Linear Algebra - Jim Hefferon
Linear Algebra - Jim Hefferon
Linear Algebra - Jim Hefferon
1 2
3 1
x 1 2
x 3 1
Jim Hefferon
2
1
Notation
R
N
C
{. . . . . . }
h. . . i
V, W, U
~v , w
~
~0, ~0V
B, D
En = h~e1 , . . . , ~en i
~ ~
,
RepB (~v )
Pn
Mnm
[S]
M N
V
=W
h, g
H, G
t, s
T, S
RepB,D (h)
hi,j
|T |
R(h), N (h)
R (h), N (h)
real numbers
natural numbers: {0, 1, 2, . . . }
complex numbers
set of . . . such that . . .
sequence; like a set but order matters
vector spaces
vectors
zero vector, zero vector of V
bases
standard basis for Rn
basis vectors
matrix representing the vector
set of n-th degree polynomials
set of nm matrices
span of the set S
direct sum of subspaces
isomorphic spaces
homomorphisms
matrices
transformations; maps from a space to itself
square matrices
matrix representing the map h
matrix entry from row i, column j
determinant of the matrix T
rangespace and nullspace of the map h
generalized rangespace and nullspace
symbol
name
iota
kappa
lambda
mu
nu
xi
omicron
pi
symbol
name
rho
sigma
tau
upsilon
phi
chi
psi
omega
symbol
Preface
In most mathematics programs linear algebra is taken in the first or second
year, following or along with at least one course in calculus. While the location
of this course is stable, lately the content has been under discussion. Some instructors have experimented with varying the traditional topics, trying courses
focused on applications, or on the computer. Despite this (entirely healthy)
debate, most instructors are still convinced, I think, that the right core material
is vector spaces, linear maps, determinants, and eigenvalues and eigenvectors.
Applications and computations certainly can have a part to play but most mathematicians agree that the themes of the course should remain unchanged.
Not that all is fine with the traditional course. Most of us do think that
the standard text type for this course needs to be reexamined. Elementary
texts have traditionally started with extensive computations of linear reduction,
matrix multiplication, and determinants. These take up half of the course.
Finally, when vector spaces and linear maps appear, and definitions and proofs
start, the nature of the course takes a sudden turn. In the past, the computation
drill was there because, as future practitioners, students needed to be fast and
accurate with these. But that has changed. Being a whiz at 55 determinants
just isnt important anymore. Instead, the availability of computers gives us an
opportunity to move toward a focus on concepts.
This is an opportunity that we should seize. The courses at the start of
most mathematics programs work at having students correctly apply formulas
and algorithms, and imitate examples. Later courses require some mathematical
maturity: reasoning skills that are developed enough to follow different types
of proofs, a familiarity with the themes that underly many mathematical investigations like elementary set and function facts, and an ability to do some
independent reading and thinking, Where do we work on the transition?
Linear algebra is an ideal spot. It comes early in a program so that progress
made here pays off later. The material is straightforward, elegant, and accessible. The students are serious about mathematics, often majors and minors.
There are a variety of argument stylesproofs by contradiction, if and only if
statements, and proofs by induction, for instanceand examples are plentiful.
The goal of this text is, along with the development of undergraduate linear
algebra, to help an instructor raise the students level of mathematical sophistication. Most of the differences between this book and others follow straight
from that goal.
One consequence of this goal of development is that, unlike in many computational texts, all of the results here are proved. On the other hand, in contrast
with more abstract texts, many examples are given, and they are often quite
detailed.
Another consequence of the goal is that while we start with a computational
topic, linear reduction, from the first we do more than just compute. The
solution of linear systems is done quickly but it is also done completely, proving
i
everything (really these proofs are just verifications), all the way through the
uniqueness of reduced echelon form. In particular, in this first chapter, the
opportunity is taken to present a few induction proofs, where the arguments
just go over bookkeeping details, so that when induction is needed later (e.g., to
prove that all bases of a finite dimensional vector space have the same number
of members), it will be familiar.
Still another consequence is that the second chapter immediately uses this
background as motivation for the definition of a real vector space. This typically
occurs by the end of the third week. We do not stop to introduce matrix
multiplication and determinants as rote computations. Instead, those topics
appear naturally in the development, after the definition of linear maps.
To help students make the transition from earlier courses, the presentation
here stresses motivation and naturalness. An example is the third chapter,
on linear maps. It does not start with the definition of homomorphism, as
is the case in other books, but with the definition of isomorphism. Thats
because this definition is easily motivated by the observation that some spaces
are just like each other. After that, the next section takes the reasonable step of
defining homomorphisms by isolating the operation-preservation idea. A little
mathematical slickness is lost, but it is in return for a large gain in sensibility
to students.
Having extensive motivation in the text helps with time pressures. I ask
students to, before each class, look ahead in the book, and they follow the
classwork better because they have some prior exposure to the material. For
example, I can start the linear independence class with the definition because I
know students have some idea of what it is about. No book can take the place
of an instructor, but a helpful book gives the instructor more class time for
examples and questions.
Much of a students progress takes place while doing the exercises; the exercises here work with the rest of the text. Besides computations, there are many
proofs. These are spread over an approachability range, from simple checks
to some much more involved arguments. There are even a few exercises that
are reasonably challenging puzzles taken, with citation, from various journals,
competitions, or problems collections (as part of the fun of these, the original
wording has been retained as much as possible). In total, the questions are
aimed to both build an ability at, and help students experience the pleasure of,
doing mathematics.
Applications, and Computers. The point of view taken here, that linear
algebra is about vector spaces and linear maps, is not taken to the exclusion
of all other ideas. Applications, and the emerging role of the computer, are
interesting, important, and vital aspects of the subject. Consequently, every
chapter closes with a few application or computer-related topics. Some of the
topics are: network flows, the speed and accuracy of computer linear reductions,
Leontief Input/Output analysis, dimensional analysis, Markov chains, voting
paradoxes, analytic projective geometry, and solving difference equations.
These are brief enough to be done in a days class or to be given as indepenii
dent projects for individuals or small groups. Most simply give a reader a feel
for the subject, discuss how linear algebra comes in, point to some accessible
further reading, and give a few exercises. I have kept the exposition lively and
given an overall sense of breadth of application. In short, these topics invite
readers to see for themselves that linear algebra is a tool that a professional
must have.
For people reading this book on their own. The emphasis on motivation
and development make this book a good choice for self-study. While a professional mathematician knows what pace and topics suit a class, perhaps an
independent student would find some advice helpful. Here are two timetables
for a semester. The first focuses on core material.
week
1
2
3
4
5
6
7
8
9
10
11
12
13
14
Mon.
1.I.1
1.I.3
1.III.1,
2.I.2
2.III.1,
2.III.2,
3.I.2
3.II.2
3.III.1
3.IV.2,
3.IV.4,
4.I.3
4.III.1
5.II.2
2
2
3
3, 4
3.V.1
Wed.
1.I.1, 2
1.II.1
1.III.2
2.II
2.III.2
2.III.3
3.II.1
3.II.2
3.III.2
3.IV.4
3.V.1, 2
4.II
5.I
5.II.3
Fri.
1.I.2, 3
1.II.2
2.I.1
2.III.1
exam
3.I.1
3.II.2
3.III.1
3.IV.1, 2
exam
4.I.1, 2
4.II
5.II.1
review
The second timetable is more ambitious (it presupposes 1.II, the elements of
vectors, usually covered in third semester calculus).
week
1
2
3
4
5
6
7
8
9
10
11
12
13
14
Mon.
1.I.1
1.I.3
2.I.1
2.III.1
2.III.4
3.I.2
3.III.1
3.IV.2
3.V.1
3.VI.2
4.I.2
4.II
5.II.1, 2
5.III.2
Wed.
1.I.2
1.III.1, 2
2.I.2
2.III.2
3.I.1
3.II.1
3.III.2
3.IV.3
3.V.2
4.I.1
4.I.3
4.II, 4.III.1
5.II.3
5.IV.1, 2
Fri.
1.I.3
1.III.2
2.II
2.III.3
exam
3.II.2
3.IV.1, 2
3.IV.4
3.VI.1
exam
4.I.4
4.III.2, 3
5.III.1
5.IV.2
Jim Hefferon
Saint Michaels College
Colchester, Vermont USA
[email protected]
April 20, 2000
Authors Note. Inventing a good exercise, one that enlightens as well as tests,
is a creative act, and hard work (at least half of the the effort on this text
has gone into exercises and solutions). The inventor deserves recognition. But,
somehow, the tradition in texts has been to not give attributions for questions.
I have changed that here where I was sure of the source. I would greatly appreciate hearing from anyone who can help me to correctly attribute others of the
questions. They will be incorporated into later versions of this book.
iv
Contents
1 Linear Systems
1.I Solving Linear Systems . . . . . . . . . .
1.I.1 Gauss Method . . . . . . . . . . . . .
1.I.2 Describing the Solution Set . . . . . .
1.I.3 General = Particular + Homogeneous
1.II Linear Geometry of n-Space . . . . . . . .
1.II.1 Vectors in Space . . . . . . . . . . . .
1.II.2 Length and Angle Measures . . . . .
1.III Reduced Echelon Form . . . . . . . . . .
1.III.1 Gauss-Jordan Reduction . . . . . . . .
1.III.2 Row Equivalence . . . . . . . . . . . .
Topic: Computer Algebra Systems . . . . . . .
Topic: Input-Output Analysis . . . . . . . . .
Topic: Accuracy of Computations . . . . . . .
Topic: Analyzing Networks . . . . . . . . . . .
2 Vector Spaces
2.I Definition of Vector Space . . . . . . .
2.I.1 Definition and Examples . . . . . .
2.I.2 Subspaces and Spanning Sets . . .
2.II Linear Independence . . . . . . . . . .
2.II.1 Definition and Examples . . . . . .
2.III Basis and Dimension . . . . . . . . . .
2.III.1 Basis . . . . . . . . . . . . . . . . .
2.III.2 Dimension . . . . . . . . . . . . . .
2.III.3 Vector Spaces and Linear Systems
2.III.4 Combining Subspaces . . . . . . .
Topic: Fields . . . . . . . . . . . . . . . . .
Topic: Crystals . . . . . . . . . . . . . . . .
Topic: Voting Paradoxes . . . . . . . . . .
Topic: Dimensional Analysis . . . . . . . .
v
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
1
1
2
11
20
32
32
38
45
45
51
61
63
67
72
.
.
.
.
.
.
.
.
.
.
.
.
.
.
79
80
80
91
102
102
113
113
119
124
131
141
143
147
152
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
159
159
159
169
176
176
184
194
194
204
211
211
214
221
230
238
238
242
250
250
255
260
269
274
280
286
4 Determinants
4.I Definition . . . . . . . . . . . . . . . .
4.I.1 Exploration . . . . . . . . . . . .
4.I.2 Properties of Determinants . . . .
4.I.3 The Permutation Expansion . . . .
4.I.4 Determinants Exist . . . . . . . .
4.II Geometry of Determinants . . . . . .
4.II.1 Determinants as Size Functions . .
4.III Other Formulas . . . . . . . . . . . . .
4.III.1 Laplaces Expansion . . . . . . .
Topic: Cramers Rule . . . . . . . . . . . .
Topic: Speed of Calculating Determinants .
Topic: Projective Geometry . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
293
294
294
299
303
312
319
319
326
326
331
334
337
5 Similarity
5.I Complex Vector Spaces . . . . . . . . . . . . . .
5.I.1 Factoring and Complex Numbers; A Review
5.I.2 Complex Representations . . . . . . . . . . .
5.II Similarity . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
347
347
348
350
351
vi
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
vii
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
Powers
. . . . .
. . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
351
353
357
365
365
368
379
379
386
399
403
405
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
A-1
A-1
A-1
A-3
A-5
A-6
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Chapter 1
Linear Systems
1.I
Systems of linear equations are common in science and mathematics. These two
examples from high school science [Onan] give a sense of how they arise.
The first example is from Physics. Suppose that we are given three objects,
one with a mass of 2 kg, and are asked to find the unknown masses. Suppose
further that experimentation with a meter stick produces these two balances.
40
50
25
50
2
15
25
Now, since the sum of moments on the left of each balance equals the sum of
moments on the right (the moment of an object is its mass times its distance
from the balance point), the two balances give this system of two equations.
40h + 15c = 100
25c = 50 + 50h
The second example of a linear system is from Chemistry. We can mix,
under controlled conditions, toluene C7 H8 and nitric acid HNO3 to produce
trinitrotoluene C7 H5 O6 N3 along with the byproduct water (conditions have to
be controlled very well, indeed trinitrotoluene is better known as TNT). In
what proportion should those components be mixed? The number of atoms of
each element present before the reaction
x C7 H8 + y HNO3
z C7 H5 O6 N3 + w H2 O
must equal the number present afterward. Applying that principle to the elements C, H, N, and O in turn gives this system.
7x = 7z
8x + 1y = 5z + 2w
1y = 3z
3y = 6z + 1w
1
1.I.1
Gauss Method
multiply row 1 by 3
1
3 x1
x1
+ 2x2
=3
+ 5x2 2x3 = 2
3x3 = 9
x1 + 6x2
=9
x1 + 5x2 2x3 = 2
3x3 = 9
x1 + 6x2
= 9
x2 2x3 = 7
3x3 = 9
The third step is the only nontrivial one. Weve mentally multiplied both sides
of the first row by 1, mentally added that to the old second row, and written
the result in as the new second row.
Now we can find the value of each variable. The bottom equation shows
that x3 = 3. Substituting 3 for x3 in the middle equation shows that x2 = 1.
Substituting those two into the top equation gives that x1 = 3 and so the system
has a unique solution: the solution set is { (3, 1, 3) }.
Most of this subsection and the next one consists of examples of solving
linear systems by Gauss method. We will use it throughout this book. It is
fast and easy. But, before we get to those examples, we will first show that
this method is also safe in that it never loses solutions or picks up extraneous
solutions.
1.4 Theorem (Gauss method) If a linear system is changed to another by
one of these operations
(1) an equation is swapped with another
(2) an equation has both sides multiplied by a nonzero constant
(3) an equation is replaced by the sum of itself and a multiple of another
then the two systems have the same set of solutions.
Each of those three operations has a restriction. Multiplying a row by 0 is
not allowed because obviously that can change the solution set of the system.
Similarly, adding a multiple of a row to itself is not allowed because adding 1
times the row to itself has the effect of multiplying the row by 0. Finally, swapping a row with itself is disallowed to make some results in the fourth chapter
easier to state and remember (and besides, self-swapping doesnt accomplish
anything).
Proof. We will cover the equation swap operation here and save the other two
cases for Exercise 29.
.
aj,1 x1 + aj,2 x2 + aj,n xn = dj
ai,1 x1 + ai,2 x2 +
..
.
am,1 x1 + am,2 x2 + am,n xn = dm
am,1 x1 + am,2 x2 +
a1,n xn = d1
..
.
aj,n xn = dj
..
.
ai,n xn = di
..
.
am,n xn = dm
The n-tuple (s1 , . . . , sn ) satisfies the system before the swap if and only if
substituting the values, the ss, for the variables, the xs, gives true statements:
a1,1 s1 +a1,2 s2 + +a1,n sn = d1 and . . . ai,1 s1 +ai,2 s2 + +ai,n sn = di and . . .
aj,1 s1 + aj,2 s2 + + aj,n sn = dj and . . . am,1 s1 + am,2 s2 + + am,n sn = dm .
In a requirement consisting of statements and-ed together we can rearrange
the order of the statements, so that this requirement is met if and only if a1,1 s1 +
a1,2 s2 + + a1,n sn = d1 and . . . aj,1 s1 + aj,2 s2 + + aj,n sn = dj and . . .
ai,1 s1 + ai,2 s2 + + ai,n sn = di and . . . am,1 s1 + am,2 s2 + + am,n sn = dm .
This is exactly the requirement that (s1 , . . . , sn ) solves the system after the row
QED
swap.
1.5 Definition The three operations from Theorem 1.4 are the elementary reduction operations, or row operations, or Gaussian operations. They are swapping, multiplying by a scalar or rescaling, and pivoting.
When writing out the calculations, we will abbreviate row i by i . For
instance, we will denote a pivot operation by ki + j , with the row that is
changed written second. We will also, to save writing, often list pivot steps
together when they use the same i .
1.6 Example A typical use of Gauss method is to solve this system.
x+ y
=0
2x y + 3z = 3
x 2y z = 3
The first transformation of the system involves using the first row to eliminate
the x in the second row and the x in the third. To get rid of the second rows
2x, we multiply the entire first row by 2, add that to the second row, and
write the result in as the new second row. To get rid of the third rows x, we
multiply the first row by 1, add that to the third row, and write the result in
as the new third row.
1 +3
21 +2
x+
y
=0
3y + 3z = 3
3y z = 3
(Note that the two 1 steps 21 + 2 and 1 + 3 are written as one operation.) In this second system, the last two equations involve only two unknowns.
To finish we transform the second system into a third system, where the last
equation involves only one unknown. This transformation uses the second row
to eliminate y from the third row.
2 +3
x+
y
3y +
=0
3z = 3
4z = 0
Now we are set up for the solution. The third row shows that z = 0. Substitute
that back into the second row to get y = 1, and then substitute back into the
first row to get x = 1.
1.7 Example For the Physics problem from the start of this chapter, Gauss
method gives this.
40h + 15c = 100
50h + 25c = 50
5/41 +2
40h +
15c = 100
(175/4)c = 175
21 +2
31 +3
(3/2)2 +3
x+ y+ z= 9
2y 5z = 17
3y 8z = 27
x+ y+
2y
z=
9
5z = 17
12 z = 32
xy
21 +2
=0
z + 2w = 4
y
+ w=0
2z + w = 5
the second equation has no leading y. To get one, we look lower down in the
system for a row that has a leading y and swap it in.
2 3
xy
y
=0
+ w=0
z + 2w = 4
2z + w = 5
(Had there been more than one row below the second with a leading y then we
could have swapped in any one.) The rest of Gauss method goes as before.
xy
y
23 +4
+
z+
= 0
w= 0
2w = 4
3w = 3
21 +3
x+
3y = 1
5y = 5
4y = 4
x+
3y = 1
5y = 5
0= 0
That examples system has more equations than variables. Gauss method
is also useful on systems with more variables than equations. Many examples
are in the next subsection.
Another way that linear systems can differ from the examples shown earlier
is that some linear systems do not have a unique solution. This can happen in
two ways.
The first is that it can fail to have any solution at all.
1.12 Example Contrast the system in the last example with this one.
x + 3y = 1
2x + y = 3
2x + 2y = 0
21 +2
x+
21 +3
3y = 1
5y = 5
4y = 2
Here the system is inconsistent: no pair of numbers satisfies all of the equations
simultaneously. Echelon form makes this inconsistency obvious.
(4/5)2 +3
x+
3y = 1
5y = 5
0= 2
21 +2
x + 2y = 8
0 = 8
The other way that a linear system can fail to have a unique solution is to
have many solutions.
1.14 Example In this system
x+ y=4
2x + 2y = 8
any pair of numbers satisfying the first equation automatically satisfies the second. The solution set {(x, y) x + y = 4} is infinite some of its members
are (0, 4), (1, 5), and (2.5, 1.5). The result of applying Gauss method here
contrasts with the prior example because we do not get a contradictory equation.
21 +2
x+y=4
0=0
2x
2z = 6
y+ z=1
y+ z=1
3y + 3z = 0
2x
2z = 6
y+ z= 1
0= 0
0 = 3
1 +3
2 +3
32 +4
We will finish this subsection with a summary of what weve seen so far
about Gauss method.
Gauss method uses the three row operations to set a system up for back
substitution. If any step shows a contradictory equation then we can stop
with the conclusion that the system has no solutions. If we reach echelon form
without a contradictory equation, and each variable is a leading variable in its
row, then the system has a unique solution and we find it by back substitution.
Finally, if we reach echelon form without a contradictory equation, and there is
not a unique solution (at least one variable is not a leading variable) then the
system has many solutions.
The next subsection deals with the third case we will see how to describe
the solution set of a system with many solutions.
Exercises
X 1.16 Use Gauss method to find the unique solution for each system.
x
z=0
2x + 3y = 13
=1
(a)
(b) 3x + y
x y = 1
x + y + z = 4
X 1.17 Use Gauss method to solve each system or conclude many solutions or no
solutions.
(a) 2x + 2y = 5
(b) x + y = 1
(c) x 3y + z = 1
x 4y = 0
x+y=2
x + y + 2z = 14
(d) x y = 1
(e)
4y + z = 20
(f ) 2x
+ z+w= 5
3x 3y = 2
2x 2y + z = 0
y
w = 1
x
+z= 5
3x
zw= 0
x + y z = 10
4x + y + 2z + w = 9
X 1.18 There are methods for solving linear systems other than Gauss method. One
often taught in high school is to solve one of the equations for a variable, then
substitute the resulting expression into other equations. That step is repeated
until there is an equation with only one variable. From that, the first number in
the solution is derived, and then back-substitution can be done. This method both
takes longer than Gauss method, since it involves more arithmetic operations and
is more likely to lead to errors. To illustrate how it can lead to wrong conclusions,
we will use the system
x + 3y = 1
2x + y = 3
2x + 2y = 0
from Example 1.12.
(a) Solve the first equation for x and substitute that expression into the second
equation. Find the resulting y.
(b) Again solve the first equation for x, but this time substitute that expression
into the third equation. Find this y.
What extra step must a user of this method take to avoid erroneously concluding
a system has a solution?
X 1.19 For which values of k are there no solutions, many solutions, or a unique
solution to this system?
x y=1
3x 3y = k
X 1.20 This system is not linear:
2 sin cos + 3 tan = 3
4 sin + 2 cos 2 tan = 10
6 sin 3 cos + tan = 9
but we can nonetheless apply Gauss method. Do so. Does the system have a
solution?
X 1.21 What conditions must the constants, the bs, satisfy so that each of these
systems has a solution? Hint. Apply Gauss method and see what happens to the
right side.
(a) x 3y = b1
(b) x1 + 2x2 + 3x3 = b1
3x + y = b2
2x1 + 5x2 + 3x3 = b2
x + 7y = b3
x1
+ 8x3 = b3
2x + 4y = b4
1.22 True or false: a system with more unknowns than equations has at least one
solution. (As always, to say true you must prove it, while to say false you must
produce a counterexample.)
1.23 Must any Chemistry problem like the one that starts this subsection a
balance the reaction problem have infinitely many solutions?
X 1.24 Find the coefficients a, b, and c so that the graph of f (x) = ax2 + bx + c passes
through the points (1, 2), (1, 6), and (2, 3).
10
11
(a) 19
(b) 21
(c) 23
(d) 29
(e) 17
X 1.35 [Am. Math. Mon., Jan. 1935] Laugh at this: AHAHA + TEHE = TEHAW.
It resulted from substituting a code letter for each digit of a simple example in
addition, and it is required to identify the letters and prove the solution unique.
1.36 [Wohascum no. 2] The Wohascum County Board of Commissioners, which has
20 members, recently had to elect a President. There were three candidates (A, B,
and C); on each ballot the three candidates were to be listed in order of preference,
with no abstentions. It was found that 11 members, a majority, preferred A over
B (thus the other 9 preferred B over A). Similarly, it was found that 12 members
preferred C over A. Given these results, it was suggested that B should withdraw,
to enable a runoff election between A and C. However, B protested, and it was
then found that 14 members preferred B over C! The Board has not yet recovered
from the resulting confusion. Given that every possible order of A, B, C appeared
on at least one ballot, how many members voted for B as their first choice?
1.37 [Am. Math. Mon., Jan. 1963] This system of n linear equations with n unknowns, said the Great Mathematician, has a curious property.
Good heavens! said the Poor Nut, What is it?
Note, said the Great Mathematician, that the constants are in arithmetic
progression.
Its all so clear when you explain it! said the Poor Nut. Do you mean like
6x + 9y = 12 and 15x + 18y = 21?
Quite so, said the Great Mathematician, pulling out his bassoon. Indeed,
the system has a unique solution. Can you find it?
Good heavens! cried the Poor Nut, I am baffled.
Are you?
1.I.2
A linear system with a unique solution has a solution set with one element.
A linear system with no solution has a solution set that is empty. In these cases
the solution set is easy to describe. Solution sets are a challenge to describe
only when they contain many elements.
2.1 Example This system has many solutions because in echelon form
2x
+z=3
xyz=1
3x y
=4
(1/2)1 +2
2x
+
z=
3
y (3/2)z = 1/2
y (3/2)z = 1/2
2x
+
z=
3
y (3/2)z = 1/2
0=
0
(3/2)1 +3
2 +3
not all of the variables are leading variables. The Gauss method theorem
showed that a triple satisfies the first system if and only if it satisfies the third.
Thus, the solution set {(x, y, z) 2x + z = 3 and x y z = 1 and 3x y = 4}
12
can also be described as {(x, y, z) 2x + z = 3 and y 3z/2 = 1/2}. However, this second description is not much of an improvement. It has two equations instead of three, but it still involves some hard-to-understand interaction
among the variables.
To get a description that is free of any such interaction, we take the variable that does not lead any equation, z, and use it to describe the variables
that do lead, x and y. The second equation gives y = (1/2) (3/2)z and
the first equation gives x = (3/2) (1/2)z. Thus, the solution
set can be de
scribed as {(x, y, z) = ((3/2) (1/2)z, (1/2) (3/2)z, z) z R}. For instance,
(1/2, 5/2, 2) is a solution because taking z = 2 gives a first component of 1/2
and a second component of 5/2.
The advantage of this description over the ones above is that the only variable
appearing, z, is unrestricted it can be any real number.
2.2 Definition The non-leading variables in an echelon-form linear system are
free variables.
In the echelon form system derived in the above example, x and y are leading
variables and z is free.
2.3 Example A linear system can end with more than one variable free. This
row reduction
x+ y+ z w= 1
y z + w = 1
3x
+ 6z 6w = 6
y + z w = 1
x+
31 +3
32 +3
2 +4
y+ z w= 1
y z + w = 1
3y + 3z 3w = 3
y + z w = 1
x+y+zw= 1
y z + w = 1
0= 0
0= 0
ends with x and y leading, and with both z and w free. To get the description
that we prefer we will start at the bottom. We first express y in terms of
the free variables z and w with y = 1 + z w. Next, moving up to the
top equation, substituting for y in the first equation x + (1 + z w) + z
w = 1 and solving for x yields x = 2 2z + 2w. Thus, the solution set is
{2 2z + 2w, 1 + z w, z, w) z, w R}.
We prefer this description because the only variables that appear, z and w,
are unrestricted. This makes the job of deciding which four-tuples are system
solutions into an easy one. For instance, taking z = 1 and w = 2 gives the
solution (4, 2, 1, 2). In contrast, (3, 2, 1, 2) is not a solution, since the first
component of any solution must be 2 minus twice the third component plus
twice the fourth.
13
=0
z + 3w = 2
3x 3y
=0
x y + 2z + 6w = 4
2x 2y
=0
z + 3w = 2
0=0
2z + 6w = 4
2x 2y
=0
z + 3w = 2
0=0
0=0
(3/2)1 +3
(1/2)1 +4
22 +4
x and z lead, y and w are free. The solution set is {(y, y, 2 3w, w) y, w R}.
For instance, (1, 1, 2, 0) satisfies the system take y = 1 and w = 0. The
four-tuple (1, 0, 5, 4) is not a solution since its first coordinate does not equal its
second.
We refer to a variable used to describe a family of solutions as a parameter
and we say that the set above is paramatrized with y and w. (The terms
parameter and free variable do not mean the same thing. Above, y and w
are free because in the echelon form system they do not lead any row. They
are parameters because they are used in the solution set description. We could
have instead paramatrized with y and z by rewriting the second equation as
w = 2/3 (1/3)z. In that case, the free variables are still y and w, but the
parameters are y and z. Notice that we could not have paramatrized with x and
y, so there is sometimes a restriction on the choice of parameters. The terms
parameter and free are related because, as we shall show later in this chapter,
the solution set of a system can always be paramatrized with the free variables.
Consequenlty, we shall paramatrize all of our descriptions in this way.)
2.5 Example This is another system with infinitely many solutions.
x + 2y
=1
2x
+z
=2
3x + 2y + z w = 4
21 +2
x+
2y
=1
4y + z
=0
4y + z w = 1
x+
2y
4y + z
31 +3
2 +3
=1
=0
w = 1
The leading variables are x, y, and w. The variable z is free. (Notice here that,
although there are infinitely many solutions, the value of one of the variables is
fixed w = 1.) Write w in terms of z with w = 1 + 0z. Then y = (1/4)z.
To express x in terms of z, substitute for y into the first equation to get x =
1 (1/2)z. The solution set is {(1 (1/2)z, (1/4)z, z, 1) z R}.
We finish this subsection by developing the notation for linear systems and
their solution sets that we shall use in the rest of this book.
2.6 Definition An mn matrix is a rectangular array of numbers with m rows
and n columns. Each number in the matrix is an entry,
14
Matrices are usually named by upper case roman letters, e.g. A. Each entry is
denoted by the corresponding lower-case letter, e.g. ai,j is the number in row i
and column j of the array. For instance,
1 2.2 5
A=
3 4 7
has two rows and three columns, and so is a 2 3 matrix. (Read that twoby-three; the number of rows is always stated first.) The entry in the second
row and first column is a2,1 = 3. Note that the order of the subscripts matters:
a1,2 6= a2,1 since a1,2 = 2.2. (The parentheses around the array are a typographic device so that when two matrices are side by side we can tell where one
ends and the other starts.)
2.7 Example We can abbreviate this linear system
=4
x1 + 2x2
x2 x3 = 0
+ 2x3 = 4
x1
with this matrix.
1 2
0 1
1 0
0
1
2
4
0
4
The vertical bar just reminds a reader of the difference between the coefficients
on the systemss left hand side and the constants on the right. When a bar
is used to divide a matrix into parts, we call it an augmented matrix. In this
notation, Gauss method goes this way.
1 2
0 4
1 2 0 4
1 2 0 4
1 +3
2 +3
0 1 1 0 2
0 1 1 0
0 1 1 0
1 0 2 4
0 2 2 0
0 0 0 0
The second row stands for y z= 0 and the first row stands for x + 2y = 4 so
the solution set is {(4 2z, z, z) z R}. One advantage of the new notation is
that the clerical load of Gauss method the copying of variables, the writing
of +s and =s, etc. is lighter.
We will also use the array notation to clarify the descriptions
of solution
sets. A description like {(2 2z + 2w, 1 + z w, z, w) z, w R} from Example 2.3 is hard to read. We will rewrite it to group all the constants together,
all the coefficients of z together, and all the coefficients of w together. We will
write them vertically, in one-column wide matrices.
2
2
2
1 1
1
{ + z +
w z, w R}
0
1
0
0
0
1
15
For instance, the top line says that x = 2 2z + 2w. The next section gives a
geometric interpretation that will help us picture the solution sets when they
are written in this way.
2.8 Definition A vector (or column vector) is a matrix with a single column.
A matrix with a single row is a row vector. The entries of a vector are its
components.
Vectors are an exception to the convention of representing matrices with
capital roman letters. We use lower-case roman or greek letters overlined with
~ . . . (boldface is also common: a or ). For
an arrow: ~a, ~b, . . . or
~ , ,
instance, this is a column vector with a third component of 7.
1
~v = 3
7
2.9 Definition The linear equation a1 x1 + a2 x2 + + an xn = d with unknowns x1 , . . . , xn is satisfied by
s1
~s = ...
sn
if a1 s1 + a2 s2 + + an sn = d. A vector satisfies a linear system if it satisfies
each equation in the system.
The style of description of solution sets that we use involves adding the
vectors, and also multiplying them by real numbers, such as the z and w. We
need to define these operations.
2.10 Definition The vector sum of ~u and ~v is this.
u1
v1
u 1 + v1
..
~u + ~v = ... + ... =
.
un
vn
u n + vn
In general, two matrices with the same number of rows and the same number
of columns add in this way, entry-by-entry.
2.11 Definition The scalar multiplication of the real number r and the vector
~v is this.
rv1
v1
r ~v = r ... = ...
vn
rvn
16
1
7
4 28
7
1 = 7
3
21
Notice that the definitions of vector addition and scalar multiplication agree
where they overlap, for instance, ~v + ~v = 2~v .
With the notation defined, we can now solve systems in the way that we will
use throughout this book.
2.13 Example This system
2x + y
w
=4
y
+ w+u=4
x
z + 2w
=0
reduces in
2 1
0 1
1 0
this way.
0 1
0
1
1 2
0
1
0
4
4
0
(1/2)1 +3
(1/2)2 +3
2
0
0
2
0
0
1
1
1/2
1 0
1 0
0 1
0
1
0
1
1 5/2
1
1
3
0
1
1/2
0
1
0
4
4
2
4
4
0
1/2
1
0
x
1
y 4 1
{ z = 0 + 3 w +
1/2 u w, u R}
0
w 0 1
1
0
0
u
Note again how well vector notation sets off the coefficients of each parameter.
For instance, the third row of the vector form shows plainly that if u is held
fixed then z increases three times as fast as w.
That format also shows plainly that there are infinitely many solutions. For
example, we can fix u as 0, let w range over the real numbers, and consider the
first component x. We get infinitely many first components and hence infinitely
many solutions.
17
Another thing shown plainly is that setting both w and u to zero gives that
this
0
x
y 4
z = 0
w 0
0
u
is a particular solution of the linear system.
2.14 Example In the same way, this system
x y+ z=1
3x
+ z=3
5x 2y + 3z = 5
reduces
1 1 1
3 0 1
5 2 3
1
1 1
31 +2
3 0 3
51 +3
5
0 3
1
2
2
1 1
1
2 +3
0 0 3
0
0 0
1
2
0
1
0
0
1
1/3
{0 + 2/3 z z R}
0
1
Before the exercises, we pause to point out some things that we have yet to
do.
The first two subsections have been on the mechanics of Gauss method.
Except for one result, Theorem 1.4 without which developing the method
doesnt make sense since it says that the method gives the right answers we
have not stopped to consider any of the interesting questions that arise.
For example, can we always describe solution sets as above, with a particular
solution vector added to an unrestricted linear combination of some other vectors? The solution sets we described with unrestricted parameters were easily
seen to have infinitely many solutions so an answer to this question could tell
us something about the size of solution sets. An answer to that question could
also help us picture the solution sets what do they look like in R2 , or in R3 ,
etc?
Many questions arise from the observation that Gauss method can be done
in more than one way (for instance, when swapping rows, we may have a choice
of which row to swap with). Theorem 1.4 says that we must get the same
solution set no matter how we proceed, but if we do Gauss method in two
different ways must we get the same number of free variables both times, so
that any two solution set descriptions have the same number of parameters?
18
Must those be the same variables (e.g., is it impossible to solve a problem one
way and get y and w free or solve it another way and get y and z free)?
In the rest of this chapter we answer these questions. The answer to each
is yes. The first question is answered in the last subsection of this section. In
the second section we give a geometric description of solution sets. In the final
section of this chapter we tackle the last set of questions.
Consequently, by the end of the first chapter we will not only have a solid
grounding in the practice of Gauss method, we will also have a solid grounding
in the theory. We will be sure of what can and cannot happen in a reduction.
Exercises
X 2.15 Find the indicated entry of the matrix, if it is defined.
A=
(a) a2,1
(b) a1,2
(c) a2,2
1
2
3
1
1
4
(d) a3,1
1
1
1 0 4
5 10
1
(a)
(b) 1
(c)
2 1 5
10 5
3
1
X 2.17 Do the indicated vector operation, if it is defined.
! !
! !
2
3
1
3
4
2
3
(a) 1 + 0
(b) 5
(c) 5 1
(d) 7
+9
1
1
5
1
4
1
1
!
1
3
2
1
1
(e)
+ 2
(f ) 6 1 4 0 + 2 1
2
3
1
3
5
X 2.18 Solve each system using matrix notation. Express the solution using vectors.
(a) 3x + 6y = 18
(b) x + y = 1
(c) x1
+ x3 = 4
x + 2y = 6
x y = 1
x1 x2 + 2x3 = 5
4x1 x2 + 5x3 = 17
(d) 2a + b c = 2
(e) x + 2y z
=3
(f ) x
+z+w=4
2a
+c=3
2x + y
+w=4
2x + y
w=2
ab
=0
x y+z+w=1
3x + y + z
=7
X 2.19 Solve each system using matrix notation. Give each solution set in vector
notation.
(a) 2x + y z = 1
(b) x
z
=1
(c) x y + z
=0
4x y
=3
y + 2z w = 3
y
+w=0
x + 2y + 3z w = 7
3x 2y + 3z + w = 0
y
w=0
(d) a + 2b + 3c + d e = 1
3a b + c + d + e = 3
X 2.20 The vector is in the set. What value of the parameters produces that vector?
5
1
(a)
,{
k k R}
5
1
19
1
2
3
2 , { 1 i + 0 j i, j R}
(b)
1
0
1
! !
!
0
1
2
(c) 4 , { 1 m + 0 n m, n R}
2
0
1
2.21 Decide
if
the
vector
is
in the set.
3
6
(a)
,{
k k R}
1
2
5
5
,{
j j R}
4
4
! ! !
2
0
1
1 ,{ 3
(c)
+ 1 r r R}
1
7
3
! !
!
1
2
3
(d) 0 , { 0 j + 1 k j, k R}
1
1
1
2.22 Paramatrize the solution set of this one-equation system.
(b)
x1 + x2 + + xn = 0
X 2.23
x + 2y
w=a
2x
+z
=b
x+ y
+ 2w = c
for x, y, z, and w, in terms of the constants a, b, and c.
(b) Use your answer from the prior part to solve this.
x + 2y
w= 3
2x
+z
= 1
x+ y
+ 2w = 2
X 2.24 Why is the comma needed in the notation ai,j for matrix entries?
X 2.25 Give the 44 matrix whose i, j-th entry is
(a) i + j;
(b) 1 to the i + j power.
2.26 For any matrix A, the transpose of A, written Atrans , is the matrix whose
columns are the rows of A. Find the transpose of each of these.
!
1
1 2 3
2 3
5 10
(a)
(b)
(c)
(d) 1
4 5 6
1
1
10 5
0
X 2.27 (a) Describe all functions f (x) = ax2 + bx + c such that f (1) = 2 and
f (1) = 6.
(b) Describe all functions f (x) = ax2 + bx + c such that f (1) = 2.
2.28 Show that any set of five points from the plane R2 lie on a common conic
section, that is, they all satisfy some equation of the form ax2 + by 2 + cxy + dx +
ey + f = 0 where some of a, . . . , f are nonzero.
2.29 Make up a four equations/four unknowns system having
(a) a one-parameter solution set;
(b) a two-parameter solution set;
(c) a three-parameter solution set.
20
1.I.3
2.7
8.9
19.3
11.3
10.8
Alcohol
Benzene
Glycerine
Water
0.81
0.90
1.26
1.00
The prior subsection has many descriptions of solution sets. They all fit a
pattern. They have a vector that is a particular solution of the system added
to an unrestricted combination of some other vectors. The solution set from
Example 2.13 illustrates.
1/2
1
0
1
1
4
+ w 3 + u 1/2 w, u R}
0
{
0
1
0
1
0
0
|
{z
}
| {z }
particular
solution
unrestricted
combination
21
~1 + + ck
~k c1 , . . . , ck R}
{~
p + c1
where p~ is any particular solution, and where the system has k free variables.
This description has two parts, the particular solution p~ and also the un~
restricted linear combination of the s.
We shall prove the theorem in two
corresponding parts, with two lemmas.
We will focus first on the unrestricted combination part. To do that, we
consider systems that have the vector of zeroes as one of the particular solutions,
~1 + + ck
~k can be shortened to c1
~1 + + ck
~k .
so that p~ + c1
3.2 Definition A linear equation is homogeneous if it has a constant of zero,
that is, if it can be put in the form a1 x1 + a2 x2 + + an xn = 0.
(These are homogeneous because all of the terms involve the same power of
their variable the first power including a 0x0 that we can imagine is on
the right side.)
3.3 Example With any linear system like
3x + 4y = 3
2x y = 1
we associate a system of homogeneous equations by setting the right side to
zeros.
3x + 4y = 0
2x y = 0
Our interest in the homogeneous system associated with a linear system can be
understood by comparing the reduction of the system
3x + 4y = 3
2x y = 1
(2/3)1 +2
3x +
4y = 3
(11/3)y = 1
(2/3)1 +2
3x +
4y = 0
(11/3)y = 0
Obviously the two reductions go in the same way. We can study how linear systems are reduced by instead studying how the associated homogeneous systems
are reduced.
22
21 +2
3x + 2y +
z=0
3x + 2y + z = 0
2 3
2z = 0
y+ z=0
y+ z=0
2z = 0
3.6 Example Some homogeneous systems have many solutions. One example
is the Chemistry problem from the first page of this book.
7x
7j
=0
8x + y 5j 2k = 0
y 3j
=0
3y 6j k = 0
7x
7z
=0
y + 3z 2w = 0
y 3z
=0
3y 6z w = 0
7x
y+
7x
7z
=0
y + 3z 2w = 0
6z + 2w = 0
0=0
(8/7)1 +2
2 +3
32 +4
(5/2)3 +4
7z
=0
3z 2w = 0
6z + 2w = 0
15z + 5w = 0
1/3
1
{
1/3 w k R}
1
has many vectors besides the zero vector (if we interpret w as a number of
molecules then solutions make sense only when w is a nonnegative multiple of
3).
We now have the terminology to prove the two parts of Theorem 3.1. The
first lemma deals with unrestricted combinations.
23
~1 , . . . ,
3.7 Lemma For any homogeneous linear system there exist vectors
~
k such that the solution set of the system is
~1 + + ck
~k c1 , . . . , ck R}
{c1
where k is the number of free variables in an echelon form version of the system.
Before the proof, we will recall the back substitution calculations that were
done in the prior subsection. Imagine that we have brought a system to this
echelon form.
x+
2y z + 2w = 0
3y + z
=0
w = 0
24
the bottom t rows in terms of free variables, then we can express the leading
variable of the next row up the t + 1-th row up from the bottom in terms
of free variables. With those two steps, the theorem will be proved because by
the base step it is true for the bottom equation, and by the inductive step the
fact that it is true for the bottom equation shows that it is true for the next
one up, and then another application of the inductive step implies it is true for
third equation up, etc.
For the base step, consider the bottom-most non-0 = 0 equation (the case
where all the equations are 0 = 0 is trivial). We call that the m-th row:
am,`m x`m + am,`m +1 x`m +1 + + am,n xn = 0
where am,`m 6= 0. (The notation here has ` stand for leading, so am,`m means
the coefficient, from the row m of the variable leading row m.) Either there
are variables in this equation other than the leading one x`m or else there are
not. If there are other variables x`m +1 , etc., then they must be free variables
because this is the bottom non-0 = 0 row. Move them to the right and divide
by am,`m
x`m = (am,`m +1 /am,`m )x`m +1 + + (am,n /am,`m )xn
to expresses this leading variable in terms of free variables. If there are no free
variables in this equation then x`m = 0 (see the tricky point noted following
this proof).
For the inductive step, we assume that for the m-th equation, and for the
(m 1)-th equation, . . . , and for the (m t)-th equation, we can express the
leading variable in terms of free variables (where 0 t < m). To prove that the
same is true for the next equation up, the (m (t + 1))-th equation, we take
each variable that leads in a lower-down equation x`m , . . . , x`mt and substitute
its expression in terms of free variables. The result has the form
am(t+1),`m(t+1) x`m(t+1) + sums of multiples of free variables = 0
where am(t+1),`m(t+1) 6= 0. We move the free variables to the right-hand side
and divide by am(t+1),`m(t+1) , to end with x`m(t+1) expressed in terms of free
variables.
Because we have shown both the base step and the inductive step, by the
QED
principle of mathematical induction the proposition is true.
~1 + + ck
~k c1 , . . . , ck R} is generated by or
We say that the set {c1
~
spanned by the set of vectors {1 , . . . , ~k }. There is a tricky point to this
definition. If a homogeneous system has a unique solution, the zero vector,
then we say the solution set is generated by the empty set of vectors. This fits
with the pattern of the other solution sets: in the proof above the solution set is
derived by taking the cs to be the free variables and if there is a unique solution
then there are no free variables.
This proof incidentally shows, as discussed after Example 2.4, that solution
sets can always be paramatrized using the free variables.
25
The next lemma finishes the proof of Theorem 3.1 by considering the particular solution part of the solution sets description.
3.8 Lemma For a linear system, where p~ is any particular solution, the solution set equals this set.
{~
p + ~h ~h satisfies the associated homogeneous system}
Proof. We will show mutual set inclusion, that any solution to the system is
in the above set and that anything in the set is a solution to the system.
For set inclusion the first way, that if a vector solves the system then it is in
the set described above, assume that ~s solves the system. Then ~s p~ solves the
associated homogeneous system since for each equation index i between 1 and
n,
QED
The two lemmas above together establish Theorem 3.1. We remember that
theorem with the slogan General = Particular + Homogeneous.
3.9 Example This system illustrates Theorem 3.1.
x + 2y z = 1
2x + 4y
=2
y 3z = 0
Gauss method
21 +2
x + 2y z = 1
x + 2y z = 1
2 3
2z = 0
y 3z = 0
y 3z = 0
2z = 0
26
21 +2 2 3
x + 2y z = 0
y 3z = 0
2z = 0
21 +2
1 +3
+ z + w = 1
y 2z w = 5
y + 2z + w = 2
21 +2 2 +3
1 +3
+ z+w=0
y 2z w = 0
0=0
{
1 z + 0 w z, w R}
0
1
However, because no particular solution of the original system exists, the general
solution set is empty there are no vectors of the form p~ + ~h because there are
no p~ s.
3.11 Corollary Solution sets of linear systems are either empty, have one
element, or have infinitely many elements.
27
Proof. Weve seen examples of all three happening so we need only prove that
those are the only possibilities.
First, notice a homogeneous system with at least one non-~0 solution ~v has
infinitely many solutions because the set of multiples s~v is infinite if s 6= 1
then s~v ~v = (s 1)~v is easily seen to be non-~0, and so s~v 6= ~v .
Now, apply Lemma 3.8 to conclude that a solution set
{~
p + ~h ~h solves the associated homogeneous system}
is either empty (if there is no particular solution p~), or has one element (if there
is a p~ and the homogeneous system has the unique solution ~0), or is infinite (if
there is a p~ and the homogeneous system has a non-~0 solution, and thus by the
QED
prior paragraph has infinitely many solutions).
This table summarizes the factors affecting the size of a general solution.
number of solutions of the
associated homogeneous system
particular
solution
exists?
yes
no
one
unique
solution
no
solutions
infinitely many
infinitely many
solutions
no
solutions
The factor on the top of the table is the simpler one. When we perform
Gauss method on a linear system, ignoring the constants on the right side and
so paying attention only to the coefficients on the left-hand side, we either end
with every variable leading some row or else we find that some variable does not
lead a row, that is, that some variable is free. (Of course, ignoring the constants
on the right is formalized by considering the associated homogeneous system.
We are simply putting aside for the moment the possibility of a contradictory
equation.)
A nice insight into the factor on the top of this table at work comes from considering the case of a system having the same number of equations as variables.
This system will have a solution, and the solution will be unique, if and only if it
reduces to an echelon form system where every variable leads its row, which will
happen if and only if the associated homogeneous system has a unique solution.
Thus, the question of uniqueness of solution is especially interesting when the
system has the same number of equations as variables.
3.12 Definition A square matrix is nonsingular if it is the matrix of coefficients of a homogeneous system with a unique solution. It is singular otherwise,
that is, if it is the matrix of coefficients of a homogeneous system with infinitely
many solutions.
28
3.13 Example The systems from Example 3.3, Example 3.5, and Example 3.9
each have an associated homogeneous system with a unique solution. Thus these
matrices are nonsingular.
1 2 1
3 2 1
3 4
6 4 0
2 4 0
2 1
0 1 3
0 1 1
The Chemistry problem from Example 3.6 is a homogeneous system with more
than one solution so its matrix is singular.
7 0 7 0
8 1 5 2
0 1 3 0
0 3 6 1
3.14 Example The first of these matrices is nonsingular while the second is
singular
1 2
1 2
3 4
3 6
because the first of these homogeneous systems has a unique solution while the
second has infinitely many solutions.
x + 2y = 0
3x + 4y = 0
x + 2y = 0
3x + 6y = 0
We have made the distinction in the definition because a system (with the same
number of equations as variables) behaves in one of two ways, depending on
whether its matrix of coefficients is nonsingular or singular. A system where
the matrix of coefficients is nonsingular has a unique solution for any constants
on the right side: for instance, Gauss method shows that this system
x + 2y = a
3x + 4y = b
has the unique solution x = b 2a and y = (3a b)/2. On the other hand, a
system where the matrix of coefficients is singular never has a unique solutions
it has either no solutions or else has infinitely many, as with these.
x + 2y = 1
3x + 6y = 2
x + 2y = 1
3x + 6y = 3
29
considering the systems left-hand side the the constants on the right-hand
side play no role in this factor. The tables other factor, determining whether a
particular solution exists, is tougher. Consider these two
3x + 2y = 5
3x + 2y = 5
3x + 2y = 5
3x + 2y = 4
with the same left sides but different right sides. Obviously, the first has a
solution while the second does not, so here the constants on the right side
decide if the system has a solution. We could conjecture that the left side of a
linear system determines the number of solutions while the right side determines
if solutions exist, but that guess is not correct. Compare these two systems
3x + 2y = 5
4x + 2y = 4
3x + 2y = 5
and 3x + 2y = 4
with the same right sides but different left sides. The first has a solution but
the second does not. Thus the constants on the right side of a system dont
decide alone whether a solution exists; rather, it depends on some interaction
between the left and right sides.
For some intuition about that interaction, consider this system with one of
the coefficients left as the parameter c.
x + 2y + 3z = 1
x+ y+ z=1
cx + 3y + 4z = 0
If c = 2 this system has no solution because the left-hand side has the third row
as a sum of the first two, while the right-hand does not. If c 6= 2 this system has
a unique solution (try it with c = 1). For a system to have a solution, if one row
of the matrix of coefficients on the left is a linear combination of other rows,
then on the right the constant from that row must be the same combination of
constants from the same rows.
More intuition about the interaction comes from studying linear combinations. That will be our focus in the second chapter, after we finish the study of
Gauss method itself in the rest of this chapter.
Exercises
X 3.15 Solve each system. Express the solution set using vectors. Identify the particular solution and the solution set of the homogeneous system.
(a) 3x + 6y = 18
(b) x + y = 1
(c) x1
+ x3 = 4
x + 2y = 6
x y = 1
x1 x2 + 2x3 = 5
4x1 x2 + 5x3 = 17
(d) 2a + b c = 2
(e) x + 2y z
=3
(f ) x
+z+w=4
2a
+c=3
2x + y
+w=4
2x + y
w=2
ab
=0
x y+z+w=1
3x + y + z
=7
3.16 Solve each system, giving the solution set in vector notation. Identify the
particular solution and the solution of the homogeneous system.
30
z
=1
y + 2z w = 3
x + 2y + 3z w = 7
(b) x
(c)
y+ z
=0
y
+w=0
3x 2y + 3z + w = 0
y
w=0
a + 2b + 3c + d e = 1
3a b + c + d + e = 3
X 3.17 For the system
2x y
w= 3
y + z + 2w = 2
x 2y z
= 1
which of these can be used as the particular solution part of some general solution?
0
2
1
3
1
4
(a)
(b)
(c)
5
1
8
0
0
1
X 3.18 Lemma 3.8 says that any particular solution may be used for p
~. Find, if
possible, a general solution to this system
x y
+w=4
2x + 3y z
=0
y+z+w=4
that uses
the
given vector
as its particular
solution.
0
5
2
0
1
1
(a)
(b)
(c)
0
7
1
4
10
1
3.19 One of these is nonsingular while the other is singular. Which is which?
(d)
1
3
1 3
(b)
4 12
4 12
X 3.20 Singular
or nonsingular?
1 2
1
2
1 2 1
(a)
(b)
(c)
(Careful!)
1 3
3 6
1 3 1
!
1 2 1
2
2 1
1
0 5
(d) 1 1 3
(e)
3 4 7
1 1 4
X 3.21 Isthe
given
vector
in the set generated by the given set?
2
1
1
(a)
,{
,
}
3
4
5
! ! !
1
2
1
0 ,{ 1 , 0 }
(b)
1
0
1
! ! ! ! !
1
1
2
3
4
(c) 3 , { 0 , 1 , 3 , 2 }
0
4
5
0
1
1
2
3
0 1 0
(d) , { , }
1
0
0
1
1
2
(a)
31
3.22 Prove that any linear system with a nonsingular matrix of coefficients has a
solution, and that the solution is unique.
3.23 To tell the whole truth, there is another tricky point to the proof of Lemma 3.7.
What happens if there are no non-0 = 0 equations? (There arent any more tricky
points after this one.)
X 3.24 Prove that if ~s and ~t satisfy a homogeneous system then so do these vectors.
(a) ~s + ~t
(b) 3~s
(c) k~s + m~t for k, m R
Whats wrong with: These three show that if a homogeneous system has one
solution then it has many solutions any multiple of a solution is another solution,
and any sum of solutions is a solution also so there are no homogeneous systems
with exactly one solution.?
3.25 Prove that if a system with only rational coefficients and constants has a
solution then it has at least one all-rational solution. Must it have infinitely many?
32
1.II
For readers who have seen the elements of vectors before, in calculus or physics,
this section is an optional review. However, later work in this book will refer to
this material often, so this section is not optional if it is not a review.
In the first section, we had to do a bit of work to show that there are only
three types of solution sets singleton, empty, and infinite. But for systems
with two equations and two unknowns, we can just see this. We picture each
two-unknowns equation as a line in R2 and then the two lines could have a
unique intersection, be parallel, or be the same.
One solution
No solutions
3x + 2y = 7
x y = 1
Infinitely many
solutions
3x + 2y = 7
3x + 2y = 4
3x + 2y = 7
6x + 4y = 14
As this shows, sometimes our results are expressed clearly in a picture. In this
section we develop the terminology and ideas we need to express our results
from the prior section, and from some future sections, geometrically. The twodimensional case is familiar enough, but to extend to systems with more than
two unknowns we shall also need some higher-dimensional geometry.
1.II.1
Vectors in Space
and make the usual correspondence with R: pick a point to label 0 and another
to label 1.
0
Now, armed with a scale and a direction, finding the point corresponding to,
say +2.17, is easy start at 0, head in the direction of 1 (i.e., the positive
direction), but dont stop there, go 2.17 times as far.
The basic idea here, combining magnitude with direction, is the key to extending to higher dimensions.
33
are equal, even though they start in different places, because they have equal
lengths and equal directions. Again: those vectors are not just alike, they are
equal.
How can things that are in different places be equal? Think of a vector as
representing a displacement (vector is Latin for carrier or traveler). These
squares undergo the same displacement, despite that those displacements start
in different places.
Sometimes, to emphasize this property vectors have of not being anchored, they
are referred to as free vectors.
These two, as free vectors, are equal;
we can think of each as a displacement of one over and two up. More generally,
two vectors in the plane are the same if and only if they have the same change
in first components and the same change in second components: the vector
extending from (a1 , a2 ) to (b1 , b2 ) equals the vector from (c1 , c2 ) to (d1 , d2 ) if
and only if b1 a1 = d1 c1 and b2 a2 = d2 c2 .
An expression like the vector that, were it to start at (a1 , a2 ), would stretch
to (b1 , b2 ) is awkward. Instead of that terminology, from among all of these
we single out the one starting at the origin as being in canonical (or natural)
position and we describe a vector by stating its endpoint when it is in canonical
34
position, as a column. For instance, the one over and two up vectors above are
denoted in this way.
1
2
More generally, the plane vector starting at (a1 , a2 ) and stretching to (b1 , b2 ) is
denoted
b1 a1
b2 a2
since the prior paragraph shows that when the vector starts at the origin, it
ends at this location.
We often just say the point
1
2
rather than the endpoint of the canonical position of that vector. That is, we
shall find it convienent to blur the distinction between a point in space and the
vector that, if it starts at the origin, ends at that point. Thus, we will refer to
both of these as Rn .
x
{ 1 x1 , x2 R}
{(x1 , x2 ) x1 , x2 R}
x2
In the prior section we defined vectors and vector operations with an algebraic motivation;
rv1
v1
w1
v1 + w1
v1
=
+
=
r
v2
rv2
v2
w2
v2 + w2
we can now interpret those operations geometrically. For instance, if ~v represents a displacement then 3~v represents a displacement in the same direction
but three times as far, and 1~v represents a displacement of the same distance
as ~v but in the opposite direction.
~v
3~v
~v
w
~
35
The long arrow is the combined displacement in this sense: if, in one minute, a
ships motion gives it the displacement relative to the earth of ~v and a passengers motion gives a displacement relative to the ships deck of w,
~ then ~v + w
~ is
the displacement of the passenger relative to the earth.
Another way to understand the vector sum is with the parallelogram rule.
Draw the parallelogram formed by the vectors ~v1 , ~v2 and then the sum ~v1 + ~v2
extends along the diagonal to the far corner.
x1 + x2
y1 + y2
x2
y2
x1
y1
The above drawings show how vectors and vector operations behave in R2 .
We can extend to R3 , or to even higher-dimensional spaces where we have no
pictures, with the obvious generalization: the free vector that, if it starts at
(a1 , . . . , an ), ends at (b1 , . . . , bn ), is represented by this column
b1 a1
..
.
bn an
(vectors are equal if they have the same representation), we arent too careful
to distinguish between a point and the vector whose canonical representation
ends at that point,
v1
..
n
R = { . v1 , . . . , vn R}
vn
and addition and scalar multiplication are component-wise.
Having considered points, we now turn to the lines. In R2 , the line through
(1, 2) and (3, 1) is comprised of (the endpoints of) the vectors in this set
1
2
t R}
{
+t
2
1
That description expresses this picture.
2
1
3
1
1
2
The vector associated with the parameter t has its whole body in the line it
is a direction vector for the line. Note that points on the line to the left of x = 1
are described using negative values of t.
36
In R3 , the line through (1, 2, 3) and (5, 5, 5) is the set of (endpoints of)
vectors of this form
1
4
{2 + t 3 t R}
3
2
and lines in even higher-dimensional spaces work in the same way.
If a line uses one parameter, so that there is freedom to move back and
forth in one dimension, then a plane must involve two. For example, the plane
through the points (1, 0, 5), (2, 1, 3), and (2, 4, 0.5) consists of (endpoints of)
the vectors in
1
1
3
{0 + t 1 + s 4 t, s R}
5
8
4.5
(the column vectors associated with the parameters
1
2
1
3
2
1
1 = 1 0
4 = 4 0
8
3
5
4.5
0.5
5
are two vectors whose whole bodies lie in the plane). As with the line, note that
some points in this plane are described with negative ts or negative ss or both.
A description of planes that is often encountered in algebra and calculus uses
a single equation
x
P = {y 2x + 3y z = 4}
z
as the condition that describes the relationship among the first, second, and
third coordinates of points in a plane. The translation from such a description
to the vector description that we favor in this book is to think of the condition
as a one-equation linear system and paramatrize x = (1/2)(4 3y + z).
2
3/2
1/2
P = {0 + 1 y + 0 z y, z R}
0
0
1
Generalizing from lines and planes, we define a k-dimensional
linear sur
p + t1~v1 + t2~v2 + + tk~vk t1 , . . . , tk R} where
face (or k-flat) in Rn to be {~
~v1 , . . . , ~vk Rn . For example, in R4 ,
2
1
0
t R}
{
+ t
3
0
0.5
0
37
is a line,
0
1
2
0
1
0
{
0 + t 0 + s 1 t, s R}
0
1
0
is a plane, and
3
0
1
2
1
0
0
0
{
2 + r 0 + s 1 + t 1 r, s, t R}
0.5
1
0
0
is a three-dimensional linear surface. Again, the intuition is that a line permits motion in one direction, a plane permits motion in combinations of two
directions, etc.
A linear surface description can be misleading about the dimension this
1
1
2
0
1
2
t, s R}
L = { + t + s
1
0
0
2
1
2
is a degenerate plane because it is actually a line.
1
1
0
1
L = {
1 + r 0 r R}
2
1
We shall see in the Linear Independence section of Chapter Two what relationships among vectors causes the linear surface they generate to be degenerate.
We finish this subsection by restating our conclusions from the first section
in geometric terms. First, the solution set of a linear system with n unknowns
is a linear surface in Rn . Specifically, it is an k-dimensional linear surface,
where k is the number of free variables in an echelon form version of the system.
Second, the solution set of a homogeneous linear system is a linear surface
passing through the origin. Finally, we can view the general solution set of any
linear system as being the solution set of its associated homogeneous system
offset from the origin by a vector, namely by any particular solution.
Exercises
X 1.1 Find the canonical name for each vector.
(a) the vector from (2, 1) to (4, 2) in R2
(b) the vector from (3, 3) to (2, 5) in R2
(c) the vector from (1, 0, 6) to (5, 0, 3) in R3
(d) the vector from (6, 8, 8) to (6, 8, 8) in R3
38
0
{ 0
+ 1 t t R}
3
4
2
X 1.6 Intersect these planes.
!
!
1
0
{ 1 t + 1 s t, s R}
1
3
1
{ 1
0
!
+
0
3
0
!
k+
2
0
4
m k, m R}
(b) { 0 + t 1
t R}, {s 1 + w 4 s, w R}
1
1
2
1
1.8 Show that the line segments (a1 , a2 )(b1 , b2 ) and (c1 , c2 )(d1 , d2 ) have the same
lengths and slopes if b1 a1 = d1 c1 and b2 a2 = d2 c2 . Is that only if?
1.9 How should R0 be defined?
X 1.10 [Math. Mag., Jan. 1957] A person traveling eastward at a rate of 3 miles per
hour finds that the wind appears to blow directly from the north. On doubling his
speed it appears to come from the north east. What was the winds velocity?
1.11 Euclid describes a plane as a surface which lies evenly with the straight lines
on itself. Commentators (e.g., Heron) have interpreted this to mean (A plane
surface is) such that, if a straight line pass through two points on it, the line
coincides wholly with it at every spot, all ways. (Translations from [Heath], pp.
171-172.) Do planes, as described in this section, have that property? Does this
description adequately define planes?
1.II.2
Weve translated the first sections results about solution sets into geometric
terms for insight into how those sets look. But we must watch out not
to be
k
t R}
of
the
forms
{~
p
+
t~
v
mislead by our own
terms;
labeling
subsets
of
R
and {~
p + t~v + sw
~ t, s R} as lines and planes doesnt make them act like
the lines and planes of our prior experience. Rather, we must ensure that the
names suit the sets. While we cant prove that the sets satisfy our intuition
we cant prove anything about intuition in this subsection well observe that
39
~
v
~
u
Put them in canonical position and, in the plane that they determine, consider
the triangle formed by ~u, ~v , and ~u ~v .
u 1 v1 + u 2 v2 + u 3 v3
)
k~u k k~v k
In higher dimensions no picture suffices but we can make the same argument
analytically. First, the form of the numerator is clear it comes from the middle
terms of the squares (u1 v1 )2 , (u2 v2 )2 , etc.
2.3 Definition The dot product (or inner product, or scalar product) of two
n-component real vectors is the linear combination of their components.
~u ~v = u1 v1 + u2 v2 + + un vn
40
Notice that the dot product of two vectors is a real number, not a vector, and
that the dot product of a vector from Rn with a vector from Rm is defined
only when n equals m. Notice also this relationship between dot product and
length: dotting a vector with itself gives its length squared ~u ~u = u1 u1 + +
un un = k~u k2 .
2.4 Remark The wording in that definition allows one or both of the two to
be a row vector instead of a column vector. Some books require that the first
vector be a row vector and that the second vector be a column vector. We shall
not be that strict.
Still reasoning with letters, but guided by the pictures, we use the next
theorem to argue that the triangle formed by ~u, ~v , and ~u ~v in Rn lies in the
planar subset of Rn generated by ~u and ~v .
2.5 Theorem (Triangle Inequality) For any ~u, ~v Rn ,
k~u + ~v k k~u k + k~v k
with equality if and only if one of the vectors is a nonnegative scalar multiple
of the other one.
This inequality is the source of the familiar saying, The shortest distance
between two points is in a straight line.
. finish
~
u + ~v
~v
start .
~
u
Proof. Well use some algebraic properties of dot product that we have not
shown, for instance that ~u (~a +~b) = ~u ~a + ~u ~b and that ~u ~v = ~v ~u. Verification
of those properties is Exercise 17. The desired inequality holds if and only if its
square holds.
41
.Q
Because the Triangle Inequality says that in any Rn , the shortest cut between
two endpoints is simply the line segment connecting them, linear surfaces have
no such bends.
Back to the definition of angle measure. The heart of the Triangle Inequalitys proof is the ~u ~v k~u k k~v k line. At first glance, a reader might wonder
if some pairs of vectors satisfy the inequality in this way: while ~u ~v is a large
number, with absolute value bigger than the right-hand side, it is a negative
large number. The next result says that no such pair of vectors exists.
2.6 Corollary (Cauchy-Schwartz Inequality) For any ~u, ~v Rn ,
|~u ~v | k~u k k~v k
with equality if and only if one vector is a scalar multiple of the other.
Proof. The Triangle Inequalitys proof shows that ~
u ~v k~u k k~v k so if ~u ~v is
positive or zero then we are done. If ~u ~v is negative then this holds.
QED
42
~u ~v
)
k~u k k~v k
(the angle between the zero vector and any other vector is defined to be a right
angle).
Thus vectors from Rn are orthogonal if and only if their dot product is zero.
2.8 Example These vectors are orthogonal.
1
1
1
=0
1
Although they are shown away from canonical position so that they dont appear
to touch, nonetheless they are orthogonal.
2.9 Example The R3 angle formula given at the start of this subsection is a
special case of the definition. Between these two
0
3
2
1
1
0
the angle is
arccos(
3
(1)(0) + (1)(3) + (0)(2)
) = arccos( )
2
2
2
2
2
2
2 13
1 +1 +0 0 +3 +2
approximately 0.94 radians. Notice that these vectors are not orthogonal. Although the yz-plane may appear to be perpendicular to the xy-plane, in fact
the two planes are that way only in the weak sense that there are vectors in each
orthogonal to all vectors in the other. Not every vector in each is orthogonal to
all vectors in the other.
Exercises
X 2.10 Find the length of each vector.
(a)
3
1
(b)
1
2
(c)
4
1
1
!
(d)
0
0
0
1
1
(e)
1
0
43
! !
1
0
1
1
(a)
(b) 2 , 4
(c)
, 4
2
0
1
1
X 2.12 During maneuvers preceding the Battle of Jutland, the British battle cruiser
Lion moved as follows (in nautical miles): 1.2 miles north, 6.1 miles 38 degrees
east of south, 4.0 miles at 89 degrees east of north, and 6.5 miles at 31 degrees
east of north. Find the distance between starting and ending positions.
2.13 Find k so that these two vectors are perpendicular.
1
1
,
2
4
k
1
4
3
1
3
1
X 2.15 (a) Find the angle between the diagonal of the unit square in R2 and one of
the axes.
(b) Find the angle between the diagonal of the unit cube in R3 and one of the
axes.
(c) Find the angle between the diagonal of the unit cube in Rn and one of the
axes.
(d) What is the limit, as n goes to , of the angle between the diagonal of the
unit cube in Rn and one of the axes?
2.16 Is there any vector that is perpendicular to itself?
X 2.17 Describe the algebraic properties of dot product.
(a) Is it right-distributive over addition: (~
u + ~v ) w
~ =~
u w
~ + ~v w?
~
(b) Is is left-distributive (over addition)?
(c) Does it commute?
(d) Associate?
(e) How does it interact with scalar multiplication?
As always, any assertion must be backed by either a proof or an example.
2.18 Verify the equality condition in Corollary 2.6, the Cauchy-Schwartz Inequality.
(a) Show that if ~
u is a negative scalar multiple of ~v then ~
u ~v and ~v ~
u are less
than or equal to zero.
(b) Show that |~
u ~v | = k~
u k k~v k if and only if one vector is a scalar multiple of
the other.
2.19 Suppose that ~
u ~v = ~
u w
~ and ~
u 6= ~0. Must ~v = w?
~
X 2.20 Does any vector have length zero except a zero vector? (If yes, produce an
example. If no, prove it.)
X 2.21 Find the midpoint of the line segment connecting (x1 , y1 ) with (x2 , y2 ) in R2 .
Generalize to Rn .
2.22 Show that if ~v 6= ~0 then ~v /k~v k has length one. What if ~v = ~0?
2.23 Show that if r 0 then r~v is r times as long as ~v . What if r < 0?
X 2.24 A vector ~v Rn of length one is a unit vector. Show that the dot product
of two unit vectors has absolute value less than or equal to one. Can less than
happen? Can equal to ?
44
X 2.38 The geometric mean of two positive reals x, y is xy. It is analogous to the
arithmetic mean (x + y)/2. Use the Cauchy-Schwartz inequality to show that the
geometric mean of any x, y R is less than or equal to the arithmetic mean.
2.39 [Am. Math. Mon., Feb. 1933] A ship is sailing with speed and direction ~v1 ;
the wind blows apparently (judging by the vane on the mast) in the direction of
a vector ~a; on changing the direction and speed of the ship from ~v1 to ~v2 the
apparent wind is in the direction of a vector ~b.
Find the vector velocity of the wind.
2.40 Verify the Cauchy-Schwartz inequality by first proving Lagranges identity:
!2
aj b j
1jn
1jn
a2j
1jn
b2j
(ak bj aj bk )2
1k<jn
and then noting that the final term is positive. (Recall the meaning
a j b j = a 1 b 1 + a 2 b 2 + + an b n
1jn
and
a j 2 = a 1 2 + a 2 2 + + an 2
1jn
1.III
45
2 2
4 3
Gauss method could derive any of these echelon form matrices.
2 2
1 1
2 0
0 1
0 1
0 1
The first results from 21 + 2 . The second comes from following (1/2)1 with
41 + 2 . The third comes from 21 + 2 followed by 22 + 1 (after the first
pivot the matrix is already in echelon form so the second one is extra work but
it is nonetheless a legal row operation).
The fact that the echelon form outcome of Gauss method is not unique
leaves us with some questions. Will any two echelon form versions of a system
have the same number of free variables? Will they in fact have exactly the same
variables free? In this section we will answer both questions yes. We will
do more than answer the questions. We will give a way to decide if one linear
system can be derived from another by row operations. The answers to the two
questions will follow from this larger result.
1.III.1
Gauss-Jordan Reduction
1
1 +3
0
0
1 2 2
1 1 2
2 +3
7 0 1 3
1
3
1
1 1
0 0 4
2
7
8
46
We can keep going to a second stage
1
(1/4)3
0
0
1 2 2
7
1 3
2
0 1
1 0 0
1 1 0 2
33 +2
2 +1
0 1 0 1 0 1 0
23 +1
0 0 1 2
0 0 1
1
1
2
21 +2
2 1
7
2 1 7
4 2 6
0 4 8
(1/2)1
1 1/2 7/2
2
0
1
(1/4)2
(1/2)2 +1
1 0 5/2
2
0 1
The answer is x = 5/2 and y = 2.
This extension of Gauss method is Gauss-Jordan reduction. It goes past
echelon form to a more refined, more specialized, matrix form.
1.3 Definition A matrix is in reduced echelon form if, in addition to being in
echelon form, each leading entry is a one and is the only nonzero entry in its
column.
