P. 1
Matrix Analysis for Scientists and Engineers by Alan J Laub

Matrix Analysis for Scientists and Engineers by Alan J Laub

|Views: 1,930|Likes:

More info:

Published by: Julio Cesar Barraza Bernaola on Nov 24, 2010
Copyright:Attribution Non-commercial

Availability:

Read on Scribd mobile: iPhone, iPad and Android.
download as PDF, TXT or read online from Scribd
See more
See less

04/25/2013

pdf

text

original

Sections

  • Introduction and Review
  • 1.1 Some Notation and Terminology
  • 1.2 Matrix Arithmetic
  • 1 .2 Matrix Arithmetic
  • 1.3 Inner Products and Orthogonality
  • 1.4 Determinants
  • Vector Spaces
  • 2.1 Definitions and Examples
  • 2.2 Subspaces
  • 2.3 Linear Independence
  • 2.4 Sums and Intersections of Subspaces
  • 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
  • 4.1 Definitions and Characterizations
  • Introduction to the Moore-Penrose Pseudoinverse
  • 4.2 Examples
  • 4.3 Properties and Applications
  • Introduction to the Singular Value Decomposition
  • 5.1 The Fundamental Theorem
  • 5.2 Some Basic Properties
  • 5.3 Row and Column Compressions
  • 5.3 Rowand Column Compressions
  • 6.1 Vector Linear Equations
  • Linear Equations
  • 6.2 Matrix Linear Equations
  • 6.3 A More General Matrix Linear Equation
  • 6.4 Some Useful and Interesting Inverses
  • Projections, Inner Product Spaces, and Norms
  • 7.1 Projections
  • 7.1.1 The four fundamental orthogonal projections
  • 7.1 .1 The four fundamental orthogonal projections
  • 7.2 Inner Product Spaces
  • 7.3 Vector Norms
  • 7.4 Matrix Norms
  • 8.1 The Linear Least Squares Problem
  • Linear Least Squares Problems
  • 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
  • 9.1 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 +1 's in JCF blocks
  • 9.4 Geometric Aspects of the JCF
  • 9.5 The Matrix Sign Function
  • 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
  • Linear Differential and Difference Equations
  • 11.1.1 Properties of the matrix exponential
  • 11.1 Differential Equations
  • 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
  • 11.1 .5 Modal decompositions
  • 11.2.1 Homogeneous linear difference equations
  • 11.2.2 Inhomogeneous linear difference equations
  • 11.2 Difference Equations
  • 11.2.3 Computation of matrix powers
  • 11.3 Higher-Order Equations
  • Generalized Eigenvalue Problems
  • 12.1 The Generalized Eigenvalue/Eigenvector Problem
  • 12.2 Canonical Forms
  • 12.3 Application to the Computation of System Zeros
  • 12.4. Symmetric Generalized Eigenvalue Problems 131
  • 12.4 Symmetric Generalized Eigenvalue Problems
  • 12.5 Simultaneous Diagonalization
  • \ 2.5.1 Simultaneous diagonalization via SVD
  • 12.5.1 Simultaneous diagonalization via SVD
  • 12.6. Higher-Order Eigenvalue Problems 135
  • 12.6 Higher-Order Eigenvalue Problems
  • 12.6.1 Conversion to first-order form
  • Kronecker Products
  • 13.1 Definition and Examples
  • 13.2 Properties of the Kronecker Product
  • 13.3 Application to Sylvester and Lyapunov Equations
  • Bibliography
  • Index

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 19104-2688.
MATLAB® is a registered trademark of The MathWorks, Inc. For MATLAB product information,
please contact The MathWorks, Inc., 3 Apple Hill Drive, Natick, MA 01760-2098 USA,
508-647-7000, Fax: 508-647-7101, 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 Cataloging-in-Publication Data
Laub, Alan J., 1948-
Matrix analysis for scientists and engineers / Alan J. Laub.
p. cm.
Includes bibliographical references and index.
ISBN 0-89871-576-8 (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 19104-2688.
MATLAB® is a registered trademark of The MathWorks, Inc. For MATLAB product information,
please contact The MathWorks, Inc., 3 Apple Hill Drive, Natick, MA 01760-2098 USA,
508-647-7000, Fax: 508-647-7101, 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 Cataloging-in-Publication Data
Laub, Alan J., 1948-
Matrix analysis for scientists and engineers / Alan J. Laub.
p. cm.
Includes bibliographical references and index.
ISBN 0-89871-576-8 (pbk.)
1. Matrices. 2. Mathematical analysis. I. Title.
QA 188.L38 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 .

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 Moore-Penrose 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 Moore-Penrose 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 Higher-Order 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 Higher-Order 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 Higher-Order Eigenvalue Problems 135
12.6.1 Conversion to first-order 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 Higher-Order Eigenvalue Problems ..
12.6.1 Conversion to first-order 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 graduate-level (or even senior-level)
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 one-quarter or one-semester 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., basis-free 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 well-equipped to pursue, either via formal courses or through self-
study, follow-on 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 "out-of-order" 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 graduate-level (or even senior-level)
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 one-quarter or one-semester 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., basis-free 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 well-equipped to pursue, either via formal courses or through self-
study, follow-on 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 "out-of-order" 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, "real-life" problems seldom yield to simple
closed-form formulas or solutions. They must generally be solved computationally and
it is important to know which types of algorithms can be relied upon and which cannot.
Some of 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 research-level questions when set
in the context of the finite-precision, finite-range floating-point arithmetic environment of
most modern computing platforms.
Some of the applications of matrix analysis mentioned briefly in this book derive
from the modern state-space approach to dynamical systems. State-space 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 state-space theory) to an appendix or introducing it "on-the-fly" 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, "real-life" problems seldom yield to simple
closed-form formulas or solutions. They must generally be solved computationally and
it is important to know which types of algorithms can be relied upon and which cannot.
Some of 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 research-level questions when set
in the context of the finite-precision, finite-range floating-point arithmetic environment of
most modem computing platforms.
Some of the applications of matrix analysis mentioned briefly in this book derive
from the modem state-space approach to dynamical systems. State-space 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 state-space theory) to an appendix or introducing it "on-the-f1y" 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 n-tuples 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 n-tuples of complex numbers represented as column vectors.
3. R
mxn
= the set of real (or real-valued) 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 complex-valued) 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 n-tuples 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 n-tuples of complex numbers represented as column vectors.
3. IR
rn
xn = the set of real (or real-valued) 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 complex-valued) 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 real-valued while a statement
like A = A
H
implies that A is complex-valued.
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 complex-valued 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 real-valued while a statement
like A = AH implies that A is complex-valued.
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 complex-valued 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 matrix-vector 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 matrix-vector product can also be computed
v1a
For large arrays of numbers, there can be important computer-architecture-related 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 matrix-vector 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 matrix-vector 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 matrix-vector product can also be computed
via
3.[ J+2.[ J+l.[ l
For large arrays of numbers, there can be important computer-architecture-related 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 matrix-vector 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 skew-symmetric, 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 - BD-Ie 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(l-xyT) = 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 skew-symmetric, 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 finite-dimensional vector spaces, including spaces formed by special classes
of matrices, but some infinite-dimensional 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 finite-dimensional vector spaces, including spaces formed by special classes
of matrices, but some infinite-dimensional 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 a-I E IF such that a . a-I = 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)=a-v + 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 jc-coordinate in the plane and u
2
with the y-coordinate. 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 x-coordinate in the plane and V2 with the y-coordinate. 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 n-tuple {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 n-dimensional 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 n-tuple ... , 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 finite-dimensional if there exists a basis X with n < +00 elements;
otherwise, V is infinite-dimensional.
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 finite-dimensional if there exists a basis X with n < +00 elements;
otherwise, V is infinite-dimensional.
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\ = r-i 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 skew-symmetric 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 right-hand 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 skew-symmetric 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:(A-A).
The first matrix on the right-hand 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 co-domain.
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 co-domain.
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-(t-r)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 = pn-l.
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 rank-one matrix A e R
mxn
can be written in the form A = xy
T
above (or xy
H
if A e C
mxn
). A rank-one 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 rank-one matrix A E ~ m x n can be written in the form A = xyT
above (or xyH if A E c
mxn
). A rank-one 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 finite-dimensional vector spaces.)
2. 'R,(A)
1
~ — J\f(A
T
). (Note: This also holds for infinite-dimensional 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 co-domain 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 co-domain 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
=XTVj-XTVj=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 finite-dimensional vector spaces.)
2. R(A).l = N(A
T
). (Note: This also holds for infinite-dimensional 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 co-domain 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 co-domain 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 one-to-one or 1-1 (also called monic or infective) ifJ\f(A) = 0. Two equivalent
characterizations of A being 1-1 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
n-r 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 one-to-one or 1-1 (also called monic or injective) if N(A) = O. Two equivalent
characterizations of A being 1-1 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 1-1 (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 1-1 (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 1-1 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 1-1 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 1-1, 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
1-1. D
Definition 3.23. A : V —» W is invertible (or bijective) if and only if it is 1-1 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 1-1 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 I-I 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 1-1 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 1-1, 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
1-1. D
Definition 3.23. A : V -+ W is invertible (or bijective) if and only if it is 1-1 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 A-I
are all 1-1 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 1-1.
Moreover, A is invertible if and only if it is both right and left invertible, i.e., both 1-1 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 1-1, 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 1-1.
Moreover, A is invertible if and only if it is both right and left invertible, i.e., both 1-1 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 1-1, 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
+ A-RA -I) = AA-
R
+ AA-RA - 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 1-1. (Proof: The only solution to 0 = Av = [
I
2
]v
is v = 0, whence A/"(A) = 0 so A is 1-1). 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 1-1 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 skew-symmetric 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 £ 1-1? Is£
onto?
4. Prove Theorem 3.4.
of R
3
and
of E
2
.
Exercises 27
2. LetA = [i]:]Rl ~ ]R2. Then A is 1-1. (Proof The only solution toO = Av = [i]v
is v = 0, whence N(A) = 0 so A is 1-1). 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 1-1 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 skew-symmetric 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 £, I-I? 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
Moore-Pen rose
Pseudoinverse
In this chapter we give a brief introduction to the Moore-Penrose pseudoinverse, a gener-
alization of the inverse of a matrix. The Moore-Penrose 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 (1-1 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 Moore-Penrose 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 Moore-Penrose
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
Moore-Penrose
Pseudoinverse
In this chapter we give a brief introduction to the Moore-Penrose pseudoinverse, a gener-
alization of the inverse of a matrix. The Moore-Penrose 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 Cl-l 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 Moore-Penrose 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 Moore-Penrose
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 Moore-Penrose 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 1-1 (independent columns) (A is left invertible).
Example 4.7. For any scalar a,
30 Chapter 4. Introduction to the Moore-Penrose 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 1-1 (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 Moore-Penrose Pseudoinverse
Note that by combining Theorems 4.12 and 4.13 we can, in theory at least, compute
the Moore-Penrose pseudoinverse of any matrix (since A A
T
and A
T
A are symmetric). This
turns out to be a poor approach in finite-precision arithmetic, however (see, e.g., [7], [11],
[23]), and better methods are suggested in text that follows.
Theorem 4.11 is suggestive of a "reverse-order" 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 reverse-order 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 Moore-Penrose Pseudo inverse
Note that by combining Theorems 4.12 and 4.13 we can, in theory at least, compute
the Moore-Penrose pseudoinverse of any matrix (since AAT and AT A are symmetric). This
turns out to be a poor approach in finite-precision arithmetic, however (see, e.g., [7], [II],
[23]), and better methods are suggested in text that follows.
Theorem 4.11 is suggestive of a "reverse-order" 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 reverse-order 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,
skew-symmetric, 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 Moore-Penrose 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,
skew-symmetric, 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 Moore-Penrose 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(m-r)
; Vi e R
«xr
j
y
2 €
Rnxfo-r^
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(m-rl, VI E IRnxr, V
2
E IRnx(n-r), and the O-subblocks 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
S-I 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 co-domain 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 = AViS-I. Then from (5.4) we see that VrVI = /; i.e., the
columns of VI are orthonormal. Choose any matrix V2 E IRmx(m-r) 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 S-I 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 co-domain 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 finite-precision 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 finite-precision 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 Q-subblocks 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 so-called reverse-order 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
n-r m-r
Figure 5.1. SVD and the four fundamental subspaces.
with the O-subblocks 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 so-called reverse-order identity matrix
(or exchange matrix) P = [e
r
, er-I, ... , 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 so-called row-reduced
echelon form which, when derived by a Gaussian elimination algorithm implemented in
finite-precision 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)(PS-1 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
canbedefinedbyT-Iu; = 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 S-I.
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 so-called row-reduced
echelon form which, when derived by a Gaussian elimination algorithm implemented in
finite-precision 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
so-called column-reduced echelon form, which is not generally a reliable procedure when
performed by Gauss transformations in finite-precision 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
so-called column-reduced echelon form, which is not generally a reliable procedure when
performed by Gauss transformations in finite-precision 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 1-1.
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 1-1.
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 1-1, 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 I-I, 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+(I-A+A)Z
=A+B+(I-A+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 1-1 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 first-order difference equations
6.2. Matrix Linear Equations 45
Remark 6.4. When A is square and nonsingular, A + = A-I and so (I - A + A) = O. Thus,
there is no "arbitrary" component, leaving only the unique solution X = A-I 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 I-lor 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+(I-A+A)y
= (I-A+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++(I-A+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 = A-I.
Example 6.9. Consider the system of linear first-order 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 continuous-time 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 system-theoretic 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
k-J
Xk = Akxo + LAk-J-j BUj
j=O
k k-J Uk-2
[
Uk-J ]
•...• 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 continuous-time 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 system-theoretic 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 left-hand side of (6.13) and let R denote the matrix on
the right-hand 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
k-l
Yk = CAkxo + L CAk-1-j BUj + DUk.
j=O
r
Yo - Duo
Yl - CBuo - Du]
Yn-] - L j : ~ CA
n
-
2
-j BUj - DUn-l
47
(6.12)
(6.13)
Let v denote the (known) vector on the left-hand side of (6.13) and let R denote the matrix on
the right-hand 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 Sherman-Morrison-Woodbury 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 = A-I - A-IB(D-
I
+ CA-IB)-ICA-I.
This result is known as the Sherman-Morrison-Woodbury 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 (A-I + D-I)-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 X-I = X. Note that the positions of the / and - / blocks may be exchanged.
4. r
l
= [
-A-I BD-
I
]
D- I .
5. r
l
= 1
6. [ / +c
BC
r
l
= [!C / 1
7. r
l
= [ A-I 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 one-dimensional 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 co-domain 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 one-dimensional 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 co-domain 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 <2-orthogonal (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 complex-valued, 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# = R-1A
T
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 IRn are Q-orthogonal (or conjugate with
respect to Q) if (x, y) Q = X T Qy = O. 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) :::: 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 complex-valued, 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 Q-orthogonality 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 p-norms, 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 p-norms, 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 p-norms:
(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 (Cauchy-Bunyakovsky-Schwarz 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 Cauchy-Bunyakovsky-Schwarz
(C-B-S) 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 C-B-S inequality is thus equivalent to the statement
Il-Mmlylb — ^
|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 p-norms:
(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 (Cauchy-Bunyakovsky-Schwarz 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 Cauchy-Bunyakovsky-Schwarz
(C-B-S) 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 C-B-S 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 ||z||2 = z z.
Theorem 7.36. All norms on C" are equivalent; i.e., there exist constants c\, c-i (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 p-norms 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/7-norms 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 1-norm," || 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 p-norms 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 Schattenp-norms are defined by
I
IIAlls.p = (at' + ... + a!)"".
Some special cases of Schatten p-norms 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 I-norm," 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|| < ||A||1|jt||. Since | | Af ij c| | < | | A| | | | f l j c| | < ||A||||fl||||j 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 p-norms 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 p-norms) 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 { A-i , . . . , > . „ } . 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 p-norms, but not necessarily
other p-norms) 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
, ||A||j, ||A||
2
,
and ||A||oo 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+(I-A+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 2-norm. 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 2-norm 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 2-norm or F-norm 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+(I-A+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 minimal2-norm. 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 2-norm 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+(I-A+A)Y,
where Y E ]R.nxk is arbitrary. The unique solution of minimum 2-norm or F-norm 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 2-norm or F-norm.
Remark 8.3. If we take B = I
m
in Theorem 8.1, then X = A
+
can be interpreted as
saying that the Moore-Penrose pseudoinverse of A is the best (in the matrix 2-norm sense)
matrix such that AX approximates the identity.
Remark 8.4. Many other interesting and useful approximation results are available for the
matrix 2-norm (and F-norm). 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 2-norm 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 Moore-Penrose pseudoinverse of A is the best (in the matrix 2-norm sense)
matrix such that AX approximates the identity.
Remark 8.4. Many other interesting and useful approximation results are available for the
matrix 2-norm (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(b-Ax)
= 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 2-norms, one could also use 1-norms or oo-norms. 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 2-norms, one could also use I-norms or oo-norms. 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 2-norm 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 2-norm 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 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 AT A. Two basic classes of algorithms are
based on SVD 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
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
so-called full-rank 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 so-called 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
= VIS-ici + V2Z2
= vls-Iufb + 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
so-called full-rank 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 so-called 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 Gram-Schmidt 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 2-norm sense) line of the form y = ax + ft that fits this
data.
(b) Find the best (in the 2-norm 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 2-norm
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 (m-n). 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 Gram-Schmidt 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 2-norm sense) line of the form y = ax + fJ that fits this
data.
(b) Find the best (in the 2-norm 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 2-norm
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 2-norm 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 2-norm 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 often-used scaling for an eigenvector is
a — \j'||;t|| so that the scaled eigenvector has norm 1. The 2-norm 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 (Cayley-Hamilton). 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 often-used scaling for an eigenvector is
a = 1/ IIx II so that the scaled eigenvector has nonn 1. The 2-nonn 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 (Cayley-Hamilton). 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 Cayley-Hamilton Theorem, we know that n(A) = 0. However, it is pos-
sible for A to satisfy a lower-order 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 Cayley-Hamilton Theorem, we know that n(A) = O. However, it is pos-
sible for A to satisfy a lower-order 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 = iJ-Z by x
H
to get x
H
Az = iJ-XH 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
X-lAX = 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
]
3-j 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
X-IAX=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
X-I 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 X-I 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 X-I 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
X-I 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. 120-124]. 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. 120-124]. 0
Transformations like T = [ _ - { ] allow us to go back and forth between a real JCF
and its complex counterpart:
T-I [ (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 X-I. Thus,
det(A) = det(X J X-I) = det(J) = n7=1 Ai.
2. Again, from Theorem 9.22 we have that A = X J X-I. Thus,
Tr(A) = Tr(X J X-I) = 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
A-3I= 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)k-l 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 right-hand 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 right-hand 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 natural-looking 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 finite-precision floating-point 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 high-quality 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 natural-looking 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 finite-precision floating-point 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 high-quality 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 X-lAX = ~
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 reverse-order 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 A-invariant 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 reverse-order 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 A-invariant 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 A-invariant 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 A-mvanant 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 A-invariant.
2. S isA-invariant 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// isA-invariant, 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 A-invariant
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 k-dimensional subspace S, i.e., R(S) = S, then S is A-invariant 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 A-invariant subspace (of dimension one).
Example 9.37. Suppose X block diagonalizes A, i.e.,
X-I 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 A-invariant 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 A-invariant.
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 A-invariant, 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 X-I 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 A-invariant
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 half-plane and P containing all Jordan blocks corresponding to
eigenvalues in the right half-plane. 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 half-plane and P containing all Jordan blocks corresponding to
eigenvalues in the right half-plane. 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 finite-word-
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 A-invariant subspace corresponding to the left half-plane eigenvalues
of A (the negative invariant subspace).
2. R(S+/) is an A-invariant subspace corresponding to the right half-plane 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 finite-word-
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(T-1AT) = T-1sgn(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 A-invariant subspace corresponding to the left half-plane eigenvalues
of A (the negative invariant subspace).
2. R(S + l) is an A -invariant subspace corresponding to the right half-plane 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 skew-Hermitian, i.e., A
H
= —A. Prove that all eigenvalues of
a skew-Hermitian 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 skew-Hermitian.
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 skew-Hermitian, i.e., AH = -A. Prove that all eigenvalues of
a skew-Hermitian 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 skew-Hermitian.
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 X-I 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 X-I is a reduction of A to JCF and suppose we can construct
the "symmetric factorization" of 1. Then A = (X SIXT)(X-
T
S2X-I) 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(n-k), i.e.,
T-IAT = [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 half-plane. 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,
skew-Hermitian, 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--+ PAP-I 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,
skew-Hermitian, 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 A-i 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(n-l) 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 one-dimensional 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 double-Francis-
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 one-dimensional 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 double-Francis-
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 quasi-upper-triangular
matrix. A quasi-upper-triangular 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 (Murnaghan-Wintner). Let A e R"
x
". Then there exists an orthogonal
matrix U such that U
T
AU = S, where S is quasi-upper-triangular.
Definition 10.7. The triangular matrix T in Theorem 10.5 is called a Schur canonical
form or Schur form. The quasi-upper-triangular 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 quasi-upper-triangular
matrix. A quasi-upper-triangular 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 (Murnaghan-Wintner). Let A E IR
n
xn. Then there exists an orthogonal
matrix U such that U
T
AU = S, where S is quasi-upper-triangular.
Definition 10.7. The triangular matrix T in Theorem 10.5 is called a Schur canonical
form or Schur fonn. The quasi-upper-triangular 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(n-l)x(n-l) 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
B-1b (= the Schur complement of B in A). But we
know that
o < det(A) = det [
b ] = det(B) det(a
nn
_ b
H
B-1b).
ann
Since det(B) > 0, we must have ann - b
H
B-1b > O. Choosing a to be the positive square
root of ann - b
H
B-1b 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
[
S-l 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(n-r) 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/.f-Arr+I' ... , I/.f-Arr+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 A-I B > 0, or D > 0 and A - BD-
I
BT > O.
Proof: The proof follows by considering, for example, the congruence
B ] [I _A-I 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 reverse-order
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_IAn-I). 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 reverse-order
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 JRn-1 denote the vector [ai, a2, ... , an-If 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 so-called condition number of A with respect to inversion and with respect
to the matrix P-norm. I f this number is large, say 0(10*), one may lose up to k digits of
precision. I n the 2-norm, 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 reverse-order 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 so-called condition number of A with respect to inversion and with respect
to the matrix p-norm. If this number is large, say O(lO
k
), one may lose up to k digits of
precision. In the 2-norm, 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 reverse-order 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 > S-I.
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 initial-value problem. We restrict our attention in this
chapter only to the so-called time-invariant 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 initial-value problem. We restrict our attention in this
chapter only to the so-called time-invariant 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
} = (sI-Ar
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) .l-I{(sl- A)-I} = erA.
Proof" We prove only (a). Part (b) follows similarly.
{+oo
= io et(-sl)e
tA
dt
(+oo
= io ef(A-sl) 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 term-by-
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(Ai-S)t x;y;H dt assuming A is diagonalizable
10 ;=1
= e(Ai-S)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(A-sl) = L Xiet(Ji-sl)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 term-by-
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 right-hand 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 initial-value 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 vector-valued function u be given
and, say, continuous. Then the solution of the linear inhomogeneous initial-value 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 right-hand 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 initial-value 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(t-to)A
xo
= Ax(t). Also, x(to) = e(to-to)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 vector-valued function u be given
and, say, continuous. Then the solution of the linear inhomogeneous initial-value 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(t-to)A
xo
+ t e(t-s)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(t-to)A Xo + Ir: Ae(t-s)A Bu(s) ds + Bu(t) = Ax(t) + Bu(t). Also,
x(t
o
} = e(to-tolA 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
Matrix-valued initial-value 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 initial-value
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 initial-value
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 initial-value 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 initial-value 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
-e-sAx(s) ds = e-SABu(s) ds.
1
t d 1t
to ds to
e-tAx(t) - e-toAx(to) = t e-
sA
Bu(s) ds
lto
x(t) = e(t-t
olA
xo
+ t e(t-s)A Bu(s) ds.
lto
11.1.4 Linear matrix differential equations
113
Matrix-valued initial-value 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 initial-value
problem
X(t) = AX(t); X(to) = C E jRnxn (11.9)
for t ::: to is given by
X(t) = e(t-to)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 initial-value
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 initial-value 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 initial-value 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(t-to)A Xo
= (ti.iU-tO)Xiyr) Xo
1=1
n
= L(Yi
H
xoeAi(t-tO»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(aieAiU-tO»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(t-s)A Bu(s) ds = t (it eAiU-S)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 X-I AX = J, where J is a JCF for A.
Then
etA = etXJX-1
= 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
X-I
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 MP-I 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
Ik-I At
(k-I)! 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 finite-precision arithmetic but is quite effective for
hand calculation in small-order 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 Cayley-Hamilton 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
lower-order 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
x-I
=[
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 finite-precision arithmetic but is quite effective for
hand calculation in small-order 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, ... , an-l 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 Cayley-Hamilton 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
lower-order powers as well. The polynomial g gives the appropriate linear combination.
Example 11.10. Let
A = [ - ~ - ~ ~ ]
o 0-1
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 small-order 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 = .c-I{(sI - A)-I} and techniques for inverse Laplace transforms. This
is quite effective for small-order 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 discrete-time analogues of the linear differential
equations of the previous section. Linear discrete-time systems, modeled by systems of
difference equations, exhibit many parallels to the continuous-time 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
m-vectors. Then the solution of the inhomogeneous initial-value 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 so-called time-invariant
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 time-invariant, 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
D-I(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 discrete-time analogues of the linear differential
equations of the previous section. Linear discrete-time systems, modeled by systems of
difference equations, exhibit many parallels to the continuous-time 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 so-called time-invariant
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 time-invariant, 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
m-vectors. Then the solution of the inhomogeneous initial-value 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 z-transforms, by analogy with the use of Laplace transforms to compute
a matrix exponential. One definition of the z-transform of a sequence {gk} is
Assuming |z| > max |A|, the z-transform 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 small-order 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
k-I
xk=AkXO+LAk-j-IBUj, 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 z-transforms, by analogy with the use of Laplace transforms to compute
a matrix exponential. One definition of the z-transform of a sequence {gk} is
+00
= LgkZ-
k
.
k=O
Assuming Izl > max IAI, the z-transform of the sequence {Ak} is then given by
AEA(A)
+00
k "'kk 1 12
Z({A})=L...-z-A =I+-A+"2A + ...
k=O z z
= (l-z-IA)-I
= z(zI - A)-I.
Methods based on the JCF are sometimes useful, again mostly for small-order prob-
lems. Assume that A E jRnxn and let X E be such that X-I AX = J, where J is a
JCF for A. Then
Ak = (XJX-I)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 X-I
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 Higher-Order Equations
It is well known that a higher-order (scalar) linear differential equation can be converted to
a first-order linear system. Consider, for example, the initial-value 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 k-I
(;)A
k
-
2
(
k ) Ak-P+I
p-l
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 = XJkX-1 = [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)k-1 (-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 Higher-Order Equations
It is well known that a higher-order (scalar) linear differential equation can be converted to
a first-order linear system. Consider, for example, the initial-value problem
(11.17)
with ¢J(t) a given function and n initial conditions
y(O) = Co, y(O) = CI, ... , in-I)(O) = Cn-I' (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 first-order 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 higher-order 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 first-order 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) = In-l)(t). Then
Xl (I) = X2(t) = y(t),
X2(t) = X3(t) = yet),
Xn-l (t) = Xn(t) = y(n-l)(t),
Xn(t) = y(n)(t) = -aoy(t) - aly(t) - ... - an_lln-l)(t) + ¢(t)
= -aOx\ (t) - a\X2(t) - ... - an-lXn(t) + ¢(t).
These equations can then be rewritten as the first-order 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 higher-order difference equation
with n initial conditions, into a linear first-order 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 skew-symmetric 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 skew-symmetric 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 S-I 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 initial-value problem
for t > 0. Suppose that A e E"
x
" is skew-symmetric and let a = \\XQ\\
2
. Show that
||*(OII
2
= af or al l f > 0.
10. Consider the n x n matrix initial-value problem
12. (a) Find the solution of the initial-value 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 initial-value problem
i(t) = Ax(t); x(O) = Xo
for t ~ O. Suppose that A E ~ n x n is skew-symmetric and let ex = Ilxol12. Show that
I/X(t)1/2 = ex for all t > O.
10. Consider the n x n matrix initial-value 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 initial-value 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(QAZ-XQBZ) = 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(A-XB)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 Murnaghan-Wintner 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 A-AB, then Z-l x isa righteigenvectorofQAZ-AQB Z.
3. ify isa left eigenvector of A -AB, then Q-H 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(Z-l x) =
o.
3. Again, the result follows easily by noting that yH (A - AB) o if and only if
(Q-H 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 Murnaghan-Wintner 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 quasi-upper-triangular.
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 quasi-upper-triangular.
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 A-invariant 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 A-invariant 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 k-dimensional
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 multi-input, multi-output 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 single-input,
with A € M
n x n
, B € R"
x m
, C e R
pxn
, and D € R
pxm
. This linear time-invariant 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 time-invariant 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(sI-A)-'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 mUlti-input, multi-output 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 single-input.
12.4. Symmetric Generalized Eigenvalue Problems 131
single-output 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 second-order
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
single-output 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 second-order
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 real-valued 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 real-valued 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 QDQ-I = QQT AQQ-I = 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 = Q-1 B-
1
Q-T QT 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 lR
nxn
be positive definite. Then A 2: B if and only if B-
1
2:
A-I.
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 so-called 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 so-called 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. Higher-Order Eigenvalue Problems 135
12.6 Higher-Order Eigenvalue Problems
Consider the second-order 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 first-order case, that we try to find a solution of (12.16) of the
form q(t) = e
xt
p, where the n-vector 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 second-order (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 second-order 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 first-order form
Let x\ = q and \i = q. Then (12.16) can be written as a first-order 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 second-order problem (12.16) can still
be converted to the first-order generalized linear system
or, since
12.6. Higher-Order Eigenvalue Problems 135
12.6 Higher-Order Eigenvalue Problems
Consider the second-order 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 first-order case, that we try to find a solution of (12.16) of the
form q(t) = eAt p, where the n-vector 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 second-order (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 second-order 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 first-order form
Let XI = q and X2 = q. Then (12.16) can be written as a first-order system (with block
companion matrix)
. [ 0
X = -M-1K
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 second-order problem (12.16) can still
be converted to the first-order generalized linear system
[
I OJ' [0 I J
o M x = -K -C x.
136 Chapter 12. Generalized Eigenvalue Problems
Many other first-order realizations are possible. Some can be useful when M, C, and/or K
have special symmetry or skew-symmetry properties that can exploited.
Higher-order analogues of (12.16) involving, say, the kth derivative of q, lead naturally
to higher-order eigenvalue problems that can be converted to first-order 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 first-order 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 first-order realizations are possible. Some can be useful when M, C, andlor K
have special symmetry or skew-symmetry properties that can exploited.
Higher-order analogues of (12.16) involving, say, the kth derivative of q, lead naturally
to higher-order eigenvalue problems that can be converted to first-order 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 first-order 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][A-U 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 Q-l AQ-T 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 Q-l =
(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 complex-valued matrices. We
restrict our attention in this chapter primarily to real-valued 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 complex-valued matrices. We
restrict our attention in this chapter primarily to real-valued 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 = A-I ® 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 Jordan-like 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 Q-l BQ = J
B
, respectively, then we get the
following Jordan-like structure:
(P ® Q)-I(A ® B)(P ® Q) = (P-
I
® Q-l)(A ® B)(P ® Q)
= (P-
1
AP) ® (Q-l 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 Jordan-like 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 fJ-j, j E I!!. Then the Kronecker sum A $ B = (1m ® A) + (B ® In) has mn
eigenvalues
Al + fJ-t, ... , AI + fJ-m, A2 + fJ-t,···, A2 + fJ-m, ... , An + fJ-m'
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 fJ-t, ... , fJ-q (q ::s: m), then Z j ® Xi E jRmn are linearly independent right
eigenvectors of A $ B corresponding to Ai + fJ-j' 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) + (fJ-Z ® 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 p-I AP = lA and Q-t 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 ® p-I)(Q-I ® In)][(lm ® A) + (B ® In)][CQ ® In)(/m <:9 P)]
= (1m ® lA) + (JB ® In)
is a Jordan-like 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 mn-vector 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 half-plane). 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 half-plane). 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.
I-Hoo 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 complex-valued 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 Sylvester-like 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 complex-valued 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 Sylvester-like 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 so-called 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 ]
D-XB
if X satisfies the so-called matrix Riccati equation
C-XA+DX-XBX=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 Moore-Penrose 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), 820-826.
[3] Bellman, R., Introduction to Matrix Analysis, Second Edition, McGraw-Hill, 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 Hessenberg-Schur Method for the Problem
AX + XB = C," IEEE Trans. Autom. Control, AC-24(1979), 909-913.
[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, "Ill-Conditioned Eigensystems and the Computation
of the Jordan Canonical Form," SIAM Rev., 18(1976), 578-619.
[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., Finite-Dimensional 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 Moore-Penrose 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),820-826.
[3] Bellman, R, Introduction to Matrix Analysis, Second Edition, McGraw-Hill, 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),57-58.
[6] Golub, G.H., S. Nash, and C. Van Loan, "A Hessenberg-Schur Method for the Problem
AX + X B = C," IEEE Trans. Autom. Control, AC-24(1979), 909-913.
[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, "Ill-Conditioned Eigensystems and the Computation
ofthe Jordan Canonical Form," SIAM Rev., 18(1976),578-619.
[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, P.R, Finite-Dimensional 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), 361-390.
[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, AC-24( 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), 801-836.
[20] Noble, B., and J.W. Daniel, Applied Linear Algebra, Third Edition, Prentice-Hall,
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,
Springer-Verlag, 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),361-390.
[15] Kenney, C.S., andAJ. 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, AJ., "A Schur Method for Solving Algebraic Riccati Equations," IEEE Trans ..
Autom. Control, AC-24( 1979), 913-921.
[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),801-836.
[20] Noble, B., and J.w. Daniel, Applied Linear Algebra, Third Edition, Prentice-Hall,
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),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,
Springer-Verlag, 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
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
LV factorization, 5
triangularization, 149
en, 1
e
mxn
, 1
e ~ x n , 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
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, 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
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
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
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
co-domain 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
quasi-upper-triangular, 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
Moore-Penrose pseudoinverse, 29
multiplication
matrix-matrix, 3
matrix-vector, 3
Mumaghan-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
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, 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
JR.n, I
JR.mxn,1
1
I
range, 20
range inclusion
Index
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, III
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, I
Schur canonical form, 98
generalized, 127
Schur complement, 6, 48, 102, 104
Schur Theorem, 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, 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
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
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. Laub University of California Davis. .Matrix Analysis Matrix Analysis for Scientists & Engineers for Scientists & Engineers Alan J. California slam.

stored. Matrix analysis for scientists and engineers / Alan J. Inc. No part of this book may be reproduced. Mathcad is a registered trademark of Mathsoft Engineering & Education. MATLAB® is a registered trademark of The MathWorks. ISBN 0-89871-576-8 (pbk.. p.. Philadelphia. Mathematica is a registered trademark of Wolfram Research. Includes bibliographical references and index. www. 3600 University City Science Center. I.lam. 508-647-7000. For information. please contact The MathWorks. Inc. Mathcad is a registered trademark of Mathsoft Engineering & Education. cm. . For MATLAB product information.Copyright Copyright © 2005 by the Society for Industrial and Applied Mathematics. Natick. Title. Mathematics. QA188138 2005 QA 188. Inc. 3 Apple Hill Drive. MA 01760-2098 USA. info@mathworks. MA 01760-2098 USA. cm.com..com. write to the Society for Industrial and Applied Mathematics. is a registered trademark. Philadelphia. Alan J. PA.com Mathematica is a registered trademark of Wolfram Research. Library of Congress Cataloging-in-Publication Data Library of Congress Cataloging-in-Publication Data Laub. 2. Mathematical analysis.L38 2005 512. Inc. Inc. Inc. I.. stored. MATLAB® is a registered trademark of The MathWorks.9'434-dc22 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. Printed in the United States of America. Inc. Matrices.. Used by permission.mathworks.com 508-647-7000. Title. PA 19104-2688. 10987654321 10987654321 All rights reserved. p.. Mathematical analysis. or transmitted in any manner without the written permission of the publisher. artist Aaron Tallon of Philadelphia. Printed in the United States of America. info@mathworks. PA. Alan J. Inc.) 1. write to the Society for Industrial and Applied of the publisher. 3600 University City Science Center. For MATLAB product information. 3 Apple Hill Drive. Includes bibliographical references and index. 1948Laub. please contact The MathWorks. 2. 2005 by the Society for Industrial and Applied Mathematics. PA 19104-2688. Used by permission 5.9'434—dc22 512. Laub.. No part of this book All rights reserved.) ISBN 0-89871-576-8 (pbk. or transmitted in any manner without the written permission may be reproduced. • slam is a registered trademark. wwwmathworks. Laub. Fax: 508-647-7101. For information. 1. Natick. Fax: 508-647-7101. Matrices. 1948Matrix analysis for scientists and engineers / Alan J.

Beverley To my wife. Beverley (who captivated me in the UBC math library captivated UBC nearly forty years ago) nearly forty .To my wife.

This page intentionally left blank This page intentionally left blank .

3.1 2. . .2 Some Basic Properties . 3. . . 5.2 Matrix Linear Equations . .3 Composition of Transformations 3..1 Definition and Examples 3. .2 Matrix Linear Equations 6. 4. .4 Structure of Linear Transformations 3. 1.1 4.. . . . ...4 Some Useful and Interesting Inverses.1 Vector Linear Equations ..3 A More General Matrix Linear Equation 6.1 The Fundamental Theorem 5.2 Matrix Arithmetic . Definitions and Examples 2..3 A More General Matrix Linear Equation 6. 6. 6. .2 Subspaces. 6. .1 Definition and Examples . . .5 Four Fundamental Subspaces . .4 Determinants Vector Spaces Vector Spaces 2. . ..1 Some Notation and Terminology 1.3 Linear Independence 2. . .3 Inner Products and Orthogonality 1.. .1 The Fundamental Theorem .1 Some Notation and Terminology 1.. 2. . . . 5..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 . .3 Properties and Applications Introduction to the Singular Value Decomposition Introduction to the Singular Value Decomposition 5.2 Some Basic Properties 5.3 Properties and Applications . .1 Vector Linear Equations 6. .4 Structure of Linear Transformations 3. . . 2. . ..5 Four Fundamental Subspaces Introduction to the Moore-Penrose Pseudoinverse Introduction to the Moore-Penrose Pseudoinverse 4.2 Examples 4.2 Subspaces 2.2 Matrix Arithmetic 1.. . 3.Contents Contents Preface Preface 1 1 Introduction and Review Introduction and Review 1. .4 Determinants 1.3 Row and Column Compressions Linear Equations Linear Equations 6. 4.. .. .3 Rowand Column Compressions 5. .3 Inner Products and Orthogonality . . . ..4 Sums and Intersections of Subspaces 2. . .2 Matrix Representation of Linear Transformations 3.2 Matrix Representation of Linear Transformations 3. 1.2 Examples.4 Sums and Intersections of Subspaces Linear Transformations Linear Transformations 3.. . .3 Linear Independence ..1 Definitions and Characterizations Definitions and Characterizations. .3 Composition of Transformations . . 4. .1 Definitions and Examples . . . .

.4 Matrix Norms . 9. . . . . .1 Properties of the matrix exponential 11.2 On the + l's in JCF blocks 9. . .5 Modal decompositions .1 7.4 Geometric Aspects of the JCF 9.1 Homogeneous linear difference equations 11.3.1 Example: Linear regression . Eigenvalues and Eigenvectors 9. .3 Equivalence Transformations and Congruence 10. . . .1. .2 Definite Matrices 10.1. . .3 Equivalence Transformations and Congruence 10.1. . . . . . .4 Geometric Aspects of the JCF 9. .3. .4 Least Squares and Singular Value Decomposition 8.1 The Linear Least Squares Problem 8. .1 Projections . . . . .3 Inhomogeneous linear differential equations 11. 8. . . . . .1 Fundamental Definitions and Properties 9. . 9.2 Geometric Solution 8. Example: Linear regression 8.5 Least Squares and QR Factorization .2.1 The four fundamental orthogonal projections The four fundamental orthogonal projections 7. .viii viii Contents Contents 7 Projections. . .3. . . .5 Modal decompositions 11. .3 Computation of matrix powers 11.1 Theoretical computation . . . . . 11. . .2 Inner Product Spaces 7. .5 Least Squares and QR Factorization 8. . 10. .2 Inhomogeneous linear difference equations 11.2 Difference Equations . . .2 Geometric Solution .1.1. . . . .2 Other least squares problems 8.2 Definite Matrices . 10.3 Linear Regression and Other Linear Least Squares Problems 8. 11. . 8. . .4 Rational Canonical Form . 11. and Norms 7.1.1 Some Basic Canonical Forms . . .1 Properties ofthe matrix exponential .1. . . .2. . .4 Rational Canonical Form 11 Linear Differential and Difference Equations 11 Linear Differential and Difference Equations 11. .3.3 Vector Norms 7. . 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 . . . . .1.3. . .2.3.3 Vector Norms 7. .1. . .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. Theoretical computation 9. . . .2. . . 11. 7.1 The Linear Least Squares Problem . .1 Projections 7.3.1 Block matrices and definiteness 10. .3 Higher-Order Equations . 10.2 Jordan Canonical Form 9.3 Determination of the JCF . . .2 On the +1's in JCF blocks 9. . .3. . . .2 Inhomogeneous linear difference equations 11. 11.1 9. . . . 8. . . . .3 Computation of matrix powers .1 8.2 Homogeneous linear differential equations 11.2 Homogeneous linear differential equations 11. . . .1. . . 9.1. .1. .3 Determination of the JCF 9. . . . . . . . 7.1 Some Basic Canonical Forms 10. . .5 The Matrix Sign Function. . . . . . . .2 Difference Equations 11.2 Inner Product Spaces 7.1 Differential Equations ILl Differential Equations .2. .3 Linear Regression and Other Linear Least Squares Problems 8.4 Linear matrix differential equations . Inner Product Spaces. 11. . . . . . . . .3 Higher-Order Equations.1. .4 Linear matrix differential equations 11. . . .4 Least Squares and Singular Value Decomposition 8.3 Inhomogeneous linear differential equations 11. .3.2 Other least squares problems .1. . . . . .6 Computation of the matrix exponential 11. .3.1 Fundamental Definitions and Properties 9.6 Computation of the matrix exponential 11. 11. .2 Jordan Canonical Form . . . .1 Homogeneous linear difference equations 11.1 Block matrices and definiteness 10.4 Matrix Norms Linear Least Squares Problems 8. . . . .2. . .

. . . . . .1 The Generalized Eigenvalue/Eigenvector Problem 12.2 Properties of the Kronecker Product . .1 Conversion to first-order form 12. . . .4 Symmetric Generalized Eigenvalue Problems . .1 Definition and Examples .3 Application to Sylvester and Lyapunov Equations 13.5 Simultaneous Diagonalization 12.5.5.5 Simultaneous Diagonalization .1 Definition and Examples 13. . 13. .Contents Contents ix ix 12 Generalized Eigenvalue Problems 12 Generalized Eigenvalue Problems 12. .1 Conversion to first-order form 125 125 125 127 127 130 131 131 133 133 133 135 135 135 139 139 139 139 140 144 144 151 153 13 Kronecker Products 13 Kronecker Products 13. . . 12. .1 The Generalized EigenvaluelEigenvector Problem 12. .2 Canonical Forms 12. .6 Higher-Order Eigenvalue Problems 12. . . .1 Simultaneous diagonalization via SVD 12. .6.4 Symmetric Generalized Eigenvalue Problems 12. . . .3 Application to the Computation of System Zeros . 12. . . . . . 12. . 12.6 Higher-Order Eigenvalue Problems .2 Canonical Forms . 12.6.1 Simultaneous diagonalization via SVD 12. . 13.3 Application to Sylvester and Lyapunov Equations Bibliography Bibliography Index Index . . . . . .2 Properties of the Kronecker Product 13. . . . . .3 Application to the Computation of System Zeros 12.

This page intentionally left blank This page intentionally left blank .

or [16]. and Strang [24] Ortega are excellent companion texts for this book. 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. students meant to learn much of the important and useful mathematics that. although Chapters 2 and 3 algebra. Upon completion of a course based on this text." However. basis-free or subspace) aspects of many of the fundamental notions. and positive definite matrices should have been covered at least once. I highly recommend MAlLAB® although other software such as a digital computer. Upon completion of a course based on this are excellent companion texts for this book. and concepts such as determinants. or computational students in engineering. the student is then well-equipped to pursue. requiring such material as prerequisite permits the early (but "out-of-order" by conventional standards) introduction of topics such as pseuthe early (but "out-of-order" by conventional standards) introduction of topics such as pseudoinverses and the singular value decomposition (SVD). even though their recollecmatrices least tion may occasionally be "hazy. but somehow didn't quite manage to do. 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). These powerful and versatile tools doinverses and the singular value decomposition (SVD). Because tools such as the SVD are not generally amenable to "hand computation. for example)." this ics. or [16]. methods. [13]. computer science. particular subject area. students meant to learn thoroughly as undergraduates. computer science. basis-free or subspace) aspects of many of the fundamental do cover some geometric (i.e. the sciences." However. The text can be used in a one-quarter or one-semester course to provide a compact overview of can be used in a one-quarter or one-semester course to provide a compact overview of much of the important and useful mathematics that. 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. singularity of matrices. I have tried throughout to emphasize only the more important and "useful" tools. The concept of matrix factorization applications and by computational utility and relevance. requiring such material as prerequisite permits tion may occasionally be "hazy. and mathematical structures. 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.e. the student is then well-equipped to pursue. for example). either via formal courses or through selftext. but somehow didn't quite manage to do. follow-on topics on the computational side (at the level of [7]. for [11]. [13]. I have tried throughout to emphasize only the and more advanced material is introduced. Noble and Daniel [20]. I highly recommend MATLAB® although other software such as xi xi . The books by Meyer [18]. The choice of topics covered in linear algebra and matrix theory is motivated both by The choice of topics covered in linear algebra and matrix theory is motivated both by applications and by computational utility and relevance. the sciences.. although Chapters 2 and 3 do cover some geometric (i. 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 text linear dynamical systems (systems of linear differential or difference equations)." this approach necessarily presupposes the availability of appropriate mathematical software on approach necessarily presupposes the availability of appropriate mathematical software on a digital computer. 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. in many cases. [23]. For this. example) or on the theoretical side (at the level of [12]. Ortega [21]. either via formal courses or through selfstudy..Preface Preface This book is intended to be used as a text for beginning graduate-level (or even senior-level) This book is intended to be used as a text for beginning graduate-level (or even senior-level) students in engineering. 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. or [25]. singularity of matrices. For this. mathematics. By matrix analysis I mean linear tools and ideas comfortably in a variety of applications. Matrices are stressed more than abstract vector spaces. in many cases. eigenvalues and eigenvectors. example) or on the theoretical side (at the level of [12]. Because tools such as the SVD are not generally amenable to "hand computation. Matrices are stressed more than abstract vector spaces. [II]. Certain topics thoroughly as undergraduates. Basic concepts such as determinants. Basic of calculus and definitely some previous exposure to matrices and linear algebra. mathematics.

they are either obvious or easily found in the literature. in particular. mathematics. 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. This is ideal not given explicitly. the student does require a certain amount of what is conventionally referred Proofs referred to as "mathematical maturity. This is ideal material from which to learn a bit about mathematical proofs and the mathematical maturity and insight gained thereby. 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 state-space from the modern state-space approach to dynamical systems. applied physics. consistent. and the course has proven to be remarkably successful at enabling students from Davis. especially the first few times it was offered. Rather. But in most engineering or scientific contexts we want to know more than that. or signal processing. It is thus crucial to acquire knowledge vocabulary a working knowledge of the vocabulary and grammar of this language. for example. completed the course. outputs. are deferred to such a course.xii xii Preface Preface Mathcad® Mathematica® or Mathcad® is also excellent. must lay firm foundation upon which and perspectives perspectives can be built in a logical. they are either obvious or easily found in the literature. diverse audience. finite-range floating-point arithmetic environment of of of most modem computing platforms. 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 per se. in particular. . many times at UCSB and twice at UC Davis. I have taught this material for many years. and states often give rise to models of very numbers models high order that must be analyzed. econometrics. are deferred to such a course. control systems with standard large numbers of interacting inputs. and a wide variety of other fields. if only they had had this course before they took linear systems. chemistry. form the foundation Some of the key algorithms of numerical linear algebra. When they are not given explicitly. Some of the key algorithms of numerical linear algebra. science. If If are linearly dependent. one must lay a firm foundation upon which subsequent applications and Rather. and thus the text can serve a rather diverse audience. prerequisites developed While prerequisites for this text are modest. The "language" in which such described models are conveniently described involves vectors and matrices. statistics. simulated. "real-life" problems seldom yield to simple "real-life" closed-form formulas or solutions. how "nearly dependent" are the vectors? If they linearly independent. Mastery of the material in this text should enable the student to read and understand the modern language of matrices used throughout mathematics." For example." Proofs are given for many theorems. linear algebra introducing "on-the-fly" algebra for elementary state-space theory) to an appendix or introducing it "on-the-f1y" when to necessary. The presentation of the material in this book is strongly influenced by computais influenced by computational issues for two principal reasons. form the foundation virtually modem upon which rests virtually all of modern scientific and engineering computation. many students who completed especially offered. and while most material is developed from basic ideas in the book. If a set of vectors is linearly independent. It is my firm conviction that such maturity is neither encouraged conviction neither nor nurtured by relegating the mathematical aspects of applications (for example. The tools of matrix analysis are also applied on a daily basis to problems in biology. and modem engineering. This is an absolutely fundamental fundamental concept. and coherent fashion. in an elementary linear algebra course. and evaluated. are there "best" linearly independent subsets? These tum out to turn be much more difficult problems and frequently involve research-level questions when set be much more difficult problems and frequently involve research-level questions when set in the context of the finite-precision. First. Indeed. They must generally be solved computationally and closed-form it is important to know which types of algorithms can be relied upon and which cannot. A second motivation for a computational emphasis is that it provides many of the essential tools for what I call "qualitative mathematics. State-space methods are State-space modem now standard in much of modern engineering where. remarked afterward that if processing." a set of vectors is either linearly independent or it is not.

