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
Matrix Analysis Matrix Analysis for Scientists & Engineers for Scientists & Engineers Alan J. Laub Alan J. Laub University of California Davis. California slam. .
PA 191042688. 3 Apple Hill Drive. 1948Matrix analysis for scientists and engineers / Alan J... Matrices. 2005 by the Society for Industrial and Applied Mathematics. p. stored. Library of Congress CataloginginPublication Data Library of Congress CataloginginPublication Data Laub.L38 2005 512.. Matrices.mathworks. Inc. cm. No part of this book All rights reserved. artist Aaron Tallon of Philadelphia.. For information. MATLAB® is a registered trademark of The MathWorks.. For MATLAB product information. PA 191042688.9'434—dc22 512.com 5086477000. 5086477000. www. MA 017602098 USA. PA. I.. stored.com Mathematica is a registered trademark of Wolfram Research. ISBN 0898715768 (pbk. Mathematical analysis. Alan J. No part of this book may be reproduced. 10987654321 10987654321 All rights reserved. write to the Society for Industrial and Applied Mathematics. . cm. Natick.. Used by permission. info@mathworks.) 1. or transmitted in any manner without the written permission of the publisher. write to the Society for Industrial and Applied of the publisher. info@mathworks. Includes bibliographical references and index. 2. PA. Mathematica is a registered trademark of Wolfram Research.) ISBN 0898715768 (pbk. Philadelphia. For MATLAB product information.com. Alan J. Title. Printed in the United States of America. Laub. or transmitted in any manner without the written permission may be reproduced. Natick. QA188138 2005 QA 188.Copyright Copyright © 2005 by the Society for Industrial and Applied Mathematics. • slam is a registered trademark. Fax: 5086477101. wwwmathworks.com. Philadelphia. MA 017602098 USA. MATLAB® is a registered trademark of The MathWorks. Mathcad is a registered trademark of Mathsoft Engineering & Education. 3600 University City Science Center. 1948Laub. please contact The MathWorks. p. Mathematical analysis.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. Laub. 3600 University City Science Center. is a registered trademark. Fax: 5086477101. Inc. 1. Inc. I. For information. Inc. Inc. Printed in the United States of America. Title. 2. please contact The MathWorks. Inc.lam. Mathematics. Inc. Mathcad is a registered trademark of Mathsoft Engineering & Education. Matrix analysis for scientists and engineers / Alan J. Includes bibliographical references and index. Used by permission 5. 3 Apple Hill Drive. Inc.
Beverley To my wife.To my wife. Beverley (who captivated me in the UBC math library captivated UBC nearly forty years ago) nearly forty .
This page intentionally left blank This page intentionally left blank .
4 Some Useful and Interesting Inverses.1 The Fundamental Theorem 5. .2 Matrix Arithmetic . . .3 Row and Column Compressions Linear Equations Linear Equations 6. 6.3 Inner Products and Orthogonality 1. .2 Matrix Arithmetic 1. .2 Matrix Linear Equations 6. 6. ..2 Matrix Representation of Linear Transformations 3.. . .3 Composition of Transformations 3.2 Subspaces 2..1 The Fundamental Theorem .. . . .1 Definitions and Examples ..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. . .. . .2 Matrix Representation of Linear Transformations 3. .. 4.5 Four Fundamental Subspaces . . ..3 A More General Matrix Linear Equation 6.2 Examples 4. .4 Structure of Linear Transformations 3.1 4. 5..3 Inner Products and Orthogonality . .3 Properties and Applications Introduction to the Singular Value Decomposition Introduction to the Singular Value Decomposition 5.2 Matrix Linear Equations . . . .3 Rowand Column Compressions 5.1 Definitions and Characterizations Definitions and Characterizations. 4. . .4 Structure of Linear Transformations 3.. . .1 Definition and Examples 3.3 Linear Independence 2.1 2..1 Definition and Examples . .. 1.Contents Contents Preface Preface 1 1 Introduction and Review Introduction and Review 1. .1 Some Notation and Terminology 1.4 Determinants 1. .4 Determinants Vector Spaces Vector Spaces 2.2 Some Basic Properties 5. . .4 Sums and Intersections of Subspaces 2.. .1 Some Notation and Terminology 1.2 Examples.3 Properties and Applications . .. 3. 1. 5. . 3.1 Vector Linear Equations . .4 Sums and Intersections of Subspaces Linear Transformations Linear Transformations 3.3 A More General Matrix Linear Equation 6.2 Some Basic Properties . . 3. 4. .3 Composition of Transformations . 6. 2.3 Linear Independence . Definitions and Examples 2.2 Subspaces. . . .5 Four Fundamental Subspaces Introduction to the MoorePenrose Pseudoinverse Introduction to the MoorePenrose Pseudoinverse 4. 2. .. . .
2 Inner Product Spaces 7. . 8. 10. . . . . .5 Modal decompositions . . .4 Least Squares and Singular Value Decomposition 8. . .1 Projections . . .2. .3 Determination of the JCF . 9. . .2 Definite Matrices 10. . . .2 Inhomogeneous linear difference equations 11. .2 Jordan Canonical Form 9. . . 7. .2. . . .3 Equivalence Transformations and Congruence 10.5 Least Squares and QR Factorization 8. . .2. . . .5 Modal decompositions 11. . .1. . .2 On the +1's in JCF blocks 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. 9. . and Norms 7.2 Other least squares problems .1 Some Basic Canonical Forms 10.3.3 HigherOrder Equations .4 Geometric Aspects of the JCF 9.3. . .2 Difference Equations . . 9.1.2 Geometric Solution 8.4 Rational Canonical Form 11 Linear Differential and Difference Equations 11 Linear Differential and Difference Equations 11.1 Projections 7. . Eigenvalues and Eigenvectors 9.1 Differential Equations ILl Differential Equations .3 Linear Regression and Other Linear Least Squares Problems 8. .4 Least Squares and Singular Value Decomposition 8. .6 Computation of the matrix exponential 11.1. . . . .1.4 Linear matrix differential equations 11.1. .1 8.1 Fundamental Definitions and Properties 9. . .2 Homogeneous linear differential equations 11. 11.1 The Linear Least Squares Problem . .3 Computation of matrix powers 11. . . .1 Properties of the matrix exponential 11. . . 11. .1 Some Basic Canonical Forms . .3 Equivalence Transformations and Congruence 10. . .1 7. . .1 Example: Linear regression . . .2 Inhomogeneous linear difference equations 11. . . 8. . . . . 11. .1 Properties ofthe matrix exponential .3 Inhomogeneous linear differential equations 11. Inner Product Spaces. . 10.1 Homogeneous linear difference equations 11.2 Jordan Canonical Form . . . .6 Computation of the matrix exponential 11. . . . . .1 The Linear Least Squares Problem 8. . . . .2 Difference Equations 11.1.2 Homogeneous linear differential equations 11. Example: Linear regression 8.2 Inner Product Spaces 7. . . .3. . .3 Inhomogeneous linear differential equations 11. .1. . . .5 The Matrix Sign Function. . .2 Other least squares problems 8.1. .1 Homogeneous linear difference equations 11. 11.1 Fundamental Definitions and Properties 9. .4 Matrix Norms Linear Least Squares Problems 8. . .3 Linear Regression and Other Linear Least Squares Problems 8.3. . .2.2 Geometric Solution .1 Block matrices and definiteness 10. . .3 Vector Norms 7. . .3 Vector Norms 7.1 Theoretical computation . .3 Determination of the JCF 9.1 The four fundamental orthogonal projections The four fundamental orthogonal projections 7.2 Definite Matrices .1. .1. .5 Least Squares and QR Factorization .1.4 Rational Canonical Form . 7. . . .4 Geometric Aspects of the JCF 9.3. . 10. . .2.2.3 Computation of matrix powers . .3. . . . . . .1.1 9. . .2 On the + l's in JCF blocks 9. . . .3 HigherOrder Equations. . . . .viii viii Contents Contents 7 Projections. . . 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 . 11. . . . . .1 Block matrices and definiteness 10. .3. . . .3. . .4 Linear matrix differential equations . .1. .1. 11. 11. .4 Matrix Norms . . .3.3. . . Theoretical computation 9. 8. . . .
. . . .1 Conversion to firstorder form 12. . .4 Symmetric Generalized Eigenvalue Problems . . . .2 Properties of the Kronecker Product 13. .5 Simultaneous Diagonalization 12.Contents Contents ix ix 12 Generalized Eigenvalue Problems 12 Generalized Eigenvalue Problems 12.1 The Generalized Eigenvalue/Eigenvector Problem 12. 12.5 Simultaneous Diagonalization . .5. . .4 Symmetric Generalized Eigenvalue Problems 12. .3 Application to the Computation of System Zeros .2 Properties of the Kronecker Product . ..6 HigherOrder Eigenvalue Problems . . . . . 13. . . .3 Application to Sylvester and Lyapunov Equations 13. 12. 12.1 Definition and Examples . 13. . .1 The Generalized EigenvaluelEigenvector Problem 12. .2 Canonical Forms . . 12. .6 HigherOrder Eigenvalue Problems 12. . . .1 Simultaneous diagonalization via SVD 12. .6. . .2 Canonical Forms 12. . . .3 Application to the Computation of System Zeros 12.3 Application to Sylvester and Lyapunov Equations Bibliography Bibliography Index Index . .1 Simultaneous diagonalization via SVD 12. . 12. . .6. .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 13.5. .1 Definition and Examples 13.
This page intentionally left blank This page intentionally left blank .
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. Basic concepts such as determinants. The text linear dynamical systems (systems of linear differential or difference equations). The books by Meyer [18]. 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). I highly recommend MATLAB® although other software such as xi xi . particular subject area. [13]. the sciences. and concepts such as determinants. followon topics on the computational side (at the level of [7].e. By matrix analysis I mean linear tools and ideas comfortably in a variety of applications. 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. Certain topics thoroughly as undergraduates. eigenvalues and eigenvectors. singularity of matrices. the student is then wellequipped to pursue. Matrices are stressed more than abstract vector spaces. for [11]. although Chapters 2 and 3 algebra. but somehow didn't quite manage to do. For this. example) or on the theoretical side (at the level of [12]. basisfree or subspace) aspects of many of the fundamental notions. and Strang [24] Ortega are excellent companion texts for this book.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. Noble and Daniel [20]. the sciences. mathematics. For this. Matrices are stressed more than abstract vector spaces. 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. students meant to learn thoroughly as undergraduates. in many cases. I have tried throughout to emphasize only the and more advanced material is introduced. example) or on the theoretical side (at the level of [12]. although Chapters 2 and 3 do cover some geometric (i. 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. These powerful and versatile tools doinverses and the singular value decomposition (SVD). 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. 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. and positive definite matrices should have been covered at least once. for example). basisfree or subspace) aspects of many of the fundamental do cover some geometric (i. The concept of matrix factorization applications and by computational utility and relevance. I have tried throughout to emphasize only the more important and "useful" tools. singularity of matrices. 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." However. or [16]. the student is then wellequipped to pursue. but somehow didn't quite manage to do. for example). or [16]. eigenvalues and eigenvectors. methods. and mathematical structures. Upon completion of a course based on this text." this approach necessarily presupposes the availability of appropriate mathematical software on approach necessarily presupposes the availability of appropriate mathematical software on a digital computer. [13]. computer science. Because tools such as the SVD are not generally amenable to "hand computation. mathematics. Basic of calculus and definitely some previous exposure to matrices and linear algebra. or computational students in engineering.. [II]. Because tools such as the SVD are not generally amenable to "hand computation. even though their recollecmatrices least tion may occasionally be "hazy." However. either via formal courses or through selfstudy. either via formal courses or through selftext.. 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). in many cases.e. computer science. I highly recommend MAlLAB® although other software such as a digital computer." this ics. Upon completion of a course based on this are excellent companion texts for this book. Ortega [21]. [23]. requiring such material as prerequisite permits tion may occasionally be "hazy. 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. students meant to learn much of the important and useful mathematics that. or [25].
If If are linearly dependent. many times at UCSB and twice at UC Davis. Indeed. and thus the text can serve a rather diverse audience. outputs. The presentation of the material in this book is strongly influenced by computais influenced by computational issues for two principal reasons. But in most engineering or scientific contexts we want to know more than that. and states often give rise to models of very numbers models high order that must be analyzed. are deferred to such a course. for example. It is my firm conviction that such maturity is neither encouraged conviction neither nor nurtured by relegating the mathematical aspects of applications (for example. they are either obvious or easily found in the literature." a set of vectors is either linearly independent or it is not. the details of most of the numerical aspects of linear algebra per se. I have taught this material for many years. form the foundation virtually modem upon which rests virtually all of modern scientific and engineering computation. control systems with standard large numbers of interacting inputs. 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. This is ideal material from which to learn a bit about mathematical proofs and the mathematical maturity and insight gained thereby. in an elementary linear algebra course. many students who completed especially offered. completed the course. they are either obvious or easily found in the literature. Mastery of the material in this text should enable the student to read and understand the modern language of matrices used throughout mathematics. 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. especially the first few times it was offered. diverse audience. econometrics. and while most material is developed from basic ideas in the book." For example. in particular. prerequisites developed While prerequisites for this text are modest. Rather. form the foundation Some of the key algorithms of numerical linear algebra. and a wide variety of other fields. Since this text is not intended for a course in numerical linear algebra per se. finiterange floatingpoint arithmetic environment of of of most modem computing platforms." Proofs are given for many theorems. mathematics. simulated. and the course has proven to be remarkably successful at enabling students from Davis. how "nearly dependent" are the vectors? If they linearly independent. A second motivation for a computational emphasis is that it provides many of the essential tools for what I call "qualitative mathematics. chemistry. Statespace methods are Statespace modem now standard in much of modern engineering where. The "language" in which such described models are conveniently described involves vectors and matrices. Some of the key algorithms of numerical linear algebra. First. the student does require a certain amount of what is conventionally referred Proofs referred to as "mathematical maturity. in particular. and coherent fashion. 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. They must generally be solved computationally and closedform it is important to know which types of algorithms can be relied upon and which cannot. science. If a set of vectors is linearly independent. When they are not given explicitly. statistics. . This is an absolutely fundamental fundamental concept. linear algebra introducing "onthefly" algebra for elementary statespace theory) to an appendix or introducing it "onthef1y" when to necessary. This is ideal not given explicitly.xii xii Preface Preface Mathcad® Mathematica® or Mathcad® is also excellent. The tools of matrix analysis are also applied on a daily basis to problems in biology. if only they had had this course before they took linear systems. and evaluated. consistent. are deferred to such a course. remarked afterward that if processing. applied physics. or signal processing. and modem engineering. "reallife" problems seldom yield to simple "reallife" closedform formulas or solutions. must lay firm foundation upon which and perspectives perspectives can be built in a logical. one must lay a firm foundation upon which subsequent applications and Rather. It is thus crucial to acquire knowledge vocabulary a working knowledge of the vocabulary and grammar of this language.
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. June 2004 — AJL.Preface Preface xiii XIII or estimation theory. too.. AJL. My fellow instructors. The concept seems to work. realized that by requiring this course as a prerequisite. etc. 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 .
and linear algebra. e. x e IR n means means where xi e IR for e n.. 3.. 2. 2.n xn Cmxn = the set of complex m x n matrices of rank r. Rnxnn denotes the set of real nonsingular n x n matrices.. Thus. the notation!! denotes the set {I. the notation n denotes the set {1. Cn = the set of ntuples of complex numbers represented as column vectors. A row vector is denoted by yT where Note: Vectors are always column vectors. 1 . xyT is an n x n matrix. x T y is a scalar while xyT is an n x n matrix. mxn = the set of complex (or complexvalued) x n matrices. IR rn xn = the set of real (or realvalued) m x n matrices. Henceforth. . but this convention makes it easy to recognize immediately throughout the text that. IR~ xn denotes the set of real = set of real of rank Thus.g..g. That a vector is always a column vector rather than a row vector is entirely arbitrary. e. . nonsingular n x n matrices. but this convention makes column vector rather than a row vector is entirely arbitrary. XTy is a scalar while it easy to recognize immediately throughout the text that.Chapter 1 Chapter 1 Introduction and Review Introduction and Review 1. IR n = the set of ntuples of real numbers represented as column vectors. 5. This is followed by a review of some basic notions in matrix analysis throughout the text. Henceforth. n}. This is followed by a review of some basic notions in matrix analysis and linear algebra. e. = the set of complex m x n matrices of rank r. where Xi E R for ii E !!. 1R. Crnxn = the set of complex (or complexvalued) m x n matrices.. e 6. That a vector is always a y E IR n and the superscript T is the transpose operation. ..1 1.1 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.. R mxn = the set of real (or realvalued) m x n matrices. x E Rn I. Thus.n xn Rmxnr = the set of real m x n matrices of rank r. A row vector is denoted by y~. Rn = the set of ntuples of real numbers represented as column vectors. where y G Rn and the superscript T is the transpose operation. en 4. n }. the set of ntuples of complex numbers represented as column vectors. The following sets appear frequently throughout subsequent chapters: The following sets appear frequently throughout subsequent chapters: 1. Note: Vectors are always column vectors. 5. Thus.
• lower Hessenberg if aij = 0 for } . A e jRmxn. There is some advantage to being conversant with both notations. ] is symmetric (and Hermitian). an equation like A = A T implies that A is realvalued while a statement like A = AH implies that A is complexvalued. For example. Introduction and Review We now classify some of the more familiar "shaped" matrices. C E jRmxm. then its Hermitian transpose (or conjugate transpose) is denoted by AH (or H If A e C mx ". For example. We henceforth adopt the convention that. Example 1. • pentadiagonal if aij = 0 for Ii . if z = a + jf$ (j = ii = R). There is some the more common notation in electrical engineering and system theory. is complexvalued symmetric but not Hermitian. A = [ . • upper triangular if aij. B E IR nxm . The notation j is used throughout the text but reminders are placed at strategic locations.[ 7 . For example. it is Transposes of block matrices can be defined in an obvious way. Example 1. A matrix A is symmetric i. text but reminders are placed at strategic locations. an equation like A = A T implies that A is realvalued while a statement otherwise noted. (7.j  7+} ] is Hermitian (but not symmetric). = 0 for j — > 1. Hermitian conjugate sometimes A*) and its (i. where the bar indicates complex j)th entry is (A H ).ii > 1. • lower triangular if a.. = 0 for i > j.e. is Hermitian (but not symmetric). and definitions block submatrices. are appropriately dimensioned subblocks. If A E em xn. The The transpose of a matrix A is denoted by AT and is the matrix whose (i. We henceforth that.J I > 2. Each of the above also has a "block" analogue obtained by replacing scalar components in the respective definitions by block submatrices. 2 2. ~ 5 is symmetric (and Hermitian). • upper triangular if a. if e Rnxn e Rmxn C e Rmxm then the (m n) x (m n) matrix [A0 ~] is block upper triangular. • upper Hessenberg if afj = 0 for — > 1. 1. • tridiagonal if aij = 0 for Ii . where the bar indicates complex sometimes A*) and its = IX jfJ (j = = v^T). i.JI > 1. • diagonal if aij7 = 0 for i i= }. j)\h entry is (AH)ij = (aji).. AT E E" xm is the (j.. A = [ 7+} 5 3· A . While \/—\ is most commonly denoted by i in mathematics texts. For example. • pentadiagonal if ai.e. j is Remark the more common notation in electrical engineering and system theory. (AT)ij = aji. Oth (A 7 ).. . • lower triangular if aij7 = 0 for i/ < }. that is. i)th entry of A. 7 = («77). then r = [ . if A E IRnxn. 7 + j ] is complexvalued symmetric but not Hermitian.. = 0 for i ^ j..2 2 Chapter 1..2. then the (m + n) x (m + n) matrix [~ Bc] is block upper triangular. a. = 0 for < j. } is While R is most commonly denoted by i in mathematics texts.jj > 1. A matrix A E IRn xn e (or A E enxn ) is A eC" x ")is • diagonal if a. = a jfJ. then A7" e jRnxm. A if A = AT and Hermitian if A = AH. then z = IX — jfi. 2 Transposes of block matrices can be defined in an obvious way. • tridiagonal if a(y = 0 for z — j\ > 1. • lower Hessenberg if a. then easy to see that if A. A = AH A complexvalued.. it is easy to see that if Aij are appropriately dimensioned subblocks. • upper Hessenberg if aij = 0 for ii .. = 0 for / — j\ > 2.. j)th entry of a matrix A is denoted by AT and is the matrix whose j)th entry A. unless if A = A T Hermitian A = A H. otherwise noted.1. = 0 for i > }. A is conjugation. Introduction and Review Chapter 1. Note that if A E R mx ". z Remark 1.2.
. importance interpretation take A = [96 85 74]x = take A = [~ ~].. . 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.2 Matrix Arithmetic 1. Theorem reader. Then the matrix product A B can be thought of as above.. That is. Theorem 1. + Xnan E jRm.3 can then also be generalized to its "row dual. if (C D)H — D C ). suppose A e Rmxn and B = [bi. . suppose A E jRmxn and [hI.. As a numerical example. Then we can quickly calculate dot products of the rows of A column Ax = [..xn~ ] Then Ax = Xjal + .1... ..'" p] E jRnxp For matrix multiplication. and is premultiplied by a row vector yTe jRlxm. vector x.. AB bi E W1.[ ~ J+2. . i. multiplication of a matrix by a scalar. The importance of this interpretation cannot be overemphasized..e..•.• a"1 E m JR " with a. recall that (CD)T = DT C T (C D)T = DT T If H H H (or (CD} = DHC H ).3. un]]Ee jRmxn with Ui t Ee jRm and V = [VI.. formulation of matrix multiplication that appears frequently in the text and is presented below as a theorem. there can be important computerarchitecturerelated advancomputerarchitecturerelated tages to preferring the latter calculation method. Ax. un Rmxn with u Rm and V = [v ... the matrixvector product Ax.. Then v E jRP.~]. A special case of matrix multiplication occurs when the second matrix is a column multiplication second i. E JRm and x = l I. Again. Matrix Arithmetic 3 1.[ ~ J+l. 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. Theorem 1.2 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.e..3. its importance cannot be overemphasized. vn Rpxn p with Vit e R . Let U = [MI. {. This gives a dual to the matrixvector result above." The details are left to the readei "row left . but equivalent.. applied p times: There is also an alternative.. suppose (linear combination) suppose A = la' .. and multiplication of matrices.. multiplication. Vn] ]Ee lR Pxn U [Uj. eRmxn has row cj e E l x ".. i=I If matrices C and D are compatible for multiplication. matrixvector if C E jRmxn has row vectors cJ E jRlxn.bhp ] e Rnxp with hi e jRn. x = ! 2 Then we can quickly calculate dot products of the rows of A [~]. It Theorem 1. Namely. n UV T = LUiVr E jRmxp. .2..[ ~ l For large arrays of numbers.
y. Then Example 1. we define their complex Euclidean inner product (or inner product.. order in which x product is important. Y}c = [ } JH [ ~ ] = [I . Note that the inner product is a scalar. y)c = y = Eni=1 xiyi but throughout the text we prefer the symmetry with the real case..e. If x and y are zero. a matrix A e en xn is said to be unitary if A H A = AA H = I. the Euclidean inner product inner for short) y is given by y is given by n T (x. i. Clearly said = an orthogonal or unitary matrix has orthonormal rows and orthonormal columns.4 4 Chapter 1. (x. 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. (x.j] [ ~ ] = 1 . indeed. Nonzero complex vectors are orthogonal if XHy = O. To illustrate. . x)c'. y) := x y = Lx. Let x = [} ]] and y = [~]. the order in which x and y appear in the complex inner (x. A EC = (orC" xn). The notation /„ is sometimes used to denote the identity matrix in IR nxn in Rnx" x nxn H H (or en "). A orthogonal and XTX = 1 and yTyy = 1. .3 1. y e IRn are said to be orthogonal if their inner product is zero.e. x)c. Similarly. Then x T x = 0 but x H X = 2. the nonzero vector x above. for short) of x and For vectors y e IRn. then we say that x and y are orthonormal. In sometimes used denote identity matrix. We list below some of (or A 6 en xn) we use the notation det A for the determinant of A. Y}c = {y. Let x = [1j and y = [1/2]. x T = O. Then (x. x)c and we see that. The more conventional definition of the complex inner product is product is important. Then XTX = 0 but XHX = 2. consider the nonzero vector x above. Similarly. There is an orthogonal or unitary matrix has orthonormal rows and orthonormal columns. where / is the n x n identity matrix. y E <en. 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.4 Determinants Determinants It is assumed that the reader is familiar with the basic theory of determinants. where I is the n x n matrix A e IRnxn is an orthogonal matrix if ATA = AAT = /.e. Nonzero complex vectors are orthogonal if x H y = 0. then we say that x and y are orthonormal. for short) by for short) by n (x'Y}c :=xHy = Lx. case.y. but throughout the text we prefer the symmetry with the real (x.. To illustrate. y)c = (y. Introduction and Review 1. Introduction and Review Chapter 1. For A E R nnxn A e IR xn It assumed of determinants. If x.y. i. i. We list below some of . i. y ) c = yHxx = L:7=1 x. 1.=1 y appear in Note that (x. y E R". What is true in the complex case is that XH x = 0 if and only if x = 0. 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. x and y are orthogonal and x Tx = 1 and yT = 1.=1 Note that the inner product is a scalar. Two nonzero vectors x. xTyy = 0.4. y)c = (y. If e C".2j while while and we see that. the Euclidean inner product (or inner product. indeed.4.. The more conventional definition of the complex inner product is H ( x . Example 1. A nxn matrix E R is an orthogonal matrix if AT A = AAT = I. rows or columns. y)c = (y. we define their complex Euclidean inner product (or inner product..3 Inner Products and Orthogonality Inner Products and Orthogonality For vectors x.4 1. consider What is true in the complex case is that x H 0 if and only if O..e. (or A E Cnxn) we use the notation det A for the determinant of A.
then det A = O.. 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.• det Ann. If det = a11a22 • • ann 12. If A e R n x n and D e R m x m .. If elements. then det [Ac ~] det D det(A B D. Determinants 5 properties the more useful properties of determinants.C).4. Multiplying A 6. Multiplying a row of A by a scalar and then adding it to another row does not change the determinant. det AT = det A (detA H = detA if A e C nxn ). Proof: This follows easily from the block LU factorization Proof" This follows easily from the block LU factorization [~ ~J=[ ~ ][ ~ 17. If A € lR~xn. 13. 3. 15. Ann (of possibly different sizes). Interchanging two rows of A changes only the sign of the determinant. Interchanging two columns of A changes only the sign of the determinant. 3.A22. then det(A. Multiplying a row of A by a scalar a results in a new matrix whose determinant is 5. A 11. 16. Multiplying a column of A by a scalar ex results in a new matrix whose determinant scalar a determinant is ex det A. A 22 . i.. Note that this is not a minimal set.. 11.B). is a det A. then det A = different det A11 det A22 . If A has a zero column or if any two columns of A are equal. If A has a zero column or if any two columns of A are equal.e. detAT = detA (det A H = det A A E C"X").1 I ][ . • • An" (of A = square diagonal blocks A11. 15. If A has a zero row or if any two rows of A are equal.• ann. If A is upper triangular. If A e IRnxn and D E lR~xm. of 5. Multiplying a row of A by a scalar and then adding it to another row does not change 7. Multiplying a row of A by a scalar ex results in a new matrix whose determinant is a det A. 7.. are consequences of one or more of the others.. 8. 9. = alla22 • • ann i. 17. det A is the product of its diagonal diagonal. . then det A = a11a22 . If A is lower triangUlar.• ann.. If A is diagonal. If A E R n x n and D e RMmxm. several more is a properties are consequences of one or more of the others. i. B E IRnxn ..1. properties 1. 10. then det A = a11a22 .e.. 11... then det A = all a22 . 14. then det A = 0. If A. the determinant. then det A = 0. If A E Rnxn. then det [~ BD] = del A det(D – CA– l 1 B). with A block diagonal (or block 13.CA..4. exdetA. 16. Interchanging two rows of A changes only the sign of the determinant. B eR n x n .1 ) = de: A.• a"n.•... If A E lR~xn and DE IR mxm det [Ac ~] detA det(D . Proof" Proof: This follows easily from the block UL factorization BD. det A 11 det A22 • • det Ann 14.. If A A A = o... Determinants 1..thendet(AB) = det A det 5. det A is the product of its diagonal 10. then det [~ BD] = det D det(A – B D – 11C ) . If A. 8.e. then det(A1) = 1detA . then det(AB) = det A det B. If A is block diagonal (or block upper triangular or block lower triangular). 4. 2. then det A = alla22 • • ann 12.. If A is lower triangular. change the determinant.
Vk € jRn xn be orthogonal matrices. [24]. A E jRnxn A2 / x™ . (b) Suppose A e IR" X "is idempotent and A i= I.. The matrix D .y E jRn. Show that A must be singular.6 6 Chapter 1. 2 0 IS I dempotent .. AB ^ BA. Tr(aA + f3B) = aTrA + fiTrB. [24].. If A e jRnxn and or is a scalar. what is det(aA)? What is det(–A)? E R a det(A)? A? If A unitary. . Show that det(I – xyT) = 1. A? 2.e. (c) Let S € Rnxn be skewsymmetric. i. TrA = L~=I au· elements. then Tr(aA fiB)= aTrA + f3TrB. denoted TrA. . i. Letx.. example.. The factorizations used above U triangular. 2sin20 J is idempotent for all #. are block analogues of these.. of Din [AC ~ l EXERCISES EXERCISES nxn 1. is defined as the sum of its diagonal A e Rnxn. what is det A? If A 3. even though in general AB i= B A. Showthatdet(lxyT) 1 – yTx. Uk E Rnxn U = VI V2 . ! [ 2cos2<9 I T 2cos2 0 (a) Show that the matrix A = _. [~ ~ ]. for example. The trace of A. of denoted Tr A. either prove the converse or provide a counterexample.• V k is an orthogonal matrix.. (b) Show that Tr(AB) = Tr(BA). Show that the product V = VI. • . i. U2 . Remark 1. see. B e JRn xn and a. Introduction and Review Chapter 1.e. Then E jRnxn skewsymmetric. if A. elements.e. i. TrS O. aII o. ft e R. Let U1.B D – l C is the Schur complement of D in [~ BD ]. II _ . . A =.. lor z r 2sm2rt # J. U1 U2 • • Uk is an 5. Suppose A E jRn xn is idempotent and A ^ I. 4. Introduction and Review Remark 1. Another such factorization is UL where V is unit upper triangular and L is lower triangular. Remark — C I B – BDIe Similarly. lower triangular with all l's on the diagonal) and an upper triangular matrix L 1's an V is called an LV factorization. The factorization of a matrix A into the product of a unit lower triangular Remark 1. Let A E jRNxn...e A – 1 B is called the Schur complement of A in[ACBD].yTx. _. ST = So Show that TrS = 0. nxn linear E R f3 E JR. Tr(Afl) = Tr(£A).5. If A is orthogonal. what is det A? If A is unitary. A . y e Rn.e.. The factorization of a matrix A into the product of a unit lower triangular matrix L (i. ST = S. ..5.. A matrix A e Wx" is said to be idempotent if A2 = A. .e. .. Another such factorization is VL U is an LU factorization. (a) Show that the trace is a linear function. 2f) 2 _ sm 2^ sin 0 sin sin 20 1 . V2. 6. Let x.6. see. TrA = Eni=1 aii.. Show that A must be singular.
(M3) e IF. (Ml) a . including spaces formed by special classes of matrices. p. afar all a. (A4) a + . 0.8 E IF.8 a for all a. An excellent reference of matrices.Chapter 2 Vector Spaces Vector Spaces In this chapter we give a brief review of some of the basic concepts of vector spaces. y Elf. aI = 1. when no confusion can arise.8) + y ffor all a. (A3) for all a e F. y € F.((. 7 . . 2. a f.8 + y) = (a +." is not written explicitly. . y Elf. • F x IF ~ F such that (Al) a (P y ) = (a + p ) y o r all a. . ^ 0. where some of the proofs that are not given here may for this and the next chapter is [10]. (M2) 1 e IF • I = for a e IF. . there exists an element (a) e F such that a + (a) = 0. The In this chapter we give a brief review of some of the basic concepts of vector spaces.8.8 . (Ml) a· p . including spaces formed by special classes emphasis is on finitedimensional vector spaces. (A2) there exists an element 0 E F such that a + 0 = a for all a E F. •) is an abelian group. p.8 e F. (A4) a + p = . there exists an element (—a) E IF such that a (—a) O.8)· yyf for all a.o r all a.8 = ft + afar all a. (M4) a • p =.8.8. y)=cip+a.ye¥. a"1 € IF • a~l = 1.((.8 + y) = a·. A field is a set F together with two operations +.1.8. for all a e IF. (A3) for all a E IF.p ) . (Al) a + (.8 +a· y for all a. (D) (D) a· p a . : IF x F —> IF such that Definition 2. A field is a set IF together with two operations +. p e F. the multiplication operator ".8 = P • a for all a. (M2) there exists an element I E F such that a . but some infinitedimensional examples are also cited. ft. +) is a group and an abelian group if (A4) also holds. (M3) for all a E ¥. yy) = (a·.) ( a . Axioms (A1)(A3) state that (F. Axioms (Al)(A3) state that (IF. where some of the proofs that are not given here may be found. Axioms (M1)(M4) state that (F \ to).. Generally speaking. The emphasis is on finitedimensional vector spaces. but some infinitedimensional examples are also cited. when no confusion can arise.1 Definitions and Examples Definition 2. not written explicitly. . Axioms (MI)(M4) state that (IF \ {0}. I = a for all a E F. there exists an element aI E F such that a .1. y Elf.y for alia. y e F. An excellent reference for this and the next chapter is [10].) is an abelian group. (A2) there exists an element 0 e IF such that a 0 = a.. be found. (M4) a·. the multiplication operator "•" is Generally speaking. +) is a group and an abelian group if (A4) also holds. ft Elf.
this causes 2. C with ordinary complex addition and multiplication is a field. (V5) 1 v = v for all v e V (1 Elf). is a vector space. In practice. 1. I.. where Z+ = {O. Remark 2.5. RMrmxn= {m x n matrices of rank r with real coefficients} is clearly not a field since... Example 2. is a field.1 in the sense of operating on different objects in different sets. 2. + apxP + .• v = a·• v + p • v for all a.p ) . when there is no possibility of confusion as to the underlying fie Id.3. (V2) ( a f3) v = a P . ft) v = a v + f3. (V5) I·• v = v for all v E V (1 e F). w E V.2. . Note that + and· in Definition 2. simply by V.( (f3' V v) o r all a.r] = the field of rational functions in the indeterminate x = {ao + f30 + atX f3t X + . 4. (Ml) does not hold unless m = n. Similar definitions hold for (en.f3i EIR .F xV »• V such that (VI) (V. f3 e F and for all v E V. 3. (MI) does not hold unless m = n.. Definition 2. Ra[x] = the field of rational functions in the indeterminate x 3. F) or. Similar definitions hold for (C". v = a . R"x" is not a field either for example. no confusion and the operator is usually not even written explicitly. . Vector Spaces Example 2. Raf. 4. (V4) a· (v + w) = a . w e V. since (M4) does not hold in general (although the other 8 axioms hold).4.. +) is an abelian group.. simply by V. v for all a. (V3) (a + f3).3 are different from the + and . in Definition Remark 2. }. ) f for all a. C). (VI) (V. (R".}.1 in the sense of operating on different objects in different sets. Moreover. underlying field.1. . 1. for example. (IRn. this causes no confusion and the·• operator is usually not even written explicitly. Moreover.. p € F and for all v e V.2. v + a.8 Chapter 2.P.4. lR~xn is not a field either since (M4) does not hold in general (although the other 8 axioms hold).2. e). f3 Elf andforall v E V. A vector space is denoted by (V. R) with addition defined by I. IF) or.l. R with ordinary addition and multiplication is a field. Example 2.3 are different from the + and • in Definition 2. IR~ xn = m x n matrices of rank r with real coefficients) is clearly not a field since.2.. where Z+ = {0. when there is no possibility of confusion as to the A vector space is denoted by (V.. p E IF andfor all v e V. IR with ordinary addition and multiplication is a field. A vector space over a field F is a set V together with two operations Definition 2. e with ordinary complex addition and multiplication is a field. Example 2. (V2) (a·.3. is a field. + f3qXq :aj. Note that + and • in Definition 2. IR) with addition defined by and scalar multiplication defined by and scalar multiplication defined by is a vector space. In practice.5. 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. (V3) (a (V4) a(v w)=av a w for all a ElF andfor all v. +) is an abelian group.qEZ +} . w for all a e F and for all v. 2.
Then (W. Let (V. Note. The latter characterization of a subspace is often the easiest way to check or prove that something is indeed a subspace (or vector space). Subspaces 2. + fJmn l yaml yamn 3. Then (W. and for all f E cf>. t\]. this question is closed under addition and scalar multiplication. IF) = (JR n . Let A € R"x". F) = (IR". and the functions are piecewise continuous (a) '0 = [to. fJ e IF andforall WI. if and only if(aw1 ßW2) E if(awl + fJw2) e W for all a. l . (V. too. Notation: When the underlying field is understood. Let O(X>. E) is a vector space with addition defined by 9 9 A+B= [ . JR). Then (x(t) : x ( t ) = Ax(t)} is a vector space (of dimension n). (E mxn JR) is a vector space with addition defined by 2. foral! a. F) be a vector space and let W ~ V. Let (V. etc.2. Then {x(t) : x(t) = Ax(t}} is a vector space (of dimension n). W = 0. E). this implies that the zero vector must be in any subspace. Notation: When the underlying field is understood. Note. t\])n continuous =: (C[to.2. (V. Let A E JR(nxn. 2. less restrictive meaning "is a subset of' is specifically flagged as such. verify that the set in or prove that something is indeed a subspace (or vector space). Let cf>('O. Special Cases: Special Cases: (a) V = [to. that since 0 e F.2 2. td. F) be an arbitrary vector space and V be an arbitrary set. h])n (b) '0 = [to. and the symbol c. g E cf> and scalar multiplication defined by and scalar multiplication defined by (af)(d) = af(d) for all a E IF. Then O(D. V) is a vector space with addition set of functions f mapping '0 to V.7.. F) is itself a vector space or.6. +00). and the functions are piecewise continuous =: (PC[to. td)n or continuous =: (C[?0. implies that the zero vector must be in any subspace.. we write W ~ V. (V." The when used with vector spaces. we write W c V.2 Subspaces Subspaces Definition 2." + fJ2I a21 + P" . td)n. F) is a Definition 2. too.2. W2 E W.. IF) is a subspace of (V. if and only subspace of (V. =: (PC[f0. IF) if and only if (W. is henceforth understood to mean "is a subspace of. The latter characterization of a subspace is often the easiest way to check Remark 2. (JRmxn. Subspaces 2. IF) be an arbitrary vector space and '0 be an arbitrary set.6. i. V) be the set of functions / mapping D to V. Then cf>('O. verify that the set in question is closed under addition and scalar multiplication. 4. . is henceforth understood to mean "is a subspace of..e. 4. y a l2 y a 22 yam 2 ya. equivalently. IF) is itself a vector space or. for all d ED. Let (V. when used with vector spaces. W f= 0. and the symbol ~. 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 = .7." ya2n ." The less restrictive meaning "is a subset of" is specifically flagged as such.e. ß E ¥ and for all w1. IF) = (JRn. F) if and only if (W. 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. JR). Let (V. w2 e Remark 2. that since 0 E IF. equivalently. V) be the 3. i. IF) be a vector space and let W c V.
Note. •••. . a = oo) is also a subspace. Then it is easily shown that aAI + fiAi is symmetric for all a. R) and for each v € R2 of the form v = [v1v2 ] identify v1 with 3. . Thus. Vk e X and scalars aI. IF) = (]R2. Vk of X and for any scalars aI. . Henceforth. too. a = 00) is also a subspace.. For ß E R define the jccoordinate in the plane and V2 with the ycoordinate. 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. As an interesting exercise.lF) = (R" X ". elements VI.. that the vertical line through the origin (i.. .e.10 Chapter 2. 1. .10./l is a subspace of V if and only if f3 = 0. A2 are symmetric. Then W is /wf a subspace of JR. W2. ak. . al VI + . define W". ak.I' and Wi.. Shifted subspaces W". in some vector space V. ak = O. Henceforth. sketch Then Wa. .Vk of X and for any scalars a1. F) = (R2./l with f3 = 0 are All lines through the origin are subspaces.. To prove two vector spaces are equal. too. F) = (JR..9.nxn..JR. ffR and S are vector spaces (or subspaces)./l = {V : v = [ ac ~ f3 ] . Then it is easily shown that ctA\ + f3A2 is Proof' Suppose AI. Consider (V. . . As an interesting exercise. .e. Note. If 12.8.} . V2. . Definition 2.R) and 1. ft E R symmetric for all a. X linearly set of Definition 2. .nxn : A We V. Example 2. and S are vector spaces (or subspaces). we drop the explicit dependence of a vector space on an underlying field. Let W = {A € R"x" : A is orthogonal}. + (XkVk = 0 implies al = 0. then R = S if and only if Definition 2.I. f3 e R.10.. Then (V.o..3 2. . JR. then R = S if and only if R C S and S C R. . .. 3. Shifted subspaces Wa.. c E JR. . explicitly stated otherwise. ak not all zero such that elements VI. • • •} be a nonempty collection of vectors u.ß with ß =1= 0 are called linear varieties. A2 are symmetric.1. v2. Definition 2.. = {A E JR. •. Then W". called linear varieties.O.ß is a subspace of V if and only if ß = O.0. unless explicitly stated otherwise.9. f3 e R. ~SandS ~ R. be an element of R.and W1/2. X is a linearly dependent set of vectors ifand only if there exist k distinct if and only if exist distinct elements v1.) and for each v E ]R2 of the form v = [~~ ] identify VI with the xcoordinate in the plane and u2 with the ycoordinate. W2. (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. . that the vertical line through the origin (i. V usually denotes a vector space with the underlying field generally being R unless Thus. .1. Proof: Suppose A\. W1/2. vk E X and scalars a1.. E ]Rnxn : not 2. For a. 2. W~V.) and let W = [A e R"x" : A is symmetric}.•..• } be a nonempty collection of vectors Vi in some vector space V. Let X {VI. we drop the explicit dependence of a vector space on an underlying field.. Consider (V. one usually proves the two inclusions separately: Note: To prove two vector spaces are equal. R"x"..3 Linear Independence Linear Independence Let X = {v1. .S.nxn. Wi. Vector Spaces Example 2. Consider (V. sketch W2. V usually denotes a vector space with the underlying field generally being JR. All lines through the origin are subspaces. W2.o.
