Matrix Analysis
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
. Laub Alan J.Matrix Analysis Matrix Analysis for Scientists & Engineers for Scientists & Engineers Alan J. California slam. Laub University of California Davis.
Inc. info@mathworks.. 2. Matrices.com.mathworks.com 5086477000.com.L38 2005 512.. Mathcad is a registered trademark of Mathsoft Engineering & Education. Fax: 5086477101. 3600 University City Science Center. stored.. or transmitted in any manner without the written permission may be reproduced. Includes bibliographical references and index. Natick. 1948Laub.lam. 2005 by the Society for Industrial and Applied Mathematics. MATLAB® is a registered trademark of The MathWorks.. Mathcad is a registered trademark of Mathsoft Engineering & Education.) ISBN 0898715768 (pbk. I.. Alan J. MA 017602098 USA. Matrices. www. PA 191042688. I. PA. Natick. Inc. Inc. Laub. 5086477000. Inc. write to the Society for Industrial and Applied of the publisher. Philadelphia. MA 017602098 USA. Laub. 10987654321 10987654321 All rights reserved. For MATLAB product information. Title. 3 Apple Hill Drive. No part of this book All rights reserved..com Mathematica is a registered trademark of Wolfram Research. Inc. Matrix analysis for scientists and engineers / Alan J.9'434—dc22 512. please contact The MathWorks. Mathematical analysis. artist Aaron Tallon of Philadelphia. Fax: 5086477101. cm. QA188138 2005 QA 188. Inc. No part of this book may be reproduced.Copyright Copyright © 2005 by the Society for Industrial and Applied Mathematics. Printed in the United States of America. Used by permission 5. cm. Mathematica is a registered trademark of Wolfram Research. 1948Matrix analysis for scientists and engineers / Alan J. p. ISBN 0898715768 (pbk. wwwmathworks. 2. Used by permission. Inc. Philadelphia. Mathematics. 3 Apple Hill Drive.) 1. . PA. is a registered trademark. Alan J. MATLAB® is a registered trademark of The MathWorks. • slam is a registered trademark.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. Title. For MATLAB product information. 3600 University City Science Center. p. Library of Congress CataloginginPublication Data Library of Congress CataloginginPublication Data Laub. write to the Society for Industrial and Applied Mathematics.. For information. Inc. Mathematical analysis. info@mathworks. please contact The MathWorks. stored. Includes bibliographical references and index. For information. Printed in the United States of America. PA 191042688. 1. or transmitted in any manner without the written permission of the publisher.
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 Sums and Intersections of Subspaces 2.3 Row and Column Compressions Linear Equations Linear Equations 6.2 Subspaces.. .Contents Contents Preface Preface 1 1 Introduction and Review Introduction and Review 1. .3 Properties and Applications .. 5. .4 Structure of Linear Transformations 3.4 Determinants Vector Spaces Vector Spaces 2.3 Inner Products and Orthogonality 1. 6.1 Definition and Examples . . .2 Matrix Linear Equations . .. 3.2 Examples 4.4 Sums and Intersections of Subspaces Linear Transformations Linear Transformations 3.2 Matrix Arithmetic .1 Definitions and Characterizations Definitions and Characterizations. . . . 6. 2. ..1 Definitions and Examples . . Definitions and Examples 2.. . .3 Linear Independence 2. 4.3 Composition of Transformations . .3 Linear Independence . .2 Subspaces 2. ..3 Properties and Applications Introduction to the Singular Value Decomposition Introduction to the Singular Value Decomposition 5. . .. .2 Examples. .3 Rowand Column Compressions 5. . . .. . .1 Definition and Examples 3.1 2. .1 The Fundamental Theorem 5. . 4. .. 4.2 Matrix Representation of Linear Transformations 3.1 The Fundamental Theorem . .4 Some Useful and Interesting Inverses.. .3 A More General Matrix Linear Equation 6. . .2 Matrix Representation of Linear Transformations 3.2 Some Basic Properties 5.. .3 Composition of Transformations 3. .5 Four Fundamental Subspaces . .. 2.4 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 4. . ..2 Matrix Linear Equations 6. . . 1.3 Inner Products and Orthogonality ..4 Determinants 1.. 6. .. 1. . 3. . .4 Structure of Linear Transformations 3.2 Matrix Arithmetic 1.1 Vector Linear Equations . . 5. . .1 Some Notation and Terminology 1.5 Four Fundamental Subspaces Introduction to the MoorePenrose Pseudoinverse Introduction to the MoorePenrose Pseudoinverse 4. . .1 Some Notation and Terminology 1.2 Some Basic Properties .1 Vector Linear Equations 6.. 3. .3 A More General Matrix Linear Equation 6.
1 Fundamental Definitions and Properties 9. . .4 Geometric Aspects of the JCF 9. . .1. 11.2 Homogeneous linear differential equations 11.1 7. 8.3 Determination of the JCF . 11. .1. .5 Least Squares and QR Factorization 8. . 7. .3 HigherOrder Equations. . . Inner Product Spaces. . . . . . . .1. . 9.3. . . .3 Vector Norms 7. . . .2 Homogeneous linear differential equations 11. 10.1 Homogeneous linear difference equations 11. . . .1 The Linear Least Squares Problem 8. .4 Least Squares and Singular Value Decomposition 8. . . . . . . .2 Definite Matrices 10.2 Jordan Canonical Form 9. . . Example: Linear regression 8.1 Projections .1 Fundamental Definitions and Properties 9.2 Jordan Canonical Form . .3. . . . .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. .5 Least Squares and QR Factorization .2 Difference Equations 11. .3 HigherOrder Equations . . . . . . .3.1. .6 Computation of the matrix exponential 11. .2.5 Modal decompositions . . .3. . .5 The Matrix Sign Function. . and Norms 7.1 8. .1. . . .viii viii Contents Contents 7 Projections. . . .3 Inhomogeneous linear differential equations 11. . .1. .4 Linear matrix differential equations . . . .4 Rational Canonical Form .5 Modal decompositions 11. . . .3 Inhomogeneous linear differential equations 11. 11. . .1. . . . .1 Example: Linear regression .1 The Linear Least Squares Problem .1. . .1.1. . .2 Geometric Solution 8. .4 Geometric Aspects of the JCF 9.2 Other least squares problems 8. . .4 Linear matrix differential equations 11. . . .2.2 Difference Equations .1 Block matrices and definiteness 10. . . . . .3.1. .2 Inner Product Spaces 7.1 Theoretical computation . .3 Equivalence Transformations and Congruence 10.3 Computation of matrix powers 11.2 On the + l's in JCF blocks 9. 11. . 7. . .3.1 Block matrices and definiteness 10.1 9. 11.2 Geometric Solution . . .1 Some Basic Canonical Forms 10. .2 Other least squares problems . .2 Inner Product Spaces 7. .2 Inhomogeneous linear difference equations 11.3 Linear Regression and Other Linear Least Squares Problems 8.1. 10. . 9.1 Differential Equations ILl Differential Equations .4 Rational Canonical Form 11 Linear Differential and Difference Equations 11 Linear Differential and Difference Equations 11.3. 11.4 Least Squares and Singular Value Decomposition 8. .1 The four fundamental orthogonal projections The four fundamental orthogonal projections 7.1 Homogeneous linear difference equations 11. . . 8.2. .1 Projections 7. .3 Linear Regression and Other Linear Least Squares Problems 8.1 Properties ofthe matrix exponential .2. .1. Eigenvalues and Eigenvectors 9. . .3 Equivalence Transformations and Congruence 10. . .1 Some Basic Canonical Forms . . Theoretical computation 9. .3 Determination of the JCF 9. . .2 Definite Matrices . 8. . .3. 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 . .6 Computation of the matrix exponential 11. . . . .1 Properties of the matrix exponential 11. .1.3 Vector Norms 7. 11. . . .2 On the +1's in JCF blocks 9. . . . 9. . . . .3. . .4 Matrix Norms Linear Least Squares Problems 8. .3 Computation of matrix powers .4 Matrix Norms . .3. . . . . . .2. 10. .2. .2 Inhomogeneous linear difference equations 11. . .
3 Application to Sylvester and Lyapunov Equations 13.6. . . .2 Canonical Forms 12. . 13. .5. . . . . . .1 Simultaneous diagonalization via SVD 12. . 12. . . .6. .1 Definition and Examples . 12. . .6 HigherOrder Eigenvalue Problems 12. . .6 HigherOrder Eigenvalue Problems . 12. .5.1 Definition and Examples 13. .3 Application to the Computation of System Zeros . 13.1 The Generalized EigenvaluelEigenvector Problem 12.4 Symmetric Generalized Eigenvalue Problems . .1 The Generalized Eigenvalue/Eigenvector Problem 12.1 Simultaneous diagonalization via SVD 12. . .2 Properties of the Kronecker Product .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. . . .2 Canonical Forms .4 Symmetric Generalized Eigenvalue Problems 12. 12. . . .5 Simultaneous Diagonalization 12. . .5 Simultaneous Diagonalization .. .3 Application to Sylvester and Lyapunov Equations Bibliography Bibliography Index Index . . . .3 Application to the Computation of System Zeros 12. 12. .1 Conversion to firstorder form 12. . . .2 Properties of the Kronecker Product 13. .Contents Contents ix ix 12 Generalized Eigenvalue Problems 12 Generalized Eigenvalue Problems 12.
This page intentionally left blank This page intentionally left blank .
I highly recommend MAlLAB® although other software such as a digital computer. students meant to learn thoroughly as undergraduates. 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. Matrices are stressed more than abstract vector spaces. for example). These powerful and versatile tools doinverses and the singular value decomposition (SVD). The text linear dynamical systems (systems of linear differential or difference equations). methods. for example). 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. in many cases. [II]. but somehow didn't quite manage to do. singularity of matrices. the sciences. but somehow didn't quite manage to do. 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. followon topics on the computational side (at the level of [7]. and mathematical structures.. eigenvalues and eigenvectors. and concepts such as determinants. Certain topics thoroughly as undergraduates. Ortega [21]. even though their recollecmatrices least tion may occasionally be "hazy. particular subject area. [23]. Because tools such as the SVD are not generally amenable to "hand computation. or [25]. mathematics. For this. The books by Meyer [18].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. the student is then wellequipped to pursue." this approach necessarily presupposes the availability of appropriate mathematical software on approach necessarily presupposes the availability of appropriate mathematical software on a digital computer. although Chapters 2 and 3 algebra. I highly recommend MATLAB® although other software such as xi xi . Noble and Daniel [20]. computer science." However. example) or on the theoretical side (at the level of [12]." However. 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.e. mathematics.e. By matrix analysis I mean linear tools and ideas comfortably in a variety of applications. [13]. I have tried throughout to emphasize only the and more advanced material is introduced. For this. or [16]. students meant to learn much of the important and useful mathematics that. I have tried throughout to emphasize only the more important and "useful" tools. eigenvalues and eigenvectors. 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. 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. Basic concepts such as determinants. computer science. and Strang [24] Ortega are excellent companion texts for this book.. in many cases. These powerful and versatile tools can then be exploited to provide a unifying foundation upon which to base subsequent topcan exploited to foundation subsequent topics. the sciences. example) or on the theoretical side (at the level of [12]. singularity of matrices. Matrices are stressed more than abstract vector spaces. 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. either via formal courses or through selfstudy. basisfree or subspace) aspects of many of the fundamental notions. either via formal courses or through selftext. Basic of calculus and definitely some previous exposure to matrices and linear algebra. requiring such material as prerequisite permits tion may occasionally be "hazy. The concept of matrix factorization applications and by computational utility and relevance. basisfree or subspace) aspects of many of the fundamental do cover some geometric (i. for [11]. or computational students in engineering. or [16]. and positive definite matrices should have been covered at least once. Upon completion of a course based on this text." this ics. [13]. Upon completion of a course based on this are excellent companion texts for this book. although Chapters 2 and 3 do cover some geometric (i. Because tools such as the SVD are not generally amenable to "hand computation. the student is then wellequipped to pursue. 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). 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).
completed the course. how "nearly dependent" are the vectors? If they linearly independent. econometrics. Since this text is not intended for a course in numerical linear algebra per se. they are either obvious or easily found in the literature. The "language" in which such described models are conveniently described involves vectors and matrices.xii xii Preface Preface Mathcad® Mathematica® or Mathcad® is also excellent. and modem engineering. mathematics. Rather. if only they had had this course before they took linear systems. remarked afterward that if processing. and the course has proven to be remarkably successful at enabling students from Davis. 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 acceptable graduate rigor for subsequent graduate studies in a variety of disciplines. The presentation of the material in this book is strongly influenced by computais influenced by computational issues for two principal reasons. are deferred to such a course. outputs. Mastery of the material in this text should enable the student to read and understand the modern language of matrices used throughout mathematics. First. many students who completed especially offered. statistics. I have taught this material for many years. This is ideal not given explicitly. and a wide variety of other fields. and thus the text can serve a rather diverse audience. Statespace methods are Statespace modem now standard in much of modern engineering where. form the foundation Some of the key algorithms of numerical linear algebra. the details of most of the numerical aspects of linear algebra per se. This is ideal material from which to learn a bit about mathematical proofs and the mathematical maturity and insight gained thereby. simulated. prerequisites developed While prerequisites for this text are modest." For example. consistent. When they are not given explicitly. the student does require a certain amount of what is conventionally referred Proofs referred to as "mathematical maturity. But in most engineering or scientific contexts we want to know more than that." Proofs are given for many theorems. Indeed. "reallife" problems seldom yield to simple "reallife" closedform formulas or solutions. they are either obvious or easily found in the literature. It is my firm conviction that such maturity is neither encouraged conviction neither nor nurtured by relegating the mathematical aspects of applications (for example. diverse audience. Some of the key algorithms of numerical linear algebra. in particular. It is thus crucial to acquire knowledge vocabulary a working knowledge of the vocabulary and grammar of this language. in an elementary linear algebra course. and states often give rise to models of very numbers models high order that must be analyzed. and while most material is developed from basic ideas in the book. especially the first few times it was offered. and evaluated. form the foundation virtually modem upon which rests virtually all of modern scientific and engineering computation. and coherent fashion. . chemistry. science. If a set of vectors is linearly independent." a set of vectors is either linearly independent or it is not. They must generally be solved computationally and closedform it is important to know which types of algorithms can be relied upon and which cannot. finiterange floatingpoint arithmetic environment of of of most modem computing platforms. A second motivation for a computational emphasis is that it provides many of the essential tools for what I call "qualitative 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. The tools of matrix analysis are also applied on a daily basis to problems in biology. must lay firm foundation upon which and perspectives perspectives can be built in a logical. applied physics. 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. control systems with standard large numbers of interacting inputs. for example. one must lay a firm foundation upon which subsequent applications and Rather. This is an absolutely fundamental fundamental concept. in particular. are deferred to such a course. linear algebra introducing "onthefly" algebra for elementary statespace theory) to an appendix or introducing it "onthef1y" when to necessary. If If are linearly dependent. or signal processing.
. June 2004 — AJL.. The concept seems to work. they would have been able to concentrate on the new ideas deficiencies they wanted to learn. they no longer had to provide as much time for "review" and could focus instead on the subject at hand. etc. rather than having to spend time making up for deficiencies in their background background in matrices and linear algebra. AJL.Preface Preface xiii XIII or estimation theory. My fellow instructors. too. realized that by requiring this course as a prerequisite.
This page intentionally left blank This page intentionally left blank .
where Xi E R for ii E !!. 5. 1R. 2. the notation!! denotes the set {I. Rn = the set of ntuples of real numbers represented as column vectors.. . xyT is an n x n matrix..1 Some Notation and Terminology Some Notation and Terminology We begin with a brief introduction to some standard notation and terminology to be used We begin with a brief introduction to some standard notation and terminology to be used throughout the text. e. XTy is a scalar while it easy to recognize immediately throughout the text that. Henceforth..n xn Cmxn = the set of complex m x n matrices of rank r.. IR rn xn = the set of real (or realvalued) m x n matrices. . The following sets appear frequently throughout subsequent chapters: The following sets appear frequently throughout subsequent chapters: 1. Note: Vectors are always column vectors. 1 . 2. . Henceforth. That a vector is always a y E IR n and the superscript T is the transpose operation. nonsingular n x n matrices. This is followed by a review of some basic notions in matrix analysis and linear algebra. but this convention makes column vector rather than a row vector is entirely arbitrary. n }. 5.g. IR~ xn denotes the set of real = set of real of rank Thus. where y G Rn and the superscript T is the transpose operation. n}. x e IR n means means where xi e IR for e n. Crnxn = the set of complex (or complexvalued) m x n matrices. and linear algebra. but this convention makes it easy to recognize immediately throughout the text that.. = the set of complex m x n matrices of rank r. the notation n denotes the set {1. x T y is a scalar while xyT is an n x n matrix.Chapter 1 Chapter 1 Introduction and Review Introduction and Review 1.g. Thus. Rnxnn denotes the set of real nonsingular n x n matrices.. A row vector is denoted by y~.n xn Rmxnr = the set of real m x n matrices of rank r. e.1 1. e. 3.. e 6. IR n = the set of ntuples of real numbers represented as column vectors. A row vector is denoted by yT where Note: Vectors are always column vectors. R mxn = the set of real (or realvalued) m x n matrices. Cn = the set of ntuples of complex numbers represented as column vectors. the set of ntuples of complex numbers represented as column vectors. Thus. en 4. mxn = the set of complex (or complexvalued) x n matrices. This is followed by a review of some basic notions in matrix analysis throughout the text. That a vector is always a column vector rather than a row vector is entirely arbitrary. x E Rn I. Thus.
The The transpose of a matrix A is denoted by AT and is the matrix whose (i. it is easy to see that if Aij are appropriately dimensioned subblocks.. ~ 5 is symmetric (and Hermitian). • tridiagonal if aij = 0 for Ii . If A E em xn. • lower triangular if a. then r = [ . where the bar indicates complex sometimes A*) and its = IX jfJ (j = = v^T). • tridiagonal if a(y = 0 for z — j\ > 1.JI > 1. } is While R is most commonly denoted by i in mathematics texts. otherwise noted. • upper triangular if aij. There is some the more common notation in electrical engineering and system theory.e.. Each of the above also has a "block" analogue obtained by replacing scalar components in the respective definitions by block submatrices. Note that if A E R mx ".j  7+} ] is Hermitian (but not symmetric). = 0 for i ^ j. then A7" e jRnxm. and definitions block submatrices.. an equation like A = A T implies that A is realvalued while a statement otherwise noted. For example. A = AH A complexvalued.. Introduction and Review We now classify some of the more familiar "shaped" matrices. • pentadiagonal if aij = 0 for Ii . z Remark 1. that is. Hermitian conjugate sometimes A*) and its (i. Oth (A 7 ). is Hermitian (but not symmetric).. A = [ .ii > 1. j)th entry of a matrix A is denoted by AT and is the matrix whose j)th entry A. • lower triangular if aij7 = 0 for i/ < }. 1. is complexvalued symmetric but not Hermitian. 2 2. it is Transposes of block matrices can be defined in an obvious way. then z = IX — jfi.J I > 2.1. We henceforth adopt the convention that. While \/—\ is most commonly denoted by i in mathematics texts. then the (m + n) x (m + n) matrix [~ Bc] is block upper triangular. • diagonal if aij7 = 0 for i i= }. Introduction and Review Chapter 1. • upper Hessenberg if aij = 0 for ii . The notation j is used throughout the text but reminders are placed at strategic locations. then easy to see that if A. A e jRmxn. then its Hermitian transpose (or conjugate transpose) is denoted by AH (or H If A e C mx ". = 0 for / — j\ > 2.2. B E IR nxm . 2 Transposes of block matrices can be defined in an obvious way.. = 0 for < j. if A E IRnxn.. = 0 for j — > 1. A matrix A E IRn xn e (or A E enxn ) is A eC" x ")is • diagonal if a. (AT)ij = aji. 7 + j ] is complexvalued symmetric but not Hermitian.. . = 0 for i > }. if e Rnxn e Rmxn C e Rmxm then the (m n) x (m n) matrix [A0 ~] is block upper triangular. ] is symmetric (and Hermitian).e. For example. a. We henceforth that. an equation like A = A T implies that A is realvalued while a statement like A = AH implies that A is complexvalued. • upper triangular if a. = 0 for i > j. j is Remark the more common notation in electrical engineering and system theory. i.2..jj > 1. A matrix A is symmetric i. • pentadiagonal if ai. unless if A = A T Hermitian A = A H. For example.. i)th entry of A. Example 1. text but reminders are placed at strategic locations. AT E E" xm is the (j. = a jfJ. A is conjugation. C E jRmxm. 7 = («77). A if A = AT and Hermitian if A = AH. • lower Hessenberg if a. are appropriately dimensioned subblocks. j)\h entry is (AH)ij = (aji). (7.2 2 Chapter 1. There is some advantage to being conversant with both notations. For example. A = [ 7+} 5 3· A . if z = a + jf$ (j = ii = R).[ 7 . • lower Hessenberg if aij = 0 for } . Example 1. • upper Hessenberg if afj = 0 for — > 1. where the bar indicates complex j)th entry is (A H ).
. eRmxn has row cj e E l x ". importance interpretation take A = [96 85 74]x = take A = [~ ~]. applied p times: There is also an alternative. suppose (linear combination) suppose A = la' .~]. n UV T = LUiVr E jRmxp. . and multiplication of matrices.•. As a numerical example... AB bi E W1. vn Rpxn p with Vit e R . matrixvector if C E jRmxn has row vectors cJ E jRlxn.2 Matrix Arithmetic 1... if (C D)H — D C )..2 Arithmetic It is assumed that the reader is familiar with the fundamental notions of matrix addition.3. suppose A e Rmxn and B = [bi. It Theorem 1. Then v E jRP. Theorem 1. i. It is deceptively simple and its full understanding is well rewarded. .. The importance of this interpretation cannot be overemphasized.3 can then also be generalized to its "row dual. the matrixvector product Ax.3.e. suppose A E jRmxn and [hI.. 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. Matrix Arithmetic 3 1...[ ~ J+l.[ ~ J+2. vector x... there can be important computerarchitecturerelated advancomputerarchitecturerelated tages to preferring the latter calculation method. 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. Then we can quickly calculate dot products of the rows of A column Ax = [.2. un]]Ee jRmxn with Ui t Ee jRm and V = [VI..xn~ ] Then Ax = Xjal + . recall that (CD)T = DT C T (C D)T = DT T If H H H (or (CD} = DHC H ). formulation of matrix multiplication that appears frequently in the text and is presented below as a theorem. and is premultiplied by a row vector yTe jRlxm." The details are left to the readei "row left . + Xnan E jRm. That is... Again. {..'" p] E jRnxp For matrix multiplication.. Ax. .• a"1 E m JR " with a... but equivalent. matrixvector product with the column x to find Ax = [50 32]' but this matrixvector product can also be computed computed via v1a 3. A special case of matrix multiplication occurs when the second matrix is a column multiplication second i. Vn] ]Ee lR Pxn U [Uj.[ ~ l For large arrays of numbers. un Rmxn with u Rm and V = [v . multiplication of a matrix by a scalar... Then the matrix product A B can be thought of as above. its importance cannot be overemphasized.1. E JRm and x = l I.bhp ] e Rnxp with hi e jRn. Let U = [MI.e. i=I If matrices C and D are compatible for multiplication. x = ! 2 Then we can quickly calculate dot products of the rows of A [~]. Theorem reader. multiplication.. This gives a dual to the matrixvector result above. . Namely. . Theorem 1.
consider What is true in the complex case is that x H 0 if and only if O. Nonzero complex vectors are orthogonal if x H y = 0. The notation /„ is sometimes used to denote the identity matrix in IR nxn in Rnx" x nxn H H (or en ").4 1. where I is the n x n matrix A e IRnxn is an orthogonal matrix if ATA = AAT = /. What is true in the complex case is that XH x = 0 if and only if x = 0.e. the nonzero vector x above. x)c and we see that.2j while while and we see that. 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. . where / is the n x n identity matrix. Introduction and Review Chapter 1. a matrix A e en xn is said to be unitary if A H A = AA H = I. y E R".3 Inner Products and Orthogonality Inner Products and Orthogonality For vectors x.. There is an orthogonal or unitary matrix has orthonormal rows and orthonormal columns. Note that the inner product is a scalar. Similarly. Y}c = {y.. Introduction and Review 1. for short) of x and For vectors y e IRn. i. If x and y are zero.4. y) := x y = Lx.3 1. y ) c = yHxx = L:7=1 x. the Euclidean inner product inner for short) y is given by y is given by n T (x.y. Similarly.. (x. 1. If e C". indeed. The more conventional definition of the complex inner product is H ( x . x T = O. . Y}c = [ } JH [ ~ ] = [I .4. y)c = (y. Nonzero complex vectors are orthogonal if XHy = O. y)c = (y.y. In sometimes used denote identity matrix. Clearly said = an orthogonal or unitary matrix has orthonormal rows and orthonormal columns. x)c'.=1 Note that the inner product is a scalar. We list below some of (or A 6 en xn) we use the notation det A for the determinant of A. (x. consider the nonzero vector x above. for short) by for short) by n (x'Y}c :=xHy = Lx. x}c.. we define their complex Euclidean inner product (or inner product. we define their complex Euclidean inner product (or inner product. Let x = [1j and y = [1/2]. then we say that x and y are orthonormal. y)c = y = Eni=1 xiyi but throughout the text we prefer the symmetry with the real case. Then XTX = 0 but XHX = 2.. A orthogonal and XTX = 1 and yTyy = 1. A EC = (orC" xn). y e IRn are said to be orthogonal if their inner product is zero. the Euclidean inner product (or inner product. x and y are orthogonal and x Tx = 1 and yT = 1. Then Example 1. Then (x. Example 1. To illustrate. i. E R are said to be orthogonal if their inner product is Two nonzero vectors x.4 4 Chapter 1.. the order in which x and y appear in the complex inner (x.4 Determinants Determinants It is assumed that the reader is familiar with the basic theory of determinants. case. but throughout the text we prefer the symmetry with the real (x. order in which x product is important. We list below some of .=1 y appear in Note that (x. Then x T x = 0 but x H X = 2. xTyy = 0.e.j] [ ~ ] = 1 . then we say that x and y are orthonormal.e. The more conventional definition of the complex inner product is product is important.y. Two nonzero vectors x. y)c = (y. (or A E Cnxn) we use the notation det A for the determinant of A. A nxn matrix E R is an orthogonal matrix if AT A = AAT = I. i. indeed. x)c.e. i. rows or columns. For A E R nnxn A e IR xn It assumed of determinants. Let x = [} ]] and y = [~]. 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. If x. 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. To illustrate.
4... If A is upper triangular. properties 1. Multiplying a column of A by a scalar ex results in a new matrix whose determinant scalar a determinant is ex det A. 16. Multiplying a row of A by a scalar ex results in a new matrix whose determinant is a det A. A 22 . B E IRnxn . is a det A. 11.• ann.B). det A is the product of its diagonal 10. are consequences of one or more of the others. 8. 16.4. then det A = alla22 • • ann 12. then det(A..1. then det A = all a22 . If A E R n x n and D e RMmxm. If A has a zero row or if any two rows of A are equal. of 5. 8. • • An" (of A = square diagonal blocks A11. If det = a11a22 • • ann 12.. If A has a zero column or if any two columns of A are equal. 11. with A block diagonal (or block 13. 15.. Interchanging two columns of A changes only the sign of the determinant. A 11. 9. then det [~ BD] = del A det(D – CA– l 1 B). 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. the determinant. 4. Determinants 5 properties the more useful properties of determinants.. Proof" Proof: This follows easily from the block UL factorization BD. 17. 3. then det A = 0.. then det A = a11a22 . exdetA.e.C). If A.• det Ann.. If A has a zero column or if any two columns of A are equal.. Note that this is not a minimal set.. 2. Interchanging two rows of A changes only the sign of the determinant. det AT = det A (detA H = detA if A e C nxn ). If A.. det A 11 det A22 • • det Ann 14. then det A = O. Multiplying a row of A by a scalar a results in a new matrix whose determinant is 5. If A is block diagonal (or block upper triangular or block lower triangular). B eR n x n . If A A A = o. Multiplying A 6. If A E Rnxn. If A is diagonal. 3. If elements.1 I ][ .• a"n. 7. If A is lower triangUlar. then det A = 0. Determinants 1. . If A is lower triangular. Multiplying a row of A by a scalar and then adding it to another row does not change 7. change the determinant.e. i. 13. If A e IRnxn and D E lR~xm. If A E lR~xn and DE IR mxm det [Ac ~] detA det(D ..CA. then det A = a11a22 . Proof: This follows easily from the block LU factorization Proof" This follows easily from the block LU factorization [~ ~J=[ ~ ][ ~ 17.A22.thendet(AB) = det A det 5. then det A = different det A11 det A22 .. several more is a properties are consequences of one or more of the others. Multiplying a row of A by a scalar and then adding it to another row does not change the determinant. Interchanging two rows of A changes only the sign of the determinant..• ann. i. then det [~ BD] = det D det(A – B D – 11C ) .. det A is the product of its diagonal diagonal. then det(AB) = det A det B. If A e R n x n and D e R m x m . 15..1 ) = de: A. = alla22 • • ann i. Ann (of possibly different sizes). 10. 14.•. then det [Ac ~] det D det(A B D. then det(A1) = 1detA . If A € lR~xn...e.. detAT = detA (det A H = det A A E C"X").
_. Let x. If A is orthogonal.y E jRn. The factorization of a matrix A into the product of a unit lower triangular Remark 1. Show that the product V = VI. 2sin20 J is idempotent for all #. Tr(aA + f3B) = aTrA + fiTrB. i. either prove the converse or provide a counterexample. Tr(Afl) = Tr(£A).. of denoted Tr A. if A. what is det A? If A is unitary. then Tr(aA fiB)= aTrA + f3TrB.. (b) Suppose A e IR" X "is idempotent and A i= I. ! [ 2cos2<9 I T 2cos2 0 (a) Show that the matrix A = _. is defined as the sum of its diagonal A e Rnxn.e.• V k is an orthogonal matrix.. V2.. 2f) 2 _ sm 2^ sin 0 sin sin 20 1 . 6. [24]. A . Show that det(I – xyT) = 1. U2 .B D – l C is the Schur complement of D in [~ BD ]. lower triangular with all l's on the diagonal) and an upper triangular matrix L 1's an V is called an LV factorization. . A? 2. i. aII o. If A e jRnxn and or is a scalar. A E jRnxn A2 / x™ . i. [24].. TrA = Eni=1 aii. denoted TrA. ft e R.Vk € jRn xn be orthogonal matrices.. Remark — C I B – BDIe Similarly. Let A E jRNxn. Suppose A E jRn xn is idempotent and A ^ I. Introduction and Review Remark 1. ST = S.. example. The trace of A. Another such factorization is UL where V is unit upper triangular and L is lower triangular. The factorization of a matrix A into the product of a unit lower triangular matrix L (i. Letx. 4. are block analogues of these. II _ . TrS O. TrA = L~=I au· elements. even though in general AB i= B A. U1 U2 • • Uk is an 5.. (a) Show that the trace is a linear function.. i. Let U1.5. elements.e.6. The matrix D . nxn linear E R f3 E JR. lor z r 2sm2rt # J. Showthatdet(lxyT) 1 – yTx. B e JRn xn and a. Introduction and Review Chapter 1. (b) Show that Tr(AB) = Tr(BA). of Din [AC ~ l EXERCISES EXERCISES nxn 1. • . ST = So Show that TrS = 0. . . Remark 1. what is det(aA)? What is det(–A)? E R a det(A)? A? If A unitary. Show that A must be singular.6 6 Chapter 1... Another such factorization is VL U is an LU factorization. The factorizations used above U triangular. (c) Let S € Rnxn be skewsymmetric. AB ^ BA. y e Rn.e.. 2 0 IS I dempotent .5. Show that A must be singular. [~ ~ ].e...e. . what is det A? If A 3. A matrix A e Wx" is said to be idempotent if A2 = A. see. . see.yTx. for example. A =.. . Then E jRnxn skewsymmetric. Uk E Rnxn U = VI V2 .e A – 1 B is called the Schur complement of A in[ACBD].
+) is a group and an abelian group if (A4) also holds. The In this chapter we give a brief review of some of the basic concepts of vector spaces. (M2) there exists an element I E F such that a . p. Axioms (MI)(M4) state that (IF \ {0}. where some of the proofs that are not given here may be found. where some of the proofs that are not given here may for this and the next chapter is [10]. (M4) a·. there exists an element (a) e F such that a + (a) = 0. Axioms (M1)(M4) state that (F \ to).8)· yyf for all a.8 . (A4) a + . p e F.8.8 E IF.p ) .((.o r all a.8 a for all a. (M4) a • p =. . there exists an element aI E F such that a . the multiplication operator "•" is Generally speaking. ft Elf. •) is an abelian group.y for alia. (A4) a + p = . not written explicitly. y € F. : IF x F —> IF such that Definition 2.8.8. ^ 0. afar all a.8 + y) = a·. y Elf.1..ye¥.8. (M2) 1 e IF • I = for a e IF. • F x IF ~ F such that (Al) a (P y ) = (a + p ) y o r all a.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.1. (A3) for all a e F. +) is a group and an abelian group if (A4) also holds. including spaces formed by special classes of matrices.8) + y ffor all a. y e F. (A2) there exists an element 0 e IF such that a 0 = a. when no confusion can arise.) is an abelian group. . ft. 2. . p. Generally speaking. A field is a set IF together with two operations +. there exists an element (—a) E IF such that a (—a) O." is not written explicitly. (Ml) a· p .) ( a .1 Definitions and Examples Definition 2. but some infinitedimensional examples are also cited. (M3) e IF.. including spaces formed by special classes emphasis is on finitedimensional vector spaces. An excellent reference for this and the next chapter is [10]. Axioms (Al)(A3) state that (IF. when no confusion can arise. a f.8 = P • a for all a. I = a for all a E F. A field is a set F together with two operations +. (Ml) a . (A2) there exists an element 0 E F such that a + 0 = a for all a E F.8 = ft + afar all a. (D) (D) a· p a . yy) = (a·.((. y)=cip+a. y Elf. (Al) a + (.8 e F. The emphasis is on finitedimensional vector spaces. for all a e IF. (A3) for all a E IF. aI = 1. An excellent reference of matrices. be found. a"1 € IF • a~l = 1. but some infinitedimensional examples are also cited. the multiplication operator ".8 +a· y for all a. Axioms (A1)(A3) state that (F. 0. . (M3) for all a E ¥. 7 . .8 + y) = (a +.
(V3) (a + f3)..1 in the sense of operating on different objects in different sets. +) is an abelian group. R) with addition defined by I. Moreover. (R". since (M4) does not hold in general (although the other 8 axioms hold).5.. Definition 2. In practice. (V2) ( a f3) v = a P . w for all a e F and for all v. where Z+ = {0.3 are different from the + and . + f3qXq :aj. (Ml) does not hold unless m = n.r] = the field of rational functions in the indeterminate x = {ao + f30 + atX f3t X + . Remark 2. IR) with addition defined by and scalar multiplication defined by and scalar multiplication defined by is a vector space.2. 1. Vector Spaces Example 2. where Z+ = {O.. simply by V. . simply by V. when there is no possibility of confusion as to the A vector space is denoted by (V..4.1 in the sense of operating on different objects in different sets. Example 2. (MI) does not hold unless m = n. 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. Example 2. IR with ordinary addition and multiplication is a field. (VI) (V. Example 2. for example. (V5) I·• v = v for all v E V (1 e F). Moreover.l.8 Chapter 2. C with ordinary complex addition and multiplication is a field. is a field. Similar definitions hold for (en.1.• v = a·• v + p • v for all a. R"x" is not a field either for example. In practice.. 4. (V3) (a (V4) a(v w)=av a w for all a ElF andfor all v. Similar definitions hold for (C". underlying field. v for all a. ft) v = a v + f3.qEZ +} .}. IF) or. v + a. Ra[x] = the field of rational functions in the indeterminate x 3. .f3i EIR . w e V.2. in Definition Remark 2. f3 Elf andforall v E V. v = a . (IRn. Note that + and • in Definition 2.. 1.3. is a field..3. no confusion and the operator is usually not even written explicitly. 3. F) or. this causes no confusion and the·• operator is usually not even written explicitly. A vector space over a field F is a set V together with two operations Definition 2.. 2.2. A vector space is denoted by (V.5. e with ordinary complex addition and multiplication is a field. f3 e F and for all v E V. RMrmxn= {m x n matrices of rank r with real coefficients} is clearly not a field since.F xV »• V such that (VI) (V. w E V. e).( (f3' V v) o r all a. +) is an abelian group. p E IF andfor all v e V.3 are different from the + and • in Definition 2. 4. + apxP + . Note that + and· in Definition 2.. this causes 2. lR~xn is not a field either since (M4) does not hold in general (although the other 8 axioms hold). p € F and for all v e V.p ) . Raf.P. IR~ xn = m x n matrices of rank r with real coefficients) is clearly not a field since. (V2) (a·. when there is no possibility of confusion as to the underlying fie Id. . (V5) 1 v = v for all v e V (1 Elf). }.4. 2. R with ordinary addition and multiplication is a field. is a vector space. (V4) a· (v + w) = a . I. ) f for all a.2. C).
is henceforth understood to mean "is a subspace of. Then (W. and the functions are piecewise continuous =: (PC[to." The when used with vector spaces. F) is itself a vector space or. Then O(D. we write W ~ V. l . if and only if(aw1 ßW2) E if(awl + fJw2) e W for all a. that since 0 E IF. IF) be a vector space and let W c V. IF) is itself a vector space or." ya2n . F) if and only if (W. i. F) is a Definition 2. if and only subspace of (V. E) is a vector space with addition defined by 9 9 A+B= [ . and the symbol ~. this implies that the zero vector must be in any subspace. JR). td)n. (E mxn JR) is a vector space with addition defined by 2. etc. W = 0. w2 e Remark 2. 4. V) is a vector space with addition set of functions f mapping '0 to V. Notation: When the underlying field is understood.2. less restrictive meaning "is a subset of' is specifically flagged as such. (V. foral! a. that since 0 e F.7.7. i. IF) if and only if (W." + fJ2I a21 + P" . Notation: When the underlying field is understood. Let (V. JR). when used with vector spaces. t\])n continuous =: (C[to. +00). V) be the 3.2. Subspaces 2. y a l2 y a 22 yam 2 ya. . g E cf> and scalar multiplication defined by and scalar multiplication defined by (af)(d) = af(d) for all a E IF. IF) = (JRn.. Note. (JRmxn. F) be an arbitrary vector space and V be an arbitrary set.2. F) be a vector space and let W ~ V. t\]. too. The latter characterization of a subspace is often the easiest way to check Remark 2." The less restrictive meaning "is a subset of" is specifically flagged as such. 2.. verify that the set in question is closed under addition and scalar multiplication. 4. IF) be an arbitrary vector space and '0 be an arbitrary set.6. and the functions are piecewise continuous (a) '0 = [to. we write W c V. Then cf>('O. td)n or continuous =: (C[?0. Then (W. + fJmn l yaml yamn 3.. and the symbol c. h])n (b) '0 = [to. =: (PC[f0. is henceforth understood to mean "is a subspace of. IF) = (JR n . verify that the set in or prove that something is indeed a subspace (or vector space). Note.2 2. F) = (IR". for all d ED.6. too. 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. Then {x(t) : x(t) = Ax(t}} is a vector space (of dimension n). Let (V. Let (V. E). W f= 0. fJ e IF andforall WI.2 Subspaces Subspaces Definition 2. V) be the set of functions / mapping D to V. Then (x(t) : x ( t ) = Ax(t)} is a vector space (of dimension n). implies that the zero vector must be in any subspace. Let O(X>.. Subspaces 2. W2 E W.e. Let cf>('O.e. Let (V. td. ß E ¥ and for all w1. Let A E JR(nxn. IF) is a subspace of (V. equivalently. The latter characterization of a subspace is often the easiest way to check or prove that something is indeed a subspace (or vector space). Let A € R"x". and for all f E cf>. this question is closed under addition and scalar multiplication. Special Cases: Special Cases: (a) V = [to. (V. 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 = . equivalently. (V.
9.3 2. + (XkVk = 0 implies al = 0. al VI + . ffR and S are vector spaces (or subspaces).•. E ]Rnxn : not 2. vk E X and scalars a1. = {A E JR. ft E R symmetric for all a. Then (V. 2. ak not all zero such that elements VI. . f3 e R. .e. Then it is easily shown that aAI + fiAi is symmetric for all a. Consider (V. W~V.e. in some vector space V.• } be a nonempty collection of vectors Vi in some vector space V.nxn. ~SandS ~ R. .lF) = (R" X ". To prove two vector spaces are equal./l with f3 = 0 are All lines through the origin are subspaces. Wi. W1/2. c E JR. For a. that the vertical line through the origin (i..R) and 1. . define W". too.10. Vk of X and for any scalars aI.} .ß is a subspace of V if and only if ß = O. sketch Then Wa. Note. . IF) = (]R2. then R = S if and only if R C S and S C R.9. . V usually denotes a vector space with the underlying field generally being R unless Thus. For ß E R define the jccoordinate in the plane and V2 with the ycoordinate. •...8. . 1.. W2. • • •} be a nonempty collection of vectors u. ak. Let X {VI.Vk of X and for any scalars a1.... .. As an interesting exercise.JR. elements VI.3 Linear Independence Linear Independence Let X = {v1. Shifted subspaces Wa. V2. . Then W".. A2 are symmetric.. Then it is easily shown that ctA\ + f3A2 is Proof' Suppose AI.I' and Wi. Vector Spaces Example 2.1. Shifted subspaces W".ß with ß =1= 0 are called linear varieties. X linearly set of Definition 2. . Then W is /wf a subspace of JR. Definition 2.10 Chapter 2.) and for each v E ]R2 of the form v = [~~ ] identify VI with the xcoordinate in the plane and u2 with the ycoordinate. Henceforth. 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. Proof: Suppose A\.0.1../l is a subspace of V if and only if f3 = 0. Let W = {A € R"x" : A is orthogonal}. R) and for each v € R2 of the form v = [v1v2 ] identify v1 with 3. .. . then R = S if and only if Definition 2. Note. V usually denotes a vector space with the underlying field generally being JR. Henceforth. sketch W2.S. . one usually proves the two inclusions separately: Note: To prove two vector spaces are equal. F) = (JR.nxn : A We V.nxn. v2. •••. . we drop the explicit dependence of a vector space on an underlying field. . ak. Example 2.. R"x". Definition 2. W2. explicitly stated otherwise. be an element of R. X is a linearly dependent set of vectors ifand only if there exist k distinct if and only if exist distinct elements v1. ak = O. 3. . too.O.o.I. As an interesting exercise. and S are vector spaces (or subspaces). Consider (V. unless explicitly stated otherwise. a = 00) is also a subspace. F) = (R2.10. Consider (V...o. All lines through the origin are subspaces.) and let W = [A e R"x" : A is symmetric}. called linear varieties./l = {V : v = [ ac ~ f3 ] . that the vertical line through the origin (i. JR. we drop the explicit dependence of a vector space on an underlying field. A2 are symmetric.and W1/2. Thus. If 12. W2. 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 X is a linearly independent set of vectors if and only if for any collection of k distinct elements v1. . a = oo) is also a subspace.. f3 e R. .