The disadvantage of using Gauss-Jordan reduction to solve a system is that the
additional row operations mean additional arithmetic. The advantage is that
the solution set can just be read off.
In any echelon form, plain or reduced, we can read off when a system has
an empty solution set because there is a contradictory equation, we can read off
when a system has a one-element solution set because there is no contradiction
and every variable is the leading variable in some row, and we can read off when
a system has an infinite solution set because there is no contradiction and at
least one variable is free.
In reduced echelon form we can read off not just what kind of solution set
the system has, but also its description. Whether or not the echelon form
47
2 6 1 2 5
2 6 1 2 5
2 +3
0 3 1 4 1 0 3 1 4 1
0 3 1 2 5
0 0 0 2 4
1 0 1/2 0 9/2
(1/2)1 (4/3)3 +2 32 +1
3
0 1 1/3 0
3 +1
(1/3)2
2
0
0
0
1
(1/2)3
Starting with the middle matrix, the echelon form version, back substitution
produces 2x4 = 4 so that x4 = 2, then another back substitution gives
3x2 + x3 + 4(2) = 1 implying that x2 = 3 (1/3)x3 , and then the final
back substitution gives 2x1 + 6(3 (1/3)x3 ) + x3 + 2(2) = 5 implying that
x1 = (9/2) + (1/2)x3 . Thus the solution set is this.
9/2
1/2
x1
x2 3 1/3
S = {
x3 = 0 + 1 x3 x3 R}
x4
2
0
Now, considering the final matrix, the reduced echelon form version, note that
adjusting the parametrization by moving the x3 terms to the other side does
indeed give the description of this infinite solution set.
Part of the reason that this works is straightforward. While a set can have
many parametrizations that describe it, e.g., both of these also describe the
above set S (take t to be x3 /6 and s to be x3 1)
9/2
3
4
1/2
3 2
8/3 1/3
{
{
0 + 6 t t R}
1 + 1 s s R}
2
0
2
0
nonetheless we have in this book stuck to a convention of parametrizing using
the unmodified free variables (that is, x3 = x3 instead of x3 = 6t). We can
easily see that a reduced echelon form version of a system is equivalent to a
parametrization in terms of unmodified free variables. For instance,
1
0
2
4
x1 = 4 2x3
0 1 1 3
x2 = 3 x3
0 0 0 0
(to move from left to right we also need to know how many equations are in the
system). So, the convention of parametrizing with the free variables by solving
48
each equation for its leading variable and then eliminating that leading variable
from every other equation is exactly equivalent to the reduced echelon form
conditions that each leading entry must be a one and must be the only nonzero
entry in its column.
Not as straightforward is the other part of the reason that the reduced
echelon form version allows us to read off the parametrization that we would
have gotten had we stopped at echelon form and then done back substitution.
The prior paragraph shows that reduced echelon form corresponds to some
parametrization, but why the same parametrization? A solution set can be
parametrized in many ways, and Gauss method or the Gauss-Jordan method
can be done in many ways, so a first guess might be that we could derive many
different reduced echelon form versions of the same starting system and many
different parametrizations. But we never do. Experience shows that starting
with the same system and proceeding with row operations in many different
ways always yields the same reduced echelon form and the same parametrization
(using the unmodified free variables).
In the rest of this section we will show that the reduced echelon form version
of a matrix is unique. It follows that the parametrization of a linear system in
terms of its unmodified free variables is unique because two different ones would
give two different reduced echelon forms.
We shall use this result, and the ones that lead up to it, in the rest of the
book but perhaps a restatement in a way that makes it seem more immediately
useful may be encouraging. Imagine that we solve a linear system, parametrize,
and check in the back of the book for the answer. But the parametrization there
appears different. Have we made a mistake, or could these be different-looking
descriptions of the same set, as with the three descriptions above of S? The prior
paragraph notes that we will show here that different-looking parametrizations
(using the unmodified free variables) describe genuinely different sets.
Here is an informal argument that the reduced echelon form version of a
matrix is unique. Consider again the example that started this section of a
matrix that reduces to three different echelon form matrices. The first matrix
of the three is the natural echelon form version. The second matrix is the same
as the first except that a row has been halved. The third matrix, too, is just a
cosmetic variant of the first. The definition of reduced echelon form outlaws this
kind of fooling around. In reduced echelon form, halving a row is not possible
because that would change the rows leading entry away from one, and neither
is combining rows possible, because then a leading entry would no longer be
alone in its column.
This informal justification is not a proof; we have argued that no two different
reduced echelon form matrices are related by a single row operation step, but
we have not ruled out the possibility that multiple steps might do. Before we go
to that proof, we finish this subsection by rephrasing our work in a terminology
that will be enlightening.
Many different matrices yield the same reduced echelon form matrix. The
three echelon form matrices from the start of this section, and the matrix they
49
1
0
0
1
We think of these matrices as related to each other. The next result speaks to
this relationship.
1.4 Lemma Elementary row operations are reversible.
Proof. For any matrix A, the effect of swapping rows is reversed by swapping
them back, multiplying a row by a nonzero k is undone by multiplying by 1/k,
and adding a multiple of row i to row j (with i 6= j) is undone by subtracting
the same multiple of row i from row j.
i j j i
ki +j ki +j
ki (1/k)i
A A
A A
QED
2
3
1
0
0
1
2
4
2
0
1
1
0 1
2
0
2
1
The technical phrase in this situation is that matrices that reduce to each other
are equivalent with respect to the relationship of row reducibility. The next
result verifies this statement using the definition of an equivalence.
1.5 Lemma Between matrices, reduces to is an equivalence relation.
Proof. We must check the conditions (i) reflexivity, that any matrix reduces to
itself, (ii) symmetry, that if A reduces to B then B reduces to A, and (iii) transitivity, that if A reduces to B and B reduces to C then A reduces to C.
Reflexivity is easy; any matrix reduces to itself in zero row operations.
That the relationship is symmetric is Lemma 1.4 if A reduces to B by
some row operations then also B reduces to A by reversing those operations.
For transitivity, suppose that A reduces to B and that B reduces to C.
Linking the reduction steps from A B with those from B C
QED
gives a reduction from A to C.
50
1.6 Definition Two matrices that are interreducible by the elementary row
operations are row equivalent.
The diagram below has the collection of all matrices as a box. Inside that
box, each matrix lies in some class. Matrices are in the same class if and only if
they are interreducible. The classes are disjoint no matrix is in two distinct
classes. The collection of matrices has been partitioned into row equivalence
classes. One of the reasons that showing the row equivalence relation is an
equivalence is useful is that any equivalence relation gives rise to a partition.
All matrices:
.A%
...
.
B$
A row equivalent
to B.
One of the classes in this partition is the cluster of matrices shown above,
expanded to include all of the nonsingular 22 matrices.
The next subsection proves that the reduced echelon form of a matrix is
unique; that every matrix reduces to one and only one reduced echelon form
matrix. Rephrased in the relation language, we shall prove that every matrix is
row equivalent to one and only one reduced echelon form matrix. In terms of the
partition in the picture what we shall prove is: every equivalence class contains
one and only one reduced echelon form matrix. So each reduced echelon form
matrix serves as a representative of its class.
After that proof we shall, as mentioned in the introduction to this section,
have a way to decide if one matrix can be derived from another by row reduction.
We can just apply the Gauss-Jordan procedure to both and see whether or not
they come to the same reduced echelon form.
Exercises
X 1.7 Use Gauss-Jordan reduction to solve each system.
(a) x + y = 2
(b) x
z=4
(c) 3x 2y = 1
xy=0
2x + 2y
=1
6x + y = 1/2
(d) 2x y
= 1
x + 3y z = 5
y + 2z = 5
X 1.8 Find the reduced echelon
form of each
! matrix.
!
1
3
1
1 0 3 1 2
2 1
2
0
4
(a)
(b)
(c) 1 4 2 1 5
1 3
1 3 3
3 4 8 1 2
!
0 1 3 2
(d) 0 0 5 6
1 5 1 5
X 1.9 Find each solution set by using Gauss-Jordan reduction, then reading off the
parametrization.
51
z
=1
y + 2z w = 3
x + 2y + 3z w = 7
(b) x
a + 2b + 3c + d e = 1
3a b + c + d + e = 3
1.10 Give two distinct echelon form versions
2 1 1
6 4 1
1 5 1
(c)
y+ z
=0
y
+w=0
3x 2y + 3z + w = 0
y
w=0
(d)
of this matrix.
!
3
2
5
X 1.11 List the reduced echelon forms possible for each size.
(a) 22
(b) 23
(c) 32
(d) 33
X 1.12 What results from applying Gauss-Jordan reduction to a nonsingular matrix?
1.13 The proof of Lemma 1.4 contains a reference to the i 6= j condition on the
row pivoting operation.
(a) The definition of row operations has an i 6= j condition on the swap operation
i j i j
i j . Show that in A
(b) Write down a 22 matrix with nonzero entries, and show that the 11 +1
operation is not reversed by 1 1 + 1 .
(c) Expand the proof of that lemma to make explicit exactly where the i 6= j
condition on pivoting is used.
1.III.2
Row Equivalence
We will close this section and this chapter by proving that every matrix is
row equivalent to one and only one reduced echelon form matrix. The ideas
that appear here will reappear, and be further developed, in the next chapter.
The underlying theme here is that one way to understand a mathematical
situation is by being able to classify the cases that can happen. We have met this
theme several times already. We have classified solution sets of linear systems
into the no-elements, one-element, and infinitely-many elements cases. We have
also classified linear systems with the same number of equations as unknowns
into the nonsingular and singular cases. We adopted these classifications because
they give us a way to understand the situations that we were investigating. Here,
where we are investigating row equivalence, we know that the set of all matrices
breaks into the row equivalence classes. When we finish the proof here, we will
have a way to understand each of those classes its matrices can be thought
of as derived by row operations from the unique reduced echelon form matrix
in that class.
To understand how row operations act to transform one matrix into another,
we consider the effect that they have on the parts of a matrix. The crucial
observation is that row operations combine the rows linearly.
52
QED
In this subsection we will use the convention that, where a matrix is named
with an upper case roman letter, the matching lower-case greek letter names
the rows.
1
1
2
2
B
=
A=
.
.
.
.
m
m
2.3 Corollary Where one matrix row reduces to another, each row of the
second is a linear combination of the rows of the first.
The proof below uses induction on the number of row operations used to
reduce one matrix to the other. Before we proceed, here is an outline of the argument (readers unfamiliar with induction may want to compare this argument
with the one used in the General = Particular + Homogeneous proof). First,
for the base step of the argument, we will verify that the proposition is true
when reduction can be done in zero row operations. Second, for the inductive
step, we will argue that if being able to reduce the first matrix to the second
in some number t 0 of operations implies that each row of the second is a
linear combination of the rows of the first, then being able to reduce the first to
the second in t + 1 operations implies the same thing. Together, this base step
and induction step prove this result because by the base step the proposition
53
is true in the zero operations case, and by the inductive step the fact that it is
true in the zero operations case implies that it is true in the one operation case,
and the inductive step applied again gives that it is therefore true in the two
operations case, etc.
Proof. We proceed by induction on the minimum number of row operations
that take a first matrix A to a second one B.
In the base step, that zero reduction operations suffice, the two matrices
~i =
are equal and each row of B is obviously a combination of As rows:
~i + + 0
~ m.
0
~1 + + 1
For the inductive step, assume the inductive hypothesis: with t 0, if a
matrix can be derived from A in t or fewer operations then its rows are linear
combinations of the As rows. Consider a B that takes t+1 operations. Because
there are more than zero operations, there must be a next-to-last matrix G so
that A G B. This G is only t operations away from A and so the
inductive hypothesis applies to it, that is, each row of G is a linear combination
of the rows of A.
If the last operation, the one from G to B, is a row swap then the rows
of B are just the rows of G reordered and thus each row of B is also a linear
combination of the rows of A. The other two possibilities for this last operation,
that it multiplies a row by a scalar and that it adds a multiple of one row to
another, both result in the rows of B being linear combinations of the rows of
G. But therefore, by the Linear Combination Lemma, each row of B is a linear
combination of the rows of A.
With that, we have both the base step and the inductive step, and so the
QED
proposition follows.
0 2 1 2 1 1 (1/2)2 1 1 2 +1 1
0 2
0 1
0
1 1
0
,
1
call the matrices A, D, G, and B. The methods of the proof show that there
are three sets of linear relationships.
1 = 0 1 + 1 2
2 = 1 1 + 0 2
1 = 0 1 + 1 2
2 = (1/2)1 + 0 2
1 = (1/2)1 + 1 2
2 = (1/2)1 + 0 2
The prior result gives us the insight that Gauss method works by taking
linear combinations of the rows. But to what end; why do we go to echelon
form as a particularly simple, or basic, version of a linear system? The answer,
of course, is that echelon form is suitable for back substitution, because we have
isolated the variables. For instance, in this matrix
2 3 7 8 0 0
0 0 1 5 1 1
R=
0 0 0 3 3 0
0 0 0 0 2 1
54
x1 has been removed from x5 s equation. That is, Gauss method has made x5 s
row independent of x1 s row.
Independence of a collection of row vectors, or of any kind of vectors, will
be precisely defined and explored in the next chapter. But a first take on it is
that we can show that, say, the third row above is not comprised of the other
rows, that 3 6= c1 1 + c2 2 + c4 4 . For, suppose that there are scalars c1 , c2 ,
and c4 such that this relationship holds.
0 0 0 3 3 0 = c1 2 3 7 8 0 0
+ c2 0 0 1 5 1 1
+ c4 0 0 0 0 2 1
The first rows leading entry is in the first column and narrowing our consideration of the above relationship to consideration only of the entries from the first
column 0 = 2c1 +0c2 +0c4 gives that c1 = 0. The second rows leading entry is in
the third column and the equation of entries in that column 0 = 7c1 + 1c2 + 0c4 ,
along with the knowledge that c1 = 0, gives that c2 = 0. Now, to finish, the
third rows leading entry is in the fourth column and the equation of entries
in that column 3 = 8c1 + 5c2 + 0c4 , along with c1 = 0 and c2 = 0, gives an
impossibility.
The following result shows that this effect always holds. It shows that what
Gauss linear elimination method eliminates is linear relationships among the
rows.
2.5 Lemma In an echelon form matrix, no nonzero row is a linear combination
of the other rows.
Proof. Let R be in echelon form. Suppose, to obtain a contradiction, that
some nonzero row is a linear combination of the others.
55
The inductive step is to show that for each row index k between 1 and i 2,
if the coefficient c1 and the coefficients c2 , . . . , ck are all zero then ck+1 is also
zero. That argument, and the contradiction that finishes this proof, is saved for
QED
Exercise 21.
We can now prove that each matrix is row equivalent to one and only one
reduced echelon form matrix. We will find it convenient to break the first half
of the argument off as a preliminary lemma. For one thing, it holds for any
echelon form whatever, not just reduced echelon form.
2.6 Lemma If two echelon form matrices are row equivalent then the leading
entries in their first rows lie in the same column. The same is true of all the
nonzero rows the leading entries in their second rows lie in the same column,
etc.
For the proof we rephrase the result in more technical terms. Define the form
of an mn matrix to be the sequence h`1 , `2 , . . . , `m i where `i is the column
number of the leading entry in row i and `i = if there is no leading entry
in that column. The lemma says that if two echelon form matrices are row
equivalent then their forms are equal sequences.
Proof. Let B and D be echelon form matrices that are row equivalent. Because
they are row equivalent they must be the same size, say nm. Let the column
number of the leading entry in row i of B be `i and let the column number of
the leading entry in row j of D be kj . We will show that `1 = k1 , that `2 = k2 ,
etc., by induction.
This induction argument relies on the fact that the matrices are row equivalent, because the Linear Combination Lemma and its corollary therefore give
that each row of B is a linear combination of the rows of D and vice versa:
b1,`1
= s1,1 0 d1,k1
+ s1,2 0 0
..
.
+ s1,m 0 0
56
First, note that `1 < k1 is impossible: in the columns of D to the left of column
k1 the entries are are all zeroes (as d1,k1 leads the first row) and so if `1 < k1
then the equation of entries from column `1 would be b1,`1 = s1,1 0+ +s1,m 0,
but b1,`1 isnt zero since it leads its row and so this is an impossibility. Next,
a symmetric argument shows that k1 < `1 also is impossible. Thus the `1 = k1
base case holds.
The inductive step is to show that if `1 = k1 , and `2 = k2 , . . . , and `r = kr ,
then also `r+1 = kr+1 (for r in the interval 1 .. m 1). This argument is saved
QED
for Exercise 22.
That lemma answers two of the questions that we have posed (i) any two
echelon form versions of a matrix have the same free variables, and consequently
(ii) any two echelon form versions have the same number of free variables. There
is no linear system and no combination of row operations such that, say, we could
solve the system one way and get y and z free but solve it another way and get
y and w free, or solve it one way and get two free variables while solving it
another way yields three.
We finish now by specializing to the case of reduced echelon form matrices.
2.7 Theorem Each matrix is row equivalent to a unique reduced echelon form
matrix.
Proof. Clearly any matrix is row equivalent to at least one reduced echelon
form matrix, via Gauss-Jordan reduction. For the other half, that any matrix
is equivalent to at most one reduced echelon form matrix, we will show that if
a matrix Gauss-Jordan reduces to each of two others then those two are equal.
Suppose that a matrix is row equivalent the two reduced echelon form matrices B and D, which are therefore row equivalent to each other. The Linear
Combination Lemma and its corollary allow us to write the rows of one, say
B, as a linear combination of the rows of the other i = ci,1 1 + + ci,m m .
The preliminary result, Lemma 2.6, says that in the two matrices, the same
collection of rows are nonzero. Thus, if 1 through r are the nonzero rows of
B then the nonzero rows of D are 1 through r . Zero rows dont contribute to
the sum so we can rewrite the relationship to include just the nonzero rows.
i = ci,1 1 + + ci,r r
()
The preliminary result also says that for each row j between 1 and r, the
leading entries of the j-th row of B and D appear in the same column, denoted
`j . Rewriting the above relationship to focus on the entries in the `j -th column
d1,`j
bi,`j
= ci,1
d2,`j
+ ci,2
..
.
dr,`j
+ ci,r
57
1 3)
. (%
27
...
1
3
.(0 1)
$
All matrices:
each class
consists of
row equivalent
matrices
(There are infinitely many matrices in the pictured class, but weve only got
room to show two.) We have proved there is one and only one reduced echelon
form matrix in each row equivalence class. So the reduced echelon form is a
canonical form for row equivalence: the reduced echelon form matrices are
representatives of the classes.
All matrices:
?
%
?
...
1
0
( 0$
1)
?!
one reduced
echelon form matrix
from each class
We can answer questions about the classes by translating them into questions
about the representatives.
58
2.8 Example We can decide if matrices are interreducible by seeing if GaussJordan reduction produces the same reduced echelon form result. Thus, these
are not row equivalent
1 3
1 3
2 6
2 5
because their reduced echelon forms are not equal.
1 3
1 0
0 0
0 1
2.9 Example Any nonsingular 33
1
0
0
0 0
1 0
0 1
2.10 Example We can describe the classes by listing all possible reduced echelon form matrices. Any 22 matrix lies in one of these: the class of matrices
row equivalent to this,
0 0
0 0
the infinitely many classes of matrices row equivalent to one of this type
1 a
0 0
where a R (including a = 0), the class of matrices row equivalent to this,
0 1
0 0
and the class of matrices row equivalent to this
1 0
0 1
(this the class of nonsingular 22 matrices).
Exercises
X 2.11 Decide if the matrices are row equivalent.
!
!
1
0
2
1 0 2
1 2
0 1
(a)
,
(b) 3 1 1 , 0 2 10
4 8
1 2
5 1 5
2 0 4
2 1 1
1 0 2
1
1 1
0 3 1
0 ,
(c) 1 1
(d)
,
0 2 10
1 2 2
2 2
5
4 3 1
1 1 1
0
1
2
(e)
,
0 0 3
1 1 1
2.12 Describe the matrices in each of the classes represented in Example 2.10.
2.13 Describe all matrices in the row equivalence class of these.
59
1 0
1 2
1 1
(b)
(c)
0 0
2 4
1 3
2.14 How many row equivalence classes are there?
2.15 Can row equivalence classes contain different-sized matrices?
2.16 How big are the row equivalence classes?
(a) Show that the class of any zero matrix is finite.
(b) Do any other classes contain only finitely many members?
2.17 Give two reduced echelon form matrices that have their leading entries in the
same columns, but that are not row equivalent.
2.18 Show that any two n n nonsingular matrices are row equivalent. Are any
two singular matrices row equivalent?
2.19 Describe all of the row equivalence classes containing these.
(a) 2 2 matrices
(b) 2 3 matrices
(c) 3 2 matrices
(d) 33 matrices
~0 is a linear combination of members of the set
2.20 (a) Show that a vector
~1 , . . . ,
~n } if and only there is a linear relationship ~0 = c0
~0 + + cn
~n
{
~0 = ~0 case.)
where c0 is not zero. (Watch out for the
(b) Derive Lemma 2.5.
2.21 Finish the proof of Lemma 2.5.
(a) First illustrate the inductive step by showing that `2 = k2 .
(b) Do the full inductive step: assume that ck is zero for 1 k < i 1, and
deduce that ck+1 is also zero.
(c) Find the contradiction.
2.22 Finish the induction argument in Lemma 2.6.
(a) State the inductive hypothesis, Also state what must be shown to follow from
that hypothesis.
(b) Check that the inductive hypothesis implies that in the relationship r+1 =
sr+1,1 1 + sr+2,2 2 + + sr+1,m m the coefficients sr+1,1 , . . . , sr+1,r are each
zero.
(c) Finish the inductive step by arguing, as in the base case, that `r+1 < kr+1
and kr+1 < `r+1 are impossible.
2.23 Why, in the proof of Theorem 2.7, do we bother to restrict to the nonzero rows?
Why not just stick to the relationship that we began with, i = ci,1 1 + +ci,m m ,
with m instead of r, and argue using it that the only nonzero coefficient is ci,i ,
which is 1?
2.24 [Trono] Three truck drivers went into a roadside cafe. One truck driver purchased four sandwiches, a cup of coffee, and seven doughnuts for $8.45. Another
driver purchased three sandwiches, a cup of coffee, and seven doughnuts for $6.30.
What did the third truck driver pay for a sandwich, a cup of coffee, and a doughnut?
2.25 The fact that Gaussian reduction disallows multiplication of a row by zero is
needed for the proof of uniqueness of reduced echelon form, or else every matrix
would be row equivalent to a matrix of all zeros. Where is it used?
2.26 The Linear Combination Lemma says which equations can be gotten from
Gaussian reduction from a given linear system.
(a)
X
X
X
60
!
1 2 3
3 0 3
1 4 5
the first and second columns add to the third.
(a) Show that remains true under any row operation.
(b) Make a conjecture.
(c) Prove that it holds.
61
i1
i2
= 0
i5
= 0
i4 + i5
= 0
i6 = 0
i3 + i4
+ 10i3
= 10
5i1
+ 4i4
= 10
2i2
+ 50i5
= 0
5i1 2i2
i3
It can be done by hand, but it would take a while and be error-prone. Using a
computer is better.
We illustrate by solving that system under Maple (for another system, a
users manual would obviously detail the exact syntax needed). The array of
coefficients can be entered in this way
> A:=array( [[1,-1,-1,0,0,0,0],
[0,1,0,-1,0,-1,0],
[0,0,1,0,-1,1,0],
[0,0,0,1,1,0,-1],
[0,5,0,10,0,0,0],
[0,0,2,0,4,0,0],
[0,5,-1,0,0,10,0]] );
(putting the rows on separate lines is not necessary, but is done for clarity).
The vector of constants is entered similarly.
> u:=array( [0,0,0,0,10,10,0] );
Systems with infinitely many solutions are solved in the same way the computer simply returns a parametrization.
Exercises
1 Use the computer to solve the two problems that opened this chapter.
(a) This is the Statics problem.
40h + 15c = 100
25c = 50 + 50h
62
63
value of
steel
value of
auto
used by
steel
used by
auto
used by
others
5 395
2 664
25 448
48
9 030
30 346
total
For instance, the dollar value of steel used by the auto industry in this year is
2, 664 million. Note that industries may consume some of their own output.
We can fill in the blanks for the external demand. This years value of the
steel used by others this year is 17, 389 and this years value of the auto used
by others is 21, 268. With that, we have a complete description of the external
demands and of how auto and steel interact, this year, to meet them.
Now, imagine that the external demand for steel has recently been going up
by 200 per year and so we estimate that next year it will be 17, 589. Imagine
also that for similar reasons we estimate that next years external demand for
autos will be down 25 to 21, 243. We wish to predict next years total outputs.
That prediction isnt as simple as adding 200 to this years steel total and
subtracting 25 from this years auto total. For one thing, a rise in steel will
cause that industry to have an increased demand for autos, which will mitigate,
to some extent, the loss in external demand for autos. On the other hand, the
drop in external demand for autos will cause the auto industry to use less steel,
and so lessen somewhat the upswing in steels business. In short, these two
industries form a system, and we need to predict the totals at which the system
as a whole will settle.
64
For that prediction, let s be next years total production of steel and let a be
next years total output of autos. We form these equations.
next years production of steel = next years use of steel by steel
+ next years use of steel by auto
+ next years use of steel by others
next years production of autos = next years use of autos by steel
+ next years use of autos by auto
+ next years use of autos by others
On the left side of those equations go the unknowns s and a. At the ends of the
right sides go our external demand estimates for next year 17, 589 and 21, 243.
For the remaining four terms, we look to the table of this years information
about how the industries interact.
For instance, for next years use of steel by steel, we note that this year the
steel industry used 5395 units of steel input to produce 25, 448 units of steel
output. So next year, when the steel industry will produce s units out, we
expect that doing so will take s (5395)/(25 448) units of steel input this is
simply the assumption that input is proportional to output. (We are assuming
that the ratio of input to output remains constant over time; in practice, models
may try to take account of trends of change in the ratios.)
Next years use of steel by the auto industry is similar. This year, the auto
industry uses 2664 units of steel input to produce 30346 units of auto output. So
next year, when the auto industrys total output is a, we expect it to consume
a (2664)/(30346) units of steel.
Filling in the other equation in the same way, we get this system of linear
equation.
5 395
s+
25 448
48
s+
25 448
2 664
a + 17 589 = s
30 346
9 030
a + 21 243 = a
30 346
Rounding to four decimal places and putting it into the form for Gauss method
gives this.
0.7880s 0.0879a = 17 589
0.0019s + 0.7024a = 21 268
The solution is s = 25 708 and a = 30 350.
Looking back, recall that above we described why the prediction of next
years totals isnt as simple as adding 200 to last years steel total and subtracting 25 from last years auto total. In fact, comparing these totals for next year
to the ones given at the start for the current year shows that, despite the drop
in external demand, the total production of the auto industry is predicted to
rise. The increase in internal demand for autos caused by steels sharp rise in
business more than makes up for the loss in external demand for autos.
65
66
67
While this code provides a first take on how Gauss method can be mechanized,
it is not ready to use. It is naive in many ways. The most glaring way is that
it assumes that a nonzero number is always found in the pivot row, pivot row
position for use as the pivot entry. To make it practical, one way in which this
code needs to be reworked is to cover the case where finding a zero in that
location leads to a row swap, or to the conclusion that the matrix is singular.
Adding some if statements to cover those cases is not hard, but we
wont pursue that here. Instead, we will consider some more subtle ways in
which the code is naive. There are pitfalls arising from the computers reliance
on finite-precision floating point arithmetic.
For example, we have seen above that we must handle as a separate case a
system that is singular. But systems that are nearly singular also require great
care. Consider this one.
x + 2y = 3
1.000 000 01x + 2y = 3.000 000 01
By eye we get the solution x = 1 and y = 1. But a computer has more trouble. A
computer that represents real numbers to eight significant places (as is common,
usually called single precision) will represent the second equation internally as
1.000 000 0x + 2y = 3.000 000 0, losing the digits in the ninth place. Instead of
reporting the correct solution, this computer will report something that is not
even close this computer thinks that the system is singular because the two
equations are represented internally as equal.
For some intuition about how the computer could think something that is
so far off, we can graph the system.
68
4
3
2
(1,1)
1
0
-1
-1
At the scale of this graph, the two lines are hard to resolve apart. This system
is nearly singular in the sense that the two lines are nearly the same line. Nearsingularity gives this system the property that a small change in the system
can cause a large change in its solution; for instance, changing the 3.000 000 01
to 3.000 000 03 changes the intersection point from (1, 1) to (3, 0). This system
changes radically depending on a ninth digit, which explains why the eightplace computer is stumped. A problem that is very sensitive to inaccuracy or
uncertainties in the input values is ill-conditioned.
The above example gives one way in which a system can be difficult to solve
on a computer. It has the advantage that the picture of nearly-equal lines
gives a memorable insight into one way that numerical difficulties can arise.
Unfortunately, though, this insight isnt very useful when we wish to solve some
large system. We cannot, typically, hope to understand the geometry of an
arbitrary large system. And, in addition, the reasons that the computers results
may be unreliable are more complicated than only that the angle between some
of the linear surfaces is quite small.
For an example, consider the system below, from [Hamming].
0.001x + y = 1
xy=0
The second equation gives x = y, so x = y = 1/1.001 and thus both variables
have values that are just less than 1. A computer using two digits represents
the system internally in this way (we will do this example in two-digit floating
point arithmetic, but a similar one with eight digits is easy to invent).
(1.0 102 )x + (1.0 100 )y = 1.0 100
(1.0 100 )x (1.0 100 )y = 0.0 100
The computers row reduction step 10001 + 2 produces a second equation
1001y = 999, which the computer rounds to two places as (1.0 103 )y =
1.0 103 . Then the computer decides from the second equation that y = 1
and from the first equation that x = 0. This y value is fairly good, but the x
is way off. Thus, another cause of unreliable output is the mixture of floating
point arithmetic and a reliance on pivots that are small.
69
An experienced programmer may respond that we should go to double precision where, usually, sixteen significant digits are retained. It is true, this will
solve many problems. However, there are some difficulties with it as a general
approach. For one thing, double precision takes longer than single precision (on
a 486 chip, multiplication takes eleven ticks in single precision but fourteen in
double precision [Programmers Ref.]) and has twice the memory requirements.
So attempting to do all calculations in double precision is just not practical. And
besides, the above systems can obviously be tweaked to give the same trouble in
the seventeenth digit, so double precision wont fix all problems. What we need
is a strategy to minimize the numerical trouble arising from solving systems
on a computer, and some guidance as to how far the reported solutions can be
trusted.
Mathematicians have made a careful study of how to get the most reliable
results. A basic improvement on the naive code above is to not simply take
the entry in the pivot row , pivot row position for the pivot, but rather to look
at all of the entries in the pivot row column below the pivot row row, and take
the one that is most likely to give reliable results (e.g., take one that is not too
small). This strategy is partial pivoting. For example, to solve the troublesome
system () above, we start by looking at both equations for a best first pivot,
and taking the 1 in the second equation as more likely to give good results.
Then, the pivot step of .0012 + 1 gives a first equation of 1.001y = 1, which
the computer will represent as (1.0100 )y = 1.0100 , leading to the conclusion
that y = 1 and, after back-substitution, x = 1, both of which are close to right.
The code from above can be adapted to this purpose.
for(pivot_row=1;pivot_row<=n-1;pivot_row++){
/* find the largest pivot in this column (in row max) */
max=pivot_row;
for(row_below=pivot_row+1;pivot_row<=n;row_below++){
if (abs(a[row_below,pivot_row]) > abs(a[max,row_below]))
max=row_below;
}
/* swap rows to move that pivot entry up */
for(col=pivot_row;col<=n;col++){
temp=a[pivot_row,col];
a[pivot_row,col]=a[max,col];
a[max,col]=temp;
}
/* proceed as before */
for(row_below=pivot_row+1;row_below<=n;row_below++){
multiplier=a[row_below,pivot_row]/a[pivot_row,pivot_row];
for(col=pivot_row;col<=n;col++){
a[row_below,col]-=multiplier*a[pivot_row,col];
}
}
}
A full analysis of the best way to implement Gauss method is outside the
scope of the book (see [Wilkinson 1965]), but the method recommended by most
70
experts is a variation on the code above that first finds the best pivot among
the candidates, and then scales it to a number that is less likely to give trouble.
This is scaled partial pivoting.
In addition to returning a result that is likely to be reliable, most welldone code will return a number, called the conditioning number of the matrix,
that describes the factor by which uncertainties in the input numbers could be
magnified to become possible inaccuracies in the results returned (see [Rice]).
The lesson of this discussion is that just because Gauss method always works
in theory, and just because computer code correctly implements that method,
and just because the answer appears on green-bar paper, doesnt mean that the
answer is reliable. In practice, always use a package where experts have worked
hard to counter what can go wrong.
Exercises
1 Using two decimal places, add 253 and 2/3.
2 This intersect-the-lines problem contrasts with the example discussed above.
4
3
2
x + 2y = 3
3x 2y = 1
(1,1)
1
0
-1
-1
Illustrate that, in the resulting system, some small change in the numbers will
produce only a small change in the solution by changing the constant in the bottom equation to 1.008 and solving. Compare it to the solution of the unchanged
system.
3 Solve this system by hand ([Rice]).
0.000 3x + 1.556y = 1.569
0.345 4x 2.346y = 1.018
(a) Solve it accurately, by hand.
(b) Solve it by rounding at each step to
four significant digits.
4 Rounding inside the computer often has an effect on the result. Assume that
your machine has eight significant digits.
(a) Show that the machine will compute (2/3) + ((2/3) (1/3)) as unequal to
((2/3) + (2/3)) (1/3). Thus, computer arithmetic is not associative.
(b) Compare the computers version of (1/3)x + y = 0 and (2/3)x + 2y = 0. Is
twice the first equation the same as the second?
5 Ill-conditioning is not only dependent on the matrix of coefficients. This example
[Hamming] shows that it can arise from an interaction between the left and right
sides of the system. Let be a small real.
3x + 2y + z =
6
2x + 2y + 2z = 2 + 4
x + 2y z = 1 +
71
(a) Solve the system by hand. Notice that the s divide out only because there
is an exact cancelation of the integer parts on the right side as well as on the
left.
(b) Solve the system by hand, rounding to two decimal places, and with =
0.001.
72
To read it, we can think of the electricity as coming out of one end of the battery
(labeled 6V OR 12V), flowing through the wires (drawn as straight lines to make
the diagram more readable), and back into the other end of the battery. If, in
making its way from one end of the battery to the other through the network of
wires, some electricity flows through a light bulb (drawn as a circle enclosing a
loop of wire), then that light lights. For instance, when the driver steps on the
brake at point A then the switch makes contact and electricity flows through
the brake lights at point B.
This network of connections and components is complicated enough that to
analyze it for instance, to find out how much electricity is used when both
the headlights and the brake lights are on then we need systematic tools.
One such tool is linear systems. To illustrate this application, we first need a
few facts about electricity and networks.
The two facts that we need about electricity concern how the electrical components act. First, the battery is like a pump for electricity; it provides a force
or push so that the electricity will flow, if there is at least one available path for
it. The second fact about the components is the observation that (in the materials commonly used in components) the amount of current flow is proportional
to the force pushing it. For each electrical component there is a constant of
proportionality, called its resistance, satisfying that potential = flow resistance.
(The units are: potential to flow is described in volts, the rate of flow itself is
given in amperes, and resistance to the flow is in ohms. These units are set up
so that volts = amperes ohms.)
For example, suppose a bulb has a resistance of 25 ohms. Wiring its ends
to a battery with 12 volts results in a flow of electrical current of 12/25 =
0.48 amperes. Conversely, with that same bulb, if we have flow of electrical
current of 2 amperes through it, then the potential difference between one end
73
of the bulb and the other end will be 2 25 = 50 volts. This is the voltage drop
across this bulb. One way to think of the above circuit is that the battery is a
voltage source, or rise, and the other components are voltage sinks, or drops,
that use up the force provided by the battery.
The two facts that we need about networks are Kirchhoffs Laws.
First Law. The flow into any spot equals the flow out.
Second Law. Around a circuit the total drop equals the total rise.
(In the above circuit the only voltage rise is at the one battery, but some circuits
have more than one rise.)
We can use these facts for a simple analysis of the circuit shown below.
There are three components; they might be bulbs, or they might be some other
component that resists the flow of electricity (resistors are drawn as zig-zags ).
When components are wired one after another, as these are, they are said to be
in series.
20 volt
potential
2 ohm
resistance
3 ohm
resistance
5 ohm
resistance
By Kirchhoffs Second Law, because the voltage rise in this circuit is 20 volts, so
too, the total voltage drop around this circuit is 20 volts. Since the resistance in
total, from start to finish, in this circuit is 10 ohms (we can take the resistance
of a wire to be negligible), we get that the current is (20/10) = 2 amperes. Now,
Kirchhoffs First Law says that there are 2 amperes through each resistor, and
so the voltage drops are 4 volts, 10 volts, and 6 volts.
Linear systems appear in the analysis of the next network. In this one, the
resistors are not in series. They are instead in parallel. This network is more
like the cars lighting diagram.
20 volts
12 ohm
8 ohm
74
We begin by labeling the branches of the network. Call the flow of current
coming out of the top of the battery and through the top wire i0 , call the
current through the left branch of the parallel portion i1 , that through the right
branch i2 , and call the current flowing through the bottom wire and into the
bottom of the battery i3 . (Remark: in labeling, we dont have to know the
actual direction of flow. We arbitrarily choose a direction to establish a sign
convention for the equations.)
i0
i1
i2
i3
The fact that i0 splits into i1 and i2 , on application of Kirchhoffs First Law,
gives that i1 + i2 = i0 . Similarly, we have that i1 + i2 = i3 . In the circuit that
loops out of the top of the battery, down the left branch of the parallel portion,
and back into the bottom of the battery, the voltage rise is 20 and the voltage
drop is i1 12, so Kirchoffs Second Law gives that 12i1 = 20. In the circuit from
the battery to the right branch and back to the battery there is a voltage rise of
20 and a voltage drop of i2 8, so Kirchoffs Second law gives that 8i2 = 20. And
finally, in the circuit that just loops around in the left and right branches of the
parallel portion (taken clockwise), there is a voltage rise of 0 and a voltage drop
of 8i2 12i1 so Kirchoffs Second Law gives 8i2 12i1 = 0.
All of these equations taken together make this system.
i0
i1 i2
= 0
i1 i2 + i3 = 0
= 20
12i1
= 20
8i2
= 0
12i1 + 8i2
75
2 ohm
5 ohm
50 ohm
10 volts
10 ohm
4 ohm
i2
i5
i3
i4
i6
Kirchoffs First Law, applied to the top node, the left node, the right node, and
the bottom node gives these equations.
i0 = i1 + i2
i1 = i3 + i5
i2 + i5 = i4
i3 + i4 = i6
Kirchhoffs Second Law, applied to the inside loop (i0 -i1 -i3 -i6 ), the outside loop,
and the upper loop not involving the battery, gives these equations.
5i1 + 10i3 = 10
2i2 + 4i4 = 10
5i1 + 50i5 2i2 = 0
We could get more equations, but these are enough to produce a solution: i0 =
7/3, i1 = 2/3, i2 = 5/3, i3 = 2/3, i4 = 5/3, i5 = 0, and i6 = 7/3.
Networks of other kinds, not just electrical ones, can also be analyzed in this
way. For instance, a network of streets in given in the exercises.
Exercises
Hint: Most of the linear systems are large enough that they are best solved on a
computer.
76
3 ohm
9 volt
2 ohm
2 ohm
(b) Compare this one with the parallel case discussed above.
3 ohm
9 volt
2 ohm
2 ohm
2 ohm
3 ohm
3 ohm
9 volt
3 ohm
2 ohm
2 ohm
4 ohm
2 ohm
Pier Bvd
North
100
75
Pier
150
150
Main
25
50
77
We can use Kirchhoffs Law, that the flow into any intersection equals the flow
out, to establish some equations modeling how traffic flows work here.
(a) Label each of the three arcs of road in the circle with a variable. For each of
the three in-out intersections, get an equation describing the traffic flow at that
node.
(b) Solve that system.
3 This is a map of a network of streets. Below we will describe the flow of cars
into, and out of, this network.
west
Jay Ln
Willow
Shelburne St
east
Winooski Ave
The hourly flow of cars into this networks entrances, and out of its exits can be
observed.
into
out of
east Winooski
100
125
west Winooski
150
150
Willow
25
50
Jay
25
Shelburne
200
125
(The total in must approximately equal the total out over a long period of time.)
Once inside the network, the traffic may proceed in different ways, perhaps
filling Willow and leaving Jay mostly empty, or perhaps flowing in some other
way. We can use Kirchhoffs Law that the flow into any intersection equals the
flow out.
(a) Determine the restrictions on the flow inside this network of streets by setting
up a variable for each block, establishing the equations, and solving them. Notice
that some streets are one-way only. (Hint: this will not yield a unique solution,
since traffic can flow through this network in various ways. You should get at
least one free variable.)
(b) Suppose some construction is proposed for Winooski Avenue East between
Willow and Jay, so traffic on that block will be reduced. What is the least
amount of traffic flow that can be allowed on that block without disrupting the
hourly flow into and out of the network?
4 Calculate the amperages in this network with more than one voltage rise.
5 ohm
3 ohm
1.5 volt
2 ohm
3 volt
6 ohm
10 ohm
5 In the circuit with the 8 ohm and 12 ohm resistors in parallel, the electric current
away from and back to the battery was found to be 25/6 amperes. Thus, the
78
Chapter 2
Vector Spaces
The first chapter began by introducing Gauss method and finished with a fair
understanding, keyed on the Linear Combination Lemma, of how it finds the
solution set of a linear system. Gauss method systematically takes linear combinations of the rows. With that insight, we now move to a general study of
linear combinations.
We need a setting for this study. At times in the first chapter, weve combined vectors from R2 , at other times vectors from R3 , and at other times vectors
from even higher-dimensional spaces. Thus, our first impulse might be to work
in Rn , leaving n unspecified. This would have the advantage that any of the
results would hold for R2 and for R3 and for many other spaces, simultaneously.
But, if having the results apply to many spaces at once is advantageous then
sticking only to Rn s is overly restrictive. Wed like the results to also apply to
combinations of row vectors, as in the final section of the first chapter. Weve
even seen some spaces that are not just a collection of all of the same-sized
column vectors or row vectors. For instance, weve seen a solution set of a
homogeneous system that is a plane, inside of R3 . This solution set is a closed
system in the sense that a linear combination of these solutions is also a solution.
But it is not just a collection of all of the three-tall column vectors; only some
of them are in this solution set.
We want the results about linear combinations to apply anywhere that linear
combinations are sensible. We shall call any such set a vector space. Our results,
instead of being phrased as Whenever we have a collection in which we can
sensibly take linear combinations . . . , will be stated as In any vector space
. . . .
Such a statement describes at once what happens in many spaces. The step
up in abstraction from studying a single space at a time to studying a class
of spaces can be hard to make. To understand its advantages, consider this
analogy. Imagine that the government made laws one person at a time: Leslie
Jones cant jay walk. That would be a bad idea; statements have the virtue of
economy when they apply to many cases at once. Or, suppose that they ruled,
Kim Ke must stop when passing the scene of an accident. Contrast that with,
Any doctor must stop when passing the scene of an accident. More general
statements, in some ways, are clearer.
79
80
2.I
We shall study structures with two operations, an addition and a scalar multiplication, that are subject to some simple conditions. We will reflect more on
the conditions later, but on first reading notice how reasonable they are. For
instance, surely any operation that can be called an addition (e.g., column vector addition, row vector addition, or real number addition) will satisfy all the
conditions in (1) below.
2.I.1
1.1 Definition A vector space (over R) consists of a set V along with two
operations + and such that
(1) if ~v , w
~ V then their vector sum ~v + w
~ is in V and
~v + w
~ =w
~ + ~v
(~v + w)
~ + ~u = ~v + (w
~ + ~u) (where ~u V )
there is a zero vector ~0 V such that ~v + ~0 = ~v for all ~v V
each ~v V has an additive inverse w
~ V such that w
~ + ~v = ~0
(2) if r, s are scalars (members of R) and ~v , w
~ V then each scalar multiple
r ~v is in V and
(r + s) ~v = r ~v + s ~v
r (~v + w)
~ = r ~v + r w
~
(rs) ~v = r (s ~v )
1 ~v = ~v .
1.2 Remark Because it involves two kinds of addition and two kinds of multiplication, that definition may seem confused. For instance, in (r + s) ~v =
r ~v + s ~v , the first + is the real number addition operator while the + to
the right of the equals sign represents vector addition in the structure V . These
expressions arent ambiguous because, e.g., r and s are real numbers so r + s
can only mean real number addition.
The best way to go through the examples below is to check all of the conditions in the definition. That check is written out in the first example. Use
it as a model for the others. Especially important are the two: ~v + w
~ is in
V and r ~v is in V . These are the closure conditions. They specify that the
addition and scalar multiplication operations are always sensible they must
be defined for every pair of vectors, and every scalar and vector, and the result
of the operation must be a member of the set (see Example 1.4).
81
1.3 Example The set R2 is a vector space if the operations + and have
their usual meaning.
y
x1 + y1
x1
rx1
x1
+ 1 =
r
=
x2
y2
x2 + y2
x2
rx2
We shall check all of the conditions in the definition.
There are five conditions in item (1). First, for closure of addition, note that
for any v1 , v2 , w1 , w2 R the result of the sum
v1
w1
v + w1
+
= 1
v2
w2
v2 + w2
is a column array with two real entries, and so is in R2 . Second, to show that
addition of vectors commutes, take all entries to be real numbers and compute
w1
v1 + w1
w1 + v1
w1
v
v1
+
=
=
=
+ 1
v2
w2
v2 + w2
w2 + v2
w2
v2
(the second equality follows from the fact that the components of the vectors
are real numbers, and the addition of real numbers is commutative). The third
condition, associativity of vector addition, is similar.
w1
u1
(v1 + w1 ) + u1
v1
+
)+
=
(
v2
w2
u2
(v2 + w2 ) + u2
v1 + (w1 + u1 )
=
v2 + (w2 + u2 )
v1
w
u1
=
+( 1 +
)
v2
w2
u2
For the fourth we must produce a zero element the vector of zeroes is it.
0
v
v1
= 1
+
v2
v2
0
Fifth, to produce an additive inverse, note that for any v1 , v2 R we have
v1
v
0
+ 1 =
v2
v2
0
so the first vector is the desired additive inverse of the second.
The checks for the five conditions in item (2) are just as routine. First, for
closure under scalar multiplication, where r, v1 , v2 R,
rv1
v
r 1 =
v2
rv2
is a column array with two real entries, and so is in R2 . This checks the second
condition.
(r + s)v1
rv1 + sv1
v
v
v
=
=r 1 +s 1
(r + s) 1 =
v2
(r + s)v2
rv2 + sv2
v2
v2
82
For the third condition, that scalar multiplication distributes from the left over
vector addition, the check is also straightforward.
rv1 + rw1
w1
r(v1 + w1 )
v1
w1
v1
=
+
)=
=r
+r
r(
v2
w2
rv2 + rw2
v2
w2
r(v2 + w2 )
The fourth
(rs)
v
(rs)v1
r(sv1 )
v1
= r (s 1 )
=
=
v2
(rs)v2
v2
r(sv2 )
rx
x2
x1 + x2
x
x1
y1 + y2 = y1 + y2
r y = ry
z1
z2
z1 + z2
z
rz
The addition and scalar multiplication operations here are just the ones of R3 ,
reused on its subset P . We say P inherits these operations from R3 . Here is a
typical addition in P .
1
1
0
1 + 0 = 1
2
1
1
This illustrates that P is closed under addition. Weve added two vectors from
P that is, with the property that the sum of their three entries is zero
and weve gotten a vector also in P . Of course, this example of closure is not
a proof of closure. To prove that P is closed under addition, take two elements
of P
x2
x1
y1 , y2
z1
z2
83
x1 + x2
y1 + y2
z1 + z2
is also in P since (x1 +x2 )+(y1 +y2 )+(z1 +z2 ) = (x1 +y1 +z1 )+(x2 +y2 +z2 ) = 0.
To show that P is closed under scalar multiplication, start with a vector from
P
x
y
z
(so that x + y + z = 0), and then for r R observe that the scalar multiple
x
rx
r y = ry
z
rz
satisfies that rx + ry + rz = r(x + y + z) = 0. Thus the two closure conditions
are satisfied. The checks for the other conditions in the definition of a vector
space are just as easy.
1.5 Example Example 1.3 shows that the set of all two-tall vectors with real
entries is a vector space. Example 1.4 gives a subset of an Rn that is also a
vector space. In contrast with those two, consider the set of two-tall columns
with entries that are integers (under the obvious operations). This is a subset
of a vector space, but it is not itself a vector space. The reason is that this set is
not closed under scalar multiplication, that is, it does not satisfy requirement (2)
in the definition. Here is a column with integer entries, and a scalar, such that
the outcome of the operation
4
2
0.5
=
3
1.5
is not a member of the set, since its entries are not all integers.
1.6 Example The singleton set
0
0
{
0}
0
0
0
0 0
r
0 = 0
0
0
84
A vector space must have at least one element, its zero vector. Thus a
one-element vector space is the smallest one possible.
1.7 Definition A one-element vector space is a trivial space.
Warning! The examples so far involve sets of column vectors with the usual
operations. But vector spaces need not be collections of column vectors, or even
of row vectors. Below are some other types of vector spaces. The term vector
space does not mean collection of columns of reals. It means something more
like collection in which any linear combination is sensible.
a0 + a1 x + a2 x + a3 x corresponds to
a2
a3
then the operations also correspond. Here is an example of corresponding additions.
1
2
3
1 2x + 0x2 + 1x3
2 3 1
2
3
+ 2 + 3x + 7x 4x
corresponds to
0 + 7 = 7
3 + 1x + 7x2 3x3
1
4
3
Things we are thinking of as the same add to the same sum. Chapter Three
makes precise this idea of vector space correspondence. For now we shall just
leave it as an intuition.
85
1.9 Example The set {f f : N R} of all real-valued functions of one natural number variable is a vector space under the operations
(f1 + f2 ) (n) = f1 (n) + f2 (n)
(r f ) (n) = r f (n)
so that if, for example, f1 (n) = n2 + 2 sin(n) and f2 (n) = sin(n) + 0.5 then
(f1 + 2f2 ) (n) = n2 + 1.
We can view this space as a generalization of Example 1.3 by thinking of
these functions as the same as infinitely-tall vectors:
n
0
1
2
3
..
.
f (n) = n2 + 1
1
2
5
10
..
.
corresponds to
1
2
5
10
..
.
{a0 + a1 x + + an xn n N and a0 , . . . , an R}
makes a vector space when given the natural +
(a0 + a1 x + + an xn ) + (b0 + b1 x + + bn xn )
= (a0 + b0 ) + (a1 + b1 )x + + (an + bn )xn
and .
r (a0 + a1 x + . . . an xn ) = (ra0 ) + (ra1 )x + . . . (ran )xn
This space differs from the space P3 of Example 1.8. This space contains not just
degree three polynomials, but degree thirty polynomials and degree three hundred polynomials, too. Each individual polynomial of course is of a finite degree,
but the set has no single bound on the degree of all of its members.
This example, like the prior one, can be thought of in terms of infinite-tuples.
For instance, we can think of 1 + 3x + 5x2 as corresponding to (1, 3, 5, 0, 0, . . . ).
However, dont confuse this space with the one from Example 1.9. Each member
of this set has a bounded degree, so under our correspondence there are no
elements from this space matching (1, 2, 5, 10, . . . ). The vectors in this space
correspond to infinite-tuples that end in zeroes.
(r f ) (x) = r f (x)
The difference between this and Example 1.9 is the domain of the functions.
86
(r f ) (x) = r f (x)
wn
both satisfy the condition that their entries add to zero then ~v + w
~ also satisfies
that condition: c1 (v1 + w1 ) + + cn (vn + wn ) = (c1 v1 + + cn vn ) + (c1 w1 +
+ cn wn ) = 0. The checks of the other conditions are just as routine.
87
w
~ + ~v = w
~ + ~v + 0 ~v
~0 = ~0 + 0 ~v
~0 = 0 ~v
The second item is easy: (1 ~v ) + ~v = (1 + 1) ~v = 0 ~v = ~0 shows that
we can write ~v for the additive inverse of ~v without worrying about possible
confusion with (1) ~v .
QED
For the third one, this r ~0 = r (0 ~0) = (r 0) ~0 = ~0 will do.
We finish this subsection with an recap, and a comment.
88
Chapter One studied Gaussian reduction. That lead us here to the study of
collections of linear combinations. We have named any such structure a vector
space. In a phrase, the point of this material is that vector spaces are the right
context in which to study linearity.
Finally, a comment. From the fact that it forms a whole chapter, and especially because that chapter is the first one, a reader could come to think that
the study of linear systems is our purpose. The truth is, we will not so much
use vector spaces in the study of linear systems as we will instead have linear
systems lead us into the study of vector spaces. The wide variety of examples
from this subsection shows that the study of vector spaces is interesting and important in its own right, aside from how it helps us understand linear systems.
Linear systems wont go away. But from now on our primary objects of study
will be vector spaces.
Exercises
1.17 Give the zero vector from each of these vector spaces.
(a) The space of degree three polynomials under the natural operations
(b) The space of 24 matrices
1
3
1
0
(c) In {aex + bex a, b R}, a space of functions of the real variable x under
the natural operations, the vector 3ex 2ex .
X 1.19 Show that each of these is a vector space.
y
L = { R4 x + y z + w = 0}
z
w
under the operations inherited from R4
X 1.20 Show that each set is not a vector space. (Hint. Start by listing two members
of each set.)
(a) Under the operations inherited from R3 , this set
x
{ y
z
R3 x + y + z = 1}
89
{ y R3 x2 + y 2 + z 2 = 1}
z
(c) Under the usual matrix operations,
a
{
b
1
c
a, b, c R}
{a0 + a1 x + a2 x2 a0 , a1 , a2 R+ }
where R+ is the set of reals greater than zero
(e) Under the inherited operations,
x
y
1.21 Define addition and scalar multiplication operations to make the complex
numbers a vector space over R.
X 1.22 Is the set of rational numbers a vector space over R under the usual addition
and scalar multiplication operations?
1.23 Show that the set of linear combinations of the variables x, y, z is a vector
space under the natural addition and scalar multiplication operations.
1.24 Prove that this is not a vector space: the set of two-tall column vectors with
real entries subject to these operations.
x1
y1
x2
y2
x1 x2
y1 y2
x
y
rx
ry
a
0
0
b
a, b R}
{
(c) This set
x
x+y
x+y
y
x, y R}
y
{ R4 x + y + w = 1}
z
w
90
a
{
b
(d) This set of 22 matrices
a
b
0
c
0
c
a, b, c R}
a + b + c = 0}
91
1.39 Is this a vector space under the natural operations: the real-valued functions
of one real variable that are differentiable?
1.40 A vector space over the complex numbers C has the same definition as a vector
space over the reals except that scalars are drawn from C instead of from R. Show
that each of these is a vector space over the complex numbers. (Recall how complex
numbers add and multiply: (a0 + a1 i) + (b0 + b1 i) = (a0 + b0 ) + (a1 + b1 )i and
(a0 + a1 i)(b0 + b1 i) = (a0 b0 a1 b1 ) + (a0 b1 + a1 b0 )i.)
(a) The set of degree two polynomials with complex coefficients
(b) This set
0
{
b
a
0
a, b C and a + b = 0 + 0i}
1.41 Find a property shared by all of the Rn s not listed as a requirement for a
vector space.
X 1.42 (a) Prove that a sum of four vectors ~v1 , . . . , ~v4 V can be associated in
any way without changing the result.
((~v1 + ~v2 ) + ~v3 ) + ~v4 = (~v1 + (~v2 + ~v3 )) + ~v4
= (~v1 + ~v2 ) + (~v3 + ~v4 )
= ~v1 + ((~v2 + ~v3 ) + ~v4 )
= ~v1 + (~v2 + (~v3 + ~v4 ))
This allows us to simply write ~v1 + ~v2 + ~v3 + ~v4 without ambiguity.
(b) Prove that any two ways of associating a sum of any number of vectors give
the same sum. (Hint. Use induction on the number of vectors.)
1.43 For any vector space, a subset that is itself a vector space under the inherited
3
operations (e.g., a plane through the
origin inside of R ) is a subspace.
2
(a) Show that {a0 + a1 x + a2 x a0 + a1 + a2 = 0} is a subspace of the vector
space of degree two polynomials.
(b) Show that this is a subspace of the 22 matrices.
a
{
c
b
0
a + b = 0}
(c) Show that a nonempty subset S of a real vector space is a subspace if and only
if it is closed under linear combinations of pairs of vectors: whenever c1 , c2 R
and ~s1 , ~s2 S then the combination c1~v1 + c2~v2 is in S.
2.I.2
One of the examples that led us to introduce the idea of a vector space
was the solution set of a homogeneous system. For instance, weve seen in
Example 1.4 such a space that is a planar subset of R3 . There, the vector space
R3 contains inside it another vector space, the plane.
2.1 Definition For any vector space, a subspace is a subset that is itself a
vector space, under the inherited operations.
92
x1
x2
x1 + x2
y1 + y2 = y1 + y2
z1
z2
z1 + z2
and scalar multiplication is also the same as it is in R3 . To show that P is a
subspace, we need only note that it is a subset and then verify that it is a space.
Checking that P satisfies the conditions in the definition of a vector space is
routine. For instance, for closure under addition, just note that if the summands
satisfy that x1 + y1 + z1 = 0 and x2 + y2 + z2 = 0 then the sum satisfies that
(x1 + x2 ) + (y1 + y2 ) + (z1 + z2 ) = (x1 + y1 + z1 ) + (x2 + y2 + z2 ) = 0.
2.3 Example The x-axis in R2 is a subspace where the addition and scalar
multiplication operations are the inherited ones.
x2
x1 + x2
x
rx
x1
+
=
r
=
0
0
0
0
0
As above, to verify that this is a subspace, we simply note that it is a subset
and then check that it satisfies the conditions in definition of a vector space.
For instance, the two closure conditions are satisfied: (1) adding two vectors
with a second component of zero results in a vector with a second component
of zero, and (2) multiplying a scalar times a vector with a second component of
zero results in a vector with a second component of zero.
2.4 Example Another subspace of R2 is
0
{
}
0
its trivial subspace.
Any vector space has a trivial subspace {~0 }. At the opposite extreme, any
vector space has itself for a subspace. These two are the improper subspaces.
Other subspaces are proper.
2.5 Example The condition in the definition requiring that the addition and
scalar multiplication operations must be the ones inherited from the larger space
is important. Consider the subset {1} of the vector space R1 . Under the operations 1+1 = 1 and r 1 = 1 that set is a vector space, specifically, a trivial space.
But it is not a subspace of R1 because those arent the inherited operations, since
of course R1 has 1 + 1 = 2.
93
2.6 Example All kinds of vector spaces, not just Rn s, have subspaces. The
3
a, b, c, d R} has a subvector space of cubic polynomials {a + bx + cx2 + dx
(1) (2)
(2) (3)
(3) (1)
94
second. For the fourth, consider the zero vector of V and note that closure of S
under linear combinations of pairs of vectors gives that (where ~s is any member
of the nonempty set S) 0 ~s + 0 ~s = ~0 is in S; showing that ~0 acts under the
inherited operations as the additive identity of S is easy. The fifth condition is
satisfied because for any ~s S, closure under linear combinations shows that
the vector 0 ~0 + (1) ~s is in S; showing that it is the additive inverse of ~s
under the inherited operations is routine.
QED
The checks for item (2) are similar and are saved for Exercise 32.
We usually show that a subset is a subspace with (2) = (1).
2.10 Remark At the start of this chapter we introduced vector spaces as collections in which linear combinations are sensible. The above result speaks
to this.
The vector space definition has ten conditions but eight of them, the ones
stated there with the bullets, simply ensure that referring to the operations
as an addition and a scalar multiplication is sensible. The proof above checks
that if the nonempty set S satisfies statement (2) then inheritance of the operations from the surrounding vector space brings with it the inheritance of these
eight properties also (i.e., commutativity of addition in S follows right from
commutativity of addition in V ). So, in this context, this meaning of sensible
is automatically satisfied.
In assuring us that this first meaning of the word is met, the result draws
our attention to the second meaning. It has to do with the two remaining
conditions, the closure conditions. Above, the two separate closure conditions
inherent in statement (1) are combined in statement (2) into the single condition
of closure under all linear combinations of two vectors, which is then extended
in statement (3) to closure under combinations of any number of vectors. The
latter two statements say that we can always make sense of an expression like
r1~s1 + r2~s2 , without restrictions on the rs such expressions are sensible in
that the vector described is defined and is in the set S.
This second meaning suggests that a good way to think of a vector space
is as a collection of unrestricted linear combinations. The next two examples
take some spaces and describe them in this way. That is, in these examples we
paramatrize, just as we did in Chapter One to describe the solution set of a
homogeneous linear system.
2.11 Example This subset of R3
x
S = {y x 2y + z = 0}
z
is a subspace under the usual addition and scalar multiplication operations of
column vectors (the check that it is nonempty and closed under linear combinations of two vectors is just like the one in Example 2.2). To paramatrize, we
95
2
1
2y z
S = { y y, z R} = {y 1 + z 0 y, z R}
0
1
z
Now the subspace is described as the collection of unrestricted linear combinations of those two vectors. Of course, in either description, this is a plane
through the origin.
2.12 Example This is a subspace of the 22 matrices
a 0
a + b + c = 0}
L={
b c
(checking that it is nonempty and closed under linear combinations is easy). To
paramatrize, express the condition as a = b c.
1 0
1 0
b c 0
b, c R}
b, c R} = {b
+c
L={
1 0
0 1
b
c
As above, weve described the subspace as a collection of unrestricted linear
combinations (by coincidence, also of two elements).
Paramatrization is an easy technique, but it is important. We shall use it
often.
2.13 Definition The span (or linear closure) of a nonempty subset S of a
vector space is the set of all linear combinations of vectors from S.
96
Proof. Call the subset S. If S is empty then by definition its span is the trivial
subspace. If S is not empty then by Lemma 2.9 we need only check that the
span [S] is closed under linear combinations. For a pair of vectors from that
~ = cn+1~sn+1 + +cm~sm , a linear combination
span, ~v = c1~s1 + +cn~sn and w
2.16 Example In any vector space V , for any vector ~v , the set {r ~v r R}
is a subspace of V . For instance, for any vector ~v R3 , the line through the
origin containing that vector, {k~v k R} is a subspace of R3 . This is true even
when ~v is the zero vector, in which case the subspace is the degenerate line, the
trivial subspace.
2.17 Example The span of this set is all of R2 .
1
1
{
,
}
1
1
Tocheck this we must show that any member of R2 is a linear combination of
these two vectors. So we ask: for which vectors (with real components x and y)
are there scalars c1 and c2 such that this holds?
1
1
x
c1
+ c2
=
1
1
y
Gauss method
c1 + c2 = x
c1 c2 = y
1 +2
c1 +
c2 =
x
2c2 = x + y
97
Since spans are subspaces, and we know that a good way to understand a
subspace is to paramatrize its description, we can try to understand a sets span
in that way.
2.18 Example Consider, in P2 , the span of the set {3x x2 , 2x}. By the
definition of span, it is the subspace of unrestricted linear combinations of the
two {c1 (3x x2 ) + c2 (2x) c1 , c2 R}. Clearly polynomials in this span must
have a constant term of zero. Is that necessary condition also sufficient?
We are asking: for which members a2 x2 + a1 x + a0 of P2 are there c1 and c2
such that a2 x2 + a1 x + a0 = c1 (3x x2 ) + c2 (2x)? Since polynomials are equal
if and only if their coefficients are equal, we are looking for conditions on a2 ,
a1 , and a0 satisfying these.
= a2
c1
3c1 + 2c2 = a1
0 = a0
Gauss method gives that c1 = a2 , c2 = (3/2)a2 + (1/2)a1 , and 0 = a0 . Thus
the only condition on polynomials in the span is the condition that we knew
of as long as a0 = 0, we can give appropriate coefficients c1 and c2 to describe
the polynomial a0 + a1 x + a2 x2 as in the span. For instance, for the polynomial
c1 = 3 and c2 = 5/2 will do. So the span of the
0 4x + 3x2 , the coefficients
1
{x
0
0
1
0
{x
0
0
H
H
A
1 A 0HH 2
{x
0
0
+y
1
0
1
0
0
+y
0
0
+z
0
1
{x
1
0
+z
...
{y 1 }
{y 1 }
{y 1 }
0
0
1
XXX PP
H
XXXPP HH @
XXPPPH
H@ 0
XXXP
X { 0 }
0
0
1
1
0
+z
...
0
1
98
The subsets are described as spans of sets, using a minimal number of members,
and are shown connected to their supersets. Note that these subspaces fall
naturally into levels planes on one level, lines on another, etc. according
to how many vectors are in a minimal-sized spanning set.
So far in this chapter we have seen that to study the properties of linear
combinations, the right setting is a collection that is closed under these combinations. In the first subsection we introduced such collections, vector spaces,
and we saw a great variety of examples. In this subsection we saw still more
spaces, ones that happen to be subspaces of others. In all of the variety weve
seen a commonality. Example 2.19 above brings it out: vector spaces and subspaces are best understood as a span, and especially as a span of a small number
of vectors. The next section studies spanning sets that are minimal.
Exercises
X 2.20 Which of these subsets of the vector space of 2 2 matrices are subspaces
under the inherited operations? For each one that is a subspace, paramatrize its
description.
For
a
(b) {
0
0
b
a
(c) {
0
0
b
a
0
c
b
(d) {
a + b = 0}
a + b = 5}
a + b = 0, c R}
X 2.21 Is this a subspace of P2 : {a0 + a1 x + a2 x2 a0 + 2a1 + a2 = 4}? If so, paramatrize its description.
X 2.22 Decide if the vector lies in the span of the set, inside of the space.
! ! !
2
1
0
(a) 0 , { 0 , 0 }, in R3
1
0
1
(b)
x x3
, {x2
, 2x +
x2 ,
x + x3
}, in P3
0 1
1 0
2 0
(c)
,{
,
}, in M22
4 2
1 1
2 3
2.23 Which of these are members of the span [{cos2 x, sin2 x}] in the vector space
of real-valued functions of one real variable?
(a) f (x) = 1
(b) f (x) = 3 + x2
(c) f (x) = sin x
(d) f (x) = cos(2x)
X 2.24 Which of these sets spans R3 ? That is, which of these sets has the property
that any three-tall vector can be expressed as a suitable linear combination of the
sets elements?
! ! !
! ! !
! !
1
0
0
2
1
0
1
3
(a) { 0 , 2 , 0 }
(b) { 0 , 1 , 0 }
(c) { 1 , 0 }
0
0
3
1
0
1
0
0
! ! ! !
! ! ! !
1
3
1
2
2
3
5
6
(d) { 0 , 1 , 0 , 1 }
(e) { 1 , 0 , 1 , 0 }
1
0
0
5
1
1
2
2
99
{
(d) This subset of M22
a
{
c
b
d
a
c
b
d
a + d = 0}
2a c d = 0 and a + 3b = 0}
(a) The even functions {f : R R f (x) = f (x) for all x}. For example, two
members of this set are f1 (x) = x2 and f2 (x) = cos(x).
(b) The odd functions {f : R R f (x) = f (x) for all x}. Two members are
f3 (x) = x3 and f4 (x) = sin(x).
2.29 Example
2.16 says that for any vector ~v in any vector space V , the set
100
101
102
2.II
Linear Independence
The prior section shows that a vector space can be understood as an unrestricted
linear combination of some of its elements
that is, as a span. For example,
the space of linear polynomials {a + bx a, b R} is spanned by the set {1, x}.
The prior section also showed that a space can have many sets that span it.
The space of linear polynomials is also spanned by {1, 2x} and {1, x, 2x}.
At the end of that section we described some spanning sets as minimal,
but we never precisely defined that word. We could take minimal to mean one
of two things. We could mean that a spanning set is minimal if it contains the
smallest number of members of any set with the same span. With this meaning
{1, x, 2x} is not minimal because it has one member more than the other two.
Or we could mean that a spanning set is minimal when it has no elements that
can be removed without changing the span. Under this meaning {1, x, 2x} is not
minimal because removing the 2x and getting {1, x} leaves the span unchanged.
The first sense of minimality appears to be a global requirement, in that to
check if a spanning set is minimal we seemingly must look at all the spanning sets
of a subspace and find one with the least number of elements. The second sense
of minimality is local in that we need to look only at the set under discussion
and consider the span with and without various elements. For instance, using
the second sense, we could compare the span of {1, x, 2x} with the span of {1, x}
and note that the 2x is a repeat in that its removal doesnt shrink the span.
In this section we will use the second sense of minimal spanning set because
of this technical convenience. However, the most important result of this book
is that the two senses coincide; we will prove that in the section after this one.
2.II.1
103
0
~v2 = 1 ,
0
2
~v3 = 1
0
the spans [{~v1 , ~v2 }] and [{~v1 , ~v2 , ~v3 }] are equal since ~v3 is in the span [{~v1 , ~v2 }].
The lemma says that if we have a spanning set then we can remove a ~v to
get a new set S with the same span if and only if ~v is a linear combination of
vectors from S. Thus, under the second sense described above, a spanning set
is minimal if and only if it contains no vectors that are linear combinations of
the others in that set. We have a term for this important property.
1.3 Definition A subset of a vector space is linearly independent if none of its
elements is a linear combination of the others. Otherwise it is linearly dependent.
Here is a small but useful observation: although this way of writing one
vector as a combination of the others
~s0 = c1~s1 + c2~s2 + + cn~sn
visually sets ~s0 off from the other vectors, algebraically there is nothing special
in that equation about ~s0 . For any ~si with a coefficient ci that is nonzero, we
can rewrite the relationship to set off ~si .
~si = (1/ci )~s0 + (c1 /ci )~s1 + + (cn /ci )~sn
When we dont want to single out any vector by writing it alone on one side of
the equation we will instead say that ~s0 , ~s1 , . . . , ~sn are in a linear relationship and
write the relationship with all of the vectors on the same side. The next result
rephrases the linear independence definition in this style. It gives what is usually
the easiest way to compute whether a finite set is dependent or independent.
1.4 Lemma A subset S of a vector space is linearly independent if and only if
for any distinct ~s1 , . . . , ~sn S the only linear relationship among those vectors
c1~s1 + + cn~sn = ~0
c1 , . . . , cn R
If the set S is linearly independent then no vector ~si can be written as a linear
combination of the other vectors from S so there is no linear relationship where
some of the ~s s have nonzero coefficients. If S is not linearly independent then
some ~si is a linear combination ~si = c1~s1 + +ci1~si1 +ci+1~si+1 + +cn~sn of
other vectors from S, and subtracting ~si from both sides of that equation gives
a linear relationship involving a nonzero coefficient, namely the 1 in front of
QED
~si .
104
1.5
Example
In the vector space of two-wide row vectors, the two-element set
{ 40 15 , 50 25 } is linearly independent. To check this, set
c1 40
15 + c2 50
25 = 0
(15/40)1 +2
40c1
50c2 = 0
(175/4)c2 = 0
Therefore c1 and c2 both equal zero and so the only linear relationship between
the two given row vectors is the
trivialrelationship.
c1 40 15 + c2 20 7.5 = 0 0
with c1 = 1 and c2 = 2.
