murnaghan–Wintner Theorem. matrix–vector. 2 Householder. 46 observability. 145 best rank k approximation to. singular. 125 equivalent. 6 block. 61 61 relations among. 91 minimal polynomial. 25 . 61 Frobenius.Index Index range of. 99 lower Hessenberg. 29 multiplication multiplication matrixmatrix. 24 nullspace. 91 of square root of a. 60 induced by a vector norm.56 natural. 146 uniqueness of solution. 2 exponential. 4 upper Hessenberg. 2 matrix exponential. 127 reciprocal.5 block. 59 1. 109 matrix norm. 33. conditions for. 146 symmetry of solution. 20 right invertible. 144 equation. 91. 26 LV LU factorization. 60 spectral. 61 Schatten. 2