Matrix Theory and

LINEAR ALGEBRA
An open text by Peter Selinger
Based on the original text by Lyryx Learning and Ken Kuttler

Creative Commons License (CC BY)

 Matrix Theory and Linear Algebra
An open text by Peter Selinger
Based on the original text by Lyryx Learning and Ken Kuttler
First edition

CONTRIBUTIONS
Ken Kuttler, Brigham Young University

Ilijas Farah, York University

Marie-Andrée B. Langlois, Dalhousie University
Peter Selinger, Dalhousie University

Lyryx Learning Team

Bruce Bauslaugh Jennifer MacKenzie
Peter Chow Tamsyn Murnaghan
Nathan Friess Bogdan Sava
Stephanie Keyowski Larissa Stone
Claude Laflamme Ryan Yee
Martha Laflamme Ehsun Zahedi

Cover photography by Pok Rie

LICENSE

Creative Commons License (CC BY): This text, including the art and illustrations, are available under
the Creative Commons license (CC BY), allowing anyone to reuse, revise, remix and redistribute the text.
To view a copy of this license, visit https:// reative ommons.org/li enses/by/4.0/
 Revision history
Current revision: Dal 2018 A
This printing: version 487897b of January 10, 2022

Extensive edits, additions, and revisions have been completed by Peter Selinger and other contributors.
Prior extensive edits, additions, and revisions were made by the editorial staff at Lyryx Learning.
All new content (text and images) is released under the same license as noted above.

• P. Selinger: updated title and front matter.

• P. Selinger: Extensive revisions and additions. Rewrote most of Chapters 9–11 from scratch.
Dal 2018 A Added application on perspective rendering, error correcting codes, least squares approximations
and curve fitting, an audio demo for Fourier series, sketching of quadratic forms, and principal
component analysis.

• P. Selinger, M.B. Langlois: Re-ordered chapters. Extensive revisions to Chapters 1–8. Added
Dal 2017 B new sections on fields, cryptography, geometric interpretation of linear transformations, recur-
rences, systems of linear differential equations.

• Lyryx: Front matter has been updated including cover, copyright, and revision pages.
2017 A • I. Farah: contributed edits and revisions, particularly the proofs in the Properties of Determinants
II: Some Important Proofs section.

• Lyryx: The text has been updated with the addition of subsections on Resistor Networks and the
2016 B Matrix Exponential based on original material by K. Kuttler.
• Lyryx: New example on Random Walks developed.

• Lyryx: The layout and appearance of the text has been updated, including the title page and newly
2016 A
designed back cover.

• Lyryx: The content was modified and adapted with the addition of new material and several
2015 A images throughout.
• Lyryx: Additional examples and proofs were added to existing material throughout.

• Original text by K. Kuttler of Brigham Young University. That version is used under
Creative Commons license CC BY (https:// reative ommons.org/li enses/by/3.0/)
2012 A
made possible by funding from The Saylor Foundation’s Open Textbook Challenge. See
Elementary Linear Algebra for more information and the original version.
Contents

Preface 1

1 Systems of linear equations 3

1.1 Geometric view of systems of equations . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Algebraic view of systems of equations . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.3 Elementary operations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.4 Gaussian elimination . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
1.5 Gauss-Jordan elimination . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
1.6 Homogeneous systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
1.7 Uniqueness of the reduced echelon form . . . . . . . . . . . . . . . . . . . . . . . . . . 38
1.8 Fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
1.9 Application: Balancing chemical reactions . . . . . . . . . . . . . . . . . . . . . . . . . 51
1.10 Application: Dimensionless variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
1.11 Application: Resistor networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56

2 Vectors in Rn 61
2.1 Points and vectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
2.2 Addition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
2.3 Scalar multiplication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
2.4 Linear combinations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
2.5 Length of a vector . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
2.6 The dot product . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
2.6.1 Definition and properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
2.6.2 The Cauchy-Schwarz and triangle inequalities . . . . . . . . . . . . . . . . . . . . 81
2.6.3 The geometric significance of the dot product . . . . . . . . . . . . . . . . . . . . 82
2.6.4 Orthogonal vectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
2.6.5 Projections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
2.7 The cross product . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
2.7.1 Right-handed systems of vectors . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
2.7.2 Geometric description of the cross product . . . . . . . . . . . . . . . . . . . . . 91
2.7.3 Algebraic definition of the cross product . . . . . . . . . . . . . . . . . . . . . . . 92
2.7.4 The box product . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95

v
vi CONTENTS

3 Lines and planes in Rn 101

3.1 Lines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
3.2 Planes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110

4 Matrices 121
4.1 Definition and equality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
4.2 Addition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
4.3 Scalar multiplication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
4.4 Matrix multiplication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
4.4.1 Multiplying a matrix and a vector . . . . . . . . . . . . . . . . . . . . . . . . . . 128
4.4.2 Matrix multiplication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
4.4.3 Properties of matrix multiplication . . . . . . . . . . . . . . . . . . . . . . . . . . 137
4.5 Matrix inverses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
4.5.1 Definition and uniqueness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
4.5.2 Computing inverses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
4.5.3 Using the inverse to solve a system of equations . . . . . . . . . . . . . . . . . . . 147
4.5.4 Properties of the inverse . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
4.5.5 Right and left inverses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
4.6 Elementary matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
4.6.1 Elementary matrices and row operations . . . . . . . . . . . . . . . . . . . . . . . 153
4.6.2 Inverses of elementary matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
4.6.3 Elementary matrices and reduced echelon forms . . . . . . . . . . . . . . . . . . 156
4.6.4 Writing an invertible matrix as a product of elementary matrices . . . . . . . . . . 158
4.6.5 More properties of inverses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
4.7 The transpose . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
4.8 Matrix arithmetic modulo p . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
4.9 Application: Cryptography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167

5 Spans, linear independence, and bases in Rn 175

5.1 Spans . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
5.2 Linear independence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
5.2.1 Redundant vectors and linear independence . . . . . . . . . . . . . . . . . . . . . 179
5.2.2 The casting-out algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
5.2.3 Alternative characterization of linear independence . . . . . . . . . . . . . . . . . 182
5.2.4 Properties of linear independence . . . . . . . . . . . . . . . . . . . . . . . . . . 185
5.2.5 Linear independence and linear combinations . . . . . . . . . . . . . . . . . . . . 186
5.2.6 Removing redundant vectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
5.3 Subspaces of Rn . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 190
 CONTENTS vii

5.4 Basis and dimension . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194

5.4.1 Definition of basis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
5.4.2 Examples of bases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196
5.4.3 Bases and coordinate systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198
5.4.4 Dimension . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 200
5.4.5 More properties of bases and dimension . . . . . . . . . . . . . . . . . . . . . . . 202
5.5 Column space, row space, and null space of a matrix . . . . . . . . . . . . . . . . . . . . 210

6 Linear transformations in Rn 217

6.1 Linear transformations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217
6.2 The matrix of a linear transformation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 220
6.3 Geometric interpretation of linear transformations . . . . . . . . . . . . . . . . . . . . . 225
6.4 Properties of linear transformations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 230
6.5 Application: Perspective rendering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235

7 Determinants 241
7.1 Determinants of 2 × 2- and 3 × 3-matrices . . . . . . . . . . . . . . . . . . . . . . . . . 241
7.2 Minors and cofactors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 243
7.3 The determinant of a triangular matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . 248
7.4 Determinants and row operations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 250
7.5 Properties of determinants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 253
7.6 Application: A formula for the inverse of a matrix . . . . . . . . . . . . . . . . . . . . . 258
7.7 Application: Cramer’s rule . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 266

8 Eigenvalues, eigenvectors, and diagonalization 269

8.1 Eigenvectors and eigenvalues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 269
8.2 Finding eigenvalues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 273
8.3 Geometric interpretation of eigenvectors . . . . . . . . . . . . . . . . . . . . . . . . . . 281
8.4 Diagonalization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 284
8.5 Application: Matrix powers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 290
8.6 Application: Solving recurrences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 293
8.7 Application: Systems of linear differential equations . . . . . . . . . . . . . . . . . . . . 299
8.7.1 Differential equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 299
8.7.2 Systems of linear differential equations . . . . . . . . . . . . . . . . . . . . . . . 300
8.7.3 Example: coupled train cars . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 302
8.8 Application: The matrix exponential . . . . . . . . . . . . . . . . . . . . . . . . . . . . 308
8.9 Properties of eigenvectors and eigenvalues . . . . . . . . . . . . . . . . . . . . . . . . . 311
8.10 The Cayley-Hamilton Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 315
viii CONTENTS

8.11 Complex eigenvalues and eigenvectors . . . . . . . . . . . . . . . . . . . . . . . . . . . 317

9 Vector spaces 325

9.1 Definition of vector spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 325
9.2 Linear combinations, span, and linear independence . . . . . . . . . . . . . . . . . . . . 334
9.3 Subspaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 342
9.4 Basis and dimension . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 349
9.5 Application: Error correcting codes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 355

10 Linear transformation of vector spaces 369

10.1 Definition and examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 369
10.2 The algebra of linear transformations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 375
10.3 Linear transformations defined on a basis . . . . . . . . . . . . . . . . . . . . . . . . . . 378
10.4 The matrix of a linear transformation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 381

11 Inner product spaces 387

11.1 Real inner product spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 387
11.2 Orthogonality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 394
11.3 The Gram-Schmidt orthogonalization procedure . . . . . . . . . . . . . . . . . . . . . . 403
11.4 Orthogonal projections and Fourier series . . . . . . . . . . . . . . . . . . . . . . . . . . 413
11.5 Application: Least squares approximations and curve fitting . . . . . . . . . . . . . . . . 424
11.6 Orthogonal functions and orthogonal matrices . . . . . . . . . . . . . . . . . . . . . . . 432
11.7 Diagonalization of symmetric matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . 437
11.8 Positive semidefinite and positive definite matrices . . . . . . . . . . . . . . . . . . . . . 442
11.9 Application: Simplification of quadratic forms . . . . . . . . . . . . . . . . . . . . . . . 448
11.10 Complex inner product spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 456
11.11 Unitary and hermitian matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 466
11.12 Application: Principal component analysis . . . . . . . . . . . . . . . . . . . . . . . . . 472

A Complex numbers 485

A.1 The complex numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 485
A.2 Geometric interpretation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 489
A.3 The fundamental theorem of algebra . . . . . . . . . . . . . . . . . . . . . . . . . . . . 491

B Answers to selected exercises 495

Index 541
Preface
Matrix Theory and Linear Algebra is an introduction to linear algebra for students in the first or second
year of university. The book contains enough material for a 2-semester course. Major topics of linear
algebra are presented in detail, and many applications are given. Although it is not a proof-oriented book,
proofs of most important theorems are provided.
Each section begins with a list of desired outcomes which a student should be able to achieve upon
completing the chapter. Throughout the text, examples and diagrams are given to reinforce ideas and
provide guidance on how to approach various problems. Students are encouraged to work through the
suggested exercises provided at the end of each section. Selected solutions to these exercises are given at
the end of the text.

Open text

This is an open text, licensed under the Creative Commons “CC BY 4.0” license. This means, among
other things, that you are permitted to copy and redistribute this textbook in any medium or format. For
example, you can download this textbook for free, print copies for yourself or others, or share it on the
internet.
The license also permits making changes. This is ideal for instructors who would like to add their
own material, change notations, or add more examples or exercises. If you make revisions, please
send them to me so that I can consider incorporating them in future versions of this book. Please see
https:// reative ommons.org/li enses/by/4.0/ for details of the licensing terms.
This textbook has a website at https://www.mathstat.dal. a/~selinger/linear-algebra/ .
There, you can find the most up-to-date version. The website also contains supplementary material, a link
to the source code and license, options for purchasing a printed version of this book, and more.

Reporting typos

Like all books, this book likely contains some typos and other errors. However, since it is an open text,
typos can easily be fixed and an updated version posted online. It is my intention to fix all typos. If
you find a typo (no matter how small), please report it to me at [email protected]. Thanks to the
following people who have already reported typos: Yaser Alkayale, Hassaan Asif, Courtney Baumgartner,
Kieran Bhaskara, Serena Drouillard, Robert Earle, Warren Fisher, Esa Hannila, Melissa Huggan, Xiaoyu
Jia, Arman Kerimbek, Peter Lake, Marie-Andrée Langlois, Brenda Le, Sarah Li, Ian MacIntosh, Li Wei
Men, Deklan Mengering, Dallas Sawtell, Alain Schaerer, Yi Shu, Bruce Smith, Asmita Sodhi, Michael St
Denis, Daniele Turchetti, Liu Yuhao, and Ziqi Zhang.

1
1. Systems of linear equations

1.1 Geometric view of systems of equations

Outcomes
A. Relate the types of solution sets of a system of two (three) variables to the intersections of
lines in a plane (the intersection of planes in 3-dimensional space)

As you may remember, linear equations like 2x + 3y = 6 can be graphed as straight lines in the coordinate
plane. We say that this equation is in two variables, in this case x and y. Suppose you have two such
equations, each of which can be graphed as a straight line, and consider the resulting graph of two lines.
What would it mean if there exists a point of intersection between the two lines? This point, which lies
on both graphs, gives x and y values for which both equations are true. In other words, this point gives the
ordered pair (x, y) that satisfies both equations. If the point (x, y) is a point of intersection, we say that (x, y)
is a solution to the two equations. In linear algebra, we often are concerned with finding the solution(s)
to a system of equations, if such solutions exist. First, we consider graphical representations of solutions
and later we will consider the algebraic methods for finding solutions.
When looking for the intersection of two lines in the plane, several situations may arise. The follow-
ing picture demonstrates the possible situations when considering two equations (two lines in the plane)
involving two variables.
y y y

x x x
One solution No solutions Infinitely many solutions

In the first diagram, there is a unique point of intersection, which means that there is only one (unique)
solution to the two equations. In the second, there are no points of intersection and no solution. There is
no solution because the two lines are parallel and they never intersect. The third situation that can occur,
as demonstrated in diagram three, is that the two lines are really the same line. For example, x + y = 1
and 2x + 2y = 2 are two equations that yield the same line when graphed. In this case there are infinitely
many points that are solutions of these two equations, as every ordered pair which is on the graph of the
line satisfies both equations.
When considering linear systems of equations, there are always three possibilities for the number of
solutions: there is exactly one solution, there are infinitely many solutions, or there is no solution. When

3
4 Systems of linear equations

we speak of solving a system of equations, we usually mean finding all of its solutions. This can mean
finding one solution (if the solution is unique), finding infinitely many solutions, or finding that there is no
solution.

Example 1.1: A graphical solution

Use a graph to solve the following system of equations:

x+y = 3
y − x = 5.

Solution. Through graphing the above equations and identifying the point of intersection, we can find the
solution(s). Remember that we must have either one solution, infinitely many, or no solutions at all. The
following graph shows the two equations, as well as the intersection. Remember, the point of intersection
represents the solution of the two equations, or the (x, y) which satisfy both equations. In this case, there
is one point of intersection at (−1, 4) which means we have one unique solution, x = −1, y = 4.
y
6 y−x = 5
5
(x, y) = (−1, 4) 4
3
2 x+y = 3
1

x
−4 −3 −2 −1 1
♠
In the above example, we investigated the intersection point of two equations in two variables, x and y.
Now we will consider the graphical solutions of three equations in two variables.
Consider a system of three equations in two variables. Again, these equations can be graphed as
straight lines in the plane, so that the resulting graph contains three straight lines. Recall the three possi-
bilities for the number of solutions: no solution, one solution, and infinitely many solutions. With three
lines, there are more complex ways of achieving these situations. For example, you can imagine the case
of three intersecting lines having no common point of intersection. Perhaps you can also imagine three
intersecting lines which do intersect at a single point. These two situations are illustrated below.
y y

x x
No solution One solution
 1.1. Geometric view of systems of equations 5

Consider the first picture above. While all three lines intersect with one another, there is no common
point of intersection where all three lines meet at one point. Hence, there is no solution to the three
equations. Remember, a solution is a point (x, y) which satisfies all three equations. In the case of the
second picture, the lines intersect at a common point. This means that there is one solution to the three
equations whose graphs are the given lines. You should take a moment now to draw the graph of a system
which results in three parallel lines. Next, try the graph of three identical lines. Which type of solution is
represented in each of these graphs?
We have now considered the graphical solutions of systems of two equations in two variables, as well
as three equations in two variables. However, there is no reason to limit our investigation to equations in
two variables. We will now consider equations in three variables.
You may recall that equations in three variables, such as 2x + 4y − 5z = 8, form a plane. Above, we
were looking for intersections of lines in order to identify any possible solutions. When graphically solving
systems of equations in three variables, we look for intersections of planes. These points of intersection
give the (x, y, z) that satisfy all the equations in the system. What types of solutions are possible when
working with three variables? Consider the following picture involving two planes, which are given by
two equations in three variables.

Notice how these two planes intersect in a line. This means that the points (x, y, z) on this line satisfy
both equations in the system. Since the line contains infinitely many points, this system has infinitely
many solutions.
It could also happen that the two planes fail to intersect. However, is it possible to have two planes
intersect at a single point? Take a moment to attempt drawing this situation, and convince yourself that it
is not possible! This means that when we have only two equations in three variables, there is no way to
have a unique solution! Hence, the only possibilities for the number of solutions of two equations in three
variables are no solution or infinitely many solutions.
Now imagine adding a third plane. In other words, consider three equations in three variables. What
types of solutions are now possible? Consider the following diagram.
6 Systems of linear equations

In this diagram, there is no point which lies in all three planes. There is no intersection between all
three planes so there is no solution. The picture illustrates the situation in which the line of intersection
of the new plane with one of the original planes forms a line parallel to the line of intersection of the first
two planes. However, in three dimensions, it is possible for two lines to fail to intersect even though they
are not parallel. Such lines are called skew lines.
Recall that when working with two equations in three variables, it was not possible to have a unique
solution. Is it possible when considering three equations in three variables? In fact, it is possible, and we
demonstrate this situation in the following picture.

In this case, the three planes have a single point of intersection. Can you think of other possibilities?
Another is that the three planes could intersect in a line, resulting in infinitely many solutions, as in the
following diagram.

We have now seen how three equations in three variables can have no solution, a unique solution, or
intersect in a line resulting in infinitely many solutions. It is also possible that all three equations describe
the same plane, which also leads to infinitely many solutions.
You can see that when working with equations in three variables, there are many more possibilities for
achieving solutions (or no solutions) than when working with two variables. It may prove enlightening to
spend time imagining (and drawing) many possible scenarios, and you should take some time to try a few.
You should also take some time to imagine (and draw) graphs of systems in more than three variables.
Equations like x + y − 2z + 4w = 8 with more than three variables are often called hyperplanes. You
may soon realize that it is tricky to draw the graphs of hyperplanes! In fact, most people cannot visualize
more than three dimensions. Fortunately, through the tools of linear algebra, we can examine systems
of equations in four variables, five variables, or even hundreds or thousands of variables, without ever
needing to graph them. Instead we will use algebra to manipulate and solve these systems of equations.
We will introduce these algebraic tools in the following sections.
 1.2. Algebraic view of systems of equations 7

Exercises

Exercise 1.1.1 Graphically, find the point (x, y) which lies on both of the lines x + 3y = 1 and 4x − y = 3.
That is, graph each line and see where they intersect.

Exercise 1.1.2 Graphically, find the point of intersection of the two lines 3x + y = 3 and x + 2y = 1. That
is, graph each line and see where they intersect.

Exercise 1.1.3 You have a system of k equations in two variables, k ≥ 2. Explain the geometric signifi-
cance of

(a) No solution.
(b) A unique solution.
(c) An infinite number of solutions.

Exercise 1.1.4 Draw a picture of three planes such that no two of the planes are parallel, but the three
planes have no common intersection.

1.2 Algebraic view of systems of equations

Outcomes
A. Recognize the difference between a linear equation and a non-linear equation.

B. Determine whether a tuple of real numbers is a solution for a system of linear equations.

C. Understand what it means for a system of linear equations to be consistent or inconsistent.

We have taken an in-depth look at graphical representations of systems of equations, as well as how to
find possible solutions graphically. Our attention now turns to working with systems algebraically.

Definition 1.2: Linear equation

A linear equation is an equation of the form

a1 x1 + a2 x2 + . . . + an xn = b.

Here, a1 , . . . , an are real numbers called the coefficients of the equation, b is a real number called
the constant term of the equation, and x1 , . . . , xn are variables.
8 Systems of linear equations

Real numbers, such as the coefficients a1 , . . . , an , the constant term b, or the values of the variables
x1 , . . . , xn , will also be called scalars. For now, the word “scalar” is just a synonym for “real number”.
Later, in Section 1.8, we will discover other kinds of scalars.

Example 1.3: Linear vs. non-linear equation

Which of the following equations are linear?

2x + 3y = 5
2
√ + 3y = 5
2x
2 x + 3y = 5
√
( 2)x + 3y = 52

Solution. The equation 2x + 3y = 5 is linear. The equation 2x√2 + 3y = 5 is not linear, because it contains

the square of a variable instead of a variable. The equation 2 x + 3y = 5 is also√ not linear, because the
2
√ is applied to one of the variables. On the other hand, the equation ( 2)x + 3y = 5 is linear,
square root
because 2 and 52 are real numbers, and can therefore be used as coefficients and constant terms. ♠
We also permit minor notational variants of linear equations. The equation 2x − 3y = 5 is linear although
Definition 1.2 does not mention subtraction, because it can be regarded as just another notation for 2x +
(−3)y = 5. Similarly, the equation 2x = 5 + 3y can be regarded as linear, because it can be easily rewritten
as 2x − 3y = 5 by bringing all the variables (and their coefficients) to the left-hand side. When we need to
emphasize that some linear equation is literally of the form a1 x1 + a2 x2 + . . . + an xn = b, we say that the
equation is in standard form. Thus, the standard form of the equation 2x = 5 + 3y is 2x + (−3)y = 5.
A solution to a linear equation is an assignment of real numbers to the variables, making the equation
true. More precisely, if r1 , . . . , rn are real numbers, the assignment x1 = r1 , . . . , xn = rn is a solution to the
equation in Definition 1.2 if the real number a1 r1 + a2 r2 + . . . + an rn is equal to the real number b. To save
space, we often write solutions in tuple notation1 as (x1 , . . . , xn ) = (r1 , . . . , rn ). When there is no doubt
about the order of the variables, we also often simply write the solution as (r1 , . . . , rn ).

Example 1.4: Solutions of a linear equation

Consider the linear equation 2x + 3y − 4z = 5. Which of the following are solutions? (a) (x, y, z) =
(1, 1, 0), (b) (x, y, z) = (0, 3, 1), (c) (x, y, z) = (1, 1, 1).

Solution. The assignment (x, y, z) = (1, 1, 0) is a solution because 2(1) + 3(1) − 4(0) = 5. The assignment
(x, y, z) = (0, 3, 1) is also a solution, because 2(0) + 3(3) − 4(1) = 5. On the other hand, (x, y, z) = (1, 1, 1)
is not a solution, because 2(1) + 3(1) − 4(1) = 1 6= 5. ♠

A system of linear equations is just several linear equations taken together.

1 The terminology “tuple” arose as follows. A collection of two items is called a “pair”, a collection of three items is called
a “triple”, followed by “quadruple”, “quintuple”, “sextuple”, and so on. You have to know Latin to know what the next ones
are called. To avoid these Latin terms, mathematicians started saying 4-tuple, 5-tuple, 6-tuple and so on, and more generally,
n-tuple for an ordered collection of n items. When n doesn’t matter or is clear from the context, we often just say “tuple”.
 1.2. Algebraic view of systems of equations 9

Definition 1.5: System of linear equations

A system of linear equations is a list of equations

a11 x1 + a12 x2 + . . . + a1n xn = b1

a21 x1 + a22 x2 + . . . + a2n xn = b2
...
am1 x1 + am2 x2 + . . . + amn xn = bm ,

where ai j and bi are scalars (i.e., real numbers). The above is a system of m equations in the n
variables, x1 , x2 . . . , xn . As before, the numbers ai j are called the coefficients and the numbers bi are
called the constant terms of the system of equations.

The relative size of m and n is not important here. We may have more variables than equations, more
equations than variables, or an equal number of equations and variables.
A solution to a system of linear equations is an assignment of real numbers to the variables that is a
solution to all of the equations in the system.

Example 1.6: Solutions of a system of linear equations

Consider the system of linear equations

2x + 3y − 4z = 5
−2x + y + 2z = −1.

Which of the following are solutions of the system? (a) (x, y, z) = (1, 1, 0), (b) (x, y, z) = (6, 3, 4), (c)
(x, y, z) = (0, 3, 1).

Solution. The assignment (x, y, z) = (1, 1, 0) is a solution of this system of equations, because it is a
solution to the first equation and the second equation. Also, (x, y, z) = (6, 3, 4) is another solution of this
system of equations (check this!). On the other hand, (x, y, z) = (0, 3, 1) is not a solution of the system,
because although it is a solution to the first equation, it is not a solution to the second equation. ♠
Recall from Section 1.1 that a system of equations either has a unique solution, infinitely many solutions,
or no solution. It is very important to us whether a system of equations has solutions or not. For this
reason, we introduce the following terminology:

Definition 1.7: Consistent and inconsistent systems

A system of linear equations is called consistent if there exists at least one solution. It is called
inconsistent if there is no solution.

If we think of each equation as a condition that must be satisfied by the variables, consistent means that
there is some choice of values for the variables which can satisfy all of the conditions. Inconsistent means
that there is no such choice of values for the variables. In the following sections, you will learn a method
for determining whether a system of equations is consistent or not, and in case it is consistent, to find all
of its solutions.
10 Systems of linear equations

Exercises

Exercise 1.2.1 Which of the following equations are linear?

(a) 2x − 3y + 4z = −10

(b) 2.123x1 + 5.541x2 − 9.101x3 = 11.012

(c) x2 + y2 + z2 = 1

(d) √1 x + 43 y = sin( π3 )
2

(e) x + yz = 3

Exercise 1.2.2 Consider the system of equations

x + 2y + 3z + 4w = 4
x+ y + z + w = 2
x + 2y + 2z + w = 2.

For each of the following tuples (x, y, z, w) of real numbers, determine whether it is a solution of the first
equation, second equation, and/or third equation. Which ones are solutions to the system of equations?

(a) (2, 0, −2, 2) (b) (2, 2, −2, 0) (c) (1, 1, −1, 1) (d) (3, 0, −1, 1) (e) (2, −2, 2, 0)

1.3 Elementary operations

Outcomes
A. Use elementary operations to simplify a system of equations.

B. Solve some systems of equations by back substitution.

C. Write a system of equations in augmented matrix form.

D. Perform elementary row operations on augmented matrices.

Our strategy for solving systems of linear equations is to successively transform a difficult system of
equations into a simpler equivalent system. Here, by an “equivalent” system of equations we mean one
that has the same solutions as the original one. We will perform the process of simplifying a system of
equations by applying certain basic steps called “elementary operations”.
 1.3. Elementary operations 11

Definition 1.8: Equivalent systems

Two systems of equations are called equivalent if they have the same solutions. This means that
every solution of the first system is also a solution of the second system, and every solution of the
second system is also a solution of the first system.

How can we know whether two systems of equations are equivalent? It turns out that the following basic
operations always transform a system of equations into an equivalent system. In fact, these operations are
the key tool we use in linear algebra to solve systems of equations.

Definition 1.9: Elementary operations

Elementary operations are the following operations:

1. Interchange the order in which the equations are listed.

2. Multiply any equation by a non-zero scalar.

3. Add a multiple of one equation to another equation.

The most important property of the elementary operations is that they do not change the solutions to the
system of equations. Before proving that this is true in general, we will first verify it in an example.

Example 1.10: Equivalent systems

Show that the systems
x + 2y = 7
−2x = −6
and
x + 2y = 7
4y = 8
are equivalent.

Solution. We can see that the second system is obtained from the first one by applying an elementary
operation, namely, adding 2 times the first equation to the second equation:

−2x + 2(x + 2y) = −6 + 2(7)

By simplifying, we obtain 4y = 8.
To verify that the two systems are indeed equivalent, let us first solve the first system. From the
second equation, we see that x = 3. Substituting x = 3 into the first equation, the equation becomes
3 + 2y = 7, which we can solve to find y = 2. Therefore, the only solution to the first system of equations
is (x, y) = (3, 2).
Now let us solve the second system. From the second equation, we find that y = 2. Substituting y = 2
into the first equation, we get x + 4 = 7, which we can solve to find x = 3. Therefore, the only solution to
12 Systems of linear equations

the second system of equations is (x, y) = (3, 2). Since the two systems have the same solutions, they are
equivalent. ♠
This example illustrates how an elementary operation applied to a system of two equations in two variables
does not affect the set of solutions. The same is true for any size of system in any number of variables.
In the following theorem, we use the notation Ei to represent the left-hand side of an equation, while bi
denotes a constant term.

Theorem 1.11: Elementary operations and solutions

Suppose you have a system of two linear equations in any number of variables

E1 = b 1
(1.1)
E2 = b 2 .

Then the following systems are equivalent to (1.1):

1.
E2 = b 2
(1.2)
E1 = b 1 .

2.
E1 = b 1
(1.3)
kE2 = kb2
for any scalar k, provided k 6= 0.

3.
E1 = b 1
(1.4)
E2 + kE1 = b2 + kb1
for any scalar k (including k = 0).

Proof.

1. By definition, a solution of (1.1) is an assignment of scalars to the variables that is a solution to

E1 = b1 and to E2 = b2 . But that is exactly the same thing as a solution of (1.2).

2. To prove that the systems (1.1) and (1.3) have the same solution set, let (x1 , . . . , xn ) be any solution
of (1.1). Then E1 = b1 and E2 = b2 are both true. Multiplying both sides of the last equation by k,
we know that kE2 = kb2 is true, and so (x1 , . . . , xn ) is a solution of (1.3). Conversely, let (x1 , . . . , xn )
be any solution of (1.3). Then E1 = b1 and kE2 = kb2 are true. Because k 6= 0, we are allowed to
divide both sides of the last equation by k, and therefore E2 = b2 is true. Hence, (x1 , . . . , xn ) is also
a solution of (1.1). Since we have shown that every solution of (1.1) is a solution of (1.3) and vice
versa, the two systems are equivalent.

3. To prove that the systems (1.1) and (1.4) have the same solution set, let (x1 , . . . , xn ) be any solution
of (1.1). Then E1 = b1 and E2 = b2 are both true. We multiply both sides of the first equation by k
to obtain kE1 = kb1 . Then kE1 + E2 = kb1 + b2 , and hence (x1 , . . . , xn ) is a solution of (1.4). For the
converse direction, assume (x1 , . . . , xn ) is a solution of E1 = b1 and kE1 + E2 = kb1 + b2 . From the
 1.3. Elementary operations 13

first equation, we have kE1 = kb1 , and subtracting this from the second equation, we get E2 = b2 ,
hence (x1 , . . . , xn ) is a solution of (1.1). Note that unlike in case 2., there was no need to divide by k,
and therefore it was not necessary to require k 6= 0.

♠
We will now use elementary operations to solve a system of three equations and three variables.

Example 1.12: Solving a system of equations with elementary operations

Solve the system of equations
x + 3y + 6z = 25
2x + 7y + 14z = 58
2y + 5z = 19.

Solution. By Theorem 1.11, we can do elementary operations on this system without changing the solution
set. We will therefore use elementary operations to try to simplify the system of equations. First, we add
(−2) times the first equation to the second equation. This yields the system

x + 3y + 6z = 25
y + 2z = 8
2y + 5z = 19.

Next, we add (−2) times the second equation to the third equation. This yields the system

x + 3y + 6z = 25
y + 2z = 8 (1.5)
z = 3.

At this point, it is easy to find the solution. The last equation tells us that z = 3. We can substitute this
value of z back into the second equation to get

y + 2(3) = 8,

which we can simplify and solve for y to find that y = 2. Finally, we can substitute the values z = 3 and
y = 2 back into the first equation to get

x + 3(2) + 6(3) = 25.

Simplifying and solving for x, we find that x = 1. Hence, the solution to the system is (x, y, z) = (1, 2, 3).
The process we followed for solving (1.5) by first computing z, then y, then x is called back substitu-
tion. Alternatively, we could have continued from (1.5) with more elementary operations as follows. Add
(−2) times the third equation to the second and then add (−6) times the third to the first. This yields

x + 3y =7
y =2
z = 3.
14 Systems of linear equations

Now add (−3) times the second to the first. This yields

x =1
y =2
z = 3,

a system which has the same solution set as the original system. This second method avoided back substi-
tution and led to the same solution set. It is your decision which you prefer to use, as both methods lead
to the correct solution, (x, y, z) = (1, 2, 3). ♠
Note how we have written each system of equations so that “like” variables line up on columns: one
column for x, one column for y, and one column for z. This makes it easier to perform elementary
operations. It is often useful to simplify the notation further, writing systems of equations in augmented
matrix notation. Recall the system of equations from Example 1.12:

x + 3y + 6z = 25
2x + 7y + 14z = 58
2y + 5z = 19.

This system can be written as an augmented matrix as follows:

 
1 3 6 25
 2 7 14 58  .
0 2 5 19

A matrix is just a 2-dimensional array of numbers. An augmented matrix has two parts separated by a
vertical line. Notice that the augmented matrix notation has exactly the same information as the original
system of equations. All the coefficients are written on the left side of the vertical line, and all the constant
terms are written on the right side of the vertical line. These two parts of the augmented matrix are also
called the coefficient matrix and the constantmatrix. Each rowof the augmented matrix corresponds to
one linear equation. For example, the top row 1 3 6 | 25 corresponds to the equation

x + 3y + 6z = 25.

Each column of  the coefficient

 matrix contains the coefficients for one particular variable. For example,
1
the first column  2  contains all of the coefficients for the variable x. If a variable does not appear in
0
an equation, the corresponding coefficient is 0. In general, the augmented matrix of a linear system of
equations is defined as follows.
 1.3. Elementary operations 15

Definition 1.13: Augmented matrix of a system of linear equations

The augmented matrix of the system of linear equations

a11 x1 + . . . + a1n xn = b1
.
..
am1 x1 + . . . + amn xn = bm

is  
a11 · · · a1n b1
 .. ... ... 
 . .
am1 · · · amn bm

We can consider elementary operations in the context of the augmented matrix. The elementary operations
can be used on the rows of an augmented matrix just as we used them on equations previously. For
example, instead of adding a multiple of one equation to another, we will now be adding a multiple of one
row to another. Note that Theorem 1.11 implies that any elementary row operation used on an augmented
matrix will not change the solutions to the corresponding system of equations. For reference, here are the
three kinds of elementary row operations, along with a shorthand notation we are going to use for them.

Definition 1.14: Elementary row operations

The elementary row operations are the following:

1. Switch two rows. (Notation: Ri ↔ R j to switch rows i and j).

2. Multiply a row by a non-zero number. (Notation: Ri ← kRi to multiply row i by k).

3. Add a multiple of one row to another row. (Notation: Ri ← Ri + kR j to add k times row j to
row i.)

We write “≃” to indicate that two augmented matrices are equivalent, i.e., that the corresponding systems
of equations have the same set of solutions.

Example 1.15: Elementary row operations

Repeat the calculations of Example 1.12, using the notations we just introduced.

Solution. We have:
     
1 3 6 25 1 3 6 25 1 3 6 25
R ← R −2R
 2 7 14 58  R2 ← ≃
R2 −2R1
 0 1 2 8  3 ≃3 2  0 1 2 8  .
0 2 5 19 0 2 5 19 0 0 1 3
The final augmented matrix corresponds to the system
x + 3y + 6z = 25
y + 2z = 8
z = 3,
16 Systems of linear equations

which is the same as (1.5). We can solve it by back substitution to obtain the solution x = 1, y = 2, and
z = 3.
Alternatively, we can continue with additional row operations:
     
1 3 6 25 R1 ← R1 −6R3 1 3 0 7 1 0 0 1
R ← R −2R
 0 1 2 8  2 ≃2 3  0 1 0 2  R1 ← ≃ R1 −3R2
 0 1 0 2 .
0 0 1 3 0 0 1 3 0 0 1 3

Notice how this notation is much more succinct than what we used in Example 1.12. ♠
We end this section with a final word of caution: logically, you can only perform one elementary row
operation at a time. For example, it would not be correct to simultaneously add R1 to R2 and add R2 to R1 .
What is permitted is to first add R1 to R2 , then then add the new R2 to R1 . Although we may sometimes try
to save space by skipping an intermediate step, as in the last example where we applied the row operations
R1 ← R1 −6R3 and R2 ← R2 −2R3 in one step, it is important to realize that logically, each row operation
must be performed separately before the next one can be done. When in doubt, the only safe course of
action is not to skip any steps.

Exercises

Exercise 1.3.1 Use elementary operations to solve the system of equations

3x + y = 3
x + 2y = 1.

Exercise 1.3.2 Use elementary operations to find the point (x, y) that lies on both lines x + 3y = 1 and
4x − y = 3.

Exercise 1.3.3 Use elementary operations to determine whether the three lines x + 2y = 1, 2x − y = 1,
and 4x + 3y = 3 have a common point of intersection. If so, find the point, and if not, tell why they don’t
have such a common point of intersection.

Exercise 1.3.4 Do the three planes, x + y − 3z = 2, 2x + y + z = 1, and 3x + 2y − 2z = 0 have a common

point of intersection? If so, find one and if not, tell why there is no such point.

Exercise 1.3.5 Solve the following system of equations by back substitution.

x + 3y − 2z = 5
y + 3z = 4
z = 1.

Exercise 1.3.6 Write the following system of linear equations as an augmented matrix. Caution: you first
have to simplify and rearrange the equations so that “like” variables are lined up in columns. Write the
 1.4. Gaussian elimination 17

variables in the order x, y, z.

x − 3z + 2y = 5
6−x = 4+y−z
2x + 3 = x + 3y.

Exercise 1.3.7 Four times the weight of Gaston is 150kg more than the weight of Ichabod. Four times
the weight of Ichabod is 660kg less than seventeen times the weight of Gaston. Four times the weight of
Gaston plus the weight of Siegfried equals 290kg. Brunhilde would balance all three of the others. Find
the weights of the four people.

1.4 Gaussian elimination

Outcomes
A. Find the echelon form of a matrix.

B. Determine whether a system of linear equations has no solution, a unique solution or an

infinite number of solutions from its echelon form.

C. Solve a system of linear equations using Gaussian elimination and back substitution.

D. Find the rank of a matrix.

E. Determine whether a consistent system of linear equations has a unique solution or an infinite
number of solutions from its rank.

In the previous section, we saw examples of how to solve a system of equations using elementary row
operations (and sometimes back substitution). But it is not clear whether every system of equations can be
solved this way. How do we know which elementary row operation to apply next? In this section, you will
learn a procedure called Gaussian elimination by which every system of linear equations can be solved
systematically.
Before we start, let’s figure out what it means to be “done”. At what point should we stop performing
row operations? The answer is that we will stop performing row operations when the system of equations
is in a special form called echelon form, which we now define.
18 Systems of linear equations

Definition 1.16: Echelon form

An entry of an augmented matrix is called a leading entry or pivot entry if it is the leftmost non-
zero entry of a row. An augmented matrix is in echelon form (also called row echelon form)
if

1. All rows of zeros are below all non-zero rows.

2. Each leading entry of a row is in a column to the right of the leading entry of any row above
it.

A column containing a pivot entry is also called a pivot column.

The word echelon comes from French échelle, which means ladder. This is because an echelon form looks
a bit like a ladder or staircase. Here are some examples of echelon forms.

Example 1.17: Matrices in echelon form

The following augmented matrices are in echelon form. We have circled the pivot entries for clarity.
     
0 5 2 1 3 5 1 4 0 0 5 0 3 0 6 2
 0 0 0 0 1 6   0 0 1 0 2 0   0 1 4 0 
     
 0 0 0 0 0 0 ,  0 0 0 1 4 0 ,  0 0 2 1 
0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0

Example 1.18: Not in echelon form

The following augmented matrices are not in echelon form.
 
0 0 0 0  
  0 2 3 3
1 2 3 3  1 2 3
  1 5 0 2 
 0 1 0 2  ,  2 4 −6  ,  
   0 0 1 0 
 0 0 0 1  4 0 7
0 0 0 0
0 0 0 0
In the first matrix, a row of zeros is above a non-zero row. In the second and third matrix, the leading
entries of some rows are not to the right of the leading entries of previous rows.

An augmented matrix can always be converted to echelon form by using elementary row operations. The
following algorithm shows how to do this.
 1.4. Gaussian elimination 19

Algorithm 1.19: Gaussian elimination

This algorithm provides a method for using row operations to take a matrix to its echelon form. We
begin with the matrix in its original form.

1. Starting from the left, find the first non-zero column. This is the first pivot column, and the
position at the top of this column will be the position of the first pivot entry. Switch rows if
necessary to place a non-zero number in the first pivot position.

2. Use row operations to make the entries below the first pivot entry (in the first pivot column)
equal to zero.

3. Ignoring the row containing the first pivot entry, repeat steps 1 and 2 with the remaining rows.
Repeat the process until there are no more non-zero rows left.

Most often we will apply this algorithm in order to solve a system of linear equations. This works by first
converting the system to echelon form, then using back substitution to find the solutions. The next few
examples show how to do this.

Example 1.20: Solving a system of equations: one solution

Solve the following system of equations:

x + 4y + 3z = 11
2x + 10y + 7z = 27
x + y + 2z = 5.

Solution. The augmented matrix for this system is

 
1 4 3 11
 2 10 7 27  .
1 1 2 5
In order to find the solution(s) to this system, we first use Algorithm 1.19 to carry the augmented matrix
to echelon form. Notice that the first column is non-zero, so this is our first pivot column. The first entry
in the first row, 1, is the first pivot entry. We will use row operations to create zeros in the entries below
the 1.
     
1 4 3 11 R ← R −2R 1 4 3 11 R ← R −R 1 4 3 11
 2 10 7 27  2 ≃2 1  0 2 1 5  3 ≃3 1  0 2 1 5 
1 1 2 5 1 1 2 5 0 −3 −1 −6
Now the entries in the first column below the pivot position are zeros. We now look for the second pivot
column, which in this case is column two. Here, the 2 in the second row and second column is in the pivot
entry. We could create a zero below the 2 with a single row operation by adding 32 times the second row
from the third row. But it is sometimes more convenient not to work with fractions, and therefore we start
instead by multiplying the third row by 2.
     
1 4 3 11 1 4 3 11 1 4 3 11
R3 ← 2R3 R3 ← R3 +3R2
 0 2 1 5  ≃  0 2 1 5  ≃  0 2 1 5 
0 −3 −1 −6 0 −6 −2 −12 0 0 1 3
20 Systems of linear equations

The final matrix is our desired echelon form.

 
1 4 3 11
 0 2 1 5  (1.6)
0 0 1 3
Now we do back substitution to solve for z, y, and then x. The last equation of the echelon form gives us
z = 3. Substituting this into the second equation, we get 2y + 3 = 5, which we can solve for y to get y = 1.
Finally, substituting y = 1 and z = 3 into the first equation, we have x + 4(1) + 3(3) = 11, which we can
solve to get x = −2. Therefore we have the solution (x, y, z) = (−2, 1, 3).
At this point, it is a good idea to double-check this solution using the original equations
x + 4y + 3z = 11
2x + 10y + 7z = 27
x + y + 2z = 5.
For example, we can double-check that (−2) + 4(1) + 3(3) is indeed equal to 11, and similarly for the
other two equations. Double-checking the solution against the original equations is an excellent way to
guard against any errors that might have happened during the row operations or back substitution.
Finally, we note that (x, y, z) = (−2, 1, 3) is the only solution to this system of equations. There cannot
be any other solutions, because by Theorem 1.11, any solution of the original system of equations would
also have to be a solution of (1.6), and the back substitution leaves us no choice except z = 3, y = 1, and
x = −2. ♠

Example 1.21: Solving a system of equations: no solution

Solve the following system of equations:

y + 2z = 2
2x + y − 2z = 3
4x − y − 10z = 4.

Solution. The augmented matrix for this system is

 
0 1 2 2
 2 1 −2 3  .
4 −1 −10 4
We use Algorithm 1.19 to carry the augmented matrix to echelon form. The first column is non-zero and
will be the first pivot column. We switch the first two rows to move a non-zero number into the pivot
position:  
R2 ↔ R1
2 1 −2 3
≃  0 1 2 2 .
4 −1 −10 4
To create zeros below the pivot entry, we subtract 2 times the first row from the third row:
 
R3 ← R3 −2R1
2 1 −2 3
≃  0 1 2 2 .
0 −3 −6 −2
 1.4. Gaussian elimination 21

This finishes the first column. The second pivot column will be column two, with the 1 in the second row
and column as the pivot entry. We add 3 times the second row to the third row to create a zero below the
pivot:  
2 1 −2 3
R3 ← R3 +3R2
≃  0 1 2 2 .
0 0 0 4
This matrix is in echelon form. Note that the final pivot entry is on the right-hand side. The last row
corresponds to the equation
0x + 0y + 0z = 4.
This equation has no solution, because for all x, y, z, the left-hand size will equal 0 and not 4. Therefore,
there is no solution to the given system of equations. In other words, the system is inconsistent. ♠

Example 1.22: Solving a system of equations: an infinite set of solutions

Solve the following system of equations:

3x − y + 5z = 8
y − 10z = 1 (1.7)
6x − y = 17.

Solution. The augmented matrix of this system is

 
3 −1 5 8
 0 1 −10 1  .
6 −1 0 17

We use Gaussian elimination to carry the augmented matrix to echelon form. The first column is the first
pivot column, and 3 is the pivot entry. We use row operating to create zeros beneath the pivot entry. We
subtract 2 times the first row from the third row and get:
 
R3 ← R3 −2R1
3 −1 5 8
≃  0 1 −10 1 
0 1 −10 1

Now, we have created zeros beneath the pivot entry in the first column, so we move on to the second pivot
column (which is the second column) and repeat the procedure. Subtracting the second row from the third
row, we get:  
3 −1 5 8
R3 ← R3 −R2
≃  0 1 −10 1 
0 0 0 0
This matrix is now in echelon form. Observe that the first two columns are pivot columns, and the third
column is not. We call the corresponding variables x and y pivot variables, and the variable z is a free
variable. The equations corresponding to this echelon form are

3x − y + 5z = 8
y − 10z = 1.
22 Systems of linear equations

Observe that the free variable z is not constrained by any equation. In fact, z can equal any number. We
choose t to be any number and let z = t. In this context t is called a parameter. We then use back
substitution to solve for the pivot variables y and x. From the second equation, we have y = 1 + 10z =
1 + 10t. From the first equation, we have 3x = 8 + y − 5z = 8 + (1 + 10t) − 5t = 9 + 5t, and therefore
x = 3 + 53 t. Therefore, the general solution of this system is
x = 3 + 53 t
y = 1 + 10t
z = t,
where t is arbitrary. The system has an infinite set of solutions which are given by these equations. For
any value of the parameter t we select, x, y, and z will be given by the above equations. For example, if
we choose t = 4 then the corresponding solution would be
x = 3 + 53 (4) = 29
3
y = 1 + 10(4) = 41
z = 4.
♠
In Example 1.22 the solution involved a parameter. It may happen that the solution to a system involves
more than one parameter, as shown in the following example.

Example 1.23: Solving a system of equations: a two parameter set of solutions

Solve the following system of equations:

x + 2y − 2z + 2w = 3
x + 2y − z + 3w = 5
x + 2y − 3z + w = 1.

Solution. The augmented matrix is  

1 2 −2 2 3
 1 2 −1 3 5  .
1 2 −3 1 1
We carry this matrix to echelon form using row operations.
     
1 2 −2 2 3 R2 ← R2 −R1 1 2 −2 2 3 1 2 −2 2 3
R ← R −R R3 ← R3 +R2
 1 2 −1 3 5  3 ≃3 1  0 0 1 1 2  ≃  0 0 1 1 2 .
1 2 −3 1 1 0 0 −1 −1 −2 0 0 0 0 0
This matrix is in echelon form and we can see that the first and third columns are pivot columns, whereas
the second and fourth columns are not. Therefore, x and z are pivot variables and y and w are free variables.
We assign parameters y = s and w = t to the free variables. Then we do back substitution to solve for z
and x. From the second equation, we have z = 2 − w = 2 − t. From the first equation, we have x =
3 − 2y + 2z − 2w = 3 − 2s + 2(2 − t) − 2t = 7 − 2s − 4t. Therefore, the general solution is given by
x = 7 − 2s − 4t
y=s
z = 2−t
w = t.
 1.4. Gaussian elimination 23

It is customary to write this solution in the form

   
x 7 − 2s − 4t
 y   s 
 = . (1.8)
 z   2−t 
w t

♠
In Examples 1.20–1.23, we have seem systems of equations with one solution, no solution, and infinitely
many solutions with one parameter as well as two parameters. Moreover, in each case, we have been
able to determine the number of solutions by looking at the echelon form of the augmented matrix. To
summarize, we have the following possibilities for a system of equations:

1. No solution: If the echelon form has a row of the form

 
0 0 0 b ,

where b 6= 0, then the system is inconsistent and has no solution.

2. One solution: For a consistent system of equations: If every column of the coefficient matrix of the
echelon form is a pivot column, the system has exactly one solution. The following is an example
of an augmented matrix in echelon form for a system of equations with one solution.
 
1 1 −2 5
 0 2 3 0 
 .
 0 0 1 2 
0 0 0 0

3. Infinitely many solutions: For a consistent system of equations: If not all columns of the coefficient
matrix of the echelon form are pivot columns, then the system has infinitely many solutions. In
this case, each variable corresponding to a non-pivot column is a free variable and can be assigned
a parameter. The remaining variables are pivot variables and can be expressed in terms of the
parameters. Therefore, the number of parameters in the general solution is equal to the number of
non-pivot columns. The following are examples of echelon forms for systems of equations with
infinitely many solutions.  
1 0 1 5
 0 1 2 −3 
 
 0 0 0 0 
0 0 0 0
or  
1 2 3 5
.
0 0 4 6

There is a special name for the number of pivot variables in a system of equations. It is called the rank
of the system.
24 Systems of linear equations

Definition 1.24: Rank of a matrix

Let A be a matrix and consider any echelon form of A. Then, the number r of pivot entries of A does
not depend on the echelon form we choose, and is called the rank of A. We denote it by rank(A).
The rank of a system of linear equations is the rank of its coefficient matrix (i.e., the matrix on the
left-hand side). It is equal to the number of pivot variables.

Example 1.25: Finding the rank of a matrix

Consider the matrix  
1 2 3
 1 5 9 
2 4 6
What is its rank?

Solution. First, we need to find an echelon form of A. Through the usual algorithm, we find that this is
 
1 0 −1
 0 3 6 .
0 0 0

Here we have two pivot entries, and therefore the rank of A is r = 2. ♠

Suppose we have a system of m equations in n variables, and suppose that n > m. Further assume that
the system is consistent. From our above discussion, we know that this system will have infinitely many
solutions. This is because there can be at most one pivot entry per row, and therefore at most m variables
can be pivot variables. It follows that there are at least n − m free variables. Therefore, the general solution
of this system has at least n − m parameters.
Notice that if n = m or n < m, it is possible for the system to have a unique solution or infinitely many
solutions. In all cases (n > m, n = m, or n < m), it is also possible for the system to be inconsistent (have
no solutions).
By refining the above argument, we get the following theorem:

Theorem 1.26: Rank and solutions of consistent system of equations

Consider a system of m equations in n variables, and assume that the coefficient matrix has rank r.
Assume further that the system is consistent.

1. If r = n, then the system has a unique solution.

2. If r < n, then the system has infinitely many solutions, with n − r parameters.

Here is a final summary of how the rank affects the number of solutions:

1. No solution. If the system of equations is inconsistent, then it has no solution, regardless of the rank.
 1.4. Gaussian elimination 25

2. Unique solution. For a consistent system, suppose r = n. Then there is a pivot position in every
column of the coefficient matrix of A. Hence, there is a unique solution.

3. Infinitely many solutions. For a consistent system, suppose r < n. Then there are less pivot positions
than columns in the coefficient matrix, meaning that not every column is a pivot column. The
columns which are not pivot columns correspond to parameters. In fact, in this case we have n − r
parameters. The system has infinitely many solutions.

Exercises

Exercise 1.4.1 Consider the following augmented matrix in which ∗ denotes an arbitrary number and 
denotes a non-zero number. Determine whether the given augmented matrix is consistent. If consistent, is
the solution unique?  
 ∗ ∗ ∗ ∗ ∗
 0  ∗ ∗ 0 ∗ 
 
 0 0  ∗ ∗ ∗ 
0 0 0 0  ∗

Exercise 1.4.2 Consider the following augmented matrix in which ∗ denotes an arbitrary number and 
denotes a non-zero number. Determine whether the given augmented matrix is consistent. If consistent, is
the solution unique?  
 ∗ ∗ ∗
 0  ∗ ∗ 
0 0  ∗

Exercise 1.4.3 Consider the following augmented matrix in which ∗ denotes an arbitrary number and 
denotes a non-zero number. Determine whether the given augmented matrix is consistent. If consistent, is
the solution unique?  
 ∗ ∗ ∗ ∗ ∗
 0  0 ∗ 0 ∗ 
 
 0 0 0  ∗ ∗ 
0 0 0 0  ∗

Exercise 1.4.4 Consider the following augmented matrix in which ∗ denotes an arbitrary number and 
denotes a non-zero number. Determine whether the given augmented matrix is consistent. If consistent, is
the solution unique?  
 ∗ ∗ ∗ ∗ ∗
 0  ∗ ∗ 0 ∗ 
 
 0 0 0 0  0 
0 0 0 0 ∗ 
26 Systems of linear equations

Exercise 1.4.5 Suppose a system of equations has fewer equations than variables. Will such a system
necessarily be consistent? If so, explain why and if not, give an example which is not consistent.

Exercise 1.4.6 If a system of equations has more equations than variables, can it have a solution? If so,
give an example and if not, explain why not.

Exercise 1.4.7 Find h such that  

2 h 4
3 6 7
is the augmented matrix of an inconsistent system.

Exercise 1.4.8 Find h such that  

1 h 3
2 4 6
is the augmented matrix of a consistent system.

Exercise 1.4.9 Find h such that  

1 1 4
3 h 12
is the augmented matrix of a consistent system.

Exercise 1.4.10 Choose h and k such that the augmented matrix shown has each of the following:

(a) one solution

(b) no solution
(c) infinitely many solutions
 
1 h 2
2 4 k

Exercise 1.4.11 Choose h and k such that the augmented matrix shown has each of the following:

(a) one solution

(b) no solution
(c) infinitely many solutions
 
1 2 2
2 h k

Exercise 1.4.12 Determine if the system is consistent. If so, is the solution unique?
x + 2y + z − w = 2
x−y+z+w = 1
2x + y − z = 1
4x + 2y + z = 5
 1.4. Gaussian elimination 27

Exercise 1.4.13 Determine if the system is consistent. If so, is the solution unique?

x + 2y + z − w = 2
x−y+z+w = 0
2x + y − z = 1
4x + 2y + z = 3

Exercise 1.4.14 Determine which matrices are in echelon form.

   
1 1 2 0   0 1 0 0 5
1 0 0
(a)  0 0 3 2  (b) (c)  0 0 1 0 4 
2 1 7
0 0 0 0 0 0 0 1 3

Exercise 1.4.15 Row reduce each of the following matrices to echelon form.
     
2 −1 3 −1 0 0 −1 −1 3 −6 −7 −8

(a) 1 0 2 1  
(b) 1 1 1 0  
(c) 1 −2 −2 −2 
1 −1 1 −2 1 1 0 −1 1 −2 −3 −4
     
2 4 5 15 4 −1 7 10 3 5 −4 2

(d) 1 2 3 9  
(e) 1 0 3 3  
(f) 1 2 −1 1 
1 2 2 6 1 −1 −2 1 1 1 −2 0
 
−2 3 −8 7
(g)  1 −2 5 −5 
1 −3 7 −8

Exercise 1.4.16 Find the general solution of the system whose augmented matrix is
   
1 2 0 2 1 2 0 2  
    1 1 0 1
(a) 1 3 4 2 (b) 2 0 1 1 (c)
1 0 4 2
1 0 2 1 3 2 1 3
   
1 0 2 1 1 2 1 0 2 1 1 2
 0 1 0 1 2 1   0 1 0 1 2 1 
(d) 
 1
 (e)  
2 0 0 1 3   0 2 0 0 1 3 
1 0 1 0 2 2 1 −1 2 2 2 0

Exercise 1.4.17 Solve the system of equations 7x + 14y + 15z = 22, 2x + 4y + 3z = 5, and 3x + 6y + 10z =
13.

Exercise 1.4.18 Solve the system of equations 3x − y + 4z = 6, y + 8z = 0, and −2x + y = −4.

Exercise 1.4.19 Solve the system of equations 9x − 2y + 4z = −17, 13x − 3y + 6z = −25, and −2x − z = 3.
28 Systems of linear equations

Exercise 1.4.20 Solve the system of equations 65x + 84y + 16z = 546, 81x + 105y + 20z = 682, and
84x + 110y + 21z = 713.

Exercise 1.4.21 Solve the system of equations 8x+2y+3z = −3, 8x+3y+3z = −1, and 4x+y+3z = −9.

Exercise 1.4.22 Suppose a system of equations has fewer equations than variables and you have found a
solution to this system of equations. Is it possible that your solution is the only one? Explain.

Exercise 1.4.23 Suppose a system of linear equations has an augmented matrix with 2 rows and 4 columns
and the last column is a pivot column. Could the system of linear equations be consistent? Explain.

Exercise 1.4.24 Suppose the coefficient matrix of a system of n equations with n variables has the property
that every column is a pivot column. Does it follow that the system of equations must have a solution? If
so, must the solution be unique? Explain.

Exercise 1.4.25 Suppose there is a unique solution to a system of linear equations. What must be true of
the pivot columns in the augmented matrix?

Exercise 1.4.26 The steady state temperature, u, of a plate solves Laplace’s equation, ∆u = 0. One way
to approximate the solution is to divide the plate into a square mesh and require the temperature at each
node to equal the average of the temperature at the four adjacent nodes. In the following picture, the
numbers represent the observed temperature at the indicated nodes. Find the temperature at the interior
nodes, indicated by x, y, z, and w. One of the equations is z = 14 (10 + 0 + w + x).

20 20
x y
10 30
z w
10 30

0 0

Exercise 1.4.27 Find the rank of the following matrices.

 
    0 0 −1 0 3
4 −16 −1 −5 3 6 5 12  1 4 1 0 −8 

(a) 1 −4 0 −1  (b) 1 2 2 5 
 (c) 
 1

4 0 1 2 
1 −4 −1 −2 1 2 1 2
−1 −4 0 −1 −2
   
1 −2 0 3 11 −2 −3 −2
 1 −2 0 4 15   1 1 1 
(d)  
 1 −2 0 3 11  (e) 


1 0 1 
0 0 0 0 0 −3 0 −3

Exercise 1.4.28 Suppose A is an m × n-matrix. Explain why the rank of A is always no larger than
min(m, n).
 1.5. Gauss-Jordan elimination 29

Exercise 1.4.29 State whether each of the following sets of data is possible for a system of equations. If
possible, describe the solution set. That is, indicate whether there exists a unique solution, no solution or
infinitely many solutions. Here, A is the coefficient matrix, and [A | B] denotes the augmented matrix of the
system.

(a) A is a 5 × 6-matrix, rank(A) = 4 and rank [A | B] = 4.

(b) A is a 3 × 4-matrix, rank(A) = 3 and rank [A | B] = 2.
(c) A is a 4 × 2-matrix, rank(A) = 4 and rank [A | B] = 4.
(d) A is a 5 × 5-matrix, rank(A) = 4 and rank [A | B] = 5.
(e) A is a 4 × 2-matrix, rank(A) = 2 and rank [A | B] = 2.

Exercise 1.4.30 Consider the system −5x + 2y − z = 0 and −5x − 2y − z = 0. Both equations equal zero
and so −5x + 2y − z = −5x − 2y − z which is equivalent to y = 0. Does it follow that x and z can equal
anything? Notice that when x = 1, z = −4, and y = 0 are plugged in to the equations, the equations do
not equal 0. Why?

1.5 Gauss-Jordan elimination

Outcomes
A. Find the reduced echelon form of a matrix.

B. Solve a system of linear equations using Gauss-Jordan elimination.

In the previous section, we saw how to solve a system of equations by using Gaussian elimination and
back substitution. The back substitution step can be quite confusing and error prone, especially when
there are parameters. For example, in Example 1.23, we had to substitute y = s, z = 2 − t, and w = t into
the equation x = 3 − 2y + 2z − 2w, which required another simplification step.
In this section, you will learn an alternative procedure called Gauss-Jordan elimination which elimi-
nates the need for back substitution, at the expense of doing a few additional row operations. The key to
this technique is a special kind of echelon form called a reduced echelon form.

Definition 1.27: Reduced echelon form

An augmented matrix is in reduced echelon form if

1. It is in echelon form.

2. Each leading entry is equal to 1.

3. All entries above a leading entry are zero.

30 Systems of linear equations

Example 1.28: Reduced echelon form

The following augmented matrices are in reduced echelon form. The leading entries have been
circled for emphasis. Note how all of the leading entries are equal to 1, and they have zeros above
them.  
  1 0 0 0
1 2 0 5 0 3  
 0 1 0 0  1 0 0 4
 0 0 1 2 0 0   
     
 0 0 0 0 1 1 ,  0 0 1 0 , 0 1 0 3
 0 0 0 1  0 0 1 2
0 0 0 0 0 0
0 0 0 0

We can carry every augmented matrix to reduced echelon form by doing elementary row operations.

Algorithm 1.29: Gauss-Jordan elimination

This algorithm provides a method for using row operations to take a matrix to its reduced echelon
form.

1. First, use Gaussian elimination (Algorithm 1.19) to reduce the matrix to echelon form.

2. Moving from right to left, consider each pivot entry. Without changing the row containing the
pivot entry, or any rows below it, use row operations to create zeros in the column above the
pivot entry. Finally, divide the row by its pivot entry, to make the pivot entry equal to 1.

Example 1.30: Gauss-Jordan elimination

Solve the system of equations from Example 1.20 using Gauss-Jordan elimination.

x + 4y + 3z = 11
2x + 10y + 7z = 27
x + y + 2z = 5.

Solution. In Example 1.20, we had already reduced the system to echelon form:
 
1 4 3 11
 0 2 1 5 .
0 0 1 3

We reduce it to reduced echelon form by performing the following row operations:

       
1 4 3 11 R2 ← R2 −R3 1 4 0 2 1 0 0 −2 R ← 1 R 1 0 0 −2
R ← R −3R R ← R −2R 2 2 2
 0 2 1 5  1 ≃1 3  0 2 0 2  1 ≃1 2  0 2 0 2  ≃  0 1 0 1 .
0 0 1 3 0 0 1 3 0 0 1 3 0 0 1 3

The resulting matrix is in reduced echelon form. Note that the final system of equations is especially easy
to solve, because the three equations are x = −2, y = 1, and z = 3. No back substitution is needed. ♠
 1.5. Gauss-Jordan elimination 31

Example 1.31: Gauss-Jordan elimination

Solve the system of equations from Example 1.23 using Gauss-Jordan elimination.

x + 2y − 2z + 2w = 3
x + 2y − z + 3w = 5
x + 2y − 3z + 1w = 1.

Solution. In Example 1.23, we had obtained the following echelon form:

 
1 2 −2 2 3
 0 0 1 1 2 .
0 0 0 0 0
We reduce this to reduced echelon form by performing the following additional step:
   
1 2 −2 2 3 1 2 0 4 7
R1 ← R1+2R2
 0 0 1 1 2  ≃  0 0 1 1 2 .
0 0 0 0 0 0 0 0 0 0
The first equation states that x = 7 − 2y − 4w, and the second equation states that z = 2 − w. Using the free
variables y and w as parameters, we obtain the following general solution:
x = 7 − 2y − 4w
y=y
z = 2−w
w = w.
Note that we did not really have to do back substitution; all we had to do is to shift parts of the equations to
the right-hand side. If the solution looks strange, because it has equations like “y = y” in it, keep in mind
that this means that y and w are arbitrary numbers, i.e., parameters. We can replace y and w by parameters
s and t on the right-hand side, as before:
x = 7 − 2s − 4t
y=s
z = 2−t
w = t.
♠

Discussion 1.32: Which procedure is better?

Which one is the better procedure to use, Gaussian elimination with back substitution, or Gauss-
Jordan elimination? The answer is: it depends. In certain applications, it is not necessary to com-
pletely solve a system of equations. Sometimes it is sufficient just to figure out whether the system
is consistent or inconsistent, or whether the solution is unique or not. In those situations, you al-
ready get the required information from the echelon form and there is no need to do the additional
steps to reduce the system to reduced echelon form. Also, in some situations, Gauss-Jordan elimi-
nation can introduce fractions into your augmented matrix, making the matrix more complicated to
work with. In such cases, it may sometimes be easier to do back substitution. But in most cases,
Gauss-Jordan elimination is simpler to do than back substitution, and therefore I recommend using
the Gauss-Jordan method most of the time.
32 Systems of linear equations

One situation where Gauss-Jordan elimination excels is when you have to solve many systems of equations
that all have the same coefficient matrix.

Example 1.33: Multiple systems sharing the same left-hand side

Solve the following two systems of equations.

x+z = 1 x+z = 2
2x + y + 3z = 2 2x + y + 3z = 5
3x + 2y + 5z = 4 3x + 2y + 5z = 8

Solution. We could certainly solve each system of equations separately. But since the left-hand sides are
the same, we will perform exactly the same row operations on both systems. We can save some work by
solving both systems together. Instead of a usual augmented matrix with only one constant vector, we
create an augmented matrix containing both constant vectors at the same time.
 
1 0 1 1 2
 2 1 3 2 5 
3 2 5 4 8

Then we row-reduce the coefficient matrix to reduced echelon form as usual. (We do not need to bother
reducing the right-hand size to reduced echelon form).
     
1 0 1 1 2 R2 ← R2−2R1 1 0 1 1 2 1 0 1 1 2
R ← R −3R R ← R −2R
 2 1 3 2 5  3 ≃3 1  0 1 1 0 1  3 ≃3 2  0 1 1 0 1  .
3 2 5 4 8 0 2 2 1 2 0 0 0 1 0

We see that the first system is inconsistent, because it contains a row of the form [0 0 0 | 1]. The second
system is consistent, and we get the general solution z = t, y = 1 − t, x = 2 − t. ♠

Exercises

Exercise 1.5.1 Determine which matrices are in reduced echelon form.

 
1 2 0
(a)
0 1 7
 
1 0 0 0
(b)  0 0 1 2 
0 0 0 0
 
1 1 0 0 0 5
(c)  0 0 1 2 0 4 
0 0 0 0 1 3
 1.6. Homogeneous systems 33

Exercise 1.5.2 Reduce each of the matrices from Exercise 1.4.15 to reduced echelon form.

Exercise 1.5.3 Use Gauss-Jordan elimination to solve the system of equations −8x + 2y + 5z = 18, −8x +
3y + 5z = 13, and −4x + y + 5z = 19.

Exercise 1.5.4 Use Gauss-Jordan elimination to solve the system of equations 3x − y − 2z = 3, y − 4z = 0,

and −2x + y = −2.

Exercise 1.5.5 Use Gauss-Jordan elimination to solve the system of equations −9x + 15y = 66, −11x +
18y = 79, −x + y = 4, and z = 3.

Exercise 1.5.6 Use Gauss-Jordan elimination to solve the system of equations −19x+8y = −108, −71x+
30y = −404, −2x + y = −12, 4x + z = 14.

Exercise 1.5.7 Solve the following two systems of equations simultaneously, by using a single augmented
matrix with two constant vectors.
x + 2y − z = 0 x + 2y − z = 1
2x + 3y + z = 3 2x + 3y + z = 7
x − y + 2z = 3 x − y + 2z = 4

1.6 Homogeneous systems

Outcomes
A. Determine whether a homogeneous system of equations has non-trivial solutions from its
rank.

B. Find the basic solutions of a homogeneous system of equations.

C. Understand the relationship between the general solution of a system of equations and that of
its associated homogeneous system.

There is a special type of system of linear equations that requires additional study. This type of system is
called a homogeneous2 system of equations. Our focus in this section is to consider what types of solutions
are possible for a homogeneous system of equations, and how the solutions of non-homogeneous systems
are related to those of their homogeneous counterparts.

2 The word “homogeneous” has 5 syllables. In scientific usage, it is not the same as the word “homogenous”.
34 Systems of linear equations

Definition 1.34: Homogeneous system of equations

A system of equations is called homogeneous if each of the constant terms is equal to 0. A homo-
geneous system therefore has the form

a11 x1 + a12 x2 + . . . + a1n xn = 0

a21 x1 + a22 x2 + . . . + a2n xn = 0
.
..
am1 x1 + am2 x2 + . . . + amn xn = 0,

where ai j are coefficients and x j are variables.

The first thing we note is that a homogeneous system is always consistent. Indeed, it always has the
solution x1 = 0, x2 = 0, . . ., xn = 0. This solution is called the trivial solution.
If the system has a solution in which not all of the x1 , . . . , xn are equal to zero, then we call this solution
non-trivial. When working with homogeneous systems of equations, since the trivial solution always
exists, we are usually interested in finding whether there are non-trivial solutions.
The following theorem is a special case of Theorem 1.26. Recall that the rank of a system is the
number of pivot variables in its echelon form.

Theorem 1.35: Rank and solutions of homogeneous system of equations

Consider a homogeneous system of m equations in n variables, and assume that the coefficient
matrix has rank r. Then the system is consistent, and

1. if r = n, then the system has only the trivial solution;

2. if r < n, then the system has infinitely many solutions.

Example 1.36: Homogeneous system with more variables than equations

True or false: Suppose a homogeneous system has more variables than equations. Then the system
has infinitely many solutions.

Solution. This is true. If the system has m equations and n variables, then the rank can be at most m. Since
m < n, the system has infinitely many solutions. Note that it is not possible for a homogeneous system to
be inconsistent, since there is always the trivial solution. ♠

Example 1.37: Homogeneous system with an equal number of variables and equations
True or false: Suppose a homogeneous system has the same number of variables as equations. Then
the system has a unique solution.

Solution. This is false in general. While it is possible for such a system to have a unique solution, it is also
possible for it to have infinitely many. Let there be n equations and n variables. Then depending on the
 1.6. Homogeneous systems 35

echelon form, the rank r could be either equal to n, in which case there is a unique solution, or less than n,
in which case there are infinitely many. ♠
We now consider an example of solving a homogeneous system of equations.

Example 1.38: Solutions to a homogeneous system of equations

Find the general solution to the following homogeneous system of equations. Does the system have
non-trivial solutions?
2x + y + z + 4w = 0
x + 2y − z + 5w = 0

Solution. Notice that this system has m = 2 equations and n = 4 variables, so n > m. Therefore by our
previous discussion, we expect this system to have infinitely many solutions. In particular, it will have
non-trivial solutions.
The process we use to find the solutions for a homogeneous system of equations is the same process
we used for non-homogeneous equations. We construct the augmented matrix and reduce it to reduced
echelon form.    
2 1 1 4 0 1 0 1 1 0
≃
1 2 −1 5 0 0 1 −1 2 0
The corresponding system of equations is

x+z+w = 0
y − z + 2w = 0.

The free variables are z and w. We set them equal to parameters z = s and w = t. Then our general solution
has the form
x = −s − t
y = s − 2t
z=s
w = t.
Hence this system has infinitely many solutions, with two parameters s and t. ♠
Let us write the solution of the last example in another form. Specifically, it can be written as
     
x −1 −1
 y     
  = s  1  + t  −2  . (1.9)
 z   1   0 
w 0 1

Notice that we have constructed a column from the coefficients of s in each equation, and another column
from the coefficients of t. We will discuss this notation more in later chapters. For now, consider what
happens when we choose the parameters to be s = 1 and t = 0. In this case, we get the solution
 
−1
 1 
 
 1 , (1.10)
0
36 Systems of linear equations

which is the same as the column of coefficients for s. This is called a basic solution of the homogeneous
system of equations. The other basic solution is obtained by setting s = 0 and t = 1. In this case,
 
−1
 −2 
 
 0 . (1.11)
1

The basic solutions of a system are columns constructed from the coefficients on parameters in the solution.
If X1 and X2 are the basic solutions (1.10) and (1.11), then the general solution (1.9) is of the form sX1 +
tX2. We say that the general solution of the homogeneous system is a linear combination of its basic
solutions.
We explore this further in the following example.

Example 1.39: Basic solutions of a homogeneous system

Consider the following homogeneous system of equations.

x + 4y + 3z = 0
(1.12)
3x + 12y + 9z = 0.

Find the basic solutions to this system.

Solution. The augmented matrix of this system and the resulting reduced echelon form are
   
1 4 3 0 1 4 3 0
≃ .
3 12 9 0 0 0 0 0

When written in equations, this system is given by

x + 4y + 3z = 0.

Notice that x is the only pivot variable, and y and z are free variables. Let y = s and z = t for parameters s
and t. Then the general solution is
x = −4s − 3t
y=s
z = t,
which can be written as      
x −4 −3
 y  = s 1  +t  0 .
z 0 1
You can see here that we have two columns of coefficients corresponding to parameters, specifically one
for s and one for t. Therefore, this system has two basic solutions! They are
   
−4 −3
X1 =  1  , X2 =  0  .
0 1
 1.6. Homogeneous systems 37

♠
We can take any non-homogeneous system of equations and get a new homogeneous system by keeping
the left-hand sides the same and setting all of the constant terms equal to 0. This is called the associated
homogeneous system of the system of equations. We end this section by investigating how the solutions
of a system of equations are related to the solutions of its associated homogeneous system.

Example 1.40: Non-homogeneous vs. homogeneous system

Solve the system of equations
x + 4y + 3z = 2
(1.13)
3x + 12y + 9z = 6.
How are the solutions related to those of the associated homogeneous system in Example 1.39?

Solution. We note that the associated homogeneous system of (1.13) is the system we saw in Exam-
ple 1.39. We solve the system (1.13) in the usual way by reducing its augmented matrix to reduced
echelon form    
1 4 3 2 1 4 3 2
≃
3 12 9 6 0 0 0 0
and then assigning parameters y = s, z = t to the free variables. From the equation x + 4y + 3z = 2, the
general solution is
x = 2 − 4s − 3t
y=s
z = t,
which can be written as        
x 2 −4 −3
 y  =  0  +s 1  +t  0 . (1.14)
z 0 0 1
We see that the general solution is almost exactly the same as that of the homogeneous system in Exam-
ple 1.39. The only difference is the additional column
 
2
 0  (1.15)
0
♠
Note that the column (1.15), by itself, is a solution of the non-homogeneous system. It is not the most
general solution, but rather the particular solution resulting from the parameters s = 0 and t = 0. We can
therefore interpret equation (1.14) as saying that the general solution of the non-homogeneous system is
equal to a particular solution of the non-homogeneous system, plus the general solution of the associated
homogeneous system. The same is true in general, and we summarize it as a theorem.

Theorem 1.41: Non-homogeneous vs. homogeneous system

Let A be a system of equations, and let B be the associated homogeneous system. Then

the general solution of A = a particular solution of A + the general solution of B.

38 Systems of linear equations

Exercises

Exercise 1.6.1 Find the basic solutions of each of the following homogeneous systems of equations.

2x + 3y + 4z = 0 x−y+z = 0 x + y − z + 2w = 0
(a) x − 2y + z = 0 (b) −x − 2y − 4z = 0 (c) x + 3y + z + 6w = 0
4x − y + 6z = 0 2x + y + 5z = 0 x + 2y + 4w = 0

Exercise 1.6.2 Which of the following homogeneous systems of linear equations have non-trivial solu-
tions?

(a) 4 equations in 3 variables, rank 3.

(b) 3 equations in 4 variables, rank 3.

(c) 4 equations in 3 variables, rank 2.

(d) 3 equations in 4 variables, rank 2.

Exercise 1.6.3 My system of equations has a solution (x, y, z) = (1, 2, 4). The associated homogeneous
system has basic solutions (x, y, z) = (1, 0, 1) and (x, y, z) = (0, 1, −1). What is the general solution of my
system of equations?

1.7 Uniqueness of the reduced echelon form

Outcomes
A. Determine whether two systems of equations are row equivalent, by comparing their reduced
echelon form.

B. For two homogeneous systems of equations that are not row equivalent, find a solution to one
system that is not a solution to the other.

We have seen in earlier sections that every matrix can be brought into reduced echelon form by a sequence
of elementary row operations. Here we will prove that the resulting matrix is unique; in other words, the
resulting matrix in reduced echelon form does not depend upon the particular sequence of elementary row
operations or the order in which they were performed.
Let A be the augmented matrix of a homogeneous system of linear equations in the variables x1 , x2 , . . . , xn
which is also in reduced echelon form. Recall that the matrix A divides the set of variables in two different
types: xi is a pivot variable when column i is a pivot column, and a free variable otherwise.
 1.7. Uniqueness of the reduced echelon form 39

Example 1.42: Pivot and free variables

Find the pivot and free variables in the following system, and find the general solution.

x + 2y − z + w = 0
x+y−z+w = 0
x + 3y − z + w = 0

Solution. The reduced echelon form of the augmented matrix is

 
1 0 −1 1 0
 0 1 0 0 0 .
0 0 0 0 0

From this, we see that columns 1 and 2 are pivot columns. Therefore, x and y are pivot variables and z and
w are free variables. We can write the solution to this system as

x = s−t
y=0
z=s
w = t.

♠
In general, all solutions can be written in terms of the free variables. In such a description, the free
variables are written as parameters, while the pivot variables are written as functions of these parameters.
Indeed, a pivot variable xi is a function of only those free variables x j with j > i. This leads to the following
observation.

Proposition 1.43: Pivot and free variables

If xi is a pivot variable of a homogeneous system of linear equations, then any solution of the system
with x j = 0 for all those free variables x j with j > i must also have xi = 0.

Using this proposition, we prove a lemma which will be used in the proof of the main result of this section.

Lemma 1.44: Solutions and the reduced echelon form of a matrix

Let A and B be two augmented matrices for two homogeneous systems of m equations in n variables,
such that A and B are each in reduced echelon form. If A and B are different, then the two systems
do not have exactly the same solutions.

Proof. With respect to the linear systems associated with the matrices A and B, there are two cases to
consider:

• Case 1: the two systems have the same pivot variables

• Case 2: the two systems do not have the same pivot variables
40 Systems of linear equations

In case 1, the two matrices will have exactly the same pivot positions. However, since A and B are not
identical, there is some row of A which is different from the corresponding row of B and yet the rows each
have a pivot in the same column position. Let i be the index of this column position. Since the matrices
are in reduced echelon form, the two rows must differ at some entry in a column j > i. Let these entries be
a in A and b in B, where a 6= b. Since A is in reduced echelon form, if x j were a pivot variable for its linear
system, we would have a = 0. Similarly, if x j were a pivot variable for the linear system of the matrix B,
we would have b = 0. Since a and b are unequal, they cannot both be equal to 0, and hence x j cannot be a
pivot variable for both linear systems. However, since the systems have the same pivot variables, x j must
then be a free variable for each system. We now look at the solutions of the systems in which x j is set
equal to 1 and all other free variables are set equal to 0. For this choice of parameters, the solution of the
system for matrix A has xi = −a, while the solution of the system for matrix B has xi = −b, so that the
two systems have different solutions.
In case 2, there is a variable xi which is a pivot variable for one matrix, let’s say A, and a free variable
for the other matrix B. The system for matrix B has a solution in which xi = 1 and x j = 0 for all other free
variables x j . However, by Proposition 1.43 this cannot be a solution of the system for the matrix A. This
completes the proof of case 2. ♠

Now, we say that the matrix B is row equivalent to the matrix A if B can be obtained from A by performing
a sequence of elementary row operations. By Theorem 1.11, we know that row equivalent systems have
exactly the same solutions. Now, we can use Lemma 1.44 to prove the main result of this section, which
is that each matrix A has a unique reduced echelon form.

Theorem 1.45: Uniqueness of the reduced echelon form

Every matrix A is row equivalent to a unique matrix in reduced echelon form.

Proof. By Gauss-Jordan elimination, we already know that every matrix is row equivalent to some reduced
echelon form. What we must show is that the resulting reduced echelon form is unique, i.e., does not
depend on the order in which row operations are performed.
Therefore, let A be an m × n-matrix and let B and C be matrices in reduced echelon form, each row
equivalent to A. We have to show that B = C.
Let A+ be the matrix A augmented with a new rightmost column consisting entirely of zeros. Similarly,
augment matrices B and C each with a rightmost column of zeros to obtain B+ and C+. Note that B+ and
C+ are augmented matrices in reduced echelon form, and that both B+ and C+ are row equivalent to A+ ,
because the addition of a column of zeros does not change the effect of any row operations.
Now, A+ , B+ , and C+ can all be considered as augmented matrices of homogeneous linear systems
in the variables x1 , x2 , . . . , xn . Because all three systems are row equivalent, they have exactly the same
solutions. By Lemma 1.44, we conclude that B+ = C+. Omitting the final column of zeros, we must also
have B = C. ♠
 1.7. Uniqueness of the reduced echelon form 41

Example 1.46: Row equivalent systems

Determine whether the following two systems of equations are row equivalent.

2x + 3y + z = 12 x + 2y = 7
x − 2y + 4z = −1 3x − y + 7z = 7
x + 2z = 3, y − z = 2.

Solution. The augmented matrices for the two systems are:

   
2 3 1 12 1 2 0 7
 1 −2 4 −1  ,  3 −1 7 7 .
1 0 2 3 0 1 −1 2

The reduced echelon forms of the two augmented matrices are:

   
1 0 2 3 1 0 2 3
 0 1 −1 2  ,  0 1 −1 2  .
0 0 0 0 0 0 0 0

Since both systems have the same reduced echelon form, they are row equivalent. ♠

Example 1.47: Non-row equivalent systems

Determine whether the following two systems of equations are row equivalent. If they are not row
equivalent, find a solution to one system that is not a solution to the other.

x − 2y − 5z = 0 2x + 2y + z = 0
x+z = 0 x + y + 3z = 0
x + y + 4z = 0, −x − y + 2z = 0.

Solution. The augmented matrices for the two systems are:

   
1 −2 −5 0 2 2 1 0
 1 0 1 0 ,  1 1 3 0 .
1 1 4 0 −1 −1 2 0

The reduced echelon forms of the two augmented matrices are:

   
1 0 1 0 1 1 0 0
 0 1 3 0 ,  0 0 1 0 .
0 0 0 0 0 0 0 0

Since the two systems have different reduced echelon forms, they are not row equivalent. Following the
proof of Lemma 1.44, we see that z is a free variable for the first system, but a pivot variable for the second
system. Therefore, there exists a solution of the first system with z = 1, namely (x, y, z) = (−1, −3, 1). But
there exists no solution for the second system with z = 1, and in particular, (x, y, z) = (−1, −3, 1) is not a
solution of the second system. ♠
42 Systems of linear equations

We finish this section by pointing out an important consequence of Theorem 1.45, namely that the rank of
a matrix is well-defined. Recall that in Definition 1.24, we defined the rank of a matrix A to be the number
of pivot entries of “any” echelon form of A. It was not clear, however, why different echelon forms of A
could not have different numbers of pivot entries. Now we can answer this question. By the Gauss-Jordan
algorithm, we know that every echelon form can be converted to a reduced echelon form without changing
the number or position of the pivots. Since the reduced echelon form is unique, it follows that all echelon
forms of A have the same number of pivot entries (and in fact the same pivot columns). Therefore, the
rank of A is a well-defined quantity.

Exercises

Exercise 1.7.1 The following are augmented matrices for four systems of equations. Determine which of
them, if any, are row equivalent.
       
1 3 5 1 12 1 2 4 −1 8 3 6 −3 1 6 1 2 4 1 10
 −1 1 −1 2 5   2 4 3 1 15   2 4 2 1 13   0 1 1 1 5 
2 0 4 −2 0 3 6 1 −1 8 0 0 1 0 2 2 1 5 0 8

Exercise 1.7.2 Find a tuple (x, y, z) that is a solution to one system of equations but not the other.
   
1 0 1 5 0 1 0 3 6 0
 2 1 4 3 0   2 1 8 2 0 
3 1 5 9 0 2 0 6 13 0

1.8 Fields

Outcomes
A. Solve systems of equations using scalars from a field other than the real numbers, such as Z2
or Z5 .

So far in this chapter, we have worked with real numbers: all of the scalars we used, for coefficients,
constant terms, variables, and parameters, were real numbers. But in fact, we have not used very many
properties of the real numbers, except for the fact that we can add, subtract, multiply, and divide them. For
example, we have never needed to take a square root or to compute a trigonometric function.
In fact, most of linear algebra only requires addition, subtraction, multiplication, and division. This
opens the door to doing linear algebra using other kinds of scalars besides the real numbers. For example,
we can do linear algebra over the rational numbers, complex numbers, or even over some more exotic
 1.8. Fields 43

number systems that you will learn about in this section. A system of scalars that one can do linear
algebra with is called a field.

Definition 1.48: Field

A field is a set K , together with two operations called addition and multiplication, and together
with two distinct elements 0 and 1, such that addition and multiplication satisfy the following prop-
erties:

(A1) Commutative law of addition: a + b = b + a;

(A2) Associative law of addition: (a + b) + c = a + (b + c);
(A3) Unit law of addition: 0 + a = a;
(A4) Additive inverse: for each a ∈ K , there exists an element (−a) ∈ K such that a + (−a) = 0;
(M1) Commutative law of multiplication: ab = ba;
(M2) Associative law of multiplication: (ab)c = a(bc);
(M3) Unit law of multiplication: 1a = a;
(M4) Multiplicative inverse: for each non-zero a ∈ K , there exists an element a−1 ∈ K such that
aa−1 = 1;
(D) Distributive law: a(b + c) = ab + ac.

Properties (A1)–(A4) are about addition, properties (M1)–(M4) are about multiplication, and property (D)
is about both addition and multiplication. Here are some examples and non-examples of fields:

Example 1.49: Some fields and some non-fields

(a) The set R of real numbers is a field.

(b) The set Q of rational numbers is a field.

(c) The set Z of integers satisfies all field properties except for (M4). It is therefore not a field.

(d) The set N = {0, 1, 2, . . .} of natural numbers satisfies all field properties except (A4) and (M4).
It is therefore not a field.

A field doesn’t have to be infinite. The following is an example of a field with only two elements.
44 Systems of linear equations

Example 1.50: The integers modulo 2

Consider the set of bits (binary digits) {0, 1}. We can multiply them as usual, and add them almost
as usual, subject to the alternative rule 1 + 1 = 0 (instead of 1 + 1 = 2). Here is a summary of the
rules for addition and multiplication:

+ 0 1 · 0 1
0 0 1 0 0 0
1 1 0 1 0 1

This particular alternative arithmetic is called “arithmetic modulo 2”. In computer science, the
addition is also called the “logical exclusive or” operation, and multiplication is also called the
“logical and” operation. You can also think of 0 as “even” and 1 as “odd”, and not that odd plus
odd makes even. For example, we can calculate like this:

1 · ((1 + 0) + 1) + 1 = 1 · (1 + 1) + 1
= 1·0+1
= 0+1
= 1.

The binary digits form a field Z2 = {0, 1}, also called the field of integers modulo 2.

You can convince yourself that the 9 properties of fields are all satisfied by the integers modulo 2. This
is a bit tedious, but it can be checked by calculations. For example, to verify (A1), we have to check that
0 + 0 = 0 + 0, 0 + 1 = 1 + 0, 1 + 0 = 0 + 1, and 1 + 1 = 1 + 1. Perhaps the most interesting properties are
(A4) and (M4). For (A4), we can set (−0) = 0 and (−1) = 1. It may be surprising that (−1) = 1, but
you can check for yourself that 1 + (−1) = 1 + 1 = 0 when calculating modulo 2. For (M4), we can set
1−1 = 1.
When solving systems of linear equations, we only used addition, subtraction, multiplication, and
division. Therefore, we can solve systems of equations using the elements of any field as the scalars,
instead of the real numbers.

Example 1.51: Solving a system of equations over Z2

Solve the following system in the integers modulo 2:

x+y = 0
x+z = 1
y + z = 1.

Solution. As usual, we write the augmented matrix of the system of equations, then reduce it to reduced
echelon form using elementary row operations. The only difference is that we will perform all arithmetic
operations modulo 2, rather than in the real numbers. The augmented matrix is:
 
1 1 0 0
 1 0 1 1 .
0 1 1 1
 1.8. Fields 45

The first pivot entry is the 1 in the upper left. We use a row operation to create a zero below it. Note that,
because we are working modulo 2, adding 1 and subtracting 1 is the same thing, so 0 − 1 = 1.
   
1 1 0 0 R ← R −R 1 1 0 0
 1 0 1 1  2 ≃2 1  0 1 1 1  .
0 1 1 1 0 1 1 1

The next pivot entry is in row 2 and column 2. We create a zero below it by subtracting row 2 from row 3,
and a zero above it by subtracting row 2 from row 1:
     
1 1 0 0 R ← R −R 1 1 0 0 1 0 1 1
 0 1 1 1  3 ≃3 2  0 1 1 1  R1 ←≃R1 −R2  0 1 1 1  .
0 1 1 1 0 0 0 0 0 0 0 0

The resulting system is in reduced echelon form. We can see that the system is consistent, because there is
no row whose left-hand side is zero and whose right-hand side is non-zero. We also see that there are two
pivot columns, and therefore two pivot variables, x and y. On the other hand, z is a free variable, so we set
it equal to a parameter: z = t. Notice that this time, the parameter t is not a real number, but an element
of Z2 . From the equation x + z = 1, we get x = 1 − z = 1 + t. Can you guess why I have written 1 + t
instead of 1 −t? This is because (−1) = 1 in the integers modulo 2. So 1 −t = 1 + (−1)t = 1 + 1t = 1 +t.
Similarly, from the equation y + z = 1, we get that y = 1 + t. Therefore, the general solution to the system
of equations is
x = 1+t
y = 1+t
z = t,
where t ∈ {0, 1} is an arbitrary parameter. Recall that this means that each time we plug in a particular
value for t, we get a solution.
There is one difference between solving equations in the real numbers and solving equations in Z2 .
In the real numbers, a system of equations has either no solution, a unique solution, or infinitely many
solutions. This is because when there is a parameter, we automatically get infinitely many solutions. By
contrast, in Z2 , there are only two scalars, and therefore only two possible values for the parameter t,
namely t = 0 and t = 1. For t = 0 we get the solution (x, y, z) = (1, 1, 0), and for t = 1 we get the solution
(x, y, z) = (0, 0, 1). Thus, when the general solution has one parameter in Z2 , there are only two solutions,
instead of infinitely many. ♠
46 Systems of linear equations

Example 1.52: A game with buttons and lights

Consider a game with 9 lights arranged in a square:

Each light is also a button. When a button is pressed, its own light, and all the lights neighboring it
(i.e., above, below, to the left and to the right) are toggled (i.e., any light that was off is turned on
and vice versa). Figure out which buttons to press to turn off all the lights if the starting position is
as shown above.

Solution. We number the lights and buttons from top to bottom, left to right, like this:
1 2 3
4 5 6 .
7 8 9
Let xi be a variable in Z2 , corresponding to the event “button i is pressed” (or more precisely, “button i is
pressed an odd number of times”, because pressing a button twice is the same as not pressing it at all. That
is why we are working modulo 2). The light in position 1 is initially on. It is toggled each time buttons 1,
2, and 4 are pressed, i.e., it is toggled x1 + x2 + x4 times. We want this light to be off in the end. So we
must have 1 + x1 + x2 + x4 = 0. Similarly, the light in position 2 is initially off. To ensure that it stays off,
we must have 0 + x1 + x2 + x3 + x5 = 0. In this way, we obtain 9 linear equations in 9 variables:
1 + x1 + x2 + x4 = 0
0 + x1 + x2 + x3 + x5 = 0
1 + x2 + x3 + x6 = 0
0 + x1 + x4 + x5 + x7 = 0
1 + x2 + x4 + x5 + x6 + x8 = 0
0 + x3 + x5 + x6 + x9 = 0
0 + x4 + x7 + x8 = 0
0 + x5 + x7 + x8 + x9 = 0
1 + x6 + x8 + x9 = 0.
If we write this system in standard form, we obtain the following augmented matrix:
 
1 1 0 1 0 0 0 0 0 1
 1 1 1 0 1 0 0 0 0 0 
 
 0 1 1 0 0 1 0 0 0 1 
 
 1 0 0 1 1 0 1 0 0 0 
 
 0 1 0 1 1 1 0 1 0 1 .
 
 0 0 1 0 1 1 0 0 1 0 
 
 0 0 0 1 0 0 1 1 0 0 
 
 0 0 0 0 1 0 1 1 1 0 
0 0 0 0 0 1 0 1 1 1
 1.8. Fields 47

We solve this system of equations by doing Gauss-Jordan elimination with scalars in Z2 . The reduced
echelon form is  
1 0 0 0 0 0 0 0 0 0
 0 1 0 0 0 0 0 0 0 0 
 
 0 0 1 0 0 0 0 0 0 1 
 
 0 0 0 1 0 0 0 0 0 1 
 
 0 0 0 0 1 0 0 0 0 1 
 .
 0 0 0 0 0 1 0 0 0 0 
 
 0 0 0 0 0 0 1 0 0 0 
 
 0 0 0 0 0 0 0 1 0 1 
0 0 0 0 0 0 0 0 1 0
The unique solution is (x1 , x2 , x3 , x4 , x5 , x6 , x7 , x8 , x9 ) = (0, 0, 1, 1, 1, 0, 0, 1, 0). This means that we must press
buttons 3, 4, 5, and 8. ♠

Example 1.53: The integers modulo 5

Consider the set Z5 = {0, 1, 2, 3, 4}, called the integers modulo 5. We define their addition and mul-
tiplication by computing the usual addition and multiplication, then “reducing” the answer modulo
5. Here, “reducing” a number means repeatedly subtracting 5 until the answer is between 0 and 4.
Imagine a clock showing 5 numbers instead of the usual 12:

0
4 1

3 2

If we want to calculate 3 o’clock plus 4 hours, we get 2 o’clock, because whenever the clock reaches
5, it resets to 0. This is how addition and multiplication modulo 5 are defined:

+ 0 1 2 3 4 · 0 1 2 3 4
0 0 1 2 3 4 0 0 0 0 0 0
1 1 2 3 4 0 1 0 1 2 3 4
2 2 3 4 0 1 2 0 2 4 1 3
3 3 4 0 1 2 3 0 3 1 4 2
4 4 0 1 2 3 4 0 4 3 2 1

We note that the integers modulo 5 form a field. Most of the properties are tedious but easy to verify.
Perhaps the most interesting of the field properties is (M4). It says that for each non-zero element a, there
is another element a−1 such that aa−1 = 1. By looking at the multiplication table, we see that 1 · 1 = 1,
2 · 3 = 1, 3 · 2 = 1, and 4 · 4 = 1. Therefore we can set 1−1 = 1, 2−1 = 3, 3−1 = 2, and 4−1 = 4.

Example 1.54: Division in Z5

What is 2 divided by 3 in Z5 ?
48 Systems of linear equations

Solution. There are no fractions in Z5 . The key to dividing is this: instead of dividing by a, multiply by
a−1 . So we have:
2/3 = 2 · 3−1 = 2 · 2 = 4.
So 2 divided by 3 equals 4. This makes sense, because 4 times 3 equals 2, when calculating modulo 5.
♠

Example 1.55: Solving a system of equations over Z5

Solve the following system of linear equations over Z5 :

2x + z = 1
x + 4y + z = 3
x + 2y + 3z = 2.

Solution. We perform the usual Gauss-Jordan algorithm on the augmented matrix. The only thing to keep
in mind is that, instead of dividing a row by a, we should multiply it by a−1 . And of course, we should
reduce all intermediate results modulo 5. For example, to change the first pivot entry from 2 to 1, we
multiply by 2−1 = 3, instead of dividing by 2.
       
2 0 1 1 R ← 3R 1 0 3 3 RR2 ← R2 −R1
← R −R
1 0 3 3
R ← 4R
1 0 3 3
 1 4 1 3  1 ≃ 1  1 4 1 3  3 ≃3 1  0 4 3 0  2 ≃ 2  0 1 2 0 
1 2 3 2 1 2 3 2 0 2 0 4 0 2 0 4
   
1 0 3 3 R1 ← R1−3R3 1 0 0 1
R3 ← R3 −2R2 R ← R −2R
≃  0 1 2 0  2 ≃2 3  0 1 0 2  .
0 0 1 4 0 0 1 4
The final matrix is in reduced echelon form, and we see that the system has the unique solution (x, y, z) =
(1, 2, 4). Please double-check the solution with respect to the original equations. ♠

Example 1.56: The integers modulo 6 are not a field

Do the integers modulo 6 form a field?

Solution. The integers modulo 6 are the set Z6 = {0, 1, 2, 3, 4, 5}, with the addition and multiplication
modulo 6:
+ 0 1 2 3 4 5 · 0 1 2 3 4 5
0 0 1 2 3 4 5 0 0 0 0 0 0 0
1 1 2 3 4 5 0 1 0 1 2 3 4 5
2 2 3 4 5 0 1 2 0 2 4 0 2 4
3 3 4 5 0 1 2 3 0 3 0 3 0 3
4 4 5 0 1 2 3 4 0 4 2 0 4 2
5 5 0 1 2 3 4 5 0 5 4 3 2 1
Then Z6 satisfies all of the field axioms except (M3). To see why (M3) fails, let a = 2, and note, by looking
at the multiplication table, that there is no b ∈ Z6 such that ab = 1. Therefore, Z6 is not a field. ♠
 1.8. Fields 49

We conclude this section with a fact that we will not prove.

Theorem 1.57: The integers modulo a prime

Let n be a positive integer. Then the set Zn = {0, . . . , n − 1} of integers modulo n, with addition and
multiplication modulo n, forms a field if and only if n is prime. Thus, for example, Z2 , Z3 , Z5 , Z7 ,
and Z11 are fields, whereas Z4 , Z6 , and Z9 are not.

Another field that is very useful in mathematics and the natural sciences is the field C of complex num-
bers. You can read about the complex numbers in Appendix A.

Example 1.58: Solving a system of equations over C

Solve the following system of equations over the complex numbers:

x − y + z = −1 + i,
x + iy + 3z = 1 + 3i.

Solution. We perform Gauss-Jordan elimination on the augmented matrix. When multiplying or dividing,
we have to use complex number arithmetic. For example,
2 2 1 − i 2 − 2i
= · = = 1 − i.
1+i 1+i 1−i 2
The row operations are:
     
1 −1 1 −1 + i R2 ← R2 −R1 1 −1 1 −1 + i R2 ← R2 /(1+i) 1 −1 1 −1 + i
≃ ≃
1 i 3 1 + 3i 0 1 + i 2 2 + 2i 0 1 1−i 2
 
R1 ← R1 +R2 1 0 2−i 1+i
≃ .
0 1 1−i 2
Therefore the general solution is z = t, y = 2 − (1 − i)t, x = 1 + i − (2 − i)t, where t ∈ C is a parameter,
i.e., t is any complex number. In vector form, the general solution is
     
x 1+i −2 + i
 y  =  2  + t  −1 + i  .
z 0 1
♠

Exercises

Exercise 1.8.1 Solve each of the following systems of equations with scalars in Z2 . If there is more
than one solution, write the general solution in parametric form and also write down all of the solutions
individually. How many solutions are there?
50 Systems of linear equations

       
0 1 1 1 0 0 1 1 1 0 1 1 0 1 0 1 1 1
(a)  1 0 0 1 1  (b)  1 0 1 1  (c)  1 0 1 0 0  (d)  1 0 1 1 
1 1 1 0 1 1 1 0 1 1 1 0 1 0 1 1 1 0

Exercise 1.8.2 Solve each of the following systems of equations with scalars in Z3 . How many solutions
does each system have?
     
0 1 1 1 2 1 2 1 0 1 2 0 2
(a)  1 2 0 1  (b)  0 2 1 1 2  (c)  1 0 1 1 
2 0 2 1 1 0 0 1 1 1 1 1 0

Exercise 1.8.3 Solve each of the following systems of equations with scalars in Z5 .
   
0 2 1 4 0 1 2 4 1
(a)  1 1 2 3 2  (b)  3 0 1 1 
2 4 0 0 4 2 4 3 2

Exercise 1.8.4 In Z7 , calculate 1−1 , 2−1 , 3−1 , 4−1 , 5−1 , and 6−1 . Hint: write down the multiplication
table.

Exercise 1.8.5 Consider a game similar to Example 1.52, with 6 lights arranged in a rectangle:

Again, each light doubles as a button. Pressing it toggles its own light, as well as all of its neighbors.
Which buttons do you have to press to turn off all the light from the starting position shown above? Is the
answer unique? Is every pattern of lights reachable from this starting position?

Exercise 1.8.6 Solve each of the following systems of equations with scalars in the complex numbers.
 
  1 i 1 + i −1 + 2i
1 1 1+i 2
(a) (b)  1 + i 2 1 − i 4 + 4i 
1 + i 2 + i 3i 3 + 2i
1 −1 + i −i 0
 1.9. Application: Balancing chemical reactions 51

1.9 Application: Balancing chemical reactions

The tools of linear algebra can also be used in the subject area of Chemistry, specifically for balancing
chemical reactions. Consider the chemical reaction

SnO2 + H2 → Sn + H2 O.

Here the elements involved are tin (Sn), oxygen (O), and hydrogen (H). A chemical reaction occurs that
transforms a combination of tin dioxide (SnO2 ) and hydrogen (H2 ) into a combination of tin (Sn) and
water (H2 O). When considering chemical reactions, we want to investigate how much of each substance
we began with and how much of each substance is involved in the result.
An important theory we will use here is the mass balance theory. It tells us that we cannot create
or delete elements within a chemical reaction. For example, in the above expression, we must have the
same number of atoms of oxygen, tin, and hydrogen on both sides of the reaction. Notice that this is not
currently the case. For example, there are two oxygen atoms on the left and only one on the right. In order
to fix this, we want to find numbers x, y, z, w such that

x SnO2 + y H2 → z Sn + w H2 O,

where both sides of the reaction have the same number of atoms of the various elements.
This is a familiar problem. We can solve it by setting up a system of equations in the variables x, y, z, w.
Thus we need
Sn : x = z
O : 2x = w
H : 2y = 2w.
We can rewrite these equations as
Sn : x − z = 0
O : 2x − w = 0
H : 2y − 2w = 0.
The augmented matrix for this system of equations is given by
 
1 0 −1 0 0
 2 0 0 −1 0  .
0 2 0 −2 0
The reduced echelon form of this matrix is
 
1 0 0 − 21 0
 
 0 1 0 −1 0  ,
0 0 1 − 21 0
and the solution is given in parametric form as
x = 21 t
y=t
z = 12 t
w = t.
52 Systems of linear equations

For example, let t = 2 and this would yield x = 1, y = 2, z = 1, and w = 2. We can put these values back
into the expression for the reaction which yields

SnO2 + 2H2 → Sn + 2H2 O.

Observe that each side of the expression contains the same number of atoms of each element. This means
that the chemical reaction is balanced. Of course, because it is a homogeneous system of equations, any
multiple of a solution is also a solution. For example,

2SnO2 + 4H2 → 2Sn + 4H2O

is also correct. It just means that we have just doubled the amount of every substance involved. In
chemistry, the numbers you are finding would typically be the number of mols of the molecules on each
side. Thus one mol of SnO2 added two mols of H2 yields one mol of Sn and two mols of H2 O.
Here is another example.

Example 1.59: Balancing a chemical reaction

Potassium is denoted by K , oxygen by O, phosphorus by P and hydrogen by H . Consider the
reaction given by
KOH + H3 PO4 → K3 PO4 + H2 O.
Balance this chemical reaction.

Solution. We will use the same procedure as above to solve this problem. We need to find values for
x, y, z, w such that
x KOH + y H3 PO4 → z K3 PO4 + w H2 O.
preserves the total number of atoms of each element. Finding these values can be done by finding the
solution to the following system of equations.

K: x = 3z
O: x + 4y = 4z + w
H: x + 3y = 2w
P: y = z.

The augmented matrix for this system is

 
1 0 −3 0 0
 1 4 −4 −1 0 
 ,
 1 3 0 −2 0 
0 1 −1 0 0

and the reduced echelon form is  

1 0 0 −1 0
 0 1 0 − 31 0 
 
 1 .
 0 0 1 −3 0 
0 0 0 0 0
 1.10. Application: Dimensionless variables 53

The general solution is given in terms of the parameter t as

x=t
y = 31 t
z = 13 t
w=t

Choose a value for t, say 3. This yields x = 3, y = 1, z = 1, and w = 3. It follows that the balanced reaction
is given by
3KOH + H3 PO4 → K3 PO4 + 3H2 O
Note that this results in the same number of atoms of each element on both sides. ♠

Exercises

Exercise 1.9.1 Balance the following chemical reactions.

(a) KNO3 + H2CO3 → K2CO3 + HNO3

(b) AgI + Na2 S → Ag2 S + NaI

(c) Ba3 N2 + H2 O → Ba(OH)2 + NH3

(d) CaCl2 + Na3 PO4 → Ca3 (PO4 )2 + NaCl

1.10 Application: Dimensionless variables

This section shows how solving systems of equations can be used to determine appropriate dimensionless
variables. It is only an introduction to this topic and considers a specific example of a simple airplane
wing shown below. We assume for simplicity that it is a flat plane at an angle to the wind which is blowing
against it with speed V as shown.

A
θ B
54 Systems of linear equations

The angle θ is called the angle of incidence, B is the span of the wing and A is called the chord.
Denote by l the lift. Then this should depend on various quantities like θ ,V , B, A and so forth. Here is a
table which indicates various quantities on which it is reasonable to expect l to depend.

Variable Symbol Units

chord A m
span B m
angle incidence θ m0 kg0 s0
speed of wind V m s−1
speed of sound V0 m s−1
density of air ρ kg m−3
viscosity µ kg s−1 m−1
lift l kg s−2 m

Here m denotes meters, s refers to seconds and kg refers to kilograms. All of these are likely familiar
except for µ , which we will discuss in further detail now.
Viscosity is a measure of how much internal friction is experienced when the fluid moves. It is roughly
a measure of how “sticky" the fluid is. Consider a piece of area parallel to the direction of motion of the
fluid. To say that the viscosity is large is to say that the tangential force applied to this area must be large
in order to achieve a given change in speed of the fluid in a direction normal to the tangential force. Thus

µ (area)(velocity gradient) = tangential force

Hence m
(units of µ ) m 2
= kg s−2 m
sm
Thus the units of µ are
kg s−1 m−1
as claimed above.
Returning to our original discussion, you may think that we would want

l = f (A, B, θ ,V ,V0 , ρ , µ )

This is very cumbersome because it depends on seven variables. Also, it is likely that without much care,
a change in the units such as going from meters to centimeters would result in an incorrect value for l.
The way to get around this problem is to look for l as a function of dimensionless variables multiplied by
something which has units of force. It is helpful because first of all, you will likely have fewer independent
variables and secondly, you could expect the formula to hold independent of the way of specifying length,
mass and so forth. One looks for
l = f (g1 , . . . , gk )ρ V 2 AB
where the units of ρ V 2 AB are
kg  m 2 2 kg × m
m =
m3 s s2
which are the units of force. Each of these gi is of the form

Ax1 Bx2 θ x3 V x4V0x5 ρ x6 µ x7 (1.16)

 1.10. Application: Dimensionless variables 55

and each gi is independent of the dimensions. That is, this expression must not depend on meters, kilo-
grams, seconds, etc. Thus, placing in the units for each of these quantities, one needs
mx1 mx2 (mx4 s−x4 )(mx5 s−x5 )(kg m−3 )x6 (kg s−1 m−1 )x7 = m0 kg0 s0
Notice that there are no units of θ because it is just the radian measure of an angle. Hence its dimensions
consist of length divided by length, thus it is dimensionless. Then this leads to the following equations for
the xi .
m : x1 + x2 + x4 + x5 − 3x6 − x7 = 0
s: −x4 − x5 − x7 = 0
kg : x6 + x7 = 0
The augmented matrix for this system is
 
1 1 0 1 1 −3 −1 0
 0 0 0 1 1 0 1 0 
0 0 0 0 0 1 1 0
The reduced echelon form is given by
 
1 1 0 0 0 0 1 0
 0 0 0 1 1 0 1 0 
0 0 0 0 0 1 1 0
and so the solutions are of the form
x1 = −x2 − x7
x3 = x3
x4 = −x5 − x7
x6 = −x7
Thus, in terms of vectors, the solution is
   
x1 −x2 − x7
 x2   x2 
   
 x3   x3 
   
 x4 = −x5 − x7 
   
 x5   x5 
   
 x6   −x7 
x7 x7
Thus the free variables are x2 , x3 , x5 , x7 . By assigning values to these, we can obtain dimensionless vari-
ables by placing the values obtained for the xi in the formula (1.16). For example, let x2 = 1 and all the
rest of the free variables are 0. This yields
x1 = −1, x2 = 1, x3 = 0, x4 = 0, x5 = 0, x6 = 0, x7 = 0
The dimensionless variable is then A−1 B1 . This is the ratio between the span and the chord. It is called
the aspect ratio, denoted as AR. Next let x3 = 1 and all others equal zero. This gives for a dimensionless
quantity the angle θ . Next let x5 = 1 and all others equal zero. This gives
x1 = 0, x2 = 0, x3 = 0, x4 = −1, x5 = 1, x6 = 0, x7 = 0
56 Systems of linear equations

Then the dimensionless variable is V −1V01 . However, it is written as V /V0 . This is called the Mach number
M . Finally, let x7 = 1 and all the other free variables equal 0. Then

x1 = −1, x2 = 0, x3 = 0, x4 = −1, x5 = 0, x6 = −1, x7 = 1

then the dimensionless variable which results from this is A−1V −1 ρ −1 µ . It is customary to write it as
Re = (AV ρ )/µ . This one is called the Reynold’s number. It is the one which involves viscosity. Thus we
would look for
l = f (Re, AR, θ , M ) kg × m / s2
This is quite interesting because it is easy to vary Re by simply adjusting the velocity or A but it is hard to
vary things like µ or ρ . Note that all the quantities are easy to adjust. Now this could be used, along with
wind tunnel experiments, to get a formula for the lift that would be reasonable. You could also consider
more variables and more complicated situations in the same way.

Exercises

Exercise 1.10.1 In this section, we observed that ρ V 2 AB has the units of force. Describe a systematic
way to obtain such combinations of the variables that will yield something that has the units of force.

1.11 Application: Resistor networks

The tools of linear algebra can be used to study the application of resistor networks. An example of an
electrical circuit is below.

2Ω

18 V
−
+ I1 4Ω

2Ω

The jagged lines ( ) denote resistors and the numbers next to them give their resistance in ohms,
written as Ω. The voltage source ( ) causes the current to flow in the direction from the longer of
3
the two lines toward the shorter . Voltage is measured in volts, written as V. The current for a circuit is
labelled Ik , and is measured in amperes, written as A.
3 By
current, we always mean the conventional current, which flows from plus to minus. It is the opposite of the electron
flow, which goes from minus to plus.
 1.11. Application: Resistor networks 57

In the above figure, the current I1 has been labelled with an arrow in the counterclockwise direction.
This is an entirely arbitrary decision and we could have chosen to label the current in the counterclockwise
direction. With our choice of direction here, we define a positive current to flow in the counterclockwise
direction and a negative current to flow in the clockwise direction.
The goal of this section is to use the values of resistors and voltage sources in a circuit to determine
the current. An essential theorem for this application is Kirchhoff’s law.

Theorem 1.60: Kirchhoff’s law

The sum of the resistance (R) times the amperes (I ) in the counterclockwise direction around a loop
equals the sum of the voltage sources (V ) in the same direction around the loop.

Kirchhoff’s law allows us to set up a system of linear equations and solve for any unknown variables.
When setting up this system, it is important to trace the circuit in the counterclockwise direction. If a
resistor or voltage source is crossed against this direction, the related term must be given a negative sign.
We will explore this in the next example where we determine the value of the current in the initial
diagram.

Example 1.61: Solving for current

Applying Kirchhoff’s Law to the diagram below, determine the value for I1 .

2Ω

18 V
−
+ I1

2Ω
4Ω

Solution. Begin in the bottom left corner, and trace the circuit in the counterclockwise direction. At the
first resistor, multiplying resistance and current gives 2I1 . Continuing in this way through all three resistors
gives 2I1 + 4I1 + 2I1 . This must equal the voltage source in the same direction. Notice that the direction
of the voltage source matches the counterclockwise direction specified, so the voltage is positive.
Therefore the equation and solution are given by

2I1 + 4I1 + 2I1 = 18,

8I1 = 18,
9
I1 = A.
4
Since the answer is positive, this confirms that the current flows counterclockwise. ♠
58 Systems of linear equations

Example 1.62: Solving for current

Applying Kirchhoff’s Law to the diagram below, determine the value for I1 .

27 V
− + 3Ω

4Ω
I1 1Ω

6Ω

Solution. Begin in the top left corner this time, and trace the circuit in the counterclockwise direction.
At the first resistor, multiplying resistance and current gives 4I1 . Continuing in this way through the four
resistors gives 4I1 + 6I1 + 1I1 + 3I1 . This must equal the voltage source in the same direction. Notice that
the direction of the voltage source is opposite to the counterclockwise direction, so the voltage is negative.
Therefore the equation and solution are given by

4I1 + 6I1 + 1I1 + 3I1 = −27,

14I1 = −27,
27
I1 = − A .
14
Since the answer is negative, this tells us that the current flows clockwise. ♠

A more complicated example follows. Two of the circuits below may be familiar; they were examined in
the examples above. However as they are now part of a larger system of circuits, the answers will differ.

Example 1.63: Unknown currents

The diagram below consists of four circuits. The current (Ik ) in the four circuits is denoted by I1 , I2 ,
I3 , I4 . Using Kirchhoff’s Law, write an equation for each circuit and solve for each current.

27 V
2Ω − + 3Ω

18 V
−
+
I2 4 Ω
I3 1Ω

2Ω 6Ω

23 V
+
−
I1 1 Ω
I4 2Ω

5Ω 3Ω
 1.11. Application: Resistor networks 59

Solution. Starting with the top left circuit, multiply the resistance by the current and sum the resulting
products. Specifically, consider the resistor labelled 2 Ω that is part of the circuits of I1 and I2 . Notice that
current I2 runs through this in a positive (counterclockwise) direction, and I1 runs through in the opposite
(negative) direction. The product of resistance and current is then 2(I2 − I1 ) = 2I2 − 2I1 . Continue in this
way for each resistor, and set the sum of the products equal to the voltage source to write the equation:
2I2 − 2I1 + 4I2 − 4I3 + 2I2 = 18.
The above process is used on each of the other three circuits, and the resulting equations are:
Upper right circuit:
4I3 − 4I2 + 6I3 − 6I4 + I3 + 3I3 = −27.
Lower right circuit:
3I4 + 2I4 + 6I4 − 6I3 + I4 − I1 = 0.
Lower left circuit:
5I1 + I1 − I4 + 2I1 − 2I2 = −23.
Notice that the voltage for the upper right and lower left circuits are negative due to the clockwise
direction they indicate. The resulting system has four equations in four variables. Simplifying and rear-
ranging with variables in order, we have:
−2I1 + 8I2 − 4I3 = 18,
−4I2 + 14I3 − 6I4 = −27,
−I1 − 6I3 + 12I4 = 0,
8I1 − 2I2 − I4 = −23.
The augmented matrix is  
−2 8 −4 0 18
 0 −4 14 −6 −27 
 .
 −1 0 −6 12 0 
8 −2 0 −1 −23
The solution to this system of equations is
I1 = −3 A,
1
I2 = A,
4
5
I3 = − A,
2
3
I4 = − A .
2
This tells us that currents I1 , I3 , and I4 travel clockwise while I2 travels counterclockwise. ♠

Exercises

Exercise 1.11.1 Consider the following diagram of four circuits.

60 Systems of linear equations

20 V
3Ω − + 1Ω

5V
−
+ I2 5 Ω
I3 1Ω

2Ω 6Ω

10 V
+
− I1 1 Ω
I4 3Ω

4Ω 2Ω

The current in amperes in the four circuits is denoted by I1 , I2 , I3 , and I4 . It is understood that a positive
current means a current flowing in the counterclockwise direction. If Ik ends up being negative, then it just
means the current flows in the clockwise direction. In the above diagram, the top left circuit should give
the equation
2I2 − 2I1 + 5I2 − 5I3 + 3I2 = 5.
Write equations for each of the other three circuits and then give a solution to the resulting system of
equations.

Exercise 1.11.2 Find I1 , I2 , and I3 , the counterclockwise currents in amperes in the three circuits of the
following diagram.

12 V
3Ω − + 7Ω

10 V
−
+ I1 5 Ω
I2 3Ω

2Ω 1Ω

2Ω
I3 4Ω

4Ω
2. Vectors in Rn

2.1 Points and vectors

Outcomes
A. Understand the geometric and algebraic meaning of points and vectors in Rn .

B. Find the position vector of a point in Rn .

C. Determine whether two vectors are equal.

In this section, we define points and vectors in n-dimensional space, and discuss some of their interpreta-
tions. We start with a brief review of Cartesian coordinate systems.