1.6 Remark Recall the Statics example that began this book. We first set the
unknown-mass objects at 40 cm and 15 cm and got a balance, and then we set
the objects at 50 cm and 25 cm and got a balance. With those two pieces of
information we could compute values of the unknown masses. Had we instead
first set the unknown-mass objects at 40 cm and 15 cm, and then at 20 cm and
7.5 cm, we would not have been able to compute the
values of the unknown
20 7.5 information is a
masses (try it). Intuitively,
the
problem
is
that
the
repeat
of the
1 +2
c1 + c2 = 0
2c2 = 0
(since polynomials are equal only if their coefficients are equal). Thus, the only
linear relationship between these two members of P2 is the trivial one.
1.8 Example In R3 , where
3
~v1 = 4 ,
5
2
~v2 = 9 ,
2
4
~v3 = 18
4
105
the set S = {~v1 , ~v2 , ~v3 } is linearly dependent because this is a relationship
0 ~v1 + 2 ~v2 1 ~v3 = ~0
where not all of the scalars are zero (the fact that some scalars are zero is
irrelevant).
1.9 Remark That example shows why, although Definition 1.3 is a clearer
statement of what independence is, Lemma 1.4 is more useful for computations.
Working straight from the definition, someone trying to compute whether S is
linearly independent would start by setting ~v1 = c2~v2 +c3~v3 and concluding that
there are no such c2 and c3 . But knowing that the first vector is not dependent
on the other two is not enough. Working straight from the definition, this
person would have to go on to try ~v2 = c3~v3 in order to find the dependence
c3 = 1/2. Similarly, working straight from the definition, a set with four vectors
would require checking three vector equations. Lemma 1.4 makes the job easier
because it allows us to get the conclusion with only one computation.
1.10 Example The empty subset of a vector space is linearly independent.
There is no nontrivial linear relationship among its members as it has no members.
1.11 Example In any vector space, any subset containing the zero vector is
linearly dependent. For example, in the space P2 of quadratic polynomials,
consider the subset {1 + x, x + x2 , 0}.
One way to see that this subset is linearly dependent is to use Lemma 1.4: we
have 0 ~v1 + 0 ~v2 + 1 ~0 = ~0, and this is a nontrivial relationship as not all of the
coefficients are zero. Another way to see that this subset is linearly dependent
is to go straight to Definition 1.3: we can express the third member of the subset
as a linear combination of the first two, namely, c1~v1 + c2~v2 = ~0 is satisfied by
taking c1 = 0 and c2 = 0 (in contrast to the lemma, the definition allows all of
the coefficients to be zero).
(There is still another way to see this that is somewhat trickier. The zero
vector is equal to the trivial sum, that is, it is the sum of no vectors. So in
a set containing the zero vector, there is an element that can be written as a
combination of a collection of other vectors from the set, specifically, the zero
vector can be written as a combination of the empty collection.)
Lemma 1.1 suggests how to turn a spanning set into a spanning set that is
minimal. Given a finite spanning set, we can repeatedly pull out vectors that
are a linear combination of the others, until there arent any more such vectors
left.
1.12 Example This set spans R3 .
1
0
1
0
3
S0 = {0 , 2 , 2 , 1 , 3}
0
0
0
1
0
106
1
3
c1
1
3/2
c2
{c3 = c3 1 + c5
0 c3 , c5 R}
0
0
c4
c5
0
1
Setting, say, c3 = 0 and c5 = 1 shows that the fifth vector is a linear combination
of the first two. Thus, Lemma 1.1 gives that this set
1
0
1
0
S1 = {0 , 2 , 2 , 1}
0
0
0
1
has the same span as S0 . Similarly, the third vector of the new set S1 is a linear
combination of the first two and we get
1
0
0
S2 = {0 , 2 , 1}
0
0
1
with the same span as S1 and S0 , but with one difference. This last set is
linearly independent (this is easily checked), and so removal of any of its vectors
will shrink the span.
We finish this subsection by recasting that example as a theorem that any
finite spanning set has a subset with the same span that is linearly independent.
To prove that result we will first need some facts about how linear independence
and dependence, which are properties of sets, interact with the subset relation
between sets.
1.13 Lemma Any subset of a linearly independent set is also linearly independent. Any superset of a linearly dependent set is also linearly dependent.
Proof. This is clear.
QED
107
1.14 Example Here are some linearly independent sets from R3 and their
supersets.
1
(1) If S1 = {0} then the span [S1 ] is the x-axis.
0
1
3
{0 , 0 }
0 0
1
0
{0 , 1}
0
0
1
0
(2) If S2 = {0 , 1} then [S2 ] is the xy-plane.
0
0
1
0
3
A linearly dependent superset: {0 , 1 , 2}
0
0 0
1
0
0
A linearly independent superset: {0 , 1 , 0}
0
0
1
1
0
0
(3) If S3 = {0 , 1 , 0} then [S3 ] is all of R3 .
0
0
1
1
0
0
2
A linearly dependent superset: {0 , 1 , 0 , 1}
0
0
1
3
There are no linearly independent supersets.
(Checking the dependence or independence of these sets is easy.)
So in general a linearly independent set can have some supersets that are dependent and some supersets that are independent. We can characterize when a
superset of a independent set is dependent and when it is independent.
1.15 Lemma Where S is a linearly independent subset of a vector space V ,
S {~v } is linearly dependent if and only if ~v [S]
for any ~v V with ~v 6 S.
108
S independent
S dependent
KS
K must be independent
K may be either
KS
K may be either
K must be dependent
109
Proof. If the finite set S is linearly independent then there is nothing to prove
so assume that S = {~s1 , . . . , ~sn } is linearly dependent. By Corollary 1.16, there
is a vector ~si that is a linear combination of ~s1 , . . . , ~si1 . Define S1 to be the
set S {~si }. Lemma 1.1 then says that the span does not shrink: [S1 ] = [S].
If S1 is linearly independent then we are finished. Otherwise repeat the
prior paragraph to derive S2 S1 such that [S2 ] = [S1 ]. Repeat this process
until a linearly independent set appears; one must eventually appear because
S is finite. (Formally, this part of the argument uses mathematical induction.
QED
Exercise 37 asks for the details.)
110
1
S={ 1
0
1
2 }
0
!
3
2
0
c1
1
1
0
+ c2
1
2
0
3
2
0
111
1.32 (a) Show that any set of four vectors in R2 is linearly dependent.
(b) Is this true for any set of five? Any set of three?
(c) What is the most number of elements that a linearly independent subset of
R2 can have?
X 1.33 Is there a set of four vectors in R3 , any three of which form a linearly independent set?
1.34 Must every linearly dependent set have a subset that is dependent and a
subset that is independent?
1.35 In R4 , what is the biggest linearly independent set you can find? The smallest?
The biggest linearly dependent set? The smallest? (Biggest and smallest mean
that there are no supersets or subsets with the same property.)
X 1.36 Linear independence and linear dependence are properties of sets. We can
thus naturally ask how those properties act with respect to the familiar elementary
set relations and operations. In this body of this subsection we have covered the
subset and superset relations. We can also consider the operations of intersection,
complementation, and union.
(a) How does linear independence relate to intersection: can an intersection of
linearly independent sets be independent? Must it be?
(b) How does linear independence relate to complementation?
(c) Show that the union of two linearly independent sets need not be linearly
independent.
(d) Characterize when the union of two linearly independent sets is linearly independent, in terms of the intersection of the span of each.
X 1.37 For Theorem 1.17,
(a) fill in the induction for the proof;
(b) give an alternate proof that starts with the empty set and builds a sequence
of linearly independent subsets of the given finite set until one appears with the
same span as the given set.
1.38 With a little calculation we can get formulas to determine whether or not a
set of vectors is linearly independent.
(a) Show that this subset of R2
a
b
,
}
c
d
! ! !
a
{ d
g
b
e
h
c
f }
i
! !
a
{ d
g
b
e }
h
linearly independent?
(d) This is an opinion question: for a set of four vectors from R4 , must there be
a formula involving the sixteen entries that determines independence of the set?
(You neednt produce such a formula, just decide if one exists.)
112
X 1.39 (a) Prove that a set of two perpendicular nonzero vectors from Rn is linearly
independent when n > 1.
(b) What if n = 1? n = 0?
(c) Generalize to more than two vectors.
1.40 Consider the set of functions from the open interval (1..1) to R.
(a) Show that this set is a vector space under the usual operations.
(b) Recall the formula for the sum of an infinite geometric series: 1+x+x2 + =
1/(1x) for all x (1..1). Why does this not express a dependence inside of the
set {g(x) = 1/(1 x), f0 (x) = 1, f1 (x) = x, f2 (x) = x2 , . . . } (in the vector space
that we are considering)? (Hint. Review the definition of linear combination.)
(c) Show that the set in the prior item is linearly independent.
This shows that some vector spaces exist with linearly independent subsets that
are infinite.
1.41 Show that, where S is a subspace of V , if a subset T of S is linearly independent in S then T is also linearly independent in V . Is that only if ?
2.III
113
The prior section ends with the statement that a spanning set is minimal when
it is linearly independent and that a linearly independent set is maximal when
it spans the space. So the notions of minimal spanning set and maximal independent set coincide. In this section we will name this notion and study some
of its properties.
2.III.1
Basis
1.1 Definition A basis for a vector space is a sequence of vectors that form a
set that is linearly independent and that spans the space.
~2 , . . . i to signify that this collec~1 ,
We denote a basis with angle brackets h
tion is a sequence the order of the elements is significant. (The requirement
that a basis be ordered will be needed, for instance, in Definition 1.13.)
1.2 Example This is a basis for R2 .
2
1
h
,
i
4
1
It is linearly independent
2
1
0
+ c2
=
c1
4
1
0
2c1 + 1c2 = 0
4c1 + 1c2 = 0
c1 = c2 = 0
and it spans R2 .
2c1 + 1c2 = x
4c1 + 1c2 = y
c2 = 2x y and c1 = (y x)/2
1
2
h
,
i
1
4
differs from the prior one because of its different order. The verification that it
is a basis is just as in the prior example.
1.4 Example The space R2 has many bases. Another one is this.
1
0
h
,
i
0
1
The verification is easy.
114
0
0
.. i
.
1
is the standard (or natural) basis. We denote these vectors by ~e1 , . . . , ~en .
Note that the symbol ~e1 means something different in a discussion of R3 than
it means in a discussion of R2 . (Calculus books call R2 s standard basis vectors
~ and ~ instead of ~e1 and ~e2 , and they call R3 s standard basis vectors ~, ~, and
~k instead of ~e1 , ~e2 , and ~e3 .)
1.6 Example We can give bases for spaces other than just those comprised
of
w=0
z+w=0
by paramatrizing.
1
1
1
0
{
0 y + 1 w y, w R}
0
1
That is, we have described the vector space of solutions as the span of a twoelement set. We can easily check that this two-vector set is also linearly independent. Thus the solution set is a subspace of R4 with a two-element basis.
115
1.11 Example Parameterization helps find bases for other vector spaces, not
just for solution sets of homogeneous systems. To find a basis for this subspace
of M22
a b
a + b 2c = 0}
{
c 0
we rewrite the condition as a = b + 2c to get this.
1 1
2 0
b + 2c b
b, c R}
b, c R} = {b
+c
{
0 0
1 0
c
0
Thus, this is a natural candidate for a basis.
1 1
2 0
h
,
i
0 0
1 0
The above work shows that it spans the space. To show that it is linearly
independent is routine.
Consider Example 1.2 again.
To show that the basis spans the space we
looked at a general vector xy from R2 . We found a formula for coefficients c1
and c2 in terms of x and y. Although we did not mention it in the example,
the formula shows that for each vector there is only one suitable coefficient pair.
This always happens.
1.12 Theorem In any vector space, a subset is a basis if and only if each
vector in the space can be expressed as a linear combination of elements of the
subset in a unique way. (We consider combinations to be the same if they differ
only in the order of summands or in the addition or deletion of terms of the
~
form 0 .)
Proof. By definition, a sequence is a basis if and only if its vectors form both
a spanning set and a linearly independent set. A subset is a spanning set if
and only if each vector in the space is a linear combination of elements of that
subset in at least one way.
Thus, to finish we need only show that a subset is linearly independent if
and only if every vector in the space is a linear combination of elements from
the subset in at most one way. Consider two expressions of a vector as a linear
combination of the members of the basis. We can rearrange the two sums, and
~ in the same
~i s, so that the two combine the same s
if necessary add some 0
~
~
~
~
~
~n . Now,
order: ~v = c1 1 + c2 2 + + cn n and ~v = d1 1 + d2 2 + + dn
equality
~1 + c2
~2 + + cn
~n = d1
~1 + d2
~2 + + dn
~n
c1
holds if and only if
~1 + + (cn dn )
~n = ~0
(c1 d1 )
holds, and so asserting that each coefficient in the lower equation is zero is the
QED
same thing as asserting that ci = di for each i.
116
0
1/2
RepB (x + x2 ) =
1/2
0 B
(note that the coordinates are scalars, not vectors). With respect to a different
basis D = h1 + x, 1 x, x + x2 , x + x3 i, the representation
0
RepD (x + x2 ) =
1
0 D
is different.
1.15 Remark This use of column notation and the term coordinates has
both a down side and an up side.
The down side is that representations look like vectors from Rn , and that
can be confusing when the vector space we are working with is Rn , especially
since we sometimes omit the subscript base. We must then infer the intent from
the context. For example, the phrase in R2 , where
3
~v =
, ...
2
refers to the plane vector that, when in canonical position, ends at (3, 2). To
find the coordinates of that vector with respect to the basis
1
0
B=h
,
i
1
2
we solve
c1
1
0
3
+ c2
=
1
2
2
117
3
RepB (~v ) =
1/2
Here, although weve ommited the subscript B from the column, the fact that
the right side it is a representation is clear from the context.
The up side of the notation and the term coordinates is that they generalize
the use that we are familiar with: in Rn and with respect to the standard
basis En , the vector starting at the origin and ending at (v1 , . . . , vn ) has this
representation.
v1
v1
..
..
RepEn ( . ) = .
vn
vn
En
Our main use of representations will come in the third chapter. The definition appears here because the fact that every vector is a linear combination of
basis vectors in a unique way is a crucial property of bases, and also to help make
two points. First, we put the elements of a basis in a fixed order so that coordinates can stated in that order. Second, for calculation of coordinates, among
other things, we shall want our bases to have only finitely many elements. We
will see that in the next subsection.
Exercises
3
X 1.16 Decide
each
!if
! isa basis
! for R . ! !
! ! !
1
3
0
1
3
0
1
2
(a) h 2 , 2 , 0 i
(b) h 2 , 2 i
(c) h 2 , 1 , 5 i
3
1
1
3
1
1
1
0
! ! !
0
1
1
(d) h 2 , 1 , 3 i
1
1
0
X 1.17 Represent
the
vector
with
respect to the basis.
1
1
1
(a)
,B=h
,
i R2
2
1
1
(b) x2 + x3 , D = h1, 1 + x, 1 + x + x2 , 1 + x + x2 + x3 i P3
0
1
(c) , E4 R4
0
1
1.18 Find a basis for P2 , the space of all quadratic polynomials. Must any such
basis contain a polynomial of each degree: degree zero, degree one, and degree
two?
1.19 Find a basis for the solution set of this system.
x1 4x2 + 3x3 x4 = 0
2x1 8x2 + 6x3 2x4 = 0
118
a
0
b
c
c 2b = 0}
119
(b) Show that no spanning subset of R3 contains fewer than three vectors. (Hint.
Recall how to calculate the span of a set and show that this method, when applied
to two vectors, cannot yield all of R3 .)
1.34 One of the exercises in the Subspaces subsection shows that the set
!
x
{ y x + y + z = 1}
z
is a vector space under these operations.
! !
!
!
!
x1
x2
x1 + x2 1
x
rx r + 1
y1 + y2
ry
y1 + y2 =
r y =
z
rz
z1
z2
z1 + z2
Find a basis.
2.III.2
Dimension
In the prior subsection we saw that a vector space can have many different
bases. For example, following the definition of a basis, we saw three different bases for R2 . So we cannot talk about the basis for a vector space.
True, some vector spaces have bases that strike us as more natural than others, for instance, R2 s basis E2 or R3 s basis E3 or P2 s basis h1, x, x2 i. But
the idea of natural
is hard to make formal. For example, with the space
120
we study only finite-dimensional vector spaces. We shall take the term vector
space to mean finite-dimensional vector space. Infinite-dimensional spaces are
interesting and important, but they lie outside of our scope.
To prove the main theorem we shall use a technical result.
~1 , . . . , ~n i is a basis
2.2 Lemma (Exchange Lemma) Assume that B = h
~1 + c2
~2 +
for a vector space, and that for the vector ~v the relationship ~v = c1
~
~
+ cn n has ci 6= 0. Then exchanging i for ~v yields another basis for the
space.
~1 , . . . , ~i1 , ~v , ~i+1 , . . . ,
~n i.
= h
Proof. Call the outcome of the exchange B
~1 + +
is linearly independent. Any relationship d1
We first show that B
~n = ~0 among the members of B,
after substitution for ~v ,
di~v + + dn
~1 + + di (c1
~1 + + ci
~i + + cn
~ n ) + + dn
~n = ~0
d1
()
QED
[B].
2.3 Theorem In any finite-dimensional vector space, all of the bases have the
same number of elements.
Proof. Fix a vector space with at least one finite basis. Choose, from among all
~1 , . . . , ~n i of minimal size. We will show that any
of this spaces bases, B = h
~
~
other basis D = h1 , 2 , . . . i also has the same number of members, n. Because
B has minimal size, D has no fewer than n vectors. We will argue that it cannot
have more.
The basis B spans the space and ~1 is in the space, so ~1 is a nontrivial linear
combination of elements of B. By the Exchange Lemma, ~1 can be swapped for
a vector from B, resulting in a basis B1 , where one element is ~ and all of the
~
n 1 other elements are s.
The prior paragraph forms the basis step for an induction argument. The
inductive step starts with a basis Bk (for 1 k < n) containing k members of D
and n k members of B. We know that D has at least n members so there is a
121
2.9 Example Recall the subspace diagram from the prior section showing the
subspaces of R3 . Each subspaces shown is described with a minimal spanning
set, for which we now have the term basis. The whole space has a basis with
three members, the plane subspaces have bases with two members, the line
subspaces have bases with one member, and the trivial subspace has a basis
with zero members. When we saw that diagram we could not show that these
are the only subspaces that this space has. We can show it now. The prior
corollary proves the only subspaces of R3 are either three-, two-, one-, or zerodimensional. Therefore, the diagram indicates all of the subspaces. There are
no subspaces somehow, say, between lines and planes.
2.10 Corollary Any linearly independent set can be expanded to make a basis.
Proof. If a linearly independent set is not already a basis then it must not
span the space. Adding to it a vector that is not in the span preserves linear
independence. Keep adding, until the resulting set does span the space, which
the prior corollary shows will happen after only a finite number of steps. QED
122
2.12 Corollary In an n-dimensional space, a set of n vectors is linearly independent if and only if it spans the space.
Proof. First we will show that a subset with n vectors is linearly independent
if and only if it is a basis. If is trivially true bases are linearly independent.
Only if holds because a linearly independent set can be expanded to a basis,
but a basis has n elements, so that this expansion is actually the set we began
with.
To finish, we will show that any subset with n vectors spans the space if and
only if it is a basis. Again, if is trivial. Only if holds because any spanning
set can be shrunk to a basis, but a basis has n elements and so this shrunken
QED
set is just the one we started with.
The main result of this subsection, that all of the bases in a finite-dimensional
vector space have the same number of elements, is the single most important
result in this book because, as Example 2.9 shows, it describes what vector
spaces and subspaces there can be. We will see more in the next chapter.
2.13 Remark The case of infinite-dimensional vector spaces is somewhat controversial. The statement any infinite-dimensional vector space has a basis
is known to be equivalent to a statement called the Axiom of Choice (see
[Blass 1984]). Mathematicians differ philosophically on whether to accept or
reject this statement as an axiom on which to base mathematics. Consequently
the question about infinite-dimensional vector spaces is still somewhat up in the
air. (A discussion of the Axiom of Choice can be found in the Frequently Asked
Questions list for the Usenet group sci.math. Another accessible reference is
[Rucker].)
Exercises
Assume that all spaces are finite-dimensional unless otherwise stated.
X 2.14 Find a basis for, and the dimension of, P2 .
2.15 Find a basis for, and the dimension of, the solution set of this system.
x1 4x2 + 3x3 x4 = 0
2x1 8x2 + 6x3 2x4 = 0
X 2.16 Find a basis for, and the dimension of, M22 , the vector space of 22 matrices.
123
a
c
b
d
124
subspaces of some vector space and that U W . Can there exist bases BU for
U and BW for W such that BU BW ? Must such bases exist?
For any basis BU for U , must there be a basis BW for W such that BU BW ?
For any basis BW for W , must there be a basis BU for U such that BU BW ?
For any bases BU , BW for U and W , must BU be a subset of BW ?
(b) Is the intersection of bases a basis? For what space?
(c) Is the union of bases a basis? For what space?
(d) What about complement?
(Hint. Test any conjectures against some subspaces of R3 .)
X 2.33 Consider how dimension interacts with subset. Assume U and W are both
subspaces of some vector space, and that U W .
(a) Prove that dim(U ) dim(W ).
(b) Prove that equality of dimension holds if and only if U = W .
(c) Show that the prior item does not hold if they are infinite-dimensional.
2.34 [Wohascum no. 47] For any vector ~v in Rn and any permutation of the
numbers 1, 2, . . . , n (that is, is a rearrangement of those numbers into a new
order), define (~v ) to be the vector whose components are v(1) , v(2) , . . . , and
v(n) (where (1) is the first
number in the rearrangement, etc.). Now fix ~v and
let V be the span of {(~v ) permutes 1, . . . , n}. What are the possibilities for
the dimension of V ?
2.III.3
We will now reconsider linear systems and Gauss method, aided by the tools
and terms of this chapter. We will make three points.
For the first point, recall the Linear Combination Lemma and its corollary: if
two matrices are related by row operations A B then each row of B
is a linear combination of the rows of A. That is, Gauss method works by taking
linear combinations of rows. Therefore, the right setting in which to study row
operations in general, and Gauss method in particular, is the following vector
space.
3.1 Definition The row space of a matrix is the span of the set of its rows. The
row rank is the dimension of the row space, the number of linearly independent
rows.
3.2 Example If
A=
2 3
4 6
{c1 2 3 + c2 4 6 c1 , c2 R}
The linear dependence of
on the first is obvious and so we can simplify
the second
this description to {c 2 3 c R}.
125
A B
or
ki
A B
or
ki +j
A B
(for i 6= j and k 6= 0) then their row spaces are equal. Hence, row-equivalent
matrices have the same row space, and hence also, the same row rank.
Proof. The row space of A is the set of all linear combinations of the rows
Thus, in the language of this chapter, Gaussian reduction works by eliminating linear dependences among rows, leaving the span unchanged, until no
nontrivial linear relationships remain (among the nonzero rows). That is, Gauss
method produces a basis for the row space.
3.5 Example From any matrix, we can produce a basis for the row space by
performing Gauss method and taking the nonzero rows of the resulting echelon
form matrix. For instance,
1 3 1
1 3 1
1 +2 62 +3
0 1 0
1 4 1
21 +3
0 0 3
2 0 5
126
3.6 Definition The column space of a matrix is the span of the set of its
columns. The column rank is the dimension of the column space, the number
of linearly independent columns.
Our interest in column spaces stems from our study of linear systems. An
example is that this system
c1 + 3c2 + 7c3 = d1
2c1 + 3c2 + 8c3 = d2
c2 + 2c3 = d3
4c1
+ 4c3 = d4
has a solution if and only if the vector of ds is a linear combination of the other
column vectors,
1
3
7
d1
2
3
8 d2
c1
0 + c2 1 + c3 2 = d3
d4
4
0
4
meaning that the vector of ds is in the column space of the matrix of coefficients.
3.7 Example Given this matrix,
1
2
0
4
3
3
1
0
7
8
2
4
to get a basis for the column space, temporarily turn the columns into rows and
reduce.
1 2 0
4
1 2 0 4
1 +2 22 +3
0 3 1 12
3 3 1 0 3
71 +3
0 0 0
0
7 8 2 4
Now turn the rows back to columns.
1
0
2 3
h
0 , 1 i
4
12
The result is a basis for the column space of the given matrix.
3.8 Definition The transpose of a matrix is the result of interchanging the
rows and columns of that matrix. That is, column j of the matrix A is row j of
Atrans , and vice versa.
127
So the instructions for the prior example are transpose, reduce, and transpose
back.
We can even, at the price of tolerating the as-yet-vague idea of vector spaces
being the same, use Gauss method to find bases for spans in other types of
vector spaces.
3.9 Example To get a basis for the span of {x2 + x4 , 2x2 + 3x4 , x2 3x4 }
in the space
P4 , think of these three polynomials
0 0 1 0 1
0 0 1 0 1
21 +2 22 +3
0 0 0 0 1
0 0 2 0 3
1 +3
0 0 0 0 0
0 0 1 0 3
and translate back to get the basis hx2 + x4 , x4 i. (As mentioned earlier, we will
make the phrase the same precise at the start of the next chapter.)
Thus, our first point in this subsection is that the tools of this chapter give
us a more conceptual understanding of Gaussian reduction.
For the second point of this subsection, consider the effect on the column
space of this row reduction.
21 +2
1 2
1 2
0 0
2 4
The column space of the left-hand matrix contains vectors with a second component that is nonzero. But the column space of the right-hand matrix is different
because it contains only vectors whose second component is zero. It is this
knowledge that row operations can change the column space that makes next
result surprising.
3.10 Lemma Row operations do not change the column rank.
Proof. Restated, if A reduces to B then the column rank of B equals the
column rank of A.
We will be done if we can show that row operations do not affect linear relationships among columns (e.g., if the fifth column is twice the second plus the
fourth before a row operation then that relationship still holds afterwards), because the column rank is just the size of the largest set of unrelated columns. But
this is exactly the first theorem of this book: in a relationship among columns,
a1,1
a1,n
0
a2,1
a2,n 0
c1 . + + cn . = .
..
.. ..
0
am,1
am,n
QED
128
Another way, besides the prior result, to state that Gauss method has something to say about the column space as well as about the row space is to consider
again Gauss-Jordan reduction. Recall that it ends with the reduced echelon form
of a matrix, as here.
1 3 1 6
1 3 0 2
2 6 3 16 0 0 1 4
1 3 1 6
0 0 0 0
Consider the row space and the column space of this result. Our first point
made above says that a basis for the row space is easy to get: simply collect
together all of the rows with leading entries. However, because this is a reduced
echelon form matrix, a basis for the column space is just as easy: take the
columns containing the leading entries, that is, h~e1 , ~e2 i. (Linear independence
is obvious. The other columns are in the span of this set, since they all have
a third component of zero.) Thus, for a reduced echelon form matrix, bases
for the row and column spaces can be found in essentially the same way
by taking the parts of the matrix, the rows or columns, containing the leading
entries.
3.11 Theorem The row rank and column rank of a matrix are equal.
Proof. First bring the matrix to reduced echelon form. At that point, the
row rank equals the number of leading entries since each equals the number
of nonzero rows. Also at that point, the number of leading entries equals the
column rank because the set of columns containing leading entries consists of
some of the ~ei s from a standard basis, and that set is linearly independent and
spans the set of columns. Hence, in the reduced echelon form matrix, the row
rank equals the column rank, because each equals the number of leading entries.
But Lemma 3.3 and Lemma 3.10 show that the row rank and column rank
are not changed by using row operations to get to reduced echelon form. Thus
the row rank and the column rank of the original matrix are also equal. QED
3.12 Definition The rank of a matrix is its row rank or column rank.
So our second point in this subsection is that the column space and row
space of a matrix have the same dimension. Our third and final point is that
the concepts that weve seen arising naturally in the study of vector spaces are
exactly the ones that we have studied with linear systems.
3.13 Theorem For linear systems with n unknowns and with matrix of coefficients A, the statements
(1) the rank of A is r
(2) the space of solutions of the associated homogeneous system has dimension n r
are equivalent.
129
So if the system has at least one particular solution then for the set of solutions,
the number of parameters equals n r, the number of variables minus the rank
of the matrix of coefficients.
Proof. The rank of A is r if and only if Gaussian reduction on A ends with r
nonzero rows. Thats true if and only if echelon form matrices row equivalent
to A have r-many leading variables. That in turn holds if and only if there are
QED
n r free variables.
d1
a1,1
a1,n
a2,1
a2,n d2
c1 . + + cn . = .
..
.. ..
am,1
am,n
dn
has a unique solution for all choices of d1 , . . . , dn R if and only if the vectors
QED
of as form a basis.
Exercises
3.16 Transpose each.
(a)
2
3
1
1
(b)
2
1
1
3
(c)
1
6
4
7
3
8
(d)
(e) 1 2
X 3.17 Decide if the vector is in the row space of the matrix.
0
1 3
2 1
(a)
, 1 0
(b) 1 0 1 , 1 1 1
3 1
1 2 7
X 3.18 Decide if the vector is in the column space.
0
0
0
130
! !
1
3
1
1
0
4 , 0
(a)
(b) 2
1 3 3
0
X 3.19 Find a basis for the row space of this matrix.
1
1
1
,
1
1
3
2
0
3
1
0
1
1
0
3
1
0
4
4
1
2
1
X 3.20 Find
of each matrix.
the rank !
!
2
1
3
1
1
2
1 3 2
3
3
6
(a) 1 1 2
(b)
(c) 5 1 1
1
0
3
2
2
4
6 4 3
!
0 0 0
(d) 0 0 0
0 0 0
X 3.21 Find
of each
a basis
for the
span
set.
(a) { 1 3 , 1 3 , 1 4 , 2 1 } M12
! ! !
1
3
1
(b) { 2 , 1 , 3 } R3
1
1
3
2
(c) {1+ x, 1 x2
,3 +
P3
2x x }
1 0
1
1 0 3
1
0
5
(d) {
,
,
} M23
3 1 1
2 1 4
1 1 9
3.22 Which matrices have rank zero? Rank one?
X 3.23 Given a, b, c R, what choice of d will cause this matrix to have the rank of
one?
a b
c d
3.24 Find the column rank of this matrix.
1 3 1 5
2 0
1
0
0
4
4
1
3.25 Show that a linear system with at least one solution has at most one solution if
and only if the matrix of coefficients has rank equal to the number of its columns.
X 3.26 If a matrix is 59, which set must be dependent, its set of rows or its set of
columns?
3.27 Give an example to show that, despite that they have the same dimension,
the row space and column space of a matrix need not be equal. Are they ever
equal?
3.28 Show that the set {(1, 1, 2, 3), (1, 1, 2, 0), (3, 1, 6, 6)} does not have the
same span as {(1, 0, 1, 0), (0, 2, 0, 3)}. What, by the way, is the vector space?
X 3.29 Show that this set of column vectors
( !
)
d1
3x + 2y + 4z = d1
d2
z = d2
there are x, y, and z such that x
d3
2x + 2y + 5z = d3
is a subspace of R3 . Find a basis.
131
2.III.4
Combining Subspaces
This subsection is optional. It is required only for the last sections of Chapter
Three and Chapter Five and for occasional exercises, and can be passed over
without loss of continuity.
This chapter opened with the definition of a vector space, and the middle consisted of a first analysis of the idea. This subsection closes the chapter
132
by finishing the analysis, in the sense that analysis means method of determining the . . . essential features of something by separating it into parts
[Macmillan Dictionary].
A common way to understand things is to see how they can be built from
component parts. For instance, we think of R3 as put together, in some way,
from the x-axis, the y-axis, and z-axis. In this subsection we will make this
precise; we will describe how to decompose a vector space into a combination of
some of its subspaces. In developing this idea of subspace combination, we will
keep the R3 example in mind as a benchmark model.
Subspaces are subsets and sets combine via union. But taking the combination operation for subspaces to be the simple union operation isnt what we
want. For one thing, the union of the x-axis, the y-axis, and z-axis is not all of
R3 , so the benchmark model would be left out. Besides, union is all wrong for
this reason: a union of subspaces need not be a subspace (it need not be closed;
for instance, this R3 vector
0
0
1
1
0 + 1 + 0 = 1
0
1
1
0
is in none of the three axes and hence is not in the union). In addition to the
members of the subspaces, we must at a minimum also include all possible linear
combinations.
4.1 Definition Where W1 , . . . , Wk are subspaces of a vector space, their sum
is the span of their union W1 + W2 + + Wk = [W1 W2 . . . Wk ].
(The notation, writing the + between sets in addition to using it between
vectors, fits with the practice of using this symbol for any natural accumulation
operation.)
~ R3 can
4.2 Example The R3 model fits with this operation. Any vector w
be written as a linear combination c1~v1 + c2~v2 + c3~v3 where ~v1 is a member of
the x-axis, etc., in this way
0
0
w1
w1
w2 = 1 0 + 1 w2 + 1 0
w3
0
w3
0
and so R3 = x-axis + y-axis + z-axis.
4.3 Example A sum of subspaces can be less than the entire space. Inside of
P4 , let L be the subspace of linear polynomials {a + bx a, b R} and let C be
L + C is not all
the subspace of purely-cubic polynomials {cx3 c R}. Then
3
of P4 . Instead, it is the subspace L + C = {a + bx + cx a, b, c R}.
4.4 Example A space can be described as a combination of subspaces in more
than one way. Besides the decomposition R3 = x-axis + y-axis + z-axis, we can
133
also write R3 = xy-plane + yz-plane. To check this, we simply note that any
w
~ R3 can be written
0
w1
w1
w2 = 1 w2 + 1 0
w3
0
w3
as a linear combination of a member of the xy-plane and a member of the
yz-plane.
The above definition gives one way in which a space can be thought of as a
combination of some of its parts. However, the prior example shows that there is
at least one interesting property of our benchmark model that is not captured by
the definition of the sum of subspaces. In the familiar decomposition of R3 , we
often speak of a vectors x part or y part or z part. That is, in this model,
each vector has a unique decomposition into parts that come from the parts
making up the whole space. But in the decomposition used in Example 4.4, we
cannot refer to the xy part of a vector these three sums
1
1
0
1
0
1
0
2 = 2 + 0 = 0 + 2 = 1 + 1
3
0
3
0
3
0
3
all describe the vector as comprised of something from the first plane plus something from the second plane, but the xy part is different in each.
That is, when we consider how R3 is put together from the three axes in
some way, we might mean in such a way that every vector has at least one
decomposition, and that leads to the definition above. But if we take it to
mean in such a way that every vector has one and only one decomposition
then we need another condition on combinations. To see what this condition
is, recall that vectors are uniquely represented in terms of a basis. We can use
this to break a space into a sum of subspaces such that any vector in the space
breaks uniquely into a sum of members of those subspaces.
4.5 Example The benchmark is R3 with its standard basis E3 = h~e1 , ~e2 , ~e3 i.
The subspace with the basis B1 = h~e1 i is the x-axis. The subspace with the
basis B2 = h~e2 i is the y-axis. The subspace with the basis B3 = h~e3 i is the
z-axis. The fact that any member of R3 is expressible as a sum of vectors from
these subspaces
x
x
0
0
y = 0 + y + 0
z
0
0
z
is a reflection of the fact that E3 spans the space this equation
1
0
0
x
y = c1 0 + c2 1 + c3 0
0
0
1
z
134
has a solution for any x, y, z R. And, the fact that each such expression is
unique reflects that fact that E3 is linearly independent any equation like the
one above has a unique solution.
4.6 Example We dont have to take the basis vectors one at a time, the same
idea works if we conglomerate them into larger sequences. Consider again the
space R3 and the vectors from the standard basis E3 . The subspace with the
basis B1 = h~e1 , ~e3 i is the xz-plane. The subspace with the basis B2 = h~e2 i is
the y-axis. As in the prior example, the fact that any member of the space is a
sum of members of the two subspaces in one and only one way
x
x
0
y = 0 + y
z
z
0
is a reflection of the fact that these vectors form a basis this system
1
0
0
x
y = (c1 0 + c3 0) + c2 1
0
1
0
z
has one and only one solution for any x, y, z R.
These examples illustrate a natural way to decompose a space into a sum
of subspaces in such a way that each vector decomposes uniquely into a sum of
vectors from the parts. The next result says that this way is the only way.
~1,1 , . . . , ~1,n i, . . . ,
4.7 Definition The concatenation of the sequences B1 = h
1
~
~
Bk = hk,1 , . . . , k,nk i is their adjoinment.
B1
B2
_
~1,1 , . . . ,
~1,n , ~2,1 , . . . , ~k,n i
Bk = h
1
k
4.8 Lemma Let V be a vector space that is the sum of some of its subspaces
V = W1 + + Wk . Let B1 , . . . , Bk be any bases for these subspaces. Then
the following are equivalent.
~ k (with w
~ i Wi ) is
(1) For every ~v V , the expression ~v = w
~1 + + w
unique.
_
_
(2) The concatenation B1 Bk is a basis for V .
~ k } (with w
~ i Wi ) form a linearly
(3) The nonzero members of {w
~ 1, . . . , w
independent set among nonzero vectors from different Wi s, every linear
relationship is trivial.
Proof. We will show that (1) = (2), that (2) = (3), and finally that
(3) = (1). For these arguments, observe that we can pass from a combination
~
of ws
~ to a combination of s
~ 1 + + dk w
~k
d1 w
~1,1 + + c1,n
~
~k,1 + + ck,n
~
= d1 (c1,1
) + + dk (ck,1
)
1 1,n1
k k,nk
~1,1 + + dk ck,n
~k,n
= d1 c1,1
k
()
135
136
0 0
0 b
a 0
c R}
b R} {
a, d R} {
{
c 0
0 0
0 d
It is the direct sum of subspaces in many other ways as well; direct sum decompositions are not unique.
4.13 Corollary The dimension of a direct sum is the sum of the dimensions
of its summands.
Proof. In Lemma 4.8, the number of basis vectors in the concatenation equals
the sum of the number of vectors in the subbases that make up the concatenaQED
tion.
4.16 Example In the space R2 , the x-axis and the y-axis are complements,
that is, R2 = x-axis y-axis. A space can have more than one pair of complementary subspaces; another pair here are the subspaces consisting of the lines
y = x and y = 2x.
4.17 Example
In the space F = {a cos
+ b sin a, b R}, the subspaces
137
4.18 Example In R3 , the xy-plane and the yz-planes are not complements,
which is the point of the discussion following Example 4.4. One complement of
the xy-plane is the z-axis. A complement of the yz-plane is the line through
(1, 1, 1).
4.19 Example Following Lemma 4.15, here is a natural question: is the simple
sum V = W1 + + Wk also a direct sum if and only if the intersection of the
subspaces is trivial? The answer is that if there are more than two subspaces
then having a trivial intersection is not enough to guarantee unique decomposition (i.e., is not enough to ensure that the spaces are independent). In R3 , let
W1 be the x-axis, let W2 be the y-axis, and let W3 be this.
q
W3 = {q q, r R}
r
The check that R3 = W1 + W2 + W3 is easy. The intersection W1 W2 W3 is
trivial, but decompositions arent unique.
x
0
0
x
xy
0
y
y = 0 + y x + x = 0 + 0 + y
z
0
0
z
0
0
z
(This example also shows that this requirement is also not enough: that all
pairwise intersections of the subspaces be trivial. See Exercise 30.)
In this subsection we have seen two ways to regard a space as built up from
component parts. Both are useful; in particular, in this book the direct sum
definition is needed to do the Jordan Form construction in the fifth chapter.
Exercises
X 4.20 Decide ifR2is the direct sum of
each
W1 and W2 .
x
x
(a) W1 = {
x R}, W2 = {
x R}
0
x
s
(b) W1 = {
s
s R}, W2 = { s
s R}
1.1s
(c) W1 = R , W2
= {~0}
t
(d) W1 = W2 = {
t R}
t
2
(e) W1 = {
1
0
x
0
x R}, W2 = { 1 + 0 y R}
0
X 4.21 Show that R3 is the direct sum of the xy-plane with each of these.
(a) the z-axis
(b) the line
z
{ z
z
z R}
138
0 0
0 b
W1 = {
c, d R}
W2 = {
b, c R}
c d
c 0
139
(b) Suppose that U and W are subspaces of a vector space. Suppose that the
~1 , . . . ,
~k i is a basis for U W . Finally, suppose that the prior
sequence h
~1 , . . . ,
~k i that is a
sequence has been expanded to give a sequence h~
1 , . . . ,
~j,
~
~
basis for U , and a sequence h1 , . . . , k ,
~ 1, . . . ,
~ p i that is a basis for W . Prove
that this sequence
~1 , . . . ,
~k ,
h~
1 , . . . ,
~j,
~ 1, . . . ,
~ pi
is a basis for for the sum U + W .
(c) Conclude that dim(U + W ) = dim(U ) + dim(W ) dim(U W ).
(d) Let W1 and W2 be eight-dimensional subspaces of a ten-dimensional space.
List all values possible for dim(W1 W2 ).
4.36 Let V = W1 . . . Wk and for each index i suppose that Si is a linearly
independent subset of Wi . Prove that the union of the Si s is linearly independent.
4.37 A matrix is symmetric if for each pair of indices i and j, the i, j entry equals
the j, i entry. A matrix is antisymmetric if each i, j entry is the negative of the j, i
entry.
(a) Give a symmetric 22 matrix and an antisymmetric 22 matrix. (Remark.
For the second one, be careful about the entries on the diagional.)
(b) What is the relationship between a square symmetric matrix and its transpose? Between a square antisymmetric matrix and its transpose?
(c) Show that Mnn is the direct sum of the space of symmetric matrices and
the space of antisymmetric matrices.
4.38 Let W1 , W2 , W3 be subspaces of a vector space. Prove that (W1 W2 ) + (W1
W3 ) W1 (W2 + W3 ). Does the inclusion reverse?
4.39 The example of the x-axis and the y-axis in R2 shows that W1 W2 = V does
not imply that W1 W2 = V . Can W1 W2 = V and W1 W2 = V happen?
X 4.40 Our model for complementary subspaces, the x-axis and the y-axis in R2 ,
has one property not used here. Where U is a subspace of Rn we define the
orthocomplement of U to be
U = {~v Rn ~v ~
u = 0 for all ~
u U}
(read U perp).
(a) Find the orthocomplement of the x-axis in R2 .
(b) Find the orthocomplement of the x-axis in R3 .
(c) Find the orthocomplement of the xy-plane in R3 .
(d) Show that the orthocomplement of a subspace is a subspace.
(e) Show that if W is the orthocomplement of U then U is the orthocomplement
of W .
(f ) Prove that a subspace and its orthocomplement have a trivial intersection.
(g) Conclude that for any n and subspace U Rn we have that Rn = U U .
(h) Show that dim(U ) + dim(U ) equals the dimension of the enclosing space.
X 4.41 Consider Corollary 4.13. Does it work both ways that is, supposing that
V = W1 + + Wk , is V = W1 . . . Wk if and only if dim(V ) = dim(W1 ) +
+ dim(Wk )?
4.42 We know that if V = W1 W2 then there is a basis for V that splits into a
basis for W1 and a basis for W2 . Can we make the stronger statement that every
basis for V splits into a basis for W1 and a basis for W2 ?
4.43 We can ask about the algebra of the + operation.
(a) Is it commutative; is W1 + W2 = W2 + W1 ?
(b) Is it associative; is (W1 + W2 ) + W3 = W1 + (W2 + W3 )?
140
Topic: Fields
141
Topic: Fields
Linear combinations involving only fractions or only integers are much easier
for computations than combinations involving real numbers, because computing
with irrational numbers is awkward. Could other number systems, like the
rationals or the integers, work in the place of R in the definition of a vector
space?
Yes and no. If we take work to mean that the results of this chapter
remain true then an analysis of which properties of the reals we have used in
this chapter gives the following list of conditions an algebraic system needs in
order to work in the place of R.
Definition. A field is a set F with two operations + and such that
(1) for any a, b F the result of a + b is in F and
a+b=b+a
if c F then a + (b + c) = (a + b) + c
(2) for any a, b F the result of a b is in F and
ab=ba
if c F then a (b c) = (a b) c
(3) if a, b, c F then a (b + c) = a b + a c
(4) there is an element 0 F such that
if a F then a + 0 = a
for each a F there is an element a F such that (a) + a = 0
(5) there is an element 1 F such that
if a F then a 1 = a
for each non-0 element a F there is an element a1 F such that
a1 a = 1.
The number system comsisting of the set of real numbers along with the
usual addition and multiplication operation is a field, naturally. Another field is
the set of rational numbers with its usual addition and multiplication operations.
An example of an algebraic structure that is not a field is the integer number
systemit fails the final condition.
Some examples are surprising. The set {0, 1} under these operations:
+
0
1
is a field (see Exercise 4).
0
0
1
1
1
0
0
1
0
0
0
1
0
1
142
We could develop Linear Algebra as the theory of vector spaces with scalars
from an arbitrary field, instead of sticking to taking the scalars only from R. In
that case, almost all of the statements in this book would carry over by replacing
R with F, and thus by taking coefficients, vector entries, and matrix entries
to be elements of F. (This says almost all because statements involving
distances or angles are exceptions.) Here are some examples; each applies to a
vector space V over a field F.
For any ~v V and a F, (i) 0 ~v = ~0, and (ii) 1 ~v + ~v = ~0, and
(iii) a ~0 = ~0.
The span (the set of linear combinations) of a subset of V is a subspace
of V .
Any subset of a linearly independent set is also linearly independent.
In a finite-dimensional vector space, any two bases have the same number
of elements.
(Even statements that dont explicitly mention F use field properties in their
proof.)
We wont develop vector spaces in this more general setting because the
additional abstraction can be a distraction. The ideas we want to bring out
already appear when we stick to the reals.
The only exception is in Chapter Five. In that chapter we must factor
polynomials, so we will switch to considering vector spaces over the field of
complex numbers. We will discuss this more, including a brief review of complex
arithmetic, when we get there.
Exercises
1 Show that the real numbers form a field.
2 Prove that these are fields:
(a) the rational numbers
(b) the complex numbers.
3 Give an example that shows that the integer number system is not a field.
4 Consider the set {0, 1} subject to the operations given above. Show that it is a
field.
5 Come up with suitable operations to make the set {0, 1, 2} a field.
Topic: Crystals
143
Topic: Crystals
Everyone has noticed that table salt comes in little cubes.
Remarkably, the explanation for the cubical external shape is the simplest one
possible: the internal shape, the way the atoms lie, is also cubical. The internal
structure is pictured below. Salt is sodium cloride, and the small spheres shown
are sodium while the big ones are cloride. (To simplify the view, only the
sodiums and clorides on the front, top, and right are shown.)
The specks of salt that we see when we spread a little out on the table consist of
many repetitions of this fundamental unit. That is, these cubes of atoms stack
up to make the larger cubical structure that we see. A solid, such as table salt,
with a regular internal structure is a crystal.
We can restrict our attention to the front face. There, we have this pattern
repeated many times.
ngstroms (an
The distance between the corners of this cell is about 3.34 A
ngstrom is 1010 meters). Obviously that unit is unwieldly for describing
A
points in the crystal lattice. Instead, the thing to do is to take as a unit the
length of each side of the square. That is, we naturally adopt this basis.
3.34
0
h
,
i
0
3.34
144
Then we can describe, say, the corner in the upper right of the picture above as
~2 .
~1 + 2
3
Another crystal from everyday experience is pencil lead. It is graphite,
formed from carbon atoms arranged in this shape.
A natural basis then would consist of the vectors that form the sides of that
unit cell. The distance along the bottom and slant is 1.42
Angstroms, so this
1.42
1.23
h
,
i
0
.71
is a good basis.
The selection of convienent bases extends to three dimensions. Another
familiar crystal formed from carbon is diamond. Like table salt, it is built from
cubes, but the structure inside each cube is more complicated than salts. In
addition to carbons at each corner,
Topic: Crystals
145
(To show the added face carbons clearly, the corner carbons have been reduced
to dots.) There are also four more carbons inside the cube, two that are a
quarter of the way up from the bottom and two that are a quarter of the way
down from the top.
(As before, carbons shown earlier have been reduced here to dots.) The distance along any edge of the cube is 2.18
Angstroms. Thus, a natural basis for
describing the locations of the carbons, and the bonds between them, is this.
2.18
0
0
h 0 , 2.18 , 0 i
0
0
2.18
Even the few examples given here show that the structures of crystals is complicated enough that some organized system to give the locations of the atoms,
and how they are chemically bound, is needed. One tool for that organization
is a convienent basis. This application of bases is simple, but it shows a context
where the idea arises naturally. The work in this chapter just takes this simple
idea and develops it.
Exercises
1 How many fundamental regions are there in one face of a speck of salt? (With a
ruler, we can estimate that face is a square that is 0.1 cm on a side.)
2 In the graphite picture, imagine that we are interested in a point 5.67
Angstroms
up and 3.14
Angstroms over from the origin.
(a) Express that point in terms of the basis given for graphite.
(b) How many hexagonal shapes away is this point from the origin?
(c) Express that point in terms of a second basis, where the first basis vector is
the same, but the second is perpendicular to the first (going up the plane) and
of the same length.
3 Give the locations of the atoms in the diamond cube both in terms of the basis,
and in
Angstroms.
4 This illustrates how the dimensions of a unit cell could be computed from the
shape in which a substance crystalizes ([Ebbing], p. 462).
(a) Recall that there are 6.0221023 atoms in a mole (this is Avagadros number).
From that, and the fact that platinum has a mass of 195.08 grams per mole,
calculate the mass of each atom.
(b) Platinum crystalizes in a face-centered cubic lattice with atoms at each lattice
point, that is, it looks like the middle picture given above for the diamond crystal.
Find the number of platinums per unit cell (hint: sum the fractions of platinums
that are inside of a single cell).
(c) From that, find the mass of a unit cell.
146
147
number with
that preference
5
4
2
8
8
2
29
Third
5 voters
Republican
1 voter
148
however, linear algebra has been used [Zwicker] to argue that a tendency toward
cyclic preference is actually present in each voters list, and that it surfaces when
there is more adding of the tendency than cancelling.
For this argument, abbreviating the choices as D, R, and T , we can describe
how a voter with preference order D > R > T contributes to the above cycle
D
1 voter
1 voter
1 voter
(the negative sign is here because the arrow describes T as preferred to D, but
this voter likes them the other way). The descriptions for the other preference
lists are in the table on page 150. Now, to conduct the election, we linearly
combine these descriptions; for instance, the Political Science mock election
5
1 voter
1 voter
1 voter
+4
1 voter
1 voter
1 voter
+ + 2
1 voter
1 voter
1 voter
149
(Read that aloud as C perp.) Consideration of those two has led to this basis
of R3 .
1
1
1
h1 , 1 , 0 i
1
0
1
We can represent votes with respect to this basis, and thereby decompose them
into a cyclic part and an acyclic part. (Note for readers who have covered the
optional section: that is, the space is the direct sum of C and C .)
For example, consider the D > R > T voter discussed above. The representation in terms of the basis is easily found,
c1 c2 c3 = 1
= 1
c1 + c2
+ c3 = 1
c1
1 +2 (1/2)2 +3
1 +3
c1 c2
2c2 +
c3 = 1
c3 = 2
(3/2)c3 = 1
1
1
1
1/3
4/3
1
2
2
1
1 = 1 + 1 + 0 = 1/3 + 2/3
3
3
3
1
0
1
1/3
2/3
1
gives the desired decomposition into a cyclic part and and an acyclic part.
1
1/3
4/3
1/3
1/3
2/3
R
T
2/3
Thus, this D > R > T voters rational preference list can indeed be seen to
have a cyclic part.
The T > R > D voter is opposite to the one just considered in that the >
symbols are reversed. This voters decomposition
1
T
1
1/3
1/3
R
1/3
4/3
2/3
R
2/3
shows that these opposite preferences have decompositions that are opposite.
We say that the first voter has positive spin since the cycle part is with the
direction we have chosen for the arrows, while the second voters spin is negative.
The fact that that these opposite voters cancel each other is reflected in the
fact that their vote vectors add to zero. This suggests an alternate way to tally
an election. We could first cancel as many opposite preference lists as possible,
and then determine the outcome by adding the remaining lists.
The rows of the table below contain the three pairs of opposite preference
lists, and the columns group those pairs by spin. For instance, the first row
contains the two voters just considered.
150
negative spin
1/3 D 1/3
R
T
1/3
4/3 D 2/3
R
T
2/3
1/3 D 1/3
R
T
1/3
2/3 D 4/3
R
T
2/3
1 D 1
R
T
1
1/3 D 1/3
R
T
1/3
2/3 D 2/3
R
T
4/3
If we conduct the election as just described then after the cancellation of as many
opposite pairs of voters as possible, there will be left three sets of preference
lists, one set from the first row, one set from the second row, and one set from
the third row. We will finish by proving that a voting paradox can happen
only if the spins of these three sets are in the same direction. That is, for a
voting paradox to occur, the three remaining sets must all come from the left
of the table or all come from the right (see Exercise 3). This shows that there
is some connection between the majority cycle and the decomposition that we
are usinga voting paradox can happen only when the tendencies toward cyclic
preference reinforce each other.
For the proof, assume that opposite preference orders have been cancelled,
and we are left with one set of preference lists from each of the three rows.
Consider the sum of these three (here, a, b, and c could be positive, negative,
or zero).
a
a
R
b
R
T
c
a + b + c
ab+c
R
a+bc
A voting paradox occurs when the three numbers on the right, a b + c and
a + b c and a + b + c, are all nonnegative or all nonpositive. On the left,
at least two of the three numbers, a and b and c, are both nonnegative or both
nonpositive. We can assume that they are a and b. That makes four cases: the
cycle is nonnegative and a and b are nonnegative, the cycle is nonpositive and
a and b are nonpositive, etc. We will do only the first case, since the second is
similar and the other two are also easy.
So assume that the cycle is nonnegative and that a and b are nonnegative.
The conditions 0 a b + c and 0 a + b + c add to give that 0 2c, which
implies that c is also nonnegative, as desired. That ends the proof.
This result only says that having all three spin in the same direction is a
necessary condition for a majority cycle. It is not also a sufficient condition; see
Exercise 4.
151
Voting theory and associated topics are the subject of current research. The
are many surprising and intriguing results, most notably the one produced by
K. Arrow [Arrow], who won the Nobel Prize in part for this work, showing, essentially, that no voting system is entirely fair. For more information, some good
introductory articles are [Gardner, 1970], [Gardner, 1974], [Gardner, 1980], and
[Neimi & Riker]. A quite readable recent book is [Taylor]. The material of this
Topic is largely drawn from [Zwicker]. (Authors Note: I would like to thank
Professor Zwicker for his kind and illuminating discussions.)
Exercises
1 Here is a reasonable way in which a voter could have a cyclic preference. Suppose
that this voter ranks each candidate on each of three criteria.
(a) Draw up a table with the rows labelled Democrat, Republican, and Third,
and the columns labelled character, experience, and policies. Inside each
column, rank some candidate as most preferred, rank another as in the middle,
and rank the remaining one as least preferred.
(b) In this ranking, is the Democrat preferred to the Republican in (at least) two
out of three criteria, or vice versa? Is the Republican preferred to the Third?
(c) Does the table that was just constructed have a cyclic preference order? If
not, make one that does.
So it is possible for a voter to have a cyclic preference among candidates. The
paradox described above, however, is that even if each voter has a straight-line
preference list, there can still be a cyclic group preference.
2 Compute the values in the table of decompositions.
3 Do the cancellations of opposite preference orders for the Political Science classs
mock election. Are all the remaining preferences from the left three rows of the
table or from the right?
4 The necessary condition that is proved abovea voting paradox can happen only
if all three preference lists remaining after cancellation have the same spinis not
also sufficient.
(a) Continuing the positive cycle case considered in the proof, use the two inequalities 0 a b + c and 0 a + b + c to show that |a b| c.
(b) Also show that c a + b, and hence that |a b| c a + b.
(c) Give an example of a vote where there is a majority cycle, and addition of
one more voter with the same spin causes the cycle to go away.
(d) Can the opposite happen; can addition of one voter with a wrong spin
cause a cycle to appear?
(e) Give a condition that is both necessary and sufficient to get a majority cycle.
5 A one-voter election cannot have a majority cycle because of the requirement
that weve imposed that the voters list must be rational.
(a) Show that a two-voter election may have a majority cycle. (We consider the
group preference a majority cycle if all three group totals are nonnegative or if
all three are nonpositivethat is, we allow some zeros in the group preference.)
(b) Show that for any number of voters greater than one, there is an election
involving that many voters that results in a majority cycle.
152
min
hr
day
sec
sec
60
24
365
= 31 536 000
min
hr
day
year
year
However, the idea of paying attention to how the quantities are measured can
be pushed beyond bookkeeping. It can be used to draw conclusions about the
nature of relationships among physical quantities.
Consider this equation expressing a relationship: dist = 16 (time)2 . If
distance is taken in feet and time in seconds then this is a true statement about
the motion of a falling body. But this equation is a correct description only in
the foot-second unit system. In the yard-second unit system it is not the case
that d = 16t2 . To get a complete equationone that holds irrespective of the
size of the unitswe will make the 16 a dimensional constant.
dist = 16
ft
(time)2
sec2
Now, the equation holds in any units system, e.g., in yards and seconds we have
this.
dist in yd = 16
16 yd
(1/3) yd
(time in sec)2 =
(time in sec)2
2
sec
3 sec2
153
arc
rad
This is the ratio of a length to a length L1 M 0 T 0 /L1 M 0 T 0 and thus angles have
the dimensional formula L0 M 0 T 0 .
Paying attention to the dimensional formulas of the physical quantities will
help us to see which relationships are possible or impossible among the quantities. For instance, suppose that we want to give the period of a pendulum as
some formula p = involving the other relevant physical quantities, length of
the string, etc. (see the table on page 154). The period is expressed in units of
timeit has dimensional formula L0 M 0 T 1 and so the quantities on the other
side of the equation must have their dimensional formulas combine in such a
way that the Ls and M s cancel and only a T is left. For instance, in that table,
the only quantities involving L are the length of the string and the acceleration
due to gravity. For these Ls to cancel, the quantities must enter the equation
in ratio, e.g., as (`/g)2 or as cos(`/g), or as (`/g)1 . In this way, simply from
consideration of the dimensional formulas, we know that that the period can be
written as a function of `/g; the formula cannot possibly involve, say, `3 and
g 2 because the dimensional formulas wouldnt cancel their Ls.
To do dimensional analysis systematically, we need two results (for proofs,
see [Bridgman], Chapter II and IV). First, each equation relating physical quantities that we shall see involves a sum of terms, where each term has the form
mp11 mp22 . . . mpkk
for numbers m1 , . . . , mk that measure the quantities.
Next, observe that an easy way to construct a dimensionally homogeneous
expression is by taking a product of dimensionless quantities, or by adding
such dimensionless terms. The second result, Buckinghams Theorem, is that
any complete relationship among quantities with dimensional formulas can be
algebraically manipulated into a form where there is some function f such that
f (1 , . . . , n ) = 0
for a complete set {1 , . . . , n } of dimensionless products. (We shall see what
makes a set of dimensionless products complete in the examples below.) We
usually want to express one of the quantities, m1 for instance, in terms of the
others, and for that we will assume that the above equality can be rewritten
2
k
. . . mp
f(2 , . . . , n )
m1 = mp
2
k
154
dimensional
formula
L0 M 0 T 1
L1 M 0 T 0
L0 M 1 T 0
L1 M 0 T 2
L0 M 0 T 0
To find which combinations of the powers in pp1 `p2 mp3 g p4 p5 yield dimensionless
products, consider this equation.
(L0 M 0 T 1 )p1 (L1 M 0 T 0 )p2 (L0 M 1 T 0 )p3 (L1 M 0 T 2 )p4 (L0 M 0 T 0 )p5 = L0 M 0 T 0
It gives three conditions on the powers.
p2
p3
p1
+ p4 = 0
=0
2p4 = 0
Note that p3 is 0the mass of the bob does not affect the period. The systems
solution space can be described in this way (p1 is taken as one of the parameters
in order to express the period in terms of the other quantities).
0
1
p1
0
p2 1/2
{
p3 = 0 p1 + 0 p5 p1 , p5 R}
0
p4 1/2
p5
1
0
Here is the linear algebra. The set of dimensionless products is the set of
products pp1 `p2 mp3 ap4 p5 subject to the conditions in the above linear system.
This forms a vector space under the + addition operation of multiplying two
such products and the scalar multiplication operation of raising such a product to the power of the scalar (see Exercise 5). The term complete set of
dimensionless products in Buckinghams Theorem means a basis for this vector
space.
We can get a basis by first taking p1 = 1 and p5 = 0, and then taking p1 = 0
and p5 = 1. The associated dimensionless products are 1 = p`1/2 g 1/2 and
2 = . The set {1 , 2 } is complete, so we have
p = `1/2 g 1/2 f()
p
= `/g f()
155
where f is a function that we cannot determine from this analysis (by other
means we know that for small angles it is approximately the constant function
f() = 2).
Thus, analysis of the relationships that are possible between the quantities
with the given dimensional formulas has given us a fair amount of information: a
pendulums period does not depend on the mass of the bob, and it rises with
the square root of the length of the string.
For the next example we try to determine the period of revolution of two
bodies in space orbiting each other under mutual gravitational attraction. An
experienced investigator could expect that these are the relevant quantities.
quantity
period of revolution p
mean radius of separation r
mass of the first m1
mass of the second m2
gravitational constant G
dimensional
formula
L0 M 0 T 1
L1 M 0 T 0
L0 M 1 T 0
L0 M 1 T 0
L3 M 1 T 2
+ 3p5 = 0
p 3 + p 4 p5 = 0
2p5 = 0
0
1
0
3/2
{
1/2 p1 + 1 p4 p1 , p4 R}
1
0
0
1/2
(p1 is taken as a parameter so that we can state the period as a function of the
other quantities). As with the pendulum example, the linear algebra here is
that the set of dimensionless products of these quantities forms a vector space,
and we want to produce a basis for that space, a complete set of dimensionless
products. One such set, gotten from setting p1 = 1 and p4 = 0, and also
1/2
setting p1 = 0 and p4 = 1 is {1 = pr3/2 m1 G1/2 , 2 = m1
1 m2 }. With
that, Buckinghams Theorem says that any complete relationship among these
quantities must be stateable this form.
1/2
p = r3/2 m1
G1/2 f(m1
1 m2 )
r3/2
f(m2 /m1 )
=
Gm1
156
dimensional
formula
L1 M 0 T 1
L3 M 1 T 0
L1 M 0 T 2
L1 M 0 T 0
Considering
(L1 M 0 T 1 )p1 (L3 M 1 T 0 )p2 (L1 M 0 T 2 )p3 (L1 M 0 T 0 )p4 = L0 M 0 T 0
gives this system
p1 3p2 + p3 + p4 = 0
=0
p2
2p3
=0
p1
with this solution space
1
0
{
1/2 p1 p1 R}
1/2
(as in the pendulum example, one of the quantities d turns out not to be involved in the relationship). There is thus one dimensionless product, 1 =
157
the relationship among the quantities. For further reading, the classic reference is [Bridgman]this brief book is a delight to read. Another source is
[Giordano, Wells, Wilde]. A description of how dimensional analysis fits into
the process of mathematical modeling is [Giordano, Jaye, Weir].
Exercises
1 [de Mestre] Consider a projectile, launched with initial velocity v0 , at an angle
. An investigation of this motion might start with the guess that these are the
relevant quantities.
dimensional
quantity formula
horizontal position x L1 M 0 T 0
vertical position y L1 M 0 T 0
initial speed v0 L1 M 0 T 1
angle of launch L0 M 0 T 0
acceleration due to gravity g L1 M 0 T 2
time t L0 M 0 T 1
(a) Show that {gt/v0 , gx/v02 , gy/v02 , } is a complete set of dimensionless products. (Hint. This can be done by finding the appropriate free variables in the
linear system that arises, but there is a shortcut that uses the properties of a
basis.)
(b) These two equations of motion for projectiles are familiar: x = v0 cos()t and
y = v0 sin()t(g/2)t2 . Algebraic manipulate each to rewrite it as a relationship
among the dimensionless products of the prior item.
2 [Einstein] conjectured that the infrared characteristic frequencies of a solid may
be determined by the same forces between atoms as determine the solids ordanary
elastic behavior. The relevant quantities are
dimensional
quantity formula
characteristic frequency L0 M 0 T 1
compressibility k L1 M 1 T 2
number of atoms per cubic cm N L3 M 0 T 0
mass of an atom m L0 M 1 T 0
Show that there is one dimensionless product. Conclude that, in any complete
relationship among quantities with these dimensional formulas, k is a constant
times 2 N 1/3 m1 . This conclusion played an important role in the early study
of quantum phenomena.
3 [Giordano, Wells, Wilde] Consider the torque produced by an engine. Torque
has dimensional formula L2 M 1 T 2 . We may first guess that it depends on the
engines rotation rate (with dimensional formula L0 M 0 T 1 ), and the volume of
air displaced (with dimensional formula L3 M 0 T 0 ).
(a) Try to find a complete set of dimensionless products. What goes wrong?
(b) Adjust the guess by adding the density of the air (with dimensional formula
L3 M 1 T 0 ). Now find a complete set of dimensionless products.
4 [Tilley] Dominoes falling make a wave. We may conjecture that the wave speed v
depends on the the spacing d between the dominoes, the height h of each domino,
and the acceleration due to gravity g.
(a) Find the dimensional formula for each of the four quantities.
158
p
{mp11 . . . mkk p1 , . . . , pk satisfy the system}
p +qk
rpk
1
r~(mp11 . . . mkk ) = mrp
. . . mk
1
Chapter 3
Isomorphisms
3.I.1
1
1 2
2
159
160
then this correspondence preserves the operations, for instance this addition
2 + 3 4 = 4 6
1
3
4
+
=
2
4
6
5 1 2 = 5
10
1
5
5
=
2
10
a0
a1
a0
a1
a0
a1 + b0
b1 = a0 + b0
a1 + b1
a0
b
a0 + b0
+ 0 =
a1
b1
a1 + b1
and
r a0
a1 = ra0
ra1
a0
ra0
=
a1
ra1
a0
a1
a2
1
(e.g., 1 + 2x + 3x2 2)
3
a0
b0
a0 + b0
a1 + b1 = a1 + b1
a2
b2
a2 + b2
a0
ra0
r a1 = ra1
a2
ra2
Section I. Isomorphisms
161
1.4 Example The vector space G = {c1 cos + c2 sin c1 , c2 R} of functions of is isomorphic to the vector space R2 under this map.
f
c
c1 cos + c2 sin 7 1
c2
We will check this by going through the conditions in the definition.
We will first verify condition (1), that the map is a correspondence between
the sets underlying the spaces.
To establish that f is one-to-one, we must prove that f (~a) = f (~b) only when
~a = ~b. If
f (a1 cos + a2 sin ) = f (b1 cos + b2 sin )
then, by the definition of f ,
b
a1
= 1
a2
b2
from which we can conclude that a1 = b1 and a2 = b2 because column vectors are
equal only when they have equal components. Weve proved that f (~a) = f (~b)
implies that ~a = ~b, which shows that f is one-to-one.
To check that f is onto we must check that any member of the codomain R2
mapped to. But thats clearany
x
R2
y
is the image, under f , of this member of the domain: x cos + y sin G.
Next we will verify condition (2), that f preserves structure.
More
162
a1 + b1
=
a2 + b2
b
a1
+ 1
=
a2
b2
= f (a1 cos + a2 sin ) + f (b1 cos + b2 sin )
A similar computation shows that f preserves scalar multiplication.
c1 x + c2 y + c3 z
f1
c1 + c2 x + c3 x2
f2
c2 + c3 x + c1 x2
f3
c1 c2 x c3 x2
f4
7
7
7
Although the first map is the more natural correspondence, below we shall
verify that the second one is an isomorphism, to underline that there are many
isomorphisms other than the obvious one that just carries the coefficients over
(showing that f1 is an isomorphism is Exercise 12).
To show that f2 is one-to-one, we will prove that if f2 (c1 x + c2 y + c3 z) =
f2 (d1 x + d2 y + d3 z) then c1 x + c2 y + c3 z = d1 x + d2 y + d3 z. The assumption
that f2 (c1 x + c2 y + c3 z) = f2 (d1 x + d2 y + d3 z) gives, by the definition of f2 , that
c2 + c3 x + c1 x2 = d2 + d3 + d1 x2 . Equal polynomials have equal coefficients, so
c2 = d2 , c3 = d3 , and c1 = d1 . Thus f2 (c1 x + c2 y + c3 z) = f2 (d1 x + d2 y + d3 z)
implies that c1 x + c2 y + c3 z = d1 x + d2 y + d3 z and therefore f2 is one-to-one.
Section I. Isomorphisms
163
The map f2 is onto because any member a + bx + cx2 of the codomain is the
image of some member of the domain, namely it is the image of cx + ay + bz.
(For instance, 2 + 3x 4x2 is f2 (4x + 2y + 3z).)
The computations for structure preservation for this map are like those in
the prior example. This map preserves addition
f2 (c1 x + c2 y + c3 z) + (d1 x + d2 y + d3 z)
p0 (x)
p1 (x)
164
71.5
~
u
d1.5 (~
v)
~
v
~
v
`
7
f` (~
v)
Section I. Isomorphisms
165
166
by breaking the sum along the final +. Then the inductive hypothesis lets us
break up the sum of the k things.
= f (c1~v1 ) + + f (ck~vk ) + f (ck+1~vk+1 )
Finally, the second half of statement (1) gives
= c1 f (~v1 ) + + ck f (~vk ) + ck+1 f (~vk+1 )
when applied k + 1 times.
QED
f
ab
a + bx 7
b
Find the image of each of these elements of the domain.
(a) 3 2x
(b) 2 + 2x
(c) x
Show that this map is an isomorphism.
1.12 Show that the natural map f1 from Example 1.5 is an isomorphism.
X 1.13 Decide whether each map is an isomorphism (of course, if it is an isomorphism
then prove it and if it isnt then state a condition that it fails to satisfy).
(a) f : M22 R given by
a
c
a
c
b
d
b
d
7 ad bc
a+b+c+d
a+b+c
7
a+b
a
a
c
b
d
Section I. Isomorphisms
167
a
c
b
d
1.14 Show that the map f : R1 R1 given by f (x) = x3 is one-to-one and onto.
Is it an isomorphism?
1.15 Refer to Example 1.1. Produce two more isomorphisms (of course, that they
satisfy the conditions in the definition of isomorphism must be verified).
1.16 Refer to Example 1.2. Produce two more isomorphisms (and verify that they
satisfy the conditions).
1.17 Show that, although R2 is not itself a subspace of R3 , it is isomorphic to the
xy-plane subspace of R3 .
1.18 Find two isomorphisms between R16 and M44 .
1.19 For what k is Mmn isomorphic to Rk ?
1.20 For what k is Pk isomorphic to Rn ?
1.21 Prove that the map in Example 1.6, from P5 to P5 given by p(x) 7 p(x 1),
is a vector space isomorphism.
1.22 Why, in Lemma 1.8, must there be a ~v V ? That is, why must V be
nonempty?
1.23 Are any two trivial spaces isomorphic?
1.24 In the proof of Lemma 1.9, what about the zero-summands case (that is, if n
is zero)?
1.25 Show that any isomorphism f : P0 R1 has the form a 7 ka for some nonzero
real number k.
