Mat67 Course Notes

Linear Algebra
As an Introduction to Abstract Mathematics

Lecture Notes for MAT67, Fall, 2007
University of California, Davis
Isaiah Lankham
Bruno Nachtergaele
Anne Schilling
Copyright c 2007 by the authors.

These lecture notes may be reproduced in their entirety for non-commercial purposes.
Contents
1 What is Linear Algebra? 1

1.1 Introduction to MAT 67 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 What is Linear Algebra? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Systems of linear equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3.1 Linear equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3.2 Non-linear equations . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.3.3 Linear transformations . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.3.4 Applications of linear equations . . . . . . . . . . . . . . . . . . . . . 7
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2 Introduction to Complex Numbers 11

2.1 Definition of complex numbers . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2 Operations on complex numbers . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.2.1 Addition and subtraction of complex numbers . . . . . . . . . . . . . 12
2.2.2 Multiplication and division of complex numbers . . . . . . . . . . . . 13
2.2.3 Complex conjugation . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.2.4 The modulus (a.k.a. norm, length, or magnitude) . . . . . . . . . . . 16
2.2.5 Complex numbers as vectors in R 2
. . . . . . . . . . . . . . . . . . . 18
2.3 Polar form and geometric interpretation for C . . . . . . . . . . . . . . . . . 19
2.3.1 Polar form for complex numbers . . . . . . . . . . . . . . . . . . . . . 19
2.3.2 Geometric multiplication for complex numbers . . . . . . . . . . . . . 20
2.3.3 Exponentiation and root extraction . . . . . . . . . . . . . . . . . . . 21
2.3.4 Some complex elementary functions . . . . . . . . . . . . . . . . . . . 22
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
ii
3 The Fundamental Theorem of Algebra and Factoring Polynomials 26
3.1 The fundamental theorem of algebra . . . . . . . . . . . . . . . . . . . . . . 26
3.2 Factoring polynomials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
4 Vector Spaces 36
4.1 Definition of vector spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
4.2 Elementary properties of vector spaces . . . . . . . . . . . . . . . . . . . . . 38
4.3 Subspaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
4.4 Sums and direct sums . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
5 Span and Bases 48

5.1 Linear span . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
5.2 Linear independence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
5.3 Bases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
5.4 Dimension . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
6 Linear Maps 62
6.1 Definition and elementary properties . . . . . . . . . . . . . . . . . . . . . . 62
6.2 Null spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
6.3 Ranges . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
6.4 Homomorphisms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
6.5 The dimension formula . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
6.6 The matrix of a linear map . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
6.7 Invertibility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
7 Eigenvalues and Eigenvectors 82

7.1 Invariant subspaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
7.2 Eigenvalues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
7.3 Diagonal matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
7.4 Existence of eigenvalues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
iii
7.5 Upper triangular matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
7.6 Diagonalization of 2 × 2 matrices and Applications . . . . . . . . . . . . . . 93
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
8 Permutations and the Determinant of a Square Matrix 99

8.1 Permutations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
8.1.1 Definition of permutations . . . . . . . . . . . . . . . . . . . . . . . . 99
8.1.2 Composition of permutations . . . . . . . . . . . . . . . . . . . . . . 103
8.1.3 Inversions and the sign of a permutation . . . . . . . . . . . . . . . . 105
8.2 Determinants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
8.2.1 Summations indexed by the set of all permutations . . . . . . . . . . 107
8.2.2 Properties of the determinant . . . . . . . . . . . . . . . . . . . . . . 109
8.2.3 Further properties and applications . . . . . . . . . . . . . . . . . . . 112
8.2.4 Computing determinants with cofactor expansions . . . . . . . . . . . 113
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
9 Inner Product Spaces 117

9.1 Inner product . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
9.2 Norms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
9.3 Orthogonality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
9.4 Orthonormal bases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
9.5 The Gram-Schmidt orthogonalization procedure . . . . . . . . . . . . . . . . 126
9.6 Orthogonal projections and minimization problems . . . . . . . . . . . . . . 128
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
10 Change of Bases 136

10.1 Coordinate vectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
10.2 Change of basis transformation . . . . . . . . . . . . . . . . . . . . . . . . . 138
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
11 The Spectral Theorem for Normal Linear Maps 144

11.1 Self-adjoint or hermitian operators . . . . . . . . . . . . . . . . . . . . . . . 144
11.2 Normal operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
11.3 Normal operators and the spectral decomposition . . . . . . . . . . . . . . . 148
iv
11.4 Applications of the spectral theorem: diagonalization . . . . . . . . . . . . . 150
11.5 Positive operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
11.6 Polar decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
11.7 Singular-value decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . 156
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
Supplementary Notes on Matrices and Linear Systems 161

12.1 From linear systems to matrix equations . . . . . . . . . . . . . . . . . . . . 161
12.1.1 Definition of and notation for matrices . . . . . . . . . . . . . . . . . 162
12.1.2 Using matrices to encode linear systems . . . . . . . . . . . . . . . . 165
12.2 Matrix arithmetic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
12.2.1 Addition and scalar multiplication . . . . . . . . . . . . . . . . . . . . 168
12.2.2 Multiplying and matrices . . . . . . . . . . . . . . . . . . . . . . . . . 172
12.2.3 Invertibility of square matrices . . . . . . . . . . . . . . . . . . . . . . 176
12.3 Solving linear systems by factoring the coefficient matrix . . . . . . . . . . . 178
12.3.1 Factorizing matrices using gaussian elimination . . . . . . . . . . . . 178
12.3.2 Solving homogenous linear systems . . . . . . . . . . . . . . . . . . . 188
12.3.3 Solving inhomogeneous linear systems . . . . . . . . . . . . . . . . . . 191
12.3.4 Solving linear systems with LU-factorization . . . . . . . . . . . . . . 195
12.4 Matrices and linear maps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 200
12.4.1 The canonical matrix of a linear map . . . . . . . . . . . . . . . . . . 200
12.4.2 Using linear maps to solve linear systems . . . . . . . . . . . . . . . . 201
12.5 Special operations on matrices . . . . . . . . . . . . . . . . . . . . . . . . . . 207
12.5.1 Transpose and conjugate transpose . . . . . . . . . . . . . . . . . . . 207
12.5.2 The trace of a square matrix . . . . . . . . . . . . . . . . . . . . . . . 208
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 210
List of Appendices
A The language of Sets and Functions 214

A.1 The language of sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215
A.1.1 Definition, notation, and examples . . . . . . . . . . . . . . . . . . . 215
A.1.2 Operations and relations on sets . . . . . . . . . . . . . . . . . . . . . 217
v
A.1.3 Cartesian products and (ordered) lists . . . . . . . . . . . . . . . . . 219
A.2 The language of functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221
A.2.1 Definition, notation, and examples . . . . . . . . . . . . . . . . . . . 221
A.2.2 Injectivity, surjectivity, and bijectivity . . . . . . . . . . . . . . . . . 223
A.2.3 Operations on functions . . . . . . . . . . . . . . . . . . . . . . . . . 225
B Summary of Algebraic Structures Encountered 227

B.1 Binary operations and scaling operations . . . . . . . . . . . . . . . . . . . . 227
B.2 Groups, fields, and vector spaces . . . . . . . . . . . . . . . . . . . . . . . . . 230
B.3 Rings and algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235
C Some Common Math Symbols & Abbreviations 237
D Summary of Notation Used 244
vi
Chapter 1
What is Linear Algebra?
1.1 Introduction to MAT 67

This class may well be one of your first mathematics classes that bridges the gap between
the mainly computation-oriented lower division classes and the abstract mathematics en-
countered in more advanced mathematics courses. The goal of this class is threefold:
1. You will learn Linear Algebra, which is one of the most wildly used mathematical
theories around. Linear Algebra finds applications in virtually every area of mathemat-
ics, including Multivariate Calculus, Differential Equations, and Probability Theory.
It is also widely applied in fields like in physics, chemistry, economics, psychology, and
engineering. You are even relying on methods from Linear Algebra every time you use
an Internet search like Google, the Global Positioning System (GPS), or a cellphone.
2. You will acquire computational skills to solve linear systems of equations, perform
operations on matrices, calculate eigenvalues, and find determinants of matrices.
3. In the setting of Linear Algebra, you will be introduced to abstraction. We will

develop the theory of Linear Algebra together, and you will learn to write proofs.
The lectures will mainly develop the theory of Linear Algebra, and the discussion sessions
will focus on the computational aspects. The lectures and the discussion sections go hand in
hand, and it is important that you attend both. The exercises for each Chapter are divided
into more computation-oriented exercises and exercises that focus on proof-writing.
1
2 CHAPTER 1. WHAT IS LINEAR ALGEBRA?
1.2 What is Linear Algebra?

Linear Algebra is the branch of mathematics aimed at solving systems of linear equations
with a finite number of unknowns. In particular, one would like to obtain answers to the
following questions:
• Characterization of solutions: Are there solutions to a given system of linear

equations? How many solutions are there?
• Finding solutions: How does the solution set look like? What are the solutions?
Linear Algebra is a systematic theory regarding the solutions of systems of linear equations.
Example 1.2.1. Let us take the following system of two linear equations in the two un-
knowns x1 and x2 : )
2x1 + x2 = 0
.
x1 − x2 = 1
This system has a unique solution for x1 , x2 ∈ R, namely x1 = 1

3
and x2 = − 32 .
This solution can be found in several different ways. One approach is to first solve for
one of the unknowns in one of the equations and then to substitute the result into the other
equation. Here, for example, we might solve to obtain
x1 = 1 + x2
from the second equation. Then, substituting this in place of x1 in the first equation, we
have
2(1 + x2 ) + x2 = 0.
From this, x2 = −2/3. Then, by further substitution,

2 1
x1 = 1 + − = .
3 3
Alternatively, we can a more systematic approach in eliminating variables. Here, for

example, we can subtracting 2 times the second equation from the first equation in order to
obtain 3x2 = −2. It is then immediate that x2 = − 32 and, by substitution this value for x2
in the first equation, that x1 = 13 .
1.2. WHAT IS LINEAR ALGEBRA? 3
Example 1.2.2. Take the following system of two linear equations in the two unknowns x1
and x2 : )
x1 + x2 = 1
.
2x1 + 2x2 = 1
Here, we can eliminate variables by adding −2 times the first equation to the second equation,
which results in 0 = −1. This is obviously a contradiction, and hence this system of equations
has no solution.
Example 1.2.3. Let us take the following system of one linear equation in the two unknowns
x1 and x2 :
x1 − 3x2 = 0.
In this case, there are infinitely many solutions given by the set {x2 = 13 x1 | x1 ∈ R}. You
can think of this solution set as a line in the Euclidean plane R2 :
x2
1 x2 = 31 x1
x1
−3 −2 −1 1 2 3
−1
In general a system of m linear equations in n unknowns x1 , x2 , . . . , xn is a collection

of equations of the form

a11 x1 + a12 x2 + · · · + a1n xn = b1 



a21 x1 + a22 x2 + · · · + a2n xn = b2 
.. .. , (1.1)
. . 




am1 x1 + am2 x2 + · · · + amn xn = bm
where the aij ’s are the coefficients (usually real or complex numbers) in front of the unknowns
xj , and the bi ’s are also fixed real or complex numbers. A solution is a set of numbers
s1 , s2 , . . . , sn such that, substituting x1 = s1 , x2 = s2 , . . . , xn = sn for the unknowns, all of
the equations in System (1.1) hold. Linear Algebra is a theory that concerns the solutions
and the structure of solutions for linear equations. As this course progresses, you will see
that there is a lot of subtlety in fully understanding the solution for such equations.
1.3 Systems of linear equations
1.3.1 Linear equations

Before going on, let us reformulate the notion of a system of linear equations into the
language of functions. This will also help us understand the adjective “linear” a bit better.
A function f is a map
f :X→Y (1.2)
from a set X to a set Y . The set X is called the domain of the function, and the set Y is
called the target space or codomain of the function. An equation is
f (x) = y, (1.3)
where x ∈ X and y ∈ Y . (If you are not familiar with the abstract notions of sets and
functions, then please consult Appendix A).
Example 1.3.1. Let f : R → R be the function f (x) = x3 − x. Then f (x) = x3 − x = 1 is

an equation. The domain and target space are both the set of real numbers R in this case.
In this setting, a system of equations is just another kind of equation.
Example 1.3.2. Let X = Y = R2 = R × R be the Cartesian product of the set of real

numbers. Then define the function f : R2 → R2 as
f (x1 , x2 ) = (2x1 + x2 , x1 − x2 ), (1.4)
and set y = (0, 1). Then the equation f (x) = y, where x = (x1 , x2 ) ∈ R2 , describes the
system of linear equations of Example 1.2.1.
The next question we need to answer is, “what is a linear equation?” Building on the
definition of an equation, a linear equation is any equation defined by a “linear” function
f that is defined on a “linear” space (a.k.a. a vector space as defined in Section 4.1). We
will elaborate on all of this in future lectures, but let us demonstrate the main features of a
“linear” space in terms of the example R2 . Take x = (x1 , x2 ), y = (y1 , y2) ∈ R2 . There are
1.3. SYSTEMS OF LINEAR EQUATIONS 5
two “linear” operations defined on R2 , namely addition and scalar multiplication:
x + y := (x1 + y1 , x2 + y2 ) (vector addition) (1.5)

cx := (cx1 , cx2 ) (scalar multiplication). (1.6)
A “linear” function on R2 is then a function f that interacts with these operations in the
following way:
f (cx) = cf (x) (1.7)

f (x + y) = f (x) + f (y). (1.8)
You should check for yourself that the function f in Example 1.3.2 has these two properties.
1.3.2 Non-linear equations

(Systems of) Linear equations are a very important class of (systems of) equations. You
will learn techniques in this class that can be used to solve any systems of linear equations.
Non-linear equations, on the other hand, are significantly harder to solve. An example is a
quadratic equation such as
x2 + x − 2 = 0, (1.9)
which, for no completely obvious reason, has exactly two solutions x = −2 and x = 1.
Contrast this with the equation
x2 + x + 2 = 0, (1.10)
which has no solutions within the set R of real numbers. Instead, it is has two complex
√ √
solutions 21 (−1 ± i 7) ∈ C where i = −1. (Complex numbers are discussed in more detail
in Chapter 2). In general, recall that the quadratic equation x2 + bx + c = 0 has the two
solutions r
b b2
x=− ± − c.
2 4
1.3.3 Linear transformations

The set R2 can be viewed as the Euclidean plane. In this context, linear functions of the
form f : R2 → R or f : R2 → R2 can be interpreted geometrically as “motions” in the plane
and are called linear transformations.
Example 1.3.3. Recall the following linear system from Example 1.2.1:
)
2x1 + x2 = 0
.
x1 − x2 = 1
Each equation can be interpreted as a straight line in the plane, with solutions (x1 , x2 ) to
the linear system given by the set of all points that simultaneously lie on both lines. In this
case, the two lines meet in only one location, which corresponds to the unique solution to
the linear system as illustrated in the following figure:
1 y =x−1
x
−1 1 2
−1
y = −2x
Example 1.3.4. The linear map f (x1 , x2 ) = (x1 , −x2 ) describes the “motion” of reflecting
a vector across the x-axis, as illustrated in the following figure:
1 (x1 , x2 )
x
1 2
−1 (x1 , −x2 )
Example 1.3.5. The linear map f (x1 , x2 ) = (−x2 , x1 ) describes the “motion” of rotating a
vector by 900 counterclockwise, as illustrated in the following figure:
(−x2 , x1 ) 2
1 (x1 , x2 )
x
−1 1 2
This example can easily be generalized to rotation by any arbitrary angle using Lemma 2.3.2.
In particular, when points in R2 are viewed as complex numbers, then we can employ the
so-called polar form for complex numbers in order to model the “motion” of rotation. (Cf.
Proof-Writing Exercise 5 on page 25.)
1.3.4 Applications of linear equations

Linear equations pop up in many different contexts. For example, you can view the derivative
df
dx
(x) of a differentiable function f : R → R as a linear approximation of f . This becomes
apparent when you look at the Taylor series of the function f (x) centered around the point
x = a (as seen in a course like MAT 21C):
df
f (x) = f (a) + (a)(x − a) + · · · . (1.11)
dx
In particular, we can graph the linear part of the Taylor series versus the original function,
as in the following figure:
f (x) f (x)
df
3 f (a) + dx
(a)(x − a)
x
1 2 3
df df
Since f (a) and dx (a) are merely real numbers, f (a) + dx (a)(x − a) is a linear function in the
single variable x.
Similar, if f : Rn → Rm is a multivariate function, then one can still view the derivative
of f as a form of linear approximation for f (as seen in a course like MAT 21D).
What if there are infinitely many variables x1 , x2 , . . .? In this case, the system of equations
has the form 
a11 x1 + a12 x2 + · · · = y1  
a21 x1 + a22 x2 + · · · = y2 .


···
Hence, the sums in each equation are infinite, and so we would have to deal with infinite
series. This, in particular, means that questions of convergence arises, where convergence
depends upon the infinite sequence x = (x1 , x2 , . . .) of variables. These questions will not
arise in this course since we are only interested in finite systems of linear equations in a finite
number of variables. Other subjects in which these questions do arise, though, include
• Differential Equations (as in a course like MAT 22B or MAT 118AB);
• Fourier Analysis (as in a course like MAT 129);
• Real and Complex Analysis (as in a course like MAT 125AB, MAT 185AB, MAT
201ABC, or MAT 202).
In courses like MAT 150ABC and MAT 250ABC, Linear Algebra is also seen to arise in the
study of such things as symmetries, linear transformations, and Lie Algebra theory.
Exercises for Chapter 1
Calculational Exercises
1. Solve the following systems of linear equations and characterize their solution set.
(I.e., determine whether there is a unique solution, no solution, etc.) Also, write each
system of linear equations as a single function f : Rn → Rm for appropriate choices of
m, n ∈ Z+ .
(a) System of 3 equations in the unknowns x, y, z, w:


x + 2y − 2z + 3w = 2 
2x + 4y − 3z + 4w = 5 .


5x + 10y − 8z + 11w = 12
(b) System of 4 equations in the unknowns x, y, z:


x + 2y − 3z = 4 


x + 3y + z = 11 
.
2x + 5y − 4z = 13 



2x + 6y + 2z = 22 
(c) System of 3 equations in the unknowns x, y, z:


x + 2y − 3z = −1 

3x − y + 2z = 7 .


5x + 3y − 4z = 2
2. Find all pairs of real numbers x1 and x2 that satisfy the system of equations
x1 + 3x2 = 2, (1.12)
x1 − x2 = 1. (1.13)
Proof-Writing Exercises
1. Let a, b, c, and d be real numbers, and consider the system of equations given by
ax1 + bx2 = 0, (1.14)

cx1 + dx2 = 0. (1.15)
Note that x1 = x2 = 0 is a solution for any choice of a, b, c, and d. Prove that if

ad − bc 6= 0, then x1 = x2 = 0 is the only solution.
Chapter 2
Introduction to Complex Numbers
Let R denote the set of real numbers, which should be a familiar collection of numbers to
anyone who has studied Calculus. In this chapter, we use R to build the equally important
set of so-called complex numbers.
2.1 Definition of complex numbers

We begin with the following definition.
Definition 2.1.1. The set of complex numbers C is defined as
C = {(x, y) | x, y ∈ R}.
Given a complex number z = (x, y), we call Re(z) = x the real part of z and Im(z) = y
the imaginary part of z.
In other words, we are defining a new collection of numbers z by taking every possible
ordered pair (x, y) of real numbers x, y ∈ R, and x is called the real part of the ordered pair
(x, y) in order to imply that the set R of real numbers should be identified with the subset
{(x, 0) | x ∈ R} ⊂ C. It is also common to use the term purely imaginary for any complex
number of the form (0, y), where y ∈ R. In particular, the complex number i = (0, 1) is
special, and it is called the imaginary unit. (The use of i is standard when denote this
complex number, though j is sometimes used if i means something else. E.g., i is used to
denote electric current in Electrical Engineering).
11
12 CHAPTER 2. INTRODUCTION TO COMPLEX NUMBERS
Note that if we write 1 = (1, 0), then we can express z = (x, y) ∈ C as
z = (x, y) = x(1, 0) + y(0, 1) = x1 + yi = x + yi.
It is often significantly easier to perform arithmetic operations on complex numbers when

written in this form, as we illustrate in the next section.
2.2 Operations on complex numbers

Even though we have formally defined C as the set of all ordered pairs of real numbers, we
can nonetheless extend the usual arithmetic operations on R so that they also make sense on
C. We discuss such extensions in this section, along with several other important operations
on complex numbers.
2.2.1 Addition and subtraction of complex numbers

Addition of complex numbers is performed component-wise, meaning that the real and imag-
inary parts are simply combined.
Definition 2.2.1. Given two complex numbers (x1 , y1 ), (x2 , y2 ) ∈ C, we define their (com-
plex) sum to be
(x1 , y1 ) + (x2 , y2 ) = (x1 + x2 , y1 + y2 ).
Example 2.2.2. (3, 2) + (17, −4.5) = (3 + 17, 2 − 4.5) = (20, −2.5).
As with the real numbers, subtraction is defined as addition with the so-called additive
inverse, where the additive inverse of z = (x, y) is defined as −z = (−x, −y).
√ √ √ √
Example 2.2.3. (π, 2) − (π/2, 19) = (π, 2) + (−π/2, − 19), where
√ √ √ √ √ √
(π, 2) + (−π/2, − 19) = (π − π/2, 2 − 19) = (π/2, 2 − 19).
The addition of complex numbers shares many of the same properties as the addition
of real numbers, including associativity, commutativity, the existence and uniqueness of an
additive identity, and the existence and uniqueness of additive inverses. We summarize these
properties in the following theorem, which you should prove for your own practice.
2.2. OPERATIONS ON COMPLEX NUMBERS 13
Theorem 2.2.4. Let z1 , z2 , z3 ∈ C be any three complex numbers. Then the following state-
ments are true.
1. (Associativity) (z1 + z2 ) + z3 = z1 + (z2 + z3 ).
2. (Commutativity) z1 + z2 = z2 + z1 .
3. (Additive Identity) There is a unique complex number, denoted 0, such that, given any
complex number z ∈ C, 0 + z = z. Moreover, 0 = (0, 0).
4. (Additive Inverses) Given any complex number z ∈ C, there is a unique complex num-
ber, denoted −z, such that z + (−z) = 0. Moreover, if z = (x, y) with x, y ∈ R, then
−z = (−x, −y).
The proof of this theorem is straightforward and relies solely on the definition of complex ad-
dition along with the familiar properties of addition for real numbers. For example, to check
commutativity, let z1 = (x1 , y1 ) and z2 = (x2 , y2 ) be complex numbers with x1 , x2 , y1, y2 ∈ R.
Then
z1 + z2 = (x1 + x2 , y1 + y2 ) = (x2 + x1 , y2 + y1 ) = z2 + z1 .
2.2.2 Multiplication and division of complex numbers

The definition of multiplication for two complex numbers is at first glance somewhat less
straightforward than that of addition.
Definition 2.2.5. Given two complex numbers (x1 , y1 ), (x2 , y2 ) ∈ C, we define their (com-
plex) product to be
(x1 , y1)(x2 , y2 ) = (x1 x2 − y1 y2 , x1 y2 + x2 y1 ).
According to this definition, i2 = −1. In other words, i is a solution of the polynomial

equation z 2 + 1 = 0, which does not have solutions in R. Solving such otherwise unsolvable
equations was largely the main motivation behind the introduction of complex numbers.
Note that the relation i2 = −1 and the assumption that complex numbers can be multiplied
like real numbers is sufficient to arrive at the general rule for multiplication of complex
numbers:
(x1 + y1 i)(x2 + y2 i) = x1 x2 + x1 y2 i + x2 y1 i + y1 y2 i2
= x1 x2 + x1 y2 i + x2 y1 i − y1 y2
= x1 x2 − y1 y2 + (x1 y2 + x2 y1 )i
As with addition, the basic properties of complex multiplication are easy enough to prove
using the definition. We summarize these properties in the following theorem, which you
should also prove for your own practice.
Theorem 2.2.6. Let z1 , z2 , z3 ∈ C be any three complex numbers. Then the following state-
ments are true.
1. (Associativity) (z1 z2 )z3 = z1 (z2 z3 ).
2. (Commutativity) z1 z2 = z2 z1 .
3. (Multiplicative Identity) There is a unique complex number, denoted 1, such that, given
any z ∈ C, 1z = z. Moreover, 1 = (1, 0).
4. (Distributivity of Multiplication over Addition) z1 (z2 + z3 ) = z1 z2 + z1 z3 .
Just as is the case for real numbers, any non-zero complex number z has a unique mul-
tiplicative inverse, which we may denote by either z −1 or 1/z.
Theorem 2.2.6 (continued).
5. (Multiplicative Inverses) Given z ∈ C with z 6= 0, there is a unique complex number,

denoted z −1 , such that zz −1 = 1. Moreover, if z = (x, y) with x, y ∈ R, then

−1 x −y
z = , 2 .
x + y x + y2
2 2
Proof. (Uniqueness.) A complex number w is an inverse of z if zw = 1 (by the commutativity

of complex multiplication this is equivalent to wz = 1). We will first prove that if w and v
are two complex numbers such that zw = 1 and zv = 1, then we necessarily have w = v.
This will then imply that any z ∈ C can have at most one inverse. To see this, we start
from zv = 1. Multiplying both sides by w, we obtain wzv = w1. Using the fact that 1 is
the multiplicative unit, that the product is commutative, and the assumption that w is an
inverse, we get zwv = v = w.
(Existence) Now assume z ∈ C with z 6= 0, and write z = x + yi for x, y ∈ R. Since
z 6= 0, at least one of x or y is not zero, and so x2 + y 2 > 0. Therefore, we can define

x −y
w= , 2 ,
x + y x + y2
2 2
and one can check that zw = 1.
Now, we can define the division of a complex number z1 by a non-zero complex number
z2 as the product of z1 and z2−1 . Explicitly, for two complex numbers z1 = x1 + iy1 and
z2 = x2 + iy2 , we have that their (complex) quotient is
z1 x1 x2 + y1 y2 + (x2 y1 − x1 y2 ) i
= .
z2 x22 + y22
Example 2.2.7. We illustrate the above definition with the following example:

(1, 2) 1·3+2·4 3·2−1·4 3+8 6−4 11 2
= , = , = , .
(3, 4) 3 2 + 42 3 2 + 42 9 + 16 9 + 16 25 25
2.2.3 Complex conjugation

Complex conjugation is an operation on C that will turn out to be very useful because it
allows us to manipulate only the imaginary part of a complex number. In particular, when
combined with the notion of modulus (as defined in the next section), it is one of the most
fundamental operations on C.
The definition and most basic properties of complex conjugation are as follows. (As in the
previous sections, you should provide a proof of the theorem below for your own practice.)
Definition 2.2.8. Given a complex number z = (x, y) ∈ C with x, y ∈ R, we define the

(complex) conjugate of z to be the complex number
z̄ = (x, −y).
Theorem 2.2.9. Given two complex numbers z1 , z2 ∈ C,
1. z1 + z2 = z1 + z2 .
2. z1 z2 = z1 z2 .
3. 1/z1 = 1/z1 , for all z1 6= 0.
4. z1 = z1 if and only if Im(z1 ) = 0.
5. z1 = z1 .
6. the real and imaginary parts of z1 can be expressed as
1 1
Re(z1 ) = (z1 + z1 ) and Im(z1 ) = (z1 − z1 ).
2 2i
2.2.4 The modulus (a.k.a. norm, length, or magnitude)

In this section, we introduce yet another operation on complex numbers, this time based
upon a generalization of the notion of absolute value of a real number. To motivate the
definition, it is useful to view the set of complex numbers as the two-dimensional Euclidean
plane, i.e., to think of C = R2 being equal as sets. The modulus, or length, of z ∈ C is
then defined as the Euclidean distance between z, as a point in the plane, and the origin
0 = (0, 0). This is the content of the following definition.
Definition 2.2.10. Given a complex number z = (x, y) ∈ C with x, y ∈ R, the modulus

of z is defined to be
p
|z| = x2 + y 2
In particular, given x ∈ R, note that

√
|(x, 0)| = x2 + 0 = |x|
under the convention that the square root function takes on its principal positive value.
Example 2.2.11. Using the above definition, we see that the modulus of the complex
number (3, 4) is
√ √ √
|(3, 4)| = 3 2 + 42 = 9 + 16 = 25 = 5,
To see this geometrically, construct a figure in the Euclidean plane, such as
5
(3, 4)
4 •
3
2
1
0 x
0 1 2 3 4 5
and apply the Pythagorean theorem to the resulting right triangle in order to find the
distance from the origin to the point (3, 4).
The following theorem lists the fundamental properties of the modulus, and especially as
it relates to complex conjugation. You should provide a proof for your own practice.
Theorem 2.2.12. Given two complex numbers z1 , z2 ∈ C,
1. |z1 z2 | = |z1 | · |z2 |.

z1 |z1 |
2. = , assuming that z2 6= 0.
z2 |z2 |
3. |z1 | = |z1 |.
4. |Re(z1 )| ≤ |z1 | and |Im(z1 )| ≤ |z1 |.
5. (Triangle Inequality) |z1 + z2 | ≤ |z1 | + |z2 |.
6. (Another Triangle Inequality) |z1 − z2 | ≥ | |z1 | − |z2 | |.
7. (Formula for Multiplicative Inverse) z1 z1 = |z1 |2 , from which
z1
z1−1 =
|z1 |2
when we assume that z1 6= 0.

2.2.5 Complex numbers as vectors in R2

When complex numbers are viewed as points in the Euclidean plane R2 , several of the
operations defined in Section 2.2 can be directly visualized as if they were operations on
vectors.
For the purposes of this chapter, we think of vectors as directed line segments that start
at the origin and end at a specified point in the Euclidean plane. These line segments may
also be moved around in space as long as the direction (which we will call the argument
in Section 2.3.1 below) and the length (a.k.a. the modulus) are preserved. As such, the
distinction between points in the plane and vectors is merely a matter of convention as long
as we at least implicitly think of each vector as having been translated so that it starts at
the origin.
As we saw in Example 2.2.11 above, the modulus of a complex number can be viewed
as the length of the hypotenuse of a certain right triangle. The sum and difference of two
vectors can also each be represented geometrically as the lengths of specific diagonals within
a particular parallelogram that is formed by copying and appropriately translating the two
vectors being combined.
Example 2.2.13. We illustrate the sum (3, 2) + (1, 3) = (4, 5) as the main, dashed diagonal
of the parallelogram in the left-most figure below. The difference (3, 2) − (1, 3) = (2, −1) can
also be viewed as the shorter diagonal of the same parallelogram, though we would properly
need to insist that this shorter diagonal be translated so that it starts at the origin. The
latter is illustrated in the right-most figure below.
y y
5 • (4, 5) 5 • (4, 5)
4 (1, 3) 4 (1, 3)
3 • 3 •
2 • (3, 2) 2 • (3, 2)
1 1
0 x 0 x
0 1 2 3 4 5 0 1 2 3 4 5
Vector Sum as Main Diagonal Vector Difference as Minor Diagonal

of Parallelogram of Parallelogram
2.3. POLAR FORM AND GEOMETRIC INTERPRETATION FOR C 19
2.3 Polar form and geometric interpretation for C

As mentioned above, C coincides with the plane R2 when viewed as a set of ordered pairs
of real numbers. Therefore, we can use polar coordinates as an alternate way to uniquely
identify a complex number. This gives rise to the so-called polar form for a complex
number, which often turns out to be a convenient representation for complex numbers.
2.3.1 Polar form for complex numbers

The following diagram summarizes the relations between cartesian and polar coordinates in
R2 :
y
z
•

 



r
y = r sin(θ)




θ 

x
| {z }
x = r cos(θ)
We call the ordered pair (x, y) the rectangular coordinates for the complex number z.
We also call the ordered pair (r, θ) the polar coordinates for the complex number z.
The radius r = |z| is called the modulus of z (as defined in Section 2.2.4 above), and the
angle θ = Arg(z) is called the argument of z. Since the argument of a complex number
describes an angle that is measured relative to the x-axis, it is important to note that θ is
only well-defined up to adding multiples of 2π. As such, we restrict θ ∈ [0, 2π) and add or
subtract multiples of 2π as needed (e.g., when multiplying two complex numbers so that their
arguments are added together) in order to keep the argument within this range of values.
It is straightforward to transform polar coordinates into rectangular coordinates using
the equations
x = r cos(θ) and y = r sin(θ).
In order to transform rectangular coordinates into polar coordinates, we first note that
p
r = x2 + y 2 is just the complex modulus. Then, θ must be chosen so that it satisfies the
bounds 0 ≤ θ < 2π in addition to the simultaneous equations
x = cos(θ) and y = sin(θ),
where we are assuming that z 6= 0.

Summarizing:
z = x + yi = r cos(θ) + r sin(θ)i = r(cos(θ) + sin(θ)i).
Part of the utility of this expression is that the size r = |z| of z is explicitly part of the very
definition since it is easy to check that | cos(θ) + sin(θ)i| = 1 for any choice of θ ∈ R.
Closely related is the exponential form for complex numbers, which does nothing more
than replace the expression cos(θ) + sin(θ)i with eiθ . The real power of this definition is
that this exponential notation turns out to be completely consistent with the usual usage of
exponential notation for real numbers.
Example 2.3.1. The complex number i in polar coordinate is expressed as eiπ/2 , whereas
the number −1 is given by eiπ .
2.3.2 Geometric multiplication for complex numbers

As discussed in Section 2.3.1 above, the general exponential form for a complex number z is
an expression of the form reiθ where r is a non-negative real number and θ ∈ [0, 2π). The
utility of this notation is immediately observed when multiplying two complex numbers:
Lemma 2.3.2. Let z1 = r1 eiθ1 , z2 = r2 eiθ2 ∈ C be complex numbers in exponential form.

Then
z1 z2 = r1 r2 ei(θ1 +θ2 ) .
Proof. By direct computation,
z1 z2 = (r1 eiθ1 )(r2 eiθ2 ) = r1 r2 eiθ1 eiθ2

= r1 r2 (cos θ1 + i sin θ1 )(cos θ2 + i sin θ2 )

= r1 r2 (cos θ1 cos θ2 − sin θ1 sin θ2 ) + i(sin θ1 cos θ2 + cos θ1 sin θ2 )

= r1 r2 cos(θ1 + θ2 ) + i sin(θ1 + θ2 ) = r1 r2 ei(θ1 +θ2 ) ,
where we have used the usual formulas for the sine and cosine of the sum of two angles.
In particular Lemma 2.3.2 shows that the modulus |z1 z2 | of the product is the product
of the moduli r1 and r2 and that the argument Arg(z1 z2 ) of the product is the sum of the
arguments θ1 + θ2 .
2.3.3 Exponentiation and root extraction

Another important use for the polar form of a complex number is in exponentiation. The
simplest possible situation here involves the use of a positive integer as a power, in which
case exponentiation is nothing more than repeated multiplication. Given the observations in
Section 2.3.2 above and using some trigonometric identities, one quickly obtains the following
fundamental result.
Theorem 2.3.3 (de Moivre’s Formula). Let z = r(cos(θ) + sin(θ)i) be a complex number in
polar form and n ∈ Z+ be a positive integer. Then
1. the exponentiation z n = r n (cos(nθ) + sin(nθ)i) and
2. the nth roots of z are given by the n complex numbers

1/n θ 2πk θ 2πk
zk = r cos + + sin + i
n n n n
where k = 0, 1, 2, . . . , n − 1.
Note, in particular, that we are not only always guaranteed the existence of an nth root for
any complex number, but that we are also always guaranteed to have exactly n of them.
This level of completeness in root extraction contrasts very sharply with the delicate care
that must be taken when one wishes to extract roots of real numbers without the aid of
complex numbers.
An important special case of de Moivre’s Formula yields an infinite family of well-studied
numbers called the roots of unity. By unity we just mean the complex number 1 = 1 + 0i,
and by the nth roots of unity we mean the n numbers

1/n 0 2πk 0 2πk
zk =1 cos + + sin + i
n n n n

2πk 2πk
= cos + sin i
n n
= e2πi(k/n) ,
where k = 0, 1, 2, . . . , n − 1. These numbers have many interesting properties and important

applications despite how simple they might appear to be.
Example 2.3.4. To find all solutions of the equation z 3 + 8 = 0 for z ∈ C, we may write
z = reiθ in polar form with r > 0 and θ ∈ [0, 2π). Then the equation z 3 + 8 = 0 becomes
z 3 = r 3 ei3θ = −8 = 8eiπ so that r = 2 and 3θ = π + 2πn. This means that there are three
distinct solutions when θ ∈ [0, 2π), namely θ = π3 , θ = π, and θ = 5π
3
.
2.3.4 Some complex elementary functions

We conclude these notes by defining three of the basic elementary functions that take com-
plex arguments. In this context, “elementary function” is used as a technical term and
essentially means something like “one of the most common forms of function encountered
when beginning to learn Calculus.” The most basic elementary functions include the fa-
miliar polynomial and algebraic functions, such as the nth root function, in addition to the
somewhat more sophisticated exponential function, the trigonometric functions, and the
logarithmic function. For the purposes of these notes, we will now define the complex ex-
ponential function and two complex trigonometric functions. Definitions for the remaining
basic elementary functions can be found in any book on Complex Analysis.
The basic groundwork for defining the complex exponential function was already
put into place in Sections 2.3.1 and 2.3.2 above. In particular, we have already defined the
expression eiθ to mean the sum cos(θ) + sin(θ)i for any real number θ. Historically, this
equivalence is a special case of the more general Euler’s formula
ex+yi = ex (cos(y) + sin(y)i),
which we here take as our definition of the complex exponential function applied to any
complex number x + yi for x, y ∈ R.

Given this exponential function, one can them define the complex sine function and
the complex cosine function as
eiz − e−iz eiz + e−iz

sin(z) = and cos(z) = .
2i 2
Remarkably, these functions retain many of their familiar properties, which should be taken
as a sign that the definitions — however abstract — have been well thought-out. We sum-
marize a few of these properties as follows.
Theorem 2.3.5. Given z1 , z2 ∈ C,
1. ez1 +z2 = ez1 ez2 and ez 6= 0 for any choice of z ∈ C.
2. sin2 (z1 ) + cos2 (z1 ) = 1.
3. sin(z1 + z2 ) = sin(z1 ) · cos(z2 ) + cos(z1 ) · sin(z2 ).
4. cos(z1 + z2 ) = cos(z1 ) · cos(z2 ) − sin(z1 ) · sin(z2 ).


1. Express the following complex numbers in the form x + yi for x, y ∈ R:
(a) (2 + 3i) + (4 + i)
(b) (2 + 3i)2 (4 + i)
2 + 3i
(c)
4+i
1 3
(d) +
i 1+i
(e) (−i)−1
√
(f) (−1 + i 3)3
2. Compute the real and imaginary parts of the following expressions, where z is the
complex number x + yi and x, y ∈ R:
1
(a)
z2
1
(b)
3z + 2
z+1
(c)
2z − 5
(d) z 3
√
3. Find r > 0 and θ ∈ [0, 2π) such that (1 − i)/ 2 = reiθ .
4. Solve the following equations for z a complex number:
(a) z 5 − 2 = 0
(b) z 4 + i = 0
(c) z 6 + 8 = 0
(d) z 3 − 4i = 0
5. Calculate the
(a) complex conjugate of the fraction (3 + 8i)4 /(1 + i)10 .

(b) complex conjugate of the fraction (8 − 2i)10 /(4 + 6i)5 .

(c) complex modulus of the fraction i(2 + 3i)(5 − 2i)/(−2 − i).
(d) complex modulus of the fraction (2 − 3i)2 /(8 + 6i)2 .
6. Compute the real and imaginary parts:
(a) e2+i
(b) sin(1 + i)
(c) e3−i
(d) cos(2 + 3i)
z
7. Compute the real and imaginary part of ee for z ∈ C.
1. Let a ∈ R and z, w ∈ C. Prove that
(a) Re(az) = aRe(z) and Im(az) = aIm(z).

(b) Re(z + w) = Re(z) + Re(w) and Im(z + w) = Im(z) + Im(w).
2. Let z ∈ C. Prove that Im(z) = 0 if and only if Re(z) = z.
3. Let z, w ∈ C. Prove the parallelogram law |z − w|2 + |z + w|2 = 2(|z|2 + |w|2).

z−w
4. Let z, w ∈ C with zw 6= 1 such that either |z| = 1 or |w| = 1. Prove that = 1.
1 − zw
5. For an angle θ ∈ [0, 2π), find the linear map fθ : R2 → R2 which describes the rotation
by the angle θ in the counterclockwise direction.
Hint: For a given angle θ find a, b, c, d ∈ R such that fθ (x1 , x2 ) = (ax1 + bx2 , cx1 + dx2 ).
Chapter 3
The Fundamental Theorem of Algebra

and Factoring Polynomials
The similarities and differences between R and C can be described as elegant and intrigu-
ing, but why are complex numbers important? One possible answer to this question is the
Fundamental Theorem of Algebra. It states that every polynomial equation in one
variable with complex coefficients has at least one complex solution. In other words, polyno-
mial equations formed over C can always be solved over C. This amazing result has several
equivalent formulations in addition to a myriad of different proofs, one of the first of which
was given by the eminent mathematician Carl Gauss in his doctoral thesis.
3.1 The fundamental theorem of algebra

The aim of this section is to provide a proof of the Fundamental Theorem of Algebra using
concepts that should be familiar to you from your study of Calculus, and so we begin by
providing an explicit formulation.
Theorem 3.1.1 (Fundamental Theorem of Algebra). Given any positive integer n ∈ Z+

and any choice of complex numbers a0 , a1 , . . . , an ∈ C with an 6= 0, the polynomial equation
an z n + · · · + a1 z + a0 = 0 (3.1)
has at least one solution z ∈ C.
26
3.1. THE FUNDAMENTAL THEOREM OF ALGEBRA 27
This is a remarkable statement. No analogous result holds for guaranteeing that a real so-
lution exists to Equation (3.1) if we restrict the coefficients a0 , a1 , . . . , an to be real numbers.
E.g., there does not exist a real number x satisfying an equation as simple as πx2 + e = 0.
Similarly, the consideration of polynomial equations having integer (resp. rational) coeffi-
cients quickly forces us to consider solutions that cannot possibly be integers (resp. rational
numbers). Thus, the complex numbers are special in this respect.
The statement of the Fundamental Theorem of Algebra can also be read as follows: Any
non-constant complex polynomial function defined on the complex plane C (when thought
of as R2 ) has at least one root, i.e., vanishes in at least one place. It is in this form that we
will provide a proof for Theorem 3.1.1.
Given how long the Fundamental Theorem of Algebra has been around, you should not
be surprised that there are many proofs of it. There have even been entire books devoted
solely to exploring the mathematics behind various distinct proofs. Different proofs arise
from attempting to understand the statement of the theorem from the viewpoint of different
branches of mathematics. This quickly leads to many non-trivial interactions with such fields
of mathematics as Real and Complex Analysis, Topology, and (Modern) Abstract Algebra.
The diversity of proof techniques available is yet another indication of how fundamental and
deep the Fundamental Theorem of Algebra really is.
To prove the Fundamental Theorem of Algebra using Differential Calculus, we will need
the Extreme Value Theorem for real-valued functions of two real variables, which we state
without proof. In particular, we formulate this theorem in the restricted case of functions
defined on the closed disk D of radius R > 0 and centered at the origin, i.e.,
D = {(x1 , x2 ) ∈ R2 | x21 + x22 ≤ R2 }.
Theorem 3.1.2 (Extreme Value Theorem). Let f : D → R be a continuous function on the

closed disk D ⊂ R2 . Then f is bounded and attains its minimum and maximum values on
D. In other words, there exist points xm , xM ∈ D such that
f (xm ) ≤ f (x) ≤ f (xM )
for every possible choice of point x ∈ D.
If we define a polynomial function f : C → C by setting f (z) = an z n + · · · + a1 z + a0 as in

28 CHAPTER 3. THE FUNDAMENTAL THEOREM OF ALGEBRA
Equation (3.1), then note that we can regard (x, y) 7→ |f (x + iy)| as a function R2 → R. By
a mild abuse of notation, we denote this function by |f ( · )| or |f |. As it is a composition of
continuous functions (polynomials and the square root), we see that |f | is also continuous.
Lemma 3.1.3. Let f : C → C be any polynomial function. Then there exists a point z0 ∈ C
where the function |f | attains its minimum value in R.
Proof. If f is a constant polynomial function, then the statement of the Lemma is trivially
true since |f | attains its minimum value at every point in C. So choose, e.g., z0 = 0.
If f is not constant, then the degree of the polynomial defining f is at least one. In this
case, we can denote f explicitly as in Equation (3.1). That is, we set
f (z) = an z n + · · · + a1 z + a0
with an 6= 0. Now, assume z 6= 0, and set A = max{|a0 |, . . . , |an−1 |}. We can obtain a lower
bound for |f (z)| as follows:
an−1 1 a0 1
|f (z)| = |an | |z|n 1 + +···+
an z an z n
∞
A X 1 A 1
≥ |an | |z|n 1 − k
= |an | |z|n 1 − .
|an | k=1 |z| |an | |z| − 1
For all z ∈ C such that |z| ≥ 2, we can further simplify this expression and obtain
2A
|f (z)| ≥ |an | |z|n 1 − .
|an ||z|
It follows from this inequality that there is an R > 0 such that |f (z)| > |f (0)|, for all z ∈ C
satisfying |z| > R. Let D ⊂ R2 be the disk of radius R centered at 0, and define a function
g : D → R, by
g(x, y) = |f (x + iy)|.
Since g is continuous, we can apply Theorem 3.1.2 in order to obtain a point (x0 , y0 ) ∈ D
such that g attains it minimum at (x0 , y0 ). By the choice of R we have that for z ∈ C \ D,
|f (z)| > |g(0, 0)| ≥ |g(x0, y0 )|. Therefore, |f | attains its minimum at z = x0 + iy0 .
We now prove the Fundamental Theorem of Algebra.

3.1. THE FUNDAMENTAL THEOREM OF ALGEBRA 29
Proof of Theorem 3.1.1. For our argument, we rely on the fact that the function |f | attains
its minimum value by Lemma 3.1.3. Let z0 ∈ C be a point where the minimum is attained.
We will show that if f (z0 ) 6= 0, then z0 is not a minimum, thus proving by contraposition
that the minimum value of |f (z)| is zero. Therefore, f (z0 ) = 0.
If f (z0 ) 6= 0, then we can define a new function g : C → C by setting
f (z + z0 )
g(z) = , for all z ∈ C.
f (z0 )
Note that g is a polynomial of degree n, and that the minimum of |f | is attained at z0 if and
only if the minimum of |g| is attained at z = 0. Moreover, it is clear that g(0) = 1.
More explicitly, g is given by a polynomial of the form
g(z) = bn z n + · · · + bk z k + 1,
with n ≥ 1 and bk 6= 0, for some 1 ≤ k ≤ n. Let bk = |bk |eiθ , and consider z of the form
z = r|bk |−1/k ei(π−θ)/k , (3.2)
with r > 0. For z of this form we have
g(z) = 1 − r k + r k+1h(r),
where h is a polynomial. Then, for r < 1, we have by the triangle inequality that
|g(z)| ≤ 1 − r k + r k+1|h(r)|.
For r > 0 sufficiently small we have r|h(r)| < 1, by the continuity of the function rh(r) and
the fact that it vanishes in r = 0. Hence
|g(z)| ≤ 1 − r k (1 − r|h(r)|) < 1,
for some z having the form in Equation (3.2) with r ∈ (0, r0 ) and r0 > 0 sufficiently small.
But then the minimum of the function |g| : C → R cannot possibly be equal to 1.
3.2 Factoring polynomials

In this section, we present several fundamental facts about polynomials, including an equiv-
alent form of the Fundamental Theorem of Algebra. While these facts should be familiar
to you, they nonetheless require careful formulation and proof. Before stating these results,
though, we first present a review of the main concepts needed in order to more carefully
work with polynomials.
Let n ∈ Z+ ∪ {0} be a non-negative integer, and let a0 , a1 , . . . , an ∈ C be complex
numbers. Then we call the expression
p(z) = an znn + · · · + a1 z + a0
a polynomial in the variable z with coefficients a0 , a1 , . . . , an . If an 6= 0, then we say

that p(z) has degree n (denoted deg(p(z)) = n), and we call an the leading term of p(z).
Moreover, if an = 1, then we call p(z) a monic polynomial. If, however, n = a0 = 0, then
we call p(z) = 0 the zero polynomial and set deg(0) = −∞.
Finally, by a root (a.k.a. zero) of a polynomial p(z), we mean a complex number z0 such
that, upon setting z = z0 , we obtain the zero polynomial p(z0 ) = 0. Note, in particular, that
every complex number is a root of the zero polynomial.
Convention dictates that
• a degree zero polynomial be called a constant polynomial,
• a degree one polynomial be called a linear polynomial,
• a degree two polynomial be called a quadratic polynomial,
• a degree three polynomial be called a cubic polynomial,
• a degree four polynomial be called a quadric polynomial,
• a degree five polynomial be called a quintic polynomial,
• and so on.
Addition and multiplication of polynomials is a direct generalization of the addition and

multiplication of real numbers, and degree interacts with these operations as follows:
3.2. FACTORING POLYNOMIALS 31
Lemma 3.2.1. Let p(z) and q(z) be non-zero polynomials. Then
1. deg (p(z) ± q(z)) ≤ max{deg(p(z)), deg(q(z))}
2. deg (p(z)q(z)) = deg(p(z)) + deg(q(z)).
Theorem 3.2.2. Given a positive integer n ∈ Z+ and any choice of a0 , a1 , . . . , an ∈ C with

an 6= 0, define the function f : C → C by setting
f (z) = an z n + · · · + a1 z + a0 , ∀ z ∈ C.
In other words, f is a polynomial function of degree n. Then
1. given any complex number w ∈ C, we have that f (w) = 0 if and only if there exists a
polynomial function g : C → C of degree n − 1 such that
f (z) = (z − w)g(z), ∀ z ∈ C.
2. there are at most n distinct complex numbers w for which f (w) = 0. In other words,
f has at most n distinct roots.
3. (Fundamental Theorem of Algebra, restated) there exist exactly n + 1 complex numbers

w0 , w1 , . . . , wn ∈ C (not necessarily distinct) such that
f (z) = w0 (z − w1 )(z − w2 ) · · · (z − wn ), ∀ z ∈ C.
In other words, every polynomial function with coefficients over C can be factored into
linear factors over C.
Proof.
1. Let w ∈ C be a complex number.
(“=⇒”) Suppose that f (w) = 0. Then, in particular, we have that
an w n + · · · + a1 w + a0 = 0.
Since this equation is equal to zero, it follows that, given any z ∈ C,
f (z) = an z n + · · · + a1 z + a0 − (an w n + · · · + a1 w + a0 )
= an (z n − w n ) + an−1 (z n−1 − w n−1) + · · · + a1 (z − w)
n−1
X n−2
X
k n−1−k
= an (z − w) z w + an−1 (z − w) z k w n−2−k + · · · + a1 (z − w)
k=0 k=0
n m−1
!
X X
= (z − w) am z k w m−k .
m=1 k=0
Thus, upon setting
n m−1
!
X X
g(z) = am z k w m−k , ∀ z ∈ C,
m=1 k=0
we have constructed a degree n − 1 polynomial function g such that
f (z) = (z − w)g(z), ∀ z ∈ C.
(“⇐=”) Suppose that there exists a polynomial function g : C → C of degree n − 1

such that
f (z) = (z − w)g(z), ∀ z ∈ C.
Then it follows that f (w) = (w − w)g(w) = 0, as desired.
2. We use induction on the degree n of f .