Points in n-dimensional space. You are probably already familiar with Cartesian coordinates, which let
you describe points in 2- or 3-dimensional space. Consider the familiar coordinate plane, with an x-axis
and a y-axis. Any point within this coordinate plane is identified by its x- and y-coordinates. For example,
the point P in the following diagram has x-coordinate 2 and y-coordinate 1. We write these coordinates
as an ordered pair P = (2, 1). Here, “ordered” means that the x-coordinate comes first, and then the y-
coordinate, i.e., (1, 2) is not the same point as (2, 1). Coordinates can be positive, negative, or zero. The
special point with coordinates (0, 0) is called the origin of the coordinate system, and also written as 0.

y
Q = (−3, 4)
4

P = (2, 1)
1
x
−3 2

The situation in 3 dimensions is analogous. Here, the coordinate system has three axes, and each point
is described by a triple of coordinates, which we can write as (x, y, z). We can extend these ideas beyond
n = 3. A coordinate system for n-dimensional space has n axes, which we may call x1 , . . . , xn (as there
are not enough letters in the alphabet to continue after z). A point of n-dimensional space is described by
an ordered n-tuple (x1 , . . . , xn ) of coordinates. For example, P = (2, 1, 0, −1) is a point in 4-dimensional
space which has x1 -coordinate 2, x2 -coordinate 1, and so on. While most people cannot really picture

61
62 Vectors in Rn

space beyond 3 dimensions, it is easy to imagine tuples of n real numbers. Thus, although we may not be
able to “see” the points in higher dimensions, we can still talk about their coordinates.

Vectors in n-dimensional space. Unlike a point, which describes a location in a coordinate system, a
vector describes an offset or a distance and direction. We usually picture a vector as an arrow, starting at
one point (called the tail of the arrow) and ending at another point (called the tip of the arrow).

Two vectors are considered equal if they have the same direction and length. Thus, all four blue arrows
in the above image describe exactly the same vector. Mathematically, a vector in 2-dimensional space is
described as an offset in the x-direction and an offset in the y-direction. For example, a certain vector v
may be described by the instruction: “move 4 units in the direction parallel to the x-axis, and move 3 units
in the direction parallel to the y-axis”. This situation is pictured here:

v
3 units

4 units

The numbers 4 and 3 are also called the x-component and the y-component of the vector. Notice that
a point has “coordinates”, but a vector has h“components”.
i We write the components of a vector as an
4
ordered column within square brackets: v = 3 . Note that components can also be negative; for example,
a negative x-component indicates to move left instead of right, and a negative y-component indicates to
move down instead of up. The vector with all components equal to 0 is called the zero vector, and is
written 0.
The situation in 3 dimensions is similar. Here, a vector is described by three components,
hxi namely, its
x-component, y-component, and z-component. The three components are written as yz . The same idea
generalizes to n-dimensional vectors when n is greater than 3.

Definition 2.1: Column vectors and Rn

A n-dimensional column vector, often simply called a vector, is an ordered list of n real numbers,
written as a column within square brackets:
 
x1
 x2 
 
 ..  .
 . 
xn

We write Rn for the set of all n-dimensional column vectors. It is also known as n-dimensional
Euclidean space.
 2.1. Points and vectors 63

Vectors are usually denoted by boldface lower-case letters such as v, w, a, b. Some people write a small
arrow above the vector, but we do not do this here.

Points vs. vectors. What is the relationship between points  and

 vectors? Algebraically, they seem to be
almost the same thing, because a point (x, y) and a vector xy are both an ordered pair of real numbers,
written in a slightly different way. On the other hand, geometrically, a point is a location in space, and has
neither a length nor direction, whereas a vector has length and direction, but is not fixed at any particular
location. Indeed, to convince yourself that despite the similarity in their notation, points and vectors
are different kinds of objects, imagine that we moved the origin of the coordinate system to a different
location. Then the coordinates of all the points would change, whereas the components of all the vectors
would remain the same. To describe the components of a vector, we require axes and a scale, but no origin.
To describe the coordinates of a point, we require axes, a scale, and an origin.

Vectors from points. If Q and P are two points in n-dimensional space, we can define a vector from Q
−→
to P. This vector is written QP, and is described by the arrow whose tail is at Q and whose tip is at P, as
in the following picture:
−→
QP
P
Q

If the point Q has coordinates (q1 , . . . , qn ) and the point P has coordinates (p1 , . . . , pn ), then the components
−→
of QP are  
p1 − q1
−→  .. 
QP =  . .
pn − qn
An important special case of this is the case when the point Q is the origin. The following definition is
concerned with that situation.

Definition 2.2: The position vector of a point

−
→
Let P be a point in n-dimensional space. The position vector of P is the vector p = 0P whose tail
is at the origin and whose tip is at P.

p P

If the point P has coordinates (p1 , . . . , pn ), then the components of the position vector are
 
p1
 
p =  ...  .
pn

Thus, the coordinates of a point are the same as the components of its position vector. For this
reason, the position vector is also sometimes called the coordinate vector of P.
64 Vectors in Rn

Points from vectors. Conversely, given any vector p, we may find a point P that has p as its position
vector. To do so geometrically, we first have the move the vector p around until its tail is at the origin. The
point P will then be located at its tip.

p
p P
p

Algebraically, if the vector p has components

 
p1
 
p =  ...  ,
pn

then the point P will have coordinates (p1 , . . . , pn ). This is just the opposite process of Definition 2.2.
So although we went to some lengths to point out that vectors and points are different geometric
objects, as soon as an origin of a coordinate system has been fixed, we can always talk about a point by
talking about its coordinate vector. We will systematically do so, and eventually the distinction between a
point and its coordinate vector will become blurred, so that we will be able to talk about Rn as “a set of
points” or “a set of vectors” interchangeably.

Equality of vectors. Two vectors are equal precisely when all corresponding components are equal. In
symbols, if    
u1 v1
   
u =  ...  and v =  ...  ,
un vn
then u = v if and only if u1 = v1 and u2 = v2 and . . . and un = vn .

Notation. In the text, it is often awkward to write column vectors, because they take up so much space.
To save space, we sometimes use a superscript “T ” to denote a column vector. For example, we write
 T
1 2 3 , or sometimes [1, 2, 3]T , to denote the vector
 
1
 2 .
3

The letter “T ” stands for “transpose”. To transpose a vector means to turn a row into a column or vice
versa.
 2.2. Addition 65

Exercises

Exercise 2.1.1 What is an element of R1 ?

−→ −→
Exercise 2.1.2 Given the points P = (2, 0, −4) and Q = (5, −2, 1), find PQ and QP.

Exercise 2.1.3 Find x and y so that u = [5x − 3y, 4]T and v = [2x − 2y, 2y]T are equal in R2 .

2.2 Addition

Outcomes
A. Compute sums and differences of vectors algebraically and geometrically.

B. Use the laws of vector addition to prove equalities between vector expressions.

Addition of vectors in Rn is defined as follows.

Definition 2.3: Addition of vectors in Rn

   
u1 v1
   
For vectors u =  ...  , v =  ...  ∈ Rn , the sum u + v ∈ Rn is defined by
un vn
     
u1 v1 u1 + v1
   ...   
u + v =  ...  +  = ... .
un vn un + vn

To add vectors, we simply add corresponding components. Therefore, in order to add vectors, they must
be the same size. For example, [1, 2, 3]T + [4, 5, 6]T = [1 + 4, 2 + 5, 3 + 6]T = [5, 7, 9]T .
The geometric significance of vector addition in Rn is given in the following proposition.
66 Vectors in Rn

Proposition 2.4: Geometry of vector addition

Let u and v be two vectors in Rn . Slide v so that the tail of v is on the tip of u. Then draw the arrow
which goes from the tail of u to the tip of v. This arrow represents the vector u + v.

u+v
v

Example 2.5: Commutative law

Let u and v be two vectors. Using the geometry of vector addition, explain why u + v = v + u.

Solution. In the following diagram, the vectors u and v form a parallelogram. Therefore, whether we line
up the tail of u with the tip of v or vice versa, we obtain the same vector, which is both u + v and v + u.

u
v
v+u
u+v
v
u

Definition 2.6: Negative

   
u1 −u1
   
The negative of a vector u =  ...  ∈ Rn is defined by −u =  ... .
un −un

Geometrically, the vector −u has the same magnitude as u, but the opposite direction.

u −u

To define the subtraction of two vectors, we simply regard u − v as an abbreviation for u + (−v),
exactly as we do with real numbers. Algebraically, this just amounts to componentwise subtraction:
     
u1 v1 u1 − v1
 ..   ..   .. 
 . − .  =  . .
un vn un − vn
 2.2. Addition 67

The following example illustrates how to subtract vectors geometrically.

Example 2.7: Graphing vector addition

Consider the following picture of vectors u and v.

v
u

Sketch a picture of u + v and u − v.

Solution. We will first sketch u + v. Begin by drawing u and then at the point of u, place the tail of v as
shown. Then u + v is the vector which results from drawing a vector from the tail of u to the tip of v.

v
u

u+v

Next consider u − v. This means u + (−v). From the above geometric description of vector addition,
−v is the vector which has the same length but which points in the opposite direction to v. Here is a picture
of −v

−v

The following picture fully represents u − v:

−v

u−v
u

Given any two vectors u and v one can create a parallelogram with sides these vectors and diagonals
u + v and v − u:

v
u
u+v

u−v u
v
68 Vectors in Rn

♠
Addition of vectors satisfies some important properties which are outlined in the following proposition.
Recall that 0 is the zero vector, the vector from Rn in which all components are equal to 0.

Proposition 2.8: Properties of vector addition

The following properties hold for vectors u, v, w ∈ Rn .

• The commutative law of addition

u + v = v + u.

• The associative law of addition

(u + v) + w = u + (v + w).

• The existence of an additive unit

u + 0 = u.

• The existence of an additive inverse

u + (−u) = 0.

Exercises

    
1 1 −1
Exercise 2.2.1 Find  2  +  5  +  2 .
3 1 −4

Exercise 2.2.2 Use the properties of vector addition from Proposition 2.8 to show the following equalities.
Justify every step.

(a) (u + v) + w = (v + w) + u.

(b) (u + 0) + (v + (−u)) = v.
 2.3. Scalar multiplication 69

2.3 Scalar multiplication

Outcomes
A. Multiply a scalar by a vector algebraically and geometrically.

B. Use the laws of scalar multiplication to prove equalities between vector expressions.

Scalar multiplication of vectors in Rn is defined as follows.

Definition 2.9: Scalar multiplication of vectors in Rn

If k ∈ R is a scalar and u ∈ Rn is a vector, then their scalar multiplication ku ∈ Rn is defined by
   
u1 ku1
   
ku = k  ...  =  ...  .
un kun

For example 3 [1, 2, 3]T = [3, 6, 9]T and −2 [1, 2, 3]T = [−2, −4, −6]T .

Example 2.10: Geometric meaning of scalar multiplication

Let u = [2, 1]T , and draw the following vectors to scale: 2u, u, 12 u, 0u, − 12 u, −u, and −2u. What is
the geometric meaning of scalar multiplication?

Solution. Here is a picture of the seven vectors. We draw their tails in different places to make their
relationship easier to see.

1
2u u 2u 0u − 12 u −u −2u

We see that the vector ku has the same direction as u when k is positive, and the opposite direction when k
is negative. Further, the length of the vector is scaled by a factor of |k|. It increases if |k| > 1 and decreases
if |k| < 1. For example, the vector 2u is exactly twice as long as u. (It is because of this scaling property
that scalars are called scalars). ♠
Just as with addition, scalar multiplication of vectors satisfies several important properties. These are
outlined in the following proposition.
70 Vectors in Rn

Proposition 2.11: Properties of scalar multiplication

The following properties hold for vectors u, v ∈ Rn and k, ℓ scalars.

• The distributive law over vector addition

k(u + v) = ku + kv.

• The distributive law over scalar addition

(k + ℓ)u = ku + ℓu.

• The associative law for scalar multiplication

k(ℓu) = (kℓ)u.

• The rule for multiplication by 1

1u = u.

Proof. We will show the proof of:

k(u + v) = ku + kv.
Assume u = [u1 , . . . , un ]T and v = [v1 , . . . , vn ]T . We have:

k(u + v) = k [u1 + v1 , . . . , un + vn ]T
= [k(u1 + v1 ), . . ., k(un + vn )]T
= [ku1 + kv1 , . . . , kun + kvn ]T
= [ku1 , . . . , kun ]T + [kv1 , . . . , kvn ]T
= ku + kv.

Exercises

Exercise 2.3.1 Consider the vectors u and v drawn below.

u
v

Draw −u, 2v, and − 21 v.

 2.4. Linear combinations 71

  
5 −8
 −1   
Exercise 2.3.2 Find −3   + 5  2 .
 2   −3 
−3 6

Exercise 2.3.3 Use the properties of scalar multiplication from Proposition 2.11 and the properties of
vector addition from Proposition 2.8 to prove the following equalities. Justify every step.

(a) (k + ℓ)(u + v) = ku + kv + ℓu + ℓv.

(b) 0u = 0.

(c) (−1)u = −u.

(d) −(ku) = k(−u) = (−k)u.

2.4 Linear combinations

Outcomes
A. Compute linear combinations of vectors algebraically and geometrically.

B. Determine whether a vector is a linear combination of given vectors.

C. Find the coefficients of one vector as a linear combination of other vectors.

Now that we have studied both vector addition and scalar multiplication, we can combine the two opera-
tions. You may remember that when we talked about the solutions to homogeneous systems of equations
in Section 1.6, we briefly mentioned that the general solution of a homogeneous system is a linear combi-
nation of its basic solutions. We now return to the concept of a linear combination.

Definition 2.12: Linear combination

A vector v is said to be a linear combination of the vectors u1 , . . ., un if there exist scalars a1 , . . . , an
such that
v = a1 u1 + . . . + an un .
The numbers a1 , . . . , an are called the coefficients of the linear combination.
72 Vectors in Rn

Example 2.13: Linear combination

We have      
−4 −3 −18
3 1 +2 0  =  3 .
0 1 2
Thus we can say that  
−18
v= 3 
2
is a linear combination of the vectors
  

−4 −3
u1 =  1  and u2 =  0  .
0 1

For the specific case of R3 , there are three special vectors which we often use. They are given by
     
1 0 0
i = 0 , j = 1 , and k = 0  .
    
0 0 1

We can write any vector u = [a1 , a2 , a3 ]T as a linear combination of these vectors, namely

u = a1 i + a2 j + a3 k.

We will use this notation from time to time.

Example 2.14: Determining if linear combination

Can v = [1, 3, 5]T be written as a linear combination of u1 = [2, 2, 6]T , u2 = [1, 6, 8]T , and u3 =
[3, 8, 18]T ? If yes, find the coefficients.

Solution. This question can be rephrased as: can we find scalars x, y, z such that

xu1 + yu2 + zu3 = v?

Multiplying out produces the system of linear equations

2x + y + 3z = 1
2x + 6y + 8z = 3
6x + 8y + 18z = 5.

Now we row reduce the corresponding augmented matrix to solve.

     
2 1 3 1 R1 ← R1−R2 0 −5 −5 −2 0 0 4 0
R ← R −3R R1 ← R1 +R3 ↔ R1
 2 6 8 3  3 ≃3 1  2 6 8 3  ≃  2 6 8 3  R2 ≃
6 8 18 5 0 5 9 2 0 5 9 2
 2.4. Linear combinations 73

  1 R1    
2 6 8 3 2
R ↔R
1 3 4 32 1
R 1 3 4 32 RR1 ← R1 −4R3
4 3 2 ← R2 −9R3
 0 0 4 0  2≃ 3 0 5 9 2  ≃  0 5 9 2  ≃
0 5 9 2 0 0 4 0 0 0 1 0
 3
  3
  3

1 3 0 2 1
R2 1 3 0 2 1 0 0 10
 0 5 0 2  5≃  0 1 2  R1 ← ≃R1 −3R2

0 5 0 1 0 25  .
0 0 1 0 0 0 1 0 0 0 1 0
We are in the case where we have a unique solution:
3
x=
10
2
y=
5
z = 0.
This means that v is a linear combination of u1 , u2 , and u3 :
3 2
v= u1 + u2 + 0u3 .
10 5
3 2
The coefficients are 10 , 5 , and 0. In fact, v is also a linear combination of just u1 and u2 . ♠
In the following example, we examine the geometric meaning of linear combinations.

Example 2.15: Graphing a linear combination of vectors

Consider the following picture of the vectors u and v
v
u

Sketch a picture of u + 2v and u − 12 v.

Solution. Both vectors are shown below.

− 12 v

u − 21 v
2v
u

u + 2v

♠
Given any two non-parallel vectors u and v in R2 , we can create a grid of their linear combinations. The
integer ones are pictured below. From this we can see that all vectors in R2 can be written as a linear
combination of u and v.
74 Vectors in Rn

3u + 3v
1u + 3v
1u + 3v
3u + 2v
0u + 3v

−1u + 3v 3u + 1v

−1u + 2v 3u + 0v

−1u + 1v v 3u − 1v
u
0 2u − 1v
−1u + 0v
1u − 1v
0u − 1v
−1u − 1v

Exercises

  
6 −13
 0   
Exercise 2.4.1 Find −7   + 6  −1 .
 4   1 
−1 6

Exercise 2.4.2 Decide whether 


4
v= 4 
−3
is a linear combination of the vectors

 

3 2
u1 =  1  and u2 =  −2  .
−1 1

If yes, find the coefficients.

Exercise 2.4.3 Decide whether 

4
v= 4 
4
 2.5. Length of a vector 75

is a linear combination of the vectors

    
3 8 2
u1 =  1  , u2 =  0  and u3 =  −2  .
−1 −1 1
If yes, find the coefficients.

2.5 Length of a vector

Outcomes
A. Compute the distance between points in n-dimensional space.

B. Compute the length of a vector algebraically and geometrically.

C. Find vectors that are a given distance from other vectors.

D. Use algebraic properties of the length operation to prove equalities.

E. Normalize a vector.

In this section, we explore what is meant by the length of a vector in Rn . We develop this concept by first
looking at the distance between two points in Rn . Consider two points P = (p1 , p2 ) and Q = (q1 , q2 ) in
the plane, as in the following picture.
P = (p1 , p2 )

Q = (q1 , q2 )

The distance between P and Q is shown in the picture as a solid line, which is the hypotenuse of a right
triangle. The lengths of the two other sides of this triangle are |p1 − q1 | and |p2 − q2 |. Therefore, the
Pythagorean Theorem implies the length of the hypotenuse (and thus the distance between P and Q)
equals q q
d(P, Q) = |p1 − q1 |2 + |p2 − q2 |2 = (p1 − q1 )2 + (p2 − q2 )2 .
Now consider two points P = (p1 , p2 , p3 ) and Q = (q1 , q2 , q3 ) in 3-dimensional space.
P = (p1 , p2 , p3 )

R = (p1 , p2 , q3 )
Q = (q1 , q2 , q3 ) S = (p1 , q2 , q3 )
76 Vectors in Rn

We will use the Pythagorean Theorem twice to find the length of the solid line connecting P and Q. First,
by the Pythagorean Theorem applied to the right triangle QSR, the length of the line joining R and Q
equals q
d(R, Q) = (p1 − q1 )2 + (p2 − q2 )2 .
Second, by the Pythagorean Theorem applied to the triangle QRP, the length of the line joining P and Q
equals q q
d(P, Q) = d(R, Q)2 + (p3 − q3 )2 = (p1 − q1 )2 + (p2 − q2 )2 + (p3 − q3 )2
This discussion motivates the following definition for the distance between points in Rn .

Definition 2.16: Distance between points

Let P = (p1 , . . . , pn ) and Q = (q1 , . . . , qn ) be two points in Rn . Then the distance between these
points is defined as q
d(P, Q) = (p1 − q1 )2 + . . . + (pn − qn )2 .
This formula is also called the distance formula. We may also write |PQ| for the distance between
P and Q.

In the following example, we use Definition 2.16 to find the distance between two points in R4 .

Example 2.17: Distance between points

Find the distance between the points P = (1, 2, −4, 6) and Q = (2, 3, −1, 0) in R4 .

Solution. Using the distance formula, we have

q p √
d(P, Q) = (1 − 2)2 + (2 − 3)2 + (−4 − (−1))2 + (6 − 0)2 = 12 + 12 + 32 + 62 = 47.

Example 2.18: The plane between two points

Describe the points in R3 that are equally distant from the two points Q = (1, 2, 3) and R = (0, 1, 2).

Solution. Let P = (p1 , p2 , p3 ) be such a point. Then P is the same distance from Q and R, thus d(P, Q) =
d(P, R). By the distance formula, we have
q q
(p1 − 1) + (p2 − 2) + (p3 − 3) = (p1 − 0)2 + (p2 − 1)2 + (p3 − 2)2 .
2 2 2

Squaring both sides, we obtain

(p1 − 1)2 + (p2 − 2)2 + (p3 − 3)2 = p21 + (p2 − 1)2 + (p3 − 2)2 ,
 2.5. Length of a vector 77

and so
(p21 − 2p1 + 1) + (p22 − 4p2 + 4) + (p23 − 6p3 + 9) = p21 + (p22 − 2p2 + 1) + (p23 − 4p3 + 4).
Simplifying, this becomes
−2p1 − 4p2 − 6p3 + 14 = −2p2 − 4p3 + 5,
which can finally be written as
2p1 + 2p2 + 2p3 = 9. (2.1)
Therefore, the points P = (p1 , p2 , p3 ) that are the same distance from Q and R form a plane whose equation
is given by (2.1). ♠
We can now use our understanding of the distance between two points to define what is meant by the
length of a vector.

Definition 2.19: Length of a vector

Let  
u1
 
u =  ... 
un
be a vector in Rn . Then the length of u, written kuk, is given by
q
kuk = u21 + . . . + u2n .

The length of a vector is also sometimes called its magnitude or its norm.

This definition corresponds to Definition 2.16, if we consider the vector u to have its tail at the point
0 = (0, . . . , 0) and its tip at the point U = (u1 , . . . , un ). Then the length of u is equal to the distance between
−→
0 and U . In general, PQ = d(P, Q).
Reconsider Example 2.17. We could have also computed the distance between P and Q as the length
−→
of the vector connecting them. This vector is PQ = [1, 1, 3, −6]T , and its length is
−→ p √

PQ = 12 + 12 + 32 + 62 = 47.

The following proposition states a few important properties of the length of vectors.

Proposition 2.20: Properties of length

The following hold for all vectors u, v and scalars k.

• kuk ≥ 0

• kuk = 0 if and only if u = 0.

• kkuk = |k| kuk.

78 Vectors in Rn

We conclude this section by giving a special name to vectors of length 1.

Definition 2.21: Unit vector

A vector u ∈ Rn is called a unit vector if it has length 1, that is, if

kuk = 1.

Let v be a non-zero vector in Rn . Then there is a unit vector u that points in the same direction as v, but
has length 1. This vector is given by
1
u= v.
kvk
We often use the term normalize to refer to this process. When we normalize a vector, we find the
corresponding unit vector.

Example 2.22: Normalizing a vector

Consider the vector v = [1, −3, 4]T . Find the unit vector u that has the same direction as v

p √
Solution. We have kvk = 12 + (−3)2 + 42 =
26, and therefore
 
1 1 T 1 3 4 T
u= v = √ [1, −3, 4] = √ , − √ , √ .
kvk 26 26 26 26
♠

Exercises

Exercise 2.5.1 Find the distance between the points P = (0, 1, 3) and Q = (2, −1, 0) in R3 .

Exercise 2.5.2 Find the distance between the points P = (1, 3, −1, 0) and Q = (2, 2, 3, 3) in R4 .

Exercise 2.5.3 Describe the points in R3 that are equally distant from the two points Q = (1, 1, 1) and
R = (−1, −1, −1).

Exercise 2.5.4 Describe the points in R3 that have distance 1 from the origin.

Exercise 2.5.5 Find the length of each of the following vectors.

 
  1
  1
−3  4 
u= , v = −2  ,
 w= 
 −2  .
2
5
1
 2.6. The dot product 79

Exercise 2.5.6 Prove the properties of Proposition 2.20.

Exercise 2.5.7 Prove that for all vectors u ∈ Rn , we have k−uk = kuk.

Exercise 2.5.8 Which of the following are unit vectors?


    1
1 1
1  1  1 1 
u= 1 , v= 1 , w=  .
2 3 2 1 
1 1
1

Exercise 2.5.9 Normalize the following vectors.

 
  5
  −2
1  −3 
u= , v= 3 , w= 
 1 .
2
2
−1

2.6 The dot product

Outcomes
A. Compute the dot product of vectors geometrically and algebraically.

B. Use properties of the dot product, including the Cauchy-Schwarz inequality and the triangle
inequality, to prove further equalities and inequalities.

C. Determine whether two vectors are orthogonal.

D. Compute the scalar and vector projection of one vector onto another.

E. Decompose a vector into orthogonal components.

There are two ways of multiplying vectors that are useful in applications. The first of these is called the
dot product, and the second is called the cross product. We will consider the dot product here, and the
cross product in the next section.
80 Vectors in Rn

2.6.1. Definition and properties

When we take the dot product of two vectors, the result is a scalar. For this reason, the dot product is also
called the scalar product. Sometimes it is also called the inner product. The definition is as follows.

Definition 2.23: Dot product

   
u1 v1
 u2   v2 
   
Let u =  .. , v =  ..  be two vectors in Rn . We define their dot product as
 .   . 
un vn

·
u v = u1 v1 + u2 v2 + . . . + un vn .

Example 2.24: Compute a dot product

·
Find u v for u = [1, 2, 0, −1]T and v = [0, 1, 2, 3]T .

Solution. We have

·
u v = (1)(0) + (2)(1) + (0)(2) + (−1)(3)
= 0 + 2 + 0 + −3
= −1.

♠
The dot product satisfies a number of important properties.

Proposition 2.25: Properties of the dot product

The dot product satisfies the following properties, where u, v, w are vectors and k, ℓ are scalars.

· ·
• u v = v u.

• u·u ≥ 0, and u·u = 0 if and only if u = 0.

• (ku + ℓv)·w = k(u·w) + ℓ(v·w).

• u·(kv + ℓw) = k(u·v) + ℓ(u·w).

• u·u = kuk . 2

The proof is left as an exercise. Note that, by the last part of the proposition, we can also use the dot
product to find the length of a vector.
 2.6. The dot product 81

Example 2.26: Length of a vector

Use a dot product to find kuk, where  
2
 1 
u= 
 4 .
2

√
Solution. By the last part of Proposition
√ √ 2.25, we have kuk = · ·
u u. We have u u = 22 + 12 + 42 + 22 =
·
25, and therefore kuk = u u = 25 = 5. ♠

2.6.2. The Cauchy-Schwarz and triangle inequalities

The Cauchy-Schwarz inequality is a fundamental inequality satisfied by the dot product. It is given in
the following proposition.

Proposition 2.27: Cauchy-Schwarz inequality

The dot product satisfies the inequality

·
|u v| ≤ kuk kvk (2.2)

Furthermore equality is obtained if and only if one of u or v is a scalar multiple of the other.

Proof. First note that if u = 0, then both sides of (2.2) are equal to zero, and so the inequality holds in this
case. Therefore, we will assume in what follows that u 6= 0. Define a function of t ∈ R by

·
f (t) = (tu + v) (tu + v).

Then by Proposition 2.25, f (t) ≥ 0 for all t ∈ R. Also from Proposition 2.25, we have

·
f (t) = tu (tu + v) + v (tu + v) ·
· ·
= t 2 u u + t(u v) + tv u + v v · ·
2 2
= t kuk + 2t(u v) + kvk . · 2

This means the graph of y = f (t) is a parabola which opens upwards and is never negative. It follows
that this function has at most one root. From the quadratic formula, we know that a quadratic function
at 2 + bt + c has one or zero roots if and only if b2 − 4ac ≤ 0. Applying this reasoning to the function f (t),
we obtain
·
(2(u v))2 − 4 kuk2 kvk2 ≤ 0,
·
which is equivalent to |u v| ≤ kuk kvk. ♠
An important consequence of the Cauchy-Schwarz inequality is the so-called triangle inequality, which
states that the length of one side of a triangle is less than or equal the sum of the lengths of the two other
sides.
82 Vectors in Rn

Proposition 2.28: Triangle inequality

For u, v ∈ Rn , we have
ku + vk ≤ kuk + kvk . (2.3)

u+v
v
u

Proof. By properties of the dot product and the Cauchy-Schwarz inequality, we have

·
ku + vk2 = (u + v) (u + v)
· · ·
= (u u) + (u v) + (v u) + (v v) ·
= kuk + 2(u·v) + kvk
2 2

≤ kuk + 2 |u·v| + kvk

2 2

≤ kuk2 + 2 kuk kvk + kvk2

= (kuk + kvk)2 .

Therefore,
ku + vk2 ≤ (kuk + kvk)2 .
Taking square roots of both sides, we obtain (2.3). ♠

Example 2.29: Triangle inequality

Use the triangle inequality to show

kuk − kvk ≤ ku − vk

holds for all vectors u, v ∈ Rn .

Solution. We have
kuk = k(u − v) + vk ≤ ku − vk + kvk ,
where we have used the triangle inequality in the last step. Note that this is an inequality between real
numbers. Bringing kvk to the other side of the equation, we have

kuk − kvk ≤ ku − vk .

2.6.3. The geometric significance of the dot product

The included angle of two vectors u and v is the angle θ between the vectors such that 0 ≤ θ ≤ π .
 2.6. The dot product 83

u
θ
v

The dot product can be used to determine the included angle between two vectors.

Proposition 2.30: The dot product and the included angle

Let u and v be two vectors in Rn , and let θ be the included angle. Then the following equation
holds.
·
u v = kuk kvk cos θ .

In words, the dot product of two vectors equals the product of the magnitude (or length) of the two vectors
multiplied by the cosine of the included angle. Note that this gives a geometric description of the dot
product that does not depend explicitly on the coordinates of the vectors.

Example 2.31: Find the angle between two vectors

Find the angle between the vectors
   
2 0
u= and v = .
2 3

Solution. By Proposition 2.30,

·
u v = kuk kvk cos θ .
Hence,
cos θ = ·
u v
.
kuk kvk
·
First, we compute u v = (2)(0) + (2)(3) = 6. Then,
√ √
kuk = √ 22 + 22 = 8,
kvk = 02 + 32 = 3.

Therefore, we have
6 1
cos θ = √ = √ .
3 8 2
π
Taking the inverse cosine of both sides of the equation, we find that θ = 4 radians, or 45 degrees. ♠

Example 2.32: Computing a dot product from an angle

Let u, v be vectors with kuk = 3 and kvk = 4. Suppose the angle between u and v is π /3. Find u v. ·
84 Vectors in Rn

Solution. From Proposition 2.30, we have

π  1
·
u v = kuk kvk cos θ = 3 · 4 · cos
3
= 3·4·
2
= 6.

2.6.4. Orthogonal vectors

Two non-zero vectors are said to be orthogonal, sometimes also called perpendicular, if the included
angle is π /2 radians (90◦ ). By convention, we also say that the zero vector is orthogonal to all vectors.

Proposition 2.33: Orthogonal vectors

Let u and v be vectors in Rn . Then u and v are orthogonal if and only if

·
u v = 0.

We also write u ⊥ v to indicate that u and v are orthogonal.

Proof. If u or v is zero, the vectors are orthogonal by definition, and the dot product is 0 in that case, so the
proposition holds. Now assume u and v are both non-zero. Then by Proposition 2.30, we have u v = 0 ·
if and only if kuk kvk cos θ if and only if cos θ = 0. Recall that the included angle is between 0 and π .
Therefore, cos θ = 0 if and only if θ = π /2. ♠

Example 2.34: Determine whether two vectors are orthogonal

Determine whether the vectors
 

2 1
u= 1  and v= 3 
−1 5

are orthogonal.

Solution. In order to determine if these two vectors are orthogonal, we compute the dot product. We have

·
u v = (2)(1) + (1)(3) + (−1)(5) = 0,

and therefore, by Proposition 2.33, the two vectors are orthogonal. ♠

2.6.5. Projections

It is sometimes important to find the component of a vector in a particular direction. Consider the following
picture:
 2.6. The dot product 85

v
u
P
0

Here, u is a non-zero vector specifying a direction, and v is any vector. We have given the label Q to
the tip of v. The point P lies at the place along u that is closest to Q, or equivalently, such that (0, P, Q)
forms a right triangle. The distance from 0 to P (measured positively in the direction of u) is called the
−
→
component of v in the direction of u, and is denoted compu (v). The vector 0P is called the projection of
v onto u, and is denoted proju (v). We wish to find formulas for these quantities.
Let θ be the included angle between v and u. From trigonometry, considering the right triangle
(0, P, Q), we know that
|0P| |0P|
cos θ = = ,
|0Q| kvk
and therefore
|0P| = kvk cos θ . (2.4)
On the other hand, from Proposition 2.30, we have

·
u v = kuk kvk cos θ ,

and therefore
kvk cos θ = ·
u v
. (2.5)
kuk
Putting equations (2.4) and (2.5) together, we obtain the desired formula for the component of v in the
direction of u:
compu (v) = |0P| =
u v
. · (2.6)
kuk
Note that it is possible for this quantity to be negative; this happens when the angle between v and u is
obtuse. In this case, v will have a negative component along u.
−→
The vector proju (v) = 0P can now be computed by re-scaling u to the correct length. Specifically, we
first normalize u by dividing it by its own length, and then multiply by |0P|. In formulas:

→ |0P|
−
proju (v) = 0P = u=
u v
u= ·
u v
u. · (2.7)
kuk kuk2 u u ·
The following definition summarizes what we have just found.
86 Vectors in Rn

Definition 2.35: Vector projection

Let u be a non-zero vector and v any vector. Then the component of v in the direction of u is
defined to be the scalar
compu (v) =
u v
. ·
kuk
The projection of v onto u is defined to be the vector

proju (v) = ·
u v
u=
u v ·u.
·
u u kuk2

These two operations are also called the scalar projection and vector projection, respectively.

Example 2.36: Find the projection of one vector onto another

Find proju (v) if    
2 1
u= 3  and v =  −2  .
−4 1

·
Solution. We can use the formula provided in Definition 2.35 to find proju (v). First, compute u v. This
is given by    
2 1

−4
·
 3   −2  = (2)(1) + (3)(−2) + (−4)(1) = −8.
1
·
Similarly, u u is given by
   
2 2

−4 −4
·
 3   3  = 22 + 32 + (−4)2 = 29.

Therefore, the projection is equal to

   
2 −16/29
8
proju (v) = −  3  =  −24/29  .
29
−4 32/29

An important application of projections is that every vector v can be uniquely written as a sum of two
orthogonal vectors, one of which is a scalar multiple of some given non-zero vector u, and the other of
which is orthogonal to u.
 2.6. The dot product 87

Theorem 2.37: Decomposition into components

Let u be a non-zero vector, and let v be any vector. Then there exist unique vectors vk and v⊥ such
that
v = vk + v⊥ (2.8)
where vk is a scalar multiple of u, and v⊥ is orthogonal to u.

Proof.

v v⊥

vk
u

To show that such a decomposition exists, let

vk = proju (v) = ·
u v
u,
·
u u
and define v⊥ = v − vk . By definition, (2.8) is satisfied, and vk is a scalar multiple of u. We must show
that v⊥ is orthogonal to u. For this, we verify that their dot product equals zero:
· ·
u v⊥ = u (v − vk )
·
= u v − u vk ·
= u v−
u v
· · ·
u u
u u ·
= u v−u v· ·
= 0.
To show uniqueness, suppose that (2.8) holds and vk = ku. Taking the dot product of both sides of (2.8)
·
with u and using u v⊥ = 0, this yields
· ·
u v = u (vk + v⊥ )
·
= u ku + u v⊥ ·
= k kuk2 ,

·
which implies k = u v/ kuk2 . Thus there can be no more than one such vector vk . Since v⊥ must equal
v − vk , it follows that there can be no more than one choice for both vk and v⊥ , proving their uniqueness.
♠

Example 2.38: Decomposition into components

Decompose the vector v into v = a + b where a is parallel to u and b is orthogonal to u.
   
−5 1
v =  3 , u =  2 .
−5 −2
88 Vectors in Rn

Solution. We can let a = vk and b = v⊥ as in the proof of Theorem 2.37. Then

   
1 11/9
a = vk = proju (v) =
u v· u =
11 
2  =  22/9 
u u· 9
−2 −22/9
and      
−5 11/9 −56/9
b = v⊥ = v − vk =  3  −  22/9  =  5/9  .
−5 −22/9 −23/9
♠

Exercises

   
1 2
 2   0 
Exercise 2.6.1 Find   
·
 3   1
.

4 3

·
Exercise 2.6.2 Let a, b be vectors. Show that (a b) = 14 (ka + bk2 − ka − bk2 ).

Exercise 2.6.3 Using the properties of the dot product, prove the parallelogram identity:

ka + bk2 + ka − bk2 = 2 kak2 + 2 kbk2

Exercise 2.6.4 Find cos θ where θ is the angle between the vectors
   
3 1
u =  −1  , v =  4 
−1 2

Exercise 2.6.5 Find cos θ where θ is the angle between the vectors
   
1 1
u =  −2  , v =  2 
1 −7

Exercise 2.6.6 Use the formula given in Proposition 2.30 to verify the Cauchy Schwarz inequality and to
show that equality occurs if and only if one of the vectors is a scalar multiple of the other.

Exercise 2.6.7 Show that the triangle with vertices A = (2, 0, −3), B = (5, −2, 1) and C = (7, 5, 3) is a
right triangle.
 2.6. The dot product 89

   
1 1
Exercise 2.6.8 Find projv (w) where w =  0 and v = 2 .
 
−2 3
   
1 1
Exercise 2.6.9 Find projv (w) where w =  2 and v = 0 .
 
−2 3
   
1 1
 2   2 
Exercise 2.6.10 Find projv (w) where w =   and v = 
 −2  
.
3 
1 0
   
1 0
Exercise 2.6.11 Find compv (w) where w = 1 and v = 3 .
  
2 3
   
1 1
Exercise 2.6.12 Find compv (w) where w = −1 and v = 1 .
  
0 2

Exercise 2.6.13 Does it make sense to speak of proj0 (w)?

Exercise 2.6.14 Decompose the vector v into v = a + b where a is parallel to u and b is orthogonal to u.
   
3 1
v =  2  , u =  −1  .
−5 2

Exercise 2.6.15 Prove the Cauchy Schwarz inequality in Rn as follows. For u, v vectors, consider

·
(u − projv (u)) (u − projv (u)) ≥ 0

Simplify using the axioms of the dot product and then put in the formula for the projection. Notice that
this expression equals 0 and you get equality in the Cauchy Schwarz inequality if and only if u = projv (u).
What is the geometric meaning of u = projv (u)?

Exercise 2.6.16 Let v, w u be vectors. Show that (w + u)⊥ = w⊥ + u⊥ , where w⊥ = w − projv (w).

Exercise 2.6.17 Show that

·
u (v − proju (v)) = 0
and conclude every vector in Rn can be written as the sum of two vectors, one which is orthogonal and
one which is parallel to the given vector.
90 Vectors in Rn

2.7 The cross product

Outcomes
A. Compute the cross product of vectors algebraically and geometrically.

B. Compute the box product of three vectors in R3 .

C. Determine whether a system of three vectors in R3 is right-handed, algebraically and geomet-

rically.

D. Find the areas of parallelograms and triangles in R3 .

E. Find the volume of a parallelepiped determined by three vectors in R3 .

F. Use properties of the cross product and the dot product to prove algebraic equalities.

Unlike the dot product, the cross product is only defined in R3 , i.e., only in 3-dimensional space. The
cross product of two vectors is a vector.
We will first discuss the geometric meaning of the cross product, and then give an algebraic description.
Both descriptions are equally important: the geometric description is essential for applications in physics
and geometry, whereas the algebraic description is necessary for computing.

2.7.1. Right-handed systems of vectors

We begin with a discussion of right-handed systems of vectors in 3-dimensional space.

Definition 2.39: Right-handed system of vectors

Three vectors, u, v, w form a right-handed system if when you extend the thumb of your right hand
in the direction of u and your index finger in the direction of v, your relaxed middle finger points
roughly in the direction of w.
u
v

u
v w

w
 2.7. The cross product 91

You should consider how a right-handed system would differ from a left-handed system. Try using your
left hand and you will see that the vector w would need to point in the opposite direction.
Recall the special vectors i = [1, 0, 0]T , j = [0, 1, 0]T , and k = [0, 0, 1]T we saw in Section 2.4. We
always assume that our coordinate system is drawn in such a way that the vectors i, j, k form a right-
handed system. Thus, if the thumb of your right hand points along the x-axis and your index finger points
along the y-axis, your middle finger should point along the z-axis.

j
i

When all three vectors lie in a plane, then we say that the vectors are coplanar. In this case, the system is
neither right-handed nor left-handed.

2.7.2. Geometric description of the cross product

The following is the geometric description of the cross product. Recall that the dot product of two vectors
results in a scalar. In contrast, the cross product results in a vector, as the cross product gives a direction
as well as a magnitude.

Definition 2.40: Geometric definition of cross product

Let u and v be two vectors in R3 . Their cross product, written u × v, is the vector defined by the
following three rules.

1. Its length is ku × vk = kuk kvk sin θ , where θ is the included angle between u and v.

2. It is orthogonal to both u and v.

3. The vectors u, v, and u × v, in that order, form a right-handed system.

We note that the length of the cross product, ku × vk, given by the formula kuk kvk sin θ , is the area of the
parallelogram determined by u and v, as shown in the following picture.

v kvk sin(θ )

θ
u
92 Vectors in Rn

2.7.3. Algebraic definition of the cross product

From its geometric description, we can prove that the cross product satisfies the following properties.

Proposition 2.41: Properties of the cross product

Let u, v, w be vectors in R3 , and k a scalar. Then the following hold.

1. u × v = −(v × u).

2. u × u = 0.

3. (ku) × v = k(u × v) = u × (kv).

4. u × (v + w) = u × v + u × w.

5. (v + w) × u = v × u + w × u.

Proof. Formula 1. follows immediately from the definition. The vectors u × v and v × u have the same
magnitude, kuk kvk sin θ , and an application of the right hand rule shows they have opposite direction.
Formula 2. is proven as follows. If k is a non-negative scalar, the direction of (ku) × v is the same as
the direction of u × v, k(u × v) and u × (kv). The magnitude is k times the magnitude of u × v which is the
same as the magnitude of k(u × v) and u × (kv). Using this yields equality in 2. In the case where k < 0,
everything works the same way except the vectors are all pointing in the opposite direction and you must
multiply by |k| when comparing their magnitudes.
The distributive laws, 3. and 4., are harder to establish. For now, we will content ourselves with
noticing that if we know that 3. is true, 4. follows. Namely, assuming 3., and using 1., we have
(v + w) × u = −u × (v + w)
= −(u × v + u × w)
= v × u + w × u.
♠
In turn, we can use the properties from Proposition 2.41 to get an algebraic description of the cross product.
We begin by determining the cross products of the special vectors i, j, and k. They are as follows:
i × j = k, j × i = −k, i × i = 0,
k × i = j, i × k = −j, j × j = 0,
j × k = i, k × j = −i, k × k = 0.
With this information and the laws of Proposition 2.41, we can compute the cross product of any two
vectors from their coordinates. Let
   
u1 v1
u =  u2  and v =  v2  .
u3 v3
Then we have:
u × v = (u1 i + u2 j + u3 k) × (v1 i + v2 j + v3 k)
 2.7. The cross product 93

= u1 v1 (i × i) + u1 v2 (i × j) + u1 v3 (i × k)
+ u2 v1 (j × i) + u2 v2 (j × j) + u2 v3 (j × k)
+ u3 v1 (k × i) + u3 v2 (k × j) + u3 v3 (k × k)
= u1 v1 0 + u1 v2 k − u1 v3 j
− u2 v1 k + u2 v2 0 + u2 v3 i
+ u3 v1 j − u3 v2 i + u3 v3 0
= (u2 v3 − u3 v2 )i + (u3 v1 − u1 v3 )j + (u1 v2 − u2 v1 )k.

The resulting formula for the cross product is summarized in the following Proposition.

Proposition 2.42: Coordinate description of cross product

The cross product can be computed as follows:
     
u1 v1 u2 v3 − u3 v2
 u2  ×  v2  =  u3 v1 − u1 v3  .
u3 v3 u1 v2 − u2 v1

We will now look at an example of how to compute a cross product.

Example 2.43: Find a cross product

Find u × v for the vectors    
1 3
u =  −1  and v =  −2  .
2 1

Solution. Using Proposition 2.42, we compute

       
1 3 (−1)(1) − (2)(−2) 3
 −1  ×  −2  =  (2)(3) − (1)(1)  =  5  .
2 1 (1)(−2) − (−1)(3) 1
♠
We use this concept in the following examples.

Example 2.44: Area of a parallelogram

Find the area of the parallelogram determined by the following vectors u and v:
   
1 3
u =  −1  , v =  −2  .
2 1

Solution. Notice that these vectors are the same as the ones given in Example 2.43. Recall from the
geometric description of the cross product that the area of the parallelogram is the magnitude of u × v.
94 Vectors in Rn

From Example 2.43, u × v = [3, 5, 1]T . Thus the area of the parallelogram is
p √
ku × vk = 32 + 52 + 12 = 35.
♠

Example 2.45: Area of a parallelogram

Find the area of the parallelogram with vertices (1, 0, 1), (2, 2, 3), (−1, 1, 3), and (0, 3, 5).

Solution. Let P = (1, 0, 1), Q = (2, 2, 3), R = (−1, 1, 3), and S = (0, 3, 5).

Q S

P R
−→ − →
First, we check that this really is a parallelogram. We have to have PQ = RS. Indeed, this is the case,
−→ −
→
as PQ = [2 − 1, 2 − 0, 3 − 1]T = [1, 2, 2]T and RS = [0 − (−1), 3 − 1, 5 − 3]T = [1, 2, 2]T . We also compute
−→ − →
PR = QS = [−2, 1, 2]T . The area of the parallelogram is
     
−→ 1 −2

2 q √
−→   
PQ × PR = 2 × 1 = −6
  
= 2 2 + (−6)2 + 52 = 65.
2 2 5
♠
We can also use this concept to find the area of a triangle, as in the following example.

Example 2.46: Area of triangle

Find the area of the triangle determined by the points (1, 2, 3), (0, 2, 5), and (5, 1, 2).

Solution. Let P = (1, 2, 3), Q = (0, 2, 5), and R = (5, 1, 2). The area of the triangle is exactly half of the
−→ −→
area of the parallelogram determined by the vectors PQ and PR.

P R
−→ −
→
We have PQ = [−1, 0, 2]T and PR = [4, −1, −1]T . The area of the parallelogram is the magnitude of the
cross product:
     
−1 4
−→ −


→  2 p √
PQ × PR = 0  ×  −1  =

 7  = 22 + 72 + 12 = 54.

2 −1 1
 2.7. The cross product 95

1
√ 3
√
Hence the area of the triangle is 2 54 = 2 6. ♠

In general, the area of the triangle determined by three points P, Q, R in R3 is given by

1 −→ − →

PQ × PR .
2

2.7.4. The box product

In this section, we explore another application of the cross product. Recall that we can use the cross
product to find the the area of a parallelogram. As we will now show, we can also use the cross product
together with the dot product to find the volume of a parallelepiped. We begin with a definition.

Definition 2.47: Parallelepiped

The parallelepiped determined by three vectors u, v, and w consists of the set of points of the form

ru + sv + tw,

where r, s,t are real numbers between 0 and 1, inclusive. The parallelepiped is a 3-dimensional
body bounded by parallelograms as shown in this picture.

w
v

Notice that the base of the parallelepiped is the parallelogram determined by the vectors u and v. There-
fore, its area is equal to ku × vk. The height of the parallelepiped is kwk cos θ , where θ is the angle
between w and u × v, as shown in this picture.

u×v w
v
θ
u

The volume of this parallelepiped is the area of the base times the height which is just

ku × vk kwk cos θ = (u × v) w. ·
This expression is known as the box product and is sometimes written as [u, v, w].
Consider what happens if you interchange v with w or u with w. Geometrically, we can see that this
merely introduces a minus sign. We find that the box product of three vectors equals the volume of the
96 Vectors in Rn

parallelepiped determined by the three vectors if the three vectors form a right-handed system, and the
negative of the volume if the vectors form a left-handed system. We summarize this in the following
proposition:

Proposition 2.48: The box product

·
Let u, v, w be three vectors in R3 that define a parallelepiped. The box product (u × v) w is equal
to:

• The volume of the parallelepiped, if u, v, w form a right-handed system.

• The negative of the volume of the parallelepiped, if u, v, w form a left-handed system.

In any case, the volume of the parallelepiped can be computed as the absolute value of the box
·
product, given by |(u × v) w|.

Example 2.49: Volume of a parallelepiped

Find the volume of the parallelepiped determined by the vectors
     
1 1 3
u=  
2 , v=  3 , w = 2 .
 
−5 −6 3

Solution. According to the above discussion, we can take the cross product of any two of these vectors,
and then the dot product with the third vector. The result will be either plus or minus the desired volume.
Therefore we can obtain the volume by taking the absolute value.
We first compute the cross product of u and v:
       
1 1 (2)(−6) − (−5)(3) 3
u × v =  2  ×  3  =  (−5)(1) − (1)(−6)  =  1 
−5 −6 (1)(3) − (2)(1) 1

Then we take the dot product of this vector with w:

   
3 3
· ·
(u × v) w =  1   2  = 9 + 2 + 3 = 14.
1 3

Thus, the volume of the parallelepiped is 14 cubic units. ♠

The following is a consequence of Proposition 2.48:
 2.7. The cross product 97

Corollary 2.50: Right- and left-handed systems of vectors

·
The box product (u × v) w is:

• Positive, if u, v, w form a right-handed system.

• Negative, if u, v, w form a left-handed system.

• Zero, if u, v, w are coplanar.

Example 2.51: Right- and left-handed systems of vectors

Which of the following systems of vectors u, v, w is right-handed? Which one is left-handed? Which
one is coplanar?

(a) u = [1, 2, 0]T , v = [0, 0, 1]T , w = [1, −1, 1]T .

(b) u = [1, 1, 1]T , v = [1, 2, 3]T , w = [0, 1, 1]T .

(c) u = [0, 1, 2]T , v = [1, 2, 2]T , w = [1, 1, 0]T .

Solution.

· ·
(a) We have [u, v, w] = (u × v) w = [2, −1, 0]T [1, −1, 1]T = 3, so the box product is positive and the
system of vectors u, v, w is right-handed.

· ·
(b) We have [u, v, w] = (u × v) w = [1, −2, 1]T [0, 1, 1]T = −1, so the box product is negative and the
system is left-handed.

· ·
(c) We have [u, v, w] = (u × v) w = [−2, 2, −1]T [1, 1, 0]T = 0, so the box product is zero and the
vectors are coplanar.

♠
We finish this section with a law involving the dot product and the cross product. It represents a funda-
mental observation that comes directly from the geometric definition of the box product.

Proposition 2.52: Box product law

·
Let u, v, and w be vectors. Then (u × v) w = u (v × w). ·
· ·
Proof. This follows from observing that both (u × v) w and u (v × w) compute the same box product,
i.e., they either both give the volume of the parallelepiped or they both give the negative of the volume.
Alternatively, we can calculate each product explicitly:

·
(u × v) w = u2 v3 w1 − u3 v2 w1 + u3 v1 w2 − u1 v3 w2 + u1 v2 w3 − u2 v1 w3 ,
·
u (v × w) = u2 v3 w1 − u3 v2 w1 + u3 v1 w2 − u1 v3 w2 + u1 v2 w3 − u2 v1 w3 .
98 Vectors in Rn

In Chapter 7, you will learn that these expressions are a special case of a determinant. ♠

Exercises

Exercise 2.7.1 Which of the following systems of vectors are right-handed? Which are left-handed?

(a) i, k, j, (b) j, k, i, (c) k, i, j, (d) k, j, i.

Exercise 2.7.2 Show that if a × u = 0 for every unit vector u, then a = 0.

  
1 3
Exercise 2.7.3 Find the area of the parallelogram determined by the vectors  2 ,  −2 .
3 1
  
1 4
Exercise 2.7.4 Find the area of the parallelogram determined by the vectors  0 ,  −2 .
3 1

Exercise 2.7.5 Find the area of the parallelogram with vertices (−2, 3, 1), (2, 1, 1), (1, 2, −1), and (5, 0, −1).

Exercise 2.7.6 Find the area of the triangle determined by the three points, (1, 0, 3), (4, 1, 0) and (−3, 1, 1).

Exercise 2.7.7 Find the area of the triangle determined by the three points, (1, 2, 3), (2, 3, 4) and (3, 4, 5).
Did something interesting happen here? What does it mean geometrically?

Exercise 2.7.8 Is u × (v × w) = (u × v) × w? What is the meaning of u × v × w? Explain. Hint: Try

(i × j) × j.

Exercise 2.7.9 Verify directly from the coordinate description of the cross product that u × v is orthogonal
to both u and v. Then show by direct computation that this coordinate description satisfies

ku × vk2 = kuk2 kvk2 − (u v)2

2 2
·
= kuk kvk (1 − cos2 (θ )),

where θ is the angle included between the two vectors. Explain why ku × vk has the correct magnitude.

·
Exercise 2.7.10 Prove the following formula by direct calculation: u × (v × w) = (u w) v − (u v) w. ·
Exercise 2.7.11 Use the formula from Exercise 2.7.10 to prove that the cross product satisfies the so-called
Jacobi identity:
u × (v × w) + v × (w × u) + w × (u × v) = 0.
 2.7. The cross product 99


  
1 1
Exercise 2.7.12 Find the volume of the parallelepiped determined by the vectors  −7 ,  −2 , and
−5 −6
 
3
 2 .
3

Exercise 2.7.13 Which of the following systems of vectors u, v, w are right-handed? Which are left-
handed? Which are coplanar?

(a) u = [1, 0, 1]T , v = [1, 2, 0]T , w = [0, 0, 1]T .

(b) u = [0, 1, 1]T , v = [−1, 2, 0]T , w = [1, 1, 2]T .

(c) u = [1, −1, 0]T , v = [1, 0, 1]T , w = [3, 1, 4]T .

(d) u = [1, 0, 0]T , v = [1, 2, 0]T , w = [2, 0, −1]T .

Exercise 2.7.14 Suppose u, v, and w are three vectors whose components are all integers. Can you
conclude the volume of the parallelepiped determined from these three vectors will always be an integer?

Exercise 2.7.15 What does it mean geometrically if the box product of three vectors equals zero?

Exercise 2.7.16 Show that

· ·
(u × v) w = (v × w) u = (w × u) v.·
Exercise 2.7.17 Use the formula from Exercise 2.7.10 to show that

· · ·
(u × v) ((v × w) × (w × z)) = ((v × w) z) ((u × v) w).

·
Exercise 2.7.18 Simplify ku × vk2 + (u v)2 − kuk2 kvk2 .

Exercise 2.7.19 This problem uses calculus. For u, v, w functions of t, prove that the derivative satisfies
the following product rules:

(u × v)′ = u′ × v + u × v′
· · ·
(u v)′ = u′ v + u v′
3. Lines and planes in Rn

3.1 Lines

Outcomes
A. Find the vector, parametric, and symmetric equations of a line.

B. Determine whether a point is on a given line.

C. Determine whether two lines intersect.

D. Find the angle between two lines.

E. Find the projection of a point onto a line.

We can use the concept of vectors and points to find equations for lines in Rn . Consider a straight line L
that passes through a point P in the direction given by a non-zero vector d.

L
Q

P d

The line L is infinitely long in both directions, although the picture only shows a finite part of it. To find
−→
an equation for this line, first suppose that Q is an arbitrary point on L. Then the vector PQ is parallel to
d. In other words, there exists some real number t such that
−→
PQ = t d.
If p is the position vector of P and q is the position vector of Q, we can write
−→
PQ = q − p.
Putting together the last two equations, we get q − p = t d, which we can write as
q = p + t d.
This is called the vector equation of the line L. The vector d is called the direction vector, and t is called
a parameter. The parameter t can be any real number; each time we plug in a different number for t, we
get a different point Q on the line. The following picture shows the effect of the parameter:

101
102 Lines and planes in Rn

d t=3
P t=3
t=2
t=1
t=0
t = −1 p
0

The following definition summarizes the above.

Definition 3.1: Vector equation of a line

Let p be a vector and d a non-zero vector. Then

q = p+t d

is the vector equation of a straight line L. Specifically, as the parameter t ranges over the real
numbers, q ranges over the position vectors of all the points Q on the line L. The vector d is called
the direction vector of the line.

Example 3.2: A line from a point and a direction vector

Find a vector equation for the line which contains the point P = (2, 0, 3) and has direction vector
d = [1, 2, 1]T .

Solution. The position vector of the point P is p = [2, 0, 3]T . The equation of the line is q = p + t d, which
we can write as      
x 2 1
 y  =  0 +t  2 .
z 3 1
♠

Example 3.3: A line from two points

Find a vector equation for the line through the points P = (1, 2, 0, 1) and R = (2, −4, 6, 3).

−→
Solution. We can use P as the base point; its position vector is p = [1, 2, 0, 1]T . We can use d = PR =
[1, −6, 6, 2] as the direction vector. Then a vector equation of the line is q = p + t d, which we can also
write as      
x 1 1
 y   2   −6 
     
 z  =  0 +t  6 .
w 1 2
♠
 3.1. Lines 103

When we write a vector equation in the form

     
x1 p1 d1
 x2   p2   d2 
     
 ..  =  .. +t  .. ,
 .   .   . 
xn pn dn

it is also called the component form of the vector equation.

Notice that the vector equation of a line is not unique. In fact, there are infinitely many vector equations
for the same line. For example, we can replace the parameter t with another parameter, say 3s or 1 − r.

Example 3.4: Change of parameter

Consider the vector equation from Example 3.2,
     
x 2 1
 y  =  0 +t  2 .
z 3 1

Find two other equations for the same line, by changing the parameter to 3s and to 1 − r.

Solution. If we let t = 3s, we get

     
x 2 1
 y  =  0  + 3s  2 
z 3 1
   
2 3
=  0 +s 6 .
 
3 3

If we let t = 1 − r, we get      
x 2 1
 y  =  0  + (1 − r)  2 
z 3 1
     
2 1 1
=  0 + 2 −r 2 
   
3 1 1
   
3 −1
=  2 + r −2  .
 
4 −1
♠
104 Lines and planes in Rn

Definition 3.5: Parametric equations of a line

A line with vector equation      
x1 p1 d1
 x2   p2   d2 
     
 . = . +t  . 
 ..   ..   .. 
xn pn dn
can also be written as a set of n scalar equations:

x1 = p1 + t d1 ,
x2 = p2 + t d2 ,
.
..
xn = pn + t dn ,

When written in this form, they are called the parametric equations of the line.

Example 3.6: Parametric equations

Find parametric equations for the line through the points P = (1, 2, 0, 1) and R = (2, −4, 6, 3).

Solution. This is a same line as in Example 3.3. We can easily convert the vector equation
     
x 1 1
 y   2   −6 
     
 z  =  0 +t  6 
w 1 2
to a set of parametric equations:
x = 1 + t,
y = 2 − 6t,
z = 6t,
w = 1 + 2t.
♠

Example 3.7: Determine whether a point is on a line

Determine whether the point P = (5, 8, 4) is on the line L given by the vector equation
     
x 1 2
 y  =  2 +t  3 .
z 1 1

Solution. The point P is on the line L if and only if there exists some t ∈ R such that
     
1 2 5
 2 +t  3  =  8 .
1 1 4
 3.1. Lines 105

Subtracting [1, 2, 1]T from both sides of the equation, this is equivalent to
   
2 4
t 3 = 6 .
  
1 3

We can write this as a set of parametric equations:

2t = 4,
3t = 6,
t = 3.

This is a system of three linear equations in one variable, and we quickly see that it is inconsistent.
Therefore, the point P does not lie on the line L. ♠

Example 3.8: Determine whether two lines intersect

Determine whether the lines
           
x 3 2 x 1 2
 y  =  1 +t  0  and  y  =  −1  + s  1 
z 0 1 z 5 −2

intersect. If yes, find the point of intersection.

Solution. The two lines intersect if and only if there exist t, s ∈ R such that
       
3 2 1 2
 1  + t  0  =  −1  + s  1  .
0 1 5 −2

Bringing s to the left-hand side, and subtracting [3, 1, 0]T from both sides of the equation, this is equivalent
to      
2 2 −2
t  0  − s  1  =  −2  .
1 −2 5
If we write this vector equation as a set of three parametric equations, it is a system of 3 linear equations
in 2 variables. The augmented matrix of the system is
 
2 −2 −2
 0 −1 −2  .
1 2 5

This system has reduced echelon form  

1 0 1
 0 1 2 ,
0 0 0
106 Lines and planes in Rn

and has the unique solution t = 1 and s = 2. Therefore, the lines intersect. (Other possible cases are: If
the system is inconsistent, the lines do not intersect. If the system has more than one solution, the lines
are identical). We find the point of intersection by plugging the parameter t = 1 into the equation of the
first line (or equivalently, but plugging s = 2 into the equation of the second line - doing it both ways is a
good way to double-check your answer). Therefore, the point of intersection is
     
3 2 5
 1 +1 0  =  1 .
0 1 1

♠
There is one other form for a line which is useful, which is the symmetric form. Consider the line given
by
x = 1 + 2t,
y = 1 − t,
z = 3 + 2t.
We can solve each equation for t:
x−1
t= 2 ,
y−1
t= −1 ,
z−3
t= 2 .
Finally, we can eliminate t from the equations by setting all three equations equal to one another:
x−1 y−1 z−3
= = .
2 −1 2
The latter is really a system of 2 equations in 3 variables. This is the symmetric form of the equation of
the line. In the following example, we look at how to convert the equation of a line from symmetric form
to parametric form.

Example 3.9: Change symmetric form to parametric form

Consider the line whose equations are given in symmetric form as
x−2 y−1 z+3
= = .
3 2 1
Find parametric and vector equations for this line.

Solution. We set all three quantities equal to t:

x−2 y−1 z+3
t= , t= , t= .
3 2 1
Solving these equations for x, y, z yields
x = 2 + 3t,
y = 1 + 2t,
z = −3 + t.
 3.1. Lines 107

These are the parametric equations for the line. The vector equation is
     
x 2 3
 y  =  1 +t  2 .
z −3 1

♠
We can use the dot product to find the angle between two intersecting lines. This is simply the smallest
angle between (any of) their direction vectors. The only subtlety here is that if u is a direction vector for a
line, then so is −u, and thus we will find pairs of complementary angles. We will take the smaller of the
two angles.

v
π −θ
−u θ
u

Example 3.10: Find the angle between two lines

Find the angle between the two lines
     
x 1 −1
L1 :  y  =  2  + t  1 
z 0 2

and      
x 0 2
L2 :  y  =  3  + s  1  .
z 2 −1

Solution. The direction vectors are

  
−1 2
u= 1  and v =  1 .
2 −1

The answer is given by

cos θ = ·
u v 1
=− ,
kuk kvk 2
which gives θ = 23π . Now the angles between any two direction vectors for these lines will either be 23π
or its complement φ = π − 23π = π3 . We choose the smaller angle, and therefore conclude that the angle
between the two lines is π3 . ♠
Finally, we will show how to use projections to find the shortest distance from a point to a line.
108 Lines and planes in Rn

Example 3.11: Shortest distance from a point to a line

 
2
Let L be the line which goes through the point P = (0, 4, −2) with direction vector d = 1 , and
2
let Q = (1, 3, 5). Find the shortest distance from Q to the line L, and find the point R on L that is
closest to Q.

Q
d
P
L
R

−→
Solution. In order to determine the shortest distance from Q to L, we will first find the vector PQ and then
−→
find the projection of this vector onto L. The vector PQ is given by
     
1 0 1
−→   
PQ = 3 − 4  =  −1  .
5 −2 7
Then, if R is the point on L closest to Q, it follows that
  
−→ 2 2
−
→ −→
PR = projd PQ = ·
d PQ
d =
15  
1 =
5 
1 .
kdk2 9
2
3
2
Now, the distance from Q to L is given by
−→ −→
→ √
−
RQ = PQ − PR = 26.
−
→ −
→
The point R is found by adding the vector PR to the position vector 0P for P as follows
     
0 2 10/3
 4  + 5  1  =  17/3  .
3
−2 2 4/3

Therefore, R = ( 10 17 4
3 , 3 , 3 ). ♠

Exercises

Exercise 3.1.1 Find the vector equation for the line through (−7, 6, 0) and (−1, 1, 4). Then, find the
parametric equations for this line.

  3.1.2 Find parametric equations for the line through the point (7, 7, 1) with direction vector
Exercise
1
d= 6 .
2
 3.1. Lines 109

Exercise 3.1.3 Parametric equations of the line are

x = t + 2,
y = 6 − 3t,
z = −t − 6.

Find a direction vector for the line and a point on the line.

Exercise 3.1.4 The equation of a line in two dimensions is written as y = x − 5. Find a vector equation
for this line.

Exercise 3.1.5 Find parametric equations for the line through (6, 5, −2, 3) and (5, 1, 2, 1).

Exercise 3.1.6 Consider the following vector equation for a line in R3 :

     
x 1 1
 y  =  2 +t  0 .
z 0 1

Find a new vector equation for the same line by doing the change of parameter t = 2 − s.

Exercise 3.1.7 Consider the line given by the following parametric equations:

x = 2t + 2,
y = 5 − 4t,
z = −t − 3.

Find symmetric equations for the line.

1
Exercise 3.1.8 Find the point on the line segment from P = (−4, 7, 5) to Q = (2, −2, −3) which is 7 of the
way from P to Q.

Exercise 3.1.9 Suppose a triangle in Rn has vertices at P, Q, and R. Consider the lines which are drawn
from a vertex to the mid point of the opposite side. Show these three lines intersect in a point and find the
coordinates of this point.

Exercise 3.1.10 Determine whether the lines

           
x 1 1 x 1 1
 y  =  1 +t  2  and  y  =  −1  + s  1 
z 2 2 z −4 −1

intersect. If yes, find the point of intersection.

Exercise 3.1.11 Determine whether the lines

           
x 2 1 x 1 1
 y  =  −1  + t  3  and  y  =  1 +s 2 
z 0 2 z 3 0
110 Lines and planes in Rn

intersect. If yes, find the point of intersection.

Exercise 3.1.12 Find the angle between the two lines

           
x 3 3 x 3 −1
 y  =  0  + t  −3  and  y  =  0 +s 2 .
z 1 0 z 1 2

Exercise 3.1.13 Let P = (1, 3

 2, 3) be a point in R . Let L be the line through the point P0 = (1, 4, 5) with
1
direction vector d = −1 . Find the shortest distance from P to L, and find the point Q on L that is
1
closest to P.

Exercise 3.1.14 Let P = (0, 2, 1) be a point in R3 . Let L be the line through the points P0 = (1, 1, 1) and
P1 = (4, 1, 2). Find the shortest distance from P to L, and find the point Q on L that is closest to P.

Exercise 3.1.15 When we computed the angle between two lines in Example 3.10, we calculated two
different angles and took the smaller of the two. Show that one can get the same answer by taking the
absolute value of the dot product, i.e., by solving

cos θ = ·
|u v|
.
kuk kvk

3.2 Planes

Outcomes
A. Find the vector and parametric equations of a plane in Rn .

B. Find the normal and standard equations of a plane in R3 .

C. Find the intersection of two planes, or of a line and a plane.

D. Find the angle between two planes, or between a line and a plane.

E. Find the shortest distance between a point and a plane.

Much like the above discussion with lines, vectors can be used to determine planes in Rn . Consider a point
P and two direction vectors d and e that are not parallel to each other. Then there is a unique plane passing
through P and containing d and e:
 3.2. Planes 111

e
P
d

The plane is infinite in each direction, although in the picture, we have only shown a small part of it. If p
is the position vector of P and q is the position vector of some other point in the plane, we have
q = p+t d+se
for some real numbers t and s. This is called the vector equation of the plane.

Definition 3.12: Vector equation of a plane

Let p be a vector and let d, e be non-zero, non-parallel vectors. Then

q = p+t d+se

is the vector equation of a plane.

The vector equation of a plane can also be written in component form

       
x1 p1 d1 e1
 x2   p2   d2   e2 
       
 ..  =  ..  + t  ..  , +s  .. 
 .   .   .   . 
xn pn dn en
and in parametric form
x1 = p1 + t d1 + s e1 ,
x2 = p2 + t d2 + s e2 ,
..
.
xn = pn + t dn + s en .
The latter set of equations are also called the parametric equations of the plane.

Example 3.13: Vector and parametric equations

Find vector and parametric equations for the plane through the points P = (1, 2, 0, 0), Q = (2, 2, 0, 1),
and R = (0, 1, 1, 0).

−→ −
→
Solution. We can use P as the base point and PQ and PR as the direction vectors. We have
           
2 1 1 0 1 −1
−→  2     
 −  2  =  0  and − →     
 1  −  2  =  −1  .

PQ =  0   0   0  PR =  1   0   1 
1 0 1 0 0 0
112 Lines and planes in Rn

Therefore the vector equation is

    
  
x 1 1 −1
 y   2   0   −1 
   +t  +s  .
 z = 0   0   1 
w 0 1 0

We can also write this as a system of parametric equations:

x = 1 + t − s,
y = 2 − s,
z = s,
w = t.

♠
Note that the vector and parametric equations of a plane are not unique. For example, in Example 3.13,
−→
we could have equally used Q or R as the base point, and/or used QR as one of the direction vectors. In
each case we would have obtained a different equation for the same plane.

Example 3.14: Determine whether a point is on a plane

Determine whether the point S = (4, 4, −2, 1) lies on the plane through the points P = (1, 2, 0, 0),
R = (2, 2, 0, 1), and Q = (0, 1, 1, 0).

Solution. We already found the parametric equations for this plane in Example 3.13. To determine whether
the point S = (4, 4, −2, 1) lies on this plane, we must substitute its coordinates into the parametric equa-
tions:
4 = 1 + t − s,
4 = 2 − s,
−2 = s,
1 = t.
This is a system of linear equations. We solve it to find that it has the unique solution (t, s) = (1, −2).
Therefore, the point S lies on the given plane, and more specifically, it is the point that corresponds to the
parameters t = 1 and s = −2. ♠

In the special case of 3 dimensions, a plane can also be described by a point and a normal vector. A
normal vector of a plane is a vector that is perpendicular to the plane.

n
Q
P
 3.2. Planes 113

Given a non-zero vector n in R3 and a point P, there exists a unique plane that contains P and has n as a
normal vector. We wish to find an equation for this plane. If Q is an arbitrary point on the plane, then by
−→ −→
·
definition, the normal vector is orthogonal to the vector PQ. Writing this as a formula, we have n PQ = 0.
−→
If p and q are the position vectors of P and Q, respectively, we have PQ = q −p, and therefore the equation
of the plane can be written as
·
n (q − p) = 0,
or equivalently,
·
n q = n p. ·
This is called the normal equation of the plane. Note that in this equation, n and p are given and fixed,
whereas q is a variable ranging over the position vectors of all points on the plane.

Definition 3.15: Normal equation of a plane in R3

Let n be a non-zero vector in R3 , and let P be a point with position vector p. Then there is a unique
plane through P with normal vector n. It is described by the equation

·
n q = n p. ·
This equation is called the normal equation of the plane.

Example 3.16: Finding the normal equation of a plane

Find the normal equation of the plane through the point P = (1, 3, 0) and orthogonal to n = [2, 1, 1]T .

Solution. Let p = [1, 3, 0]T be the position vector of P, and let q = [x, y, z]T be the position vector of some
· ·
arbitrary point Q in the plane. The normal equation is n q = n p, which we can write in component form:
       
2 x 2 1

1
·
 1   y  =  1   3 .
z 1
· 0

·
We can pre-compute the dot product on the right-hand side: n p = 1(2) + 3(1) = 0(1) = 5. Therefore,
the normal equation can also be written as
   
2 x

1 z
·
 1   y  = 5.

♠
Notice that the last equation in Example 3.16 can also be written in the form

2x + y + z = 5.

This last form is called the standard equation of the plane.

114 Lines and planes in Rn

Definition 3.17: Standard equation of a plane in R3

Let n = [a, b, c]T be the normal vector for a plane that contains the point P = (x0 , y0 , z0 ). The
standard equation of the plane is given by

ax + by + cz = d,

where a, b, c, d ∈ R and d = ax0 + by0 + cz0 .

Example 3.18: Normal and standard equations

Find normal and standard equations for the plane through the points P = (0, 1, 3), Q = (2, −1, 0),
and R = (1, 2, 2).

Solution. We first need to find a normal vector for the plane. Since the normal vector must be perpendicular
−→ −
→
to the plane, it must be orthogonal to both PQ and PR. We can therefore use the cross product to compute
a normal vector for the plane:
     
2 1 5
−→ − →
n = PQ × PR =  −2  ×  1  =  −1  .
−3 −1 4

n
R
P
Q

Now we can easily obtain the normal equation from any point on the plane (say P) and the normal vector
we just calculated:        
5 x 5 0

4
·
 −1   y  =  −1   1 
z 4 3
·
We get the standard equation by computing the dot products on the left- and right-hand sides:
5x − y + 4z = 11.
It is worthwhile to double-check the answer by substituting each of the three original points P, Q, and R
into this equation. For example, for Q = (2, −1, 0), we obtain 5(2) − (−1) + 4(0), which is indeed 11.
♠

Example 3.19: Find the normal vector of a plane

Find a normal vector for the plane 2x + 3y − z = 7.
 3.2. Planes 115

Solution. The standard equation 2x + 3y − z = 7 can be rewritten as a normal equation

   
2 x

−1 z
·
 3   y  = 7.

Therefore, 

2
n= 3 
−1
is a normal vector for the plane. ♠

Example 3.20: Determine whether a point is on a plane

Let n = [1, 2, 3]T be the normal vector for a plane which contains the point P = (2, 1, 4). Determine
if the point Q = (5, 4, 1) is in this plane.

Solution. By Definition 3.15, Q is a point in the plane if and only if

·
n q = n p,·
where p and q are the position vectors of P and Q, respectively. Given n, P, and Q as above, this equation
becomes        
1 5 1 2

3
·
 2   4  =  2   1 .
1 3
· 4
Since both sides of the equation are equal to 16, the equation is true. So the point Q is indeed in the plane
determined by n and P. ♠

Example 3.21: Vector equation from normal equation

Find a vector equation for the plane x + 3y − 2z = 7.

Solution. This is the same thing as finding the general solution of a system of one linear equation in 3
variables. Since there is only a single equation x + 3y − 2z = 7, it is already in echelon form. The variables
y and z are free, so we set them equal to parameters: z = t and y = s. The variable x is a pivot variable, and
we get x = 7 + 2t − 3s. So the general solution of the equation is
       
x 7 2 −3
 y  =  0 +t  0 +s  1 .
z 0 1 0

This is also a vector equation for the plane. ♠

116 Lines and planes in Rn

Example 3.22: Intersection of two planes

Find the intersection of the planes x − 2y + z = 0 and 2x − 3y − z = 4.

Solution. Finding the intersection means finding all of the points (x, y, z) that are on both planes simulta-
neously. This is the same as solving the system of equations

x − 2y + z = 0,
2x − 3y − z = 4.

We solve the system by Gauss-Jordan elimination:

     
1 −2 1 0 R2 ← R2 −2R1 1 −2 1 0 R1 ← R1 +2R2 1 0 −5 8
≃ ≃ .
2 −3 −1 4 0 1 −3 4 0 1 −3 4

Therefore, the general solution is      

x 8 5
 y  =  4 +t  3 ,
z 0 1
where t is a parameter. This is the parametric equation of a line. Therefore, the two planes intersect in a
line. Specifically, the intersection is the line through the point (8, 4, 0) with direction vector [5, 3, 1]T . ♠

Example 3.23: Intersection of a line and a plane

Find the intersection of the line
     
x 1 −1
 y  =  2 +t  1 
z 0 2

and the plane 2x + 2y − z = 2.

Solution. Let us write

      

1 −1 2 x
p =  2 , d =  1 , n =  2 , and q =  y  .
0 2 −1 z

·
Then the equation of the line is q = p + t d and the equation of the plane is n q = 2. Substituting the
·
first equation into the second one, we get n (p + t d) = 2. Using distributivity of the dot product, we can
· · · ·
write this last equation as n p + t(n d) = 2. By computing the dot products n p = 6 and n d = −2, this
equation simplifies to 6 − 2t = 2, or t = 2. Therefore, the line intersects the plane when t = 2, or at the
point        
x 1 −1 −1
 y  =  2 +2 1  =  4 .
z 0 2 4
 3.2. Planes 117

An alternative method is to directly substitute the parametric equation of the line, x = 1 − t, y = 2 + t, and
z = 2t, into the equation of the plane, 2x + 2y − z = 2. In this case, we get 2(1 − t) + 2(2 + t) − (2t) = 2,
which we can solve for t to obtain t = 2. ♠

The next few examples are concerned with calculating angles between planes, angles between lines and
planes, and finding the distance between points and planes.

Example 3.24: Find the angle between two planes

Find the angle between the planes 7x − y = 5 and 4x + 3y + 5z = 3.

Solution. The angle between two planes is the same thing as the angle between their normal vectors.

n1 n2

The normal vectors are n1 = [7, −1, 0]T and n2 = [4, 3, 5]T . The angle between them is given by

cos θ = ·
n1 n2
=
25 1
= .
kn1 k kn2 k 50 2

Therefore, the angle is arccos( 12 ) = π /3, or 60 degrees. ♠

Example 3.25: Find the angle between a line and a plane

Find the angle between the line
     
x 1 2
 y  =  2  + t  −1 
z 0 −2

and the plane 2x + 2y − z = 2.

Solution. To get the angle θ between the plane and the line, we can compute the angle φ between the
direction vector of the line and the normal vector of the plane, and then take θ = π2 − φ .
118 Lines and planes in Rn

n d

φ
θ

The direction vector of the line is [2, −1, −2]T and the normal vector of the plane is [2, 2, −1]. We have

cos φ =
n d· 4
= ,
knk kdk 9
π
and therefore φ = arccos( 49 ) ≈ 1.11 radians. We have θ = 2 − φ ≈ 0.46 radians, or about 26.4 degrees.
♠

Example 3.26: Shortest distance from a point to a plane

Find the shortest distance from the point P = (3, 2, 3) to the plane given by 2x + y + 2z = 2, and find
the point Q on the plane that is closest to P.

Solution. In this problem, we are going to use the projection of one vector onto another, which was
introduced in Section 2.6.5. Pick an arbitrary point R on the plane. Then, it follows that
−→ −
→
QP = projn RP
−→

and QP is the shortest distance from P to the plane. Further, the position vector of the point Q can be
−→
computed as q = p − QP, where p is the position vector of P.

P
−
→ −→
RP QP
n

R
Q

 
2
From the above scalar equation, we have that n = 1 . Now, choose any point on the plane, for example,
2
R = (1, 0, 0) (notice that this satisfies 2x + y + 2z = 2). Then,
     
3 1 2
−
→      
RP = 2 − 0 = 2 .
3 0 3
 3.2. Planes 119

−→ −
→
Next, compute QP = projn RP.
  
→!
− 2 2
−→ −
→
QP = projn RP = ·
n RP
n =
12  
1
4
=  1 .
knk2 9
2
3
2
−→

Then, QP = 4 so the shortest distance from P to the plane is 4. To find the point Q on the plane that is
closest to P, we have      
3 2 1
−→   4  1 
q = p − QP = 2 − 1 = 2 .
3 3
3 2 1
Therefore, Q = ( 13 , 23 , 13 ). ♠

Exercises

Exercise 3.2.1 Find vector and parametric equations for the plane through the points P = (0, 1, 1), Q =
(−1, 2, 1), and R = (1, 1, 2).

Exercise 3.2.2 Consider the following vector equation for a plane in R4 :

       
x 1 1 −1
 y   2   0   −1 
       
 z  =  0 +t  0 +s  1 .
w 0 1 0
Find a new vector equation for the same plane by doing the change of parameters t = 1 − r1 , s = r1 + r2 .

Exercise 3.2.3 Determine which of the following points lie on the plane through the points P = (2, 6, 1),
R = (1, 4, 1), and Q = (1, 2, −1).

(a) S1 = (1, 2, 4).

(b) S2 = (1, 5, 2).
(c) S3 = (0, 0, 0).

Exercise 3.2.4 Use cross products to find the normal vector to the plane going through the points P =
(1, 2, 3), Q = (−2, 1, 8) and R = (2, 2, 2).

Exercise 3.2.5 Find normal and standard equations of the plane through the point P = (1, 1, 2) and
orthogonal to n = [1, 0, −1]T .

Exercise 3.2.6 Find normal and standard equations for the plane through the points P = (2, 1, 0), Q =
(1, −1, 0), and R = (1, 1, −1).
120 Lines and planes in Rn

Exercise 3.2.7 Find a vector equation for the plane 2x + y − z = 1.

Exercise 3.2.8 The chapter mentions that the normal equation and standard equation of a plane only
work in R3 , and not in general Rn . Why does the equation ax + by + cz + dw = e not describe a plane in
R4 ?

Exercise 3.2.9 Find the intersection between the planes x + 3y + 4z = 3 and 2x + 5y − z = 2. Is the
intersection a line, a plane, or empty?

Exercise 3.2.10 Find the intersection of the line

     
x 1 2
 y  =  1  + t  −1 
z 1 2
and the plane x + 3y + z = 6. Is the intersection a point, a line, or empty?

Exercise 3.2.11 Find the angle between the planes x + y = 5 and 2x + y − z = 4.

Exercise 3.2.12 Find the angle between the line

     
x 0 1
 y  =  3 +t  1 
z 7 4
and the plane 4x + 7y + 4z = 6.

Exercise 3.2.13 In Example 3.25, we calculated the angle θ between a line and a plane by calculating
the angle φ between the direction vector of the line and the normal vector of the plane according to the
formula
cos φ =
n d ·
knk kdk
π
and then taking θ = 2 −φ.

(a) Explain what happens when the dot product is negative. How should we adjust the formula to ensure
that θ is always between 0 and π2 ?
(b) Show that one can get the answer in a single step with the formula

sin θ = ·
|n d|
.
knk kdk

Exercise 3.2.14 Find the shortest distance from the point P = (1, 1, −1) to the plane given by x+2y+2z =
6, and find the point Q on the plane that is closest to P.

Exercise 3.2.15 Use Exercise 2.7.15 to find an equation of a plane containing the two vectors p and q
and the point 0. Hint: If (x, y, z) is a point in this plane, the volume of the parallelepiped determined by
(x, y, z) and the vectors p, q equals 0.
4. Matrices

4.1 Definition and equality

Outcomes
A. Identify the dimension and entries of a matrix.

B. Check equality of matrices.

We have solved systems of equations by writing them in terms of an augmented matrix and then doing
row operations. It turns out that matrices are important not only for systems of equations but also for many
other purposes.

Definition 4.1: Matrix

A matrix is a rectangular array of numbers
 
a11 a12 · · · a1n
 a21 a22 · · · a2n 
 
A =  .. . . . . .. ,
 . .. . 
am1 am2 · · · amn

where the ai j are scalars, called the entries or components of A. The size or dimension of a matrix
is defined as m × n, where m is the number of rows and n is the number of columns.

For example, here is a 3 × 4-matrix (pronounced “three-by-four matrix”):

 
1 2 3 4
 5 2 8 7 .
6 −9 1 2

This is a 3 ×4-matrix because there are three rows and four columns. When specifying the size of a matrix,
we always list the number of rows before the number of columns.
Entries of the matrix are identified according to their position. The (i, j)-entry of a matrix is the entry
in the i th row and j th column, and is often denoted ai j . For example, in the above matrix, the (2, 3)-entry
 
is the entry in the second row and the third column, and is equal to 8. We sometimes use A = ai j as

121
122 Matrices

a short-hand notation for the entire m × n-matrix whose (i, j)-entry is equal to ai j for all i = 1, . . . m and
j = 1, . . . , n.
There are various operations which are done on matrices of appropriate sizes. Matrices can be added
and subtracted, multiplied by a scalar, and multiplied by other matrices. We will never divide a matrix by
another matrix, but we will see later how matrix inverses play a similar role.

Definition 4.2: Equality of matrices

Two matrices are equal  if they
 have
 the same size and the same corresponding entries. More
precisely, if A = ai j and B = bi j are two m × n-matrices, then A = B means that ai j = bi j for all
i = 1, . . . m and j = 1, . . ., n.

For example,  
0 0  
 0 0 = 0 0
6
0 0
0 0
because they are different sizes. Also,
   
0 1 1 0
6=
3 2 2 3

because, although they are the same size, their corresponding entries are not identical.
There are special names for matrices of certain dimensions: some matrices are called square matrices,
columns vectors, or row vectors.

Definition 4.3: Square matrix

A matrix of size n × n is called a square matrix. In other words, A is a square matrix if it has the
same number of rows and columns.

Definition 4.4: Column vectors and row vectors

A matrix of size n × 1 is called a column vector. A matrix of size 1 × n is called a row vector.
Here is an example of a column vector X and a row vector Y :
 
x1
   
X =  ...  , Y = y1 · · · yn .
xn

We have already encountered column vectors in Chapter 2. When we use the term vector without fur-
ther qualification, we always mean a column vector. Also recall from Definition 2.1 that the set of n-
dimensional column vectors is called Rn .
 4.2. Addition 123

Exercises

Exercise 4.1.1 Can a column vector ever be equal to a row vector?

Exercise 4.1.2 Find scalars x, y, z such that the following two matrices are equal.
   
x −1 2 y
and .
2 4 z 4

Exercise 4.1.3 What are the dimensions of the following matrices?

 
  1 0
  3 4 1
1 −2 0  4 0 
A= , B = 1 3 1 ,
 C=
 2
.
4 3 2 0 
6 2 2
0 0

 
1 2 1
Exercise 4.1.4 What is the (2, 3)-entry of the matrix  −4 4 7 ?
6 −5 3

4.2 Addition

Outcomes
A. Perform the operations of matrix addition and subtraction.

B. Identify when these operations are not defined.

C. Apply the algebraic properties of matrix addition to manipulate an algebraic expression in-
volving matrices.

To add two matrices, the matrices have to be of the same size. The addition works by simply adding
corresponding entries of the matrices.

Definition 4.5: Addition of matrices

   
Let A =  ai j and B = bi j be two m × n-matrices. Then A + B = C where C is the m × n-matrix
C = ci j defined by
ci j = ai j + bi j
124 Matrices

Example 4.6: Addition of matrices

Add the following matrices.
   
1 2 3 5 2 3
A= , B=
1 0 4 −6 2 1

Solution. Notice that both A and B are of size 2 × 3. Since A and B are of the same size, the addition is
possible. Using Definition 4.5, the addition is done as follows.
       
1 2 3 5 2 3 1+5 2+2 3+3 6 4 6
A+B = + = = .
1 0 4 −6 2 1 1 + (−6) 0 + 2 4 + 1 −5 2 5
♠
On the other hand, the matrices
 
1 2  
 3 4  −1 4 8
and
2 8 5
5 2
cannot be added, because one has size 3 × 2 while the other has size 2 × 3.

Definition 4.7: The zero matrix

The m × n zero matrix is the m × n-matrix in which all entries are equal to zero. It is denoted by 0.

Note there is a zero matrix for every size. For example, there is a 2 × 3 zero matrix, a 3 × 4 zero matrix,
and so on.

Example 4.8: The zero matrix

 
0 0 0
The 2 × 3 zero matrix is 0 = .
0 0 0

Definition 4.9: Negative of a matrix and subtraction

 
The negative of a matrix A = ai j is defined to be −A = [−ai j ]. In other words, it is obtained
by negating every entry of A. To subtract two matrices, we simply add the negative of the second
matrix to the first one, i.e., A − B = A + (−B). This is just the same as componentwise subtraction.

Example 4.10: Subtraction

Subtract the following matrices.
   
1 2 3 5 2 3
A= , B=
1 0 4 −6 2 1
 4.2. Addition 125

Solution.
       
1 2 3 5 2 3 1−5 2−2 3−3 −4 0 0
A−B = − = = .
1 0 4 −6 2 1 1 − (−6) 0 − 2 4 − 1 7 −2 3

♠
Addition of matrices obeys the same properties as addition of vectors.

Proposition 4.11: Properties of matrix addition

Let A, B and C be matrices of the same size. Then, the following properties hold.

• The commutative law of addition

A + B = B + A.

• The associative law of addition

(A + B) +C = A + (B +C).

• The existence of an additive unit

A + 0 = A.

• The existence of an additive inverse

A + (−A) = 0.

Proof. To prove the commutative law of addition, let A and B be matrices of the same size. We want to
show that A + B = B + A. To do so, we use the definition of matrix addition given in Definition 4.5. We
have    
A + B = ai j + bi j = bi j + ai j = B + A.
The proof of the other properties are similar, and are left as an exercise. ♠

Exercises

Exercise 4.2.1 For the following pairs of matrices, determine if the sum A + B and the difference A − B
are defined. If so, calculate them.
   
1 0 0 1
(a) A = , B= .
0 1 1 0
   
2 1 2 −1 0 3
(b) A = , B= .
1 1 0 0 1 4
126 Matrices

 
1 0  
2 7 −1
(c) A =  −2 3 , B= .
0 3 4
4 2

Exercise 4.2.2 For each matrix A, find the matrix −A.

 
1 2
(a) A =
2 1
 
−2 3
(b) A =
0 2
 
0 1 2
(c) A =  1 −1 3 
4 2 0

   
1 2 −1 0 3 0
Exercise 4.2.3 Let A = and B = .
−1 4 0 1 −1 1
Find a matrix X such that (A + X ) − (B + 0) = B + A. Hint: first use the properties of matrix addition to
simplify the equation and solve for X .

Exercise 4.2.4 Using only the properties given in Proposition 4.11, show that if A + B = 0, then B = −A.

Exercise 4.2.5 Using only the properties given in Proposition 4.11, show A + B = A implies B = 0.

4.3 Scalar multiplication

Outcomes
A. Multiply a matrix by a scalar, and take linear combinations of matrices.

B. Identify when these operations are not defined.

C. Apply the algebraic properties of matrix addition and scalar multiplication to manipulate an
algebraic expression involving matrices.

The multiplication of a scalar by a matrix is called the scalar multiplication of matrices. The new matrix
is obtained by multiplying every entry of the original matrix by the given scalar, as in the following
example.    
1 2 3 4 3 6 9 12
3 5 2 8 7  =  15 6 24 21  .
6 −9 1 2 18 −27 3 6
 4.3. Scalar multiplication 127

The formal definition of scalar multiplication is as follows.

Definition 4.12: Scalar multiplication of a matrix

   
If k is a scalar and A = ai j is a matrix, then kA = kai j .

Example 4.13: Linear combination of matrices

Find 2A − 3B, where    
1 2 3 5 2 3
A= and B =
1 0 4 −6 2 1

Solution.
         
1 2 3 5 2 3 2 4 6 15 6 9 −13 −2 −3
2A − 3B = 2 −3 = − = .
1 0 4 −6 2 1 2 0 8 −18 6 3 20 −6 5
♠
Scalar multiplication of matrices obeys the same properties as scalar multiplication of vectors.

Proposition 4.14: Properties of scalar multiplication

Let A and B be matrices of the same size, and let k, ℓ be scalars. Then the following properties hold.

• The distributive law over matrix addition

k(A + B) = kA + kB.

• The distributive law over scalar addition

(k + ℓ)A = kA + ℓA.

• The associative law for scalar multiplication

k(ℓA) = (kℓ)A.

• The rule for multiplication by 1

1A = A.

Exercises

Exercise 4.3.1 For each matrix A, find the products (−2)A, 0A, and 3A.
128 Matrices

 
1 2
(a) A =
2 1
 
−2 3
(b) A =
0 2
 
0 1 2
(c) A =  1 −1 3 
4 2 0

Exercise 4.3.2 Find scalars x, y, z, w such that

         
1 2 0 1 0 2 1 0 1 2
x +y +z +w = .
3 5 1 1 −1 2 0 2 6 5

Exercise 4.3.3 Using only the properties given in Propositions 4.11 and 4.14, show that 0A = 0. Here the
0 on the left is the scalar 0 and the 0 on the right is the zero matrix of appropriate size.

4.4 Matrix multiplication

Outcomes
A. Use two different methods to multiply a matrix and a vector.

B. Multiply two matrices using the componentwise method, the column method, or the row
method.

C. Identify when these operations are not defined.

D. Write a system of linear equations in vector form and matrix form.

E. Demonstrate that matrix multiplication is not commutative.

F. Use algebraic properties of matrix multiplication to solve matrix equations.

4.4.1. Multiplying a matrix and a vector

One of the most important uses of a matrix is to multiply a matrix by a vector. In fact, this is one of the
reasons matrices were invented. Let us start by considering an alternative way of writing a system of linear
equations.
 4.4. Matrix multiplication 129

Definition 4.15: The vector form of a system of linear equations

Suppose we have a system of equations given by

a11 x1 + . . . + a1n xn = b1 ,
a21 x1 + . . . + a2n xn = b2 ,
.
..
am1 x1 + . . . + amn xn = bm .

We can express this system in vector form, which is as follows:

       
a11 a12 a1n b1
 a21   a22   a2n   b2 
       
x1  ..  + x2  ..  + . . . + xn  ..  =  .. .
 .   .   .   . 
am1 am2 amn bm

Notice that each vector used here is one column from the corresponding augmented matrix. There is
one vector for each variable in the system, along with the constant vector. The left-hand side is a linear
combination of column vectors. Linear combinations of column vectors are so important that we introduce
a special notation for them.

Definition 4.16: The product of a matrix and a vector, by columns

The product of an m × n-matrix A and an n-dimensional column vector x is an m-dimensional
column vector, defined as a linear combination of the columns of A as follows:
        
a11 a12 · · · a1n x1 a11 a12 a1n
 a21 a22 · · · a2n   x2   a21   a22   a2n 
        
Ax =  .. .. . . ..   ..  = x1  ..  + x2  ..  + . . . + xn  ..  .
 . . . .  .   .   .   . 
am1 am2 · · · amn xn am1 am2 amn

In other words, we can think of the vector x as encoding instructions for how to take a linear combination
of the columns of A. The product Ax is computed by taking x1 times the first column of A, plus x2 times
the second column of A, and so on. For this to work, A must have the same number of columns as x has
components.

Example 4.17: Multiplying a matrix and a vector, by columns

Compute the product  
  7
1 2 3  8 .
4 5 6
9
130 Matrices

Solution. We have
 
  7        
1 2 3  8  = 7 1 2 3 50
+8 +9 = .
4 5 6 4 5 6 122
9

♠
There is another way of looking at the product of a matrix and a vector. Instead of looking at the columns
of A, we can look at the rows.

Proposition 4.18: The product of a matrix and a vector, by rows

Let A be an m × n-matrix and let x be an n-dimensional column vector. The product Ax can also be
written like this:
    
a11 a12 · · · a1n x1 a11 x1 + a12 x2 + . . . + a1n xn
 a21 a22 · · · a2n   x2   a21 x1 + a22 x2 + . . . + a2n xn 
    
Ax =  .. . ..
. .   ..  = 
. .
. .
 . . . .  .   . 
am1 am2 · · · amn xn am1 x1 + am2 x2 + . . . + amn xn

Example 4.19: Multiplying a matrix and a vector, by rows

Compute the product  
  7
1 2 3  8 
4 5 6
9
by rows.

Solution. We have
 
  7    
1 2 3  8  = 1 · 7 + 2 · 8 + 3 · 9 50
= .
4 5 6 4·7+5·8+6·9 122
9

Note that the is exactly the same answer as before. ♠

When we use Definition 4.16, we calculate the product by looking at one column of A at a time. When we
use Proposition 4.18, we calculate the product by looking at one row of A at a time. As the above examples
show, both methods give exactly the same answer. Please convince yourself that this is true in general.
This ability to switch back and forth between a column-based viewpoint and a row-based viewpoint is one
of the central tools of linear algebra.
Using the above operation, we can also write a system of linear equations in matrix form. In this
form, we express the system as a matrix multiplied by a vector.
 4.4. Matrix multiplication 131

Definition 4.20: The matrix form of a system of linear equations

Suppose we have a system of equations given by

a11 x1 + . . . + a1n xn = b1 ,
a21 x1 + . . . + a2n xn = b2 ,
.
..
am1 x1 + . . . + amn xn = bm .

Then we can express this system in matrix form, which is as follows:

    
a11 a12 · · · a1n x1 b1
 a21 a22 · · · a2n   x2   b2 
    
 .. ... . . . ...   .  =  . .
 .   ..   .. 
am1 am2 · · · amn xn bm

The matrix form of a system of equations is therefore written as Ax = b, where A is the coefficient matrix
of the system, x is an n-dimensional column vector constructed from the variables of the system, and b is
an m-dimensional column vector constructed from the constant terms of the system. Any system of linear
equations can be written in this form.

4.4.2. Matrix multiplication

The multiplication of a matrix and a vector from the previous section is a special case of the operation of
multiplying two matrices, which we now define.

Definition 4.21: Matrix multiplication

   
Let A = ai j be an m × n-matrix, and let B = b jk be an n × p-matrix. Then their product is the
m × p-matrix AB = [cik ] whose (i, k)-entry is defined by

cik = ai1 b1k + ai2 b2k + . . . + ain bnk .

For matrices A and B, in order to form the product AB, the number of columns of A must equal the number
of rows of B. Consider a product AB where A has dimensions m × n and B has dimensions n × p. Then the
dimensions of the product are given by
these must match!
d
( m × n) (n × p) = m × p.

Note that the two outside numbers give the dimensions of the product. If the two middle numbers do not
match, we cannot multiply the matrices.
132 Matrices

To better visualize the rule of matrix multiplication, suppose

   
a11 a12 · · · a1n b11 b12 · · · b1p
 a21 a22 · · · a2n   b21 b22 · · · b2p 
   
A =  .. .. . . ..  and B =  .. .. .. . .
 . . . .   . . . .. 
am1 am2 · · · amn bn1 bn2 · · · bnp

Then their product  

c11 c12 · · · c1p
 c21 c22 · · · c2p 
 
AB =  .. .. . . . 
 . . . .. 
cm1 cm2 · · · cmp
is an m × p-matrix whose (i, k)-entry is defined by

cik = ai1 b1k + ai2 b2k + . . . + ain bnk .

Note that we can also write this as

 
b1k
 
 b2k 

cik = ai1 ai2 · · · ain  .. .
 . 
bnk

In other words, the (i, k)-entry of the matrix product AB is a kind of dot product of the i th row of A with
the k th column of B.

Example 4.22: Matrix multiplication

Find AB, where  
  1 2 0
1 2 1
A= and B =  0 3 1 .
0 2 1
−2 1 1

Solution. First, let us note that since A has size 2 × 3 and B has size 3 × 3, the product AB is well-defined
and has size 2 × 3. Let C = AB. We compute each of the six entries of C:

• The (1, 1)-entry is the first row of A times the first column of B: c11 = 1 · 1 + 2 · 0 + 1 · (−2) = −1.

• The (1, 2)-entry is the first row of A times the second column of B: c12 = 1 · 2 + 2 · 3 + 1 · 1 = 9.

• The (1, 3)-entry is the first row of A times the third column of B: c13 = 1 · 0 + 2 · 1 + 1 · 1 = 3.

• The (2, 1)-entry is the second row of A times the first column of B: c21 = 0 · 1 + 2 · 0 + 1 · (−2) = −2.

• The (2, 2)-entry is the second row of A times the second column of B: c22 = 0 · 2 + 2 · 3 + 1 · 1 = 7.

• The (2, 3)-entry is the second row of A times the third column of B: c23 = 0 · 0 + 2 · 1 + 1 · 1 = 3.
 4.4. Matrix multiplication 133

Therefore, we have  
−1 9 3
AB = .
−2 7 3
♠
As this example shows, calculating matrix products one component at a time can be an extremely repetitive
and tedious process. Fortunately, we can speed this up by considering whole columns at once.

Proposition 4.23: Matrix multiplication, column method

Let A be an m × n-matrix, and let B be an n × p-matrix. Suppose that the columns of B are
b1 , b2 , . . . , b p . Then the columns of AB are

Ab1 , Ab2 , . . . , Ab p .

In other words, the k th column of the matrix product AB is equal to A times the k th column of B.

Example 4.24: Matrix multiplication by the column method

Find the matrix product AB by the column method, where
 
  1 2 0
1 2 1
A= and B =  0 3 1 .
0 2 1
−2 1 1

Solution. We multiply A by each of the columns of B:

 
  1        
1 2 1   1 2 1 −1
0 = 1 +0 −2 = ,
0 2 1 0 2 1 −2
−2
 
  2        
1 2 1   1 2 1 9
3 = 2 +3 +1 = ,
0 2 1 0 2 1 7
1
 
  0        
1 2 1   1 2 1 3
1 = 0 +1 +1 = .
0 2 1 0 2 1 3
1
The resulting three column vectors form the columns of AB. Thus,
 
−1 9 3
AB = .
−2 7 3

♠
Of course, the answer in Example 4.24 is the same as that in Example 4.22. Please convince yourself that
both methods of matrix multiplication give the same answer, since they each ultimately calculate the same
134 Matrices

thing. Nevertheless, with a bit of practice, the column method is much faster, and you can even learn to
multiply matrices in your head! The key to understanding the column method is that each column of B
provides instructions for taking a linear combination of the columns of A. The method works especially
well if B contains many zeros and ones.
Since column vectors are simply n × 1-matrices, and row vectors are 1 × m-matrices, we can also
multiply a column vector by a row vector or vice versa.

Example 4.25: Column vector times row vector

 
1  
Multiply 2  1 2 1 0 .

1

Solution. Here we are multiplying a 3 × 1-matrix by a 1 × 4-matrix, so the result will be a 3 × 4-matrix.
Using the column method, we can compute this product as follows:
 
First column Second column Third column Fourth column
  z }| {
z }| { z  }| { z }| {
     
1     1   1   1   1   1 2 1 0
 2  1 2 1 0 =  2  1 ,  2  2 ,  2  1 ,  2  0   =  2 4 2 0 
 
1  1 1 1 1  1 2 1 0

Example 4.26: Row vector times column vector

 
  1
Multiply 1 2 3  2 .
−1

Solution. Here we are multiplying a 1 × 3-matrix by a 3 × 1-matrix, so the result will be a 1 × 1-matrix,
or in other words, a scalar. (We regard a scalar and a 1 × 1-matrix as the same thing). We have:
 
  1
1 2 3  2  = 1 · 1 + 2 · 2 + 3 · (−1) = 2.
−1
Therefore, multiplying a row vector by a column vector works very similarly to an ordinary dot product
(except that the dot product is defined between two column vectors, not a row vector and a column vector).
♠

Example 4.27: A multiplication that is not defined

Find BA, where  
1 2 0  
1 2 1
B= 0 3 1  and A = .
0 2 1
−2 1 1
 4.4. Matrix multiplication 135

Solution. The product BA is not defined, since B is a 3 × 3-matrix and A is a 2 × 3-matrix. Since the
number of columns of B does not match the number of rows of A, the product is not defined. ♠
Notice that the matrices in Example 4.27 are the same as those in Example 4.22. This demonstrates an
important property of matrix multiplication: it is possible that AB is defined by BA is undefined. Even if
AB and BA are both defined, they may not be equal, as the following example shows. Therefore, matrix
multiplication is not commutative.

Example 4.28: Matrix multiplication is not commutative

Compute AB and BA, where
   
1 2 0 1
A = and B =
3 4 1 0

Are they equal?

Solution. We have     
1 2 0 1 2 1
AB = =
3 4 1 0 4 3
and     
0 1 1 2 3 4
BA = = .
1 0 3 4 1 2
Therefore, AB and BA are not equal. Matrix multiplication is not commutative. ♠
We have seen two methods for matrix multiplication: one component at a time, and by the column method.
There is also a third method, called the row method. It is exactly symmetric to the column method.

Proposition 4.29: Matrix multiplication, row method

Let A be an m × n-matrix, and let B be an n × p-matrix. Suppose that the rows of A are a1 , a2 , . . . , am .
Then the rows of AB are
a1 B, a2 B, . . . , am B.
In other words, the i th column of the matrix product AB is equal to the i th column of A times B.

Example 4.30: Matrix multiplication by the row method

Find the matrix product AB by the row method, where
 
  1 2 0
1 2 1
A= and B =  0 3 1 .
0 2 1
−2 1 1

Solution. We multiply each of the rows of A by B:

 
  1 2 0        
1 2 1  0 3 1  = 1 1 2 0 + 2 0 3 1 + 1 −2 1 1 = −1 9 3 ,
−2 1 1
136 Matrices

 
 1 2 0        
0 2 1  0 3 1  = 0 1 2 0 + 2 0 3 1 + 1 −2 1 1 = −2 7 3 .
−2 1 1
The resulting two row vectors form the rows of AB. Thus,
 
−1 9 3
AB = .
−2 7 3

Once again this is the same answer as in Examples 4.24 and 4.22. All three methods give the same result.
But notice how in the row method, each row of A provides instructions for taking a linear combinations of
the rows of B. ♠

We finish this section by introducing an important square matrix called the identity matrix.

Definition 4.31: Identity matrix

The identity matrix of size n × n has ones along the diagonal, and zeros everywhere else. In other
words, it is the matrix δi j where δi j = 1 if i = j and δi j = 0 otherwise. The identity matrix is
always a square matrix. Here are some identity matrices of various sizes.
 
  1 0 0 0
  1 0 0
  1 0  0 1 0 0 
1 , ,  0 1 0 ,   0 0 1 0 .

0 1
0 0 1
0 0 0 1

When it is necessary to distinguish which size of identity matrix is being discussed, we will use the
notation In for the n × n identity matrix.

Example 4.32: Multiplying by the identity matrix

Calculate AI , where I is the 2 × 2 identity matrix and
 
a11 a12
A =  a21 a22  .
a31 a32

Solution. We need to calculate  

a11 a12  
  1 0
AI = a21 a22 .
0 1
a31 a32
Using the column method, we find that the first column of AI is
       
a11 a12   a11 a12 a11
 a21 a22  1 = 1  a21  + 0  a22  =  a21  ,
0
a31 a32 a31 a32 a31
 4.4. Matrix multiplication 137

which is exactly the same as the first column of A. Similarly, the second column of AI is
       
a11 a12   a11 a12 a12
 a21 a22  0 = 0  a21  + 1  a22  =  a22  ,
1
a31 a32 a31 a32 a32

which is exactly the same as the second column of A. Therefore

 
a11 a12
AI =  a21 a22  = A.
a31 a32

♠
The calculation of the last example generalizes to matrices of all sizes, and is summarized in the following
proposition.

Proposition 4.33: Multiplying by the identity matrix

Let A be any m × n-matrix. Then
Im A = A = AIn .

We can also raise a square matrix to a power. For example, A5 means A · A · A · A · A.

Example 4.34: Raising a matrix to a power

Compute A3 , where  
1 2
A= .
−2 3

Solution. We have
        
3 1 2 1 2 1 2 −3 8 1 2 −19 18
A = A·A·A = = = .
−2 3 −2 3 −2 3 −8 5 −2 3 −18 −1

4.4.3. Properties of matrix multiplication

We have already seen that matrix multiplication is not in general commutative, i.e., AB and BA may be
different, even if they are both defined. Sometimes it can happen that AB = BA for specific matrices A and
B. In this case, we say that A and B commute.
The following are some properties of matrix multiplication. Notice that these properties hold only
when the size of matrices are such that the products are defined.
138 Matrices

Proposition 4.35: Properties of matrix multiplication

The following properties hold for matrices A, B,C of appropriate dimensions and for scalars r.

• The associative law of multiplication

(AB)C = A(BC).

• The existence of multiplicative units

Im A = A = AIn ,

where A is an m × n-matrix.

• Compatibility with scalar multiplication

(rA)B = r(AB) = A(rB).

• The distributive laws of multiplication over addition

A(B +C) = AB + AC,

(B +C)A = BA +CA.

Proof. First, we will prove the associative law. In the proof, it will be useful to use summation notation.
We write
n
∑ xi = x1 + x2 + . . . + xn
i=1
for the sum of the n numbers x1 , . . . , xn . Assume A is an m × n-matrix, B is an n × p-matrix, and C is a
p × q-matrix. Then both (AB)C and A(BC) are m × q-matrices. We must show that they have the same
entries. The (i, ℓ)-entry of the matrix (AB)C is
p p n p n
((AB)C)iℓ = ∑ (AB)ik ckℓ = ∑ ( ∑ ai j b jk )ckℓ = ∑ ∑ ai j b jk ckℓ .
k=1 k=1 j=1 k=1 j=1

On the other hand, the (i, ℓ)-entry of the matrix A(BC) is

n n p n p
(A(BC))iℓ = ∑ ai j (BC) jℓ = ∑ ai j ( ∑ b jk ckℓ ) = ∑ ∑ ai j b jk ckℓ .
j=1 j=1 k=1 j=1 k=1

Both sums are equal, since they are both summing over all the terms where j = 1, . . . , n and k = 1, . . . , p.
Therefore, (AB)C = A(BC). The fact that identity matrices act as multiplicative units was already men-
tioned in Proposition 4.33. We leave compatibility with scalar multiplication as an exercise. To prove the
first distributive law, assume A is an m × n-matrix, and B and C are n × p-matrices. Then both A(B +C)
and AB + AC are m × p-matrices. We have
n n n n
(A(B +C))ik = ∑ ai j (B +C) jk = ∑ ai j (b jk + c jk ) = ∑ ai j b jk + ∑ ai j c jk = (AB + AC)ik.
j=1 j=1 j=1 j=1
 4.4. Matrix multiplication 139

Thus A(B +C) = AB + AC as claimed. The proof of the other distributive law is similar. ♠

Exercises
 
1 2 0
Exercise 4.4.1 Let A = . Multiply A by each of the following vectors.
−1 1 1
         
1 0 2 0 1
 
(a) 0 , (b)   
1 , (c) 1 , (d) 0 , (e) 0  .
   
0 −1 0 0 3

Exercise 4.4.2 Write the following system of equations in vector form and matrix form.
2x + 3y + z = 4,
x − 2z = 3,
2y + z = 1.

Exercise 4.4.3 Compute the following by columns and by rows. Convince yourself that both method give
the same result.   
a11 a12 a13 x1
 a21 a22 a23   x2  .
a31 a32 a33 x3

Exercise 4.4.4 Write the vector  

x1 − x2 + 2x3
 2x3 + x1 
 
 3x3 
3x4 + 3x2 + x1
 
x1
 x2 
in the form A  
 x3  where A is an appropriate matrix.
x4

Exercise 4.4.5 Write the vector  

x1 + 3x2 + 2x3
 2x3 + x1 
 
 6x3 
x4 + 3x2 + x1
 
x1
 x2 
in the form A  
 x3  where A is an appropriate matrix.
x4
140 Matrices

Exercise 4.4.6 Consider the matrices

         
1 2 3 3 −1 2 1 2 −1 2 2
A= , B= , C= , D= , E= .
2 1 7 −3 2 1 3 1 2 −3 3
Find the following if possible. If it is not possible explain why.
(a) −3A
(b) 3B − A
(c) AC
(d) CB
(e) AE
(f) EA

Exercise 4.4.7 Consider the matrices

 
1 2        
  2 −5 2 1 2 −1 1 1
A= 3 2 , B= , C= , D= , E=
−3 2 1 5 0 4 −3 3
1 −1
Find the following if possible. If it is not possible explain why.
(a) −3A
(b) 3B − A
(c) AC
(d) CA
(e) AE
(f) EA
(g) BE
(h) DE
 
−1 −1
Exercise 4.4.8 Let A = . Find all 2 × 2-matrices B such that AB = 0.
3 3
   
1 2 1 2
Exercise 4.4.9 Let A = and B = . Is it possible to find k such that AB = BA? If so,
3 4 3 k
what should k equal?
   
1 2 1 2
Exercise 4.4.10 Let A = and B = . Is it possible to choose k such that AB = BA? If so,
3 4 1 k
what should k equal?

Exercise 4.4.11 For each pair of matrices, find the (1, 2)-entry and (2, 3)-entry of the product AB.
 4.4. Matrix multiplication 141

   
1 2 −1 4 6 −2

(a) A = 3 4 
0 ,B=  7 2 1 
2 5 1 −1 0 0
   
1 3 1 2 3 0
(b) A =  0 2 4 , B =  −4 16 1 
1 0 5 0 2 2

Exercise 4.4.12 Compute A4 , where  

1 −3
A= .
2 −5
Hint: you can save some work by calculating A2 times A2 .

Exercise 4.4.13 Find 2 × 2-matrices A, B, and C such that A 6= 0, C 6= B, but AC = AB.

Exercise 4.4.14 Find 2 × 2-matrices A and B such that A 6= 0 and B 6= 0 but AB = 0.

Exercise 4.4.15 Find 3 × 3-matrices A and B such that AB 6= BA.

Exercise 4.4.16 Give an example of a matrix A such that A2 = I and yet A 6= I and A 6= −I.

Exercise 4.4.17 A matrix A is called idempotent if A2 = A. Let

 
2 0 2
A= 1 1 2 .
−1 0 −1

Show that A is idempotent.

Exercise 4.4.18 Suppose A and B are square matrices of the same size. Which of the following are
necessarily true?

(a) (A + B)2 = A2 + 2AB + B2 .

(b) (A − B)2 = A2 − 2AB + B2 .

(c) (AB)2 = A2 B2 .

(d) (A + B)2 = A2 + AB + BA + B2 .

(e) A2 B2 = A(AB)B.

(f) (A + B)3 = A3 + 3A2 B + 3AB2 + B3 .

(g) (A + B)(A − B) = A2 − B2 .
142 Matrices

4.5 Matrix inverses

Outcomes
A. Determine whether a matrix is invertible, and compute the inverse if it exists.

B. Solve a system of linear equations using matrix algebra.

C. Prove algebraic properties of matrix inverses.

D. Determine whether a matrix is a left inverse, right inverse, or inverse of another matrix.

4.5.1. Definition and uniqueness

We now define a matrix operation which in some ways plays the role of division. We cannot divide by a
matrix, but we can multiply by the inverse of a matrix, which is almost as good.

Definition 4.36: The inverse of a matrix

Let A and B be n × n-matrices. We say that B is an inverse of A if

BA = I and AB = I.

If this is the case, we also write B = A−1 . When a matrix has an inverse, it is called invertible.

Example 4.37: Verifying the inverse of a matrix

   
1 1 2 −1
Let A = . Check that B = is an inverse of A.
1 2 −1 1

Solution. To check this, multiply

    
1 1 2 −1 1 0
AB = = =I
1 2 −1 1 0 1

and     
2 −1 1 1 1 0
BA = = = I.
−1 1 1 2 0 1
This shows that B is indeed an inverse of A. ♠
Unlike multiplication of scalars, it can happen that A 6= 0 but A does not have an inverse. This is illustrated
in the following example.
 4.5. Matrix inverses 143

Example 4.38: A non-zero matrix with no inverse

 
1 1
Let A = . Show that A is not invertible.
1 1

Solution. One might think A has an inverse because it does not equal zero. However, note that
    
1 1 −1 0
= .
1 1 1 0
If an inverse A−1 existed, we would have the following:
   
−1 −1
= I
1 1
 
−1 −1
= (A A)
1
  
−1 −1
= A A
1
 
0
= A−1
0
 
0
= .
0
This says that    
−1 0
= ,
1 0
which is impossible! Therefore, A does not have an inverse. ♠
Can a matrix have more than one inverse? It turns out that this is not the case: the following theorem
shows that if A has an inverse, then the inverse is unique. We can therefore speak of “the” inverse, rather
than just “an” inverse, of A.

Theorem 4.39: Uniqueness of inverse

Suppose A is an n × n-matrix such that both B and C are inverses of A. Then B = C.

Proof. By assumption, both B and C are inverses of A, so we have AB = I, BA = I, AC = I, and CA = I.

Using the associative and unit properties of matrix multiplication, we have:
B = BI = B(AC) = (BA)C = IC = C.
Therefore, B = C, as desired. ♠

4.5.2. Computing inverses

In Example 4.37, we verified that a matrix A had an inverse. But we did not actually compute the inverse:
the inverse B was already given, and we merely checked that AB = I and BA = I. We now explore a method
for finding the inverse when it is not already known what it is.
144 Matrices

Example 4.40: Finding the inverse of a matrix

Find the inverse of the matrix  
1 −2
A= .
2 −3

 
x z
Solution. To find A−1 , we need to find a matrix such that
y w
    
1 −2 x z 1 0
= .
2 −3 y w 0 1
We can multiply these two matrices, and see that in order for this equation to be true, we must solve the
systems of equations
x − 2y = 1,
2x − 3y = 0,
and
z − 2w = 0,
2z − 3w = 1.
Writing the augmented matrix for these two systems gives
 
1 −2 1
2 −3 0
for the first system and  
1 −2 0
2 −3 1
for the second one. Note that both systems have A as their coefficient matrix. Since both systems have the
same coefficient matrix, they both require exactly the same row operations, and we can use the method
of Example 1.33 to solve both systems at the same time. To do so, we create a single augmented matrix
containing both of the right-hand sides:
 
1 −2 1 0
.
2 −3 0 1
Then we perform row operations until the coefficient matrix is in reduced echelon form:
     
1 −2 1 0 R2 ← R2 −2R1 1 −2 1 0 R1 ← R1 +2R2 1 0 −3 2
≃ ≃ . (4.1)
2 −3 0 1 0 1 −2 1 0 1 −2 1
This corresponds to the following reduced echelon forms for the two original systems of equations:
   
1 0 −3 1 0 2
and .
0 1 −2 0 1 1
The solution of the first system is x = −3 and y = −2. The solution for the second system is z = 2 and
w = 1. If we take the values found for x, y, z, and w and put them into our inverse matrix, we see that the
inverse is    
−1 x z −3 2
A = = .
y w −2 1
 4.5. Matrix inverses 145

Notice that this is exactly the right-hand side in the last augmented matrix of (4.1). In other words, all we
really had to do to find the inverses were the row operations in (4.1). The inverse can be read off directly
from the result. ♠
The example suggests a general method for finding the inverse of a matrix, which we summarize in the
following algorithm.

Algorithm 4.41: Finding the inverse of a matrix

Suppose A is an n × n-matrix. To find A−1 if it exists, form the augmented n × 2n-matrix

[A | I] .

If possible, do row operations until you obtain an n × 2n-matrix of the form

[I | B] .

If this can be done, then A is invertible and A−1 = B. If it is not possible (i.e., if the reduced echelon
form of A has less than n pivot entries), then A is not invertible.

This algorithm shows how to find the inverse if it exists. It also tells us if A does not have an inverse.