. . {1.en} = ]Rn. If the set of vectors is independent. Example 2. . e2 ..14.•}} be a collection of vectors vi. o Definition 2. T V is nonsingular. . = {v : where N = {I.. t1] (recall that etA denotes the matrix exponential.. which is discussed in more detail in efA Chapter 11)..11. Let X = {VI. Let V = Rn and define = ]Rn and el = 0 0 . en} = Rn..'" . X = [v1 v2 . Then the span of of X is defined as X is defined as Sp(X) = Sp{VI. E V. ~ HHi] } Ime~ly i is a i" linearly independent set.14. and there exists a E ]Rk such that VT V is singular. If the set of vectors is independent. Then consider the rows of etA B as vectors in Cm [t0. Sp(X) = V. Why? independent..en = 0 0 0 o SpIel.3. 2. Why? However.. }. 2. Linear Independence Example 2. . 2. Then {[ Then I. Then Sp{e1.13. to be studied further in what follows. then = O. Vi EX. = [ v 1 . Independence of these vectors turns out to be equivalent to a concept called controllability. (since 2vI .2. e k. 1. . + (XkVk ."I [ i1i1l ]} [[ s a linearly is a Iin=ly dependent set de~ndent ~t (since 2v\ — V2 + v3 = 0). LetV = 11 11 ~..13. Vi e span of Definition 2.12. ... . to be studied further in what follows.. and consider the matrix V = [VI. 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. Example 2. V V2.. A set of vectors X is a basis for V if and only if 1. Vk] E Rnxk. (Xi ElF.. tIl 2. An equivalent condition for linear dependence is that the k x k matrix condition VT V is singular.11... ii E If. e2.. Definition 2. An equivalent condition for linear independence is that the matrix V TV is nonsingular. . and X (of and 2.3. .. Sp(X) = V. Then consider the rows of etA B as vectors in em [to. ..12.v2 + V3 = 0). 1£t V = R3.. then a = 0.. Let A E ]Rnxn and 5 e R"xm. V2. A e R xn B E ]Rnxm. Independence of these vectors turns out to be equivalent to a concept Chapter 11). kEN}.Vk] e ]Rnxk. and there exists a e R* such that Va = 0. An equivalent condition for linear independence is that the matrix Va = 0.. linear dependence x such that Va = 0. Linear Independence 2.. . X is a linearly independent set (of basis vectors)... e2 = 0 1 0 . A set of vectors X is a basis for V if and only ij Definition 2. Howe. } = (Xl VI + . The linear v E ]Rn. called consider Let Vif e R".}.
for]Rn [e\.16.. .dimensional or have dimension n and we write dim (V) = n or dim V — n. Example 2. We represents B..17. . In Rn..E~n} such that v= where ~Ibl + . ] l = = Theorem 2..b. n } such that for V..[ ~  ] + 4· [ ~ l To see this.. The scalars {Ei}are called the components (or sometimes the coordinates) components coordinates) Definition 2. .. .. el + 2 .. In]Rn.16.. . .. . VI ] : = vlel + V2e2 + . We say that the vector x of of of (b1...18. The number of elements in a basis of a vector space is independent of the Theorem 2. components represents the vector v with respect to the basis B.19. r I [ .. en} is a basis for IR" (sometimes called the natural basis). For example. + ~nbn = Bx. while We can also determine components of v with respect to another basis.12 12 Chapter 2. Definition 2. with respect to the basis with respect to the basis {[~l[!J} we have we have [ ~ ~ ] = 3. write [ ] = XI • [ ~ + ] X2 • [ _! ] =[ ~ = [ ~ Then Then ! ][ ~~ l 1 [ ~~ ] = [ . . n for Then for all e there exists a unique ntuple {E1 ... B ~ [b".19. . bn be a basis (with a specific order associated with the basis vectors) b1. bn]} and are unique. For example. particular basis considered...15. Vector Spaces Example 2. x ~ D J Definition 2. . If V= 0) V is Definition 2. For .... {~i } of v with respect to the basis {b l .. + vne n · Vn We can also determine components of v with respect to another basis. e2.l.18. {el.. . en} natural Now let b l . For be n. V is said to X for be ndimensional or have dimension n and we write dim(V) n or dim V n. . If a basis X for a vector space V(Jf 0) has n elements. n unique. 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'.. while [ ~ ] = I .
+ 7^ =: L R. S S. n 5 S. Theorem 2.22. The sum and intersection Definition 2. otherwise.21. U + S = T (in general ft. is not necessarily a subspace.4. Sums and Intersections of Subspaces 2. j e n. 2. and 1.20. j E ~.23.4 Sums and Intersections of Subspaces Subspaces Definition 2. s e S}. The collection of Eij matrices can be called the "natural basis matrices." 3.2. 1. R S = (in general. V for an arbitrary index set A).24. 2. determine !n(n 1) symmetric basis matrices. Theorem 2. J)th location. Example 2. U\ + 1. 2. Let (V. R C S. is not necessarily a subspace. R j ) = 0 am/ Ri = T).24. 2.4." The collection of E. j)th location. Remark 2. a eA CiEA f] n *R. Theorem 2. S. V (in general. A vector space V is finitedimensional if there exists a basis X with n < +00 elements.) (To see why. RI \ h Rk =: ]T ft/ C V. U S. determine 1/2n(n + 1) symmetric basis matrices. otherwise. 1. dim{A E ~nxn :: A is upper (lower) triangular} = 1/2n(n+ 1). dim{A € Rnxn A AT} = {1/2(n 1 (To see why. « The subspaces R. t1]) . i E m. JF') be a vector space and let 71. dim{A E ~nxn :: A = AT} = !n(n + 1). 2. s E 5}. Definition 2.=1 K k 1=1 2. 4. where Eij is a matrix all of whose elements are 0 except for a 1 in the (i. V is infinitedimensional.+00.j matrices can be called the "natural basis matrices.. H. The subspaces Rand S are said to be complements of each other in T. dim(Rn)=n. and 2. we define dim(O) = 0. The sum and intersection ofR and S are defined respectively by: of R.a S. dim(~mXn) = mn. n n S = 0. dim(~n) = n. 1. V. n (^ ft. 1.. tJJ) = +00.18 says that dim(V) = the number of elements in a basis. Remark 2. Thus. and because the 0 vector is in any vector space. vector space V is finitedimensional if there exists a basis X with n < +00 elements. Thus. Ra C V/or an arbitrary index set A). S c V. R D S C V (in general.20.21. ft n 5 = {v : v E 7^ and v E 5}. 72.) 2 5.18 says that dim (V) the number of elements in a basis. R S C V (in general. The union of two subspaces. . Theorem 2. R H S = {v : v e R and v e S}. A consistency. 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. Note: Check that a basis for Rmxn is given by the mn matrices Eij. n S {r s : r E U. Sums and Intersections of Subspaces 13 13 consistency. we define dim(O) = O. dim{A e Rnxn A is upper (lower) triangular} = !n(n 1). R. and because the 0 vector is in any vector space. and S are said to be complements of each other in T.4 2. The union of two subspaces. y>f (L L .= T). for finite k). Example 2. R + S = {r + s : r e R. . 2. and S are defined respectively by: 1. Let (V.. K + S S. dim(C[to. i e m.23. R = 0. V (in general. for finite k).) = 0 and ]P ft. 5. V is infinitedimensional. where Efj is a matrix all of whose elements are 0 except for a 1 in the (i.22. F) be a vector space and let R. dim(R mxn ) mn. Note: Check that a basis for ~mxn is given by the mn matrices Eij. V.
.. Then it may be checked that U + .. Theorem 2... Suppose T = R O S.. ft. . let R be the set of skewsymmetric matrices in (V.c = jRnnxn jRn xn. e jRnxn 4. and SI.r2 £ Rand 52 . validity of the formula given in Theorem 2. jR). *2.14 14 Chapter 2. . Example 2. Theorem 2. XI. Xk} must be a linearly independent set.27. and let R"x". Let x\.r2 S2 . We discuss more about orthogonal complements elsewhere in the text.. Vn thonormal if and only if A E R"x" is orthogonal. together with Examples 2. For arbitrary subspaces ft. . . Let VI.25. every t E T can be written uniquely in the form tt = r + s with r E Rand s E S.25.26. But as t = r1 + s1 = r2 + S2.20. ft be the set of symmetric matrices in R" x ". . x/c E R" be nonzero mutually orthogonal vectors. Then any other distinct line through the origin is and let R be any line through the origin.. Avn are also orjRn. Show that {XI.29. . 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).5.r .29. while U n £ is the set of diagonal matrices in Rnxn.vn be orthonormal vectors in R". Example 2....27. ft. Using the fact that dim {diagonal (diagonal matrices} = n. Then r1 — r2 = s2— SI.. Then show that one of the vectors 1. 2. S2 E S. one can easily verify the validity = n.. Vk} is a linearly dependent set. mutually [x\.. D Theorem 2. unique ft.s\. 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 . . we must have r\ ri and SI rl . .. the set in jRnxn. jRn xn . R). where rl.27. every t € can be written uniquely in the form r s with r e R and s e S. = dim(ft) + Proof: To Proof: To prove the first part. Consider the vectors VI — [2 1f and V2 = [3 1f. .. F) (R n x n . The complement of R (or S) is not unique. IF) = (jRnxn.28. and let S Let (V. dim(T) = dim(R) + dim(S).. .. r2 e R.. Since ft fl 0. 1 TIT The first matrix on the righthand side above is in S while the second is in R.27. X2. Suppose {VI. For example. S of a vector space V. Show that Av\.c the Example 2.c jRnxn. consider V = R2 unique. Then V = U $ S.si e S. we must have rl = r2 and s\ = si from S2 from which uniqueness follows. EXERCISES EXERCISES 1. v = [4 l]r jR2. dim(R + S) = dim(R) + dim(S)  dim(R n S).. . Xk} must be a linearly independent set. Find the components of the vector v = [4 If with respect to this basis. . ft S) = jR2 and let ft be any line through the origin. r2 E Rand s1. For arbitrary subspaces R.. where r1. Then Theorem 2. AVn are orv\.26. jRnxn. 0 The statement of the second part is a special case of the next theorem.20.. triangular + L = R xn un. Vector Spaces Remark 2. Then any other distinct line through the origin is a complement of R. Av" •. But r1 –r2 E ft and S2 — SI E S. {vi.. Let U be the subspace of upper triangular matrices in E" x" and let £ be the subspace of lower triangUlar matrices in Rnxn. of the formula given in Theorem 2. Since R n S = 0.28. S of a vector space V. 3. which uniqueness follows. 0 S. Vector Spaces Chapter 2. vd must be a linear combination of the others..2 and 2. suppose an arbitrary vector t E T can be written in two ways t e as t S2. Xk E jRn 2. ... S2 e rl Sl r2 Then r. Example 2. Suppose =R EB Then 1. 2. Among all the complements there is a unique one orthogonal to R.
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
.. F) is linear and further suppose that {Vi. . i e n} and {Wj..n.} are the usual (natural) bases. is arbitrary).e. £V = W A since x was arbitrary.• + E nVnn = V x (where u. Note that A = Mat £ depends on the particular bases for V and W. + ~nLvn =~IWal+"'+~nWan = WAx... i e ~}.. 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." to Rm usually causes no Thinking of A both as a matrix and as a linear transformation from Rn to lR. w ] and L is the ith column of A. if v = ~I VI + • • + ~n v = Vx (where v. F) —>• (W. and hence jc. Then the {w j. then LVx = Lv = ~ILvI + . In other words. Thinking of both as a matrix and as a linear transformation from JR. i. We identify A the equation £V = W A becomes simply £ = A. W = R m and [ v i . Thus. Thus. . When V = R". j E raj...} are bases for V and W.e.. respectively. i E n}. When V = JR. then arbitrary). Li near Transformations Chapters. with respect to {w }•. L IF) ~ (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. w m] and where W = [WI. Linear Transformations 3.. + amiWm where W = [w\. Thus.18 Chapter 3. {W jj' j e !!!.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. 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. IF) veniently in matrix form.. j E m} are the usual (natural) bases WA linea LV L = A. suppose £ : (V. i. if V = E1v1 + .. transformation with its matrix representation...m and {Vi.m usually causes no naturally confusion. j e m}. Change of basis then corresponds naturally to appropriate matrix multiplication. is by its action on a basis. W = lR. + . {u. n A= al : ] E JR. j E !!!. z'th V This could be reflected by subscripts.2 3... say. In other words. Thus. [ w . Specifically. in the notation. and hence x. LV WA since x was arbitrary. for V and W) is the representation of £i>. but this is usually not done.
Two Special Cases: Two Special Inner Product: Let x. Inner Product: n xTy = Lx. we have C A B . Then we can define a new transformation C as follows: C The above diagram illustrates the composition of transformations C = AB.=1 Outer Product: Let x e Rm. If dimZ// = p. in the same order in both the diagram and the equation. Then we can define a new transformation C as follows: to W. then composition of transformations corresponds to standard matrix mUltiplication. That is.3. and dim W m.3. Then their inner product is the scalar E ~n. . Composition of Transformations 19 19 3. V. y e ~n. If dimU = p. expressed mxp nxp formula cij = L k=1 n aikbkj. . xx T XX ). 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). dim V = n. the arrows above are reversed as follows: C However.y. y e Rn. and dim W = m.3. The above is sometimes expressed componentwise by the C — A B . Then their outer product is the m x n E ~m. then composition of transformations corresponds to standard matrix multiplication. Note that in The above diagram illustrates the composition of transformations C = AB. Note that in most texts. Composition ofTransformations 3. That is. A rankone symmetric matrix can be written in above (or xy if A E C xyH e c ). and if we associate matrices with the transformations in the usual way..3 Composition of Transformations Composition Consider three vector spaces U. the form XXT (or xx HH). 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. and W and transformations B from U to V and A from Wand V to W. dimV = n. and if we associate matrices with the transformations in the usual way.
—/=== .an]. . is the set {v E V : Av = O}. {[ ~~i ]. of of denoted Im(A).i •. 2. The nullspace of A. The set is said to 3. ..3. { t > . Note that N(A) and R(A) are.6. LinearTransformations Chapter 3. R(A) = {Av : v E V}.. an} . Theorem 3. orthonormal set. Note that N(A) and R(A) are. if i f= j.. essentially following immediately from the definition. .. then Proof: The proof of this theorem is easy. {[ ~ J..Vk } With Vi E. denoted N(A). See also the last paragraph of Section 3. an M.7. .. 0 nition.3. essentially following immediately from the defiProof: The proof of this theorem is easy. {v1. ~. subspaces of different spaces. See also the of Section 3. is the set {w e w = Av for some v e V}. Definition 3. is an orthonormal set. The nullspace of kernel of and A is also known as the kernel of A and denoted Ker (A).4 3. ." • orthonormal set. vi ^/v'k vk ~~~ ] .. ~ 3 . Then Let A : V —>• be a linear transformation. e ~mxn. Theorem 3. [: J} is an orthogonal set. V. is an orthogonal set.7. the same symbol (A) is Note that in Theorem and throughout the text. . subspaces of different spaces. then then R(A) = Sp{al. The range of A.. Li near Transformations 3... Then 1. If in of = [a\. N(A) c V.•. N(A) S.. D Remark 3.. R(A) C W.2. Example 3. Let A : V + W be a linear transformation. Vk} with u.. an].[ :~~ J} . I ~VI VI ^/v. . is an orthonormal set.5 and throughout the text. denotedlZ( 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. Let {VI. . in general. Let A : V + be a linear transformation. 3.4 Structure of Linear Transformations Structure of Linear Transformations Let A : V —> W be a linear transformation. Note that in Theorem 3. e Rn. is the set {w E W : w = Av for some v E V}. R ( A ) S. 2. 1.. Let A E Rmxn. 1. Definition3. Definition 3. where 8ij is the Kronecker delta defined by Kronecker delta defined by 8 = {I0 ij ifi=j. Equivalently. The nullspace of The of denoted N(A). € 1Tlln is an orthogonal set. in general. usual (natural) bases.8...4.20 20 Chapter 3.IN. (A).5. W. vd of u. The range of A is also known as the image of A and — {Av e V}.. then ~ . If A is written in terms of its columns as A = [ai. denoted Im(A).8... denotedR(A). . . IS an orthogonaI set. 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. . is the {v e V Av = 0}. vk] be a set of nonzero vectors Vi E ~n. If {VI. 2.. then {I —/==.2. The range of A. ~}  ISisan 3.
Note that there is nothing special about the two vectors in the basis defining S being orthogonal. Structure of Li near Transformations 3. The proofs of the other results are left as Proof: left exercises. (n n S)~ = n~ + S~. S1. k =X . . = S. Let 3.=1 XI.10. . the computation involved is simply to find all nontrivial (i.10. Then the orthogonal complement of S is defined as the set c ]Rn.. vk} e ]Rn vector. Set vector.. ]Rn. Any set of vectors will do. 3.e. n~... Theorem 311 Let Theorem 3. Then n.9. Let R S C Rn The S <. 4. S 5. nonzero) solutions of the system of equations 3xI 4xI + 5X2 + 7X3 = 0. Example 3. S \B S~ = ]Rn. of course.3. Let {VI. 2. (S~)l.4.. Vk} be an orthonormal basis for S and let x E Rn be an arbitrary {v1. Then the of defined T S~={VE]Rn: V S = 0 for all s e S}.4. Then it can be shown that Working from the definition.. n S~. . then give rise to redundant equations). Set XI X2 = L (xT Vi)Vi. (n + S)~ = nl. Rn. + X2 + X3 = 0.11. n <. including dependent spanning vectors (which would. 6. Let S <.= {v e Rn : vTs=OforallsES}. if and only if S~ <. Structure of Linear Transformations 21 21 Definition 3. . Proof: We prove and discuss only item 2 here.
x~ e S.l = N(A ). (Note: This also holds for infinitedimensional vector spaces.. The proof of the second part is similar and is left as an exercise. In other words.22 22 Chapter 3.. R" N(A) 0 ft(Ar ». every vector in the codomain space R m can be written ina unique way asw = x+y.e..) N(A)1" spaces. Thus. When thought of as a linear transformation from IR n to Rm. Let A : IRn > Rm. many properties of A can be developed in terms of the four fundamental subspaces to IRm. Vk and hence to any linear combination of these we see that X2 is orthogonal to VI. Clearly.(A)1~ — J\f(ATT ). Let A : R" + IRm. E S and X2. It can write vectors in a unique way with respect to the corresponding subspaces. where x e U(A) and y e ft(A)1. every vector v in the domain space R" can be written in a unique way as v = x + y. x. Suppose. Ax = 0 if and only if x orthogonal is orthogonal to all vectors of the form AT y.13. +x~).. and x2 = x~.. (Note: This also holds for infinitedimensional vector spaces. (w e Rm : w T A = 0} is called the left nullspace of A. we form AT v. ft(Ar) (i. i.l where x € M(A) and y € J\f(A)± = R(AT) (i. Thus. In other words. R(A).11 can be combined to give two very funTheorem 3. 0 The proof of the second part is similar and is left as an exercise.) 2. 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. Similarly.. E N(A) and E N(A). that x = XI for example. X2 is orthogonal to any vector in S. But yT Ax = ( A T ) x. . .. Then X2 = x. that x = x1 + x2. we see that x2 is orthogonal to v1.l = 0 since the only vector s E S orthogonal to S1 = IRn. IRn = M(A) EB R(A T)). the right nullspace is A/"(A) while the left nullspace is N(A T ). Li near Transformations Chapters. every vector w in the codomain space IRm can be written in a unique way as w = x+y.e. But then (x'1 —XI)TT (x. See also Theorem 2. Then 1.= Af(AT) ) (i.l = Rn. E R(A) and E R(A).12.e. for example.e. Then {v E R" : Av = 0} is sometimes called the right nullspace of A. many properties of A can be developed in terms of the four fundamental subspaces . (Note: This holds only for finitedimensional vector spaces.5 3.26. the right nullspace is N(A) while the left nullspace is J\f(AT). Then Theorem 3. {w E IR m : WT A = O} is called the left nullspace of A. 0 x1 — x'1 andx2 = x2. When thought of as a linear transformation from E" Consider a general matrix A € E^ x ". But then (x. + x~. right nullspace of A. x~ X2 = (x. Theorem 3. – x1) = 0 since 0 by definition of S. Similarly. XI) (which follows by rearranging the equation XI +X2 = x.. Linear Transformations Then x\ E <S and. x'2 E S1. D Definition 3.12.. Ax = 0 if and only if x equivalent to yT Ax = 0 for all y. 'R. We have thus shown that S + S. Then Ax = 0 and this is an and equivalent to yT Ax = 0 for all v. But yT Ax = (ATyy{ x. Let A : Rn + IRm.X2) 0 since (x'1 — x1) (x' 2 — x2) = 0 by definition of ST. Then {v e IRn : A v = O} is sometimes called the Definition 3.e. transformation A. (Note: This for finitedimensional 1.12 and part 2 of Theorem 3. We S n S1 =0 the e orthogonal everything in (i.e. Thus. including itself) is O. Then T (x. Then Theorem 3.1 in the next section. everything in S (i. x e R(AT).5 Four Fundamental Subspaces Four Fundamental Subspaces Consider a general matrix A E lR. = x'1+ x'2. Let A : IRn —> Rm. 3.XITVj =XTVjXTVj=O.13.l N(A T (i. . See also Theorem 2.e. since Then XI e S and. Thus.XI/ (x~ .14 (Decomposition Theorem). Vk and hence to any linear combination of these vectors.l = R(A T}. We have thus shown that vectors.e. Let A : Rn + Rm. standing Figure 3. where x\.xn. Ax = Proof: To prove the first part. y. Since x was arbitrary. Theorem 3.11 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.x1) (x'1 xd x2 — X2 = — (x'1 — x1) (which follows by rearranging the equation x1+x2 = x'1 + x'2). It is also easy to see directly that.. x 1 E Sand x2.) 2. where XI.l.. D Theorem 3. Rm = 7l(A) 0 M(AT)).l.l. can write vectors in a unique way with respect to the corresponding subspaces.•. we decompositions.) Proof: To x E N(A). i.. XI = x. since T x 2 Vj = XTVj . . Let A : IRn > IRm.14 (Decomposition Theorem). x E R(A r ) Since x 1 have established thatN(A). take an arbitrary x e A/"(A). N(A). Clearly. Then R(A r ). 2. Suppose. including itself) is 0. We also have that S U S.26. X2 is orthogonal to any vector in S. established that N(A) U(AT ).l = 7£(AT).1 in the next section. IRm = R(A) EBN(A T».12 and part 2 of Theorem 3.. when we have such direct sum decompositions. every vector v in the domain space IRn can be written in a unique way as v = x 7. .
The row rank of A is column rank of of independent row rank of . 1. Four fundamental subspaces. Then rank(A) = dim R(A). properties 7£(A). and in illustrating concepts such as controllability and observability. Two equivalent 2.16.3.(A)^. IR n > IRm. Let A : E" + Rm. Definition 3. Let and W be vector spaces and let A : motion. 3.1.16. Figure 3. This is sometimes called 3.(A) = W.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.5. Four Fundamental Subspaces 23 23 A r N(A)1 r EB {OJ X {O}Gl nr m r Figure 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 . 'R. and N(A)1.1.1 makes many key properties seem almost N(A)T. be a linear transforDefinition 3. N(A).5. A is onto (also called epic or surjective) ifR(A) = W. R(A). 1. Four Fundamental Subspaces 3. 2. A is onetoone or 11 (also called monic or infective) ifJ\f(A) = 0. A is onto (also called epic or surjective) ifR. t= V2 ===} AVI t= AV2 . the column rank of A (maximum number of independent columns). Figure 3. fundamental subspaces. A is onetoone or 11 (also called monic or injective) if N(A) = O. A f ( A ) . R(A)1. Let V and W be vector spaces and let A : V + W be a linear transforDefinition 3. mation. rank(A) dimftCA).15.15.
rank(B)}. r*i *i E N(A)L. . the subspaces themselves are not necessarily in the same vector space.. The dual notion to rank is the nullity R(AT) of independent rows). The basic results are contained in the following easily proved following theorem.11 and 3. 1 1 Xl E A/^A) . dimension of the domain of A. dimA/'(A) ± (Note: 1 T T ). . then {TVI. . denoted nullity(A) or corank(A). . we include here a few miscellaneous results about ranks of sums completeness. Like the theorem.19 suggests looking at the general problem of the four fundamental Part 4 of Theorem 3. The last equality AXI x\ e N(A)L and jc E N(A).. and is defined as dimN(A). Let A. dimensions.19. by definition there is a vector x E ]Rn such that Ax = w. Then Ajti = W = TXI since Xl e A/^A). 1.18. 3.. of A. if {ui. R(A) : ]Rn ~ ]Rm. + rank(B)  n :s rank(AB) :s min{rank(A). nullity(B) :s nullity(AB) :s nullity(A) 4. . (Note: Since 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. .17 we see immediately that Proof: From Theorems 3. LinearTransformations Chapter3. Finally. . Tvrr]} is a basis for R(A).11 and 3. Tv abasis 7?. O:s rank(A 2. it is a statement about equality of dimensions. this theorem is sometimes colloquially stated "row rank of A = column N(A)L = R(A A/^A) " = 7l(A ). where Ax — w. take any W e R(A). e ]Rnxn. B E R" xn . and is defined as dim A/"(A).. . {Tv\. where n is the ]Rn > ]Rm. iv} abasis forA/'CA) .. 0 For completeness. + nullity(B). Then dim K(A) = dimNCA)L. rank(AB) = rank(BA) = rank(A) and N(BA) = N(A). and products of matrices.") of A... the subspaces themselves are not necessarily in the same vector space. colloquially of = rank of A. Proof: From Theorems 3. Clearly T is 11 (since A/"(T) = 0). x x e R" x\ X2. To see that T is also onto.17 we see immediately that n = dimN(A) = dimN(A) + dimN(A)L + dim R(A) . u. sometimes denoted nullity(A) or corank(A).19. Then 3. Let A : R" ~ Rm.(A).andx22 e A/"(A).24 24 Chapter 3. following follows we apply this and several previous results. Let A : Rn > Rm.17. Part 4 of Theorem 3.(A) = dimA/^A^ 1 if that if {VI. Write x = Xl + X2. Linear Transformations dim 7£(A r ) (maximum number of independent rows). We thus have that dim R(A) = dimN(A)L since it is easily shown T dim7?.18. Theorem 3. v r } is a basis for N(A)L. of Corollary 3." 0 of D The following corollary is immediate. rank(A) + B) :s rank(A) + rank(B).17. dimA/"(A) + dimft(A) = dimension of the domain of A.19 suggests looking atthe general problem of the four fundamental subspaces of matrix products. . Then N(T) = To w E 7£(A). if B is nonsingular. shows that T is onto. .. Theorem 3. 3. Then dimN(A) + dim R(A) = n.") 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 A. .
R(AT) 3.2 N(B). suppose AXI = Ax^. which implies that dim A/^A).(A).20. the transformations A. Four Fundamental Subspaces Theorem 3. A is AT AXI AT AX2. 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.22.22. RCAB) S. A : IRn »• IR n is invertible or Note that if A is invertible. Conversely. especially when dealing with pseudoinverses and linear least squares problems. A : V —» W is invertible (or bijective) if and only if it is 11 and onto. N«AB)T) . N(A) = N(A T A). A : W1 + E" is invertible or nonsingular if and only z/rank(A) = n. Also. Conversely.ti = AT Ax2. Then y = Ax.17. 4. then dim V — dim W. A A T. 1. then A/"(A) = 0. AT. R(B T ). i. A is 11 if and only z/rank(A) = n (A has linearly independent columns or is said to have full column rank. equivalently. then N(A) = 0. Then 3. N(A T ) = N(AA T ). Conversely. 3. A is onto if and only if rank (A) = m (A has linearly independent rows or is said to have full row rank.e.. 2. Theorem 3. y E R(A).(A) — m — rank (A). D D 11. have full row rank. equivalently.21. A is onto if and only //"rank(A) — m (A has linearly independent rows or is said to 1. Let A : IRn + IRm. 4. AX2. since ArA is invertible. linear least squares problems. AT A is nonsingular). : R n » Rm. N(AB) . equivalently. suppose Ax\ dim R(A T). Let jc = AT(AAT)~]y Y E Rn. Note that if A is invertible.—n = dim 7£(A r ). A is 11 if and only ifrank(A) = n (A has linearly independent columns or is said 2. Note that in the special case when A E R"x".17. and hence dim R(A) n by Theorem 3. R«AB)T) S. Then 3. Definition 3. RCA). Then A r A. Let A E Rmxn. 1. XI = X2 AT A A 11.2 N(A T ). equivalently. A € IR~xn. = R(A T A).21. Let e IRmxn. 2.20 and is also easily proved. Proof' Proof of part 1: If A is onto. The transformations AT are all 11 and onto between the two spaces N(A)1. terms of rank and invertibility. Definition 3. AT A nonsingular). A Proof of part 2: If A is 11. Let A E Rmxn. A"1 ± are all 11 and onto between the two spaces M(A) and 7£(A). and AI A. AA is nonsingular). Conversely. Thus. 1. 25 25 The next theorem is closely related to Theorem 3. dim7?. and hence dim 7£(A) = n by Theorem 3.23. Similar remarks apply to A and A~T. to have full column rank. especially when dealing with pseudoinverses and is extremely useful in text that follows. It The next theorem is closely related to Theorem 3. B E Rnxp. AATT is nonsingular).23. Four Fundamental Subspaces 3. x AT (AAT)I e IRn. Ar.3. then dim V = dim W. 2. which implies that dimN(A)11 = n — Proof of part 2: If A is 11.and R(A). 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. Also. . which implies x\ = x^. It is extremely useful in text that follows. nonsingular ifand only ifrank(A) = n. let y e Rm Proof: Proof of part 1: If A is onto.5. e IRmxn. e IRnxp.20 and is also easily proved. let y E IRm be arbitrary. dim R(A) = m = rank(A).20. R(A) = R(AA T ). A : V + W is invertible (or bijective) if and only if it is 11 and onto. Theorem 3. e 7?.5. Then Theorem 3. so A is onto.
25. 2. Let A = [1 2] : E2 »• E1I. it may still be right or left invertible.24. it is clear that there are infinitely many right inverse. Obviously A has full row rank can always find v E ]R2 such that [1 2][ ~~] = a).: AA R = w Iw W + V such that AA~R = Iw. Let Theorem 3. 1. then A is invertible. then one (Proo!' = [1 2]:]R2 + ]R . Theorem 3. i.e. linear Transformations If a linear transformation is not invertible.22 we see that if A : E" + Em is onto. then a left inverse is given by A R = AT (AAT) left T L = A. Let A : V + Then 1. Let A : V + W. 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. D Example 3. i. (A R + A RA . Li near Transformations Chapters. (A R + A R A — /) must be a right inverse and.I) must be a right inverse and. A is invertible if and only if it is both right and left invertible. But this implies that A~RA = /.24.. in which case A~l = A~R = A~L. where Iv denotes the identity transfonnation on V.25 that A is invertible.. It then follows from Theorem 3. If Proof: proof of second Proof: We prove the first part and leave the proof of the second to the reader. i. that A~R is a left inverse. A is left invertible if and only ifit is 11. then A is invertible. If there exists a unique left inverse A~L such that A~LA = I. Let A : V > W. 1. A right invertible if and only if it onto. € ]R . therefore.26 Chapter 3.. Then A is onto. A~ (A A)~ A . can always find v e E2 such that [1 2][^] = a). therefore.e. Definition 3. is left invertible if and if it and left invertible. Then Definition 3. Let + V.26. —> transformation if left + 2. where Iw denotes the identity transfonnation on W. 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.R + AARA = I A +IA  A since AA R = I = I.. Defileft If linear concepts left nitions of these concepts are followed by a theorem characterizing left and right invertible transformations.e. characterizing all solutions of the linear matrix equation AR = I. both 11 and Moreover. Then > 1.L = (ATTA)I1AT.I = A~R.both 11 and is if and if onto.22 ]Rn >• ]Rm Note: From Theorem 3. Obviously A has full row rank (= 1) and A .. It then follows from Theorem 3. Also. A R A = I. are infinitely A. If there exists a unique right inverse A~R such that AA~R = I. by uniqueness it must be A R + A R A — = A R. 1. if A is 11.R = If left A L A L A = 2.R AA. A R the case that A~R + A~RA . Similarly. 0 a left inverse.R + ARA I) = AA. Theorem 3.25 that A is invertible.R = _~] (=1) and A~R = [ _j j is a right inverse.. Notice the and leave the following: following: A(A. (Proof: Take any a E E1I. in A I = A R = A L. right inverses for A. 3.. by uniqueness it must be Thus. Let A : V » V. then a right inverse is given by A~R = AT(AAT) I.27. such that A~LA = Iv where Iv denotes the identity transformation on V. A is said to be right invertible if there exists a right inverse transformation A~RR : if A. .26.e. A is right invertible if and only if it is onto. i. A.
and let R denote the subspace of skewsymmetric matrices. £. Show that. 3. Prove Theorem 3. The matrix A = 1 1 2 1 [ 3 1 when considered as a linear transformation onIE \ is neither 11 nor onto. Let A = [i]:]Rl > ]R2. Consider the vector space R nx " over E. It is now obvious that A has full column rank (=1) and A~L = [3 .3.Exercises 27 2. Y E Enx" define their inner product by (X. 2. Let A = [8 5 i) and consider A as a linear transformation mapping E3 to ]R2. EXERCISES EXERCISES 3 4 1. Y) = Tr(X Tr F). and let 7£ denote the subspace of skewsymmetric matrices.4.4. 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 . R = S J. 3. y) = Tr(X Y). whence A/"(A) = 0 so A is 11). respect to this inner product. 2 . Then A is 11. it is clear that there are A L = [3 — 1] infinitely many left inverses for A. is neither 11 nor onto. (Proof Theonly solution toO = Av = [i]v 2. Consider the differentiation operator C defined in Example 3. II? Is £. In Chapter 6 we characterize all left inverses of a matrix by characterizing all solutions of the linear matrix equation LA = I. For matrices X. . below bases for its four fundamental subspaces. Prove Theorem 3. The matrix 3. let denote the subspace of symmetric 2. In Chapter 6 we characterize all left inverses of a infinitely many left inverses for A. with Y e ]Rnxn (X. Again. (Proof: The only solution to 0 = Av = [I2]v is v = 0. Consider the vector space ]Rnxn over ]R. J E2.2. respect to this inner product. Let A = [~ .1] is a left inverse. ThenAis 11. LetA [J] : E1 ~ E2. We give below bases for its four fundamental subspaces. let S denote the subspace of symmetric matrices. 4. Is £. For matrices matrices. Consider differentiation £ 11? Is£ onto? onto? 4. 'R. whence N(A) = 0 so A is 11). — S^. consider A linear transformation ]R3 1. We give when considered as a linear on ]R3. matrix characterizing all solutions of the linear matrix equation LA = I.
prove it. Prove Theorem 3. Chapter 3. Linear Transformations Chapters. homogeneous linear system Ax = 0? homogeneous linear system Ax = O? n 3.4.1 to illustrate the four fundamental subspaces associated with AT e associated ATE nxm IR from IR m R". Are they equal? Is this true in general? If this is true in general.12.2. linearly independent solutions 10. If E 1R~9X48. Let A = [ J o]. Rnxm thought of as a transformation from Rm to IRn. if not.1 11. . Prove Theorem 3. Suppose A E IR m xn has a left inverse.11. ~ ~ 3 8. Theorem 6. Linear Transformations 7. Determine A/"(A) and 7£(A). How many linearly independent solutions can be found to the 10.28 5. provide a counterexample.4. Let = [~ 9. 3. Are they equal? Is this true in general? DetennineN(A) and R(A). left T Suppose e Rmxn 9. Modify Figure 3. Show that AT has a right inverse. Suppose A € Mg 9x48 .Il. Determine bases for the four fundamental subspaces of the matrix Detennine fundamental A=[~2 5 5 ~].
. where and are arbitrary finiteConsider a linear transformation A : X + y. as noted in the proof of Theorem 3.l —>• Tl(A) by Tx = Ax for all x E NCA). 29 . Then A+ is the MoorePenrose where y = y\ pseudoinverse of A. the definition neither provides nor suggests a good computational strategy good computational strategy for determining A +. Then A+ is the MoorePenrose j2 with y\ E RCA) and yi E RCA). Definition 4.1 Definitions and Characterizations Definitions and Characterizations Consider a linear transformation A : X —>• y. With A and T as defined above. neither provides Unfortunately. Although X and y were arbitrary vector spaces above. Define a transformation T : Af(A)1. Then. define a transformation A + y + X by where Y = YI + Yz with Yl e 7£(A) and Yz e Tl(A}L.l.1 4. as noted in the proof of Theorem 3. where X Xand Y y are arbitrary finitedimensional vector spaces. a generIn this chapter we give a brief introduction to the MoorePenrose pseudoinverse. pseudoinverse of A.l. let us henceforth consider the Although X and Y were arbitrary vector spaces above. the MoorePenrose pseudoinverse of A. which was proved by Penrose in 1955.m We A+ A e lP1. characterization of A is given in the next theorem.1. the MoorePenrose pseudoinverse of A. as is shown in the following text. case X = W1 and Y = Rm. can be used to give our first definition of A .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. see [22]. which was proved by Penrose in 1955. A purely algebraic y + characterization of A+ is given in the next theorem. T is bijective (11 and onto). This transformation T~ + can be used to give our first definition of A+.+ R(A) by dimensional Define transformation T : N(A).1." X ".l.. problems. let us henceforth consider the X ~n lP1. define a transformation A+ : Y —»• X by Definition 4. 4. a generalization of the inverse of a matrix. T is bijective Cll and onto).17. for determining A+ . see [22]. brings great notational and conceptual clarity matrix and. as is shown in the following text.17.. With A and T as defined above. and hence we can RCA) —>• J\f(A}~L This transformation can define a unique inverse transformation Tl 1 :: 7£(A) + NCA). 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. We have thus defined A+ for all A E IR™xn. and hence we Then. The MoorePenrose pseudoinverse is defined for any matrix and.
A~ = [3 . Also. (P4) (GA)T = GA. A L = [3 — 1]) satisfy properties (PI). p.2. whose proof Still another characterization of A + is given in the following theorem. and (P4) but not (P3). Furthermore. Unfortunately. Note that other left inverses (for example. However. Example 4. Theorem 4. the Penrose properties do offer the great virtue of providing a tional algorithm. the Penrose properties do offer the great virtue of providing a checkable criterion in the following sense. Still another characterization of A+ is given in the following theorem. if a t= 0.2 nor its proof suggests a computawith Definition 4. Also. Example 4. Let A E lR.2) 4. 19]." xn. as a right or left inverse satisfies no fewer than three of the four properties. terizations.1]) satisfy properties (PI). Then G = A+ if and only if Theorem 4.7. Example 4. While not generally suitable for computer implementation. X+ = AT(AATT) I if A is onto (independent rows) (A is right invertible). Example 4. it must be A+.7. If G satisfies all four. A+ = (AT A)I AT if A is 11 (independent columns) (A is left invertible). Then Theorem 4. one need simply verify the four Penrose conditions (P1)(P4).2 4. For any scalar a. Such a verification is often relatively satisfies all four.1. L Note that other left inverses (for example. Verify directly that A+ = Example 4. For any scalar a. (P3) (AGf (P3) (AG)T = AG.6. and (P4) but not (P3).3. this can be found in [1..6. as with Definition 4. If G the pseudoinverse of A. if a =0. Given a matrix G that is a candidate for being checkable criterion in the following sense. whose proof can be found in [1.4. A t = AT (AA )~ if A is onto (independent rows) (A is Example 4.1) = limAT(AAT +8 2 1)1. Let A e R™xn. Introduction to the MoorePenrose Pseudoinverse Theorem 4. neither the statement of Theorem 4. neither the statement of Theorem 4. (P2) GAG = G. (P2) GAG G.30 Chapter 4. A + always exists and is unique. Given a matrix G that is a candidate for being the pseudoinverse of A. (P4) (GA)T = GA. Let A E lR.2. Note that the inverse of a nonsingular matrix satisfies all four Penrose properties. AG. Consider A = [']. a right or left inverse satisfies no fewer than three of the four properties. Then A+ [a [! = lim (AT A + 82 1) I AT 6+0 6+0 (4. (P2). Consider A = f ] satisfies (P1)(P4). = Furthermore.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.3. Example 4. Let A e R?xn Then G = A + if and only if (Pl) AGA = A. Verify directly that A+ = [ ~] satisfies (PI)(P4). (PI) AGA = A. it must be A +.4. one need simply verify the four Penrose conditions (P1)(P4). . Such a verification is often relatively straightforward. p. Introduction to the MoorePenrose Pseudoinverse Chapter 4." xn. this characterization can be useful for hand calculation of small examples.2 nor its proof suggests a computational algorithm. (P2). Example 4.5. 19]. characterization can be useful for hand calculation of small examples. A+ always exists and is unique.1. However. While not generally suitable for computer implementation. straightforward. then by uniqueness. then by uniqueness. (4. Unfortunately. A+ = (AT A)~ AT if A is 11 (independent columns) (A is left invertible). Note that the inverse of a nonsingular matrix satisfies all four Penrose properties.5.
Many of these are used in the text that follows.9. The interested reader can consult the proof in [1. For A e Rmxn 1. Properties and Applications Example 4.10. 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. simply verify that the expression above does indeed satisfy each c the four Penrose conditions. 4.11. elements are determined according to Example 4. Let A E R m x "and suppose UUEejRmxm. Then orthogonal if MT = M.13. Let S E Rnxn be symmetric with U TSU = D.4. Example 4. Properties and Applications 4. are used in the text that follows. Example 4. p.4.12. [~ ~ r ~ =[ 0 Example 4.12. Many of these This section presents some miscellaneous useful results on pseudoinverses. The interested reader can consult the proof in [1. Example 4..11.9. 2. 0 the four Penrose conditions. Then Proof: For the proof. .13. if v i= 0.3 4. The Proof: Both results can be proved using the limit characterization of Theorem 4. simply verify that the expression above does indeed satisfy each of Proof: For the proof. e jRmxn and suppose Rmxm R n are orthogonal (M is T 1 1 orthogonal if M M ). D Theorem 4.). Let S e jRnxn be symmetric with UT SU = D. For any vector e jRn. 27]. 31 31 Example 4. A+ = (AT A)+ AT = AT (AA T)+.8.VVEejRnxnx " are orthogonal (M is 4. p. For all A E jRmxn. (A T )+ = (A+{.4.8. . if v = O.10.3. where D+ is again a diagonal matrix whose diagonc D is diagonal.3 Properties and Applications Properties and Applications This section presents some miscellaneous useful results on pseudoinverses. 27]. [~ r 1 =[ 4 4 I I ~l 4 I I 4 4. Theorem 4.7. Then S+ UD+U T where D+ is again a diagonal matrix whose diagonal elements are determined according to Example 4. 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. .3. For any vector v E M". Proof: Both results can be proved using the limit characterization of Theorem 4. Theorem 4. Then S+ = U D+UT. where U is orthogonal and D is diagonal. where U is orthogonal an Theorem 4.7.
however (see.11 nets of matrices such as exists for inverses of nroducts Unfortunately.13 we can. Proof: Proof: For the proof. where BI = A+AB and A) = ABIB{.15. 4. e.15. BB+ f r The by taking BI = B.11 is suggestive of a "reverseorder" property for pseudoinverses of prodTheorem 4.At = A in Theorem 4.13 can. (AB)+ = B+ A + if and only if 1. If e lR~xm.15. compute 4. see [9]. necessary and sufficient conditions under which the reverseorder property does hold are known and we quote a couple of moderately useful results for reference.32 Chapter 4. B E Rrrxm. [9].12 Note that by combining Theorems 4. and better methods are suggested in text that follows. Proof: Proof: For the proof. Ir Similarly. then (AB)+ = B+ A+. If A e Rnrxr. (AB)+ = B{ Ai.15. TTnfortnnatelv. A\ = A in Theorem 4. N(A+) 5. 2. Introduction to the MoorePenrose Pseudoinverse Chapter 4. [] sufficient reverseorder However.17. Introduction to the MoorePenrose Pseudo inverse 4. 3. A+ = (AT A)I AT. For e Rmxn . Then (AB)+ = 1+ = I while while B+ A+ = [~ ~J ~ = ~. since e lR~xr. This A AT AT turns out to be a poor approach in finiteprecision arithmetic.14. [II]. [5]. where BI = A+ AB and AI = AB\B+. (A+)+ = A. whence A+A = f r. (AB)+ = B+A+ if and only if 4. n(A+) 4. D takingB t = B. then AkA+ = A+ Ak and (Ak)+ = (A+)kforall integers k > O. (AT A)+ = A+(A T)+. = N(AA+) = N«AA T)+) = N(AA T) = N(A T). [11]. then A+ = (ATA)~lAT.. 0 D E lR~xr. xm + T B e Wr . The result then follows by E lR.16.. see [5]. As an example consider A = [0 1J and B = [ : J. = n(A T) = n(A+ A) = n(A TA).• Similarly. (AA T )+ = (A T)+ A+. If A is normal. poor (see. properties Theorem 4. 1. 0 D Theorem 4. Then As an example consider A = [0 I] and B = LI.14. (AB)+ = B?A+. [7]. in theory at least. 4.xm. Theorem 4. (AB)+ = B+A+. [23]). 4. we B+ BT(BBT)I. For all A E lR mxn . n(A T AB) ~ nCB) .16. e. 0 The following theorem gives some additional useful properties of pseudoinverses. .g. the MoorePenrose pseudoinverse of any matrix (since AAT and AT A are symmetric). we have B = B (BBT)~\ whence BB+ = Ir.12 and 4. n(BB T AT) ~ n(AT) and 2.. Proof' A+ A Proof: Since A E Rnrxr. [7]. in peneraK ucts of matrices such as exists for inverses of products. Theorem 4.g.17. Theorem 4. in general.