If n = 1, then f (z) = a1 z + a0 is a linear function, and the equation a1 z + a0 = 0 has
the unique solution z = −a0 /a1 . Thus, the result holds for n = 1.
Now, suppose that the result holds for n − 1. In other words, assume that every
polynomial function of degree n − 1 has at most n − 1 roots. Using the Fundamental
Theorem of Algebra (Theorem 3.1.1), we know that there exists a complex number
w ∈ C such that f (w) = 0. Moreover, from Part 1 above, we know that there exists a
polynomial function g of degree n − 1 such that
f (z) = (z − w)g(z), ∀ z ∈ C.
It then follows by the induction hypothesis that g has at most n − 1 distinct roots, and
so f must have at most n distinct roots.
3. This part follows from an induction argument on n that is virtually identical to that
of Part 2, and so the proof is left as an exercise to the reader.
1. Let n ∈ Z+ be a positive integer, let w0 , w1 , . . . , wn ∈ C be distinct complex numbers,
and let z0 , z1 , . . . , zn ∈ C be any complex numbers. Then one can prove that there is
a unique polynomial p(z) of degree at most n such that, for each k ∈ {0, 1, . . . , n},
p(wk ) = zk .
(a) Find the unique polynomial of degree at most 2 that satisfies p(0) = 0, p(1) = 1,
and p(2) = 2.
(b) Can your result in Part (a) be easily generalized to find the unique polynomial of
degree at most n satisfying p(0) = 0, p(1) = 1, . . . , p(n) = n?
2. Given any complex number α ∈ C, show that the coefficients of the polynomial
(z − α)(z − α)
are real numbers.
1. Let m, n ∈ Z+ be positive integers with m ≤ n. Prove that there is a degree n
polynomial p(z) with complex coefficients such that p(z) has exactly m distinct roots.
2. Given a polynomial p(z) = an znn + · · · + a1 z + a0 with complex coefficients, define the

conjugate of p(z) to be the new polynomial
p(z) = an znn + · · · + a1 z + a0 .
(a) Prove that p(z) = p(z).
(b) Prove that p(z) has real coefficients if and only if p(z) = p(z).
(c) Given polynomials p(z), q(z), and r(z) such that p(z) = q(z)r(z), prove that
p(z) = q(z)r(z).
3. Let p(z) be a polynomial with real coefficients, and let α ∈ C be a complex number.
Prove that p(α) = 0 if and only p(α) = 0.
Chapter 4
Vector Spaces
Now that important background has been developed, we are finally ready to begin the study
of Linear Algebra by introducing vector spaces. To get a sense of how important vector
spaces are, try flipping to a random page in these notes. There is very little chance that you
will flip to a page that does not have at least one vector space on it.
4.1 Definition of vector spaces

As we have seen in Chapter 1, a vector space is a set V with two operations defined upon
it: addition of vectors and multiplication by scalars. These operations must satisfy certain
properties, which we are about to discuss in more detail. The scalars are taken from a field
F, where for the remainder of these notes F stands either for the real numbers R or for the
complex numbers C. The sets R and C are examples of fields. The abstract definition of a
field along with further examples can be found in Appendix B.
Vector addition can be thought of as a function + : V × V → V that maps two vectors
u, v ∈ V to their sum u + v ∈ V . Scalar multiplication can similarly be described as a
function F × V → V that maps a scalar a ∈ F and a vector v ∈ V to a new vector av ∈ V .
(More information on these kinds of functions, also known as binary operations, can be found
in Appendix B.) It is when we place the right conditions on these operations that we turn
V into a vector space.
Definition 4.1.1. A vector space over F is a set V together with the operations of addition
V × V → V and scalar multiplication F × V → V satisfying each of the following properties.
36
4.1. DEFINITION OF VECTOR SPACES 37
1. Commutativity: u + v = v + u for all u, v ∈ V ;
2. Associativity: (u + v) + w = u + (v + w) and (ab)v = a(bv) for all u, v, w ∈ V and

a, b ∈ F;
3. Additive identity: There exists an element 0 ∈ V such that 0 + v = v for all v ∈ V ;
4. Additive inverse: For every v ∈ V , there exists an element w ∈ V such that v+w = 0;
5. Multiplicative identity: 1v = v for all v ∈ V ;
6. Distributivity: a(u + v) = au + av and (a + b)u = au + bu for all u, v ∈ V and

a, b ∈ F.
A vector space over R is usually called a real vector space, and a vector space over
C is similarly called a complex vector space. The elements v ∈ V of a vector space are
called vectors.
Even though Definition 4.1.1 may appear to be an extremely abstract definition, vector
spaces are fundamental objects in mathematics because there are countless examples of them.
You should expect to see many examples of vector spaces throughout your mathematical life.
Example 4.1.2. Consider the set Fn of all n-tuples with elements in F. This is a vector
space with addition and scalar multiplication defined componentwise. That is, for u =
(u1 , u2 , . . . , un ), v = (v1 , v2 , . . . , vn ) ∈ Fn and a ∈ F, we define
u + v = (u1 + v1 , u2 + v2 , . . . , un + vn ),
au = (au1 , au2 , . . . , aun ).
It is easy to check that each properties of Definition 4.1.1 is satisfied. In particular, the
additive identity 0 = (0, 0, . . . , 0), and the additive inverse of u is −u = (−u1 , −u2 , . . . , −un ).
An important cases of Example 4.1.2 is Rn , and especially when n = 2 or n = 3. We

have already seen in Chapter 1 that there is a geometric interpretation for elements of R2
and R3 as points in the Euclidean plane and Euclidean space, respectively.
Example 4.1.3. Let F∞ be the set of all sequences over F, i.e.,
F∞ = {(u1 , u2 , . . .) | uj ∈ F for j = 1, 2, . . .}.

38 CHAPTER 4. VECTOR SPACES
Addition and scalar multiplication are defined as expected, namely,
(u1 , u2, . . .) + (v1 , v2 , . . .) = (u1 + v1 , u2 + v2 , . . .),

a(u1 , u2 , . . .) = (au1 , au2 , . . .).
You should verify that F∞ becomes a vector space under these operations.
Example 4.1.4. Verify that V = {0} is a vector space! (Here, 0 denotes the zero vector in
any vector space.)
Example 4.1.5. Let F[z] be the set of all polynomial functions p : F → F with coefficients
in F. As discussed in Chapter 3, p(z) is a polynomial if there exist a0 , a1 , . . . , an ∈ F such
that
p(z) = an z n + an−1 z n−1 + · · · + a1 z + a0 . (4.1)
Addition and scalar multiplication on F[z] are defined pointwise as
(p + q)(z) = p(z) + q(z),

(ap)(z) = ap(z),
where p, q ∈ F[z] and a ∈ F. For example, if p(z) = 5z + 1 and q(z) = 2z 2 + z + 1, then

(p + q)(z) = 2z 2 + 6z + 2 and (2p)(z) = 10z + 2.
It can be easily verified that, under these operations, F[z] forms a vector space over F. The
additive identity in this case is the zero polynomial, for which all coefficients are equal to zero,
and the additive inverse of p(z) in Equation (4.1) is −p(z) = −an z n −an−1 z n−1 −· · ·−a1 z−a0 .
Example 4.1.6. Extending Example 4.1.5, let D ⊂ R be a subset of R, and let C(D) denote
the set of all continuous functions with domain D and codomain R. Then, under the same
operations of pointwise addition and scalar multiplication, one can show that C(D) also
forms a vector space.
4.2 Elementary properties of vector spaces

We are going to prove several important, yet simple, properties of vector spaces. From now
on, V will denote a vector space over F.
4.2. ELEMENTARY PROPERTIES OF VECTOR SPACES 39
Proposition 4.2.1. Every vector space has a unique additive identity.
Proof. Suppose there are two additive identities 0 and 0′ . Then
0′ = 0 + 0′ = 0,
where the first equality holds since 0 is an identity and the second equality holds since 0′ is
an identity. Hence 0 = 0′ , proving that the additive identity is unique.
Proposition 4.2.2. Every v ∈ V has a unique additive inverse.
Proof. Suppose w and w ′ are additive inverses of v so that v + w = 0 and v + w ′ = 0. Then
w = w + 0 = w + (v + w ′ ) = (w + v) + w ′ = 0 + w ′ = w ′ .
Hence w = w ′, as desired.
Since the additive inverse of v is unique, as we have just shown, it will from now on
be denoted by −v. We also define w − v to mean w + (−v). We will, in fact, show in
Proposition 4.2.5 below that −v = −1v.
Proposition 4.2.3. 0v = 0 for all v ∈ V .
Note that the 0 on the left-hand side in Proposition 4.2.3 is a scalar, whereas the 0 on
the right-hand side is a vector.
Proof. For v ∈ V , we have by distributivity that
0v = (0 + 0)v = 0v + 0v.
Adding the additive inverse of 0v to both sides, we obtain
0 = 0v − 0v = (0v + 0v) − 0v = 0v.
Proposition 4.2.4. a0 = 0 for every a ∈ F.

Proof. As in the proof of Proposition 4.2.3, if a ∈ F, then
a0 = a(0 + 0) = a0 + a0.
Adding the additive inverse of a0 to both sides, we obtain 0 = a0, as desired.
Proposition 4.2.5. (−1)v = −v for every v ∈ V .
Proof. For v ∈ V , we have
v + (−1)v = 1v + (−1)v = (1 + (−1))v = 0v = 0,
which shows that (−1)v is the additive inverse −v of v.
4.3 Subspaces
As mentioned in the last section, there are countless examples of vector spaces. One partic-
ularly important source of new vector spaces comes from looking at subsets of a set that is
already known to be a vector space.
Definition 4.3.1. Let V be a vector space over F, and let U ⊂ V be a subset of V . Then
we call U a subspace of V if U is a vector space over F under the same operations that
make V into a vector space over F.
To check that a subset U ⊂ V is a subspace, it suffices to check only a few of the

conditions of a vector space.
Lemma 4.3.2. Let U ⊂ V be a subset of a vector space V over F. Then U is a subspace of

V if and only if the following three conditions hold.
1. additive identity: 0 ∈ U;
2. closure under addition: u, v ∈ U implies u + v ∈ U;
3. closure under scalar multiplication: a ∈ F, u ∈ U implies that au ∈ U.

4.3. SUBSPACES 41
Proof. Condition 1 implies that the additive identity exists. Condition 2 implies that vector
addition is well-defined and, Condition 3 ensures that scalar multiplication is well-defined.
All other conditions for a vector space are inherited from V since addition and scalar mul-
tiplication for elements in U are the same when viewed as elements in either U or V .
Remark 4.3.3. Note that if we require U ⊂ V to be a nonempty subset of V , then condition

1 of Lemma 4.3.2 already follows from condition 3 since 0u = 0 for u ∈ U.
Example 4.3.4. In every vector space V , the subsets {0} and V are easily verified to be
subspaces. We call these the trivial subspaces of V .
Example 4.3.5. {(x1 , 0) | x1 ∈ R} is a subspace of R2 .
Example 4.3.6. U = {(x1 , x2 , x3 ) ∈ F3 | x1 + 2x2 = 0} is a subspace of F3 . To see this, we

need to check the three conditions of Lemma 4.3.2.
The zero vector (0, 0, 0) ∈ F3 is in U since it satisfies the condition x1 + 2x2 = 0. To
show that U is closed under addition, take two vectors v = (v1 , v2 , v3 ) and u = (u1 , u2, u3 ).
Then, by the definition of U, we have v1 + 2v2 = 0 and u1 + 2u2 = 0. Adding these two
equations, it is not hard to see that the vector v + u = (v1 + u1, v2 + u2 , v3 + u3 ) satisfies
(v1 + u1 ) + 2(v2 + u2 ) = 0. Hence v + u ∈ U. Similarly, to show closure under scalar
multiplication, take u = (u1 , u2, u3 ) ∈ U and a ∈ F. Then au = (au1 , au2 , au3 ) satisfies the
equation au1 + 2au2 = a(u1 + 2u2 ) = 0, and so au ∈ U.
Example 4.3.7. U = {p ∈ F[z] | p(3) = 0} is a subspace of F[z]. Again, to check this, we

need to verify the three conditions of Lemma 4.3.2.
Certainly the zero polynomial p(z) = 0z n +0z n−1 +· · ·+0z +0 is in U since p(z) evaluated
at 3 is 0. If f (z), g(z) ∈ U, then f (3) = g(3) = 0 so that (f + g)(3) = f (3) + g(3) = 0 + 0 = 0.
Hence f + g ∈ U, which proves closure under addition. Similarly, (af )(3) = af (3) = a0 = 0
for any a ∈ F, which proves closure under scalar multiplication.
Example 4.3.8. As in Example 4.1.6, let D ⊂ R be a subset of R, and let C ∞ (D) denote the
set of all smooth (a.k.a. continuously differentiable) functions with domain D and codomain
R. Then, under the same operations of pointwise addition and scalar multiplication, one can
show that C ∞ (D) is a subspace of C(D).
0
z
−2 4
2
0
−4 −2
4 x
−4 2
0
−2
−4
y
Figure 4.1: The intersection U ∩ U ′ of two subspaces is a subspace
Example 4.3.9. The subspaces of R2 consist of {0}, all lines through the origin, and R2
itself. The subspaces of R3 are {0}, all lines through the origin, all planes through the origin,
and R3 . In fact, these exhaust all subspaces of R2 and R3 , respectively. To prove this, we
will need further tools such as the notion of bases and dimensions to be discussed soon.
In particular, this shows that lines and planes that do not pass through the origin are not
subspaces (which is not so hard to show!).
Note that if U and U ′ are subspaces of V , then their intersection U ∩ U ′ is also a

subspace (see Proof-writing Exercise 2 on page 47 and Figure 4.1). However, the union
of two subspaces is not necessarily a subspace. Think, for example, of the union of two lines
in R2 , as in Figure 4.2.
4.4 Sums and direct sums

Throughout this section, V is a vector space over F, and U1 , U2 ⊂ V denote subspaces.
Definition 4.4.1. Let U1 , U2 ⊂ V be subspaces of V . Define the (subspace) sum of U1

4.4. SUMS AND DIRECT SUMS 43
/ U ∪ U′
u+v ∈
u
v
U′
U
Figure 4.2: The union U ∪ U ′ of two subspaces is not necessarily a subspace
and U2 to be the set

U1 + U2 = {u1 + u2 | u1 ∈ U1 , u2 ∈ U2 }.
Check as an exercise that U1 + U2 is a subspace of V . In fact, U1 + U2 is the smallest

subspace of V that contains both U1 and U2 .
Example 4.4.2. Let
U1 = {(x, 0, 0) ∈ F3 | x ∈ F},
U2 = {(0, y, 0) ∈ F3 | y ∈ F}.
Then
U1 + U2 = {(x, y, 0) ∈ F3 | x, y ∈ F}. (4.2)
If, alternatively, U2 = {(y, y, 0) ∈ F3 | y ∈ F}, then Equation (4.2) still holds.
If U = U1 +U2 , then, for any u ∈ U, there exist u1 ∈ U1 and u2 ∈ U2 such that u = u1 +u2 .
If it so happens that u can be uniquely written as u1 + u2 , then U is called the direct sum
of U1 and U2 .
Definition 4.4.3. Suppose every u ∈ U can be uniquely written as u = u1 + u2 for u1 ∈ U1

and u2 ∈ U2 . Then we use
U = U1 ⊕ U2
to denote the direct sum of U1 and U2 .

Example 4.4.4. Let
U1 = {(x, y, 0) ∈ R3 | x, y ∈ R},
U2 = {(0, 0, z) ∈ R3 | z ∈ R}.
Then R3 = U1 ⊕ U2 . However, if instead
U2 = {(0, w, z) | w, z ∈ R},
then R3 = U1 + U2 but is not the direct sum of U1 and U2 .
Example 4.4.5. Let
U1 = {p ∈ F[z] | p(z) = a0 + a2 z 2 + · · · + a2m z 2m },

U2 = {p ∈ F[z] | p(z) = a1 + a3 z 3 + · · · + a2m+1 z 2m+1 }.
Then F[z] = U1 ⊕ U2 .
Proposition 4.4.6. Let U1 , U2 ⊂ V be subspaces. Then V = U1 ⊕ U2 if and only if the

following two conditions hold:
1. V = U1 + U2 ;
2. If 0 = u1 + u2 with u1 ∈ U1 and u2 ∈ U2 , then u1 = u2 = 0.
Proof.
(“=⇒”) Suppose V = U1 ⊕ U2 . Then Condition 1 holds by definition. Certainly 0 = 0 + 0,
and, since by uniqueness this is the only way to write 0 ∈ V , we have u1 = u2 = 0.
(“⇐=”) Suppose Conditions 1 and 2 hold. By Condition 1, we have that, for all v ∈ V ,
there exist u1 ∈ U1 and u2 ∈ U2 such that v = u1 + u2 . Suppose v = w1 + w2 with w1 ∈ U1
and w2 ∈ U2 . Subtracting the two equations, we obtain
0 = (u1 − w1 ) + (u2 − w2 ),
where u1 − w1 ∈ U1 and u2 − w2 ∈ U2 . By Condition 2, this implies u1 − w1 = 0 and

u2 − w2 = 0, or equivalently u1 = w1 and u2 = w2 , as desired.
Proposition 4.4.7. Let U1 , U2 ⊂ V be subspaces. Then V = U1 ⊕ U2 if and only if the

following two conditions hold:
1. V = U1 + U2 ;
2. U1 ∩ U2 = {0}.
Proof.
(“=⇒”) Suppose V = U1 ⊕ U2 . Then Condition 1 holds by definition. If u ∈ U1 ∩ U2 , then
0 = u + (−u) with u ∈ U1 and −u ∈ U2 (why?). By Proposition 4.4.6, we have u = 0 and
−u = 0 so that U1 ∩ U2 = {0}.
(“⇐=”) Suppose Conditions 1 and 2 hold. To prove that V = U1 ⊕ U2 holds, suppose that
0 = u1 + u2 , where u1 ∈ U1 and u2 ∈ U2 . (4.3)
By Proposition 4.4.6, it suffices to show that u1 = u2 = 0. Equation (4.3) implies that

u1 = −u2 ∈ U2 . Hence u1 ∈ U1 ∩ U2 , which in turn implies that u1 = 0. It then follows that
u2 = 0 as well.
Everything in this section can be generalized to m subspaces U1 , U2 , . . . Um , with the

notable exception of Proposition 4.4.7. To see, this consider the following example.
Example 4.4.8. Let
U1 = {(x, y, 0) ∈ F3 | x, y ∈ F},
U2 = {(0, 0, z) ∈ F3 | z ∈ F},
U3 = {(0, y, y) ∈ F3 | y ∈ F}.
Then certainly F3 = U1 + U2 + U3 , but F3 6= U1 ⊕ U2 ⊕ U3 since, for example,
(0, 0, 0) = (0, 1, 0) + (0, 0, 1) + (0, −1, −1).
But U1 ∩ U2 = U1 ∩ U3 = U2 ∩ U3 = {0} so that the analog of Proposition 4.4.7 does not

hold.
1. For each of the following sets, either show that the set is a vector space or explain why
it is not a vector space.
(a) The set R of real numbers under the usual operations of addition and multiplica-
tion.
(b) The set {(x, 0) | x ∈ R} under the usual operations of addition and multiplication
on R2 .
(c) The set {(x, 1) | x ∈ R} under the usual operations of addition and multiplication
on R2 .
(d) The set {(x, 0) | x ∈ R, x ≥ 0} under the usual operations of addition and
multiplication on R2 .
(e) The set {(x, 1) | x ∈ R, x ≥ 0} under the usual operations of addition and
multiplication on R2 .
(" # )
a a+b
(f) The set | a, b ∈ R under the usual operations of addition and
a+b a
multiplication on R2×2 .
(" # )
a a+b+1
(g) The set | a, b ∈ R under the usual operations of addition
a+b a
and multiplication on R2×2 .
2. Show that the space V = {(x1 , x2 , x3 ) ∈ F3 | x1 + 2x2 + 2x3 = 0} forms a vector space.
3. For each of the following sets, either show that the set is a subspace of C(R) or explain
why it is not a subspace.
(a) The set {f ∈ C(R) | f (x) ≤ 0, ∀ x ∈ R}.

(b) The set {f ∈ C(R) | f (0) = 0}.
(c) The set {f ∈ C(R) | f (0) = 2}.
(d) The set of all constant functions.
(e) The set {α + β sin(x) | α, β ∈ R}.
4. Give an example of a nonempty subset U ⊂ R2 such that U is closed under scalar

multiplication but is not a subspace of R2 .
5. Let F[z] denote the vector space of all polynomials having coefficient over F, and define
U to be the subspace of F[z] given by
U = {az 2 + bz 5 | a, b ∈ F}.
Find a subspace W of F[z] such that F[z] = U ⊕ W .
1. Let V be a vector space over F. Then, given a ∈ F and v ∈ V such that av = 0, prove
that either a = 0 or v = 0.
2. Let V be a vector space over F, and suppose that W1 and W2 are subspaces of V .
Prove that their intersection W1 ∩ W2 is also a subspace of V .
3. Prove or give a counterexample to the following claim:
Claim. Let V be a vector space over F, and suppose that W1 , W2 , and W3 are subspaces
of V such that W1 + W3 = W2 + W3 . Then W1 = W2 .
Claim. Let V be a vector space over F, and suppose that W1 , W2 , and W3 are subspaces
of V such that W1 ⊕ W3 = W2 ⊕ W3 . Then W1 = W2 .
Chapter 5
Span and Bases
Intuition probably tells you that the plane R2 is of dimension two and that the space we live
in R3 is of dimension three. You have probably also learned in physics that space-time has
dimension four and that string theories are models that can live in ten dimensions. In this
chapter we will give a mathematical definition of the dimension of a vector space. For this
we will first need the notions of linear span, linear independence, and the basis of a vector
space.
5.1 Linear span

As before, let V denote a vector space over F. Given vectors v1 , v2 , . . . , vm ∈ V , a vector
v ∈ V is a linear combination of (v1 , . . . , vm ) if there exist scalars a1 , . . . , am ∈ F such
that
v = a1 v1 + a2 v2 + · · · + am vm .
Definition 5.1.1. The linear span (or simply span) of (v1 , . . . , vm ) is defined as
span(v1 , . . . , vm ) := {a1 v1 + · · · + am vm | a1 , . . . , am ∈ F}.
Lemma 5.1.2. Let V be a vector space and v1 , v2 , . . . , vm ∈ V . Then
1. vj ∈ span(v1 , v2 , . . . , vm ).
2. span(v1 , v2 , . . . , vm ) is a subspace of V .
48
5.1. LINEAR SPAN 49
3. If U ⊂ V is a subspace such that v1 , v2 , . . . vm ∈ U, then span(v1 , v2 , . . . , vm ) ⊂ U.

Proof. Property 1 is obvious. For Property 2, note that 0 ∈ span(v1 , v2 , . . . , vm ) and that
span(v1 , v2 , . . . , vm ) is closed under addition and scalar multiplication. For Property 3, note
that a subspace U of a vector space V is closed under addition and scalar multiplication.
Hence, if v1 , . . . , vm ∈ U, then any linear combination a1 v1 + · · · + am vm must also be an
element of U.
Lemma 5.1.2 implies that span(v1 , v2 , . . . , vm ) is the smallest subspace of V containing

each of v1 , v2 , . . . , vm .
Definition 5.1.3. If span(v1 , . . . , vm ) = V , then we say that (v1 , . . . , vm ) spans V and we
call V finite-dimensional. A vector space that is not finite-dimensional is called infinite-
dimensional.
Example 5.1.4. The vectors e1 = (1, 0, . . . , 0), e2 = (0, 1, 0, . . . , 0), . . . , en = (0, . . . , 0, 1)
span Fn . Hence Fn is finite-dimensional.
Example 5.1.5. The vectors v1 = (1, 1, 0) and v2 = (1, −1, 0) span a subspace of R3 . More
precisely, if we write the vectors in R3 as 3-tuples of the form (x, y, z), then span(v1 , v2 ) is
the xy-plane in R3 .
Example 5.1.6. Recall that if p(z) = am z m +am−1 z m−1 +· · ·+a1 z+a0 ∈ F[z] is a polynomial
with coefficients in F such that am 6= 0, then we say that p(z) has degree m. By convention,
the degree of the zero polynomial p(z) = 0 is −∞. We denote the degree of p(z) by deg(p(z)).
Define
Fm [z] = set of all polynomials in F[z] of degree at most m.
Then Fm [z] ⊂ F[z] is a subspace since Fm [z] contains the zero polynomial and is closed
under addition and scalar multiplication. In fact, Fm [z] is a finite-dimensional subspace of
F[z] since
Fm [z] = span(1, z, z 2 , . . . , z m ).
At the same time, though, note that F[z] itself is infinite-dimensional. To see this, assume
the contrary, namely that
F[z] = span(p1 (z), . . . , pk (z))
for a finite set of k polynomials p1 (z), . . . , pk (z). Let m = max(deg p1 (z), . . . , deg pk (z)).
Then z m+1 ∈ F[z], but z m+1 ∈
/ span(p1 (z), . . . , pk (z)).
50 CHAPTER 5. SPAN AND BASES
5.2 Linear independence

We are now going to define the notion of linear independence of a list of vectors. This concept
will be extremely important in the sections that follow, and especially when we introduce
bases and the dimension of a vector space.
Definition 5.2.1. A list of vectors (v1 , . . . , vm ) is called linearly independent if the only
solution for a1 , . . . , am ∈ F to the equation
a1 v1 + · · · + am vm = 0
is a1 = · · · = am = 0. In other word, the zero vector can only trivially be written as a linear
combination of (v1 , . . . , vm ).
Definition 5.2.2. A list of vectors (v1 , . . . , vm ) is called linearly dependent if it is not

linearly independent. That is, (v1 , . . . , vm ) is linear dependent if there exist a1 , . . . , am ∈ F,
not all zero, such that
a1 v1 + · · · + am vm = 0.
Example 5.2.3. The vectors (e1 , . . . , em ) of Example 5.1.4 are linearly independent. To see
this, note that the only solution to the vector equation
0 = a1 e1 + · · · + am em = (a1 , . . . , am )
is a1 = · · · = am = 0. Alternatively, we can reinterpret this vector equation as the homoge-

neous linear system 
a1 = 0 



a2 = 0 
.. .. .. ,
. . . 




am = 0
which clearly has only the trivial solution. (See Section 12.3.2 for the appropriate definitions.)
Example 5.2.4. The vectors v1 = (1, 1, 1), v2 = (0, 1, −1), and v3 = (1, 2, 0) are linearly
5.2. LINEAR INDEPENDENCE 51
dependent. To see this, we need to consider the vector equation
a1 v1 + a2 v2 + a3 v3 = a1 (1, 1, 1) + a2 (0, 1, −1) + a3 (1, 2, 0)

= (a1 + a3 , a1 + a2 + 2a3 , a1 − a2 ) = (0, 0, 0).
Solving for a1 , a2 , and a3 , we see, for example, that (a1 , a2 , a3 ) = (1, 1, −1) is a nonzero
solution. Alternatively, we can reinterpret this vector equation as the homogeneous linear
system 
a1 + a3 = 0  
a1 + a2 + 2a3 = 0 .


a1 − a2 = 0
Using the techniques of Section 12.3, we see that solving this linear system is equivalent to
solving the following linear system:
)
a1 + a3 = 0
.
a2 + a3 = 0
Note that this new linear system clearly has infinitely many solutions. In particular, the set
of all solutions is given by
N = {(a1 , a2 , a3 ) ∈ Fn | a1 = a2 = −a3 } = span((1, 1, −1)).
Example 5.2.5. The vectors (1, z, . . . , z m ) in the vector space Fm [z] are linearly indepen-
dent. Requiring that
a0 1 + a1 z + · · · + am z m = 0
means that the polynomial on the left should be zero for all z ∈ F. This is only possible for
a0 = a1 = · · · = am = 0.
An important consequence of the notion of linear independence is the fact that any vector
in the span of a given list of linearly independent vectors can be uniquely written as a linear
combination.
Lemma 5.2.6. The list of vectors (v1 , . . . , vm ) is linearly independent if and only if every
v ∈ span(v1 , . . . , vm ) can be uniquely written as a linear combination of (v1 , . . . , vm ).
Proof.
(“=⇒”) Assume that (v1 , . . . , vm ) is a linearly independent list of vectors. Suppose there are
two ways of writing v ∈ span(v1 , . . . , vm ) as a linear combination of the vi :
v = a1 v1 + · · · am vm ,
v = a′1 v1 + · · · a′m vm .
Subtracting the two equations yields 0 = (a1 − a′1 )v1 + · · · + (am − a′m )vm . Since (v1 , . . . , vm )
is linearly independent, the only solution to this equation is a1 − a′1 = 0, . . . , am − a′m = 0,
or equivalently a1 = a′1 , . . . , am = a′m .
(“⇐=”) Now assume that, for every v ∈ span(v1 , . . . , vm ), there are unique a1 , . . . , am ∈ F
such that
v = a1 v1 + · · · + am vm .
This implies, in particular, that the only way the zero vector v = 0 can be written as a
linear combination of v1 , . . . , vm is with a1 = · · · = am = 0. This shows that (v1 , . . . , vm ) are
linearly independent.
It is clear that if (v1 , . . . , vm ) is a list of linearly independent vectors, then the list
(v1 , . . . , vm−1 ) is also linearly independent.
For the next lemma, we introduce the following notation: If we want to drop a vector vj
from a given list (v1 , . . . , vm ) of vectors, then we indicate the dropped vector by a hat. I.e.,
we write
(v1 , . . . , v̂j , . . . , vm ) = (v1 , . . . , vj−1, vj+1 , . . . , vm ).
Lemma 5.2.7 (Linear Dependence Lemma). If (v1 , . . . , vm ) is linearly dependent and

v1 6= 0, then there exists an index j ∈ {2, . . . , m} such that the following two conditions hold.
1. vj ∈ span(v1 , . . . , vj−1 ).
2. If vj is removed from (v1 , . . . , vm ), then span(v1 , . . . , v̂j , . . . , vm ) = span(v1 , . . . , vm ).
Proof. Since (v1 , . . . , vm ) is linearly dependent there exist a1 , . . . , am ∈ F not all zero such
that a1 v1 + · · · + am vm = 0. Since by assumption v1 6= 0, not all of a2 , . . . , am can be zero.
5.2. LINEAR INDEPENDENCE 53
Let j ∈ {2, . . . , m} be largest such that aj 6= 0. Then we have
a1 aj−1
vj = − v1 − · · · − vj−1 , (5.1)
aj aj
which implies Part 1.

Let v ∈ span(v1 , . . . , vm ). This means, by definition, that there exist scalars b1 , . . . , bm ∈ F
such that
v = b1 v1 + · · · + bm vm .
The vector vj that we determined in Part 1 can be replaced by Equation (5.1) so that v
is written as a linear combination of (v1 , . . . , v̂j , . . . , vm ). Hence, span(v1 , . . . , v̂j , . . . , vm ) =
span(v1 , . . . , vm ).
Example 5.2.8. The list (v1 , v2 , v3 ) = ((1, 1), (1, 2), (1, 0)) of vectors spans R2 . To see
this, take any vector v = (x, y) ∈ R2 . We want to show that v can be written as a linear
combination of (1, 1), (1, 2), (1, 0), i.e., that there exist scalars a1 , a2 , a3 ∈ F such that
v = a1 (1, 1) + a2 (1, 2) + a3 (1, 0),
or equivalently that
(x, y) = (a1 + a2 + a3 , a1 + 2a2 ).
Clearly a1 = y, a2 = 0, and a3 = x − y form a solution for any choice of x, y ∈ R, and so

R2 = span((1, 1), (1, 2), (1, 0)). However, note that
2(1, 1) − (1, 2) − (1, 0) = (0, 0), (5.2)
which shows that the list ((1, 1), (1, 2), (1, 0)) is linearly dependent. The Linear Dependence
Lemma 5.2.7 thus states that one of the vectors can be dropped from ((1, 1), (1, 2), (1, 0))
and that the resulting list of vectors will still span R2 . Indeed, by Equation (5.2),
v3 = (1, 0) = 2(1, 1) − (1, 2) = 2v1 − v2 ,
and so span((1, 1), (1, 2), (1, 0)) = span((1, 1), (1, 2)).
The next results shows that linearly independent lists of vectors that span a finite-
dimensional vector space are the smallest possible spanning sets.

Theorem 5.2.9. Let V be a finite-dimensional vector space. Suppose that (v1 , . . . , vm ) is a
linearly independent list of vectors that spans V , and let (w1 , . . . , wn ) be any list that spans
V . Then m ≤ n.
Proof. The proof uses the following iterative procedure: start with an arbitrary list of vectors
S0 = (w1 , . . . , wn ) such that V = span(S0 ). At the k th step of the procedure, we construct
a new list Sk by replacing some vector wjk by the vector vk such that Sk still spans V .
Repeating this for all vk then produces a new list Sm of length n that contains each of
v1 , . . . , vm , which then proves that m ≤ n. Let us now discuss each step in this procedure in
detail.
Step 1. Since (w1 , . . . , wn ) spans V , adding a new vector to the list makes the new list
linearly dependent. Hence (v1 , w1 , . . . , wn ) is linearly dependent. By Lemma 5.2.7, there
exists an index j1 such that
wj1 ∈ span(v1 , w1 , . . . , wj1 −1 ).
Hence S1 = (v1 , w1 , . . . , ŵj1 , . . . , wn ) spans V . In this step, we added the vector v1 and
removed the vector wj1 from S0 .
Step k. Suppose that we already added v1 , . . . , vk−1 to our spanning list and removed the
vectors wj1 , . . . , wjk−1 in return. Call this list Sk−1 , and note that V = span(Sk−1 ). Add the
vector vk to Sk−1 . By the same arguments as before, adjoining the extra vector vk to the
spanning list Sk−1 yields a list of linearly dependent vectors. Hence, by Lemma 5.2.7, there
exists an index jk such that Sk−1 with vk added and wjk removed still spans V . The fact
that (v1 , . . . , vk ) is linearly independent ensures that the vector removed is indeed among
the wj . Call the new list Sk , and note that V = span(Sk ).
The final list Sm is S0 but with each v1 , . . . , vm added and each wj1 , . . . , wjm removed.
Moreover, note that Sm has length n and still spans V . It follows that m ≤ n.
5.3 Bases
A basis of a finite-dimensional vector space is a spanning list that is also linearly independent.
We will see that all bases for finite-dimensional vector spaces have the same length. This
5.3. BASES 55
length will be then be called the dimension of our vector space.
Definition 5.3.1. A list of vectors (v1 , . . . , vm ) is a basis for the finite-dimensional vector
space V if (v1 , . . . , vm ) is linearly independent and V = span(v1 , . . . , vm ).
If (v1 , . . . , vm ) forms a basis of V , then, by Lemma 5.2.6, every vector v ∈ V can be

uniquely written as a linear combination of (v1 , . . . , vm ).
Example 5.3.2. (e1 , . . . , en ) is a basis of Fn . There are, of course, other bases. For example,
((1, 2), (1, 1)) is a basis of F2 . Note that the list ((1, 1)) is also linearly independent, but it
does not span F2 and hence is not a basis.
Example 5.3.3. (1, z, z 2 , . . . , z m ) is a basis of Fm [z].
Theorem 5.3.4 (Basis Reduction Theorem). If V = span(v1 , . . . , vm ), then either

(v1 , . . . , vm ) is a basis of V or some vi can be removed to obtain a basis of V .
Proof. Suppose V = span(v1 , . . . , vm ). We start with the list S = (v1 , . . . , vm ) and iteratively
run through all vectors vk for k = 1, 2, . . . , m to determine whether to keep or remove them
from S:
Step 1. If v1 = 0, then remove v1 from S. Otherwise, leave S unchanged.
Step k. If vk ∈ span(v1 , . . . , vk−1 ), then remove vk from S. Otherwise, leave S unchanged.
The final list S still spans V since, at each step, a vector was only discarded if it was already
in the span of the previous vectors. The process also ensures that no vector is in the span
of the previous vectors. Hence, by the Linear Dependence Lemma 5.2.7, the final list S is
linearly independent. It follows that S is a basis of V .
Example 5.3.5. To see how Basis Reduction Theorem 5.3.4 works, consider the list of
vectors
S = ((1, −1, 0), (2, −2, 0), (−1, 0, 1), (0, −1, 1), (0, 1, 0)).
This list does not form a basis for R3 as it is not linearly independent. However, it is clear
that R3 = span(S) since any arbitrary vector v = (x, y, z) ∈ R3 can be written as the
following linear combination over S:
v = (x + z)(1, −1, 0) + 0(2, −2, 0) + (z)(−1, 0, 1) + 0(0, −1, 1) + (x + y + z)(0, 1, 0).

In fact, since the coefficients of (2, −2, 0) and (0, −1, 1) in this linear combination are both
zero, it suggests that they add nothing to the span of the subset
B = ((1, −1, 0), (−1, 0, 1), (0, 1, 0))
of S. Moreover, one can show that B is a basis for R3 , and it is exactly the basis produced
by applying the process from the proof of Theorem 5.3.4 (as you should be able to verify).
Corollary 5.3.6. Every finite-dimensional vector space has a basis.
Proof. By definition, a finite-dimensional vector space has a spanning list. By the Basis
Reduction Theorem 5.3.4, any spanning list can be reduced to a basis.
Theorem 5.3.7 (Basis Extension Theorem). Every linearly independent list of vectors
in a finite-dimensional vector space V can be extended to a basis of V .
Proof. Suppose V is finite-dimensional and that (v1 , . . . , vm ) is linearly independent. Since

V is finite-dimensional, there exists a list (w1 , . . . , wn ) of vectors that spans V . We wish to
adjoin some of the wk to (v1 , . . . , vm ) in order to create a basis of V .
Step 1. If w1 ∈ span(v1 , . . . , vm ), then let S = (v1 , . . . , vm ). Otherwise, S = (v1 , . . . , vm , w1 ).
Step k. If wk ∈ span(S), then leave S unchanged. Otherwise, adjoin wk to S.
After each step, the list S is still linearly independent since we only adjoined wk if wk was
not in the span of the previous vectors. After n steps, wk ∈ span(S) for all k = 1, 2, . . . , n.
Since (w1 , . . . , wn ) was a spanning list, S spans V so that S is indeed a basis of V .
Example 5.3.8. Take the two vectors v1 = (1, 1, 0, 0) and v2 = (1, 0, 1, 0) in R4 . One may
easily check that these two vectors are linearly independent, but they do not form a basis
of R4 . We know that (e1 , e2 , e3 , e4 ) spans R4 . (In fact, it is even a basis.) Following the
algorithm outlined in the proof of the Basis Extension Theorem, we see that e1 6∈ span(v1 , v2 ).
Hence, we adjoin e1 to obtain S = (v1 , v2 , e1 ). Note that now
e2 = (0, 1, 0, 0) = 1v1 + 0v2 + (−1)e1
so that e2 ∈ span(v1 , v2 , e1 ), and so we leave S unchanged. Similarly,
e3 = (0, 0, 1, 0) = 0v1 + 1v2 + (−1)e1 ,

5.4. DIMENSION 57
and hence e3 ∈ span(v1 , v2 , e1 ), which means that we again leave S unchanged. Finally,
e4 6∈ span(v1 , v2 , e1 ), and so we adjoin it to obtain a basis (v1 , v2 , e1 , e4 ) of R4 .
5.4 Dimension
We now come to the important definition of the dimension of a finite-dimensional vector
space. Intuitively, we know that R2 has dimension 2, that R3 has dimension 3, and, more
generally, that Rn has dimension n. This is precisely the length every basis for each of these
vector spaces, which prompts the following definition.
Definition 5.4.1. We call the length of any basis for V (which is well-defined by Theo-
rem 5.4.2 below) the dimension of V , and we denote this by dim(V ).
Note that Definition 5.4.1 only makes sense if, in fact, every basis for a given finite-
dimensional vector space has the same length. This is true by the following theorem.
Theorem 5.4.2. Let V be a finite-dimensional vector space. Then any two bases of V have
the same length.
Proof. Let (v1 , . . . , vm ) and (w1 , . . . , wn ) be two bases of V . Both span V . By Theorem 5.2.9,
we have m ≤ n since (v1 , . . . , vm ) is linearly independent. By the same theorem, we also
have n ≤ m since (w1 , . . . , wn ) is linearly independent. Hence n = m, as asserted.
Example 5.4.3. dim(Fn ) = n and dim(Fm [z]) = m + 1. Note that dim(Cn ) = n as a

complex vector space, whereas dim(Cn ) = 2n as an real vector space. This comes from the
fact that we can view C itself as an real vector space of dimension 2 with basis (1, i).
Theorem 5.4.4. Let V be a finite-dimensional vector space with dim(V ) = n. Then:
1. If U ⊂ V is a subspace of V , then dim(U) ≤ dim(V ).
2. If V = span(v1 , . . . , vn ), then (v1 , . . . , vn ) is a basis of V .
3. If (v1 , . . . , vn ) is linearly independent in V , then (v1 , . . . , vn ) is a basis of V .
Point 1 implies, in particular, that every subspace of a finite-dimensional vector space is

finite-dimensional. Points 2 and 3 show that if the dimension of a vector space is known to
be n, then, to check that a list of n vectors is a basis, it is enough to check whether it spans
V (resp. is linearly independent).
Proof. To prove Point 1, let (u1 , . . . , um ) be a basis of U. This list is linearly independent
in both U and V . By the Basis Extension Theorem 5.3.7, we can extend (u1 , . . . , um ) to a
basis for V , which is of length n since dim(V ) = n. This implies that m ≤ n, as desired.
To prove Point 2, suppose that (v1 , . . . , vn ) spans V . Then, by the Basis Reduction
Theorem 5.3.4, this list can be reduced to a basis. However, every basis of V has length n;
hence, no vector needs to be removed from (v1 , . . . , vn ). It follows that (v1 , . . . , vn ) is already
a basis of V .
Point 3 is proven in a very similar fashion. Suppose (v1 , . . . , vn ) is linearly independent.
By the Basis Extension Theorem 5.3.7, this list can be extended to a basis. However, every
basis has length n; hence, no vector needs to be added to (v1 , . . . , vn ). It follows that
(v1 , . . . , vn ) is already a basis of V .
We conclude this chapter with some additional interesting results on bases and dimen-
sions. The first one combines the concepts of basis and direct sum.
Theorem 5.4.5. Let U ⊂ V be a subspace of a finite-dimensional vector space V . Then

there exists a subspace W ⊂ V such that V = U ⊕ W .
Proof. Let (u1 , . . . , um ) be a basis of U. By Theorem 5.4.4(1), we know that m ≤ dim(V ).

Hence, by the Basis Extension Theorem 5.3.7, (u1 , . . . , um ) can be extended to a basis
(u1 , . . . , um, w1 , . . . , wn ) of V . Let W = span(w1 , . . . , wn ).
To show that V = U ⊕ W , we need to show that V = U + W and U ∩ W = {0}. Since
V = span(u1 , . . . , um , w1 , . . . , wn ) where (u1 , . . . , um ) spans U and (w1 , . . . , wn ) spans W , it
is clear that V = U + W .
To show that U ∩W = {0}, let v ∈ U ∩W . Then there exist scalars a1 , . . . , am , b1 , . . . , bn ∈
F such that
v = a1 u1 + · · · + am um = b1 w1 + · · · + bn wn ,
or equivalently that
a1 u1 + · · · + am um − b1 w1 − · · · − bn wn = 0.
Since (u1 , . . . , um, w1 , . . . , wn ) forms a basis of V and hence is linearly independent, the only
solution to this equation is a1 = · · · = am = b1 = · · · = bn = 0. Hence v = 0, proving that
indeed U ∩ W = {0}.
5.4. DIMENSION 59
Theorem 5.4.6. If U, W ⊂ V are subspaces of a finite-dimensional vector space, then
dim(U + W ) = dim(U) + dim(W ) − dim(U ∩ W ).
Proof. Let (v1 , . . . , vn ) be a basis of U ∩ W . By the Basis Extension Theorem 5.3.7, there
exist (u1 , . . . , uk ) and (w1 , . . . , wℓ ) such that (v1 , . . . , vn , u1 , . . . , uk ) is a basis of U and
(v1 , . . . , vn , w1 , . . . , wℓ ) is a basis of W . It suffices to show that
B = (v1 , . . . , vn , u1, . . . , uk , w1 , . . . , wℓ )
is a basis of U + W since then
dim(U + W ) = n + k + ℓ = (n + k) + (n + ℓ) − n = dim(U) + dim(W ) − dim(U ∩ W ).
Clearly span(v1 , . . . , vn , u1 , . . . , uk , w1 , . . . , wℓ ) contains U and W , and hence U + W . To

show that B is a basis, it remains to show that B is linearly independent. Suppose
a1 v1 + · · · + an vn + b1 u1 + · · · + bk uk + c1 w1 + · · · + cℓ wℓ = 0, (5.3)
and let u = a1 v1 + · · · + an vn + b1 u1 + · · · + bk uk ∈ U. Then, by Equation (5.3), we also

have that u = −c1 w1 − · · · − cℓ wℓ ∈ W , which implies that u ∈ U ∩ W . Hence, there
exist scalars a′1 , . . . , a′n ∈ F such that u = a′1 v1 + · · · + a′n vn . Since there is a unique linear
combination of the linearly independent vectors (v1 , . . . , vn , u1 , . . . , uk ) that describes u, we
must have b1 = · · · = bk = 0 and a1 = a′1 , . . . , an = a′n . Since (v1 , . . . , vn , w1 , . . . , wℓ ) is also
linearly independent, it further follows that a1 = · · · = an = c1 = · · · = cℓ = 0. Hence,
Equation (5.3) only has the trivial solution, which implies that B is a basis.

1. Show that the vectors v1 = (1, 1, 1), v2 = (1, 2, 3), and v3 = (2, −1, 1) are linearly
independent in R3 . Write v = (1, −2, 5) as a linear combination of v1 , v2 , and v3 .
2. Consider the complex vector space V = C3 and the list (v1 , v2 , v3 ) of vectors in V ,
where
v1 = (i, 0, 0), v2 = (i, 1, 0), v3 = (i, i, −1) .
(a) Prove that span(v1 , v2 , v3 ) = V .

(b) Prove or disprove: (v1 , v2 , v3 ) is a basis for V .
3. Determine the dimension of each of the following subspaces of F4 .
(a) {(x1 , x2 , x3 , x4 ) ∈ F4 | x4 = 0}.