}. Then {[ Then I. (Xi ElF. . and there exists a E ]Rk such that VT V is singular. .Vk] e ]Rnxk. } = (Xl VI + .2. .3.. en} = Rn. + (XkVk .13.v2 + V3 = 0).'" .}. An equivalent condition for linear independence is that the matrix Va = 0. A e R xn B E ]Rnxm.. e k.12. linear dependence x such that Va = 0. 2. Definition 2. The linear v E ]Rn. . 1. ii E If. o Definition 2. 1£t V = R3. V2. .. . Vi e span of Definition 2. Independence of these vectors turns out to be equivalent to a concept Chapter 11). called consider Let Vif e R".. e2 . to be studied further in what follows.11. and there exists a e R* such that Va = 0. 2.. . X is a linearly independent set (of basis vectors). If the set of vectors is independent. A set of vectors X is a basis for V if and only if 1. tIl 2..3.en = 0 0 0 o SpIel.•}} be a collection of vectors vi.. = [ v 1 . An equivalent condition for linear independence is that the matrix V TV is nonsingular. Linear Independence Example 2. Independence of these vectors turns out to be equivalent to a concept called controllability..... . e2 = 0 1 0 . e2.. 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."I [ i1i1l ]} [[ s a linearly is a Iin=ly dependent set de~ndent ~t (since 2v\ — V2 + v3 = 0). .. Then the span of of X is defined as X is defined as Sp(X) = Sp{VI. Let X = {VI. kEN}. Let A E ]Rnxn and 5 e R"xm.. Then consider the rows of etA B as vectors in em [to. Vi EX.14. An equivalent condition for linear dependence is that the k x k matrix condition VT V is singular.. LetV = 11 11 ~. and consider the matrix V = [VI. E V.. t1] (recall that etA denotes the matrix exponential. Why? independent. {1. Vk] E Rnxk. Howe. If the set of vectors is independent.en} = ]Rn.12. and X (of and 2.. = {v : where N = {I.11. then a = 0. 2. X = [v1 v2 . ~ HHi] } Ime~ly i is a i" linearly independent set. (since 2vI . to be studied further in what follows.. V V2. Sp(X) = V. Let V = Rn and define = ]Rn and el = 0 0 . ... Then consider the rows of etA B as vectors in Cm [t0.. . Sp(X) = V..14.. Example 2. Why? However. Then Sp{e1. Linear Independence 2. . Example 2. T V is nonsingular. which is discussed in more detail in efA Chapter 11). then = O. A set of vectors X is a basis for V if and only ij Definition 2.13..
For example. ..15. . n } such that for V.19. + ~nbn = Bx. . Vector Spaces Example 2.. . en} is a basis for IR" (sometimes called the natural basis).b.[ ~  ] + 4· [ ~ l To see this.16.. We represents B. en} natural Now let b l .. In Rn. n for Then for all e there exists a unique ntuple {E1 .. bn]} and are unique. VI ] : = vlel + V2e2 + . For example.. .12 12 Chapter 2. . e2. We say that the vector x of of of (b1..dimensional or have dimension n and we write dim (V) = n or dim V — n. + vne n · Vn We can also determine components of v with respect to another basis. write [ ] = XI • [ ~ + ] X2 • [ _! ] =[ ~ = [ ~ Then Then ! ][ ~~ l 1 [ ~~ ] = [ . B ~ [b". Definition 2.E~n} such that v= where ~Ibl + . . For . The number of elements in a basis of a vector space is independent of the Theorem 2. particular basis considered. {el... .. The number of elements in a basis of a vector space is independent of the particular basis considered. while We can also determine components of v with respect to another basis.. .. V is said to X for be ndimensional or have dimension n and we write dim(V) n or dim V n.. {~i } of v with respect to the basis {b l . If V= 0) V is Definition 2. For be n.. The scalars {Ei}are called the components (or sometimes the coordinates) components coordinates) Definition 2.... If a basis X for a vector space V(Jf 0) has n elements. with respect to the basis with respect to the basis {[~l[!J} we have we have [ ~ ~ ] = 3.18. x ~ D J Definition 2.. r I [ . Example 2.18. components represents the vector v with respect to the basis B.. Then for all v E V there exists a unique ntuple {~I'. . In]Rn. .16....l. el + 2 .. n unique. bn be a basis (with a specific order associated with the basis vectors) b1.19. .. ] l = = Theorem 2. while [ ~ ] = I .17. for]Rn [e\.
ft n 5 = {v : v E 7^ and v E 5}. . i e m.21. dim{A E ~nxn :: A = AT} = !n(n + 1). The sum and intersection ofR and S are defined respectively by: of R. U\ + 1. Thus.4 2.) = 0 and ]P ft. K + S S. Let (V. Theorem 2. s E 5}. Thus. The collection of Eij matrices can be called the "natural basis matrices. otherwise.24. Let (V. vector space V is finitedimensional if there exists a basis X with n < +00 elements. Example 2. 5. for finite k). n S {r s : r E U. Remark 2.+00.21.j matrices can be called the "natural basis matrices. otherwise. « The subspaces R. V (in general. n (^ ft. and S are said to be complements of each other in T. determine !n(n 1) symmetric basis matrices. we define dim(O) = O. U S. 72. Ra C V/or an arbitrary index set A). V (in general. R = 0. V is infinitedimensional. 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. Definition 2.18 says that dim(V) = the number of elements in a basis. is not necessarily a subspace." 3. R C S. n n S = 0. n 5 S.. dim{A € Rnxn A AT} = {1/2(n 1 (To see why. dim{A E ~nxn :: A is upper (lower) triangular} = 1/2n(n+ 1). j e n. Sums and Intersections of Subspaces 13 13 consistency. A vector space V is finitedimensional if there exists a basis X with n < +00 elements. The union of two subspaces. The subspaces Rand S are said to be complements of each other in T. U + S = T (in general ft. Note: Check that a basis for ~mxn is given by the mn matrices Eij. j)th location. 1. RI \ h Rk =: ]T ft/ C V. where Eij is a matrix all of whose elements are 0 except for a 1 in the (i. The union of two subspaces. V for an arbitrary index set A). V. H. dim(~mXn) = mn. 2. The sum and intersection Definition 2.23. dim(Rn)=n. 2. determine 1/2n(n + 1) symmetric basis matrices. Remark 2. and 2. R S C V (in general. A consistency. J)th location. 4.4 Sums and Intersections of Subspaces Subspaces Definition 2. .. t1]) .= T). i E m.22. + 7^ =: L R.18 says that dim (V) the number of elements in a basis. R. S. dim(R mxn ) mn.20. where Efj is a matrix all of whose elements are 0 except for a 1 in the (i. dim{A e Rnxn A is upper (lower) triangular} = !n(n 1). F) be a vector space and let R. 2. dim(~n) = n.2. y>f (L L . Example 2. S S. Note: Check that a basis for Rmxn is given by the mn matrices Eij.4. for finite k). and because the 0 vector is in any vector space. 2. 1.24. R j ) = 0 am/ Ri = T). dim(C[to.a S. and 1. tJJ) = +00. and S are defined respectively by: 1. a eA CiEA f] n *R. R + S = {r + s : r e R. we define dim(O) = 0.=1 K k 1=1 2.20. and because the 0 vector is in any vector space. 1. S c V. 2. is not necessarily a subspace.4. V is infinitedimensional." The collection of E. V. s e S}.23. 1. R D S C V (in general. 2. R H S = {v : v e R and v e S}.22. Theorem 2. Theorem 2. Theorem 2..) 2 5. R S = (in general. JF') be a vector space and let 71. Sums and Intersections of Subspaces 2. j E ~.) (To see why.
. suppose an arbitrary vector t E T can be written in two ways t e as t S2.. mutually [x\.r2 £ Rand 52 . Xk} must be a linearly independent set. Avn are also orjRn. 2.c the Example 2. and let S Let (V... which uniqueness follows. jRn xn . .14 14 Chapter 2. where rl. Then any other distinct line through the origin is and let R be any line through the origin. .27. r2 e R. every t E T can be written uniquely in the form tt = r + s with r E Rand s E S.. Suppose {VI. For arbitrary subspaces ft.. 0 S. and SI. we must have r\ ri and SI rl . {vi. triangular + L = R xn un. . Theorem 2. F) (R n x n . vd must be a linear combination of the others..5.27. . 1 TIT The first matrix on the righthand side above is in S while the second is in R.. jR). ft. validity of the formula given in Theorem 2. v = [4 l]r jR2.r . XI. every t € can be written uniquely in the form r s with r e R and s e S. 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 Rnxn. let R be the set of skewsymmetric matrices in (V. where r1.. Then V = U $ S.27. S of a vector space V. EXERCISES EXERCISES 1. 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 .29.. R). Show that Av\. Let x\. ..25. Example 2. one can easily verify the validity = n. Then Theorem 2. . Vector Spaces Chapter 2. 2. We discuss more about orthogonal complements elsewhere in the text.. Find the components of the vector v = [4 If with respect to this basis.2 and 2. 3. Vector Spaces Remark 2. Using the fact that dim {diagonal (diagonal matrices} = n.. Then it may be checked that U + . Vk} is a linearly dependent set. jRnxn. . .c = jRnnxn jRn xn. Then any other distinct line through the origin is a complement of R. Then r1 — r2 = s2— SI.26.. of the formula given in Theorem 2.s\. 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). But as t = r1 + s1 = r2 + S2. IF) = (jRnxn.28. S2 e rl Sl r2 Then r. For example.. = dim(ft) + Proof: To Proof: To prove the first part. Let VI. Example 2. we must have rl = r2 and s\ = si from S2 from which uniqueness follows. and let R"x"..20.29. Consider the vectors VI — [2 1f and V2 = [3 1f.. unique ft. D Theorem 2. ft S) = jR2 and let ft be any line through the origin.c jRnxn. consider V = R2 unique. while U n £ is the set of diagonal matrices in Rnxn. Xk} must be a linearly independent set. Vn thonormal if and only if A E R"x" is orthogonal. dim(T) = dim(R) + dim(S). . . dim(R + S) = dim(R) + dim(S)  dim(R n S)... S of a vector space V.26.. r2 E Rand s1.. ft be the set of symmetric matrices in R" x ". 2. . Suppose =R EB Then 1.. X2.20. Since ft fl 0. For arbitrary subspaces R.. e jRnxn 4. *2. ft. Suppose T = R O S. Xk E jRn 2.si e S. 0 The statement of the second part is a special case of the next theorem. . Among all the complements there is a unique one orthogonal to R. Show that {XI. . x/c E R" be nonzero mutually orthogonal vectors..27.. . AVn are orv\. Av" •. But r1 –r2 E ft and S2 — SI E S. the set in jRnxn. together with Examples 2. Then show that one of the vectors 1.vn be orthonormal vectors in R".r2 S2 . The complement of R (or S) is not unique.25. Since R n S = 0.. S2 E S. Theorem 2.
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
i E n}. {u. In other words. Thus. Thus.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.} are the usual (natural) bases. + amiWm where W = [w\. When V = R".2 3. then LVx = Lv = ~ILvI + . j E !!!.m and {Vi. Thus... n A= al : ] E JR. say. w ] and L is the ith column of A. if V = E1v1 + . if v = ~I VI + • • + ~n v = Vx (where v. j E m} are the usual (natural) bases WA linea LV L = A..m usually causes no naturally confusion. j E raj. + ~nLvn =~IWal+"'+~nWan = WAx.. Thinking of both as a matrix and as a linear transformation from JR. and hence x.e. i e n} and {Wj. in the notation. W = R m and [ v i .. but this is usually not done. 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. j e m}. and hence jc. is by its action on a basis. i e ~}. Li near Transformations Chapters. We thus commonly identify A as a linear transformation with its matrix representation. Then the {w j. is arbitrary).. with respect to {w }•. Linear Transformations 3. + .18 Chapter 3. respectively.mxn a mn represents L since represents £ since LVi = aliwl =Wai.. w m] and where W = [WI." to Rm usually causes no Thinking of A both as a matrix and as a linear transformation from Rn to lR. Note that A = Mat £ depends on the particular bases for V and W. transformation with its matrix representation. z'th V This could be reflected by subscripts. then arbitrary). L IF) ~ (W.n. LV WA since x was arbitrary. When V = JR. i. .e.. {W jj' j e !!!.. i.• + E nVnn = V x (where u. Change of basis then corresponds naturally to appropriate matrix multiplication.} are bases for V and W. We identify A the equation £V = W A becomes simply £ = A. F) is linear and further suppose that {Vi. suppose £ : (V.. for V and W) is the representation of £i>.. In other words. F) —>• (W.. 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. Specifically.. [ w . Thus. . £V = W A since x was arbitrary... W = lR.. IF) veniently in matrix form.
The above is sometimes expressed componentwise by the C — A B . . 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).3 Composition of Transformations Composition Consider three vector spaces U. Note that in most texts. and dim W m. If dimZ// = p. y E Rn.=1 Outer Product: Let x e Rm. xx T XX ). dimV = n. Then we can define a new transformation C as follows: C The above diagram illustrates the composition of transformations C = AB. That is. then composition of transformations corresponds to standard matrix mUltiplication. Note that in The above diagram illustrates the composition of transformations C = AB. y e Rn. Composition ofTransformations 3.3. Then their outer product is the m x n E ~m. Composition of Transformations 19 19 3. Inner Product: n xTy = Lx. 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.3. That is. expressed mxp nxp formula cij = L k=1 n aikbkj. and if we associate matrices with the transformations in the usual way.3. the arrows above are reversed as follows: C However. and W and transformations B from U to V and A from Wand V to W. V. the form XXT (or xx HH). If dimU = p. we have C A B . in the same order in both the diagram and the equation. dim V = n. Two Special Cases: Two Special Inner Product: Let x. Then we can define a new transformation C as follows: to W.y.. and if we associate matrices with the transformations in the usual way. Then their inner product is the scalar E ~n. y e ~n. A rankone symmetric matrix can be written in above (or xy if A E C xyH e c ). and dim W = m. . then composition of transformations corresponds to standard matrix multiplication.
Note that N(A) and R(A) are. (A). The nullspace of The of denoted N(A). Definition 3. Vk} with u. in general. 3.4 Structure of Linear Transformations Structure of Linear Transformations Let A : V —> W be a linear transformation. of of denoted Im(A). [: J} is an orthogonal set.•. LinearTransformations Chapter 3.. an} . if i f= j. 1. . Theorem 3. {[ ~~i ]. Note that in Theorem 3.. Then 1. 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. N(A) S. See also the last paragraph of Section 3. R(A) C W. subspaces of different spaces. denotedlZ( A). .. Let A : V + W be a linear transformation.2. If A is written in terms of its columns as A = [ai. .. is the set {v E V : Av = O}.. in general.3.. V. 2.. . R(A) = {Av : v E V}.. . ~. ~ 3 .. . an M. {[ ~ J.. Equivalently. . then then R(A) = Sp{al. is the set {w e w = Av for some v e V}. {v1. . is the {v e V Av = 0}. . vi ^/v'k vk ~~~ ] . Li near Transformations 3. usual (natural) bases.. Theorem 3. where 8ij is the Kronecker delta defined by Kronecker delta defined by 8 = {I0 ij ifi=j.5 and throughout the text. See also the of Section 3. Let A : V + be a linear transformation. If in of = [a\..8.7. . e Rn.. If {VI. essentially following immediately from the definition.[ :~~ J} . 2. then ~ .. The range of A.7. I ~VI VI ^/v. { t > . Definition 3. the same symbol (A) is used to denote both a linear transformation and its matrix representation with respect to the used to denote both a linear transformation and its matrix representation with respect to the usual (natural) bases. an].20 20 Chapter 3. Definition3.. denoted Im(A). Note that N(A) and R(A) are.5.8. Then Let A : V —>• be a linear transformation.4 3. IS an orthogonaI set. is an orthonormal set.i •..2. . is an orthonormal set. The range of A. e ~mxn.. vk] be a set of nonzero vectors Vi E ~n. essentially following immediately from the defiProof: The proof of this theorem is easy. The range of A is also known as the image of A and — {Av e V}. Let A E Rmxn.. Let {VI.an]. ~}  ISisan 3. W. N(A) c V. is the set {w E W : w = Av for some v E V}." • orthonormal set. subspaces of different spaces. then Proof: The proof of this theorem is easy.. 2. is an orthogonal set. denotedR(A).IN. 1. then {I —/==. The nullspace of kernel of and A is also known as the kernel of A and denoted Ker (A). € 1Tlln is an orthogonal set. orthonormal set.6. vd of u.3. Example 3. denoted N(A). —/=== . 0 nition.4.Vk } With Vi E.. The nullspace of A. the same symbol (A) is Note that in Theorem and throughout the text. D Remark 3. The set is said to 3. R ( A ) S.
n S~.. Set XI X2 = L (xT Vi)Vi. . = S. Then the of defined T S~={VE]Rn: V S = 0 for all s e S}. Proof: We prove and discuss only item 2 here. ]Rn. if and only if S~ <. Then it can be shown that Working from the definition.9. .. Structure of Linear Transformations 21 21 Definition 3.10. Rn..= {v e Rn : vTs=OforallsES}. Then the orthogonal complement of S is defined as the set c ]Rn. (S~)l. S \B S~ = ]Rn.3. 6. Let S <. S 5.10. 4. vk} e ]Rn vector. Theorem 311 Let Theorem 3. (n + S)~ = nl.4. 2. The proofs of the other results are left as Proof: left exercises.e. Any set of vectors will do..=1 XI. including dependent spanning vectors (which would. Note that there is nothing special about the two vectors in the basis defining S being orthogonal. Let R S C Rn The S <. then give rise to redundant equations). . nonzero) solutions of the system of equations 3xI 4xI + 5X2 + 7X3 = 0. S1. (n n S)~ = n~ + S~. n <. Structure of Li near Transformations 3. Set vector. k =X . the computation involved is simply to find all nontrivial (i. .. Then n. Example 3. Let {VI. n~. Vk} be an orthonormal basis for S and let x E Rn be an arbitrary {v1. of course. 3.11. Let 3. + X2 + X3 = 0..4.
D Theorem 3. y. we form AT v.XI/ (x~ . every vector w in the codomain space IRm can be written in a unique way as w = x+y. standing Figure 3. But yT Ax = (ATyy{ x. .l = 0 since the only vector s E S orthogonal to S1 = IRn. 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. Let A : Rn + Rm. we see that x2 is orthogonal to v1.12 and part 2 of Theorem 3. X2 is orthogonal to any vector in S. x e R(AT). Then {v E R" : Av = 0} is sometimes called the right nullspace of A. The proof of the second part is similar and is left as an exercise. Thus.5 Four Fundamental Subspaces Four Fundamental Subspaces Consider a general matrix A E lR. x... In other words.. It can write vectors in a unique way with respect to the corresponding subspaces.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. But then (x. since T x 2 Vj = XTVj ..e. x~ X2 = (x. (Note: This for finitedimensional 1.e. where XI.) Proof: To x E N(A). R(A).l where x € M(A) and y € J\f(A)± = R(AT) (i. and x2 = x~. We S n S1 =0 the e orthogonal everything in (i.5 3. that x = XI for example. including itself) is 0. We also have that S U S.e.. Then Ax = 0 and this is an and equivalent to yT Ax = 0 for all v. Let A : IRn > Rm. Thus. (Note: This holds only for finitedimensional vector spaces. Clearly. Vk and hence to any linear combination of these vectors. Then Theorem 3.e. everything in S (i.13. since Then XI e S and.x1) (x'1 xd x2 — X2 = — (x'1 — x1) (which follows by rearranging the equation x1+x2 = x'1 + x'2). See also Theorem 2. many properties of A can be developed in terms of the four fundamental subspaces . N(A).12 and part 2 of Theorem 3. Then {v e IRn : A v = O} is sometimes called the Definition 3.XITVj =XTVjXTVj=O.13. Then 1. including itself) is O. But yT Ax = ( A T ) x. E N(A) and E N(A). . Let A : IRn —> Rm. Theorem 3. where x e U(A) and y e ft(A)1. R" N(A) 0 ft(Ar ». 'R. for example.l. can write vectors in a unique way with respect to the corresponding subspaces. Thus. take an arbitrary x e A/"(A). X2 is orthogonal to any vector in S. (w e Rm : w T A = 0} is called the left nullspace of A. Vk and hence to any linear combination of these we see that X2 is orthogonal to VI. right nullspace of A.1 in the next section. ft(Ar) (i. When thought of as a linear transformation from IR n to Rm. Suppose. established that N(A) U(AT ).l = Rn. Theorem 3. Linear Transformations Then x\ E <S and. the right nullspace is N(A) while the left nullspace is J\f(AT). Similarly. every vector in the codomain space R m can be written ina unique way asw = x+y.. = x'1+ x'2. . i. when we have such direct sum decompositions. x 1 E Sand x2.e.l = 7£(AT).12. IRm = R(A) EBN(A T». Suppose. Then X2 = x. +x~)..12.14 (Decomposition Theorem). But then (x'1 —XI)TT (x. Similarly. every vector v in the domain space R" can be written in a unique way as v = x + y... + x~. (Note: This also holds for infinitedimensional vector spaces. In other words.l = N(A ). Ax = Proof: To prove the first part.e..l = R(A T}. Ax = 0 if and only if x orthogonal is orthogonal to all vectors of the form AT y. every vector v in the domain space IRn can be written in a unique way as v = x 7. It is also easy to see directly that. Let A : R" + IRm.1 in the next section. Rm = 7l(A) 0 M(AT)). 3. 0 The proof of the second part is similar and is left as an exercise. (Note: This also holds for infinitedimensional vector spaces. When thought of as a linear transformation from E" Consider a general matrix A € E^ x ". E S and X2.26.26. Let A : IRn > IRm.l N(A T (i. Ax = 0 if and only if x equivalent to yT Ax = 0 for all y. Let A : Rn + IRm. We have thus shown that vectors. 0 x1 — x'1 andx2 = x2.xn.) 2. that x = x1 + x2. Then Theorem 3. x'2 E S1. x E R(A r ) Since x 1 have established thatN(A). Since x was arbitrary. transformation A. See also Theorem 2.l. We have thus shown that S + S.. we decompositions.22 22 Chapter 3. Thus. where x\. Then R(A r ). E R(A) and E R(A). x~ e S.•. IRn = M(A) EB R(A T)).e. Li near Transformations Chapters. XI) (which follows by rearranging the equation XI +X2 = x. D Definition 3.= Af(AT) ) (i.14 (Decomposition Theorem). 2. XI = x.l. . .) 2.11 can be combined to give two very funTheorem 3. Clearly.) N(A)1" spaces..e. i. the right nullspace is A/"(A) while the left nullspace is N(A T ).X2) 0 since (x'1 — x1) (x' 2 — x2) = 0 by definition of ST. {w E IR m : WT A = O} is called the left nullspace of A..(A)1~ — J\f(ATT ). many properties of A can be developed in terms of the four fundamental subspaces to IRm. – x1) = 0 since 0 by definition of S. Then T (x.
16.(A) = W. 1. Figure 3. R(A). the column rank of A (maximum number of independent columns).1. Four fundamental subspaces.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. properties 7£(A). A is onto (also called epic or surjective) ifR. and N(A)1. N(A). fundamental subspaces.1 makes many key properties seem almost N(A)T. Definition 3. be a linear transforDefinition 3. Figure 3. IR n > IRm. A f ( A ) .1. Four Fundamental Subspaces 23 23 A r N(A)1 r EB {OJ X {O}Gl nr m r Figure 3. mation. A is onetoone or 11 (also called monic or infective) ifJ\f(A) = 0. Four Fundamental Subspaces 3.16. 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 . Two equivalent 2. rank(A) dimftCA).15.(A)^. Let and W be vector spaces and let A : motion. R(A)1. A is onto (also called epic or surjective) ifR(A) = W. A is onetoone or 11 (also called monic or injective) if N(A) = O. Let A : E" + Rm. 1. Then rank(A) = dim R(A). t= V2 ===} AVI t= AV2 .5.3.15. This is sometimes called 3. 2. The row rank of A is column rank of of independent row rank of . 3. Let V and W be vector spaces and let A : V + W be a linear transforDefinition 3. and in illustrating concepts such as controllability and observability. 'R.
{Tv\. then {TVI. The basic results are contained in the following easily proved following theorem. Tvrr]} is a basis for R(A). . . we include here a few miscellaneous results about ranks of sums completeness. To see that T is also onto. of Corollary 3. Part 4 of Theorem 3.. rank(B)}.. by definition there is a vector x E ]Rn such that Ax = w. Let A : R" ~ Rm. . .19. Theorem 3. LinearTransformations Chapter3. shows that T is onto. following follows we apply this and several previous results. Proof: From Theorems 3. Then N(T) = To w E 7£(A).19 suggests looking atthe general problem of the four fundamental subspaces of matrix products.. dimension of the domain of A. .(A) = dimA/^A^ 1 if that if {VI. Theorem 3. We thus have that dim R(A) = dimN(A)L since it is easily shown T dim7?. Let A. (Note: Since 3. nullity(B) :s nullity(AB) :s nullity(A) 4. . if B is nonsingular. rank(A) + B) :s rank(A) + rank(B).17. + rank(B)  n :s rank(AB) :s min{rank(A). Then dim K(A) = dimNCA)L... The dual notion to rank is the nullity R(AT) of independent rows). The last equality AXI x\ e N(A)L and jc E N(A). . Like the theorem. R(A) : ]Rn ~ ]Rm. + nullity(B). . of A.17. Then 3. it is a statement about equality of dimensions.11 and 3.17 we see immediately that n = dimN(A) = dimN(A) + dimN(A)L + dim R(A) . . 1 1 Xl E A/^A) . dimA/'(A) ± (Note: 1 T T ). where Ax — w.. Tv abasis 7?. colloquially of = rank of A. 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. .18.19.") 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. dimensions. v r } is a basis for N(A)L.18. Clearly T is 11 (since A/"(T) = 0). where n is the ]Rn > ]Rm. rank(AB) = rank(BA) = rank(A) and N(BA) = N(A). and is defined as dimN(A). 3.. Linear Transformations dim 7£(A r ) (maximum number of independent rows). 3. take any W e R(A). O:s rank(A 2. dimA/"(A) + dimft(A) = dimension of the domain of A. 0 For completeness. Let A : Rn > Rm." 0 of D The following corollary is immediate. Write x = Xl + X2. . Then dimN(A) + dim R(A) = n. iv} abasis forA/'CA) . B E R" xn . r*i *i E N(A)L.") of A.andx22 e A/"(A). this theorem is sometimes colloquially stated "row rank of A = column N(A)L = R(A A/^A) " = 7l(A ). sometimes denoted nullity(A) or corank(A). the subspaces themselves are not necessarily in the same vector space.19 suggests looking at the general problem of the four fundamental Part 4 of Theorem 3.24 24 Chapter 3. x x e R" x\ X2. and products of matrices. and is defined as dim A/"(A). e ]Rnxn. of A. denoted nullity(A) or corank(A).17 we see immediately that Proof: From Theorems 3.. the subspaces themselves are not necessarily in the same vector space.(A). Then Ajti = W = TXI since Xl e A/^A). 1. Finally.11 and 3. if {ui. u.
dim R(A) = m = rank(A). AT A is nonsingular).5.2 N(B). It The next theorem is closely related to Theorem 3. Let A E Rmxn. A is 11 if and only ifrank(A) = n (A has linearly independent columns or is said 2. Proof' Proof of part 1: If A is onto. Note that if A is invertible. AA is nonsingular).3.22. dim7?. and hence dim R(A) n by Theorem 3. R(A) = R(AA T ). y E R(A). Also.ti = AT Ax2. N«AB)T) . Theorem 3. A Proof of part 2: If A is 11. have full row rank. equivalently.21.(A) — m — rank (A). i. which implies that dimN(A)11 = n — Proof of part 2: If A is 11. the transformations A. linear least squares problems. R(AT) 3.(A). N(AB) . Let A : IRn + IRm. suppose Ax\ dim R(A T). Conversely. = R(A T A). e IRmxn. Conversely. especially when dealing with pseudoinverses and linear least squares problems. It is extremely useful in text that follows.17. .e. suppose AXI = Ax^. Let A E Rmxn. which implies that dim A/^A).and R(A). Then 3. AATT is nonsingular). A is AT AXI AT AX2. A : V + W is invertible (or bijective) if and only if it is 11 and onto. Conversely. Thus. e 7?. 2. Theorem 3. equivalently. 4. Then A r A.17. to have full column rank. A is 11 if and only z/rank(A) = n (A has linearly independent columns or is said to have full column rank. The transformations AT are all 11 and onto between the two spaces N(A)1. 3. then dim V = dim W. since ArA is invertible. XI = X2 AT A A 11. and AI A.5.23. then A/"(A) = 0. x AT (AAT)I e IRn. Also. A € IR~xn. Ar.20. Let e IRmxn. 2.20. A"1 ± are all 11 and onto between the two spaces M(A) and 7£(A). A is onto if and only if rank (A) = m (A has linearly independent rows or is said to have full row rank. D D 11. 1. Then y = Ax. 1. Then 3.—n = dim 7£(A r ). Then Theorem 3. e IRnxp.21. then N(A) = 0. R«AB)T) S. AT A nonsingular). B E Rnxp. A A T. A : IRn »• IR n is invertible or Note that if A is invertible. Note that in the special case when A E R"x".22.23. N(A) = N(A T A). terms of rank and invertibility. 4. A : W1 + E" is invertible or nonsingular if and only z/rank(A) = n. Conversely. let y E IRm be arbitrary. Definition 3.2 N(A T ). Four Fundamental Subspaces 3. equivalently. and hence dim 7£(A) = n by Theorem 3. RCA). Definition 3. A is onto if and only //"rank(A) — m (A has linearly independent rows or is said to 1. so A is onto. Similar remarks apply to A and A~T.20 and is also easily proved. nonsingular ifand only ifrank(A) = n. which implies x\ = x^. 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.20 and is also easily proved. let y e Rm Proof: Proof of part 1: If A is onto. especially when dealing with pseudoinverses and is extremely useful in text that follows. equivalently. : R n » Rm. AT. 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. N(A T ) = N(AA T ).. RCAB) S. R(B T ). 2. A : V —» W is invertible (or bijective) if and only if it is 11 and onto. then dim V — dim W. Let jc = AT(AAT)~]y Y E Rn. Four Fundamental Subspaces Theorem 3. 25 25 The next theorem is closely related to Theorem 3. 1. AX2.
If there exists a unique left inverse A~L such that A~LA = I. if A is 11.24. linear Transformations If a linear transformation is not invertible.25. Notice the and leave the following: following: A(A. Let A : V + W. Obviously A has full row rank can always find v E ]R2 such that [1 2][ ~~] = a). A R A = I. Definition 3.26 Chapter 3.22 we see that if A : E" + Em is onto. by uniqueness it must be A R + A R A — = A R. —> transformation if left + 2. such that A~LA = Iv where Iv denotes the identity transformation on V.R = If left A L A L A = 2.R AA..I) must be a right inverse and.. A~ (A A)~ A . where Iv denotes the identity transfonnation on V. Then > 1.27. D Example 3. A is said to be right invertible if there exists a right inverse transformation A~RR : if A.. in A I = A R = A L. (A R + A RA .I = A~R.e.R + AARA = I A +IA  A since AA R = I = I. are infinitely A. A is left invertible if and only ifit is 11. i. that A~R is a left inverse. Similarly.. A. 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. 0 a left inverse. it may still be right or left invertible.e. then one (Proo!' = [1 2]:]R2 + ]R . A is invertible if and only if it is both right and left invertible.25 that A is invertible. 1. But this implies that A~RA = /. € ]R . then a right inverse is given by A~R = AT(AAT) I. Then Definition 3.24.26. Theorem 3. It then follows from Theorem 3. 1. it is clear that there are infinitely many right inverse. therefore. in which case A~l = A~R = A~L. Let + V.22 ]Rn >• ]Rm Note: From Theorem 3. i. It then follows from Theorem 3.. Also. both 11 and Moreover. (Proof: Take any a E E1I. characterizing all solutions of the linear matrix equation AR = I. i. by uniqueness it must be Thus.. 2. (A R + A R A — /) must be a right inverse and.L = (ATTA)I1AT. A R the case that A~R + A~RA .R = _~] (=1) and A~R = [ _j j is a right inverse. Let Theorem 3. A right invertible if and only if it onto. is left invertible if and if it and left invertible.. then a left inverse is given by A R = AT (AAT) left T L = A. where Iw denotes the identity transfonnation on W.both 11 and is if and if onto. Li near Transformations Chapters. . Let A : V » V. A is right invertible if and only if it is onto.e. Let A : V + Then 1.26. Obviously A has full row rank (= 1) and A . Then A is onto. If there exists a unique right inverse A~R such that AA~R = I. right inverses for A. can always find v e E2 such that [1 2][^] = a). therefore. In Chapter 6 we characterize all right inverses of a matrix by Chapter characterize characterizing all solutions of the linear matrix equation A R = I. then A is invertible.25 that A is invertible. Let A = [1 2] : E2 »• E1I.e. Let A : V > W. 3. Defileft If linear concepts left nitions of these concepts are followed by a theorem characterizing left and right invertible transformations. then A is invertible.: AA R = w Iw W + V such that AA~R = Iw. 1. i. Theorem 3. If Proof: proof of second Proof: We prove the first part and leave the proof of the second to the reader.R + ARA I) = AA.
with Y e ]Rnxn (X. Y E Enx" define their inner product by (X. 3. For matrices X. In Chapter 6 we characterize all left inverses of a infinitely many left inverses for A. respect to this inner product. Let A = [~ . Prove Theorem 3. it is clear that there are A L = [3 — 1] infinitely many left inverses for A. (Proof: The only solution to 0 = Av = [I2]v is v = 0. Is £. Show that. The matrix 3. In Chapter 6 we characterize all left inverses of a matrix by characterizing all solutions of the linear matrix equation LA = I. R = S J. Y) = Tr(X Tr F). .4. 3. is neither 11 nor onto. Let A = [i]:]Rl > ]R2. £. II? Is £. 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 . below bases for its four fundamental subspaces. Consider the vector space ]Rnxn over ]R. LetA [J] : E1 ~ E2. y) = Tr(X Y). — S^. 4. and let 7£ denote the subspace of skewsymmetric matrices. Prove Theorem 3. We give when considered as a linear on ]R3.1] is a left inverse. let denote the subspace of symmetric 2. 2 . Then A is 11. The matrix A = 1 1 2 1 [ 3 1 when considered as a linear transformation onIE \ is neither 11 nor onto. (Proof Theonly solution toO = Av = [i]v 2. We give below bases for its four fundamental subspaces. consider A linear transformation ]R3 1. J E2. 2.4. whence A/"(A) = 0 so A is 11). Let A = [8 5 i) and consider A as a linear transformation mapping E3 to ]R2. Consider the differentiation operator C defined in Example 3.Exercises 27 2. Consider the vector space R nx " over E. respect to this inner product. For matrices matrices. and let R denote the subspace of skewsymmetric matrices. let S denote the subspace of symmetric matrices. Consider differentiation £ 11? Is£ onto? onto? 4. It is now obvious that A has full column rank (=1) and A~L = [3 . 'R. EXERCISES EXERCISES 3 4 1. It is now obvious that A has full column is v 0.2. ThenAis 11.3. Again. matrix characterizing all solutions of the linear matrix equation LA = I. whence N(A) = 0 so A is 11).
How many linearly independent solutions can be found to the 10. Linear Transformations Chapters. . Are they equal? Is this true in general? If this is true in general. Suppose A E IR m xn has a left inverse. Theorem 6. Let A = [ J o]. if not.1 to illustrate the four fundamental subspaces associated with AT e associated ATE nxm IR from IR m R". Prove Theorem 3.12. provide a counterexample. Linear Transformations 7. Determine A/"(A) and 7£(A). prove it.1 11. Chapter 3. Show that AT has a right inverse. 3. ~ ~ 3 8. Let = [~ 9.4. Rnxm thought of as a transformation from Rm to IRn. Determine bases for the four fundamental subspaces of the matrix Detennine fundamental A=[~2 5 5 ~]. left T Suppose e Rmxn 9. Are they equal? Is this true in general? DetennineN(A) and R(A).11. Suppose A € Mg 9x48 .28 5. If E 1R~9X48. homogeneous linear system Ax = 0? homogeneous linear system Ax = O? n 3. linearly independent solutions 10.4.2. Modify Figure 3. Prove Theorem 3.Il.
1. as is shown in the following text. Although X and y were arbitrary vector spaces above. brings great notational and conceptual clarity matrix and.17. This transformation T~ + can be used to give our first definition of A+. let us henceforth consider the X ~n lP1. the definition neither provides nor suggests a good computational strategy good computational strategy for determining A +. T is bijective (11 and onto). pseudoinverse of A.1 4. Then A+ is the MoorePenrose where y = y\ pseudoinverse of A. as is shown in the following text.17. let us henceforth consider the Although X and Y were arbitrary vector spaces above.. problems. 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. a generIn this chapter we give a brief introduction to the MoorePenrose pseudoinverse.1 Definitions and Characterizations Definitions and Characterizations Consider a linear transformation A : X —>• y.l. and hence we Then. the MoorePenrose pseudoinverse of A.+ R(A) by dimensional Define transformation T : N(A). Define a transformation T : Af(A)1. see [22].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. With A and T as defined above. A purely algebraic y + characterization of A+ is given in the next theorem.. for determining A+ . 4.. as noted in the proof of Theorem 3. 29 . case X = W1 and Y = Rm.l. Definition 4.l.l —>• Tl(A) by Tx = Ax for all x E NCA). where X Xand Y y are arbitrary finitedimensional vector spaces. characterization of A is given in the next theorem. define a transformation A + y + X by where Y = YI + Yz with Yl e 7£(A) and Yz e Tl(A}L.1. as noted in the proof of Theorem 3. the MoorePenrose pseudoinverse of A. Then A+ is the MoorePenrose j2 with y\ E RCA) and yi E RCA). see [22]. T is bijective Cll and onto). can be used to give our first definition of A . where and are arbitrary finiteConsider a linear transformation A : X + y. and hence we can RCA) —>• J\f(A}~L This transformation can define a unique inverse transformation Tl 1 :: 7£(A) + NCA). a generalization of the inverse of a matrix." X ". which was proved by Penrose in 1955.m We A+ A e lP1. Then. We have thus defined A+ for all A E IR™xn. define a transformation A+ : Y —»• X by Definition 4. The MoorePenrose pseudoinverse is defined for any matrix and. which was proved by Penrose in 1955. neither provides Unfortunately. With A and T as defined above.
2 4. Example 4. Then Theorem 4. A+ = (AT A)I AT if A is 11 (independent columns) (A is left invertible). Example 4. Introduction to the MoorePenrose Pseudoinverse Theorem 4. However. one need simply verify the four Penrose conditions (P1)(P4). (P3) (AGf (P3) (AG)T = AG. Introduction to the MoorePenrose Pseudoinverse Chapter 4.30 Chapter 4. the Penrose properties do offer the great virtue of providing a tional algorithm. Given a matrix G that is a candidate for being the pseudoinverse of A..2 nor its proof suggests a computational algorithm. Also. For any scalar a. Example 4. A L = [3 — 1]) satisfy properties (PI).5.1. (P2) GAG = G. Let A E lR. this characterization can be useful for hand calculation of small examples. Note that the inverse of a nonsingular matrix satisfies all four Penrose properties. neither the statement of Theorem 4. this can be found in [1. p. Such a verification is often relatively satisfies all four. While not generally suitable for computer implementation. it must be A+.2.6. Note that other left inverses (for example. (P4) (GA)T = GA. Theorem 4. If G satisfies all four. straightforward. 19]. For any scalar a.5. A~ = [3 . (4. whose proof Still another characterization of A + is given in the following theorem. Consider A = [']. Furthermore. as with Definition 4.2. Let A e R™xn. (PI) AGA = A.2 nor its proof suggests a computawith Definition 4. Verify directly that A+ = Example 4.3. and (P4) but not (P3). Example 4.2) 4. terizations. Also. Example 4. it must be A +. 19].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.7. Verify directly that A+ = [ ~] satisfies (PI)(P4). p.4. Given a matrix G that is a candidate for being checkable criterion in the following sense. if a t= 0. characterization can be useful for hand calculation of small examples.6. A+ always exists and is unique. Example 4. Let A e R?xn Then G = A + if and only if (Pl) AGA = A. Such a verification is often relatively straightforward. if a =0.1) = limAT(AAT +8 2 1)1. one need simply verify the four Penrose conditions (P1)(P4). .4. Consider A = f ] satisfies (P1)(P4). Then G = A+ if and only if Theorem 4.7." xn. whose proof can be found in [1. Let A E lR. A+ = (AT A)~ AT if A is 11 (independent columns) (A is left invertible). Unfortunately. However. AG. A t = AT (AA )~ if A is onto (independent rows) (A is Example 4. then by uniqueness. (P2). Note that the inverse of a nonsingular matrix satisfies all four Penrose properties. and (P4) but not (P3).1. neither the statement of Theorem 4.1]) satisfy properties (PI). the Penrose properties do offer the great virtue of providing a checkable criterion in the following sense. L Note that other left inverses (for example. While not generally suitable for computer implementation. Unfortunately. (P2). a right or left inverse satisfies no fewer than three of the four properties. If G the pseudoinverse of A. A + always exists and is unique.3. X+ = AT(AATT) I if A is onto (independent rows) (A is right invertible). = Furthermore. then by uniqueness. (P4) (GA)T = GA. Then A+ [a [! = lim (AT A + 82 1) I AT 6+0 6+0 (4." xn. (P2) GAG G. Still another characterization of A+ is given in the following theorem. as a right or left inverse satisfies no fewer than three of the four properties.