June 2004 — AJL. My fellow instructors. etc. too.Preface Preface xiii XIII or estimation theory. . -AJL. realized that by requiring this course as a prerequisite. rather than having to spend time making up for deficiencies in their background background in matrices and linear algebra.. 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. The concept seems to work.

This page intentionally left blank This page intentionally left blank .

n xn Cmxn = the set of complex m x n matrices of rank r. 2.. the notation!! denotes the set {I.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.. Thus. Rn = the set of n-tuples of real numbers represented as column vectors.1 1. e.. mxn = the set of complex (or complex-valued) x n matrices. xyT is an n x n matrix. IR~ xn denotes the set of real = set of real of rank Thus. the notation n denotes the set {1. XTy is a scalar while it easy to recognize immediately throughout the text that. This is followed by a review of some basic notions in matrix analysis and linear algebra. That a vector is always a column vector rather than a row vector is entirely arbitrary. . Crnxn = the set of complex (or complex-valued) m x n matrices.. Note: Vectors are always column vectors. That a vector is always a y E IR n and the superscript T is the transpose operation. Thus. and linear algebra. en 4. x E Rn I. e 6. the set of n-tuples of complex numbers represented as column vectors.Chapter 1 Chapter 1 Introduction and Review Introduction and Review 1. where Xi E R for ii E !!. IR n = the set of n-tuples of real numbers represented as column vectors. The following sets appear frequently throughout subsequent chapters: The following sets appear frequently throughout subsequent chapters: 1. A row vector is denoted by y~. 5. but this convention makes it easy to recognize immediately throughout the text that. = the set of complex m x n matrices of rank r. IR rn xn = the set of real (or real-valued) m x n matrices. R mxn = the set of real (or real-valued) m x n matrices. Thus. 2. Cn = the set of n-tuples of complex numbers represented as column vectors.g. x T y is a scalar while xyT is an n x n matrix. where y G Rn and the superscript T is the transpose operation. Henceforth. n }. . 5. 3. nonsingular n x n matrices. e. 1 . .n xn Rmxnr = the set of real m x n matrices of rank r. Rnxnn denotes the set of real nonsingular n x n matrices. 1R. n}. x e IR n means means where xi e IR for e n. but this convention makes column vector rather than a row vector is entirely arbitrary. Henceforth. A row vector is denoted by yT where Note: Vectors are always column vectors. e.. This is followed by a review of some basic notions in matrix analysis throughout the text...g.

~ 5 is symmetric (and Hermitian). if z = a + jf$ (j = ii = R). A matrix A is symmetric i. then the (m + n) x (m + n) matrix [~ Bc] is block upper triangular. 7 = («77).. then its Hermitian transpose (or conjugate transpose) is denoted by AH (or H If A e C mx ". The The transpose of a matrix A is denoted by AT and is the matrix whose (i. • tridiagonal if a(y = 0 for |z — j\ > 1. Oth (A 7 ). A if A = AT and Hermitian if A = AH. • diagonal if aij7 = 0 for i i= }. If A E em xn. 2 Transposes of block matrices can be defined in an obvious way.e.jj > 1. For example.. = 0 for i > j. j is Remark the more common notation in electrical engineering and system theory. unless if A = A T Hermitian A = A H.j - 7+} ] is Hermitian (but not symmetric). Hermitian conjugate sometimes A*) and its (i. Each of the above also has a "block" analogue obtained by replacing scalar components in the respective definitions by block submatrices. • pentadiagonal if ai. and definitions block submatrices. that is. where the bar indicates complex j)th entry is (A H ). then easy to see that if A.. • upper triangular if a. C E jRmxm. Example 1. i)th entry of A.e. = 0 for j — > 1. • upper triangular if aij. A = AH A complex-valued. text but reminders are placed at strategic locations. • tridiagonal if aij = 0 for Ii . then r = [ ..[ 7 . • lower Hessenberg if aij = 0 for } . otherwise noted. j)\h entry is (AH)ij = (aji). a. i. • upper Hessenberg if afj = 0 for — > 1. • upper Hessenberg if aij = 0 for ii . For example. an equation like A = A T implies that A is real-valued while a statement like A = AH implies that A is complex-valued.JI > 1. While \/—\ is most commonly denoted by i in mathematics texts. an equation like A = A T implies that A is real-valued while a statement otherwise noted. z Remark 1.1. (7. is Hermitian (but not symmetric). • lower Hessenberg if a. where the bar indicates complex sometimes A*) and its = IX jfJ (j = = v^T). A = [ . Introduction and Review We now classify some of the more familiar "shaped" matrices. it is easy to see that if Aij are appropriately dimensioned subblocks.. Introduction and Review Chapter 1. 2 2. • lower triangular if aij7 = 0 for i/ < }. then A7" e jRnxm.2. 7 + j ] is complex-valued symmetric but not Hermitian. are appropriately dimensioned subblocks. A = [ 7+} 5 3· A . A is conjugation. We henceforth adopt the convention that. then z = IX -— jfi.J I > 2. For example. A matrix A E IRn xn e (or A E enxn ) is A eC" x ")is • diagonal if a... if A E IRnxn. = a jfJ. B E IR nxm .. There is some the more common notation in electrical engineering and system theory. = 0 for |/ — j\ > 2.2 2 Chapter 1. = 0 for < j. ] is symmetric (and Hermitian). = 0 for i ^ j. For example. (AT)ij = aji. = 0 for i > }. it is Transposes of block matrices can be defined in an obvious way. • lower triangular if a. j)th entry of a matrix A is denoted by AT and is the matrix whose j)th entry A.. The notation j is used throughout the text but reminders are placed at strategic locations. Note that if A E R mx ".. . if e Rnxn e Rmxn C e Rmxm then the (m n) x (m n) matrix [A0 ~] is block upper triangular. A e jRmxn. • pentadiagonal if aij = 0 for Ii . is complex-valued symmetric but not Hermitian. 1. AT E E" xm is the (j. } is While R is most commonly denoted by i in mathematics texts.ii > 1.2. There is some advantage to being conversant with both notations. We henceforth that. Example 1.

xn~ ] Then Ax = Xjal + . suppose A e Rmxn and B = [bi. multiplication. eRmxn has row cj e E l x "..• a"1 E m JR " with a. un Rmxn with u Rm and V = [v . suppose (linear combination) suppose A = la' ..3 can then also be generalized to its "row dual. . {.3. multiplication of a matrix by a scalar.e. n UV T = LUiVr E jRmxp.[ ~ J+l. and multiplication of matrices. It is deceptively simple and its full understanding is well rewarded.~]. matrix-vector if C E jRmxn has row vectors cJ E jRlxn. vector x. 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.[ ~ J+2.. x = ! 2 Then we can quickly calculate dot products of the rows of A [~]. its importance cannot be overemphasized. suppose A E jRmxn and [hI. E JRm and x = l I. As a numerical example. Theorem reader..2 Matrix Arithmetic 1.. Let U = [MI.. if (C D)H — D C ). but equivalent. . ." The details are left to the readei "row left . recall that (CD)T = DT C T (C D)T = DT T If H H H (or (CD} = DHC H ).. Ax. This gives a dual to the matrix-vector result above. i=I If matrices C and D are compatible for multiplication. Namely. Then the matrix product A B can be thought of as above.bhp ] e Rnxp with hi e jRn. applied p times: There is also an alternative.1. Theorem 1. the matrix-vector product Ax.. Vn] ]Ee lR Pxn U [Uj. and is premultiplied by a row vector yTe jRlxm.. Again. Then v E jRP. there can be important computer-architecture-related advancomputer-architecture-related tages to preferring the latter calculation method..[ ~ l For large arrays of numbers..3..2 Arithmetic It is assumed that the reader is familiar with the fundamental notions of matrix addition.2.. A special case of matrix multiplication occurs when the second matrix is a column multiplication second i. Then we can quickly calculate dot products of the rows of A column Ax = [...'" p] E jRnxp For matrix multiplication. Theorem 1. i. It Theorem 1. The importance of this interpretation cannot be overemphasized. AB bi E W1. formulation of matrix multiplication that appears frequently in the text and is presented below as a theorem. vn Rpxn p with Vit e R .•. importance interpretation take A = [96 85 74]x = take A = [~ ~]. + Xnan E jRm. un]]Ee jRmxn with Ui t Ee jRm and V = [VI..... A very important way to view this product is interpret weighted to interpret it as a weighted sum (linear combination) of the columns of A. .. . That is.. Matrix Arithmetic 3 1.e. matrix-vector product with the column x to find Ax = [50 32]' but this matrix-vector product can also be computed computed via v1a 3. .

To illustrate. where / is the n x n identity matrix. If x and y are zero.e. Nonzero complex vectors are orthogonal if x H y = 0. y e IRn are said to be orthogonal if their inner product is zero. x)c and we see that. Similarly. In sometimes used denote identity matrix. A EC = (orC" xn). xTyy = 0. A orthogonal and XTX = 1 and yTyy = 1. y)c = (y. for short) of x and For vectors y e IRn. consider What is true in the complex case is that x H 0 if and only if O. y)c = (y. Clearly said = an orthogonal or unitary matrix has orthonormal rows and orthonormal columns. x)c. Nonzero complex vectors are orthogonal if XHy = O.2j while while and we see that. Then Example 1.3 Inner Products and Orthogonality Inner Products and Orthogonality For vectors x.j] [ ~ ] = 1 . (x. Introduction and Review 1. we define their complex Euclidean inner product (or inner product.y. Y}c = {y. i. Let x = [} ]] and y = [~]. Similarly. 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. then we say that x and y are orthonormal. Then (x. A nxn matrix E R is an orthogonal matrix if AT A = AAT = I.. x)c'. 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. the nonzero vector x above. for short) by for short) by n (x'Y}c :=xHy = Lx. y)c = (y. case. y) := x y = Lx. 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. i. Let x = [1j and y = [1/2].e. indeed. Then x T x = 0 but x H X = 2.4. x and y are orthogonal and x Tx = 1 and yT = 1. Two nonzero vectors x. Introduction and Review Chapter 1. 1. a matrix A e en xn is said to be unitary if A H A = AA H = I. (or A E Cnxn) we use the notation det A for the determinant of A.4 4 Chapter 1. the Euclidean inner product (or inner product. x T = O. If x. y E R".y. (x. . For A E R nnxn A e IR xn It assumed of determinants.. The notation /„ is sometimes used to denote the identity matrix in IR nxn in Rnx" x nxn H H (or en "). the Euclidean inner product inner for short) y is given by y is given by n T (x. y ) c = yHxx = L:7=1 x.3 1. but throughout the text we prefer the symmetry with the real (x.=1 y appear in Note that (x. x}c. Example 1. What is true in the complex case is that XH x = 0 if and only if x = 0. i. consider the nonzero vector x above. the order in which x and y appear in the complex inner (x.e. The more conventional definition of the complex inner product is product is important.. There is an orthogonal or unitary matrix has orthonormal rows and orthonormal columns. .. y)c = y = Eni=1 xiyi but throughout the text we prefer the symmetry with the real case. Then XTX = 0 but XHX = 2.. where I is the n x n matrix A e IRnxn is an orthogonal matrix if ATA = AAT = /. We list below some of (or A 6 en xn) we use the notation det A for the determinant of A.4 1. Note that the inner product is a scalar.4. i. we define their complex Euclidean inner product (or inner product. order in which x product is important. E R are said to be orthogonal if their inner product is Two nonzero vectors x. To illustrate. We list below some of . y E <en. If e C". indeed. then we say that x and y are orthonormal. rows or columns.. Y}c = [ } JH [ ~ ] = [I . The more conventional definition of the complex inner product is H ( x .4 Determinants Determinants It is assumed that the reader is familiar with the basic theory of determinants.e.y.=1 Note that the inner product is a scalar.

. Interchanging two rows of A changes only the sign of the determinant. If A.. Interchanging two rows of A changes only the sign of the determinant.. then det A = a11a22 . then det [Ac ~] det D det(A B D.. Proof" Proof: This follows easily from the block UL factorization BD.. . 16.e. Multiplying a column of A by a scalar ex results in a new matrix whose determinant scalar a determinant is ex det A. 4. 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. of 5. Determinants 1.4..• ann. exdetA.. If A A A = o.. 7. If A e R n x n and D e R m x m . Note that this is not a minimal set. then det A = all a22 . 8. A 11. det A is the product of its diagonal 10.1 ) = de: A... properties 1. det A 11 det A22 • • det Ann 14..CA. then det(A. i. Interchanging two columns of A changes only the sign of the determinant. If A. Multiplying a row of A by a scalar and then adding it to another row does not change 7. det AT = det A (detA H = detA if A e C nxn ). If A has a zero column or if any two columns of A are equal. det A is the product of its diagonal diagonal. If A is lower triangUlar.1.• a"n. detAT = detA (det A H = det A A E C"X"). 8. then det [~ BD] = del A det(D – CA– l 1 B). If A € lR~xn.• det Ann.e..C). 15. then det A = alla22 • • ann 12. Multiplying A 6. Multiplying a row of A by a scalar and then adding it to another row does not change the determinant. 10. i. 13. then det(AB) = det A det B. 11. If A is block diagonal (or block upper triangular or block lower triangular). 3.1 I ][ . several more is a properties are consequences of one or more of the others.4. the determinant. If det = a11a22 • • ann 12.thendet(AB) = det A det 5. If elements. Multiplying a row of A by a scalar ex results in a new matrix whose determinant is a det A.• ann. are consequences of one or more of the others.B). then det A = different det A11 det A22 . 16. If A has a zero column or if any two columns of A are equal..A22.. If A E Rnxn. then det A = O. If A is upper triangular. If A e IRnxn and D E lR~xm. 14. A 22 . If A E lR~xn and DE IR mxm det [Ac ~] detA det(D .. 2. then det [~ BD] = det D det(A -– B D – 11C ) . Multiplying a row of A by a scalar a results in a new matrix whose determinant is 5. 15. If A has a zero row or if any two rows of A are equal. • • An" (of A = square diagonal blocks A11. If A E R n x n and D e RMmxm. Ann (of possibly different sizes).. If A is diagonal. Determinants 5 properties the more useful properties of determinants. then det A = 0.. 17. Proof: This follows easily from the block LU factorization Proof" This follows easily from the block LU factorization [~ ~J=[ ~ ][ ~ 17. is a det A.e. B eR n x n . 11... = alla22 • • ann i. then det A = 0. change the determinant. If A is lower triangular. 3. then det A = a11a22 . then det(A-1) = 1detA . B E IRnxn .•. 9. with A block diagonal (or block 13.

B e JRn xn and a. i. TrA = L~=I au· elements.. Introduction and Review Chapter 1. lower triangular with all l's on the diagonal) and an upper triangular matrix L 1's an V is called an LV factorization. Show that A must be singular. Letx.• V k is an orthogonal matrix. denoted TrA. A matrix A e Wx" is said to be idempotent if A2 = A. then Tr(aA fiB)= aTrA + f3TrB.. (a) Show that the trace is a linear function. The factorization of a matrix A into the product of a unit lower triangular matrix L (i.yTx.. see.e. if A.6 6 Chapter 1. lor z r 2sm2rt # J. 6. Then E jRnxn skew-symmetric.e. Tr(aA + f3B) = aTrA + fiTrB.. AB ^ BA.. Remark — C -I B – BD-Ie Similarly. II _ . The trace of A. TrS O.. . Another such factorization is VL U is an LU factorization. Let A E jRNxn. see. A E jRnxn A2 / x™ . (b) Suppose A e IR" X "is idempotent and A i= I.. U1 U2 • • Uk is an 5. are block analogues of these. A . • .e. A? 2. The factorizations used above U triangular. either prove the converse or provide a counterexample... .. of Din [AC ~ l EXERCISES EXERCISES nxn 1. what is det A? If A 3. If A is orthogonal. The matrix D . . Let U1. Introduction and Review Remark 1.e A – 1 B is called the Schur complement of A in[ACBD]. i. Remark 1. ft e R.5. 2sin20 J is idempotent for all #. (b) Show that Tr(AB) = Tr(BA). . [~ ~ ].. Let x. aII o. The factorization of a matrix A into the product of a unit lower triangular Remark 1. ST = -So Show that TrS = 0. Show that det(I – xyT) = 1. i.e.e.. ST = -S. i.B D – l C is the Schur complement of D in [~ BD ]. A =. of denoted Tr A. V2. what is det(aA)? What is det(–A)? E R a det(-A)? A? If A unitary.5. Show that the product V = VI. (c) Let S € Rnxn be skew-symmetric. .6.. U2 . ! [ 2cos2<9 I T 2cos2 0 (a) Show that the matrix A = _.y E jRn.Vk € jRn xn be orthogonal matrices. 2 0 IS I dempotent . for example. Suppose A E jRn xn is idempotent and A ^ I. _. Showthatdet(l-xyT) 1 – yTx. elements. nxn linear E R f3 E JR. 2f) 2 |_ sm 2^ sin 0 sin sin 20 1 . [24]. TrA = Eni=1 aii. what is det A? If A is unitary. Tr(Afl) = Tr(£A). Uk E Rnxn U = VI V2 . even though in general AB i= B A. 4. y e Rn. Another such factorization is UL where V is unit upper triangular and L is lower triangular.. example. . is defined as the sum of its diagonal A e Rnxn. [24].. Show that A must be singular. If A e jRnxn and or is a scalar.

8 = ft + afar all a. ft. including spaces formed by special classes of matrices. Axioms (M1)-(M4) state that (F \ to). ft Elf. not written explicitly.8 +a· y for all a. (M4) a·. (Ml) a . I = a for all a E F. a-I = 1.1. there exists an element (-a) e F such that a + (-a) = 0. but some infinite-dimensional examples are also cited. y Elf. (Ml) a· p . (Al) a + (. 2. . a f. (M2) 1 e IF • I = for a e IF.8 + y) = (a +. p e F.8 .) ( a . : IF x F —> IF such that Definition 2. p. when no confusion can arise.8.8. ^ 0. Generally speaking. 0. but some infinite-dimensional examples are also cited. (A4) a + .o r all a. The emphasis is on finite-dimensional vector spaces. be found. including spaces formed by special classes emphasis is on finite-dimensional vector spaces.8 = P • a for all a. • F x IF ~ F such that (Al) a (P y ) = (a + p ) y o r all a. there exists an element (—a) E IF such that a (—a) O. A field is a set IF together with two operations +. y e F. .. yy) = (a·. An excellent reference of matrices. (A2) there exists an element 0 E F such that a + 0 = a for all a E F.((.y for alia. (D) (D) a· p a .8) + y ffor all a.8 e F.) is an abelian group.8 E IF. +) is a group and an abelian group if (A4) also holds.1 Definitions and Examples Definition 2. Axioms (MI)-(M4) state that (IF \ {0}. there exists an element a-I E F such that a . The In this chapter we give a brief review of some of the basic concepts of vector spaces. (M3) e IF.Chapter 2 Vector Spaces Vector Spaces In this chapter we give a brief review of some of the basic concepts of vector spaces. afar all a. Axioms (Al)-(A3) state that (IF. (A4) a + p = . (M3) for all a E ¥. y)=ci-p+a. for all a e IF. where some of the proofs that are not given here may be found.8 + y) = a·. y Elf. An excellent reference for this and the next chapter is [10]. the multiplication operator ". where some of the proofs that are not given here may for this and the next chapter is [10]. . y € F. . Axioms (A1)-(A3) state that (F. (A3) for all a E IF.8)· yyf for all a.8.1.8. (M2) there exists an element I E F such that a . (A2) there exists an element 0 e IF such that a 0 = a. •) is an abelian group.((. the multiplication operator "•" is Generally speaking. A field is a set F together with two operations +. 7 . (M4) a • p =. p. when no confusion can arise..ye¥. . a"1 € IF • a~l = 1. +) is a group and an abelian group if (A4) also holds. (A3) for all a e F." is not written explicitly.8 a for all a. y Elf.p ) .

w for all a e F and for all v. v for all a.3 are different from the + and • in Definition 2. Note that + and· in Definition 2. in Definition Remark 2. .1 in the sense of operating on different objects in different sets.8 Chapter 2. (V2) ( a f3) v = a P .4. e).5. where Z+ = {0. v + a-.3. }.. IR) with addition defined by and scalar multiplication defined by and scalar multiplication defined by is a vector space. simply by V. p € F and for all v e V. (V2) (a·. p E IF andfor all v e V.2. Example 2.f3i EIR . Definition 2.2. Note that + and • in Definition 2. f3 Elf andforall v E V. In practice.l. when there is no possibility of confusion as to the underlying fie Id.. (V5) I·• v = v for all v E V (1 e F).qEZ +} . Similar definitions hold for (en. (Ml) does not hold unless m = n. f3 e F and for all v E V. Remark 2. when there is no possibility of confusion as to the A vector space is denoted by (V. + f3qXq :aj.. R) with addition defined by I.1. no confusion and the operator is usually not even written explicitly. simply by V. Moreover. + apxP + . for example. R"x" is not a field either for example. Similar definitions hold for (C".F xV -»• V such that (VI) (V.P.r] = the field of rational functions in the indeterminate x = {ao + f30 + atX f3t X + . w e V. 1. IF) or. this causes no confusion and the·• operator is usually not even written explicitly. ft) v = a v + f3.p ) -. IR with ordinary addition and multiplication is a field. In practice. (R". A vector space is denoted by (V. C).5. (VI) (V. (V3) (a (V4) a-(v w)=a-v a w for all a ElF andfor all v.. . e with ordinary complex addition and multiplication is a field. 3. where Z+ = {O.1 in the sense of operating on different objects in different sets. lR~xn is not a field either since (M4) does not hold in general (although the other 8 axioms hold).. Moreover.. 4. (V3) (a + f3). RMrmxn= {m x n matrices of rank r with real coefficients} is clearly not a field since.• v = a·• v + p • v for all a. I. A vector space over a field F is a set V together with two operations Definition 2. since (M4) does not hold in general (although the other 8 axioms hold). 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. . this causes 2. Ra[x] = the field of rational functions in the indeterminate x 3. C with ordinary complex addition and multiplication is a field. Example 2. w E V.( (f3' V v) o r all a. (V5) 1 v = v for all v e V (1 Elf). ) f for all a. v = a . +) is an abelian group.}. 4. is a field. IR~ xn = m x n matrices of rank r with real coefficients) is clearly not a field since. Vector Spaces Example 2.. R with ordinary addition and multiplication is a field. F) or. 2. 1. +) is an abelian group. Raf.. underlying field. is a vector space.4. Example 2. (V4) a· (v + w) = a . (IRn.2. (MI) does not hold unless m = n.. is a field.3 are different from the + and .3.2. 2.

JR). is henceforth understood to mean "is a subspace of. this implies that the zero vector must be in any subspace. fJ e IF andforall WI. w2 e Remark 2. Subspaces 2. t\]. l . and the symbol ~. if and only if(aw1 ßW2) E if(awl + fJw2) e W for all a." ya2n . Let (V. (E mxn JR) is a vector space with addition defined by 2. F) if and only if (W. W f= 0. we write W c V. ß E ¥ and for all w1. and the symbol c.6. IF) is itself a vector space or. Then (W. JR).2 Subspaces Subspaces Definition 2. F) is itself a vector space or. Special Cases: Special Cases: (a) V = [to. that since 0 E IF. t\])n continuous =: (C[to. and the functions are piecewise continuous =: (PC[to. Let (V. implies that the zero vector must be in any subspace. equivalently. + fJmn l yaml yamn 3. Then (x(t) : x ( t ) = Ax(t)} is a vector space (of dimension n).. Notation: When the underlying field is understood. F) is a Definition 2. Note. Let cf>('O. Then {x(t) : x(t) = Ax(t}} is a vector space (of dimension n). i. 4. (JRmxn. g E cf> and scalar multiplication defined by and scalar multiplication defined by (af)(d) = af(d) for all a E IF. Let (V." The when used with vector spaces. Note. E). E) is a vector space with addition defined by 9 9 A+B= [ . Then (W. IF) = (JR n . W = 0.7. Then O(D. etc. 4. too." + fJ2I a21 + P" . verify that the set in question is closed under addition and scalar multiplication. F) = (IR". F) be a vector space and let W ~ V. is henceforth understood to mean "is a subspace of. too. V) be the 3. The latter characterization of a subspace is often the easiest way to check or prove that something is indeed a subspace (or vector space). 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. verify that the set in or prove that something is indeed a subspace (or vector space). (V. td)n or continuous =: (C[?0. h])n (b) '0 = [to. IF) = (JRn.. IF) be a vector space and let W c V. this question is closed under addition and scalar multiplication. W2 E W. 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 = . and for all f E cf>. foral! a.. +00). less restrictive meaning "is a subset of' is specifically flagged as such. (V. i. IF) is a subspace of (V. Let A E JR(nxn. 2. and the functions are piecewise continuous (a) '0 = [to. that since 0 e F. equivalently. =: (PC[f0. (V. Let O(X>. we write W ~ V. Notation: When the underlying field is understood.2. Subspaces 2. Then cf>('O..6." The less restrictive meaning "is a subset of" is specifically flagged as such. td)n.2.2.2 2. V) be the set of functions / mapping D to V. Let A € R"x". if and only subspace of (V. Let (V. IF) if and only if (W.e. IF) be an arbitrary vector space and '0 be an arbitrary set.e. F) be an arbitrary vector space and V be an arbitrary set. when used with vector spaces. for all d ED. y a l2 y a 22 yam 2 ya.7. td. V) is a vector space with addition set of functions f mapping '0 to V. The latter characterization of a subspace is often the easiest way to check Remark 2. .

•••. ~SandS ~ R..nxn.R) and 1. R"x".. v2. 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. • • •} be a nonempty collection of vectors u. Let X {VI.. Then it is easily shown that ctA\ + f3A2 is Proof' Suppose AI. Note.. .9. + (XkVk = 0 implies al = 0. ak not all zero such that elements VI. To prove two vector spaces are equal. .ß with ß =1= 0 are called linear varieties.10. Example 2. sketch Then Wa..ß is a subspace of V if and only if ß = O. that the vertical line through the origin (i. . too. IF) = (]R2. sketch W2. V usually denotes a vector space with the underlying field generally being R unless Thus. Consider (V.8.nxn..10.e. . As an interesting exercise. ft E R symmetric for all a. Then W". . Then W is /wf a subspace of JR. A2 are symmetric.I' and Wi. define W". called linear varieties.) and let W = [A e R"x" : A is symmetric}. a = oo) is also a subspace. Let W = {A € R"x" : A is orthogonal}. Vk e X and scalars aI. Proof: Suppose A\. Consider (V. . Consider (V. For a. Wi.Vk of X and for any scalars a1. Vk of X and for any scalars aI.O. W~V. X is a linearly dependent set of vectors ifand only if there exist k distinct if and only if exist distinct elements v1. All lines through the origin are subspaces. . . ./l is a subspace of V if and only if f3 = 0.JR. Definition 2. one usually proves the two inclusions separately: Note: To prove two vector spaces are equal. explicitly stated otherwise. . we drop the explicit dependence of a vector space on an underlying field.S. ak. be an element of R.3 2.} .10 Chapter 2. Shifted subspaces Wa. = {A E JR./l with f3 = 0 are All lines through the origin are subspaces. 1. Henceforth. Then it is easily shown that aAI + fiAi is symmetric for all a. . that the vertical line through the origin (i. . unless explicitly stated otherwise. al VI + .e.nxn : A We V. F) = (JR.and W1/2.lF) = (R" X ". As an interesting exercise.. .. then R = S if and only if Definition 2. ak = O.. Shifted subspaces W". Thus. V2. . Vector Spaces Example 2..9. (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 v1.1.) and for each v E ]R2 of the form v = [~~ ] identify VI with the x-coordinate in the plane and u2 with the y-coordinate. Note. . ffR and S are vector spaces (or subspaces). c E JR.. E ]Rnxn : not 2. f3 e R. If 12. then R = S if and only if R C S and S C R.. elements VI. a = 00) is also a subspace. Henceforth. ak.• } be a nonempty collection of vectors Vi in some vector space V.. R) and for each v € R2 of the form v = [v1v2 ] identify v1 with 3.o. . . vk E X and scalars a1. W1/2.o. Definition 2. we drop the explicit dependence of a vector space on an underlying field. in some vector space V. W2. . f3 e R. JR...I.3 Linear Independence Linear Independence Let X = {v1. too. •. and S are vector spaces (or subspaces). X linearly set of Definition 2. 3. W2. W2.0. A2 are symmetric.•.1. For ß E R define the jc-coordinate in the plane and V2 with the y-coordinate.. V usually denotes a vector space with the underlying field generally being JR. 2./l = {V : v = [ ac ~ f3 ] . F) = (R2. . Then (V.

. e k. e2 = 0 1 0 . to be studied further in what follows. o Definition 2.. Why? independent. . + (XkVk . then a = 0.12. Vi EX. ~ HHi] } Ime~ly i is a i" linearly independent set. = [ v 1 . A set of vectors X is a basis for V if and only ij Definition 2. 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.. 2.2. . .. linear dependence x such that Va = 0. then = O. } = (Xl VI + . en} = Rn. .. An equivalent condition for linear dependence is that the k x k matrix condition VT V is singular. Sp(X) = V. Howe. Independence of these vectors turns out to be equivalent to a concept called controllability. (Xi ElF. = {v : where N = {I. which is discussed in more detail in efA Chapter 11).12. X = [v1 v2 . Then consider the rows of etA B as vectors in Cm [t0.v2 + V3 = 0).3. Example 2. Definition 2. The linear v E ]Rn.11."I [ i1i1l ]} [[ s a linearly is a Iin=ly dependent set de~ndent ~t (since 2v\ — V2 + v3 = 0). and X (of and 2.13. 1.. to be studied further in what follows... {1. Sp(X) = V. kEN}. Vi e span of Definition 2.14. Vk] E Rnxk. e2 . 1£t V = R3. and there exists a E ]Rk such that VT V is singular.. . 2. LetV = 11 11 ~.. .'" . Then {[ Then I. . .13.14. Let X = {VI.Vk] e ]Rnxk. An equivalent condition for linear independence is that the matrix Va = 0. ii E If. Linear Independence Example 2. Independence of these vectors turns out to be equivalent to a concept Chapter 11). V2.. An equivalent condition for linear independence is that the matrix V TV is nonsingular.. Let A E ]Rnxn and 5 e R"xm.•}} be a collection of vectors vi. e2.. and there exists a e R* such that Va = 0. called consider Let Vif e R". Why? However. Let V = Rn and define = ]Rn and el = 0 0 ... .}.... Then Sp{e1. A set of vectors X is a basis for V if and only if 1.. . t1] (recall that etA denotes the matrix exponential. Linear Independence 2.en = 0 0 0 o SpIel. T V is nonsingular.. A e R xn B E ]Rnxm. }. If the set of vectors is independent. Example 2. Then consider the rows of etA B as vectors in em [to... If the set of vectors is independent. . X is a linearly independent set (of basis vectors). E V. tIl 2. (since 2vI . V V2. .en} = ]Rn.. and consider the matrix V = [VI.3. Then the span of of X is defined as X is defined as Sp(X) = Sp{VI.11. 2..

....18. r I [ . + ~nbn = Bx. In]Rn.dimensional or have dimension n and we write dim (V) = n or dim V — n. .. components represents the vector v with respect to the basis B.. Definition 2.l.19. en} natural Now let b l .. n for Then for all e there exists a unique n-tuple {E1 . Then for all v E V there exists a unique n-tuple {~I'. for]Rn [e\. Example 2. en} is a basis for IR" (sometimes called the natural basis).. The number of elements in a basis of a vector space is independent of the Theorem 2.. el + 2 . For example. .. {el. n unique. .[ -~ - ] + 4· [ -~ l To see this. .. The number of elements in a basis of a vector space is independent of the particular basis considered.18. We say that the vector x of of of (b1. If V= 0) V is Definition 2. .17..b. . . If a basis X for a vector space V(Jf 0) has n elements. B ~ [b". VI ] : = vlel + V2e2 + . e2. {~i } of v with respect to the basis {b l . bn]} and are unique. while We can also determine components of v with respect to another basis.16... . particular basis considered. while [ ~ ] = I .. x ~ D J Definition 2.16. For be n. In Rn. + vne n · Vn We can also determine components of v with respect to another basis..19. The scalars {Ei}are called the components (or sometimes the coordinates) components coordinates) Definition 2. bn be a basis (with a specific order associated with the basis vectors) b1. with respect to the basis with respect to the basis {[-~l[-!J} we have we have [ ~ ~ ] = 3.. We represents B. .15... write [ ] = XI • [ ~ + ] X2 • [ _! ] =[ ~ = [ -~ Then Then -! ][ ~~ l -1 [ ~~ ] = [ -.. V is said to X for be n-dimensional or have dimension n and we write dim(V) n or dim V n. . ] l = = Theorem 2. . .. For example. n } such that for V. Vector Spaces Example 2..E~n} such that v= where ~Ibl + .12 12 Chapter 2... For .

= T). a eA CiEA f] n *R. 2. R + S = {r + s : r e R. ft n 5 = {v : v E 7^ and v E 5}. 2. Remark 2. and because the 0 vector is in any vector space. The union of two subspaces. U + S = T (in general ft. Note: Check that a basis for ~mxn is given by the mn matrices Eij.20. . j E ~. and 1." 3. Remark 2. 2. V. i e m. y>f (L L . and because the 0 vector is in any vector space. we define dim(O) = O. H. Theorem 2. 2.4. Theorem 2.) 2 5. Thus. 1. R j ) = 0 am/ Ri = T).) (To see why. and S are said to be complements of each other in T. A consistency. and 2.. dim(Rn)=n.23. S.. R S C V (in general.-) = 0 and ]P ft. Note: Check that a basis for Rmxn is given by the mn matrices Eij. 5. n 5 S. j e n. Let (V.22.18 says that dim (V) the number of elements in a basis. R C S. F) be a vector space and let R. otherwise. dim{A e Rnxn A is upper (lower) triangular} = !n(n 1). JF') be a vector space and let 71.4 2.20.4. and S are defined respectively by: 1. Example 2.j matrices can be called the "natural basis matrices. Let (V. i E m. s e S}.a S. The sum and intersection ofR and S are defined respectively by: of R. n (^ ft. 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. K + S S. for finite k). Example 2. dim{A E ~nxn :: A = AT} = !n(n + 1). + 7^ =: L R. Theorem 2. The union of two subspaces.24. for finite k). 2. R H S = {v : v e R and v e S}. 72.. V (in general. 1. 1. determine 1/2n(n + 1) symmetric basis matrices. U\ + 1. J)th location.=1 K k 1=1 2. dim(~n) = n. The collection of Eij matrices can be called the "natural basis matrices. Definition 2.21. dim{A € Rnxn A AT} = {1/2(n 1 (To see why. Sums and Intersections of Subspaces 13 13 consistency. The subspaces Rand S are said to be complements of each other in T. V is infinite-dimensional.22. S c V. R D S C V (in general.+00.18 says that dim(V) = the number of elements in a basis. dim(~mXn) = mn. V. j)th location. we define dim(O) = 0. where Eij is a matrix all of whose elements are 0 except for a 1 in the (i.4 Sums and Intersections of Subspaces Subspaces Definition 2. V for an arbitrary index set A). Ra C V/or an arbitrary index set A). RI -\ h Rk =: ]T ft/ C V. t1]) . is not necessarily a subspace. dim(C[to. n S {r s : r E U. U S. S S. otherwise. Theorem 2. R S = (in general. Thus. V (in general.2.21. tJJ) = +00. dim(R mxn ) mn. 4. R = 0. s E 5}. A vector space V is finite-dimensional if there exists a basis X with n < +00 elements.23. is not necessarily a subspace. where Efj is a matrix all of whose elements are 0 except for a 1 in the (i. determine !n(n 1) symmetric basis matrices. Sums and Intersections of Subspaces 2. The sum and intersection Definition 2. R.24. 2. n n S = 0. V is infinite-dimensional. dim{A E ~nxn :: A is upper (lower) triangular} = 1/2n(n+ 1)." The collection of E. vector space V is finite-dimensional if there exists a basis X with n < +00 elements. 1. . « The subspaces R.

Av" •. S2 E S. AVn are orv\. consider V = R2 unique.26.27. S of a vector space V. . For example.. 1 TIT The first matrix on the right-hand side above is in S while the second is in R.27. {vi. *2.. We discuss more about orthogonal complements elsewhere in the text. 2.. e jRnxn 4. Example 2. . . Then r1 — r2 = s2— SI. while U n £ is the set of diagonal matrices in Rnxn. Xk E jRn 2. Theorem 2.c jRnxn. For arbitrary subspaces R.. 0 S. EXERCISES EXERCISES 1. Xk} must be a linearly independent set. Among all the complements there is a unique one orthogonal to R. Suppose {VI.. together with Examples 2.26...c = jRnnxn jRn xn. Vector Spaces Chapter 2.29. Example 2. the set in jRnxn.27.. But as t = r1 + s1 = r2 + S2.29. S2 e rl Sl r2 Then r.si e S. r2 E Rand s1. Then show that one of the vectors 1. and let S Let (V.vn be orthonormal vectors in R". ft.r2 £ Rand 52 . Since ft fl 0. Suppose T = R O S. . Let U be the subspace of upper triangular matrices in E" x" and let £ be the subspace of lower triangUlar matrices in Rnxn. Then Theorem 2.. 3. 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:(A-A). ft S) = jR2 and let ft be any line through the origin. = dim(ft) + Proof: To Proof: To prove the first part.. mutually [x\. Xk} must be a linearly independent set. validity of the formula given in Theorem 2. IF) = (jRnxn.. Vk} is a linearly dependent set. we must have rl = r2 and s\ = si from S2 from which uniqueness follows. let R be the set of skew-symmetric matrices in (V.28.25.. suppose an arbitrary vector t E T can be written in two ways t e as t S2. Example 2.2 and 2. ft. .20. x/c E R" be nonzero mutually orthogonal vectors. and SI.. v = [4 l]r jR2. of the formula given in Theorem 2. 0 The statement of the second part is a special case of the next theorem.27. XI.25. S of a vector space V. . jRn xn . But r1 –r2 E ft and S2 — SI E S... jRnxn.14 14 Chapter 2. Then V = U $ S.. triangular + L = R xn un. Using the fact that dim {diagonal (diagonal matrices} = n.. Show that {XI. 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 .28.. every t € can be written uniquely in the form r s with r e R and s e S. . 2. dim(R + S) = dim(R) + dim(S) - dim(R n S).. F) (R n x n . Vector Spaces Remark 2. Then it may be checked that U + . . Theorem 2. Then any other distinct line through the origin is a complement of R.20. Vn thonormal if and only if A E R"x" is orthogonal. . every t E T can be written uniquely in the form tt = r + s with r E Rand s E S.r2 S2 . .. . Show that Av\. X2.. D Theorem 2. Consider the vectors VI — [2 1f and V2 = [3 1f. For arbitrary subspaces ft.. Then any other distinct line through the origin is and let R be any line through the origin. r2 e R.. unique ft. vd must be a linear combination of the others. The complement of R (or S) is not unique. Avn are also orjRn. where rl. jR). R). dim(T) = dim(R) + dim(S)..r . . we must have r\ r-i and SI rl . Let VI.. ft be the set of symmetric matrices in R" x ". one can easily verify the validity = n. . Find the components of the vector v = [4 If with respect to this basis. . which uniqueness follows. Since R n S = 0.c the Example 2. where r1. 2.s\. Let x\. Suppose =R EB Then 1.5. and let R"x"..

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 co-domain.

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-(t-r)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 = pn-l. w = -pn-1. I:- : —> Define C.: V -+ W by I:- p = p', where'I denotes differentiation with respect to x. Lp — p', where denotes differentiation x.

17

{W jj' j e !!!. j E !!!. is arbitrary). 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. Change of basis then corresponds naturally to appropriate matrix multiplication. We identify A the equation £V = W A becomes simply £ = A.m and {Vi. for V and W) is the representation of £i>. respectively.. F) is linear and further suppose that {Vi.. F) —>• (W." to Rm usually causes no Thinking of A both as a matrix and as a linear transformation from Rn to lR..18 Chapter 3. [ w .. Thus. W = R m and [ v i . L IF) ~ (W. Linear Transformations 3. with respect to {w }•. + . Li near Transformations Chapters. IF) veniently in matrix form. transformation with its matrix representation.. i. then arbitrary).n.e..2 Matrix Representation of Linear Transformations Matrix Representation of Linear Transformations Linear transformations between vector spaces with specific bases can be represented conLinear transformations between vector spaces with specific bases can be represented conSpecifically. z'th V This could be reflected by subscripts. LV WA since x was arbitrary. i.. j E raj.. Thus. Note that A = Mat £ depends on the particular bases for V and W. w ] and L is the ith column of A. Specifically.. W = lR. . + amiWm where W = [w\. £V = W A since x was arbitrary. and hence x. + ~nLvn =~IWal+"'+~nWan = WAx. j E m} are the usual (natural) bases WA linea LV L = A. in the notation. suppose £ : (V. say. is by its action on a basis. Thus. In other words. and hence jc.. When V = JR. n A= al : ] E JR. w m] and where W = [WI. i e n} and {Wj. j e m}. Thus. Thinking of both as a matrix and as a linear transformation from JR.m usually causes no naturally confusion. .} are the usual (natural) bases... {u.} are bases for V and W. In other words. if V = E1v1 + . 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. We thus commonly identify A as a linear transformation with its matrix representation.2 3. i e ~}. i E n}.mxn a mn represents L since represents £ since LVi = aliwl =Wai... When V = R". but this is usually not done.e.• + E nVnn = V x (where u.. if v = ~I VI + • • + ~n v = Vx (where v. then LVx = Lv = ~ILvI + .. Then the {w j.

That is. The above is sometimes expressed componentwise by the C — A B .3. Note that in most texts. Inner Product: n xTy = Lx. the form XXT (or xx HH).3. xx T XX ). Outer Product: matrix matrix mxn E R Note that any rank-one matrix A e ~mxn can be written in the form A = xyT = xyT H mxn mxn). we have C A B . dimV = n. If dimU = p. dim V = n.. y E Rn. If dimZ// = p.3. and if we associate matrices with the transformations in the usual way. Composition of Transformations 19 19 3.y.=1 Outer Product: Let x e Rm. . and dim W = m. the arrows above are reversed as follows: C However. then composition of transformations corresponds to standard matrix mUltiplication. Then their outer product is the m x n E ~m. That is. then composition of transformations corresponds to standard matrix multiplication. Then their inner product is the scalar E ~n. Composition ofTransformations 3.3 Composition of Transformations Composition Consider three vector spaces U. and if we associate matrices with the transformations in the usual way. V. and dim W m. y e ~n. Then we can define a new transformation C as follows: C The above diagram illustrates the composition of transformations C = AB. expressed mxp nxp formula cij = L k=1 n aikbkj. Two Special Cases: Two Special Inner Product: Let x. . A rank-one symmetric matrix can be written in above (or xy if A E C xyH e c ). Note that in The above diagram illustrates the composition of transformations C = AB. 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. and W and transformations B from U to V and A from Wand V to W. Then we can define a new transformation C as follows: to W. y e Rn. in the same order in both the diagram and the equation.

Note that N(A) and R(A) are. ~.Vk } With Vi E. vi ^/v'k vk ~~~ ] . denoted N(A). orthonormal set. See also the last paragraph of Section 3. 2. The nullspace of kernel of and A is also known as the kernel of A and denoted Ker (A). then {I —/==. Let A : V --+ W be a linear transformation.[ -:~~ J} . the same symbol (A) is Note that in Theorem and throughout the text.2. . R(A) C W. e Rn. is the {v e V Av = 0}...20 20 Chapter 3. 2.. Definition 3. then Proof: The proof of this theorem is easy. subspaces of different spaces. is the set {v E V : Av = O}. in general. is an orthogonal set. usual (natural) bases.. {[ ~ J. is the set {w E W : w = Av for some v E V}. if i f= j. . ~} | ISisan 3. R(A) = {Av : v E V}.4 3. ..3. Let {VI. The nullspace of The of denoted N(A). {v1. where 8ij is the Kronecker delta defined by Kronecker delta defined by 8 = {I0 ij ifi=j. Then Let A : V —>• be a linear transformation.7. Theorem 3. Equivalently. € 1Tlln is an orthogonal set. .. I ~VI VI ^/v.. [-: J} is an orthogonal set. Let A E Rmxn. . —/=== . Definition 3.6.3. 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. {[ ~~i ]. Note that in Theorem 3.IN.5.. V. denoted Im(A). The nullspace of A. an} . Vk} with u.. If {VI. Example 3. essentially following immediately from the defiProof: The proof of this theorem is easy.4 Structure of Linear Transformations Structure of Linear Transformations Let A : V —> W be a linear transformation.. denotedR(A).an]. 3." • orthonormal set. be orthogonal if' vjvj 0 for i ^ j and orthonormal if vf vj 8ij' where 8tj is the be orthogonal if vr v j = 0 for i f= j and orthonormal if vr v j = 8ij. The range of A.. D Remark 3. . Li near Transformations 3. . Then 1. 2. denotedlZ( A)... vd of u. Definition3. essentially following immediately from the definition. N(A) c V.8. The range of A. Note that N(A) and R(A) are. .. in general. e ~mxn. is an orthonormal set. The range of A is also known as the image of A and — {Av e V}. .i •..7. 0 nition.-. Let A : V --+ be a linear transformation..2.•. is an orthonormal set. If in of = [a\. then then R(A) = Sp{al. . subspaces of different spaces. R ( A ) S. an]. an M. LinearTransformations Chapter 3. vk] be a set of nonzero vectors Vi E ~n. The set is said to 3.. If A is written in terms of its columns as A = [ai. then ~ . IS an orthogonaI set. ~ 3 . { t > . . of of denoted Im(A). Theorem 3. See also the of Section 3. is the set {w e w = Av for some v e V}. 1. 1.. N(A) S.4. W. (A)..5 and throughout the text.8.