(b) {(x1 , x2 , x3 , x4 ) ∈ F4 | x4 = x1 + x2 }.
(c) {(x1 , x2 , x3 , x4 ) ∈ F4 | x4 = x1 + x2 , x3 = x1 − x2 }.
(d) {(x1 , x2 , x3 , x4 ) ∈ F4 | x4 = x1 + x2 , x3 = x1 − x2 , x3 + x4 = 2x1 }.
(e) {(x1 , x2 , x3 , x4 ) ∈ F4 | x1 = x2 = x3 = x4 }.
4. Determine the value of λ ∈ R for which each list of vectors is linear dependent.
(a) ((λ, −1, −1), (−1, λ, −1), (−1, −1, λ)) as a subset of R3 .

(b) sin2 (x), cos(2x), λ as a subset of C(R).
5. Consider the real vector space V = R4 . For each of the following five statements,
provide either a proof or a counterexample.
(a) dim V = 4.
(b) span((1, 1, 0, 0), (0, 1, 1, 0), (0, 0, 1, 1)) = V .
(c) The list ((1, −1, 0, 0), (0, 1, −1, 0), (0, 0, 1, −1), (−1, 0, 0, 1)) is linearly independent.
(d) Every list of four vectors v1 , . . . , v4 ∈ V , such that span(v1 , . . . , v4 ) = V , is linearly
independent.
5.4. DIMENSION 61
(e) Let v1 and v2 be two linearly independent vectors in V . Then, there exist vectors
u, w ∈ V , such that (v1 , v2 , u, w) is a basis for V .
1. Let V be a vector space over F, and suppose that the list (v1 , v2 , . . . , vn ) of vectors
spans V , where each vi ∈ V . Prove that the list
(v1 − v2 , v2 − v3 , v3 − v4 , . . . , vn−2 − vn−1 , vn−1 − vn , vn )
also spans V .
2. Let V be a vector space over F, and suppose that (v1 , v2 , . . . , vn ) is a linearly indepen-
dent list of vectors in V . Given any w ∈ V such that
(v1 + w, v2 + w, . . . , vn + w)
is a linearly dependent list of vectors in V , prove that w ∈ span(v1 , v2 , . . . , vn ).
3. Let V be a finite-dimensional vector space over F with dim(V ) = n for some n ∈ Z+ .

Prove that there are n one-dimensional subspaces U1 , U2 , . . . , Un of V such that
V = U1 ⊕ U2 ⊕ · · · ⊕ Un .
4. Let V be a finite-dimensional vector space over F, and suppose that U is a subspace

of V for which dim(U) = dim(V ). Prove that U = V .
5. Let Fm [z] denote the vector space of all polynomials with degree less than or equal to
m ∈ Z+ and having coefficient over F, and suppose that p0 , p1 , . . . , pm ∈ Fm [z] satisfy
pj (2) = 0. Prove that (p0 , p1 , . . . , pm ) is a linearly dependent list of vectors in Fm [z].
6. Let U and V be five-dimensional subspaces of R9 . Prove that U ∩ V 6= {0}.
7. Let V be a finite-dimensional vector space over F, and suppose that U1 , U2 , . . . , Um are

any m subspaces of V . Prove that
dim(U1 + U2 + · · · + Um ) ≤ dim(U1 ) + dim(U2 ) + · · · + dim(Um ).

Chapter 6
Linear Maps
As discussed in Chapter 1, one of the main goals of Linear Algebra is the characterization
of solutions to a system of m linear equations in n unknowns x1 , . . . , xn ,


a11 x1 + · · · + a1n xn = b1 

.. .. ..
. . . ,


a x +···+a x = b 
m1 1 mn n m
where each of the coefficients aij and bi is in F. Linear maps and their properties are what
give us insight into the characteristics of solutions to linear systems.
6.1 Definition and elementary properties

Throughout this chapter, V and W denote vector spaces over F. We are going to study
functions from V into W that have the special properties given in the following definition.
Definition 6.1.1. A function T : V → W is called linear if
T (u + v) = T (u) + T (v), for all u, v ∈ V , (6.1)

T (av) = aT (v), for all a ∈ F and v ∈ V . (6.2)
The set of all linear maps from V to W is denoted by L(V, W ). We also write T v for T (v).
Moreover, if V = W , then we write L(V, V ) = L(V ) and call T ∈ L(V ) a linear
operator on V .
62
6.1. DEFINITION AND ELEMENTARY PROPERTIES 63
Example 6.1.2.
1. The zero map 0 : V → W mapping every element v ∈ V to 0 ∈ W is linear.
2. The identity map I : V → V defined as Iv = v is linear.
3. Let T : F[z] → F[z] be the differentiation map defined as T p(z) = p′ (z). Then, for
two polynomials p(z), q(z) ∈ F[z], we have
T (p(z) + q(z)) = (p(z) + q(z))′ = p′ (z) + q ′ (z) = T (p(z)) + T (q(z)).
Similarly, for a polynomial p(z) ∈ F[z] and a scalar a ∈ F, we have
T (ap(z)) = (ap(z))′ = ap′ (z) = aT (p(z)).
Hence T is linear.
4. Let T : R2 → R2 be the map given by T (x, y) = (x − 2y, 3x + y). Then, for

(x, y), (x′, y ′ ) ∈ R2 , we have
T ((x, y) + (x′ , y ′)) = T (x + x′ , y + y ′ ) = (x + x′ − 2(y + y ′), 3(x + x′ ) + y + y ′ )

= (x − 2y, 3x + y) + (x′ − 2y ′, 3x′ + y ′ ) = T (x, y) + T (x′ , y ′ ).
Similarly, for (x, y) ∈ R2 and a ∈ F, we have
T (a(x, y)) = T (ax, ay) = (ax − 2ay, 3ax + ay) = a(x − 2y, 3x + y) = aT (x, y).
Hence T is linear. More generally, any map T : Fn → Fm defined by
T (x1 , . . . , xn ) = (a11 x1 + · · · + a1n xn , . . . , am1 x1 + · · · + amn xn )
with aij ∈ F is linear.
5. Not all functions are linear! For example, the exponential function f (x) = ex is not
linear since e2x 6= 2ex when x 6= 0. Also, the function f : F → F given by f (x) = x − 1
is not linear since f (x + y) = (x + y) − 1 6= (x − 1) + (y − 1) = f (x) + f (y).
64 CHAPTER 6. LINEAR MAPS
An important result is that linear maps are already completely determined if their values
on basis vectors are specified.
Theorem 6.1.3. Let (v1 , . . . , vn ) be a basis of V and (w1 , . . . , wn ) be an arbitrary list of

vectors in W . Then there exists a unique linear map
T :V →W such that T (vi ) = wi , ∀ i = 1, 2, . . . , n.
Proof. First we verify that there is at most one linear map T with T (vi ) = wi . Take any
v ∈ V . Since (v1 , . . . , vn ) is a basis of V there are unique scalars a1 , . . . , an ∈ F such that
v = a1 v1 + · · · + an vn . By linearity, we have
T (v) = T (a1 v1 + · · · + an vn ) = a1 T (v1 ) + · · · + an T (vn ) = a1 w1 + · · · + an wn , (6.3)
and hence T (v) is completely determined. To show existence, use Equation (6.3) to define
T . It remains to show that this T is linear and that T (vi ) = wi . These two conditions are
not hard to show and are left to the reader.
The set of linear maps L(V, W ) is itself a vector space. For S, T ∈ L(V, W ) addition is
defined as
(S + T )v = Sv + T v, for all v ∈ V .
For a ∈ F and T ∈ L(V, W ), scalar multiplication is defined as
(aT )(v) = a(T v), for all v ∈ V .
You should verify that S + T and aT are indeed linear maps and that all properties of a
vector space are satisfied.
In addition to the operations of vector addition and scalar multiplication, we can also
define the composition of linear maps. Let V, U, W be vector spaces over F. Then, for
S ∈ L(U, V ) and T ∈ L(V, W ), we define T ◦ S ∈ L(U, W ) by
(T ◦ S)(u) = T (S(u)), for all u ∈ U.
The map T ◦ S is often also called the product of T and S denoted by T S. It has the
following properties:
6.2. NULL SPACES 65
1. Associativity: (T1 T2 )T3 = T1 (T2 T3 ), for all T1 ∈ L(V1 , V0 ), T2 ∈ L(V2 , V1 ) and T3 ∈

L(V3 , V2 ).
2. Identity: T I = IT = T , where T ∈ L(V, W ) and where I in T I is the identity map

in L(V, V ) whereas the I in IT is the identity map in L(W, W ).
3. Distributive property: (T1 + T2 )S = T1 S + T2 S and T (S1 + S2 ) = T S1 + T S2 , where

S, S1 , S2 ∈ L(U, V ) and T, T1 , T2 ∈ L(V, W ).
Note that the product of linear maps is not always commutative. For example, if we take
T ∈ L(F[z], F[z]) to be the differentiation map T p(z) = p′ (z) and S ∈ L(F[z], F[z]) to be the
map Sp(z) = z 2 p(z), then
(ST )p(z) = z 2 p′ (z) but (T S)p(z) = z 2 p′ (z) + 2zp(z).
6.2 Null spaces

Definition 6.2.1. Let T : V → W be a linear map. Then the null space (a.k.a. kernel)
of T is the set of all vectors in V that are mapped to zero by T . I.e.,
null (T ) = {v ∈ V | T v = 0}.
Example 6.2.2. Let T ∈ L(F[z], F[z]) be the differentiation map T p(z) = p′ (z). Then
null (T ) = {p ∈ F[z] | p(z) is constant}.
Example 6.2.3. Consider the linear map T (x, y) = (x − 2y, 3x + y) of Example 6.1.2. To
determine the null space, we need to solve T (x, y) = (0, 0), which is equivalent to the system
of linear equations )
x − 2y = 0
.
3x + y = 0
We see that the only solution is (x, y) = (0, 0) so that null (T ) = {(0, 0)}.
Proposition 6.2.4. Let T : V → W be a linear map. Then null (T ) is a subspace of V .

Proof. We need to show that 0 ∈ null (T ) and that null (T ) is closed under addition and
scalar multiplication. By linearity, we have
T (0) = T (0 + 0) = T (0) + T (0)
so that T (0) = 0. Hence 0 ∈ null (T ). For closure under addition, let u, v ∈ null (T ). Then
T (u + v) = T (u) + T (v) = 0 + 0 = 0,
and hence u + v ∈ null (T ). Similarly, for closure under scalar multiplication, let u ∈ null (T )
and a ∈ F. Then
T (au) = aT (u) = a0 = 0,
and so au ∈ null (T ).
Definition 6.2.5. The linear map T : V → W is called injective if, for all u, v ∈ V , the
condition T u = T v implies that u = v. In other words, different vectors in V are mapped to
different vectors in W .
Proposition 6.2.6. Let T : V → W be a linear map. Then T is injective if and only if

null (T ) = {0}.
Proof.
(“=⇒”) Suppose that T is injective. Since null (T ) is a subspace of V , we know that 0 ∈
null (T ). Assume that there is another vector v ∈ V that is in the kernel. Then T (v) = 0 =
T (0). Since T is injective, this implies that v = 0, proving that null (T ) = {0}.
(“⇐=”) Assume that null (T ) = {0}, and let u, v ∈ V be such that T u = T v. Then
0 = T u − T v = T (u − v) so that u − v ∈ null (T ). Hence u − v = 0, or, equivalently, u = v.
This shows that T is indeed injective.
Example 6.2.7.
1. The differentiation map p(z) 7→ p′ (z) is not injective since p′ (z) = q ′ (z) implies that
p(z) = q(z) + c, where c ∈ F is a constant.
2. The identity map I : V → V is injective.

6.3. RANGES 67
3. The linear map T : F[z] → F[z] given by T (p(z)) = z 2 p(z) is injective since it is easy
to verify that null (T ) = {0}.
4. The linear map T (x, y) = (x − 2y, 3x + y) is injective since null (T ) = {(0, 0)}, as we
calculated in Example 6.2.3.
6.3 Ranges
Definition 6.3.1. Let T : V → W be a linear map. The range of T , denoted by range (T ),
is the subset of vectors in W that are in the image of T . I.e.,
range (T ) = {T v | v ∈ V } = {w ∈ W | there exists v ∈ V such that T v = w}.
Example 6.3.2. The range of the differentiation map T : F[z] → F[z] is range (T ) = F[z]
since, for every polynomial q ∈ F[z], there is a p ∈ F[z] such that p′ = q.
Example 6.3.3. The range of the linear map T (x, y) = (x − 2y, 3x + y) is R2 since, for any
(z1 , z2 ) ∈ R2 , we have T (x, y) = (z1 , z2 ) if (x, y) = 71 (z1 + 2z2 , −3z1 + z2 ).
Proposition 6.3.4. Let T : V → W be a linear map. Then range (T ) is a subspace of W .
Proof. We need to show that 0 ∈ range (T ) and that range (T ) is closed under addition and
scalar multiplication. We already showed that T 0 = 0 so that 0 ∈ range (T ).
For closure under addition, let w1 , w2 ∈ range (T ). Then there exist v1 , v2 ∈ V such that
T v1 = w1 and T v2 = w2 . Hence
T (v1 + v2 ) = T v1 + T v2 = w1 + w2 ,
and so w1 + w2 ∈ range (T ).
For closure under scalar multiplication, let w ∈ range (T ) and a ∈ F. Then there exists
a v ∈ V such that T v = w. Thus
T (av) = aT v = aw,
and so aw ∈ range (T ).
Definition 6.3.5. A linear map T : V → W is called surjective if range (T ) = W . A linear

map T : V → W is called bijective if T is both injective and surjective.
Example 6.3.6.
1. The differentiation map T : F[z] → F[z] is surjective since range (T ) = F[z]. However,
if we restrict ourselves to polynomials of degree at most m, then the differentiation
map T : Fm [z] → Fm [z] is not surjective since polynomials of degree m are not in the
range of T .
2. The identity map I : V → V is surjective.
3. The linear map T : F[z] → F[z] given by T (p(z)) = z 2 p(z) is not surjective since, for
example, there are no linear polynomials in the range of T .
4. The linear map T (x, y) = (x − 2y, 3x + y) is surjective since range (T ) = R2 , as we

calculated in Example 6.3.3.
6.4 Homomorphisms
It should be mentioned that linear maps between vector spaces are also called vector space
homomorphisms. Instead of the notation L(V, W ), one often sees the convention
HomF (V, W ) = {T : V → W | T is linear}.
A homomorphism T : V → W is also often called
• Monomorphism iff T is injective;
• Epimorphism iff T is surjective;
• Isomorphism iff T is bijective;
• Endomorphism iff V = W ;
• Automorphism iff V = W and T is bijective.

6.5. THE DIMENSION FORMULA 69
6.5 The dimension formula
The next theorem is the key result of this chapter. It relates the dimension of the kernel and
range of a linear map.
Theorem 6.5.1. Let V be a finite-dimensional vector space and T : V → W be a linear

map. Then range (T ) is a finite-dimensional subspace of W and
dim(V ) = dim(null (T )) + dim(range (T )). (6.4)
Proof. Let V be a finite-dimensional vector space and T ∈ L(V, W ). Since null (T ) is a sub-
space of V , we know that null (T ) has a basis (u1 , . . . , um ). This implies that dim(null (T )) =
m. By the Basis Extension Theorem, it follows that (u1 , . . . , um ) can be extended to a basis
of V , say (u1 , . . . , um, v1 , . . . , vn ), so that dim(V ) = m + n.
The theorem will follow by showing that (T v1 , . . . , T vn ) is a basis of range (T ) since this
would imply that range (T ) is finite-dimensional and dim(range (T )) = n, proving Equa-
tion (6.4).
Since (u1 , . . . , um , v1 , . . . , vn ) spans V , every v ∈ V can be written as a linear combination
of these vectors; i.e.,
v = a1 u1 + · · · + am um + b1 v1 + · · · + bn vn ,
where ai , bj ∈ F. Applying T to v, we obtain
T v = b1 T v1 + · · · + bn T vn ,
where the terms T ui disappeared since ui ∈ null (T ). This shows that (T v1 , . . . , T vn ) indeed
spans range (T ).
To show that (T v1 , . . . , T vn ) is a basis of range (T ), it remains to show that this list is
linearly independent. Assume that c1 , . . . , cn ∈ F are such that
c1 T v1 + · · · + cn T vn = 0.
By linearity of T , this implies that
T (c1 v1 + · · · + cn vn ) = 0,
and so c1 v1 + · · · + cn vn ∈ null (T ). Since (u1 , . . . , um ) is a basis of null (T ), there must exist

scalars d1 , . . . , dm ∈ F such that
c1 v1 + · · · + cn vn = d1 u1 + · · · + dm um .
However, by the linear independence of (u1 , . . . , um , v1 , . . . , vn ), this implies that all coeffi-
cients c1 = · · · = cn = d1 = · · · = dm = 0. Thus, (T v1 , . . . , T vn ) is linearly independent, and
we are done.
Example 6.5.2. Recall that the linear map T : R2 → R2 defined by T (x, y) = (x−2y, 3x+y)
has null (T ) = {0} and range (T ) = R2 . It follows that
dim(R2 ) = 2 = 0 + 2 = dim(null (T )) + dim(range (T )).
Corollary 6.5.3. Let T ∈ L(V, W ).
1. If dim(V ) > dim(W ), then T is not injective.
2. If dim(V ) < dim(W ), then T is not surjective.
Proof. By Theorem 6.5.1, we have that
dim(null (T )) = dim(V ) − dim(range (T ))

≥ dim(V ) − dim(W ) > 0.
Since T is injective if and only if dim(null (T )) = 0, T cannot be injective.

Similarly,
dim(range (T )) = dim(V ) − dim(null (T ))

≤ dim(V ) < dim(W ),
and so range (T ) cannot be equal to W . Hence, T cannot be surjective.

6.6. THE MATRIX OF A LINEAR MAP 71
6.6 The matrix of a linear map

Now we will see that every linear map T ∈ L(V, W ), with V and W finite-dimensional vector
spaces, can be encoded by a matrix, and, vice versa, every matrix defines such a linear map.
Let V and W be finite-dimensional vector spaces, and let T : V → W be a linear map.
Suppose that (v1 , . . . , vn ) is a basis of V and that (w1 , . . . , wm ) is a basis for W . We have seen
in Theorem 6.1.3 that T is uniquely determined by specifying the vectors T v1 , . . . , T vn ∈ W .
Since (w1 , . . . , wm ) is a basis of W , there exist unique scalars aij ∈ F such that
T vj = a1j w1 + · · · + amj wm for 1 ≤ j ≤ n. (6.5)
We can arrange these scalars in an m × n matrix as follows:

 
a11 . . . a1n
 . .. 
M(T ) = 

.. . 
.
am1 . . . amn
Often, this is also written as A = (aij )1≤i≤m,1≤j≤n . As in Section 12.1.1, the set of all m × n
matrices with entries in F is denoted by Fm×n .
Remark 6.6.1. It is important to remember that M(T ) not only depends on the linear map
T but also on the choice of the basis (v1 , . . . , vn ) for V and the choice of basis (w1 , . . . , wm )
for W . The j th column of M(T ) contains the coefficients of the j th basis vector vj when
expanded in terms of the basis (w1 , . . . , wm ), as in Equation (6.5).
Example 6.6.2. Let T : R2 → R2 be the linear map given by T (x, y) = (ax+ by, cx+ dy) for
some a, b, c, d ∈ R. Then, with respect to the canonical basis of R2 given by ((1, 0), (0, 1)),
the corresponding matrix is " #
a b
M(T ) =
c d
since T (1, 0) = (a, c) gives the first column and T (0, 1) = (b, d) gives the second column.
More generally, suppose that V = Fn and W = Fm , and denote the standard basis for
V by (e1 , . . . , en ) and the standard basis for W by (f1 , . . . , fm ). Here, ei (resp. fi ) is the
n-tuple (resp. m-tuple) with a one in position i and zeroes everywhere else. Then the matrix
M(T ) = (aij ) is given by

aij = (T ej )i ,
where (T ej )i denotes the ith component of the vector T ej .

Example 6.6.3. Let T : R2 → R3 be the linear map defined by T (x, y) = (y, x + 2y, x + y).
Then, with respect to the standard basis, we have T (1, 0) = (0, 1, 1) and T (0, 1) = (1, 2, 1)
so that  
0 1
 
M(T ) = 1 2 .
1 1
However, if alternatively we take the bases ((1, 2), (0, 1)) for R2 and ((1, 0, 0), (0, 1, 0), (0, 0, 1))
for R3 , then T (1, 2) = (2, 5, 3) and T (0, 1) = (1, 2, 1) so that
 
2 1
 
M(T ) = 5 2 .
3 1
Example 6.6.4. Let S : R2 → R2 be the linear map S(x, y) = (y, x). With respect to the
basis ((1, 2), (0, 1)) for R2 , we have
S(1, 2) = (2, 1) = 2(1, 2) − 3(0, 1) and S(0, 1) = (1, 0) = 1(1, 2) − 2(0, 1),
and so "#
2 1
M(S) = .
−3 −2
Given vector spaces V and W of dimensions n and m, respectively, and given a fixed choice
of bases, note that there is a one-to-one correspondence between linear maps in L(V, W ) and
matrices in Fm×n . If we start with the linear map T , then the matrix M(T ) = A = (aij ) is
defined via Equation (6.5). Conversely, given the matrix A = (aij ) ∈ Fm×n , we can define a
linear map T : V → W by setting
m
X
T vj = aij wi .
i=1
Recall that the set of linear maps L(V, W ) is a vector space. Since we have a one-to-one
correspondence between linear maps and matrices, we can also make the set of matrices
Fm×n into a vector space. Given two matrices A = (aij ) and B = (bij ) in Fm×n and given a
scalar α ∈ F, we define the matrix addition and scalar multiplication componentwise:
A + B = (aij + bij ),
αA = (αaij ).
Next, we show that the composition of linear maps imposes a product on matrices,
also called matrix multiplication. Suppose U, V, W are vector spaces over F with bases
(u1 , . . . , up ), (v1 , . . . , vn ) and (w1 , . . . , wm ), respectively. Let S : U → V and T : V → W be
linear maps. Then the product is a linear map T ◦ S : U → W .
Each linear map has its corresponding matrix M(T ) = A, M(S) = B and M(T S) = C.
The question is whether C is determined by A and B. We have, for each j ∈ {1, 2, . . . p},
that
(T ◦ S)uj = T (b1j v1 + · · · + bnj vn ) = b1j T v1 + · · · + bnj T vn

Xn Xn X m

= bkj T vk = bkj aik wi
k=1 k=1 i=1
Xm n
X
= aik bkj wi .
i=1 k=1
Hence, the matrix C = (cij ) is given by

n
X
cij = aik bkj . (6.6)
k=1
Equation (6.6) can be used to define the m × p matrix C as the product of a m × n matrix
A and a n × p matrix B, i.e.,
C = AB. (6.7)
Our derivation implies that the correspondence between linear maps and matrices respects
the product structure.
Proposition 6.6.5. Let S : U → V and T : V → W be linear maps. Then
M(T S) = M(T )M(S).

Example 6.6.6. With notation as in Examples 6.6.3 and 6.6.4, you should be able to verify
that    
2 1 " # 1 0
  2 1  
M(T S) = M(T )M(S) = 5 2 = 4 1 .
−3 −2
3 1 3 1
Given a vector v ∈ V , we can also associate a matrix M(v) to v as follows. Let (v1 , . . . , vn )
be a basis of V . Then there are unique scalars b1 , . . . , bn such that
v = b1 v1 + · · · bn vn .
The matrix of v is then defined to be the n × 1 matrix

 
b1
.
M(v) =  .
 . .
bn
Example 6.6.7. The matrix of a vector x = (x1 , . . . , xn ) ∈ Fn in the standard basis

(e1 , . . . , en ) is the column vector or n × 1 matrix

x1
.
M(x) = 
.
.
xn
since x = (x1 , . . . , xn ) = x1 e1 + · · · + xn en .
The next result shows how the notion of a matrix of a linear map T : V → W and the
matrix of a vector v ∈ V fit together.
Proposition 6.6.8. Let T : V → W be a linear map. Then, for every v ∈ V ,
M(T v) = M(T )M(v).
Proof. Let (v1 , . . . , vn ) be a basis of V and (w1 , . . . , wm ) be a basis for W . Suppose that,
with respect to these bases, the matrix of T is M(T ) = (aij )1≤i≤m,1≤j≤n . This means that,
for all j ∈ {1, 2, . . . , n},

m
X
T vj = akj wk .
k=1
The vector v ∈ V can be written uniquely as a linear combination of the basis vectors as
v = b1 v1 + · · · + bn vn .
Hence,
T v = b1 T v1 + · · · + bn T vn
Xm m
X
= b1 ak1 wk + · · · + bn akn wk
k=1 k=1
m
X
= (ak1 b1 + · · · + akn bn )wk .
k=1
This shows that M(T v) is the m × 1 matrix

 
a11 b1 + · · · + a1n bn
 .. 
M(T v) = 
 . .

am1 b1 + · · · + amn bn
It is not hard to check, using the formula for matrix multiplication, that M(T )M(v) gives
the same result.
Example 6.6.9. Take the linear map S from Example 6.6.4 with basis ((1, 2), (0, 1)) of R2 .
To determine the action on the vector v = (1, 4) ∈ R2 , note that v = (1, 4) = 1(1, 2) + 2(0, 1).
Hence, " #" # " #
2 1 1 4
M(Sv) = M(S)M(v) = = .
−3 −2 2 −7
This means that
Sv = 4(1, 2) − 7(0, 1) = (4, 1),
which is indeed true.

6.7 Invertibility
Definition 6.7.1. A linear map T : V → W is called invertible if there exists a linear map
S : W → V such that
T S = IW and ST = IV ,
where IV : V → V is the identity map on V and IW : W → W is the identity map on W .

We say that S is an inverse of T .
Note that if the linear map T is invertible, then the inverse is unique. Suppose S and R
are inverses of T . Then
ST = IV = RT,
T S = IW = T R.
Hence,
S = S(T R) = (ST )R = R.
We denote the unique inverse of an invertible linear map T by T −1 .
Proposition 6.7.2. A linear map T ∈ L(V, W ) is invertible if and only if T is injective and
surjective.
Proof.
(“=⇒”) Suppose T is invertible.
To show that T is injective, suppose that u, v ∈ V are such that T u = T v. Apply the
inverse T −1 of T to obtain T −1 T u = T −1 T v so that u = v. Hence T is injective.
To show that T is surjective, we need to show that, for every w ∈ W , there is a v ∈ V
such that T v = w. Take v = T −1 w ∈ V . Then T (T −1w) = w. Hence T is surjective.
(“⇐=”) Suppose that T is injective and surjective. We need to show that T is invertible.
We define a map S ∈ L(W, V ) as follows. Since T is surjective, we know that, for every
w ∈ W , there exists a v ∈ V such that T v = w. Moreover, since T is injective, this v is
uniquely determined. Hence, define Sw = v.
We claim that S is the inverse of T . Note that, for all w ∈ W , we have T Sw = T v = w
so that T S = IW . Similarly, for all v ∈ V , we have ST v = Sw = v so that ST = IV .
6.7. INVERTIBILITY 77
It remains to show that S is a linear map. For all w1 , w2 ∈ W , we have
T (Sw1 + Sw2 ) = T Sw1 + T Sw2 = w1 + w2 ,
and so Sw1 + Sw2 is the unique vector v in V such that T v = w1 + w2 = w. Hence,
Sw1 + Sw2 = v = Sw = S(w1 + w2 ).
The proof that S(aw) = aSw is similar. For w ∈ W and a ∈ F, we have
T (aSw) = aT (Sw) = aw
so that aSw is the unique vector in V that maps to aw. Hence, S(aw) = aSw.
Example 6.7.3. The linear map T (x, y) = (x − 2y, 3x + y) is both injective, since null (T ) =
{0}, and surjective, since range (T ) = R2 . Hence, T is invertible by Proposition 6.7.2.
Definition 6.7.4. Two vector spaces V and W are called isomorphic if there exists an
invertible linear map T ∈ L(V, W ).
Theorem 6.7.5. Two finite-dimensional vector spaces V and W over F are isomorphic if
and only if dim(V ) = dim(W ).
Proof.
(“=⇒”) Suppose V and W are isomorphic. Then there exists an invertible linear map
T ∈ L(V, W ). Since T is invertible, it is injective and surjective, and so null (T ) = {0} and
range (T ) = W . Using the Dimension Formula, this implies that
dim(V ) = dim(null (T )) + dim(range (T )) = dim(W ).
(“⇐=”) Suppose that dim(V ) = dim(W ). Let (v1 , . . . , vn ) be a basis of V and (w1 , . . . , wn )
be a basis of W . Define the linear map T : V → W as
T (a1 v1 + · · · + an vn ) = a1 w1 + · · · + an wn .
Since the scalars a1 , . . . , an ∈ F are arbitrary and (w1 , . . . , wn ) spans W , this means that
range (T ) = W and T is surjective. Also, since (w1 , . . . , wn ) is linearly independent, T is
injective (since a1 w1 + · · · + an wn = 0 implies that all a1 = · · · = an = 0 and hence only the

zero vector is mapped to zero). It follows that T is both injective and surjective; hence, by
Proposition 6.7.2, T is invertible. Therefore, V and W are isomorphic.
We close this chapter by considering the case of linear maps having equal domain and
codomain. As in Definition 6.1.1, a linear map T ∈ L(V, V ) is called a linear operators
on V . As the following remarkable theorem shows, the notions of injectivity, surjectivity,
and invertibility of a linear operator T are the same — as long as V is finite-dimensional. A
similar result does not hold for infinite-dimensional vector spaces. For example, the set of all
polynomials F[z] is an infinite-dimensional vector space, and we saw that the differentiation
map on F[z] is surjective but not injective.
Theorem 6.7.6. Let V be a finite-dimensional vector space and T : V → V be a linear map.

Then the following are equivalent:
1. T is invertible.
2. T is injective.
3. T is surjective.
Proof. By Proposition 6.7.2, Part 1 implies Part 2.

Next we show that Part 2 implies Part 3. If T is injective, then we know that null (T ) =
{0}. Hence, by the Dimension Formula, we have
dim(range (T )) = dim(V ) − dim(null (T )) = dim(V ).
Since range (T ) ⊂ V is a subspace of V , this implies that range (T ) = V , and so T is

surjective.
Finally, we show that Part 3 implies Part 1. Since T is surjective by assumption, we have
range (T ) = V . Thus, by again using the Dimension Formula,
dim(null (T )) = dim(V ) − dim(range (T )) = 0,
and so null (T ) = {0}, from which T is injective. By Proposition 6.7.2, an injective and
surjective linear map is invertible.

1. Define the map T : R2 → R2 by T (x, y) = (x + y, x).
(a) Show that T is linear.

(b) Show that T is surjective.
(c) Find dim (null (T )).
(d) Find the matrix for T with respect to the canonical basis of R2 .
(e) Show that the map F : R2 → R2 given by F (x, y) = (x + y, x + 1) is not linear.
2. Let T ∈ L(R2 ) be defined by

! ! !
x y x
T = , for all ∈ R2 .
y −x y
(a) Show that T is surjective.

(b) Find dim (null (T )).
(c) Find the matrix for T with respect to the canonical basis of R2 .
(d) Show that the map F : R2 → R2 given by F (x, y) = (x + y, x + 1) is not linear.
3. Consider the complex vector spaces C2 and C3 with their canonical bases, and define
S ∈ L(C3 , C2 ) be the linear map defined by S(v) = Av, ∀v ∈ C3 , where A is the matrix
by the matrix !
i 1 1
A = M(S) = .
2i −1 −1
Find a basis for null(S).
4. Give an example of a function f : R2 → R having the property that
∀ a ∈ R, ∀ v ∈ R, f (av) = af (v)
but such that f is not a linear map.

5. Show that the linear map T : F4 → F2 is surjective if
null(T ) = {(x1 , x2 , x3 , x4 ) ∈ F4 | x1 = 5x2 , x3 = 7x4 }.
6. Show that no linear map T : F5 → F2 can have as its null space the set
{(x1 , x2 , x3 , x4 , x5 ) ∈ F5 | x1 = 3x2 , x3 = x4 = x5 }.
7. Describe the set of solutions x = (x1 , x2 , x3 ) ∈ R3 of the system of equations


x1 − x2 + x3 = 0 

x1 + 2x2 + x3 = 0 .


2x1 + x2 + 2x3 = 0
1. Let V and W be vector spaces over F with V finite-dimensional, and let U be any
subspace of V . Given a linear map S ∈ L(U, W ), prove that there exists a linear map
T ∈ L(V, W ) such that, for every u ∈ U, S(u) = T (u).
2. Let V and W be vector spaces over F, and suppose that T ∈ L(V, W ) is injec-
tive. Given a linearly independent list (v1 , . . . , vn ) of vectors in V , prove that the
list (T (v1 ), . . . , T (vn )) is linearly independent in W .
3. Let U, V , and W be vector spaces over F, and suppose that the linear maps S ∈ L(U, V )
and T ∈ L(V, W ) are both injective. Prove that the composition map T ◦ S is injective.
4. Let V and W be vector spaces over F, and suppose that T ∈ L(V, W ) is surjective.
Given a spanning list (v1 , . . . , vn ) for V , prove that span(T (v1 ), . . . , T (vn )) = W .
5. Let V and W be vector spaces over F with V finite-dimensional. Given T ∈ L(V, W ),

prove that there is a subspace U of V such that
U ∩ null(T ) = {0} and range(T ) = {T (u) | u ∈ U}.

6. Let V be a vector spaces over F, and suppose that there is a linear map T ∈ L(V, V )
such that both null(T ) and range(T ) are finite-dimensional subspaces of V . Prove that
V must also be finite-dimensional.
7. Let U, V , and W be finite-dimensional vector spaces over F with S ∈ L(U, V ) and

T ∈ L(V, W ). Prove that
dim(null(T ◦ S)) ≤ dim(null(T )) + dim(null(S)).
8. Let V be a finite-dimensional vector space over F with S, T ∈ L(V, V ). Prove that

T ◦ S is invertible if and only if both S and T are invertible.
9. Let V be a finite-dimensional vector space over F with S, T ∈ L(V, V ), and denote by

I the identity map on V . Prove that T ◦ S = I if and only if S ◦ T = I.
Chapter 7
Eigenvalues and Eigenvectors
In this chapter we study linear operators T : V → V on a finite-dimensional vector space

V . We are interested in understanding when there is a basis B for V such that that matrix
M(T ) of T with respect to B has a particularly nice form. In particular, we would like M(T )
to be either upper triangular or diagonal. This quest leads us to the notions of eigenvalues
and eigenvectors of a linear operator, which is one of the most important concepts in Linear
Algebra and beyond. For example, quantum mechanics is largely based upon the study of
eigenvalues and eigenvectors of operators on infinite-dimensional vector spaces.
7.1 Invariant subspaces

To begin our study, we will look at subspaces U of V that have special properties under an
operator T ∈ L(V, V ).
Definition 7.1.1. Let V be a finite-dimensional vector space over F with dim(V ) ≥ 1, and
let T ∈ L(V, V ) be an operator in V . Then a subspace U ⊂ V is called an invariant
subspace under T if
Tu ∈ U for all u ∈ U.
That is, U is invariant under T if the image of every vector in U under T remains within U.
We denote this as T U = {T u | u ∈ U} ⊂ U.
Example 7.1.2. The subspaces null (T ) and range (T ) are invariant subspaces under T . To
see this, let u ∈ null (T ). This means that T u = 0. But, since 0 ∈ null (T ), this implies that
82
7.2. EIGENVALUES 83
T u = 0 ∈ null (T ). Similarly, let u ∈ range (T ). Since T v ∈ range (T ) for all v ∈ V , we

certainly also have that T u ∈ range (T ).
Example 7.1.3. Take the linear operator T : R3 → R3 corresponding to the matrix

 
1 2 0
 
 1 1 0
0 0 2
with respect to the basis (e1 , e2 , e3 ). Then span(e1 , e2 ) and span(e3 ) are both invariant
subspaces under T .
An important special case is the case of Definition 7.1.1 involves one-dimensional invariant
subspaces under an operator T ∈ L(V, V ). If dim(U) = 1, then there exists a nonzero vector
u ∈ V such that
U = {au | a ∈ F}.
In this case, we must have

T u = λu for some λ ∈ F.
This motivates the definitions of eigenvectors and eigenvalues of a linear operator, as given
in the next section.
7.2 Eigenvalues
Definition 7.2.1. Let T ∈ L(V, V ). Then λ ∈ F is an eigenvalue of T if there exists a
nonzero vector u ∈ V such that
T u = λu.
The vector u is called an eigenvector of T corresponding to the eigenvalue λ.
Finding the eigenvalues and eigenvectors of a linear operator is one of the most important
problems in Linear Algebra. We will see later that this so-called “eigen-information” has
many uses and applications. (As an example, quantum mechanics is based upon understand-
ing the eigenvalues and eigenvectors of operators on specifically defined vector spaces. These
vector spaces are often infinite-dimensional, though, and so we do not consider them further
in these notes.)
84 CHAPTER 7. EIGENVALUES AND EIGENVECTORS
Example 7.2.2.
1. Let T be the zero map defined by T (v) = 0 for all v ∈ V . Then every vector u 6= 0 is
an eigenvector of T with eigenvalue 0.
2. Let I be the identity map defined by I(v) = v for all v ∈ V . Then every vector u 6= 0
is an eigenvector of T with eigenvalue 1.
3. The projection map P : R3 → R3 defined by P (x, y, z) = (x, y, 0) has eigenvalues 0

and 1. The vector (0, 0, 1) is an eigenvector with eigenvalue 0, and both (1, 0, 0) and
(0, 1, 0) are eigenvectors with eigenvalue 1.
4. Take the operator R : F2 → F2 defined by R(x, y) = (−y, x). When F = R, R can be

interpreted as counterclockwise rotation by 900 . From this interpretation, it is clear
that no non-zero vector in R2 mapped to a scalar multiple of itself. Hence, for F = R,
the operator R has no eigenvalues.
For F = C, though, the situation is significantly different! In this case, λ ∈ C is an
eigenvalue of R if
R(x, y) = (−y, x) = λ(x, y)
so that y = −λx and x = λy. This implies that y = −λ2 y, i.e., that λ2 = −1.
The solutions are hence λ = ±i. One can check that (1, −i) is an eigenvector with
eigenvalue i and that (1, i) is an eigenvector with eigenvalue −i.
Eigenspaces are important examples of invariant subspaces. Let T ∈ L(V, V ), and let
λ ∈ F be an eigenvalue of T . Then
Vλ = {v ∈ V | T v = λv}
is called an eigenspace of T . Equivalently,
Vλ = null (T − λI).
Note that Vλ 6= {0} since λ is an eigenvalue if and only if there exists a nonzero vector u ∈ V
such that T u = λu. We can reformulate this as follows:
• λ ∈ F is an eigenvalue of T if and only if the operator T − λI is not injective.

7.2. EIGENVALUES 85
Since the notion of injectivity, surjectivity, and invertibility are equivalent for operators on
a finite-dimensional vector space, we can equivalently say either of the following:
• λ ∈ F is an eigenvalue of T if and only if the operator T − λI is not surjective.
• λ ∈ F is an eigenvalue of T if and only if the operator T − λI is not invertible.
We close this section with two fundamental facts about eigenvalues and eigenvectors.
Theorem 7.2.3. Let T ∈ L(V, V ), and let λ1 , . . . , λm ∈ F be m distinct eigenvalues of T with

corresponding nonzero eigenvectors v1 , . . . , vm . Then (v1 , . . . , vm ) is linearly independent.
Proof. Suppose that (v1 , . . . , vm ) is linearly dependent. Then, by the Linear Dependence
Lemma, there exists an index k ∈ {2, . . . , m} such that
vk ∈ span(v1 , . . . , vk−1 )
and such that (v1 , . . . , vk−1 ) is linearly independent. This means that there exist scalars
a1 , . . . , ak−1 ∈ F such that
vk = a1 v1 + · · · + ak−1 vk−1 . (7.1)
Applying T to both sides yields, using the fact that vj is an eigenvector with eigenvalue λj ,
λk vk = a1 λ1 v1 + · · · + ak−1 λk−1 vk−1 .
Subtracting λk times Equation (7.1) from this, we obtain
0 = (λk − λ1 )a1 v1 + · · · + (λk − λk−1 )ak−1 vk−1 .
Since (v1 , . . . , vk−1 ) is linearly independent, we must have (λk − λj )aj = 0 for all j =
1, 2, . . . , k − 1. By assumption, all eigenvalues are distinct, so λk − λj 6= 0, which im-
plies that aj = 0 for all j = 1, 2, . . . , k − 1. But then, by Equation (7.1), vk = 0, which
contradicts the assumption that all eigenvectors are nonzero. Hence (v1 , . . . , vm ) is linearly
independent.
Corollary 7.2.4. Any operator T ∈ L(V, V ) has at most dim(V ) distinct eigenvalues.
Proof. Let λ1 , . . . , λm be distinct eigenvalues of T , and let v1 , . . . , vm be corresponding

nonzero eigenvectors. By Theorem 7.2.3, the list (v1 , . . . , vm ) is linearly independent. Hence
m ≤ dim(V ).
7.3 Diagonal matrices

Note that if T has n = dim(V ) distinct eigenvalues, then there exists a basis (v1 , . . . , vn ) of
V such that
T vj = λj vj , for all j = 1, 2, . . . , n.
Then any v ∈ V can be written as a linear combination v = a1 v1 + · · · + an vn of v1 , . . . , vn .

Applying T to this, we obtain
T v = λ1 a1 v1 + · · · + λn an vn .
Hence the vector  

a1
.
M(v) = 
.
.
an
is mapped to  
λ1 a1
 . 
M(T v) =  . 
 . .
λn an
This means that the matrix M(T ) for T with respect to the basis of eigenvectors (v1 , . . . , vn )
is diagonal, and so we call T diagonalizable:
 
λ1 0
 .. 
M(T ) = 
 . .

0 λn
We summarize the results of the above discussion in the following Proposition.
Proposition 7.3.1. If T ∈ L(V, V ) has dim(V ) distinct eigenvalues, then M(T ) is diagonal
with respect to some basis of V . Moreover, V has a basis consisting of eigenvectors of T .
7.4. EXISTENCE OF EIGENVALUES 87
7.4 Existence of eigenvalues

In what follows, we want to study the question of when eigenvalues exist for a given operator
T . To answer this question, we will use polynomials p(z) ∈ F[z] evaluated on operators
T ∈ L(V, V ) (or, equivalently, on square matrices A ∈ Fn×n ). More explicitly, given a
polynomial
p(z) = a0 + a1 z + · · · + ak z k
we can associate the operator
p(T ) = a0 IV + a1 T + · · · + ak T k .
Note that, for p(z), q(z) ∈ F[z], we have
(pq)(T ) = p(T )q(T ) = q(T )p(T ).
The results of this section will be for complex vector spaces. This is because the proof of
the existence of eigenvalues relies on the Fundamental Theorem of Algebra from Chapter 3,
which makes a statement about the existence of zeroes of polynomials over C.
Theorem 7.4.1. Let V 6= {0} be a finite-dimensional vector space over C, and let T ∈
L(V, V ). Then T has at least one eigenvalue.
Proof. Let v ∈ V with v 6= 0, and consider the list of vectors
(v, T v, T 2v, . . . , T n v),
where n = dim(V ). Since the list contains n + 1 vectors, it must be linearly dependent.
Hence, there exist scalars a0 , a1 , . . . , an ∈ C, not all zero, such that
0 = a0 v + a1 T v + a2 T 2 v + · · · + an T n v.
Let m be largest index for which am 6= 0. Since v 6= 0, we must have m > 0 (but possibly
m = n). Consider the polynomial
p(z) = a0 + a1 z + · · · + am z m .
By Theorem 3.2.2 (3) it can be factored as
p(z) = c(z − λ1 ) · · · (z − λm ),
where c, λ1 , . . . , λm ∈ C and c 6= 0.
Therefore,
0 = a0 v + a1 T v + a2 T 2 v + · · · + an T n v = p(T )v
= c(T − λ1 I)(T − λ2 I) · · · (T − λm I)v,
and so at least one of the factors T − λj I must be noninjective. In other words, this λj is
an eigenvalue of T .
Note that the proof of Theorem 7.4.1 only uses basic concepts about linear maps, which
is the same approach as in a popular textbook called Linear Algebra Done Right by Sheldon
Axler. Many other textbooks rely on significantly more difficult proofs using concepts like the
determinant and characteristic polynomial of a matrix. At the same time, it is often prefer-
able to use the characteristic polynomial of a matrix in order to compute eigen-information
of an operator; we discuss this approach in Chapter 8.
Note also that Theorem 7.4.1 does not hold for real vector spaces. E.g., as we saw in
Example 7.2.2, the rotation operator R on R2 has no eigenvalues.
7.5 Upper triangular matrices
As before, let V be a complex vector space.

Let T ∈ L(V, V ) and (v1 , . . . , vn ) be a basis for V . Recall that we can associate a matrix
M(T ) ∈ Cn×n to the operator T . By Theorem 7.4.1, we know that T has at least one
eigenvalue, say λ ∈ C. Let v1 6= 0 be an eigenvector corresponding to λ. By the Basis
Extension Theorem, we can extend the list (v1 ) to a basis of V . Since T v1 = λv1 , the first
7.5. UPPER TRIANGULAR MATRICES 89
column of M(T ) with respect to this basis is

 
λ
 
0
..
.
.
0
What we will show next is that we can find a basis of V such that the matrix M(T ) is upper
triangular.
Definition 7.5.1. A matrix A = (aij ) ∈ Fn×n is called upper triangular if aij = 0 for
i > j.
Schematically, an upper triangular matrix has the form

 
∗ ∗
 . 
 ..  ,
 
0 ∗
where the entries ∗ can be anything and every entry below the main diagonal is zero.
Some of the reasons why upper triangular matrices are so fantastic are that
1. the eigenvalues are on the diagonal (as we will see later);
2. it is easy to solve the corresponding system of linear equations by back substitution

(as discussed in Section 12.3).
The next proposition tells us what upper triangularity means in terms of linear operators
and invariant subspaces.
Proposition 7.5.2. Suppose T ∈ L(V, V ) and that (v1 , . . . , vn ) is a basis of V . Then the
following statements are equivalent:
1. the matrix M(T ) with respect to the basis (v1 , . . . , vn ) is upper triangular;
2. T vk ∈ span(v1 , . . . , vk ) for each k = 1, 2, . . . , n;
3. span(v1 , . . . , vk ) is invariant under T for each k = 1, 2, . . . , n.

Proof. The equivalence of Condition 1 and Condition 2 follows easily from the definition
since Condition 2 implies that the matrix elements below the diagonal are zero.
Obviously, Condition 3 implies Condition 2. To show that Condition 2 implies Condi-
tion 3, note that any vector v ∈ span(v1 , . . . , vk ) can be written as v = a1 v1 + · · · + ak vk .
Applying T , we obtain
T v = a1 T v1 + · · · + ak T vk ∈ span(v1 , . . . , vk )
since, by Condition 2, each T vj ∈ span(v1 , . . . , vj ) ⊂ span(v1 , . . . , vk ) for j = 1, 2, . . . , k and

since the span is a subspace of V .
The next theorem shows that complex vector spaces indeed have some basis for which
the matrix of a given operator is upper triangular.
Theorem 7.5.3. Let V be a finite-dimensional vector space over C and T ∈ L(V, V ). Then
there exists a basis B for V such that M(T ) is upper triangular with respect to B.
Proof. We proceed by induction on dim(V ). If dim(V ) = 1, then there is nothing to prove.

Hence, assume that dim(V ) = n > 1 and that we have proven the result of the theorem
for all T ∈ L(W, W ), where W is a complex vector space with dim(W ) ≤ n − 1. By
Theorem 7.4.1, T has at least one eigenvalue λ. Define
U = range (T − λI),
and note that
1. dim(U) < dim(V ) = n since λ is an eigenvalue of T and hence T − λI is not surjective;
2. U is an invariant subspace of T since, for all u ∈ U, we have
T u = (T − λI)u + λu,
which implies that T u ∈ U since (T − λI)u ∈ range (T − λI) = U and λu ∈ U.
Therefore, we may consider the operator S = T |U , which is the operator obtained by re-
stricting T to the subspace U. By the induction hypothesis, there exists a basis (u1, . . . , um )
7.5. UPPER TRIANGULAR MATRICES 91
of U with m ≤ n − 1 such that M(S) is upper triangular with respect to (u1 , . . . , um ). This
means that
T uj = Suj ∈ span(u1 , . . . , uj ), for all j = 1, 2, . . . , m.
Extend this to a basis (u1 , . . . , um , v1 , . . . , vk ) of V . Then
T vj = (T − λI)vj + λvj , for all j = 1, 2, . . . , k.
Since (T − λI)vj ∈ range (T − λI) = U = span(u1 , . . . , um ), we have that
T vj ∈ span(u1 , . . . , um, v1 , . . . , vj ), for all j = 1, 2, . . . , k.
Hence, T is upper triangular with respect to the basis (u1 , . . . , um , v1 , . . . , vk ).
The following are two very important facts about upper triangular matrices and their
associated operators.
Proposition 7.5.4. Suppose T ∈ L(V, V ) is a linear operator and that M(T ) is upper
triangular with respect to some basis of V . Then
1. T is invertible if and only if all entries on the diagonal of M(T ) are nonzero.
2. The eigenvalues of T are precisely the diagonal elements of M(T ).