10.3 Properties and Applications Properties and Applications This section presents some miscellaneous useful results on pseudoinverses. 2. Let S E Rnxn be symmetric with U TSU = D. . [~ ~ r ~ =[ 0 Example 4.8. 0 the four Penrose conditions. For any vector v E M".3. 31 31 Example 4. where U is orthogonal and D is diagonal. Many of these This section presents some miscellaneous useful results on pseudoinverses..10. Example 4. p. Theorem 4. e jRmxn and suppose Rmxm R n are orthogonal (M is T 1 1 orthogonal if M M ). The interested reader can consult the proof in [1. elements are determined according to Example 4. simply verify that the expression above does indeed satisfy each of Proof: For the proof. 27].9.7. Then Proof: For the proof. The interested reader can consult the proof in [1. 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. D Theorem 4. Let S e jRnxn be symmetric with UT SU = D. 4. 27]. . The Proof: Both results can be proved using the limit characterization of Theorem 4. Properties and Applications Example 4.11. where D+ is again a diagonal matrix whose diagonc D is diagonal.9. Many of these are used in the text that follows. Example 4. Example 4. simply verify that the expression above does indeed satisfy each c the four Penrose conditions.11. if v i= 0. Then S+ UD+U T where D+ is again a diagonal matrix whose diagonal elements are determined according to Example 4. (A T )+ = (A+{. if v = O.12.3.12. Let A E R m x "and suppose UUEejRmxm. For all A E jRmxn.7. are used in the text that follows. p.3 4. For A e Rmxn 1. [~ r 1 =[ 4 4 I I ~l 4 I I 4 4.4.13.4. For any vector e jRn. Theorem 4. . Then S+ = U D+UT.8. Proof: Both results can be proved using the limit characterization of Theorem 4.13. A+ = (AT A)+ AT = AT (AA T)+. Then orthogonal if MT = M.4. where U is orthogonal an Theorem 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.).VVEejRnxnx " are orthogonal (M is 4. Properties and Applications 4.
Proof: Proof: For the proof.g. then A+ = (ATA)~lAT. For e Rmxn .15. e. n(A+) 4. 0 The following theorem gives some additional useful properties of pseudoinverses. necessary and sufficient conditions under which the reverseorder property does hold are known and we quote a couple of moderately useful results for reference. since e lR~xr. the MoorePenrose pseudoinverse of any matrix (since AAT and AT A are symmetric). Theorem 4. (AB)+ = B+A+. 4. The result then follows by E lR. [5]. Theorem 4. in peneraK ucts of matrices such as exists for inverses of products.14.17. [9]. [] sufficient reverseorder However. (AA T )+ = (A T)+ A+. BB+ f r The by taking BI = B. then (AB)+ = B+ A+. If e lR~xm. in theory at least. Theorem 4. compute 4.32 Chapter 4. properties Theorem 4. (AB)+ = B+ A + if and only if 1.14. where BI = A+AB and A) = ABIB{. Then (AB)+ = 1+ = I while while B+ A+ = [~ ~J ~ = ~.12 Note that by combining Theorems 4. 1. = N(AA+) = N«AA T)+) = N(AA T) = N(A T). n(A T AB) ~ nCB) . B E Rrrxm. 4. = n(A T) = n(A+ A) = n(A TA). Introduction to the MoorePenrose Pseudoinverse Chapter 4. If A e Rnrxr.. 2.15.15.11 nets of matrices such as exists for inverses of nroducts Unfortunately. n(BB T AT) ~ n(AT) and 2. As an example consider A = [0 1J and B = [ : J.16. Then As an example consider A = [0 I] and B = LI.. [7]. however (see. If A is normal. poor (see. Proof' A+ A Proof: Since A E Rnrxr. [11]. [II].xm. 0 D Theorem 4.12 and 4.11 is suggestive of a "reverseorder" property for pseudoinverses of prodTheorem 4. whence A+A = f r. in general. see [9]. 3. (A+)+ = A. we B+ BT(BBT)I. and better methods are suggested in text that follows. e.15.16. then AkA+ = A+ Ak and (Ak)+ = (A+)kforall integers k > O.17. Introduction to the MoorePenrose Pseudo inverse 4. A+ = (AT A)I AT. Ir Similarly. (AB)+ = B+A+ if and only if 4. A\ = A in Theorem 4.13 we can. 4. (AB)+ = B?A+. xm + T B e Wr . D takingB t = B. (AT A)+ = A+(A T)+. see [5]. This A AT AT turns out to be a poor approach in finiteprecision arithmetic. N(A+) 5.13 can.g. (AB)+ = B{ Ai.At = A in Theorem 4. [23]). we have B = B (BBT)~\ whence BB+ = Ir. where BI = A+ AB and AI = AB\B+. Proof: Proof: For the proof.. [7].• Similarly. 0 D E lR~xr. TTnfortnnatelv. . For all A E lR mxn .
• 1 2 x. A e IRPxn thatN(A) S. whereupon y = Bx = AA+Bx E R(A). whereupon there exists a vector x e IR m such that Bx = y. e IRnxm. if A is symmetric. Let A G M"xn. so Proof: Suppose R(B) c U(A) and take arbitrary jc E Rm. 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. a matrix can be none of the preceding but still be normal.Exercises 33 Note: Recall that A E IRn xn is normal if A A = A T A.1 D. Then Bx E H(B) S. If jc. Then we have there exists a vector y e IRP such that Ay = Bx. Then Bx e R(B) c H(A).4 to compute the pseudoinverse of U .i l . U(A) if and only if AA+B = B. such as A=[ b a a b] for scalars a. Proof: Suppose K(B) S. prove that 7£(A) = 7£(AA r ) using only definitions and elementary properties of the MoorePenrose pseudoinverse. For A E IRmxn. Use Theorem 4. that B = AA+ B.18. we have shown that B = AA+B.i ]. Then R(B) c R(A) if and only if Suppose e IRnxp. Y e R". prove that R(A+) = R(A T). prove that RCA) = R(AAT) using only definitions and elementary 3.. where one of the Penrose properties is used above. = B. € IRm xm D 6. R(A) and take arbitrary x e IRm.4 to compute the pseudoinverse of \ 2 1. 5 e JRn x m . we have shown where one of the Penrose properties is used above. b e R for scalars a. To prove the converse. RCA). For example. A E IRn xn B E E n xm 6. Since x was arbitrary. Suppose A E Rnxp. Then B and take arbitrary y E R(B). assume that AA + B To prove the converse. or orthogonal. b E E. if A is symmetric. Then there exists a vector x E Rm such that Bx = y. However. Since x was arbitrary. or orthogonal. 4. so there exists a vector y E Rp such that Ay = Bx. then it is normal. Theorem 4. such as preceding but still be normal. problems. Note: Recall that A e R" xn is normal if AATT = AT A. y E IRn. and D E E mxm and suppose further that D is nonsingular. For A e Rmxn.i D. 0 EXERCISES EXERCISES 1. 2. For A E Rpxn and BE R mx ". ft(A+) ft(Ar 5. A+ 0 A+BD. show that JV(A) C A/"(S) if and only if BA+ A = B. Then we have Bx = Ay = AA + Ay = AA + Bx. (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. fiA+A B. then it is normal. assume that AA+B = B and take arbitrary y e K(B). B E E M X m . a matrix can be none of the skewsymmetric. For A e R m x n . Then K(B) S. N(B) and 5 € IRmxn. For example. Use Theorem 4. However. skewsymmetric. (xyT)+ = (xTx)+(yTy) yx T 3. show that (xyT)+ = (x Tx)+(yT y)++yxT. properties of the MoorePenrose pseudoinverse. A E IRmxn.].
This page intentionally left blank This page intentionally left blank .
recall. . e n} be a set of corresponding orthonormal eigenvectors 0= Ur+1 = . VI «xr j V U2 e ^x(mr) . Ch. . where S = [J °0].. i. V2 = [Vr+I. vectors.. (Note: The rest of the proof follows [24. y22 €E Rnxfor^ and the 0JM^/ocJb in E~are compatibly IRnx(nr). More S > o r > O.3) = Ulsvt· The submatrix sizes are all determined by r (which must be S min{m.. 5. More where ~ = [~ specifically. < min{m.. ..1 The Fundamental Theorem Theorem A Theorem 5. Let A e R™ x n ... . rcfr)..... . Pre. Then there exist orthogonal matrices U E Rmxm and E IR~xn.e.} with UI ::::: . U2 E IRrnx(mrl..Vn ].. Premultiplying by vt gives vt ATAVi = vt VI S2 = S2. . ..1. role throughout (numerical) linear algebra and its applications. i... Let {u.] . [24.} be a set of corresponding orthonormal eigenvectors and let VI = [v\. r write A r A VI = ViS2. Premultiplying by Vf gives Vf A T A VI write ATAVi = VI S2. recall. The SVD plays a key conceptual and computational of this important matrix factorization. Preand postmultiplying by SI gives the emotion S~l eives the equation (5.4) 35 .• = Un.1. Proof: Since AT A ::::: ( (AT A s symmetric and nonnegative definite..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 ). . for example.. . and a\ > • • • > Ur > 0. Theorem 5. Proof: Since A r A > 00 A r A i is symmetric and nonnegative definite. vn]. dimensioned. for example.2) (5. Letting — diag(cri. (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. .Vv r). .. and the Osubblocks in are compatibly dimensioned.e. S = diagfcri.. (5. «}). Ch. the latter VfV^S2 = S2. n}). its eigenvalues are all real and nonnegative.1) rxr A = [U I U2) [ ~ 0 0 ][ ] 2 T VI VT (5. we have n = U~VT. ii E !!.and postmultiplying by equality following from the orthonormality of the r. U\ E IRmxr. i e !!. Vi eE RIRnxr. . .. UI e Wnxr. . We In this chapter we give a brief introduction to the singular value decomposition (SVD).. we can and let V\ [VI.. e IRmxm and V E IR nxn such that V € Rnxn such that UI > diag(ul. u r ) e R . The SVD plays a key conceptual and computational role throughout (numerical) linear algebra and its applications. 6]).) Denote the set of eigenvalues of AT A by {of / E n} with a\ > • • > a > 0 = o>+i = • • an. 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. .) Denote the set of eigenvalues of AT A by {U?. Let {Vi. the latter equality following from the orthonormality of the Vi vectors.we can Vi = [vr+ ..\.LettingSS = diag(uI. its eigenvalues are all real and nonnegative.o>) E IRrxr.• ::::: Urr > as an exercise. . 6]). specifically.
1.. Referring to the equation V I = A VI SI defining U\. . . Remark 5. Introduction to the Singular Value Decomposition Chapter 5.16. Ui e (5. Remark 5. . . the IRmxr VI AViSI. The analogous complex case in which A E C™ x " is quite straightforward. Note that V and V can be interpreted as changes of basis in both the domain Remark 5. whence Vi ATAV22 = O. . The !:ingular value decomposition is not unique. Now V20 Vf A T A V 0. of the proof of Theorem 5.3.. Then T V AV =[ =[ VrAVI VIAVI VrAVI vIA VI Vr AVz vI AVz ] ~] since A V2 =0.1 reveals that proof Theorem any orthonormal basis for jV(A) can be used for V2 • £lny orthonormal basis for N(A) can be used for V2.. Remark 5. Referring to the equation U\ = A V\ S l defining VI. we see that Mat C is "i:.. in fact.2). Remark 5.4) we see that VrVI = /. Remark 5. The latter equality follows from the orthogonality of the columns of VI and V 2. Choose U2 £ IRmx(mr) [U\ U2] orthogonal. For example. cr. 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.? (AA I min{m. . .denote A thought of as aalinear transformation mapping IR n to IRm..... The latter equality follows from the orthogonality of the S and vI AVi = vI VI S = O. an examination decomposition Remark 5..).5. Then from (5. U be interpreted changes domain and codomain spaces with respect to which A then has a diagonal matrix representation. Choose any matrix V2 E ^ 77IX( ™~ r) such that [VI V2] is VI columns orthonormal.1 we see that ai(A) = A(2 (ATA) = £(A). Introduction to the Singular Value Decomposition Turning now to the eigenvalue equations corresponding to the eigenvalues or+\. singular 2. VT as AV = V"i:. u ]} for Rm (see the discussion in Section 3.16. Then Specifically. 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). . eigenvectors of AA TT).36 36 Chapter 5. 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 ). VT be an SVD of A as in Theorem 5.'. D to be S completes the proof... .1. Let A = V"i:.. an we have that ATAV2z = VzO = 0. VTAV = [~ Q].5.. See also m Remark 5.4) see UfU\ = columns of U\ are orthonormal. Thus..r zero singular values. . See also m [v\. matrix VI E M mx/ " by U\ = AViS~l. except for Hermitian transposes replacing transposes. vn } for IR and {u\. Specifically.. The decomposition is A = V"i:. } for R" {VI.3.. U V identical. 0 Definition 5. 1. where V and V are unitary and the proof is essentially proof decomposition A = t/E V H. i.4. we see that U{ AV\ = S and 1/2 A VI = U^UiS = 0. •. and defining this matrix U\ andU UT A V [Q ~].... of A A). From the proof of Theorem 5. let C. and codomain spaces with respect to which A then has a diagonal matrix representation.e. n} .(AT A) = is denoted ~(A).. singular unique. AV2 = 0. . V H. Then T rewriting A = V"i:.(A) At. analogous complex e IC~ xn straightforward. Definition 5.2). we see that.2. A. an we Turning now to the eigenvalue equations corresponding to the eigenvalues ar+l. responding to multiple cr/'s. of values of i I proof of A.2. 3. The set {ai. Thus. m for IR (see the discussion in Section 3. Now define the ATA V AV O.. n] — Note that there are also min{m.4. we see that V r A VI = since A V2 = O.(AATT).. with respect to the bases A = U^V as A V Mat £ is S the U E we see respect n and {u I.. ar}} is called the set of (nonzero) singular values of the matrix A and called [a\. C denote A thought of as linear transformation mapping W to W. to be ~ completes the proof. A = t/E VT SVD of A 5.
[1 0 ] . VT AV = A > O. Example 5. then A A. that "full SVD" (5. [25]. V2 (see Theorem 5. f/2. The Fundamental Theorem 5. 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. corner f/E V T r r T Ti isaanS V D o f AT. C/ [U\ Ui] orthogonal.8.10. if A = UI:VT is an SVD of A.5.3). A _ [ 1  0 ~ ] cose = [ .e. is the matrix I: and the span of the columns of UI.2) can always be constructed from a "compact SVD" (5.6.8.g. [7]. Then A = V A VTT is an A.2) can always be constructed from ¥2 Theorem too.11). Better algorithms exist that work directly on A via a sequence of orthogonal transformations. symmetric V orthogonal matrix of eigenvectors that diagonalizes A. is an SVD.9. and E U\. SVDof A.. Example 5. Computing AA AATT is numerically poor in finiteprecision arithmetic. [11].10. [7].. Let V be an orthogonal 5. F/vamnlp 5. U arbitrary 2 x orthogonal 5. Example A . however. 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). Note. The Fundamental Theorem 37 37 • any U22can be used so long as [U I U2] is orthogonal.. 0 Example 5.1. SVD of A.9.3).sine sin e cose J[~ ~J[ cose sine Sine] cose ' where e is arbitrary. VI:TU/ s n SVD of A VS C . e j8 the case).1. that aa"full SVD" (5. VI.7. 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). i.. Computing an SVD by working directly with the eigenproblem for AT A or 5. A=U is an SVD. [11]. What is unique. For example. too.e. SVD" (5.g. is an SVD. orthogonal transformations.U I U T . [25]..6. U2. Vi. e. see. AT A Remark 5. i. Let A e IRnxn" be symmetric and positive definite. e. see. VT A V = A = VAV eigenvectors > 0. 01  where U is an arbitrary 2x2 2 orthogonal matrix.
The elegance of the dyadic decomposition (5. the following properties hold: 1. the following properties hold: the notation of Theorem 5. . Let V = [UI. (b) R(U2) = R(A)1.5) as a sum of outer products Remark 5. Then A has the dyadic (or outer product) expansion product) expansion r A = Laiuiv.. [HI. Let A e jRrnxn have a singular value decomposition A = VLV T Using Theorem 5.13. Vn]. U2 = [Ur+I...1.1.7) explain why it is conventional to write the SVD and the key vector relations (5.£V. The relationship to the four fundamental subspaces is summarized knowledge of the rank r. Introduction to the Singular Value Decomposition Chapter 5. Theorem 5.7) AT Uj = aivi for i E r.. Remark 5. .6) and (5. for example. 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. .12.]. . u r ]. Part 4 of the above theorem provides a numerically superior method for Remark 5.£VTT as in Let A e jRmxn have a singular value decomposition A = UHV as in Theorem Theorem 5. . rank(A) = r = the number of nonzero singular values of A.11.. A = UZV. Let A E Rmxn have a singular value decomposition A = U'£ VT. 4. ..12.1. The singular vectors satisfy the relations AVi = ajui. say. The elegance of the dyadic decomposition (5. vn].8) where where .5) 3. . vr ]. . rank(A) = r = the number of nonzero singular values of A.. Introduction to the Singular Value Decomposition 5. andV2 = [Vr+I.. as A = UZV rather than. urn].. i=1 (5. . vn].= N(A T ).14. Then (5.5) as a sum of outer products and the key vector relations (5.. .6) and (5. . nicely in Figure 5.. Remark 5. Let A E E mx " have a singular value decomposition A = U. Then A has the dyadic (or outer 2.38 38 Chapter 5.1. um] and V = [VI.. 1. (c) R(VI) = N(A)1. A = U. . .2 Some Basic Properties Some Basic Properties Theorem 5. Let U =. 2. urn] and V = [v\.6) (5. say. . reduction to row or column echelon form. (5..13. . Then (a) R(VI) = R(A) = N(A T / . Then TheoremS.. reduction to row or column echelon form.1.. The singular vectors satisfy the relations 3. for example. Note that each subspace requires on. Note that each subspace requires knowledge of the rank r.. The relationship to the four fundamental subspaces is summarized nicely in Figure 5.7) explain why it is conventional to write the SVD as A = U'£VTT rather than. Using the notation of Theorem 5. (d) R(V2) = N(A) = R(AT)1.11.= R(A T ). VI = [VI. LetUI = [UI.2 5. . ..
=1 U. if we let the columns of U and V with the Qsubblocks appropriately sized.11.15.1.. Some Basic Properties 39 39 A r r E9 {O} / {O)<!l nr mr Figure 5.10) .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. . SVD and the four fundamental subspaces. 0 D (5. Note that none of the expressions above quite qualifies as an SVD of A + Remark 5. Furthermore. e^. then = L r 1 v. However. which is clearly orthogonal and symmetric. (or exchange matrix) P = [e erI. Some Basic Properties 5.. Proof: The proof follows easily by verifying the four Penrose conditions.11.. Proof' The proof follows easily by verifying the four Penrose conditions.. if we let the columns of U and V be as defined in Theorem 5. then be as defined in Theorem 5. ... . e\\. which is clearly orthogonal and symmetric. Remark 5. However. ed.5.er^\. SVD and the four fundamental subspaces. with the Osubblocks appropriately sized. a simple if we insist that the singular values be ordered from largest to smallest. Figure 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. e2.2. Furthermore.u.2. a simple reordering accomplishes the task: reordering accomplishes the task: (5. ..
notice that H(A) = K(AV) = R(UI S) and the matrix UiS e Rm xr has full K(UiS) and the matrix VI S E lR.17 and Remark 5. Recall the linear transformation T used in the proof of Theorem 3.. A "full SVD" can be similarly constructed.. Since T is determined by its action on a basis. while the matrix representation for the inverse linear transformation T~ with respect to the same bases is SI. .. finiteprecision arithmetic. .vvr}}is aa is r basisforN(A). notice that R(A) R(A V) This time. . From Section 3.[ SVr ] 0 mxn E lR.mxn. when derived by a Gaussian elimination algorithm implemented in finiteprecision arithmetic. ..1). Both compressions are analogous to the socalled rowreduced echelon form which. v } and {MI .2.3 Rowand Column Compressions Row and Column Compressions Row compression Let A E R have an SVD given by (5. since [u\..3 5.16. and since {VI. w. premultiplication of A by VT is an orthogonal transformation that "compresses" A by row transformations.. Such a compression is analogous to the "compresses" A by I. In other words.1. Then Again.2. is not generally as reliable a procedure. Introduction to the Singular Value Decomposition A+ = (VI p)(PS1 p)(PVr) is the matrix version of (5.1). .11). is the matrix version of (5. Then AV = V:E = [VI U2] [~ ~ ] =[VIS 0] ElR.urr}} e r.40 40 Then Then Chapters. Similarly.olumn transformations. then TlI can be defined by T^'M. the isabasisfor7£(. premultiplication of A by UT is an orthogonal transformation that rank. while the matrix representation for the inverse linear transformation TlI with respect to S. .r) and the matrix SVr e lR. / E~. . Then Let A E lR. Column compression Column compression Again. Such a row compression can also be accomplished by orthogonal row transformations performed directly on A to reduce it to the form 0 . and since ( v \ .i / E~. is not generally as reliable a procedure..l.. Introduction to the Singular Value Decomposition Chapter 5.... then T can be defined by TVj = OjUj .17 and in Definition 4. . in Definition 4. then T can be defined by TV.11). Notice that N(A) . = tv. Remark 5.mxn have an SVD given by (5.M(UT Notice that M(A) = N(V T A) = N(svr> and the matrix SVf E Rrxll" has full row A/"(SV. urr]} is clearly S.4). . when derived by a Gaussian elimination algorithm implemented in echelon form which. In other words. u is clearly matrix representation for T with respect to the bases { v \ . . Similarly. Since T is determined by its action on a basis. mxr has full column rank. Such a row compression can also be accomplished "compresses" A by row transformations. the matrix representation for T with respect to the bases {VI. From Section 3. then T~ canbedefinedbyTIu. Both compressions are analogous to the socalled rowreduced where R is upper triangular. .. vrr} and {u I.i e r.1).. where R is upper triangular. 5.16. = cr.. postmultiplication of A by V is an orthogonal transformation that "compresses" A by column transformations. Recall the linear transformation T used in the proof of Theorem 3.1. This time. . postmultiplication of A by V is an orthogonal transformation column rank. r x has full row rank.. the same bases is 5""1.. A "full SVD" can be similarly constructed. since {UI.1). In other words. Such a compression is analogous to the .. have an SVD given by (5. . . let A E lR. = ^u. .. basis forJ\f(A)±.. let A e Rmxn have an SVD given by (5. by orthogonal row transformations performed directly on A to reduce it to the form [~]. . u is a basis forR(A). In other words. Then VT A = :EVT = [~ ~ ] [ ~i ] D _ . .
[7]. Use the SVD to determine a polar factorization of A.1 starting from the observation that AAT ~ O. i.e. Let E ~~xn.[11]. A E IRnxn indefinite.1 starting from the observation that AAT > 0. see. Let A e E"xn be symmetric but indefinite. (b) Suppose that W and Y are nonsingular but not necessarily orthogonal. xyT 5. Let A e Rmxn and suppose W eRmxm and Y e Rnxn are orthogonal. 3. see. € IRmxn. 2. A = Q P 7. y E ~n Determine A e ~~ 4. [11]. . which is not generally a reliable procedure when socalled columnreduced echelon form. Prove Theorem 5. Let X E M mx ". [25]. [25]. Let x e Rm. Prove Theorem 5. EXERCISES EXERCISES 1. y e Rn be nonzero vectors.Exercises Exercises 41 41 socalled columnreduced echelon form. Use the SVD to determine a where Q is orthogonal and P p T > O. 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). for example. (a) Show that and W A F have the same singular values (and hence the same rank).. of defined by A defined by A = xyT. For details. 4. [7]. Note: this is analogous to the polar form where Q is orthogonal and P = PT > 0. = o. Determine an SVD of the matrix A E R™ xn E IRm. Note: this is analogous to the polar form iO z = rel&ofaa complex scalar z (where i = j = V^T). If XTX = 0.. A = QP 7. an SVD A.e. for performed by Gauss transformations in finiteprecision arithmetic. [23]. z of complex scalar z (where i j J=I). If XT X = 2.. Determine an SVD of A. 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. Determine SVDs of the matrices (a) (b) [ ] [ ~l 1 0 1 6. For details. show that X = 0. which is not generally a reliable procedure when performed by Gauss transformations in finiteprecision arithmetic. [23]. Determine SVDs of the matrices 5. Let A € R" X M .
This page intentionally left blank This page intentionally left blank .
A/"(A) = 0. only if rank(A) < n. n this is possible only ifm < n (since m dimT^(A) = rank(A) < min{m.e. 43 .. b]) = rank(A). 3. There exists a solution to (6.3) for all b E ]Rm if and only if A is nonsingular.. 4. b E lRm. there exists a solution if and only ifrank([A. equivalently. A E lR mxm and A has neither a 0 singular value nor a 0 eigenvalue. A E ]Rn xn.1 Vector Linear Equations Vector Linear Equations We begin with a review of some of the principal results associated with vector linear systems.3) is unique if and only if N(A) = 0. We begin with a review of some of the principal results associated with vector linear systems.3) for all b e W1 if and only if the columns of A are linearly independent. (6. i.3) 1. A E lRmxn. 6. and this is possible only ifm > n. A are linearly independent. (6. there exists a solution if and only j/"rank([A. Consider the system of linear equations Theorem 6.1. There exists a solution to (6.3} for e R m if only ifU(A) = W".. A is 11. i.3) if and only b E R(A)..3) for all b e W" if and only if is nonsingular. 2. and this is possible only ifm ::: n. b]) = rank(A). onto. equivalently.1) are studied and include. There exists a unique to (6. b E ]Rn. There exists a unique solution to (6. n. General linear systems of the form (6. There exists a nontrivial solution to the homogeneous system Ax = 0 if and only if Ax = 0 if 6. 1. A solution to (6.e.e.3) for all b E lRm if and only if the columns of 5. 3.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.3) for all b E lRm if and only ifR(A) = lRm. A solution to (6. equivalently. i. the familiar vector system are studied and include. the familiar vector system Ax = b.3) if and only ififbeH(A).1.e. (6. 5.. n}). 4. rank(A) < n. There exists at most one solution to (6.. General linear systems of the form equations. There exists a solution to (6. A is 11. i. Theorem 6.3) is unique if and only ifJ\f(A) = 0.e. N(A) 0.2) 6. Consider the system of linear equations Ax = b. A is 2. and this is possible only ifm :::: n (since m = dim R(A) = rank(A) :::: min{m.e.1 6. A G M m x m and A has neither a singular value nor a eigenvalue. equivalently. i. as a special case. as a special case. and onto. i. There exists at most one solution to (6.
premultiply by A: Proof: To verify that (6. A is not 11. (6. Linear Equations Chapter 6. Note that the results of Theorem of solutions to the general matrix linear system (6. For example. we prove part 6. E JR. i. Furthermore.1 follow from those below for the special case k = 1. specializing even further to the case m = n. we must have the case of a nonunique solution. The matrix linear equation AX = B.5). Then any matrix eRmxk of the form of the form X = A+ B + (/ .2 6.2 (Existence). AZ :::: B. (6.3.e. A is not II. Therefore. 0 . 0 Theorem 6.5).5) is a solution of is a solution of AX=B. to algebra.A+ A)Y. 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. of a matrix. Theorem 6.mxn. Let Z be an arbitrary solution of That all solutions arc of this form can be seen as follows. The matrix criterion is Theorem 4.3. Note that some parts of the theorem follow directly from others. a solution exists if and only if has a solution if and only ifR(B) S.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 (6.1 follow from those below for the special case = 1.4) has a solution if and only ifl^(B) C 7£(A). Let Z be an arbitrary solution of (6.e. B E JR. The matrix criterion is Theorem 4.1). i. R(A)..e. while results for (6.. note that x = 0 is always a solution to the homogeneous system. That all solutions are of this form can be seen as follows.mxk.6) are of this form. mxn .nxk is arbitrary.1). The matrix linear equation Theorem 6..6) Furthermore. Proof: To verify that (6. equivalently. a solution exists if and only if AA+B = B. all solutions of (6.2) follow by specializing even further to the case m = n. +B = Theorem 6.6). Note that the results of Theorem 6.18.6). which implies rank(A) < n must have the case of a nonunique solution. i.5) is a solution. to prove part 6.2 (Existence). AA+B B. equivalently.18. BE JR. D 6. which implies rank(A) < n by part 0 by part 3. For example. 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. all solutions of (6. Then we can write (6. 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. where Y E JR..6) are of this form. (6. AZ — B. and this is clearly of the form (6. Then we can write Z=A+AZ+(IA+A)Z =A+B+(IA+A)Z and this is clearly of the form (6.2)follow by 6.44 Chapter 6. A E JR.mxk and suppose that AA+B = B. Therefore.5) is a solution. while results for (6. Note that some parts of the theorem follow directly from others. i.e . Let A e Rmxn. note that x 0 is always a solution to the homogeneous system.
mxn. there is no "arbitrary" component. equivalently. X• = A~ B. (6.8) . 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).7.7) has a unique solution if and only if M(A) = 0.nxm is arbitrary. Example 6. this can occur if and only if rank(A) = r m (since equivalent to AA+Im = 1m.A+A) = O. There is a unique right inverse if and only if A+A = I/ e E"xm arbitrary. D 0 Example 6.7) is unique if and only if A+A = /.) Theorem 6.6. A+ A where Y E lR. Butrank(A) = n if and only if A is 11 or N(A) = 0. BE lR. Suppose A E lR. in which case A must be invertible and R = AI.S. Computation: Since y is arbitrary. It particular (6. leaving only the unique solution X = AI1B. Matrix Linear Equations 45 Remark 6. if there exists a unique. and (N(A) = 0). Ax — 0. N(A) = O. A A = f/E VT.) that minimizes TrXT X. where y e lR. Consider Example 6. All right inverses r < m) A (A+ of A are then of the form of A R = A+ 1m + (In . Characterize all right inverses of a matrix A E lR.8. y E R" A + A t= I. it is easy to see that all solutions are generated y from a basis for 7£(7 .7) has a unique solution if and only if unique if and only if A + A = I. A e E"x". (6. then it is easily R(I — A + A). find all A e ]Rmx". Matrix Linear Equations 6.5.6) +B Remark 6. r checked that 1.j jcj. But if A has an SVD given by A = U h VT. it is not unique. there exists a nonzero solution if and only if A+A /= I. recall that TrXT X = £\ •xlj.6 (Uniqueness). Solution: x=A+O+(IA+A)y = (IA+A)y. equivalently.9. wherer = rank(A) (recallr ::: h). A solution of the matrix linear equation Theorem 6.7. A A+ AI Remark (/ — A + A) 0. Thus. Example 6. Consider the system of linear firstorder difference equations (6. Find all solutions of the homogeneous system Ax = 0.nxn. When A is square and nonsingular.A+A). A R (AA(A) = A"1.4. / . vD Example 6. Here. Characterize AR = Im solutions R of the equation AR = 1m. Clearly. where r rank(A) (recall r < n).5. Proof: Proof: The first equivalence is immediate from Theorem 6. we write 1m to emphasize the m x m identity Im matrix. matrix. Hence.n is arbitrary. nonzero solution. Example 6. A solution of the matrix linear equation AX = B. 7£(A) and this is 7£(/m) c R(A) equivalent to AA + 1m Im. Clearly. Clearly.3. A E lR.A+A = Vz V2 and R(Vz2V^) = R(Vz) = N(A). But rank(A) = n that A+ A = / if r — n.2.. Remark 6. A+ = A"1 and so (I .2. (TrO denotes the trace of a matrix. It can be shown that the particular solution X = A+B is the solution of (6. rank(A) = < A This is equivalent to either rank (A) = r < n or A being singular.mxk (6. recall that TrX r = Li. The second follows by noting thatA+A = I can occur only ifr = n.6) that minimizes TrX7 (Tr() denotes the trace of a matrix.A+ A)Y =A++(IA+A)Y.9.A+ A V2 V[ and U(V = K(V2) = N(A). if and only if A is Ilor _/V(A) = O.mxn.6 (Uniqueness). Solution: There exists a right inverse if and only if R(Im) S. equivalently.
The matrices A = [ ° Q and f ^ provide an lability and reachability are equivalent. The condition dual to reachability is called observability: When does knowledge of {u 7 r/:b and {Yj l'.ra>l).: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.10) for k ~ 1. from the fundamental Existence Theorem. Theorem l'/:b Clearly. does there exA related question is the following: Given an arbitrary initial vector XQ.• A k kJ B] [ ~o (6. from the fundamental Existence Theorem.10.e. does there exist an input sequence {u j an input sequence {"y}"~o such that xn = O? In linear system theory. The matrices A = [~ ~]1and B5 == [~] 1 providean example of a system that is controllable but not reachable. The general solution of (6..8) is given by kJ Xk = Akxo + LAkJj BUj j=O UkJ ] Uk2 (6. . .8) of Example 6.e. AB. does there exist an input sequence {ujj 1jj^ such that Xk takes an arbitrary value in 1R"? In linear system theory.~ I).8) is controllable if and only if if controllability. this is called such that Xn = 0? linear system theory.9) ~Axo+[B. . We might now ask the question: Given XQ = 0. B) is if(AT .. m known as the state vector at time while Uk is the input (control) vector.2.. B T] is observable [reconsrrucrible] [controllablcl if and T) observable [reconstructive]. .8) is reachable if and only if if R([ B. if and only if or. we of reachability. B) iJ reachable [controllable] ifand only if (A . A related question is the following: Given an arbitrary initial vector Xo. linear differential equations). this is called controllability. The answers are cast in terms that are dual in the linear algebra sense as well. The vector Jt* in linear system theory is e IR nx " fieR" (n ~ I. We now introduce an output vector yk to the system (6. does there exist an input sequence for k > 1. We now introduce an output vector Yk to the system (6. A n . Since > 1. There are many other algebraically equivalent conditions.9 by appending the equation by appending the equation (6. A n .2. reachability always implies controllability and.. Since m ~ I. Example 6. standard conditions with analogues for continuoustime models (i. (A.T. The condition The answers are cast in terms that are dual in the linear algebra sense as well.9 Example 6.. equivalently. We might now ask the question: Given Xo 0. The above are standard conditions with analogues for continuoustime models (i. Again from Theorem 6.10.. The general known as the state vector at time k while Uk is the input (control) vector.11) with and D (p ~ 1).J B]) = 1R" or.8) of Example 6.. if and only if rank [B. The linear differential equations). Theorem 6.8) is given by solution of (6.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. this is a question {u }y~Q Xk of reacbability. equivalently. controlA 1 lability and reachability are equivalent.AB •. we have the notion of suffice to determine (uniquely) xo? As a dual to controllability.J B] = n. There are many other algebraically equivalent conditions. we have the notion of reconstructibility: When does knowledge of {u jy }"~Q and {. if A is nonsingular. see that (6. Linear Equations Xk with A E R"xn and B E IR nxmxm(rc>l. example of a system that is controllable but not reachable.46 46 Equations Chapter 6. overall system that are dual in the systemtheoretic sense to reachability and controllability. AB. we see that (6. We can then pose some new questions about the with C and (p > 1). this is a question va [Uj }k~:b such that x^ takes an arbitrary value in W ? In linear system theory.. We can then pose some new questions about the overall system that are dual in the systemtheoretic sense to reachability and controllability.
Then. and C E Rpxti.13) Let v denote the (known) vector on the lefthand side of (6. Listed below is a small collection of useful matrix identities.14) requires the notion A compact matrix criterion for uniqueness of solutions to (6. by definition. Verification of each identity is recommended as an exercise for the reader. equivalently. .11.3 A More General Matrix Linear Equation A More General Matrix Linear Equation AXC=B (6. C E jRmxn. Let A E Rmxn. (6. sociated e jRnxn.2 j BUj . mxm and D E jRm Invertibility is assumed for any component or subblock whose inverse is and D € E xm.27. in which case the general solution is of the form (6.6.4 6.6. so a solution exists. asbelow is a small collection of useful matrix identities.12) j=O Yo .DUnl 6. or.15) E jRnxp where Y € Rn*p is arbitrary. Listed In many applications. 6. indicated. is stated and proved in Theorem 13. Invertibility is assumed for any component or subblock whose inverse is indicated. in which case the general solution is of the has a solution if and only if AA + BC+C = B. notice that To derive a condition for observability. B E Rnxm. by definition. Then the equation e jRmxn. particularly for block matrices. arbitrary. so a solution exists. the solution is then unique if and only if N(R) ==0. Then.CBuo .13) and let R denote the matrix on Let denote the (known) vector on the lefthand side of (6. 0. e Tl(R). A E Rnxn. Thus. the coefficient matrices of interest are square and nonsingular. the has a solution if and only if AA+BC+C = B. if and only if r Yn]  Lj:~ CA n . By the fundamental Uniqueness Theorem. the solution is then unique if and only if N(R) Uniqueness Theorem.Duo Yl . v E R(R). B e jRmx q . e Rmxn. associated with matrix inverses. E jRnxm. Theorem 6. B E Rmxq. Theorem 6.14) Theorem 6. if and only if or.4 Some Useful and Interesting Inverses Some Useful and Interesting Inverses In many applications.4 Some Useful and Interesting Inverses 6.4 Some Useful and Interesting Inverses 47 To derive a condition for observability.14) requires the notion of the Kronecker product of matrices for its statement.13) and let denote the matrix on the righthand side. Such a criterion (CC+ <g) A+ A — I) is stated and proved in Theorem 13. the coefficient matrices of interest are square and nonsingular. notice that Yk = CAkxo Thus. Such a criterion (C C+ ® A +A = I) of the Kronecker product of matrices for its statement. particularly for block matrices. +L kl CAk1j BUj + DUk.3 6.Du] (6.. In these identities. By the fundamental the righthand side.6. and C e jRpxq. A compact matrix criterion for uniqueness of solutions to (6.27. Verification of each identity is recommended as an exercise for the reader. equivalently.
As in Example 6. 5.I B)I (E is the inverse of the Schur complement of A). It has many This result is known as the ShermanMorrisonWoodbury formula.48 Chapter 6.B D.I .I C) I.I EXERCISES EXERCISES 1.8. X. 1. 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 Mm xn .c E E") that arise in optimization (A + xx T ) — (with symmetric A e lRnxn and x e lRn) that arise in optimization theory. characterize all left inverses of a matrix A e lR ". Let A € E mx ". (A BDCr1 = AI .CA.1. Rmxk and suppose has an SVD as in Theorem 5.4. Note that the positions of the / and — / blocks may be exchanged.. [ / +c 7. where E = (D . This result follows easily from the block LU factorization in property 16 of Section 1. mx . result follows easily from the block LU factorization in property 16 of Section 1. Note that the positions of the / and . It has many applications (and is frequently "rediscovered") including.AIB(DlI + CAIB)ICAI. = = 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. This where E = (D — CA B) (E is the inverse of the Schur complement of A). BB EelR fflxk and suppose AAhas an SVD as in Theorem 5. As in Example 6. l = l = [!C / [~ ~ l = [ AI +_~~!~CAI A~BE = D. Let A E lRmxn.4. formulas for applications (and is frequently "rediscovered") including. (A + BDC)I = A~l .1. Linear Equations Chapter 6.8.I ] D. Linear Equations 1.. characterize all solutions of the matrix linear equation 7Z(B) c 7£(A). 1./ blocks may be exchanged.BDI l = [ AI BD. l 8. ization in property 17 of Section 1. 2. formulas for the inverse of a sum of matrices such as (A + D)lor (AI1 + DI)I.A~lB(D~ CA~lB)~[CA~l This result is known as the ShermanMorrisonWoodbury formula. r A~I [~ ~ r [D~I~AI D~I 1 ~r ~~B 1 r l [~ ~ r [D~CF +~~I~. theory.. 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 .4.4. for example. 2. where F = (A . that X~l [~ !/ [~ ~ r [~ ~ l [~ ~/ r [~ ~ 1 l l l = [ ~ 4. [~ ~ r l 3. 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. Assuming 2. for example. BC 6. Assuming R(B) ~ R(A).
e. .xy) T 1 49 = I  1 xTy 1 xy . Show that 3. Hint: Show that ct <= M(B). Let A E 1R~ " and let A 1 have columns Cl. Let A e R"xxn and let A"1 have columns c\. Show that 4.. y E E" and suppose further that XTy i= 1.. Show that (/ ..l ~i e.y Assume that Yji i= 0 for some i/ and j. .10..e. 6. T 4. Let x. As in Example 6. in Example 6. Show that the matrix B = A .x xTy). check directly that the condition for reconstructibility takes the 6. € IRn and suppose that x T y i= 1. c and individual elements y. A with yl subtracted from its (ij)th element) is singular. Show that cxJ C ' where c 1/(1 — T y). Show that the matrix B — A — —eie T : (i. 5. y E E" and suppose further that XTy ^ 1. Let x.... check directly condition for reconstructibility the form form N[ fA J CA n 1 ~ N(A n ). A with — subtracted from its (zy)th element) is singular. (i. Let jc. where C = 1/(1 . l' Hint: Show that Ci E N(B). Assume that x/( 7^ 0 for some and j. . y e IRn and suppose further that x T y ^ 1..10.Cn and individual elements Yij.e. .Exercises Exercises 3.