]. 2. b E E. If jc. Let A G M"xn. skewsymmetric. if A is symmetric. Then B and take arbitrary y E R(B).Exercises 33 Note: Recall that A E IRn xn is normal if A A = A T A. For A E IRmxn. Then there exists a vector x E Rm such that Bx = y. For A e R m x n . Then Bx e R(B) c H(A). (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.18. € IRm xm D 6.i D. 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. For A e Rmxn.4 to compute the pseudoinverse of U . R(A) and take arbitrary x e IRm. A e IRPxn thatN(A) S.1 D. Note: Recall that A e R" xn is normal if AATT = AT A. Then we have Bx = Ay = AA + Ay = AA + Bx. For A E Rpxn and BE R mx ". such as A=[ b a a b] for scalars a. we have shown that B = AA+B. we have shown where one of the Penrose properties is used above. Then R(B) c R(A) if and only if Suppose e IRnxp. Then we have there exists a vector y e IRP such that Ay = Bx. 0 EXERCISES EXERCISES 1. Use Theorem 4. assume that AA+B = B and take arbitrary y e K(B). N(B) and 5 € IRmxn. To prove the converse. Suppose A E Rnxp.i l . whereupon y = Bx = AA+Bx E R(A). 5 e JRn x m . Since x was arbitrary. so Proof: Suppose R(B) c U(A) and take arbitrary jc E Rm. or orthogonal.4 to compute the pseudoinverse of \ 2 1. However. Proof: Suppose K(B) S. Y e R". b e R for scalars a. show that (xyT)+ = (x Tx)+(yT y)++yxT. then it is normal. U(A) if and only if AA+B = B. Theorem 4.• 1 2 x. y E IRn. prove that 7£(A) = 7£(AA r ) using only definitions and elementary properties of the MoorePenrose pseudoinverse. or orthogonal. A+ 0 A+BD. A E IRn xn B E E n xm 6. and D E E mxm and suppose further that D is nonsingular. if A is symmetric. a matrix can be none of the preceding but still be normal. such as preceding but still be normal. Then Bx E H(B) S. that B = AA+ B. For example. prove that R(A+) = R(A T). B E E M X m . Since x was arbitrary. For example.i ]. properties of the MoorePenrose pseudoinverse. = B. a matrix can be none of the skewsymmetric.. (xyT)+ = (xTx)+(yTy) yx T 3. problems. Use Theorem 4. then it is normal. 4. show that JV(A) C A/"(S) if and only if BA+ A = B. so there exists a vector y E Rp such that Ay = Bx. where one of the Penrose properties is used above. whereupon there exists a vector x e IR m such that Bx = y. assume that AA + B To prove the converse. e IRnxm. A E IRmxn. However. RCA). fiA+A B. Then K(B) S. ft(A+) ft(Ar 5. prove that RCA) = R(AAT) using only definitions and elementary 3.
This page intentionally left blank This page intentionally left blank .
.. Let {u. Preand postmultiplying by SI gives the emotion S~l eives the equation (5.. .} with UI ::::: . .4) 35 . (5. Vi eE RIRnxr.e. U2 E IRrnx(mrl.. . Ch. Theorem 5.LettingSS = diag(uI. . r write A r A VI = ViS2.. Premultiplying by Vf gives Vf A T A VI write ATAVi = VI S2.1.2) (5. 6]). Proof: Since A r A > 00 A r A i is symmetric and nonnegative definite. Pre. More where ~ = [~ specifically. . vn]. for example. where S = [J °0]... the latter equality following from the orthonormality of the Vi vectors.1.• = Un. i. We In this chapter we give a brief introduction to the singular value decomposition (SVD). . ii E !!. u r ) e R ..Vn ].. for example. rcfr). recall. . its eigenvalues are all real and nonnegative.. Proof: Since AT A ::::: ( (AT A s symmetric and nonnegative definite. .) Denote the set of eigenvalues of AT A by {of / E n} with a\ > • • > a > 0 = o>+i = • • an.Vv r). 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.e..1) rxr A = [U I U2) [ ~ 0 0 ][ ] 2 T VI VT (5. 5.. . . UI e Wnxr. < min{m.. «}). (Note: The rest of the proof follows [24. vectors.. Premultiplying by vt gives vt ATAVi = vt VI S2 = S2.we can Vi = [vr+ .. Then there exist orthogonal matrices U E Rmxm and E IR~xn. i. 6]). Letting — diag(cri.. the latter VfV^S2 = S2. and the Osubblocks in are compatibly dimensioned.and postmultiplying by equality following from the orthonormality of the r.] .. specifically. [24. we have n = U~VT. recall. i e !!. . and a\ > • • • > Ur > 0. .... e IRmxm and V E IR nxn such that V € Rnxn such that UI > diag(ul.) Denote the set of eigenvalues of AT A by {U?. Let {Vi.. role throughout (numerical) linear algebra and its applications.\. . Let A e R™ x n . dimensioned.. S = diagfcri.. V2 = [Vr+I. . (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. VI «xr j V U2 e ^x(mr) .. Ch. y22 €E Rnxfor^ and the 0JM^/ocJb in E~are compatibly IRnx(nr).3) = Ulsvt· The submatrix sizes are all determined by r (which must be S min{m.} be a set of corresponding orthonormal eigenvectors and let VI = [v\. The SVD plays a key conceptual and computational of this important matrix factorization. its eigenvalues are all real and nonnegative. n}). More S > o r > O.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). U\ E IRmxr.o>) E IRrxr. e n} be a set of corresponding orthonormal eigenvectors 0= Ur+1 = .u ). The SVD plays a key conceptual and computational role throughout (numerical) linear algebra and its applications.1 The Fundamental Theorem Theorem A Theorem 5. we can and let V\ [VI. ..• ::::: Urr > as an exercise.
For example.. i.r zero singular values. an we Turning now to the eigenvalue equations corresponding to the eigenvalues ar+l. Now define the ATA V AV O.. u ]} for Rm (see the discussion in Section 3.(AT A) = is denoted ~(A).4. and codomain spaces with respect to which A then has a diagonal matrix representation.? (AA I min{m. Then T V AV =[ =[ VrAVI VIAVI VrAVI vIA VI Vr AVz vI AVz ] ~] since A V2 =0.denote A thought of as aalinear transformation mapping IR n to IRm. C denote A thought of as linear transformation mapping W to W.3.1 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. n] — Note that there are also min{m. of values of i I proof of A. The !:ingular value decomposition is not unique. m for IR (see the discussion in Section 3. of the proof of Theorem 5..5.. VT be an SVD of A as in Theorem 5..4. A.2. The latter equality follows from the orthogonality of the columns of VI and V 2. U V identical... and defining this matrix U\ andU UT A V [Q ~].2. A = t/E VT SVD of A 5.). 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). matrix VI E M mx/ " by U\ = AViS~l. an examination decomposition Remark 5. Then Specifically. Thus. Introduction to the Singular Value Decomposition Turning now to the eigenvalue equations corresponding to the eigenvalues or+\.2). 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 ). we see that Mat C is "i:. to be ~ completes the proof. 1. except for Hermitian transposes replacing transposes.16. . . AV2 = 0. See also m [v\. Referring to the equation U\ = A V\ S l defining VI. cr. From the proof of Theorem 5. D to be S completes the proof. Let A = V"i:. responding to multiple cr/'s... . The analogous complex case in which A E C™ x " is quite straightforward. VT as AV = V"i:. .. Remark 5. The set {ai.. we see that U{ AV\ = S and 1/2 A VI = U^UiS = 0. Choose any matrix V2 E ^ 77IX( ™~ r) such that [VI V2] is VI columns orthonormal. 0 Definition 5. The decomposition is A = V"i:. 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. singular 2. of A A).(A) At.'. Thus. we see that. Ui e (5. U be interpreted changes domain and codomain spaces with respect to which A then has a diagonal matrix representation. Remark 5..1. eigenvectors of AA TT). analogous complex e IC~ xn straightforward. V H. the IRmxr VI AViSI. an we have that ATAV2z = VzO = 0. 3. whence Vi ATAV22 = O. Specifically.36 36 Chapter 5. . . vn } for IR and {u\. .. Choose U2 £ IRmx(mr) [U\ U2] orthogonal.4) see UfU\ = columns of U\ are orthonormal. .5.e. Then T rewriting A = V"i:.. Remark 5. . n} . VTAV = [~ Q]. let C. Then from (5..(AATT). Introduction to the Singular Value Decomposition Chapter 5. •. in fact.. with respect to the bases A = U^V as A V Mat £ is S the U E we see respect n and {u I. Now V20 Vf A T A V 0. singular unique. The latter equality follows from the orthogonality of the S and vI AVi = vI VI S = O. ar}} is called the set of (nonzero) singular values of the matrix A and called [a\.. } for R" {VI. where V and V are unitary and the proof is essentially proof decomposition A = t/E V H. Note that V and V can be interpreted as changes of basis in both the domain Remark 5..1. .4) we see that VrVI = /.3..16.. . Remark 5.1 we see that ai(A) = A(2 (ATA) = £(A). Definition 5. Remark 5. we see that V r A VI = since A V2 = O..2).. Referring to the equation V I = A VI SI defining U\.. See also m Remark 5.
The Fundamental Theorem 5. 01  where U is an arbitrary 2x2 2 orthogonal matrix. VT AV = A > O.[1 0 ] . Note.1. too.9. SVD" (5. [11]. SVDof A. e. [7].3). Computing AA AATT is numerically poor in finiteprecision arithmetic. Then A = V A VTT is an A.5. see. VI:TU/ s n SVD of A VS C . [25]. 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).U I U T . [11].8. however.10.2) can always be constructed from ¥2 Theorem too. e j8 the case). U2.sine sin e cose J[~ ~J[ cose sine Sine] cose ' where e is arbitrary. [7]..g.6. The Fundamental Theorem 37 37 • any U22can be used so long as [U I U2] is orthogonal. What is unique. Vi. and E U\. AT A Remark 5.e. i. Let A e IRnxn" be symmetric and positive definite.. 0 Example 5. symmetric V orthogonal matrix of eigenvectors that diagonalizes A.11). F/vamnlp 5. C/ [U\ Ui] orthogonal.1. V2 (see Theorem 5. e.e.6. VI. is an SVD. VT A V = A = VAV eigenvectors > 0. For example. A=U is an SVD. if A = UI:VT is an SVD of A.9. then A A. that "full SVD" (5.g. corner f/E V T r r T Ti isaanS V D o f AT. see. [25].10. Example A . that aa"full SVD" (5. U arbitrary 2 x orthogonal 5. is the matrix I: and the span of the columns of UI. SVD of A. is an SVD. A _ [ 1  0 ~ ] cose = [ . 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).2) can always be constructed from a "compact SVD" (5. orthogonal transformations. Let V be an orthogonal 5. Computing an SVD by working directly with the eigenproblem for AT A or 5.. i.7.. Example 5.8.3).. f/2. Example 5. Better algorithms exist that work directly on A via a sequence of orthogonal transformations. 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.
6) and (5.. u r ].1.5) as a sum of outer products Remark 5. Let A E Rmxn have a singular value decomposition A = U'£ VT. 2...= N(A T ). . i=1 (5.. Let A E E mx " have a singular value decomposition A = U. .. reduction to row or column echelon form.6) (5. .1.. LetUI = [UI. Let V = [UI.1..= R(A T ). 4. The elegance of the dyadic decomposition (5. urn]. Part 4 of the above theorem provides a numerically superior method for Remark 5.. rank(A) = r = the number of nonzero singular values of A. for example.1. vr ]. nicely in Figure 5. Introduction to the Singular Value Decomposition 5. say.£VTT as in Let A e jRmxn have a singular value decomposition A = UHV as in Theorem Theorem 5. (5. . . the following properties hold: the notation of Theorem 5. The singular vectors satisfy the relations AVi = ajui. .1.12.5) as a sum of outer products and the key vector relations (5. urn] and V = [v\.. (c) R(VI) = N(A)1.7) AT Uj = aivi for i E r. as A = UZV rather than.13. The relationship to the four fundamental subspaces is summarized nicely in Figure 5. rank(A) = r = the number of nonzero singular values of A. VI = [VI. Then (a) R(VI) = R(A) = N(A T / . U2 = [Ur+I..14. . 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. um] and V = [VI.. Vn].7) explain why it is conventional to write the SVD as A = U'£VTT rather than. Introduction to the Singular Value Decomposition Chapter 5. Then (5. Note that each subspace requires on.8) where where . Let A e jRrnxn have a singular value decomposition A = VLV T Using Theorem 5.£V. .. . vn]. Using the notation of Theorem 5. Then TheoremS. andV2 = [Vr+I. . Let U =. .2 Some Basic Properties Some Basic Properties Theorem 5.11. . for example. A = UZV. Then A has the dyadic (or outer 2. 1. Remark 5.13. .12. [HI. . say. (b) R(U2) = R(A)1. (d) R(V2) = N(A) = R(AT)1.6) and (5. The elegance of the dyadic decomposition (5. the following properties hold: 1.. A = U. Theorem 5. Then A has the dyadic (or outer product) expansion product) expansion r A = Laiuiv.38 38 Chapter 5.. .. Note that each subspace requires knowledge of the rank r.2 5.11. The relationship to the four fundamental subspaces is summarized knowledge of the rank r. ..].7) explain why it is conventional to write the SVD and the key vector relations (5. reduction to row or column echelon form.. vn]. Remark 5. .5) 3. The singular vectors satisfy the relations 3.
1. 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. a simple reordering accomplishes the task: reordering accomplishes the task: (5. However.1. Proof' The proof follows easily by verifying the four Penrose conditions. Furthermore. SVD and the four fundamental subspaces.5... if we let the columns of U and V with the Qsubblocks appropriately sized.. e^.2. then = L r 1 v. Proof: The proof follows easily by verifying the four Penrose conditions.. (or exchange matrix) P = [e erI. which is clearly orthogonal and symmetric.u. Furthermore.. then be as defined in Theorem 5. . Some Basic Properties 5.15.11) 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.10) . ed. SVD and the four fundamental subspaces. Figure 5. Some Basic Properties 39 39 A r r E9 {O} / {O)<!l nr mr Figure 5.2. .er^\.11. e\\. if we let the columns of U and V be as defined in Theorem 5.15. e2..11. 0 D (5. Remark 5. . However. a simple if we insist that the singular values be ordered from largest to smallest. which is clearly orthogonal and symmetric.. with the Osubblocks appropriately sized. . Note that none of the expressions above quite qualifies as an SVD of A + Remark 5.=1 U.
Remark 5.. . since [u\. From Section 3. Introduction to the Singular Value Decomposition A+ = (VI p)(PS1 p)(PVr) is the matrix version of (5.. In other words. Notice that N(A) . the isabasisfor7£(. . A "full SVD" can be similarly constructed. In other words. A "full SVD" can be similarly constructed. / E~. = ^u. premultiplication of A by VT is an orthogonal transformation that "compresses" A by row transformations. let A E lR. u is clearly matrix representation for T with respect to the bases { v \ ..17 and Remark 5.. have an SVD given by (5. Such a row compression can also be accomplished by orthogonal row transformations performed directly on A to reduce it to the form 0 . .11). finiteprecision arithmetic.i e r.. by orthogonal row transformations performed directly on A to reduce it to the form [~].1)..r) and the matrix SVr e lR. . Then VT A = :EVT = [~ ~ ] [ ~i ] D _ . while the matrix representation for the inverse linear transformation T~ with respect to the same bases is SI. 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. Both compressions are analogous to the socalled rowreduced where R is upper triangular. . Then Again. postmultiplication of A by V is an orthogonal transformation that "compresses" A by column transformations...vvr}}is aa is r basisforN(A).mxn. then T can be defined by TVj = OjUj .. Since T is determined by its action on a basis. Such a compression is analogous to the . mxr has full column rank. then T can be defined by TV. Such a row compression can also be accomplished "compresses" A by row transformations. vrr} and {u I. Column compression Column compression Again.1). Then AV = V:E = [VI U2] [~ ~ ] =[VIS 0] ElR. r x has full row rank. Then Let A E lR.[ SVr ] 0 mxn E lR. Both compressions are analogous to the socalled rowreduced echelon form which.17 and in Definition 4. v } and {MI . .2. Recall the linear transformation T used in the proof of Theorem 3.i / E~. w.16. 5. Introduction to the Singular Value Decomposition Chapter 5...1).1.3 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 finiteprecision arithmetic. Recall the linear transformation T used in the proof of Theorem 3. Similarly.. In other words. .olumn transformations. .16. premultiplication of A by UT is an orthogonal transformation that rank. .2. Since T is determined by its action on a basis. is the matrix version of (5.M(UT Notice that M(A) = N(V T A) = N(svr> and the matrix SVf E Rrxll" has full row A/"(SV. This time. where R is upper triangular... .40 40 Then Then Chapters. From Section 3. = cr. Similarly. = tv. basis forJ\f(A)±.. when derived by a Gaussian elimination algorithm implemented in echelon form which. since {UI. urr]} is clearly S.. . is not generally as reliable a procedure.l. then TlI can be defined by T^'M. u is a basis forR(A).1). in Definition 4.. and since {VI..4). In other words.11).1. notice that R(A) R(A V) This time.urr}} e r. and since ( v \ .mxn have an SVD given by (5. . . is not generally as reliable a procedure..3 5. Such a compression is analogous to the "compresses" A by I. the same bases is 5""1. the matrix representation for T with respect to the bases {VI. let A e Rmxn have an SVD given by (5. . . then T~ canbedefinedbyTIu. while the matrix representation for the inverse linear transformation TlI with respect to S. . . postmultiplication of A by V is an orthogonal transformation column rank.
Use the SVD to determine a where Q is orthogonal and P p T > O.1 starting from the observation that AAT ~ O. Determine an SVD of A. [7]. see. 3.. [11]. EXERCISES EXERCISES 1. for performed by Gauss transformations in finiteprecision arithmetic.e. z of complex scalar z (where i j J=I). A = Q P 7. Let A e E"xn be symmetric but indefinite. = o. For details. [7]. 4.. Let A e Rmxn and suppose W eRmxm and Y e Rnxn are orthogonal. For details. of defined by A defined by A = xyT. € IRmxn. see. If XT X = 2. Let E ~~xn. xyT 5. . i. which is not generally a reliable procedure when performed by Gauss transformations in finiteprecision arithmetic.1 starting from the observation that AAT > 0. [25]. y E ~n Determine A e ~~ 4. Let X E M mx ". Prove Theorem 5. Use the SVD to determine a polar factorization of A. A E IRnxn indefinite. [23].Exercises Exercises 41 41 socalled columnreduced echelon form. 2. y e Rn be nonzero vectors. which is not generally a reliable procedure when socalled columnreduced echelon form. 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). 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. If XTX = 0. an SVD A. Determine SVDs of the matrices 5. (a) Show that and W A F have the same singular values (and hence the same rank). [25]. Note: this is analogous to the polar form where Q is orthogonal and P = PT > 0. Determine SVDs of the matrices (a) (b) [ ] [ ~l 1 0 1 6.. Let x e Rm. Note: this is analogous to the polar form iO z = rel&ofaa complex scalar z (where i = j = V^T). [23].[11]. for example. (b) Suppose that W and Y are nonsingular but not necessarily orthogonal. show that X = 0. A = QP 7.e. Let A € R" X M . Prove Theorem 5. Determine an SVD of the matrix A E R™ xn E IRm.
This page intentionally left blank This page intentionally left blank .
There exists a unique solution to (6.e. i. b E lRm.e.3) is unique if and only ifJ\f(A) = 0. 5. There exists at most one solution to (6. and this is possible only ifm ::: n. there exists a solution if and only j/"rank([A. (6.. There exists a solution to (6.e. i. n. and onto..3) 1.. equivalently. i. b]) = rank(A). and this is possible only ifm :::: n (since m = dim R(A) = rank(A) :::: min{m. A solution to (6. (6. n this is possible only ifm < n (since m dimT^(A) = rank(A) < min{m. there exists a solution if and only ifrank([A.3) for all b E ]Rm if and only if A is nonsingular.3) if and only ififbeH(A). There exists a solution to (6. b]) = rank(A). equivalently.3) for all b E lRm if and only ifR(A) = lRm.. as a special case. We begin with a review of some of the principal results associated with vector linear systems. A is 11.1. There exists a solution to (6. N(A) 0. 43 . the familiar vector system Ax = b.3) if and only b E R(A). A G M m x m and A has neither a singular value nor a eigenvalue.1 6.3} for e R m if only ifU(A) = W". General linear systems of the form equations. A solution to (6. Consider the system of linear equations Ax = b. (6.e.3) for all b e W" if and only if is nonsingular. 3. There exists a nontrivial solution to the homogeneous system Ax = 0 if and only if Ax = 0 if 6.1) are studied and include. the familiar vector system are studied and include. rank(A) < n.2) 6. 3. Theorem 6. 4. A E ]Rn xn. General linear systems of the form (6. A/"(A) = 0. onto. A is 11.. 4. only if rank(A) < n.3) for all b E lRm if and only if the columns of 5. b E ]Rn. 1. A is 2. i.1 Vector Linear Equations Vector Linear Equations We begin with a review of some of the principal results associated with vector linear systems. equivalently.3) is unique if and only if N(A) = 0. There exists at most one solution to (6.3) for all b e W1 if and only if the columns of A are linearly independent. A are linearly independent. A E lRmxn.e.e.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. n}). A E lR mxm and A has neither a 0 singular value nor a 0 eigenvalue. as a special case. Consider the system of linear equations Theorem 6.1.. There exists a unique to (6. i. i. 6. 2. equivalently. and this is possible only ifm > n.
1).nxk is arbitrary. Let A e Rmxn. Note that some parts of the theorem follow directly from others. equivalently. i.. +B = Theorem 6. 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.2 6. Furthermore.18. of a matrix.e.6) Furthermore. AZ — B. Let Z be an arbitrary solution of That all solutions arc of this form can be seen as follows. AA+B B. Then any matrix eRmxk of the form of the form X = A+ B + (/ . Note that the results of Theorem 6. The matrix linear equation Theorem 6. Let Z be an arbitrary solution of (6.2) follow by specializing even further to the case m = n. BE JR.1 follow from those below for the special case k = 1.6) are of this form.e.3. we prove part 6. B E JR. Note that the results of Theorem of solutions to the general matrix linear system (6. The matrix linear equation AX = B. note that x 0 is always a solution to the homogeneous system.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 (6.2 (Existence).mxk. A is not II. (6. all solutions of (6. to prove part 6.1 follow from those below for the special case = 1.mxk and suppose that AA+B = B. i.6).. Proof: To verify that (6. 0 . and this is clearly of the form (6. equivalently..44 Chapter 6. R(A).6) are of this form. E JR. to algebra. which implies rank(A) < n by part 0 by part 3. premultiply by A: Proof: To verify that (6. Note that some parts of the theorem follow directly from others. i. we must have the case of a nonunique solution. Therefore.5) is a solution. while results for (6. The matrix criterion is Theorem 4. 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.5) is a solution. a solution exists if and only if has a solution if and only ifR(B) S. i. (6. For example.1).4) has a solution if and only ifl^(B) C 7£(A). while results for (6..2)follow by 6.5). AZ :::: B. 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. Then we can write Z=A+AZ+(IA+A)Z =A+B+(IA+A)Z and this is clearly of the form (6.3. specializing even further to the case m = n. 0 Theorem 6. a solution exists if and only if AA+B = B. note that x = 0 is always a solution to the homogeneous system. where Y E JR. (6. A is not 11. A E JR.A+ A)Y. For example. The matrix criterion is Theorem 4. Then we can write (6. Theorem 6. Linear Equations Chapter 6.e. mxn .e .mxn.5). all solutions of (6. That all solutions are of this form can be seen as follows. which implies rank(A) < n must have the case of a nonunique solution.5) is a solution of is a solution of AX=B. D 6.6).18.2 (Existence).
A+ A)Y =A++(IA+A)Y.A+A).6 (Uniqueness).) that minimizes TrXT X. recall that TrXT X = £\ •xlj. y E R" A + A t= I.6. in which case A must be invertible and R = AI. BE lR. Clearly. then it is easily R(I — A + A).A+ A V2 V[ and U(V = K(V2) = N(A).4. / .mxn. 7£(A) and this is 7£(/m) c R(A) equivalent to AA + 1m Im. There is a unique right inverse if and only if A+A = I/ e E"xm arbitrary. A E lR.7) is unique if and only if A+A = /.6) +B Remark 6.5.) Theorem 6. It particular (6. leaving only the unique solution X = AI1B.2.S. Clearly.nxn. A+ = A"1 and so (I . we write 1m to emphasize the m x m identity Im matrix. A A = f/E VT. (6.3. Matrix Linear Equations 45 Remark 6. Butrank(A) = n if and only if A is 11 or N(A) = 0. A solution of the matrix linear equation AX = B. Clearly. A R (AA(A) = A"1. A A+ AI Remark (/ — A + A) 0. wherer = rank(A) (recallr ::: h).j jcj. nonzero solution. A e E"x". Computation: Since y is arbitrary. Consider the system of linear firstorder difference equations (6. Example 6.A+A = Vz V2 and R(Vz2V^) = R(Vz) = N(A).8. When A is square and nonsingular.2. and (N(A) = 0). D 0 Example 6. A+ A where Y E lR.9. rank(A) = < A This is equivalent to either rank (A) = r < n or A being singular.mxn.7) has a unique solution if and only if unique if and only if A + A = I.nxm is arbitrary. (6.5.. Matrix Linear Equations 6. if and only if A is Ilor _/V(A) = O. Remark 6. Proof: Proof: The first equivalence is immediate from Theorem 6.9. Here. Characterize AR = Im solutions R of the equation AR = 1m.A+A) = O. find all A e ]Rmx". Solution: x=A+O+(IA+A)y = (IA+A)y.7. But if A has an SVD given by A = U h VT. this can occur if and only if rank(A) = r m (since equivalent to AA+Im = 1m. A solution of the matrix linear equation Theorem 6.mxk (6. it is not unique. matrix.6 (Uniqueness). where r rank(A) (recall r < n). equivalently. Suppose A E lR. But rank(A) = n that A+ A = / if r — n. Solution: There exists a right inverse if and only if R(Im) S. vD Example 6.7. (TrO denotes the trace of a matrix. Example 6. if there exists a unique.n is arbitrary. Example 6. Thus. Consider Example 6. X• = A~ B. there exists a nonzero solution if and only if A+A /= I. 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). Hence.8) . N(A) = O. equivalently. It can be shown that the particular solution X = A+B is the solution of (6. it is easy to see that all solutions are generated y from a basis for 7£(7 . The second follows by noting thatA+A = I can occur only ifr = n. equivalently.7) has a unique solution if and only if M(A) = 0. r checked that 1. where y e lR. Ax — 0. recall that TrX r = Li. there is no "arbitrary" component. Characterize all right inverses of a matrix A E lR. Find all solutions of the homogeneous system Ax = 0. All right inverses r < m) A (A+ of A are then of the form of A R = A+ 1m + (In .6) that minimizes TrX7 (Tr() denotes the trace of a matrix.
8) of Example 6. this is a question {u }y~Q Xk of reacbability.. The above are standard conditions with analogues for continuoustime models (i. AB. We now introduce an output vector yk to the system (6. The matrices A = [ ° Q and f ^ provide an lability and reachability are equivalent. does there exist an input sequence for k > 1.10. A related question is the following: Given an arbitrary initial vector Xo. linear differential equations).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. from the fundamental Existence Theorem. The linear differential equations). The general solution of (6.8) of Example 6. Theorem l'/:b Clearly. Example 6... controlA 1 lability and reachability are equivalent. we of reachability. There are many other algebraically equivalent conditions. B T] is observable [reconsrrucrible] [controllablcl if and T) observable [reconstructive].8) is given by solution of (6. does there exA related question is the following: Given an arbitrary initial vector XQ. Since m ~ I. The vector Jt* in linear system theory is e IR nx " fieR" (n ~ I.2.10) for k ~ 1. (A. we have the notion of reconstructibility: When does knowledge of {u jy }"~Q and {. if A is nonsingular.. if and only if rank [B. We can then pose some new questions about the overall system that are dual in the systemtheoretic sense to reachability and controllability. We might now ask the question: Given XQ = 0.9 by appending the equation by appending the equation (6.8) is reachable if and only if if R([ B. does there exist an input sequence {u j an input sequence {"y}"~o such that xn = O? In linear system theory. if and only if or. We now introduce an output vector Yk to the system (6. equivalently. 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. this is called controllability.9) ~Axo+[B. There are many other algebraically equivalent conditions. We might now ask the question: Given Xo 0. we see that (6. standard conditions with analogues for continuoustime models (i.10. Again from Theorem 6. does there exist an input sequence {ujj 1jj^ such that Xk takes an arbitrary value in 1R"? In linear system theory. A n . The matrices A = [~ ~]1and B5 == [~] 1 providean example of a system that is controllable but not reachable.T.ra>l). from the fundamental Existence Theorem.~ I). ..J B]) = 1R" or. A n .. We can then pose some new questions about the with C and (p > 1). equivalently.8) is given by kJ Xk = Akxo + LAkJj BUj j=O UkJ ] Uk2 (6. B) iJ reachable [controllable] ifand only if (A . . The general known as the state vector at time k while Uk is the input (control) vector. m known as the state vector at time while Uk is the input (control) vector.: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. .11) with and D (p ~ 1).2.8) is controllable if and only if if controllability. Since > 1. reachability always implies controllability and. AB. we have the notion of suffice to determine (uniquely) xo? As a dual to controllability.e. B) is if(AT .AB •.. example of a system that is controllable but not reachable. this is called such that Xn = 0? linear system theory. The condition 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 7 r/:b and {Yj l'.• A k kJ B] [ ~o (6.e.. Linear Equations Xk with A E R"xn and B E IR nxmxm(rc>l.9 Example 6.J B] = n. see that (6. . Theorem 6. this is a question va [Uj }k~:b such that x^ takes an arbitrary value in W ? In linear system theory.46 46 Equations Chapter 6.
C E jRmxn. the coefficient matrices of interest are square and nonsingular. Invertibility is assumed for any component or subblock whose inverse is indicated. so a solution exists.Duo Yl . e Tl(R).15) E jRnxp where Y € Rn*p is arbitrary.27. +L kl CAk1j BUj + DUk. By the fundamental the righthand side. Listed In many applications.DUnl 6. 0. Theorem 6. if and only if or.14) Theorem 6. B E Rmxq. Then. by definition. notice that Yk = CAkxo Thus. e Rmxn. particularly for block matrices.3 6. equivalently. the coefficient matrices of interest are square and nonsingular. In these identities. A compact matrix criterion for uniqueness of solutions to (6. equivalently. By the fundamental Uniqueness Theorem.6. and C E Rpxti. or. the solution is then unique if and only if N(R) Uniqueness Theorem. B E Rnxm. v E R(R).12) j=O Yo . Theorem 6. particularly for block matrices.14) requires the notion A compact matrix criterion for uniqueness of solutions to (6. Then the equation e jRmxn. asbelow is a small collection of useful matrix identities. (6.13) and let R denote the matrix on Let denote the (known) vector on the lefthand side of (6. associated with matrix inverses. in which case the general solution is of the has a solution if and only if AA + BC+C = B. Such a criterion (CC+ <g) A+ A — I) is stated and proved in Theorem 13.13) and let denote the matrix on the righthand side.6. notice that To derive a condition for observability. so a solution exists.4 Some Useful and Interesting Inverses 6. E jRnxm.4 6. Let A E Rmxn.. mxm and D E jRm Invertibility is assumed for any component or subblock whose inverse is and D € E xm.Du] (6. A E Rnxn.11. .2 j BUj .4 Some Useful and Interesting Inverses 47 To derive a condition for observability.3 A More General Matrix Linear Equation A More General Matrix Linear Equation AXC=B (6.14) requires the notion of the Kronecker product of matrices for its statement. Listed below is a small collection of useful matrix identities.13) Let v denote the (known) vector on the lefthand side of (6.4 Some Useful and Interesting Inverses Some Useful and Interesting Inverses In many applications. Verification of each identity is recommended as an exercise for the reader. and C e jRpxq. B e jRmx q . by definition. the has a solution if and only if AA+BC+C = B.27. if and only if r Yn]  Lj:~ CA n . arbitrary. Such a criterion (C C+ ® A +A = I) of the Kronecker product of matrices for its statement. sociated e jRnxn. is stated and proved in Theorem 13. Verification of each identity is recommended as an exercise for the reader.CBuo . indicated. Then. in which case the general solution is of the form (6. Thus. 6. the solution is then unique if and only if N(R) ==0.6.
where E = (D . 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 Chapter 6.I EXERCISES EXERCISES 1. for example. theory. Let A € E mx ". As in Example 6. 1. BB EelR fflxk and suppose AAhas an SVD as in Theorem 5. (A BDCr1 = AI . formulas for the inverse of a sum of matrices such as (A + D)lor (AI1 + DI)I.B D. = = 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. It has many applications (and is frequently "rediscovered") including. 1. l = l = [!C / [~ ~ l = [ AI +_~~!~CAI A~BE = D. mx .48 Chapter 6. Note that the positions of the / and — / blocks may be exchanged. characterize all solutions of the matrix linear equation AX=B in terms of the SVD of A in terms of the SVD of A. characterize all left inverses of a matrix A e lR ".I C) I.1. This where E = (D — CA B) (E is the inverse of the Schur complement of A). 2. It has many 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 . result follows easily from the block LU factorization in property 16 of Section 1..8. [ / +c 7. (A + BDC)I = A~l .BDI l = [ AI BD. r A~I [~ ~ r [D~I~AI D~I 1 ~r ~~B 1 r l [~ ~ r [D~CF +~~I~. where F = (A . [~ ~ r l 3. Rmxk and suppose has an SVD as in Theorem 5.8. for example.4. Linear Equations 1.. characterize all left inverses of a matrix A E Mm xn . As in Example 6. characterize all solutions of the matrix linear equation 7Z(B) c 7£(A).I B)I (E is the inverse of the Schur complement of A).1. Assuming 2./ blocks may be exchanged.AIB(DlI + CAIB)ICAI. 5.4. Note that the positions of the / and .c E E") that arise in optimization (A + xx T ) — (with symmetric A e lRnxn and x e lRn) that arise in optimization theory.I .. This result follows easily from the block LU factorization in property 16 of Section 1.I ] D. Assuming R(B) ~ R(A). ization in property 17 of Section 1. Let A E lRmxn.4. X. 2.CA. that X~l [~ !/ [~ ~ r [~ ~ l [~ ~/ r [~ ~ 1 l l l = [ ~ 4. l 8.A~lB(D~ CA~lB)~[CA~l This result is known as the ShermanMorrisonWoodbury formula. BC 6. formulas for applications (and is frequently "rediscovered") including.4.
e.Exercises Exercises 3.. y E E" and suppose further that XTy ^ 1. c and individual elements y. Show that 3.. where C = 1/(1 .. Show that the matrix B — A — —eie T : (i. . check directly that the condition for reconstructibility takes the 6. Show that the matrix B = A . Show that (/ . As in Example 6. Hint: Show that ct <= M(B). A with yl subtracted from its (ij)th element) is singular.10.10.e. Show that cxJ C ' where c 1/(1 — T y).. Assume that x/( 7^ 0 for some and j. A with — subtracted from its (zy)th element) is singular. ...y Assume that Yji i= 0 for some i/ and j.. Let A E 1R~ " and let A 1 have columns Cl. . T 4. check directly condition for reconstructibility the form form N[ fA J CA n 1 ~ N(A n ). y e IRn and suppose further that x T y ^ 1. l' Hint: Show that Ci E N(B).Cn and individual elements Yij.l ~i e. y E E" and suppose further that XTy i= 1. (i.x xTy). Let jc. Let A e R"xxn and let A"1 have columns c\. 6.e. € IRn and suppose that x T y i= 1. Let x. in Example 6. .. Show that 4. Let x.xy) T 1 49 = I  1 xTy 1 xy . 5.
This page intentionally left blank This page intentionally left blank .
Theorem 7. Define pX. Let V be a vector space with V = X EEl Y. P2 = P.3. V by by PX. Px. Infact.yp2 = P.y is called the (oblique) projection on X along 3^.Chapter 7 Chapter 7 Projections. P isaprojectionifandonlyifl P isaprojection..1 7.1). A linear transformation P is a projection if and only if it is idempotent.1. i.x — I — Px.e. every v E V Definition 7. Figure 7.2.1 displays the projection of von both and 3^ in the case = Figure 7. and Norms Spaces..1.3.26. i.1. Also.y. P is a projection if and only if I —P is a projection.yV = x for all v E V.y • V —>• c V has a unique decomposition v = x + y with x e X and y E y. Let V be a vector space with V X 0 y. By Theorem 2.y Theorem 7. Proof: Suppose P is a projection. Py. Oblique projections.26.y is called the (oblique) projection on X along y.x = I px. y = Px. say on X along y (using the notation of Definition 7. px. A linear transformation P is a projection if and only if it is idempotent. Proof: Suppose P is a projection. Oblique projections.1. Infact.1). y x Figure 7.y.e. every v e V has a unique decomposition v x y with x E and y e y. and Norms 7. Px.y is linear and P# y — px.2. Inner Product Projections. 51 51 .1 displays the projection of v on both X and Y in the case V = ]R2. Theorem 7. PX. Define PX y : V + X <. Theorem 7. Figure 7.y is linear and pl. Py.1 Projections Definition 7. Also. say on X along Y (using the notation of Definition 7. Inner Product Spaces. By Theorem 2.
P) = 0. we have (py)T ((I . we must have pT (I .P2v = 0 so Y E y.. Let x = Pv.1 . First note that v E X. Then v if v € Pv (I . say. T Since x and y were arbitrary.. (I . D Essentially the same argument shows that I — P is the projection on y along X. We now prove that V = X $ y.P)v. suppose P = P.xx by Theorem 7. mental subspaces.P)x = xTP(I . PX. i=r+l PN(A) 1. Conversely. Thus.. let A E Rmxn with SVD A = U!:VTT = A = UT. we have ( P y f I (/ .4. Conversely. p2 = P.px. Then symmetric projection matrix and let x be arbitrary.3. 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. V = X $ Y and the projection on X along Y is P. since Px e U(P).3. then Pv = v. Write x = Px (I — P)x. then v = O. Thus. while Py = P(l . and P must be an orthogonal projection. Px E R(P). y = (I . suppose p2 = P.V Theorems 5. V Pv .P)x = x T P(l .P)x E XL. then Pv = 0. Then U\SVr Then r PR(A) AA+ U\U[ Lu. along XXL} and let x. Then Px = p 2v = Pv = x so x e X.P)x = (I . R" be Proof: Let P be an orthogonal projection (on X. It is easy to check that X and Y are subspaces. Projections. 0 Definition 7.P is the projection on Y along X. Hence that V X 0 y. First note that iftfveX. Then Px = P2v = Pv = x so x E X.P)x.xl. while Py = P(I P}v = x Pv . Note that (I . Then Pv = P(x + y) = Px = x. .xJ. Since Py E X. along 1) and let jc. arbitrary.5. Write x = P x + (I . PN(A)J. * called an orthogonal projecDefinition 7. Projections. Essentially the same argument shows that / .4.P)v.)x = PXJ. Moreover.P)x = (I . A 6 jRmxII UtSVf. Note that (/ .P)x e X1. It is easy to check that X and 3^ are subspaces. P = P. with the second equality following since PTP is symmetric. then Pv = O. Then Pv = P(x + y) = Px = x. and Norms Let v e V be arbitrary. Hence if v E X ny. If v e y. P E E"xn is the matrix of an orthogonal projection (onto K(P)} if and only 7.1 and 5. Let X n y. Thus. Let X = {v E V : Pv = v} Px = x = Pv. Inner Product Spaces.11. Conversely. . Then x T pT (I .XLtion and we then use the notation P x = PX.A+A V2V{ L i=r+l i=l 11 ViVf.P}x = 0. Py e X.5. Inner Product Spaces. yy Ee jR" be arbitrary.=1 m PR(A). then (/ . and Norms Chapter 7.P)v.X^X = Px±. In the special case where y X1. suppose symmetric projection matrix and let x be arbitrary. Now let v E V be arbitrary. Now let u e V be arbitrary. Conversely.p 2 v 0 so y e Thus.xx by Theorem 7.52 52 Chapter 7.P)x = O. If v E Y. P Proof: Let P be an orthogonal projection (on X.AA+ U2 U ! LUiUT.PX. px. Since x and y were arbitrary. Thus.P)x 6 R(P)1and P must be an orthogonal projection.L 1. D 0 7. Hence PT = PTP = P.P)x E ft(P)1 xTPT(I . X 0 y and the projection on X along y is P. Thus. Thus. then v = 0.P)v.1. (I . Hence pT = pT P = P. In the special case where Y = X^. P)x = y PT(I P)x = 0.1 The four fundamental orthogonal projections The four fundamental orthogonal projections Using the notation of Theorems 5.P)x = O. say.11. Let X = {v e V : Pv = v} and y {v € V : Pv 0}. P e jRnxn is the matrix of an orthogonal projection (onto R(P)) if and only ifP2 PT if p2 = p = pT..uT. We now prove and Y = {v E V : Pv = OJ.XL iss called an orthogonal projection and we then use the notation PX = PX. y = (I . suppose P is a is a with the second equality following since pT P is symmetric. we must have P (I — P) = O.1 7. Moreover. then Pv v. p 2v = PPv = Let u E V be arbitrary. P2v = P Pv — 2 2 Px = x = Pv.P)x = yTTpT (I .P)v = = Pv.XL Theorem 7. Then v = Pv + (I .1 5.