1.26 These prove that isomorphism is an equivalence relation.
(a) Show that the identity map id : V V is an isomorphism. Thus, any vector
space is isomorphic to itself.
(b) Show that if f : V W is an isomorphism then so is its inverse f 1 : W V .
Thus, if V is isomorphic to W then also W is isomorphic to V .
(c) Show that a composition of isomorphisms is an isomorphism: if f : V W is
an isomorphism and g : W U is an isomorphism then so also is g f : V U .
Thus, if V is isomorphic to W and W is isomorphic to U , then also V is isomorphic to U .
1.27 Suppose that f : V W preserves structure. Show that f is one-to-one if and
only if the unique member of V mapped by f to ~0W is ~0V .
1.28 Suppose that f : V W is an isomorphism. Prove that the set {~v1 , . . . , ~vk }
V is linearly dependent if and only if the set of images {f (~v1 ), . . . , f (~vk )} W is
linearly dependent.
1.29 Show that each type of map from Example 1.7 is an automorphism.
(a) Dilation ds by a nonzero scalar s.
(b) Rotation t through an angle .
(c) Reflection f` over a line through the origin.
Hint. For the second and third items, polar coordinates are useful.
1.30 Produce an automorphism of P2 other than the identity map, and other than
a shift map p(x) 7 p(x k).
1.31 (a) Show that a function f : R1 R1 is an automorphism if and only if it
has the form x 7 kx for some k 6= 0.
(b) Let f be an automorphism of R1 such that f (3) = 7. Find f (2).
168
x
ax + by
7
y
cx + dy
for some a, b, c, d R with ad bc 6= 0. Hint. Exercises in prior subsections
have shown that
b
a
is not a multiple of
d
c
if and only if ad bc 6= 0.
(d) Let f be an automorphism of R2 with
f(
1
)=
3
2
1
and
Find
f(
1
)=
4
0
.
1
0
f(
).
1
1.32 Refer to Lemma 1.8 and Lemma 1.9. Find two more things preserved by
isomorphism.
1.33 We show that isomorphisms can be tailored to fit in that, sometimes, given
vectors in the domain and in the range we can produce an isomorphism associating
those vectors.
~1 ,
~2 ,
~3 i be a basis for P2 so that any p
(a) Let B = h
~ P2 has a unique
~1 + c2
~2 + c3
~3 , which we denote in this way.
representation as p
~ = c1
!
c1
RepB (~
p) = c2
c3
Show that the RepB () operation is a function from P2 to R3 (this entails showing
that with every domain vector ~v P2 there is an associated image vector in R3 ,
and further, that with every domain vector ~v P2 there is at most one associated
image vector).
(b) Show that this RepB () function is one-to-one and onto.
(c) Show that it preserves structure.
(d) Produce an isomorphism from P2 to R3 that fits these specifications.
x + x 7
2
1
0
0
and
1 x 7
0
1
0
and
r (~
u, w)
~ = (r~
u, rw)
~
Section I. Isomorphisms
169
(d) Suppose that U and W are subspaces of a vector space V such that V =
U W . Show that the map f : U W V given by
f
(~
u, w)
~ 7 ~
u+w
~
is an isomorphism. Thus if the internal direct sum is defined then the internal
and external direct sums are isomorphic.
3.I.2
~ 1 + c2 w
~ 2 ) = f 1 c1 f (~v1 ) + c2 f (~v2 )
f 1 (c1 w
= f 1 ( f c1~v1 + c2~v2 )
= c1~v1 + c2~v2
= c1 f 1 (w
~ 1 ) + c2 f 1 (w
~ 2)
Finally, to check transitivity, that if V is isomorphic to W via some map f
and if W is isomorphic to U via some map g then also V is isomorphic to U ,
consider the composition map g f : V U . As stated in the appendix, the
170
= g c1 f (~v1 ) + c2 f (~v2 )
QED
Finite-dimensional
vector spaces:
.V %
...
.
W$
V
=W
The next result gives a simple criteria describing which spaces are in each class.
2.2 Theorem Vector spaces are isomorphic if and only if they have the same
dimension.
This theorem follows from the next two lemmas.
2.3 Lemma If spaces are isomorphic then they have the same dimension.
Proof. We shall show that an isomorphism of two spaces gives a correspondence
between their bases. That is, where f : V W is an isomorphism and a basis
~1 ), . . . , f (
~n )i
~1 , . . . , ~n i, then the image set D = hf (
for the domain V is B = h
is a basis for the codomain W . (The other half of the correspondencethat for
any basis of W the inverse image is a basis for V follows on recalling that if
f is an isomorphism then f 1 is also an isomorphism, and applying the prior
sentence to f 1 .)
To see that D spans W , fix a w
~ W , use the fact that f is onto and so there
is a ~v V with w
~ = f (~v ), and expand ~v as a combination of basis vectors.
~n )
~1 + + vn
~n ) = v1 f (~1 ) + + vn f (
w
~ = f (~v ) = f (v1
For linear independence of D, if
~1 ) + + cn f (
~n ) = f (c1
~1 + + cn
~n )
~0W = c1 f (
then, since f is one-to-one and so the only vector sent to ~0W is ~0V , we have
~1 + + cn
~n , implying that all the cs are zero.
QED
that ~0V = c1
Section I. Isomorphisms
171
2.4 Lemma If spaces have the same dimension then they are isomorphic.
Proof. To show that any two spaces of dimension n are isomorphic, we can
simply show that any one is isomorphic to Rn . Then we will have shown that
they are isomorphic to each other, by the transitivity of isomorphism (which
was established in Theorem 2.1).
~n i for the
~1 , . . . ,
Let V be an n-dimensional space. Fix a basis B = h
domain V and consider as a function the representation of the members of that
domain with respect to the basis.
v1
RepB .
~
~
~v = v1 1 + + vn n 7 ..
vn
(This is well-defined since every ~v has one and only one such representationsee
Remark 2.5 below. )
This function is one-to-one because if
~1 + + un
~n ) = RepB (v1
~1 + + vn
~n )
RepB (u1
then
u1
v1
.. ..
. = .
un
vn
~1 + +
and so u1 = v1 , . . . , un = vn , and therefore the original arguments u1
~
~
~
un n and v1 1 + + vn n are equal.
This function is onto; any n-tall vector
w1
..
w
~ = .
wn
~ 1 + + vn
~n ).
is the image of some ~v V , namely w
~ = RepB (v1
Finally, this function preserves structure.
~1 + + (run + svn )
~n )
RepB (r ~u + s ~v ) = RepB ( (ru1 + sv1 )
ru1 + sv1
..
=
.
run + svn
v1
u1
..
..
=r . +s .
un
vn
172
Thus the function is an isomorphism, and we can say that any n-dimensional
space is isomorphic to the n-dimensional space Rn . Consequently, as noted at
QED
the start, any two spaces with the same dimension are isomorphic.
2.5 Remark The parenthetical comment in that proof about the role played
by the one and only one representation result requires some explanation. We
need to show that each vector in the domain is associated by RepB with one
and only one vector in the codomain.
A contrasting example, where an association doesnt have this property, is
illuminating. Consider this subset of P2 , which is not a basis.
A = {1 + 0x + 0x2 , 0 + 1x + 0x2 , 0 + 0x + 1x2 , 1 + 1x + 2x2 }
~ 4 . If, mimicing above proof, we try to
Call those four polynomials
~ 1, . . . ,
~ 1 + c2
~ 2 + c3
~ 3 + c4
~ 4 , and associate p~
write the members of P2 as p~ = c1
with the four-tall vector with components c1 , . . . , c4 then there is a problem.
For, consider p~(x) = 1 + x + x2 . The set A spans the space P2 , so there is at
least one four-tall vector associated with p~. But A is not linearly independent
so vectors do not have unique decompositions. In this case, both
2 + 1~
3 + 0~
4
p~(x) = 1~
1 + 1~
and
p~(x) = 0~
1 + 0~
2 1~
3 + 1~
4
and so there is more than one four-tall vector associated with p~.
1
0
1
and 0
1
1
0
1
If we are trying to think of this association as a function then the problem is
that, for instance, with input p~ the association does not have a well-defined
output value.
Any map whose definition appears possibly ambiguous must be checked to
see that it is well-defined. For the above proof that check is Exercise 19.
That ends the proof of Theorem 2.2. We say that the isomorphism classes
are characterized by dimension because we can describe each class simply by
giving the number that is the dimension of all of the spaces in that class.
This subsections results give us a collection of representatives of the isomorphism classes.
2.6 Corollary A finite-dimensional vector space is isomorphic to one and only
one of the Rn .
2.7 Remark The proofs above pack many ideas into a small space. Through
the rest of this chapter well consider these ideas again, and fill them out. For a
taste of this, we will close this section by indicating how we can expand on the
proof of Lemma 2.4.
Section I. Isomorphisms
173
1 0
0 1
0 0
0 0
B=h
,
,
,
i
0 0
0 0
1 0
0 1
the isomorphism given in the lemma, the representation map, simply carries the
entries over.
a
b
a b f1
7
c
c d
d
One way to understand the map f1 is this: we fix the basis B for the domain
~1 with ~e1 , and
~2 with ~e2 ,
and the basis E4 for the codomain, and associate
etc. We then extend this association to all of the vectors in two spaces.
a b
c d
~2 + c
~3 + d
~4
~ 1 + b
= a
a
b
f1
c
d
We say that the map has been extended linearly from the bases to the spaces.
We can do the same thing with different bases, for instance, taking this basis
for the domain.
2 0
0 2
0 0
0 0
A=h
,
,
,
i
0 0
0 0
2 0
0 2
Associating corresponding members of A and E4 , and extending linearly,
a b
c d
2 + (c/2)~
3 + (d/2)~
4
= (a/2)~
1 + (b/2)~
a/2
b/2
f2
c/2
d/2
and
1
0
0
0
0 1 0 0
D = h , , ,
i
0
0
0
1
0
0
1
0
174
b
f3
a b
~
~
~
~
~
~
~
~
7 a1 + b2 + c3 + d4 =
a1 + b2 + c3 + d4 =
c d
d
c
gives still another isomorphism.
So there is a connection between the maps between spaces and bases for
those spaces. We will explore that connection in later sections.
We now finish this section with a summary.
Recall that in the first chapter, we defined two matrices as row equivalent
if they can be derived from each other by elementary row operations (this was
the meaning of same-ness that was of interest there). We showed that is an
equivalence relation and so the collection of matrices is partitioned into classes,
where all the matrices that are row equivalent fall together into a single class.
Then, for insight into which matrices are in each class, we gave representatives
for the classes, the reduced echelon form matrices.
In this section, except that the appropriate notion of same-ness here is vector
space isomorphism, we have followed much the same outline. First we defined
isomorphism, saw some examples, and established some basic properties. Then
we showed that it is an equivalence relation, and now we have a set of class
representatives, the real vector spaces R1 , R2 , etc.
Finite-dimensional
vector spaces:
?R3
?R0
%
.V ?R2
...
.
W$
?R1
One representative
per class
?R4
Section I. Isomorphisms
X
X
X
175
(a) R2 , R4
(b) P5 , R5
(c) M23 , R6
(d) P5 , M23
(e) M2k , Ck
2
2.10 Consider the isomorphism RepB () : P1 R where B = h1, 1 + xi. Find the
image of each of these elements of the domain.
(a) 3 2x;
(b) 2 + 2x;
(c) x
2.11 Show that if m 6= n then Rm
6 Rn .
=
176
3.II
Homomorphisms
The definition of isomorphism has two conditions. In this section we will consider the second one, that the map must preserve the algebraic structure of the
space. We will focus on this condition by studying maps that are required only
to preserve structure; that is, maps that are not required to be correspondences.
Experience shows that this kind of map is tremendously useful in the study
of vector spaces. For one thing, as we shall see in the second subsection below,
while isomorphisms describe how spaces are the same, these maps describe how
spaces can be thought of as alike.
3.II.1
Definition
x2
x1 + x2
x2
x1
x1
+
x
x
1
2
= ( y1 ) + ( y2 )
( y1 + y2 ) = ( y1 + y2 ) =
y1 + y2
z1
z2
z1 + z2
z1
z2
and it preserves scalar multiplication.
rx1
x1
x1
rx1
= r ( y1 )
(r y1 ) = ( ry1 ) =
ry1
z1
rz1
z1
Note that this map is not an isomorphism, since it is not one-to-one. For
instance, both ~0 and ~e3 in R3 are mapped to the zero vector in R2 .
1.3 Example The domain and codomain can be other than spaces of column
vectors. Both of these maps are homomorphisms.
177
(1) f1 : P2 P3 given by
a0 + a1 x + a2 x2 7 a0 x + (a1 /2)x2 + (a2 /3)x3
a b
c d
7 a + d
y 7g
3x + 2y 4.5z + 1
z
is not linear; for instance,
0
1
g(0 + 0) = 4 while
0
0
0
1
g(0) + g(0) = 5
0
0
(to show that a map is not linear we need only produce one example of a
linear combination that is not preserved).
(2) The first of these two maps t1 , t2 : R3 R2 is linear while the second is
not.
x
x
5x 2y
5x 2y
t2
1
y 7t
and y 7
x+y
xy
z
z
(Finding an example that the second fails to preserve structure is easy.)
What distinguishes the homomorphisms is that the coordinate functions are
linear combinations of the arguments. See also Exercise 22.
178
Obviously, any isomorphism is a homomorphisman isomorphism is a homomorphism that is also a correspondence. So, one way to think of the homomorphism idea is that it is a generalization of isomorphism, motivated by the
observation that many of the properties of isomorphisms have only to do with
the map respecting structure and not to do with it being a correspondence. As
examples, these two results from the prior section do not use one-to-one-ness or
onto-ness in their proof, and therefore apply to any homomorphism.
1.6 Lemma A homomorphism sends a zero vector to a zero vector.
1.7 Lemma Each of these is a necessary and sufficient condition for f : V W
to be a homomorphism.
(1) for any c1 , c2 R and ~v1 , ~v2 V ,
f (c1 ~v1 + c2 ~v2 ) = c1 f (~v1 ) + c2 f (~v2 )
x/2
f 0
x
7
x + y
y
3y
satisfies that check
x1 /2
x2 /2
r1 (x1 /2) + r2 (x2 /2)
0
0
r1 (x1 + y1 ) + r2 (x2 + y2 ) = r1 x1 + y1 + r2 x2 + y2
3y1
3y2
r1 (3y1 ) + r2 (3y2 )
and so it is a homomorphism.
(Sometimes, such as with Lemma 1.15 below, it is less awkward to check preservation of addition and preservation of scalar multiplication separately, but this
is purely a matter of taste.)
However, some of the results that we have seen for isomorphisms fail to hold
for homomorphisms in general. An isomorphism between spaces gives a correspondence between their bases, but a homomorphisms need not; Example 1.2
shows this and another example is the zero map between any two nontrivial
spaces. Instead, a weaker but still very useful result holds.
179
~1 + + cn
~n ,
extending linearly to all of the domain. That is, where ~v = c1
~ 1 + + cn w
~ n . This is well-defined because, with respect to the
let h(~v ) be c1 w
basis, the representation of each domain vector ~v is unique.
This map is a homomorphism since it preserves linear combinations; where
~1 + + cn
~n and v~2 = d1
~ 1 + + dn
~n , we have this.
v~1 = c1
~1 + + (r1 cn + r2 dn )
~n )
h(r1~v1 + r2~v2 ) = h((r1 c1 + r2 d1 )
~ 1 + + (r1 cn + r2 dn )w
~n
= (r1 c1 + r2 d1 )w
= r1 h(~v1 ) + r2 h(~v2 )
: V W is another homomorphism such
And, this map is unique since if h
agree on all of the vectors in the domain.
~i ) = w
~ i for each i then h and h
that h(
~1 + + cn
~n )
v ) = h(c
1
h(~
~1 ) + + cn h(
~n )
= c1 h(
~ 1 + + cn w
~n
= c1 w
= h(~v )
are the same map.
Thus, h and h
QED
1.10 Example This result says that we can construct homomorphisms by fixing a basis for the domain and specifying where the map sends those basis vectors. For instance, if we specify a map h : R2 R2 that acts on the standard
basis E2 in this way
1
1
0
4
h(
)=
and h(
)=
1
1
4
0
then the action of h on any other member of the domain is also specified. For
instance, the value of h on this argument
3
1
0
1
0
5
h(
) = h(3
2
) = 3 h(
) 2 h(
)=
2
0
1
0
1
5
is a direct consiquence of the value of h on the basis vectors. (Later in this
chapter we shall develop a scheme, using matrices, that is a convienent way to
do computations like this one.)
Just as isomorphisms of a space with itself are useful and interesting, so too
are homomorphisms of a space with itself.
180
a b
a c
7
c d
b d
is a linear transformation of M22 . Note that this transformation is one-to-one
and onto, and so in fact is an automorphism.
We finish this subsection about maps by recalling that we can linearly combine maps. For instance, for these maps from R2 to itself
f
g
2x
x
0
x
7
and
7
3x 2y
y
5x
y
we can take the linear combination 5f 2g to get this.
10x
x 5f 2g
7
5x 10y
y
1.15 Lemma For vector spaces V and W , the set of linear functions from V
to W is itself a vector space, a subspace of the space of all functions from V to
W . It is denoted L(V, W ).
Proof. This set is non-empty because it contains the zero homomorphism. So
to show that it is a subspace we need only check that it is closed under linear
combinations. Let f, g : V W be linear. Then their sum is linear
= c1 f + g (~v1 ) + c2 f + g (~v2 )
181
QED
x
2x + y
(d) h( y ) =
3y 4z
z
X 1.17 Decide
if each
map h : M22 R is linear.
a b
(a) h(
)=a+d
c d
a
(b) h(
c
(c) h(
x
1
(c) h( y ) =
1
z
b
) = ad bc
d
a
c
b
) = 2a + 3b + c d
d
a
c
b
) = a2 + b 2
d
(d) h(
182
X 1.22 Stating that a function is linear is different than stating that its graph is a
line.
(a) The function f1 : R R given by f1 (x) = 2x 1 has a graph that is a line.
Show that it is not a linear function.
(b) The function f2 : R2 R given by
x
y
7 x + 2y
does not have a graph that is a line. Show that it is a linear function.
X 1.23 Part of the definition of a linear function is that it respects addition. Does a
linear function respect subtraction?
~1 , . . . ,
~n i is a basis
1.24 Assume that h is a linear transformation of V and that h
of V . Prove each statement.
~i ) = ~0 for each basis vector then h is the zero map.
(a) If h(
~i ) =
~i for each basis vector then h is the identity map.
(b) If h(
~i ) = r
~i for each basis vector then
(c) If there is a scalar r such that h(
h(~v ) = r ~v for all vectors in V .
X 1.25 Consider the vector space R+ where vector addition and scalar multiplication
are not the ones inherited from R but rather are these: a + b is the product of
a and b, and r a is the r-th power of a. (This was shown to be a vector space
in an earlier exercise.) Verify that the natural logarithm map ln : R+ R is a
homomorphism between these two spaces. Is it an isomorphism?
X 1.26 Consider this transformation of R2 .
x
y
x/2
y/3
x
y
2
(x /4) + (y 2 /9) = 1}
X 1.27 Imagine a rope wound around the earths equator so that it fits snugly (suppose that the earth is a sphere). How much extra rope must be added to raise the
circle to a constant six feet off the ground?
X 1.28 Verify that this map h : R3 R
!
x
y
z
x
y
z
3
1
1
= 3x y z
is linear. Generalize.
1.29 Show that every homomorphism from R1 to R1 acts via multiplication by a
scalar. Conclude that every nontrivial linear transformation of R1 is an isomorphism. Is that true for transformations of R2 ? Rn ?
1.30 (a) Show that for any scalars a1,1 , . . . , am,n this map h : Rn Rm is a homomorphism.
x1
a1,1 x1 + + a1,n xn
..
..
. 7
.
xn
am,1 x1 + + am,n xn
183
(b) Show that for each i, the i-th derivative operator di /dxi is a linear transformation of Pn . Conclude that for any scalars ck , . . . , c0 this map is a linear
transformation of that space.
dk
dk1
d
f 7
f
+
c
f + + c1 f + c0 f
k1
dxk
dxk1
dx
1.31 Lemma 1.15 shows that a sum of linear functions is linear and that a scalar
multiple of a linear function is linear. Show also that a composition of linear
functions is linear.
X 1.32 Where f : V W is linear, suppose that f (~v1 ) = w
~ 1 , . . . , f (~vn ) = w
~ n for
some vectors w
~ 1, . . . , w
~ n from W .
(a) If the set of w
~ s is independent, must the set of ~v s also be independent?
(b) If the set of ~v s is independent, must the set of w
~ s also be independent?
(c) If the set of w
~ s spans W , must the set of ~v s span V ?
(d) If the set of ~v s spans V , must the set of w
~ s span W ?
1.33 Generalize Example 1.14 by proving that the matrix transpose map is linear.
What is the domain and codomain?
1.34 (a) Where ~
u, ~v Rn , the
line segment connecting them is defined to be
the set ` = {t ~
u + (1 t) ~v t [0..1]}. Show that the image, under a homomorphism h, of the segment between ~
u and ~v is the segment between h(~
u) and
h(~v ).
(b) A subset of Rn is convex if, for any two points in that set, the line segment
joining them lies entirely in that set. (The inside of a sphere is convex while the
skin of a sphere is not.) Prove that linear maps from Rn to Rm preserve the
property of set convexity.
X 1.35 Let h : Rn Rm be a homomorphism.
(a) Show that the image under h of a line in Rn is a (possibly degenerate) line
in Rn .
(b) What happens to a k-dimensional linear surface?
1.36 Prove that the restriction of a homomorphism to a subspace of its domain is
another homomorphism.
1.37 Assume that h : V W is linear.
(a) Show that the rangespace of this map {h(~v ) ~v V } is a subspace of the
codomain W .
(b) Show that the nullspace of this map {~v V h(~v ) = ~0W } is a subspace of
the domain V .
(c) Show that if U is a subspace of the domain V then its image {h(~
u) ~
u U}
is a subspace of the codomain W . This generalizes the first item.
(d) Generalize the second item.
1.38 Consider the set of isomorphisms from a vector space to itself. Is this a
subspace of the space L(V, V ) of homomorphisms from the space to itself?
~1 , . . . ,
~n i is a basis? That is, can we still get
1.39 Does Theorem 1.9 need that h
a well-defined and unique homomorphism if we drop either the condition that the
~ be linearly independent, or the condition that it span the domain?
set of s
1.40 Let V be a vector space and assume that the maps f1 , f2 : V R1 are linear.
(a) Define a map F : V R2 whose component functions are the given linear
ones.
f1 (~v )
~v 7
f2 (~v )
184
3.II.2
The image h(S) is nonempty because S is nonempty. Thus, to show that h(S)
is a subspace of the codomain W , we need only show that it is closed under
linear combinations of two vectors. If h(~s1 ) and h(~s2 ) are members of h(S) then
c1 h(~s1 ) + c2 h(~s2 ) = h(c1 ~s1 ) + h(c2 ~s2 ) = h(c1 ~s1 + c2 ~s2 ) is also a member
QED
of h(S) because it is the image of c1 ~s1 + c2 ~s2 from S.
2.2 Definition The rangespace of h : V W is
R(h) = {h(~v ) ~v V }
sometimes denoted h(V ). The dimension of the rangespace is the maps rank .
(We shall soon see the connection between the rank of a map and the rank of a
matrix.)
2.3 Example Recall that the derivative map d/dx : P3 P3 given by a0 +
a1 x + a2 x2 + a3 x3 7 a1 + 2a2 x + 3a3 x2 is linear.
The rangespace R(d/dx) is
the set of quadratic polynomials {r + sx + tx2 r, s, t R}. Thus, the rank of
this map is three.
2.4 Example With this homomorphism h : M22 P3
a b
7 (a + b + 2d) + 0x + cx2 + cx3
c d
185
an image vector in the range can have any constant term, must have an x
coefficient of zero, and must have the same coefficient
of x2 as of x3 . That is,
2
3
the rangespace is R(h) = {r + 0x + sx + sx r, s R} and so the rank is two.
The prior result shows that, in passing from the definition of isomorphism to
the more general definition of homomorphism, omitting the onto requirement
doesnt make an essential difference. Any homomorphism is onto its rangespace.
However, omitting the one-to-one condition does make a difference. A
homomorphism may have many elements of the domain map to a single element
in the range. The general picture is below. There is a homomorphism and its
domain, codomain, and range. The homomorphism is many-to-one, and two
elements of the range are shown that are each the image of more than one
member of the domain.
domain V
codomain W
R(h)
.
(Recall that for a map h : V W , the set of elements of the domain that are
~ is the inverse image of w.
~ It
mapped to w
~ in the codomain {~v V h(~v ) = w}
~ this notation is used even if h has no inverse function, that
is denoted h1 (w);
is, even if h is not one-to-one.)
2.5 Example Consider the projection : R3 R2
x
y 7 x
y
z
which is a homomorphism but is not one-to-one. Picturing R2 as the xy-plane
inside of R3 allows us to see (~v ) as the shadow of ~v in the plane. In these
terms, the preservation of addition property says that
~
v1 above (x1 , y1 )
plus
~
v2 above (x2 , y2 )
equals
~
v1 + ~
v2 above (x1 + x2 , y1 + y2 ).
Briefly, the shadow of a sum equals the sum of the shadows. (Preservation of
scalar multiplication has a similar interpretation.)
186
w
~2
w
~1
w
~1 + w
~2
v
u1
) = a and h( 1 ) = b
h(
u2
v2
then
u 1 + v1
h(
)=a+b
u 2 + v2
187
(u1 + v1 , u2 + v2 )
(u1 , u2 )
(v1 , v2 )
an a vector
plus
a b vector
equals
an a + b vector
2.8 Remark We wont describe how every homomorphism that we will use in
this book is an analogy, both because the formal sense we make of alike in this
way . . . is a homomorphism exists such that . . . , and because many vector
spaces are hard to draw (e.g., a space of polynomials). Nonetheless, the idea
that a homomorphism between two spaces expresses how the domains vectors
fall into classes that act like the the ranges vectors, is a good way to view
homomorphisms.
We derive two insights from examples 2.5, 2.6, and 2.7.
First, in all three, each inverse image shown is a linear surface. In particular,
the inverse image of the ranges zero vector is a line or plane through the origin
a subspace of the domain. The next result shows that this insight extends to
any vector space, not just spaces of column vectors (which are the only spaces
where the term linear surface is defined).
2.9 Lemma For any homomorphism, the inverse image of a subspace of the
range is a subspace of the domain. In particular, the inverse image of the trivial
subspace of the range is a subspace of the domain.
Proof. Let h : V W be a homomorphism and let S be a subspace of the
188
2.12 Example The map from Example 2.4 has this nullspace.
a
b
a, b R}
N (h) = {
0 (a + b)/2
Now for the second insight from the above pictures. In Example 2.5, each
of the vertical lines is squashed down to a single point, in passing from the
domain to the range, takes all of these one-dimensional vertical lines and zeroes
them out, leaving the range one dimension smaller than the domain. Similarly,
in Example 2.6, the two-dimensional domain is mapped to a one-dimensional
range by breaking the domain into lines (here, they are diagonal lines), and
compressing each of those lines to a single member of the range. Finally, in
Example 2.7, the domain breaks into planes which get zeroed out, and so the
map starts with a three-dimensional domain but ends with a one-dimensional
rangethis map subtracts two from the dimension. (Notice that, in this
third example, the codomain is two-dimensional but the range of the map is
only one-dimensional, and it is the dimension of the range that is of interest.)
2.13 Theorem A linear maps rank plus its nullity equals the dimension of its
domain.
~1 , . . . , ~k i be a basis for
Proof. Let h : V W be linear and let BN = h
~1 , . . . , ~k , ~k+1 , . . . ,
~n i for the
the nullspace. Extend that to a basis BV = h
~
~
entire domain. We shall show that BR = hh(k+1 ), . . . , h(n )i is a basis for the
rangespace. Then counting the size of these bases gives the result.
~k+1 ) +
To see that BR is linearly independent, consider the equation ck+1 h(
~n ) = ~0W . This gives that h(ck+1
~k+1 + + cn
~n ) = ~0W and so
+ cn h(
~
~
ck+1 k+1 + +cn n is in the nullspace of h. As BN is a basis for this nullspace,
there are scalars c1 , . . . , ck R satisfying this relationship.
~1 + + ck
~k = ck+1
~k+1 + + cn
~n
c1
But BV is a basis for V so each scalar equals zero. Therefore BR is linearly
independent.
189
To show that BR spans the rangespace, consider h(~v ) R(h) and write ~v
~1 + + cn
~n of members of BV . This gives
as a linear combination ~v = c1
~1 )+ +ck h(
~k )+ck+1 h(
~k+1 )+ +cn h(~n )
~1 + +cn
~n ) = c1 h(
h(~v ) = h(c1
~
~
and since 1 , . . . , k are in the nullspace, we have that h(~v ) = ~0 + + ~0 +
~k+1 ) + + cn h(
~n ). Thus, h(~v ) is a linear combination of members of
ck+1 h(
QED
BR , and so BR spans the space.
2.14 Example Where h : R3 R4 is
x
x
h
y 7 0
y
z
0
190
7 y
y
0
gives the obvious correspondence between R2 and the xy-plane inside of R3 .
We will close this section by adapting some results about isomorphisms to
this setting.
2.20 Theorem In an n-dimensional vector space V , then these
(1) h is nonsingular, that is, one-to-one
(2) h has a linear inverse
(3) N (h) = {~0 }, that is, nullity(h) = 0
(4) rank(h) = n
~n i is a basis for V then hh(
~1 ), . . . , h(
~n )i is a basis for R(h)
~1 , . . . ,
(5) if h
are equivalent statements about a linear map h : V W .
Proof. We will first show that (1) (2). We will then show that (1) =
(3) = (4) = (5) = (2).
For (1) = (2), suppose that the linear map h is one-to-one, and so has an
inverse. The domain of that inverse is the range of h and so a linear combination of two members of that domain has the form c1 h(~v1 ) + c2 h(~v2 ). On that
combination, the inverse h1 gives this.
191
Next, (3) = (4) is true since rank plus nullity equals the dimension of the
domain.
~n )i is a basis for the rangespace
~1 ), . . . , h(
For (4) = (5), to show that hh(
we need only show that it is a spanning set, because by assumption the range
has dimension n. Consider h(~v ) R(h). Expressing ~v as a linear combination
~1 + c2
~2 + + cn
~n ), which gives that
of basis elements produces h(~v ) = h(c1
~
~
h(~v ) = c1 h(1 ) + + cn h(n ), as desired.
~1 , . . . , ~n i is a basis
Finally, for the (5) = (2) implication, assume that h
~
~
~ R(h) a the
for V so that hh(1 ), . . . , h(n )i is a basis for R(h). Then every w
~1 ) + + cn h(
~n ). Define a map from R(h) to
unique representation w
~ = c1 h(
V by
~1 + c2
~2 + + cn
~n
w
~ 7 c1
(uniqueness of the representation makes this well-defined). Checking that it is
QED
linear and that it is the inverse of h are easy.
Weve now seen that a linear map shows how the structure of the domain is
like that of the range. Such a map can be thought to organize the domain space
into inverse images of points in the range. In the special case that the map is
one-to-one, each inverse image is a single point and the map is an isomorphism
between the domain and the range.
Exercises
X 2.21 Let h : P3 P4 be given by p(x) 7 x p(x). Which of these are in the
nullspace? Which are in the rangespace?
(a) x3
(b) 0
(c) 7
(d) 12x 0.5x3
(e) 1 + 3x2 x3
X 2.22 Find the nullspace, nullity, rangespace, and rank of each map.
(a) h : R2 P3 given by
a
b
7 a + ax + ax2
a
c
b
d
7 a + d
b
d
a
c
7 a + b + c + dx2
192
f(
x
) = 2x + y
y
x=1
a0 + a1 x + + an xn 7
a0 + a1 x + + an xn dx.
x=0
2.33 (a) Prove that a homomorphism is onto if and only if its rank equals the
dimension of its codomain.
(b) Conclude that a homomorphism between vector spaces with the same dimension is one-to-one if and only if it is onto.
2.34 Show that a linear map is nonsingular if and only if it preserves linear independence.
2.35 Corollary 2.16 says that for there to be an onto homomorphism from a vector
space V to a vector space W , it is necessary that the dimension of W be less
than or equal to the dimension of V . Prove that this condition is also sufficient;
use Theorem 1.9 to show that if the dimension of W is less than or equal to the
dimension of V , then there is a homomorphism from V to W that is onto.
2.36 Let h : V R be a homomorphism, but not the zero homomorphism. Prove
~1 , . . . ,
~n i is a basis for the nullspace and if ~v V is not in the nullspace
that if h
~
~n i is a basis for the entire domain V .
then h~v , 1 , . . . ,
X 2.37 Recall that the nullspace is a subset of the domain and the rangespace is a
subset of the codomain. Are they necessarily distinct? Is there a homomorphism
that has a nontrivial intersection of its nullspace and its rangespace?
193
2.38 Prove that the image of a span equals the span of the images. That is, where
h : V W is linear, prove that if S is a subset of V then h([S]) equals [h(S)]. This
generalizes
Lemma 2.1 since it shows that if U is any subspace of V then its image
{h(~
u) ~
u U } is a subspace of W , because the span of the set U is U .
X 2.39 (a) Prove that for any linear map h : V W and any w
~ W , the set
h1 (w)
~ has the form
{~v + ~n ~n N (h)}
for ~v V with h(~v ) = w
~ (if h is not onto then this set may be empty). Such a
set is a coset of N (h) and is denoted ~v + N (h).
2
(b) Consider the map t : R2
R givenby
x
ax + by
t
7
y
cx + dy
for some scalars a, b, c, and d. Prove that t is linear.
(c) Conclude from the prior two items that for any linear system of the form
ax + by = e
cx + dy = f
the solution set can be written (the vectors are members of R2 )
{~
p + ~h ~h satisfies the associated homogeneous system}
where p
~ is a particular solution of that linear system (if there is no particular
solution then the above set is empty).
n
m
(d) Show that this map
h :R
R is linear
x1
a1,1 x1 + + a1,n xn
..
..
. 7
.
xn
am,1 x1 + + am,n xn
for any scalars a1,1 , . . . , am,n . Extend the conclusion made in the prior item.
(e) Show that the k-th derivative map is a linear transformation of Pn for each
k. Prove that this map is a linear transformation of that space
dk
dk1
d
f 7
f
+
c
f + + c1 f + c0 f
k1
dxk
dxk1
dx
for any scalars ck , . . . , c0 . Draw a conclusion as above.
2.40 Prove that for any transformation t : V V that is rank one, the map given
by composing the operator with itself t t : V V satisfies t t = r t for some
real number r.
2.41 Show that for any space V of dimension n,
the dual space
L(V, R) = {h : V R h is linear}
is isomorphic to Rn . It is often denoted V . Conclude that V
=V.
2.42 Show that any linear map is the sum of maps of rank one.
2.43 Is is homomorphic to an equivalence relation? (Hint: the difficulty is to
decide on an appropriate meaning for the quoted phrase.)
2.44 Show that the rangespaces and nullspaces of powers of linear maps t : V V
form descending
V R(t) R(t2 ) . . .
and ascending
{~0} N (t) N (t2 ) . . .
chains. Also show that if k is such that R(tk ) = R(tk+1 ) then all following
rangespaces are equal: R(tk ) = R(tk+1 ) = R(tk+2 ) . . . . Similarly, if N (tk ) =
N (tk+1 ) then N (tk ) = N (tk+1 ) = N (tk+2 ) = . . . .
194
3.III
The prior section shows that a linear map is determined by its action on a basis.
In fact, the equation
~1 + + cn
~n ) = c1 h(~1 ) + + cn h(~n )
h(~v ) = h(c1
shows that, if we know the value of the map on the vectors in a basis, then we
can compute the value of the map on any vector ~v at all just by finding the cs
to express ~v with respect to the basis.
This section gives the scheme that computes, from the representation of a
vector in the domain RepB (~v ), the representation of that vectors image in the
~1 ), . . . , h(
~n ).
codomain RepD (h(~v )), using the representations of h(
3.III.1
1
0
1
and D = h0 , 2 , 0i
0
0
1
as the bases for these spaces) that is determined by this action on the vectors
in the domains basis.
1
2
h
7 1
0
1
1
1
h
7 2
4
0
To compute the action of this map on any vector at all from the domain, we
~2 ) with respect to the codomains basis:
~1 ) and h(
first express h(
0
1
1
1
1 = 0 0 1 2 + 1 0
2
0
1
1
0
0
~1 )) = 1/2
so RepD (h(
1
D
and
1
1
0
1
2 = 1 0 1 2 + 0 0
0
0
0
1
1
~2 )) = 1
so RepD (h(
0 D
195
(these are easy to check). Then, as described in the preamble, for any member
~
~v of the domain, we can express the image h(~v ) in terms of the h()s.
2
1
+ c2
)
0
4
2
1
) + c2 h(
)
= c1 h(
0
4
0
1
1
1
0
1
1
= c1 (0 0 2 + 1 0) + c2 (1 0 1 2 + 0 0)
2
0
1
0
0
0
1
1
0
1
1
= (0c1 + 1c2 ) 0 + ( c1 1c2 ) 2 + (1c1 + 0c2 ) 0
2
0
0
1
h(~v ) = h(c1
Thus,
0c1 + 1c2
c1
then RepD ( h(~v ) ) = (1/2)c1 1c2 .
with RepB (~v ) =
c2
1c1 + 0c2
For instance,
2
4
1
4
)=
then RepD ( h(
) ) = 5/2.
with RepB (
8
2 B
8
1
This is a formula that computes how h acts on any argument.
We will express computations like the one above with a matrix notation.
0
1/2
1
0c1 + 1c2
1
c
1
1
= (1/2)c1 1c2
c2 B
1c1 + 0c2
0 B,D
D
In the middle is the argument ~v to the map, represented with respect to the
domains basis B by a column vector with components c1 and c2 . On the right
is the value h(~v ) of the map on that argument, represented with respect to the
codomains basis D by a column vector with components 0c1 + 1c2 , etc. The
matrix on the left is the new thing. It consists of the coefficients from the vector
on the right, 0 and 1 from the first row, 1/2 and 1 from the second row, and
1 and 0 from the third row.
This notation simply breaks the parts from the right, the coefficients and the
cs, out separately on the left, into a vector that represents the maps argument
and a matrix that we will take to represent the map itself.
196
1.2 Definition Suppose that V and W are vector spaces of dimensions n and
m with bases B and D, and that h : V W is a linear map. If
h1,1
h1,n
h2,1
h2,n
~1 )) =
~n )) =
RepD (h(
.. , . . . , RepD (h(
..
.
.
hm,1
hm,n
then
h1,1
h2,1
RepB,D (h) =
hm,1
h1,2
h2,2
..
.
...
...
hm,2
...
h1,n
h2,n
hm,n B,D
..
..
.
.
~
~n ) )
RepB,D (h) =
RepD ( h(
RepD ( h(1 ) )
..
..
.
.
Observe that the number of columns of the matrix is the dimension of the
domain of the map, and the number of rows is the dimension of the codomain.
1.3 Example If h : R3 P1 is given by
a1
a2 7h (2a1 + a2 ) + (a3 )x
a3
then where
0
0
2
B = h0 , 2 , 0i and
1
0
0
D = h1 + x, 1 + xi
0
2 7h 2
0
2
0 7h 4
0
1/2
1
RepD (2) =
RepD (x) =
1/2 D
1 D
RepD (4) =
2
2 D
197
showing that this is the matrix representing h with respect to the bases.
1/2 1
2
RepB,D (h) =
1/2 1 2 B,D
We will use lower case letters for a map, upper case for the matrix, and
lower case again for the entries of the matrix. Thus for the map h, the matrix
representing it is H, with entries hi,j .
1.4 Theorem Assume that V and W are vector spaces of dimensions m and
n with bases B and D, and that h : V W is a linear map. If h is represented
by
RepB,D (h) =
..
.
hm,1
hm,2
...
hm,n
B,D
and ~v V is represented by
c1
c2
RepB (~v ) = .
..
cn
RepD ( h(~v ) ) =
..
.
hm,1 c1 + hm,2 c2 + + hm,n cn
Proof. Exercise 28.
QED
We will think of the matrix RepB,D (h) and the vector RepB (~v ) as combining
to make the vector RepD (h(~v )).
1.5 Definition The matrix-vector product of a mn matrix and a n1 vector
is this.
a1,1 c1 + a1,2 c2 + + a1,n cn
a1,1 a1,2 . . . a1,n
c
1
. =
..
..
.
.
cn
am,1 am,2 . . . am,n
am,1 c1 + am,2 c2 + + am,n cn
198
The point of Definition 1.2 is to generalize Example 1.1, that is, the point
of the definition is Theorem 1.4, that the matrix describes how to get from
the representation of a domain vector with respect to the domains basis to
the representation of its image in the codomain with respect to the codomains
basis. With Definition 1.5, we can restate this as: application of a linear map is
represented by the matrix-vector product of the maps representative and the
vectors representative.
1.6 Example With the matrix from Example 1.3 we can calculate where that
map sends this vector.
4
~v = 1
0
This vector is represented, with respect to the domain basis B, by
0
1/2
RepB (~v ) =
2 B
and so this is the representation of the value h(~v ) with respect to the codomain
basis D.
0
1/2 1
2
1/2
RepD (h(~v )) =
1/2 1 2 B,D
2 B
(1/2) 0 + 1 (1/2) + 2 2
9/2
=
=
(1/2) 0 1 (1/2) 2 2 D
9/2 D
To find h(~v ) itself, not its representation, take (9/2)(1 + x) (9/2)(1 + x) = 9.
1.7 Example Let : R3 R2 be projection onto the xy-plane. To give a
matrix representing this map, we first fix bases.
1
1
1
2
1
B = h0 , 1 , 0 i
D=h
,
i
1
1
0
0
1
For each vector in the domains basis, we find its image under the map.
1
1
1
1 7 1
0 7 1
0 7 1
0
1
0
0
1
0
Then we find the representation of each image with respect to the codomains
basis
1
1
1
0
1
1
)=
RepD (
)=
RepD (
)=
RepD (
0
1
1
1
0
1
199
(these are easily checked). Finally, adjoining these representations gives the
matrix representing with respect to B, D.
1 0 1
RepB,D () =
1 1 1 B,D
We can illustrate Theorem 1.4 by computing the matrix-vector product representing the following statement about the projection map.
2
2
( 2 ) =
2
1
Representing this vector from the domain with respect to the domains basis
2
1
RepB (2) = 2
1
1 B
gives this matrix-vector product.
2
1 0
RepD ( (1) ) =
1 1
1
1
1
2 = 0
1 B,D
2 D
1 B
200
~
v
cos
0 t
sin
1 t
7
7
sin
1
cos
0
Then we represent these images with respect to the codomains basis. Because
this basis is E2 , vectors are represented by themselves. Finally, adjoining the
representations gives the matrix representing the map.
cos sin
RepE2 ,E2 (t ) =
sin
cos
The advantage of this scheme is that just by knowing how to represent the image
of the two basis vectors, we get a formula that tells us the image of any vector
at all; here a vector rotated by = /6.
t/6
3
3.598
3/2
1/2
3
7
2
0.232
2
3/2
1/2
(Again, we are using the fact that, with respect to E2 , vectors represent themselves.)
We have already seen the addition and scalar multiplication operations of
matrices and the dot product operation of vectors. Matrix-vector multiplication
is a new operation in the arithmetic of vectors and matrices. Nothing in Definition 1.5 requires us to view it in terms of representations. We can get some
insight into this operation by turning away from what is being represented, and
instead focusing on how the entries combine.
1.9 Example In the definition the width of the matrix equals the height of
the vector. Hence, the first product below is defined while the second is not.
1 0 0
1
1 0 0
1
0 =
4 3 1
6
4 3 1
0
2
One reason that this product is not defined is purely formal: the definition requires that the sizes match, and these sizes dont match. (Behind the formality,
though, we have a reason why it is left undefinedthe matrix represents a map
with a three-dimensional domain while the vector represents a member of a
two-dimensional space.)
201
..
..
1
.
.
c2
ai,1 c1 + ai,2 c2 + . . . + ai,n cn
ai,1 ai,2 . . . ai,n
=
.
.
.
..
..
.
.
cn
Looked at in this
Matrix-vector
h1,1 h1,2
h2,1 h2,2
..
.
hm,1
hm,2
1.10 Example
1 0
2 0
c1
h1,1 c1 + h1,2 c2 + + h1,n cn
. . . h1,n
. . . h2,n
c2 h2,1 c1 + h2,2 c2 + + h2,n cn
.. =
..
.
.
...
hm,n
cn
h1,1
h1,n
h2,1
h2,n
= c1 . + + cn .
..
..
hm,1
hm,n
2
1
1
0
1
1
1 = 2
1
+1
=
3
2
0
3
7
1
The result has the columns of the matrix weighted by the entries of the
vector. This way of looking at it brings us back to the objective stated at the
~1 ) + + cn h(
~n ).
~1 + + cn
~n ) as c1 h(
start of this section, to compute h(c1
We began this section by noting that the equality of these two enables us
~n ).
~1 ), . . . , h(
to compute the action of h on any argument knowing only h(
We have developed this into a scheme to compute the action of the map by
taking the matrix-vector product of the matrix representing the map and the
vector representing the argument. In this way, any linear map is represented
with respect to some bases by a matrix. In the next subsection, we will show
the converse, that any matrix represents a linear map.
Exercises
X 1.11 Multiply, where it is defined, the matrix
!
1
3
1
0 1 2
1
1
0
by each vector.
!
!
2
0
2
(a) 1
(b)
(c) 0
2
0
0
1.12 Perform, if possible, each matrix-vector multiplication.
202
(a)
2
3
1
1/2
4
2
(b)
1
2
2
1
0
1
1 1
1
1
1!
0
0
3
1
! !
1
3
2
x
y
z
(c)
1
2
1!
1
1
3
1
!
=
8
4
4
x 7 1 + 2x,
and
x2 7 x x3
R1
203
~1 7
~2 ,
~2 7
~3 ,
~3 7 ~0,
~4 7 ~0
(c)
1.21 Example 1.8 shows how to represent the rotation transformation of the plane
with respect to the standard basis. Express these other transformations also with
respect to the standard basis.
(a) the dilation map ds , which multiplies all vectors by the same scalar s
(b) the reflection map f` , which reflects all all vectors across a line ` through the
origin
X 1.22 Consider a linear transformation of R2 determined by these two.
1
1
2
0
1
0
1
0
0
5
1
1
B=h
,
i
1
1
2
1
D=h
,
i
2
1
(d) Using B from the prior item, represent the transformation with respect to
B, B.
1.23 Suppose that h : V W is nonsingular so that by Theorem 2.20, for any
~1 , . . . ,
~n i V the image h(B) = hh(
~1 ), . . . , h(
~n )i is a basis for
basis B = h
W.
(a) Represent the map h with respect to B, h(B).
(b) For a member ~v of the domain, where the representation of ~v has components
c1 , . . . , cn , represent the image vector h(~v ) with respect to the image basis h(B).
1.24 Give a formula for the product of a matrix and ~ei , the column vector that is
all zeroes except for a single one in the i-th position.
X 1.25 For each vector space of functions of one real variable, represent the derivative
transformation with respect
to B, B.
matrix.
1 0
0 0
a
b
(a)
(b)
(c) a matrix of the form
0 0
3 2
2a 2b
X 1.27 Can one matrix represent two different linear maps? That is, can RepB,D (h) =
RepB,
(h)?
D
1.28 Prove Theorem 1.4.
X 1.29 Example 1.8 shows how to represent rotation of all vectors in the plane through
an angle about the origin, with respect to the standard bases.
(a) Rotation of all vectors in three-space through an angle about the x-axis is a
transformation of R3 . Represent it with respect to the standard bases. Arrange
the rotation so that to someone whose feet are at the origin and whose head is
at (1, 0, 0), the movement appears clockwise.
204
(b) Repeat the prior item, only rotate about the y-axis instead. (Put the persons
head at ~e2 .)
(c) Repeat, about the z-axis.
(d) Extend the prior item to R4 . (Hint: rotate about the z-axis can be restated
as rotate parallel to the xy-plane.)
1.30 (Schurs Triangularization Lemma)
(a) Let U be a subspace of V and fix bases BU BV . What is the relationship
between the representation of a vector from U with respect to BU and the
representation of that vector (viewed as a member of V ) with respect to BV ?
(b) What about maps?
~1 , . . . ,
~n i for V and observe that the spans
(c) Fix a basis B = h
~1 }] [{
~1 ,
~2 }] [B] = V
[{~0}] = {~0} [{
form a strictly increasing chain of subspaces. Show that for any linear map
h : V W there is a chain W0 = {~0} W1 Wm = W of subspaces of
W such that
~1 , . . . ,
~i }]) Wi
h([{
for each i.
(d) Conclude that for every linear map h : V W there are bases B, D so the
matrix representing h with respect to B, D is upper-triangular (that is, each
entry hi,j with i > j is zero).
(e) Is an upper-triangular representation unique?
3.III.2
The prior subsection shows that the action of a linear map h is described by
a matrix H, with respect to appropriate bases, in this way.
h1,1 v1 + + h1,n vn
v1
h
..
~v = ... 7
= h(~v )
.
H
vn
hm,1 v1 + + hm,n vn
In this subsection, we will show the converse, that each matrix represents a
linear map.
Recall that, in the definition of the matrix representation of a linear map,
the number of columns of the matrix is the dimension of the maps domain and
the number of rows of the matrix is the dimension of the maps codomain. Thus,
for instance, a 23 matrix cannot represent a map from R5 to R4 . The next
result says that, beyond this restriction on the dimensions, there are no other
limitations: the 23 matrix represents a map from any three-dimensional space
to any two-dimensional space.
2.1 Theorem Any matrix represents a homomorphism between vector spaces
of appropriate dimensions, with respect to any pair of bases.
205
h1,1
h2,1
H=
hm,1
h1,2
h2,2
..
.
...
...
hm,2
...
h1,n
h2,n
hm,n
fix any n-dimensional domain space V and any m-dimensional codomain space
~1 , . . . , ~n i and D = h~1 , . . . , ~m i for those spaces.
W . Also fix bases B = h
Define a function h : V W by: where ~v in the domain is represented as
v1
..
RepB (~v ) = .
vn
h1,1 v1 + + h1,n vn
..
RepD ( h(~v ) ) =
.
hm,1 v1 + + hm,n vn
~1 + + vn
~n ) is defined to be (h1,1 v1 + + h1,n vn ) ~1 +
that is, h(~v ) = h(v1
+ (hm,1 v1 + + hm,n vn ) ~m . (This is well-defined by the uniqueness of the
representation RepB (~v ).)
Observe that h has simply been defined to make it the map that is represented with respect to B, D by the matrix H. So to finish, we need only check
that h is linear. If ~v , ~u V are such that
v1
u1
RepB (~v ) = ... and RepB (~u) = ...
vn
un
QED
2.2 Example Which map the matrix represents depends on which bases are
used. If
1
0
0
1
1 0
,
i, and B2 = D2 = h
,
i,
H=
, B1 = D1 = h
0
1
1
0
0 0
206
=
= 2
c2
c1 B
0 D
c2
2
These two are different. The first is projection onto the x axis, while the second
is projection onto the y axis.
So not only is any linear map described by a matrix but any matrix describes
a linear map. This means that we can, when convenient, handle linear maps
entirely as matrices, simply doing the computations, without have to worry that
a matrix of interest does not represent a linear map on some pair of spaces of
interest. (In practice, when we are working with a matrix but no spaces or
bases have been specified, we will often take the domain and codomain to be Rn
and Rm and use the standard bases. In this case, because the representation
is transparentthe representation with respect to the standard basis of ~v is
~v the column space of the matrix equals the range of the map. Consequently,
the column space of H is often denoted by R(H).)
With the theorem, we have characterized linear maps as those maps that act
in this matrix way. Each linear map is described by a matrix and each matrix
describes a linear map. We finish this section by illustrating how a matrix can
be used to tell things about its maps.
2.3 Theorem The rank of a matrix equals the rank of any map that it represents.
Proof. Suppose that the matrix H is mn. Fix domain and codomain spaces
~1 , . . . , ~n i and D. Then H
V and W of dimension n and m, with bases B = h
represents some linear map h between those spaces with respect to these bases
whose rangespace
~1 + + cn
~n ) c1 , . . . , cn R}
{h(~v ) ~v V } = {h(c1
~1 ) + + cn h(
~n ) c1 , . . . , cn R}
= {c1 h(
207
~1 ) + + cn h(
~n ) if and only if ~0 = c1 Rep (h(
~1 )) + + cn Rep (h(
~n )).
c1 h(
D
D
Hence, a subset of the rangespace is linearly independent if and only if the corresponding subset of the column space is linearly independent. This means that
the size of the largest linearly independent subset of the rangespace equals the
size of the largest linearly independent subset of the column space, and so the
QED
two spaces have the same dimension.
2.4 Example Any map represented
1
1
0
0
by
2
2
0
0
2
1
3
2
The above results end any confusion caused by our use of the word rank to
mean apparently different things when applied to matrices and when applied to
maps. We can also justify the dual use of nonsingular. Weve defined a matrix
to be nonsingular if it is square and is the matrix of coefficients of a linear system
with a unique solution, and weve defined a linear map to be nonsingular if it is
one-to-one.
2.6 Corollary A square matrix represents nonsingular maps if and only if it
is a nonsingular matrix. Thus, a matrix represents an isomorphism if and only
if it is square and nonsingular.
Proof. Immediate from the prior result.
QED
208
1
0
1 2
3 6
is not nonsingular because this matrix is not nonsingular.
Weve now seen that the relationship between maps and matrices goes both
ways: fixing bases, any linear map is represented by a matrix and any matrix
describes a linear map. That is, by fixing spaces and bases we get a correspondence between maps and matrices. In the rest of this chapter we will explore
this correspondence. For instance, weve defined for linear maps the operations
of addition and scalar multiplication and we shall see what the corresponding
matrix operations are. We shall also see the matrix operation that represent
the map operation of composition. And, we shall see how to find the matrix
that represents a maps inverse.
Exercises
X 2.9 Decide if the vector is in the column space of the matrix.
! !
1
1
1
2
2 1
1
4 8
0
1
1
1 , 0
(a)
,
(b)
,
(c)
2 5
3
2 4
1
1 1
1
0
X 2.10 Decide if each vector lies in the range of the map from R3 to R2 represented
with respect
to the
bases
standard
by the matrix.
1 1 3
1
2 0 3
1
(a)
,
(b)
,
0 1 4
3
4 0 6
1
X 2.11 Consider this matrix, representing a transformation of R2 , and these bases for
that space.
1
1
1
1
1
B=h
0
1
,
i
1
0
D=h
1
1
,
i
1
1
0
1
0
0
209
!
1 3 0
0 1 0
1 0 1
2.14 Example 2.8 gives a matrix that is nonsingular, and is therefore associated
with maps that are nonsingular.
(a) Find the set of column vectors representing the members of the nullspace of
any map represented by this matrix.
(b) Find the nullity of any such map.
(c) Find the set of column vectors representing the members of the rangespace
of any map represented by this matrix.
(d) Find the rank of any such map.
(e) Check that rank plus nullity equals the dimension of the domain.
X 2.15 Because the rank of a matrix equals the rank of any map it represents, if
1
0
0
0
3
0
0
2
0
1
1
0
1
0
3
1
2.20 The fact that for any linear map the rank plus the nullity equals the dimension
of the domain shows that a necessary condition for the existence of a homomorphism between two spaces, onto the second space, is that there be no gain in
dimension. That is, where h : V W is onto, the dimension of W must be less
than or equal to the dimension of V .
(a) Show that this (strong) converse holds: no gain in dimension implies that
210
H=
1
2
0
0
0
1
0
0
0
3.IV
211
Matrix Operations
The prior section shows how matrices represent linear maps. A good strategy, on
seeing a new idea, is to explore how it interacts with some already-established
ideas. In the first subsection we will ask how the representation of the sum
of two maps f + g related to the representations F and G of the two maps,
and how the representation of a scalar product r h of a map is related to the
representation H of that map. In later subsections we will see how to represent
map composition and map inverse.
3.IV.1
Recall that for two maps f and g with the same domain and codomain, the
map sum f + g has the natural definition.
f +g
~v 7 f (~v ) + g(~v )
The easiest way to see how the representations of the maps combine to represent
the map sum is with an example.
1.1 Example Suppose that f, g : R2 R3 are represented with respect to the
bases B and D by these matrices.
1 3
0
0
G = RepB,D (g) = 1 2
F = RepB,D (f ) = 2 0
1 0 B,D
2
4 B,D
Then, for any ~v V represented with respect to B, computation of the representation of f (~v ) + g(~v )
0v1 + 0v2
0
0
1v1 + 3v2
1 3
v
v
1
1
2 0
+ 1 2
= 2v1 + 0v2 + 1v1 2v2
v2
v2
1v1 + 0v2
2v1 + 4v2
2
4
1 0
gives this representation of f + g (~v ).
1v1 + 3v2
(1 + 0)v1 + (3 + 0)v2
(2 1)v1 + (0 2)v2 = 1v1 2v2
(1 + 2)v1 + (0 + 4)v2
3v1 + 4v2
Thus, the action of f + g is described by this matrix-vector product.
1v1 + 3v2
1 3
v1
1 2
= 1v1 2v2
v
3v1 + 4v2 D
3 4 B,D 2 B
This matrix is the entry-by-entry sum of original matrices, e.g., the 1, 1 entry
of RepB,D (f + g) is the sum of the 1, 1 entry of F and the 1, 1 entry of G.
212
1 0
v1
v1
so that ~v =
7
= t(~v )
RepB,D (t) =
v2 B
v1 + v2 D
1 1 B,D
then the scalar multiple map 5t acts in this way.
5v1
v1
7
= 5 t(~v )
~v =
v2 B
5v1 + 5v2 D
Therefore, this is the matrix representing 5t.
5 0
RepB,D (5t) =
5 5 B,D
1.3 Definition The sum of two same-sized matrices is their entry-by-entry
sum. The scalar multiple of a matrix is the result of entry-by-entry scalar
multiplication.
1.4 Remark These extend the vector addition and scalar multiplication operations that we defined in the first chapter.
1.5 Theorem Let h, g : V W be linear maps represented with respect to
bases B, D by the matrices H and G, and let r be a scalar. Then the map
h + g : V W is represented with respect to B, D by H + G, and the map
r h : V W is represented with respect to B, D by rH.
Proof. Exercise 8; generalize the examples above.
QED
0 0 0
Z=
0 0 0
no matter which domain and codomain bases are used.
Exercises
X 1.7 Perform
the indicated
operations,
if defined.
5 1 2
2 1 4
(a)
+
6
1
1
3 0 5
(b) 6
2
1
1
2
1
3
(c)
(d) 4
(e) 3
2
0
1
3
1
3
2
1
2
3
1
0
2
0
1
3
+5
+2
213
1
3
1
2
4
1
1
0
4
5
214
3.IV.2
Matrix Multiplication
QED
To see how the representation of the composite arises out of the representations
of the two compositors, consider an example.
2.2 Example Let h : R4 R2 and g : R2 R3 , fix bases B R4 , C R2 ,
D R3 , and let these be the representations.
1 1
4 6 8 2
G = RepC,D (g) = 0 1
H = RepB,C (h) =
5 7 9 3 B,C
1 0 C,D
To represent the composition g h : R4 R3 we fix a ~v , represent h of ~v , and
then represent g of that. The representation of h(~v ) is the product of hs matrix
and ~v s vector.
v1
v2
4 6 8 2
1 1
4v1 + 6v2 + 8v3 + 2v4
RepD ( g(h(~v )) ) = 0 1
5v1 + 7v2 + 9v3 + 3v4 C
1 0 C,D
215
v1
14+15 16+17 18+19 12+13
v2
= 0 4 + 1 5 0 6 + 1 7 0 8 + 1 9 0 2 + 1 3
v3
1 4 + 0 5 1 6 + 0 7 1 8 + 0 9 1 2 + 0 3 B,D
v4 D
Thus, the matrix representing gh has the rows of G combined with the columns
of H.
2.3 Definition The matrix-multiplicative product of the mr matrix G and
the rn matrix H is the mn matrix P , where
pi,j = gi,1 h1,j + gi,2 h2,j + + gi,r hr,j
that is, the i, j-th entry of the product is
the j-th column.
..
.
...
g
.
.
.
g
g
GH =
i,1
i,2
i,r
..
.
2.4 Example
14+15
0 4 + 1 5
14+05
. . .
...
=
hr,j
..
.
pi,j
..
.
. . .
2.5 Example
2 0
4 6 1
5
8 2
21+05 23+07
2
3
= 4 1 + 6 5 4 3 + 6 7 = 34
7
81+25 83+27
18
5
3
2
6
54
38
2.6 Theorem A composition of linear maps is represented by the matrix product of the representatives.
Proof. (This argument parallels Example 2.2.) Let h : V W and g : W X
be represented by H and G with respect to bases B V , C W , and D X,
of sizes n, r, and m. For any ~v V , the k-th component of RepC ( h(~v ) ) is
hk,1 v1 + + hk,n vn
and so the i-th component of RepD ( g h (~v ) ) is this.
gi,1 (h1,1 v1 + + h1,n vn ) + gi,2 (h2,1 v1 + + h2,n vn )
+ + gi,r (hr,1 v1 + + hr,n vn )
216
QED
Vw.r.t.
gh
C
g
G&
h%
H
GH
Xw.r.t.
Above the arrows, the maps show that the two ways of going from V to X,
straight over via the composition or else by way of W , have the same effect
gh
~v 7 g(h(~v ))
~v 7 h(~v ) 7 g(h(~v ))
(this is just the definition of composition). Below the arrows, the matrices
indicate that the product does the same thingmultiplying GH into the column
vector RepB (~v ) has the same effect as multiplying the column first by H and
then multiplying the result by G.
RepB,D (g h) = GH = RepC,D (g) RepB,C (h)
The definition of the matrix-matrix product operation does not restrict us
to view it as a representation of a linear map composition. We can get insight
into this operation by studying it as a mechanical procedure. The striking thing
is the way that rows and columns combine.
One aspect of that combination is that the sizes of the matrices involved is
significant. Briefly, mr times rn equals mn.
2.7 Example This product is not defined
1 2
0
0 0
0 10 1.1
0 2
because the number of columns on the left does not equal the number of rows
on the right.
217
In terms of the underlying maps, the fact that the sizes must match up reflects
the fact that matrix multiplication is defined only when a corresponding function
composition
g
1 ,j
+ gi,
2 ,j
+ + gi,
r ,j
the boxed subscripts on the gs are column indicators while those on the hs
indicate rows. That is, summation takes place over the columns of G but over
the rows of H; left is treated differently than right, so GH may be unequal to
HG. Matrix multiplication is not commutative.
2.9 Example Matrix multiplication hardly ever commutes. Test that by multiplying randomly chosen matrices both ways.
1 2
5 6
19 22
5 6
1 2
23 34
=
=
3 4
7 8
43 50
7 8
3 4
31 46
2.10 Example Commutativity can fail more dramatically:
5 6
1 2 0
23 34 0
=
7 8
3 4 0
31 46 0
while
1
3
2
4
0
5
0
7
6
8
218
219
We have now seen how the representation of the composition of two linear
maps is derived from the representations of the two maps. We have called
the combination the product of the two matrices. This operation is extremely
important. Before we go on to study how to represent the inverse of a linear
map, we will explore it some more in the next subsection.
Exercises
X 2.14 Compute, or state not defined.
(a)
3
4
(c)
2
7
1
2
7
4
0
0
0
1
8
1
3
X 2.15 Where
A=
5
0.5
1
2
(b)
1
4
5
1
4
1
0
1
0
(d)
B=
5
3
5
4
1
3
2 1 1!
2
1
3
3
1
3
2
4
1
1
2
5
C=
2
4
1
1
3
1
220
X 2.24 (a) How does matrix multiplication interact with scalar multiplication: is
r(GH) = (rG)H? Is G(rH) = r(GH)?
(b) How does matrix multiplication interact with linear combinations: is F (rG +
sH) = r(F G) + s(F H)? Is (rF + sG)H = rF H + sGH?
2.25 We can ask how the matrix product operation interacts with the transpose
operation.
(a) Show that (GH)trans = H trans Gtrans .
(b) A square matrix is symmetric if each i, j entry equals the j, i entry, that is,
if the matrix equals its own transpose. Show that the matrices HH trans and
H trans H are symmetric.
X 2.26 Rotation of vectors in R3 about an axis is a linear map. Show that linear
maps do not commute by showing geometrically that rotations do not commute.
2.27 In the proof of Theorem 2.12 some maps are used. What are the domains and
codomains?
2.28 How does matrix rank interact with matrix multiplication?
(a) Can the product of rank n matrices have rank less than n? Greater?
(b) Show that the rank of the product of two matrices is less than or equal to
the minimum of the rank of each factor.
2.29 Is commutes with an equivalence relation among nn matrices?
X 2.30 (This will be used in the Matrix Inverses exercises.) Here is another property
of matrix multiplication that might be puzzling at first sight.
(a) Prove that the composition of the projections x , y : R3 R3 onto the x
and y axes is the zero map despite that neither one is itself the zero map.
(b) Prove that the composition of the derivatives d2 /dx2 , d3 /dx3 : P4 P4 is
the zero map despite that neither is the zero map.
(c) Give a matrix equation representing the first fact.
(d) Give a matrix equation representing the second.
When two things multiply to give zero despite that neither is zero, each is said to
be a zero divisor.
2.31 Show that, for square matrices, (S + T )(S T ) need not equal S 2 T 2 .
X 2.32 Represent the identity transformation id : V V with respect to B, B for any
basis B. This is the identity matrix I. Show that this matrix plays the role in
matrix multiplication that the number 1 plays in real number multiplication: HI =
IH = H (for all matrices H for which the product is defined).
2.33 In real number algebra, quadratic equations have at most two solutions. That
is not so with matrix algebra. Show that the 2 2 matrix equation T 2 = I has
more than two solutions, where I is the identity matrix (this matrix has ones in
its 1, 1 and 2, 2 entries and zeroes elsewhere; see Exercise 32).
2.34 (a) Prove that for any 22 matrix T there are scalars c0 , . . . , c4 such that the
combination c4 T 4 + c3 T 3 + c2 T 2 + c1 T + I is the zero matrix (where I is the 22
identity matrix, with ones in its 1, 1 and 2, 2 entries and zeroes elsewhere; see
Exercise 32).
(b) Let p(x) be a polynomial p(x) = cn xn + + c1 x + c0 . If T is a square
matrix we define p(T ) to be the matrix cn T n + + c1 T + I (where I is the
appropriately-sized identity matrix). Prove that for any square matrix there is
a polynomial such that p(T ) is the zero matrix.
(c) The minimal polynomial m(x) of a square matrix is the polynomial of least
degree, and with leading coefficient 1, such that m(T ) is the zero matrix. Find
221
3/2
1/2
1/2
3/2
(This is the representation with respect to E2 , E2 , the standard basis, of a rotation
through /6 radians counterclockwise.)
2.35 The infinite-dimensional space P of all finite-degree polynomials gives a memorable example of the non-commutativity of linear maps. Let d/dx : P P be the
usual derivative and let s : P P be the shift map.
s
a0 + a1 x + + an xn 7 0 + a0 x + a1 x2 + + an xn+1
Show that the two maps dont commute d/dx s 6= s d/dx; in fact, not only is
(d/dx s) (s d/dx) not the zero map, it is the identity map.
2.36 Recall the notation for the sum of the sequence of numbers a1 , a2 , . . . , an .
n
X
a i = a 1 + a 2 + + an
i=1
r
X
gi,k hk,j
k=1
3.IV.3
1 1
4 6
0 1 5 7
1 0
8
9
! 9 13
2
= 5 7
3
4 6
17
9
8
5
3
2
We can view this as the left matrix acting by multiplying its rows, one at a time,
into the columns of the right matrix. Of course, another perspective is that the
right matrix uses its columns to act on the left matrixs rows. Below, we will
examine actions from the left and from the right for some simple matrices.
The first case, the action of a zero matrix, is very easy.
222
0 0
0 0
1 3
1 1
2
0
=
1
0
0
0
0
0
2
1
3
0
4
0
0
0 0
=
0
0 0
0
0
0
1
5
0
7
0
7
6
= 0
8
0
8
0
0
Acting from the left, an i, j unit matrix copies row j of the multiplicand into
row i of the result. From the right an i, j unit matrix copies column i of the
multiplicand into column j of the result.
1
4
7
2
5
8
3
0
6 0
9
0
1
0
0 = 0
0
0
1
4
7
3.4 Example Rescaling these matrices simply rescales the result. This is the
action from the left of the matrix that is twice the one in the prior example.
0 2
14
5
6
0 0
=0
7 8
0 0
0
16
0
0
And this is the action of the matrix that is minus three times the one from the
prior example.
1 2 3
0 3
0 3
4 5 6 0 0 = 0 12
7 8 9
0 0
0 21
Next in complication are matrices with two nonzero entries. There are two
cases. If a left-multiplier has entries in different rows then their actions dont
interact.
1 0 0
1 2 3
1
0 0 2 4 5 6 = (0
0 0 0
7 8 9
0
1
= 0
0
1
= 14
0
223
0
0 0
0 + 0 0
0
0 0
2 3
0
0
0 0 + 14 16
0 0
0
0
2
3
16 18
0
0
0
0
0
0
1 2 3
2) 4 5 6
0
7 8 9
0
18
0
But if the left-multipliers nonzero entries are in the same row then that row of
the result is a combination.
3.6 Example
1 0 2
1 2 3
1
0 0 0 4 5 6 = (0
0 0 0
7 8 9
0
1
= 0
0
15
=0
0
0
0 0
0 + 0 0
0
0 0
2 3
14 16
0 0 + 0
0
0 0
0
0
18 21
0
0
0
0
0
0
0
2
1 2 3
0) 4 5 6
0
7 8 9
18
0
0
..
..
..
..
.
.
.
.
~
~
G
~hn
G ~hn
= G h 1
h1
..
..
..
..
.
.
.
.
and the rows of GH are formed by taking the rows of G times H
~g1 H
~g1
..
..
H =
.
.
~gr
~gr H
(ignoring the extra parentheses).
224
Proof. We will exhibit the 22 case, and leave the general case as an exercise.
GH =
g1,1
g2,1
g1,2
g2,2
h1,1
h2,1
h1,2
h2,2
h1,2
g1,1 h1,1 + g1,2 h2,1
h1,1
G
=
G
h2,1
h2,2
g2,1 h1,1 + g2,2 h2,1
is indeed the same as the right side of GH, except for the extra parentheses (the
ones marking the columns as column vectors). The other equation is similarly
QED
easy to recognize.
An application of those observations is that there is a matrix that just copies
out the rows and columns.
3.8 Definition The main diagonal (or principle diagonal or diagonal) of a
square matrix goes from the upper left to the lower right.
3.9 Definition An identity matrix is
cept for ones in the main diagonal.
1
0
Inn =
0
3.10 Example The 33
the left
1 0
0 1
0 0
and from the right.
2 3
1 3
7 1
0
0
0
2 3
0 1 3
1
7 1
6
1
8 0
0
0
0
1
0
6
2
8 = 1
0
7
3 6
3 8
1 0
0
2
0 = 1
1
7
3 6
3 8
1 0
1 2
1 2
0 2 1 0
0 2
1 1 0 1 = 1 1
4 3
4 3
225
0
...