Structure of Li near Transformations 3. Any set of vectors will do. n S~. Example 3. The proofs of the other results are left as Proof: left exercises.. Then the orthogonal complement of S is defined as the set c ]Rn.= {v e Rn : vTs=OforallsES}. Let R S C Rn The S <. the computation involved is simply to find all nontrivial (i. Vk} be an orthonormal basis for S and let x E Rn be an arbitrary {v1. nonzero) solutions of the system of equations 3xI -4xI + 5X2 + 7X3 = 0. + X2 + X3 = 0. 6. (n n S)~ = n~ + S~. Set XI X2 = L (xT Vi)Vi. S1. Theorem 311 Let Theorem 3. . Let {VI.10. k =X .3. if and only if S~ <.e..=1 -XI. Proof: We prove and discuss only item 2 here. then give rise to redundant equations). vk} e ]Rn vector.10. S \B S~ = ]Rn. Then n. . including dependent spanning vectors (which would. Then the of defined T S~={VE]Rn: V S = 0 for all s e S}. Let 3.4. Set vector. (S~)l. Let S <. Rn. = S.. 4. . of course.4.9. 3. n~... (n + S)~ = nl. Structure of Linear Transformations 21 21 Definition 3. Then it can be shown that Working from the definition. n <. ]Rn. 2. Note that there is nothing special about the two vectors in the basis defining S being orthogonal. .11. S 5..

. -XI) (which follows by rearranging the equation XI +X2 = x. Then Theorem 3. Then 1. x~ -X2 = -(x.5 3. x E R(A r ) Since x 1 have established thatN(A). everything in S (i. Vk and hence to any linear combination of these we see that X2 is orthogonal to VI. including itself) is O. D Theorem 3. ft(Ar) (i. D Definition 3.l = Rn.•. (Note: This holds only for finite-dimensional vector spaces. where x e U(A) and y e ft(A)1..13. Rm = 7l(A) 0 M(AT)). . Let A : IRn —> Rm. right nullspace of A. 2. established that N(A) U(AT ). Ax = 0 if and only if x equivalent to yT Ax = 0 for all y. We have thus shown that vectors.l. (Note: This also holds for infinite-dimensional vector spaces. that x = XI for example. We S n S1 =0 the e orthogonal everything in (i. Ax = Proof: To prove the first part. Since x was arbitrary.l = 0 since the only vector s E S orthogonal to S1 = IRn.26. many properties of A can be developed in terms of the four fundamental subspaces to IRm.5 Four Fundamental Subspaces Four Fundamental Subspaces Consider a general matrix A E lR. x e R(AT).) 2. Suppose. the right nullspace is N(A) while the left nullspace is J\f(AT).l where x € M(A) and y € J\f(A)± = R(AT) (i.) Proof: To x E N(A)..12 and part 2 of Theorem 3. we see that x2 is orthogonal to v1. Li near Transformations Chapters. X2 is orthogonal to any vector in S.x1) (x'1 xd x2 — X2 = — (x'1 — x1) (which follows by rearranging the equation x1+x2 = x'1 + x'2).xn.11 can be combined to give two very fundamental decompositions damental and useful decompositions of vectors in the domain and co-domain of a linear and transformation A. that x = x1 + x2. .XITVj =XTVj-XTVj=O.. Linear Transformations Then x\ E <S and.l = 7£(AT). Then {v E R" : Av = 0} is sometimes called the right nullspace of A. when we have such direct sum decompositions. But yT Ax = (ATyy{ x. 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. See also Theorem 2.e. 0 The proof of the second part is similar and is left as an exercise. y. (w e Rm : w T A = 0} is called the left nullspace of A. It can write vectors in a unique way with respect to the corresponding subspaces. Suppose. Then R(A r ). every vector v in the domain space R" can be written in a unique way as v = x + y.1 in the next section. X2 is orthogonal to any vector in S. XI = x.e. where XI. Then Ax = 0 and this is an and equivalent to yT Ax = 0 for all v.1 in the next section. But yT Ax = ( A T ) x. Let A : IRn -> IRm. i. In other words. IRm = R(A) EBN(A T». Let A : Rn -+ IRm. Similarly. We have thus shown that S + S. many properties of A can be developed in terms of the four fundamental subspaces . take an arbitrary x e A/"(A). 'R. The proof of the second part is similar and is left as an exercise.XI/ (x~ . Similarly. (Note: This also holds for infinite-dimensional vector spaces.e.12.26. Let A : Rn -+ Rm.22 22 Chapter 3. can write vectors in a unique way with respect to the corresponding subspaces.. + x~. -– x1) = 0 since 0 by definition of S. (Note: This for finite-dimensional 1. It is also easy to see directly that.. But then (x.l N(A T (i. Thus.e.l. IRn = M(A) EB R(A T)). +x~). Ax = 0 if and only if x orthogonal is orthogonal to all vectors of the form AT y. Vk and hence to any linear combination of these vectors. We also have that S U S. See also Theorem 2. When thought of as a linear transformation from E" Consider a general matrix A € E^ x ". x~ e S. E S and X2..l. every vector w in the co-domain space IRm can be written in a unique way as w = x+y. . 3. Thus. x'2 E S1. In other words.(A)1~ — J\f(ATT ).. standing Figure 3. Then Theorem 3. since T x 2 Vj = XTVj . E N(A) and E N(A).e. R(A).14 (Decomposition Theorem). Then {v e IRn : A v = O} is sometimes called the Definition 3. every vector v in the domain space IRn can be written in a unique way as v = x 7.e. the right nullspace is A/"(A) while the left nullspace is N(A T ). {w E IR m : WT A = O} is called the left nullspace of A.. But then (x'1 —XI)TT (x. x 1 E Sand x2. Theorem 3.13. including itself) is 0. transformation A. Let A : IRn -> Rm. every vector in the co-domain space R m can be written ina unique way asw = x+y.14 (Decomposition Theorem).. we decompositions. Clearly. Let A : R" -+ IRm. i.) 2.. and x2 = x~. Clearly. = x'1+ x'2. since Then XI e S and.X2) 0 since (x'1 — x1) (x' 2 — x2) = 0 by definition of ST. we form AT v..) N(A)1" spaces. Then T (x. N(A). Thus.12.11 can be combined to give two very funTheorem 3.. for example.= Af(AT) ) (i.e. E R(A) and E R(A).e. where x\. Theorem 3. .l = R(A T}. x.l = N(A ). R" N(A) 0 ft(Ar ». Then X2 = x. When thought of as a linear transformation from IR n to Rm. 0 x1 — x'1 andx2 = x2. Thus.12 and part 2 of Theorem 3..

Two equivalent 2. A is onto (also called epic or surjective) ifR. fundamental subspaces. A is onto (also called epic or surjective) ifR(A) = W. and in illustrating concepts such as controllability and observability. IR n -> IRm. N(A).16.(A)^. Two equivalent characterizations of A being 1-1 that are often easier to verify in practice are the characterizations of A being 1-1 that are often easier to verify in practice are the following: following: (a) AVI = AV2 (b) VI ===} VI = V2 . 3.5. A is one-to-one or 1-1 (also called monic or injective) if N(A) = O. Four Fundamental Subspaces 3. be a linear transforDefinition 3. Then rank(A) = dim R(A). Four fundamental subspaces.16. Let A : E" -+ Rm.1 makes many key properties seem almost N(A)T. Figure 3. 1.1. The row rank of A is column rank of of independent row rank of . This is sometimes called 3. R(A)1-. Let and W be vector spaces and let A : motion. rank(A) dimftCA). Definition 3. A is one-to-one or 1-1 (also called monic or infective) ifJ\f(A) = 0.5. 2. 'R. Let V and W be vector spaces and let A : V -+ W be a linear transforDefinition 3. t= V2 ===} AVI t= AV2 .3.15. and N(A)1-.(A) = W. properties 7£(A). Four Fundamental Subspaces 23 23 A r N(A)1- r EB {OJ X {O}Gl n-r m -r Figure 3.1. A f ( A ) .15. 1. Figure 3. the column rank of A (maximum number of independent columns). mation.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. R(A).

nullity(B) :s nullity(AB) :s nullity(A) 4.19 suggests looking at the general problem of the four fundamental Part 4 of Theorem 3. . To see that T is also onto.11 and 3. . Finally. . if {ui.(A) = dimA/^A^ 1 if that if {VI. + rank(B) - n :s rank(AB) :s min{rank(A).. dimA/'(A) ± (Note: 1 T T ). the subspaces themselves are not necessarily in the same vector space. R(A) : ]Rn ~ ]Rm. shows that T is onto. Let A : R" ~ Rm. dimA/"(A) + dimft(A) = dimension of the domain of A. LinearTransformations Chapter3. . Let A. Let A : Rn -> Rm. Part 4 of Theorem 3. of Corollary 3.19. .18. e ]Rnxn. rank(B)}. dimension of the domain of A. . Theorem 3. O:s rank(A 2.") of A. then {TVI. Like the theorem. The dual notion to rank is the nullity R(AT) of independent rows). and products of matrices. 1. the subspaces themselves are not necessarily in the same vector space. 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.19 suggests looking atthe general problem of the four fundamental subspaces of matrix products. 3... where n is the ]Rn -> ]Rm. B E R" xn .andx22 e A/"(A). and is defined as dimN(A). u. {Tv\. Theorem 3.. We thus have that dim R(A) = dimN(A)-L since it is easily shown T dim7?. . 3.24 24 Chapter 3.. and is defined as dim A/"(A). rank(AB) = rank(BA) = rank(A) and N(BA) = N(A). sometimes denoted nullity(A) or corank(A). . Proof: From Theorems 3. Linear Transformations dim 7£(A r ) (maximum number of independent rows). 1 1 Xl E A/^A) .11 and 3. The basic results are contained in the following easily proved following theorem. colloquially of = rank of A. of A. r*i *i E N(A)-L. of A. x x e R" x\ X2. v r } is a basis for N(A)-L. . Then dim K(A) = dimNCA)-L. it is a statement about equality of dimensions.18.") 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. by definition there is a vector x E ]Rn such that Ax = w. this theorem is sometimes colloquially stated "row rank of A = column N(A)-L = R(A A/^A) " = 7l(A ).. . Write x = Xl + X2. The last equality AXI x\ e N(A)-L and jc E N(A). rank(A) + B) :s rank(A) + rank(B). 0 For completeness. Then 3. ..17. take any W e R(A).(A).. Then Ajti = W = TXI since Xl e A/^A). + nullity(B). where Ax — w. Tvrr]} is a basis for R(A).17 we see immediately that Proof: From Theorems 3.17 we see immediately that n = dimN(A) = dimN(A) + dimN(A)-L + dim R(A) .17. dimensions. iv} abasis forA/'CA) . we include here a few miscellaneous results about ranks of sums completeness. Clearly T is 1-1 (since A/"(T) = 0). (Note: Since 3. denoted nullity(A) or corank(A). Tv abasis 7?. if B is nonsingular. Then dimN(A) + dim R(A) = n. Then N(T) = To w E 7£(A).19. following follows we apply this and several previous results." 0 of D The following corollary is immediate.

e. Thus. Let e IRmxn. D D 1-1.22. R(B T ). equivalently. RCAB) S.23. and A-I A. nonsingular ifand only ifrank(A) = n. x AT (AAT)-I e IRn. y E R(A). = R(A T A). 1. Then 3. Conversely. Conversely. A is AT AXI AT AX2. equivalently.20. Then Theorem 3. 1. R(A) = R(AA T ).and R(A). which implies that dimN(A)1-1 = n — Proof of part 2: If A is 1-1. and hence dim 7£(A) = n by Theorem 3. A is 1-1 if and only ifrank(A) = n (A has linearly independent columns or is said 2. equivalently. The transformations AT are all 1-1 and onto between the two spaces N(A)1. then dim V — dim W.3. Theorem 3. RCA). especially when dealing with pseudoinverses and linear least squares problems. AATT is nonsingular). dim7?. and hence dim R(A) n by Theorem 3. especially when dealing with pseudoinverses and is extremely useful in text that follows. have full row rank. A A -T.22. N(AB) . AT A is nonsingular). A : V —» W is invertible (or bijective) if and only if it is 1-1 and onto. let y e Rm Proof: Proof of part 1: If A is onto.—n = dim 7£(A r ). Let A : IRn -+ IRm. 3. A : V -+ W is invertible (or bijective) if and only if it is 1-1 and onto. 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. A € IR~xn. Then y = Ax. Let A E Rmxn. We now characterize 1-1 and onto transformations and provide characterizations in We now characterize I-I and onto transformations and provide characterizations in terms of rank and invertibility. 2. A is 1-1 if and only z/rank(A) = n (A has linearly independent columns or is said to have full column rank. Similar remarks apply to A and A~T.5.17. Definition 3. dim R(A) = m = rank(A). . so A is onto. the transformations A. 1. A"1 ± are all 1-1 and onto between the two spaces M(A) and 7£(A). e IRnxp. AX2. XI = X2 AT A A 1-1. Four Fundamental Subspaces 3. It is extremely useful in text that follows. then N(A) = 0. A : IRn -»• IR n is invertible or Note that if A is invertible. Also. since ArA is invertible. Proof' Proof of part 1: If A is onto. Let A E Rmxn. which implies that dim A/^A). R(AT) 3. then A/"(A) = 0. Then 3. N(A) = N(A T A). A is onto if and only //"rank(A) — m (A has linearly independent rows or is said to 1. AT A nonsingular). Note that in the special case when A E R"x". Conversely. Let jc = AT(AAT)~]y Y E Rn. equivalently. linear least squares problems. Also. 2. then dim V = dim W.(A). let y E IRm be arbitrary. Four Fundamental Subspaces Theorem 3. A : W1 -+ E" is invertible or nonsingular if and only z/rank(A) = n. Conversely.20 and is also easily proved. Note that if A is invertible. suppose Ax\ dim R(A T). A is onto if and only if rank (A) = m (A has linearly independent rows or is said to have full row rank. e IRmxn. N«AB)T) .. 4. terms of rank and invertibility. A Proof of part 2: If A is 1-1. which implies x\ = x^. N(A T ) = N(AA T ).21.2 N(A T ). e 7?.(A) — m — rank (A).17. Then A r A. 25 25 The next theorem is closely related to Theorem 3. AT.20. Ar. It The next theorem is closely related to Theorem 3. B E Rnxp.5. Theorem 3.2 N(B). Definition 3. suppose AXI = Ax^. AA is nonsingular). : R n -» Rm. 4.20 and is also easily proved. 2. R«AB)T) S. i.ti = AT Ax2. to have full column rank.21.23.

If there exists a unique right inverse A~R such that AA~R = I. 1. Notice the and leave the following: following: A(A. If there exists a unique left inverse A~L such that A~LA = I.I) must be a right inverse and.R AA. Theorem 3. therefore. Then Definition 3. then one (Proo!' = [1 2]:]R2 -+ ]R .L = (ATTA)-I1AT. i..both 1-1 and is if and if onto. then a left inverse is given by A -R = AT (AAT) left T L = A. i. Let A : V -> W.24. A is left invertible if and only ifit is 1-1.22 ]Rn ->• ]Rm Note: From Theorem 3. Defileft If linear concepts left nitions of these concepts are followed by a theorem characterizing left and right invertible transformations.R = _~] (=1) and A~R = [ _j j is a right inverse.25 that A is invertible. 1..26. Let -+ V. it may still be right or left invertible. then A is invertible. Let A : V -+ W. in which case A~l = A~R = A~L. by uniqueness it must be A -R + A -R A — = A -R. linear Transformations If a linear transformation is not invertible..e. if A is 1-1. D Example 3.. Let A : V -+ Then 1. A~ (A A)~ A . (A R + A RA . It then follows from Theorem 3. —> transformation if left -+ 2. A is invertible if and only if it is both right and left invertible.26 Chapter 3. right inverses for A.e..R + AA-RA = I A +IA - A since AA -R = I = I. Theorem 3. Let A : V -» V. A right invertible if and only if it onto. Let Theorem 3.e. that A~R is a left inverse. then a right inverse is given by A~R = AT(AAT) -I. A -R the case that A~R + A~RA . A is right invertible if and only if it is onto. are infinitely A. A.: AA -R = w Iw W -+ V such that AA~R = Iw. Let A = [1 2] : E2 -»• E1I.24.25.R + A-RA -I) = AA. (Proof: Take any a E E1I. Similarly. 3.26. where Iv denotes the identity transfonnation on V.27.25 that A is invertible. it is clear that there are infinitely many right inverse.I = A~R. It then follows from Theorem 3. Then A is onto. Definition 3. . 0 a left inverse. where Iw denotes the identity transfonnation on W. by uniqueness it must be Thus. can always find v e E2 such that [1 2][^] = a).e. Obviously A has full row rank can always find v E ]R2 such that [1 2][ ~~] = a).. characterizing all solutions of the linear matrix equation AR = I. 2. then A is invertible. Then -> 1. Obviously A has full row rank (= 1) and A . Li near Transformations Chapters. But this implies that A~RA = /. A is said to be right invertible if there exists a right inverse transformation A~RR : if A. i.. (A -R + A -R A — /) must be a right inverse and. is left invertible if and if it and left invertible. If Proof: proof of second Proof: We prove the first part and leave the proof of the second to the reader. 1. € ]R . in A -I = A -R = A -L. Also. both 1-1 and Moreover.R = If left A -L A -L A = 2. A -R A = I. such that A~LA = Iv where Iv denotes the identity transformation on V.22 we see that if A : E" -+ Em is onto. i. In Chapter 6 we characterize all right inverses of a matrix by Chapter characterize characterizing all solutions of the linear matrix equation A R = I. therefore. 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.

The matrix A = 1 1 2 1 [ 3 1 when considered as a linear transformation onIE \ is neither 1-1 nor onto.4. matrix characterizing all solutions of the linear matrix equation LA = I. respect to this inner product. In Chapter 6 we characterize all left inverses of a matrix by characterizing all solutions of the linear matrix equation LA = I. In Chapter 6 we characterize all left inverses of a infinitely many left inverses for A. Show that. ThenAis 1-1. is neither 1-1 nor onto. I-I? Is £. £. For matrices X. EXERCISES EXERCISES 3 4 1. J E2. — S^. Consider differentiation £ 1-1? Is£ onto? onto? 4. We give when considered as a linear on ]R3. Again. Y E Enx" define their inner product by (X. Prove Theorem 3.4. Y) = Tr(X Tr F). 'R. whence A/"(A) = 0 so A is 1-1). 3. below bases for its four fundamental subspaces. it is clear that there are A -L = [3 — 1] infinitely many left inverses for A. Consider the vector space R nx " over E. R = S J. 2.1] is a left inverse. consider A linear transformation ]R3 1.2. let S denote the subspace of symmetric matrices. let denote the subspace of symmetric 2. . For matrices matrices.3. Consider the differentiation operator C defined in Example 3. whence N(A) = 0 so A is 1-1). 4. LetA [J] : E1 ~ E2. (Proof: The only solution to 0 = Av = [I2]v is v = 0. Let A = [8 5 i) and consider A as a linear transformation mapping E3 to ]R2. Then A is 1-1. Consider the vector space ]Rnxn over ]R. Let A = [i]:]Rl -> ]R2. 3. 2 . and let 7£ denote the subspace of skew-symmetric matrices. It is now obvious that A has full column rank (=1) and A~L = [3 . The matrix 3. (Proof Theonly solution toO = Av = [i]v 2. We give below bases for its four fundamental subspaces. Let A = [~ . 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 . and let R denote the subspace of skew-symmetric matrices. It is now obvious that A has full column is v 0. Prove Theorem 3.Exercises 27 2. Is £. with Y e ]Rnxn (X. respect to this inner product. y) = Tr(X Y).

. Determine bases for the four fundamental subspaces of the matrix Detennine fundamental A=[~2 5 5 ~]. Are they equal? Is this true in general? If this is true in general. Linear Transformations 7.4. Let = [~ 9. Chapter 3.4. ~ ~ 3 8. Linear Transformations Chapters. Let A = [ J o].2. prove it. 3. Show that AT has a right inverse. if not. provide a counterexample. If E 1R~9X48.11. left T Suppose e Rmxn 9. Determine A/"(A) and 7£(A). Theorem 6. Suppose A E IR m xn has a left inverse. Suppose A € Mg 9x48 .28 5. Are they equal? Is this true in general? DetennineN(A) and R(A). How many linearly independent solutions can be found to the 10. Prove Theorem 3. homogeneous linear system Ax = 0? homogeneous linear system Ax = O? n 3. Modify Figure 3. Rnxm thought of as a transformation from Rm to IRn.Il.1 to illustrate the four fundamental subspaces associated with AT e associated ATE nxm IR from IR m R".12. linearly independent solutions 10.1 11. Prove Theorem 3.

and hence we Then. which was proved by Penrose in 1955. the Moore-Penrose pseudoinverse of A. With A and T as defined above.. 4.1. the definition neither provides nor suggests a good computational strategy good computational strategy for determining A +.. see [22]. A purely algebraic y + characterization of A+ is given in the next theorem. as noted in the proof of Theorem 3. case X = W1 and Y = Rm.---+ R(A) by dimensional Define transformation T : N(A).1. as is shown in the following text. Then. Then A+ is the Moore-Penrose where y = y\ pseudoinverse of A. Although X and y were arbitrary vector spaces above.m We A+ A e lP1.1 Definitions and Characterizations Definitions and Characterizations Consider a linear transformation A : X —>• y. pseudoinverse of A.l. let us henceforth consider the Although X and Y were arbitrary vector spaces above. 29 . as noted in the proof of Theorem 3. 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. Define a transformation T : Af(A)1. T is bijective Cl-l and onto).17.. see [22]." X ".l —>• Tl(A) by Tx = Ax for all x E NCA). Definition 4.17.l. a generIn this chapter we give a brief introduction to the Moore-Penrose pseudoinverse.1 4. Then A+ is the Moore-Penrose j2 with y\ E RCA) and yi E RCA). where and are arbitrary finiteConsider a linear transformation A : X ---+ y.l. which was proved by Penrose in 1955. the Moore-Penrose pseudoinverse of A.Chapter 4 Chapter 4 Introduction to the Introduction to the Moore-Penrose Moore-Pen rose Pseudoinverse Pseudoinverse In this chapter we give a brief introduction to the Moore-Penrose pseudoinverse. brings great notational and conceptual clarity matrix and. With A and T as defined above. The Moore-Penrose pseudoinverse is defined for any matrix and. as is shown in the following text. for determining A+ . and hence we can RCA) —>• J\f(A}~L This transformation can define a unique inverse transformation T-l 1 :: 7£(A) ---+ NCA). where X Xand Y y are arbitrary finitedimensional vector spaces. a generalization of the inverse of a matrix. neither provides Unfortunately. T is bijective (1-1 and onto). problems. characterization of A is given in the next theorem. This transformation T~ + can be used to give our first definition of A+. define a transformation A + y ---+ X by where Y = YI + Yz with Yl e 7£(A) and Yz e Tl(A}L. can be used to give our first definition of A . define a transformation A+ : Y —»• X by Definition 4. let us henceforth consider the X ~n lP1. We have thus defined A+ for all A E IR™xn.

(P2) GAG = G. (4. Given a matrix G that is a candidate for being the pseudoinverse of A. Let A E lR.1. Unfortunately. Such a verification is often relatively straightforward.1]) satisfy properties (PI).2 4. it must be A+. A + always exists and is unique. p. Then A+ [a [! = lim (AT A + 82 1) -I AT 6--+0 6--+0 (4. 19]. neither the statement of Theorem 4. and (P4) but not (P3). terizations. (PI) AGA = A. Example 4. straightforward. Theorem 4.2. A -L = [3 — 1]) satisfy properties (PI). (P2). Let A e R™xn. it must be A +. 19]. (P4) (GA)T = GA. However. Example 4. Also. as with Definition 4. If G the pseudoinverse of A. Unfortunately. L Note that other left inverses (for example. Note that other left inverses (for example. Note that the inverse of a nonsingular matrix satisfies all four Penrose properties. Introduction to the Moore-Penrose Pseudoinverse Chapter 4. X+ = AT(AATT) -I if A is onto (independent rows) (A is right invertible). A+ = (AT A)~ AT if A is 1-1 (independent columns) (A is left invertible).. Example 4. Then G = A+ if and only if Theorem 4. . While not generally suitable for computer implementation. Note that the inverse of a nonsingular matrix satisfies all four Penrose properties.30 Chapter 4.6. A+ = (AT A)-I AT if A is 1-1 (independent columns) (A is left invertible). this characterization can be useful for hand calculation of small examples.1) = limAT(AAT +8 2 1)-1.7." xn.5. A~ = [3 . Example 4.1.2 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. Verify directly that A+ = Example 4. if a t= 0. Given a matrix G that is a candidate for being checkable criterion in the following sense. For any scalar a." xn. characterization can be useful for hand calculation of small examples. A t = AT (AA )~ if A is onto (independent rows) (A is Example 4. (P2) GAG G. While not generally suitable for computer implementation. the Penrose properties do offer the great virtue of providing a checkable criterion in the following sense. whose proof Still another characterization of A + is given in the following theorem.2) 4.3. (P3) (AGf (P3) (AG)T = AG. Furthermore. Let A e R?xn Then G = A + if and only if (Pl) AGA = A. A+ always exists and is unique.5. neither the statement of Theorem 4. and (P4) but not (P3). Let A E lR. Consider A = [']. Consider A = f ] satisfies (P1)-(P4). this can be found in [1. However. then by uniqueness. as a right or left inverse satisfies no fewer than three of the four properties. Then Theorem 4. if a =0. Also. (P2). AG.6. the Penrose properties do offer the great virtue of providing a tional algorithm.2 nor its proof suggests a computawith Definition 4. (P4) (GA)T = GA. whose proof can be found in [1. For any scalar a.4.2 nor its proof suggests a computational algorithm. Still another characterization of A+ is given in the following theorem.4. one need simply verify the four Penrose conditions (P1)-(P4). Such a verification is often relatively satisfies all four. If G satisfies all four. then by uniqueness. p. a right or left inverse satisfies no fewer than three of the four properties.2. Example 4. Introduction to the Moore-Penrose Pseudoinverse Theorem 4.3. = Furthermore. Example 4. Verify directly that A+ = [| ~] satisfies (PI)-(P4). one need simply verify the four Penrose conditions (P1)-(P4).7.

Theorem 4. where D+ is again a diagonal matrix whose diagonc D is diagonal. e jRmxn and suppose Rmxm R n are orthogonal (M is T -1 1 orthogonal if M M ). [~ ~ r ~ =[ 0 Example 4. (A T )+ = (A+{. Then Proof: For the proof. The Proof: Both results can be proved using the limit characterization of Theorem 4.3.3 4. Let S e jRnxn be symmetric with UT SU = D. where U is orthogonal and D is diagonal. 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+{.10. The especially illuminating. For any vector v E M". [~ r 1 =[ 4 4 I I ~l 4 I I 4 4. 27]. elements are determined according to Example 4.9. Let S E Rnxn be symmetric with U TSU = D. if v = O. For A e Rmxn 1.13. 27]. The interested reader can consult the proof in [1. .4. Properties and Applications 4. Theorem 4. simply verify that the expression above does indeed satisfy each c the four Penrose conditions. Then orthogonal if MT = M. Example 4.12.. For all A E jRmxn. 2.7. p.10.7. . For any vector e jRn. 31 31 Example 4. Example 4. Then S+ UD+U T where D+ is again a diagonal matrix whose diagonal elements are determined according to Example 4. Let A E R m x "and suppose UUEejRmxm.3. p.11.12. The interested reader can consult the proof in [1.).4. . Many of these are used in the text that follows.3 Properties and Applications Properties and Applications This section presents some miscellaneous useful results on pseudoinverses.11. Proof: Both results can be proved using the limit characterization of Theorem 4. if v i= 0.4. Many of these This section presents some miscellaneous useful results on pseudoinverses. The proof of the first result is not particularly easy and does not even have the virtue of being proof of the first result is not particularly easy and does not even have the virtue of being especially illuminating. simply verify that the expression above does indeed satisfy each of Proof: For the proof.8.13. are used in the text that follows. where U is orthogonal an Theorem 4. Properties and Applications Example 4.9.8. A+ = (AT A)+ AT = AT (AA T)+. D Theorem 4. 0 the four Penrose conditions. Then S+ = U D+UT.VVEejRnxnx " are orthogonal (M is 4. 4. Example 4.

Theorem 4.32 Chapter 4. [] sufficient reverse-order However. 0 The following theorem gives some additional useful properties of pseudoinverses. [7].12 and 4. then A+ = (ATA)~lAT. As an example consider A = [0 1J and B = [ : J. TTnfortnnatelv. then (AB)+ = B+ A+.14. e. properties Theorem 4. (AB)+ = B?A+. n(A T AB) ~ nCB) . Introduction to the Moore-Penrose Pseudoinverse Chapter 4. 3.g.• Similarly. For e Rmxn . 0 D Theorem 4. in theory at least. [7]. since e lR~xr.. 0 D E lR~xr. in peneraK ucts of matrices such as exists for inverses of products.xm.16. we have B = B (BBT)~\ whence BB+ = Ir. in general. the Moore-Penrose pseudoinverse of any matrix (since AAT and AT A are symmetric)..15. B E Rrrxm. Theorem 4. (AA T )+ = (A T)+ A+. see [5]. [11]. Proof' A+ A Proof: Since A E Rnrxr. (AB)+ = B+A+ if and only if 4.15.g. This A AT AT turns out to be a poor approach in finite-precision arithmetic. xm + T B e Wr . [5]. 1. Proof: Proof: For the proof. N(A+) 5. we B+ BT(BBT)-I. BB+ f r The by taking BI = B.11 nets of matrices such as exists for inverses of nroducts Unfortunately. A\ = A in Theorem 4. 2. necessary and sufficient conditions under which the reverse-order property does hold are known and we quote a couple of moderately useful results for reference.14.17. (AB)+ = B{ Ai. n(A+) 4. 4. [II]. see [9]. If A is normal. For all A E lR mxn . e. [23]). Ir Similarly.15. Proof: Proof: For the proof. [9]. If e lR~xm. Theorem 4. (AB)+ = B+A+. n(BB T AT) ~ n(AT) and 2. 4. .13 can.At = A in Theorem 4. however (see.12 Note that by combining Theorems 4.11 is suggestive of a "reverse-order" property for pseudoinverses of prodTheorem 4. where BI = A+ AB and AI = AB\B+. Introduction to the Moore-Penrose Pseudo inverse 4. and better methods are suggested in text that follows. compute 4.15. then AkA+ = A+ Ak and (Ak)+ = (A+)kforall integers k > O. D takingB t = B. Then (AB)+ = 1+ = I while while B+ A+ = [~ ~J ~ = ~. (AT A)+ = A+(A T)+. whence A+A = f r. (AB)+ = B+ A + if and only if 1.17.. If A e Rnrxr. = N(AA+) = N«AA T)+) = N(AA T) = N(A T). A+ = (AT A)-I AT.16.13 we can. 4. = n(A T) = n(A+ A) = n(A TA). The result then follows by E lR. (A+)+ = A. Then As an example consider A = [0 I] and B = LI. where BI = A+AB and A) = ABIB{. poor (see.

A E IRmxn. fiA+A B. that B = AA+ B. properties of the Moore-Penrose pseudoinverse.4 to compute the pseudoinverse of \ 2 1. A E IRn xn B E E n xm 6. b E E. B E E M X m . 5 e JRn x m . ft(A+) ft(Ar 5. For A E IRmxn. if A is symmetric. Let A G M"xn. (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.i ]. A e IRPxn thatN(A) S. show that (xyT)+ = (x Tx)+(yT y)++yxT. € IRm xm D 6. or orthogonal. For A e R m x n .• 1 2 x. RCA). However. Note: Recall that A e R" xn is normal if AATT = AT A. = B.]. or orthogonal. and D E E mxm and suppose further that D is nonsingular. such as A=[ -b a a b] for scalars a. Theorem 4. 0 EXERCISES EXERCISES 1. show that JV(A) C A/"(S) if and only if BA+ A = B..4 to compute the pseudoinverse of U . Since x was arbitrary. For example. If jc. prove that 7£(A) = 7£(AA r ) using only definitions and elementary properties of the Moore-Penrose pseudoinverse. Since x was arbitrary. a matrix can be none of the skew-symmetric. b e R for scalars a. assume that AA+B = B and take arbitrary y e K(B). so there exists a vector y E Rp such that Ay = Bx. we have shown where one of the Penrose properties is used above. Then we have there exists a vector y e IRP such that Ay = Bx. a matrix can be none of the preceding but still be normal. Suppose A E Rnxp. 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. skew-symmetric. Then Bx E H(B) S.1 D. Then we have Bx = Ay = AA + Ay = AA + Bx. if A is symmetric. Then R(B) c R(A) if and only if Suppose e IRnxp. 4. Proof: Suppose K(B) S.i l . prove that RCA) = R(AAT) using only definitions and elementary 3. e IRnxm.18. Use Theorem 4. y E IRn. 2. prove that R(A+) = R(A T). For A E Rpxn and BE R mx ". (xyT)+ = (xTx)+(yTy) yx T 3. A+ 0 -A+BD. where one of the Penrose properties is used above. problems. assume that AA + B To prove the converse. Then B and take arbitrary y E R(B).Exercises 33 Note: Recall that A E IRn xn is normal if A A = A T A. U(A) if and only if AA+B = B. whereupon y = Bx = AA+Bx E R(A). we have shown that B = AA+B. then it is normal. N(B) and 5 € IRmxn. Then Bx e R(B) c H(A). such as preceding but still be normal. then it is normal. Y e R". R(A) and take arbitrary x e IRm.i D. whereupon there exists a vector x e IR m such that Bx = y. so Proof: Suppose R(B) c U(A) and take arbitrary jc E Rm. For example. However. Use Theorem 4. For A e Rmxn. To prove the converse. Then there exists a vector x E Rm such that Bx = y. Then K(B) S.

This page intentionally left blank This page intentionally left blank .

..• ::::: Urr > as an exercise. ii E !!. i e !!. More S > o r > O. S = diagfcri. vectors. [24.. Vi eE RIRnxr. e IRmxm and V E IR nxn such that V € Rnxn such that UI > diag(ul. for example. .} be a set of corresponding orthonormal eigenvectors and let VI = [v\. U2 E IRrnx(m-rl.1) rxr A = [U I U2) [ ~ 0 0 ][ ] 2 T VI VT (5. . Then there exist orthogonal matrices U E Rmxm and E IR~xn. Let A e R™ x n . Premultiplying by Vf gives Vf A T A VI write ATAVi = VI S2. n}). (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. vn]. The SVD plays a key conceptual and computational of this important matrix factorization.... where S = [J °0]. i. Ch.. . . y22 €E Rnxfo-r^ and the 0-JM^/ocJb in E~are compatibly IRnx(n-r). 6]). role throughout (numerical) linear algebra and its applications. Pre-and postmultiplying by S-I gives the emotion S~l eives the equation (5. its eigenvalues are all real and nonnegative. Pre. for example. and a\ > • • • > Ur > 0. Proof: Since A r A > 00 A r A i is symmetric and nonnegative definite.1 The Fundamental Theorem Theorem A Theorem 5. 6]). .\. VI «xr j V U2 e ^x(m-r) . Let {u.. the latter equality following from the orthonormality of the Vi vectors.. Ch. «}). (5.) Denote the set of eigenvalues of AT A by {U?.. r write A r A VI = ViS2. .• = Un.. < min{m. recall. (Note: The rest of the proof follows [24.u ).Vv r). its eigenvalues are all real and nonnegative. We In this chapter we give a brief introduction to the singular value decomposition (SVD). . specifically... .. rcfr). ..e. . and the O-subblocks in are compatibly dimensioned.. V2 = [Vr+I. .2) (5..} with UI ::::: ....1. . Letting — diag(cri.. The SVD plays a key conceptual and computational role throughout (numerical) linear algebra and its applications.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).3) = Ulsvt· The submatrix sizes are all determined by r (which must be S min{m. dimensioned. 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. .] .o>) E IRrxr.. Theorem 5. Let {Vi.and postmultiplying by equality following from the orthonormality of the r. u r ) e R .LettingSS = diag(uI. U\ E IRmxr. e n} be a set of corresponding orthonormal eigenvectors 0= Ur+1 = .) Denote the set of eigenvalues of AT A by {of / E n} with a\ > • • > a > 0 = o>+i = • • an.e. the latter VfV^S2 = S2. . UI e Wnxr... Premultiplying by vt gives vt ATAVi = vt VI S2 = S2. we have n = U~VT.we can Vi = [vr+ .Vn ]. 5. recall. Proof: Since AT A ::::: ( (AT A s symmetric and nonnegative definite.. . we can and let V\ [VI. i.1.4) 35 . More where ~ = [~ specifically.

the IRmxr VI AViS-I.1.2. with respect to the bases A = U^V as A V Mat £ is S the U E we see respect n and {u I.. The latter equality follows from the orthogonality of the S and vI AVi = vI VI S = O... •.2). .)..'. See also m [v\.r zero singular values. Now V20 Vf A T A V 0. 0 Definition 5. ar}} is called the set of (nonzero) singular values of the matrix A and called [a\. we see that Mat C is "i:. } for R" {VI. Then Specifically. singular 2. Definition 5.4. analogous complex e IC~ xn straightforward.? (AA I min{m. U V identical.. . . of values of i I proof of A. an we have that ATAV2z = VzO = 0. to be ~ completes the proof..(AT A) = is denoted ~(A)... The set {ai.denote A thought of as aalinear transformation mapping IR n to IRm.2.(A) At.e.. let C.4. except for Hermitian transposes replacing transposes. eigenvectors of AA TT). A.16. VT be an SVD of A as in Theorem 5. The analogous complex case in which A E C™ x " is quite straightforward.. u ]} for Rm (see the discussion in Section 3.3. Then T V AV =[ =[ VrAVI VIAVI VrAVI vIA VI Vr AVz vI AVz ] ~] since A V2 =0. Choose U2 £ IRmx(m-r) [U\ U2] orthogonal. The latter equality follows from the orthogonality of the columns of VI and V 2. Specifically. The !:ingular value decomposition is not unique.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. VT as AV = V"i:.16.. Ui e (5. we see that U{ AV\ = S and 1/2 A VI = U^UiS = 0. Let A = V"i:.1. . we see that. n] — Note that there are also min{m. of the proof of Theorem 5. responding to multiple cr/'s.2). Then T rewriting A = V"i:.5. an we Turning now to the eigenvalue equations corresponding to the eigenvalues ar+l. U be interpreted changes domain and co-domain spaces with respect to which A then has a diagonal matrix representation. V H.3. and co-domain spaces with respect to which A then has a diagonal matrix representation. . VTAV = [~ Q]. cr. For example. 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 ). Now define the ATA V AV O. i. Introduction to the Singular Value Decomposition Chapter 5.1 we see that ai(A) = A(2 (ATA) = £(A). D to be S completes the proof.4) see UfU\ = columns of U\ are orthonormal. C denote A thought of as linear transformation mapping W to W. Remark 5.. m for IR (see the discussion in Section 3. Remark 5.4) we see that VrVI = /. in fact. The decomposition is A = V"i:. Remark 5. Remark 5. n} . where V and V are unitary and the proof is essentially proof decomposition A = t/E V H. A = t/E VT SVD of A 5.. there may be nonuniqueness associated with the columns of V\ (and hence U\) cor• there may be nonuniqueness associated with the columns of VI (and hence VI) corresponding to multiple O'i'S. See also m Remark 5. matrix VI E M mx/ " by U\ = AViS~l. vn } for IR and {u\. Referring to the equation V I = A VI S-I defining U\. From the proof of Theorem 5. Thus. and defining this matrix U\ andU UT A V [Q ~]. 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). Choose any matrix V2 E ^ 77IX( ™~ r) such that [VI V2] is VI columns orthonormal.. we see that V r A VI = since A V2 = O.. Thus.. 1. Introduction to the Singular Value Decomposition Turning now to the eigenvalue equations corresponding to the eigenvalues or+\. .5. . .. .. Then from (5. singular unique.. an examination decomposition Remark 5..36 36 Chapter 5. . AV2 = 0. whence Vi ATAV22 = O..(AATT). Remark 5... of A A). 3. Referring to the equation U\ = A V\ S l defining VI. Note that V and V can be interpreted as changes of basis in both the domain Remark 5. .

Then A = V A VTT is an A.10. i. Example A .9. U arbitrary 2 x orthogonal 5. [11]. Let V be an orthogonal 5. see. A _ [ 1 - 0 -~ ] cose = [ . symmetric V orthogonal matrix of eigenvectors that diagonalizes A. [11]. however. if A = UI:VT is an SVD of A.e.10. that "full SVD" (5.2) can always be constructed from a "compact SVD" (5. SVDof A.g. [25]. e j8 the case). Example 5.. V2 (see Theorem 5. VI:TU/ s n SVD of A VS C .6. The Fundamental Theorem 5. is an SVD.U I U T . A=U is an SVD. is the matrix I: and the span of the columns of UI. 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. AT A Remark 5.6. e. 01 - where U is an arbitrary 2x2 2 orthogonal matrix. What is unique.8.3). [7]. i. then A A.7. For example. Example 5.9.sine sin e cose J[~ ~J[ cose sine Sine] -cose ' where e is arbitrary. Computing an SVD by working directly with the eigenproblem for AT A or 5. orthogonal transformations.. see.5.. 0 Example 5.11). f/2.g.. corner f/E V T r r T Ti isaanS V D o f AT.[1 0 ] . Better algorithms exist that work directly on A via a sequence of orthogonal transformations.2) can always be constructed from ¥2 Theorem too.3). Computing AA AATT is numerically poor in finite-precision arithmetic. VT AV = A > O. 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). The Fundamental Theorem 37 37 • any U22can be used so long as [U I U2] is orthogonal. [25]. SVD of A. and E U\. SVD" (5.1. F/vamnlp 5.8. 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). U2. Note. is an SVD. Let A e IRnxn" be symmetric and positive definite.. C/ [U\ Ui] orthogonal. too.1. that aa"full SVD" (5. [7]. VT A V = A = VAV eigenvectors > 0.e. e. VI. Vi.

. Introduction to the Singular Value Decomposition 5. Part 4 of the above theorem provides a numerically superior method for finding (orthonormal) bases for the four fundamental subspaces compared to methods based finding (orthonormal) bases for the four fundamental subspaces compared to methods based on. Let A E Rmxn have a singular value decomposition A = U'£ VT. Theorem 5. (c) R(VI) = N(A)1. urn] and V = [v\. andV2 = [Vr+I. vn]. Part 4 of the above theorem provides a numerically superior method for Remark 5.. u r ].38 38 Chapter 5.= N(A T ). The singular vectors satisfy the relations 3.. Let A E E mx " have a singular value decomposition A = U. . A = U. Let A e jRrnxn have a singular value decomposition A = VLV T Using Theorem 5.11.1. .12. .7) explain why it is conventional to write the SVD and the key vector relations (5.5) as a sum of outer products Remark 5. (b) R(U2) = R(A)1.. [HI.2 Some Basic Properties Some Basic Properties Theorem 5. .. .13. Then A has the dyadic (or outer product) expansion product) expansion r A = Laiuiv. VI = [VI.5) as a sum of outer products and the key vector relations (5. 4.8) where where . um] and V = [VI.. say.. A = UZV. The relationship to the four fundamental subspaces is summarized nicely in Figure 5. The elegance of the dyadic decomposition (5. Remark 5. Introduction to the Singular Value Decomposition Chapter 5.14. The singular vectors satisfy the relations AVi = ajui. Then TheoremS.6) and (5. vr ].. Remark 5. say. . Using the notation of Theorem 5..£V.2 5. Note that each subspace requires knowledge of the rank r.. the following properties hold: 1.11. Then A has the dyadic (or outer 2.6) and (5.£VTT as in Let A e jRmxn have a singular value decomposition A = UHV as in Theorem Theorem 5. for example. . Then (a) R(VI) = R(A) = N(A T / . . . 1. rank(A) = r = the number of nonzero singular values of A. LetUI = [UI. . reduction to row or column echelon form.6) (5.. Let V = [UI. (d) R(V2) = N(A) = R(AT)1-. i=1 (5.]. vn]. . Note that each subspace requires on. rank(A) = r = the number of nonzero singular values of A.13. urn]..7) AT Uj = aivi for i E r. . The relationship to the four fundamental subspaces is summarized knowledge of the rank r...1. (5. U2 = [Ur+I. as A = UZV rather than. 2. Then (5.1. ...12. nicely in Figure 5.5) 3. . the following properties hold: the notation of Theorem 5. for example.= R(A T ). reduction to row or column echelon form.1.1. The elegance of the dyadic decomposition (5. .. Let U =. . Vn]..7) explain why it is conventional to write the SVD as A = U'£VTT rather than.