Example 4.42: Finding the inverse of a matrix

 
1 2 2
Let A =  1 0 2 . Find A−1 if it exists.
3 1 −1

Solution. We set up the augmented matrix and reduce it to reduced echelon form.
 
1 2 2 1 0 0
[A | I] =  1 0 2 0 1 0 
3 1 −1 0 0 1
R2 ← R2 −R1
 
1 2 2 1 0 0
R3 ← R3 −3R1
≃  0 −2 0 −1 1 0 
0 −5 −7 −3 0 1
R1 ← 7R1
 
R3 ← −2R3
7 14 14 7 0 0
≃  0 −2 0 −1 1 0 
0 10 14 6 0 −2
R1 ← R1 +7R2
 
7 0 14 0 7 0
R3 ← R3 +5R2
≃  0 −2 0 −1 1 0 
0 0 14 1 5 −2
 
7 0 0 −1 2 2
R1 ← R1 −R3
≃  0 −2 0 −1 1 0 
0 0 14 1 5 −2
146 Matrices

R1 ← 71 R1  
R2 ← − 12 R2 1 0 0 − 17 2 2
R3 ← 1  7 7 
14 R3  1 
≃  0 1 0 2 − 12 0 .
 
1 5
0 0 1 14 14 − 17

Notice that the last augmented matrix is of the form [I | B], where the left-hand side is the 3 × 3 identity
matrix. Therefore, the inverse is the 3 × 3-matrix on the right-hand side, given by
 
− 17 2 2
 7 7 
−1  1 
A =  2 − 12 0 .
 
1 5 1
14 14 − 7

♠
When looking for the inverse of a matrix, it can happen that the left-hand side cannot be row reduced to
the identity matrix. The following is an example of this situation.

Example 4.43: A non-invertible matrix

 
1 −2 2
Let A =  1 0 2 . Find A−1 if it exists.
2 −2 4

Solution. We write the augmented matrix

 
1 −2 2 1 0 0
[A | I] =  1 0 2 0 1 0 
2 −2 4 0 0 1
 
and proceed to do row operations attempting to obtain I | A−1 . After a few row operations, we have
 
1 −2 2 1 0 0
 0 2 0 −1 1 0 .
0 0 0 −1 −1 1

At this point, we see that the coefficient matrix has rank 2, i.e., there are only two pivot entries. This
means there is no way to obtain I on the left-hand side of this augmented matrix. Hence, there is no way
to complete the algorithm, and the inverse of A does not exist. ♠
If the algorithm provides an inverse, it is always possible to double-check that your answer is correct. To
do so, use the method demonstrated in Example 4.37. Check that the products AA−1 and A−1 A both equal
the identity matrix. Through this method, you can always ensure that you have calculated A−1 properly.
 4.5. Matrix inverses 147

4.5.3. Using the inverse to solve a system of equations

One way in which the inverse of a matrix is useful is to find the solution of a system of linear equations.
Recall from Definition 4.20 that we can write a system of equations in matrix form, which is in the form
Ax = b.
Suppose we find the inverse A−1 of the matrix A. Then we can multiply both sides of this equation by A−1
on the left and simplify to obtain
x = A−1 b.
Therefore we can find x, the solution to the system, by computing x = A−1 b. Note that once we have found
A−1 , we can easily get the solution for different right-hand sides (different b). It is always just A−1 b.

Example 4.44: Using the inverse to solve a system of equations

Consider the following system of equations. Use the inverse of a suitable matrix to solve this system.

x+z = 1
x−y+z = 3
x+y−z = 2

Solution. First, we can write the system in matrix form

    
1 0 1 x 1
Ax =  1 −1 1   y  =  3  = b.
1 1 −1 z 2
The inverse of A is  
1 1
0 2 2
 
A−1 =  1 −1 0 .
1 − 12 − 12
From here, the solution to the system Ax = b is found by x = A−1 b, i.e.,
      
1 1 5
x 0 2 2 1 2
      
 y  =  1 −1 0   3  =  −2  .
z 1 − 12 − 12 2 − 32
♠
 
0
What if the right-hand side had been b = 1 ? In this case, the solution would be given by

3
      
1 1
x 0 2 2 0 2
      
  
y = 1 −1 0   1  =  −1  .
z 1 − 12 − 12 3 −2
This illustrates that for a system Ax = b where A−1 exists, it is easy to find the solution when the vector b
is changed.
148 Matrices

4.5.4. Properties of the inverse

The following are some algebraic properties of matrix inverses.

Proposition 4.45: Properties of the inverse

Let A and B be n × n-matrices, I the n × n-identity matrix. Then the following hold.

1. I is invertible and I −1 = I .

2. If A and B are invertible then AB is invertible, and (AB)−1 = B−1 A−1 .

3. If A is invertible then so is A−1 , and (A−1 )−1 = A.

4. If A is invertible then so is Ak , and (Ak )−1 = (A−1 )k .

5. If A is invertible and p is a non-zero scalar, then pA is invertible and (pA)−1 = 1p A−1 .

4.5.5. Right and left inverses

So far, we have only talked about the inverses of square matrices. But what about matrices that are not
square? Can they be invertible? It turns out that non-square matrices can never be invertible. However,
they can have left inverses or right inverses.

Definition 4.46: Left and right inverses

Let A be an m × n-matrix and B an n × m-matrix. We say that B is a left inverse of A if

BA = I.

We say that B is a right inverse of A if

AB = I.
If A has a left inverse, we also say that A is left invertible. Similarly, if A has a right inverse, we
say that A is right invertible.

Example 4.47: Right inverse

Let  
  1 0
1 0 0
A= and B =  0 1 .
0 1 0
0 0
Show that B is a right inverse, but not a left inverse, of A.
 4.5. Matrix inverses 149

Solution. We compute
 
 1 0   
1 0 0  1 0
AB = 0 1  = = I,
0 1 0 0 1
0 0
   
1 0   1 0 0
1 0 0
BA =  0 1  =  0 1 0  6= I.
0 1 0
0 0 0 0 0
Therefore, B is a right inverse, but not a left inverse, of A. ♠
Recall from Definition 4.36 that B is called an inverse of A if it is both a left inverse and a right inverse. A
crucial fact is that invertible matrices are always square.

Theorem 4.48: Invertible matrices are square

Let A be an m × n-matrix.

• If A is left invertible, then m ≥ n.

• If A is right invertible, then m ≤ n.

• If A is invertible, then m = n.

In particular, only square matrices can be invertible.

Proof. To prove the first claim, assume that A is left invertible, i.e., assume that BA = I for some n × m-
matrix B. We must show that m ≥ n. Assume, for the sake of obtaining a contradiction, that this is not the
case, i.e., that m < n. Then the matrix A has more columns than rows. It follows that the homogeneous
system of equations Ax = 0 has a non-trivial solution; let x be such a solution. We obtain a contradiction
by a similar method as in Example 4.38. Namely, we have

x = I x = (BA)x = B(Ax) = B0 = 0,

contradicting the fact that x was non-trivial. Since we got a contradiction from the assumption that m < n,
it follows that m ≥ n.
The second claim is proved similarly, but exchanging the roles of A and B. The third claim follows
directly from the first two claims, because every invertible matrix is both left and right invertible. ♠
Of course, not all square matrices are invertible. In particular, zero matrices are not invertible, along with
many other square matrices.
150 Matrices

Exercises

Exercise 4.5.1 For each of the following pairs of matrices, determine whether B is an inverse of A.

(a)    
2 4 1 3 −4
A= , B= .
1 3 2 −1 2

(b)    
1 −2 −1 2
A= , B= .
4 −7 −4 7

(c)    
4 1 3 1 −1 −1
A =  2 1 2 , B= 0 1 −2  .
1 0 1 −1 1 2

Exercise 4.5.2 Suppose AB = AC and A is an invertible n × n-matrix. Does it follow that B = C? Explain
why or why not.

Exercise 4.5.3 Suppose AB = AC and A is a non-invertible n × n-matrix. Does it follow that B = C?

Explain why or why not.

Exercise 4.5.4 For each of the following matrices, find the inverse if possible. If the inverse does not exist,
explain why.
         
2 1 0 1 2 1 2 1 0 1 2
A= , B= , C= , D= , E= .
−1 3 5 3 3 0 4 2 1 2 5

Exercise 4.5.5 For each of the following matrices, find the inverse if possible. If the inverse does not exist,
explain why.
 
      1 2 0 2
1 2 3 1 0 3 1 2 3  1 1 2 0 
A =  2 1 4 , B =  2 3 4 , C =  2 1 4 , D =   2 1 −3 2  .

1 0 2 1 0 2 4 5 10
1 2 1 2

 
a b
Exercise 4.5.6 Let A be a 2 × 2 invertible matrix, with A = . Find a formula for A−1 in terms of
c d
a, b, c, d.

Exercise 4.5.7 Using the inverse of the matrix, find the solution to the systems:
 4.5. Matrix inverses 151

(a)     
2 4 x 1
= ,
1 1 y 2

(b)     
2 4 x 2
= .
1 1 y 0

Now give the solution in terms of a and b to

    
2 4 x a
= .
1 1 y b

Exercise 4.5.8 Using the inverse of the matrix, find the solution to the systems:

(a)     
1 0 3 x 1
 2 3 4  y  =  0 ,
1 0 2 z 1

(b)     
1 0 3 x 3
 2 3 4   y  =  −1  .
1 0 2 z −2

Now give the solution in terms of a, b, and c to the following:

    
1 0 3 x a
 2 3 4  y  =  b .
1 0 2 z c

Exercise 4.5.9 Show that if A is an n × n invertible matrix and X and B are n × 1-matrices such that
AX = B, then X = A−1 B.

Exercise 4.5.10 Prove that if A−1 exists and AX = 0 then X = 0.

Exercise 4.5.11 Show (AB)−1 = B−1 A−1 by verifying that

AB(B−1 A−1 ) = I and B−1 A−1 (AB) = I.

Exercise 4.5.12 Is it possible to have matrices A and B such that AB = I, while BA = 0? If it is possible,
give an example of such matrices. If it is not possible, explain why.

Exercise 4.5.13 Show that (ABC)−1 = C−1 B−1 A−1 by verifying that

(ABC)(C−1B−1 A−1 ) = I and (C−1 B−1 A−1 )(ABC) = I.

152 Matrices

Exercise 4.5.14 If A is invertible, show that A2 is invertible and (A2 )−1 = (A−1 )2 .

Exercise 4.5.15 If A is invertible, show (A−1 )−1 = A. Hint: Use the uniqueness of inverses.

Exercise 4.5.16 Determine whether B is a right inverse, left inverse, both, or neither of A.
(a)  
  1 −1
4 2 1
A= , B =  −1 1 .
2 1 1
−1 2
(b)  
1 0  
1 2 −1
A =  0 1 , B= .
0 1 0
0 2
(c)    
1 2 7 −2
A= , B= .
3 7 −3 1
(d)    
1 2 −1 2
A= , B= .
3 1 1 −1

Exercise 4.5.17 Show that right inverses are not unique by giving an example of matrices A, B,C such
that both B and C are right inverses of A, but B 6= C.

Exercise 4.5.18 Solve the following system of equations by using the inverse of a suitable matrix.
8x + 2y + 3z = −1
y − 2z = 2
x+z = 1

Exercise 4.5.19 Suppose that A, B,C, D are n × n-matrices, and that all relevant matrices are invertible.
Further, suppose that (A + B)−1 = CB−1 . Solve this equation for A (in terms of B and C), B (in terms of A
and C), and C (in terms of A and B).

Exercise 4.5.20 Which of the following matrices is right invertible? Find a right inverse if one exists. If
possible, find two different right inverses.
 
    1 2  
1 2 3 1 1 2 1 2
A= B= C= 1 0  D=
0 1 0 2 2 4 2 3
0 1

Exercise 4.5.21 Which of the following matrices is left invertible? Find a left inverse if one exists. If
possible, find two different left inverses.
   
    1 0 1 0
1 2 1 1 1
A= B= C= 0 1  D= 2 0 
0 1 0 2 3
2 1 3 0
 4.6. Elementary matrices 153

4.6 Elementary matrices

Outcomes
A. Use multiplication by elementary matrices to apply row operations.

B. Find the elementary matrix corresponding to a particular row operation.

C. Write the reduced echelon form of a matrix A in the form R = UA, where U is invertible.

D. Write a matrix as a product of elementary matrices.

4.6.1. Elementary matrices and row operations

Recall from Definition 1.14 that there are three kinds of elementary row operations on matrices:

1. Switch two rows.

2. Multiply a row by a non-zero number.

3. Add a multiple of one row to another row.

The purpose of this section is to show that each of these row operations corresponds to a special type of
invertible matrix called an elementary matrix.

Example 4.49: Elementary matrix for switching two rows

Let  
1 0 0
E =  0 0 1 .
0 1 0
What is the effect of multiplying E by an arbitrary 3 × n-matrix A?

Solution. Consider an arbitrary 3 × n-matrix

 
a11 a12 · · · a1n
A =  a21 a22 · · · a2n  .
a31 a32 · · · a3n

We compute the product EA by the row method:

    
1 0 0 a11 a12 · · · a1n a11 a12 · · · a1n
EA =  0 0 1   a21 a22 · · · a2n  =  a31 a32 · · · a3n  .
0 1 0 a31 a32 · · · a3n a21 a22 · · · a2n
154 Matrices

So the effect of multiplying A by E on the left is exactly the same as switching rows 2 and 3. We say that
E is the elementary matrix for switching rows 2 and 3. ♠

Example 4.50: Elementary matrix for multiplying a row by a non-zero number

Let  
1 0 0
E =  0 k 0 .
0 0 1
What is the effect of multiplying E by an arbitrary 3 × n-matrix A?

Solution. We compute the product EA by the row method:

    
1 0 0 a11 a12 · · · a1n a11 a12 · · · a1n
EA =  0 k 0   a21 a22 · · · a2n  =  ka21 ka22 · · · ka2n  .
0 0 1 a31 a32 · · · a3n a31 a32 · · · a3n
So the effect of multiplying A by E on the left is exactly the same as multiplying row 2 by the scalar k. We
say that E is the elementary matrix for multiplying row 2 by k. ♠

Example 4.51: Elementary matrix for adding a multiple of one row to another row
Let  
1 0 0
E =  0 1 0 .
0 k 1
What is the effect of multiplying E by an arbitrary 3 × n-matrix A?

Solution. Once again we compute the product EA:

    
1 0 0 a11 a12 · · · a1n a11 a12 ··· a1n
EA =  0 1 0   a21 a22 · · · a2n  =  a21 a22 ··· a2n .
0 k 1 a31 a32 · · · a3n a31 + ka21 a32 + ka22 · · · a3n + ka2n
So the effect of multiplying A by E on the left is exactly the same as adding k times row 2 to row 3. We
say that E is the elementary matrix for adding k times row 2 to row 3. ♠
As these examples show, performing each type of elementary row operation is the same as multiplying (on
the left) by a certain invertible matrix. These matrices are called the elementary matrices. In the above
examples, we have only considered 3 × 3-elementary matrices, but they exist for other sizes too. The fol-
lowing definition makes this precise. It also shows how to calculate the elementary matrix corresponding
to any elementary row operation.

Definition 4.52: Elementary matrices and row operations

Let E be an n × n-matrix. Then E is an elementary matrix if it is the result of applying one
elementary row operation to the n × n identity matrix.
 4.6. Elementary matrices 155

Example 4.53: Finding an elementary matrix

Consider the elementary row operation of adding 5 times row 3 to row 1 of a 4 × n-matrix. Find the
elementary matrix E corresponding to this row operation.

Solution. Following Definition 4.52, all we have to do is apply the desired row operation to the 4 × 4-
identity matrix:    
1 0 0 0 1 0 5 0
 0 1 0 0  R1 ← R1 +5R3  0 1 0 0 
  ≃   = E.
 0 0 1 0   0 0 1 0 
0 0 0 1 0 0 0 1
♠
We can double-check that multiplying E by any 4 × n-matrix does indeed have the desired effect:
    
1 0 5 0 a11 a12 · · · a1n a11 + 5a31 a12 + 5a32 · · · a1n + 5a3n
 0 1 0 0   a21 a22 · · · a2n   a21 a22 ··· a2n 
   =  .
 0 0 1 0   a31 a32 · · · a3n   a31 a32 ··· a3n 
0 0 0 1 a41 a42 · · · a4n a41 a42 ··· a4n
The fact that this always works is the content of the following theorem.

Theorem 4.54: Multiplication by an elementary matrix and row operations

Performing any of the three elementary row operations on a matrix A is the same as taking the
product EA, where E is the elementary matrix obtained by applying the desired row operation to
the identity matrix.

4.6.2. Inverses of elementary matrices

Suppose we have applied a row operation to a matrix A. Consider the row operation required to return A to
its original form, i.e., to undo the row operation. It turns out that this action is described by the inverse of
an elementary matrix. The following theorem ensures that the inverse of each elementary matrix is itself
an elementary matrix.

Theorem 4.55: Inverses of elementary matrices

Every elementary matrix is invertible and its inverse is also an elementary matrix.

In fact, the inverse of an elementary matrix is constructed by doing the reverse row operation on I. E −1 is
obtained by performing the row operation which would carry E back to I.

• If E is obtained by switching rows i and j, then E −1 is also obtained by switching rows i and j.
• If E is obtained by multiplying row i by the scalar k, then E −1 is obtained by multiplying row i by
the scalar 1k .
156 Matrices

• If E is obtained by adding k times row i to row j, then E −1 is obtained by subtracting k times row i
from row j.

Example 4.56: Inverse of an elementary matrix

Find E −1 , where E is the elementary matrix
 
1 0
E=
0 2

Solution. E is obtained from the 2 × 2 identity matrix by multiplying the second row by 2. In order to
carry E back to the identity, we need to multiply the second row of E by 12 . Hence, E −1 is given by
 
−1 1 0
E = .
0 12
♠

4.6.3. Elementary matrices and reduced echelon forms

Suppose an m × n-matrix A is row reduced to its reduced echelon form. By tracking each row operation
completed, this row reduction can be performed through multiplication by elementary matrices. The
following theorem uses this fact.

Theorem 4.57: The form R = UA

Let A be any m × n-matrix and let R be its reduced echelon form. Then there exists an invertible
m × m-matrix U such that
R = UA.
Specifically, U can be computed as the product (from right to left) of the elementary matrices of all
row operations used to convert A to reduced echelon form.

Example 4.58: The form R = UA

Let  
0 1
A =  1 0 .
2 0
Find the reduced echelon form of A and write it in the form R = UA, where U is invertible.

Solution. To find the reduced echelon form R, we row reduce A. For each step, we will record the appro-
priate elementary matrix. First, switch rows 1 and 2.
   
0 1 1 0
↔ R2
 1 0  R1 ≃  0 1 .
2 0 2 0
 4.6. Elementary matrices 157

 
0 1 0
The corresponding elementary matrix is E1 =  1 0 0 , i.e.,
0 0 1
    
0 1 0 0 1 1 0
 1 0 0  1 0  =  0 1 .
0 0 1 2 0 2 0

Next, subtract 2 times the first row from the third row.
   
1 0 1 0
R3 ← R3 −2R1
 0 1  ≃  0 1 .
2 0 0 0
 
1 0 0
The corresponding elementary matrix is E2 =  0 1 0 , i.e.,
−2 0 1
    
1 0 0 1 0 1 0
 0 1 0  0 1  =  0 1 .
−2 0 1 2 0 0 0

Notice that the resulting matrix is R, the required reduced echelon form of A. We can then write

R = E2 E1 A
= UA.

It remains to compute U :
    
1 0 0 0 1 0 0 1 0
U = E2 E1 =  0 1 0   1 0 0  =  1 0 0 .
−2 0 1 0 0 1 0 −2 1

We can verify that R = UA holds for this matrix U :

    
0 1 0 0 1 1 0
UA =  1 0 0   1 0  =  0 1  = R.
0 −2 1 2 0 0 0

♠
While the process used in the above example is reliable and simple when only a few row operations are
used, it becomes cumbersome in a case where many row operations are needed to carry A to R. The
following theorem provides an alternate way to find the matrix U .

Theorem 4.59: Finding the matrix U

Let A be an m × n-matrix and let R be its reduced echelon form. Then R = UA, where U is an
invertible m × m-matrix found by forming the augmented matrix [A | I] and row reducing to [R | U ].
158 Matrices

Let’s revisit the above example using the process outlined in Theorem 4.59.

Example 4.60: The form R = UA, revisited

 
0 1
Let A =  1 0 . Use the process of Theorem 4.59 to find U such that R = UA.
2 0

Solution. First, we set up the augmented matrix [A | I]:

 
0 1 1 0 0
 1 0 0 1 0 .
2 0 0 0 1
Now, we row reduce until the left-hand side is in reduced echelon form:
   
0 1 1 0 0 1 0 0 1 0
R1 ↔ R2
 1 0 0 1 0  ≃  0 1 1 0 0 
2 0 0 0 1 2 0 0 0 1
 
1 0 0 1 0
R3 ← R3 −2R1
≃  0 1 1 0 0 .
0 0 0 −2 1
The left-hand side of this augmented matrix is R, and the right-hand side is U . Comparing this to the
matrices R and U we found in Example 4.58, we see that the same matrices are obtained regardless of
which process is used. ♠

4.6.4. Writing an invertible matrix as a product of elementary matrices

Recall from Algorithm 4.41 that an n × n-matrix A is invertible if and only if A can be carried to the
n × n identity matrix using elementary row operations. Combining this with our discussion of elementary
matrices we see that A is invertible if and only if it can be written as a product of elementary matrices.
This is the content of the following theorem.

Theorem 4.61: Product of elementary matrices

Let A be an n × n-matrix. Then A is invertible if and only if it can be written as a product of
elementary matrices.

Proof. If A is an invertible n × n-matrix, then its reduced echelon form is the n × n identity matrix I. By
Theorem 4.57, we can write I = UA, where U = Ek · · · E2 E1 is a product of elementary matrices. Then

A = U −1 = E1−1 E2−1 · · · Ek−1 .

By Theorem 4.55, if Ei is an elementary matrix, then so is Ei−1 . Therefore, A has been written as a product
of elementary matrices. Conversely, if A can be written as a product of elementary matrices, then A is
clearly invertible, because each elementary matrix is invertible. ♠
 4.6. Elementary matrices 159

Example 4.62: Product of elementary matrices

 
0 1 0
Let A =  1 1 0 . Write A as a product of elementary matrices.
0 −2 1

Solution. Following the process of Theorem 4.61, we first row-reduce A to its reduced echelon form,
recording each row operation as an elementary matrix.
     
0 1 0 1 1 0 0 1 0
R1 ↔ R2
 1 1 0  ≃  0 1 0  with elementary matrix E1 =  1 0 0  ,
0 −2 1 0 −2 1 0 0 1
     
1 1 0 1 0 0 1 −1 0
R1 ← R1 −R2
 0 1 0  ≃  0 1 0  with elementary matrix E2 =  0 1 0 ,
0 −2 1 0 −2 1 0 0 1
     
1 0 0 R3 ← R3 +2R2
1 0 0 1 0 0
 0 1 0  ≃  0 1 0  with elementary matrix E3 =  0 1 0  .
0 −2 1 0 0 1 0 2 1
Notice that the reduced echelon form of A is I. Hence I = UA where U = E3 E2 E1 . It follows that
A = U −1 = E1−1 E2−1 E3−1 , and so we have succeeded in writing A as a product of elementary matrices
   
0 1 0 1 1 0 1 0 0
A = E1−1 E2−1 E3−1 =  1 0 0   0 1 0   0 1 0 .
0 0 1 0 0 1 0 −2 1

In particular, it follows that A is invertible. ♠

4.6.5. More properties of inverses

In this section, we will use elementary matrices to prove a useful theorem about the inverse of a square
matrix. We start with an observation about the echelon form of a right invertible matrix.

Lemma 4.63: Echelon form of a right invertible matrix

Suppose that A is right invertible. Then the reduced echelon form of A does not have a row of zeros.

Proof. Let R be the reduced echelon form of A. Then by Theorem 4.57, we can write R = UA for some
invertible square matrix U . By assumption, we have AB = I, and therefore

R(BU −1 ) = (UA)(BU −1) = U (AB)U −1 = U IU −1 = UU −1 = I.

If R had a row of zeros, then so would the product R(BU −1). But since the identity matrix I does not have
a row of zeros, neither does R. ♠
160 Matrices

Theorem 4.64: Right invertible square matrices are invertible

Suppose A and B are square matrices such that AB = I . Then it follows that BA = I , and therefore
B = A−1 . In particular, a square matrix is right invertible if and only if it is left invertible if and only
if it is invertible.

Proof. Assume A and B are square matrices such that AB = I. Let R be the reduced echelon form of A.
Then by Theorem 4.57, we can write R = UA where U is an invertible matrix. Since AB = I, we know by
Lemma 4.63 that R does not have a row of zeros. Since R is a square reduced echelon form with no row
of zeros, each column must be a pivot column, and it follows that R = I. Hence, UA = I, and therefore A
is left invertible. Moreover, we have

B = IB = (UA)B = U (AB) = U I = U ,

and therefore B = U . It follows that BA = UA = I, as claimed.

To prove the last claim, note that we just proved that for square matrices, AB = I implies BA = I.
Therefore, every right inverse of A is also a left inverse, and therefore an inverse. But of course, we also
have that BA = I implies AB = I. This is just the same theorem, with the roles of A and B interchanged.
Therefore, every left inverse of A is also a right inverse. ♠
This theorem is very useful, because it allows us to test only one of the products AB or BA in order to
check that B is the inverse of A, saving us half of the work. It is important to stress, however, that this
only works for square matrices. As we saw in Example 4.47, non-square matrices can be right invertible
without being left invertible, or vice versa.

Exercises

Exercise 4.6.1 For each of the following pairs of matrices, suppose a row operation is applied to A and
the result is B. Find the elementary matrix E that represents this row operation.

(a)    
2 3 1 2
A= , B= .
1 2 2 3

(b)    
4 0 8 0
A= , B= .
2 1 2 1

(c)    
1 −3 1 −3
A= , B= .
0 5 2 −1

Exercise 4.6.2 For each of the following pairs of matrices, suppose a row operation is applied to A and
the result is B.
 4.6. Elementary matrices 161

• Find the elementary matrix E such that EA = B.

• Find the inverse of E, E −1 , such that E −1 B = A.

(a)    
1 2 1 1 2 1
A= 0 5 1 , B =  2 −1 4  .
2 −1 4 0 5 1

(b)    
1 2 1 1 2 1
A= 0 5 1 , B =  0 10 2  .
2 −1 4 2 −1 4

(c)    
1 2 1 1 2 1
A= 0 5 1 , B= 0 5 1 .
2 −1 4 1 − 12 2

(d)    
1 2 1 1 2 1
A= 0 5 1 , B= 2 4 5 .
2 −1 4 2 −1 4

Exercise 4.6.3 Find the reduced echelon form of each of the following matrices A, and write it in the form
R = UA where U is invertible.
     
1 3 1 0 1   1 1 0
0 1 −2
(a)  0 2 4  , (b)  3 −1  , (c) , (d)  0 1 1  .
1 3 2
2 6 −2 2 6 1 0 1

Exercise 4.6.4 Write each of the following matrices as a product of elementary matrices, if possible, or
else say why it is not possible.
   
    1 2 0 0 1 2
1 3 2 1
(a) , (b) , (c)  0 1 3  , (d)  1 3 7  .
0 2 5 3
2 4 1 1 1 3
162 Matrices

4.7 The transpose

Outcomes
A. Calculate the transpose of a matrix.

B. Determine whether a matrix is symmetric, antisymmetric, or neither.

C. Manipulate algebraic expressions involving the transpose of matrices.

Another important operation on matrices is that of taking the transpose. The transpose of a matrix is
obtained by turning the rows into columns and vice versa.

Definition 4.65: The transpose of a matrix

Let A be an m × n-matrix. Then the transpose of A, denoted AT , is the n × m-matrix whose (i, j)-
entry is the ( j, i)-entry of A.
 
a11 · · · a1n  
 a21 · · · a2n  a11 a21 · · · am1
  T  . ... . . . ... 
A =  .. . . ..  , A =  .. .
 . . . 
a1n a2n · · · amn
am1 · · · amn

Example 4.66: The transpose of a matrix

Find the transpose of the following matrix:
 
1 2 6
A= .
3 5 4

Solution. The transpose is  

1 3
AT =  2 5  .
6 4
Notice that A is a 2 × 3-matrix, while AT is a 3 × 2-matrix. ♠
 T
We have already used a special case of the transpose since Chapter 2, when we wrote 1 2 3 as a
space-saving notation for the column vector  
1
 2 .
3
The transpose of a matrix satisfies the following properties:
 4.7. The transpose 163

Proposition 4.67: Properties of the transpose

Let A and B be matrices of appropriate sizes, and r a scalar. Then the following hold.

1. (AT )T = A.

2. (A + B)T = AT + BT .

3. (rA)T = rAT .

4. (AB)T = BT AT .

5. 0T = 0.

6. I T = I .

7. (A−1 )T = (AT )−1 , if A is invertible.

Recall that a column vector is the same thing as a n × 1-matrix. Using the transpose, we can make precise
the connection between the dot product and the matrix product. Namely, let
   
v1 w1
   
v =  ...  and w =  ... 
vn wn
by column vectors. Then
 
w1
  
·
v w = v1 w1 + . . . + vn wn = v1 · · · vn  ...  = vT w.
wn

In other words, the dot product of column vectors v and w is the same thing as the matrix product vT w.
We can also use the notion of transpose to define what it means for a matrix to be symmetric and
antisymmetric.

Definition 4.68: Symmetric and antisymmetric matrices

An n × n-matrix A is said to be symmetric if AT = A. It is said to be antisymmetric (sometimes
also called skew symmetric) if AT = −A.

Example 4.69: Symmetric and antisymmetric matrices

Let      
2 1 3 0 1 3 0 1 3
A= 1 5 −3  , B =  −1 0 2 , and C =  1 5 −3  .
3 −3 7 −3 −2 0 1 3 0
Then A is symmetric because AT = A, B is antisymmetric because BT = −B, and C is neither
symmetric nor antisymmetric because CT is equal to neither C nor −C.
164 Matrices

Exercises
   
Exercise 4.7.1 Let X = −1 −1 1 and Y = 0 1 2 . Find X T Y and XY T if possible.

Exercise 4.7.2 Consider the matrices

 
1 2        
  2 −5 2 1 2 −1 1 1
A= 3 2 , B= , C= , D= , E= .
−3 2 1 5 0 4 −3 3
1 −1
Find the following if possible. If it is not possible explain why.

(a) −3AT .
(b) 3B − AT .
(c) E T B.
(d) EE T .
(e) BT B.
(f) CAT .
(g) DT BE.

Exercise 4.7.3 Which of the following matrices are symmetric, antisymmetric, both, or neither?
       
0 1 2 1 1 2 0 0
A= , B= , C= , D= .
−1 0 1 3 −2 0 0 0

Exercise 4.7.4 Suppose A is a matrix that is both symmetric and antisymmetric. Show that A = 0.

Exercise 4.7.5 Let A be an n × n-matrix. Show A equals the sum of a symmetric and an antisymmetric
matrix. Hint: Show that 12 (AT + A) is symmetric and then consider using this as one of the matrices.

Exercise 4.7.6 Show that the main diagonal of every antisymmetric matrix consists of only zeros. Recall
that the main diagonal consists of every entry of the matrix which is of the form aii .

Exercise 4.7.7 Show that for m × n-matrices A, B and scalars r, s, the following holds:
(rA + sB)T = rAT + sBT .

Exercise 4.7.8 Let A be a real m × n-matrix and let u ∈ Rn and v ∈ Rm . Show (Au) v = u (AT v). · ·
Exercise 4.7.9 Show that if A is an invertible n × n-matrix, then so is AT and (AT )−1 = (A−1 )T .

Exercise 4.7.10 Suppose A is invertible and symmetric. Show that A−1 is symmetric.
 4.8. Matrix arithmetic modulo p 165

4.8 Matrix arithmetic modulo p

Outcomes
A. Perform matrix operations over the field Z p .

In Section 1.8, you learned that most of linear algebra can be done over scalars from any field K, and not
just the real numbers. You also learned that Z p , the set of integers modulo p, is a field whenever p is a
prime number.
Indeed, all of the operations on matrices that we covered in this chapter make sense over any field:
addition, scalar multiplication, matrix multiplication, inverses, elementary matrices, and the transpose.

Example 4.70: A matrix product over Z5

Compute the matrix product AB over the field Z5 , where
 
  3 1
1 0 4
A= and B =  4 0 .
2 3 1
2 2

Solution. For example, the (1, 1)-entry of AB is calculated by multiplying the first row of A by the first
column of B, i.e.,
c11 = 1 · 3 + 0 · 4 + 4 · 2 = 3 + 0 + 3 = 1,
keeping in mind that all arithmetic operations are done in Z5 . We repeat the same for the other entries and
obtain  
  3 1  
1 0 4  1 4
AB = 4 0  = .
2 3 1 0 4
2 2
♠

Example 4.71: An inverse over Z7

Compute the inverse of the matrix  
1 4 2
A= 0 6 2 
5 0 3
with scalars in the field Z7 .

Solution. We use exactly the method of Algorithm 4.41, i.e., we set up the augmented matrix [A | I] and
reduce it to reduced echelon form. The only thing we have to keep in mind is that all operations are done
166 Matrices

modulo 7. Also, as usual, instead of dividing by a scalar, we must multiply by its inverse.
 
1 4 2 1 0 0
[A | I] =  0 6 2 0 1 0 
5 0 3 0 0 1
 
1 4 2 1 0 0
R3 ← R3 +2R1
≃  0 6 2 0 1 0 
0 1 0 2 0 1
 
1 4 2 1 0 0
R2 ↔ R3
≃  0 1 0 2 0 1 
0 6 2 0 1 0
 
R3 ← R3 +R2
1 4 2 1 0 0
≃  0 1 0 2 0 1 
0 0 2 2 1 1
 
−1 1 4 2 1 0 0
R3 ← 2 R3 = 4R3
≃  0 1 0 2 0 1 
0 0 1 1 4 4
 
1 0 2 0 0 3
R1 ← R1 −4R2
≃  0 1 0 2 0 1 
0 0 1 1 4 4
 
R1 ← R1 −2R3
1 0 0 5 6 2
≃  0 1 0 2 0 1 .
0 0 1 1 4 4
Therefore, the inverse is  
5 6 2
A−1 =  2 0 1  .
1 4 4
As usual, we can double-check that we didn’t make any mistakes by calculating
    
1 4 2 5 6 2 1 0 0
AA−1 =  0 6 2   2 0 1  =  0 1 0  = I.
5 0 3 1 4 4 0 0 1
So indeed, we have calculated the inverse correctly. ♠

Exercises

Exercise 4.8.1 Let  

    2 1 0
1 3 1 0 2
A= , B= , C= 1 0 2 
5 2 0 2 1
0 3 1
and compute the following over Z11 :
 4.9. Application: Cryptography 167

(a) 3A,

(b) A2 ,

(c) AB,

(d) BC,

(e) C−1 .

4.9 Application: Cryptography

Cryptography is about encoding a message so that it is hard for a third party to read. The original message
is called the plaintext and the encrypted message is called the ciphertext. The process of turning a
plaintext into the corresponding ciphertext is called encryption, and the process of turning a ciphertext
into the corresponding plaintext is called decryption. An encryption and decryption method is also called
a cipher. Modern ciphers are designed in such a way that the cipher itself is not secret, but the encryption
depends on a secret key. A cipher should be designed so that decryption is easy for a person who knows
the key, but difficult for everybody else. The art of designing ciphers is called cryptography, and the art
of breaking ciphers is called cryptanalysis.
In order to be able to define ciphers using algebraic operations, we start by encoding strings as se-
quences of numbers. To that end, we assign a number to each letter of the alphabet, as well as the special
symbols “space”, “comma”, and “period”, according to the following scheme.

Space A B C D ... Z Comma Period

0 1 2 3 4 ... 26 27 28

In practical applications, one would probably use a larger set of symbols and a standard encoding such as
ASCII or UTF-8. But the above 29 symbols will be sufficient for our purposes. It will also come in handy
that 29 is prime.

Example 4.72: Representing strings as sequences of numbers

Convert the string “Attack at dawn” to a sequence of numbers. Convert the sequence of numbers
9, 0, 12, 9, 11, 5, 0, 3, 15, 4, 5, 19, 28 to a string.

Solution. We have A = 1, T = 20, T = 20, A = 1, C = 3, K = 11, Space = 0, and so on. Continuing in

this way, the encoding of “Attack at dawn” is 1, 20, 20, 1, 3, 11, 0, 1, 20, 0, 4, 1, 23, 14. Conversely, we have
9 = I, 0 = Space, 12 = L, 9 = I, 11 = K, 5 = E, and so on. We find that the decoded string is “I like codes.”
♠
There are many different ways to define ciphers. Some of the oldest known ciphers date back thousands of
years. An example of such a “classic” cipher is a substitution cipher, where each letter of the alphabet is
replaced by a different letter, for example A 7→ D, B 7→ E, and so on. Substitution ciphers have the property
168 Matrices

that changing one letter of the plaintext always changes exactly one letter of the ciphertext. This is not a
desirable property, because it makes the cipher easy to break. Therefore, modern ciphers are designed to
satisfy a property called diffusion: changing one letter of the plaintext should change many letters of the
ciphertext.
In a block cipher, the plaintext is first divided into blocks of equal size, and then each block is en-
crypted separately. The block size is the number of plaintext symbols in each block. If the length of the
plaintext is not divisible by the block size, we pad the final block with additional spaces. In the context
of a block cipher, the diffusion property means that changing one symbol of a plaintext block potentially
affects every symbol of the ciphertext block. The following is an example of a block cipher.

Definition 4.73: Hill cipher

The Hill cipher of block size n has as its key an invertible n × n-matrix A with scalars from Z29 .
Each ciphertext block c1 , . . . , cn is computed from the corresponding plaintext block p1 , . . . , pn by
matrix multiplication modulo Z29 :
   
c1 p1
 ..   . 
 .  = A  ..  .
cn pn

The matrix A is called the encryption matrix of the cipher. It inverse A−1 is called the decryption
matrix.

Example 4.74: Hill cipher: encryption

Encrypt the message “Meet me tomorrow” using the Hill cipher with block size 3 and encryption
matrix  
2 4 1
A =  3 1 5 .
1 3 2

Solution. We start by converting the message “Meet me tomorrow” to a sequence of scalars. We have
M = 13, E = 5, and so on. The encoded plaintext is 13, 5, 5, 20, 0, 13, 5, 0, 20, 15, 13, 15, 18, 18, 15, 23. Next,
we divide the plaintext into blocks of length 3. Since the length of the plaintext is not a multiple of three,
we pad the final block with spaces, i.e., with zeros.

Plaintext blocks: (13, 5, 5), (20, 0, 13), (5, 0, 20), (15, 13, 15), (18, 18, 15), (23, 0, 0).

To compute the ciphertext, we regard each plaintext block as a 3-dimensional column vector and multiply
by the encryption matrix A. All calculations are done modulo 29. For example, for the first block, we have
      
13 2 4 1 13 22
A  5  =  3 1 5   5  =  11  ,
5 1 3 2 5 9
 4.9. Application: Cryptography 169

so the first ciphertext block is (22, 11, 9). We repeat the same with the remaining plaintext blocks.
           
20 24 5 1 15 10
A  0  =  9  , A  0  =  28  , A  13  =  17  ,
13 17 20 16 15 26
       
18 7 23 17
A  18  =  2  , A  0  =  11  .
15 15 0 23

Therefore, we have found the following ciphertext blocks:

Ciphertext blocks: (22, 11, 9), (24, 9, 17), (1, 28, 16), (10, 17, 26), (7, 2, 15), (17, 11, 23).

Finally, we can convert the ciphertext to a list of symbols: “ VKIXIQA.PJQZGBOQKW ”. ♠

Example 4.75: Hill cipher: decryption

Decrypt the message “ RNOLFPHHCIGH DE ” using the Hill cipher with block size 3 and encryption
matrix  
2 4 1
A =  3 1 5 .
1 3 2

Solution. The process is analogous to encryption, except that we need to use the decryption matrix A−1
instead of A. We first calculate A−1 , keeping in mind that scalars are from the field Z29 . The method is the
same as in Example 4.71; we skip the individual steps in the interest of brevity.
   
2 4 1 1 0 0 1 0 0 23 20 11  
[A | I] =  3 1 5 0 1 0  ≃ . . . ≃  0 1 0 4 17 28  = I | A−1 .
1 3 2 0 0 1 0 0 1 26 8 11

Next, we convert the 15 ciphertext symbols “ RNOLFPHHCIGH DE ” to scalars and divide them into blocks
of length 3:
Ciphertext blocks: (18, 14, 15), (12, 6, 16), (8, 8, 3), (9, 7, 8), (0, 4, 5).
Now we decrypt each ciphertext block by a matrix multiplication with A−1 .
           
18 18 12 21 8 0
A−1  14  =  5  , A−1  6  =  18  , A−1  8  =  20  ,
15 20 16 14 3 15
       
9 0 0 19
A−1  7  =  2  , A−1  4  =  5  .
8 1 5 0

This yields the following plaintext blocks:

Plaintext blocks: (18, 5, 20), (21, 18, 14), (0, 20, 15), (0, 2, 1), (19, 5, 0).
170 Matrices

Converting these back to letters, and omitting the trailing space, we find that the plaintext is “return to
base”. ♠
It is important to note that, despite its good diffusion properties, the Hill cipher is not secure. The cipher
has many weaknesses. For one, because A0 = 0, a block of spaces in the plaintext will always be encrypted
as a block of spaces in the ciphertext, regardless of the encryption matrix A. More importantly, the cipher
is subject to a so-called known plaintext attack. If an eavesdropper intercepts some ciphertext for which
a small amount of the corresponding plaintext happens to be known, it is immediately possible to recover
the key and therefore decrypt the rest of the ciphertext. Carrying out this attack only requires some basic
knowledge of linear algebra. The following example illustrates how this is done.

Example 4.76: Cryptanalysis of the Hill cipher: known plaintext attack

Eve intercepts the following encrypted message sent by Alice:

“ EFNOR.AHIFNEPL.TSZS,RSKT.ZBBRFVUPFVZLFHNTV ” .

Eve knows that Alice uses a Hill cipher with block length 3, but she does not know the secret
encryption matrix. Eve also knows that Alice begins all of her correspondence with “My dear
love”. Decrypt the message.

Solution. The first three blocks of the ciphertext are “ EFNOR.AHI ”, i.e.,

Ciphertext blocks: (5, 6, 14), (15, 18, 28), (1, 8, 9).

Eve also knows that the first three blocks of the plaintext are “ MY DEAR L ”, i.e.,

Plaintext blocks: (13, 25, 0), (4, 5, 1), (18, 0, 12).

These facts allow Eve to deduce the following information about the unknown decryption matrix A−1 :
           
5 13 15 4 1 18
A−1  6  =  25  , A−1  18  =  5  , A−1  8  =  0  .
14 0 28 1 9 12

Since Eve remembers the column method of matrix multiplication, she knows that these three equations
can be written as a single equation in matrix form:
   
5 15 1 13 4 18
A−1  6 18 8  =  25 5 0  .
14 28 9 0 1 12

Note that this equation is of the form A−1C = P. (Here, C stands for “ciphertext” and P for “plaintext”).
Multiplying both sides of the equation by C−1 on the right, we get A−1 = PC−1 . Thus, assuming that C is
invertible, Eve can easily compute the decryption matrix A−1 . Eve computes:
 −1  
5 15 1 19 8 23
C−1 =  6 18 8  =  0 5 2  .
14 28 9 22 1 0
 4.9. Application: Cryptography 171

This allows Eve to compute the decryption matrix:

    
13 4 18 19 8 23 5 26 17
A−1 = PC−1 =  25 5 0   0 5 2  =  11 22 5  .
0 1 12 22 1 0 3 17 2

Armed with the decryption matrix A−1 , Eve can now decrypt Alice’s entire message, using the same
method as in Example 4.75. The plaintext is “My dear love, run away with me at midnight”. ♠
As the example shows, the Hill cipher is not secure at all. The main problem is that the cipher is linear, i.e.,
each component of a ciphertext block is a simple linear combination of the components of the plaintext
block. This linearity property enables Eve to break the cipher by solving a system of linear equations.
For this reason, all modern block ciphers have a non-linear component. Often this takes the form of
so-called S-boxes. An S-box is an operation that scrambles the symbols of the alphabet in a non-linear
way. For example, consider the following S-box, which is an operation from Z29 to Z29 :

x 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28
S(x) 17 9 27 2 20 12 21 26 16 18 4 24 23 7 19 14 28 29 1 15 10 22 6 5 25 11 13 3 8

The inputs of the S-box are shown in the top row, and the corresponding outputs in the bottom row. For
example, this S-box maps the input 7 to the output 26. We write S(7) = 26.

Definition 4.77: A toy block cipher

Consider the following block cipher on the alphabet Z29 with block size 3. The key consists of
12 elements k1 , . . . , k12 of Z29 . To encrypt a plaintext block, regard the block as a 3-dimensional
column vector. Then repeat the following steps 3 times. All operations are carried out modulo 29.

• Key mixing: add the next three components of the key to the components of the vector.
 
1 2 3
• Diffusion: multiply the vector by the fixed 3 × 3-matrix A =  3 1 2 .
2 3 1
• S-box application: apply the S-box to each component of the vector.

Finally, apply one more key mixing step at the end. The resulting vector is the ciphertext block.
The cipher can be visualized as follows:

k1 S k4 S k7 S k10

plaintext k2 A S k5 A S k8 A S k11 ciphertext

k3 S k6 S k9 S k12

Note that the three basic steps (key mixing, diffusion, and S-box application) are repeated several times;
each such repetition is called a round of the block cipher. The more rounds a block cipher has, the better
172 Matrices

its diffusion and non-linearity properties. The final round is short: it only consists of a key mixing step,
with no final diffusion or S-box application. The reason is that performing a final diffusion and S-box
application would not add anything to the security of the cipher. An attacker could simply undo these last
two steps, since they do not depend on the key.
The matrix A is called the diffusion matrix of the cipher. Note that, unlike for the Hill cipher, the
matrix A is fixed once and for all and is not part of the key. Instead, the key consists of scalars that are
added to the current block at the beginning of each round.

Example 4.78: Toy block cipher: encryption

Encrypt the message “I like math” using the block cipher of Definition 4.77 and the key
1, 1, 3, 3, 5, 5, 7, 7, 9, 9, 11, 11.

Solution. We first represent the plaintext as a sequence of blocks, padding the final block with zeros:

Plaintext blocks: (9, 0, 12), (9, 11, 5), (0, 13, 1), (20, 8, 0).

To encrypt the first block, we start with the vector [9, 0, 12]T and apply the following steps:
Round 1:

• Key mixing: the first three components of the key are 1, 1, 3. We add them to the plaintext.
     
9 1 10
 0  + 1  =  1 .
12 3 15

• Diffusion: multiply by the matrix A.

    
1 2 3 10 28
 3 1 2  1  =  3 .
2 3 1 15 9

• S-box application: apply the S-box to each component of the vector.

   
S(28) 8
 S(3)  =  2  .
S(9) 18

Round 2:

• Key mixing: the next three components of the key are 3, 5, 5.

     
8 3 11
 2  + 5  =  7 .
18 5 23
 4.9. Application: Cryptography 173

• Diffusion:     
1 2 3 11 7
 3 1 2   7  =  28  .
2 3 1 23 8

• S-box application:    
S(7) 26
 S(28)  =  8  .
S(8) 16

Round 3:

• Key mixing: the next three components of the key are 7, 7, 9.

     
26 7 4
 8  +  7  =  15  .
16 9 25

• Diffusion:     
1 2 3 4 22
 3 1 2   15  =  19  .
2 3 1 25 20

• S-box application:    
S(22) 6
 S(19)  =  15  .
S(20) 10

Round 4 (the final round is abbreviated):

• Key mixing: the next three components of the key are 9, 11, 11.
     
6 9 15
 15  +  11  =  26  .
10 11 21

Therefore, the first ciphertext block is (15, 26, 21). We repeat the same procedure with the remaining
plaintext blocks, and obtain the following ciphertext blocks:

Ciphertext blocks: (15, 26, 21), (7, 24, 1), (2, 16, 23), (7, 20, 22).

The corresponding ciphertext is “ OZUGXABPWGTV ”. ♠

Are ciphers like this actually used in the real world? The answer is yes. While the cipher of Definition 4.77
is greatly simplified, it has the same basic structure as modern real-world block ciphers (such as AES, the
Advanced Encryption Standard). Naturally, these real-world ciphers differ in some details, such as the
alphabet size, the block size, the number of rounds, the design of the S-boxes, the way the key is computed,
and the precise order in which the operations are applied. However, their basic structure is very similar to
174 Matrices

our toy cipher, and indeed, all such ciphers rely on key mixing, diffusion, and non-linear S-boxes as their
key components.
For example, AES uses an alphabet size of 256 instead of 29 (i.e., it operates on bytes, rather than
elements of Z29 ). Although Z256 is not a field (because 256 is not prime), it nevertheless turns out that
there exists a field with 256 elements, and AES uses it for its algebraic operations. Our toy cipher’s
block size of 3 is much too small to achieve effective diffusion; modern real-world ciphers use block sizes
between 16 and 32 bytes (128 to 256 bits). The design of the S-boxes is a bit of a black art; at minimum,
they must be designed to withstand two common types of cryptanalysis known as linear cryptanalysis
and differential cryptanalysis. Among other things, this means that the S-box should be “as far from
linear” as possible.
A detailed discussion of the design and cryptanalysis of modern block ciphers is far beyond the scope
of this book, but we hope that you have gotten a taste of this fascinating subject, and the role that linear
algebra over finite fields plays in it.

Exercises

Exercise 4.9.1 Encrypt the message “Rendezvous at dawn” using the Hill cipher with block size 3 and
encryption matrix  
2 1 1
A =  1 3 1 .
1 1 4

Exercise 4.9.2 Decrypt the message “ ERM DXYBJUWW.JWQLD,HL ” using the Hill cipher with block size 3
and encryption matrix  
2 1 1
A =  1 3 1 .
1 1 4

Exercise 4.9.3 Eve intercepts the following encrypted message sent by Bob:

“ TGVXKHGSW,JU,JHYJSCDSBQIRPEV ”

Eve knows that Alice uses a Hill cipher with block length 2, but she does not know the secret encryption
matrix. Eve also knows that Bob begins all of his letters with “Hello”. Decrypt the message.

Exercise 4.9.4 Encrypt the message “Lost contact” using the block cipher of Definition 4.77 and the key
2, 3, 4, 1, 1, 1, 5, 5, 5, 4, 3, 2.

Exercise 4.9.5 Decrypt the message “ NRQEUAPOM GLFN ”, using the block cipher of Definition 4.77 and
the key 1, 1, 1, 2, 2, 2, 3, 3, 3, 4, 4, 4. Hint: To decrypt, we must perform all the encryption steps in reverse.
To undo a key mixing step, we subtract the relevant key components. To undo an S-box application, we
apply the S-box in reverse. To undo a diffusion step, multiply by the inverse of the diffusion matrix.
5. Spans, linear independence, and bases in Rn

5.1 Spans

Outcomes
A. Determine the span of a set of vectors.

B. Determine if a vector is contained in a specified span.

Let u and v be two non-parallel vectors in Rn . We can picture the set of their linear combinations as
follows:

3u + 3v
1u + 3v
1u + 3v
3u + 2v
0u + 3v

−1u + 3v 3u + 1v

−1u + 2v 3u + 0v

−1u + 1v v 3u − 1v
u
0 2u − 1v
−1u + 0v
1u − 1v
0u − 1v
−1u − 1v

As the picture shows, the linear combinations of u and v form a 2-dimensional plane through the origin.
We say that this plane is spanned by the vectors u and v. This concept generalizes to more than two
vectors. For example, three vectors may span a 3-dimensional space (although sometimes, they span only
a 2-dimensional space, or even a line). This motivates the following definition.

175
176 Spans, linear independence, and bases in Rn

Definition 5.1: Span of a set of vectors

The set of all linear combinations of the vectors u1 , . . . , uk in Rn is known as the span of these
vectors and is written as span {u1 , . . . , uk }. Using set notation, we can write

span {u1 , . . . , uk } = {a1 u1 + . . . + ak uk | a1 , . . . , ak ∈ R} .

Example 5.2: Vectors in a span

   
1 3
Let u = 1 and v = 2 . Which of the following vectors are elements of span {u, v}?
  
1 1
   
2 1

(a) w = 3 , (b) z = 1  .

4 2

Solution. (a) For a vector to be in span {u, v}, it must be a linear combination of u and v. Therefore,
w ∈ span {u, v} if and only if we can find find scalars a, b such that a u + b v = w. We must therefore solve
the equation      
1 3 2
a 1 +b 2 = 3 .
    
1 1 4
We write this as an augmented matrix and solve.
   
1 3 2 1 0 5
 1 2 3  ≃ . . . ≃  0 1 −1  .
1 1 4 0 0 0

The solution is a = 5 and b = −1. This means that w = 5u + (−1)v. Therefore, w is an element of
span {u, v}.
(b) We repeat the same method with the vector z. This time, we have to find a, b such that a u + b v = z.
The system of equations is    
1 3 1 1 3 1
 1 2 1  ≃ . . . ≃  0 −1 0  ,
1 1 2 0 0 1
which is inconsistent. Therefore, there is no solution. We conclude that z is not an element of span {u, v}.
♠

Example 5.3: Describing the span

  
1 3
Describe the span of the vectors u = 1 and v = 2  in R3 .
  
1 1
 5.1. Spans 177

Solution. Let w = [x, y, z]T be any vector. Proceeding as in the previous example, we know that w is an
element of span {u, v} if and only if the equation
     
1 3 x
a 1 +b 2  =  y 
1 1 z

is consistent. Note that the variables of this equation are a, b; we regard x, y, z as constants for the moment.
We write the augmented matrix of this system and reduce to echelon form:
     
1 3 x R2←R2 −R1 1 3 x 1 3 x
R ←R −R R ←R −2R
 1 2 y  2 ≃3 1  0 −1 y − x  3 ≃3 2  0 −1 y−x ,
1 1 z 0 −2 z − x 0 0 (z − x) − 2(y − x)

From the echelon form, we see that the system is consistent if and only if (z − x) − 2(y − x) = 0, or
equivalently x − 2y + z = 0. Therefore, the vector w is in span {u, v} if and only if x − 2y + z = 0. In other
words, the span of u and v is the plane x − 2y + z = 0. ♠

Example 5.4: Span of redundant vectors

     
1 3 11
Let u =  1 , v =  2 , and w =  8 . Show that span {u, v, w} = span {u, v}.
1 1 5

Solution. Observe that w = 2 u + 3 v. Therefore, w is already in the span of u and v. Two sets are equal if
they have the same elements, i.e., each element of the first set is an element of the second set and vice versa.
Therefore, to show span {u, v, w} = span {u, v}, we must show (a) that every element of span {u, v, w} is
an element of span {u, v} and (b) vice versa.
(a) Let z be an arbitrary element of span {u, v, w}. Then, by definition of span, there exist scalars a, b, c
such that

z = a u + b v + c w.

But as observed above, we have w = 2 u + 3 v, and therefore we can also write

z = a u + b v + c(2 u + 3 v)
= (a + 2c)u + (b + 3c)v.

It follows that z is a linear combination of u and v, and therefore, z ∈ span {u, v}.
(b) Clearly every linear combination of u and v is also a linear combination of u, v, and w, namely,
taking the coefficient of w to be 0. Therefore, every element of span {u, v} is an element of span {u, v, w}.
Because we have shown that every element of span {u, v, w} is an element of span {u, v} and vice
versa, it follows that span {u, v, w} and span {u, v} are the same set of vectors. ♠
In the situation of the last example, we say that the vector w is redundant; it does not contribute anything
to span {u, v, w}. Geometrically, the three vectors u, v, and w lie in a plane. Since the two vectors u and v
178 Spans, linear independence, and bases in Rn

are sufficient to span this plane, the third vector w is not really needed. We will study this situation more
systematically in the next section.

Example 5.5: Span of the empty set

We talked about the span of k vectors u1 , . . . , uk . What if k = 0? What is the span of an empty set
of vectors?

Solution. Consider what happens when we compute the sum of three numbers. We usually write this as
b1 + b2 + b3 . We can also compute the sum of three numbers by starting from 0 and then adding each of
the three numbers to it. I.e., the sum can be computed as 0 + b1 + b2 + b3 . Similarly, we can write the
sum of two numbers as 0 + b1 + b2 , and the sum of just one number as 0 + b1 . Continuing the pattern, it
follows that the sum of zero numbers should be 0:

Sum of 3 numbers: 0 + b1 + b2 + b3 .
Sum of 2 numbers: 0 + b1 + b2 .
Sum of 1 numbers: 0 + b1 .
Sum of 0 numbers: 0.

The sum of zero numbers is also called the empty sum. It is equal to the unit of addition, i.e., 0. By an
analogous argument, the empty sum of vectors is equal to the unit of vector addition, i.e., to the zero vector
0. In general, if we have k vectors u1 , . . . , uk , the span consists of all vectors of the form a1 u1 + . . . + ak uk ,
which is a sum of k vectors. In case k = 0, the span consists only of the empty sum, i.e., the zero vector 0,
which we also call the empty linear combination. Therefore, the span of the empty set of vectors is {0}.
♠

Exercises

Exercise 5.1.1 Consider the vectors

       
1 1 1 −1
u1 =  1  , u2 =  2  , u3 =  −3  , u4 =  1  .
−2 −2 −2 2

Which of the following vectors are in span {u1 , . . . , u4 }? For each vector that is in the span, exhibit a linear
combination of u1 , . . . , u4 that equals this vector.
     
1 2 1
(a) x =  2  , (b) y =  −3  , (c) z =  5  .
−1 −4 −2
   
1 2
Exercise 5.1.2 Describe the span of the vectors u =  0  and v =  −1  in R3 .
2 1
 5.2. Linear independence 179

Exercise 5.1.3 Describe the span of the following vectors in R4 :

       
1 5 0 3
 1   1   −2   1 
u1 =    
 −1  , u2 =  1  , u3 =  2
 ,
 u4 =  
 1 .
−1 0 1 1


    
2 1 1
Exercise 5.1.4 Let u =  1 , v =  3 , and w =  0 . Show that span {u, v, w} = span {u, v}.
0 5 −1

Exercise 5.1.5 Suppose {u1 , . . ., uk } is a set of vectors from Rn . Show that 0 ∈ span {u1 , . . . , uk }.

Exercise 5.1.6 In this exercise, we use scalars from the field Z5 of integers modulo 5 instead of real
numbers (see Section 1.8, “Fields”). Consider the vectors
       
1 0 2 3
u1 =  4  , u2 =  2  , u3 =  1  , u4 =  1  .
3 3 3 0
Which of the following vectors are in span {u1 , . . . , u4 }? For each vector that is in the span, exhibit a linear
combination of u1 , . . . , u4 that equals this vector.
     
0 1 2
(a) x = 1 ,   (b) y = 1 ,  (c) z = 0  .

4 2 4

5.2 Linear independence

Outcomes
A. Find the redundant vectors in a set of vectors.

B. Determine whether a set of vectors is linearly independent.

C. Find a linearly independent subset of a set of spanning vectors.

D. Write a vector as a unique linear combination of a set of linearly independent vectors.

5.2.1. Redundant vectors and linear independence

In Example 5.4, we encountered three vectors u, v, and w such that span {u, v, w} = span {u, v}. If this
happens, then the vector w does not contribute anything to the span of {u, v, w}, and we say that w is
redundant. The following definition generalizes this notion.
180 Spans, linear independence, and bases in Rn

Definition 5.6: Redundant vectors, linear dependence, and linear independence

Consider a sequence of k vectors u1 , . . . , uk . We say that the vector u j is redundant if it can be
written as a linear combination of earlier vectors in the sequence, i.e., if

u j = a1 u1 + a2 u2 + . . . + a j−1 u j−1

for some scalars a1 , . . . , a j−1 . We say that the sequence of vectors u1 , . . . , uk is linearly depen-
dent if it contains one or more redundant vectors. Otherwise, we say that the vectors are linearly
independent.

Example 5.7: Redundant vectors

Find the redundant vectors in the following sequence of vectors. Are the vectors linearly indepen-
dent?
           
0 1 1 2 0 3
 0   2   1   3   1   3 
u1 =          
 0  , u2 =  2  , u3 =  1  , u4 =  3  , u5 =  2  , u6 =  2  .
 

0 3 1 4 3 2

Solution.

• The vector u1 is redundant, because it is a linear combination of earlier vectors. (Although there are
no earlier vectors, recall from Example 5.5 that the empty sum of vectors is equal to the zero vector
0. Therefore, u1 is indeed an (empty) linear combination of earlier vectors.)

• The vector u2 is not redundant, because it cannot be written as a linear combination of u1 . This is
because the system of equations  
0 1
 0 2 
 
 0 2 
0 3
has no solution.

• The vector u3 is not redundant, because it cannot be written as a linear combination of u1 and u2 .
This is because the system of equations
 
0 1 1
 0 2 1 
 
 0 2 1 
0 3 1

has no solution.

• The vector u4 is redundant, because u4 = u2 + u3 .

 5.2. Linear independence 181

• The vector u5 is not redundant, because This is because the system of equations
 
0 1 1 2 0
 0 2 1 3 1 
 
 0 2 1 3 2 
0 3 1 4 3

has no solution.

• The vector u6 is redundant, because u6 = u2 + 2u3 − u5 .

In summary, the vectors u1 , u4 , and u6 are redundant, and the vectors u2 , u3 , and u5 are not. It follows
that the vectors u1 , . . . , u6 are linearly dependent. ♠

5.2.2. The casting-out algorithm

The last example shows that it can be a lot of work to find the redundant vectors in a sequence of k vectors.
Doing so in the naive way require us to solve up to k systems of linear equations! Fortunately, there is a
much faster and easier method, the so-called casting-out algorithm.

Algorithm 5.8: Casting-out algorithm

Input: a list of k vectors u1 , . . . , uk ∈ Rn .
Output: the set of indices j such that u j is redundant.
Algorithm: Write the vectors u1 , . . . , uk as the columns of an n × k-matrix, and reduce to echelon
form. Every non-pivot column, if any, corresponds to a redundant vector.

Example 5.9: Casting-out algorithm

Use the casting-out algorithm to find the redundant vectors among the vectors from Example 5.7.

Solution. Following the casting-out algorithm, we write the vectors u1 , . . . , u6 as the columns of a matrix
and reduce to echelon form.
   
0 1 1 2 0 3 0 1 1 2 0 3
 0 2 1 3 1 3   3 
  ≃ . . . ≃  0 0 1 1 −1 .
 0 2 1 3 2 2   0 0 0 0 1 −1 
0 3 1 4 3 2 0 0 0 0 0 0

The pivot columns are columns 2, 3, and 5. The non-pivot columns are columns 1, 4, and 6. Therefore,
the vectors u1 , u4 , and u6 are redundant. Note that this is the same answer we got in Example 5.7. ♠
The above version of the casting-out algorithm only tells us which of the vectors (if any) are redundant,
but it does not give us a specific way to write the redundant vectors as linear combinations of previous
vectors. However, we can easily get this additional information if we reduce the matrix all the way to
reduced echelon form. We call this version of the algorithm the extended casting-out algorithm.
182 Spans, linear independence, and bases in Rn

Algorithm 5.10: Extended casting-out algorithm

Input: a list of k vectors u1 , . . . , uk ∈ Rn .
Output: the set of indices j such that u j is redundant, and a set of coefficients for writing each
redundant vector as a linear combination of previous vectors.
Algorithm: Write the vectors u1 , . . ., uk as the columns of an n × k-matrix, and reduce to reduced
echelon form. Every non-pivot column, if any, corresponds to a redundant vector. If u j is a re-
dundant vector, then the entries in the j th column of the reduced echelon form are coefficients for
writing u j as a linear combination of previous non-redundant vectors.

Example 5.11: Extended casting-out algorithm

Use the casting-out algorithm to find the redundant vectors among the vectors from Example 5.7,
and write each redundant vector as a linear combination of previous non-redundant vectors.

Solution. Once again, we write the vectors u1 , . . . , u6 as the columns of a matrix. This time we use the
extended casting-out algorithm, which means we reduce the matrix to reduced echelon form instead of
echelon form.    
0 1 1 2 0 3 0 1 0 1 0 1
 0 2 1 3 1 3   0 0 1 1 0 2 
   
 0 2 1 3 2 2  ≃ . . . ≃  0 0 0 0 1 −1  .
0 3 1 4 3 2 0 0 0 0 0 0
As before, the non-pivot columns are columns 1, 4, and 6, and therefore, the vectors u1 , u4 , and u6 are
redundant. The non-redundant vectors are u2 , u3 , and u5 . Moreover, the entries in the sixth column are 1,
2, and −1. Note that this means that the sixth column can be written as 1 times the second column plus 2
times the third column plus (−1) times the fourth column. The same coefficients can be used to write u6
as a linear combination of previous non-redundant columns, namely:

u6 = 1 u2 + 2 u3 − 1 u5 .

Also, the entries in the fourth column are 1 and 1, which are the coefficients for writing u4 as a linear
combination of previous non-redundant columns, namely:

u4 = 1 u2 + 1 u3 .

Finally, there are no non-zero entries in the first column. This means that u1 is the empty linear combina-
tion
u1 = 0.
♠

5.2.3. Alternative characterization of linear independence

Our definition of redundant vectors depends on the order in which the vectors are written. This is because
each redundant vector must be a linear combination of earlier vectors in the sequence. For example, in
 5.2. Linear independence 183

the sequence of vectors      

1 3 11
u =  1 , v =  2 , w =  8 ,
1 1 5
the vector w is redundant, because it is a linear combination of earlier vectors w = 2 u + 3 v. Neither u nor
v are redundant. On the other hand, in the sequence of vectors
     
1 11 3

u= 1 , w=   8 , v = 2 ,
 
1 5 1

v is redundant because v = 13 w − 23 u, but neither u nor w are redundant. Note that none of the vectors have
changed; only the order in which they are written is different. Yet w is the redundant vector in the first
sequence, and v is the redundant vector in the second sequence.
Because we defined linear independence in terms of the absence of redundant vectors, you may sus-
pect that the concept of linear independence also depends on the order in which the vectors are written.
However, this is not the case. The following theorem gives an alternative characterization of linear inde-
pendence that is more symmetric (it does not depend on the order of the vectors).

Theorem 5.12: Characterization of linear independence

Let u1 , . . . , uk be vectors. Then u1 , . . . , uk are linearly independent if and only if the homogeneous
equation
a1 u1 + . . . + ak uk = 0
has only the trivial solution.

Proof. Let A be the n × k-matrix whose columns are u1 , . . . , uk . We know from the theory of homogeneous
systems that the system a1 u1 + . . . + ak uk = 0 has no non-trivial solution if and only if every column of
the echelon form of A is a pivot column. By the casting-out algorithm, this is the case if and only if none
of the vectors u1 , . . . , uk are redundant, i.e., if and only if the vectors are linearly independent. ♠

Example 5.13: Characterization of linear independence

Use the method of Theorem 5.12 to determine whether the following vectors are linearly indepen-
dent in R4 .        
1 0 1 2
 1   1   2   3 
u1 =        
 2  , u2 =  1  , u3 =  3  , u4 =  3  .
0 1 2 1

Solution. We must check whether the equation

         
1 0 1 2 0
 1     2   3   0 
a1   + a2  1  + a3   + a4  = 
 2   1   3   3   0 
0 1 2 1 0
184 Spans, linear independence, and bases in Rn

has a non-trivial solution. If it does, the vectors are linearly dependent. On the other hand, if there is only
the trivial solution, the vectors are linearly independent. We write the augmented matrix and solve:
   
1 0 1 2 0 1 0 1 2 0
 1 1 2 3 0   
  ≃ ... ≃  0 1 1 1 0 .
 2 1 3 7 0   0 0 1 0 0 
0 1 2 1 0 0 0 0 2 0
Since every column is a pivot column, there are no free variables; the system of equations has a unique
solution, which is a1 = a2 = a3 = a4 = 0, i.e., the trivial solution. Therefore, the vectors u1 , . . . , u4 are
linearly independent. ♠

Example 5.14: Characterization of linear independence

Use the method of Theorem 5.12 to determine whether the following vectors are linearly indepen-
dent in R3 .      
1 1 0
u1 = 1 , u2 = 3 , u3 = 4  ,
    
0 1 2

Solution. As in the previous example, we must check whether the equation a1 u1 + a2 u2 + a3 u3 = 0 has a
non-trivial solution. Once again, we write the augmented matrix and solve:
   
1 1 0 0 1 1 0 0
 1 3 4 0  ≃ ... ≃  0 2 4 0 .
0 1 2 0 0 0 0 0
Since column 3 is not a pivot column, a3 is a free variable. Therefore, the system has a non-trivial solution,
and the vectors are linearly dependent.
With a small amount of extra work, we can find an actual non-trivial solution of a1 u1 + a2 u2 + a3 u3 =
0. All we have to do is set a3 = 1 and do a back substitution. We find that (a1 , a2 , a3 ) = (2, −2, 1) is a
solution. In other words,
2u1 − 2u2 + u3 = 0.
We can also use this information to write u3 as a linear combination of previous vectors, namely, u3 =
−2u1 + 2u2 . ♠
The characterization of linear independence in Theorem 5.12 is mostly useful for theoretical reasons.
However, it can also help in solving problems such as the following.

Example 5.15: Related sets of vectors

Let u, v, w be linearly independent vectors in Rn . Are the vectors u + v, 2u + w, and v − 5w linearly
independent?

Solution. By Theorem 5.12, to check whether the vectors are linearly independent, we must check whether
the equation
a(u + v) + b(2u + w) + c(v − 5w) = 0 (5.1)
 5.2. Linear independence 185

has non-trivial solutions. If it does, the vectors are linearly dependent, if it does not, they are linearly
independent. We can simplify the equation as follows:

(a + 2b)u + (a + c)v + (b − 5c)w = 0. (5.2)

Since u, v, and w are linearly independent, we know, again by Theorem 5.12, that equation (5.2) only has
the trivial solution. Therefore,

a + 2b = 0,
a + c = 0,
b − 5c = 0.

We can solve this system of three equations in three variables, and we find that it has the unique solution
a = b = c = 0. Therefore, a = b = c = 0 is the only solution to equation (5.1), which means that the vectors
u + v, 2u + w, and v − 5w are linearly independent. ♠

5.2.4. Properties of linear independence

The following are some properties of linearly independent sets.

Proposition 5.16: Properties of linear independence

1. Linear independence and reordering. If a sequence u1 , . . . , uk of k vectors is linearly inde-

pendent, then so is any reordering of the sequence (i.e., whether or not the vectors are linearly
independent does not depend on the order in which the vectors are written down).

2. Linear independence of a subset. If u1 , . . . , uk are linearly independent, then so are u1 , . . . , u j

for any j < k.

3. Linear independence and dimension. Let u1 , . . . , uk be a sequence of k vectors in Rn . If

k > n, then the vectors are linearly dependent (i.e., not linearly independent).

Proof.

1. This follows from Theorem 5.12, because whether or not the equation a1 u1 + . . . + ak uk = 0 has a
non-trivial solution does not depend on the order in which the vectors are written.

2. If one of the vectors in the sequence u1 , . . . , u j were redundant, then it would be redundant in the
longer sequence u1 , . . . , uk as well.

3. Let A be the n × k-matrix that has the vectors u1 , . . . , uk as its columns and suppose that k > n. Then
the rank of A is at most n, so the echelon form of A has some non-pivot columns. Therefore, the
system a1 u1 + . . . + ak uk = 0 has non-trivial solutions, and the vectors are linearly dependent by
Theorem 5.12.

♠
186 Spans, linear independence, and bases in Rn

Example 5.17: Linear dependence

Are the following vectors linearly independent?
     
1 2 3
, , .
4 3 2

Solution. Since these are 3 vectors in R2 , they are linearly dependent by Proposition 5.16. No calculation
is necessary. ♠

5.2.5. Linear independence and linear combinations

In general, there is more than one way of writing a given vector as a linear combination of some spanning
vectors. For example, consider
       
1 0 1 1
u1 = 1 , u2 = 1 , u3 = 2 , v = 3  .
      
0 1 1 2

We can write v in many different ways as a linear combination of u1 , . . . , u3 , for example

v = −u1 + 2u3 ,
v= u2 + u3 ,
v= u1 + 2u2 ,
v = 2u1 + 3u2 − u3 .

However, when the vectors u1 , . . . , uk are linearly independent, this does not happen. In this case, the linear
combination is always unique, as the following theorem shows.

Theorem 5.18: Unique linear combination

Assume u1 , . . . , uk are linearly independent. Then every vector v ∈ span {u1 , . . . , uk } can be written
as a linear combination of u1 , . . . , uk in a unique way.

Proof. We already know that every vector v ∈ span {u1 , . . . , uk } can be written as a linear combination of
u1 , . . . , uk , because that is the definition of span. So what must be proved is the uniqueness. Suppose,
therefore, that there are two ways of writing v as such a linear combination, i.e., that

v = a1 u1 + a2 u2 + . . . + ak uk and
v = b1 u1 + b2 u2 + . . . + bk uk .

Subtracting one equation from the other, we get

0 = (a1 − b1 )u1 + (a2 − b2 )u2 + . . . + (ak − bk )uk .

Since u1 , . . . , uk are linearly independent, we know by Theorem 5.12 that the last equation only has the
trivial solution, i.e., a1 − b1 = 0, a2 − b2 = 0, . . . , ak − bk = 0. It follows that a1 = b1 , a2 = b2 , . . . , ak = bk .
 5.2. Linear independence 187

We have shown that any two ways of writing v as a linear combination of u1 , . . . , uk are equal. Therefore,
there is only one way of doing so. ♠

5.2.6. Removing redundant vectors

Consider the span of some vectors u1 , . . . , uk . As we just saw in the previous subsection, the span is
especially nice when the vectors u1 , . . . , uk are linearly independent, because in that case, every element v
of the span can be uniquely written in the form v = a1 u1 + . . . + ak uk .
But what if we have a span of some vectors u1 , . . . , uk that are not linearly independent? It turns out
that we can always find some linearly independent vectors that span the same set. In fact, this can be done
by simply removing the redundant vectors from u1 , . . . , uk . This is the subject of the following theorem.

Theorem 5.19: Removing redundant vectors

Let u1 , . . . , uk be a sequence of vectors, and suppose that u j1 , . . . , u jℓ is the subsequence of vectors
that is obtained by removing all of the redundant vectors. Then u j1 , . . . , u jℓ are linearly independent
and 
span u j1 , . . . , u jℓ = span {u1 , . . . , uk } .

Proof. Remove the redundant vectors one by one, from right to left. Each time a redundant vector is
removed, the span does not change; the proof of this is similar to Example 5.4. Moreover, the resulting
sequence of vectors span u j1 , . . . , u jℓ is linearly independent, because if any of these vectors were a
linear combination of earlier ones, then it would have been redundant in the original sequence of vectors,
and would have therefore been removed. ♠

Example 5.20: Finding a linearly independent set of spanning vectors

Find a subset of {u1 , . . . , u4 } that is linearly independent and has the same span as {u1 , . . . , u4 }.
       
1 −2 1 3
 0   0   2   4 
u1 =      
 −2  , u2 =  4  , u3 =  2  , u4 =  2  .
 

3 −6 1 5

Solution. We use the casting-out algorithm to find the redundant vectors:

   
1 −2 1 3 1 −2 1 3
 0 0 2 4    0 0 2 4 
 ≃ ... ≃  .
 −2 4 2 2   0 0 0 0 
3 −6 1 5 0 0 0 0

Therefore, the redundant vectors are u2 and u4 . We remove them (“cast them out”) and are left with u1 and
u3 . Therefore, by Theorem 5.19, {u1 , u3 } is linearly independent and span {u1 , u3 } = span {u1 , . . . , u4 }.
♠
188 Spans, linear independence, and bases in Rn

Exercises

Exercise 5.2.1 Which of the following vectors are redundant? If there are redundant vectors, write each
of them as a linear combination of previous vectors.
       
1 2 1 1
u1 =  0  , u2 =  0  , u3 =  2  , u4 =  6  .
1 2 1 1

Exercise 5.2.2 Which of the following vectors are redundant? If there are redundant vectors, write each
of them as a linear combination of previous vectors.
       
−1 −3 0 0
 −2       
u1 =   , u2 =  −4  , u3 =  −1  , u4 =  2  .
 2   3   4   −3 
3 3 3 −6

Exercise 5.2.3 Use the method of Theorem 5.12 to determine whether the following vectors are linearly
independent. If they are linearly dependent, find a non-trivial linear combination of the vectors that is
equal to 0.        
2 −1 −3 5
 
u1 = 3 , u2 =  0 , u3 = −4 , u4 = 6  .
   
2 2 −2 2

Exercise 5.2.4 Use the method of Theorem 5.12 to determine whether the following vectors are linearly
independent. If they are linearly dependent, find a non-trivial linear combination of the vectors that is
equal to 0.        
1 1 3 1
 −1   6   5   0 
u1 =      
 0  , u2 =  7  , u3 =  8  , u4 =  1  .
 

1 1 3 1

Exercise 5.2.5 Are the following vectors linearly independent? If not, write one of them as a linear
combination of the others.
     
1 1 1
u =  3 , v =  4 , w =  1 .
1 2 −1

Exercise 5.2.6 Find a linearly independent set of vectors that has the same span as the given vectors.
       
2 1 3 3
u1 =  0  , u2 =  3  , u3 =  3  , u4 =  −3  .
3 5 8 1
 5.2. Linear independence 189

Exercise 5.2.7 Find a linearly independent set of vectors that has the same span as the given vectors.
       
1 2 1 1
 3   6   0   2 
u1 =      
 3  , u2 =  6  , u3 =  −3  , u4 =  1  .
 

1 2 1 1

Exercise 5.2.8 Here are some vectors in R4 .

         
1 1 1 1 1
 1   2   −2   2   −1 
u1 =      
 −1  , u2 =  −1  , u3 =  −1  , u4 =  
 0 , u5 =  
 −1  .
1 1 1 1 1
Explain why these vectors cannot possibly be linearly independent. Then obtain a linearly independent
subset of these vectors that has the same span as these vectors.

Exercise 5.2.9 Here are some vectors in R4 .

         
1 −3 1 2 1
 −1   3   0   −9   0 
u1 =     
 −1  , u2 =  3  , u3 =  −1
,
 u4 =  
 −2  , u5 =  
 0 .
1 −3 1 2 1
Explain why these vectors cannot possibly be linearly independent. Then find a non-trivial linear combi-
nation of these vectors that equals 0.

Exercise 5.2.10 Here are some vectors.

         
1 2 2 5 12
u1 =  1  , u2 =  2  , u3 =  7  , u4 =  7 , u5 =  17  .
−2 −4 −4 −10 −24
Describe the span of these vectors as the span of as few vectors as possible.

Exercise 5.2.11 Here are some vectors.

         
1 1 1 −1 1
u1 =  2  , u2 =  3  , u3 =  −2  , u4 =  0  , u5 =  3  .
−2 −2 −2 2 −1
Describe the span of these vectors as the span of as few vectors as possible.

Exercise 5.2.12 In this exercise, we use scalars from the field Z3 of integers modulo 3 instead of real
numbers (see Section 1.8, “Fields”). Use the extended casting-out algorithm to determine which of the
following vectors are redundant. If there are redundant vectors, write each of them as a linear combination
of previous vectors.
       
2 1 2 0
u1 = 1 , u2 = 1 , u3 = 0 , u4 = 1  .
      
0 2 2 1
190 Spans, linear independence, and bases in Rn

Exercise 5.2.13 Let u, v, w be linearly independent vectors in Rn . Are the vectors u + v, 2u + w, and
w − 2v linearly independent?

Exercise 5.2.14 Let u, v, w be linearly independent vectors in Rn . Are the vectors u + v, u + w, and w + v
linearly independent?

Exercise 5.2.15 Suppose A is an m × n-matrix and {w1 , . . . , wk } is a linearly independent set of vectors
in Rm . Now suppose Azi = wi . Show {z1 , . . . , zk } is also linearly independent.

5.3 Subspaces of Rn

Outcomes
A. Determine whether a subset of Rn is a subspace.

B. Recognize that spans are subspaces of Rn .

C. Recognize that solution sets of homogeneous systems of equations are subspaces of Rn .

As we saw earlier, the span of 0 vectors in Rn is a point, namely the set {0}. The span of one non-zero
vector is a line through the origin, and the span of two linearly independent vectors is a plane through the
origin.

Span of 0 vectors: a point Span of one vector: a line Span of two vectors: a plane

We also call these sets, respectively, a 0-dimensional subspace, a 1-dimensional subspace, and a 2-
dimensional subspace of Rn . The purpose of this section is to generalize this concept of subspace to
arbitrary dimensions.

Definition 5.21: Subspace

A subset V of Rn is called a subspace of Rn if

1. V contains the zero vector of Rn , i.e., 0 ∈ V ;

2. V is closed under addition, i.e., for all u, w ∈ V , we have u + w ∈ V ;

3. V is closed under scalar multiplication, i.e., for all u ∈ V and scalars k, we have k u ∈ V .
 5.3. Subspaces of Rn 191

Notice that the subset V = {0} is a subspace of Rn (called the zero subspace). Every line or plane through
the origin is a subspace. Moreover, the entire space Rn is a subspace of itself. A subspace that is not the
entire space Rn is referred to as a proper subspace of Rn .

Proposition 5.22: Spans are subspaces

Let u1 , . . . , uk be vectors in Rn . Then span {u1 , . . . , uk } is a subspace of Rn .

Proof. Let S = span {u1 , . . . , uk }. To verify that S is a subspace of Rn , we must check that the three
conditions of Definition 5.21 hold.
• We have 0 ∈ S because 0 = 0u1 + . . . + 0uk .
• Suppose u, w ∈ S. By definition of span, there exist scalars a1 , . . . , ak and b1 , . . . , bk such that u =
a1 u1 + . . . + ak uk and w = b1 u1 + . . . + bk uk . Therefore,
u + w = u = (a1 + b1 )u1 + . . . + (ak + bk )uk .
It follows that u + w ∈ S = span {u1 , . . . , uk }, so that S is closed under addition.
• Suppose u ∈ S and t is a scalar. Then by definition of span, there exists scalars a1 , . . . , ak such that
u = a1 u1 + . . . + ak uk . Then
t u = (ta1)u1 + . . . + (tak )uk ,
and thus t u ∈ S. It follows that S is closed under scalar multiplication.
Since S = span {u1 , . . . , uk } satisfies all three conditions, it follows that it is a subspace of Rn . ♠

Example 5.23: A line in R3

In R3 , let L be the line through the origin that is parallel to the vector
 
−5
d =  1 .
−4

Show that L is a subspace of R3 .

Solution. The line L is simply the span of the vector d, i.e., L = span {d}. Therefore, it is a subspace by
Proposition 5.22. ♠

Proposition 5.24: Solution space of a homogeneous system of equations

Consider a homogeneous system of equations Ax = 0, where A is an m × n-matrix. Then the set of
solutions,
V = {x ∈ Rn | Ax = 0} ,
is a subspace of Rn . It is called the solution space of the system.

Proof. To show that V is a subspace of Rn , we check the three conditions of Definition 5.21.
192 Spans, linear independence, and bases in Rn

• We have 0 ∈ V because A0 = 0.

• To show that V is closed under addition, suppose u, w ∈ V . Then by definition of V , Au = 0 and

Aw = 0. Therefore,
A(u + w) = Au + Aw = 0 + 0 = 0.
It follows that u + w ∈ V .

• To show that V is closed under scalar multiplication, suppose u ∈ V and t is a scalar. Then by
definition of V , we have Au = 0. It follows that

A(t u) = t(Au) = t 0 = 0.

Therefore, t u ∈ V .

Since the solution space V = {x ∈ Rn | Ax = 0} satisfies all three conditions, it is a subspace of Rn . ♠

Example 5.25: A plane in R3

Show that the plane 2x + 3y − z = 0 is a subspace of R3 .

Solution. Since 2x + 3y − z = 0 is a homogeneous equation, its solution space

   
 x  
 y   2x + 3y − z = 0
  
z 

is a subspace of R3 by Proposition 5.24. ♠

Example 5.26: Non-examples

Which of the following are subspaces of R3 ?

(a) The line      

x 1 1
 y  =  2 +t  0 .
z 3 1

(b) The plane 2x + 3y − z = 5.

(c) The set of vectors    

 x  
W=   
y  x, y, z ≥ 0 .
 
z 

Solution. None of them are subspaces. Neither the line (a) nor the plane (b) contains the origin 0, so they
fail to satisfy the first condition of subspaces. The set of vectors in (c) contains 0. It is also closed under
 5.3. Subspaces of Rn 193

addition. However, it fails to be closed under scalar multiplication. For example, let u = [1, 1, 1]T . Then
u ∈ W , but (−1)u 6∈ W . ♠

Exercises

Exercise 5.3.1 Which of the following sets are subspaces of R3 ? Explain.

   
 u1  
  
(a) V1 = u = u2  |u1 | ≤ 4 .
 
u3 
   
 u1  
(b) V2 = u =  u2   ui ≥ 0 for each i = 1, 2, 3 .
 
u3 
   
 u1  
(c) V3 = u =  u2   u3 + u1 = 2u2 .
 
u3 
   
 u1  
(d) V4 = u =  u2   u3 ≥ u1 .
 
u3 
   
 u1  
  
(e) V5 = u = u2  u3 = u1 = 0 .
 
u3 

Exercise 5.3.2 Let w ∈ R4 be a given fixed vector. Let

    
 u  


1  

 u2  
M= u=




u3 
 ∈ R 4


w u·= 0



.
u4 

Is M a subspace of R4 ? Explain.

Exercise 5.3.3 Let w, v be given vectors in R4 and define

    
 u1  

  

 u2  
M= u=



 
u3 
4 

· ·
∈ R  w u = 0 and v u = 0 .



u4 

Is M a subspace of R4 ? Explain.

Exercise 5.3.4 In this exercise, we use scalars from the field Z2 of integers modulo 2 instead of real
numbers (see Section 1.8, “Fields”). Which of the following sets are subspaces of (Z2 )3 ?
194 Spans, linear independence, and bases in Rn

       
 1 0 1 0 
(a) V1 =  1  ,  1 , 0 , 0  .
 
0 1 1 0
       
 1 1 1 0 
(b) V2 =  1  ,  1 , 0 , 0  .
 
1 0 0 0
 
 0 
(c) V3 =  0  .
 
0
       
 1 1 0 0 
(d) V3 =  1  ,  0 , 1 , 0  .
 
1 0 0 1

Exercise 5.3.5 Suppose V ,W are subspaces of Rn . Let V ∩W be the set of all vectors that are in both V
and W . Show that V ∩W is also a subspace.

Exercise 5.3.6 Let V be a subset of Rn . Show that V is a subspace if and only if it is non-empty and the
following condition holds: for all u, v ∈ V and all scalars a, b ∈ R,
au + bv ∈ V .

Exercise 5.3.7 Let u1 , . . . , uk be vectors in Rn , and let S = span {u1 , . . ., uk }. Show that S is the smallest
subspace of Rn that contains u1 , . . . , uk . Specifically, this means you have to show: if V is any other
subspace of Rn such that u1 , . . . , uk ∈ V , then S ⊆ V .

5.4 Basis and dimension

Outcomes
A. Find a basis for a subspace of Rn .
B. Use the casting-out algorithm to find a basis for a subspace given as a span.
C. Use basic solutions to find a basis for a subspace given as the solution space of a homogeneous
system of equations.
D. Find the coordinates of a vector with respect to a basis.
E. Find the dimension of a subspace of Rn .
F. Extend a set of linearly independent vectors to a basis.
G. Shrink a spanning set to a basis by removing redundant vectors.
H. Determine whether k vectors form a basis of a k-dimensional space.
 5.4. Basis and dimension 195

5.4.1. Definition of basis

We saw in Proposition 5.22 that spans are subspaces of Rn . Interestingly, the converse is also true: every
subspace of Rn is the span of some finite set of vectors.

Theorem 5.27: Subspaces are spans

Let V be a subspace of Rn . Then there exist linearly independent vectors {u1 , . . . , uk } in V such that

V = span {u1 , . . . , uk } .

Proof. We proceed as follows.

0. If V = {0}, then V is the empty span, and we are done.

1. Otherwise, V contains some non-zero vector. Pick a non-zero vector u1 in V . If V = span {u1 }, we
are done.

2. Otherwise, pick a vector u2 in V that is not in span {u1 }. If V = span {u1 , u2 }, we are done.

3. Otherwise, pick a vector u3 in V that is not in span {u1 , u2 }. If V = span {u1 , u2 , u3 }, we are done.

4. Otherwise, pick a vector u4 in V that is not in span {u1 , u2 , u4 }, and so on.

Continue in this way. Note that after the j th step of this process, the vectors u1 , . . . , u j are linearly inde-
pendent. This is because, by construction, no vector is in the span of the previous vectors, and therefore
no vector is redundant. By Proposition 5.16(3), there can be at most n linearly independent vectors in Rn .
Therefore the process must stop after k steps for some k ≤ n. But then V = span {u1 , . . . , uk }, as desired.
♠
In summary, every subspace of Rn is spanned by a finite, linearly independent collection of vectors. Such
a collection of vectors is called a basis of the subspace.

Definition 5.28: Basis of a subspace

Let V be a subspace of Rn . Then {u1 , . . . , uk } is a basis for V if the following two conditions hold:

1. span {u1 , . . . , uk } = V , and

2. u1 , . . . , uk are linearly independent.

Note that the plural of basis is bases.

196 Spans, linear independence, and bases in Rn

5.4.2. Examples of bases

Proposition 5.29: Standard basis of Rn

Let ei be the vector in Rn whose i th component is 1 and all of whose other components are 0. In
other words, ei is the i th column of the identity matrix.
       
1 0 0 0
 0   1   0   0 
       
       
e1 =  0  , e2 =  0  , e3 =  1  , . . . , en =  0  .
 ..   ..   ..   .. 
 .   .   .   . 
0 0 0 1

Then {e1 , e2 , . . . , en } is a basis for Rn . It is called the standard basis of Rn .

Proof. To see that it is a basis of Rn , first notice that the vectors e1 , e2 , . . . , en span Rn . Indeed, every
vector v = [x1 , . . . , xn ]T ∈ Rn can be written as v = x1 e1 + . . . + xn en . Second, the vectors e1 , e2 , . . . , en are
evidently linearly independent, because none of these vectors can be written as a linear combination of
previous vectors. Since the vectors span Rn and are linearly independent, they form a basis of Rn . ♠

Example 5.30: A non-standard basis of R3

Check that the vectors
     
1 0 −1
u1 =  2  , u2 =  1  , and u3 =  0 
1 0 1

form a basis of R3 .

Solution. We must check that the vectors u1 , u2 , u3 are linearly independent and span R3 . To check linear
independence, we use the casting-out algorithm.
   
1 0 −1 1 0 −1
 2 1 0  ≃  0 1 2 .
1 0 1 0 0 2
Since all columns are pivot columns, there are no redundant vectors, so u1 , u2 , u3 are linearly independent.
To check that they span all of R3 , let w = [x, y, z]T be an arbitrary element of R3 . We must show that w is
a linear combination of u1 , u2 , u3 . This amounts to solving the system of equations
a1 u1 + a2 u2 + a3 u3 = w,
or in augmented matrix form,
   
1 0 −1 x 1 0 −1 x
 2 1 0 y  ≃  0 1 2 y − 2x  .
1 0 1 z 0 0 2 z−x
 5.4. Basis and dimension 197

The system is clearly consistent, so it has a solution, and therefore w is indeed a linear combination of
u1 , u2 , u3 . Since w was an arbitrary vector of R3 , it follows that u1 , u2 , u3 span R3 . ♠
Generalizing the last example, we find that a set of n vectors forms a basis of Rn if and only if the matrix
having those vectors as its columns is invertible. This is the content of the following proposition.

Proposition 5.31: Invertible matrices and bases of Rn

Let A be an n × n-matrix. Then the columns of A form a basis of Rn if and only if A is invertible.

We turn to the question of finding bases for subspaces of Rn .

Example 5.32: Basis of a span

Let          
2 −1 1 3 −1
u1 =  0  , u2 =  0  , u3 =  3  , u4 =  5  , u5 =  1  .
−2 1 5 7 3
Find a basis of span {u1 , . . . , u5 }.

Solution. Let S = span {u1 , . . . , u5 }. By Theorem 5.19, we know that if we remove the redundant vectors
from {u1 , . . . , u5 }, then the remaining vectors will be linearly independent and will still span S. In other
words, the remaining vectors will be a basis for S. We use the casting-out algorithm to identity the
redundant vectors:
     
2 −1 1 3 −1 2 −1 1 3 −1 2 −1 1 3 −1
 0 0 3 5 1  ≃  0 0 3 5 1  ≃  0 0 3 5 1 .
−2 1 5 7 3 0 0 6 10 2 0 0 0 0 0

Since columns 2, 4, and 5 are the non-pivot columns, it follows that the vectors u2 , u4 , and u5 are redun-
dant. Therefore, the desired basis is {u1 , u3 }. ♠

Example 5.33: Basis of the solution space of a homogeneous system of equations

Find a basis for the solution space of the system of equations

x + y − z + 3w − 2v = 0,
x + y + z − 11w + 8v = 0,
4x + 4y − 3z + 5w − 3v = 0.

Solution. We solve the system of equations in the usual way:

     
1 1 −1 3 −2 0 1 1 −1 3 −2 0 1 1 0 −4 3 0
 1 1 1 −11 8 0  ≃  0 0 2 −14 10 0  ≃  0 0 1 −7 5 0  .
4 4 −3 5 −3 0 0 0 1 −7 5 0 0 0 0 0 0 0
198 Spans, linear independence, and bases in Rn

From the reduced echelon form, we see that y, w, and v are free variables. The general solution is:
       
x −3 4 −1
 y   0   0   1 
       
 z  = t  −5  + s  7  + r  0  .
       
 w   0   1   0 
v 1 0 0
Thus, the solution space is spanned by the vectors
     

 −3 4 −1 


 
 0  
  0  
  1 

 −5  ,  7 , 0  .
      

  0   1   0 


 

1 0 0
Moreover, these vectors are evidently linearly independent, because each vector contains a 1 in a position
where all the previous vectors have 0 (and therefore, none of the vectors can be written as a linear com-
bination of previous vectors). It follows that the above three vectors form a basis of the solution space.
♠
Note that the basis vectors of the solution space are exactly what we called the basic solutions in Sec-
tion 1.6.

5.4.3. Bases and coordinate systems

Let V be a subspace of Rn . A basis of V is essentially the same thing as a coordinate system for V .
To see why, let B = {u1 , . . . , uk } be some basis of V . This means that the vectors u1 , . . . , uk are linearly
independent and span V . Because the basis vectors are spanning, every vector v ∈ V can be written as a
linear combination of basis vectors
v = a1 u1 + . . . + ak uk .
Moreover, because the basis vectors are linearly independent, it follows by Theorem 5.18 that the coeffi-
cients a1 , . . . , ak are unique. We say that a1 , . . . , ak are the coordinates of v with respect to the basis B,
and we write  
a1
[v]B =  a2  .
a3

1
2
u3
u2
−2 1
−1
0 u1
1
−1 2
−1
−2
Basis as coordinate system
 5.4. Basis and dimension 199

Example 5.34: Find a vector from its coordinates in a basis

Find the vector v that has coordinates
 
1
[v]B =  −1 
2

with respect to the basis B = {u1 , u2 , u3 } of R3 , where

     
1 0 −1
u1 =  2  , u2 =  1  , and u3 =  0  .
1 0 1

Solution. This simply means that v = 1u1 − 1u2 + 2u3 . We calculate

       
1 0 −1 −1
v = 1 2 −1 1 +2 0  =  1 .
1 0 1 3

♠
In case the basis is the standard basis, the coordinates are just the usual ones, as the following example
illustrates:

Example 5.35: Find a vector from its coordinates in the standard basis
Find the vector v that has coordinates
 
1
[v]B =  −1  ,
2

where B is the standard basis of R3 .

Solution. The standard basis is

     
1 0 0
e1 =  0  , e2 =  1  , and e3 =  0  .
0 0 1

We have to calculate  
1
v = 1e1 − 1e2 + 2e3 =  −1  .
2
We see that the coordinates of any vector with respect to the standard basis are just the usual components
of the vector. ♠
We can also ask to find the coordinates of a given vector in a given basis.
200 Spans, linear independence, and bases in Rn

Example 5.36: Find the coordinates of a vector with respect to a basis

Find the coordinates of the vector 
1
v= 2 
3
with respect to the basis B = {u1 , u2 , u3 } of R3 , where
     
1 0 −1
u1 =  2  , u2 =  1  , and u3 =  0  .
1 0 1

Solution. To find the coordinates, we must solve the system of equations v = a1 u1 + a2 u2 + a3 u3 . We

solve:      
1 0 −1 1 1 0 −1 1 1 0 0 2
 2 1 0 2  ≃  0 1 2 0  ≃  0 1 0 −2  .
1 0 1 3 0 0 2 2 0 0 1 1
Therefore, the unique solution is (a1 , a2 , a3 ) = (2, −2, 1). The coordinates of v with respect to the basis B
are    
a1 2
[v]B =  a2  =  −2  .
a3 1
♠

5.4.4. Dimension

One of the most important properties of bases is that any two bases for the same space must be of the same
size. To show this, we will need the the following fundamental result, called the Exchange Lemma. This
lemma states that spanning sets have at least as many vectors as linearly independent sets.

Lemma 5.37: Exchange Lemma

Suppose u1 , . . . , ur are linearly independent elements of span {v1 , . . . , vs }. Then r ≤ s.

Proof. Since each u j is an element of span {v1 , . . . , vs }, there exist scalars ai j such that
u j = a1 j v1 + . . . + as j vs .
 
Let A = ai j . Note that this matrix has s rows and r columns, i.e., it is an s × r-matrix. Now suppose,
for the sake of obtaining a contradiction, that r > s. Then by Theorem 1.35, the system Ax = 0 has a
non-trivial solution x, i.e., there exists x 6= 0 such that Ax = 0. In other words, for all i = 1, . . ., s,
ai1 x1 + . . . + air xr = 0.
Therefore,
x1 u1 + . . . + xr ur = x1 (a11 v1 + . . . + as1 vs ) + . . . + xr (a1r v1 + . . . + asr vs )
 5.4. Basis and dimension 201

= (a11 x1 + . . . + a1r xr )v1 + . . . + (as1 x1 + . . . + asr xr )vs

= 0 v1 + . . . + 0 vs
= 0.
This contradicts the assumption that u1 , . . . , ur are linearly independent. Since we assumed r > s and
obtained a contradiction, it follows that r ≤ s, as desired. ♠
Armed with the Exchange Lemma, we are now ready to show that any two bases of a space are of the
same size.

Theorem 5.38: Bases are of the same size

Let V be a subspace of Rn , and let B1 and B2 be bases of V . Suppose B1 contains s vectors and B2
contains r vectors. Then s = r.

Proof. This follows right away from the Exchange Lemma. Indeed, observe that B1 = {u1 , . . . , us } is a
spanning set for V while B2 = {v1 , . . . , vr } is linearly independent, so s ≥ r. Similarly B2 = {v1 , . . . , vr } is
a spanning set for V while B1 = {u1 , . . . , us } is linearly independent, so r ≥ s. ♠
Because every basis of V has the same number of vectors, we give this number a special name. It is called
the dimension of V .

Definition 5.39: Dimension of a subspace

Let V be a subspace of Rn . Then the dimension of V , written dim(V ), is defined to be the number
of vectors in a basis.

Example 5.40: Dimension of Rn

What is the dimension of Rn ?

Solution. The standard basis of Rn is {e1 , . . . , en }. Since it has n vectors, so dim(Rn ) = n. ♠

Example 5.41: Dimension of a subspace

Let    
 x  

V =  y  ∈ R3  x − y + 2z = 0 .
  
z
What is the dimension of V ?

Solution. We know that V is a subspace of R3 , because it is the solution space of a system of a homoge-
neous system of equations (in this case, one equation in three variables). We can take y = t and z = s as
the free variables and solve for x = y − 2z = t − 2s. Therefore, a general element of V is of the form
       
x t − 2s 1 −2
 y  =  t  = t  1  +s 0 .
z s 0 1
202 Spans, linear independence, and bases in Rn

Thus,    
 1 −2 
V = span  1  ,  0  .
 
0 1
Since the two spanning vectors are linearly independent, they form a basis of V , and thus dim(V ) = 2.
♠
Note that the dimension of the solution space of a system of equations is equal to the number of parameters
in the general solution, which is equal to the number of free variables. For this reason, the dimension is
also sometimes called the number of degrees of freedom.

Example 5.42: Dimension of a span

Let            

 1 1 8 −6 1 1 

   3   19   −15   3   5 
2
W = span 
 −1
,     
  −1  ,  −8  , 
,  , 
  0   0
 .


 6 
 
1 1 8 −6 1 1
What is the dimension of W ?

Solution. Let
           
1 1 8 −6 1 1
 2   3   19     3   5 
u1 =    , u3 =   , u4 =  −15  , u5 =   , u6 =  ,
 −1  , u2 =  −1   −8   6   0   0 
1 1 8 −6 1 1

so that W = span {u1 , . . . , u6 }. We use the casting-out algorithm to remove any redundant vectors from
u1 , . . . , u6 . The remaining vectors will be linearly independent, and therefore a basis of the span.
   
1 1 8 −6 1 1 1 0 5 −3 0 −2
 2 3 19 −15 3 5   2 
  ≃  0 1 3 −3 0 .
 −1 −1 −8 6 0 0   0 0 0 0 1 1 
1 1 8 −6 1 1 0 0 0 0 0 0

Therefore, the vectors u3 , u4 , and u6 are redundant, and {u1 , u2 , u5 } is a basis of W . It follows that
dim(W ) = 3. ♠

5.4.5. More properties of bases and dimension

We begin by noting that every subspace V of Rn has a basis.

Theorem 5.43: Existence of bases

Every subspace of Rn has a basis.

Proof. This is just a restatement of Theorem 5.27. ♠

 5.4. Basis and dimension 203

Of course, the theorem does not mean that the basis is unique. Usually, a subspace of Rn will have many
different bases. The theorem just states that there exists at least one.
Sometimes, when we are looking for a basis of a space, we may already have a number of linearly
independent vectors. We would like to obtain a basis by adding some additional linearly independent
vectors to the ones we already have. The following lemma guarantees that this can always be done.

Lemma 5.44: Linearly independent set can be extended to a basis

Let V be a subspace of Rn , and let u1 , . . . , uℓ be linearly independent elements of V . Then it is
possible to extend {u1 , . . . , uℓ } to a basis of V . In other words, there exist zero or more vectors
w1 , . . . , ws such that
{u1 , . . . , uℓ , w1 , . . . , ws }
is a basis of V .

Proof. By Theorem 5.43, we know that V has some basis, say {v1 , . . . , vk }. However, this may not be the
basis we are looking for, because maybe it does not contain the vectors u1 , . . . , uℓ . Consider the sequence
of ℓ + k vectors
u1 , . . . , uℓ , v1 , . . . , vk .
Since V is spanned by the vectors v1 , . . . , vk , it is certainly also spanned by the larger set of vectors
u1 , . . . , uℓ , v1 , . . . , vk . From Theorem 5.19, we know that we can obtain a basis of V by removing the re-
dundant vectors from u1 , . . . , uℓ , v1 , . . . , vk . On the other hand, u1 , . . . , uℓ are linearly independent, so none
of them can be redundant. It follows that the resulting basis of V contains all of the vectors u1 , . . . , uℓ . In
other words, we have found a basis of V that is an extension of u1 , . . . , uℓ , which is what had to be shown.
♠

Example 5.45: Extending a linearly independent set to a basis

Extend {u1 , u2 } to a basis of R4 , where
   
1 1
 1   2 
u1 =  
 −1  , u2 =  
 −2  .
2 4

Solution. Let {e1 , . . . , e4 } be the standard basis of R4 . We obtain the desired basis by applying the casting-
out algorithm to u1 , u2 , e1 , e2 , e3 , e4 :
   
1 1 1 0 0 0 1 1 1 0 0 0
 1 2 0 1 0 0   
  ≃  0 1 −1 1 0 0  .
 −1 −2 0 0 1 0   0 0 0 1 1 0 
2 4 0 0 0 1 0 0 0 0 2 1
204 Spans, linear independence, and bases in Rn

Therefore, we cast out the vectors e1 and e4 and keep the rest. The resulting basis is
       

 1 1 0 0  
   2   1   0 
 1       
{u1 , u2 , e2 , e3 } =  , , , .

 −1   −2   0   1  
 
2 4 0 0

Example 5.46: Extending a linearly independent set to a basis

Let   
 x  

  

y 
V=    x + 2y + z − w = 0 .

 z   

 
w 
Note that u1 , u2 ∈ V , where    
1 −2
 −1   1 
u1 =  
 1 , u2 =  
 1 .
0 1
Extend {u1 , u2 } to a basis of V .

Solution. We first find a basis of V by solving the linear equation x + 2y + z − w = 0. Taking y = r, z = s,

and w = t as the free variables, we get x = −2y − z + w = −2r − s +t, and therefore the general solution is
         
x −2r − s + t −2 −1 1
 y   r   1   0   0 
  =   = r     
 z   s   0 +s 1  +t  0 .
w t 0 0 1

Therefore, the vectors {v1 , v2 , v3 } form a basis of V , where

     
−2 −1 1
 1     0 
v1 =   , v2 =  0  , v3 =  
 0   1   0 .
0 0 1

However, this is not the basis we are looking for, because it does not extend {u1 , u2 }. To get a basis of V
that extends {u1 , u2 }, we perform the casting-out algorithm on the vectors u1 , u2 , v1 , v2 , v3 :
   
1 −2 −2 −1 1 1 −2 −2 −1 1
 −1 1 1 0 0   1 1 −1 
  ≃ ... ≃  0 1 .
 1 1 0 1 0   0 0 1 1 −2 
0 1 0 0 1 0 0 0 0 0

Casting out v2 and v3 , we find that the desired basis is {u1 , u2 , v1 }. ♠

 5.4. Basis and dimension 205

We also have a kind of opposite of Lemma 5.44: every spanning set can be shrunk to a basis.

Lemma 5.47: Spanning set can be shrunk to a basis

Let V be a subspace of Rn , and let u1 , . . . , uℓ be a set of vectors spanning V . Then it is possible to
obtain a basis of V by “shrinking” the set, i.e., by removing zero or more vectors from it.

Proof. This is merely a restatement of Theorem 5.19. We obtain the linearly independent subset by remov-
ing the redundant vectors, which can be achieved by the casting-out algorithm. See also Example 5.20.
♠
The following proposition tells us something about the size of a linearly independent set of vectors or the
size of a spanning set of vectors.

Proposition 5.48: Size of a linearly independent or spanning set of vectors

Let V be a k-dimensional subspace of Rn . Then

(a) Every linearly independent set of vectors in V has at most k vectors.

(b) Every spanning set of vectors in V has at least k vectors.

Proof. Both properties follow from the Exchange Lemma (Lemma 5.37). Since V is k-dimensional, it has
some basis consisting of k vectors v1 , . . . , vk .

(a) Suppose u1 , . . . ur are linearly independent vectors in V . Since u1 , . . . ur are linearly independent and
v1 , . . . , vk are spanning, the Exchange Lemma implies that r ≤ k.
(b) Suppose the vectors u1 , . . . us span V . Since v1 , . . . , vk are linearly independent and u1 , . . . us are
spanning, the Exchange Lemma implies that k ≤ s.

♠
The next proposition often comes in handy when we need to check that some set of vectors is a basis for
a subspace V , where the dimension of V is already known. If dim(V ) = k, we know that any basis has to
have size k. Interestingly, to check that a set of k vectors is a basis of V , it is sufficient to check either that
it is linearly independent or that it is spanning. This can save half the work in checking that some set of
vectors is a basis (but it only works if the number of vectors is exactly k, the dimension of V ).

Proposition 5.49: Basis test for k vectors in k-dimensional space

Let V be a k-dimensional subspace of Rn , and consider k vectors u1 , . . . , uk in V .

• If u1 , . . . , uk are linearly independent, then they form a basis for V .

• If u1 , . . . , uk span V , then they form a basis for V .

Proof. The first claim is an easy consequence of Lemma 5.44. Assume that u1 , . . . , uk are linearly inde-
pendent. By Lemma 5.44, we can add zero or more vectors to u1 , . . . , uk to obtain a basis of V . On the
206 Spans, linear independence, and bases in Rn

other hand, since V is k-dimensional, every basis must have exactly k elements, so that the only possibility
is that we have added zero vectors. Therefore, {u1 , . . . , uk } is already a basis of V , as claimed.
To prove the second claim, assume that V = span {u1 , . . . , uk }. By Theorem 5.27, there exists a linearly
independent subset of {u1 , . . . , uk } that spans V , i.e., that is a basis for V . But since dim(V ) = k, every
basis must have exactly k elements, so that the only possible such subset is {u1 , . . . , uk } itself. Therefore,
{u1 , . . . , uk } is a basis of V , as claimed. ♠
It is important to note that Proposition 5.49 does not say that every linearly independent set of vectors in
V is a basis. For example, a set of k − 1 or fewer linearly independent vectors will not be spanning. Also,
the proposition does not say that every spanning set of vectors in V is a basis. For example, a set of k + 1
or more spanning vectors will not be linearly independent. Rather, what the proposition is saying is that if
we have exactly k vectors in a k-dimensional space, then linear independence implies spanning and vice
versa.

Example 5.50: Basis test for 3 vectors in 3-dimensional space

Do the vectors      
1 1 1
u1 =  2  , u2 =  1  , and u3 =  1 
3 1 2
form a basis of R3 ?

Solution. This is similar to Example 5.30. But because we know that R3 is a 3-dimensional space, and
because we have exactly 3 vectors, by Proposition 5.49, we only need to check either whether u1 , u2 , u3
are linearly independent or whether they are spanning. We check whether they are linearly independent
by using the casting-out algorithm.
   
1 1 1 1 1 1
 2 1 1  ≃  0 −1 −1  .
3 1 2 0 0 −1

Since the matrix has rank 3, the vectors u1 , u2 , u3 are linearly independent. Therefore, by Proposition 5.49,
they form a basis of R3 . ♠
The following proposition is also a consequence of Lemma 5.44. It says that smaller subspaces have
smaller dimension.

Proposition 5.51: Subspace of a subspace

Let V and W be subspaces of Rn , and suppose that V ⊆ W . Then dim(V ) ≤ dim(W ), with equality
only when V = W .

Proof. Consider any basis u1 , . . . , uk of V . Because V ⊆ W , the vectors u1 , . . . , uk are linearly independent
elements of W , and therefore can be extended to a basis of W by Lemma 5.44. The resulting basis
of W has at least k elements, i.e., dim(V ) ≤ dim(W ). To prove the last claim, assume moreover that
dim(V ) = dim(W ). In that case, dim(W ) = k, so that the k linearly independent vectors u1 , . . . , uk form a
 5.4. Basis and dimension 207

basis of W by Proposition 5.49. Since both V and W are spanned by u1 , . . . , uk , we must have V = W , as
claimed. ♠

Exercises

Exercise 5.4.1 For each of the following subspaces of R4 , find a basis and determine the dimension.
       

 2 −1 5 −1 
     2   1 
1 0
(a) V1 = span       
 1  ,  −1  ,  3  ,  −2 .


 
 
1 −1 3 −2
       

 0 −1 2 0  
   −1   3   1 
1
(b) V2 = span       
 1  ,  −2  ,  5  ,  2 .


 
 
−1 2 −5 −2
       

 −2 −9 −33 −22  
   4   15   10 
1
(c) V3 = span     
 1  ,  3  ,  12  , 
   .


 8 
 
−3 −9 −36 −24
         

 −1 −4 −3 −1 −7 

  1   3   2   1   5 
(d) V4 = span    ,   ,   ,   ,   .

 −1   −2   −1   −2   −3  
 
−2 −4 −2 −4 −6

Exercise 5.4.2 Find a basis and the dimension of each of the following subspaces of Rn .
  
 4u + v − 5w  

(a) S1 =  12u + 6v − 6w   u, v, w ∈ R .
  
4u + 4v + 4w
  
 2u + 6v + 7w  

  

−3u − 9v − 12w  
(b) S2 =   2u + 6v + 6w  
  u, v, w ∈ R .

  

  
u + 3v + 3w
  
 2u + v  

(c) S3 =   
6v − 3u + 3w  u, v, w ∈ R .
 
3v − 6u + 3w 

Exercise 5.4.3 Find a basis and the dimension of each of the following subspaces of Rn .
208 Spans, linear independence, and bases in Rn

   
 u  

(a) W1 =   

v  u + v = 0 and u − 2w = 0 .
w 
   
 u  

(b) W2 =  v   u + v + w = 0 .
 
w 
   
 u  
   
 

 v  
(c) S =    u + v = w + x and u + w = v + x .

 w  

  
x

Exercise 5.4.4 Find the vector v that has coordinates

 
2
[v]B =  1 
−3

with respect to the basis B = {u1 , u2 , u3 } of R3 , where

     
2 1 −1
u1 =  4  , u2 =  −1  , and u3 =  0  .
1 0 3

Exercise 5.4.5 Find the coordinates of each of v, w with respect to the basis B = {u1 , u2 , u3 }, where
         
4 −1 1 0 2
v = 3 , w = −1 , u1 = 0 , u2 = 1 , u3 = 2  .
        
8 3 3 1 1

Exercise 5.4.6 Extend {u1 , u2 } to a basis of R3 , where

   
3 0
u1 =  3  , u2 =  −1  .
−6 2

Exercise 5.4.7 Let    

 x  

    

y
   x + y + z + 2w = 0 .
V =  z 

  

  
w
Note that u1 , u2 ∈ V , where    
1 −2
 1   −2 
u1 =  
 0 , u2 =  
 −2  .
−1 3
 5.4. Basis and dimension 209

Is {u1 , u2 } a basis of V ? If not, extend it to a basis of V by adding additional basis vectors.

Exercise 5.4.8 Shrink {u1 , u2 , u3 , u4 } to a basis of R4 by removing redundant vectors, where

       
1 0 2 −7
u1 =  2  , u2 =  −1  .u3 =  3  , u4 =  2  .
3 2 8 1

Exercise 5.4.9 Use one of the basis tests of Proposition 5.49 to determine whether the vectors
     
4 −2 1
u1 =  −2  , u2 =  4  , and u3 =  −2 
1 1 4

form a basis of R3 .

Exercise 5.4.10 Use one of the basis tests of Proposition 5.49 to determine whether the vectors
     
2 −1 −1
u1 =  −1  , u2 =  2  , and u3 =  −1 
−1 −1 2

form a basis of R3 .

Exercise 5.4.11 In this exercise, we use scalars from the field Z5 of integers modulo 5 instead of real
numbers (see Section 1.8, “Fields”). Find a basis and the dimension of each of the following subspaces
of (Z5 )n .
       

 2 1 4 2 
        
3   4   1   2 
(a) V1 = span  , , ,
 1   3   2   1 .

 
 
4 2 3 0
  
 u  

(b) V2 =   

v  2u + v = 0 and u + 4w = 0 .
w 
  
 u  

(c) V3 =  v   u + 2v + 3w = 0 .
 
w 

Exercise 5.4.12 In this exercise, we use scalars from the field Z7 of integers modulo 7 instead of real
numbers. Find the coordinates of v with respect to the basis B = {u1 , u2 , u3 } in (Z7 )3 , where
       
3 2 1 4
v = 1 , u1 = 4 , u2 = 3 , u3 = 1  .
      
4 2 3 2
210 Spans, linear independence, and bases in Rn

Exercise 5.4.13 In this exercise, we use scalars from the field Z2 of integers modulo 2 instead of real
numbers. Extend {u1 , u2 } to a basis of (Z2 )4 , where
   
1 0
 0   1 
u1 =    
 1  , u2 =  1  .
1 1

Exercise 5.4.14 True or false? Explain.

(a) Every set of 5 vectors in R5 is linearly independent.

(b) Every set of 4 vectors in R5 is linearly independent.

(c) Every set of 6 vectors in R5 is linearly dependent.

(d) No set of 4 vectors spans R5 .

(e) Every linearly independent set of 5 vectors in R5 is a basis of R5 .

(f) Every linearly independent set of 4 vectors in R5 is a basis of R5 .

(g) Some linearly independent set of 4 vectors in R5 is a basis of R5 .

(h) Every spanning set of 6 vectors in R5 is a basis of R5 .

(i) Every linearly independent set of 4 vectors in R5 spans a 4-dimensional subspace of R5 .

Exercise 5.4.15 If you have 6 vectors in R5 , is it possible they are linearly independent? Explain.

Exercise 5.4.16 Suppose V and W both have dimension equal to 7 and they are subspaces of R10 . What
are the possibilities for the dimension of V ∩ W ? Hint: Remember that a linear independent set can be
extended to form a basis.

5.5 Column space, row space, and null space of a matrix

Outcomes
A. Find a basis for the column space, row space, and null space of a matrix.

B. Find the rank and nullity of a matrix.

There are three important spaces we can associate to a matrix. They are called the column space, row
space, and null space, and are defined as follows.
 5.5. Column space, row space, and null space of a matrix 211

Definition 5.52: Column space, row space, null space

Let A be an m × n-matrix. The column space of A, written col(A), is the span of the columns. The
row space of A, written row(A), is the span of the rows. The null space of A, written null(A), is the
set
null(A) = {x | Ax = 0} .

Note that the column space is a subspace of Rm and the null space is a subspace of Rn . The row space, on
the other hand, is a set of row vectors. It can be regarded as a subspace of Rn , but only if we regard Rn as
the set of n-dimensional row vectors (and not column vectors, as usual).
Before we give an example, recall that two matrices are called row equivalent if one can be obtained
from the other by performing a sequence of elementary row operations. The point of elementary row
operations is that they do not affect the row space or the null space of the matrix. (They do, however,
affect the column space). The following proposition makes this more precise.