This page intentionally left blank This page intentionally left blank .
By Theorem 2. A linear transformation P is a projection if and only if it is idempotent. Also.y is linear and pl.. A linear transformation P is a projection if and only if it is idempotent. say on X along Y (using the notation of Definition 7. Define PX y : V + X <.yV = x for all v E V..yp2 = P. V by by PX. Oblique projections. Let V be a vector space with V X 0 y.Chapter 7 Chapter 7 Projections.x = I px.1 displays the projection of von both and 3^ in the case = Figure 7.x — I — Px. say on X along y (using the notation of Definition 7.1 displays the projection of v on both X and Y in the case V = ]R2.1.26. Px. Px.1 7. y x Figure 7. Define pX.1). Theorem 7.y.y.y is linear and P# y — px.y • V —>• c V has a unique decomposition v = x + y with x e X and y E y.e.1.y Theorem 7. 51 51 . Infact. Infact.2. P is a projection if and only if I —P is a projection.1). Let V be a vector space with V = X EEl Y.y is called the (oblique) projection on X along y. every v E V Definition 7. i. Also. and Norms Spaces. P2 = P. Proof: Suppose P is a projection.y is called the (oblique) projection on X along 3^. Py. Figure 7. Inner Product Spaces. Py. Theorem 7. Figure 7. y = Px. By Theorem 2. Proof: Suppose P is a projection.26.3.e. px. Inner Product Projections. every v e V has a unique decomposition v x y with x E and y e y. PX. P isaprojectionifandonlyifl P isaprojection.1.1.1 Projections Definition 7.3. and Norms 7. Oblique projections. Theorem 7. i.2.
P)x E XL. Inner Product Spaces.4. P E E"xn is the matrix of an orthogonal projection (onto K(P)} if and only 7. while Py = P(l . Thus.xJ. Then x T pT (I . Then Px = P2v = Pv = x so x E X. Then v if v € Pv (I .px. Essentially the same argument shows that / .XLtion and we then use the notation P x = PX. Px E R(P). we must have pT (I .P) = 0.XL iss called an orthogonal projection and we then use the notation PX = PX. Since x and y were arbitrary. T Since x and y were arbitrary.1 The four fundamental orthogonal projections The four fundamental orthogonal projections Using the notation of Theorems 5. P)x = y PT(I P)x = 0.1 5. V Pv . let A E Rmxn with SVD A = U!:VTT = A = UT. Projections. Conversely. P Proof: Let P be an orthogonal projection (on X. Inner Product Spaces..)x = PXJ. . Now let v E V be arbitrary. Then Px = p 2v = Pv = x so x e X. while Py = P(I P}v = x Pv . In the special case where y X1. First note that iftfveX. In the special case where Y = X^.P)x = xTP(I .3. then (/ .AA+ U2 U ! LUiUT.p 2 v 0 so y e Thus. we have ( P y f I (/ .P}x = 0. and Norms Let v e V be arbitrary.=1 m PR(A). y = (I . PX.X^X = Px±.11.3.5. we have (py)T ((I ..P)v = = Pv. Py e X. Then Pv = P(x + y) = Px = x.P)x = (I .xl. we must have P (I — P) = O.L 1.XL Theorem 7. P = P. (I .5.PX. P2v = P Pv — 2 2 Px = x = Pv. p2 = P. suppose p2 = P. Then v = Pv + (I . then Pv v. A 6 jRmxII UtSVf. Write x = P x + (I . Moreover.P)x = yTTpT (I . suppose P is a is a with the second equality following since pT P is symmetric. 0 Definition 7.. Write x = Px (I — P)x. Now let u e V be arbitrary. It is easy to check that X and Y are subspaces. First note that v E X. Then Pv = P(x + y) = Px = x. then Pv = 0. Projections. 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. We now prove and Y = {v E V : Pv = OJ. Let X n y. and Norms Chapter 7.uT. then v = 0. Moreover. Thus. arbitrary. X 0 y and the projection on X along y is P.xx by Theorem 7.4. Note that (/ . Conversely. Since Py E X. Then U\SVr Then r PR(A) AA+ U\U[ Lu. Thus. say.P)x e X1. Then symmetric projection matrix and let x be arbitrary.. Hence PT = PTP = P.P)v.52 52 Chapter 7. Thus. along 1) and let jc.P)x = x T P(l .1. P e jRnxn is the matrix of an orthogonal projection (onto R(P)) if and only ifP2 PT if p2 = p = pT. Note that (I . . px. p 2v = PPv = Let u E V be arbitrary.P)v.P)v. then v = O. Let X = {v e V : Pv = v} and y {v € V : Pv 0}.P)x = (I . If v E Y. with the second equality following since PTP is symmetric. i=r+l PN(A) 1. y = (I . yy Ee jR" be arbitrary.P)x = O.P)x = O. Hence if v E X ny.P2v = 0 so Y E y.P)x E ft(P)1 xTPT(I . V = X $ Y and the projection on X along Y is P. then Pv = O.A+A V2V{ L i=r+l i=l 11 ViVf. Thus. It is easy to check that X and 3^ are subspaces. (I . Thus. and P must be an orthogonal projection. D Essentially the same argument shows that I — P is the projection on y along X. Conversely. Conversely. Hence pT = pT P = P. If v e y.P)x. R" be Proof: Let P be an orthogonal projection (on X. Let x = Pv. Let X = {v E V : Pv = v} Px = x = Pv. suppose symmetric projection matrix and let x be arbitrary. Hence that V X 0 y. PN(A)J. * called an orthogonal projecDefinition 7.1 . We now prove that V = X $ y.P)v.P is the projection on Y along X. suppose P = P.1 and 5. since Px e U(P).V Theorems 5.1 7. D 0 7.xx by Theorem 7. say.11. then Pv = v. along XXL} and let x. mental subspaces.P)x 6 R(P)1and P must be an orthogonal projection.
X on Specifically. orthogonal: that z and u. e Rn Solution: Think of the vector w as an element of the onedimensional subspace IZ(w).2. Then X = PN(A)u + PN(A)X .2. the vector z that is orthogonal to w and such that Pv Moreover. Projections 53 Example 7. 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.2. in fact. . {VI.~) w. Determine the orthogonal projection of a vector e M" on another nonzero Example 7. See Figure 7.7. The expression for x\ is simply the orthogonal projection of XI projection of rather x on S. Vk} was an orthornormal Example 7. An arbitrary vector x e IRn was chosen and a formula for x\ appeared rather mysteriously. Projections 7.(:. Recall the diagram of the four fundamental subspaces. Then Let x e W be an arbitrary vector. IR n Rm 1 n Let X E IR be an arbitrary vector.7. Recall the proof of Theorem 3. Vk} was an orthomormal basis for a subset S of W1.. See Figure 7. Orthogonal projection on a "line. Then the desired projection is simply Then the desired projection is simply Pn(w)v = ww+v wwTv (using Example 4..7.8. A direct calculation shows z = Pn(w)"' = (l .2.1.8) (using Example 4.1. A direct calculation shows that and ware. Recall the proof of Theorem 3. Specifically.8.A+ A)x 2 = A+ Ax + (I = VI vt x + V Vi x (recall VVT = I). The indicated direct Example 7.11. The indicated direct sum decompositions of the domain E" and codomain IRm are given easily as follows. in fact.6." Figure 7." Example 7.Pn(w»v = v . There. . Recall the diagram of the four fundamental subspaces. Orthogonal projection on a "line.8) = (WTV) W... orthogonal: v z Pv w Figure 7... T W W Moreover.11. Determine the orthogonal projection of a vector v E IR n on another nonzero vector w E IRn.6. . . There. are. An arbitrary vector x E R" was chosen and a formula for XI basis for a subset of IRn. Example 7. Solution: Think of the vector w as an element of the onedimensional subspace R( w). { v \ .
x) ::: Qfor aU x 6V and (x. Y2 E V and/or all a. Inner Product Spaces. Ay) = {AT x. y\) + /3(jt. V = IRn. Then Similarly. Inner Product Spaces. Yl) + f3(x.12.54 Chapter 7. Then {^. y) = (y. 3. Yl. y)Q = XT Qy.13. definite defines Definition 7.9. Y2) for all x.y E V.) ) :: V x V + IR is a real inner is a real inner Definition 7.12. Example 7.13.x)forallx. where Q = QT > 0 is an arbitrary Q = Q T > is an Example 7. . y} = XTy is the "usual" Euclidean inner product or dot product. [ 5~2 + 7. Projections.10. Let V = IRn. and Norms Chapter 7. n x n positive definite matrix. Then ('. aYI + PY2) = a(x. y^} for all jc. such that {x. Let V = R". .AA+)y = U1Ur y + U2U[ Y (recall UU T = I). cryi + ^2) = a(x. y e V. If e IR mx ". {*. and Norms Similarly. 2.11. 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). Then Y = PR(A)Y + PR(A)~Y = AA+y + ( l . 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 . Example 7. f3ftE IR. (jc. (x. j2 ^ V and for alia. ATE IR nxm transformation Definition 7. (x. Then { • • V x V if product if 1. Let V be a vector space over R. Let V = E". y) x T Y is the "usual" Euclidean inner product or Example 7. respectively. x) for all x. y) = (y. If A E Rm xn. Then (x. y) for all x € Rm and for all y e R". Example 7. only ifx = O.2 Inner Product Inner Product Spaces Definition 7. 3. e R. (x. then AT e Rn xm is the unique linear transformation or map T E IRm andfor IRn. yi.(A . y) Q = X T Qy. as follows: and a vector in J\f(A). x } = 0 if and only ifx = 0. Let Example 7. let y e IR m be an arbitrary vector. . defines a "weighted" inner product. (x.9. (x. let Y E ]Rm be an arbitrary vector. Then (x.11. > Ofor all E V ( x x) =0 if 2. respectively.. Projections. Let V be a vector space over IR.10.
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
D 0 \\X\\2\\y\\2Note: This is not the classical algebraic proof of the CauchyBunyakovskySchwarz Note: This is not the classical algebraic proof of the CauchyBunyakovskySchwarz (CBS) inequality (see.33. 11·111 and 1I·IIClO XHUHUx . if U E enxn is unitary. \\Ux\\l XHX = \\x\\\). 217]).Q — (xhH Qx) 2. Some weighted pnorms: (a) IIxll1. Let x. In other words. Since yHxx = x Hy. i. The CBS inequality is thus equivalent to the statement I..e. On the vector space (C[to.l. e.g = (x QXY denoted  • c). y e C" may be defined by Remark 7..D = E^rf/l*/!.D = L~=ld. where 4 > O. then Remark 7. The angle 0 between two nonzero vectors x. [20. Remark 7. its determinant must be nonnegative. define the vector norm On the vector space «e[to. JC. > 0. it is particularly easy to remember. t\])n. p.tl Theorem 7. i.e. if U € C"x" is unitary. 217]). IIQ)' 1 3.31 and Remark 7. tO~t:5. define the vector norm 11111 = max 1/(t)I· to:::. denoted II . R). Since Proof: Consider the matrix [x y] e en x2 . Projections. In other words. define the vector norm 1111100 = max II/(t) 11 00 . Inner Product Spaces. and  .. [20. Since is a nonnegative definite matrix. Remark 7. Theorem 7.1^ IIUxll2 = IIxll2 (Proof IIUxili = x U Ux = xHx = IIxlli)· However. 112 is unitarily invariant. +=1.34. On the vector space (C[to. then H H \\Ux\\2 \\x\\2 (Proof.lx. Since yH = x H y.32. ttl. Then Theorem 7.30 (HOlder Inequality). The norm  • 2 is unitarily invariant.(x Hyy)(yH x).31 (CauchyBunyakovskySchwarz Inequality). and Norms 2. A particular case of the Holder inequality is of special interest.g.. Remark 7.> where Q = QH > 0 (this norm is more commonly = QH > Ikllz.32 are true for general inner product spaces. Let x.. o ~ (x Hxx)(yH y) .. whered. y e C". e.58 58 Chapter 7. it is particularly easy to remember. However. Projections. its determinant must be nonnegative. „ . 0 < e — 5. y E en. Then with equality if and only if x and yare linearly dependent. Ther. Proof' Consider the matrix [x y] E C" x2 .32 are true for general inner product spaces. p. Theorem 7.30 (Holder Inequality). t \ ] R). y E en may be defined by cos# = 1I. we see immediately that IXH y\ ~ 0 < ( x H ) ( y H y ) — ( x H ) ( y H x ) . and Norms Chapter 7.33. Then Fhcorem 7.^ cos e = IlMmlylb 0 ~ 0 < I' The CBS inequality is thus equivalent to the statement ~ ^  COS 0 < 1. 1Ft). The norm II . The angle e between two nonzero vectors x. Let x. Some weighted pnorms: 2.t~JI On the vector space ((C[to. However. (CBS) inequality (see. with equality if and only if x and y are linearly dependent.. we see immediately that \XH yl < IIxll2l1yllz. y e en. Theorem 7. (b) IIx IIz.31 and Remark 7. is a nonnegative definite matrix.g.~~1~1112'. Inner Product Spaces. However. y E C".32.31 (CauchyBunyakovskySchwarz Inequality). Remark 7. 1Ft). define the vector norm 3. Let x.  . 1cose 1~ 1.34. ttlr. p q I I A particular case of the HOlder inequality is of special interest.
IIxl12 :::: Jn Ilxll oo . we conclude this section with a theorem about convergence of vectors. and essentially obvious. 2.35. then we have the Pythagorean Identity Ilx ± YII~ = IIxll~ + IIYII~.e. there exist constants CI. Let II·• II be a vector norm and suppose v. If x... Extension to the complex case is straightforward and essentially obvious. i» (1) v(2\ . C2 (possibly 7. Theorem 7. Theorem 7.4.e. Finally. e C". i.36. the motivation for In this section we introduce the concept of matrix norm. Remark 7. All norms on C" are equivalent.39... ci (possibly depending onn) such that depending on n) such that Example 7.4 Matrix Norms Matrix Norms In this section we introduce the concept of matrix norm.37.. Definition 7. there exist Example 7.4 7. Similar remarks apply to the unitary invariance of norms of real are not unitarily invariant. Similar remarks apply to the unitary invariance of norms of real vectors under orthogonal transformation. 7.   R mx " ~ E is a matrix norm if it satisfies the following three Definition 7.e. Attention is confined to the vector space (IRmnxn. then we have the Pythagorean Identity Remark 7. the proof of which follows easily from liz II~ = ZH z. 3. i. IIxlioo :::: IIxllz. Extension to the complex case is straightforward what arises in the majority of applications.35. Matrix Norms 59 59 are not unitarily invariant. i. BE IRmxn. The using matrix norms is to have a notion of either the size of or the nearness of matrices. 2 the proof of which follows easily from z2 _ z_//. while the latter is needed to make sense of "convergence" of matrices. II·• II : IR mxn > IR is a matrix norm if it satisfies the following three properties: properties: 1. Let \\ \\ be a vector norm and suppose v. convergence in terms of vector norms. .. convergence in terms of vector norms. there exist vectors x for which equality holds: vectors x for which equality holds: Ilxlll :::: Jn Ilxlb Ilxll2:::: IIxll» IIxlloo :::: IIxll» Ilxlll :::: n IIxlloo. Matrix Norms 7. this is called the triangle inequality.4. The former notion is useful for perturbation analysis. Then 7.) . 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. If y € C" are orthogonal. the following inequalities are all tight bounds. Attention is confined to the vector space (W xn R) since that is what arises in the majority of applications. For x E en. E en. ConFinally... y E en are orthogonal. v(2)..e.e. All norms on en are equivalent. IIAII ~ Ofor all A E IR mxn and IR IIAII = 0 if and only if A = O. lIaAl1 = lalliAliforall A E mxn andfor all a E IR.e. IIA + BII :::: IIAII + IIBII for all A. about convergence of real numbers.37.7. the motivation for using matrix norms is to have a notion of either the size of or the nearness of matrices. As with vectors. For x G C". vectors under orthogonal transformation. i. v(l). As with vectors. there exist constants c\... i.36. IR) since that is "convergence" of matrices. the following inequalities are all tight bounds.38. Then lim k4+00 V(k) = v if and only if lim k~+oo II v(k)  v II = O.39.38.. (As with vectors. while the latter is needed to make sense of former notion is useful for perturbation analysis. we conclude this section with a theorem about convergence of vectors. i.
11·115.  5 2 = II IIF and  • 5i00 = II .mxn.60 Chapter 7.. where r mxn = rank(A). The "matrix analogue of the vector 1norm. The norm  .43. The "maximum row sum" norm is 2. ai..43." theorem and requires a proof. Example 7. The Schattenpnorms are defined by E lR.<110#0 IIxllq Example 7.40. 112' The norm II • 115.00 =  • 2. e R mx ". Let A E R . Let A E lR._ Then "mixed" norms can also be defined by e lR."  A\\ = ^ \ai} .. Let A E K m x ". (where r = laiiK^/i. 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. IIAlioo = max rE!!l. I Some special cases of Schatten /?norms are equal to norms defined previously." Each is a "computable.42. For example. 1.mxn. The following three special cases are important because they are "computable.42. The "maximum column sum" norm is 2. Example 7.41. to estimate the size of a matrix product A B in terms of the sizes of A and B individually. Inner Product Spaces. and Norms Example 7. is a norm. Let A E lR.40. 5>1 is often called the trace norm. ^wncic = rank(A)).  .. Example 7. pnorms previously.60 max _P IIAxll = max Ilxli p IIxllp=1 IIAxll p . Projections. Then the Frobenius norm (or matrix Euclidean norm) is 7. (A' A)) 1 ~ (T.mxn IIAII p. J=1 3. matrix = Ilxllp. + a!)"". Example 7.jj laij." IIAliss = Li. The spectral norm is 3. Example 7. defined by IIAIIF ~ (t. tTL T Note: IIA+llz = l/ar(A). (AA ')). (t laUI). Projections. Schatten/7norms IIAlls. IIAII2 = Amax(A A) = A~ax(AA ) = a1(A). I.p = (at' + .44.q = max IIAxil p 11. is a norm. IIAII P t altA)) 1 ~ (T. \\F and 11'115.mxn. Example 7. Inner Product Spaces.2 =  .44. Then the matrix pnorms are defined by A e Rmxn. and Norms Chapter 7. .) I ~ (t. I.1 is often called the trace norm. The "matrix analogue of the vector Inorm.
For example. For A following inequalities are all tight.e.. Example 7. A = max^o IIxll. •II p for all p are consistent matrix norms. 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. Not every consistent matrix norm is subordinate to a vector norm. more generally. •1122is consistent with II . while E ]Rnxn.e. \\v consistent with it. IIAII2::S IIAIIF. i. also caUedoperator norms. IIAxll1 = max . IIAlioo ::s .. e.e. exercise.60 .jii II A IIF.j is a matrix norm but it is not consistent. IIABIII.. If II . 2.45.q = maxx.jii IIAII I. so not exist a vector norm  . B e Rnxk. although there are analogues for.jii IIAlb IIAIIF ::s .. although there are analogues for. II such that AF is given by max^o ".46.60 Ilx i. HAjcJI^ ::s \\A\\m Ilxli v. For example. which equality holds: which equality holds: IIAIII ::s . B E ]Rnxk. For example. p for all p are consistent matrix norms. also called oper(or.oo 2 while IIAIII. take A = B = [: is a matrix norm but it is not consistent.. Matrix Norms 7.. Then the norms \\ . \\m is a consistent matrix norm. there exist matrices A for i.~~i'.so II . 2.60 IIx II Ilxll=1 IIAxll p .::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. wec1earlyhave IIAxll < IIAllllxll· Since Afijc < IIAlIllBxll < Afljt. Since IIABxl1 ::s Afljc ::s IIAIIIIBllllxll. if II A B II < II A 1111 fi whenever the matrix product is defined. II In II p = 1 for all p. there exists a vector norm II • IIv Theorem 7. 11^4^11 P (or. For such subordinate norms.jii IIAII I . Example 7. If \\ • 11m is a consistent matrix norm. consistent with it. )). IIAIIF ::s . IIAxliv < IIAlim \\x\\v' Not every consistent matrix norm is subordinate to a vector norm. Definition 7. inner products or outer products of vectors. There exists a vector x* such that IIAx*11 = IIAllllx*11 if the matrix norm is subordinate to the vector norm..1100 = max laijl x.  • /7and II . Notice that this difficulty did not arise for vectors.60 \^ • Useful Results The following miscellaneous results about matrix norms are collected for future reference. A matrix norm 11·11\\is said to be consistent mutuallyconsistentifIlABII.45.. IIAllp. II A 1100 ::s n IIAII I . Then :]. more generally. e.but there does  is consistent with F. 1. atornorms. a vector norm. •II F. l. the IIIn II F = .g. consider  • \\F. II F' ThenA^ < A II Filx 112. it follows that all subordinate norms are consistent. IIAIII ::s .. We thus need the following definition..jii. There exists a vector x* such that Ajt* = A jc* if the matrix norm is Theorem 7. i. II· II F and 1. Let A e Rmxn. .oo J1. \\ are Definition 7.jii IIAIIF. i.e.jii IIAlloo. We thus need the following definition. Theorem 7.e. 1. Matrix Norms 61 61 Notice that this difficulty did not arise for vectors.= max IIAxl1 x. we clearly have Ajc ::s A1jt. Theorem 7. Let A E ]Rmxn.ooIlBIII. not exist a vector norm II •  such that IIAIIF is given by max x .47. i. there exists a vector norm \\ . e R" x ". q For such subordmate norms. IIAII2 ::s . A A 2.48.and II \\ •lIy y are mutually consistent if \\ A B \\ a < IIAllfllIBlly.jii IIAII2.4. Then the norms II • \\a II· Ilfl' and .e.4. II". The "mixed" norm "mixed" norm II· 11 100 . IIAlioo ::s .48. IIAIIF ::s. For example.47. but there does consider II .jii IIAlioo' .. \\ • \\p.Then II Ax 1122 ::s II AFjc2. 2.46.jii IIAlb IIAIII ::s n IIAlloo. inner products or outer products of vectors. IIAII2 ::s. take A = B = \ \ Afl li00 = 2whileA li00 B 1>00 = 1. The following miscellaneous results about matrix norms are collected for future reference. it follows that all subordinate norms are consistent. IIAxl1 IIAII = max ... reader The interested reader is invited to prove each of them as an exercise.7.g. subordinate to the vector norm. .
i.. Then 7. Theorem 7. If P is an orthogonal projection. prove directly that V22Vl is an I — +A V V/ is an orthogonal projection.1. > . EeIRmxn..] 4. prove that P+ = P.. „ } The spectral radius of A is the scalar p(A) = max IA. Inner Product Spaces. space. 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. The spectral radius of A is the scalar by {Ai .  • 2 and  • \\F 8.. Then k~+oo lim A (k) = A if and only if k~+oo lim IIA (k)  A II = o. IIF and II . Chapter 7. A(I). 7. Prove that / . B) = space.49. For A eRmxa . but not necessarily The norms  • \\F and  • 2 (as well as all the Schatten pnorms.l. Show that the matrix norms II .c — v + = 0.y + 2z = O. [2 3 4]r R3 spanned by the plane 3. p+ = P. and Norms max laijl :::: IIAII2 :::: ~ max laijl. 3. spanned by the plane 3x . where V2 is defined as in Theorem 5. l. . . If P projection.. IIQAZlia = A fora = 2 or F. Suppose P and Q are orthogonal projections and P + Q = I.Q — Q must be an orthogonal matrix. and Norms Chapter 7. scalars. For A E IR mxn . 3. A (2) . matrices Q zR and e M" ". 4. 112 (as well as all the Schatten /?norms. . 112 and II . IIF are unitarily invariant.. for all A E IRmxn and for all orthogonal unitarily invariant.I. The norms II . where ¥2 is defined as in Theorem 5. i . Projections... \\ \\bea Rmx".] l. ..49.1..e. e Rmx" mxm x mxm and Z E IRnxn . Inner Product Spaces.. Prove that E"xn with the inner product (A. 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. Prove that P . but not necessarily other pnorms) are unitarily invariant. EXERCISES EXERCISES 1. Definition: Let A E IRnxn and denote its set of eigenvalues (not necessarily distinct) by P.62 62 3. Definition: Let A e Rnxn and denote its set of eigenvalues (not necessarily distinct) 8. Projections. B) = Tr ATB is a real inner product IR n x" AT B (A. 6. 2.A+A is an orthogonal projection.. Also. Suppose that a matrix A E IR mxn has linearly independent columns. 1. (MZa or F.e. A(2). i. A(A A) 1 AT 5. .An}. orthogonal projection.. A (1) . Let II ·11 be a matrix norm and suppose A. must be an orthogonal matrix. Find the (orthogonal) projection of the vector [2 3 4f onto the subspace of 1R3 5.
Let A = xy . where a and ft take the value 1. HA^. Determine IIAIIF' IIAIII> IIAlb and Aoo in terms of \\x\\a and/or \\y\\p. 9. Aj. If M is a magic square matrix. Let 9. where both x.Exercises Exercises 63 63 Let Let A=[~ 14 0 12 5 ~]. y E IR n are nonzero.) that  M Up = for all/?. it can be proved that IIMllp = ss for all p. 10. A2..) T 10. Determine IIAIIF' IIAII d . IIAlb IIAlloo. Determine AF. Let A = xyT. all of whose Determine AF. appropriate. is called a "magic square" matrix. and p(A). Let A=[~4 9 2 ~ ~]. (An n x n matrix. or oo as and II A 1100 in terms of IIxlla and/or IlylljJ. and peA). where ex and {3 take the value 1. and p(A). (An n x n matrix. Determine AF.. or (Xl as appropriate. y e R" are nonzero. Determine IIAIIF' IIAII Ilt. H A H ^ and peA). IIAlb IIAlloo. where both x. 2. H A I I A2. all of whose columns and rows as well as main diagonal and antidiagonal sum to s = n(n2 + 1) /2. \\A\\ A2.2. columns and rows as well as main diagonal and antidiagonal sum to s = n (n 2 l)/2. .
This page intentionally left blank This page intentionally left blank .
1 8. Now. 2. AT — A T Ax = AT b latter form is commonly known as the normal equations.b E 'R(A)L so these two vectors are orthogonal. A vector x E X if and only if ATrr = 0. Hence. (PR(A)b . (Pn(A)b — AJC) is clearly in 7£(A). x e X if and only if latter form is commonly known as the normal equations. is a solution of the normal equations.PR(A)bll~ + IIPR(A)b  Axll~ from the Pythagorean identity (Remark 7. A vector x X if and onlv if x is of the x=A+b+(IA+A)y. vector x e X if and only if AT where b — Ax is the residual associated with x.2) 65 .e.35).PR(A))b = PR(A).e. where r = b . The equations ATrr = 0 can be rewritten in the form A TAx = ATb and the x.1 The Linear Least Squares Problem The Linear Least Squares Problem Problem: Suppose A E Rmx" with m 2: nand b E jRm is aagiven vector. see Section 8.Chapter 8 Chapter 8 Linear Least Squares Linear Least Squares Problems Problems 8.x — b\\\ (and hence p ( x ) = from the Pythagorean identity (Remark 7. while (b . so these two vectors are orthogonal. Hence.2. IIrll~ = lib .Ax) is clearly in 'R(A).. (8.35). Solution: The set X has a number of easily verified properties: The set X has a number of easily verified properties: 1.. write the residual in the form r = (b . x E X if and only if x is a solution of the normal equations. IIAx . whereyEjRnisarbitrary. i..bll~ (and hence p(x) = \\Ax .b 112) assumes its minimum value if and only if II Ax —b\\2) assumes its minimum value if and only if (8.bll 2 is minimized}. For further details. while Now.Axll~ = lib .1) To see why this must be so. For further details. 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 . Thus.PR(A)b) = (I .PR(A)b) + (PR(A)b  Ax).Ax is the residual associated 1. A.2. The linear least Problem: Suppose A e jRmxn with > n and b <= Rm is given vector. i. Thus. A vector x E X if and only if x is of the form 2. write the residual r in the form To see why this must be so. see Section 8.
of linear least squares solutions. which follows since the two vectors are orthogonal.e. there is no "existence condition" such as R(B) S. The only difference is that in the case of linear least squares solutions. 5. The minimum value of p x ) is then clearly equal to lib . 3. X. Linear Least Squares Problems Chapter 8.. and only if A+A = I or.. X = {x"} = {A+b}.0)z) is clearly in 4. equivalently. There is a unique solution to the least squares problem. The general solution to e ]R. consider two arbitrary vectors jci = A+b + (I — A + A) y (I . consider two arbitrary vectors Xl = A + b 3. To see why. By Theorem 6. all solutions of (8. the last inequality following by Theorem 7.2. The minimum value of p ((x) is then clearly equal to where y E ]R. Linear Least Squares Problems and this equation always has a solution since AA+b e 7£(A). By Theorem 6. i. Just as for the solution of linear equations.A+A)y and *2 = A+b + (I — A+A)z in X. Then the convex combination 8x.e.PR(A)bll z = ~ 11(1 Ilbll z. problem to the matrix case. The unique solution of minimum 2norm or Fnorm is where Y € ]R. if 5. There is a unique solution to the least squares problem. if and only if rank(A) n..1]. x" = + b is the unique vector 4.66 Chapter 8. Then the convex combination and Xz = A+b (I . x* minimizes the residual p ( x ) and is the vector of minimum 2norm that does so.e. X = {x*} = {A+b}. then equality holds and the least squares If the existence condition happens to be satisfied. 1]. AA+)bI1 2 the last inequality following by Theorem 7.3. 0*i (1 #)* = A+b (I . To see why. Notice that solutions of the linear least squares problem look exactly the same as solutions of the linear system AX = B.1. equivalently. X is convex. we can generalize the linear least squares problem to the matrix case. In fact.A+A)(Oy (1 . + (1 . if and only if A + A lor. Notice that solutions of the linear least squares problem look exactly the Remark 8. we can generalize the linear least squares Just as for the solution of linear equations. The unique solution of minimum 2norm or Fnorm is X = A+B..8)xz2 = A+b ++ (I A+ A)(8y ++ (1 8)z) is clearly in X. In fact.e. if and only if rank (A) = n. x* minimizes the residual p(x) that solves this "double minimization" problem. Let A E E mx " and B € Rmxk. This follows immediately from and is the vector of minimum 2norm that does so.2.2) are of the form solutions of (8.1) and which follows since the two vectors are orthogonal. X has a unique element x* of minimal 2norm. If the existence condition happens to be satisfied. where y e W is arbitrary. Remark 8.23. X = A+B. Let 8 e [0. all and this equation always has a solution since AA+b E R(A).mxn XElR Plxk min IIAX  Bib is of the form is of the form X=A+B+(IA+A)Y.nxk is arbitrary.2) are of the form x = A+ AA+b + (I  A+ A)y =A+b+(IA+A)y.3. i. X is convex.n is arbitrary. then equality holds and the least squares .23. where Y E R" xfc is arbitrary. The Theorem 8. there is no "existence condition" such as K(B) c R(A).. i. 7£(A). i.mxk. The only difference is that in the case same as solutions of the linear system AX = B.. Let 6 E [0. x* = A+b is the unique vector that solves this "double minimization" problem. has a unique element x" of minimal2norm. This follows immediately from convexity or directly from the fact that all x e X are of the form (8. BE ]R.1) and convexity or directly from the fact that all x E X are of the form (8.A+ A)z 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
as in Theorem 5. VT = U. [7]. Then GI defining y = logy. The subvector z2 is arbitrary. 8. . respectively. Sometimes a problem in which the Ci'S appear nonlinearly nonlinearly can be converted into a linear problem.. ] II: = II [ The last equality follows from the fact that if v [~~]. it is shown that solution [4].5) is that the coefficients Ci appear linearly. Z2 while the minimum value of \\Ax — b II ~ is l^llr while the minimum value of II Ax . then taking logarithms yields the equation logy = logci + cjt. The former is much more expensive but is generally more reliable and offers considerable theoretical offers insight. the subvectors can have different lengths). norm. This that orthogonality is not what is used here. are based on orthogonal polynomials. we assume that A has an SVD given by A = UT. etc. quantity above is clearly minimized by taking z\ = S'c. c\ logci. the two are equivalent.can be arbitrarily nonlinear.. etc. fact. (8. 's ¢i. the subvectors can have different lengths).4 8. of linear least squares problems via the normal equations can be a very poor numerical method in finiteprecision arithmetic. [7]. Linear Least Squares Problems difficulties because of numerical ill conditioning for large n. we assume that A has an SVD given by A U\SVf via the SVD. S~lc\. appear functions </>. Since the standard Kalman filter essentially amounts method in finiteprecision arithmetic.SVr U~VT Theorem 5. For example. This explains why it is convenient to work above with the square of the norm rather than the concerned. Ib is unitarily invariant =11~zcll~ wherez=VTx. [11]. 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 can be converted into a linear problem. piecewise polynomial functions. Two basic classes of algorithms are A itself S VD and QR (orthogonalupper triangular) factorization. then u^ = i>i \\\ \\vi\\\ (note The last equality follows from the fact that if v = [£ ].6) via the SVD. then II v II ~ = II viii ~ + II v211 ~ (note that orthogonality is not what is used here.4 Least Squares and Singular Value Decomposition Least Squares and Singular Value Decomposition In the numerical linear algebra literature (e. Specifically.1. problem. Better numerical methods are based on algorithms that behavior in practice (and it does). Specifically. bE IR m . e C2 / then taking logarithms yields the equation log y = log c. z.b\\^ is II czll ~. Then c. since II . In this section we investigate solution of the linear least squares problem min II Ax x b11 2 . Numerically better approaches ill difficulties n. arbitrary.[ ~~ ] II: sz~~ c.g.70 70 Chapter 8. c. respectively. The key feature in (8. [23]). if the fitting function is of the form y t) Y = ff( (t) = c\eC2i. splines. Since the standard Kalman filter essentially amounts to sequential updating of normal equations. [4]. c. The basis functions coefficients c. As far as the minimization is concerned. Linear Least Squares Problems Chapter 8.1.c=UTb = II [~ ~] [ ~~ ] . + c2f. the last equivalent. A E IRmxn . The former based on SVD and QR (orthogonalupper triangular) factorization. insight. In fact. For example. it can be expected to exhibit such poor numerical behavior in practice (and it does).. Better numerical methods are based on algorithms that AT work directly and solely on A itself rather than AT A. We now note that IIAx  bll~ = IIU~VT x =  bll~ II ~ VT X  U T bll.
e.A + A)_y.mxm. can be quite reliable. 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. If we label the product of such orthogonal row transformations as the orthogonal matrix QT E R m x m .~xn. Another expression for the minimum residual is  (/ — AA + )b 2 . we again look at the solution of the linear least squares problem (8. V2z is an arbitrary vector in R(V2 = N(A). to reduce A in the following way.1).. and there is thus "no V2 part" to the solution. This matrix factorization is much cheaper to compute time in terms of the QR factorization.5 Least Squares and QR Factorization Least Squares and QR Factorization In this section. This follows easily since (7 . Z VISici The last equality follows from The last equality follows from c = UTb = [ ~ f: ]= [ ~~ l Note that since Z2 is arbitrary. of course. an important special case of the linear least squares problem is the socalled fullrank problem. A e R™ X M . A finite sequence of simple orthogonal transformations. It is then possible. where y e ffi.e. This follows easily since Another expression for the minimum residual is II (I . To simplify the exposition. Least Squares and QR Factorization Now transform back to the original coordinates: Now transform back to the original coordinates: x = Vz 71 71 = [VI V2 1[ ~~ ] = VIZ I + V2Z2 = = + V2Z2 vlsIufb + V2 2. we again look at the solution of the linear least squares problem (8. Thus.. x has Note that since 12 is arbitrary. This matrix factorization is much cheaper to compute than an SVD and.. with appropriate numerical enhancements. of course. i. Least Squares and QR Factorization B.6) but this time in terms of the QR factorization.AA+)bll~ .S. an important special case of the linear least squares problem is the Finally. 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.8.e. can be quite reliable.5. Thus. i.AA+)bllz. i. = +b + (/ — A + A) y. Finally. we add the simplifying assumption that A has full column To simplify the exposition. In this case the SVD of A is given by socalled fullrank problem. via a sequence of socalled Householder or Givens transformations.1). than an SVD and.AA+)b\\22 = \\U2Ufb\\l = bTU2U^U22V!b = bTU2U*b = \\U?b\\22.6) but this In this section. is orthogonal to all vectors in 7l(A}L b E R(A). we add the simplifying assumption that A has full column rank. If we label the product of such orthogonal row transformations as the to triangular form. via a sequence of socalled Householder or Givens rank.5 8. i. to reduce A in the following way. 11(1. This agrees. A e 1R™ X ". with appropriate numerical enhancements. with (8.~xn. It is then possible. A E ffi. we have QT € ffi.e. where y E Rm is arbitrary. This agrees.m is arbitrary. V2 Z 2 is an arbitrary vector in 7Z(V2)) = A/"(A). (8. UZV = [U t/2][o]^i r > and there is thus "no V2 part" to the solution. 8..11U2U!b"~ = bTUZV!V UJb = bTVZV!b = IIV!bll~. x has been written in the form x = A+b + (I .7) . with (8. In this case the SVD of A is given by A A = V:EVTT = [VI{ Vzl[g]Vr. A E ffi.
Suppose qi and q2 are two orthonormal vectors and b is a fixed vector.72 Chapter 8. we see that in (8. and qz are two orthonormal vectors and b is a fixed vector.7). data. Qz] [ (8. sense). we have = R~l Qf b = +b and the minimum residual is IIC?^!^ EXERCISES EXERCISES 1. Suppose q. (8. check directly that (I .9) Any of (8. (3. Both Q I and Qz2 have orthonormal columns. i.. Note that Any of (8. yt): (1. we have x = R. Both Q\ and <2 have orthonormal columns. all in R". are orthogonal vectors. all in ffi. 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). b e ffi. (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 € ffi.8).3).7).1Q\b = A+b and the minimum residual is II Qr bllz' is \\C2\\2.[ ~~ ] If:..+ A)y and A+b are orthogonal vectors. (b) Let r denote the "error vector" b .~xn is upper triangular. Yi): 2. xn..A+A)y and A +b 1.e. 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. Equivalently. Now write Q = [Q\ Qz].aql . n . (2.flq2 Show that r is orthogonal to (b) Let r denote the "error vector" b — ctq\ — {3qz. both ql and q2 .Show that r is orthogonal to both^i and q2. 112 is unitarily invariant ~ ] x . or (8. we see that A=Q[~J = [QI = QIR.m IX(m ~" ) .8). data. 2.e.9) are variously referred to as QR factorizations of A. b E Em. m and any e ffi.mxn and where R e M£ x " is upper triangular. Note that (8.Cl and the minimum residual The last quantity above is clearly minimized by taking x = R lIc\ and the minimum residual is Ilczllz.9) is essentially what is accomplished by the GramSchmidt process. (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. Now write Q = [QI Q2].9) is essentially what is accomplished by the GramSchmidt process.. check directly that (I ..2). where Q\ e R mx " and Qz € K" x(mn). Consider the following set of measurements (*. The last quantity above is clearly minimized by taking x = R.I of the columns of yields the orthonormal columns of Q\. and any y E R". 3.Equivalently. where QI E ffi. Linear Least Squares Problems Chapter 8. or (8. Now note that Now note that IIAx  bll~ = IIQ T Ax = II [ QTbll~ since II . i.. Multiplying through by Q in (8. by writing (8.7). For A E Wmxn .7). Multiplying through by Q Q2 E ffi.1). Consider the following set of measurements (Xi. Linear Least Squares Problems where E ffi. 3.9) are variously referred to as QR factorizations of A. (8. R~l) ) of the columns of A yields the orthonormal columns of QI.8) ~ ] (8.
9). and suppose A = QR.yII2 as 8 approaches O? (b) Now consider the perturbation E2 = [~ (b) Now consider the perturbation EI = \0 s~\ of A. 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. Solve the perturbed problem min II A 2 z . then A+ == R. where Q is orthogonal. not necessarily nonsingular. Find all solutions of the linear least squares problem min II Ax .bl1 2 when A = [~ ~ ] and b = [ !1 x The solution is (a) Consider a perturbation EI = [~ pi of A. (a) Consider a perturbation E\ = [0 ~] of A.bll 2 x when A = [ ~ 5. 7. where again 8 is a small of A. where again 8 is a small positive number. Prove that A+ = R+ QT.xn can be factored in the form (8. where 8 is a small positive number.:. Let A e R"x". Solve the perturbed problem positive number.bib z n where A2 = A + E2. where AI = A + E I .IlQf.9). Solve the perturbed version of the above problem. and suppose A where is 1. Consider the problem of finding the minimum 2norm solution of the linear least 5. of 2norm solution of least «rmarp« problem squares nrr»h1<=>m min II Ax .• What happens to IIx* . verify that if A E ~. Let A E ~nxn. then A+ R~ Q\. Use the four Penrose conditions and the fact that Q\ has orthonormal columns to 6. Solve the perturbed version of the above problem.z2 as 8 approaches O? where A2 — A E 2 What happens to \\x* — zll2 as 8 approaches 0? 6. not necessarily nonsingular. A+ R+QT . What happens to IIx* — y 2 as 8 approaches 0? where AI = A + E\.Exercises Exercises 73 4. Find all solutions of the linear least squares problem 4. What happens to jt* . where 8 is a small positive number.
This page intentionally left blank This page intentionally left blank .