.6. Recall the diagram of the four fundamental subspaces. See Figure 7. the vector z that is orthogonal to w and such that Pv Moreover.~) w.7. See Figure 7. are. e Rn Solution: Think of the vector w as an element of the onedimensional subspace IZ(w). { v \ . There. Projections 53 Example 7.2.Pn(w»v = v . . orthogonal: that z and u.11.2. Then the desired projection is simply Then the desired projection is simply Pn(w)v = ww+v wwTv (using Example 4.2. The expression for x\ is simply the orthogonal projection of XI projection of rather x on S.1.8) (using Example 4. in fact." Example 7.1. Orthogonal projection on a "line." Figure 7. T W W Moreover. The indicated direct sum decompositions of the domain E" and codomain IRm are given easily as follows.2. . Vk} was an orthomormal basis for a subset S of W1. Vk} was an orthornormal Example 7.7. {VI.7.A+ A)x 2 = A+ Ax + (I = VI vt x + V Vi x (recall VVT = I)... An arbitrary vector x E R" was chosen and a formula for XI basis for a subset of IRn. IR n Rm 1 n Let X E IR be an arbitrary vector.8) = (WTV) W.8. orthogonal: v z Pv w Figure 7. . Then X = PN(A)u + PN(A)X .. The indicated direct Example 7. Example 7. Then Let x e W be an arbitrary vector. Projections 7. Recall the proof of Theorem 3. Recall the diagram of the four fundamental subspaces. X on Specifically. Orthogonal projection on a "line. Recall the proof of Theorem 3. 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. There. ..8. Determine the orthogonal projection of a vector v E IR n on another nonzero vector w E IRn. in fact..(:. A direct calculation shows that and ware. Solution: Think of the vector w as an element of the onedimensional subspace R( w). Determine the orthogonal projection of a vector e M" on another nonzero Example 7. Specifically.6.11. A direct calculation shows z = Pn(w)"' = (l . An arbitrary vector x e IRn was chosen and a formula for x\ appeared rather mysteriously.
Then Similarly. V = IRn.12. {*. and Norms Similarly. Yl. j2 ^ V and for alia. y) x T Y is the "usual" Euclidean inner product or Example 7. Inner Product Spaces. Then { • • V x V if product if 1. Projections. Example 7. Let V be a vector space over R. Y2 E V and/or all a. as follows: and a vector in J\f(A).y E V. y) for all x € Rm and for all y e R".12. e R. Let V = E". Inner Product Spaces.10. x) ::: Qfor aU x 6V and (x. such that {x. If A E Rm xn. Let V = R".54 Chapter 7. y) = (y.13. then AT e Rn xm is the unique linear transformation or map T E IRm andfor IRn.(A . x) for all x. Then ('. y)Q = XT Qy. f3ftE IR. respectively. aYI + PY2) = a(x. Then Y = PR(A)Y + PR(A)~Y = AA+y + ( l . cryi + ^2) = a(x. (x.10. let y e IR m be an arbitrary vector. Let Example 7. only ifx = O. 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 . where Q = QT > 0 is an arbitrary Q = Q T > is an Example 7. Then {^. (x.11. ATE IR nxm transformation Definition 7. y) = (y. . Example 7. y\) + /3(jt. respectively. (x. y e V.) ) :: V x V + IR is a real inner is a real inner Definition 7.. > Ofor all E V ( x x) =0 if 2.AA+)y = U1Ur y + U2U[ Y (recall UU T = I).9. Yl) + f3(x. .9. yi. 2. . and Norms Chapter 7. [ 5~2 + 7. (x. n x n positive definite matrix.13.2 Inner Product Inner Product Spaces Definition 7. definite defines Definition 7. Then (x. (jc.11. Y2) for all x. Let V be a vector space over IR. y^} for all jc. Then (x. defines a "weighted" inner product. If e IR mx ". 3. 3.x)forallx. Projections. Ay) = {AT x. (x. Example 7. let Y E ]Rm be an arbitrary vector. Let V = IRn. 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). y) Q = X T Qy. y} = XTy is the "usual" Euclidean inner product or dot product. x } = 0 if and only ifx = 0.
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
Let x. [20.33. JC.t~JI On the vector space ((C[to. we see immediately that IXH y\ ~ 0 < ( x H ) ( y H y ) — ( x H ) ( y H x ) . 1Ft). if U € C"x" is unitary. In other words. 11·111 and 1I·IIClO XHUHUx . and Norms Chapter 7. Then Theorem 7. (b) IIx IIz.g = (x QXY denoted  • c).58 58 Chapter 7. y E en. Since yH = x H y. define the vector norm 1111100 = max II/(t) 11 00 . The angle e between two nonzero vectors x. Since Proof: Consider the matrix [x y] e en x2 . i. Inner Product Spaces. Ther. Remark 7.~~1~1112'. ttl. (CBS) inequality (see.32 are true for general inner product spaces.. On the vector space (C[to.g. > 0. „ . 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.31 (CauchyBunyakovskySchwarz Inequality). e.30 (HOlder Inequality). However. However. Projections. A particular case of the Holder inequality is of special interest. 1Ft). IIQ)' 1 3.l. o ~ (x Hxx)(yH y) . with equality if and only if x and y are linearly dependent. it is particularly easy to remember. p.> where Q = QH > 0 (this norm is more commonly = QH > Ikllz. denoted II .D = E^rf/l*/!. Since yHxx = x Hy.34. p q I I A particular case of the HOlder inequality is of special interest. 1cose 1~ 1. y e en. whered.30 (Holder Inequality). t\])n. The angle 0 between two nonzero vectors x. its determinant must be nonnegative. Then Fhcorem 7. Then with equality if and only if x and yare linearly dependent.  . Since is a nonnegative definite matrix.Q — (xhH Qx) 2. is a nonnegative definite matrix. 217]). y E C". define the vector norm 3. define the vector norm On the vector space «e[to.. t \ ] R). e. However. The CBS inequality is thus equivalent to the statement I.^ cos e = IlMmlylb 0 ~ 0 < I' The CBS inequality is thus equivalent to the statement ~ ^  COS 0 < 1. then Remark 7. Let x. 0 < e — 5. The norm  • 2 is unitarily invariant. \\Ux\\l XHX = \\x\\\). Inner Product Spaces.31 and Remark 7. Let x. The norm II . its determinant must be nonnegative. we see immediately that \XH yl < IIxll2l1yllz.. Remark 7. ttlr.32. Theorem 7. y e C".32 are true for general inner product spaces.e. Projections.. then H H \\Ux\\2 \\x\\2 (Proof.D = L~=ld. Proof' Consider the matrix [x y] E C" x2 . y e C" may be defined by Remark 7. Some weighted pnorms: (a) IIxll1. where 4 > O. +=1. Some weighted pnorms: 2. Theorem 7. R). and  . Remark 7.e.(x Hyy)(yH x). tO~t:5. y E en may be defined by cos# = 1I. Remark 7. On the vector space (C[to.31 (CauchyBunyakovskySchwarz Inequality).tl Theorem 7. 112 is unitarily invariant..31 and Remark 7. Theorem 7.g. [20.32. 217]).33.34.lx.1^ IIUxll2 = IIxll2 (Proof IIUxili = x U Ux = xHx = IIxlli)· However. In other words. it is particularly easy to remember. i.. p. define the vector norm 11111 = max 1/(t)I· to:::. Let x.. if U E enxn is unitary. and Norms 2.
there exist constants c\.38. the following inequalities are all tight bounds. there exist constants CI. As with vectors.. For x G C". Theorem 7. IIxl12 :::: Jn Ilxll oo .e. convergence in terms of vector norms. As with vectors. and essentially obvious.e. 2.. i» (1) v(2\ . we conclude this section with a theorem about convergence of vectors. Then 7. If x. i. Matrix Norms 59 59 are not unitarily invariant. Theorem 7.7. (As with vectors. we conclude this section with a theorem about convergence of vectors..e. the motivation for In this section we introduce the concept of matrix norm.. 2 the proof of which follows easily from z2 _ z_//. this is called the triangle inequality.37.4 Matrix Norms Matrix Norms In this section we introduce the concept of matrix norm.) .4. convergence in terms of vector norms. C2 (possibly 7. the motivation for using matrix norms is to have a notion of either the size of or the nearness of matrices. about convergence of real numbers.. Extension to the complex case is straightforward what arises in the majority of applications. Similar remarks apply to the unitary invariance of norms of real are not unitarily invariant. Then lim k4+00 V(k) = v if and only if lim k~+oo II v(k)  v II = O. The using matrix norms is to have a notion of either the size of or the nearness of matrices. Remark 7. Attention is confined to the vector space (IRmnxn. i. E en.. there exist vectors x for which equality holds: vectors x for which equality holds: Ilxlll :::: Jn Ilxlb Ilxll2:::: IIxll» IIxlloo :::: IIxll» Ilxlll :::: n IIxlloo. while the latter is needed to make sense of "convergence" of matrices.. Let \\ \\ be a vector norm and suppose v. The former notion is useful for perturbation analysis. BE IRmxn.   R mx " ~ E is a matrix norm if it satisfies the following three Definition 7. Finally. v(2)..37. vectors under orthogonal transformation.4.35. there exist Example 7. Let II·• II be a vector norm and suppose v. All norms on en are equivalent. the following inequalities are all tight bounds. 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. Matrix Norms 7.4 7. Extension to the complex case is straightforward and essentially obvious.e.. y E en are orthogonal. Similar remarks apply to the unitary invariance of norms of real vectors under orthogonal transformation.. i. ConFinally. IIAII ~ Ofor all A E IR mxn and IR IIAII = 0 if and only if A = O. Definition 7. II·• II : IR mxn > IR is a matrix norm if it satisfies the following three properties: properties: 1.38.36. i. i.. .39. the proof of which follows easily from liz II~ = ZH z. IIxlioo :::: IIxllz. while the latter is needed to make sense of former notion is useful for perturbation analysis. v(l). then we have the Pythagorean Identity Remark 7. i.e.39. e C". For x E en. ci (possibly depending onn) such that depending on n) such that Example 7. All norms on C" are equivalent. IR) since that is "convergence" of matrices.36. IIA + BII :::: IIAII + IIBII for all A. then we have the Pythagorean Identity Ilx ± YII~ = IIxll~ + IIYII~. If y € C" are orthogonal. 7. 3.e.35. Attention is confined to the vector space (W xn R) since that is what arises in the majority of applications.. lIaAl1 = lalliAliforall A E mxn andfor all a E IR.
<110#0 IIxllq Example 7. and Norms Example 7. e R mx ". matrix = Ilxllp. Example 7. Example 7. and Norms Chapter 7.44.00 =  • 2. where r mxn = rank(A). The "maximum column sum" norm is 2. Then the Frobenius norm (or matrix Euclidean norm) is 7. The following three special cases are important because they are "computable. Schatten/7norms IIAlls.mxn. For example." theorem and requires a proof. I. + a!)"".p = (at' + .  . ^wncic = rank(A))." Each is a "computable. (A' A)) 1 ~ (T.60 Chapter 7._ Then "mixed" norms can also be defined by e lR. (where r = laiiK^/i. IIAII P t altA)) 1 ~ (T.44. Projections. to estimate the size of a matrix product A B in terms of the sizes of A and B individually. I Some special cases of Schatten /?norms are equal to norms defined previously..jj laij. 11·115.  5 2 = II IIF and  • 5i00 = II . The norm  .2 =  . Let A E lR. Example 7.40. The "matrix analogue of the vector 1norm. Let A E K m x "..42.41.60 max _P IIAxll = max Ilxli p IIxllp=1 IIAxll p . The "maximum row sum" norm is 2. Projections. The "matrix analogue of the vector Inorm. Example 7." IIAliss = Li. tTL T Note: IIA+llz = l/ar(A). Inner Product Spaces. The Schattenpnorms are defined by E lR.mxn. \\F and 11'115. Example 7. 5>1 is often called the trace norm.40.43. IIAII2 = Amax(A A) = A~ax(AA ) = a1(A)..) I ~ (t. Then the matrix pnorms are defined by A e Rmxn. (AA ')). 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. 1. J=1 3. is a norm. is a norm. The spectral norm is 3.mxn IIAII p.. ai. I.43. Let A E R . defined by IIAIIF ~ (t.42.q = max IIAxil p 11. pnorms previously. IIAlioo = max rE!!l. (t laUI). Example 7. Let A E lR. 112' The norm II • 115.mxn. Inner Product Spaces. .1 is often called the trace norm."  A\\ = ^ \ai} .
atornorms. We thus need the following definition.jii IIAII I.Then II Ax 1122 ::s II AFjc2. \\v consistent with it. \\ are Definition 7. the IIIn II F = .  • /7and II . there exist matrices A for i. e. IIAII2 ::s . II such that AF is given by max^o ".::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. IIAIII ::s .e. •II p for all p are consistent matrix norms. p for all p are consistent matrix norms. 2. consistent with it.oo 2 while IIAIII. more generally.46. IIAxl1 IIAII = max .jii IIAlioo' . IIAII2::S IIAIIF. Not every consistent matrix norm is subordinate to a vector norm. 2.60 Ilx i. 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. B E ]Rnxk. so not exist a vector norm  . IIAxll1 = max . The "mixed" norm "mixed" norm II· 11 100 . For example. We thus need the following definition. i. IIAllp. Example 7. For example. For example. A = max^o IIxll. IIAlioo ::s .. II F' ThenA^ < A II Filx 112. For such subordinate norms. For A following inequalities are all tight.60 \^ • Useful Results The following miscellaneous results about matrix norms are collected for future reference.. IIAxliv < IIAlim \\x\\v' Not every consistent matrix norm is subordinate to a vector norm. Since IIABxl1 ::s Afljc ::s IIAIIIIBllllxll.4. i.. subordinate to the vector norm. If II .g.45. inner products or outer products of vectors. inner products or outer products of vectors. Example 7. 1. .and II \\ •lIy y are mutually consistent if \\ A B \\ a < IIAllfllIBlly. Theorem 7.. \\ • \\p.oo J1. IIAlioo ::s .q = maxx. i. Definition 7. Notice that this difficulty did not arise for vectors. •1122is consistent with II . If \\ • 11m is a consistent matrix norm.e..48.7.e.jii IIAIIF. Matrix Norms 7. II· II F and 1..jii IIAlloo. Then the norms \\ . HAjcJI^ ::s \\A\\m Ilxli v.60 IIx II Ilxll=1 IIAxll p .j is a matrix norm but it is not consistent. . \\m is a consistent matrix norm. i. There exists a vector x* such that Ajt* = A jc* if the matrix norm is Theorem 7. but there does consider II . take A = B = [: is a matrix norm but it is not consistent. For example. 11^4^11 P (or. also called oper(or. II In II p = 1 for all p. consider  • \\F. IIAIIF ::s. IIABIII. 2.ooIlBIII. Matrix Norms 61 61 Notice that this difficulty did not arise for vectors. also caUedoperator norms.so II . II A 1100 ::s n IIAII I . although there are analogues for. there exists a vector norm II • IIv Theorem 7. Then :].jii.g.jii IIAlb IIAIII ::s n IIAlloo. exercise. 1. ))..but there does  is consistent with F. a vector norm. Let A E ]Rmxn.jii IIAII2.jii II A IIF..~~i'. Then the norms II • \\a II· Ilfl' and .. if II A B II < II A 1111 fi whenever the matrix product is defined. A matrix norm 11·11\\is said to be consistent mutuallyconsistentifIlABII. not exist a vector norm II •  such that IIAIIF is given by max x . we clearly have Ajc ::s A1jt.60 . i. q For such subordmate norms. e. •II F. reader The interested reader is invited to prove each of them as an exercise.e. wec1earlyhave IIAxll < IIAllllxll· Since Afijc < IIAlIllBxll < Afljt. A A 2. take A = B = \ \ Afl li00 = 2whileA li00 B 1>00 = 1. which equality holds: which equality holds: IIAIII ::s .45. II". although there are analogues for.. there exists a vector norm \\ . more generally.4.47.. Let A e Rmxn.46.= max IIAxl1 x.e. it follows that all subordinate norms are consistent. it follows that all subordinate norms are consistent.48.jii IIAlb IIAIIF ::s .jii IIAII I . IIAII2 ::s.e. e R" x ". Theorem 7..1100 = max laijl x. The following miscellaneous results about matrix norms are collected for future reference. B e Rnxk. There exists a vector x* such that IIAx*11 = IIAllllx*11 if the matrix norm is subordinate to the vector norm.47. while E ]Rnxn.. l. IIAIIF ::s .
B) = space. . and Norms max laijl :::: IIAII2 :::: ~ max laijl. Let II ·11 be a matrix norm and suppose A. matrices Q zR and e M" ". 1. spanned by the plane 3x .62 62 3. 3. Prove that E"xn with the inner product (A.49. scalars. B) = Tr ATB is a real inner product IR n x" AT B (A..] l. For A E IR mxn . ..1. must be an orthogonal matrix. Find the (orthogonal) projection of the vector [2 3 4f onto the subspace of 1R3 5. l.Q — Q must be an orthogonal matrix. 6. A(I). Inner Product Spaces. > .A+A is an orthogonal projection. Definition: Let A E IRnxn and denote its set of eigenvalues (not necessarily distinct) by P... .I. i. The spectral radius of A is the scalar by {Ai .. ..49. Suppose that a matrix A E IR mxn has linearly independent columns. IIQAZlia = A fora = 2 or F. where ¥2 is defined as in Theorem 5.. i. 2. EeIRmxn. 4. [2 3 4]r R3 spanned by the plane 3. prove directly that V22Vl is an I — +A V V/ is an orthogonal projection..1. 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. e Rmx" mxm x mxm and Z E IRnxn .. Theorem 7. If P projection. Projections. IIF are unitarily invariant. prove that P+ = P. (MZa or F. but not necessarily other pnorms) are unitarily invariant.An}. For A eRmxa . and Norms Chapter 7. Suppose P and Q are orthogonal projections and P + Q = I.y + 2z = O. but not necessarily The norms  • \\F and  • 2 (as well as all the Schatten pnorms. . Prove that P . orthogonal projection. 3.c — v + = 0. for all A E IRmxn and for all orthogonal unitarily invariant. 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. space.. Prove that / . Inner Product Spaces. The norms II . 7. „ } The spectral radius of A is the scalar p(A) = max IA. Then k~+oo lim A (k) = A if and only if k~+oo lim IIA (k)  A II = o. p+ = P. If P is an orthogonal projection. 112 (as well as all the Schatten /?norms.  • 2 and  • \\F 8. Also.e. EXERCISES EXERCISES 1. Then 7. Definition: Let A e Rnxn and denote its set of eigenvalues (not necessarily distinct) 8. Show that the matrix norms II . A(2).. i . Chapter 7. A (2) . Projections. A (1) .. \\ \\bea Rmx".e. where V2 is defined as in Theorem 5..l. IIF and II . 112 and II . A(A A) 1 AT 5.] 4.
HA^. where both x. columns and rows as well as main diagonal and antidiagonal sum to s = n (n 2 l)/2. is called a "magic square" matrix. Let A = xyT. 10. Aj. H A H ^ and peA). all of whose Determine AF. all of whose columns and rows as well as main diagonal and antidiagonal sum to s = n(n2 + 1) /2.) that  M Up = for all/?. (An n x n matrix. and p(A). \\A\\ A2. 9. Determine IIAIIF' IIAII d . where both x. (An n x n matrix. H A I I A2. Determine AF.) T 10. IIAlb IIAlloo. Let 9. and p(A). where a and ft take the value 1. Determine IIAIIF' IIAIII> IIAlb and Aoo in terms of \\x\\a and/or \\y\\p. or oo as and II A 1100 in terms of IIxlla and/or IlylljJ.Exercises Exercises 63 63 Let Let A=[~ 14 0 12 5 ~].2. 2. If M is a magic square matrix. y e R" are nonzero. where ex and {3 take the value 1.. A2. IIAlb IIAlloo. Let A = xy . . it can be proved that IIMllp = ss for all p. Determine AF. or (Xl as appropriate. appropriate.. y E IR n are nonzero. Let A=[~4 9 2 ~ ~]. and peA). Determine IIAIIF' IIAII Ilt.
This page intentionally left blank This page intentionally left blank .
so these two vectors are orthogonal. The linear least Problem: Suppose A e jRmxn with > n and b <= Rm is given vector. For further details.Chapter 8 Chapter 8 Linear Least Squares Linear Least Squares Problems Problems 8.Ax is the residual associated 1. while (b .. Hence. see Section 8..PR(A)b) + (PR(A)b  Ax).bll~ (and hence p(x) = \\Ax .e. write the residual r in the form To see why this must be so.PR(A))b = PR(A). For further details. write the residual in the form r = (b .1) To see why this must be so. (Pn(A)b — AJC) is clearly in 7£(A). 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 . A.PR(A)b) = (I . The equations ATrr = 0 can be rewritten in the form A TAx = ATb and the x.2. A vector x X if and onlv if x is of the x=A+b+(IA+A)y.. (PR(A)b . Now. i. 2. Hence. A vector x E X if and only if x is of the form 2. IIrll~ = lib . i.1 8.PR(A)bll~ + IIPR(A)b  Axll~ from the Pythagorean identity (Remark 7.e. IIAx .35). Thus. x E X if and only if x is a solution of the normal equations.2.x — b\\\ (and hence p ( x ) = from the Pythagorean identity (Remark 7. AT — A T Ax = AT b latter form is commonly known as the normal equations. Solution: The set X has a number of easily verified properties: The set X has a number of easily verified properties: 1. where r = b . (8.b 112) assumes its minimum value if and only if II Ax —b\\2) assumes its minimum value if and only if (8.35).Axll~ = lib .Ax) is clearly in 'R(A). while Now.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. Thus. vector x e X if and only if AT where b — Ax is the residual associated with x. whereyEjRnisarbitrary.2) 65 . x e X if and only if latter form is commonly known as the normal equations.bll 2 is minimized}. A vector x E X if and only if ATrr = 0.b E 'R(A)L so these two vectors are orthogonal. see Section 8. is a solution of the normal equations.
X. equivalently. BE ]R. To see why.. 3. X = {x"} = {A+b}. By Theorem 6.nxk is arbitrary. Just as for the solution of linear equations. If the existence condition happens to be satisfied. There is a unique solution to the least squares problem.1. then equality holds and the least squares . consider two arbitrary vectors jci = A+b + (I — A + A) y (I .PR(A)bll z = ~ 11(1 Ilbll z. Let 8 e [0. all and this equation always has a solution since AA+b E R(A). The only difference is that in the case same as solutions of the linear system AX = B.mxk.0)z) is clearly in 4. if 5. x* minimizes the residual p(x) that solves this "double minimization" problem. The Theorem 8. 0*i (1 #)* = A+b (I .mxn XElR Plxk min IIAX  Bib is of the form is of the form X=A+B+(IA+A)Y. The minimum value of p x ) is then clearly equal to lib .1) and which follows since the two vectors are orthogonal. Notice that solutions of the linear least squares problem look exactly the Remark 8. if and only if A + A lor.e.2) are of the form x = A+ AA+b + (I  A+ A)y =A+b+(IA+A)y. if and only if rank(A) n.A+ A)z in X. X has a unique element x* of minimal 2norm. This follows immediately from and is the vector of minimum 2norm that does so.. 5. which follows since the two vectors are orthogonal. By Theorem 6. The minimum value of p ((x) is then clearly equal to where y E ]R. x" = + b is the unique vector 4.1].A+A)(Oy (1 . where Y E R" xfc is arbitrary. equivalently.23. Linear Least Squares Problems and this equation always has a solution since AA+b e 7£(A). Then the convex combination and Xz = A+b (I . X is convex. 7£(A).e. Remark 8. the last inequality following by Theorem 7. The general solution to e ]R. Linear Least Squares Problems Chapter 8. and only if A+A = I or. The unique solution of minimum 2norm or Fnorm is X = A+B.2) are of the form solutions of (8. 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. problem to the matrix case. In fact. i.e. consider two arbitrary vectors Xl = A + b 3.2. 1]. Then the convex combination 8x. The unique solution of minimum 2norm or Fnorm is where Y € ]R. i. i. Let 6 E [0.66 Chapter 8. of linear least squares solutions.2.A+A)y and *2 = A+b + (I — A+A)z in X. we can generalize the linear least squares problem to the matrix case. There is a unique solution to the least squares problem. there is no "existence condition" such as K(B) c R(A). x* minimizes the residual p ( x ) and is the vector of minimum 2norm that does so. X is convex.. there is no "existence condition" such as R(B) S. then equality holds and the least squares If the existence condition happens to be satisfied. In fact. To see why. + (1 . all solutions of (8. AA+)bI1 2 the last inequality following by Theorem 7.1) and convexity or directly from the fact that all x E X are of the form (8. i. This follows immediately from convexity or directly from the fact that all x e X are of the form (8.. x* = A+b is the unique vector that solves this "double minimization" problem. X = {x*} = {A+b}. Let A E E mx " and B € Rmxk. if and only if rank (A) = n. where y e W is arbitrary. X = A+B.. The only difference is that in the case of linear least squares solutions.n is arbitrary. has a unique element x" of minimal2norm.23.e.3.3.8)xz2 = A+b ++ (I A+ A)(8y ++ (1 8)z) is clearly in X..
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
4 8. 8. insight. [23]). The key feature in (8. Better numerical methods are based on algorithms that AT work directly and solely on A itself rather than AT A. c. since II . (8.. it can be expected to exhibit such poor numerical behavior in practice (and it does).can be arbitrarily nonlinear. appear functions </>. z. [11]. This explains why it is convenient to work above with the square of the norm rather than the concerned. Linear Least Squares Problems difficulties because of numerical ill conditioning for large n. The former is much more expensive but is generally more reliable and offers considerable theoretical offers insight. the subvectors can have different lengths). S~lc\. As far as the minimization is concerned. . bE IR m . it is shown that solution [4]. For example. norm. c.5) is that the coefficients Ci appear linearly. ] II: = II [ The last equality follows from the fact that if v [~~]. Two basic classes of algorithms are A itself S VD and QR (orthogonalupper triangular) factorization.c=UTb = II [~ ~] [ ~~ ] . c\ logci. We now note that IIAx  bll~ = IIU~VT x =  bll~ II ~ VT X  U T bll. e C2 / then taking logarithms yields the equation log y = log c. A E IRmxn . In fact. VT = U. Linear Least Squares Problems Chapter 8. the subvectors can have different lengths).g.. Sometimes a problem in which the Ci'S appear nonlinearly nonlinearly can be converted into a linear problem. are based on orthogonal polynomials. Ib is unitarily invariant =11~zcll~ wherez=VTx. In this section we investigate solution of the linear least squares problem min II Ax x b11 2 .1. [7].. Since the standard Kalman filter essentially amounts to sequential updating of normal equations. the last equivalent. arbitrary. Specifically. [4]. then II v II ~ = II viii ~ + II v211 ~ (note that orthogonality is not what is used here. Then GI defining y = logy. For example.1.6) via the SVD. problem. Numerically better approaches ill difficulties n. + c2f. piecewise polynomial functions.SVr U~VT Theorem 5.[ ~~ ] II: sz~~ c. This that orthogonality is not what is used here.70 70 Chapter 8. etc. etc. Then c. [7]. splines. Specifically. Better numerical methods are based on algorithms that behavior in practice (and it does). of linear least squares problems via the normal equations can be a very poor numerical method in finiteprecision arithmetic.4 Least Squares and Singular Value Decomposition Least Squares and Singular Value Decomposition In the numerical linear algebra literature (e. respectively. fact. Since the standard Kalman filter essentially amounts method in finiteprecision arithmetic. The basis functions coefficients c. as in Theorem 5. Z2 while the minimum value of \\Ax — b II ~ is l^llr while the minimum value of II Ax . if the fitting function is of the form can be converted into a linear problem. then u^ = i>i \\\ \\vi\\\ (note The last equality follows from the fact that if v = [£ ].b\\^ is II czll ~. then taking logarithms yields the equation logy = logci + cjt. the two are equivalent. quantity above is clearly minimized by taking z\ = S'c. 's ¢i. The subvector z2 is arbitrary. respectively. we assume that A has an SVD given by A U\SVf via the SVD. we assume that A has an SVD given by A = UT. The former based on SVD and QR (orthogonalupper triangular) factorization. C2 problem. = log c" and C2 = cj_ results in a standard linear least squares y — log y. if the fitting function is of the form y t) Y = ff( (t) = c\eC2i.
with appropriate numerical enhancements. 8.6) but this In this section. A e R™ X M . i.. This matrix factorization is much cheaper to compute time in terms of the QR factorization. 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.5 Least Squares and QR Factorization Least Squares and QR Factorization In this section.AA+)bll~ .1).AA+)b\\22 = \\U2Ufb\\l = bTU2U^U22V!b = bTU2U*b = \\U?b\\22. we have QT € ffi.5. A finite sequence of simple orthogonal transformations. is orthogonal to all vectors in 7l(A}L b E R(A). 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. Thus. we again look at the solution of the linear least squares problem (8. with (8.S.e.AA+)bllz. In this case the SVD of A is given by A A = V:EVTT = [VI{ Vzl[g]Vr. to reduce A in the following way. i.7) .. It is then possible. i.11U2U!b"~ = bTUZV!V UJb = bTVZV!b = IIV!bll~. and there is thus "no V2 part" to the solution. V2 Z 2 is an arbitrary vector in 7Z(V2)) = A/"(A). This agrees. i. we add the simplifying assumption that A has full column rank.~xn. V2z is an arbitrary vector in R(V2 = N(A). than an SVD and. A E ffi. an important special case of the linear least squares problem is the Finally. x has Note that since 12 is arbitrary. Another expression for the minimum residual is  (/ — AA + )b 2 . A e 1R™ X ".e. where y e ffi. This agrees. can be quite reliable. UZV = [U t/2][o]^i r > and there is thus "no V2 part" to the solution. In this case the SVD of A is given by socalled fullrank problem. Finally. A E ffi. = +b + (/ — A + A) y. It is then possible.5 8. can be quite reliable.mxm.8. we add the simplifying assumption that A has full column To simplify the exposition. Least Squares and QR Factorization B. If we label the product of such orthogonal row transformations as the to triangular form. This follows easily since (7 . of course. 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 2.6) but this time in terms of the QR factorization. x has been written in the form x = A+b + (I . with appropriate numerical enhancements. where y E Rm is arbitrary. If we label the product of such orthogonal row transformations as the orthogonal matrix QT E R m x m . This follows easily since Another expression for the minimum residual is II (I . an important special case of the linear least squares problem is the socalled fullrank problem.m is arbitrary.e. via a sequence of socalled Householder or Givens rank. To simplify the exposition. to reduce A in the following way. with (8. via a sequence of socalled Householder or Givens transformations. we again look at the solution of the linear least squares problem (8.A + A)_y. (8.. This matrix factorization is much cheaper to compute than an SVD and. 11(1.~xn. of course.e. Thus.1). Z VISici The last equality follows from The last equality follows from c = UTb = [ ~ f: ]= [ ~~ l Note that since Z2 is arbitrary..
7). n .e.8). or (8. or (8. we see that A=Q[~J = [QI = QIR. (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. 112 is unitarily invariant ~ ] x . Qz] [ (8. Consider the following set of measurements (*.. 3. i.e.[ ~~ ] If:. check directly that (I . Both Q I and Qz2 have orthonormal columns. (b) Let r denote the "error vector" b . Note that Any of (8. where Q\ e R mx " and Qz € K" x(mn). Consider the following set of measurements (Xi.. and qz are two orthonormal vectors and b is a fixed vector.mxn and where R e M£ x " is upper triangular. data. Multiplying through by Q in (8. and any y E R". Linear Least Squares Problems where E ffi..9) are variously referred to as QR factorizations of A. Now write Q = [Q\ Qz]. m and any e ffi. we have x = R. Suppose qi and q2 are two orthonormal vectors and b is a fixed vector.9) are variously referred to as QR factorizations of A. Both Q\ and <2 have orthonormal columns. R~l) ) of the columns of A yields the orthonormal columns of QI.~xn is upper triangular. yt): (1. i.9) is essentially what is accomplished by the GramSchmidt process. 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. Linear Least Squares Problems Chapter 8.72 Chapter 8. (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. For A E Wmxn . all in R". Equivalently.3). Suppose q. we see that in (8. Note that (8. b E Em.9) is essentially what is accomplished by the GramSchmidt process.A+A)y and A +b 1..+ A)y and A+b are orthogonal vectors.flq2 Show that r is orthogonal to (b) Let r denote the "error vector" b — ctq\ — {3qz.1). Now note that Now note that IIAx  bll~ = IIQ T Ax = II [ QTbll~ since II .I of the columns of yields the orthonormal columns of Q\.8). (8. check directly that (I .8) ~ ] (8. (3. 2. (8. sense).m IX(m ~" ) .9) Any of (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). we have = R~l Qf b = +b and the minimum residual is IIC?^!^ EXERCISES EXERCISES 1. data. 3. are orthogonal vectors. Yi): 2. For € ffi. b e ffi.7).Cl and the minimum residual The last quantity above is clearly minimized by taking x = R lIc\ and the minimum residual is Ilczllz.Show that r is orthogonal to both^i and q2. Now write Q = [QI Q2].2). xn.. (2.Equivalently.7). Multiplying through by Q Q2 E ffi. all in ffi. where QI E ffi.1Q\b = A+b and the minimum residual is II Qr bllz' is \\C2\\2.. both ql and q2 . by writing (8.7). The last quantity above is clearly minimized by taking x = R.aql .
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. not necessarily nonsingular. Solve the perturbed version of the above problem. (a) Consider a perturbation E\ = [0 ~] of A. where again 8 is a small of A.z2 as 8 approaches O? where A2 — A E 2 What happens to \\x* — zll2 as 8 approaches 0? 6.:. Solve the perturbed problem positive number. Solve the perturbed problem min II A 2 z . of 2norm solution of least «rmarp« problem squares nrr»h1<=>m min II Ax . then A+ == R. and suppose A where is 1.Exercises Exercises 73 4.bl1 2 when A = [~ ~ ] and b = [ !1 x The solution is (a) Consider a perturbation EI = [~ pi of A. Solve the perturbed version of the above problem. A+ R+QT . where 8 is a small positive number. Use the four Penrose conditions and the fact that Q\ has orthonormal columns to 6. Let A e R"x".9). then A+ R~ Q\.• What happens to IIx* .9). Find all solutions of the linear least squares problem min II Ax . and suppose A = QR. verify that if A E ~. where Q is orthogonal. Find all solutions of the linear least squares problem 4. Consider the problem of finding the minimum 2norm solution of the linear least 5. Prove that A+ = R+ QT.IlQf. 7. What happens to jt* .xn can be factored in the form (8.yII2 as 8 approaches O? (b) Now consider the perturbation E2 = [~ (b) Now consider the perturbation EI = \0 s~\ of A. where again 8 is a small positive number. where 8 is a small positive number.bll 2 x when A = [ ~ 5.bib z n where A2 = A + E2. where AI = A + E I . Let A E ~nxn. What happens to IIx* — y 2 as 8 approaches 0? where AI = A + E\. not necessarily nonsingular.
This page intentionally left blank This page intentionally left blank .
then so is ax [ay] for any nonzero scalar a E C. for proved easily from the Jordan canonical fonn to be discussed in the text to follow (see. [21D or directly using elementary properties of inverses and determinants (see.1. Note that if x [y] is a right [left] eigenvector of A. It can be proved from elementary properties of determinants that if A E C" ". for example. Definition 9. (9. as a matter of convenience. Let A = [~g ~g].Chapter 9 Chapter 9 Eigenvalues and Eigenvalues and Eigenvectors Eigenvectors 9. such that Ax = AX. we use both forms results in at most a change of sign and.t so that the scaled eigenvector has norm 1. for example. as a matter of convenience. called an eigenvalue.4. the Fundamental Theorem of Algebra says that x 75 . Thus.2. It can be The following classical theorem can be very useful in hand calculation. we use both forms throughout the text.2.31 = 0.1. [3]). A nonzero vector x E en is a right eigenvector of A E e nxn if there exists Definition 9. we see immediately that XH is a left eigenvector of A H associated with A. A nonzero vector x e C" is a right eigenvector of A e Cnxn if there exists a scalar A.1). Example 9. This of A.— 3. Then n(k) = X2 + 2A. example.) throughout the text. Thus. then n(A) is a polynomial of degree n. [3]). Theorem 9. n(A) = 0. n(A) = O.1) Similarly.Al) is called the characteristic polynomial Definition 9. a nonzero vector y E en is a left eigenvector corresponding to an eigenvalue Similarly. a nonzero vector y e C" is a left eigenvector corresponding to an eigenvalue a if Mif (9.A). (Note that the characteristic polynomial can also be defined as det(Al .1). It is an easy exercise to 2 verify that n(A) = A + 2A . we see immediately that x H is a left eigenBy taking Hermitian transposes in (9. This results in at most a change of sign and.1 9. for example. Let A [~ ~]./ — A). The 2norm is the most common nonn used for such scaling. verify that n(A) = A2 2A . Theorem 9.2) By taking Hennitian transposes in (9. Note that if x [y] is a right [left] eigenvector of A. [21]) or directly using elementary properties of inverses and determinants (see. e C. It can be proved easily from the Jordan canonical form to be discussed in the text to follow (see. then vector of AH associated with I. The polynomialn (A. The polynomial n (A) = det(A—A./) is called the characteristic polynomial of A.4. The 2nonn is the most common a — \j'. Then n(A) A2 + 2A 3. the Fundamental Theorem of Algebra says that 7t (X) is a polynomial of degree n. For any A e Cnxn .31 O. For any A E e nxn . such that a scalar A E e.1 Fundamental Definitions and Properties Fundamental Definitions and Properties Definition 9.3 (CayleyHamilton).) = det (A ..} The following classical theorem can be very useful in hand calculation. 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. then It can be proved from elementary properties of detenninants that if A e enxn .3 (CayleyHamilton). (Note that the characteristic polynomial can also be defined as det(A. norm used for such scaling. called an eigenvalue. It is an easy exercise to Example 9.
A.5..~. all roots of its characteristic polynomialn(A). we get the interesting fact that del (A) = A] • A2 • • An (see also Theorem 9.AI) :::: m.A) . degree such that a (A) =0. if If A € A(A) has algebraic multiplicity m. if we denote the geometric multiplicity of A by g. it can also be . we say that X is an eigenvalue of A of algebraic multiplicity m. A is said to be defective if it does not have n linearly independent (right or left) eigenvectors. A is said to be defective if it does not have n linearly independent (right or left) eigenvectors.. Specifically. but that A(A) = A(A) only if A E 1Ftnxn..5. neftnhion ~. Note.. 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. The minimal polynomial Of A l::: K""" ix (hI' polynomilll a(A) of least degree such that a(A) ~ O. Xn. say. Thus. eigenvalues of A. Definition 9. y of AT y is a left eigenvector of A corresponding to A e A(A). Equivalently. the set of Definition 9.n =0o. However. The spectrum of A e nxn is the set of all eigenvalues of A.e. E A(A). the set of all roots of its characteristic polynomial n(X). For example. such a polynomial is said to be monic and we of the highest power of A to be +1. less than) its algebraic multiplicity.AI) = dimN(A . we say that A is an eigenvalue of A Definition 9.25). that by elementary properties of the determinant. we always have A(A) = A(AT). independent eigenvectors = n . (9. then 1 :::: dimN(A .ft Definition 5.2. A matrix A E Wnxn is said to be defective if it has an eigenvalue whose Definition 9. • AM(see and set X = 0 in this identity.2aA + aa2+ f322 and A has eigenvalues a f3j (where A has eigenvalues a ± fij (where j = i = R). But it also clearly satisfies the smaller degree polynomial equation isfies (1 . and hence further guarantee the existence of corresponding nonzero eigenvectors. of we always have A(A) = A(A r ). Hence the roots of 7r(A).7.e. For example.A) (9. But it also clearly satisfies the smaller degree polynomial equation (it. then there is an easily checked relationship between the left and right If A € R"x". we get the interesting fact that det(A) = AI . The geometric multiplicity of A is the number of associated independent eigenvectors = n — rank(A — A/) = dim J\f(A — XI)..6. say. sible for A to satisfy a lowerorder polynomial.8. possibly repeated. as solutions of the determinant equation n(A) = det(A  AI) = 0.4) and set A = 0 in this identity.. The spectrum of A E C"x" is the set of all eigenvalues of A. These roots. IfXA is a root of multiplicity m ofjr(X). guarantee the existence of corresponding nonzero eigenvectors.76 Chapter 9.6. geometric multiplicity is not equal to (i. it can also be generally write a(A) as a monic polynomial throughout the text). eigenvalues of A. These roots. then we must have I < g < m.. A.AI) = (A] .. if left of A A E A(A). and hence further are the eigenvalues of A and imply the singularity of the matrix A — AI. A matrix A e 1Ft x" is said to be defective if it has an eigenvalue whose geometric multiplicity is not equal to (i. 2aA + 2 + ft and Example 9.. we denote the geometric multiplicity of A by g. i. Thll minimal polynomial of A G l!if. Specifically. a .. but that A(A) = A(A) only if A e R"x".AI). If is a root of multiplicity m of n(A). must occur in complex conjugate pairs. Eigenvalues and Eigenvectors n(A) has n roots. Moreover.. Definition 9. Example 9.rank(A .e. If A E A(A) has algebraic multiplicity m. then A satisfies (Je — I)2 = 0.. as solutions of the determinant equation 7r(A) has n roots. Definition 9. ft E 1Ft and let A = [ _^ !]. then we must have 1 :::: g :::: m. possibly repeated. if A = [~ ~]. we know that n(A) = 0. (An . then n(X) has real coefficients.. i. too. c form form A e C" " A]. Then n(A) = A22. An. If e Wxn.2». such a polynomial is said to be monic and we generally write et (A) as a monic polynomial throughout the text). if A = \1Q ®].1) .e... checked eigenvectors of A and AT (take Hermitian transposes of both sides of (9.:. The geometric multiplicity ofX is the number of associated of algebraic multiplicity m.XI. Theorem If A E 1Ftnxn. f3 e R and let A = [~f3 £ ]. . Then if we write (9. Equivalently..1)2 = O. . less than) its algebraic multiplicity. Let a. Thus.nxn is the polynomial o/(X) oJ IPll. Then jr(A.. if eigenvectors of A and AT (take Hermitian transposes of both sides of (9. Let the eigenvalues of A E en xxn be denoted X\. must occur in complex conjugate pairs..5.3) are the eigenvalues of A and imply the singularity of the matrix A .e. then y is a right eigenvector of AT corresponding to I € A (A). i. then I < dimA/"(A — A/) < m. The spectrum of A is denoted A (A). From the CayleyHamilton Theorem. The spectrum of A is denoted A(A). Let a.2)). then A satsible for A to satisfy a lowerorder polynomial. i •>/—!)• If A E 1Ftnxn.8. However. n(A).) A.3) in the n(A) = det(A .7. Eigenvalues and Eigenvectors Chapter 9. Moreover. the n(A) coefficients. it is posn(A) = O..
eigenvector is numerically unstable. The above definitions are illustrated below for a series of matrices.11. 0 0 0 2 0 0 0 2 ] h'" a(A) (A ..) directly (without knowing eigenvalues and as Unfortunately. be an eigenvalue of A with corresponding right Theorem 9.5) = (A  2)2 and g = 2. Let A e C« x " ana [et A.2)' ""d g 2. Example 9."(A) ~ ~ ~ ~ (A . Bezout algorithm.1.) divides every nonzero polynomial fi(k} for which ft (A) = 0.1.10. this algorithm..2)' ""d g ~ ~ ~ 1. sociated eigenvector structure). Unfortunately. . Let A E cc nxn and let Ai be an eigenvalue of A with corresponding right eigenvector jc.10. 0 0 0 2 A~U 0 0 ] ha<a(A) (A . of which has an eigenvalue 2 of algebraic multiplicity 4.9.11.. a(A) divides n(A). A[~  2 0 I 2 0 0 0 0 0 0 !] ~ ~ ~ ha.*. 0 g At this point. shown that a (A. Furthermore. Theorem 9. a(X) n(X). Proof' Since AXi AiXi.2)2 ""d g 3. left Aj E A (A) such that Xj 1= A.e. Then yfx{ = O. Fundamental Definitions and Properties 9. e l\(A) yj Xi. A~[~ A~U 2 0 0 I 2 2 ] ha< a(A) (A . a(A) directly (without knowing eigenvalues and asThere is an algorithm to determine or (A. such is not the case.. The matrix A~U has a(A) I 2 0 0 2 0 0 0 !] (9. every nonzero polynomial f3(A) particular. Then Xi = 0. We denote 7r(A) (A . Fundamental Definitions and Properties 77 77 a(A) f3(A) O. In particular. g. algorithm.2)4. one might speculate that g plus the degree of a must always be five. let Yj be a left eigenvector corresponding to any A. n(A) = (A — 2)4..2) andg ~ 4. the geometric multiplicity by g. called the Bezout algorithm.. YY Proof: Since Axt = A. i. Example 9. each 4. Unfortunately. such that Aj =£ Ai.