0
a1,1
0
0
a2,2 . . .
..
.
0
0
. . . an,n
3.13 Example From the left, the action of multiplication by a diagonal matrix
is to rescales the rows.
2 0
2 1 4 1
4 2
8 2
=
0 1
1 3 4 4
1 3 4 4
From the right such a matrix rescales the columns.
1
2
3
1
0
2
0
2
2
0 0
3 4
2 0 =
6 4
0 2
2
4
The second generalization of identity matrices is that we can put a single one
in each row and column in ways other than putting them down the diagonal.
3.14 Definition A permutation matrix is square and is all zeros except for a
single one in each row and column.
3.15 Example From the left these matrices permute rows.
0
1
0
0
0
1
1
1
0 4
0
7
2
5
8
3
7
6 = 1
9
4
8
2
5
9
3
6
1
2
0 = 5
0
8
3
6
9
1
4
7
1
4
7
2
5
8
3
0
6 1
9
0
0
0
1
226
1 0 0
0
2
1 1
0 2
0 3 0 0 1/3 1 1 = 0 1
0 0 1
1
0
2 0
1 0
1 1
3 3
2 0
We have seen how to produce a matrix that will swap rows. Multiplying by this
permutation matrix swaps the first and third rows.
0 0 1
0 2 1 1
1 0 2 0
0 1 0 0 1 3 3 = 0 1 3 3
1 0 0
1 0 2 0
0 2 1 1
To see how to perform a pivot, we observe something about those two examples. The matrix that rescales the second row by a factor of three arises in
this way from the identity.
1 0 0
1 0 0
32
0 1 0 0 3 0
0 0 1
0 0 1
Similarly, the matrix that swaps first and third
0
1 0 0
1 3
0 1 0 0
1
0 0 1
3.17 Example The 33
1
0
0
will, when it acts
1
0
0
0 1
1 0
0 0
1 0
0 0
22 +3
1 0 0 1
0 2
0 1
0 0
1
1 0 0
2 1
0
0
0
1
22 + 3 .
0
3
7
3.18 Definition The elementary reduction matrices are obtained from identity
matrices with one Gaussian operation. We denote them:
ki
227
QED
3.20 Example This is the first system, from the first chapter, on which we
performed Gauss method.
3x3 = 9
x1 + 5x2 2x3 = 2
=3
(1/3)x1 + 2x2
It can be reduced with matrix
0 0 1
0
0 1 0 1
1/3
1 0 0
triple the first row,
3 0
0 1
0 0
and then add 1
1
1
0
0
1/3
0 1
0
1
1/3 2 0 3
0 3 9
5 2 2 = 1
5 2 2
2 0 3
0
0 3 9
1
3
2 = 1
9
0
2 0
5 2
0 3
0 0
1 6 0
1 0 1 5 2
0 0 3
0 1
6 0
5 2
0 3
the second.
1 6
0
9
2 = 0 1 2
9
0 0
3
9
2
9
9
7
9
1 6 0
9
1 0
0
1 6
0
0 1 0 0 1 2 7 = 0 1 2
9
0 0
3
0 0 1
0 0 1/3
and to finish, clear the third column and
1 6 0
1 0 0
1
0 1 0 0 1 2 0
0
0 0 1
0 0 1
9
7
3
1 0 0 3
6 0 9
1 2 7 = 0 1 0 1
0 1 3
0 0 1 3
228
We have observed the following result, which we shall use in the next subsection.
3.22 Corollary For any matrix H there are elementary reduction matrices
R1 , . . . , Rr such that Rr Rr1 R1 H is in reduced echelon form.
Until now we have taken the point of view that our primary objects of study
are vector spaces and the maps between them, and have adopted matrices only
for computational convenience. This subsection show that this point of view
isnt the whole story. Matrix theory is a fascinating and fruitful area.
In the rest of this book we shall continue to focus on maps as the primary
objects, but we will be pragmaticif the matrix point of view gives some clearer
idea then we shall use it.
Exercises
X 3.23 Predict the result of each multiplication by an elementary reduction matrix,
and then
by multiplying
itout.
check
3 0
1 2
4 0
1 2
1
0
1 2
(a)
(b)
(c)
0 0
3 4
0 2
3 4
2 1
3 4
1 2
1 1
1 2
0 1
(e)
3 4
0
1
3 4
1 0
X 3.24 The need to take linear combinations of rows and columns in tables of numbers
arises often in practice. For instance, this is a map of part of Vermont and New
York.
(d)
Swanton
In part because of Lake Champlain,
there are no roads connecting some
pairs of towns. For instance, there
is no way to go from Winooski to
Grand Isle without going through
Colchester. (Of course, many other
roads and towns have been left off
to simplify the graph. From top to
bottom of this map is about forty
miles.)
Grand Isle
Colchester
Winooski
Burlington
(a) The incidence matrix of a map is the square matrix whose i, j entry is the
number of roads from city i to city j. Produce the incidence matrix of this map
(take the cities in alphabetical order).
(b) A matrix is symmetric if it equals its transpose. Show that an incidence
matrix is symmetric. (These are all two-way streets. Vermont doesnt have
many one-way streets.)
(c) What is the significance of the square of the incidence matrix? The cube?
229
X 3.25 The need to take linear combinations of rows and columns in tables of numbers
arises often in practice. For instance, this table gives the number of hours of each
type done by each worker, and the associated pay rates. Use matrices to compute
the wages due.
regular overtime
wage
Alan
40
12
regular
$25.00
Betty
overtime $45.00
35
6
Catherine
40
18
Donald
28
0
3.26 Find the product of this matrix with its transpose.
cos sin
sin
cos
X 3.27 Prove that the diagonal matrices form a subspace of Mnn . What is its
dimension?
3.28 Does the identity matrix represent the identity map if the bases are unequal?
3.29 Show that every multiple of the identity commutes with every square matrix.
Are there other matrices that commute with all square matrices?
3.30 Prove or disprove: nonsingular matrices commute.
X 3.31 Show that the product of a permutation matrix and its transpose is an identity
matrix.
3.32 Show that if the first and second rows of G are equal then so are the first and
second rows of GH. Generalize.
3.33 Describe the product of two diagonal matrices.
3.34 Write
1
0
3 3
as the product of two elementary reduction matrices.
X 3.35 Show that if G has a row of zeros then GH (if defined) has a row of zeros.
Does that work for columns?
3.36 Show that the set of unit matrices forms a basis for Mnm .
3.37 Find the formula for the n-th power of this matrix.
1 1
1 0
X 3.38 The trace of a square matrix is the sum of the entries on its diagonal (its
significance appears in Chapter Five). Show that trace(GH) = trace(HG).
X 3.39 A square matrix is upper triangular if its only nonzero entries lie above, or
on, the diagonal. Show that the product of two upper triangular matrices is upper
triangular. Does this hold for lower triangular also?
3.40 A square matrix is a Markov matrix if each entry is between zero and one
and the sum along each row is one. Prove that a product of Markov matrices is
Markov.
X 3.41 Give an example of two matrices of the same rank with squares of differing
rank.
3.42 Combine the two generalizations of the identity matrix, the one allowing entires to be other than ones, and the one allowing the single one in each row and
column to be off the diagonal. What is the action of this type of matrix?
230
3.43 On a computer multiplications are more costly than additions, so people are
interested in reducing the number of multiplications used to compute a matrix
product.
(a) How many real number multiplications are needed in formula we gave for the
product of a mr matrix and a rn matrix?
(b) Matrix multiplication is associative, so all associations yield the same result.
The cost in number of multiplications, however, varies. Find the association
requiring the fewest real number multiplications to compute the matrix product
of a 510 matrix, a 1020 matrix, a 205 matrix, and a 51 matrix.
(c) (Very hard.) Find a way to multiply two 2 2 matrices using only seven
multiplications instead of the eight suggested by the naive approach.
3.44 [Putnam, 1990, A-5] If A and B are square matrices of the same size such
that ABAB = 0, does it follow that BABA = 0?
3.45 [Am. Math. Mon., Dec. 1966] Demonstrate these four assertions to get an alternate proof that column rank equals row rank.
(a) ~
y~
y = ~0 iff ~
y = ~0.
(b) A~
x = ~0 iff Atrans A~
x = ~0.
(c) dim(R(A)) = dim(R(Atrans A)).
(d) col rank(A) = col rank(Atrans ) = row rank(A).
3.46 [Ackerson] Prove (where A is an nn matrix and so defines a transformation
of any n-dimensional space V with respect to B, B where B is a basis) dim(R(A)
N (A)) = dim(R(A)) dim(R(A2 )). Conclude
(a) N (A) R(A) iff dim(N (A)) = dim(R(A)) dim(R(A2 ));
(b) R(A) N (A) iff A2 = 0;
(c) R(A) = N (A) iff A2 = 0 and dim(N (A)) = dim(R(A)) ;
(d) dim(R(A) N (A)) = 0 iff dim(R(A)) = dim(R(A2 )) ;
(e) (Requires the Direct Sum subsection, which is optional.) V = R(A) N (A)
iff dim(R(A)) = dim(R(A2 )).
3.IV.4
Inverses
x
x
7 y
y
0
231
x
x
7 y 7
:
y
y
0
We say is a left inverse map of or, what is the same thing, that is a right
inverse map of . However, composition in the other order doesnt give the
identity maphere is a vector that is not sent to itself under this composition.
0
0
0 7
7 0
:
0
0
1
In fact, the projection has no left inverse at all. For, if f were to be a left
inverse of then we would have
x
x
y 7 x 7f y
f :
y
z
z
for all of the infinitely many zs. But no function f can send a single argument
to more than one value.
(An example of a function with no inverse on either side, is the zero transformation on R2 .) Some functions have a two-sided inverse map, another function
that is the inverse of the first, both from the left and from the right. For instance, the map given by ~v 7 2 ~v has the two-sided inverse ~v 7 (1/2) ~v . In
this subsection we will focus on two-sided inverses. The appendix shows that a
function (linear or not) has a two-sided inverse if and only if it is both one-to-one
and onto. The appendix also shows that if a function f has a two-sided inverse
then it is unique, and so it is called the inverse, and is denoted f 1 . So our
purpose in this subsection is, where a linear map h has an inverse, to find the
relationship between RepB,D (h) and RepD,B (h1 ).
4.2 Definition A matrix G is a left inverse matrix of the matrix H if GH is
the identity matrix. It is a right inverse matrix if HG is the identity. A matrix
H with a two-sided inverse is an invertible matrix. That two-sided inverse is
called the inverse matrix and is denoted H 1 .
Because of the correspondence between linear maps and matrices, statements
about map inverses translate into statements about matrix inverses.
4.3 Lemma If a matrix has both a left inverse and a right inverse then the
two are equal.
4.4 Theorem A matrix is invertible if and only if it is nonsingular.
232
Proof. (For both results.) Given a matrix H, fix spaces of appropriate dimension for the domain and codomain. Fix bases for these spaces. With respect to
these bases, H represents a map h. The statements are true about the map and
QED
therefore they are true about the matrix.
Here is the arrow diagram giving the relationship between map inverses and
matrix inverses. It is a special case of the diagram for function composition and
matrix multiplication.
Ww.r.t.
h%
H
Vw.r.t.
&h
H 1
id
Vw.r.t.
Beyond its place in our general program of seeing how to represent map
operations, another reason for our interest in inverses comes from solving linear
systems. A linear system is equivalent to a matrix equation, as here.
1 1
x1
3
x1 + x2 = 3
()
=
2 1
x2
2x1 x2 = 2
2
By fixing spaces and bases (e.g., R2 , R2 and E2 , E2 ), we take the matrix H to
represent some map h. Then solving the system is the same as asking: what
domain vector ~x is mapped by h to the result d~ ? If we could invert h then we
~ to get Rep (~x).
could solve the system by multiplying RepD,B (h1 ) RepD (d)
B
4.6 Example We can find a left inverse for the matrix just given
m n
1 1
1 0
=
p q
2 1
0 1
by using Gauss method to solve the resulting linear system.
m + 2n
m n
=1
=0
p + 2q = 0
p q=1
x
1/3 1/3
3
5/3
=
=
y
2/3 1/3
2
4/3
233
4.7 Remark Why solve systems this way, when Gauss method takes less
arithmetic (this assertion can be made precise by counting the number of arithmetic operations, as computer algorithm designers do)? Beyond its conceptual
appeal of fitting into our program of discovering how to represent the various
map operations, solving linear systems by using the matrix inverse has at least
two advantages.
First, once the work of finding an inverse has been done, solving a system
with the same coefficients but different constants is easy and fast: if we change
the entries on the right of the system () then we get a related problem
1 1
x
5
=
2 1
y
1
wtih a related solution method.
x
1/3
=
y
2/3
1/3
1/3
5
2
=
1
3
3.01
1 1
x1
=
x2
2
2 1
can be solved with the inverse.
1/3 1/3
3.01
(1/3)(3.01) + (1/3)(2)
=
2/3 1/3
2
(2/3)(3.01) (1/3)(2)
to show that x1 changes by 1/3 of the tweak while x2 moves by 2/3 of that
tweak. This sort of analysis is used, for example, to decide how accurately data
must be specified in a linear model to ensure that the solution has a desired
accuracy.
We finish by describing the computational procedure usually used to find
the inverse matrix.
4.8 Lemma A matrix is invertible if and only if it can be written as the product of elementary reduction matrices. The inverse can be computed by applying
to the identity matrix the same row steps, in the same order, as are used to
Gauss-Jordan reduce the invertible matrix.
Proof. A matrix H is invertible if and only if it is nonsingular and thus GaussJordan reduces to the identity. By Corollary 3.22 this reduction can be done
with elementary matrices Rr Rr1 . . . R1 H = I. This equation gives the two
halves of the result.
234
First, elementary matrices are invertible and their inverses are also elemen1
, etc.,
tary. Applying Rr1 to the left of both sides of that equation, then Rr1
1
1
gives H as the product of elementary matrices H = R1 Rr I (the I is
here to cover the trivial r = 0 case).
Second, matrix inverses are unique and so comparison of the above equation
with H 1 H = I shows that H 1 = Rr Rr1 . . . R1 I. Therefore, applying R1
to the identity, followed by R2 , etc., yields the inverse of H.
QED
4.9 Example To find the inverse of
1
2
1
1
21 +2
1 1
1 0
1 1 1 0
2 1 0 1
0 3 2 1
1/32
1 1
1
0
0 1 2/3 1/3
2 +1
1 0 1/3 1/3
0 1 2/3 1/3
This calculation has found the inverse.
1
1/3
1 1
=
2/3
2 1
4.10 Example This
0 3 1 1
1 0
1 0
1 1 0 0
1/3
1/3
1 0
1 0 1 0
0 0
1 2
0 3 1 1 0 0
1 0
0 1
1 1 0 0 0 1
1 0
1 0 1 0
1 +3
0 3 1 1 0 0
0 1 1 0 1 1
..
.
4.11 Example A non-invertible matrix is detected by the fact that the left
half wont reduce to the identity.
1 1 1 0 21 +2 1 1 1 0
2 2 0 1
0 0 2 1
235
This procedure will find the inverse of a general nn matrix. The 22 case
is handy.
4.12 Corollary The inverse for a 22 matrix exists and equals
a b
c d
1
=
1
ad bc
d b
c a
if and only if ad bc 6= 0.
Proof. This computation is Exercise 22.
QED
1
1 3
3 1
2 1/2
2
4
0
2 4
(a)
(b)
(c)
(d)
0 2
3
1
1
2
1 1 0
!
0
1
5
2
2
3
4
(e) 0 2
(f ) 1 2 3
2
3
2
4 2 3
X 4.17 What matrix has this one for its inverse?
1 3
2 5
236
4.18 How does the inverse operation interact with scalar multiplication and addition of matrices?
(a) What is the inverse of rH?
(b) Is (H + G)1 = H 1 + G1 ?
X 4.19 Is (T k )1 = (T 1 )k ?
4.20 Is H 1 invertible?
4.21 For each real number let t : R2 R2 be represented with respect to the
standard bases by this matrix.
cos
sin
sin
cos
1 0 1
H=
0 1 0
X
X
has infinitely many right inverses. Show also that it has no left inverse.
4.24 In Example 4.1, how many left inverses has ?
4.25 If a matrix has infinitely many right-inverses, can it have infinitely many
left-inverses? Must it have?
4.26 Assume that H is invertible and that HG is the zero matrix. Show that G is
a zero matrix.
4.27 Prove that if H is invertible then the inverse commutes with a matrix GH 1 =
H 1 G if and only if H itself commutes with that matrix GH = HG.
4.28 Show that if T is square and if T 4 is the zero matrix then (I T )1 =
I + T + T 2 + T 3 . Generalize.
4.29 Let D be diagonal. Describe D2 , D3 , . . . , etc. Describe D1 , D2 , . . . , etc.
Define D0 appropriately.
4.30 Prove that any matrix row-equivalent to an invertible matrix is also invertible.
4.31 The first question below appeared as Exercise 28.
(a) Show that the rank of the product of two matrices is less than or equal to
the minimum of the rank of each.
(b) Show that if T and S are square then T S = I if and only if ST = I.
4.32 Show that the inverse of a permutation matrix is its transpose.
4.33 The first two parts of this question appeared as Exercise 25.
(a) Show that (GH)trans = H trans Gtrans .
(b) A square matrix is symmetric if each i, j entry equals the j, i entry (that is, if
the matrix equals its transpose). Show that the matrices HH trans and H trans H
are symmetric.
(c) Show that the inverse of the transpose is the transpose of the inverse.
(d) Show that the inverse of a symmetric matrix is symmetric.
4.34 The items starting this question appeared as Exercise 30.
(a) Prove that the composition of the projections x , y : R3 R3 is the zero
map despite that neither is the zero map.
(b) Prove that the composition of the derivatives d2 /dx2 , d3 /dx3 : P4 P4 is
the zero map despite that neither map is the zero map.
(c) Give matrix equations representing each of the prior two items.
237
When two things multiply to give zero despite that neither is zero, each is said to
be a zero divisor. Prove that no zero divisor is invertible.
4.35 In real number algebra, there are exactly two numbers, 1 and 1, that are
their own multiplicative inverse. Does H 2 = I have exactly two solutions for 22
matrices?
4.36 Is the relation is a two-sided inverse of transitive? Reflexive? Symmetric?
4.37 [Am. Math. Mon., Nov. 1951] Prove: if the sum of the elements of a square
matrix is k, then the sum of the elements in each row of the inverse matrix is 1/k.
238
3.V
Change of Basis
Representations, whether of vectors or of maps, vary with the bases. For instance, with respect to the two bases E2 and
1
1
B=h
,
i
1
1
for R2 , the vector ~e1 has two different representations.
1
1/2
RepB (~e1 ) =
RepE2 (~e1 ) =
0
1/2
Similarly, with respect to E2 , E2 and E2 , B, the identity map has two different
representations.
1 0
1/2 1/2
RepE2 ,B (id) =
RepE2 ,E2 (id) =
0 1
1/2 1/2
With our point of view that the objects of our studies are vectors and maps, in
fixing bases we are adopting a scheme of tags or names for these objects, that
are convienent for computation. We will now see how to translate among these
nameswe will see exactly how representations vary as the bases vary.
3.V.1
In converting RepB (~v ) to RepD (~v ) the underlying vector ~v doesnt change.
Thus, this translation is accomplished by the identity map on the space, described so that the domain space vectors are represented with respect to B and
the codomain space vectors are represented with respect to D.
Vw.r.t.
idy
Vw.r.t.
(The diagram is vertical to fit with the ones in the next subsection.)
1.1 Definition The change of basis matrix for bases B, D V is the representation of the identity map id : V V with respect to those bases.
..
..
.
.
~
~n )
RepB,D (id) =
RepD (
RepD (1 )
..
..
.
.
239
1.2 Lemma Left-multiplication by the change of basis matrix for B, D converts a representation with respect to B to one with respect to D. Conversly, if
left-multiplication by a matrix changes bases M RepB (~v ) = RepD (~v ) then M
is a change of basis matrix.
Proof. For the first sentence, for each ~
v , as matrix-vector multiplication represents a map application, RepB,D (id) RepB (~v ) = RepD ( id(~v ) ) = RepD (~v ). For
the second sentence, with respect to B, D the matrix M represents some linear
QED
map, whose action is ~v 7 ~v , and is therefore the identity map.
1/2
2
)) =
RepD ( id(
1
3/2 D
1/2
1
RepD ( id(
)) =
0
1/2 D
RepB,D (id) =
1/2
3/2
1/2
1/2
We can see this matrix at work by finding the two representations of ~e2
0
1
0
1/2
RepB (
)=
RepD (
)=
1
2
1
1/2
and checking that the conversion goes as expected.
1/2 1/2
1
1/2
=
3/2
1/2
2
1/2
We finish this subsection by recognizing that the change of basis matrices
are familiar.
1.4 Lemma A matrix changes bases if and only if it is nonsingular.
Proof. For one direction, if left-multiplication by a matrix changes bases then
the matrix represents an invertible function, simply because the function is
inverted by changing the bases back. Such a matrix is itself invertible, and so
nonsingular.
To finish, we will show that any nonsingular matrix M performs a change of
basis operation from any given starting basis B to some ending basis. Because
the matrix is nonsingular, it will Gauss-Jordan reduce to the identity, so there
are elementatry reduction matrices such that Rr R1 M = I. Elementary
matrices are invertible and their inverses are also elementary, so multiplying
from the left first by Rr 1 , then by Rr1 1 , etc., gives M as a product of
240
241
Exercises
X 1.6 In R2 , where
2
2
,
i
1
4
find the change of basis matrices from D to E2 and from E2 to D. Multiply the
two.
X 1.7 Find the change of basis matrix for B, D R2 .
1
1
(a) B = E2 , D = h~e2 , ~e1 i
(b) B = E2 , D = h
,
i
2
4
D=h
1
1
1
2
0
1
(c) B = h
,
i, D = E2
(d) B = h
,
i, D = h
,
i
2
4
1
2
4
3
1.8 For the bases in Exercise 7, find the change of basis matrix in the other direction,
from D to B.
X 1.9 Find the change of basis matrix for each B, D P2 .
(a) B = h1, x, x2 i, D = hx2 , 1, xi
(b) B = h1, x, x2 i, D = h1, 1+x, 1+x+x2 i
2
2
(c) B = h2, 2x, x i, D = h1 + x , 1 x2 , x + x2 i
2
X 1.10 Decide
ifeach changes
bases
on R . To what basis
is E2 changed?
5 0
2 1
1
4
1 1
(a)
(b)
(c)
(d)
0 4
3 1
2
8
1
1
1.11 Find bases such that this matrix represents the identity map with respect to
those bases.
!
3
1
4
2 1 1
0
0
4
1.12 Conside the vector space of real-valued functions with basis hsin(x), cos(x)i.
Show that h2 sin(x)+cos(x), 3 cos(x)i is also a basis for this space. Find the change
of basis matrix in each direction.
1.13 Where does this matrix
cos(2)
sin(2)
sin(2) cos(2)
send the standard basis for R2 ? Any other bases? Hint. Consider the inverse.
X 1.14 What is the change of basis matrix with respect to B, B?
1.15 Prove that a matrix changes bases if and only if it is invertible.
1.16 Finish the proof of Lemma 1.4.
X 1.17 Let H be a nn nonsingular matrix. What basis of Rn does H change to the
standard basis?
X 1.18 (a) In P3 with basis B = h1 + x, 1 x, x2 + x3 , x2 x3 i we have this represenatation.
0
1
2
3
RepB (1 x + 3x x ) =
1
2 B
Find a basis D giving this different representation for the same polynomial.
1
0
2
3
RepD (1 x + 3x x ) =
2
0 D
242
(b) State and prove that any nonzero vector representation can be changed to
any other.
Hint. The proof of Lemma 1.4 is constructiveit not only says the bases change,
it shows how they change.
be bases for V and D, D
be bases for
1.19 Let V, W be vector spaces, and let B, B
W . Where h : V W is linear, find a formula relating RepB,D (h) to RepB,
D
(h).
X 1.20 Show that the columns of an n n change of basis matrix form a basis for
Rn . Do all bases appear in that way: can the vectors from any Rn basis make the
columns of a change of basis matrix?
X 1.21 Find a matrix having this effect.
1
3
4
1
That is, find a M that left-multiplies the starting vector to yield the ending vector.
Is therea
matrix
having
these
two
effects?
1
1
2
1
1
1
2
1
(a)
7
7
(b)
7
7
3
1
1
1
3
1
6
1
Give a necessary and sufficient condition for there to be a matrix such that ~v1 7 w
~1
and ~v2 7 w
~ 2.
3.V.2
The first subsection shows how to convert the representation of a vector with
respect to one basis to the representation of that same vector with respect to
another basis. Here we will see how to convert the representation of a map with
respect to one pair of bases to the representation of that map with respect to
a different pair. That is, we want the relationship between the matrices in this
arrow diagram.
Vw.r.t.
idy
Vw.r.t.
Ww.r.t.
H
idy
Ww.r.t.
To move from the lower-left of this diagram to the lower-right we can either go
straight over, or else up to VB then over to WD and then down. Restated in
= Rep (h) either by simply using
terms of the matrices, we can calculate H
B,D
and D,
or else by first changing bases with Rep (id) then multiplying
B
B,B
by H = RepB,D (h) and then changing bases with RepD,D (id). This equation
summarizes.
= Rep (id) H Rep (id)
H
B,B
D,D
()
243
(To compare this equation with the sentence before it, remember that the equation is read from right to left because function composition is read right to left
and matrix multiplication represent the composition.)
2.1 Example The matrix
cos(/6)
T =
sin(/6)
sin(/6)
cos(/6)
3/2
1/2
1/2
3/2
1
3
(3 + 3)/2
(1 + 3 3)/2
idy
R2w.r.t.
E2
R2w.r.t.
T
idy
E2
R2w.r.t.
T
Note that RepE2 ,D (id) can be calculated as the matrix inverse of RepD,E
2 (id).
1 2
1 0
3/2
1/2
RepB,
D
(t) =
0 3
1 2
3/2
1/2
(5 3)/6 (3 +2 3)/3
=
3/3
(1 + 3)/6
Although the new matrix is messier-appearing, the map that it represents is the
same. For instance, to replicate the effect of t in the picture, start with B,
1
1
)=
RepB (
3
1 B
apply T,
(5 3)/6
(1 + 3)/6
(11 + 3 3)/6
(3 +2 3)/3
1
=
3/3
(1 + 3 3)/6 D
D
1 B
B,
244
1+3 3
11 + 3 3
1
2
(3 + 3)/2
=
0
3
(1 + 3 3)/2
6
6
to see that it is the same result as above.
2.2 Example On R3 the map
y+z
x
t
x + z
y 7
x+y
z
that is represented with respect to the standard
0 1
RepE3 ,E3 (t) = 1 0
1 1
1
1
0
1 0 0
1
1
1
if B = h1 , 1 , 1i
then RepB,B (t) = 0 1 0
0
0 2
0
2
1
in a way that is simpler, in that the action of a diagonal matrix is easy to
understand.
Naturally, we usually prefer basis changes that make the representation easier to understand. When the representation with respect to equal starting and
ending bases is a diagonal matrix we say the map or matrix has been diagonalized. In Chaper Five we shall see which maps and matrices are diagonalizable,
and where one is not, we shall see how to get a representation that is nearly
diagonal.
We finish this subsection by considering the easier case where representations are with respect to possibly different starting and ending bases. Recall
that the prior subsection shows that a matrix changes bases if and only if it
is nonsingular. That gives us another version of the above arrow diagram and
equation ().
are matrix equivalent if there
2.3 Definition Same-sized matrices H and H
245
%
.H
...
.
$
H
All matrices:
H matrix equivalent
to H
We can get some insight into the classes by comparing matrix equivalence with
row equivalence (recall that matrices are row equivalent when they can be re = P HQ, the matrices P and
duced to each other by row operations). In H
Q are nonsingular and thus each can be written as a product of elementary
reduction matrices (Lemma 4.8). Left-multiplication by the reduction matrices
making up P has the effect of performing row operations. Right-multiplication
by the reduction matrices making up Q performs column operations. Therefore,
matrix equivalence is a generalization of row equivalencetwo matrices are row
equivalent if one can be converted to the other by a sequence of row reduction
steps, while two matrices are matrix equivalent if one can be converted to the
other by a sequence of row reduction steps followed by a sequence of column
reduction steps.
Thus, if matrices are row equivalent then they are also matrix equivalent
(since we can take Q to be the identity matrix and so perform no column
operations). The converse, however, does not hold.
2.5 Example These two
1
0
0
0
1
0
1
0
are matrix equivalent because the second can be reduced to the first by the
column operation of taking 1 times the first column and adding to the second.
They are not row equivalent because they have different reduced echelon forms
(in fact, both are already in reduced form).
We will close this section by finding a set of representatives for the matrix
equivalence classes.
2.6 Theorem Any mn matrix of rank k is matrix equivalent to the mn
matrix that is all zeros except that the first k diagonal entries are ones.
1 0 ... 0 0 ... 0
0 1 . . . 0 0 . . . 0
..
0 0 . . . 1 0 . . . 0
0 0 . . . 0 0 . . . 0
..
.
0
0 ...
0 ...
246
I Z
Z Z
Proof. As discussed above, Gauss-Jordan reduce the given matrix and combine
all the reduction matrices used there to make P . Then use the leading entries to
do column reduction and finish by swapping columns to put the leading ones on
the diagonal. Combine the reduction matrices used for those column operations
QED
into Q.
1 2
0 0
2 4
1 1
1 1
2 2
1 1 0
1 0 0
1 2 1
0 1 0 0 1 0 0 0 1
0 0 1
2 0 1
2 4 2
1
1 2 0
1 = 0 0 1
2
0 0 0
0
1
0
1 2 0 0
1 0 0 0
1 2 0 0
1 0 0 0
0 1 0 0
0
1
0
0
0 0 1 1
0 0 1 0 0 0 1 1 = 0 0 1 0
0 0 0 0
0 0 0 0
0 0 0 1
0 0 0 1
Finish by swapping columns.
1
1 0 0 0
0 0 1 0 0
0
0 0 0 0
0
0
0
1
0
0
1
0
0
0
1 0 0 0
0
= 0 1 0 0
0
0 0 0 0
1
1 0 2
1 1 0
1 2 1 1
0 0 1
0
1 0 0 0 1 1
0 1 0
2 0 1
2 4 2 2
0 0 0
0
1 0 0 0
0
= 0 1 0 0
1
0 0 0 0
1
2.8 Corollary Two same-sized matrices are matrix equivalent if and only if
they have the same rank. That is, the matrix equivalence classes are characterized by rank.
Proof. Two same-sized matrices with the same rank are equivalent to the same
QED
block partial-identity matrix.
247
2.9 Example Now that we know that the block partial-identity matrices form
canonical representatives of the matrix-equivalence classes, we can see what the
classes look like and how many classes there are. Consider the 22 matrices.
There are only three possible ranks: zero, one, or two. Thus the 22 matrices
fall into three matrix-equivalence classes.
All 22
matrices:
0 0 .
? 10 01
%
1 0
.
0 0
.
.
0 0
the three
equivalence
classes
Each class just consists of all the 22 matrices with the same rank.
In this subsection we have seen how to change the representation of a map
with respect to a first pair of bases to one with respect to a second pair. That
led to a definition describing when matrices are equivalent in this way. Finally
we noted that, with the proper choice of (possibly different) starting and ending
bases, any map can be represented in block partial-identity form.
One of the nice things about this representation is that, in some sense, we
can completely understand the map when it is expressed in this way: if the
~1 , . . . , ~n i and D = h~1 , . . . , ~m i then the map sends
bases are B = h
~1 + + ck
~k + ck+1
~k+1 + + cn
~n 7 c1~1 + + ck ~k + ~0 + + ~0
c1
where k is the maps rank. Thus, we can understand any linear map as a kind
of projection.
c1
c1
..
..
.
.
ck
7 ck
0
ck+1
.
.
.
..
.
cn B
0 D
Of course, understanding a map expressed in this way requires that we understand the relationship between B and D. However, despite that difficulty,
this is a good classification of linear maps.
Exercises
X 2.10 Decide
if these
matrices are
matrix equivalent.
1 3 0
2 2
1
(a)
,
2 3 0
0 5 1
(b)
0
1
3
,
1
4
0
0
5
248
1 3
1
3
,
2 6
2 6
X 2.11 Find the canonical representative of the matrix-equivalence class of each matrix.
0 1 0
2
2 1 0
4
(a)
(b) 1 1 0
4 2 0
3 3 3 1
2.12 Suppose that, with respect to
1
1
B = E2
D=h
,
i
1
1
(c)
1 2
3 4
Use change
of
basis
torepresent
t with respect to each pair.
matrices
0
1
1
2
=h
=h
(a) B
,
i, D
,
i
1
1
0
1
X
X
X
X
X
X
= h 1 , 1 i, D
=h 1 , 2 i
(b) B
2
0
2
1
2.13 What size are P and Q?
2.14 Use Theorem 2.6 to show that a square matrix is nonsingular if and only if it
is equivalent to an identity matrix.
2.15 Show that, where A is a nonsingular square matrix, if P and Q are nonsingular
square matrices such that P AQ = I then QP = A1 .
2.16 Why does Theorem 2.6 not show that every matrix is diagonalizable (see
Example 2.2)?
2.17 Must matrix equivalent matrices have matrix equivalent transposes?
2.18 What happens in Theorem 2.6 if k = 0?
2.19 Show that matrix-equivalence is an equivalence relation.
2.20 Show that a zero matrix is alone in its matrix equivalence class. Are there
other matrices like that?
2.21 What are the matrix equivalence classes of matrices of transformations on R1 ?
R3 ?
2.22 How many matrix equivalence classes are there?
2.23 Are matrix equivalence classes closed under scalar multiplication? Addition?
2.24 Let t : Rn Rn represented by T with respect to En , En .
(a) Find RepB,B (t) in this specific case.
T =
1
3
1
1
1
1
B=h
,
i
2
1
~1 , . . . ,
~n i.
(b) Describe RepB,B (t) in the general case where B = h
2.25 (a) Let V have bases B1 and B2 and suppose that W has the basis D. Where
h : V W , find the formula that computes RepB2 ,D (h) from RepB1 ,D (h).
(b) Repeat the prior question with one basis for V and two bases for W .
2.26 (a) If two matrices are matrix-equivalent and invertible, must their inverses
be matrix-equivalent?
(b) If two matrices have matrix-equivalent inverses, must the two be matrixequivalent?
249
(c) If two matrices are square and matrix-equivalent, must their squares be
matrix-equivalent?
(d) If two matrices are square and have matrix-equivalent squares, must they be
matrix-equivalent?
X 2.27 Square matrices are similar if they represent the same transformation, but
each with respect to the same ending as starting basis. That is, RepB1 ,B1 (t) is
similar to RepB2 ,B2 (t).
(a) Give a definition of matrix similarity like that of Definition 2.3.
(b) Prove that similar matrices are matrix equivalent.
(c) Show that similarity is an equivalence relation.
(d) Show that if T is similar to T then T 2 is similar to T2 , the cubes are similar,
etc. Contrast with the prior exercise.
(e) Prove that there are matrix equivalent matrices that are not similar.
250
3.VI
Projection
The prior section describes the matrix equivalence canonical form as expressing a
projection and so this section takes the natural next step of studying projections.
However, this section is optional; only the last two sections of Chapter Five
require this material. In addition, this section requires some optional material
from the subsection on length and angle measure in n-space.
We have described the projection from R3 into its xy plane subspace as
a shadow map. This shows why, but it also shows that some shadows fall
upward.
1
2
2
1
2
0
1
2
0
1
2
1
So perhaps a better description is: the projection of ~v is the p~ in the plane with
the property that someone standing on p~ and looking straight up or down sees
~v . In this section we will generalize this to other projections, both orthogonal
(i.e., straight up and down) and nonorthogonal.
3.VI.1
The picture shows someone who has walked out on the line until the tip of
~v is straight overhead. That is, where the line is described as the span of
some nonzero vector ` = {c ~s c R}, the person has walked out to find the
coefficient cp~ with the property that ~v cp~ ~s is orthogonal to cp~ ~s.
251
~
v
~
v cp
s
~~
cp
s
~~
We can solve for this coefficient by noting that because ~v cp~~s is orthogonal to
a scalar multiple of ~s it must be orthogonal to ~s itself, and then the consequent
fact that the dot product (~v cp~~s) ~s is zero gives that cp~ = ~v ~s/~s ~s.
1.1 Definition The orthogonal projection of ~v into the line spanned by a
nonzero ~s is this vector.
proj[~s ] (~v ) =
~v ~s
~s
~s ~s
Exercise 19 checks that the outcome of the calculation depends only on the line
and not on which vector ~s happens to be used to describe that line.
1.2 Remark The wording of that definition says spanned by ~s instead the
more formal the span of the set {~s }. This casual first phrase is common.
1.3 Example In R2 , to orthogonally project into the line y = 2x, we first pick
a direction vector for this line. For instance,
1
~s =
2
will do. With that, calculation of a projection is routine.
2
1
2
~
v=
3
2
3
1 = 8 1 = 8/5
1
2
1
2
16/5
x
y
z
0
1
0
0
1
0
0
0
1 = y
0
0
0
1
0
252
The picture above with the stick figure walking out on the line until ~v s tip
is overhead is one way to think of the orthogonal projection of a vector into a
line. We finish this subsection with two other ways.
1.5 Example A railroad car left on an east-west track without its brake is
pushed by a wind blowing toward the northeast at fifteen miles per hour; what
speed will the car reach?
For the wind we use a vector of length 15 that points toward the northeast.
p
15p1/2
~v =
15 1/2
The car can only be affected by the part of the wind blowing in the east-west
directionthe part of ~v in the direction of the x-axis is this (the picture has the
same perspective as the railroad car picture above).
p
north
15 1/2
p~ =
0
p
~
east
p
So the car will reach a velocity of 15 1/2 miles per hour toward the east.
Thus, another way to think of the picture that precedes the definition is that
it shows ~v as decomposed into two parts, the part with the line (here, the part
with the tracks, p~), and the part that is orthogonal to the line (shown here lying
on the north-south axis). These two are not interacting or independent, in
the sense that the east-west car is not at all affected by the north-south part
of the wind (see Exercise 11). So the orthogonal projection of ~v into the line
spanned by ~s can be thought of as the part of ~v that lies in the direction of ~s.
Finally, another useful way to think of the orthogonal projection is to have
the person stand not on the line, but on the vector that is to be projected to the
line. This person has a rope over the line and pulls it tight, naturally making
the rope orthogonal to the line.
253
That is, we can think of the projection p~ as being the vector in the line that is
closest to ~v (see Exercise 17).
1.6 Example A submarine is tracking a ship moving along the line y = 3x+2.
Torpedo range is one-half mile. Can the sub stay where it is, at the origin on
the chart below, or must it move to reach a place where the ship will pass within
range?
north
east
The formula for projection into a line does not immediately apply because the
line doesnt pass through the origin, and so isnt the span of any ~s. To adjust
for this, we start by shifting the entire map down two units. Now the line is
y = 3x, which is a subspace, and we can project to get the point p~ of closest
approach, the point on the line through the origin closest to
0
~v =
2
the subs shifted position.
1
0
3
2
1
3/5
=
p~ =
3
9/5
1
1
3
3
The distance between ~v and p~ is approximately 0.63 miles and so the sub must
move to get in range.
This subsection has developed a natural projection map, orthogonal projection into a line. As suggested by the examples, it is often called for in applications. The next subsection shows how the definition of orthogonal projection
into a line gives us a way to calculate especially convienent bases for vector
spaces, again something that is common in applications. The final subsection
completely generalizes projection, orthogonal or not, into any subspace at all.
Exercises
X 1.7 Project the first vector orthogonally into the line spanned by the second vector.
! !
! !
1
1
1
3
2
3
2
3
2
(a)
,
(b)
,
(c) 1 ,
(d) 1 , 3
1
2
1
0
4
1
4
12
X 1.8 Project the vector orthogonally into the line.
254
2
3
c R} (b) 1 , the line y = 3x
(a) 1 , {c 1
1
4
3
1.9 Although the development of Definition 1.1 is guided by the pictures, we are
not restricted to spaces that we can draw. In R4 project this vector into this line.
1
1
` = {c c R}
1
1
1
2
~v =
1
3
X 1.10 Definition 1.1 uses two vectors ~s and ~v . Consider the transformation of R2
resulting from fixing
~s =
3
1
and projecting ~v into the line that is the span of ~s. Apply it to these vectors.
1
0
(a)
(b)
2
4
Show that in general the projection tranformation is this.
x1
x2
X
X
255
first one in a way that is bound to the coordinates (that is, it fixes a basis and
then computes) and the second in a way that is more conceptual.
(a) Produce a matrix that describes the functions action.
(b) Show also that this map can be obtained by first rotating everything in the
plane /4 radians clockwise, then projecting into the x-axis, and then rotating
/4 radians counterclockwise.
1.21 For ~a, ~b Rn let ~v1 be the projection of ~a into the line spanned by ~b, let ~v2 be
the projection of ~v1 into the line spanned by ~a, let ~v3 be the projection of ~v2 into
the line spanned by ~b, etc., back and forth between the spans of ~a and ~b. That is,
~vi+1 is the projection of ~vi into the span of ~a if i + 1 is even, and into the span of ~b
if i + 1 is odd. Must that sequence of vectors eventually settle downmust there
be a sufficiently large i such that ~vi+2 equals ~vi and ~vi+3 equals ~vi+1 ? If so, what
is the earliest such i?
3.VI.2
Gram-Schmidt Orthogonalization
This subsection is optional. It requires material from the prior, also optional,
subsection. The work done here will only be needed in the final two sections of
Chapter Five.
The prior subsection suggests that projecting into the line spanned by ~s
decomposes a vector ~v into two parts
~
v
~
v proj[~s ] (~
v)
proj[~s ] (~
v)
that are orthogonal and so are not interacting. We now make that phrase
precise.
2.1 Definition Vectors ~v1 , . . . , ~vk Rn are mutually orthogonal when any two
are orthogonal: if i 6= j then the dot product ~vi ~vj is zero.
2.2 Theorem If the vectors in a set {~v1 , . . . , ~vk } Rn are mutually orthogonal
and nonzero then that set is linearly independent.
Proof. Consider a linear relationship c1~
v1 + c2~v2 + + ck~vk = ~0. If i [1..k]
then taking the dot product of ~vi with both sides of the equation
~vi (c1~v1 + c2~v2 + + ck~vk ) = ~vi ~0
ci (~vi ~vi ) = 0
shows, since ~vi is nonzero, that ci is zero.
QED
256
Of course, the converse of Corollary 2.3 does not holdnot every basis of
every subspace of Rn is made of mutually orthogonal vectors. However, we can
get the partial converse that for every subspace of Rn there is at least one basis
consisting of mutually orthogonal vectors.
~2 of this basis for R2 are not orthogonal.
~1 and
2.4 Example The members
B=h
~2
4
1
,
i
2
3
~1
However, we can derive from B a new basis for the same space that does have
mutually orthogonal members. For the first member of the new basis we simply
~1 .
use
4
~1 =
2
~2 its part in the
For the second member of the new basis, we take away from
direction of ~1 ,
~2 =
1
3
proj[~1 ] (
1
)=
3
1
3
2
1
1
2
~
2
257
0
0
0
2/3
2/3
~2 = 2 proj[~1 ] (2) = 2 2/3 = 4/3
0
0
0
2/3
2/3
Finally, we get ~3 by taking the third given vector and subtracting the part of
it in the direction of ~1 , and also the part of it in the direction of ~2 .
1
1
1
1
~3 = 0 proj[~1 ] (0) proj[~2 ] (0) = 0
3
3
3
1
Again the corollary gives that
1
2/3
1
h1 , 4/3 , 0 i
1
2/3
1
is a basis for the space.
The next result verifies that the process used in those examples works with
any basis for any subspace of an Rn (we are restricted to Rn only because we
have not given a definition of orthogonality for other vector spaces).
~1 , . . .
~k i is a basis
2.7 Theorem (Gram-Schmidt orthogonalization) If h
n
for a subspace of R then, where
~1
~1 =
~2 proj
~2 =
[~
1 ] (2 )
~3 proj (
~
~
~3 =
[~
1 ] 3 ) proj[~
2 ] (3 )
..
.
~k proj (
~
~
~k =
[~
1 ] k ) proj[~
k1 ] (k )
the ~ s form an orthogonal basis for the same subspace.
Proof. We will use induction to check that each ~
i is nonzero, is in the span of
~i i and is orthogonal to all preceding vectors: ~1 ~i = = ~i1 ~i = 0.
~1 , . . .
h
With those, and with Corollary 2.3, we will have that h~1 , . . . ~k i is a basis for
~k i.
~1 , . . .
the same space as h
We shall cover the cases up to i = 3, which give the sense of the argument.
Completing the details is Exercise 23.
~1 makes it a nonzero vector
The i = 1 case is trivialsetting ~1 equal to
~1 is a member of a basis, it is obviously in the desired span, and the
since
orthogonal to all preceding vectors condition is vacuously met.
258
~2 ~1
~
~2 2 ~1
~1
~1 =
~1 ~1
~1 ~1
This expansion shows that ~2 is nonzero or else this would be a non-trivial linear
~ (it is nontrivial because the coefficient of
~2 is 1) and
dependence among the s
also shows that ~2 is in the desired span. Finally, ~2 is orthogonal to the only
preceding vector
~2 proj (
~
~1 ~2 = ~1 (
[~
1 ] 2 )) = 0
because this projection is orthogonal.
The i = 3 case is the same as the i = 2 case except for one detail. As in the
i = 2 case, expanding the definition
~
~3 3
~3 =
~1
~
~3 3
=
~1
~3
~1
~1
~1
~2
~3
~1 ~
1
~1
~2
~2
~2
~2
~2 ~1
~2 ~
~1
2
~2
~1 ~1
~
~
~3 proj (
~1 ~3 = ~1
[~
1 ] 3 ) proj[~
2 ] (3 )
~
~3 )
~3 proj (
1 proj[~2 ] (
= ~1
[~
1 ] 3 ) ~
=0
(Heres the difference from the i = 2 casethe second line has two kinds of
terms. The first term is zero because this projection is orthogonal, as in the
i = 2 case. The second term is zero because ~1 is orthogonal to ~2 and so is
orthogonal to any vector in the line spanned by ~2 .) The check that ~3 is also
QED
orthogonal to the other preceding vector ~2 is similar.
Beyond having the vectors in the basis be orthogonal, we can do more; we
can arrange for each vector to have length one by dividing each by its own length
(we can normalize the lengths).
2.8 Example Normalizing the length of each vector in the orthogonal basis of
Example 2.6 produces this orthonormal basis.
1/ 6
1/3
1/ 2
h1/3 , 2/ 6 , 0 i
1/ 2
1/ 3
1/ 6
Besides its intuitive appeal, and its analogy with the standard basis En for Rn ,
an orthonormal basis also simplifies some computations. See Exercise 17, for
example.
Exercises
2.9 Perform the Gram-Schmidt process on each of these bases for R2 .
259
1
2
0
1
0
1
,
i
(b) h
,
i
(c) h
,
i
1
1
1
3
1
0
Then turn those orthogonal bases into orthonormal bases.
X 2.10 Perform the Gram-Schmidt process on each of these bases for R3 .
! ! !
! ! !
2
1
0
1
0
2
(a) h 2 , 0 , 3 i
(b) h 1 , 1 , 3 i
2
1
1
0
0
1
Then turn those orthogonal bases into orthonormal bases.
X 2.11 Find an orthonormal basis for this subspace of R3 : the plane x y + z = 0.
2.12 Find an orthonormal basis for this subspace of R4 .
x
y
{ x y z + w = 0 and x + z = 0}
z
w
(a) h
2.13 Show that any linearly independent subset of Rn can be orthogonalized without changing its span.
X 2.14 What happens if we apply the Gram-Schmidt process to a basis that is already
orthogonal?
2.15 Let h~1 , . . . , ~k i be a set of mutually orthogonal vectors in Rn .
(a) Prove that for any ~v in the space, the vector ~v (proj[~1 ] (~v )+ +proj[~vk ] (~v ))
is orthogonal to each of ~1 , . . . , ~k .
(b) Illustrate the prior item in R3 by using ~e1 as ~1 , using ~e2 as ~2 , and taking
~v to have components 1, 2, and 3.
(c) Show that proj[~1 ] (~v ) + + proj[~vk ] (~v ) is the vector in the span of the set
of ~s that is closest to ~v . Hint. To the illustration done for the prior part,
add a vector d1~1 + d2~2 and apply the Pythagorean Theorem to the resulting
triangle.
2.16 Find a vector in R3 that is orthogonal to both of these.
!
!
1
2
5
2
1
0
X 2.17 One advantage of orthogonal bases is that they simplify finding the representation of a vector with respect to that basis.
(a) For this vector and this non-orthogonal basis for R2
2
1
1
B=h
,
i
3
1
0
first represent the vector with respect to the basis. Then project the vector into
~1 ] and [
~2 ].
the span of each basis vector [
(b) With this orthogonal basis for R2
~v =
K=h
1
1
,
i
1
1
represent the same vector ~v with respect to the basis. Then project the vector
into the span of each basis vector. Note that the coefficients in the representation
and the projection are the same.
(c) Let K = h~1 , . . . , ~k i be an orthogonal basis for some subspace of Rn . Prove
that for any ~v in the subspace, the i-th component of the representation RepK (~v )
is the scalar coefficient (~v ~i )/(~i ~i ) from proj[~i ] (~v ).
260
B2 = {~e1 , ~e2 }
3.VI.3
This subsection, like the others in this section, is optional. It also requires
material from the optional earlier subsection on Direct Sums.
The prior subsections project a vector into a line by decomposing it into two
parts: the part in the line proj[~s ] (~v ) and the rest ~v proj[~s ] (~v ). To generalize
projection to arbitrary subspaces, we follow this idea.
3.1 Definition For any direct sum V = M N and any ~v V , the projection
of ~v into M along N is
~
projM,N (~v ) = m
where ~v = m
~ + ~n with m
~ M, ~n N .
261
0 0
a b
c, d R}
a, b R}
N ={
M ={
c d
0 0
To project
A=
3
0
1
4
0 1
0
1 0
0 0
,
i
BN = h
,
BM = h
0 0
0 0
1 0
0
The concatenation of these
B = BM
1 0
0
BN = h
,
0 0
0
1
0
,
0
1
0
0
,
0
0
0
i
1
0
i
1
is a basis for the entire space, because the space is the direct sum, so we can
use it to represent A.
3 1
1 0
0 1
0 0
0 0
=3
+1
+0
+4
0 0
0 0
1 0
0 1
0 4
Now the projection of A into M along N is found by keeping the M part of this
sum and dropping the N part.
3 1
1 0
0 1
3 1
)=3
+1
=
projM,N (
0 4
0 0
0 0
0 0
3.3 Example Both subscripts on projM,N (~v ) are significant. The first subscript M matters because the result of the projection is an m
~ M , and changing
this subspace would change the possible results. For an example showing that
the second subscript matters, fix this plane subspace of R3 and its basis
1
0
x
BM = h0 , 2i
M = {y y 2z = 0}
0
1
z
and compare the projections along two different subspaces.
0
0
= {k 1 k R}
N
N = {k 0 k R}
2
1
262
1
0
2
projM,N (~v ) = 2 0 + (9/5) 2 = 18/5
0
1
9/5
Therefore projection along different subspaces may yield different results.
These pictures compare the two maps. Both show that the projection is
indeed into the plane and along the line.
263
M
Notice that the projection along N is not orthogonalthere are members of the
plane M that are not orthogonal to the dotted line. But the projection along
is orthogonal.
N
A natural question is: what is the relationship between the projection operation defined above, and the operation of orthogonal projection into a line?
The second picture above suggests the answerorthogonal projection into a
line is a special case of the projection defined above; it is just projection along
a subspace perpendicular to the line.
N
M
264
0
3
2
cos =
10+13+02
so 0.94 rad
1+1+0 0+9+4
Instead M is the z-axis, since proceeding as in the prior example and taking
the natural basis for the xy-plane gives this.
x
x
x
1
0
0
0
y =
} = {y x = 0 and y = 0}
M = {y
0 1 0
0
z
z
z
The two examples that weve seen since Definition 3.4 illustrate the first
sentence in that definition. The next result justifies the second sentence.
3.7 Lemma Let M be a subspace of Rn . The orthogonal complement of M is
also a subspace. The space is the direct sum of the two Rn = M M . And,
for any ~v Rn , the vector ~v projM (~v ) is perpendicular to every vector in M .
Proof. First, the orthogonal complement M is a subspace of Rn because, as
265
Exercise 17 from the prior subsection proves this about any orthogonal basis: each vector ~v in the space is the sum of its orthogonal projections onto the
lines spanned by the basis vectors.
~v = proj[~1 ] (~v ) + + proj[~n ] (~v )
()
To check this, represent the vector ~v = r1~1 + + rn~n , apply ~i to both sides
~v ~i = (r1~1 + + rn~n ) ~i = r1 0 + + ri (~i ~i ) + + rn 0, and
solve to get ri = (~v ~i )/(~i ~i ), as desired.
Since obviously any member of the span of h~k+1 , . . . , ~n i is orthogonal to
any vector in M , to show that this is a basis for M we need only show the
other containmentthat any w
~ M is in the span of this basis. The prior
paragraph does this. On projections into basis vectors from M , any w
~ M
~
~
gives proj[~1 ] (w
~ ) = 0, . . . , proj[~k ] (w
~ ) = 0 and therefore () gives that w
~ is a
linear combination of ~k+1 , . . . , ~n . Thus this is a basis for M and Rn is the
direct sum of the two.
The final sentence is proved in much the same way. Write ~v = proj[~1 ] (~v ) +
+ proj[~n ] (~v ). Then projM (~v ) is gotten by keeping only the M part and
dropping the M part projM (~v ) = proj[~k+1 ] (~v ) + + proj[~k ] (~v ). Therefore
~v projM (~v ) consists of a linear combination of elements of M and so is
QED
perpendicular to every vector in M .
We can find the orthogonal projection into a subspace by following the steps
of the proof, but the next result gives a convienent formula.
3.8 Theorem Let ~v be a vector in Rn and let M be a subspace of Rn
~1 , . . . ,
~k i. If A is the matrix whose columns are the s
~ then
with basis h
~
~
projM (~v ) = c1 1 + + ck k where the coefficients ci are the entries of the
vector (Atrans A)Atrans ~v . That is, projM (~v ) = A(Atrans A)1 Atrans ~v .
Proof. The vector projM (~
v ) is a member of M and so it is a linear combination
~1 + + ck
~k . Since As columns are the s,
~ that can
of basis vectors c1
k
be expressed as: there is a ~c R such that projM (~v ) = A~c (this is expressed
compactly with matrix multiplication as in Example 3.5 and 3.6). Because
~v projM (~v ) is perpendicular to each member of the basis, we have this (again,
expressed compactly).
A
~v
~c = Atrans A
gives the formula for the projection matrix as projM (~v ) = A ~c.
3.9 Example To orthogonally project this vector into this subspace
1
x
~v = 1
P = {y x + z = 0}
1
z
QED
266
0
A = 1
0
1
0
1
0 1
trans 1 trans
0
1
1 0
A
A
= 1 0
A A
1/2 0
0 1
0 1
1/2 0 1/2
1
0
= 0
1/2 0 1/2
1
0
With the matrix, calculating the orthogonal projection of any vector into P is
easy.
1/2 0 1/2
1
0
1
0 1 = 1
projP (~v ) = 0
1/2 0 1/2
1
0
Exercises
X 3.10 Project
the vectors into
M along N .
3
x
(a)
, M ={
x + y = 0},
2
y
N ={
x
y
x 2y = 0}
1
x
x
(b)
, M ={
x y = 0}, N = {
2x + y = 0}
2
y
y
!
!
!
3
x
1
(c) 0 , M = { y x + y = 0}, N = {c 0 c R}
1
z
1
X 3.11 Find M .
x
x
(a) M = {
x + y = 0}
(b) M = {
2x + 3y = 0}
y
y
x
x
x y = 0}
(d) M = {~0 }
(e) M = {
x = 0}
y
y
!
!
x
x
(f ) M = { y
x + 3y + z = 0}
(g) M = { y x = 0 and y + z = 0}
z
z
3.12 This subsection shows how to project orthogonally in two ways, the method of
Example 3.2 and 3.3, and the method of Theorem 3.8. To compare them, consider
the plane P specified by 3x + 2y z = 0 in R3 .
(a) Find a basis for P .
(b) Find P and a basis for P .
(c) Represent this vector with respect to the concatenation of the two bases from
the prior item.
!
1
~v = 1
2
(c) M = {
267
(d) Find the orthogonal projection of ~v into P by keeping only the P part from
the prior item.
(e) Check that against the result from applying Theorem 3.8.
X 3.13 We have three ways to find the orthogonal projection of a vector into a line,
the Definition 1.1 way from the first subsection of this section, the Example 3.2
and 3.3 way of representing the vector with respect to a basis for the space and
then keeping the M part, and the way of Theorem 3.8. For these cases, do all
three ways.
1
x
(a) ~v =
, M ={
x + y = 0}
3
y
!
!
0
x
(b) ~v = 1 , M = { y x + z = 0 and y = 0}
2
z
3.14 Check that the operation of Definition 3.1 is well-defined. That is, in Example 3.2 and 3.3, doesnt the answer depend on the choice of bases?
3.15 What is the orthogonal projection into the trivial subspace?
3.16 What is the projection of ~v into M along N if ~v M ?
3.17 Show that if M Rn is a subspace with orthonormal basis h~1 , . . . , ~n i then
the orthogonal projection of ~v into M is this.
(~v ~1 ) ~1 + + (~v ~n ) ~n
X 3.18 Prove that the map p : V V is the projection into M along N if and only
if the map id p is the projection into N along M . (Recall the definition of the
difference of two maps: (id p) (~v ) = id(~v ) p(~v ) = ~v p(~v ).)
X 3.19 Show that if a vector is perpendicular to every vector in a set then it is
perpendicular to every vector in the span of that set.
3.20 True or false: the intersection of a subspace and its orthogonal complement is
trivial.
3.21 Show that the dimensions of orthogonal complements add to the dimension
of the entire space.
X 3.22 Suppose that ~v1 , ~v2 Rn are such that for all complements M, N Rn , the
projections of ~v1 and ~v2 into M along N are equal. Must ~v1 equal ~v2 ? (If so, what
if we relax the condition to: all orthogonal projections of the two are equal?)
X 3.23 Let M, N be subspaces of Rn . The perp operator acts on subspaces; we can
ask how it interacts with other such operations.
(a) Show that two perps cancel: (M ) = M .
(b) Prove that M N implies that N M .
(c) Show that (M + N ) = M N .
X 3.24 The material in this subsection allows us to express a geometric relationship
that we have not yet seen between the rangespace and the nullspace of a linear
map.
(a) Represent f : R3 R given by
!
v1
v2
v3
268
!
v1
v2
v3
~
~i ) = i i = 1, 2, . . . , r
t(
~0 i = r + 1, r + 2, . . . , n
where r is the rank of t.
(d) Conclude that every projection is a projection along a subspace.
(e) Also conclude that every projection has a representation
RepB,B (t) =
I
Z
Z
Z
269
30
16
60
34
90
51
Naturally, because of randomness, the exact proportion is not found with this
sample indeed, there is no solution to this system.
30m = 16
60m = 34
90m = 51
That is, the vector of experimental data is not in the subspace of solutions.
16
30
34 6 {m 60 m R}
51
90
However, as described above, we expect that there is an m that nearly works.
An orthogonal projection of the data vector into the line subspace gives our best
guess at m.
30
16
34 60
30
30
90
51
7110
60 =
60
12600
30
30
90
90
60 60
90
90
The estimate (m = 7110/12600 0.56) is higher than 1/2, but not by much, so
probably the penny is fair enough for flipping purposes.
The line with the slope m 0.56 is called the line of best fit for this data.
heads
60
c
b
c
b
30
c
b
30
60
90
flips
270
Minimizing the distance between the given vector and the vector used as the
right-hand side minimizes the total of these vertical lengths (these have been
distorted, exaggerated by a factor of ten, to make them more visible).
c
b
c
b
c
b
Because it involves minimizing this total distance, we say that the line has been
obtained through fitting by least-squares.
In the previous example, the line that we use, whose slope is our best guess
of the true ratio of heads to flips, must pass through (0, 0). We can also handle
cases where the line is not required to pass through the origin.
For example, the different denominations of U.S. money have different average times in circulation (the $2 bill is left off as a special case). How long should
we expect a $25 bill to last?
denomination
average life (years)
1
1.5
5
2
10
3
20
5
50
9
100
20
The plot (see below) looks roughly linear. It isnt a perfect line, i.e., the linear
system with equations b + 1m = 1.5, . . . , b + 100m = 20 has no solution, but
we can again use orthogonal projection to find a best approximation. Consider
the matrix of coefficients of that linear system and also its vector of constants,
the experimentally-determined values.
1.5
1
1
2
1
5
3
1 10
~
v
=
A=
5
1 20
9
1 50
20
1 100
The ending result in the subsection on Projection into a Subspace says that
coefficients b and m so that the linear combination of the columns of A is as
close as possible to the vector ~v are the entries of (Atrans A)1 Atrans ~v . Some
calculation gives an intercept of b = 1.05 and a slope of m = 0.18.
avg life
c
b
15
c
b
5
c
c b
b
c
b
10
c
b
30
50
70
90
denom
Plugging x = 25 into the equation of the line shows that such a bill should last
between five and six years.
271
We close with an example [Oakley & Baker] that cautions about overusing
least-squares fitting. These are the world record times for the mens mile race
that were in force on January first of the given years. We want to project when
a 3:40 mile will be run.
year
seconds
1
1
A = ...
1
1
280.0
1860
268.8
1870
..
..
~
v
=
.
.
229.0
1980
226.32
1990
bc
270
bc
bc
bc
bc
c
b
250
c
b
c
b
c
b
c
b
230
1870
1890
1910
1930
1950
c
b
c
b
1970
c
b
c
b
1990
year
When will a 220 second mile be run? Solving 220 = 970.68 0.37x gives an
estimate of the year 2027.
This example is amusing, but serves as a caution because the linearity of the
data will break down someday the tool of fitting by orthogonal projection
should be applied judicioulsy.
Exercises
The calculations here are most practically done on a computer. In addition, some
of the problems require more data, available in your library, on the net, or in the
Answers to the Exercises.
1 Use least-squares to judge if the coin in this experiment is fair.
flips 8 16 24 32 40
heads 4
9
13 17 20
2 For the mens mile record, rather than give each of the many records and its
exact date, weve smoothed the data somewhat by taking a periodic sample. Do
the longer calculation and compare the conclusions.
3 Find the line of best fit for the mens 1500 meter run. How does the slope compare
with that for the mens mile (the distances are close; a mile is about 1609 meters)?
4 Find the line of best fit for the records for womens mile.
5 Do the lines of best fit for the mens and womens miles cross?
272
6 When the space shuttle Challenger exploded in 1986, one of the criticisms made of
NASAs decision to launch was in the way the analysis of number of O-ring failures
versus temperature was made (of course, O-ring failure caused the explosion). Four
O-ring failures will cause the rocket to explode. NASA had data from 24 previous
flights.
temp F
53 75 57 58 63 70 70 66 67 67 67
failures
3
2
1
1
1
1
1
0
0
0
0
68 69 70 70 72 73 75 76 76 78 79 80
0
0
0
0
0
0
0
0
0
0
0
0
81
0
Venus
0.72
Earth
1.00
Mars
1.52
Jupiter
5.20
Saturn
9.54
Uranus
19.2
(a) Plot the number of the planet (Mercury is 1, etc.) versus the distance. Note
that it does not look like a line, and so finding the line of best fit is not fruitful.
(b) It does, however look like an exponential curve. Therefore, plot the number
of the planet versus the logarithm of the distance. Does this look like a line?
(c) The asteroid belt between Mars and Jupiter is thought to be what is left of a
planet that broke apart. Renumber so that Jupiter is 6, Saturn is 7, and Uranus
is 8, and plot against the log again. Does this look better?
(d) Use least squares on that data to predict the location of Neptune.
(e) Repeat to predict where Pluto is.
(f ) Is the formula accurate for Neptune and Pluto?
This method was used to help discover Neptune (although the second item is misleading about the history; actually, the discovery of Neptune in position 9 prompted
people to look for the missing planet in position 5). See [Gardner, 1970]
8 William Bennett has proposed an Index of Leading Cultural Indicators for the
US ([Bennett], in 1993). Among the statistics cited are the average daily hours
spent watching TV, and the average combined SAT scores.
TV
SAT
1960
5:06
975
1965
5:29
969
1970
5:56
948
1975
6:07
910
1980
6:36
890
1985
7:07
906
1990
6:55
900
1992
7:04
899
Suppose that a cause and effect relationship is proposed between the time spent
watching TV and the decline in SAT scores (in this article, Mr. Bennett does not
argue that there is a direct connection).
(a) Find the line of best fit relating the independent variable of average daily
TV hours to the dependent variable of SAT scores.
273
(b) Find the most recent estimate of the average daily TV hours (Bennetts cites
Neilsen Media Research as the source of these estimates). Estimate the associated SAT score. How close is your estimate to the actual average? (Warning: a
change has been made recently in the SAT, so you should investigate whether
some adjustment needs to be made to the reported average to make a valid
comparison.)
274
These maps are nicer, more regular, in that for each map all of the domain is
spread out by the same factor.
Because the only transformations of R1 are multiplications by a scalar, these
pictures are possibly misleading by being too simple. In higher-dimensional
spaces more can happen. For instance, this linear transformation of R2 , which
rotates all vectors counterclockwise, is not a simple scalar multiplication.
x
y
x cos y sin
x sin + y cos
275
And neither is this transformation of R3 , which projects vectors into the xzplane.
x
y
z
x
0
z
ki
I Mi (k)
I Pi,j
ki +j
I Ci,j (k)
1 0 0
0 1 0
0 0 0 E ,E
x
x
3
3
y
y
0
z
Therefore, we will have a description of the geometric action of h if we just
describe the geometric actions of the three kinds of elementary matrices. The
pictures below sticks to the elementary transformations of R2 only, for ease of
drawing.
The action of a matrix of the form Mi (k) (that is, the action of the transformation of R2 that is induced by this matrix) is to stretch vectors by a factor
of k along the i-th axis. This is a dilation. This map stretches by a factor of 3
along the x-axis.
x
y
3x
y
Note that if 0 k < 1 or if k < 0 then the i-th component goes the other way;
here, toward the left.
x
y
2x
y
The action of a matrix of the form Pi,j (that is, of the transformation induced
by this matrix) is to interchange the i-th and j-th axes; in two dimensions there
is only the single case P1,2 , which reflects vectors about the line y = x.
276
y
x
1 0
2 1
E2 , E2
x
x
2x + y
y
The picture
x
y
u
~
~
v
x
2x + y
h(~
u)
h(~
v)
shows that any Ci,j (k) affects vectors depending on their i-th component; in
this example, the vector ~v with the larger first component is affected moreit
is pushed further vertically, since h(~v ) is 4 higher than ~v while h(~u) is only 2
higher than ~u. Another way to see the action of this map is to see where it
sends the unit square.
u
~
~
v
w
~
x
y
x
2x + y
h(~
v)
h(~
u)
h(w)
~
In this picture, vectors with a first component of 0, like ~u, are not pushed
vertically at all but vectors with a positive first component are slid up. In
general, for any Ci,j (k), the sliding happens in such a way that vectors with the
same i-th component are slid by the same amount. Here, ~v and w
~ are each slid
up by 2. The resulting shape, a rhombus, has the same base and height as the
square (and thus the same area) but the right angles are gone. Because of this
action, this kind of map is called a skew.
Recall that under a linear map, the image of a subspace is a subspace. Thus
a linear transformation maps lines through the origin to lines through the origin
(the dimension of the image space cannot be greater than the dimension of the
domain space, so a line cant map onto, say, a plane). Note, however, that all
four sides of the above rhombus are straight, not just the two sides lying in lines
through the origin. A skew in fact a linear map of any kind maps any
line to a line. Exercise 6 asks for a proof of this. That is, linear transformations
respect the linear structures of a space. This is the reason for the assertion
made above that, even on higher-dimensional spaces, linear maps are nice or
regular.
277
But in calculus we dont focus on the map overall, we focus on the local effect
of the map. The picture on the right looks more closely at what this map does
near x = 1. At x = 1 the derivative is y 0 (1) = 3, so that near x = 1 we have
that y 3 x; in other words, (1.0012 + 1.001) (12 + 1) 3 (0.001). That
is, in a neighborhood of x = 1, this map carries the domain to the codomain by
stretching by a factor of 3 it is, locally, approximately, a dilation. This shows
a small interval in the domain (x x .. x + x) carried over to an interval in
the codomain (y y .. y + y) that is three times as wide: y 3 x.
y = 2
x = 1
(When the above picture is drawn in the traditional cartesian way then the prior
sentence is usually rephrased as: the derivative y 0 (1) = 3 gives the slope of the
line tangent to the graph at the point (1, 2).)
Calculus considers the map that locally approximates the change x 7
3 x, instead of the actual change map x 7 y(1 + x) y(1), because the
local map is easy to work with. Specifically, if the input change is doubled, or
tripled, etc., then the resulting output change will double, or triple, etc.
3(r x) = r (3x)
(for r R) and adding changes in input adds the resulting output changes.
3(x1 + x2 ) = 3x1 + 3x2
In short, whats easy to work with about x 7 3 x is that it is linear.
278
g(f (x))
f (x)
x
279
(d) Sketch the geometric effect matrix, that is, sketch how H is expressed as a
combination of dilations, flips, skews, and projections (the identity is a trivial
projection).
2 What combination of dilations, flips, skews, and projections produces a rotation
counterclockwise by 2/3 radians?
3 What combination of dilations, flips, skews, and projections produces the map
h : R3 R3 represented with respect to the standard bases by this matrix?
!
1 2 1
3 6 0
1 2 2
4 Show that any linear transformation of R1 is the map that multiplies by a scalar
x 7 kx.
5 Show that for any permutation (that is, reordering) p of the numbers 1, . . . , n,
the map
x1
xp(1)
x2 xp(2)
. 7 .
.. ..
xn
xp(n)
can be accomplished with a composition of maps, each of which only swaps a single
pair of coordinates. Hint: it can be done by induction on n. (Remark: in the fourth
chapter we will show this and we will also show that the parity of the number of
swaps used is determined by p. That is, although a particular permutation could
be accomplished in two different ways with two different numbers of swaps, either
both ways use an even number of swaps, or both use an odd number.)
6 Show that linear maps preserve the linear structures of a space.
(a) Show that for any linear map from Rn to Rm , the image of any line is a line.
The image may be a degenerate line, that is, a single point.
(b) Show that the image of any linear surface is a linear surface. This generalizes
the result that under a linear map the image of a subspace is a subspace.
(c) Linear maps preserve other linear ideas. Show that linear maps preserve
betweeness: if the point B is between A and C then the image of B is between
the image of A and the image of C.
7 Use a picture like the one that appears in the discussion of the Chain Rule to
answer: if a function f : R R has an inverse, whats the relationship between how
the function locally, approximately dilates space, and how its inverse dilates
space (assuming, of course, that it has an inverse)?