Remark 5. which is clearly orthogonal and symmetric. if we let the columns of U and V be as defined in Theorem 5.. Some Basic Properties 39 39 A r r E9 {O} / {O)<!l n-r m-r Figure 5.5.=1 U.u. e2. a simple if we insist that the singular values be ordered from largest to smallest. Figure 5. then be as defined in Theorem 5. SVD and the four fundamental subspaces.1. a simple reordering accomplishes the task: reordering accomplishes the task: (5. which is clearly orthogonal and symmetric. 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. Proof' The proof follows easily by verifying the four Penrose conditions.11. Proof: The proof follows easily by verifying the four Penrose conditions..15.2. then = L r 1 -v.. Furthermore. Note that none of the expressions above quite qualifies as an SVD of A + Remark 5. . 0 D (5.11. ed. (or exchange matrix) P = [e er-I. However.. Furthermore. e\\.. if we let the columns of U and V with the Q-subblocks appropriately sized.11) This can also be written in matrix terms by using the so-called reverse-order identity matrix This can also be written in matrix terms by using the so-called reverse-order identity matrix (or exchange matrix) P = \err.er^\.. SVD and the four fundamental subspaces.10) ..2. However. . . with the O-subblocks appropriately sized.15. e^. Some Basic Properties 5.1. .

let A E lR. . . Notice that N(A) . . Remark 5.. Introduction to the Singular Value Decomposition Chapter 5.olumn transformations. Such a compression is analogous to the "compresses" A by I.1).3 5. where R is upper triangular.2.l. .17 and Remark 5.mxn. v } and {MI . is not generally as reliable a procedure.. Then AV = V:E = [VI U2] [~ ~ ] =[VIS 0] ElR. 5..1. the same bases is 5""1..1).1). From Section 3. since {UI. mxr has full column rank. then T can be defined by TV. Column compression Column compression Again. notice that R(A) R(A V) This time. Such a row compression can also be accomplished "compresses" A by row transformations.[ SVr ] 0 mxn E lR.. is not generally as reliable a procedure. Introduction to the Singular Value Decomposition A+ = (VI p)(PS-1 p)(PVr) is the matrix version of (5. Since T is determined by its action on a basis. .vvr}}is aa is r basisforN(A).r) and the matrix SVr e lR.. when derived by a Gaussian elimination algorithm implemented in finite-precision arithmetic..16. In other words. urr]} is clearly S. when derived by a Gaussian elimination algorithm implemented in echelon form which. by orthogonal row transformations performed directly on A to reduce it to the form [~]. Then Let A E lR.. Similarly.17 and in Definition 4. = ^-u. and since ( v \ .. / E~. Both compressions are analogous to the so-called row-reduced echelon form which. premultiplication of A by UT is an orthogonal transformation that rank. .. .2.. This time. Such a compression is analogous to the . . while the matrix representation for the inverse linear transformation T-lI with respect to S. in Definition 4. Recall the linear transformation T used in the proof of Theorem 3. In other words.. premultiplication of A by VT is an orthogonal transformation that "compresses" A by row transformations. then T-lI can be defined by T^'M. In other words. then T can be defined by TVj = OjUj .4). From Section 3. Then Again. the matrix representation for T with respect to the bases {VI. basis forJ\f(A)±. while the matrix representation for the inverse linear transformation T~ with respect to the same bases is S-I.. postmultiplication of A by V is an orthogonal transformation that "compresses" A by column transformations. w. . have an SVD given by (5. vrr} and {u I.1.3 Rowand Column Compressions Row and Column Compressions Row compression Let A E R have an SVD given by (5.16.. A "full SVD" can be similarly constructed. finite-precision arithmetic.M(UT Notice that M(A) = N(V T A) = N(svr> and the matrix SVf E Rrxll" has full row A/"(SV. = tv. . postmultiplication of A by V is an orthogonal transformation column rank. . Both compressions are analogous to the so-called row-reduced where R is upper triangular.1).. . r x has full row rank.11). u is clearly matrix representation for T with respect to the bases { v \ .i / E~. since [u\.. 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.i e r. is the matrix version of (5.mxn have an SVD given by (5. = cr. Recall the linear transformation T used in the proof of Theorem 3. Then VT A = :EVT = [~ ~ ] [ ~i ] D _ . and since {VI. Similarly. u is a basis forR(A). .urr}} e r. A "full SVD" can be similarly constructed. In other words. let A e Rmxn have an SVD given by (5.. Such a row compression can also be accomplished by orthogonal row transformations performed directly on A to reduce it to the form 0 . . then T~ canbedefinedbyT-Iu.40 40 Then Then Chapters. . the isabasisfor7£(. Since T is determined by its action on a basis.11)... . .

For details. 3. Use the SVD to determine a where Q is orthogonal and P p T > O. of defined by A defined by A = xyT.[11]. y E ~n Determine A e ~~ 4. € IRmxn.e. i.Exercises Exercises 41 41 so-called column-reduced echelon form. Determine an SVD of A. . Note: this is analogous to the polar form where Q is orthogonal and P = PT > 0.1 starting from the observation that AAT > 0. Determine SVDs of the matrices 5. (b) Suppose that W and Y are nonsingular but not necessarily orthogonal. Let X E M mx ". Let A e Rmxn and suppose W eRmxm and Y e Rnxn are orthogonal. see. 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). If XT X = 2. Determine an SVD of the matrix A E R™ xn E IRm. A E IRnxn indefinite. an SVD A. [25]. Determine SVDs of the matrices (a) (b) [ ] [ ~l -1 0 -1 6. see. for performed by Gauss transformations in finite-precision arithmetic. [7]. [23]. A = Q P 7. [7]. xyT 5. which is not generally a reliable procedure when performed by Gauss transformations in finite-precision arithmetic.. Note: this is analogous to the polar form iO z = rel&ofaa complex scalar z (where i = j = V^T).. = o. Use the SVD to determine a polar factorization of A. for example. which is not generally a reliable procedure when so-called column-reduced echelon form. show that X = 0. 2. Let E ~~xn. 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. A = QP 7. y e Rn be nonzero vectors. (a) Show that and W A F have the same singular values (and hence the same rank). If XTX = 0. 4. [11]. [25].e. For details. z of complex scalar z (where i j J=I). Prove Theorem 5. [23]. EXERCISES EXERCISES 1.. Let A € R" X M . Let A e E"xn be symmetric but indefinite. Prove Theorem 5.1 starting from the observation that AAT ~ O. Let x e Rm.

This page intentionally left blank This page intentionally left blank .

There exists a solution to (6. equivalently. 4..e. i. There exists a nontrivial solution to the homogeneous system Ax = 0 if and only if Ax = 0 if 6. and this is possible only ifm > n. A solution to (6.3} for e R m if only ifU(A) = W". as a special case. 2. There exists at most one solution to (6.3) 1. A is 1-1.1 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 ifJ\f(A) = 0.3) is unique if and only if N(A) = 0.. n this is possible only ifm < n (since m dimT^(A) = rank(A) < min{m..3) for all b E ]Rm if and only if A is nonsingular. and this is possible only ifm :::: n (since m = dim R(A) = rank(A) :::: min{m. N(A) 0. and this is possible only ifm ::: n.e.3) for all b e W" if and only if is nonsingular. A solution to (6. We begin with a review of some of the principal results associated with vector linear systems.. equivalently. i. Consider the system of linear equations Ax = b. There exists a unique solution to (6. (6. b]) = rank(A). 1. (6. General linear systems of the form (6.2) 6. A is 2. i. b E ]Rn.3) if and only ififbeH(A).3) for all b e W1 if and only if the columns of A are linearly independent. and onto. 3. 6. the familiar vector system Ax = b. equivalently.1.1) are studied and include. the familiar vector system are studied and include. as a special case. i. n. 3. Consider the system of linear equations Theorem 6.1 6. A E ]Rn xn. A is 1-1. rank(A) < n. A G M m x m and A has neither a singular value nor a eigenvalue.e. (6. there exists a solution if and only j/"rank([A. i. There exists a unique to (6..e. 5. there exists a solution if and only ifrank([A.. i. There exists a solution to (6.1. There exists at most one solution to (6.3) for all b E lRm if and only ifR(A) = lRm. A/"(A) = 0. General linear systems of the form equations. Theorem 6. 4.3) for all b E lRm if and only if the columns of 5. onto. n}).3) if and only b E R(A). b]) = rank(A). equivalently.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. 43 . A E lR mxm and A has neither a 0 singular value nor a 0 eigenvalue. A are linearly independent. A E lRmxn. There exists a solution to (6. b E lRm.e.e. only if rank(A) < n.

A is not 1-1. note that x = 0 is always a solution to the homogeneous system. i.6) Furthermore.e. 0 . R(A).mxn. mxn . Let Z be an arbitrary solution of That all solutions arc of this form can be seen as follows. i.1). of a matrix.e . all solutions of (6. we must have the case of a nonunique solution. A is not I-I.. AZ — B.1).6).A+ A)Y.mxk. Theorem 6. (6. and this is clearly of the form (6. That all solutions are of this form can be seen as follows. AZ :::: B. a solution exists if and only if AA+B = B.18. to algebra. Let Z be an arbitrary solution of (6.2 (Existence). Furthermore. specializing even further to the case m = n. i. which implies rank(A) < n by part 0 by part 3. Then we can write (6.18. Therefore.2) follow by specializing even further to the case m = n.. 0 Theorem 6. A E JR.1 follow from those below for the special case = 1. a solution exists if and only if has a solution if and only ifR(B) S. Linear Equations Chapter 6.2 6. +B = Theorem 6. to prove part 6. we prove part 6. all solutions of (6.6).2 (Existence). E JR. (6. AA+B B. i. Let A e Rmxn. The matrix criterion is Theorem 4. where Y E JR.5). 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. which implies rank(A) < n must have the case of a nonunique solution. Proof: To verify that (6. B E JR. The matrix linear equation Theorem 6.4) has a solution if and only ifl^(B) C 7£(A). The matrix criterion is Theorem 4.5) is a solution of is a solution of AX=B. (6.6) are of this form. D 6. equivalently. Then we can write Z=A+AZ+(I-A+A)Z =A+B+(I-A+A)Z and this is clearly of the form (6. Note that some parts of the theorem follow directly from others. note that x 0 is always a solution to the homogeneous system. 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. BE JR. Note that the results of Theorem 6.5) is a solution..e.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 (6.5). while results for (6. premultiply by A: AX = AA+ B + A(I = B A+ A)Y + (A - AA+ A)Y by hypothesis = B since AA + A = A by the first Penrose condition. Note that the results of Theorem of solutions to the general matrix linear system (6.5) is a solution. The matrix linear equation AX = B. For example.44 Chapter 6..1 follow from those below for the special case k = 1. For example. Note that some parts of the theorem follow directly from others.e.6) are of this form. equivalently.3.nxk is arbitrary. Then any matrix eRmxk of the form of the form X = A+ B + (/ . Therefore. while results for (6.2)follow by 6.mxk and suppose that AA+B = B.3. premultiply by A: Proof: To verify that (6.

A+ A)Y =A++(I-A+A)Y.mxn.mxk (6. When A is square and nonsingular.6. equivalently. BE lR.6 (Uniqueness). It can be shown that the particular solution X = A+B is the solution of (6.6) that minimizes TrX7 (Tr(-) denotes the trace of a matrix. Clearly.6) +B Remark 6.4. Matrix Linear Equations 45 Remark 6.8) . A R (AA(A) = A"1.7) has a unique solution if and only if M(A) = 0. it is not unique.8. Here.A+ A V2 V[ and U(V = K(V2) = N(A). and (N(A) = 0). Consider the system of linear first-order difference equations (6. A e E"x".nxn. 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). equivalently. But if A has an SVD given by A = U h VT.3. Thus. There is a unique right inverse if and only if A+A = I/ e E"xm arbitrary. matrix. Example 6. (6.A+A) = O.mxn. Find all solutions of the homogeneous system Ax = 0. then it is easily R(I — A + A). Clearly. Ax — 0. Proof: Proof: The first equivalence is immediate from Theorem 6. find all A e ]Rmx". N(A) = O.9. X• = A~ B.) Theorem 6.n is arbitrary. where r rank(A) (recall r < n). A+ = A"1 and so (I . A solution of the matrix linear equation AX = B. wherer = rank(A) (recallr ::: h). A solution of the matrix linear equation Theorem 6. D 0 Example 6. rank(A) = < A This is equivalent to either rank (A) = r < n or A being singular. this can occur if and only if rank(A) = r m (since equivalent to AA+Im = 1m. All right inverses r < m) A (A+ of A are then of the form of A R = A+ 1m + (In .7) is unique if and only if A+A = /. we write 1m to emphasize the m x m identity Im matrix. there exists a nonzero solution if and only if A+A /= I. (6. vD Example 6. y E R" A + A t= I.A+A = Vz V2 and R(Vz2V^) = R(Vz) = N(A). nonzero solution. A A+ A-I Remark (/ — A + A) 0. The second follows by noting thatA+A = I can occur only ifr = n.nxm is arbitrary. Remark 6. (TrO denotes the trace of a matrix. Characterize all right inverses of a matrix A E lR. where y e lR. It particular (6. Clearly.2. equivalently. Hence. in which case A must be invertible and R = A-I. it is easy to see that all solutions are generated y from a basis for 7£(7 .j jcj. Solution: x=A+O+(I-A+A)y = (I-A+A)y.9. recall that TrXT X = £\ •xlj. if and only if A is I-lor _/V(A) = O. Butrank(A) = n if and only if A is 1-1 or N(A) = 0.A+A). Characterize AR = Im solutions R of the equation AR = 1m. Consider Example 6. if there exists a unique.7.6 (Uniqueness). A+ A where Y E lR.2..7) has a unique solution if and only if unique if and only if A + A = I. Computation: Since y is arbitrary. A E lR.5. A A = f/E VT. Solution: There exists a right inverse if and only if R(Im) S. / .S. Example 6.7.) that minimizes TrXT X. leaving only the unique solution X = A-I1B.5. Matrix Linear Equations 6. But rank(A) = n that A+ A = / if r — n. r checked that 1. there is no "arbitrary" component. Example 6. 7£(A) and this is 7£(/m) c R(A) equivalent to AA + 1m Im. recall that TrX r = Li. Suppose A E lR.

.10.8) of Example 6.8) is given by solution of (6. The matrices A = [~ ~]1and B5 == [~] 1 providean example of a system that is controllable but not reachable.ra>l). we have the notion of reconstructibility: When does knowledge of {u jy }"~Q and {.~ I). Again from Theorem 6. We might now ask the question: Given XQ = 0. equivalently. A n . we see that (6. The vector Jt* in linear system theory is e IR nx " fieR" (n ~ I. We might now ask the question: Given Xo 0.10) for k ~ 1. m known as the state vector at time while Uk is the input (control) vector. B T] is observable [reconsrrucrible] [controllablcl if and T) observable [reconstructive].11) with and D (p ~ 1). A n . does there exist an input sequence {ujj 1jj^ such that Xk takes an arbitrary value in 1R"? In linear system theory. (A. we of reachability. B) iJ reachable [controllable] ifand only if (A . A related question is the following: Given an arbitrary initial vector Xo. AB. does there exist an input sequence {u j an input sequence {"y}"~o such that xn = O? In linear system theory.2.9) ~Axo+[B. controlA 1 lability and reachability are equivalent. this is a question va [Uj }k~:b such that x^ takes an arbitrary value in W ? In linear system theory.10. see that (6..8) is reachable if and only if if R([ B. Example 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. The matrices A = [ ° Q and f ^ provide an lability and reachability are equivalent.e.e.46 46 Equations Chapter 6. reachability always implies controllability and. There are many other algebraically equivalent conditions.T. from the fundamental Existence Theorem. if A is nonsingular. AB.. The answers are cast in terms that are dual in the linear algebra sense as well. We can then pose some new questions about the overall system that are dual in the system-theoretic sense to reachability and controllability.... we have the notion of suffice to determine (uniquely) xo? As a dual to controllability.AB •. Linear Equations Xk with A E R"xn and B E IR nxmxm(rc>l. We now introduce an output vector yk to the system (6.. linear differential equations). The above are standard conditions with analogues for continuous-time models (i. . Theorem 6. Theorem l'/:b Clearly.9 Example 6. We now introduce an output vector Yk to the system (6. There are many other algebraically equivalent conditions. The condition The answers are cast in terms that are dual in the linear algebra sense as well.8) is controllable if and only if if controllability. does there exA related question is the following: Given an arbitrary initial vector XQ.8) is given by k-J Xk = Akxo + LAk-J-j BUj j=O Uk-J ] Uk-2 (6. standard conditions with analogues for continuous-time models (i.. overall system that are dual in the system-theoretic sense to reachability and controllability. The general known as the state vector at time k while Uk is the input (control) vector. from the fundamental Existence Theorem.2. B) is if(AT .. The condition dual to reachability is called observability: When does knowledge of {u 7 r/:b and {Yj l'.J B] = n. We can then pose some new questions about the with C and (p > 1). The linear differential equations). if and only if or. this is a question {u }y~Q Xk of reacbability.: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. Since m ~ I. . does there exist an input sequence for k > 1. this is called such that Xn = 0? linear system theory. Since > 1. this is called controllability.8) of Example 6. if and only if rank [B. . The general solution of (6.9 by appending the equation by appending the equation (6.• A k k-J B] [ ~o (6. example of a system that is controllable but not reachable.J B]) = 1R" or. equivalently.

in which case the general solution is of the form (6.13) Let v denote the (known) vector on the left-hand side of (6. particularly for block matrices.14) Theorem 6. Theorem 6. the coefficient matrices of interest are square and nonsingular. C E jRmxn. e Rmxn. particularly for block matrices. 0.13) and let R denote the matrix on Let denote the (known) vector on the left-hand side of (6. A E Rnxn. and C e jRpxq. (6.15) E jRnxp where Y € Rn*p is arbitrary. Verification of each identity is recommended as an exercise for the reader. +L k-l CAk-1-j BUj + DUk. Then the equation e jRmxn. Theorem 6. Such a criterion (CC+ <g) A+ A — I) is stated and proved in Theorem 13.4 Some Useful and Interesting Inverses 6.4 6.27. Verification of each identity is recommended as an exercise for the reader. and C E Rpxti.12) j=O Yo . By the fundamental Uniqueness Theorem.4 Some Useful and Interesting Inverses Some Useful and Interesting Inverses In many applications. equivalently. in which case the general solution is of the has a solution if and only if AA + BC+C = B. e Tl(R). 6. the solution is then unique if and only if N(R) ==0. Then. or. so a solution exists. v E R(R).3 6. indicated.11. Such a criterion (C C+ ® A +A = I) of the Kronecker product of matrices for its statement.6. Let A E Rmxn. Listed below is a small collection of useful matrix identities. . the coefficient matrices of interest are square and nonsingular.27. notice that To derive a condition for observability. B E Rmxq. B e jRmx q . Listed In many applications.DUn-l 6. is stated and proved in Theorem 13. A compact matrix criterion for uniqueness of solutions to (6. by definition.Duo Yl . sociated e jRnxn. by definition. B E Rnxm.CBuo . the solution is then unique if and only if N(R) Uniqueness Theorem. notice that Yk = CAkxo Thus. E jRnxm.14) requires the notion A compact matrix criterion for uniqueness of solutions to (6. mxm and D E jRm Invertibility is assumed for any component or subblock whose inverse is and D € E xm.Du] (6. equivalently.4 Some Useful and Interesting Inverses 47 To derive a condition for observability.2 -j BUj .6. Invertibility is assumed for any component or subblock whose inverse is indicated. Thus. Then.13) and let denote the matrix on the right-hand side.6. By the fundamental the right-hand side. asbelow is a small collection of useful matrix identities. arbitrary. associated with matrix inverses. In these identities.3 A More General Matrix Linear Equation A More General Matrix Linear Equation AXC=B (6. if and only if or. so a solution exists.. the has a solution if and only if AA+BC+C = B.14) requires the notion of the Kronecker product of matrices for its statement. if and only if r Yn-] - Lj:~ CA n .

I .4. l = l = [!C / [~ ~ l = [ A-I +_~~!~CA-I -A~BE = D. l 8. [~ ~ r l 3.48 Chapter 6. for example.8. r A~I [~ ~ r [-D~I~A-I D~I 1 ~r ~~B 1 r l [~ ~ r [-D~CF +-~~I~.. characterize all solutions of the matrix linear equation 7Z(B) c 7£(A). 2. where F = (A .4. 5. It has many applications (and is frequently "rediscovered") including.1.. = = 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 X-I = X. Let A E lRmxn. where E = (D . for example. mx .8.1. BC 6. X. Assuming 2. Linear Equations Chapter 6.BD-I l = [ -A-I BD.CA. (A + BDC)-I = A~l . (A BDCr1 = A-I . ization in property 17 of Section 1. theory.I ] D. Linear Equations 1. As in Example 6.I C) -I. This result follows easily from the block LU factorization in property 16 of Section 1. 1.I EXERCISES EXERCISES 1. It has many This result is known as the Sherman-Morrison-Woodbury formula. 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. 2. It also the inverse of a sum of matrices such as (A + D)"1 or (A" + D"1) It also yields very efficient "updating" or "downdating" formulas in expressions such as yields very efficient "updating" or "downdating" formulas in expressions such as T (A + JUT ) -I1 (with symmetric A E R"x" and . BB EelR fflxk and suppose AAhas an SVD as in Theorem 5. Let A € E mx ".4. As in Example 6. This where E = (D — CA B) (E is the inverse of the Schur complement of A). [ / +c 7.B D. formulas for applications (and is frequently "rediscovered") including. Rmxk and suppose has an SVD as in Theorem 5. Note that the positions of the / and ./ blocks may be exchanged. characterize all solutions of the matrix linear equation AX=B in terms of the SVD of A in terms of the SVD of A. Assuming R(B) ~ R(A). characterize all left inverses of a matrix A e lR ". Note that the positions of the / and — / blocks may be exchanged. 1.. formulas for the inverse of a sum of matrices such as (A + D)-lor (A-I1 + D-I)-I.I B)-I (E is the inverse of the Schur complement of A).A-IB(D-lI + CA-IB)-ICA-I. characterize all left inverses of a matrix A E Mm xn . that X~l [~ !/ [~ ~ r [~ -~ l [~ ~/ r [~ -~ 1 l l l = [ ~ 4. result follows easily from the block LU factorization in property 16 of Section 1.c E E") that arise in optimization (A + xx T ) — (with symmetric A e lRnxn and x e lRn) that arise in optimization theory.A~lB(D~ CA~lB)~[CA~l This result is known as the Sherman-Morrison-Woodbury formula.4.

. Show that the matrix B — A — —eie T : (i.Cn and individual elements Yij. where C = 1/(1 . Show that 4.10.. Let x. Show that -cxJ C ' where c 1/(1 — T y). .. T 4. y e IRn and suppose further that x T y ^ 1. As in Example 6. 6. c and individual elements y. Let x. € IRn and suppose that x T y i= 1. Let jc.e. A with yl subtracted from its (ij)th element) is singular.. y E E" and suppose further that XTy ^ 1. Let A E 1R~ " and let A -1 have columns Cl. check directly condition for reconstructibility the form form N[ fA J CA n 1 ~ N(A n ). . . y E E" and suppose further that XTy i= 1. in Example 6... Show that the matrix B = A .x xTy). Show that 3.10. A with — subtracted from its (zy)th element) is singular.xy) T -1 49 = I - 1 xTy -1 xy . l' Hint: Show that Ci E N(B). Assume that x/( 7^ 0 for some and j. (i.l ~i e.y Assume that Yji i= 0 for some i/ and j. . Hint: Show that ct <= M(B). 5.e.. Let A e R"xxn and let A"1 have columns c\.e. Show that (/ . check directly that the condition for reconstructibility takes the 6.Exercises Exercises 3..

This page intentionally left blank This page intentionally left blank .

x = I -px.y. Let V be a vector space with V = X EEl Y. Proof: Suppose P is a projection. P2 = P. Let V be a vector space with V X 0 y.1). every v E V Definition 7. Proof: Suppose P is a projection. every v e V has a unique decomposition v x y with x E and y e y.1 displays the projection of von both and 3^ in the case = Figure 7. P is a projection if and only if I —P is a projection.y is linear and P# y — px. Px.yV = x for all v E V..y is linear and pl. Infact.y is called the (oblique) projection on X along y. say on X along y (using the notation of Definition 7. i.1 displays the projection of v on both X and Y in the case V = ]R2.2.Chapter 7 Chapter 7 Projections. and Norms 7. A linear transformation P is a projection if and only if it is idempotent.e.. Py. Theorem 7.e. Oblique projections.1.x — I — Px. y = Px. V by by PX. Inner Product Spaces.1 Projections Definition 7.y is called the (oblique) projection on X along 3^.26. i. Define pX.1 7.y- Theorem 7. px.yp2 = P. Define PX y : V ---+ X <. Theorem 7.1).26. P isaprojectionifandonlyifl -P isaprojection. say on X along Y (using the notation of Definition 7. y x Figure 7.1.1. Px.y. By Theorem 2. Theorem 7. Py. and Norms Spaces. Also.1.3. Also. Inner Product Projections.3. Figure 7.2. 51 51 . Infact. A linear transformation P is a projection if and only if it is idempotent. PX. By Theorem 2. Figure 7. Oblique projections.y • V —>• c V has a unique decomposition v = x + y with x e X and y E y.

suppose symmetric projection matrix and let x be arbitrary.P)x e X1-.P)x = (I . along 1-) and let jc. Then Px = p 2v = Pv = x so x e X. then (/ .X^X = Px±.P)x = O. Moreover.. let A E Rmxn with SVD A = U!:VTT = A = UT. arbitrary.P)x = O. then Pv = v. Projections.p 2 v 0 so y e Thus. PN(A)J. P E E"xn is the matrix of an orthogonal projection (onto K(P)} if and only 7.P)x 6 R(P)1and P must be an orthogonal projection. p2 = P.P) = 0. say.P)x = xTP(I . mental subspaces. V = X $ Y and the projection on X along Y is P. Conversely. Then symmetric projection matrix and let x be arbitrary.P)x. while Py = P(l -.P)x E ft(P)1 xTPT(I .P)x = yTTpT (I . yy Ee jR" be arbitrary. Now let v E V be arbitrary. Hence PT = PTP = P.P}x = 0.1 and 5.5. Thus. Conversely. Write x = Px (I — P)x. Projections. Px E R(P).1 The four fundamental orthogonal projections The four fundamental orthogonal projections Using the notation of Theorems 5. Let X n y.AA+ U2 U ! LUiUT. A 6 jRmxII UtSVf. Note that (I .P)v.A+A V2V{ L i=r+l i=l 11 ViVf. Then U\SVr Then r PR(A) AA+ U\U[ Lu. Then Px = P2v = Pv = x so x E X. then v = 0.. First note that v E X.P)x = x T P(l .uT. 0 Definition 7. It is easy to check that X and Y are subspaces. First note that iftfveX. If v E Y. In the special case where Y = X^.P is the projection on Y along X. . P2v = P Pv — 2 2 Px = x = Pv. p 2v = PPv = Let u E V be arbitrary.11.11. then Pv = O.. we have ( P y f I (/ . P e jRnxn is the matrix of an orthogonal projection (onto R(P)) if and only ifP2 PT if p2 = p = pT. Hence that V X 0 y. . Then v = Pv + (I . along XXL} and let x. Then v if v € Pv (I . Thus. say. y = (I . suppose P = P.xl. D 0 7. suppose P is a is a with the second equality following since pT P is symmetric. Inner Product Spaces. T Since x and y were arbitrary.L 1. X 0 y and the projection on X along y is P.5. while Py = P(I P}v = x Pv . Note that (/ .xx by Theorem 7. (I .=1 m PR(A).PX. we must have P (I — P) = O.P)v. We now prove and Y = {v E V : Pv = OJ. V Pv . Since Py E X.P)x E XL. with the second equality following since PTP is symmetric. P Proof: Let P be an orthogonal projection (on X. px.P)v.3.V Theorems 5. Py e X.1. Then Pv = P(x + y) = Px = x. y = (I . Thus. suppose p2 = P. Thus. We now prove that V = X $ y. i=r+l PN(A) 1.xJ. Moreover. we must have pT (I . Inner Product Spaces. PX. Thus.P)x = (I . we have (py)T ((I . In the special case where y X1-.1 7. then v = O.3.XL Theorem 7. * called an orthogonal projecDefinition 7.P)v. and Norms Let v e V be arbitrary. P = P. Hence pT = pT P = P. then Pv v. Now let u e V be arbitrary.P)v = = Pv. Conversely. Conversely. Hence if v E X ny.P2v = 0 so Y E y. A+A VIV{ r LViVT are easily checked to be (unique) orthogonal projections onto the respective four fundaare easily checked to be (unique) orthogonal projections onto the respective four fundamental subspaces.1 5. P)x = y PT(I P)x = 0. If v e y. Let x = Pv. D Essentially the same argument shows that I — P is the projection on y along X. since Px e U(P). Let X = {v e V : Pv = v} and y {v € V : Pv 0}.1 . Write x = P x + (I .52 52 Chapter 7.xx by Theorem 7. then Pv = 0.XLtion and we then use the notation P x = PX. and Norms Chapter 7.4. R" be Proof: Let P be an orthogonal projection (on X. (I . Then Pv = P(x + y) = Px = x. Then x T pT (I . Since x and y were arbitrary.px. Thus. It is easy to check that X and 3^ are subspaces. and P must be an orthogonal projection. Let X = {v E V : Pv = v} Px = x = Pv..)x = PXJ.4.XL iss called an orthogonal projection and we then use the notation PX = PX. Essentially the same argument shows that / .

Determine the orthogonal projection of a vector v E IR n on another nonzero vector w E IRn. 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.A+ A)x 2 = A+ Ax + (I = VI vt x + V Vi x (recall VVT = I).7.1. { v \ . A direct calculation shows z = Pn(w)"' = (l .11.2. Recall the proof of Theorem 3. IR n Rm 1 n Let X E IR be an arbitrary vector. T W W Moreover.7.~) w. Then the desired projection is simply Then the desired projection is simply Pn(w)v = ww+v wwTv (using Example 4.8) (using Example 4." Example 7. . Vk} was an orthomormal basis for a subset S of W1. The indicated direct sum decompositions of the domain E" and co-domain IRm are given easily as follows.. Orthogonal projection on a "line. See Figure 7. Projections 53 Example 7. The expression for x\ is simply the orthogonal projection of XI projection of rather x on S. Specifically.2. There. Projections 7...6.7. orthogonal: v z Pv w Figure 7. .. X on Specifically. See Figure 7. e Rn Solution: Think of the vector w as an element of the one-dimensional subspace IZ(w).(:..8) = (WTV) W." Figure 7. A direct calculation shows that and ware. Recall the diagram of the four fundamental subspaces.Pn(w»v = v . Then Let x e W be an arbitrary vector. There. are. Example 7. Solution: Think of the vector w as an element of the one-dimensional subspace R( w). in fact.. in fact. The indicated direct Example 7. Recall the proof of Theorem 3. Orthogonal projection on a "line.2. An arbitrary vector x E R" was chosen and a formula for XI basis for a subset of IRn. Vk} was an orthornormal Example 7. An arbitrary vector x e IRn was chosen and a formula for x\ appeared rather mysteriously.8.1.11. {VI.8. .6. . Recall the diagram of the four fundamental subspaces. Determine the orthogonal projection of a vector e M" on another nonzero Example 7. orthogonal: that z and u. the vector z that is orthogonal to w and such that Pv Moreover.2. Then X = PN(A)u + PN(A)X .

10. (x.x)forallx. x) for all x. Yl. only ifx = O. f3ftE IR. . y) for all x € Rm and for all y e R".y E V. aYI + PY2) = a(x. Let V = R". x } = 0 if and only ifx = 0. y) Q = X T Qy. V = IRn. (jc. respectively. {*.54 Chapter 7. respectively. yi. Then { • • V x V if product if 1. ATE IR nxm transformation Definition 7. Let V be a vector space over IR.13. 3. Let V = IRn.13. Then (x. y^} for all jc. such that {x. Projections.. x) ::: Qfor aU x 6V and (x. > Ofor all E V ( x x) =0 if 2. cryi + ^2) = a(x. Inner Product Spaces. . y} = XTy is the "usual" Euclidean inner product or dot product. Then Y = PR(A)Y + PR(A)~Y = AA+y + ( l . and Norms Chapter 7. where Q = QT > 0 is an arbitrary Q = Q T > is an Example 7.9.9.12. 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 . Then (x.11. Y2 E V and/or all a. Ay) = {AT x. (x.2 Inner Product Inner Product Spaces Definition 7. defines a "weighted" inner product. let Y E ]Rm be an arbitrary vector. Inner Product Spaces. Example 7. Then ('. y) = (y. Let Example 7. y e V. y)Q = XT Qy. and Norms Similarly. j2 ^ V and for alia. 2. Let V = E". 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). [ 5~2 + 7. y\) + /3(jt. n x n positive definite matrix. If A E Rm xn. Example 7. e R.(A . y) = (y.AA+)y = U1Ur y + U2U[ Y (recall UU T = I).) ) :: V x V -+ IR is a real inner is a real inner Definition 7. . Projections. Then Similarly. (x. (x.11. Y2) for all x. definite defines Definition 7.12.10. If e IR mx ". Example 7. y) x T Y is the "usual" Euclidean inner product or Example 7. 3. then AT e Rn xm is the unique linear transformation or map T E IRm andfor IRn. (x. Then {^. as follows: and a vector in J\f(A). let y e IR m be an arbitrary vector. Let V be a vector space over R. Yl) + f3(x.

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# = R-1A 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 Q-orthogonality (Q is a positive definite matrix). Two vectors x, y E IRn are Q-orthogonal (or conjugate with a positive definite matrix). Two vectors x, y e W are <2-orthogonal (or conjugate with T X Qy O. Q-orthogonality 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 complex-valued, the complex inner product is to be understood. if the vectors involved are complex-valued, 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 Q-orthogonality 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 p-norms, are defined by 1. For e C", the HOlder norms, or p-norms, 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-

(C-B-S) inequality (see. its determinant must be nonnegative. ttlr. i. t \ ] R).32. Then Theorem 7. then Remark 7. The angle e between two nonzero vectors x. Theorem 7. denoted II . [20. R).g. On the vector space (C[to. y e en. || . 217]). then H H \\Ux\\2 \\x\\2 (Proof. Ther. and Norms Chapter 7. -+-=1. 217]). define the vector norm On the vector space «e[to. we see immediately that \XH yl < IIxll2l1yllz. Since yHxx = x Hy. Let x. o ~ (x Hxx)(yH y) .||. e.. is a nonnegative definite matrix. y e C".31 (Cauchy-Bunyakovsky-Schwarz Inequality). whered.32.> where Q = QH > 0 (this norm is more commonly = QH > Ikllz. Theorem 7.. p. Theorem 7. Inner Product Spaces. The C-B-S inequality is thus equivalent to the statement I. In other words. Proof' Consider the matrix [x y] E C" x2 .1^ IIUxll2 = IIxll2 (Proof IIUxili = x U Ux = xHx = IIxlli)· However. tO~t:5. 0 < e — 5-. y E en.31 and Remark 7. Some weighted p-norms: (a) IIxll1. Inner Product Spaces. However.D = E^rf/l*/!. Since yH = x H y. define the vector norm 11111 = max 1/(t)I· to:::. D 0 \\X\\2\\y\\2Note: This is not the classical algebraic proof of the Cauchy-Bunyakovsky-Schwarz Note: This is not the classical algebraic proof of the Cauchy-Bunyakovsky-Schwarz (C-B-S) inequality (see. define the vector norm 3. define the vector norm 1111100 = max II/(t) 11 00 . Then Fhcorem 7. Remark 7.Q — (xhH Qx) 2. 112 is unitarily invariant. it is particularly easy to remember. ||JC||. Since is a nonnegative definite matrix. However. Projections.34. [20.31 (Cauchy-Bunyakovsky-Schwarz Inequality).33. Since Proof: Consider the matrix [x y] e en x2 . p.58 58 Chapter 7. where 4 > O. it is particularly easy to remember. (b) IIx IIz. and Norms 2. 1Ft).e.lx.31 and Remark 7. i. The angle 0 between two nonzero vectors x. we see immediately that IXH y\ ~ 0 < ( x H ) ( y H y ) — ( x H ) ( y H x ) ... 1Ft). A particular case of the Holder inequality is of special interest. The norm || • ||2 is unitarily invariant.34. t\])n. with equality if and only if x and y are linearly dependent. p q I I A particular case of the HOlder inequality is of special interest. Let x.D = L~=ld. Remark 7. Then with equality if and only if x and yare linearly dependent. On the vector space (C[to. Projections.g = (x QXY denoted || • ||c).30 (HOlder Inequality).(x Hyy)(yH x). its determinant must be nonnegative. ttl.. However.e.t~JI On the vector space ((C[to.32 are true for general inner product spaces. and || . Remark 7. > 0.~~1~1112'. e. In other words. Remark 7.g.33.l. y E C". Some weighted p-norms: 2. The norm II . 1cose 1|~ 1. if U € C"x" is unitary.32 are true for general inner product spaces. y E en may be defined by cos# = 1I. 11·111 and 1I·IIClO XHUHUx . \\Ux\\l XHX = \\x\\\).tl Theorem 7.30 (Holder Inequality). Let x. y e C" may be defined by Remark 7. Let x. „ |.^|| cos e = Il-Mmlylb 0 ~ 0 < I' The C-B-S inequality is thus equivalent to the statement ~ ^ | COS 0 < 1.. if U E enxn is unitary. IIQ)' 1 3.

Attention is confined to the vector space (IRmnxn. Similar remarks apply to the unitary invariance of norms of real vectors under orthogonal transformation. IIA + BII :::: IIAII + IIBII for all A.. 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. Then lim k4+00 V(k) = v if and only if lim k~+oo II v(k) - v II = O. and essentially obvious.4. ConFinally.. For x E en. For x G C". || || R mx " ~ E is a matrix norm if it satisfies the following three Definition 7. i. i. there exist constants CI. then we have the Pythagorean Identity Ilx ± YII~ = IIxll~ + IIYII~.. c-i (possibly depending onn) such that depending on n) such that Example 7. i.38.35. this is called the triangle inequality. Attention is confined to the vector space (W xn R) since that is what arises in the majority of applications.. E en.e. e C". the proof of which follows easily from liz II~ = ZH z.37.. convergence in terms of vector norms. Theorem 7.e. vectors under orthogonal transformation. i. BE IRmxn... Extension to the complex case is straightforward and essentially obvious. Remark 7.38. 3. The former notion is useful for perturbation analysis. Then 7.) .35. v(2). convergence in terms of vector norms.e. . 7. Theorem 7.. 2 the proof of which follows easily from ||z||2 _ z_//. then we have the Pythagorean Identity Remark 7. IIAII ~ Ofor all A E IR mxn and IR IIAII = 0 if and only if A = O. If x. the following inequalities are all tight bounds. there exist Example 7.4 Matrix Norms Matrix Norms In this section we introduce the concept of matrix norm. the motivation for In this section we introduce the concept of matrix norm.39.7. while the latter is needed to make sense of former notion is useful for perturbation analysis. Finally. about convergence of real numbers. there exist vectors x for which equality holds: vectors x for which equality holds: Ilxlll :::: Jn Ilxlb Ilxll2:::: IIxll» IIxlloo :::: IIxll» Ilxlll :::: n IIxlloo. there exist constants c\. i.36. All norms on en are equivalent. (As with vectors.. C2 (possibly 7.e.. As with vectors.e. we conclude this section with a theorem about convergence of vectors. As with vectors. lIaAl1 = lalliAliforall A E mxn andfor all a E IR. II·• II : IR mxn -> IR is a matrix norm if it satisfies the following three properties: properties: 1. Definition 7. while the latter is needed to make sense of "convergence" of matrices. Matrix Norms 59 59 are not unitarily invariant. The using matrix norms is to have a notion of either the size of or the nearness of matrices.4.. we conclude this section with a theorem about convergence of vectors.36. If y € C" are orthogonal. i» (1) v(2\ . Extension to the complex case is straightforward what arises in the majority of applications. i.e. IIxlioo :::: IIxllz. IR) since that is "convergence" of matrices.. Matrix Norms 7.4 7. v(l). IIxl12 :::: Jn Ilxll oo . All norms on C" are equivalent.39. the motivation for using matrix norms is to have a notion of either the size of or the nearness of matrices. Similar remarks apply to the unitary invariance of norms of real are not unitarily invariant. the following inequalities are all tight bounds. 2. y E en are orthogonal.37. Let \\ \\ be a vector norm and suppose v. Let II·• II be a vector norm and suppose v.

Let A E lR. Let A E K m x ". IIAII2 = Amax(A A) = A~ax(AA ) = a1(A).42. I. and Norms Chapter 7. tTL T Note: IIA+llz = l/ar(A).43." IIAliss = Li. e R mx ".41. (AA ')). (A' A)) 1 ~ (T. defined by IIAIIF ~ (t. 1.mxn IIAII p. The norm || ._ Then "mixed" norms can also be defined by e lR. Then the matrix p-norms are defined by A e Rmxn.44.40.1 is often called the trace norm. The "matrix analogue of the vector I-norm.p = (at' + . The "maximum row sum" norm is 2. Then the Frobenius norm (or matrix Euclidean norm) is 7. ^wncic = rank(A)). The spectral norm is 3. IIAII P t altA)) 1 ~ (T. \\F and 11'115. Inner Product Spaces. The "matrix analogue of the vector 1-norm. IIAlioo = max rE!!l.44. ||5>1 is often called the trace norm. The following three special cases are important because they are "computable.jj laij. Example 7. ai. Projections. I Some special cases of Schatten /?-norms are equal to norms defined previously.2 = || . Projections. . Example 7.60 max -_P IIAxll = max Ilxli p IIxllp=1 IIAxll p . (where r = laiiK^/i. Inner Product Spaces. 11·115.. (t laUI). and Norms Example 7. to estimate the size of a matrix product A B in terms of the sizes of A and B individually. I. is a norm.<110#0 IIxllq Example 7.43.mxn..00 = || • ||2." || A\\ = ^ \ai} |..) I ~ (t. J=1 3. where r mxn = rank(A). Let A E R . Let A E lR. + a!)"". Example 7. Example 7. Example 7.40. || 5 2 = II IIF and || • ||5i00 = II ." Each is a "computable. is a norm. p-norms previously.60 Chapter 7.mxn. 112' The norm II • 115. || . The "maximum column sum" norm is 2. For example." theorem and requires a proof.q = max IIAxil p 11.mxn. Schatten/7-norms IIAlls. 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. Example 7. The Schattenp-norms are defined by E lR. matrix = Ilxllp..42.

)). II such that ||A||F is given by max^o ". e. subordinate to the vector norm.. IIAIIF ::s . A A 2.e. Let A E ]Rmxn. l. We thus need the following definition. 11^4^11 P (or. it follows that all subordinate norms are consistent.jii IIAlloo. there exists a vector norm II • IIv Theorem 7. 2. e R" x ". For such subordinate norms. e. \\m is a consistent matrix norm. i. reader The interested reader is invited to prove each of them as an exercise. take A = B = [: is a matrix norm but it is not consistent. i. For example. Let A e Rmxn. II In II p = 1 for all p. Example 7. •II ||F. For example. . .g. there exist matrices A for i. \\v consistent with it. IIAlioo ::s .e. but there does consider II . while E ]Rnxn.45.e.46. atornorms. Theorem 7.oo 2 while IIAIII. || • ||/7and II ||.ooIlBIII..jii IIAIIF. IIAII2 ::s. not exist a vector norm II • || such that IIAIIF is given by max x . ||A|| = max^o IIxll. IIAIII ::s . If \\ • 11m is a consistent matrix norm. so not exist a vector norm || . Then the norms \\ . For example.so II ||.. also called oper(or. 1. it follows that all subordinate norms are consistent.q = maxx.45. 1. B e Rnxk.jii IIAlioo' .48.60 \^ • Useful Results The following miscellaneous results about matrix norms are collected for future reference. Matrix Norms 61 61 Notice that this difficulty did not arise for vectors. IIAII2::S IIAIIF. IIAxll1 = max . the IIIn II F = .= max IIAxl1 x. For example. Theorem 7.47. There exists a vector x* such that IIAx*11 = IIAllllx*11 if the matrix norm is subordinate to the vector norm. For A following inequalities are all tight. q For such subordmate norms..4. IIAllp. IIABIII.60 IIx II Ilxll=1 IIAxll p . \\ are Definition 7. exercise. IIAxl1 IIAII = max . consistent with it. although there are analogues for. Notice that this difficulty did not arise for vectors. i.. more generally. i. B E ]Rnxk. p for all p are consistent matrix norms. Not every consistent matrix norm is subordinate to a vector norm.47.Then II Ax 1122 ::s II ||A||F||jc||2.. consider || • \\F. which equality holds: which equality holds: IIAIII ::s .. We thus need the following definition.jii IIAlb IIAIIF ::s . II· II F and 1.jii. Then the norms II • \\a II· Ilfl' and .e.1100 = max laijl x.e. inner products or outer products of vectors.and II \\ •lIy y are mutually consistent if \\ A B \\ a < IIAllfllIBlly. IIAII2 ::s .60 Ilx i.but there does || is consistent with F. although there are analogues for.4..jii IIAII I.46. The "mixed" norm "mixed" norm II· 11 100 . Then :]. •II ||p for all p are consistent matrix norms.e.. IIAlioo ::s . The following miscellaneous results about matrix norms are collected for future reference.g. There exists a vector x* such that ||Ajt*|| = ||A|| ||jc*|| if the matrix norm is Theorem 7. \\ • \\p. IIAxliv < IIAlim \\x\\v' Not every consistent matrix norm is subordinate to a vector norm. Since IIABxl1 ::s ||A||||fljc|| ::s IIAIIIIBllllxll.::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.. II".~~i'.j is a matrix norm but it is not consistent. we clearly have ||Ajc|| ::s ||A||1|jt||. Then The p -norms are examples of matrix norms that are subordinate to (or induced by) The p-norms are examples of matrix norms that are subordinate to (or induced by) a vector norm. HAjcJI^ ::s \\A\\m Ilxli v. II A 1100 ::s n IIAII I . take A = B = \ \ ||Afl|| li00 = 2while||A|| li00 ||B|| 1>00 = 1. also caUedoperator norms.7. wec1earlyhave IIAxll < IIAllllxll· Since ||Afijc|| < IIAlIllBxll < ||A||||fl||||jt||. IIAIIF ::s. Example 7.60 .48. •1122is consistent with II ||. A matrix norm 11·11\\is said to be consistent mutuallyconsistentifIlABII. more generally.. Definition 7. i. there exists a vector norm \\ . II F' Then||A^|| < A II Filx 112. if II A B II < II A 1111 fi|| whenever the matrix product is defined. Matrix Norms 7..jii IIAlb IIAIII ::s n IIAlloo.oo J1.jii II A IIF. inner products or outer products of vectors.jii IIAII2.jii IIAII I . 2.. If II . 2. a vector norm.

Suppose P and Q are orthogonal projections and P + Q = I. Projections. A (2) . B) = space.y + 2z = O. Then 7. A(I). A(A A) -1 AT 5.A+A is an orthogonal projection. p+ = P. A(2). „ } The spectral radius of A is the scalar p(A) = max IA. . . 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. \\ -\\bea Rmx". Inner Product Spaces.] l. . 1. IIF are unitarily invariant. for all A E IRmxn and for all orthogonal unitarily invariant. .An}. If P is an orthogonal projection.. B) = Tr ATB is a real inner product IR n x" AT B (A. > . Find the (orthogonal) projection of the vector [2 3 4f onto the subspace of 1R3 5.l. Theorem 7. 112 and II . Definition: Let A e Rnxn and denote its set of eigenvalues (not necessarily distinct) 8. space.49. 3. EeIRmxn. spanned by the plane 3x ..62 62 3. where ¥2 is defined as in Theorem 5.] 4. The spectral radius of A is the scalar by {A-i .Q — Q must be an orthogonal matrix. Show that the matrix norms II . EXERCISES EXERCISES 1. Then k~+oo lim A (k) = A if and only if k~+oo lim IIA (k) - A II = o. 2... but not necessarily The norms || • \\F and || • ||2 (as well as all the Schatten p-norms.. 112 (as well as all the Schatten /?-norms.1. . Prove that P . A (1) . i. must be an orthogonal matrix.I.. Suppose that a matrix A E IR mxn has linearly independent columns. Chapter 7. 6.e.. 4. [2 3 4]r R3 spanned by the plane 3.49.. || • ||2 and || • \\F 8. e Rmx" mxm x mxm and Z E IRnxn .. i .. If P projection. Definition: Let A E IRnxn and denote its set of eigenvalues (not necessarily distinct) by P. The norms II . IIF and II . For A E IR mxn . prove that P+ = P. orthogonal projection. scalars.c — v + = 0. and Norms max laijl :::: IIAII2 :::: ~ max laijl. prove directly that V22Vl is an I — +A V V/ is an orthogonal projection. where V2 is defined as in Theorem 5. 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. Projections. Prove that / ... For A eRmxa . matrices Q zR and e M" ". 3. 7.e. Prove that E"xn with the inner product (A. ||(MZ||a or F. but not necessarily other p-norms) are unitarily invariant. IIQAZlia = ||A|| fora = 2 or F.. Also. i. and Norms Chapter 7. Inner Product Spaces.1. l. Let II ·11 be a matrix norm and suppose A.