Proof of Proposition 7.5.4, Part 1. Let (v1 , . . . , vn ) be a basis of V such that
 
λ1 ∗
 .. 
M(T ) = 
 . 

0 λn
is upper triangular. The claim is that T is invertible if and only if λk 6= 0 for all k =
1, 2, . . . , n. Equivalently, this can be reformulated as follows: T is not invertible if and only
if λk = 0 for at least one k ∈ {1, 2, . . . , n}.
Suppose λk = 0. We will show that this implies the non-invertibility of T . If k = 1, this
is obvious since then T v1 = 0, which implies that v1 ∈ null (T ) so that T is not injective and
hence not invertible. So assume that k > 1. Then
T vj ∈ span(v1 , . . . , vk−1 ), for all j ≤ k,

since T is upper triangular and λk = 0. Hence, we may define S = T |span(v1 ,...,vk ) to be the
restriction of T to the subspace span(v1 , . . . , vk ) so that
S : span(v1 , . . . , vk ) → span(v1 , . . . , vk−1 ).
The linear map S is not injective since the dimension of the domain is larger than the
dimension of its codomain, i.e.,
dim(span(v1 , . . . , vk )) = k > k − 1 = dim(span(v1 , . . . , vk−1 )).
Hence, there exists a vector 0 6= v ∈ span(v1 , . . . , vk ) such that Sv = T v = 0. This implies

that T is also not injective and therefore also not invertible.
Now suppose that T is not invertible. We need to show that at least one λk = 0. The
linear map T not being invertible implies that T is not injective. Hence, there exists a vector
0 6= v ∈ V such that T v = 0, and we can write
v = a1 v1 + · · · + ak vk
for some k, where ak 6= 0. Then
0 = T v = (a1 T v1 + · · · + ak−1 T vk−1 ) + ak T vk . (7.2)
Since T is upper triangular with respect to the basis (v1 , . . . , vn ), we know that a1 T v1 + · · · +
ak−1 T vk−1 ∈ span(v1 , . . . , vk−1). Hence, Equation (7.2) shows that T vk ∈ span(v1 , . . . , vk−1 ),
which implies that λk = 0.
Proof of Proposition 7.5.4, Part 2. Recall that λ ∈ F is an eigenvalue of T if and only if

the operator T − λI is not invertible. Let (v1 , . . . , vn ) be a basis such that M(T ) is upper
triangular. Then  
λ1 − λ ∗
 .. 
M(T − λI) =   . .

0 λn − λ
Hence, by Proposition 7.5.4(1), T − λI is not invertible if and only if λ = λk for some k.
7.6. DIAGONALIZATION OF 2 × 2 MATRICES AND APPLICATIONS 93
7.6 Diagonalization of 2 × 2 matrices and Applications

" #
a b
Let A = ∈ F2×2 , and recall that we can define a linear operator T ∈ L(F2 ) on F2 by
c d
" #
v1
setting T (v) = Av for each v = ∈ F2 .
v2
One method for finding the eigen-information of T is to analyze the solutions of the
matrix equation Av = λv for λ ∈ F and v ∈ F2 . In particular, using the definition of
eigenvector and eigenvalue, v is an eigenvector associated to the eigenvalue λ if and only if
Av = T (v) = λv.
A simpler method involves the equivalent matrix equation (A−λI)v = 0, where I denotes
the identity map on F2 . In particular, 0 6= v ∈ F2 is an eigenvector for T associated to the
eigenvalue λ ∈ F if and only if the system of linear equations
)
(a − λ)v1 + bv2 = 0
(7.3)
cv1 + (d − λ)v2 = 0
has a non-trivial solution. Moreover, System (7.3) has a non-trivial solution if and only if the
polynomial p(λ) = (a − λ)(d − λ) − bc evaluates to zero. (See see Proof-writing Exercise 12
on page 98.)
In other words, the eigenvalues for T are exactly the λ ∈ F for which p(λ) = 0, and
" #the
v1
eigenvectors for T associated to an eigenvalue λ are exactly the non-zero vectors v = ∈
v2
F2 that satisfy System (7.3).
" #
−2 −1
Example 7.6.1. Let A = . Then p(λ) = (−2 − λ)(2 − λ) − (−1)(5) = λ2 + 1,
5 2
which is equal to zero exactly when λ = ±i. Moreover, if λ = i, then the System (7.3)
becomes )
(−2 − i)v1 − v2 = 0
,
5v1 + (2 − i)v2 = 0
" #
v1
which is satisfied by any vector v = ∈ C2 such that v2 = (−2 − i)v1 . Similarly, if
v2
λ = −i, then the System (7.3) becomes

)
(−2 + i)v1 − v2 = 0
,
5v1 + (2 + i)v2 = 0
" #
v1
which is satisfied by any vector v = ∈ C2 such that v2 = (−2 + i)v1 .
v2
" #
−2 −1
It follows that, given A = , the linear operator on C2 defined by T (v) = Av
5 2
has eigenvalues λ = ±i, with associated eigenvectors as described above.
Example 7.6.2. Take the rotation Rθ : R2 → R2 by an angle θ ∈ [0, 2π) given by the matrix
" #
cos θ − sin θ
Rθ =
sin θ cos θ
Then we obtain the eigenvalues by solving the polynomial equation
p(λ) = (cos θ − λ)2 + sin2 θ

= λ2 − 2λ cos θ + 1 = 0,
where we have used the fact that sin2 θ + cos2 θ = 1. Solving for λ in C, we obtain
√ p
λ = cos θ ± cos2 θ − 1 = cos θ ± − sin2 θ = cos θ ± i sin θ = e±iθ .
We see that, as an operator over the real vector space R2 , the"operator

# Rθ only has eigenvalues
x1
when θ = 0 or θ = π. However, if we interpret the vector ∈ R2 as a complex number
x2
z = x1 + ix2 , then z is an eigenvector if Rθ : C → C maps z 7→ λz = e±iθ z. Moreover, from
Section 2.3.2, we know that multiplication by e±iθ corresponds to rotation by the angle ±θ.

1. Let T ∈ L(F2 , F2 ) be defined by
T (u, v) = (v, u)
for every u, v ∈ F. Compute the eigenvalues and associated eigenvectors for T .
2. Let T ∈ L(F3 , F3 ) be defined by
T (u, v, w) = (2v, 0, 5w)
for every u, v, w ∈ F. Compute the eigenvalues and associated eigenvectors for T .
3. Let n ∈ Z+ be a positive integer and T ∈ L(Fn , Fn ) be defined by
T (x1 , . . . , xn ) = (x1 + · · · + xn , . . . , x1 + · · · + xn )
for every x1 , . . . , xn ∈ F. Compute the eigenvalues and associated eigenvectors for T .
4. Find eigenvalues and associated eigenvectors for the linear operators on F2 defined by
each given 2 × 2 matrix.
" # " # " #
3 0 10 −9 0 3
(a) (b) (c)
8 −1 4 −2 4 0
" # " # " #
−2 −7 0 0 1 0
(d) (e) (f)
1 2 0 0 0 1
" #
a b
Hint: Use the fact that, given a matrix A = ∈ F2×2 , λ ∈ F is an eigenvalue
c d
for A if and only if (a − λ)(d − λ) − bc = 0.
5. For each matrix A below, find eigenvalues for the induced linear operator T on Fn
without performing any calculations. Then describe the eigenvectors v ∈ Fn associated
to each eigenvalue λ by looking at solutions to the matrix equation (A − λI)v = 0,
where I denotes the identity map on Fn .

   
" # − 13 0 0 0 1 3 7 11
 1
  1

−1 6  0 −3 0 0   3 8 
(a) , (b)  , (c)  0 2 
0 5  0 0 1 0   0 0 0 4 
   
1
0 0 0 2 0 0 0 2
6. For each matrix A below, describe the invariant subspaces for the induced linear op-
erator T on F2 that maps each v ∈ F2 to T (v) = Av.
" # " # " # " #
4 −1 0 1 2 3 1 0
(a) , (b) , (c) , (d)
2 1 −1 0 0 2 0 0
7. Let T ∈ L(R2 ) be defined by

! ! !
x y x
T = , for all ∈ R2 .
y x+y y
Define two real numbers λ+ and λ− as follows:

√ √
1+ 5 1− 5
λ+ = , λ− = .
2 2
(a) Find the matrix of T with respect to the canonical basis for R2 (both as the
domain and the codomain of T ; call this matrix A).
(b) Verify that λ+ and λ− are eigenvalues of T by showing that v+ and v− are eigen-
vectors, where ! !
1 1
v+ = , v− = .
λ+ λ−
(c) Show that (v+ , v− ) is a basis of R2 .
(d) Find the matrix of T with respect to the basis (v+ , v− ) for R2 (both as the domain
and the codomain of T ; call this matrix B).
1. Let V be a finite-dimensional vector space over F with T ∈ L(V, V ), and let U1 , . . . , Um
be subspaces of V that are invariant under T . Prove that U1 + · · · + Um must then also
be an invariant subspace of V under T .
2. Let V be a finite-dimensional vector space over F with T ∈ L(V, V ), and suppose that
U1 and U2 are subspaces of V that are invariant under T . Prove that U1 ∩ U2 is also
an invariant subspace of V under T .
3. Let V be a finite-dimensional vector space over F with T ∈ L(V, V ) invertible and

λ ∈ F \ {0}. Prove λ is an eigenvalue for T if and only if λ−1 is an eigenvalue for T −1 .
4. Let V be a finite-dimensional vector space over F, and suppose that T ∈ L(V, V ) has
the property that every v ∈ V is an eigenvector for T . Prove that T must then be a
scalar multiple of the identity function on V .
5. Let V be a finite-dimensional vector space over F, and let S, T ∈ L(V ) be linear

operators on V with S invertible. Given any polynomial p(z) ∈ F[z], prove that
p(S ◦ T ◦ S −1 ) = S ◦ p(T ) ◦ S −1 .
6. Let V be a finite-dimensional vector space over C, T ∈ L(V ) be a linear operator on

V , and p(z) ∈ C[z] be a polynomial. Prove that λ ∈ C is an eigenvalue of the linear
operator p(T ) ∈ L(V ) if and only if T has an eigenvalue µ ∈ C such that p(µ) = λ.
7. Let V be a finite-dimensional vector space over C with T ∈ L(V ) a linear operator

on V . Prove that, for each k = 1, . . . , dim(V ), there is an invariant subspace Uk of V
under T such that dim(Uk ) = k.
Claim. Let V be a finite-dimensional vector space over F, and let T ∈ L(V ) be a linear
operator on V . If the matrix for T with respect to some basis on V has all zeros on
the diagonal, then T is not invertible.

Claim. Let V be a finite-dimensional vector space over F, and let T ∈ L(V ) be a linear
operator on V . If the matrix for T with respect to some basis on V has all non-zero
elements on the diagonal, then T is invertible.
10. Let V be a finite-dimensional vector space over F, and let S, T ∈ L(V ) be linear
operators on V . Suppose that T has dim(V ) distinct eigenvalues and that, given any
eigenvector v ∈ V for T associated to some eigenvalue λ ∈ F, v is also an eigenvector
for S associated to some (possibly distinct) eigenvalue µ ∈ F. Prove that T ◦ S = S ◦ T .
11. Let V be a finite-dimensional vector space over F, and suppose that the linear operator
P ∈ L(V ) has the property that P 2 = P . Prove that V = null(P ) ⊕ range(P ).
12. (a) Let a, b, c, d ∈ F and consider the system of equations given by
ax1 + bx2 = 0 (7.4)

cx1 + dx2 = 0. (7.5)
Note that x1 = x2 = 0 is a solution for any choice of a, b, c, and d. Prove that

this system of equations has a non-trivial solution if and only if ad − bc = 0.
" #
a b
(b) Let A = ∈ F2×2 , and recall that we can define a linear operator T ∈ L(F2 )
c d
" #
v1
on F2 by setting T (v) = Av for each v = ∈ F2 .
v2
Show that the eigenvalues for T are exactly the λ ∈ F for which p(λ) = 0, where
p(z) = (a − z)(d − z) − bc.
Hint: Write the eigenvalue equation Av = λv as (A − λI)v = 0 and use the first
part.
Chapter 8
Permutations and the Determinant of

a Square Matrix
There are many operations that can be applied to a square matrix. This chapter is devoted
to one particularly important operation called the determinant. In effect, the determinant
can be thought of as a single number that is used to check for many of the different properties
that a matrix might possess.
In order to define the determinant operation, we will first need to define permutations.
8.1 Permutations
Permutations appear in many different mathematical concepts, and so we give a general
introduction to them in this section.
8.1.1 Definition of permutations

Given a positive integer n ∈ Z+ , a permutation of an (ordered) list of n distinct objects is
any reordering of this list. When describing the reorderings themselves, though, the nature
of the objects involved is more or less irrelevant. E.g., we can imagine interchanging the
second and third items in a list of five distinct objects — no matter what those items are —
and this defines a particular permutation that can be applied to any list of five objects.
Since the nature of the objects being rearranged (i.e., permuted) is immaterial, it is
common to use the integers 1, 2, . . . , n as the standard list of n objects. Alternatively, one
99
100 CHAPTER 8. PERMUTATIONS AND DETERMINANTS
can also think of these integers as labels for the items in any list of n distinct elements. This
gives rise to the following definition.
Definition 8.1.1. A permutation π of n elements is a one-to-one and onto function having

the set {1, 2, . . . , n} as both its domain and codomain.
In other words, a permutation is a function π : {1, 2, . . . , n} −→ {1, 2, . . . , n} such that,

for every integer i ∈ {1, . . . , n}, there exists exactly one integer j ∈ {1, . . . , n} for which
π(j) = i. We will usually denote permutations by Greek letters such as π (pi), σ (sigma),
and τ (tau). The set of all permutations of n elements is denoted by Sn and is typically
referred to as the symmetric group of degree n. (In particular, the set Sn forms a group
under function composition as discussed in Section 8.1.2.)
Given a permutation π ∈ Sn , there are several common notations used for specifying
how π permutes the integers 1, 2, . . . , n. The important thing to keep in mind when working
with these different notations is that π is a function defined on the finite set {1, 2, . . . , n},
with notation being used as a convenient short-hand for keeping track of how π permutes
the elements in this set.
Definition 8.1.2. Given a permutation π ∈ Sn , denote πi = π(i) for each i ∈ {1, . . . , n}.
Then the two-line notation for π is given by the 2 × n matrix
!
1 2 ··· n
π= .
π1 π2 · · · πn
In other words, given a permutation π ∈ Sn and an integer i ∈ {1, . . . , n}, we are denoting
the image of i under π by πi instead of using the more conventional function notation π(i).
Then, in order to specify the image of each integer i ∈ {1, . . . , n} under π, we list these
images in a two-line array as shown above. (One can also use so-called one-line notation
for π, which is given by simply ignoring the top row and writing π = π1 π2 · · · πn .)
It is important to note that, although we represent permutations as 2 × n matrices, you
should not think of permutations as linear transformations from an n-dimensional vector
space into a two-dimensional vector space. Moreover, the composition operation on permu-
tation that we describe in Section 8.1.2 below does not correspond to matrix multiplication.
The use of matrix notation in denoting permutations is merely a matter of convenience.
8.1. PERMUTATIONS 101
Example 8.1.3. Suppose that we have a set of five distinct objects and that we wish to
describe the permutation that places the first item into the second position, the second item
into the fifth position, the third item into the first position, the fourth item into the third
position, and the fifth item into the fourth position. Then, using the notation developed
above, we have the permutation π ∈ S5 such that
π1 = π(1) = 3, π2 = π(2) = 1, π3 = π(3) = 4, π4 = π(4) = 5, π5 = π(5) = 2.
In two-line notation, we would write π as

!
1 2 3 4 5
π= .
3 1 4 5 2
It is relatively straightforward to find the number of permutations of n elements, i.e., to

determine cardinality of the set Sn . To construct an arbitrary permutation of n elements,
we can proceed as follows: First, choose an integer i ∈ {1, . . . , n} to put into the first
position. Clearly, we have exactly n possible choices. Next, choose the element to go in the
second position. Since we have already chosen one element from the set {1, . . . , n}, there are
now exactly n − 1 remaining choices. Proceeding in this way, we have n − 2 choices when
choosing the third element from the set {1, . . . , n}, then n − 3 choices when choosing the
fourth element, and so on until we are left with exactly one choice for the nth element. This
proves the following theorem.
Theorem 8.1.4. The number of elements in the symmetric group Sn is given by
|Sn | = n · (n − 1) · (n − 2) · · · · · 3 · 2 · 1 = n!
We conclude this section with several examples, including a complete description of the
one permutation in S1 , the two permutations in S2 , and the six permutations in S3 . For
your own practice, you should (patiently) attempt to list the 4! = 24 permutations in S4 .
Example 8.1.5.
1. Given any positive integer n ∈ Z+ , the identity function id : {1, . . . , n} −→ {1, . . . , n}

given by id(i) = i, ∀ i ∈ {1, . . . , n}, is a permutation in Sn . This function can be
thought of as the trivial reordering that does not change the order at all, and so we
call it the trivial or identity permutation.
2. If n = 1, then, by Theorem 8.1.4, |Sn | = 1! = 1. Thus, S1 contains only the identity

permutation.
3. If n = 2, then, by Theorem 8.1.4, |Sn | = 2! = 2 · 1 = 2. Thus, there is only one

non-trivial permutation π in S2 , namely the transformation interchanging the first and
the second elements in a list. As a function, π(1) = 2 and π(2) = 1, and, in two-line
notation, ! !
1 2 1 2
π= = .
π1 π2 2 1
4. If n = 3, then, by Theorem 8.1.4, |Sn | = 3! = 3 · 2 · 1 = 6. Thus, there are five

non-trivial permutation in S3 . Using two-line notation, we have that
( ! ! ! ! ! !)
1 2 3 1 2 3 1 2 3 1 2 3 1 2 3 1 2 3
S3 = , , , , ,
1 2 3 1 3 2 2 1 3 2 3 1 3 1 2 3 2 1
Keep in mind the fact that each element in S3 is simultaneously both a function and
a reordering operation. E.g., the permutation
! !
1 2 3 1 2 3
π= =
π1 π2 π3 2 3 1
can be read as defining the reordering that, with respect to the original list, places
the second element in the first position, the third element in the second position, and
the first element in the third position. This permutation could equally well have been
identified by describing its action on the (ordered) list of letters a, b, c. In other words,
! !
1 2 3 a b c
= ,
2 3 1 b c a
regardless of what the letters a, b, c might happen to represent. In particular, if we set

a = 2, b = 1, and c = 3, then the above equally becomes

! !
1 2 3 2 1 3
= .
2 3 1 1 3 2
8.1.2 Composition of permutations
Let n ∈ Z+ be a positive integer and π, σ ∈ Sn be permutations. Then, since π and σ

are both functions from the set {1, . . . , n} to itself, we can compose them to obtain a new
function π ◦ σ (read as “pi after sigma”) that takes on the values
(π ◦ σ)(1) = π(σ(1)), (π ◦ σ)(1) = π(σ(2)), ... (π ◦ σ)(n) = π(σ(n)).
In two-line notation, we can write π ◦ σ as

! ! !
1 2 ··· n 1 2 ··· n 1 2 ··· n
or or .
π(σ(1)) π(σ(2)) · · · π(σ(n)) πσ(1) πσ(2) · · · πσ(n) πσ1 πσ2 · · · πσn
Example 8.1.6. From S3 , suppose that we have the permutations π and σ given by
π(1) = 2, π(2) = 3, π(3) = 1 and σ(1) = 1, σ(2) = 3, σ(3) = 2.
Then note that

(π ◦ σ)(1) = π(σ(1)) = π(1) = 2,
(π ◦ σ)(2) = π(σ(2)) = π(3) = 1,
(π ◦ σ)(3) = π(σ(3)) = π(2) = 3.
In other words,
! ! ! !
1 2 3 1 2 3 1 2 3 1 2 3
◦ = = .
2 3 1 1 3 2 π(1) π(3) π(2) 2 1 3
Similar computations (which you should check for your own practice) yield compositions
such as ! ! ! !
1 2 3 1 2 3 1 2 3 1 2 3
◦ = = ,
1 3 2 2 3 1 σ(2) σ(3) σ(1) 3 2 1
! ! ! !
1 2 3 1 2 3 1 2 3 1 2 3
◦ = = ,
2 3 1 1 2 3 σ(1) σ(2) σ(3) 2 3 1
and ! ! ! !
1 2 3 1 2 3 1 2 3 1 2 3
◦ = = .
1 2 3 2 3 1 id(2) id(3) id(1) 2 3 1
In particular, note that the result of each composition above is a permutation, that compo-
sition is not a commutative operation, and that composition with id leaves a permutation
unchanged. Moreover, since each permutation π is a bijection, one can always construct an
inverse permutation π −1 such that π ◦ π −1 = id. E.g.,
! ! ! !
1 2 3 1 2 3 1 2 3 1 2 3
◦ = = .
2 3 1 3 1 2 π(3) π(1) π(2) 1 2 3
We summarize the basic properties of composition on the symmetric group in the follow-
ing theorem.
Theorem 8.1.7. Let n ∈ Z+ be a positive integer. Then the set Sn has the following
properties.
1. Given any two permutations π, σ ∈ Sn , the composition π ◦ σ ∈ Sn .
2. (Associativity of Composition) Given any three permutations π, σ, τ ∈ Sn ,
(π ◦ σ) ◦ τ = π ◦ (σ ◦ τ ).
3. (Identity Element for Composition) Given any permutation π ∈ Sn ,
π ◦ id = id ◦ π = π.
4. (Inverse Elements for Composition) Given any permutation π ∈ Sn , there exists a

unique permutation π −1 ∈ Sn such that
π ◦ π −1 = π −1 ◦ π = id.
In other words, the set Sn forms a group under composition.
Note that the composition of permutations is not commutative in general. In particular,

for n ≥ 3, it is easy to find permutations π and σ such that π ◦ σ 6= σ ◦ π.
8.1.3 Inversions and the sign of a permutation

Let n ∈ Z+ be a positive integer. Then, given a permutation π ∈ Sn , it is natural to ask how
“out of order” π is in comparison to the identity permutation. One method for quantifying
this is to count the number of so-called inversion pairs in π as these describe pairs of
objects that are out of order relative to each other.
Definition 8.1.8. Let π ∈ Sn be a permutation. Then an inversion pair (i, j) of π is a

pair of positive integers i, j ∈ {1, . . . , n} for which i < j but π(i) > π(j).
Note, in particular, that the components of an inversion pair are the positions where the
two “out of order” elements occur.
Example 8.1.9. We classify all inversion pairs for elements in S3 :

!
1 2 3
• id = has no inversion pairs since no elements are “out of order”.
1 2 3
!
1 2 3
• π= has the single inversion pair (2, 3) since π(2) = 3 > 2 = π(3).
1 3 2
!
1 2 3
• π= has the single inversion pair (1, 2) since π(1) = 2 > 1 = π(2).
2 1 3
!
1 2 3
• π= has the two inversion pairs (1, 3) and (2, 3) since we have that both
2 3 1
π(1) = 2 > 1 = π(3) and π(2) = 3 > 1 = π(3).
!
1 2 3
• π= has the two inversion pairs (1, 2) and (1, 3) since we have that both
3 1 2
π(1) = 3 > 1 = π(2) and π(1) = 3 > 2 = π(3).
!
1 2 3
• π= has the three inversion pairs (1, 2), (1, 3), and (2, 3), as you can check.
3 2 1
Example 8.1.10. As another example, for each i, j ∈ {1, . . . , n} with i < j, we define the
transposition tij ∈ Sn by
!
1 2 ··· i ··· j ··· n
tij = .
1 2 ··· j ··· i ··· n
In other words, tij is the permutation that interchanges i and j while leaving all other integers
fixed in place. One can check that the number of inversions pairs in tij is exactly 2(j − i) − 1.
Thus, the number of inversions in a transposition is always odd. E.g.,
!
1 2 3 4
t13 =
3 2 1 4
has inversion pairs (1, 2), (1, 3), and (2, 3).
For the purposes of using permutations in Linear Algebra, the significance of inversion
pairs is mainly due to the following fundamental definition.
Definition 8.1.11. Let π ∈ Sn be a permutation. Then the sign of π, denoted by sign(π),

is defined by

+1, if the number of inversions in π is even
sign(π) = (−1)# of inversion pairs in π = .
−1, if the number of inversions in π is odd
We call π an even permutation if sign(π) = +1, whereas π is called an odd permutation

if sign(π) = −1.
Example 8.1.12. Based upon the computations in Example 8.1.9 above, we have that
! ! !
1 2 3 1 2 3 1 2 3
sign = sign = sign = +1
1 2 3 2 3 1 3 1 2
8.2. DETERMINANTS 107
and that ! ! !
1 2 3 1 2 3 1 2 3
sign = sign = sign = −1.
1 3 2 2 1 3 3 2 1
Similarly, from Example 8.1.10, it follows that any transposition is an odd permutation.
We summarize some of the most basic properties of the sign operation on the symmetric
group in the following theorem.
Theorem 8.1.13. Let n ∈ Z+ be a positive integer. Then,
1. for id ∈ Sn the identity permutation,
sign(id) = +1.
2. for tij ∈ Sn a transposition with i, j ∈ {1, . . . , n} and i < j,
sign(tij ) = −1. (8.1)
3. given any two permutations π, σ ∈ Sn ,
sign(π ◦ σ) = sign(π) sign(σ), (8.2)

sign(π −1 ) = sign(π). (8.3)
4. the number of even permutations in Sn , when n ≥ 2, is exactly 21 n!.
5. the set An of even permutations in Sn forms a group under composition.
8.2 Determinants
Now that we have developed the appropriate background material on permutations, we are
finally ready to define the determinant and explore its many important properties.
8.2.1 Summations indexed by the set of all permutations

Given a positive integer n ∈ Z+ , we being with the following definition:
Definition 8.2.1. Given a square matrix A = (aij ) ∈ Fn×n , the determinant of A is

defined to be
X
det(A) = sign(π)a1,π(1) a2,π(2) · · · an,π(n) , (8.4)
π ∈ Sn
where the sum is over all permutations of n elements (i.e., over the symmetric group).
Note that each permutation in the summand of (8.4) permutes the n columns of the n×n
matrix.
Example 8.2.2. Suppose that A ∈ F2×2 is the 2 × 2 matrix

" #
a11 a12
A= .
a21 a22
To calculate the determinant of A, we first list the two permutations in S2 :

! !
1 2 1 2
id = and σ= .
1 2 2 1
The permutation id has sign 1, and the permutation σ has sign −1. Thus, the determinant
of A is given by
det(A) = a11 a22 − a12 a21 .
Were one to attempt to compute determinants directly using Equation (8.4), then one
would need to sum up n! terms, where each summand is itself a product of n factors. This
is an incredibly inefficient method for finding determinants since n! increases in size very
rapidly as n increases. E.g., 10! = 3628800. Thus, even if you could compute one summand
per second without stopping, it would still take you well over a month to compute the
determinant of a 10 × 10 matrix using Equation (8.4). Fortunately, there are properties of
the determinant (as summarized in Section 8.2.2 below) that can be used to greatly reduce the
size of such computations. These properties of the determinant follow from general properties
that hold for any summation taken over the symmetric group, which are in turn themselves
based upon properties of permutations and the fact that addition and multiplication are
commutative operations in the field F (which, as usual, we take to be either R or C).
Let T : Sn → V be a function defined on the symmetric group Sn that takes values in
some vector space V . E.g., T (π) could be the term corresponding to the permutation π in
Equation 8.4. Then, since the sum

X
T (π)
π ∈ Sn
is finite, we are free to reorder the summands. In other words, the sum is independent
of the order in which the terms are added, and so we are free to permute the term order
without affecting the value of the sum. Some commonly used reorderings of such sums are
the following:
X X
T (π) = T (σ ◦ π) (8.5)
π ∈ Sn π ∈ Sn
X
= T (π ◦ σ) (8.6)
π ∈ Sn
X
= T (π −1 ), (8.7)
π ∈ Sn
where σ is a fixed permutation.

Equation (8.5) follows from the fact that, if π runs through each permutation in Sn exactly
once, then σ ◦ π similarly runs through each permutation but in a potentially different order.
I.e., the action of σ upon something like Equation (8.4) is that σ merely permutes the
permutations that index the terms. Put another way, there is a one-to-one correspondence
between permutations in general and permutations composed with σ.
Similar reasoning holds for Equations (8.6) and (8.7).
8.2.2 Properties of the determinant

We summarize some of the most basic properties of the determinant below. The proof of the
following theorem uses properties of permutations, properties of the sign function on permu-
tations, and properties of sums over the symmetric group as discussed in Section 8.2.1 above.
In thinking about these properties, it is useful to keep in mind that, using Equation (8.4),
the determinant of an n × n matrix A is the sum over all possible ways of selecting n entries
of A, where exactly one element is selected from each row and from each column of A.
Theorem 8.2.3 (Properties of the Determinant). Let n ∈ Z+ be a positive integer, and

suppose that A = (aij ) ∈ Fn×n is an n × n matrix. Then
1. det(0n×n ) = 0 and det(In ) = 1, where 0n×n denotes the n × n zero matrix and In
denotes the n × n identity matrix.
2. det(AT ) = det(A), where AT denotes the transpose of A.
3. denoting by A(·,1) , A(·,2) , . . . , A(·,n) ∈ Fn the columns of A, det(A) is a linear function

of column A(·,i) , for each i ∈ {1, . . . , n}. In other words, if we denote

A = A(·,1) | A(·,2) | · · · | A(·,n)
then, given any scalar z ∈ F and any vectors a1 , a2 , . . . , an , c, b ∈ Fn ,
det [a1 | · · · | ai−1 | zai | · · · | an ] = z det [a1 | · · · | ai−1 | ai | · · · | an ] ,

det [a1 | · · · | ai−1 | b + c | · · · | an ] = det [a1 | · · · | b | · · · | an ] + det [a1 | · · · | c | · · · | an ] .
4. det(A) is an antisymmetric function of the columns of A. In other words, given any

positive integers 1 ≤ i < j ≤ n and denoting A = A(·,1) | A(·,2) | · · · | A(·,n) ,

det(A) = − det A(·,1) | · · · | A(·,j) | · · · | A(·,i) | · · · | A(·,n) .
5. if A has two identical columns, det(A) = 0.
6. if A has a column of zero’s, det(A) = 0.
7. Properties 3–6 also hold when rows are used in place of columns.
8. given any other matrix B ∈ Fn×n ,
det(AB) = det(A) det(B).
9. if A is either upper triangular or lower triangular,
det(A) = a11 a22 · · · ann .
Proof. First, note that Properties 1, 3, 6, and 9 follow directly from the sum given in
Equation (8.4). Moreover, Property 5 follows directly from Property 4, and Property 7
follows directly from Property 2. Thus, we need only prove Properties 2, 4, and 8.
Proof of 2. Since the entries of AT are obtained from those of A by interchanging the
row and column indices, it follows that det(AT ) is given by
X
det(AT ) = sign(π) aπ(1),1 aπ(2),2 · · · aπ(n),n .
π ∈ Sn
Using the commutativity of the product in F and Equation (8.3), we see that
X
det(AT ) = sign(π −1 ) a1,π−1 (1) a2,π−1 (2) · · · an,π−1 (n) ,
π ∈ Sn
which equals det(A) by Equation (8.7).

Proof of 4. Let B = A(·,1) | · · · | A(·,j) | · · · | A(·,i) | · · · | A(·,n) be the matrix obtained
from A by interchanging the ith and the jth column. Then note that
X
det(B) = sign(π) a1,π(1) · · · aj,π(i) · · · ai,π(j) · · · an,π(n) .
π ∈ Sn
Define π̃ = π ◦ tij , and note that π = π̃ ◦ tij . In particular, π(i) = π̃(j) and π(j) = π̃(i), from
which
X
det(B) = sign(π̃ ◦ tij ) a1,π̃(1) · · · ai,π̃(i) · · · aj,π̃(j) · · · an,π̃(n) .
π ∈ Sn
It follows from Equations (8.2) and (8.1) that sign(π̃ ◦ tij ) = −sign (π̃). Thus, using Equa-
tion (8.6), we obtain det(B) = − det(A).
Proof of 8. Using the standard expression for the matrix entries of the product AB in
terms of the matrix entries of A = (aij ) and B = (bij ), we have that
X n
X n
X
det(AB) = sign(π) ··· a1,k1 bk1 ,π(1) · · · an,kn bkn ,π(n)
π ∈ Sn k1 =1 kn =1
X n n
X X
= ··· a1,k1 · · · an,kn sign (π)bk1 ,π(1) · · · bkn ,π(n) .
k1 =1 kn =1 π ∈ Sn
P
Note that, for fixed k1 , . . . , kn ∈ {1, . . . , n}, the sum π ∈ Sn sign (π)bk1 ,π(1) · · · bkn ,π(n) is the
determinant of a matrix composed of rows k1 , . . . , kn of B. Thus, by property 5, it follows
that this expression vanishes unless the ki are pairwise distinct. In other words, the sum
over all choices of k1 , . . . , kn can be restricted to those sets of indices σ(1), . . . , σ(n) that are
labeled by a permutation σ ∈ Sn . In other words,
X X
det(AB) = a1,σ(1) · · · an,σ(n) sign(π) bσ(1),π(1) · · · bσ(n),π(n) .
σ ∈ Sn π ∈ Sn
Now, proceeding with the same arguments as in the proof of Property 4 but with the role
of tij replaced by an arbitrary permutation σ, we obtain
X X
det(AB) = sign(σ) a1,σ(1) · · · an,σ(n) sign(π ◦ σ −1 ) b1,π◦σ−1 (1) · · · bn,π◦σ−1 (n) .
σ ∈ Sn π ∈ Sn
Using Equation (8.6), this last expression then becomes (det(A))(det(B)).
Note that Properties 3 and 4 of Theorem 8.2.3 effectively summarize how multiplica-
tion by an Elementary Matrix interacts with the determinant operation. These Properties
together with Property 9 facilitate numerical computation of determinants for very large
matrices.
8.2.3 Further properties and applications

There are many, many applications of Theorem 8.2.3. We conclude these notes with a few
consequences that are particularly useful when computing with matrices. In particular,
we use the determinant to list several characterizations for matrix invertibility, and, as a
corollary, give a method for using determinants to calculate eigenvalues. You should provide
a proof of the these results for your own benefit.
Theorem 8.2.4. Let n ∈ Z+ and A ∈ Rn×n . Then the following statements are equivalent:
1. A is invertible.
 
x1
.
2. denoting x =  .
 . , the matrix equation Ax = 0 has only the trivial solution x = 0.
xn
   
x1 b1
. .
3. denoting x =  . .
 . , the matrix equation Ax = b has a solution for all b =  .  ∈ R .
n
xn bn
4. A can be factored into a product of elementary matrices.
5. det(A) 6= 0.
6. the rows (or columns) of A form a linearly independent set in Rn .
7. zero is not an eigenvalue of A.
8. the linear operator T : Rn → Rn defined by T (x) = Ax, for every x ∈ Rn , is bijective.
1
Moreover, should A be invertible, then det(A−1 ) = .
det(A)
Given a matrix A ∈ Cn×n and a complex number λ ∈ C, the expression
P (λ) = det(A − λIn )
is called the characteristic polynomial of A. Note that P (λ) is a basis independent

polynomial of degree n. Thus, as with the determinant, we can consider P (λ) to be associated
with the linear map that has matrix A with respect to some basis. Since the eigenvalues
of A are exactly those λ ∈ C such that A − λI is not invertible, the following is then an
immediate corollary.
Corollary 8.2.5. The roots of the polynomial P (λ) = det(A−λI) are exactly the eigenvalues
of A.
8.2.4 Computing determinants with cofactor expansions

As noted in Section 8.2.1, it is generally impractical to compute determinants directly with
Equation (8.4). In this section, we briefly describe the so-called cofactor expansions of a de-
terminant. When properly applied, cofactor expansions are particularly useful for computing
determinants by hand.
Definition 8.2.6. Let n ∈ Z+ and A ∈ Rn×n . Then, for each i, j ∈ {1, 2, . . . , n}, the
i − j minor of A, denoted Mij , is defined to be the determinant of the matrix obtained by
removing the ith row and j th column from A. Moreover, the i − j cofactor of A is defined to
be
Aij = (−1)i+j Mij .
Cofactors themselves, though, aren’t terribly useful unless put together in the right way.
Definition 8.2.7. Let n ∈ Z+ and A = (aij ) ∈ Rn×n . Then, for each i, j ∈ {1, 2, . . . , n}, the
n
X X n
ith row (resp. j th column) cofactor expansion of A is the sum aij Aij (resp. aij Aij ).
j=1 i=1
Theorem 8.2.8. Let n ∈ Z+ and A ∈ Rn×n . Then every row and column factor expansion
of A is equal to the determinant of A.
Since the determinant of a matrix is equal to every row or column cofactor expansion, one
can compute the determinant using a convenient choice of expansions until the calculation
is reduced to one or more 2 × 2 determinants. We close with an example.
Example 8.2.9. By first expanding along the second column, we obtain

1 2 −3 4
−4 1 −3
1 3 4
−4 2 1 3
= (−1)1+2 (2) 3 0 −3 + (−1)2+2
(2) 3 0 −3 .
3 0 0 −3
2 −2
3 2 −2 3
2 0 −2 3
Then, each of the resulting 3 × 3 determinants can be computed by further expansion:

−4 1 3
3 −3 −4 3
1+2 3+2
3 0 −3 = (−1) (1) + (−1) (−2) = −15 + 6 = −9.
2 3 3 −3
2 −2 3

1 −3 4
−3 4 1 −3
2+1 2+3
3 0 −3 = (−1) (3) + (−1) (−3) = 3 + 12 = 15.
−2 3 2 −2
2 −2 3
It follows that the original determinant is then equal to −2(−9) + 2(15) = 48.
1. Let A ∈ C3×3 be given by  
1 0 i
 
A =  0 1 0 .
−i 0 −1
(a) Calculate det(A).
(b) Find det(A4 ).
2. (a) For each permutation π ∈ S3 , compute the number of inversions in π, and classify
π as being either an even or an odd permutation.
(b) Use your result from Part (a) to construct a formula for the determinant of a 3 ×3
matrix.
3. (a) For each permutation π ∈ S4 , compute the number of inversions in π, and classify
π as being either an even or an odd permutation.
(b) Use your result from Part (a) to construct a formula for the determinant of a 4 ×4
matrix.
4. Solve for the variable x in the following expression:

 
" #! 1 0 −3
x −1  
det = det 2 x −6  .
3 1−x
1 3 x−5
5. Prove that the following determinant does not depend upon the value of θ:
 
sin(θ) cos(θ) 0
 
det  − cos(θ) sin(θ) 0
sin(θ) − cos(θ) sin(θ) + cos(θ) 1
6. Given scalars α, β, γ ∈ F, prove that the following matrix is not invertible:

 
sin2 (α) sin2 (β) sin2 (γ)
 2 
cos (α) cos2 (β) cos2 (γ)
1 1 1
Hint: Compute the determinant.
1. Let a, b, c, d, e, f ∈ F be scalars, and suppose that A and B are the following matrices:
" # " #
a b d e
A= and B = .
0 c 0 f
" #!
b a−c
Prove that AB = BA if and only if det = 0.
e d−f
2. Given a square matrix A, prove that A is invertible if and only if AT A is invertible.
3. Prove or give a counterexample: For any n ≥ 1 and A, B ∈ Rn×n , one has
det(A + B) = det(A) + det(B).
4. Prove or give a counterexample: For any r ∈ R, n ≥ 1 and A ∈ Rn×n , one has
det(rA) = r det(A).
Chapter 9
Inner Product Spaces
The abstract definition of a vector space only takes into account algebraic properties for the
addition and scalar multiplication of vectors. For vectors in Rn , for example, we also have
geometric intuition involving the length of a vector or the angle formed by two vectors. In
this chapter we discuss inner product spaces, which are vector spaces with an inner product
defined upon them. Inner products are what allow us to abstract notions such as the length
of a vector. We will also abstract the concept of angle via a condition called orthogonality.
9.1 Inner product

In this section, V is a finite-dimensional, nonzero vector space over F.
Definition 9.1.1. An inner product on V is a map
h·, ·i : V × V → F
(u, v) 7→ hu, vi
with the following four properties.
1. Linearity in first slot: hu + v, wi = hu, wi + hv, wi and hau, vi = ahu, vi for all
u, v, w ∈ V ;
2. Positivity: hv, vi ≥ 0 for all v ∈ V ;
3. Positive definiteness: hv, vi = 0 if and only if v = 0;
117
118 CHAPTER 9. INNER PRODUCT SPACES
4. Conjugate symmetry: hu, vi = hv, ui for all u, v ∈ V .
Remark 9.1.2. Recall that every real number x ∈ R equals its complex conjugate. Hence,
for real vector spaces, conjugate symmetry of an inner product becomes actual symmetry.
Definition 9.1.3. An inner product space is a vector space over F together with an inner
product h·, ·i.
Example 9.1.4. Let V = Fn and u = (u1 , . . . , un ), v = (v1 , . . . , vn ) ∈ Fn . Then we can

define an inner product on V by setting
n
X
hu, vi = ui v i .
i=1
For F = R, this reduces to the usual dot product, i.e.,
u · v = u1 v1 + · · · + un vn .
Example 9.1.5. Let V = F[z] be the space of polynomials with coefficients in F. Given
f, g ∈ F[z], we can define their inner product to be
Z 1
hf, gi = f (z)g(z)dz,
0
where g(z) is the complex conjugate of the polynomial g(z).
For a fixed vector w ∈ V , one can define a map T : V → F by setting T v = hv, wi.
Note that T is linear by Condition 1 of Definition 9.1.1. This implies, in particular, that
h0, wi = 0 for every w ∈ V . By conjugate symmetry, we also have hw, 0i = 0.
Lemma 9.1.6. The inner product is anti-linear in the second slot, that is, hu, v + wi =
hu, vi + hu, wi and hu, avi = ahu, vi for all u, v, w ∈ V .
Proof. For additivity, note that
hu, v + wi = hv + w, ui = hv, ui + hw, ui

= hv, ui + hw, ui = hu, vi + hu, wi.
9.2. NORMS 119
Similarly, for anti-homogeneity, note that
hu, avi = hav, ui = ahv, ui = ahv, ui = ahu, vi.
We close this section by noting that the convention in physics is often the exact opposite
of what we have defined above. In other words, an inner product in physics is traditionally
linear in the second slot and anti-linear in the first slot.
9.2 Norms
The norm of a vector in an arbitrary inner product space is the analog of the length or
magnitude of a vector in Rn . We formally define this concept as follows.
Definition 9.2.1. Let V be a vector space over F. A map
k·k:V →R
v 7→ kvk
is a norm on V if the following three conditions are satisfied.
1. Positive definiteness: kvk = 0 if and only if v = 0;
2. Positive Homogeneity: kavk = |a| kvk for all a ∈ F and v ∈ V ;
3. Triangle inequality: kv + wk ≤ kvk + kwk for all v, w ∈ V .
Remark 9.2.2. Note that, in fact, kvk ≥ 0 for each v ∈ V since
0 = kv − vk ≤ kvk + k − vk = 2kvk.
Next we want to show that a norm can always be defined from an inner product h·, ·i via
the formula
p
kvk = hv, vi for all v ∈ V . (9.1)
Properties 1 and 2 follow easily from Conditions 1 and 3 of Definition 9.1.1. The triangle
inequality requires more careful proof, though, which we give in Theorem 9.3.4 below.
v
x3
x2
x1
Figure 9.1: The length of a vector in R3 via Equation 9.2
If we take V = Rn , then the norm defined by the usual dot product is related the usual
notion of length of a vector. Namely, for v = (x1 , . . . , xn ) ∈ Rn , we have
q
kvk = x21 + · · · + x2n . (9.2)
We illustrate this for the case of R3 in Figure 9.1.

While it is always possible to start with an inner product and use it to define a norm, the
converse requires more care. In particular, one can prove that a norm can be used to define
an inner product via the Equation (9.1) if and only if the norm satisfies the Parallelogram
Law (Theorem 9.3.6).
9.3 Orthogonality
Using the inner product, we can now define the notion of orthogonality, prove that the
Pythagorean theorem holds in any inner product space, and use the Cauchy-Schwarz in-
p
equality to prove the triangle inequality. In particular, this will show that kvk = hv, vi
does indeed defines a norm.
9.3. ORTHOGONALITY 121
Definition 9.3.1. Two vectors u, v ∈ V are orthogonal (denoted u⊥v) if hu, vi = 0.
Note that the zero vector is the only vector that is orthogonal to itself. In fact, the zero
vector is orthogonal to every vector v ∈ V .
Theorem 9.3.2 (Pythagorean Theorem). If u, v ∈ V with u⊥v, then
ku + vk2 = kuk2 + kvk2 .
Proof. Suppose u, v ∈ V such that u⊥v. Then
ku + vk2 = hu + v, u + vi = kuk2 + kvk2 + hu, vi + hv, ui

= kuk2 + kvk2.
Note that the converse of the Pythagorean Theorem holds for real vector spaces since, in
that case, hu, vi + hv, ui = 2Rehu, vi = 0.
Given two vectors u, v ∈ V with v 6= 0, we can uniquely decompose u into two pieces:
one piece parallel to v and one piece orthogonal to v. This is called an orthogonal decom-
position. More precisely, we have
u = u1 + u2 ,
where u1 = av and u2 ⊥v for some scalar a ∈ F. To obtain such a decomposition, write

u2 = u − u1 = u − av. Then, for u2 to be orthogonal to v, we need
0 = hu − av, vi = hu, vi − akvk2 .
Solving for a yields a = hu, vi/kvk2 so that

hu, vi hu, vi
u= v+ u− v . (9.3)
kvk2 kvk2
This decomposition is particularly useful since it allows us to provide a simple proof for
the Cauchy-Schwarz inequality.
Theorem 9.3.3 (Cauchy-Schwarz Inequality). Given any u, v ∈ V , we have
|hu, vi| ≤ kukkvk.
Furthermore, equality holds if and only if u and v are linearly dependent, i.e., are scalar
multiples of each other.
Proof. If v = 0, then both sides of the inequality are zero. Hence, assume that v 6= 0, and
consider the orthogonal decomposition
hu, vi
u= v+w
kvk2
where w⊥v. By the Pythagorean theorem, we have

2
hu, 2 2
kuk2 =
vi
v + kwk2 = |hu, vi| + kwk2 ≥ |hu, vi| .
kvk2 kvk2 kvk2
Multiplying both sides by kvk2 and taking the square root then yields the Cauchy-Schwarz
inequality.
Note that we get equality in the above arguments if and only if w = 0. But, by Equa-
tion (9.3), this means that u and v are linearly dependent.
The Cauchy-Schwarz inequality has many different proofs. Here is another one.
Alternate proof of Theorem 9.3.3. Given u, v ∈ V , consider the norm square of the vector
u + reiθ v:
0 ≤ ku + reiθ vk2 = kuk2 + r 2 kvk2 + 2Re(reiθ hu, vi).
Since hu, vi is a complex number, one can choose θ so that eiθ hu, vi is real. Hence, the right
hand side is a parabola ar 2 + br + c with real coefficients. It will lie above the real axis,
i.e. ar 2 + br + c ≥ 0, if it does not have any real solutions for r. This is the case when the
discriminant satisfies b2 − 4ac ≤ 0. In our case this means
4|hu, vi|2 − 4kuk2 kvk2 ≤ 0.
Moreover, equality only holds if r can be chosen such that u + reiθ v = 0, which means that
u and v are scalar multiples.
9.3. ORTHOGONALITY 123
u+v
v
u+v‘
v‘
u
Figure 9.2: The triangle inequality in R2
Now that we have proven the Cauchy-Schwarz inequality, we are finally able to verify the
p
triangle inequality. This is the final step in showing that kvk = hv, vi does indeed defines
a norm. We illustrate the triangle inequality in Figure 9.2.
Theorem 9.3.4 (Triangle Inequality). For all u, v ∈ V we have
ku + vk ≤ kuk + kvk.
Proof. By a straightforward calculation, we obtain
ku + vk2 = hu + v, u + vi = hu, ui + hv, vi + hu, vi + hv, ui

= hu, ui + hv, vi + hu, vi + hu, vi = kuk2 + kvk2 + 2Rehu, vi.
Note that Rehu, vi ≤ |hu, vi| so that, using the Cauchy-Schwarz inequality, we obtain
ku + vk2 ≤ kuk2 + kvk2 + 2kukkvk = (kuk + kvk)2 .
Taking the square root of both sides now gives the triangle inequality.
Remark 9.3.5. Note that equality holds for the triangle inequality if and only if v = ru or
u = rv for some r ≥ 0. Namely, equality in the proof happens only if hu, vi = kukkvk, which
is equivalent to u and v being scalar multiples of one another.
u−v u+v
v
h
u g
Figure 9.3: The parallelogram law in R2
Theorem 9.3.6 (Parallelogram Law). Given any u, v ∈ V , we have
ku + vk2 + ku − vk2 = 2(kuk2 + kvk2 ).
Proof. By direct calculation,
ku + vk2 + ku − vk2 = hu + v, u + vi + hu − v, u − vi
= kuk2 + kvk2 + hu, vi + hv, ui + kuk2 + kvk2 − hu, vi − hv, ui
= 2(kuk2 + kvk2 ).
Remark 9.3.7. We illustrate the parallelogram law in Figure 9.3.
9.4 Orthonormal bases

We now define the notions of orthogonal basis and orthonormal basis for an inner product
space. As we will see later, orthonormal bases have many special properties that allow us to
simplify various calculations.
9.4. ORTHONORMAL BASES 125
Definition 9.4.1. Let V be an inner product space with inner product h·, ·i. A list of
nonzero vectors (e1 , . . . , em ) in V is called orthogonal if
hei , ej i = 0, for all 1 ≤ i 6= j ≤ m.
The list (e1 , . . . , em ) is called orthonormal if
hei , ej i = δi,j , for all i, j = 1, . . . , m,
where δij is the Kronecker delta symbol. I.e., δij = 1 if i = j and is zero otherwise.
Proposition 9.4.2. Every orthogonal list of nonzero vectors in V is linearly independent.
Proof. Let (e1 , . . . , em ) be an orthogonal list of vectors in V , and suppose that a1 , . . . , am ∈ F
are such that
a1 e1 + · · · + am em = 0.
Then
0 = ka1 e1 + · · · + am em k2 = |a1 |2 ke1 k2 + · · · + |am |2 kem k2
Note that kek k > 0, for all k = 1, . . . , m, since every ek is a nonzero vector. Also, |ak |2 ≥ 0.
Hence, the only solution to a1 e1 + · · · + am em = 0 is a1 = · · · = am = 0.
Definition 9.4.3. An orthonormal basis of a finite-dimensional inner product space V is
an orthonormal list of vectors that is basis for V .
Clearly, any orthonormal list of length dim(V ) is an orthonormal basis for V .
Example 9.4.4. The canonical basis for Fn is an orthonormal basis.
Example 9.4.5. The list (( √12 , √12 ), ( √12 , − √12 )) is an orthonormal basis for R2 .
The next theorem allows us to use inner products to find the coefficients of a vector v ∈ V
in terms of an orthonormal basis. This result highlights how much easier it is to compute
with an orthonormal basis.
Theorem 9.4.6. Let (e1 , . . . , en ) be an orthonormal basis for V . Then, for all v ∈ V , we
have
v = hv, e1 ie1 + · · · + hv, en ien
Pn
and kvk2 = k=1 |hv, ek i|2.
Proof. Let v ∈ V . Since (e1 , . . . , en ) is a basis for V , there exist unique scalars a1 , . . . , an ∈ F
such that
v = a1 e1 + · · · + an en .
Taking the inner product of both sides with respect to ek then yields hv, ek i = ak .
9.5 The Gram-Schmidt orthogonalization procedure

We now come to a fundamentally important algorithm, which is called the Gram-Schmidt
orthogonalization procedure. This algorithm makes it possible to construct, for each
list of linearly independent vectors (resp. basis), a corresponding orthonormal list (resp.
orthonormal basis).
Theorem 9.5.1. If (v1 , . . . , vm ) is a linearly independent list of vectors in V , then there

exists an orthonormal list (e1 , . . . , em ) such that
span(v1 , . . . , vk ) = span(e1 , . . . , ek ), for all k = 1, . . . , m. (9.4)
Proof. The proof is constructive, that is, we will actually construct vectors e1 , . . . , em having
the desired properties. Since (v1 , . . . , vm ) is linearly independent, vk 6= 0 for each k =
v1
1, 2, . . . , m. Set e1 = kv1 k
. Then e1 is a vector of norm 1 and satisfies Equation (9.4) for
k = 1. Next, set
v2 − hv2 , e1 ie1
e2 = .
kv2 − hv2 , e1 ie1 k
This is, in fact, the normalized version of the orthogonal decomposition Equation (9.3). I.e.,
w = v2 − hv2 , e1 ie1 ,
where w⊥v2 . Note that ke2 k = 1 and span(e1 , e2 ) = span(v1 , v2 ).