Proposition 5.53: Effect of row operations

Let A and B be row equivalent matrices. Then row(A) = row(B) and null(A) = null(B).

Proof. The fact that elementary row operations do not change the null space is a special case of Theo-
rem 1.11, applied to a homogeneous system. To prove that they do not change the row space is also easy;
we just need to look at each kind of elementary row operation. For example, adding a multiple of one row
to another clearly does not change the span of the rows. ♠

Example 5.54: Basis of column space, row space, and null space
Find a basis for the column space, row space, and null space of the matrix
 
1 2 1 3 2
A =  1 3 6 0 2 .
3 7 8 6 6

Solution. The column space of A is the span of the columns of A, i.e.,

         
 1 2 1 3 2 
col(A) = span  
1 ,  3 ,  6 ,  0 ,   2  .
 
3 7 8 6 6

To find a basis for the column space, we use the casting-out algorithm. The reduced echelon form of A is
 
1 0 −9 9 2
 0 1 5 −3 0  . (5.3)
0 0 0 0 0

Note that the first two columns of the reduced echelon form are pivot columns. Therefore, by the casting-
out algorithm, the first two columns of A form a basis for the column space. Thus, the following is a basis
212 Spans, linear independence, and bases in Rn

for the column space:    

 1 2 
Basis of col(A):  1 ,  3  .
 
3 7
The rows space of A is the span of the rows of A, i.e.,
     
row(A) = span 1 2 1 3 2 , 1 3 6 0 2 , 3 7 8 6 6
We could find a basis of the row space by writing all three rows as column vectors and using the casting-
out algorithm. However, there is an easier way. By Proposition 5.53, the row space of A is equal to the row
space of the reduced echelon form (5.3). Moreover, the non-zero rows of the reduced echelon form are
clearly linearly independent (no non-zero row can be a linear combination of other rows below it, because
each non-zero row has a pivot entry). Therefore, the non-zero rows of the reduced echelon form form a
basis of the row space.
   
Basis of row(A): 1 0 −9 9 2 , 0 1 5 −3 0 .
Finally, the null space of A is just the solution space of the homogeneous system Ax = 0. Thus, finding
a basis of the null space is the same as finding a set of basic solutions. From the reduced echelon form,
we can easily find the general solution of Ax = 0, using three parameters r, s, t corresponding to the three
non-pivot columns of (5.3). The general solution is:
     
9 −9 −2
 −5   3   0 
     
x = r    
 1  +s 0  +t  0 .

 0   1   0 
0 0 1
Therefore, the following is a basis of the null space:
     

 9 −9 −2 


      

 −5   3   0 
Basis of null(A):  1 ,  0 ,  0  .
      
 0
   1   0 


 

0 0 1
♠
As the example shows, all three bases, for the column space, the row space, and the null space of A, can
be easily determined from the reduced echelon form. In the next proposition, we use this information to
determine the dimensions of these three spaces. Recall from Definition 1.24 that the rank of a matrix is
equal to the number of pivot entries of its reduced echelon form.

Proposition 5.55: Dimension of column space, row space, and null space
Let A be an m × n-matrix. Then the dimensions of the column space, row space, and null space of
A are as follows:
dim(col(A)) = rank(A),
dim(row(A)) = rank(A),
dim(null(A)) = n − rank(A).
 5.5. Column space, row space, and null space of a matrix 213

Proof. Let r = rank(A). Following the same method as in Example 5.54, we can use the casting-out
algorithm to find a basis for the column space. Since the reduced echelon form of A has r pivot columns,
the basis has r elements, and therefore dim(col(A)) = r. Also, the reduced echelon form has r non-zero
rows (since each non-zero row contains exactly one pivot entry). These form a basis of the row space, and
therefore dim(row(A)) = r. Finally, the dimension of the null space is equal to the number of parameters
in the general solution of the system of equations Ax = 0. There is one parameter for each non-pivot
column, and since A has n columns and r pivot columns, it follows that dim(null(A)) = n − r. ♠
Among other things, the proposition states that the “row rank” of a matrix (the dimension of its row space)
is always equal to the “column rank” (the dimension of the column space). This fact is not at all obvious
when one first considers the definition of a matrix. It is often called the rank theorem and is one of
the deep and mysterious facts of linear algebra. It means, for example, that if we do elementary column
operations instead of elementary row operations, we end up with exactly the same number of pivots. Since
the “row rank” and “column rank” are always equal, we are justified in simply calling this quantity the
“rank” of the matrix.
There is also a name for the dimension of the null space. It is called the nullity of the matrix, and is
written nullity(A). The last part of Proposition 5.55 is also called the rank-nullity theorem, and is often
written in the form
rank(A) + nullity(A) = n.

Example 5.56: Rank and nullity

Find the rank and nullity of the matrix
 
1 2 1 3 2
 1 3 6 0 2 
A=
 1
.
2 1 3 2 
1 3 2 4 0

Solution. The reduced echelon form of A is

 
13
1 0 0 0 2
 
 0 1 0 2 − 25 
 
 ,
 0 0 1 −1 1 
 2 
0 0 0 0 0

and so rank(A) = 3 and nullity(A) = 5 − 3 = 2. ♠

We conclude this section with two useful theorems about matrices.

214 Spans, linear independence, and bases in Rn

Theorem 5.57:
The following are equivalent for an m × n-matrix A.

1. rank(A) = n.

2. row(A) = Rn , i.e., the rows of A span Rn .

3. The columns of A are linearly independent in Rm .

4. The n × n-matrix AT A is invertible.

5. A is left invertible, i.e., there exists an n × m-matrix B such that BA = I .

6. The system Ax = 0 has only the trivial solution.

Theorem 5.58:
The following are equivalent for an m × n-matrix A.

1. rank(A) = m.

2. col(A) = Rm , i.e., the columns of A span Rm .

3. The rows of A are linearly independent in Rn .

4. The m × m-matrix AAT is invertible.

5. A is right invertible, i.e., there exists an n × m-matrix B such that AB = I .

6. The system Ax = b is consistent for every b ∈ Rm .

Exercises

Exercise 5.5.1 Determine the rank and nullity and find a basis of the column space, row space, and null
space of each of the following matrices.

(a)  
1 3 2
A= 3 9 6 
1 3 2

(b)  
1 3 0 2
B= 3 9 1 7 
1 3 1 3
 5.5. Column space, row space, and null space of a matrix 215

(c)  
1 0 3
 3 1 10 
C=
 1

1 4 
1 −1 2

(d)  
0 0 −1 0 1
 1 2 3 −2 −18 
D=
 1

2 2 −1 −11 
−1 −2 −2 1 11

(e)  
1 0 3 0
 3 1 10 0 
E=
 −1

1 −2 1 
1 −1 2 −2

Exercise 5.5.2 Find null(A) for the following matrices.

(a)  
2 3
A=
4 6

(b)  
1 0 −1
A =  −1 1 3 
3 2 1

(c)  
2 4 0
A =  3 6 −2 
1 2 −2

(d)  
2 −1 3 5
 2 0 1 2 
A=
 6

4 −5 −6 
0 2 −4 −6

Exercise 5.5.3 In this exercise, we use scalars from the field Z5 of integers modulo 5 instead of real
numbers (see Section 1.8, “Fields”). Determine the rank and nullity and find a basis of the column space,
row space, and null space of the following matrix over Z5 .
 
2 4 1
A= 1 2 3 
3 1 2
216 Spans, linear independence, and bases in Rn

Exercise 5.5.4 Show that if A is an m × n-matrix, then null(A) is a subspace of Rn .

Exercise 5.5.5 Let A be an m × n-matrix. Show that col(A) = {Au | u ∈ Rn }.

Exercise 5.5.6 Show that rank(A) = rank(AT ).

Exercise 5.5.7 For invertible matrices B and C of appropriate size, show that rank(A) = rank(BA) =
rank(AC).

Exercise 5.5.8 Suppose A is an m × n-matrix and B is an n × p-matrix. Show that

nullity(AB) ≤ nullity(A) + nullity(B).

Hint: Consider the subspace col(B) ∩ null(A) and suppose a basis for this subspace is {w1 , . . . , wk }.
Let {z1 , . . . , zk } be such that Bzi = wi . Now suppose {u1 , . . . , ur } is a basis for null(B), and argue that
null(AB) ⊆ span {u1 , . . . , ur , z1 , . . . , zk }.
6. Linear transformations in Rn

6.1 Linear transformations

Outcomes
A. Determine whether a vector function T : Rn → Rm is a linear transformation.

In calculus, a function (or map) f : R → R is a rule that maps a real number x ∈ R to a real number
f (x) ∈ R. In linear algebra, we can generalize this concept to vectors. A vector function T : Rn → Rm is
a rule that inputs an n-dimensional vector v ∈ Rn and outputs an m-dimensional vector T (v) ∈ Rm . The
following are some examples of vector functions:
 2       
  x x x+y x  x+z 
x            e
T1 = x + y , T2 y = x + y + z , T3 y = √ . (6.1)
y 2 y
y z 0 z

Of these, the first is a function T1 : R2 → R3 , the second is a function T2 : R3 → R3 , and the third is a
function T3 : R3 → R2 . We can evaluate a vector function by applying it to a vector, for example,
 2     2   
  1 1   0 0
1     0
T1 = 1 + 2 = 3 , T1 = 1+1 = 1 ,
  
2 1
22 4 12 1

and so on. The study of arbitrary vector functions and their derivatives and integrals is the subject of
multivariable calculus. In linear algebra, we will only be concerned with linear vector functions, which
are also called linear transformations or linear maps. They are defined as follows.

Definition 6.1: Linear transformation

A vector function T : Rn → Rm is called a linear transformation, or simply linear, if it satisfies
the following two conditions:

1. T preserves addition, i.e., for all v, w ∈ Rn , we have T (v + w) = T (v) + T (w);

2. T preserves scalar multiplication, i.e, for all v ∈ Rn and scalars k, we have T (kv) = kT (v).

217
218 Linear transformations in Rn

Example 6.2: Linear and non-linear transformations

Which of the vector functions in (6.1) are linear transformations?

Solution.
 
1
(a) The function T1 is not a linear transformation. For example, let v = . Then
0
     
  1   4 1
1 2
T1 (v) = T1 =  1  and T1 (2v) = =  2  = 6 2 1 .
0 0
0 0 0

Since T1 (2v) 6= 2T1 (v), the vector function T1 does not preserve scalar multiplication, and therefore
it is not a linear transformation.

(b) The function T2 is a linear transformation. For example, to prove that T2 preserves addition, consider
two arbitrary vectors    
x1 x2
v =  y1  and w =  y2  .
z1 z2
We have
   
x1 + x2 (x1 + x2 ) + (y1 + y2 )
T2 (v + w) = T2  y1 + y2  =  (x1 + x2 ) + (y1 + y2 ) + (z1 + z2 ) 
z1 + z2 0

and
     
x1 + y1 x2 + y2 (x1 + y1 ) + (x2 + y2 )
T2 (v) + T2 (w) =  x1 + y1 + z1  +  x2 + y2 + z2  =  (x1 + y1 + z1 ) + (x2 + y2 + z2 )  .
0 0 0

Since the two sides are evidently equal, T2 preserves addition. The fact that it preserves scalar
multiplication can be shown by a similar calculation.
   
0 1
(c) The function T3 is not a linear transformation. For example, consider v = 1 and w = 1 .
  
0 0
Then  
1  
e
T3 (v + w) = T3  2  = √ ,
2
0
and      
1 e e+1
T3 (v) + T3 (w) = + = .
1 1 2
Since T3 (v + w) 6= T3 (v) + T3 (w), the vector function T3 does not preserve addition, and therefore it
is not linear.
 6.1. Linear transformations 219

♠
An easy fact about linear transformation is that they preserve the origin, i.e., they satisfy T (0) = 0. This
can be seen, for example, by considering T (0) = T (0 + 0) = T (0) + T (0) and then subtracting T (0) from
both sides of the equation. This gives an easier way to see that T3 in the above example is not a linear
transformation, since T3 (0) 6= 0. On the other hand, of course not every function that preserves the origin
is linear. For example, T1 is not linear although it satisfies T1 (0) = 0.
The following characterization of linearity is often useful, as it permits us to check just one property
instead of two.

Proposition 6.3: Alternative characterization of linear transformations

A vector function T : Rn → Rm is linear if and only if it satisfies the following condition, for all
v, w ∈ Rn and scalars a, b:
T (av + bw) = aT (v) + bT (w).

Proof. First, assume that T is linear. Then from preservation of addition and scalar multiplication, we
have T (av + bw) = T (av) + T (bw) = aT (v) + bT (w). Conversely, assume that T satisfies T (av + bw) =
aT (v) + bT (w) for all vectors v, w and scalars a, b. Then we get preservation of addition by setting
a = b = 1, and preservation of scalar multiplication by setting b = 0. ♠

Exercises

Exercise 6.1.1 Which of the following vector functions are linear transformations?
       
  2x + y x x + y2 x  
x             0
T1 = x − 2y , T2 y = (x + y)z , T3 y = .
y 0
−x − y z 0 z

Exercise 6.1.2 Consider the following functions T : R3 → R2 . Explain why each of these functions T is
not linear.
 
x  
  x + 2y + 3z + 1
(a) T y =
2y − 3x + z
z
 
x  2 + 3z

x + 2y
(b) T  y  =
2y + 3x + z
z
 
x  
  sin x + 2y + 3z
(c) T y =
2y + 3x + z
z
220 Linear transformations in Rn

 
x  
x + 2y + 3z
(d) T  y  =
2y + 3x − ln z
z

Exercise 6.1.3 Let A be an m × n-matrix. Show the vector function T : Rn → Rm defined by T (v) = Av is
a linear transformation.

Exercise 6.1.4 Let u ∈ Rn be a fixed vector. Show that the function T defined by T (v) = v − proju (v) is a
linear transformation.

Exercise 6.1.5 Let u ∈ Rn be a fixed non-zero vector. The function T defined by T (v) = u + v has the
effect of translating all vectors by adding u. Show this is not a linear transformation.

6.2 The matrix of a linear transformation

Outcomes
A. Find the matrix corresponding to a linear transformation T : Rn → Rm .

An important example of linear transformations are the so-called matrix transformations.

Proposition 6.4: Matrix transformations are linear transformations

Let A be an m × n-matrix, and consider the vector function T : Rn → Rm defined by T (v) = Av.
Then T is a linear transformation.

Proof. This follows from the laws of matrix multiplication. Namely, by the distributive law, we have
A(v+w) = Av +Aw, showing that T preserves addition. And by the compatibility of matrix multiplication
and scalar multiplication, we have A(kv) = k(Av), showing that T preserves scalar multiplication. ♠
In fact, matrix transformations are not just an example of linear transformations, but they are essentially
the only example. One of the central theorems in linear algebra is that all linear transformations T : Rn →
Rm are in fact matrix transformations. Therefore, a matrix can be regarded as a notation for a linear
transformation, and vice versa. This is the subject of the following theorem.

Theorem 6.5: Linear transformations are matrix transformations

Let T : Rn → Rm be any linear transformation. Then there exists an m × n-matrix A such that for all
v ∈ Rn ,
T (v) = Av.
In other words, T is a matrix transformation.
 6.2. The matrix of a linear transformation 221

Proof. Suppose T : Rn → Rm is a linear transformation and consider the standard basis {e1 , . . . , en } of Rn .
For all i, define ui = T (ei ), and let A be the matrix that has u1 , . . . , un as its columns. We claim that A is
the desired matrix, i.e., that T (v) = Av holds for all v ∈ Rn .
To see this, let  
x1
 
v =  ... 
xn
be some arbitrary element of Rn . Then v = x1 e1 + . . . + xn en , and we have:

T (v) = T (x1 e1 + . . . + xn en )
= T (x1 e1 ) + . . . + T (xn en ) by linearity
= x1 T (e1 ) + . . . + xn T (en ) by linearity
= x1 u1 + . . . + xn un by definition of ui
= Av by the column method of matrix multiplication.

In summary, the matrix corresponding to the linear transformation T has as its columns the vectors
T (e1 ), . . . , T (en ), i.e., the images of the standard basis vectors. We can visualize this matrix as follows:
 
| |
A =  T (e1 ) · · · T (en )  .
| |

Example 6.6: The matrix of a linear transformation

Suppose T : R3 → R2 is a linear transformation where
     
1   0   0  
   1    9   1
T 0 = , T 1 = , and T 0 = .
2 −3 1
0 0 1

Find the matrix A such that T (v) = Av for all v.

Solution. By Theorem 6.5, the columns of A are T (e1 ), T (e2 ), and T (e3 ). Therefore,
 
1 9 1
A= .
2 −3 1

♠
222 Linear transformations in Rn

Example 6.7: The matrix of a linear transformation

Let T : R3 → R2 be the linear transformation defined by
 
x  
x + y
T  y  = ,
x + 2y − z
z

for all x, y, z ∈ R. Find the matrix of this linear transformation.

Solution. We compute the images of the standard basis vectors:

 
1  
    1
T (e1 ) = T 0 = ,
1
0
 
0  
    1
T (e2 ) = T 1 = ,
2
0
 
0  
0
T (e3 ) = T  0  = .
−1
1
The matrix A has T (e1 ), T (e2 ), and T (e3 ) as its columns. Therefore,
 
1 1 0
A= .
1 2 −1
♠

Example 6.8: Matrix of a projection map

 
1
Let u = 2  and let T : R3 → R3 be the projection map defined by

3

T (v) = proju (v)

for all v ∈ R3 .

(a) Is T a linear transformation?

(b) If yes, find the matrix of T .

Solution.

(a) Recall the formula for the projection of v onto u:

proju (v) = ·
u v
u.
·
u u
 6.2. The matrix of a linear transformation 223

In the situation we are interested in, u is a fixed vector, and v is the input to the function T . Given
any two vectors v, w, and using the distributive laws of the dot product and scalar multiplication, we
have:
 
proju (v + w) = ·
u (v + w)
u= ·
u v u w
+ · u= ·
u v
u+ ·
u w
u = proju (v) + proju (w).
u u· ·
u u u u · ·
u u ·
u u

Therefore, the function T (v) = proju (v) preserves addition. Also, given any scalar k, we have
   
proju (kv) = ·
u (kv)
u= k
u v· u=k
u v ·
u = k proju (v).
u u · u u· u u ·
Therefore, the function T preserves scalar multiplication. It follows that T is a linear transformation.

(b) To find the matrix of T , we must compute the images of the standard basis vectors T (e1 ), . . . , T (e3 ).
We compute
 
1
T (e1 ) = proju (e1 ) =
u e1 ·
u =
1  
2 ,
u u · 14
3
 
1
T (e2 ) = proju (e2 ) =
u e2 ·
u =
2  
2 ,
u u · 14
3
 
1
T (e3 ) = proju (e3 ) =
u e3 ·
u =
3  
2 .
u u · 14
3

Hence the matrix of T is  

1 2 3
1 
A= 2 4 6 .
14
3 6 9

Exercises

Exercise 6.2.1 For each of the following vector functions T : Rn → Rn , show that T is a linear transfor-
mation and find the corresponding matrix A such that T (x) = Ax.

(a) T multiplies the j th component of v by a non-zero number b.

(b) T adds b times the j th component of v to the i th component.

(c) T switches the i th and j th components.

224 Linear transformations in Rn

Exercise 6.2.2 Assume that T : Rn → Rm is a linear transformation, and that u1 , . . . , un is a basis of Rn .

For all i = 1, . . . , n, let vi = T (ui ). Let A be the matrix that has u1 , . . . , un as its columns, and let B be the
matrix that has v1 , . . . , vn as its columns. Show that A is invertible and the matrix of T is BA−1 .

Exercise 6.2.3 Suppose T is a linear transformation such that

           
1 5 −1 1 0 5
T  2  =  1  , T  −1  =  1  , and T  −1  =  3  .
−6 3 5 5 2 −2

Find the matrix of T . Hint: use Exercise 6.2.2.

Exercise 6.2.4 Suppose T is a linear transformation such that

           
1 1 −1 2 0 6
T  1  =  3  , T  0  =  4  , and T  −1  =  1  .
−8 1 6 1 3 −1

Find the matrix of T . Hint: use Exercise 6.2.2.

Exercise 6.2.5 Consider the following linear transformations T : R3 → R2 . For each, determine the
matrix A such that T (x) = Ax.
 
x  
x + 2y + 3z
(a) T y =
2y − 3x + z
z
 
x  
  7x + 2y + z
(b) T y =
3x − 11y + 2z
z
 
x  
3x + 2y + z
(c) T y =
x + 2y + 6z
z
 
x  
  2y − 5x + z
(d) T y =
x+y+z
z

Exercise 6.2.6 Find the matrix for T (w) = projv (w), where v = [1, −2, 3]T .

Exercise 6.2.7 Find the matrix for T (w) = projv (w), where v = [1, 5, 3]T .
 6.3. Geometric interpretation of linear transformations 225

6.3 Geometric interpretation of linear transformations

Outcomes

A. Find the matrix of rotations, reflections, scalings, and shearings in R2 and R3 .

In this section, we will examine some special examples of linear transformations in R2 and R3 including
rotations and reflections.

Example 6.9: Rotation by 90◦ in R2

Consider the linear transformation T : R2 → R2 that is given by a counterclockwise rotation by 90
degrees. Find the matrix A corresponding to this linear transformation. Find a formula for T .

Solution. To visualize a vector function on R2 , it is often useful to consider a pair of before-and-after

pictures such as the following:

h i
e2 T (e1 ) = 01

T
e1 7−→ h i
T (e2 ) = −1
0

“before” “after”

The picture illustrates how the function T rotates the entire plane (including the pink letter “F”) by 90
degrees counterclockwise. The picture also illustrates that when we apply the rotation T to the first and
second standard basis vectors e1 and e2 , we obtain the vectors
   
0 −1
T (e1 ) = and T (e2 ) = .
1 0

The matrix of T has these vectors as its columns. Therefore, the matrix of T is
 
0 −1
A= .
1 0
226 Linear transformations in Rn

Finally, we can use this to find a formula for the counterclockwise 90 degree rotation T :
      
x 0 −1 x −y
T = = .
y 1 0 y x

To illustrate how this works, consider the top right corner of the letter “F”. It has the coordinates (0.6, 1).
Applying the function T to the coordinate vector, we get
      
0.6 0 −1 0.6 −1
T = = .
1 1 0 1 0.6

These are precisely the coordinates of the corresponding point on the letter “F” after the rotation. ♠

Example 6.10: Reflection about the y-axis in R2

Let T : R2 → R2 be a reflection about the y-axis. Find the matrix A corresponding to this linear
transformation, and a formula for T .

Solution. The before-and-after picture for a reflection about the y-axis looks like this:

h i
e2 T (e2 ) = 01

T
e1 7−→ h i
T (e1 ) = −1
0

We see that    
−1 0
T (e1 ) = −e1 = and T (e2 ) = e2 = .
0 1
Therefore, the matrix of T is  
−1 0
A= .
0 1
The formula for a reflection about the y-axis is:
      
x −1 0 x −x
T = = .
y 0 1 y y

♠
 6.3. Geometric interpretation of linear transformations 227

Example 6.11: Rotation by an arbitrary angle in R2

Find the matrix A for a counterclockwise rotation by angle θ in R2 .

Solution. The before-and-after picture is as follows:

h i
sin θ
T (e2 ) = −cos θ
h i
e2 T (e1 ) = cos θ
sin θ

T θ
e1 7−→

Thus the matrix of T is  

cos θ − sin θ
A= .
sin θ cos θ
♠

Example 6.12: More linear transformations of the plane

Describe the linear transformation that is given by each of the following matrices. Draw a before-
and-after picture for each.
     1   
0 1 2 0 2 0 1 1
(a) A = , (b) B = , (c) C = , (d) D = .
1 0 0 2 0 2 0 1

Solution. To draw each before-and-after picture, we can start by drawing the images of the two standard
basis vectors e1 and e2 , which are the columns of the transformation matrix. We have also drawn the image
of the letter “F”, to better illustrate the effect of each transformation.
228 Linear transformations in Rn

(a) (b)
T (e2 )
e2 T (e1 ) e2
A B
e1 7−→ e1 7−→
T (e2 ) T (e1 )

(c) (d) T (e2 )

T (e2 )
e2 e2
C D
e1 7−→ e1 7−→
T (e1 ) T (e1 )

The transformation A is a reflection about the line x = y. The transformation B is a scaling by a factor of
2. The transformation C is also a scaling, but by a different factor in the x- and y-directions. It scales the
x-direction by a factor of 12 (or equivalently, shrinks it by a factor of 2), and scales the y-direction by a
factor of 2. The transformation D is called a shearing. It keeps one line (the x-axis) fixed, while shifting
all other points by varying distances along lines that are parallel to the x-axis. ♠

Example 6.13: Rotation in R3

Find the matrix of a rotation by angle θ about the z-axis in 3-dimensional space, counterclockwise
when viewed from above.

Solution. Here is the before-and-after picture. A rotation in 3-dimensional space is usually harder to
visualize than in the plane, but fortunately, the rotation is about the z-axis, so all the “action” is taking
place in the xy-plane.

   
0 − sin θ
e3 T (e3 ) = 0 T (e2 ) = cos θ
1 0
e2
T
7−→
e1 θ
 
cos θ
T (e1 ) = sin θ
0
 6.3. Geometric interpretation of linear transformations 229

Therefore, the matrix of the rotation is

 
cos θ − sin θ 0
A =  sin θ cos θ 0  .
0 0 1
♠

Exercises

Exercise 6.3.1 Find the matrix for the linear transformation that rotates every vector in R2 by an angle
of π /3.

Exercise 6.3.2 Find the matrix for the linear transformation that reflects every vector in R2 about the
x-axis.

Exercise 6.3.3 Find the matrix for the linear transformation that reflects every vector in R2 about the line
y = −x.

Exercise 6.3.4 Find the matrix for the linear transformation that stretches R2 by a factor of 3 in the
vertical direction.

Exercise 6.3.5 Find the matrix of the linear transformation that reflects every vector in R3 about the
xy-plane.

Exercise 6.3.6 Find the matrix of the linear transformation that reflects every vector in R3 about plane
x = z.

Exercise 6.3.7 Describe the linear transformation that is given by each of the following matrices. Draw
a before-and-after picture for each.
       
1 −1 0 2 1 0 1 0
(a) A = , (b) B = , (c) C = , (d) D = .
1 1 1 0 −1 1 0 0

 
a
Exercise 6.3.8 Let u = be a unit vector in R2 . Find the matrix that reflects all vectors about this
b
vector, as shown in the following picture.

u
230 Linear transformations in Rn

6.4 Properties of linear transformations

Outcomes
A. Use properties of linear transformations to solve problems.

B. Find the composite of transformations and the inverse of a transformation.

We begin by noting that linear transformations preserve the zero vector, negation, and linear combinations.

Proposition 6.14: Properties of linear transformations

Let T : Rn → Rm be a linear transformation. Then

• T preserves the zero vector: T (0) = 0.

• T preserves negation: T (−v) = −T (v).

• T preserves linear combinations:

T (a1 v1 + a2 v2 + . . . + ak vk ) = a1 T (v1 ) + a2 T (v2 ) + . . . + ak T (vk ).

Example 6.15: Linear combination

Let T : R3 → R4 be a linear transformation such that
   
  4   4
1  4  4  5 
T  3  =  
 0  and T
 0  = 
 −1
.

1 5
−2 5
 
−7
Find T  3 .
−9

   
−7 −7
Solution. Using the third property in Proposition 6.14, we can find T  3  by writing  3  as
−9 −9
   
1 4
a linear combination of  3  and  0 . By solving the appropriate system of equations, we find that
1 5
     
−7 1 4
 3  =  3 −2 0 .
−9 1 5
 6.4. Properties of linear transformations 231

Therefore,
     
−7 1 4
T  3  = T  3  − 2  0 
−9 1 5
     
    4 4 −4
1 4  4   5   

= T 3  − 2T  0  =     =  −6  .
 0  − 2  −1   2 
1 5
−2 5 −12

♠
Suppose that we first apply a linear transformation T to a vector, and then the linear transformation S to
the result. The resulting two-step transformation is also a linear transformation, called the composition of
T and S.

Definition 6.16: Composition of linear transformations

Let S : Rk → Rn and T : Rn → Rm be linear transformations. Then the composition of S and T (also
called the composite transformation of S and T ) is the linear transformation

T ◦ S : Rk → Rm

that is defined by
(T ◦ S)(v) = T (S(v)),
for all v ∈ Rk .

Notice that the resulting vector will be in Rm . Be careful to observe the order of transformations. The
composite transformation T ◦ S means that we are first applying S, and then T . Composition of linear
transformations is written from right to left. The composition T ◦ S is sometimes pronounced “T after S”.

Theorem 6.17: Matrix of a composite transformation

Let S : Rk → Rn and T : Rn → Rm be linear transformations. Let A be the matrix corresponding
to S, and let B be the matrix corresponding to T . Then the matrix corresponding to the composite
linear transformation T ◦ S is BA.

Proof. For all v ∈ Rk , we have

(T ◦ S)(v) = T (S(v)) = B(Av) = (BA)v.

Therefore, BA is the matrix corresponding to T ◦ S. ♠

Example 6.18: Two rotations

Find the matrix for a counterclockwise rotation by angle θ + φ in two different ways, and compare.
232 Linear transformations in Rn

Solution. Let Aθ be the matrix of a rotation by θ , and let Aφ be the matrix of a rotation by angle φ . We
calculated these matrices in Example 6.11. Then a rotation by the angle θ + φ is given by the product of
these two matrices:
  
cos θ − sin θ cos φ − sin φ
Aθ Aφ =
sin θ cos θ sin φ cos φ
 
cos θ cos φ − sin θ sin φ − cos θ sin φ − sin θ cos φ
= .
sin θ cos φ + cos θ sin φ cos θ cos φ − sin θ sin φ

On the other hand, we can compute the matrix for a rotation by angle θ + φ directly:
 
cos(θ + φ ) − sin(θ + φ )
Aθ +φ = .
sin(θ + φ ) cos(θ + φ )

The fact that these matrices are equal amounts to the well-known trigonometric identities for the sum of
two angles, which we have here derived using linear algebra concepts:

sin(θ + φ ) = sin θ cos φ + cos θ sin φ ,

cos(θ + φ ) = cos θ cos φ − sin θ sin φ .

Example 6.19: Multiple rotations in R3

Find the matrix of the linear transformation T : R3 → R3 that is given as follows: a rotation by 30
degrees about the z-axis, followed by a rotation by 45 degrees about the x-axis.

Solution. It would be quite difficult to picture the transformation T in one step. Fortunately, we don’t have
to do this. All we have to do is find the matrix for each rotation separately, then multiply the two matrices.
We have the be careful to multiply the matrices in the correct order.
Let B be the matrix for a 30-degree rotation about the z-axis. It is given exactly as in Example 6.13:
   √ 
3 1
cos 30◦ − sin 30◦ 0 − 0
   2 √2 

B =  sin 30 ◦ ◦ 0   1 3
0 
cos 30 = 2 2 .
0 0 1 0 0 1

Let C be the matrix for a 45-degree rotation about the x-axis. It is analogous to Example 6.13, except that
the rotation takes place in the yz-plane instead of the xy-plane.
   
1 0 0 1 0 0
   


C=  0 cos 45 ◦ − sin 45◦  =  0 √1
  − √1  .
2 2 
0 sin 45◦ cos 45◦ 0 √1 √1
2 2

Finally, to apply the linear transformation T to a vector v, we must first apply B and then C. This means
that T (v) = C(Bv). Therefore, the matrix corresponding to T is CB. Note that it is important that we
 6.4. Properties of linear transformations 233

multiply the matrices corresponding to each subsequent rotation from right to left.
  √   √ 
3
1 0 0 3 1 − 21 0
  2 − √ 2 0
  2 √ 
   √ 
A = CB =  0 √1 − √1   1 3
0  = 
1 √3 − √1 .
 2 2  2 2  2 2 2√ 2 2 
1 1 1 3 1
0 √ √ 0 0 1 √ √ √
2 2 2 2 2 2 2

♠
We can also consider the inverse of a linear transformation. The inverse of T , if it exists, is a linear
transformation that undoes the effect of T .

Definition 6.20: Inverse of a transformation

Let T , S : Rn → Rn be linear transformations. Suppose that for each v ∈ Rn ,

(S ◦ T )(v) = v

and
(T ◦ S)(v) = v.
Then S is called the inverse of T , and we write S = T −1 .

Example 6.21: Inverse of a transformation

What is the inverse of a counterclockwise rotation by the angle θ in R2 ?

Solution. The inverse is a clockwise rotation by the same angle. ♠

It is perhaps not entirely unexpected that the matrix of T −1 is exactly the inverse of the matrix of T , if it
exists.

Theorem 6.22: Matrix of the inverse transformation

Let T : Rn → Rn be a linear transformation and let A be the corresponding n × n-matrix. Then T
has an inverse if and only if the matrix A is invertible. In this case, the matrix of T −1 is A−1 .

Example 6.23: Matrix of the inverse transformation

Find the inverse of the linear transformation T : R2 → R2 given by
   
x 2x + y
T = .
y 7x + 4y

Solution. The easiest way to do this is to find the matrix of T . We have

       
1 2 0 1
T = and T = .
0 7 1 4
234 Linear transformations in Rn

Therefore, the matrix of T is  

2 1
A= .
7 4
The inverse of A is  
−1 4 −1
A = .
−7 2
Therefore, T −1 is the linear transformation defined by
      
−1 x 4 −1 x 4x − y
T = = .
y −7 2 y −7x + 2y

Exercises

Exercise 6.4.1 Find the matrix for the linear transformation that reflects every vector in R2 about the
x-axis and then reflects about the y-axis.

Exercise 6.4.2 Find the matrix for the linear transformation that rotates every vector in R2 by an angle
of 2π /3 and then reflects about the x-axis.

Exercise 6.4.3 Find the matrix for the linear transformation that rotates every vector in R2 by an angle
of π /6 and then reflects about the x-axis followed by a reflection about the y-axis.

Exercise 6.4.4 Find the matrix for the linear transformation that reflects every vector in R2 about the
x-axis and then rotates by an angle of π /4.

Exercise 6.4.5 Find the matrix of the linear transformation that rotates every vector in R3 counterclock-
wise about the z-axis when viewed from the positive z-axis by an angle of 30 degrees and then reflects
about the xy-plane.

Exercise 6.4.6 Prove the three properties in Proposition 6.14, using only the definition of a linear trans-
formation (i.e., the fact that it preserves addition and scalar multiplication).
 
3 1
Exercise 6.4.7 Let T be the linear transformation with matrix A = and S the linear transfor-
  −1 2
 
0 −2 2
mation with matrix B = . Find the matrix of S ◦ T . Compute (S ◦ T )(v) for v = .
4 2 −1
   
1 2
Exercise 6.4.8 Let T be a linear transformation and suppose T = . Suppose S is the
   −4 −3

1 2 1
linear transformation with matrix B = . Find (S ◦ T )(v) for v = .
−1 3 −4
 6.5. Application: Perspective rendering 235

Exercise 6.4.9 What is the inverse of a reflection? Rotation? Shearing? Scaling?

 
2 1
Exercise 6.4.10 Let T be a linear transformation with matrix A = . Find the matrix of T −1 .
5 2
   
x 4x − 3y
Exercise 6.4.11 Let T be the linear transformation given by T = . Find the matrix
y 2x − 2y
of T −1 .
     
1 9 0
Exercise 6.4.12 Let T be a linear transformation and suppose T = and T =
  2 8 −1
−4
. Find the matrix of T and the matrix of T −1 .
−3

6.5 Application: Perspective rendering

As an application of linear transformations, we consider the problem of perspective rendering. Imagine

some object has been described by coordinates in 3-dimensional space, and we wish to make an image of
the object as it would be seen by a human eye or by a camera. The process of computing such an image is
known as rendering.
Conceptually, the rendering process makes use of a camera, which we will assume is located at the
origin of a 3-dimensional coordinate system called the camera coordinate system, and an image plane,
which we will assume is the plane z = 1 in camera coordinates. The 3-dimensional space also contains
one or more objects that we wish to render. We can consider the object to be described by a set of points.
For each point p on the object, we draw a straight line from p to the camera, and let p′ be the point where
this line intersects the image plane. The point p of the object is rendered as the point p′ in the image. This
process is illustrated in the following figure:
p

p' Object
y

x z

Camera
Image plane
236 Linear transformations in Rn

Object coordinates

It is convenient to describe each object in its own coordinate system, called the object coordinate system.
To illustrate this concept, we will consider a cube of side length 2, centered at the origin. The 8 corners of
this cube have the following coordinates in the object coordinate system:
               
1 1 1 1 −1 −1 −1 −1
 1  ,  1  ,  −1  ,  −1  ,  1  ,  1  ,  −1  ,  −1  .
1 −1 1 −1 1 −1 1 −1

For later reference, let us call this the standard cube. The following picture shows the standard cube
within its object coordinate system:

x
y

Conversion to camera coordinates

Before we render an object, we need to place it in some appropriate location relative to the camera. We do
this by specifying four vectors q, ax , ay , and az in R3 . Here, q is the origin of the object coordinate system,
relative to the camera coordinate system. The vectors ax , ay , and az are the axes of the object coordinate
system, relative to the camera coordinate system, as shown in the following illustration:

az

ax ay
y
q
z
x

Camera
  
x px
Thus, given a point with object coordinates v =  y , we can find its camera coordinates p =  py  by
z pz
the following formula:
p = q + xax + yay + zaz .
 6.5. Application: Perspective rendering 237

If we write A for the 3 × 3-matrix whose columns are ax , ay , and az , we can also write this formula more
succinctly as
p = q + Av.

Example 6.24: Converting object coordinates to camera coordinates

Let    
0 0.8 0.6 0
q =  0.5  , A= 0 0 1 .
5 0.6 −0.8 0
Convert each of the 8 corners of the standard cube from object coordinates to camera coordinates.

Solution. Let v1 , . . . , v8 be the object coordinates of the 8 corners of the cube:

               
1 1 1 1 −1 −1 −1 −1
 1  ,  1  ,  −1  ,  −1  ,  1  ,  1  ,  −1  ,  −1  .
1 −1 1 −1 1 −1 1 −1

We convert each of them to camera coordinates using the formula pi = q + Avi :

      
0 0.8 0.6 0 1 1.4
p1 =  0.5  +  0 0 1   1  =  1.5  ,
5 0.6 −0.8 0 1 4.8
      
0 0.8 0.6 0 1 1.4
p2 =  0.5  +  0 0 1   1  =  −0.5  ,
5 0.6 −0.8 0 −1 4.8
      
0 0.8 0.6 0 1 0.2
p3 =  0.5  +  0 0 1   −1  =  1.5  ,
5 0.6 −0.8 0 1 6.4

and so on. Continuing in the same fashion, we find p1 , . . . , p8 :

               
1.4 1.4 0.2 0.2 −0.2 −0.2 −1.4 −1.4
 1.5  ,  −0.5  ,  1.5  ,  −0.5  ,  1.5  ,  −0.5  ,  1.5  ,  −0.5  . (6.2)
4.8 4.8 6.4 6.4 3.6 3.6 5.2 5.2

We can also write f : R3 → R3 for the function that converts object coordinates to camera coordinates,
i.e.,
f (v) = q + Av.
We note that this is not a linear function, because f (0) 6= 0. The function f is called an affine function,
which means that it is a linear function v 7→ Av followed by a translation v 7→ q + v.
238 Linear transformations in Rn

Rendering


px
Once we know the camera coordinates p =  py  of a point, we need to render the point, i.e., find its
pz
coordinates in the image plane.
p

y
p′
z
x

Camera

Image plane

Since the camera is located at the origin, the line that passes through the camera and the point p has the
parametric equation  
t px
r = tp =  t py  .
t pz
1
Since the image plane is the plane z = 1, we must set t such that t pz = 1, i.e., t = pz . Therefore, the
coordinates of the rendered point are
 
px /pz
1
p′ = p =  py /pz  .
pz
1
Finally, since the image plane
 is 2-dimensional,
 we can forget the now useless z-coordinate, and render
px /pz
the point at the coordinates in the 2-dimensional image plane.
py /pz

Example 6.25: Rendering

Render the cube from Example 6.24.

Solution. We must apply the rendering function

 
px  
   px /pz
g py =
py /pz
pz
 6.5. Application: Perspective rendering 239

to each of the corners of the cube from (6.2). We have

 
1.4    
   1.4/4.8 0.292
g 1.5 = = ,
1.5/4.8 0.312
4.8
 
1.4    
   1.4/4.8 0.292
g −0.5 = = ,
−0.5/4.8 −0.104
4.8
 
0.2    
0.2/6.4 0.031
g  1.5  = = ,
1.5/6.4 0.234
6.4
and so on. The 8 rendered points are:
               
0.292 0.292 0.031 0.031 −0.056 −0.056 −0.269 −0.269
, , , , , , , .
0.312 −0.104 0.234 −0.078 0.417 −0.139 0.288 −0.096
Drawing these in the 2-dimensional image plane, we get the following picture, which is the final perspective-
rendered image of the cube:

Animation

We placed our object in the camera coordinate system using a coordinate transformation function
f (v) = q + Av.
One of the advantages of using such a coordinate transformation (as opposed to specifying the object
points directly in the camera coordinate system) is that this makes it very easy to move the objects around,
rotate them, scale and shrink them, etc. For example:
1. To move the object to a different location, we only have to change the vector q.
2. To rotate the object about its own z-axis, we only have to replace v by Rθ v, where Rθ is the matrix
for a rotation about the z-axis by angle θ :
 
cos θ − sin θ 0
Rθ =  sin θ cos θ 0  .
0 0 1
240 Linear transformations in Rn

Similarly to Rθ , we can also insert other transformation matrices (for example, we could rotate the object
about its x-axis instead of its z-axis, scale the object, etc). We can even make an animation by rendering
the object repeatedly for different values of these parameters.

Example 6.26: An animated cube

Make an animation of a rotating, moving cube. The animation is 5 seconds long (i.e., time t ranges
from 0 to 5). The location of the cube at time t , in camera coordinates, is given by
 
−3 + 3t
q(t) =  −3  .
8 + 3t

The transformation matrix A is as in Example 6.24. Moreover, the cube should make one quarter
rotation about its z-axis during the time of the animation, i.e., it should be transformed by Rθ , where
π
θ = 10 t . Compute 6 frames of the animation, for t = 0, t = 1, . . . , t = 5.

Solution. For each of the animation frames t ∈ {0, 1, 2, 3, 4, 5}, we do a calculation very similar to that of
Examples 6.24 to convert the cube coordinates to camera coordinates, using the coordinate transformation

f (v) = q(t) + ARθ v,

π
where θ = 10 t. We then render each of the frames using the same method as in Example 6.25. We skip
the detailed calculations, which are best done by computer (though they could be done by hand, of course,
as we did in Examples 6.24 and 6.25). The final rendered frames look like this:

t =2 t = 3 t = 4t = 5
t =1
t =0

Note that there is a bit of distortion in the first and last cubes. This is because the camera is very close to
the image plane (the scene has been “filmed” with a wide-angle camera). The distortion goes away if you
close one eye and bring the other eye very close to the page. ♠
7. Determinants

7.1 Determinants of 2 × 2- and 3 × 3-matrices

Outcomes
A. Calculate the determinant of 2 × 2-matrices and 3 × 3-matrices.

Let A be an n × n-matrix. The determinant of A, denoted by det(A), is a very important number which
we will explore throughout this chapter.
The determinant of a2 × 2-matrix is given by the following formula.

Definition 7.1: Determinant of a 2 × 2-matrix

 
a b
Let A = . Then
c d
det(A) = ad − bc.

Example 7.2: A 2 × 2 determinant

 
2 4
Find det(A) for the matrix A = .
−1 6

Solution. We have det(A) = 2 · 6 − (−1) · 4 = 12 + 4 = 16. ♠

The determinant is also often denoted by enclosing the matrix with two vertical lines. Thus
   
a b  a b 
det =   = ad − bc.
c d c d 

Definition 7.3: Determinant of a 3 × 3-matrix

 
a11 a12 a13
Let A =  a21 a22 a23 . Then
a31 a32 a33

det(A) = a11 a22 a33 + a12 a23 a31 + a13 a21 a32 − a31 a22 a13 − a32 a23 a11 − a33 a21 a12 .

241
242 Determinants

The following picture may help in memorizing the formula:

 
a11 a12 a13 a11 a12
a 
 21 a22 a23  a21 a22
a31 a32 a33 a31 a32

Here, we have written down the matrix A, then repeated the first two columns next to it. The blue lines
correspond to the positive terms of the determinant: a11 a22 a33 , a12 a23 a31 , and a13 a21 a32 . The pink lines
correspond to the negative terms: a31 a22 a13 , a32 a23 a11 , and a33 a21 a12 .

Example 7.4: A 3 × 3 determinant

Find det(A), where  
0 1 2
A= 3 1 0 .
1 1 −1

Solution. We have
 
 0 1 2 
 

det(A) =  3 1 0  = 0 · 1 · (−1) + 1 · 0 · 1 + 2 · 3 · 1 − 1 · 1 · 2 − 1 · 0 · 0 − (−1) · 3 · 1 = 7.

 1 1 −1 

Exercises

Exercise 7.1.1 Find the determinants of the following matrices.

       
1 3 0 3 4 3 −3 4
(a) (b) (c) (d)
0 2 0 2 6 2 −1 2

Exercise 7.1.2 Find the following determinants.

       
 1 2 3   1 0 2   3 4 1   0 −2 1 
       
(a)  4 5 6  (b)  2 5 3  (c)  0 −1 1 
 (d)  4 1 −3 
 7 8 9   −1 0 0   1 2 1   −1 3 1 
 7.2. Minors and cofactors 243

7.2 Minors and cofactors

Outcomes
A. Compute minors and cofactors of matrices.

B. Use cofactor expansion to compute the determinant of an n × n-matrix.

Determinants of larger matrices can be computed in terms of the determinants of smaller matrices. We
begin with the following definition.

Definition 7.5: The ij th minor of a matrix

Let A be an n × n-matrix. The ij th minor of A, denoted by Mi j , is the determinant of the (n − 1) ×
(n − 1)-matrix that is obtained by deleting the i th row and the j th column of A.

Hence, there is a minor associated with each entry of A. The following example illustrates this definition.

Example 7.6: Finding minors of a matrix

Let  
1 2 3
A= 4 3 2 
3 2 1
Find the minors M12 and M33 .

Solution. First we will find M12 . By definition, this is the determinant of the 2 × 2-matrix that results when
we delete the first row and the second column of A. This minor is given by
 
 1 2 3   
   4 2 
M12 =  4 3 2  =   = 4 · 1 − 3 · 2 = − 2.
 3 2 1  3 1 

Similarly, M33 is the determinant of the 2 × 2-matrix that is obtained by deleting the third row and the third
column of A. This minor is therefore
 
 1 2 3   
   1 2 
M33 =  4 3 2  =   = − 5.
 3 2 1  4 3 

We now define the ij th cofactor of a matrix A, which is either plus or minus the ij th minor.
244 Determinants

Definition 7.7: The ij th cofactor of a matrix

Suppose A is an n × n-matrix. The ij th cofactor, denoted by Ci j , is defined to be

Ci j = (−1)i+ j Mi j

In other words, the ij th cofactor is equal to the corresponding minor if i + j is even, and the negative of the
minor if i + j is odd. For remembering the signs, the following picture is sometimes helpful:
 
+ − + −
 − + − + 
 
 + − + − .
− + − +

Example 7.8: Finding cofactors of a matrix

Consider the matrix  
1 2 3
A= 4 3 2 
3 2 1
Find the cofactors C12 and C33 .

Solution. We have already computed the corresponding minors in Example 7.6. For the cofactors, we
have:
C12 = (−1)1+2 M12 = − M12 = − (−2) = 2,
C33 = (−1)3+3 M33 = + M33 = + (−5) = − 5.
Note that 1 + 2 is odd, so C12 = −M12 . On the other hand, 3 + 3 is even, so C33 = M33 . ♠
You may wish to find the remaining cofactors of the above matrix. Remember that there is a cofactor for
every entry in the matrix.
We have now established the tools we need to find the determinant of an n × n-matrix.

Definition 7.9: The determinant of an n × n-matrix

Let A be an n × n-matrix. Then det(A) is calculated by picking a row (or column) and taking the
product of each entry in that row (column) with its cofactor and adding these products together.
This process is known as expanding along the i th row (or column).

In formulas, the process of expanding along the i th row is given as follows:

det(A) = ai1Ci1 + ai2Ci2 + . . . + ainCin .

Similarly, the process of expanding along the j th column is:

det(A) = a1 jC1 j + a2 jC2 j + . . . + an jCn j .

 7.2. Minors and cofactors 245

When calculating the determinant, you can choose to expand any row or any column. Regardless of which
row or column you expand, you will always get the same number, which is the determinant of the matrix
A. This method of evaluating a determinant by expanding along a row or a column is also called cofactor
expansion or Laplace expansion.

Example 7.10: Finding a determinant by cofactor expansion

Find det(A) using the method of cofactor expansion, where
 
1 2 3
A =  4 3 2 .
3 2 1

Solution. First, we will calculate det(A) by expanding along the first column. Using Definition 7.9, the
determinant is

det(A) = a11C11 + a21C21 + a31C31

= a11 M11 − a21 M21 + a31 M31
     
 3 2   2 3   
= 1  −4 +3 2 3 
2 1   2 1   3 2 

= 1(−1) − 4(−4) + 3(−5)

= 0.

As mentioned in Definition 7.9, we can choose to expand along any row or column. Let’s try now by
expanding along the second row. The calculation is as follows.
     
 2 3   1 3   1 2 
det(A) = − 4  +3   
 3 1  − 2  3 2  = − 4(−4) + 3(−8) − 2(−4) = 0.
2 1 

You can see that for both methods, we obtained det(A) = 0. ♠

You should try to compute the above determinant by expanding along other rows and columns. This is a
good way to check your work, because you should come up with the same number each time!

Theorem 7.11: The determinant is well-defined

Expanding an n × n-matrix along any row or column always gives the same answer, which is the
determinant.

Example 7.12: Determinant of a four by four matrix

 
 1 2 3 0 
 
 2 4 2 3 
Calculate  .

 0 3 0 5 
 3 0 3 2 
246 Determinants

Solution. Using the cofactor method, we can expand this determinant along any row or column. But notice
that the third row contains two zeros. This makes the cofactor expansion particularly convenient. So let us
expand along the third row. We have:
 
 1 2 3 0         
   2 3 0   1 3 0   1 2 0   1 2 3 
 2 4 2 3         
  = 0 4 2 3 −3 2 2 3 +0 2 4 3 −5 2 4 2 .
 0 3 0 5         
   0 3 2   3 3 2   3 0 2   3 0 3 
 3 0 3 2 

Note that we only need to compute two of the 3 × 3 determinants, since the remaining two are multiplied
by 0. We can compute each 3 × 3 determinant using the method of Definition 7.3. We find:
 
 1 2 3 0     
   1 3 0   1 2 3 
 2 4 2 3     
  = −3  2 2 3  − 5  2 4 2  = −3 · 10 − 5 · (−24) = 90.
 0 3 0 5     
   3 3 2   3 0 3 
 3 0 3 2 

♠
We remark that the cofactor expansion is mainly useful for calculating determinants of small matrices, or
matrices containing many zeros. Indeed, imagine calculating the determinant of a 10 × 10-matrix by the
cofactor method. This requires calculating the determinants of ten 9 × 9-matrices, each of which requires
calculating the determinants of nine 8 × 8-matrices, each of which requires calculating the determinants of
eight 7 × 7-matrices, and so on. Calculating the determinant of a 10 × 10-matrix by the cofactor method
would therefore require 10 · 9 · 8 · . . . · 2 · 1 = 3628800 steps!
In the next few sections, we will explore some important properties and characteristics of the determi-
nant, including a much more efficient method of calculating determinants of large matrices.

Exercises

 
1 2 4
Exercise 7.2.1 Let A =  0 1 3 . Find the following minors and cofactors:
−2 5 1

(a) M11 ,

(b) M21 ,

(c) M32 ,

(d) C11 ,

(e) C21 ,

(f) C32 .
 7.2. Minors and cofactors 247

 
0 −1 3 1
 1 0 2 2 
Exercise 7.2.2 Let A = 
 2
. Find M11 , M21 , M32 , C11 , C21 , and C32 .
3 −1 0 
1 1 0 1

Exercise 7.2.3 Compute the determinants of the following matrices using cofactor expansion along any
row or column.
 
1 2 0
(a)  3 −2 2 
0 3 1
 
1 −2 2
(b)  3 0 0 
4 3 1
 
1 2 −2 2
 1 3 2 3 
(c) 
 4 0

1 0 
1 2 1 2

Exercise 7.2.4 Find the following determinant by expanding (a) along the first row and (b) along the
second column.  
 1 2 1 
 
 2 1 3 
 
 2 1 1 

Exercise 7.2.5 Find the following determinant by expanding (a) along the first column and (b) along the
third row.  
 2 3 1 1 
 
 4 3 1 2 
 
 1 1 0 1 
 
 3 2 1 2 

Exercise 7.2.6 Find the following determinant by expanding (a) along the second row and (b) along the
first column.  
 1 2 −1 1 
 
 0 1 2 1 
 
 0 2 1 3 

 1 4 0 2 

Exercise 7.2.7 Compute the determinant by cofactor expansion. Pick the easiest row or column to use.
 
 1 0 0 1 
 
 2 1 1 0 
 
 0 0 0 2 
 
 2 1 3 1 
248 Determinants

7.3 The determinant of a triangular matrix

Outcomes
A. Calculate the determinant of an upper or lower triangular matrix.

There is a certain type of matrix for which finding the determinant is a very simple procedure: a triangular
matrix.

Definition 7.13: Triangular matrices

An square matrix A is upper triangular if ai j = 0 whenever i > j. In other words, a matrix is upper
triangular if the entries below the main diagonal are 0. Thus, an upper triangular matrix looks as
follows, where ∗ refers to any non-zero number:
 
∗ ∗ ··· ∗ ∗
 0 ∗ ··· ∗ ∗ 
 
 .. .. . . .. .. 
 . . . . . .
 
 0 0 ··· ∗ ∗ 
0 0 ··· 0 ∗

Similarly, a square matrix is lower triangular if all entries above the main diagonal are 0.

The following theorem provides a useful way to calculate the determinant of a triangular matrix.

Theorem 7.14: Determinant of a triangular matrix

Let A be an upper or lower triangular matrix. Then det(A) is equal to the product of the entries on
the main diagonal. Written as a formula, we have

det(A) = a11 a22 · · · ann .

Example 7.15: Determinant of a triangular matrix

Compute det(A), where  
1 2 3 16
 0 2 6 −7 
A=
 0
.
0 3 33 
0 0 0 −1

Solution. By Theorem 7.14, it suffices to take the product of the elements on the main diagonal. Thus

det(A) = 1 · 2 · 3 · (−1) = − 6.
 7.3. The determinant of a triangular matrix 249

♠
For comparison, let us compute the determinant without Theorem 7.14, i.e., by using cofactor expansion.
If we expand the determinant along the first column, we get:
       
 2 6 −7   2 3 16   2 3 16   2 3 16 
       
det(A) = 1  0 3 33  − 0  0 3 33  + 0  2 6 −7  − 0  2 6 −7  .
 0 0 −1   0 0 −1   0 0 −1   0 3 33 

The only non-zero term in the expansion is

 
 2 6 −7 
 
1  0 3 33 .

 0 0 −1 

We can in turns expand this 3 × 3 determinant by the cofactor method along the first column:
 
 2 6 −7       
   3 33   6 −7   6 −7 
 0 3 33  = 2      
   0 −1  − 0  0 −1  + 0  3 33  .
 0 0 −1 

Again, the only non-zero term is the first term. In summary,

 
 1 2 3 16   
   2 6 −7   
 0 2 6 −7     
det(A) =   = 1  0 3 33  = 1 · 2  3 33  = 1 · 2 · 3 · (−1) = − 6.
    0 −1 
 0 0 3 33   0 0 −1 
 0 0 0 −1 

Of course this is just the same as the product of the diagonal entries of A, which is the point of Theo-
rem 7.14.

Exercises

Exercise 7.3.1 Find the determinant of the following matrices.

 
1 −34
(a) A =
0 2
 
4 0 0
(b) B =  3 −2 0 
14 1 5
 
2 3 15 0
 0 4 1 7 
(c) C = 
 0 0 −3

5 
0 0 0 1
250 Determinants

7.4 Determinants and row operations

Outcomes
A. Determine the effect of a row operation on the determinant of a matrix.

B. Use row operations to calculate a determinant.

Recall that there are three kinds of elementary row operations on matrices:

1. Switch two rows.

2. Multiply a row by a non-zero number.

3. Add a multiple of one row to another row.

The following theorem examines the effect of these row operations on the determinant of a matrix.

Theorem 7.16: Effect of row operations on the determinant

Let A be an n × n-matrix.

1. If B is obtained from A by switching two rows, then

det(B) = − det(A).

2. If B is obtained from A by multiplying one row by a non-zero scalar k, then

det(B) = k det(A).

3. If B is obtained from A by adding a multiple of one row to another row, then

det(B) = det(A).

Notice that the second part of this theorem is true when we multiply one row of the matrix by k. If we
were to multiply two rows of A by k to obtain B, we would have det(B) = k2 det(A).

Example 7.17: Using row operations to calculate a determinant

Use row operations to calculate the following determinant:
 
 1 5 5 
 
 0 0 −3  .
 
 0 2 7 
 7.4. Determinants and row operations 251

Solution. If we switch the second and third rows, we obtain a triangular matrix, of which the determinant
is easy to compute. By Theorem 7.16, switching two rows negates the determinant. We therefore have:
   
 1 5 5   1 5 5 
   
 0 0 −3  = −  0 2 7  = − (1 · 2 · (−3)) = 6.
  
 0 2 7   0 0 −3 

Example 7.18: Using row operations to calculate a determinant

Use row operations to calculate the following determinant:
 
 1 4 −2 
 
 1 8 1  .

 2 4 −9 

Solution. We can use elementary row operations to reduce this matrix to triangular form:
       
1 4 −2 1 4 −2 1 4 −2 1 4 −2
R2 ← R2 −R1 R3 ← R3 −2R1 R3 ← R3 +R2
 1 8 1  ≃  0 4 3  ≃  0 4 3  ≃  0 4 3 .
2 4 −9 2 4 −9 0 −4 −5 0 0 −2

Each of the row operations is of the form “add a multiple of one row to another row”, and therefore does
not change the determinant. We therefore have:
       
 1 4 −2   1 4 −2   1 4 −2   1 4 −2 
      
 1 8 1  =  0 4 3  =  0 4 3  =  0 4 3  = 1 · 4 · (−2) = −8.
       
 2 4 −9   2 4 −9   0 −4 −5   0 0 −2 

♠
In general, we can convert any square matrix to triangular form using elementary row operations. In fact,
it is always possible to do so using only elementary operations of the first and third kind (swap two rows
or add a multiple of one row to another). This gives us a very efficient way to compute determinants. If
the matrices are large, this method is much more efficient than the cofactor method.

Example 7.19: Using row operations to calculate a determinant

Use elementary row operations of the first and third kind to calculate the following determinant:
 
 0 2 1 4 

 2 2 −4 −1 
 
 1 1 −2 −1  .
 
 1 3 2 5 
252 Determinants

Solution. We use elementary row operations to reduce the matrix to triangular form:
     
0 2 1 4 1 1 −2 −1 1 1 −2 −1
 2 2 −4 −1  R1 ↔ R3  2 2 −4 −1  R2 ← R2 −2R1  0 0 0 1 
  ≃   ≃  
 1 1 −2 −1   0 2 1 4   0 2 1 4 
1 3 2 5 1 3 2 5 1 3 2 5
   
1 1 −2 −1 1 1 −2 −1
R4 ← R4 −R1  0 0 0 1   1 
≃   R4 ←≃R4 −R3  0 0 0 
 0 2 1 4   0 2 1 4 
0 2 4 6 0 0 3 2
   
1 1 −2 −1 1 1 −2 −1
R2 ↔ R3  0 2 1 4  R3 ↔ R4  0 2 1 4 
≃   ≃  
 0 0 0 1   0 0 3 2 
0 0 3 2 0 0 0 1
By Theorem 7.16, the determinant changes signs each time we swap two rows. The determinant is un-
changed when we add a multiple of one row to another. Therefore, we have
     
 0 2 1 4   1 1 −2 −1   1 1 −2 −1 
    
 2 2 −4 −1   2 2 −4 −1   0 0 0 1 
  = −  = − 
 1 1 −2 −1   0 2 1 4   0 2 1 4 
     
 1 3 2 5   1 3 2 5   1 3 2 5 
   
 1 1 −2 −1   1 1 −2 −1 
   
 0 0 0 1   0 0 0 1 
= −   = − 
 0 2 1 4   0 2
 1 4 
 0 2 4 6   0 0 3 2 
   
 1 1 −2 −1   1 1 −2 −1 
   
 0 2 1 4   0 2 1 4 
= +  = −  = −6.
 0 0 0 1   0 0 3 2 
 0 0 3 2   0 0 0 1 
In practice, the last calculation could have been done in a single step. All we had to do is count the number
of swap operations we performed during the row operations. If there is an odd number of swap operations,
the sign of the determinant changes; otherwise, it stays the same. ♠

Exercises

Exercise 7.4.1 Use row operations to calculate the following determinants:

   
     1 2 1 2   1 4 1 2 
 1 2 1   2 1 3    
    3 1 −2 3   3 2 −2 3 
  
(a)  2 3 2  , (b)  2 4 2  , (c)  , (d)  .
 −4 1 2   1 4 −5   −1 0 3 1   −1 0
 3 3 

 2 3 2 −2   2 1 2 −2 
 7.5. Properties of determinants 253

7.5 Properties of determinants

Outcomes
A. Use the determinant of a square matrix to decide whether the matrix is invertible.

B. From the determinants of two matrices, calculate the determinant of their product.

C. From the determinant of a matrix, calculate the determinant of its inverse.

D. From the determinant of a matrix, calculate the determinant of its transpose.

E. Calculate the determinant of kA, if the determinant of A is known.

F. Without calculation, find the determinant of a matrix containing a row or column of zeros, or
a matrix containing a row (or column) that is a scalar multiple of another row (or column).

G. Use algebraic properties to reason about determinants.

One reason that the determinant is such an important quantity is that it permits us to tell whether a square
matrix is invertible.

Theorem 7.20: Determinants and invertible matrices

Let A be an n × n-matrix. Then A is invertible if and only if det(A) 6= 0.

Proof. We know that every matrix A can be converted to echelon form by elementary row operations. We
also know from Theorem 7.16 that no elementary row operation changes whether the determinant is zero
or not. Let R be an echelon form of A. Because R is an echelon form, it is also an upper triangular matrix.
Case 1: A is invertible. In that case, the rank of R is n, and every diagonal entry of R is a pivot entry
(therefore non-zero). It follows that det(R) 6= 0, which implies det(A) 6= 0. Case 2: A is not invertible.
In that case, the triangular matrix R contains a row of zeros. It follows that det(R) = 0, and therefore
det(A) = 0. ♠

Example 7.21: Determinants and invertible matrices

Determine which of the following matrices are invertible by computing their determinants.
 
    1 2 −5
3 6 2 3
A= , B= , C= 2 0 2 .
2 4 5 1
3 1 0

Solution. We have det(A) = 3 · 4 − 2 · 6 = 0 and det(B) = 2 · 1 − 5 · 3 = −13. Therefore, B is invertible and

A is not invertible. A quick way to compute the determinant of C is to expand it along the third row. We
254 Determinants

have    
 2 −5   1 −5 

det(C) = 3   
−1  = 3 · 4 − 1 · 12 = 0.
0 2  2 2 
Therefore, C is not invertible. ♠

As an application of Theorem 7.20, we note that the determinant of an n × n-matrix can be used to predict
whether a homogeneous system of equations has non-trivial solutions.

Corollary 7.22: Determinants and homogeneous systems

Let A be an n × n-matrix. Then the homogeneous system Av = 0 has non-trivial solutions if and
only if det(A) = 0.

Proof. We know from Theorem 1.35 that the homogeneous system has a non-trivial solution if and only if
rank(A) < n. This is the case if and only if A is not invertible, i.e., if and only if det(A) = 0. ♠
Another reason the determinant is important is that it compatible with matrix product.

Theorem 7.23: Determinant of a product

Let A and B be n × n-matrices. Then

det(AB) = det(A) det(B)

Proof. We first prove this in case A = E is an elementary matrix. Remember from Section 4.6 that ele-
mentary matrices correspond to elementary row operations.

1. If E is an elementary matrix for swapping two rows, then det(E) = −1. Also, by Theorem 7.16(1),
det(EB) = − det(B). Therefore det(EB) = det(E) det(B).

2. If E is an elementary matrix for multiplying a row by a non-zero scalar k, then det(E) = k. Also, by
Theorem 7.16(2), det(EB) = k det(B). Therefore det(EB) = det(E) det(B).

3. If E is an elementary matrix for adding a multiple of one row to another, then det(E) = 1. Also, by
Theorem 7.16(3), det(EB) = det(B). Therefore det(EB) = det(E) det(B).

Now consider the case where A is an arbitrary matrix. Case 1: A is invertible. Then by Theorem 4.61,
we can write A as a product of elementary matrices A = E1 E2 · · · Ek . By repeatedly using the formula
det(EB) = det(E) det(B) that we proved above, we have

det(AB) = det(E1 E2 · · · Ek B) = det(E1 ) det(E2 ) · · · det(Ek ) det(B) = det(A) det(B).

Case 2: A is not invertible. Then AB is also not invertible (because if C were an inverse of AB, we
would have ABC = I, and therefore, BC would be an inverse of A). Therefore, by Theorem 7.20, we have
det(A) = 0 and det(AB) = 0. It follows that det(AB) = det(A) det(B). ♠
 7.5. Properties of determinants 255

Example 7.24: The determinant of a product

Compare det(AB) and det(A) det(B), where
   
1 2 3 2
A= and B = .
−3 2 4 1

Solution. We first compute AB:

    
1 2 3 2 11 4
AB = = .
−3 2 4 1 −1 −4

The three determinants are

     
 11 4   1 2   3 2 
det(AB) =   = −40,


det(A) =   = 8, and det(B) =   = −5.
−1 −4 −3 2  4 1 

Therefore det(A) det(B) = 8 · (−5) = −40 = det(AB). ♠

The following proposition summarizes some properties of determinants we have discussed so far, as well
as additional properties.

Proposition 7.25: Properties of determinants

Let A, B be n × n-matrices. Then:

1. det(AB) = det(A) det(B).

2. det(I) = 1.

3. A is invertible if and only if det(A) 6= 0. Moreover, if this is the case, then

1
det(A−1 ) = .
det(A)

4. det(kA) = kn det(A).

5. det(AT ) = det(A).

Proof. Property 1 is a restatement of Theorem 7.23. Property 2 follows from Theorem 7.14, because the
identity matrix is an upper triangular matrix. Property 3: The first part is Theorem 7.20. For the second
part, assume A is invertible. Then by properties 1 and 2, det(A) det(A−1 ) = det(AA−1 ) = det(I) = 1. The
claim follows by dividing both sides of the equation by det(A). Property 4 follows from Theorem 7.16(2),
because kA is obtained from A by multiplying all n rows by k. Each time we multiple one row by k, the
determinant is multiplied by k. Property 5 follows because expanding det(A) along columns amounts to
the same thing as expanding det(AT ) along rows. ♠
We end this section with a few useful ways of spotting matrices of determinant 0.
256 Determinants

Theorem 7.26: Special matrices with zero determinant

Let A be an n × n-matrix.

1. If A has a row consisting only of zeros, or a column consisting only of zeros, then det(A) = 0.

2. If A has a row that is a scalar multiple of another row, or a column that is a scalar multiple of
another column, then det(A) = 0.

Proof. The first property follows by cofactor expansion: simply expand the determinant along the row
or column that consists only of zeros. For the second property, assume that A has a row that is a scalar
multiple of another row. We can then perform an elementary row operation to create a row of zeros.
By Theorem 7.16(3), the determinant is unchanged, so that det(A) = 0. In the case that A has a column
that is a scalar multiple of another column, we apply the same reasoning to AT and use the fact that
det(A) = det(AT ). ♠

Exercises

Exercise 7.5.1 Let  

a b
A= .
c d
An operation is done to get from A to a matrix B. In each case, identify which operation was done and
explain how it will affect the value of the determinant.

(a)  
a c
B=
b d

(b)  
c d
B=
a b

(c)  
a b
B=
a+c b+d

(d)  
a b
B=
2c 2d

(e)  
b a
B=
d c
 7.5. Properties of determinants 257

Exercise 7.5.2 Let A be an n × n-matrix and suppose there are n − 1 rows such that the remaining row is
a linear combinations of these n − 1 rows. Show det(A) = 0.

Exercise 7.5.3 Let A be an n × n-matrix. Show that if det(A) 6= 0 and Ax = 0, then x = 0.

Exercise 7.5.4 Using only Theorems 7.14 and 7.23, show that det(kA) = kn det(A) for an n × n-matrix A
and scalar k.

Exercise 7.5.5 Construct two random 2 × 2-matrices A and B and verify that det(A) det(B) = det(AB).

Exercise 7.5.6 Is it true that det(A + B) = det(A) + det(B)? If this is so, explain why. If it is not so, give
a counterexample.

Exercise 7.5.7 An n × n-matrix is called nilpotent if there exists some positive integer k such that Ak = 0.
If A is a nilpotent matrix, what are the possible values of det(A)?

Exercise 7.5.8 A square matrix is said to be orthogonal if AT A = I. Thus the inverse of an orthogonal
matrix is its transpose. What are the possible values of det(A) if A is an orthogonal matrix?

Exercise 7.5.9 Let A and B be two n × n-matrices. We say that A is similar to B, in symbols A ∼ B, if there
exists an invertible matrix P such that A = P−1 BP. Show that if A ∼ B, then det(A) = det(B).

Exercise 7.5.10 Find the determinant of

 
1 1 1
A =  1 a a2  .
1 b b2

For which values of a and b is this matrix invertible? Hint: after you compute the determinant, you can
factor out (a − 1) and (b − 1) from it.

Exercise 7.5.11 Assume A, B, and C are n × n-matrices and ABC is invertible. Use determinants to show
that each of A, B, and C is invertible.

Exercise 7.5.12 Suppose A is an upper triangular matrix. Show that A−1 exists if and only if all elements
of the main diagonal are non-zero. Is it true that A−1 will also be upper triangular? Explain. Could the
same be concluded for lower triangular matrices?
258 Determinants

Exercise 7.5.13 Specify whether each statement is true or false. If true, provide a proof. If false, provide
a counterexample.

(a) If A is a 3 × 3-matrix with determinant zero, then one column must be a multiple of some other
column.

(b) If any two columns of a square matrix are equal, then the determinant of the matrix equals zero.

(c) For two n × n-matrices A and B, det(A + B) = det(A) + det(B).

(d) For an n × n-matrix A, det(3A) = 3 det(A).

(e) If A−1 exists, then det(A−1 ) = det(A)−1.

(f) If B is obtained by multiplying a single row of A by 4, then det(B) = 4 det(A).

(g) For an n × n-matrix A, we have det(−A) = (−1)n det(A).

(h) If A is a real n × n-matrix, then det(AT A) ≥ 0.

(i) If Ak = 0 for some positive integer k, then det(A) = 0.

(j) If Ax = 0 for some x 6= 0, then det(A) = 0.

7.6 Application: A formula for the inverse of a matrix

Outcomes
A. Find the cofactor matrix and the adjugate of a matrix.

B. Find the inverse of a matrix using the adjugate formula.

The determinant of a matrix also provides a way to find the inverse of a matrix. Recall the definition of
the inverse of a matrix from Definition 4.36. If A is an n × n-matrix, we say that A−1 is the inverse of A if
AA−1 = I and A−1 A = I.
We now define a new matrix called the cofactor matrix of A. The cofactor matrix of A is the matrix
whose ij th entry is the ij th cofactor of A.
 7.6. Application: A formula for the inverse of a matrix 259

Definition 7.27: The cofactor matrix

Let A be an n × n-matrix. Then the cofactor matrix of A, denoted cof(A), is defined by
 
C11 C12 · · · C1n
   C21 C22 · · · C2n 
 
cof(A) = Ci j =  .. ... . . . ...  ,
 . 
Cn1 Cn2 · · · Cnn

where Ci j is the ij th cofactor of A.

We will use the cofactor matrix to create a formula for the inverse of A. First, we define the adjugate of
A, denoted adj(A), to be the transpose of the cofactor matrix:

adj(A) = cof(A)T .

The adjugate is also sometimes called the classical adjoint of A.

Example 7.28: Cofactor matrix and adjugate

Find the cofactor matrix and the adjugate of A, where
 
1 2 3
A =  3 0 1 .
1 2 1

Solution. We first find cof(A). To do so, we need to compute the cofactors of A. We have:
 
 1 2 3   
   0 1 
C11 = + M11 =  3 0 1  =   = − 2,
 1 2 1  2 1 
 
 1 2 3   
   3 1 

C12 = − M12 = −  3 0 1  = −     = − 2,
1 1 
 1 2 1 
 
 1 2 3   
   3 0 

C13 = + M13 =  3 0 1  =     = 6,
1 2 
 1 2 1 
 
 1 2 3   
   2 3 

C21 = − M21 = −  3 0 1  = −     = 4,
2 1 
 1 2 1 

and so on. Continuing in this way, we find the cofactor matrix

 
−2 −2 6
cof(A) =  4 −2 0 .
2 8 −6
260 Determinants

Finally, the adjugate is the transpose of the cofactor matrix:

 
−2 4 2
adj(A) = cof(A)T =  −2 −2 8 .
6 0 −6

♠
The following theorem provides a formula for A−1 using the determinant and the adjugate of A.

Theorem 7.29: Formula for the inverse

Let A be an n × n-matrix. Then

A adj(A) = adj(A) A = det(A) I.

Moreover, if A is invertible if and only if det(A) 6= 0. In this case, we have:

1
A−1 = adj(A).
det(A)

We call this the adjugate formula for the matrix inverse.

Proof. Recall that the (i, j)-entry of adj(A) is equal to C ji . Thus the (i, j)-entry of B = A adj(A) is:

Bi j = ai1 adj(A)1 j + ai2 adj(A)2 j + . . . + ain adj(A)n j

= ai1C j1 + ai2C j2 + . . . + ainC jn .

By the cofactor expansion theorem, we see that this expression for Bi j is equal to the determinant of the
matrix obtained from A by replacing its jth row by [ai1 , ai2 , . . . ain ], i.e., by its ith row.
If i = j then this matrix is A itself and therefore Bii = det(A). If on the other hand i 6= j, then this
matrix has its ith row equal to its jth row, and therefore Bi j = 0 in this case. Thus we obtain:

A adj(A) = det(A)I.

By a similar argument (using columns instead of rows), we can verify that:

adj(A) A = det(A)I.

This proves the first part of the theorem. For the second part, assume that A is invertible. Then by
Theorem 7.20, det(A) 6= 0. Dividing the formula from the first part of the theorem by det(A), we obtain
   
1 1
A adj(A) = adj(A) A = I,
det(A) det(A)

and therefore
1
A−1 = adj(A).
det(A)
This completes the proof. ♠
 7.6. Application: A formula for the inverse of a matrix 261

Example 7.30: Finding the inverse using a formula

Use the adjugate formula to find the inverse of the matrix
 
1 2 3
A =  3 0 1 .
1 2 1

Solution. We must compute

1
A−1 = adj(A).
det(A)
We will start by computing the determinant. We expand along the second row:
 
 1 2 3     
   2 3   1 2 
det(A) =  3 0 1  = − 3  −1 
 1 2  = − 3(−4) − 1(0) = 12.
 1 2 1  2 1 

We have already calculated the adjugate adj(A) in Example 7.28:

 
−2 4 2
adj(A) =  −2 −2 8 .
6 0 −6

Therefore, the inverse of A is given by

   
−2 4 2 − 16 1
3
1
6
1 1  −2 −2
 
 =  −1

.
A−1 = adj(A) = 8 − 16 2
det(A) 12    6 3 
1
6 0 −6 2 0 − 12

Since it is very easy to make a mistake in this calculation, we double-check our answer by computing
A−1 A:     
− 16 1
3
1
6 1 2 3 1 0 0
    
−1 
A A =  −6 −6 1 1 2   =  0 1 0  = I.
3  3 0 1   
1 1
2 0 −2 1 2 1 0 0 1
♠

Example 7.31: Finding the inverse using a formula

Use the adjugate formula to find the inverse of the matrix
 
0 2 1
A =  −1 2 2 .
2 −2 −2
262 Determinants

Solution. We start by calculating the determinant:

 
 0 2 1     
   −1   
det(A) =  −1 2 2  = − 2 
2  + 1  −1 2  = − 2 · (−2) + 1 · (−2) = 2.
2 −2   2 −2 
 2 −2 −2 

Next, we compute the cofactor matrix:

  


 −1



 −1

 
 2 2  − 2   2 
  −2 −2   2 −2   2 −2  
   
        0 2 −2
  2 1   0 1   0 2  

cof(A) =  −    −   =  2 −2 4 .
 −2 −2   2 −2  2 −2  

  2 −1 2
       
 2 1   0 1   0 2 
  −    
 2 2  −1 2   −1 2 

The adjugate is the transpose of the cofactor matrix:

 
0 2 2
adj(A) = cof(A)T =  2 −2 −1  .
−2 4 2

We therefore have
   
0 2 2 0 1 1
1 1   
A−1 = adj(A) =  2 −2 −1  =  1 −1 − 21  .
det(A) 2
−2 4 2 −1 2 1

Once again, we double-check our work by computing A−1 A:

    
0 1 1 0 2 1 1 0 0
    
A−1 A =  1 −1 − 12   −1 2 2  =  0 1 0 .
−1 2 1 2 −2 −2 0 0 1

♠
It is always a good idea to double-check your work. At the end of the calculation, it is very easy to compute
A−1 A and check whether it is equal to I. If they are not equal, be sure to go back and double-check each
step. One common mistake is to forget to take the transpose of the cofactor matrix, so be sure not to forget
this step.
In practice, it is usually much faster to compute the inverse by the method of Section 4.5.2, because
this only requires solving a single system of equations, rather than computing a large number of cofactors.
However, there are some situations where the adjugate formula is useful. One such situation is when the
matrix has complicated entries that are functions rather than numbers. The following example illustrates
this.
 7.6. Application: A formula for the inverse of a matrix 263

Example 7.32: Inverse for non-constant matrix

Let  
et 0 0
A(t) =  0 cost sint  .
0 − sint cost
Show that A(t)−1 exists and find it.

Solution. First note that

det(A(t)) = et (cos2 t + sin2 t) = et 6= 0.
Therefore A(t)−1 exists for all values of the variable t. The cofactor matrix is
 
1 0 0
cof(A(t)) =  0 et cost et sint  .
0 −et sint et cost

The adjugate is the transpose of the cofactor matrix, and therefore the inverse is
   −t 
1 0 0 e 0 0
1 1
A(t)−1 = adj(A(t)) = t  0 et cost −et sint  =  0 cost − sint  .
det(A(t)) e
0 et sint et cost 0 sint cost

♠
Another situation where the adjugate formula is useful is the case of a 2 × 2-matrix. In this case both the
determinant and the adjugate are especially easy to compute. For a 2 × 2-matrix
 
a b
A = ,
c d

we have

det(A) = ad − bc
 
d −b
adj(A) = .
−c a

Therefore, A is invertible if and only if ad − bc 6= 0, and in that case, the inverse is given by
 
−1 1 d −b
A = . (7.1)
ad − bc −c a

Example 7.33: Inverse of a 2 × 2-matrix

Find the inverse of  
7 5
A = .
2 2
264 Determinants

Solution. We use formula (7.1) to compute the inverse:

   
1 5
1 2 −5 −4
A−1 =   =  2 .
7 · 2 − 2 · 5 −2 7 − 12 7
4

Exercises

Exercise 7.6.1 Find the cofactor matrix and the adjugate of each of the following matrices.
   
  1 0 1 1 1 3
1 0
A= , B =  −1 1 −2  , C =  2 3 1 .
3 2
2 −1 3 1 −1 1

Exercise 7.6.2 For each of the following matrices, determine whether it is invertible by checking whether
the determinant is non-zero. If the determinant is non-zero, use the adjugate formula to find the inverse.
         
1 2 3 1 2 0 1 3 3 1 2 3 1 0 3
A =  0 2 1 , B =  0 2 1 , C =  2 4 1 , D =  0 2 1 , E =  1 0 1 .
3 1 0 3 1 1 0 1 1 2 6 7 3 1 0

Exercise 7.6.3 Determine whether each of the following matrices is invertible. If so, use the adjugate
formula to find the inverse. If the inverse does not exist, explain why.
   
  1 2 3 1 2 1
1 1
A= , B =  0 2 1 , C =  2 3 0 .
1 2
4 1 1 0 1 2

Exercise 7.6.4 Use the adjugate formula to find the inverse of the matrix
 
3 0 3
A =  −1 2 −3  .
−5 4 −3

Exercise 7.6.5 Use the adjugate formula to find the inverse of the matrix
 
1 1 0
A= 3 1 2 .
2 −2 5
 7.6. Application: A formula for the inverse of a matrix 265

Exercise 7.6.6 Consider the matrix

 
1 0 0
A =  0 cost − sint  .
0 sint cost

Does there exist a value of t for which this matrix fails to be invertible? Explain.

Exercise 7.6.7 Consider the matrix  

1 t t2
A =  0 1 2t  .
t 0 2
Does there exist a value of t for which this matrix fails to be invertible? Explain.

Exercise 7.6.8 Consider the matrix

 
et cosht sinht
A =  et sinht cosht  .
et cosht sinht

Does there exist a value of t for which this matrix fails to be invertible? Explain.

Exercise 7.6.9 Consider the matrix

 t 
e e−t cost e−t sint
A =  et −e−t cost − e−t sint −e−t sint + e−t cost  .
et 2e−t sint −2e−t cost

Does there exist a value of t for which this matrix fails to be invertible? Explain.

Exercise 7.6.10 Use the adjugate formula to find the inverse of the matrix
 t 
e 0 0
A= 0 cost sint .
0 cost − sint cost + sint

Exercise 7.6.11 Find the inverse, if it exists, of the matrix

 t 
e cost sint
A =  et − sint cost  .
et − cos t − sint
266 Determinants

7.7 Application: Cramer’s rule

Outcomes
A. Use Cramer’s rule to solve a system of equations with invertible coefficient matrix.

Another application of determinants is Cramer’s rule for solving a system of equations. Recall that we
can represent a system of linear equations in the form Ax = b, where x is a vector of variables. Cramer’s
rule gives a formula for the solutions x in the special case that the coefficient matrix A is a square invertible
matrix. Note that Cramer’s rule does not apply if you have a system of equations in which there is a
different number of equations than variables (in other words, when A is not square), or when A is not
invertible.

Theorem 7.34: Cramer’s rule

Suppose A is an invertible n × n-matrix and we wish to solve the system Ax = b, where x =
[x1 , . . . , xn ]. Then xi can be computed by the rule

det(Ai )
xi = ,
det(A)

where Ai is the matrix obtained by replacing the i th column of A with b.

Proof. Since A is invertible, the solution to the system Ax = b is given by x = A−1 b. By Theorem 7.29,
we have
1
A−1 = adj(A),
det(A)
and therefore
1
x= adj(A)b.
det(A)
Let xi be the i th component of x and b j be the j th component of b. Recall that the ij th entry of adj(A) is
C ji , the ji th cofactor of A. By definition of matrix multiplication, we have

1
xi = (C1i b1 + . . . +Cni bn ).
det(A)
By the formula for the expansion of a determinant along a column, this is equal to
 
 ∗ · · · b1 · · · ∗ 
1  . ..

..  ,
xi =  .. . . 
det(A) 
 ∗ · · · bn · · · ∗ 

where the i th column of A is replaced with the column vector b. But this last formula is exactly Cramer’s
rule. ♠
 7.7. Application: Cramer’s rule 267

Example 7.35: Using Cramer’s rule

Use Cramer’s rule to solve the system of equations
    
1 2 1 x 3
 3 2 1  y  =  5 .
1 4 1 z 6

Solution. The matrices A1 , A2 , and A3 are obtained by respectively replacing the first, second, and third
column of A by b. We compute
   
 1 2 1   3 2 1 
   
det(A) =  3 2 1  = 4, det(A1 ) =  5 2 1  = 4
 1 4 1   6 4 1 
   
 1 3 1   1 2 3 
   
det(A2 ) =  3 5 1  = 6, det(A3 ) =  3 2 5  = −4.
 1 6 1   1 4 6 
Then by Cramer’s rule,

det(A1 ) 4 det(A2 ) 6 3 det(A3 ) −4

x= = = 1, y= = = , and z= = = −1.
det(A) 4 det(A) 4 2 det(A) 4

Thus, the solution is (x, y, z) = (1, 32 , −1). ♠

Cramer’s rule is sometimes useful in situations where row operations would be difficult to do. One such
situation is when a system of equations involves functions rather than numbers, as in the following exam-
ple.

Example 7.36: Using Cramer’s rule for non-constant matrix

Solve the following system of equations for z.
    
1 0 0 x 1
 0 et cost et sint   y  =  t  .
0 −et sint et cost z t2

Solution. We are asked to find the value of z in the solution. By Cramer’s rule, we have
 
 1 0 1 
 
 0 e cost t 
t
 
 0 −et sint t 2  et (t 2 cost − t sint)
z =   = = e−t (t 2 cost − t sint).
 1 0 0  e2t
 
 0 et cost et sint 
 
 0 −et sint et cost 

♠
268 Determinants

Exercises

Exercise 7.7.1 True or false? “Cramer’s rule is useful for finding solutions to systems of linear equations
in which there is an infinite set of solutions.”

Exercise 7.7.2 Use Cramer’s rule to find the solution to

x + 2y = 1
2x − y = 2

Exercise 7.7.3 Use Cramer’s rule to find the solution to

x + 2y + z = 3
2x − y − z = 1
x+z = 1

Exercise 7.7.4 Use Cramer’s rule to solve the system of equations

    
2 1 1 x 3
 2 2 −1   y  =  −2  .
2 1 2 z 5

Exercise 7.7.5 Find the value of y in the following system of equations:

    
1 1 1 x t
 1 t t2   y  =  s  .
1 s s2 z 1
8. Eigenvalues, eigenvectors, and diagonalization
In this chapter, we introduce the theory of eigenvalues and eigenvectors, and the technique of diagonal-
ization. In the same way that Gaussian elimination is a fundamental tool that permits us to solve many
different kinds of problems, diagonalization is also one of the fundamentals tool of linear algebra. It has a
great number of applications in every field of mathematics, science, and engineering.

8.1 Eigenvectors and eigenvalues

Outcomes
A. Determine whether a vector is an eigenvector of a matrix.

B. Given an eigenvector, find the corresponding eigenvalue.

C. Given an eigenvalue, find the corresponding eigenvectors.

D. Find a basis for the eigenspace of a given eigenvalue.

When we multiply a square matrix A by a non-zero vector v, we obtain another vector Av. Most of the
time, the vectors Av and v are unrelated; they could point in completely different directions. However,
sometimes it can happen that Av is a scalar multiple of v. In that case, v is called an eigenvector of A. We
will see later in this chapter that we can learn a lot about the matrix A by considering its eigenvectors.

Definition 8.1: Eigenvalues and eigenvectors

Let A be an n × n-matrix. Suppose that v ∈ Rn is a non-zero vector such that Av is a scalar multiple
of v. In other words, suppose that there exists a scalar λ such that

Av = λ v.

Then v is called an eigenvector of A, and λ is called the corresponding eigenvalue.

269
270 Eigenvalues, eigenvectors, and diagonalization

Example 8.2: Eigenvalues and eigenvectors

Consider the matrix  
3 −2 2
A= 1 2 1 .
0 2 1
Which of the following vectors are eigenvectors of A? Find the corresponding eigenvalues.
       
2 0 1 0
v1 = −1 , v2 = 1 , v3 = 1 , v4 = 0  .
      
−2 1 1 0

Solution. We compute
       
4 0 3 0
Av1 =  −2  , Av2 =  3  , Av3 =  4  , Av4 =  0  .
−4 3 3 0
• We see that Av1 is a scalar multiple of v1 , namely Av1 = 2v1 . Therefore, v1 is an eigenvector of A
with corresponding eigenvalue λ = 2.
• Similarly, Av2 = 3v2 , so v2 is an eigenvector of A with corresponding eigenvalue λ = 3.
• On the other hand, Av3 is not a scalar multiple of v3 . Hence, v3 is not an eigenvector of A.
• Finally, although Av4 is a scalar multiple of v4 , the zero vector is not considered an eigenvector.
♠

Example 8.3: Find eigenvectors for the given eigenvalue

Let  
2 0 0
A =  −1 3 1 .
2 −2 0
Find the eigenvectors corresponding to the eigenvalue λ = 2.

Solution. We have to solve the equation Av = 2v. We can use algebra to rewrite this as
Av = 2v ⇐⇒ Av − 2v = 0
⇐⇒ (A − 2I)v = 0
   
0 0 0 0
⇐⇒  −1 1 1 v =  0 .
2 −2 −2 0
This is a homogeneous system of equations with general solution
   
1 1
v = s 1 +t 0 ,
  
0 1
 8.1. Eigenvectors and eigenvalues 271

where s and t are parameters. These (except the zero vector) are exactly the eigenvectors corresponding to
the eigenvalue λ = 2. ♠
As the last example shows, the eigenvectors for a given eigenvalue λ , plus the zero vector, form a subspace
of Rn . This is called the eigenspace of λ .

Definition 8.4: Eigenspace

Let A be an n × n-matrix, and let λ be an eigenvalue of A. The eigenspace of λ is the set

Eλ = {v | Av = λ v} .

It is a subspace of Rn .

Instead of finding all eigenvectors for a given eigenvalue, it is often sufficient to find a basis for the
eigenspace. We also sometimes call the basis vectors of the eigenspace basic eigenvectors.

Example 8.5: Basis of eigenspace

Let  
2 0 0
A =  −1 3 1 .
2 −2 0
The matrix A has eigenvalues λ = 1 and λ = 2. Find a basis for each eigenspace.

Solution. We already found a basis for the eigenspace E2 in Example 8.3.

   
 1 1 
Basis of E2 :  
1 ,  0  .
 
0 1
To find a basis for the eigenspace E1 , we proceed analogously. We must solve the equation Av = 1v. We
have:
Av = 1v ⇐⇒ Av − v = 0
⇐⇒ (A − I)v = 0
   
1 0 0 0
⇐⇒  −1 2 1 v =  0 .
2 −2 −1 0
This is a homogeneous system of rank 2, with general solution
 
0
v = t  −1  .
2
Thus, the following is a basis for the eigenspace E1 :
 
 0 
Basis of E1 :  −1  .
 
2
272 Eigenvalues, eigenvectors, and diagonalization

Exercises

Exercise 8.1.1 Consider the matrix

 
1 −2 −2
A =  2 −3 −2  .
−2 2 1

Which of the following vectors are eigenvectors of A? Find the corresponding eigenvalues.
       
1 1 2 0
v1 =  1  , v2 =  1  , v3 =  2  , v4 =  −1  .
−1 0 −1 1

Exercise 8.1.2 Let  

1 0 0
A =  −5 −1 5  .
−3 0 4
Find the eigenvectors corresponding to the eigenvalue λ = 4.

Exercise 8.1.3 Let  

7 −4 8
A =  −1 4 −2  .
−2 2 −1
Find the eigenvectors corresponding to the eigenvalue λ = 3.

Exercise 8.1.4 Let  

4 0 3
A =  −3 1 −3  .
0 0 1
This matrix has eigenvalues λ = 1 and λ = 4. Find a basis for each eigenspace.

Exercise 8.1.5 Let  

2 4 −4
A =  −1 6 −9  .
0 0 −3
This matrix has eigenvalues λ = 3 and λ = 4. Find a basis for each eigenspace.
 8.2. Finding eigenvalues 273

Exercise 8.1.6 Suppose A is a 3×3-matrix with eigenvalues λ1 = 1, λ2 = 0, and λ3 = 2 and corresponding

eigenvectors      
−1 1 −1
v1 =  −2  , v2 =  1  , and v3 =  −4  .
−2 1 −3
(By “corresponding”, we mean that v1 corresponds to λ1 , v2 corresponds to λ2 , and so on). Find
 
3
A  −4  .
3

Exercise 8.1.7 Let A be an n × n-matrix, and assume λ is an eigenvalue of A. Show that λ 2 is an

eigenvalue of A2 .

Exercise 8.1.8 Let A be an invertible n × n-matrix, and assume λ is an eigenvalue of A. Show that λ 6= 0
and that λ −1 is an eigenvalue of A−1 .

Exercise 8.1.9 If A is an n × n-matrix and c is a non-zero constant, compare the eigenvalues of A and cA.

Exercise 8.1.10 Let A, B be invertible n × n-matrices which commute. That is, AB = BA. Suppose v is an
eigenvector of B. Show that then Av must also be an eigenvector for B.

Exercise 8.1.11 Suppose A is an n × n-matrix and it satisfies Am = A for some m a positive integer larger
than 1. Show that if λ is an eigenvalue of A then λ equals either 0, 1, or −1.

Exercise 8.1.12 Show that if Av = λ v and Aw = λ w, then whenever k, p are scalars,

A(kv + pw) = λ (kv + pw)

Does this imply that kv + pw is an eigenvector? Explain.

8.2 Finding eigenvalues

Outcomes
A. Find the characteristic polynomial, eigenvalues, and eigenvectors of a matrix.

B. Find the eigenvalues of a triangular matrix.

In the previous section, we saw how to find the eigenvectors corresponding to a given eigenvalue λ , if
λ is already known. But we have not yet seen how to find the eigenvalues of a matrix. However, the
calculations in Examples 8.3 and 8.5 suggest a way forward. We can see that the following are equivalent:
274 Eigenvalues, eigenvectors, and diagonalization

1. λ is an eigenvalue of A.

2. There exists a non-zero vector v such that Av = λ v.

3. The homogeneous system of equations (A − λ I)v = 0 has a non-trivial solution.

Indeed, the equivalence between 1 and 2 is just the definition of an eigenvalue, and the equivalence between
2 and 3 is just algebra. By Corollary 7.22, we know that the system (A − λ I)v = 0 has a non-trivial solution
if and only if det(A − λ I) = 0. Therefore, we have proved the following theorem:

Theorem 8.6: Eigenvalues

Let A be a square matrix, and let λ be a scalar. Then λ is an eigenvalue of A if and only if

det(A − λ I) = 0.

Example 8.7: Finding the eigenvalues

Find the eigenvalues of the matrix  
−5 2
A= .
−7 4

Solution. By Theorem 8.6, a scalar λ is an eigenvalue of A if and only if det(A − λ I) = 0. We calculate

the determinant:
 
 −5 − λ 2 
det(A − λ I) =  
−7 4−λ 
= (−5 − λ )(4 − λ ) + 14
= λ 2 + λ − 6.

Therefore, λ is an eigenvalue if and only if λ 2 + λ − 6 = 0. We can find the roots of this equation using
the quadratic formula, or equivalently, by factoring the left-hand side:

λ 2 + λ − 6 = 0 ⇐⇒ (λ + 3)(λ − 2) = 0.

Therefore, the eigenvalues are λ = −3 and λ = 2. ♠

Example 8.8: Finding the eigenvalues

Find the eigenvalues of the matrix
 
5 −4 4
A =  2 −1 2  .
0 0 2
 8.2. Finding eigenvalues 275

Solution. Once again, we calculate det(A − λ I):

 
 5−λ −4 4 
 
det(A − λ I) =  2 −1 − λ 2 
 0 0 2−λ 
= (5 − λ )(−1 − λ )(2 − λ ) − 2(−4)(2 − λ )
= −λ 3 + 6λ 2 − 11λ + 6
= (3 − λ )(1 − λ )(2 − λ ).
The eigenvalues are the roots of this polynomial, i.e., the solutions of the equation (λ −3)(λ −1)(2 − λ ) =
0. Therefore, the eigenvalues of A are λ = 1, λ = 2, and λ = 3. ♠

Example 8.9: No real eigenvalue

Find the eigenvalues of the matrix  
0 −1
A= .
1 0

Solution. We have  
 −λ −1 
det(A − λ I) =   = λ 2 + 1.

1 −λ
Since λ 2 + 1 = 0 does not have any solutions in the real numbers, the matrix A has no real eigenvalues.
(However, if we were working over the field of complex numbers rather than real numbers, this matrix
would have eigenvalues λ = ±i). ♠
As the examples show, the quantity det(A − λ I) is always a polynomial in the variable λ . A polynomial
is an expression of the form
p(λ ) = an λ n + an−1 λ n−1 + . . . + a1 λ + a0 ,
where a0 , . . . , an are constants called the coefficients of the polynomial. The polynomial det(A − λ I) has
a special name:

Definition 8.10: Characteristic polynomial

Let A be a square matrix. The expression

p(λ ) = det(A − λ I)

is called the characteristic polynomial of A.

Example 8.11: Characteristic polynomial

Find the characteristic polynomial of the matrix
 
3 0 2
A= 6 4 3 .
−4 0 −3
276 Eigenvalues, eigenvectors, and diagonalization

Solution. The characteristic polynomial is

det(A − λ I)
 
 3−λ 0 2 
 

=  6 4−λ 3 

 −4 0 −3 − λ 
   
 4−λ 3   6 4−λ 

= (3 − λ )   + 2  
0 −3 − λ  −4 0 
= (3 − λ )(4 − λ )(−3 − λ ) + 8(4 − λ )
= −λ 3 + 4λ 2 + λ − 4. ♠
It is time to summarize the method for finding the eigenvalues and eigenvectors of a matrix.

Procedure 8.12: Finding eigenvalues and eigenvectors

Let A be an n × n-matrix. To find the eigenvalues and eigenvectors of A:

1. Calculate the characteristic polynomial det(A − λ I).

2. The eigenvalues are the roots of the characteristic polynomial.

3. For each eigenvalue λ , find a basis for the eigenvectors by solving the homogeneous system

(A − λ I)v = 0.

To double-check your work, make sure that Av = λ v for each eigenvalue λ and associated eigen-
vector v.

Example 8.13: Finding eigenvalues and eigenvectors

Find the eigenvalues and eigenvectors of the matrix
 
3 0 2
A= 6 4 3 .
−4 0 −3

Solution. We already found the characteristic polynomial in Example 8.11. It is

p(λ ) = det(A − λ I) = −λ 3 + 4λ 2 + λ − 4.
Finding the roots of a cubic polynomial can be a bit tricky, but with some trial and error, we can find that
λ = 1 is a root. We can therefore factor out (λ − 1):
p(λ ) = (λ − 1)(−λ 2 + 3λ + 4).
Then we can use the quadratic formula to find the remaining two roots:
√
−3 ± 9 + 16
λ= ,
−2
 8.2. Finding eigenvalues 277

which yields the two roots λ = −1 and λ = 4. Therefore, we have

p(λ ) = −(λ − 1)(λ + 1)(λ − 4),

and the eigenvalues of A are λ = 1, λ = −1, and λ = 4. We now find the eigenvectors for each eigenvalue.

• For λ = 1: We must solve (A − I)v = 0, i.e.,

 
2 0 2
 6 3 3  v = 0.
−4 0 −4

The basic solution is  

−1
v1 =  1  .
1

• For λ = −1: We must solve (A + I)v = 0, i.e.,

 
4 0 2
 6 5 3  v = 0.
−4 0 −2

The basic solution is  

−1
v2 =  0  .
2

• For λ = 4: We must solve (A − 4I)v = 0, i.e.,

 
−1 0 2
 6 0 3  v = 0.
−4 0 −7

The basic solution is  

0
v3 =  1  .
0

Example 8.14: A zero eigenvalue

Let  
2 2 −2
A =  1 3 −1  .
−1 1 1
Find the eigenvalues and eigenvectors of A.
278 Eigenvalues, eigenvectors, and diagonalization

Solution. To find the eigenvalues of A, we first compute the characteristic polynomial.

 
 2−λ 2 −2 
 
det(A − λ I) =  1 3 − λ −1  = −λ 3 + 6λ 2 − 8λ .
 −1 1 1−λ 

You can verify that the roots of this polynomial are λ1 = 0, λ2 = 2, λ3 = 4. Notice that while eigenvectors
can never equal 0, it is possible to have an eigenvalue equal to 0. Now we will find the basic eigenvectors.

• For λ1 = 0: We must solve the equation (A − 0I)v = 0. This equation becomes Av = 0. We write
the augmented matrix for this system and reduce to echelon form:
   
2 2 −2 0 1 0 −1 0
 1 3 −1 0  ≃ . . . ≃  0 1 0 0 .
−1 1 1 0 0 0 0 0

The basic solution is 

1
v1 =  0  .
1

• For λ2 = 2: We solve the equation (A − 2I)v = 0:

   
0 2 −2 0 1 0 0 0
 1 1 −1 0  ≃ . . . ≃  0 1 −1 0  .
−1 1 −1 0 0 0 0 0

The basic solution is 

0
v2 =  1  .
1

• For λ3 = 4: We solve the equation (A − 4I)v = 0:

   
−2 2 −2 0 1 −1 0 0
 1 −1 −1 0  ≃ . . . ≃  0 0 1 0 .
−1 1 −3 0 0 0 0 0

The basic solution is 

1
v3 =  1  .
0

Thus we have found the eigenvectors v1 for λ1 , v2 for λ2 , and v3 for λ3 . We can double-check our answers
by checking the equation Av = λ v in each case:
    
2 2 −2 1 0
Av1 =  1 3 −1   0  =  0  = 0v1 ,
−1 1 1 1 0
 8.2. Finding eigenvalues 279

    
2 2 −2 0 0
Av2 =  1 3 −1   1 = 2  = 2v2 ,
−1 1 1 1 2
    
2 2 −2 1 4
Av3 =  1 3 −1   1  =  4  = 4v3 .
−1 1 1 0 0

Therefore, our eigenvectors and eigenvalues are correct. ♠

We conclude this section by considering the eigenvalues of a triangular matrix. Recall from Definition 7.13
that a matrix is upper triangular if all entries below the main diagonal are zero, and lower triangular if
all entries above the main diagonal are zero.

Example 8.15: Eigenvalues of a triangular matrix

Find the eigenvalues of  
1 2 4
A =  0 4 7 .
0 0 6

Solution. We calculate det(A − λ I) = 0 as follows:

 
1−λ 2 4
det(A − λ I) = det  0 4−λ 7  = (1 − λ )(4 − λ )(6 − λ ).
0 0 6−λ

Solving the equation (1 − λ )(4 − λ )(6 − λ ) = 0 results in the eigenvalues λ1 = 1, λ2 = 4, and λ3 = 6.

Thus the eigenvalues are the entries on the main diagonal of A. ♠
Clearly, the same is true for any (upper or lower) triangular matrix. We therefore have the following
proposition:

Proposition 8.16: Eigenvalues of a triangular matrix

Let A be an upper or lower triangular matrix. Then the eigenvalues of A are the entries on the main
diagonal.

Exercises

Exercise 8.2.1 Find the characteristic polynomial of the matrix

 
3 −2
.
1 0
280 Eigenvalues, eigenvectors, and diagonalization

Use the quadratic formula to find the eigenvalues.

Exercise 8.2.2 Find the characteristic polynomial, eigenvalues, and basic eigenvectors of the matrix
 
9 10
.
−5 −6

Exercise 8.2.3 Find the characteristic polynomial, eigenvalues, and basic eigenvectors of the matrix
 
0 3 −1
 −2 4 −2  .
2 −3 3

One eigenvalue is 1.

Exercise 8.2.4 Find the characteristic polynomial, eigenvalues, and basic eigenvectors of the matrix
 
3 0 −2
 −2 1 2 .
0 0 1

One eigenvalue is 3.

Exercise 8.2.5 Find the characteristic polynomial, eigenvalues, and basic eigenvectors of the matrix
 
9 2 8
 2 −6 −2  .
−8 2 −5

One eigenvalue is −3.

Exercise 8.2.6 Which of the following matrices have no real eigenvalue?

     
1 1 1 −1 0 1
A= , B= , C= .
1 −1 1 1 1 0

Exercise 8.2.7 Find the eigenvalues and eigenvectors of the following triangular matrix:
 
3 2 2
 0 1 −2  .
0 0 −1

Exercise 8.2.8 Is it possible for a non-zero matrix to have only 0 as an eigenvalue?

 8.3. Geometric interpretation of eigenvectors 281

8.3 Geometric interpretation of eigenvectors

Outcomes
A. Visualize the effect of a linear transformation by considering its eigenvectors and eigenvalues.

Consider the matrix  

2 1
A= .
1 2
In Chapter 6, we saw that this matrix corresponds to a linear transformation T : R2 → R2 defined by
T (x) = Ax. We also saw how to visualize this linear transformation as a before-and-after picture. For this,
we considered the images of the first and second standard basis vectors:
 
2
T (e1 ) = Ae1 = ,
1
 
1
T (e2 ) = Ae2 = .
2
Here is the before-and-after picture for this transformation: h i
T (e2 ) = 12

e2 h i
T (e1 ) = 21
T
e1 7−→

Although we can see from this picture that the letter “F” is being distorted somehow, it is perhaps not very
obvious what exactly this linear transformation does.
We can get a much better idea by computing the eigenvectors and eigenvalues of A. A short calculation
shows that the basic eigenvectors are
   
1 −1
v1 = and v2 = ,
1 1
with corresponding eigenvalues λ1 = 3 and λ2 = 1. Consider the effect of the linear transformation T on
the eigenvectors:
    
2 1 1 3
T (v1 ) = = = 3v1 ,
1 2 1 3
282 Eigenvalues, eigenvectors, and diagonalization

    
2 1 −1 −1
T (v2 ) = = = v2 .
1 2 1 1
So each eigenvector is mapped to a scalar multiple of itself. This gives us a hint for how to draw a more
useful before-and-after picture. Rather than tracking the movement of the standard basis vectors e1 and
~e2 , let us track the movement of the eigenvectors v1 and v2 instead:

T (v1 ) = 3v1

v2 v1 T (v2 ) = v2
T
7−→

Thus, the linear transformation described by the matrix A is revealed to be just a scaling by a factor of 3
along the direction of  
1
v1 = .
1
In summary, the geometric meaning of an eigenvector is that it is mapped to a multiple of itself. Thus,
when viewed from the point of view of its action on the eigenvectors, a linear transformation behaves like
a scaling of each eigenvector. We can say that each eigenvector describes a direction of scaling, and each
corresponding eigenvalue giving the corresponding (positive or negative) scaling factor.

Example 8.17: Visualize a linear transformation

Visualize the linear transformation T : R2 → R2 that is described by the matrix
 
0 1
A=
1 0

by considering the eigenvectors and eigenvalues.

Solution. The characteristic polynomial is λ 2 − 1, and so the eigenvalues are λ1 = 1 and λ2 = −1. By
solving each equation (A − λ I)v = 0, we find that the corresponding basic eigenvectors are
   
1 −1
v1 = and v2 =
1 1
(check this!). We get the following before-and-after picture:
 8.3. Geometric interpretation of eigenvectors 283

T (v1 ) = v1
v2 v1

T
7−→

T (v2 ) = −v2

We see that this linear transformation is a reflection about the vector v1 . ♠

Example 8.18: Visualize a linear transformation

Visualize the linear transformation T : R2 → R2 that is described by the matrix
 
−2 3
A=
0 1

by considering the eigenvectors and eigenvalues.

Solution. The characteristic polynomial is (−2 − λ )(1 − λ ), and therefore, the eigenvalues are λ1 = −2
and λ2 = 1. We find that the corresponding basic eigenvectors are
   
1 1
v1 = and v2 = ,
0 1

respectively. We get the following before-and-after picture:

v2
T (v2 ) = v2
T
7−→
v1 T (v1 ) = −2v1
284 Eigenvalues, eigenvectors, and diagonalization

This particular linear transformation keeps the vector v2 fixed, while scaling by a factor of −2 in the
direction of v1 . It could be described as a kind of slanted reflection with scaling. ♠

Exercises

Exercise 8.3.1 For each of the following matrices, find the eigenvectors and eigenvalues. Use this to
visualize the linear transformation T : R2 → R2 that is described by the matrix.
         
1 2 1 0 1 1 1 1 1 0
A= , B= , C= , D= , E= .
2 1 2 −1 0 2 1 1 −1 2

Exercise 8.3.2 If A is the matrix of a linear transformation that rotates all vectors in R2 by 60◦ , explain
why A cannot have any real eigenvalues. Is there an angle such that rotation by this angle would have a
real eigenvalue? What eigenvalues would be obtainable in this way?

Exercise 8.3.3 Let T be the linear transformation that reflects vectors about the x-axis. Find a matrix for
T and then find its eigenvalues and eigenvectors.

Exercise 8.3.4 Let T be the linear transformation that reflects vectors about the line x = y. Find a matrix
of T and then find eigenvalues and eigenvectors.

Exercise 8.3.5 Let T be the linear transformation that reflects all vectors in R3 about the xy-plane. Find
a matrix for T and then obtain its eigenvalues and eigenvectors.

8.4 Diagonalization

Outcomes
A. Compute sums, products, and powers of diagonal matrices.

B. Determine whether a square matrix is diagonalizable, and diagonalize it if possible.

A square matrix D is called a diagonal matrix if all entries except those on the main diagonal are zero.
Such matrices look like the following:
 
d11 0 · · · 0
 0 d22 · · · 0 
 
 .. .. . . ..  .
 . . . . 
0 0 · · · dnn
 8.4. Diagonalization 285

Diagonal matrices are particularly easy to work with. For example, the sum of two diagonal matrices is
diagonal. Also, the product of two diagonal matrices is diagonal, and is computed by taking the product
of corresponding diagonal entries.

Example 8.19: Sums, products, and powers of diagonal matrices

Compute A + B, AB, and A4 , where
   
2 0 0 1 0 0
A= 0 3 0  and B =  0 −2 0  .
0 0 4 0 0 2

Solution. We have    
2+1 0 0 3 0 0
A+B =  0 3−2 0 = 0 1 0 ,
0 0 4+2 0 0 6
   
2·1 0 0 2 0 0
AB =  0 3 · (−2) 0 = 0 −6 0  ,
0 0 4·2 0 0 8
and    
24 0 0 16 0 0
A4 =  0 34 0  =  0 81 0  .
0 0 44 0 0 256
Notice that all operations are computed componentwise on the diagonal. Therefore, multiplication of
diagonal matrices is much simpler than multiplication of general matrices. ♠
One of the most important problem solving techniques in linear algebra is diagonalization. In a nutshell,
the point of diagonalization is to simplify a problem by replacing an arbitrary matrix by a diagonal matrix.
We say that two square matrices A and B are similar if there exists an invertible matrix P such that
P−1 AP = B. A matrix is diagonalizable if it is similar to a diagonal matrix. This is summarized in the
following definition.

Definition 8.20: Diagonalizable matrix

Let A be an n × n-matrix. Then A is said to be diagonalizable if there exists an invertible matrix P
and a diagonal matrix D such that
P−1 AP = D.

The key connection between diagonalizability, eigenvectors, and eigenvalues is the following theorem.

Theorem 8.21: Diagonalization and eigenvectors

An n × n-matrix A is diagonalizable if and only if A has n linearly independent eigenvectors.

Moreover, in this case, let P be the invertible matrix whose columns are n linearly independent
eigenvectors of A, and let D be the diagonal matrix whose diagonal entries are the corresponding
eigenvalues. Then P−1 AP = D.
286 Eigenvalues, eigenvectors, and diagonalization

Proof. Assume that A has n linearly independent eigenvectors v1 , . . . , vn . Let λ1 , . . ., λn be the correspond-
ing eigenvalues, so that
Avi = λi vi (8.1)
for all i = 1, . . ., n. Let P be the matrix that has v1 , . . . , vn as its columns. Then P is invertible because
v1 , . . . , vn are linearly independent. Let D be the diagonal matrix that has λ1 , . . . , λn as its diagonal entries.
By the column method of matrix multiplication, the i th column of AP is Avi . Also by the column method
of matrix multiplication, the i th column of PD is λi vi . Therefore, by (8.1), the matrices AP and PD have
the same columns, i.e.,
AP = PD.
It follows that P−1 AP = D, as desired.
Conversely, assume that A is diagonalizable. Then there exists an invertible matrix P and a diagonal
matrix D such that P−1 AP = D, or equivalently, AP = PD. Let v1 , . . . , vn be the columns of P and let
λ1 , . . . , λn be the diagonal entries of D. Again we find that the i th column of AP is Avi and the i th column
of PD is λi vi , and therefore Avi = λi vi holds for all i. It follows that v1 , . . . , vn are eigenvectors of A. Since
P is invertible, v1 , . . . , vn are linearly independent, so A has n linearly independent eigenvectors. ♠

Example 8.22: Diagonalizing a matrix

Diagonalize the matrix  
3 0 2
A= 6 4 3 .
−4 0 −3
In other words, find an invertible matrix P and a diagonal matrix D such that P−1 AP = D.

Solution. By Theorem 8.21, we use the eigenvectors of A as the columns of P and the corresponding eigen-
values as the diagonal entries of D. We already found the eigenvectors and -values of A in Example 8.13.
They were      
−1 −1 0
v1 =  
1 . v2 =  0 . and v3 = 1  .
 
1 2 0
with corresponding eigenvalues λ1 = 1, λ2 = −1, and λ3 = 4. Therefore we can use
   
−1 −1 0 1 0 0
P= 1 0 1  and D =  0 −1 0  .
1 2 0 0 0 4
To double-check that P−1 AP is indeed equal to D, we first compute the inverse of P:
 
−2 0 −1
P−1 =  1 0 1 .
2 1 1
Then
     
−2 0 −1 3 0 2 −1 −1 0 1 0 0
P−1 AP =  1 0 1  6 4 3  1 0 1  =  0 −1 0  = D.
2 1 1 −4 0 −3 1 2 0 0 0 4
 8.4. Diagonalization 287

Alternatively, we could have checked that AP = PD, which would not have required computing P−1 . ♠

Example 8.23: Diagonalizing a matrix

Diagonalize the matrix  
2 0 0
A= 1 4 −1  .
−2 −4 4

Solution. First, we will find the characteristic polynomial of A:

 
 2−λ 0 0 

det(A − λ I) =  1 4 − λ −1 
 −2 −4 4 − λ 
 
 4 − λ −1 
= (2 − λ )  
−4 4 − λ 
= (2 − λ ) ((4 − λ )(4 − λ ) − 4)
= (2 − λ )(12 − 8λ + λ 2 )
= (2 − λ )(2 − λ )(6 − λ ).
Therefore, the eigenvalues are λ = 2 and λ = 6. Next, we need to find the eigenvectors. We first find the
eigenvectors for λ = 2. We solve (A − 2I)v = 0:
   
0 0 0 0 1 2 −1 0
 1 2 −1 0  ≃  0 0 0 0 .
−2 −4 2 0 0 0 0 0
The general solution is    
−2 1
t  1 +s 0 ,
 
0 1
where t, s are parameters. Thus, the basic eigenvectors for λ = 2 are
   
−2 1
v1 =  1  and v2 = 0  .

0 1
Doing a similar calculation, we find that the basic eigenvector for λ = 6 is
 
0
v3 =  1  .
−2
By Theorem 8.21, we use the eigenvectors of A as the columns of P and the corresponding eigenvalues as
the diagonal entries of D. Therefore,
   
−2 1 0 2 0 0
P= 1 0 1  and D =  0 2 0  .
0 1 −2 0 0 6
288 Eigenvalues, eigenvectors, and diagonalization

We can double-check this answer by computing

     
− 14 12 1
4  2 0 0 −2 1 0 2 0 0
    
P AP = 
−1
 2 1
1 1 
2  1 4 −1  
 1 0 1  =  0 2 0  = D.
  
1 1 1
4 2 −4 −2 −4 4 0 1 −2 0 0 6

Notice that the eigenvalues on the main diagonal of D must be in the same order as the corresponding
eigenvectors in P. Since the eigenvectors v1 and v2 are both for the eigenvalue λ = 2, the entry 2 appears
twice in the matrix D. ♠
The following example shows that not all matrices are diagonalizable.

Example 8.24: A matrix that cannot be diagonalized

 
1 1
Show that the matrix A = cannot be diagonalized.
0 1

Solution. Through the usual procedure, we find that the characteristic polynomial is (1− λ )2 , and therefore
the only eigenvalue is λ = 1. To find the eigenvectors, we solve the equation (A − I)v = 0:
 
0 1 0
.
0 0 0
The general solution is  
1
v=t .
0
Because the solution space is 1-dimensional, there is only one basic eigenvector:
 
1
v1 = .
0
Since the matrix A has only one basic eigenvector, we cannot find two linearly independent eigenvectors.
Therefore, by Theorem 8.21, A cannot be diagonalized. ♠

Exercises
   
−2 0 0 1 0 0
Exercise 8.4.1 Let D =  0 1 0  and E =  0 −1 0 . Find D + E, DE, and D7 .
0 0 2 0 0 3

Exercise 8.4.2 Find the eigenvalues and eigenvectors of the matrix

 
−13 −28 28
 4 9 −8  .
−4 −8 9
 8.4. Diagonalization 289

One eigenvalue is 3. Diagonalize if possible.

Exercise 8.4.3 Find the eigenvalues and eigenvectors of the matrix

 
5 −18 −32
 0 5 4 .
2 −5 −11

One eigenvalue is 1. Diagonalize if possible.

Exercise 8.4.4 Find the eigenvalues and eigenvectors of the matrix

 
8 0 10
 −6 −3 −6  .
−5 0 −7

One eigenvalue is −3. Diagonalize if possible.

Exercise 8.4.5 Find the eigenvalues and eigenvectors of the matrix

 
−1 −2 2
 0 5 −8  .
0 4 −7

One eigenvalue is 1. Diagonalize if possible.