H A I I ||A||2. y E IR n are nonzero.Exercises Exercises 63 63 Let Let A=[~ 14 0 12 5 ~]. 2. is called a "magic square" matrix. Let A=[~4 9 2 ~ ~]. all of whose columns and rows as well as main diagonal and antidiagonal sum to s = n(n2 + 1) /2. . 10. ||A||2. or (Xl as appropriate.) that || M Up = for all/?.. y e R" are nonzero. and p(A). and p(A). it can be proved that IIMllp = ss for all p..2. If M is a magic square matrix. Determine ||A||F. and peA). \\A\\ ||A||2. H A H ^ and peA). HA^. all of whose Determine ||A||F. Let A = xy . or oo as and II A 1100 in terms of IIxlla and/or IlylljJ. (An n x n matrix. Determine IIAIIF' IIAII Ilt. ||A||j. where ex and {3 take the value 1. Let 9.) T 10. IIAlb IIAlloo. Determine ||A||F. Determine IIAIIF' IIAIII> IIAlb and ||A||oo in terms of \\x\\a and/or \\y\\p. where a and ft take the value 1. where both x. Determine IIAIIF' IIAII d . appropriate. Let A = xyT. IIAlb IIAlloo. (An n x n matrix. columns and rows as well as main diagonal and antidiagonal sum to s = n (n 2 l)/2. where both x. 9.

This page intentionally left blank This page intentionally left blank .

(8.bll~ (and hence p(x) = \\Ax .2. see Section 8.e. ||A. where r = b . is a solution of the normal equations. A vector x E X if and only if x is of the form 2.PR(A)b) = (I . Hence.b 112) assumes its minimum value if and only if II Ax —b\\2) assumes its minimum value if and only if (8. The linear least squares problem consists of finding an element of the set squares problem consists of finding an element of the set x = {x E jRn : p(x) = IIAx .x — b\\\ (and hence p ( x ) = from the Pythagorean identity (Remark 7.35).Ax is the residual associated 1. IIAx .PR(A)bll~ + IIPR(A)b - Axll~ from the Pythagorean identity (Remark 7.Axll~ = lib . write the residual in the form r = (b . x e X if and only if latter form is commonly known as the normal equations.2.1 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. Solution: The set X has a number of easily verified properties: The set X has a number of easily verified properties: 1. while (b . i. The linear least Problem: Suppose A e jRmxn with > n and b <= Rm is given vector. while Now.e. i. A vector x X if and onlv if x is of the x=A+b+(I-A+A)y.bll 2 is minimized}. vector x e X if and only if AT where b — Ax is the residual associated with x.Ax) is clearly in 'R(A).. Now. A vector x E X if and only if ATrr = 0. IIrll~ = lib . so these two vectors are orthogonal..1) To see why this must be so. For further details.35).PR(A)b) + (PR(A)b - Ax). For further details. (PR(A)b . see Section 8. Hence. x E X if and only if x is a solution of the normal equations. The equations ATrr = 0 can be rewritten in the form A TAx = ATb and the x.-b E 'R(A)-L so these two vectors are orthogonal. AT — A T Ax = AT b latter form is commonly known as the normal equations.PR(A))b = PR(A). Thus. 2. Thus. whereyEjRnisarbitrary. (Pn(A)b — AJC) is clearly in 7£(A).1 8.. write the residual r in the form To see why this must be so.Chapter 8 Chapter 8 Linear Least Squares Linear Least Squares Problems Problems 8.2) 65 .

if and only if rank (A) = n.3..23. X is convex.A+A)(Oy (1 . X has a unique element x* of minimal 2-norm.. The unique solution of minimum 2-norm or F-norm is X = A+B. + (1 . X is convex.A+ A)z in X.66 Chapter 8. we can generalize the linear least squares Just as for the solution of linear equations. The unique solution of minimum 2-norm or F-norm is where Y € ]R. 1]. consider two arbitrary vectors jci = A+b + (I — A + A) y (I .1. AA+)bI1 2 the last inequality following by Theorem 7. This follows immediately from convexity or directly from the fact that all x e X are of the form (8. X = {x*} = {A+b}.e.n is arbitrary.A+A)y and *2 = A+b + (I — A+A)z in X. The only difference is that in the case of linear least squares solutions. Let A E E mx " and B € Rmxk. The minimum value of p ((x) is then clearly equal to where y E ]R. x" = + b is the unique vector 4. X. i. Linear Least Squares Problems and this equation always has a solution since AA+b e 7£(A). To see why. i. Then the convex combination and Xz = A+b (I . then equality holds and the least squares If the existence condition happens to be satisfied. and only if A+A = I or.2) are of the form x = A+ AA+b + (I - A+ A)y =A+b+(I-A+A)y.e. 5. if 5. x* minimizes the residual p ( x ) and is the vector of minimum 2-norm that does so. problem to the matrix case. In fact.. The minimum value of p x ) is then clearly equal to lib . There is a unique solution to the least squares problem. 7£(A). where Y E R" xfc is arbitrary. X = {x"} = {A+b}.mxn XElR Plxk min IIAX - Bib is of the form is of the form X=A+B+(I-A+A)Y. 0*i (1 #)* = A+b (I .1) and which follows since the two vectors are orthogonal. BE ]R. has a unique element x" of minimal2-norm. of linear least squares solutions. where y e W is arbitrary. Let 8 e [0.1) and convexity or directly from the fact that all x E X are of the form (8. which follows since the two vectors are orthogonal. The general solution to e ]R. if and only if A + A lor. Then the convex combination 8x.2. By Theorem 6. x* = A+b is the unique vector that solves this "double minimization" problem.. then equality holds and the least squares . X = A+B. equivalently. Remark 8. all and this equation always has a solution since AA+b E R(A). i.8)xz2 = A+b ++ (I -A+ A)(8y ++ (1 8)z) is clearly in X.2.23.2) are of the form solutions of (8. if and only if rank(A) n. If the existence condition happens to be satisfied..3. Just as for the solution of linear equations. Notice that solutions of the linear least squares problem look exactly the same as solutions of the linear system AX = B. there is no "existence condition" such as R(B) S. In fact.nxk is arbitrary. there is no "existence condition" such as K(B) c R(A). 3. Notice that solutions of the linear least squares problem look exactly the Remark 8. equivalently. x* minimizes the residual p(x) that solves this "double minimization" problem. Let 6 E [0. To see why. This follows immediately from and is the vector of minimum 2-norm that does so. the last inequality following by Theorem 7. The only difference is that in the case same as solutions of the linear system AX = B. The Theorem 8.0)z) is clearly in 4.. we can generalize the linear least squares problem to the matrix case. Linear Least Squares Problems Chapter 8.1].e. By Theorem 6.e. consider two arbitrary vectors Xl = A + b 3. i.mxk.PR(A)bll z = ~ 11(1 Ilbll z. all solutions of (8. There is a unique solution to the least squares problem.

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 2-norm or F -norm. F-norm. 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 Moore-Penrose pseudoinverse of A is the best (in the matrix 2-norm 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 2-norm (and F-norm). 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(b-Ax) = 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 2-norms, one could also use 1-norms or oo-norms. The latter two are computationally of 2-norms, one could also use I-norms or oo-norms. 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 2-norm 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

c. Then c.can be arbitrarily nonlinear. Then GI defining y = logy. [11]. the subvectors can have different lengths). C2 problem. Better numerical methods are based on algorithms that behavior in practice (and it does).b\\^ is II czll ~. (8. c.SVr U~VT Theorem 5.1. As far as the minimization is concerned. insight. then taking logarithms yields the equation logy = logci + cjt.70 70 Chapter 8. the subvectors can have different lengths).4 Least Squares and Singular Value Decomposition Least Squares and Singular Value Decomposition In the numerical linear algebra literature (e. Numerically better approaches ill difficulties n. piecewise polynomial functions. VT = U.c=UTb = II [~ ~] [ ~~ ] . In this section we investigate solution of the linear least squares problem min II Ax x b11 2 . Since the standard Kalman filter essentially amounts to sequential updating of normal equations. We now note that IIAx - bll~ = IIU~VT x = - bll~ II ~ VT X - U T bll.5) is that the coefficients Ci appear linearly. since II . if the fitting function is of the form y t) Y = ff( (t) = c\eC2i. Specifically.6) via the SVD. Sometimes a problem in which the Ci'S appear nonlinearly nonlinearly can be converted into a linear problem. it is shown that solution [4]. This explains why it is convenient to work above with the square of the norm rather than the concerned. Specifically. the two are equivalent. of linear least squares problems via the normal equations can be a very poor numerical method in finite-precision arithmetic. fact. e C2 / then taking logarithms yields the equation log y = log c. quantity above is clearly minimized by taking z\ = S-'c. c\ logci. 8. . In fact.. [7]. we assume that A has an SVD given by A U\SVf via the SVD. respectively. the last equivalent. The basis functions coefficients c.. appear functions </>. as in Theorem 5. For example. S~lc\.1. Better numerical methods are based on algorithms that AT work directly and solely on A itself rather than AT A.g. then ||u||^ = ||i>i \\\ \\vi\\\ (note The last equality follows from the fact that if v = [£ ]. A E IRmxn . etc. Ib is unitarily invariant =11~z-cll~ wherez=VTx. [7]. 's ¢i. For example. respectively. The key feature in (8. Linear Least Squares Problems difficulties because of numerical ill conditioning for large n.4 8. Z2 while the minimum value of \\Ax — b II ~ is l^llr while the minimum value of II Ax . are based on orthogonal polynomials. if the fitting function is of the form can be converted into a linear problem. The former is much more expensive but is generally more reliable and offers considerable theoretical offers insight. norm. + c2f. ] II: = II [ The last equality follows from the fact that if v [~~]. Two basic classes of algorithms are A itself S VD and QR (orthogonal-upper triangular) factorization. then II v II ~ = II viii ~ + II v211 ~ (note that orthogonality is not what is used here. z. [4]. The former based on SVD and QR (orthogonal-upper triangular) factorization. it can be expected to exhibit such poor numerical behavior in practice (and it does). problem. splines. [23]). we assume that A has an SVD given by A = UT.[ ~~ ] II: sz~~ c. bE IR m . The subvector z2 is arbitrary. arbitrary. This that orthogonality is not what is used here. Since the standard Kalman filter essentially amounts method in finite-precision arithmetic.. etc. = log c" and C2 = cj_ results in a standard linear least squares y — log y. Linear Least Squares Problems Chapter 8.

to reduce A in the following way. A finite sequence of simple orthogonal transformations. Z VIS-ici The last equality follows from The last equality follows from c = UTb = [ ~ f: ]= [ ~~ l Note that since Z2 is arbitrary. is orthogonal to all vectors in 7l(A}L b E R(A). A E ffi. where y E Rm is arbitrary. 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 vls-Iufb + V2 2.1). This follows easily since ||(7 . x has been written in the form x = A+b + (I . V2z is an arbitrary vector in R(V2 = N(A). = +b + (/ — A + A) y. 11(1.. we add the simplifying assumption that A has full column To simplify the exposition. Thus. than an SVD and. i.e. an important special case of the linear least squares problem is the Finally. where y e ffi.AA+)b\\22 = \\U2Ufb\\l = bTU2U^U22V!b = bTU2U*b = \\U?b\\22. A e R™ X M . x has Note that since 12 is arbitrary. 8. This agrees. Least Squares and QR Factorization B. i. i. Thus. In this case the SVD of A is given by A A = V:EVTT = [VI{ Vzl[g]Vr. It is then possible.1). via a sequence of so-called Householder or Givens rank. A e 1R™ X ". This agrees. Another expression for the minimum residual is || (/ — AA + )b|| 2 . with appropriate numerical enhancements.S. we again look at the solution of the linear least squares problem (8.A + A)_y.11U2U!b"~ = bTUZV!V UJb = bTVZV!b = IIV!bll~. with appropriate numerical enhancements. we add the simplifying assumption that A has full column rank. If we label the product of such orthogonal row transformations as the orthogonal matrix QT E R m x m .. we have QT € ffi. V2 Z 2 is an arbitrary vector in 7Z(V2)) = A/"(A). of course. an important special case of the linear least squares problem is the so-called full-rank problem.5.6) but this In this section. 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. and there is thus "no V2 part" to the solution. A E ffi.7) .5 8. we again look at the solution of the linear least squares problem (8. with (8.mxm.. via a sequence of so-called Householder or Givens transformations. In this case the SVD of A is given by so-called full-rank problem.6) but this time in terms of the QR factorization. i. This matrix factorization is much cheaper to compute time in terms of the QR factorization. of course.e. It is then possible.AA+)bll~ .5 Least Squares and QR Factorization Least Squares and QR Factorization In this section. 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.e. This follows easily since Another expression for the minimum residual is II (I .~xn. (8. This matrix factorization is much cheaper to compute than an SVD and. can be quite reliable.e.~xn.8. UZV = [U t/2][o]^i r > and there is thus "no V2 part" to the solution. Finally. with (8. can be quite reliable. If we label the product of such orthogonal row transformations as the to triangular form.m is arbitrary.. To simplify the exposition.AA+)bllz. to reduce A in the following way.

n . Consider the following set of measurements (*. (a) Find the best (in the 2-norm sense) line of the form y = ax + fJ that fits this (a) Find the best (in the 2-norm sense) line of the form y = ax + ft that fits this data. (b) Let r denote the "error vector" b . both ql and q2 . (8.7).I of the columns of yields the orthonormal columns of Q\. i. Multiplying through by Q in (8. Suppose q. and qz are two orthonormal vectors and b is a fixed vector. Multiplying through by Q Q2 E ffi.72 Chapter 8..e. Equivalently. sense).[ ~~ ] If:. by writing (8. Note that (8.A+A)y and A +b 1. xn. Both Q I and Qz2 have orthonormal columns. Consider the following set of measurements (Xi. (3. Linear Least Squares Problems Chapter 8.flq2 Show that r is orthogonal to (b) Let r denote the "error vector" b — ctq\ — {3qz. 3. Now write Q = [Q\ Qz].8). 112 is unitarily invariant ~ ] x . yt): (1.7).. we see that A=Q[~J = [QI = QIR. data. or (8. are orthogonal vectors. n • (a) Find the optimal linear combination aq^ + (3q2 that is closest to b (in the 2-norm (a) Find the optimallinear combination aql + fiq2 that is closest to b (in the 2-norm sense). For € ffi.Cl and the minimum residual The last quantity above is clearly minimized by taking x = R lIc\ and the minimum residual is Ilczllz. we have = R~l Qf b = +b and the minimum residual is IIC?^!^- EXERCISES EXERCISES 1. The last quantity above is clearly minimized by taking x = R.+ A)y and A+b are orthogonal vectors. Yi): 2. data. or (8. all in R".7).1Q\b = A+b and the minimum residual is II Qr bllz' is \\C2\\2.9) are variously referred to as QR factorizations of A. i. and any y E R". check directly that (I .8).7). b e ffi.. Note that Any of (8. 3.Equivalently. (2. Now write Q = [QI Q2].3).2).~xn is upper triangular. Suppose qi and q2 are two orthonormal vectors and b is a fixed vector. we have x = R. R~l) ) of the columns of A yields the orthonormal columns of QI. (8.aql . m and any e ffi.9) is essentially what is accomplished by the Gram-Schmidt process.e. Both Q\ and <2 have orthonormal columns. where QI E ffi. all in ffi.. we see that in (8. 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.9) is essentially what is accomplished by the Gram-Schmidt process.8) ~ ] (8. check directly that (I .1). (b) Find the best (in the 2-norm sense) line of the form jc = ay + (3 that fits this (b) Find the best (in the 2-norm sense) line of the form x = ay fJ that fits this data.9) Any of (8. Linear Least Squares Problems where E ffi.. For A E Wmxn . Qz] [ (8. 2.Show that r is orthogonal to both^i and q2.9) are variously referred to as QR factorizations of A.m IX(m ~" ) .. Now note that Now note that IIAx - bll~ = IIQ T Ax = II [ QTbll~ since II . b E Em. where Q\ e R mx " and Qz € K" x(m-n).mxn and where R e M£ x " is upper triangular.

where Q is orthogonal.xn can be factored in the form (8. Prove that A+ = R+ QT. Consider the problem of finding the minimum 2-norm solution of the linear least 5.bl1 2 when A = [~ ~ ] and b = [ !1 x The solution is (a) Consider a perturbation EI = [~ pi of A. and suppose A where is 1. Let A e R"x". Find all solutions of the linear least squares problem 4. What happens to IIx* — y ||2 as 8 approaches 0? where AI = A + E\. Let A E ~nxn. Solve the perturbed version of the above problem. of 2-norm solution of least «rmarp« problem squares nrr»h1<=>m min II Ax .:. not necessarily nonsingular. where again 8 is a small 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 version of the above problem.z||2 as 8 approaches O? where A2 — A E 2 What happens to \\x* — zll2 as 8 approaches 0? 6. where 8 is a small positive number. then A+ R~ Q\. and suppose A = QR. Solve the perturbed problem positive number. where 8 is a small positive number.• What happens to IIx* . What happens to ||jt* . Find all solutions of the linear least squares problem min II Ax .IlQf. Solve the perturbed problem min II A 2 z .9). not necessarily nonsingular.9).bib z n where A2 = A + E2. (a) Consider a perturbation E\ = [0 ~] of A.yII2 as 8 approaches O? (b) Now consider the perturbation E2 = [~ (b) Now consider the perturbation EI = \0 s~\ of A. then A+ == R. Use the four Penrose conditions and the fact that Q\ has orthonormal columns to 6. 7. where AI = A + E I . where again 8 is a small positive number. verify that if A E ~.bll 2 x when A = [ ~ 5. A+ R+QT .Exercises Exercises 73 4.

This page intentionally left blank This page intentionally left blank .

For any A E e nxn .1) Similarly.3 (Cayley-Hamilton).4.1). (Note that the characteristic polynomial can also be defined as det(A. [3]). Example 9. such that a scalar A E e. Note that if x [y] is a right [left] eigenvector of A. verify that n(A) = A2 2A .3 (Cayley-Hamilton). One often-used scaling for an eigenvector is One often-used 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 nonzero vector y E en is a left eigenvector corresponding to an eigenvalue Similarly.4. Thus. the Fundamental Theorem of Algebra says that 7t (X) is a polynomial of degree n. [21D or directly using elementary properties of inverses and determinants (see. for example. for proved easily from the Jordan canonical fonn to be discussed in the text to follow (see. n(A) = O. Let A = [~g ~g]. (Note that the characteristic polynomial can also be defined as det(Al . Theorem 9. Note that if x [y] is a right [left] eigenvector of A. It is an easy exercise to Example 9. Theorem 9. we use both forms results in at most a change of sign and.1 Fundamental Definitions and Properties Fundamental Definitions and Properties Definition 9. This results in at most a change of sign and. A nonzero vector x E en is a right eigenvector of A E e nxn if there exists Definition 9.1 9. as a matter of convenience. It is an easy exercise to 2 verify that n(A) = A + 2A . the Fundamental Theorem of Algebra says that x 75 . [21]) or directly using elementary properties of inverses and determinants (see. This of A. norm used for such scaling. Let A [-~ -~].2) By taking Hennitian transposes in (9. For any A e Cnxn . as a matter of convenience. It can be proved easily from the Jordan canonical form to be discussed in the text to follow (see.2.31 = 0. called an eigenvalue. then It can be proved from elementary properties of detenninants that if A e enxn . then vector of AH associated with I. The 2-nonn is the most common a — \j'||. (9.1. such that Ax = AX. The polynomialn (A. we use both forms throughout the text.1. for example. A nonzero vector x e C" is a right eigenvector of A e Cnxn if there exists a scalar A. example./ — A).31 O. The 2-norm is the most common nonn used for such scaling.— 3.1).Al) is called the characteristic polynomial Definition 9. The polynomial n (A) = det(A—A. called an eigenvalue. we see immediately that x H is a left eigenBy taking Hermitian transposes in (9. [3]).} The following classical theorem can be very useful in hand calculation.t|| so that the scaled eigenvector has norm 1. then n(A) is a polynomial of degree n. Thus.Chapter 9 Chapter 9 Eigenvalues and Eigenvalues and Eigenvectors Eigenvectors 9. e C. Then n(k) = X2 + 2A. It can be proved from elementary properties of determinants that if A E C" ". a nonzero vector y e C" is a left eigenvector corresponding to an eigenvalue a if Mif (9.A). n(A) = 0.. Definition 9. we see immediately that XH is a left eigenvector of A H associated with A.) = det (A . for example./) is called the characteristic polynomial of A.2. then so is ax [ay] for any nonzero scalar a E C.) throughout the text. Then n(A) A2 + 2A 3. It can be The following classical theorem can be very useful in hand calculation.

Definition 9.e. The minimal polynomial Of A l::: K""" ix (hI' polynomilll a(A) of least degree such that a(A) ~ O. but that A(A) = A(A) only if A E 1Ftnxn. (9.4) and set A = 0 in this identity..:. Specifically. Moreover.5. then I < dimA/"(A — A/) < m. that by elementary properties of the determinant. too.~. of we always have A(A) = A(A r ). . but that A(A) = A(A) only if A e R"x". Eigenvalues and Eigenvectors Chapter 9.XI. These roots.. then y is a right eigenvector of AT corresponding to I € A (A). Note.7. The geometric multiplicity of A is the number of associated independent eigenvectors = n — rank(A — A/) = dim J\f(A — XI).3) in the n(A) = det(A . ft E 1Ft and let A = [ _^ !]...5.n =0o. However.2». If e Wxn.e.AI) = (A] . sible for A to satisfy a lower-order polynomial. if If A € A(A) has algebraic multiplicity m. eigenvalues of A. we say that X is an eigenvalue of A of algebraic multiplicity m. If A E A(A) has algebraic multiplicity m..8. But it also clearly satisfies the smaller degree polynomial equation (it. n(A). Hence the roots of 7r(A). as solutions of the determinant equation 7r(A) has n roots. y of AT y is a left eigenvector of A corresponding to A e A(A).. Equivalently. then n(X) has real coefficients. then 1 :::: dimN(A . guarantee the existence of corresponding nonzero eigenvectors. we denote the geometric multiplicity of A by g. all roots of its characteristic polynomialn(A). neftnhion ~.AI) :::: m. it is posn(A) = O. E A(A). Equivalently. such a polynomial is said to be monic and we generally write et (A) as a monic polynomial throughout the text).1) .. Then jr(A. the n(A) coefficients. we know that n(A) = 0. eigenvalues of A. i •>/—!)• If A E 1Ftnxn..2aA + aa2+ f322 and A has eigenvalues a f3j (where A has eigenvalues a ± fij (where j = i = R). checked eigenvectors of A and AT (take Hermitian transposes of both sides of (9.25)..6.e. The spectrum of A E C"x" is the set of all eigenvalues of A. less than) its algebraic multiplicity.) A. But it also clearly satisfies the smaller degree polynomial equation isfies (1 . say. f3 e R and let A = [~f3 £ ]. c form form A e C" " A]. if left of A A E A(A). must occur in complex conjugate pairs. Moreover. A is said to be defective if it does not have n linearly independent (right or left) eigenvectors. Then if we write (9. then we must have I < g < m.3) are the eigenvalues of A and imply the singularity of the matrix A . if we denote the geometric multiplicity of A by g. For example. we get the interesting fact that del (A) = A] • A2 • • An (see also Theorem 9. then there is an easily checked relationship between the left and right If A € R"x".e. Let the eigenvalues of A E en xxn be denoted X\. it can also be generally write a(A) as a monic polynomial throughout the text). must occur in complex conjugate pairs. Xn. IfXA is a root of multiplicity m ofjr(X). Eigenvalues and Eigenvectors n(A) has n roots.A) . Example 9. The spectrum of A is denoted A(A). a . A matrix A e 1Ft x" is said to be defective if it has an eigenvalue whose geometric multiplicity is not equal to (i. Definition 9.. degree such that a (A) =0. A matrix A E Wnxn is said to be defective if it has an eigenvalue whose Definition 9..rank(A .A) (9. if eigenvectors of A and AT (take Hermitian transposes of both sides of (9.AI).e. we get the interesting fact that det(A) = AI . the set of all roots of its characteristic polynomial n(X).7. possibly repeated. we always have A(A) = A(AT).nxn is the polynomial o/(X) oJ IPll. Definition 9. The spectrum of A e nxn is the set of all eigenvalues of A.ft Definition 5. less than) its algebraic multiplicity.. Then n(A) = A22. such a polynomial is said to be monic and we of the highest power of A to be +1.. From the Cayley-Hamilton Theorem. These roots. then we must have 1 :::: g :::: m.5. independent eigenvectors = n . the set of Definition 9. then A satsible for A to satisfy a lower-order polynomial. and hence further guarantee the existence of corresponding nonzero eigenvectors.8. we say that A is an eigenvalue of A Definition 9. Let a.. Thus. Theorem If A E 1Ftnxn. then A satisfies (Je — I)2 = 0. i. For example. An.2..1)2 = O. The geometric multiplicity ofX is the number of associated of algebraic multiplicity m. possibly repeated.AI) = dimN(A . and hence further are the eigenvalues of A and imply the singularity of the matrix A — AI. i. it can also be . A is said to be defective if it does not have n linearly independent (right or left) eigenvectors. i.. However. 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..76 Chapter 9. Let a. as solutions of the determinant equation n(A) = det(A - AI) = 0. if A = [~ ~]. Specifically. If is a root of multiplicity m of n(A). geometric multiplicity is not equal to (i. say.. A. . if A = \1Q ®].6. The spectrum of A is denoted A (A). 2aA + 2 + ft and Example 9. Thll minimal polynomial of A G l!if.. (An . • AM(see and set X = 0 in this identity. Thus. A.2)).

.10. be an eigenvalue of A with corresponding right Theorem 9. e l\(A) yj Xi.11. every nonzero polynomial f3(A) particular.2) andg ~ 4.2)2 ""d g 3. Bezout algorithm. algorithm. let Yj be a left eigenvector corresponding to any A.*. a(A) divides n(A). We denote 7r(A) (A . Proof' Since AXi AiXi.) divides every nonzero polynomial fi(k} for which ft (A) = 0.. YY Proof: Since Axt = A.. such that Aj =£ Ai. this algorithm.. called the Bezout algorithm. Fundamental Definitions and Properties 9. Then yfx{ = O.9. Furthermore. a(X) n(X). 0 0 0 2 0 0 0 2 ] h'" a(A) (A . The matrix A~U has a(A) I 2 0 0 2 0 0 0 !] (9. Unfortunately. Example 9. 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. of which has an eigenvalue 2 of algebraic multiplicity 4. such is not the case. each 4. 0 g At this point. The above definitions are illustrated below for a series of matrices. the geometric multiplicity by g.1.2)' ""d g ~ ~ ~ 1. In particular.2)4. sociated eigenvector structure). eigenvector is numerically unstable. one might speculate that g plus the degree of a must always be five. g..e.10. n(A) = (A — 2)4. Unfortunately.-. left Aj E A (A) such that Xj 1= A. Example 9."(A) ~ ~ ~ ~ (A . shown that a (A.2)' ""d g 2. 0 0 0 2 A~U 0 0 ] ha<a(A) (A .1. Let A e C« x " ana [et A. a(A) directly (without knowing eigenvalues and asThere is an algorithm to determine or (A. A~[~ A~U 2 0 0 I 2 2 ] ha< a(A) (A .5) = (A - 2)2 and g = 2. Then Xi = 0.11. Theorem 9. Fundamental Definitions and Properties 77 77 a(A) f3(A) O.. i. A-[~ - 2 0 I 2 0 0 0 0 0 0 !] ~ ~ ~ ha.

or both. we must have Subtracting (9. jc. Let A E C"x" be Hermitian. from which conclude A.7) Taking Hermitian transposes in (9.. since y" A = Similarly.13. . is real to get H Az AxH z. (9. be real. i..13.7. results.6) from (9. we find 0 = (Ai . However..n with corresponding right eigenvectors x\. .... 's. Since Ai .5). x) is an arbitrary eigenvalue/eigenvector pair such that Ax = AX.. we must have that x H = 0.. for example. ^ /z. Premultiply the equation Az = iJ-Z by x H to get x HAz = /^XHZZ = XxHz. i. we have that XxH = AXH x. However. [21. p. it cannot be the case that YiH Xi = 0 as well. 0 Let us now return to the general case.. 1 ?. .14) and would thus have to be 0.7) yields Using the fact that A is Hermitian. Let A e nxn be Hermitian..JC.JC by ZH to get ZH Ax = X z Hx . 0 If A E cnx " has distinct eigenvalues.is real. Let us now return to the general case. 's..e. The same right eigenvectors XI. is If A e C nxn has distinct eigenvalues. 0 The proof of Theorem 9. since x is an Using the fact that A is Hermitian. A. Eigenvalues and Eigenvectors yy.. for [21. we find 0 = (A. the proof see.— A y )j^jc.. i. Then all eigenvalues of A must Theorem 9. since x is an eigenvector.e. D A =1= iJ-. the two vectors must be orthogonal. 0 D Theorem 9. since YY A = AjXjyf. Since yf*Xi =1= 0 for each i. Xi is orthogonal to all y/s for which j =1= i.14. then by Theorem 9. Let A e C"x" be Hermitian and suppose A and /JL are distinct eigenvalues Theorem 9. A = AH. Theorem 9..7) yields Taking Hermitian transposes in (9.. Then and z must be orthogonal.Aj)YY xi.14..5)..xnn • Then {XI. i.12. Cnxn have distinct eigenvalues AI. we can choose the normalization of the *. so that y H Xi = 1 for each i.e. Then [x\.78 Similarly. for i € !1.. The same result holds for the corresponding left eigenvectors. 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. Take the Hermitian Proof: Premultiply the equation Ax = AX by ZH to get ZH Ax = AZ H x.11. Eigenvalues and Eigenvectors Chapter 9. However. Then all eigenvalues of A must be real. 118]. Since XxHz. i. YyXi =0. XH x /= 0. x) is an arbitrary eigenvalue/eigenvector pair such that Ax = A. and if Ai E A(A).11. .are distinct eigenvalues of A with corresponding right eigenvectors x and z. we have xHX =1= 0. then by Theorem 9. Since A. so that YitH x.. orthogonal. . An with corresponding Theorem 9.11 is very similar to two other fundamental and important The proof of Theorem 9.— A..12. = A.11 is very similar to two other fundamental and important results.e. . A is real. contradicting the fact that it is an eigenvector. eigenvector. . 118]. yr .. or the y... Let A E nxn be Hermitian and suppose X and iJ. we must have that XHzz = 0.e. it cannot be the case that yf*xt = 0 as orthogonal to all yj's for which j ^ i. Since Xi ^ 0 and would thus have to be 0.14) well... Then (9.. we must have yfxt = O. or the Yi 's. the two vectors must be orthogonal. respectively. and if A. However. result holds for the corresponding left eigenvectors. Then Proof: Suppose (A. c Proof: Suppose (A.Aj ^ 0. e A(A). from which we conclude I = A.e.=1= 0. Let A €. or else xf would be orthogonal to n linearly independent vectors (by Theorem 9. c Proof: Premultiply the equation Ax = A. or else Xi would be orthogonal to n linearly independent vectors (by Theorem 9. Chapter 9. A = AH. contradicting the fact that it is an eigenvector.. we have that IXHxx = XxHx. x is a linearly independent set. AXH z. i. xn}} is a linearly independent set.6) Subtracting (9. Proof: Proof: For the proof see. Theorem 9. = 1 for/ E n. Then x and z must be of A with corresponding right eigenvectors and respectively. A. x .. p. Since equation Az i^z XH get X H Az = iJ-XH A. or both. Let A E cnxn have distinct eigenvalues A . we can choose the normalization of the Xi'S.6) from (9..

1. .j. Similarly.(-1 + 2j) I)x2 = 0 to get For A2 — 1 + 2j. is expressed by the equation while YiH Xj = oij.10) = XAX. 10) -(A.15.8) while y^Xj = 5. .. Let A e en xn have distinct eigenvalues AI..ci can be set arbitrarily. -1 ± 2j}... solve the 3 x 3 linear system (A . xn].. + 5).I = = LAixiyr i=1 (9..11) Example 9. . is expressed by the equation yHX = I.. .(-2)l)xI = 0 to get Note that one component of . y' E !!...15. . . i E !!:: Finally. let Y — [y\. Fundamental Definitions and Properties 9. 2)(A. suppose that the left and right eigenvectors have been normalized so that yf1 Xi = 1. . An) e ]Rnxn. solve the linear system (A — (—1 + 2j)I)x2 = 0 to get yi X2 =[ 3+ j ] 3 ~/ . Similarly. from which we find A(A) = {—2.2 + 9)" + 10) = -()" + 2)(). Yn] ing right eigenvectors form a matrix X = [XI. . solve the 3 x 3 linear system (A — (—2}I)x\ = 0 to get For A-i = -2. Furthermore.(-2)1) = 1). Finally. This time we have chosen the arbitrary scale factor for y\ so that \ = 1. can be written in matrixform as diag(A.9) =A = XAyH yRAX n (9.. / en. . corresponding to these eigenvalues. .. Furthermore. We can now find the right and left eigenvectors which we find A (A) = {-2. Let Example 9.A.... let A = right eigenvectors have been normalized so that YiH Xi = 1."" yn] be the matrix of corresponding left eigenvectors. A. To get the corresponding left eigenvector YI.16. 2A.. / e n.. . To get the corresponding left eigenvector y\.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. — 1 2 j } . suppose that the left and be the matrix of corresponding left eigenvectors. solve the (since dimN(A . Let 2 5 -3 -3 -2 -4 ~ ] .AI) -(A.22 + 2)" + 5). and this then determines the other two (since dimA/XA — (—2)7) = 1). For A2 = -1 + 2j.-*/.. j e n. Then rr(A) = det(A .16.1. i E !!.3 4A2 9 A.I . Let A E C"x" have distinct eigenvalues A. These matrix equations can be combined to yield the following matrix factorizations: These matrix equations can be combined to yield the following matrix factorizations: X-lAX and and A (9.. / en. can be written in matrix form as AX=XA (9. and this then determines the other two Note that one component of XI can be set arbitrarily. For Al = —2. i E!!. An and let the corresponding right eigenvectors form a matrix X [x\. Fundamental Definitions and 79 Theorem 9. let Y = [YI. let A = diag(AJ. from Then n(X) det(A . Then AJC. = A.nand let the correspondTheorem 9. Xn) E Wtxn. solve the linear system (A . .. We can now find the right and left eigenvectors corresponding to these eigenvalues.. Then AXi = AiXi. . xn]./) = -(A 3 + 4A.

j ] 3+j .) 19X + 12) = -(A.2 and simply can also note that x$ = X2 and Y3 = Y2. For example.. we For XT./) _(A + 8A from which we find A(A) = {—1. Then. we could proceed to solve linear systems as for A2. Let Example 9. = -I . we can also note that X3 =x2' and yi jj. Eigenvalues and Eigenvectors Chapter 9. Finally. To see this. similar argument yields the result for left eigenvectors.15 can also be verified.'s directly. Proceeding as in the previous example. we could have found them instead by computing instead of detennining the Yi'S directly. Then. -2 It is then easy to verify that It is then easy to verify that -2 .!. + 4). + 3)(A. Let A = [-~ -~ ~] .A. itit is gtruightforw!U"d to compute straightforward to comput~ X~[~ and and I 0 -I -i ] 1 x-.L 4 !. + 1)(A. we could have found them instead by computing X-I and reading off its rows. Proceeding as in the previous example.=. To see this.17. X~l Example 9. —4}. —3.=.c2 = ^. note that we could have solved directly only for *i and x2 (and XT.AI) = -(A33 + 8A 22+ 19A + 12) = -(A + I)(A + 3)(A + 4).( -I + 2 j) I) = 0 and nonnalize Y22 so that y"xX2 = 1 to get For A3 = — 1 — 2j.!.80 Chapter 9. Other results in Theorem 9. we could proceed to solve linear systems as for A. use the fact that A.L Other results in Theorem 9.2 However. instead of determining the j. det(A . Then Jl"(A) = det(A . A.A similar argument yields the result conjugate the equation AX2 — A2X2 to get AX2 A2X2.2*2 to get Ax^ = ^2X2. X-IAX=A= [ -2 0 0 -1+2j o 0 Finally.±1 4 4 4 l+j . -3. Now define the matrix X of right eigenvectors: Now define the matrix of right eigenvectors: 3+j 3-j 3.~q 1 3 2 2 0 -2 -3 ] 2 ~ y' . = x2). -4}. o -3 Then 7r(A.17. 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 . However. note that we could have solved directly only for XI and X2 (and X3 = X2). For example. is from which we find A (A) = {-I.15 can also be verified. for left eigenvectors.2j. use the fact that A33 = A2 and simply conjugate the equation 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. A Theorem 9.g. A = T0 6 2 has only one right eigenvector corresponding to the eigenvalue 0. from the theorem statement follows.20. or. A is diagonalizable). Theorem 9. Proof: Suppose (A.lIAT)(T~lx)x) = X ( T ~ lIxx). Then suppose X-I AX = A. -4). Eigenvalues (but not eigenvectors) are invariant under a similarity transTheorem 9. formation T. or sin*. which is equivalent to the dyadic expansion sion 3 A = LAixiyr i=1 ~(-I)[ ~ W~l+(-3)[ j ][~ ~ 1 . but A = [~ has two independent right eigenvectors associated with the eigenvalue o. which is equivalent to the dyadic expanWe also have X~l AX = A = diag( -1. ( x ) is a polynomial. A = [~ Oj ] have all the same eigenvectors (unless. in general..1 AT) =X(THyf. For example. For left eigenvectors we have a similar statement. Fundamental Definitions and Properties 81 81 We also have X-I AX = A = diag(—1. x) but not conversely.18.3 0 -~l +(-4) [ -. I 3 I (. 2 3 I (. namely yH A AyH if and only if Hy)H(T~1AT) = A(T Hy)H. say.1. What is true is that the independent right eigenvectors associated with the eigenvalue 0. A is diagonalizable). of Chapter 11. since T is nonsingular. Remark 9. ] [~ ~ (-I) [ I (.9. —3. 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. What is true is that the eigenvalue/eigenvector pair (A. x) maps to (f(A). X) is an eigenvalue/eigenvector pair such that Ax = AX. If / is an analytic function (e. representable by a power series X^^o fln*n)> then it is easy to show that representable L~:O anxn). since T Proof: Suppose (A. Remark 9.. we have the equivalent statement (T. If f is an analytic function (e. -3. Then.but f(A) does not necessarily have all the same eigenvectors (unless. then easy to show that the eigenvalues of f(A) (defined as L~:OanAn) are f(A). The following theorem is useful when solving systems of linear differential equations. or ex. eigenvalue/eigenvector pair (A.) .19. For left eigenvectors we have a similar statement. but /(A) does not necessarily the eigenvalues of /(A) (defined as X^o^-A") are /(A). Eigenvalues (but not eigenvectors) are invariant under a similarity transformation T.txiYiH. The following theorem is useful when solving systems of linear differential equations. For example. but A2 = f0 0~1]has two has only one right eigenvector corresponding to the eigenvalue 0. D D yHA = XyH ifandon\yif(T(T Hy)H (T. or eX. Then.19. or sinx. I I ~J I 2 0 0 0 3 3 3 I I (. from which equivalent statement (T~ AT)(T. Fundamental Definitions and Properties 9. —4). x) but not conversely. say. jc) is an eigenvalue/eigenvector pair such that Ax = Xx. e jRnxn n = LeA. ff(x) is a polynomial.1. x) maps to (/(A). namely the theorem statement follows.g.18. where A is diagonal. Let A E R" xn and suppose X~~1AX — A.I = A(T.20. i=1 .

.= Xdiag(/(A..e.22..20 and its corollary in which A is not necessarily diagonalizable.21. . . eigenvectors Xi.IXiYiH. Eigenvalues and Eigenvectors Chapter 9. i=1 0 The following corollary is immediate from the theorem upon setting t = I. I. kn E C (not necessarily distinct).12) where each of the lordan block matrices 1i .. 1q)..20 and Corollary 9. 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. we have Proof' Starting from the definition. Corollary 9.. analytic on the spectrum of A... to have a version of Theorem 9. then e A has eigenvalues e A There are extensions to Theorem 9. lordan Canonical Form (JCF): For all A E C"x" with eigenvalues X\. ..2 9.. from which such a result is then available and presented later in this chapter... . ii E ~. and right eigenvectors xt•. there exists X € c~xn such that (not necessarily distinct).. If A e ]Rn xn is diagonalizable with eigenvalues Ai.82 Proof: Starting from the definition.. and the same eigenvectors.21 for any function that isis There are extensions to Theorem 9.-. canonical form. ( A ) = Xf(A)X.. i E ~. If A E Rnx" is diagonalizable with eigenvalues A. ..21 for any function that analytic on the spectrum of A. f ( X t t ) ) X ~ It is desirable.. . and right Corollary 9. from which such a result is then available and presented later in this chapter.. € n_. i.22. we have Chapter 9.. Theorem 9. i E ~. i.Il . The following corollary is immediate from the theorem upon setting t = I. / € n_.e. /' en. 9.20 and Corollary 9. It is necessary first to consider the notion of Jordan canonical form. f(An))X.13) 1i = o o Ai o Ai . of course. . ..21. of course. then eA has eigenvalues e X"i . ff(A) = X f ( A ) X ~ l I = Xdiag(J(AI).. 0 (9. An e C 1. there exists X E C^x" such that X-I AX = 1 = diag(ll.2 Jordan Canonical Form Jordan Canonical Form Theorem 9. to have a version of Theorem 9. Eigenvalues and Eigenvectors n = LeA. It is necessary first to consider the notion of Jordan A is not necessarily diagonalizable. and the same eigenvectors.i).. Jordan Canonical Form (/CF): For all A e c nxn with eigenvalues AI.20 and its corollary in which It is desirable.

. [21.. pp.2. . for example.. 83 83 Form: 2. Jordan Canonical Form and L. 1q is of form where each of the Jordan block matrices 11. there exists X € R" xn such that (9. complicated.. .. . ~: ] and I = \0 A in the case of complex conjugate eigenvalues a ± jfJi E A(A). With 1 -j o -j o 1 o o o -j ~ -~] 0 1 ' .An n (not € jRnxn Xi. .. Real Jordan Canonical Form: For all A E R n x " with eigenvalues AI.. Jq is of the form of in the case of real eigenvalues A.JfJ =[ (X -fJ fJ ] (X = M.9. . For nontrivial Jordan blocks.14) J\.=1 ki = n. Proof: proof D 0 Transformations like T = [ _~ -"•{"]allow us to go back and forth between aareal JCF Transformations like T = I" _. 120-124]... X (not necessarily X E lR. Proof: For the proof see..~xn necessarily distinct). the situation is only a bit more complicated.2. Jordan Canonical Form 9.. . ] T (X . and where M. = [ _»' ^ 1 and h2 = [6 ~] in the case of complex conjugate eigenvalues where Mi = [ _~. (Xii±jpieA(A>). { ] allow us to go back and forth between real JCF and its complex counterpart: T-I [ (X + jfJ o O. e A (A).

. det(A) = nAi. The characteristic polynomials of the Jordan blocks defined in Theorem Definition 9.-.(A. Again.25. Then Theorem 9. 1)2(A . From Theorem 9. J(2) has elementary divisors (A while /( 2) haselementary divisors (A . Suppose A e lR. -(A1).2). 2 . x Theorem 9.84 it is easily checked that it is easily checked that Chapter 9. Then c n 1. Tr(A) = Tr(XJX~ ) = TrC/X"1 X) = Tr(/) = £"=1 A. Tr(A) = 2.I)(A (A2)2. Suppose A E E (A.. Eigenvalues and Eigenvectors T. i=1 n 2. Eigenvalues and Eigenvectors Chapter 9. Thus.22 are called the elementary divisors or invariant factors of A. 1).26.2).jf3 0 0 et .— I) (A. From Theorem 9.) = det(7) = ]~["=l A. 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 . D 0 Example 9. 2(A(A. Let A e C" " with eigenvalues AI.2). 2)2. X X-I. and (A .) = (A.22 are called the elementary divisors or invariant factors of A.23.24. Let A E nxn with eigenvalues AI. det(A) = det(XJX. 1 det(A) = det(X J X-I) det(J) = n7=1 Ai.)i. 1). 1. The characteristic polynomial of a matrix is the product of its elementary Theorem 9.2)2. The minimal polynomial of a matrix is the product of the elementary divisors of highest degree corresponding to distinct eigenvalues.(A 1). .1*) = Tr(J) = L7=1 Ai. and (A -(A . .2)2.22 l Tr(A) = Tr(X J X-I) Tr(JX.-1)z. Then AAhas two possible JCFs (not counting reorderings of the a (A.22 we have that A X J X-I.. . " Xn. highest degree corresponding to distinct eigenvalues. 1)4(A 2) and 2 2 et(A) = (A . ..(A.25. The characteristic polynomial of a matrix is the product of its elementary divisors.22 we have that A = X J X ~ . .23.2)3 3and is known to have :rr(A) Example 9. 2. . from Theorem 9. An. . — 2) ... I) .1)4(A .22 we have that A = X JJX ~ l .. and (A . i=1 l Proof: Proof: 1. Thus. and 2).. 9.I [ "+ jfi 0 0 0 et + jf3 0 0 0 0 et .-1)2.1)2. from Theorem 9. .24.. Thus. The characteristic polynomials of the Jordan blocks defined in Theorem 9.7x7 is known to have 7r(A) = (A . Theorem 9.26..-. The minimal polynomial of a matrix is the product of the elementary divisors of divisors.jf3 0 ]T~[~ l h M Definition 9. I) .2)2.2(A(A .

To get a third vector X3 such that X = [Xl KJ_ X3] reduces A to JCF.. The more interesting (and difficult) case occurs when Ai multiplicity A.A. we find that 2£2 + ~3 = O. a(A.rank(A . and rank(A al) vectors. is not sufficient to Example 9. 9. The matrices A uniquely.) = (A. a (A).3.e.— al) == 4. is of algebraic multiplicity greater than one.27.nxn). both denote a solution to the linear system (A . X principal Definition 9..7) for distinct A.. Knowing TT (A. Thus. 1.e. we find that 2~2 + £ 3= 0 .. a)7. the associated number of linearly A.3.. and rank (A A.. a)\ .nxn number of eigenvectors.).29..A. of algebraic multiplicity 1. we need the notion of principal vector. c . X e A(A) if and only if (A XI)kx = 0 and (A U}k~l x ^ 0. when Ai is simple. suppose suppose A = [3 2 0 o Then Then 3 0 A-3I= U2 I] o o 0 0 n has rank 1. Thus. An analogous definition holds for a left principal vector of degree k. of algebraic multiplicity 1. of course. so the eigenvalue 3 has two eigenvectors associated with it. a(A). i.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.9.l) independent right — — A. eigenvectors dimN(A — A..l). For example.] are eigenvectors (and are independent).. Remark 9. it then has precisely one eigenvector. To get a third vector JC3 such that X [x\ X2 XT.AI)k-lx i= o. left k. i. If we let [~l ~2 ~3]T associated If [^i £2 &]T denote a solution to the linear system (A — 3l)~ = 0. 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 4.3/)£ = 0.28. i. it The straightforward case is.is simple.e.— a) and rank(A . Let A E C"xn (or R"x"). Determination of the JCF 9. A e nxn ]R. Definition 9.) = (A.e. of course. Determination of the JCF 85 &5 Example 9. and rank(A -—Ai l) for distinct Ai is not sufficient to rr(A). associated independent right (or left) eigenvectors is given by dim A^(A .ulx = 0 and (A -. The straightforward case is.a(A) = (A . three eigen7r(A. i.7) = n .27.(7)..29. when X.28. determine the JCF of A uniquely. For each distinct eigenvalue Ai. both are eigenvectors (and are independent). three eigenboth have rr(A) = (A .— a) . Then x is a right principal vector of degree k degree associated with A E A (A) ifand only if(A -. Remark 9.