78 Similarly.14. 0 The proof of Theorem 9. Let A E cnxn have distinct eigenvalues A . Then all eigenvalues of A must be real. c Proof: Suppose (A. Then Proof: Suppose (A.e. = A. Take the Hermitian Proof: Premultiply the equation Ax = AX by ZH to get ZH Ax = AZ H x. eigenvector. However. p. Let A €. Theorem 9..Aj ^ 0. it cannot be the case that yf*xt = 0 as orthogonal to all yj's for which j ^ i. A is real. and if Ai E A(A).11. i. we can choose the normalization of the *. the proof see. for [21. . AXH z. . 0 D Theorem 9. Xi is orthogonal to all y/s for which j =1= i. x) is an arbitrary eigenvalue/eigenvector pair such that Ax = AX. i. then by Theorem 9.6) from (9.13. Since equation Az i^z XH get X H Az = iJXH A. A = AH. we must have Subtracting (9.n with corresponding right eigenvectors x\. contradicting the fact that it is an eigenvector. i. (9. then by Theorem 9. The same right eigenvectors XI..are distinct eigenvalues of A with corresponding right eigenvectors x and z..13.12. yr .5). D A =1= iJ. The same result holds for the corresponding left eigenvectors. YyXi =0. c Proof: Premultiply the equation Ax = A. i. Let A e nxn be Hermitian. Eigenvalues and Eigenvectors yy. Eigenvalues and Eigenvectors Chapter 9.. A = AH. for i € !1.5)....e. we have that IXHxx = XxHx. so that YitH x.12. 1 ?. x) is an arbitrary eigenvalue/eigenvector pair such that Ax = A. since YY A = AjXjyf. . or both. i. Since Xi ^ 0 and would thus have to be 0.. result holds for the corresponding left eigenvectors. 's..e. ^ /z. so that y H Xi = 1 for each i..JC. and if A.. we find 0 = (A. XH x /= 0. we find 0 = (Ai . However.e. Then all eigenvalues of A must Theorem 9. . or else xf would be orthogonal to n linearly independent vectors (by Theorem 9. Then x and z must be of A with corresponding right eigenvectors and respectively. [21. the two vectors must be orthogonal. .. is If A e C nxn has distinct eigenvalues. we can choose the normalization of the Xi'S. Since XxHz. we must have yfxt = O. = 1 for/ E n. Then [x\.. jc. A. orthogonal. or else Xi would be orthogonal to n linearly independent vectors (by Theorem 9. 0 If A E cnx " has distinct eigenvalues.. since y" A = Similarly. xn}} is a linearly independent set. we have that XxH = AXH x. since x is an Using the fact that A is Hermitian. However. Then and z must be orthogonal. or the Yi 's.. the two vectors must be orthogonal.14) and would thus have to be 0. x ... from which conclude A. 118]. or both... Let A e C"x" be Hermitian and suppose A and /JL are distinct eigenvalues Theorem 9. e A(A).e..is real.11 is very similar to two other fundamental and important The proof of Theorem 9.11 is very similar to two other fundamental and important results. Let A E C"x" be Hermitian. Since A. we must have that XHzz = 0. Premultiply the equation Az = iJZ by x H to get x HAz = /^XHZZ = XxHz. be real.— A.e. since x is an eigenvector... Cnxn have distinct eigenvalues AI. from which we conclude I = A. is real to get H Az AxH z. i. for example. we have xHX =1= 0..6) from (9. 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. Chapter 9. 118]. 's. Let us now return to the general case. x is a linearly independent set. Then (9.7..JC by ZH to get ZH Ax = X z Hx .7) Taking Hermitian transposes in (9. Since Ai . Let A E nxn be Hermitian and suppose X and iJ. A.=1= 0. respectively.11.7) yields Taking Hermitian transposes in (9. .Aj)YY xi. it cannot be the case that YiH Xi = 0 as well. p. Proof: Proof: For the proof see. results. Theorem 9.. An with corresponding Theorem 9.— A y )j^jc. or the y. However...7) yields Using the fact that A is Hermitian.xnn • Then {XI.14) well. . Since yf*Xi =1= 0 for each i.. we must have that x H = 0.6) Subtracting (9. 0 Let us now return to the general case. contradicting the fact that it is an eigenvector..14.
8) while y^Xj = 5. i E !!. . corresponding to these eigenvalues.. xn]. solve the linear system (A — (—1 + 2j)I)x2 = 0 to get yi X2 =[ 3+ j ] 3 ~/ . let Y — [y\... For A2 = 1 + 2j. Furthermore.. . y' E !!. . Furthermore. Similarly. solve the 3 x 3 linear system (A . let A = diag(AJ. . and this then determines the other two Note that one component of XI can be set arbitrarily.22 + 2)" + 5). is expressed by the equation yHX = I. let Y = [YI. 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.. Fundamental Definitions and Properties 9. .15.(1 + 2j) I)x2 = 0 to get For A2 — 1 + 2j. Similarly. / e n. . 1 ± 2j}."" yn] be the matrix of corresponding left eigenvectors.11) Example 9. can be written in matrix form as AX=XA (9.. — 1 2 j } .. We can now find the right and left eigenvectors corresponding to these eigenvalues.. suppose that the left and right eigenvectors have been normalized so that yf1 Xi = 1.15. .. Xn) E Wtxn.10) = XAX. An) e ]Rnxn. j e n. 2)(A.. from which we find A(A) = {—2. = A. / en. . Fundamental Definitions and 79 Theorem 9..1.9.2 + 9)" + 10) = ()" + 2)(). solve the (since dimN(A . Let 2 5 3 3 2 4 ~ ] . Let A E C"x" have distinct eigenvalues A.A. is expressed by the equation while YiH Xj = oij... .. can be written in matrixform as diag(A. / en. 10) (A.1. + 5). . An and let the corresponding right eigenvectors form a matrix X [x\. . . Finally. Let Example 9. Then rr(A) = det(A . solve the 3 x 3 linear system (A — (—2}I)x\ = 0 to get For Ai = 2. i E !!:: Finally. To get the corresponding left eigenvector y\. Yn] ing right eigenvectors form a matrix X = [XI. 2A. i E!!..16. xn]. and this then determines the other two (since dimA/XA — (—2)7) = 1). For Al = —2./) = (A 3 + 4A. suppose that the left and be the matrix of corresponding left eigenvectors.AI) (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. Let A e en xn have distinct eigenvalues AI. Then AJC. To get the corresponding left eigenvector YI..nand let the correspondTheorem 9..3 4A2 9 A. Then AXi = AiXi. A. .ci can be set arbitrarily.I . let A = right eigenvectors have been normalized so that YiH Xi = 1.16. solve the linear system (A ..9) =A = XAyH yRAX n (9...*/. from Then n(X) det(A .j.(2)1) = 1). We can now find the right and left eigenvectors which we find A (A) = {2.I = = LAixiyr i=1 (9.(2)l)xI = 0 to get Note that one component of . This time we have chosen the arbitrary scale factor for y\ so that \ = 1.
we can also note that X3 =x2' and yi jj. instead of determining the j.2*2 to get Ax^ = ^2X2.L 4 !.!.AI) = (A33 + 8A 22+ 19A + 12) = (A + I)(A + 3)(A + 4).A similar argument yields the result conjugate the equation AX2 — A2X2 to get AX2 A2X2.( I + 2 j) I) = 0 and nonnalize Y22 so that y"xX2 = 1 to get For A3 = — 1 — 2j.17. + 4). X~l Example 9. Now define the matrix X of right eigenvectors: Now define the matrix of right eigenvectors: 3+j 3j 3. XIAX=A= [ 2 0 0 1+2j o 0 Finally.~q 1 3 2 2 0 2 3 ] 2 ~ y' . For example. Finally. we could have found them instead by computing instead of detennining the Yi'S directly. we For XT. for left eigenvectors.2 However. Proceeding as in the previous example. + 1)(A.c2 = ^.17.80 Chapter 9. Then.) 19X + 12) = (A. Then. we could proceed to solve linear systems as for A. Let Example 9. 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 . 2 It is then easy to verify that It is then easy to verify that 2 . —3. is from which we find A (A) = {I. Proceeding as in the previous example. itit is gtruightforw!U"d to compute straightforward to comput~ X~[~ and and I 0 I i ] 1 x.. Let A = [~ ~ ~] . To see this. For example. = x2). note that we could have solved directly only for *i and x2 (and XT.2j.=. Eigenvalues and Eigenvectors Chapter 9. use the fact that A. we could have found them instead by computing XI and reading off its rows. —4}./) _(A + 8A from which we find A(A) = {—1.!.±1 4 4 4 l+j . similar argument yields the result for left eigenvectors. = I .j ] 3+j . Other results in Theorem 9. 3. 4}. we could proceed to solve linear systems as for A2.=. o 3 Then 7r(A. A. note that we could have solved directly only for XI and X2 (and X3 = X2). + 3)(A.'s directly. det(A .A.15 can also be verified.15 can also be verified. To see this. Then Jl"(A) = det(A . use the fact that A33 = A2 and simply conjugate the equation A. However.L Other results in Theorem 9.2 and simply can also note that x$ = X2 and Y3 = Y2.
representable by a power series X^^o fln*n)> then it is easy to show that representable L~:O anxn). The following theorem is useful when solving systems of linear differential equations. x) maps to (/(A). 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. i=1 . What is true is that the eigenvalue/eigenvector pair (A. Eigenvalues (but not eigenvectors) are invariant under a similarity transTheorem 9. A is diagonalizable). but A2 = f0 0~1]has two has only one right eigenvector corresponding to the eigenvalue 0. in general. 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. —3.. Proof: Suppose (A. I 3 I (. 3. For left eigenvectors we have a similar statement. —4).lIAT)(T~lx)x) = X ( T ~ lIxx). A Theorem 9.19. which is equivalent to the dyadic expansion sion 3 A = LAixiyr i=1 ~(I)[ ~ W~l+(3)[ j ][~ ~ 1 . Then. from which equivalent statement (T~ AT)(T. or ex. ff(x) is a polynomial. then easy to show that the eigenvalues of f(A) (defined as L~:OanAn) are f(A). ( x ) is a polynomial. Remark 9.I = A(T. Remark 9. Eigenvalues (but not eigenvectors) are invariant under a similarity transformation T. namely the theorem statement follows.1.9. For example. but A = [~ has two independent right eigenvectors associated with the eigenvalue o. For left eigenvectors we have a similar statement.1. or sinx.20. where A is diagonal. Fundamental Definitions and Properties 9. or.20. A = [~ Oj ] have all the same eigenvectors (unless. 4).18. we have the equivalent statement (T. x) maps to (f(A). A is diagonalizable). If / is an analytic function (e.18. from the theorem statement follows.but f(A) does not necessarily have all the same eigenvectors (unless. or sin*. say. I I ~J I 2 0 0 0 3 3 3 I I (. e jRnxn n = LeA..) . eigenvalue/eigenvector pair (A. jc) is an eigenvalue/eigenvector pair such that Ax = Xx. For example. 2 3 I (. A = T0 6 2 has only one right eigenvector corresponding to the eigenvalue 0. Theorem 9. Then suppose XI AX = A. x) but not conversely. Then. of Chapter 11.3 0 ~l +(4) [ . ] [~ ~ (I) [ I (.19.g. but /(A) does not necessarily the eigenvalues of /(A) (defined as X^o^A") are /(A). D D yHA = XyH ifandon\yif(T(T Hy)H (T. What is true is that the independent right eigenvectors associated with the eigenvalue 0. x) but not conversely. Let A E R" xn and suppose X~~1AX — A. since T Proof: Suppose (A.g. say. formation T. If f is an analytic function (e. which is equivalent to the dyadic expanWe also have X~l AX = A = diag( 1.1 AT) =X(THyf.txiYiH. X) is an eigenvalue/eigenvector pair such that Ax = AX. since T is nonsingular. The following theorem is useful when solving systems of linear differential equations. or eX. namely yH A AyH if and only if Hy)H(T~1AT) = A(T Hy)H. Fundamental Definitions and Properties 81 81 We also have XI AX = A = diag(—1.
of course.21. I... canonical form.12) where each of the lordan block matrices 1i . It is necessary first to consider the notion of Jordan canonical form. and right eigenvectors xt•. It is necessary first to consider the notion of Jordan A is not necessarily diagonalizable.. of course. .= Xdiag(/(A. An e C 1. to have a version of Theorem 9. ...20 and Corollary 9.20 and its corollary in which It is desirable. i.. we have Proof' Starting from the definition.21 for any function that isis There are extensions to Theorem 9. € n_.. eigenvectors Xi.22.. . from which such a result is then available and presented later in this chapter. to have a version of Theorem 9...e.82 Proof: Starting from the definition.21. .. ff(A) = X f ( A ) X ~ l I = Xdiag(J(AI). i. Eigenvalues and Eigenvectors Chapter 9. then eA has eigenvalues e X"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. Jordan Canonical Form (/CF): For all A e c nxn with eigenvalues AI. Theorem 9. Eigenvalues and Eigenvectors n = LeA.. / € n_. 9. there exists X E C^x" such that XI AX = 1 = diag(ll.. 0 (9. f(An))X.e. .2 9. .. . Corollary 9.20 and Corollary 9.13) 1i = o o Ai o Ai .22. If A E Rnx" is diagonalizable with eigenvalues A.Il .. . f ( X t t ) ) X ~ It is desirable. . /' en. from which such a result is then available and presented later in this chapter. i E ~. we have Chapter 9. then e A has eigenvalues e A There are extensions to Theorem 9. and the same eigenvectors. lordan Canonical Form (JCF): For all A E C"x" with eigenvalues X\. and right Corollary 9.. If A e ]Rn xn is diagonalizable with eigenvalues Ai.21 for any function that analytic on the spectrum of A... kn E C (not necessarily distinct)..20 and its corollary in which A is not necessarily diagonalizable. there exists X € c~xn such that (not necessarily distinct).2 Jordan Canonical Form Jordan Canonical Form Theorem 9.IXiYiH. i E ~. and the same eigenvectors. The following corollary is immediate from the theorem upon setting t = I..i). i=1 0 The following corollary is immediate from the theorem upon setting t = I. ( A ) = Xf(A)X.. 1q). ii E ~.. analytic on the spectrum of A.
.. .~xn necessarily distinct).. the situation is only a bit more complicated. Real Jordan Canonical Form: For all A E R n x " with eigenvalues AI.9. for example. 120124]. 1q is of form where each of the Jordan block matrices 11.JfJ =[ (X fJ fJ ] (X = M.. Jordan Canonical Form and L. ~: ] and I = \0 A in the case of complex conjugate eigenvalues a ± jfJi E A(A).2. . . there exists X € R" xn such that (9. (Xii±jpieA(A>).2. ] T (X .. For nontrivial Jordan blocks.. Proof: For the proof see. Jq is of the form of in the case of real eigenvalues A.14) J\. X (not necessarily X E lR. { ] allow us to go back and forth between real JCF and its complex counterpart: TI [ (X + jfJ o O.. [21. and where M. .An n (not € jRnxn Xi.... pp. = [ _»' ^ 1 and h2 = [6 ~] in the case of complex conjugate eigenvalues where Mi = [ _~. e A (A). Jordan Canonical Form 9.=1 ki = n. complicated. 83 83 Form: 2.. . With 1 j o j o 1 o o o j ~ ~] 0 1 ' .. Proof: proof D 0 Transformations like T = [ _~ "•{"]allow us to go back and forth between aareal JCF Transformations like T = I" _.
. Eigenvalues and Eigenvectors Chapter 9.1)z.84 it is easily checked that it is easily checked that Chapter 9.25. Let A E nxn with eigenvalues AI. Thus.7x7 is known to have 7r(A) = (A .jf3 0 0 et . I) .) = (A. The characteristic polynomial of a matrix is the product of its elementary divisors. 1.1)2.2)3 3and is known to have :rr(A) Example 9.2(A(A . . .(A.2).2). highest degree corresponding to distinct eigenvalues. Theorem 9. Tr(A) = 2. From Theorem 9. The characteristic polynomial of a matrix is the product of its elementary Theorem 9...24.) = det(7) = ]~["=l A. . . . .jf3 0 ]T~[~ l h M Definition 9.23. x Theorem 9. .)i. Thus. Again.2)2. Tr(A) = Tr(XJX~ ) = TrC/X"1 X) = Tr(/) = £"=1 A.. and (A . 1).24. and (A (A .22 we have that A X J XI. J(2) has elementary divisors (A while /( 2) haselementary divisors (A .2)2. Suppose A E E (A.23. Thus.1*) = Tr(J) = L7=1 Ai.1)2. Suppose A e lR. I) . 2 .22 l Tr(A) = Tr(X J XI) Tr(JX.22 we have that A = X JJX ~ l ..(A.. and (A .. D 0 Example 9. 2.22 are called the elementary divisors or invariant factors of A. Let A e C" " with eigenvalues AI. The minimal polynomial of a matrix is the product of the elementary divisors of divisors. i=1 l Proof: Proof: 1.26. i=1 n 2.22 are called the elementary divisors or invariant factors of A. from Theorem 9. Then c n 1.26. The characteristic polynomials of the Jordan blocks defined in Theorem Definition 9. 1)2(A .1)4(A .. det(A) = det(XJX. — 2) . The characteristic polynomials of the Jordan blocks defined in Theorem 9.. from Theorem 9..— I) (A. From Theorem 9. (A1).2)2.25. Then AAhas two possible JCFs (not counting reorderings of the a (A. 9. 2(A(A.(A 1). An. . 2)2.. 1 det(A) = det(X J XI) det(J) = n7=1 Ai. det(A) = nAi. " Xn. The minimal polynomial of a matrix is the product of the elementary divisors of highest degree corresponding to distinct eigenvalues. Then Theorem 9. Eigenvalues and Eigenvectors T.I [ "+ jfi 0 0 0 et + jf3 0 0 0 0 et .. and 2). 1). 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 .I)(A (A2)2. X XI.22 we have that A = X J X ~ .2). 1)4(A 2) and 2 2 et(A) = (A .
a(A. Then x is a right principal vector of degree k degree associated with A E A (A) ifand only if(A .e..a(A) = (A . so the eigenvalue 3 has two eigenvectors associated with it. Thus.A. we find that 2£2 + ~3 = O.e.27. An analogous definition holds for a left principal vector of degree k.. i. of course.7) for distinct A. it then has precisely one eigenvector. Definition 9.rank(A . is not sufficient to Example 9.l) independent right — — A.AI)klx i= o. it The straightforward case is.] are eigenvectors (and are independent).nxn number of eigenvectors. X e A(A) if and only if (A XI)kx = 0 and (A U}k~l x ^ 0. If we let [~l ~2 ~3]T associated If [^i £2 &]T denote a solution to the linear system (A — 3l)~ = 0.9.29. For example.. 1. of algebraic multiplicity 1.(7). when Ai is simple.) = (A.A. A e nxn ]R. suppose suppose A = [3 2 0 o Then Then 3 0 A3I= U2 I] o o 0 0 n has rank 1.).3.ulx = 0 and (A . Knowing TT (A.) = (A. we find that 2~2 + £ 3= 0 .e. 9. Determination of the JCF 9. three eigen7r(A.. and rank (A A. eigenvectors dimN(A — A. Remark 9.. The more interesting (and difficult) case occurs when Ai multiplicity A. both denote a solution to the linear system (A . c . associated independent right (or left) eigenvectors is given by dim A^(A . For each distinct eigenvalue Ai. The matrices A uniquely. a(A).7) = n .l). of course. we need the notion of principal vector.. of algebraic multiplicity 1.e. i.3/)£ = 0. To get a third vector JC3 such that X [x\ X2 XT.. and rank(A al) vectors.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. To get a third vector X3 such that X = [Xl KJ_ X3] reduces A to JCF. a)\ . the associated number of linearly A. determine the JCF of A uniquely.— al) == 4. three eigenboth have rr(A) = (A . X principal Definition 9. Let A E C"xn (or R"x").— a) and rank(A .27.is simple. i. Thus. is of algebraic multiplicity greater than one. both are eigenvectors (and are independent). left k.28.nxn). The straightforward case is. i. Remark 9. and rank(A —Ai l) for distinct Ai is not sufficient to rr(A). when X..29.28..— a) .3. Determination of the JCF 85 &5 Example 9.. a)7. a (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.
The case k = 1 corresponds to the "usual" eigenvector. 9. determine all eigenvalues of A e R" x " nxn ). The other solution necessary is the desired principal vector of degree 2. combination of jc(1) vectors to get a righthand side that is in 7£(A — XI). the principal vector second of degree 2: of degree (A . If.XI)0 = 0. For each independent jc (1) . Principal vectors are sometimes also called generalized eigenvectors. (It may be necessary to take a linear of x(l) R(A .A/)x(2) = x(l). principal vectors of degree 1) associated with A. the definition of principal vector is satisfied. First. solutions solutions to the homogeneous equation (A .X I ) . for get righthand example. Then the equation AX = XJ can be written that reduces a matrix A to this JCF. Denote by x(l) and x(2) the two columns of a matrix X E lR~X2 2x2 2 Jordan block{h0 h1. there is only one eigenvector. The phrase "of grade k" is often used synonymously with "of degree k. which simply says that x(!) is a right Ax(1) = hx(1) x (1) (2) x(2). Thus.A1)2 x(2) = (A . Denote by x(1) and x(2) the two columns of a matrix X e R2. S. since (A . wefind(A. solve (A .X I ) ( l ) = (A AI)O o.XI. by (A .'A1)22xx(l) = (A . different term will be assigned a much different meaning in Chapter 12. If the algebraic multiplicity of If A principal need X is greater than its geometric multiplicity. ji of dimension k or larger. If we premultiply (9. Theother solution (A . 2. of course.A1)X(l) = O. One of these solutions (A — AI)2 x (2) x(l) (1= 0). we find (A If we premultiply XI) x = (A XI)x = 0.e. (A — uf. A E A(A) following: (or C ). Then for each distinct X e A (A) perform the following: z (2) w c 1.XI)2x^ = 0. The number of eigenvectors depends on the rank of A .A1)x(2) = x(l). E lR nxn This suggests a "general" procedure. Eigenvalues and Eigenvectors synonymously "of 2.17) by (A . The second column yields the following equation for x . if of .AI). for example.3. Exercise 7. Solve (A .17) The first column yields the equation Ax(!) = AX(!).3. For example.XI). See.) .A/) = — multiplicity of rank(A — XI) = n . is. this rank is n . Then the equation AX = X J can be written A [x(l) x(2)] = [x(l) X(2)] [~ ~ J. k = eigenvector.2." "of often 3.AI).1 Theoretical computation Theoretical computation To motivate the development of a procedure for determining principal vectors. eigenvector. consider a determining 2 x Jordan [~ i]. there are two linearly independent n — o.x2 A JCF. x(l). if X. principal vectors still need to be computed from succeeding steps.1. See. I) associated This step finds all the eigenvectors (i.1 9. of The number of linearly independent solutions at this step depends on the rank of 2 (A . x(l) (^ 0). (9. 4. A right (or left) principal vector of degree k is associated with a Jordan block J. Thus. but the latter generalized eigenvectors..A1)X(l) = O. Eigenvalues and Eigenvectors Chapter 9. of — AI.86 Chapter 9. of k 5.
for example.31.1. A2 = 1. with the distinct eigenvalues 1 and 2. There are significant numerical difficulties inherent in attempting to compute a JCF. . . . . Then Theorem 9.. Principal vectors associated with different Jordan blocks are linearly indeTheorem 9. Continue in this way until the total number of independent eigenvectors and principal vectors is equal to the algebraic multiplicity of A. where the chain of vectors x(i) is constructed as above. Theorem 9. There are significant numerical difficulties inherent in attempting generally prove unreliable..3. 1 . 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. Determination of eigenvectors more extensive treatments. see. solve (A AI)x(3) 87 = x(2). Symbolic Toolbox. For more extensive treatments. X(k)]... Attempts to do such calculations in finiteprecision floatingpoint arithmetic or 3. of algebraic multiplicity and Theorem 9. this naturallooking procedure can fail to find all Jordan vectors. X(k)} is a linearly independent set. say). and h3 = 2. say). x(k)]. Notice that highquality mathematical software such as MATLAB readable [8] to learn why. For each independent X(2) from step 2.2/)x3(1)= 0 yields (A . 4.33. [20] and [21]. Continue in this way until the total number of independent eigenvectors and principal 4. . vectors is equal to the algebraic multiplicity of A.3. {x(l). Attempts to do such calculations in finiteprecision floatingpoint arithmetic generally prove unreliable. Notice that highquality mathematical software such as MATLAB does not offer a j cf command.30.(1) (A . this naturallooking procedure can fail to find all Jordan vectors. pendent. For each independent x(2) from step 2. Unfortunately. find the eigenvectors associated The eigenvalues of A are A1 = I. and A3 = 2. solve 3. For Unfortunately. Principal vectors associated with different Jordan blocks are linearly independent.. [20] and [21]. although a jordan command is available in MATLAB's Symbolic Toolbox.. . Let X = x ( l ) . see. h2 = 1.32. . .. for example. . (x (1) . 0 The eigenvalues of A are AI = 1. First. Theorem 9. Suppose A e Ckxk has an eigenvalue A. First. Theorem 9. and the interested student is strongly urged to consult the classical and very to compute a JCF. Determination of the JCF 9. Let X = [[x(l). x (k) } is a linearly independent set.2I)x~1) = 0 yields . Determination of the JCF 3.32. Let A=[~ 0 2 ] . find the eigenvectors associated with the distinct eigenvalues 1 and 2. Suppose A E C kxk has an eigenvalue A of algebraic multiplicity kkand suppose further that rank(A — AI) = k — 1. Example 9.30.33.31. and the interested student is strongly urged to consult the classical and very readable [8] to learn why. where the chain of suppose further that rank(A . although a j ardan command is available in MATLAB'S does not offer a jcf command.9. Let Example 9. .AI) = k . Then vectors x(i) is constructed as above.
we consider below the case of a single Jordan block.. . Then A 4l.. (1) toeet x. 0 !b. d. . For the sake of definiteness.2 9. Then Let D = diag(d" .1I)xl ) = xiI) to get (A – l/)x. d.. solve To find a principal vector of degree 2 associated with the multiple eigenvalue 1.88 (A .. consider below the case of a single Jordan block. dn be a nonsingular "scaling" matrix.11)x?J = 0 yields (A.so long as they are nonzero. solve 2 (A . . 0 = 0 A dn dn I 2 0 dn dn I A 0 ).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 .3.3.. but the result clearly holds for any JCF. For the sake of defmiteness. Suppose A € Rnxn and SupposedA E jRnxn and Let D diag(d1. but the result clearly holds for any JCF. 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. 0 0 D'(X' AX)D = D' J D = j ). Now let Now let X (2) =[ 0 ] ~ .. . . d n)) be a nonsingular "scaling" matrix.(2) = x. Eigenvalues and Eigenvectors To find a principal vector of degree 2 associated with the multiple eigenvalue 1.l/)x. we 1 's but can be arbitrary — so long as they are nonzero. =0 yields (1) Chapter 9..
.. Specifically. dimN(A — AJ)Vi = ni. where AS is defined as the set {As:: s e S}. Then jH.n = N(A = N (A .Amtm c and minimal polynomial a(A) = (A .35. ..A[)V) '" (A .A. Geometric Aspects of the JCF 89 di's Appropriate choice of the di 's then yields any desired nonzero superdiagonal elements. Such a decomposition is given in the following theorem.4 9. the reverseorder identity matrix (or exchange matrix) 0 p = pT = p[ = 0 I 0 (9. similarity transformation XD [d[x[.34.... Such a decomposition is given in the following associated direct sum decomposition of jH. Am distinct.. It is thus natural to expect an associated direct sum decomposition of R. j is obtained from A via the and principal vectors that reduces A to its JCF... where AS is defined as the transformation. Specifically. . Theorem 9. Let IF and suppose + transformation. (A .Am I) Vm . It is thus natural to expect an with respect to which the matrix is diagonal or block diagonal. . Suppose e jH.AlIt) E6 . Suppose A E R"x" has characteristic polynomial 9./) w = «.. set {As s E S}. A subspace S c V is A invariant if AS c S. Then AI.AmItm . Note that dimM(A ..Am)Vm with Ai..A1I) v) E6 ..... A subspace S ~ V is Ainvariant if AS ~ S. . Let V be a vector space over F and suppose A : V —>• V is a linear Definition 9.35. In a similar fashion. .4. .nxn n(A) = (A .9. dnxn}.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...xn]] of eigenvectors = [x[. x n eigenvectors and principal vectors that reduces A to its lCF.34.. Definition 9.nxn (or nxn to JCF provides change of basis with respect to which the matrix is diagonal or block diagonal. E6 N(A ..n.A[)n) . E6 N (A . interpreted This result can also be interpreted in terms of the matrix X = [x\... . mdistinct. . the reverseorder identity matrix (or exchange matrix) In a similar fashion. J is obtained from A via the similarity transformation XD = \d\x\. Geometric Aspects of the JCF 9. dnxn]....4. A.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.
34."" Jik.) and each Jik is a Jordan block corresponding to A. 7.34. the eigenvectors and principal vectors associated with A.A.e...as in Theorem 9.. . so by (9. then S <S is Ainvariant if and only if there exists M E ]Rkxk such that eRkxk (9..Xm] ] Ee]R~xnxnisis such that X^AX ==diag(J1. we have that A A.. .90 Chapter 9. If V is a vector space over IF such that V = N\ EB . We would then get a block diagonal representation for A with full blocks rather than the highly structured Jordan blocks. Jm).36. Eigenvalues and Eigenvectors Chapter 9.2. Jm)..) span an Ainvariant of A".Ji ..2. Rewriting in the form ~ J.. where X [ X i . Then N(p(A)) and 1. the eigenvectors and principal vectors associated with Ai) span an Ainvariant subspace of]Rn. Ainvariant. e E"x". Let Yi E <enxn .. K(S) <S.38.. . e A(A). and S e R" xk s\. via SVD).. .. Finally. Suppose X block diagonalizes A. partition . //*. i. Note that AXi = A*. Xm R"n such that XI AX diag(7i...i. Sk If R" R..e. = X. i. so the columns of A.e.37. we have that AXi Theorem 9. Equivalently. If F = NI ® • • 0 m A// is Ainvariant. but we restrict our attention here to only the Jordan block case. Note that A A".. Eigenvalues and Eigenvectors If V is taken to be ]Rn over Rand S E ]Rn x* is a matrix whose columns SI. The equation Ax Example 9. is Ainvariant. where each Ji = diag(/. Example 9.lt. A". Let peA) = «o/ + o?i A + • + <xqA be a polynomial in A.19): /th Example 9../)"' by SVD..li. /. . Let 7. S is Ainvariant if and only if S .g. This follows easily by comparing the ith columns of each side of (9. each Ji = diag(JiI.). 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..19) the columns of Xi (i. s/t span a /^dimensional subspace <S. Then N(p(A)) and R(p(A)) 7£(p(A)) are Ainvariant. we could choose bases for N(A — A.. i /= 1. . Other such "canonical" forms are discussed in text that follows. The Jordan canonical form is a special case of the above theorem. so by (9. Let p(A) = CloI + ClIA + '"• •+ ClqAqq be a polynomial in A. i. 9. of W.36. XI AX = [~ J 2 ]. If A has distinct The Jordan canonical form is a special case of the above theorem.. be a Jordan basis for N (AT . R(S) == S. 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).39..39.. . Other representation for A with full blocks rather than the highly structured Jordan blocks... 9. We would then get a block diagonal example (note that the power ni could be replaced by Vi). for N(A . = Xi.. If A has distinct eigenvalues A.. example (note that the power n.38.• EB Nm.) and each /.. partition Equivalently...Ai/)n. could be replaced by v.19) the columns attention here to only the Jordan block case. diagonal representation. /.is A T invariant.19) AS = SM. .e../)"'. .span an Ainvariant subspace. then a basis for V can be chosen with respect to which A has a block diagonal representation.. A invariant if only ifS1 1. AT Theorem 9. (i. € C"x"' be a Jordan basis for N(AT — A.e. 2... so the columns of Xi span an Amvanant subspace. such "canonical" forms are discussed in text that follows.. Suppose A"== [Xl . eigenvalues Ai 9. Suppose A E ]Rnxn.. is not necessarily diagonalizable. then a basis for V can be chosen with respect to which A has a block N. where each Theorem 9. We could also use other block diagonal decompositions (e. then is Ainvariant if and only if there span a kdimensional subspace S. = 1.* is a Jordan block corresponding to Ai E A(A).37.
is given by eigenvalues in the right halfplane. for a k x k Jordan block 7..41. of defined Definition 9.41.. Then the sign of A.9. Then compatibly. m ••• . Then A = XJX.40.5. Suppose A E C"x" has no eigenvalues on the imaginary axis.I = XJy H = [XI. A called the matrix sign function.= Ai. Let z E C with Re(z) ^ O.YiH. Ym]H = LX..S. The Matrix Sign Function 91 91 compatibly.40.JiYi . 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 . It is a generalization of the sign (or signum) of a scalar.. 9. Definition 9. and let e cnxn be a Jordan canonical form for A. . Then the sign of A.5 9. ifRe(z) < O. Jm) [YI . 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.. A survey of the matrix sign function and some of its applications can be found in [15]. . The Matrix Sign Function 9. is given by sgn(A) = X [ / 0] 0 / X I .. . Definition 9. Xm] diag(JI. denoted sgn(A). . Definition 9. i=1 H In a similar fashion we can compute m etA = LXietJ.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. It is a generalization of the sign (or signum) of a scalar. . Then the sign of z is defined by Re(z) {+1 sgn(z) = IRe(z) I = 1 ifRe(z) > 0. E f= 0. associated with an eigenvalue A. denoted sgn(A).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.
R(S — /) Ainvariant of (the negative invariant subspace).. 6. respectively.. ± 1. sgn(cA) = sgn(c) sgn(A) for all nonzero real scalars c.42. 5. .xn and left eigenvectors Yl. and let = sgn(A). Let e C" be an arbitrary vector. Theorem 9. Let A E Cnxn have distinct eigenvalues AI.43. S = sgn(A). sgn(TlAT) = T1sgn(A)T foralinonsingularT E C"x". . 2. sgn(AH) = (sgn(A))". and let — sgn(A). Xn and left eigenvectors y\. 4.. distinct right eigenvectors Xi.. sgn(T1AT) Tlsgn(A)TforallnonsingularT e enxn 6. AS = SA. respectively. yn. AS = SA. Yn. Theorem 9.. Then the following hold: following 1. S = sgn(A)..92 92 Chapter 9. The JCF definition of the matrix sign function does not generally lend itself to reliable computation on a finitewordgenerally itself length digital computer. The JCF definition of the here is especially useful in deriving many of its key properties. projection subspace of 4. Suppose A E C"x" has no eigenvalues on the imaginary axis. . 5. We state some of the more useful properties of the matrix sign function as theorems. Show that v can be expressed (uniquely) as a linear combination arbitrary vector.n 1. 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).. negA == (l .S) /2 is a projection onto the negative invariant subspace of A. sgn(cA) = sgn(c) sgn(A)/or c. Their straightforward proofs are left to the exercises.. . ).. . but the one given There are other equivalent definitions of the matrix sign function. Xn with corresponding right eigenA e nxn ). positive = (/ + of A. Their left exercises.. 3. In fact. 7l(S l) is an Ainvariant subspace corresponding to the left halfplane eigenvalues left halfplane I.. but the one given here is especially useful in deriving many of its key properties. positive of P. Find the appropriate expression for v as a linear combination expression of the left eigenvectors as well. S2 = I. 2.. Show that v can be expressed (uniquely) as a linear combination e of the right eigenvectors. 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). posA == (l + S)/2 is a projection onto the positive invariant subspace of A. .43. its reliable numerical calculation is an interesting topic in calculation its own right. respectively. 3. There are other equivalent definitions of the matrix sign function. ••. ••• .42. sgn(A") = (sgn(A»H. Suppose A E enxn has no eigenvalues on the imaginary axis. Then the following hold: following e 1. 3. We state some of the more useful properties of the matrix sign function as theorems. EXERCISES EXERCISES 1. S2 = I.1> . e nxn Theorem 9. 4. Eigenvalues and Eigenvectors Chapter 9. 2. e C"x" Theorem 9. negA = (/ — S)/2 3. Let v E en be an vectors Xl. S is diagonalizable with eigenvalues equal to del..
Show that x is also a left eigenvector for A.22. n are nonzero vectors with with xTTyy = 0. AH = A. Suppose A E C"x" is skewHermitian. 3}. 2. 5. but then the equation (A . where x. Show that x is also a left eigenvector for A. Determine the JCF of A.e. k .1) element of J. What are the eigenvalues of this slightly perturbed matrix? matrix? . Suppose a matrix A E lR. 3}. Determine the JCF of A. y E lR. Suppose a matrix A E R 16x 16 has 16 eigenvalues at 0 and its JCF consists of a single A e lR. right eigenvectors and right principal vectors if necessary. Prove that all eigenvalues of a skewHermitian matrix must be pure imaginary.l]r as an eigenvector.22. Let 7. Let A e R" xn be of the form A = 1+ xyT.16 Jordan form specified 9. eigenvectors and if and (real) JCFs of the following matrices: (a) 2 1 ] 0 ' [ 1 6./)jc = x can't be solved. Determine the JCF of A.. x. Determine all possible € R 5x5 {2. Let A = [H 1]· 2 2" Find a nonsingular matrix X such that XI AX = J. 3. 4. i. 10. Determine the JCF of A. y E lR. Prove that all eigenvalues of 2. Characterize all left eigenvectors.. y e R" are nonzero vectors 10.e. 11. Let A be an eigenvalue of A with corresponding 3. if A is skewHermitian. JCFs for A. Suppose A e rc nxn is Hermitian. JCFs for A. where J is the JCF 1 J=[~ 0 1~]. 16x 16 has eigenvalues at 0 its JCF consists of single Jordan block of the form specified in Theorem 9.nxn A = xyT. a skewHermitian matrix must be pure imaginary. nxn be of the form A = / + xyT. where x. Prove the same result if A is skewHermitian. multiples of e\ E lR.30 must be 8. Let A E lR. Let A e R"x" be of the form A = xyT. Show that all right eigenvectors of the Jordan block matrix in Theorem 9. ~ 0 Hint: Use[1 1 — I]T an Hint: Use[— 1 1 . Let A be an eigenvalue of A with corresponding right eigenvector x. Prove the same result right eigenvector x. The vectors [0 1 Ifand[l 0 of [0 — l] r and[1 0]r (2) (1) are both eigenvectors. What are the eigenvalues of this slightly perturbed is added to the (16. 5. Characterize all left eigenvectors.30 must be multiples of el e R*. 2. where J is the JCF Find a nonsingular matrix X such that X AX = J. = O. AH = —A. where x. 9.n T T x yy = 0. Suppose A € rc nxn is skewHermitian. Determine the JCFs of the following matrices: 6. 2. y e R" are nonzero vectors with A E lR.Exercises 93 93 2. (A — I)x(2) x(1) 8. Determine the JCFs of the following matrices: <a) Uj n 2 1 2 =n 7. Determine the eigenvalues. i.1) element of J. Suppose 10~16 is added to the (16. x O. Suppose the small number 10. Suppose A E C"x" is Hermitian.5x5 has eigenvalues {2. eigenvalues. Show that all right eigenvectors of the Jordan block matrix in Theorem 9.
what can you say further. TIAT = [A011 A22 0 ] . Show that every matrix A e jRnxn can be factored in the form A Si$2.18) is useful. 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. say S1. 14. say Si. about when the equation for X is what can you say further. Suppose Au =1= 0 and that we we e jRnxn and All e jRkxk 1 < ::s: n. Suppose A e sgn(A) = 1. Hint: Use the factorization in the previous exercise. Prove Theorem 9. Then = ( X SIXT)(X. where Si 12.43. Prove that 17. 13. Find a matrix equation that X must satisfy for this to be possible. 15. about when the equation for is solvable? solvable? 15. xn has all its eigenvalues in the left halfplane. is nonsingular.18) is useful.e. 16. Prove that every matrix A E W x" is similar to its transpose and determine a similarity 13. Eigenvalues and Eigenvectors Chapter 9.42. Hint: Use the factorization in the previous exercise. 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).S2X~l) would required symmetric factorization of A. Then A = (XS i X T ) ( X ~ T T S2XI) would be the the "symmetric factorization" of J. 16. is nonsingular. Eigenvalues and Eigenvectors 12. Prove Theorem 9. and S2 are real symmetric matrices and one of them.. Prove that every matrix e jRn xn is similar to its transpose and determine a similarity transformation explicitly. Prove that 17.42. en . X e R*x <«*). Prove Theorem 9. i. where SI and £2 are real symmetric matrices and one of them. sgn(A) = /. Consider the block upper triangular matrix 14. If n = 2 and k = 1.94 Chapter 9. The transformation P in (9. If n = 2 and k = 1. Thus. JCF. Show that every matrix A E R"x" can be factored in the form A = SIS2. it suffices to prove the result for the JCF.43. The transformation P in (9. 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. Find a matrix equation that X must satisfy for this to be possible. Suppose A E C"xn has all its eigenvalues in the left halfplane. transformation explicitly. Prove Theorem 9. it suffices to prove the result for the required symmetric factorization of A. in terms of All and A22. in terms of AU and A 22. Thus.