Exercise 8.4.6 Find the eigenvalues and eigenvectors of the matrix

 
2 −1 6
 −4 −1 −6  .
−2 1 −6

One eigenvalue is 0. Diagonalize if possible.

Exercise 8.4.7 Find the eigenvalues and eigenvectors of the matrix

 
3 −1 0
 1 4 −1 
−1 −1 4

One eigenvalue is 3. Diagonalize if possible.

290 Eigenvalues, eigenvectors, and diagonalization

8.5 Application: Matrix powers

Outcomes
A. Use diagonalization to raise a matrix to a high power.

B. Use diagonalization to compute a square root of a matrix.

Suppose we have a matrix A and we want to find A50 . One could try to multiply A with itself 50 times,
but this is a lot of work (try it!). However, diagonalization allows us to compute high powers of a matrix
relatively easily. Suppose A is diagonalizable, so that P−1 AP = D. We can rearrange this equation to write
A = PDP−1 . Now, consider A2 . Since A = PDP−1 , it follows that
A2 = (PDP−1 )2 = PDP−1 PDP−1 = PD2 P−1 .
Similarly,
A3 = (PDP−1 )3 = PDP−1 PDP−1 PDP−1 = PD3 P−1 .
In general,
An = (PDP−1 )n = PDn P−1 .
Therefore, we have reduced the problem to finding Dn . But as we saw in Example 8.19, computing a
power of a diagonal matrix is easy. To compute Dn , we only need to raise every entry on the diagonal to
the power of n. Through this method, we can compute large powers of matrices.

Example 8.25: Raising a matrix to a high power

 
2 1 0
Let A =  0 1 0 . Find A50 .
−1 −1 1

Solution. First, we will diagonalize A. Following the usual steps, we find that the eigenvalues are λ = 1
and λ = 2. The basic eigenvectors corresponding to λ = 1 are
   
0 −1
v1 =  0  and v2 =  1  ,
1 0
and the basic eigenvector corresponding to λ = 2 is
 
−1
v3 =  0  .
1
Now we construct P by using the basic eigenvectors of A as the columns of P. Thus
 
0 −1 −1
P= 0 1 0 .
1 0 1
 8.5. Application: Matrix powers 291

The inverse of P is  
1 1 1
P−1 =  0 1 0 .
−1 −1 0
Then
     
1 1 1 2 1 0 0 −1 −1 1 0 0
P−1 AP =  0 1 0  0 1 0  0 1 0  =  0 1 0  = D.
−1 −1 0 −1 −1 1 1 0 1 0 0 2

Now it follows by rearranging the equation that A = PDP−1 , and therefore, as noted above,
   50  
0 −1 −1 1 0 0 1 1 1
A50 = PD50 P−1 =  0 1 0   0 150 0  0 1 0 
1 0 1 0 0 2 50 −1 −1 0
 
250 −1 + 250 0
=  0 1 0 .
1−2 50 1−2 50 1 ♠
Thus, through diagonalization, we have efficiently computed a high power of A. The following example
shows that we can also use the same technique for finding a square root of a matrix.

Example 8.26: Square root of a matrix

 
1 3 3
Let A =  −1 5 3 . Find a square root of A, i.e., find a matrix B such that A = B2 .
1 −1 1

Solution. We first diagonalize A. The characteristic polynomial is

 
 1−λ 3 3 

 −1 5 − λ 3  = (1 − λ )(5 − λ )(1 − λ ) + 9 + 3 − 3(5 − λ ) + 3(1 − λ ) + 3(1 − λ )

 1 −1 1 − λ 
= −λ 3 + 7λ 2 − 14λ + 8,

with roots λ = 1, λ = 2, and λ = 4. The corresponding eigenvectors are

     
1 0 1
v1 =  
1 , v2 =  1 , and v3 = 1  ,
 
−1 −1 0

respectively. Therefore we have P−1 AP = D, where

     
1 0 1 1 −1 −1 1 0 0
P= 1 1 1  , P−1 =  −1 1 0 , and D =  0 2 0  .
−1 −1 0 0 1 1 0 0 4
292 Eigenvalues, eigenvectors, and diagonalization

We can equivalently write A = PDP−1 . Finding a square root of a diagonal matrix is easy:
 
1 √0 0
1
D2 =  0 2 0 .
0 0 2
1 1 1
If we now define B = PD 2 P−1 , we clearly have B2 = PD 2 P−1 PD 2 P−1 = PDP−1 = A. So the desired
square root of A is
     
1 0 1 1 √0 0 1 −1 −1 1√ 1√ 1
1
B = PD 2 P−1 =  1 1 1  0 2 0   −1 1 0  =  1 − √2 1 + √2 1  .
−1 −1 0 0 0 2 0 1 1 −1 + 2 1 − 2 1

Finally, we verify that we have computed B correctly by squaring it and double-checking that we really
get A.
    
1√ 1√ 1 1√ 1√ 1 1 3 3
B2 =  1 − √2 1 + √2 1   1 − √2 1 + √2 1  =  −1 5 3  = A.
−1 + 2 1 − 2 1 −1 + 2 1 − 2 1 1 −1 1

We note that the square root of a matrix is not unique. In fact, D has 8 different square roots, all of the
form  
±1 0
√ 0
 0 ± 2 0 .
0 0 ±2
It follows that A has 8 different square roots as well. We leave it as an exercise to compute them all. ♠
The same method can also be used to compute other powers of a matrix, for example a cube root.

Exercises

 
1 2
Exercise 8.5.1 Let A = . Find A10 by diagonalization.
2 1
 
1 4 1
Exercise 8.5.2 Let A =  0 2 5 . Find A50 by diagonalization.
0 0 5
 
1 −2 −1
Exercise 8.5.3 Let A =  2 −1 1 . Find A100 by diagonalization.
−2 3 1
 
−5 −6
Exercise 8.5.4 Let A = . Find a square root of A, i.e., find a matrix B such that B2 = A.
9 10
 8.6. Application: Solving recurrences 293

 
−2 0 6
Exercise 8.5.5 Let A =  −3 1 6 . Find a square root of A.
−3 0 7

8.6 Application: Solving recurrences

Outcomes
A. Solve a linear recurrence relation using diagonalization.

Consider the following sequence of integers, called the Fibonacci sequence:

0, 1, 1, 2, 3, 5, 8, 13, 21, . . .

The first two Fibonacci numbers are 0 and 1. Every subsequent Fibonacci number is the sum of the
previous two numbers. For example, 0 + 1 = 1, 1 + 1 = 2, 1 + 2 = 3, 2 + 3 = 5, and so on. Thus, if we
write Fn for the n th Fibonacci number, then the Fibonacci sequence is given by the following conditions:

F0 = 0,
F1 = 1,
Fn+2 = Fn + Fn+1 , for all n ≥ 0.

The condition Fn+2 = Fn +Fn+1 is known as a recurrence relation, or simply as a recurrence, because we
compute each member of the sequence from previous members (the word “recurrence” comes from Latin
“recurrere”, which means “to go back”). The conditions F0 = 0 and F1 = 1 are known as the base cases of
the recurrence. Note that we start counting from zero, i.e., we call F0 = 0 the “zeroth Fibonacci number”,
F1 = 1 the “first Fibonacci number”, and so on. Counting from zero will help simplify our calculations
later.

Example 8.27: Computing a Fibonacci number

Compute the 10 th Fibonacci number F10 .

Solution. To compute the 10 th Fibonacci number using the recurrence, we have to compute all the previous
Fibonacci numbers as well.

F0 = 0,
F1 = 1,
F2 = F0 + F1 = 0+1 = 1,
F3 = F1 + F2 = 1+1 = 2,
F4 = F2 + F3 = 1+2 = 3,
F5 = F3 + F4 = 2+3 = 5,
294 Eigenvalues, eigenvectors, and diagonalization

F6 = F4 + F5 = 3 + 5 = 8,
F7 = F5 + F6 = 5 + 8 = 13,
F8 = F6 + F7 = 8 + 13 = 21,
F9 = F7 + F8 = 13 + 21 = 34,
F10 = F8 + F9 = 21 + 34 = 55.

Thus, the 10 th Fibonacci number is 55. ♠

Suppose we want to compute the 100 th Fibonacci number. As the previous example shows, computing
this by using the recurrence relation would be lot of work. We will therefore explore how to use linear
algebra, and in particular diagonalization, to find a closed formula for the n th Fibonacci number. By a
closed formula, we mean a formula to calculate Fn directly in one step, i.e., without using a recurrence.
The process of finding a closed formula is called solving the recurrence.
The first step in solving the recurrence is to replace the recurrence relation Fn+2 = Fn + Fn+1 , which
requires two previous terms of the sequence, by another recurrence relation requiring only one previous
term. To that end, we define vn to be the vector consisting of the n th and n + 1 st Fibonacci numbers, for
all n ≥ 0:  
Fn
vn = .
Fn+1
Using the recurrence relation for Fn+2 , we have
        
Fn+1 Fn+1 0 1 Fn 0 1
vn+1 = = = = vn .
Fn+2 Fn + Fn+1 1 1 Fn+1 1 1

Therefore, to compute vn+1 , we only need to know vn (and not vn−1 ). Let
 
0 1
A= .
1 1

Since vn is obtained from v0 by applying the matrix A n times, we have vn = An v0 for all n ≥ 0. We can
therefore get a formula for vn , and thus for Fn , by diagonalizing the matrix A.

Problem 8.28: Diagonalizing A

 
0 1
Diagonalize the matrix A = .
1 1

Solution. The characteristic polynomial is

 
 −λ 1 
p(λ ) =   = (−λ )(1 − λ ) − 1 = λ 2 − λ − 1.

1 1−λ

The eigenvalues are the roots of the characteristic polynomial. We find them by using the quadratic
formula. The eigenvalues are √ √
1+ 5 1− 5
λ1 = and λ2 = .
2 2
 8.6. Application: Solving recurrences 295

To simplify later calculations, we note that λ1 + λ2 = 1, or equivalently,

1 − λ1 = λ2 . (8.2)

We also note that λ1 λ2 = −1, or equivalently,

1
λ2 = − . (8.3)
λ1
To find the eigenvector corresponding to the eigenvalue λ1 , we solve the equation
 
−λ1 1
v = 0.
1 1 − λ1
By (8.2) and (8.3), this is equivalent to
 
−λ1 1
v = 0,
0 − λ1
1

and we find the basic solution  

1
u= .
λ1
By a similar method, we find that the basic eigenvector corresponding to the eigenvalue λ2 is
 
1
w= .
λ2

Therefore, by Theorem 8.21, A is diagonalizable, and we have P−1 AP = D, where

   
1 1 λ1 0
P= and D = .
λ1 λ2 0 λ2
For later reference, we note that the inverse of P is given by
 
−1 1 −λ2 1
P =√ .
5 λ1 −1
♠
We are now ready to find our formula for the n th Fibonacci number.

Problem 8.29: Solving the recurrence

Find a formula for the n th Fibonacci number.

Solution. Since  
Fn
vn = ,
Fn+1
we know that the n th Fibonacci number is the first component of vn , i.e.,
 
Fn = 1 0 vn .
296 Eigenvalues, eigenvectors, and diagonalization

Putting together all of the above calculations, we then have:

 
Fn = 1 0 vn
 
= 1 0 An v0
 
= 1 0 PDn P−1 v0
  n    
  1 1 λ1 0 1 −λ2 1 0
= 1 0 √
λ1 λ2 0 λ2n 5 λ 1 −1 1
 n  
1   λ
1 0 1
= √ 1 1 n
5 0 λ2 −1
1
= √ (λ1n − λ2n ).
5

So the n th Fibonacci number is

√ !n √ !n !
1 1 1+ 5 1− 5
Fn = √ (λ1n − λ2n ) = √ − .
5 5 2 2

Example 8.30: Computing the 100 th Fibonacci number

Calculate the 100 th Fibonacci number without using the recurrence.

Solution. Note that this calculation requires about 25 digits of precision to give the correct result. We have
1
F100 = √ (λ1100 − λ2100 )
5
1
= √ (1.6180339887498948482045868100 − (−0.6180339887498948482045868)100)
5
= 354224848179261915075.

♠
We can use the same method to solve other linear recurrences. Here is another example:

Example 8.31: Solving a recurrence

Consider the sequence of numbers defined by the recurrence

b0 = 1,
b1 = 2,
bn+2 = 6bn + bn+1 , for all n ≥ 0.

Find the first 5 members of this sequence. Then solve the recurrence and find b20 .
 8.6. Application: Solving recurrences 297

Solution. The first 5 members of the sequence are:

b0 = 1,
b1 = 2,
b2 = 6b0 + b1 = 6 + 2 = 8,
b3 = 6b1 + b2 = 12 + 8 = 20,
b4 = 6b2 + b3 = 48 + 20 = 68.
To solve the recurrence, let  
bn
wn = ,
bn+1
so that for all n ≥ 0,
        
bn+1 bn+1 0 1 bn 0 1
wn+1 = = = = wn .
bn+2 6bn + bn+1 6 1 bn+1 6 1
We then diagonalize the matrix  
0 1
B= .
6 1
The characteristic polynomial is
p(λ ) = λ 2 − λ − 6,
with roots λ = −2 and λ = 3. The respective eigenvectors are
 
1
v= for the eigenvalue λ = −2,
−2
 
1
u= for the eigenvalue λ = 3.
3

Therefore, B = PDP−1 , where

   
1 1 −2 0
P= and D= .
−2 3 0 3
The inverse of P is  
−1 1 3 −1
P = .
5 2 1
Finally, we use this information to solve the recurrence:
 
bn = 1 0 wn
 
= 1 0 B n w0
 
= 1 0 PDn P−1 w0
     
  1 1 (−2)n 0 1 3 −1 1
= 1 0 n
−2 3 0 3 5 2 1 2
  
1  (−2) 0n 1
= 1 1 n
5 0 3 4
n
(−2) + 4 · 3 n
=
5
298 Eigenvalues, eigenvectors, and diagonalization

Finally, we calculate

(−2)20 + 4 · 320 1048576 + 4 · 3486784401

b20 = = = 2789637236.
5 5

Exercises

Exercise 8.6.1 Consider the sequence of numbers defined by the recurrence

a0 = 0,
a1 = 1,
an+2 = 2an+1 + 3an , for all n ≥ 0.

Find the first 5 members of this sequence. Then solve the recurrence and find a20 .

Exercise 8.6.2 Consider the sequence of numbers defined by the recurrence

b0 = 1,
b1 = 2,
b2 = 3,
bn+3 = 2bn+2 + bn+1 − 2bn , for all n ≥ 0.

Find the first 5 members of this sequence. Then solve the recurrence and find b20 .

Exercise 8.6.3 Consider two sequences of numbers c0 , c1 , c2 , . . . and d0 , d1 , d2 , . . ., defined by the following
mutual recurrence relation.
c0 = 0, d0 = 1
cn+1 = 4cn − 3dn , dn+1 = 2cn − dn , for all n ≥ 0.

Find the first 5 members of both sequences. Then solve the recurrence. Hint: first find a matrix A such that
   
cn+1 cn
=A .
dn+1 dn
 8.7. Application: Systems of linear differential equations 299

8.7 Application: Systems of linear differential equations

Outcomes
A. Solve a system of second order linear differential equations.

8.7.1. Differential equations

Recall from calculus that if y = f (x) is a function, then y′ = f ′ (x) is its derivative and y′′ = f ′′ (x) is its
second derivative. For example,
y = sin(x),
y′ = cos(x),
y′′ = − sin(x).
Note that if y = sin(x), the second derivative y′′ is exactly the negative of y, i.e.,
y′′ = −y.
This last equation is called a differential equation. Unlike an ordinary equation, which is about an
unknown number, a differential equation is about an unknown function. Typically, a differential equation
mentions the function and one or more of its derivatives. A differential equation that only mentions the
first derivative y′ is called a first-order differential equation. A differential equation that also mentions
the second derivative y′′ is called a second-order differential equation. If the equation is a linear function
of y and its derivatives, it is called a linear differential equation.
We say that the function y = sin(x) is a solution of the differential equation y′′ = −y. It is not the only
solution. Another solution is y = cos(x), because in that case, y′′ = − cos(x), and therefore y′′ = −y. From
calculus, we know that the general solution to the differential equation y′′ = −y is given by
y = a sin(x) + b cos(x),
where a and b are any real numbers, i.e., parameters. Using the terminology of linear algebra, we can say
that the general solution of the equation y′′ = −y is a linear combination of the basic solutions y = sin(x)
and y = cos(x). More generally, we have the following theorem from calculus:

Theorem 8.32: Solutions of y′′ = qy

Let q be a positive real number. The differential equation
√ √
y′′ = −qy has basic solutions y = sin( q x) and y = cos( q x),
y′′ = 0 has basic solutions y = 1√ and y = x, √
y′′ = qy has basic solutions y=e q x and y = e− q x .

Proof. By taking derivatives, it is easy to check that each of the six functions is a solution of the cor-
√ √ √
responding differential equation. For example, for y = sin( q x), we have y′ = q cos( q x) and y′′ =
√
−q sin( q x). Therefore, y′′ = −qy.
300 Eigenvalues, eigenvectors, and diagonalization

Note that we can obtain the general solution of each of the differential equations as a linear combination
of the basic solutions. Thus, the general solution of y′′ = −qy is
√ √
y = a sin( qx) + b cos( qx),

the general solution of y′′ = 0 is

y = a + bx,
and the general solution of y′′ = qy is √ √
y = ae qx
+ be− qx
,
where a and b are parameters. The fact that each of these solutions is indeed the most general one is
proved in a calculus course. ♠

8.7.2. Systems of linear differential equations

In the same way that a system of linear equations consists of several linear equations about several vari-
ables, a system of differential equations consists of several differential equations about several unknown
functions and their derivatives. For example, the following is a system of second order linear differential
equations:
y′′ = 4y − 3z,
z′′ = 6y − 5z.
In reading these equations, it is important to understand that we are looking for two unknown functions
y = f (x) and z = g(x) such that their second derivatives satisfy both of the equations y′′ = 4y − 3z and
z′′ = 6y − 5z. The reason that this is in principle a difficult problem is that the equation for y′′ mentions
not only y, but also z, and the equation for z′′ mentions not only z, but also y. Therefore, it is not possible
to solve this system one function at a time. We say that the variables y and z are coupled.
The following example shows how we can use diagonalization to decouple the variables in a system
of differential equations. This makes it possible to solve the equations.

Example 8.33: A system of linear differential equations

Solve the following system of second order linear differential equations:

y′′ = 4y − 3z,
z′′ = 6y − 5z.

Solution. We start by writing the system in matrix form:

 ′′    
y 4 −3 y
′′ = .
z 6 −5 z
Let us write    
y 4 −3
v= and A= .
z 6 −5
With these notations, the system of differential equations take the form

v′′ = Av. (8.4)

 8.7. Application: Systems of linear differential equations 301

Our next step is to diagonalize the matrix A. Following the usual diagonalization procedure, we find that
A = PDP−1 , where    
1 1 1 0
P= and D = .
1 2 0 −2
With this, our equation takes the form v′′ = PDP−1 v, which we can also write as P−1 v′′ = DP−1 v. We
now introduce a change of variables. Let w = P−1 v. Then our system of differential equations can be
written as
w′′ = Dw. (8.5)
Note that the equation (8.5) is of exactly the same form as the equation (8.4), but with the crucial difference
that the matrix in (8.5) is diagonal. Let us give a name to the components of w:
 
u
w= .
v

Then the equation (8.5) can be written as

 ′′    
u 1 0 u
′′ = ,
v 0 −2 v

or equivalently,
u′′ = u,
v′′ = −2v.
Note that the variables u and v are not coupled! This happened because the matrix D is diagonal. We can
therefore use Theorem 8.32 to solve the equations for u and for v separately. By Theorem 8.32, the general
solution for the equation u′′ = u is
u = aex + be−x ,
and the general solution for the equation v′′ = −2v is
√ √
v = c sin( 2 x) + d cos( 2 x).

Here, a, b, c, and d are parameters. Therefore, the general solution for (8.5) is
   x + be−x
  x   −x     
u √ae √ e e 0
√ 0
√
w= = =a +b +c +d .
v c sin( 2 x) + d cos( 2 x) 0 0 sin( 2 x) cos( 2 x)

We have therefore found the four basic solution of (8.5). But what about our original equation (8.4)? We
can undo our change of variables. Since w = P−1 v, we have v = Pw. Therefore, the general solution to
our original system of differential equations is
 x   −x     
e e 0
√ 0
√
v = Pw = a P +bP +cP +dP
0 0 sin( 2 x) cos( 2 x)
       
x 1 −x 1
√ 1 √ 1
= ae +be + c sin( 2 x) + d cos( 2 x) .
1 1 2 2

♠
302 Eigenvalues, eigenvectors, and diagonalization

8.7.3. Example: coupled train cars

One of the reasons that differential equations are important is that the laws of nature often take the form
of differential equations. For example, Newton’s second law of motion asserts that the acceleration of an
object is equal to the total force on the object divided by the mass of the object. In physics, it is common
to use t instead of x for the independent variable, and x instead of y for the dependent variable, so that we
write x = f (t) instead of y = f (x). If x is the position of the object at time t, then the object’s acceleration
is x′′ , and Newton’s second law takes the form
F
x′′ = .
m
This is a differential equation. In the following example, we will need another law of physics, namely
Hooke’s law about the force exerted by a spring. A spring is an object made from an elastic material
(often in the shape of a coil), which returns to its original shape after being stretched or compressed.
Hooke’s law states that the force exerted by a spring to both of its ends is equal to

F = kx.

Here x is the extension of the spring, i.e., the change in length of the spring, relative to its relaxed (natural)
N
length. Also, k is a constant called the spring constant, measured in units of m (Newtons per meter). Of
course the direction of the force on one end of the spring is the opposite of the direction on the other
end. Hooke’s law is not a differential equation, because it does not mention any derivatives. It is just an
ordinary equation. Nevertheless, both the force F and the extension x can vary with time, i.e., they can
both be functions of t.

Example 8.34: Coupled train cars

Consider a train made up of three cars of mass 1 kg, 2 kg, and 1 kg, which are aligned on a linear
N
track and connected by springs. Assume that each spring has the same spring constant k = 2 m .

k k
1 kg 2 kg 1 kg

Find and solve the equations of motion of this system.

Solution. Let us start by defining appropriate coordinates. Let x be the position of the first car, y the
position of the second car, and z the position of the third car, measured in meters from left to right, relative
to each car’s natural resting position. The x-, y-, and z-axes are shown in the following picture. All three
axes are parallel to the train tracks, but they have their origins in different places. The coordinates are
chosen so that when x = 0, y = 0, and z = 0, then all three cars are at rest and both springs are in their
 8.7. Application: Systems of linear differential equations 303

natural relaxed state.

x-axis y-axis z-axis

x=0 y=0 z=0

Then the extension of the left spring is y − x, and therefore the left spring’s contracting force is
F1 = k(y − x).
Similarly, the extension of the right spring is z − y, and therefore its contracting force is
F2 = k(z − y).
The total force acting on the left car is F1 , the total force acting on the middle car is F2 − F1 , and the total
force acting on the right car is −F2 . By Newton’s second law, the acceleration of each car is given by
x′′ = mF11 , y′′ = F2m−F
2
1
, and z′′ = −F2
m3 . We therefore have the following equations of motion:

k
x′′ = (y − x),
m1
k
y′′ = (x − 2y + z),
m2
k
z′′ = (z − y).
m3
Let us ignore the physical units and plug in the masses m1 = 1, m2 = 2, and m3 = 1 and the spring constant
k = 2. Then the equations of motion are:
x′′ = 2(y − x),
y′′ = x − 2y + z,
z′′ = 2(z − y),
or equivalently in matrix form:
    
x′′ −2 2 0 x
 y′′  =  1 −2 1  y .
z′′ 0 2 −2 z
We can also write this as v′′ = Av, where
   
x −2 2 0
v= y  and A =  1 −2 1 .
z 0 2 −2
To solve the equation, we diagonalize the matrix A. Using the usual method for diagonalization, we find
that A = PDP−1 , where
   
−1 1 1 −2 0 0
P =  0 −1 1  and D =  0 −4 0  .
1 1 1 0 0 0
304 Eigenvalues, eigenvectors, and diagonalization

The equation v′′ = Av then becomes v′′ = PDP−1 v, or equivalently P−1 v′′ = DP−1 v. We then diagonalize
the equation by performing the change of variables w = P−1 v. The equation becomes

w′′ = Dw.

If the components of w are called u, v, and w, we can write this as

 ′′    
u −2 0 0 u
 v′′  =  0 −4 0   v  ,
w′′ 0 0 0 w

or equivalently,
u′′ = −2u,
v′′ = −4v,
w′′ = 0.
Since the variables are now decoupled, we can solve each differential equation individually.

• Solutions for λ = −2: By Theorem 8.32, the basic solutions for u′′ = −2u are
√ √
u = sin( 2t) and u = cos( 2t).

This translates into the following basic solutions for w:

   
√ 1 √ 1
w = sin( 2t) 0  and w = cos( 2t) 0  .

0 0

Using v = Pw to change to the original variables, we get the basic solutions

   
√ −1 √ −1
v = sin( 2t)  0  and v = cos( 2t)  0  .
1 1

For example, the first of these basic solutions, written in the coordinates x, y, and z, gives
√
x = − sin( 2t),
y = 0, √
z = sin( 2t).

This corresponds to a periodic oscillation of the train where the middle car
√is stationary, and the left
car moves left when the right car moves right. Each oscillation takes 2π / 2 ≈ 4.4 seconds. Here is
 8.7. Application: Systems of linear differential equations 305

a “movie” showing one oscillation:

t = 0s

t = 1.1 s

t = 2.2 s

t = 3.3 s

t = 4.4 s

The other basic solution, with cos instead of sin, is the same motion, just starting at a different offset
in time. Note how the eigenvector of A,  
−1
 0 ,
1
describes the relative motion of the three cars, i.e., the first and last cars are moving in opposite di-
rections, whereas the middle car is stationary.
√ The corresponding eigenvalue λ = −2 determines the
frequency. The frequency, which is 2/2π oscillations per second, is also called an eigenfrequency
or resonance frequency of the system.
• Solutions for λ = −4: By Theorem 8.32, the basic solutions for v′′ = −4v are

v = sin(2t) and v = cos(2t).

This translates into the following basic solutions for w:

   
0 0
w = sin(2t)  1  and w = cos(2t)  1  .
0 0
We change this to the original variables using v = Pw, and get the basic solutions
   
1 1
v = sin(2t)  −1  and v = cos(2t)  −1  .
1 1
Writing the first of these basic solutions in the coordinates x, y, and z, we get
x = sin(2t),
y = − sin(2t),
z = sin(2t).
306 Eigenvalues, eigenvectors, and diagonalization

This corresponds to a periodic oscillation of the train where the outer cars move right at the same
time that the middle car moves left. Each oscillation takes 2π /2 ≈ 3.14 seconds. Here is a “movie”
of the motion:

t = 0s

t = 0.79 s

t = 1.57 s

t = 2.36 s

t = 3.14 s

As before, the other basic solution, using cos instead of sin, is the same motion, but shifted in time.
Also, the eigenvector  
1
 −1 
1
describes the relative motion of the three cars; here the first and last car move in the same direction
while the middle car moves in the opposite direction. The eigenfrequency of this oscillation, at
2/2π oscillations per second, is slightly higher than the first one, due to the larger magnitude of the
eigenvalue λ = −4.
• Solutions for λ = 0: The last eigenvalue is zero. By Theorem 8.32, the general solution of w′′ = 0
is
w = a + bt,
where a and b are arbitrary constants. This translates into the following solution for w:
 
0
w = (a + bt) 0  ,

1
and by the change of variables v = Pw, we get
 
1
v = (a + bt)  1  .
1
Writing this in terms of the coordinates x, y, and z, we get
x = a + bt,
y = a + bt,
z = a + bt.
 8.7. Application: Systems of linear differential equations 307

This simply describes a linear motion: the three cars are moving down the track at constant speed b
from some initial starting position a.

t = 0s

t = 1s

t = 2s

t = 3s

As before, the eigenvector  

1
 1 
1
describes the relative motion of the three cars (in this case all moving in the same direction). The
eigenvalue λ = 0 indicates linear motion.

What we have described here are the basic solutions of the system. The solutions corresponding to each of
the eigenvalues are also called modes of the system. Thus, the train has three modes, each corresponding
to a particular eigenvalue of the matrix A. In the first mode, the middle car is stationary and the other two
cars oscillate in opposite directions. In the second mode, the middle car oscillates in the opposite direction
of the two outer cars. The third mode is a linear movement along the track.
As always, the general solution is a linear combination of basic solutions; for example, the cars might
be oscillating in both the first and second modes at their respective frequencies, while also moving down
the tracks. ♠

Exercises

Exercise 8.7.1 Solve the following system of second order linear differential equations:
y′′ = −5y − 6z,
z′′ = 3y + 4z.

Exercise 8.7.2 Solve the following system of second order linear differential equations:
y′′ = 4y − 3z,
z′′ = 2y − z.
308 Eigenvalues, eigenvectors, and diagonalization

Exercise 8.7.3 Solve the following system of second order linear differential equations:

x′′ = −2y + 2z,

y′′ = x − z,
z′′ = x − 2y + z.

Exercise 8.7.4 Solve the following system of first order linear differential equations.

y′ = 4y + 6z,
z′ = −3y − 5z.

Hint: The method is similar to that of second-order equations. Use the fact, known from calculus, that the
equation f ′ = k f has basic solution f (x) = ekx . Here k is any constant (positive, negative, or zero).

Exercise 8.7.5 Consider three coupled train cars as in Example 8.34, except that all three cars have mass
N
1 kg and both spring have spring constant k = 1 m . Find and solve the equations of motion.

8.8 Application: The matrix exponential

Outcomes

A. Compute eA , sin A, and cos A, for a square matrix A.

B. Apply any analytic function to a square matrix.

From calculus, recall that a function is called analytic if it can be defined by a power series. For example:
1 1 1
ex = 1 + x + x2 + x3 + x4 + . . .
2 3! 4!
1 3 1 5 1 7
sin x = x − x + x − x ± . . .
3! 5! 7!
1 2 1 4 1 6
cos x = 1 − x + x − x ± . . .
2 4! 6!
We know from calculus that the above power series converge for all real numbers x. Since it makes sense
to compute the n th power of a square matrix, in principle it also makes sense to plug a matrix into a power
series. For a square matrix A, we can define
1 1 1
eA = I + A + A2 + A3 + A4 + . . .
2 3! 4!
1 3 1 5 1 7
sin A = A − A + A − A ± . . .
3! 5! 7!
1 2 1 4 1 6
cos A = I − A + A − A ± . . .
2 4! 6!
 8.8. Application: The matrix exponential 309

The goal of this section is to investigate whether these power series converge, and if yes, how to compute
the sum of the series. We begin with the case of a diagonal matrix.

Example 8.35: Exponential of a diagonal matrix

Compute eD , where  
x 0
D= .
0 y

Solution. By definition,
1 1
eD = I + D + D2 + D3 + . . .
 2  3!     
1 0 x 0 1 x2 0 1 x3 0
= + + + +...
0 1 0 y 2 0 y2 3! 0 y3
 
1 + x + 12 x2 + 3!
1 3
x +... 0
=
0 1 + y + 21 y2 + 3!1 y3 + . . .
 x 
e 0
= .
0 ey
Therefore, the exponential of a diagonal matrix is computed by taking the exponential of each diagonal
entry. Note that this proves, in particular, that the sum converges. ♠
The same argument also works for applying other analytic functions to diagonal matrices, for example:
   
x 0 sin x 0
sin = ,
0 y 0 sin y
   
x 0 cos x 0
cos = .
0 y 0 cos y
But how can we compute the matrix exponential of a non-diagonal matrix? This can be done by diagonal-
ization. The following theorem shows how:

Theorem 8.36: Matrix functions by diagonalization

Suppose A is a diagonalizable square matrix, with A = PDP−1 . Then

eA = PeD P−1 ,
sin A = P(sinD)P−1 ,
cos A = P(cos D)P−1 .

Proof. We have
1 1
eA = I + A + A2 + A3 + . . .
2 3!
1 1
= I + PDP−1 + (PDP−1)2 + (PDP−1)3 + . . .
2 3!
310 Eigenvalues, eigenvectors, and diagonalization

1 1
= PIP−1 + PDP−1 + PD2 P−1 + PD3 P−1 + . . .
2 3!
1 2 1 3 −1
= P(I + D + D + D + . . .)P
2 3!
= PeD P−1 .

The proof for sin and cos is similar. Indeed, the same method works for any analytic function. ♠

Example 8.37: A matrix exponential

Compute eA , where  
−8 10
A= .
−5 7

Solution. We first diagonalize A. Following the usual method, we find that the eigenvalues are λ = −3
and λ = 2, with corresponding eigenvectors
   
2 1
and .
1 1

Therefore, we can diagonalize A as A = PDP−1 , where

     
2 1 −3 0 −1 1 −1
P= , D= , and P = .
1 1 0 2 −1 2

By Theorem 8.36, we have

   −3    −3 
A D −1 2 1 e 0 1 −1 2e − e2 −2e−3 + 2e2
e = Pe P = = .
1 1 0 e2 −1 2 e−3 − e2 −e−3 + 2e2

Exercises

Exercise 8.8.1 Compute eA and cos A for the following matrices:

 
    0 1 1
1 0 4 2
(a) A = , (b) A = , (c) A =  1 0 −1  .
0 2 −3 −1
−1 1 2

Exercise 8.8.2 The cube root is an analytic function. Compute the cube root of the matrix
 
22 −21
A= .
14 −13
 8.9. Properties of eigenvectors and eigenvalues 311

Exercise 8.8.3 Use matrix exponentials to find the solution to the first-order linear differential equation
 ′   
x 0 −1 x
=
y 6 5 y

with initial value    

x(0) 2
= .
y(0) 2
Hint: form the matrix exponential eAt and then the solution is eAt v0 , where v0 is the initial vector.

8.9 Properties of eigenvectors and eigenvalues

Outcomes
A. Know that eigenvectors corresponding to distinct eigenvalues are linearly independent.

B. Compute the algebraic and geometric multiplicity of an eigenvalue.

C. Determine whether a matrix is diagonalizable from the geometric multiplicities of its eigen-
values.

In this section, we state some useful properties of eigenvectors and eigenvalues. The first question we
consider is whether eigenvectors for different eigenvalues are linearly independent. This is indeed the
case, as the following proposition shows:

Proposition 8.38: Eigenvectors for different eigenvalues are linearly independent

Let A be a square matrix, and suppose that A has distinct eigenvalues λ1 , . . . , λk with corresponding
eigenvectors v1 , . . . , vk . Then v1 , . . . , vk are linearly independent.

Proof. Suppose, for the sake of obtaining a contradiction, that v1 , . . . , vk are linearly dependent. Let m the
the smallest index such that vm is redundant, i.e., such that vm is a linear combination of previous vectors.
Say
vm = a1 v1 + . . . + am−1 vm−1 . (8.6)
Multiplying the equation by A, we get

Avm = a1 Av1 + . . . + am−1 Avm−1 ,

and therefore, since v1 , . . . , vm are eigenvectors,

λm vm = a1 λ1 v1 + . . . + am−1 λm−1 vm−1 . (8.7)

312 Eigenvalues, eigenvectors, and diagonalization

Subtracting λm times equation (8.6) from (8.7), we get

0 = a1 (λ1 − λm )v1 + . . . + am−1 (λm−1 − λm )vm−1 .
Since v1 , . . . , vm−1 are, by assumption, linearly independent (because vm was the leftmost redundant vec-
tor), it follows that a1 (λ1 − λm ) = 0, . . . , am−1 (λm−1 − λm−1 ) = 0. Since the eigenvalues λ1 , . . . , λm are,
by assumption, distinct, it follows that a1 , . . . , am−1 = 0. But then (8.6) implies that vm = 0, contradicting
the assumption that vm is an eigenvector (and therefore non-zero). ♠
An immediate consequence of this proposition is that an n × n-matrix with n distinct eigenvalues is diag-
onalizable.

Corollary 8.39: Distinct eigenvalues

Let A be an n × n-matrix and suppose it has n distinct eigenvalues. Then A is diagonalizable.

Proof. Each of the n eigenvalues has an eigenvector, and by Proposition 8.38, they are linearly independent.
Then A is diagonalizable by Theorem 8.21. ♠
The next issue we consider is that of “repeated” eigenvalues. There are two senses in which an eigenvalue
can occur “more than once”. The first is if the eigenvalue appears as a repeated root of the characteristic
polynomial. For example, if the characteristic polynomial is p(λ ) = (1 − λ )(1 − λ )(3 − λ ), then we say
that the root λ = 1 appears with multiplicity two, and the root λ = 3 appears with multiplicity one. We
call this the algebraic multiplicity of the eigenvalue.
The second sense in which an eigenvalue can occur “more than once” is when an eigenvalue has more
than one linearly independent eigenvector. In other words, when the eigenspace has dimension greater
than 1. We call this the geometric multiplicity of the eigenvalue.
The following definition summarizes these concepts.

Definition 8.40: Algebraic and geometric multiplicity

Let λ̂ be an eigenvalue of a square matrix A. Then the algebraic multiplicity of λ̂ is the largest
power k such that (λ̂ − λ )k is a factor of the characteristic polynomial. The geometric multiplicity
of λ̂ is the dimension of its eigenspace Eλ̂ .

One would hope that the algebraic and geometric multiplicities are always equal. Unfortunately, this is
not the case, as the following example shows.

Example 8.41: Algebraic and geometric multiplicity

Let  
3 1 0 0 0
 0 3 0 0 0 
 
A=
 0 0 4 0 0 .

 0 0 0 4 0 
0 0 0 0 5
Find the algebraic and geometric multiplicity of each eigenvalue of A.
 8.9. Properties of eigenvectors and eigenvalues 313

Solution. The characteristic polynomial is (3 − λ )2 (4 − λ )2 (5 − λ ). Therefore, the eigenvalues are 3, 4,

and 5, with algebraic multiplicity 2, 2, and 1, respectively. To compute the geometric multiplicity, we need
to find each eigenspace. For λ = 3, we must solve (A − 3I)v = 0, or equivalently,
 
0 1 0 0 0 0
 0 0 0 0 0 0 
 
 0 0 1 0 0 0 .
 
 0 0 0 1 0 0 
0 0 0 0 2 0

This system has rank 4, and the only basic solution is [1, 0, 0, 0, 0]T . Thus, the eigenspace E3 is 1-
dimensional, and the geometric multiplicity of λ = 3 is 1. A similar calculation show that λ = 4 has
geometric multiplicity 2 and λ = 5 has geometric multiplicity 1. The information is summarized in the
following table:

Eigenvalue λ =3 λ =4 λ =5
Algebraic multiplicity 2 2 1
Geometric multiplicity 1 2 1

♠
In the example, the geometric multiplicity is either smaller than or equal to the algebraic multiplicity. The
following proposition states that this is always the case.

Proposition 8.42: Algebraic and geometric multiplicity

Let λ̂ be an eigenvalue of a matrix A, with algebraic multiplicity k and geometric multiplicity m.
Then
1 ≤ m ≤ k.

Proof. It is clear that m ≥ 1, because each eigenvalue, by definition, must have at least one associated
eigenvector. Therefore, the eigenspace is at least 1-dimensional. We must show m ≤ k. Assume that A is
an n × n-matrix. By assumption, the geometric multiplicity of λ̂ is m, so the eigenspace Eλ̂ has dimension
m. So there exist m linearly independent eigenvectors v1 , . . . , vm for the eigenvalue λ̂ . Extend v1 , . . . , vm to
a basis v1 , . . . , vn of Rn , and let P be the invertible matrix that has v1 , . . . , vn as its columns. Let B = P−1 AP.
Since the first m columns of P are eigenvectors of A for the eigenvalue λ̂ , it follows that B is of the form
 
λ̂ 0 · · · 0 ∗ · · · ∗
 0 λ̂ · · · 0 ∗ · · · ∗ 
 
 . . . . . . 
 .. .. . . .. . . .. 
 
 0 0 · · · λ̂ ∗ · · · ∗  ,
 
 0 0 ··· 0 ∗ ··· ∗ 
 
 . . . 
 .. .. . . ... . . . ... 
0 0 ··· 0 ∗ ··· ∗
i.e., the first m columns of B are like those of a diagonal matrix. Then from the cofactor method for
computing determinants, we know that det(B − λ I) contains the factor (λ̂ − λ )m . But since A and B are
314 Eigenvalues, eigenvectors, and diagonalization

similar matrices, they have the same characteristic polynomial. Therefore, det(A − λ I) also has (λ̂ − λ )m
as a factor. It follows, by definition of algebraic multiplicity, that m ≤ k, as desired. ♠
We know from Theorem 8.21 that an n × n-matrix is diagonalizable if and only if it has n linearly inde-
pendent eigenvectors. We can re-state this in terms of geometric multiplicity as follows.

Proposition 8.43: Geometric multiplicity and diagonalization

An n × n-matrix A is diagonalizable if and only if the sum of the geometric multiplicities of all the
eigenvalues of A is n.

Proof. By Proposition 8.38, eigenvectors corresponding to different eigenvalues are linearly independent.
Therefore, by taking a basis of each eigenspace, we can obtain exactly as many linearly independent eigen-
vectors as the sum of the dimensions of all the eigenspaces, i.e., the sum of the geometric multiplicities of
all eigenvalues. By Theorem 8.21, A is diagonalizable if and only if this number is n. ♠

Exercises

Exercise 8.9.1 Find the algebraic and geometric multiplicity of each eigenvalue of the following matrices.
Which of the matrices are diagonalizable?
   
1 1 −7 8
(a) A = , (b) B = ,
−1 3 −4 5
     
2 −1 −1 3 1 −1 −2 0 1
(c) C =  0 4 2  , (d) D =  0 2 1  , (e) E =  −1 −1 1  .
0 −1 1 0 −1 4 −2 1 0

Exercise 8.9.2 Determine which of the following matrices are diagonalizable.

(a) A is a 3 × 3-matrix with eigenvalues −1 and 3. The eigenvalue −1 has algebraic multiplicity 2 and
geometric multiplicity 1. The eigenvalue 3 has algebraic and geometric multiplicity 1.

(b) B is a 4 × 4-matrix with eigenvalues 2 and −2. The eigenvalue 2 has algebraic and geometric
multiplicity 1. The eigenvalue −2 has algebraic and geometric multiplicity 3.

(c) C is a 5 × 5-matrix with eigenvalues 1 and 3. The eigenvalue 1 has algebraic and geometric multi-
plicity 2, and the eigenvalue 3 has algebraic and geometric multiplicity 1.
 8.10. The Cayley-Hamilton Theorem 315

8.10 The Cayley-Hamilton Theorem

Outcomes

A. For a square matrix A, find a polynomial p(x) such that p(A) = 0.

In this section, we will consider the so-called Cayley-Hamilton theorem. It states that every square
matrix is a root of its own characteristic polynomial. We use the following notation. If

p(λ ) = an λ n + an−1 λ n−1 + . . . + a1 λ + a0

is a polynomial, we denote by p(A) the matrix defined by

p(A) = an An + an−1 An−1 + . . . + a1 A + a0 I.

The explanation for the last term is that A0 is interpreted as I, the identity matrix.

Theorem 8.44: Cayley-Hamilton theorem

Let A be a square matrix and let p(λ ) = det(A− λ I) be its characteristic polynomial. Then p(A) = 0.

Before we prove this theorem, we consider an example.

Example 8.45: Cayley-Hamilton theorem

Let  
3 4
A= .
−1 2
Find the characteristic polynomial p(λ ), and compute p(A).

Solution. The characteristic polynomial is

 
 3−λ 4 
p(λ ) = det(A − λ I) =   = (3 − λ )(2 − λ ) − (−1)4 = λ 2 − 5λ + 10.

−1 2 − λ

Applying the characteristic polynomial to A, we get

 2    
2 3 4 3 4 10 0
p(A) = A − 5A + 10I = −5 +
−1 2 −1 2 0 10
       
5 20 15 20 10 0 0 0
= − + = ,
−5 0 −5 10 0 10 0 0

just as predicted by the Cayley-Hamilton theorem. ♠

316 Eigenvalues, eigenvectors, and diagonalization

The remainder of this section is devoted to the proof of the Cayley-Hamilton theorem. Readers who are
not interested in the proof can skip this material. We begin with a lemma:

Lemma 8.46: Polynomials with matrix coefficients

Let A0 , . . . , Am be n × n-matrices and assume that for all scalars λ ,

A0 + A1 λ + . . . + Am λ m = 0.

Then each Ai = 0.

Proof. Multiply by λ −m to obtain

A0 λ −m + A1 λ −m+1 + . . . + Am−1 λ −1 + Am = 0.

Now let |λ | → ∞ to obtain Am = 0. With this, multiply by λ to obtain

A0 λ −m+1 + A1 λ −m+2 + . . . + Am−1 = 0.

Now let |λ | → ∞ to obtain Am−1 = 0. Continue multiplying by λ and letting λ → ∞ to obtain Ai = 0 for
all i. ♠
The following is a simple consequence of the lemma.

Corollary 8.47:
Let Ai and Bi be n × n-matrices and suppose that

A0 + A1 λ + . . . + Am λ m = B0 + B1 λ + . . . + Bm λ m

for all λ . Then for any n × n-matrix C,

A0 + A1C + . . . + AmCm = B0 + B1C + . . . + BmCm .

Proof. Subtracting the right-hand side from the left-hand side and using Lemma 8.46, we get that Ai = Bi
for all i. But then the conclusion immediately follows. ♠
With this preparation, it is now relatively easy to prove the Cayley-Hamilton theorem.
Proof of the Cayley-Hamilton Theorem. Let A be an n × n-matrix, and let p(λ ) = det(A − λ I) be its
characteristic polynomial. Let adj(A − λ I) be the adjugate of the matrix A − λ I (see Section 7.6 for the
definition of the adjugate). Since each of the entries of the adjugate is a cofactor of A − λ I, the entries are
polynomials in λ of degree at most n − 1. Therefore, the adjugate can be written in the form

adj(A − λ I) = C0 +C1 λ + . . . +Cn−1 λ n−1 .

By Theorem 7.29, we have

det(A − λ I) I = (A − λ I) adj(A − λ I),
or equivalently,
p(λ ) I = (A − λ I) (C0 +C1 λ + . . . +Cn−1 λ n−1 ).
 8.11. Complex eigenvalues and eigenvectors 317

Since this equation holds for all λ , Corollary 8.47 may be used. Therefore, if λ is replaced with A, the
two sides will be equal. Thus

p(A) I = (A − A) (C0 +C1 A + . . . +Cn−1 An−1 ) = 0.

It follows that p(A) = 0, concluding the proof of the Cayley-Hamilton Theorem. ♠

Exercises

Exercise 8.10.1 Let  

5 7
A= .
−4 3
Find the characteristic polynomial p(λ ), and compute p(A).

Exercise 8.10.2 Let  

1 2 0
A =  0 2 −1  .
0 1 4
Find the characteristic polynomial p(λ ), and compute p(A).

Exercise 8.10.3

(a) Let A be a 2 × 2-matrix. Prove that A2 is a linear combination of A and I.

(b) Given an example of a 3 × 3-matrix A such that A2 is not a linear combination of A and I.

8.11 Complex eigenvalues and eigenvectors

Outcomes
A. Find the complex eigenvalues and eigenvectors of a matrix.

B. Diagonalize a matrix over the complex numbers.

An n × n-matrix is diagonalizable if and only if it has n linearly independent eigenvectors. But as we

saw in Example 8.9, if we work over the real numbers, it can sometimes happen that a matrix has no
eigenvalues, and therefore no eigenvectors, at all. For example, the matrix
 
0 −1
A=
1 0
318 Eigenvalues, eigenvectors, and diagonalization

has characteristic polynomial λ 2 +1. Since the equation λ 2 +1 does not have any roots in the real numbers,
there are no real eigenvalues.
On the other hand, the fundamental theorem of algebra tell us that over the complex numbers, every
non-constant polynomial has a root. In fact, every polynomial of degree n factors into n linear factors.
Therefore, every matrix has at least one eigenvalue over the complex numbers. Some matrices are diago-
nalizable over the complex numbers but not over the real numbers. An introduction to complex numbers
and the fundamental theorem of algebra can be found in Appendix A.

Example 8.48: Complex eigenvalues and eigenvectors

Find the eigenvectors and eigenvalues of
 
0 −1
A=
1 0

over the complex numbers. Diagonalize A if possible.

Solution. The characteristic polynomial is λ 2 + 1. This has no roots in the real numbers, but it has two
roots λ = i and λ = −i in the complex numbers. To find the eigenvectors for λ = i, we solve (A −iI)v = 0:
     
−i −1 0 R1 ↔ R2 1 −i 0 R2 ← R2 +iR1 1 −i 0
≃ ≃ .
1 −i 0 −i −1 0 0 0 0
Thus, the basic eigenvector for λ = i is  
i
v1 = .
1
Similarly, to find the eigenvectors for λ = −i, we solve (A + iI)v = 0:
     
i −1 0 R1 ↔ R2 1 i 0 R2 ← R2 −iR1 1 i 0
≃ ≃ .
1 i 0 i −1 0 0 0 0
Thus, the basic eigenvector for λ = −i is
 
−i
v2 = .
1
Since we have found two linearly independent eigenvectors, the matrix A is diagonalizable. We have
A = PDP−1 , where    
i −i i 0
P= and D = .
1 1 0 −i
♠

Example 8.49: Complex eigenvalues and eigenvectors

Find the eigenvectors and eigenvalues of
 
1 −1
A=
1 1

over the complex numbers. Diagonalize A if possible.

 8.11. Complex eigenvalues and eigenvectors 319

Solution. The characteristic polynomial is

 
 1 − λ −1 
det(A − λ I) =   = (1 − λ )2 + 1 = λ 2 − 2λ + 2.

1 1−λ
To find the roots, we use the quadratic formula. The roots are given by:
√ √
−b ± b2 − 4ac 2 ± −4
λ= = = 1 ± i.
2a 2
Note that since the discriminant b2 − 4ac is negative, there are no real solutions. However, we find two
complex solutions λ = 1 + i and λ = 1 − i. To find the eigenvectors for λ = 1 + i, we solve the equation
(A − (1 + i)I)v = 0:    
−i −1 0 1 −i 0
≃ ... ≃ .
1 −i 0 0 0 0
The basic eigenvector for λ = 1 + i is  
i
v1 = .
1
Similarly, the basic eigenvector for λ = 1 − i is
 
−i
v2 = .
1
Since we have found two linearly independent eigenvectors, the matrix A is diagonalizable. We have
A = PDP−1 , where    
i −i 1+i 0
P= and D = .
1 1 0 1−i
♠

Example 8.50: Diagonalize a matrix over the complex numbers

Diagonalize the matrix  
−3 −2 4
A= 2 1 0 .
−2 −2 3

Solution. The characteristic polynomial is

p(λ ) = (−3 − λ )(1 − λ )(3 − λ ) − 16 + 8(1 − λ ) + 4(3 − λ ) = −λ 3 + λ 2 − 3λ − 5.
By trial and error, we find that λ = −1 is one of the roots. We factor out (λ + 1):
p(λ ) = (λ + 1)(−λ 2 + 2λ − 5).
We then use the quadratic formula to find the other two eigenvalues, i.e., the roots of −λ 2 + 2λ − 5. They
are: √
−2 ± −16
λ= = 1 ± 2i.
−2
Eigenvectors:
320 Eigenvalues, eigenvectors, and diagonalization

• For λ = −1, we solve (A − (−1)I)v = 0:

   
−2 −2 4 0 1 1 0 0
 2 2 0 0  ≃ ... ≃  0 0 1 0 .
−2 −2 4 0 0 0 0 0
The basic eigenvector is 

−1
v1 =  1  .
0
• For λ = 1 + 2i, we solve (A − (1 + 2i)I)v = 0:
R1 ← −R1 /2
  R2 ← R2 /2  
−4 − 2i −2 4 0 R3 ← −R3 /2
2 + i 1 −2 0
 2 −2i 0 0  ≃  1 −i 0 0 
−2 −2 2 − 2i 0 1 1 i−1 0
 
1 −i 0 0
R1 ↔ R2
≃  2 + i 1 −2 0 
1 1 i−1 0
R2 ← R2 −(2+i)R1
 
1 −i 0 0
R3 ← R3 −R1
≃  0 2i −2 0 
0 1+i i−1 0
R2 ← R2 /2i  
R3 ← R3 /(1+i)
1 −i 0 0
≃  0 1 i 0 
0 1 i 0
R1 ← R1 +iR2
 
1 0 −1 0
R3 ← R3 −R2
≃  0 1 i 0 .
0 0 0 0
The basic eigenvector is  
1
v2 =  −i  .
1
• For λ = 1 − 2i, we solve (A − (1 − 2i)I)v = 0:
   
−4 + 2i −2 4 0 1 0 −1 0
 2 2i 0 0  ≃ ... ≃  0 1 −i 0  .
−2 −2 2 + 2i 0 0 0 0 0
The basic eigenvector is  
1
v3 =  i  .
1
Therefore, A = PDP−1 , where
   
−1 1 1 −1 0 0
P =  1 −i i  and D =  0 1 + 2i 0 .
0 1 1 0 0 1 − 2i
 8.11. Complex eigenvalues and eigenvectors 321

♠
In Example 8.48, the complex eigenvalues were i and −i. In Example 8.49, the complex eigenvalues were
1 + i and 1 − i. In Example 8.50, the complex eigenvalues were 1 + 2i and 1 − 2i, and there was also a real
eigenvalue of −1. Is it a coincidence that the complex eigenvalues always come in conjugate pairs? The
following proposition states that this is always the case.

Proposition 8.51: Complex conjugate eigenvalues

Let A be a square matrix whose entries are real numbers. If λ is an eigenvalue of A, then so is λ .

Proof. Assume λ is an eigenvalue of A with corresponding eigenvector v. Then Av = λ v. Taking complex

conjugates of both sides of the equation, we have Av = λ v, and therefore Av = λ v. Since A is matrix
with real entries, we have A = A, and therefore Av = λ v. It follows that λ is an eigenvalue of A with
corresponding eigenvector v. ♠

It is important to note that even over the complex numbers, not all matrices are diagonalizable. On the one
hand, the characteristic polynomial of an n × n-matrix always factors into n linear factors over the complex
numbers. Therefore, the sum of the algebraic multiplicities of the eigenvalues is always n. However, it
can still happen that the geometric multiplicity of some eigenvalue is less than its algebraic multiplicity.
In that case, the matrix is not diagonalizable, even over the complex numbers. We have:

Proposition 8.52: Diagonalizability criterion

A square matrix A is diagonalizable over the complex numbers if and only if the geometric multi-
plicity of each eigenvalue is equal to its algebraic multiplicity.

Proof. Let A be an n × n-matrix. By the fundamental theorem of algebra, the characteristic polynomial
factors into n linear factors. Therefore, the sum of the algebraic multiplicities of all the eigenvalues is n.
We know by Proposition 8.42 that the geometric multiplicity of each eigenvalue less than or equal to its
algebraic multiplicity. If the geometric multiplicity of each eigenvalue is equal to its algebraic multiplicity,
then the sum of the geometric multiplicities is n, and therefore A is diagonalizable by Proposition 8.43.
On the other hand, if the geometric multiplicity of some eigenvalue is less than its algebraic multiplicity,
then the sum of the geometric multiplicities is less than n, and A is not diagonalizable. ♠

Example 8.53: Non-diagonalizable matrix

 
1 1
Show that the matrix A = cannot be diagonalized, even over the complex numbers.
0 1

Solution. The characteristic polynomial is (1 − λ )2 , and therefore the only eigenvalue is λ = 1, with
algebraic multiplicity 2. On the other hand, the eigenspace for λ = 1 is 1-dimensional:
   
0 1 0
v=
0 0 0
322 Eigenvalues, eigenvectors, and diagonalization

has a 1-dimensional solution space. Therefore, we can find only one basic eigenvector, and the matrix is
not diagonalizable. ♠
To finish up this chapter, we will consider an application of complex eigenvalues. We will solve a recur-
rence as in Section 8.6. But this time, although the recurrence relation only uses real numbers, complex
numbers will be required to solve it.

Example 8.54: Solving a recurrence using complex eigenvalues

Consider the sequence of numbers defined by the recurrence

f0 = 1,
f1 = 3,
fn+2 = 2 fn+1 − 2 fn , for all n ≥ 0.

Solve the recurrence, i.e., find a closed formula for fn .

Solution. The first few members of the sequence are:

1, 3, 4, 2, −4, −12, −16, −8, 16, . . .

To solve the recurrence, let  

fn
vn = ,
fn+1
so that for all n ≥ 0,
        
fn+1 fn+1 0 1 fn 0 1
vn+1 = = = = vn .
fn+2 2 fn+1 − 2 fn −2 2 fn+1 −2 2

We then diagonalize the matrix  

0 1
A= .
−2 2
The characteristic polynomial is
p(λ ) = λ 2 − 2λ + 2.
The eigenvalues are the roots of the characteristic polynomial. We compute them using the quadratic
formula: √
2 ± 4 − 8 2 ± 2i
λ1 , λ2 = = = 1 ± i.
2 2
To find the eigenvectors for λ1 = 1 + i, we solve the equation (A − (1 + i)I)v = 0:
     
−1 − i 1 0 R1 ← (1−i)R1 −2 1 − i 0 −2 1 − i 0
≃ ≃ .
−2 1−i 0 −2 1 − i 0 0 0 0

A basic eigenvector for λ = 1 + i is  

1−i
v= .
2
 8.11. Complex eigenvalues and eigenvectors 323

Similarly, a basic eigenvector for λ = 1 − i is

 
1+i
u= .
2
Since we have found two linearly independent eigenvectors, the matrix A is diagonalizable. We have
A = PDP−1 , where    
1−i 1+i 1+i 0
P= and D = .
2 2 0 1−i
The inverse of P is  
−1 1 2i 1 − i
P = .
4 −2i 1 + i
Finally, we use this information to solve the recurrence:
 
  fn
fn = 1 0
fn+1
 
= 1 0 vn
 
= 1 0 An v0
 
= 1 0 PDn P−1 v0
     
  1−i 1+i (1 + i)n 0 1 2i 1 − i 1
= 1 0
2 2 0 (1 − i)n 4 −2i 1 + i 3
   
1  (1 + i)n 0 3−i
= 1−i 1+i n
4 0 (1 − i) 3+i
1
= ((1 − i)(1 + i)n(3 − i) + (1 + i)(1 − i)n(3 + i))
4
1
= ((2 − 4i)(1 + i)n + (2 + 4i)(1 − i)n)
4
1
= ((1 − 2i)(1 + i)n + (1 + 2i)(1 − i)n) .
2
We can use this, for example, to calculate the 8 th element of the sequence:
1
1 32
f8 = (1 − 2i)(1 + i)8 + (1 + 2i)(1 − i)8 = ((1 − 2i)16 + (1 + 2i)16) = = 16.
2 2 2
♠

Exercises

Exercise 8.11.1 Find the (real or complex) eigenvalues and eigenvectors of the following matrices. Diag-
onalize each matrix if possible.
   
    2 −1 3 1 1 1
2 1 −1 −2
A= , B= , C= 0 1 1  , D =  −1 2 0 
−1 2 1 1
−1 1 0 1 −1 1
324 Eigenvalues, eigenvectors, and diagonalization

Exercise 8.11.2 I know a certain  2 × 2-matrix A. My matrix has complex eigenvalue λ = 1 + 2i and
 real
1
corresponding eigenvector v = .
i

(a) Find another eigenvalue and corresponding eigenvector of A.

(b) Diagonalize A.

(c) What is the secret matrix A?

Exercise 8.11.3 Solve the recurrence

a0 = 0,
a1 = 1,
an+2 = 4an+1 − 5an , for all n ≥ 0.
9. Vector spaces
In Chapter 2, we considered Rn , the set of n-dimensional column vectors. We now introduce a more
general concept of “vector”, as an element of an abstract vector space. Basically, vectors are entities that
can be added and scaled. While some vectors look like lists of numbers (for example, column vectors, row
vectors), other kinds of vectors don’t look like lists of numbers at all (for example, functions, polynomials).
Part of the power of linear algebra comes from our ability to find vector spaces in many unexpected places.
Much of the content of this chapter will be a repetition of things we have already seen in Chapter 2
in the context of Rn . For example, we will be talking about linear combinations, linear independence,
spanning sets, bases, subspaces, linear transformations, and so on. We initially introduced these concepts
in the context of the vector space Rn , so that they would be easier to understand. We will now see that
they in fact apply to all vector spaces.

9.1 Definition of vector spaces

Outcomes
A. Develop the concept of a vector space through axioms.

B. Use the vector space axioms to determine if a set and its operations constitute a vector space.

C. Encounter several examples of vector spaces.

Definition 9.1: Vector space

Let K be a field. A vector space over K is a set V equipped with two operations of addition and
scalar multiplication, such that the following properties hold:

(A1) Commutative law of addition: u + v = v + u.

(A2) Associative law of addition: (u + v) + w = u + (v + w).
(A3) The existence of an additive unit: there exists an element 0 ∈ V such that for all u, u + 0 = u.
(A4) The law of additive inverses: u + (−u) = 0.
(SM1) The distributive law over vector addition: k(u + v) = ku + kv.
(SM2) The distributive law over scalar addition: (k + ℓ)u = ku + ℓu.
(SM3) The associative law for scalar multiplication: k(ℓu) = (kℓ)u.
(SM4) The rule for multiplication by one: 1u = u.

325
326 Vector spaces

The above definition is concerned about two operations: vector addition, denoted by v + w, and scalar
multiplication, denoted by kv or sometimes k · v. In the law of additive inverses, we have written −u for
(−1)u. Often, the scalars will be real numbers, but it is also possible to use scalars from a different field
K. We also use the term K-vector space to refer to a vector space over a field K. When K = R, we also
speak of a real vector space, and when K = C, we speak of a complex vector space. If the field is clear
from the context, we often don’t mention it at all, and just speak of a “vector space”. The elements of a
vector space are called vectors.
Our first example of a vector space is of course Rn .

Example 9.2: Rn is a vector space

The set Rn of n-dimensional real column vectors, with the usual operations of vector addition and
scalar multiplication, is a vector space.
More generally, if K is a field, the set K n of n-dimensional column vectors with components in K is
a K -vector space.

Proof. Properties (A1)–(A4) hold by Proposition 2.8, and properties (SM1)–(SM4) hold by Proposi-
tion 2.11. ♠
We now consider some other examples of vector spaces.

Example 9.3: Vector space of polynomials of degree 2

Let P2 be the set of all polynomials of degree at most 2 with coefficients from a field K , i.e.,
expressions of the form
p(x) = ax2 + bx + c,
where a, b, c ∈ K . Define addition and scalar multiplication of polynomials in the usual way, i.e.,

(ax2 + bx + c) + (a′ x2 + b′ x + c′ ) = (a + a′ )x2 + (b + b′ )x + (c + c′ )

k(ax2 + bx + c) = ka x2 + kb x + kc.

Then P2 is a vector space.

Proof. To show that P2 is a vector space, we verify the 8 vector space axioms. Let

p(x) = a2 x2 + a1 x + a0 ,
q(x) = b2 x2 + b1 x + b0 ,
r(x) = c2 x2 + c1 x + c0

be polynomials in P2 and let k, ℓ be scalars.

(A1) We prove the commutative law of addition.

p(x) + q(x) = (a2 x2 + a1 x + a0 ) + (b2 x2 + b1 x + b0 )

= (a2 + b2 )x2 + (a1 + b1 )x + (a0 + b0 )
 9.1. Definition of vector spaces 327

= (b2 + a2 )x2 + (b1 + a1 )x + (b0 + a0 )

= (b2 x2 + b1 x + b0 ) + (a2 x2 + a1 x + a0 )
= q(x) + p(x).

(A2) We prove the associative law of addition.

(p(x) + q(x)) + r(x) = ((a2x2 + a1 x + a0 ) + (b2 x2 + b1 x + b0 )) + (c2 x2 + c1 x + c0 )
= ((a2 + b2 )x2 + (a1 + b1 )x + (a0 + b0 )) + (c2x2 + c1 x + c0 )
= ((a2 + b2 ) + c2 )x2 + ((a1 + b1 ) + c1 )x + ((a0 + b0 ) + c0 )
= (a2 + (b2 + c2 ))x2 + (a1 + (b1 + c1 ))x + (a0 + (b0 + c0 ))
= (a2 x2 + a1 x + a0 ) + ((b2 + c2 )x2 + (b1 + c1 )x + (b0 + c0 ))
= (a2 x2 + a1 x + a0 ) + ((b2 x2 + b1 x + b0 ) + (c2 x2 + c1 x + c0 ))
= p(x) + (q(x) + r(x)).

(A3) To prove the existence of an additive unit, let 0(x) = 0x2 + 0x + 0, the so-called zero polynomial.
Then
p(x) + 0(x) = (a2 x2 + a1 x + a0 ) + (0x2 + 0x + 0)
= (a2 + 0)x2 + (a1 + 0)x + (a0 + 0)
= a2 x2 + a1 x + a0
= p(x).

(A4) We prove the law of additive inverses.

p(x) + (−p(x)) = (a2 x2 + a1 x + a0 ) + (−a2 x2 − a1 x − a0 )
= (a2 − a2 )x2 + (a1 − a1 )x + (a0 − a0 )
= 0x2 + 0x + 0
= 0(x).

(SM1) We prove the distributive law over vector addition.

k(p(x) + q(x)) = k((a2x2 + a1 x + a0 ) + (b2 x2 + b1 x + b0 ))
= k((a2 + b2 )x2 + (a1 + b1 )x + (a0 + b0 ))
= k(a2 + b2 )x2 + k(a1 + b1 )x + k(a0 + b0 )
= (ka2 + kb2 )x2 + (ka1 + kb1 )x + (ka0 + kb0 )
= (ka2 x2 + ka1 x + ka0 ) + (kb2 x2 + kb1 x + kb0 )
= kp(x) + kq(x).

(SM2) We prove the distributive law over scalar addition.

(k + ℓ)p(x) = (k + ℓ)(a2x2 + a1 x + a0 )
= (k + ℓ)a2 x2 + (k + ℓ)a1 x + (k + ℓ)a0
= (ka2 x2 + ka1 x + ka0 ) + (ℓa2x2 + ℓa1 x + ℓa0 )
= kp(x) + ℓp(x).
 328 Vector spaces

(SM3) We prove the associative law for scalar multiplication.

k(ℓp(x)) = k(ℓ(a2x2 + a1 x + a0 ))
= k(ℓa2 x2 + ℓa1 x + ℓa0 )
= kℓa2 x2 + kℓa1 x + kℓa0
= (kℓ)(a2x2 + a1 x + a0 )
= (kℓ)p(x).

(SM4) Finally, we prove the rule for multiplication by one.

1p(x) = 1(a2 x2 + a1 x + a0 )
= 1a2 x2 + 1a1 x + 1a0
= a2 x2 + a1 x + a0
= p(x).

Since the operations of addition and scalar multiplication on P2 satisfy the 8 vector space axioms, P2 is a
vector space. ♠
Our next example of a vector space is the set of all n × m-matrices.

Example 9.4: Vector space of matrices

Let Mm,n be the set of all m × n-matrices with entries in a field K , together with the usual operations
of matrix addition and scalar multiplication. Then Mm,n is a vector space.

Proof. The properties (A1)–(A4) hold by Proposition 4.11, and the properties (SM1)–(SM4) hold by
Proposition 4.14. ♠
We now examine an example of a set that does not satisfy all of the above axioms, and is therefore not a
vector space.

Example 9.5: Not a vector space

Let V denote the set of 2 × 3-matrices. Let us define a non-standard addition in V by A ⊕ B = A for
all matrices A, B ∈ V . Let scalar multiplication in V be the usual scalar multiplication of matrices.
Show that V is not a vector space.

Solution. In order to show that V is not a vector space, it suffices to find one of the 8 axioms that is not
satisfied. We will begin by examining the axioms for addition until one is found which does not hold. In
fact, for this example, the very first axiom fails. Let
   
1 0 0 0 0 0
A= , B= .
0 0 0 1 0 0

Then A ⊕ B = A and B ⊕ A = B. Since A 6= B, we have A ⊕ B 6= B ⊕ A for these two matrices, so property

(A1) is false. ♠
 9.1. Definition of vector spaces 329

Our next example looks a little different.

Example 9.6: Vector space of functions

Let X be a nonempty set, K a field, and define FuncX,K to be the set of functions defined on X and
valued in K . In other words, the elements of FuncX,K are functions f : X → K . The sum of two
functions is defined by
( f + g)(x) = f (x) + g(x),
and the scalar multiplication is defined by

(k f )(x) = k( f (x)).

Then FuncX,K is a vector space.

Proof. To verify that FuncX,K is a vector space, we must prove the 8 axioms of vector spaces. Let f , g, h
be functions in FuncX,K , and let k, ℓ be scalars. Recall that two functions f , g are equal if for all x ∈ X , we
have f (x) = g(x).

(A1) We prove the commutative law of addition. For all x ∈ X , we have

( f + g)(x) = f (x) + g(x) = g(x) + f (x) = (g + f )(x).
Therefore, f + g = g + f .
(A2) We prove the associative law of addition. For all x ∈ X , we have
(( f + g) + h)(x) = ( f + g)(x) + h(x) = ( f (x) + g(x)) + h(x)
= f (x) + (g(x) + h(x)) = ( f (x) + (g + h)(x)) = ( f + (g + h))(x).
Therefore, ( f + g) + h = f + (g + h).
(A3) To prove the existence of an additive unit, let 0 denote the function that is given by 0(x) = 0. This is
called the zero function. It is an additive unit because for all x,
( f + 0)(x) = f (x) + 0(x) = f (x),
and so f + 0 = f .
(A4) We prove the law of additive inverses. Let − f = (−1) f be the function that satisfies (− f )(x) =
− f (x). Then for all x,
( f + (− f ))(x) = f (x) + (− f )(x) = f (x) + − f (x) = 0.
Therefore f + (− f ) = 0.
(SM1) We prove the distributive law over vector addition. For all x, we have
(k( f + g))(x) = k( f + g)(x) = k( f (x) + g(x))
= k f (x) + kg(x) = (k f + kg)(x),
and so k( f + g) = k f + kg.
 330 Vector spaces

(SM2) We prove the distributive law over scalar addition.

((k + ℓ) f )(x) = (k + ℓ) f (x) = k f (x) + ℓ f (x) = (k f + ℓ f )(x),

and so (k + ℓ) f = k f + ℓ f .

(SM3) We prove the associative law for scalar multiplication.

((kℓ) f )(x) = (kℓ) f (x) = k(ℓ f (x)) = (k(ℓ f ))(x),

so (kℓ f ) = k(ℓ f ).

(SM4) Finally, we prove the rule for multiplication by one. For all x ∈ X , we have

(1 f )(x) = 1 f (x) = f (x),

and therefore 1 f = f .

It follows that FuncX,K satisfies all the required axioms and is a vector space. ♠
For the next two examples of vector spaces, we leave the proofs as an exercise.

Example 9.7: Infinite sequences

Let K be a field. A sequence of elements of K is an infinite list

(a0 , a1 , a2 , a3 . . .),

where ai ∈ K for all i. We also use the notation (ai )i∈N , or occasionally (ai ), to denote such a
sequence. Let SeqK be the set of sequences of elements of K . We add two sequences by adding
their i th elements:
(ai )i∈N + (bi )i∈N = (ai + bi )i∈N .
We scale a sequence by scaling each of its elements:

k(ai )i∈N = (kai )i∈N .

Then SeqK is a vector space.

Example 9.8: Vector space of polynomials of unbounded degree

Let K be a field, and let P be the set of all polynomials (of any degree) with coefficients from K ,
i.e., expressions of the form

p(x) = an xn + an−1 xn−1 + . . . + a1 x + a0 ,

where n ≥ 0 and a0 , . . . , an ∈ K . Addition and scalar multiplication of polynomials are defined in

the usual way . Then P is a vector space.
 9.1. Definition of vector spaces 331

We conclude this section by deriving some initial consequences of the vector space axioms.

Proposition 9.9: Elementary consequences of the vector space axioms

In any vector space, the following are true:

(a) The additive unit is unique. In other words, whenever u + v = u, then v = 0.

(b) Additive inverses are unique. In other words, whenever u + v = 0, then v = −u.

(c) 0u = 0 for all vectors u.

(d) The following cancellation law holds: if u + w = v + w, then u = v.

Proof. We prove the first three properties, and leave the last one as an exercise. Assume V is any vector
space over a field K.

(a) Consider arbitrary vectors u, v ∈ V and assume

u + v = u.

Applying the law (A1) (commutative law) to the left-hand side, we have

v + u = u.

Adding −u to both sides of the equation, we have

(v + u) + (−u) = u + (−u).

Applying the law (A2) (associative law) to the left-hand side, we have

v + (u + (−u)) = u + (−u).

Applying the law (A4) (additive inverse law) to both sides of the equation, we have

v + 0 = 0.

Applying the law (A3) (additive unit law) to the left-hand side, we have

v = 0.

This proves that whenever u + v = u, then v = 0, or in other words, v = 0 is the only element acting
as an additive unit.

(b) Consider arbitrary vectors u, v ∈ V and assume

u + v = 0.

Applying the law (A1) (commutative law) to the left-hand side, we have

v + u = 0.
332 Vector spaces

Adding −u to both sides of the equation, we have

(v + u) + (−u) = 0 + (−u).
Applying the law (A2) (associative law) to the left-hand side, we have
v + (u + (−u)) = 0 + (−u).
Applying the law (A4) (additive inverse law) to the left-hand side, we have
v + 0 = 0 + (−u).
Applying the law (A1) (commutative law) to the right-hand side, we have
v + 0 = −u + 0.
Applying the law (A3) (additive unit law) to both sides of the equation, we have
v = −u.
This proves that whenever u + v = 0, then v = −u, or in other words, v = −u is the only element
acting as an additive inverse of u.
(c) First, note that the scalar 0 ∈ K satisfies the property 0 + 0 = 0, by property (A3) of the definition of
a field. Now let u ∈ V be any vector. Using the vector space law (SM2) (distributive law over scalar
addition) and 0 + 0 = 0, we have
0u + 0u = (0 + 0)u = 0u.
Next, we use a small trick: add −(0u) to both sides of the equation. This gives
(0u + 0u) + (−(0u)) = 0u + (−(0u)).
Applying the additional laws (A2), (A4), and (A3), we have
0u + (0u + (−(0u))) = 0u + (−(0u)),
0u + 0 = 0,
0u = 0.
This proves that 0u = 0 holds for all vectors u, as desired.
(d) This is left as an exercise.

Exercises

Exercise 9.1.1 Consider the set R2 with the following non-standard addition operation ⊕:
(a, b) ⊕ (c, d) = (a + d, b + c).
 9.1. Definition of vector spaces 333

Scalar multiplication is defined in the usual way. Is this a vector space? Explain why or why not.

Exercise 9.1.2 Consider R2 with the following non-standard addition operation ⊕:

(a, b) ⊕ (c, d) = (0, b + d).

Scalar multiplication is defined in the usual way. Is this a vector space? Explain why or why not.

Exercise 9.1.3 Consider R2 with the following non-standard scalar multiplication:

c ⊙ (a, b) = (a, cb).

Vector addition is defined as usual. Is this a vector space? Explain why or why not.

Exercise 9.1.4 Consider R2 with the following non-standard addition operation ⊕:

(a, b) ⊕ (c, d) = (a − c, b − d).

Scalar multiplication is defined as usual. Is this a vector space? Explain why or why not.

Exercise 9.1.5 Prove that the set SeqK from Example 9.7 is a vector space. Hint: this is a special case of
Example 9.6, if you realize that a sequence (ai )i∈N is the same thing as a function a : N → K.

Exercise 9.1.6 Prove that the set P from Example 9.8 is a vector space.

Exercise 9.1.7 Let V be the set of functions defined on a set X that have values in a vector space W . Is
this a vector space? Explain.

Exercise 9.1.8 Consider the set R2 with the following non-standard operations of addition and scalar
multiplication:
(a, b) ⊕ (c, d) = (a + c − 1, b + d − 1),
k ⊙ (c, d) = (kc + (1 − k), kd + (1 − k)).
Show that R2 is a vector space with these operations. Hint: the zero vector is not (0, 0), but (1, 1).

Exercise 9.1.9 Consider the set R of real numbers. Addition of real numbers is defined in the usual way,
and scalar multiplication is just multiplication of one real number by another. In other words, x + y means
to add the two numbers and xy means to multiply them. Show that R, with these operations, is a real vector
space.

Exercise 9.1.10
√ Let K = Q be the field of rational numbers, and let V be the set of real numbers of the
form a + b 2, where a and b are rational numbers. Show that with the usual operations, V is a Q-vector
space.

Exercise 9.1.11 Let K = Q be the field of rational numbers, and let V = R be the set of real numbers.
Show that with the usual operations, V is a Q-vector space.
334 Vector spaces

Exercise 9.1.12 Let P3 be the set of all polynomials of degree 3 or less. That is, these are of the form
ax3 + bx2 + cx + d. Addition and scalar multiplication of polynomials are defined as usual. Show that P3
is a vector space.

Exercise 9.1.13 Let X = {1, 2, . . . , n}, and consider the space FuncX,R of real-valued functions defined
on X . Explain how FuncX,R can be considered as Rn .

Exercise 9.1.14 Prove the cancellation law from Proposition 9.9.

9.2 Linear combinations, span, and linear independence

Outcomes
A. Determine if a vector is within a given span.

B. Determine if a set is spanning.

C. Determine if a set is linearly independent.

In this section, we will again explore concepts introduced earlier in terms of Rn and extend them to apply
to abstract vector spaces.
We can now revisit many of the concepts first introduced in Chapter 2 in the context of general vec-
tor spaces. We will look at linear combinations, span, and linear independence in this section, and at
subspaces, bases, and dimension in the next section.

Definition 9.10: Linear combination

Let V be a vector space over a field K . Let u1 , . . . , un ∈ V . A vector v ∈ V is called a linear
combination of u1 , . . . , un if there exist scalars a1 , . . . , an ∈ K such that

v = a1 u1 + . . . + an un .

Example 9.11: Linear combination of matrices

 
1 3
Write the matrix A = as a linear combination of
−1 2
       
1 0 1 0 0 1 0 −1
, , , and .
0 1 0 −1 1 0 1 0
 9.2. Linear combinations, span, and linear independence 335

Solution. We must find coefficients a, b, c, d such that

         
1 3 1 0 1 0 0 1 0 −1
= a +b +c +d ,
−1 2 0 1 0 −1 1 0 1 0

or equivalently,    
1 3 a+b c−d
= .
−1 2 c+d a−b
This yields a system of four equations in four variables:

a+b = 1,
c+d = −1,
c−d = 3,
a−b = 2.

We can easily solve the system of equations to find the unique solution a = 32 , b = − 12 , c = 1, d = −2.
Therefore          
1 3 3 1 0 1 1 0 0 1 0 −1
= − +1 −2 .
−1 2 2 0 1 2 0 −1 1 0 1 0
♠

Example 9.12: Linear combination of polynomials

Write the polynomial p(x) = 7x2 + 4x − 3 as a linear combination of

q1 (x) = x2 , q2 (x) = (x + 1)2 , and q3 (x) = (x + 2)2 .

Solution. Note that q2 (x) = (x + 1)2 = x2 + 2x + 1 and q3 (x) = (x + 2)2 = x2 + 4x + 4. We must find
coefficients a, b, c such that p(x) = aq1 (x) + bq2 (x) + cq3 (x), or equivalently,

7x2 + 4x − 3 = ax2 + b(x2 + 2x + 1) + c(x2 + 4x + 4).

Collecting equal powers of x, we can rewrite this as

7x2 + 4x − 3 = (a + b + c)x2 + (2b + 4c)x + (b + 4c).

Since two polynomials are equal if and only if each corresponding coefficient is equal, this yields a system
of three equations in three variables
a + b + c = 7,
2b + 4c = 4,
b + 4c = −3.
We can easily solve this system of equations and find that the unique solution is a = 52 , b = 7, c = − 52 .
Therefore
5 5
p(x) = q1 (x) + 7 q2 (x) − q3 (x).
2 2
♠
336 Vector spaces

As in Chapter 2, the span of a set of vectors is defined as the set of all of its linear combinations. We
generalize the concept of span to consider spans of arbitrary (possibly finite, possibly infinite) sets of
vectors.

Definition 9.13: Span of a set of vectors

Let V be a vector space over some field K , and let S be a set of vectors (i.e., a subset of V ). The
span of S is the set of all linear combinations of elements of S. In symbols, we have

span S = {a1 u1 + . . . + ak uk | u1 , . . . , uk ∈ S and a1 , . . . , ak ∈ K } .

It is important not to misunderstand this definition. Even when the set S is infinite, each individual element
v ∈ span S is a linear combination of only finitely many elements u1 , . . . , uk of S. The definition does not
talk about infinite linear combinations
a1 u1 + a2 u2 + a3 u3 + . . .
Indeed, such infinite sums do not typically exist. However, different elements v, w ∈ span S can be linear
combinations of a different (finite) number of vectors of S. For example, it is possible that v is a linear
combination of 10 elements of S, and w is a linear combination of 100 elements of S.

Example 9.14: Spans of sequences

Consider the vector space SeqK of infinite sequences. For every k ∈ N, let ek be the sequence whose
k th element is 1 and that is 0 everywhere else, i.e.,

e0 = (1, 0, 0, 0, 0, . . .),
e1 = (0, 1, 0, 0, 0, . . .),
e2 = (0, 0, 1, 0, 0, . . .),

and so on. Let S = ek | k ∈ N . Which of the following sequences are in span S?

(a) f = (1, 1, 1, 0, 0, 0, 0, 0, . . .) (followed by infinitely many zeros),

(b) g = (1, 2, 0, 5, 0, 0, 0, 0, . . .) (followed by infinitely many zeros),

(c) h = (1, 1, 1, 1, 1, 1, 1, 1, . . .) (followed by infinitely many ones),

(d) k = (1, 0, 1, 0, 1, 0, 1, 0, . . .) (forever alternating between 1 and 0).

Solution.

(a) We have f ∈ span S, because f = e0 + e1 + e2 .

(b) We have g ∈ span S, because g = 1e0 + 2e1 + 5e3 .
(c) The sequence h is not in span S, because each element of span S is, by definition, a linear combina-
tions of finitely many elements of S. No linear combinations of finitely many ek can end in infinitely
many ones. Note that we are not permitted to write an infinite sum such as e0 + e1 + e2 + . . .. Such
infinite sums are not defined in vector spaces.
 9.2. Linear combinations, span, and linear independence 337

(d) The sequence k is not in span S, for the same reason. We would need to add infinitely many se-
quences of the form ek to get a sequence that contains infinitely many non-zero elements. However,
this is not permitted by the definition of span. ♠

Example 9.15: Span of polynomials


Let p(x) = 7x2 + 4x − 3. Is p(x) ∈ span x2 , (x + 1)2 , (x + 2)2 ?

Solution. The answer is yes, because we found in Example 9.12 that p(x) = 52 x2 + 7 (x+1)2 − 52 (x+2)2 .
♠
We say that a set of vectors S is a spanning set for V if V = span S.

Example 9.16: Spanning set


Let S = x2 , (x + 1)2 , (x + 2)2 . Show that S is a spanning set for P2 , the vector space of all
polynomials of degree at most 2.

Solution. This is analogous to Example 9.12. Consider an arbitrary element p(x) = p2 x2 + p1 x + p0 of

P2 . We must show that p(x) ∈ span S, i.e., that there exists a, b, c ∈ K such that
p(x) = ax2 + b(x + 1)2 + c(x + 2)2 .
We can equivalently rewrite this equation as
p2 x2 + p1 x + p0 = (a + b + c)x2 + (2b + 4c)x + (b + 4c),
which yields the system of equations
   
a + b + c = p2 1 1 1 p2 1 0 0 p2 − 34 p1 + 12 p0
2b + 4c = p1 ≃  0 2 4 p1  ≃  0 1 0 p1 − p0 .
b + 4c = p0 0 1 4 p0 1 1
0 0 1 2 p0 − 4 p1

Since the system has rank 3, it has a solution. Therefore, p(x) ∈ span S. Since p(x) was an arbitrary
element of P2 , it follows that S is a spanning set for P2 . ♠
To define the concept of linear independence in a general vector space, it will be convenient to base our
definition on the “alternative” characterization of Theorem 5.12. Here too, we generalize the definition to
an arbitrary (finite or infinite) set of vectors.

Definition 9.17: Linear independence

Let V be a vector space over some field K . A finite set of vectors {u1 , . . . , uk } is called linearly
independent if the equation
a1 u1 + . . . + ak uk = 0
has only the trivial solution a1 , . . . , ak = 0. An infinite set S of vectors is called linearly independent
if every finite subset of S is linearly independent. A set of vectors is called linearly dependentif it
is not linearly independent.
338 Vector spaces

Example 9.18: Linearly independent polynomials

Determine whether the polynomials x2 , x2 + 2x − 1, and 2x2 − x + 3 are linearly independent.

Solution. According to the definition of linear independence, we must solve the equation

ax2 + b(x2 + 2x − 1) + c(2x2 − x + 3) = 0.

If there is a non-trivial solution, the polynomials are linearly dependent. If there is only the trivial solution,
they are linearly independent. We first rearrange the left-hand side to collect equal powers of x:

(a + b + 2c)x2 + (2b − c)x + (3c − b) = 0.

This turns into a system of 3 equations in 3 variables:

   
a + b + 2c = 0 1 1 2 0 1 0 0 0
2b − c = 0 ≃  0 2 −1 0  ≃  0 1 0 0 .
3c − b = 0 0 −1 3 0 0 0 1 0

Since the system has rank 3, there are no free variables. The only solution is a = b = c = 0, and the
polynomials are linearly independent. ♠

Example 9.19: Linearly independent sequences

Let K be a field, and consider again the sequences from Example 9.14,

e0 = (1, 0, 0, 0, 0, . . .),
e1 = (0, 1, 0, 0, 0, . . .),
e2 = (0, 0, 1, 0, 0, . . .),

and so on. Let S = e0 , e1 , e2 , . . . . This is an infinite subset of SeqK . Show that S is linearly
independent.

Solution. Since S is an infinite set, we have to show that every finite subset of S is linearly independent.
So consider a finite subset n o
k1 k2 kn
e ,e ,...,e ⊆S
and assume that
a1 ek1 + a2 ek2 + . . . + an ekn = 0. (9.1)
We have to show that a1 , . . . , an = 0. Consider some index i ∈ {1, . . . , n}. Then the kith element of a1 ek1 +
. . .+an ekn is equal to ai by the left-hand side of (9.1), but itis also equal to 0 by the right-hand side of (9.1).
It follows that ai = 0 for all i ∈ {1, . . . , n}, and therefore ek1 , ek2 , . . . , ekn is linearly independent. Since
 k k
e 1 , e 2 , . . . , ekn was an arbitrary finite subset of S, it follows, by definition, that S is linearly independent.
♠
 9.2. Linear combinations, span, and linear independence 339

Example 9.20: Linearly dependent matrices

Determine whether the following elements of Mm,n are linearly independent:
     
−1 0 1 1 1 3
M1 = , M2 = , M3 = .
1 −1 1 2 5 4

Solution. To determine whether {M1 , M2 , M3 } is linearly independent, we look for solutions to

aM1 + bM2 + cM3 = 0.

Notice that this equation has non-trivial solutions, for example a = 2, b = 3 and c = −1. Therefore the
matrices are linearly dependent. ♠

Example 9.21: Linearly independent functions

In the vector space FuncR,R of real-valued functions on the real numbers, show that the functions
f (x) = sin x and g(x) = cos x are linearly independent.

Solution. Assume A sin x + B cos x = 0. Note that this is an equality of functions, which means that it
is true for all x. In particular, substituting x = 0 into the equation, and using the fact that sin 0 = 0 and
cos 0 = 1, we have
0 = A sin 0 + B cos 0 = A · 0 + B · 1 = B,
π
and therefore B = 0. On the other hand, substituting x = 2 into the equation, and using the fact that
sin π2 = 1 and cos π2 = 0, we have
π π
0 = A sin + B cos = A · 1 + B · 0 = A,
2 2
and therefore A = 0. Therefore, the equation A sin x + B cos x = 0 only has the trivial solution A = B = 0,
and it follows that sin x and cos x are linearly independent. ♠
The properties of linear independence that were discussed in Chapter 2 remain true in the general setting
of vector spaces. For example, the first two parts of Proposition 5.16 apply without change. (The third
part specifically mentions Rn , but can be generalized to any vector space of dimension n). We also have
the usual characterization of linear dependence in terms of redundant vectors:

Proposition 9.22: Linear dependence and redundant vectors

Let V be a vector space, and let u1 , u2 , . . . be a (finite or infinite) sequence of vectors in V . If
u1 , u2 , . . . are linearly dependent, then at least one of the vectors can be written as a linear combina-
tion of earlier vectors in the sequence:

u j = a1 u1 + a2 u2 + . . . + a j−1 u j−1 ,

for some j.
340 Vector spaces

Proof. Suppose that the vectors are linearly dependent. Then the equation b1 u1 + . . . + bk uk = 0 has a
non-trivial solution for some k. In other words, there exist scalars b1 , . . . , bk , not all equal to zero, such that
b1 u1 + . . . + bk uk = 0. Let j be the largest index such that b j 6= 0. Then b1 u1 + . . . + b j u j = 0. Dividing by
b
b j and solving for u j , we have u j = − bb1j u1 − . . . − bj−1
j
u j−1 , so u j can be written as a linear combination
of earlier vectors as claimed. ♠

Example 9.23: Polynomials of increasing degree

Consider a sequence of non-zero polynomials p1 (x), . . . , pk (x) of increasing degree, i.e., such that
the degree of each pi (x) is strictly larger than that of pi−1 (x). Show that p1 (x), . . . , pk (x) are linearly
independent in the vector space P.

Solution. A polynomial of degree n cannot be a linear combination of polynomials of degree less than
n. Therefore, none of the polynomials p1 (x), . . ., pk (x) can be written as a linear combination of earlier
polynomials. By Proposition 9.22, p1 (x), . . . , pk (x) are linearly independent. ♠
Theorems 5.18 and 5.19 also remain true in the setting of general vector spaces. The original proofs can be
used without change. Thus, if u1 , . . . , uk are linearly independent, then every vector v ∈ span {u1 , . . . , uk }
can be uniquely written as a linear combination of u1 , . . . , uk . Also, given any finite set of vectors, we can
find a subset of the vectors that is linearly independent and has the same span.
We finish this section with a useful observation about linear independence. Namely, given a linearly
independent set of vectors and one more vector that is not in their span, then we can add the vector to the
set and it will remain linearly independent.

Proposition 9.24: Adding to a linearly independent set

Suppose {u1 , . . . , uk } is linearly independent and v ∈
/ span {u1 , . . . , uk }. Then the set

{u1 , . . . , uk , v}

is also linearly independent.

Proof. Assume, on the contrary, that the set were linearly dependent. Then by Proposition 9.22, one of
the vectors can be written as a linear combination of earlier vectors. This vector cannot be one of the ui ,
because u1 , . . . , uk are linearly independent. It also cannot be v, because v ∈
/ span {u1 , . . . , uk }. Therefore,
our assumption cannot be true, and the set is linearly independent. ♠

Exercises

Exercise 9.2.1 Let V be a vector space and suppose {u1 , . . . , uk } is a set of vectors in V . Show that 0 is in
span {u1 , . . . , uk }.
 9.2. Linear combinations, span, and linear independence 341


Exercise 9.2.2 Determine whether p(x) = 4x2 − x is in span x2 + x, x2 − 1, − x + 2 .

Exercise 9.2.3 Determine whether p(x) = −x2 + x + 2 is in span x2 + x + 1, 2x2 + x .

Exercise 9.2.4
 
1 3
(a) Write A = as a linear combination of
0 0
       
1 0 0 1 1 0 0 1
, , , .
0 1 1 0 1 1 1 1

(b) Show that the above set of four matrices is a spanning set for M2,2 , the vector space of all 2 × 2-
matrices.

Exercise 9.2.5 Let K be a field, and consider the vector space SeqK of infinite sequences of scalars. A
sequence a = (ai )i∈N is called finitely supported if all but finitely many elements of the sequence are zero.
In other words, a is finitely supported if there exists some N ∈ N such that ak = 0 for all k ≥ N. Let
e0 , e1 , e2 , . . . be the sequences from Example 9.14. Show that a ∈ span e0 , e1 , e2 , . . . if and only if a is
finitely supported.

Exercise 9.2.6 For each of the following sets of polynomials, determine whether the set is linearly inde-
pendent. If it is linearly dependent, write one polynomial as a linear combination of the other polynomials
in the set.

(a) x + 1, x2 + 2, x2 − x − 3 .

(b) x2 + x, − 2x2 − 4x − 6, 2x − 2 .

Exercise 9.2.7 Determine whether each of the following sets of matrices is linearly independent. If it is
linearly dependent, write one matrix as a linear combination of the other matrices in the set.
     
1 2 −7 2 4 0
(a) , , .
0 1 −2 −3 1 2
       
1 0 0 1 1 0 0 0
(b) , , , .
0 1 0 1 1 0 1 1

Exercise 9.2.8 Consider polynomials

 3
ai x + bi x2 + ci x + di | i = 1, 2, 3, 4 .
Show that this collection of polynomials is linearly independent if and only if
 
a1 b1 c1 d1
 a2 b2 c2 d2 
 
 a3 b3 c3 d3 
a4 b4 c4 d4
342 Vector spaces

is an invertible matrix.

Exercise 9.2.9 Assume u, v, w are linearly independent elements of some vector space V . Consider the
set of vectors 
R = 2u − w, w + v, 3v + 12 u .
Determine whether R is linearly independent.

9.3 Subspaces

Outcomes
A. Determine whether a set of vectors is a subspace of a given vector space.

B. Determine whether two sets of vectors span the same subspace.

In this section we will consider subspaces of general vector spaces.

Definition 9.25: Subspace

Let V be a vector space over a field K . A subset W ⊆ V is said to be a subspace of V if the following
conditions hold:

1. 0 ∈ W , where 0 is the additive unit of V .

2. W is closed under addition: Whenever u, v ∈ W , then u + v ∈ W .

3. W is closed under scalar multiplication: Whenever k ∈ K and u ∈ W , then ku ∈ W .

Example 9.26: Subspaces of R3

As we have seen in Section 5.3, the subspaces of R3 are:

• the zero subspace {0};

• lines through the origin;

• planes through the origin;

• R3 itself.
 9.3. Subspaces 343

Example 9.27: Space of continuous functions

Let V = FuncR,R , the vector space of functions from real numbers to real numbers. Let W ⊆ V be
the subset of continuous functions. Then W is a subspace of V .

Proof. We know from calculus that:

1. the zero function, defined by f (x) = 0 for all x, is continuous;

2. if f , g are continuous functions, then f + g is continuous;

3. if f is a continuous function and k a constant, then k f is continuous.

It follows that W contains 0, and is closed under addition and scalar multiplication. Therefore, W is a
subspace of V . ♠

Example 9.28: Space of differentiable functions

Let V = FuncR,R . Recall from calculus that a function f : R → R is called differentiable if the
derivative f ′ (x) exists for all x ∈ R. Let W ⊆ V be the subset of differentiable functions. Then W is
a subspace of V .

Proof. We know from calculus that:

1. The zero function, defined by f (x) = 0 for all x, is differentiable. In fact, its derivative is f ′ (x) = 0.

2. If f , g are differentiable functions, then h = f + g is differentiable. In fact, h′ (x) = f ′ (x) + g′ (x).

3. if f is a differentiable function and k a constant, then h = k f is differentiable. In fact, h′ = k f ′ .

It follows that W contains 0, and is closed under addition and scalar multiplication. Therefore, W is a
subspace of V . ♠

Example 9.29: Space of sequences satisfying a linear recurrence

Let V = SeqR , the vector space of sequences of real numbers. Let

W = {a ∈ SeqR | for all n ≥ 0, an+2 = an + an+1 } .

In other words, W is the set of all sequences satisfying the recurrence relation an+2 = an + an+1 .
Then W is a subspace of V .

Proof. Before we prove that W is a subspace, let us first consider an example. The following sequences
are elements of W , because they both satisfy the recurrence:

a = (1, 1, 2, 3, 5, 8, 13, 21, . . .) ,

b = (1, 3, 4, 7, 11, 18, 29, 47, . . .) .
344 Vector spaces

Note that if we add these sequences, we get

a + b = (2, 4, 6, 10, 16, 26, 42, 68, . . .) ,

which again satisfies the recurrence. Therefore, the set W is closed under the addition of these particular
sequences a and b. We now prove the properties in general.

1. Let z be the zero sequence, defined by zn = 0 for all n. Then z satisfies the recurrence relation, since
for all n ≥ 0, zn+2 = 0 = zn + zn+1 . Therefore z ∈ W .

2. To show that W is closed under addition, consider any two sequences a, b ∈ W , and let c = a + b.
Then for all n ≥ 0,

cn+2 = an+2 + bn+2 = (an + an+1 ) + (bn + bn+1 ) = (an + bn ) + (an+1 + bn+1 ) = cn + cn+1 ,

so c satisfies the recurrence. It follows that c ∈ W , and therefore W is closed under addition.

3. To show that W is closed under scalar multiplication, consider any k ∈ R and a ∈ W , and let c = ka.
Then for all n ≥ 0,

cn+2 = kan+2 = k(an + an+1 ) = kan + kan+1 = cn + cn+1 ,

so c satisfies the recurrence. It follows that c ∈ W , and therefore W is closed under scalar multipli-
cation.

Example 9.30: Solution space of a linear differential equation

Let V = FuncR,R . Recall from calculus that a differential equation is an equation about an un-
known function and its derivatives. For example

f ′′ = − f

is a differential equation. The functions f (x) = sin x, f (x) = cos x, and f (x) = 0 are examples of
solutions of this differential equation. Let W be the set of all functions that are solutions of the
differential equation f ′′ = − f . Then W is a subspace of V .

Proof.

1. The zero function f (x) = 0 is a solution of the differential equation, and therefore an element of W .

2. To show that W is closed under addition, let f , g ∈ W and consider h = f + g. Then f ′′ = − f and
g′′ = −g, and therefore h′′ = f ′′ + g′′ = − f + (−g) = −h. Therefore, h ∈ W , and W is closed under
addition.

3. To show that W is closed under scalar multiplication, let k ∈ R and f ∈ W , and consider h = k f .
Then f ′′ = − f , and therefore h′′ = k f ′′ = k(− f ) = −h. It follows that h ∈ W , and therefore W is
closed under scalar multiplication.
 9.3. Subspaces 345

Example 9.31: Subspace of polynomials

Consider P2 , the vector space of polynomials of degree at most 2, with coefficients in a field K . Fix
some element r ∈ K , and let W ⊆ P2 be the subset of polynomials that have r as a root. Then W is
a subspace of P2 .

Proof. We can express W as follows:

W = {p ∈ P2 | p(r) = 0} .

We need to show that W is a subspace.

1. The zero polynomial, given by 0(x) = 0, satisfies 0(r) = 0, so 0 ∈ W .

2. To show that W is closed under addition, assume p, q ∈ W , and let s = p + q. Then p(r) = 0 and
q(r) = 0, therefore s(r) = p(r) + q(r) = 0. It follows that s ∈ W .

3. To show that W is closed under scalar multiplication, assume p(x) ∈ W and k be a scalar. Then
(kp)(r) = k(p(r)) = k0 = 0, and therefore kp ∈ W .

Example 9.32: Trivial subspaces of V

Let V be an arbitrary vector space over a field K . Then {0} is a subspace of V , called the zero
subspace. Also, V is a subspace of itself.

Proof. Clearly {0} contains 0, and is closed under addition and scalar multiplication because 0 + 0 = 0
and k0 = 0 for all k. Similarly, V contains 0 and is closed under addition and scalar multiplication, because
addition and scalar multiplication are operations on V . Therefore, both {0} and V are subspaces of V . ♠
The interest of subspaces lies in the fact that they are vector spaces in their own right, as stated in the
following proposition.

Proposition 9.33: Subspaces are vector spaces

Let W be a subspace of a vector space V . Then W satisfies the vector space axioms (A1)–(A4)
and (SM1)–(SM4), with respect to the same operations (addition and scalar multiplication) as those
defined on V .

Proof. Since W is a subspace, it is closed under addition and scalar multiplication. This ensures that
addition and scalar multiplication are well-defined operations on W . The axioms (A1), (A2), (A4), and
(SM1)–(SM4) all obviously hold in W , because they hold in V (two elements of W are equal in W if and
only if they are equal in V ). The axiom (A3) holds because 0 ∈ W . ♠
346 Vector spaces

We end this section with an observation about spans.

Proposition 9.34: Span is smallest subspace containing given vectors

Let V be a vector space over some field K , and consider a set of vectors S ⊆ V . Then span S is the
smallest subspace of V containing S. More explicitly, we have:

(a) The set span S is a subspace of V , and S ⊆ span S.

(b) If W is any other subspace of V such that S ⊆ W , then span S ⊆ W .

Proof.

(a) To show that span S is a subspace, first note that 0 ∈ span S, because 0 is the empty linear combi-
nation. Also, if v, u ∈ span S, then by definition of span, there exist v1 , . . . , vk , u1 , . . . , uℓ ∈ S and
a1 , . . . , ak , b1 , . . . , bℓ ∈ K such that
v = a1 v1 + . . . + ak vk ,
u = b1 u 1 + . . . + bℓ u ℓ .
Then
v + u = a1 v1 + . . . + ak vk + b1 u1 + . . . + bℓ u,
and therefore v + u ∈ span S. It follows that span S is closed under addition. The proof for scalar
multiplication is similar. Finally, every v ∈ S is trivially a linear combination of itself, v = 1v, and
therefore S ⊆ span S.
(b) Consider any other subspace W of V such that S ⊆ W . To show that span S ⊆ W , consider an
arbitrary element v ∈ span S. By definition of span, there exist v1 , . . . , vk ∈ S and a1 , . . . , ak ∈ K such
that v = a1 v1 + . . . + ak vk . By assumption, v1 , . . . , vk ∈ W . Since W is closed under addition and
scalar multiplication, it follows that v ∈ W . Since v was an arbitrary element of span S, it follows
that span S ⊆ W .

♠
While the last proposition looks technical, it can actually be useful for proving that two sets of vectors
span the same subspace. The following in an example of this.

Example 9.35: Equal spans

 
Show that the sets S = x2 − 2x + 1, x − 1 and T = x2 − 1, x2 − x span the same subspace of
P2 . In other words, show that
span S = span T .

Solution. To show that two sets are equal, we must show that each is a subset of the other. So we will
show span S ⊆ span T and span T ⊆ span S. By Proposition 9.34, it is sufficient to show that S ⊆ span T
and T ⊆ span S, i.e., we must show that every element of S is a linear combination of elements of T and
vice versa.
 9.3. Subspaces 347

1. S ⊆ span T . We have

x2 − 2x + 1 = (−1)(x2 − 1) + 2(x2 − x) ∈ span T ,

x − 1 = 1(x2 − 1) − 1(x2 − x) ∈ span T .

Since each element of S is an element of span T , it follows that S ⊆ span T . By Proposition 9.34,
this implies that span S ⊆ span T .

2. T ⊆ span S. We have

x2 − 1 = 1(x2 − 2x + 1) − 2(x − 1) ∈ span S,

x2 − x = 1(x2 − 2x + 1) − 1(x − 1) ∈ span S.

Since each element of T is an element of span S, it follows that T ⊆ span S. By Proposition 9.34,
this implies that span T ⊆ span S.

Exercises

Exercise 9.3.1 Consider the set of symmetric n × n-matrices, i.e., matrices satisfying A = AT . Show that
this set of symmetric matrices is a subspace of Mn,n , the vector space of n × n-matrices.
 
Exercise 9.3.2 Consider the set of all vectors xy ∈ R2 such that x + y ≥ 0. Is this a subspace of R2 ?
 
Exercise 9.3.3 Consider the set of all vectors xy ∈ R2 such that xy = 0. Is this a subspace of R2 ?

Exercise 9.3.4 Let V be the set of those polynomials ax2 + bx + c ∈ P2 such that a + b + c = 0. Is V a
subspace of P2 ? Explain.

Exercise 9.3.5 Let U ,W be subspaces of a vector space V and consider U + W defined as the set of all
vectors that can be written of the form u + w, where u ∈ U and w ∈ W . Show that U +W is a subspace of
V.

Exercise 9.3.6 Let U ,W be subspaces of a vector space V . Then U ∩W consists of all vectors which are
in both U and W . Show that U ∩W is a subspace of V .

Exercise 9.3.7 Let U ,W be subspaces of a vector space V . Then U ∪W consists of all vectors which are
in either U or W . Show that U ∪W is not necessarily a subspace of V by giving an example where U ∪W
fails to be a subspace.
n  o

Exercise 9.3.8 Let U = [x, y, z]T ∈ R3  |x| ≤ 4 . Is U a subspace of R3 ?
348 Vector spaces

Exercise 9.3.9 Let W be the subset of M2,2 given by

nh i o
W= a b  a, b, c, d ∈ K, a + b = c + d
c d
Is W a subspace of M2,2 ?
n  h i h i o
 1 0 1 1
Exercise 9.3.10 Let U = A ∈ M2,2  A 1 −1 = 0 −1 A . Show that U is a subspace of M2,2 .

Exercise 9.3.11 Let W be the subset of P3 given by


W = ax3 + bx2 + cx + d | a, b, c, d ∈ K, d = 0
Is W a subspace of P3 ?

Exercise 9.3.12 Let W be the subset of P3 given by

W = {p(x) | p(2) = 1}
Is W a subspace of P3 ?

Exercise 9.3.13 Let W be the subset of SeqR consisting of all sequences that are alternating, i.e., where
ai ≥ 0 for even i and ai ≤ 0 for odd i, or vice versa. Is W a subspace of SeqR ?

Exercise 9.3.14 Let W be the subset of SeqK consisting of all sequences that satisfy the recurrence relation
an+3 = an+2 + 2an+1 − an . Is W a subspace of SeqK ?

Exercise 9.3.15 A sequence (ai )i∈N is called periodic if there exists some k > 0 such that for all i, ai+k = ai .
The number k is called a period of the sequence. For example, the following is a periodic sequence with
period 3:
(1, 5, −7, 1, 5, −7, 1, 5, −7, 1, 5, −7, 1, . . .).

(a) Show that the set of all periodic sequences of a fixed period k forms a subspace of SeqK .
(b) More difficult: show that the set of all periodic sequences of all periods forms a subspace of SeqK .

Exercise 9.3.16 A function f : R → R is called symmetric if f (x) = f (−x) for all x ∈ R. Let W be the
subset of FuncR,R consisting of all symmetric functions. Is W a subspace of FuncR,R ?

Exercise 9.3.17 A function f : R → R is said to vanish at infinity if limx→∞ f (x) = 0 and limx→−∞ f (x) =
0. Let W be the subset of FuncR,R consisting of all functions that vanish at infinity. Prove that W is a
subspace of FuncR,R .
 
Exercise 9.3.18 Show that the sets S = x + 2, (x + 2)2 and T = x2 − 4, x2 + x − 2 span the same
subspace of P2 .
nh i h io nh i h io
Exercise 9.3.19 Show that the sets S = 1 0 0 1 and T = 1 1 1 1
0 1 , 1 0 1 −1 , −1 1 span the
same subspace of M2,2 .
 9.4. Basis and dimension 349

9.4 Basis and dimension

Outcomes
A. Find a basis of a given vector space.

B. Determine the dimension of a vector space.

C. Extend a linearly independent set of vectors to a basis.

D. Shrink a spanning set of vectors to a basis.

Definition 9.36: Basis

Let V be a vector space. A set B of vectors is called a basis of V if

1. B is a spanning set for V , and

2. B is linearly independent.

Example 9.37: Bases of P2

Consider the vector space P2 of polynomials of degree at most 2 with coefficients in a field K .

• 1, x, x2 is a basis of P2 .

• x2 , (x + 1)2 , (x + 2)2 is a basis of P2 .

• 1, x − 1, (x − 1)2 is a basis of P2 .

Unlike Rn , a vector space like P2 does not necessarily have a “standard” basis. One basis might be
useful for one application, and another basis for a different application.

Proof. It is easy to verify that each set of vectors is linearly independent and spanning. See Examples 9.16,
9.18, and 9.23 for similar calculations. ♠

Example 9.38: An infinite basis

Consider the vector space P of all polynomials with coefficients in a field K . The following is a
basis for P: 
1, x, x2 , x3 , x4 , . . . .
Note that this basis is infinite.

Proof. The polynomials 1, x, x2 , x3 , x4 , . . . are linearly independent by Proposition 9.22. Namely, if they
were linearly dependent, then one of the polynomials could be written as a linear combination of earlier
350 Vector spaces

ones. However, this is not possible because a polynomial of degree n cannot be a linear combination of
polynomials of degree less than n.
To show that the polynomials 1, x, x2 , x3 , x4 , . . . are a spanning set, consider an arbitrary element p(x)
of P. Then by definition, p(x) is of the form
p(x) = an xn + an−1 xn−1 + . . . + a1 x + a0 ,
forsome n ≥ 0 and a0 , . . . , an ∈ K. But then p(x) is a linear combination of 1, . . . , xn , i.e., it is in the span
of 1, x, x2 , x3 , x4 , . . . . ♠

Example 9.39: Not a basis

Consider the vector space SeqK of infinite sequences. As before, let ek be the sequence whose k th
element is 1 and that is 0 everywhere else, i.e.,

e0 = (1, 0, 0, 0, 0, . . .),
e1 = (0, 1, 0, 0, 0, . . .),
e2 = (0, 0, 1, 0, 0, . . .),

and so on. Then the set  0 1 2

e , e , e , ...
is not a basis of SeqK . Indeed, although we saw in Example 9.19 that the sequences e0 , e1 , e2 , . . .
are linearly independent, Example 9.14 shows that they are not spanning. Indeed,

W = span e0 , e1 , e2 , . . .

is a subspace of SeqK , consisting exactly of the finitely supported

 0 1 sequences,
i.e., those sequences
2
that have only finitely many non-zero components. Thus, e , e , e , . . . is a basis of W .

The following theorem ensures that every vector space has a basis. We will not prove this theorem, because
when the spaces are infinite-dimensional, the proof uses mathematics that is beyond the scope of this book.
The proof uses a reasoning principle called the axiom of choice, which allows us to prove the existence
of a basis even in cases where we cannot find an actual concrete example of a basis. For example, it is
not possible to give a specific example of a basis for the space SeqK , even though the following theorem
guarantees that such a basis exists.

Theorem 9.40: Existence of bases

Every vector space has a basis.

The Exchange Lemma, which we proved in the context of Rn in Section 5.4, is true in general vector
spaces.

Lemma 9.41: Exchange Lemma

Let V be a vector space over a field K . Suppose u1 , . . . , ur are linearly independent elements of
span {v1 , . . . , vs }. Then r ≤ s.
 9.4. Basis and dimension 351

The proof is exactly the same as that of Lemma 5.37, so we do not repeat it here. As in Section 5.4, an
important consequence of the Exchange Lemma is that any two bases of a vector space have the same size.

Theorem 9.42: Bases are of the same size

Let V be a vector space over some field K , and let B1 and B2 be bases of V . Then either B1 and B2
are both finite and have the same number of elements, or else B1 and B2 are both infinite.

Proof. We first show that B1 and B2 are either both finite or both infinite. Assume one of them, say B1 ,
is finite and contains s vectors. Since B1 is spanning and B2 is linearly independent, it follows from the
Exchange Lemma that B2 cannot contain more than s vectors, and in particular, B2 must be finite. So the
sets are either both finite or both infinite. If they are both finite, say of size s and r, then by the Exchange
Lemma, we have s ≤ r and r ≤ s, hence r = s. ♠
This allows us to define the dimension of a vector space.

Definition 9.43: Dimension

Let V be a vector space over a field K . If V has a basis consisting of n vectors, we say that V has
dimension n, and we write dim(V ) = n. In this case we also say that V is finite-dimensional. If V
has an infinite basis, we say that V is infinite-dimensional, and we write dim(V ) = ∞.

Note that the dimension is well-defined by Theorems 9.40 and 9.42, since these theorems ensure that
every vector space has a basis (and therefore a dimension), and that any two bases are of the same size
(and therefore a vector space cannot have more than one dimension).
We now calculate the dimensions of some vector spaces we encountered in Sections 9.1 and 9.3.

• The space Rn has dimension n.

• The space P2 has dimension 3. We found several bases for this space in Example 9.37.

• The space Mm,n has dimension mn. A possible basis consists of all the matrices that contain a single
1 and zeros everywhere else.

• The space FuncX,K is infinite-dimensional if X is an infinite set. If X is a finite set of n elements,

then this space is n-dimensional. In that case, a basis is given by the set of functions whose value is
1 for one input and 0 for all other inputs.

• The space SeqK is infinite-dimensional. We found an infinite linearly independent set in Exam-
ple 9.19, showing that the space cannot be finite-dimensional.

• The space P is infinite-dimensional. We found a basis for this space in Example 9.38.

• The subspace of Func

 R,R2 consisting of the continuous functions is infinite-dimensional. For exam-
3
ple, the functions 1, x, x , x , . . . form an infinite, linearly independent set of continuous functions.

• The subspace
 of Func R,R consisting of the differentiable functions is infinite-dimensional. Again,
the set 1, x, x , x3 , . . . is an infinite linearly independent set in this space.
2
352 Vector spaces

Example 9.44: Space of sequences satisfying a linear recurrence

In Example 9.29, we considered the space W of sequences of real numbers that satisfy the recurrence
an+2 = an + an+1 . What is the dimension of this space?

Solution. The space is 2-dimensional. The easiest way to see this is to observe that a sequence a ∈ W is
determined by its first two elements. We can say that the first two elements of the sequence are parameters,
and all the other elements are then computed by the recurrence relation. Specifically, suppose a0 = x and
a1 = y. Using the recurrence relation to compute the remaining elements, we have

a = (x, y, x + y, x + 2y, 2x + 3y, 3x + 5y, . . .)

= x(1, 0, 1, 1, 2, 3, . . .) + y(0, 1, 1, 2, 3, 5, . . .).

Since this is the general form of the elements of W , and since the two sequences starting with 1,0 and 0,1
are clearly linearly independent, it follows that

{(1, 0, 1, 1, 2, 3, . . .), (0, 1, 1, 2, 3, 5, . . .)}

is a basis of W . ♠

Example 9.45: Solution space of a linear differential equation

In Example 9.30, we considered the space of solutions of the differential equation f ′′ = − f . What
is the dimension of this space?

Solution. From calculus, we know that the general solution of the differential equation f ′′ = − f is

f (x) = A sin x + B cos x,

where A, B are constants. We also know, from Example 9.21, that sin x and cos x are linearly independent.
It follows that {sin x, cos x} is a basis for the solution space. The solution space is therefore 2-dimensional.
♠
We conclude this section by stating two properties of bases that generalize Theorem 5.19 and Lemma 5.44:
every linearly independent set can be extended to a basis by adding 0 or more vectors, and every spanning
set can be reduced to a basis by removing 0 or more vectors.

Proposition 9.46: Extending a linearly independent set to a basis

Let V be a vector space, and let S ⊆ V be a linearly independent set of vectors. Then S can be
extended to a basis of V , i.e., there exists a basis B of V such that S ⊆ B.

Proposition 9.47: Shrinking a spanning set to a basis

Let V be a vector space, and let S ⊆ V be a spanning set of V . Then S can be shrunk to a basis, i.e.,
there exists a basis B of V such that B ⊆ S.
 9.4. Basis and dimension 353

Example 9.48: Extending a linearly independent set to a basis

Let S ⊆ M22 be the linearly independent set given by
   
1 0 0 1
S= ,
0 0 0 0

Enlarge S to a basis of M22 .

Solution. We can obtain a basis of M22 by adding two more linearly independent matrices
   
0 0 0 0
and .
1 0 0 1

The resulting basis is        

1 0 0 1 0 0 0 0
B= , , , .
0 0 0 0 1 0 0 1
♠

Example 9.49: Shrinking a spanning set to a basis

Consider the spanning set S ⊆ P2 given by

S = 1, x, 2x + 1, x2 + 1, x2 + 2

Shrink S to a basis of P2 .

Solution. We use a version of the casting-out method. We examine each element S from left to right and
cast out the elements that are linear combinations of previous elements. Clearly the first two elements, 1
and x, are linearly independent. The next element, 2x + 1, is redundant because it is a linear combination
of 1 and x. The next element x2 +1 is linearly independent of 1 and x. The final element x2 +2 is redundant
because it is a linear combination of 1 and x2 + 1. Therefore, the following subset of S is a basis of P2 :

B = 1, x, x2 + 1 .

Exercises

Exercise 9.4.1 Let P3 be the vector space of polynomials of degree at most 3. Determine which of the
following are bases for this vector space.
 3
(a) x + 1, x2 + x, 2x3 + x2 , 2x3 − x2 − 3x + 1 .
354 Vector spaces


(b) x + 1, x3 + x2 + 2x, x2 + x, x3 + x2 + x .

Exercise 9.4.2 Determine whether the following is a basis for P2 , the vector space of polynomials of
degree at most 2.  2
x + x + 1, 2x2 + 2x + 1, x + 1 .

Exercise 9.4.3 Find a basis for the following subspace of P2 :


W = span 1 + x + x2 , 1 + 2x, 1 + 5x − 3x2 .

Exercise 9.4.4 Find a basis for the following subspace of P3 :


W = span 1 + x − x2 + x3 , 1 + 2x + 3x3 , − 1 + 3x + 5x2 + 7x3 , 1 + 6x + 4x2 + 11x3 .

Exercise 9.4.5 Extend the following linearly independent set of polynomials to a basis of P3 :
 3
x + x2 − x − 1, 3x3 + 2x2 + 2x − 1 .

Exercise 9.4.6 Let V be a 5-dimensional vector space. If you have 5 linearly independent vectors in V ,
can you conclude that the vectors span V ?