S.X I ) ( l ) = (A AI)O o. the principal vector second of degree 2: of degree (A . of course.AI).3. Solve (A .XI). x(l).1 Theoretical computation Theoretical computation To motivate the development of a procedure for determining principal vectors. Principal vectors are sometimes also called generalized eigenvectors. The other solution necessary is the desired principal vector of degree 2. principal vectors of degree 1) associated with A. Eigenvalues and Eigenvectors Chapter 9. Then for each distinct X e A (A) perform the following: z (2) w c 1. wefind(A-. since (A . solutions solutions to the homogeneous equation (A . which simply says that x(!) is a right Ax(1) = hx(1) x (1) (2) x(2). Then the equation AX = XJ can be written that reduces a matrix A to this JCF. (It may be necessary to take a linear of x(l) R(A .AI). See. by (A . The case k = 1 corresponds to the "usual" eigenvector.A/)x(2) = x(l). For example.X I ) .17) by (A . Exercise 7.XI)2x^ = 0. I) associated This step finds all the eigenvectors (i. determine all eigenvalues of A e R" x " nxn ). If. Eigenvalues and Eigenvectors synonymously "of 2. ji of dimension k or larger. Then the equation AX = X J can be written A [x(l) x(2)] = [x(l) X(2)] [~ ~ J. For each independent jc (1) . if X.3.e. (9. If the algebraic multiplicity of If A principal need X is greater than its geometric multiplicity. of The number of linearly independent solutions at this step depends on the rank of 2 (A . for get right-hand example. First. but the latter generalized eigenvectors.XI.86 Chapter 9. (A — uf. is. One of these solutions (A — AI)2 x (2) x(l) (1= 0).A/) = — multiplicity of rank(A — XI) = n . this rank is n . different term will be assigned a much different meaning in Chapter 12. See. the definition of principal vector is satisfied. A right (or left) principal vector of degree k is associated with a Jordan block J. consider a determining 2 x Jordan [~ i]. 2. The phrase "of grade k" is often used synonymously with "of degree k. Denote by x(l) and x(2) the two columns of a matrix X E lR~X2 2x2 2 Jordan block{h0 h1. 9.. solve (A .x2 A JCF. of k 5.A1)2 x(2) = (A -. Thus.'A1)22xx(l) = (A . E lR nxn This suggests a "general" procedure.1 9. we find (A If we premultiply XI) x = (A XI)x = 0. eigenvector.17) The first column yields the equation Ax(!) = AX(!). there are two linearly independent n — o. principal vectors still need to be computed from succeeding steps. Thus.XI)0 = 0." "of often 3. x(l) (^ 0).2. k = eigenvector. If we premultiply (9.1. 4. if of . The second column yields the following equation for x .A1)X(l) = O.A1)x(2) = x(l).A1)X(l) = O. of — AI.) . there is only one eigenvector. for example. combination of jc(1) vectors to get a right-hand side that is in 7£(A — XI). A E A(A) following: (or C ). Theother solution (A . Denote by x(1) and x(2) the two columns of a matrix X e R2. The number of eigenvectors depends on the rank of A .

3. say). X(k)]. 1 . find the eigenvectors associated The eigenvalues of A are A1 = I.. Notice that high-quality mathematical software such as MATLAB readable [8] to learn why. find the eigenvectors associated with the distinct eigenvalues 1 and 2. Notice that high-quality mathematical software such as MATLAB does not offer a j cf command.2I)x~1) = 0 yields . . and the interested student is strongly urged to consult the classical and very readable [8] to learn why. . Determination of eigenvectors more extensive treatments. Determination of the JCF 3. First. solve 3. of algebraic multiplicity and Theorem 9. . Attempts to do such calculations in finite-precision floating-point arithmetic or 3.3. with the distinct eigenvalues 1 and 2. where the chain of suppose further that rank(A . see. solve (A AI)x(3) 87 = x(2)..32. say). Continue in this way until the total number of independent eigenvectors and principal vectors is equal to the algebraic multiplicity of A. . Example 9. Unfortunately..31. vectors is equal to the algebraic multiplicity of A. and A3 = 2. although a j ardan command is available in MATLAB'S does not offer a jcf command. Let Example 9. .33. Then Theorem 9.1. Theorem 9. [20] and [21]. First. for example.2/)x3(1)= 0 yields (A .(1) (A . For more extensive treatments.30. . 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. Attempts to do such calculations in finite-precision floating-point arithmetic generally prove unreliable. this natural-looking procedure can fail to find all Jordan vectors. 4. and the interested student is strongly urged to consult the classical and very to compute a JCF. where the chain of vectors x(i) is constructed as above. Symbolic Toolbox.. Then vectors x(i) is constructed as above.. .32.33. Let A=[~ 0 2 ] . There are significant numerical difficulties inherent in attempting to compute a JCF. .AI) = k . There are significant numerical difficulties inherent in attempting generally prove unreliable. A2 = 1. pendent. although a jordan command is available in MATLAB's Symbolic Toolbox. Continue in this way until the total number of independent eigenvectors and principal 4. (x (1) . x(k)].30. Let X = x ( l ) . Suppose A E C kxk has an eigenvalue A of algebraic multiplicity kkand suppose further that rank(A — AI) = k — 1. For each independent x(2) from step 2. Suppose A e Ckxk has an eigenvalue A.9. Theorem 9. Let X = [[x(l). h2 = 1. [20] and [21]. Principal vectors associated with different Jordan blocks are linearly independent. . and h3 = 2.. X(k)} is a linearly independent set. for example. {x(l). x (k) } is a linearly independent set. Theorem 9. For Unfortunately. 0 The eigenvalues of A are AI = 1.. Determination of the JCF 9. Principal vectors associated with different Jordan blocks are linearly indeTheorem 9. see. .31. For each independent X(2) from step 2. this natural-looking procedure can fail to find all Jordan vectors.

For the sake of definiteness.. . Then Let D = diag(d" . d n)) be a nonsingular "scaling" matrix. solve 2 (A . we 1 's but can be arbitrary — so long as they are nonzero. Eigenvalues and Eigenvectors To find a principal vector of degree 2 associated with the multiple eigenvalue 1.. .3. solve To find a principal vector of degree 2 associated with the multiple eigenvalue 1. d. .11)x?J = 0 yields (A..l/)x. we consider below the case of a single Jordan block. . 0 = 0 A dn dn I 2 0 dn dn I A- 0 ).2 9. 0 0 D-'(X-' AX)D = D-' J D = j ). d.(2) = x.. Now let Now let X (2) =[ 0 ] ~ . but the result clearly holds for any JCF.. For the sake of defmiteness.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 . 0 !b. (1) toeet x. Suppose A € Rnxn and SupposedA E jRnxn and Let D diag(d1.3. 0 1 = [xiI) 0 xl" xl"] ~ [ ~ -5 ] and X-lAX 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. but the result clearly holds for any JCF.. Then A 4l. consider below the case of a single Jordan block. .88 (A . dn be a nonsingular "scaling" matrix..so long as they are nonzero.1I)xl ) = xiI) to get (A – l/)x. =0 yields (1) Chapter 9.

(A . Then jH. ... It is thus natural to expect an with respect to which the matrix is diagonal or block diagonal. A subspace S ~ V is A-invariant if AS ~ S. .A.. Let IF and suppose --+ transformation. x n eigenvectors and principal vectors that reduces A to its lCF.. .. It is thus natural to expect an associated direct sum decomposition of R.. dnxn}.18) 0 I 1 0 0 can be used to put the superdiagonal elements in the subdiagonal instead if that is desired: to superdiagonal elements in instead desired: A I 0 0 A 0 A 0 A 0 0 A 0 p-[ A p= 0 1 0 0 A 0 I A A 0 0 0 A 9.A[)n) . J is obtained from A via the similarity transformation XD = \d\x\.. .. . interpreted This result can also be interpreted in terms of the matrix X = [x\. . E6 N (A ..nxn (or nxn to JCF provides change of basis with respect to which the matrix is diagonal or block diagonal.Am)Vm with A-i. .AmItm . Specifically. the reverse-order identity matrix (or exchange matrix) 0 p = pT = p-[ = 0 I 0 (9.Am I) Vm . Definition 9. dimN(A — AJ)Vi = ni.. similarity transformation XD [d[x[.n = N(A = N (A . Geometric Aspects of the JCF 89 di's Appropriate choice of the di 's then yields any desired nonzero superdiagonal elements.. the reverse-order identity matrix (or exchange matrix) In a similar fashion. Suppose A E R"x" has characteristic polynomial 9.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 9.35.AlIt) E6 ... A subspace S c V is A -invariant if AS c S. Geometric Aspects of the JCF 9.9.A[)V) '" (A .A1I) v) E6 . Suppose e jH. dnxn]. Theorem 9. mdistinct. set {As s E S}.35..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. In a similar fashion. where AS is defined as the transformation... where AS is defined as the set {As:: s e S}... Such a decomposition is given in the following theorem.xn]] of eigenvectors = [x[. Such a decomposition is given in the following associated direct sum decomposition of jH.34. E6 N(A ..nxn n(A) = (A ..4. A.. Specifically. Note that dimM(A .4 9. Am distinct. j is obtained from A via the and principal vectors that reduces A to its JCF...34.n.4./) w = «. Then AI.-.. .

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.Xm] ] Ee]R~xnxnisis such that X^AX ==diag(J1. This follows easily by comparing the ith columns of each side of (9. we have that AXi Theorem 9. Xm R"n such that X-I AX diag(7i.. is A-invariant. is not necessarily diagonalizable. R(S) == S.lt.e. and S e R" xk s\..36... we have that A A. .. ..19): /th Example 9. the eigenvectors and principal vectors associated with Ai) span an A-invariant subspace of]Rn. then a basis for V can be chosen with respect to which A has a block diagonal representation. We would then get a block diagonal representation for A with full blocks rather than the highly structured Jordan blocks.. we could choose bases for N(A — A. X-I AX = [~ J 2 ].. 9. .34.2. Then N(p(A)) and 1.. Jm).* is a Jordan block corresponding to Ai E A(A). Let peA) = «o/ + o?i A + • + <xqA be a polynomial in A.. Then N(p(A)) and R(p(A)) 7£(p(A)) are A-invariant. A".. then a basis for V can be chosen with respect to which A has a block N.. Suppose X block diagonalizes A.90 Chapter 9. i.37. example (note that the power n.. = 1.g. 7. Example 9. . such "canonical" forms are discussed in text that follows. 9.. but we restrict our attention here to only the Jordan block case..Ji .. If V is a vector space over IF such that V = N\ EB ./)"'. diagonal representation. = Xi.span an A-invariant subspace. where each Ji = diag(/. where each Theorem 9. of W.. S is A-invariant if and only if S . The equation Ax = A* = x A defining a right eigenvector x of an eigenvalue AX = x A defining a right eigenvector x of an eigenvalue A x X says that * spans an A-invariant subspace (of dimension one).e. Eigenvalues and Eigenvectors If V is taken to be ]Rn over Rand S E ]Rn x* is a matrix whose columns SI.39. Jm).19) the columns of Xi (i. eigenvalues Ai 9. The Jordan canonical form is a special case of the above theorem. so by (9.) and each Jik is a Jordan block corresponding to A. = X. then S <S is A-invariant if and only if there exists M E ]Rkxk such that eRkxk (9. Eigenvalues and Eigenvectors Chapter 9. i /= 1. i. then is A-invariant if and only if there span a k-dimensional subspace S. the eigenvectors and principal vectors associated with A. . e A(A). . Sk If R" R..38. Finally.. Let p(A) = CloI + ClIA + '"• •+ ClqAqq be a polynomial in A. /..).37. 2..34. via SVD).i. Note that A A"..e. where X [ X i .e. Other such "canonical" forms are discussed in text that follows..39. e E"x"... . be a Jordan basis for N (AT .. Suppose A"== [Xl . A -invariant if only ifS1 1.-/)"' by SVD.• EB Nm.. € C"x"' be a Jordan basis for N(AT — A.. We would then get a block diagonal example (note that the power ni could be replaced by Vi). /. If A has distinct The Jordan canonical form is a special case of the above theorem. (i. A-invariant. each Ji = diag(JiI.A. so the columns of Xi span an A-mvanant subspace.as in Theorem 9. The equation Ax Example 9.36.38. Note that AXi = A*. //*.. Suppose A E ]Rnxn.19) the columns attention here to only the Jordan block case. We could also use other block diagonal decompositions (e. If F = NI ® • • 0 m A// is A-invariant. Other representation for A with full blocks rather than the highly structured Jordan blocks. so the columns of A.Ai/)n.19) AS = SM. partition . s/t span a /^-dimensional subspace <S. . Rewriting in the form ~ J.. i.) span an A-invariant of A". If A has distinct eigenvalues A. K(S) <S. for N(A .. Equivalently."" Jik.) and each /..2. Let 7. could be replaced by v..is A T -invariant.. .li. AT Theorem 9..e. so by (9. partition Equivalently.. Let Yi E <enxn .

S. Then the sign of A. E f= 0.5 9.lt 2 e At 2! 0 exp t 0 0 0 1 A teAt eAt 0 0 0 0 0 block Ji associated A = A.9. . of defined Definition 9.5.40.. Then A = XJX. Jm) [YI . denoted sgn(A).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 half-plane and P containing all Jordan blocks corresponding to eigenvalues of A in the left half-plane and P containing all Jordan blocks corresponding to eigenvalues in the right half-plane.. Suppose A E C"x" has no eigenvalues on the imaginary axis. Then the sign of z is defined by Re(z) {+1 sgn(z) = IRe(z) I = -1 ifRe(z) > 0.. 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 . 9. is given by sgn(A) = X [ -/ 0] 0 / X -I . .41. It is a generalization of the sign (or signum) of a scalar. Definition 9.41. . . ifRe(z) < O.. Let z E C with Re(z) ^ O. for a k x k Jordan block 7. A survey of the matrix sign function and some of its applications can be found in [15]. i=1 H In a similar fashion we can compute m etA = LXietJ.40. The Matrix Sign Function 9.. m ••• . Then compatibly. A called the matrix sign function. associated with an eigenvalue A.= Ai. Then the sign of A. It is a generalization of the sign (or signum) of a scalar. Xm] diag(JI.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. . The Matrix Sign Function 91 91 compatibly.JiYi . denoted sgn(A). is given by eigenvalues in the right half-plane.YiH. Ym]H = LX. and let e cnxn be a Jordan canonical form for A. Definition 9.. Definition 9.

sgn(AH) = (sgn(A))".43. . 5. Let A E Cnxn have distinct eigenvalues AI. of A (the negative invariant subspace). projection subspace of 4. 6. sgn(cA) = sgn(c) sgn(A)/or c. its reliable numerical calculation is an interesting topic in calculation its own right. sgn(T-lAT) = T-1sgn(A)T foralinonsingularT E C"x".. S2 = I. Let v E en be an vectors Xl. 2. sgn(T-1AT) T-lsgn(A)TforallnonsingularT e enxn 6... 2. 7l(S -l) is an A-invariant subspace corresponding to the left half-plane eigenvalues left half-plane I. 4. ). EXERCISES EXERCISES 1. and let = sgn(A). Xn and left eigenvectors y\. Then the following hold: following 1. negA == (l ..92 92 Chapter 9... Find the appropriate expression for v as a linear combination expression of the left eigenvectors as well. 2. . In fact.42. .. Xn with corresponding right eigenA e nxn ). 4. yn. Eigenvalues and Eigenvectors Chapter 9.. Their straightforward proofs are left to the exercises. S2 = I. Show that v can be expressed (uniquely) as a linear combination arbitrary vector. 3. positive = (/ + of A.1> . respectively. The JCF definition of the here is especially useful in deriving many of its key properties. We state some of the more useful properties of the matrix sign function as theorems. 3. S = sgn(A). Show that v can be expressed (uniquely) as a linear combination e of the right eigenvectors. distinct right eigenvectors Xi.42. S = sgn(A). ± 1. e nxn Theorem 9. Suppose A E C"x" has no eigenvalues on the imaginary axis. sgn(A") = (sgn(A»H. R(S+/) is an A-invariant subspace corresponding to the right half-plane eigenvalues R(S + l) A -invariant half-plane of (the positive invariant of A (the positive invariant subspace). and let — sgn(A). positive of P. Eigenvalues and Eigenvectors where the negative and positive identity matrices are of the same dimensions as N and p.n 1. . Theorem 9. Yn. We state some of the more useful properties of the matrix sign function as theorems. posA == (l + S)/2 is a projection onto the positive invariant subspace of A. Suppose A E enxn has no eigenvalues on the imaginary axis. e C"x" Theorem 9. but the one given There are other equivalent definitions of the matrix sign function. ••• . AS = SA. R(S — /) A-invariant of (the negative invariant subspace). Then the following hold: following e 1. ••.xn and left eigenvectors Yl. negA = (/ — S)/2 3.. There are other equivalent definitions of the matrix sign function.S) /2 is a projection onto the negative invariant subspace of A.43. Theorem 9.. Their left exercises.. 3. .. respectively. The JCF definition of the matrix sign function does not generally lend itself to reliable computation on a finite-wordgenerally itself length digital computer. AS = SA.. but the one given here is especially useful in deriving many of its key properties. sgn(cA) = sgn(c) sgn(A) for all nonzero real scalars c. respectively. .. S is diagonalizable with eigenvalues equal to del. Let e C" be an arbitrary vector. 5.

11. where x. Show that x is also a left eigenvector for A. 5. y E lR. 3}. Prove that all eigenvalues of a skew-Hermitian matrix must be pure imaginary. where x. Show that all right eigenvectors of the Jordan block matrix in Theorem 9.n T T x yy = 0. 10. Suppose the small number 10. Determine the JCF of A. where J is the JCF Find a nonsingular matrix X such that X AX = J.30 must be multiples of el e R*.5x5 has eigenvalues {2. Let A e R"x" be of the form A = xyT. 4. i. multiples of e\ E lR. Determine the JCF of A. 3. Show that x is also a left eigenvector for A. Characterize all left eigenvectors. What are the eigenvalues of this slightly perturbed is added to the (16.nxn A = xyT. AH = —A. Suppose A e rc nxn is Hermitian. Suppose 10~16 is added to the (16. 3}./)jc = x can't be solved. Prove the same result if A is skew-Hermitian. y e R" are nonzero vectors with A E lR. Let A E lR. (A — I)x(2) x(1) 8. Determine the eigenvalues..Exercises 93 93 2. ~ 0 Hint: Use[-1 1 — I]T an Hint: Use[— 1 1 . x O. right eigenvectors and right principal vectors if necessary. but then the equation (A .e. where x. 5. x.22. = O. Let A e R" xn be of the form A = 1+ xyT. Let A = [H -1]· 2 2" Find a nonsingular matrix X such that X-I AX = J.. 2. Suppose A E C"x" is Hermitian. Characterize all left eigenvectors. Determine the JCFs of the following matrices: 6. Suppose A E C"x" is skew-Hermitian. Let A be an eigenvalue of A with corresponding 3. n are nonzero vectors with with xTTyy = 0. The vectors [0 1 -Ifand[l 0 of [0 — l] r and[1 0]r (2) (1) are both eigenvectors.1) element of J. a skew-Hermitian matrix must be pure imaginary.30 must be 8. k . y e R" are nonzero vectors 10. y E lR. Suppose A € rc nxn is skew-Hermitian. 2. Prove the same result right eigenvector x. 9. 2. Suppose a matrix A E lR. JCFs for A. i. Suppose a matrix A E R 16x 16 has 16 eigenvalues at 0 and its JCF consists of a single A e lR. JCFs for A. Determine the JCFs of the following matrices: <a) Uj n -2 -1 2 =n 7.16 Jordan form specified 9. Determine the JCF of A. 16x 16 has eigenvalues at 0 its JCF consists of single Jordan block of the form specified in Theorem 9. where J is the JCF 1 J=[~ 0 1~]. Let 7. Prove that all eigenvalues of 2. Show that all right eigenvectors of the Jordan block matrix in Theorem 9. nxn be of the form A = / + xyT. if A is skew-Hermitian.1) element of J. eigenvalues. What are the eigenvalues of this slightly perturbed matrix? matrix? .e.22.l]r as an eigenvector. Determine the JCF of A. Let A be an eigenvalue of A with corresponding right eigenvector x. Determine all possible € R 5x5 {2. AH = -A. eigenvectors and if and (real) JCFs of the following matrices: (a) 2 -1 ] 0 ' [ 1 6.

Suppose A E C"xn has all its eigenvalues in the left half-plane. If n = 2 and k = 1. Prove Theorem 9. en . Hint: Use the factorization in the previous exercise. 15. The transformation P in (9. transformation explicitly. Find a matrix equation that X must satisfy for this to be possible.42. Hint: Use the factorization in the previous exercise. The transformation P in (9. 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(n-k). 14. in terms of All and A22. about when the equation for X is what can you say further.e. Then = ( X SIXT)(X. Thus. Prove that 17. xn has all its eigenvalues in the left half-plane. it suffices to prove the result for the JCF. T-IAT = [A011 A22 0 ] . 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. Prove that 17. X e R*x <«-*). Eigenvalues and Eigenvectors 12. Eigenvalues and Eigenvectors Chapter 9. Prove that every matrix e jRn xn is similar to its transpose and determine a similarity transformation explicitly. it suffices to prove the result for the required symmetric factorization of A. about when the equation for is solvable? solvable? 15. Consider the block upper triangular matrix 14. say S1. Prove Theorem 9. Suppose A e sgn(A) = -1. where Si 12. Show that every matrix A e jRnxn can be factored in the form A Si$2. 16. 13. is nonsingular. sgn(A) = -/.94 Chapter 9. Prove Theorem 9.43. and S2 are real symmetric matrices and one of them. in terms of AU and A 22. Show that every matrix A E R"x" can be factored in the form A = SIS2. If n = 2 and k = 1. Prove Theorem 9. Then A = (XS i X T ) ( X ~ T T S2X-I) would be the the "symmetric factorization" of J. where SI and £2 are real symmetric matrices and one of them. Thus. Find a matrix equation that X must satisfy for this to be possible. say Si. Suppose Au =1= 0 and that we we e jRnxn and All e jRkxk 1 < ::s: n. 16. Prove that every matrix A E W x" is similar to its transpose and determine a similarity 13.18) is useful. is nonsingular. JCF.43.S2X~l) would required symmetric factorization of A.. what can you say further. i.42. Hint: Suppose A = Xl X-I 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 1.18) is useful.

Xn) (the columns ofX are orthonormal eigenvectors for A).e. and orthogonal. . it is called an orthogonal equivalence if P and Q are orthogonal matrices. it is called an "canonical form. the transformation A H> PAP" 1 is called similarity.. AA = AHA).. An.2. Normal matrices include Hermitian. then there exists a unitary matrix £7 such that A = AH E en xxn has eigenvalues AI. respectively). !] Theorem 10. If a matrix A is not normal.1. respectively). Let A = AH e C"x" have (real) eigenvalues A... the for real scalars a and h. such as A = [_~ most "diagonal" we can get is the JCF described in Chapter 9.1.. . the transformation A f--+ P ApT is called 2..2. . the transformation A i-» PAPT is called If an orthogonal similarity (or unitary similarity in the complex case). What other U HAU = D. find P e lR.9. What other matrices are "diagonalizable" under unitary similarity? The answer is given in Theorem matrices are "diagonalizable" under unitary similarity? The answer is given in Theorem 10. An. Xn Theorem 10.. 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. Q are unitary.2.1 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. n ).. . skewsymmetric. an orthogonal similarity (or unitary similarity in the complex case). This is proved in Theorem 10. = H E en xn exists a unitary matrix X such that X H AX = D = diag(Al. The following results are typical of what can be achieved under a unitary similarity.2. such as A = [ _ab ^1 for real scalars a and b. and orthogonal.An. . matrix if and only if it is normal (i. if A E IR mxn find E R™ and Q E lR~xn such that P AQ has a form. V and Q 1.. If a matrix A is not normal. orthonormal eigenvectors for A). . .9. If P"1 . Two special cases are of interest: Two special cases are of interest: 1. Problem: Let V and W be vector spaces and suppose A : V —>• W is a linear transformation.Chapter 10 Chapter 10 Canonical Forms Canonical Forms 10.. if A e Rmxn . .. An) (the columns of X are exists a unitary matrix X such that XHAX = D = diag(A. . orthogonal equivalence if P and are orthogonal matrices. If A = A H 6 C" " has eigenvalues AI. 95 95 . If W = V and <2== p. = V and if pT is orthogonal. An)... This is proved in Theorem 10.." In matrix terms..j.j." The transformation A f--+ P AQ is called an equivalence. ..1 10. then there exists a unitary matrix U such that UH AU — D.1... where it is proved that a general matrix A E C"x" is unitarily similar to a diagonal 10. skew-Hermitian. . as well as other matrices that merely satisfy the definition. We can also consider the case A e Cm xn and unitary equivalence if P and Remark 10. as well as other matrices that merely satisfy the symmetric.I. where D = diag(A. most "diagonal" we can get is the JCF described in Chapter 9. where D = diag(AJ.. . and unitary matrices (and their "real" counterparts: symmetric. . Then there AI. . and unitary matrices (and their "real" counterparts: symmetric. . Normal matrices include Hermitian." The transformation A M» PAQ is called an equivalence. . A. If W = V and if Q = PT is orthogonal. Find bases in V and W with respect to which Mat A has a "simple form" or "canonical Find bases in V and W with respect to which Mat A has a "simple form" or "canonical xm form.:xm and Q e Rnnxn such that PAQ has a "canonical form. AAHH = AH A). skewskew-Hermitian.." In matrix terms. If The following results are typical of what can be achieved under a unitary similarity. the transformation A f--+ PAP-I is called aasimilarity. We can also consider the case A E emxn and unitary equivalence if P and <2 are unitary. . the definition. Remark 10.e.

.. k = For simplicity. . ..3 called Theorem 10.2)-block X2 .. Write V H matrix such that V X I = [ ~]. . We also get 0 in the (2. XH AX induction noting that XH AX is Hermitian. Thus. 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.. we consider the real case.. An... . Then there exist n . Xk H Hk.T. 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. Then [XI V 2] is unitary.. X 1 XI e E". ..... .l)-block. Xk].... Construct a sequence of Householder matrices (also known Proof: Let X [XI.1) we have used the fact that AXI = k\x\. D 0 (2. X = Given a unit vector x\ E JRn.• • Hk and Hk'" HI...I)-block..2) Al X~AX2 XfAX 2 0 Al ] 0 XfAX z 0 l In (10.2).. . Xn such that [x\.Hv. simplicity. the construction of X2 E JRnx(n-l) such that X — z e ]R" (". When combined with the fact that In (l0. I)-block x"xi = 1.2 is then a special case of Theorem 10.3. . xn] = [x\ ] [XI X22] is unitary. following general result. and normalize it such that x~ XI = XI 1.. (l.. . .96 96 Chapter 10.2)-block must have eigenvalues A2. Canonical Forms Proof: eigenvector corresponding AI. (/ € k) U2 X i U2 = Xi . . Let X\ e Cnxk have orthonormal columns and suppose U is a unitary Theorem 10. [XI U2] is unitary... . Construct a sequence of Householder matrices (also known HI. Write UH = [U\ U ] [VI Vz] 0 2 with Ui E Cnxk . Now XHAX =[ xH I XH ] A [XI 2 X 2] =[ =[ =[ x~Axl XfAxl X~AX2 XfAX 2 ] (10...k rows of the unitary matrix U. xn] = 1.k But the latter are orthonormal since they are the last n . 10. Then there exist n — 1 additional vectors X2... we get A-i remaining in the (l. [£i. Let XI E Cnxk have orthonormal columns and suppose V is a unitary matrix such that UX\ = \ 1. ~nf.3 for k = 1. n . .1) (10. %n] XI .) [XI X2]] is orthogonal is frequently required. xn such that X = (XI..1 additional vectors x2.. Let the unit vector x\ be denoted by [~I. Hk as elementary reflectors) H\. The proof is completed easily by induction upon noting proof that the (2. When combined with the fact that x~ XI = 1. We illustrate the construction of the necessary Householder matrix for k — 1. Let V = XI. Canonical Forms Chapter 10. xf*x\ = Proof' Let x\ be a right eigenvector corresponding to X\. [Xi f/2] unitary..2)-block noting that x\ is orthogonal to all vectors in X2..xd.. where R E kxk is upper triangular. . . orthogonal (l. D The construction called for in Theorem 10. In (10. xd = [ ~ l U = where R is upper triangular (and nonsingular since x\. . VI € Cnxk [Xi U ] Proof: Let X\I = [x\..1) we have used the fact that Ax\ = AIXI. we get 0 in the (l. Hk in the usual way (see below) such that Hk . 0 Thus. ... . HdxI. where R € Ckxk is upper triangular. . A. Then VH = / / . . k 1. . —k U.. ....2)-block by XI Xz.l)-block by Al (2.3... . . Xk are orthonormal).

In fact.. • • . '. Further details on Householder matrices.1 ± 1.•» '. . sponding to the Ai'S).. i=1 (10.. To see that U effects the U symmetric U U = U = I. Then there exists an 10. . where Pi = PR(x. n A = LAiPi. where u -^UU [t-\ 1. (onto the one-dimensional eigenspaces correPi one-dimensional eigenspaces sponding to the A.1 and UT X\ = 1 ± £1. so U is orthogonal. A Note that Theorem 10. = PUM = xixf = xxixT since xj xi — 1.e. [25]. . it is easily verified that u T u = ± 2£i and u T Xl = 1 ± '.It can easily be checked — 2uu+ — u u T . where u = ['. Some Basic Canonical Forms 10. Thus.. X n ).+uu T . An. [23].2uu+ = I . U orthogonal. Theorem 10. i.Xn.3) spectral which is often called the spectral representation of A. £«] r It can checked T 2 that U is symmetric and U TU = U 2 = I. necessary compression of Xl.1. .i. i=l theoretical The following pair of theorems form the theoretical foundation of the double-Francisdouble-FrancisQR algorithm used to compute matrix eigenvalues in a numerically stable and reliable way. .. U effects necessary compression of jci. An). quently in applications. [11].2 for Hermitian matrices) can be written from Theorem 10... [23]... . .2 is worth stating separately since it is applied fre10.4.1.3) is actually a often weighted sum of orthogonal projections P. .. The real version of Theorem 10. 's). including the choice of sign and the complex case. £2.nf.4.. it is easily verified that UT U = 2 ± 2'. . x where P. consulted standard numerical linear algebra can be consulted in standard numerical linear algebra texts such as [7]. 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.1. . [11].2 is worth stating separately since it is applied frequently in applications.2 for Hermitian matrices) can be written n A = XDX T = LAiXiXT.4 implies that a symmetric matrix A (with the obvious analogue from Theorem 10.10. Then there exists an AT E jRnxn have eigenvalues AI. A in (10. Let A = AT e E nxn have eigenvalues k\.2. [7].) — xiXt = i j since xT Xi = 1. . XTAX = D = diag(Xi. [25].e. 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 ..

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 when we can go further and reduce a matrix via unitary similarity to diagonal form.6 is called a real Schur canonical form or real Schur form (RSF). A matrix A E C"x" is unitarily similar to a diagonal matrix if and only if Theorem 10. what is true. so A is normal.5 is called a Schur canonical form or Schur form. the next theorem shows that every to place such eigenvalues on the diagonal of T. where D is diagonal.2 except that Proof: The proof of this theorem is essentially the same as that of Theorem 10. where T is upper triangular. diagonal of T (or S). but if A has a complex conjugate pair of eigenvalues. Then there exists an orthogonal 10. it is thus unitarily similar to an upper triangular matrix. [17]).98 98 Chapter 10. Let A e C"x". matrix U such that U AU = S. Theorem 10. the next theorem shows that every A E IR xn is also orthogonally similar (i. Canonical Forms Theorem 10.9. A quasi-upper-triangular matrix is block upper triangular with 1 x 1 diagonal matrix. then complex arithmetic is clearly needed if A has a complex conjugate pair of eigenvalues. then complex arithmetic is clearly needed to place such eigenValues on the diagonal of T. real arithmetic) to a quasi-upper-triangular A e Wnxn is also orthogonally similar (i.6 (Murnaghan-Wintner). D ur In the case of A e IRn ". where S is quasi-upper-triangular. it is of interest to know when we can go further and reduce a matrix via unitary similarity to diagonal form. it is of interest to know While every matrix can be reduced to Schur form (or RSF).2 except that in this case (using the notation U rather than X) the (l. real arithmetic) to a quasi-upper-triangular matrix.. following theorem answers this question. However. However.. The quasi-upper-triangular matrix S in Theorem 10. Theorem 10. Then there exists an orthogonal Let A e IR n ". Then there exists a unitary matrix U such that U H AU = T. Then Proof: Suppose U is a unitary matrix such that U H AU = D. is that the first Schur vectors span the same all applications (see. The triangular matrix T in Theorem 10.5 (Schur). The matrix 10.6 is called a real form or Schur fonn. AHA = AA H). it is thus unitarily similar to an upper triangular matrix. .5 is called a Schur canonical Definition 10. Example 10. complex conjugate pairs of eigenvalues. . for example. However.e. The columns of a unitary [orthogonal} Schur canonical form or real Schur fonn (RSF). AH A = AAH ).8.5 (Schur). what is true. Proof: The proof of this theorem is essentially the same as that of Theorem lO.e.2)-block wf AU2 is not 0. The following theorem answers this question.e. but In the case of A E R"xxn .8. Let A E R"xxn.6 T T matrix U such that U AU = S. Let A E cnxn Then there exists a unitary matrix U such that Theorem 10. for example. The triangular matrix T in Theorem 10. and sufficient for virtually is real if U is orthogonal) is an eigenvector. The columns of a unitary [orthogonal] matrix U that reduces a matrix to [real} Schur fonn are called Schur vectors. 0 in this case (using the notation U rather than X) the (l. While every matrix can be reduced to Schur form (or RSF).. Proof: Suppose U is a unitary matrix such that U H AU = D. Its real JCF is is in RSF. Canonical Forms Chapter 10. Definition 10. is that the first k Schur vectors span the same Ainvariant subspace as the eigenvectors corresponding to the first eigenvalues along the invariant subspace as the eigenvectors corresponding to the first k eigenvalues along the diagonal of T (or S). [17]). A quasi-upper-triangular 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.2)-block AU2 is not O.7. However. matrix U that reduces a matrix to [real] Schur form are called Schur vectors.9. The matrix s~ [ -2 0 -2 5 4 0 is in RSF.e. A matrix A e c nxn is unitarily similar to a diagonal matrix if and only if A is normal (i. where S is quasi-upper-triangular. UH AU = T. Then AAH = U VUHU VHU H = U DDHU H == U DH DU H == AH A so A is normal. where D is diagonal. where T is upper triangular.. A is normal (i. and sufficient for virtually all applications (see.7. The quasi-upper-triangular matrix S in Theorem 10.

. Then T (Theorem It is then a routine exercise to show that T must. this section that may be stated in the real case for simplicity. Proof: Proof: Let U be a unitary matrix that diagonalizes A as in Theorem 10.=1 But clearly n LA. nonzero x E lR. Let A = AH e enxn with eigenvalues X{ :::: A2 :::: . A symmetric matrix A e Wxn 1. Then n x HAx = (U HX)H U H AU(U Hx) = yH Dy = LA. Similarly.n • We write A :::: 0.A ~ O. We (or negative if— A nonnegative definite.10. we write A :::: B if and only ifA — B>QorB — A < 0.12 ~ AlyH Y = AIX HX .nxn is Definition 10. it is said to be indefinite.10..• :::: An. if A and B are symmetric matrices. Remark 10.2 10.. nonpositive definite (or negative semidefinite) if -A is nonnegative definite. we write A > B if and only if A — B > B .B :::: 0 or B . negative definite if .11.. this is generally true for all results in the remainder of of superscript //s Ts. positive definite if and only if xTT Ax > 0 for all nonzero x G lR. We write A > O. CM j]i. Indeed.12. Remark 10. in fact. If a matrix is neither definite nor semidefinite.2. we write A > B if and only if A . x eC". If neither semidefinite.10.A is positive definite. we write A > B if and only if A . 11'/. We write A < 0. 3. We write A ~ 0. If A E C"x" is Hermitian.2. 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. Definite Matrices 99 Conversely. where T is an upper triangular matrix (Theorem 10. 111. Then for all E en. Thenfor all Let A = AH E Cnxn with eigenvalues AI > A2 > • • > An. A symmetric matrix A E lR. positive definite if and only ifx Ax > Qfor all nonzero x E W1 We write A > O.13.=1 . B — A < 0. 2. let y = U H x. U diagonalizes A 10. . all the above definitions hold except that A e nxn Remark 10. Furthermore. Definite Matrices 10. Similarly. where x is an arbitrary vector in en. if—A 4. negative positive definite. It T 0 D 10. superscript H s replace T s. if A and B are symmetric matrices. be diagonal..11. We write A < O. suppose A is normal and let U be a unitary matrix such that U H AU = T. Furthermore. and denote the components of y by v UHx. e Theorem 10.2. Then 11.n.2 Definite Matrices Definite Matrices Definition 10. Also.5). write A < O. i En. A U U HA U T.A < O.12.2. We write A > 0. i € n.B > 0 or or Also. indefinite.

Theorem 10. 3. Theorem 1O. XHAx > 0 for all nonzero = AH E enxn E en. from which the theorem follows. of positive.. If A = AH e C"x" is positive definite. 2. Theorem 10. The determinant of the I x leading submatrix is 0 and consider the matrix A = [~ 2x 0 (cf. I Proof: E C" Proof: For all x € en we have Let x be an eigenvector corresponding to Amax (A HA).18.17. determinant the determinant of the 2x2 2 leading submatrix is also 0 (cf. Remark 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.19.18. x E C". A symmetric matrix A € R"x" is nonnegative definite if and only if any of following equivalent of the following three equivalent conditions hold: 1.19. A can be wrirren in [he/orm MT M. Canonical Forms and and n LAillJilZ::: i=l AnyHy = An xHx . Then ^pjp2 = ^^(A" HA). All eigenvalues of A are positive. Note that the determinants of all principal "ubm!ltriC[!!l mu"t bB nonnBgmivB R. where M E IRb<n and k ~ ranlc(A) — ranlc(M).1. A can be written in the form MT M.16. The determinants of all principal submatrices of A are nonnegative. Then 111~~1~2 Let jc be an eigenvector corresponding to Xmax(AHA). The determinant of the 1x1 1 leading submatrix is 0 and 1. of all principal submatrices of 2.soO < X n < ••• < A. 0 D Remark XHHAx Remark 10.15.I. consider the matrix A — [0 _l~]. whence IIAxll2 ! H IIAliz = max . For example. Let A E enxn Then \\A\\2 = Ar1ax(AH A).17). Let A e C"x". Corollary Corollary 10. ::::: AI. the .@mllrk 10.14. All eigenvalues of A are nonnegative. 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.l3 provides upper (AO and lower (An) bounds for (A. xfO IIxll2 I 0 Definition submatrixofan n x -k) x -k) Definition 10. Theorem 10.= Amax{A A).13 provides (A 1) Rayleigh quotient of jc. of obtained and E ~nxn positive definite if and only if any of the Theorem 10. where M e R"x" is nonsingular. 2. Then IIAII2 = ^m(AH A}. where M 6 R ix " and k > rank(A) "" rank(M).100 100 Chapter 10.. whence Ar1ax (A A). 3. so 0 An ::::: . form MT E ~n xn E ~n xn definite if and only if Theorem 10. However. not just those of the leading principal submatrices. Not@th!ltthl!dl!termin!lntl:ofnllprincip!ll eubmatrioes muet bQ nonnogativo in Theorem 10.17). A leading principal submatrix of order n — k is obtained by deleting the last k rows and columns. Theorem 10. Canonical Forms Chapter 10.. All eigenvalues of A are positive. The determinants of all leading principal submatrices of A are positive. All eigenvalues of A are nonnegaTive.1. 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. A can be written in the form MT M. .w) x HAx > the Rayleigh quotient.18. 3.

concerns the notion of the "square root" of a matrix. in fact.3 is not unique. For example. basic definitions.nxn . BM. nxm 2.20. Write the matrix A in the form the form By our induction hypothesis. if A E lR. any matrix S of c e s 9 the " °* ™ the form [ ssinOe _ ccosOe ] IS a square root. 0 Recall that A :::: B if the matrix A .. For example. p.10. Write the matrix A in Proof: The proof is by induction. Then there exists a positive definite. B e Rnxn be symmetric. A e R"x be nonnegative definite. For example.2.we say 181]). Ll E C1-""1^""^ and . That is. .21. The following theorem is useful in "comparing" symmetric matrices.20. if A = lz.. [16. p. The following Recall that A > B if the matrix A — B is nonnegative definite. for example. Then A has unique nonnegative Theorem 10. rankS = rankA definite definite if positive definite). In general.3 is not unique. Definite Matrices 101 101 principal submatrix consisting of the (2. where L\ e c(n-l)x(n-l) is nonsingular and lower triangular as = L1Lf.17 is available and is A stronger form of the third characterization in Theorem 10.22. with positive diagonal elements such that positive Proof: The proof is by induction. then MT AM > MT TBM. MT AM> M. assume the result is true for matrices of order n . Then A has aaunique nonnegative definite square root S. [ fz -ti o o l [~ 0] ~ 0 v'3 ..nxn"be nonnegative definite. If >BandMe jRnxm.23. The factor M in Theorem 10. standard theorem stated 181]). j proof (see. It is stated and proved below for the more general known as the Cholesky factorization.2) element is. Let A e c nxn be Hermitian unique nonsingular lower triangular matrix L nonsingular A = LLH. if € E" xn we say that e jRn x that S E R nxn"isisa asquare root of AA ifS2 2 =— A. negative and is nonpositive definite.18. Theorem 10. 10rm [COSO _ Sino] . For example. Hermitian case.is a square root. in fact. and positive definite. 1. E jRnxn MT AM > M BM. It is stated and proved below for the more general Hermitian case. Definite Matrices 10.B is nonnegative definite. then MT AM :::: MTTBM. 1f A :::: Band M E Rnxm. It concerns the notion of the "square root" of a matrix. Let A E lR. Its proof is straightforward from basic definitions. The factor M in Theorem 10. The case n = 1 is trivially true. A stronger form of the third characterization in Theorem 10. That is. if then M can be then M can be [1 0]. In general. Moreover. nxn Theorem 10. E <C Theorem 10.17 is available and is known as the Cholesky factorization. assume the result is true for matrices of order — 1 so that B may be written as B = L\L^. Its proof is straightforward from theorem is useful in "comparing" symmetric matrices. The following standard theorem is stated without proof (see. Let A. = LLH. SA = AS and rankS = rank A (and hence S is positive = AS S S. 2. for example.18.1 so that B By our induction hypothesis. negative and A is nonpositive principal submatrix consisting of the (2. [16. If A > B and M e Rm . any matrix of nonsymmetric) have infinitely many square roots. If A> Band E jR~xm. if Remark 10. Theorem 10.23. Remark 10. The case = is trivially true.2.2) element is. definite if A is positive definite). if = /2.22. matrices (both symmetric and nonsymmetric) have infinitely many square roots. matrices (both symmetric and square root of if S A.

Ch.4) Proof: proof Proof: A classical proof can be consulted in. 0 Note that the greater freedom afforded by the equivalence transformation of Theorem afforded 10. [21. yields a far "simpler" canonical form (10. Take P =[ S~ 'f [I ] and Q = V to complete the proof. the unitary equivunitary alence known as the SVD.3 Equivalence Transformations and Congruence Equivalence Transformations and Congruence Theorem 10.2) in its complex version.4) [7. Many similar results are also (10. [4. Alternatively.b H B-1b > O. Gaussian or elementary row and column operations. Ch. Choosing a be det(fi) > HB~lb completes the proof. . Then there exist matrices P E C: xm and Q e C"nx" such E c. ann Since det(B) > 0.xn.102 102 Chapter 10.p. Performing the indicated matrix multiplication and equating the corresponding submatrices. available.4). p. multiplication where a is positive. as opposed to the more restrictive situation of a similarity transformation. 5]. But we = ann — b LIH L\lb = ann — bH B~lb B A). They are more stably computable than (lOA) and more efficiently computable than a full SVD. Substituting in the involving we find 2 a2 = ann . Choosing a to be the positive square ann .• Clearly we see we L I C = b and ann = c HC a 2 c is given simply by c = C. Alternatively. Let A € C™*71. But know that o < det(A) = det [ ~ b ] = det(B) det(a nn _ b H B-1b). Then [ S-l o 0 ] [ I Uf U H ] AV = [I 0 0 ] 0 . [21. Two such forms are stated here. of course. Substituting in the expression involving a. we must have ann —bHB lb > 0. are generally unreliable.4) efficiently available. However.4) and the SVD.3 10. for example (10.b H B-1b (= the Schur complement of B in A). Canonical Forms with positive diagonal elements. the SVD is relatively expensive to compute and other canonical forms exist that are intermediate between (l0. 131]. say.b B-1b completes D 10. 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. we see that we must have L\c = b and ann = CHc + a 2. Then E c~xn such exist e C™ x m that that PAQ=[~ ~l (l0. Canonical Forms Chapter 10.24..131]. numerical procedures for computing such procedures an equivalence directly via. suppose A has an SVD of the form (5. we find by L^b.24. of ann — b 0 root of «„„ . 2].b HL\H L11b = ann . The numerically preferred equivalence is. see. However.lb. for example.

HE C xn E e~ xn. In(A) = In(X AX). where R E Crrxr is upper triangular and S e C rx( " r) is arbitrary but in general nonzero. 0 2. v. 0.10. Example 10. and £ denote the numbers of positive. n The signature of A is given by sig(A) = n . v. then rank(A) = n + v. Then the inertia of A is the triple of inertia of of negative.30. 0 D x Theorem 10. see [4] for details. see [4].31 guarantees that rank and signature of a a matrixare preserved under Theorem 10. Note that congruence preserves the property of being Hermitian. When A has full column rank but is "near" a rank deficient matrix. . Then there exists a unitary matrix Q e e mxm and a Theorem 10.rrxr is upper (or lower) triangular with positive diagonal elements. i. The H. It is of interest to ask what other properties of a matrix are then X H AX is also Hermitian. It is of interest to ask what other properties of a matrix are preserved under congruence.31 guarantees that rank and signature of matrix are preserved under congruence. Theorem 10. Then there exist unitary matrices U e Cmxm and V E Cnxn such that unitary matrices U E e mxm and V e e nxn such that (10. v.In[! 1o o o 0 0 00] -10 =(2. Let A e e~xn. [21. Proof: For the proof. In(A) 3.e. see [4]. (TT.27. Let A E C™ x ". Again. i. D Proof: For the proof. if A is Hermitian. Let A = AH e C"x" and let 7t. We then have the following. Note that a congruence is a similarity if and only ifX is unitary.28. of A. then XH AX is also Hermitian.28. Proof: For the proof. 134]. Equivalence Transformations and Congruence 103 103 Theorem 10. and ~ denote the numbers of positive. Definition 10. a congruence.xr E erx(n-r) arbitrary general nonzero. Let A e C™ ". see. respectively. Again. £).. see. upper Proof: For the proof. D Theorem 10.26. n. respectively.30. It turns out that the principal property so preserved is the sign preserved under congruence.0.xr is upper (or lower) triangular with positive diagonal elements.1). Equivalence Transformations and Congruence 10. see [4]. [21. . and zero eigenvalues.27.XH AX Definition 10. The signature of is Example 10. nxn E e X E e~xn. then rank(A) rr v. 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. congruence. If A AH e e x " then A> 0 if and only if In(A) = (n. l.v. phenomena at a cost considerably less than a full SVD. v.6) E e. It turns out that the principal property so preserved is the sign of each eigenvalue. Let A = AH E e nxn and let rr. Let A E e~xn.e.25 (Complete Orthogonal Decomposition). Then H HAX). numbers In(A) (n.3. v. p. sig(A) = rr — v. for example. Definition 10.t h e n A > 0 if and only if In (A) = (n. for example. of each eigenvalue. If A = A" E C nnxn. see [4]. if A is Hermitian. 134]. Definition 10.0). and eigenvalues. Note that congruence preserves the property of being Hermitian.25 (Complete Orthogonal Decomposition).3. where R e €. When A has full column rank but is "near" a rank deficient matrix.29. 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.31 (Sylvester's Law of Inertia). Note that a congruence is a similarity if and only if X is unitary. Theorem 10.31 (Sylvester's Law of Inertia). The transformation A i-> XH AX is called a congruence. Remark 10. Then is the numbers In(A) = (rr. Proof: For the proof. p. D 0 Remark 10.5) where R E e. negative. of A. v. If In(A) = (rr. see [4] for details. 0). In(A) = ln(X Proof: For the proof. 2..29. Let A e Cnxn and X e Cnnxn. We then have the following. Let A = A He ennxn and X e Cnnxn.26. £).1. Then there exist Theorem 10.