) = det (A . [21]) or directly using elementary properties of inverses and determinants (see. Theorem 9. then vector of AH associated with I. n(A) = O. Definition 9. Note that if x [y] is a right [left] eigenvector of A.3 (CayleyHamilton). Note that if x [y] is a right [left] eigenvector of A. then n(A) is a polynomial of degree n.31 = 0. (Note that the characteristic polynomial can also be defined as det(Al . The 2nonn is the most common a — \j'.1 Fundamental Definitions and Properties Fundamental Definitions and Properties Definition 9./ — A). for proved easily from the Jordan canonical fonn to be discussed in the text to follow (see. Let A [~ ~]. [3]). Theorem 9.1).1 9.1). It is an easy exercise to 2 verify that n(A) = A + 2A . such that a scalar A E e. It can be proved easily from the Jordan canonical form to be discussed in the text to follow (see.Chapter 9 Chapter 9 Eigenvalues and Eigenvalues and Eigenvectors Eigenvectors 9. we see immediately that XH is a left eigenvector of A H associated with A. Thus.1) Similarly. 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.A).1. then so is ax [ay] for any nonzero scalar a E C. This results in at most a change of sign and.2.t so that the scaled eigenvector has norm 1.1. as a matter of convenience. [21D or directly using elementary properties of inverses and determinants (see. It can be The following classical theorem can be very useful in hand calculation. The 2norm is the most common nonn used for such scaling. the Fundamental Theorem of Algebra says that 7t (X) is a polynomial of degree n. called an eigenvalue. such that Ax = AX. for example. for example.4. we see immediately that x H is a left eigenBy taking Hermitian transposes in (9.. Then n(k) = X2 + 2A. Example 9. as a matter of convenience. norm used for such scaling. we use both forms results in at most a change of sign and. A nonzero vector x e C" is a right eigenvector of A e Cnxn if there exists a scalar A.— 3.Al) is called the characteristic polynomial Definition 9.3 (CayleyHamilton).4. For any A e Cnxn . It can be proved from elementary properties of determinants that if A E C" ". then It can be proved from elementary properties of detenninants that if A e enxn .2) By taking Hennitian transposes in (9. we use both forms throughout the text. a nonzero vector y e C" is a left eigenvector corresponding to an eigenvalue a if Mif (9. n(A) = 0. [3]). It is an easy exercise to Example 9. This of A. verify that n(A) = A2 2A .) throughout the text. The polynomialn (A./) is called the characteristic polynomial of A. Let A = [~g ~g]. called an eigenvalue. (9. Then n(A) A2 + 2A 3.} The following classical theorem can be very useful in hand calculation. Thus. e C. A nonzero vector x E en is a right eigenvector of A E e nxn if there exists Definition 9. the Fundamental Theorem of Algebra says that x 75 .2. The polynomial n (A) = det(A—A.31 O. for example. For any A E e nxn . example. a nonzero vector y E en is a left eigenvector corresponding to an eigenvalue Similarly. (Note that the characteristic polynomial can also be defined as det(A.
8.2». If A E A(A) has algebraic multiplicity m.~. if we denote the geometric multiplicity of A by g. Example 9. checked eigenvectors of A and AT (take Hermitian transposes of both sides of (9. i. the n(A) coefficients. But it also clearly satisfies the smaller degree polynomial equation isfies (1 . The geometric multiplicity ofX is the number of associated of algebraic multiplicity m...e. A. then n(X) has real coefficients.nxn is the polynomial o/(X) oJ IPll. less than) its algebraic multiplicity. the set of Definition 9. but that A(A) = A(A) only if A e R"x".3) are the eigenvalues of A and imply the singularity of the matrix A . Note. Eigenvalues and Eigenvectors n(A) has n roots. Then n(A) = A22. f3 e R and let A = [~f3 £ ]. Theorem If A E 1Ftnxn. A matrix A e 1Ft x" is said to be defective if it has an eigenvalue whose geometric multiplicity is not equal to (i. i. A is said to be defective if it does not have n linearly independent (right or left) eigenvectors. Thll minimal polynomial of A G l!if. such a polynomial is said to be monic and we generally write et (A) as a monic polynomial throughout the text)...:..5. i. The spectrum of A E C"x" is the set of all eigenvalues of A. then we must have 1 :::: g :::: m. But it also clearly satisfies the smaller degree polynomial equation (it. such a polynomial is said to be monic and we of the highest power of A to be +1.. independent eigenvectors = n .. .. say.e. For example. we get the interesting fact that det(A) = AI . say. sible for A to satisfy a lowerorder polynomial. Eigenvalues and Eigenvectors Chapter 9. If is a root of multiplicity m of n(A). n(A). The spectrum of A e nxn is the set of all eigenvalues of A. must occur in complex conjugate pairs. then I < dimA/"(A — A/) < m. we always have A(A) = A(AT).76 Chapter 9. • AM(see and set X = 0 in this identity. as solutions of the determinant equation n(A) = det(A  AI) = 0.3) in the n(A) = det(A . It can be shown that or(l) is essentially unique (unique if we force the coefficient It can be shown that a(Je) is essentially unique (unique if we force the coefficient of the highest power of A to be + 1. a .5. it is posn(A) = O. However. then A satsible for A to satisfy a lowerorder polynomial. From the CayleyHamilton Theorem.AI) = (A] . it can also be . Moreover. if If A € A(A) has algebraic multiplicity m. y of AT y is a left eigenvector of A corresponding to A e A(A). Then jr(A.. A matrix A E Wnxn is said to be defective if it has an eigenvalue whose Definition 9.7. then there is an easily checked relationship between the left and right If A € R"x".n =0o. Let the eigenvalues of A E en xxn be denoted X\.AI) :::: m.rank(A . as solutions of the determinant equation 7r(A) has n roots. it can also be generally write a(A) as a monic polynomial throughout the text). of we always have A(A) = A(A r ).2)). eigenvalues of A. too. E A(A). we say that A is an eigenvalue of A Definition 9. These roots.4) and set A = 0 in this identity. If e Wxn. Let a.. Moreover.A) (9. Definition 9. Definition 9...2aA + aa2+ f322 and A has eigenvalues a f3j (where A has eigenvalues a ± fij (where j = i = R). then 1 :::: dimN(A . The spectrum of A is denoted A(A). The spectrum of A is denoted A (A).25).8.) A. eigenvalues of A. Specifically. must occur in complex conjugate pairs.. However..AI). For example.e. and hence further guarantee the existence of corresponding nonzero eigenvectors.1)2 = O. Equivalently. all roots of its characteristic polynomialn(A).ft Definition 5. less than) its algebraic multiplicity. possibly repeated. ft E 1Ft and let A = [ _^ !]. A is said to be defective if it does not have n linearly independent (right or left) eigenvectors. IfXA is a root of multiplicity m ofjr(X). The geometric multiplicity of A is the number of associated independent eigenvectors = n — rank(A — A/) = dim J\f(A — XI).7. (An . then A satisfies (Je — I)2 = 0. Thus. if left of A A E A(A).. if eigenvectors of A and AT (take Hermitian transposes of both sides of (9. we say that X is an eigenvalue of A of algebraic multiplicity m. i •>/—!)• If A E 1Ftnxn. if A = \1Q ®]. then we must have I < g < m.1) . Definition 9. possibly repeated.. c form form A e C" " A]. An. Equivalently. . (9. we get the interesting fact that del (A) = A] • A2 • • An (see also Theorem 9.5.6. 2aA + 2 + ft and Example 9.2. the set of all roots of its characteristic polynomial n(X).. but that A(A) = A(A) only if A E 1Ftnxn. Let a. guarantee the existence of corresponding nonzero eigenvectors. we denote the geometric multiplicity of A by g. degree such that a (A) =0. Xn.e. These roots. if A = [~ ~]. geometric multiplicity is not equal to (i. then y is a right eigenvector of AT corresponding to I € A (A). Specifically. that by elementary properties of the determinant. The minimal polynomial Of A l::: K""" ix (hI' polynomilll a(A) of least degree such that a(A) ~ O. A. Thus. Hence the roots of 7r(A).6..AI) = dimN(A . neftnhion ~. and hence further are the eigenvalues of A and imply the singularity of the matrix A — AI.XI.A) . we know that n(A) = 0.e. Then if we write (9..
such that Aj =£ Ai. g. eigenvector is numerically unstable.2)' ""d g 2. left Aj E A (A) such that Xj 1= A. shown that a (A. 0 0 0 2 0 0 0 2 ] h'" a(A) (A .. sociated eigenvector structure).. this algorithm. Bezout algorithm.1. 0 g At this point.. The matrix A~U has a(A) I 2 0 0 2 0 0 0 !] (9. Fundamental Definitions and Properties 77 77 a(A) f3(A) O.11. each 4. 0 0 0 2 A~U 0 0 ] ha<a(A) (A . Let A E cc nxn and let Ai be an eigenvalue of A with corresponding right eigenvector jc.*.) directly (without knowing eigenvalues and as Unfortunately.2)' ""d g ~ ~ ~ 1. A[~  2 0 I 2 0 0 0 0 0 0 !] ~ ~ ~ ha.. such is not the case. n(A) = (A — 2)4.1. Example 9. Let A e C« x " ana [et A. i. the geometric multiplicity by g. . e l\(A) yj Xi. a(A) divides n(A).2)2 ""d g 3.e.2)4. Then Xi = 0.) divides every nonzero polynomial fi(k} for which ft (A) = 0. let Yj be a left eigenvector corresponding to any A.9. Unfortunately.5) = (A  2)2 and g = 2. a(X) n(X).."(A) ~ ~ ~ ~ (A . be an eigenvalue of A with corresponding right Theorem 9. one might speculate that g plus the degree of a must always be five. Fundamental Definitions and Properties 9. every nonzero polynomial f3(A) particular. The above definitions are illustrated below for a series of matrices. A~[~ A~U 2 0 0 I 2 2 ] ha< a(A) (A .11. Then yfx{ = O.. Theorem 9.10. Furthermore. Proof' Since AXi AiXi. Unfortunately.10.2) andg ~ 4. of which has an eigenvalue 2 of algebraic multiplicity 4. called the Bezout algorithm. a(A) directly (without knowing eigenvalues and asThere is an algorithm to determine or (A. We denote 7r(A) (A . algorithm. YY Proof: Since Axt = A. In particular. Example 9.
. Let A e nxn be Hermitian.. However. D A =1= iJ. Theorem 9. Proof: Proof: For the proof see. we must have Subtracting (9. An with corresponding Theorem 9. we have xHX =1= 0. x) is an arbitrary eigenvalue/eigenvector pair such that Ax = AX. Let us now return to the general case.14) and would thus have to be 0.. Theorem 9. The same result holds for the corresponding left eigenvectors. since YY A = AjXjyf.11. for example. i.e. or else Xi would be orthogonal to n linearly independent vectors (by Theorem 9. Then (9. we must have that x H = 0. Eigenvalues and Eigenvectors Chapter 9..xnn • Then {XI.— A. x is a linearly independent set. 0 D Theorem 9.6) from (9. [21.12.6) from (9. XH x /= 0.. c Proof: Suppose (A. or both. 118].14) well. 0 If A E cnx " has distinct eigenvalues. result holds for the corresponding left eigenvectors. Then all eigenvalues of A must Theorem 9.e. since x is an eigenvector. yr . . Let A E nxn be Hermitian and suppose X and iJ. (9.11 is very similar to two other fundamental and important The proof of Theorem 9. so that y H Xi = 1 for each i.. 1 ?..e.5). A. i..7) yields Using the fact that A is Hermitian.n with corresponding right eigenvectors x\. Since yf*Xi =1= 0 for each i.e. Let A E cnxn have distinct eigenvalues A . p. is If A e C nxn has distinct eigenvalues. so that YitH x. AXH z. 118]... A = AH. the proof see. respectively. Then Proof: Suppose (A. or both.e.=1= 0. we can choose the normalization of the Xi'S. Since A. ^ /z. A = AH..JC by ZH to get ZH Ax = X z Hx . However.Aj ^ 0. Then x and z must be of A with corresponding right eigenvectors and respectively.. c Proof: Premultiply the equation Ax = A. i. and if A. Then and z must be orthogonal. 's.. The same right eigenvectors XI. Since Xi ^ 0 and would thus have to be 0.Aj)YY xi. xn}} is a linearly independent set. Since Ai .e.14. Let A e C"x" be Hermitian and suppose A and /JL are distinct eigenvalues Theorem 9. Cnxn have distinct eigenvalues AI.7. Since equation Az i^z XH get X H Az = iJXH A. Take the Hermitian A transpose of this equation and use the facts that A is Hermitian and A is real to get xXHAz == of equation facts A. contradicting the fact that it is an eigenvector.. is real to get H Az AxH z. A. Let A €. results. the two vectors must be orthogonal. Chapter 9. Let A E C"x" be Hermitian. then by Theorem 9.7) Taking Hermitian transposes in (9. 0 Let us now return to the general case.is real. from which conclude A. for i € !1.. .JC.— A y )j^jc.13. Take the Hermitian Proof: Premultiply the equation Ax = AX by ZH to get ZH Ax = AZ H x.. the two vectors must be orthogonal. x . .. and if Ai E A(A). Premultiply the equation Az = iJZ by x H to get x HAz = /^XHZZ = XxHz.. Since XxHz. from which we conclude I = A.11 is very similar to two other fundamental and important results. then by Theorem 9.11.. it cannot be the case that yf*xt = 0 as orthogonal to all yj's for which j ^ i. Eigenvalues and Eigenvectors yy. or else xf would be orthogonal to n linearly independent vectors (by Theorem 9. e A(A). . A is real. x) is an arbitrary eigenvalue/eigenvector pair such that Ax = A. or the Yi 's. we have that XxH = AXH x. . since x is an Using the fact that A is Hermitian. i. orthogonal. ... jc. Xi is orthogonal to all y/s for which j =1= i... i. be real. we must have yfxt = O. p. contradicting the fact that it is an eigenvector. 0 The proof of Theorem 9.. However. or the y.14. for [21. = A. . = 1 for/ E n.13.6) Subtracting (9... Then all eigenvalues of A must be real. it cannot be the case that YiH Xi = 0 as well.12..78 Similarly. YyXi =0. we must have that XHzz = 0. since y" A = Similarly.5). i.are distinct eigenvalues of A with corresponding right eigenvectors x and z. 's. we can choose the normalization of the *. we have that IXHxx = XxHx.. Then [x\. eigenvector..7) yields Taking Hermitian transposes in (9. However. we find 0 = (Ai . we find 0 = (A.
/ e n.. i E !!:: Finally. solve the linear system (A — (—1 + 2j)I)x2 = 0 to get yi X2 =[ 3+ j ] 3 ~/ . corresponding to these eigenvalues."" yn] be the matrix of corresponding left eigenvectors.. Let A E C"x" have distinct eigenvalues A. i E!!. Let Example 9. suppose that the left and right eigenvectors have been normalized so that yf1 Xi = 1..11) Example 9. These matrix equations can be combined to yield the following matrix factorizations: These matrix equations can be combined to yield the following matrix factorizations: XlAX and and A (9. let A = diag(AJ. and this then determines the other two (since dimA/XA — (—2)7) = 1).15. . Let 2 5 3 3 2 4 ~ ] . For Al = —2. Then rr(A) = det(A .ci can be set arbitrarily.9) =A = XAyH yRAX n (9. Yn] ing right eigenvectors form a matrix X = [XI.I = = LAixiyr i=1 (9. i E !!. — 1 2 j } . xn]. . + 5). let Y — [y\. solve the (since dimN(A .8) while y^Xj = 5.I . Furthermore. . xn]. 10) (A. A. To get the corresponding left eigenvector y\.(2)1) = 1). . Fundamental Definitions and Properties 9. 1 ± 2j}. . An and let the corresponding right eigenvectors form a matrix X [x\..*/.1. Finally.j. can be written in matrix form as AX=XA (9.16.. / en. Xn) E Wtxn.(1 + 2j) I)x2 = 0 to get For A2 — 1 + 2j.. solve the linear system (A . Similarly.22 + 2)" + 5). . . To get the corresponding left eigenvector YI. can be written in matrixform as diag(A...A. solve the 3 x 3 linear system (A — (—2}I)x\ = 0 to get For Ai = 2.nand let the correspondTheorem 9. 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. Similarly..10) = XAX.. let A = right eigenvectors have been normalized so that YiH Xi = 1. / en. .3 4A2 9 A. is expressed by the equation while YiH Xj = oij.15.. We can now find the right and left eigenvectors corresponding to these eigenvalues. .. Let A e en xn have distinct eigenvalues AI.. Furthermore. solve the 3 x 3 linear system (A . Fundamental Definitions and 79 Theorem 9.. 2A.. ...AI) (A. is expressed by the equation yHX = I.2 + 9)" + 10) = ()" + 2)(). suppose that the left and be the matrix of corresponding left eigenvectors. from Then n(X) det(A .1.16. For A2 = 1 + 2j. j e n.(2)l)xI = 0 to get Note that one component of . from which we find A(A) = {—2.. y' E !!. This time we have chosen the arbitrary scale factor for y\ so that \ = 1. Then AJC. let Y = [YI.. and this then determines the other two Note that one component of XI can be set arbitrarily. . We can now find the right and left eigenvectors which we find A (A) = {2. An) e ]Rnxn. .9. = A. Then AXi = AiXi../) = (A 3 + 4A. 2)(A. .
Then Jl"(A) = det(A . itit is gtruightforw!U"d to compute straightforward to comput~ X~[~ and and I 0 I i ] 1 x. we For XT.!. Let A = [~ ~ ~] . use the fact that A33 = A2 and simply conjugate the equation A.'s directly.( I + 2 j) I) = 0 and nonnalize Y22 so that y"xX2 = 1 to get For A3 = — 1 — 2j.2 and simply can also note that x$ = X2 and Y3 = Y2. Proceeding as in the previous example.. Then. XIAX=A= [ 2 0 0 1+2j o 0 Finally. —4}. Proceeding as in the previous example.c2 = ^. we could proceed to solve linear systems as for A2. 3. instead of determining the j.=.2*2 to get Ax^ = ^2X2. we could have found them instead by computing XI and reading off its rows. similar argument yields the result for left eigenvectors.2 However. —3./) _(A + 8A from which we find A(A) = {—1. + 3)(A. use the fact that A.=. To see this. + 4).2j.) 19X + 12) = (A.L Other results in Theorem 9.80 Chapter 9. Now define the matrix X of right eigenvectors: Now define the matrix of right eigenvectors: 3+j 3j 3. 4}. Finally.15 can also be verified. = I . For example. 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 . Let Example 9.A. A. we could proceed to solve linear systems as for A. note that we could have solved directly only for XI and X2 (and X3 = X2).~q 1 3 2 2 0 2 3 ] 2 ~ y' .AI) = (A33 + 8A 22+ 19A + 12) = (A + I)(A + 3)(A + 4). we can also note that X3 =x2' and yi jj.±1 4 4 4 l+j . we could have found them instead by computing instead of detennining the Yi'S directly. Then. For example.!. for left eigenvectors. note that we could have solved directly only for *i and x2 (and XT.15 can also be verified.L 4 !.A similar argument yields the result conjugate the equation AX2 — A2X2 to get AX2 A2X2. 2 It is then easy to verify that It is then easy to verify that 2 . o 3 Then 7r(A.j ] 3+j . To see this. Eigenvalues and Eigenvectors Chapter 9. However. = x2). X~l Example 9. is from which we find A (A) = {I.17. Other results in Theorem 9.17. + 1)(A. det(A .
then easy to show that the eigenvalues of f(A) (defined as L~:OanAn) are f(A). I 3 I (. For example. 2 3 I (. Eigenvalues (but not eigenvectors) are invariant under a similarity transformation T. 3.20. Proof: Suppose (A. —4). What is true is that the independent right eigenvectors associated with the eigenvalue 0. Theorem 9. say.18. formation T.but f(A) does not necessarily have all the same eigenvectors (unless. X) is an eigenvalue/eigenvector pair such that Ax = AX. For left eigenvectors we have a similar statement.g. For left eigenvectors we have a similar statement. —3. say.1. since T Proof: Suppose (A. A = [~ Oj ] have all the same eigenvectors (unless. Let A E R" xn and suppose X~~1AX — A. i=1 . If / is an analytic function (e. but A2 = f0 0~1]has two has only one right eigenvector corresponding to the eigenvalue 0. 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 9.3 0 ~l +(4) [ . D D yHA = XyH ifandon\yif(T(T Hy)H (T. x) but not conversely.g. which is equivalent to the dyadic expanWe also have X~l AX = A = diag( 1. If f is an analytic function (e. we have the equivalent statement (T. in general. or eX. ( x ) is a polynomial. eigenvalue/eigenvector pair (A.. The following theorem is useful when solving systems of linear differential equations. where A is diagonal. Then suppose XI AX = A. 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 11.20. Remark 9.19.I = A(T. A is diagonalizable). namely yH A AyH if and only if Hy)H(T~1AT) = A(T Hy)H. A Theorem 9. jc) is an eigenvalue/eigenvector pair such that Ax = Xx.1 AT) =X(THyf. The following theorem is useful when solving systems of linear differential equations. x) maps to (/(A). or. What is true is that the eigenvalue/eigenvector pair (A. e jRnxn n = LeA. ] [~ ~ (I) [ I (. A = T0 6 2 has only one right eigenvector corresponding to the eigenvalue 0.) . of Chapter 11. but /(A) does not necessarily the eigenvalues of /(A) (defined as X^o^A") are /(A). from which equivalent statement (T~ AT)(T. which is equivalent to the dyadic expansion sion 3 A = LAixiyr i=1 ~(I)[ ~ W~l+(3)[ j ][~ ~ 1 . I I ~J I 2 0 0 0 3 3 3 I I (.lIAT)(T~lx)x) = X ( T ~ lIxx). or ex.txiYiH. Eigenvalues (but not eigenvectors) are invariant under a similarity transTheorem 9. Then. A is diagonalizable). representable by a power series X^^o fln*n)> then it is easy to show that representable L~:O anxn). Fundamental Definitions and Properties 9. 4).18. Then. For example. x) but not conversely. or sinx.19. ff(x) is a polynomial. x) maps to (f(A). but A = [~ has two independent right eigenvectors associated with the eigenvalue o. Remark 9. Fundamental Definitions and Properties 81 81 We also have XI AX = A = diag(—1.1. or sin*. since T is nonsingular..9. from the theorem statement follows. namely the theorem statement follows.
2 9. Theorem 9. eigenvectors Xi. .. .. Eigenvalues and Eigenvectors n = LeA. If A E Rnx" is diagonalizable with eigenvalues A. Eigenvalues and Eigenvectors Chapter 9. we have Proof' Starting from the definition.20 and its corollary in which A is not necessarily diagonalizable. . € n_..e.12) where each of the lordan block matrices 1i . from which such a result is then available and presented later in this chapter..IXiYiH.. ff(A) = X f ( A ) X ~ l I = Xdiag(J(AI). of course... and right eigenvectors xt•. It is necessary first to consider the notion of Jordan canonical form. and right Corollary 9.22.e. then e A has eigenvalues e A There are extensions to Theorem 9. / € n_..2 Jordan Canonical Form Jordan Canonical Form Theorem 9.. 1q)..22. .82 Proof: Starting from the definition.Il . . canonical form.20 and Corollary 9. Corollary 9. i.21. to have a version of Theorem 9. i=1 0 The following corollary is immediate from the theorem upon setting t = I. to have a version of Theorem 9... 0 (9. analytic on the spectrum of A. Jordan Canonical Form (/CF): For all A e c nxn with eigenvalues AI. . 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. ( A ) = Xf(A)X..20 and Corollary 9.i)..21 for any function that analytic on the spectrum of A. lordan Canonical Form (JCF): For all A E C"x" with eigenvalues X\.13) 1i = o o Ai o Ai ..21 for any function that isis There are extensions to Theorem 9. there exists X € c~xn such that (not necessarily distinct). 9.. i E ~.= Xdiag(/(A. f ( X t t ) ) X ~ It is desirable. kn E C (not necessarily distinct). The following corollary is immediate from the theorem upon setting t = I. ... /' en. An e C 1. i E ~.. then eA has eigenvalues e X"i . there exists X E C^x" such that XI AX = 1 = diag(ll. i. It is necessary first to consider the notion of Jordan A is not necessarily diagonalizable. If A e ]Rn xn is diagonalizable with eigenvalues Ai.. and the same eigenvectors.20 and its corollary in which It is desirable. ii E ~. f(An))X.. . we have Chapter 9. . from which such a result is then available and presented later in this chapter..21. and the same eigenvectors. I. of course.
For nontrivial Jordan blocks.. ] T (X .. 1q is of form where each of the Jordan block matrices 11.9. complicated.. Proof: For the proof see. X (not necessarily X E lR. and where M. Real Jordan Canonical Form: For all A E R n x " with eigenvalues AI.An n (not € jRnxn Xi. With 1 j o j o 1 o o o j ~ ~] 0 1 ' . 83 83 Form: 2.. = [ _»' ^ 1 and h2 = [6 ~] in the case of complex conjugate eigenvalues where Mi = [ _~. .... Proof: proof D 0 Transformations like T = [ _~ "•{"]allow us to go back and forth between aareal JCF Transformations like T = I" _. Jordan Canonical Form 9. { ] allow us to go back and forth between real JCF and its complex counterpart: TI [ (X + jfJ o O. for example. .. Jordan Canonical Form and L. Jq is of the form of in the case of real eigenvalues A.2.14) J\.=1 ki = n.. 120124].~xn necessarily distinct). . [21. e A (A).JfJ =[ (X fJ fJ ] (X = M. there exists X € R" xn such that (9. pp.. . ~: ] and I = \0 A in the case of complex conjugate eigenvalues a ± jfJi E A(A). . the situation is only a bit more complicated. (Xii±jpieA(A>).. .2.
1. 1). 2)2.24. i=1 n 2.. 2(A(A..2)2. Thus. Let A e C" " with eigenvalues AI. The characteristic polynomials of the Jordan blocks defined in Theorem 9.26.) = det(7) = ]~["=l A. Suppose A E E (A. Suppose A e lR.1)4(A . " Xn.I)(A (A2)2. 2 .24.22 we have that A = X J X ~ . .1)2.26. The characteristic polynomials of the Jordan blocks defined in Theorem Definition 9. (A1).jf3 0 0 et . The characteristic polynomial of a matrix is the product of its elementary Theorem 9..2).25.1)z. 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 . and (A . The minimal polynomial of a matrix is the product of the elementary divisors of highest degree corresponding to distinct eigenvalues. Then c n 1. From Theorem 9. Thus.. highest degree corresponding to distinct eigenvalues. I) . . and (A . The minimal polynomial of a matrix is the product of the elementary divisors of divisors. Theorem 9. from Theorem 9. Again.. i=1 l Proof: Proof: 1. Eigenvalues and Eigenvectors Chapter 9. .2)2.— I) (A. 1 det(A) = det(X J XI) det(J) = n7=1 Ai. and 2).1*) = Tr(J) = L7=1 Ai.2)3 3and is known to have :rr(A) Example 9.. X XI.(A 1).(A.1)2. An.) = (A.22 we have that A X J XI. The characteristic polynomial of a matrix is the product of its elementary divisors. x Theorem 9..25. Tr(A) = 2.(A. Let A E nxn with eigenvalues AI.84 it is easily checked that it is easily checked that Chapter 9. from Theorem 9. J(2) has elementary divisors (A while /( 2) haselementary divisors (A . 2... .22 we have that A = X JJX ~ l . 9.jf3 0 ]T~[~ l h M Definition 9.22 are called the elementary divisors or invariant factors of A. .23.22 l Tr(A) = Tr(X J XI) Tr(JX.22 are called the elementary divisors or invariant factors of A. — 2) .23. D 0 Example 9.)i. det(A) = det(XJX.2).. Thus.2(A(A .I [ "+ jfi 0 0 0 et + jf3 0 0 0 0 et . Eigenvalues and Eigenvectors T. I) . 1)2(A . From Theorem 9.2).2)2. Then AAhas two possible JCFs (not counting reorderings of the a (A.. Then Theorem 9. . 1). and (A (A . 1)4(A 2) and 2 2 et(A) = (A .. det(A) = nAi. . Tr(A) = Tr(XJX~ ) = TrC/X"1 X) = Tr(/) = £"=1 A. .7x7 is known to have 7r(A) = (A .
1.l).. X e A(A) if and only if (A XI)kx = 0 and (A U}k~l x ^ 0. both are eigenvectors (and are independent).— a) and rank(A .3/)£ = 0.) = (A.nxn).. A e nxn ]R. the associated number of linearly A. we need the notion of principal vector. To get a third vector X3 such that X = [Xl KJ_ X3] reduces A to JCF.nxn number of eigenvectors.28.7) = n . suppose suppose A = [3 2 0 o Then Then 3 0 A3I= U2 I] o o 0 0 n has rank 1. Determination of the JCF 9. eigenvectors dimN(A — A. is of algebraic multiplicity greater than one.) = (A.27. X principal Definition 9.e. a)\ . Thus. Determination of the JCF 85 &5 Example 9. The more interesting (and difficult) case occurs when Ai multiplicity A. 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.A.a(A) = (A . For example..is simple. The matrices A uniquely. An analogous definition holds for a left principal vector of degree k. a)7. i. a(A. of algebraic multiplicity 1.9. and rank (A A. so the eigenvalue 3 has two eigenvectors associated with it. If we let [~l ~2 ~3]T associated If [^i £2 &]T denote a solution to the linear system (A — 3l)~ = 0. To get a third vector JC3 such that X [x\ X2 XT.rank(A .. of course. we find that 2£2 + ~3 = O. and rank(A al) vectors.] are eigenvectors (and are independent). The straightforward case is.. three eigen7r(A. it then has precisely one eigenvector. For each distinct eigenvalue Ai. i. Definition 9. we find that 2~2 + £ 3= 0 .ulx = 0 and (A . a (A). both denote a solution to the linear system (A .27.— a) . Then x is a right principal vector of degree k degree associated with A E A (A) ifand only if(A . c .29.3 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.).. when X. three eigenboth have rr(A) = (A . Thus. is not sufficient to Example 9. i. determine the JCF of A uniquely. and rank(A —Ai l) for distinct Ai is not sufficient to rr(A).e. Knowing TT (A. Remark 9..l) independent right — — A.. i.AI)klx i= o. associated independent right (or left) eigenvectors is given by dim A^(A . of algebraic multiplicity 1.29.3. Let A E C"xn (or R"x"). it The straightforward case is. Remark 9.7) for distinct A..— al) == 4.A.28.e.3.(7).e. 9.. of course. when Ai is simple. left k.
for get righthand example. x(l). there is only one eigenvector. solutions solutions to the homogeneous equation (A . The number of eigenvectors depends on the rank of A .A1)X(l) = O. The phrase "of grade k" is often used synonymously with "of degree k. See. which simply says that x(!) is a right Ax(1) = hx(1) x (1) (2) x(2). principal vectors of degree 1) associated with A. k = eigenvector. E lR nxn This suggests a "general" procedure. solve (A .1 Theoretical computation Theoretical computation To motivate the development of a procedure for determining principal vectors.86 Chapter 9. Then the equation AX = XJ can be written that reduces a matrix A to this JCF. of k 5. (A — uf.3. by (A .e.x2 A JCF. One of these solutions (A — AI)2 x (2) x(l) (1= 0). Then for each distinct X e A (A) perform the following: z (2) w c 1.XI. Solve (A .AI). (9.XI)2x^ = 0. if X. If we premultiply (9. Denote by x(l) and x(2) the two columns of a matrix X E lR~X2 2x2 2 Jordan block{h0 h1..1.1 9. For each independent jc (1) . Thus. Principal vectors are sometimes also called generalized eigenvectors. of course. A E A(A) following: (or C ). ji of dimension k or larger.XI).X I ) ( l ) = (A AI)O o. Theother solution (A . 2. wefind(A. A right (or left) principal vector of degree k is associated with a Jordan block J. this rank is n . Denote by x(1) and x(2) the two columns of a matrix X e R2. is. eigenvector.X I ) . (It may be necessary to take a linear of x(l) R(A . I) associated This step finds all the eigenvectors (i. See. The case k = 1 corresponds to the "usual" eigenvector. 9.3. since (A .A1)X(l) = O.A/) = — multiplicity of rank(A — XI) = n . First. 4. but the latter generalized eigenvectors. Eigenvalues and Eigenvectors synonymously "of 2. combination of jc(1) vectors to get a righthand side that is in 7£(A — XI). determine all eigenvalues of A e R" x " nxn ). Exercise 7.A/)x(2) = x(l). principal vectors still need to be computed from succeeding steps.17) The first column yields the equation Ax(!) = AX(!). Thus. of The number of linearly independent solutions at this step depends on the rank of 2 (A .A1)2 x(2) = (A . The second column yields the following equation for x . For example.AI). S. the principal vector second of degree 2: of degree (A .'A1)22xx(l) = (A .XI)0 = 0. If." "of often 3. consider a determining 2 x Jordan [~ i]. if of . different term will be assigned a much different meaning in Chapter 12. x(l) (^ 0). there are two linearly independent n — o.A1)x(2) = x(l).17) by (A . The other solution necessary is the desired principal vector of degree 2. Eigenvalues and Eigenvectors Chapter 9.2. for example. If the algebraic multiplicity of If A principal need X is greater than its geometric multiplicity. we find (A If we premultiply XI) x = (A XI)x = 0. Then the equation AX = X J can be written A [x(l) x(2)] = [x(l) X(2)] [~ ~ J. the definition of principal vector is satisfied. of — AI.) .
Let A=[~ 0 2 ] . Let X = x ( l ) . . vectors is equal to the algebraic multiplicity of A. There are significant numerical difficulties inherent in attempting to compute a JCF. . Attempts to do such calculations in finiteprecision floatingpoint arithmetic or 3. and A3 = 2. 0 The eigenvalues of A are AI = 1. 4.31. Principal vectors associated with different Jordan blocks are linearly indeTheorem 9. Notice that highquality mathematical software such as MATLAB readable [8] to learn why.2/)x3(1)= 0 yields (A . where the chain of suppose further that rank(A . see. although a j ardan command is available in MATLAB'S does not offer a jcf command. and the interested student is strongly urged to consult the classical and very readable [8] to learn why. . Let Example 9.. A2 = 1. Notice that highquality mathematical software such as MATLAB does not offer a j cf command. . for example.32.2I)x~1) = 0 yields . for example. For each independent x(2) from step 2. [20] and [21]. For each independent X(2) from step 2.3. h2 = 1. Symbolic Toolbox..1. with the distinct eigenvalues 1 and 2. First. Then Theorem 9. say). Determination of the JCF 3. this naturallooking procedure can fail to find all Jordan vectors.30.AI) = k . Theorem 9. Theorem 9. 1 . Continue in this way until the total number of independent eigenvectors and principal 4. [20] and [21]. . 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. . Continue in this way until the total number of independent eigenvectors and principal vectors is equal to the algebraic multiplicity of A. . solve 3. find the eigenvectors associated with the distinct eigenvalues 1 and 2. x(k)]. X(k)} is a linearly independent set. (x (1) . say). Suppose A E C kxk has an eigenvalue A of algebraic multiplicity kkand suppose further that rank(A — AI) = k — 1. pendent. Example 9.. and the interested student is strongly urged to consult the classical and very to compute a JCF. Determination of the JCF 9. see. . Unfortunately. this naturallooking procedure can fail to find all Jordan vectors. Then vectors x(i) is constructed as above. Attempts to do such calculations in finiteprecision floatingpoint arithmetic generally prove unreliable.3..9.(1) (A . Principal vectors associated with different Jordan blocks are linearly independent.. Let X = [[x(l).30. and h3 = 2. Theorem 9. solve (A AI)x(3) 87 = x(2).31. For more extensive treatments. For Unfortunately. x (k) } is a linearly independent set. of algebraic multiplicity and Theorem 9.32. find the eigenvectors associated The eigenvalues of A are A1 = I. X(k)]. First. . Determination of eigenvectors more extensive treatments. where the chain of vectors x(i) is constructed as above. There are significant numerical difficulties inherent in attempting generally prove unreliable. although a jordan command is available in MATLAB's Symbolic Toolbox..33.33.. . {x(l). Suppose A e Ckxk has an eigenvalue A.
we 1 's but can be arbitrary — so long as they are nonzero. solve To find a principal vector of degree 2 associated with the multiple eigenvalue 1.2 9... we consider below the case of a single Jordan block. . Then A 4l. d.. solve 2 (A .(2) = x. 0 = 0 A dn dn I 2 0 dn dn I A 0 ). but the result clearly holds for any JCF.l/)x. . 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. (1) toeet x. d. .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 . but the result clearly holds for any JCF. Eigenvalues and Eigenvectors To find a principal vector of degree 2 associated with the multiple eigenvalue 1.. consider below the case of a single Jordan block.1I)xl ) = xiI) to get (A – l/)x. ..3.88 (A . .. 0 !b.so long as they are nonzero.11)x?J = 0 yields (A. Then Let D = diag(d" . d n)) be a nonsingular "scaling" matrix.3. =0 yields (1) Chapter 9. 0 0 D'(X' AX)D = D' J D = j ). Now let Now let X (2) =[ 0 ] ~ .. Suppose A € Rnxn and SupposedA E jRnxn and Let D diag(d1. dn be a nonsingular "scaling" matrix. For the sake of defmiteness. For the sake of definiteness.
A1I) v) E6 ... Note that dimM(A .A[)n) . Theorem 9. J is obtained from A via the similarity transformation XD = \d\x\. the reverseorder identity matrix (or exchange matrix) 0 p = pT = p[ = 0 I 0 (9... .. set {As s E S}.... where AS is defined as the transformation.A[)V) '" (A ./) w = «. A.. Suppose e jH.Amtm c and minimal polynomial a(A) = (A . Let V be a vector space over F and suppose A : V —>• V is a linear Definition 9.. Such a decomposition is given in the following theorem. . A subspace S ~ V is Ainvariant if AS ~ S. j is obtained from A via the and principal vectors that reduces A to its JCF. Specifically.35. interpreted This result can also be interpreted in terms of the matrix X = [x\. A subspace S c V is A invariant if AS c S..4 9.. In a similar fashion. Suppose A E R"x" has characteristic polynomial 9. .34. Such a decomposition is given in the following associated direct sum decomposition of jH.. Geometric Aspects of the JCF 9.nxn n(A) = (A .Am)Vm with Ai. E6 N (A .4.. Definition 9. the reverseorder identity matrix (or exchange matrix) In a similar fashion. dnxn}... . .A.AmItm . .n = N(A = N (A . dnxn].Am I) Vm ... Let IF and suppose + transformation.35.. (A . Specifically.AlIt) E6 . Then AI. similarity transformation XD [d[x[. mdistinct. E6 N(A .. Then jH.. Am distinct.nxn (or nxn to JCF provides change of basis with respect to which the matrix is diagonal or block diagonal.. where AS is defined as the set {As:: s e S}.4.n. dimN(A — AJ)Vi = ni..xn]] of eigenvectors = [x[. .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. . It is thus natural to expect an associated direct sum decomposition of R. x n eigenvectors and principal vectors that reduces A to its lCF...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 9.34. Geometric Aspects of the JCF 89 di's Appropriate choice of the di 's then yields any desired nonzero superdiagonal elements.9. It is thus natural to expect an with respect to which the matrix is diagonal or block diagonal.
19): /th Example 9. i.) and each /. Equivalently. Let peA) = «o/ + o?i A + • + <xqA be a polynomial in A. Eigenvalues and Eigenvectors Chapter 9. e E"x".39. diagonal representation.. where X [ X i .. we have that AXi Theorem 9.. where each Theorem 9.. for N(A ./)"' by SVD. R(S) == S.19) the columns of Xi (i./)"'. If V is a vector space over IF such that V = N\ EB . Then N(p(A)) and R(p(A)) 7£(p(A)) are Ainvariant. then S <S is Ainvariant if and only if there exists M E ]Rkxk such that eRkxk (9.i. Let Yi E <enxn . A". Let 7.. such "canonical" forms are discussed in text that follows. .2.e. . We would then get a block diagonal representation for A with full blocks rather than the highly structured Jordan blocks. Other representation for A with full blocks rather than the highly structured Jordan blocks. partition Equivalently. each Ji = diag(JiI.. so by (9. but we restrict our attention here to only the Jordan block case... If A has distinct The Jordan canonical form is a special case of the above theorem. then a basis for V can be chosen with respect to which A has a block diagonal representation. Ainvariant. Jm). A invariant if only ifS1 1..34.2. 7. then a basis for V can be chosen with respect to which A has a block N. The Jordan canonical form is a special case of the above theorem. we could choose bases for N(A — A.. . . Sk If R" R. //*. Example 9. so by (9.e. Xm R"n such that XI AX diag(7i..38.. example (note that the power n.37.lt. We could also use other block diagonal decompositions (e.li.. Eigenvalues and Eigenvectors If V is taken to be ]Rn over Rand S E ]Rn x* is a matrix whose columns SI. 2.37. eigenvalues Ai 9. Then N(p(A)) and 1. . is Ainvariant... Note that A A". /. and S e R" xk s\.36.. S is Ainvariant if and only if S .Ji .. . .) span an Ainvariant of A".as in Theorem 9. /.g.38. Suppose A"== [Xl . so the columns of A. Note that AXi = A*.34.19) the columns attention here to only the Jordan block case..)... i /= 1. XI AX = [~ J 2 ].) and each Jik is a Jordan block corresponding to A. .e. is not necessarily diagonalizable.90 Chapter 9.. Let p(A) = CloI + ClIA + '"• •+ ClqAqq be a polynomial in A. the eigenvectors and principal vectors associated with Ai) span an Ainvariant subspace of]Rn.. s/t span a /^dimensional subspace <S..36. = 1. via SVD). Suppose X block diagonalizes A.e.* is a Jordan block corresponding to Ai E A(A).e. be a Jordan basis for N (AT . could be replaced by v. of W. 9. If A has distinct eigenvalues A. € C"x"' be a Jordan basis for N(AT — A..• EB Nm. This follows easily by comparing the ith columns of each side of (9.. Jm). Finally. K(S) <S. The equation Ax Example 9.is A T invariant. i. so the columns of Xi span an Amvanant subspace. We would then get a block diagonal example (note that the power ni could be replaced by Vi).. = Xi... Suppose A E ]Rnxn. (i.."" Jik. e A(A). the eigenvectors and principal vectors associated with A. Other such "canonical" forms are discussed in text that follows. i. If F = NI ® • • 0 m A// is Ainvariant. 9..19) AS = SM.. . Rewriting in the form ~ J...Xm] ] Ee]R~xnxnisis such that X^AX ==diag(J1. AT Theorem 9.Ai/)n.39. = X. 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.A..span an Ainvariant subspace. 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). we have that A A. partition ... then is Ainvariant if and only if there span a kdimensional subspace S. where each Ji = diag(/.