Exercise 9.4.7 Let V be a 5-dimensional vector space. If you have 6 vectors in V , is it possible that they
are linearly independent? Explain.

Exercise 9.4.8 Find a basis for the vector space of symmetric 3 × 3-matrices, i.e., matrices satisfying
A = AT . What is the dimension of this space?

Exercise 9.4.9 Let W be the subspace of P3 (over the field R) consisting of all polynomials p(x) that
satisfy p(3) = 0. Find a basis for W . What is the dimension of W ?
n  h i h i o

Exercise 9.4.10 Find a basis for U = A ∈ M2,2  A 11 −10 = 10 −11 A . What is the dimension of
U?

Exercise 9.4.11

(a) Let k be a positive integer, let Wk ⊆ SeqK be the subspace consisting of all sequences that are
periodic with period k (see Exercise 9.3.15). Find a basis for Wk . What is its dimension?
(b) More difficult: find a basis for the infinite-dimensional vector space consisting of all periodic se-
quences of all periods.

√
Exercise 9.4.12 Let K = Q, the field of rational numbers. Consider vectors of the form a + b 2 where
a, b are rational numbers. Show that this collection of vectors is a vector space over Q and give a basis
for this vector space. What is its dimension?
 9.5. Application: Error correcting codes 355

9.5 Application: Error correcting codes

Outcomes
A. Determine the block length, message length, Hamming distance, and rate of a code.

B. Determine whether a code is m-error detecting and m-error correcting.

C. Find generator and check matrices for a linear code.

D. Use the syndrome method to correct errors in code blocks.

E. Construct and use Hamming codes.

Binary codes

When transmitting or storing information on digital media, the information is usually encoded as a se-
quence of bits, i.e., 0s and 1s. For example, in the ASCII code, each symbol is encoded as a sequence of
8 bits. The letter “A” is encoded as 01000001, the letter “B” is encoded as 01000010, and so on. We can
think of a sequence of n bits as an n-dimensional column vector over the field Z2 , i.e., as an element of
Zn2 . For our purposes, it is convenient to continue writing bit sequences horizontally, but we will consider
this to be merely an alternate notation for a column vector.
One issue with digital data is that the data can sometimes be corrupted. DVDs may be scratched,
magnetic storage may depolarize, and data sent by radio transmission may be subject to interference. This
can result in errors in the data, such as some 0s being changed to 1s or vice versa. One of the ways to deal
with such errors is to introduce redundancy in the way the data is encoded.
A very simple example of redundant encoding is the so-called 3-repetition code. This simply means
to repeat each bit 3 times. Thus, the bit 0 is encoded as 000 and the bit 1 as 111. For example, the bit
string 100101 is encoded as 111000000111000111. With this encoding, single bit errors are easy to detect
and correct. Assuming that at most one error occurs within each 3-bit block, the blocks can be decoded by
“majority decision”. Namely, if the bits in a block are not the same, we assume that the error occurred in
the bit that is in the minority. The following table shows the decoding scheme:

Received Likely error Corrected Decoded

000 none 000 0
100 bit 1 000 0
010 bit 2 000 0
001 bit 3 000 0
011 bit 1 111 1
101 bit 2 111 1
110 bit 3 111 1
111 none 111 1
356 Vector spaces

Example 9.50: Decoding the 3-repetition code

A message was encoded with the 3-repetition code. The receiver receives the following data:
001011000100111110. What was the likely message?

Solution. We start by dividing the received data into blocks of 3 bits:

001 011 000 100 111 110.
We then decode each block separately, using the majority rule. The first block 001 has a majority of zeros,
so the likely error is in the third bit and the likely decoding is 0. The second block 011 has a majority of
ones, so the likely error is in the first bit and the likely decoding is 1. Continuing this way, we find that the
likely original message was 01 00 11. ♠
Since the 3-repetition code can be used to correct some errors, it is called an error correcting code. This
particular code only works if there are not too many errors; namely, it can correct at most one error per
code block. The main drawback of the 3-repetition code is that it increases the message length by a factor
of 3. In this section, we will use linear algebra over the field Z2 to construct better error correcting codes.
Most error correcting codes do not work by encoding individual bits. Rather, the codes work by
dividing the message into message blocks of length k, and encoding each such message block by a code
block of length n. This leads us to the following definition:

Definition 9.51: Code

Let n and k be positive integers with k ≤ n. A code with message length k and block length n is
a set C of 2k different vectors in Zn2 . We call the elements of C the code blocks. Since there are 2k
different code blocks, they can be used to encode message blocks of length k. We say that the rate
of the code is nk , because for every n bits of encoded data transmitted, k bits of decoded data are
obtained.

Example 9.52: A simple code

Consider the following code with message length k = 2 and block length n = 5:
Message block Code block
00 00000
01 00111
10 11100
11 11011

(a) Encode the message 011110.

(b) Decode the message 111000011111011.

(c) What is the rate of this code? Is it better or worse than the rate of the 3-repetition code?

Solution. (a) We divide the message into blocks of length 2: 01 11 10. Then we encode each block
separately: 00111 11011 11100. (b) We divide the message into blocks of length 5: 11100 00111 11011.
 9.5. Application: Error correcting codes 357

Then we decode each block separately: 10 01 11. (c) The rate is 25 = 0.4. It is slightly higher, and therefore
better, than the rate of the 3-repetition code, which is 13 ≈ 0.33. ♠

Hamming distance and error correction

The error correction capabilities of a code depend on a property called the Hamming distance of the code,
which we now define.

Definition 9.53: Hamming weight and Hamming distance

• Let v be a vector in Zn2 . The Hamming weight of v, denoted W (v), is the number of compo-
nents of v that are equal to 1.

• Let v, w be two vectors in Zn2 . The Hamming distance between v and w, denoted D(v, w), is
the number of components where v and w differ. We can also express this as the Hamming
weight of v − w, i.e., D(v, w) = W (v − w).

• Finally, we say that the Hamming distance of a code is equal to the smallest Hamming
distance between any two code blocks.

A code with block length n, message length k, and Hamming distance d is also called an (n, k, d)-
code.

Example 9.54: Hamming distance

Calculate the Hamming distance between 00111 and 11100. What is the Hamming distance of the
3-repetition code? What is the Hamming distance of the code in Example 9.52?

Solution. The vectors 00111 and 11100 differ in 4 places, so their Hamming distance is D(00111, 11100) =
4. This is also equal to the Hamming weight of 00111 − 11100 = 11011.
The 3-repetition code has only two code blocks: 000 and 111. Since their Hamming distance is 3, the
Hamming distance of the code is also 3.
To calculate the Hamming distance of the code from Example 9.52, we calculate the Hamming distance
between all pairs of code blocks:
D(00000, 00111) = 3,
D(00000, 11100) = 3,
D(00000, 11011) = 4,
D(00111, 11100) = 4,
D(00111, 11011) = 3,
D(11100, 11011) = 3.
Since the smallest distance between any two code blocks is 3, the Hamming distance of the code is 3. ♠
The significance of a code’s Hamming distance is explained by the following definition and proposition.
358 Vector spaces

Definition 9.55: Error detection and error correction

• We say that a code C is m-error detecting if for all valid code blocks v, whenever w is
obtained from v by introducing up to m bit errors, then w is not a valid code block.

• We say that a code C is m-error correcting if for all valid code blocks v, whenever w is
obtained from v by introducing up to m bit errors, then v is the only valid code block within
Hamming distance m of w.

Proposition 9.56: Error detection and error correction

Consider a code with Hamming distance d . Then the code is:

• m-error detecting if m ≤ d − 1;

• m-error correcting if 2m ≤ d − 1.

Proof. Assume up to m errors have happened. In other words, let v be a valid code block, and let w be
the code block obtained from v by introducing up to m errors. To prove the first claim, assume m ≤ d − 1.
Then D(v, w) ≤ m ≤ d − 1. Since d is the minimum distance between any two valid code blocks, w cannot
be a valid code block. Hence, the errors can be detected. To prove the second claim, assume 2m ≤ d − 1.
Then D(v, w) ≤ m. Assume that there exists another valid code block u within Hamming distance m of w,
i.e., assume D(w, u) ≤ m. Then

D(v, u) ≤ D(v, w) + D(w, u) ≤ m + m ≤ d − 1.

Therefore, the Hamming distance of v and u is at most d − 1, contradicting the assumption that the code
has Hamming distance d. Hence, there is no such code block u, and the errors can be corrected. ♠

Example 9.57: Error detection and error correction

Consider a code with Hamming distance 3. How many errors per code block can the code detect?
How many can it correct? Also answer this question for Hamming distance 2, 4 and 5.

Solution. By Proposition 9.56, a code with Hamming distance 3 can detect up to 2 errors and correct up
to 1 error. The answers for other Hamming distances are summarized in the following table:

Hamming distance Errors detected Errors corrected

2 1 0
3 2 1
4 3 1
5 4 2

♠
 9.5. Application: Error correcting codes 359

Example 9.58: Decoding

The following message has been encoded using the code of Example 9.52. It contains some errors,
but no more than 1 error per code block. Can you decode the message?

11101 01000 11011 11111 00011

Solution. Since the code of Example 9.52 has Hamming distance 3, it can correct up to 1 error per code
block, so we are able to decode the message uniquely. For each code block, we must find the unique valid
code block that is within Hamming distance 1 or less.
Received Error Corrected Decoded
11101 bit 5 11100 10
01000 bit 2 00000 00
11011 none 11011 11
11111 bit 3 11011 11
00011 bit 3 00111 01
The decoded message is 10 00 11 11 01. ♠
We can say that a “good” error correcting code is one that has a high rate (of message length divided by
block length) and a large Hamming distance.

Linear codes

To construct a code of message length n, we need to specify a set of 2n code blocks. If n is large, it is
not really feasible to write down a list of all the code blocks, and to check all their Hamming distances by
hand. For example, a code of block length n = 10 requires 210 = 1024 code blocks, and we need to check
more than half a million Hamming distances. Instead, we will focus on a particular class of codes that is
much easier to describe. These are the linear codes.

Definition 9.59: Linear code

A linear code is a subspace of Zn2 . If the subspace is k-dimensional, then the code has message
length k and block length n.

The advantage of a linear code is that to specify the code blocks, we only need to list k basis elements,
rather than all 2k elements of the code.

Example 9.60: Linear code

Show that the code from Example 9.52 is linear. What is a basis for the code?

Solution. Consider all linear combinations of 11100 and 00111 (with scalars in Z2 ):
0(11100) + 0(00111) = 00000,
0(11100) + 1(00111) = 00111,
1(11100) + 0(00111) = 11100,
1(11100) + 1(00111) = 11011.
360 Vector spaces

This shows that the code of Example 9.52 is span {11100, 00111}, and hence a subspace of Z52 . A basis
for the code is {11100, 00111} ♠
From Section 5.5, we know that every subspace of Zn2 , and therefore every linear code, is the column space
of some matrix G, and also the null space of some matrix H. Such matrices are called a generator matrix
and a check matrix for the code, respectively.

Definition 9.61: Generator matrix and check matrix

Consider a linear code C with block length n and message length k, i.e., a k-dimensional subspace
of Zn2 .

• A generator matrix for the code is an n × k-matrix G such that C is the column space of G.

• A check matrix for the code is an (n − k) × n-matrix H such that C is the null space of H .

Example 9.62: Generator matrix and check matrix

Find a generator matrix and check matrix for the linear code from Example 9.52.

Solution. We already found in Example 9.60 that {11100, 00111} is a basis for this code. We can obtain
a generator matrix by using the basis vectors as columns:
 
1 0
 1 0 
 
G=  1 1 .

 0 1 
0 1

To find a check matrix, assume that [a, b, c, d, e] is a row of the check matrix. Then for every code block
v ∈ C, we must have [a, b, c, d, e] v = 0. Since the code blocks are spanned by 11100 and 00111, it suffices
to consider the two equations
 
1
 1 
  
a b c d e  
 1  = 1a + 1b + 1c + 0d + 0e = 0
 0 
0
and  
0
 0 
  
a b c d e 
 1  = 0a + 0b + 1c + 1d + 1e = 0.

 1 
1
Solving this system of equations, we find that the following is a basis for the solution space:
     
1 1 0 0 0 , 1 0 1 1 0 , 1 0 1 0 1
 9.5. Application: Error correcting codes 361

We can use these basic solutions as the rows of the check matrix:
 
1 1 0 0 0
H= 1 0 1 1 0 
1 0 1 0 1 ♠

Encoding and the generator matrix

Let C be a linear code with generator matrix G. Since C is the column space of the generator matrix, there
is a simple encoding method: we can simply multiply the generator matrix G by a message block to obtain
the corresponding code block. In other words, if u ∈ Zk2 is a message block, we can use v = Gu ∈ Zn2 as
the code block.

Example 9.63: Using the generator matrix for encoding

Use the generator matrix  
1 0
 1 0 
 
G=
 1 1 

 0 1 
0 1
to encode the message 01 11 10.

h i h i h i
Solution. The message blocks are u1 = 01 , u2 = 11 , and u3 = 10 . We obtain the corresponding
code blocks by multiplication with the generator matrix:
     
0 1 1
 0   1   1 
     
v1 = Gu1 =      
 1  v2 = Gu2 =  0  v3 = Gu3 =  1 
 1   1   0 
1 1 0
Therefore the encoded message is 00111 11011 11100. Note that this is the same answer as in Exam-
ple 9.52(a). ♠

Decoding and the check matrix

Let C be a linear code with check matrix H. By definition, C is the null space of the check matrix. This
means that a vector v ∈ Zn2 is a valid code block (i.e., a code block without errors) if and only if Hv = 0.
So we can easily use the check matrix to determine whether a code block contains errors or not.
However, something even better is true. The value of Hv, when it is not zero, tell us not only that
an error has occurred, but also which error has occurred! To see why, consider a code block v′ possibly
containing errors. Then v′ = v + e, where v is a valid code block and e is an error pattern. The error
pattern is the vector e that has a 1 in every component in which an error occurred, and a 0 everywhere else.
Then we have
Hv′ = H(v + e) = Hv + He = 0 + He = He.
362 Vector spaces

Therefore, the value of Hv′ only depends on the error pattern, and not on v. The value s = Hv’ = He
is called the syndrome of the error. A code is error correcting, for a class of error patterns, if and only
if each such error pattern has a different syndrome. In that case, we can simply make a table of all the
syndromes and corresponding error patterns, as an efficient method for correcting errors. Such a table is
called a syndrome table and the corresponding decoding method is called syndrome decoding.

Example 9.64: Syndrome decoding

Consider the linear code of Example 9.52 with check matrix
 
1 1 0 0 0
H =  1 0 1 1 0 .
1 0 1 0 1

Make a syndrome table for all single-bit errors. Then use your syndrome table to decode the mes-
sage 11101 01000 11011 11111 00011.

Solution. Since we are only interested in single-bit errors, there are six error patterns to consider: 00000
(no error), 10000, 01000, 00100, 00010, and 00001. For each of these error patterns e, we compute the
corresponding syndrome He. For example,
     
0   1   0  
 0  0  0  1  1  1
     
H 0  =  0 , H  0  =  1 , H  0  =  0 ,
   
 0  0  0  1  0  0
0 0 0

and so on. We put this information in a table (writing the vectors horizontally as usual in this section).
This is the syndrome table:
Error pattern e Syndrome He
00000 000
10000 111
01000 100
00100 011
00010 010
00001 001
To decode the message 11101 01000 11011 11111 00011, we first calculate the syndrome of each code
block by multiplying it by the check matrix. The syndrome table then tells us the corresponding error
pattern, which we can use to correct the code block.
For example, consider the first code block 11101. Multiplying by H, we get the syndrome 001:
 
1  
 1  0
 
H 1 = 0 
 0  1
1
 9.5. Application: Error correcting codes 363

The syndrome table then tells us that the corresponding error pattern is 00001. Therefore, the corrected
code block is 11101 + 00001 = 11100. The decoded block is 10. We proceed in the same way for all the
code blocks:
Received code block v′ Syndrome Hv′ Error pattern e Corrected code block v = v′ + e Decoded
11101 001 00001 11100 10
01000 100 01000 00000 00
11011 000 00000 11011 11
11111 011 00100 11011 11
00011 011 00100 00111 01

Therefore, the decoded message is 10 00 11 11 01. Note that this is the same answer we got in Exam-
ple 9.58. But the syndrome table gives a more systematic method of finding the error patterns, which we
previously had to do by guessing, or by comparing to all possible code blocks. ♠

Example 9.65: An invalid code block

Using the code of the previous example, suppose you have received the code block 10010. Can you
decode it?

Solution. The syndrome of this code block is

 
1  
 0  1
 
H
 0  =  0 .

 1  1
0

Since the syndrome 101 is not in our syndrome table, it is not the syndrome of any single-bit error. There-
fore, the code block 10010 must contain more than one error. Since our code is only 1-error correcting,
this error cannot be corrected. (In fact, there are two valid code words within Hamming distance 2, namely
00000 and 11011). ♠

Hamming codes

The syndrome decoding method immediately gives us the following theorem:

Theorem 9.66: Check matrices of 1-error correcting codes

Consider a linear code with check matrix H . Then the code is 1-error correcting if and only if all
columns of the check matrix are non-zero and distinct.

Proof. The syndrome of the zero error pattern is always H0 = 0. Let ei be the error pattern containing a
single bit error in the i th component. Then its syndrome is Hei , which is the i th column of H.
If H has distinct, non-zero columns, then each possible single bit error has a different syndrome.
Therefore, the errors can be corrected.
364 Vector spaces

Conversely, if H has a column that is zero, or two columns that are equal, then the corresponding error
patterns have the same syndrome, and therefore cannot be corrected. ♠
This theorem was discovered by Richard Hamming in 1950. Hamming then realized that the best single-
error correcting binary codes can be constructed by letting the check matrix have all possible non-zero
columns. The resulting codes are called Hamming codes.

Definition 9.67: Hamming code

Let r ≥ 2 The r th Hamming code is a linear code whose check matrix H is an r × (2r − 1)-matrix
that has all possible non-zero column vectors as its columns.

From the size of the check matrix, we know that the code length is n = 2r − 1. Since the check matrix
has rank r, its null space has dimension k = n − r. Thus, the r th Hamming code is an (n, k, 3)-code, where
n = 2r − 1 and k = 2r − 1 − r.
It is customary to order the columns of the check matrix so that the last r columns are the standard
basis vectors. In other words, the check matrix is usually taken to be of the form
 
H= A I

where I is the r × r-identity matrix and A is the r × k-matrix consisting of the remaining columns of H. In
that case, we can take the n × k-matrix  
I
G=
A
as the generator matrix, where I is the k × k-identity matrix. Specifically, for this choice of G, we have
HG = 0, ensuring that the column space of G is contained in the null space of H. Moreover, since
rank G = k, the column space of G is k-dimensional, and therefore equal to the null space of H, so that G
is indeed a correct generator matrix for the code.

Example 9.68: Hamming code for r = 3

Construct a check matrix and generator matrix for the Hamming code with r = 3. Then encode the
message 0101 1110 0111.

Solution. The check matrix must be a 3 × 7-matrix whose columns are all the possible non-zero vectors
of length 3. Moreover, we will follow the convention of using the standard basis vectors as the last 3
columns. Such a matrix is  
1 1 0 1 1 0 0
H =  1 0 1 1 0 1 0 .
0 1 1 1 0 0 1
 9.5. Application: Error correcting codes 365

The corresponding generator matrix G is

 
1 0 0 0
 0 1 0 0 
 
 0 0 1 0 
 
G=
 0 0 0 1 .

 1 1 0 1 
 
 1 0 1 1 
0 1 1 1

To encode the message 0101 1110 0111, we multiply the generator matrix by each code block:
     
0 1 0
   1     1     1 
0   1   0  
 0   1   
 1     1     1   1 
G           
 0  =  1 , G 1  =  0 , G 1  =  1 .
 0   0   0 
1   0   1  
 1   0   0 
0 0 1

Thus, the encoded message is 0101010 1110000 0111001. Note that the fact that the top part of the
generator matrix G is the identity matrix has the pleasant effect that the k first bits of each code block are
the corresponding message block. This makes decoding especially convenient (after any errors have been
corrected first). ♠

Example 9.69: Decoding a Hamming code

Using the Hamming (7, 4, 3)-code of the previous example, correct the errors in the message
1011101 1101001 0001101 1110000 and decode it.

Solution. We do not need to make a syndrome table, because the syndromes are exactly the columns of the
check matrix H. More precisely, the i th column of the check matrix is the syndrome of the error pattern
containing a single-bit error in the i th bit. We use the check matrix
 
1 1 0 1 1 0 0
H = 1 0 1 1 0 1 0 
0 1 1 1 0 0 1

to calculate the syndrome of each code block:

       
1 1 0 1
 0   1   0   1 
               
 1  1  0  1  0  0  1  0
       
H  1  =  1 , H 
  1
 
 =  0 , H
 1  =  1 ,
 H
 0  =  0 .

 1  1  0  1  1  0  0  0
       
 0   0   0   0 
1 1 1 0
366 Vector spaces

From these syndromes, we can read off the error locations and correct the errors:

Received code block Syndrome Error position Error pattern Corrected code block Decoded
1011101 111 bit 4 0001000 1010101 1010
1101001 101 bit 2 0100000 1001001 1001
0001101 010 bit 6 0000010 0001111 0001
1110000 000 none 0000000 1110000 1110

Thus, the decoded message is 1010 1001 0001 1110. ♠

Example 9.70: Hamming code for r = 2

What is the Hamming code for r = 2? Have you seen this code before?

Solution. The check matrix and generator matrix for the Hamming code with r = 2 are
 
  1
1 1 0
H= 
and G = 1  .
1 0 1
1

The code has block size n = 3 and message length k = 1. The encoding function is
   
  0   1
G 0 = 0 , G 1 = 1 .
  
0 1

This is exactly the 3-repetition code. ♠

We end this section by remarking that the Hamming codes are very efficient. The rate of the r th Hamming
code is
k 2r − 1 − r r
= = 1 − ,
n 2r − 1 2r − 1
which is very close to 1 when r is large. The following table lists the block sizes, message sizes, and rates
of all Hamming codes up to r = 8.

r Block size n Message size k Rate

2 3 1 0.333
3 7 4 0.571
4 15 11 0.733
5 31 26 0.839
6 63 57 0.905
7 127 120 0.945
8 255 247 0.969

Of course, since the codes are only 1-error correcting, one cannot increase the block size indefinitely,
or else the probability of having two or more errors in a block becomes too large. There exist more
sophisticated error correcting codes with larger Hamming distances, which can correct many errors per
code block. You might learn more about such codes in a course on applied modern algebra.
 9.5. Application: Error correcting codes 367

Exercises

Exercise 9.5.1 Encode the message 0110 using the 3-repetition code. Decode the message 100 011 000
110 101 110 after correcting single-bit errors.

Exercise 9.5.2 Use the code of Example 9.52 to encode the message 01 11 10 00 01. Use the syndrome
method to correct all single-bit errors in the message 10100 00001 11001. What is the decoded message?

Exercise 9.5.3 Consider the code

C = {000000, 011100, 111010, 101001, 100110, 110101, 010011, 001111}.

Is this a linear code? What are the message length and block length of this code? What is its Hamming
distance? How many errors per code block can this code detect? How many can it correct?

Exercise 9.5.4 Consider a linear code with check matrix

 
1 1 0 0 0
H =  1 0 1 1 0 .
1 0 1 0 1

Find a generator matrix for this code (hint: the columns of G should form a basis for the null space of H).
List all possible code blocks. What is the Hamming distance of the code? Make a syndrome table for this
code.

Exercise 9.5.5 Use the generator matrix

 
1 0 0
 1 1 1 
 
 1 0 1 
G=



 0 1 0 
 1 1 0 
0 0 1

to encode the message 111 110 101 001. Find a check matrix for this code and make a syndrome ta-
ble. Is the code 1-error correcting? Use the syndrome method to correct and decode the message
111000 101111 100111 110100.

Exercise 9.5.6 Construct check and generator matrices of a Hamming code for r = 4. What is the block
length and message length of this code? Encode the message 00110011000 10100000001. Decode the
message 001000010010000 010000001000001 after correcting single-bit errors.
10. Linear transformation of vector spaces

10.1 Definition and examples

Outcomes
A. Understand the definition of a linear transformation in the context of vector spaces.

B. Determine whether a function is linear or not.

In Chapter 6, we defined a linear transformation T : Rn → Rm to be a function that preserves addition and

scalar multiplication. We now revisit this concept in the more general setting of vector spaces V and W .

Definition 10.1: Linear transformation

Let V and W be vector spaces over some field K . A function T : V → W is called a linear transfor-
mation from V to W if it satisfies the following two conditions:

1. T preserves addition, i.e., for all v, w ∈ V , we have T (v + w) = T (v) + T (w);

2. T preserves scalar multiplication, i.e, for all v ∈ V and k ∈ K , we have T (kv) = kT (v).

A linear transformation is also sometimes called a linear function or a linear map. A linear
transformation T : V → V (i.e., when V = W ) is also sometimes called an operator.

Our first example of a linear transformation is a matrix transformation. We have already seen this in
Section 6.2.

Example 10.2: Matrix transformation

Let A be an m × n-matrix. Then the function T : Rm → Rn defined by T (v) = Av is a linear trans-
formation, called a matrix transformation. This was proved in Proposition 6.4.

There are many interesting examples of linear transformations on vector spaces other than Rn . We will
consider a few such examples.

369
370 Linear transformation of vector spaces

Example 10.3: Derivative operator

Let Pn be the vector space of real polynomials of degree at most n. The function D : Pn → Pn−1 is
defined by
D(p(x)) = p′ (x),
where p′ (x) denotes the derivative of the polynomial p(x). The function D is called the derivative
operator.

(a) Compute D(x3 ), D(2x2 + x), and D(ax3 + bx2 + cx + d).

(b) Show that D is a linear transformation.

Solution.

(a) In each case, we simply take the derivative:

D(x3 ) = 3x2 ,
D(2x2 + x) = 4x + 1,
D(ax3 + bx2 + cx + d) = 3ax2 + 2bx + c.

(b) First, we note that if p(x) is a polynomial of degree at most n, then its derivative p′ (x) is a polynomial
of degree at most n − 1. Therefore, the derivative operator D is a well-defined function from Pn to
Pn−1 . To show that it preserves addition, consider any two polynomials p(x), q(x) ∈ Pn . From
calculus, we know that the derivative of p(x) + q(x) is p′ (x) + q′ (x). Therefore,
D(p(x) + q(x)) = (p(x) + q(x))′ = p′ (x) + q′ (x) = D(p(x)) + D(q(x)),
and so D preserves addition. To show that it preserves scalar multiplication, consider p(x) ∈ Pn and
k ∈ R. From calculus, we know that the derivative of kp(x) is kp′ (x), and therefore
D(kp(x)) = (kp(x))′ = kp′ (x) = kD(p(x)).
Hence, D preserves scalar multiplication. It follows that D is a linear transformation.

♠
It is important to understand that we are not claiming that the derivative p′ (x) of a polynomial p(x) is a
linear function. It is of course a polynomial. Rather, what the above example shows is that the act of
taking the derivative is a linear operation, i.e., the derivative of a sum is the sum of the derivatives, and the
derivative of a constant times a function is a constant times the derivative.

Example 10.4: A differential equation

Solve the equation p(x) = x3 + D(p(x)), where D : P3 → P2 is the derivative operator of Exam-
ple 10.3.

Solution. Every element of P3 is of the form p(x) = ax3 + bx2 + cx + d. We can write the equation
p(x) = x3 + D(p(x)) as
(ax3 + bx2 + cx + d) = x3 + (3ax2 + 2bx + c).
 10.1. Definition and examples 371

For the left-hand side and right-hand side to be equal, we must have a = 1, b = 3a, c = 2b, and d = c. This
yields the unique solution (a, b, c, d) = (1, 3, 6, 6), or p(x) = x3 + 3x2 + 6x + 6. ♠

Example 10.5: The shift and unshift operators

Consider the vector space SeqK of infinite sequences of elements of K . The function shift : SeqK →
SeqK is defined by shifting the entire sequence to the left and dropping the first element:

shift(a0 , a1 , a2 , a3 , . . .) = (a1 , a2 , a3 , a4 , . . .).

The function unshift : SeqK → SeqK is defined by shifting the entire sequence to the right and
adding 0 as the new first element:

unshift(a0 , a1 , a2 , a3 , . . .) = (0, a0, a1 , a2 , . . .).

(a) Compute shift(1, 2, 3, . . .), unshift(shift(1, 1, 1, . . .)), and shift(unshift(1, 1, 1, . . .)).

(b) Show that shift and unshift are linear transformations.

Solution.

(a) We have
shift(1, 2, 3, 4, . . .) = (2, 3, 4, 5, . . .),
unshift(shift(1, 1, 1, 1, . . .)) = unshift(1, 1, 1, 1, . . .) = (0, 1, 1, 1, . . .),
shift(unshift(1, 1, 1, 1, . . .)) = shift(0, 1, 1, 1, . . .) = (1, 1, 1, 1, . . .).

(b) To show that shift is a linear transformation, we show that it preserves addition and scalar multipli-
cation. Let a = (a0 , a1 , a2 , . . .) and b = (b0 , b1 , b2 , . . .). Then
shift(a + b) = shift(a0 + b0 , a1 + b1 , a2 + b2 , . . .)
= (a1 + b1 , a2 + b2 , a3 + b3 , . . .)
= (a1 , a2 , a3 , . . .) + (b1 , b2 , b3 , . . .)
= shift(a) + shift(b),
shift(ka) = shift(ka0 , ka1 , ka2 , . . .)
= (ka1 , ka2 , ka3 , . . .)
= k(a1 , a2 , a3 , . . .)
= k shift(a).
Therefore, shift is a linear transformation. The proof for unshift is similar.

Example 10.6: Properties of shift and unshift

Show that for all sequences a ∈ SeqK , we have shift(unshift(a)) = a. On the other hand, show that
in general, unshift(shift(a)) 6= a.
372 Linear transformation of vector spaces

Solution. For a = (a0 , a1 , a2 , . . .), we have

shift(unshift(a)) = shift(unshift(a0 , a1 , a2 , . . .)) = shift(0, a0, a1 , . . .) = (a0 , a1 , a2 , . . .) = a.

On the other hand, we have

unshift(shift(a)) = unshift(shift(a0 , a1 , a2 , . . .)) = unshift(a1 , a2 , a3 , . . .) = (0, a1 , a2 , . . .).

The latter is not equal to a unless a0 = 0. ♠

Example 10.7: Recurrence as a linear equation

Find the general solution to the following equation, where a ∈ SeqK :

shift(shift(a)) = shift(a) + a.

Where have you seen this equation before?

Solution. For a = (a0 , a1 , a2 , . . .), we have

shift(shift(a)) = (a2 , a3 , a4 , . . .),

shift(a) = (a1 , a2 , a3 , . . .),
a = (a0 , a1 , a2 , . . .).

Therefore a is a solution of the equation shift(shift(a)) = shift(a) + a if and only if

a2 = a1 + a0 ,
a3 = a2 + a1 ,
a4 = a3 + a2 ,

and so on. In other words, a is a solution if and only if an+2 = an+1 + an holds for all n ≥ 0. This is nothing
but the recurrence relation of Example 9.29. We already calculated the general solution in Example 9.44.
The general solution is
a = (x, y, x + y, x + 2y, 2x + 3y, 3x + 5y, . . .),
and a basis for the solution space is

{(1, 0, 1, 1, 2, 3, . . .), (0, 1, 1, 2, 3, 5, . . .)} .

♠
We conclude this section by stating some elementary properties of linear transformations. “Elementary”
means that these properties follow directly from the definition, i.e., from the fact that linear transformations
preserve addition and scalar multiplication.
 10.1. Definition and examples 373

Proposition 10.8: Properties of linear transformations

Let V and W be vector spaces over a field K , and let T : V → W be a linear transformation. Then

• T preserves the zero vector: T (0) = 0.

• T preserves additive inverses: T (−v) = −T (v).

• T preserves linear combinations:

T (a1 v1 + a2 v2 + . . . + ak vk ) = a1 T (v1 ) + a2 T (v2 ) + . . . + ak T (vk ).

Proof. To prove the first property, let k = 0 in the equation T (kv) = kT (v). Since 0v = 0 and 0T (v) = 0
by Proposition 9.9, we therefore have T (0) = 0. Similarly, to prove the second property, let k = −1 in the
equation T (kv) = kT (v). Finally, the third property is a direct consequence of the fact that T preserves
addition and scalar multiplication:

T (a1 v1 + a2 v2 + . . . + ak vk ) = T (a1 v1 ) + T (a2 v2 ) + . . . + T (ak vk )

= a1 T (v1 ) + a2 T (v2 ) + . . . + ak T (vk ).

Exercises

Exercise 10.1.1 Consider the following functions T : R3 → R2 . For each of these functions T , show that
it is not linear either by showing that T does not preserve addition, or by showing that it does not preserve
scalar multiplication, or by showing that it does not preserve the zero vector.
 
x  
x + 2y + 3z + 1
(a) T  y  = .
2y − 3x + z
z
 
x  
  x + 2y2 + 3z
(b) T y = .
2y + 3x + z
z
 
x  
sin x + 2y + 3z
(c) T  y  = .
2y + 3x + z
z

Exercise 10.1.2 Consider the following functions T : R3 → R2 . For each function T , show that T is a
linear transformation. Do this by showing that T is a matrix transformation, i.e., find a matrix A such that
T (v) = Av.
374 Linear transformation of vector spaces

 
x  
x + 2y + 3z
(a) T y = .
2y − 3x + z
z
 
x  
  7x + 2y + z
(b) T y = .
3x − 11y + 2z
z
 
x  
  3x + 2y + z
(c) T y = .
x + 2y + 6z
z
 
x  
2y − 5x + z
(d) T y = .
x+y+z
z

Exercise 10.1.3 Which of the following functions are linear transformations?

(a) The function f : SeqK → SeqK defined by f ((a0 , a1 , a2 , . . .)) = (−a1 , −a3 , −a5 , . . .). In words, the
function f removes the even-numbered elements, and negates the odd-numbered elements.
(b) The function f : M2,2 → M2,2 defined by f (A) = AB. Here, M2,2 is the vector space of 2 ×2-matrices
with real entries, and B is a fixed matrix.
(c) The function f : M2,2 → M2,2 defined by f (A) = A + B, where B is a fixed matrix.
(d) The function f : M2,2 → M2,2 defined by f (A) = AT .

Exercise 10.1.4 Recall the vector space P of polynomials with coefficients in a field K. Consider the
function M : P → P defined by M(p(x)) = xp(x).
(a) Compute M(x3 ), M(2x2 + x), and M(ax2 + bx + c).
(b) Show that M is a linear transformation.

Exercise 10.1.5 Recall the vector space P of polynomials with coefficients in a field K. Consider the
function S : P → P defined by S(p(x)) = p(x + 1).
(a) Compute S(x3 ), S(2x2 + x), and S(ax2 + bx + c).
(b) Show that S is a linear transformation.

Exercise 10.1.6 Consider the shift function shift : SeqK → SeqK from Example 10.5. Find a basis for the
solution space of each of the following equations:
(a) shift(a) = a.
(b) shift(a) = −a.
(c) shift(shift(a)) = shift(a) + 2a.
 10.2. The algebra of linear transformations 375

10.2 The algebra of linear transformations

Outcomes
A. Use algebraic properties of linear transformations to manipulate expressions.

Two linear transformations are considered to be equal if they act in the same way on all vectors. This is
the content of the following definition.

Definition 10.9: Equal transformations

Let S and T be linear transformations from V to W . We say that S and T are equal, and we write
S = T , if for all v ∈ V ,
S(v) = T (v).

We now consider several operations on linear transformations. These include addition and scalar multipli-
cation of linear transformations, as well as the zero transformation.

Definition 10.10: Addition and scalar multiplication of linear transformations

Let V and W be vector spaces over a field K .

(a) The zero transformation 0 : V → W is defined by 0(v) = 0 for all v ∈ V .

(b) If T , S : V → W are linear transformations, then their sum T + S : V → W is the linear trans-
formation defined by
(T + S)(v) = T (v) + S(v)
for all v ∈ V .

(c) If T : V → W is a linear transformation and k ∈ K , the linear transformation kT : V → W is

defined by
(kT )(v) = k(T (v))
for all v ∈ V .

Proposition 10.11: Addition and scalar multiplication of linear transformations

If T , S : V → W are linear transformations, then so are T + S and kT .

Proof. To show that T + S is linear, we must verify that it preserves addition and scalar multiplication. Let
v, u ∈ V . Then we have
(T + S)(v + u) = T (v + u) + S(v + u) by definition of T + S,
= (T (v) + T (u)) + (S(v) + S(u)) by linearity of T and S,
= (T (v) + S(v)) + (T (u) + S(u)) by the associative and commutative laws of vectors,
= (T + S)(v) + (T + S)(u) by definition of T + S.
376 Linear transformation of vector spaces

Therefore, T + S preserves addition. Also, for v ∈ V and ℓ ∈ K, we have

(T + S)(ℓv) = T (ℓv) + S(ℓv) by definition of T + S

= ℓT (v) + ℓS(v) by linearity of T and S
= ℓ(T (v) + S(v)) by the distributive law over vector addition
= ℓ((T + S)(v)) by definition of T + S.

Therefore, T + S preserves scalar multiplication, and hence T + S is a linear transformation. The proof
that kT is a linear transformation is left as an exercise. ♠
There operations satisfy the following properties:

Proposition 10.12: Properties of addition and scalar multiplication of linear transformations

Let V ,W be vector spaces over a field K , let T , S, R : V → W be linear transformations, and let
k, ℓ ∈ K be scalars. Then the following hold:

• Commutative law of addition: T + S = S + T .

• Associative law of addition: (T + S) + R = T + (S + R).
• The existence of an additive unit: there exists an element 0 ∈ V such that for all T , T + 0 = T .
• The law of additive inverses: T + (−T ) = 0.
• The distributive law over vector addition: k(T + S) = kT + kS.
• The distributive law over scalar addition: (k + ℓ)T = kT + ℓT .
• The associative law for scalar multiplication: k(ℓT ) = (kℓ)T .
• The rule for multiplication by one: 1T = T .

But these 8 properties are just the vector space laws (A1)–(A4) and (SM1)–(SM4)! Therefore, the set of
linear transformations from V to W , with the above operations of addition and scalar multiplication, forms
a vector space.

Definition 10.13: Vector space of linear transformations

Let V ,W be vector spaces over a field K . We define LinV ,W to be the vector space of all linear
transformations from V to W , with the above operations of addition and scalar multiplication.

Another important operation is the composition of linear transformations. We have already encountered
this in Definition 6.16 for the case of Rn . Here, we generalize it to linear transformations of arbitrary
vector spaces. We also consider the identity transformation on a vector space, which forms the unit for
composition.
 10.2. The algebra of linear transformations 377

Definition 10.14: Composition of linear transformations

Let V ,U ,W be vector spaces over a field K , and let S : V → U and T : U → W be linear transforma-
tions.

(a) The composition T ◦ S : V → W is the linear transformation defined by

(T ◦ S)(v) = T (S(v))

for all v ∈ V .

(b) The identity transformation 1V : V → V is defined by 1V (v) = v for all v ∈ V . We often

omit the subscript and just write 1 : V → V when V is clear from the context.

Composition of linear transformations satisfies the following properties:

Proposition 10.15: Properties of composition of linear transformations

Composition of linear transformations satisfies the following properties, for all R, S, T as appropriate
for each law:

• Associative law of composition: (T ◦ S) ◦ R = T ◦ (S ◦ R).

• Unit laws of composition: T ◦ 1V = T and 1W ◦ T = T , where T : V → W .
• Distributive laws: (T + S) ◦ R = T ◦ R + S ◦ R and P ◦ (T + S) = P ◦ T + P ◦ S.
• Zero laws: 0 ◦ R = 0 and R ◦ 0 = 0.
• Compatibility with scalar multiplication: (kT ) ◦ S = k(T ◦ S) = T ◦ (kS), where k ∈ K .

Finally, we consider the notion of the inverse of a linear transformation.

Definition 10.16: Inverse of a linear transformation

Let V ,U be vector spaces over a field K , and let S : V → U and T : U → V be linear transformations.
We say that S and T are inverses if

T ◦ S = 1V and S ◦ T = 1U .

In this case, we also write S = T −1 and T = S−1 .

Proposition 10.17: Uniqueness of inverses

Inverses are unique. In other words, if S and S′ are two inverses of T , then S = S′ .

Proof. Suppose both S and S′ are inverses of T . Consider S ◦ T ◦ S′ . Since S and T are inverses, this is
equal to S′ , but since T and S′ are inverses, it is also equal to S. Therefore, S = S′ . ♠
378 Linear transformation of vector spaces

Proposition 10.18: Properties of inverses of linear transformations

• If S : V → U and T : U → W are both invertible, then so is T ◦ S, and we have

(T ◦ S)−1 = S−1 ◦ T −1 .

• 1−1 = 1.

• If S : V → U is invertible and k is a non-zero scalar, then

(kS)−1 = k−1 S−1 .

Exercises

Exercise 10.2.1 Suppose V and W are vector spaces, T , S : V → W and R, Q : W → W are linear trans-
formations, and k is a scalar. Which of the following equalities are valid?

(a) R ◦ (T + kS) = R ◦ T + kR ◦ S.

(b) (R + Q) ◦ (R + Q) = R ◦ R + 2R ◦ Q + Q ◦ Q.

(c) (R + Q) ◦ (T + S) = R ◦ T + R ◦ S + Q ◦ T + Q ◦ S.

(d) If T , S, and T + S are invertible, then (T + S)−1 = T −1 + S−1 .

Exercise 10.2.2 Finish the proof of Proposition 10.11, i.e., prove that if T : V → W is a linear transfor-
mation and k a scalar, then kT : V → W is a linear transformation.

10.3 Linear transformations defined on a basis

Outcomes
A. Check whether two linear transformations are equal by considering their action on a spanning
set.

B. Specify a linear transformation by considering its action on a basis.

Recall that, by definition, two linear transformations S, T : V → W are equal if and only if for all v ∈ V , we
have S(v) = T (v). However, this is not a very practical way of checking whether S = T , as it theoretically
 10.3. Linear transformations defined on a basis 379

requires checking S(v) = T (v) for each one of infinitely many vectors v. The following proposition states
that it is sufficient to check the actions of S and T on a spanning set of vectors.

Proposition 10.19: Equality of linear transformations

Let V and W be vector spaces over a field K , and let S, T : V → W be linear transformations.
Moreover, let X ⊆ V be a spanning set of V , i.e., such that V = span X . If S(v) = T (v) for all v ∈ X ,
then S = T .

Proof. Assume that S(v) = T (v) holds for all v ∈ X . To show that S = T , let u ∈ V be an arbitrary vector.
Since X is a spanning set, we can write u = a1 v1 + . . . + an vn , for some v1 , . . . , vn ∈ X and a1 , . . . , an ∈ K.
By assumption, S(vi ) = T (vi ) for all i, because vi ∈ X . Then we have

S(u) = S(a1v1 + . . . + an vn )
= a1 S(v1 ) + . . . + an S(vn )
= a1 T (v1 ) + . . . + an T (vn )
= T (a1 v1 + . . . + an vn )
= T (u).

Since u ∈ V was arbitrary, it follows that S = T . ♠

Therefore, if we know how a linear transformation acts on a spanning set (and in particular, on a basis),
then we know how it acts on the entire space. There is also a kind of converse to this: given a basis of V ,
we can map the basis vectors to any elements of W we like, and this will always determine a unique linear
transformations. This is the content of the following theorem.

Theorem 10.20: Linear transformation defined on a basis

Suppose V and W are vector spaces over a field K . Suppose v1 , v2 , . . . , vn is a basis of V , and
w1 , w2 , . . . , wn are any vectors in W (which may or may not be distinct). Then there exists a unique
linear transformation T : V → W such that

T (v1 ) = w1 , T (v2 ) = w2 , ... T (vn ) = wn .

Proof. To show that such a linear transformation T exists, we first define a function T : V → W as follows.
Given any v ∈ V , there exists a unique set of coordinates a1 , . . . , an ∈ K such that

v = a1 v1 + . . . + an vn .

Then define
T (v) = a1 w1 + . . . + an wn .
This defines a function T : V → W . Next, we must check that T is linear. To show that T preserves
addition, consider v, v′ ∈ V , with v = a1 v1 + . . . + an vn and v′ = b1 v1 + . . . + bn vn . Then

T (v + v′ ) = T ((a1 + b1 )v1 + . . . + (an + bn )vn )

= (a1 + b1 )w1 + . . . + (an + bn )wn
380 Linear transformation of vector spaces

= (a1 w1 + . . . + an wn ) + (b1 w1 + . . . + bn wn )
= T (v) + T (v′ ).

Therefore, T preserves addition. To show that T preserves scalar multiplication, consider v = a1 v1 + . . . +

an vn and k ∈ K. Then

T (kv) = T (ka1v1 + . . . + kan vn )

= ka1 w1 + . . . + kan wn
= k(a1 w1 + . . . + an wn )
= kT (v).

Therefore, T preserves scalar multiplication. It follows that T is linear. Next, we must show that T satisfies
the condition of the theorem, i.e., that T (vi ) = wi for each i. But this is clearly the case, because in this
case, ai = 1 and a j = 0 for all j 6= i. We have shown that there exists a linear function T satisfying all of
the conditions required by the theorem.
Finally, the only thing left to show is uniqueness. But this follows from Proposition 10.19. Namely, if
T ′ is another linear transformation such that T ′ (vi ) = wi for all i, then T and T ′ agree on v1 , . . . , vn , which
is a basis and hence a spanning set. By Proposition 10.19, T = T ′ . ♠

Example 10.21: Linear transformation defined on a basis


Recall that x2 , (x + 1)2 , (x + 2)2 is a basis of P2 . Consider the linear function T : P2 → M22
defined by
     
2 1 1 2 0 1 2 0 0
T (x ) = , T ((x + 1) ) = , T ((x + 2) ) = .
0 0 0 1 1 1

Find T (4x).

i h h i h i
Solution. Let v1 = x2 ,v2 = (x + 1)2 ,
v3 = 1
(x + 2)2 , 1 0 1
w1 = 0 0 , w2 = 0 1 , and w3 = 1 1 . We 0 0
must first find a, b, c such that 4x = av1 + bv2 + cv3 . We do this by solving a system of equations, using
the same method as in Example 9.12. We find that a = −3, b = 4, and c = −1. Therefore
 
−3 1
T (4x) = T (−3v1 + 4v2 − v3 ) = − 3w1 + 4w2 − w3 = .
−1 3

Exercises

Exercise 10.3.1 Let T : P2 → R be a linear transformation such that

T (x2 ) = 1, T (x2 + x) = 5, and T (x2 + x + 1) = −1.

 10.4. The matrix of a linear transformation 381

Find T (ax2 + bx + c).

Exercise 10.3.2 Let vectors v1 , . . . , vn ∈ Rn and w1 , . . . , wn ∈ Rm be given. Let A be the matrix whose
columns are v1 , . . . , vn , and assume that A−1 exists. Show that there exists a linear transformation T such
that T (vi ) = wi for i = 1, . . . , n.

10.4 The matrix of a linear transformation

Outcomes
A. Find the matrix of a linear transformation with respect to general bases in vector spaces.

In Section 6.2, we saw that linear transformations T : Rn → Rm are in one-to-one correspondence with
m ×n-matrices. Here, we will generalize this correspondence to arbitrary finite-dimensional vector spaces.
There is an important difference, however. While Rn comes with a natural coordinate system (i.e., every
vector in Rn has a first component, second component, and so on), there is no distinguished coordinate
system on an arbitrary vector space. To define the matrix of a linear transformation T : V → W , we must
first choose a basis, or equivalently a coordinate system, for V and for W . Different choices of basis will
give rise to different matrices.
Let V be a vector space with basis B = {v1 , . . . , vn }. Recall from Section 5.4.3 that the coordinates of
a vector v with respect to the basis B are the unique scalars a1 , . . . , an such that

v = a1 v1 + . . . + an vn .

As before, we write  
a1
[v]B =  a2 
a3
to denote the coordinates of v with respect to the basis B. We will now see how to use bases and coordinates
to encode any linear map between finite-dimensional vector spaces as a matrix.

Proposition 10.22: The matrix of a linear transformation

Let V and W be finite-dimensional vector spaces over a field K . Assume B is a basis of V and C is
a basis of W . Let T : V → W be a linear transformation. Then there exists a unique m × n-matrix A
such that for all v ∈ V ,
A [v]B = [T v]C .
Moreover, A can be computed as follows: the i th column of A holds the coordinates of T (vi ), where
vi is the i th vector of the basis B, and the coordinates are computed with respect to the basis C.
382 Linear transformation of vector spaces

Proof. Let B = {v1 , . . . , vn } and C = {w1 , . . . , wm }. By Theorem 10.20, the linear transformation T is
completely determined by the images of the basis vectors, T (v1 ), . . ., T (vn ) ∈ W . Since {w1 , . . . , wm } is a
basis of W , we can write each T (vi ) as a linear combination of w1 , . . . , wm :
T (v1 ) = a11 w1 + a21 w2 + . . . + am1 wm ,
T (v2 ) = a12 w1 + a22 w2 + . . . + am2 wm ,
···
T (vn ) = a1n w1 + a2n w2 + . . . + amn wm .
Let  
a11 a12 · · · a1n
 a21 a22 · · · a2n 
 
A= .. .. . . . .
 . . . .. 
am1 am2 · · · amn
Then A is an m × n-matrix. We must prove that it has the desired property, i.e., that A [v]B = [T v]C , for
all v ∈ V . Since both the left-hand side and the right-hand side are linear functions of v, it suffices to
check that this property holds for basis vectors. Consider, therefore, one of the basis vectors vi . Note that
vi = 0v1 + 0v2 + . . . + 1vi + . . . + 0vn . Therefore, the coordinates of vi with respect to the basis B are
 
0
 ... 
 
[vi ]B =  1  = ei ,
 .. 
.
0
where ei is the usual i th standard basis vector. On the other hand, since T (vi ) = a1i w1 + . . . + ami wm , we
have " #
a1i
[T (vi )]C = .. = Aei .
.
ami
Here, in the last equation, we have used the fact that Aei is the same thing as the i th column of A. We
therefore have [T (vi )]C = Aei = A [vi ]B , as desired. ♠

Definition 10.23: The matrix of a linear transformation

The matrix A of Proposition 10.22 is called the matrix of the linear transformation T with respect
to the bases B and C. We also write
A = [T ]C,B .
Therefore,
[T ]C,B [v]B = [T v]C
for all v ∈ V .

Example 10.24: Finding the matrix of a linear transformation


2 , x3 of
Find the matrix of thederivative
operator D : P 3 → P 2 with respect to the basis B = 1, x, x
P3 and the basis C = 1, x, x2 of P2 .
 10.4. The matrix of a linear transformation 383

Solution. We first find the images of each basis vector of the basis B, and we write each of them as a linear
combination of basis vectors from the basis C. Let us denote the basis vectors of B as v1 = 1, v2 = x,
v3 = x2 , and v4 = x3 , and the basis vectors of C as w1 = 1, w2 = x, and w3 = x2 . We have

D(v1 ) = D(1) = 0 = 0 + 0x + 0x2 = 0w1 + 0w2 + 0w3 ,

D(v2 ) = D(x) = 1 = 1 + 0x + 0x2 = 1w1 + 0w2 + 0w3 ,
D(v3 ) = D(x2 ) = 2x = 0 + 2x + 0x2 = 0w1 + 2w2 + 0w3 ,
D(v4 ) = D(x3 ) = 3x2 = 0 + 0x + 3x2 = 0w1 + 0w2 + 3w3 .

Therefore, we have  
0 1 0 0
A = [D]C,B =  0 0 2 0  .
0 0 0 3
♠

Example 10.25: Same linear transformation, different bases

operator D : P3 → P ′
Find
 the matrix of the derivative 2 with respect to the basis B =
1, x + 1, x2 + x + 1, x3 + x2 + x + 1 of P3 and the basis C′ = 1, x − 1, x2 − 1 of P2 .

Solution. This is the same linear transformation as in the previous example, but we are given different
bases. Let us denote the basis vectors of B′ as v1 = 1, v2 = x + 1, v3 = x2 + x + 1, and v4 = x3 + x2 + x + 1,
and the basis vectors of C′ as w1 = 1, w2 = x − 1, and w3 = x2 − 1. We must write each D(vi ) as a linear
combination of w1 , w2 , and w3 , which requires solving a system of linear equations for each of them. We
have:
D(v1 ) = D(1) = 0 = 0w1 + 0w2 + 0w3 ,
D(v2 ) = D(x + 1) = 1 = 1w1 + 0w2 + 0w3 ,
2
D(v3 ) = D(x + x + 1) = 2x + 1 = 3w1 + 2w2 + 0w3 ,
D(v4 ) = D(x3 + x2 + x + 1) = 3x2 + 2x + 1 = 6w1 + 2w2 + 3w3 .
Therefore, the matrix is  
0 1 3 6
[D]C′ ,B′ =  0 0 2 2 .
0 0 0 3
♠
The last two examples illustrate that a linear transformation can have many different matrices, because the
matrix depends not only on the linear transformation, but also on the given bases. The art of linear algebra
often lies in choosing “convenient” bases for a given application. Often, a “convenient” basis is one that
gives rise to simple matrices, for example, matrices containing many zeros, or matrices that are diagonal.
384 Linear transformation of vector spaces

Example 10.26: Finding a convenient basis

Let T : P3 → M2,2 be the linear transformation defined by
 
3 2 a+d b−c
T (ax + bx + cx + d) =
b+c a−d

for all ax3 + bx2 + cx + d ∈ P3 . Let B = x3 , x2 , x, 1 and
       
1 0 0 1 0 0 0 0
C= , , ,
0 0 0 0 1 0 0 1

be bases of P3 and M2,2 , respectively.

(a) Find [T ]C,B .

(b) Find a basis C′ of M2,2 such that [T ]C′,B is the identity matrix.

Solution. (a) We have

         
1 0 1 0 0 1 0 0 0 0
T (x3 ) = = 1 +0 +0 +1 ,
0 1 0 0 0 0 1 0 0 1
         
0 1 1 0 0 1 0 0 0 0
T (x2 ) = = 0 +1 +1 +0 ,
1 0 0 0 0 0 1 0 0 1
         
0 −1 1 0 0 1 0 0 0 0
T (x) = = 0 −1 +1 +0 ,
1 0 0 0 0 0 1 0 0 1
         
1 0 1 0 0 1 0 0 0 0
T (1) = = 1 +0 +0 −1 .
0 −1 0 0 0 0 1 0 0 1

Therefore,  
1 0 0 1
 0 1 −1 0 
[T ]C,B = 
 0
.
1 1 0 
1 0 0 −1

(b) Since the matrices T (x3 ), T (x2 ), T (x), and T (1) form a basis of M2,2 , we can take C′ to consist of
these four matrices, i.e.,
       
′ 1 0 0 1 0 −1 1 0
C = , , , .
0 1 1 0 1 0 0 −1

Then  
1 0 0 0
 0 1 0 0 
[T ]C′ ,B = 
 0

0 1 0 
0 0 0 1
is the identity matrix. ♠
 10.4. The matrix of a linear transformation 385

We end this section with some properties of matrices of linear transformations.

Proposition 10.27: Properties of matrices of linear transformations.

Let V , W , and U be finite-dimensional vector spaces with respective bases B, C, and D. Suppose
T , T ′ : V → W and S : W → U are linear transformations. The following hold:

(a) [S ◦ T ]D,B = [S]D,C [T ]C,B .

(b) [1V ]B,B = I .

 
(c) T is invertible if and only if [T ]C,B is invertible, and in that case, T −1 B,C = ([T ]C,B )−1 .

(d) [0]C,B = 0.

(e) [T + T ′ ]C,B = [T ]C,B + [T ′ ]C,B .

(f) [kT ]C,B = k [T ]C,B .

Exercises

     
2 3 2 5
Exercise 10.4.1 Let B = , be a basis of R and let x = be a vector in R2 . Find
−1 2 −7
[x]B .
       
 1 2 −1  5
Exercise 10.4.2 Let B =  −1  ,  1  ,  0  be a basis of R3 and let x =  −1  be a vector
 
2 2 2 4
2
in R . Find [x]B .

Exercise 10.4.3 Suppose T : R3 → R3 is a linear transformation such that

           
1 3 0 1 0 1
T  0  =  3 , T  1  =  2 , T  0  =  3 .
0 3 0 3 1 −1

Let E = {e1 , e2 , e3 } be the standard basis of R3 , and let

     
 1 1 2 
B = {v1 , v2 , v3 } =  
0 ,  
1 ,  1 
 
0 0 1
be another basis.

(a) Find the matrix of T with respect to E, i.e., find [T ]E,E .

386 Linear transformation of vector spaces

(b) Find [T ]B,B .

Exercise 10.4.4 Let T : R2 → R2 be a linear transformation defined by

   
a a+b
T = .
b a−b
Consider the two bases    
1 −1
B1 = ,
0 1
and    
1 1
B2 = , .
1 −1
Find the matrix MB2 ,B1 of T with respect to the bases B1 and B2 .
 
1 2
Exercise 10.4.5 Let M = , and consider the linear transformation T : M2,2 → M2,2 given by
−2 1
T (A) = MAM. Find the matrix of T with respect to the basis
       
1 0 0 1 0 0 0 0
B= , , , .
0 0 0 0 1 0 0 1

Exercise 10.4.6 Consider the linear transformation T : P3 → P3 given by T (p(x)) = p(x+1). Find [T ]B,B ,

where B = 1, x, x2 , x3 .
 
1
Exercise 10.4.7 Let v =  −2  and consider the linear function T (w) = projv (w). Find the matrix of
3
T with respect to the standard basis of R3 .
 
1
Exercise 10.4.8 Let v =  −2  and consider the linear function T (w) = projv (w). Find the matrix of
3
T with respect to the basis
     
 1 2 3 
B = {v1 , v2 , v3 } =  −2  ,  1  ,  0  .
 
3 0 1

Exercise 10.4.9 Suppose that V and W are finite-dimensional vector spaces with bases B and C, respec-
tively. Let T : V → W be a linear transformation such that

T (vi ) = wi

for i = 1, . . . , n. Let M be the matrix whose columns are [v1 ]B , . . . , [vn ]B , and let N be the matrix whose
columns are [w1 ]C , . . . , [wn ]C . Suppose that M is invertible. Show that [T ]C,B = NM −1 .
11. Inner product spaces
One structure that we often use in Rn , but which is missing from abstract vector spaces, is the dot product.
In Chapter 2, we saw how to use dot products to compute the length of a vector, the angle between two
vectors, and to decide when two vectors are orthogonal. We also used dot products to define the projection
of one vector onto another, and to find shortest distances between various objects (such as points and
planes, two lines, etc).
In this chapter, we will consider inner product spaces. An inner product space is essentially an abstract
vector spaces that has been equipped with an operation that works “like” the dot product.

11.1 Real inner product spaces

Outcomes
A. Check whether an operation is an inner product.

B. Give an example of a vector space on which more than one inner product can be defined.

C. Calculate the inner product of vectors in various examples of inner product spaces.

D. Calculate the norm of a vector and the angle between two vectors.

E. Use the Cauchy-Schwarz inequality and the triangle inequality to reason about the size of
inner products and norms of vectors.

In this section, we fix K to be the field of real numbers.

Definition 11.1: Real inner product space

A (real) inner product space is a real vector space V equipped with an operation that assigns to
any pair of vectors u, v ∈ V a real number hu, vi, called the inner product of u and v. This operation
must satisfy the following properties:

1. Symmetry: hu, vi = hv, ui.

2. Linearity: hu, kv + ℓwi = khu, vi + ℓhu, wi.

3. The positive definite property: hu, ui ≥ 0, and moreover, hu, ui = 0 if and only if u = 0.

387
388 Inner product spaces

Example 11.2: Rn with the usual dot product

V = Rn , with the usual dot product

·
hu, vi = u v = u1 v1 + . . . + un vn

is an inner product space.

Proof. The relevant properties were shown in Proposition 2.25. ♠

There exist other inner product operations on Rn besides the dot product. The following is an example.

Example 11.3: R2 with a non-standard inner product

 
2 1 1
Let V = R , and consider the matrix A = . Define an operation h−, −i by
1 2
  
T
  1 1 v1
hu, vi = u Av = u1 u2 = u1 v1 + u1 v2 + u2 v1 + 2u2 v2 .
1 2 v2

Then R2 , with this operation, is an inner product space.

Proof. We must show the three properties are satisfied. For symmetry, note that a scalar is equal to its own
transpose. Therefore
hu, vi = uT Av = (uT Av)T = vT AT u.
But this is equal to hv, ui because A is a symmetric matrix, i.e., AT = A. For linearity, note that, by
properties of matrix multiplication,
hu, kv + ℓwi = uT A(kv + ℓw) = kuT Av + ℓuT Aw = khu, vi + ℓhu, wi.
For the positive definite property, we have
hu, ui = u21 + 2u1 u2 + 2u22 = (u21 + 2u1 u2 + u22 ) + u22 = (u1 + u2 )2 + u22 ≥ 0,
because the sum of two squares is always non-negative. Moreover, if equality holds in the last equation,
then we must have u1 + u2 = 0 and u2 = 0, and this is only possible if u = 0. ♠

Example 11.4: Continuous functions on an interval

Let a < b be real numbers, and let

[a, b] = {x | a ≤ x ≤ b}

be the closed interval. Let V = C[a, b] be the vector space of all continuous functions f : [a, b] → R.
Given two functions f , g ∈ C[a, b], define
Z b
h f , gi = f (x)g(x) dx.
a

This is an inner product.

 11.1. Real inner product spaces 389

Proof. This follows from well-known properties of integrals. For symmetry, we have
Z b Z b
h f , gi = f (x)g(x) dx = g(x) f (x) dx = hg, f i.
a a

For linearity, we have

Z b Z b Z b
h f , kg + ℓhi = f (x)(kg(x) + ℓh(x)) dx = k f (x)g(x) dx + ℓ f (x)h(x) dx = kh f , gi + ℓh f , hi.
a a a

For the positive definite property, we have

Z b
hf, fi = f (x)2 dx ≥ 0.
a

This is ≥ 0 because a < b and the integral over a non-negative function is non-negative. Moreover, since f
is continuous, we know from calculus that the last integral can only be equal to 0 if f is the constant zero
function. ♠

Example 11.5: Polynomials

Let P be the vector space of polynomials (of any degree) with real coefficients. Let a < b be real
numbers, and consider the inner product defined by
Z b
hp, qi = p(x)q(x) dx.
a

This is an inner product space.

Proof. The proof is the same as for C[a, b]. ♠

Example 11.6: Real Hilbert space

Let HilbR be the vector space of all infinite sequences of real numbers a = (a0 , a1 , a2 , . . .) satisfying

a20 + a21 + a22 + . . . < ∞.

These are called the square summable sequences. (One needs to do a little bit of work to show that
it is indeed a vector space; in particular, to show that the sum of two square summable sequences is
square summable.) On this space, we can define an inner product as follows:

ha, bi = a0 b0 + a1 b1 + a2 b2 + . . . .

The details will be worked out in Exercise 11.1.6. This inner product space is called (real) Hilbert
space.

We now look at some properties of inner products. The first thing to note is that while the axioms require
linearity in the right component, symmetry ensures that linearity in the left component also holds.
390 Inner product spaces

Proposition 11.7: Left linearity

Let V be an inner product space. Then hku + ℓv, wi = khu, wi + ℓhv, wi.

Proof. This follows directly from symmetry. We have

hku + ℓv, wi = hw, ku + ℓvi = khw, ui + ℓhw, vi = khu, wi + ℓhv, wi.

♠
Another important property of inner products is the Cauchy-Schwarz inequality. We have already encoun-
tered this in the context of the dot product in Section 2.6.

Theorem 11.8: Cauchy-Schwarz inequality

Let u, v be vectors in an inner product space. Then

hu, vi2 ≤ hu, ui · hv, vi. (11.1)

Proof. The proof is almost identical to that of Proposition 2.27. First note that if u = 0, then both sides of
(11.1) are equal to zero, and so there is nothing to show. Therefore, we will assume in what follows that
u 6= 0. Define a function of t ∈ R by

f (t) = htu + v,tu + vi.

Then by the positive definite property, we know that f (t) ≥ 0 for all t ∈ R. Also from linearity and
symmetry, we have

f (t) = htu,tu + vi + hv,tu + vi

= t 2hu, ui + thu, vi + thv, ui + hv, vi
= t 2hu, ui + 2thu, vi + hv, vi.

This means the graph of y = f (t) is a parabola which opens upwards and is never negative. It follows
that this function has at most one root. From the quadratic formula, we know that a quadratic function
at 2 + bt + c has one or zero roots if and only if b2 − 4ac ≤ 0. Applying this reasoning to the function f (t),
we obtain
(2hu, vi)2 − 4hu, uihv, vi < 0,
which is equivalent to hu, vi2 ≤ hu, ui · hv, vi. ♠
Finally, we can use the inner product to define the norm of a vector.

Definition 11.9: Norm

Let u be a vector in an inner product space. Then the norm of u is defined as
p
kuk = hu, ui.
 11.1. Real inner product spaces 391

Example 11.10: Norm in C[−1, 1]

Calculate the norm of f (x) = x2 in C[−1, 1].

Solution. In the vector space C[−1, 1], the inner product is defined as
Z 1
h f , gi = f (x)g(x) dx.
−1

We therefore have  1
Z 1 Z 1
2 1 2
hf, fi = f (x) dx = x dx = x54
= .
−1 −1 5 −1 5
p q
2
Therefore, k f k = hf, fi = 5. ♠

A vector u in an inner product space is called normalized or a unit vector if kuk = 1. We note that if v is
any non-zero vector in an inner product space, then
1
u= v
kvk

is normalized.
By the Cauchy-Schwarz inequality, we have the following properties:

Proposition 11.11: Inner product and norm

Let u, v be vectors in an inner product space. Then

|hu, vi| ≤ kuk kvk .

Proof. The Cauchy-Schwarz inequality states that hu, vi2 ≤ kuk2 kvk2 . The claim follows by taking the
square root of both sides of the equation. ♠

Proposition 11.12: Triangle inequality

Let u, v be vectors in an inner product space. Then

ku + vk ≤ kuk + kvk .

This is called the triangle inequality.

Proof. The proof is essentially the same as that of Proposition 2.28. We have

ku + vk2 = hu + v, u + vi
= hu, ui + hu, vi + hv, ui + hv, vi
= kuk2 + 2hu, vi + kvk2
392 Inner product spaces

≤ kuk2 + 2 |hu, vi| + kvk2

≤ kuk2 + 2 kuk kvk + kvk2
= (kuk + kvk)2 .

The triangle inequality follows by taking square roots of both sides. ♠

Generalizing the situation in Rn , we can define the angle between any two vectors in an inner product
space. Note that by the Cauchy-Schwarz inequality,

|hu, vi|
≤ 1,
kuk kvk
and therefore,
hu, vi
−1 ≤ ≤ 1.
kuk kvk
This ensures that the following definition is well-defined.

Definition 11.13: Angle between two vectors

Let u, v be vectors in an inner product space. The angle between u and v is, by definition, the unique
θ such that 0 ≤ θ ≤ π and
hu, vi
cos θ = .
kuk kvk

Example 11.14: Find the angle between two vectors

Find the angle between 1 and x2 in C[−1, 1].

Solution. We have
Z 1
h1, 1i = 1 · 1 dx = 2,
−1
Z 1
2
hx2 , x2 i = x2 · x2 dx =
,
−1 5
Z 1
2 2
h1, x i = 1 · x2 dx = .
−1 3
Therefore √
2
h1, x2 i 3q 5
cos θ = = √ = .
k1k kx2 k 2 25 3
√
5
The angle θ is cos−1 ( 3 ), which is approximately 0.7297 radians or 41.81 degrees.
♠
 11.1. Real inner product spaces 393

Exercises

Exercise 11.1.1 For each matrix A, determine whether the formula hu, vi = uT Av determines an inner
product on R2 .
         
1 0 3 1 1 1 1 0 1 1
(a) A = , (b) A = , (c) A = , (d) A = , (e) A = .
0 2 1 3 0 2 0 −1 1 1

Exercise 11.1.2 Consider the inner product space C[0, 1] as in Example 11.4. Compute the following
inner products:
(a) h1, xi, (b) hx, x2 i, (c) h1 + x, 2 + x2 i.

Exercise 11.1.3 Consider the inner product space C[0, 1] as in Example 11.4. Compute the following
norms:
(a) k1k , (b) kxk , (c) x2 + 1 .

Exercise 11.1.4 For u, v vectors in R3 , define the product u ∗ v = u1 v1 + 2u2 v2 + 3u3 v3 . Show that

|u ∗ v| ≤ (u ∗ u)1/2 (v ∗ v)1/2 .

Hint: first show that the operation hu, vi = u ∗ v is an inner product on R3 , then use Proposition 11.11.

Exercise 11.1.5 In C[−1, 1], find (a) the angle between x and x2 , (b) the angle between x and x3 .

Exercise 11.1.6 . In this exercise, we will work out the details of Example 11.6. We must show that HilbR
is a vector space. We will do this by showing that it is a subspace of SeqR . Further, we must show that the
inner product is well-defined. This requires some knowledge of convergent series from calculus.

(a) Assume a = (a0 , a1 , . . .) and b = (b0 , b1 , . . .) are square summable sequences. Show that the series
a0 b0 + a1 b1 + a2 b2 + . . . converges to a real number. Hint: consider the series |a0 b0 | + |a1 b2 | + . . .
and use the Cauchy-Schwarz inequality and the absolute convergence test.

(b) Using the result of part (a), show that HilbR is a subspace of SeqR .

(c) Show that the operation ha, bi = a0 b0 + a1 b1 + a2 b2 + . . . is an inner product on HilbR .

394 Inner product spaces

11.2 Orthogonality

Outcomes
A. Determine whether two vectors in an inner product space are orthogonal.

B. Find the orthogonal complement of a set of vectors.

C. Determine whether a set of vectors is orthogonal and/or orthonormal.

D. Check whether a basis is orthogonal and/or orthonormal.

E. Calculate the Fourier coefficients of a vector with respect to an orthogonal basis.

Definition 11.15: Orthogonality

Let u and v be vectors in an inner product space. We say that u and v are orthogonal if

hu, vi = 0.

We also write u ⊥ v to indicate that u and v are orthogonal.

We note that the zero vector 0 is orthogonal to all vectors, because hu, 0i = 0 follows from the linearity
of the inner product. We also note that u ⊥ v if and only if v ⊥ u; this follows from the symmetry of the
inner product.

Definition 11.16: Orthogonal complement

Let S be a subset of an inner product space V . The orthogonal complement of S is the set

S⊥ = {v ∈ V | hv, wi = 0 for all w ∈ S } .

Proposition 11.17: Orthogonal complement

If S is any subset of an inner product space V , then S⊥ is a subspace.

Proof. We clearly have 0 ∈ S⊥, because 0 is orthogonal to all vectors. To show that S⊥ is closed under
addition, assume v, v′ ∈ S⊥. We have to show v + v′ ∈ S⊥ . So take an arbitrary w ∈ S. Then we have
hv + v′ , wi = hv, wi + hv′ , wi = 0 + 0 = 0.
Therefore, v + v′ ∈ S⊥ . Finally, to show that S⊥ is closed under scalar multiplication, assume v ∈ S⊥ and
k ∈ R. We have to show kv ∈ S⊥. So take an arbitrary w ∈ S. Then we have
hkv, wi = khv, wi = k0 = 0.
 11.2. Orthogonality 395

Therefore, kv ∈ S⊥ . It follows that S⊥ is a subspace of V . ♠

Here is an illustration of a subspace S of R3 and its orthogonal complement S⊥ :

S
S⊥

Example 11.18: Orthogonal complement

Consider the inner product space P3 of polynomials of degree at most 3, with the inner product
defined by
Z 1
h f , gi = f (x)g(x) dx.
−1

Find the orthogonal complement of x2 .

Solution. We have to compute the set of all polynomials of the form p(x) = ax3 + bx2 + cx + d that are
orthogonal to x2 . So let us compute the inner product:
Z 1
2
hp(x), x i = (ax3 + bx2 + cx + d)x2 dx
−1
Z 1
= ax5 + bx4 + cx3 + dx2 dx
−1
2 2
= 0a + b + 0c + d.
5 3
Setting this equal to 0, we see that hp(x), x2 i = 0 if and only if 52 b + 23 d = 0, or equivalently, 3b + 5d = 0.
The basic solutions are
           
a 1 a 0 a 0
 b   0   b   5   b   0 
  =  ,   =      
 c   0   c   0 ,  c  =  1 ,
d 0 d −3 d 0

giving the following basis for the space of polynomials orthogonal to x2 :

 3
x , 5x2 − 3, x .

♠
We now consider orthogonal sets and bases.
396 Inner product spaces

Definition 11.19: Orthogonal and orthonormal sets of vectors

A set of vectors {u1 , . . . , uk } in an inner product space is called an orthogonal set if the vectors are
non-zero and pairwise orthogonal, i.e., for all i, ui 6= 0 and for all i 6= j, ui ⊥ u j .
Moreover, the set of vectors {u1 , . . . , uk } is called orthonormal if it is orthogonal and each vector
is normalized, i.e., kui k = 1.

The interest of orthogonal and orthonormal sets of vectors lies, among other things, in the fact that they
are automatically linearly independent.

Proposition 11.20: Orthogonal set is linearly independent

If {u1 , . . . , uk } is an orthogonal set of vectors, then u1 , . . . , uk are linearly independent.

Proof. Assume {u1 , . . . , uk } is an orthogonal set. To show that u1 , . . . , uk are linearly independent, assume

a1 u1 + . . . + ak uk = 0.

We must show that a1 , . . . , ak = 0. So pick some i ∈ {1, . . . , k}. We compute

ha1 u1 + . . . + ak uk , ui i = a1 hu1 , ui i + . . . + ai hui , ui i + . . . + ak huk , ui i

= 0 + . . . + ai hui , ui i + . . . + 0
= ai hui , ui i.

On the other hand,

ha1 u1 + . . . + ak uk , ui i = h0, ui i = 0.
It follows that ai hui , ui i = 0. Since hui , ui i = 6 0, it follows that ai = 0. Since the choice of i was arbitrary,
it follows that a1 , . . . , ak = 0, and u1 , . . . , uk are linearly independent. ♠

Example 11.21: An orthogonal set of functions

Consider the inner product space C[0, 2π ]. The following functions form an (infinite) orthogonal
set:

g0 (x) = 1
f1 (x) = sin x
g1 (x) = cos x
f2 (x) = sin 2x
g2 (x) = cos 2x
.
..
fk (x) = sin kx
gk (x) = cos kx
.
..
 11.2. Orthogonality 397

Proof. We have to check that any two functions are orthogonal to each other. This is true because of
trigonometric identities. We have the following trigonometric formulas:
1
sin α sin β = (cos(α − β ) − cos(α + β )),
2
1
cos α cos β = (cos(α − β ) + cos(α + β )),
2
1
sin α cos β = (sin(α − β ) + sin(α + β )).
2
Using these, we can compute the relevant inner products. Assume i 6= j. Then
Z 2π
hgi , g j i = cos(ix) cos( jx) dx
0
Z
1 2π
= cos((i − j)x) + cos((i + j)x) dx
2 0
 2π
1 1 1
= sin((i − j)x) + sin((i + j)x)
2 (i − j) (i + j) 0
= 0,
and therefore gi ⊥ g j . The proofs of fi ⊥ f j and fi ⊥ g j are similar. ♠

Definition 11.22: Orthogonal and orthonormal bases

Let V be an inner product space, and let W be a subspace of V . We say that {u1 , . . . , uk } is an
orthogonal basis for W if it is an orthogonal set and spans W .
If, moreover, each ui is normalized, we say that {u1 , . . . , uk } is an orthonormal basis for W .

We note that, by Proposition 11.20, every orthogonal (or orthonormal) basis is automatically linearly
independent, and therefore an actual basis of W .

Example 11.23: Orthogonal, orthonormal, and non-orthogonal bases

Consider R2 as an inner product space with the usual dot product.
   
1 0
• , is an orthonormal basis of R2 .
0 1
    
1 1 1 1
• √ ,√ is an orthonormal basis of R2 .
2 1 2 −1
   
2 0
• , is an orthogonal basis of R2 , but not orthonormal.
0 3
   
1 1
• , is an orthogonal basis of R2 , but not orthonormal.
1 −1
   
1 1
• , is a basis of R2 , but neither orthogonal nor orthonormal.
0 1
398 Inner product spaces

So why are we interested in orthogonal and orthonormal bases? The main reason is that finding coordinates
is much easier when the bases are orthogonal (and even better when they are orthonormal). We have the
following property:

Proposition 11.24: Fourier coefficients

Let B = {u1 , . . . , un } be an orthogonal basis of some space W , and suppose v ∈ W . Then

v = a1 u1 + . . . + an un ,

where
hui , vi
ai = .
hui , ui i
In this situation, the coordinates a1 , . . . , an are also called the Fourier coefficients of v (with respect
to the orthogonal basis B).
In case B is an orthonormal basis, the formula is even simpler. In that case:

ai = hui , vi.

Proof. Since B is a basis of W , we know that there exist coefficients a1 , . . . , an such that v = a1 u1 + . . . +
an un . It remains to verify that the coefficients satisfy the required formula. This is a simple calculation.
We have

hui , vi = hui , a1 u1 + . . . + an un i
= a1 hui , u1 i + . . . + ai hui , ui i + . . . + an hui , un i
= 0 + . . . + ai hui , ui i + . . . + 0
= ai hui , ui i.

Since ui 6= 0, we have hui , ui i =

6 0 by the positive definite property. We can therefore divide both sides of
the equation by hui , ui i to obtain
hui , vi
ai = ,
hui , ui i
as desired. Finally, if the basis is orthonormal, then hui , ui i = 1, and the denominator disappears, so that
ai = hui , vi. ♠

Example 11.25: Fourier coefficients

Suppose√B = {u1 , u2 , u3 } is an orthogonal basis for an inner product space V , such that ku1 k = 1,
ku2 k = 2, and ku3 k = 2. Moreover, suppose that v ∈ V is a vector such that hv, u1 i = 3, hv, u2 i =
−1, and hv, u3 i = 2. Find the coordinates of v with respect to B.

Solution. We have to find a1 , a2 , a3 such that

v = a1 u1 + a2 u2 + a3 u3 .
 11.2. Orthogonality 399

We have sufficient information to compute a1 , a2 , a3 . Namely,

hu1 , vi hu1 , vi 3
a1 = = 2
= = 3,
hu1 , u1 i ku1 k 1
hu2 , vi hu2 , vi −1
a2 = = 2
= = − 0.5,
hu2 , u2 i ku2 k 2
hu3 , vi hu3 , vi 2
a3 = = 2
= = 0.5.
hu3 , u3 i ku3 k 4

♠
Proposition 11.24 shows that if {u1 , . . . , un } is an orthogonal basis, then we can solve v = a1 u1 +. . .+an un
without having to solve a system of equations. While this is a useful thing to be able to do, perhaps it is
merely a convenience (solving a system of equations would also be fine). However, there is another very
useful property of Fourier coefficients. The coefficient ai only depends on the basis vector ui , and not on
any of the other basis vectors. This is useful in situations where only part of an orthogonal basis is known.
We can calculate the corresponding coefficients without having to know the rest of the basis.

Example 11.26: Finding partial coordinates

Suppose B = {u1 , u2 , u3 , u4 } is an orthogonal basis of R4 . We have been told that
   
1 0
 2   1 
u1 =  
 −1  and u2 =  2  ,
 

0 3

but it is not known what u3 and u4 are. Find the first two coordinates of the vector
 
1
 0 
v=  2 


with respect to the basis B.

Solution. We have v = a1 u1 + a2 u2 + a3 u3 + a4 u4 , where

hu1 , vi −1 hu2 , vi 4
a1 = = and a2 = = .
hu1 , u1 i 6 hu2 , u2 i 14

So the first two coordinates are − 61 and 27 . ♠

400 Inner product spaces

A word of warning is in order: Fourier coefficients do not work when the basis is not orthogonal. Consider
the following picture, where u1 and u2 are not orthogonal.

a2

u2 v

u1 a1

The coordinates of v with respect to u1 , u2 are (2, 1), because v = 2u1 + 1u2 , as indicated by the shaded
parallelogram. On the other hand, the formula for the Fourier coefficients is concerned with the orthogonal
projections of v onto u1 and u2 , as indicated by the dashed lines. It yields the coefficients
hu1 , vi
a1 = ≈ 2.5,
hu1 , u1 i
hu2 , vi
a2 = ≈ 1.8.
hu2 , u2 i
These are not the same as the coordinates of v, because the dashed lines are orthogonal to u1 and u2 ,
instead of parallel to them. In summary, the Fourier coefficients of a vector are equal to its coordinates
only if the basis is orthogonal.
When u1 , . . . , uk are orthogonal, we have a convenient formula for the norm of a1 u1 + . . . + ak uk :

Proposition 11.27: Norm of orthogonal linear combination

Let V be an inner product space, and suppose that u1 , . . . , uk are orthogonal. Then

ka1 u1 + . . . + ak uk k2 = a21 ku1 k2 + . . . + a2k kuk k2 .

Proof. We have
ka1 u1 + . . . + ak uk k2 = ha1 u1 + . . . + ak uk , a1 u1 + . . . + ak uk i
= a21 hu1 , u1 i + a1 a2 hu1 , u2 i + . . . + a1 an hu1 , un i
+ a2 a1 hu2 , u1 i + a22 hu2 , u2 i + . . . + a2 an hu2 , un i
.. .. ..
+ . . .
+ an a1 hun , u1 i + an a2 hun , u2 i + . . . + a2n hun , un i
= a21 ku1 k2 + . . . + a2k kuk k2 .
Here we have used the fact that hui , u j i = 0 when i 6= j. ♠
We conclude this section by making a connection between arbitrary inner products and dot products. Once
an orthonormal basis has been chosen on an inner product space, computing inner products is essentially
the same as computing dot products. The following proposition makes this more precise.
 11.2. Orthogonality 401

Proposition 11.28: Inner product and dot product

Let V be an inner product space, and suppose that B = {u1 , . . . un } is an orthonormal basis. Then
for any pair of vectors v, w ∈ V , their inner product is equal to the dot product of their coordinate
vectors with respect to the basis B, i.e.

·
hv, wi = [v]B [w]B .

Proof. By definition of coordinate vectors, we have

   
a1 b1
   
[v]B =  ...  and [w]B =  ...  ,
an bn

where v = a1 u1 + . . . + an un and w = b1 u1 + . . . + bn un . We calculate

hv, wi = ha1 u1 + . . . + an un , b1 u1 + . . . + bn un i
= a1 b1 hu1 , u1 i + a1 b2 hu1 , u2 i + . . . + a1 bn hu1 , un i
+ a2 b1 hu2 , u1 i + a2 b2 hu2 , u2 i + . . . + a2 bn hu2 , un i
.. .. ..
+ . . .
+ an b1 hun , u1 i + an b2 hun , u2 i + . . . + an bn hun , un i
= a1 b1 + a2 b2 + . . . + an bn
·
= [v]B [w]B .

Here we have used the fact that hui , u j i = 1 when i = j and hui , u j i = 0 otherwise, which holds because B
is orthonormal. ♠

Exercises

 
1 1 0
Exercise 11.2.1 Let A =  1 2 0 , and consider R3 with the inner product given by hu, vi = uT Av.
0 0 2
Which of the following vectors are orthogonal to each other?
       
1 −1 7 10
u1 =  1  , u2 =  2  , u3 =  −5  , u4 =  −2  .
1 −2 −2 −7

Exercise 11.2.2 On C[−1, 1], which of the following functions are orthogonal to each other?

f1 (x) = x, f2 (x) = x2 , f3 (x) = x3 − x, f4 (x) = 1 − x4 .

402 Inner product spaces

Exercise 11.2.3 Consider the inner product space P3 of polynomials of degree at most 3, with the inner
product defined by
Z 1
h f , gi = f (x)g(x) dx.
−1

(a) Find the orthogonal complement of x2 , x .

(b) Find the orthogonal complement of {x + 1}.

Exercise 11.2.4 Consider R3 as an inner product space with the usual dot product. For each of the
following bases of R3 , state whether it is orthonormal, orthogonal, or neither.
     
 1 0 0 
(a)  0  ,  1  ,  0  .
 
0 0 1
     
 1 0 0 
(b)  0 , 1 , 0  .
   
 
1 1 1
     
 1 0 −2 
(c)  0  ,  1  ,  0  .
 
2 0 1
     
 3 4 
 5 0 5 
 4     3 
(d)  5  ,  0  ,  − 5  .

 

0 −1 0

Exercise 11.2.5 Suppose

√ B = {u1 , u√
2 , u3 } is an orthogonal basis for an inner product space V , such that
ku1 k = 2, ku2 k = 3, and ku3 k = 5. Moreover, suppose that v ∈ V is a vector such that hv, u1 i = 1,
hv, u2 i = 2, and hv, u3 i = −4. Find the coordinates of v with respect to B.

Exercise 11.2.6 Suppose B = {u1 , u2 , u3 } is an orthogonal basis of R3 . We have been told that
 
1
u1 =  1  ,
0

but it is not known what u2 and u3 are. Find the first coordinate of the vector
 
1
v= 0 
2

with respect to the basis B.

 11.3. The Gram-Schmidt orthogonalization procedure 403

11.3 The Gram-Schmidt orthogonalization procedure

Outcomes
A. Use the Gram-Schmidt procedure to find an orthogonal basis of a subspace of an inner product
space.

B. Find an orthonormal basis of a subspace.

Although we have already seen some potential uses for orthogonal bases, we have not yet seen very many
examples of such bases. In this section, we will look at the Gram-Schmidt orthogonalization procedure, a
method for turning any basis into an orthogonal one.
The basic idea is very simple: if two vectors v1 , v2 are not orthogonal, then we can make them orthog-
onal by replacing v2 by a vector of the form u2 = v2 − tv1 , for a suitable parameter t.

v1 v2

u2 = v2 − tv1

But what is the correct value of t? It turns out that this value is uniquely determined by the requirement
that v1 and u2 must be orthogonal. We calculate
hv1 , u2 i = hv1 , v2 − tv1 i = hv1 , v2 i − thv1 , v1 i.
Setting this equal to 0 yields the unique solution
hv1 , v2 i
t=
hv1 , v1 i
Note that this is exactly the same thing as the Fourier coefficient of v2 in the direction of v1 . The following
proposition summarizes what we have found so far. For consistency with our later notation, we also
rename the first basis vector v1 to u1 .

Proposition 11.29: Gram-Schmidt orthogonalization procedure for 2 vectors

Let {v1 , v2 } be a basis for some subspace W of an inner product space V . Define vectors u1 , u2 as
follows:

u 1 = v1 ,
hu1 , v2 i
u 2 = v2 − u1 .
hu1 , u1 i

Then {u1 , u2 } is an orthogonal basis of W .

404 Inner product spaces

Example 11.30: Gram-Schmidt orthogonalization procedure for 2 vectors

In R3 with the usual dot product, find an orthogonal basis for
   
 1 0 
span  1  ,  1  .
 
0 1

   
1 0
Solution. Let v1 = 1 and v2 = 1 . We calculate
  
0 1
 
1
u 1 = v1 =  1 ,
0
     
0 1 −1/2
hu1 , v2 i 1
u 2 = v2 − u1 =  1  −  1  =  1/2  .
hu1 , u1 i 2
1 0 1
   
 1 −1/2 
Therefore the desired orthogonal basis is  1  ,  1/2  . ♠
 
0 1
The procedure for finding an orthogonal basis of a k-dimensional space is very similar. We adjust each
basis vector vi by subtracting a suitable linear combination of previous orthogonal basis vectors.

Proposition 11.31: Gram-Schmidt orthogonalization procedure for k vectors

Let {v1 , . . . , vk } be a basis for some subspace W of an inner product space V . Define vectors
u1 , . . . , uk as follows:

u1 = v1 ,
hu1 , v2 i
u2 = v2 − u1 ,
hu1 , u1 i
hu1 , v3 i hu2 , v3 i
u3 = v3 − u1 − u2 ,
hu1 , u1 i hu2 , u2 i
.
..
hu1 , vk i hu2 , vk i huk−1 , vk i
uk = vk − u1 − u2 − . . . − uk−1 .
hu1 , u1 i hu2 , u2 i huk−1 , uk−1 i

Then {u1 , . . . , uk } is an orthogonal basis of W .

Proof. First, it is clear that {v1 , . . . , vk } and {u1 , . . . , uk } span the same subspace, as each vi is a linear
combination of u1 , . . . , ui and conversely, each ui is a linear combination of v1 , . . . , vi . So the only thing
we must check is that {u1 , . . . , uk } is an orthogonal set. In other words, we must show that hu j , ui i = 0 for
 11.3. The Gram-Schmidt orthogonalization procedure 405

all j < i. We prove this by induction on i, i.e., we assume it is already true for all pairs of indices smaller
than i. To show hu j , ui i = 0, we calculate:
hu1 ,vi i hu j ,vi i hui−1 ,vi i
hu j , ui i = hu j , vi − hu1 ,u1 i u1 − ... − hu j , u j i u j − ... − hui−1 ,ui−1 i ui−1 i
hu1 ,vi i hu j ,vi i hui−1 ,vi i
= hu j , vi i − hu1 ,u1 i hu j , u1 i − ... − hu j ,u j i hu j , u j i − ... − hui−1 ,ui−1 i hu j , ui−1 i
hu j ,vi i
= hu j , vi i − 0 − . . . − hu j ,u j i hu j , u j i − ... − 0
= hu j , vi i − hu j , vi i
= 0.

It follows that the set {u1 , . . . , uk } is orthogonal, as desired. ♠

Example 11.32: Gram-Schmidt orthogonalization procedure

In R4 , find an orthogonal basis for
     

 1 1 1 
     1 
1   1
span  , ,   .

 1   1   0 

 
1 0 0

Solution. Let     
1 1 1
 1   1   1 
v1 =  
 1 , v2 =  
 1 , and v3 =  
 0 .
1 0 0
We calculate
 
1
 1 
u1 = v1 =  
 1 ,
1
 
   
1 1 1/4
hu1 , v2 i  1     
u2 = v2 − u1 =   − 3  1  =  1/4  ,
hu1 , u1 i  1  4 1   1/4 
0 1 −3/4
hu1 , v3 i hu2 , v3 i
u3 = v3 − u1 − u2
hu1 , u1 i hu2 , u2 i
       
1 1 1/4 1/3
 1       
=   − 2  1  − 1/2  1/4  =  1/3  .
 0  4 1  3/4  1/4   −2/3 
0 1 −3/4 0
406 Inner product spaces

     

 1 1/4 1/3 
    1/3 
1   1/4 
Therefore the orthogonal basis is {u1 , u2 , u3 } =  1 , 
,   . ♠

 1/4   −2/3 

 
1 −3/4 0
The Gram-Schmidt procedure is sensitive to reordering the vectors. For example, if we order the original
basis vectors in Example 11.32 in the opposite order, we end up with a different orthogonal basis at the
end. Sometimes this can simplify the calculations, as the following example shows.

Example 11.33: Gram-Schmidt orthogonalization procedure: reordering the vectors

In R4 , find an orthogonal basis for
     

 1 1 1 
     1 
1   1
span  , ,   .

 0   1   1 

 
0 0 1

Solution. Note that these are the same basis vectors as in Example 11.32, but listed in a different order.
Let      
1 1 1
 1   1   1 
v1 =      
 0  , v2 =  1  , and v3 =  1  .
0 0 1
We calculate


1
 1 
u1 = v1 =  
 0 ,
0
     
1 1 0
hu1 , v2 i  1     
u 2 = v2 − u1 =   − 2 1 
= 
0 ,
hu1 , u1 i  1  2 0   1 
0 0 0
hu1 , v3 i hu2 , v3 i
u 3 = v3 − u1 − u2
hu1 , u1 i hu2 , u2 i
       
1 1 0 0
 1       0 
=   − 2 1  − 1 0  =  .
 1  2 0  1 1   0 
1 0 0 1
     

 1 0 0 

 
1   0   0
This time, we end up with the orthogonal basis {u1 , u2 , u3 } =  
,  ,   . ♠

 0   1   0 

 
0 0 1
In the next example, we will consider Rn , but with a non-standard inner product.
 11.3. The Gram-Schmidt orthogonalization procedure 407

Example 11.34: Gram-Schmidt orthogonalization procedure, non-standard inner product

Let  
1 2 −2
A= 2 6 −1  ,
−2 −1 9
and consider the vector space R3 with the inner product given by hv, wi = vT Aw. Let
     
1 0 0
v1 =  0  , v2 =  1  , and v3 =  0  .
0 0 1

Apply the Gram-Schmidt procedure to v1 , v2 , v3 to find an orthogonal basis {u1 , u2 , u3 } for R3 with
respect to the above inner product.

Solution. As usual, we start with


1
u1 = v1 =  0  .
0

Next, we calculate

hu1 , v2 i = uT1 Av2 = 2,

hu1 , u1 i = uT1 Au1 = 1.

Therefore,

    
0 1 −2
hu1 , v2 i 2
u2 = v2 − u1 =  1  −  0  =  1  .
hu1 , u1 i 1
0 0 0

Finally, we calculate

hu1 , v3 i = uT1 Av3 = −2,

hu2 , v3 i = uT2 Av3 = 3,
hu2 , u2 i = uT2 Au2 = 2.

Therefore,
       
0 1 −2 5
hu1 , v3 i hu2 , v3 i −2   3 
u3 = v3 − u1 − u2 =  0  − 0 − 1  =  −3/2  .
hu1 , u1 i hu2 , u2 i 1 2
1 0 0 1

So the desired orthogonal basis is

     
 1 −2 5 
{u1 , u2 , u3 } =  
0 ,  1  
, −3/2  .
 
0 0 1
408 Inner product spaces

Note that it is not orthogonal with respect to the dot product, but with respect to the inner product defined
above. ♠

Example 11.35: Legendre polynomials

Consider the vector space P of polynomials, with the inner product
Z 1
hp, qi = p(x)q(x) dx.
−1

Use the Gram-Schmidt procedure to find an orthogonal basis for span 1, x, x2 , x3 .

Solution. Let v1 = 1, v2 = x, v3 = x2 , and v4 = x3 . We follow the Gram-Schmidt procedure:

u1 = v1 = 1.

Before we calculate u2 , we have to evaluate two integrals:

Z 1  1 2 1
hu1 , v2 i = 1 · x dx = 2 x −1 = 0,
−1
Z 1
hu1 , u1 i = 1 · 1 dx = [x]1−1 = 2.
−1

Then
hu1 , v2 i 0
u 2 = v2 − u1 = x − · 1 = x.
hu1 , u1 i 2

To calculate u2 , we first evaluate three integrals:

Z 1  1 3 1 2
hu1 , v3 i = 1 · x2 dx =
3 x −1 = 3 ,
−1
Z 1  1
hu2 , v3 i = x · x2 dx = 14 x4 −1 = 0,
−1
Z 1  1 3 1 2
hu2 , u2 i = x · x dx = 3 x −1 = 3 .
−1

Then
hu1 , v3 i hu2 , v3 i 2/3 1
u3 = v3 − u1 − u2 = x2 − · 1 − 0 · x = x2 − .
hu1 , u1 i hu2 , u2 i 2 3

To calculate u3 , we first evaluate four integrals:

Z 1  1 4 1
hu1 , v4 i = 1 · x3 dx = 4 x −1 = 0,
−1
Z 1  1 5 1 2
hu2 , v4 i = x · x3 dx = 5 x −1 = 5 ,
−1
 11.3. The Gram-Schmidt orthogonalization procedure 409

Z 1 1 1
hu3 , v4 i = (x2 − 13 ) · x3 dx = 6 1 4
6 x − 12 x −1
= 0,
−1
Z 1 1 
hu3 , u3 i = (x2 − 13 )2 dx = 5 − 2 x3 + 1 x 1 8
5 x 9 9 −1 = 45 .
−1

Then
hu1 , v4 i hu2 , v4 i hu3 , v4 i
u4 = v4 − u1 − u2 − u3
hu1 , u1 i hu2 , u2 i hu3 , u3 i
0 2/5 0 1 3
= x3 − · 1 − ·x − · (x2 − ) = x3 − x.
2 2/3 8/45 3 5

Thus, we obtain the orthogonal basis {u1 , u2 , u3 , u4 } = 1, x, x2 − 13 , x3 − 35 x . ♠
The orthogonal polynomials from Example 11.35 are known (up to scalar multiples) as Legendre poly-
nomials. We can continue in the same fashion applying the Gram-Schmidt procedure to the polynomials
1, x, x2 , x3 , x4 , x5 , x6 , . . . to get an infinite sequence of orthogonal polynomials. The first few elements of
this sequence are:

p0 (x) = 1,
p1 (x) = x,
1
p2 (x) = x2 − ,
3
3
p3 (x) = x3 − x,
5
6 3
p4 (x) = x − x2 + ,
4
7 35
10 5
p5 (x) = x5 − x3 + x,
9 21
15 5 5
p6 (x) = x6 − x4 + x2 − ,
11 11 231
21 105 3 35
p7 (x) = x7 − x5 + x − x,
13 143 429
28 14 28 2 7
p8 (x) = x8 − x6 + x4 − x + .
15 13 143 1287
410 Inner product spaces

p(x)

x2 − 13

x4 − 67 x2 + 35
3

x6 − 15 4 5 2 5
11 x + 11 x − 231 x
x5 − 10 3 5
9 x + 21 x −1 1

x3 − 35 x

The Gram-Schmidt procedure yields an orthogonal basis. If we want to compute an orthonormal basis,
we also have to normalize each basis vector. Since normalization usually involves dividing by a square
root, it is best to do this at the end, i.e., after the entire Gram-Schmidt procedure is complete, rather than
normalizing each ui immediately after it is found. Note that the Gram-Schmidt procedure itself does not
involve computing any square roots.

Example 11.36: Finding an orthonormal basis

In R4 , find an orthogonal basis for
     

 1 1 1 

     
1   1   1
span  1 ,  ,  .

 1   0 

 
1 0 0

Solution. In Example 11.32, we already found an orthogonal basis

     

 1 1/4 1/3 

     1/3 
 1   1/4   
{u1 , u2 , u3 } =   ,  ,

 1 1/4   −2/3  
 
1 −3/4 0
 11.3. The Gram-Schmidt orthogonalization procedure 411

for this space. So all that is left to do is to normalize each vector. The orthonormal basis is
      
  
 1 1/4 √ 1/3 

u1 u2 u3 1   2   3  1/3 
1
 , √  1/4   
, ,
ku1 k ku2 k ku3 k
=
 2  1  3  1/4  , √2  −2/3  .

 

1 −3/4 0
Alternatively, we could have also normalized the orthogonal basis we found in Example 11.33. In that
case, we obtain the orthonormal basis
      

 1 0 0 

 1    0   0 
1 
√ 
 ,  ,   .
 0   1   0 
 2
 

0 0 1
♠

Example 11.37: Orthonormal basis of polynomials

Find an orthonormal basis for the space of Example 11.35.


Solution. In Example 11.35, we found the orthogonal basis {u1 , u2 , u3 , u4 } = 1, x, x2 − 13 , x3 − 35 x . We
also computed
2 8
hu1 , u1 i = 2, hu2 , u2 i = , and hu3 , u3 i = .
3 45
We also need to compute hu4 , u4 i:
Z 1 Z 1  
3 3 1 8
hu4 , u4 i = (x3 − 53 x)2 dx = x6 − 65 x4 + 259 2
x dx = 71 x7 − 25
6 5
x + 25 x −1 = .
−1 −1 175
Therefore, the orthonormal basis is:
  ( r r r )
u1 u2 u3 u4 1 3 45 2 1 175 3 3
, , , = √ , x, (x − ), (x − x) .
ku1 k ku2 k ku3 k ku4 k 2 2 8 3 8 5
♠
We finish this section by remarking that the formula
hu, vi
u
hu, ui
is exactly what we called the projection of v onto u in Section 2.6.5, except that we have generalized this
concept from Rn to an arbitrary inner product space. We can define
hu, vi
proju (v) = u.
hu, ui
With this definition, the Gram-Schmidt procedure can also be expressed more succinctly as follows.
u1 = v1 ,
u2 = v2 − proju1 (v2 ),
u3 = v3 − proju1 (v3 ) − proju2 (v3 ),
..
.
uk = vk − proju1 (vk ) − proju2 (vk ) − . . . − projuk−1 (vk ).
412 Inner product spaces

Exercises

Exercise 11.3.1 In R3 with the usual dot product, find an orthogonal basis for
   
 1 2 
span  2  ,  6  .
 
3 0

Exercise 11.3.2 In R4 with the usual dot product, find an orthogonal basis for
   

 0 3  
   
1 0
span   
 1  ,  2  .


 
 
0 −1

Exercise 11.3.3 In R4 with the usual dot product, find an orthogonal basis for
     

 1 1 2 
     
0 3 4
span      
 1  ,  1  ,  2  .

 
 
0 −1 2

Exercise 11.3.4 Let  

3 −1 0
A =  −1 5 2  ,
0 2 3
and consider the vector space R3 with the inner product given by hv, wi = vT Aw. Let
     
1 −1 2
 
v1 = 0 , v2 =  1  , and v3 = 2  .
2 −5 3
Apply the Gram-Schmidt procedure to v1 , v2 , v3 to find an orthogonal basis {u1 , u2 , u3 } for R3 with respect
to the above inner product.

Exercise 11.3.5 Let  

3 2 0 0
 2 5 1 0 
A=
 0
,
1 3 1 
0 0 1 3
and consider the vector space R4 with the inner product given by hv, wi = vT Aw. Let
     
1 2 2
 −1   0   0 
v1 =    
 1  , v2 =  0  , and v3 =  2  ,
 

1 2 3
 11.4. Orthogonal projections and Fourier series 413

and let W = span {v1 , v2 , v3 }. Apply the Gram-Schmidt procedure to v1 , v2 , v3 to find an orthogonal basis
{u1 , u2 , u3 } for W with respect to the above inner product.

Exercise 11.3.6 Consider the inner product space C[0, 2], with the inner product given by
Z 2
hp, qi = f (x)g(x) dx.
0

Use the Gram-Schmidt procedure to find an orthogonal basis for span 1, x, x2 .

Exercise 11.3.7 Find an orthonormal basis for the subspace of R3 from Exercise 11.3.1.

11.4 Orthogonal projections and Fourier series

Outcomes
A. Compute the orthogonal projection of a vector onto a subspace.

B. Find the least squares approximation of a function by a polynomial of a given degree.

C. Calculate the generalized Fourier series of a vector.

D. Use Fourier series to approximate a function by polynomials or trigonometric functions.

We now reconsider a problem that we briefly encountered, only in the context of R3 , in Section 3.2:
how to find the shortest distance between a point and a subspace. The method we used in Section 3.2
(see Example 3.26) relies on the existence of normal vectors and does not generalize beyond R3 . The
following proposition gives a much better method for solving this problem, provided that we have an
orthogonal basis of the subspace.
414 Inner product spaces

Proposition 11.38: Orthogonal projection onto a subspace

Let V be an inner product space, and let W be a subspace of V . Assume {u1 , . . . , uk } is an orthogonal
basis of W , and v ∈ V is any vector. Then the following vector v′ is the element of W that is closest
to v, i.e., such that kv − v′ k is as small as possible.

hu1 , vi hu2 , vi huk , vi

v′ = u1 + u2 + . . . + uk .
hu1 , u1 i hu2 , u2 i huk , uk i

Moreover, the vector v − v′ is orthogonal to W .

v′
u2
0
u1 W

The vector v′ is called the orthogonal projection of v onto W . We also say that v′ is the best
approximation of v in W .

Proof. In case v ∈ W , then by Proposition 11.24, we have

hu1 , vi hu2 , vi huk , vi

v= u1 + u2 + . . . + uk ,
hu1 , u1 i hu2 , u2 i huk , uk i

and therefore v′ = v. In that case kv − v′ k = 0, which is clearly as small as possible, and v − v′ = 0 is

orthogonal to W , so we are done.
Now assume that v 6∈ W . Let W ′ = span {u1 , . . . , uk , v}. Apply the Gram-Schmidt procedure to the k +1
vectors u1 , . . . , uk , v to obtain an orthogonal basis of W ′ . Since the first k vectors are already orthogonal,
the Gram-Schmidt procedure does not change them, and we therefore obtain an orthogonal basis of W ′
of the form u1 , . . . , uk , uk+1 . Since v ∈ W ′ , we can write v as a linear combination of these basis vectors,
i.e., v = a1 u1 + . . . + ak uk + ak+1 uk+1 . Moreover, by Proposition 11.24, we know that each a j is the
hu j ,vi
corresponding Fourier coefficient of v, i.e., a j = hu j , u j i .
Now consider any w ∈ W . Since {u1 , . . . , uk } is a basis of W , we can write w = x1 u1 + . . . + xk uk .
Using Proposition 11.27, we have

kv − wk2 = k(a1 − x1 )u1 + . . . + (ak − xk )uk + ak+1 uk+1 k2

= (a1 − x1 )2 ku1 k2 + . . . + (ak − xk )2 kuk k2 + a2k+1 kuk+1 k2 .

Here, a1 , . . . , ak+1 are the fixed coordinates of v, and x1 , . . . , xk depend on w. Therefore, kv − wk takes its
 11.4. Orthogonal projections and Fourier series 415

smallest value when x j = a j , for j = 1, . . . , k. So the minimum value of kv − wk occurs when

hu1 , vi hu2 , vi huk , vi

w = a1 u1 + . . . + ak uk = u1 + u2 + . . . + uk = v′ .
hu1 , u1 i hu2 , u2 i huk , uk i

This is what had to be shown. Finally, to show that v − v′ is orthogonal to W , note that v − v′ = ak+1 uk+1 ,
which is orthogonal to each u1 , . . . , uk , and therefore to W . ♠

Example 11.39: Orthogonal projection onto a subspace

Consider R4 with the usual dot product. Let W = span {u1 , u2 }, where
   
1 −1
 1   
u1 =   , u2 =  −1  .
 2   1 
0 3

Note that u1 and u2 are orthogonal. Let


1
 2 
v= 
 3 .
4

Find the best approximation of v in W .

Solution. We calculate hu1 , vi = 9, hu1 , u1 i = 6, hu2 , vi = 12, and hu2 , u2 i = 12. Therefore, by Proposi-
tion 11.38, the desired vector is
     
1 −1 1/2
hu1 , vi hu2 , vi 9 1    
 + 12  −1  =  1/2  .

v′ = u1 + u2 =      
hu1 , u1 i hu2 , u2 i 6 2 12 1 4 
0 3 3

♠
Note how straightforward the calculations in this example are. All we had to do is calculate a few inner
products. It is not even necessary to solve a system of equations. Such is the power of orthogonal bases.
The next example shows that we can use exactly the same method to find approximations of functions.
416 Inner product spaces

Example 11.40: Approximating a function by a polynomial

Let V = C[−1, 1] be the vector space of continuous functions, with the inner product given by
Z 1
h f , gi = f (x)g(x) dx.
−1

Consider the function f ∈ V given by

(
1+x if x < 0,
f (x) = 1 − |x| =
1−x if x ≥ 0.

Find the closest approximation to f by a polynomial of degree at most 2. Graph both f and the
approximating polynomial.


Solution. Let W = span 1, x, x2 be the subspace of V consisting of polynomials of degree at most 2.
What we are looking for is an element g ∈ W such that k f − gk is as small as possible. We can solve this
problem using Proposition 11.38.
First we need an orthogonal basis for W . We found such an orthogonal basis in Example 11.35, namely
the Legendre polynomials p0 (x) = 1, p1 (x) = x, and p2 (x) = x2 − 31 . By Proposition 11.38, the desired
approximation g ∈ W is given by:
hp0 , f i hp1 , f i hp2 , f i
g = p0 + p1 + p2 .
hp0 , p0 i hp1 , p1 i hp2 , p2 i
We already computed the inner products hp0 , p0 i = 2, hp1 , p1 i = 23 , and hp2 , p2 i = 8
45 in Example 11.35.
We calculate the remaining inner products:
Z 1 Z 0 Z 1
1 1
hp0 , f i = 1 · f (x) dx = 1 · (1 + x) dx + 1 · (1 − x) dx =
+ = 1,
−1 −1 0 2 2
Z 1 Z 0 Z 1
1 1
hp1 , f i = x · f (x) dx = x · (1 + x) dx + x · (1 − x) dx = − + = 0,
−1 −1 0 6 6
Z 1
1
hp2 , f i = (x2 − ) · f (x) dx
−1 3
Z 0 Z 1
1 1 1 1 1
= (x − ) · (1 + x) dx + (x2 − ) · (1 − x) dx = − −
2
= − .
−1 3 0 3 12 12 6
Therefore,
hp0 , f i hp1 , f i hp2 , f i
g = p0 + p1 + p2
hp0 , p0 i hp1 , p1 i hp2 , p2 i
1 0 1/6
= p0 + p1 − p2
2 2/3 8/45
1 15 2 1
= − (x − )
2 16 3
13 15 2
= − x .
16 16
The following graph shows the function f (x) as well as the polynomial g(x):
 11.4. Orthogonal projections and Fourier series 417

f (x) = 1 − |x|
13
g(x) = 16 − 15
16 x
2

x
−1 1

♠
When approximating a function f by a polynomial, as in the last example, there is no need to stop with
polynomials of degree 2. We can also ask what is the best approximation of f by a polynomial of degree
3, of degree 4, of degree 5, and so on. By increasing the degree of the polynomials, we get better and
better approximations to f . This leads us to the concept of a generalized Fourier series.

Definition 11.41: Generalized Fourier series

Let V be an inner product space and let {u1 , u2 , u3 , . . .} be an infinite orthogonal set of vectors. Let
v ∈ V . The generalized Fourier series of v (with respect to u1 , u2 , u3 , . . .) consists of the following
sequence of vectors:

hu1 , vi
v1 = u1 ,
hu1 , u1 i
hu1 , vi hu2 , vi
v2 = u1 + u2 ,
hu1 , u1 i hu2 , u2 i
hu1 , vi hu2 , vi hu3 , vi
v3 = u1 + u2 + u3 ,
hu1 , u1 i hu2 , u2 i hu3 , u3 i
hu1 , vi hu2 , vi hu3 , vi hu4 , vi
v4 = u1 + u2 + u3 + u4 ,
hu1 , u1 i hu2 , u2 i hu3 , u3 i hu4 , u4 i
.
..

The vectors vi are also called generalized Fourier approximations of v.

By Proposition 11.38, we know that each vi is the best approximation of v in the subspace span {u1 , . . . , ui }.
In particular, kv − vi+1 k ≤ kv − vi k, so each vi is potentially a better approximation of v than the previous
one. In a course on analysis1 , you will learn that in many situations, the sequence v1 , v2 , v3 , . . . can be
shown to converge to v.

Example 11.42: Generalized Fourier series

Proceeding as in Example 11.40, find the generalized Fourier approximations of f (x) = 1 − |x| up
to degree 8.
1 “Algebra” is the name for those subjects where all sums are finite. This includes, for example, linear algebra and abstract
algebra. “Analysis” is the name for those subjects where sums are potentially infinite (and may or may not converge). This
includes, for example, calculus, complex analysis, and functional analysis.
418 Inner product spaces

Solution. We use the Legendre polynomials p0 , . . . , p8 from Section 11.3 and calculate the relevant inner
products, each of which requires solving an integral. As these integrals get a bit complicated, it is best to
use a computer algebra system to compute them.

hp0 , f i = 1, hp0 , p0 i = 2,
2
hp1 , f i = 0, hp1 , p1 i = 3,
hp2 , f i = − 16 , hp2 , p2 i = 8
45 ,
8
hp3 , f i = 0, hp3 , p3 i = 175 ,
1 128
hp4 , f i = 105 , hp4 , p4 i = 11025 ,
128
hp5 , f i = 0, hp5 , p5 i = 43659 ,
1 512
hp6 , f i = − 924 , hp6 , p6 i = 693693 ,
512
hp7 , f i = 0, hp7 , p7 i = 2760615 ,
1 32768
hp8 , f i = 6435 , hp8 , p8 i = 703956825 .

We therefore have the following approximations:

1
f0 = f1 = p0 ,
2
1 15
f2 = f3 = p0 − p2 ,
2 16
1 15 105
f4 = f5 = p0 − p2 + p4 ,
2 16 128
1 15 105 3003
f6 = f7 = p0 − p2 + p4 − p6 ,
2 16 128 2048
1 15 105 3003 109395
f8 = f9 = p0 − p2 + p4 − p6 + p8 .
2 16 128 2048 32768
The following graph shows the function f as well as its approximations f0 , f2 , f4 , f6 , and f8 . It can be
seen that each successive approximation is closer to the function f than the previous one.

f0

f4
f8
x
f6
−1 1 f
2

♠
 11.4. Orthogonal projections and Fourier series 419

Example 11.43: Generalized Fourier series

The following graph shows the generalized Fourier approximations of the function
(
0 if x < 0,
f (x) =
1 if x ≥ 0

by polynomials up to degree 7. (Note: the function f is not continuous, so not technically an

element of C[−1, 1], but we ignore this here. Instead of the vector space of continuous functions,
we can work in the vector space of piecewise continuous functions).

f1
f f5
f7
f3

f0

x
−1 1

Comparing the last two examples, we see that the function in Example 11.43 is much harder to ap-
proximate well by polynomials. This is due to the discontinuity in the latter function. Nevertheless, the
sequence of approximations eventually converges to f .

Example 11.44: Generalized Fourier series

The following graph shows the generalized Fourier approximations of the function


0 if x < − 21 ,
f (x) = x + 12 if − 12 ≤ x < 12 ,


1 if x ≥ 21

by polynomials up to degree 7.
420 Inner product spaces

f1
f f7
f5
f3

f0

x
−1 1

So far, we have worked with Legendre polynomials, but there are of course other examples of or-
thogonal sets of functions. An important such set is given by sine and cosine waves. We will see that
every periodic function can be decomposed into sine and cosine waves of varying frequencies. This was
Fourier’s original discovery, and the corresponding series are just known as Fourier series (i.e., not “gen-
eralized”).

Example 11.45: Fourier series

Consider the inner product space C[0, 2π ]. Recall from Example 11.21 that the following functions
form an orthogonal set:

u0 = 1
u1 = sin x
u2 = cos x
u3 = sin 2x
u4 = cos 2x
u5 = sin 3x
u6 = cos 3x
...

Consider the function f (x) = x − π, where x ∈ [0, 2π ]. Find its Fourier series.

Solution. Following Definition 11.41, we must calculate a number of inner products. We have hu0 , u0 i =
2π . Also, for all i ≥ 1, we have hui , ui i = π . We note the following antiderivatives, for k ≥ 1:
Z Z
x 1 x 1
x sin kx dx = − cos kx + 2 sin kx, x cos kx dx = sin kx + 2 cos kx.
k k k k
Using these formulas, we can compute the following inner products quite easily:
Z 2π
hu0 , f i = (x − π ) = 0,
0
 11.4. Orthogonal projections and Fourier series 421

Z 2π
hu1 , f i = (x − π ) sin x = − 2π ,
0
Z 2π
hu2 , f i = (x − π ) cos x = 0,
0
Z 2π
2π
hu3 , f i = (x − π ) sin 2x = − ,
0 2
Z 2π
hu4 , f i = (x − π ) cos 2x = 0,
0
Z 2π
2π
hu5 , f i = (x − π ) sin 3x = − ,
0 3
Z 2π
hu6 , f i = (x − π ) cos 3x = 0,
0

and so on. We therefore find the following Fourier series:

f1 = −2 sin x,
2
f3 = −2 sin x − sin 2x,
2
2 2
f5 = −2 sin x − sin 2x − sin 3x,
2 3
2 2 2
f7 = −2 sin x − sin 2x − sin 3x − sin 4x,
2 3 4
2 2 2 2
f9 = −2 sin x − sin 2x − sin 3x − sin 4x − sin 5x,
2 3 4 5
and so on. ♠
The following graph illustrates the successive approximations of this Fourier series.

f7 f
f5
f3
f1

x
π 2π 3π 4π 5π 6π

Note that, although the function f is defined on the interval [0, 2π ], we have extended it periodically for
[2π , 4π ], [4π , 6π ], and so on. This makes sense because all of the orthogonal functions 1, sin x, cos x,
sin 2x, and so on, have period 2π .
We can also illustrate the same information differently, by showing the individual sine waves making
up the wave form of the function f (x). In the context of audio signals, these sine waves are also called the
harmonics of the signal.
422 Inner product spaces





 − 21 sin(x)







 + − 22 sin(2x)






= + − 23 sin(3x)







 + − 24 sin(4x)









 + ···

Example 11.46: Fourier series

The following graph shows the first few Fourier approximations of the function
(
x− π if 0 ≤ x < π,
f (x) = 3π 2
2 − x if π ≤ x ≤ 2π.

f
f8
f6
f4
f2
x
π 2π 3π 4π 5π 6π

Note how rapidly this Fourier series converges to f . The following image shows the function f (extended
periodically outside the interval [0, 2π ]) as a sum of its harmonics. Compared to Example 11.45, we see
that the higher harmonics have much smaller amplitudes, which explains the rapid convergence.





 − π4 cos(x)







 + − 94π cos(3x)






= + − 254π cos(5x)







 + − 494π cos(7x)









 + ···
 11.4. Orthogonal projections and Fourier series 423

Example 11.47: Fourier series

The following graph shows the first few Fourier approximations of the function
(
1 if 0 ≤ x < π,
f (x) =
−1 if π ≤ x ≤ 2π.

f1 f3 f5 f7
f
x
π 2π 3π 4π 5π 6π

Once again, here is an image showing the function f as a sum of its harmonics:



 4

 π sin(x)







 + 4
sin(3x)

 3π




= + 4
5π sin(5x)







 + 4
sin(7x)

 7π







 + ···

Watch the video at http://y2u.be/3IAMpH4xF9Q for a demonstration of what the functions from Exam-
ples 11.45–11.47, and their harmonics, sound like as audio signals.