Canonical Forms Chapter 10.1 Block matrices and definiteness Theorem 10. . the number 0/0 's is (... .. 1/. . Define the nn x n matrix U UH AV = diag(Ai. 's is 7i.104 104 Chapter 10. By Theorem 10.0). the number of — 's is v. if and if either A> and D . -1. and D ..f-Arr+v. Suppose A = AT and D = DT.. . and the numberofO's is~. . v. Define the x n matrix vv = diag(I/~.. ..1 10. . ... Theorem 10.BT A+B:::: o. A w ). Then Remark Remark 10. Note the symmetric Schur complements of A (or D) in the theorem.. .2 there exists a unitary matrix V such that VHAU = diag(AI.... . left AT D DT. 0 D 10. An of Jr Proof: Let AI . AA+B = B. and the final £ are 0. where the number of E c~xn XH AX = diag(1.4 Rational Canonical Form Rational Canonical Form rational One final canonical form to be mentioned is the rational canonical form. ..BT A~l B > 0.BT A+B > 0. D > and . . . Canonical Forms Theorem 10. AA+B = B. Suppose A = AT and D = DT. .. Then = AT D = DT. Then there exists a matrix AH C"xn In(A) = (Jr..34. X UW desired 10. .I BT > O. -1.BD^BT > 0.4 10.....3. 0).3. where the number of X 1's is Jr. the number of -Il's is v.f-Arr+I' . £). O. .32. -I. v. . . if ifA>0.33. . . 1. Let A = AHeE cnxn with In(A) = (jt. I. 0... B D ] > - ° if and only if A:::: 0. X e C"nxn such that XHAX = diag(l. and D . . . Proof: proof Proof: The proof follows by considering. Proof: Consider the congruence with Proof: Consider proof Theorem and proceed as in the proof of Theorem 10. ifand only ifeither A > 0 and D . .BD. . 0 D Then it is easy to check that X = V VV yields the desired result.. I/.BT A-I > 0. Xw denote the eigenvalues of A and order them such that the first TTare ~ O.. Theorem positive.35. -1. or D > 0 and A . I/~. Proof: AI. I. . An).0. the congruence B ] [I D ~ 0 _A-I B I ° JT [ A BT ~ ][ ~ 0 D The details are straightforward and are left to the reader.. .33. the next v are negative. for example.1).. .

since a matrix is similar to its transpose (see exercise 13 in Chapter 9). consider the companion matrix illustrate. if its Jordan canonical form and characteristic polynomial are the same or.18).11) .4.Then it can be shown (see [12]) that A mial is 7r(A) = A" . : ~ ! ~01]. Using the reverse-order identity similarity P given by (9. Notice that in all cases a companion matrix is nonsingular if and only if aO i= O. has only one block associated with each distinct eigenvalue.37.10.7) is called a cornpanion rnatrix or Definition 10.(ao + «A + . equivalently. Using the reverse-order This matrix is a special case of a matrix in lower Hessenberg form.7) Definition 10. In fact.4. 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 ~'). if its Jordan canonical form has only one block associated with each distinct eigenvalue..18). o (10. the Moreover. To illustrate. consider the companion matrix (l0. 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. For In fact. A matrix A E lRn Xn is said to be nonderogatory ifits minimal polynomial if its minimal polynomial and characteristic polynomial are the same or. To Companion matrices also appear in the literature in several equivalent forms. the inverse of a nonsingular companion matrix is again in companion form.10) o 1 o 1 o o o o o o (10. + an_IAn-I). l 0 0 -~ ao -~ ao _!!l (10. For £*Yamr\1j=» example.7) is called a companion matrix or is said to be in companion forrn. Rational Canonical Form 105 105 Definition A matrix A e M"x" is said to be Definition 10. since a matrix is similar to its transpose (see exercise 13 in Chapter 9). the inverse of a nonsingular companion matrix is again in companion form. is said to be in cornpanion form. 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 . Rational Canonical Form 10. A matrix A E lRnxn of the form (10. A matrix A e E nx " of the form (10.9) Moreover. A is easily seen to be similar to the following matrix identity similarity P given by (9.37..8) This matrix is a special case of a matrix in lower Hessenberg form.36. 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. equivalently.

then it is not similar to a companion matrix of the form (10. an-If and let e M"" \a\. the largest and smallest singular values can also be written in the equivalent form Remark 10. = ~ (y .. companion an arbitrary matrix to companion form are numerically unstable. Algorithms to reduce an arbitrary matrix to companion form are numerically unstable. Explicit formulas for all the associated right and left singular vectors can Remark 10. n .. + a. 3.39. 02. Such matrices are said to be in rational canonical form Frobenius rational canonical form (or Frobenius canonical form). form). If a companion matrix of the form (10. Moreover. see [14]. For details. if ao = 0. then it is not similar to a companion matrix of the form (10._1 and y = 1 + + a. matrix is not a companion matrix unless a = O. and so forth [14]. Then A in (10. among which. Then + ai + . 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. Ifao ^ 0. associated at least one eigenvalue. and perhaps surprisingly. each of whose diagonal blocks is a companion matrix. . Let al ~ GI ~ • • ~ an be the singular values of the companion matrix Theorem 10. Moreover. [12]. for example. For example.7) is singular. Let a = a\ + a\ + • • • + a%_{ and y = 1 + «.7). Such matrices are said to be in each of whose diagonal blocks is a companion matrix.4aJ) .38..106 Chapter 10. among which. I — T = T) Note that / . it can be shown that a derogatory matrix is similar to a block diagonal matrix.• > an be the singular values of the companion matrix A in (10.caa T = (I + aaT) -I . and hence the pseudoinverse of a singular companion + matrix is not a companion matrix unless a = 0... if ao = 1 inverse can still be computed. stable ones are nearly unstable. at least one eigenvalue. see haps surprisingly. is the fact that their singular values can be found in closed form.. . Explicit formulas for all the associated right and left singular vectors can also be derived easily. i.caa T ca o J. is the fact that their singular values can be found in closed form. If A E R nx " is derogatory. in n general and especially as n increases. with a similar result for companion matrices of the form (10. see.10).1. Theorem 10.38. However.10).39. . the largest and smallest singular values can also be written in the equivalent form If ao =1= 0. their eigenstructure is extremely ill conditioned. then its pseudoIf singular. companion matrices are known to possess many undesirable numerical properties. For example.e..4ao ' 1 2) - a? = 1 for i = 2.7). . Leta = ar aJ al 2_ 2 ( y + Jy 2. has more than one Jordan block associated with If A € JRnxn derogatory. in matrices are known to possess many undesirable numerical properties. ... stable ones are nearly unstable. Let a\ > a2 > . 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. nonsingular ones are nearly singular.. Algorithms to reduce but unfortunately they are often very difficult to work with numerically. a. Canonical Forms with a similar result for companion matrices of the form (10. a2.. Companion matrices have many other interesting properties. also be derived easily.. Let a E JRn-1 denote the vector [ai.. Canonical Forms Chapter 10.Jy2 ..7). and perCompanion matrices have many other interesting properties.7). i. a n -i] and l c I+~T a.Q + a.e. especially nonsingular ones are nearly singular. and so forth [14]..

Let A € C n xn be normal with eigenvalues y1 . (A) = IA. 9. show that A-I must also be positive definite. In the 2-norm. Find a unitary matrix U such that [~ M CC x 2 Find a unitary matrix U such that 6.. • • > on > 0. A [ must also be positive 7. then p(A) = IIAII2' Show that the converse is true if n = 2.. For example.. Show that a. 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.-(A)| for e n.. Theorem 10. A E cc nxn peA) = max). . (A) |A. An and singular 0'1 > 0'2 ~ 4. Show that if a triangular matrix is normal. If A e jRn xn 8. EXERCISES EXERCISES 1. Let A G Cnx" and define p(A) = maxx€A(A) IAI.5 to find a unitary matrix Q that reduces A e C"x" to lower triangular form.. 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 . Is [ ^ A E jRnxn is definite.• ~ an ~ O. K\ (A) (10.. then K2(A) ^ T~I. 5.18) U A E cc nxn Theorem 10. then it must be diagonal. Suppose A e E"x" is positive definite.. 1. can be determined explicitly as determined explicitly y+J y 2 . by the theorem. 6. E jRnxn be symmetric.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. . It is easy to show that y/2/ao < K2(A) < -£-. Remark 10. 3. this condition number is the ratio of largest to smallest singular values which.40. when solving linear equations numerical sensitivity Kp(A) = systems of equations of the form (6. In the 2-norm. say 0(10*).EA(A) I'MpeA) 3. Show that the converse radius of A.11). Show that [ * }. Let R. Let R..18) and the matrix U in identity in (9. Show that [~ R > S-I.38 yields some understanding of why difficult numerical Remark 10.2). is true if n = 2. then peA) = ||A||2. Theorem 10. Use the reverse-order identity matrix P introduced in (9. Show that if A is normal. this condition number is the ratio of largest to smallest singular precision. If this number is large. Then p(A) is called the spectral radius of A. Prove that if A e M"x" is normal. yn and singular values a\ ~ a2 > . If this number is large. Let A 7.4a5 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). when solving linear behavior might be expected for companion matrices. 2. Show that a. If A E Wxn is positive definite. A E en x n eigenvalues A].(A)I for ii E!l. A E jRnxn N(A) = A/"(A ). one measure of numerical sensitivity is KP(A) = A A -] > the so-called condition number of A with respect to inversion and with respect II ^ IIpp II A~l IIpp'me so-called condition number of A with respect to inversion and with respect to the matrix p-norm. Let A = I J : ]eEC 22x2. say O(lO k ). and when GO is small or y is large (or both).38 yields some understanding of why difficult numerical behavior might be expected for companion matrices...40. one may lose up to k digits of to the matrix P-norm. one may lose up to k digits of precision. Show that if A is normal.11).. For example. S 6 E nxn be symmetric. It is not unusual for y to be large for large n. then Af(A) = N(A Tr ). .

1 1. (a) Chapter 10. Canonical Forms Chapter 10. .j 1+ j ] -1 . Find the inertia of the following matrices: following 10.j 1+ j ] -2 ' (d) [ .108 108 10. Canonical Forms [~ ~ l (b) [ -2 1.

1 and linearity of the transpose. We restrict our attention in this chapter only to the so-called time-invariant case. For all A JR.1 Differential Equations Differential Equations = Ax(t). Definition 11.2) k=O The series (11. Forall A EG R" XM .1.3) which thus also converges for all A and uniformly in t.Chapter 11 Chapter 11 Linear Differential and Linear Differential and Difference Equations Difference Equations 11. unique. The solution of (11. where the matrix A E Rnxn is constant chapter only to the so-called time-invariant case.1 11. It can be described conveniently in terms of the matrix exponential.nxn.1) for t 2: to. Proof This follows immediately from Definition 11. where the matrix A e JR. Proof: This follows immediately from Definition 11. T T 109 109 .1) is then known always to exist and be and does not depend on t. This is known as an initial-value problem. For all A e Rnxn.2) can be shown to converge for all A (has radius of convergence equal The series (11. The solution of (11.1. Proof: This follows immediately from Definition 11.1 11.1) is then known always to exist and be unique. = Xo In this section we study solutions of the linear homogeneous system of differential equations In this section we study solutions of the linear homogeneous system of differential equations x(t) x(to) E JR.1.1 Properties of the matrix exponential Properties of the matrix exponential 1.1) involves the matrix to +(0). eO = I. 11.1 by setting = 0. The solution of (11.2) can be shown to converge for all A (has radius of convergence equal to +00).n (11. For all A E JR.nxn is defined by Definition 11. the matrix exponential e A e JR. (e(eAf = -e A e^. Proof This follows immediately from Definition 11. The solution of (11.Ak.1.1) involves the matrix (11. This is known as an initial-value problem.nxn. 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 . We restrict our attention in this for t > IQ.1 and linearity of the transpose. k. A) • 2.1 by setting AA =O. It can be described conveniently in terms of the matrix exponential. (11. e° = I.nxn is constant and does not depend on t.

T E JR.A)-I} = «M. on (t + T)*.e. 4. all A € R"x" and for all t € lR. AB = BA. For all A E JRnxn and for all t. Linear Differential and Difference Equations Chapter 11. Proof" We prove only (a).. Proof: Simply take T = -t in property 3.tA . (a) C{etA = (sI-Arl. B E R" xn and for all t E JR. Proof' Note that Proof: Note that et(A+B) = I t + teA + B) + -(A + B)2 + . 5. ) . r e R.e. Then for E JRnxn t E R. i.110 110 Chapter 11..1 {(j/-A). Then for 6..A)-I. = I + (t + T)A + (t + T)2 A 2 + .l{e tA}} = (sI .. Let £ denote the Laplace transform and £-1 the inverse Laplace transform. Compare like powers of t in the first equation and the second or third and use the Compare like powers of t in the first equation and the second or third and use the k binomial theorem on (A + B/ and the commutativity of A and B. binomial theorem on (A B) and the commutativity of A and B.. (a) .. and B commute. For all A. Proof" Simply take T = — t in property 3.. AB = B A. 6.. ) ( I + T A + T2!2 A 2 +. 2 2! and and while while e e tB tA = ( 1+ tB t2 2 + 2iB 2 +. ForaH A E R" x " and for all t e JR.... {+oo = io et(-sl)e tA dt since A and (-sf) commute =io (+oo ef(A-sl) dt . Proof: We prove only (a).1 } = erA.. (etA)-1 = e.. ) . Part (b) follows similarly. (e'A)~l e~'A. Proof" Note that Proof: Note that e(t+r)A = etA erA = erAe tA . ) ( 1+ tA + t2!A 2 +. For all e JRnxn and for all E R. Compare like powers of A in the above two equations and use the binomial theorem Compare like powers of A in the above two equations and use the binomial theorem on(t+T)k. (b) . 2! and and e e tA rA 2 = ( I + t A + t2! A 2 +. et(A+B) =^e'Ae'B = e'Be'A and and B commute. et(A+B) =-etAe tB = etBe tA if and only if A all e JRnxn and all e R.l-I{(sl. (b) £.. Part (b) follows similarly. = e'A erA = elAe'A . For all e R"x" and for all t. Let denote the Laplace transform and £~! the inverse Laplace transform. i. Linear Differential and Difference Equations e(t+r)A e(t+T)A 3.

etA . For any consistent matrix norm.=1 = ~[fo+oo e(Ai-S)t dt]x. employed I e(t+~t)AAt.. it can be differentiated term-byterm from which the result follows immediately. Notice in the proof that we have assumed. s ..X i y. e'A Proof: Since the series (11.1. Notice in the proof that we have assumed.H using the JCF.. £(e'A) 7. All succeeding steps in the proof then follow in straightforward way. ) = L'lt IIA 21111e tA IIe~tIIAII... the formal definition d dt _(/A) = lim ~t-+O e(t+M)A _ etA L'lt can be employed as follows.. for convenience. for convenience. that A is diagonalizable.H = '"' assuming Re s > Re Ai for i E !! = (sI .. .H L. For all A E JRnxn and for all E JR. Differential Equations 11.A) ~' is called the resolvent of A and is defined for all s not in A (A).A)-I. using the JCF. the scalar dyadic decomposition can be replaced by If this is not the case. The matrix (s I . A2 + (~~)2 A tA II tA Il 1 (_ 2! + . ) 3 II I ( ~.. 1h(e tA ) = AetA = etA A.3) is uniformly convergent.l)e . If this is not the case. ..Ae II = I L'lt (M)2 + ~ A 2 +.y. that A is diagonalizable. For all A e R"x" and for all t e R. the scalar dyadic decomposition can be replaced by et(A-sl) =L .3) is uniformly convergent.A"I i=1 . Differential Equations 111 111 = {+oo 10 n t 1 e(Ai-S)t x. ) = I ( Ae + = tA ~..Ae tA etA) .1.All succeeding steps in the proof then follow in aastraightforward way.H dt assuming A is diagonalizable ..... Alternatively.Ae tA I = III (etAe~tA L'lt = = /A) ..Ae tA tA tA I I e tA .u . = (sl -A).AetAil Ae tA I ~t (e~tAetA I (M A I ~t (e~tA .. ) 3! 4! L'ltiIAIl < L'lt1lA21111e (1 + + (~t IIAII2 + .=1 m Xiet(Ji-sl)y. A 2etA + .y...11... it can be differentiated term-byProof: Since the series (11.1 The matrix (s I — A) -I is called the resolvent of A and is defined for all s not in A (A). ) etA I < MIIA21111e - L'lt (L'lt)2 + -IIAII + --IIAI12 + . .

(11. The general formula formula d dt l q (t) pet) f(x. the right-hand side above clearly goes to 0 as t:. 0 uniqueness theorem for ordinary differential equations. 11..4). lo t (11. the limit exists and equals Ae'A •. x(to) = Xo E IRn (11.4). say.5) and use property 7 of the matrix exponential to get x ((t) = Proof: Differentiate (11.5) Proof: Differentiate (11.t goes to O.5) and use property 7 of the matrix exponential to get x t ) = Ae(t-to)A xo fundamental Ae(t~to)Axo = Ax(t). or one can use the limit exists and equals Ae t A A similar proof yields the limit et A A.¥o + 0 = XQ so. the right-hand side above clearly goes to 0 as At goes to 0. Let A E Rnxn . B e IR xm and let the vector-valued function u be given Theorem and. t ) . or one can use the fact that A commutes with any polynomial of A of finite degree and hence with e'A. D Ir: Remark 11.4) for t ::: to is given by (11.6). The formula can be derived by means of an integrating factor "trick" as follows. Also. D 11. The formula can be derived by means of an integrating factor "trick" direct differentiation. Let A E IR n xn. by the fundamental existence and uniqueness theorem for ordinary differential equations.2 Homogeneous linear differential equations Homogeneous equations x(t) Theorem 11.Ax = Bu by e. continuous. (11. 0 ordinary differential equations. by the fundamental existence and x(t0) — e(fo~t°')AXQ — Xo uniqueness theorem for ordinary differential equations. Linear Differential and Difference Equations Chapter 11. Then the solution of the linear inhomogeneous initial-value problem and. Then the solution of the linear inhomogeneous initial-value problem x(t) = Ax(t) + Bu(t).i~t()Oc() nnd uniqu()Oc:s:s theorem for *('o)} = <?(f°~fo)/1.5) is the solution of (11.3 Inhomogeneous linear differential equations Inhomogeneous equations Theorem 11. t) dx = l af(x t) ' dx pet) at (t) q + dq(t) dp(t) f(q(t).2. The proof above simply verifies the variation of parameters formula by direct differentiation. A similar proof yields the limit e'A A. The general Proof: Differentiate (11. Linear Differential and Difference Equations For fixed t.1.3.112 112 Chapter 11. t ) . The proof above simply verifies the variation of parameters formula by Remark 11.6).7) Proof: Differentiate (11.7) is the solution of (1l. x(to) = Xo E IR n (11. x(to) = e(to-to)A Xo = XQ so.4. fact that A commutes with any polynomial of A of finite degree and hence with etA. B E Wnxm and let the vector-valued function u be given Let A e IR nxn .7) and again use property 7 of the matrix exponential. say. Premultiply the equation x .dt dt is used to get x ( t ) = Ae(t-to)Ax0 + f'o Ae('-s)ABu(s) ds + Bu(t) = Ax(t) + Bu(t). (11.tA to get as follows. continuous.f(p(t)..5) is the solution of (11.7) and again use property 7 of the matrix exponential. The solution ofthe linear homogeneous initial-value problem Let A e Rnxn.4. Thus.1. Also. Premultiply the equation x — Ax = Bu by e~ to get (11. Thus.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(t-to)A xo + t e(t-s)A Bu(s) ds.7) is the solution of (11. Ae(t-s)A Bu(s) to get x(t) = Ae{'-to)A Xo + Bu(t) = Ax(t) = x(to e(to-tolA Xo + = Xo fundilm()ntill ()lI. (11. the For fixed t.8) . The solution of the linear homogeneous initial-value problem = Ax(l).

.4 Linear matrix differential equations Linear matrix differential equations Matrix-valued initial-value problems also occur frequently.2. E ]R. Theorem 11. the When C is symmetric in (11. e jRnxn. punov differential equation. C e IR" ". B e R m x m .sA Bu(s) ds x(t) = e(t-tolA xo + lto t e(t-s)A Bu(s) ds. problem problem X(t) = AX(t) + X(t)B. the Theorem 11. Theorem 11.11) X(t) = etACe = e ratB has the solution X ( t ) — atACe tB .nxn.12) is known as a LyaX t) punov differential equation. (11. differential equation. X t) X 0 D Corollary 11.1. 11.11) is known as a Sylvester Sylvester differential equation. The solution of the matrix linear homogeneous initial-value e jRnxn.8) over the interval [to. The fact that X((t) satisfies the initial condition is trivial. e-tAx(t) . X(O) =C (11. Let A E Wlxn. Theorem 11. The initial-value problem (11.7. For convenience.6. we can have coefficient matrices on both the right and left. and C e Rnxm. Let A E Rnxn.12) X(t) = etACetAT has the solution X(t} = etACetAT. t]: 113 1 Thus.10) coefficient In the matrix case. t]: Now integrate (11.6. and hence t d -e-sAx(s) ds = to ds 1t to e-SABu(s) ds. X(to) =C E jRnxn (11. t exponential. and the proof is essentially the same. The of nrohlcm problem X(t) = AX(t)..5. and the proof is essentially the same.nxm.1.11. Let A. E ]R. Then the matrix initial-value E jRmxm.1. the Proof: Differentiate etACe tB property Proof: Differentiate etACetB with respect to t and use property 7 of the matrix exponential. X((t) is symmetric and (11.7. the following theorem is stated with initial time to = 0. Differential Equations 11. Corollary 11. The first is an obvious generalization of Theorem 11.12). X(O) = C (11.4 11.e-toAx(to) = lto t e. following to = O.9) for t ::: to is given by for t > to is given by X(t) = e(t-to)Ac. Then the matrix initial-value problem X(t) = AX(t) + X(t)AT.1. Differential Equations [to.2.

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.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. Similarly. are called the modal directions.5 Modal decompositions Let A and suppose. if A is diagonalizable in geneml.1. Then the solution x(t) of (11.x. i=1 In the last equality we have used the fact that yf*Xj = Sfj.5 11.iU-tO)Xiyr) Xo 1=1 n = L(YiHxoeAi(t-tO»Xi.114 114 Chapter 11.4) can be written x(t) = e(t-to)A Xo E jRnxn E Wxn = (ti. Let A E jRnxn and suppose X e jR~xn is such that X-I AX = J.1 n Le A• X'YiH . Linear Differential and Difference Equations 11. ~ 11.y. In the last equality we have used the fact that YiHXj = flij.li y t as discussed in Chapter 9).1. Then the solution x(t) of (11. 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. The decomposition above expresses the solution x (t) as a weighted sum of its directions. ~ 1=1 I t. the rest of this subsection is easily generalized by using the JCF and the decomposition able.H . in the inhomogeneous case we can write t e(t-s)A Bu(s) ds i~ = t i=1 (it eAiU-S)YiH Bu(s) dS) Xi. where J is a JCF for A. that it is diagonalizable (if A is not diagonalizLet A and suppose. Then Then i=1 n = L(aieAiU-tO»Xi. modal velocities and directions.1 . Then Then etA = etXJX-1 = XetJX. 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 *.e'J. where J is a JCF for A. The decomposition above expresses the solution x(t) as a weighted sum of its modal velocities and directions. that it is diagonalizable (if A is not diagonalizable. in the inhomogeneous case we can write Similarly. Linear Differential and Difference Equations Chapter 11.4) can be written A = L X. for convenience.

e. let . Differential Equations 115 If A is diagonalizable. e'u e l N tu x lH = diag(e At .1. it is then easy to compute etA via the formula etA = XetJ X-I' Xe tl X If is etA etA tj since et I is simply a diagonal matrix. aareal version of the above can be worked out. A.7. or grade) p if if matrix M e jRnxn is nilpotent of degree (or index. Mp~l ^ O.EeCkxk be aaJordan block of the form Ji <Ckxk be Jordan block of the form A Ji = 1 o o o =U+N. degree k. . Finally. the problem clearly reduces simply to the computation of problem clearly reduces the exponential of a Jordan block. e lN finite. real version of the above can be worked out.I e IN =I+tN+-N 2 + . O. and so forth. l's For the matrix N defined above. (1. Mp = 0. AI e I. ••• . ext}. N has 1's along only its second superdiagonal.1. i. or grade) MP = 0... Thus. and N kforth. + N k2! (k .. o A o o A Clearly A/ and N commute.eAt).. of In the more general case.I)! I o t 1 o Thus.0.!etN by property 4 of the matrix exponential. the series expansion of e'N is finite. its first superdiagonal (and O's elsewhere). N22 has l's along only its second superdiagonal.8. k) O's k k N = 0. is complex. A matrix M E M nx " is nilpotent of degree (or index. Nk~lI has a 1 in its (1. But e tN is almost as easy since N The diagonal part is easy: e e = diag(e '. Thus. eAt teAt eAt o 2I e 12 At Ik-I At e (k-I)! 0 ell. Differential Equations 11. teAl eAt = 0 0 0 2I e 12 At teAl 0 eAt In the case when A is complex. k) element and has O's everywhere else. To be specific... t2 t k.. while MP-I t=. e ttJi = eO.11. elN is is nilpotent of degree k. it is easy to check that while N has 1's along only its first superdiagonal (and O's elsewhere). nilpotent Definition 11.

.s are distinct. compute f(A) = etA. compute f(A) = e'A. The method is stated and illustrated for the exponential function but applies equally well to other functions. Let A = [-~ o -~0-1~ ] t . k = 0. g(-I) = f(-1) ==> ao . Define the Ai nr=1 n where ao.116 Chapter 11. The motivation for this method is the Cayley-Hamilton Theorem.1 in the power series for et A can be written in terms of these greater n— e' A lower-order powers as well. The motivation for this method is known.. the function g is known and /(A) = g(A). an-i solution of the n equations: g(k)(Ai) = f(k)(Ai).. Let A = [ ~_\ J]. Linear Differential and Difference Equations Example 11. With the aiS then kth superscript (&) X. where t is a fixed scalar. . .. . .. ==> 2a2 = t 2 e. . Then A(A) = {-2.10. .) = -(A + 1)3. the function g is known and f(A) = g(A). in fact. f(A) n(A) etK.1. Thus. -2} and Example 11.nxn and /(A) = etx. Let A Then A (A) = {-2. Here. — 1.Ai t'. n .2a2 = te. .10. -2} and etA Xe tJ =[=i a = x-I =[ =[ 2 1 ] exp t ] [ [ -2 0 -~ ] [ -1 -1 2 -1 2 ] 2 1 e~2t te. the unique OTQ. The method is stated and illustrated for the hand calculation in small-order problems. + I) 3 ..9. -(A. . The polynomial g gives the appropriate linear combination.2t ][ -1 ] Interpolation method Interpolation method This method is numerically unstable in finite-precision arithmetic but is quite effective for effective hand calculation in small-order problems. .a l +a2 = e==> at . a. Let Example 11. Suppose the characteristic polynomial of A can be written as n(A)) = Yi?=i (A . which says that all powers of A greater than A n . . I... i Em. t fixed Given A € E.. characteristic of n(X (^ ~~ ^i)"'» where the A.t • g'(-1) = f'(-1) g"(-I) = 1"(-1) . so m = 1 and nl Let g(X) = UQ + alA + a2A2. functions. and /(A) = etA. They are. an-l are n constants that are to be determined. ni . Theorem 9. Then the three equations for the a.s known.t .I. I. Then jr(A. .1 can be expressed as linear combinations of Ak for k = 0. so m = 1 and n{ = 3. all the Ak — expressed k 1. lower-order g Example 11.s are given by g(A) — ao aiS a\X o^A.2. Given A E jRnxn and f(A) = etA.9. .2t e. terms of order greater than n .3.. the superscript (k) denotes the fcth derivative with respect to A. Linear Differential and Difference Equations Chapter 11.

we find 117 Thus.2t aL = + 2te. 1.11.c-I{(sI — A)-I} is quite effective for small-order problems. we find ao = e. There is an extensive literature on approximating certain nonlinear functions by rational functions. Then rr(A) = f\ + o\2 so m = and (A i 2) «i nL = 2.2t .s.2t . Use etA = £~l{(sl . 2.2t _ Other methods Other methods 1.2t [ ~ o ] + te. Thus.1. te. Let g(A.A)^ 1 } and techniques for inverse Laplace transforms. 2. Let A _* Example 11. t ff>\ tk TU^^ _/"i\ Example 11.) = «o + ofiA. Use Pade approximation..2t I [-4 4] -I 0 _ [ - e. This etA = . Then the defining equations for the aiS are given by 6] g(-2) = f(-2) ==> ao ==> al 2al = e.2t ) 2te. g'(-2) = f'(-2) = te- Solving for the a. we find Solving for the a. Differential Equations 11 . f(A) = etA = g(A) = aoI + al A = (e.-s are given by Let g(A) ao + aLA.2t -te. 2t .1.11. Let A = [ ~4 J] and /(A) = eO-.2t . Differential Equations Solving for the ai s.2t + 2te. The matrix analogue yields e A ~ functions rational eA = . we find Solving for the aiS. but general nonsymbolic computational effective small-order techniques numerically problem equivalent techniques are numerically unstable since the problem is theoretically equivalent to knowing precisely a JCE JCF. Then 7r(X) = (A+ 2)22 so m = 11and [::::~ 4i and f(A) = ea. s. Then the defining equations for the a.11.

where the matrix A in (11. Unfortunately. 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. Let A E Rnxn. exhibit many parallels to the continuous-time differential equation difference equations. Linear discrete-time systems. Linear Differential and Difference Equations Chapter 11. + opAP and N(A) = vol + vIA + D~ (A)N(A). This can be arranged by scaling A. where D(A) = 001 + olA + . convenience. The solution ofthe linear homogeneous system of difference equations equations (11.2. and this observation is exploited frequently.13. case. 11.2.• + Vq A q. B e Rnxm and suppose {«*}£§ « a given sequence of m-vectors. Again. for example. and since we want to keep the formulas "clean" (i. no double subscripts).2. we have chosen ko = 0 for want to keep the formulas "clean" (i.e.13) for k > 0 is given by for k 2:: 0 is given by Proof: The proof is almost immediate upon substitution of (11. Numerical loss of accuracy can occur in this procedure from the successive squarings. and since we consider an arbitrary "initial time" ko.13). a Fade approximation for polynomials of various orders. exhibit many parallels to the continuous-time differential equation case.13. Proof: The proof is almost immediate upon substitution of (11. Let A e Rnxn. We could also case. in the matrix case the exponential is accurate only in a neighborhood of the origin.e. e (e( 3. 11. Linear discrete-time systems. by 22' 2* )A A multiplying it by 1/2k for sufficiently large k and using the fact that A = / { ]I //2')A )\ * .13) is constant and does not depend on k. Linear Differential and Difference Equations D-I(A)N(A). of matrix functions such as e A and 10g(A) remains a fertile area for research.2 11. modeled by systems of difference equations.14. [19]. we restrict our attention only to the so-called time-invariant case. no double subscripts). We could also consider an arbitrary "initial time" ko. say.118 118 l Chapter 11. Reduce A to (real) Schur form S via the unitary similarity U and use eA = U e SsUH Ue U H and successive recursions up the superdiagonals of the (quasi) upper triangular matrix and successive recursions up the superdiagonals of the (quasi) upper triangular matrix e s. Many methods are outlined in.2 Inhomogeneous linear difference equations Inhomogeneous linear difference equations E jRnxn. a Pad6 approximation for denominator the exponential is accurate only in a neighborhood of the origin. but since the system is time-invariant. The solution of the linear homogeneous system ofdifference Let A e jRn xn.13).2 Difference Equations Difference Equations In this section we outline solutions of discrete-time analogues of the linear differential In this section we outline solutions of discrete-time analogues of the linear differential equations of the previous section..1 11..15) . [19]. Reduce A to (real) Schur form S via the unitary similarity U and use e A 3. Numerical loss of accuracy can occur in this procedure from the successive squarings. E jRnxm {udt~ is of Theorem 11. Many methods are outlined in.14) into (11. This can be arranged by scaling A. and this observation is exploited frequently. = P = multiplying it by 1/2* for sufficiently large k and using the fact that e = ( e j . Unfortunately. we have chosen ko = 0 for convenience. 11. say.14) into (11. Again. but since the system is time-invariant. Then the solution of the inhomogeneous initial-value problem m-vectors. for example. Reliable and efficient computation of matrix functions such as e A and log(A) remains a fertile area for research.12.. eS . Reliable and efficient computation 4... Then the solution of the inhomogeneous initial-value problem (11. where the matrix A in (11.. modeled by systems of equations of the previous section. 4.13) is constant and does not depend on k. in the matrix case this means when || A|| is sufficiently small.1 Homogeneous linear difference equations Homogeneous linear difference equations Theorem 11. 0 D Remark 11. •• vq A . we restrict our attention only to the so-called time-invariant Remark 11. by this means when IIAII is sufficiently small.

Difference Equations 11. sometimes useful Ak. is to use z-transforms.2. k:::.A)-I. in general... 0 D 11.. it is then easy to compute Ak via the formula Ak = XJkXX-I Ak Ak — X Jk If diagonalizable.-z-A =I+-A+"2 A + . +00 k=O z z = (l-z-IA)-I = z(zI . Then Ak = (XJX-I)k = XJkX. One solution method.=1 H l If A is diagonalizable. LXi Jtyi .1 _I tA~X. since /* is simply a diagonal matrix.. Assume that A e M" xn and let X e jR~xn be such that X-I AX = /.16) Proof: The proof is again almost immediate Proof: The proof is again almost immediate upon substitution of (11. One definition of the z-transform of a sequence is +00 Z({gk}t~) = LgkZ-k.2. a matrix exponential. which is numerically unstable but sometimes useful for hand calculation. j=O (11.15).O. substitution of (11.15). One definition of the z-transform of a sequence {gk} is a matrix exponential.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. by analogy with the use of Laplace transforms to compute z-transforms.3 11. where J is a E jRnxn and X E R^n J. k=O Assuming Izl > max |A|. Difference Equations 119 119 is given by k-I xk=AkXO+LAk-j-IBUj. Jk .11.16) into (11. Then JCF for A.y. again mostly for small-order probsmall-order lems. the z-transform of the sequence {Ak is then given by Assuming |z| > max IAI. the z-transform of the sequence {Ak}} is then given by AEA(A) X€A(A) k "'kk 1 12 Z({A})=L.16) into (11.H m if A is diagonalizable. based Methods based on the JCF are sometimes useful.2.2. X~1 AX JCF for A.

1 (-2 . Let A Ak = XJkX-1 = [=i -4 a [2 1 J]. 0 A Writing /..is complex.2) . 11.. e Cpxp be a Jordan block of the form o . the initial-value problem initial-value (11. Then Then 1 ] [(_2)k 1 0 k(-2)kk(-2) 1 ] [ _ [ (_2/.. 18]. Linear Differential and Difference Equations In the general case. To be specific. the problem again reduces to the computation of the power of a To Ji E Cpxp Jordan block. Ch.15. y(O) = CI.1. but again no universally "best" method exists.6 can also methods 11. it is then straightforward to apply the binomial theorem to (AI + N)k and verify that straightforward N)k (XI verify Ak kA k-I Ak k 2 (.3 Higher-Order Equations Higher-Order Equations differential It is well known that a higher-order (scalar) linear differential equation can be converted to higher-order a first-order linear system. Consider. for example.(^ . Let A = [_J Example 11.2 0 0 0 0 kA k . real version of the above can be worked out. it is commute. For an erudite discussion of the state of the art.3 11.l8) . In the case when A.. ..1(2k . and is to be interpreted as 0 if k < q. let 7.1 Ak ( .2k) -k( _2)k-1 ] k( -2l+ (-2l. A is complex.• = AI and noting that AI and the nilpotent matrix Writing Ji = XI + N and noting that XI and the nilpotent matrix N commute. ) Ak. The symbol ( ) has the usual definition of . Example 11.1 Ak The symbol (: ) has the usual definition of q!(kk~q)! and is to be interpreted as 0 if k < q.1. see [11.)A - ( k ) Ak-P+I p-l 0 J/ = kA k. Linear Differential and Difference Equations Chapter 11. the problem again reduces to the computation of the power of a In the general case.17) with ¢J(t) a given function and n initial conditions 4>(t} y(O) = Co. in-I)(O) = Cn-I' (1l.15. 1 -1 1 -2 1 ] Basic analogues of other methods such as those mentioned in Section 11.120 Chapter 11.6 be derived for the computation of matrix powers. [11. aareal version of the above can be worked out. . but again no universally "best" method be derived for the computation of matrix powers.

EXERCISES EXERCISES 1. where I + get.. Show that e % / + 1. at least for computational purposes. aly(t) . into a linear first-order difference equation with (vector) initial with n initial conditions. is often well worth avoiding. as mentioned before. .an-lXn(t) + ¢(t). the companion Note that det(A! . let a = XT y. the companion matrix A in (11... where !(eat . 3. However. Cn -\ . Define a vector x (t) E ]Rn with Here. . Let .. Show that e P ~ ! + 1. However. = X3(t) = yet). Cl.I) g(t.a\X2(t) . and. +h a\X+ ao. Let 3. Define a vector x (?) e R" with components *i(0 = y ( t ) . v (m) denotes the mth derivative of y with respect to t. A similar procedure holds for the conversion of a higher-order difference equation A similar procedure holds for the conversion of a higher-order difference equation with n initial conditions. condition.. . let a = xTy. a)xyT. x2(t) = y ( t ) .Exercises 121 121 Here.. Show that etA 2..a)= { a t nxn p if a if a 1= 0.. Let P E lR nxn be a projection. be a projection.. as mentioned before. = Xn(t) = y(n-l)(t). Suppose x. ..19) possesses many nasty numerical properties for even moderately sized n and.. y(m) denotes the mth derivative of y with respect to t. Show that e'A 1+ g ( t . X2(t) yet)...19) possesses many nasty numerical properties for even moderately sized n matrix A in (11.718P. into a linear first-order difference equation with (vector) initial condition. is often well worth avoiding. c\. xn(t) = In-l)(t). Xn(t) y { n ~ l ) ( t ) . Further.A) = A. a)xyT. 2. . These equations can then be rewritten as the first-order linear system These equations can then be rewritten as the first-order linear system 0 0 x(t) = 0 0 1 0 0 0 -ao -a\ x(t)+ [ 0 1 -a n-\ n ~(t) r. Then Xl (I) X2(t) = X2(t) = y(t). .an_lln-l)(t) Xn-l (t) Xn(t) = y(n)(t) = -aoy(t) - + ¢(t) = -aOx\ (t) . Let P € R 1. y € R" and let A = xyT. Further.19) The initial conditions take the form ^(0) = c [CQ. •.."+ an_1A n-~ + . .. Suppose x. y E lRn and let A = xyT. . Then components Xl (t) yet).. C M _I] The initial conditions take the form X (0) = C = [co. (11.718P. = O. Note that det(X7 — A) = An + an-\Xn 1l H alA + ao. at least for computational purposes.

Show that E jRmxn e = [eoI A sinh 1 X ] ~I . Show that -).A 1 .. Linear Differential and Difference where X e M'nx" is arbitrary. (b) Suppose S is symplectic and let A. Let K denote the skew-symmetric matrix 0 [ -In In ] 0 ' In A E jR2nx2n where /„ denotes the n x n identity matrix. must (a) Suppose E is Hamiltonian and let A. Let a. Show that 1/).. f3 E R and Then show that Then show that ectt _eut cos f3t sin f3t ectctrt e sin ~t cos/A J. Find a general expression for Find a general expression for 7. be an eigenvalue of H. Show S~1 H S Hamiltonian. (d) Suppose 5. 6.122 122 Chapter 11. Show that 1 /A. Hamiltonian if K~1ATK = -A and to be symplectic if K -I ATK = A --I.. Let 5.A and to be symplectic K~l AT K . 4. must also be an eigenvalue of H. Show that S-I HS must be Suppose and symplectic. also eigenvalue of (c) Suppose that H is Hamiltonian and S is symplectic. Show that —A. Let (a) Solve the differential equation (a) Solve the differential equation i = Ax . .. be an eigenvalue of S.be an eigenvalue of H. must (b) Suppose S is symplectic and let). Find eM when A = Find etA = 8. Linear Differential and Difference Equations Chapter 11. Let denote the skew-symmetric matrix 4. also be an eigenvalue of H. ft € lR and Let a. x(O) =[ ~ J. must also be an eigenValue of S.be an eigenvalue of S. H (d) Suppose H is Hamiltonian. Hamiltonian. Show that eH must be symplectic. (a) Suppose H is Hamiltonian and let). A matrix A e R 2nx2n is said to be K -I AT K = .

(d) Find the limiting distribution of the $40 trillion as the universe ends. Europe (E). If £0 = 1 and z\ If Zo = 1 and ZI = 2. For Europe and Asia. Consider the n x n matrix initial-value problem X(t) = AX(t) . k ---* +00 (i. Consider the initial-value problem 9. (a) Find the solution of the initial-value problem (a) Find the solution of the initial-value problem . 12. goes to Asia.11 in [24]. and the Americas (R).. (b) Consider the difference equation (b) Consider the difference equation Zk+2 + 2Zk+1 + Zk = O.3. a quarter goes to Europe.) 12. of Cf or all t. (b) Find the eigenvalues and right eigenvectors of M. and the Americas (R). 11. The year is 2004 and there are three large "free trade zones" in the world: Asia (A). i. Show that ||*(OII2 = aforallf > O.YeO) = O..3. what is the value of ZIOOO? What is the value of Zk in 2. Consider the n x n matrix initial-value problem 10.X(t)A.. i.e. 10. as k —»• +00 (i. X(O) = c. x(O) = Xo for t ~ O.Exercises Exercises (b) Solve the differential equation (b) Solve the differential equation i 123 = Ax + b. Suppose that A E ~nxn is skew-symmetric and let ex = Ilxol12. and a quarter year half of the Americas' money stays home. I/X(t)1/2 = ex for all t > 0.e.Yet) + 2y(t) + yet) = 0. as (d) Find the limiting distribution of the $40 trillion as the universe ends. 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?. For Europe and Asia. half stays home and half goes to the Americas. yeO) = 1. Suppose certain multinational companies have Europe (E). and a quarter goes to Asia. Suppose certain multinational companies have total assets of $40 trillion of which $20 trillion is in E and $20 trillion is in R.. what is the value of ZIQOO? What is the value of Zk in general? general? . half stays home and half goes to the Americas. x(O) =[ ~ l 9.11 in [24]. The year is 2004 and there are three large "free trade zones" in the world: Asia (A). a quarter goes to Europe. 11.e. Each year half of the Americas' money stays home. Show that for t > 0. around the time the Cubs win a World Series). (c) Find the distribution of the companies' assets at year k. 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. Suppose that e E"x" is skew-symmetric and let a = \\XQ\\2.e. (Exercise adapted from Problem 5. Consider the initial-value problem i(t) = Ax(t). (c) Find the distribution of the companies' assets at year k. around the time the Cubs win a World Series). .) (Exercise adapted from Problem 5.