Now, suppose that e1 , . . . , ek−1 have been constructed such that (e1 , . . . , ek−1 ) is an or-
thonormal list and span(v1 , . . . , vk−1 ) = span(e1 , . . . , ek−1 ). Then define
vk − hvk , e1 ie1 − hvk , e2 ie2 − · · · − hvk , ek−1 iek−1

ek = .
kvk − hvk , e1 ie1 − hvk , e2 ie2 − · · · − hvk , ek−1 iek−1k
Since (v1 , . . . , vk ) is linearly independent, we know that vk 6∈ span(v1 , . . . , vk−1). Hence, we

9.5. THE GRAM-SCHMIDT ORTHOGONALIZATION PROCEDURE 127
also know that vk 6∈ span(e1 , . . . , ek−1 ). It follows that the norm in the definition of ek is not
zero, and so ek is well-defined (i.e., we are not dividing by zero). Note that a vector divided
by its norm has norm 1 so that kek k = 1. Furthermore,

vk − hvk , e1 ie1 − hvk , e2 ie2 − · · · − hvk , ek−1 iek−1
hek , ei i = , ei
kvk − hvk , e1 ie1 − hvk , e2 ie2 − · · · − hvk , ek−1 iek−1 k
hvk , ei i − hvk , ei i
= = 0,
kvk − hvk , e1 ie1 − hvk , e2 ie2 − · · · − hvk , ek−1 iek−1 k
for each 1 ≤ i < k. Hence, (e1 , . . . , ek ) is orthonormal.

From the definition of ek , we see that vk ∈ span(e1 , . . . , ek ) so that span(v1 , . . . , vk ) ⊂
span(e1 , . . . , ek ). Since both lists (e1 , . . . , ek ) and (v1 , . . . , vk ) are linearly independent, they
must span subspaces of the same dimension and therefore are the same subspace. Hence
Equation (9.4) holds.
Example 9.5.2. Take v1 = (1, 1, 0) and v2 = (2, 1, 1) in R3 . The list (v1 , v2 ) is linearly
independent (as you should verify!). To illustrate the Gram-Schmidt procedure, we begin
by setting
v1 1
e1 = = √ (1, 1, 0).
kv1 k 2
Next, set
v2 − hv2 , e1 ie1
e2 = .
kv2 − hv2 , e1 ie1 k
The inner product hv2 , e1 i = √1 h(1, 1, 0), (2, 1, 1)i = √3 , so
2 2
3 1
u2 = v2 − hv2 , e1 ie1 = (2, 1, 1) − (1, 1, 0) = (1, −1, 2).
2 2
q √
Calculating the norm of u2 , we obtain ku2 k = 14 (1 + 1 + 4) = 26 . Hence, normalizing this
vector, we obtain
u2 1
e2 = = √ (1, −1, 2).
ku2 k 6
The list (e1 , e2 ) is therefore orthonormal and has the same span as (v1 , v2 ).
Corollary 9.5.3. Every finite-dimensional inner product space has an orthonormal basis.
Proof. Let (v1 , . . . , vn ) be any basis for V . This list is linearly independent and spans V .
Apply the Gram-Schmidt procedure to this list to obtain an orthonormal list (e1 , . . . , en ),
which still spans V by construction. By Proposition 9.4.2, this list is linearly independent
and hence a basis of V .
Corollary 9.5.4. Every orthonormal list of vectors in V can be extended to an orthonormal

basis of V .
Proof. Let (e1 , . . . , em ) be an orthonormal list of vectors in V . By Proposition 9.4.2, this list
is linearly independent and hence can be extended to a basis (e1 , . . . , em , v1 , . . . , vk ) of V by
the Basis Extension Theorem. Now apply the Gram-Schmidt procedure to obtain a new or-
thonormal basis (e1 , . . . , em , f1 , . . . , fk ). The first m vectors do not change since they already
are orthonormal. The list still spans V and is linearly independent by Proposition 9.4.2 and
therefore forms a basis.
Recall Theorem 7.5.3: given an operator T ∈ L(V, V ) on a complex vector space V , there
exists a basis B for V such that the matrix M(T ) of T with respect to B is upper triangular.
We would like to extend this result to require the additional property of orthonormality.
Corollary 9.5.5. Let V be an inner product space over F and T ∈ L(V, V ). If T is upper-
triangular with respect to some basis, then T is upper-triangular with respect to some or-
thonormal basis.
Proof. Let (v1 , . . . , vn ) be a basis of V with respect to which T is upper-triangular. Apply

the Gram-Schmidt procedure to obtain an orthonormal basis (e1 , . . . , en ), and note that
span(e1 , . . . , ek ) = span(v1 , . . . , vk ), for all 1 ≤ k ≤ n.
We proved before that T is upper-triangular with respect to a basis (v1 , . . . , vn ) if and only if
span(v1 , . . . , vk ) is invariant under T for each 1 ≤ k ≤ n. Since these spans are unchanged by
the Gram-Schmidt procedure, T is still upper triangular for the corresponding orthonormal
basis.
9.6 Orthogonal projections and minimization problems

Definition 9.6.1. Let V be a finite-dimensional inner product space and U ⊂ V be a subset
(but not necessarily a subspace) of V . Then the orthogonal complement of U is defined
9.6. ORTHOGONAL PROJECTIONS AND MINIMIZATION PROBLEMS 129
to be the set
U ⊥ = {v ∈ V | hu, vi = 0 for all u ∈ U }.
Note that, in fact, U ⊥ is always a subspace of V (as you should check!) and that
{0}⊥ = V and V ⊥ = {0}.
In addition, if U1 and U2 are subsets of V satisfying U1 ⊂ U2 , then U2⊥ ⊂ U1⊥ .

Remarkably, if U ⊂ V is a subspace of V , then we can say quite a bit more about U ⊥ .
Theorem 9.6.2. If U ⊂ V is a subspace of V , then V = U ⊕ U ⊥ .
Proof. We need to show two things:
1. V = U + U ⊥ .
2. U ∩ U ⊥ = {0}.
To show Condition 1 holds, let (e1 , . . . , em ) be an orthonormal basis of U. Then, for all
v ∈ V , we can write
v = hv, e1 ie1 + · · · + hv, em iem + v − hv, e1 ie1 − · · · − hv, em iem . (9.5)

| {z } | {z }
u w
The vector u ∈ U, and
hw, ej i = hv, ej i − hv, ej i = 0, for all j = 1, 2, . . . , m,
since (e1 , . . . , em ) is an orthonormal list of vectors. Hence, w ∈ U ⊥ . This implies that

V = U + U ⊥.
To prove that Condition 2 also holds, let v ∈ U ∩ U ⊥ . Then v has to be orthogonal to
every vector in U, including to itself, and so hv, vi = 0. However, this implies v = 0 so that
U ∩ U ⊥ = {0}.
Example 9.6.3. R2 is the direct sum of any two orthogonal lines, and R3 is the direct sum
of any plane and any line orthogonal to the plane (as illustrated in Figure 9.4). For example,
R2 = {(x, 0) | x ∈ R} ⊕ {(0, y) | y ∈ R},

R3 = {(x, y, 0) | x, y ∈ R} ⊕ {(0, 0, z) | z ∈ R}.
Figure 9.4: R3 as a direct sum of a plane and a line, as in Example 9.6.3
Another fundamental fact about the orthogonal complement of a subspace is as follows.

Theorem 9.6.4. If U ⊂ V is a subspace of V , then U = (U ⊥ )⊥ .
Proof. First we show that U ⊂ (U ⊥ )⊥ . Let u ∈ U. Then, for all v ∈ U ⊥ , we have hu, vi = 0.
Hence, u ∈ (U ⊥ )⊥ by the definition of (U ⊥ )⊥ .
Next we show that (U ⊥ )⊥ ⊂ U. Suppose 0 6= v ∈ (U ⊥ )⊥ such that v 6∈ U, and decompose
v according to Theorem 9.6.2, i.e., as
v = u1 + u2 ∈ U ⊕ U ⊥
with u1 ∈ U and u2 ∈ U ⊥ . Then u2 6= 0 since v 6∈ U. Furthermore, hu2 , vi = hu2, u2 i =

6 0.
But then v is not in (U ⊥ )⊥ , which contradicts our initial assumption. Hence, we must have
that (U ⊥ )⊥ ⊂ U.
By Theorem 9.6.2, we have the decomposition V = U ⊕ U ⊥ for every subspace U ⊂ V .

This allows us to define the orthogonal projection PU of V onto U.
Definition 9.6.5. Let U ⊂ V be a subspace of a finite-dimensional inner product space.
Every v ∈ V can be uniquely written as v = u + w where u ∈ U and w ∈ U ⊥ . Define
PU : V → V,
v 7→ u.
Note that PU is called a projection operator since it satisfies PU2 = PU . Further, since we
also have
range (PU ) = U,
null (PU ) = U ⊥ ,
it follows that range (PU )⊥null (PU ). Therefore, PU is called an orthogonal projection.
The decomposition of a vector v ∈ V as given in Equation (9.5) yields the formula
PU v = hv, e1 ie1 + · · · + hv, em iem , (9.6)
where (e1 , . . . , em ) is any orthonormal basis of U. Equation (9.6) is a particularly useful tool
for computing such things as the matrix of PU with respect to the basis (e1 , . . . , em ).
Let us now apply the inner product to the following minimization problem: Given a
subspace U ⊂ V and a vector v ∈ V , find the vector u ∈ U that is closest to the vector v.
In other words, we want to make kv − uk as small as possible. The next proposition shows
that PU v is the closest point in U to the vector v and that this minimum is, in fact, unique.
Proposition 9.6.6. Let U ⊂ V be a subspace of V and v ∈ V . Then
kv − PU vk ≤ kv − uk for every u ∈ U.
Furthermore, equality holds if and only if u = PU v.
Proof. Let u ∈ U and set P := PU for short. Then
kv − P vk2 ≤ kv − P vk2 + kP v − uk2

= k(v − P v) + (P v − u)k2 = kv − uk2 ,
where the second line follows from the Pythagorean Theorem 9.3.2 since v − P v ∈ U ⊥ and
P v − u ∈ U. Furthermore, equality only holds if kP v − uk2 = 0, which is equivalent to
P v = u.
Example 9.6.7. Consider the plane U ⊂ R3 through 0 and perpendicular to the vector
u = (1, 1, 1). Using the standard norm on R3 , we can calculate the distance of the point
v = (1, 2, 3) to U using Proposition 9.6.6. In particular, the distance d between v and U
is given by d = kv − PU vk. Let ( √13 u, u1, u2) be a basis for R3 such that (u1 , u2) is an
orthonormal basis of U. Then, by Equation (9.6), we have
1
v − PU v = ( hv, uiu + hv, u1 iu1 + hv, u2 iu2) − (hv, u1iu1 + hv, u2iu2 )
3
1
= hv, uiu
3
1
= h(1, 2, 3), (1, 1, 1)i(1, 1, 1)
3
= (2, 2, 2).
√
Hence, d = k(2, 2, 2)k = 2 3.
1. Let (e1 , e2 , e3 ) be the canonical basis of R3 , and define
f1 = e1 + e2 + e3
f2 = e2 + e3
f3 = e3 .
(a) Apply the Gram-Schmidt process to the basis (f1 , f2 , f3 ).
(b) What do you obtain if you instead applied the Gram-Schmidt process to the basis
(f3 , f2 , f1 )?
2. Let C[−π, π] = {f : [−π, π] → R | f is continuous} denote the inner product space

of continuous real-valued functions defined on the interval [−π, π] ⊂ R, with inner
product given by
Z π
hf, gi = f (x)g(x)dx, for every f, g ∈ C[−π, π].
−π
Then, given any positive integer n ∈ Z+ , verify that the set of vectors

1 sin(x) sin(2x) sin(nx) cos(x) cos(2x) cos(nx)
√ , √ , √ ,..., √ , √ , √ ,..., √
2π π π π π π π
is orthonormal.
3. Let R2 [x] denote the inner product space of polynomials over R having degree at most
two, with inner product given by
Z 1
hf, gi = f (x)g(x)dx, for every f, g ∈ R2 [x].
0
Apply the Gram-Schmidt procedure to the standard basis {1, x, x2 } for R2 [x] in order
to produce an orthonormal basis for R2 [x].
4. Let v1 , v2 , v3 ∈ R3 be given by v1 = (1, 2, 1), v2 = (1, −2, 1), and v3 = (1, 2, −1).
Apply the Gram-Schmidt procedure to the basis (v1 , v2 , v3 ) of R3 , and call the resulting
orthonormal basis (u1 , u2, u3 ).
5. Let P ⊂ R3 be the plane containing 0 perpendicular to the vector (1, 1, 1). Using the
standard norm, calculate the distance of the point (1, 2, 3) to P .
6. Give an orthonormal basis for null(T ), where T ∈ L(C4 ) is the map with canonical
matrix  
1 1 1 1
 
1 1 1 1
 .
1 1 1 1
 
1 1 1 1
1. Let V be a finite-dimensional inner product space over F. Given any vectors u, v ∈ V ,
prove that the following two statements are equivalent:
(a) hu, vi = 0
(b) kuk ≤ ku + αvk for every α ∈ F.
2. Let n ∈ Z+ be a positive integer, and let a1 , . . . , an , b1 , . . . , bn ∈ R be any collection of

2n real numbers. Prove that
n
!2 n
! n
!
X X X b2
ak bk ≤ ka2k k
k=1 k=1 k=1

k
3. Prove or disprove the following claim:
Claim. There is an inner product h· , ·i on R2 whose associated norm k · k is given by

the formula
k(x1 , x2 )k = |x1 | + |x2 |
for every vector (x1 , x2 ) ∈ R2 , where | · | denotes the absolute value function on R.
4. Let V be a finite-dimensional inner product space over R. Given u, v ∈ V , prove that
ku + vk2 − ku − vk2
hu, vi = .
4
5. Let V be a finite-dimensional inner product space over C. Given u, v ∈ V , prove that
ku + vk2 − ku − vk2 ku + ivk2 − ku − ivk2

hu, vi = + i.
4 4
6. Let V be a finite-dimensional inner product space over F, and let U be a subspace of

V . Prove that the orthogonal complement U ⊥ of U with respect to the inner product
h· , ·i on V satisfies
dim(U ⊥ ) = dim(V ) − dim(U).
7. Let V be a finite-dimensional inner product space over F, and let U be a subspace of

V . Prove that U = V if and only if the orthogonal complement U ⊥ of U with respect
to the inner product h· , ·i on V satisfies U ⊥ = {0}.
8. Let V be a finite-dimensional inner product space over F, and suppose that P ∈ L(V )
is a linear operator on V having the following two properties:
(a) Given any vector v ∈ V , P (P (v)) = P (v). I.e., P 2 = P .

(b) Given any vector u ∈ null(P ) and any vector v ∈ range(P ), hu, vi = 0.
Prove that P is an orthogonal projection.
9. Prove or give a counterexample: For any n ≥ 1 and A ∈ Cn×n , one has
null(A) = (range(A))⊥ .
10. Prove or give a counterexample: The Gram-Schmidt process applied to an an orthonor-

mal list of vectors reproduces that list unchanged.
Chapter 10
Change of Bases
In Section 6.6, we saw that linear operators on an n-dimensional vector space are in one-
to-one correspondence with n × n matrices. This correspondence, however, depends upon
the choice of basis for the vector space. In this chapter we address the question of how the
matrix for a linear operator changes if we change from one orthonormal basis to another.
10.1 Coordinate vectors
Let V be a finite-dimensional inner product space with inner product h·, ·i and dimension
dim(V ) = n. Then V has an orthonormal basis e = (e1 , . . . , en ), and, according to Theo-
rem 9.4.6, every v ∈ V can be written as
n
X
v= hv, ei iei .
i=1
This induces a map
[ · ]e : V → Fn
 
hv, e1 i
 . 
v 7→  . 
 . ,
hv, en i
136
10.1. COORDINATE VECTORS 137
which maps the vector v ∈ V to the n × 1 column vector of its coordinates with respect to
the basis e. The column vector [v]e is called the coordinate vector of v with respect to
the basis e.
Example 10.1.1. Recall that the vector space R1 [x] of polynomials over R of degree at
most 1 is an inner product space with inner product defined by
Z 1
hf, gi = f (x)g(x)dx.
0
√
Then e = (1, 3(−1 + 2x)) forms an orthonormal basis for R1 [x]. The coordinate vector of
the polynomial p(x) = 3x + 2 ∈ R1 [x] is, e.g.,
# "
1 7
[p(x)]e = √ .
2 3
Note also that the map [ · ]e is an isomorphism (meaning that it is an injective and
surjective linear map) and that it is also inner product preserving. Denote the usual inner
product on Fn by
Xn
hx, yiFn = xk y k .
k=1
Then
hv, wiV = h[v]e , [w]e iFn , for all v, w ∈ V ,
since
n
X n
X
hv, wiV = hhv, eiiei , hw, ej iej i = hv, ei ihw, ej ihei , ej i
i,j=1 i,j=1
Xn n
X
= hv, ei ihw, ej iδij = hv, eiihw, eii = h[v]e , [w]e iFn .
i,j=1 i=1
It is important to remember that the map [ · ]e depends on the choice of basis e = (e1 , . . . , en ).
138 CHAPTER 10. CHANGE OF BASES
10.2 Change of basis transformation
Recall that we can associate a matrix A ∈ Fn×n to every operator T ∈ L(V, V ). More
precisely, the j th column of the matrix A = M(T ) with respect to a basis e = (e1 , . . . , en ) is
obtained by expanding T ej in terms of the basis e. If the basis e is orthonormal, then the
coefficient of ei is just the inner product of the vector with ei . Hence,
M(T ) = (hT ej , ei i)1≤i,j≤n,
where i is the row index and j is the column index of the matrix.
Conversely, if A ∈ Fn×n is a matrix, then we can associate a linear operator T ∈ L(V, V )
to A by setting
n
X n X
X n
Tv = hv, ej iT ej = hT ej , ei ihv, ej iei
j=1 j=1 i=1
n n
!
X X X
= aij hv, ej i ei = (A[v]e )i ei ,
i=1 j=1 i=1
where (A[v]e )i denotes the ith component of the column vector A[v]e . With this construction,
we have M(T ) = A. The coefficients of T v in the basis (e1 , . . . , en ) are recorded by the column
vector obtained by multiplying the n × n matrix A with the n × 1 column vector [v]e whose
components ([v]e )j = hv, ej i.
Example 10.2.1. Given " #

1 −i
A= ,
i 1
we can define T ∈ L(V, V ) with respect to the canonical basis as follows:
" # " #" # " #
z1 1 −i z1 z1 − iz2
T = = .
z2 i 1 z2 iz1 + z2
Suppose that we want to use another orthonormal basis f = (f1 , . . . , fn ) for V . Then, as
10.2. CHANGE OF BASIS TRANSFORMATION 139
Pn Pn
before, we have v = i=1 hv, fi ifi . Comparing this with v = j=1 hv, ej iej , we find that
n n n
!
X X X
v= hhv, ej iej , fi ifi = hej , fiihv, ej i fi .
i,j=1 i=1 j=1
Hence,
[v]f = S[v]e ,
where
S = (sij )ni,j=1 with sij = hej , fi i.
The j th column of S is given by the coefficients of the expansion of ej in terms of the basis
f = (f1 , . . . , fn ). The matrix S describes a linear map in L(Fn ), which is called the change
of basis transformation.
We may also interchange the role of bases e and f . In this case, we obtain the matrix
R = (rij )ni,j=1, where
rij = hfj , ei i.
Then, by the uniqueness of the expansion in a basis, we obtain
[v]e = R[v]f
so that
RS[v]e = [v]e , for all v ∈ V .
Since this equation is true for all [v]e ∈ Fn , it follows that either RS = I or R = S −1 . In
particular, S and R are invertible. We can also check this explicitly by using the properties
of orthonormal bases. Namely,
n
X n
X
(RS)ij = rik skj = hfk , ei ihej , fk i
k=1 k=1
Xn
= hej , fk ihei , fk i = h[ej ]f , [ei ]f iFn = δij .
k=1
Matrix S (and similarly also R) has the interesting property that its columns are orthonor-
mal to one another. This follows from the fact that the columns are the coordinates of
orthonormal vectors with respect to another orthonormal basis. A similar statement holds
for the rows of S (and similarly also R).
Example 10.2.2. Let V = C2 , and choose the orthonormal bases e = (e1 , e2 ) and f =
(f1 , f2 ) with
" # " #
1 1
e1 = , e2 = ,
0 0
" # " #
1 1 1 −1
f1 = √ , f2 = √ .
2 1 2 1
Then " # " #

he1 , f1 i he2 , f1 i 1 1 1
S= =√
he1 , f2 i he2 , f2 i 2 −1 1
and " " #
#
hf1 , e1 i hf2 , e1 i
1 1 −1
R= =√ .
hf1 , e2 i hf2 , e2 i 2 1 1
One can then check explicitly that indeed
" #" # " #
1 1 −1 1 1 1 0
RS = = = I.
2 1 1 −1 1 0 1
So far we have only discussed how the coordinate vector of a given vector v ∈ V changes
under the change of basis from e to f . The next question we can ask is how the matrix
M(T ) of an operator T ∈ L(V ) changes if we change the basis. Let A be the matrix of T
with respect to the basis e = (e1 , . . . , en ), and let B be the matrix for T with respect to the
basis f = (f1 , . . . , fn ). How do we determine B from A? Note that
[T v]e = A[v]e
so that
[T v]f = S[T v]e = SA[v]e = SAR[v]f = SAS −1 [v]f .
This implies that

B = SAS −1 .
Example 10.2.3. Continuing Example 10.2.2, let

" #
1 1
A=
1 1
be the matrix of a linear operator with respect to the basis e. Then the matrix B with
respect to the basis f is given by
" #" #" # " #" # " #
1 1 1 1 1 1 −1 1 1 1 2 0 2 0
B = SAS −1 = = = .
2 −1 1 1 1 1 1 2 −1 1 2 0 0 0
1. Consider R3 with two orthonormal bases: the canonical basis e = (e1 , e2 , e3 ) and the
basis f = (f1 , f2 , f3 ), where
1 1 1
f1 = √ (1, 1, 1), f2 = √ (1, −2, 1), f3 = √ (1, 0, −1) .
3 6 2
Find the matrix, S, of the change of basis transformation such that
[v]f = S[v]e , for all v ∈ R3 ,
where [v]b denotes the column vector of v with respect to the basis b.
2. Let v ∈ C4 be the vector given by v = (1, i, −1, −i). Find the matrix (with respect to
the canonical basis on C4 ) of the orthogonal projection P ∈ L(C4 ) such that
null(P ) = {v}⊥ .
3. Let U be the subspace of R3 that coincides with the plane through the origin that is
perpendicular to the vector n = (1, 1, 1) ∈ R3 .
(a) Find an orthonormal basis for U.

(b) Find the matrix (with respect to the canonical basis on R3 ) of the orthogonal
projection P ∈ L(R3 ) onto U, i.e., such that range(P ) = U.
4. Let V = C4 with its standard inner product. For θ ∈ R, let

 
1
 iθ 
e 
vθ =   4
e2iθ  ∈ C .
 
e3iθ
Find the canonical matrix of the orthogonal projection onto the subspace {vθ }⊥ .
1. Let V be a finite-dimensional vector space over F with dimension n ∈ Z+ , and sup-
pose that b = (v1 , v2 , . . . , vn ) is a basis for V . Prove that the coordinate vectors
[v1 ]b , [v2 ]b , . . . , [vn ]b with respect to b form a basis for Fn .
2. Let V be a finite-dimensional vector space over F, and suppose that T ∈ L(V ) is a

linear operator having the following property: Given any two bases b and c for V , the
matrix M(T, b) for T with respect to b is the same as the matric M(T, c) for T with
respect to c. Prove that there exists a scalar α ∈ F such that T = αidV , where idV
denotes the identity map on V .
Chapter 11
The Spectral Theorem for Normal

Linear Maps
In this chapter we come back to the question of when a linear operator on an inner product
space V is diagonalizable. We first introduce the notion of the adjoint (a.k.a. hermitian
conjugate) of an operator, and we then use this to define so-called normal operators. The
main result of this chapter is the Spectral Theorem, which states that normal operators are
diagonal with respect to an orthonormal basis. We use this to show that normal opera-
tors are “unitarily diagonalizable” and generalize this notion to finding the singular-value
decomposition of an operator.
11.1 Self-adjoint or hermitian operators

Let V be a finite-dimensional inner product space over C with inner product h·, ·i. A linear
operator T ∈ L(V ) is uniquely determined by the values of
hT v, wi, for all v, w ∈ V .
This means, in particular, that if T, S ∈ L(V ) and
hT v, wi = hSv, wi for all v, w ∈ V ,
then T = S. To see this, take w to be the elements of an orthonormal basis of V .
144
11.1. SELF-ADJOINT OR HERMITIAN OPERATORS 145
Definition 11.1.1. Given T ∈ L(V ), the adjoint (a.k.a. hermitian conjugate) of T is

defined to be the operator T ∗ ∈ L(V ) for which
hT v, wi = hv, T ∗ wi, for all v, w ∈ V
Moreover, we call T self-adjoint (a.k.a. hermitian) if T = T ∗ .
The uniqueness of T ∗ is clear by the previous observation.
Example 11.1.2. Let V = C3 , and let T ∈ L(C3 ) be defined by T (z1 , z2 , z3 ) = (2z2 +

iz3 , iz1 , z2 ). Then
h(y1, y2 , y3 ), T ∗ (z1 , z2 , z3 )i = hT (y1, y2 , y3 ), (z1 , z2 , z3 )i

= h(2y2 + iy3 , iy1 , y2 ), (z1 , z2 , z3 )i
= 2y2 z1 + iy3 z1 + iy1 z2 + y2 z3
= h(y1 , y2, y3 ), (−iz2 , 2z1 + z3 , −iz1 )i
so that T ∗ (z1 , z2 , z3 ) = (−iz2 , 2z1 + z3 , −iz1 ). Writing the matrix for T in terms of the
canonical basis, we see that
  

0 2 i 0 −i 0
   
M(T ) =  i 0 0 and M(T ∗ ) =  2 0 1 .
0 1 0 −i 0 0
Note that M(T ∗ ) can be obtained from M(T ) by taking the complex conjugate of each
element and then transposing. This operation is called the conjugate transpose of M(T ),
and we denote it by (M(T ))∗ .
We collect several elementary properties of the adjoint operation into the following propo-
sition. You should provide a proof of these results for your own practice.
Proposition 11.1.3. Let S, T ∈ L(V ) and a ∈ F.
1. (S + T )∗ = S ∗ + T ∗ .
2. (aT )∗ = aT ∗ .
146 CHAPTER 11. THE SPECTRAL THEOREM FOR NORMAL LINEAR MAPS
3. (T ∗ )∗ = T .
4. I ∗ = I.
5. (ST )∗ = T ∗ S ∗ .
6. M(T ∗ ) = M(T )∗ .
When n = 1, note that the conjugate transpose of a 1 × 1 matrix A is just the complex
conjugate of its single entry. Hence, requiring A to be self-adjoint (A = A∗ ) amounts to
saying that this sole entry is real. Because of the transpose, though, reality is not the same
as self-adjointness when n > 1, but the analogy does nonetheless carry over to the eigenvalues
of self-adjoint operators.
Proposition 11.1.4. Every eigenvalue of a self-adjoint operator is real.
Proof. Suppose λ ∈ C is an eigenvalue of T and that 0 6= v ∈ V is a corresponding eigenvector

such that T v = λv. Then
λkvk2 = hλv, vi = hT v, vi = hv, T ∗ vi

= hv, T vi = hv, λvi = λhv, vi = λkvk2 .
This implies that λ = λ.

" #
2 1+i
Example 11.1.5. The operator T ∈ L(V ) defined by T (v) = v is self-adjoint,
1−i 3
and it can be checked (e.g., using the characteristic polynomial) that the eigenvalues of T
are λ = 1, 4.
11.2 Normal operators

Normal operators are those that commute with their own adjoint. As we will see, this
includes many important examples of operations.
Definition 11.2.1. We call T ∈ L(V ) normal if T T ∗ = T ∗ T .
Given an arbitrary operator T ∈ L(V ), we have that T T ∗ 6= T ∗ T in general. However,

both T T ∗ and T ∗ T are self-adjoint, and any self-adjoint operator T is normal. We now give
a different characterization for normal operators in terms of norms.
11.2. NORMAL OPERATORS 147
Proposition 11.2.2. Let V be a complex inner product space, and suppose that T ∈ L(V )
satisfies
hT v, vi = 0, for all v ∈ V .
Then T = 0.
Proof. You should be able to verify that
1
hT u, wi = {hT (u + w), u + wi − hT (u − w), u − wi
4
+ihT (u + iw), u + iwi − ihT (u − iw), u − iwi} .
Since each term on the right-hand side is of the form hT v, vi, we obtain 0 for each u, w ∈ V .
Hence T = 0.
Proposition 11.2.3. Let T ∈ L(V ). Then T is normal if and only if
kT vk = kT ∗ vk, for all v ∈ V .
Proof. Note that
T is normal ⇐⇒ T ∗ T − T T ∗ = 0
⇐⇒ h(T ∗ T − T T ∗)v, vi = 0, for all v ∈ V
⇐⇒ hT T ∗ v, vi = hT ∗ T v, vi, for all v ∈ V
⇐⇒ kT vk2 = kT ∗ vk2 , for all v ∈ V .
Corollary 11.2.4. Let T ∈ L(V ) be a normal operator.
1. null (T ) = null (T ∗ ).
2. If λ ∈ C is an eigenvalue of T , then λ is an eigenvalue of T ∗ with the same eigenvector.
3. If λ, µ ∈ C are distinct eigenvalues of T with associated eigenvectors v, w ∈ V , respec-

tively, then hv, wi = 0.
Proof. Note that Part 1 follows from Proposition 11.2.3 and the positive definiteness of the
norm.
To prove Part 2, first verify that if T is normal, then T −λI is also normal with (T −λI)∗ =
T ∗ − λI. Therefore, by Proposition 11.2.3, we have
0 = k(T − λI)vk = k(T − λI)∗ vk = k(T ∗ − λI)vk,
and so v is an eigenvector of T ∗ with eigenvalue λ.

Using Part 2, note that
(λ − µ)hv, wi = hλv, wi − hv, µwi = hT v, wi − hv, T ∗wi = 0.
Since λ − µ 6= 0 it follows that hv, wi = 0, proving Part 3.
11.3 Normal operators and the spectral decomposition

Recall that an operator T ∈ L(V ) is diagonalizable if there exists a basis B for V such
that B consists entirely of eigenvectors for T . The nicest operators on V are those that are
diagonalizable with respect to some orthonormal basis for V . In other words, these are the
operators for which we can find an orthonormal basis for V that consisting of eigenvectors
for T . The Spectral Theorem for finite-dimensional complex inner product spaces states that
this can be done precisely for normal operators.
Theorem 11.3.1 (Spectral Theorem). Let V be a finite-dimensional inner product space

over C and T ∈ L(V ). Then T is normal if and only if there exists an orthonormal basis for
V consisting of eigenvectors for T .
Proof.
(“=⇒”) Suppose that T is normal. Combining Theorem 7.5.3 and Corollary 9.5.5, there
exists an orthonormal basis e = (e1 , . . . , en ) for which the matrix M(T ) is upper triangular,
i.e.,  
a11 · · · a1n
 .. .. 
M(T ) = 
 . . 
.
0 ann
We will show that M(T ) is, in fact, diagonal, which implies that the basis elements e1 , . . . , en
are eigenvectors of T .
11.3. NORMAL OPERATORS AND THE SPECTRAL DECOMPOSITION 149
Since M(T ) = (aij )ni,j=1 with aij = 0 for i > j, we have T e1 = a11 e1 and T ∗ e1 =
Pn
k=1 a1k ek . Thus, by the Pythagorean Theorem and Proposition 11.2.3,
n
X n
X
2 2 2 ∗ 2 2
|a11 | = ka11 e1 k = kT e1 k = kT e1 k = k a1k ek k = |a1k |2 ,
k=1 k=1
from which it follows that |a12 | = · · · = |a1n | = 0. Repeating this argument, kT ej k2 = |ajj |2
P
and kT ∗ ej k2 = nk=j |ajk |2 so that aij = 0 for all 2 ≤ i < j ≤ n. Hence, T is diagonal with
respect to the basis e, and e1 , . . . , en are eigenvectors of T .
(“⇐=”) Suppose there exists an orthonormal basis (e1 , . . . , en ) for V that consists of eigen-
vectors for T . Then the matrix M(T ) with respect to this basis is diagonal. Moreover,
M(T ∗ ) = M(T )∗ with respect to this basis must also be a diagonal matrix. It follows that
T T ∗ = T ∗ T since their corresponding matrices commute:
M(T T ∗ ) = M(T )M(T ∗ ) = M(T ∗ )M(T ) = M(T ∗ T ).
The following corollary is the best possible decomposition of a complex vector space V
into subspaces that are invariant under a normal operator T . On each subspace null (T −λi I),
the operator T acts just like multiplication by scalar λi . In other words,
T |null (T −λi I) = λi Inull (T −λi I) .
Corollary 11.3.2. Let T ∈ L(V ) be a normal operator, and denote by λ1 , . . . , λm the distinct
eigenvalues of T .
1. V = null (T − λ1 I) ⊕ · · · ⊕ null (T − λm I).
2. If i 6= j, then null (T − λi I)⊥null (T − λj I).
As we will see in the next section, we can use Corollary 11.3.2 to decompose the canonical
matrix for a normal operator into a so-called “unitary diagonalization”.
11.4 Applications of the spectral theorem: diagonal-

ization
Let e = (e1 , . . . , en ) be a basis for an n-dimensional vector space V , and let T ∈ L(V ). In
this section we denote the matrix M(T ) of T with respect to basis e by [T ]e . This is done
to emphasize the dependency on the basis e. In other words, we have that
[T v]e = [T ]e [v]e , for all v ∈ V ,
where  
v1
.
[v]e = 
.
.
vn
is the coordinate vector for v = v1 e1 + · · · + vn en with vi ∈ F.
The operator T is diagonalizable if there exists a basis e such that [T ]e is diagonal, i.e.,
if there exist λ1 , . . . , λn ∈ F such that
 
λ1 0
 .. 
[T ]e = 
 . .

0 λn
The scalars λ1 , . . . , λn are necessarily eigenvalues of T , and e1 , . . . , en are the corresponding

eigenvectors. We summarize this in the following proposition.
Proposition 11.4.1. T ∈ L(V ) is diagonalizable if and only if there exists a basis (e1 , . . . , en )
consisting entirely of eigenvectors of T .
We can reformulate this proposition using the change of basis transformations as follows.
Suppose that e and f are bases of V such that [T ]e is diagonal, and let S be the change of
basis transformation such that [v]e = S[v]f . Then S[T ]f S −1 = [T ]e is diagonal.
Proposition 11.4.2. T ∈ L(V ) is diagonalizable if and only if there exists a invertible

11.4. APPLICATIONS OF THE SPECTRAL THEOREM: DIAGONALIZATION 151
matrix S ∈ Fn×n such that  

λ1 0
 .. 
S[T ]f S −1 =
 . ,

0 λn
where [T ]f is the matrix for T with respect to a given arbitrary basis f = (f1 , . . . , fn ).
On the other hand, the Spectral Theorem tells us that T is diagonalizable with respect
to an orthonormal basis if and only if T is normal. Recall that
[T ∗ ]f = [T ]∗f
for any orthonormal basis f of V . As before,
A∗ = (aji )nij=1, for A = (aij )ni,j=1 ,
is the conjugate transpose of the matrix A. When F = R, note that A∗ = AT is just the
transpose of the matrix, where AT = (aji)ni,j=1 .
The change of basis transformation between two orthonormal bases is called unitary in
the complex case and orthogonal in the real case. Let e = (e1 , . . . , en ) and f = (f1 , . . . , fn )
be two orthonormal bases of V , and let U be the change of basis matrix such that [v]f = U[v]e ,
for all v ∈ V . Then
hei , ej i = δij = hfi , fj i = hUei , Uej i.
Since this holds for the basis e, it follows that U is unitary if and only if
hUv, Uwi = hv, wi for all v, w ∈ V . (11.1)
This means that unitary matrices preserve the inner product. Operators that preserve the
inner product are often also called isometries. Orthogonal matrices also define isometries.
By the definition of the adjoint, hUv, Uwi = hv, U ∗ Uwi, and so Equation 11.1 implies
that isometries are characterized by the property
U ∗ U = I, for the unitary case,

O T O = I, for the orthogonal case.
The equation U ∗ U = I implies that U −1 = U ∗ . For finite-dimensional inner product spaces,

the left inverse of an operator is also the right inverse, and so
UU ∗ = I if and only if U ∗ U = I,
(11.2)
OO T = I if and only if O T O = I.
It is easy to see that the columns of a unitary matrix are the coefficients of the elements of
an orthonormal basis with respect to another orthonormal basis. Therefore, the columns are
orthonormal vectors in Cn (or in Rn in the real case). By Condition (11.2), this is also true
for the rows of the matrix.
The Spectral Theorem tells us that T ∈ L(V ) is normal if and only if [T ]e is diagonal
with respect to an orthonormal basis e for V , i.e., if there exists a unitary matrix U such
that  
λ1 0
 .. 
UT U ∗ = 
 . .

0 λn
Conversely, if a unitary matrix U exists such that UT U ∗ = D is diagonal, then
T T ∗ − T ∗ T = U ∗ (DD − DD)U = 0
since diagonal matrices commute, and hence T is normal.

Let us summarize some of the definitions that we have seen in this section.
Definition 11.4.3. Given a square matrix A ∈ Fn×n , we call
1. symmetric if A = AT .
2. Hermitian if A = A∗ .
3. orthogonal if AAT = I.
4. unitary if AA∗ = I.
Note that every type of matrix in Definition 11.4.3 is an example of a normal operator.
11.4. APPLICATIONS OF THE SPECTRAL THEOREM: DIAGONALIZATION 153
An example of a normal operator N that is neither Hermitian nor unitary is

" #
−1 −1
N =i .
−1 1
You can easily verify that NN ∗ = N ∗ N and that iN is symmetric (not Hermitian).
Example 11.4.4. Consider the matrix

" #
2 1+i
A=
1−i 3
from Example 11.1.5. To unitarily diagonalize A, we need to find a unitary matrix U and a
diagonal matrix D such that A = UDU −1 . To do this, we need to first find a basis for C2
that consists entirely of orthonormal eigenvectors for the linear map T ∈ L(C2 ) defined by
T v = Av, for all v ∈ C2 .
To find such an orthonormal basis, we start by finding the eigenspaces
" of T . We already
#
1 0
determined that the eigenvalues of T are λ1 = 1 and λ2 = 4, so D = . It follows that
0 4
C2 = null (T − I) ⊕ null (T − 4I)

= span((−1 − i, 1)) ⊕ span((1 + i, 2)).
Now apply the Gram-Schmidt procedure to each eigenspace in order to obtain the columns
of U. Here,
" #" #" #−1
−1−i 1+i
√ √ 1 0 −1−i√ 1+i
√
A = UDU −1 = 3 6 3 6
√1 √2 0 4 √1 √2
3 6 3 6
" #" #" #
−1−i 1+i −1+i √1
√
3
√
6
1 0 √
3 3
= .
√1 √2 0 4 1−i
√ √2
3 6 6 6
As an application, note that such diagonal decomposition allows us to easily compute

powers and the exponential of matrices. Namely, if A = UDU −1 with D diagonal, then we
have
An = (UDU −1 )n = UD n U −1 ,
∞ ∞
!
X 1 k X 1 k
exp(A) = A =U D U −1 = U exp(D)U −1 .
k! k!
k=0 k=0
Example 11.4.5. Continuing Example 11.4.4,

" # " #
1 0 6 5 + 5i
A2 = (UDU −1 )2 = UD 2 U −1 =U U∗ = ,
0 16 5 − 5i 11
" # " #
2 1+i
1 0 (1 + 2n−1 ) (−1 + 22n )
An = (UDU −1 )n = UD n U −1 = U 2n
U∗ = 3
1−i
3
1
,
0 2 3
(−1 + 22n ) 3
(1 + 22n+1 )
" # " #
4 4 4
e 0 1 2e + e e − e + i(e − e)
exp(A) = U exp(D)U −1 =U 4
U −1 = .
0 e 3 e4 − e + i(e − e4 ) e + 2e4
11.5 Positive operators

Recall that self-adjoint operators are the operator analog for real numbers. Let us now define
the operator analog for positive (or, more precisely, nonnegative) real numbers.
Definition 11.5.1. An operator T ∈ L(V ) is called positive (denoted T ≥ 0) if T = T ∗

and hT v, vi ≥ 0 for all v ∈ V .
(If V is a complex vector space, then the condition of self-adjointness follows from the
condition hT v, vi ≥ 0 and hence can be dropped).
Example 11.5.2. Note that, for all T ∈ L(V ), we have T ∗ T ≥ 0 since T ∗ T is self-adjoint
and hT ∗ T v, vi = hT v, T vi ≥ 0.
Example 11.5.3. Let U ⊂ V be a subspace of V and PU be the orthogonal projection onto

U. Then PU ≥ 0. To see this, write V = U ⊕ U ⊥ and v = uv + u⊥ v for each v ∈ V , where
uv ∈ U and uv ∈ U . Then hPU v, wi = huv , uw + uw i = huv , uw i = huv + u⊥
⊥ ⊥ ⊥
v , uw i = hv, PU wi
so that PU∗ = PU . Also, setting v = w in the above string of equations, we obtain hPU v, vi =
huv , uv i ≥ 0, for all v ∈ V . Hence, PU ≥ 0.
11.6. POLAR DECOMPOSITION 155
If λ is an eigenvalue of a positive operator T and v ∈ V is an associated eigenvector, then

hT v, vi = hλv, vi = λhv, vi ≥ 0. Since hv, vi ≥ 0 for all vectors v ∈ V , it follows that λ ≥ 0.
√
This fact can be used to define T by setting
√ p
T ei = λi ei ,
where λi are the eigenvalues of T with respect to the orthonormal basis e = (e1 , . . . , en ). We
know that these exist by the Spectral Theorem.
11.6 Polar decomposition

Continuing the analogy between C and L(V ), recall the polar form of a complex number
z = |z|eiθ , where |z| is the absolute value or modulus of z and eiθ lies on the unit circle in
R2 . In terms of an operator T ∈ L(V ), where V is a complex inner product space, a unitary
operator U takes the role of eiθ , and |T | takes the role of the modulus. As in Section 11.5,
√
T ∗ T ≥ 0 so that |T | := T ∗ T exists and satisfies |T | ≥ 0 as well.
Theorem 11.6.1. For each T ∈ L(V ), there exists a unitary U such that
T = U|T |.
This is called the polar decomposition of T .
Sketch of proof. We start by noting that
kT vk2 = k |T | vk2,
√ √
since hT v, T vi = hv, T ∗T vi = h T ∗ T v, T ∗ T vi. This implies that null (T ) = null (|T |). By
the Dimension Formula, this also means that dim(range (T )) = dim(range (|T |)). Moreover,
we can define an isometry S : range (|T |) → range (T ) by setting
S(|T |v) = T v.
The trick is now to define a unitary operator U on all of V such that the restriction of U
onto the range of |T | is S, i.e.,

U|range (|T |) = S.
Note that null (|T |)⊥range (|T |), i.e., for v ∈ null (|T |) and w = |T |u ∈ range (|T |),
hw, vi = h|T |u, vi = hu, |T |vi = hu, 0i = 0
since |T | is self-adjoint.
Pick an orthonormal basis e = (e1 , . . . , em ) of null (|T |) and an orthonormal basis f =
(f1 , . . . , fm ) of (range (T ))⊥ . Set S̃ei = fi , and extend S̃ to all of null (|T |) by linearity. Since
null (|T |)⊥range (|T |), any v ∈ V can be uniquely written as v = v1 +v2 , where v1 ∈ null (|T |)
and v2 ∈ range (|T |) . Now define U : V → V by setting Uv = S̃v1 + Sv2 . Then U is an
isometry. Moreover, U is also unitary, as shown by the following calculation application of
the Pythagorean theorem:
kUvk2 = kS̃v1 + Sv2 k2 = kS̃v1 k2 + kSv2 k2

= kv1 k2 + kv2 k2 = kvk2 .
Also, note that U|T | = T by construction since U|null (|T |) is irrelevant.
11.7 Singular-value decomposition
The singular-value decomposition generalizes the notion of diagonalization. To unitarily

diagonalize T ∈ L(V ) means to find an orthonormal basis e such that T is diagonal with
respect to this basis, i.e.,
 
λ1 0
 .. 
M(T ; e, e) = [T ]e = 
 . ,

0 λn
where the notation M(T ; e, e) indicates that the basis e is used both for the domain and
codomain of T . The Spectral Theorem tells us that unitary diagonalization can only be
11.7. SINGULAR-VALUE DECOMPOSITION 157
done for normal operators. In general, we can find two orthonormal bases e and f such that
 
s1 0
 .. 
M(T ; e, f ) = 
 . ,

0 sn
which means that T ei = si fi even if T is not normal. The scalars si are called singular
values of T . If T is diagonalizable, then these are the absolute values of the eigenvalues.
Theorem 11.7.1. All T ∈ L(V ) have a singular-value decomposition. That is, there exist
orthonormal bases e = (e1 , . . . , en ) and f = (f1 , . . . , fn ) such that
T v = s1 hv, e1 if1 + · · · + sn hv, en ifn ,
where si are the singular values of T .
Proof. Since |T | ≥ 0, it is also also self-adjoint. Thus, by the Spectral Theorem, there is an
orthonormal basis e = (e1 , . . . , en ) for V such that |T |ei = si ei . Let U be the unitary matrix
in the polar decomposition of T . Since e is orthonormal, we can write any vector v ∈ V as
v = hv, e1 ie1 + · · · + hv, en ien ,
and hence
T v = U|T |v = s1 hv, e1 iUe1 + · · · + sn hv, en iUen .
Now set fi = Uei for all 1 ≤ i ≤ n. Since U is unitary, (f1 , . . . , fn ) is also an orthonormal
basis, proving the theorem.
1. Consider R3 with two orthonormal bases: the canonical basis e = (e1 , e2 , e3 ) and the
basis f = (f1 , f2 , f3 ), where
1 1 1
f1 = √ (1, 1, 1), f2 = √ (1, −2, 1), f3 = √ (1, 0, −1) .
3 6 2
Find the canonical matrix, A, of the linear map T ∈ L(R3 ) with eigenvectors f1 , f2 , f3
and eigenvalues 1, 1/2, −1/2, respectively.
2. For each of the following matrices, verify that A is Hermitian by showing that A = A∗ ,
find a unitary matrix U such that U −1 AU is a diagonal matrix, and compute exp(A).
" # " # " #

4 1−i 3 −i 6 2 + 2i
(a) A = (b) A = (c) A =
1+i 5 i 3 2 − 2i 4
 
  2 √i2 √ −i
" # 5 0 0 2 
0 3+i 
   −i
2 0 
(d) A = (e) A =  0 −1 −1 + i  (f) A =  √
2 
3 − i −3  
0 −1 − i 0 √i 0 2
2
3. For each of the following matrices, either find a matrix P (not necessarily unitary)
such that P −1 AP is a diagonal matrix, or show why no such matrix exists.
     