280
p0,n
p0,n+1
1 .5 0 0 0 0
0 0 .5 0 0 0 p1,n p1,n+1
0 .5 0 .5 0 0 p2,n p2,n+1
0 0 .5 0 .5 0 p3,n p3,n+1
0 0 0 .5 0 0 p4,n p4,n+1
p5,n
p5,n+1
0 0 0 0 .5 1
For instance, the probability of being in state s0 after flip n + 1 is p0,n+1 =
p0,n + 0.5 p1,n . With the initial condition that the player starts with three
dollars, calculation gives this.
n =
0
0
0
0
1
0
0
n = 1
0
0
.5
.5
0
n = 2
0
.25
.5
0
.25
n = 3
.125
.375
.25
.25
n=4
.125
.1875
.3125
0
.375
n = 24
.39600
.00276
.00447
0
.59676
For instance, after the fourth flip there is a probability of 0.50 that the game
is already over the player either has no money left or has won five dollars.
As this computational exploration suggests, the game is not likely to go on for
long, with the player quickly ending in either state s0 or state s5 . (Because a
player who enters either of these two states never leaves, they are said to be
absorbtive. An argument that involves taking the limit as n goes to infinity will
show that when the player starts with $3, there is a probability of 0.60 that the
player eventually ends with $5 and consequently a probability of 0.40 that the
player ends the game with $0. That argument is beyond the scope of this Topic,
however; here we will just look at a few computations for applications.)
This game is an example of a Markov chain, named for work by A.A. Markov
at the start of this century. The vectors of ps are probability vectors. The
matrix is a transition matrix. A Markov chain is historyless in that, with a
fixed transition matrix, the next state depends only on the current state and
not on any states that came before. Thus a player, say, who starts in state s3 ,
281
then goes to state s2 , then to s1 , and then to s2 has exactly the same chance
at this point of moving next to state s3 as does a player whose history was to
start in s3 , then go to s4 , then to s3 , and then to s2 .
Here is a Markov chain from sociology. A study ([Macdonald & Ridge],
p. 202) divided occupations in the United Kingdom into upper level (executives
and professionals), middle level (supervisors and skilled manual workers), and
lower level (unskilled). To determine the mobility across these levels in a generation, about two thousand men were asked, At which level are you, and at
which level was your father when you were fourteen years old? This equation
summarizes the results.
pU,n+1
.60 .29 .16
pU,n
.26 .37 .27 pM,n = pM,n+1
pL,n
pL,n+1
.14 .34 .57
For instance, a child of a lower class worker has a .27 probability of growing up to
be middle class. Notice that the Markov model assumption about history seems
reasonablewe expect that while a parents occupation has a direct influence
on the occupation of the child, the grandparents occupation has no such direct
influence. With the initial distribution of the respondentss fathers given below,
this table lists the distributions for the next five generations.
n = 0
.12
.32
.56
n = 1
.23
.34
.42
n = 2
.29
.34
.37
n = 3
.31
.34
.35
n = 4
.32
.33
.34
n = 5
.33
.33
.34
One more example, from a very important subject, indeed. The World Series
of American baseball is played between the team winning the American League
and the team winning the National League (we follow [Brunner] but see also
[Woodside]). The series is won by the first team to win four games. That means
that a series is in one of twenty-four states: 0-0 (no games won yet by either
team), 1-0 (one game won for the American League team and no games for the
National League team), etc. If we assume that there is a probability p that the
American League team wins each game then we have the following transition
matrix.
p0-0,n+1
p0-0,n
0
0
0
0 ...
p
0
0
0 . . .
p1-0,n p1-0,n+1
1 p
0
0
0 . . . p0-1,n p0-1,n+1
0
p
0
0 . . .
p2-0,n = p2-0,n+1
0
1p
p
0 . . . p1-1,n p1-1,n+1
0
0
1 p 0 . . .
p0-2,n p0-2,n+1
..
..
..
..
..
..
.
.
.
.
.
.
An especially interesting special case is p = 0.50; this table lists the resulting
components of the n = 0 through n = 7 vectors. (The code to generate this
table in the computer algebra system Octave follows the exercises.)
282
00
10
01
20
11
02
30
21
12
03
40
31
22
13
04
41
32
23
14
42
33
24
43
34
n=1
0
0.5
0.5
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
n=2
0
0
0
0.25
0.5
0.25
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
n=3
0
0
0
0
0
0
0.125
0.375
0.375
0.125
0
0
0
0
0
0
0
0
0
0
0
0
0
0
n=4
0
0
0
0
0
0
0
0
0
0
0.0625
0.25
0.375
0.25
0.0625
0
0
0
0
0
0
0
0
0
n=5
0
0
0
0
0
0
0
0
0
0
0.0625
0
0
0
0.0625
0.125
0.3125
0.3125
0.125
0
0
0
0
0
n=6
0
0
0
0
0
0
0
0
0
0
0.0625
0
0
0
0.0625
0.125
0
0
0.125
0.15625
0.3125
0.15625
0
0
n=7
0
0
0
0
0
0
0
0
0
0
0.0625
0
0
0
0.0625
0.125
0
0
0.125
0.15625
0
0.15625
0.15625
0.15625
283
motivated by the warmer climate, by lower wages, and by less unionization. Here is
the transition matrix for large firms in Electric and Electronic Equipment ([Kelton],
p. 43)
NE
NC
S
W
Z
NE
0.787
0
0
0
0.021
NC
0
0.966
0.063
0
0.009
S
0
0.034
0.937
0.074
0.005
W
0.111
0
0
0.612
0.010
Z
0.102
0
0
0.314
0.954
For example, a firm in the Northeast region will be in the West region next year
with probability 0.111. (The Z entry is a birth-death state. For instance, with
probability 0.102 a large Electric and Electronic Equipment firm from the Northeast will move out of this system next year: go out of business, move abroad, or
move to another category of firm. There is a 0.021 probability that a firm in the
National Census of Manufacturers will move into Electronics, or be created, or
move in from abroad, into the Northeast. Finally, with probability 0.954 a firm
out of the categories will stay out, according to this research.)
(a) Does the Markov model assumption of lack of history seem justified?
(b) Assume that the initial distribution is even, except that the value at Z is
0.9. Compute the vectors for n = 1 through n = 4.
(c) Suppose that the initial distribution is this.
NE
NC
S
W
Z
0.0000 0.6522 0.3478 0.0000 0.0000
Calculate the distributions for n = 1 through n = 4.
(d) Find the distribution for n = 50 and n = 51. Has the system settled down
to an equilibrium?
4 This model has been suggested for some kinds of learning ([Wickens], p. 41). The
learner starts in an undecided state sU . Eventually the learner has to decide to do
either response A (that is, end in state sA ) or response B (ending in sB ). However,
the learner doesnt jump right from being undecided to being sure A is the correct
thing to do (or B). Instead, the learner spends some time in a tentative-A
state, or a tentative-B state, trying the response out (denoted here tA and tB ).
Imagine that once the learner has decided, it is final, so once sA or sB is entered
it is never left. For the other state changes, imagine a transition is made with
probability p, in either direction.
(a) Construct the transition matrix.
(b) Take p = 0.25 and take the initial vector to be 1 at sU . Run this for five
steps. What is the chance of ending up at sA ?
(c) Do the same for p = 0.20.
(d) Graph p versus the chance of ending at sA . Is there a threshold value for p,
above which the learner is almost sure not to take longer than five steps?
5 A certain town is in a certain country (this is a hypothetical problem). Each year
ten percent of the town dwellers move to other parts of the country. Each year
one percent of the people from elsewhere move to the town. Assume that there
are two states sT , living in town, and sC , living elsewhere.
(a) Construct the transistion matrix.
(b) Starting with an initial distribution sT = 0.3 and sC = 0.7, get the results
for the first ten years.
(c) Do the same for sT = 0.2.
284
Computer Code
This is the code for the computer algebra system Octave that was used
to generate the table of World Series outcomes. First, this script is kept
in the file markov.m. (The sharp character # marks the rest of a line as a
comment.)
# Octave script file to compute chance of World Series outcomes.
function w = markov(p,v)
q = 1-p;
A=[0,0,0,0,0,0, 0,0,0,0,0,0, 0,0,0,0,0,0, 0,0,0,0,0,0; # 0-0
p,0,0,0,0,0, 0,0,0,0,0,0, 0,0,0,0,0,0, 0,0,0,0,0,0; # 1-0
q,0,0,0,0,0, 0,0,0,0,0,0, 0,0,0,0,0,0, 0,0,0,0,0,0; # 0-1_
0,p,0,0,0,0, 0,0,0,0,0,0, 0,0,0,0,0,0, 0,0,0,0,0,0; # 2-0
0,q,p,0,0,0, 0,0,0,0,0,0, 0,0,0,0,0,0, 0,0,0,0,0,0; # 1-1
0,0,q,0,0,0, 0,0,0,0,0,0, 0,0,0,0,0,0, 0,0,0,0,0,0; # 0-2__
0,0,0,p,0,0, 0,0,0,0,0,0, 0,0,0,0,0,0, 0,0,0,0,0,0; # 3-0
0,0,0,q,p,0, 0,0,0,0,0,0, 0,0,0,0,0,0, 0,0,0,0,0,0; # 2-1
0,0,0,0,q,p, 0,0,0,0,0,0, 0,0,0,0,0,0, 0,0,0,0,0,0; # 1-2_
0,0,0,0,0,q, 0,0,0,0,0,0, 0,0,0,0,0,0, 0,0,0,0,0,0; # 0-3
0,0,0,0,0,0, p,0,0,0,1,0, 0,0,0,0,0,0, 0,0,0,0,0,0; # 4-0
0,0,0,0,0,0, q,p,0,0,0,0, 0,0,0,0,0,0, 0,0,0,0,0,0; # 3-1__
0,0,0,0,0,0, 0,q,p,0,0,0, 0,0,0,0,0,0, 0,0,0,0,0,0; # 2-2
0,0,0,0,0,0, 0,0,q,p,0,0, 0,0,0,0,0,0, 0,0,0,0,0,0; # 1-3
0,0,0,0,0,0, 0,0,0,q,0,0, 0,0,1,0,0,0, 0,0,0,0,0,0; # 0-4_
0,0,0,0,0,0, 0,0,0,0,0,p, 0,0,0,1,0,0, 0,0,0,0,0,0; # 4-1
0,0,0,0,0,0, 0,0,0,0,0,q, p,0,0,0,0,0, 0,0,0,0,0,0; # 3-2
0,0,0,0,0,0, 0,0,0,0,0,0, q,p,0,0,0,0, 0,0,0,0,0,0; # 2-3__
0,0,0,0,0,0, 0,0,0,0,0,0, 0,q,0,0,0,0, 1,0,0,0,0,0; # 1-4
0,0,0,0,0,0, 0,0,0,0,0,0, 0,0,0,0,p,0, 0,1,0,0,0,0; # 4-2
0,0,0,0,0,0, 0,0,0,0,0,0, 0,0,0,0,q,p, 0,0,0,0,0,0; # 3-3_
0,0,0,0,0,0, 0,0,0,0,0,0, 0,0,0,0,0,q, 0,0,0,1,0,0; # 2-4
0,0,0,0,0,0, 0,0,0,0,0,0, 0,0,0,0,0,0, 0,0,p,0,1,0; # 4-3
0,0,0,0,0,0, 0,0,0,0,0,0, 0,0,0,0,0,0, 0,0,q,0,0,1]; # 3-4
285
w = A * v;
endfunction
v0=[1;0;0;0;0;0;0;0;0;0;0;0;0;0;0;0;0;0;0;0;0;0;0;0]
p=.5
v1=markov(p,v0)
v2=markov(p,v1)
...
286
Q1
P3
Q3
(Euclid never explicitly states this principle but he uses it often [Casey].) In
modern terms, picking the plane up . . . means taking a map from the plane
to itself. We, and Euclid, are considering only certain transformations of the
plane, ones that may possibly slide or turn the plane but not bend or stretch
it. Accordingly, we define a function f : R2 R2 to be distance-preserving (or
a rigid motion, or isometry) if for all points P1 , P2 R2 , the map satisfies the
condition that the distance from f (P1 ) to f (P2 ) equals the distance from P1 to
P2 . We define a plane figure to be a set of points in the plane and we say that
two figures are congruent if there is a distance-preserving map from the plane
to itself that carries one figure onto the other.
Many statements from Euclidean geometry follow easily from these definitions. Some are: (i) collinearity is invariant under any distance-preserving map
(that is, if P1 , P2 , and P3 are collinear then so are f (P1 ), f (P2 ), and f (P3 )),
(ii) betweeness is invariant under any distance-preserving map (if P2 is between
P1 and P3 then so is f (P2 ) between f (P1 ) and f (P3 )), (iii) the property of
being a triangle is invariant under any distance-preserving map (if a figure is a
triangle then the image of that figure is also a triangle), (iv) and the property of
being a circle is invariant under any distance-preserving map. In 1872, F. Klein
suggested that Euclidean geometry can be characterized as the study of properties that are invariant under distance-preserving maps. (This forms part of
Kleins Erlanger Program, which proposes the organizing principle that each
kind of geometryEuclidean, projective, etc.can be described as the study
of the properties that are invariant under some group of transformations. The
word group here means more than just collection, but that lies outside of our
scope.)
We can use linear algebra to characterize the distance-preserving maps of
the plane.
First, there are distance-preserving transformations of the plane that are not
linear. The obvious example is this translation.
x
x
1
x+1
7
+
=
y
y
0
y
However, this example turns out to be the only example, in the sense that if f
is distance-preserving and sends ~0 to ~v0 then the map ~v 7 f (~v ) ~v0 is linear.
287
That will follow immediately from this statement: a map t that is distancepreserving and sends ~0 to itself is linear. To prove this statement, let
a
c
t(~e2 ) =
t(~e1 ) =
b
d
for some a, b, c, d R. Then to show that t is linear, it suffices to show that it
can be represented by a matrix, that is, that t acts in this way for all x, y R.
ax + cy
x
t
7
()
~v =
bx + dy
y
Recall that if we fix three non-collinear points then any point in the plane can
be described by giving its distance from those three. So any point ~v in the
domain is determined by its distance from the three fixed points ~0, ~e1 , and ~e2 .
Similarly, any point t(~v ) in the codomain is determined by its distance from the
three fixed points t(~0), t(~e1 ), and t(~e2 ) (these three are not collinear because, as
mentioned above, collinearity is invariant and ~0, ~e1 , and ~e2 are not collinear).
In fact, because t is distance-preserving, we can say more: for the point ~v in the
plane that is determined by being the distance d0 from ~0, the distance d1 from
~e1 , and the distance d2 from ~e2 , its image t(~v ) must be the unique point in the
codomain that is determined by being d0 from t(~0), d1 from t(~e1 ), and d2 from
t(~e2 ). Because of the uniqueness, checking that the action in () works in the
d0 , d1 , and d2 cases
x ~
x
ax + cy ~
~
dist(
, 0) = dist(t(
), t(0)) = dist(
, 0)
y
y
bx + dy
(t is assumed to send ~0 to itself)
x
ax + cy
a
x
), t(~e1 )) = dist(
,
)
dist(
, ~e1 ) = dist(t(
y
bx + dy
b
y
and
x
ax + cy
c
x
), t(~e2 )) = dist(
,
)
dist(
, ~e2 ) = dist(t(
y
bx + dy
d
y
288
lengths of vectors, because then for all p~ and ~q the distance between the two is
preserved kt(~
p ~q )k = kt(~
p) t(~q )k = k~
p ~q k. For that check, let
~v =
x
y
t(~e1 ) =
a
b
t(~e2 ) =
c
d
b
a
a b
a b
RepE2 ,E2 (t) =
RepE2 ,E2 (t) =
b a
b a
We can geometrically describe these two cases. Let be the angle between the
x-axis and the image of ~e1 , measured counterclockwise.
The first matrix above represents, with respect to the standard bases, a
rotation of the plane by radians.
289
x cos y sin
x
t
7
x sin + y cos
y
The second matrix above represents a reflection of the plane through the line
bisecting the angle between ~e1 and t(~e1 ). (This picture shows ~e1 reflected up
into the first quadrant and ~e2 reflected down into the fourth quadrant.)
a
x
x cos + y sin
t
7
y
x sin y cos
b
a
Note that in this second case, the right angle from ~e1 to ~e2 has a counterclockwise
sense but the right angle between the images of these two has a clockwise sense,
so the sense gets reversed. Geometers speak of a distance-preserving map as
direct if it preserves sense and as opposite if it reverses sense.
So, we have characterized the Euclidean study of congruence into the consideration of the properties that are invariant under combinations of (i) a rotation
followed by a translation (possibly the trivial translation), or (ii) a reflection
followed by a translation (a reflection followed by a non-trivial translation is a
glide reflection).
Another idea, besides congruence of figures, encountered in elementary geometry is that figures are similar if they are congruent after a change of scale.
These two triangles are similar since the second is the same shape as the first,
but 3/2-ths the size.
P2
P1
Q2
Q1
P3
Q3
From the above work, we have that figures are similar if there is an orthonormal
matrix T such that the points ~q on one are derived from the points p~ by ~q =
(kT )~v + p~0 for some nonzero real number k and constant vector p~0 .
Although many of these ideas were first explored by Euclid, mathematics is
timeless and they are very much in use today. One application of rigid motions
is in computer graphics. We can, for example, take this top view of a cube
290
Frame 1:
.2 radians
Frame 2:
.4 radians
Frame 3:
.6 radians
Frame 1:
110 percent
Frame 2:
120 percent
Frame 3:
130 percent
1/
3
2/ 3
(c)
2/ 3
1/ 3
2 Write down the formula for each of these distance-preserving maps.
(a) the map that rotates /6 radians, and then translates by ~e2
(b) the map that reflects about the line y = 2x
(c) the map that reflects about y = 2x and translates over 1 and up 1
3 (a) The proof that a map that is distance-preserving and sends the zero vector to itself incidentally shows that such a map is one-to-one and onto (the
point in the domain determined by d0 , d1 , and d2 corresponds to the point
in the codomain determined by those three numbers). Therefore any distancepreserving map has an inverse. Show that the inverse is also distance-preserving.
(b) Using the definitions given in this Topic, prove that congruence is an equivalence relation between plane figures.
291
4 In practice the matrix for the distance-preserving linear transformation and the
translation are often combined into one. Check that these two computations yield
the same first two components.
! !
a c e
x
a c
x
e
b d f
y
+
b d
y
f
0 0 1
1
(These are homogeneous coordinates; see the Topic on Projective Geometry).
5 (a) Verify that the properties described in the second paragraph of this Topic
as invariant under distance-preserving maps are indeed so.
(b) Give two more properties that are of interest in Euclidean geometry from
your experience in studying that subject that are also invariant under distancepreserving maps.
(c) Give a property that is not of interest in Euclidean geometry and is not
invariant under distance-preserving maps.
Chapter 4
Determinants
In the first chapter of this book we considered linear systems and we picked out
the special case of systems with the same number of equations as unknowns,
those of the form T ~x = ~b where T is a square matrix. We noted a distinction
between two classes of T s. While such systems may have a unique solution or
no solutions or infinitely many solutions, if a particular T is associated with a
unique solution in any system, such as the homogeneous system ~b = ~0, then
T is associated with a unique solution for every ~b. We call such a matrix of
coefficients nonsingular. The other kind of T , where every linear system for
which it is the matrix of coefficients has either no solution or infinitely many
solutions, we call singular.
Through the second and third chapters the value of this distinction has been
a theme. For instance, we now know that nonsingularity of an nn matrix T
is equivalent to each of these:
a system T ~x = ~b has a solution, and that solution is unique;
Gauss-Jordan reduction of T yields an identity matrix;
the rows of T form a linearly independent set;
the columns of T form a basis for Rn ;
any map that T represents is an isomorphism;
an inverse matrix T 1 exists.
So when we look at a particular square matrix, the question of whether it
is nonsingular is one of the first things that we ask. This chapter develops
a formula to determine this. (Since we will restrict the discussion to square
matrices, in this chapter we will usually simply say matrix in place of square
matrix.)
More precisely, we will develop infinitely many formulas, one for 11 matrices, one for 22 matrices, etc. Of course, these formulas are related that
is, we will develop a family of formulas, a scheme that describes the formula for
each size.
293
294
4.I
Chapter 4. Determinants
Definition
a b
is nonsingular iff ad bc 6= 0
c d
The 33
a
d
g
b c
e f is nonsingular iff aei + bf g + cdh hf a idb gec 6= 0
h i
With these cases in mind, we posit a family of formulas, a, ad bc, etc. For each
n the formula gives rise to a determinant function detnn : Mnn R such that
an nn matrix T is nonsingular if and only if detnn (T ) 6= 0. (We usually omit
the subscript because if T is nn then det(T ) could only mean detnn (T ).)
4.I.1
Exploration
c
d
h
come one from each row and one from each column. But these observations
perhaps seem more puzzling than enlightening. For instance, we might wonder
why some of the terms are added while others are subtracted.
A good problem solving strategy is to see what properties a solution must
have and then search for something with those properties. So we shall start by
asking what properties we require of the formulas.
At this point, our primary way to decide whether a matrix is singular is
to do Gaussian reduction and then check whether the diagonal of resulting
echelon form matrix has any zeroes (that is, to check whether the product
down the diagonal is zero). So, we may expect that the proof that a formula
Section I. Definition
295
a
b
det(
) = a(kb + d) (ka + c)b = ad bc
ka + c kb + d
shows that it is indeed unchanged, and the other 22 pivot k2 + 1 gives the
same result. The 33 pivot k3 + 2 leaves the determinant unchanged
a
b
c
det(kg + d kh + e ki + f ) = a(kh + e)i + b(ki + f )g + c(kg + d)h
g
h
i
h(ki + f )a i(kg + d)b g(kh + e)c
= aei + bf g + cdh hf a idb gec
as do the other 33 pivot operations.
So there seems to be promise in the plan. Of course, perhaps the 4 4
determinant formula is affected by pivoting. We are exploring a possibility here
and we do not yet have all the facts. Nonetheless, so far, so good.
The next step is to compare det(T) with det(T ) for the operation
i j
T T
c d
det(
) = cb ad
a b
does not yield ad bc. This 1 3 swap inside of a 33 matrix
g h i
det(d e f ) = gec + hf a + idb bf g cdh aei
a b c
296
Chapter 4. Determinants
also does not give the same determinant as before the swap again there is a
sign change. Trying a different 33 swap 1 2
d e f
det(a b c ) = dbi + ecg + f ah hcd iae gbf
g h i
also gives a change of sign.
Thus, row swaps appear to change the sign of a determinant. This modifies our plan, but does not wreck it. We intend to decide nonsingularity by
considering only whether the determinant is zero, not by considering its sign.
Therefore, instead of expecting determinants to be entirely unaffected by row
operations, will look for them to change sign on a swap.
To finish, we compare det(T) to det(T ) for the operation
ki
T T
a
b
det(
) = a(kd) (kc)b = k (ad bc)
kc kd
and the other case has
a
b
e
det( d
kg kh
c
f ) = ae(ki) + bf (kg) + cd(kh)
ki
(kh)f a (ki)db (kg)ec
= k (aei + bf g + cdh hf a idb gec)
and the other two are similar. These lead us to suspect that multiplying a row
by k multiplies the determinant by k. This fits with our modified plan because
we are asking only that the zeroness of the determinant be unchanged and we
are not focusing on the determinants sign or magnitude.
In summary, to develop the scheme for the formulas to compute determinants, we look for determinant functions that remain unchanged under the
pivoting operation, that change sign on a row swap, and that rescale on the
rescaling of a row. In the next two subsections we will find that for each n such
a function exists and is unique.
For the next subsection, note that, as above, scalars come out of each row
without affecting other rows. For instance, in this equality
3 3
9
1 1
3
1 ) = 3 det(2 1
1 )
det(2 1
5 10 5
5 10 5
the 3 isnt factored out of all three rows, only out of the top row. The determinant acts on each row of independently of the other rows. When we want to use
this property of determinants, we shall write the determinant as a function of
~2 , . . .
~n ), instead of as det(T ) or det(t1,1 , . . . , tn,n ). The
the rows: det(~
1 ,
definition of the determinant that starts the next subsection is written in this
way.
Section I. Definition
297
Exercises
X 1.1 Evaluate the determinant
of each.!
2
0 1
4 0
1
3
1
3
1 1
1
(a)
(b)
(c) 0 0
1 1
1 0 1
1 3 1
1.2 Evaluate the determinant
of each. !
2
1
1
2 3 4
2
0
5
2
(a)
(b) 0
(c) 5 6 7
1 3
1 3
4
8 9 1
X 1.3 Verify that the determinant of an upper-triangular 33 matrix is the product
down the diagonal.
a
det( 0
0
b
e
0
c
f ) = aei
i
to decide
if each
or nonsingular.
2 1
0
1
4 2
(a)
(b)
(c)
3 1
1 1
2 1
1.5 Singular or nonsingular? Use the determinant to decide.
!
2 1 1
1 0 1
2
1
0
(a) 3 2 2
(b) 2 1 1
(c) 3 2 0
0 1 4
4 1 3
1
0
0
X 1.6 Each pair of matrices differ by one row operation. Use this operation to compare
det(A) with
det(B).
1 2
1
2
(a) A =
B=
2 3
0 1
!
3 1 0
3 1 0
(b) A = 0 0 1 B = 0 1 2
0 1 2
0 0 1
!
1 1
3
1 1
3
2
6 B = 1
1
3
(c) A = 2
1
0
4
1
0
4
1.7 Show this.
!
1
1
1
b
c ) = (b a)(c a)(c b)
det( a
a2 b2 c2
X 1.8 Which real numbers x make this matrix singular?
12 x
8
4
8x
1.9 Do the Gaussian reduction to check the formula for 33 matrices stated in the
preamble to this section.
a
d
g
b
e
h
c
f
i
298
Chapter 4. Determinants
x
det( x1
x2
1
1 )=0
1
x1 6= x2
X 1.11 Many people know this mnemonic for the determinant of a 33 matrix: first
repeat the first two columns and then sum the products on the forward diagonals
and subtract the products on the backward diagonals. That is, first write
!
h1,1 h1,2 h1,3 h1,1 h1,2
h2,1 h2,2 h2,3 h2,1 h2,2
h3,1 h3,2 h3,3 h3,1 h3,2
and then calculate this.
h1,1 h2,2 h3,3 + h1,2 h2,3 h3,1 + h1,3 h2,1 h3,2
h3,1 h2,2 h1,3 h3,2 h2,3 h1,1 h3,3 h2,1 h1,2
(a) Check that this agrees with the formula given in the preamble to this section.
(b) Does it extend to other-sized determinants?
1.12 The cross product of the vectors
!
!
x1
y1
~x = x2
~
y = y2
x3
y3
is the vector computed as this determinant.
!
~e1 ~e2 ~e3
~
x~
y = det( x1 x2 x3 )
y1 y2 y3
Note that the first row is composed of vectors, the vectors from the standard basis
for R3 . Show that the cross product of two vectors is perpendicular to each vector.
1.13 Prove that each statement holds for 22 matrices.
(a) The determinant of a product is the product of the determinants det(ST ) =
det(S) det(T ).
(b) If T is invertible then the determinant of the inverse is the inverse of the
determinant det(T 1 ) = ( det(T ) )1 .
Matrices T and T 0 are similar if there is a nonsingular matrix P such that T 0 =
P T P 1 . (This definition is in Chapter Five.) Show that similar 22 matrices have
the same determinant.
X 1.14 Prove that the area of this region in the plane
x2
y2
x1
det(
y1
Compare with this.
x2
det(
y2
x1
y1
x2
)
y2
x1
)
y1
Section I. Definition
299
1.15 Prove that for 22 matrices, the determinant of a matrix equals the determinant of its transpose. Does that also hold for 33 matrices?
X 1.16 Is the determinant function linear is det(xT +yS) = xdet(T )+ydet(S)?
1.17 Show that if A is 33 then det(c A) = c3 det(A) for any scalar c.
1.18 Which real numbers make
cos sin
sin
cos
singular? Explain geometrically.
1.19 [Am. Math. Mon., Apr. 1955] If a third order determinant has elements 1, 2,
. . . , 9, what is the maximum value it may have?
4.I.2
Properties of Determinants
300
Chapter 4. Determinants
2.3 Lemma A matrix with two identical rows has a determinant of zero. A
matrix with a zero row has a determinant of zero. A matrix is nonsingular if
and only if its determinant is nonzero. The determinant of an echelon form
matrix is the product down its diagonal.
Proof. To verify the first sentence, swap the two equal rows. The sign of the
determinant changes, but the matrix is unchanged and so its determinant is
unchanged. Thus the determinant is zero.
The second sentence is clearly true if the matrix is 11. If it has at least
two rows then apply property (1) of the definition with the zero row as row j
and with k = 1.
~i , . . . ,
~i + ~0, . . . )
det(. . . ,
~i , . . . , ~0, . . . ) = det(. . . ,
The first sentence of this lemma gives that the determinant is zero.
For the third sentence, where T T is the Gauss-Jordan reduction,
by the definition the determinant of T is zero if and only if the determinant of
T is zero (although they could differ in sign or magnitude). A nonsingular T
Gauss-Jordan reduces to an identity matrix and so has a nonzero determinant.
A singular T reduces to a T with a zero row; by the second sentence of this
lemma its determinant is zero.
Finally, for the fourth sentence, if an echelon form matrix is singular then it
has a zero on its diagonal, that is, the product down its diagonal is zero. The
third sentence says that if a matrix is singular then its determinant is zero. So
if the echelon form matrix is singular then its determinant equals the product
down its diagonal.
If an echelon form matrix is nonsingular then none of its diagonal entries is
zero so we can use property (3) of the definition to factor them out (again, the
vertical bars | | indicate the determinant operation).
t1,1
t1,2
t2,2
..
1 t1,2 /t1,1
t1,n
t2,n
1
0
tn,n
..
t1,n /t1,1
t2,n /t2,2
Next, the Jordan half of Gauss-Jordan elimination, using property (1) of the
definition, leaves the identity matrix.
1 0
0
0 1
0
0
1
Therefore, if an echelon form matrix is nonsingular then its determinant is the
QED
product down its diagonal.
Section I. Definition
301
2 4 2
1 3 = 0
4
= 10
5
with Gauss method wont give a big savings because the 2 2 determinant
formula is so easy. However, a 33 determinant is usually easier to calculate
with Gauss method than with the formula given earlier.
2 2
2 2 6 2 2
6
6
4 4 3 = 0 0 9 = 0 3 5 = 54
0 0 9
0 3 5 0 3 5
2.5 Example Determinants of matrices any bigger than 33 are almost always
most quickly done with this Gauss method procedure.
1 0 1
1 0 1
1 0 1 3
3
3
0 1 1
0 1 1
0 1 1 4
4
4
= 0 0 1 3 = (5) = 5
0 0 0 5 = 0 0 0
5
0 0 0
0 0 1 3
0 1 0 1
5
The prior example illustrates an important point. Although we have not yet
found a 44 determinant formula, if one exists then we know what value it gives
to the matrix if there is a function with properties (1)-(4) then on the above
matrix the function must return 5.
2.6 Lemma For each n, if there is an n n determinant function then it is
unique.
Proof. For any n n matrix we can perform Gauss method on the matrix,
keeping track of how the sign alternates on row swaps, and then multiply down
the diagonal of the echelon form result. By the definition and the lemma, all nn
determinant functions must return this value on this matrix. Thus all nn determinant functions are equal, that is, there is only one input argument/output
QED
value relationship satisfying the four conditions.
The if there is an nn determinant function emphasizes that, although we
can use Gauss method to compute the only value that a determinant function
could possibly return, we havent yet shown that such a determinant function
exists for all n. In the rest of the section we will produce determinant functions.
Exercises
For these, assume that an nn determinant function exists for all n.
X 2.7 Use Gauss method to find each determinant.
302
Chapter 4. Determinants
2
(b)
1
1
0 0 1
1 1 0
0 1 0
1 1 0
2.8 Use Gauss method to find each.
1 1 0
2 1
(b) 3 0 2
(a)
1 1
5 2 2
2.9 For which values of k does this system have a unique solution?
(a) 3
0
1
1
1
0
4
+ zw=2
y 2z
=3
x + kz
=4
zw=2
h1,1
h1,2
h1,3
h2,2
h2,3
(c) h2,1
5h
5h3,2
5h3,3
3,1
X 2.11 Find the determinant of a diagonal matrix.
2.12 Describe the solution set of a homogeneous linear system if the determinant
of the matrix of coefficients is nonzero.
X 2.13 Show that this determinant is zero.
y + z x + z x + y
y
z
x
1
1
1
2.14 (a) Find the 11, 22, and 33 matrices with i, j entry given by (1)i+j .
(b) Find the determinant of the square matrix with i, j entry (1)i+j .
2.15 (a) Find the 11, 22, and 33 matrices with i, j entry given by i + j.
(b) Find the determinant of the square matrix with i, j entry i + j.
X 2.16 Show that determinant functions are not linear by giving a case where |A +
B| 6= |A| + |B|.
2.17 The second condition in the definition, that row swaps change the sign of a
determinant, is somewhat annoying. It means we have to keep track of the number
of swaps, to compute how the sign alternates. Can we get rid of it? Can we replace
it with the condition that row swaps leave the determinant unchanged? (If so then
we would need new 1 1, 2 2, and 3 3 formulas, but that would be a minor
matter.)
2.18 Prove that the determinant of any triangular matrix, upper or lower, is the
product down its diagonal.
2.19 Refer to the definition of elementary matrices in the Mechanics of Matrix
Multiplication subsection.
(a) What is the determinant of each kind of elementary matrix?
Section I. Definition
303
(b) Prove that if E is any elementary matrix then |ES| = |E||S| for any appropriately sized S.
(c) (This question doesnt involve determinants.) Prove that if T is singular then
a product T S is also singular.
(d) Show that |T S| = |T ||S|.
(e) Show that if T is nonsingular then |T 1 | = |T |1 .
2.20 Prove that the determinant of a product is the product of the determinants
|T S| = |T | |S| in this way. Fix the n n matrix S and consider the function
d : Mnn R given by T 7 |T S|/|S|.
(a) Check that d satisfies property (1) in the definition of a determinant function.
(b) Check property (2).
(c) Check property (3).
(d) Check property (4).
(e) Conclude the determinant of a product is the product of the determinants.
2.21 A submatrix of a given matrix A is one that can be obtained by deleting some
of the rows and columns of A. Thus, the first matrix here is a submatrix of the
second.
!
3
4
1
3 1
0
9
2
2 5
2 1
5
Prove that for any square matrix, the rank of the matrix is r if and only if r is the
largest integer such that there is an rr submatrix with a nonzero determinant.
X 2.22 Prove that a matrix with rational entries has a rational determinant.
2.23 [Am. Math. Mon., Feb. 1953] Find the element of likeness in (a) simplifying a
fraction, (b) powdering the nose, (c) building new steps on the church, (d) keeping
emeritus professors on campus, (e) putting B, C, D in the determinant
1
3
a
B
4.I.3
a
1
a3
D
a2
a
1
a3
a3
a2
.
a
1
304
Chapter 4. Determinants
conflict: here are two Gauss method reductions of the same matrix, the first
without any row swap
1 2 31 +2 1 2
0 2
3 4
and the second with a swap.
1 2 1 2 3 4 (1/3)1 +2 3
1 2
0
3 4
4
2/3
Following Definition 2.1 gives that both calculations yield the determinant 2
since in the second one we keep track of the fact that the row swap changes
the sign of the result of multiplying down the diagonal. But if we follow the
supposition and change the second condition then the two calculations yield
different values, 2 and 2. That is, under the supposition the outcome would not
be well-defined no function exists that satisfies the changed second condition
along with the other three.
Of course, observing that Definition 2.1 does the right thing in this one
instance is not enough; what we will do in the rest of this section is to show
that there is never a conflict. The natural way to try this would be to define
the determinant function with: The value of the function is the result of doing
Gauss method, keeping track of row swaps, and finishing by multiplying down
the diagonal. (Since Gauss method allows for some variation, such as a choice
of which row to use when swapping, we would have to fix an explicit algorithm.)
Then we would be done if we verified that this way of computing the determinant
satisfies the four properties. For instance, if T and T are related by a row swap
then we would need to show that this algorithm returns determinants that are
negatives of each other. However, how to verify this is not evident. So the
development below will not proceed in this way. Instead, in this subsection we
will define a different way to compute the value of a determinant, a formula,
and we will use this way to prove that the conditions are satisfied.
The formula that we shall use is based on an insight gotten from property (2)
of the definition of determinants. This property shows that determinants are
not linear.
3.1 Example For this matrix det(2A) 6= 2 det(A).
2 1
A=
1 3
Instead, the scalar comes out of each of the two rows.
4 2
= 2 2 1 = 4 2 1
1 3
2 6
2 6
Since scalars come out a row at a time, we might guess that determinants
are linear a row at a time.
Section I. Definition
305
~ . . . , ~n ) = det(~
1 , . . . , ~v , . . . , ~n ) + det(~
1 , . . . , w,
~ ...,
~n )
det(~
1 , . . . , ~v + w,
~i+1 , . . . , ~n } is linearly dependent then all three matrices
If the set {~
1 , . . . , ~i1 ,
are singular and so all three determinants are zero and the equality is trivial.
Therefore assume that the set is linearly independent. This set of n-wide row
vectors has n 1 members, so we can make a basis by adding one more vector
~
~i1 , ,
~i+1 , . . . ,
~n i. Express ~v and w
~ with respect to this basis
h~
1 , . . . ,
~ + vi+1
~1 + + vi1
~i1 + vi
~i+1 + + vn
~n
~v = v1
~ + wi+1
w
~ = w1
~1 + + wi1
~i1 + wi
~i+1 + + wn
~n
giving this.
~ + + (vn + wn )~
1 + + (vi + wi )
n
~v + w
~ = (v1 + w1 )~
By the definition of determinant, the value of det(~
1 , . . . , ~v + w,
~ . . . , ~n ) is un1 to ~v + w.
~
changed by the pivot operation of adding (v1 + w1 )~
~ + + (vn + wn )~
1 = (v2 + w2 )~
2 + + (vi + wi )
n
~v + w
~ (v1 + w1 )~
Then, to the result, we can add (v2 + w2 )~
2 , etc. Thus
~ ...,
~n )
det(~
1 , . . . , ~v + w,
~ . . . , ~n )
= det(~
1 , . . . , (vi + wi ) ,
~ . . . , ~n )
1 , . . . , ,
= (vi + wi ) det(~
~ . . . , ~n ) + wi det(~
~ . . . , ~n )
1 , . . . , ,
1 , . . . , ,
= vi det(~
(using (2) for the second equality). To finish, bring vi and wi back inside in
~ and use pivoting again, this time to reconstruct the expressions of ~v
front of
~1
and w
~ in terms of the basis, e.g., start with the pivot operations of adding v1
~ and w1
to vi
~1 to wi
~1 , etc.
QED
Multilinearity allows us to expand a determinant into a sum of determinants,
each of which involves a simple matrix.
306
Chapter 4. Determinants
3.4 Example We can use multilinearity to split this determinant into two,
first breaking up the first row
2 1 2 0 0 1
+
=
4 3 4 3 4 3
and then separating each of those two, breaking along
2 0 2 0 0 1 0
+
+
+
=
4 0 0 3 4 0 0
1
3
We are left with four determinants, such that in each row of each matrix there
is a single entry from the original matrix.
3.5 Example In the same way, a 3 3 determinant separates into a sum of
many simpler determinants. We start by splitting along the first row, producing
three determinants (the zero in the 1, 3 position is underlined to set it off visually
from the zeroes that appear in the splitting).
1 1 2 0 0 0 1 0 0 0
3 0 = 4 3 0 + 4 3 0 + 4 3
1 5 2 1 5 2 1 5 2 1
1
0
5
Each of these three will itself split in three along the second row. Each of
the resulting nine splits in three along the third row, resulting in twenty seven
determinants
0 0 1
2 0 0 2 0 0 2 0 0 2 0 0
= 4 0 0 + 4 0 0 + 4 0 0 + 0 3 0 + + 0 0 0
0 0 5
2 0 0 0 1 0 0 0 5 2 0 0
such that each row contains a single entry from the starting matrix.
So an nn determinant expands into a sum of nn determinants where each
row of each summands contains a single entry from the starting matrix. However, many of these summand determinants are zero.
3.6 Example In each of these three matrices from the above expansion, two
of the rows have their entry from the starting matrix in the same column, e.g.,
in the first matrix, the 2 and the 4 both come from the first column.
0 1 0
0 0 1
2 0 0
0 0 0
0 3 0
4 0 0
0 0 5
0 0 5
0 1 0
Any such matrix is singular, because in each, one row is a multiple of the other
(or is a zero row). Thus, any such determinant is zero, by Lemma 2.3.
Section I. Definition
307
Therefore, the above expansion of the 33 determinant into the sum of the
twenty seven determinants simplifies to the sum of these six.
2 1 1 2 0 0 2 0 0
4 3 0 = 0 3 0 + 0 0 0
2 1 5 0 0 5 0 1 0
0 1 0 0 1 0
+ 4 0 0 + 0 0 0
0 0 5 2 0 0
0 0 1 0 0 1
+ 4 0 0 + 0 3 0
0 1 0 2 0 0
We can bring out the scalars.
1 0 0
1 0 0
= (2)(3)(5) 0 1 0 + (2)(0)(1) 0 0 1
0 0 1
0 1 0
0 1 0
0 1 0
+ (1)(4)(5) 1 0 0 + (1)(0)(2) 0 0 1
0 0 1
1 0 0
0 0 1
0 0 1
+ (1)(4)(1) 1 0 0 + (1)(3)(2) 0 1 0
1 0 0
0 1 0
To finish, we evaluate those six determinants by row-swapping them to the
identity matrix, keeping track of the resulting sign changes.
= 30 (+1) + 0 (1)
+ 20 (1) + 0 (+1)
4 (+1) 6 (1) = 12
That example illustrates the key idea. Weve applied multilinearity to a 33
determinant to get 33 separate determinants, each with one distinguished entry
per row. We can drop most of these new determinants because the matrices
are singular, with one row a multiple of another. We are left with the oneentry-per-row determinants also having only one entry per column (one entry
from the original determinant, that is). And, since we can factor scalars out, we
can further reduce to only considering determinants of one-entry-per-row-andcolumn matrices where the entries are ones.
These are permutation matrices. Thus, the determinant can be computed
in this three-step way (Step 1) for each permutation matrix, multiply together
the entries from the original matrix where that permutation matrix has ones,
(Step 2) multiply that by the determinant of the permutation matrix and
(Step 3) do that for all permutation matrices and sum the results together.
308
Chapter 4. Determinants
0 0 1 0
3
2 0 1 0 0
=
1 1 0 0 0
4
0 0 0 1
3.7 Definition An n-permutation is a sequence consisting of an arrangement
of the numbers 1, 2, . . . , n.
3.8 Example The 2-permutations are 1 = h1, 2i and 2 = h2, 1i. These are
the associated permutation matrices.
1
2
1 0
0 1
P2 =
=
=
P1 =
2
1
0 1
1 0
We sometimes write permutations as functions, e.g., 2 (1) = 2, and 2 (2) = 1.
Then the rows of P2 are 2 (1) = 2 and 2 (2) = 1 .
The 3-permutations are 1 = h1, 2, 3i, 2 = h1, 3, 2i, 3 = h2, 1, 3i, 4 =
h2, 3, 1i, 5 = h3, 1, 2i, and 6 = h3, 2, 1i. Here are two of the associated permutation matrices.
1
3
1 0 0
0 0 1
P5 = 1 = 1 0 0
P2 = 3 = 0 0 1
2
2
0 1 0
0 1 0
For instance, the rows of P5 are 5 (1) = 3 , 5 (2) = 1 , and 5 (3) = 2 .
3.9 Definition The
t1,1 t1,2
t2,1 t2,2
..
tn,1 tn,2
. . . t1,n
. . . t2,n
= t1,1 (1) t2,1 (2) tn,1 (n) |P1 |
Section I. Definition
309
read aloud as the sum, over all permutations , of terms having the form
t1,(1) t2,(2) tn,(n) |P |. This phrase is just a restating of the three-step
process (Step 1) for each permutation matrix, compute t1,(1) t2,(2) tn,(n)
(Step 2) multiply that by |P | and (Step 3) sum all such terms together.
3.10 Example The familiar formula for the determinant of a 22 matrix can
be derived in this way.
t1,1 t1,2
1 0
0 1
+ t1,2 t2,1
= t1,1 t2,2
0 1
1 0
= t1,1 t2,2 t1,2 t2,1
(the second permutation matrix takes one row swap to pass to the identity).
Similarly, the formula for the determinant of a 33 matrix is this.
t2,1 t2,2 t2,3 = t1,1 t2,2 t3,3 |P | + t1,1 t2,3 t3,2 |P | + t1,2 t2,1 t3,3 |P |
1
2
3
310
Chapter 4. Determinants
We finish with a summary (although the final subsection contains the unfinished business of proving the two theorems). Determinant functions exist,
are unique, and we know how to compute them. As for what determinants are
about, perhaps these lines [Kemp] help make it memorable.
Determinant none,
Solution: lots or none.
Determinant some,
Solution: just one.
Exercises
These summarize the notation used in this book for the 2- and 3- permutations.
i
1 2
1 2 3
i
1 (i) 1 2
1 (i) 1 2 3
2 (i) 2 1
2 (i) 1 3 2
3 (i) 2 1 3
4 (i) 2 3 1
5 (i) 3 1 2
6 (i) 3 2 1
X 3.14 Compute
the
determinant
by
using
the
permutation expansion.
1 2 3
2
2
1
1 0
(a) 4 5 6
(b) 3
7 8 9
2 0 5
X 3.15 Compute these both with Gauss method and with the permutation expansion
formula.
0 1 4
2 1
(b) 0 2 3
(a)
3 1
1 5 1
X 3.16 Use the permutation expansion formula to derive the formula for 33 determinants.
3.17 List all of the 4-permutations.
3.18 A permutation, regarded as a function from the set {1, .., n} to itself, is oneto-one and onto. Therefore, each permutation has an inverse.
(a) Find the inverse of each 2-permutation.
(b) Find the inverse of each 3-permutation.
3.19 Prove that f is multilinear if and only if for all ~v , w
~ V and k1 , k2 R, this
holds.
f (~
1 , . . . , k1~v1 + k2~v2 , . . . ,
~n ) = k1 f (~
1 , . . . , ~v1 , . . . ,
~n ) + k2 f (~
1 , . . . , ~v2 , . . . ,
~n )
3.20 Find the only nonzero term in the permutation expansion of this matrix.
0 1 0 0
1 0 1 0
0 1 0 1
0 0 1 0
Compute that determinant by finding the signum of the associated permutation.
3.21 How would determinants change if we changed property (4) of the definition
to read that |I| = 2?
3.22 Verify the second and third statements in Corollary 3.13.
Section I. Definition
311
X 3.23 Show that if an nn matrix has a nonzero determinant then any column vector
~v Rn can be expressed as a linear combination of the columns of the matrix.
3.24 True or false: a matrix whose entries are only zeros or ones has a determinant
equal to zero, one, or negative one.
3.25 (a) Show that there are 120 terms in the permutation expansion formula of
a 55 matrix.
(b) How many are sure to be zero if the 1, 2 entry is zero?
3.26 How many n-permutations are there?
3.27 A matrix A is skew-symmetric if
Atrans =
A, as in this matrix.
0
3
A=
3 0
Show that nn skew-symmetric matrices with nonzero determinants exist only for
even n.
X 3.28 What is the smallest number of zeros, and the placement of those zeros, needed
to ensure that a 44 matrix has a determinant of zero?
X 3.29 If we have n data points (x1 , y1 ), (x2 , y2 ), . . . , (xn , yn ) and want to find a
polynomial p(x) = an1 xn1 + an2 xn2 + + a1 x + a0 passing through those
points then we can plug in the points to get an n equation/n unknown linear
system. The matrix of coefficients for that system is called the Vandermonde
matrix. Prove that the determinant of the transpose of that matrix of coefficients
1
1
...
1
x2
...
xn
x1
x 2
2
2
x2
...
xn
1
..
x1 n1 x2 n1 . . . xn n1
equals the product, over all indices i, j {1, . . . , n} with i < j, of terms of the
form xj xi . (This shows that the determinant is zero, and the linear system has
no solution, if and only if the xi s in the data are not distinct.)
3.30 A matrix can be divided into blocks, as here,
!
1 2
0
3 4
0
0 0 2
which shows four blocks, the square 22 and 11 ones in the upper left and lower
right, and the zero blocks in the upper right and lower left. Show that if a matrix
can be partitioned as
J
Z2
T =
Z1 K
where J and K are square, and Z1 and Z2 are all zeroes, then |T | = |J| |K|.
X 3.31 Prove that for any nn matrix T there are at most n distinct reals r such
that the matrix T rI has determinant zero (we shall use this result in Chapter
Five).
3.32 [Math. Mag., Jan. 1963, Q307] The nine positive digits can be arranged into
33 arrays in 9! ways. Find the sum of the determinants of these arrays.
3.33 [Math. Mag., Jan. 1963, Q237] Show that
x 2 x 3 x 4
x + 1 x 1 x 3 = 0.
x 4 x 7 x 10
312
Chapter 4. Determinants
3.34 [Am. Math. Mon., Jan. 1949] Let S be the sum of the integer elements of a
magic square of order three and let D be the value of the square considered as a
determinant. Show that D/S is an integer.
3.35 [Am. Math. Mon., Jun. 1931] Show that the determinant of the n2 elements
in the upper left corner of the Pascal triangle
1
1
1
1
.
.
1
2
3
.
1
3
.
.
1
.
.
.
.
4.I.4
Determinants Exist
This subsection is optional. It consists of proofs of two results from the prior
subsection. These proofs involve the properties of permutations, which will not
be used later, except in the optional Jordan Canonical Form subsection.
The prior subsection attacks the problem of showing that for any size there
is a determinant function on the set of square matrices of that size by using
multilinearity to develop the permutation expansion.
..
0 1 0 0
1 0 0 0
P =
0 0 1 0
0 0 0 1
Section I. Definition
313
1
1 2 0
P
0
0
or with three.
0
3 1 1
P
0
0
0
0
1
0
1
0
0
0
0
0
2 3
0
1
0
0
1
0
0
1
0
0
0
0
1
0
0
0
0
1
0
1
0
0
1
0
0
0
0
0
1 3
0
1
1
0
0
0
0
1
0
0
0
0
1
0
0
0
0
1
1
0
0
..
..
.
.
(j) k j (k)
P =
(k) (j)
..
..
.
.
314
Chapter 4. Determinants
then the swap changes the total number of inversions by one either removing
or producing one inversion, depending on whether (j) > (k) or not, since
inversions involving rows not in this pair are not affected. Consequently, the
total number of inversions changes from odd to even or from even to odd.
If the rows are not adjacent then they can be swapped via a sequence of
adjacent swaps, first bringing row k up
..
..
.
.
(k)
(j)
(j)
(j+1)
j+1 j
(j+2) k k1 k1 k2
. . . (j+1)
..
..
.
.
(k1)
(k)
..
..
.
.
and then bringing row j down.
...
k1 k
..
.
(k)
(j+1)
(j+2)
..
.
(j)
..
.
Each of these adjacent swaps changes the number of inversions from odd to even
or from even to odd. There are an odd number (k j) + (k j 1) of them.
The total change in the number of inversions is from even to odd or from odd
to even.
QED
4.4 Definition The signum of a permutation sgn() is +1 if the number of
inversions in P is even, and is 1 if the number of inversions is odd.
4.5 Example With the subscripts from Example 3.8 for the 3-permutations,
sgn(1 ) = 1 while sgn(2 ) = 1.
4.6 Corollary If a permutation matrix has an odd number of inversions then
swapping it to the identity takes an odd number of swaps. If it has an even
number of inversions then swapping to the identity takes an even number of
swaps.
Proof. The identity matrix has zero inversions. To change an odd number to
zero requires an odd number of swaps, and to change an even number to zero
QED
requires an even number of swaps.
Section I. Definition
315
We still have not shown that the permutation expansion is well-defined because we have not considered row operations on permutation matrices other than
row swaps. We will finesse this problem: we will define a function d : Mnn R
by altering the permutation expansion formula, replacing |P | with sgn()
X
t1,(1) t2,(2) . . . tn,(n) sgn()
d(T ) =
permutations
(this gives the same value as the permutation expansion because the prior result
shows that det(P ) = sgn()). This formulas advantage is that the number of
inversions is clearly well-defined just count them. Therefore, we will show
that a determinant function exists for all sizes by showing that d is it, that is,
that d satisfies the four conditions.
4.7 Lemma The function d is a determinant. Hence determinants exist for
every n.
Proof. Well must check that it has the four properties from the definition.
d(I) =
perms
all of the summands are zero except for the product down the diagonal, which
is one.
ki
For property (3) consider d(T) where T T.
X
X
t1,(1) kti,(i) tn,(n) sgn()
t1,(1) ti,(i) tn,(n) sgn() =
perms
i j
For (2), let T T.
X
t1,(1) ti,(i) tj,(j) tn,(n) sgn()
d(T) =
perms
To convert to unhatted ts, for each consider the permutation that equals
except that the i-th and j-th numbers are interchanged, (i) = (j) and (j) =
(i). Replacing the in t1,(1) ti,(i) tj,(j) tn,(n) with this gives
t1,(1) tj,(j) ti,(i) tn,(n) . Now sgn() = sgn() (by Lemma 4.3)
and so we get
X
316
Chapter 4. Determinants
where the sum is over all permutations derived from another permutation
by a swap of the i-th and j-th numbers. But any permutation can be derived
from some other permutation by such a swap, in one and only one way, so this
summation is in fact a sum over all permutations, taken once and only once.
Thus d(T) = d(T ).
ki +j
To do property (1) let T T and consider
X
d(T) =
perms
X
t1,(1) ti,(i) kti,(j) tn,(n) sgn()
= k
+ d(T )
We finish by showing that the terms t1,(1) ti,(i) ti,(j) . . . tn,(n) sgn()
add to zero. This sum represents d(S) where S is a matrix equal to T except
that row j of S is a copy of row i of T (because the factor is ti,(j) , not tj,(j) ).
Thus, S has two equal rows, rows i and j. Since we have already shown that d
changes sign on row swaps, as in Lemma 2.3 we conclude that d(S) = 0. QED
We have now shown that determinant functions exist for each size. We
already know that for each size there is at most one determinant. Therefore,
the permutation expansion computes the one and only determinant value of a
square matrix.
We end this subsection by proving the other result remaining from the prior
subsection, that the determinant of a matrix equals the determinant of its transpose.
4.8 Example Writing out the permutation expansion of the general 33 matrix
and of its transpose, and comparing corresponding terms
0 0 1
a b c
d e f = + cdh 1 0 0 +
0 1 0
g h i
Section I. Definition
317
letters)
0
d g
e h = + dhc 0
1
f i
1
0
0
0
1 +
0
shows that the corresponding permutation matrices are transposes. That is,
there is a relationship between these corresponding permutations. Exercise 15
shows that they are inverses.
4.9 Theorem The determinant of a matrix equals the determinant of its transpose.
Proof. Call the matrix T and denote the entries of T trans with ss so that
and we can finish the argument by manipulating the expression on the right
to be recognizable as the determinant of the transpose. We have written all
permutation expansions (as in the middle expression above) with the row indices
ascending. To rewrite the expression on the right in this way, note that because
is a permutation, the row indices in the term on the right (1), . . . , (n) are
just the numbers 1, . . . , n, rearranged. We can thus commute to have these
ascend, giving s1,1 (1) sn,1 (n) (if the column index is j and the row index
is (j) then, where the row index is i, the column index is 1 (i)). Substituting
on the right gives
X
s1,1 (1) sn,1 (n) sgn(1 )
=
1
as required.
QED
Exercises
These summarize the notation used in this book for the 2i
1 2
1 2
i
1 (i) 1 2
1 (i) 1 2
2 (i) 2 1
2 (i) 1 3
3 (i) 2 1
4 (i) 2 3
5 (i) 3 1
6 (i) 3 2
and 3- permutations.
3
3
2
3
1
2
1
318
Chapter 4. Determinants
4.10 Give the permutation expansion of a general 22 matrix and its transpose.
X 4.11 This problem appears also in the prior subsection.
(a) Find the inverse of each 2-permutation.
(b) Find the inverse of each 3-permutation.
X 4.12 (a) Find the signum of each 2-permutation.
(b) Find the signum of each 3-permutation.
4.13 What is the signum of the n-permutation = hn, n 1, . . . , 2, 1i?
4.14 Prove these.
(a) Every permutation has an inverse.
(b) sgn(1 ) = sgn()
(c) Every permutation is the inverse of another.
4.15 Prove that the matrix of the permutation inverse is the transpose of the matrix
of the permutation P1 = P trans , for any permutation .
X 4.16 Show that a permutation matrix with m inversions can be row swapped to
the identity in m steps. Contrast this with Corollary 4.6.
X 4.17 For any permutation let g() be the integer defined in this way.
g() =
[(j) (i)]
i<j
(This is the product, over all indices i and j with i < j, of terms of the given
form.)
(a) Compute the value of g on all 2-permutations.
(b) Compute the value of g on all 3-permutations.
(c) Prove this.
g()
sgn() =
|g()|
Many authors give this formula as the definition of the signum function.
4.II
319
Geometry of Determinants
4.II.1
x1
y1
is familiar from the construction of the sum of the two vectors. One way to
compute the area that it encloses is to draw this rectangle and subtract the
area of each subregion.
A
area of parallelogram
= area of rectangle area of A area of B
area of F
= (x1 + x2 )(y1 + y2 ) x2 y1 x1 y1 /2
x2 y2 /2 x2 y2 /2 x1 y1 /2 x2 y1
= x1 y2 x2 y1
y2
y1 C
E
x2
F
x1
The fact that the area equals the value of the determinant
x1 x2
y1 y2 = x1 y2 x2 y1
is no coincidence. The properties in the definition of determinants make reasonable postulates for a function that measures the size of the region enclosed
by the vectors in the matrix.
For instance, this shows the effect of multiplying one of the box-defining
vectors by a scalar (the scalar used is k = 1.4).
w
~
w
~
~
v
k~
v
320
Chapter 4. Determinants
w
~
~
v
~
v
Although the region on the right, the box formed by v and k~v + w,
~ is more
slanted than the shaded region, the two have the same base and the same height
and hence the same area. This illustrates that size(~v , k~v + w)
~ = size(~v , w).
~
Generalized, size(. . . , ~v , . . . , w,
~ . . . ) = size(. . . , ~v , . . . , k~v + w,
~ . . . ), which is a
restatement of the determinant postulate.
Of course, this picture
~
e2
~
e1
shows that size(~e1 , ~e2 ) = 1, and we naturally extend that to any number of
dimensions size(~e1 , . . . , ~en ) = 1, which is a restatement of the property that the
determinant of the identity matrix is one.
With that, because property (2) of determinants is redundant (as remarked
right after the definition), we have that all of the properties of determinants are
reasonable to expect of a function that gives the size of boxes. We can now cite
the work done in the prior section to show that the determinant exists and is
unique to be assured that these postulates are consistent and sufficient (we do
not need any more postulates). That is, weve got an intuitive justification to
interpret det(~v1 , . . . , ~vn ) as the size of the box formed by the vectors. (Comment. An even more basic approach, which also leads to the definition below,
is [Weston].)
1.1 Example The volume of this parallelepiped, which can be found by the
usual formula from high school geometry, is 12.
1
2
0
2
0
1
0
3
1
2 0
0 3
2 1
1
0 = 12
1
321
1.2 Remark Although property (2) of the definition of determinants is redundant, it raises an important point. Consider these two.
~v
~v
~
u
1
= 10
3
~
u
4
= 10
2
The only difference between them is in the order in which the vectors are taken.
If we take ~u first and then go to ~v , follow the counterclockwise arc shown, then
the sign is positive. Following a clockwise arc gives a negative sign. The sign
returned by the size function reflects the orientation or sense of the box. (We
see the same thing if we picture the effect of scalar multiplication by a negative
scalar.)
Although it is both interesting and important, the idea of orientation turns
out to be tricky. It is not needed for the development below, and so we will pass
it by. (See Exercise 27.)
1.3 Definition The box (or parallelepiped) formed by h~v1 ,. . . , ~vn i (where each
vector is from Rn ) includes all of the set {t1~v1 + + tn~vn t1 , . . . , tn [0..1]}.
The volume of a box is the absolute value of the determinant of the matrix with
those vectors as columns.
1.4 Example Volume, because it is an absolute value, does not depend on
the order in which the vectors are given. The volume of the parallelepiped in
Exercise 1.1, can also be computed as the absolute value of this determinant.
0 2 0
3 0 3 = 12
1 2 1
The definition of volume gives a geometric interpretation to something in
the space, boxes made from vectors. The next result relates the geometry to
the functions that operate on spaces.
1.5 Theorem A transformation t : Rn Rn changes the size of all boxes by
the same factor, namely the size of the image of a box |t(S)| is |T | times the
size of the box |S|, where T is the matrix representing t with respect to the
standard basis. That is, for all nn matrices, the determinant of a product is
the product of the determinants |T S| = |T | |S|.
The two sentences state the same idea, first in map terms and then in matrix
terms. Although we tend to prefer a map point of view, the second sentence,
the matrix version, is more convienent for the proof and is also the way that
we shall use this result later. (Alternate proofs are given as Exercise 23 and
Exercise 28.)
322
Chapter 4. Determinants
Proof. The two statements are equivalent because |t(S)| = |T S|, as both give
the size of the box that is the image of the unit box En under the composition
t s (where s is the map represented by S with respect to the standard basis).
First consider the case that |T | = 0. A matrix has a zero determinant if and
only if it is not invertible. Observe that if T S is invertible, so that there is an
M such that (T S)M = I, then the associative property of matrix multiplication
T (SM ) = I shows that T is also invertible (with inverse SM ). Therefore, if T
is not invertible then neither is T S if |T | = 0 then |T S| = 0, and the result
holds in this case.
Now consider the case that |T | =
6 0, that T is nonsingular. Recall that any
nonsingular matrix can be factored into a product of elementary matrices, so
that T S = E1 E2 Er S. In the rest of this argument, we will verify that if E
is an elementary matrix then |ES| = |E| |S|. The result will follow because
then |T S| = |E1 Er S| = |E1 | |Er | |S| = |E1 Er | |S| = |T | |S|.
If the elementary matrix E is Mi (k) then Mi (k)S equals S except that row i
has been multiplied by k. The third property of determinant functions then
gives that |Mi (k)S| = k |S|. But |Mi (k)| = k, again by the third property
because Mi (k) is derived from the identity by multiplication of row i by k, and
so |ES| = |E| |S| holds for E = Mi (k). The E = Pi,j = 1 and E = Ci,j (k)
QED
checks are similar.
1.6 Example Application of the map t represented with respect to the standard bases by
1 1
2 0
will double sizes of boxes, e.g., from this
~
v
w
~
1
=3
2
to this
t(~
v)
3
=6
2
t(w)
~
QED
323
Exercises
1.8 Find
volume
the
of the region formed.
1
1
(a) h
,
i
3
4
! ! !
2
3
8
(b) h 1 , 2 , 3 i
0
4
8
1
2
1
0
2 2 3 1
(c) h , , , i
0
2
0
0
1
2
5
7
X 1.9 Is
!
4
1
2
inside of the box formed by these three?
3
3
1
2
6
1
!
1
0
5
x
xy
x
2x
x
3x y
(a)
7
(b)
7
(c) y 7 x + y + z
y
3y
y
2x + y
z
y 2z
1.13 What is the area of the image of the rectangle [2..4] [2..5] under the action
of this matrix?
2
3
4 1
1.14 If t : R3 R3 changes volumes by a factor of 7 and s : R3 R3 changes volumes by a factor of 3/2 then by what factor will their composition changes volumes?
1.15 In what way does the definition of a box differ from the defintion of a span?
X 1.16 Why doesnt this picture contradict Theorem 1.5?
2 1
0 1
area is 2
determinant is 2
area is 5
324
Chapter 4. Determinants
X 1.19 Let T be the matrix representing (with respect to the standard bases) the
map that rotates plane vectors counterclockwise thru radians. By what factor
does T change sizes?
X 1.20 Must a transformation t : R2 R2 that preserves areas also preserve lengths?
X 1.21 What is the volume of a parallelepiped in R3 bounded by a linearly dependent
set?
X 1.22 Find the area of the triangle in R3 with endpoints (1, 2, 1), (3, 1, 4), and
(2, 2, 2). (Area, not volume. The triangle defines a planewhat is the area of the
triangle in that plane?)
X 1.23 An alternate proof of Theorem 1.5 uses the definition of determinant functions.
(a) Note that the vectors forming S make a linearly dependent set if and only if
|S| = 0, and check that the result holds in this case.
(b) For the |S| 6= 0 case, to show that |T S|/|S| = |T | for all transformations,
consider the function d : Mnn R given by T 7 |T S|/|S|. Show that d has
the first property of a determinant.
(c) Show that d has the remaining three properties of a determinant function.
(d) Conclude that |T S| = |T | |S|.
1.24 Give a non-identity matrix with the property that Atrans = A1 . Show that
if Atrans = A1 then |A| = 1. Does the converse hold?
1.25 The algebraic property of determinants that factoring a scalar out of a single
row will multiply the determinant by that scalar shows that where H is 33, the
determinant of cH is c3 times the determinant of H. Explain this geometrically,
that is, using Theorem 1.5,
X 1.26 Matrices H and G are said to be similar if there is a nonsingular matrix P
such that H = P 1 GP (we will study this relation in Chapter Five). Show that
similar matrices have the same determinant.
1.27 We usually represent vectors in R2 with respect to the standard basis so
vectors in the first quadrant have both coordinates positive.
~
v
RepE2 (~v ) =
+3
+2
+
+
B=h
0
1
,
i
1
0
gives the same counterclockwise cycle. We say these two bases have the same
orientation.
(a) Why do they give the same cycle?
(b) What other configurations of unit vectors on the axes give the same cycle?
(c) Find the determinants of the matrices formed from those (ordered) bases.
(d) What other counterclockwise cycles are possible, and what are the associated
determinants?
(e) What happens in R1 ?
(f ) What happens in R3 ?
A fascinating general-audience discussion of orientations is in [Gardner].
325
1.28 This question uses material from the optional Determinant Functions Exist
subsection. Prove Theorem 1.5 by using the permutation expansion formula for
the determinant.
X 1.29 (a) Show that this gives the equation of a line in R2 thru (x2 , y2 ) and (x3 , y3 ).
y
1
x2
y2
1
x3
y3 = 0
1
(b) [Petersen] Prove that the area of a triangle with vertices (x1 , y1 ), (x2 , y2 ),
and (x3 , y3 ) is
x
1 1
y1
2
1
x2
y2
1
x3
y3 .
1
(c) [Math. Mag., Jan. 1973] Prove that the area of a triangle with vertices at
(x1 , y1 ), (x2 , y2 ), and (x3 , y3 ) whose coordinates are integers has an area of N
or N/2 for some positive integer N .
326
4.III
Chapter 4. Determinants
Other Formulas
4.III.1
Laplaces Expansion
1
t1,1 t1,2 t1,3
0
t3,1 t3,2 t3,3
expansion
1 0 0
0 0
0 1 0
0 1
0 0 1
0 0
we can, for instance, factor out the entries from the first row
1 0 0
1 0 0
0 1 0
0 1 0
0 0 1
0 0 1
1 0 0
1 0
1 0 0
1
1 0 0
1
0
1
0
0 0
0 1
1 0
0 0
0 1
1 0
0
1
0
1
0
0
327
The point of the swapping (one swap to each of the permutation matrices on
the second line and two swaps to each on the third line) is that the three lines
simplify to three terms.
t2,2 t2,3
t2,1 t2,3
t2,1 t2,2
t1,2
+ t1,3
= t1,1
t3,2 t3,3
t3,1 t3,3
t3,1 t3,2
The formula given in Theorem 1.5, which generalizes this example, is a recurrence the determinant is expressed as a combination of determinants. This
formula isnt circular because, as here, the determinant is expressed in terms of
determinants of matrices of smaller size.
1.2 Definition For any nn matrix T , the (n 1)(n 1) matrix formed by
deleting row i and column j of T is the i, j minor of T . The i, j cofactor Ti,j of
T is (1)i+j times the determinant of the i, j minor of T .
1.3 Example The 1, 2 cofactor of the matrix from Example 1.1 is the negative
of the second 22 determinant.
t
t
T1,2 = 1 2,1 2,3
t3,1 t3,3
1.4 Example Where
1
T = 4
7
these are the 1, 2 and 2, 2 cofactors.
4 6
=6
T1,2 = (1)1+2
7 9
2
5
8
3
6
9
1 3
= 12
T2,2 = (1)2+2
7 9
1 2 3
|T | = 4 5 6
7 8 9
QED
328
Chapter 4. Determinants
5 6
+ 2 (1) 4 6 + 3 (+1) 4 5 = 3 + 12 9 = 0
|T | = 1 (+1)
7 8
7 9
8 9
Alternatively, we can expand down the second column.
1 3
1 3
4 6
= 12 60 + 48 = 0
+ 8 (1)
+ 5 (+1)
|T | = 2 (1)
4 6
7 9
7 9
1.7 Example A row or column with many zeroes suggests a Laplace expansion.
1 5 0
3 1 0
We finish by applying this result to derive a new formula for the inverse
of a matrix. With Theorem 1.5, the determinant of an n n matrix T can
be calculated by taking linear combinations of entries from a row and their
associated cofactors.
ti,1 Ti,1 + ti,2 Ti,2 + + ti,n Ti,n = |T |
()
Recall that a matrix with two identical rows has a zero determinant. Thus, for
any matrix T , weighing the cofactors by entries from the wrong row row k
with k 6= i gives zero
ti,1 Tk,1 + ti,2 Tk,2 + + ti,n Tk,n = 0
()
because it represents the expansion along the row k of a matrix with row i equal
to row k. This equation summarizes () and ().
..
..
..
.
.
.
tn,1
tn,2
...
tn,n
T1,n
T2,n
...
Tn,n
...
|T |
Note that the order of the subscripts in the matrix of cofactors is opposite to
the order of subscripts in the other matrix; e.g., along the first row of the matrix
of cofactors the subscripts are 1, 1 then 2, 1, etc.
1.8 Definition The matrix adjoint to the square matrix T is
adj(T ) =
..
.
T1,n
where Tj,i is the j, i cofactor.
T2,n
...
Tn,n
329
QED
1.10 Example If
1 0
T = 2 1
1 0
then the adjoint adj(T ) is
1
1
1
1
1
1
0
4
1
1
1
2
0 4
0 1
1 4
1 1
1 0
1 0
4
1
4
3
=
1
1
0
1
0 4
3 9
0
1
1 0 4
1
0 4
3 0
0
2 1 1 3 3 9 = 0 3 0
1 0 1
1 0
1
0
0 3
1.11 Corollary If |T | =
6 0 then T 1 = (1/|T |) adj(T ).
1.12 Example The inverse of the matrix from Example 1.10 is (1/3)adj(T ).
1/3
0/3 4/3
1/3 0 4/3
1
3
T 1 = 3/3 3/3 9/3 = 1
1/3 0/3
1/3
1/3 0 1/3
The formulas from this section are often used for by-hand calculation and
are sometimes useful with special types of matrices. However, they are not the
best choice for computation with arbitrary matrices because they require more
arithmetic than, for instance, the Gauss-Jordan method.
Exercises
X 1.13 Find the cofactor.