2. then there exists a unitary matrix U such that UH AU — D. skewsymmetric. !] Theorem 10. Let A = AH e C"x" have (real) eigenvalues A. .. Normal matrices include Hermitian.2. most "diagonal" we can get is the JCF described in Chapter 9. if A e Rmxn . Then there AI. ." In matrix terms. If a matrix A is not normal. respectively).. Q are unitary. . the for real scalars a and h. Xn Theorem 10. . matrix if and only if it is normal (i." In matrix terms. If The following results are typical of what can be achieved under a unitary similarity. An). and orthogonal. n ).. A. the transformation A f+ PAPI is called aasimilarity.. = H E en xn exists a unitary matrix X such that X H AX = D = diag(Al. 95 95 . An. such as A = [_~ most "diagonal" we can get is the JCF described in Chapter 9. . the transformation A f+ P ApT is called 2. If W = V and <2== p. Normal matrices include Hermitian.1. orthogonal equivalence if P and are orthogonal matrices.9.. AA = AHA). 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. An) (the columns of X are exists a unitary matrix X such that XHAX = D = diag(A.1. . respectively). and unitary matrices (and their "real" counterparts: symmetric.1. as well as other matrices that merely satisfy the symmetric. AAHH = AH A). We can also consider the case A E emxn and unitary equivalence if P and <2 are unitary.. 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.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. where D = diag(A.. an orthogonal similarity (or unitary similarity in the complex case).9. . . skewHermitian. where D = diag(AJ.. Problem: Let V and W be vector spaces and suppose A : V —>• W is a linear transformation...e. We can also consider the case A e Cm xn and unitary equivalence if P and Remark 10.1 10. If a matrix A is not normal.. the transformation A i» PAPT is called If an orthogonal similarity (or unitary similarity in the complex case). The following results are typical of what can be achieved under a unitary similarity. where it is proved that a general matrix A E C"x" is unitarily similar to a diagonal 10. then there exists a unitary matrix £7 such that A = AH E en xxn has eigenvalues AI." The transformation A M» PAQ is called an equivalence. Xn) (the columns ofX are orthonormal eigenvectors for A). = V and if pT is orthogonal. such as A = [ _ab ^1 for real scalars a and b. .2.. If W = V and if Q = PT is orthogonal. . An..2. If A = A H 6 C" " has eigenvalues AI. This is proved in Theorem 10. as well as other matrices that merely satisfy the definition. find P e lR. the transformation A H> PAP" 1 is called similarity.. V and Q 1.e. and orthogonal. 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... Two special cases are of interest: Two special cases are of interest: 1. the definition. if A E IR mxn find E R™ and Q E lR~xn such that P AQ has a form. . What other U HAU = D.An.Chapter 10 Chapter 10 Canonical Forms Canonical Forms 10. . it is called an "canonical form. it is called an orthogonal equivalence if P and Q are orthogonal matrices.:xm and Q e Rnnxn such that PAQ has a "canonical form. . skewskewHermitian.j.j.. Remark 10. This is proved in Theorem 10. .I... . orthonormal eigenvectors for A). . If P"1 .. ." The transformation A f+ P AQ is called an equivalence. and unitary matrices (and their "real" counterparts: symmetric.
%n] XI . Canonical Forms Chapter 10. [XI U2] is unitary. We illustrate the construction of the necessary Householder matrix for k — 1.2)block by XI Xz. the construction of X2 E JRnx(nl) such that X — z e ]R" (".. [£i. . VI € Cnxk [Xi U ] Proof: Let X\I = [x\.1) (10.2 is then a special case of Theorem 10. . .. Then there exist n . ..k But the latter are orthonormal since they are the last n .Hv. Let X\ e Cnxk have orthonormal columns and suppose U is a unitary Theorem 10. 0 Thus. Then VH = / / . X 1 XI e E". xn] = [x\ ] [XI X22] is unitary. xd = [ ~ l U = where R is upper triangular (and nonsingular since x\. . Write V H matrix such that V X I = [ ~]. n . D 0 (2.3 called Theorem 10. —k U. . we consider the real case. xn such that X = (XI. Write UH = [U\ U ] [VI Vz] 0 2 with Ui E Cnxk .1) we have used the fact that Ax\ = AIXI. Canonical Forms Proof: eigenvector corresponding AI. where R E kxk is upper triangular. Let the unit vector x\ be denoted by [~I.. orthogonal (l. In (10. we get 0 in the (l.. . we get Ai remaining in the (l.. 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 V2.2) Al X~AX2 XfAX 2 0 Al ] 0 XfAX z 0 l In (10.. Xk H Hk. . Thus. X = Given a unit vector x\ E JRn.96 96 Chapter 10. .1) we have used the fact that AXI = k\x\. xn] = 1.. Construct a sequence of Householder matrices (also known HI. (l... Xn such that [x\. .. where R € Ckxk is upper triangular. following general result. Let XI E Cnxk have orthonormal columns and suppose V is a unitary matrix such that UX\ = \ 1.. . . An. .l)block by Al (2...T. Then there exist n — 1 additional vectors X2. k 1. . ..2)block must have eigenvalues A2. D The construction called for in Theorem 10...2). . .. 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.• • Hk and Hk'" HI.. Now XHAX =[ xH I XH ] A [XI 2 X 2] =[ =[ =[ x~Axl XfAxl X~AX2 XfAX 2 ] (10. and normalize it such that x~ XI = XI 1. .3. I)block x"xi = 1. The proof is completed easily by induction upon noting proof that the (2.2)block noting that x\ is orthogonal to all vectors in X2. Hk as elementary reflectors) H\... simplicity.. xf*x\ = Proof' Let x\ be a right eigenvector corresponding to X\. . .. Hk in the usual way (see below) such that Hk ...1 additional vectors x2..3 for k = 1. ~nf... XH AX induction noting that XH AX is Hermitian.. Xk]..l)block... When combined with the fact that In (l0... Let V = XI. (/ € k) U2 X i U2 = Xi .3. .. k = For simplicity. Then [XI V 2] is unitary... [Xi f/2] unitary. HdxI. 10. Xk are orthonormal). Construct a sequence of Householder matrices (also known Proof: Let X [XI...xd. A.2)block X2 .k rows of the unitary matrix U.. We also get 0 in the (2. . When combined with the fact that x~ XI = 1. .) [XI X2]] is orthogonal is frequently required.. ..I)block.
2 is worth stating separately since it is applied frequently in applications..•» '.It can easily be checked — 2uu+ — u u T .2 for Hermitian matrices) can be written n A = XDX T = LAiXiXT.. The real version of Theorem 10..e. quently in applications.1..2 is worth stating separately since it is applied fre10.. [23]. U effects necessary compression of jci.. X n ). x where P. it is easily verified that UT U = 2 ± 2'.2. Further details on Householder matrices. . n A = LAiPi. = PUM = xixf = xxixT since xj xi — 1. . '. [11]. A in (10.4.. necessary compression of Xl. .. .4 implies that a symmetric matrix A (with the obvious analogue from Theorem 10.e.10. [11]. Then there exists an AT E jRnxn have eigenvalues AI. 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. . [25]. Some Basic Canonical Forms 10. A Note that Theorem 10. An.. . Thus. . An). Theorem 10. 's).3) spectral which is often called the spectral representation of A.1. . it is easily verified that u T u = ± 2£i and u T Xl = 1 ± '. To see that U effects the U symmetric U U = U = I. £«] r It can checked T 2 that U is symmetric and U TU = U 2 = I. Then there exists an 10. i=1 (10. where u ^UU [t\ 1.Xn.2 for Hermitian matrices) can be written from Theorem 10. including the choice of sign and the complex case. . where Pi = PR(x. [25]. U orthogonal. so U is orthogonal. where u = ['.i. sponding to the Ai'S)... XTAX = D = diag(Xi. £2. 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.nf. . [23]. 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 . (onto the onedimensional eigenspaces correPi onedimensional eigenspaces sponding to the A.2uu+ = I . consulted standard numerical linear algebra can be consulted in standard numerical linear algebra texts such as [7]. • • . Let A = AT e E nxn have eigenvalues k\.) — xiXt = i j since xT Xi = 1.1.1 and UT X\ = 1 ± £1.+uu T . In fact..3) is actually a often weighted sum of orthogonal projections P.4. . i.1 ± 1. [7].
AH A = AAH ). The matrix 10. Example 10. Let A E R"xxn.8. it is of interest to know While every matrix can be reduced to Schur form (or RSF). where S is quasiuppertriangular.7... However.6 is called a real Schur canonical form or real Schur form (RSF). The quasiuppertriangular matrix S in Theorem 10.2)block wf AU2 is not 0.6 is called a real form or Schur fonn. Proof: Suppose U is a unitary matrix such that U H AU = D.e. Proof: The proof of this theorem is essentially the same as that of Theorem lO.e. it is thus unitarily similar to an upper triangular matrix.. 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. and sufficient for virtually all applications (see. but if A has a complex conjugate pair of eigenvalues.6 (MurnaghanWintner). Let A E cnxn Then there exists a unitary matrix U such that Theorem 10. matrix U that reduces a matrix to [real] Schur form are called Schur vectors. AHA = AA H). matrix U such that U AU = S. UH AU = T.98 98 Chapter 10. However. where T is upper triangular. for example.5 is called a Schur canonical form or Schur form. 0 in this case (using the notation U rather than X) the (l. A matrix A e c nxn is unitarily similar to a diagonal matrix if and only if A is normal (i. where S is quasiuppertriangular. but In the case of A E R"xxn .5 (Schur). what is true. . the next theorem shows that every A E IR xn is also orthogonally similar (i. [17]). The triangular matrix T in Theorem 10. complex conjugate pairs of eigenvalues. However.2 except that in this case (using the notation U rather than X) the (l. The quasiuppertriangular matrix S in Theorem 10. where D is diagonal.2)block AU2 is not O. The matrix s~ [ 2 0 2 5 4 0 is in RSF. D ur In the case of A e IRn ". 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. Theorem 10. A matrix A E C"x" is unitarily similar to a diagonal matrix if and only if Theorem 10. the next theorem shows that every to place such eigenvalues on the diagonal of T. A quasiuppertriangular matrix is block upper triangular with 1 x 1 diagonal matrix. where T is upper triangular. 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. and sufficient for virtually is real if U is orthogonal) is an eigenvector. Then there exists a unitary matrix U such that U H AU = T. The triangular matrix T in Theorem 10.e. real arithmetic) to a quasiuppertriangular matrix. Canonical Forms Theorem 10. The columns of a unitary [orthogonal] matrix U that reduces a matrix to [real} Schur fonn are called Schur vectors. A is normal (i. Let A e C"x". then complex arithmetic is clearly needed if A has a complex conjugate pair of eigenvalues. for example. Then there exists an orthogonal Let A e IR n ". where D is diagonal. Then there exists an orthogonal 10.2 except that Proof: The proof of this theorem is essentially the same as that of Theorem 10.8.9. Its real JCF is is in RSF. so A is normal. real arithmetic) to a quasiuppertriangular A e Wnxn is also orthogonally similar (i. following theorem answers this question. 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). it is thus unitarily similar to an upper triangular matrix. Theorem 10. Canonical Forms Chapter 10.9.e. then complex arithmetic is clearly needed to place such eigenValues on the diagonal of T. Definition 10. diagonal of T (or S). Then Proof: Suppose U is a unitary matrix such that U H AU = D. [17]). The when we can go further and reduce a matrix via unitary similarity to diagonal form. The columns of a unitary [orthogonal} Schur canonical form or real Schur fonn (RSF).. Then AAH = U VUHU VHU H = U DDHU H == U DH DU H == AH A so A is normal. what is true.5 (Schur). However.6 T T matrix U such that U AU = S. is that the first Schur vectors span the same all applications (see.5 is called a Schur canonical Definition 10. The following theorem answers this question.7.
12 ~ AlyH Y = AIX HX . nonpositive definite (or negative semidefinite) if A is nonnegative definite. e Theorem 10. U diagonalizes A 10. 111. if—A 4. 3. we write A > B if and only if A — B > B . We write A < O. If a matrix is neither definite nor semidefinite.2 Definite Matrices Definite Matrices Definition 10. Definite Matrices 10. Proof: Proof: Let U be a unitary matrix that diagonalizes A as in Theorem 10. If A E C"x" is Hermitian.=1 .A is positive definite. this section that may be stated in the real case for simplicity. 2. If neither semidefinite.n. nonzero x E lR.12. if A and B are symmetric matrices. where T is an upper triangular matrix (Theorem 10. 11'/. CM j]i. it is said to be indefinite. suppose A is normal and let U be a unitary matrix such that U H AU = T. i En. We (or negative if— A nonnegative definite. in fact. Indeed. A symmetric matrix A E lR.n • We write A :::: 0. We write A ~ 0. Then for all E en. Furthermore.2.13. Then n x HAx = (U HX)H U H AU(U Hx) = yH Dy = LA.2 10. Thenfor all Let A = AH E Cnxn with eigenvalues AI > A2 > • • > An. positive definite if and only if xTT Ax > 0 for all nonzero x G lR..10. and denote the components of y by v UHx. write A < O.10.10.A < O. let y = U H x. we write A > B if and only if A . It T 0 D 10. negative positive definite. Also. Then T (Theorem It is then a routine exercise to show that T must. Definite Matrices 99 Conversely. B — A < 0.12. We write A > O. indefinite. we write A > B if and only if A . We write A < 0. be diagonal.=1 But clearly n LA. .11. this is generally true for all results in the remainder of of superscript //s Ts.nxn is Definition 10. A U U HA U T.2. superscript H s replace T s. Let A = AH e enxn with eigenvalues X{ :::: A2 :::: .11. we write A :::: B if and only ifA — B>QorB — A < 0.A ~ O. Similarly. i € n.5).• :::: An.2..2.B :::: 0 or B . 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. Similarly. Remark 10.. all the above definitions hold except that A e nxn Remark 10.. where x is an arbitrary vector in en. x eC". Then 11. Furthermore. negative definite if . if A and B are symmetric matrices. We write A > 0. Remark 10.. A symmetric matrix A e Wxn 1.B > 0 or or Also. positive definite if and only ifx Ax > Qfor all nonzero x E W1 We write A > O.
Theorem 10. If A = AH e C"x" is positive definite. determinant the determinant of the 2x2 2 leading submatrix is also 0 (cf. All eigenvalues of A are positive. Then 111~~1~2 Let jc be an eigenvector corresponding to Xmax(AHA).17).100 100 Chapter 10.1. 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. of obtained and E ~nxn positive definite if and only if any of the Theorem 10. ::::: AI. whence Ar1ax (A A).@mllrk 10. All eigenvalues of A are nonnegative. The determinant of the 1x1 1 leading submatrix is 0 and 1. The determinants of all principal submatrices of A are nonnegative. 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. from which the theorem follows.= Amax{A A).. 2.15.18.18.. 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 Proof: E C" Proof: For all x € en we have Let x be an eigenvector corresponding to Amax (A HA).I. . Canonical Forms 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. not just those of the leading principal submatrices.17.17). A can be wrirren in [he/orm MT M. Theorem 10. However. where M 6 R ix " and k > rank(A) "" rank(M). The determinant of the I x leading submatrix is 0 and consider the matrix A = [~ 2x 0 (cf. 3. Note that the determinants of all principal "ubm!ltriC[!!l mu"t bB nonnBgmivB R. Theorem 10. 3. Let A e C"x". 3.13 provides (A 1) Rayleigh quotient of jc.. Corollary Corollary 10.w) x HAx > the Rayleigh quotient. of positive. 2. XHAx > 0 for all nonzero = AH E enxn E en. xfO IIxll2 I 0 Definition submatrixofan n x k) x k) Definition 10. where M e R"x" is nonsingular. Canonical Forms and and n LAillJilZ::: i=l AnyHy = An xHx . the .l3 provides upper (AO and lower (An) bounds for (A. For example. A can be written in the form MT M. x E C". where M E IRb<n and k ~ ranlc(A) — ranlc(M).18.19. 0 D Remark XHHAx Remark 10. Not@th!ltthl!dl!termin!lntl:ofnllprincip!ll eubmatrioes muet bQ nonnogativo in Theorem 10. All eigenvalues of A are nonnegaTive. Remark 10. Then IIAII2 = ^m(AH A}. of all principal submatrices of 2. form MT E ~n xn E ~n xn definite if and only if Theorem 10. consider the matrix A — [0 _l~]. Let A E enxn Then \\A\\2 = Ar1ax(AH A).16. Theorem 1O. A leading principal submatrix of order n — k is obtained by deleting the last k rows and columns.14. so 0 An ::::: .19. Then ^pjp2 = ^^(A" HA). A can be written in the form MT M.1. whence IIAxll2 ! H IIAliz = max .soO < X n < ••• < A. Theorem 10. The determinants of all leading principal submatrices of A are positive.
For example. Hermitian case.20.. E <C Theorem 10. MT AM> M.2) element is. basic definitions. if A E lR. For example. E jRnxn MT AM > M BM. Let A E lR. If >BandMe jRnxm. [ fz ti o o l [~ 0] ~ 0 v'3 . definite if A is positive definite).B is nonnegative definite. That is.is a square root. Remark 10. where L\ e c(nl)x(nl) is nonsingular and lower triangular as = L1Lf. Write the matrix A in the form the form By our induction hypothesis.. Theorem 10.18. then MT AM :::: MTTBM. Its proof is straightforward from theorem is useful in "comparing" symmetric matrices. Then A has unique nonnegative Theorem 10. For example. j proof (see. If A > B and M e Rm .17 is available and is known as the Cholesky factorization. negative and A is nonpositive principal submatrix consisting of the (2. In general. if = /2. Let A. then MT AM > MT TBM. Ll E C1""1^""^ and . The following theorem is useful in "comparing" symmetric matrices. any matrix S of c e s 9 the " °* ™ the form [ ssinOe _ ccosOe ] IS a square root. any matrix of nonsymmetric) have infinitely many square roots. if then M can be then M can be [1 0].22. SA = AS and rankS = rank A (and hence S is positive = AS S S. That is.21.22.2. if A = lz. Then there exists a positive definite. The factor M in Theorem 10. if Remark 10. and positive definite. with positive diagonal elements such that positive Proof: The proof is by induction. concerns the notion of the "square root" of a matrix. for example.3 is not unique. p. assume the result is true for matrices of order n . BM. matrices (both symmetric and square root of if S A. [16. Definite Matrices 101 101 principal submatrix consisting of the (2.17 is available and is A stronger form of the third characterization in Theorem 10. [16. 1f A :::: Band M E Rnxm.23. matrices (both symmetric and nonsymmetric) have infinitely many square roots. nxn Theorem 10. rankS = rankA definite definite if positive definite). negative and is nonpositive definite. The factor M in Theorem 10. if € E" xn we say that e jRn x that S E R nxn"isisa asquare root of AA ifS2 2 =— A. 0 Recall that A :::: B if the matrix A . A e R"x be nonnegative definite.23. The following Recall that A > B if the matrix A — B is nonnegative definite. Theorem 10. Then A has aaunique nonnegative definite square root S. nxm 2. for example. B e Rnxn be symmetric.20.2. For example. in fact. Its proof is straightforward from basic definitions. It concerns the notion of the "square root" of a matrix. assume the result is true for matrices of order — 1 so that B may be written as B = L\L^. The case n = 1 is trivially true.nxn"be nonnegative definite. = LLH. The case = is trivially true.2) element is. Let A e c nxn be Hermitian unique nonsingular lower triangular matrix L nonsingular A = LLH. In general. p.1 so that B By our induction hypothesis.18. Moreover. Definite Matrices 10. If A> Band E jR~xm.we say 181]).nxn . 1. 2. The following standard theorem is stated without proof (see.. 10rm [COSO _ Sino] . in fact. Write the matrix A in Proof: The proof is by induction. standard theorem stated 181]). A stronger form of the third characterization in Theorem 10. It is stated and proved below for the more general Hermitian case. . It is stated and proved below for the more general known as the Cholesky factorization.10.3 is not unique.
4) [7. Ch. we find by L^b. Alternatively.• Clearly we see we L I C = b and ann = c HC a 2 c is given simply by c = C. Then [ Sl o 0 ] [ I Uf U H ] AV = [I 0 0 ] 0 . Substituting in the expression involving a.4). of ann — b 0 root of «„„ . Substituting in the involving we find 2 a2 = ann . Two such forms are stated here. Let A € C™*71. 0 Note that the greater freedom afforded by the equivalence transformation of Theorem afforded 10. say. suppose A has an SVD of the form (5. 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. However. [21. multiplication where a is positive.3 Equivalence Transformations and Congruence Equivalence Transformations and Congruence Theorem 10. for example.lb. numerical procedures for computing such procedures an equivalence directly via. we must have ann —bHB lb > 0. But we = ann — b LIH L\lb = ann — bH B~lb B A).3 10. Choosing a to be the positive square ann .p.24. However.24.4) and the SVD. of course. Ch. are generally unreliable. .b H B1b (= the Schur complement of B in A).4) efficiently available. Gaussian or elementary row and column operations. They are more stably computable than (lOA) and more efficiently computable than a full SVD. for example (10. 131]. The numerically preferred equivalence is. ann Since det(B) > 0. Take P =[ S~ 'f [I ] and Q = V to complete the proof. Alternatively. But know that o < det(A) = det [ ~ b ] = det(B) det(a nn _ b H B1b). Performing the indicated matrix multiplication and equating the corresponding submatrices.b HL\H L11b = ann .2) in its complex version. Canonical Forms with positive diagonal elements. [21. as opposed to the more restrictive situation of a similarity transformation. yields a far "simpler" canonical form (10.b H B1b > O.131]. see. Many similar results are also (10..b B1b completes D 10. Canonical Forms Chapter 10. available. [4. we see that we must have L\c = b and ann = CHc + a 2. Then there exist matrices P E C: xm and Q e C"nx" such E c. the SVD is relatively expensive to compute and other canonical forms exist that are intermediate between (l0.xn. 5]. Then E c~xn such exist e C™ x m that that PAQ=[~ ~l (l0. p.4) Proof: proof Proof: A classical proof can be consulted in. 2].102 102 Chapter 10. Choosing a be det(fi) > HB~lb completes the proof. the unitary equivunitary alence known as the SVD.
Definition 10. for example.xr E erx(nr) arbitrary general nonzero. and zero eigenvalues. Theorem 10. Remark 10. a congruence. It turns out that the principal property so preserved is the sign of each eigenvalue.e.25 (Complete Orthogonal Decomposition). . 0). nxn E e X E e~xn. for example.xr is upper (or lower) triangular with positive diagonal elements. We then have the following. Let A e Cnxn and X e Cnnxn.26. if A is Hermitian. The H. We then have the following. Again. 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. Let A e e~xn. see. The transformation A i> XH AX is called a congruence. Then H HAX). When A has full column rank but is "near" a rank deficient matrix. Let A E C™ x ". and ~ denote the numbers of positive. 0 D x Theorem 10. D Proof: For the proof. and £ denote the numbers of positive. see [4]. In(A) = ln(X Proof: For the proof.. v. where R E Crrxr is upper triangular and S e C rx( " r) is arbitrary but in general nonzero. then rank(A) = n + v.0. then XH AX is also Hermitian. 134]. Theorem 10. It is of interest to ask what other properties of a matrix are then X H AX is also Hermitian. Then there exists a unitary matrix Q e e mxm and a Theorem 10. [21. Then the inertia of A is the triple of inertia of of negative.3. respectively. v.rrxr is upper (or lower) triangular with positive diagonal elements.1).29. Then is the numbers In(A) = (rr. 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. see [4]. 134]. see [4]. Let A e C™ ".5) where R E e. Note that congruence preserves the property of being Hermitian. 0 2. respectively. see [4]. Definition 10. In(A) = In(X AX).26. Equivalence Transformations and Congruence 10. HE C xn E e~ xn. When A has full column rank but is "near" a rank deficient matrix.31 guarantees that rank and signature of a a matrixare preserved under Theorem 10.30. p.In[! 1o o o 0 0 00] 10 =(2. Definition 10. Then there exist Theorem 10. l.30. £). phenomena at a cost considerably less than a full SVD. It turns out that the principal property so preserved is the sign preserved under congruence. and eigenvalues. v. If In(A) = (rr. Example 10. If A AH e e x " then A> 0 if and only if In(A) = (n.0). It is of interest to ask what other properties of a matrix are preserved under congruence. Let A = A He ennxn and X e Cnnxn.t h e n A > 0 if and only if In (A) = (n. D 0 Remark 10. v. v. then rank(A) rr v. p. 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. .28. i. Equivalence Transformations and Congruence 103 103 Theorem 10.1. n The signature of A is given by sig(A) = n . where R e €. upper Proof: For the proof. [21. see [4] for details. D Theorem 10. of each eigenvalue.25 (Complete Orthogonal Decomposition). i. Proof: For the proof. Let A E e~xn. The signature of is Example 10. Note that a congruence is a similarity if and only if X is unitary. see. Let A = AH E e nxn and let rr. 0. of A. congruence.10.31 (Sylvester's Law of Inertia). sig(A) = rr — v.. (TT. negative.31 (Sylvester's Law of Inertia). Proof: For the proof. v. Again. Let A = AH e C"x" and let 7t. n.v. if A is Hermitian.27. of A. Proof: For the proof.31 guarantees that rank and signature of matrix are preserved under congruence.3.XH AX Definition 10. If A = A" E C nnxn.e. numbers In(A) (n. £). Note that congruence preserves the property of being Hermitian. Note that a congruence is a similarity if and only ifX is unitary.28. 2. see [4] for details.27. In(A) 3.29.6) E e.
. .1). . Theorem 10. 1. . . the number 0/0 's is (. or D > 0 and A . I. Then Remark Remark 10. for example. Define the x n matrix vv = diag(I/~. if ifA>0. B D ] >  ° if and only if A:::: 0.. . X e C"nxn such that XHAX = diag(l. and the numberofO's is~.. the next v are negative. 0).. 1.. . Proof: AI. 0 D Then it is easy to check that X = V VV yields the desired result. .. Then = AT D = DT.. Let A = AHeE cnxn with In(A) = (jt.. .BT AI > 0. 1.. the number of — 's is v.fArr+I' . left AT D DT. where the number of X 1's is Jr.... X UW desired 10. .. where the number of E c~xn XH AX = diag(1. I/~.BD^BT > 0. A w ).. .3.0. . An). v. AA+B = B. and D .I BT > O. 0..BT A+B:::: o.. v. 1. D > and . . £).BT A+B > 0. Theorem positive. By Theorem 10.2 there exists a unitary matrix V such that VHAU = diag(AI. ifand only ifeither A > 0 and D . 's is 7i.4 10. . . and D . O.104 104 Chapter 10. I..35...33.. and the final £ are 0.0). Canonical Forms Chapter 10. .BT A~l B > 0. Then there exists a matrix AH C"xn In(A) = (Jr. Suppose A = AT and D = DT.. the number of Il's is v. .34. .. An of Jr Proof: Let AI . . I..4 Rational Canonical Form Rational Canonical Form rational One final canonical form to be mentioned is the rational canonical form. Canonical Forms Theorem 10.1 Block matrices and definiteness Theorem 10. .. . Xw denote the eigenvalues of A and order them such that the first TTare ~ O. Define the nn x n matrix U UH AV = diag(Ai. .33. . 1/. AA+B = B.3. 0 D 10. .. Proof: Consider the congruence with Proof: Consider proof Theorem and proceed as in the proof of Theorem 10. Proof: proof Proof: The proof follows by considering.1 10.32. . if and if either A> and D .. . the congruence B ] [I D ~ 0 _AI B I ° JT [ A BT ~ ][ ~ 0 D The details are straightforward and are left to the reader. . .BD. I/.fArr+v. Suppose A = AT and D = DT. Note the symmetric Schur complements of A (or D) in the theorem.
Using the reverseorder identity similarity P given by (9.7) is called a cornpanion rnatrix or Definition 10..4. o (10.4. A is easily seen to be similar to the following matrix identity similarity P given by (9. + an_IAnI).37. the inverse of a nonsingular companion matrix is again in companion form. equivalently. the inverse of a nonsingular companion matrix is again in companion form.10. To Companion matrices also appear in the literature in several equivalent forms. : ~ ! ~01].7) Definition 10. 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 /= 0.9) Moreover. A matrix A e E nx " of the form (10. Companion matrices also appear in the literature in several equivalent forms.Then it can be shown (see [12]) that A mial is 7r(A) = A" . 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 (10.7) is called a companion matrix or is said to be in companion forrn.18). consider the companion matrix (l0. since a matrix is similar to its transpose (see exercise 13 in Chapter 9).18).. Notice that in all cases a companion matrix is nonsingular if and only if aO i= O. since a matrix is similar to its transpose (see exercise 13 in Chapter 9). Using the reverseorder This matrix is a special case of a matrix in lower Hessenberg form. is said to be in cornpanion form. 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 ~'). Rational Canonical Form 10.37. A matrix A E lRnxn of the form (10.11) . equivalently. To illustrate. For £*Yamr\1j=» example. if its Jordan canonical form and characteristic polynomial are the same or.10) o 1 o 1 o o o o o o (10. In fact. 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.(ao + «A + . l 0 0 ~ ao ~ ao _!!l (10.8) This matrix is a special case of a matrix in lower Hessenberg form.36. 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 . For In fact. if its Jordan canonical form has only one block associated with each distinct eigenvalue. has only one block associated with each distinct eigenvalue. Rational Canonical Form 105 105 Definition A matrix A e M"x" is said to be Definition 10. the Moreover. consider the companion matrix illustrate.
stable ones are nearly unstable.1. Moreover. stable ones are nearly unstable. 02. has more than one Jordan block associated with If A € JRnxn derogatory. 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. Explicit formulas for all the associated right and left singular vectors can Remark 10. Explicit formulas for all the associated right and left singular vectors can also be derived easily... Such matrices are said to be in each of whose diagonal blocks is a companion matrix.e. If a companion matrix of the form (10. see [14].. matrix is not a companion matrix unless a = O.7) is singular. a n i] and l c I+~T a. Then + ai + ... n . However. 3. . and so forth [14]. For example. then its pseudoIf singular. Canonical Forms Chapter 10.7). Algorithms to reduce but unfortunately they are often very difficult to work with numerically.Jy2 . companion matrices are known to possess many undesirable numerical properties. Companion matrices have many other interesting properties.38. I — T = T) Note that / .10).. is the fact that their singular values can be found in closed form. For example. also be derived easily. . companion an arbitrary matrix to companion form are numerically unstable. Algorithms to reduce an arbitrary matrix to companion form are numerically unstable. Let a\ > a2 > . Let al ~ GI ~ • • ~ an be the singular values of the companion matrix Theorem 10. and perhaps surprisingly. if ao = 1 inverse can still be computed.. each of whose diagonal blocks is a companion matrix. the largest and smallest singular values can also be written in the equivalent form If ao =1= 0.. Theorem 10. Leta = ar aJ al 2_ 2 ( y + Jy 2.39. Such matrices are said to be in rational canonical form Frobenius rational canonical form (or Frobenius canonical form).7). and perCompanion matrices have many other interesting properties.• > an be the singular values of the companion matrix A in (10. form). see haps surprisingly. at least one eigenvalue. 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 (10.7). a2. for example.106 Chapter 10. anIf and let e M"" \a\. then it is not similar to a companion matrix of the form (10.Q + a. Let a = a\ + a\ + • • • + a%_{ and y = 1 + «. and hence the pseudoinverse of a singular companion + matrix is not a companion matrix unless a = 0.e. is the fact that their singular values can be found in closed form. among which. If A E R nx " is derogatory. [12].caa T = (I + aaT) I .. their eigenstructure is extremely ill conditioned.38. 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.39. and so forth [14]. associated at least one eigenvalue. especially nonsingular ones are nearly singular.caa T ca o J. = ~ (y . . Moreover.. Let a E JRn1 denote the vector [ai. i. .. i.. Ifao ^ 0.7). + a.10)..4aJ) . see. For details.. among which. nonsingular ones are nearly singular. Then A in (10.4ao ' 1 2)  a? = 1 for i = 2. with a similar result for companion matrices of the form (10. it can be shown that a derogatory matrix is similar to a block diagonal matrix.. in matrices are known to possess many undesirable numerical properties. in n general and especially as n increases._1 and y = 1 + + a. if ao = 0. the largest and smallest singular values can also be written in the equivalent form Remark 10. a. .
Show that [ * }. K\ (A) (10..18) U A E cc nxn Theorem 10. 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. Find a unitary matrix U such that [~ M CC x 2 Find a unitary matrix U such that 6.2). say O(lO k ). 3. this condition number is the ratio of largest to smallest singular values which. by the theorem. A [ must also be positive 7. then peA) = A2. Suppose A e E"x" is positive definite. For example.• ~ an ~ O. Let A 7. Show that if A is normal. S 6 E nxn be symmetric. one may lose up to k digits of precision.11). Prove that if A e M"x" is normal. yn and singular values a\ ~ a2 > . Let A G Cnx" and define p(A) = maxx€A(A) IAI.18) and the matrix U in identity in (9. .11). is true if n = 2..38 yields some understanding of why difficult numerical Remark 10. A E cc nxn peA) = max). 5. It is easy to show that y/2/ao < K2(A) < £. Let A € C n xn be normal with eigenvalues y1 . 6. 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. when solving linear equations numerical sensitivity Kp(A) = systems of equations of the form (6.. Let R. Let R. Show that [~ R > SI. when solving linear behavior might be expected for companion matrices. Is [ ^ A E jRnxn is definite. If this number is large. It is not unusual for y to be large for large n.(A) for e n. Let A = I J : ]eEC 22x2. can be determined explicitly as determined explicitly y+J y 2 . then it must be diagonal. Remark 10. .5 to find a unitary matrix Q that reduces A e C"x" to lower triangular form. (A) = IA. then p(A) = IIAII2' Show that the converse is true if n = 2.. In the 2norm.. Show that a. . (A) A.38 yields some understanding of why difficult numerical behavior might be expected for companion matrices. EXERCISES EXERCISES 1. Theorem 10. If A E Wxn is positive definite. For example. • • > on > 0. Show that a.(A)I for ii E!l. then K2(A) ^ T~I. Show that the converse radius of A. one may lose up to k digits of to the matrix Pnorm. 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 . If this number is large. A E en x n eigenvalues A]. In the 2norm. and when GO is small or y is large (or both)...40. 1. this condition number is the ratio of largest to smallest singular precision. A E jRnxn N(A) = A/"(A ). Theorem 10.. If A e jRn xn 8. E jRnxn be symmetric. 2.4a5 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). 9.40.. Show that if a triangular matrix is normal. Show that if A is normal.. Use the reverseorder identity matrix P introduced in (9. An and singular 0'1 > 0'2 ~ 4.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 p(A) is called the spectral radius of A.EA(A) I'MpeA) 3. then Af(A) = N(A Tr ). say 0(10*). show that AI must also be positive definite.
(a) Chapter 10. Canonical Forms Chapter 10.1 1. Find the inertia of the following matrices: following 10.j 1+ j ] 2 ' (d) [ .108 108 10. Canonical Forms [~ ~ l (b) [ 2 1. .j 1+ j ] 1 .
e° = I.nxn is constant and does not depend on t.1 Properties of the matrix exponential Properties of the matrix exponential 1. Forall A EG R" XM . The solution of (11. Proof: This follows immediately from Definition 11. The solution of (11.2) k=O The series (11.n (11. The solution of (11.1) involves the matrix to +(0).1 Differential Equations Differential Equations = Ax(t).1) involves the matrix (11. k. Proof This follows immediately from Definition 11.2) can be shown to converge for all A (has radius of convergence equal to +00). For all A JR.1) for t 2: to. It can be described conveniently in terms of the matrix exponential. (e(eAf = e A e^. This is known as an initialvalue problem.nxn is defined by Definition 11. Proof: This follows immediately from Definition 11. A) • 2.Ak.1.Chapter 11 Chapter 11 Linear Differential and Linear Differential and Difference Equations Difference Equations 11. the matrix exponential e A E Rnxn is defined by the power series power series e = A L +00 1 . unique.1 and linearity of the transpose.1.1 11. The solution of (11.nxn. which thus also converges for all A and uniformly in t. It can be described conveniently in terms of the matrix exponential. (11.nxn.2) can be shown to converge for all A (has radius of convergence equal The series (11. where the matrix A E Rnxn is constant chapter only to the socalled timeinvariant case.3) which thus also converges for all A and uniformly in t. the matrix exponential e A e JR. Definition 11. T T 109 109 .1 by setting AA =O.1) is then known always to exist and be unique. = 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.1. 11. We restrict our attention in this for t > IQ.1) is then known always to exist and be and does not depend on t. For all A e Rnxn.1 by setting = 0. where the matrix A e JR. Proof This follows immediately from Definition 11. For all A E JR.1 and linearity of the transpose.1.1 11. eO = I. We restrict our attention in this chapter only to the socalled timeinvariant case. This is known as an initialvalue problem.
.A)I} = «M. Proof" Simply take T = — t in property 3. binomial theorem on (A B) and the commutativity of A and B. ) . For all e R"x" and for all t.. ) ( 1+ tA + t2!A 2 +. ForaH A E R" x " and for all t e JR. For all A.l{e tA}} = (sI . 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. 4... = e'A erA = elAe'A . et(A+B) =etAe tB = etBe tA if and only if A all e JRnxn and all e R.. i. Let denote the Laplace transform and £~! the inverse Laplace transform.1 {(j/A). For all A E JRnxn and for all t. (b) .A)I. Linear Differential and Difference Equations e(t+r)A e(t+T)A 3.. ) ( I + T A + T2!2 A 2 +. et(A+B) =^e'Ae'B = e'Be'A and and B commute. r e R. For all e JRnxn and for all E R. 2 2! and and while while e e tB tA = ( 1+ tB t2 2 + 2iB 2 +. Then for 6. (a) C{etA = (sIArl.. Let £ denote the Laplace transform and £1 the inverse Laplace transform. T E JR. (b) £. Proof" We prove only (a). {+oo = io et(sl)e tA dt since A and (sf) commute =io (+oo ef(Asl) dt . (e'A)~l e~'A.e. Then for E JRnxn t E R. i. Part (b) follows similarly. 5. on (t + T)*. ) . Proof" Note that Proof: Note that e(t+r)A = etA erA = erAe tA .. AB = B A. (etA)1 = e. and B commute. (a) ..e.lI{(sl. AB = BA.110 110 Chapter 11. = I + (t + T)A + (t + T)2 A 2 + . Proof: We prove only (a). Part (b) follows similarly.1 } = erA. 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.... B E R" xn and for all t E JR. Proof' Note that Proof: Note that et(A+B) = I t + teA + B) + (A + B)2 + .tA ... 2! and and e e tA rA 2 = ( I + t A + t2! A 2 +. all A € R"x" and for all t € lR. Proof: Simply take T = t in property 3. 6. Linear Differential and Difference Equations Chapter 11.
the scalar dyadic decomposition can be replaced by et(Asl) =L .1 The matrix (s I — A) I is called the resolvent of A and is defined for all s not in A (A). For all A e R"x" and for all t e R.H = '"' assuming Re s > Re Ai for i E !! = (sI . Alternatively.11.3) is uniformly convergent.y.l)e . that A is diagonalizable. ) = I ( Ae + = tA ~.Ae tA I = III (etAe~tA L'lt = = /A) . A2 + (~~)2 A tA II tA Il 1 (_ 2! + .. = (sl A)..u ..Ae tA tA tA I I e tA . the scalar dyadic decomposition can be replaced by If this is not the case. A 2etA + .. Differential Equations 11.X i y.y..Ae tA etA) . The matrix (s I .A)I. ) = L'lt IIA 21111e tA IIe~tIIAII.All succeeding steps in the proof then follow in aastraightforward way.3) is uniformly convergent.. for convenience. ) etA I < MIIA21111e  L'lt (L'lt)2 + IIAII + IIAI12 + . 1h(e tA ) = AetA = etA A... ) 3! 4! L'ltiIAIl < L'lt1lA21111e (1 + + (~t IIAII2 + . ) 3 II I ( ~.. All succeeding steps in the proof then follow in straightforward way..A) ~' is called the resolvent of A and is defined for all s not in A (A). e'A Proof: Since the series (11. that A is diagonalizable. it can be differentiated termbyterm from which the result follows immediately...Ae II = I L'lt (M)2 + ~ A 2 +.etA .. the formal definition d dt _(/A) = lim ~t+O e(t+M)A _ etA L'lt can be employed as follows.H L... employed I e(t+~t)AAt. Notice in the proof that we have assumed.H dt assuming A is diagonalizable .=1 = ~[fo+oo e(AiS)t dt]x. for convenience..1. using the JCF. For all A E JRnxn and for all E JR. If this is not the case.AetAil Ae tA I ~t (e~tAetA I (M A I ~t (e~tA . it can be differentiated termbyProof: Since the series (11. . For any consistent matrix norm. Differential Equations 111 111 = {+oo 10 n t 1 e(AiS)t x..H using the JCF.=1 m Xiet(Jisl)y.. s .1. .A"I i=1 .. £(e'A) 7. . Notice in the proof that we have assumed..