19 −9 −6 −1 4 −2 5 0 0
     
(a) A =  25 −11 −9  (b) A =  −3 4 0  (c) A =  1 5 0 
17 −9 −4 −3 1 3 0 1 5
     
0 0 0 −i 1 1 0 0 i
     
(d) A =  0 0 0  (e) A =  −i 1 1  (f) A =  4 0 i 
3 0 1 −i 1 1 0 0 i
11.7. SINGULAR-VALUE DECOMPOSITION 159
4. Let r ∈ R and let T ∈ L(C2 ) be the linear map with canonical matrix
!
1 −1
T = .
−1 r
(a) Find the eigenvalues of T .

(b) Find an orthonormal basis of C2 consisting of eigenvectors of T .
(c) Find a unitary matrix U such that UT U ∗ is diagonal.
5. Let A be the complex matrix given by:

 
5 0 0
 
A= 0 −1 −1 + i 
0 −1 − i 0
(a) Find the eigenvalues of A.

(b) Find an orthonormal basis of eigenvectors of A.
√
(c) Calculate |A| = A∗ A.
(d) Calculate eA .
6. Let θ ∈ R, and let T ∈ L(C2 ) have canonical matrix

!
1 eiθ
M(T ) = .
e−iθ −1
(a) Find the eigenvalues of T .

(b) Find an orthonormal basis for C2 that consists of eigenvectors for T .
1. Prove or give a counterexample: The product of any two self-adjoint operators on a
finite-dimensional vector space is self-adjoint.
2. Prove or give a counterexample: Every unitary matrix is invertible.

3. Let V be a finite-dimensional vector space over F, and suppose that T ∈ L(V ) satisfies
T 2 = T . Prove that T is an orthogonal projection if and only if T is self-adjoint.
4. Let V be a finite-dimensional inner product space over C, and suppose that T ∈ L(V )
has the property that T ∗ = −T . (We call T a skew Hermitian operator on V .)
(a) Prove that the operator iT ∈ L(V ) defined by (iT )(v) = i(T (v)), for each v ∈ V ,
is Hermitian.
(b) Prove that the canonical matrix for T can be unitarily diagonalized.
(c) Prove that T has purely imaginary eigenvalues.
5. Let V be a finite-dimensional vector space over F, and suppose that S, T ∈ L(V ) are
positive operators on V . Prove that S + T is also a positive operator on T .
6. Let V be a finite-dimensional vector space over F, and let T ∈ L(V ) be any operator
on V . Prove that T is invertible if and only if 0 is not a singular value of T .
Supplementary Notes on Matrices
and Linear Systems
As discussed in Chapter 1, there are many ways in which you might try to solve a system
of linear equation involving a finite number of variables. These supplementary notes are
intended to illustrate the use of Linear Algebra in solving such systems. In particular, any
arbitrary number of equations in any number of unknowns — as long as both are finite —
can be encoded as a single matrix equation. As you will see, this has many computational
advantages, but, perhaps more importantly, it also allows us to better understand linear
systems abstractly. Specifically, by exploiting the deep connection between matrices and so-
called linear maps, one can completely determine all possible solutions to any linear system.
These notes are also intended to provide a self-contained introduction to matrices and
important matrix operations. As you read the sections below, remember that a matrix is,
in general, nothing more than a rectangular array of real or complex numbers. Matrices are
not linear maps. Instead, a matrix can (and will often) be used to define a linear map.
12.1 From linear systems to matrix equations
We begin this section by reviewing the definition of and notation for matrices. We then
review several different conventions for denoting and studying systems of linear equations,
the most fundamental being as a single matrix equation. This point of view has a long
history of exploration, and numerous computational devices — including several computer
programming languages — have been developed and optimized specifically for analyzing
matrix equations.
161
162 CHAPTER 12. SUPPLEMENTARY NOTES
12.1.1 Definition of and notation for matrices

Let m, n ∈ Z+ be positive integers, and, as usual, let F denote either R or C. Then we begin
by defining an m × n matrix A to be a rectangular array of numbers
 
a11 · · · a1n  
 . . . 
A = (aij )m,n
i,j=1 = (A
(i,j) m,n
)i,j=1 = 
 .
. .. ..  m numbers


am1 · · · amn 
| {z }
n numbers
where each element aij ∈ F in the array is called an entry of A (specifically, aij is called
the “i, j entry”). We say that i indexes the rows of A as it ranges over the set {1, . . . , m}
and that j indexes the columns of A as it ranges over the set {1, . . . , n}. We also say that
the matrix A has size m × n and note that it is a (finite) sequence of doubly-subscripted
numbers for which the two subscripts in no way depend upon each other.
Definition 12.1.1. Given positive integers m, n ∈ Z+ , we use Fm×n to denote the set of all
m × n matrices having entries over F
" #
1 0 2
Example 12.1.2. The matrix A = ∈ C2×3 , but A ∈
/ R2×3 since the “2, 3” entry
−1 3 i
of A is not in R.
Given the ubiquity of matrices in both abstract and applied mathematics, a rich vo-
cabulary has been developed for describing various properties and features of matrices. In
addition, there is also a rich set of equivalent notations. For the purposes of these notes, we
will use the above notation unless the size of the matrix is understood from context or is
unimportant. In this case, we will drop much of this notation and denote a matrix simply
as
A = (aij ) or A = (aij )m×n .
To get a sense of the essential vocabulary, suppose that we have an m×n matrix A = (aij )
with m = n. Then we call A a square matrix. The elements a11 , a22 , . . . , ann in a square
matrix form the main diagonal of A, and the elements a1n , a2,n−1 , . . . , an1 form what is
sometimes called the skew main diagonal of A. Entries not on the main diagonal are
12.1. FROM LINEAR SYSTEMS TO MATRIX EQUATIONS 163
also often called off-diagonal entries, and a matrix whose off-diagonal entries are all zero is
called a diagonal matrix. It is common to call a12 , a23 , . . . , an−1,n the superdiagonal of A
and a21 , a32 , . . . , an,n−1 the subdiagonal of A. The motivation for this terminology should
be clear if you create a sample square matrix and trace the entries within these particular
subsequences of the matrix.
Square matrices are important because they are fundamental to applications of Linear
Algebra. In particular, virtually every use of Linear Algebra either involves square matrices
directly or employs them in some indirect manner. In addition, virtually every usage also
involves the notion of vector, where here we mean either an m × 1 matrix (a.k.a. a row
vector) or a 1 × n matrix (a.k.a. a column vector).
Example 12.1.3. Suppose that A = (aij ), B = (bij ), C = (cij ), D = (dij ), and E = (eij )
are the following matrices over F:
     
3 " # 1 5 2 6 1 3
4 −1 h i
     
A =  −1 , B = , C = 1, 4, 2 , D =  −1 0 1 , E =  −1 1 2 .
0 2
1 3 2 4 4 1 3
Then we say that A is a 3 × 1 matrix (a.k.a. a column vector), B is a 2 × 2 square matrix,

C is a 1 × 3 matrix (a.k.a. a row vector), and both D and E are square 3 × 3 matrices.
Moreover, only B is an upper-triangular matrix (as defined below), and none of the matrices
in this example are diagonal matrices.
We can discuss individual entries in each matrix. E.g.,
• the 2th row of D is d21 = −1, d22 = 0, and d23 = 1.
• the main diagonal of D is the sequence d11 = 1, d22 = 0, d33 = 4.
• the skew main diagonal of D is the sequence d13 = 2, d22 = 0, d31 = 3.
• the off-diagonal entries of D are (by row) d12 , d13 , d21 , d23 , d31 , and d32 .
• the 2th column of E is e12 = e22 = e32 = 1.
• the superdiagonal of E is the sequence e12 = 1, e23 = 2.
• the subdiagonal of E is the sequence e21 = −1, e32 = 1.

A square matrix A = (aij ) ∈ Fn×n is called upper triangular (resp. lower triangular)
if aij = 0 for each pair of integers i, j ∈ {1, . . . , n} such that i > j (resp. i < j). In other
words, A is triangular if it has the form
   
a11 a12 a13 · · · a1n a11 0 0 ··· 0
   
 0 a22 a23 · · · a2n   a21 a22 0 · · · 0 
   
 0 a33 · · · a3n   
 0  or  a31 a32 a33 · · · 0  .
 .. .. .. . . 
..   .. 
 . . . . .   ... ..
.
..
.
..
. . 
  
0 0 0 · · · ann an1 an2 an3 · · · ann
Note that a diagonal matrix is simultaneously both an upper triangular matrix and a lower
triangular matrix.
Two particularly important examples of diagonal matrices are defined as follows: Given
any positive integer n ∈ Z+ , we can construct the identity matrix In and the zero matrix
0n×n by setting
   
1 0 0 ··· 0 0 0 0 0 ··· 0 0
   
0 1 0 · · · 0 0 0 0 0 · · · 0 0
   
0 0 1 · · · 0 0 0 0 0 · · · 0 0
   
In =  . 
.. .. . . .. ..  and 0 n×n = . . . . . .. ..  ,
 .. . . . . .  .. .. .. . . .
  
   
0 0 0 · · · 1 0 0 0 0 · · · 0 0
0 0 0 ··· 0 1 0 0 0 ··· 0 0
where each of these matrices is understood to be a square matrix of size n × n. The zero
matrix 0m×n is analogously defined for any m, n ∈ Z+ and has size m × n. I.e.,
 
0 0 0 ··· 0 0  
 

0 0 0 · · · 0 0 
 


0 0 0 · · · 0 0 

0m×n = . .. .. . . .. ..  m rows
 .. . . . . . 

 
 

0 0 0 · · · 0 0 



0 0 0 ··· 0 0
| {z }
n columns
12.1.2 Using matrices to encode linear systems

Let m, n ∈ Z+ be positive integers. Then a system of m linear equations in n unknowns
x1 , . . . , xn looks like

a11 x1 + a12 x2 + a13 x3 + · · · + a1n xn = b1 




a21 x1 + a22 x2 + a23 x3 + · · · + a2n xn = b2 



a31 x1 + a32 x2 + a33 x3 + · · · + a3n xn = b3 , (12.3)


.. 

. 




am1 x1 + am2 x2 + am3 x3 + · · · + amn xn = bm
where each aij , bi ∈ F is a scalar for i = 1, 2, . . . , m and j = 1, 2, . . . , n. In other words, each

scalar b1 , . . . , bm ∈ F is being written as a linear combination of the unknowns x1 , . . . , xn
using coefficients from the field F. To solve System (12.3) means to describe the set of
all possible values for x1 , . . . , xn (when thought of as scalars in F) such that each of the m
equations in System (12.3) is satisfied simultaneously.
Rather than dealing directly with a given linear system, it is often convenient to first
encode the system using less cumbersome notation. Specifically, System (12.3) can be sum-
marized using exactly three matrices. First, we collect the coefficients from each equation
into the m × n matrix A = (aij ) ∈ Fm×n , which we call the coefficient matrix for the linear
system. Similarly, we assemble the unknowns x1 , x2 , . . . , xn into an n × 1 column vector
x = (xi ) ∈ Fn , and the right-hand sides b1 , b2 , . . . , bm of the equation are used to form an
m × 1 column vector b = (bi ) ∈ Fm . In other words,
     
a11 a12 ··· a1n x1 b1
     
 a21 a22 ··· a2n   x2   b2 
A=
 .. .. .. ..     
 , x =  ..  , and b =  ..  .
 . . . .  .  . 
am1 am2 · · · amn xn bm
Then the left-hand side of the ith equation in System (12.3) can be recovered by taking the
dot product (a.k.a. Euclidean inner product) of x with the ith row in A:
h i n
X
ai1 ai2 · · · ain · x = aij xj = ai1 x1 + ai2 x2 + ai3 x3 + · · · + ain xn .
j=1
In general, we can extend the dot product between two vectors in order to form the
product of any two matrices (as in Section 12.2.2). For the purposes of this section, though,
it suffices to simply define the product of the matrix A ∈ Fm×n and the vector x ∈ Fn to be
    
a11 a12 ··· a1n x1 a11 x1 + a12 x2 + · · · + a1n xn
    
 a21 a22 ··· a2n   x2   a21 x1 + a22 x2 + · · · + a2n xn 
Ax = 
 .. .. .. ..    
  ..  =  .. .
 (12.4)
 . . . .  .   . 
am1 am2 · · · amn xn am1 x1 + am2 x2 + · · · + amn xn
Then, since each entry in the resulting m × 1 column vector Ax ∈ Fm corresponds exactly to
the left-hand side of each equation in System 12.3, we have effectively encoded System (12.3)
as the single matrix equation
 
a11 x1 + a12 x2 + · · · + a1n xn  
  b 1
 a21 x1 + a22 x2 + · · · + a2n xn   . 
Ax = 
 ..  =  ..  = b.
   (12.5)
 . 
bm
am1 x1 + am2 x2 + · · · + amn xn
Example 12.1.4. The linear system


x1 + 6x2 + 4x5 − 2x6 = 14 

x3 + 3x5 + x6 = −3 .


x4 + 5x5 + 2x6 = 11
has three equations and involves the six variables x1 , x2 , . . . , x6 . One can check that possible
solutions to this system include
       
x1 14 x1 6
       
x2   0  x2   1 
       
x  −3 x  −9
 3    3  
  =   and   =   .
x4   11  x4  −5
       
x   0  x   2 
 6    6  
x6 0 x6 3
Note that, in describing these solutions, we have used the six unknowns x1 , x2 , . . . , x6 to
form the 6 × 1 column vector x = (xi ) ∈ F6 . We can similarly form the coefficient matrix
A ∈ F3×6 and the 3 × 1 column vector b ∈ F3 , where
     
1 6 0 0 4 −2 b1 14
     
A = 0 0 1 0 3 1  and b2  = −3 .
0 0 0 1 5 2 b3 11
You should check that, given these matrices, each of the solutions given above satisfies
Equation (12.5).
We close this section by mentioning another common conventions for encoding linear
systems. Specifically, rather than attempt to solve Equation (12.3) directly, one can instead
look at the equivalent problem of describing all coefficients x1 , . . . , xn ∈ F for which the
following vector equation is satisfied:
         
a11 a12 a13 a1n b1
         
 a21   a22   a23   a2n   b2 
         
a  a  a  a  b 
x1  31  + x2  32  + x3  33  + · · · + xn  3n  =  3  . (12.6)
 .   .   .   .   . 
 ..   ..   ..   ..   .. 
         
am1 am2 am3 amn bm
This approach emphasizes analysis of the so-called column vectors A(·,j) (j = 1, . . . , n) of

the coefficient matrix A in the matrix equation Ax = b. (See in Section 12.2 for more details
about how Equation (12.6) is formed). Conversely, it is also common to directly encounter
Equation (12.6) when studying certain questions about vectors in Fn .
It is important to note that System (12.3) differs from Equations (12.5) and (12.6) only in
terms of notation. The common aspect of these different representations is that the left-hand
side of each equation in System (12.3) is a linear sum. Because of this, it is also common to
rewrite System (12.3) using more compact notation such as
n
X n
X n
X n
X
a1k xk = b1 , a2k xk = b2 , a3k xk = b3 , . . . , amk xk = bm .
k=1 k=1 k=1 k=1
12.2 Matrix arithmetic

In this section, we examine algebraic properties of the set Fm×n (where m, n ∈ Z+ ). Specifi-
cally, Fm×n forms a vector space under the operations of component-wise addition and scalar
multiplication, and it is isomorphic to Fmn as a vector space.
We also define a multiplication operation between matrices of compatible size and show
that this multiplication operation interacts with the vector space structure on Fm×n in a
natural way. In particular, Fn×n forms an algebra over F with respect to these operations.
(See Section B.3 for the definition of an algebra.)
12.2.1 Addition and scalar multiplication

Let A = (aij ) and B = (bij ) be m × n matrices over F (where m, n ∈ Z+ ), and let α ∈ F.
Then matrix addition A+B = ((a+b)ij )m×n and scalar multiplication αA = ((αa)ij )m×n
are both defined component-wise, meaning
(a + b)ij = aij + bij and (αa)ij = αaij .
Equivalently, A + B is the m × n matrix given by

     
a11 · · · a1n b11 · · · b1n a11 + b11 · · · a1n + b1n
 . .. ..   . .. ..   .. .. .. 
 .. .   . .   ,
 . + . . = . . . 
am1 · · · amn bm1 · · · bmn am1 + bm1 · · · amn + bmn
and αA is the m × n matrix given by

   
a11 · · · a1n αa11 · · · αa1n
 . ..   .. 
α .. ..  =  ... ..
. 
 . .   . .
am1 · · · amn αam1 · · · αamn
Example 12.2.1. With notation as in Example 12.1.3,

 
7 6 5
 
D + E =  −2 1 3  ,
7 3 7
12.2. MATRIX ARITHMETIC 169
and no two other matrices from Example 12.1.3 can be added since their sizes are not
compatible. Similarly, we can make calculations like
   
−5 4 −1 0 0 0
   
D − E = D + (−1)E =  0 −1 −1  and 0D = 0E =  0 0 0  = 03×3 .
−1 1 1 0 0 0
It is important to note that, while these are not the only ways of defining addition
and scalar multiplication operations on Fm×n , the above operations have the advantage of
endowing Fm×n with a reasonably natural vector space structure. As a vector space, Fm×n
is seen to have dimension mn since we can build the standard basis matrices
E11 , E12 , . . . , E1n , E21 , E22 , . . . , E2n , . . . , Em1 , Em2 , . . . , Emn
by analogy to the standard basis for Fmn . That is, each Ekℓ = ((e(k,ℓ) )ij ) satisfies

1, if i = k and j = ℓ
(e(k,ℓ) )ij = .
0, otherwise
This allows us to build a vector space isomorphism Fm×n → Fmn using a bijection that
simply “lays each matrix out flat”. In other words, given A = (aij ) ∈ Fm×n ,
 
a11 · · · a1n
 . .. .. 
 .. . 
mn
 .  7→ (a11 , a12 , . . . , a1n , a21 , a22 , . . . , a2n , . . . , am1 , am2 , . . . , amn ) ∈ F .
am1 · · · amn
Example 12.2.2. The vector space R2×3 of 2 × 3 matrices over R has standard basis
( " # " # " #
1 0 0 0 1 0 0 0 1
E11 = , E12 = , E13 = ,
0 0 0 0 0 0 0 0 0
" # " # " #)
0 0 0 0 0 0 0 0 0
E21 = , E22 = , E23 = ,
1 0 0 0 1 0 0 0 1
which is seen to naturally correspond with the standard basis {e1 , . . . , e6 } for R6 , where
e1 = (1, 0, 0, 0, 0, 0), e2 = (0, 1, 0, 0, 0, 0), . . . , e6 = (0, 0, 0, 0, 0, 1).
Of course, it is not enough to just assert that Fm×n is a vector space since we have yet
to verify that the above defined operations of addition and scalar multiplication satisfy the
vector space axioms. The proof of the following theorem is straightforward and something
that you should work through for practice with matrix notation.
Theorem 12.2.3. Given positive integers m, n ∈ Z+ and the operations of matrix addition
and scalar multiplication as defined above, the set Fm×n of all m × n matrices satisfies each
of the following properties.
1. (associativity of matrix addition) Given any three matrices A, B, C ∈ Fm×n ,
(A + B) + C = A + (B + C).
2. (additive identity for matrix addition) Given any matrix A ∈ Fm×n ,
A + 0m×n = 0m×n + A = A.
3. (additive inverses for matrix addition) Given any matrix A ∈ Fm×n , there exists a
matrix −A ∈ Fm×n such that
A + (−A) = (−A) + A = 0m×n .
4. (commutativity of matrix addition) Given any two matrices A, B ∈ Fm×n ,
A + B = B + A.
5. (associativity of scalar multiplication) Given any matrix A ∈ Fm×n and any two scalars
α, β ∈ F,
(αβ)A = α(βA).
6. (multiplicative identity for scalar multiplication) Given any matrix A ∈ Fm×n and
denoting by 1 the multiplicative identity of F,
1A = A.
7. (distributivity of scalar multiplication) Given any two matrices A, B ∈ Fm×n and any
two scalars α, β ∈ F,
(α + β)A = αA + βA and α(A + B) = αA + αB.
In other words, Fm×n forms a vector space under the operations of matrix addition and scalar
multiplication.
As a consequence of Theorem 12.2.3, every property that holds for an arbitrary vector
space can be taken as a property of Fm×n specifically. We highlight some of these properties
in the following corollary to Theorem 12.2.3.
Corollary 12.2.4. Given positive integers m, n ∈ Z+ and the operations of matrix addition
and scalar multiplication as defined above, the set Fm×n of all m × n matrices satisfies each
of the following properties:
1. Given any matrix A ∈ Fm×n , given any scalar α ∈ F, and denoting by 0 the additive
identity of F,
0A = A and α0m×n = 0m×n .
2. Given any matrix A ∈ Fm×n and any scalar α ∈ F,
αA = 0 =⇒ either α = 0 or A = 0m×n .
3. Given any matrix A ∈ Fm×n and any scalar α ∈ F,
−(αA) = (−α)A = α(−A).
In particular, the additive inverse −A of A is given by −A = (−1)A, where −1 denoted

the additive inverse for the additivity identity of F.
While one could prove Corollary 12.2.4 directly from definitions, the point of recognizing
Fm×n as a vector space is that you get to use these results without worrying about their
proof. Moreover, there is no need to separate prove for each of Rm×n and Cm×n .
12.2.2 Multiplying and matrices
Let r, s, t ∈ Z+ be positive integers, A = (aij ) ∈ Fr×s be an r×s matrix, and B = (bij ) ∈ Fs×t
be an s × t matrix. Then matrix multiplication AB = ((ab)ij )r×t is defined by
s
X
(ab)ij = aik bkj .
k=1
In particular, note that the “i, j entry” of the matrix product AB involves a summation
over the positive integer k = 1, . . . , s, where s is both the number of columns in A and the
number of rows in B. Thus, this multiplication is only defined when the “middle” dimension
of each matrix is the same:
  

 a · · · a b · · · b1t 

 11 1s 11
 . . 
. . . .
(aij )r×s (bij )s×t = r  .
. . . ..  ..

. . ..  s


 
 a bs1 · · · bst 
r1 · · · ars
| {z } | {z }
s t
P Ps 
s 
a1k bk1 · · · k=1 a1k bkt  
 k=1 . .. .. 
= 

.. . . 
 r
Ps Ps 
k=1a brk k1 ··· k=1 a b 
rk kt
| {z }
t
Alternatively, if we let n ∈ Z+ be a positive integer, then another way of viewing matrix

multiplication is through the use of the standard inner product on Fn = F1×n = Fn×1 . In
particular, we define the dot product (a.k.a. Euclidean inner product) of the row vector
x = (x1j ) ∈ F1×n and the column vector y = (yi1 ) ∈ Fn×1 to be



h y
i  11  X n
 .. 
x · y = x11 , · · · , x1n ·  .  = x1k yk1 ∈ F.
k=1
yn1
We can then decompose matrices A = (aij )r×s and B = (bij )s×t into their constituent row
vectors by fixing a positive integer k ∈ Z+ and setting
h i h i
A(k,·) = ak1 , · · · , aks ∈ F1×s and B (k,·) = bk1 , · · · , bkt ∈ F1×t .
Similarly, fixing ℓ ∈ Z+ , we can also decompose A and B into the column vectors
   
a1ℓ b1ℓ
 .   . 
A (·,ℓ) 
=  .. ∈F
r×1
and B (·,ℓ) 
=  ..  s×1
∈F .
arℓ bsℓ
It follows that the product AB is the following matrix of dot products:

 
A(1,·) · B (·,1) · · · A(1,·) · B (·,t)
 .. .. .. 
AB = 
 . . .  ∈ Fr×t .

(r,·) (·,1) (r,·) (·,t)
A ·B ··· A ·B
Example 12.2.5. With notation as in Example 12.1.3, you should sit down and use the
above definitions in order to verify that the following matrix products hold.
   
3 h i 3 12 6
   
AC =  −1  1, 4, 2 =  −1 −4 −2  ∈ F3×3 ,
1 1 4 2
 
h i 3
 
CA = 1, 4, 2 ·  −1  = 3 − 4 + 2 = 1 ∈ F,
1
" #" # " #
2 4 −1 4 −1 16 −6
B = BB = = ∈ F2×2 ,
0 2 0 2 0 4
 
h i 6 1 3
h i
  1×3
CE = 1, 4, 2  −1 1 2  = 10, 7, 17 ∈ F , and
4 1 3
    
1 5 2 3 0
    
DA =  −1 0 1   −1  =  −2  ∈ F3×1 .
3 2 4 1 11
Note, though, that B cannot be multiplied by any of the other matrices, nor does it make
sense to try to form the products AD, AE, DC, and EC due to the inherent size mismatches.
As illustrated in Example 12.2.5 above, matrix multiplication is not a commutative op-
eration (since, e.g., AC ∈ F3×3 while CA ∈ F1×1 ). Nonetheless, despite the complexity of its
definition, the matrix product otherwise satisfies many familiar properties of a multiplication
operation. We summarize the most basic of these properties in the following theorem.
Theorem 12.2.6. Let r, s, t, u ∈ Z+ be positive integers.
1. (associativity of matrix multiplication) Given A ∈ Fr×s , B ∈ Fs×t , and C ∈ Ft×u ,
A(BC) = (AB)C.
2. (distributivity of matrix multiplication) Given A ∈ Fr×s , B, C ∈ Fs×t , and D ∈ Ft×u ,
A(B + C) = AB + AC and (B + C)D = BD + CD.
3. (compatibility with scalar multiplication) Given A ∈ Fr×s , B ∈ Fs×t , and α ∈ F,
α(AB) = (αA)B = A(αB).
Moreover, given any positive integer n ∈ Z+ , Fn×n is an algebra over F.

As with Theorem 12.2.3, you should work through a proof of each part of Theorem 12.2.6 (and
especially of the first part) in order to practice manipulating the indices of entries correctly.
We state and prove a useful followup to Theorems 12.2.3 and 12.2.6 as an illustration.
Theorem 12.2.7. Let A, B ∈ Fn×n be upper triangular matrices and c ∈ R be any real
scalar. Then each of the following properties hold:
1. cA is upper triangular.
2. A + B is upper triangular.
3. AB is upper triangular.
In other words, the set of all m × n upper triangular matrices forms an algebra over F.
Moreover, each of the above statements still holds when upper triangular is replaced by
lower triangular.
Proof. The proofs of Parts 1 and 2 are straightforward and follow directly form the appro-
priate definitions. Moreover, the proof of the case for lower triangular matrices follows from
the fact that a matrix A is upper triangular if and only if AT is lower triangular, where AT
denotes the transpose of A. (See Section 12.5.1 for the definition of transpose.)
To prove Part 3, we start from the definition of the matrix product. Denoting A = (aij )
and B = (bij ), note that AB = ((ab)ij ) is an n × n matrix having “i-j entry” given by
n
X
(ab)ij = aik bkj .
k=1
Since A and B are upper triangular, we have that aik = 0 when i > k and that bkj = 0
when k > j. Thus, to obtain a non-zero summand aik bkj 6= 0, we must have both aik 6= 0,
which implies that i ≤ k, and bkj 6= 0, which implies that k ≤ j. In particular, these two
conditions are simultaneously satisfiable only when i ≤ j. Therefore, (ab)ij = 0 when i > j,
from which AB is upper triangular.
At the same time, you should be careful not to blithely perform operations on matrices as
you would with numbers. The fact that matrix multiplication is not a commutative operation
should make it clear that significantly more care is required with matrix arithmetic. As
another example, given a positive integer n ∈ Z+ , the set Fn×n has what are called zero
divisors. That is, there exist non-zero matrices A, B ∈ Fn×n such that AB = 0n×n :
" #2 " #" # " #
0 1 0 1 0 1 0 0
= = = 02×2 .
0 0 0 0 0 0 0 0
Moreover, note that there exist matrices A, B, C ∈ Fn×n such that AB = AC but B 6= C:
" #" # " #" #
0 1 1 0 0 1 0 1
= 02×2 = .
0 0 0 0 0 0 0 0
As a result, we say that the set Fn×n fails to have the so-called cancellation property.
This failure is a direct result of the fact that there are non-zero matrices in Fn×n that have
no multiplicative inverse. We discuss matrix invertibility at length in the next section and
define a special subset GL(n, F) ⊂ Fn×n upon which the cancellation property does hold.
12.2.3 Invertibility of square matrices

In this section, we explore the opposite of matrix multiplication. More specifically, we
characterize square matrices for which multiplicative inverses exist.
Definition 12.2.8. Given a positive integer n ∈ Z+ , we say that the square matrices
A ∈ Fn×n is invertible (a.k.a. nonsingular) if there exists a square matrix B ∈ Fn×n
such that
AB = BA = In .
We use GL(n, F) to denote the set of all invertible n × n matrices having entries from F.
One can prove that, if the multiplicative inverse of a matrix exists, then the inverse is
unique. As such, we will usually denote the so-called inverse matrix of A ∈ GL(n, F) by
A−1 . Even though this notation is analogous to the notation for the multiplicative inverse
of a scalar, you should not take this to mean that it is possible to “divide” by a matrix.
Moreover, note that the zero matrix 0n×n ∈/ GL(n, F). This means that GL(n, F) is not a
vector subspace of Fn×n .
Since matrix multiplication is not a commutative operation, care must when working
with the multiplicative inverses of invertible matrices. In particular, many of the algebraic
properties for multiplicative inverses of scalars, when properly modified, continue to hold.
We summarize the most basic of these properties in the following theorem.
Theorem 12.2.9. Let n ∈ Z+ be a positive integer and A, B ∈ GL(n, F). Then
1. the inverse matrix A−1 ∈ GL(n, F) and satisfies (A−1 )−1 = A.

2. the matrix power Am ∈ GL(n, F) and satisfies (Am )−1 = (A−1 )m , where m ∈ Z+ is
any positive integer.
3. the matrix αA ∈ GL(n, F) and satisfies (αA)−1 = α−1 A−1 , where α ∈ F is any non-zero
scalar.
4. the product AB ∈ GL(n, F) and has inverse given by the formula
(AB)−1 = B −1 A−1 .
Moreover, GL(n, F) has the cancellation property. In other words, given any three ma-
trices A, B, C ∈ GL(n, F), if AB = AC, then B = C.
At the same time, it is important to note that the zero matrix is not the only non-
invertible matrix. As an illustration of the subtlety involved in understanding invertibility,
we give the following theorem for the 2 × 2 case.
" #
a11 a12
Theorem 12.2.10. Let A = ∈ F2×2 . Then A is invertible if and only if A
a21 a22
satisfies
a11 a22 − a12 a21 6= 0.
Moreover, if A is invertible, then

 a22 −a12 
 a11 a22 − a12 a21 a11 a22 − a12 a21 
 
A−1 = .
 −a21 a11 
a11 a22 − a12 a21 a11 a22 − a12 a21
A more general theorem holds for larger matrices, but its statement requires substantially
more machinery than could reasonably be included here. We nonetheless state this result
for completeness and refer the reader to Chapter 8 for the definition of the determinant.
Theorem 12.2.11. Let n ∈ Z+ be a positive integer, and let A = (aij ) ∈ Fn×n be an n × n

matrix. Then A is invertible if and only if det(A) 6= 0. Moreover, if A is invertible, then the
“i, j entry” of A−1 is Aij / det(A). Here, Aij = (−1)i+j Mij , and Mij is the determinant of
the matrix obtained when both the ith row and j th column are removed from A.
We close this section by noting that the set GL(n, F) of all invertible n × n matrices
over F is often called the general linear group. This set has so many important uses in
mathematics that there are many equivalent notations for it, including GLn (F) and GL(Fn ),
and sometimes simply GL(n) or GLn if it is not important to emphasis the dependence
on F. Note, moreover, that the usage of the term “group” in the name “general linear
group” is highly technical. This is because GL(n, F) forms a nonabelian group under matrix
multiplication. (See Section B.2 for the definition of a group.)
12.3 Solving linear systems by factoring the coefficient

matrix
There are many ways in which you might try to solve a given system of linear equations.
This section is primarily devoted to describing two particularly popular techniques, both
of which involve factoring the coefficient matrix for the system into a product of simpler
matrices. These techniques are also at the heart of many frequently used numerical (i.e.,
computer-assisted) applications of Linear Algebra.
You should note that the factorization of complicated objects into simpler components is
an extremely common problem solving technique in mathematics. E.g., we will often factor
a given polynomial into several polynomials of lower degree, and one can similarly use the
prime factorization for an integer in order to simply certain numerical computations.
12.3.1 Factorizing matrices using gaussian elimination

In this section, we discuss a particularly significant factorization for matrices known as
Gaussian elimination (a.k.a. Gauss-Jordan elimination). Gaussian elimination can
be used to express any matrix as a product involving one matrix in so-called reduced
row-echelon form and one or more so-called elementary matrices. Then, once such
a factorization has been found, we can immediately solve any linear system that has the
factorized matrix as its coefficient matrix. Moreover, the underlying technique for arriving
at such a factorization is essentially an extension of the techniques already familiar to you
for solving small systems of linear equations by hand.
Let m, n ∈ Z+ denote positive integers, and suppose that A ∈ Fm×n is an m × n matrix
12.3. SOLVING LINEAR SYSTEMS BY FACTORING THE COEFFICIENT MATRIX179
over F. Then, following Section 12.2.2, we will make extensive use of A(i,·) and A(·,j) in
denote the row vectors and column vectors of A, respectively.
Definition 12.3.1. Let A ∈ Fm×n be an m × n matrix over F. Then we say that A is in

row-echelon form (abbreviated REF) if the rows of A satisfy the following conditions:
(1) either A(1,·) is the zero vector or the first non-zero entry in A(1,·) (when read from left
to right) is a one.
(2) for i = 1, . . . , m, if any row vector A(i,·) is the zero vector, then each subsequent row
vector A(i+1,·) , . . . , A(m,·) is also the zero vector.
(3) for i = 2, . . . , m, if some A(i,·) is not the zero vector, then the first non-zero entry (when
read from left to right) is a one and occurs to the right of the initial one in A(i−1,·) .
The initial leading one in each non-zero row is called a pivot. We furthermore say that A
is in reduced row-echelon form (abbreviated RREF) if
(4) for each column vector A(·,j) containing a pivot (j = 2, . . . , n), the pivot is the only
non-zero element in A(·,j).
The motivation behind Definition 12.3.1 is that matrix equations having their coefficient
matrix in RREF (and, in some sense, also REF) are particularly easy to solve. Note, in
particular, that the only square matrix in RREF without zero rows is the identity matrix.
Example 12.3.2. The following matrices are all in REF:

       
1 1 1 1 1 1 1 0 1 1 0 1 1 1 0 0
       
A1 = 0 1 1 1 , A2 = 0 1 1 0 , A3 = 0 1 1 0 , A4 = 0 0 1 0 ,
0 0 1 1 0 0 1 0 0 0 0 1 0 0 0 1
       
1 0 1 0 0 0 1 0 0 0 0 1 0 0 0 0
       
A5 = 0 0 0 1 , A6 = 0 0 0 1 , A7 = 0 0 0 0 , A8 = 0 0 0 0 .
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
However, only A4 through A8 are in RREF, as you should verify. Moreover, if we take the
transpose of each of these matrices (as defined in Section 12.5.1), then only AT6 , AT7 , and
AT8 are in RREF.
Example 12.3.3.
1. Consider the following matrix in RREF:

 
1 0 0 0
 
0 1 0 0

A= .
0 0 1 0
 
0 0 0 1
Given any vector b = (bi ) ∈ F4 , the matrix equation Ax = b corresponds to the system
of equations 
x1 = b1 



x2 = b2 
.
x3 = b3 



x4 = b4 
Since A is in RREF (in fact, A = I4 is the 4 × 4 identity matrix), we can immediately
conclude that the matrix equation Ax = b has the solution x = b for any choice of b.
Moreover, as we will see in Section 12.4.2, x = b is the only solution to this system.
2. Consider the following matrix in RREF:

 
1 6 0 0 4 −2
 
0 0 1 0 3 1 
A=
0 0 0 1 5
.

 2 
0 0 0 0 0 0
Given any vector b = (bi ) ∈ F4 , the matrix equation Ax = b corresponds to the system
of equations

x1 + 6x2 + 4x5 − 2x6 = b1 



x3 + 3x5 + x6 = b2 
.
x4 + 5x5 + 2x6 = b3 



0 = b4 
Since A is in RREF, we can immediately conclude a number of facts about solutions

to this system. First of all, solutions exist if and only if b4 = 0. Moreover, by “solving
for the pivots”, we see that the system reduces to

x1 = b1 −6x2 − 4x5 + 2x6 

x3 = b2 − 3x5 − x6 ,


x4 = b3 − 5x5 − 2x6
and so there is only enough information to specify values for x1 , x3 , and x4 in terms
of the otherwise arbitrary values for x2 , x5 , and x6 .
In this context, x1 , x3 , and x4 are called leading variables since these are the variable
corresponding to the pivots in A. We similarly call x2 , x5 , and x6 free variables since
the leading variables have been expressed in terms of these remaining variable. In
particular, given any scalars α, β, γ ∈ F, it follows that the vector
           
x1 b1 − 6α − 4β + 2γ b1 −6α 4β 2γ
           
x2   α  0  α   0   0 
           
x   b − 3β  b   0  −3β   0 
 3  2   2      
x= = = + + + 
x4   b3 − 5β − 2γ  b3   0  −5β  −2γ 
           
x   β  0  0   β   0 
 6          
x6 γ 0 0 0 γ
must satisfy the matrix equation Ax = b. One can also verify that every solution to
the matrix equation must be of this form. In then follows that the set of all solutions
should somehow be “three dimensional”.
As the above examples illustrate, a matrix equation having coefficient matrix in RREF
corresponds to a system of equations that can be solved with only a small amount of com-
putation. Somewhat amazingly, any matrix can be factored into a product that involves
exactly one matrix in RREF and one or more of the matrices defined as follows.
Definition 12.3.4. A square matrix E ∈ Fm×m is called an elementary matrix if it has

one of the following forms:
1. (row exchange, a.k.a. “row swap”, matrix) E is obtained from the identity matrix Im
(r,·) (s,·)
by interchanging the row vectors Im and Im for some particular choice of positive
integers r, s ∈ {1, 2, . . . , m}. I.e., in the case that r < s,

 
1 0 0 ··· 0 0 0 ··· 0 0 0 ··· 0
 
 0 1 0 ··· 0 0 0 ··· 0 0 0 ··· 0 
 

 0 0 1 ··· 0 0 0 ··· 0 0 0 ··· 0  
 .. .. .. . . .. .. .. . . .. .. .. . . .. 

 . . . . . . . . . . . . . 

 
 0 0 0 ··· 1 0 0 ··· 0 0 0 ··· 0 
 

 0 0 0 ··· 0 0 0 ··· 0 1 0 ··· 0  th
 ← r row
 0 0 0 ··· 0 0 1 ··· 0 0 0 ··· 0  
E=
 .. .. .. . . .. .. .. . . .. .. .. . . .. 

 . . . . . . . . . . . . . 

 

 0 0 0 ··· 0 0 0 ··· 1 0 0 ··· 0  

 0 0 0 ··· 0 1 0 ··· 0 0 0 ··· 0  th
 ← s row.
 
 0 0 0 ··· 0 0 0 ··· 0 0 1 ··· 0 
 
 .. .. .. . . .. .. .. . . .. .. .. . . .. 
 . . . . . . . . . . . . . 
0 0 0 ··· 0 0 0 ··· 0 0 0 ··· 1
2. (row scaling matrix) E is obtained from the identity matrix Im by replacing the row
(r,·) (r,·)
vector Im with αIm for some choice of non-zero scalar 0 6= α ∈ F and some choice
of positive integer r ∈ {1, 2, . . . , m}. I.e.,
 
1 0 ··· 0 0 0 ··· 0
 
0 1 · · · 0 0 0 · · · 0
 
 .. .. . . .. .. .. . . .. 
. . . . . . . .
 
0 0 · · · 1 0 0 · · · 0
 
E = Im + (α − 1)Err = 
0 0 · · · 0 α 0 · · · 0 th
  ← r row,
0 0 · · · 0 0 1 · · · 0
 
. .. . . .. .. .. . . .. 
 .. . . . . . . .
 
0 0 ··· 0 0 0 ··· 1
where Err is the matrix having “r, r entry” equal to one and all other entries equal to
zero. (Recall that Err was defined in Section 12.2.1 as a standard basis vector for the
vector space Fm×m .)
3. (row combination, a.k.a. “row sum”, matrix) E is obtained from the identity matrix
(r,·) (r,·) (s,·)
Im by replacing the row vector Im with Im + αIm for some choice of scalar α ∈ F
and some choice of positive integers r, s ∈ {1, 2, . . . , m}. I.e., in the case that r < s,
 
1 0 0 ··· 0 0 0 ··· 0 0 0 ··· 0
 
 0 1 0 ··· 0 0 0 ··· 0 0 0 ··· 0 
 
 0 0 1 ··· 0 0 0 ··· 0 0 0 ··· 0 
 
 .. .. .. . . .. .. .. . . .. .. .. . . .. 
 . . . . . . . . . . . . . 
 
 0 0 0 ··· 1 0 0 ··· 0 0 0 ··· 0 
 
 
 0 0 0 ··· 0 1 0 ··· 0 α 0 ··· 0  ← r th row
 
 0 0 0 ··· 0 0 1 ··· 0 0 0 ··· 0 
E = Im + αErs =  
 .. .. .. . . .. .. .. . . .. .. .. . . .. 
 . . . . . . . . . . . . . 
 
 
 0 0 0 ··· 0 0 0 ··· 1 0 0 ··· 0 
 
 0 0 0 ··· 0 0 0 ··· 0 1 0 ··· 0 
 
 
 0 0 0 ··· 0 0 0 ··· 0 0 1 ··· 0 
 
 .. .. .. . . .. .. .. . . .. .. .. . . .. 
 . . . . . . . . . . . . . 
0 0 0 ··· 0 0 0 ··· 0 0 0 ··· 1
↑
sth column
where Ers is the matrix having “r, s entry” equal to one and all other entries equal to
zero. (Ers was also defined in Section 12.2.1 as a standard basis vector for Fm×m .)
The “elementary” in the name “elementary matrix” comes from the correspondence be-
tween these matrices and so-called “elementary operations” on systems of equations. In
particular, each of the elementary matrices is clearly invertible (in the sense defined in Sec-
tion 12.2.3), just as each “elementary operation” is itself completely reversible. We illustrate
this correspondence in the following example.
Example 12.3.5. Define A, x, and b by

  
  
2 5 3 x1 4
     
A = 1 2 3 , x = x2  , and b = 5 .
1 0 8 x3 9
We illustrate the correspondence between elementary matrices and “elementary” operations

on the system of linear equations corresponding to the matrix equation Ax = b, as follows.
System of Equations Corresponding Matrix Equation
2x1 + 5x2 + 3x3 = 5

x1 + 2x2 + 3x3 = 4 Ax = b
x1 + 8x3 = 9
To begin solving this system, one might want to either multiply the first equation through
by 1/2 or interchange the first equation with one of the other equations. From a computa-
tional perspective, it is preferable to perform an interchange since multiplying through by
1/2 would unnecessarily introduce fractions. Thus, we choose to interchange the first and
second equation in order to obtain

 
0 1 0
x1 + 2x2 + 3x3 = 4  
E0 Ax = E0 b, where E0 = 1 0 0 .
2x1 + 5x2 + 3x3 = 5
0 0 1
x1 + 8x3 = 9
Another reason for performing the above interchange is that it now allows us to use more
convenient “row combination” operations when eliminating the variable x1 from all but one
of the equations. In particular, we can multiply the first equation through by −2 and add it
to the second equation in order to obtain
 
x1 + 2x2 + 3x3 = 4 1 0 0
 
x2 − 3x3 = −3 E1 E0 Ax = E1 E0 b, where E1 = −2 1 0 .
x1 + 8x3 = 9 0 0 1
Similarly, in order to eliminate the variable x1 from the third equation, we can next multiply
the first equation through by −1 and add it to the third equation in order to obtain
 
x1 + 2x2 + 3x3 = 4 1 0 0
 
x2 − 3x3 = −3 E2 E1 E0 Ax = E2 E1 E0 b, where E2 =  0 1 0 .
−2x2 + 5x3 = 5 −1 0 1
Now that the variable x1 only appears in the first equation, we can somewhat similarly iso-
late the variable x2 by multiplying the second equation through by 2 and adding it to the
third equation in order to obtain
 
x1 + 2x2 + 3x3 = 4 1 0 0
 
x2 − 3x3 = −3 E3 · · · E0 Ax = E3 · · · E0 b, where E3 = 0 1 0 .
−x3 = −1 0 2 1
Finally, in order to complete the process of transforming the coefficient matrix into REF,
we need only rescale row three by −1. This corresponds to multiplying the third equation
through by −1 in order to obtain
 
x1 + 2x2 + 3x3 = 4 1 00
 
x2 − 3x3 = −3 E4 · · · E0 Ax = E4 · · · E0 b, where E4 = 0 10 .
x3 = 1 0 0 −1
Now that the coefficient matrix is in REF, we can alreAdy solve for the variables x1 , x2 , and
x3 using a process called back substitution. In other words, it should be clear from the
third equation that x3 = 1. Using this value and solving for x2 in the second equation, it
then follows that
x2 = −3 + 3x3 = −3 + 3 = 0.
Similarly, by solving the first equation for x1 , it follows that
x1 = 4 − 2x2 − 3x3 = 4 − 3 = 1.
From a computational perspective, this process of back substitution can be applied to

solve any system of equations when the coefficient matrix of the corresponding matrix equa-
tion is in REF. However, from an algorithmic perspective, it is often more useful to continue
“row reducing” the coefficient matrix in order to produce a coefficient matrix in full RREF.
Here, there are several next natural steps that we could now perform in order to move
toward RREF. Since we have so far worked “from the top down, from left to right”, we
choose to now work “from bottom up, from right to left”. In other words, we now multi-
ply the third equation through by 3 and then add it to the second equation in order to obtain
 
x1 + 2x2 + 3x3 = 4 1 0 0
 
x2 = 0 E5 · · · E0 Ax = E5 · · · E0 b, where E5 = 0 1 3 .
x3 = 1 0 0 1
Next, we can multiply the third equation through by −3 and add it to the first equation in
order to obtain
x1 + 2x2 = 1  
1 0 −3
x2 = 0  
E6 · · · E0 Ax = E6 · · · E0 b, where E6 = 0 1 0  .
x3 = 1
0 0 1
Finally, we can multiply the second equation through by −2 and add it to the first equation
in order to obtain
x1 = 1  
1 −2 0
x2 = 0  
E7 · · · E0 Ax = E7 · · · E0 b, where E7 = 0 1 0 .
x3 = 1
0 0 1
Now it should be extremely clear that we obtained a correct solution when using back
substitution on the linear system
E4 · · · E0 Ax = E4 · · · E0 b.
However, in many applications, it is not enough to merely find a solution. Instead, it is

important to describe every solution. As we will see in the remaining sections of these notes,
this requires the machinery of Linear Algebra. In particular, we will need the theory of
vector spaces and linear maps.
To close this section, we take a closer look at the following expression obtained from the
above analysis:
E7 E6 · · · E1 E0 A = I3 .
Because of the way in which we have defined elementary matrices, it should be clear that
each of the matrices E0 , E1 , . . . , E7 is invertible. Thus, we can use Theorem 12.2.9 in order
to “solve” for A:
A = (E7 E6 · · · E1 E0 )−1 I3 = E0−1 E1−1 · · · E7−1 I3 .
In effect, since the inverse of an elementary matrix is itself easily seen to be an elemen-
tary matrix, this has factored A into the product of eight elementary matrices (namely,
E0−1 , E1−1 , . . . , E7−1 ) and one matrix in RREF (namely, I3 ). Moreover, because each elemen-
tary matrix is invertible, we can conclude that x solves Ax = b if and only if x solves
(E7 E6 · · · E1 E0 A) x = (I3 ) x = (E7 E6 · · · E1 E0 ) b.
Consequently, given any linear system, one can use Gaussian elimination in order to reduce
the problem to solving a linear system whose coefficient matrix is in RREF.
Similarly, we can conclude that the inverse of A is
 
13 −5 −3
 
A−1 = E7 E6 · · · E1 E0 = −40 16 9  .
5 −2 1
Having computed this product, one could essentially “reuse” much of the above computation
in order to solve the matrix equation Ax = b′ for several different right-hand sides b′ ∈ F3 .
The process of “resolving” a linear system is a common practice in applied mathematics.
12.3.2 Solving homogenous linear systems

In this section, we study the solutions for an important special class of linear systems. As we
will see in the next section, though, solving any linear system is fundamentally dependent
upon knowing how to solve these so-called homogeneous systems.
As usual, we use m, n ∈ Z+ to denote arbitrary positive integers.
Definition 12.3.6. The system of linear equations, System (12.3), is called a homogeneous
system if the right-hand side of each equation is zero. In other words, a homogeneous system
corresponds to a matrix equation of the form
Ax = 0,
where A ∈ Fm×n is an m × n matrix and x is an n-tuple of unknowns. We also call the set
N = {v ∈ Fn | Av = 0}
the solution space for the homogeneous system corresponding to Ax = 0.
When describing the solution space for a homogeneous linear system, there are three
important cases to keep in mind:
Definition 12.3.7. The system of linear equations System (12.3) is called
1. overdetermined if m > n.
2. square if m = n.
3. underdetermined if m < n.
In particular, we can say a great deal about underdetermined homogeneous systems, which
we state as a corollary to the following more general result.
Theorem 12.3.8. Let N be the solution space for the homogeneous linear system corre-
sponding to the matrix equation Ax = 0, where A ∈ Fm×n . Then
1. the zero vector 0 ∈ N.
2. N is a subspace of the vector space Fn .