This page intentionally left blank This page intentionally left blank .

characteristic hence nonreal eigenvalues must occur in complex conjugate pairs. Definition 12.3. B E C MX if there exists a scalar A E C. The roots ofn('A) are the eigenvalues of the associated nomial of the matrix pair (A. B E enxn. B E E" xn . The standard eigenvalue problem considered in Chapter 9 obviously where A.'AB is singular. the characteristic polynomial is obviously real. The polynomial n('A) = det(A — A.Chapter 12 Chapter 12 Generalized Eigenvalue Generalized Eigenvalue Problems 12. 125 125 . if x [y] is a right [left] ax [ay] for any eigenvector.XB is called a matrix pencil (or pencil of the matrices A and B). eigenvector. B e jRnxn. B).1) Ax = 'ABx.4. Remark 12. then so is ax [ay] for any nonzero scalar a E <C. B). B) with A. and A.1 The Generalized Eigenvalue/Eigenvector Problem The Generalized Eigenvalue/Eigenvector Problem Ax = 'ABx.1.5) is called the characteristic polynomial of the matrix pair (A.4. In this chapter we consider the generalized eigenvalue problem In we the generalized eigenvalue problem where A. . When A.1. e C. B) with A.'AB) is called the characteristic polyDefinition 12. (A.2. Similarly. a. called a generalized eigenvalue. eigenvalues for the generalized eigenvalue problem occur pencil — XB problem occur where the matrix pencil A . Definition 12. B). and Remark 12. hence nonreal eigenvalues must occur in complex conjugate pairs. The matrix A . such that that (12. corresponds to the special case that B = I. B e C" xn The standard eigenvalue problem considered in Chapter 9 obviously corresponds to the special case that B = I.1 12.) = det(A . As with the standard eigenvalue problem.3. B e enxn" if there exists a scalar 'A.2) When the context is such that no confusion can arise. e e. As with the standard eigenvalue problem. The matrix A — 'AB is called a matrix pencil (or pencil of the matrices A Definition 12. A E en Definition 12.) are the eigenvalues of the associated generalized eigenvalue problem. The polynomial 7r(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. called a generalized eigenvalue. Definition 12. generalized eigenvalue problem. A nonzero vector x e C" is a right generalized eigenvector of the pair generalized eigenvector of (A. the adjective "generalized" "generalized" standard eigenvalue [y] is usually dropped.2. The roots ofn(X.

f3 / 0. ^ 0. 1 and O.LA) == 0./. I1 and |. I (of multiplicity 1). the pencil A — XB is said to be 12. with its reciprocal eigenvalue being 0 in Case 3 of the reciprocal pencil B — /. (3 = 0. There are two eigenvalues. only the case of regular pencils is considered in the remainder of this chapter.AB) and there are several cases to consider.XB. Associated with any matrix pencil A .XB.nA. There is only one eigenvalue. A similar reciprocal symmetry holds for Case 2.B. k E !!. is singular.A.XB) is not identically zero. ft =I. However. All A E C are eigenvalues since det(B . Case = ft ^ 0. 1 and 0. B k e n. =I. ft ^ O. then rr(A) is a polynomial nonsingular). There are two eigenvalues. If = of degree n. (3 = O. there is a second eigenvalue "at infinity" for Case 3 of of . eigenvalues — AB.LA) = (1 . I and O. There are two eigenvalues.O. That is to say. I). {3 =I.5.AB Definition 12. I multiplicity 1). There are two eigenvalues. eigenvalues associated with the pencil A . All A E C are eigenvalues since det(A .LA and corresponding generalized eigenvalue problem. and hence there are n eigenvalues associated with the pencil A .6.{3 = 0. pencil . All A 6 C are eigenvalues since det(B — uA) = O. Then the characteristic polynomial is ft det(A .5. It is instructive to consider the reciprocal pencil associated with the example in It reciprocal Remark 12. A similar reciprocal symmetry holds for Case 2. regular. when B is singular. There are two eigenvalues. Case 4: a = 0. There are two eigenvalues.X B is a reciprocal pencil B — n.B) == O.3) where a and (3 are scalars. when B =I. If del (A .0. 1 and ~. Case 2: = 0. there may be 0. Case 1: ^ 0. det(B . There are two eigenvalues. Case 2: a = 0. n(X) Remark 12.3). zero. All A e C are eigenvalues since det(A — AB) =0.A and corAssociated with any matrix pencil — AB is a reciprocal pencil . Case 1: a =I. Note that if AA(A) n J\f(B) ^ 0. Note appear. There is only one eigenvalue. {3 =I. Case 3: Case 4: = 0. (3 = O. If B = I (or in general when B is nonsingular).AB. suppose associated — AB.(3A) ±.L)({3 .0. f3 = O. {3 =I. otherwise. or infinitely many B = I. With A and B as in (12. the pencil A . 1 Case 3: a =I. Clearly the reciprocal pencil has eigenvalues responding generalized /. Case 1: a =I. (12. Note that A and/or B may still be singular. {3 = 0. f3 = 0. I and ^. in particular. Case 2: a = 0. While While there are applications in system theory and control where singular pencils appear. There are two eigenvalues.126 126 Chapter 12.6. At least for the case of regular pencils.L) and there are again four cases to consider. reciprocal Case of reciprocal .KB always has pencil — AB ./. If B is singular.LA. the associated matrix pencil is singular (as in Case N(A) n N(B) =Isingular 4 above). However.I.AHa .0. {3 ^ 0. If det(A — AB) not regular. f3 = O. Generalized Eigenvalue Problems Remark 12. Case 4: a = 0. only the case of regular pencils is considered in the remainder of this chapter. For example.a/. and ~.O. Case Case 3: a = 0.5. 1 and 0. Generalized Eigenvalue Problems Chapter 12. it is said to be singular. Case 4: a = 0. it is apparent where the "missing" eigenvalues have "missing" gone in Cases 2 and 3.0.O. A — A. 1 (of multiplicity 1)./. the characteristic polynomial is = (I .L = (JL = £./. Case 1: a ^ 0.

f i always has precisely eigenvalues. [7. the result follows easily by noting that yH(A — XB) — 0 if and only if yH (A . Let A. [7. in fact. There is also an analogue of the Murnaghan-Wintner Theorem for real matrices. see. 12.7]. 3. the result follows. since the generalized eigenvalue problem is then easily seen to be equivalent eigenvalues. which is the generally preferred method for solving the generalized eigenvalue problem. ify is a left of -AB.7]. and the first theorem deals with what happens to eigenvalues lencies rather than similarities. see. since the generalized eigenvalue problem is then easily seen to be equivalent to the standard eigenvalue problem B.7. However. fl.2. of A-AB. Proof: Proof: 1. 7. canonical forms are available for the generalized Just as for the standard eigenvalue problem.7]. lencies rather than similarities.7. the pencil A --AAB always has precisely n .AB) o if and only if (Q-H y ) H Q ( A –_ B ) Z = Q. where Ta and TfJ are upper triangular. If B is nonsingular. . Q. B E Cnxn Then there exist unitary matrices Q. Numerical methods that work directly on A and are discussed in standard textbooks on numerical linear algebra. ifx isa right eigenvector of A—XB.AQBZ) = det[Q(A . Canonical Forms 12. By Theorem 12. Sec. and the first theorem deals with what happens to eigenvalues and eigenvectors under equivalence. for example.7] or [25. E nxn with Q and nonsingular. the pencil A fewer than eigenvalues.l W AW). in fact. [7. the result follows. fewer than n eigenvalues. work directly on A and B are discussed in standard textbooks on numerical linear algebra.7] or [25. There is also an analogue of the Murnaghan-Wintner Theorem for real matrices. det(QAZ . this turns to the standard eigenvalue problem B~1Ax = Xx (or AB~1w = Xw). [7. with the understanding onal elements of Ta to the corresponding diagonal elements of Tp. canonical forms are available for the generalized eigenvalue problem. Then 12. Let A.l Ax Ax (or AB. for example. 7.. where Ta and Tp are upper triangular.AB and QAZ . which is the generally preferred method for theoretical foundation for the QZ algorithm. c 3.XB)Z] = detQ det Z det(A -. 6. B e cnxn . the eigenvalues of the problems A — XB and QAZ — XQBZ are the same (the two 1. we now deal with equivaa matrices. 2. Then 1. The result follows by noting that (A -AB)x = 0 if and only if Q(A -AB)Z(Z-l x) = The result follows by noting that (A –yB)x . Since det 0 and det Z are nonzero. the eigenvalues of the pencil A . for example. the eigenvalues of the problems A . Let A. for example. and det Z are nonzero. Let A. Since the latter involves a pair of matrices. the theoretical foundation for the QZ algorithm. Theorem 12.7].7] or [25. with the understanding that a zero diagonal element of Tp corresponds to an infinite generalized eigenvalue. Canonical Forms 127 B is nonsingular. However. the eigenvalues ofthe pencil A — XB are then the ratios of the diagonal elements of Ta to the corresponding diagonal elements of TfJ .2 Canonical Forms Canonical Forms Just as for the standard eigenvalue problem. that a zero diagonal element of TfJ corresponds to an infinite generalized eigenvalue. Sec. see. Z e Cnxn with Q and Z nonsingular.7.7. QBZ = TfJ . 7. 0 ( Q ~ H y)H Q(A X AB)Z = O.2 12.8. o. Q. det(QAZ-XQBZ) = det[0(A . Theorem 12. Numerical methods that if B is even moderately ill conditioned with respect to inversion.AB).2. 2. 6. Sec. solving the generalized eigenvalue problem. Sec. 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.8. see. B. to ifx is a Z-l x is a righteigenvectorofQAZ-AQB Z. ify isa left eigenvector of A —KB. then Q-Hy isa left eigenvector ofQAZ -AQBZ. Since det Q XB). 6. the The first canonical form is an analogue of Schur's Theorem and forms.AB)Z] = det gdet Zdet(A 1. then Z~lx isa right eigenvector of QAZ—XQ B Z. D The first canonical form is an analogue of Schur's Theorem and forms. 6. Sec. and eigenvectors under equivalence.AQBZ are the same (the two problems problems are said to be equivalent).12.AB are then the ratios of the diagBy Theorem 12. Sec.Oif andonly if Q(A-XB)Z(Z~lx) = 0. Again. 7. Q~H y isa lefteigenvectorofQAZ — XQBZ. E c nxn such that QAZ = Ta . Then there exist unitary matrices Q.7] [25. Z e Cnxn such that 12. Sec.

Let A. B E c nxn pencil — AB Theorem 12.I. including analogues of principal vectors and description of of so forth.12 (Kronecker Canonical Form). Then there exist 12.. real eigenvalues. I . [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 .9.A. while the full KeF in all its generality applies also to "rectangular" and singular KCF "rectangular" pencils. B e c mxn .9..11. T.)"N). Let A. KCF. When S has a 2 x 2 diagonal block. we present only statements of the basic theorems and some examples.XB is regular. L l" L~. . of — XB. of eigenvalues are given as above by the ratios of diagonal elements of S to corresponding elements of T. Otherwise. is beyond the scope of this book. quasi-upper-triangular. A full description of the KeF. where T is upper triangular and S is quasi-upper-triangular..AB where J is a Jordan canonical form corresponding to the finite eigenvalues of A -A. Z e R"xn such B E jRnxn. Let A.. Example 12. There is also an analogue of the Jordan canonical form called the Kronecker canonical fonn Kronecker form (KeF). .AB)Q = [~ ~ ] . .2)2 with characteristic polynomial (A — 2)2 has a finite eigenvalue 2 of multiplicty 2 and three 2 2 infinite eigenvalues.11.12 mxm nxn mxm nxn E C nonsingular nonsingular matrices P e c and Q e c QE C such that peA .AB. form (KCF). B e Cnxn and suppose the pencil A .128 Chapter 12. Then there exist orthogonal matrices Q. Generalized Eigenvalue Problems Chapter 12. In this chapter.10.'. Generalized Eigenvalue Problems Theorem 12. mxn E C • Theorem 12. The first theorem pertains only to "square" regular pencils.A [~ ~ l of . 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. The matrix pencil 12. B e Rnxn. J .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 . Q € c nxn"such that nonsingular E C" such that peA .AB)Q = diag(LII' . Then there x exist nonsingular matrices P.• L. E jRnxn 12. QBZ = T. thnt that QAZ = S.

and L^ is the (k + I) x k where N is nilpotent. i.12. suppose S e Rn* xk is a matrix whose columns span a k-dimensional E ~nxk ^-dimensional subspace S of ~n.13.2. 000 Just as sets of eigenvectors span A-invariant subspaces in the case of the standard eigenvectors eigenproblem (recall Definition 9. The /( are called the left minimal indices while the r. Specifically. Then V is a E ~nxn suppose pencil — AB deflating subspace if deflating subspace if dim(AV + BV) = dimV.e. both N and J are in Jordan canonical form.. i. Consider a 13 x 12 block diagonal matrix whose diagonal blocks are -A 0] I o -A I . are called the right minimal indices. next two correspond to correspond J = 21 0 2 [ o 0 while the nilpotent matrix N in this example is N [ ~6~]. there is a matrix characterization of deflating subspace.XB is regular.e. generalized eigenproblem. Definition 12. R ( S <S. Then SS is aadeflating subspace for the pencil A . Let A. 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. B e Wlxn and suppose the pencil A . LQ. Example 12.14.— XBif S Rn.. Then is deflating subspace for the pencil A AB if and only if there exists M E Rkxk such that e ~kxk AS = BSM.2.4) eigenvalue characterization Just as in the standard eigenvalue case. Left Left or right minimal indices can take the value O. corresponds LQ. Lo L6 one column. there is an analogous geometric concept for the eigenproblem generalized eigenproblem.The next two blocks second block L\ one the block is L\. The first block of zeros actually corresponds to LQ. L6. 0. Such a matrix is in KCF. n(S)) = S. both Nand J are in Jordan canonical form. The second block is L\ while the third block is LI. LQ.5) . Lo. while each LQ has "zero rows" and L6. LQ . (12. where each LQ has "zero columns" and one row. Canonical Forms 12.35). Lo. Canonical Forms 129 where N is nilpotent. (12. Lo.

3 12. multi-output systems. where x(= x(t)) is called the state vector.6). where the "system pencil" (12.4) becomes dim (A V + V) = dim V.6». Generalized Eigenvalue Problems If B = /. the (finite) zeros of this system are given by the (finite) complex numbers In general. E jRPxn. C e Rpxn. vector. Numerically. these values are the generalized eigenvalues of the drops rank. D=O. This linear time-invariant statespace model is often used in multivariable control theory. where x(= x(t)) is called the state space model is often used in multivariable control theory. Example 12. In general. 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 multi-input. Similarly. one must be well for general mUlti-input.8. the (finite) zeros of this system are given by the (finite) complex numbers where the "system pencil" z. The connection between system zeros and the corresponding system pencil is nonThe connection between system zeros and the corresponding system pencil is nontrivial. see. = Cx + Du E jRnxn.5) becomes AS = SM as before. B] .5) becomes AS = SM as before. Let A=[ -4 2 C = [I 2]. Numerically. (12. for example. which is clearly equivalent to If B = I. one must be careful first to "deflate out" the infinite zeros (infinite eigenvalues of (12. we offer some insight below into the special case of a single-input. there AV ~ V. zeros). however. If the pencil is not regular.8.6). [26].8. and D € Rpxm. lEthe pencil is not regular. B € R" xm . which has a root at —2. then (12. (12. for example. [26]. 12. and y is the vector of outputs or observables. Then the transfer matrix (see [26]) of this system is Then the transfer matrix (see [26)) of this system is g(5)=C(sI-A)-'B+D= 5 55 2 + 14 ' + 3s + 2 which clearly has a zero at —2. and E jRPxm. In the special case p = m. Checking the finite eigenvalues of the pencil (12. we offer some insight below into the special case of a single-input. and y is the vector of outputs or observables. these values are the generalized eigenvalues of the (n + m) x (n m) pencil.8. However. u is the vector of inputs or controls.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. This linear with A € M n x n . E jRnxm. Similarly. trivial. is a concept analogous to deflating subspace called a reducing subspace. In the special case p = m. (n + m) x (n + m) pencil. This is accomcareful first to "deflate out" the infinite zeros (infinite eigenvalues of (12. A-c M D "'" 5A + 14. u is the vector of inputs or controls. For details.6)). multi-output systems.6) drops rank. we which clearly has a zero at -2. Checking the finite eigenvalues of the pencil (12. Let Example 12. However. 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). then (12.4) becomes dim(AV + V) = dimV. we find the characteristic polynomial to be find the characteristic polynomial to be det [ which has a root at -2. however.130 Chapter 12. which is clearly equivalent to AV c V.15.15. there is a concept analogous to deflating subspace called a reducing subspace. see. For details.

the second-order A.l xn. Suppose z € is such that Suppose Z E C is such that [ A . However.7) we get get x = -(A . or g ( z ) y = 0 by the definition of g. g. . then from (12.9)). no pole/zero cancellations). B e ffi. Hence g(z) = 0. and D e R r T(s7 . or g(z)y 0 by the definition of g. we have _c T (A .12. Specifically.. M K where M is a symmetric positive definite "mass matrix" and K is a symmetric "stiffness definite "stiffness matrix. the problem (12.zl)x + by = 0.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. b e ffi.10) is equivalent Since B is positive definite it is nonsingular. g(s) Furthermore. z is a zero of g.8) c T x +dy = O. system of differential equations differential Mx+Kx=O. A pole/zero Assuming z is not an eigenvalue of A (i. and D = d E R..zl)-lby + dy = 0.n. 12.A)~ ! Z? + d denote the system transfer function (matrix).A to the standard eigenvalue problem B-l1Ax = AJC. Thus. (12. e ffi." is a frequently employed model of structures or vibrating systems and yields a frequently generalized eigenvalue problem ofthe form (12.8). Symmetric Generalized Eigenvalue Problems 131 131 1 single-output system. Thus. B E Rnxn arises when A = A and B = BT > O. Now _y 1= 0 (else x = 0 from (12. and v(s) and n(s) are relatively prime TT(S) v(s) TT(S) (i. Symmetric Generalized Eigenvalue Problems 12. Hence g(z) 0.zl)-lby. Then there exists a nonzero solution to or or (A .zI cT b ] d is singular.4.4 12.7) (12. let g(.. of the Since B is positive definite it is nonsingular.4 Symmetric Generalized Eigenvalue Problems Symmetric Generalized Eigenvalue Problems Ax = ABx A very important special case of the generalized eigenvalue problem (12.e.8).10) for A. (12.e.10). the problem (12.4. symmetric. "pole/zero cancellations"). Now y ^ 0 (else x z i.9». For example. there are no "pole/zero cancellations").e. we have Substituting this in (12. 0 from (12.10) is equivalent B. relatively where n(s) is the characteristic polynomial of A. let B = b E Rn.9) Substituting this in (12. C = c T E R l x n .nxn A AT and B the B1 0.

the eigenvalues of B l A are always real (and are approximately 2. Finally. of course. Let A Example 12. Let A = [~ . but since real-valued matrices are commonly used in most applications. Generalized Eigenvalue Problems Example 12. . then = C T > 0.18. if A = A > 0. .. we have restricted our attention to that case only. it has a Cholesky factorization B = LL T.18.12) Since C = C T the eigenproblem (12. Moreover.fi Then it is easily checked that Then it is easily checked thai c = L~lAL~T = [ 0. The Cholesky factor for the matrix B in Example 12. E !!.1926 and -3. Example 12.23).. •. the eigenvalues are also all positive. Proof: Since B > 0.1926 whose eigenvalues are approximately 2. with corresponding eigenSince C = C T. Finally.16 is Example 12.11) can be rewritten as the equivalent problem 1 Letting C = L ~I AL ~T and z = L1 x. generalized case A and B are Hermitian. l = [i ~ J B ThenB~ A Then A B~Il = [-~ ~ J B~I A approximately Nevertheless.1926 and —3. Xj)B T T = xr BXj = (zi L ~l)(LLT)(L ~T Zj) = Dij. Zn Zj = Dij.. (12.fi 1] . B E Rnxn with A = AT and B = BT > 0. . Then the generalized A. with corresponding eigenvectors zi. = L ~Tzi.1926 in Example 12. are eigenvectors of the original generalized eigenvalue problem and satisfy and satisfy (Xi. the eigenvalue problem eigenvalue problem Ax = ABx has n real eigenvalues. then C = C T > 0. and the n corresponding right eigenvectors can be chosen to be orthogonal with respect to the inner product (x. The Cholesky factor for the matrix B in Example 12.16. y)BB = XT By.12) has n real eigenvalues. Let A.. Generalized Eigenvalue Problems Chapter 12.11) can then be rewritten as AL J and Z = LT x.5 ] -1. then product y) x T By.23).1926 as expected. B e jRnxn A AT and B BT > O. (12. the eigenproblem (12.5 ' -3. zi Then x.12) has n real eigenvalues.17.. Moreover. so the eigenvalues are positive. if A > 0. it has a Cholesky factorization B = LLT. Then the eigenvalue problem Ax = ABx = ALL Tx (12. where L is nonsingular (Theorem 10.11) can then be rewritten as = Cz = AZ.16). ii € n. Then the eigenvalue problem (Theorem 10. zn satisfying vectors Z I. Theorem 12.16.16 is D 0 L=[~ . The material of this section can.5 2. and are Hermitian. but since real-valued matrices are commonly used in most applications..5 2. positive. if A = AT> 0...132 132 Chapter 12. are eigenvectors of the original generalized eigenvalue problem Xi Zi. so the eigenvalues are positive.. if orthogonal > 0. where L is nonsingular Proof: Since B > 0. be generalized easily to the case where A material of can. (12... we have restricted our attention to that case only.

There are many matrices (A.g.19 is very useful for reducing many statements about pairs of symmetric Theorem 12. Now D > 0 by Theorem 10. Then there exists a nonsingular matrix Q such that where D is diagonal. it does preserve the eigenvalues of A . the diagonal elements of D are the eigenvalues of B 1A.5.5. so it does not preserve eigenvalues of and B Note that Q is not in general orthogonal.e.19.19 is There are situations in which forming C = L~1AL~T as in the proof of Theorem 12.19 is numerically problematic. QD~ QT < QQT.1A. A~l :::: B-l1. B) can be simultaneously diagonalized by the same matrix. To illustrate. But then D. Then A 2: B if and only if B~ 2: A-I.. let . haveA(D) = A(B. since A 2: B. Let A. i.e.1 12. individually. In such cases. the diagonal elements of D are the eigenvalues of B. B E E"x" with 12.< / (this is trivially true 0 since the two matrices are diagonal). To illustrate. Then A > B if and only if B-l1 > Theorem 12. Proof: Let B = LLT be the Cholesky factorization of B and set C = L~1AL -T. we restrict our attention only to the real case.e.20. Let A. where D is diagonal. straightforward way. Then there exists a nonsingular matrix Q such that A = AT and B = BT > 0. i.T P. Proof: By Theorem 12. In numerically problematic. simultaneous reduction can also be accomplished via an SVD. Now D > 0 by Theorem 10. we Note that Q is not in general orthogonal. we restrict our attention only to the real case. when L is highly ill conditioned with respect to inversion. Theorem 12. Let Q = L .lI QT :::: Q QT. D > I. = QQT AQQ~l = L-TPPTL~-IA = L~TL~1A L -T P pT L 1 A L -T L -I A QQT AQQ-I 0 D = B-1A.l Q~T QT Q~ B~ AQ. there exists an orthogonal matrix P such that P CP = D. where D is C is symmetric.1A = Q-1l B~1Q-T QT AQ = Q-11B. Since Proof: Let T C is symmetric. normal matrices can be diagonalized by a unitary similarity.19.5 Simultaneous Diagonalization Simultaneous Diagonalization Recall that many matrices can be diagonalized by a similarity." The following is typical. there exists Q E lR~xn such that QT AQ = D and QT BQ = [. e. Proof: By Theorem 12. A -I < B~ .20. simultaneous reduction can also be accomplished via an SVD. D 2: [.5. Again. normal maRecall that many matrices can be diagonalized by a similarity.. there exists Q e E"x" such that QT AQ = D and QT BQ = I. In particular. A-1. Simultaneous Diagonalization 12. with the complex case following in a Again. it does preserve the eigenvalues of A — XB.. where D is diagonal. where D is diagonal. There are many such results and we present only a representative (but important and useful) theorem here. we B~ 1 A. Then and and QT BQ Finally. B e M" xn be positive definite. But then D"1I :::: [(this is trivially true 10.19 (Simultaneous Reduction to Diagonal Form). i. Let A = QT AQandB = QT BQ. This can be seen directly. B E lRnxn be positive definite. Infact. by Theorem where D is diagonal. by Theorem 10. e. there exists an orthogonal matrix P such that pTe p = D. Since LLT be the Cholesky factorization of and setC L -I AL~T. since QDQ~l have A(D) = A(B~1A).5.31. since A > B. B) can be simultaneously diagonalized by the same matrix. This can be seen directly.19 is very useful for reducing many statements about pairs of symmetric matrices to "the diagonal case. so it does not preserve eigenvalues of A and B individually. Let Q = L~T P..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. However.1AQ.'AB.1A). D since the two matrices are diagonal).g.31. Let A.5 12. Thus. Thus.19 e ][~nxn A AT and B BT > O. such results and we present only a representative (but important and useful) theorem here." The following is typical. Theorem 12. However. since QDQ-I Finally. Q D. Also. with the complex case following in a straightforward way. matrices to "the diagonal case. Also. \ 2. Then B.21 we have that QT AQ > QT BQ. LetA QT AQ and B QT Then/HA Q~ B. Then diagonal. = pT L -I(LLT)L -T P = pT P = [. i. In particular. Theorem 12. In fact.21 we have that QT AQ 2: QT BQ. It turns out that in some cases a pair of matrices (A.e. let such cases. when L is highly iII conditioned with respect to inversion. It turns out that in some cases a pair of trices can be diagonalized by a unitary similarity... Simultaneous Diagonalization 133 12.12.

i.14) can be rewritten in the form LALAx = XLBz = Letting x = LBT Z we see 02. respectively.22.3].15) is called a generalized singular value problem and algorithms exist to problem generalized solve it (and hence equivalently (12.21 are possible.e.. but in writing = PDDp D diagonal. for generalizations results 12. This is analogous to finding the singular values of a matrix M by Sec.21 example.13) can be computed without explicitly forming the without Remark product indicated matrix product or the inverse by using the so-called generalized singular value decomposition (GSVD).21.. when A = AT > 0. Then the matrix Q U performs the simultaneous L e 1R~ xn diagonalization. example. Further. Then the matrix Q == LLBTu performs the simultaneous diagonal. The case when A is symmetric but indefinite is not so A = AT::: O. To check this.butin writing A — PDDP T = PD(PD) with D is diagonal and P orthogonal. Various generalizations of the results in Remark 12. The SVD in (12. A can be written as A = PDP T. at least in real arithmetic. Sec. operations performed directly on M rather than by forming the matrix MT M and solving performed MT forming the eigenproblem MT MX = AX. without forming the products LALTA or LBLTB explicitly. Generalized Eigenvalue Problems us assume that both A and B are positive definite. D may have pure imaginary elements. A straightforward. Remark 12. D b . (12.e. For example. Compute the SVD Cholesky factorizations A B.134 134 Chapter 12. for LB i. which is thus equivalent to the generalized eigenvalue problem ALBL~LBT z. PDPT ~ ~ ~ ~ T PD(PD{ with where Disdiagonaland P is orthogonal.. let A = LAL~ and B = LsLTB be Cholesky factorizations of A and B.15) The problem (12. [7. note that T QT AQ = U Li/(LAL~)Li/U = UTULVTVLTUTU i/ = while L2 QT BQ = U T LB1(LBL~)Li/U = UTU = I.14) rewritten the LAL~x = ALBz = A L g L ^ L g 7 z . Generalized Eigenvalue Problems Chapter 12. 8.7.13» via arithmetic operations performed only on LA LA (12. eigenproblem MT M x Xx.13)) and LB separately. respectively. products LA L ~ LBL~ see.14) Letting x = LB z we see that (12. Remark 12. see. which is thus to the generalized eigenvalue problem 02. let A = LALTA and B — LBL~ us assume that both A and B are positive definite.13) where E E R£ x " isisdiagonal. Further. Note that LB A and thus the singular values of L B 1 LA can be found from the eigenvalue problem 02.

..C + K is singular. there are 2n eigenvalues for the second-order (or A2 M + AC + K.. Suppose. and = = KT.. Higher-Order Eigenvalue Problems 12.6. where q(t} e W1 and M. KT > 0). Then (12.16) Consider the second-order system of differential equations Consider the second-order system of differential equations q(t) E ~n E ~nxn.6. that we try to find a solution of (12. the second-order problem (12..12. Suppose K has eigenvalues eigenvalues IL I ::: ... .. E2". C = 0.16) we get (12. p. K = KT ::: 0).2M + A. we thus seek values of A. seek A A2 M + AC + To get a nonzero solution /?. are to be determined. (12.6 Higher-Order Eigenvalue Problems Higher-Order Eigenvalue Problems Mq+Cq+Kq=O. Since the determinantal equation o = det(A 2 M + AC + K) = A2n + . ± Wk. Since the determinantal equation is singular. Higher-Order Eigenvalue Problems 135 12..16) or.16) can be written as a first-order system (with block Let XI q and X2 Then (12. n.. then all solutions of q + Kq = 0 are oscillatory. If r n (i.2M + A. . or if it is desired to avoid the calculation of M lI because M is too ill conditioned with respect to inversion..16) can be written as a first-order system (with block companion matrix) X . where the n-vector p and scalar A. by analogy with the first-order case.16) arises frequently in applications: 0. for which the matrix A.16) of the p A are to be determined. Substituting in form q(t) = ext p. (12. k = r + 1.16) can still M second-order generalized linear be converted to the first-order generalized linear system converted I [ o M OJ'x = [0 -K I -C Jx. 12. ::: ILr ::: 0 > ILr+ I ::: .16) arises frequently in applications: M = I. K e Rnxn. .C + K. If M is singular. ::: ILn· Let a>k = IILk I!. yields a polynomial of degree 2rc.1 12. quadratic) eigenvalue problem A.2 K are are ± jWk.6. = [ -M-1K 0 x (t) E ~2n. M Mwhere x(t) €. r.1 Conversion to first-order form Conversion to first-order form Let x\ = q and \i = q. A special case of (12. .6.6 12.. If r = n (i. Suppose K = KT. Substituting in q(t) = eAt p.e. Assume for simplicity that M is nonsingular.e. (A 2 M + AC + K) p = O. k = 1.• Then the 2n eigenvalues of the second-order eigenvalue problem A2 I /+ K Let Wk = | fjik 12 Then the 2n eigenvalues of the second-order eigenvalue problem A. then all solutions of q K q 0 are oscillatory. and A special case of (12. C. polynomial 2n. since eAt :F 0. .

Show that the finite generalized eigenvalues of E lR " finite eigenvalues of e R™ x m the pencil [~ ~J-A[~ ~J are the eigenvalues of the matrix A — BD 1 C. the kth derivative of q.. Some can be useful when M. such results show that zeros are invariant under state feedback or output injection. Similar procedures hold for the general kthblock companion matrix analogue of (11. E Rnxm and E E 4. lead naturally naturally involving.16) involving.19). C. say. 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 ". andlor K Many other first-order realizations are possible. to higher-order eigenvalue problems that can be converted to first-order form using a kn x kn to higher-order eigenvalue problems that can be converted to first-order form using aknxkn block companion matrix analogue of (11. (A similar result is also true for "nonsquare" pencils.. G E enxn". EXERCISES EXERCISES nx 1. G e Cmxn • Are the nonzero singular values of FG and GF the same? same? wx E ]Rnxn. F 6 Rm *" G R" x .19).1 2. Hint: Consider the equivalence I G][A-UO F0]' B][I l [01 C (A similar result is also true for "nonsquare" pencils. Show that the generalized eigenvalues of the pencils ues of the pencils e e [~ ~J-A[~ ~J and and [ A + B~ + GC ~] _ A [~ ~] are identical for all F E E"1xn and all G E R" xmm . Are the FG and GF the 3. In the parlance of control theory. Suppose A e Rnxn and D E lR::! xm.) . and/or K have special symmetry or skew-symmetry properties that can exploited. B e lRn*m. Let F. which can be converted to various first-order systems of dimension kn. Let € C M X • Show that the nonzero eigenvalues of and G F are the same. properties Higher-order analogues of (12.136 136 Chapter 12. Let F e Cnxm . and C e lRmxn. In the parlance of control theory. . Show that the generalized eigenval". Similar procedures hold for the general k\horder difference equation order difference equation which can be converted to various first-order systems of dimension kn. Generalized Eigenvalue Problems Many other first-order realizations are possible. C. Generalized Eigenvalue Problems Chapter 12. Some can be useful when M. derivative q. Show that the nonzero eigenvalues of FG and GF are the same.B D. Suppose A € Rnxn.

respectively. positive. Such QT BQ a transformation is called contragredient. (b) Show that Q~l = ^~^UT LTB. 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. positive Cholesky = LA L ~ = L B L ~. (c) Show that the eigenvalues of A B are the same as those of 1. A B B are positive definite with Cholesky factorizations A = L&LTA and B = L#Lg. B E e jRnxn Q-l AQ-T ]Rnx" in such a way that Q~l AQ~T and QT BQ are simultaneously diagonal. Q-l = ~-!UTL~.Exercises Exercises 137 137 desired 5. A and B to the same diagonal matrix. and let UWT be an SVD of L~LA'. .2 and hence are AB E2 positive. respectively. 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. Consider the case where both A and transformation contragredient.

This page intentionally left blank This page intentionally left blank .

1.1. Then A@B =[ 3~ ~]~U J. pointing out the restrict our attention in this chapter primarily to real-valued matrices. Let A e R mx ".1 Definition and Examples Definition and Examples Definition 13. 2B 2B ~J. Forany B e!F pxq /z @ B = [~ In Replacing 12 by /„ yields a block diagonal matrix with n copies of B along the I2 diagonal with n copies of along the diagonal. Let B be an arbitrary 2x2 matrix.2. We restrict our attention in this chapter primarily to real-valued matrices. the same definition holds if A and B are complex-valued matrices.Chapter 13 Chapter 13 Kronecker Products Kronecker Products 13.1) amnB Obviously.. Foranyfl E lRX(7. extension to the complex case only where it is not obvious. Then 0 b ll b12 B @/z = l b" b~l 139 0 b2 2 0 b21 0 0 b12 0 b 22 l . B e lR pxq. Then 3. Example 13. Note that B <g> A / A <g> B. We Obviously. Let B be an arbitrary 2 x 2 matrix. Let A = [~ 2 2 nand B = [.A @ B..2. / 2 <8>fl = [o ~ l\ 2.1. Example 13.. 4 3 4 3 4 9 4 2 6 2 6 6 6 2 2 Note that B @ A i. pointing out the extension to the complex case only where it is not obvious. Let A E lRmxn B E R Definition 13.1 13. Then the Kronecker product (or tensor Then the Kronecker product (or tensor product) of A and B is defined as the matrix product) of A and B is defined as the matrix allB A@B= [ : amlB alnB ] : E lRmpxnq. the same definition holds if A and B are complex-valued matrices. n 2. (13.

XIYn. simply note that (A ® B)(A -1 ® B.6.3. For all A and B. Simply verify that ~[ =AC0BD. If A and B are nonsingular. Let A e R mx ". . D Corollary 13. simply verify using the definitions of transpose and Kronecker verify transpose Kronecker 0 product.. . .5.1 ) Theorem 13. Let E ~mxn. Then 13..6. Let Jt € Rm. 0 . then A® B is symmetric. C e ~nxp. .140 Chapter 13. Theorem 13. mn . (A ® Bl = AT ® BT. L~=l al. .m xm are symmetric. 5 E R r x i . (13. Then 13.5.2 13. Foral! Proof' Proof: For the proof.2 Properties of the Kronecker Product Properties of the Kronecker Product (A 0 B)(C 0 D) = AC 0 BD (E ~mrxpt). C E R" x ^ and D E ~sxt.kCkPBD L~=1 amkckpBD ] 0 Theorem 13. 4. XmY T]T = [XIYJ. If A e R"xn and B E !R.. X2Yl. 5. If E ]Rn xn e Rmxm are Theorem 13. = 1 ® 1 = I.1. Proof: Proof: Using Theorem 13.. E R". (A ® B)-I = Bare 13. Let* eR m .4.3. y e !R.3. B In x E ~m.3. Then X ® Y = [ XIY T . xmYnf E !R. Kronecker Products Kronecker Products The extension to arbitrary B and /„ is obvious. B e ~rxs.. If A-I ® B. and D e Rsxt.2) Proof: Simply verify that Proof. A® 13.n. .. y eR".

.• :::: U rr > 0 and let B E IRfx Corollary e R™x" singular a\ > • • > a > e have singular values T\ > • • > <s > 0.3 since A and B are normal by Theorem 13.10...4 by Theorem 13.7. mxm /zave Theorem 13.. If A E IR"xn and B eRmxm are normal.. Lgf A E E mxn have a singular value decomposition VA ~A Theorem 13.3.10. 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 :::: . if x\. TTzen ?/ze mn eigenvalues of A 0 Bare Moreover. . j € m. 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 rTs > 0 and ^iT\ > • • • > ffr <s Qand rank(A 0 B) = (rankA)(rankB) = rank(B 0 A) . if A and fi have Jordan form . xp are linearly independent right eigenvectors of A corresponding AI.[Cos</> cos</>O Then It IS easl'1y seen that . 0 Zj E€ IR mn "are linearly independent right eigenvectors of A 0 B corresponding to Ai JL 7 i e /?.. . q Corollary 13. In general.. • • zq independent of to A . Theorem 13. Example 13. If A E E"xn is orthogonal and B E Mmxm is orthogonal. Properties of the Kronecker Product Theorem 13. we can take p thus get the complete eigenstructure of A 0 B. Then A <g)B (or B 0<8> A) has rs singular values U.8. Ap (p ::::: and ZI.p (p < n). eigenvectors of A® B corresponding to A. . Let A E lR.8.. 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).Zq are linearly independent right eigenvectors of B corresponding to JLI.i .12.13. then A <g> B is € IR nxn orthogonal and e IR m x m 15 then 0 is orthogonal..9. If A e IR nxn am/ B E IR mxm are normal. . A.•. .2... ."xn have singular values UI :::: .. elements of ~A 0 ~B and the corresponding right and left singular vectors). then Xi <8> Zj ffi. = (A 0 B)(A 0 B)T 0 Corollary 13. 0 If A and Bare diagonalizable in Theorem 13.JLqq (q < m)... and zi.. if A and B have Jordan form thus get the complete eigenstructure of A <8> B. If Corollary 13. and let eigenvalues jJij. ./u.-. Properties of the Kronecker Product 13... 7 E m. A0 B e±jeH</» e±jefJ -</».m are linearly independent right corresponding to JJL\ . Sine] and B . ••.n..• :::: TS > O.7.12. 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. matrix A ® 5 is then also orthogonal with eigenvalues e^'^+'W and e ± ^ (6> ~^ > \ Theorem 13.12..2.. j e q.and let BB E e IRR mxwhave e IR nxn have eigenvalues A.c. i / E e!!.11. we can take p = nand q = m and n and q —m and If A and B are diagonalizable in Theorem 13.j. then . L et A E xamp Ie 139 Let A = [ _eose cose andB .• 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. ... if Xl. then A 0 B is normal. Let A E R nx "have eigenvalues Ai. \Ju (q ::::: m).. i E l!! 7 E 1· Proof: proof Proof: The basic idea of the proof is as follows: follows: (A 0 B)(x 0 z) = Ax 0 Bz =AX 0 JLZ = AJL(X 0 z). then A® B is normal. .. The 4 x 4 orthogonal e±j9 orthogonal eigenvalues e±j(i>. In general. <I :::: .. xp are linearly independent right eigenvectors of A corresponding Moreover.

in of A and B. Let A~U Then Then 2 2 !]andB~[ . A ® B i= B © A. denoted A EEl B. denoted A © B. i. Then reducing A and B to real Schur form).I ® Q-l)(A ® B)(P ® Q) = (P. Corollary 13. E IR nxn E IR mxm. A EEl B ^ B EEl A. then we get the JA and Q~] BQ following Jordan-like structure: following Jordan-like structure: (P ® Q)-I(A ® B)(P ® Q) = (P.. to Schur (triangular) form. A Schur form for A ® B can be derived similarly. suppose P and Q are unitary matrices that reduce A and B. is the mn mn matrix (Im ® A) + (B ® /„). Note that. in general. Then (P ® Q)H (A ® B)(P ® Q) = (pH ® QH)(A ® B)(P ® Q) = (pH AP) ® (QH BQ) = TA ® TR . 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). det(A ® B) = (det A)m(det Bt = det(B ® A).142 142 Chapter 13. Example 13. Kronecker Products decompositions given by p. to Schur (triangular) form.e. then we get the decompositions given by P~lI AP = J A and Q-l BQ = JB.e. Then the Kronecker sum (or tensor sum) . Then 13. 1. nxn mxm Definition 13. while upper triangular..14.14. For example. are unitary matrices that reduce A and 5. respectively.13. suppose P and Schur form for A ® B can be derived similarly. respectively. of A and B.13. Note that. Let A e Rn Xn and B e Rm xrn. Example 13. respectively. is generally not quite in Jordan form and needs Note that JA® JB.1 AP) ® (Q-l BQ) = JA ® JB · Note that h ® JR. general. 1. 2.15. ~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.15. Let 1. i. . E IR E IR Kronecker Definition 13. is the mn x mn matrix Urn <g> A) + (B ® In). 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). Kronecker Products Chapter 13. respectively. eigenvalues are zero or nonzero). with A EEl B. For example. Let A e Rn xn and B e Rrn xm.AP J B . while upper triangular. Tr(A ® B) = (TrA)(TrB) = Tr(B ® A).

. . . . Xp (p ::s: n). .i e n. Properties of the Kronecker Product 2. + fJ-j' € p. . is a Jordan-like structure for A © B.13.. Then J can be written in the very compact form J Theorem 13. In general. . we can take p nand q and If A and B are diagonalizable in Theorem 13. A2 + fJ-t. 0 I M 0 where M = [ where M = o M a f3 -f3 a J. A2 + fJ-m. zq are linearly independent eigenvectors of corresponding to fJ-t. (I} ® M) + (E^®l2) = M 0 Ek. xp are linearly independent right eigenvectors of A corresponding to AI.\ .•• . . . 0 If A and Bare diagonalizable in Theorem 13.. Ap (p < and ZI. Recall the real JCF M I M 143 143 0 I M I 0 o 1= 0 E jR2kx2k.. e jRmxm eigenvalues /z. f^q (q ::s: ra). An + fJ-m' Moreover. we can take p = n and q = m and thus get the complete eigenstructure of A $ In general. fJ-q (q < m).. . if A and B have Jordan form p-I l B . if x\.2. 7 e I!!. Let A E E"x" have eigenvalues Ai. if A and have Jordan form thus get the complete eigenstructure of A 0 B. .. i E !!.. and z\. . . and let B E Rmx'" have e jRnxn eigenvalues A.16. j E fl· eigenvectors of A $ B corresponding to A. Properties of the Kronecker Product 13... respectively... . Proof: The basic idea of the proof is as follows: Proof: The basic idea of the proof is as follows: [(1m ® A) + (B ® In)](Z ® X) = (Z ® Ax) = (Z + (Bz ® X) ® Ax) + (fJ-Z ® X) = (A + fJ-)(Z ® X). then decompositions given by P~1AP = lA and Q"1 BQ = JB. respectively.16. then decompositions given JA and Q-t 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 ® p-I)(Q-I ® In)][(lm ® A) In)][CQ ® In)(/m <:9 P)] + (JB ® In) is a Jordan-like structure for A $ B. Define 0 0 0 0 o o Ek = 0 o Then 1 can be written in the very compact form 1 = (4 <8>M) + (Ek ® h) = M $ E k . .. .. eigenvectors of A® B corresponding to Ai + [ij. j E ra. Recall the real JCF 2. ii E E. . . . Zq are linearly independent right eigenvectors of B AI. .2..... Then the Kronecker sum A® B = (1m (g>A) + (B ® In) has mn (Im ® A) + (B <g> /„) /za^ ran eigenvalues fJ-j.xp are linearly independent right eigenvectors of A corresponding Moreover.···.. TTzen r/ze Kronecker sum A $ B eigenvalues e/genva/wes Al + fJ-t.-. AI + fJ-m. then Zj ® Xi E€ jRmn" are linearly independent right Zj <8> Xi W1 are linearly independent right corresponding f j i . if XI.. j e q.16..

=1 A special case of (13. B e Rmxm . Again. = C.3) mxm E IRnxn E IR E IRnxm.J. to Schur (triangular) form. 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. where [(Q <8>In)(lm ® P)] = (Q ® P) is unitary by Theorem 13. This equation is now often called a Sylvester equation is now often equation in honor of 1. When C is symmetric.3) is the symmetric equation AX +XAT = C (13. The first important question to ask regarding (13.8.3) in terms of their easily seen z'th columns that ith columns that m AXi + Xb. Lyapunov equations also to be symmetric and (13.3) is.1. . Then ((Q ® /„)(/« ® P)]"[(/m <8> A) + (B ® /B)][(e (g) /„)(/„.3 and Corollary 13.4) obtained by taking B = AT. suppose P and are unitary A Schur form for A © B can be derived similarly. . Then to real Schur fonn). i. (13. it is easily seen by equating the writing (13. pH AP = TA matrices that reduce A and B. arise naturally in stability theory. 13.e. Lyapunovequations arise naturally in stability theory. i. [(Q ® /„)(/« ® P)] = (<2 ® P) is unitary by Theorem 13.3) in tenns of their columns.5) as an "ordinary" linear system. When does a solution exist? By writing the matrices in (13. Sylvester where A e R"x".XB.144 Chapter 13.5) [ blml The coefficient matrix in (13.3) is.8. the solution X E Wnx" is easily shown taking B = AT. Kronecker Products A Schur fonn for A EB B can be derived similarly. 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). Sylvester who studied general linear matrix equations of the form equation in honor of J..4) is known as a Lyapunov equation. Kronecker Products Chapter 13. respectively. suppose P and Q are unitary fonn. ® P)] = (/m <8> rA) + (7* (g) /„).5) clearly can be written as the Kronecker sum (Im * A) + (BT ® In). Again..e. Sylvester who studied general linear matrix equations of the fonn k LA.3 13. The following definition is very helpful in completing the writing of (13.. and C e M" xm . an "ordinary" linear system. =C.5) as (B T 0 /„). When symmetric. The following definition is very helpful in completing the writing of (13. When does a solution exist? The first important question to ask regarding (13.4) is known as a Lyapunov equation. = AXi + l:~>j. .5) clearly can be written as the Kronecker sum (1m 0 A) + The coefficient matrix in (13.Xj. solution e IR xn also to be symmetric and (13.3 Application to Sylvester and Lyapunov Equations Application to Sylvester and Lyapunov Equations In this section we study the linear matrix equation In this section we study the linear matrix equation AX+XB=C.3 and Corollary 13.

An equivalent linear system is then solved in which the triangular form equivalent linear system is then solved in which the triangular form of the reduced and can be exploited to solve successively for the columns of a suitably of the reduced A and B can be exploited to solve successively for the columns of a suitably transformed solution matrix X. (A)+ Aj(B) =I 00 for all i.7) has a unique solution if and only if A and . this algorithm takes only 0 (n 3) transformed solution matrix X. Definition 13.19. +00): (with X(0) = C) on [0. elegant connections between matrix theory and stability theory for differential equations. B e Rmxm..18. Aj(A) + A. vec(C) = Using Definition 13.17.16.. .n denote the columns ofC E Rnxm so that C = [ n .10) . where A. They culminate in Theorem 13. n :::: m. this algorithm takes only O(n3 ) operations rather than the O(n6)) that would be required by solving (13. n > m. Application to Sylvester and Lyapunov Equations 13.8) by Theorem 13. say. the linear system (13. Then the Sylvester equation G jRmxm. Let A e lRnxn.17.17.3) (or symmetric Lyapunov equations of the form Sylvester equations of the form (13.B have no eigenvalues in common.. Suppose further that A and B E Rn . ofC e jRnxm [CI.24. 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. the eigenvalues of [(1m ® A) + (BT <8> /„)] are + Mj.5) can be rewritten in the form [(1m ® A) + (B T ® In)]vec(X) = vec(C). one of many The next few theorems are classical. B E Rmxm. Now integrate the differential equation X = AX + X B solution to (13.3. .6).6) if and only if [(1m ® A) + (B T ® In)] is nonsingular. 77ie/i Theorem 13. The next few theorems are classical. (13. Theorem C E jRnxm. Then the (unique) solution of the Sylvester equation parts in the open left half-plane). has a unique solution if and only if A and —B have no eigenvalues in common. They culminate in Theorem 13.3. (13.6) There exists a unique solution to (13. and ^j Theorem 13. First A and B are reduced to (real) Schur form..6) if and only if [(Im ® A) + (BT ® /„)] is nonsingular. We thus have the following theorem. Assuming that. and C e Rnxm. the eigenvalues of [(/m <g> A) + (BT ® In)] are Ai A. But [(Im <8>A) + (B TT ® In)] isisnonsingular ififand only ififitithas no zero eigenvalues.6) directly with Gaussian elimination. 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 half-plane). where From Theorem 13.4)) are generally not solved using the mn x mn "vee" formulation (13. Schur form. and C e R" xm . .X(O) = A 10 roo X(t)dt + ([+00 X(t)dt) 10 B.. Let A e jRnxn.1S.e A (A).. The most commonly preferred numerical algorithm is described in [2]. A(fi). From Theorem 13. E!!. The most (13. E R E jRnxm. But [(1m ® A) + (B (g) /„)] nonsingular and only has no zero eigenvalues.13. so there exists unique Proof: Since A and B are stable.4» are generally not solved using the mn x mn "vec" formulation (13. c ]. i.17. the linear system (13. c E jRn the Then vec(C) is defined to be the mn-vector formed by stacking the columns ofC on top of by C ::~~::~: ::d~~:::O:[]::::fonned "ocking the colunuu of on top of one another..-(B) ^ solution to(13. E jRmxm. Then the (unique) solution of the Sylvester equation AX+XB=C (13.5) can be rewritten in the form Using Definition 13. Let Ci( €