T =
1
1
0
(a) T2,3
(b) T3,2
(c) T1,3
X 1.14 Find the determinant by expanding
3 0
1 2
1 3
0
1
2
2
3
1
2
0
330
Chapter 4. Determinants
2
1 4
1
4 3
3 1
1 1
(a) 1 0 2
(b)
(c)
(d) 1 0 3
2
4
5 0
1
0 1
1
8 9
X 1.17 Find the inverse of each matrix in the prior question with Theorem 1.9.
1.18 Find the matrix adjoint to this one.
2 1 0 0
1 2 1 0
0 1 2 1
0 0 1 2
X 1.19 Expand across the first row to derive the formula for the determinant of a 22
matrix.
X 1.20 Expand across the first row to derive the formula for the determinant of a 33
matrix.
X 1.21 (a) Give a formula for the adjoint of a 22 matrix.
(b) Use it to derive the formula for the inverse.
X 1.22 Can we compute a determinant by expanding down the diagonal?
1.23 Give a formula for the adjoint of a diagonal matrix.
X 1.24 Prove that the transpose of the adjoint is the adjoint of the transpose.
1.25 Prove or disprove: adj(adj(T )) = T .
1.26 A square matrix is upper triangular if each i, j entry is zero in the part above
the diagonal, that is, when i > j.
(a) Must the adjoint of an upper triangular matrix be upper triangular? Lower
triangular?
(b) Prove that the inverse of a upper triangular matrix is upper triangular, if an
inverse exists.
1.27 This question requires material from the optional Determinants Exist subsection. Prove Theorem 1.5 by using the permutation expansion.
1.28 Prove that the determinant of a matrix equals the determinant of its transpose
using Laplaces expansion and induction on the size of the matrix.
1.29 [Am. Math. Mon., Jun. 1949] Show that
1 1
1 1
1
F n = 0
0 0
.
.
1
0
1
1
.
1
1
0
1
.
1
0
1
0
.
1
1
0
1
.
. . .
. . .
. . .
. . .
. . .
331
1
2
6
+ x2
=
3
1
8
1
3
+ x2
2
1
1
3
2
1
x1
1
3
1
3
2
1
2
1
2
1
The second box is formed from x1 13 and 21 , and one of the properties of the
size function (that is, the determinant function) is that the size of the second
box is therefore
x
1timesthe
size of the first box. Since the third box is formed
from x1 13 + x2 21 and 21 , and sizes are unchanged by side operations (that
332
Chapter 4. Determinants
is, the determinant is unchanged by adding x2 times the second column to the
first column), the size of the third box equals the size of the second box.
6 2
= x1 1 2
3 1
8 1
Solving gives the value of one of the
8
x1 =
1
3
variables.
2
1
10
=
=2
5
2
1
The theorem that generalizes this example, Cramers Rule, is: if |A| =
6 0
then the system A~x = ~b has the unique solution xi = |Bi |/|A| where the matrix
Bi is formed from A by replacing column i with the vector ~b. Exercise 3 asks
for a proof.
For instance, to solve this system for x2
1 0 4
x1
2
2 1 1 x2 = 1
x3
1 0 1
1
we do this computation.
1 2
4
2 1 1
1 1 1
18
=
x2 =
3
1 0 4
2 1 1
1 0 1
Cramers Rule, with practice, allows us to solve two equations/two unknown
systems by eye. It is also sometimes used for three equations/three unknowns
systems. But computing large determinants takes a long time so that solving
large systems by Cramers Rule is impractical.
Exercises
1 Use Cramers Rule to solve each for each of the variables.
x y= 4
2x + y = 2
(a)
(b)
x + 2y = 7
x 2y = 2
2 Use Cramers Rule to solve this system for z.
2x + y + z = 1
3x
+z=4
xyz=2
3 Prove Cramers Rule.
333
4 Suppose that a linear system with as many equations as unknowns, and with
integer coefficients and constants, has a matrix of coefficients with determinant 1.
Prove that the entries in the solution are all integers. (Remark. This is often used
to invent linear systems for exercises. If an instructor makes the linear system with
this property then the solution is not some disagreeable fraction.)
5 Use Cramers Rule to give a formula for the solution of a two equation/two
unknown linear system.
6 Can Cramers Rule tell the difference between a system with no solutions and
one with infinitely many?
334
Chapter 4. Determinants
..
.
permutations
335
The outermost loop (not shown) runs through N 1 rows. For each row, the
nested I and J loops shown perform arithmetic on the entries in A that are
below and to the right of the pivot entry. Assume that the pivot is found in
the expected place, that is, that COL = ROW . Then there are (N ROW )2
entries below and to the right of the pivot. On average, ROW will be N/2.
Thus, we estimate that the arithmetic will be performed about (N/2)2 times,
that is, will run in a time proportional to the square of the number of equations.
Taking into account the outer loop that is not shown, we get the estimate that
the running time of the algorithm is proportional to the cube of the number of
equations.
Finding the fastest algorithm to compute the determinant is a topic of current research. Algorithms are known that run in time between the second and
third power.
Speed estimates like these help us to understand how quickly or slowly an
algorithm will run. Algorithms that run in time proportional to the size of the
data set are fast, algorithms that run in time proportional to the square of the
size of the data set are less fast, but typically quite usable, and algorithms that
run in time proportional to the cube of the size of the data set are still reasonable
in speed. However, algorithms that run in time (greater than or equal to) the
factorial of the size of the data set are not practical.
There are other methods besides the two discussed here that are also used
for computation of determinants. Those lie outside of our scope. Nonetheless,
this contrast of the two methods for computing determinants makes the point
that although in principle they give the same answer, in practice the idea is to
select the one that is fast.
Exercises
Most of these problems presume access to a computer.
1 Computer systems generate random numbers (of course, these are only pseudorandom, in that they are generated by an algorithm, but they pass a number of
reasonable statistical tests for randomness).
(a) Fill a 55 array with random numbers (say, in the range [0..1)). See if it is
singular. Repeat that experiment a few times. Are singular matrices frequent
or rare (in this sense)?
336
Chapter 4. Determinants
(b) Time your computer algebra system at finding the determinant of ten 55
arrays of random numbers. Find the average time per array. Repeat the prior
item for 1515 arrays, 2525 arrays, and 3535 arrays. (Notice that, when an
array is singular, it can sometimes be found to be so quite quickly, for instance
if the first row equals the second. In the light of your answer to the first part,
do you expect that singular systems play a large role in your average?)
(c) Graph the input size versus the average time.
2 Compute the determinant of each of these by hand using the two methods discussed above.
2 1
0
0
3
1
1
2 1
2
0
(b) 1 0 5 (c) 1 3
(a)
5 3
0 1 2 1
1 2 2
0 0 2 1
Count the number of multiplications and divisions used in each case, for each of
the methods. (On a computer, multiplications and divisions take much longer than
additions and subtractions, so algorithm designers worry about them more.)
3 What 1010 array can you invent that takes your computer system the longest
to reduce? The shortest?
4 Write the rest of the FORTRAN program to do a straightforward implementation
of calculating determinants via Gauss method. (Dont test for a zero pivot.)
Compare the speed of your code to that used in your computer algebra system.
5 The FORTRAN language specification requires that arrays be stored by column, that is, the entire first column is stored contiguously, then the second column, etc. Does the code fragment given take advantage of this, or can it be
rewritten to make it faster, by taking advantage of the fact that computer fetches
are faster from contiguous locations?
337
What is there in the room, for instance where the ceiling meets the left and
right walls, are lines that are parallel. However, what a viewer sees is lines
that, if extended, would intersect. The intersection point is called the vanishing
point. This aspect of perspective is also familiar as the image of a long stretch
of railroad tracks that appear to converge at the horizon.
To depict the room, da Vinci has adopted a model of how we see, of how we
project the three dimensional scene to a two dimensional image. This model is
only a first approximation it does not take into account that our retina is
curved and our lens bends the light, that we have binocular vision, or that our
brains processing greatly affects what we see but nonetheless it is interesting,
both artistically and mathematically.
The projection is not orthogonal, it is a central projection from a single
point, to the plane of the canvas.
A
B
C
(It is not an orthogonal projection since the line from the viewer to C is not
orthogonal to the image plane.) As the picture suggests, the operation of central projection preserves some geometric properties lines project to lines.
However, it fails to preserve some others equal length segments can project
to segments of unequal length; the length of AB is greater than the length of
BC because the segment projected to AB is closer to the viewer and closer
things look bigger. The study of the effects of central projections is projective
geometry. We will see how linear algebra can be used in this study.
338
Chapter 4. Determinants
There are three cases of central projection. The first is the projection done
by a movie projector.
projector P
image I
source S
We can think that each source point is pushed from the domain plane outward to the image point in the codomain plane. This case of projection has a
somewhat different character than the second case, that of the artist pulling
the source back to the canvas.
painter P
source S
image I
In the first case S is in the middle while in the second case I is in the middle.
One more configuration is possible, with P in the middle. An example of this
is when we use a pinhole to shine the image of a solar eclipse onto a piece of
paper.
source S
pinhole P
image I
All three projection cases appear here. The first picture below shows P acting
like a movie projector by pushing points from part of S out to image points on
the lower half of I. The middle picture shows P acting like the artist by pulling
points from another part of S back to image points in I. In the third picture, P
acts like the pinhole. This picture is the trickiestthe points that are projected
near to the vanishing point are the ones that are far out to the bottom left of
S. Points in S that are near to the vertical dotted line are sent high up on I.
339
There are two awkward things about this situation. The first is that neither of
the two points in the domain nearest to the vertical dotted line (see below) has
an image because a projection from those two is along the dotted line that is
parallel to the codomain plane (we sometimes say that these two are projected
to infinity). The second awkward thing is that the vanishing point in I isnt
the image of any point from S because a projection to this point would be along
the dotted line that is parallel to the domain plane (we sometimes say that the
vanishing point is the image of a projection from infinity).
For a better model, put the projector P at the origin. Imagine that P is
covered by a glass hemispheric dome. As P looks outward, anything in the line
of vision is projected to the same spot on the dome. This includes things on
the line between P and the dome, as in the case of projection by the movie
projector. It includes things on the line further from P than the dome, as in
the case of projection by the painter. It also includes things on the line that lie
behind P , as in the case of projection by a pinhole.
1
` = {k 2 k R}
3
From this perspective P , all of the spots on the line are seen as the same point.
Accordingly, for any nonzero vector ~v R3 , we define the associated point v
in the projective plane to be the set {k~v k R and k 6= 0} of nonzero vectors
lying on the same line through the origin as ~v . To describe a projective point
we can give any representative member of the line, so that the projective point
shown above can be represented in any of these three ways.
1
1/3
2
2
2/3
4
3
1
6
340
Chapter 4. Determinants
While drawing a point as a pair of antipodal spots is not as natural as the onespot-per-point dome mode, on the other hand the awkwardness of the dome
model is gone, in that if as a line of view slides from north to south, no sudden
changes happen on the picture. This model of central projection is uniform
the three cases are reduced to one.
So far we have described points in projective geometry. What about lines?
What a viewer at the origin sees as a line is shown below as a great circle, the
intersection of the model sphere with a plane through the origin.
(One of the projective points on this line is shown to bring out a subtlety.
Because two antipodal spots together make up a single projective point, the
great circles behind-the-paper part is the same set of projective points as its
in-front-of-the-paper part.) Just as we did with each projective point, we will
also describe a projective line with a triple of reals. For instance, the members
341
project to a line that we can described with the triple 1 1 1 (we use row
vectors to typographically distinguish lines from points). In general, for any
~ we define the associated line in the projective
nonzero three-wide row vector L
~
~ k R and k 6= 0} of nonzero multiples of L.
plane, to be the set L = {k L
The reason that this description of a line as a triple is convienent is that
in the projective plane, a point v and a line L are incident the point lies
on the line, the line passes throught the point if and only if a dot product
of their representatives v1 L1 + v2 L2 + v3 L3 is zero (Exercise 4 shows that this
~
is independent of the choice of representatives ~v and L).
For instance, the
projective point described above by the column
vector
with
components 1, 2,
342
Chapter 4. Determinants
()
Note though that projective points on the equator dont project up to the plane.
Instead, these project out to infinity. We can thus think of projective space
as consisting of the Euclidean plane with some extra points adjoined the
Euclidean plane is embedded in the projective plane. These extra points, the
equatorial points, are the ideal points or points at infinity and the equator is the
ideal line or line at infinity (note that it is not a Euclidean line, it is a projective
line).
The advantage of the extension to the projective plane is that some of the
awkwardness of Euclidean geometry disappears. For instance, the projective
lines shown above in () cross at antipodal spots, a single projective point, on
the spheres equator. If we put those lines into () then they correspond to
Euclidean lines that are parallel. That is, in moving from the Euclidean plane to
the projective plane, we move from having two cases, that lines either intersect
or are parallel, to having only one case, that lines intersect (possibly at a point
at infinity).
The projective case is nicer in many ways than the Euclidean case but has
the problem that we dont have the same experience or intuitions with it. Thats
one advantage of doing analytic geometry, where the equations can lead us to
the right conclusions. Analytic projective geometry uses linear algebra. For
instance, for three points of the projective plane t, u, and v, setting up the
equations for those points by fixing vectors representing each, shows that the
three are collinear incident in a single line if and only if the resulting threeequation system has infinitely many row vector solutions representing that line.
That, in turn, holds if and only if this determinant is zero.
t 1 u 1 v1
t 2 u 2 v2
t 3 u 3 v3
Thus, three points in the projective plane are collinear if and only if any three
representative column vectors are linearly dependent. Similarly (and illustrating
the Duality Principle), three lines in the projective plane are incident on a
single point if and only if any three row vectors representing them are linearly
dependent.
343
The following result is more evidence of the niceness of the geometry of the
projective plane, compared to the Euclidean case. These two triangles are said
to be in perspective from P because their corresponding vertices are collinear.
O
T1
U1
T2
V1
U2
V2
We will prove this theorem, using projective geometry. (These are drawn as
Euclidean figures because it is the more familiar image. To consider them as
projective figures, we can imagine that, although the line segments shown are
parts of great circles and so are curved, the model has such a large radius
compared to the size of the figures that the sides appear in this sketch to be
straight.)
For this proof, we need a preliminary lemma [Coxeter]: if W , X, Y , Z are
four points in the projective plane (no three of which are collinear) then there
is a basis B for R3 such that
1
0
0
1
~ = 0 RepB (~x) = 1 RepB (~y ) = 0 RepB (~z) = 1
RepB (w)
0
0
1
1
where w,
~ ~x, ~y , ~z are homogeneous coordinate vectors for the projective points.
The proof is straightforward. Because W, X, Y are not on the same projective
line, any homogeneous coordinate vectors w
~ 0 , ~x0 , ~y0 do not line on the same
plane through the origin in R3 and so form a spanning set for R3 . Thus any
homogeneous coordinate vector for Z can be written as a combination ~z0 =
~ = aw
~ 0 , ~x = b ~x0 , ~y = c ~y0 , and ~z = ~z0 will do,
aw
~ 0 + b ~x0 + c ~y0 . Then w
for B = hw,
~ ~x, ~y i.
Now, to prove of Desargues Theorem, use the lemma to fix homogeneous
coordinate vectors and a basis.
1
0
0
1
RepB (~t1 ) = 0 RepB (~u1 ) = 1 RepB (~v1 ) = 0 RepB (~o) = 1
0
0
1
1
344
Chapter 4. Determinants
Because the projective point T2 is incident on the projective line OT1 , any
homogeneous coordinate vector for T2 lies in the plane through the origin in R3
that is spanned by homogeneous coordinate vectors of O and T1 :
1
1
RepB (~t2 ) = a 1 + b 0
1
0
for some scalars a and b. That is, the homogenous coordinate vectors of members
T2 of the line OT1 are of the form on the left below, and the forms for U2 and
V2 are similar.
t2
1
1
RepB (~v2 ) = 1
RepB (~u2 ) = u2
RepB (~t2 ) = 1
1
1
v2
The projective line T1 U1 is the image of a plane through the origin in R3 . A
quick way to get its equation is to note that any vector in it is linearly dependent
on the vectors for T1 and U1 and so this determinant is zero.
1 0 x
0 1 y = 0
=
z=0
0 0 z
The equation of the plane in R3 whose image is the projective line T2 U2 is this.
t2 1 x
1 u2 y = 0
=
(1 u2 ) x + (1 t2 ) y + (t2 u2 1) z = 0
1 1 z
Finding the intersection of the two is routine.
t2 1
T1 U1 T2 U2 = 1 u2
0
(This is, of course, the homogeneous coordinate vector of a projective point.)
The other two intersections are similar.
1 t2
0
U1 V1 U2 V2 = u2 1
T 1 V 1 T2 V 2 = 0
v2 1
1 v2
The proof is finished by noting that these projective points are on one projective
line because the sum of the three homogeneous coordinate vectors is zero.
Every projective theorem has a translation to a Euclidean version, although
the Euclidean result is often messier to state and prove. Desargues theorem
illustrates this. In the translation to Euclidean space, the case where O lies on
345
the ideal line must be treated separately for then the lines T1 T2 , U1 U2 , and V1 V2
are parallel.
The parenthetical remark following the statement of Desargues Theorem
suggests thinking of the Euclidean pictures as figures from projective geometry
for a model of very large radius. That is, just as a small area of the earth appears
flat to people living there, the projective plane is also locally Euclidean.
Although its local properties are the familiar Euclidean ones, there is a global
property of the projective plane that is quite different. The picture below shows
a projective point. At that point is drawn an xy-axis. There is something
interesting about the way this axis appears at the antipodal ends of the sphere.
In the northern hemisphere, where the axis are drawn in black, a right hand put
down with fingers on the x-axis will have the thumb point along the y-axis. But
the antipodal axis, drawn in gray, has just the opposite: a right hand placed with
its fingers on the x-axis will have the thumb point in the wrong way, instead,
a left hand comes out correct. Briefly, the projective plane is not orientable: in
this geometry, left and right handedness are not fixed properties of figures.
At the end of the circuit, the x arrow from the xy-axis sticks out in the other
direction. Another example of the same effect is that a clockwise spiral, on taking the same trip, would switch to counterclockwise. (This orientation reversal
appeared earlier, in the pinhole/eclipse picture.)
This exhibition of the existence of a non-orientable space raises the question
of whether our space is orientable: is is possible for a right handed astronaut to
take a trip to Mars and return left handed? An excellent nontechnical reference
is [Gardner]. An intriguing science fiction story about orientation reversal is
[Clarke].
So projective geometry is mathematically interesting and rewarding, in addition to the natural way in which it arises in art. It is more than just a technical
device to shorten some proofs. For an overview, see [Courant & Robbins]. The
approach weve taken here, the analytic approach, leads to quick theorems and
most importantly for us illustrates the power of linear algebra (see [Hanes],
346
Chapter 4. Determinants
[Ryan], and [Eggar]). But, note that the other possible approach, the synthetic
approach of deriving the results from an axiom system, is both extraordinarily
beautiful and is also the historical route of development. Two fine sources for
this approach are [Coxeter] or [Seidenberg]. An interesting, easy, application is
[Davies]
Exercises
1 What is the equation of this point?
!
1
0
0
(a) Find the line incident on these points in the projective plane.
! !
1
2
3
4
5
6
3 , 4
3 Find the formula for the line incident on two projective points. Find the formula
for the point incident on two projective lines.
4 Prove that the definition of incidence is independent of the choice of the representatives of p and L. That is, if p1 , p2 , p3 , and q1 , q2 , q3 are two triples of
homogeneous coordinates for p, and L1 , L2 , L3 , and M1 , M2 , M3 are two triples
of homogeneous coordinates for L, prove that p1 L1 + p2 L2 + p3 L3 = 0 if and only
if q1 M1 + q2 M2 + q3 M3 = 0.
5 Give a drawing to show that central projection does not preserve circles, that a
circle may project to an ellipse. Can a (non-circular) ellipse project to a circle?
6 Give the formula for the correspondence between the non-equatorial part of the
antipodal modal of the projective plane, and the plane z = 1.
7 (Pappuss Theorem) Assume that T0 , U0 , and V0 are collinear and that T1 , U1 ,
and V1 are collinear. Consider these three points: (i) the intersection V2 of the lines
T0 U1 and T1 U0 , (ii) the intersection U2 of the lines T0 V1 and T1 V0 , and (iii) the
intersection T2 of U0 V1 and U1 V0 .
(a) Draw a (Euclidean) picture.
(b) Apply the lemma used in Desargues Theorem to get simple homogeneous
coordinate vectors for the T s and V0 .
(c) Find the resulting homogeneous coordinate vectors for U s (these must each
involve a parameter as, e.g., U0 could be anywhere on the T0 V0 line).
(d) Find the resulting homogeneous coordinate vectors for V1 . (Hint: it involves
two parameters.)
(e) Find the resulting homogeneous coordinate vectors for V2 . (It also involves
two parameters.)
(f ) Show that the product of the three parameters is 1.
(g) Verify that V2 is on the T2 U2 line..
Chapter 5
Similarity
While studying matrix equivalence, we have shown that for any any homomorphism there are bases B and D such that the representation matrix has a block
partial-identity form.
RepB,D (h) =
Identity
Zero
Zero
Zero
~1 + + cn
~n to
This representation lets us think of the map as sending c1
~
~
~
~
c1 1 + + ck k + 0 + + 0, where n is the dimension of the domain and k is
the dimension of the range. So, under this representation the action of the map
is easy to understand because most of the matrix entries are zero.
This chapter considers the special case where the domain and the codomain
are equal, that is, where the homomorphism is a transformation. In this case
we naturally ask to find a single basis B so that RepB,B (t) is as simple as
possible (we will take simple to mean that it has many zeroes). A matrix
having the above block partial-identity form is not always possible here. But,
we will develop a form that comes close the representation will be nearly
diagonal.
5.I
This chapter will require that we factor polynomials. Of course, many polynomials do not factor over the real numbers. For instance, x2 + 1 does not factor
into the product of two linear polynomials with real coefficients. For that reason, we shall from now on take our scalars from the complex numbers. In this
chapter the cs in c1~v1 + c2~v2 + + cn~vn will be complex numbers.
So we are shifting from studying vector spaces over the real numbers to
vector spaces over the complex numbers. As a consequence, in this chapter
vector and matrix entries are complex. (The real numbers are a subset of the
complex numbers, and a quick glance through this chapter shows that most of
347
348
Chapter 5. Similarity
the examples use only pure-real numbers. Nonetheless, the critical theorems
require the use of the complex number system to go through.) Therefore, the
first section of this chapter is a quick review of complex numbers.
The idea of taking scalars from a structure other than the real numbers is
an interesting one. However, in this book we are moving to this more general
context only for the pragmatic reason that we must do so in order to develop
the representation. We will not go into using other sets of scalars in more
detail because it could distract from our task. For the more general approach,
delightful presentations are in [Halmos] or [Hoffman & Kunze].
5.I.1
This subsection is a review only and we take the main results as known.
For proofs, see [Birkhoff & MacLane] or [Ebbinghaus].
Just as integers have a division operation e.g., 4 goes 5 times into 21
with remainder 1 so do polynomials.
1.1 Theorem (Division Theorem for Polynomials) Let c(x) be a polynomial. If m(x) is a non-zero polynomial then there are quotient and remainder
polynomials q(x) and r(x) such that
c(x) = m(x) q(x) + r(x)
where the degree of r(x) is strictly less than the degree of m(x).
In this book constant polynomials, including the zero polynomial, are said to
have degree 0. (This is not the standard definition, but it is convienent here.)
The point of the integer division statement 4 goes 5 times into 21 with
remainder 1 is that the remainder is less than 4 while 4 goes 5 times, it does
not go 6 times. In the same way, the point of the polynomial division statement
is its final clause.
1.2 Example If c(x) = 2x3 3x2 + 4x and m(x) = x2 + 1 then q(x) = 2x 3
and r(x) = 2x + 3. Note that r(x) has a lower degree than m(x).
1.3 Corollary The remainder when c(x) is divided by x is the constant
polynomial r(x) = c().
Proof. The remainder must be a constant polynomial. because it is of degree
less than the divisor x , To determine the constant, taking m(x) from the
theorem to be x and substituting for x yields c() = ( ) q() +
QED
r(x).
If a divisor m(x) goes into a dividend c(x) evenly, meaning that r(x) is the
zero polynomial, then m(x) is a factor of c(x). Any root of the factor (any
R such that m() = 0) is a root of c(x) since c() = m() q() = 0. The
prior corollary immediately yields the following converse.
349
b + b2 4ac
b b2 4ac
2 =
1 =
2a
2a
(if the discriminant b2 4ac is negative then the polynomial has no real number
roots). A polynomial that cannot be factored into two lower-degree polynomials
with real number coefficients is irreducible over the reals.
1.5 Theorem Any constant or linear polynomial is irreducible over the reals.
A quadratic polynomial is irreducible over the reals if and only if its discriminant
is negative. No cubic or higher-degree polynomial is irreducible over the reals.
1.6 Corollary Any polynomial with real coefficients can be factored into linear
and irreducible quadratic polynomials. That factorization is unique; any two
factorizations have the same powers of the same factors.
Note the analogy with the prime factorization of integers. In both cases, the
uniqueness clause is very useful.
1.7 Example Because of uniqueness we know, without multiplying them out,
that (x + 3)2 (x2 + 1)3 does not equal (x + 3)4 (x2 + x + 1)2 .
1.8 Example By uniqueness, if c(x) = m(x) q(x) then where c(x) = (x
3)2 (x + 2)3 and m(x) = (x 3)(x + 2)2 , we know that q(x) = (x 3)(x + 2).
While x2 + 1 has no real roots and so doesnt factor over the real numbers,
if we imagine a root traditionally denoted i so that i2 + 1 = 0 then x2 + 1
factors into a product of linears (x i)(x + i).
So we adjoin this root i to the reals and close the new system with respect
to addition, multiplication, etc. (i.e., we also add 3 + i, and 2i, and 3 + 2i, etc.,
putting in all linear combinations of 1 and i). We then get a new structure, the
complex numbers, denoted C.
In C we can factor (obviously, at least some) quadratics that would be irreducible if we were to stick to the real numbers. Surprisingly, in C we can
not only factor x2 + 1 and its close relatives, we can factor any quadratic. Any
quadratic polynomial factors over the complex numbers.
b b2 4ac
b + b2 4ac
2
x
ax + bx + c = a x
2a
2a
1.9 Example The second degree polynomial x2 +x+1 factors over the complex
numbers into the product of two first degree polynomials.
1 3
1
1
3
3
1 + 3
x
= x ( +
i) x (
i)
x
2
2
2
2
2
2
350
Chapter 5. Similarity
5.I.2
Complex Representations
1 + 1i 2 0i
1 + 0i 1 0i
i
2 + 3i
3i
i
1 + 7i 1 1i
=
9 5i 3 + 3i
Everything else from prior chapters that we can, we shall also carry over
unchanged. For instance, we shall call this
0 + 0i
1 + 0i
0 + 0i
0 + 0i
h . , . . . , . i
..
..
0 + 0i
1 + 0i
the standard basis for Cn as a vector space over C and again denote it En .
5.II
5.II.1
351
Similarity
Definition and Examples
idy
Vw.r.t.
Ww.r.t.
H
idy
Ww.r.t.
idy
Vw.r.t.
Vw.r.t.
idy
Vw.r.t.
To move from the lower left to the lower right we can either go straight over, or
up, over, and then down. In matrix terms,
1
RepD,D (t) = RepB,D (id) RepB,B (t) RepB,D (id)
(recall that a representation of composition like this one reads right to left).
1.1 Definition The matrices T and S are similar if there is a nonsingular P
such that T = P SP 1 .
Since nonsingular matrices are square, the similar matrices T and S must be
square and of the same size.
1.2 Example With these two,
2 1
P =
1 1
S=
2
1
3
1
0 1
T =
1 1
1.3 Example The only matrix similar to the zero matrix is itself: P ZP 1 =
P Z = Z. The only matrix similar to the identity matrix is itself: P IP 1 =
P P 1 = I.
352
Chapter 5. Similarity
Since matrix similarity is a special case of matrix equivalence, if two matrices are similar then they are equivalent. What about the converse: must
matrix equivalent square matrices be similar? The answer is no. The prior
example shows that the similarity classes are different from the matrix equivalence classes, because the matrix equivalence class of the identity consists of
all nonsingular matrices of that size. Thus, for instance, these two are matrix
equivalent but not similar.
1 0
1 2
T =
S=
0 1
0 3
So some matrix equivalence classes split into two or more similarity classes
similarity gives a finer partition than does equivalence. This picture shows some
matrix equivalence classes subdivided into similarity classes.
%
.A
...
.
B$
A matrix-equivalent to,
S=
1
2
3
6
T =
0
11/2
0
5
P =
4
3
2
2
check that T = P SP 1 .
X 1.5 Example 1.3 shows that the only matrix similar to a zero matrix is itself and
that the only matrix similar to the identity is itself.
(a) Show that the 11 matrix (2), also, is similar only to itself.
(b) Is a matrix of the form cI for some scalar c similar only to itself?
(c) Is a diagonal matrix similar only to itself?
1.6 Show that these matrices are not similar.
!
1 0 4
1 0 1
1 1 3
0 1 1
2 1 7
3 1 2
353
1
2
X
X
X
X
X
3
0
1
1
1
2
and pick two bases, and represent t with respect to then T = RepB,B (t) and
S = RepD,D (t). Then compute the P and P 1 to change bases from B to D and
back again.
1.9 Explain Example 1.3 in terms of maps.
1.10 Are there two matrices A and B that are similar while A2 and B 2 are not
similar?
1.11 Prove that if two matrices are similar and one is invertible then so is the
other.
1.12 Show that similarity is an equivalence relation.
1.13 Consider a matrix representing, with respect to some B, B, reflection across
the x-axis in R2 . Consider also a matrix representing, with respect to some D, D,
reflection across the y-axis. Must they be similar?
1.14 Prove that similarity preserves determinants and rank. Does the converse
hold?
1.15 Is there a matrix equivalence class with only one matrix similarity class inside?
One with infinitely many similarity classes?
1.16 Can two different diagonal matrices be in the same similarity class?
1.17 Prove that if two matrices are similar then their k-th powers are similar when
k > 0. What if k 0?
1.18 Let p(x) be the polynomial cn xn + + c1 x + c0 . Show that if T is similar to
S then p(T ) = cn T n + + c1 T + c0 I is similar to p(S) = cn S n + + c1 S + c0 I.
1.19 List all of the matrix equivalence classes of 11 matrices. Also list the similarity classes, and describe which similarity classes are contained inside of each
matrix equivalence class.
1.20 Does similarity preserve sums?
1.21 Show that if T I and N are similar matrices then T and N + I are also
similar.
5.II.2
Diagonalizability
The prior subsection defines the relation of similarity and shows that, although similar matrices are necessarily matrix equivalent, the converse does not
hold. Some matrix-equivalence classes break into two or more similarity classes
(the nonsingular n n matrices, for instance). This means that the canonical form for matrix equivalence, a block partial-identity, cannot be used as a
354
Chapter 5. Similarity
?
%
...
?$
?
? !
?
?
is diagonalizable.
2
0
4
1
2
1
1
0
1 2
4 2
1 2
=
3
1 1
1 1
1 1
0 0
N=
1 0
is the zero matrix. Thus, for any map n that N represents (with respect to the
same basis for the domain as for the codomain), the composition n n is the
zero map. This implies that no such map n can be diagonally represented (with
respect to any B, B) because no power of a nonzero diagonal matrix is zero.
That is, there is no diagonal matrix in N s similarity class.
That example shows that a diagonal form will not do for a canonical form
we cannot find a diagonal matrix in each matrix similarity class. However, the
canonical form that we are developing has the property that if a matrix can
be diagonalized then the diagonal matrix is the canonical representative of the
similarity class. The next result characterizes which maps can be diagonalized.
355
..
..
1
0
.
.
..
.
..
~
~n )) =
RepB,B (t) =
RepB (t(
.
.
RepB (t(1 ))
..
..
..
n
0
.
.
3
T =
0
2
1
1 0
RepB,B (t) =
0 2
~2 ) = 2 .
~1 ) = 1 and t(
that is, such that t(
3
3 2 ~
~1
1 = 1
0
0 1
2 ~
~2
= 2
1 2
b
3 2
b1
=x 1
b2
b2
0 1
has solutions b1 and b2 , which are not both zero. Rewrite that as a linear system.
(3 x) b1 +
2 b2 = 0
(1 x) b2 = 0
()
In the bottom equation the two numbers multiply to give zero only if at least
one of them is zero so there are two possibilities, b2 = 0 and x = 1. In the b2 = 0
possibility, the first equation gives that either b1 = 0 or x = 3. Since the case
of both b1 = 0 and b2 = 0 is disallowed, we are left looking at the possibility of
x = 3. With it, the first equation in () is 0 b1 + 2 b2 = 0 and so associated
with 3 are vectors with a second component of zero and a first component that
is free.
b
3 2
b1
=3 1
0
0
0 1
356
Chapter 5. Similarity
0
In the x = 1 possibility, the first equation in () is 2 b1 + 2 b2 = 0, and so
associated with 1 are vectors whose second component is the negative of their
first component.
b1
3 2
b1
=1
b1
b1
0 1
Thus, another solution is 2 = 1 and a second basis vector is this.
1
~
2 =
1
To finish, drawing the similarity diagram
R2w.r.t.
idy
R2w.r.t.
E2
R2w.r.t.
T
idy
E2
R2w.r.t.
D
and noting that the matrix RepB,E2 (id) is easy leads to this diagonalization.
3 0
0 1
1
0
1
3 2
1 1
1
0 1
0 1
1
triangular
2 1
5 4
(a)
(b)
0
2
0 1
X 2.8 What form do the powers of a diagonal matrix have?
2.9 Give two same-sized diagonal matrices that are not similar. Must any two
different diagonal matrices come from different similarity classes?
2.10 Give a nonsingular diagonal matrix. Can a diagonal matrix ever be singular?
X 2.11 Show that the inverse of a diagonal matrix is the diagonal of the the inverses,
if no element on that diagonal is zero. What happens when a diagonal entry is
zero?
357
1
1
3 2
1
1
3 0
=
0 1
0 1
0 1
0 1
is a bit jarring because for P we must take the first matrix, which is shown as an
inverse, and for P 1 we take the inverse of the first matrix, so that the two 1
powers cancel and this matrix is shown without a superscript 1.
(a) Check that this nicer-appearing equation holds.
3
0
0
1
1
0
1
1
3
0
2
1
1
0
1
1
(b) Is the previous item a coincidence? Or can we always switch the P and the
P 1 ?
2.13 Show that the P used to diagonalize in Example 2.5 is not unique.
2.14 Find a formula for the powers of this matrix Hint: see Exercise 8.
3 1
4 2
X 2.15 Diagonalize
these.
1 1
0 1
(a)
(b)
0 0
1 0
2.16 We can ask how diagonalization interacts with the matrix operations. Assume
that t, s : V V are each diagonalizable. Is ct diagonalizable for all scalars c?
What about t + s? t s?
X 2.17 Show that matrices of this form are not diagonalizable.
1 c
c 6= 0
0 1
2.18 Show
isdiagonalizable.
that
each of these
1 2
x y
(a)
(b)
x, y, z scalars
2 1
y z
5.II.3
y 7 y
0
z
x, y, z C
358
Chapter 5. Similarity
2 0
T = RepB,B (t) =
0 0
then 2 is an eigenvalue of T , associated with these eigenvectors.
c0
2c0
c
c 2 0
=
} = { 0 c0 C, c0 6= 0}
{ 0
c1
c1
2c1
0
0 0
On the other hand, representing t with respect to D = h2 + 1x, 1 + 0xi gives
3 0
S = RepD,D (t) =
3 2
359
S associated with the eigenvalue 2 are these.
0
c0
2c0
0
c1 C, c1 6= 0}
=
}={
c1
2c1
2
c1
1 2 1
T = 2 0 2
1 2 3
~ bring everyTo find the scalars x such that T ~ = x~ for non-~0 eigenvectors ,
thing to the left-hand side
z1
1 2 1
z1
2 0 2 z2 x z2 = ~0
z3
z3
1 2 3
and factor (T xI)~ = ~0. (Note that it says T xI; the expression T x doesnt
make sense because T is a matrix while x is a scalar.) This homogeneous linear
system
0
1x
2
1
z1
2
0x
2 z2 = 0
z3
0
1
2
3x
has a non-~0 solution if and only if the matrix is nonsingular. We can determine
when that happens.
0 = |T xI|
1 x
2
1
0x
2
= 2
1
2
3 x
= x3 4x2 + 4x
= x(x 2)2
360
Chapter 5. Similarity
0
z1
a
10
2
1
z1
z2 = 0
2
0 0 2 z2 = 0
=
z3
z3
0
a
1
2
30
for a scalar parameter a 6= 0 (a is non-0 because eigenvectors must be non-~0).
In the same way,
0
z1
b
21
2
1
z1
2
z
b
2 0 2 z2 = 0
=
=
2
z3
z3
0
b
1
2
23
with b 6= 0.
3.8 Example If
S=
0
1
3
x
1
= (x )(x 3)
0
3x
so S has eigenvalues of 1 = and 2 = 3. To find associated eigenvectors, first
plug in 1 for x:
0
z1
a
1
z1
=
=
=
z2
z2
0
0
0
3
for a scalar a 6= 0, and then plug in 2 :
0
3
1
z1
=
z2
0
0
33
z1
(1/)b
=
z2
b
where b 6= 0.
3.9 Definition The characteristic polynomial of a square matrix T is the determinant of the matrix T xI, where x is a variable. The characteristic
equation is |T xI| = 0. The characteristic polynomial of a transformation t is
the polynomial of any RepB,B (t).
Exercise 30 checks that the characteristic polynomial of a transformation is
well-defined, that is, any choice of basis yields the same polynomial.
3.10 Lemma A linear transformation on a nontrivial vector space has at least
one eigenvalue.
361
Notice the familiar form of the sets of eigenvectors in the above examples.
3.11 Definition The eigenspace
of a transformation t associated with the
QED
3.13 Example In Example 3.8 the eigenspace associated with the eigenvalue
and the eigenspace associated with the eigenvalue 3 are these.
a
b/
a R}
V3 = {
b R}
V = {
0
b
3.14 Example In Example 3.7, these are the eigenspaces associated with the
eigenvalues 0 and 2.
a
b
V2 = {b b R}.
V0 = {0 a R},
a
b
3.15 Remark The characteristic equation is 0 = x(x 2)2 so in some sense 2
is an eigenvalue twice. However there are not twice as many eigenvectors,
in that the dimension of the eigenspace is one, not two. The next example shows
a case where a number, 1, is a double root of the characteristic equation and
the dimension of the associated eigenspace is two.
3.16 Example With respect to the
1
0
0
0 0
1 0
0 0
362
Chapter 5. Similarity
represents projection.
x
x
y 7 y
0
z
x, y, z C
Its eigenspace associated with the eigenvalue 0 and its eigenspace associated
with the eigenvalue 1 are easy to find.
0
c1
V1 = {c2 c1 , c2 C}
V0 = { 0 c3 C}
0
c3
By the lemma, if two eigenvectors ~v1 and ~v2 are associated with the same
eigenvalue then any linear combination of those two is also an eigenvector associated with that same eigenvalue. But, if two eigenvectors ~v1 and ~v2 are
associated with different eigenvalues then the sum ~v1 + ~v2 need not be related
to the eigenvalue of either one. In fact, just the opposite. If the eigenvalues are
different then the eigenvectors are not linearly related.
3.17 Theorem For any set of distinct eigenvalues of a map or matrix, a set of
associated eigenvectors, one per eigenvalue, is linearly independent.
Proof. We will use induction on the number of eigenvalues. If there is no eigenvalue or only one eigenvalue then the set of associated eigenvectors is empty or is
a singleton set with a non-~0 member, and in either case is linearly independent.
For induction, assume that the theorem is true for any set of k distinct eigenvalues, suppose that 1 , . . . , k+1 are distinct eigenvalues, and let ~v1 , . . . , ~vk+1
be associated eigenvectors. If c1~v1 + + ck~vk + ck+1~vk+1 = ~0 then after multiplying both sides of the displayed equation by k+1 , applying the map or matrix
to both sides of the displayed equation, and subtracting the first result from the
second, we have this.
2 2 2
0
1 1
4 8 3
are distinct: 1 = 1, 2 = 2, and 3 = 3. A set of associated eigenvectors like
2
9
2
{1 , 4 , 1}
0
4
2
is linearly independent.
363
QED
Exercises
3.20 For
and
each, find
the characteristic
polynomial
the eigenvalues.
(a)
10
4
9
2
1
4
(b)
2
3
(c)
0
7
3
0
(d)
0
0
0
0
1 0
(e)
0 1
X 3.21 For each matrix, find the characteristic equation, and the eigenvalues and
associated
eigenvectors.
3
0
3
2
(a)
(b)
8 1
1 0
3.22 Find the characteristic equation, and the eigenvalues and associated eigenvectors for this matrix. Hint. The eigenvalues are complex.
2
5
1
2
3.23 Find the characteristic polynomial, the eigenvalues, and the associated eigenvectors of this matrix.
!
1 1 1
0 0 1
0 0 1
X 3.24 For each matrix, find the characteristic equation, and the eigenvalues and
associated eigenvectors.
!
3
2 0
0
1
0
3
0
0
1
(a) 2
(b) 0
0
0
5
4 17 8
X 3.25 Let t : P2 P2 be
a0 + a1 x + a2 x2 7 (5a0 + 6a1 + 2a2 ) (a1 + 8a2 )x + (a0 2a2 )x2 .
Find its eigenvalues and the associated eigenvectors.
3.26 Find the eigenvalues and eigenvectors of this map t : M2 M2 .
a
c
b
d
2c
b 2c
a+c
d
X 3.27 Find the eigenvalues and associated eigenvectors of the differentiation operator
d/dx : P3 P3 .
3.28 Prove that the eigenvalues of a triangular matrix (upper or lower triangular)
are the entries on the diagonal.
X 3.29 Find the formula for the characteristic polynomial of a 22 matrix.
3.30 Prove that the characteristic polynomial of a transformation is well-defined.
X 3.31 (a) Can any non-~0 vector in any nontrivial vector space be a eigenvector?
That is, given a ~v 6= ~0 from a nontrivial V , is there a transformation t : V V
and a scalar R such that t(~v ) = ~v ?
(b) Given a scalar , can any non-~0 vector in any nontrivial vector space be an
eigenvector associated with the eigenvalue ?
364
Chapter 5. Similarity
X 3.32 Suppose that t : V V and T = RepB,B (t). Prove that the eigenvectors of T
associated with are the non-~0 vectors in the kernel of the map represented (with
respect to the same bases) by T I.
3.33 Prove that if a, . . . , d are all integers and a + b = c + d then
a
c
b
d
!
1
2
2
2
2
2
3 6 6
3.41 Suppose that P is a nonsingular nn matrix. Show that the similarity transformation map tP : Mnn Mnn sending T 7 P T P 1 is an isomorphism.
3.42 [Math. Mag., Nov. 1967] Show that if A is an n square matrix and each row
(column) sums to c then c is a characteristic root of A.
5.III
365
Nilpotence
The goal of this chapter is to show that every square matrix is similar to one
that is a sum of two kinds of simple matrices. The prior section focused on the
first kind, diagonal matrices. We now consider the other kind.
5.III.1
Self-Composition
Note that this power notation for the linear transformation functions dovetails
with the notation that weve used earlier for their square matrix representations
because if RepB,B (t) = T then RepB,B (tj ) = T j .
1.1 Example For the derivative map d/dx : P3 P3 given by
d/dx
a + bx + cx2 + dx3 7 6d
and any higher power is the zero map.
1.2 Example This transformation of the space of 22 matrices
b a
a b
t
7
d 0
c d
366
Chapter 5. Similarity
a b
c d
a b
c d
t2
t3
a b
0 0
b
0
a
0
367
c2 7 c2
c3
0
has C3 R() = R( 2 ) = and {~0 } N () = N ( 2 ) = .
1.6 Example Let t : P2 P2 be the map c0 + c1 x + c2 x2 7 2c0 + c2 x. As the
lemma describes, on iteration the rangespace shrinks
rank(tj )
As sketched, on iteration the rank falls and with it the nullity grows until the
two reach a steady state. This state must be reached by the n-th iterate. The
steady states distance above zero is the dimension of the generalized rangespace
and its distance below n is the dimension of the generalized nullspace.
1.7 Definition Let t be a transformation on an n-dimensional space. The
generalized rangespace (or the closure of the rangespace) is R (t) = R(tn ) The
generalized nullspace (or the closure of the nullspace) is N (t) = N (tn ).
368
Chapter 5. Similarity
Exercises
1.8 Give the chains of rangespaces and nullspaces for the zero and identity transformations.
1.9 For each map, give the chain of rangespaces and the chain of nullspaces, and
the generalized rangespace and the generalized nullspace.
(a) t0 : P2 P2 , a + bx + cx2 7 b + cx2
(b) t1 : R2 R2 ,
a
b
0
a
!
a
b
c
a
a
b
5.III.2
Strings
This subsection is optional, and requires material from the optional Direct Sum
subsection.
The prior subsection shows that as j increases, the dimensions of the R(tj )s
fall while the dimensions of the N (tj )s rise, in such a way that this rank and
nullity split the dimension of V . Can we say more; do the two split a basis
is V = R(tj ) N (tj )?
The answer is yes for the smallest power j = 0 since V = R(t0 ) N (t0 ) =
V {~0}. The answer is also yes at the other extreme.
2.1 Lemma Where t : V V is a linear transformation, the space is the direct
sum V = R (t) N (t). That is, both dim(V ) = dim(R (t)) + dim(N (t))
and R (t) N (t) = {~0 }.
369
Proof. We will verify the second sentence, which is equivalent to the first. The
first clause, that the dimension n of the domain of tn equals the rank of tn plus
the nullity of tn , holds for any transformation and so we need only verify the
second clause.
Assume that ~v R (t) N (t) = R(tn ) N (tn ), to prove that ~v is ~0.
Because ~v is in the nullspace, tn (~v ) = ~0. On the other hand, because R(tn ) =
R(tn+1 ), the map t : R (t) R (t) is a dimension-preserving homomorphism
and therefore is one-to-one. A composition of one-to-one maps is one-to-one,
and so tn : R (t) R (t) is one-to-one. But now because only ~0 is sent by
a one-to-one linear map to ~0 the fact that tn (~v ) = ~0 implies that ~v = ~0. QED
2.2 Note Technically we should distinguish the map t : V V from the map
t : R (t) R (t) because the domains or codomains might differ. The second
one is said to be the restriction of t to R(tk ). We shall use later a point from
that proof about the restriction map, namely that it is nonsingular.
In contrast to the j = 0 and j = n cases, for intermediate powers the space
V might not be the direct sum of R(tj ) and N (tj ). The next example shows
that the two can have a nontrivial intersection.
2.3 Example Consider the transformation of C2 defined by this action on the
elements of the standard basis.
0
0
0
0 0
1
n
n
7
7
N = RepE2 ,E2 (n) =
1
1
0
1 0
0
The vector
0
~e2 =
1
is in both the rangespace and nullspace. Another way to depict this maps
action is with a string.
~e1 7 ~e2 7 ~0
2.4 Example A map n
: C4 C4 whose action on E4 is given by the string
~e1 7 ~e2 7 ~e3 7 ~e4 7 ~0
n2 ) N (
n2 ) = [{~e3 , ~e4 }], and
has R(
n) N (
n) equal to the span [{~e4 }], has R(
3
3
n ) = [{~e4 }]. The matrix representation is all zeros except for
has R(
n ) N (
some subdiagonal ones.
0 0 0 0
1 0 0 0
= RepE ,E (
n) =
N
4 4
0 1 0 0
0 0 1 0
370
Chapter 5. Similarity
2.5 Example Transformations can act via more than one string. A transfor~1 , . . . , ~5 i by
mation t acting on a basis B = h
~1
7
~
7
4
~2
7
~5
~3 7 ~0
~0
is represented by a matrix that is all zeros except for blocks of subdiagonal ones
0 0 0 0 0
1 0 0 0 0
RepB,B (t) =
0 1 0 0 0
0 0 0 0 0
0 0 0 1 0
(the lines just visually organize the blocks).
In those three examples all vectors are eventually transformed to zero.
2.6 Definition A nilpotent transformation is one with a power that is the zero
map. A nilpotent matrix is one with a power that is the zero matrix. In either
case, the least such power is the index of nilpotency.
2.7 Example In Example 2.3 the index of nilpotency is two. In Example 2.4
it is four. In Example 2.5 it is three.
2.8 Example The differentiation map d/dx : P2 P2 is nilpotent of index
three since the third derivative of any quadratic polynomial is zero. This maps
action is described by the string x2 7 2x 7 2 7 0 and taking the basis
B = hx2 , 2x, 2i gives this representation.
0 0 0
RepB,B (d/dx) = 1 0 0
0 1 0
Not all nilpotent matrices are all zeros except for blocks of subdiagonal ones.
from Example 2.4, and this four-vector basis
2.9 Example With the matrix N
1
0
1
0
0 2 1 0
D = h
1 , 1 , 1 , 0i
0
0
0
1
a change of basis operation produces this
1 0 1 0
0 0 0 0
1 0
0 2 1 0 1 0 0 0 0 2
1 1 1 0 0 1 0 0 1 1
0 0 0 1
0 0 1 0
0 0
1
1 0
1 0
1 0
1 0
= 3 2 5 0
2 1 3 0
1 0
2
1 2 0
0 1
371
The new matrix is nilpotent; its fourth power is the zero matrix since
P 1 P N
P 1 P N
P 1 P N
P 1 = P N
4 P 1
P 1 )4 = P N
(P N
4 is the zero matrix.
and N
The goal of this subsection is Theorem 2.13, which shows that the prior
example is prototypical in that every nilpotent matrix is similar to one that is
all zeros except for blocks of subdiagonal ones.
2.10 Definition Let t be a nilpotent transformation on V . A t-string generated by ~v V is a sequence h~v , t(~v ), . . . , tk1 (~v )i. This sequence has length k.
A t-string basis is a basis that is a concatenation of t-strings.
~3 i and h
~4 ,
~5 i, of length
~1 , ~2 ,
2.11 Example In Example 2.5, the t-strings h
three and two, can be concatenated to make a basis for the domain of t.
2.12 Lemma If a space has a t-string basis then the longest string in it has
length equal to the index of nilpotency of t.
Proof. Suppose not. Those strings cannot be longer; if the index is k then
tk sends any vector including those starting the string to ~0. So suppose
instead that there is a transformation t of index k on some space, such that
the space has a t-string basis where all of the strings are shorter than length
k. Because t has index k, there is a vector ~v such that tk1 (~v ) 6= ~0. Represent
~v as a linear combination of basis elements and apply tk1 . We are supposing
that tk1 sends each basis element to ~0 but that it does not send ~v to ~0. That
QED
is impossible.
We shall show that every nilpotent map has an associated string basis. Then
our goal theorem, that every nilpotent matrix is similar to one that is all zeros
except for blocks of subdiagonal ones, is immediate, as in Example 2.5.
Looking for a counterexample a nilpotent map without an associated
string basis that is disjoint will suggest the idea for the proof. Consider the
map t : C5 C5 with this action.
0 0 0 0 0
~e1 7
0 0 0 0 0
~e3 7 ~0
7
1 1 0 0 0
Rep
(t)
=
~e2
E5 ,E5
0 0 0 0 0
~
~e4 7 ~e5 7 0
0 0 0 1 0
Even after ommitting the zero vector, these three strings arent disjoint, but
that doesnt end hope of finding a t-string basis. It only means that E5 will not
do for the string basis.
To find a basis that will do, we first find the number and lengths of its
strings. Since ts index of nilpotency is two, Lemma 2.12 says that at least one
372
Chapter 5. Similarity
string in the basis has length two. Thus the map must act on a string basis in
one of these two ways.
~1
7
~
7
3
~5
7
~1
~
3
~4
~5
~2
7 ~0
~
4
7 ~0
~0
7
7
~2 7 ~0
~0
~0
~0
Now, the key point. A transformation with the left-hand action has a nullspace
of dimension three since thats how many basis vectors are sent to zero. A
transformation with the right-hand action has a nullspace of dimension four.
Using the matrix representation above, calculation of ts nullspace
x
x
N (t) = {
z x, z, r C}
0
r
shows that it is three-dimensional, meaning that we want the left-hand action.
~4 from R(t) N (t)
~2 and
To produce a string basis, first pick
0
0
~2 = 1
0
0
0
0
~4 = 0
0
1
~2 , ~4 } is linearly independent).
(other choices are possible, just be sure that {
~2 , ~4 }.
~5 pick a vector from N (t) that is not in the span of {
For
1
1
~
5 =
0
0
0
~3 such that t(
~1 ) =
~2 and t(
~3 ) =
~4 .
~1 and
Finally, take
0
1
~1 = 0
0
0
0
0
~3 = 0
1
0
373
~1 , . . . ,
~5 i, the matrix of t is as desired.
Now, with respect to B = h
0
1
RepB,B (t) =
0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
0
7 1 7 1 7 ~0
7 1 7 1 7 ~0
3 7 1 7
..
.
3 7 1 7 7 1 7 1 7 ~0
2 7 ~0
..
.
2 7 ~0
Proof. Fix a vector space V ; we will argue by induction on the index of nilpotency of t : V V . If that index is 1 then t is the zero map and any basis is
~n 7 ~0. For the inductive step, assume that the
~1 7 ~0, . . . ,
a string basis
theorem holds for any transformation with an index of nilpotency between 1
and k 1 and consider the index k case.
First observe that the restriction to the rangespace t : R(t) R(t) is also
nilpotent, of index k 1. Apply the inductive hypothesis to get a string basis
for R(t), where the number and length of the strings is determined by t.
~1 ), . . . , th1 (
~1 )i _ h
~2 , . . . , th2 (
~2 )i _ _ h
~i , . . . , thi (
~i )i
~1 , t(
B = h
(In the illustration these are the basis vectors of kind 1, so there are i strings
shown with this kind of basis vector.)
Second, note that taking the final nonzero vector in each string gives a basis
~1 ), . . . , thi (
~i )i for R(t) N (t). (These are illustrated with 1s in
C = hth1 (
squares.) For, a member of R(t) is mapped to zero if and only if it is a linear
combination of those basis vectors that are mapped to zero. Extend C to a
basis for all of N (t).
_
C = C h~1 , . . . , ~p i
374
Chapter 5. Similarity
~ are the vectors of kind 2 so that C is the set of squares.) While many
(The s
~ their number p is determined by the map t as it
choices are possible for the s,
is the dimension of N (t) minus the dimension of R(t) N (t).
_
Finally, B C is a basis for R(t)+N (t) because any sum of something in the
rangespace with something in the nullspace can be represented using elements
of B for the rangespace part and elements of C for the part from the nullspace.
Note that
2.14 Corollary Every nilpotent matrix is similar to a matrix that is all zeros
except for blocks of subdiagonal ones. That is, every nilpotent map is represented with respect to some basis by such a matrix.
This form is unique in the sense that if a nilpotent matrix is similar to two
such matrices then those two simply have their blocks ordered differently. Thus
this is a canonical form for the similarity classes of nilpotent matrices provided
that we order the blocks, say, from longest to shortest.
2.15 Example The matrix
M=
1 1
1 1
Mp
1
M=
1
0
M2 =
0
1
1
0
0
p
N
(M )
x
x C}
{
x
C2
The calculation also describes how a map m represented by M must act on any
string basis. With one map application the nullspace has dimension one and so
one vector of the basis is sent to zero. On a second application, the nullspace
has dimension two and so the other basis vector is sent to zero. Thus, the action
~2 7 ~0 and the canonical form of the matrix is this.
~1 7
of the map is
0 0
1 0
375
We can exhibit such a m-string basis and the change of basis matrices witnessing the matrix similarity. For the basis, take M to represent m with respect
~1 so that m(
~1 ) =
~2 .
~2 N (m) and also pick a
to the standard bases, pick a
~1 = 1
~2 = 1
1
0
(If we take M to be a representative with respect to some nonstandard bases
then this picking step is just more messy.) Recall the similarity diagram.
m
C2w.r.t. E2 C2w.r.t. E2
M
idyP
idyP
2
Cw.r.t.
C2w.r.t.
= RepB,E2 (id) =
1 1
0 1
P = (P
1 1
1
0
1
1
1 1
1 1
1 1
0 0
=
0 1
1 1
0 1
1 0
2.16 Example The matrix
0
1
0
1
0 0
0 0
1 1
1 0
0 1
0
0
0
0
1 1
0
0
1 1
Np
0 0 0
0
0
1 0 0
0
0
1 1 1 1 1
0 1 0
0
0
1 0 1 1
1
0 0 0 0 0
0 0 0 0 0
1 0 0 0 0
1 0 0 0 0
0 0 0 0 0
zero matrix
N(N p )
0
0
{
u v u, v C}
u
v
0
y
{
z y, z, u, v C}
u
v
C5
376
Chapter 5. Similarity
That table shows that any string basis must satisfy: the nullspace after one map
application has dimension two so two basis vectors are sent directly to zero,
the nullspace after the second application has dimension four so two additional
basis vectors are sent to zero by the second iteration, and the nullspace after
three applications is of dimension five so the final basis vector is sent to zero in
three hops.
~2 7
~3 7 ~0
~1 7
~
~
4 7 5 7 ~0
To produce such a basis, first pick two independent vectors from N (n)
0
0
0
0
~5 = 0
~3 = 1
1
1
0
1
~4 N (n2 ) such that n(
~2 ) =
~3 and n(
~4 ) =
~5
~2 ,
then add
0
0
1
1
~4 = 0
~2 = 0
0
1
0
0
~1 ) =
~2 .
~1 N (n3 ) = C5 ) such that n(
and finish by adding
1
0
~1 = 1
0
0
Exercises
X 2.17 What is the index of nilpotency of the left-shift operator, here acting on the
space of triples of reals?
(x, y, z) 7 (0, x, y)
X 2.18 For each string basis state the index of nilpotency and give the dimension of
the rangespace and nullspace of each iteration of the nilpotent map.
~1 7
~2 7 ~0
(a)
~
~4 7 ~0
3 7
~
~2 7
~3 7 ~0
(b) 1 7
~
4 7 ~0
~5 7 ~0
~6 7 ~0
~1 7
~2 7
~3 7 ~0
(c)
Also give the canonical form of the matrix.
2.19 Decide which of these matrices are nilpotent.
(a)
2
1
377
4
2
(b)
3
1
1
3
(c)
3
3
3
45 22 19
(e) 33 16 14
69 34 29
X 2.20 Find the canonical form ofthis matrix.
0 1 1 0
0 0 1 1
0 0 0 0
0 0 0 0
0 0 0 0
2
2
2
1
1
1
(d)
1
3
5
1
0
2
4
1
7
1
1
0
0
0
0
0
0
1
1
1
1/2 1/2
1
0
1
(a)
(b) 0 1 1
(c)
1/2 1/2
0 1 1
1
1
1
Put each in canonical form.
2.23 Describe the effect of left or right multiplication by a matrix that is in the
canonical form for nilpotent matrices.
2.24 Is nilpotence invariant under similarity? That is, must a matrix similar to a
nilpotent matrix also be nilpotent? If so, with the same index?
X 2.25 Show that the only eigenvalue of a nilpotent matrix is zero.
2.26 Is there a nilpotent transformation of index three on a two-dimensional space?
2.27 In the proof of Theorem 2.13, why isnt the proofs base case that the index
of nilpotency is zero?
X 2.28 Let t : V V be a linear transformation and suppose ~v V is such that
tk (~v ) = ~0 but tk1 (~v ) 6= ~0. Consider the t-string h~v , t(~v ), . . . , tk1 (~v )i.
(a) Prove that t is a transformation on the span of the set of vectors in the string,
that is, prove that t restricted to the span has a range that is a subset of the
span. We say that the span is a t-invariant subspace.
(b) Prove that the restriction is nilpotent.
(c) Prove that the t-string is linearly independent and so is a basis for its span.
(d) Represent the restriction map with respect to the t-string basis.
2.29 Finish the proof of Theorem 2.13.
2.30 Show that the terms nilpotent transformation and nilpotent matrix, as
given in Definition 2.6, fit with each other: a map is nilpotent if and only if it is
represented by a nilpotent matrix. (Is it that a transformation is nilpotent if an
only if there is a basis such that the maps representation with respect to that
basis is a nilpotent matrix, or that any representation is a nilpotent matrix?)
2.31 Let T be nilpotent of index four. How big can the rangespace of T 3 be?
2.32 Recall that similar matrices have the same eigenvalues. Show that the converse
does not hold.
2.33 Prove a nilpotent matrix is similar to one that is all zeros except for blocks of
super-diagonal ones.
378
Chapter 5. Similarity
X 2.34 Prove that if a transformation has the same rangespace as nullspace. then the
dimension of its domain is even.
2.35 Prove that if two nilpotent matrices commute then their product and sum are
also nilpotent.
2.36 Consider the transformation of Mnn given by tS (T ) = ST T S where S is
an nn matrix. Prove that if S is nilpotent then so is tS .
2.37 Show that if N is nilpotent then I N is invertible. Is that only if also?
5.IV
379
Jordan Form
This section uses material from three optional subsections: Direct Sum, Determinants Exist, and Other Formulas for the Determinant.
The chapter on linear maps shows that every h : V W can be represented
by a partial-identity matrix with respect to some bases B V and D W .
This chapter revisits this issue in the special case that the map is a linear
transformation t : V V . Of course, the general result still applies but with
the codomain and domain equal we naturally ask about having the two bases
also be equal. That is, we want a canonical form to represent transformations
as RepB,B (t).
After a brief review section, we began by noting that a block partial identity
form matrix is not always obtainable in this B, B case. We therefore considered
the natural generalization, diagonal matrices, and showed that if its eigenvalues
are distinct then a map or matrix can be diagonalized. But we also gave an
example of a matrix that cannot be diagonalized and in the section prior to this
one we developed that example. We showed that a linear map is nilpotent
if we take higher and higher powers of the map or matrix then we eventually
get the zero map or matrix if and only if there is a basis on which it acts via
disjoint strings. That led to a canonical form for nilpotent matrices.
Now, this section concludes the chapter. We will show that the two cases
weve studied are exhaustive in that for any linear transformation there is a
basis such that the matrix representation RepB,B (t) is the sum of a diagonal
matrix and a nilpotent matrix in its canonical form.
5.IV.1
Recall that the set of square matrices is a vector space under entry-by-entry
addition and scalar multiplication and that this space Mnn has dimension n2 .
2
Thus, for any nn matrix T the n2 +1-member set {I, T, T 2 , . . . , T n } is linearly
2
dependent and so there are scalars c0 , . . . , cn2 such that cn2 T n + + c1 T + c0 I
is the zero matrix.
1.1 Remark This observation is small but important. It says that every transformation exhibits a generalized nilpotency: the powers of a square matrix cannot climb forever without a repeat.
1.2 Example Rotation of plane vectors /6 radians counterclockwise is represented with respect to the standard basis by
3/2
1/2
T =
3/2
1/2
and verifying that 0T 4 + 0T 3 + 1T 2 2T 1I equals the zero matrix is easy.
380
Chapter 5. Similarity
has a smaller degree than either and still sends the map or matrix to zero. Thus m(x) m(x)
is
the zero polynomial and the two are equal. (The leading coefficient requirement
also prevents a minimal polynomial from being the zero polynomial.)
1.6 Example We can see that m(x) = x2 2x 1 is minimal for the matrix
of Example 1.2 by computing the powers of T up to the power n2 = 4.
1/2 3/2
0 1
1/2 3/2
2
3
4
T =
T =
T =
1 0
3/2
1/2
3/2 1/2
Next, put c4 T 4 + c3 T 3 + c2 T 2 + c1 T + c0 I equal to the zero matrix
+ (1/2)c2 + ( 3/2)c1 + c0 = 0
(1/2)c
4
=0
(3/2)c4 c3 (3/2)c2 (1/2)c1
=0
( 3/2)c4 + c3 + ( 3/2)c2 + (1/2)c1
+ (1/2)c2 + ( 3/2)c1 + c0 = 0
(1/2)c4
and use Gauss method.
c4
c2
c3 + 3c2 +
3c1 2c0 = 0
2c1 + 3c0 = 0
Setting c4 , c3 , and c2 to zero forces c1 and c0 to also come out as zero. To get
a leading one, the most we can do is to set c4 and c3 to zero. Thus the minimal
polynomial is quadratic.
381
= k (t2 (1 + 2 )t + 1 2 ) (~v )
The third equality holds because the scalar 2 comes out of the second term, as
QED
t is linear.
In particular, if a minimial polynomial m(x) for a transformation t factors
as m(x) = (x 1 )q1 (x ` )q` then m(t) = (t 1 )q1 (t ` )q` is
the zero map. Since m(t) sends every vector to zero, at least one of the maps
t i sends some nonzero vectors to zero. So, too, in the matrix case if m is
minimal for T then m(T ) = (T 1 I)q1 (T ` I)q` is the zero matrix and at
least one of the matrices T i I sends some nonzero vectors to zero. Rewording
both cases: at least some of the i are eigenvalues. (See Exercise 29.)
Recall how we have earlier found eigenvalues. We have looked for such that
T ~v = ~v by considering the equation ~0 = T ~v x~v = (T xI)~v and computing the
determinant of the matrix T xI. That determinant is a polynomial in x, the
characteristic polynomial, whose roots are the eigenvalues. The major result
of this subsection, the next result, is that there is a connection between this
characteristic polynomial and the minimal polynomial. This results expands
on the prior paragraphs insight that some roots of the minimal polynomial
are eigenvalues by asserting that every root of the minimal polynomial is an
eigenvalue and further that every eigenvalue is a root of the minimal polynomial
(this is because it says 1 qi and not just 0 qi ).
382
Chapter 5. Similarity
2x2 + 3x 1
3x2 + 4x + 1
2
x2 + 2
=
4x2 + x + 1
3
1 2
3
x +
4
4
0
1
x+
1
1
2
1
t1,2
...
t1,1 x
t2,1
t
x
2,2
C= .
..
..
.
tn,n x
Recall that the product of the adjoint of a matrix with the matrix itself is the
determinant of that matrix times the identity.
c(x) I = adj(C)C = adj(C)(T xI) = adj(C)T adj(C) x
()
The entries of adj(C) are polynomials, each of degree at most n 1 since the
minors of a matrix drop a row and column. Rewrite it, as suggested above, as
adj(C) = Cn1 xn1 + + C1 x + C0 where each Ci is a matrix of scalars. The
left and right ends of equation () above give this.
cn Ixn + cn1 Ixn1 + + c1 Ix + c0 I = (Cn1 T )xn1 + + (C1 T )x + C0 T
Cn1 xn Cn2 xn1 C0 x
383
Combining the prior two lemmas gives that the minimal polynomial divides
the characteristic polynomial. Thus, any root of the minimal polynomial is
also a root of the characteristic polynomial. That is, so far we have that if
m(x) = (x 1 )q1 . . . (x i )qi then c(x) must has the form (x 1 )p1 . . . (x
i )pi (x i+1 )pi+1 . . . (x ` )p` where each qj is less than or equal to pj . The
proof of the Cayley-Hamilton Theorem is finished by showing that in fact the
characteristic polynomial has no extra roots i+1 , etc.
1.11 Lemma Each linear factor of the characteristic polynomial of a square
matrix is also a linear factor of the minimal polynomial.
Proof. Let T be a square matrix with minimal polynomial m(x) and assume
that x is a factor of the characteristic polynomial of T , that is, assume that
is an eigenvalue of T . We must show that x is a factor of m, that is, that
m() = 0.
In general, where is associated with the eigenvector ~v , for any polynomial function f (x), application of the matrix f (T ) to ~v equals the result of
multiplying ~v by the scalar f (). (For instance, if T has eigenvalue associated with the eigenvector ~v and f (x) = x2 + 2x + 3 then (T 2 + 2T + 3) (~v ) =
T 2 (~v ) + 2T (~v ) + 3~v = 2 ~v + 2 ~v + 3 ~v = (2 + 2 + 3) ~v .) Now, as m(T ) is
QED
the zero matrix, ~0 = m(T )(~v ) = m() ~v and therefore m() = 0.
384
Chapter 5. Similarity
1.12 Example We can use the Cayley-Hamilton Theorem to help find the
minimal polynomial of this matrix.
2 0 0 1
1 2 0 2
T =
0 0 2 1
0 0 0 1
First, its characteristic polynomial c(x) = (x 1)(x 2)3 can be found with the
usual determinant. Now, the Cayley-Hamilton Theorem says that T s minimal
polynomial is either (x 1)(x 2) or (x 1)(x 2)2 or (x 1)(x 2)3 . We can
decide among the choices just by computing:
1 0 0 1
0 0 0 1
0 0 0 0
1 1 0 2 1 0 0 2 1 0 0 1
(T 1I)(T 2I) =
0 0 1 1 0 0 0 1 = 0 0 0 0
0 0 0 0
0 0 0 1
0 0 0 0
and
0
1
2
(T 1I)(T 2I) =
0
0
0
0
0
0
0
0
0
0
0
0
1
1
0 0
0
0
0
0
0
0
0 1
0
0
0 2
=
0 1 0
0 1
0
0
0
0
0
0
0
0
0
0
0
0
0
!
3 0 0
3 0 0
3 0 0
2 0 1
(a) 1 3 0
(b) 1 3 0
(c) 1 3 0
(d) 0 6 2
0 0 4
0 0 3
0 1 3
0 0 2
1
4
0
0
0
!
0
3
0
0
0
2 2 1
4 1 0
0
(e) 0 6 2
(f ) 0
3 9 4 2 1
0 0 2
1
5
4
1
4
1.15 Find the minimal polynomial of this matrix.
0
0
1
1
0
0
0
1
0
385
0
0
..
0
...
...
0
0
386
Chapter 5. Similarity
(c) Decide if these are similar.
1
2
3
3
4
1
1
1
1.32 (a) Show that a matrix is invertible if and only if the constant term in its
minimal polynomial is not 0.
(b) Show that if a square matrix T is not invertible then there is a nonzero
matrix S such that ST and T S both equal the zero matrix.
X 1.33 (a) Finish the proof of Lemma 1.7.
(b) Give an example to show that the result does not hold if t is not linear.
5.IV.2
This subsection moves from the canonical form for nilpotent matrices to the
one for all matrices.
We have shown that if a map is nilpotent then all of its eigenvalues are zero.
We can now prove the converse.
2.1 Lemma A linear transformation whose only eigenvalue is zero is nilpotent.
Proof. If a transformation t on an n-dimensional space has only the single
eigenvalue of zero then its characteristic polynomial is xn . The Cayley-Hamilton
Theorem says that a map satisfies its characteristic polynimial so tn is the zero
QED
map. Thus t is nilpotent.
We have a canonical form for nilpotent matrices, that is, for each matrix
whose single eigenvalue is zero: each such matrix is similar to one that is all
zeroes except for blocks of subdiagonal ones. (To make this representation
unique we can fix some arrangement of the blocks, say, from longest to shortest.)
We next extend this to all single-eigenvalue matrices.
Observe that if ts only eigenvalue is then t s only eigenvalue is 0
because t(~v ) = ~v if and only if (t ) (~v ) = 0 ~v . The natural way to extend
the results for nilpotent matrices is to represent t in the canonical form N ,
and try to use that to get a simple representation T for t. The next result says
that this try works.
2.2 Lemma If the matrices T I and N are similar then T and N + I are
also similar, via the same change of basis matrices.