Exercises

Exercise 11.4.1 Consider R3 with the usual dot product. Let

     
−1 −1 −1
u1 =  1  , u2 =  −2  , and v =  5 .
1 1 3

Note that u1 and u2 are orthogonal. Find the best approximation of v in span {u1 , u2 }.
424 Inner product spaces

Exercise 11.4.2 Consider R4 with the usual dot product. Let

      
0 1 1 6
 0   1   −1   −2 
u1 =    
 1  , u2 =  0  , u3 =  3
 ,
 and v =  
 −5  .
3 0 −1 5

Note that u1 , u2 , and u3 are orthogonal. Find the best approximation of v in span {u1 , u2 , u3 }.

Exercise 11.4.3 In the inner product space V = C[−1, 1], consider the function f ∈ V given by
(
1 if x < 0,
f (x) =
1 − x if x ≥ 0.

Find the closest approximation to f by a polynomial of degree at most 0, 1, 2, 3, and 4. Graph both f and
the approximating polynomials.

Exercise 11.4.4 In the inner product space C[−π , π ], consider the orthogonal set of functions from Ex-
ample 11.21: {1, sin x, cos x, sin 2x, cos 2x, sin 3x, cos 3x, . . .}. Let f (x) = x2 , where x ∈ [−π , π ]. Find the
Fourier series of f .

11.5 Application: Least squares approximations and curve

fitting

Outcomes
A. Find least squares approximations for a system of equations.

B. Find best fit lines and parabolas for a set of data points.

In this section, we will consider the problem of finding approximate solutions to a system of linear equa-
tions Av = b. This can be useful when the system is inconsistent, but we would still like to find a “best”
answer. For example, consider the following system of equations.

0.1x + 0.2y = 0.3,

0.2x + 0.5y = 0.7,
0.7x + 0.2y = 0.9.

This system has the solution (x, y) = (1, 1), so it is clearly consistent. Now imagine that we introduce
some small inaccuracies into the equations. The inaccuracies might perhaps be due to round-off errors, or
 11.5. Application: Least squares approximations and curve fitting 425

due to measurement errors if the coefficients are obtained from experimental data. We might end up with
the following system of equations:

0.1000001x + 0.2y = 0.3,

0.2x + 0.5y = 0.7,
0.7x + 0.2y = 0.9.

Except for a tiny error in one of the coefficients, this is the same system of equations as before. We would
expect that such a small error does not affect the result much. However, this last system of equations
is inconsistent; it has no solutions at all. You can see this by observing that (x, y) = (1, 1) is still the
unique solution to the last two equations; substituting this into the first equation, we get 0.3000001 = 0.3,
which almost, but not exactly, true. What we would like to find in a situation like this is an “approximate
solution”, i.e., numbers x and y such that each of the three equations “almost” holds. We can formulate
the problem more precisely as follows:

Problem 11.48: Least squares approximation problem

Given a (possibly inconsistent) system of equations Av = b, find v such that

kAv − bk

is as small as possible. We call such a vector v a least squares approximation for the system of
equations.

To see why this is called a “least squares” approximation, consider a system of equations

a11 x1 + · · · + a1n xn = b1 ,
a21 x1 + · · · + a2n xn = b2 ,
···
am1 x1 + · · · + amn xn = bm .

We can write this in matrix form Av = b, where

     
a11 · · · a1n b1 x1
   ..  ,  
A =  ... . . . ...  , b =  .  and v =  ...  .
am1 · · · amn bm xn

Then  
a11 x1 + · · · + a1n xn − b1
Av − b =  ··· ,
am1 x1 + · · · + amn xn − bm
and therefore

kAv − bk2 = (a11 x1 + · · · + a1n xn − b1 )2 + . . . + (am1 x1 + · · · + amn xn − bm )2 .

Therefore, minimizing kAv − bk is the same as minimizing the sum of the squares of the errors of all the
equations, where the error of each equation is defined to be the difference between its left-hand side and
right-hand side.
426 Inner product spaces

We note that kAv − bk = 0 if and only if Av = b. Therefore, if the system of equations Av = b is

consistent, then its least squares approximations are exactly the solutions of the system of equations in the
usual sense.
The least squares approximation problem has a very elegant solution, provided by the following propo-
sition.

Proposition 11.49: Solution of the least squares approximation problem

A vector v is a least squares approximation of the system of equations Av = b if and only if

AT Av = AT b.

Proof. Let a1 , . . . an be the columns of the matrix A. Recall that span {a1 , . . . an } is called the column space
of A, which we write as col(A). If  
x1
 
v =  ...  ,
xn
is any vector, then by the definition of matrix multiplication, we have

Av = x1 a1 + . . . + xn an .

Therefore, a vector is of the form Av if and only if it is an element of col(A). In particular, the equation
Av = b has a solution if and only if b is an element of the column space of A. For v to be a least squares
approximation, we want kAv − bk to be as small as possible. This means that we are looking for the
element of col(A) that is closest to b.

b Av − b

Av
0 col(A)

From Proposition 11.38, we know that this happens when Av − b is orthogonal to col(A). Since col(A) =
span {a1 , . . . an }, this is equivalent to saying that Av − b is orthogonal to each of the vectors a1 , . . . an .
Therefore,
aTi (Av − b) = 0
for i = 1, . . . , n. Since aT1 , . . . aTn are the rows of the matrix AT , this system of n equations is equivalent to
the single equation
AT (Av − b) = 0,
or equivalently,
AT Av = AT b.
 11.5. Application: Least squares approximations and curve fitting 427

Example 11.50: Least squares approximation

Find the least squares approximation for the system of equations

2x + 2y + 2z = 1,
x − y − z = −2,
−x − y + 2z = 4,
2x + 2y − z = −8.

Solution. We first write the system in matrix form Av = b, where

   
2 2 2 1
 1 −1 −1   
A=  and b =  −2  .
 −1 −1 2   4 
2 2 −1 −8
By Proposition 11.49, the least squares approximation is given by the solution of the system of equations
AT Av = AT b. We calculate:
 
  2 2 2  
2 1 −1 2  10 8 −1
1 −1 −1 
AT A =  2 −1 −1 2  −1 −1
 =  8 10 1 ,
2 
2 −1 2 −1 −1 1 10
2 2 −1
 
  1  
2 1 −1 2   −20
−2  
AT b =  2 −1 −1 2  4  = −16 .

2 −1 2 −1 20
−8
Therefore, we must solve the system of equations
    
10 8 −1 x −20
 8 10 1   y  =  −16  .
−1 1 10 z 20
After some row operations, we find that the unique solution is (x, y, z) = (−1, −1, 2). We can double-check
this answer as follows. We calculate
         
2 2 2   1 0 1 −1
 1 −1 −1  −1        
Av − b =    −1  −  −2  =  −2  −  −2  =  0 
 −1 −1 2   4   6   4   2 
2
2 2 −1 −8 −6 −8 2
and check that this vector is orthogonal to every column of A. Since this is the case, our answer is correct.
♠
An important application of least squares approximations is curve fitting: finding the “best” function of a
given type (for example, linear, quadratic) to fit a given series of data points. The next two examples show
how to use least square approximations to solve curve fitting problems.
428 Inner product spaces

Example 11.51: Least squares line

A company has collected daily data on temperature and ice cream sales over a period of one week.
The data is as follows:
Date Peak temperature Sales
July 1 17 ◦ C $320
July 2 22 ◦ C $570
July 3 26 C◦ $850
July 4 20 C◦ $470
July 5 24 ◦ C $750
July 6 23 C◦ $620
July 7 23 C◦ $680
The company is interested in predicting how temperature will affect future sales. Find a function
of the form y = a + bx that best fits the data, where x is temperature in degrees Celsius and y is
sales in dollars. By “best fit”, we mean that the sum of the square of the errors should be as small
as possible (where each error is the difference between the dollar amount predicted by the formula
y = a + bx and the actual dollar amount). Such a function is called a least squares line or a linear
regression for the data.

Solution. We first write down a system of equations that expresses the relationship y = a + bx for each of
the seven data points (x, y):
a + 17b = 320,
a + 22b = 570,
a + 26b = 850,
a + 20b = 470,
a + 24b = 750,
a + 23b = 620,
a + 23b = 680.
This is a system of seven equations in two variables (the variables are a and b). This system is likely
inconsistent, because there are more equations than variables, and also because it is unlikely that the
relationship between temperature and sales is exactly (as opposed to approximately) linear. Instead, we
find the least squares approximation for the system of equations. We let
   
1 17 320
 1 22   570 
   
 1 26   850 
   
A=  
 1 20  and b =  470  .

 1 24   750 
   
 1 23   620 
1 23 680

By Proposition 11.49, we must solve the equation AT Av = AT b. We calculate

   
T 7 155 T 4260
A A= and A b = .
155 3483 97380
 11.5. Application: Least squares approximations and curve fitting 429

So the system of equations we must solve is

    
7 155 a 4260
= .
155 3483 b 97380
After some row operations, we find the unique solutions (a, b) = (−720, 60). This means that the desired
linear approximation is y = −720 + 60x. The following plot shows this function along with the original
data points.
y (sales)
1000

800

600

400

200 x (temperature)
15 20 25

The following table compares the observed data to the computed linear regression (sorted by increasing
temperature). It also shows the error for each data point.
Temperature Actual sales Best linear fit Error
17 320 300 −20
20 470 480 +10
22 570 600 +30
23 620 660 +40
23 680 660 −20
24 750 720 −30
26 850 840 −10
The sum of the squares of the errors is 202 + 102 + 302 + 402 + 202 + 302 + 102 = 4400. ♠

Example 11.52: Least squares parabola

Find a quadratic polynomial that is the best fit for the following data points:

(x1 , y1 ) = (3, 23.5),

(x2 , y2 ) = (4, 13.5),
(x3 , y3 ) = (5, 12.5),
(x4 , y4 ) = (6, 5.5),
(x5 , y5 ) = (7, 9.0),
(x6 , y6 ) = (8, 8.0),
(x7 , y7 ) = (9, 19.0).
430 Inner product spaces

Solution. We are looking for an equation of the form y = a + bx + cx2 . Substituting each of the seven data
points into the equation, we obtain 7 equations in the unknowns a, b, and c:
a + 3b + 9c = 23.5,
a + 4b + 16c = 13.5,
a + 5b + 25c = 12.5,
a + 6b + 36c = 5.5,
a + 7b + 49c = 9.0,
a + 8b + 64c = 8.0,
a + 9b + 81c = 19.0.
We write this in matrix form at Av = b, where
   
1 3 9 23.5
 1 4 16   13.5 
   
 1 5 25   12.5 
   
A=  1 6 36

 and b = 
 5.5 .

 1 7 49   9.0 
   
 1 8 64   8.0 
1 9 81 19.0
To find the least squares approximation, we calculate
   
7 42 280 91
AT A =  42 280 2016  and AT b =  518 
280 2016 15316 3430
and solve the system of equations AT Av = AT b, i.e.,
    
7 42 280 a 91
 42 280 2016   b  =  518  .
280 2016 15316 c 3430
After doing some row operations, we find that the unique solution is (a, b, c) = (67, −19, 1.5). Therefore,
the desired quadratic approximation is y = 67 − 19x + 1.5x2 . The following plot shows this function along
with the original data points.
y
25

20

15

10

0 x
0 5 10
 11.5. Application: Least squares approximations and curve fitting 431

The following table shows the original data, the values of the best fit parabola, and the error for each data
point.
x Actual y Best fit parabola Error
3 23.5 23.5 0.0
4 13.5 15.0 +1.5
5 12.5 9.5 −3.0
6 5.5 7.0 +1.5
7 9.0 7.5 −1.5
8 8.0 11.0 +3.0
9 19.0 17.5 −1.5
The sum of the squares of the errors is 02 + 1.52 + 32 + 1.52 + 1.52 + 32 + 1.52 = 27. ♠

Exercises

Exercise 11.5.1 Find the least squares approximation for the system of equations

x + 2y + 2z = 5,
x + y − z = 11,
x + 2y − z = −18,
2x − y + 2z = 0.

Exercise 11.5.2 Find the least squares approximation for the system of equations
   
−1 2 1   1
 −1 0 −1  x  1 
   
 2 0 2  y  =  3 .
   
 0 0 2  z  −2 
1 2 2 4

Exercise 11.5.3 Consider the points (x1 , y1 ) = (−1, 0), (x2 , y2 ) = (0, 3), (x3 , y3 ) = (1, 3), (x4 , y4 ) = (2, 5),
(x5 , y5 ) = (3, 9). Find the least squares line for these points.

Exercise 11.5.4 Consider the points (x1 , y1 ) = (−1, 4), (x2 , y2 ) = (0, −2), (x3 , y3 ) = (1, 4), (x4 , y4 ) =
(2, 2). Find the least squares parabola for these points.
432 Inner product spaces

11.6 Orthogonal functions and orthogonal matrices

Outcomes
A. Determine whether a linear transformation is an isometry and/or orthogonal.

B. Determine whether a matrix is orthogonal.

Definition 11.53: Isometries and orthogonal maps

Let V ,W be real inner product spaces, and let T : V → W be a linear transformation. We say that T
is an isometry if for all u, v ∈ V ,
hT (u), T (v)i = hu, vi.
Moreover, T is called an orthogonal transformation, or simply orthogonal, if it is an isometry
and invertible.

p p
In particular, if T is an isometry, we have kT vk = hT v, T vi = hv, vi = kvk, so isometries preserve
the norm. (This is fact where the name “isometry” comes from: from Greek “isos”, meaning “equal”, and
“metron”, meaning “measure”). Conversely, any norm-preserving linear function is an isometry, as the
following proposition shows.

Proposition 11.54: Norm-preserving linear maps are isometries

Let T : V → W be a linear function such that for all v ∈ V , kT vk = kvk. Then T is an isometry.

Proof. We first note that for all u, v ∈ V ,

1 1 1
hu, vi = (hu, vi + hv, ui) = (hu + v, u + vi − hu, ui − hv, vi) = (ku + vk2 − kuk2 − kvk2 ).
2 2 2
Now suppose T is norm-preserving and linear. Then for all u, v ∈ V ,
1
hT (u), T (v)i = (kT (u) + T (v)k2 − kT (u)k2 − kT (v)k2 )
2
1
= (kT (u + v)k2 − kT (u)k2 − kT (v)k2 )
2
1
= (ku + vk2 − kuk2 − v2 )
2
= hu, vi.

Therefore, T is an isometry. ♠
We also note that if T is an isometry and u ⊥ v, then T (u) ⊥ T (v), because hT (u), T (v)i = hu, vi = 0.
So isometries preserve right angles. In fact, isometries also preserve arbitrary angles, due to the formula
hu,vi hT (u),T (v)i
cos θ = kuk kvk = kT (u)k kT (v)k .
 11.6. Orthogonal functions and orthogonal matrices 433

Example 11.55: Orthogonal transformations on Rn

Let T1 , T2 , T3 : R2 → R2 be the linear transformations with matrices
     
0 −1 2 0 1 1
[T1 ] = , [T2 ] = , and [T3 ] = ,
1 0 0 2 0 1

respectively. T1 is a counterclockwise rotation by 90◦ , T2 is a scaling by a factor of 2, and T3 is a

shearing. Then T1 is orthogonal, but T2 and T3 are not. Note that T1 , being a rotation, preserves
lengths and angles. T2 preserves angles but not lengths. T3 preserves neither lengths nor angles.

T1 T2
7−→ 7−→

orthogonal not orthogonal

T3
7−→

not orthogonal

Proposition 11.56: The matrix of an orthogonal transformation

Let V ,W be finite-dimensional real inner product spaces and let T : V → W be a linear transforma-
tion. Let B and C be orthonormal bases of V and W , respectively, and let P = [T ]C,B be the matrix
of T with respect to the bases B and C. Then

(a) T is an isometry if and only if PT P = I .

(b) T is orthogonal if and only if PT P = I and dimV = dimW .

·
Proof. (a) Recall from Proposition 11.28 that for all u, v ∈ V , we have hu, vi = [u]B [v]B . Consider two
vectors u, v ∈ V and let t = [u]B and s = [v]B . We have:

hT (u), T (v)i = hu, vi ⇐⇒ · ·

[T (u)]C [T (v)]C = [u]B [v]B
⇐⇒ · ·
([T ]C,B [u]B ) ([T ]C,B [v]B ) = [u]B [v]B
⇐⇒ ·
(Pt) (Ps) = t s ·
⇐⇒ tT PT Ps = tT s
⇐⇒ tT PT Ps = tT Is.
434 Inner product spaces

Therefore, T is an isometry if and only if tT PT Ps = tT Is holds for all t, s ∈ Rm . If PT P = I, then this is

clearly true. Conversely, assume tT PT Ps = tT Is holds for all t, s ∈ Rm . Then by taking t and s to be the i th
and j th standard basis vector, it follows that the (i, j)-entry of PT P is equal to the (i, j)-entry of I, for all i
and j. Therefore, PT P = I.
(b) By definition, T is orthogonal if and only if it is an isometry and invertible. Therefore, if T is
orthogonal, then PT P = I by part (a) and dimV = dimW because every invertible matrix is square (see
Theorem 4.48). Conversely, assume that PT P = I and dimV = dimW . Then T is an isometry by part (a).
Also, by Theorem 4.64, since P is square and PT is a left inverse of P, the matrix P, and therefore the
linear transformation T , is invertible. Therefore, T is orthogonal. ♠

Example 11.57: Orthogonal transformations

Which of the following matrices define orthogonal transformations? Which ones define isometries?
 
      1 −2
0 1 1 1 1 1 1 1
A= , B= , C= √ , D =  2 −1  .
−1 0 1 0 2 1 −1 3
2 2

Solution. We have     
T 0 1 0 −1 1 0
A A= = = I,
−1 0 1 0 0 1
and therefore A defines an isometry and (since it is also a square matrix) an orthogonal transformation.
We have     
T 1 1 1 1 2 1
B B= = 6= I,
1 0 1 0 1 1
so the linear transformation defined by B is neither an isometry nor orthogonal. A similar pair of calcula-
tions shows that CT C = I and DT D = I, so both C and D are isometries. Since C is a square matrix, it is
also orthogonal. D is not a square matrix, and therefore not orthogonal. ♠

In light of Proposition 11.56, we say that an n × n-matrix P is orthogonal if PT P = I.

Definition 11.58: Orthogonal matrix

An n × n-matrix P is called orthogonal if PT P = I .

The following proposition gives several equivalent ways of checking whether an n × n-matrix P is orthog-
onal.
 11.6. Orthogonal functions and orthogonal matrices 435

Proposition 11.59: Conditions for orthogonal matrices

The following are equivalent for an n × n-matrix P:

(a) P is orthogonal.

(b) PT P = I .

(c) P is invertible and P−1 = PT .

(d) PPT = I .

(e) PT is orthogonal.

(f) The columns of P form an orthonormal set of vectors.

(g) The rows of P form an orthonormal set of vectors.

Proof. The equivalence (a) ⇔ (b) is just the definition of orthogonality, bearing in mind that P is assumed
to be a square matrix. The equivalences (b) ⇔ (c) ⇔ (d) follow from Theorem 4.64, since a square matrix
is invertible if and only if it is left invertible if and only if it is right invertible. The equivalence (d) ⇔
(e) is again just the definition of orthogonality, this time applied to PT . To show (b) ⇔ (f), let a1 , . . . , an
be the columns of P, and let e1 , . . . , en be the standard basis vectors of Rn . Then we have PT P = I if and
only if for all i, j, the (i, j)-entry of PT P is equal to the (i, j)-entry of I. But the (i, j)-entry of PT P is
·
eTi PT Pe j = aTi a j = ai a j . On the other hand, the (i, j)-entry of I is 1 if i = j and 0 if i 6= j. It follows that
PT P = I if and only if for all i, j, (
1 when i = j and
·
ai a j =
0 otherwise.
But this is exactly what it means for a1 , . . . , an to form an orthonormal set. The proof of (d) ⇔ (g) is
completely analogous, using rows instead of columns. ♠

Example 11.60: Orthogonal matrices

Determine which of the following matrices are orthogonal by checking whether the columns form
an orthonormal set.
       
0 −1 0 2 −1 2 3 4 0 2 1 2
1 1
A= 1 0 0  , B =  −1 2 2  , C =  4 −3 0  , D =  1 2 2  .
3 3
0 0 1 2 2 −1 0 0 5 2 2 1

Solution. By Proposition 11.59, it suffices to check whether the columns of each matrix are orthonormal.
This is the case for A and B. For example, the columns of B are
     
2 −1 2
1 1 1
b1 =  −1  , b2 =  2  , and b3 =  2  ,
3 3 3
2 2 −1
436 Inner product spaces

and we have
       
2 2 2 −1
1 1
b1 · b1 =  −1 
9
2
·  −1  = 1,
2
b1 · b2 =  −1 
9
2
·  2  = 0,
2
       
−1 −1 2 2
1 1
b2 · b2 =  2 
9
2
·  2  = 1,
2
b1 · b3 =  −1 
9
2
·  2  = 0,
−1
       
2 2 −1 2
1 1
b3 · b3 =  2 
9
−1
·  2  = 1,
−1
b2 · b3 =  2 
9
2
·  2  = 0.
−1
So the columns of B are orthonormal. The columns of C are orthogonal, but not normalized, so C is not an
orthogonal matrix. The columns of D are normalized, but not orthogonal, so D is not an orthogonal matrix
either. ♠

Proposition 11.61: Properties of orthogonal matrices

(a) If P, Q are orthogonal n × n-matrices, then PQ is orthogonal.

(b) The identity matrix is orthogonal.

(c) If P is orthogonal, then so is P−1 .

(d) If P is an orthogonal n × n-matrix and Q is an orthogonal m × m-matrix, then the (n + m) ×

(n + m)-matrix  
P 0
0 Q
is orthogonal.

Exercises

Exercise 11.6.1 Which of the following linear transformations T : R3 → R3 are orthogonal?

(a) A rotation by 30 degrees about the z-axis.

(b) A reflection about the plane x = y.

(c) A scaling by a factor of 2.

(d) A scaling by a factor of −1.

 11.7. Diagonalization of symmetric matrices 437

Exercise 11.6.2 Which of the following matrices define orthogonal transformations? Which ones define
isometries?
 
      1 2
1 0 1 1 1 1 1 −2 1
A= , B= , C= √ , D=  2 2 .
0 −1 2 −1 1 5 2 1 4
2 −1

Exercise 11.6.3 Determine which of the following matrices are orthogonal by checking whether the
columns form an orthonormal set.
       
1 2 2 0 0 2 2 −1 2 3 4 0
1 1 1
A =  2 1 −2  , B =  0 2 0 , C =  −1 2 −2  , D =  4 −3 0  .
3 3 5
−2 2 −1 2 0 0 2 −2 1 0 0 5

11.7 Diagonalization of symmetric matrices

Outcomes
A. Compute the real eigenvalues of a symmetric matrix.

B. Compute an orthogonal basis of eigenvectors for a symmetric matrix.

C. Orthogonally diagonalize a symmetric matrix.

In Chapter 8, we saw that some matrices are diagonalizable, and others are not. In this section, we will see
that the theory of diagonalization is much nicer when the matrix to be diagonalized is symmetric. Recall
that a matrix A is symmetric if A = AT . We begin with some observations about the eigenvalues and
eigenvectors of a symmetric matrix.

Proposition 11.62: Eigenvalues and eigenvectors of symmetric matrices

Let A be a symmetric matrix with real entries. Then

(a) All eigenvalues of A are real (i.e., not complex).

(b) Eigenvectors for distinct eigenvalues of A are orthogonal.

Proof. (a) Suppose λ is a (possibly complex) eigenvalue of A, with (possibly complex) eigenvector v. We
will show that λ is in fact real. Let λ be the complex conjugate of λ , and let v be the complex conjugate
of v, i.e., the vector obtained from v by taking the complex conjugate of each of its entries. Since v is an
eigenvector for eigenvalue λ of A, we have

vT Av = λ vT v. (11.2)
438 Inner product spaces

Taking the complex conjugate of both sides of the equation, and using the fact that A = A (since A is real),
we get
vT Av = λ vT v. (11.3)
Then, taking the transpose of both sides of the equation, and using the fact that AT = A (since A is sym-
metric), we have
vT Av = λ vT v. (11.4)
Comparing equations (11.2) and (11.4), we find that λ vT v = λ vT v. Since vT v is a non-zero scalar, it
follows that λ = λ , i.e., λ is real.
(b) Suppose v is an eigenvector for eigenvalue λ , w is an eigenvector for eigenvalue µ , and λ 6= µ . We
evaluate vT Aw in two different ways. On the on hand, we have

vT Aw = vT (Aw) = vT (µ w) = µ vT w.

On the other hand, we have

vT Aw = (vT A)w = (AT v)T w = (Av)T w = (λ v)T w = λ vT w.

Therefore, µ vT w = λ vT w, or equivalently (λ − µ )vT w = 0. Since by assumption, λ − µ 6= 0, we must

have vT w = 0, i.e., v ⊥ w. ♠

Recall that an n × n-matrix A is diagonalizable if there exists an invertible matrix P and a diagonal matrix
D such that D = P−1 AP. We say that A is orthogonally diagonalizable if P can, moreover, be chosen
to be orthogonal. Orthogonal diagonalizability is a convenient property, because when P is orthogonal,
then P−1 = PT , and we can interchangeably write D = P−1 AP or D = PT AP. The following is the main
theorem about the diagonalization of real symmetric matrices.

Theorem 11.63: Diagonalization of symmetric matrices

Every real symmetric matrix A is orthogonally diagonalizable.

Proof. By induction on the size of the matrix. For a 1 × 1-matrix, there is nothing to show, as it is already
diagonal. Now consider a real symmetric n × n-matrix A with n ≥ 2. By the fundamental theorem of
algebra, the characteristic polynomial has at least one root, so that A has at least one (possibly complex)
eigenvalue λ . By Proposition 11.62(a), λ is real. Since we can solve the system of equations (A− λ I)v = 0
over the real numbers, there exists a real eigenvector v for the eigenvalue λ . We can assume without loss of
generality that v is normalized, because else we could replace v by k1vk v. By the Gram-Schmidt method, we
can find an orthonormal basis {u1 , . . . , un } of Rn such that u1 = v. Let Q be the orthogonal matrix that has
u1 , . . . , un as its columns, and consider B = Q−1 AQ. Since Be1 = Q−1 AQe1 = Q−1 Au1 = λ Q−1 u1 = λ e1 ,
the matrix B is of the form  
λ b12 · · · b1n
 0 b22 · · · b2n 
 
B =  .. .. . . ..  .
 . . . . 
0 bn2 · · · bnn
 11.7. Diagonalization of symmetric matrices 439

Moreover, since B = Q−1 AQ = QT AQ, the matrix B is symmetric. Therefore b12 , . . . , b1n = 0, and B is of
the form  
λ 0 ··· 0
 0 
 
B =  .. ,
 . C 
0
where C is a symmetric matrix of dimension n − 1. By induction hypothesis, C is orthogonally diagonal-
izable, i.e., there exists an orthogonal matrix R such that R−1CR = D is diagonal. Let
 
1 0 ··· 0
 0 
 
S =  .. .
 . R 
0

Then S is orthogonal by Proposition 11.61(4). Let E = S−1 BS. Then

   
λ 0 ··· 0 λ 0 ··· 0
 0   0 
−1    
E = S BS =  .. −1
 =  .. .
 . R CR   . D 
0 0

Therefore, E is diagonal. Let P = QS. Then P is orthogonal by Proposition 11.61(1). Moreover, we have

P−1 AP = S−1 Q−1 AQS = S−1 BS = E.

Therefore, A is orthogonally diagonalizable. ♠

Example 11.64: Diagonalization of symmetric matrices

Orthogonally diagonalize the matrix  
3 2
A= ,
2 6
i.e., find an orthogonal matrix P and a diagonal matrix D such that D = P−1 AP.

Solution. We proceed in much the same way as in Chapter 8, except that at a crucial moment, we ensure
that P is orthogonal. We start by calculating the characteristic polynomial of A:
 
 3−λ 2 
det(A − λ I) =  = (3 − λ )(6 − λ ) − 4 = λ 2 − 9λ + 14.
2 6−λ 

We find the roots using the quadratic formula. The roots of the characteristic polynomial, and therefore
the eigenvalues of A, are λ = 7 and λ = 2. For the eigenvalue λ = 7, we find the eigenvector
 
1
v= ,
2
440 Inner product spaces

and for the eigenvalue λ = 2, we find the eigenvector

 
−2
w= .
1
We note that these two eigenvectors are orthogonal to each other, exactly as predicted by Proposition 11.62.
So the vectors {v, w} form an orthogonal set. We turn this into an orthonormal set by normalizing each
eigenvector, i.e., the normalized eigenvectors are
   
1 1 1 1 1 −2
u1 = v = √ and u2 = w = √ .
kvk 5 2 kwk 5 1
We let P be the matrix that has columns u1 and u2 , i.e.,
 
1 1 −2
P= √ .
5 2 1
Note that since u1 and u2 are orthonormal, P is automatically orthogonal by Proposition 11.59. Moreover,
we have  
−1 T 1 1 2
P = P = √ .
5 −2 1
Finally,      
−1 T 1 1 2 3 2 1 −2 7 0
P AP = P AP = = .
5 −2 1 2 6 2 1 0 2
♠

Example 11.65: Diagonalization of symmetric matrices

Find an orthogonal matrix P and a diagonal matrix D such D = P−1 AP, where
 
3 1 −2
A= 1 3 2 .
−2 2 0

Solution. Again, we start by calculating the characteristic polynomial and its roots:
 
 3−λ 1 −2 
 
det(A − λ I) =  1 3 − λ 2  = − λ 3 + 6λ 2 − 32.
 −2 2 −λ 
The roots are λ = 4 and λ = −2. To find the eigenvectors for λ = 4, we solve the system of equations
 
−1 1 −2
(A − 4I)v =  1 −1 2  v = 0.
−2 2 −4
The solution space is 2-dimensional, with basis
  
1 0
v1 = 1 
 and v2 =  2  .
0 1
 11.7. Diagonalization of symmetric matrices 441

To find the eigenvectors for λ = −2, we solve the system of equations

 
5 1 −2
(A + 2I)v =  1 5 2  v = 0.
−2 2 2

The solution space is 1-dimensional, with basis

 
1
v3 =  −1  .
2

As predicted by Proposition 11.62, v3 is orthogonal to v1 and v2 . However, there is a slight complication:

v1 and v2 are not orthogonal to each other. This is because they are eigenvectors for the same eigenvalue
(λ = 4), not for distinct eigenvalues. We found basis {v1 , v2 } for the eigenspace for λ = 4, but it doesn’t
happen to be an orthogonal basis. However, we can fix this by applying the Gram-Schmidt method to
{v1 , v2 }. This yields the orthogonal basis {u1 , u2 } for the eigenspace for λ = 4, where
   
1 −1
u1 =  1  and u2 =  1  .
0 1

We now have an orthogonal basis {u1 , u2 , v3 } of eigenvectors of the matrix A. We normalize the three
vectors and use them as the columns of P:
 
√1 − √ 1 √1
 2 3 6 
 √1 1 
P= √ − √1  .
 2 3 6 
0 √1 √2
3 6

It seems that inverting P will not be all that easy, but in fact, since P is orthogonal, the inverse is just the
transpose:  
√1 √1 0
 2 2 
−1 T  √1 1 √1

P = P =  − √ .
 3 3 3 
√1 − √1 √2
6 6 6
We have 
4 0 0
P−1 AP = D =  0 4 0  .
0 0 −2
♠
442 Inner product spaces

Exercises

Exercise 11.7.1 For each of the following symmetric matrices, find the eigenvalues, an orthonormal basis
for each eigenspace, and then orthogonally diagonalize the matrix.
   
    1 1 0 2 1 2
−1 2 3 1
(a) A = , (b) A = , (c) A =  1 2 1  , (d) A =  1 2 2  .
2 2 1 3
0 1 1 2 2 5

Exercise 11.7.2 Prove the converse of Theorem 11.63: if a matrix A is orthogonally diagonalizable, then
A is symmetric.

11.8 Positive semidefinite and positive definite matrices

Outcomes
A. Determine whether a matrix is positive semidefinite and/or positive definite, either directly or
by looking at the eigenvalues.

B. Determine whether a matrix is positive semidefinite and/or positive definite using Descartes’
rule of signs.

C. Determine whether a matrix defines an inner product.

In Example 11.3, we saw that it is sometimes possible to define an inner product on Rn from an n × n-
matrix A by the formula
hu, vi = uT Av.
We will now explore in more detail under what conditions this formula defines an inner product.

Definition 11.66: Positive semidefinite and positive definite matrices

Let A be an n × n-matrix over the real numbers.

• A is called symmetric if A = AT .

• A is called positive semidefinite if it is symmetric and for all v ∈ Rn , we have vT Av ≥ 0.

• A is called positive definite if it is positive semidefinite and vT Av = 0 if and only if v = 0.

Equivalently, the positive definite property can also be stated as follows: A is positive definite if it is
symmetric and for all v ∈ Rn with v 6= 0, we have vT Av > 0.
 11.8. Positive semidefinite and positive definite matrices 443

Example 11.67: Positive semidefinite and positive definite matrices

Which of the following matrices are positive semidefinite? Which ones are positive definite?
       
2 −1 1 2 1 −1 2 0
A= , B= , C= , D= .
−1 1 2 1 −1 1 2 1

 
x
Solution. The matrix A is positive definite. To see why, consider any vector v = . Then
y
  
  2 −1 x
T
v Av = x y = 2x2 − 2xy + y2 = x2 + (x − y)2 ≥ 0.
−1 1 y

This inequality implies that A is positive semidefinite. Moreover, vT Av = 0 if and only if x = 0 and
x − y = 0, which implies x = y = 0. So v = 0 is the only solution of vT Av = 0, and A is positive definite.
The matrix
 B isnot positive semidefinite (and therefore not positive definite either). For example,
1
consider v = . Then
−1
  
T
  1 2 1
v Bv = 1 −1 = 1 − 2 − 2 + 1 = − 2 < 0,
2 1 −1
showing that B is not positive semidefinite.
 
x
The matrix C is positive semidefinite, but not positive definite. To see why, consider v = . Then
y
  
  1 −1 x
T
v Cv = x y = x2 − 2xy + y2 = (x − y)2 ≥ 0.
−1 1 y
Therefore,
  C is positive semidefinite. However, it is not positive definite because, for example, for v =
1
, we have v 6= 0 but vT Cv = 0.
1
The matrix D is not symmetric, and therefore neither positive semidefinite nor positive definite. ♠
The interest of positive definite matrices lies in the following proposition.

Proposition 11.68: Positive definite matrices and inner products on Rn

Let A be an n × n-matrix over the real numbers. Then the formula

hu, vi = uT Av

defines an inner product on Rn if and only if A is positive definite. Moreover, every inner product
on Rn arises from some positive definite matrix A in this way.

Proof. Define hu, vi = uT Av. We will check each of the three properties of an inner product from Defini-
tion 11.1. Linearity holds for all matrices A, because
hu, kvi = uT A(kv) = k(uT Av) = khu, vi
444 Inner product spaces

and
hu, v + v′ i = uT A(v + v′ ) = uT Av + uT Av′ = hu, vi + hu, v′ i.
For symmetry, first observe that for all u, v,

hv, ui = vT Au = (vT Au)T = uT AT v.

Therefore hu, vi = hv, ui if and only if uT Av = uT AT v. This holds for all u, v if and only if A = AT .
Therefore, symmetry holds if and only if A is symmetric. Finally, the positive definite property of the
inner product holds, by definition, if and only if A is positive definite.
To prove the second part, consider any inner  product
 hu, vi on Rn . Let e1 , . . . , en be the standard basis
vectors, and define ai j = hei , e j i. Then A = ai j is a matrix. We claim that hu, vi = uT Av for all u, v.
Indeed, let u = u1 e1 + . . . + un en and v = v1 e1 + . . . + vn en . Then

hu, vi = hu1 e1 + . . . + un en , v1 e1 + . . . + vn en i
= u1 v1 he1 , e1 i + u1 v2 he1 , e2 i + . . . + un vn hen , en i
= u1 v1 a11 + u1 v2 a12 + . . . + un vn ann
= uT Av.

By the first part, A is positive definite. ♠

Given a symmetric matrix A, it is not always an easy task to determine whether A is positive definite (or
semidefinite) by using the definition directly. This would require checking the condition vT Av ≥ 0 for all
vectors v, of which there are infinitely many. The following proposition gives us a more practical method
for determining whether a matrix is positive (semi)definite.

Proposition 11.69: Characterization of positive (semi)definite matrices using eigenvalues

Let A be a symmetric n × n-matrix, and let λ1 , . . . , λn be its eigenvalues. Then A is positive semidef-
inite if and only if λ1 , . . . , λn ≥ 0. Moreover, A is positive definite if and only if λ1 , . . . , λn > 0.

Proof. By Theorem 11.63, we know that A is orthogonally diagonalizable, i.e., D = PT AP, where P is an
orthogonal matrix and  
λ1 · · · 0
 
D =  ... . . . ...  .
0 · · · λn

Let v = [x1 , . . . , xn ]T be any vector, and let w = Pv. Then wT Aw = vT PT APv = vT Dv, so that A is pos-
itive (semi)definite if and only if D is positive (semi)definite. Also, we have vT Dv = λ1 x21 + . . . + λn x2n .
Therefore vT Dv ≥ 0 for all v if and only if λ1 , . . . , λn ≥ 0. Moreover, vT Dv > 0 for all v 6= 0 if and only if
λ1 , . . . , λn > 0, as claimed. ♠
 11.8. Positive semidefinite and positive definite matrices 445

Example 11.70: Using eigenvalues to check whether a matrix is positive definite

Use eigenvalues to determine whether the matrix
 
5 0 2
A= 0 5 1 
2 1 1

is positive definite, positive semidefinite, or neither.

Solution. The characteristic polynomial is

 
 5−λ 0 2 
 

det(A − λ I) =  0 5−λ 1  = − λ 3 + 11λ 2 − 30λ = − λ (λ − 5)(λ − 6).

 2 1 1−λ 

The eigenvalues are λ1 = 0, λ2 = 5, and λ3 = 6. Therefore, the matrix A is positive semidefinite, but not
positive definite. ♠
While Proposition 11.69 gives us a method to determine whether a matrix A is positive definite (or semidef-
inite), it still requires finding all of the eigenvalues of A, i.e., to factor the characteristic polynomial. This
can be a difficult calculation, especially if the degree of the polynomial is large or if the roots are not inte-
gers. Fortunately, there is a better way to determine whether a matrix is positive definite (or semidefinite)
by directly looking at the characteristic polynomial, without having to calculate its roots. The method was
found by René Descartes in 1637 and is called Descartes’ rule of signs. The general form of Descartes’
rule of signs is actually more complicated than what we consider here. We state a version that has been
specialized to characteristic polynomials of symmetric matrices.
Let a0 , . . . , an be a sequence of real numbers. We say that a0 , . . . , an have strongly alternating signs
if a0 > 0, a1 < 0, a2 > 0, and so on (i.e., ai > 0 when i is even, and ai < 0 when i is odd). We say that
a0 , . . . , an have weakly alternating signs if a0 ≥ 0, a1 ≤ 0, a2 ≥ 0, and so on (i.e., ai ≥ 0 when i is even,
and ai ≤ 0 when i is odd).

Proposition 11.71: Descartes’ rule of signs for positive (semi)definite matrices

Let A be a symmetric n × n-matrix, and let

p(λ ) = an λ n + an−1 λ n−1 + . . . + a1 λ + a0

be its characteristic polynomial. Then:

• A is positive definite if and only if a0 , . . . , an have strongly alternating signs.

• A is positive semidefinite if and only if a0 , . . . , an have weakly alternating signs.

Proof. We first prove a general fact about polynomials. Suppose d1 , . . . , dn are real numbers and

(x + d1 )(x + d2 ) · · ·(x + dn ) = bn xn + bn−1 xn−1 + . . . + b1 x + b0 .

Then d1 , . . . , dn > 0 if and only if b0 , . . . , bn > 0. Moreover, d1 , . . . , dn ≥ 0 if and only if b0 , . . . , bn ≥ 0.

446 Inner product spaces

Proof: To prove the first claim, assume d1 , . . . , dn > 0. It is easy to see that by multiplying out (x +
d1 )(x + d2 ) · · · (x + dn ), we can only obtain positive coefficients, so b0 , . . . , bn > 0. Conversely, assume
b0 , . . . , bn > 0. Then for every x ≥ 0, we clearly have bn xn + bn−1 xn−1 + . . . + b1 x + b0 > 0, and therefore
no such x ≥ 0 can be a root of this polynomial. In other words, all of the roots must be negative. Since the
roots are −d1 , . . . , −dn , it follows that d1 , . . . , dn > 0. The proof of the second claim (using “≥” instead of
“>”) is similar.
We are now ready to prove Proposition 11.71. By Theorem 11.63, we know that A is diagonalizable,
i.e., A = PDP−1 for some real diagonal matrix D. Note that A and D have the same characteristic polyno-
mial. If d1 , . . . , dn are the diagonal entries of D, the characteristic polynomial can therefore be written in
two different ways:

p(λ ) = (d1 − λ )(d2 − λ ) · · · (dn − λ ) = a0 + a1 λ + a2 λ 2 + a3 λ 3 + . . . + an λ n .

Now let λ = −x and consider

p(−x) = (d1 + x)(d2 + x) · · · (dn + x) = a0 − a1 x + a2 x2 − a3 x3 + . . . ± an xn .

We have: A is positive definite if and only if d1 , . . . , dn > 0, if and only if a0 , −a1 , a2 , −a3 , . . . > 0, if and
only if a0 , . . . , an are strongly alternating. Moreover, A is positive semidefinite if and only if d1 , . . . , dn ≥ 0,
if and only if a0 , −a1 , a2 , −a3 , . . . ≥ 0, if and only if a0 , . . . , an are weakly alternating. ♠

Example 11.72: Using Descartes’ rule of signs to check if a matrix is positive definite
Use Descartes’ rule of signs to check whether the matrix
 
5 0 2
A= 0 5 1 
2 1 1

is positive definite, positive semidefinite, or neither.

Solution. We already found the characteristic polynomial in Example 11.70: it is −λ 3 + 11λ 2 − 30λ . The
coefficients are a0 = 0, a1 = −30, a2 = 11, and a3 = −1. Note that we have included all of the coefficients,
even ones that are zero. Since the signs are weakly alternating, but not strongly alternating, the matrix is
positive semidefinite, but not positive definite. ♠

Example 11.73: Using Descartes’ rule of signs to check if a matrix is positive definite
Use Descartes’ rule of signs to determine which of the following matrices are positive definite
and/or positive semidefinite.
     
2 1 2 2 1 −2 2 −1 2
A =  1 1 0 , B =  1 1 −1  , C =  −1 2 0  .
2 0 2 −2 −1 2 2 0 3
 11.8. Positive semidefinite and positive definite matrices 447

Solution. The characteristic polynomials are:

det(A − λ I) = −λ 3 + 5λ 2 − 3λ − 2,
det(B − λ I) = −λ 3 + 5λ 2 − 2λ + 0,
det(C − λ I) = −λ 3 + 7λ 2 − 11λ + 1.
For A, the coefficients are not weakly alternating, so A is not positive semidefinite. For B, the coefficients
are weakly, but not strongly alternating (note that a0 = 0), and therefore B is positive semidefinite, but not
positive definite. For C, the coefficients are strongly alternating, so C is positive definite. ♠

Exercises

Exercise 11.8.1 Determine by direct calculation (i.e., without calculating the characteristic polynomial
or the eigenvalues) which of the following matrices are positive definite, positive semidefinite, or neither.
   
      2 1 1 1 0 0
1 1 1 −2 2 3
A= , B= , C= , D =  1 1 0 , E =  0 2 0 .
1 1 −2 5 1 3
1 0 1 0 0 −1

Exercise 11.8.2 Calculate the eigenvalues of each symmetric matrix, then determine for each matrix
whether it is positive definite, positive semidefinite, or neither.
   
    2 1 0 3 −4 2
2 2 4 −6
A= , B= , C= 1 1 −1  , D =  −4 4 0 .
2 5 −6 9
0 −1 2 2 0 4
is positive definite, positive semidefinite, or neither.

Exercise 11.8.3 Which of the following formulas define an inner product on R3 ? Here, u = [u1 , u2 , u3 ]T
and v = [v1 , v2 , v3 ]T .

(a) hu, vi = u1 v1 + 2u1 v2 + 2u2 v1 + 3u2 v2 + 3u3 v3 .

(b) hu, vi = u1 v1 − u1 v2 − u2 v1 + 2u2 v2 + u3 v3 .
(c) hu, vi = u1 v1 − u1 v2 − u2 v1 + u2 v2 + 4u3 v3 .
(d) hu, vi = u1 v1 − u1 v2 − u2 v1 + 3u2 v2 − u2 v3 − u3 v2 + u3 v3 .

Exercise 11.8.4 Use Descartes’ rule of signs to determine which of the following matrices are positive
definite and/or positive semidefinite.
 
  2 1 −1 1
  2 −1 0
2 2  1 1 0 0 
A= , B =  −1 1 0  , C =  
.
2 1 −1 0 2 −1 
0 0 1
1 0 −1 1
448 Inner product spaces

11.9 Application: Simplification of quadratic forms

Outcomes
A. Determine whether or not a function of several variables is a quadratic form.

B. Convert a quadratic form to and from matrix form.

C. Apply a change of variables to a quadratic form.

D. Diagonalize a quadratic form.

E. Find the principal axes of a quadratic form.

F. Sketch quadratic curves, such as circles and ellipses.

In this section, we will explore an application of the diagonalization of symmetric matrices, namely, the
simplification of quadratic forms. Quadratic forms are special kinds of functions that arise, for example,
in calculus when we approximate some quantity up to terms of second order.

Definition 11.74: Quadratic form

A quadratic form is a polynomial in n variables in which each term is of degree 2. For example,
the following is a quadratic form in 3 variables:

f (x, y, z) = 3x2 + 3y2 + 2xy − 4xz + 4yz.

More generally, a quadratic form is a function of the form

f (x1 , . . . , xn ) = q1 x21 + . . . + qn x2n + q12 x1 x2 + . . . + qi j xi x j + . . . + qn−1,n xn−1 xn .

The numbers q1 , . . . , qn , q12 , . . . , qn−1,n , which may be positive, negative, or zero, are called the
coefficients of the quadratic form.

Example 11.75: Quadratic forms

Which of the following are quadratic forms?

(a) f (x, y, z) = 2xy + 3xz − 5yz.

(b) g(x, y, z) = x2 + 2xy + y2 + 3.

(c) h(x, y, z) = (x + y)2 − z2 .

Solution. The function f is a quadratic form. The function g is not a quadratic form, because the constant
term, +3, is not of degree 2. It should be either a coefficient times the square of a variable, or a coefficient
 11.9. Application: Simplification of quadratic forms 449

times the product of two variables. The function h is a quadratic form. We can simplify it to h(x, y, z) =
x2 + 2xy + y2 − z2 . ♠

Definition 11.76: Matrix form of a quadratic form

Let A be a symmetric n × n-matrix, and let v = [x1 , . . . , xn ]T . Then

f (x1 , . . . , xn ) = vT Av

is a quadratic form in n variables. Conversely, every quadratic form in n variables can be uniquely
written in this way. We call this the matrix form of the quadratic form.

Example 11.77: Matrix form of a quadratic form

Find the coefficients of the quadratic form f (x, y, z) = vT Av, where
 
1 2 4
A= 2 0 −1  .
4 −1 −2

Solution. We have
  
  1 2 4 x
f (x, y, z) = x y z  2 0 −1   y  = x2 − 2z2 + 4xy + 8xz − 2yz.
4 −1 −2 z
Note that there is a term 2xy and a term 2yx, which together yield 4xy. Similarly the term 4xz and 4zx are
combined into 8xz, and the terms −yz and −zy are combined into −2yz. ♠

Example 11.78: Matrix form of a quadratic form

Write the quadratic form

f (x, y, z) = 5x2 − y2 + z2 + 2xy − 4xz + 3yz

in matrix form.

Solution. We can write this as

  
  5 1 −2 x
f (x, y, z) = x y z  1 −1 1.5   y  .
−2 1.5 1 z
Note: since the matrix A must be symmetric, we have no choice but to split the term 2xy evenly into 1xy
and 1yx. This explains why the (1, 2)- and (2, 1)-entries of the matrix are 1. Also, −4xz has been split into
−2xz and −2zx, and 3yz has been split into 1.5yz and 1.5zy. In general, the matrix of the quadratic form
f (x, y, z) = ax2 + by2 + cz2 + dxy + exz + f yz
450 Inner product spaces

is  
d e
a 2 2
 
A=

d
2 b f
2
.

e f
2 2 c
♠
We will now turn to the question of how to simplify quadratic forms. The primary tool we have for doing
so is a change of variables. This means replacing the variables x1 , . . . , xn by new variables y1 , . . . , yn that
are linear combinations of x1 , . . . , xn .

Example 11.79: Change of variables

Apply the change of variables

x = u
y = v−w
z = w

to the quadratic form

3x2 + y2 + 2xy + 2xz + 2yz

Solution. We have
3x2 + y2 + 2xy + 2xz + 2yz
= 3u2 + (v − w)2 + 2u(v − w) + 2uw + 2(v − w)w
= 3u2 + v2 − w2 + 2uv.
♠

The simplest kind of quadratic form is one that involves only squared variables, and no products of two
different variables. We call such quadratic forms diagonal.

Definition 11.80: Diagonal quadratic form

A quadratic form is diagonal if it is of the form

f (x1 , . . . , xn ) = q1 x21 + . . . + qn x2n .

Proposition 11.81: Diagonalization of quadratic forms

Every quadratic form can be made diagonal by a change of variables.

Proof. We first write the quadratic form in matrix form, i.e.,

f (x1 , . . . , xn ) = vT Av,
 11.9. Application: Simplification of quadratic forms 451

where A is a symmetric matrix and v = [x1 , . . . , xn ]T . By Theorem 11.63, A is orthogonally diagonalizable.

So let P be an orthogonal matrix such that
 
λ1 · · · 0
 
P−1 AP = PT AP = D =  ... . . . ... 
0 · · · λnn

is diagonal. Let w = [y1 , . . . , yn ]T be a new set of variables such that v = Pw. Then

vT Av = wT PT APw = wT Dw = λ1 y21 + . . . + λn y2n .

So the quadratic form is diagonal in the variables y1 , . . . , yn . ♠

Example 11.82: Diagonalization of quadratic forms

Consider the quadratic form
f (x, y) = 3x2 + 4xy + 6y2 .
Perform a change of variables so that the quadratic form becomes diagonal.

Solution. We first write f (x, y) in matrix form:


 
 3 2  x
f (x, y) = x y .
2 6 y
 
3 2
Next, we orthogonally diagonalize the matrix A = . In Example 11.64, we found that P−1 AP = D,
2 6
where    
1 1 −2 7 0
P= √ and D = .
5 2 1 0 2
Since P is orthogonal, we also have PT AP = D. Next, let u, v be new variables such that
     √ 
x u (u − 2v)/√5
=P =
y v (2u + v)/ 5

Then the change of variables gives

     
  x   T u   u
f (x, y) = x y A = u v P AP = u v D = 7u2 + 2v2 .
y v v

Example 11.83: Diagonalization of quadratic forms

Diagonalize the quadratic form

f (x, y, z) = 3x2 + 3y2 + 2xy − 4xz + 4yz.

452 Inner product spaces

Solution. The matrix form of f (x, y, z) is

  
  3 1 −2 x
f (x, y, z) = x y z  1 3 2   y  = vT Av.
−2 2 0 z

We orthogonally diagonalize the matrix A. In Example 11.64, we found that P−1 AP = D, where
 1   
√ − √13 √1 4 0 0
2 6
 
P =  √12 √13 − √16  and D =  0 4 0 .
0 √1 √2 0 0 −2
3 6

Let u, v, w be new variables such that    

x u
 y  = P v .
z w
Then the change of variables gives


  x
f (x, y, z) = x y z A y 
z
 
  T u
= u v w P AP  v 
w
 
  u
= u v w D v 

w
= 4u2 + 4v2 − 2w2 .

Next, we turn our attention to the task of sketching the solutions of quadratic equations in 2 or more
variables.

Example 11.84: Sketching circles and ellipses

Sketch the following curves:

(a) x2 + y2 = 1,

(b) x2 + 2y2 = 1.
 11.9. Application: Simplification of quadratic forms 453

Solution. (a) The curve x2 + y2 = 1 is the familiar unit circle.

y
1

x
−1 1

−1

√
(b) The curve x2 + 2y2 = 1 is the same, except that it has been stretched by a factor √ of 1/ 2 in the
y-direction. In other words, it is an ellipse with x-intercepts ±1 and y-intercepts ±1/ 2.

√1 y
2

x
−1 1

− √12

♠
√
In general,
√ for a, b > 0, the curve ax2 + by2 = 1 is an ellipse with x-intercepts ±1/ a and y-intercepts
±1/ b. Similarly, for 2 2 2
√ a, b, c > 0, the equation
√ ax + by + cz = 1√describes a 3-dimensional ellipsoid
with x-intercepts ±1/ a, y-intercepts ±1/ b, and z-intercepts ±1/ c:
z

√1
c

y
x √1 √1
a b

Each ellipse or ellipsoid has a set of principal axes, which are its axes of symmetry. When the quadratic
forms are diagonal, as in the above examples, the principal axes are the standard coordinate axes. When
454 Inner product spaces

the quadratic forms are not diagonal, the principal axes are the eigenvectors of the matrix A. This is the
content of the following proposition.

Proposition 11.85: Principal axes of a quadratic form

Consider a quadratic form in matrix form, f (v) = vT Av, where A is a positive definite n × n-matrix.
Let {u1 , . . . , un } be an orthonormal set of eigenvectors of A, and let λ1 , . . . , λn be the corresponding
eigenvalues. Then the solutions of the equation

vT Av = 1

form an n-dimensional ellipsoid whose principal axes are parallel to u1 , . . . , un and whose ui -
intercepts are ± √1 .
λi

Proof. Let P be the orthogonal matrix whose columns are u1 , . . . , un . From the proof of Proposition 11.81,
we know that the equation vT Av = 1 is equivalent to λ1 y21 + . . . + λn y2n = 1, where y1 , . . . , yn are variables
such that  
y1
 
v = P  ...  . (11.5)
yn
Since A is positive definite, we have λ1 , . . . , λn > 0 by Proposition 11.69. We therefore know that in
the (y1 , . . . , yn )-coordinate
√ system, the equation λ1 y21√ + . . . + λn y2n = 1 describes an ellipsoid whose y1 -
intercepts are ±1/ λ1 , whose y2 -intercepts are ±1/ λ2 , and so on. The only thing that remains to do
is to figure out the direction of the coordinate axes. The y1 -axis points in the direction of the point with
coordinates (y1 , . . . , yn ) = (1, 0, . . ., 0). Using the change of variables formula (11.5), we find that this
corresponds to the first column of P, i.e., v = u1 . Similarly, the y2 -axis points in the direction of u2 , and
so on. ♠

Example 11.86: Sketching a quadratic curve

Sketch the curve 3x2 + 4xy + 6y2 = 1.

Solution. The matrix for the quadratic form 3x2 + 4xy + 6y2 is
 
3 2
A= .
2 6

From Example 11.82, we know that the normalized eigenvectors of A are

   
1 1 1 −2
u1 = √ and u2 = √ ,
5 2 5 1
 11.9. Application: Simplification of quadratic forms 455

with respective eigenvalues λ1 = 7 and λ2 = 2. Therefore, by Proposition 11.85, the√curve 3x2 +√ 4xy +
2
6y = 1 is an ellipse with principal axes u1 and u2 , and with respective intercepts ±1/ 7 and ±1/ 2.
y

1 1 u1

u2 √1
7
1
√1
2 x
−1 1
− √1
7
− √1
2 −1
−1
−1

Exercises

Exercise 11.9.1 Which of the following are quadratic forms?

(a) f1 (x, y, z) = (x + y + z)2 .

(b) f2 (x, y, z) = x2 + y2 + z2 .
(c) f3 (x, y, z) = (x + y)2 + (x + z)2 + (y + z)2 .
(d) f4 (x, y, z) = x2 − y2 .
(e) f5 (x, y, z) = x2 + 2xyz + (y + z)2 .
(f) f6 (x, y, z) = x(y + z).
(g) f7 (x, y, z) = x2 y2 z2 .

Exercise 11.9.2 Find the coefficients of the quadratic form f (x, y, z) = vT Av, where
 
0 1 2
A =  1 −3 −4  .
2 −4 5
456 Inner product spaces

Exercise 11.9.3 Write the quadratic form

f (x, y, z) = (x + y)2 + 2(x − z)2 + 3yz
in matrix form.

Exercise 11.9.4 Apply the change of variables x = u + 2w, y = v, z = w − v to the quadratic form
x2 + 7y2 + 5z2 − 4xy − 4xz + 11yz.

Exercise 11.9.5 Perform a change of variables so that the quadratic form f (x, y) = 3x2 − 2xy + 3y2 be-
comes diagonal.

Exercise 11.9.6 Diagonalize the quadratic form

f (x, y, z) = 3x2 + 3y2 + 4z2 + 2xz − 2yz.

Exercise 11.9.7 Find the principal axes of the following curves, and sketch them:
(a) x2 + 12 y2 = 1,

(b) 2x2 + 4xy + 5y2 = 1.

(c) 3x2 + 2xy + 3y2 = 1.

Exercise 11.9.8 Find the principal axes of the ellipsoid 2x2 + 2y2 + 3z2 + 2xz − 2yz.

11.10 Complex inner product spaces

Outcomes
A. Compute dot products in Cn .

B. Use properties of the complex dot product to prove equalities and inequalities.

C. Compute the adjoint of a matrix.

D. Check whether an operation is a complex inner product.

E. Determine whether vectors in a complex inner product space are orthogonal.

F. Calculate the complex Fourier coefficients of a vector.

G. Use the Gram-Schmidt procedure to find an orthogonal basis of a subspace of a complex inner
product space.

H. Compute the orthogonal projection of a complex vector onto a subspace.

 11.10. Complex inner product spaces 457

So far, in this chapter, the field K was always R, the set of real numbers. The reason we have not con-
sidered inner products over other fields K is that the positive definite property requires hu, ui ≥ 0, and the
requirement that a scalar is “greater or equal to 0” does not make sense if K is, say, the field of integers
modulo p.
In this section, we will consider inner product spaces over the complex numbers. It turns out that the
theory of complex inner product spaces is similar, but not completely identical, to that of real inner product
spaces. To explain the difference, consider the definition of the dot product. In Rn , the dot product of two
vectors    
x1 y1
   
v =  ...  and w =  ... 
xn yn
is defined to be
·
v w = x1 y1 + . . . + xn yn .
One of the most important properties of the dot product is positivity: for all v, we have

·
v v = x21 + . . . + x2n ≥ 0.

The reason positivity holds is that the square of a real number is always greater or equal to 0. If we blindly
replaced x1 , . . . , xn by complex numbers and kept the same definition of dot product, positivity would no
longer hold. This is because for a complex number z, it is not in general true that z2 ≥ 0. In fact, z2 may
not be a real number, and even in cases where z2 is real, it may not be positive. For example, if z = i, then
z2 = −1.
Fortunately, all is not lost: the complex numbers actually do have a useful positivity property. Namely,
if z = a + bi is a complex number and z = a − bi is its complex conjugate, then

zz = (a − bi)(a + bi) = a2 + b2 ≥ 0.

So instead of squaring a complex number, we should multiply it by its conjugate. With this in mind, we
arrive at the following definition of dot product on Cn :

Definition 11.87: The dot product on Cn

Let   
v1 w1
   
v =  ...  and w =  ... 
vn wn
be vectors in Cn . Their (complex) dot product is defined to be

·
v w = v1 w1 + . . . + vn wn .

The complex dot product satisfies properties that are similar to, but not exactly the same as, the properties
satisfied by the real dot product.
458 Inner product spaces

Proposition 11.88: Properties of the complex dot product

The dot product satisfies the following properties, where u, v, w ∈ Cn and k, ℓ ∈ C.

· ·
• Conjugate symmetry: u v = v u.

• Linearity on the right: u·(kv + ℓw) = k(u·v) + ℓ(u·w).

• Antilinearity on the left: (ku + ℓv)·w = k(u·w) + ℓ(v·w).

• The positive definite property: u·u ≥ 0, and u·u = 0 if and only if u = 0.

We note that the complex dot product can be equivalently expressed as a matrix product, namely
 
w1
  
·
v w = v1 · · · vn  ...  = vT w.
wn
Here, v denotes the complex conjugate of a vector (i.e., taking the complex conjugate of each component
of a vector), and (−)T denotes the transpose as usual. As a matter of fact, when working with complex
vectors and matrices, it turns out that we should almost always take the complex conjugate at the same
time as taking the transpose. For this reason, we introduce a special name and notation for the conjugate
transpose of a vector or matrix.

Definition 11.89: Adjoint of a matrix

Let A be a complex n × m-matrix. The adjoint of A, denoted A∗ , is the transpose of the complex
conjugate of A. In symbols:
A∗ = AT .

With this definition, we can also write the dot product as

·
v w = v∗ w.

We are now ready to state the definition of a complex inner product, which is a generalization of the
complex dot product.

Definition 11.90: Complex inner product space

A complex inner product space is a complex vector space V equipped with an operation that
assigns to any pair of vectors u, v ∈ V a complex number hu, vi, called the inner product of u and
v, satisfying the following properties:

1. Conjugate symmetry: hu, vi = hv, ui.

2. Linearity on the right: hu, kv + ℓwi = khu, vi + ℓhu, wi.

3. Antilinearity on the left: hku + ℓv, wi = khu, wi + ℓhv, wi.

4. The positive definite property: hu, ui ≥ 0, and hu, ui = 0 if and only if u = 0.

 11.10. Complex inner product spaces 459

Note that conjugate symmetry implies that hu, ui is a real number for every vector u. Namely, let z = hu, ui.
Then by conjugate symmetry, we have

z = hu, ui = hu, ui = z.

Since z is equal to its own conjugate, it must be a real number. Therefore, the positive definite property
makes sense: when we require that hu, ui ≥ 0, we are talking about a real number that is greater than or
equal to 0. (It would not in general make sense to ask whether a complex number is greater than or equal
to 0).
The space Cn with the complex dot product is evidently an example of a complex inner product space.
Here is another example:

Example 11.91: Complex-valued continuous functions

Let a < b be real numbers, and let C[a, b] be the space of continuous, complex-valued functions
f : [a, b] → C. Given two such functions f , g ∈ C[a, b], we define their inner product as
Z b
h f , gi = f (x)g(x) dx.
a

With this operation, C[a, b] is a complex inner product space.

Armed with this definition of complex inner products, we can now pretty much redo everything we did for
real inner products in the complex case. The only thing we have to be careful about is to put the complex
conjugate operation in the correct places.
p
• The norm of a vector in a complex inner product space is defined to be kuk = hu, ui. This
definition makes sense because hu, ui ≥ 0.

• The Cauchy-Schwarz inequality |hu, vi| ≤ kuk kvk and the triangle inequality ku + vk ≤ kuk +
kvk hold in complex inner product spaces.

• Two vectors u, v in a complex inner product space are called orthogonal, in symbols u ⊥ v, if
hu, vi = 0.

• A vector u in a complex inner product space is called normalized if kuk = 1.

• A set of vectors {u1 , . . . , un } is called an orthogonal set if the vectors are non-zero and pairwise
orthogonal, and an orthonormal set if the vectors are moreover normalized.

• If u1 , . . . , uk are orthogonal, then ka1 u1 + . . . + ak uk k2 = |a1 |2 ku1 k2 +. . .+|ak |2 kuk k2 . The absolute
value signs are necessary because ai ai = |ai |2 .
460 Inner product spaces

Example 11.92: Orthogonal vectors

Consider C2 with the complex dot product, and the following vectors:
     
1 −i i
u= , v= , and w = .
i 1 1

Are u and v orthogonal? Are u and w orthogonal?

Solution. We have  
∗
  −i
hu, vi = u v = 1 −i = −i − i = −2i 6= 0,
1
so u and v are not orthogonal. We have
 
∗
  i
hu, wi = u w = 1 −i = i − i = 0,
1
so u and w are orthogonal. Note that it is crucial here that we did not forget to take the complex conjugate
of u, or else we would have gotten a different answer. ♠
In some formulas, we must be careful about whether we use hv, wi or hw, vi. Although this did not make
any difference in the case of real inner products, it does make a difference for complex inner products,
because in general, hv, wi =6 hw, vi. In particular, we have to be careful about this in the formulas for
Fourier coefficients, projections, and the Gram-Schmidt procedure.

Example 11.93: Complex Fourier coefficients

Let {u1 , . . . , un } be an orthogonal set of vectors in a complex inner product space, and let

v = a1 u1 + . . . + an un .

Which of the following formulas is correct?

hv, ui i hui , vi
ai = or ai = ?
hui , ui i hui , ui i

Solution. To check whether the first formula is correct, we calculate

hv, ui i ha1 u1 + . . . + ai ui + . . . + an un , ui i
=
hui , ui i hui , ui i
hu1 , ui i hui , ui i hun , ui i
= a1 + . . . + ai + . . . + an
hui , ui i hui , ui i hui , ui i
= a1 · 0 + . . . + ai · 1 + . . . + an · 0
= ai .
Note that because of antilinearity, this formula came out to be ai , and not ai . Therefore, the first formula
is not correct. The check the second formula, we calculate
hui , vi hui , a1 u1 + . . . + ai ui + . . . + an un i
=
hui , ui i hui , ui i
 11.10. Complex inner product spaces 461

hui , u1 i hui , ui i hui , un i

= a1 + . . . + ai + . . . + an
hui i hui , ui i hui , ui i
= a1 · 0 + . . . + ai 1 + . . . + an · 0
= ai .

Therefore, the second formula is correct. ♠

Since this is an important result, we state it as a proposition.

Proposition 11.94: Complex Fourier coefficients

If {u1 , . . . , un } is an orthogonal set of vectors in a complex inner product space and

v = a1 u1 + . . . + an un ,

then
hui , vi hv, ui i
ai = and ai = .
hui , ui i hui , ui i

Example 11.95: Calculating complex Fourier coefficients

Suppose that B = {u√ 1 , u2 , u3 } is an orthogonal basis for a complex inner product space V , such that
ku1 k = 1, ku2 k = 5, and ku3 k = 2. Moreover, suppose that v ∈ V is a vector such that hu1 , vi = i,
hu2 , vi = −2, and hu3 , vi = 1 − 2i. Find the coordinates of v with respect to B.

Solution. We need to find a1 , a2 , a3 such that v = a1 u1 + a2 u2 + a3 u3 . By Proposition 11.94, we have

hu1 , vi hu1 , vi i
a1 = = = = i,
hu1 , u1 i ku1 k2 1
hu2 , vi hu2 , vi −2 2
a2 = = = = − ,
hu2 , u2 i ku2 k2 5 5
hu3 , vi hu3 , vi 1 − 2i 1 1
a3 = = = = − i.
hu3 , u3 i ku3 k2 4 4 2

♠
The Gram-Schmidt orthogonalization procedure works without changes in complex inner product spaces,
as long as we are careful not to confuse hui , v j i with hv j , ui i.
462 Inner product spaces

Proposition 11.96: Complex Gram-Schmidt orthogonalization procedure

Let {v1 , . . . , vk } be a basis for some subspace W of a complex inner product space V . Define vectors
u1 , . . . , uk as follows:

u1 = v1 ,
hu1 , v2 i
u2 = v2 − u1 ,
hu1 , u1 i
hu1 , v3 i hu2 , v3 i
u3 = v3 − u1 − u2 ,
hu1 , u1 i hu2 , u2 i
.
..
hu1 , vk i hu2 , vk i huk−1 , vk i
uk = vk − u1 − u2 − . . . − uk−1 .
hu1 , u1 i hu2 , u2 i huk−1 , uk−1 i

Then {u1 , . . . , uk } is an orthogonal basis of W .

Example 11.97: Complex Gram-Schmidt orthogonalization procedure

Consider C3 with the complex dot product. Let
   
1+i i
v1 =  1  and v2 =  −1 
i −3

Use the Gram-Schmidt procedure to find an orthogonal basis for span {v1 , v2 }.

Solution. We start with 


1+i
u1 = v1 =  1  .
i
Next, we calculate
 
 i

hu1 , v2 i = u∗1 v2 = 1 − i 1 −i  −1  = (1 − i)i + 1(−1) + (−i)(−3) = 4i
−3
and
 
  1 + i
hu1 , u1 i = u∗1 u1 = 1 − i 1 −i  1  = (1 − i)(1 + i) + 1 · 1 + (−i)(i) = 4.
i
Therefore
     
i 1+i 1
hu1 , v2 i 4i
u2 = v2 − u1 = v2 − u1 = v2 − iu1 =  −1  − i  1  =  −1 − i  .
hu1 , u1 i 4
−3 i −2
 11.10. Complex inner product spaces 463

The desired orthogonal basis is {u1 , u2 }. We double-check that u1 and u2 are indeed orthogonal:
 
  1
hu1 , u2 i = u∗1 u2 = 1 − i 1 −i  −1 − i  = (1 − i)1 + 1(−1 − i) + (−i)(−2) = 0.
−2

♠
Orthogonal projections also work in complex inner product spaces.

Proposition 11.98: Orthogonal projection onto a subspace

Let V be a complex inner product space, and let W be a subspace of V . Assume {u1 , . . . , uk } is an
orthogonal basis of W , and v ∈ V is any vector. Let

hu1 , vi hu2 , vi huk , vi

v′ = u1 + u2 + . . . + uk .
hu1 , u1 i hu2 , u2 i huk , uk i

Then v′ is the best approximation of v in W , i.e., it is the element of W such that kv − v′ k is as

small as possible. Moreover, the vector v − v′ is orthogonal to W . We say that v′ is the orthogonal
projection of v onto W .

Example 11.99: Orthogonal projection onto a subspace

Consider the subspace of C3 spanned by
   
1 i
u1 =  i  and u2 =  1  .
0 1
 
0
With respect to the complex dot product, find the best approximation of v =  2  in this subspace.
4

Solution. First notice that hu1 , u2 i = 0, so that u1 and u2 are orthogonal. Therefore, the best approximation
is given by
hu1 , vi hu2 , vi
v′ = u1 + u2 .
hu1 , u1 i hu2 , u2 i
We calculate the relevant inner products:
 
 0 
hu1 , vi = 1 −i 0  2  = − 2i,
4
 
  0
hu2 , vi = −i 1 1  2  = 6,
4
464 Inner product spaces

 
 1 
hu1 , u1 i = 1 −i 0  i  = 2,
0
 
  i
hu2 , u2 i = −i 1 1  1  = 3.
1
Therefore,

    
1 i i
hu1 , vi hu2 , vi −2i 6
v′ = u1 + u2 . = u1 + u2 = −i i +2 1  =  3 .
hu1 , u1 i hu2 , u2 i 2 3
0 1 2
To double-check the answer, we can check that v − v′ is indeed orthogonal to u1 and u2 . We have
 
−i
v − v′ =  −1 
2
and
 
 −i 
hu1 , v − v′ i = 1 −i 0  −1  = 0,
2
 
  −i
hu2 , v − v′ i = −i 1 1  −1  = 0.
2
♠
We finish this section with some remarks on the differences between the notations used in mathematics and
in physics. Complex inner product spaces are very important in physics because they are the foundation
of quantum mechanics. The adjoint of a matrix A is usually denoted A∗ in mathematics and A† in physics.
In quantum mechanics, a column vector v is often written |vi, and the corresponding row vector v∗ is then
written as hv|. This is the so-called Dirac notation. With this convention, an inner product v∗ w is hv||wi,
which is usually written as hv|wi. Also, the matrix vw∗ is written |vihw| and is called an outer product. In
mathematics, it is customary for inner products to be linear in the left component and antilinear in the right
component. In physics, it is customary to use the opposite convention, i.e., inner products are antilinear in
the left component and linear in the right component. In this book, we have used the physics convention
of antilinearity in the left component, because it is the better convention.

Exercises

Exercise 11.10.1 Compute the following complex dot products in Cn :

           
    2 i i 2 i 1 + 2i
1 1+i
(a)
i · 2

, (b) 1
i
  ·1
−3
 , (c)  1
−3
  1 ,
i
· 1
·
(d)  i + 1   −2  .
−i
 11.10. Complex inner product spaces 465

Exercise 11.10.2 Compute the norm of each of the following vectors in Cn :

   
  1 1+i
1
(a) , (b)  0  , (c)  1 − i  .
i
1 + 2i 3i

· · ·
Exercise 11.10.3 Suppose u, v ∈ Cn and u u = 1, v v = 2, and u v = 2i. Then compute:

·
(a) (2u + v) (u − v).

(b) (iu − v)·(u − 2iv).

Exercise 11.10.4 Compute the adjoint of the following matrices.

 
i    
1 i 2 1 + i −i
(a)  i − 1  , (b) , (c) .
−i 2 1 − i 3 2 + 2i
2

Exercise 11.10.5 Suppose u and v are vectors in a complex inner product space such that hu, ui = 3,
hv, vi = 2, and hu, vi = i + 1. Then compute:

(a) hu + 2v, u − ivi.

(b) hu − iv, 2u + vi.

Exercise 11.10.6 In C[0, 2π ], consider the functions f (x) = sin x+i cos x, g(x) = 1, and h(x) = x. Compute
the following:
(a) h f , gi, (b) h f , hi, (c) h f , f i, (d) k f k .

Exercise 11.10.7 Which of the following vectors, if any, are orthogonal?

       
1 1 0 0
 
u1 = i , u2 = −i , u3 =   1 , u4 = 0  .
 
0 1 −i 1

Exercise 11.10.8 Suppose√ that B = {u1 , u2 , u3 } is√an orthogonal basis for a complex inner product space
V , such that ku1 k = 2, ku2 k = 2, and ku3 k = 3. Moreover, suppose that v ∈ V is a vector such that
hu1 , vi = 1, hu2 , vi = 3i, and hu3 , vi = 1 − i. Find the coordinates of v with respect to B.

Exercise 11.10.9 Consider C3 with the complex dot product. Let

   
0 1
v1 =  i  and v2 =  1 + i 
2 3i + 2
466 Inner product spaces

Use the Gram-Schmidt procedure to find an orthogonal basis for span {v1 , v2 }. Then find an orthonormal
basis.

Exercise 11.10.10 Consider the subspace of C3 spanned by the orthogonal vectors

   
2 i

u1 = i  and u2 = 3  .

1 i
 
−2i
Find the best approximation of v =  3  in this subspace.
4i

11.11 Unitary and hermitian matrices

Outcomes
A. Determine whether a complex linear transformation is an isometry and/or unitary.

B. Determine whether a complex matrix is unitary.

C. Determine whether a complex matrix is hermitian.

D. Calculate the eigenvalues and eigenvectors of a hermitian matrix.

E. Compute an orthogonal basis of eigenvectors for a hermitian matrix.

F. Unitarily diagonalize a hermitian matrix.

In the context of real inner product spaces, we studied orthogonal functions, orthogonal matrices, and
symmetric matrices. The corresponding concepts in the context of complex inner product spaces are
unitary functions, unitary matrices, and hermitian matrices. We will introduce these concepts in this
section. Since the proofs are similar to those in Section 11.6, we omit most of them.

Definition 11.100: Isometries and unitary maps of complex inner product spaces
Let V ,W be complex inner product spaces. A linear transformation T : V → W is called an isometry
if for all u, v ∈ V ,
hT (u), T (v)i = hu, vi.
An isometry that is also invertible is called a unitary transformation, or simply unitary.

In the case of real inner product spaces, we found in Proposition 11.56 that a square matrix P is the matrix
of an orthogonal transformation (with respect to orthonormal bases) if and only if PT P = I. In the complex
case, we have an analogous property, except that we must use the adjoint instead of the transpose.
 11.11. Unitary and hermitian matrices 467

Proposition 11.101: The matrix of a unitary transformation

Let T : V → W be a linear transformation between finite-dimensional complex inner product spaces
V and W . Let B and C be orthonormal bases of V and W , respectively, and let P = [T ]C,B be the
matrix of T with respect to the bases B and C. Then

(a) T is an isometry if and only if P∗ P = I .

(b) T is unitary if and only if P∗ P = I and dimV = dimW .

This motivates the following definition. We therefore define a unitary matrix to be an n × n-matrix satis-
fying P∗ P = I.

Definition 11.102: Unitary matrix

An n × n-matrix P is called unitary if P∗ P = I .

We have the following analogue of Proposition 11.59:

Proposition 11.103: Conditions for unitary matrices

The following are equivalent for a complex n × n-matrix P:

(a) P is unitary.

(b) P∗ P = I .

(c) P is invertible and P−1 = P∗ .

(d) PP∗ = I .

(e) P∗ is unitary.

(f) The columns of P form an orthonormal set of vectors.

(g) The rows of P form an orthonormal set of vectors.

Example 11.104: Unitary matrices

Determine which of the following matrices are unitary.
     
1 0 1 1 i 1 1 i
A= , B= √ , C= √ .
0 i 2 i 1 2 −i 1

Solution. We have     
∗ 1 0 1 0 1 0
A A= = = I,
0 −i 0 i 0 1
468 Inner product spaces

so A is unitary. Similarly, we have

    
∗ 1 1 i 1 −i 1 2 0
B B= = = I,
2 i 1 −i 1 2 0 2

so B is unitary as well. On the other hand,

    
∗ 1 1 i 1 i 1 2 2i
C C= = 6= I,
2 −i 1 −i 1 2 −2i 2

so C is not unitary. Equivalently, we could have checked whether the columns of A, B, and C form
an orthonormal set of vectors (they do, in the case of A and B, but don’t, in the case of C. See also
Example 11.92). ♠
Of course, if P happens to be a matrix with real entries, then P is unitary if and only if it is orthogonal,
T
because in that case P∗ = P = PT .
Recall that a matrix A is called symmetric if A = AT . In the complex world, we are more often
interested in the property A = A∗ . A matrix with this property is called hermitian (after the French
mathematician Charles Hermite, 1822–1901).

Definition 11.105: Hermitian matrix

A complex n × n-matrix A is called hermitian if A = A∗ .

Example 11.106: Hermitian vs. symmetric matrices

Which of the following matrices are hermitian? Which ones are symmetric?
       
1 2 1 2i 1 2i i 2
A= , B= , C= , D= .
2 1 2i 1 −2i 1 2 i

Solution. The matrix A is symmetric and also hermitian. The matrix B is symmetric but not hermitian. In
fact, we have  
∗ 1 −2i
B = 6= B.
−2i 1
The matrix C is hermitian but not symmetric. The matrix D is symmetric but not hermitian. ♠
 
A matrix A = ai j is hermitian if and only if ai j = a ji , for all i, j. In particular, the diagonal entries of a
hermitian matrix are always real, and the off-diagonal entries come in complex conjugate pairs. If all of
the entries in a matrix A are real, then A is hermitian if and only if it is symmetric.
Hermitian matrices are of interest, among other things, because their eigenvalues are always real.
Moreover, eigenvectors for distinct eigenvalues are orthogonal. The following proposition is analogous to
Proposition 11.62.
 11.11. Unitary and hermitian matrices 469

Proposition 11.107: Eigenvalues and eigenvectors of hermitian matrices

Let A be a hermitian matrix. Then

(a) All eigenvalues of A are real.

(b) Eigenvectors for distinct eigenvalues of A are orthogonal.

Proof. (a) Suppose λ is an eigenvalue of A, with eigenvector v. We will evaluate v∗ Av in two different
ways:

v∗ Av = v∗ (Av) = v∗ (λ v) = λ v∗ v,
v∗ Av = (v∗ A)v = (v∗ A∗ )v = (Av)∗ v = (λ v)∗ v = λ v∗ v.

Therefore, λ v∗ v = λ v∗ v. Since v∗ v is non-zero, this implies λ = λ , i.e., λ is real.

(b) Suppose v and w are eigenvectors for eigenvalues λ and µ , respectively, and λ 6= µ . From part (a), we
know that both λ and µ are real. By evaluating v∗ Aw in two different ways, we find that

v∗ Aw = v∗ (Aw) = v∗ (µ w) = µ v∗ w,
v∗ Aw = (v∗ A)w = (A∗ v)∗ w = (Av)∗ w = (λ v)∗ w = λ v∗ w.

Therefore, µ v∗ w = λ v∗ w, or equivalently (λ − µ )v∗ w = 0. Since by assumption, λ − µ 6= 0, we must

have v∗ w = 0, i.e., v ⊥ w. ♠
We say that a matrix A is unitarily diagonalizable if there exists a unitary matrix P and a diagonal
matrix D such that D = P−1 AP. The following is the main theorem about the diagonalization of hermitian
matrices. It is analogous to Theorem 11.63 in the real case.

Theorem 11.108: Diagonalization of hermitian matrices

Every hermitian matrix A is unitarily diagonalizable as D = P−1 AP. Moreover, the entries of D are
real.

Example 11.109: Diagonalization of hermitian matrices

The matrix  
3 −2i
A=
2i 6
is hermitian. Unitarily diagonalize A, i.e., find a unitary matrix P and a real diagonal matrix D such
that D = P−1 AP.

Solution. The characteristic polynomial is

 
 3 − λ −2i 
det(A − λ I) =   = (3 − λ )(6 − λ ) + 4i 2 = λ 2 − 9λ + 14.

2i 6−λ
470 Inner product spaces

Its roots are λ1 = 2 and λ2 = 7. For the eigenvalue λ1 = 2, we find the normalized eigenvector
 
1 2i
v1 = √ .
5 1
For the eigenvalue λ2 = 7, we find the normalized eigenvector
 
1 1
v2 = √ .
5 2i
Note that these eigenvectors are orthogonal to each other, confirming Proposition 11.107. We therefore
have D = P−1 AP, where    
2 0 1 2i 1
D= and P = √ .
0 7 5 1 2i
Note that P is unitary and D is real diagonal. ♠

Exercises

Exercise 11.11.1 Which of the following linear transformations T : C2 → C2 is unitary?

   
x x
(a) T = .
y iy
   
x iy
(b) T = .
y −ix
   
x x + iy
(c) T   =  .
y ix + y
" #!  
x √1 x + √i y
(d) T = 2 2 .
y √1 x − √i y
2 2

Exercise 11.11.2 Determine which of the following matrices are unitary.

     
0 i 1 4 3i 1 1 2i
A= , B= , C= √ .
i 0 5 3 4i 5 2i 1

Exercise 11.11.3 Determine which of the following matrices are unitary.

 
    1 1 1 1
1 i 0 1 2 2i
1 1 1  1 i −1 −i 
A= 
0 1 i , B= −2i 2i 1  , C =  .
2 3 2  1 −1 1 −1 
i 0 1 2i i 2
1 −i −1 i
 11.11. Unitary and hermitian matrices 471

Exercise 11.11.4 Assume A and B are unitary 3 × 3-matrices. Then which of the following matrices are
unitary?

(a) AB (b) A + B (c) iA (d) − A (e) 2A ( f ) B−1 (g) AT (h) A∗ (i) A2 ( j) ABA−1

Exercise 11.11.5 Assume A and B are 3 × 3-matrices, and assume A is unitary and B is invertible. Then
which of the following matrices are unitary?

(a) AB (b) A + B ( f ) BAB−1 (g) ABA−1 (h) BB∗ (i) BB−1

Exercise 11.11.6 Which of the following matrices are hermitian? Which ones are symmetric?
       
0 1 i 2 0 i 1 1+i
A= , B= , C= , D= .
1 0 2 −i −i 0 1−i 2

Exercise 11.11.7 Which of the following matrices are hermitian? Which ones are symmetric?
     
0 −2 1 + i 3 i −1 + 2i i i i
A =  −2 3 i  , B =  −i 0 −1 + i  , C =  i i i  .
1−i i 2 −1 − 2i −1 − i 2 i i i

Exercise 11.11.8 Assume A and B are hermitian 3 × 3-matrices, and A is invertible. Then which of the
following matrices are hermitian?

(a) AB (b) A + B (c) iA (d) − A (e) 2A ( f ) A∗ BA (g) A−1 BA (h) BB∗ (i) A2

 
2 1+i
Exercise 11.11.9 Unitarily diagonalize the hermitian matrix A = .
1−i 3
 
0 −1 + 2i
Exercise 11.11.10 Unitarily diagonalize the hermitian matrix A = .
−1 − 2i 4
 
2 1+i 1−i
Exercise 11.11.11 Unitarily diagonalize the hermitian matrix A =  1 − i 3 −2i .
1 + i 2i 3
472 Inner product spaces

11.12 Application: Principal component analysis

Outcomes
A. Compute the principal components of a matrix A.

B. Compute the centroid of a collection of data points.

C. Find the k-dimensional subspace that best approximates a given collection of data points.

D. Find the k-dimensional affine subspace that best approximates a given collection of data
points.

E. Compute the total squared distance of the data points to the best fit subspace (or best fit affine
subspace).

In this section, we will explore an application of the diagonalization of symmetric matrices called princi-
pal component analysis. Imagine we are given a collection of data points such as the following:

(11.6)

Although these points are spread out in two dimensions, they seem to be located pretty close to a 1-
dimensional subspace. Probably the best way to interpret this particular data set is to think of the points
as being “essentially” on a line, up to some small random errors.
More generally, suppose we are given a collection of data points in n-dimensional space, and we
are looking for a k-dimensional subspace that all data points are close to. This is an important way to
make sense of high-dimensional data. For example, it would be very difficult to visualize data in a 100-
dimensional space. However, it we knew that the data points lie very close to a 2-dimensional subspace,
then we could project all of the points to the subspace to obtain a 2-dimensional image of the data.
To state the problem more precisely, let us introduce the following notation. If W is a subspace of Rn
and v ∈ Rn is a vector, let us write d(v,W ) for the shortest distance from v to W (i.e., the distance from v
 11.12. Application: Principal component analysis 473

to W along a line that is perpendicular to W ). Moreover, if W is a subspace of Rn and v1 , . . . , vm ∈ Rn are

the position vectors of m points, we define the total squared distance of the points to the subspace to be
the quantity
D = d(v1 ,W )2 + . . . + d(vm ,W )2 .
Then the problem we would like to solve can be stated as follows:

Problem 11.110: Subspace fitting problem

Given vectors v1 , . . . , vm ∈ Rn and given an integer k ≤ n, find the k-dimensional subspace W ⊆ Rn
that minimizes the total squared distance, i.e., such that D is as small as possible.

The following proposition gives us a method for solving the subspace fitting problem. It turns out that
the key ingredient in solving this problem is the diagonalization of symmetric matrices. The method was
discovered by Gale Young in 1937.

Proposition 11.111: Solution of the subspace fitting problem

Given vectors v1 , . . . , vm ∈ Rn and k ≤ n, the optimal solution to the subspace fitting problem can
be computed as follows:

1. Let A be the m × n-matrix whose rows are vT1 , . . . , vTm . (Or equivalently, AT is the n × m-matrix
whose columns are v1 , . . . , vm .)

2. Let B = AT A. Then B is a positive semidefinite n × n-matrix.

3. By Proposition 11.69, all eigenvalues of B are real and non-negative. Let λ1 , . . . , λn be the
eigenvalues of B, listed according to their multiplicity and in decreasing order, i.e., so that
λ1 ≥ λ2 ≥ . . . ≥ λn ≥ 0. Let u1 , . . . , un be the corresponding eigenvectors.

4. Then W = span {u1 , . . . , uk } is the solution to the subspace fitting problem. Moreover, the
total squared distance of the points to this subspace is

D = λk+1 + . . . + λn .

Example 11.112: Subspace fitting in R2

Consider the following collection of points in R2 :
                 
2 −1 2 −6 6 0 1 −2 −7
, , , , , , , , .
−3 0 3 −7 11 −1 6 −3 −6

Find the 1-dimensional subspace that best approximates this collection of points. What is the total
squared distance of the points to the subspace?

Solution. We follow the steps outlined in Proposition 11.111.

474 Inner product spaces

1. We have  
T 2 −1 2 −6 6 0 1 −2 −7
A = .
−3 0 3 −7 11 −1 6 −3 −6

2. We calculate  
T 135 162
B=A A= .
162 270

3. The eigenvalues of B are λ1 = 378 and λ2 = 27, with corresponding eigenvectors

   
2 3
u1 = and u2 = .
3 −2

4. The desired subspace W is spanned by the eigenvector corresponding to the largest eigenvalue, i.e.,
W = span {u1 }. The total squared distance is λ2 = 27.

The space W is shown in the following illustration, along with the original points:

10

10

Of course, the example was rigged to ensure that the eigenvalues are integers. In real life, the entries of A
and B, as well as the eigenvalues and components of the eigenvectors are usually arbitrary real numbers.
♠

Example 11.113: Subspace fitting in R3

Consider the following collection of points in R3 :
                   
 −7 0 2 10 −2 −8 5 −6 9 −2 
 4  ,  3  ,  −5  ,  −4  ,  5  ,  −1  ,  4  ,  9  ,  −6  ,  −7  .
 
5 3 −4 1 4 −5 2 6 3 −8

(a) Find the 1-dimensional subspace that best approximates this collection of points.

(b) Find the 2-dimensional subspace that best approximates this collection of points.

(c) What is the 3-dimensional subspace that best approximates this collection of points?

In each case, what is the total squared distance of the points to the subspace?

Solution. Again, we follow the steps from Proposition 11.111. We can do the calculations for parts (a),
(b), and (c) at the same time.
 11.12. Application: Principal component analysis 475

1. We have  
−7 0 2 10 −2 −8 5 −6 9 −2
AT =  4 3 −5 −4 5 −1 4 9 −6 −7  .
5 3 −4 1 4 −5 2 6 3 −8

2. We calculate  
367 −154 16
B = AT A =  −154 274 170  .
16 170 205

3. The eigenvalues of B are λ1 = 513, λ2 = 306, and λ3 = 27, with corresponding eigenvectors
     
−2 2 −1
u1 =  2  , u2 =  1  and u3 =  −2  .
1 2 2

For part (a), the desired 1-dimensional subspace is spanned by the eigenvector corresponding to the
largest eigenvalue, i.e., it is span {u1 }. The total squared distance is λ2 + λ3 = 306 + 27 = 333.
For part (b), the desired 2-dimensional subspace is spanned by the eigenvectors corresponding to the
two largest eigenvalues, i.e., it is span {u1 , u2 }. The total squared distance is λ3 = 27.
Finally, in part (c), the desired 3-dimensional subspace is spanned by all three eigenvectors; it is of
course R3 itself, since it is the only 3-dimensional subspace. The total squared distance is 0, since all
points lie in the subspace. ♠
The vectors u1 , . . . , un that appear in the solution of the subspace fitting problem are called the principal
components of the matrix A.

Definition 11.114: Principal components

Let A be an m × n-matrix. The principal components of A are the (normalized, orthogonal) eigen-
vectors u1 , . . . , un of the positive semidefinite n × n-matrix AT A. They are usually listed in order of
decreasing eigenvalues.

The first principal component u1 gives the direction in which the rows of A show the most variability.
The second principal component u2 gives the direction in which the rows of A show the most remaining
variability that is orthogonal to u1 . The third principal component u3 gives the direction of most variability
that is orthogonal to u1 and u2 , and so on.
476 Inner product spaces

Subspace fitting vs. curve fitting

In the particular case where n = 2 and k = 1, we are looking for a 1-dimensional subspace, i.e., a line
through the origin, which best fits the given 2-dimensional data, as in the illustration (11.6) above or as in
Example 11.112. On its face, the subspace fitting problem in this case seems similar to the linear curve
fitting problem we solved in Section 11.5. However, there is a subtle but important difference: in linear
curve fitting, we were seeking to minimize the distances of the points from the line in the y-direction,
whereas in subspace fitting, we are seeking to minimize the distances of the points from the subspace in
the direction perpendicular to the subspace. The following pair of pictures illustrates the difference:

minimize
minimize
distances
vertical
perpendicular
distances
to subspace
Linear curve fitting Subspace fitting

Affine fitting

So far, we have been looking to approximate a given collection of points by a subspace, which necessarily
passes through the origin. But sometimes the points may not be near the origin, as in this example:

In this case, approximating the points by a subspace passing through the origin does not make much sense.
Instead, we should be looking for an affine subspace. An affine subspace is similar to a subspace, except
it does not necessarily contain the origin.
 11.12. Application: Principal component analysis 477

Definition 11.115: Affine subspace

Let V be a vector space. A subset A ⊆ V is called an affine subspace of V if A is either empty, or
else of the form
A = v +W = {v + w | w ∈ W } ,
where v ∈ V and W is a subspace of V .

For example, in Chapter 3, we considered lines and planes in Rn that pass through a given point (not
necessarily the origin). These are examples of affine subspaces of Rn . The affine subspace fitting problem
is analogous to the subspace fitting problem:

Problem 11.116: Affine subspace fitting problem

Given the position vectors v1 , . . . , vm of m points in Rn , and given an integer k ≤ n, find the k-
dimensional affine subspace A ⊆ Rn that minimizes the total squared distance from the points to
A.

It turns out that the optimal solution to the affine subspace fitting problem can be computed by first com-
puting the centroid of the points, shifting the whole problem so that the centroid is at the origin, and then
solving an ordinary subspace fitting problem.

Definition 11.117: Centroid

Given m vectors v1 , . . . , vm , their centroid is the vector
1
v= (v1 + . . . + vm ).
m
It is also sometimes called the average or the center of mass of the vectors.

Proposition 11.118: Solution of the affine subspace fitting problem

Given vectors v1 , . . . , vm ∈ Rn and k ≤ n, the optimal solution to the affine subspace fitting problem
can be computed as follows:

1. Compute the centroid v = m1 (v1 + . . . + vm ) of the vectors.

2. Let wi = vi − v, for all i = 1, . . ., n.

3. Compute the solution W to the (ordinary) subspace fitting problem for w1 , . . . , wm , as in

Proposition 11.111.

Then the best solution to the affine subspace problem is v +W .

478 Inner product spaces

Example 11.119: Affine subspace fitting problem

Consider the following collection of points in R2 :
                 
10 2 5 8 2 3 4 10 1
, , , , , , , , .
−6 10 −1 3 5 3 11 −1 12

Find the 1-dimensional affine subspace that best approximates this collection of points. What is the
total squared distance of the points to the subspace?

Solution. We start by computing the centroid:

   
1 1 45 5
v = (v1 + . . . + v9 ) = = .
9 9 36 4

Next, we shift all vectors by −v to get a new collection of vectors w1 , . . . , w9 centered at the origin. For
example,
     
10 5 5
w1 = v1 − v = − = ,
−6 4 −10
     
2 5 −3
w2 = v2 − v = − = ,
10 4 6

and so on. We get

                 
5 −3 0 3 −3 −2 −1 5 −4
{w1 , . . . , w9 } = , , , , , , , , .
−10 6 −5 −1 1 −1 7 −5 8

Next we, we proceed as in Proposition 11.111 to find the best subspace fitting w1 , . . . , w9 . We have
 
T 5 −3 0 3 −3 −2 −1 5 −4
A =
−10 6 −5 −1 1 −1 7 −5 8

and  
T 98 −136
B=A A= .
−136 302
The eigenvalues of B are λ1 = 370 and λ2 = 30, with corresponding eigenvectors
   
1 2
u1 = and u2 =
−2 1

Thus, the best-fitting 1-dimensional subspace for w1 , . . . , w9 is W = span {u1 }, and the best-fitting 1-
dimensional affine subspace for v1 , . . . , v9 is
    
5 1 
v +W = {v + w | w ∈ W } = +t t ∈R .
4 −2 
 11.12. Application: Principal component analysis 479

Note that this is the equation of a line passing through the centroid v, and with direction vector u1 . The
points v1 , . . . , v9 , their centroid, and the affine subspace v +W are shown in the following illustration:

10
centroid

10

Application to U.S. Senate voting data

The United States Senate votes on a lot of things: motions, resolutions, amendments, and bills, among
other things. Many of these votes are roll call votes, which means that the vote of every individual
senator is recorded (as opposed to a voice vote, where only the outcome is recorded). Roll call data
for the last 3 decades is publicly available and can be downloaded from the U.S. Senate website at
https://www.senate.gov/legislative/votes.htm .
We will now explore how to use linear algebra, and in particular principal component analysis, to
gain some useful information from the voting records.2 I have made a spreadsheet containing the votes
of 99 senators for the first 200 roll call votes of 2007. Each row in the spreadsheet corresponds to a
senator, listed in alphabetical order from Daniel Akaka of Hawaii to Ron Wyden of Oregon. I omitted
one senator who died during 2007. Each column of the spreadsheet corresponds to a vote. For exam-
ple, the first roll call vote of 2007 was on a resolution to honour President Gerald Ford (it passed 88
to 0). Each cell of the spreadsheet contains the number 1 if the senator voted “yes”, the number −1
if the senator voted “no”, and the number 0 if the senator did not vote. The spreadsheet is available
from https://www.mathstat.dal. a/~selinger/linear-algebra/ under “Supplementary materi-
als”. Here are the first few rows and columns of the spreadsheet.
Akaka, Daniel (D) HI 1 1 1 1 1 −1 1 . . .
Alexander, Lamar (R) TN 0 1 −1 1 −1 −1 1 . . .
Allard, A. (R) CO 1 1 −1 −1 −1 1 1 ...
Baucus, Max (D) MT 1 1 1 1 1 −1 1 . . .
Bayh, Evan (D) IN 1 1 1 −1 1 −1 1 . . .
Bennett, Robert (R) UT 1 1 −1 1 1 −1 1 . . .
.. .. .. .. .. .. .. .. . .
. . . . . . . . .
The human mind is not very well equipped to deal with such massive amounts of data. Rather than
listing 122 motions that Senator X supported and 78 motions that she opposed, we like to come up with
2 Thisexample was inspired by Examples 11.2.13 and 11.3.15 of “Coding the Matrix: Linear Algebra through Computer
Science Applications” by Philip N. Klein.
480 Inner product spaces

abstractions, such as Senator X is “conservative”, “pro choice”, “pro business”, “hawkish”, etc. However,
the problem with abstractions is that they do not necessarily mean anything in the real world. In the real
world, a senator’s record is just a sequence of votes.
We will represent each senator by a vector in R200 , which corresponds to a row of the above table. For
example, to Senator Akaka, we associate the vector
 
1
 1 
 
 1 
 
 1 
 
 1  ∈ R200 .
 
 −1 
 
 1 
 
..
.

Thus, we can represent each senator (or more precisely, each senator’s voting record) as a point in 200-
dimensional space. In this way, the voting data can be interpreted as 99 points in R200 .
Unfortunately, 200 dimensions are impossible to visualize. But what if the voting records of all the
senators lie on (or at least close to) a much-smaller-dimensional affine subspace? This is actually not an
unreasonable expectation; after all, there are probably only a handful of issues most senators care about.
For example, if a certain senator supports gun control, he will be likely to vote a certain way on measures
that affect gun control. If another senator supports the gun lobby, she is likely to vote the opposite way.
We can thus consider this as an instance of the affine subspace problem: we are looking for a low-
dimensional affine subspace that is close to all 99 points. Following the method of Proposition 11.118,
we first find the centroid of the points, and then we compute a certain 99 × 200-matrix A and a positive
semidefinite 200 × 200-matrix B = AT A. Using software, we can find the eigenvalues and -vectors of B.
The first few eigenvalues (in decreasing order) are:

λ1 = 7255.65, λ2 = 519.16, λ3 = 430.60, λ4 = 278.05, and λ5 = 230.56.

All of the remaining eigenvalues are less than 200, and the sum of the remaining eigenvalues is λ6 + . . . +
λ200 = 3913.46. This means that the vast majority of the voting behavior of each senator is determined
by a single dimension, given by the eigenvector corresponding to the eigenvalue λ1 . In other words, there
is a 1-dimensional affine subspace that all 99 points are pretty close to. If we project each senator to this
 11.12. Application: Principal component analysis 481

affine subspace, we get the following picture:

McConnell, KY 2007 U.S. Senate voting data, projection to the first principal component:

Whitehouse, RI
Lieberman, CT
Alexander, TN

Coleman, MN

Feinstein, CA
Johnson, SD
McCain, AZ

Clinton, NY
Bennett, UT
Graham, SC
DeMint, SC

Salazar, CO
Snowe, ME
Warner, VA

Nelson, NE
Specter, PA

Nelson, FL
Allard, CO

Obama, IL
Bond, MO

Smith, OR
Hatch, UT

Hagel, NE

Pryor, AR
Enzi, WY

Webb, VA
Cornyn, TX

Voinovich, OH
Brownback, KS

Collins, ME

McCaskill, MO
Murkowski, AK

Reid, NV
Schumer, NY
Sanders, VT
Sessions, AL

Vitter, LA

Cochran, MS

Kennedy, MA
Roberts, KS
Shelby, AL

Lugar, IN

Baucus, MT
Coburn, OK

Biden, DE
Kyl, AZ

Conrad, ND
centroid

For convenience, Republican senators have been shown in red and Democratic and independent senators
in blue. Not all senators have been named, because in some areas they are clustered very densely. An
interpretation of the principal component then immediately suggests itself: it appears to be the “conser-
vative” vs. “liberal” axis. We can use this picture to assist in answering questions such as: “Which party
votes more uniformly?”, “Which state are the most liberal Republicans from?”, “Which state are the most
conservative Democrats from?”, “Was Obama really a radical?”, and “Was McCain really a maverick?”.
If we repeat the same calculation for the 2017 senate, we get the following picture:

2017 U.S. Senate voting data, projection to the first principal component:
McConnell, KY

Manchin, WV

Feinstein, CA

Franken, MN
Donnelly, IN

Warren, MA
Sanders, VT
Collins, ME

Cardin, MD
Bennet, CO
Warner, VA
Heller, NV

Durbin, IL
Kaine, VA
Rubio, FL
Risch, ID

Paul, KY

McCaskill, MO
King, ME
Cruz, TX

Tester, MT
Portman, OH

Klobuchar, MN

Gillibrand, NY
Duckworth, IL
McCain, AZ

Heitkamp, ND
Inhofe, OK

Isakson, GA

Carper, DE
Cortez Masto, NV

Schumer, NY

centroid

We can use this to help answer questions such as “Has the senate become more partisan between 2007
and 2017?”.
If we instead project the data onto the first two principal components, we get the following picture for
482 Inner product spaces

the 2007 data:

Coburn, OK 2007 U.S. Senate voting data, projection to the first two principal components:
DeMint, SC

Johnson, SD

McCaskill, MO
Vitter, LA

McCain, AZ
Allard, CO

Graham, SC

Obama, IL
Cornyn, TX

Sessions, AL

Clinton, NY
Schumer, NY
Brownback, KS

Biden, DE

Sanders, VT
Whitehouse, RI
Feinstein, CA
Hagel, NE

Reid, NV
Voinovich, OH

Bayh, IN
Shelby, AL

Lieberman, CT
McConnell, KY

Baucus, MT
centroid

Salazar, CO
Pryor, AR
Nelson, NE
Smith, OR
Cochran, MS
Warner, VA
Bond, MO

Snowe, ME
Collins, ME
Lugar, IN

Specter, PA
Murkowski, AK

Coleman, MN
Bennett, UT

The picture clearly shows senators clustering in certain areas. We can use this to help answer certain ques-
tions, for example, “How different was Sanders’s voting record from Clinton’s?”. However, although the
2-dimensional picture seems to reveal more detail, its interpretation is less clear. While it seems obvious
that the horizontal axis corresponds to a conservative vs. liberal world view, it is much less obvious what
the political meaning of the vertical axis is. Maybe it is related to some issue that does not typically follow
party lines, such as North vs. South, rich states vs. poor states, pro-immigration vs. anti-immigration, and
so on. To find a convincing interpretation of the vertical axis, further investigation of the data would be
required (such as, looking at the actual content of the votes in question).
Finally, a word of caution. Whenever we use mathematics to try to draw real-world conclusions
from data, these conclusions should be taken with an extra-large grain of salt. People have an outsized
tendency to trust mathematics and to take its results as infallible. We therefore have a special responsibility
not to overstate any conclusions, and to point out potential pitfalls with the analysis. No matter how
wonderful principal complement analysis is, we must keep in mind that what we are still only looking
at a 2-dimensional projection of a 200-dimensional space. Therefore it is inevitable that lots of details
and nuances are lost. We could get a completely different picture by looking at a different 2-dimensional
projection.
To see how the data can sometimes be misleading, consider the question “How similar is Senator
Tim Johnson, Democrat of South Dakota, to Senators Olympia Snowe and Susan Collins of Maine?”.
In the 1-dimensional picture, it looked as if they were very similar. We could easily rationalize this by
pointing out that Johnson is the most conservative Democrat, and Snowe and Collins are the most liberal
 11.12. Application: Principal component analysis 483

Republicans. However, the 2-dimensional picture reveals an interesting nuance, which is that the voting
records of Johnson is not all that similar to that of Snowe and Collins. It is entirely possible that if we add
a third or fourth dimension to the picture, many more additional such details will emerge. In summary,
while principal component analysis is a useful tool, it is just one tool among many, and we always need to
exercise our best judgement in drawing conclusions from data.

Exercises

Exercise 11.12.1 Consider the following collection of points in R2 :

                   
1 −2 3 2 −2 3 −3 0 0 0
, , , , , , , , , .
−1 1 −2 −2 1 −1 0 3 0 −2
Find the 1-dimensional subspace that best approximates this collection of points. What is the total squared
distance of the points to the subspace? Sketch the subspace and the points.

Exercise 11.12.2 Consider the following collection of points in R3 :

                      
 0 1 0 0 6 1 0 0 2 4 1 0 
 0  ,  0  ,  0   5  ,  −2  ,  0  ,  2  ,  1  ,  0  ,  0  ,  −3  ,  0  .
 
9 −3 1 −4 13 0 0 1 0 0 0 4

(a) Find the 1-dimensional subspace that best approximates this collection of points.

(b) Find the 2-dimensional subspace that best approximates this collection of points.

(c) What is the 3-dimensional subspace that best approximates this collection of points?

In each case, what is the total squared distance of the points to the subspace?

Exercise 11.12.3 Find the principal components of the matrix

 
0 0 0 0 −1 0 3 0 1
A= 0 1 0 0 2 3 0 2 −1  .
1 −1 −3 3 −1 0 0 −2 2

Exercise 11.12.4 Compute the centroid of the following collection of points:

                   
2 0 4 −4 0 −2 0 −2 3 −5
, , , , , , , , , .
−1 1 −3 −4 −2 −1 −2 −4 −1 −5

Exercise 11.12.5 Consider the following collection of points in R2 :

                   
4 10 4 −2 10 4 5 4 −2 3
, , , , , , , , , .
−3 −9 −7 3 −3 −3 −3 1 −3 −3
484 Inner product spaces

Compute the centroid, and then find the 1-dimensional affine subspace that best approximates this collec-
tion of points. What is the total squared distance of the points to the subspace?

Exercise 11.12.6 Consider the following collection of points in R3 :

              

 0 6 2 3 3 5 4 5 1 3 3 3 1 3 

              
 −2 ,  2 ,  2 ,  2 ,  0 ,  0 ,  −2 , 0 , 0 , 0 ,  0 ,  0 ,  0 ,  −2  .
              

 

2 2 6 2 5 2 −2 2 2 −1 1 3 2 2

Find the 1- and 2-dimensional affine subspaces that best approximate this collection of points. What is the
total squared distance of the points to each subspace?
A. Complex numbers
Throughout history, mankind has invented more and more complicated number systems in an effort to
make algebra easier.

• In the beginning, there were the natural numbers N = {1, 2, 3, . . .}. However, after a while, it
became a problem that certain equations, such as x + 5 = 3, do not have a solution in the natural
numbers.

• To solve this problem, zero and negative numbers were invented, resulting in the set of integers
Z = {. . . , −3, −2, −1, 0, 1, 2, 3, . . .}. In the integers, the equation x + 5 = 3 has a solution, namely
x = −2. But some other equations, such as 2x = 1, still do not have a solution in the integers.

• To solve this problem, the rational numbers Q were invented. In the rational numbers, the equation
2x = 1 has a solution, namely x = 21 . However, some other equations, such as x2 = 2, still do not
have a solution in the rational numbers.

• To solve this problem, the real

√ numbers R were invented. In the real numbers, the equation x2 = 2
has a solution, namely x = 2. However, some other equations, such as x2 = −1, still don’t have a
solution in the real numbers.

• To solve this problem, the complex numbers were invented.

The purpose of this section is to summarize the most important facts about the complex numbers. We
will also see that the above process does not continue. In the complex numbers, all non-trivial polynomial
equations have a solution, and therefore, no additional numbers are “missing”. This property is called the
fundamental theorem of algebra. Gauss is usually credited with giving a proof of this theorem in 1797
but many others worked on it and the first completely correct proof was due to Argand in 1806.

A.1 The complex numbers

Outcomes
A. Add, subtract, multiply, and divide complex numbers.

B. Find the conjugate and the magnitude of a complex number.

C. Apply algebraic properties of the complex numbers to simplify equations.

485
486 Complex numbers

Definition A.1: The complex numbers

Let i be an imaginary number such that i 2 = −1. A complex number is a number of the form

z = a + bi,

where a and b are real numbers. The set of all complex numbers is denoted C.

The form z = a + bi is called the standard form or Cartesian form of the complex number z. We refer to
a as the real part and to b as the imaginary part of z.
Addition, subtraction, and multiplication of complex numbers are defined in the obvious way, keep-
ing in mind that i 2 = −1. Namely, we have

(a + bi) + (c + di) = (a + c) + (b + d)i,

(a + bi) − (c + di) = (a − c) + (b − d)i,
(a + bi)(c + di) = ac + adi + bci + bdi 2 = (ac − bd) + (ad + bc)i.

Example A.2: Addition, subtraction, and multiplication of complex numbers

• (3 + 5i) + (2 − 3i) = (3 + 2) + (5 − 3)i = 5 + 2i.

• (3 + 5i) − (2 − 3i) = (3 − 2) + (5 + 3)i = 1 + 8i.

• (3 + 5i)(2 − 3i) = 6 − 9i + 10i − 15i 2 = 21 + i.

Division of complex numbers is more complicated. We first note that it is easy to divide a complex number
by a real number. Namely,
a + bi a b
= + i.
r r r
But how can we divide by a complex number? We use the following trick. Let z = a + bi be a complex
number, and consider the product (a + bi)(a − bi). It is equal to

(a + bi)(a − bi) = a2 − b2 i 2 = a2 + b2 .

Therefore, (a + bi)(a − bi) is always a real number, and therefore easy to divide by. Therefore, we can
compute the multiplicative inverse of a complex number z = a + bi as follows:
1 1 1 a − bi a − bi
z−1 = = = = 2 .
z a + bi a + bi a − bi a + b2

Example A.3: Inverse of a complex number

1 2 − 5i 2 5
= 2 2
= − i.
2 + 5i 2 + 5 29 29
 A.1. The complex numbers 487

2 5
You should verify that this is indeed the inverse, by multiplying 2 + 5i by 29 − 29 i and checking that the
answer is indeed 1.
Division of complex numbers can then be defined in terms of the multiplicative inverse, i.e., wz = zw−1 .

Example A.4: Division of complex numbers

5 + 7i (5 + 7i)(3 + 4i) −13 + 41i 13 41
= = 2 = − + i.
3 − 4i (3 − 4i)(3 + 4i) 3 + 42 25 25

As a special case of division, note that

1 1(−i)
i −1 = = = −i.
i i(−i)

The complex numbers form a field, i.e., they satisfy the nine field axioms. See Section 1.8 for the
definition of a field.

Proposition A.5: The complex numbers form a field

The complex numbers, with the operations of addition, subtraction, multiplication, and division,
form a field. Specifically, this means that they satisfy the following properties:

(A1) Commutative law of addition: z + w = w + z;

(A2) Associative law of addition: (z + w) + u = z + (w + u);

(A3) Unit law of addition: 0 + z = z;

(A4) Additive inverse: z + (−z) = 0;

(M1) Commutative law of multiplication: zw = wz;

(M2) Associative law of multiplication: (zw)u = z(wu);

(M3) Unit law of multiplication: 1z = z;

(M4) Multiplicative inverse: when z is non-zero: zz−1 = 1;

(D) Distributive law: z(w + u) = zw + zu.

Another useful operation on complex numbers is the complex conjugate. Let z = a + bi be a complex
number. Then the conjugate of z, written z, is given by

z = a − bi.

Note that if z = a + bi is a complex number, then zz = (a + bi)(a − bi) = a2 + b2 . Therefore, zz is always

a real number and zz ≥ 0. We define the magnitude of z to be
√ p
|z| = zz = a2 + b2 .
488 Complex numbers

The magnitude is also sometimes called the absolute value or the modulus of the complex number.

Example A.6: Conjugate and magnitude of a complex number

• 3 + 5i = 3 − 5i.

• i = −i.

• 7 = 7.
√ √
• |3 + 5i| = 32 + 52 = 34.

• |i| = 1.

• |−6| = 6.

√
Note that for a real number a, we have a = a. Also, when a is real, the magnitude |a| = a2 is just the
usual absolute value of real numbers. The following two propositions list some basic properties of the
conjugate and of the magnitude.

Proposition A.7: Properties of the conjugate

Let z and w be complex numbers. Then, the following properties of the conjugate hold.

• z ± w = z ± w.

• (zw) = z w.

• z−1 = z−1 .

• z/w = z/w.

• z = z.

• z is real if and only if z = z.

Proposition A.8: Properties of the magnitude

Let z, w be complex numbers. The following properties hold.

• |z| ≥ 0, and |z| = 0 if and only if z = 0.

• |zw| = |z| |w|.

• |z| = |z|.

• |z/w| = |z| / |w|.

• Triangle inequality: |z + w| ≤ |z| + |w|.

 A.2. Geometric interpretation 489

Exercises

Exercise A.1.1 Let z = 2 + 7i and let w = 3 − 8i. Compute z + w, z − 2w, zw, and wz .

Exercise A.1.2 Let z = 1 − 4i. Compute z, z−1 , and |z|.

Exercise A.1.3 Let z = 3 + 5i and w = 2 − i. Compute zw, |zw|, and z−1 w.

Exercise A.1.4 Use the properties of complex numbers to prove that if z is a complex number, then there
exists a complex number w with |w| = 1 and wz = |z|.

Exercise A.1.5 I claim that 1 = −1. Here is why.

√ √ q √
2
−1 = i = −1 −1 = (−1)2 = 1 = 1.

What is wrong with this argument?

A.2 Geometric interpretation

Outcomes
A. View complex numbers as points in the plane.

B. Understand the geometric meaning of addition, subtraction, multiplication, and the complex
conjugate.

C. Understand the geometric meaning of the magnitude and argument of a complex number.

Just as a real number can be considered as a point on the line, a complex number z = a + bi can be
considered as a point (a, b) in the plane whose x coordinate is a and whose y coordinate is b. For example,
in the following picture, the complex number z = 3 + 2i can be represented as the point in the plane with
coordinates (3, 2).
2i z = 3 + 2i
r
i
θ
−1 0 1 2 3
490 Complex numbers

The magnitude r = |z| of a complex number is its distance from the origin. We also define the argument
of z to be the angle θ between the x-axis and the line from the origin to z, counted positively in the
counterclockwise direction. The magnitude r and argument θ are shown in the above picture.
Addition of complex numbers is like vector addition. The effect of multiplying two complex numbers
is to multiply their magnitudes and add their arguments. For example, the following picture illustrates the
multiplication (3 + 2i)(1 + i) = 1 + 5i.

5i 1 + 5i
4i
3i
2i 3 + 2i 2i 2i
i i 1+i i

−1 0 1 2 3 −1 0 1 2 3 −1 0 1 2 3

The number 3 + 2i The number 1 + i The product (3 + 2i)(1 + i) = 1 + 5i

To take the multiplicative inverse of a complex number, we take the reciprocal of the magnitude and negate
the argument.
i
z

θ
−θ
1
1
r z−1
The effect of taking the complex conjugate is to reflect the given complex number about the x axis (or
equivalently, keep the magnitude unchanged and negate the argument).

i
z

θ
−θ
1
r

z̄
 A.3. The fundamental theorem of algebra 491

Exercises

Exercise A.2.1 Draw the complex numbers z = 2 + i and w = −2 + 3i as points in the plane. Then use
the geometric interpretation to find z + w, z − w, zw, z−1 , z, and |z|.

Exercise A.2.2 Use the geometric interpretation to find a complex number z such that z2 = i. Can you
find two such numbers?

Exercise A.2.3 Use the geometric interpretation to find 3 different complex numbers z such that z3 = −1.
Hint: these numbers will lie on the unit circle.

A.3 The fundamental theorem of algebra

Outcomes
A. Find the complex roots of a quadratic polynomial.

B. In special cases, find the complex roots of a polynomial of degree 3 or more.

C. Factor a polynomial into linear factors.

The complex numbers were invented so that equations such as z2 +1 = 0 would have solutions. In fact, this
equation has two complex solutions, namely z = i and z = −i. However, something much more general
(and surprising) is true: every non-trivial polynomial equation has a solution in the complex numbers. To
understand this statement, recall that a polynomial is an expression of the form

p(z) = an zn + an−1 zn−1 + . . . + a1 z + a0 .

The constants a0 , . . . , an are called the coefficients of the polynomial. If an is the largest non-zero coeffi-
cient, we say that the polynomial has degree n. A polynomial of degree 0 is of the form p(z) = a0 , and is
also called a constant polynomial. Recall that a root of a polynomial is a number z such that p(z) = 0.
The fundamental theorem of algebra is the following:

Theorem A.9: Fundamental theorem of algebra

Every non-constant polynomial p(z) with real or complex coefficients has a complex root.

The proof of this theorem is beyond the scope of this book. Note that the theorem does not say that the
roots are always easy to find. To find the roots of a polynomial of degree 2, we can use the quadratic
formula. However, if the degree is greater than 2, we may sometimes have to use fancier methods, such
492 Complex numbers

as Newton’s method from calculus, or even a computer algebra system, to locate the roots. We give some
examples.

Example A.10: Roots of a quadratic polynomial

Find the roots of the polynomial p(z) = z2 − 2z + 2.