The formula can be derived by means of an integrating factor "trick" as follows. The formula can be derived by means of an integrating factor "trick" direct differentiation. lo t (11. the limit exists and equals Ae'A •.8) .5) and use property 7 of the matrix exponential to get x t ) = Ae(tto)A xo fundamental Ae(t~to)Axo = Ax(t).4) for t ::: to is given by (11. x(to) = Xo E IRn (11. Premultiply the equation x .7) is the solution of (1l. (11. Also.2. Let A E Rnxn . A similar proof yields the limit e'A A. continuous.6). 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. 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. D Ir: Remark 11.t goes to O. 0 ordinary differential equations. x(to) = e(toto)A Xo = XQ so.6). t ) . Then the solution of the linear inhomogeneous initialvalue problem and. say.4.5) Proof: Differentiate (11.3 Inhomogeneous linear differential equations Inhomogeneous equations Theorem 11. (11. by the fundamental existence and uniqueness theorem for ordinary differential equations. 0 uniqueness theorem for ordinary differential equations.1.7) Proof: Differentiate (11. the righthand side above clearly goes to 0 as At goes to 0.5) is the solution of (11. fact that A commutes with any polynomial of A of finite degree and hence with etA.4). or one can use the limit exists and equals Ae t A A similar proof yields the limit et A A. Thus. The proof above simply verifies the variation of parameters formula by Remark 11.. The proof above simply verifies the variation of parameters formula by direct differentiation.f(p(t). Linear Differential and Difference Equations Chapter 11. The solution of the linear homogeneous initialvalue problem = Ax(l). (11.3.5) and use property 7 of the matrix exponential to get x ((t) = Proof: Differentiate (11. t ) .7) and again use property 7 of the matrix exponential.5) is the solution of (11.4. D 11. (11. Linear Differential and Difference Equations For fixed t. t) dx = l af(x t) ' dx pet) at (t) q + dq(t) dp(t) f(q(t). continuous.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. Premultiply the equation x — Ax = Bu by e~ to get (11.dt dt is used to get x ( t ) = Ae(tto)Ax0 + f'o Ae('s)ABu(s) ds + Bu(t) = Ax(t) + Bu(t).7) is the solution of (11. Let A E IR n xn. Thus.2 Homogeneous linear differential equations Homogeneous equations x(t) Theorem 11.1. say. The general formula formula d dt l q (t) pet) f(x. B e IR xm and let the vectorvalued function u be given Theorem and.7) and again use property 7 of the matrix exponential.Ax = Bu by e.4).112 112 Chapter 11..i~t()Oc() nnd uniqu()Oc:s:s theorem for *('o)} = <?(f°~fo)/1. 11.¥o + 0 = XQ so. the righthand side above clearly goes to 0 as t:. Then the solution of the linear inhomogeneous initialvalue problem x(t) = Ax(t) + Bu(t). Also. The general Proof: Differentiate (11. B E Wnxm and let the vectorvalued function u be given Let A e IR nxn . by the fundamental existence and x(t0) — e(fo~t°')AXQ — Xo uniqueness theorem for ordinary differential equations.tA to get as follows. the For fixed t. The solution ofthe linear homogeneous initialvalue problem Let A e Rnxn.
sA Bu(s) ds x(t) = e(ttolA xo + lto t e(ts)A Bu(s) ds. and the proof is essentially the same. X t) X 0 D Corollary 11. X(to) =C E jRnxn (11. The first is an obvious generalization of Theorem 11. For convenience.12). problem problem X(t) = AX(t) + X(t)B. The fact that X((t) satisfies the initial condition is trivial.etoAx(to) = lto t e. Theorem 11.1. following to = O. the Proof: Differentiate etACe tB property Proof: Differentiate etACetB with respect to t and use property 7 of the matrix exponential. etAx(t) . E ]R.12) X(t) = etACetAT has the solution X(t} = etACetAT. punov differential equation.1. the When C is symmetric in (11.11. Then the matrix initialvalue problem X(t) = AX(t) + X(t)AT.8) over the interval [to.5. Differential Equations 11. the Theorem 11.7.6.. differential equation. and the proof is essentially the same. C e IR" ".12) is known as a LyaX t) punov differential equation.7. and hence t d esAx(s) ds = to ds 1t to eSABu(s) ds. Then the matrix initialvalue E jRmxm.4 Linear matrix differential equations Linear matrix differential equations Matrixvalued initialvalue problems also occur frequently. Differential Equations [to. . t exponential. t]: 113 1 Thus.nxm. Theorem 11. X(O) = C (11.9) for t ::: to is given by for t > to is given by X(t) = e(tto)Ac. Let A E Rnxn. we can have coefficient matrices on both the right and left.2.11) X(t) = etACe = e ratB has the solution X ( t ) — atACe tB . E ]R. t]: Now integrate (11. Theorem 11.1. X(O) =C (11.2. X((t) is symmetric and (11. The of nrohlcm problem X(t) = AX(t).1.4 11. 11.nxn.10) coefficient In the matrix case. Let A E Wlxn. The solution of the matrix linear homogeneous initialvalue e jRnxn.11) is known as a Sylvester Sylvester differential equation. Corollary 11. Let A. the following theorem is stated with initial time to = 0. B e R m x m .6. and C e Rnxm. e jRnxn. (11. The initialvalue problem (11.
4) can be written x(t) = e(tto)A Xo E jRnxn E Wxn = (ti. that it is diagonalizable (if A is not diagonalizLet A and suppose.1.1 n Le A• X'YiH . Then the solution x(t) of (11.iUtO)Xiyr) Xo 1=1 n = L(YiHxoeAi(ttO»Xi. are called the modal directions.1. Then Then i=1 n = L(aieAiUtO»Xi.y. for convenience. Similarly. The decomposition above expresses the solution x (t) as a weighted sum of its directions.H . Linear Differential and Difference Equations Chapter 11. ~ 11. The decomposition above expresses the solution x(t) as a weighted sum of its modal velocities and directions. i=1 In the last equality we have used the fact that yf*Xj = Sfj. 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 *. Then the solution x(t) of (11. 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. Let A E jRnxn and suppose X e jR~xn is such that XI AX = J.e'J. Linear Differential and Difference Equations 11.x.1 . for convenience. in the inhomogeneous case we can write Similarly. 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.114 114 Chapter 11. the rest of this subsection is easily generalized by using the JCF and the decomposition able.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 A = L X. modal velocities and directions. if A is diagonalizable in geneml. 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). where J is a JCF for A. that it is diagonalizable (if A is not diagonalizable. ~ 1=1 I t.5 11. Then Then etA = etXJX1 = XetJX. In the last equality we have used the fact that YiHXj = flij. where J is a JCF for A.li y t as discussed in Chapter 9).
t2 t k. To be specific.eAt). O. But e tN is almost as easy since N The diagonal part is easy: e e = diag(e '. of In the more general case. ext}. the problem clearly reduces simply to the computation of problem clearly reduces the exponential of a Jordan block.I e IN =I+tN+N 2 + . Nk~lI has a 1 in its (1.. + N k2! (k . aareal version of the above can be worked out. real version of the above can be worked out. Differential Equations 11.0.7. ••• ..I)! I o t 1 o Thus. the series expansion of e'N is finite. elN is is nilpotent of degree k.1.1. e'u e l N tu x lH = diag(e At .8.. teAl eAt = 0 0 0 2I e 12 At teAl 0 eAt In the case when A is complex. A. degree k. its first superdiagonal (and O's elsewhere).. or grade) p if if matrix M e jRnxn is nilpotent of degree (or index.e. 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. l's For the matrix N defined above. N has 1's along only its second superdiagonal.EeCkxk be aaJordan block of the form Ji <Ckxk be Jordan block of the form A Ji = 1 o o o =U+N. nilpotent Definition 11. let . i. A matrix M E M nx " is nilpotent of degree (or index. Mp~l ^ O. Differential Equations 115 If A is diagonalizable. Finally. Mp = 0.!etN by property 4 of the matrix exponential. e lN finite. eAt teAt eAt o 2I e 12 At IkI At e (kI)! 0 ell..11. Thus. AI e I. and N kforth. is complex. k) element and has O's everywhere else. it is easy to check that while N has 1's along only its first superdiagonal (and O's elsewhere). while MPI t=. Thus. N22 has l's along only its second superdiagonal. . or grade) MP = 0. (1.. o A o o A Clearly A/ and N commute. e ttJi = eO.. and so forth. k) O's k k N = 0.
. all the Ak — expressed k 1. the function g is known and /(A) = g(A).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.10. Let Example 11. + I) 3 . . compute f(A) = etA.3. . anl are n constants that are to be determined. Given A E jRnxn and f(A) = etA.9. With the aiS then kth superscript (&) X. The method is stated and illustrated for the hand calculation in smallorder problems. k = 0. Linear Differential and Difference Equations Example 11.s known.2t e. Let A = [~ o ~01~ ] t .) = (A + 1)3. so m = 1 and n{ = 3. .. characteristic of n(X (^ ~~ ^i)"'» where the A. the function g is known and f(A) = g(A). The motivation for this method is known..nxn and /(A) = etx.I. Then jr(A. . in fact. The method is stated and illustrated for the exponential function but applies equally well to other functions..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 11.1. 2} and Example 11. Then A(A) = {2.t • g'(1) = f'(1) g"(I) = 1"(1) . so m = 1 and nl Let g(X) = UQ + alA + a2A2.2. which says that all powers of A greater than A n . compute f(A) = e'A. Here.s are given by g(A) — ao aiS a\X o^A. . g(I) = f(1) ==> ao . the unique OTQ. . . I. Let A = [ ~_\ J]. and /(A) = etA. They are. 2} and etA Xe tJ =[=i a = xI =[ =[ 2 1 ] exp t ] [ [ 2 0 ~ ] [ 1 1 2 1 2 ] 2 1 e~2t te.Ai t'. . ni . The polynomial g gives the appropriate linear combination. where t is a fixed scalar..a l +a2 = e==> at . Define the Ai nr=1 n where ao.116 Chapter 11.. — 1.. Linear Differential and Difference Equations Chapter 11. f(A) n(A) etK.t .. (A.1 can be expressed as linear combinations of Ak for k = 0.. i Em. a.s are distinct.. Suppose the characteristic polynomial of A can be written as n(A)) = Yi?=i (A . terms of order greater than n . Thus. The motivation for this method is the CayleyHamilton Theorem.10.. functions. . . I. Let A Then A (A) = {2.2a2 = te.9. . Theorem 9. Then the three equations for the a. the superscript (k) denotes the fcth derivative with respect to A. ani solution of the n equations: g(k)(Ai) = f(k)(Ai). t fixed Given A € E. ==> 2a2 = t 2 e. n .
This etA = . we find 117 Thus. Thus.s are given by Let g(A) ao + aLA. Use Pade approximation. Then the defining equations for the a.2t aL = + 2te.2t I [4 4] I 0 _ [  e.2t te. 2.2t . we find Solving for the a. te.2t + 2te.) = «o + ofiA. Use etA = £~l{(sl . Let A _* Example 11. we find Solving for the aiS. s. Let A = [ ~4 J] and /(A) = eO.1. Then rr(A) = f\ + o\2 so m = and (A i 2) «i nL = 2. Differential Equations 11 .A)^ 1 } and techniques for inverse Laplace transforms. f(A) = etA = g(A) = aoI + al A = (e.11..11.2t _ Other methods Other methods 1. g'(2) = f'(2) = te Solving for the a. Differential Equations Solving for the ai s.cI{(sI — A)I} is quite effective for smallorder problems. t ff>\ tk TU^^ _/"i\ Example 11. we find ao = e. Then 7r(X) = (A+ 2)22 so m = 11and [::::~ 4i and f(A) = ea.2t . 2t . There is an extensive literature on approximating certain nonlinear functions by rational functions.2t ) 2te. 2. The matrix analogue yields e A ~ functions rational eA = .s. 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.11. Let g(A.2t .1. 1. Then the defining equations for the aiS are given by 6] g(2) = f(2) ==> ao ==> al 2al = e.2t [ ~ o ] + te.
we restrict our attention only to the socalled timeinvariant case.118 118 l Chapter 11. Then the solution of the inhomogeneous initialvalue problem (11.e. The solution of the linear homogeneous system ofdifference Let A e jRn xn.• + Vq A q. exhibit many parallels to the continuoustime differential equation case. Numerical loss of accuracy can occur in this procedure from the successive squarings. = P = multiplying it by 1/2* for sufficiently large k and using the fact that e = ( e j .2 Inhomogeneous linear difference equations Inhomogeneous linear difference equations E jRnxn.13).14) into (11. convenience. e (e( 3. Reliable and efficient computation 4.2 11. Let A e Rnxn. and this observation is exploited frequently.13). case. modeled by systems of equations of the previous section. 4.13) is constant and does not depend on k. [19]. say. say. we have chosen ko = 0 for convenience. This can be arranged by scaling A. Again. no double subscripts). Linear discretetime systems. 0 D Remark 11. but since the system is timeinvariant.. [19]. exhibit many parallels to the continuoustime differential equation difference equations.. eS . •• vq A . 11. in the matrix case the exponential is accurate only in a neighborhood of the origin. Again. no double subscripts). in the matrix case this means when  A is sufficiently small. Reduce A to (real) Schur form S via the unitary similarity U and use e A 3.13) is constant and does not depend on k.13. Many methods are outlined in. This can be arranged by scaling A. a Fade approximation for polynomials of various orders. Unfortunately. Numerical loss of accuracy can occur in this procedure from the successive squarings.12. a Pad6 approximation for denominator the exponential is accurate only in a neighborhood of the origin..14) into (11.2.e. We could also consider an arbitrary "initial time" ko.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. for example. by 22' 2* )A A multiplying it by 1/2k for sufficiently large k and using the fact that A = / { ]I //2')A )\ * .2. + opAP and N(A) = vol + vIA + D~ (A)N(A). We could also case. Proof: The proof is almost immediate upon substitution of (11.2. B e Rnxm and suppose {«*}£§ « a given sequence of mvectors. Reliable and efficient computation of matrix functions such as e A and log(A) remains a fertile area for research. Linear Differential and Difference Equations Chapter 11.15) .. The solution ofthe linear homogeneous system of difference equations equations (11... Then the solution of the inhomogeneous initialvalue problem mvectors. Linear Differential and Difference Equations DI(A)N(A). but since the system is timeinvariant. and since we consider an arbitrary "initial time" ko. we have chosen ko = 0 for want to keep the formulas "clean" (i. and since we want to keep the formulas "clean" (i. Many methods are outlined in. 11. 11. and this observation is exploited frequently. of matrix functions such as e A and 10g(A) remains a fertile area for research. where the matrix A in (11.14. E jRnxm {udt~ is of Theorem 11. Linear discretetime systems. modeled by systems of difference equations.1 11. where the matrix A in (11. we restrict our attention only to the socalled timeinvariant Remark 11. for example.13.13) for k > 0 is given by for k 2:: 0 is given by Proof: The proof is almost immediate upon substitution of (11. Unfortunately. 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. where D(A) = 001 + olA + . Let A E Rnxn. 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. by this means when IIAII is sufficiently small.1 Homogeneous linear difference equations Homogeneous linear difference equations Theorem 11.
.. where J is a E jRnxn and X E R^n J.16) into (11..H m if A is diagonalizable. Assume that A e M" xn and let X e jR~xn be such that XI AX = /. based Methods based on the JCF are sometimes useful.15). Difference Equations 119 119 is given by kI xk=AkXO+LAkjIBUj. One solution method.y. sometimes useful Ak.2. X~1 AX JCF for A. k=O Assuming Izl > max A. Then Ak = (XJXI)k = XJkX. substitution of (11.=1 H l If A is diagonalizable. 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. j=O (11. the ztransform of the sequence {Ak is then given by Assuming z > max IAI. One definition of the ztransform of a sequence {gk} is a matrix exponential.2.16) Proof: The proof is again almost immediate Proof: The proof is again almost immediate upon substitution of (11. Difference Equations 11.3 11. One definition of the ztransform of a sequence is +00 Z({gk}t~) = LgkZk.11.. LXi Jtyi . since /* is simply a diagonal matrix. Jk . by analogy with the use of Laplace transforms to compute ztransforms.zA =I+A+"2 A + .2. +00 k=O z z = (lzIA)I = z(zI . again mostly for smallorder probsmallorder lems. k:::.16) into (11.A)I. it is then easy to compute Ak via the formula Ak = XJkXXI Ak Ak — X Jk If diagonalizable.1 _I tA~X.O.2. in general. a matrix exponential.15). is to use ztransforms. which is numerically unstable but sometimes useful for hand calculation.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. Then JCF for A.
.. see [11. To be specific. The symbol ( ) has the usual definition of . A is complex. and is to be interpreted as 0 if k < q. 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 (.1 Ak The symbol (: ) has the usual definition of q!(kk~q)! and is to be interpreted as 0 if k < q. ) Ak..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. 0 A Writing /. for example.3 11.1 Ak ( . [11. Linear Differential and Difference Equations In the general case.1. Let A Ak = XJkX1 = [=i 4 a [2 1 J].1. Example 11.17) with ¢J(t) a given function and n initial conditions 4>(t} y(O) = Co. aareal version of the above can be worked out. Consider.l8) .2) . the problem again reduces to the computation of the power of a In the general case. the initialvalue problem initialvalue (11. inI)(O) = CnI' (1l. real version of the above can be worked out.(^ . Then Then 1 ] [(_2)k 1 0 k(2)kk(2) 1 ] [ _ [ (_2/.1(2k . . the problem again reduces to the computation of the power of a To Ji E Cpxp Jordan block. but again no universally "best" method be derived for the computation of matrix powers.• = AI and noting that AI and the nilpotent matrix Writing Ji = XI + N and noting that XI and the nilpotent matrix N commute.120 Chapter 11. 1 1 1 2 1 ] Basic analogues of other methods such as those mentioned in Section 11. Let A = [_J Example 11.6 can also methods 11..is complex.15. Linear Differential and Difference Equations Chapter 11. For an erudite discussion of the state of the art. In the case when A.1 (2 .15. 11..6 be derived for the computation of matrix powers.2k) k( _2)k1 ] k( 2l+ (2l.)A  ( k ) AkP+I pl 0 J/ = kA k. Ch.2 0 0 0 0 kA k . 18]. y(O) = CI. let 7. e Cpxp be a Jordan block of the form o .. it is commute. but again no universally "best" method exists.
Define a vector x (t) E ]Rn with Here. Cn \ .. . . Then Xl (I) X2(t) = X2(t) = y(t).. Let P E lR nxn be a projection.. Further. EXERCISES EXERCISES 1. X2(t) yet). . = Xn(t) = y(nl)(t). Show that e % / + 1.. Xn(t) y { n ~ l ) ( t ) . into a linear firstorder difference equation with (vector) initial with n initial conditions. Let . Cl. is often well worth avoiding. Suppose x. 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. Suppose x.A) = A. let a = xTy.19) possesses many nasty numerical properties for even moderately sized n matrix A in (11. 3. Show that etA 2. as mentioned before. Let 3. into a linear firstorder difference equation with (vector) initial condition.. y(m) denotes the mth derivative of y with respect to t. 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. condition. at least for computational purposes. Show that e'A 1+ g ( t . Then components Xl (t) yet). Note that det(X7 — A) = An + an\Xn 1l H alA + ao. the companion Note that det(A! . be a projection. let a = XT y.. However. is often well worth avoiding. where !(eat . c\. Define a vector x (?) e R" with components *i(0 = y ( t ) .. Let P € R 1. x2(t) = y ( t ) . where I + get. . y € R" and let A = xyT. a)xyT. a)xyT. and. 2. (11. aly(t) ..718P. the companion matrix A in (11. However. = O. •.. v (m) denotes the mth derivative of y with respect to t. xn(t) = Inl)(t).Exercises 121 121 Here.an_llnl)(t) Xnl (t) Xn(t) = y(n)(t) = aoy(t)  + ¢(t) = aOx\ (t) . at least for computational purposes... C M _I] The initial conditions take the form X (0) = C = [co...a\X2(t) . . .718P. = X3(t) = yet).anlXn(t) + ¢(t).I) g(t."+ an_1A n~ + .. . as mentioned before. +h a\X+ ao. Further.19) possesses many nasty numerical properties for even moderately sized n and.a)= { a t nxn p if a if a 1= 0. Show that e P ~ ! + 1... . y E lRn and let A = xyT.
Show that eH must be symplectic. must also be an eigenValue of S. Let denote the skewsymmetric matrix 4.A and to be symplectic K~l AT K . f3 E R and Then show that Then show that ectt _eut cos f3t sin f3t ectctrt e sin ~t cos/A J. Let a. also be an eigenvalue of H. Show that 1 /A. Let K denote the skewsymmetric matrix 0 [ In In ] 0 ' In A E jR2nx2n where /„ denotes the n x n identity matrix. Show that —A. (a) Suppose H is Hamiltonian and let). must also be an eigenvalue of H. Find eM when A = Find etA = 8. Let (a) Solve the differential equation (a) Solve the differential equation i = Ax . Show that ). must (b) Suppose S is symplectic and let).. Show that 1/). (b) Suppose S is symplectic and let A.122 122 Chapter 11. A matrix A e R 2nx2n is said to be K I AT K = . Show S~1 H S Hamiltonian. must (a) Suppose E is Hamiltonian and let A. Let 5.. be an eigenvalue of S. 4. Hamiltonian.be an eigenvalue of S. Show that SI HS must be Suppose and symplectic.A 1 . Hamiltonian if K~1ATK = A and to be symplectic if K I ATK = A I.. (d) Suppose 5. Linear Differential and Difference Equations Chapter 11. be an eigenvalue of H. x(O) =[ ~ J. Find a general expression for Find a general expression for 7. also eigenvalue of (c) Suppose that H is Hamiltonian and S is symplectic. H (d) Suppose H is Hamiltonian. ft € lR and Let a.be an eigenvalue of H. 6. Show that E jRmxn e = [eoI A sinh 1 X ] ~I .. . Linear Differential and Difference where X e M'nx" is arbitrary.
and the Americas (R).X(t)A. what is the value of ZIOOO? What is the value of Zk in 2. If £0 = 1 and z\ If Zo = 1 and ZI = 2. (b) Consider the difference equation (b) Consider the difference equation Zk+2 + 2Zk+1 + Zk = O. (d) Find the limiting distribution of the $40 trillion as the universe ends. around the time the Cubs win a World Series).Exercises Exercises (b) Solve the differential equation (b) Solve the differential equation i 123 = Ax + b. (c) Find the distribution of the companies' assets at year k. of Cf or all t. (b) Find the eigenvalues and right eigenvectors of M. i. 10. Show that for t > 0. 11. as (d) Find the limiting distribution of the $40 trillion as the universe ends.3.11 in [24]. a quarter goes to Europe. I/X(t)1/2 = ex for all t > 0. x(O) = Xo for t ~ O. 11. and a quarter year half of the Americas' money stays home..11 in [24].e. 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). and a quarter goes to Asia. For Europe and Asia.e. Suppose that A E ~nxn is skewsymmetric and let ex = Ilxol12. Each year half of the Americas' money stays home.3. Each total assets of $40 trillion of which $20 trillion is in E and $20 trillion is in R. k * +00 (i. (Exercise adapted from Problem 5. Suppose certain multinational companies have Europe (E).) 12. as k —»• +00 (i. . (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.Yet) + 2y(t) + yet) = 0. 12. Consider the initialvalue problem i(t) = Ax(t). Consider the initialvalue problem 9. The year is 2004 and there are three large "free trade zones" in the world: Asia (A). x(O) =[ ~ l 9. what is the value of ZIQOO? What is the value of Zk in general? general? .YeO) = O. Suppose that e E"x" is skewsymmetric and let a = \\XQ\\2. Suppose certain multinational companies have total assets of $40 trillion of which $20 trillion is in E and $20 trillion is in R. X(O) = c.. Show that *(OII2 = aforallf > O.) (Exercise adapted from Problem 5. yeO) = 1.. 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?. (c) Find the distribution of the companies' assets at year k. Consider the n x n matrix initialvalue problem X(t) = AX(t) . Europe (E).e. half stays home and half goes to the Americas. and the Americas (R). Consider the n x n matrix initialvalue problem 10.e.. goes to Asia. For Europe and Asia. (a) Find the solution of the initialvalue problem (a) Find the solution of the initialvalue problem . i. a quarter goes to Europe. half stays home and half goes to the Americas.
This page intentionally left blank This page intentionally left blank .
then so is ax [ay] for any nonzero scalar a E <C.Chapter 12 Chapter 12 Generalized Eigenvalue Generalized Eigenvalue Problems 12. The roots ofn(X. (A.1 The Generalized Eigenvalue/Eigenvector Problem The Generalized Eigenvalue/Eigenvector Problem Ax = 'ABx. B E E" xn . called a generalized eigenvalue.XB is called a matrix pencil (or pencil of the matrices A and B). The standard eigenvalue problem considered in Chapter 9 obviously where A.3. When A. Definition 12. and Remark 12. 125 125 .4. Similarly. Remark 12.1) Ax = 'ABx. Definition 12. characteristic hence nonreal eigenvalues must occur in complex conjugate pairs. B) with A. generalized eigenvalue problem.2. the adjective "generalized" "generalized" standard eigenvalue [y] is usually dropped. e C. B e enxn" if there exists a scalar 'A.4. B e C" xn The standard eigenvalue problem considered in Chapter 9 obviously corresponds to the special case that B = I. e e.1 12.2) When the context is such that no confusion can arise. As with the standard eigenvalue problem. B E enxn. if x [y] is a right [left] ax [ay] for any eigenvector.'AB is singular. The matrix A . B). The polynomial n('A) = det(A — A.1.3. . The matrix A — 'AB is called a matrix pencil (or pencil of the matrices A Definition 12.2. such that that (12. In this chapter we consider the generalized eigenvalue problem In we the generalized eigenvalue problem where A. B).) are the eigenvalues of the associated generalized eigenvalue problem. The roots ofn('A) are the eigenvalues of the associated nomial of the matrix pair (A.1. the characteristic polynomial is obviously real. A nonzero vector x e C" is a right generalized eigenvector of the pair generalized eigenvector of (A.5) is called the characteristic polynomial of the matrix pair (A. eigenvector. As with the standard eigenvalue problem. eigenvalues for the generalized eigenvalue problem occur pencil — XB problem occur where the matrix pencil A . hence nonreal eigenvalues must occur in complex conjugate pairs. B E C MX if there exists a scalar A E C.'AB) is called the characteristic polyDefinition 12. B) with A. corresponds to the special case that B = I. B e jRnxn. 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. The polynomial 7r(A. a.) = det(A . Definition 12. B). A E en Definition 12. called a generalized eigenvalue. and A.
Clearly the reciprocal pencil has eigenvalues responding generalized /. ft ^ O. f3 = 0. when B is singular.O.0. 1 and 0.AHa . there is a second eigenvalue "at infinity" for Case 3 of of . then rr(A) is a polynomial nonsingular). With A and B as in (12. For example. (12. That is to say.3) where a and (3 are scalars.L) and there are again four cases to consider. Case 2: = 0. All A E C are eigenvalues since det(A .AB) and there are several cases to consider. otherwise. 1 and ~. If = of degree n.AB Definition 12. k E !!. Case 1: a =I. the pencil A — XB is said to be 12./.0. in particular. 1 (of multiplicity 1).6. the pencil A . Case 1: ^ 0. ^ 0. B k e n.XB) is not identically zero. and ~. A — A. A similar reciprocal symmetry holds for Case 2. Case 2: a = 0. Case 1: a =I.XB. Case 1: a ^ 0. only the case of regular pencils is considered in the remainder of this chapter. 1 Case 3: a =I. A similar reciprocal symmetry holds for Case 2. There are two eigenvalues. it is said to be singular. There are two eigenvalues. when B =I. regular. If det(A — AB) not regular. only the case of regular pencils is considered in the remainder of this chapter. There are two eigenvalues. 1 and O. If del (A .B.126 126 Chapter 12. Case 4: a = 0. There is only one eigenvalue.5. is singular.nA. I and ^. n(X) Remark 12. Generalized Eigenvalue Problems Remark 12.O./. All A 6 C are eigenvalues since det(B — uA) = O. However.{3 = 0. (3 = O.6. {3 ^ 0.O. 1 and 0. f3 = O. All A E C are eigenvalues since det(B . eigenvalues — AB.5. Generalized Eigenvalue Problems Chapter 12. Note that if AA(A) n J\f(B) ^ 0. Case 3: Case 4: = 0.KB always has pencil — AB . f3 / 0. Associated with any matrix pencil A . There is only one eigenvalue. (3 = 0. {3 =I. it is apparent where the "missing" eigenvalues have "missing" gone in Cases 2 and 3. with its reciprocal eigenvalue being 0 in Case 3 of the reciprocal pencil B — /. There are two eigenvalues. Note appear. zero.a/.3).AB. There are two eigenvalues. Case 2: a = 0. Case 4: a = 0./. =I.0. Then the characteristic polynomial is ft det(A .X B is a reciprocal pencil B — n. I multiplicity 1). While While there are applications in system theory and control where singular pencils appear.L = (JL = £. If B is singular. Note that A and/or B may still be singular. Case = ft ^ 0. I (of multiplicity 1). I1 and .LA) = (1 . and hence there are n eigenvalues associated with the pencil A ./. pencil . f3 = O. I and O. However. (3 = O. Case Case 3: a = 0. {3 =I.(3A) ±. There are two eigenvalues. there may be 0.A and corAssociated with any matrix pencil — AB is a reciprocal pencil . There are two eigenvalues. At least for the case of regular pencils. reciprocal Case of reciprocal .XB. All A e C are eigenvalues since det(A — AB) =0. There are two eigenvalues. Case 4: a = 0. It is instructive to consider the reciprocal pencil associated with the example in It reciprocal Remark 12.L)({3 . or infinitely many B = I. det(B .5.A. ft =I.LA. I). the characteristic polynomial is = (I . eigenvalues associated with the pencil A .0.LA and corresponding generalized eigenvalue problem.B) == O. the associated matrix pencil is singular (as in Case N(A) n N(B) =Isingular 4 above). {3 = 0. If B = I (or in general when B is nonsingular).I. suppose associated — AB. {3 =I.LA) == 0.
of AAB. since the generalized eigenvalue problem is then easily seen to be equivalent to the standard eigenvalue problem B. 6. 2. 7. 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 . 6. Sec. B e cnxn . Numerical methods that if B is even moderately ill conditioned with respect to inversion. the pencil A fewer than eigenvalues. Sec. for example. ifx isa right eigenvector of A—XB.AB and QAZ . and eigenvectors under equivalence. to ifx is a Zl x is a righteigenvectorofQAZAQB Z. where Ta and Tp are upper triangular. the eigenvalues ofthe pencil A — XB are then the ratios of the diagonal elements of Ta to the corresponding diagonal elements of TfJ .AQBZ) = det[Q(A .7].7.. 12.7] or [25. Q. ify isa left eigenvector of A —KB. Sec.7. since the generalized eigenvalue problem is then easily seen to be equivalent eigenvalues. lencies rather than similarities. for example.Oif andonly if Q(AXB)Z(Z~lx) = 0. Let A.2. B. then QHy isa left eigenvector ofQAZ AQBZ. Let A. [7. Q~H y isa lefteigenvectorofQAZ — XQBZ. Sec. the eigenvalues of the pencil A . see. E nxn with Q and nonsingular. [7. 7. the The first canonical form is an analogue of Schur's Theorem and forms.AB are then the ratios of the diagBy Theorem 12. which is the generally preferred method for solving the generalized eigenvalue problem. o. However. [7.2.AB) o if and only if (QH y ) H Q ( A –_ B ) Z = Q. canonical forms are available for the generalized Just as for the standard eigenvalue problem. this turns to the standard eigenvalue problem B~1Ax = Xx (or AB~1w = Xw). then Z~lx isa right eigenvector of QAZ—XQ B Z. the eigenvalues of the problems A — XB and QAZ — XQBZ are the same (the two 1. see.2 Canonical Forms Canonical Forms Just as for the standard eigenvalue problem. Since the latter involves a pair of matrices. 6. Z e Cnxn such that 12. det(QAZ . If B is nonsingular. with the understanding onal elements of Ta to the corresponding diagonal elements of Tp. 3. 0 ( Q ~ H y)H Q(A X AB)Z = O.8.2 12. Sec. QBZ = TfJ . Then 1. with the understanding that a zero diagonal element of Tp corresponds to an infinite generalized eigenvalue.AQBZ are the same (the two problems problems are said to be equivalent). see.7. Again. E c nxn such that QAZ = Ta . B E Cnxn Then there exist unitary matrices Q. Sec. 2. 7. Let A.8. for example.7].l Ax Ax (or AB. 7. solving the generalized eigenvalue problem. Since det Q XB). det(QAZXQBZ) = det[0(A . Z e Cnxn with Q and Z nonsingular. the theoretical foundation for the QZ algorithm. 6. Canonical Forms 12. that a zero diagonal element of TfJ corresponds to an infinite generalized eigenvalue.7].7] [25. which is the generally preferred method for theoretical foundation for the QZ algorithm. [7. Let A. Then there exist unitary matrices Q. the pencil A AAB always has precisely n . see.AB). c 3. and the first theorem deals with what happens to eigenvalues and eigenvectors under equivalence. Numerical methods that work directly on A and are discussed in standard textbooks on numerical linear algebra.7]. Sec. Theorem 12. the result follows. fewer than n eigenvalues. Proof: Proof: 1. f i always has precisely eigenvalues. work directly on A and B are discussed in standard textbooks on numerical linear algebra. and the first theorem deals with what happens to eigenvalues lencies rather than similarities. where Ta and TfJ are upper triangular.AB)Z] = det gdet Zdet(A 1.l W AW). However. By Theorem 12. for example. Canonical Forms 127 B is nonsingular. canonical forms are available for the generalized eigenvalue problem. in fact. D The first canonical form is an analogue of Schur's Theorem and forms. we now deal with equivaa matrices. Theorem 12. the eigenvalues of the problems A . in fact.12. There is also an analogue of the MurnaghanWintner Theorem for real matrices. 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. the result follows. the result follows easily by noting that yH(A — XB) — 0 if and only if yH (A . Then 12. .7] or [25. fl. Since det 0 and det Z are nonzero.7] or [25.XB)Z] = detQ det Z det(A . There is also an analogue of the MurnaghanWintner Theorem for real matrices. and det Z are nonzero. Q. ify is a left of AB.7.
In this chapter. B e c mxn .9. . . we present only statements of the basic theorems and some examples. J .. Generalized Eigenvalue Problems Theorem 12. B e Cnxn and suppose the pencil A . including analogues of principal vectors and description of of so forth.. Otherwise.12 mxm nxn mxm nxn E C nonsingular nonsingular matrices P e c and Q e c QE C such that peA . real eigenvalues. T.10. mxn E C • Theorem 12. Example 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 . KCF. thnt that QAZ = S. .I.XB is regular.12 (Kronecker Canonical Form). where T is upper triangular and S is quasiuppertriangular. Let A. Z e R"xn such B E jRnxn.AB)Q = [~ ~ ] .AB. form (KCF). [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 .11. quasiuppertriangular. A full description of the KeF. Generalized Eigenvalue Problems Chapter 12.9. Then there exist 12.)"N). 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). QBZ = T. B e Rnxn. Then there x exist nonsingular matrices P.'. E jRnxn 12. Then there exist orthogonal matrices Q. The first theorem pertains only to "square" regular pencils.A. of eigenvalues are given as above by the ratios of diagonal elements of S to corresponding elements of T.128 Chapter 12. Let A. L l" L~.. Let A. Q € c nxn"such that nonsingular E C" such that peA .• L.. B E c nxn pencil — AB Theorem 12. When S has a 2 x 2 diagonal block. 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.AB where J is a Jordan canonical form corresponding to the finite eigenvalues of A A.2)2 with characteristic polynomial (A — 2)2 has a finite eigenvalue 2 of multiplicty 2 and three 2 2 infinite eigenvalues. of — XB. I . is beyond the scope of this book.11.AB)Q = diag(LII' .A [~ ~ l of . The matrix pencil 12.
corresponds LQ.XB is regular. both N and J are in Jordan canonical form.e. LQ.2. Example 12. Canonical Forms 12. Such a matrix is in KCF. Left Left or right minimal indices can take the value O.. Specifically. while each LQ has "zero rows" and L6. 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. R ( S <S. i. n(S)) = S. Let A.e. suppose S e Rn* xk is a matrix whose columns span a kdimensional E ~nxk ^dimensional subspace S of ~n. Lo. B e Wlxn and suppose the pencil A .2. Lo L6 one column. next two correspond to correspond J = 21 0 2 [ o 0 while the nilpotent matrix N in this example is N [ ~6~].14. where each LQ has "zero columns" and one row. and L^ is the (k + I) x k where N is nilpotent.4) eigenvalue characterization Just as in the standard eigenvalue case.35). generalized eigenproblem. The /( are called the left minimal indices while the r. Lo.5) .13. Consider a 13 x 12 block diagonal matrix whose diagonal blocks are A 0] I o A I . (12.The next two blocks second block L\ one the block is L\. Then is deflating subspace for the pencil A AB if and only if there exists M E Rkxk such that e ~kxk AS = BSM. are called the right minimal indices. The second block is L\ while the third block is LI. Then V is a E ~nxn suppose pencil — AB deflating subspace if deflating subspace if dim(AV + BV) = dimV. there is a matrix characterization of deflating subspace. Canonical Forms 129 where N is nilpotent. Lo. both Nand J are in Jordan canonical form. 000 Just as sets of eigenvectors span Ainvariant subspaces in the case of the standard eigenvectors eigenproblem (recall Definition 9. L6. LQ. Then SS is aadeflating subspace for the pencil A . Definition 12.. (12. LQ .12. there is an analogous geometric concept for the eigenproblem generalized eigenproblem.— XBif S Rn. The first block of zeros actually corresponds to LQ. i. 0.
u is the vector of inputs or controls. However.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. (12. which is clearly equivalent to If B = I. and y is the vector of outputs or observables.4) becomes dim(AV + V) = dimV. where x(= x(t)) is called the state space model is often used in multivariable control theory. 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. [26].6) drops rank. which has a root at —2.130 Chapter 12. vector. Checking the finite eigenvalues of the pencil (12. multioutput systems. we offer some insight below into the special case of a singleinput.6). multioutput systems. for example. In the special case p = m. where the "system pencil" (12.8. see.5) becomes AS = SM as before. we which clearly has a zero at 2. these values are the generalized eigenvalues of the (n + m) x (n m) pencil. is a concept analogous to deflating subspace called a reducing subspace. one must be well for general mUltiinput. which is clearly equivalent to AV c V. Ac M D "'" 5A + 14. 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. In the special case p = m.6). Example 12. one must be careful first to "deflate out" the infinite zeros (infinite eigenvalues of (12.5) becomes AS = SM as before.8. however. we offer some insight below into the special case of a singleinput. 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). Generalized Eigenvalue Problems If B = /. E jRnxm.6». Similarly. u is the vector of inputs or controls. Checking the finite eigenvalues of the pencil (12. B] .4) becomes dim (A V + V) = dim V.6)).15. (n + m) x (n + m) pencil. In general. and D € Rpxm. then (12. and y is the vector of outputs or observables. For details. Numerically.8. there is a concept analogous to deflating subspace called a reducing subspace. B € R" xm . D=O. For details. If the pencil is not regular. and E jRPxm. 12. Let A=[ 4 2 C = [I 2]. = Cx + Du E jRnxn. [26]. Numerically.8. trivial. (12. This linear with A € M n x n . 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. however. there AV ~ V. C e Rpxn. E jRPxn.15. for example. Similarly.3 12. zeros). these values are the generalized eigenvalues of the drops rank. This is accomcareful first to "deflate out" the infinite zeros (infinite eigenvalues of (12. the (finite) zeros of this system are given by the (finite) complex numbers where the "system pencil" z. Let Example 12. see. then (12. where x(= x(t)) is called the state vector. we find the characteristic polynomial to be find the characteristic polynomial to be det [ which has a root at 2. The connection between system zeros and the corresponding system pencil is nonThe connection between system zeros and the corresponding system pencil is nontrivial. However. lEthe pencil is not regular.
z is a zero of g. B E Rnxn arises when A = A and B = BT > O.4.7) (12. there are no "pole/zero cancellations").8) c T x +dy = O.4 12.nxn A AT and B the B1 0. Thus. Symmetric Generalized Eigenvalue Problems 12. and D e R r T(s7 .. let B = b E Rn. the problem (12.12. Now _y 1= 0 (else x = 0 from (12. g. relatively where n(s) is the characteristic polynomial of A.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. we have Substituting this in (12.10).10) is equivalent B. Symmetric Generalized Eigenvalue Problems 131 131 1 singleoutput system. Suppose z € is such that Suppose Z E C is such that [ A . B~11A is not necessarily B~ Ax = AX. Now y ^ 0 (else x z i.A)~ ! Z? + d denote the system transfer function (matrix). 12. Hence g(z) = 0. B e ffi.e.. symmetric.zI cT b ] d is singular.8).9».n.. of the Since B is positive definite it is nonsingular. Then there exists a nonzero solution to or or (A . g(s) Furthermore.9) Substituting this in (12. b e ffi. e ffi.8). However. (12.e.zl)lby + dy = 0. C = c T E R l x n . let g(. system of differential equations differential Mx+Kx=O.l xn. 0 from (12.e. then from (12. and D = d E R. we have _c T (A . and v(s) and n(s) are relatively prime TT(S) v(s) TT(S) (i. M K where M is a symmetric positive definite "mass matrix" and K is a symmetric "stiffness definite "stiffness matrix. no pole/zero cancellations). Hence g(z) 0.zl)lby. or g ( z ) y = 0 by the definition of g. Specifically.9))." is a frequently employed model of structures or vibrating systems and yields a frequently generalized eigenvalue problem ofthe form (12.10) for A.7) we get get x = (A .zl)x + by = 0. "pole/zero cancellations"). (12. or g(z)y 0 by the definition of g. the problem (12. the secondorder A.10) is equivalent Since B is positive definite it is nonsingular.4. . A pole/zero Assuming z is not an eigenvalue of A (i.A to the standard eigenvalue problem Bl1Ax = AJC.4 Symmetric Generalized Eigenvalue Problems Symmetric Generalized Eigenvalue Problems Ax = ABx A very important special case of the generalized eigenvalue problem (12. For example. Thus.
then C = C T > 0. where L is nonsingular (Theorem 10.fi Then it is easily checked that Then it is easily checked thai c = L~lAL~T = [ 0. Theorem 12. are eigenvectors of the original generalized eigenvalue problem and satisfy and satisfy (Xi. it has a Cholesky factorization B = LL T. Moreover. if A = AT> 0. Finally. Example 12. with corresponding eigenvectors zi... generalized case A and B are Hermitian.5 ' 3. the eigenvalue problem eigenvalue problem Ax = ABx has n real eigenvalues..12) has n real eigenvalues.11) can then be rewritten as AL J and Z = LT x. we have restricted our attention to that case only. but since realvalued matrices are commonly used in most applications. and are Hermitian. (12.5 2. be generalized easily to the case where A material of can. Finally. Xj)B T T = xr BXj = (zi L ~l)(LLT)(L ~T Zj) = Dij..17.23).5 ] 1. so the eigenvalues are positive. Then the generalized A.11) can be rewritten as the equivalent problem 1 Letting C = L ~I AL ~T and z = L1 x.18.16. E !!. .16. Let A Example 12. y)BB = XT By.1926 as expected. are eigenvectors of the original generalized eigenvalue problem Xi Zi. Let A. Zn Zj = Dij. zn satisfying vectors Z I. but since realvalued matrices are commonly used in most applications. with corresponding eigenSince C = C T.16). (12. Proof: Since B > 0. Let A = [~ .16 is D 0 L=[~ .16 is Example 12. the eigenvalues of B l A are always real (and are approximately 2. (12. then = C T > 0.11) can then be rewritten as = Cz = AZ. Generalized Eigenvalue Problems Example 12..18.1926 and —3. then product y) x T By. of course.12) Since C = C T the eigenproblem (12. B E Rnxn with A = AT and B = BT > 0. it has a Cholesky factorization B = LLT.5 2.. .. Moreover. and the n corresponding right eigenvectors can be chosen to be orthogonal with respect to the inner product (x.. positive. the eigenvalues are also all positive. l = [i ~ J B ThenB~ A Then A B~Il = [~ ~ J B~I A approximately Nevertheless. we have restricted our attention to that case only. so the eigenvalues are positive. the eigenproblem (12.132 132 Chapter 12. if A = A > 0. The Cholesky factor for the matrix B in Example 12.1926 whose eigenvalues are approximately 2. where L is nonsingular Proof: Since B > 0. •.. = L ~Tzi.. Generalized Eigenvalue Problems Chapter 12.1926 in Example 12. B e jRnxn A AT and B BT > O. zi Then x. if A > 0. Then the eigenvalue problem (Theorem 10. ii € n.12) has n real eigenvalues. if orthogonal > 0.fi 1] .23).1926 and 3. Then the eigenvalue problem Ax = ABx = ALL Tx (12. The Cholesky factor for the matrix B in Example 12. The material of this section can. ..