Proof. With N = P (T I)P 1 = P T P 1 P (I)P 1 we have N =
2 1
T =
1 4
387
1 1
T 3I =
1
1
the only eigenvalue is 0, and T 3I is nilpotent. The null spaces are routine
to find; to ease this computation we take T to represent the transformation
t : C2 C2 with respect to the standard basis (we shall maintain this convention
for the rest of the chapter).
y
y C}
N ((t 3)2 ) = C2
N (t 3) = {
y
The dimensions of these null spaces show that the action of an associated map
~2 7 ~0. Thus, the canonical form for t 3 with
~1 7
t 3 on a string basis is
one choice for a string basis is
1
2
0 0
B=h
,
i
RepB,B (t 3) = N =
1 0
1
2
and by Lemma 2.2, T is similar to this matrix.
Rept (B, B) = N + 3I =
3 0
1 3
C2w.r.t. E2 C2w.r.t. E2
T 3I
idyP
idyP
2
Cw.r.t.
t3
C2w.r.t.
N
describes that to move from the lower left to the upper left we multiply by
1
1 2
1
= RepB,E2 (id) =
P = RepE2 ,B (id)
1 2
and to move from the upper right to the lower right we multiply by this matrix.
1
1/2 1/2
1 2
=
P =
1/4 1/4
1 2
So the similarity is expressed by
3 0
1/2
=
1 3
1/4
which is easily checked.
1/2
1/4
2
1
1
1 2
4
1 2
388
Chapter 5. Similarity
4 1 0 1
0 3 0 1
T =
0 0 4 0
1 0 0 5
and so has the single eigenvalue 4. The nullities of t 4 are: the null space of
t 4 has dimension two, the null space of (t 4)2 has dimension three, and the
null space of (t 4)3 has dimension four. Thus, t 4 has the action on a string
~2 7
~3 7 ~0 and
~4 7 ~0. This gives the canonical form N for
~1 7
basis of
t 4, which in turn gives the form for t.
4 0 0 0
1 4 0 0
N + 4I =
0 1 4 0
0 0 0 4
An array that is all zeroes, except for some number down the diagonal
and blocks of subdiagonal ones, is a Jordan block. We have shown that Jordan
block matrices are canonical representatives of the similarity classes of singleeigenvalue matrices.
2.5 Example The 3 3 matrices whose only eigenvalue is 1/2 separate into
three similarity classes. The three classes have these canonical representatives.
1/2
0
0
1/2
0
0
1/2
0
0
0
1
1
1/2
0
1/2
0
1/2
0
0
0
1/2
0
0
1/2
0
1
1/2
In particular, this matrix
1/2
0
0
0
1/2
1
0
0
1/2
belongs to the similarity class represented by the middle one, because we have
adopted the convention of ordering the blocks of subdiagonal ones from the
longest block to the shortest.
We will now finish the program of this chapter by extending this work to
cover maps and matrices with multiple eigenvalues. The best possibility for
general maps and matrices would be if we could break them into a part involving
their first eigenvalue 1 (which we represent using its Jordan block), a part with
2 , etc.
This ideal is in fact what happens. For any transformation t : V V , we
shall break the space V into the direct sum of a part on which t1 is nilpotent,
plus a part on which t 2 is nilpotent, etc. More precisely, we shall take three
389
steps to get to this sections major theorem and the third step shows that
V = N (t 1 ) N (t ` ) where 1 , . . . , ` are ts eigenvalues.
Suppose that t : V V is a linear transformation. Note that the restriction
of t to a subspace M need not be a linear transformation on M because there may
be an m
~ M with t(m)
~ 6 M . To ensure that the restriction of a transformation
to a part of a space is a transformation on the partwe need the next condition.
2.6 Definition Let t : V V be a transformation. A subspace M is t invariant if whenever m
~ M then t(m)
~ M (shorter: t(M ) M ).
Two examples are that the generalized null space N (t) and the generalized
range space R (t) of any transformation t are invariant. For the generalized null
space, if ~v N (t) then tn (~v ) = ~0 where n is the dimension of the underlying
space and so t(~v ) N (t) because tn ( t(~v ) ) is zero also. For the generalized
~ for some w
~ and then t(~v ) = tn+1 (w)
~ =
range space, if ~v R (t) then ~v = tn (w)
n
~ ) shows that t(~v ) is also a member of R (t).
t ( t(w)
Thus the spaces N (t i ) and R (t i ) are t i invariant. Observe
also that t i is nilpotent on N (t i ) because, simply, if ~v has the property
that some power of t i maps it to zero that is, if it is in the generalized
null space then some power of t i maps it to zero. The generalized null
space N (t i ) is a part of the space on which the action of t i is easy
to understand.
The next result is the first of our three steps. It establishes that t j leaves
t i s part unchanged.
2.7 Lemma A subspace is t invariant if and only if it is t invariant for any
scalar . In particular, where i is an eigenvalue of a linear transformation t,
then for any other eigenvalue j , the spaces N (t i ) and R (t i ) are
t j invariant.
Proof. For the first sentence we check the two implications of the if and only
if separately. One of them is easy: if the subspace is t invariant for any
then taking = 0 shows that it is t invariant. For the other implication suppose
that the subspace is t invariant, so that if m
~ M then t(m)
~ M , and let
be any scalar. The subspace M is closed under linear combinations and so if
t(m)
~ M then t(m)
~ m
~ M . Thus if m
~ M then (t ) (m)
~ M , as
required.
The second sentence follows straight from the first. Because the two spaces
are t i invariant, they are therefore t invariant. From this, applying the first
QED
sentence again, we conclude that they are also t j invariant.
The second step of the three that we will take to prove this sections major
result makes use of an additional property of N (t i ) and R (t i ), that
they are complementary. Recall that if a space is the direct sum of two others
V = N R then any vector ~v in the space breaks into two parts ~v = ~n + ~r
where ~n N and ~r R, and recall also that if BN and BR are bases for N
390
Chapter 5. Similarity
_
..
..
.
.
(t(~
))
Rep
(t(~
q ))
Rep
RepB,B (t) =
1
B
B
..
..
.
.
zeroes
of the
left of
QED
To see that t has been decomposed into its action on the parts, observe
that the restrictions of t to the subspaces N and R are represented, with
respect to the obvious bases, by the matrices T1 and T2 . So, with subspaces
that are invariant and complementary, we can split the problem of examining a
linear transformation into two lower-dimensional subproblems. The next result
illustrates this decomposition into blocks.
2.9 Lemma If T is a matrices with square submatrices T1 and T2
T1 Z2
T =
Z1 T2
where the Zs are blocks of zeroes, then |T | = |T1 | |T2 |.
Proof. Suppose that T is nn, that T1 is pp, and that T2 is q q. In the
permutation formula for the determinant
X
t1,(1) t2,(2) tn,(n) sgn()
|T | =
permutations
391
perms 2
of p+1,...,p+q
equals |T | =
tn,(n) sgn().
QED
2.10 Example
0
2
0
0
0
0
3
0
0 2 0 3 0
= 36
=
0 1 2 0 3
3
From Lemma 2.9 we conclude that if two subspaces are complementary and
t invariant then t is nonsingular if and only if its restrictions to both subspaces
are nonsingular.
Now for the promised third, final, step to the main result.
2.11 Lemma If a linear transformation t : V V has the characteristic polynomial (x 1 )p1 . . . (x ` )p` then (1) V = N (t 1 ) N (t ` )
and (2) dim(N (t i )) = pi .
Proof. Because dim(V ) is the degree p1 + + p` of the characteristic polynomial, to establish statement (1) we need only show that statement (2) holds
and that N (t i ) N (t j ) is trivial whenever i 6= j.
For the latter, by Lemma 2.7, both N (ti ) and N (tj ) are t invariant.
Notice that an intersection of t invariant subspaces is t invariant and so the
restriction of t to N (t i ) N (t j ) is a linear transformation. But both
t i and t j are nilpotent on this subspace and so if t has any eigenvalues
on the intersection then its only eigenvalue is both i and j . That cannot
be, so this restriction has no eigenvalues: N (t i ) N (t j ) is trivial
(Lemma 3.10 shows that the only transformation without any eigenvalues is on
the trivial space).
392
Chapter 5. Similarity
J1
zeroes
J2
..
J`1
zeroes
J`
where each J is the Jordan block associated with the eigenvalue of the original matrix (that is, is all zeroes except for s down the diagonal and some
subdiagonal ones).
Proof. Given an nn matrix T , consider the linear map t : Cn Cn that it
represents with respect to the standard bases. Use the prior lemma to write
Cn = N (t 1 ) N (t ` ) where 1 , . . . , ` are the eigenvalues of t.
Because each N (t i ) is t invariant, Lemma 2.8 and the prior lemma show
that t is represented by a matrix that is all zeroes except for square blocks along
the diagonal. To make those blocks into Jordan blocks, pick each Bi to be a
QED
string basis for the action of t i on N (t i ).
393
2.13 Example This matrix has the characteristic polynomial (x 2)2 (x 6).
2 0 1
T = 0 6 2
0 0 2
We will handle the eigenvalues 2 and 6 separately.
Computation of the powers, and the null spaces and nullities, of T 2I is
routine. (Recall from Example 2.3 the convention of taking T to represent a
transformation, here t : C3 C3 , with respect to the standard basis.)
power p
1
(T
0
0
0
0
0
0
2I)p
0 1
4 2
0 0
0 0
16 8
0
0
64
0
32
0
N ((t 2)p )
x
{ 0 x C}
0
{z/2 x, z C}
nullity
1
z
same
So the generalized null space N (t 2) has dimension two. Weve noted that
the restriction of t 2 is nilpotent on this subspace. From the way that the
~2 7 ~0.
~1 7
nullities grow we know that the action of t 2 on a string basis
Thus the restriction can be represented in the canonical form
1
2
0 0
= RepB,B (t 2)
B2 = h 1 , 0 i
N2 =
1 0
2
0
where many choices of basis are possible. Consequently, the action of the restriction of t to N (t 2) is represented by this matrix.
2 0
J2 = N2 + 2I = RepB2 ,B2 (t) =
1 2
The second eigenvalues computations are easier. Because the power of x 6
in the characteristic polynomial is one, the restriction of t 6 to N (t 6) must
~3 7 ~0 and since
be nilpotent of index one. Its action on a string basis must be
it is the zero map, its canonical form N6 is the 11 zero matrix. Consequently,
the canonical form J6 for the action of t on N (t6) is the 11 matrix with the
single entry 6. For the basis we can use any nonzero vector from the generalized
null space.
0
B6 = h1i
0
394
Chapter 5. Similarity
Taken together, these two give that the Jordan form of T is
2 0 0
RepB,B (t) = 1 2 0
0 0 6
2 2 1
T = 0 6 2
0 0 2
which has the same characteristic polynomial (x 2)2 (x 6).
While the characteristic polynomial is the same,
power p
1
(T 2I)p
0 2 1
0 4 2
0 0
0 8
0
0
16 8
0 0
N ((t 2)p )
{z/2 x, z C}
nullity
2
z
same
here the action of t2 is stable after only one application the restriction of of
t 2 to N (t 2) is nilpotent of index only one. (So the contrast with the prior
example is that while the characteristic polynomial tells us to look at the action
of the t 2 on its generalized null space, the characteristic polynomial does not
describe completely its action and we must do some computations to find, in
this example, that the minimal polynomial is (x 2)(x 6).) The restriction of
~2 7 ~0,
~1 7 ~0 and
t 2 to the generalized null space acts on a string basis as
and we get this Jordan block associated with the eigenvalue 2.
2 0
J2 =
0 2
For the other eigenvalue, the arguments for the second eigenvalue of the
prior example apply again. The restriction of t 6 to N (t 6) is nilpotent
of index one (it cant be of index less than one, and since x 6 is a factor of
the characteristic polynomial to the power one it cant be of index more than
one either). Thus t 6s canonical form N6 is the 11 zero matrix, and the
associated Jordan block J6 is the 11 matrix with entry 6.
Therefore, T is diagonalizable.
2 0 0
1
0
3
_
B = B2 B6 = h0 , 1 , 4i
RepB,B (t) = 0 2 0
0 0 6
0
2
0
(Checking that the third vector in B is in the nullspace of t 6 is routine.)
395
1 4
0
0
3
0
T =
0 4 1
3 9 4
1
5
4
0
0
0
2
1
0
0
1
4
(T 3I)p
3
1
16
16
0
64
64
N ((t 3)p )
nullity
(u + v)/2
4
0
0
0
0
0
0
0 (u + v)/2
4 4 0
0
{ (u + v)/2 u, v C} 2
u
9 4 1 1
v
5
4
1
1
z
16
0
0 0
z
0
0
0 0
3
{
16
16 0 0
z z, u, v C}
32
16 0 0
u
v
16 16 0 0
64
0
0 0
0
0
0 0
same
64 64 0 0
128 64 0 0
64
64 0 0
(T + 1I)p
0 4
0 4
0 4
3 9
8
8
5
16
16
16
40
24
0
0
0
0
0 0
4 3
4
0
0
1
0
0
0
0
16 8
16
1
5
8
24
N ((t + 1)p )
(u + v)
u, v C}
{
v
nullity
same
396
Chapter 5. Similarity
shows that the restriction of t + 1 to its generalized null space acts on a string
~5 7 ~0.
~4 7 ~0 and
basis via the two separate strings
Therefore T is similar to this Jordan form matrix.
1 0 0 0 0
0 1 0 0 0
0
0 3 0 0
0
0 1 3 0
0
0 0 0 3
We close with the statement that the subjects considered earlier in this
Chpater are indeed, in this sense, exhaustive.
2.16 Corollary Every square matrix is similar to the sum of a diagonal matrix
and a nilpotent matrix.
Exercises
2.17 Do the check for Example 2.3.
2.18 Each matrix is in Jordan form. State its characteristic polynomial and its
minimal polynomial.
2 0
0
3 0 0
3 0
1
0
0
(a)
(b)
(c) 1 2
(d) 1 3 0
1 3
0
1
0 0 1/2
0 1 3
!
3 0 0 0
4 0
0
0
5 0 0
0
1 3 0 0
1 4 0
(e)
(f )
(g) 0 2 0
0 0 3 0
0 0 4
0
0 0 3
0 0 1 3
0 0
1
4
5 0 0 0
5 0 0 0
0 2 0 0
0 2 0 0
(h)
(i)
0 0 2 0
0 1 2 0
0 0 0 3
0 0 0 3
X 2.19 Find the Jordan form from the given data.
(a) The matrix T is 55 with the single eigenvalue 3. The nullities of the powers
are: T 3I has nullity two, (T 3I)2 has nullity three, (T 3I)3 has nullity
four, and (T 3I)4 has nullity five.
(b) The matrix S is 55 with two eigenvalues. For the eigenvalue 2 the nullities
are: S 2I has nullity two, and (S 2I)2 has nullity four. For the eigenvalue
1 the nullities are: S + 1I has nullity one.
2.20 Find the change of basis matrices for each example.
(a) Example 2.13
(b) Example 2.14
(c) Example 2.15
X 2.21 Find
the Jordan
form and a Jordan basis for each matrix.
10 4
(a)
25 10
5
(b)
9
4
(c) 2
5
4
7
!
0 0
1 3
0 4
X
X
397
5
4
3
0
3
(d) 1
1
2
1
!
9
7
3
(e) 9 7 4
4
4
4
!
2
2
1
1
(f ) 1 1
1 2
2
7
1
2
2
1 4 1 1
(g)
2 1
5
1
1
1
2
8
2.22 Find all possible Jordan forms of a transformation with characteristic polynomial (x 1)2 (x + 2)2 .
2.23 Find all possible Jordan forms of a transformation with characteristic polynomial (x 1)3 (x + 2).
2.24 Find all possible Jordan forms of a transformation with characteristic polynomial (x 2)3 (x + 1) and minimal polynomial (x 2)2 (x + 1).
2.25 Find all possible Jordan forms of a transformation with characteristic polynomial (x 2)4 (x + 1) and minimal polynomial (x 2)2 (x + 1).
2.26 Diagonalize
these.
1 1
0 1
(a)
(b)
0 0
1 0
2.27 Find the Jordan matrix representing the differentiation operator on P3 .
2.28 Decide if these two are similar.
1 1
1
0
4 3
1
1
2.29 Find the Jordan form of this matrix.
0 1
1
0
398
Chapter 5. Similarity
(d) Show that the trace of a map is the sum of its eigenvalues (counting multiplicities).
(e) Show that the trace of a nilpotent map is zero. Does the converse hold?
2.36 To use Definition 2.6 to check whether a subspace is t invariant, we seemingly
have to check all of the infinitely many vectors in a (nontrivial) subspace to see if
they satisfy the condition. Prove that a subspace is t invariant if and only if its
~ is in the subspace.
subbasis has the property that for all of its elements, t()
X 2.37 Is t invariance preserved under intersection? Under union? Complementation?
Sums of subspaces?
2.38 Give a way to order the Jordan blocks if some of the eigenvalues are complex
numbers. That is, suggest a reasonable ordering for the complex numbers.
2.39 Let Pj (R) be the vector space over the reals of degree j polynomials. Show
that if j k then Pj (R) is an invariant subspace of Pk (R) under the differentiation
operator. In P7 (R), does any of P0 (R), . . . , P6 (R) have an invariant complement?
2.40 In Pn (R), the vector space (over the reals) of degree n polynomials,
and
399
T ~v =
..
.
c1 31 ~1
c2 32 ~2
+ + cn 3n ~n
T k~v = c1 k1 ~1 + c2 k2 ~2 + + cn kn ~n
If one of the eigenvaluse, say, 1 , has a larger absolute value than any of the
other eigenvalues then its term will dominate the above expression. Put another
way, dividing through by k1 gives this,
k
k
T k~v
= c1 ~1 + c2 2k ~2 + + cn nk ~n
k
1
1
1
and, because 1 is assumed to have the largest absolute value, as k gets larger
the fractions go to zero. Thus, the entire expression goes to c1 ~1 .
That is (as long as c1 is not zero), as k increases, the vectors T k~v will
tend toward the direction of the eigenvectors associated with the dominant
eigenvalue, and, consequently, the ratios of the lengths k T k~v k/k T k1~v k will
tend toward that dominant eigenvalue.
For example (sample computer code for this follows the exercises), because
the matrix
3 0
T =
8 1
is triangular, its eigenvalues are just the entries on the diagonal, 3 and 1.
Arbitrarily taking ~v to have the components 1 and 1 gives
~v
1
1
T ~v
3
7
2
T ~v
9
17
9
T ~v
19 683
39 367
10
T ~v
59 049
118 097
and the ratio between the lengths of the last two is 2.999 9.
400
Chapter 5. Similarity
Two implementation issues must be addressed. The first issue is that, instead
of finding the powers of T and applying them to ~v , we will compute ~v1 as T ~v and
then compute ~v2 as T ~v1 , etc. (i.e., we never separately calculate T 2 , T 3 , etc.).
These matrix-vector products can be done quickly even if T is large, provided
that it is sparse. The second issue is that, to avoid generating numbers that are
so large that they overflow our computers capability, we can normalize the ~vi s
at each step. For instance, we can divide each ~vi by its length (other possibilities
are to divide it by its largest component, or simply by its first component). We
thus implement this method by generating
w
~ 0 = ~v0 /k~v0 k
~0
~v1 = T w
w
~ 1 = ~v1 /k~v1 k
~2
~v2 = T w
..
.
w
~ k1 = ~vk1 /k~vk1 k
~k
~vk = T w
until we are satisfied. Then the vector ~vk is an approximation of an eigenvector,
~ k1 k =
and the approximation of the dominant eigenvalue is the ratio k~vk k/kw
k~vk k.
One way we could be satisfied is to iterate until our approximation of the
eigenvalue settles down. We could decide, for instance, to stop the iteration
process not after some fixed number of steps, but instead when k~vk k differs
from k~vk1 k by less than one percent, or when they agree up to the second
significant digit.
The rate of convergence is determined by the rate at which the powers of
k2 /1 k go to zero, where 2 is the eigenvalue of second largest norm. If that
ratio is much less than one then convergence is fast, but if it is only slightly
less than one then convergence can be quite slow. Consequently, the method of
powers is not the most commonly used way of finding eigenvalues (although it
is the simplest one, which is why it is here as the illustration of the possibility of
computing eigenvalues without solving the characteristic polynomial). Instead,
there are a variety of methods that generally work by first replacing the given
matrix T with another that is similar to it and so has the same eigenvalues, but
is in some reduced form such as tridiagonal form: the only nonzero entries are
on the diagonal, or just above or below it. Then special techniques can be used
to find the eigenvalues. Once the eigenvalues are known, the eigenvectors of T
can be easily computed. These other methods are outside of our scope. A good
reference is [Goult, et al.]
Exercises
1 Use ten iterations to estimate the largest eigenvalue of these matrices, starting
from the vector with components 1 and 2. Compare the answer with the one
obtained by solving the characteristic equation.
(a)
1
0
5
4
(b)
3
1
401
2
0
2 Redo the prior exercise by iterating until k~vk k k~vk1 k has absolute value less
than 0.01 At each step, normalize by dividing each vector by its length. How many
iterations are required? Are the answers significantly different?
3 Use ten iterations to estimate the largest eigenvalue of these matrices, starting
from the vector with components 1, 2, and 3. Compare the answer with the one
obtained by solving the characteristic equation.
!
4
0 1
1
2
2
2
2
2
(a) 2 1 0
(b)
2 0 1
3 6 6
4 Redo the prior exercise by iterating until k~vk k k~vk1 k has absolute value less
than 0.01. At each step, normalize by dividing each vector by its length. How
many iterations does it take? Are the answers significantly different?
5 What happens if c1 = 0? That is, what happens if the initial vector does not to
have any component in the direction of the relevant eigenvector?
6 How can the method of powers be adopted to find the smallest eigenvalue?
Computer Code
This is the code for the computer algebra system Octave that was used
to do the calculation above.
>T=[3, 0;
8, -1]
T=
3
0
8 -1
>v0=[1; 2]
v0=
1
1
>v1=T*v0
v1=
3
7
>v2=T*v1
v2=
9
17
>T9=T**9
T9=
19683 0
39368 -1
>T10=T**10
T10=
59049 0
118096 1
>v9=T9*v0
v9=
19683
402
Chapter 5. Similarity
39367
>v10=T10*v0
v10=
59049
118096
>norm(v10)/norm(v9)
ans=2.9999
(It has been lightly edited to remove blank lines, etc. Remark: we are ignoring the power of Octave here; there are built-in functions to automatically
apply quite sophisticated methods to find eigenvalues and eigenvectors. Instead, we are using just the system as a calculator.)
403
.90 .01
pn
pn+1
=
rn+1
rn
.10 .99
Now, stable level means that pn+1 = pn and rn+1 = rn , so that the matrix
equation ~vn+1 = T ~vn becomes ~v = T ~v . We are therefore looking for eigenvectors
for T that are associated with the eigenvalue 1. The equation (I T )~v = ~0 is
.10 .01
p
0
=
.10 .01
r
0
which gives the eigenspace: vectors with the restriction that p = .1r. Coupled
with additional information, that the total world population of this species is is
p + r = 110 000, we find that the stable state is p = 10, 000 and r = 100, 000.
If we start with a park population of ten thousand animals, so that the rest of
the world has one hundred thousand, then every year ten percent (a thousand
animals) of those inside will leave the park, and every year one percent (a
thousand) of those from the rest of the world will enter the park. It is stable,
self-sustaining.
Now imagine that we are trying to gradually build up the total world population of this species. We can try, for instance, to have the world population
grow at a rate of 1% per year. In this case, we can take a stable state for
the parks population to be that it also grows at 1% per year. The equation
~vn+1 = 1.01 ~vn = T ~vn leads to ((1.01 I) T )~v = ~0, which gives this system.
.11 .01
p
0
=
.10 .02
r
0
The matrix is nonsingular, and so the only solution is p = 0 and r = 0. Thus,
there is no (usable) initial population that we can establish at the park and
expect that it will grow at the same rate as the rest of the world.
404
Chapter 5. Similarity
Knowing that an annual world population growth rate of 1% forces an unstable park population, we can ask which growth rates there are that would
allow an initial population for the park that will be self-sustaining. We consider
~v = T ~v and solve for .
.9
.01
= ( .9)( .99) (.10)(.01) = 2 1.89 + .89
0 =
.10
.99
A shortcut to factoring that quadratic is our knowledge that = 1 is an eigenvalue of T , so the other eigenvalue is .89. Thus there are two ways to have a
stable park population (a population that grows at the same rate as the population of the rest of the world, despite the leaky park boundaries): have a world
population that is does not grow or shrink, and have a world population that
shrinks by 11% every year.
So this is one meaning of eigenvalues and eigenvectorsthey give a stable state for a system. If the eigenvalue is 1 then the system is static. If
the eigenvalue isnt 1 then the system is either growing or shrinking, but in a
dynamically-stable way.
Exercises
1 What initial population for the park discussed above should be set up in the case
where world populations are allowed to decline by 11% every year?
2 What will happen to the population of the park in the event of a growth in world
population of 1% per year? Will it lag the world growth, or lead it? Assume
that the inital park population is ten thousand, and the world population is one
hunderd thousand, and calculate over a ten year span.
3 The park discussed above is partially fenced so that now, every year, only 5% of
the animals from inside of the park leave (still, about 1% of the animals from the
outside find their way in). Under what conditions can the park maintain a stable
population now?
4 Suppose that a species of bird only lives in Canada, the United States, or in
Mexico. Every year, 4% of the Canadian birds travel to the US, and 1% of them
travel to Mexico. Every year, 6% of the US birds travel to Canada, and 4%
go to Mexico. From Mexico, every year 10% travel to the US, and 0% go to
Canada.
(a) Give the transition matrix.
(b) Is there a way for the three countries to have constant populations?
(c) Find all stable situations.
405
The is an example of a recurrence relation (it is called that because the values
of f are calculated by looking at other, prior, values of f ). From it, we can
easily answer Fibonaccis twelve-month question.
month
pairs
0
1
1
1
2
2
3
3
4
5
5
8
6
13
7
21
8
34
9
55
10
89
11
144
12
233
The sequence of numbers defined by the above equation (of which the first few
are listed) is the Fibonacci sequence. The material of this chapter can be used
to give a formula with which we can can calculate f (n + 1) without having to
first find f (n), f (n 1), etc.
For that, observe that the recurrence is a linear relationship and so we can
give a suitable matrix formulation of it.
1 1
f (n)
f (n + 1)
f (1)
1
=
where
=
1 0
f (n 1)
f (n)
f (0)
1
Then, where we write T for the matrix and ~vn for the vector with components
f (n+1) and f (n), we have that ~vn = T n~v0 . The advantage of this matrix formulation is that by diagonalizing T we get a fast way to compute its powers: where
T = P DP 1 we have T n = P Dn P 1 , and the n-th power of the diagonal
matrix D is the diagonal matrix whose entries that are the n-th powers of the
entries of D.
The characteristic
equation ofT is 2 1. The quadratic formula gives
1+ 5
5
1
1 5
1+ 5
1
0
1 1
5
2
2
5
2
2
=
1 5
1
1+ 5
1 0
1
1
0
2
5
2 5
406
Chapter 5. Similarity
n
f (1)
f (n + 1)
1 1
=
f (0)
f (n)
1 0
!
!
n
1 5
1+ 5
1
1 5
1+ 5
0
f (1)
5
2
5
2
2
2
=
n
1
1+ 5
1 5
f (0)
1
1
0
2
5
2 5
We can compute f (n) from the second component of that equation.
"
!n
!n #
1+ 5
1 5
1
f (n) =
2
2
5
Notice that f is dominated by its first term because (1 5)/2 is less than
one, so its powers go to zero. Although we have extended the elementary model
of population growth by adding a delay period before the onset of fertility, we
nonetheless still get an (asmyptotically) exponential function.
In general, a linear recurrence relation has the form
f (n + 1) = an f (n) + an1 f (n 1) + + ank f (n k)
(it is also called a difference equation). This recurrence relation is homogeneous
because there is no constant term; i.e, it can be put into the form 0 = f (n +
1) + an f (n) + an1 f (n 1) + + ank f (n k). This is said to be a relation
of order k. The relation, along with the initial conditions f (0), . . . , f (k)
completely determine a sequence. For instance, the Fibonacci relation is of
order 2 and it, along with the two initial conditions f (0) = 1 and f (1) = 1,
determines the Fibonacci sequence simply because we can compute any f (n) by
first computing f (2), f (3), etc. In this Topic, we shall see how linear algebra
can be used to solve linear recurrence relations.
First, we define the vector space in which we are working. Let V be the set
of functions f from the natural numbers N = {0, 1, 2, . . . } to the real numbers.
(Below we shall have functions with domain {1, 2, . . . }, that is, without 0, but
it is not an important distinction.)
Putting the initial conditions aside for a moment, for any recurrence, we can
consider the subset S of V of solutions. For example, without initial conditions,
in addition to the function f given above, the Fibonacci relation is also solved by
the function g whose first few values are g(0) = 1, g(1) = 1, g(2) = 3, g(3) = 4,
and g(4) = 7.
The subset S is a subspace of V . It is nonempty because the zero function
is a solution. It is closed under addition since if f1 and f2 are solutions, then
an+1 (f1 + f2 )(n + 1) + + ank (f1 + f2 )(n k)
= (an+1 f1 (n + 1) + + ank f1 (n k))
+ (an+1 f2 (n + 1) + + ank f2 (n k))
= 0.
407
f (0)
f (1)
f 7 .
..
f (k)
Exercise 3 shows that this map is linear. Because, as noted above, any solution
of the recurrence is uniquely determined by the k initial conditions, this map is
one-to-one and onto. Thus it is an isomorphism, and thus S has dimension k,
the order of the recurrence.
So (again, without any initial conditions), we can describe the set of solutions of any linear homogeneous recurrence relation of degree k by taking linear
combinations of only k linearly independent functions. It remains to produce
those functions.
For that, we express the recurrence f (n + 1) = an f (n) + + ank f (n k)
with a matrix equation.
an
1
..
.
0
an1
0
1
0
..
.
an2
0
0
1
...
...
ank+1
0
..
.
...
ank
f (n + 1)
f (n)
0
f (n 1)
f (n)
..
..
.
.
..
f (n k + 1)
f (n k)
.
0
In trying to find the characteristic function of the matrix, we can see the pattern
in the 22 case
an an1
= 2 an an1
1
and 33 case.
an
1
0
an1
an2
0 = 3 + an 2 + an1 + an2
408
Chapter 5. Similarity
an
..
.
an1
1
0
..
.
an2
0
...
...
ank+1
0
..
.
...
ank
0
..
.
c2 ((1 5)/2)n . Including the initial conditions for the cases n = 0 and n = 1
gives
c2 = 1
c1 +
5/2)c1 + (1 5/2)c2 = 1
409
After moving the small disk to the far needle, the mid-sized disk to the middle
needle, and then moving the small disk to the middle needle we have this.
Now we can move the big disk over. Then, to finish, we repeat the process of
moving the smaller disks, this time so that they end up on the third needle, on
top of the big disk.
So the thing to see is that to move the very largest disk, the bottom disk,
at a minimum we must: first move the smaller disks to the middle needle, then
move the big one, and then move all the smaller ones from the middle needle to
the ending needle. Those three steps give us this recurence.
T (n + 1) = T (n) + 1 + T (n) = 2T (n) + 1 where T (1) = 1
We can easily get the first few values of T .
n
T (n)
1
1
2
3
3
7
4
15
5
31
6
63
7
127
8
255
9
511
10
1023
410
Chapter 5. Similarity
411
Computer Code
This code allows the generation of the first few values of a function defined by a recurrence and initial conditions. It is in the Scheme dialect of
LISP (specifically, it was written for A. Jaffers free scheme interpreter SCM,
although it should run in any Scheme implementation).
First, the Tower of Hanoi code is a straightforward implementation of
the recurrence.
(define (tower-of-hanoi-moves n)
(if (= n 1)
1
(+ (* (tower-of-hanoi-moves (- n 1))
2)
1) ) )
(Note for readers unused to recursive code: to compute T (64), the computer
is told to compute 2 T (63) 1, which requires, of course, computing T (63).
The computer puts the times 2 and the plus 1 aside for a moment to
do that. It computes T (63) by using this same piece of code (thats what
recursive means), and to do that is told to compute 2 T (62) 1. This
keeps up (the next step is to try to do T (62) while the other arithmetic is
held in waiting), until, after 63 steps, the computer tries to compute T (1).
It then returns T (1) = 1, which now means that the computation of T (2)
can proceed, etc., up until the original computation of T (64) finishes.)
The next routine calculates a table of the first few values. (Some language
notes: () is the empty list, that is, the empty sequence, and cons pushes
something onto the start of a list. Note that, in the last line, the procedure
proc is called on argument n.)
(define (first-few-outputs proc n)
(first-few-outputs-helper proc n ()) )
;
(define (first-few-outputs-aux proc n lst)
(if (< n 1)
lst
(first-few-outputs-aux proc (- n 1) (cons (proc n) lst)) ) )
412
Chapter 5. Similarity
1152921504606846975 2305843009213693951 4611686018427387903
9223372036854775807 18446744073709551615)
This is a list of T (1) through T (64). (The 120 mSec came on a 50 mHz 486
running in an XTerm of XWindow under Linux. The session was edited to
put line breaks between numbers.)
Appendix
Introduction
Mathematics is made of arguments (reasoned discourse that is, not pottery
throwing). This section is a reference to the most used techniques. A reader
having trouble with, say, proof by contradiction, can turn here for an outline of
that method.
But this section gives only a sketch. For more, these are classics: Propositional Logic by Copi, Induction and Analogy in Mathematics by P
olya, and
Naive Set Theory by Halmos.
Propositions
The point at issue in an argument is the proposition. Mathematicians usually
write the point in full before the proof and label it either Theorem for major
points, Lemma for results chiefly used to prove others, or Corollary for points
that follow immediately from a prior result.
Propositions can be complex, with many subparts. The truth or falsity of
the entire proposition depends both on the truth value of the parts, and on the
words used to assemble the statement from its parts.
Not. For example, where P is a proposition, it is not the case that P is true
provided P is false. Thus n is not prime is true only when n is the product of
smaller integers.
We can picture the not operation with a Venn diagram:
'$
. . .. . .. . .. . .. . .. . .. . .. . .. . .. . .. . .. . .. . .. . .. . .. .
..............
. .............
. .. . .. . . . . . . . .
.. .. .. .. .. .. .. .. .. .. . .
. . . .. . .. . .. . ..
.
. .. .. .. .. .. .. .. .. .
. .. . .. . .. . .. .
.. . . . . . . . P
.........
. .. .. .. .. .. .. ..
. .. .. .. .. .. .. ..
.. . . . . . . .
.. . . . . . . .
. . . .. . .. . ..
. .. .. .. .. .. .. ..
.
. .. .. .. .. .. .. .. ..
.. .. .. .. .. .. .. .. ..
. . .. . . . . . .
.
.
.
.
.
.
. . . . .. . .. . .. .
. . . . .
.
.
.
. .. .. .. .. .. .. .. .. .. .. . .
.. . . . . . . . . . . . . . . . . . .. .. .. .. .. .. .. .. .. .. .. .. ..
................................
&%
Where the box encloses all natural numbers, and inside the circle are the primes,
the dots are numbers satisfying not P .
To prove a not P statement holds, show P is false.
A-1
A-2
And. Consider the statement form P and Q. For the statement to be true
both halves must hold: 7 is prime and so is 3 is true, while 7 is prime and 3
is not is false.
Here is the Venn diagram for P and Q.
'$
'$
P
...
. .. .. .. .. .
.. .. .. .. .. .. ..
.
. ......
. .. .. .. .. .. .. ..
.. .. .. .. .. .. ..
.....
. .. . .
&%
&%
To prove P and Q, prove each half holds.
Or. A P or Q is true when either half holds: 7 is prime or 4 is prime is
true, while 7 is not prime or 4 is prime is false. We take or inclusively so that
if both halves are true 7 is prime or 4 is not then the statement as a whole
is true. (In everyday speech sometimes or is meant in an exclusive way: Eat
your vegetables or no dessert does not intend both halves to hold.)
The Venn diagram for or includes all of both circles.
'$
. .
.
. '$
....... .......
...................
. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .
.....
. . . . . . . . . . . . . . . .Q
. .P
. .. . .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .
.. . .. . . . . . . . . . . . . . . . . . . . . .
. .. . .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .
.......................
. .. . . . . . . . . . . . . . . . . . . .
. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. . .
.....
.....
&%
&%
To prove P or Q show that at all times at least one half holds (perhaps
sometimes one half and sometimes the other, but at all times at least one).
Implication. A P implies Q statement (perhaps instead phrased if P then
Q or P = Q) is true unless P is true while Q is false. Thus 7 is prime
implies 4 is not is true (contrary to its use in casual speech, in mathematics if
P then Q does not connote that P precedes Q or causes Q) while 7 is prime
implies 4 is also prime is false.
More subtly, P = Q is always true when P is false: 4 is prime implies 7
is prime and 4 is prime implies 7 is not are both true statements, sometimes
said to be vacuously true. (We adopt this convention because we want if a
number is a perfect square then it is not prime to be always true, for instance
when the number is 5 or when the number is 6.)
The diagram
$
'
'$
P
&%
&
%
shows Q holds whenever P does (another phrasing is P is sufficient to give Q).
Notice again that if P does not hold, Q may or may not be in force.
There are two main ways to establish an implication. The first way is direct:
assume P is true and, using that assumption, prove Q. For instance, to show
A-3
if a number is divisible by 5 then twice that number is divisible by 10, assume
the number is 5n and deduce that 2(5n) = 10n. The second way is indirect:
prove the contrapositive statement: if Q is false then P is false (rephrased, Q
can only be false when P is also false). As an example, to show if a number is
prime then it is not a perfect square, argue that if it were a square p = n2 then
it could be factored p = n n where n < p and so wouldnt be prime (of course,
p = 0 or p = 1 dont give n < p, but they are non-primes by definition)
Two comments about implication.
First, a P = Q result can sometimes be improved by weakening P or
strengthening Q. Thus, if a number is divisible by p2 then its square is also
divisible by p2 could be upgraded either by relaxing its hypothesis: if a number
is divisible by p then its square is divisible by p2 , or by tightening its conclusion:
if a number is divisible by p2 then its square is divisible by p4 .
Second, after showing P implies Q, a good next step is to look into whether
there are cases where Q holds but P does not. The idea is to better understand
the relationship between P and Q, with an eye toward strengthening the proposition.
Equivalence.
An implication cannot be improved when not only does P
imply Q, but also Q implies P , ( some ways to say this are: P if and only if
Q, P iff Q, P and Q are logically equivalent, P is necessary and sufficient
to give Q, P Q). For example, a number is divisible by a prime if and
only if that number squared is divisible by the prime squared.
The picture here shows that P and Q hold in exactly the same cases.
'$
P
&%
Although in simple cases a chain P R S Q may
be practical, we often show P = Q and Q = P separately. For instance,
P Q can be shown with both of the contrapositives: not Q implies not
P and not P implies not Q.
Quantifiers
Compare these two statements about natural numbers: there is an x such
that x is divisible by x2 is true, while for all numbers x, that x is divisible
by x2 is false. Clearly the prefixes deserve some study. We call the delimiting
prefixes quantifiers.
For all.
.
The for all prefix is called the universal quantifier and symbolized
A-4
Venn diagrams arent very helpful with quantifiers, but in a sense the box we
draw to border the diagram represents the universal quantifier since it dilineates
the universe of possible members.
. . . . . . . . . . . . . . . .
.. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. ..
................................
. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. ..
.. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. ..
................................
.
.. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. ..
................................
. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. ..
.. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. ..
................................
.
.. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. ..
................................
To show a statement holds in all cases, we must show it holds in each case.
Thus, to prove every number divisible by p has its square divisible by p2 , take a
single number of the form pn and square it: (pn)2 = p2 n2 . Hence the statement
holds for each number divisible by p. (This is a typical element proof. This
kind of argument requires that no properties are assumed for that element other
than those in the hypothesisfor instance, this is a wrong argument: if n is
divisible by a prime, say 2, so that n = 2k then n2 = (2k)2 = 4k 2 and the square
of the number is divisible by the square of the prime. That is an argument
about the case p = 2, but it isnt a proof for general p.)
There exists. The other common quantifier is there exists, symbolized .
This quantifier is in some sense the opposite of for all. For instance, contrast
these two definitions of primality of an integer p: (1) for all n, if n is not 1 and
n is not p then n does not divide p, and (2) it is not the case that there exists
an n (with n 6= 1 and n 6= p) such that n divides p.
As noted above, Venn diagrams are not much help with quantifiers, but a
picture of there is a number such that P would show both that there can be
more than one and that not all numbers need satisfy P .
. '$
P .
.
.
&%
An existence proposition can be proved by producing something satisfying
5
the property: one time, to settle the question of primality of 22 + 1, Euler
produced the divisor 641. But there are proofs that show that something exists
without saying how to find it; Euclids argument, given in the next subsection,
shows there are infinitely many primes without naming them. In general, while
demonstrating existence is better than nothing, giving an example is better than
that, and an exhaustive list of all instances is great. Still, mathematicians take
what they can get.
Finally, along with Are there any? we often ask How many? That
is why the issue of uniqueness often arises in conjunction with questions of
existence. Many times the two arguments are simpler if separated, so note that
just as proving something exists does not show it is unique, neither does proving
something is unique show it exists.
A-5
Techniques of Proof
Induction. Many proofs are iterative, Heres why its true for 1, it then
follows for 2, from there to 3, and so on . . . . These are called proofs by
induction. Such a proof has two steps. In the base step the proposition is
established for some first number, often 0 or 1. Then in the inductive step we
assume the proposition holds for numbers up to some k then it holds for the
next number.
Examples explain it best.
We prove that 1 + 2 + 3 + + n = n(n + 1)/2.
For the base step we must show the formula holds when n = 1. Thats
easy, the sum of the first 1 numbers equals 1(1 + 1)/2.
For the inductive step, assume the formula holds for 1, 2, . . . , k
1 = 1(1 + 1)/2
and
1 + 2 = 2(2 + 1)/2
k(k + 1)
(k + 1)(k + 2)
+ (k + 1) =
.
2
2
The idea of induction is simple. Weve shown the proposition holds for 1,
and weve shown that if it holds for 1 then it holds for 2, and if it holds for 1
and 2 then it holds for 3, etc. Thus it holds for any natural number greater
than or equal to 1.
We prove that every integer greater than 1 is a product of primes.
The base step is easy, 2 is the product of a single prime.
For the inductive step assume each of 2, 3, . . . , k is a product of primes,
aiming to show k + 1 is also. There are two cases: (1) if k + 1 is not
divisible by a number smaller than itself then it is a prime and so the
product of primes, and (2) if k + 1 factors then each factor can be written
as a product of primes by the inductive hypothesis and so k + 1 can be
rewritten as a product of primes. That ends the proof.
(Remark. The Prime Factorization Theorem of Number Theory says that
not only does a factorization exist, but that it is unique. Weve shown the
easy half.)
A-6
The other thing is that we sometimes use induction to go down, say, from
10 to 9 to 8, etc., down to 0. This is OKnext number could mean next
lowest number. Of course, at the end we have not shown the fact for all natural
numbers, only for those less than or equal to 10.
Contradiction. Another technique of proof is to show something is true by
showing it cant be false.
The classic example is Euclids, that there are infinitely many primes.
Suppose there are only finitely many primes p1 , . . . , pk . Consider p1
p2 . . . pk + 1. None of the primes on this supposedly exhaustive list divides
that number evenly, each leaves a remainder of 1. But every number is
a product of primes so this cant be. Thus there cannot be only finitely
many primes.
2n2 = m2 .
Factor out the 2s: n = 2kn n
and m = 2km m.
Rewrite:
2 (2kn n
)2 = (2km m)
2.
The Prime Factorization Theorem says there must be the same number
of factors of 2 on both sides, but there are an odd number (1 + 2kn ) on
the left and an even number (2km ) on the right. Thats a contradiction
so a rational with a square of 2 cannot be.
A-7
The Principle of Extensionality is that two sets with the same elements are
equal. Hence repeats collapse {7, 7} = {7} and order doesnt matter {2, } =
{, 2}.
We use for the subset relationship: {2, } {2, , 7} and for subset or
equality (if A is a subset of B but A 6= B then A is a proper subset of B). These
symbols may be flipped to signify the reverse relationship: {2, , 5} {2, 5}.
Because of Extensionality, to prove A = B just show they have the same
members. Usually we show mutual inclusion: both A B and A B.
Set operations. Venn diagrams are handy here. For instance, x P can be
pictured
'$
P
.x
&%
and P Q looks like this.
$
'
'$
P
&%
&
%
Note this is also the diagram for implication. Thats because P Q means
x P = x Q.
In general, for every propositional logic operator there is an associated set
operator. For instance, the complement of P is P comp = {x not(x P )}
'$
. . .. . .. . .. . .. . .. . .. . .. . .. . .. . .. . .. . .. . .. . .. . .. .
. .. . .. . .. . . . . . . . .
.. .. .. .. .. .. .. .. .. .. .. .. . .
..........
. .. . .. .. .. .. .. .. .. ..
.
.........
.. .. .. .. .. .. .. .. .. .
.. . . . . . . . .
.
.
.
.
P
........
. . . .
. .. .. .. .. .. .. ..
. .. .. .. .. .. .. ..
.. .. .. .. .. .. .. ..
.. .. .. .. .. .. .. ..
.. . . .
........
.
. .. .. .. .. .. .. .. ..
.. .. .. .. .. .. .. .. ..
.........
.
.
.
.
.
.
. . . . .
. .. . .. .. .. .. .. .. .. ..
. .. .. .. .. .. .. .. .. .. .. . .
...... . . .
.. .. .. .. .. .. .. .. .. .. .. .. .. .. .. . .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. ..
.
&%
&%
&%
...
. .. .. .. .. .
.......
.. .. .. .. .. .. .. .
.
. ......
. .. .. .. .. .. .. .
.......
.....
. .. . .
&%
&%
A-8
When two sets share no members their intersection is the empty set {},
symbolized . Any set has the empty set for a subset by the vacuously true
property of the definition of implication.
Sequences. We shall also use collections where order does matter and where
repeats do not collapse. These are sequences, denoted with angle brackets:
h2, 3, 7i =
6 h2, 7, 3i. A sequence of length 2 is sometimes called an ordered pair
and written with parentheses: (, 3). We also sometimes say ordered triple,
ordered 4-tuple, etc. The set of ordered n-tuples of elements of a set A is
denoted An . Thus the set of pairs of reals is R2 .
Functions. Functions are used to express relationships; for instance, Galelio
rolled balls down inclined planes to find the function relating elapsed time to
distance covered.
We first see functions in elementary Algebra, where they are presented as
formulas (e.g., f (x) = 16x2 100), but progressing to more advanced Mathematics reveals more general functionstrigonometric ones, exponential and
logarithmic ones, and even constructs like absolute value that involve piecing
together partsand we see that functions arent formulas, instead the key idea
is that a function associates with each input x a single output f (x).
Consequently, a function (or map) is defined to be a set of ordered pairs
(x, f (x) ) such that x suffices to determine f (x), that is x1 = x2 = f (x1 ) =
f (x2 ) (this requirement is referred to by saying a function is well-defined).
Each input x is one of the functions arguments and each output f (x) is a
value. The set of all arguments is f s domain and the set of output values is its
range. Often we dont need to produce exactly the range, and we instead work
with a superset of the range, the codomain. The notation for a function f with
domain X and codomain Y is f : X Y .
X'$ 'Y$
f
)
range of f
&% &%
f
We sometimes instead use the notation x 7 16x2 100, read x maps under
f to 16x2 100, or 16x2 100 is the image of x.
Complicated maps, like x 7 sin(1/x), can be thought of as combinations
of simple maps, for instance here applying the function g(y) = sin(y) to the
image of f (x) = 1/x. The composition of g : Y Z with f : X Y , denoted
g f : X Z, is the map sending x X to g( f (x) ) Z. This definition only
makes sense if the range of f is a subset of the domain of g.
Observe that the identity map id : Y Y, defined by id(y) = y, has the
property that for any f : X Y , the composition id f is equal to the map f .
More
A-9
So an identity map plays the same role with respect to function composition
that the number 0 plays in real number addition, or that the number 1 plays in
multiplication.
In line with that analogy, define a left inverse of a map f : X Y to be a
function g : range(f ) X such that g f is the identity map on X. Of course,
a right inverse of f is a h : Y X such that f h is the identity.
A map that is both a left and right inverse of f is called simply an inverse.
An inverse, if one exists, is unique because if both g1 and g2 are inverses of f
then g1 (x) = g1 (f g2 )(x) = (g1 f ) g2 (x) = g2 (x) (the middle equality
comes from the associativity of function composition), so we often call it the
inverse, written f 1 . For instance, the inverse of the function f : R R given
by f (x) = 2x 3 is the function f 1 : R R given by f 1 (x) = (x + 3)/2.
The superscript f 1 notation for function inverse can be confusing it
doesnt mean 1/f (x). It is used because it fits into a larger scheme. Functions that have the same codomain as domain can be iterated, so that where
f : X X, we can consider the composition of f with itself: f f , and f f f ,
etc. Naturally enough, we write f f as f 2 and f f f as f 3 , etc. Note
that the familiar exponent rules for real numbers obviously hold: f i f j = f i+j
and (f i )j = f ij . The relationship with the prior paragraph is that, where f is
invertible, writing f 1 for the inverse and f 2 for the inverse of f 2 , etc., gives
that these familiar exponent rules continue to hold, once f 0 is defined to be the
identity map.
If the codomain Y equals the range of f then we say that the function is onto.
A function has a right inverse if and only if it is onto (this is not hard to check).
If no two arguments share an image, if x1 6= x2 implies that f (x1 ) 6= f (x2 ),
then the function is one-to-one. A function has a left inverse if and only if it is
one-to-one (this is also not hard to check).
By the prior paragraph, a map has an inverse if and only if it is both onto
and one-to-one; such a function is a correspondence. It associates one and only
one element of the domain with each element of the range (for example, finite
sets must have the same number of elements to be matched up in this way).
Because a composition of one-to-one maps is one-to-one, and a composition of
onto maps is onto, a composition of correspondences is a correspondence.
We sometimes want to shrink the domain of a function. For instance, we
may take the function f : R R given by f (x) = x2 and, in order to have an
inverse, limit input arguments to nonnegative reals f: R+ R. Technically,
f is a different function than f ; we call it the restriction of f to the smaller
domain.
A final point on functions: neither x nor f (x) need be a number. As an
example, we can think of f (x, y) = x + y as a function that takes the ordered
pair (x, y) as its argument.
Relations. Some familiar operations are obviously functions: addition maps
(5, 3) to 8. But what of < or = ? We here take the approach of rephrasing
3 < 5 to (3, 5) is in the relation <. That is, define a binary relation on a set
A to be a set of ordered pairs of elements of A. For example, the < relation is
A-10
the set {(a, b) a < b}; some elements of that set are (3, 5), (3, 7), and (1, 100).
Another binary relation on the natural numbers is equality; this relation is
formally written as the set {. . . , (1, 1), (0, 0), (1, 1), . . . }.
Still another example is closer than 10, the set {(x, y) |x y| < 10}. Some
members of that relation are (1, 10), (10, 1), and (42, 44). Neither (11, 1) nor
(1, 11) is a member.
Those examples illustrate the generality of the definition. All kinds of relationships (e.g., both numbers even or first number is the second with the
digits reversed) are covered under the definition.
Equivalence Relations.
We shall need to say, formally, that two objects
are alike in some way. While these alike things arent identical, they are related
(e.g., two integers that give the same remainder when divided by 2).
A binary relation {(a, b), . . . } is an equivalence relation when it satisfies
(1) reflexivity: any object is related to itself;
(2) symmetry: if a is related to b then b is related to a;
(3) transitivity: if a is related to b and b is related to c then a is related to c.
(To see these conditions formalize being the same, read them again, replacing
is related to with is like.)
Some examples (on the integers): = is an equivalence relation, < does
not satisfy symmetry, same sign is a equivalence, while nearer than 10 fails
transitivity.
Partitions. In same sign {(1, 3), (5, 7), (1, 1), . . . } there are two kinds
of pairs, the first with both numbers positive and the second with both negative.
So integers fall into exactly one of two classes, positive or negative.
A partition of a set S is a collection of subsets {S1 , S2 , . . . } such that every
element of S is in one and only one Si : (1) S1 S2 . . . = S and (2) if i is not
equal to j then Si Sj = . Picture S being decomposed into distinct parts.
S1
S:
S2
S5
%
...
S3
$
!
S4
Thus, the first paragraph says same sign partitions the integers into the positives and the negatives.
.2
...
.4
.3
.1
.
1
.
0
.2
...
A-11
Similarly, the equivalence relation = partitions the integers into one-element
sets.
. . . .1 .0 .1 .2
...
%
?
...
$
?!
A-12
Usually when we pick representatives we have some natural scheme in mind. In
that case we call them the canonical class representatives.
For example, when considering the even and odd natural numbers,
.1
...
.1
.3
.3
.
2
.
...
.2
?1
.3
.3
.
2
...
?0
.2
. 35%. . .
9
.
15$
3
5
equivalent to
? 12
? 23
? 8
7
%
? 35
...
$
!
?2
1
9
.
15
Bibliography
[Abbot] Edwin Abbott, Flatland, Dover.
[Ackerson] R. H. Ackerson, A Note on Vector Spaces, American Mathematical
Monthly, volume 62 number 10 (Dec. 1955), p. 721.
[Agnew] Jeanne Agnew, Explorations in Number Theory, Brooks/Cole, 1972.
[Am. Math. Mon., Jun. 1931] C. A. Rupp (proposer), H. T. R. Aude (solver), problem
3468, American Mathematical Monthly, volume 37 number 6 (June-July 1931),
p. 355.
[Am. Math. Mon., Feb. 1933] V. F. Ivanoff (proposer), T. C. Esty (solver), problem
3529, American Mathematical Monthly, volume 39 number 2 (Feb. 1933),
p. 118.
[Am. Math. Mon., Jan. 1935] W. R. Ransom (proposer), Hansraj Gupta (solver), Elementary problem 105, American Mathematical Monthly, volume 42 number
1 (Jan. 1935), p. 47.
[Am. Math. Mon., Jan. 1949] C. W. Trigg (proposer), R. J. Walker (solver), Elementary problem 813, American Mathematical Monthly, volume 56 number 1 (Jan.
1949), p. 33.
[Am. Math. Mon., Jun. 1949] Don Walter (proposer), Alex Tytun (solver), Elementary problem 834, American Mathematical Monthly, volume 56 number 6
(June-July 1949), p. 409.
[Am. Math. Mon., Nov. 1951] Albert Wilansky, The Row-Sums of the Inverse Matrix,
American Mathematical Monthly, volume 58 number 9 (Nov. 1951), p. 614.
[Am. Math. Mon., Feb. 1953] Norman Anning (proposer), C. W. Trigg (solver), Elementary problem 1016, American Mathematical Monthly, volume 60 number
2 (Feb. 1953), p. 115.
[Am. Math. Mon., Apr. 1955] Vern Haggett (proposer), F. W. Saunders (solver), Elementary problem 1135, American Mathematical Monthly, volume 62 number
4 (Apr. 1955), p. 257.
[Am. Math. Mon., Jan. 1963] Underwood Dudley, Arnold Lebow (proposers), David
Rothman (solver), Elemantary problem 1151, American Mathematical
Monthly, volume 70 number 1 (Jan. 1963), p. 93.
[Am. Math. Mon., Dec. 1966] Hans Liebeck, A Proof of the Equality of Column Rank
and Row Rank of a Matrix American Mathematical Monthly, volume 73 number 10 (Dec. 1966), p. 1114.
[Anton] Howard Anton, Elementary Linear Algebra, John Wiley & Sons, 1987.
[Arrow] Kenneth J. Arrow, Social Choice and Individual Values, Wiley, 1963.
[Ball & Coxeter] W.W. Rouse Ball, Mathematical Recreations and Essays, revised by
H.S.M. Coxeter, MacMillan, 1962.
[Bennett] William J. Bennett, Quantifying Americas Decline, in Wall Street Journal,
15 March, 1993.
[Birkhoff & MacLane] Garrett Birkhoff, Saunders MacLane, Survey of Modern Algebra, third edition, Macmillan, 1965.
[Blass 1984] A. Blass, Existence of Bases Implies the Axiom of Choice, pp. 31-33, Axiomatic Set Theory, J. E. Baumgartner, ed., American Mathematical Society,
Providence RI, 1984.
[Bridgman] P.W. Bridgman, Dimensional Analysis, Yale University Press, 1931.
[Casey] John Casey, The Elements of Euclid, Books I to VI and XI, ninth edition,
Hodges, Figgis, and Co., Dublin, 1890.
[Clark & Coupe] David H. Clark, John D. Coupe, The Bangor Area Economy Its
Present and Future, report to the city of Bangor ME, Mar. 1967.
[Clarke] Arthur C. Clarke, Technical Error, Fantasy, December 1946, reprinted in
Great SF Stories 8 (1946), DAW Books, 1982.
[Con. Prob. 1955] The Contest Problem Book, 1955 number 38.
[Coxeter] H.S.M. Coxeter, Projective Geometry, second edition, Springer-Verlag,
1974.
[Courant & Robbins] Richard Courant, Herbert Robbins, What is Mathematics?, Oxford University Press, 1978.
[Cullen] Charles G. Cullen, Matrices and Linear Transformations, second edition,
Dover, 1990.
[Dalal, et. al.] Siddhartha R. Dalal, Edward B. Fowlkes, & Bruce Hoadley, Lesson
Learned from Challenger: A Statistical Perspective, Stats: the Magazine for
Students of Statistics, Fall 1989, p. 3.
[Davies] Thomas D. Davies, New Evidence Places Peary at the Pole, National Geographic Magazine, vol. 177 no. 1 (Jan. 1990), p. 44.
[de Mestre] Neville de Mestre, The Mathematics of Projectiles in Sport, Cambridge
University Press, 1990.
[De Parville] De Parville, La Nature, Paris, 1884, part I, p. 285-286 (citation from
[Ball & Coxeter]).
[Duncan] W.J. Duncan, Method of Dimensions, in Encyclopaedic Dictionary of
Physics Macmillan, 1962.
[Ebbing] Darrell D. Ebbing, General Chemistry, fourth edition, Houghton Mifflin,
1993.
[Ebbinghaus] H. D. Ebbinghaus, Numbers, Springer-Verlag, 1990.
[Einstein] A. Einstein, Annals of Physics, v. 35, 1911, p. 686.
[Eggar] M.H. Eggar, Pinhole Cameras, Perspecitve, and Projective Geometry, American Mathematicsl Monthly, August-September 1998, p. 618-630.
[Kemeny & Snell] John G. Kemeny, J. Laurie Snell, Finite Markov Chains, D. Van
Nostrand, 1960.
[Kemp] Franklin Kemp Linear Equations, American Mathematical Monthly, volume
89 number 8 (Oct. 1982), p. 608.
[Knuth] Donald E. Knuth, The Art of Computer Programming, Addison Wesley, 1988.
[Lang] Serge Lang, Linear Algebra, Addison-Wesley, 1966.
[Leontief 1951] Wassily W. Leontief, Input-Output Economics, Scientific American,
volume 185 number 4 (Oct. 1951), p. 15.
[Leontief 1965] Wassily W. Leontief, The Structure of the U.S. Economy, Scientific
American, volume 212 number 4 (Apr. 1965), p. 25.
[Lieberman] David Lieberman, The Big Bucks Ad Battles Over TVs Most Expensive
Minutes, TV Guide, Jan. 26 1991, p. 11.
[Macdonald & Ridge] Kenneth Macdonald, John Ridge, Social Mobility, in British
Social Trends Since 1900, A.H. Halsey, Macmillian, 1988.
[Macmillan Dictionary] William D. Halsey, Macmillan, 1979.
[Math. Mag., Sept. 1952] Dewey Duncan (proposer), W. H. Quelch (solver), Mathematics Magazine, volume 26 number 1 (Sept-Oct. 1952), p. 48.
[Math. Mag., Jan. 1957] M. S. Klamkin (proposer), Trickie T-27, Mathematics Magazine, volume 30 number 3 (Jan-Feb. 1957), p. 173.
[Math. Mag., Jan. 1963, Q237] D. L. Silverman (proposer), C. W. Trigg (solver),
Quickie 237, Mathematics Magazine, volume 36 number 1 (Jan. 1963).
[Math. Mag., Jan. 1963, Q307] C. W. Trigg (proposer). Quickie 307, Mathematics
Magazine, volume 36 number 1 (Jan. 1963), p. 77.
[Math. Mag., Nov. 1967] Clarence C. Morrison (proposer), Quickie, Mathematics
Magazine, volume 40 number 4 (Nov. 1967), p. 232.
[Math. Mag., Jan. 1973] Marvin Bittinger (proposer), Quickie 578, Mathematics
Magazine, volume 46 number 5 (Jan. 1973), p. 286, 296.
[Munkres] James R. Munkres, Elementary Linear Algebra, Addison-Wesley, 1964.
[Neimi & Riker] Richard G. Neimi, William H. Riker, The Choice of Voting Systems,
Scientific American, June 1976, p. 21-27.
[Nering] Evar D. Nering, Linear Algebra and Matrix Theory, second edition, John
Wiley, 1970.
[Niven & Zuckerman] I. Niven, H. Zuckerman, An Introduction to the Theory of Numbers, third edition, John Wiley, 1972.
[Oakley & Baker] Cletus O. Oakley, Justine C. Baker, Least Squares and the 3 : 40
Mile, Mathematics Teacher, Apr. 1977.
[Ohanian] Hans OHanian, Physics, volume one, W. W. Norton, 1985.
[Onan] Onan, Linear Algebra.
[Petersen] G. M. Petersen, Area of a Triangle, American Mathematical Monthly, volume 62 number 4 (Apr. 1955), p. 249.
[Polya] G. Polya, Patterns of Plausible Inference,
[Taylor] Alan D. Taylor, Mathematics and Politics: Strategy, Voting, Power, and
Proof, Springer-Verlag, 1995.
[Tilley] Burt Tilley, private communication, 1996.
[Trono] Tony Trono, compiler, University of Vermont Mathematics Department High
School Prize Examinations 1958-1991, mimeographed printing, 1991.
[USSR Olympiad no. 174] The USSR Mathematics Olympiad, number 174.
[Weston] J. D. Weston, Volume in Vector Spaces, American Mathematical Monthly,
volume 66 number 7 (Aug./Sept. 1959), p. 575-577.
[Weyl] Hermann Weyl, Symmetry, Princeton University Press, 1952.
[Wickens] Thomas D. Wickens, Models for Behavior, W.H. Freeman, 1982.
[Wilkinson 1965] Wilkinson, 1965
[Wohascum no. 2] The Wohascum County Problem Book problem number 2.
[Wohascum no. 47] The Wohascum County Problem Book problem number 47.
[Yaglom] I.M. Yaglom, Felix Klein and Sophus Lie: Evolution of the Idea of Symmetry
in the Nineteenth Century, translated by Sergei Sossinsky, Birkh
auser, 1988.
[Zwicker] William S. Zwicker, The Voters Paradox, Spin, and the Borda Count, Mathematical Social Sciences, vol. 22 (1991), p. 187-227.
Index
accuracy
of Gauss method, 6771
rounding error, 68
addition
vector, 80
additive inverse, 80
adjoint matrix, 328
angle, 42
antipodal, 340
antisymmetric matrix, 139
arrow diagram, 216, 232, 238, 242, 351
augmented matrix, 14
automorphism, 163
dilation, 164
reflection, 164
rotation, 164
back-substitution, 5
basis, 113124
change of, 238
definition, 113
orthogonal, 256
orthogonalization, 257
orthonormal, 258
standard, 114, 350
standard over the complex numbers, 350
string, 371
best fit line, 269
block matrix, 311
box, 321
orientation, 321
sense, 321
volume, 321
C language, 67
canonical form
for matrix equivalence, 245
for nilpotent matrices, 374
for row equivalence, 57
of a matrix, 358
of a transformation, 357
elementary
matrix, 226, 275
elementary reduction operations, 4
pivoting, 4
rescaling, 4
swapping, 4
elementary row operations, 4
entry, 13
equivalence
class, A-11
canonical representative, A-12
representative, A-11
equivalence relation, A-10, A-11
isomorphism, 169
matrix equivalence, 244
matrix similarity, 351
row equivalence, 50
equivalent statements, A-3
Erlanger Program, 286
Euclid, 286
even functions, 99, 138
even polynomials, 398
external direct sum, 168
Fibonacci sequence, 405
field, 141142
definition, 141
finite-dimensional vector space, 119
flat, 36
form, 55
free variable, 12
full column rank, 131
full row rank, 131
function, A-8
inverse image, 185
codomain, A-8
composition, 215, A-8
correspondence, A-9
domain, A-8
even, 99
identity, A-8
inverse, 231, A-9
left inverse, 231
multilinear, 305
odd, 99
one-to-one, A-9
onto, A-9
range, A-8
restriction, A-9
right inverse, 231
structure preserving, 161, 165
seehomomorphism, 176
two-sided inverse, 231
well-defined, A-8
zero, 177
Fundamental Theorem
of Linear Algebra, 268
Gauss method, 2
accuracy, 6771
back-substitution, 5
elementary operations, 4
Gauss-Jordan, 46
Gauss-Jordan, 46
generalized nullspace, 367
generalized rangespace, 367
Geometry of Linear Maps, 274279
Gram-Schmidt process, 255260
homogeneous coordinate vector, 340
homogeneous coordinates, 291
homogeneous equation, 21
homomorphism, 176
composition, 215
matrix representing, 194204
nonsingular, 190, 207
nullity, 188
nullspace, 188
rangespace, 184
rank, 206
zero, 177
ideal line, 342
ideal point, 342
identity
function, A-8
matrix, 224
ill-conditioned, 68
implication, A-2
improper subspace, 92
incidence matrix, 228
index
of nilpotency, 370
induced map, 275
induction, 23, A-5
inner product, 39
Input-Output Analysis, 6366
internal direct sum, 135
invariant
subspace, 377
invariant subspace
definition, 389
inverse, 231, A-9
additive, 80
exists, 231
left, 231, A-9
matrix, 329
right, 231, A-9
two-sided, A-9
inverse function, 231
inverse image, 185
inversion, 313
isometry, 286
isomorphism, 159175
characterized by dimension, 172
definition, 161
of a space with itself, 163
Jordan block, 388
Jordan form, 379398
represents similarity classes, 392
kernel, 188
Kirchhoffs Laws, 73
Klein, F., 286
Laplace expansion, 326330
computes determinant, 327
leading variable, 5
least squares, 269273
length, 39
Leontief, W., 63
line
best fit, 269
in projective plane, 341
line at infinity, 342
line of best fit, 269273
linear
transpose operation, 131
linear combination, 52
Linear Combination Lemma, 52
linear equation, 2
coefficients, 2
constant, 2
homogeneous, 21
inconsistent systems, 269
satisfied by a vector, 15
solution of, 2
Gauss method, 3
Gauss-Jordan, 46
solutions of
Cramers rule, 332
system of, 2
linear map
dilation, 275
reflection, 289
rotation, 274, 288
seehomomorphism, 176
skew, 276
trace, 397
linear recurrence, 406
linear recurrences, 405412
linear relationship, 103
linear surface, 36
linear transformation
seetransformation, 180
linearly dependent, 103
linearly independent, 103
LINPACK, 61
map
distance-preserving, 286
extended linearly, 173
induced, 275
self composition, 365
Maple, 61
Markov chains, 280285
Markov matrix, 284
Mathematica, 61
mathematical induction, 23
MATLAB, 61
matrix, 13
adjoint, 328
antisymmetric, 139
augmented, 14
block, 246, 311
change of basis, 238
characteristic polynomial, 360
cofactor, 327
column, 13
column space, 126
conditioning number, 70
determinant, 294, 299
diagonal, 209, 225
diagonalizable, 354
diagonalized, 244
elementary reduction, 226, 275
entry, 13
equivalent, 244
identity, 220, 224
incidence, 228
induced map, 275
inverse, 329
main diagonal, 224
Markov, 229, 284
matrix-vector product, 197
minimal polynomial, 220, 380
minor, 327
multiplication, 215
nilpotent, 370
nonsingular, 27, 207
orthogonal, 288
orthonormal, 286291
permutation, 225
rank, 206
representation, 196
row, 13
row equivalence, 50
row rank, 124
row space, 124
scalar multiple, 212
similar, 324
similarity, 351
singular, 27
skew-symmetric, 311
submatrix, 303
sum, 212
symmetric, 118, 139, 213, 220, 228,
268
trace, 213, 229, 397
transpose, 19, 126, 213
triangular, 204, 229, 330
unit, 222
Vandermonde, 311
matrix equivalence, 242249
canonical form, 245
definition, 244
matrix:form, 55
mean
arithmetic, 44
geometric, 44
method of powers, 399402
minimal polynomial, 220, 380
minor, 327
morphism, 161
multilinear, 305
multiplication
matrix-matrix, 215
matrix-vector, 197
mutual inclusion, A-7
natural representative, A-12
networks, 7278
Kirchhoffs Laws, 73
nilpotent, 368378
canonical form for, 374
definition, 370
matrix, 370
transformation, 370
nilpotentcy
index, 370
nonsingular, 207, 231
homomorphism, 190
matrix, 27
normalize, 258
nullity, 188
nullspace, 188
closure of, 367
generalized, 367
Octave, 61
odd functions, 99, 138
order
of a recurrence, 406
orientation, 321, 324
orthogonal, 42
basis, 256
complement, 263
mutually, 255
projection, 263
orthogonal matrix, 288
orthogonalization, 257
orthonormal basis, 258
orthonormal matrix, 286291
parallelepiped, 321
parallelogram rule, 35
parameter, 13
partial pivoting, 69
partition, A-10A-12
matrix equivalence classes, 244, 247
row equivalence classes, 50
partitions
into isomorphism classes, 170
permutation, 308
inversions, 313
matrix, 225
signum, 314
seesignum, 314
signum, 314
similar, 298, 324
canonical form, 392
similar matrices, 351
similarity, 351364
similarity transformation, 364
singular
matrix, 27
size, 319, 321
skew, 276
skew-symmetric, 311
span, 95
of a singleton, 99
spin, 149
square root, 398
stable populations, 403404
standard basis, 114
Statics problem, 5
string, 371
basis, 371
of basis vectors, 369
structure
preservation, 176
submatrix, 303
subspace, 91101
closed, 93
complementary, 136
definition, 91
direct sum, 135
improper, 92
independence, 135
invariant, 389
orthocomplement, 139
proper, 92
sum, 132
sum
of matrices, 212
of subspaces, 132
vector, 15, 34, 80
summation notation
for permutation expansion, 308
swapping rows, 4
symmetric matrix, 118, 139, 213, 220
system of linear equations, 2
Gauss method, 2
solving, 2
trace, 213, 229, 397
transformation
dual, 193
finite dimensional, 119
homomorphism, 176
isomorphism, 161
map, 176
over complex numbers, 347
subspace, 91
trivial, 84, 114
volume, 321
voting paradox, 147
majority cycle, 147
rational preference order, 147
voting paradoxes, 147151
spin, 149
well-defined, A-8
Wheatstone bridge, 75
zero
zero
zero
zero
zero
divisor, 220
divison, 237
divisor, 220
homomorphism, 177
vector, 22, 80