This is an amazing theorem. Since N is a subspace of Fn , we know that either N will contain
exactly one element (namely, the zero vector) or N will contain infinitely many elements.
Corollary 12.3.9. Every homogeneous system of linear equations is solved by the zero vector.
Moreover, every underdetermined homogeneous system has infinitely many solution.
We call the zero vector the trivial solution for a homogeneous linear system. The fact that
every homogeneous linear system has the trivial solution thus reduces solving such a system
to determining if solutions other than the trivial solution exist.
One method for finding the solution space of a homogeneous system is to first use Gaus-
sian elimination (as demonstrated in Example 12.3.5) in order to factor the coefficient matrix
of the system. Then, because the original linear system is homogeneous, the homogeneous
system corresponding to the resulting RREF matrix will have the same solutions as the
original system. In other words, if a given matrix A satisfies
Ek Ek−1 · · · E0 A = A0 ,
where each Ei an elementary matrix and A0 is an RREF matrix, then the matrix equation
Ax = 0 has the exact same solution set as A0 x = 0 since E0−1 E1−1 · · · Ek−1 0 = 0.
Example 12.3.10. In the following examples, we illustrate the process of determining the
solution space for a homogeneous linear system having coefficient matrix in RREF.
1. Consider the matrix equation Ax = 0, where A is the matrix given by

 
1 0 0
 
0 1 0
A=
0
.
 0 1

0 0 0
This corresponds to an overdetermined homogeneous system of linear equations. More-

over, since there are no free variables (as defined in Example 12.3.3), it should be clear
that this system has only the trivial solution. Thus, N = {0}.

 
1 0 1
 
0 1 1
A=
0
.
 0 0

0 0 0
This corresponds to an overdetermined homogeneous system of linear equations. Un-

like the above example, we see that x3 is a free variable for this system, and so we
would expect the solution space to contain more than just the zero vector. As in Ex-
ample 12.3.3, we can solve for the leading variables in terms of the free variable in
order to obtain )
x1 = − x3
,
x2 = − x3
It follows that, given any scalar α ∈ F, every vector of the form
     
x1 −α −1
     
x = x2  = −α = α −1
x3 α 1
is a solution to Ax = 0. Therefore,

N = (x1 , x2 , x3 ) ∈ F3 | x1 = −x3 , x2 = −x3 = span ((−1, −1, 1)) .

 
1 1 1
 
A = 0 0 0 .
0 0 0
This corresponds to a square homogeneous system of linear equations with two free
variables. Thus, using the same technique as in the previous example, we can solve
for the leading variable in order to obtain x1 = −x2 − x3 . It follows that, given any
scalars α, β ∈ F, every vector of the form

       
x1 −α − β −1 −1
       
x = x2  =  α  = α  1  + β  0 
x3 β 0 1
is a solution to Ax = 0. Therefore,

N = (x1 , x2 , x3 ) ∈ F3 | x1 + x2 + x3 = 0 = span ((−1, 1, 0), (−1, 0, 1)) .
12.3.3 Solving inhomogeneous linear systems

In this section, we demonstrate the relationship between arbitrary linear systems and ho-
mogeneous linear systems. Specifically, we will see that it takes little more work to solve a
general linear system than it does to solve the homogeneous system associated to it.
As usual, we use m, n ∈ Z+ to denote arbitrary positive integers.
Definition 12.3.11. The system of linear equations System (12.3) is called an inhomoge-
neous system if the right-hand side of at least one equation is not zero. In other words, an
inhomogeneous system corresponds to a matrix equation of the form
Ax = b,
where A ∈ Fm×n is an m × n matrix, x is an n-tuple of unknowns, and b ∈ Fm is a vector

having at least one non-zero component. We also call the set
U = {v ∈ Fn | Av = b}
the solution set for the linear system corresponding to Ax = b.
As illustrated in Example 12.3.3, the zero vector cannot be a solution for an inhomo-
geneous system. Consequently, the solution set U for an inhomogeneous linear system will
never be a subspace of any vector space. Instead, it will be a related algebraic structure as
described in the following theorem.
Theorem 12.3.12. Let U be the solution space for the inhomogeneous linear system corre-
sponding to the matrix equation Ax = b, where A ∈ Fm×n and b ∈ Fm is a vector having at
least one non-zero component. Then, given any element u ∈ U, we have that
U = u + N = {u + n | n ∈ N} ,
where N is the solution space to Ax = 0. In other words, if B = (n(1) , n(2) , . . . , n(k) ) is a list
of vectors forming a basis for N, then every element of U can be written in the form
u + α1 n(1) + α2 n(2) + . . . + αk n(k)
for some choice of scalars α1 , α2 , . . . , αk ∈ F.
As a consequence of this theorem, we can conclude that inhomogeneous linear systems behave
a lot like homogeneous systems. The main difference is that inhomogeneous systems are not
necessarily solvable. This, then, creates three possibilities: an inhomogeneous linear system
will either have no solution, a unique solution, or infinitely many solutions. An important
special case is as follows.
Corollary 12.3.13. Every overdetermined inhomogeneous linear system will necessarily be

unsolvable for some choice of values for the right-hand sides of the equations.
The solution set U for an inhomogeneous linear system is called an affine subspace
of Fn since it is a genuine subspace of Fn that has been “offset” by a vector u ∈ Fn . Any
set having this structure might also be called a coset (when used in the context of Group
Theory) or a linear manifold (when used in a geometric context such as a discussion of
intersection hyperplanes).
In order to actually find the solution set for an inhomogeneous linear system, we rely on
Theorem 12.3.12. Given an m × n matrix A ∈ Fm×n and a non-zero vector b ∈ Fm , we call
Ax = 0 the associated homogeneous matrix equation to the inhomogeneous matrix
equation Ax = b. Then, according to Theorem 12.3.12, U can be found by first finding
the solution space N for the associated equation Ax = 0 and then finding any so-called
particular solution u ∈ Fn to Ax = b.
As with homogeneous systems, one can first use Gaussian elimination in order to factorize
A, and so we restrict the following examples to the special case of RREF matrices.
Example 12.3.14. The following examples use the same matrices as in Example 12.3.10.
1. Consider the matrix equation Ax = b, where A is the matrix given by

 
1 0 0
 
0 1 0
A=
0

 0 1

0 0 0
and b ∈ F4 has at least one non-zero component. Then Ax = b corresponds to an

overdetermined inhomogeneous system of linear equations and will not necessarily be
solvable for all possible choices of b.
In particular, note that the bottom row A(4,·) of A corresponds to the equation
0 = b4 ,
from which Ax = b has no solution unless the fourth component of b is zero. Further-
more, the remaining rows of A correspond to the equations
x1 = b1 , x2 = b2 , and x3 = b3 .
It follows that, given any vector b ∈ Fn with fourth component zero, x = b is the only
solution to Ax = b. In other words, U = {b}.

 
1 0 1
 
0 1 1
A=
0

 0 0

0 0 0
and b ∈ F4 . This corresponds to an overdetermined inhomogeneous system of linear

equations. Note, in particular, that the bottom two rows of the matrix corresponds to
the equations 0 = b3 and 0 = b4 , from which Ax = b has no solution unless the third
and fourth component of the vector b are both zero. Furthermore, we conclude from
the remaining rows of the matrix that x3 is a free variable for this system and that
)
x1 = b1 − x3
.
x2 = b2 − x3
It follows that, given any scalar α ∈ F, every vector of the form

       
x1 b1 − α b1 −1
       
x = x2  = b2 − α = b2  + α −1 = u + αn
x3 α 0 1
is a solution to Ax = b. Recall from Example 12.3.10 that the solution space for the
associated homogeneous matrix equation Ax = 0 is

N = (x1 , x2 , x3 ) ∈ F3 | x1 = −x3 , x2 = −x3 = span ((−1, −1, 1)) .
Thus, in the language of Theorem 12.3.12, we have that u is a particular solution for
Ax = b and that (n) is a basis for N. Therefore, the solution set for Ax = b is

U = (b1 , b2 , 0) + N = (x1 , x2 , x3 ) ∈ F3 | x1 = b1 − x3 , x2 = b2 − x3 .

 
1 1 1
 
A = 0 0 0
0 0 0
and b ∈ F4 . This corresponds to a square inhomogeneous system of linear equations

with two free variables. As above, this system has no solutions unless b2 = b3 = 0,
and, given any scalars α, β ∈ F, every vector of the form
         
x1 b1 − α − β b1 −1 −1
          (1) (2)
x = x2  =  α  =  0  + α  1  + β  0  = u + αn + βn
x3 β 0 0 1
is a solution to Ax = b. Recall from Example 12.3.10, that the solution space for the
associated homogeneous matrix equation Ax = 0 is

N = (x1 , x2 , x3 ) ∈ F3 | x1 + x2 + x3 = 0 = span ((−1, 1, 0), (−1, 0, 1)) .
Thus, in the language of Theorem 12.3.12, we have that u is a particular solution for
Ax = b and that (n(1) , n(2) ) is a basis for N. Therefore, the solution set for Ax = b is

U = (b1 , 0, 0) + N = (x1 , x2 , x3 ) ∈ F3 | x1 + x2 + x3 = b1 .
12.3.4 Solving linear systems with LU-factorization

Let n ∈ Z+ be a positive integer, and suppose that A ∈ Fn×n is an upper triangular matrix
and that b ∈ Fn is a column vector. Then, in order to solve the matrix equation Ax = b,
there is no need to apply Gaussian elimination. Instead, we can exploit the triangularity of
A in order to directly obtain a solution.
Using the notation in System (12.3), note that the last equation in the linear system
corresponding to Ax = b can only involve the single unknown xn , and so we can obtain the
solution
bn
xn =
ann
as long as ann 6= 0. If ann = 0, then we must be careful to distinguish between the two cases
in which bn = 0 or bn 6= 0. Thus, for reasons that will become clear below, we assume that
the diagonal elements of A are all nonzero. Under this assumption, there is no ambiguity in
substitution the solution for xn into the penultimate (a.k.a. second-to-last) equation. Since
A is upper triangular, the penultimate equation involves only the single unknown xn−1 , and
so we obtain the solution
bn−1 − an−1,n xn
xn−1 = .
an−1,n−1
We can then similarly substitute the solutions for xn and xn−1 into the ante penultimate
(a.k.a. third-to-last) equation in order to solve for xn−2 , and so on until a complete solution
is found. In particular, P
b1 − nk=2 ank xk
x1 = .
a11
As in Example 12.3.5, we call this process back substitution. Given an arbitrary linear
system, back substitution essentially allows us to halt the Gaussian elimination procedure
and immediately obtain a solution for the system as soon as an upper triangular matrix
(possibly in REF or even RREF) has been obtained from the coefficient matrix.
A similar procedure can be applied when A is lower triangular. Again using the notation
in System (12.3), the first equation contains only x1 , and so
b1
x1 = .
a11
We are again assuming that the diagonal entries of A are all nonzero. Then, acting similarly
to back substitution, we can substitute the solution for x1 into the second equation in order
to obtain
b2 − a21 x1
x2 = .
a22
Continuing this process, we have created a forward substitution procedure. In particular,
Pn−1
bn − k=1 ank xk
xn = .
ann
More generally, suppose that A ∈ Fn×n is an arbitrary square matrix for which there
exists a lower triangular matrix L ∈ Fn×n and an upper triangular matrix U ∈ Fn×n such
that A = LU. When such matrices exist, we call A = LU an LU-factorization (a.k.a. LU-
decomposition) of A. The benefit of such a factorization is that it allows us to exploit the
triangularity of L and U when solving linear systems having coefficient matrix A.
To see this, suppose that A = LU is an LU-factorization for the matrix A ∈ Fn×n and
that b ∈ Fn is a column vector. (As above, we also assume that the none of the diagonal
entries in either L or U is zero.) Furthermore, set y = Ux, where x is the as yet unknown
solution of Ax = b. Then, by substitution, y must satisfy
Ly = b,
and so, since L is lower triangular, we can immediately solve for y via forward substitution.
In other words, we are using the associative of matrix multiplication (cf. Theorem 12.2.6) in
order to conclude that
(A)x = (LU)x = L(Ux) = L(y)
Then, once we have obtained y ∈ Fn , we can apply back substitution in order to solve for x
in the matrix equation

Ux = y.
In general, one can only obtain an LU-factorization for a matrix A ∈ Fn×n when there
exist elementary “row combination” matrices E1 , E2 , . . . , Ek ∈ Fn×n and an upper triangular
matrix U such that
Ek Ek−1 · · · E1 A = U.
There are various generalizations of LU-factorization that allow for more than just elementary
“row combinations” matrices in this product, but we do not mention them here. Instead,
we provide a detailed example that illustrates how to obtain an LU-factorization and then
how to use such a factorization in solving linear systems.
Example 12.3.15. Consider the matrix A ∈ F3×3 given by

 
2 3 4
 
A = 4 5 10 .
4 8 2
Using the techniques illustrated in Example 12.3.5, we have the following matrix product:
      
1 0 0 1 0 0 1 0 0 2 3 4 2 3 4
      
0 1 0  0 1 0 −2 1 0 4 5 10 = 0 −1 3  = U.
0 2 1 −2 0 1 0 0 1 4 8 2 0 0 −2
In particular, we have found three elementary “row combination” matrices, which, when
multiplied by A, produce an upper triangular matrix U.
Now, in order to produce a lower triangular matrix L such that A = LU, we rely on two
facts about lower triangular matrices. First of all, any lower triangular matrix with entirely
no-zero diagonal is invertible, and, second, the product of lower triangular matrices is always
lower triangular. (Cf. Theorem 12.2.7.) More specifically, we have that
   −1  −1  −1  
2 3 4 1 0 0 1 0 0 1 0 0 2 3 4
         
4 5 10 = −2 1 0  0 1 0 0 1 0 0 −1 3  .
4 8 2 0 0 1 −2 0 1 0 2 1 0 0 −2
where
 −1  −1  −1    
1 0 0 1 0 0 1 0 0 1 0 0 1 0 0 1 0 0
         
−2 1 0  0 1 0  0 1 0 = 2 1 0 0 1 0 0 1 0
0 0 1 −2 0 1 0 2 1 0 0 1 2 0 1 0 −2 1
 
1 0 0
 
= 2 1 0 .
2 −2 1
We call the resulting lower triangular matrix L and note that A = LU, as desired.
Now, define x, y, and b by
     
x1 y1 6
     
x = x2  , y = y2  , and b = 16 .
x3 y3 2
Applying forward substitution to Ly = b, we obtain the solution


y1 = b1 = 6 

y2 = b2 − 2y1 = 4 .


y3 = b3 − 2y1 + 2y2 = −2
Then, given this unique solution y to Ly = b, we can apply backward substitution to Ux = y

in order to obtain 
2x1 = y1 − 3x2 − 4x3 = 8 

−1x2 = y2 − 2x3 = 2 .


−2x3 = y3 = −2
It follows that the unique solution to Ax = b is

x1 = 4 

x2 = −2 .


x3 = 1
In summary, we have given an algorithm for solving any matrix equation Ax = b in which
A = LU, where L is lower triangular, U is upper triangular, and both L and U have nothing
but non-zero entries along their diagonals.

We note in closing that the simple procedures of back and forward substitution can also
be used for computing the inverses of lower and upper triangular matrices. E.g., the inverse
U = (uij ) of the matrix
 
2 3 4
 
U −1 = 0 −1 3 
0 0 −2
must satisfy
   
2u11 + 3u21 + 4u31 2u12 + 3u22 + 4u32 2u13 + 3u23 + 4u33 1 0 0
   
 −u21 + 3u31 −u22 + 3u32 −u23 + 3u33  = U −1 U = I3 = 0 1 0 ,
−2u31 −2u32 −2u33 0 0 1
from which we obtain the linear system


2u11 +3u21 +4u31 = 1 



2u12 +3u22 +4u32 = 0 




2u13 +3u23 +4u33 = 0 




−u21 +3u31 = 0 


−u22 +3u32 = 1


−u23 +3u33 = 0 




−2u31 = 0 




−2u32 = 0 




−2u33 = 1
in the nine variables u11 , u12 , . . . , u33 . Since this linear system has upper triangular coefficient
matrix, we can apply back substitution in order to directly solve for the entries in U.
The only condition we imposed upon our triangular matrices above was that all diagonal
entries were non-zero. It should be clear to you that this non-zero diagonal restriction is
a necessary and sufficient condition for a triangular matrix to be non-singular. Moreover,
once the inverses of both L and U in an LU-factorization have been obtained, then we can
immediately calculate the inverse for A = LU by applying Theorem 12.2.9(4):
A−1 = (LU)−1 = U −1 L−1 .

12.4 Matrices and linear maps

As discussed in Chapter 1, the machinery of Linear Algebra can be used to solve systems of
linear equations involving a finite number of unknowns. This section is devoted to illustrating
how linear maps are one of the most fundamental tools for gaining insight into the solutions
to such systems.
Moreover, by involving these special functions, we are able to work in an arbitrarily high
number of dimensions with little more work than that required for two dimensions.
12.4.1 The canonical matrix of a linear map

Let m, n ∈ Z+ be positive integers. Then, given a choice of bases for the vector spaces Fn
and Fm , there is a duality between matrices and linear maps. In other words, as discussed
in Section 6.6, every linear map in the set L(Fn , Fm ) uniquely corresponds to exactly one
m × n matrix in Fm×n . However, you should not take this to mean that matrices and linear
maps are interchangeable or indistinguishable ideas. By itself, a matrix in the set Fm×n is
nothing more than a collection of mn scalars that have been arranged in a rectangular shape.
It is only when a matrix appears as part of some larger context that the theory of linear
maps becomes applicable. In particular, one can gain insight into the solutions of matrix
equations when the coefficient matrix is viewed as the matrix associated to a linear map
under a convenient choice of bases for Fn and Fm .
Given a positive integer, k ∈ Z+ , one particularly convenient choice of basis for Fk is the
so-called standard basis (a.k.a. the canonical basis) e1 , e2 , . . . , ek , where each ei is the
k-tuple having zeros for each of its component other than in the ith position:
ei = (0, 0, . . . , 0, 1, 0, . . . , 0).
↑
i
Then, taking the vector spaces Fn and Fm under their canonical bases, we say that the
matrix A ∈ Fm×n associated to the linear map T ∈ L(Fn , Fm ) is the canonical matrix for
T . One reason for this choice of basis is that it gives us the particularly nice formula
T (x) = Ax, ∀ x ∈ Fn . (12.7)

12.4. MATRICES AND LINEAR MAPS 201
In other words, one can compute the action of the linear map upon any vector in Fn by simply
multiplying the vector by the associated canonical matrix A. There are many circumstances
in which one might wish to use non-standard bases for either Fn or Fm , but the trade-off is
that Equation (12.7) will no longer hold as stated. (To modify Equation (12.7) for use with
non-standard bases, one needs to use coordinate vectors as described in Chapter 10.)
The utility of Equation (12.7) cannot be overly emphasized. To get a sense of this,
consider once again the generic matrix equation (Equation (12.5))
Ax = b,
which involves a given matrix A = (aij ) ∈ Fm×n , a given vector b ∈ Fm , and the n-tuple
of unknowns x. To provide a solution to this equation means to provide a vector x ∈ Fn
for which the matrix product Ax is exactly the vector b. In light of Equation (12.7), the
question of whether such a vector x ∈ Fn exists is equivalent to asking whether or not the
vector b is in the range of the linear map T .
While the encoding of System (12.3) into Equation (12.5) might be considered a matter
of mere notational equivocation, the above reinterpretation of Equation (12.5) using linear
maps is a genuine change of viewpoint. Solving System (12.3) (and thus Equation (12.5))
essentially amounts to understanding how m distinct objects interact in an ambient space
having n-dimensions. (In particular, solutions to System (12.3) correspond to the points
of intersect of m hyperplanes in Fn .) On the other hand, questions about a linear map
genuinely involve understanding a single object, i.e., the linear map itself. Such a point of
view is both extremely flexible and extremely fruitful, as we illustrate in the next section.
12.4.2 Using linear maps to solve linear systems
Encoding a linear system as a matrix equation is more than just a notational trick. Perhaps
most fundamentally, the resulting linear map viewpoint can then be used to provide unpar-
alleled insight into the exact structure of solutions to the original linear system. (In general,
the more that can be said with absolute certainty when solving a problem, the better.) We
illustrate this in the following series of revisited examples.
Example 12.4.1. Consider the following inhomogeneous linear system from Example 1.2.1:
)
2x1 + x2 = 0
,
x1 − x2 = 1
where x1 and x2 are unknown real numbers. To solve this system, we can first form the
matrix A ∈ R2×2 and the column vector b ∈ R2 such that
" # " #" # " #
x1 2 1 x1 0
A = = = b,
x2 1 −1 x2 1
In other words, we have reinterpreted solving the original linear system as asking when the
column vector " #" # " #
2 1 x1 2x1 + x2
=
1 −1 x2 x1 − x2
is equal to the column vector b. Equivalently, this corresponds to asking what input vector
results in b being an element of the range of the linear map T : R2 → R2 defined by
" #! " #
x1 2x1 + x2
T = .
x2 x1 − x2
More precisely, T is the linear map having canonical matrix A.

It should be clear that b is in the range of T , since, from Example 1.2.1,
" #! " #
1/3 0
T = .
−2/3 1
In addition, note that T is a a bijective function. (This can be proven, for example, by
noting that the canonical matrix A for T is invertible.) Since T is bijective, this means that
" # " #
x1 1/3
x= =
x2 −2/3
is the only possible input vector that can result in the output vector b, and so we have
verified that x is the unique solution to the original linear system. Moreover, this technique
can be trivially generalized to any number of equations.
Example 12.4.2. Consider the matrix A and the column vectors x and b from Exam-
ple 12.3.5:      
2 5 3 x1 4
     
A = 1 2 3 , x = x2  , and b = 5 .
1 0 8 x3 9
Here, asking if the matrix equation Ax = b has a solution corresponds is equivalent to asking
if b is an element of the range of the linear map T : F3 → F3 defined by
   
x1 2x1 + 5x2 + 3x3
   
T x2  =  x1 + 2x2 + 3x3  .
x3 2x1 + 8x3
In order to answer this corresponding question regarding the range of T , we take a closer
look at the following expression obtained in Example 12.3.5:
A = E0−1 E1−1 · · · E7−1 ,
Here, we have factored A into the product of eight elementary matrices. From the linear
map point of view, this means that can apply the results of Section 6.6 in order to obtain
the factorization
T = S0 ◦ S1 ◦ · · · ◦ S7 ,
where Si is the (invertible) linear map having canonical matrix Ei−1 for i = 0, . . . , 7.
This factorization of the linear map T into a composition of invertible linear maps fur-
thermore implies that T itself is invertible. In particular, T is surjective, and so b must be an
element of the range of T . Moreover, T is also injective, and so b has exactly one pre-image.
Thus, the solution that was found for Ax = b in Example 12.3.5 is unique.
In the above examples, we used the bijectivity of a linear map in order to prove the
uniqueness of solutions to linear systems. As discussed in Section 12.3, though, many linear
systems that do not have unique solutions. Instead, there are exactly two other possibilities:
if a linear system does not have a unique solution, then it will either have no solution or
it will have infinitely many solutions. Fundamentally, this is because finding solutions to a
linear system is equivalent to describing the pre-image (a.k.a. pullback) of an element in the
codomain of a linear map.
In particular, based upon the discussion in Section 12.3.2, it should be clear that solving
a homogeneous linear system corresponds to describing the null space of some corresponding
linear map. In other words, given any matrix A ∈ Fm×n , finding the solution space N to the
matrix equation Ax = 0 (as defined in Section 12.3.2) is the same thing as finding null(T ),
where T ∈ L(Fn , Fm ) is the linear map having canonical matrix A. (Recall from Section 6.2
that null(T ) is a subspace of Fn .) Thus, the fact that every homogeneous linear system has
the trivial solution then is equivalent to the fact that the image of the zero vector under
any linear map always results in the zero vector, and determining whether or not the trivial
solution is unique can be viewed as a dimensionality question about the null space of a
corresponding linear map.
We close this section by illustrating this, along with the case for inhomogeneous systems,
in the following examples.
Example 12.4.3. The following examples use the same matrices as in Example 12.3.10.

 
1 0 0
 
0 1 0
A=
0

 0 1

0 0 0
and b ∈ F4 is a column vector. Here, asking if this matrix equation has a solution
corresponds to asking if b is an element of the range of the linear map T : F3 → F4
defined by  
  x1
x1  
  x2 
T x2  =  
x  .
 3
x3
0
From the linear map point of view, it should be extremely clear that Ax = b has
a solution if and only if the fourth component of b is zero. In particular, T is not
surjective, so Ax = b cannot have a solution for every possible choice of b.
However, it should also be clear that T is injective, from which null(T ) = {0}. Thus,
when b = 0, the homogeneous matrix equation Ax = 0 has only the trivial solution,
and so we can apply Theorem 12.3.12 in order to verify that Ax = b has a unique
solution for any b having fourth component equal to zero.

 
1 0 1
 
0 1 1
A=
0

 0 0

0 0 0
defined by  
  x1 + x3
x1  
   x2 + x3 
.
T x2  =  
 0 
x3
0
From the linear map point of view, it should be extremely clear that Ax = b has a
solution if and only if the third and fourth components of b are zero. In particular,
2 = dim(range (T )) < dim(F4 ) = 4 so that T cannot be surjective, and so Ax = b
cannot have a solution for every possible choice of b.
In addition, it should also be clear that T is not injective. E.g.,

 
  0
−1  
  0
T −1 =  
0 .
 
1
0
Thus, {0} ( null(T ), and so the homogeneous matrix equation Ax = 0 will necessarily
have infinitely many solution since dim(null(T )) > 0. Using the Dimension Formula,
dim(null(T )) = dim(F3 ) − dim(range (T )) = 3 − 2 = 1,
and so the solution space for Ax = 0 is a one-dimensional subspace of F3 . Moreover,

by applying Theorem 12.3.12, we see that that Ax = b must then also have infinitely
many solutions for any b having third and fourth components equal to zero.

 
1 1 1
 
A = 0 0 0
0 0 0
defined by    
x1 x1 + x2 + x3
   
T x2  =  0 .
x3 0
From the linear map point of view, it should be extremely clear that Ax = b has a
solution if and only if the second and third components of b are zero. In particular,
1 = dim(range (T )) < dim(F3 ) = 3 so that T cannot be surjective, and so Ax = b
cannot have a solution for every possible choice of b.
In addition, it should also be clear that T is not injective. E.g.,

   
1/2 0
   
T 1/2 = 0 .
−1 0
Thus, {0} ( null(T ), and so the homogeneous matrix equation Ax = 0 will necessarily
have infinitely many solution since dim(null(T )) > 0. Using the Dimension Formula,
dim(null(T )) = dim(F3 ) − dim(range (T )) = 3 − 1 = 2,
and so the solution space for Ax = 0 is a two-dimensional subspace of F3 . Moreover,

by applying Theorem 12.3.12, we see that that Ax = b must then also have infinitely
many solutions for any b having second and third components equal to zero.
12.5. SPECIAL OPERATIONS ON MATRICES 207
12.5 Special operations on matrices

In this section, we define three important operations on matrices called the transpose, conju-
gate transpose, and the trace. These will then be seen to interact with matrix multiplication
and invertibility in order to form special classes of matrices that are extremely important to
applications of Linear Algebra.
12.5.1 Transpose and conjugate transpose

Given positive integers m, n ∈ Z+ and any matrix A = (aij ) ∈ Fm×n , we define the trans-
pose AT = ((aT )ij ) ∈ Fn×m and the conjugate transpose A∗ = ((a∗ )ij ) ∈ Fn×m by
(aT )ij = aji and (a∗ )ij = aji ,
where aji denotes the complex conjugate of the scalar aji ∈ F. In particular, if A ∈ Rm×n ,
then note that AT = A∗ .
Example 12.5.1. With notation as in Example 12.1.3,

 
" # 1
h i 4 0  
AT = 3 −1 1 , B T = , C T =  4 ,
−1 2
2
   
1 −1 3 6 −1 4
   
DT =  5 0 2 , and E T =  1 1 1 .
2 1 4 3 2 3
One of the motivations for defining the operations of transpose and conjugate transpose
is that they interact with the usual arithmetic operations on matrices in a natural manner.
We summarize the most fundamental of these interactions in the following theorem.
Theorem 12.5.2. Given positive integers m, n ∈ Z+ and any matrices A, B ∈ Fm×n ,
1. (AT )T = A and (A∗ )∗ = A.
2. (A + B)T = AT + B T and (A + B)∗ = A∗ + B ∗ .

3. (αA)T = αAT and (αA)∗ = αA∗ , where α ∈ F is any scalar.
4. (AB)T = B T AT .
5. if m = n and A ∈ GL(n, F), then AT , A∗ ∈ GL(n, F) with respective inverses given by
(AT )−1 = (A−1 )T and (A∗ )−1 = (A−1 )∗ .
Another motivation for defining the transpose and conjugate transpose operations is that
they allow us to define several very special classes of matrices.
Definition 12.5.3. Given a positive integer n ∈ Z+ , we say that the square matrix A ∈ Fn×n
1. is symmetric if A = AT .
2. is Hermitian if A = A∗ .
3. is orthogonal if A ∈ GL(n, R) and A−1 = AT . Moreover, we define the (real)

orthogonal group to be the set O(n) = {A ∈ GL(n, R) | A−1 = AT }.
4. is unitary if A ∈ GL(n, C) and A−1 = A∗ . Moreover, we define the (complex)

unitary group to be the set U(n) = {A ∈ GL(n, C) | A−1 = A∗ }.
A lot can be said about these classes of matrices. Both O(n) and U(n), for example, form
a group under matrix multiplication. Additionally, real symmetric and complex Hermitian
matrices always have real eigenvalues. Moreover, given any matrix A ∈ Rm×n , AAT is a
symmetric matrix with real, non-negative eigenvalues. Similarly, for A ∈ Cm×n , AA∗ is
Hermitian with real, non-negative eigenvalues.
12.5.2 The trace of a square matrix

Given a positive integer n ∈ Z+ and any square matrix A = (aij ) ∈ Fn×n , we define the
trace of A to be the scalar n
X
trace(A) = akk ∈ F.
k=1
Example 12.5.4. With notation as in Example 12.1.3 above,
trace(B) = 4 + 2 = 6, trace(D) = 1 + 0 + 4 = 5, and trace(E) = 6 + 1 + 3 = 10.
Note, in particular, that the traces of A and C are not defined since these are not square
matrices.
We summarize some of the most basic properties of the trace operation in the following
theorem, including its connection to the transpose operations defined in the previous section.
Theorem 12.5.5. Given a positive integer m, n ∈ Z+ and square matrices A, B ∈ Fn×n ,
1. trace(αA) = α trace(A), for any scalar α ∈ F.
2. trace(A + B) = trace(A) + trace(B).
3. trace(AT ) = trace(A) and trace(A∗ ) = trace(A).

n X
X n
∗
4. trace(AA ) = |akℓ |2 . In particular, trace(AA∗ ) = 0 if and only if A = 0n×n .
k=1 ℓ=1
5. trace(AB) = trace(BA). More generally, given matrices A1 , . . . , Am ∈ Fn×n , the trace

operation has the so-called cyclic property, meaning that
trace(A1 · · · Am ) = trace(A2 · · · Am A1 ) = · · · = trace(Am A1 · · · Am−1 ).
Moreover, if we define a linear map T : Fn → Fn by setting T (v) = Av for each v ∈ Fn and

n
X
if T has distinct eigenvalues λ1 , . . . , λn , then trace(A) = λk .
k=1

1. In each of the following, find matrices A, x, and b such that the given system of linear
equations can be expressed as the single matrix equation Ax = b.

 4x1 − 3x3 + x4 = 1 

2x1 − 3x2 + 5x3 = 7  

 5x1 + x2 − 8x4 = 3 
(a) 9x1 − x2 + x3 = −1 (b)

 2x1 − 5x2 + 9x3 − x4 = 0 
x1 + 5x2 + 4x3 = 0 

3x2 − x3 + 7x4 = 2 
2. In each of the following, express the matrix equation as a system of linear equations.
    
     3 −2 0 1 w 0
3 −1 2 x1 2     
 5 0 2 −2     
(a)
 4
   
3 7   x2  =  −1  (b)   x  =  0 
 3 1 4 7     
  y   0 
−2 1 5 x3 4
−2 5 1 6 z 0
3. Suppose that A, B, C, D, and E are matrices over F having the following sizes:
A is 4 × 5, B is 4 × 5, C is 5 × 2, D is 4 × 2, E is 5 × 4.
Determine whether the following matrix expressions are defined, and, for those that
are defined, determine the size of the resulting matrix.
(a) BA (b) AC + D (c) AE + B (d) AB + B (e) E(A + B) (f) E(AC)
4. Suppose that A, B, C, D, and E are the following matrices:

 
3 0 " # " #
  4 −1 1 4 2
A =  −1 2 , B = , C= ,
0 2 3 1 5
1 1
   
1 5 2 6 1 3
   
D =  −1 0 1 , and E =  −1 1 2 .
3 2 4 4 1 3
are defined, compute the resulting matrix.
(a) D + E (b) D − E (c) 5A (d) −7C (e) 2B − C

(f) 2E − 2D (g) −3(D + 2E) (h) A − A (i) AB (j) BA
(k) (3E)D (l) (AB)C (m) A(BC) (n) (4B)C + 2B (o) D − 3E
(p) CA + 2E (q) 4E − D (r) DD
5. Suppose that A, B, and C are the following matrices and that a = 4 and b = −7.
     
1 5 2 6 1 3 1 5 2
     
A =  −1 0 1 , B =  −1 1 2 , and C =  −1 0 1  .
3 2 4 4 1 3 3 2 4
Verify computationally that
(a) A + (B + C) = (A + B) + C (b) (AB)C = A(BC)

(c) (a + b)C = aC + bC (d) a(B − C) = aB − aC
(e) a(BC) = (aB)C = B(aC) (f) A(B − C) = AB − AC
(g) (B + C)A = BA + CA (h) a(bC) = (ab)C
(i) B − C = −C + B
6. Suppose that A is the matrix " #

3 1
A= .
2 1
Compute p(A), where p(z) is given by
(a) p(z) = z − 2 (b) p(z) = 2z 2 − z + 1

(c) p(z) = z 3 − 2z + 4 (d) p(z) = z 2 − 4z + 1
7. Define matrices A, B, C, D, and E by

 
" # " # 2 −3 5
3 1 4 −1  
A= , B= , C =  9 −1 1 ,
2 1 0 2
1 5 4
   
1 5 2 6 1 3
   
D =  −1 0 1 , and E =  −1 1 2 .
3 2 4 4 1 3
(a) Factor each matrix into a product of elementary matrices and an RREF matrix.
(b) Find, if possible, the LU-factorization of each matrix.
(c) Determine whether or not each of these matrices is invertible, and, if possible,
compute the inverse.
8. Suppose that A, B, C, D, and E are the following matrices:

 
3 0 " # " #
  4 −1 1 4 2
A =  −1 2 , B = , C= ,
0 2 3 1 5
1 1
   
1 5 2 6 1 3
   
D =  −1 0 1 , and E =  −1 1 2 .
3 2 4 4 1 3
are defined, compute the resulting matrix.
(a) 2AT + C (b) D T − E T (c) (D − E)T
(d) B T + 5C T (e) 12 C T − 41 A (f) B − B T
(g) 3E T − 3D T (h) (2E T − 3D T )T (i) CC T
(j) (DA)T (k) (C T B)AT (l) (2D T − E)A
(m) (BAT − 2C)T (n) B T (CC T − AT A) (o) D T E T − (ED)T
(p) trace(DD T ) (q) trace(4E T − D) (r) trace(C T AT + 2E T )
1. Let n ∈ Z+ be a positive integer and ai,j ∈ F be scalars for i, j = 1, . . . , n. Prove that
the following two statements are equivalent:
(a) The trivial solution x1 = · · · = xn = 0 is the only solution to the homogeneous

system of equations 
n
X 
a1,k xk =0 




k=1 
..
. .
n


X 

an,k xk =0 


k=1
(b) For every choice of scalars c1 , . . . , cn ∈ F, there is a solution to the system of

equations 
X n

a1,k xk = c1 




k=1 
..
. .
n


X 

an,k xk = cn 


k=1
2. Let A and B be any matrices.
(a) Prove that if both AB and BA are defined, then AB and BA are both square
matrices.
(b) Prove that if A has size m × n and ABA is defined, then B has size n × m.
3. Suppose that A is a matrix satisfying AT A = A. Prove that A is then a symmetric

matrix and that A = A2 .
4. Suppose A is an upper triangular matrix and that p(z) is any polynomial. Prove or
give a counterexample: p(A) is a upper triangular matrix.
Appendix A
The language of Sets and Functions
The power of mathematics is to a large extent the power of abstraction. Abstraction, in

short, consists in replacing concrete objects with a limited list of their relevant properties.
In mathematics, all reasoning is then based solely on those properties. As a consequence,
the arguments and results one thus obtains apply to all situations where the properties one
has assumed are present.
The french mathematician Henri Poincaré expressed this succinctly by saying that math-
ematics is the art of giving the same name to different things.
For this abstraction to yield reliable results, our reasoning must be reliable. We have
to apply logical thinking, and we must avoid using hidden assumptions in addition to the
properties we have explicitly assumed. That is, we have to adhere to rigorous arguments.
Mathematics is powerful also because its practitioners strive to communicate their results
and arguments in clear language so that others can independently verify that they are free
of logical flaws and use the results in their own work of creating and applying mathematics.
Abstract mathematics starts then with describing the list of properties that will define a
object in a definition. Short definitions may be given in ordinary sentences and paragraphs.
More complex ones, or especially Important ones or often displayed in a formal statement
called Definition. The generic name for a result is theorem and the rigorous arguments
that lead to the result is called a proof of the theorem. When one defines a new notion
or object, its definition typically involves other notions that should have been previously
defined. When one thinks this through, one soon realizes that the demands of mathematics—
defining everything carefully and adhere to logical arguments of the highest standard—, is
214
A.1. THE LANGUAGE OF SETS 215
not a trivial task. In particular, one has to start somewhere. In everyday mathematics that
starting point is the notion set and a small number of elementary operations with sets that
allow one to construct new sets from old ones. This is why we start this appendix with a
brief discussion of the language of sets. A particular construction with sets then leads to the
notion of function. Sets and functions form the basis of all modern mathematics.
A.1 The language of sets

A.1.1 Definition, notation, and examples
We begin by introducing the concept of a set. Intuitively, a set is any finite or infinite
collection of objects for which the object are well-defined and distinct. This means two
things: there is no ambiguity in deciding whether or not a given object belongs to a set, and
the objects in a set must be distinguishable from each other. More concretely, we define the
concept of “set” as follows:
Definition A.1.1. A set S is any (unordered) collection of (distinct) objects whose mem-
bership in S is well-defined. Given an object s, we say that s is an element of S, denoted
s ∈ S, if s is a member of S. Otherwise, we write s ∈
/ S.
Example A.1.2. Some examples of sets include
1. the empty set (a.k.a. the null set), which is denoted by either {} or ∅. This is the
set with no objects inside of it, which is certainly valid under the definition of set. In
particular, given any object s, s ∈/ ∅.
2. so-called singleton sets. These are sets that contain only a single element. E.g., the
set {37} containing the number 37 is a singleton set, while the set {3, 7} containing
both 3 and 7 is not.
3. the set {α, β, γ} containing the first three lower case Greek letters. Since the elements
in a set are unordered, we could also write this set as {α, γ, β} or {γ, β, α}, etc.
Since the elements in a set are also required to be distinct, note that something like
{α, β, α, γ} would not be considered a set. However, it is often convenient to just agree
that {α, β, α, γ} = {α, β, γ}, unless the context dictates otherwise.
216 APPENDIX A. THE LANGUAGE OF SETS AND FUNCTIONS
4. the sets Z, R, and C, of integers, real numbers, and complex numbers, respectively.
5. the single set {Z, R, C} that contains as object the sets of numbers from the previous
example. There is nothing in the definition that restricts the objects in a set from
themselves also being sets.
6. the singleton set {{}} = {∅} containing the empty set ∅. Here, the set {∅} contains
the empty set ∅ ∈ {∅} as an element, and so {∅} is itself not empty.
The notation may look daunting, but think of {{}} as being like an empty grocery
bag nested inside of another grocery bag. Since the “outer” grocery bag contains the
“inner” empty grocery bag, the “outer” bag is itself not considered empty.
Similarly, {{{}}}, {{{{}}}}, etc., are all perfectly valid singleton sets, whereas some-
thing like { {{}} , {{{}}} } is a set with the two elements {∅} and {{∅}}.
7. the set B of all Davis bookstores holding a book sale at a fixed moment in time.
Even though it would potentially take a great deal of effort to explicitly list its elements,
B nonetheless qualifies as a set. E.g., one could determine whether or not a particular
bookstore b is an element of B by telephoning them to ask about book sales.
At the same time, though, note that the collection B′ of all interesting Davis bookstores
holding sales during a fixed moment in time would not be a set. The problem with B′
is that there is no well-defined membership rule unless we can first rigorously define
what it means for a bookstore to be “interesting”.
Note that the bookstore example only make sense if we first define the set of all Davis
bookstores U. Each bookstore b ∈ U is then well-defined as an object before being tested
for membership in a set like B. More generally, there is an all-encompassing universal set,
often dictated implicitly by the context, that contains every object s that we might wish to
test for membership in a given set S. A common practice is then to specify S by giving some
type of pattern or constructive algorithm that distinguishing objects within the universal
set. We illustrate the three most common such methods in the following example.
Example A.1.3.
1. The simplest form of pattern uses list notation (a.k.a. roster notation) in order to
either explicitly or implicitly specify each individual element in a set. The set {α, β, γ}
from Example A.1.2(3) above is an example of the former, while the set of positive
integers Z+ = {1, 2, 3, . . .} is an example of the latter. Note that a pattern is also
allowed to be bi-directional, as in the set of integers Z = {. . . , −2, −1, 0, 1, 2, . . .}.
2. Patterns can also be given using so-called set builder notation. In this notation,
generic objects within some universal set are specified by giving one or more conditions
that must satisfied. An example is the set of rational numbers,
na o
Q= | a, b ∈ Z, b 6= 0 ,
b
in which one starts with all (admittedly ill-defined) fractions a/b and then restricts to
the case of fractions formed from integers without zero division. This notation is read
a
“Q is the set of all fractions such that a and b are integers and b 6= 0.”
b
3. When dealing with numbers, a common variant of set builder notation is so-called
interval notation. In this notation, elements are selected from some universal set
based upon their relative size or rank. For example,
(1, 3] = {x ∈ R | 1 < x ≤ 3}.
A.1.2 Operations and relations on sets

Given how fundamental sets are to the mathematical thought process, you should not be
surprised to know that there is a rich vocabulary for describing both how sets can be related
to each other and how sets can be operated upon by other sets. In this section, we briefly
look at the most basic of such concepts. We then discuss the most fundamental method for
adjoining sets, called the Cartesian product, in the next section.
You are highly encouraged to provide careful proofs for each of the theorems in this
section. Doing so will help you in better understanding the concepts involved.
Definition A.1.4. Given two sets S and T , we say that S is a subset of T (denoted S ⊂ T )
if, given any element s ∈ S, s is also an element of T . Moreover, if S ⊂ T but S 6= T , then
we call S a proper subset of T , or that T contains S, and denote this by S ( T .
Equivalently, one can also say that S is contained in T , which is denoted T ⊃ S.
Example A.1.5. It is easy to see that Z ⊂ R ⊂ C, while Z 6⊃ (1, 3] ⊂ R. What might

not be so easy to see, though, is that ∅ ⊂ Z. (In fact, as we will see below, ∅ is actually a
subset of any set. So, e.g., we could have also taken ∅ ⊂ R or ∅ ⊂ C.) This is because every
element in ∅ (of which there are none) is an element in Z. In other words, it is trivially true
that every element of ∅ is contained in Z. Similarly, you should make sure understand why
the following statements hold:
{} ∈ {{}}, and {} ⊂ {{}}, but {{}} ∈ {{{}}}, and {{}} 6⊂ {{{}}}.
and also
{} ∈ {{}} ∈ {{{}}} ∈ {{{}}} ∈ {{{{}}}} ∈ {{{{{}}}}} ∈ · · · ,
which, to extend the analogy of Example A.1.2(6), can be thought of as successively nesting
more and more grocery bags within one other.
By the same reasoning, the empty set is also a subset of itself. In other words, ∅ ⊂ ∅.
This is because, once again, every element of ∅ (of which there are none) is contained in ∅.
Since ∅ has no elements, we have in particular that ∅ 6∈ ∅.
Set containment has some elementary properties we can summarize in the following the-
orem. Note that, unless otherwise specified, an arbitrary set is allowed to be empty.
Theorem A.1.6. Let R, S, and T be any sets.
1. (the empty set is a subset of every set) Then ∅ ⊂ R.
2. (set containment is reflexive) Then R ⊂ R.
3. (set containment is antisymmetric) If R ⊂ S and S ⊂ R, then R = S.
4. (set containment is transitive) If R ⊂ S and S ⊂ T , then R ⊂ T .
The following operations allow us to combine sets in various natural ways.
Definition A.1.7. Let S and T be sets with respect to universal set U. Then we define
1. the union of S and T to be the set S ∪ T , where
S ∪ T = {x ∈ U | x ∈ S or x ∈ T }.
2. the intersection of S and T to be the set S ∩ T , where
S ∩ T = {x ∈ U | x ∈ S and x ∈ T }.
3. the complement of S to be the set Se (sometimes also denoted as S or S C ), where
Se = {x ∈ U | x ∈
/ S}.
4. the set difference of S from T to be the set T \ S = T ∩ Se (sometimes also denoted

as T − S), where
T \ S = {x ∈ U | x ∈ T but x ∈
/ S}.
These operations often behave in a fairly intuitive manner. E.g., if S ⊂ T , then you should
be able to prove that S ∪ T = T and S ∩ T = S. Similarly, if we also have a subset R ⊂ T
for which R ∩ S = ∅, then S \ R = S and R \ S = R.
Even more important are the ways in which these operations interact with each other.
We summarize the most essential interactions in the following theorem.
Theorem A.1.8. Let R, S, and T be any sets. Then
1. (distributivity) R ∩ (S ∪ T ) = (R ∩ S) ∪ (R ∩ T ) and R ∪ (S ∩ T ) = (R ∪ S) ∩ (R ∪ T ).
^
2. (De Morgan’s Laws) A e∩B
∪B = A ^
e and A e ∪ B.
∩B =A e
3. (relative complements) A\(B∪C) = (A\B)∩(A\C) and A\(B∩C) = (A\B)∪(A\C).
A.1.3 Cartesian products and (ordered) lists

In Section A.1.2 above, several methods were given for two combining sets in order to form a
new set. In this section, we introduce yet another method. Unlike the operations described in
Section A.1.2, though, the so-called Cartesian product of sets builds a new set of entirely
distinct objects and is not based upon comparison of the objects involved.
Definition A.1.9. Let S and T be sets. Then the Cartesian product of S and T is
defined to be the set S × T of all ordered pairs that can be formed with elements from S in
the first position and elements from T in the second position. In other words,
S × T = {(s, t) | s ∈ S and t ∈ T } .
Given elements s ∈ S and t ∈ T , we call the ordered pair (s, t) a two-tuple.
As with many of the operations performed on sets (including union and intersection), the
formation of “tuples” can easily be extended to make sense for any number of sets.
Definition A.1.10. Let n ∈ Z+ be a positive integer, and let S1 , S2 , . . . , Sn be any sets.