Theorem 12.e. There are many such results and we present only a representative (but important and useful) theorem here.. Let Q = L .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. i.e. D > I.1A. = pT L I(LLT)L T P = pT P = [. Let A. we restrict our attention only to the real case. there exists Q E lR~xn such that QT AQ = D and QT BQ = [.20. by Theorem 10. Proof: By Theorem 12. Let A. Again. such results and we present only a representative (but important and useful) theorem here. In numerically problematic. when L is highly ill conditioned with respect to inversion.. Thus.5. Infact. individually. A1. Q D. B e M" xn be positive definite. normal matrices can be diagonalized by a unitary similarity. it does preserve the eigenvalues of A .20.T P. There are many matrices (A. In particular. where D is C is symmetric. simultaneous reduction can also be accomplished via an SVD. where D is diagonal. B) can be simultaneously diagonalized by the same matrix. Then diagonal. In such cases. since A > B. haveA(D) = A(B..1A = Q1l B~1QT QT AQ = Q11B. where D is diagonal.19 (Simultaneous Reduction to Diagonal Form).< / (this is trivially true 0 since the two matrices are diagonal). Now D > 0 by Theorem 10. D since the two matrices are diagonal). Also. To illustrate. Let A = QT AQandB = QT BQ.g." The following is typical. since QDQ~l have A(D) = A(B~1A). Then A 2: B if and only if B~ 2: AI. Proof: Let B = LLT be the Cholesky factorization of B and set C = L~1AL T.31. the diagonal elements of D are the eigenvalues of B. B E lRnxn be positive definite.. Then there exists a nonsingular matrix Q such that where D is diagonal. LetA QT AQ and B QT Then/HA Q~ B. But then D. Let A.5 Simultaneous Diagonalization Simultaneous Diagonalization Recall that many matrices can be diagonalized by a similarity. Then there exists a nonsingular matrix Q such that A = AT and B = BT > 0. so it does not preserve eigenvalues of A and B individually. normal maRecall that many matrices can be diagonalized by a similarity. In particular. Since Proof: Let T C is symmetric. Then A > B if and only if Bl1 > Theorem 12.1A).1 12.g. matrices to "the diagonal case. there exists an orthogonal matrix P such that pTe p = D.19 is very useful for reducing many statements about pairs of symmetric Theorem 12. Thus.5.5 12.21 we have that QT AQ > QT BQ. e.5. we restrict our attention only to the real case. It turns out that in some cases a pair of matrices (A.l Q~T QT Q~ B~ AQ. so it does not preserve eigenvalues of and B Note that Q is not in general orthogonal. there exists an orthogonal matrix P such that P CP = D. where D is diagonal. B) can be simultaneously diagonalized by the same matrix. since A 2: B.21 we have that QT AQ 2: QT BQ. i. A~l :::: Bl1.19. Simultaneous Diagonalization 12. with the complex case following in a Again. straightforward way. simultaneous reduction can also be accomplished via an SVD. In fact.. Theorem 12.19 is very useful for reducing many statements about pairs of symmetric matrices to "the diagonal case. since QDQI Finally. let such cases. A I < B~ .1AQ.19 is There are situations in which forming C = L~1AL~T as in the proof of Theorem 12. we Note that Q is not in general orthogonal.19. Also. To illustrate.12. Now D > 0 by Theorem 10. Simultaneous Diagonalization 133 12.'AB. It turns out that in some cases a pair of trices can be diagonalized by a unitary similarity. i. However.19 e ][~nxn A AT and B BT > O. Proof: By Theorem 12. the diagonal elements of D are the eigenvalues of B 1A. by Theorem where D is diagonal.31. Then B. This can be seen directly. D 2: [. Then and and QT BQ Finally. i.e.. with the complex case following in a straightforward way. = 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.lI QT :::: Q QT.19 is numerically problematic. there exists Q e E"x" such that QT AQ = D and QT BQ = I. Theorem 12. This can be seen directly.e. we B~ 1 A.5. B E E"x" with 12. when L is highly iII conditioned with respect to inversion. But then D"1I :::: [(this is trivially true 10. it does preserve the eigenvalues of A — XB. \ 2. Since LLT be the Cholesky factorization of and setC L I AL~T. Let Q = L~T P. QD~ QT < QQT. let ." The following is typical. e. However.
D b . example. when A = AT > 0.21.15) is called a generalized singular value problem and algorithms exist to problem generalized solve it (and hence equivalently (12.13» via arithmetic operations performed only on LA LA (12.21 example.. without forming the products LALTA or LBLTB explicitly. for LB 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). Remark 12.7. let A = LALTA and B — LBL~ us assume that both A and B are positive definite.e.e. products LA L ~ LBL~ see. but in writing = PDDp D diagonal. The case when A is symmetric but indefinite is not so A = AT::: O. operations performed directly on M rather than by forming the matrix MT M and solving performed MT forming the eigenproblem MT MX = AX.14) rewritten the LAL~x = ALBz = A L g L ^ L g 7 z . Generalized Eigenvalue Problems Chapter 12.15) The problem (12. A can be written as A = PDP T.13) where E E R£ x " isisdiagonal. Various generalizations of the results in Remark 12. Note that LB A and thus the singular values of L B 1 LA can be found from the eigenvalue problem 02. [7. i. D may have pure imaginary elements. let A = LAL~ and B = LsLTB be Cholesky factorizations of A and B. The SVD in (12.21 are possible. see. (12. Then the matrix Q U performs the simultaneous L e 1R~ xn diagonalization. Sec. note that T QT AQ = U Li/(LAL~)Li/U = UTULVTVLTUTU i/ = while L2 QT BQ = U T LB1(LBL~)Li/U = UTU = I.14) Letting x = LB z we see that (12. Remark 12..13)) and LB separately. Compute the SVD Cholesky factorizations A B. PDPT ~ ~ ~ ~ T PD(PD{ with where Disdiagonaland P is orthogonal.134 134 Chapter 12. A straightforward.butin writing A — PDDP T = PD(PD) with D is diagonal and P orthogonal. which is thus equivalent to the generalized eigenvalue problem ALBL~LBT z. at least in real arithmetic. respectively. 8. for generalizations results 12. For example. Further. Further. Generalized Eigenvalue Problems us assume that both A and B are positive definite. To check this. respectively. which is thus to the generalized eigenvalue problem 02. This is analogous to finding the singular values of a matrix M by Sec.22. eigenproblem MT M x Xx.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].
e..16) or. ::: ILn· Let a>k = IILk I!. . seek A A2 M + AC + To get a nonzero solution /?. C.16) we get (12.2M + A.12.6.• 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.6 12. If r = n (i. Suppose K has eigenvalues eigenvalues IL I ::: .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.6. = [ M1K 0 x (t) E ~2n. and A special case of (12. K e Rnxn. or if it is desired to avoid the calculation of M lI because M is too ill conditioned with respect to inversion.e.C + K is singular. ± Wk. ::: ILr ::: 0 > ILr+ I ::: .. p. (A 2 M + AC + K) p = O. Since the determinantal equation is singular. then all solutions of q + Kq = 0 are oscillatory. and = = KT. by analogy with the firstorder case. yields a polynomial of degree 2rc.6.1 12. If M is singular.16) of the p A are to be determined.. where q(t} e W1 and M. HigherOrder Eigenvalue Problems 135 12.. Suppose. .16) can be written as a firstorder system (with block companion matrix) X . k = 1.. (12. C = 0. . the secondorder problem (12. Substituting in q(t) = eAt p. 12. there are 2n eigenvalues for the secondorder (or A2 M + AC + K. are to be determined. where the nvector p and scalar A.... .6. then all solutions of q K q 0 are oscillatory. KT > 0).2M + A.. M Mwhere x(t) €. A special case of (12. for which the matrix A.6 HigherOrder Eigenvalue Problems HigherOrder Eigenvalue Problems Mq+Cq+Kq=O.16) arises frequently in applications: M = I. . Then (12. that we try to find a solution of (12. k = r + 1. we thus seek values of A. Suppose K = KT. (12.. Since the determinantal equation o = det(A 2 M + AC + K) = A2n + .2 K are are ± jWk.. K = KT ::: 0). n. E2".C + K. If r n (i.16) Consider the secondorder system of differential equations Consider the secondorder system of differential equations q(t) E ~n E ~nxn. quadratic) eigenvalue problem A.1 Conversion to firstorder form Conversion to firstorder form Let x\ = q and \i = q.16) arises frequently in applications: 0. Assume for simplicity that M is nonsingular. Substituting in form q(t) = ext p.. r.16) can be written as a firstorder system (with block Let XI q and X2 Then (12. HigherOrder Eigenvalue Problems 12. since eAt :F 0. polynomial 2n.
and C e lRmxn. In the parlance of control theory. Some can be useful when M. F 6 Rm *" G R" x . Generalized Eigenvalue Problems Chapter 12. .19). (A similar result is also true for "nonsquare" pencils.1 2.16) involving. Similar procedures hold for the general k\horder difference equation order difference equation which can be converted to various firstorder systems of dimension kn. Let € C M X • Show that the nonzero eigenvalues of and G F are the same. Generalized Eigenvalue Problems Many other firstorder realizations are possible. derivative q. G E enxn". Suppose A e Rnxn and D E lR::! xm. B e lRn*m.) . E Rnxm and E E 4. 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 ". Show that the nonzero eigenvalues of FG and GF are the same. and/or K have special symmetry or skewsymmetry properties that can exploited. andlor K Many other firstorder realizations are possible.. such results show that zeros are invariant under state feedback or output injection.B D. Let F. C. 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. lead naturally naturally involving. EXERCISES EXERCISES nx 1. which can be converted to various firstorder systems of dimension kn. Show that the generalized eigenval". 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 (11.19). 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 . Some can be useful when M. C. the kth derivative of q. Suppose A € Rnxn. say. Are the FG and GF the 3. G e Cmxn • Are the nonzero singular values of FG and GF the same? same? wx E ]Rnxn.136 136 Chapter 12. Similar procedures hold for the general kthblock companion matrix analogue of (11. properties Higherorder analogues of (12. Let F e Cnxm .. Hint: Consider the equivalence I G][AUO F0]' B][I l [01 C (A similar result is also true for "nonsquare" pencils. In the parlance of control theory.
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. and let UWT be an SVD of L~LA'. Ql = ~!UTL~. positive Cholesky = LA L ~ = L B L ~. (b) Show that Q~l = ^~^UT LTB. 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.2 and hence are AB E2 positive. A and B to the same diagonal matrix. respectively. A B B are positive definite with Cholesky factorizations A = L<A and B = L#Lg. (c) Show that the eigenvalues of A B are the same as those of 1. B E e jRnxn Ql AQT ]Rnx" in such a way that Q~l AQ~T and QT BQ are simultaneously diagonal. Consider the case where both A and transformation contragredient. positive.Exercises Exercises 137 137 desired 5. . respectively. Such QT BQ a transformation is called contragredient.
This page intentionally left blank This page intentionally left blank .
. pointing out the extension to the complex case only where it is not obvious.2.. Example 13.1) amnB Obviously.A @ B. the same definition holds if A and B are complexvalued matrices.. Then 3. 2B 2B ~J. 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. n 2. Foranyfl E lRX(7. Example 13. 4 3 4 3 4 9 4 2 6 2 6 6 6 2 2 Note that B @ A i. Let A e R mx ". Let B be an arbitrary 2x2 matrix. We Obviously. Let A E lRmxn B E R Definition 13.1. 1. B e lR pxq.1. the same definition holds if A and B are complexvalued matrices. Let A = [~ 2 2 nand B = [. / 2 <8>fl = [o ~ l\ 2. (13. Then A@B =[ 3~ ~]~U J.1 13. Note that B <g> A / A <g> B.2. Let B be an arbitrary 2 x 2 matrix. We restrict our attention in this chapter primarily to realvalued matrices. Then 0 b ll b12 B @/z = l b" b~l 139 0 b2 2 0 b21 0 0 b12 0 b 22 l . extension to the complex case only where it is not obvious. pointing out the restrict our attention in this chapter primarily to realvalued matrices. 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.Chapter 13 Chapter 13 Kronecker Products Kronecker Products 13.1 Definition and Examples Definition and Examples Definition 13.
Then 13.3.3.5. 5 E R r x i . C e ~nxp. 4.. B In x E ~m.2) Proof: Simply verify that Proof. . simply note that (A ® B)(A 1 ® B. Proof: Proof: Using Theorem 13.. (A ® B)I = Bare 13. and D e Rsxt. . xmYnf E !R.3. y e !R. (13.. Then 13. 5. XIYn. Then X ® Y = [ XIY T . D Corollary 13.2 13.m xm are symmetric. XmY T]T = [XIYJ..1 ) Theorem 13. then A® B is symmetric.n. A® 13. If AI ® B. Theorem 13.5. Foral! Proof' Proof: For the proof. If A and B are nonsingular.kCkPBD L~=1 amkckpBD ] 0 Theorem 13. Let* eR m . simply verify using the definitions of transpose and Kronecker verify transpose Kronecker 0 product. L~=l al. . = 1 ® 1 = I. y eR". Let Jt € Rm. B e ~rxs.1.4. Let E ~mxn. Simply verify that ~[ =AC0BD.2 Properties of the Kronecker Product Properties of the Kronecker Product (A 0 B)(C 0 D) = AC 0 BD (E ~mrxpt).. For all A and B.. If E ]Rn xn e Rmxm are Theorem 13. . E R". If A e R"xn and B E !R.6. C E R" x ^ and D E ~sxt. (A ® Bl = AT ® BT. mn . . 0 . X2Yl. .140 Chapter 13.3. Kronecker Products Kronecker Products The extension to arbitrary B and /„ is obvious. Let A e R mx ".6.
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)..8. . . In general.JLqq (q < m)./u. then Xi <8> Zj ffi. . 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). 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. Let A E R nx "have eigenvalues Ai.3 since A and B are normal by Theorem 13. Theorem 13.... Properties of the Kronecker Product 13. If A E E"xn is orthogonal and B E Mmxm is orthogonal."xn have singular values UI :::: .• :::: TS > O. Ap (p ::::: and ZI.12.c. xp are linearly independent right eigenvectors of A corresponding Moreover.n. = (A 0 B)(A 0 B)T 0 Corollary 13... matrix A ® 5 is then also orthogonal with eigenvalues e^'^+'W and e ± ^ (6> ~^ > \ Theorem 13. TTzen ?/ze mn eigenvalues of A 0 Bare Moreover. . ••. 0 If A and Bare diagonalizable in Theorem 13.. elements of ~A 0 ~B and the corresponding right and left singular vectors). A. then A <g> B is € IR nxn orthogonal and e IR m x m 15 then 0 is orthogonal. j e q.. xp are linearly independent right eigenvectors of A corresponding AI.p (p < n).10. eigenvectors of A® B corresponding to A. if Xl. <I :::: . Then A <g)B (or B 0<8> A) has rs singular values U. q Corollary 13. .. we can take p = nand q = m and n and q —m and If A and B are diagonalizable in Theorem 13.. then A 0 B is normal. A0 B e±jeH</» e±jefJ </».. if A and fi have Jordan form .• 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.Zq are linearly independent right eigenvectors of B corresponding to JLI.• :::: U rr > 0 and let B E IRfx Corollary e R™x" singular a\ > • • > a > e have singular values T\ > • • > <s > 0. Let A E lR. .3.. Then the mn eigenvalues of A® B are eigenvalues JL j. If A e IR nxn am/ B E IR mxm are normal. .7.7... L et A E xamp Ie 139 Let A = [ _eose cose andB . and let eigenvalues jJij. Properties of the Kronecker Product Theorem 13.. • • zq independent of to A .. i / E e!!.10.12. mxm /zave Theorem 13.9.and let BB E e IRR mxwhave e IR nxn have eigenvalues A.. we can take p thus get the complete eigenstructure of A 0 B.... In general. 7 E m. .4 by Theorem 13. The 4 x 4 orthogonal e±j9 orthogonal eigenvalues e±j(i>. Lgf A E E mxn have a singular value decomposition VA ~A Theorem 13.i .. Then A 0 B (or B A) has rs singular values have singular values <I :::: .8. if A and B have Jordan form thus get the complete eigenstructure of A <8> B. if x\. Sine] and B . \Ju (q ::::: m). If Corollary 13. . Example 13. If A E IR"xn and B eRmxm are normal. 0 Zj E€ IR mn "are linearly independent right eigenvectors of A 0 B corresponding to Ai JL 7 i e /?. then A® B is normal.m are linearly independent right corresponding to JJL\ .11.•. :::: U rTs > 0 and ^iT\ > • • • > ffr <s Qand rank(A 0 B) = (rankA)(rankB) = rank(B 0 A) .2.. and zi.12.j... 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.. j € m..2.[Cos</> cos</>O Then It IS easl'1y seen that . .13. then .
Let A e Rn Xn and B e Rm xrn. then we get the JA and Q~] BQ following Jordanlike structure: following Jordanlike structure: (P ® Q)I(A ® B)(P ® Q) = (P. Tr(A ® B) = (TrA)(TrB) = Tr(B ® A).I ® Ql)(A ® B)(P ® Q) = (P. A EEl B ^ B EEl A. respectively. is the mn x mn matrix Urn <g> A) + (B ® In).15. respectively. Then the Kronecker sum (or tensor sum) . 1. Let A e Rn xn and B e Rrn xm. E IR E IR Kronecker Definition 13. suppose P and Schur form for A ® B can be derived similarly.14. det(A ® B) = (det A)m(det Bt = det(B ® A).e. For example. respectively.. general. 1. ~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. denoted A EEl B. Example 13. Then reducing A and B to real Schur form). . 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). Then (P ® Q)H (A ® B)(P ® Q) = (pH ® QH)(A ® B)(P ® Q) = (pH AP) ® (QH BQ) = TA ® TR . Let A~U Then Then 2 2 !]andB~[ .1 AP) ® (Ql BQ) = JA ® JB · Note that h ® JR. suppose P and Q are unitary matrices that reduce A and B. Corollary 13. i. are unitary matrices that reduce A and 5.. Note that. Then 13. in general. with A EEl B. of A and B.142 142 Chapter 13. Example 13. nxn mxm Definition 13. eigenvalues are zero or nonzero). Let 1. in of A and B. is the mn mn matrix (Im ® A) + (B ® /„). then we get the decompositions given by P~lI AP = J A and Ql BQ = JB. For example.13. E IR nxn E IR mxm. A Schur form for A ® B can be derived similarly. to Schur (triangular) form. A ® B i= B © A. is generally not quite in Jordan form and needs Note that JA® JB. denoted A © B. i.e. Kronecker Products Chapter 13.15. Note that.14. while upper triangular. to Schur (triangular) form.13. while upper triangular. 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). respectively. Kronecker Products decompositions given by p. 2.AP J B .
. ii E E. Then the Kronecker sum A® B = (1m (g>A) + (B ® In) has mn (Im ® A) + (B <g> /„) /za^ ran eigenvalues fJj.13. . .. we can take p = n and q = m and thus get the complete eigenstructure of A $ In general. 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 . xp are linearly independent right eigenvectors of A corresponding to AI. . .•• . A2 + fJm.. if A and B have Jordan form pI l B .2. and z\. TTzen r/ze Kronecker sum A $ B eigenvalues e/genva/wes Al + fJt.. zq are linearly independent eigenvectors of corresponding to fJt.\ . In general. Xp (p ::s: n). 0 I M 0 where M = [ where M = o M a f3 f3 a J... j e q.i e n. . . e jRmxm eigenvalues /z. Properties of the Kronecker Product 2. i E !!..xp are linearly independent right eigenvectors of A corresponding Moreover.. is a Jordanlike structure for A © B. if XI. .···. . .. (I} ® M) + (E^®l2) = M 0 Ek. respectively. 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). Let A E E"x" have eigenvalues Ai.. if A and have Jordan form thus get the complete eigenstructure of A 0 B. Recall the real JCF M I M 143 143 0 I M I 0 o 1= 0 E jR2kx2k. . + fJj' € p. eigenvectors of A® B corresponding to Ai + [ij. A2 + fJt. Zq are linearly independent right eigenvectors of B AI.. j E ra. . 7 e I!!.. . AI + fJm. Recall the real JCF 2. then Zj ® Xi E€ jRmn" are linearly independent right Zj <8> Xi W1 are linearly independent right corresponding f j i . j E fl· eigenvectors of A $ B corresponding to A... . Then J can be written in the very compact form J Theorem 13.16.. 0 If A and Bare diagonalizable in Theorem 13... . Properties of the Kronecker Product 13... 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. An + fJm' Moreover. f^q (q ::s: ra). . then decompositions given by P~1AP = lA and Q"1 BQ = JB. . . .16.. fJq (q < m). if x\. and let B E Rmx'" have e jRnxn eigenvalues A.16. . respectively.2.. Ap (p < and ZI. we can take p nand q and If A and B are diagonalizable in Theorem 13.
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). pH AP = TA matrices that reduce A and B. [(Q ® /„)(/« ® P)] = (<2 ® P) is unitary by Theorem 13.3) is. Sylvester who studied general linear matrix equations of the fonn k LA.3) in tenns of their columns. Lyapunovequations arise naturally in stability theory.5) clearly can be written as the Kronecker sum (1m 0 A) + The coefficient matrix in (13..Xj.144 Chapter 13.3) in terms of their easily seen z'th columns that ith columns that m AXi + Xb.4) obtained by taking B = AT.5) as (B T 0 /„).5) [ blml The coefficient matrix in (13.3 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.4) is known as a Lyapunov equation. Again. (13.3) is the symmetric equation AX +XAT = C (13. it is easily seen by equating the writing (13. an "ordinary" linear system. When does a solution exist? By writing the matrices in (13. Then ((Q ® /„)(/« ® P)]"[(/m <8> A) + (B ® /B)][(e (g) /„)(/„. suppose P and Q are unitary fonn. to Schur (triangular) form. where [(Q <8>In)(lm ® P)] = (Q ® P) is unitary by Theorem 13. Then to real Schur fonn).3 13.e.3 and Corollary 13. When does a solution exist? The first important question to ask regarding (13. . = AXi + l:~>j. = C.J.4) is known as a Lyapunov equation. 13. and C e M" xm . This equation is now often called a Sylvester equation is now often equation in honor of 1. i. ® P)] = (/m <8> rA) + (7* (g) /„).8. The first important question to ask regarding (13. Sylvester who studied general linear matrix equations of the form equation in honor of J. the solution X E Wnx" is easily shown taking B = AT. respectively. Kronecker Products Chapter 13. i.1.5) clearly can be written as the Kronecker sum (Im * A) + (BT ® In). Again.3 and Corollary 13.5) as an "ordinary" linear system. The following definition is very helpful in completing the writing of (13. When symmetric. arise naturally in stability theory. =C. suppose P and are unitary A Schur form for A © B can be derived similarly. .. When C is symmetric. solution e IR xn also to be symmetric and (13.=1 A special case of (13.XB.3) is.8.3) mxm E IRnxn E IR E IRnxm.. Sylvester where A e R"x". 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.e. Lyapunov equations also to be symmetric and (13. The following definition is very helpful in completing the writing of (13. . Kronecker Products A Schur fonn for A EB B can be derived similarly. B e Rmxm .
7) has a unique solution if and only if A and . AX+XB=C (13. B e Rmxm.(B) ^ solution to(13. B E Rmxm.3) (or symmetric Lyapunov equations of the form (13. where A. But [(1m ® A) + (B (g) /„)] nonsingular and only has no zero eigenvalues. 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. the eigenvalues of [(/m <g> A) + (BT ® In)] are Ai A. ofC e jRnxm [CI. elegant connections between matrix theory and stability theory for differential equations. They culminate in Theorem 13..B have no eigenvalues in common. the linear system (13. From Theorem 13.6) if and only if [(Im ® A) + (BT ® /„)] is nonsingular. xn Theorem 13. n > m. Now integrate the differential equation X AX XB (with X(O) C) on [0.18. + IJLJ. They culminate in Theorem 13. Let A e jRnxn. E R E jRnxm. (13..6). 77ie/i Theorem 13. Then the (unique) solution of the Sylvester equation AX+XB=C (13. There exists a unique solution to (13.. this algorithm takes only O(n3 ) operations rather than the O(n6)) that would be required by solving (13. First A and B are reduced to commonly preferred numerical algorithm is described in [2]..17. The next few theorems are classical.10) . i. Sylvester equations of the form (13.24. We thus have the following theorem.3. A further enhancement to this algorithm is available in [6] whereby Gaussian elimination. A(fi).8) by Theorem 13.17.16. has a unique solution if and only if A and —B have no eigenvalues in common. c ]. and C e R" xm . .. so there exists unique Proof: Since A and B are stable.. .1S. one of many The next few theorems are classical.6) directly with operations rather than the O(n 6 that would be required by solving (13. Definition 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.e A (A). 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).16..6) if and only if [(1m ® A) + (B T ® In)] is nonsingular. Cm}.. j j so there exists aaunique for all i. 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. . say. Let A e lRnxn. say. and ^j Theorem 13. j j E!!!. Then the Sylvester equation G jRmxm.3) (or symmetric Lyapunov equations of the form Sylvester equations of the form (13. vec(C) = Using Definition 13. Then the (unique) solution of the Sylvester equation parts in the open left halfplane).6) There exists a unique solution to (13.17. e m. Suppose further that A and B E Rn . the linear system (13.and Mj Ee A(B).e. But [(Im <8>A) + (B TT ® In)] isisnonsingular ififand only ififitithas no zero eigenvalues.8) can be written as can be written as (13. where From Theorem 13. +00): IHoo lim XU) .6) directly with Gaussian elimination.9) Proof: Since A and B are stable. this algorithm takes only 0 (n 3) transformed solution matrix X.3. one of many elegant connections between matrix theory and stability theory for differential equations.24. Assuming that.18. E jRmxm. . +00): (with X(0) = C) on [0. n :::: m. Let Ci( € E. Now integrate the differential equation X = AX + X B solution to (13.n denote the columns ofC E Rnxm so that C = [ n .4» are generally not solved using the mn x mn "vec" formulation (13. Application to Sylvester and Lyapunov Equations 13. and C e Rnxm. We thus have the following theorem. the eigenvalues of [(1m ® A) + (BT <8> /„)] are + Mj. The most (13.17. .19.6).X(O) = A 10 roo X(t)dt + ([+00 X(t)dt) 10 B. (13. 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).13. First A and B are reduced to (real) Schur form.. ii e n_.. Ai E A(A).8)by Theorem 13. Assuming that.4)) are generally not solved using the mn x mn "vee" formulation (13. E!!. Theorem C E jRnxm. Aj(A) + A. Schur form. Application to Sylvester and Lyapunov Equations 145 145 Definition 13. (real) Schur form. (A)+ Aj(B) =I 00 for all i.5) can be rewritten in the form Using Definition 13.18. A.
ATT have A —A. . Theorem Substituting in (13. using the solution X ((t) = elACe tB from Theorem 11. 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.24.8).. (13. v E". Many useful results exist concerning the relationship between stability and Lyapunov equations. Two basic results due to Lyapunov are the following.. . Hence. _* ]). Theorem 13..22. where C Proof: asymptotically l3. . we have that lim X ((t) = 0.. C e jRnxn further asymptotically stable. Let A. If the matrix A E Wxn has eigenvalues A. By Theorems 13.19. +00 r—>+oo t—v+oo X t ) = etACelB X t ) — O.21.12).12) Theorem 13... —kn.6.. then that solution is symmetric.!„. results = 0. If matrix A e jRn xn eigenvalues )"" . TTzen r/ze AX+XAT =C (13.23 solution Proof: Suppose A is asymptotically stable.6. Remark 13. . A. If C is has unique if and only if and —A T eigenvalues in common.AT has eigen— AT eigenvalues AI. If symmetric and (13. then that solution is symmetric.I . the first of which follows immediately from Theorem 13. C E R"x". Now let v be an arbitrary nonzero vector in jRn.21 l3.21 and 13.10) 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. symmetric and ( 13. Theorem 13. Then the (unique) solution o/the Lyapunov equation of the AX+XAT=C can be written as can be written as (13. Then the Lyapunov equation e jRnxn.146 146 Chapter 13.]. . X B = is that [ J _Cfi ] be similar to [~ _OB] (via the similarity [ Let Theorem 13.23 a solution to (13.11) has a unique solution if and only if A and .11) has a unique solution.A T have no eigenvalues in common. . Lef A.11) has a unique solution. Kronecker Products Using the results of Section 11. a sufficient condition that guarantees that A and . sufficient —A common eigenvalues A asymptotically no common eigenvalues is that A be asymptotically stable. Remark 13.An. 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 _~ ]). Thus..20.. then .C E R"x" and suppose further that A is asymptotically stable..19.23. . An. it can be shown easily that lim elA = lim elB = O.. 1>+00 1 . Kronecker Products Chapter 13.13) where C = C T < O.13) exists and takes the form (13. Then Then .1..
Then 0> yHCy = yH AXy + yHXAT Y = (A + I)yH Xy. D asymptotically stable. defined. A must be asymptotically stable. B. Let A E Rmxn. Then vector y. result. 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. Then the equation 13. the integrand above is positive. in which the solution is of the form is of the form (13. for the solution of the simple Sylvesterlike equation introduced in Theorem 6. most of which derive from one key 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). C. e jRrnxn. the AXB =C (13. Hence Since C > 0 and etA is nonsingular for all t. B e jRPxq.14) xp E jRn has a solution X e R. Since A was arbitrary.11. where Y e Rnxp is arbitrary.3. v TXv > 0 and thus X is positive definite. Application to Sylvester and Lyapunov Equations 13. D Remark 13.13. Theorem 13.16) . 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(.26. suppose X = XT > 0 and let A. B.26.3. e A(A) with corresponding left eigenvector y. Application to Sylvester and Lyapunov Equations 147 147 Since — C > 0 and etA is nonsingular for all the integrand above is positive. For any three matrices A.14) as Proof: Write (13. The Lyapunov equation AX X A = C can also be written using the Remark 13. A must be Since yHXy > 0. where Y E jRnxp is arbitrary. The vec operator has many useful properties. The solution of (13. 14) is unique if BB+ ® A+A = [. and C for which the matrix product ABC is defined.14) as (B T ® A)vec(X) = vec(C) (13. suppose X = XT > 0 and let A E A (A) with corresponding left eigenConversely. vec(ABC) = (C T ® A)vec(B). 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).yr) = <8> x. in which case the general solution has a if only ifAA + C B+ C. most of which derive from one key result. For any three matrices A. we must have A + I = 2 Re A < 0 .27. nx p if and only if A A+CB+BB = C.t. The Proof: Write (13. we must have A + A = 2 R e A < O.14) is unique if BB+ ® A+ A = I.25. Since yH Xy > 0. Theorem 13. B E Rpx(}. the complexvalued equation AHX + XA = C is equivalent to [(/ ® AH) vec(C). D tions or from the fact that vec(xyT) = y ® x. Conversely.25. However.27. Since A was arbitrary.11. and C E Rmxq. Hence vT Xv > 0 and thus X is positive definite. 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. and C for which the matrix product ABC is Theorem 13. the complexvalued equation H X X A = C is equivalent to However.15) of (13.
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
518–521 [Erratum. FiniteDimensional Vector Spaces. NJ. [3] Bellman. 1991. Cambridge Univ. 1972..5758.F. York. and C. New [3] Bellman. N. 909913. [11] Higham. 1996. P. [9] Greville. AX + XB = C." IEEE [7] Golub.. "Note on the Generalized Inverse of the Product of Matrices. [5] Cline.H. and G.w. S..H. T.. G. FiniteDimensional Vector Spaces. and C.. Second Edition. C. [8] Golub. R. AC24(1979). 57–58. Van Loan." SIAM Rev.F.N.R. New [1] York. R. [4] Bjorck. NY." SIAM Rev. G. Baltimore. New York. PA. [11] Higham. T. "Note on the Generalized Inverse of the Product of Matrices. "Algorithm 432. UK. 18(1976). 249]. 1970. Solution of the Matrix Equation [2] Bartels. 8(1966). Control. Cambridge. Univ. 2002. 15(1972). NY..578619. SIAM 9(1967). Johnson. to Second Edition. Solution Equation AX + X B = C.E.. [4] Bjorck.... Autom. Van Loan. and C. 578619. 1991.. Regression and the MoorePenrose Pseudoinverse. Press. Second Edition. Wilkinson. "Algorithm 432. Third Edition. ACM.H. Press. NJ. Cambridge. [2] Bartels.E. Cambridge Univ. RA. 8(1966). R. RH. 1996. G. SIAM. Matrix Analysis.. and J. 1996.A.. [10] Halmos.H.H. and c.A. G.N. and C.. PA. 1985. 1972.E. Numerical Methods for Least Squares Problems. Second Edition. 151 151 . Philadelphia. R. Johnson. Accuracy and Stability ofNumerical Algorithms.518521 [Erratum. Stewart.. 249]. Numerical Methodsfor Least Squares Problems. PA. "A HessenbergSchur Method for the Problem AX + XB = C.. R. G.. "Note on the Generalized Inverse of a Matrix Product. Cambridge. SIAM.1. Matrix Analysis. [9] Greville.R. S. and G." Cornm. 6(1964)..E. AC24(1979). "A HessenbergSchur Method for the Problem [6] Golub. Wilkinson. Academic Press." Comm. Second Edition.. 1958. [12] Hom. [5] Cline." SIAM Rev. Accuracy of'Numerical Algorithms.R. A." IEEE Trans. Princeton. of the SIAM 18(1976). N. Press. MD. York. 1958.820826. 1996. Van Loan.. Press. UK. Press.H. Philadelphia. "IllConditioned Eigensystems and Computation ofthe Jordan Canonical Form. Philadelphia. McGrawHill. SIAM. R. NY. Johns Hopkins [7] Golub. Horn. and lH. G..H.. Introduction to Matrix Analysis. 6(1964). AX X B = C. Matrix Computations. R. Baltimore.R. UK. A. 2002.Bibliography [1] Albert. Princeton. 1970. [10] Halmos. Van Nostrand. Analysis. Third Edition. "Note on the Generalized Inverse of a Matrix Product. 820826.R. Stewart.. [6] Golub. Horn.. Topics in Matrix Analysis. Nash. RA. Univ. MD.W. Van Loan... Cambridge Univ. Johns Hopkins Univ. Nash. SIAM. UK. and C. NY.H. "IllConditioned Eigensystems and the Computation [8] Golub. Van Nostrand. 1985. 9(1967). Matrix Computations... Cambridge. 15(1972). Johnson. McGrawHill. and C.J." SIAM Rev." SIAM Rev. SIAM Rev. [13] Hom. PR. Second Edition. Philadelphia.
Third Edition. New York. 1988. NY.. SpringerVerlag." SIAM 20(1978). New York. Third Edition. 913–921.. and M. of [25] Watkins. Signals. A. PA. NY. [22] Penrose. New York. SpringerVerlag. FL. Academic Press. [24] Strang. San Diego. 1988. Linear Multivariable Control.W. A Geometric Approach. B. Academic Press. Englewood Cliffs. 801836. [21] Ortega. Daniel.152 152 Bibliography Bibliography [14] Kenney. Second Edition. Philadelphia. "A Generalized Inverse for Matrices. Tismenetsky. [23] Stewart. [20] Noble." IEEE Trans. 1973. [21] Ortega.S." Math. "Controllability and Stability Radii for Companion Form Systems.. Linear Algebra and Its Applications.. A Second Course. "Controllability and Stability Radii for Companion Fonn [14] Kenney. San Diego. and M. Systems...M... Analysis Applied Linear [18] Meyer. 1(1988). Applied Linear Englewood Cliffs. .S. Theory of Second Edition with Applications.. NY.J. W. 1985. 913921. 1985. Tismenetsky." Math. Soc. and A. [18] Meyer. and A. Introduction to Matrix Computations. R. Laub.D. Laub.." IEEE Trans .. 1988.801836. c. NY. 2002. Applied Linear Algebra. [19] Moler. and Systems.J. The Theory of Matrices. Van Loan. NY. Second Edition. [20] Noble. Daniel.. [23] Stewart.. NJ. [26] Wonham. Autom.w.B. C. of a Matrix. New York. C. "The Matrix Sign Function. and J." IEEE Trans. Wiley[25] D. New York. 1985. Control. P. SIAM. AC24( 1979). New York.. Control. of Control. 1973. C. NJ. CA.406413.S. Third Edition.. [22] Pemose. 20(1978). Fundamentals of Matrix Computations..J. [26] Wonham. 1979). "A Generalized Inverse for Matrices. P. Control. G. Linear Algebra and Its Applications. [15] Kenney. Harcourt Brace Jovanovich. Plenum. Second Edition with [16] Lancaster.W. AJ. Laub." Proc. 51(1955). Orlando. and J. 1987.. andAJ. "The Matrix Sign Function. 40(1995).. Harcourt Brace [24] Strang. Philadelphia. 2002.. W. NY. to Academic Press. of Control.. 1988. [15] Kenney.. G. Applications. "Nineteen Dubious Ways to Compute the Exponential [19] Moler. New York.. Autom. Soc. "A Schur Method for Solving Algebraic Riccati Equations. [16] Lancaster... D. FL. [17] Laub. Laub. G. Third Edition. 51(1955). 361390. "Nineteen Dubious Ways to Compute the Exponential of a Matrix. 40(1995). Third Edition. PA. Orlando. Interscience. 1(1988).M. 1330–1348. 2000. R. Third Edition.B." SIAM Rev." IEEE Trans.13301348. C. [17] Laub. Academic Press." Proc. 1987. Cambridge Philos.D. Van Loan. Interscience. CA.F. Autom.S. Matrix Analysis and Applied Linear Algebra. Jovanovich. 2000... C. SIAM. 406–413. C. Matrix Theory. W. 1985. G. PrenticeHall. "A Schur Method for Solving Algebraic Riccati Equations. and c. J..361390.. C. Control.P. and C. PrenticeHall. Plenum. and AJ.
4 determinant. 106 singular values of. 81 mation. 11 basis. 48 LV factorization. 17 eigenvalue. 90 algebraic multiplicity. 90 matrix characterization of. 85 of a principal vector. 39. 115 exponential of a matrix. 84 equivalence transformation. 104 diagonalization. 89 matrix characterization of. 5 properties of. 39. 23 function of a matrix. 128 generalized real Schur form. 149 congruence. 106 complement complement of a subspace. 125 generalized real Schur form. 58 basis. 75 chain chain of eigenvectors. 1 companion matrix companion matrix inverse of. 109 exponential of a matrix. 103 congruence. 48 inverse of. 75 Cayley–Hamilton Theorem. C".58 ity. 103 conjugate transpose. 150 inverse of. 109112 field. 125 of a matrix pencil. 114118 inverse of. 17 co–domain. 12 direct sum direct sum of subspaces. 58 CayleyHamilton Theorem. 75 of a matrix. 1 CauchyBunyakovskySchwarz InequalCauchy–Bunyakovsky–Schwarz Inequality. 114–118 computation of. 4–6 dimension. 106 singular vectors of. 106 singular values of. 12 dimension. 137 controllability. 89 A–invariant subspace. 75 of a matrix pencil. 115 exponential of a Jordan block. 5 triangularization. 23 four fundamental subspaces.Index Index Ainvariant subspace. 7 field. 23 vector. 23 rank. 106 singular vectors of. 76 defective. 76 algebraic multiplicity. 75 invariance under similarity transforinvariance under similarity transformation. 95 equivalence transformation. 87 characteristic polynomial characteristic polynomial of a matrix. 106 pseudoinverse of. 127 exchange matrix. 1 . 2 definiteness of. 58 angle between vectors. 125 generalized eigenvalue. 95 equivalent generalized eigenvalue probequivalent generalized eigenvalue problems. 95 unitary. four fundamental subspaces. 81 generalized eigenvalue. 91. 150 diagonalization. 21 153 . 95 orthogonal. 46 properties of. 101 Cholesky factorization. 17 domain. 125 Cholesky factorization. 21 orthogonal.81 elementary divisors. 2 block matrix. 128 e (pmxn mxn en. 2 conjugate transpose. 5 LU factorization. 81 function of a matrix. 105 inverse of. 13 domain. 81. 110 inverse of. 85 determinant. 76 angle between vectors. 105 pseudoinverse of. 1 vector. 12 natural. 12 block matrix. 13 orthogonal. 84 elementary divisors. 4 of a block matrix. 87 of eigenvectors. 5 of a block matrix. 13 of subspaces. 109–112 properties of. 101 codomain. 137 contragredient transformation. 89 exchange matrix. 2 contragredient transformation. 89 exponential of a Jordan block. 46 defective. 95 unitary. equivalent matrix pencils. 76 degree degree of a principal vector. 11 natural. 109 computation of. 13 of a subspace. 110 properties of. 91. 127 lems. 17 column column rank. 75 eigenvalue. 104 definiteness of. 81. 46 controllability. 127 equivalent matrix pencils. i 1 (p/nxn 1 e~xn.
45 linear independence.2 7. 44 characterization of all solutions. 47 Index Index Kronecker product. 71 71 decomsolution via singular value decomposition. 142 transpose of. 10 linear least squares problem. 120 firstorder higherorder eigenvalue problems higher–order problems conversion to first–order form. 20 nullspace j. 18 nonsingular. 140 Kronecker sum. 149 leading principal submatrix. 2 idempotent. 67 linear transformation. 140 pseudoinverse of. 121 conversion firstorder higher–order differential equations higherorder conversion to first–order form. 142 determinant eigenvalues of. 65 uniqueness of solution. 17 invertible. 112 equations. 26 invertible. 51 identity matrix. 4. 82 Kronecker canonical form (KCF). 148 singUlar values of. 65 problem. inner product inner product complex. 75 left generalized eigenvector. 136 firstorder i. 22 left left principal vector. 54 real. 141 eigenvectors of. 54 usual. 17 co–domain codomain of. 141 singular trace of. 120 higher–order for homogeneous linear difference homogeneous equations. 51 idempotent. 70 statement of. 6. 66 linear regression. 125 left invertible. 65 solution via QR factorization. 4 complex Euclidean. 112 for inhomogeneous linear difference for inhomogeneous linear difference equations. 100 left eigenvector. 82 Jordan canonical form (JCF). 142 eigenvalues of. 118 for homogeneous linear differential homogeneous equations.2 i. 54 weighted. 19 domain of. 2 Jordan block. 44 existence of solutions. 44 existence of solutions. 76 Holder Inequality. generalized decomposition. 139 determinant of. 20 Kronecker . 143 eigenvectors of. 4 inertia. 10 linear equations linear equations characterization of all solutions. 10 linear independence. 25 nulls pace of. 141 eigenvectors products of. 25 left invertible. 134 134 geometric geometric multiplicity. 129 Kronecker canonical Kronecker delta. 127 generalized generalized singular value decomposition.154 generalized Schur form. 85 linear dependence. matrix representation of. 143 exponential of. 143 eigenvectors of. 44 uniqueness of solutions. general solution of. 67 residual of. 4 Euclidean. 84 inverses of block matrices. 54 invariant factors. 2 transpose. 17 composition of. 119 for inhomogeneous linear differeninhomogeneous differential equations. 66 geometric solution of. higherorder higher–order difference equations conversion to first–order form. 26 invertible. 55 complex Euclidean. 58 Hermitian transpose. left nullspace. 45 uniqueness of solutions. 109 initial–value for higherorder equations. 103 initialvalue problem.
76 MoorePenrose Moore–Penrose pseudoinverse. 98 negative definite. matrix–matrix. 2 nearest singular matrix to. and asymptotic stability. 20 left. 57 nullity.5 block. 91 of square root of a. 6 block. 2 lower triangular. 126 singUlar. 109 Hamiltonian. conditions for. 62 unitarily invariant. 56 induced.60 2–. 127 reciprocal. 2 quasiuppertriangular. 81. 2 symplectic. 22 22 right. 146 uniqueness of solution.60 mutually consistent. 92 92 nonnegative definite. 25 onto. 2 unitary.60 2.60 1–. 109 matrix exponential. 99 definite. 22 observability. 113 Lyapunov differential equation. 91. 20 right invertible. 2 upper Hessenberg. 99 negative invariant subspace.