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.. A called the matrix sign function.. . The Matrix Sign Function 9. Let z E C with Re(z) ^ O.I = XJy H = [XI.. 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. is given by sgn(A) = X [ / 0] 0 / X I .. of defined Definition 9. Suppose A E C"x" has no eigenvalues on the imaginary axis. i=1 H In a similar fashion we can compute m etA = LXietJ. . Jm) [YI . 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 . denoted sgn(A).YiH.5. is given by eigenvalues in the right halfplane. 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]. associated with an eigenvalue A.= Ai. . Then A = XJX. . .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. m ••• . for a k x k Jordan block 7. 9. It is a generalization of the sign (or signum) of a scalar. Then compatibly. Xm] diag(JI. Then the sign of A. and let e cnxn be a Jordan canonical form for A.5 9.40. Definition 9. The Matrix Sign Function 91 91 compatibly.41. Ym]H = LX.. Then the sign of A.S.41.40. 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. denoted sgn(A).JiYi . E f= 0.9. Definition 9..
. projection subspace of 4. sgn(AH) = (sgn(A))". In fact. negA == (l .42. sgn(A") = (sgn(A»H.. 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). respectively.. S = sgn(A). sgn(TlAT) = T1sgn(A)T foralinonsingularT E C"x". Yn. R(S — /) Ainvariant of (the negative invariant subspace). positive = (/ + of A. and let = sgn(A). Find the appropriate expression for v as a linear combination expression of the left eigenvectors as well. Let A E Cnxn have distinct eigenvalues AI.. 2.. respectively. 5. 3. The JCF definition of the matrix sign function does not generally lend itself to reliable computation on a finitewordgenerally itself length digital computer. 3. sgn(T1AT) Tlsgn(A)TforallnonsingularT e enxn 6. but the one given here is especially useful in deriving many of its key properties. AS = SA. posA == (l + S)/2 is a projection onto the positive invariant subspace of A. S2 = I. respectively. S = sgn(A). .S) /2 is a projection onto the negative invariant subspace of A. S2 = I.. Let e C" be an arbitrary vector. 5. 6.42. negA = (/ — S)/2 3. Let v E en be an vectors Xl. Suppose A E C"x" has no eigenvalues on the imaginary axis. e nxn Theorem 9. Xn and left eigenvectors y\. sgn(cA) = sgn(c) sgn(A) for all nonzero real scalars c.43. 2. . Then the following hold: following 1. There are other equivalent definitions of the matrix sign function. sgn(cA) = sgn(c) sgn(A)/or c.. Show that v can be expressed (uniquely) as a linear combination e of the right eigenvectors.n 1. positive of P. 4. Xn with corresponding right eigenA e nxn ). We state some of the more useful properties of the matrix sign function as theorems.1> . yn. ••• .. of A (the negative invariant subspace). ± 1. . ••. S is diagonalizable with eigenvalues equal to del.92 92 Chapter 9.xn and left eigenvectors Yl. Their straightforward proofs are left to the exercises. . but the one given There are other equivalent definitions of the matrix sign function. The JCF definition of the here is especially useful in deriving many of its key properties.. EXERCISES EXERCISES 1.. and let — sgn(A). Their left exercises. . 7l(S l) is an Ainvariant subspace corresponding to the left halfplane eigenvalues left halfplane I. We state some of the more useful properties of the matrix sign function as theorems. .. its reliable numerical calculation is an interesting topic in calculation its own right. 2. distinct right eigenvectors Xi. ). Show that v can be expressed (uniquely) as a linear combination arbitrary vector. Theorem 9.. Theorem 9. Eigenvalues and Eigenvectors where the negative and positive identity matrices are of the same dimensions as N and p. AS = SA. Suppose A E enxn has no eigenvalues on the imaginary axis. 3. 4. Then the following hold: following e 1...43. e C"x" Theorem 9. Eigenvalues and Eigenvectors Chapter 9.
k . y e R" are nonzero vectors with A E lR.22. 11. eigenvectors and if and (real) JCFs of the following matrices: (a) 2 1 ] 0 ' [ 1 6. Determine the JCF of A. Determine the JCF of A. Let A e R" xn be of the form A = 1+ xyT. The vectors [0 1 Ifand[l 0 of [0 — l] r and[1 0]r (2) (1) are both eigenvectors. a skewHermitian matrix must be pure imaginary. y E lR. n are nonzero vectors with with xTTyy = 0.1) element of J. = O. Prove that all eigenvalues of a skewHermitian matrix must be pure imaginary. Determine the JCF of A. where x. eigenvalues. What are the eigenvalues of this slightly perturbed matrix? matrix? . What are the eigenvalues of this slightly perturbed is added to the (16. ~ 0 Hint: Use[1 1 — I]T an Hint: Use[— 1 1 . where J is the JCF Find a nonsingular matrix X such that X AX = J.30 must be multiples of el e R*. Let A = [H 1]· 2 2" Find a nonsingular matrix X such that XI AX = J. 10. Suppose A E C"x" is Hermitian. Let 7. 16x 16 has eigenvalues at 0 its JCF consists of single Jordan block of the form specified in Theorem 9. x. Show that x is also a left eigenvector for A. x O.1) element of J. Suppose A € rc nxn is skewHermitian. Prove that all eigenvalues of 2. i.5x5 has eigenvalues {2. Suppose a matrix A E lR.n T T x yy = 0. Let A be an eigenvalue of A with corresponding 3. nxn be of the form A = / + xyT. 4. AH = A. where J is the JCF 1 J=[~ 0 1~]. right eigenvectors and right principal vectors if necessary. Suppose the small number 10. y e R" are nonzero vectors 10. Let A E lR. JCFs for A. 3}. where x. Let A be an eigenvalue of A with corresponding right eigenvector x.. Show that all right eigenvectors of the Jordan block matrix in Theorem 9. multiples of e\ E lR. i. 3./)jc = x can't be solved. Determine the JCFs of the following matrices: 6. 5.l]r as an eigenvector. 9. Prove the same result if A is skewHermitian. 2. Determine the JCFs of the following matrices: <a) Uj n 2 1 2 =n 7.e. but then the equation (A . Characterize all left eigenvectors. JCFs for A. 5.Exercises 93 93 2. Let A e R"x" be of the form A = xyT. Suppose A e rc nxn is Hermitian. if A is skewHermitian. (A — I)x(2) x(1) 8. where x. Determine the JCF of A. Determine the eigenvalues. 2. Show that all right eigenvectors of the Jordan block matrix in Theorem 9. Determine all possible € R 5x5 {2. Show that x is also a left eigenvector for A.30 must be 8. y E lR. AH = —A.e. Characterize all left eigenvectors. Suppose A E C"x" is skewHermitian. Prove the same result right eigenvector x.22. 3}. Suppose a matrix A E R 16x 16 has 16 eigenvalues at 0 and its JCF consists of a single A e lR. Suppose 10~16 is added to the (16.16 Jordan form specified 9..nxn A = xyT. 2.
say Si. and S2 are real symmetric matrices and one of them. i. X e R*x <«*).18) is useful.S2X~l) would required symmetric factorization of A. about when the equation for X is what can you say further.. Hint: Use the factorization in the previous exercise. 16. sgn(A) = /. Suppose A e sgn(A) = 1. where SI and £2 are real symmetric matrices and one of them. Thus. 16. say S1. is nonsingular. JCF. where Si 12. 14. Show that every matrix A e jRnxn can be factored in the form A Si$2. Prove that every matrix e jRn xn is similar to its transpose and determine a similarity transformation explicitly. Eigenvalues and Eigenvectors 12. Then = ( X SIXT)(X. what can you say further. xn has all its eigenvalues in the left halfplane. Prove that every matrix A E W x" is similar to its transpose and determine a similarity 13. Prove that 17.e. 15. Prove Theorem 9. in terms of AU and A 22. Prove that 17. TIAT = [A011 A22 0 ] . If n = 2 and k = 1. The transformation P in (9. Consider the block upper triangular matrix 14. Prove Theorem 9. If n = 2 and k = 1. Then A = (XS i X T ) ( X ~ T T S2XI) would be the the "symmetric factorization" of J. Suppose Au =1= 0 and that we we e jRnxn and All e jRkxk 1 < ::s: n. about when the equation for is solvable? solvable? 15. 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. Suppose Al2 ^ and want to block diagonalize A via the similarity transformation want to block diagonalize A via the similarity transformation where X E IRkx(nk). it suffices to prove the result for the required symmetric factorization of A. Eigenvalues and Eigenvectors Chapter 9. Find a matrix equation that X must satisfy for this to be possible.42.18) 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. in terms of All and A22.42. Prove Theorem 9. en . is nonsingular. Find a matrix equation that X must satisfy for this to be possible. it suffices to prove the result for the JCF.94 Chapter 9. Show that every matrix A E R"x" can be factored in the form A = SIS2.43. The transformation P in (9. transformation explicitly. Thus. Hint: Use the factorization in the previous exercise. Prove Theorem 9.43. Suppose A E C"xn has all its eigenvalues in the left halfplane. 13.
where D = diag(A.. matrix if and only if it is normal (i. AAHH = AH A). orthonormal eigenvectors for A). Xn) (the columns ofX are orthonormal eigenvectors for A). find P e lR. then there exists a unitary matrix U such that UH AU — D. if A e Rmxn . most "diagonal" we can get is the JCF described in Chapter 9. where it is proved that a general matrix A E C"x" is unitarily similar to a diagonal 10. We can also consider the case A E emxn and unitary equivalence if P and <2 are unitary..1. if A E IR mxn find E R™ and Q E lR~xn such that P AQ has a form..I. Xn Theorem 10. The following results are typical of what can be achieved under a unitary similarity.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. . An. 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." The transformation A M» PAQ is called an equivalence..9. and orthogonal. Then there AI. . An). . the transformation A f+ P ApT is called 2.:xm and Q e Rnnxn such that PAQ has a "canonical form. = V and if pT is orthogonal. If P"1 . where D = diag(AJ." In matrix terms. We can also consider the case A e Cm xn and unitary equivalence if P and Remark 10. What other matrices are "diagonalizable" under unitary similarity? The answer is given in Theorem matrices are "diagonalizable" under unitary similarity? The answer is given in Theorem 10. such as A = [ _ab ^1 for real scalars a and b. If W = V and if Q = PT is orthogonal. If The following results are typical of what can be achieved under a unitary similarity.1." The transformation A f+ P AQ is called an equivalence. . it is called an "canonical form.j. V and Q 1.. .. the transformation A i» PAPT is called If an orthogonal similarity (or unitary similarity in the complex case).. as well as other matrices that merely satisfy the definition.2. Normal matrices include Hermitian.. respectively). . 95 95 ..1. What other U HAU = D. Let A = AH e C"x" have (real) eigenvalues A. and orthogonal. This is proved in Theorem 10.2. . Normal matrices include Hermitian.Chapter 10 Chapter 10 Canonical Forms Canonical Forms 10. . such as A = [_~ most "diagonal" we can get is the JCF described in Chapter 9. . An. A. an orthogonal similarity (or unitary similarity in the complex case). it is called an orthogonal equivalence if P and Q are orthogonal matrices. ..j.. Q are unitary. the for real scalars a and h. n ). . then there exists a unitary matrix £7 such that A = AH E en xxn has eigenvalues AI. . as well as other matrices that merely satisfy the symmetric. This is proved in Theorem 10. !] Theorem 10. skewskewHermitian. . and unitary matrices (and their "real" counterparts: symmetric... . skewHermitian.9. . Remark 10.. If a matrix A is not normal. Two special cases are of interest: Two special cases are of interest: 1. .1 10." In matrix terms.. the transformation A H> PAP" 1 is called similarity.. the definition. respectively). If W = V and <2== p.. = H E en xn exists a unitary matrix X such that X H AX = D = diag(Al.An.e. and unitary matrices (and their "real" counterparts: symmetric. Problem: Let V and W be vector spaces and suppose A : V —>• W is a linear transformation.2. the transformation A f+ PAPI is called aasimilarity..2.. An) (the columns of X are exists a unitary matrix X such that XHAX = D = diag(A. AA = AHA). skewsymmetric. . orthogonal equivalence if P and are orthogonal matrices. If A = A H 6 C" " has eigenvalues AI. 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.e.. If a matrix A is not normal.
Canonical Forms Proof: eigenvector corresponding AI. simplicity. ... [£i. Write V H matrix such that V X I = [ ~]. k = For simplicity. HdxI. (/ € k) U2 X i U2 = Xi . we consider the real case... We also get 0 in the (2. the construction of X2 E JRnx(nl) such that X — z e ]R" (". X 1 XI e E"...2) Al X~AX2 XfAX 2 0 Al ] 0 XfAX z 0 l In (10. xn such that X = (XI. Xk]..1) we have used the fact that AXI = k\x\. We illustrate the construction of the necessary Householder matrix for k — 1.. Then VH = / / . I)block x"xi = 1. . Now XHAX =[ xH I XH ] A [XI 2 X 2] =[ =[ =[ x~Axl XfAxl X~AX2 XfAX 2 ] (10. . where R € Ckxk is upper triangular. The proof is completed easily by induction upon noting proof that the (2. ~nf.....l)block by Al (2..I)block.. . . .. and normalize it such that x~ XI = XI 1. Then there exist n . Hk in the usual way (see below) such that Hk . . Then there exist n — 1 additional vectors X2. XH AX induction noting that XH AX is Hermitian. 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.. [Xi f/2] unitary.xd... . . orthogonal (l. . Let X\ e Cnxk have orthonormal columns and suppose U is a unitary Theorem 10.. where R E kxk is upper triangular.. . Construct a sequence of Householder matrices (also known HI. .• • Hk and Hk'" HI.3 called Theorem 10.2)block X2 .1) we have used the fact that Ax\ = AIXI. xn] = 1.2)block must have eigenvalues A2..3.2). Construct a sequence of Householder matrices (also known Proof: Let X [XI. . Hk as elementary reflectors) H\. An. When combined with the fact that x~ XI = 1. Canonical Forms Chapter 10... Thus. 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.2)block by XI Xz..l)block... .1 additional vectors x2. we get 0 in the (l. . .. When combined with the fact that In (l0. xf*x\ = Proof' Let x\ be a right eigenvector corresponding to X\. Let V = XI. .. xn] = [x\ ] [XI X22] is unitary. A...k rows of the unitary matrix U..3 for k = 1.... D The construction called for in Theorem 10. —k U. X = Given a unit vector x\ E JRn.Hv.. n . Xn such that [x\. VI € Cnxk [Xi U ] Proof: Let X\I = [x\. k 1. D 0 (2.T. ..1) (10. . .2)block noting that x\ is orthogonal to all vectors in X2. Write UH = [U\ U ] [VI Vz] 0 2 with Ui E Cnxk . .96 96 Chapter 10.3. Xk H Hk.2 is then a special case of Theorem 10.. Xk are orthonormal). [XI U2] is unitary... . following general result. In (10.. Let XI E Cnxk have orthonormal columns and suppose V is a unitary matrix such that UX\ = \ 1... xd = [ ~ l U = where R is upper triangular (and nonsingular since x\. . 0 Thus.k But the latter are orthonormal since they are the last n . . 10. %n] XI . (l.) [XI X2]] is orthogonal is frequently required. we get Ai remaining in the (l. Let the unit vector x\ be denoted by [~I. Then [XI V 2] is unitary.
3) spectral which is often called the spectral representation of A.1 ± 1. where u = ['.1 and UT X\ = 1 ± £1. . [23]. it is easily verified that u T u = ± 2£i and u T Xl = 1 ± '.4. (onto the onedimensional eigenspaces correPi onedimensional eigenspaces sponding to the A. where Pi = PR(x. Some Basic Canonical Forms 10.1. consulted standard numerical linear algebra can be consulted in standard numerical linear algebra texts such as [7]. 's)..e.2. In fact...e. n A = LAiPi. . Thus. i. = PUM = xixf = xxixT since xj xi — 1. sponding to the Ai'S).. U effects necessary compression of jci. so U is orthogonal. it is easily verified that UT U = 2 ± 2'. [11]. 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.•» '. 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. .It can easily be checked — 2uu+ — u u T . U orthogonal.4. An. x where P. A in (10.2uu+ = I . . . A Note that Theorem 10. .) — xiXt = i j since xT Xi = 1. '. . X n ). Let A = AT e E nxn have eigenvalues k\. quently in applications.Xn.1.i. [7]. .. . The real version of Theorem 10.. An). . [23]. £2. .2 is worth stating separately since it is applied frequently in applications.2 for Hermitian matrices) can be written from Theorem 10.2 for Hermitian matrices) can be written n A = XDX T = LAiXiXT.10. [25].3) is actually a often weighted sum of orthogonal projections P. XTAX = D = diag(Xi.+uu T .. 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 . Theorem 10..1. • • . including the choice of sign and the complex case. where u ^UU [t\ 1.. necessary compression of Xl.2 is worth stating separately since it is applied fre10.nf. i=1 (10. [25]. [11]. Then there exists an 10. Then there exists an AT E jRnxn have eigenvalues AI.4 implies that a symmetric matrix A (with the obvious analogue from Theorem 10.. 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.. Further details on Householder matrices..
The quasiuppertriangular matrix S in Theorem 10. complex conjugate pairs of eigenvalues. The triangular matrix T in Theorem 10. matrix U that reduces a matrix to [real] Schur form are called Schur vectors. Then AAH = U VUHU VHU H = U DDHU H == U DH DU H == AH A so A is normal. Let A e C"x".2 except that in this case (using the notation U rather than X) the (l. for example. and sufficient for virtually all applications (see. where S is quasiuppertriangular. real arithmetic) to a quasiuppertriangular A e Wnxn is also orthogonally similar (i.8.6 (MurnaghanWintner). The quasiuppertriangular matrix S in Theorem 10.2)block AU2 is not O.98 98 Chapter 10. is that the first Schur vectors span the same all applications (see. Let A E R"xxn.7. The when we can go further and reduce a matrix via unitary similarity to diagonal form. [17]). what is true.6 T T matrix U such that U AU = S.8.5 is called a Schur canonical form or Schur form.e. Example 10. Its real JCF is h[ 1 1 1 0 0 n n Note that only the first Schur vector (and then only if the corresponding first eigenvalue Note that only the first Schur vector (and then only if the corresponding first eigenvalue is real if U is orthogonal) is an eigenvector. The columns of a unitary [orthogonal] matrix U that reduces a matrix to [real} Schur fonn are called Schur vectors.5 (Schur). the next theorem shows that every A E IR xn is also orthogonally similar (i. The triangular matrix T in Theorem 10.. The matrix 10. 0 in this case (using the notation U rather than X) the (l. [17]). However. Its real JCF is is in RSF. . A matrix A e c nxn is unitarily similar to a diagonal matrix if and only if A is normal (i. A matrix A E C"x" is unitarily similar to a diagonal matrix if and only if Theorem 10. but if A has a complex conjugate pair of eigenvalues. Proof: Suppose U is a unitary matrix such that U H AU = D.5 is called a Schur canonical Definition 10. A is normal (i.. Then there exists an orthogonal Let A e IR n ". However. Let A E cnxn Then there exists a unitary matrix U such that Theorem 10. matrix U such that U AU = S. then complex arithmetic is clearly needed to place such eigenValues on the diagonal of T. 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. it is of interest to know when we can go further and reduce a matrix via unitary similarity to diagonal form. Theorem 10.6 is called a real form or Schur fonn. UH AU = T.9. it is thus unitarily similar to an upper triangular matrix. it is thus unitarily similar to an upper triangular matrix. it is of interest to know While every matrix can be reduced to Schur form (or RSF). The columns of a unitary [orthogonal} Schur canonical form or real Schur fonn (RSF). The matrix s~ [ 2 0 2 5 4 0 is in RSF. where S is quasiuppertriangular.e. following theorem answers this question. the next theorem shows that every to place such eigenvalues on the diagonal of T. but In the case of A E R"xxn . While every matrix can be reduced to Schur form (or RSF). Definition 10. where T is upper triangular. and sufficient for virtually is real if U is orthogonal) is an eigenvector. what is true. Canonical Forms Theorem 10. D ur In the case of A e IRn ". Proof: The proof of this theorem is essentially the same as that of Theorem lO. However.e.7. where D is diagonal. 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). Theorem 10..2 except that Proof: The proof of this theorem is essentially the same as that of Theorem 10. AHA = AA H). Then Proof: Suppose U is a unitary matrix such that U H AU = D. then complex arithmetic is clearly needed if A has a complex conjugate pair of eigenvalues. where D is diagonal. real arithmetic) to a quasiuppertriangular matrix.e.9.. where T is upper triangular. for example.5 (Schur). diagonal of T (or S). Then there exists a unitary matrix U such that U H AU = T. However. . so A is normal.6 is called a real Schur canonical form or real Schur form (RSF). AH A = AAH ).2)block wf AU2 is not 0. The following theorem answers this question. Then there exists an orthogonal 10. Canonical Forms Chapter 10. A quasiuppertriangular matrix is block upper triangular with 1 x 1 diagonal matrix.
. Similarly. negative positive definite. 3. Proof: Proof: Let U be a unitary matrix that diagonalizes A as in Theorem 10.10. If a matrix is neither definite nor semidefinite. Then n x HAx = (U HX)H U H AU(U Hx) = yH Dy = LA. Definite Matrices 10. we write A :::: B if and only ifA — B>QorB — A < 0.2 10. this section that may be stated in the real case for simplicity.nxn is Definition 10.12 ~ AlyH Y = AIX HX . nonpositive definite (or negative semidefinite) if A is nonnegative definite.2 Definite Matrices Definite Matrices Definition 10. x eC". and denote the components of y by v UHx... it is said to be indefinite. be diagonal.10. we write A > B if and only if A .2. Let A = AH e enxn with eigenvalues X{ :::: A2 :::: . B — A < 0. U diagonalizes A 10. Definite Matrices 99 Conversely.B > 0 or or Also. Similarly. A symmetric matrix A e Wxn 1.12. We write A > 0. e Theorem 10. 2.A ~ O. if—A 4.n. suppose A is normal and let U be a unitary matrix such that U H AU = T. if A and B are symmetric matrices.2. A symmetric matrix A E lR.• :::: An. in fact. It T 0 D 10. Remark 10. Indeed. let y = U H x.2. where T is an upper triangular matrix (Theorem 10.10. superscript H s replace T s.5). positive definite if and only if xTT Ax > 0 for all nonzero x G lR. Remark 10.A is positive definite. Thenfor all Let A = AH E Cnxn with eigenvalues AI > A2 > • • > An.12. We write A < O. positive definite if and only ifx Ax > Qfor all nonzero x E W1 We write A > O. Furthermore. we write A > B if and only if A — B > B . If neither semidefinite.B :::: 0 or B . indefinite. this is generally true for all results in the remainder of of superscript //s Ts.n • We write A :::: 0. Furthermore. We write A < 0.=1 But clearly n LA. 111. i € n.11.2. We write A > O. We (or negative if— A nonnegative definite.. negative definite if . 11'/.. all the above definitions hold except that A e nxn Remark 10. nonnegative definite (or positive semidefinite) if and only if XT Ax :::: 0 for all (or positive if and only if x T Ax > for all nonzero x e W. Then for all E en. we write A > B if and only if A . i En. We write A ~ 0. . nonzero x E lR. Then 11. A U U HA U T. write A < O.11.=1 . CM j]i.13.A < O. If A E C"x" is Hermitian. where x is an arbitrary vector in en. Then T (Theorem It is then a routine exercise to show that T must. if A and B are symmetric matrices. Also.
Theorem 10. A leading principal submatrix of order n — k is obtained by deleting the last k rows and columns. The determinant of the 1x1 1 leading submatrix is 0 and 1. 3.1. whence Ar1ax (A A).19. ::::: AI.18. 2.1.@mllrk 10. The determinants of all leading principal submatrices of A are positive. 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. Theorem 1O. . 3.18. All eigenvalues of A are nonnegative. Let A e C"x". where M e R"x" is nonsingular.soO < X n < ••• < A.. The determinant of the I x leading submatrix is 0 and consider the matrix A = [~ 2x 0 (cf. Theorem 10. determinant the determinant of the 2x2 2 leading submatrix is also 0 (cf. Canonical Forms and and n LAillJilZ::: i=l AnyHy = An xHx . Canonical Forms Chapter 10.15. XHAx > 0 for all nonzero = AH E enxn E en. For example. 3. Then ^pjp2 = ^^(A" HA). Then 111~~1~2 Let jc be an eigenvector corresponding to Xmax(AHA). Theorem 10. of obtained and E ~nxn positive definite if and only if any of the Theorem 10.17). so 0 An ::::: . Remark 10. not just those of the leading principal submatrices. Then IIAII2 = ^m(AH A}. Corollary Corollary 10. 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. where M E IRb<n and k ~ ranlc(A) — ranlc(M). 0 D Remark XHHAx Remark 10. form MT E ~n xn E ~n xn definite if and only if Theorem 10. 2. A can be written in the form MT M.w) x HAx > the Rayleigh quotient. I Proof: E C" Proof: For all x € en we have Let x be an eigenvector corresponding to Amax (A HA). 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..18. Let A E enxn Then \\A\\2 = Ar1ax(AH A).I.19. Theorem 10. the . Note that the determinants of all principal "ubm!ltriC[!!l mu"t bB nonnBgmivB R.16. of positive. Not@th!ltthl!dl!termin!lntl:ofnllprincip!ll eubmatrioes muet bQ nonnogativo in Theorem 10. from which the theorem follows. All eigenvalues of A are positive.17. If A = AH e C"x" is positive definite.14. whence IIAxll2 ! H IIAliz = max .17). However. All eigenvalues of A are nonnegaTive. The determinants of all principal submatrices of A are nonnegative.l3 provides upper (AO and lower (An) bounds for (A. A can be wrirren in [he/orm MT M. 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. where M 6 R ix " and k > rank(A) "" rank(M). All eigenvalues of A are positive. consider the matrix A — [0 _l~].. of all principal submatrices of 2. A can be written in the form MT M.100 100 Chapter 10.13 provides (A 1) Rayleigh quotient of jc. xfO IIxll2 I 0 Definition submatrixofan n x k) x k) Definition 10. x E C".= Amax{A A).
If A > B and M e Rm .18. Ll E C1""1^""^ and .nxn"be nonnegative definite. if = /2.22. [16. BM. The factor M in Theorem 10. [ fz ti o o l [~ 0] ~ 0 v'3 . Theorem 10. j proof (see. definite if A is positive definite). then MT AM > MT TBM. 0 Recall that A :::: B if the matrix A . negative and A is nonpositive principal submatrix consisting of the (2. Let A. Then there exists a positive definite.23. That is. if € E" xn we say that e jRn x that S E R nxn"isisa asquare root of AA ifS2 2 =— A. MT AM> M. negative and is nonpositive definite. That is. The following Recall that A > B if the matrix A — B is nonnegative definite. for example.. assume the result is true for matrices of order n . E jRnxn MT AM > M BM. Then A has aaunique nonnegative definite square root S.2) element is. for example. Definite Matrices 10.22. where L\ e c(nl)x(nl) is nonsingular and lower triangular as = L1Lf.B is nonnegative definite.10. In general. matrices (both symmetric and square root of if S A. 2. Its proof is straightforward from basic definitions. Moreover. Theorem 10.is a square root. B e Rnxn be symmetric.. Remark 10..20. assume the result is true for matrices of order — 1 so that B may be written as B = L\L^.21.2. and positive definite. [16. In general. with positive diagonal elements such that positive Proof: The proof is by induction. The following standard theorem is stated without proof (see. Definite Matrices 101 101 principal submatrix consisting of the (2. any matrix S of c e s 9 the " °* ™ the form [ ssinOe _ ccosOe ] IS a square root. in fact. Let A E lR. The case = is trivially true. It is stated and proved below for the more general Hermitian case. any matrix of nonsymmetric) have infinitely many square roots. = LLH. Hermitian case. SA = AS and rankS = rank A (and hence S is positive = AS S S. For example. if A = lz. If >BandMe jRnxm. A e R"x be nonnegative definite. p. For example. rankS = rankA definite definite if positive definite).1 so that B By our induction hypothesis. Write the matrix A in Proof: The proof is by induction. Its proof is straightforward from theorem is useful in "comparing" symmetric matrices.we say 181]). The factor M in Theorem 10. nxn Theorem 10. basic definitions. A stronger form of the third characterization in Theorem 10. standard theorem stated 181]). if then M can be then M can be [1 0]. if Remark 10. if A E lR.18. Let A e c nxn be Hermitian unique nonsingular lower triangular matrix L nonsingular A = LLH. It is stated and proved below for the more general known as the Cholesky factorization.2) element is. p. 1f A :::: Band M E Rnxm. .3 is not unique.20. It concerns the notion of the "square root" of a matrix. concerns the notion of the "square root" of a matrix. Then A has unique nonnegative Theorem 10. For example. If A> Band E jR~xm.17 is available and is A stronger form of the third characterization in Theorem 10. The following theorem is useful in "comparing" symmetric matrices.2. Write the matrix A in the form the form By our induction hypothesis. 10rm [COSO _ Sino] . nxm 2.23.17 is available and is known as the Cholesky factorization. then MT AM :::: MTTBM. For example. The case n = 1 is trivially true. in fact.3 is not unique. E <C Theorem 10. matrices (both symmetric and nonsymmetric) have infinitely many square roots. 1.nxn .
2]. Take P =[ S~ 'f [I ] and Q = V to complete the proof. Substituting in the involving we find 2 a2 = ann . However. [21.b H B1b (= the Schur complement of B in A).4) [7.b H B1b > O. 131]. Alternatively. for example (10. However. we see that we must have L\c = b and ann = CHc + a 2. suppose A has an SVD of the form (5.4). Then there exist matrices P E C: xm and Q e C"nx" such E c.2) in its complex version. say.• Clearly we see we L I C = b and ann = c HC a 2 c is given simply by c = C. for example. multiplication where a is positive. ann Since det(B) > 0. the SVD is relatively expensive to compute and other canonical forms exist that are intermediate between (l0. But know that o < det(A) = det [ ~ b ] = det(B) det(a nn _ b H B1b). yields a far "simpler" canonical form (10. Canonical Forms with positive diagonal elements. Ch. Substituting in the expression involving a. Let A € C™*71.p. 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. Canonical Forms Chapter 10. Then [ Sl o 0 ] [ I Uf U H ] AV = [I 0 0 ] 0 . 5].24.3 Equivalence Transformations and Congruence Equivalence Transformations and Congruence Theorem 10.lb. Two such forms are stated here. [21. Alternatively. see. we find by L^b..131].102 102 Chapter 10. numerical procedures for computing such procedures an equivalence directly via. Ch.4) and the SVD. we must have ann —bHB lb > 0.24. Performing the indicated matrix multiplication and equating the corresponding submatrices.b B1b completes D 10. They are more stably computable than (lOA) and more efficiently computable than a full SVD. [4.3 10. The numerically preferred equivalence is. of ann — b 0 root of «„„ . Choosing a to be the positive square ann . Then E c~xn such exist e C™ x m that that PAQ=[~ ~l (l0.xn. p. are generally unreliable. Gaussian or elementary row and column operations.b HL\H L11b = ann . Choosing a be det(fi) > HB~lb completes the proof.4) Proof: proof Proof: A classical proof can be consulted in.4) efficiently available. of course. . the unitary equivunitary alence known as the SVD. But we = ann — b LIH L\lb = ann — bH B~lb B A). 0 Note that the greater freedom afforded by the equivalence transformation of Theorem afforded 10. as opposed to the more restrictive situation of a similarity transformation. Many similar results are also (10. available.
Then the inertia of A is the triple of inertia of of negative. Definition 10.In[! 1o o o 0 0 00] 10 =(2. respectively. negative. Let A E e~xn.. where R e €. Example 10.25 (Complete Orthogonal Decomposition). v. 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. D Proof: For the proof. Proof: For the proof.28. then rank(A) = n + v. HE C xn E e~ xn. Then there exist Theorem 10.v. Note that congruence preserves the property of being Hermitian. Let A E C™ x ". and £ denote the numbers of positive. The transformation A i> XH AX is called a congruence. v.31 (Sylvester's Law of Inertia).31 guarantees that rank and signature of a a matrixare preserved under Theorem 10. When A has full column rank but is "near" a rank deficient matrix. if A is Hermitian. Theorem 10. n The signature of A is given by sig(A) = n . phenomena at a cost considerably less than a full SVD. .27. In(A) = ln(X Proof: For the proof.26. Proof: For the proof. if A is Hermitian. Let A = AH E e nxn and let rr. of A. 0). a congruence.31 guarantees that rank and signature of matrix are preserved under congruence. It is of interest to ask what other properties of a matrix are preserved under congruence. nxn E e X E e~xn. upper Proof: For the proof. Proof: For the proof.26. Equivalence Transformations and Congruence 10. Let A = A He ennxn and X e Cnnxn. (TT.1. see [4] for details. for example. If In(A) = (rr. It is of interest to ask what other properties of a matrix are then X H AX is also Hermitian. v.31 (Sylvester's Law of Inertia).. of each eigenvalue. where R E Crrxr is upper triangular and S e C rx( " r) is arbitrary but in general nonzero. i. In(A) = In(X AX). If A = A" E C nnxn. 134]. £). see [4]. numbers In(A) (n. It turns out that the principal property so preserved is the sign of each eigenvalue. see [4]. D 0 Remark 10.29. n. Note that congruence preserves the property of being Hermitian. Let A = AH e C"x" and let 7t. D Theorem 10. Again. Then H HAX). 0 2.e. Note that a congruence is a similarity if and only ifX is unitary. see [4]. Remark 10. In(A) 3. 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. The signature of is Example 10. Let A e e~xn.xr is upper (or lower) triangular with positive diagonal elements.e. v.3.27.3. Definition 10. Again. 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. of A. and zero eigenvalues. l. then rank(A) rr v. 2.0). We then have the following.xr E erx(nr) arbitrary general nonzero. If A AH e e x " then A> 0 if and only if In(A) = (n.30. . v. The H. £). respectively. see. Then is the numbers In(A) = (rr. When A has full column rank but is "near" a rank deficient matrix. Theorem 10. We then have the following. 0. and eigenvalues. p.25 (Complete Orthogonal Decomposition).XH AX Definition 10.t h e n A > 0 if and only if In (A) = (n. Definition 10. [21.6) E e. see. see [4] for details.30.0. v. i. sig(A) = rr — v. see [4].28. Let A e C™ ". for example. It turns out that the principal property so preserved is the sign preserved under congruence. Equivalence Transformations and Congruence 103 103 Theorem 10. 0 D x Theorem 10.1). then XH AX is also Hermitian.29. p. [21. congruence. and ~ denote the numbers of positive. Note that a congruence is a similarity if and only if X is unitary. Then there exists a unitary matrix Q e e mxm and a Theorem 10. 134].5) where R E e. Let A e Cnxn and X e Cnnxn.rrxr is upper (or lower) triangular with positive diagonal elements.10.
.. and D . . where the number of E c~xn XH AX = diag(1. D > and .4 Rational Canonical Form Rational Canonical Form rational One final canonical form to be mentioned is the rational canonical form.. and the final £ are 0. By Theorem 10...32.. or D > 0 and A . .BT AI > 0. .4 10.I BT > O. 0). AA+B = B. and D .1 10. Xw denote the eigenvalues of A and order them such that the first TTare ~ O. 1. Proof: proof Proof: The proof follows by considering. Then = AT D = DT.BT A+B > 0. Suppose A = AT and D = DT. .104 104 Chapter 10. Define the x n matrix vv = diag(I/~. the number 0/0 's is (... . An).1).33. the number of — 's is v. . v. . . Suppose A = AT and D = DT. Then there exists a matrix AH C"xn In(A) = (Jr.. B D ] >  ° if and only if A:::: 0. I/.BT A~l B > 0. X e C"nxn such that XHAX = diag(l. Proof: AI. X UW desired 10. the number of Il's is v. Define the nn x n matrix U UH AV = diag(Ai.fArr+I' . 1. .3. . 0..35. 0 D 10.0). Canonical Forms Chapter 10.. . .BD.... the next v are negative. . A w ). . I.. .. . ifand only ifeither A > 0 and D . . .. v. if ifA>0.. if and if either A> and D . 0 D Then it is easy to check that X = V VV yields the desired result. left AT D DT. I.. Proof: Consider the congruence with Proof: Consider proof Theorem and proceed as in the proof of Theorem 10. Let A = AHeE cnxn with In(A) = (jt. .3. . . Then Remark Remark 10.0. . . where the number of X 1's is Jr. 1..34. and the numberofO's is~.. .BT A+B:::: o. . the congruence B ] [I D ~ 0 _AI B I ° JT [ A BT ~ ][ ~ 0 D The details are straightforward and are left to the reader. 1. Theorem 10. An of Jr Proof: Let AI . ..fArr+v.2 there exists a unitary matrix V such that VHAU = diag(AI.BD^BT > 0. £). Note the symmetric Schur complements of A (or D) in the theorem.. I. for example. AA+B = B.. 1/.33. ..1 Block matrices and definiteness Theorem 10. 's is 7i. Canonical Forms Theorem 10. I/~. Theorem positive.. O.
4.4. : ~ ! ~01].37. the inverse of a nonsingular companion matrix is again in companion form.18). Notice that in all cases a companion matrix is nonsingular if and only if aO i= O. A is easily seen to be similar to the following matrix identity similarity P given by (9. 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. Companion matrices also appear in the literature in several equivalent forms. For In fact.(ao + «A + . 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. Using the reverseorder This matrix is a special case of a matrix in lower Hessenberg form. 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. consider the companion matrix illustrate. Rational Canonical Form 105 105 Definition A matrix A e M"x" is said to be Definition 10.8) This matrix is a special case of a matrix in lower Hessenberg form. is said to be in cornpanion form. the inverse of a nonsingular companion matrix is again in companion form.7) is called a cornpanion rnatrix or Definition 10. the Moreover. A matrix A e E nx " of the form (10. Rational Canonical Form 10. + an_IAnI). has only one block associated with each distinct eigenvalue. equivalently.11) . since a matrix is similar to its transpose (see exercise 13 in Chapter 9). consider the companion matrix (l0.9) Moreover.. 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 . A matrix A E lRnxn of the form (10. o (10. if its Jordan canonical form and characteristic polynomial are the same or. l 0 0 ~ ao ~ ao _!!l (10.10) o 1 o 1 o o o o o o (10. Using the reverseorder identity similarity P given by (9. In fact.7) Definition 10. To illustrate.37.18).7) is called a companion matrix or is said to be in companion forrn. For £*Yamr\1j=» example.10. 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 ~')..36. equivalently.Then it can be shown (see [12]) that A mial is 7r(A) = A" . since a matrix is similar to its transpose (see exercise 13 in Chapter 9). To Companion matrices also appear in the literature in several equivalent forms. if its Jordan canonical form has only one block associated with each distinct eigenvalue.
Let a E JRn1 denote the vector [ai. If A E R nx " is derogatory. Let a = a\ + a\ + • • • + a%_{ and y = 1 + «. see haps surprisingly. with a similar result for companion matrices of the form (10. Canonical Forms Chapter 10. Leta = ar aJ al 2_ 2 ( y + Jy 2. If a companion matrix of the form (10. Then A in (10.e. is the fact that their singular values can be found in closed form. Algorithms to reduce but unfortunately they are often very difficult to work with numerically. For details. + a.Q + a. Explicit formulas for all the associated right and left singular vectors can Remark 10. Let al ~ GI ~ • • ~ an be the singular values of the companion matrix Theorem 10. For example.. Let a\ > a2 > . Theorem 10. = ~ (y . associated at least one eigenvalue. also be derived easily. in matrices are known to possess many undesirable numerical properties. is the fact that their singular values can be found in closed form. the largest and smallest singular values can also be written in the equivalent form If ao =1= 0. i. among which. companion an arbitrary matrix to companion form are numerically unstable. Such matrices are said to be in each of whose diagonal blocks is a companion matrix. Ifao ^ 0..caa T ca o J. in n general and especially as n increases. companion matrices are known to possess many undesirable numerical properties. Then + ai + .caa T = (I + aaT) I . matrix is not a companion matrix unless a = O. the largest and smallest singular values can also be written in the equivalent form Remark 10. has more than one Jordan block associated with If A € JRnxn derogatory. and hence the pseudoinverse of a singular companion + matrix is not a companion matrix unless a = 0. nonsingular ones are nearly singular.... anIf and let e M"" \a\.e. and so forth [14]. . 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. it can be shown that a derogatory matrix is similar to a block diagonal matrix.. stable ones are nearly unstable.39. . if ao = 1 inverse can still be computed.. 02.10). and perCompanion matrices have many other interesting properties.. [12].. Moreover.. each of whose diagonal blocks is a companion matrix. Moreover. . and so forth [14]. if ao = 0. a n i] and l c I+~T a.4aJ) . and perhaps surprisingly. see [14]. a2. Such matrices are said to be in rational canonical form Frobenius rational canonical form (or Frobenius canonical form). their eigenstructure is extremely ill conditioned. then its pseudoIf singular.106 Chapter 10. form). .38.• > an be the singular values of the companion matrix A in (10. see. at least one eigenvalue. .7).. especially nonsingular ones are nearly singular. i. among which.10).1.7).4ao ' 1 2)  a? = 1 for i = 2. a.._1 and y = 1 + + a.7). then it is not similar to a companion matrix of the form (10.7) is singular.. Canonical Forms with a similar result for companion matrices of the form (10. I — T = T) Note that / .7). Algorithms to reduce an arbitrary matrix to companion form are numerically unstable.38. Companion matrices have many other interesting properties.39.Jy2 . then it is not similar to a companion matrix of the form (10. 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. However. n . For example. for example. stable ones are nearly unstable.. 3. Explicit formulas for all the associated right and left singular vectors can also be derived easily..