Solution. The quadratic formula gives √

2 ± −4
z= .
2
Of course, in the real numbers, the square root of −4 does not exist, so p(z) has no roots in the real
numbers. However, in the complex numbers, the square root of −4 exists and is equal to ±2i. Thus, the
roots of p(z) are:
2 ± 2i
z= = 1 ± i.
2
Indeed, we can double-check that 1 + i and 1 − i are in fact roots:
p(1 + i) = (1 + i)2 − 2(1 + i) + 2 = (1 + 2i + (−1)) − 2 − 2i + 2 = 0,
p(1 − i) = (1 − i)2 − 2(1 − i) + 2 = (1 − 2i + (−1)) − 2 + 2i + 2 = 0.
♠
Example A.11: Roots of a cubic polynomial
Find the roots of the polynomial p(z) = z3 − 4z2 + 9z − 10.

Solution. By the intermediate value theorem of calculus, we know that a cubic polynomial with real
coefficients always has at least one real root. This is because p(z) goes to −∞ when z → −∞ and to ∞
when z → ∞. By trial and error, we find that z = 2 is a root of this polynomial. We can therefore factor out
(z − 2) from this polynomial:
p(z) = z3 − 4z2 + 9z − 10 = (z − 2)(z2 − 2z + 5).
Now we can use the quadratic formula to find the roots of z2 − 2z + 5. We find
√
2 ± −16 2 ± 4i
z= = = 1 ± 2i.
2 2
Thus, the three complex roots of p(z) are z = 2, z = 1 + 2i, and z = 1 − 2i. ♠
The following proposition is an important and useful consequence of the fundamental theorem of algebra:

Proposition A.12: Factoring a polynomial

Let p(z) be a polynomial of degree n with real or complex coefficients. Then p(z) can be factored
into n linear factors over the complex numbers, i.e., p(z) can be written in the form

p(z) = a(z − b1 )(z − b2) · · · (z − bn ),

where b1 , . . . , bn are (not necessarily distinct) roots of p(z).

 A.3. The fundamental theorem of algebra 493

Proof. If n = 0, then p(z) = a and there is nothing to show. Otherwise, by the fundamental theorem of
algebra, p(z) has at least one complex root, say b1 . From calculus, we know that we can factor out (z − b1 )
from p(z), i.e., we can find a polynomial q(z) of degree n − 1 such that

p(z) = (z − b1 )q(z),

We can repeatedly apply the same procedure to q(z) until p(z) has been factored into linear factors. ♠

Example A.13: Factoring a polynomial

Factor p(z) = z3 − 4z2 + 9z − 10 into linear factors.

Solution. From Example A.11, we know that p(z) has three distinct roots b1 = 2, b2 = 1 + 2i, and b3 =
1 − 2i. We can therefore write
p(z) = a(z − b1 )(z − b2 )(z − b3).
Since the leading term is z3 , we find that a = 1. Therefore

p(z) = (z − 2) (z − 1 − 2i) (z − 1 + 2i).

Exercises

Exercise A.3.1 Find the roots of the polynomial p(z) = z2 − 6z + 13.

Exercise A.3.2 Find the roots of p(z) = z3 + 5z2 + 4z − 10. Hint: one of the roots is z = 1.

Exercise A.3.3 Factor z3 − 3z + 2 into a product of linear factors.

Exercise A.3.4 Let p(x) = an xn + an−1 xn−1 + . . . + a1 x + a0 be a polynomial with real coefficients, i.e.,
such that all the ak are real numbers. Suppose that z is a root of p. Show that z is also a root of p.
B. Answers to selected exercises
x + 3y = 1  10 1

1.1.1 , Solution is: x = 13 , y = 13 .
4x − y = 3

3x + y = 3
1.1.2 , Solution is: [x = 1, y = 0]
x + 2y = 1

1.1.4

1.2.1 All of them except (c) and (e).

1.2.2 (a) is a solution to equations 1 and 2, (b) is a solution to equations 2 and 3, (c) is a solution to
equations 1, 2, and 3, (d) is a solution to 1 and 3, and (e) is a solution to equations 1, 2, and 3. Only (c)
and (e) are solutions to the system of equations.

1.3.1 (x, y) = (1, 0).

x + 3y = 1 10 1


1.3.2 , Solution is: (x, y) = 13 , 13
4x − y = 3

x + 2y = 1

3 1
1.3.3 2x − y = 1 , Solution is: (x, y) = 5, 5
4x + 3y = 3

x + y − 3z = 2 x + 4z = 0
1.3.4 No solution exists. 2x + y + z = 1 , after elementary operations: y − 7z = 0 . Thus one of
3x + 2y − 2z = 0 0 =1
the equations says 0 = 1 in an equivalent system of equations.

1.3.5 (x, y, z) = (4, 1, 1).

 
x + 2y − 3z = 5 1 2 −3 5
1.3.6 −x − y + z = −2 As an augmented matrix:  −1 −1 1 −2  .
x − 3y = −3. 1 −3 0 −3

495
496 Answers to selected exercises

4g − I = 150
4I − 17g = −660
1.3.7 , Solution is : {g = 60, I = 90, b = 200, s = 50}
4g + s = 290
g+I +s−b = 0

1.4.1 The solution exists but is not unique.

1.4.2 A solution exists and is unique.

1.4.4 There might be a solution. If so, there are infinitely many.

1.4.5 No. Consider x + y + z = 2 and x + y + z = 1.

1.4.6 These can have a solution. For example, x + y = 1, 2x + 2y = 2, 3x + 3y = 3 even has an infinite set
of solutions.

1.4.7 h = 4.

1.4.8 Any h will work.

1.4.9 Any h will work.

1.4.10 If h 6= 2 there will be a unique solution for any k. If h = 2 and k 6= 4, there are no solutions. If h = 2
and k = 4, then there are infinitely many solutions.

1.4.11 If h 6= 4, then there is exactly one solution. If h = 4 and k 6= 4, then there are no solutions. If h = 4
and k = 4, then there are infinitely many solutions.

1.4.12
 There is no solution.
 The system is inconsistent. You can 
see this from the augmented matrix.
1 2 1 −1 2 1 2 1 −1 2
 1 −1 1 1 1   0 −3 0 2 −1 
 , echelon form:  .
 2 1 −1 0 1   0 0 −3 0 −2 
4 2 1 0 5 0 0 0 0 1

1.4.13 The solution is: x = 13 − 13 t, y = 23 + 23 t, z = 13 , w = t.

1.4.14 (a) Yes. (b) No. (c) Yes.

 
1 1
1 0 2 2
 
1.4.16 (b) The echelon form is   0 1 . Therefore, the solution is of the form z = t, y =
 − 41 3
4
0 0 0 0
3 1

1 1
4 + t 4 , x = 2 − 2 t, where t ∈ R.
 
1 0 4 2
(c) The echelon form is and so the solution is z = t, y = −1 + 4t, x = 2 − 4t.
0 1 −4 −1
 497

 
1 0 0 0 9 3
 0 1 0 0 −4 0 
(d) The echelon form is  
 0 0 1 0 −7 −1  and so x5 = t, x4 = 1 − 6t, x3 = −1 + 7t, x2 = 4t,
0 0 0 1 6 1
x1 = 3 − 9t.
 
1 0 2 0 − 12 5
2
 1 3 
 0 1 0 0 
(e) The echelon form is  2 2 . Therefore, let x = t, x = s. Then the other
3 1  5 3
 0 0 0 1 2 −2 
0 0 0 0 0 0
variables are given by x4 = − 2 − 2 t, x2 = 2 − t 2 , x1 = 25 + 12 t − 2s.
1 3 3 1

1.4.17 Solution is: [x = 1 − 2t, z = 1, y = t].

1.4.18 Solution is: [x = 2 − 4t, y = −8t, z = t].

1.4.19 Solution is: [x = −1, y = 2, z = −1].

1.4.20 Solution is: [x = 2, y = 4, z = 5].

1.4.21 Solution is: [x = 1, y = 2, z = −5].

1.4.22 No. Consider x + y + z = 2 and x + y + z = 1.

1.4.23 No. This would lead to 0 = 1.

1.4.24 Yes. It has a unique solution.

1.4.25 The last column must not be a pivot column. The remaining columns must each be pivot columns.

1
4 (20 + 30 + w + x) − y = 0
1
1.4.26 You need 4 (y + 30 + 0 + z) − w = 0
. Solution is: [x = 15, y = 20, z = 10, w = 15].
1
4 (20 + y + z + 10) − x = 0
1
4 (x + w + 0 + 10) − z = 0

1.4.28 The rank is the number of pivot entries in the echelon form. There is at most one pivot entry in each
row and column. Therefore, the rank cannot be larger than the number of rows or the number of columns;
in other words, the rank is at most min(m, n).

1.4.29 (a) The echelon form has 4 non-zero rows and 6 columns, so there are 2 free variables, and the
system has infinitely many solutions.

(b) Such a system of equations of equations does not exist. If you add in another column, the rank does
not get smaller.
498 Answers to selected exercises

(c) Such a system of equations does not exist, because the rank cannot equal 4 if there are only two
columns.

(d) The echelon form has 4 non-zero rows on the left-hand side, but 5 non-zero rows if we also include
the right-hand side. Therefore the system is inconsistent, i.e., it has no solutions.

(e) The echelon form has 2 non-zero rows, so there are 2 pivot variables and no free variables. There-
fore, the system has a unique solution.

1.4.30 These are not legitimate row operations. They do not preserve the solution set of the system.

1.5.1 (a) This one is not.

(b) This one is.

(c) This one is.

1.5.3 Solution is: [x = −1, y = −5, z = 4]

1.5.4 Solution is: [x = 2t + 1, y = 4t, z = t]

1.5.5 Solution is: [x = 1, y = 5, z = 3]

1.5.6 Solution is: [x = 4, y = −4, z = −2]

1.5.7 The rank of the coefficient matrix is 3, so both systems have a unique solution. The solution of the
first system is (x, y, z) = (1, 0, 1), and the solution of the second system is (x, y, z) = (1, 1, 2).

1.6.2 The systems (b), (c), and (d) have non-trivial solutions, because the rank is less than the number of
variables.

1.6.3 The general solution is (x, y, z) = (1, 2, 4) + s(1, 0, 1) + t(0, 1, −1), or equivalently, (x, y, z) = (1 +
s, 2 + t, 4 + s − t), where s and t are parameters.
 
1 0 1 1
1.8.6 (a) Reduced echelon form: . General solution: x = 1 − t, y = 1 − it, z = t.
0 1 i 1
(b) Solution: x = 1, y = 1 + i, z = i.

1.11.1 The other three equations are

4I1 + I1 − I4 + 2I1 − 2I2 = −10

6I3 − 6I4 + I3 + I3 + 5I3 − 5I2 = −20
2I4 + 3I4 + 6I4 − 6I3 + I4 − I1 = 0.
 499

Then the system is

2I2 − 2I1 + 5I2 − 5I3 + 3I2 = 5
4I1 + I1 − I4 + 2I1 − 2I2 = −10
6I3 − 6I4 + I3 + I3 + 5I3 − 5I2 = −20
2I4 + 3I4 + 6I4 − 6I3 + I4 − I1 = 0.
The solution is:
750
I1 = −
373
1421
I2 = −
1119
3061
I3 = −
1119
1718
I4 = − .
1119

1.11.2 We have

2I1 + 5I1 + 3I1 − 5I2 = 10

I2 − I3 + 3I2 + 7I2 + 5I2 − 5I1 = −12
2I3 + 4I3 + 4I3 + I3 − I2 = 0.

Simplifying this yields

10I1 − 5I2 = 10
−5I1 + 16I2 − I3 = −12
−I2 + 11I3 = 0.

The solution is given by

218 154 14
I1 = , I2 = − , I3 = − .
295 295 295

2.1.2    
5−2 3
−→ − → − →
PQ = 0Q − 0P =  −2 − 0  =  −2 
1 − (−4) 5
   
2−5 −3
−→ − → − →
QP = 0P − 0Q =  0 − (−2)  =  2 
(−4) − 1 −5

2
2.1.3 We need 5x − 3y = 2x − 2y and 4 = 2y. The unique solution is x = 3 and y = 2.
 
1
2.2.1  9 .
0
500 Answers to selected exercises

2.2.2 (a) (u + v) + w = (v + u) + w = v + (u + w) = v + (w + u) = (v + w) + u. Here we have used the

commutative law, the associative law, the commutative law, and the associative law, in that order.

(b) (u + 0) + (v + (−u)) = u + (v + (−u)) = u + ((−u) + v) = (u + (−u)) + v = 0 + v = v + 0 = v.

Here we have used the additive unit, the commutative law, the associative law, the additive inverse,
the commutative law, and the additive unit, in that order.

2.3.1

u
− 21 v
v
−u
2v

 
−55
 13 
2.3.2  
 −21 .
39

2.3.3 (a) (k + ℓ)(u + v) = (k + ℓ)u + (k + ℓ)v = ku + kv + ℓu + ℓv. Here we used the distributive law
over vector addition in the first step, and the distributive law over scalar addition in the second step.

(b) We have 0u = (0 + 0)u = 0u + 0u by properties of scalars and by the distributive law, respectively.
Adding −(0u) to both sides of the equation, and using the additive unit law and associativity, we
have 0 = 0u.

(c) We have (−1)u = (−1)u + 0 = (−1)u + (u + (−u)) = ((−1)u + u) + (−u) = ((−1)u + 1u) +
(−u) = ((−1) + 1)u + (−u) = 0u + (−u) = 0 + (−u) = −u. Here, we have used the additive unit
law, the additive inverse law, the associative law, the rule for multiplication by 1, the distributive
law, properties of scalars, the property of part (b), and the additive unit law, respectively.

2.4.2      
4 3 2
 4  = 2  1  −  −2  .
−3 −1 1

2.4.3 The system        

4 3 8 2
 4  = a1  1  + a2  0  + a3  −2 
4 −1 −1 1
has no solution.
√
2.5.1 17.
√
2.5.2 27.
 501

2.5.3 The form the plane x + y + z = 0.

2.5.4 They form a sphere of radius 1, given by the equation x2 + y2 + z2 = 1.

√ √ √
2.5.5 kuk = 13, kvk = 30, kwk = 22.

2.5.7 This follows from the last property of Proposition 2.20: k−uk = k(−1)uk = |−1| kuk = kuk.

2.5.8 Only w is a unit vector.

2.5.9 
  5
  −2
1 1 1 1 1 1 1  −3 
u= √ , v= √  3 , w=  .
kuk 5 2 kvk 17 kwk 6 1 
2
−1
   
1 2
 2   0 
2.6.1   
·
 3   1
 = 17.

4 3
 T  T
2.6.4 cos θ =
3 −1 −1
√ √ ·
1 4 2 −3
=√ √ .
9 + 1 + 1 1 + 16 + 4 11 21

−10
2.6.5 cos θ = √ √ .
1 + 4 + 1 1 + 4 + 49

·
2.6.6 This formula says that u v = kuk kvk cos θ where θ is the included angle between the two vectors.
Thus
·
ku vk = kuk kvk kcos θ k ≤ kuk kvk
and equality holds if and only if θ = 0 or π . This means that the two vectors either point in the same
direction or opposite directions. Hence one is a multiple of the other.
−
→ − → − →
2.6.7 This triangle has sides AB, BC, CA. The direction of the sides does not matter, since scalar multipli-
cation preserves orthogonality.
−→
PQ = (5, −2, 1) − (2, 0, 3) = (3, −2, 4),
−→
QR = (7, 5, 3) − (5, −2, 1) = (2, 7, 2),
−→
RP = (2, 0, −3) − (7, 5, 3) = (−5, −5, −6).

If any pair of these is orthogonal then we are done:

−
→ − →
AB · BC = (3, −2, 4) · (2, 7, 2) = 6 − 14 + 8 = 0.

The triangle has a right angle at B.

502 Answers to selected exercises


  
5
1 − 14
2.6.8
v w·v=
−5    5 
 2  =  − 7 .
v v· 14
3 − 15
14


  
1 − 12
2.6.9
v w·v=
−5   
 0  =  0 .

v v· 10
3 − 32
   1

h iT h iT 1 − 14

2.6.10
v w·v=
1 2 3 0 · 1 2 −2 1 


2  
 
=
2
− 14 

.
v v· 1+4+9  3   3
− 14 
0 0

2.6.13 No, it does not. The vector 0 has no direction. The formula for proj0 (w) doesn’t make sense either.

2.6.14    
−3/2 9/2
a =  3/2  , b =  1/2  .
−3 −2

2.6.15 ! !
u− ·
v u
2
v · u− ·
v u
2
v = kuk2 − 2(v u)2 · 1
2
+ (v u)2· 1
≥ 0,
kvk kvk kvk kvk2
and so
kuk2 kvk2 ≥ (v u)2 . ·
We get equality exactly when u = projv (u) =
v u ·
v, or in other words, when u is a multiple of v.
kvk2

2.6.16

w − projv (w) + u − projv (u) = w + u − (projv (w) + projv (u))

= w + u − projv (w + u).

This follows because

projv (w) + projv (u) = · v + v·u v

v w
2 2
kvk kvk
v·(w + u)
= v
kvk2
= projv (w + u).

u·v
·
2.6.17 u (v−proju (v)) = u v −
kuk2
· · · ·
u u = u v −u v = 0. Therefore, we can write v = (v−proju (v)) +
proju (v). The first is orthogonal to u and the second is a multiple of u so is parallel to u.
 503

2.7.1 (a) Left-handed. (b) Right-handed. (c) Right-handed. (d) Left-handed.

2.7.2 If a 6= 0, then the condition says that ka × uk = kak sin θ = 0 for all angles θ . Hence a = 0.
     
1 3 8 √
2.7.3  2  ×  −2  =  8 . The area of the parallelogram is 8 3.
3 1 −8
     
1 4 6 √ √
2.7.4  0  ×  −2  =  11 . The area is of the parallelogram is 36 + 121 + 4 = 161.
3 1 −2
 
4
−→ − →
2.7.5 Let P = (−2, 3, 1), Q = (2, 1, 1), R = (1, 2, −1), and S = (5, 0, −1). We have PQ × PR =  −2  ×
0
   
3 5 −→
 −1  =  8 . The area of the parallelogram is −→ √ 2 √
PQ × PR = 5 + 82 + 22 = 93.
−2 2
    
3 −4 1
−→ − →
2.7.6 Let P = (1, 0, 3), Q = (4, 1, 0), and R = (−3, 1, 1). PQ × PR =  1  ×  1  =  18 . The
−3 −2 7
1 −→ −

→ 1 p 1 √
area of the triangle is PQ × PR = 1 + 182 + 72 = 374.
2 2 2
     
1 2 0
−→ − →
2.7.7 Let P = (1, 2, 3), Q = (2, 3, 4), and R = (3, 4, 5). PQ × PR =  1  ×  2  =  0 . The area of
1 2 0
the triangle is 0. It means the three points are on a line.

2.7.8 (i × j) × j = k × j = −i. However, i × (j × j) = 0 and so the cross product is not associative. The
expression u × v × w has no meaning.

2.7.10 Let u = [u1 , u2 , u3 ]T , v = [v1 , v2 , v3 ]T , and w = [w1 , w2 , w3 ]T . Then

     
u1 v2 w3 − v3 w2 u2 (v1 w2 − v2 w1 ) − u3 (v3 w1 − v1 w3 )
u × (v × w) =  u2  ×  v3 w1 − v1 w3  =  u3 (v2 w3 − v3 w2 ) − u1 (v1 w2 − v2 w1 ) 
u3 v1 w2 − v2 w1 u1 (v3 w1 − v1 w3 ) − u2 (v2 w3 − v3 w2 )
 
u2 v1 w2 − u2 v2 w1 − u3 v3 w1 + u3 v1 w3
=  u3 v2 w3 − u3 v3 w2 − u1 v1 w2 + u1 v2 w1 
u1 v3 w1 − u1 v1 w3 − u2 v2 w3 + u2 v3 w2

and
   
v1 w1
· ·
(u w) v − (u v) w = (u1 w1 + u2 w2 + u3 w3 )  v2  − (u1 v1 + u2 v2 + u3 v3 )  w2 
v3 w3
504 Answers to selected exercises

 
u1 w1 v1 + u2 w2 v1 + u3 w3 v1 − u1 v1 w1 − u2 v2 w1 − u3 v3 w1
=  u1 w1 v2 + u2 w2 v2 + u3 w3 v2 − u1 v1 w2 − u2 v2 w2 − u3 v3 w2  .
u1 w1 v3 + u2 w2 v3 + u3 w3 v3 − u1 v1 w3 − u2 v2 w3 − u3 v3 w3
We can see by careful inspection that these are equal.

2.7.11
u × (v × w) + v × (w × u) + w × (u × v)
· · · · ·
= ((u w) v − (u v) w) + ((v u) w − (v w) u) + ((w v) u − (w u) v) ·
· · · · ·
= ((u w) v − (w u) v) + ((v u) w − (u v) w) + ((w v) u − (v w) u)
= 0.
·
         
1 −7 −5 8 −5

2 3
·
2.7.12  1  ×  −2   −6  =  −23   −6  = 113.
3 5 3
·
2.7.13 · ·
(a) We have [u, v, w] = (u × v) w = [−2, 1, 2]T [0, 0, 1]T = 2, so the box product is positive.
This means that the system of vectors u, v, w is right-handed.

· ·
(b) We have [u, v, w] = (u × v) w = [−2, −1, 1]T [1, 1, 2]T = −1, so the box product is negative and the
system is left-handed.

· ·
(c) We have [u, v, w] = (u × v) w = [−1, −1, 1]T [3, 1, 4]T = 0, so the box product is zero. This means
that the volume of the parallelepiped spanned by u, v, w is empty, i.e., the vectors are coplanar.

· ·
(d) We have [u, v, w] = (u × v) w = [0, 0, 2]T [2, 0, −1]T = −2, so the box product is negative and the
system is left-handed.

2.7.14 Yes. It will involve the sum of a product of integers and so it will be an integer.

2.7.15 It means that if you place them so that they all have their tails at the same point, the three will lie
in the same plane.

2.7.17 First, we have

· ·
(v × w) × (w × z) = ((v × w) z) w − ((v × w) w) z.
Since v × w is orthogonal to w, their dot product (v × w)·w is zero, and therefore
(v × w) × (w × z) = ((v × w)·z) w.
But then

· · · ·
(u × v) ((v × w) × (w × z)) = (u × v) (((v × w) z) w) = ((v × w) z) ((u × v) w). ·
2.7.18 We have
ku × vk2 = kuk2 kvk2 sin2 θ
= kuk2 kvk2 (1 − cos2 θ )
= kuk2 kvk2 − kuk2 kvk2 cos2 θ
= kuk2 kvk2 − (u v)2 , ·
which implies the expression equals 0.
 505

3.1.6 We have          
x 1 1 3 −1
 y  =  2  + (2 − s)  0  =  2  + s  0  .
z 0 1 2 −1

3.1.15 From trigonometry, we have the following properties of the cosine function:

• cos θ is positive for 0 ≤ θ ≤ π2 , and negative for π

2 ≤ θ ≤ π.

• cos(π − θ ) = − cos θ .

By the method in Example 3.10, we calculated θ such that

cos θ = ·
u v
.
kuk kvk

If 0 ≤ θ ≤ π2 , the answer is θ . If π
2 ≤ θ ≤ π , the answer is φ = π − θ . But in the last case, the dot product
is negative and we have

cos φ = cos(π − θ ) = − cos θ = ·

−u v
= ·
|u v|
.
kuk kvk kuk kvk
So in either case we get the correct answer by taking the absolute value.

3.2.2 We have
             
x 1 1 −1 2 −2 −1
 y   2      
−1   2     
    + (1 − r1 )  0 
+ (r1 + r2 )   + r1  −1  + r2  −1  .
 z = 0   0   1 = 0   1   1 
w 0 1 0 1 −1 0

3.2.8 The general solution of ax + by + cz + dw = e involves at least three parameters. But the vector
equation of a plane only has two parameters, therefore ax + by + cz + dw = e does not describe a plane. (It
describes a 3-dimensional so-called hyperplane inside R4 ).

3.2.13 (a) If the dot product is negative, φ will be greater than π2 , and therefore θ will end up being
negative. We can fix this, similarly to Exercise 3.1.15, by taking the absolute value of the dot product, i.e.,
by solving
cos φ =
|n d| · (2.1)
knk kdk
(b) From trigonometry, we know that sin θ = sin( π2 − φ ) = cos φ . Together with (2.1), this gives the desired
formula.

·
3.2.15 x (a × b) = 0.

4.1.1 Yes, a 1×1-matrix is both a row vector and a column vector. However, a column vector of dimension
2 or greater can never be equal to a row vector, because one is an n × 1-matrix and the other is a 1 × n-
matrix.
506 Answers to selected exercises

4.1.2 x = 2, y = −1, z = 2.

4.1.3 2 × 3, 3 × 3, 4 × 2.

4.1.4 7.
 
0 6 0
4.2.3 The equation simplifies to X = B + B, so X = .
2 −2 2

4.2.4 Suppose A + B = 0. Then we have −A = (−A) + 0 = (−A) + (A + B) = ((−A) + A) + B = (A +

(−A)) + B = 0 + B = B + 0 = B. Here, we have used the additive unit law, the assumption A + B = 0,
associativity, commutativity, the additive inverse law, commutativity, and the additive unit law, in that
order.

4.2.5 Suppose A+B = A. Then B = B+0 = 0+B = (A+(−A))+B = ((−A)+A)+B = (−A)+(A+B) =

(−A) + A = A + (−A) = 0. Here, we have used the additive unit law, commutativity, the additive inverse
law, commutativity, associativity, the assumption A + B = A, commutativity, and the additive inverse law,
in that order.

4.3.2 By definition of scalar multiplication, this is equivalent to the matrix equation

   
x + 0y + 0z + w 2x + y + 2z + 0w 1 2
= .
3x + y − z + 0w 5x + y + 2z + w 6 5

By definition of equality of matrices, this is equivalent to the system of four scalar equations x + 0y + 0z +
w = 1, 2x + y + 2z + 0w = 2, 3x + y − z + 0w = 6, 5x + y + 2z + w = 5. Solving the system of equations,
we find x = 1, y = 2, z = −1, w = 0.

4.3.3 0A = (0 + 0)A = 0A + 0A. Now add −(0A) to both sides. Then 0 = 0A.

4.4.1          
1 2 4 0 1
(a) , (b) , (c) , (d) , (e) .
−1 0 −1 0 2

4.4.3 By columns:
       
a11 a12 a13 a11 x1 + a12 x2 + a13 x3
x1  a21  + x2  a22  + x3  a23  =  a21 x1 + a22 x2 + a23 x3  .
a31 a32 a33 a31 x1 + a32 x2 + a33 x3

This is exactly the same as the “by rows” formula.

 
1 −1 2 0
 1 0 2 0 
4.4.4 A = 
 0

0 3 0 
1 3 0 3
 507

 
1 3 2 0
 1 0 2 0 
4.4.5 A = 
 0

0 6 0 
1 3 0 1
 
−3 −6 −9
4.4.6 (a) .
−6 −3 −21
 
8 −5 3
(b) .
−11 5 −4
(c) Not possible.
 
−3 3 4
(d) .
6 −1 7
(e) Not possible.

(f) Not possible.

 
−3 −6
4.4.7 (a)  −9 −6 .
−3 3
(b) Not possible.
 
11 2
(c)  13 6 .
−4 2
(d) Not possible.
 
7
(e)  9 .
−2
(f) Not possible.

(g) Not possible.

 
2
(h) .
−5
      
−1 −1 x y −x − z −w − y 0 0
4.4.8 = = . Solution is: w = −y, x = −z. So the
3 3 z w 3x+ 3z 3w + 3y 0 0
x y
matrices are of the form .
−x −y
508 Answers to selected exercises

4.4.9
    
1 2 1 2 7 2k + 2
= ,
3 4 3 k 15 4k + 6
    
1 2 1 2 7 10
= .
3 k 3 4 3k + 3 4k + 6

Thus you must have 3k + 3 = 15 and 2k + 2 = 10. Solution is: k = 4.

4.4.10
    
1 2 1 2 3 2k + 2
= ,
3 4 1 k 7 4k + 6
    
1 2 1 2 7 10
= .
1 k 3 4 3k + 1 4k + 2

However, 7 6= 3 and so there is no possible choice of k which will make these matrices commute.
 
−71 168
4.4.12 .
−112 265
     
1 −1 1 1 2 2
4.4.13 A = ,B= ,C= .
−1 1 1 1 2 2
   
1 −1 1 1
4.4.14 A = ,B= .
−1 1 1 1
       
0 1 0 1 2 0 3 4 0 2 1 0
4.4.15 A =  1 0 0 , B =  3 4 0 . Then AB =  1 2 0  and BA =  4 3 0 .
0 0 0 0 0 0 0 0 0 0 0 0
 
1 0
4.4.16 A =
0 −1

4.4.18 (a) Not necessarily true.

(b) Not necessarily true.

(c) Not necessarily true.

(d) Necessarily true.

(e) Necessarily true.

(f) Not necessarily true.

(g) Not necessarily true.

 509

4.5.1 (a) Yes. (b) No. (c) Yes.

4.5.2 Yes B = C. Multiply AB = AC on the left by A−1 .

     
1 1 1 2 2 3
4.5.3 No. For example, let A = ,B= , and C = . Then AB = AC but B 6= C.
1 1 3 4 2 3
" # " # " #
3
− 17 − 53 15 0 13
4.5.4 A−1 = 7
1 2 , B−1 = , C−1 = . D−1 does not exist because the reduced
7 7" 1 0 1 − 32
#
1 12
echelon form of D is , which has a row of zeros. E −1 does not exist because E is not a square
0 0
matrix.
   
−2 4 −5 −2 0 3
   
4.5.5 A−1 =  0 1 −2 , B−1 =  0 13 − 23 . C−1 does not exist because the reduced eche-
1 −2 3 1 0 −1
 
  −1 1 1 1
5 2 2 2
1 0 3  3 1 1 5 
  
lon form of C is  0 1 23 , which has a row of zeros. D−1 =  2 − 2 − 2 .

 −1 0 0 1 
0 0 0
−2 − 34 1
4
9
4

" #−1 " #

d b
a b ad−bc − ad−bc
4.5.6 = c a
.
c d − ad−bc ad−bc

   
x 1
   
4.5.8 (a)  y  =  − 23 .
z 0
   
x −12
(b)  y = 1 .
z 5
   
x 3c − 2a
   
 y  =  13 b − 23 c .
z a−c

4.5.9 Multiply both sides of AX = B on the left by A−1 .

4.5.10 Multiply on both sides on the left by A−1 . Thus 0 = A−1 0 = A−1 (AX ) = (A−1 A)X = IX = X .

4.5.11 (AB)B−1A−1 = A(BB−1 )A−1 = AA−1 = I and B−1 A−1 (AB) = B−1 (A−1 A)B = B−1 IB = B−1 B = I.
510 Answers to selected exercises

4.5.12 It is not possible, because in that case, we would have A = IA = (AB)A = A(BA) = A0 = 0, and
therefore AB = 0, contradicting AB = I.

4.5.14 This follows from Problem 4.5.11 by taking A = B.

4.5.15 A−1 A = AA−1 = I and so A is an inverse of A−1 . Since (A−1 )−1 , by definition, is also an inverse of
A−1 , we have (A−1 )−1 = A by uniqueness.

4.5.16 (a) B is a right inverse of A, (b) B is a left inverse of A, (c) B is both a right inverse and left inverse
of A, (d) B is neither a right inverse nor a left inverse of A.
   
  1 0 1 0
1 0 0
4.5.17 A = , B =  0 1 , C =  0 1 .
0 1 0
0 0 1 1

4.5.18 The system can be written in matrix form as Av = b, where

   
8 2 3 −1
A =  0 1 −2  and b =  2  .
1 0 1 1
The inverse of A is  
1 −2 −7
A−1 =  −2 5 16  .
−1 2 8
Therefore,  
−12
v = A−1 b =  28  .
13

4.5.19 To solve for A, we invert both sides of the equation (A + B)−1 = CB−1 and use matrix algebra to
get A + B = (CB−1 )−1 = (B−1 )−1C−1 = BC−1 . Therefore, A = BC−1 − B.
To solve for B, we note that A = BC−1 − B = B(C−1 − I). Multiplying both sides of the equation on
the right by the inverse of C−1 − I, we get B = A(C−1 − I)−1 .
To solve for C, we take the original equation (A + B)−1 = CB−1 and right-multiply both sides of the
equation B. This yields C = (A + B)−1 B.

4.5.20 The matrix A is right invertible. Two possible right inverses are
   
1 −2 −2 1
 0 1  and  0 1 .
0 0 1 −1
The matrices B and C are not right invertible. The matrix D is right invertible with inverse
 
−3 2
.
2 −1
Since D is square, its right inverse is actually an inverse, and therefore unique.
 511

     
0 1 2 0 1 0
4.6.1 (a) E = , (b) E = , (c) E = .
1 0 0 1 2 1
 
0 −1 −2
4.7.1 X T Y =  0 −1 −2 , XY T = 1.
0 1 2
 
−3 −9 −3
4.7.2 (a) .
−6 −6 3
 
5 −18 5
(b) .
−11 4 4
 
(c) −7 1 5 .
 
1 3
(d) .
3 9
 
13 −16 1
(e)  −16 29 −8 .
1 −8 5
 
5 7 −1
(f) .
5 15 5
(g) Not possible because B is a 2 × 3-matrix and E is a 2 × 1-matrix, cannot multiply BE.

4.7.3 A is antisymmetric, B is symmetric, C is neither, and D is both.

4.7.4 We have A = AT = −A. Therefore, each entry ai j of A is equal to its own negation. This implies that
ai j = 0, and therefore A is the zero matrix.

4.7.5 A = 12 (A + AT ) + 12 (A − AT ).

4.7.6 If A is antisymmetric then A = −AT . It follows that aii = −aii and so each aii = 0.

4.7.7 This follows from properties 2 and 3 of Proposition 4.67.

· ·
4.7.8 (Au) v = (Au)T v = (uT AT )v = uT (AT v) = u (AT v).

4.7.9 We need to show that (A−1 )T is the inverse of AT . From properties of the transpose,

AT (A−1 )T = (A−1 A)T = I T = I,

(A−1 )T AT = (AA−1 )T = I T = I.

Hence AT is invertible and (AT )−1 = (A−1 )T .

512 Answers to selected exercises

4.7.10 We have (A−1 )T = (AT )−1 = A−1 , and therefore A−1 is equal to its own transpose, hence symmet-
ric.

4.9.1 Ciphertext: “ ZRUJPZVAEJTWOXGJZV ”.

4.9.2 The decryption matrix is  

16 22 5
A−1 =  22 26 17  .
5 17 2
Plaintext: “Spies are at the gate”.

4.9.3 The first two plaintext blocks are (8, 5), (12, 12) and the first two ciphertext blocks are (20, 7), (22, 24).
Eve solves the equation    
−1 20 22 8 12
A =
7 24 5 12
to find the secret decryption matrix  
−1 3 5
A = .
1 2
The plaintext is “Hello, password is kiwifruit”.

4.9.4 Ciphertext: “ VXUAJYY,AFJE ”.

4.9.5 The inverse of the diffusion matrix is

 
11 2 21
A−1 =  21 11 2  .
2 21 11

The plaintext is: “Eat more fruit”.

5.1.1 (a) x is not in the span. (b) y = 7u1 − 5u2 . (c) z = −3u1 + 4u2 .

5.1.2 It is the plane 2x + 3y − z = 0.

5.1.3 It is the hyperplane given by the equation x − 2y − 3z + 2w = 0.

5.1.5 0u1 + . . . + 0uk = 0.

5.1.6 (a) x = 3u2 . (b) y is not in the span. (c) z = 2u1 + 1u2 .

5.2.1 u2 and u4 are redundant. We have u2 = 2u1 and u4 = 3u3 − 2u1 .

5.2.3 The vectors are linearly dependent. We have u2 + 3u3 + 2u4 = 0.

5.2.5 The vectors are linearly dependent. We have w = 3u − 2v.

 513

   
2 1 3 3 1 4 5 −2
5.2.6  0 3 3 −3  ≃  0 1 1 −1 . Linearly independent subset: {u1 , u2 }.
3 5 8 1 0 0 0 0

5.2.10    
1 2 2 5 12 1 1 2 5 12
 1 2 7 7 17  ≃  0 0 5 2 5  .
−2 −4 −4 −10 −24 0 0 0 0 0
Linearly independent subset: {u1 , u3 }. Since the rank is 2, this is the smallest possible.

5.2.12 We write the vectors as the columns of a matrix and reduce to reduced echelon form over Z3 :
   
2 1 2 0 1 0 2 2
 1 1 0 1  ≃ ... ≃  0 1 1 2 .
0 2 2 1 0 0 0 0

u3 and u4 are redundant. We have u3 = 2u1 + 1u2 and u4 = 2u1 + 2u2 .

5.2.13 From a(u + v) + b(2u + w) + c(w − 2v) = 0 we get (a + 2b)u + (a − 2c)v + (b + c)w = 0. Since
u, v, w are linearly independent, this last system has only the trivial solution, so a + 2b = 0, a − 2c = 0, and
b + c = 0. However, these three equations have non-trivial solutions, for example (a, b, c) = (2, −1, 1). So
the vectors u + v, 2u + w, and w − 2v are linearly dependent.

5.2.14 From a(u + v) + b(u + w) + c(w + v) = 0 we get (a + b)u + (a + c)v + (b + c)w = 0. Since u, v, w
are linearly independent, this last system has only the trivial solution, so a +b = 0, a +c = 0, and b +c = 0.
Solving, we find the unique solution (a, b, c) = (0, 0, 0). So the vectors u + v, u + w, and w + v are linearly
independent.

5.2.15 Assume a1 z1 + . . . + ak zk = 0. Multiplying both sides of the equation by A, we get

a1 Az1 + . . . + ak Azk = a1 w1 + . . . + ak wk = 0.

Since the wi are linearly independent, it follows that each ai = 0. Therefore the zi are linearly independent
as well.
   
1 1
 0   0 
5.3.1 (a) No. We have     / V1 .
 0  ∈ V1 but 10  0  ∈
0 0
   
1 1
 1   1 
(b) This is not a subspace. The vector   is in V2 . However, (−1) 
 1  
 is not.
1 
1 1
(c) This is a subspace. It contains the zero vector and is closed with respect to vector addition and scalar
multiplication.
514 Answers to selected exercises

     
0 0 0
 0    
0   0 
(d) This is not a subspace. The vector   is in V4 . However (−1) 
 1   =  is not.
1   −1 
0 0 0
(e) This is a subspace. It contains the zero vector and is closed with respect to vector addition and scalar
multiplication.

5.3.2 Yes, this is a subspace because it contains the zero vector and is closed with respect to vector addition
· · ·
and scalar multiplication. For example, if u, v ∈ M, then w u = 0 and w v = 0, therefore w (u + v) = 0,
therefore u + v ∈ M.

5.3.3 Yes, this is a subspace.

5.3.4 (a) V1 is a subspace.

(b) V2 is not a subspace: not closed under addition. For example,

         
1 1 1 1 0
 1  ∈ V2 ,  1  ∈ V2 , but  1  +  1  =  0  ∈
6 V2 .
1 0 1 0 1

(c) V3 is a subspace.

(d) V4 is not a subspace. For example, it does not contain the zero vector.

5.3.5 Because 0 ∈ V and 0 ∈ W , we have 0 ∈ V ∩W . To show that V ∩W is closed under addition, assume
u, v ∈ V ∩W . Then u, v ∈ V , and since V is a subspace we have u + v ∈ V . Similarly u, v ∈ W , and since
W is a subspace we have u + v ∈ W . It follows that u + v is in both V and W , and therefore in V ∩ W .
To show that V ∩W is closed under scalar multiplication, assume u ∈ V ∩W and k ∈ R. Then u ∈ V , and
therefore ku ∈ V . Similarly ku ∈ W , and therefore also ku ∈ V ∩W .
   

 2 2 
    
3  ,  2  , dimension 2.
5.4.11 (a) Basis   1   1 

 
 
4 0
 
 1 
(b) Basis  3  , dimension 1.
 
1
   
 2 3 
(c) Basis  0 , 1  , dimension 2.
 
 
1 0

5.4.14 (a) No. For example, the vectors {0, 0, 0, 0, 0} are linearly dependent.

(b) No. For example, the vectors {0, 0, 0, 0} are linearly dependent.
 515

(c) Yes, by Proposition 5.48.

(d) Correct, by Proposition 5.48.

(e) Yes, by Proposition 5.49.

(f) No, in fact no such set is a basis of R5 , since every basis of R5 consists of 5 vectors by Theorem 5.38.

(g) No, as noted in the previous answer.

(h) No, no basis of R5 can have 6 elements.

(i) Yes, because a linearly independent set of vectors is a basis of the subspace it spans.

5.4.15 No. As a 5-dimensional space, R5 is spanned by 5 vectors. By the Exchange Lemma, any linearly
independent set can have size at most 5.
   
 2 1 
5.5.3 Rank 2, nullity 1, basis of column space  1 , 3  , basis of row space {[1, 2, 0] , [0, 0, 1]},
 
 
  3 2
 3 
basis of null space  1  .
 
0

5.5.4 Clearly 0 ∈ null(A), since A0 = 0. If x, y ∈ null(A), then A(x + y) = Ax + Ay = 0 + 0 = 0, and

therefore x + y ∈ null(A). Similarly, if x ∈ null(A) and k is a scalar, then A(kx) = k(Ax) = k0 = 0, so
kx ∈ null(A). So null(A) contains 0, and is closed under addition and scalar multiplication. It is therefore
a subspace of Rn .

5.5.5 Let  
a11 · · · a1n
 
A =  ... . . . ...  .
am1 · · · amn
Since col(A) is the span of the columns of A, we have v ∈ col(A) if and only if there exists scalars u1 , . . . , un
such that    
a11 a1n
   
v = u1  ...  + . . . + un  ...  .
am1 amn
But this equation is equivalent to
  
a11 · · · a1n u1
  
v =  ... . . . ...   ...  = Au.
am1 · · · amn un

Therefore, v ∈ col(A) if and only if v is of the form Au, for some u ∈ Rn . In other words, col(A) =
{Au | u ∈ Rn }.
516 Answers to selected exercises

5.5.6 The row space of A is the same as the column space of AT (except that it uses row vectors instead of
column vectors). Therefore, rank(A) = dim(row(A)) = dim(col(AT )) = rank(AT ).

5.5.7 From the theory of elementary matrices, we know that B can be written as a product of elementary
matrices B = E1 · · · Ek . It follows that BA = E1 · · · Ek A. Since each elementary matrix corresponds to an
elementary row operation, it follows that BA and A are row equivalent. Therefore, BA and A have the
same row space. It follows that rank(BA) = dim(row(BA)) = dim(row(A)) = rank(A). This proves the
first claim. To show the claim about AC, first note that by the above argument, rank(CT AT ) = rank(AT ),
because CT is invertible. Then rank(AC) = rank(A) follows by taking the transpose.

5.5.8 Let {w1 , . . . , wk } be a basis of col(B) ∩ null(A). Then for each i, we have wi ∈ col(B), and therefore
by Exercise 5.5.5, we can find zi such that wi = Bzi . Also let {u1 , . . . , ur } be a basis for null(B). Now
assume x ∈ null(AB). Then ABx = 0, and therefore Bx ∈ null(A) ∩ col(B). We therefore have

Bx = c1 w1 + . . . + ck wk = B(c1 z1 + . . . + ck zk ).

This implies
x − (c1 z1 + . . . + ck zk ) ∈ null(B)
and so it is of the form
x − (c1 z1 + . . . + ck zk ) = d1 u1 + . . . + dr ur .
It follows that
x ∈ span {z1 , . . . , zk , u1 , . . . , ur } .
Since we have shown that every element of null(AB) is in the span of these k + r vectors, it follows that

dim(null(AB)) ≤ k + r
= dim(col(B) ∩ null(A)) + dim(null(B))
≤ dim(null(A)) + dim(null(B)),

and therefore nullity(AB) ≤ nullity(A) + nullity(B).

6.1.1 T1 and T3 are linear, and T2 is not. The transformation T3 is called the zero transformation.

6.1.3 We have T (av + bw) = A(av + bw) = a(Av) + b(Aw) = aT (v) + bT (w) by properties of matrix
multiplication. Therefore, T is linear by Proposition 6.3.

6.1.4 We have

T (av + bw) = av + bw − ·
u (av + bw)
u
kuk2
= av − a ·
u v
2
u + bw − b ·
u w
u
kuk kuk2
= aT (v) + bT (w).

Therefore, T is linear.
 517

6.1.5 If T were a linear transformation, it should satisfy T (0) = 0, but it does not. Also T (v + w) 6=
T v + T w.

6.2.1 (a) The matrix of T is the elementary matrix that is like the identity matrix, except that the ( j, j)-
entry is b.
(b) The matrix of T is the elementary matrix that is like the identity matrix, except that the (i, j)-entry
is b.
(c) The matrix of T is the elementary matrix that switches the i th and the j th rows.

6.2.2 Since u1 , . . . , un is a basis of Rn , we know that A is invertible by Proposition 5.31. For each i =
1, . . . , n, since Aei = ui , we have A−1 ui = ei , and therefore

BA−1 ui = Bei = vi = T (ui ).

Now let w be some arbitrary vector in Rn . Since u1 , . . . , un is a basis, there exists a1 , . . . , an such that
w = a1 u1 + . . . + an un . Then

BA−1 w = BA−1 (a1 u1 + . . . + an un )

= a1 (BA−1u1 ) + . . . + an (BA−1un )
= a1 T (u1 ) + . . . + an T (un )
= T (a1 u1 + . . . + an un ) = T (w).

Therefore, BA−1 is the matrix of T .

6.2.3
  −1     
5 1 5 1 −1 0 5 1 5 3 2 1 37 17 11
 1 1 3   2 −1 −1  =  1 1 3   2 2 1  =  17 7 5  .
3 5 −2 −6 5 2 3 5 −2 4 1 1 11 14 6

6.2.4
  −1     
1 2 6 1 −1 0 1 2 6 6 3 1 52 21 9
 3 4 1  1 0 −1  =  3 4 1   5 3 1  =  44 23 8  .
1 1 −1 −8 6 3 1 1 −1 6 2 1 5 4 1

6.2.6 Recall that the desired matrix has i th column equal to proju (ei ) = ·
u ei
u. Therefore, the matrix is
kuk2
 
1 −2 3
1 
−2 4 −6  .
14
3 −6 9

6.2.7  
1 5 3
1 
5 25 15  .
35
3 15 9
518 Answers to selected exercises

 π


  1 √ 
cos 3
− sin π3
− 1
3
6.3.1 π π = 1√
2 2
1
sin 3 cos 3 2 3 2

6.3.8 First, we compute v′ , the projection of v onto u.

v
v′

T (v)

We have
v′ = proju v =
u v · u.
kuk2
·
But since u is a unit vector, this simplifies to v′ = (u v)u. From the above picture, we see that T (v) =
·
v + 2(v′ − v) = 2v′ − v = 2(u v)u − v. To get the matrix of this linear transformation, we must compute
the image of the standard basis vectors:
     2 
a 1 2a − 1
·
T (e1 ) = 2(u e1 )u − e1 = 2a
b
−
0
=
2ab
     
a 0 2ab
·
T (e2 ) = 2(u e2 )u − e2 = 2b
b
−
1
=
2b2 − 1
.

Therefore, the matrix is  

2a2 − 1 2ab
A= .
2ab 2b2 − 1

6.4.2 " #"

# " √ #

2π 2π
1 0 cos 3
− sin 3
− 12 − 21 3
2π 2π = √
0 −1 sin 3 cos 3 − 21 3 1
2

6.4.3 " #"

# " 1√ #

π
−1 0 cos 6
− sin π6 −2 3 1
2
π π

= 1 1
√
0 1 sin 6 cos 6 2 2 3

6.4.4 "

#" # " √ √ #

π
cos 4
− sin π4 1 0 1
2 1
2
π π

= 21 √ 2 √
1
sin 4 cos 4 0 −1 2 2 −2 2

6.4.5  

  √ 
1 0 0 cos π6 − sin π6 0 1
2 3 − 12 0
 

  √ 
 0 1 0   sin π6 cos π6 0  =  12 1
2 3 0 
0 0 −1 0 0 1 0 0 −1
 519

6.4.6 (a) T (0) = T (00) = 0T (0) = 0. (b) T (−v) = T ((−1)v) = (−1)T (v) = −T (v). (c) T (a1 v1 + . . . +
ak vk ) = T (a1 v1 ) + . . . + T (ak vk ) = a1 T (v1 ) + . . . + ak T (vk ).

6.4.7 The matrix of S ◦ T is given by

    
0 −2 3 1 2 −4
BA = = .
4 2 −1 2 10 8

Now,     
2 −4 2 8
(S ◦ T )(v) = BAv = = .
10 8 −1 12

6.4.8 We have
    
1 2 2 −4
(S ◦ T )(v) = S(T (v)) = B(T (v)) = = .
−1 3 −3 −11

6.4.9 The inverse of a reflection is a reflection, namely, itself. (For example, reflecting twice about the
x-axis returns each vector to its original position). The inverse of a rotation is a rotation by the same angle
in the opposite direction. The inverse of a shearing is a shearing in the opposite direction. The inverse of
a scaling by factor a is a scaling by factor 1/a.

6.4.10 The matrix of T −1 is A−1 .  −1  

2 1 −2 1
=
5 2 5 −2
   
4 −3 −1 −1 1 −3/2
6.4.11 The matrix of T is A = . The matrix of T is A = .
2 −2 1 −2

7.2.3 (a) −14. (b) 24. (c) 0.

7.2.4  
 1 2 1 
 
 2 1 3  = 6.
 
 2 1 1 

7.2.5  
 2 3 1 1 
 
 4 3 1 2 
  = 2.
 1 1 0 1 
 
 3 2 1 2 

7.2.6  
 1 2 −1 1 

 0 1 2 1 
 = 6.
 0 2 1 3 

 1 4 0 2 
520 Answers to selected exercises

7.2.7  
 1 0 0 1 
 
 2 1 1 0 
  = −4.
 0 0 0 2 
 
 2 1 3 1 

7.3.1 (a) det(A) = 2. (b) det(B) = −40. (c) det(C) = −24.

7.4.1 (a)  
 1 2 1 
 
 2 3 2  = −6.
 
 −4 1 2 

(b)  
 2 1 3 
 
 2 4 2  = −32.
 
 1 4 −5 

(c) One can row reduce this, using only row operations of the third kind, to
 
1 2 1 2
 0 −5 −5 −3 
 
 9 .
 0 0 2 5

0 0 0 − 63
10

Therefore, the determinant is −63.

(d) One can row reduce this, using only row operations of the third kind, to
 
1 4 1 2
 0 −10 −5 −3 
 
 19 .
 0 0 2 5

0 0 0 − 211
20

Thus the determinant is 211.

7.5.1 (a) The transpose was taken and det(B) = det(A).

(b) Two rows were switched and det(B) = − det(A).

(c) The first row was added to the second row and det(B) = det(A).

(d) The second row was multiplied by 2 and det(B) = 2 det(A).

(e) Two columns were switched and det(B) = − det(A).

7.5.2 By assumption, we can obtain a row of zeros by doing row operations. Row operations do not
change whether the determinant is zero, so the determinant must have been zero all along.
 521

7.5.3 If det(A) 6= 0, then A−1 exists. Therefore x = A−1 Ax = A−1 0 = 0.

7.5.4 The matrix kI has k down the main diagonal and has determinant equal to kn by Theorem 7.14.
Using Theorem 7.23, it follows that det(kA) = det(kIA) = det(kI) det(A) = kn det(A).

7.5.5   
1 2 −1 2
det = −8,
3 4 −5 6
   
1 2 −1 2
det det = −2 × 4 = −8.
3 4 −5 6
   
1 0 −1 0
7.5.6 This is not true at all. Consider A = and B = . Then det(A) = 1, det(B) = 1,
0 1 0 −1
and det(A + B) = 0.

7.5.7 Since Ak = 0, we have det(A)k = det(Ak ) = det(0) = 0. Therefore, it must be the case that det(A) = 0.

7.5.8 If A is orthogonal, we have det(A)2 = det(AT ) det(A) = det(AT A) = det(I) = 1. Therefore the only
possible values for det(A) are ±1.

1
7.5.9 det(A) = det(P−1 BP) = det(P−1 ) det(B) det(P) = det(P) det(B) det(P) = det(B).

7.5.10 The determinant is

 
 1 1 1       
   a a 2   1 1   1 1 
 1 a a2  =    +  = ab2 − ba2 − b2 + b + a2 − a = (a − 1)(b − 1)(b − a).
   b b2  −  b b2   a a2 
 1 b b 2

Therefore the determinant is 0 if a = 1, b = 1, or a = b. In all other cases, the determinant is non-zero and
the matrix is invertible.

7.5.11 This follows because det(ABC) = det(A) det(B) det(C) and if this product is non-zero, then each
determinant in the product is non-zero. Therefore, each of these matrices is invertible.

7.5.12 The given condition is what it takes for the determinant to be non-zero. Recall that the determinant
of an upper triangular matrix is just the product of the entries on the main diagonal. The inverse will
also be upper triangular; this can be seen by noting that every invertible upper triangular can be written
as a product of upper triangular elementary matrices; the inverse of each such elementary matrix is upper
triangular, and therefore so is their product. The analogous statement about lower triangular matrices is
also true.
 
1 1 0
7.5.13 (a) False. Consider  1 2 1 .
0 1 1
(b) True.
522 Answers to selected exercises

(c) False.

(d) False.

(e) True.

(f) True.

(g) True.

(h) True.

(i) True.

(j) True.

7.6.2 (a) det(A) = −13 and so A is invertible. This inverse is

  


 0 1 
  
 0 2  T
 2 1  −   
  1 0   3 0   3 1  
   T
        −1 3 −6
 
1   2 3        1 
−  1 3  − 1 2   = 3 −9 5 

−13  1 0   3 0   3 1    −13
  −4 −1 2
  2 3      
 1 3   1 2  
     
 2 1  − 0 1   0 2 
 
1 3 4
− 13
 13 13 
 3
=  − 13 9 1 .
13 13 
6 5 2
13 − 13 − 13

 T  
1
1 3 −6 7 − 72 2
7
1   
(b) det(B) = 7, so B is invertible. The inverse is   −2 1 5 
 =

3
7 7
1
− 17 .

7
6 5 2
2 −1 2 −7 7 7
 
1 0 −3
 
(c) det(C) = 3, so C is invertible. The inverse is 
 −3
2 1
3
5 .
3 
2 1 2
3 −3 −3

(d) det(D) = 0, so D is not invertible.

 
− 21 3
2 0
 
(e) det(E) = 2, and so E is invertible. The inverse is 

3
2 − 92 1 
.
1
2 − 12 0
 
2 −1
7.6.3 (a) det(A) = 1, so A is invertible. The inverse is A−1 = .
−1 1
 523



−1 −1 4
1 
(b) det(B) = −15, so B is invertible. The inverse is B−1 = −4 11 1 .
15
8 −7 −2

(c) det(C) = 0, so C is not invertible.

7.6.4 We have
   
6 12 6 6 12 −6
det(A) = 36, cof(A) =  12 6 −12  , adj(A) =  12 6 6 ,
−6 6 6 6 −12 6

and therefore    
1 1
6 12 −6 6 3 − 16
1  12
  
A−1 = 6 6  =  1 1 1 .
36    3 6 6 
1
6 −12 6 6 − 31 1
6

7.6.5 We have
   
−9 11 8 −9 −2 5
det(A) = 2, cof(A) =  −2 2 2 , adj(A) =  11 2 −5  ,
5 −5 −4 8 2 −4

and therefore    
−9 −2 5 − 92 −1 5
2
1
 11
  
A−1 = 2 −5  
 = 
11
1 5
−2 .
2 2 
8 2 −4 4 1 −2

7.6.6 No. It has non-zero determinant det(A) = cos2 t + sin2 t = 1 for all t, so it is invertible for all t.
 
 1 t t2 
  √
7.6.7 det(A) =  0 1 2t  = t 3 + 2, and so A has no inverse when t = − 3 2.

 t 0 2 

7.6.8 Since the matrix A has two identical rows, we have det(A) = 0 for all t. So this matrix is non-
invertible for all t.

7.6.9  
et e−t cost e−t sint
det  et −e−t cost − e−t sint −e−t sint + e−t cost  = 5e−t 6= 0
et 2e−t sint −2e−t cost
and so this matrix is always invertible.
524 Answers to selected exercises

7.6.10  
et 0 0
det  0 cost sint  = et .
0 cost − sint cost + sint
Hence the inverse is
 T  −t 
1 0 0 e 0 0
e−t  0 et cost + et sint −et (cost − sint)  =  0 cost + sint − sint  .
0 −et sint et cost 0 sint − cost cost

7.6.11  −1  
1 −t 1 −t
et cost sint 2e 0 2e
   
 et − sint cost  =  1 1 1 1 .
   2 cost + 2 sint − sint 2 sint − 2 cost 
1 1
t
e − cost − sint 2 sint − 2 cost cost − 12 cost − 12 sint

7.7.1 False. Cramer’s rule only works when the coefficient matrix is invertible. In these cases, the solution
is always unique.

7.7.2 Solution is: (x, y) = (1, 0).

7.7.3 Solution is: (x, y, z) = (1, 1, 0). For example,

 
 1 3 1 

 2 1 −1 
 
 1 1 1 
y=   =1
 1 2 1 

 2 −1 −1 
 
 1 0 1 

7.7.4 The solution is (x, y, z) = (1, −1, 2).

7.7.5 By Cramer’s rule, we have

 
 1 t 1 
 
 1 s t2 
 
 1 1 s2  s3 + t 3 + 1 − 2 − ts2 − t 2
y=  = 2 2 .
 1 1 1  ts + t + s − t − st 2 − s2
 
 1 t t2 
 
 1 s s2 

8.1.7 If v is an eigenvector corresponding to the eigenvalue λ , then A2 v = A(Av) = A(λ v) = λ (Av) = λ 2 v.

Therefore, v is an eigenvector of A2 with eigenvalue λ 2 .

8.1.8 We have λ A−1 v = A−1 λ v = A−1 Av = v. Since v 6= 0, this implies λ 6= 0. Moreover, it implies
A−1 v = λ −1 v. Thus, λ −1 is an eigenvalue of A−1 .
 525

8.1.9 Say Av = λ v. Then cAv = cλ v and so the eigenvalues of cA are just cλ where λ is an eigenvalue of
A.

8.1.10 Suppose v is an eigenvector of B, i.e., Bv = λ v. Then BAv = ABv = Aλ v = λ Av, and therefore Av
is an eigenvector of B.

8.1.11 Let v be the eigenvector. Then Am v = λ m v and Am v = Av = λ v. Therefore λ m = λ . Hence if

λ 6= 0, we must have λ m−1 = 1, which implies that λ = ±1.

8.1.12 The formula follows from properties of matrix multiplication. However, this vector might not be
an eigenvector because it might equal 0 and eigenvectors cannot equal 0.
 
0 1
8.2.8 Yes. works.
0 0

8.3.2 A rotation by 60◦ cannot map any non-zero vector to a scalar multiple of itself. The only rotations
of R2 that have real eigenvalues are rotations by 180◦ and 0◦ . The former has eigenvalue −1, and the latter
has eigenvalue 1 (with all vectors being eigenvectors).


1 0
8.3.3 The matrix of T is . The eigenvectors and eigenvalues are:
0 −1
   
0 1
for eigenvalue −1, for eigenvalue 1.
1 0


0 1
8.3.4 The matrix of T is . The eigenvectors and eigenvalues are:
1 0
   
1 −1
for eigenvalue −1, for eigenvalue 1.
1 1
 
1 0 0
8.3.5 The matrix of T is  0 1 0  The eigenvectors and eigenvalues are:
0 0 −1
     
 0   1 0 
 0  for eigenvalue −1,  0  ,  1  for eigenvalue 1.
   
1 0 0

8.4.2 The eigenvectors and eigenvalues are:

     
 2 −2   7 
 0  ,  1  for eigenvalue 1,  −2  for eigenvalue 3.
   
1 0 2
526 Answers to selected exercises

The matrix P needed to diagonalize the above matrix is

 
2 −2 7
 0 1 −2 
1 0 2

and the diagonal matrix D is  

1 0 0
 0 1 0 .
0 0 3

8.4.3 The eigenvalues are −1 and 1. The eigenvectors corresponding to the eigenvalues are:
   
 10   7 
 −2  for eigenvalue −1,  −2  for eigenvalue 1.
   
3 2

Since there are only 2 linearly independent eigenvectors, this matrix is not diagonalizable.

8.4.4 The eigenvectors and eigenvalues are:

     
 0   −2   −1 
 1  for eigenvalue −3,  1  for eigenvalue 3,  0  for eigenvalue −2.
     
0 1 1

The matrix P needed to diagonalize the above matrix is

 
0 −2 −1
 1 1 0 
0 1 1

and the diagonal matrix D is  

−3 0 0
 0 3 0 .
0 0 −2

8.5.1 First we write A = PDP−1 .

" # " #" #" #
1 2 −1 1 −1 0 − 12 1
2
= 1 1
2 1 1 1 0 3 2 2

Therefore A10 = PD10 P−1 .

" #10 " #" #" # " #
1 2 −1 1 (−1)10 0 − 12 1
2 1 310 + 1 310 − 1
= 1 1
= .
2 1 1 1 0 310 2 2
2 310 − 1 310 + 1
 
−1 −2
8.5.4 B = .
3 4
 527

 
0 0 2
8.5.5 A =  −1 1 2 .
−1 0 3

8.8.3 The solution is  

At 8e2t − 6e3t
e C= .
18e3t − 16e2t

8.9.1 (a) λ = 2 has algebraic multiplicity 2 and geometric multiplicity 1. Since the sum of the geometric
multiplicities of all eigenvalues is 1, the matrix is not diagonalizable.

(b) λ = 1 has algebraic and geometric multiplicity 1; λ = −3 has algebraic and geometric multiplicity
1. Since the sum of the geometric multiplicities is 2, the matrix is diagonalizable.

(c) λ = 2 has algebraic and geometric multiplicity 2, λ = 3 has algebraic and geometric multiplicity 1.
Since the sum of the geometric multiplicities is 3, the matrix is diagonalizable.

(d) λ = 3 has algebraic multiplicity 3 and geometric multiplicity 2. Since the sum of the geometric
multiplicities of all eigenvalues is 2, the matrix is not diagonalizable.

(e) λ = −1 has algebraic multiplicity 3 and geometric multiplicity 1. Since the sum of the geometric
multiplicities of all eigenvalues is 1, the matrix is not diagonalizable.

8.9.2 (a) Not diagonalizable.

(b) Diagonalizable.

(c) Not diagonalizable.

8.10.3 (a) The characteristic polynomial of A is a quadratic polynomial, and therefore it is of the form
p(λ ) = λ 2 + bλ + c, for some r, s ∈ R. By the Cayley-Hamilton theorem, p(A) = 0, therefore
A2 = −bA − cI. This proves that A2 is a linear combination of A and I.
 
0 1 0
(b) A =  0 0 1 .
0 0 0

8.11.1 (a) The characteristic polynomial is λ 2 − 4λ + 5, the eigenvalues are λ1 = 2 + i and λ2 = 2 − i,

and the corresponding basic eigenvectors are v1 = [1, i ]T and v2 = [1, −i ]T , respectively. The matrix
A is diagonalizable as A = PDP−1 , where
   
1 1 2+i 0
P= and D = .
i −i 0 2−i

(b) The characteristic polynomial is λ 2 + 1, the eigenvalues are λ1 = i and λ2 = −i, and the corre-
sponding basic eigenvectors are v1 = [1 − i, −1]T and v2 = [1 + i, −1]T , respectively. Therefore
B = PDP−1 , where    
1−i 1+i i 0
P= and D = .
−1 −1 0 −i
528 Answers to selected exercises

(c) The characteristic polynomial is x2 − 2x + 2 − (x3 − 2x2 + 2x) −λ 3 − λ 2 + 2 = (1 − λ )(λ 2 − 2λ + 2).

The eigenvalues are λ1 = 1, λ2 = 1 + i, and λ3 = 1 − i. The corresponding basic eigenvectors are
v1 = [1, 1, 0]T , v2 = [−1 − 2i, −i, 1]T , and v2 = [1 + 2i, i, 1]T , respectively. Therefore C = PDP−1 ,
where    
1 −1 − 2i 1 + 2i 1 0 0
P= 1 −i i  and D =  0 1 + i 0  .
0 1 1 0 0 1−i

(d) The characteristic polynomial is −λ 3 + 4λ 2 − 5λ + 2 = −(λ − 1)(λ − 1)(λ − 2). Therefore the
eigenvalues are λ = 1 with algebraic multiplicity 2 and λ = 2 with algebraic multiplicity 1. The
eigenspace for λ = 1 is 1-dimensional; it is spanned by [1, 1, −1]T . Since the geometric multiplicity
of the eigenvalue 1 is less than its algebraic multiplicity, the matrix is not diagonalizable, not even
over the complex numbers.
 
1
8.11.2 (a) By Proposition 8.51, the other eigenvalue is 1 − 2i with corresponding eigenvector .
−i

(b) From the eigenvalues and eigenvectors, we know that A = PDP−1 , where
   
1 1 1 + 2i 0
P= and D = .
i −i 0 1 − 2i
 
1 2
(c) A = PDP−1 = .
−2 1

9.1.13 Let f (i) be the i th component of a vector x ∈ Rn . Thus a typical element in Rn is ( f (1), . . ., f (n)).

9.2.1 0 = 0u1 + . . . + 0uk .


9.2.5 Let S = e0 , e1 , e2 , . . . . To prove the left-to-right implication, assume a ∈ span S. Then there exists

some finite subset ek1 , ek2 , . . . , ekn of S and scalars b1 , . . . , bn such that a = b1 ek1 + . . . + bn ekn . Let N ∈ N
be a number that is greater than k1 , . . . , kn . Since each of the sequences ek1 , . . . , ekn is zero after the first N
elements, the same is true for their linear combination a. Thus, a is finitely supported.
To prove the right-to-left direction, assume a is finitely supported. Let N ∈ N such that ak = 0 for
all k ≥ N. Then a = (a0 , a1 , . . . , aN−1 , 0, 0, 0, . . .), with infinitely many zeros following. Therefore, a =
a0 e0 + a1 e1 + . . . + aN−1 eN−1 , and it follows that a ∈ span S.

9.2.8 Let pi (x) denote the i th of these polynomials. Suppose C1 p1 (x) + . . . +C4 p4 (x) = 0. Then collecting
terms according to the exponent of x, we have

C1 a1 +C2 a2 +C3 a3 +C4 a4 = 0,

C1 b1 +C2 b2 +C3 b3 +C4 b4 = 0,
C1 c1 +C2 c2 +C3 c3 +C4 c4 = 0,
C1 d1 +C2 d2 +C3 d3 +C4 d4 = 0.

The matrix of coefficients is just the transpose of the above matrix. There exists a non-trivial solution if
and only if the determinant of this matrix equals 0.
 529

9.2.9 To determine whether R is linearly independent, we must solve the equation

a(2u − w) + b(w + v) + c(3v + 12 u) = 0.

If the only solution is the trivial solution, the set is linearly independent. We rewrite the equation as
follows.
(2a + 21 c)u + (b + 3c)v + (−a + b)w = 0.
Since u, v, w are linearly independent, the coefficients in the last equation must all equal 0. In other words:

2a + 12 c = 0,
b + 3c = 0,
−a + b = 0.

We solve and find that the unique solution is a = b = c = 0. Therefore, the set R is linearly independent.

9.3.2 No. It is not closed under scalar multiplication.

9.3.3 No. It is not closed under addition.

   
1 1
9.3.8 This is not a subspace. 1 is in it, but 5 1 is not.
1 1

9.4.1 (a) Yes. Suppose

c1 (x3 + 1) + c2 (x2 + x) + c3 (2x3 + x2 ) + c4 (2x3 − x2 − 3x + 1) = 0.

Then collect equal powers of x:

(c1 + 2c3 + 2c4 )x3 + (c2 + c3 − c4 )x2 + (c2 − 3c4 )x + (c1 + c4 ) = 0.

Does the system

c1 + 2c3 + 2c4 = 0
c2 + c3 − c4 = 0
c2 − 3c4 = 0
c1 + c4 = 0
have a non-trivial solution? The only solution is

c1 = 0, c2 = 0, c3 = 0, c4 = 0,

and therefore, the polynomials are linearly independent. Since there are 4 linearly independent
polynomials in a 4-dimensional space, they form a basis.

(b) Yes.

9.4.6 Yes, because the set of 5 linearly independent vectors can be extended to a basis B of V . But since
V is 5-dimensional, B has only 5 elements, which must be the original 5 vectors.
530 Answers to selected exercises

9.4.7 No. Since V has a spanning set of size 5, the 6 vectors cannot be linearly independent by the
Exchange Lemma.

9.4.11 (a) In case k = 3, the following sequences form a basis for W3 :

(1, 0, 0, 1, 0, 0, 1, 0, 0, 1, . . .),
(0, 1, 0, 0, 1, 0, 0, 1, 0, 0, . . .),
(0, 0, 1, 0, 0, 1, 0, 0, 1, 0, . . .).

Therefore, W3 is a 3-dimensional space. For general k, the situation is analogous and the dimension
of Wk is k.

(b) Let U be the set of all periodic sequences of all periods. We can find an (infinite) spanning set for U
by taking all the basis vectors for all of the spaces Wk :

(1, 1, 1, 1, 1, 1, 1, 1, 1, 1, . . .) (period 1),

(1, 0, 1, 0, 1, 0, 1, 0, 1, 0, . . .) (period 2),
(0, 1, 0, 1, 0, 1, 0, 1, 0, 1, . . .) (period 2),
(1, 0, 0, 1, 0, 0, 1, 0, 0, 1, . . .) (period 3),
(0, 1, 0, 0, 1, 0, 0, 1, 0, 0, . . .) (period 3),
(0, 0, 1, 0, 0, 1, 0, 0, 1, 0, . . .) (period 3),
(1, 0, 0, 0, 1, 0, 0, 0, 1, 0, . . .) (period 4),
(0, 1, 0, 0, 0, 1, 0, 0, 0, 1, . . .) (period 4),
(0, 0, 1, 0, 0, 0, 1, 0, 0, 0, . . .) (period 4),
(0, 0, 0, 1, 0, 0, 0, 1, 0, 0, . . .) (period 4),

and so on. However, these sequences are not linearly independent. For example, we can obtain the
sequence (1, 0, 1, 0, 1, 0, 1, 0, . . .) of period 2 as a linear combination of two sequences of period 4,
namely (1, 0, 0, 0, 1, 0, 0, 0, . . .) and (0, 0, 1, 0, 0, 0, 1, 0, . . .). By Proposition 9.47, we know that it is
possible to shrink the above spanning set to a basis by removing certain sequences. How exactly to
do this is an interesting question. One way to construct a basis is to keep exactly those sequences of
period k that start with ℓ zeros, where gcd(ℓ, k) = 1. Proving that this really works is an interesting
project.
√
9.4.12 When we add two number of the √form a + b 2, we get another number of the same form. When
we multiply a number
√ of the form a + b 2 by a (rational) scalar, we get another number of the same form.
Also, 0 = 0 + 0 2 is of the required form. The 8 axiomsnof a vector space are satisfied because all of
√ o
them are laws of the arithmetic of real numbers. A basis is 1, 2 . By definition, the span of these gives
√
the collection of vectors. To prove
√that they are linearly independent, assume √ a + b 2 = 0, where a, b are
rational numbers. If b 6= 0, then 2 = − ba , which cannot happen because 2 is irrational. If a 6= 0, then
√
√1 = − b , which again cannot happen since √1 is irrational. Hence both a, b = 0. Therefore, 1 and 2 are
2 a 2
linearly independent over the rational numbers, and form a basis. The dimension of the space is 2.

9.5.6 The block length is n = r − 1 = 15 and the message length is k = n − r = 11. The check matrix and
generator matrix are not unique, because there are different ways of ordering the columns of the check
 531

matrix. Here is one possible answer:

 
1 0 0 0 0 0 0 0 0 0 0
 00 1
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0

0 0 0 1 0 0 0 0 0 0 0
0 0 0 0 1 0 0 0 0 0 0
   
1 1 0 1 1 0 1 0 1 0 1 1 0 0 0 0 0 0 0 0 1 0 0 0 0 0
0 0 0 0 0 0 1 0 0 0 0
1 0 1 1 0 1 1 0 0 1 1 0 1 0 0 ,  
H =0 1 1 1 0 0 0 1 1 1 1 0 0 1 0 G=0 0 0 0 0 0 0 1 0 0 0 .
0 0 0 0 0 0 0 0 1 0 0
0 0 0 0 1 1 1 1 1 1 1 0 0 0 1 0 0 0 0 0 0 0 0 0 1 0
 
0 0 0 0 0 0 0 0 0 0 1
1 1 0 1 1 0 1 0 1 0 1
 
1 0 1 1 0 1 1 0 0 1 1
0 1 1 1 0 0 0 1 1 1 1
0 0 0 0 1 1 1 1 1 1 1
With these matrices, the encoding of 00110011000 10100000001 is 001100110000110 101101110110101.
The error syndrome for 001000010010000 is 1010, which corresponds to an error in bit 2. We find
that the corrected code block is 011000010010000, and the decoding is 01100001001. The error syn-
drome for 010000100000001 is 0110, which corresponds to an error in bit 3. The corrected code block is
011000100000001 and the decoding is 01100010000. However, if you have used a different check matrix
and/or generator matrix, your answers might differ.
 
0  
  1
10.1.1 (a) We have T 0 = , so T does not preserve the zero vector.
0
0
   
0   0  
2 8
(b) We have T  1  = and T  2  = , so T does not preserve scalar multiplication.
2 4
0 0
   
π /2   π  
1 0
(c) We have T  0  = and T  0  = , so T does not preserve scalar multiplication.
π /2 π
0 0

10.1.3 Yes, yes, no, yes.

10.1.6 (a) Let a = (a0 , a1 , a2 , . . .). Then the equation shift(a) = a is equivalent to

(a1, a2 , a3 , . . .) = (a0 , a1 , a2 , . . .),

which translates to the recurrence an+1 = an . The only free variable is a0 , and the only solutions are
the constant sequences. They form a 1-dimensional space with basis {(1, 1, 1, . . .)}.

(b) The equation shift(a) = −a is equivalent to

(a1 , a2 , a3 , . . .) = (−a0 , −a1 , −a2 , . . .),

which translates to the recurrence an+1 = −an . The only free variable is a0 , and the solutions form
a 1-dimensional space with basis {(1, −1, 1, −1, . . .)}.
532 Answers to selected exercises

(c) The equation shift(shift(a)) = shift(a) + 2a is equivalent to

(a2 , a3 , a4 , . . .) = (a1 + 2a0 , a2 + 2a1 , a3 + 2a2 , . . .),
which translates to the recurrence an+2 = an+1 + 2an . The only free variables are a0 and a1 . The
most obvious basis vectors are obtained by letting (a0 , a1 ) = (1, 0) and (a0 , a1 ) = (0, 1), giving the
sequences (1, 0, 2, 2, 6, 10, 22, . . .) and (0, 1, 1, 3, 5, 11, 21, . . .). Thus,
{(1, 0, 2, 2, 6, 10, 22, . . .), (0, 1, 1, 3, 5, 11, 21, . . .)}
is a basis for the solutions. Another basis, which is slightly more convenient, is
{(1, −1, 1, −1, 1, −1, . . .), (1, 2, 4, 8, 16, 32, . . .)} .

10.2.1 Yes, no, yes, no.

10.3.1 By linearity we have T (x2 ) = 1, T (x) = T (x2 + x − x2 ) = T (x2 + x) − T (x2 ) = 5 − 1 = 4, and

T (1) = T ((x2 + x + 1) − (x2 + x)) = T (x2 + x + 1) − T (x2 + x) = −1 − 5 = −6. Thus T (ax2 + bx + c) =
aT (x2 ) + bT (x) + cT (1) = a + 4b − 6c.

10.3.2 The matrix A is invertible if and only if its rank is n, which means that v1 , . . . , vn are linearly
independent and therefore a basis of Rn . The existence of T then follows from Theorem 10.20.
 
2
10.4.2 [x]B =  1 .
−1

10.4.3 (a) The i th column of [T ]E,E is [T (ei )]E = T (ei ), so

 
3 1 1
[T ]E,E =  3 2 3  .
3 3 −1

(b) The i th column of [T ]B,B is [T (vi )]B . This requires a calculation:

           
1 3 1 1 2 −3
T (v1 ) = T  0  =  3  = −3  0  + 0  1  + 3  1  , therefore [T (v1 )]B =  0 .
0 3 0 0 1 3
           
1 4 1 1 2 −7
T (v2 ) = T  1  =  5  = −7  0  − 1  1  + 6  1  , therefore [T (v2 )]B =  −1  .
0 6 0 0 1 6
           
2 8 1 1 2 −11
T (v3 ) = T  1  =  11  = −11  0  + 3  1  + 8  1  , therefore [T (v3 )]B =  3 .
1 8 0 0 1 8
We therefore have  
−3 −7 −11
[T ]B,B =  0 −1 3  .
3 6 8
 533

 
1 −1
10.4.4 MB2B1 = .
0 1

10.4.7 Recall that projv (w) =

v w
kvk2
· v. The desired matrix has i th column equal to projv (ei ). Therefore, the
desired matrix is  
1 −2 3
1 
−2 4 −6  .
14
3 −6 9

10.4.8 We have T (v1 ) = v1 , T (v2 ) = 0, and T (v3 ) = 0. Therefore

 
1 0 0
[T ]B,B =  0 0 0  .
0 0 0

10.4.9 Since M is invertible, its columns [v1 ]B , . . . , [vn ]B are linearly independent and span Rn ; it follows
that v1 , . . . , vn is a basis of V . To show that [T ]C,B = NM −1 , it is sufficient to check that NM −1 [vi ]B = [wi ]C ,
for all i = 1, . . . , n. But by assumption, [vi ]B is the i th column of M, so that [vi ]B = Mei , where ei is the
i th basis vector. Therefore M −1 [vi ]B = ei . On the other hand, Nei is the i th column of N, i.e., [wi ]C . We
therefore have
NM −1 [vi ]B = Nei = [wi ]C ,
as desired.

11.1.1 (a) Yes, it is an inner product. Since A = AT , symmetry and linearity follow as in Example 11.3.
For the positive definite property, note that hu, ui = u21 + 2u22 ≥ 0, and equality holds if and only if
u1 , u2 = 0.

(b) Yes, it is an inner product. Since A = AT , symmetry and linearity follow as in Example 11.3. For
the positive definite property, we have hu, ui = 3u21 + 2u1 u2 + 3u22 = (u1 + u2 )2 + 2u21 + 2u22 ≥ 0.
Equality holds if and only if u1 + u2 , u1 , and u2 are all zero, which is the case if and only if u = 0.
       
1 0 0 1
(c) No, it is not symmetric. For example, h , i = 1 but h , i = 0.
0 1 1 0
   
0 0
(d) No, it is not positive definite. For example, h , i = −1.
1 1
     
1 1 1
(e) No, it is not positive definite. For example, h , i = 0 although 6= 0.
−1 −1 −1
R1 R1
11.1.2 (a) h1, xi = 0 1 · x dx = 0 x dx = 12 .
R1 R1 3 1
(b) hx, x2 i = 0 x · x2 dx = 0 x dx = 4 .
R1 R1 43
(c) h1 + x, 2 + x2 i = 2 2 3
0 (1 + x)(2 + x ) dx = 0 2 + x + 2x + x dx = 12 .

R1
11.1.3 (a) k1k2 = h1, 1i = 0 1 dx = 1, therefore k1k = 1.
534 Answers to selected exercises

R1 2
(b) kxk2 = hx, xi = 1
0 x dx = 3 , therefore kxk =
√1 .
3
2 R q
(c) x2 + 1 = hx2 + 1, x2 + 1i = 01 x4 + 2x2 + 1 dx = 15 + 23 + 1 = 28
15 , therefore x2 + 1 = 28
15 .

11.1.4 The operation hu, vi = u ∗ v is an inner product on R3 . Symmetry and linearity are straightforward
to check, as is the positive definite property. Therefore, the claimed inequality holds by Proposition 2.27.
R1 2 2 2 2
R1 4 2 2
R1 3
11.1.5 (a) We have hx, xi = −1 x dx = 3 , hx , x i = −1 x dx = 5 , and hx, x i = −1 x dx = 0. Therefore
hx, x2 i
cos θ = = 0.
kxk kx2 k
Therefore, the angle θ is π /2 radians, or 90 degrees. In other words, x and x2 are orthogonal in
C[−1, 1].
R1 2 R1 1 4 R
(b) We have hx, xi = 2
−1 x dx = 3 , hx3 , x3 i = = 27 , and hx, x3 i = −1
6
−1 x dx x dx = 25 . Therefore
2 √
hx, x3 i 5 21
cos θ = 3
= q q = .
kxk kx k 2 2 5
3 7
√
The angle θ is cos−1 ( 521 ), which is approximately 0.4115 radians or 23.58 degrees.

11.1.6 (a) By assumption, a and b are square summable. Let N = a20 + a21 + . . . and M = b20 + b21 + . . .. By
the Cauchy-Schwarz inequality, for all n, we have
q q √ √
|a0 | |b0 | + . . . + |an | |bn | ≤ |a0 |2 + . . . + |an |2 |b0 |2 + . . . + |bn |2 ≤ N M.
Therefore the series |a0 b0 | + |a1 b1 | + . . . is bounded. By the absolute convergence test from calculus, it
follows that the series a0 b0 + a1 b1 + . . . converges.
(b) It is clear that the zero sequence is square summable, and also that a scalar multiple of a square
summable sequence is square summable. Hence HilbR contains the zero vector and is closed under scalar
multiplication. To show that it is closed under addition, assume a, b ∈ HilbR , and let c = a + b. We must
show that c is square summable. But
c20 + c21 + . . . = (a0 + b0 )2 + (a1 + b1 )2 + . . . = (a20 + a21 + . . .) + (b20 + b21 + . . .) + (2a0 b0 + 2a1 b1 + . . .).
The series a20 + a21 + . . . and b20 + b21 + . . . converge by assumption, and the series 2a0 b0 + 2a1 b1 + . . .
converges by part (a). It follows that HilbR is closed under addition.
(c) Symmetry and linearity follow straightforwardly from properties of convergent series. For example,
ha, kb + ℓci = a0 (kb0 + ℓc0 ) + a1 (kb1 + ℓc1 ) + . . .
= k(a0 b0 + a1 b1 + . . .) + ℓ(a0c0 + a1 c1 + . . .)
= kha, bi + ℓha, ci.
As for the positive definite property, note that
ha, ai = a20 + a21 + . . . ≥ 0.
Moreover, since all terms in the series are ≥ 0, it follows that ha, ai = 0 if and only if ai = 0 for all i.
 535

11.2.1 We have hu1 , u2 i = uT1 Au2 = 0, hu1 , u3 i = uT1 Au3 = −5, hu1 , u4 i = uT1 Au4 = 0, hu2 , u3 i = uT2 Au3 =
0, hu2 , u4 i = uT2 Au4 = 32, and hu3 , u4 i = uT3 Au4 = 54. Therefore, u1 ⊥ u2 , u1 ⊥ u4 , and u2 ⊥ u3 . None
of the other pairs of vectors are orthogonal.

11.2.2 f1 ⊥ f2 , f1 ⊥ f4 , f2 ⊥ f3 , and f3 ⊥ f4 .

11.2.3 (a) Let p(x) = ax3 +bx2 +cx +d. Then hp(x), x2 i = 52 b + 32 d and hp(x), xi = 25 a + 23 c. Therefore,

p(x) is in the orthogonal complement of x2 , x if and only if 25 b+ 23 d = 0 and 25 a+ 23 c = 0. It follows
that b = − 53 d and a = − 53 c. The general solution is p(x) = − 53 cx3 − 53 dx2 + cx + d. A basis for the

orthogonal complement is − 35 x3 + x, − 35 x2 + 1 .

(b) A basis for the orthogonal complement of {x + 1} is 5x3 − 1, 3x2 − 1, 3x − 1 .

11.2.4 (a) Orthonormal (therefore also orthogonal). (b) Neither orthogonal nor orthonormal. (c) Orthog-
onal (not orthonormal). (d) Orthonormal (therefore also orthogonal).

11.2.5 v = 14 u1 + 23 u2 − 45 u3 .

hu1 ,vi
11.2.6 We have v = a1 u1 + a2 u2 + a3 u3 where a1 = hu 1 , u 1 i = 12 . So the first coordinate is 21 .

  
1 1
11.3.1 u1 =  2 , u2 =  4 .
3 −3
  
0 3
 1   −1 
11.3.2 u1 =  , u2 = 
 1  
.
1 
0 −1
    
1 0 0
 0   3   1 
11.3.3 u1 =  , u2 = 
 1  
, u3 =  .
0   0 
0 −1 3
    
1 1 −1
11.3.4 u1 =  0 , u2 =  1 , u3 =  1 .
2 −1 0
     
1 1 −1
 −1   1   1 
11.3.5 u1 = 
 1 
, u2 = 

, u3 =  .
−1   1 
1 1 0

11.3.6 u1 = 1, u2 = x − 1, x2 − 2x + 32 .
536 Answers to selected exercises

    
   1 1 1 
u1 u2 1
11.3.7 The orthonormal basis is , = √  2 , √  4  .
ku1 k ku2 k  14 26 −3 
3
 
−2
11.4.1 The best approximation is v′ = 3u1 − u2 =  5 .
2
 
1
 3 
11.4.2 The best approximation is v′ = u1 + 2u2 − u3 =  
 −2 .
4

11.4.3 Let p0 , p1 , . . . be the Legendre polynomials from Section 11.3. The approximating polynomials
are:
3
f0 (x) = p0 ,
4
3 1
f1 (x) = p0 − p1 ,
4 2
3 1 15
f2 (x) = p0 − p1 − p2 ,
4 2 32
3 1 15
f3 (x) = p0 − p1 − p2 + 0p3 ,
4 2 32
3 1 15 105
f4 (x) = p0 − p1 − p2 + 0p3 + p4 .
4 2 32 256

f0

f1
f4
f x
−1 f
1 2 = f 3

π2
11.4.4 3 − 41 cos x + 44 cos 2x − 49 cos 3x + 16
4 4
cos 4x − 25 cos 5x ± . . ..

11.5.1 (x, y, z) = (−1, −1, 2).

11.5.2 (x, y, z) = (2, 2, −1).

11.5.3 y = 2 + 2x.

11.5.4 y = 1 − x + x2 .
 537

11.6.1 (a) Yes. (b) Yes. (c) No. (d) Yes.

11.6.2 A and C are orthogonal (and therefore isometries). D is an isometry but not orthogonal. B is neither
orthogonal nor an isometry.

11.6.3 A and D are orthogonal, but B and C are not.

   
−2 0 1 2 1
11.7.1 (a) A = PDP−1
where D = ,P= √ .
0 3 5 −1 2
   
2 0 1 1 1
(b) A = PDP−1 where D = ,P= √ .
0 4 2 −1 1
 
  √1 √1 √1
0 0 0  3 2 6 
(c) A = PDP−1 where D =  0 1 0 , P =   − √1 0 √2 .
3 6 
0 0 3 √1 − √1 √1
3 2 6
 
  √1 √1 √1
1 0 0  2 3 6 
(d) A = PDP−1 where D =  0 1 0 , P = 
 − √12 √1 √1 .

3 6
0 0 7 0 − √13 √2
6

11.7.2 Suppose A is orthogonally diagonalizable. Then A = PDP−1 , where D is diagonal and P is orthog-
onal. Since P is orthogonal, we have P−1 = PT , and therefore A = PDPT . Moreover, since D is diagonal,
we have D = DT . It follows that AT = (PDPT )T = (PT )T DT PT = PDPT = A, so A is symmetric.

11.8.1 (a) Positive semidefinite. (b) Positive definite. (c) Not symmetric (therefore neither). (d) Positive
semidefinite. (e) Neither.

11.8.2 (a) Eigenvalues: {1, 6}. Positive definite. (b) Eigenvalues: {0, 13}. Positive semidefinite. (c)
Eigenvalues: {0, 2, 3}. Positive semidefinite. (d) Eigenvalues: {−1, 4, 8}. Neither.

11.8.3 (a) No (neither positive definite nor semidefinite). (b) Yes. (c) No (positive semidefinite but not
definite). (d) Yes.

11.8.4 The characteristic polynomials are:

det(A − λ I) = λ 2 − 3λ − 2,
det(B − λ I) = −λ 3 + 4λ 2 − 4λ + 1,
det(C − λ I) = λ 4 − 6λ 3 + 9λ 2 − 3λ + 0.

For A, the coefficients are not weakly alternating, so A is not positive semidefinite. For B, the coefficients
are strongly alternating, so B is positive definite. For C, the coefficients are weakly, but not strongly
alternating, so C is positive semidefinite, but not positive definite.

11.9.1 All of them except (e) and (g).

538 Answers to selected exercises

11.9.3 f (x, y, z) = vT Av, where  

3 1 −2
A= 1 1 1.5  .
−2 1.5 2

11.9.4 3u2 + v2 + vw + w2 .

11.9.5 x = u + v and y = u − v gives f = 4u2 + 8v2 .

   
1 1 0
11.9.7 (a) Eigenvalues: λ1 = 1, λ2 = 2. Principal axes: u1 = and u2 = . u1 -intercept: 1,
0 1
√
u2 -intercept: 2.
   
√1
2 √1
1
(b) Eigenvalues: λ1 = 1, λ2 = 6. Principal axes: u1 = and u2 = . u1 -intercept: 1,
5 −1 5 2
u2 -intercept: √1 .
6
   
√1
1 √1
1 √1 ,
(c) Eigenvalues: λ1 = 2, λ2 = 4. Principal axes: u1 = and u2 = . u1 -intercept:
2 −1 2 1 2
u2 -intercept: 12 .

√
2
y y y
1 1 u2 1 u2
1
u2 √1 2
6
1
x x x
u1
−1 1 −1 1u −1 1
1
1 √1 u1
2
−1 −1 −1

    
1 1 1
11.9.8 The principal axes are √13  −1 , √1  1 ,
2
√1  −1 .
6
−1 0 2

11.10.1 (a) 1 − i, (b) 1 + 5i, (c) 1 − 5i, (d) 0.

√ √ √
11.10.2 (a) 2, (b) 6, (c) 13.

11.10.3 (a) −6i, (b) 5i.

11.10.7 u1 ⊥ u2 , u1 ⊥ u4 , and u2 ⊥ u3 .

11.10.9 Orthogonal: [0, i, 2]T , [1, 2, i]T . Orthonormal: √1 [0, i, 2]T , √1 [1, 2, i]T .
5 6
 539

11.11.1 (a) Yes. (b) Yes. (c) No. (d) Yes.

11.11.2 (a) Yes. (b) No. (c) Yes.

11.11.3 (a) No. (b) Yes. (c) Yes.

11.11.4 All except (b) and (e) are unitary.

11.11.5 Only (f) is unitary.

11.11.8 All except (a), (c), and (d) are hermitian.

   
√1
1+i √1
1
11.11.9 Eigenvalues: λ1 = 1 and λ2 = 4. Normalized eigenvectors: v1 = , v2 = .
3 −1 3 1−i
   
1 0 1 1+i 1
D= , P= √ .
0 4 3 −1 1 − i

w 50 37
A.1.1 z + w = 5 − i, z − 2w = −4 + 23i, zw = 62 + 5i, and = − − i.
z 53 53

|z| |z| |z|

A.1.4 If z = 0, let w = 1. If z 6= 0, let w = . Note that wz = z = |z| and |w| = = 1.
z z |z|
√
A.1.5 The problem is that there is no single −1. In the complex numbers, −1 has two square roots,
namely i and −i. Since each complex number has two square roots, and generally neither of them is
√
real, the notation z does not have a fixed meaning in the complex
positive or even√ numbers. Therefore,
√ √ √√ √
the equation z w = zw cannot be used. At best, we could maybe say z w = ± zw.

A.2.2 Since i has magnitude 1 and argument π /2 (or 90◦ ), the number z must have magnitude 1 and
argument π /4. It therefore lies at 45◦ on the unit circle. The solution is z = 1+i
√ . A second solution is
2
−z = − 1+i
√ , whose argument is −3π /4 or −135◦ . Note that if we double this angle, we get −270◦ , which
2
is the same as +90◦ .
i
z = 1+i
√
2

−1 1
−z = − 1+i
√
2
−i
540 Answers to selected exercises

A.2.3 The three solutions can be found on the unit circle at 60◦ , 180◦ , and 300◦ . If we triple any of
◦ ◦
√ we get 180 (up to multiples of 360 ). Thus, the three cube roots of −1 are z = −1 and
these angles,
z = 0.5 ± 0.75 i.
√
0.5 + 0.75 i

−1

√
0.5 − 0.75 i

√
6 ± 16
A.3.1 By the quadratic formula, z = = 3 ± 2i.
2

A.3.2 p(z) factors as (z − 1)(z2 + 6z + 10). The roots are z = 1, z = −3 + i, and z = −3 − i.

A.3.3 By trial and error, we find the roots z = 1 and z = −2. Moreover, z = 1 is a double root. We have
z3 − 3z + 2 = (z − 1)(z − 1)(z + 2).

A.3.4 If p(z) = 0 and a0 , . . . , ak are real, we have

p(z) = an zn + an−1 zn−1 + . . . + a1 z + a0

= an zn + an−1 zn−1 + . . . + a1 z + a0
= an zn + an−1 zn−1 + . . . + a1 z + a0
= an zn + an−1 zn−1 + . . . + a1 z + a0
= p(z)
= 0.
Index
absolute value of triangle, 94
of a complex number, see magnitude associated homogeneous system, 37
addition associative law
closed under, 342 of addition, 43, 68, 325
in a field, 43 of multiplication, 43
in a vector space, 325 of scalar multiplication, 70, 325
of complex numbers, 486 augmented matrix, see also matrix, 15
of linear transformations, 375 average
of matrices, 123 of vectors, 477
of polynomials, 326, 330 axiom of choice, 350
of two angles, 232
of vectors, 65, 325 back substitution, 13
additive inverse, 43, 68, 325 base case
additive unit, 43, 68, 325 of a recurrence, 293
adjugate formula, 260 basic eigenvector, 271
adjugate of a matrix, 259 basic solution, 36, 198
affine function, 237 of differential equation, 299
affine subspace, 477 basis, 195, 349
affine subspace fitting, 477 and invertible matrix, 197
algebra basis test, 205
fundamental theorem of, 485 by extending linearly independent set, 203, 352
algebraic multiplicity by shrinking spanning set, 205, 352
of an eigenvalue, 312, 321 existence, 202
alternating signs of a solution space, 197
strongly, 445 of a span, 197
weakly, 445 of a vector space, 349
analytic function, 308 of Rn , 196, 197
angle orthogonal, 397
between line and plane, 117 orthonormal, 397
between two lines, 107 size of, 201, 351
between two planes, 117 standard, 196
between vectors, 83, 392 basis test, 205
in inner product space, 392 bit, 44
antilinearity block
of inner product, 458 block length, 356
antisymmetric matrix, 163 code block, 356
approximation, 414, 463 message block, 356
Fourier, 417 block cipher, 168
least squares, 425 block size, 168
of functions, 415, 417 box product, 95
area volume of parallelepiped, 96
of parallelogram, 91 byte, 174

541
542 INDEX

C[a, b], 388 substitution cipher, 167

calculus ciphertext, 167
multivariable, 217 classical adjoint, 259
camera, 235 closed formula, 294
camera coordinates, 235 closed under
cancellation law addition, 342
of addition, 331 scalar multiplication, 342
Cartesian coordinates, 61 code, 356
Cartesian form error correcting, see error correcting code
of a complex number, 486 code block, 356
casting-out algorithm, 181 coefficient, 7, 9
extended, 182 of a linear combination, 71
Cauchy-Schwarz inequality of a polynomial, 275, 491
complex, 459 of a quadratic form, 448
for dot product, 81 coefficient matrix, 14, 131
in inner product space, 390 cofactor, 244
Cayley-Hamilton theorem, 315 cofactor expansion, 245
center of mass, 477 cofactor matrix, 259
centroid, 477 column method
change of parameters matrix multiplication, 133
line, 103 column space, 211, 426
plane, 119 column vector, 62, 122
change of variables commutative law
for a differential equation, 301 of addition, 43, 68, 325
for a quadratic form, 450 of multiplication, 43
characteristic polynomial, 275 commuting matrices, 137
Cayley-Hamilton theorem, 315 complex conjugate, 487
root, 276 properties, 488
check matrix, 360 complex dot product, 457
chemical reactions properties, 458
balancing, 51 complex inner product space, see inner product space
choice complex number, 49, 485, 486
axiom of, 350 absolute value, see magnitude
cipher, 167 addition, 486
block cipher, 168 Cartesian form, 486
ciphertext, 167 conjugate, 487
decryption, 167 properties, 488
diffusion, 168 division, 487
diffusion matrix, 172 field axioms, 487
encryption, 167 geometric interpretation, 489
Hill cipher, 168 imaginary part, 486
key, 167 inverse, 486
known plaintext attack, 170 magnitude, 487, 490
plaintext, 167 properties, 488
round, 171 modulus, see magnitude
S-box, 171 multiplication, 486
 INDEX 543

real part, 486 cryptanalysis, 167, 170

standard form, 486 differential, 174
subtraction, 486 linear, 174
complex vector space, 326 cryptography, 167
component current, 56
of a matrix, 121 curve fitting, 427, 476
of a vector, 62, 85 least squares line, 428
principal component, 475 least squares parabola, 429
component form linear regression, 428
line, 103
plane, 111 decryption, 167
componentwise method decryption matrix, 168
matrix multiplication, 131 degree
composite transformation, 231 of a polynomial, 491
composition of linear transformations, 231, 377 of freedom, 202
conjugate derivative, 299
of a complex number, 487 Descartes, René, 445
properties, 488 determinant, 241
conjugate symmetry and invertibility, 253
of inner product, 458 cofactor expansion, 245
consistent system, 9 cofactor of a matrix, 244
constant matrix, 14 Cramer’s rule, 266
constant polynomial, 491 expanding along row or column, 244
constant term, 7, 9 formula for matrix inverse, 260
continuous function, 343, 388 Laplace expansion, 245
coordinate minor of a matrix, 243
of a point, 61 of product, 254
with respect to basis, 198, 381 properties of, 255
coordinate system diagonal matrix, 284
and basis, 198, 381 diagonal of a matrix, 248
camera coordinates, 235 diagonal quadratic form, 450
Cartesian, 61 diagonalizable matrix, 285, 438
object coordinates, 236 orthogonally diagonalizable, 438
coordinate vector, 63 unitarily diagonalizable, 469
coplanar vectors, 91 diagonalization, 285
cosine orthogonal diagonalization, 438
of a matrix, 309 unitary diagonalization, 469
coupled variables, 300 differentiable function, 343
Cramer’s rule, 266 differential cryptanalysis, 174
cross product, 91 differential equation, 299, 344
area of parallelogram, 91 basic solution, 299
area of triangle, 94 change of variables, 301
coordinate description, 92 coupled variables, 300
geometric description, 91 eigenfrequency, 305
Jacobi identity, 98 first order, 299
properties, 92 linear, 299
544 INDEX

mode, 307 calculating, 276

resonance frequency, 305 complex, 318
second order, 299 elementary matrix, 153, 154
solution, 299 inverse, 155
system of, 300 elementary operation, 11
diffusion, 168 elementary row operation, 15, 153, 250
diffusion matrix, 172 empty linear combination, 178
dimension empty sum, 178
of a matrix, 121 encryption, 167
of a span, 202 encryption matrix, 168
of a vector, 62 entry of a matrix, 121
of solution space, 201 equality
of subspace of Rn , 201 of linear transformations, 375, 379
of vector space, 351 of matrices, 122
Dirac notation, 464 of vectors, 64
direction vector, 101, 102 equality of functions, 329
distance equation
point to line, 108 differential, see differential equation
point to plane, 118 equivalence
point to point, 76 row equivalence, 40
total squared distance, 473 system of linear equations, 11
distance formula, 76 error correcting code, 356
distributive law block length, 356
of fields, 43 check matrix, 360
over scalar addition, 70, 325 code block, 356
over vector addition, 70, 325 error pattern, 361
division generator matrix, 360
of complex numbers, 487 Hamming code, 364
dot product, 80 Hamming distance, 357
complex, 457 Hamming weight, 357
from matrix product, 163 linear code, 359
properties, 80 m-error correcting, 358
m-error detecting, 358
echelon form, 18 message block, 356
algorithm, 19 message length, 356
reduced, 29 (n, k, d)-code, 357
eigenfrequency, 305 rate, 356
eigenspace, 271 redundancy, 355
eigenvalue, 269 repetition code, 355
algebraic multiplicity, 312, 321 syndrome, 362
calculating, 276 syndrome decoding, 362
complex, 318 syndrome table, 362
geometric multiplicity, 312, 321 error pattern, 361
of triangular matrix, 279 Euclidean space, 62
eigenvector, 269 exchange lemma
basic, 271 in a vector space, 350
 INDEX 545

in Rn , 200 geometric multiplicity

exponential function of an eigenvalue, 312, 321
of a matrix, 309 Gram-Schmidt procedure, 404
extended casting-out algorithm, 182 complex, 462
Fibonacci sequence, 293 Hamming code, 364
field, 43 Hamming distance, 357
integers modulo p, 44 Hamming weight, 357
of complex numbers, 49, 487 Hamming, Richard, 364
finite-dimensional space, 351 harmonic, 421
finitely supported sequence, 341, 350 hermitian matrix, 468
fitting Hilbert space, 389
affine subspace fitting, 477 Hill cipher, 168
curve fitting, 427 homogeneous system, 34
line fitting, 428 associated, 37
subspace fitting, 473 dimension of solution space, 201
formula solution space, 191
closed, 294 Hooke’s law, 302
Fourier approximation, 417 hyperplane, 6, 505
Fourier coefficients, 398
Fourier series, 420 idempotent matrix, 141
generalized, 417 identity matrix, 136
free variable, 21, 23 identity transformation, 377
freedom (i, j)-entry of a matrix, 121
degree of, 202 image plane, 235
frequency imaginary part
resonance, 305 of a complex number, 486
function, 217 included angle, 82
affine, 237 inconsistent system, 9
continuous, 343, 388 infinite-dimensional space, 351
differentiable, 343 inner product, see also dot product, 387
equality of, 329 complex, 458
linear, see linear transformation inner product space
symmetric, 348 Cauchy-Schwarz inequality, 390, 459
vanishing at infinity, 348 complex, 458
vector function, 217 Hilbert space, 389
vector space of, 329 norm, 390, 459
zero function, 329 of continuous functions, 388
FuncX,K , 329 of polynomials, 389
fundamental theorem of algebra, 485 real, 387
Rn , 388
Gale Young, 473 triangle inequality, 391, 459
Gauss-Jordan elimination, 30 integer, 485
Gaussian elimination, 19 integers modulo p, 44, 47
general solution, 37 intersection
generalized Fourier series, 417 of a line and a plane, 116
generator matrix, 360 of two lines, 105
546 INDEX

of two planes, 116 syndrome, 362

inverse syndrome decoding, 362
additive, 43, 68, 325 syndrome table, 362
in a field, 43 linear combination
left inverse, 148 coefficient, 71
modulo p, 43 empty, 178
multiplicative, 43 in a vector space, 334
of a complex number, 486 in Rn , 71
of a linear transformation, 233, 377 of basic solutions, 36
of a matrix, 142, 149 of differential equation, 299
adjugate formula, 260 of matrices, 127, 334
does not exist, 146 of polynomials, 335
right inverse, 148 of vectors, 334
invertible matrix, 142 linear cryptanalysis, 174
and determinant, 253 linear dependence, 180, 337
isometry, 432, 466 linear differential equation, 299
linear equation, 7
Jacobi identity, 98 solution, 8
K-vector space, 326 standard form, 8
key linear function, see linear transformation
cryptography, 167 linear independence, 180
Kirchhoff’s law, 57 alternative characterization, 183
known plaintext attack, 170 casting-out algorithm, 181
extended, 182
Laplace expansion, 245 extending to basis, 203, 352
laws of nature, 302 in a vector space, 337
leading entry, 18 properties, 185
least squares approximation, 425 linear map, see linear transformation
least squares line, 428 linear regression, 428
least squares parabola, 429 linear transformation, 217, 369
left inverse, 148 addition, 375
Legendre polynomial, 408, 409 properties, 376
length of a vector, 77 composition, 231, 377
line properties, 377
angle between, 107 defined on a basis, 379
change of parameter, 103 equality of, 375, 379
component form, 103 identity transformation, 377
parametric equations, 104 inverse, 233, 377
symmetric form equation, 106 isometry, 432, 466
vector equation, 101, 102 matrix of
line fitting, 428 in Rn , 220
linear code, 359 w.r.t. basis, 382
check matrix, 360 matrix transformation, 220
error pattern, 361 on Rn , 217
generator matrix, 360 on vector spaces, 369
Hamming code, 364 orthogonal, 432
 INDEX 547

properties of addition, 376 properties, 255

properties of composition, 377 diagonal matrix, 284
properties of scalar multiplication, 376 diagonal of, 248
reflection, 226, 228 diagonalizable, 285, 438
rotation, 225, 227, 228, 231 orthogonally, 438
scalar multiplication, 375 unitarily, 469
properties, 376 diagonalization, 285
scaling, 228 diffusion, 172
shearing, 228 dimension of, 121
unitary, 466 eigenspace, 271
zero transformation, 375, 516 eigenvalue, 269
linearity calculating, 276
of inner product, 387, 458 complex, 318
LinV ,W , 376 eigenvector, 269
logical operation basic, 271
and, 44 calculating, 276
exclusive or, 44 complex, 318
lower triangular matrix, 248 elementary matrix, 153, 154
eigenvalues, 279 inverse, 155
elementary row operation, 15, 153, 250
magnitude encryption, 168
of a complex number, 487, 490 entry of, 121
properties, 488 equality, 122
of a vector, 77 exponential of, 309
map, see function finding the inverse, 145
matrix, 14, 121 hermitian, 468
addition, 123 idempotent, 141
properties, 125 identity, 136
adjugate, 259 (i, j)-entry, 121
antisymmetric, 163 inverse, 142, 149
augmented matrix, 14, 15 adjugate formula, 260
Cayley-Hamilton theorem, 315 does not exist, 146
characteristic polynomial, 275 left inverse, 148
root, 276 right inverse, 148
classical adjoint, 259 invertible, 142
coefficient matrix, 14, 131 and determinant, 253
cofactor matrix, 259 left inverse, 148
cofactor of, 244 lower triangular, 248
column space, 211, 426 eigenvalues of, 279
commuting matrices, 137 minor, 243
component of, 121 multiplication, 131
constant matrix, 14 column method, 133
decryption, 168 componentwise method, 131
determinant determinant of product, 254
and invertibility, 253 matrix times vector, 129
of product, 254 properties, 138
548 INDEX

row method, 135 eigenvalues of, 279

negative, 124 vector space of, 328
nilpotent, 257 zero matrix, 124
null space, 211 matrix form
nullity, 213 of a system of equations, 131
of a linear transformation matrix form of quadratic form, 449
in Rn , 220 matrix transformation, 220, 369
w.r.t. basis, 382 message block, 356
of a reflection, 226, 228 message length, 356
of a rotation, 225, 227, 228, 231 minor, 243
of a scaling, 228 Mm,n , 328
of a shearing, 228 mode of resonance, 307
orthogonal, 257, 434 modulo, 44
orthogonally diagonalizable, 438 modulus
positive definite, 442 of a complex number, see magnitude
positive semidefinite, 442 multiplication
properties of addition, 125 complex inner product, 458
properties of determinants, 255 in a field, 43
properties of multiplication, 138 inner product, 387
properties of scalar multiplication, 127 matrix times vector, 129
properties of transpose, 162 of complex numbers, 486
raising to a power, 290 of matrices, 131
rank, 24, 212 determinant of product, 254
rank theorem, 213 outer product, 464
rank-nullity theorem, 213 scalar multiplication, see scalar multiplication
right inverse, 148 multiplicative inverse, 43
row equivalence, 40, 211 of a complex number, 486
row operation, 15, 153, 250 multiplicative unit, 43
row space, 211 multiplicity
scalar multiplication, 127 of eigenvalue
properties, 127 algebraic, 312, 321
similar, 257, 285 geometric, 312, 321
sine and cosine of, 309 of root of a polynomial, 312
size of, 121 multivariable calculus, 217
skew symmetric, 163
square, 122 natural number, 485
square root of, 291 negative
subtraction, 124 of a matrix, 124
symmetric, 163, 347, 354, 437, 442 of a vector, 66
transpose, 162 nilpotent matrix, 257
properties, 162 (n, k, d)-code, 357
triangular, 248 non-trivial solution, 34
eigenvalues of, 279 norm
unitarily diagonalizable, 469 in Cn , 459
unitary, 467 in inner product space, 390
upper triangular, 248 in Rn , 77
 INDEX 549

normal equation of a plane, 113 particular solution, 37

normal vector of a plane, 112 periodic sequence, 348, 354
normalized vector, 391, 459 perpendicular vectors, 84
null space, 211 pivot column, 18
nullity, 213 pivot entry, 18
number pivot variable, 21, 23
complex, see complex number plaintext, 167
Fibonacci, 293 plane
integer, 485 angle between, 117
natural, 485 change of parameters, 119
rational, 485 component form, 111
real, 485 hyperplane, 6, 505
image plane, 235
object coordinates, 236 normal equation, 113
offset, 62 normal vector, 112
operator, see also linear transformation, 369 parametric equations, 111
origin of a coordinate system, 61 standard equation, 114
orthogonal basis, 397 vector equation, 111
Gram-Schmidt procedure, 404 point, 61
complex, 462 polynomial, 275, 326, 330, 491
orthogonal complement, 394 addition, 326, 330
orthogonal matrix, 257, 434 characteristic, see characteristic polynomial
orthogonal projection coefficient, 275, 491
onto subspace, 414 constant, 491
complex, 463 degree, 491
orthogonal set, 396, 459 Legendre polynomial, 408, 409
orthogonal transformation, 432 root, 276, 491
orthogonal vectors, 84, 394, 459 scalar multiplication, 326, 330
orthogonalization, 404 zero, 327
complex, 462 position vector, 63
orthogonally diagonalizable matrix, 438 positive definite matrix, 442
orthonormal basis, 397 positive definite property, 387, 458
orthonormal set, 396, 459 positive semidefinite matrix, 442
outer product, 464 power
P, 330 of a matrix, 290
basis of, 349 principal axis, 453
P2 , 326 principal component, 475
basis of, 349 product, see multiplication
parallelepiped, 95 projection
volume, 96 as a linear transformation, 222
parallelogram in inner product space, 411
area of, 91 onto subspace, 414
parameter, 22, 23, 101 complex, 463
parametric equations point to line, 108
of a line, 104 point to plane, 118
of a plane, 111 vector to vector, 85, 411
550 INDEX

proper subspace, 191 redundancy, 355

properties of addition redundant vector, 177, 179, 180
complex numbers, 487 removing, 187
field, 43 reflection
linear transformations, 376 about a given vector, 229
matrices, 125 matrix of, 226, 228
vectors, 68 René Descartes, 445
properties of complex conjugate, 488 rendering, 235
properties of composition repetition code, 355
linear transformations, 377 resistance, 56
properties of cross product, 92 resonance
properties of determinants, 255 frequency, 305
properties of dot product, 80 mode, 307
complex, 458 Richard Hamming, 364
properties of linear independence, 185 right inverse, 148
properties of multiplication right-handed system of vectors, 90
complex numbers, 487 Rn , 62
field, 43 roll call vote, 479
matrices, 138 root
properties of scalar multiplication of a polynomial, 276, 491
linear transformation, 376 rotation
matrix, 127 matrix of, 225, 227, 228, 231
vector, 70 rounds of a cipher, 171
row echelon form, see echelon form
quadratic form, 448 row equivalence, 40, 211
change of variables, 450 row method
coefficient, 448 matrix multiplication, 135
diagonal, 450 row operation, 15, 153, 250
matrix form, 449 row space, 211
principal axes, 453 row vector, 122
rank, 24, 34, 42, 212 rule for multiplication by 1, 70, 325
rank theorem, 213 S-box, 171
rank-nullity theorem, 213
scalar, 8
rate of a code, 356
for scaling, 69
rational number, 485
scalar multiplication
real inner product space, 387
closed under, 342
real number, 485 in a vector space, 325
real part
of a matrix, 127
of a complex number, 486
of a vector, 69, 325
real vector space, 326
of linear transformation, 375
recurrence, 293
of polynomials, 326, 330
as a subspace, 343, 352
scalar product, see dot product
base case, 293
scalar projection, 86
solving, 294
scaling
reduced echelon form, 29 matrix of, 228
algorithm, 30
 INDEX 551

semidefinite, 442 standard form

senate, 479 of a complex number, 486
SeqK , 330 of a linear equation, 8
sequence, 330 strongly alternating signs, 445
defined by a recurrence, 293 subspace, 190, 342
Fibonacci, 293 affine, 477
finitely supported, 341, 350 basis, see basis
periodic, 348, 354 dimension, 201
square summable, 389 eigenspace, 271
zero sequence, 344 is a span, 195
series of Rn , 190
Fourier series, 417, 420 of vector space, 342
shearing proper, 191
matrix of, 228 solution space of a homogeneous system, 191
sign span is a subspace, 191
strongly alternating, 445 zero subspace, 191
weakly alternating, 445 subspace fitting, 473
similar matrices, 257, 285 affine, 477
sine total squared distance, 473
of a matrix, 309 substitution cipher, 167
size of a matrix, 121 subtraction
skew lines, 6 of complex numbers, 486
skew symmetric matrix, 163 of matrices, 124
solution of vectors, 66
basic, 36, 198 sum, see addition
basis of, 198 of two angles, 232
non-trivial, 34 summation notation, 138
trivial, 34, 183 symmetric form, 106
solution space, 191 symmetric function, 348
span, 175, 176, 336 symmetric matrix, 163, 347, 354, 437, 442
as a subspace, 191 symmetry
dimension of, 202 conjugate symmetry, 458
spanning set, 337 of inner product, 387
shrink to basis, 205, 352 syndrome, 362
spring (mechanics), 302 syndrome decoding, 362
extension, 302 syndrome table, 362
Hooke’s law, 302 system of differential equations, 300
spring constant, 302 system of linear equations, 9
square matrix, 122 consistent, 9
square root Cramer’s rule, 266
of a matrix, 291 equivalence, 11
square summable sequence, 389 homogeneous, 34
squared distance, 473 dimension of solution space, 201
standard basis, 196 solution space, 191
standard cube, 236 inconsistent, 9
standard equation of a plane, 114 matrix form, 131
552 INDEX

non-trivial solution, 34 basis, see basis

solution, 9 cancellation law, 331
trivial solution, 34, 183 casting-out algorithm, 181
vector form, 129 extended, 182
column vector, 62, 122
tail of a vector, 62 commutative law of addition, 68, 325
tip of a vector, 62 component, 85, 86
total squared distance, 473 coplanar, 91
transpose of a matrix, 162 corresponding unit vector, 78
properties, 162 cross product
triangle properties, 92
area of, 94 dimension, 62
triangle inequality direction vector, 101, 102
complex, 459 distributive law, 70, 325
in complex numbers, 488 dot product, 80
in inner product space, 391 complex, 457
in Rn , 81 from matrix product, 163
triangular matrix, 248 properties, 80
eigenvalues, 279 eigenvector, 269
trigonometry complex, 318
sum of two angles, 232 empty linear combination, 178
trivial solution, 34, 183 equality, 64
tuple, 8 from a point to a point, 63
U.S. Senate, 479 geometric meaning, 62
unit in a vector space, 326
of addition, see additive unit, see additive unit length, 77
of multiplication, see multiplicative unit linearly dependent, 180, 337
unit vector, 78, 391 linearly independent, 180
unitarily diagonalizable matrix, 469 alternative characterization, 183
unitary matrix, 467 in a vector space, 337
unitary transformation, 466 properties, 185
United States Senate, 479 magnitude, 77
upper triangular matrix, 248 negative, 66
eigenvalues, 279 norm, 77, 390, 459
normal vector, 112
vanish at infinity, 348 normalized, 391, 459
variable, 7 normalizing a vector, 78
free, 21, 23, 38 orthogonal, 84, 394, 459
pivot, 21, 23, 38 perpendicular, 84
vector, 62 projection of, 85, 86, 411
addition, 65, 325 properties of addition, 68
properties, 68 properties of cross product, 92
additive inverse, 68, 325 properties of dot product, 80
additive unit, 68, 325 complex, 458
associative law of addition, 68, 325 properties of scalar multiplication, 70
associative law of scalar multiplication, 70, 325 redundant, 177, 179, 180
 INDEX 553

removing, 187 zero matrix, 124

right-handed system of, 90 zero polynomial, 327
row vector, 122 zero sequence, 344
rule for multiplication by 1, 70, 325 zero subspace, 191, 342
scalar multiplication, 69, 325 zero transformation, 375, 516
properties, 70 zero vector, 62, 68, 178
span, 175, 176, 336
as a subspace, 191
subtraction, 66
tail, 62
tip, 62
unit vector, 78, 391
zero vector, 68, 178
vector equation
of a line, 101, 102
component form, 103
of a plane, 111
component form, 111
vector form
of a system of equations, 129
vector function, 217
linear, see linear transformation
vector projection, 86
vector space, 325
basis, 349
complex, 326
dimension, 351
finite-dimensional, 351
infinite-dimensional, 351
linear independence, 337
of linear transformations, 376
of matrices, 328
of polynomials, 326, 330
of sequences, 330
real, 326
span, 336
spanning set, 337
voltage, 56
volume
of parallelepiped, 96
voting, 479

weakly alternating signs, 445

Young, Gale, 473

zero function, 329