Show that a.(A)I for ii E!l.. ...11). If this number is large.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.38 yields some understanding of why difficult numerical behavior might be expected for companion matrices. A [ must also be positive 7. Find a unitary matrix U such that [~ M CC x 2 Find a unitary matrix U such that 6. Let A G Cnx" and define p(A) = maxx€A(A) IAI. this condition number is the ratio of largest to smallest singular precision. say O(lO k ). Show that a. is true if n = 2. (A) = IA. 6. It is not unusual for y to be large for large n. Show that the converse radius of A. E jRnxn be symmetric. For example. Let R. 5.5 to find a unitary matrix Q that reduces A e C"x" to lower triangular form. yn and singular values a\ ~ a2 > . EXERCISES EXERCISES 1.. then Af(A) = N(A Tr ).11). K\ (A) (10.. Use the reverseorder identity matrix P introduced in (9. Let A = I J : ]eEC 22x2. For example. then it must be diagonal. one may lose up to k digits of precision. 1. 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 . . when solving linear behavior might be expected for companion matrices. 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. can be determined explicitly as determined explicitly y+J y 2 . Theorem 10. Theorem 10. If A E Wxn is positive definite. Suppose A e E"x" is positive definite. 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. (A) A. Remark 10. Let R..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).38 yields some understanding of why difficult numerical Remark 10.18) U A E cc nxn Theorem 10. Show that [~ R > SI.. then K2(A) ^ T~I. Prove that if A e M"x" is normal. Show that if A is normal. S 6 E nxn be symmetric.18) and the matrix U in identity in (9. .. A E cc nxn peA) = max). 2. 3.. Show that if A is normal. one may lose up to k digits of to the matrix Pnorm.40. In the 2norm.. Is [ ^ A E jRnxn is definite. If A e jRn xn 8. Show that if a triangular matrix is normal. show that AI must also be positive definite. then p(A) = IIAII2' Show that the converse is true if n = 2. If this number is large.. • • > on > 0. then peA) = A2. 9.(A) for e n. It is easy to show that y/2/ao < K2(A) < £.EA(A) I'MpeA) 3.40. and when GO is small or y is large (or both). Show that [ * }.2). In the 2norm. say 0(10*). An and singular 0'1 > 0'2 ~ 4. A E jRnxn N(A) = A/"(A ). this condition number is the ratio of largest to smallest singular values which. A E en x n eigenvalues A]. Then p(A) is called the spectral radius of A. Let A € C n xn be normal with eigenvalues y1 . Let A 7.• ~ an ~ O. when solving linear equations numerical sensitivity Kp(A) = systems of equations of the form (6. by the theorem.
(a) Chapter 10.108 108 10.1 1.j 1+ j ] 2 ' (d) [ . Canonical Forms Chapter 10. .j 1+ j ] 1 . Find the inertia of the following matrices: following 10. Canonical Forms [~ ~ l (b) [ 2 1.
eO = I. For all A E JR.2) can be shown to converge for all A (has radius of convergence equal to +00).1) involves the matrix to +(0).n (11.1) for t 2: to.1. where the matrix A E Rnxn is constant chapter only to the socalled timeinvariant case.1 and linearity of the transpose.1) is then known always to exist and be unique. T T 109 109 . For all A e Rnxn. The solution of (11.1 by setting = 0. (e(eAf = e A e^. For all A JR.1 and linearity of the transpose. A) • 2. e° = I.1 Differential Equations Differential Equations = Ax(t).Ak. the matrix exponential e A e JR.1) is then known always to exist and be and does not depend on t.1. The solution of (11.nxn. Definition 11. This is known as an initialvalue problem.1 by setting AA =O.1 Properties of the matrix exponential Properties of the matrix exponential 1. Forall A EG R" XM .1) involves the matrix (11. It can be described conveniently in terms of the matrix exponential. 11. = 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 11.2) k=O The series (11. The solution of (11. unique. We restrict our attention in this for t > IQ.nxn is defined by Definition 11. Proof: This follows immediately from Definition 11. where the matrix A e JR. Proof This follows immediately from Definition 11. Proof: This follows immediately from Definition 11. We restrict our attention in this chapter only to the socalled timeinvariant case. It can be described conveniently in terms of the matrix exponential.3) which thus also converges for all A and uniformly in t. the matrix exponential e A E Rnxn is defined by the power series power series e = A L +00 1 . k. Proof This follows immediately from Definition 11.nxn.1.2) can be shown to converge for all A (has radius of convergence equal The series (11. which thus also converges for all A and uniformly in t.1. This is known as an initialvalue problem.Chapter 11 Chapter 11 Linear Differential and Linear Differential and Difference Equations Difference Equations 11.nxn is constant and does not depend on t.1 11. The solution of (11. (11.
Part (b) follows similarly.. Proof' Note that Proof: Note that et(A+B) = I t + teA + B) + (A + B)2 + . For all A E JRnxn and for all t. 2 2! and and while while e e tB tA = ( 1+ tB t2 2 + 2iB 2 +. ) . T E JR.. (a) . AB = BA. Linear Differential and Difference Equations e(t+r)A e(t+T)A 3. ForaH A E R" x " and for all t e JR. For all e JRnxn and for all E R. Proof" Simply take T = — t in property 3. For all e R"x" and for all t. (a) C{etA = (sIArl. 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. 2! and and e e tA rA 2 = ( I + t A + t2! A 2 +. 5.. on (t + T)*. AB = B A.. ) ( I + T A + T2!2 A 2 +.. ) .. = e'A erA = elAe'A . Proof: We prove only (a).A)I. For all A.1 {(j/A). all A € R"x" and for all t € lR. and B commute... Proof" We prove only (a). Proof" Note that Proof: Note that e(t+r)A = etA erA = erAe tA . = I + (t + T)A + (t + T)2 A 2 + .. (etA)1 = e.lI{(sl. Then for E JRnxn t E R. ) ( 1+ tA + t2!A 2 +.tA .e.1 } = erA. Part (b) follows similarly.e.. i. binomial theorem on (A B) and the commutativity of A and B.. et(A+B) =etAe tB = etBe tA if and only if A all e JRnxn and all e R.l{e tA}} = (sI . B E R" xn and for all t E JR. Let denote the Laplace transform and £~! the inverse Laplace transform. 6. 4.110 110 Chapter 11. Then for 6. r e R. et(A+B) =^e'Ae'B = e'Be'A and and B commute. Let £ denote the Laplace transform and £1 the inverse Laplace transform. (b) £.A)I} = «M.. i. Linear Differential and Difference Equations Chapter 11. {+oo = io et(sl)e tA dt since A and (sf) commute =io (+oo ef(Asl) dt . (e'A)~l e~'A. (b) . 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. Proof: Simply take T = t in property 3...
l)e .u .H using the JCF. ) etA I < MIIA21111e  L'lt (L'lt)2 + IIAII + IIAI12 + .. using the JCF.. it can be differentiated termbyProof: Since the series (11.A)I. A2 + (~~)2 A tA II tA Il 1 (_ 2! + . The matrix (s I . 1h(e tA ) = AetA = etA A.11. for convenience. For all A E JRnxn and for all E JR. .. = (sl A)..Ae tA I = III (etAe~tA L'lt = = /A) . ) 3 II I ( ~.. ) 3! 4! L'ltiIAIl < L'lt1lA21111e (1 + + (~t IIAII2 + ..=1 m Xiet(Jisl)y. If this is not the case. Differential Equations 111 111 = {+oo 10 n t 1 e(AiS)t x. employed I e(t+~t)AAt. ) = I ( Ae + = tA ~.... it can be differentiated termbyterm from which the result follows immediately... For all A e R"x" and for all t e R. £(e'A) 7. Differential Equations 11. that A is diagonalizable. All succeeding steps in the proof then follow in straightforward way.1.1. Alternatively..H = '"' assuming Re s > Re Ai for i E !! = (sI . the formal definition d dt _(/A) = lim ~t+O e(t+M)A _ etA L'lt can be employed as follows. Notice in the proof that we have assumed.All succeeding steps in the proof then follow in aastraightforward way.Ae tA tA tA I I e tA . A 2etA + .H L.=1 = ~[fo+oo e(AiS)t dt]x.X i y.. that A is diagonalizable. s . e'A Proof: Since the series (11.H dt assuming A is diagonalizable .etA .A"I i=1 . for convenience.3) is uniformly convergent.. the scalar dyadic decomposition can be replaced by If this is not the case.3) is uniformly convergent....Ae tA etA) . ) = L'lt IIA 21111e tA IIe~tIIAII...y.1 The matrix (s I — A) I is called the resolvent of A and is defined for all s not in A (A). For any consistent matrix norm.Ae II = I L'lt (M)2 + ~ A 2 +.AetAil Ae tA I ~t (e~tAetA I (M A I ~t (e~tA .y. Notice in the proof that we have assumed.A) ~' is called the resolvent of A and is defined for all s not in A (A).. . . the scalar dyadic decomposition can be replaced by et(Asl) =L .
D 11. 11. D Ir: Remark 11.. B E Wnxm and let the vectorvalued function u be given Let A e IR nxn . x(to) = e(toto)A Xo = XQ so.6). Linear Differential and Difference Equations Chapter 11. the limit exists and equals Ae'A •.2 Homogeneous linear differential equations Homogeneous equations x(t) Theorem 11. or one can use the fact that A commutes with any polynomial of A of finite degree and hence with e'A.3. The formula can be derived by means of an integrating factor "trick" direct differentiation.112 112 Chapter 11.3 Inhomogeneous linear differential equations Inhomogeneous equations Theorem 11. B e IR xm and let the vectorvalued function u be given Theorem and.4. by the fundamental existence and uniqueness theorem for ordinary differential equations.i~t()Oc() nnd uniqu()Oc:s:s theorem for *('o)} = <?(f°~fo)/1. The proof above simply verifies the variation of parameters formula by direct differentiation. t ) . say. Thus.4) for t ::: to is given by (11. the righthand side above clearly goes to 0 as At goes to 0.4.4). 0 ordinary differential equations. Thus.f(p(t). 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.8) .t goes to O.tA to get as follows. The solution of the linear homogeneous initialvalue problem = Ax(l). 0 uniqueness theorem for ordinary differential equations. The general Proof: Differentiate (11. fact that A commutes with any polynomial of A of finite degree and hence with etA. Then the solution of the linear inhomogeneous initialvalue problem and. or one can use the limit exists and equals Ae t A A similar proof yields the limit et A A. x(to) = Xo E IRn (11. by the fundamental existence and x(t0) — e(fo~t°')AXQ — Xo uniqueness theorem for ordinary differential equations.1. A similar proof yields the limit e'A A. (11. The proof above simply verifies the variation of parameters formula by Remark 11. The formula can be derived by means of an integrating factor "trick" as follows.2. Also.7) and again use property 7 of the matrix exponential.5) and use property 7 of the matrix exponential to get x t ) = Ae(tto)A xo fundamental Ae(t~to)Axo = Ax(t).6).. Then the solution of the linear inhomogeneous initialvalue problem x(t) = Ax(t) + Bu(t).5) is the solution of (11. lo t (11.7) is the solution of (11. Let A E Rnxn . (11.Ax = Bu by e. the For fixed t. The general formula formula d dt l q (t) pet) f(x. (11. Premultiply the equation x — Ax = Bu by e~ to get (11. Let A E IR n xn. Premultiply the equation x .7) Proof: Differentiate (11. t) dx = l af(x t) ' dx pet) at (t) q + dq(t) dp(t) f(q(t).5) is the solution of (11.¥o + 0 = XQ so.6) for t ::: to is given by the variation of parameters formula for t > IQ is given by the variation of parameters formula x(t) = e(tto)A xo + t e(ts)A Bu(s) ds. (11. the righthand side above clearly goes to 0 as t:.dt dt is used to get x ( t ) = Ae(tto)Ax0 + f'o Ae('s)ABu(s) ds + Bu(t) = Ax(t) + Bu(t). t ) . x(to) = Xo E IR n (11. Also.4).7) is the solution of (1l.5) and use property 7 of the matrix exponential to get x ((t) = Proof: Differentiate (11. The solution ofthe linear homogeneous initialvalue problem Let A e Rnxn. continuous. continuous.7) and again use property 7 of the matrix exponential.5) Proof: Differentiate (11.1. Linear Differential and Difference Equations For fixed t. say.
the When C is symmetric in (11. X((t) is symmetric and (11.6. Differential Equations [to.11) is known as a Sylvester Sylvester differential equation.9) for t ::: to is given by for t > to is given by X(t) = e(tto)Ac. X t) X 0 D Corollary 11.11.1. X(to) =C E jRnxn (11. 11. E ]R.6.2. e jRnxn. and C e Rnxm.10) coefficient In the matrix case. problem problem X(t) = AX(t) + X(t)B. The fact that X((t) satisfies the initial condition is trivial. following to = O. Let A E Rnxn.7. X(O) =C (11.7. and the proof is essentially the same.8) over the interval [to. the following theorem is stated with initial time to = 0. Then the matrix initialvalue E jRmxm. X(O) = C (11.2.11) X(t) = etACe = e ratB has the solution X ( t ) — atACe tB . The first is an obvious generalization of Theorem 11.4 Linear matrix differential equations Linear matrix differential equations Matrixvalued initialvalue problems also occur frequently. The initialvalue problem (11.5. (11. t]: Now integrate (11. the Theorem 11. punov differential equation. E ]R.etoAx(to) = lto t e. Let A E Wlxn. differential equation.sA Bu(s) ds x(t) = e(ttolA xo + lto t e(ts)A Bu(s) ds.4 11. Then the matrix initialvalue problem X(t) = AX(t) + X(t)AT.1. For convenience. Theorem 11.nxn. C e IR" ". t exponential.1. Theorem 11. etAx(t) .12) is known as a LyaX t) punov differential equation. the Proof: Differentiate etACe tB property Proof: Differentiate etACetB with respect to t and use property 7 of the matrix exponential. and hence t d esAx(s) ds = to ds 1t to eSABu(s) ds. Let A.nxm.1. . t]: 113 1 Thus.12). Corollary 11. and the proof is essentially the same. The solution of the matrix linear homogeneous initialvalue e jRnxn.12) X(t) = etACetAT has the solution X(t} = etACetAT. Differential Equations 11. The of nrohlcm problem X(t) = AX(t). B e R m x m . Theorem 11.. we can have coefficient matrices on both the right and left.
that it is diagonalizable (if A is not diagonalizable. Then the solution x(t) of (11. where J is a JCF for A. 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. where J is a JCF for A.e'J.1 n Le A• X'YiH . Similarly.4) can be written x(t) = e(tto)A Xo E jRnxn E Wxn = (ti.5 Modal decompositions Let A and suppose. for convenience. if A is diagonalizable in geneml.114 114 Chapter 11.x.li y t as discussed in Chapter 9). ~ 11.5 11. The decomposition above expresses the solution x(t) as a weighted sum of its modal velocities and directions. in the inhomogeneous case we can write Similarly. Linear Differential and Difference Equations Chapter 11. modal velocities and directions. The decomposition above expresses the solution x (t) as a weighted sum of its directions.y. 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 *.1.1 .iUtO)Xiyr) Xo 1=1 n = L(YiHxoeAi(ttO»Xi. Then the solution x(t) of (11. Then Then etA = etXJX1 = XetJX. 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. Then Then i=1 n = L(aieAiUtO»Xi. i=1 In the last equality we have used the fact that yf*Xj = Sfj. In the last equality we have used the fact that YiHXj = flij. ~ 1=1 I t.4) can be written A = L X. 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). for convenience.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. are called the modal directions. Linear Differential and Difference Equations 11. Let A E jRnxn and suppose X e jR~xn is such that XI AX = J. that it is diagonalizable (if A is not diagonalizLet A and suppose.1.H . the rest of this subsection is easily generalized by using the JCF and the decomposition able.
its first superdiagonal (and O's elsewhere). nilpotent Definition 11.. Nk~lI has a 1 in its (1. Thus.e. Mp~l ^ O. o A o o A Clearly A/ and N commute. and N kforth.eAt). the series expansion of e'N is finite. ••• . and so forth.11. the problem clearly reduces simply to the computation of problem clearly reduces the exponential of a Jordan block..7. elN is is nilpotent of degree k. A matrix M E M nx " is nilpotent of degree (or index.. Finally. k) O's k k N = 0.!etN by property 4 of the matrix exponential. i. . aareal version of the above can be worked out. while MPI t=. N22 has l's along only its second superdiagonal. Thus. To be specific. degree k. t2 t k. Mp = 0. O. it is easy to check that while N has 1's along only its first superdiagonal (and O's elsewhere).EeCkxk be aaJordan block of the form Ji <Ckxk be Jordan block of the form A Ji = 1 o o o =U+N.0.I e IN =I+tN+N 2 + . AI e I.. Differential Equations 115 If A is diagonalizable.1. l's For the matrix N defined above. e lN finite. e'u e l N tu x lH = diag(e At .I)! I o t 1 o Thus. or grade) p if if matrix M e jRnxn is nilpotent of degree (or index.8. Differential Equations 11. + N k2! (k . e ttJi = eO. or grade) MP = 0.. (1. 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. teAl eAt = 0 0 0 2I e 12 At teAl 0 eAt In the case when A is complex. But e tN is almost as easy since N The diagonal part is easy: e e = diag(e '. ext}.. k) element and has O's everywhere else. A. is complex. let .1. eAt teAt eAt o 2I e 12 At IkI At e (kI)! 0 ell. N has 1's along only its second superdiagonal.. of In the more general case. real version of the above can be worked out.
9. the unique OTQ. ==> 2a2 = t 2 e. I. the function g is known and /(A) = g(A).nxn and /(A) = etx.) = (A + 1)3. The method is stated and illustrated for the hand calculation in smallorder problems. i Em. Let Example 11.9. ni . (A. . functions..s are distinct. all the Ak — expressed k 1. 2} and etA Xe tJ =[=i a = xI =[ =[ 2 1 ] exp t ] [ [ 2 0 ~ ] [ 1 1 2 1 2 ] 2 1 e~2t te. k = 0.2t e. so m = 1 and n{ = 3. in fact. Given A E jRnxn and f(A) = etA. Theorem 9. . The motivation for this method is the CayleyHamilton Theorem. compute f(A) = e'A. Then jr(A. Let A = [ ~_\ J]. a. t fixed Given A € E. Let A = [~ o ~01~ ] t . the superscript (k) denotes the fcth derivative with respect to A.116 Chapter 11.10.s are given by g(A) — ao aiS a\X o^A. Linear Differential and Difference Equations Example 11.10.3. + I) 3 . compute f(A) = etA. g(I) = f(1) ==> ao .. Let A Then A (A) = {2. The motivation for this method is known. terms of order greater than n . .1.2a2 = te. characteristic of n(X (^ ~~ ^i)"'» where the A..Ai t'.t • g'(1) = f'(1) g"(I) = 1"(1) .a l +a2 = e==> at . Linear Differential and Difference Equations Chapter 11. where t is a fixed scalar. lowerorder g Example 11.s known.. Define the Ai nr=1 n where ao. f(A) n(A) etK. . n . Here.1 can be expressed as linear combinations of Ak for k = 0. which says that all powers of A greater than A n .. Thus.2.. so m = 1 and nl Let g(X) = UQ + alA + a2A2. anl are n constants that are to be determined.1 in the power series for et A can be written in terms of these greater n— e' A lowerorder powers as well. The method is stated and illustrated for the exponential function but applies equally well to other functions..t . . They are.. I. . The polynomial g gives the appropriate linear combination. Then A(A) = {2. Suppose the characteristic polynomial of A can be written as n(A)) = Yi?=i (A .. . .I. — 1. . 2} and Example 11.. . . ani solution of the n equations: g(k)(Ai) = f(k)(Ai). . With the aiS then kth superscript (&) X. and /(A) = etA. Then the three equations for the 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. the function g is known and f(A) = g(A).
11. t ff>\ tk TU^^ _/"i\ Example 11.2t .2t [ ~ o ] + te.2t I [4 4] I 0 _ [  e. but general nonsymbolic computational effective smallorder techniques numerically problem equivalent techniques are numerically unstable since the problem is theoretically equivalent to knowing precisely a JCE JCF.2t ) 2te. 2.2t _ Other methods Other methods 1.1.2t aL = + 2te. Thus.2t + 2te.1. we find 117 Thus.11.2t te. we find Solving for the aiS. g'(2) = f'(2) = te Solving for the a. we find Solving for the a. Differential Equations 11 . The matrix analogue yields e A ~ functions rational eA = . te. Then the defining equations for the a. Use etA = £~l{(sl . 1.s are given by Let g(A) ao + aLA. s. Then the defining equations for the aiS are given by 6] g(2) = f(2) ==> ao ==> al 2al = e.. Use Pade approximation.2t .11. Let g(A.s. Let A = [ ~4 J] and /(A) = eO. This etA = .A)^ 1 } and techniques for inverse Laplace transforms.2t . f(A) = etA = g(A) = aoI + al A = (e. 2. we find ao = e. Differential Equations Solving for the ai s. 2t . There is an extensive literature on approximating certain nonlinear functions by rational functions.cI{(sI — A)I} is quite effective for smallorder problems. Let A _* Example 11.) = «o + ofiA. Then rr(A) = f\ + o\2 so m = and (A i 2) «i nL = 2. Then 7r(X) = (A+ 2)22 so m = 11and [::::~ 4i and f(A) = ea.
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. but since the system is timeinvariant. •• vq A . we have chosen ko = 0 for want to keep the formulas "clean" (i. 11. 11. Linear Differential and Difference Equations Chapter 11. Linear discretetime systems. The solution ofthe linear homogeneous system of difference equations equations (11. where the matrix A in (11.e..2. = P = multiplying it by 1/2* for sufficiently large k and using the fact that e = ( e j . a Pad6 approximation for denominator the exponential is accurate only in a neighborhood of the origin. Reduce A to (real) Schur form S via the unitary similarity U and use e A 3.. say.13) for k > 0 is given by for k 2:: 0 is given by Proof: The proof is almost immediate upon substitution of (11.• + Vq A q. [19]. by this means when IIAII is sufficiently small.13. we restrict our attention only to the socalled timeinvariant Remark 11. in the matrix case this means when  A is sufficiently small. Many methods are outlined in. but since the system is timeinvariant. + opAP and N(A) = vol + vIA + D~ (A)N(A). say. where D(A) = 001 + olA + . 0 D Remark 11.15) .. for example. we restrict our attention only to the socalled timeinvariant case.1 Homogeneous linear difference equations Homogeneous linear difference equations Theorem 11.2. by 22' 2* )A A multiplying it by 1/2k for sufficiently large k and using the fact that A = / { ]I //2')A )\ * . convenience. Reliable and efficient computation of matrix functions such as e A and log(A) remains a fertile area for research. The solution of the linear homogeneous system ofdifference Let A e jRn xn. We could also case. Many methods are outlined in.13).. no double subscripts). we have chosen ko = 0 for convenience.118 118 l Chapter 11. and since we consider an arbitrary "initial time" ko. We could also consider an arbitrary "initial time" ko. Let A E Rnxn. Unfortunately. Then the solution of the inhomogeneous initialvalue problem mvectors. [19]. B e Rnxm and suppose {«*}£§ « a given sequence of mvectors. modeled by systems of equations of the previous section. E jRnxm {udt~ is of Theorem 11. Linear discretetime systems.13) is constant and does not depend on k.13) is constant and does not depend on k. Again. Reliable and efficient computation 4. exhibit many parallels to the continuoustime differential equation case. where the matrix A in (11. modeled by systems of difference equations. 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. 4. 11. This can be arranged by scaling A. for example. Then the solution of the inhomogeneous initialvalue problem (11. and this observation is exploited frequently. case.e.13).14) into (11.. no double subscripts). Proof: The proof is almost immediate upon substitution of (11. in the matrix case the exponential is accurate only in a neighborhood of the origin. Unfortunately.1 11.14) into (11. e (e( 3. of matrix functions such as e A and 10g(A) remains a fertile area for research. Linear Differential and Difference Equations DI(A)N(A). Numerical loss of accuracy can occur in this procedure from the successive squarings. Again.12.2 Inhomogeneous linear difference equations Inhomogeneous linear difference equations E jRnxn.2. 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. and since we want to keep the formulas "clean" (i. This can be arranged by scaling A. Let A e Rnxn. a Fade approximation for polynomials of various orders.13. Numerical loss of accuracy can occur in this procedure from the successive squarings. and this observation is exploited frequently.14.2 11. exhibit many parallels to the continuoustime differential equation difference equations. eS ..
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. +00 k=O z z = (lzIA)I = z(zI .O. Assume that A e M" xn and let X e jR~xn be such that XI AX = /. it is then easy to compute Ak via the formula Ak = XJkXXI Ak Ak — X Jk If diagonalizable.=1 H l If A is diagonalizable.2.3 11. is to use ztransforms. k:::. substitution of (11. a matrix exponential.y.11.1 _I tA~X.A)I.16) Proof: The proof is again almost immediate Proof: The proof is again almost immediate upon substitution of (11. Then JCF for A. again mostly for smallorder probsmallorder lems. the ztransform of the sequence {Ak}} is then given by AEA(A) X€A(A) k "'kk 1 12 Z({A})=L.2. since /* is simply a diagonal matrix.16) into (11. Then Ak = (XJXI)k = XJkX. based Methods based on the JCF are sometimes useful... where J is a E jRnxn and X E R^n J. X~1 AX JCF for A. the ztransform of the sequence {Ak is then given by Assuming z > max IAI.zA =I+A+"2 A + .16) into (11. One definition of the ztransform of a sequence {gk} is a matrix exponential. by analogy with the use of Laplace transforms to compute ztransforms.15).. which is numerically unstable but sometimes useful for hand calculation. sometimes useful Ak.15).2. k=O Assuming Izl > max A. j=O (11. Jk . Difference Equations 119 119 is given by kI xk=AkXO+LAkjIBUj. LXi Jtyi . One solution method.H m if A is diagonalizable. in general.. 0 D 11.2. Difference Equations 11. One definition of the ztransform of a sequence is +00 Z({gk}t~) = LgkZk.
120 Chapter 11. but again no universally "best" method be derived for the computation of matrix powers.6 be derived for the computation of matrix powers.. 18]. and is to be interpreted as 0 if k < q.15.is complex.. inI)(O) = CnI' (1l. ) Ak. Ch. but again no universally "best" method exists. for example. 1 1 1 2 1 ] Basic analogues of other methods such as those mentioned in Section 11.6 can also methods 11.1 Ak The symbol (: ) has the usual definition of q!(kk~q)! and is to be interpreted as 0 if k < q. For an erudite discussion of the state of the art.l8) .• = AI and noting that AI and the nilpotent matrix Writing Ji = XI + N and noting that XI and the nilpotent matrix N commute.17) with ¢J(t) a given function and n initial conditions 4>(t} y(O) = Co.1 (2 . aareal version of the above can be worked out. In the case when A.1 Ak ( . y(O) = CI. Linear Differential and Difference Equations Chapter 11. Consider. . the initialvalue problem initialvalue (11.1. 0 A Writing /. the problem again reduces to the computation of the power of a In the general case. The symbol ( ) has the usual definition of .15. Example 11.(^ .2k) k( _2)k1 ] k( 2l+ (2l. real version of the above can be worked out. 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 (. e Cpxp be a Jordan block of the form o .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. 11. let 7.1(2k . . see [11..1. Linear Differential and Difference Equations In the general case..3 11. the problem again reduces to the computation of the power of a To Ji E Cpxp Jordan block. Then Then 1 ] [(_2)k 1 0 k(2)kk(2) 1 ] [ _ [ (_2/.)A  ( k ) AkP+I pl 0 J/ = kA k.2) .2 0 0 0 0 kA k . [11. it is commute. A is complex. Let A = [_J Example 11. To be specific. Let A Ak = XJkX1 = [=i 4 a [2 1 J]..
19) possesses many nasty numerical properties for even moderately sized n and. . a)xyT. +h a\X+ ao. Cl. Then components Xl (t) yet).. Let . Define a vector x (?) e R" with components *i(0 = y ( t ) . (11.19) The initial conditions take the form ^(0) = c [CQ... the companion matrix A in (11... as mentioned before. where I + get. = Xn(t) = y(nl)(t). is often well worth avoiding. 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.. y(m) denotes the mth derivative of y with respect to t.. be a projection. However. . aly(t) . as mentioned before. Then Xl (I) X2(t) = X2(t) = y(t). condition. the companion Note that det(A! .718P. where !(eat .19) possesses many nasty numerical properties for even moderately sized n matrix A in (11..an_llnl)(t) Xnl (t) Xn(t) = y(n)(t) = aoy(t)  + ¢(t) = aOx\ (t) .. . c\. y E lRn and let A = xyT. at least for computational purposes. Define a vector x (t) E ]Rn with Here. . Show that e % / + 1. let a = xTy. a)xyT. However.. is often well worth avoiding.Exercises 121 121 Here.. into a linear firstorder difference equation with (vector) initial with n initial conditions. Xn(t) y { n ~ l ) ( t ) . Cn \ ."+ an_1A n~ + . into a linear firstorder difference equation with (vector) initial condition. y € R" and let A = xyT. Suppose x..a\X2(t) ... Further. Let P E lR nxn be a projection. Let P € R 1. Note that det(X7 — A) = An + an\Xn 1l H alA + ao.a)= { a t nxn p if a if a 1= 0. Show that e'A 1+ g ( t . .718P. xn(t) = Inl)(t). X2(t) yet). let a = XT y. = X3(t) = yet). •. C M _I] The initial conditions take the form X (0) = C = [co. and. . EXERCISES EXERCISES 1. Show that e P ~ ! + 1. at least for computational purposes.A) = A. Suppose x. x2(t) = y ( t ) . . v (m) denotes the mth derivative of y with respect to t. ..anlXn(t) + ¢(t).I) g(t. Further. Show that etA 2. 2. = O. Let 3. 3. 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..
Show that 1/). Let (a) Solve the differential equation (a) Solve the differential equation i = Ax . (a) Suppose H is Hamiltonian and let). Show that ).. Let K denote the skewsymmetric matrix 0 [ In In ] 0 ' In A E jR2nx2n where /„ denotes the n x n identity matrix. (d) Suppose 5. Show that 1 /A. Hamiltonian if K~1ATK = A and to be symplectic if K I ATK = A I. . 6. Let denote the skewsymmetric matrix 4..be an eigenvalue of H. Show that E jRmxn e = [eoI A sinh 1 X ] ~I . Show that SI HS must be Suppose and symplectic.be an eigenvalue of S. be an eigenvalue of S. Let 5. Show that —A. (b) Suppose S is symplectic and let A. also be an eigenvalue of H. must (b) Suppose S is symplectic and let).. also eigenvalue of (c) Suppose that H is Hamiltonian and S is symplectic. ft € lR and Let a. be an eigenvalue of H.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. H (d) Suppose H is Hamiltonian.122 122 Chapter 11. Hamiltonian. must also be an eigenValue of S.. Show that eH must be symplectic. must (a) Suppose E is Hamiltonian and let A. must also be an eigenvalue of H. x(O) =[ ~ J.A 1 . Linear Differential and Difference where X e M'nx" is arbitrary. Find a general expression for Find a general expression for 7. A matrix A e R 2nx2n is said to be K I AT K = . Linear Differential and Difference Equations Chapter 11. Let a. Find eM when A = Find etA = 8. 4. Show S~1 H S Hamiltonian.
(c) Find the distribution of the companies' assets at year k. half stays home and half goes to the Americas. . 10.e. I/X(t)1/2 = ex for all t > 0. If £0 = 1 and z\ If Zo = 1 and ZI = 2. Consider the n x n matrix initialvalue problem X(t) = AX(t) . what is the value of ZIOOO? What is the value of Zk in 2. Show that *(OII2 = aforallf > O. i. Each year half of the Americas' money stays home. Consider the initialvalue problem i(t) = Ax(t). Suppose that A E ~nxn is skewsymmetric and let ex = Ilxol12. x(O) =[ ~ l 9. a quarter goes to Europe. and the Americas (R). (c) Find the distribution of the companies' assets at year k. 11. For Europe and Asia.e. (d) Find the limiting distribution of the $40 trillion as the universe ends. and the Americas (R). i.Yet) + 2y(t) + yet) = 0.3. around the time the Cubs win a World Series). of Cf or all t. For Europe and Asia.) 12. what is the value of ZIQOO? What is the value of Zk in general? general? . Consider the n x n matrix initialvalue problem 10. (a) Find the solution of the initialvalue problem (a) Find the solution of the initialvalue problem . x(O) = Xo for t ~ O. Suppose certain multinational companies have Europe (E). Consider the initialvalue problem 9. and a quarter year half of the Americas' money stays home. as k —»• +00 (i. Europe (E). The year is 2004 and there are three large "free trade zones" in the world: Asia (A). (b) Consider the difference equation (b) Consider the difference equation Zk+2 + 2Zk+1 + Zk = O. Show that for t > 0. (Exercise adapted from Problem 5. around the time the Cubs win a World Series). goes to Asia.e. 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?. (b) Find the eigenvalues and right eigenvectors of M.e.11 in [24]. yeO) = 1.) (Exercise adapted from Problem 5.YeO) = O. Suppose that e E"x" is skewsymmetric and let a = \\XQ\\2. Each total assets of $40 trillion of which $20 trillion is in E and $20 trillion is in R. (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. 11. and a quarter goes to Asia. as (d) Find the limiting distribution of the $40 trillion as the universe ends. 12. a quarter goes to Europe.3... Suppose certain multinational companies have total assets of $40 trillion of which $20 trillion is in E and $20 trillion is in R.X(t)A.. X(O) = c. k * +00 (i.Exercises Exercises (b) Solve the differential equation (b) Solve the differential equation i 123 = Ax + b.11 in [24]. The year is 2004 and there are three large "free trade zones" in the world: Asia (A). half stays home and half goes to the Americas..
This page intentionally left blank This page intentionally left blank .
1 12. e C. such that that (12.1. Similarly. B) with A.) = det(A . As with the standard eigenvalue problem. B). B) with A. B E enxn. a.4.'AB) is called the characteristic polyDefinition 12. (A. The matrix A . hence nonreal eigenvalues must occur in complex conjugate pairs. B e jRnxn. B E C MX if there exists a scalar A E C. the characteristic polynomial is obviously real. B e enxn" if there exists a scalar 'A. and Remark 12. The roots ofn('A) are the eigenvalues of the associated nomial of the matrix pair (A. called a generalized eigenvalue. then so is ax [ay] for any nonzero scalar a E <C. e e. When A. eigenvalues for the generalized eigenvalue problem occur pencil — XB problem occur where the matrix pencil A . Definition 12.1 The Generalized Eigenvalue/Eigenvector Problem The Generalized Eigenvalue/Eigenvector Problem Ax = 'ABx. The polynomial 7r(A. 125 125 . Remark 12. The roots ofn(X. .3. The standard eigenvalue problem considered in Chapter 9 obviously where A. A nonzero vector x e C" is a right generalized eigenvector of the pair generalized eigenvector of (A.Chapter 12 Chapter 12 Generalized Eigenvalue Generalized Eigenvalue Problems 12. Definition 12.1.3. B). B e C" xn The standard eigenvalue problem considered in Chapter 9 obviously corresponds to the special case that B = I.2. the adjective "generalized" "generalized" standard eigenvalue [y] is usually dropped.2) When the context is such that no confusion can arise. The matrix A — 'AB is called a matrix pencil (or pencil of the matrices A Definition 12. called a generalized eigenvalue. Definition 12.XB is called a matrix pencil (or pencil of the matrices A and B). In this chapter we consider the generalized eigenvalue problem In we the generalized eigenvalue problem where A. a nonzero vector y e C" is a left generalized eigenvector corresponding to an E en generalized eigenvector eigenvalue 'X if eigenvalue A if (12.5) is called the characteristic polynomial of the matrix pair (A.'AB is singular. eigenvector. As with the standard eigenvalue problem.1) Ax = 'ABx. B). The polynomial n('A) = det(A — A. B E E" xn . characteristic hence nonreal eigenvalues must occur in complex conjugate pairs.2.4. A E en Definition 12. corresponds to the special case that B = I. generalized eigenvalue problem.) are the eigenvalues of the associated generalized eigenvalue problem. if x [y] is a right [left] ax [ay] for any eigenvector. and A.
B. Then the characteristic polynomial is ft det(A . in particular.5. or infinitely many B = I. Case Case 3: a = 0.I. Case 2: = 0. the associated matrix pencil is singular (as in Case N(A) n N(B) =Isingular 4 above). f3 = 0. Generalized Eigenvalue Problems Chapter 12. it is apparent where the "missing" eigenvalues have "missing" gone in Cases 2 and 3. (3 = O.0. 1 Case 3: a =I. eigenvalues associated with the pencil A . It is instructive to consider the reciprocal pencil associated with the example in It reciprocal Remark 12.O. {3 = 0.XB./.{3 = 0. There are two eigenvalues. For example. the characteristic polynomial is = (I .AB. 1 and ~. Note appear.6.LA and corresponding generalized eigenvalue problem.X B is a reciprocal pencil B — n. f3 = O. (3 = O. there is a second eigenvalue "at infinity" for Case 3 of of .126 126 Chapter 12. Case 4: a = 0. Generalized Eigenvalue Problems Remark 12. All A E C are eigenvalues since det(A . There are two eigenvalues. zero.A. otherwise. 1 and 0. I1 and . There are two eigenvalues./. Clearly the reciprocal pencil has eigenvalues responding generalized /. If del (A . det(B . {3 =I. Case 4: a = 0. with its reciprocal eigenvalue being 0 in Case 3 of the reciprocal pencil B — /. only the case of regular pencils is considered in the remainder of this chapter. regular. Note that A and/or B may still be singular.XB) is not identically zero. Case 3: Case 4: = 0.L)({3 .0. ^ 0.3).O.AB) and there are several cases to consider. Case 1: a =I.0. With A and B as in (12.AB Definition 12. B k e n. Case 2: a = 0. A similar reciprocal symmetry holds for Case 2. There are two eigenvalues. ft =I.XB. There is only one eigenvalue. All A E C are eigenvalues since det(B . f3 = O. There is only one eigenvalue.(3A) ±. it is said to be singular. reciprocal Case of reciprocal .6.LA) = (1 . If det(A — AB) not regular. is singular. 1 and 0. and ~. Case 1: ^ 0.KB always has pencil — AB . If B is singular. and hence there are n eigenvalues associated with the pencil A . when B =I. I (of multiplicity 1). eigenvalues — AB. 1 and O.B) == O. suppose associated — AB. While While there are applications in system theory and control where singular pencils appear. {3 =I. pencil . k E !!. then rr(A) is a polynomial nonsingular).AHa . I). 1 (of multiplicity 1). Case = ft ^ 0. n(X) Remark 12. All A e C are eigenvalues since det(A — AB) =0. However. f3 / 0. =I. ft ^ O.O. At least for the case of regular pencils. If B = I (or in general when B is nonsingular). when B is singular. There are two eigenvalues.0.L = (JL = £. There are two eigenvalues.LA) == 0.nA. {3 ^ 0. (12.5. the pencil A . A similar reciprocal symmetry holds for Case 2. (3 = 0.A and corAssociated with any matrix pencil — AB is a reciprocal pencil ./.5. Associated with any matrix pencil A . Case 2: a = 0. That is to say.a/. All A 6 C are eigenvalues since det(B — uA) = O. Case 1: a ^ 0. There are two eigenvalues.LA. only the case of regular pencils is considered in the remainder of this chapter./. the pencil A — XB is said to be 12. Note that if AA(A) n J\f(B) ^ 0. There are two eigenvalues. If = of degree n. However. {3 =I.L) and there are again four cases to consider. Case 1: a =I. I and O. there may be 0.3) where a and (3 are scalars. I and ^. A — A. I multiplicity 1). Case 4: a = 0.
Sec.7. lencies rather than similarities. 12. Theorem 12.2 Canonical Forms Canonical Forms Just as for the standard eigenvalue problem. Sec. Let A. 6. Let A. the result follows. If B is nonsingular. Q. QBZ = TfJ .AB)Z] = det gdet Zdet(A 1.AQBZ are the same (the two problems problems are said to be equivalent). 6. the result follows easily by noting that yH(A — XB) — 0 if and only if yH (A . and eigenvectors under equivalence. [7. the pencil A AAB always has precisely n . Q~H y isa lefteigenvectorofQAZ — XQBZ. see. Since det Q XB). By Theorem 12. then Z~lx isa right eigenvector of QAZ—XQ B Z. canonical forms are available for the generalized Just as for the standard eigenvalue problem.7.7]. 7. and the first theorem deals with what happens to eigenvalues and eigenvectors under equivalence. see. det(QAZ . Then 12.7] or [25.7]. c 3. Canonical Forms 127 B is nonsingular.AB). However. det(QAZXQBZ) = det[0(A . B E Cnxn Then there exist unitary matrices Q. 6.AB are then the ratios of the diagBy Theorem 12. Let A. where Ta and TfJ are upper triangular. Proof: Proof: 1. Sec. 0 ( Q ~ H y)H Q(A X AB)Z = O.Oif andonly if Q(AXB)Z(Z~lx) = 0. 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 . the eigenvalues ofthe pencil A — XB are then the ratios of the diagonal elements of Ta to the corresponding diagonal elements of TfJ .7. and the first theorem deals with what happens to eigenvalues lencies rather than similarities. the result follows. with the understanding that a zero diagonal element of Tp corresponds to an infinite generalized eigenvalue.. the eigenvalues of the problems A — XB and QAZ — XQBZ are the same (the two 1.8. f i always has precisely eigenvalues. 2. canonical forms are available for the generalized eigenvalue problem. 7. Q. ifx isa right eigenvector of A—XB. 6. [7. for example. . 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.AB) o if and only if (QH y ) H Q ( A –_ B ) Z = Q. B.l W AW). the The first canonical form is an analogue of Schur's Theorem and forms.2.2. where Ta and Tp are upper triangular. Theorem 12. fewer than n eigenvalues. work directly on A and B are discussed in standard textbooks on numerical linear algebra. ify is a left of AB. Sec. Numerical methods that if B is even moderately ill conditioned with respect to inversion.7].AQBZ) = det[Q(A .7] or [25. the eigenvalues of the problems A . However. B e cnxn . 3. Z e Cnxn such that 12. of AAB. Z e Cnxn with Q and Z nonsingular. which is the generally preferred method for theoretical foundation for the QZ algorithm.12. for example. Then 1. in fact. Since the latter involves a pair of matrices. we now deal with equivaa matrices. There is also an analogue of the MurnaghanWintner Theorem for real matrices. Since det 0 and det Z are nonzero. for example. Sec. see. that a zero diagonal element of TfJ corresponds to an infinite generalized eigenvalue. fl. in fact. 7. Then there exist unitary matrices Q. this turns to the standard eigenvalue problem B~1Ax = Xx (or AB~1w = Xw). to ifx is a Zl x is a righteigenvectorofQAZAQB Z. [7. ify isa left eigenvector of A —KB. Sec. with the understanding onal elements of Ta to the corresponding diagonal elements of Tp.7.7] [25. then QHy isa left eigenvector ofQAZ AQBZ. Let A. which is the generally preferred method for solving the generalized eigenvalue problem. Canonical Forms 12.7]. Sec. E c nxn such that QAZ = Ta . E nxn with Q and nonsingular. the pencil A fewer than eigenvalues. see.l Ax Ax (or AB. since the generalized eigenvalue problem is then easily seen to be equivalent to the standard eigenvalue problem B.XB)Z] = detQ det Z det(A .AB and QAZ . D The first canonical form is an analogue of Schur's Theorem and forms. o.7] or [25. Numerical methods that work directly on A and are discussed in standard textbooks on numerical linear algebra. Again. the eigenvalues of the pencil A . the theoretical foundation for the QZ algorithm. solving the generalized eigenvalue problem. for example.8. and det Z are nonzero. 7. There is also an analogue of the MurnaghanWintner Theorem for real matrices.2 12. [7. 2. since the generalized eigenvalue problem is then easily seen to be equivalent eigenvalues.
There is also an analogue of the Jordan canonical form called the Kronecker canonical fonn Kronecker form (KeF)..)"N). quasiuppertriangular.11. Q € c nxn"such that nonsingular E C" such that peA . real eigenvalues. J .A [~ ~ l of .AB)Q = diag(LII' . Then there x exist nonsingular matrices P. while the full KeF in all its generality applies also to "rectangular" and singular KCF "rectangular" pencils. Example 12. B e Rnxn. of eigenvalues are given as above by the ratios of diagonal elements of S to corresponding elements of T. including analogues of principal vectors and description of of so forth. Otherwise. B e Cnxn and suppose the pencil A .12 (Kronecker Canonical Form). The first theorem pertains only to "square" regular pencils. B e c mxn . where T is upper triangular and S is quasiuppertriangular. 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. Z e R"xn such B E jRnxn. KCF. T.I.2)2 with characteristic polynomial (A — 2)2 has a finite eigenvalue 2 of multiplicty 2 and three 2 2 infinite eigenvalues.AB. Generalized Eigenvalue Problems Theorem 12. is beyond the scope of this book. L l" L~. Then there exist 12. QBZ = T. ..AB)Q = [~ ~ ] . [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 . thnt that QAZ = S. .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 . The matrix pencil 12. When S has a 2 x 2 diagonal block. Generalized Eigenvalue Problems Chapter 12. Then there exist orthogonal matrices Q. E jRnxn 12.11.9.AB where J is a Jordan canonical form corresponding to the finite eigenvalues of A A. B E c nxn pencil — AB Theorem 12.'.9.A. of — XB.128 Chapter 12.. . In this chapter. A full description of the KeF. Let A. I .• L. mxn E C • Theorem 12..XB is regular. Let A.12 mxm nxn mxm nxn E C nonsingular nonsingular matrices P e c and Q e c QE C such that peA .10. form (KCF). we present only statements of the basic theorems and some examples. Let A.
while each LQ has "zero rows" and L6. LQ. LQ. i. Lo. suppose S e Rn* xk is a matrix whose columns span a kdimensional E ~nxk ^dimensional subspace S of ~n. i.e. generalized eigenproblem. (12. 000 Just as sets of eigenvectors span Ainvariant subspaces in the case of the standard eigenvectors eigenproblem (recall Definition 9. 0. Let A.2. The second block is L\ while the third block is LI. Specifically. Lo.4) eigenvalue characterization Just as in the standard eigenvalue case. Lo.The next two blocks second block L\ one the block is L\. Then V is a E ~nxn suppose pencil — AB deflating subspace if deflating subspace if dim(AV + BV) = dimV. Left Left or right minimal indices can take the value O. there is an analogous geometric concept for the eigenproblem generalized eigenproblem. and L^ is the (k + I) x k where N is nilpotent.14. there is a matrix characterization of deflating subspace. LQ . Definition 12. both N and J are in Jordan canonical form.13.35).— XBif S Rn. (12.5) . n(S)) = S. 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. Then is deflating subspace for the pencil A AB if and only if there exists M E Rkxk such that e ~kxk AS = BSM..2. L6. are called the right minimal indices. R ( S <S. next two correspond to correspond J = 21 0 2 [ o 0 while the nilpotent matrix N in this example is N [ ~6~]. Canonical Forms 12.XB is regular.e. where each LQ has "zero columns" and one row.. both Nand J are in Jordan canonical form. Then SS is aadeflating subspace for the pencil A . Such a matrix is in KCF.12. Consider a 13 x 12 block diagonal matrix whose diagonal blocks are A 0] I o A I . The /( are called the left minimal indices while the r. Canonical Forms 129 where N is nilpotent. The first block of zeros actually corresponds to LQ. B e Wlxn and suppose the pencil A . corresponds LQ. Lo L6 one column. Example 12.
D=O. for example. however. trivial. Ac M D "'" 5A + 14. see. For details. u is the vector of inputs or controls. these values are the generalized eigenvalues of the (n + m) x (n m) pencil. we offer some insight below into the special case of a singleinput. The connection between system zeros and the corresponding system pencil is nonThe connection between system zeros and the corresponding system pencil is nontrivial. 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.8.5) becomes AS = SM as before.6)). the (finite) zeros of this system are given by the (finite) complex numbers In general. see. which has a root at —2. (n + m) x (n + m) pencil. If the pencil is not regular. then (12.15. Numerically.6) drops rank. and E jRPxm. vector. zeros).130 Chapter 12. = Cx + Du E jRnxn.4) becomes dim (A V + V) = dim V. For details. is a concept analogous to deflating subspace called a reducing subspace. E jRnxm. u is the vector of inputs or controls. where the "system pencil" (12. Checking the finite eigenvalues of the pencil (12. lEthe pencil is not regular. This linear timeinvariant statespace model is often used in multivariable control theory. we offer some insight below into the special case of a singleinput.15. then (12. [26]. where x(= x(t)) is called the state vector. In the special case p = m. However.6». Checking the finite eigenvalues of the pencil (12. In the special case p = m. one must be well for general mUltiinput.5) becomes AS = SM as before. Similarly. the (finite) zeros of this system are given by the (finite) complex numbers where the "system pencil" z. we which clearly has a zero at 2. C e Rpxn. This is accomcareful first to "deflate out" the infinite zeros (infinite eigenvalues of (12. we find the characteristic polynomial to be find the characteristic polynomial to be det [ which has a root at 2. This linear with A € M n x n . multioutput systems. B] . and D € Rpxm. for example.8. However. multioutput systems.6).3 12. which is clearly equivalent to AV c V. Similarly. B € R" xm . however. 12. Generalized Eigenvalue Problems If B = /.8. 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 —2.6). these values are the generalized eigenvalues of the drops rank. which is clearly equivalent to If B = I. (12.3 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. Example 12. there is a concept analogous to deflating subspace called a reducing subspace. one must be careful first to "deflate out" the infinite zeros (infinite eigenvalues of (12. where x(= x(t)) is called the state space model is often used in multivariable control theory. and y is the vector of outputs or observables. Let A=[ 4 2 C = [I 2].4) becomes dim(AV + V) = dimV. E jRPxn. In general. Numerically. [26]. 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).8. and y is the vector of outputs or observables. there AV ~ V. Let Example 12. (12.
9». . z is a zero of g. b e ffi. 0 from (12. Hence g(z) 0.9)). the problem (12.zl)x + by = 0. let B = b E Rn. and v(s) and n(s) are relatively prime TT(S) v(s) TT(S) (i.l xn. Now y ^ 0 (else x z i.7) we get get x = (A .4.e.10) is equivalent B.4. Thus.10) is equivalent Since B is positive definite it is nonsingular..A)~ ! Z? + d denote the system transfer function (matrix). Suppose z € is such that Suppose Z E C is such that [ A . Specifically. Now _y 1= 0 (else x = 0 from (12. no pole/zero cancellations).e. Then there exists a nonzero solution to or or (A .A to the standard eigenvalue problem Bl1Ax = AJC. system of differential equations differential Mx+Kx=O.nxn A AT and B the B1 0. However. "pole/zero cancellations"). we have _c T (A . the secondorder A.4 Symmetric Generalized Eigenvalue Problems Symmetric Generalized Eigenvalue Problems Ax = ABx A very important special case of the generalized eigenvalue problem (12. Symmetric Generalized Eigenvalue Problems 131 131 1 singleoutput system. 12.zI cT b ] d is singular. (12.zl)lby + dy = 0.. Thus.. or g ( z ) y = 0 by the definition of g. symmetric." is a frequently employed model of structures or vibrating systems and yields a frequently generalized eigenvalue problem ofthe form (12. For example.12. and D e R r T(s7 . then from (12. we have Substituting this in (12.n. M K where M is a symmetric positive definite "mass matrix" and K is a symmetric "stiffness definite "stiffness matrix.4 12.10) for A. the problem (12.e. there are no "pole/zero cancellations"). Hence g(z) = 0. C = c T E R l x n . e ffi. A pole/zero Assuming z is not an eigenvalue of A (i.zl)lby. (12. Symmetric Generalized Eigenvalue Problems 12.7) (12. or g(z)y 0 by the definition of g.10). B e ffi.8). relatively where n(s) is the characteristic polynomial of A.9) Substituting this in (12. g(s) Furthermore. and D = d E R.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. B~11A is not necessarily B~ Ax = AX.8) c T x +dy = O. g. let g(. of the Since B is positive definite it is nonsingular. B E Rnxn arises when A = A and B = BT > O.8).
are eigenvectors of the original generalized eigenvalue problem Xi Zi.11) can be rewritten as the equivalent problem 1 Letting C = L ~I AL ~T and z = L1 x. zn satisfying vectors Z I. . if A = AT> 0. so the eigenvalues are positive.1926 and —3. it has a Cholesky factorization B = LL T.18..5 ' 3.5 2. B E Rnxn with A = AT and B = BT > 0.23). we have restricted our attention to that case only.. then product y) x T By. with corresponding eigenSince C = C T. E !!.5 2. the eigenvalues are also all positive. (12.. zi Then x. then C = C T > 0. if orthogonal > 0. Then the eigenvalue problem Ax = ABx = ALL Tx (12.12) has n real eigenvalues..1926 and 3. so the eigenvalues are positive. we have restricted our attention to that case only. Finally. where L is nonsingular Proof: Since B > 0.. (12. Finally. = L ~Tzi.16).17. Example 12.16 is D 0 L=[~ . l = [i ~ J B ThenB~ A Then A B~Il = [~ ~ J B~I A approximately Nevertheless. Then the eigenvalue problem (Theorem 10.16.11) can then be rewritten as AL J and Z = LT x. ii € n. if A = A > 0. then = C T > 0. Theorem 12. Generalized Eigenvalue Problems Example 12. Zn Zj = Dij. where L is nonsingular (Theorem 10. Moreover. The material of this section can. .fi 1] . Then the generalized A. of course. but since realvalued matrices are commonly used in most applications.. . Let A. it has a Cholesky factorization B = LLT. the eigenvalue problem eigenvalue problem Ax = ABx has n real eigenvalues..1926 as expected.1926 whose eigenvalues are approximately 2. are eigenvectors of the original generalized eigenvalue problem and satisfy and satisfy (Xi. and are Hermitian.23).11) can then be rewritten as = Cz = AZ. Moreover...1926 in Example 12. the eigenproblem (12. Xj)B T T = xr BXj = (zi L ~l)(LLT)(L ~T Zj) = Dij. Generalized Eigenvalue Problems Chapter 12. generalized case A and B are Hermitian. be generalized easily to the case where A material of can.18. Let A = [~ .5 ] 1. Proof: Since B > 0.. Let A Example 12.132 132 Chapter 12. The Cholesky factor for the matrix B in Example 12. y)BB = XT By.fi Then it is easily checked that Then it is easily checked thai c = L~lAL~T = [ 0. the eigenvalues of B l A are always real (and are approximately 2.16.12) has n real eigenvalues. B e jRnxn A AT and B BT > O. but since realvalued matrices are commonly used in most applications. •. (12. The Cholesky factor for the matrix B in Example 12. and the n corresponding right eigenvectors can be chosen to be orthogonal with respect to the inner product (x.. positive. if A > 0. with corresponding eigenvectors zi.16 is Example 12.12) Since C = C T the eigenproblem (12.
A I < B~ . Also. In numerically problematic.e. There are many matrices (A.19 (Simultaneous Reduction to Diagonal Form). To illustrate.5 12. let such cases. This can be seen directly. Then there exists a nonsingular matrix Q such that where D is diagonal.e. Then A > B if and only if Bl1 > Theorem 12. we restrict our attention only to the real case. \ 2. B E E"x" with 12. = 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.19 is very useful for reducing many statements about pairs of symmetric matrices to "the diagonal case. B e M" xn be positive definite. But then D"1I :::: [(this is trivially true 10. so it does not preserve eigenvalues of A and B individually. normal maRecall that many matrices can be diagonalized by a similarity. Proof: By Theorem 12. Then diagonal.. where D is diagonal. Simultaneous Diagonalization 12. Thus. where D is C is symmetric. by Theorem 10.20. normal matrices can be diagonalized by a unitary similarity. In particular. it does preserve the eigenvalues of A . there exists Q E lR~xn such that QT AQ = D and QT BQ = [. by Theorem where D is diagonal. Simultaneous Diagonalization 133 12. let . i.. e. But then D.g. Then there exists a nonsingular matrix Q such that A = AT and B = BT > 0.1A). such results and we present only a representative (but important and useful) theorem here. we restrict our attention only to the real case. it does preserve the eigenvalues of A — XB.21 we have that QT AQ 2: QT BQ. since QDQ~l have A(D) = A(B~1A). Now D > 0 by Theorem 10. e. Theorem 12.e. there exists an orthogonal matrix P such that pTe p = D. matrices to "the diagonal case. Proof: Let B = LLT be the Cholesky factorization of B and set C = L~1AL T. Theorem 12. Thus. D 2: [. D > I.19. A~l :::: Bl1. B) can be simultaneously diagonalized by the same matrix. i. However.. It turns out that in some cases a pair of matrices (A. Let Q = L~T P.19 is very useful for reducing many statements about pairs of symmetric Theorem 12. Also. the diagonal elements of D are the eigenvalues of B. with the complex case following in a straightforward way. Since LLT be the Cholesky factorization of and setC L I AL~T. Q D.< / (this is trivially true 0 since the two matrices are diagonal). so it does not preserve eigenvalues of and B Note that Q is not in general orthogonal.1A. with the complex case following in a Again. Again.e. individually.12. B) can be simultaneously diagonalized by the same matrix. since A > B.l Q~T QT Q~ B~ AQ. Then A 2: B if and only if B~ 2: AI. Let A = QT AQandB = QT BQ.. This can be seen directly. when L is highly iII conditioned with respect to inversion.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. It turns out that in some cases a pair of trices can be diagonalized by a unitary similarity." The following is typical. Let A.5 Simultaneous Diagonalization Simultaneous Diagonalization Recall that many matrices can be diagonalized by a similarity.5.T P.1A = Q1l B~1QT QT AQ = Q11B. straightforward way. Proof: By Theorem 12. In such cases. we Note that Q is not in general orthogonal. there exists Q e E"x" such that QT AQ = D and QT BQ = I. There are many such results and we present only a representative (but important and useful) theorem here. simultaneous reduction can also be accomplished via an SVD. Infact. Let Q = L . haveA(D) = A(B.5. since QDQI Finally.. Then and and QT BQ Finally. where D is diagonal. Theorem 12.19." The following is typical. In fact. To illustrate. Then B. the diagonal elements of D are the eigenvalues of B 1A. Let A. B E lRnxn be positive definite. i.. where D is diagonal. QD~ QT < QQT. LetA QT AQ and B QT Then/HA Q~ B.1AQ. Since Proof: Let T C is symmetric. D since the two matrices are diagonal).19 is There are situations in which forming C = L~1AL~T as in the proof of Theorem 12.lI QT :::: Q QT. Let A. Now D > 0 by Theorem 10.5.5. since A 2: B. However.20.1 12.21 we have that QT AQ > QT BQ. we B~ 1 A.31. = pT L I(LLT)L T P = pT P = [.19 e ][~nxn A AT and B BT > O.31. simultaneous reduction can also be accomplished via an SVD. there exists an orthogonal matrix P such that P CP = D.19 is numerically problematic. A1. when L is highly ill conditioned with respect to inversion. In particular.'AB. i.g.
let A = LAL~ and B = LsLTB be Cholesky factorizations of A and B. example. see. (12. PDPT ~ ~ ~ ~ T PD(PD{ with where Disdiagonaland P is orthogonal. [7. respectively.21 are possible. without forming the products LALTA or LBLTB explicitly. which is thus to the generalized eigenvalue problem 02.134 134 Chapter 12. Generalized Eigenvalue Problems Chapter 12.3]. at least in real arithmetic. Various generalizations of the results in Remark 12. which is thus equivalent to the generalized eigenvalue problem ALBL~LBT z.15) The problem (12. Generalized Eigenvalue Problems us assume that both A and B are positive definite.butin writing A — PDDP T = PD(PD) with D is diagonal and P orthogonal. D b . Then the matrix Q == LLBTu performs the simultaneous diagonal. Further.e. The SVD in (12.13» via arithmetic operations performed only on LA LA (12. when A = AT > 0.. A can be written as A = PDP T.13) where E E R£ x " isisdiagonal. 8. D may have pure imaginary elements.e.21.13)) and LB separately. let A = LALTA and B — LBL~ us assume that both A and B are positive definite. To check this.14) can be rewritten in the form LALAx = XLBz = Letting x = LBT Z we see 02.7. This is analogous to finding the singular values of a matrix M by Sec.22. operations performed directly on M rather than by forming the matrix MT M and solving performed MT forming the eigenproblem MT MX = AX. Note that LB A and thus the singular values of L B 1 LA can be found from the eigenvalue problem 02. i.. Compute the SVD Cholesky factorizations A B. note that T QT AQ = U Li/(LAL~)Li/U = UTULVTVLTUTU i/ = while L2 QT BQ = U T LB1(LBL~)Li/U = UTU = I. respectively. Remark 12. Then the matrix Q U performs the simultaneous L e 1R~ xn diagonalization.15) is called a generalized singular value problem and algorithms exist to problem generalized solve it (and hence equivalently (12. eigenproblem MT M x Xx. The case when A is symmetric but indefinite is not so A = AT::: O. for LB i. for generalizations results 12.14) Letting x = LB z we see that (12. Further.14) rewritten the LAL~x = ALBz = A L g L ^ L g 7 z . Remark 12. A straightforward. products LA L ~ LBL~ see..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). Sec. For example.21 example. but in writing = PDDp D diagonal.
16) arises frequently in applications: 0. A special case of (12..12. and = = KT.2 K are are ± jWk. since eAt :F 0. Substituting in q(t) = eAt p.16) Consider the secondorder system of differential equations Consider the secondorder system of differential equations q(t) E ~n E ~nxn.16) we get (12. n. then all solutions of q K q 0 are oscillatory. for which the matrix A. are to be determined. r.. or if it is desired to avoid the calculation of M lI because M is too ill conditioned with respect to inversion.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.e. = [ M1K 0 x (t) E ~2n.6 HigherOrder Eigenvalue Problems HigherOrder Eigenvalue Problems Mq+Cq+Kq=O. and A special case of (12. Assume for simplicity that M is nonsingular.. ::: ILn· Let a>k = IILk I!. the secondorder problem (12. .. M Mwhere x(t) €. k = 1. there are 2n eigenvalues for the secondorder (or A2 M + AC + K. KT > 0). If r n (i...1 12. . (12. then all solutions of q + Kq = 0 are oscillatory.2M + A. where the nvector p and scalar A.6.C + K.6. (12. k = r + 1. HigherOrder Eigenvalue Problems 135 12. we thus seek values of A.. p.16) can be written as a firstorder system (with block Let XI q and X2 Then (12.6. ± Wk. yields a polynomial of degree 2rc. K = KT ::: 0)..16) of the p A are to be determined. polynomial 2n. 12. Suppose K = KT. quadratic) eigenvalue problem A. C = 0.6 12. by analogy with the firstorder case.• 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. .. E2". K e Rnxn.. . HigherOrder Eigenvalue Problems 12. that we try to find a solution of (12. seek A A2 M + AC + To get a nonzero solution /?.e.1 Conversion to firstorder form Conversion to firstorder form Let x\ = q and \i = q.16) arises frequently in applications: M = I. where q(t} e W1 and M. Since the determinantal equation o = det(A 2 M + AC + K) = A2n + . ::: ILr ::: 0 > ILr+ I ::: .6. If M is singular. Suppose.C + K is singular.16) or..16) can be written as a firstorder system (with block companion matrix) X . If r = n (i. Since the determinantal equation is singular. Suppose K has eigenvalues eigenvalues IL I ::: . (A 2 M + AC + K) p = O. .2M + A. C. Substituting in form q(t) = ext p.. Then (12.
19). derivative q. 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. and C e lRmxn. Show that the nonzero eigenvalues of FG and GF are the same. In the parlance of control theory. .19). and/or K have special symmetry or skewsymmetry properties that can exploited. properties Higherorder analogues of (12. say. such results show that zeros are invariant under state feedback or output injection.1 2. E Rnxm and E E 4.136 136 Chapter 12.. B e lRn*m. Similar procedures hold for the general kthblock companion matrix analogue of (11. Suppose A e Rnxn and D E lR::! xm.. andlor K Many other firstorder realizations are possible. Generalized Eigenvalue Problems Chapter 12. C. Let F e Cnxm . Suppose A € Rnxn. 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 . G e Cmxn • Are the nonzero singular values of FG and GF the same? same? wx E ]Rnxn. lead naturally naturally involving. Let € C M X • Show that the nonzero eigenvalues of and G F are the same. Let F. Are the FG and GF the 3. In the parlance of control theory.16) involving. G E enxn". Some can be useful when M. F 6 Rm *" G R" x . Hint: Consider the equivalence I G][AUO F0]' B][I l [01 C (A similar result is also true for "nonsquare" pencils. Similar procedures hold for the general k\horder difference equation order difference equation which can be converted to various firstorder systems of dimension kn. the kth derivative of q.) . EXERCISES EXERCISES nx 1.B D. (A similar result is also true for "nonsquare" pencils. Show that the generalized eigenval". Some can be useful when M. 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. 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 ". which can be converted to various firstorder systems of dimension kn. Generalized Eigenvalue Problems Many other firstorder realizations are possible.
Consider the case where both A and transformation contragredient. respectively. positive. A and B to the same diagonal matrix. and let UWT be an SVD of L~LA'.Exercises Exercises 137 137 desired 5. A B B are positive definite with Cholesky factorizations A = L<A and B = L#Lg. Such QT BQ a transformation is called contragredient. 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. 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. (c) Show that the eigenvalues of A B are the same as those of 1.2 and hence are AB E2 positive. Ql = ~!UTL~. B E e jRnxn Ql AQT ]Rnx" in such a way that Q~l AQ~T and QT BQ are simultaneously diagonal. (b) Show that Q~l = ^~^UT LTB. respectively. positive Cholesky = LA L ~ = L B L ~. .
This page intentionally left blank This page intentionally left blank .
Let B be an arbitrary 2 x 2 matrix. Then 0 b ll b12 B @/z = l b" b~l 139 0 b2 2 0 b21 0 0 b12 0 b 22 l . 1. Let B be an arbitrary 2x2 matrix. B e lR pxq.1. / 2 <8>fl = [o ~ l\ 2. 4 3 4 3 4 9 4 2 6 2 6 6 6 2 2 Note that B @ A i. Note that B <g> A / A <g> B. Let A e R mx "..1. 2B 2B ~J. Then 3.1 13. Let A E lRmxn B E R Definition 13. extension to the complex case only where it is not obvious. the same definition holds if A and B are complexvalued 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. pointing out the restrict our attention in this chapter primarily to realvalued matrices. Then A@B =[ 3~ ~]~U J.. We restrict our attention in this chapter primarily to realvalued matrices.. the same definition holds if A and B are complexvalued matrices.1 Definition and Examples Definition and Examples Definition 13.A @ B. Let A = [~ 2 2 nand B = [. (13. We Obviously. Example 13.2.Chapter 13 Chapter 13 Kronecker Products Kronecker Products 13. Foranyfl E lRX(7. n 2.2. pointing out the extension to the complex case only where it is not obvious. 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. Example 13.1) amnB Obviously.
Then 13.6. Proof: Proof: Using Theorem 13.5. Let Jt € Rm.3. 0 . B e ~rxs. B In x E ~m. then A® B is symmetric. L~=l al. xmYnf E !R. 5. y eR". 5 E R r x i .3. XIYn. . E R"... If AI ® B. .. = 1 ® 1 = I.1. 4. . Kronecker Products Kronecker Products The extension to arbitrary B and /„ is obvious.5. XmY T]T = [XIYJ.m xm are symmetric.. Let E ~mxn. and D e Rsxt.2 13. simply verify using the definitions of transpose and Kronecker verify transpose Kronecker 0 product. . If A and B are nonsingular. Then X ® Y = [ XIY T .kCkPBD L~=1 amkckpBD ] 0 Theorem 13. simply note that (A ® B)(A 1 ® B.3. (13. .. A® 13.6. . mn .2 Properties of the Kronecker Product Properties of the Kronecker Product (A 0 B)(C 0 D) = AC 0 BD (E ~mrxpt). Theorem 13. (A ® B)I = Bare 13.1 ) Theorem 13. Then 13. For all A and B. C e ~nxp.n. (A ® Bl = AT ® BT. Simply verify that ~[ =AC0BD.3. y e !R. Let* eR m . Let A e R mx ". If E ]Rn xn e Rmxm are Theorem 13. Foral! Proof' Proof: For the proof. X2Yl.2) Proof: Simply verify that Proof.4.. C E R" x ^ and D E ~sxt. D Corollary 13.140 Chapter 13. If A e R"xn and B E !R.
c.. if A and fi have Jordan form . then A 0 B is normal.. then . 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).3.10. i / E e!!. If A e IR nxn am/ B E IR mxm are normal. If A E E"xn is orthogonal and B E Mmxm is orthogonal.. . q Corollary 13.... . Let A G IR mx " have a singular value decomposition l/^E^Vj an^ let and /ef singular decomposition UB^B^BB e IR pxq fi E ^pxq have a singular value decomposition V B ~B VI./u. elements of ~A 0 ~B and the corresponding right and left singular vectors)..2. In general.10.Zq are linearly independent right eigenvectors of B corresponding to JLI... \Ju (q ::::: m).7.•.p (p < n).12.. A0 B e±jeH</» e±jefJ </». then A® B is normal. Properties of the Kronecker Product Theorem 13. :::: U rTs > 0 and ^iT\ > • • • > ffr <s Qand rank(A 0 B) = (rankA)(rankB) = rank(B 0 A) . if Xl. 7 E m. In general.m are linearly independent right corresponding to JJL\ ...• :::: U rr > 0 and let B E IRfx Corollary e R™x" singular a\ > • • > a > e have singular values T\ > • • > <s > 0.[Cos</> cos</>O Then It IS easl'1y seen that . = (A 0 B)(A 0 B)T 0 Corollary 13.12.13. <I :::: . .• 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.and let BB E e IRR mxwhave e IR nxn have eigenvalues A.11. mxm /zave Theorem 13. Ap (p ::::: and ZI.. . if x\. If Corollary 13. Let A E lR. eigenvectors of A® B corresponding to A. Lgf A E E mxn have a singular value decomposition VA ~A Theorem 13.. 0 Zj E€ IR mn "are linearly independent right eigenvectors of A 0 B corresponding to Ai JL 7 i e /?. ••.8. L et A E xamp Ie 139 Let A = [ _eose cose andB . TTzen ?/ze mn eigenvalues of A 0 Bare Moreover. matrix A ® 5 is then also orthogonal with eigenvalues e^'^+'W and e ± ^ (6> ~^ > \ Theorem 13.2. and zi. Example 13. then A <g> B is € IR nxn orthogonal and e IR m x m 15 then 0 is orthogonal.. xp are linearly independent right eigenvectors of A corresponding AI.j.. 0 If A and Bare diagonalizable in Theorem 13. If A E IR"xn and B eRmxm are normal.."xn have singular values UI :::: . Let A E R nx "have eigenvalues Ai. Properties of the Kronecker Product 13.i . .3 since A and B are normal by Theorem 13. .. then Xi <8> Zj ffi. Then A <g)B (or B 0<8> A) has rs singular values U. The 4 x 4 orthogonal e±j9 orthogonal eigenvalues e±j(i>. . xp are linearly independent right eigenvectors of A corresponding Moreover.8. .n..4 by Theorem 13.. . Then the mn eigenvalues of A® B are eigenvalues JL j.. Then A 0 B (or B A) has rs singular values have singular values <I :::: . Sine] and B .9.• :::: TS > O. 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). we can take p = nand q = m and n and q —m and If A and B are diagonalizable in Theorem 13.JLqq (q < m). Theorem 13... 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. • • zq independent of to A . we can take p thus get the complete eigenstructure of A 0 B. j e q. and let eigenvalues jJij.7..12. if A and B have Jordan form thus get the complete eigenstructure of A <8> B. A. j € m... .
is generally not quite in Jordan form and needs Note that JA® JB. Tr(A ® B) = (TrA)(TrB) = Tr(B ® A). Then 13. Example 13. in general. in of A and B.1 AP) ® (Ql BQ) = JA ® JB · Note that h ® JR. suppose P and Schur form for A ® B can be derived similarly. Kronecker Products Chapter 13. pH AP = TA and QH BQ = TB (and similarly if and are orthogonal similarities PHAP = TA and QHBQ = TB (and similarly if P and Q are orthogonal similarities reducing A and B to real Schur form). to Schur (triangular) form. denoted A © B.13. 2. Then reducing A and B to real Schur form). Let A e Rn xn and B e Rrn xm. respectively. Corollary 13.13. i. respectively. Let 1. then we get the decompositions given by P~lI AP = J A and Ql BQ = JB.e.AP J B . 1. Let A e Rn Xn and B e Rm xrn. eigenvalues are zero or nonzero).14. while upper triangular.15. Then the Kronecker sum (or tensor sum) . Let A~U Then Then 2 2 !]andB~[ . is generally not quite in Jordan form and needs further reduction (to an ultimate Jordan form that also depends on whether or not certain further reduction (to an ultimate Jordan form that also depends on whether or not certain eigenvalues are zero or nonzero). A Schur form for A ® B can be derived similarly.I ® Ql)(A ® B)(P ® Q) = (P. Example 13. is the mn x mn matrix Urn <g> A) + (B ® In). Then (P ® Q)H (A ® B)(P ® Q) = (pH ® QH)(A ® B)(P ® Q) = (pH AP) ® (QH BQ) = TA ® TR . respectively. Note that.142 142 Chapter 13. For example. A EEl B ^ B EEl A. with A EEl B. E IR E IR Kronecker Definition 13.14. denoted A EEl B. is the mn mn matrix (Im ® A) + (B ® /„). det(A ® B) = (det A)m(det Bt = det(B ® A). of A and B. . 1. For example.. i. ~l 2 2 1 3 AfflB = (h®A)+(B®h) = 1 3 0 1 0 4 0 3 0 0 0 0 0 0 0 0 0 0 0 0 2 2 0 0 0 3 4 2 0 0 2 0 0 2 0 0 2 0 0 0 1 0 0 + 0 2 0 0 2 0 0 0 0 3 0 0 0 3 0 0 0 3 The reader is invited to compute B 0 A = (/3 ® B) + (A 0 h) and note the difference The reader is invited to compute B EEl A = (h ® B) (A <g> /2) and note the difference with A © B. Kronecker Products decompositions given by p. E IR nxn E IR mxm. nxn mxm Definition 13. Note that. suppose P and Q are unitary matrices that reduce A and B.15. while upper triangular. respectively. general. then we get the JA and Q~] BQ following Jordanlike structure: following Jordanlike structure: (P ® Q)I(A ® B)(P ® Q) = (P. A ® B i= B © A. are unitary matrices that reduce A and 5.. to Schur (triangular) form.e.
···. . then Zj ® Xi E€ jRmn" are linearly independent right Zj <8> Xi W1 are linearly independent right corresponding f j i . .2. ... .. zq are linearly independent eigenvectors of corresponding to fJt. . Zq are linearly independent right eigenvectors of B AI..i e n... + fJj' € p. . .xp are linearly independent right eigenvectors of A corresponding Moreover. . In general.. Ap (p < and ZI. (I} ® M) + (E^®l2) = M 0 Ek. and let B E Rmx'" have e jRnxn eigenvalues A. we can take p nand q and If A and B are diagonalizable in Theorem 13.2. 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). if XI. . fJq (q < m).. . A2 + fJt. i E !!.. . is a Jordanlike structure for A © B. if A and have Jordan form thus get the complete eigenstructure of A 0 B. . Then the Kronecker sum A® B = (1m (g>A) + (B ® In) has mn (Im ® A) + (B <g> /„) /za^ ran eigenvalues fJj. Properties of the Kronecker Product 2.•• . j e q..... . eigenvectors of A® B corresponding to Ai + [ij. 0 If A and Bare diagonalizable in Theorem 13. if A and B have Jordan form pI l B . ii E E. . Xp (p ::s: n).16. 0 I M 0 where M = [ where M = o M a f3 f3 a J.. f^q (q ::s: ra). respectively.. Recall the real JCF M I M 143 143 0 I M I 0 o 1= 0 E jR2kx2k. e jRmxm eigenvalues /z.. Recall the real JCF 2.16. respectively. and z\. An + fJm' Moreover. A2 + fJm. if x\. we can take p = n and q = m and thus get the complete eigenstructure of A $ In general. j E fl· eigenvectors of A $ B corresponding to A. . 7 e I!!. 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. Then J can be written in the very compact form J Theorem 13.. Properties of the Kronecker Product 13.. AI + fJm. then decompositions given by P~1AP = lA and Q"1 BQ = JB..13. TTzen r/ze Kronecker sum A $ B eigenvalues e/genva/wes Al + fJt.. . Let A E E"x" have eigenvalues Ai. . j E ra. .16. 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.\ .
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. [(Q ® /„)(/« ® P)] = (<2 ® P) is unitary by 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).3) is.5) as an "ordinary" linear system. B e Rmxm . When C is symmetric. solution e IR xn also to be symmetric and (13. . 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. Lyapunov equations also to be symmetric and (13. This equation is now often called a Sylvester equation is now often equation in honor of 1.3 and Corollary 13. The first important question to ask regarding (13. Kronecker Products Chapter 13. Kronecker Products A Schur fonn for A EB B can be derived similarly.8.5) [ blml The coefficient matrix in (13. = AXi + l:~>j. .3) is the symmetric equation AX +XAT = C (13. Sylvester where A e R"x". The following definition is very helpful in completing the writing of (13. arise naturally in stability theory.1.XB. 13. pH AP = TA matrices that reduce A and B.5) clearly can be written as the Kronecker sum (Im * A) + (BT ® In). ® P)] = (/m <8> rA) + (7* (g) /„).4) is known as a Lyapunov equation..e. Sylvester who studied general linear matrix equations of the form equation in honor of J. Lyapunovequations arise naturally in stability theory. When does a solution exist? The first important question to ask regarding (13. i. =C. Again.e.3) mxm E IRnxn E IR E IRnxm. The following definition is very helpful in completing the writing of (13.144 Chapter 13. (13. i.. When symmetric. suppose P and Q are unitary fonn. Sylvester who studied general linear matrix equations of the fonn k LA.8.3 and Corollary 13.Xj.3 13. an "ordinary" linear system.=1 A special case of (13.3) in tenns of their columns. and C e M" xm . When does a solution exist? By writing the matrices in (13.J. where [(Q <8>In)(lm ® P)] = (Q ® P) is unitary by Theorem 13.3) in terms of their easily seen z'th columns that ith columns that m AXi + Xb. respectively. Then to real Schur fonn). the solution X E Wnx" is easily shown taking B = AT. .3) is. Again.5) as (B T 0 /„).4) obtained by taking B = AT.5) clearly can be written as the Kronecker sum (1m 0 A) + The coefficient matrix in (13. Then ((Q ® /„)(/« ® P)]"[(/m <8> A) + (B ® /B)][(e (g) /„)(/„.4) is known as a Lyapunov equation. suppose P and are unitary A Schur form for A © B can be derived similarly. = C. it is easily seen by equating the writing (13. to Schur (triangular) form..
B have no eigenvalues in common. where From Theorem 13. +00): IHoo lim XU) .24. j j E!!!. +00): (with X(0) = C) on [0. There exists a unique solution to (13.6) if and only if [(1m ® A) + (B T ® In)] is nonsingular.. Ai E A(A). .6) if and only if [(Im ® A) + (BT ® /„)] is nonsingular. The most commonly preferred numerical algorithm is described in [2]. this algorithm takes only O(n3 ) operations rather than the O(n6)) that would be required by solving (13.18.6) There exists a unique solution to (13.e A (A).17.8) can be written as can be written as (13. 77ie/i Theorem 13. We thus have the following theorem. + IJLJ.8)by Theorem 13. Schur form..10) . e m.3. 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. AX+XB=C (13. Let Ci( € E.17. Sylvester equations of the form (13. one of many elegant connections between matrix theory and stability theory for differential equations. Then the Sylvester equation G jRmxm. . Theorem C E jRnxm.5) can be rewritten in the form [(1m ® A) + (B T ® In)]vec(X) = vec(C). Now integrate the differential equation X AX XB (with X(O) C) on [0. n > m. Cm}.17.18..18. elegant connections between matrix theory and stability theory for differential equations.4» are generally not solved using the mn x mn "vec" formulation (13.3. But [(Im <8>A) + (B TT ® In)] isisnonsingular ififand only ififitithas no zero eigenvalues. E jRmxm. ii e n_. the linear system (13. xn Theorem 13. say. n :::: m.6) directly with Gaussian elimination. E R E jRnxm. Assuming that. ofC e jRnxm [CI. Application to Sylvester and Lyapunov Equations 13. (13. the linear system (13.4)) are generally not solved using the mn x mn "vee" formulation (13. Then the (unique) solution of the Sylvester equation parts in the open left halfplane). 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). Suppose further that A and B E Rn .9) Proof: Since A and B are stable. Application to Sylvester and Lyapunov Equations 145 145 Definition 13. E!!.6). Assuming that. Definition 13. First A and B are reduced to commonly preferred numerical algorithm is described in [2].6). so there exists unique Proof: Since A and B are stable.3) (or symmetric Lyapunov equations of the form (13. c ].17. Then the (unique) solution of the Sylvester equation AX+XB=C (13.. say. Let A e jRnxn.7) has a unique solution if and only if A and . 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. vec(C) = Using Definition 13.. j j so there exists aaunique for all i.8) by Theorem 13.19.e. one of many The next few theorems are classical. . (13.X(O) = A 10 roo X(t)dt + ([+00 X(t)dt) 10 B. where A.16. the eigenvalues of [(1m ® A) + (BT <8> /„)] are + Mj.24. i. and ^j Theorem 13.16. and C e Rnxm. A(fi). A further enhancement to this algorithm is available in [6] whereby 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 the larger of A or B is initially reduced only to upper Hessenberg rather than triangular Schur form. A. From Theorem 13.and Mj Ee A(B).(B) ^ solution to(13. has a unique solution if and only if A and —B have no eigenvalues in common.. We thus have the following theorem.. But [(1m ® A) + (B (g) /„)] nonsingular and only has no zero eigenvalues.6) directly with operations rather than the O(n 6 that would be required by solving (13. They culminate in Theorem 13. this algorithm takes only 0 (n 3) transformed solution matrix X.5) can be rewritten in the form Using Definition 13.n denote the columns ofC E Rnxm so that C = [ n .13. the eigenvalues of [(/m <g> A) + (BT ® In)] are Ai A. They culminate in Theorem 13. (A)+ Aj(B) =I 00 for all i. The most (13. Let A e lRnxn.1S. ..3) (or symmetric Lyapunov equations of the form Sylvester equations of the form (13. Now integrate the differential equation X = AX + X B solution to (13. First A and B are reduced to (real) Schur form. B e Rmxm. (real) Schur form. and C e R" xm . Aj(A) + A. . The next few theorems are classical. B E Rmxm.
C E R"x".19.23 a solution to (13. . 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. Then the (unique) solution o/the Lyapunov equation of the AX+XAT=C can be written as can be written as (13. then that solution is symmetric. . the first of which follows immediately from Theorem 13. then that solution is symmetric. Remark 13.A T have no eigenvalues in common. . v E". it can be shown easily that lim elA = lim elB = O. Remark 13. If the matrix A E Wxn has eigenvalues A. ..11) has a unique solution if and only if A and . results = 0. Many useful results exist concerning the relationship between stability and Lyapunov equations.6. If matrix A e jRn xn eigenvalues )"" .12). Thus.11) has a unique solution.. Kronecker Products Chapter 13.C E R"x" and suppose further that A is asymptotically stable. we have that lim X ((t) = 0. If C is has unique if and only if and —A T eigenvalues in common. 1>+00 1 .11) has a unique solution. .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 (13.146 146 Cha