Then the Cartesian product of S1 , S2 , . . . , Sn is defined to be the set S1 × S2 × · · · × Sn of
all ordered arrangements of elements that can be formed with elements from S1 in the first
position, elements from S2 in the second position, and so on. In other words,
S1 × S2 × · · · × Sn = {(s1 , s2 , . . . , sn ) | s1 ∈ S1 , s2 ∈ S2 , . . . , and sn ∈ Sn } .
Given elements s1 ∈ S1 , s2 ∈ S2 , . . . , sn ∈ Sn , we call the ordered arrangement (s1 , s2 , . . . , sn )

an n-tuple.
Definition A.1.9 even makes sense when “two” is replaced by “infinity”. (In this context,
you should take “infinity” to mean something like, “an object that is larger than any positive
integer”. In a more general context, it would be important to carefully define the concept of
“cardinality” so that “countable infinity” could be distinguished from “uncountable infinity”.
However, such a discussion is beyond the scope of this Appendix.)
Definition A.1.11. Let S1 , S2 , . . . , Sn , . . . be an infinite number of sets. Then the Carte-

sian product of S1 , S2 , . . . is defined to be the set S1 × S2 × · · · × Sn × · · · of all infinite
ordered arrangements of elements that can be formed such that, for each positive integer
i ∈ Z+ , elements from Si appear in the ith position. In other words,
S1 × S2 × · · · × Sn × · · · = {(s1 , s2 , . . . , sn , . . .) | s1 ∈ S1 , s2 ∈ S2 , . . .} .
Given elements s1 ∈ S1 , s2 ∈ S2 , . . . , sn ∈ Sn , . . ., we call (s1 , s2 , . . . , sn , . . .) a sequence.
You have undoubtedly encountered Cartesian products and n-tuples before. E.g., there
is the so-called Euclidean line R1 , which consists of all “1-tuples” of real numbers. (You
A.2. THE LANGUAGE OF FUNCTIONS 221
can think of the “1-tuples” in R1 as being distinct from the real numbers in R. More often
than not, though, there is no confusion in thinking of R1 and R as the same set.) You have
probably also spent a considerable amount of time dealing with objects in the Euclidean
plane R2 = R × R, which consists of all two-tuples of real numbers. There is even a good
chance that you have encountered Euclidean space R3 = R × R × R, which consists of all
three-tuples of real numbers. One can similarly define so-called Euclidean n-space Rn by
taking the Cartesian product of the set R with itself n times.
Note that Rn is formed by taking each of the sets in Definition A.1.10 to be the same. This
is a very special case that you will encounter frequently. Of particular note, the “n-tuples”
in this construction allow us to form finite, ordered subsets with repeated values.
Definition A.1.12. Let n ∈ Z+ be a positive integer and S be a set. Then the n-fold
Cartesian product of S is defined to be the set S n = S
| ×S ×
{z· · · × S} .
n times
Given elements s1 , s2 , . . . , sn ∈ S, we call the n-tuple (s1 , s2 , . . . , sn ) a list. Moreover, given

any i ∈ {1, 2, . . . , n}, we denote the list formed by removing si from (s1 , s2 , . . . , sn ) by
(s1 , s2 , . . . , sbi , . . . , sn ) = (s1 , s2 , . . . , si−1 , si+1 , . . . , sn ) ∈ S n−1 .
A.2 The language of functions
A.2.1 Definition, notation, and examples

While sets are already a useful concept, their utter ubiquity and indispensability in abstract
mathematics largely stems from the definition of a function. In particular, a function
(a.k.a. map) is a special type of relationship between two sets that we define as follows.
Definition A.2.1. Let X and Y be sets. Then a function f : X → Y from X into Y

is any rule that assigns to each element x ∈ X exactly one element y ∈ Y . We call X the
domain of the function f and Y the codomain (a.k.a. target space) of the function f .
If x ∈ X is assigned the value y ∈ Y , then we call y the image (a.k.a. output value) of
f
the input value x under f and denote this relationship either by f (x) = y or by x 7→ y.
Moreover, given two functions f : X → Y and g : X → Y , we say that f and g are equal
and write f = g if, for all x ∈ X, f (x) = g(x).
Given a function f : X → Y , note that it is common to simply refer to it by its “name” f ,

and especially if the domain and codomain are understood from the context. (Alternatively,
if the “name” is unimportant, then it is common to refer to the function simply as X → Y .)
Sometimes, however, it is also convenient to use the so-called dot notation “f (·)”. In this
case, a raised dot “·” is used as a placeholder for “input” to the function. Dot notation is
especially convenient when a function’s name is more easily given as some type of formula.
We illustrate both Definition A.2.1 and dot notation in the following examples.
Example A.2.2.
1. Consider the function f : R → R defined by f (t) = et for each t ∈ R. Since et

only represents the value of this function for a particular input value t ∈ R, it is an
(incredibly common) abuse of notation to simply refer to f as “et ”. More properly,
one could use dot notation to denote f either as “e· ” or as “exp(·)”.
Consider also the function g : R → R defined by g(t) = (et + e−t )/2 for each t ∈ R. As
with the previous example, (et + e−t )/2 only represents the value of g under a specific
input value t ∈ R. In dot notation, g is rendered either as “(e· +e−· )/2” or as “cosh(·)”.
As explained at the end of Section A.2.3, one can also relate the functions f and g by
writing
g(·) = (f (·) + f (−·))/2.
2. Consider the function f : R → R defined by f (t) = |t| for each t ∈ R. In other words,
f (t) is the absolute value of t ∈ R. To render f in dot notation, one writes “| · |”.
Similar considerations lead to the use of “k · k” and “h·, ·i” when denoting an otherwise
unspecified norm and inner product, respectively, on an inner product space V . Note,
in particular, that the domain of h·, ·i is the Cartesian product V × V of V with itself.
In other words, the two dots in “h·, ·i” are understood to be independent of each other
and do not mean that the same input value is necessarily being repeated. Instead, the
input to h·, ·i can be any ordered pair (u, v) ∈ V × V , with “hu, vi” used to denote the
image of (u, v) under h·, ·i. Moreover, note that hu, vi ∈ C is a single complex number.
3. Finally, with notation as in the previous example, one can also define the functions
hx, ·i and h·, yi. Here, x, y ∈ V are any two fixed elements in V , and both functions
are understood to have domain V .
E.g., given any vector u ∈ V , we use hu, yi ∈ C to denote the image of u under h·, yi.
Every function f : X → Y has several special sets associated to it. E.g., by the graph
of f we mean the set
Γ(f ) = {(x, f (x)) | x ∈ X},
which is a subset of the Cartesian product X × Y . (In fact, you have probably spent a good
deal of time sketching the graphs Γ(f ) ⊂ R2 that are associated to functions f having both
domain and codomain R.) One can, in fact, define the function by specifying its graph as a
subset of X × Y , and the abstract notion of function can be defined as a subset of X × Y
with suitable properties (which properties?). The range of a function f is defined to be the
set
range (f ) = {y ∈ Y | y = f (x) for some x ∈ X}.
Given any y ∈ Y , we also define the pre-image (a.k.a. pullback) of y to be the set
f −1 (y) = {x ∈ X | f (x) = y}.
You should be able to prove that f −1 (y) = ∅ for every y ∈ Y \ range (f ).
A.2.2 Injectivity, surjectivity, and bijectivity

There are three basic adjectives that are applied to functions, each of which distinctly places
extra conditions on how elements in the domain are assigned values in the codomain. The
resulting fundamental classes of functions are summarized in the following definition.
Definition A.2.3. Let f : X → Y be a function. Then we call f
1. injective (a.k.a. one-to-one) if f assigns at most one element from the domain to
each element in the codomain. In other words, f is an injection if, for each y ∈ Y ,
the cardinality |f −1 (y)| of the pullback of y is at most one. Another way of expressing
this is the implication f (x) = f (y) =⇒ x = y.
2. a surjective (a.k.a. onto) if f assigns at least one element from the domain to each
element in the codomain. In other words, f is a surjection if, for each y ∈ Y , the
cardinality |f −1(y)| of the pullback of y is at least one.
3. bijective (a.k.a. invertible) if f is both an injective function and a surjective function.

In other words, f is a bijection if, for each y ∈ Y , the cardinality |f −1(y)| of the
pullback of y is exactly one. In other words: for every x ∈ X there is exactly one
y ∈ Y , and for every y ∈ Y there is exactly one x ∈ X, such that f (x) = y.
In particular, if f is a bijection, then the assignment of values from the domain to the
codomain is called a one-to-one correspondence since exactly one element in the domain
corresponds to exactly one element in the codomain. Consequently, it is then possible to
literally “undo” the assignment of values under f . This yields the inverse of the function f ,
which we denote by f −1 . (Since each pullback of a bijection is a singleton set, this is only a
minor abuse of notation.) In other words, given a bijection f : X → Y , the inverse function
f −1 : Y → X is defined by the rule f −1 (y) = x if and only if f (x) = y, for all y ∈ Y .
We conclude this section with some distinguished examples of functions that arise fre-
quently in abstract mathematics.
Example A.2.4. Let X and Y be sets. Then,
1. if X = Y , we can define the identity map idX : X → Y by setting idX (x) = x for
each input value x ∈ X. Clearly, idX is invertible regardless of the structure on X,
and, moreover, idX is equal to its own inverse function id−1
X : Y → X.
2. if X ⊂ Y , we can define the inclusion map inclX : X → Y by setting inclX (x) = x

for each input value x ∈ X. Clearly, inclX is a one-to-one function, but it cannot be
onto unless X = Y .
3. for each y ∈ Y , we can define the constant map consty : X → Y by setting

consty (x) = y for each input value x ∈ X. Note, in particular, that consty is one-to-one
if and only if Y = {y} is a singleton set, and consty is onto if and only if X = {x} is a
singleton set. It follows that consty is invertible if and only if X = Y = {y}.
4. if Y = {0, 1}, we can define the Kronecker delta function (a.k.a. Kronecker delta
symbol) δ : X × X → Y by setting, for each pair of input values x1 , x2 ∈ X,

1 if x = x ,
1 2
δ(x1 , x2 ) =
0 otherwise.
As you will see, the Kronecker delta function is a often a very convenient notational
tool.
Quite often, the Kronecker delta function is taken with X = Z+ or X = Z+ ∪ {0}, in
which case δ(i, j) is denotes as δi,j for each i, j ∈ X.
In this last example, note that the image of each ordered pair (x1 , x2 ) ∈ X × X under δ
is denoted as δ(x1 , x2 ) rather than as δ ((x1 , x2 )). The standard convention is to “drop” (i.e.,
not to write) the extra parentheses when the domain of a function is a Cartesian product.
A.2.3 Operations on functions

We conclude these notes by mentioning several important operations that are often applied
to functions. These operations will see many applications throughout your study of abstract
mathematics.
Definition A.2.5. Let f : X → Y and g : Y → Z be functions. Then
1. we can define the composition of f with g to be the function g ◦ f : X → Z, whose

values are given by (g ◦ f )(x) = g(f (x)) for each x ∈ X.
2. we can define the restriction of f to a subset W ⊂ X to be the function f |W : X → Y ,

whose values are given by f |W (w) = f (w) for each w ∈ W .
If one thinks of a function as “acting” on its argument, and of the inverse function as
undoing this action, then the following property of the inverse of the composition of two
functions is intuitively clear.
Theorem A.2.6. Let f : X → Y and g : Y → Z be bijections. Then
(g ◦ f )−1 = f −1 ◦ g −1 .
To remember this result, you can also think of inverting a function as reversing the arrow in
f : X → Y . Then, g ◦ f : X → Y → Z and (g ◦ f )−1 : Z → Y → X.
Note that composition and restriction are operations that can be applied to functions with
arbitrary domain and codomain. If the codomain is equipped with an algebraic operation, it
is possible to naturally extend this operation to functions. We illustrate this in the following
example.
Example A.2.7. Consider the functions f : R → R and g : R → R defined by f (t) = sin(t)

and g(t) = cos(t) for each t ∈ R. Then we can define a new function f + g : R → R. This
new function has the name “f + g” — or, in dot notation, one of (f + g)(·), f (·) + g(·), or
sin(·) + cos(·) — and, for each input t ∈ R, the output is defined to be
(f + g)(t) = f (t) + g(t) = sin(t) + cos(t).
One can similarly define functions such as

p
22g(·)/7 = (22/7) cos(·), f 1/2 (·) = sin(·), and (f 2 + g 2 )(·) = const1 .
Note, though, that care must be taken with negative exponents. The function f −1 , in this
example, is the so-called “arc sine” function sin−1 (·) = Arcsin(·) and not the function
1
(f (·))−1 = .
sin(·)
This inconsistency in notation is deeply rooted in tradition and rarely leads to confusion.
Appendix B
Summary of Algebraic Structures

Encountered
Loosely speaking, an algebraic structure is any set upon which “arithmetic-like” opera-
tions have been defined. The importance of such structures in abstract mathematics cannot
be overstated. By recognized a given set S as an instance of a well-known algebraic structure,
every result that is known about that abstract algebraic structure is then automatically also
known to hold for S. This utility is, in large part, the main motivation behind abstraction.
Before reviewing the algebraic structures that are most important to the study of Linear
Algebra, we first carefully define what it means for an operation to be “arithmetic-like”.
B.1 Binary operations and scaling operations

When discussing an arbitrary nonempty set S, you should never assume that S has any type
of “structure” (algebraic or otherwise) unless the context suggests differently. Put another
way, the elements in S can only every really be related to each other in a subjective manner.
E.g., if we take S = {Alice, Bob, Carol}, then there is nothing intrinsic in the definition of
S that suggests how these names should objectively be related to one another.
If, on the other hand, we take S = R, then you have no doubt been conditioned to expect
that a great deal of “structure” already exists within S. E.g., given any two real numbers
r1 , r2 ∈ R, one can form the sum r1 + r2 , the difference r1 − r2 , the product r1 r2 , the quotient
r1 /r2 (assuming r2 6= 0), the maximum max{r1 , r2 }, the minimum min{r1 , r2 }, the average
227
228 APPENDIX B. SUMMARY OF ALGEBRAIC STRUCTURES ENCOUNTERED
(r1 + r2 )/2, and so on. Each of these operations follows the same pattern: take two real
numbers and “combine” (or “compare”) them in order to form a new real number.
Moreover, each of these operations imposes a sense of “structure” within R by relating
real numbers to each other. We can abstract this to an arbitrary nonempty set as follows:
Definition B.1.1. A binary operation on a nonempty set S is any function that has as
its domain S × S and as its codomain S.
In other words, a binary operation on S is any rule f : S × S → S that assigns exactly one
element f (s1 , s2 ) ∈ S to each pair of elements s1 , s2 ∈ S. We illustrate this definition in the
following examples.
Example B.1.2.
1. Addition, subtraction, and multiplication are all examples of familiar binary operations
on R. Formally, one would denote these by something like
+ : R × R → R, − : R × R → R, and ∗ : R × R → R, respectively.
Then, given two real numbers r1 , r2 ∈ R, we would denote their sum by +(r1 , r2 ),
their difference by −(r1 , r2 ), and their product by ∗(r1 , r2 ). (E.g., +(17, 32) = 49,
−(17, 32) = −15, and ∗(17, 32) = 544.) However, this level of notational formality can
be rather inconvenient, and so we often resort to writing +(r1 , r2 ) as the more familiar
expression r1 + r2 , −(r1 , r2 ) as r1 − r2 , and ∗(r1 , r2 ) as either r1 ∗ r2 or r1 r2 .
2. The division function ÷ : R × (R \ {0}) → R is not a binary operation on R since it

does not have the proper domain. However, division is a binary operation on R \ {0}.
3. Other binary operations on R include the maximum function max : R × R → R, the

minimum function min : R × R → R, and the average function (· + ·)/2 : R × R → R.
4. An example of a binary operation f on the set S = {Alice, Bob, Carol} is given by


s if s1 alphabetically precedes s2 ,
1
f (s1 , s2 ) =
Bob otherwise.
This is because the only requirement for a binary operation is that exactly one element
of S is assigned to every ordered pair of elements (s1 , s2 ) ∈ S × S.
B.1. BINARY OPERATIONS AND SCALING OPERATIONS 229
Even though one could define any number of binary operations upon a given nonempty
set, we are generally only interested in operations that satisfy additional “arithmetic-like”
conditions. In other words, the most interesting binary operations are those that, in some
sense, abstract the salient properties of common binary operations like addition and multipli-
cation on R. We make this precise with the definition of a so-called “group” in Section B.2.
At the same time, though, binary operations can only be used to impose “structure”
within a set. In many settings, it is equally useful to additional impose “structure” upon a
set. Specifically, one can define relationships between elements in an arbitrary set as follows:
Definition B.1.3. A scaling operation (a.k.a. external binary operation) on a non-
empty set S is any function that has as its domain F × S and as its codomain S, where F
denotes an arbitrary field. (As usual, you should just think of F as being either R or C).
In other words, a scaling operation on S is any rule f : F × S → S that assigns exactly one
element f (α, s) ∈ S to each pair of elements α ∈ F and s ∈ S. This abstracts the concept
of “scaling” an object in S without changing what “type” of object it already is. As such,
f (α, s) is often written simply as αs. We illustrate this definition in the following examples.
Example B.1.4.
1. Scalar multiplication of n-tuples in Rn is probably the most familiar scaling operation

to you. Formally, scalar multiplication on Rn is defined as the following function:
(α, (x1 , . . . , xn )) 7−→ α(x1 , . . . , xn ) = (αx1 , . . . , αxn ), ∀ α ∈ R, ∀ (x1 , . . . , xn ) ∈ R.
In other words, given any α ∈ R and any n-tuple (x1 , . . . , xn ) ∈ R, their scalar multipli-
cation results in a new n-tuple denoted by α(x1 , . . . , xn ). This new n-tuple is virtually
identical to the original, each component having just been “rescaled” by α.
2. Scalar multiplication of continuous functions is another familiar scaling operation.

Given any real number α ∈ R and any function f ∈ C(R), their scalar multiplica-
tion results in a new function that is denoted by αf , where αf is defined by the rule
(αf )(r) = α(f (r)), ∀ r ∈ R.
In other words, this new continuous function αf ∈ C(R) is virtually identical to the
original function f ; it just “rescales” the image of each r ∈ R under f by α.
3. The division function ÷ : R × (R \ {0}) → R is a scaling operation on R \ {0}. In

particular, given two real number r1 , r2 ∈ R and any non-zero real number s ∈ R \ {0},
we have that ÷(r1 , s) = r1 (1/s) and ÷(r2 , s) = r2 (1/s), and so ÷(r1 , s) and ÷(r2 , s)
can be viewed as different “scalings” of the multiplicative inverse 1/s of s.
This is actually a special case of the previous example. In particular, we can define a
function f ∈ C(R \ {0}) by f (s) = 1/s, for each s ∈ R \ {0}. Then, given any two real
numbers r1 , r2 ∈ R, the functions r1 f and r2 f can be defined by
r1 f (·) = ÷(r1 , ·) and r2 f (·) = ÷(r2 , ·), respectively.
4. Strictly speaking, there is nothing in the definition that precludes S from equalling F.
Consequently, addition, subtraction, and multiplication can all be seen as examples of
scaling operations on R.
As with binary operations, it is easy to define any number of scaling operations upon
a given nonempty set S. However, we are generally only interested in operations that are
essentially like scalar multiplication on Rn , and it is also quite common to additionally impose
conditions for how scaling operations should interact with any binary operations that might
also be defined upon S. We make this precise when we present an alternate formulation of
the definition for a vector space in Section B.2.
Put another way, the definitions for binary operation and scaling operation are not par-
ticularly useful when taken as is. Since these operations are allowed to be any functions
having the proper domains, there is no immediate sense of meaningful abstraction. Instead,
binary and scaling operations become useful when additionally conditions are placed upon
them so that they can be used to abstract “arithmetic-like” properties. In other words,
we are usually only interested in operations that abstract the salient properties of familiar
operations for combining things like numbers, n-tuples, and functions.
B.2 Groups, fields, and vector spaces

We begin this section with the following definition, which is unequivocably one of the most
fundamental and ubiquitous notions in all of abstract mathematics.
B.2. GROUPS, FIELDS, AND VECTOR SPACES 231
Definition B.2.1. Let G be a nonempty set, and let ∗ be a binary operation on G. (In
other words, ∗ : G × G → G is a function with ∗(a, b) denoted by a ∗ b, for each a, b ∈ G.)
Then G is said to form a group under ∗ if the following three conditions are satisfied:
1. (associativity) Given any three elements a, b, c ∈ G,
(a ∗ b) ∗ c = a ∗ (b ∗ c).
2. (existence of an identity element) There is an element e ∈ G such that, given any

element a ∈ G,
a ∗ e = e ∗ a = a.
3. (existence of inverse elements) Given any element a ∈ G, there is an element b ∈ G

such that
a ∗ b = b ∗ a = e.
You should recognize these three conditions (which are sometimes collectively referred
to as the group axioms) as properties that are satisfied by the operation of addition on
R. This is not an accident. In particular, given real numbers α, β ∈ R, the group axioms
form the minimal set of assumptions needed in order to solve the equation x + α = β for
the variable x, and it is in this sense that the group axioms are an abstraction of the most
fundamental properties of addition of real numbers.
A similar remark holds regarding multiplication on R \ {0} and solving the equation
αx = β for the variable x. Note, however, that this cannot be extended to all of R.
Because the group axioms are so general, they are particularly useful in building more
complicated algebraic structures. This is done by adding any number of additional axioms,
the most fundamental of which is as follows.
Definition B.2.2. Let G be a group under binary operation ∗. Then G is called an abelian
group (a.k.a. commutative group) if, given any two elements a, b ∈ G, a ∗ b = b ∗ a.
Examples of groups are everywhere in abstract mathematics. We now give some of the
more important examples that occur in Linear Algebra. Please note, though, that these
examples are primarily aimed at motivating the definitions of more complicated algebraic
structures. (In general, groups can be much “stranger” than those below.)
Example B.2.3.
1. If G ∈ {Z, Q, R, C}, then G forms an abelian group under the usual definition of
addition.
Note, though, that the set Z+ of positive integers does not form a group under addition
since, e.g., it does not contain an additive identity element.
2. Similarly, if G ∈ { Q \ {0}, R \ {0}, C \ {0}}, then G forms an abelian group under

the usual definition of multiplication.
Note, though, that Z \ {0} does not form a group under multiplication since only ±1
have multiplicative inverses.
3. If m, n ∈ Z+ are positive integers and F denotes either R or C, then the set Fm×n of
all m × n matrices forms an abelian group under matrix addition.
Note, though, that Fm×n does not form a group under matrix multiplication unless
m = n = 1, in which case F1×1 = F.
4. Similarly, if n ∈ Z+ is a positive integer and F denotes either R or C, then the set

GL(n, F) of invertible n × n matrices forms a group under matrix multiplications. This
group, which is often called the general linear group, is non-abelian when n ≥ 2.
Note, though, that GL(n, F) does not form a group under matrix addition for any
/ GL(n, F).
choice of n since, e.g., the zero matrix 0n×n ∈
In the above examples, you should notice two things. First of all, it is important to
specify the operation under which a set might or might not be a group. Second, and perhaps
more importantly, all but one example is an abelian group. Most of the important sets in
Linear Algebra possess some type of algebraic structure, and abelian groups are the principal
building block of virtually every one of these algebraic structures. In particular, fields and
vector spaces (as defined below) and rings and algebra (as defined in Section B.3) can all be
described as “abelian groups plus additional structure”.
Given an abelian group G, adding “additional structure” amounts to imposing one or
more additional operation on G such that each new operations is “compatible” with the
preexisting binary operation on G. As our first example of this, we add another binary
operation to G in order to obtain the definition of a field:
B.2. GROUPS, FIELDS, AND VECTOR SPACES 233
Definition B.2.4. Let F be a nonempty set, and let + and ∗ be binary operations on F .
Then F forms a field under + and ∗ if the following three conditions are satisfied:
1. F forms an abelian group under +.
2. Denoting the identity element for + by 0, F \ {0} forms an abelian group under ∗.
3. (∗ distributes over +) Given any three elements a, b, c ∈ F ,
a ∗ (b + c) = a ∗ b + a ∗ c.
You should recognize these three conditions (which are sometimes collectively referred
to as the field axioms) as properties that are satisfied when the operations of addition
and multiplication are taken together on R. This is not an accident. As with the group
axioms, the field axioms form the minimal set of assumptions needed in order to abstract
fundamental properties of these familiar arithmetic operations. Specifically, the field axioms
guarantee that, given any field F , three conditions are always satisfied:
1. Given any a, b ∈ F , the equation x + a = b can be solved for the variable x.
2. Given any a ∈ F \ {0} and b ∈ F , the equation a ∗ x = b can be solved for x.
3. The binary operation ∗ (which is like multiplication on R) can be distributed over (i.e.,
is “compatible” with) the binary operation + (which is like addition on R).
Example B.2.5. It should be clear that, if F ∈ {Q, R, C}, then F forms a field under the
usual definitions of addition and multiplication.
Note, though, that the set Z of integers does not form a field under these operations since
Z \ {0} fails to form a group under multiplication. Similarly, none of the other sets from
Example B.2.3 can be made into a field.
In some sense Q, R, and C are the only easily describable fields. While there are many
other interesting and useful examples of fields, none of them can be described using entirely
familiar sets and operations. This is because the field axioms are extremely specific in
describing algebraic structure. As we will see in the next section, though, we can build
a much more general algebraic structure called a “ring” by still requiring that F form an
abelian group under + but simultaneously relaxing the requirement that F simultaneously
form an abelian group under ∗.
For now, though, we close this section by taking a completely different point of view.
Rather than place an additional (and multiplication-like) binary operation on an abelian
group, we instead impose a special type of scaling operation called scalar multiplication.
In essence, scalar multiplication imparts useful algebraic structure on an arbitrary nonempty
set S by indirectly imposing the algebraic structure of F as an abelian group under multi-
plication. (Recall that F can be replaced with either R or C.)
Definition B.2.6. Let S be a nonempty set, and let ∗ be a scaling operation on S. (In
other words, ∗ : F × S → S is a function with ∗(α, s) denoted by α ∗ s or even just αs, for
every α ∈ F and s ∈ S.) Then ∗ is called scalar multiplication if it satisfies the following
two conditions:
1. (existence of a multiplicative identity element for ∗) Denote by 1 the multiplicative

identity element for F. Then, given any s ∈ S, 1 ∗ s = s.
2. (multiplication in F is quasi-associative with respect to ∗) Given any α, β ∈ F and any

s ∈ S,
(αβ) ∗ s = α ∗ (β ∗ s).
Note that we choose to have the multiplicative part of F “act” upon S because we are
abstracting scalar multiplication as it is intuitively defined in Example B.1.4 on both Rn and
C(R). This is because, by also requiring a “compatible” additive structure (called vector
addition), we obtain the following alternate formulation for the definition of a vector space.
Definition B.2.7. Let V be an abelian group under the binary operation +, and let ∗ be
a scalar multiplication operation on V with respect to F. Then V forms a vector space
over F with respect to + and ∗ if the following two conditions are satisfied:
1. (∗ distributes over +) Given any α ∈ F and any u, v ∈ V ,
α ∗ (u + v) = α ∗ u + α ∗ v.
2. (∗ distributes over addition in F) Given any α, β ∈ F and any v ∈ V ,
(α + β) ∗ v = α ∗ v + β ∗ v.
B.3. RINGS AND ALGEBRAS 235
B.3 Rings and algebras

In this section, we briefly mention two other common algebraic structures. Specifically, we
we first “relax” the definition of a field in order to define a ring, and we then combine the
definitions of ring and vector space in order to define an algebra. In some sense, groups,
rings, and fields are the most fundamental algebraic structures, with vector spaces and
algebras being particularly important variants within the study of Linear Algebra and its
applications.
Definition B.3.1. Let R be a nonempty set, and let + and ∗ be binary operations on R.
Then R forms an (associative) ring under + and ∗ if the following three conditions are
satisfied:
1. R forms an abelian group under +.
2. (∗ is associative) Given any three elements a, b, c ∈ R, a ∗ (b ∗ c) = (a ∗ b) ∗ c.
3. (∗ distributes over +) Given any three elements a, b, c ∈ R,
a ∗ (b + c) = a ∗ b + a ∗ c and (a + b) ∗ c = a ∗ c + b ∗ c.
As with the definition of group, there are many additional properties that can be added to
a ring; here, each additional property makes a ring more field-like in some way.
Definition B.3.2. Let R be a ring under the binary operations + and ∗. Then we call R
• commutative if ∗ is a commutative operation; i.e., given any a, b ∈ R, a ∗ b = b ∗ a.
• unital if there is an identity element for ∗; i.e., if there exists an element i ∈ R such
that, given any a ∈ R, a ∗ i = i ∗ a = a.
• a commutative ring with identity (a.k.a. CRI) if it’s both commutative and unital.
In particular, note that a commutative ring with identity is almost a field; the only
thing missing is the assumption that every element has a multiplicative inverse. It is this
one difference that results in many familiar sets being CRIs (or at least unital rings) but
not fields. E.g., Z is a CRI under the usual operations of addition and multiplication, yet,
because of the lack of multiplicative inverses for all elements except ±1, Z is not a field.
In some sense, Z is the prototypical example of a ring, but there are many other familiar
examples. E.g., if F is any field, then the set of polynomials F [z] with coefficients from F
is a CRI under the usual operations of polynomial addition and multiplication, but again,
because of the lack of multiplicative inverses for every element, F [z] is itself not a field.
Another important example of a ring comes from Linear Algebra. Given any vector space
V , the set L(V ) of all linear maps from V into V is a unital ring under the operations of
function addition and composition. However, L(V ) is not a CRI unless dim(V ) ∈ {0, 1}.
Alternatively, if a ring R forms a group under ∗ (but not necessarily an abelian group),
then R is sometimes called a skew field (a.k.a. division ring). Note that a skew field is also
almost a field; the only thing missing is the assumption that multiplication is commutative.
Unlike CRIs, though, there are no simple examples of skew fields that are not also fields.
As you can probably imagine, many other properties that can be appended to the defini-
tion of a ring, some of which are more useful than others. We close this section by defining
the concept of an algebra over a field. In essence, an algebra is a vector space together
with a “compatible” ring structure. Consequently, anything that can be done with either a
ring or a vector space can also be done with an algebra.
Definition B.3.3. Let A be a nonempty set, let + and × be binary operations on A, and
let ∗ be scalar multiplication on A with respect to F. Then A forms an (associative)
algebra over F with respect to +, ×, and ∗ if the following three conditions are satisfied:
1. A forms an (associative) ring under + and ×.
2. A forms a vector space over F with respect to + and ∗.
3. (∗ is quasi-associative and homogeneous with respect to ×) Given any element α ∈ F

and any two elements a, b ∈ R,
α ∗ (a × b) = (α ∗ a) × b and α ∗ (a × b) = a × (α ∗ b).
Two particularly important examples of algebras were already defined above: F [z] (which
is unital and commutative) and L(V ) (which is, in general, just unital). On the other hand,
there are also many important sets in Linear Algebra that are not algebras. E.g., Z is a ring
that cannot easily be made into an algebra, and R3 is a vector space but cannot easily be
made into a ring (since the cross product operation from Vector Calculus is not associative).
Appendix C
Some Common Mathematical

Symbols and Abbreviations
(with History)
This Appendix contains a fairly long list of common mathematical symbols as well as a list
of some common Latin abbreviations and phrases. While you will not necessarily need all
of the included symbols for your study of Linear Algebra, this list will hopefully nonetheless
give you an idea of where much of our modern mathematical notation comes from.
Binary Relations
= (the equals sign) means “is the same as” and was first introduced in the 1557 book
The Whetstone of Witte by physician and mathematician Robert Recorde (c. 1510–
1558). He wrote, “I will sette as I doe often in woorke use, a paire of parralles, or
Gemowe lines of one lengthe, thus: =====, bicause noe 2 thynges can be moare
equalle.” (Recorde’s equals sign was significantly longer than the one in modern usage
and is based upon the idea of “Gemowe” or “identical” lines, where “Gemowe” means
“twin” and comes from the same root as the name of the constellation “Gemini”.)
Robert Recorde also introduced the plus sign, “+”, and the minus sign, “−”, in The
Whetstone of Witte.
237
238 APPENDIX C. SOME COMMON MATH SYMBOLS & ABBREVIATIONS
< (the less than sign) means “is strictly less than”, and > (the greater than sign)
means “is strictly greater than”. These first appeared in the book Artis Analyti-
cae Praxis ad Aequationes Algebraicas Resolvendas (“The Analytical Arts Applied
to Solving Algebraic Equations”) by mathematician and astronomer Thomas Harriot
(1560–1621), which was published posthumously in 1631.
Pierre Bouguer (1698–1758) later refined these to ≤ (“is less than or equals”) and ≥
(“is greater than or equals”) in 1734. Bouger is sometimes called “the father of naval
architecture” due to his foundational work in the theory of naval navigation.
:= (the equal by definition sign) means “is equal by definition to”. This is a com-
mon alternate form of the symbol “=Def ”, the latter having first appeared in the 1894
book Logica Matematica by logician Cesare Burali-Forti (1861–1931). Other common
def
alternate forms of the symbol “=Def ” include “ =” and “≡”, with “≡” being especially
common in applied mathematics.
≈ (the approximately equals sign) means “is approximately equal to” and was first in-
troduced in the 1892 book Applications of Elliptic Functions by mathematician Alfred
Greenhill (1847–1927).
.
Other modern symbols for “approximately equals” include “=” (read as “is nearly
equal to”), “∼
=” (read as “is congruent to”), “≃” (read as “is similar to”), “≍” (read
as “is asymptotically equal to”), and “∝” (read as “is proportional to”). Usage varies,
and these are sometimes used to denote varying degrees of “approximate equality”
within a given context.
Some Symbols from Mathematical Logic

∴ (three dots) means “therefore” and first appeared in print in the 1659 book Teusche
Algebra (“Teach Yourself Algebra”) by mathematician Johann Rahn (1622–1676).
Teusche Algebra also contains the first use of the obelus, “÷”, to denote division.
∵ (upside-down dots) means “because” and seems to have first appeared in the 1805
book The Gentleman’s Mathematical Companion. However, it is much more common
(and less ambiguous) to just abbreviate “because” as “b/c”.
239
∋ (the such that sign) means “under the condition that” and first appeared in the 1906
edition of Formulaire de mathematiqués by the logician Giuseppe Peano (1858–1932).
However, it is much more common (and less ambiguous) to just abbreviate “such that”
as “s.t.”.
There are two good reasons to avoid using “∋” in place of “such that”. First of all, the
abbreviation “s.t.” is significantly more suggestive of its meaning than is “∋”. Perhaps
more importantly, though, the symbol “∋” is now commonly used to mean “contains as
an element”, which is a logical extension of the usage of the unquestionably standard
symbol “∈” to mean “is contained as an element in”.
⇒ (the implies sign) means “logically implies that”, and ⇐ (the is implied by sign)
means “is logically implied by”. Both have an unclear historical origin. (E.g., “if it’s
raining, then it’s pouring” is equivalent to saying “it’s raining ⇒ it’s pouring.”)
⇐⇒ (the iff symbol) means “if and only if” (abbreviated “iff”) and is used to connect
two logically equivalent mathematical statements. (E.g., “it’s raining iff it’s pouring”
means simultaneously that “if it’s raining, then it’s pouring” and that “if it’s pouring,
then it’s raining”. In other words, the statement “it’s raining ⇐⇒ it’s pouring” means
simultaneously that “it’s raining ⇒ it’s pouring” and “it’s raining ⇐ it’s pouring”.)
The abbreviation “iff” is attributed to the mathematician Paul Halmos (1916–2006).
∀ (the universal quantifier) means “for all” and was first used in the 1935 publication
Untersuchungen ueber das logische Schliessen (“Investigations on Logical Reasoning”)
by logician Gerhard Gentzen (1909–1945). He called it the All-Zeichen (“all character”)
by analogy to the symbol “∃”, which means “there exists”.
∃ (the existential quantifier) means “there exists” and was first used in the 1897
edition of Formulaire de mathematiqués by the logician Giuseppe Peano (1858–1932).
(the Halmos tombstone or Halmos symbol) means “Q.E.D.”, which is an abbre-

viation of the Latin phrase quod erat demonstrandum (“which was to be proven”).
“Q.E.D.” has been the most common way to symbolize the end of a logical argument
for many centuries, but the modern convention of the “tombstone” is now generally
preferred both because it is easier to write and because it is visually more compact.
The symbol “” was first made popular by mathematician Paul Halmos (1916–2006).
Some Notation from Set Theory

⊂ (the is included in sign) means “is a subset of” and ⊃ (the includes sign) means
“has as a subset”. Both symbols were introduced in the 1890 book Vorlesungen über die
Algebra der Logik (“Lectures on the Algebra of the Logic”) by logician Ernst Schröder
(1841–1902).
∈ (the is in sign) means “is an element of” and first appeared in the 1895 edition of
Formulaire de mathematiqués by the logician Giuseppe Peano (1858–1932). Peano
originally used the Greek letter “ǫ” (viz. the first letter of the Latin word est for “is”).
The modern stylized version of this symbol was later introduced in the 1903 book
Principles of Mathematics by logician and philosopher Betrand Russell (1872–1970).
It is also common to use the symbol “∋” to mean “contains as an element”, which is
not to be confused with the more archaic usage of “∋” to mean “such that”.
∪ (the union sign) means “take the elements that are in either set”, and ∩ (the inter-
section sign) means “take the elements that the two sets have in common”. These
were both introduced in the 1888 book Calcolo geometrico secondo l’Ausdehnungslehre
di H. Grassmann preceduto dalle operazioni della logica deduttiva (“Geometric Calcu-
lus based upon the teachings of H. Grassman, preceded by the operations of deductive
logic”) by logician Giuseppe Peano (1858–1932).
∅ (the null set or empty set) means “the set without any elements in it” and was first
used in the 1939 book Éléments de mathématique by Nicolas Bourbaki. (Bourbaki is
the collective pseudonym for a group of primarily European mathematicians who have
written many mathematics books together.) It was borrowed simultaneously from the
Norwegian, Danish and Faroese alphabets by group member André Weil (1906–1998).
∞ (infinity) denotes “a quantity or number of arbitrarily large magnitude” and first

appeared in print in the 1655 publication De Sectionibus Conicus (“Tract on Conic
Sections”) by mathematician John Wallis (1616–1703).
Possible explanations for Wallis’ choice of “∞” include its resemblance to the symbol
“oo” (used by ancient Romans to denote the number 1000), to the final letter of the
Greek alphabet ω (used symbolically to mean the “final” number), and to a simple
curve called a “lemniscate”, which can be endlessly traversed with little effort.
241
Some Important Numbers in Mathematics

π (the ratio of the circumference to the diameter of a circle) denotes the num-
ber 3.141592653589 . . ., and was first used in the 1706 book Synopsis palmariorum
mathesios (“A New Introduction to Mathematics”) by mathematician William Jones
(1675–1749). The use of π to denote this number was then popularized by the great
mathematician Leonhard Euler (1707–1783) in his 1748 book Introductio in Analysin
Infinitorum. (It is speculated that Jones chose the letter “π” because it is the first
letter in the Greek word perimetron, πǫριµǫτ ρoν, which roughly means “around”.)
e = limn→∞ (1+ n1 )n (the natural logarithm base, also sometimes called Euler’s num-
ber) denotes the number 2.718281828459 . . ., and was first used in the 1728 manuscript
Meditatio in Experimenta explosione tormentorum nuper instituta (“Meditation on ex-
periments made recently on the firing of cannon”) by Leonhard Euler. (It is speculated
that Euler chose “e” because it is the first letter in the Latin word for “exponential”.)
The mathematician Edmund Landau (1877–1938) once wrote that, “The letter e may
now no longer be used to denote anything other than this positive universal constant.”
√
i = −1 (the imaginary unit) was first used in the 1777 memoir Institutionum calculi
integralis (“Foundations of Integral Calculus”) by Leonhard Euler.
The five most important numbers in mathematics are widely considered to be (in order)
0, 1, i, π, and e. These numbers are even remarkably linked by the equation eiπ +1 = 0,
which the physicist Richard Feynman (1918–1988) once called “the most remarkable
formula in mathematics”.
P
γ = limn→∞ ( nk=1 k1 − ln n) (the Euler-Mascheroni constant, also known as just
Euler’s constant), denotes the number 0.577215664901 . . ., and was first used in
the 1792 book Adnotationes ad Euleri Calculum Integralem (“Annotations to Euler’s
Integral Calculus”) by geometer Lorenzo Mascheroni (1750–1800).
The number γ is widely considered to be the sixth most important important number
in mathematics due to its frequent appearance in formulas from number theory and
applied mathematics. However, as of this writing, it is still not even known whether
or not γ is even an irrational number.
Some Common Latin Abbreviations and Phrases

i.e. (id est) means “that is” or “in other words”. (It is used to paraphrase a statement
that was just made, not to mean “for example”, and is always followed by a comma.)
e.g. (exempli gratia) means “for example”. (It is usually used to give an example of a
statement that was just made and is always followed by a comma.)
viz. (videlicet) means “namely” or “more specifically”. (It is used to clarify a statement
that was just made by providing more information and is never followed by a comma.)
etc. (et cetera) means “and so forth” or “and so on”. (It is used to suggest that the reader
should infer further examples from a list that has already been started and is usually
not followed by a comma.)
et al. (et alii ) means “and others”. (It is used in place of listing multiple authors past the
first and is never followed by a comma.) The abbreviation “et al.” can also be used
in place of et alibi, which means “and elsewhere”.
cf. (conferre) means “compare to” or “see also”. (It is used either to draw a comparison
or to refer the reader to somewhere else that they can find more information, and it is
never followed by a comma.)
q.v. (quod vide) means “which see” or “go look it up if you’re interested”. (It is used to
cross-reference a different written work or a different part of the same written work,
and it is never followed by a comma.) The plural form of “q.v.” is “q.q.”
v.s. (vide supra) means “see above”. (It is used to imply that more information can be
found before the current point in a written work and is never followed by a comma.)
N.B. (Nota Bene) means “note well” or “pay attention to the following”. (It is used to
imply that the wise reader will pay especially careful attention to what follows and is
never followed by a comma. Cf. the abbreviation “verb. sap.”)
verb. sap. (verbum sapienti sat est) means “a word to the wise is enough” or “enough has already
been said”. (It is used to imply that, while something may still be left unsaid, enough
has been said for the reader to infer the entire meaning.)
243
vs. (versus) means “against” or “in contrast to”. (It is used to contrast two things and is
never followed by a comma.) The abbreviation “vs.” is also often written as “v.”
c. (circa) means “around” or “near”. (It is used when giving an approximation, usually
for a date, and is never followed by a comma.) The abbreviation “c.” is also commonly
written as “ca.”, “cir.”, or “circ.”
ex lib. (ex libris) means “from the library of”. (It is used to indicate ownership of a book and
is never followed by a comma.).
• vice versa means “the other way around” and is used to indicate that an implication
can logically be reversed. (This is sometimes abbreviated as “v.v.”)
• a fortiori means “from the stronger” or “more importantly”.
• a priori means “from before the fact” and refers to reasoning that is done while an
event still has yet to happen.
• a posteriori means “from after the fact” and refers to reasoning that is done after an
event has already happened.
• ad hoc means “to this” and refers to reasoning that is specific to an event as it is
happening. (Such reasoning is regarded as not being generalizable to other situations.)
• ad infinitum means “to infinity” or “without limit”.
• ad nauseam means “causing sea-sickness” or “to excessive”.
• mutatis mutandis means “changing what needs changing” or “with the necessary
changes having been made”.
• non sequitur means “it does not follow” and refers to something that is out of place in
a logical argument. (This is sometimes abbreviated as “non seq.”)
• Me transmitte sursum, Caledoni! means, “Beam me up, Scotty!”
• Illud Latine dici non potest means “You can’t say that in Latin”.
• Quid quid latine dictum sit, altum videtur means something like, “Anything that is
said in Latin will sound profound.”
Appendix D
Summary of Notation Used

In addition to the notation for sets and functions (as reviewed in Appendix A), the nota-
tion for matrices and linear systems, and the common mathematical symbols reviewed in
Appendix C, the following notation is used frequently in the study of Linear Algebra.
Special Sets
• The set of positive integers is denoted by Z+ = {1, 2, 3, 4, . . .}.
• The set of integers is denoted by Z = {. . . , −2, −1, 0, 1, 2, . . .}.
• The set of real numbers is denoted by R.
• The set of complex numbers is denoted by C = {x + yi | x, y ∈ R}.

(F is often used to denote a set that can equally well be chosen as either R or C.)
• The set of polynomials of degree at most n in the variable z and with coefficients
over F is denoted by Fn [z] = {a0 + a1 z + a2 z 2 + · · · + an z n | a0 , a1 , . . . , an ∈ F}.
• The set of polynomials of all degrees in z with coefficients over F is denoted by F[z].
• The set of matrices of size m × n over F is denoted by Fm×n .
• The general linear group of n × n invertible matrices over F is denoted by GL(n, F).
• The set of continuous functions with domain D ⊂ R and codomain R is denoted by

C(D), and the set of smooth (a.k.a. infinitely differentiable) functions with domain
D ⊂ R and codomain R is denoted by C ∞ (D).
244
245
Complex Numbers
Given z = x + yi ∈ C with x, y ∈ R, and where i denotes the imaginary unit, we denote
• the additive inverse of z by −z = (−x) + (−y)i.

−1 x −y
• the multiplicative inverse of z by z = + i, assuming z 6= 0.
x2 + y 2 x2 + y 2
• the complex conjugate of z by z = x + (−y)i.
• the real part of z by Re(z) = x.
• the imaginary part of z by Im(z) = y.

p
• the modulus of z by |z| = x2 + y 2.
• the argument of z by Arg(z) = min { θ | x = cos(θ), y = sin(θ)}.

θ≥0
Vector Spaces
Let V be an arbitrary vector space, and let U1 and U2 be subspaces of V . Then we denote
• the additive identity of V by 0.
• the additive inverse of each v ∈ V by −v.
• the (subspace) sum of U1 and U2 by U1 + U2 .
• the direct sum of U1 and U2 by U1 ⊕ U2 .
• the span of v1 , v2 , . . . , vn ∈ V by span (v1 , v2 , . . . , vn ).
• the dimension of V by dim(V ), where



0 if V = {0} is the zero vector space,


dim(V ) = n if every basis for V has n ∈ Z+ elements in it,



∞ otherwise.
• the change of basis map with respect to a given basis B for V by [ · ]B : V → Rn ,

where V is assumed to be n-dimensional.
246 APPENDIX D. SUMMARY OF NOTATION USED
Linear Maps
Let U, V , and W denote vector spaces over the field F. Then we denote
• the vector space of all linear maps from V into W by L(V, W ) or HomF (V, W ).
• the vector space of all linear operators on V by L(V ) or HomF (V ).
• the composition (a.k.a. product) of S ∈ L(U, V ) and T ∈ L(V, W ) by T ◦ S (or,

equivalently, T S), where (T ◦ S)(u) = T (S(u)) for each u ∈ U.
• the null space (a.k.a. kernel) of T ∈ L(V, W ) by null (T ) = {v ∈ V | T (v) = 0}.
• the range of T ∈ L(V, W ) by range (T ) = {w ∈ W | w = T (v) for some v ∈ V }.
• the eigenspace of T ∈ L(V ) associated to eigenvalue λ ∈ C by Vλ = null (T − λidV ),

where idV denotes the identity map on V .
• the matrix of T ∈ L(V, W ) with respect to the basis B on V and with respect to the
basis C on W by M(T, B, C) (or simply as M(T )).
Inner Product Spaces

Let V be an arbitrary inner product space, and let U be a subspace of V . Then we denote
• the inner product on V by h·, ·i.

p
• the norm on V induced by h·, ·i as k · k = h·, ·i.
• the orthogonal complement of U by U ⊥ = {v ∈ V | hu, vi = 0, ∀ u ∈ V }.
• the orthogonal projection onto U by PU , which, for each v ∈ V , is defined by

PU (v) = u such that v = u + w for u ∈ U and w ∈ U ⊥ .
• the adjoint of the operator T ∈ L(V ) by T ∗ , where T ∗ satisfies hT (v), wi = hv, T ∗(w)i
for each v, w ∈ V .
√ √ √
• the square root of the positive operator T ∈ L(V ) by T , which satisfies T = T T .
√
• the positive part of the operator T ∈ L(V ) by |T | = T ∗ T .

Mat67 Course Notes

Uploaded by

Mat67 Course Notes

Uploaded by

Linear Algebra

As an Introduction to Abstract Mathematics

Copyright c 2007 by the authors.

1 What is Linear Algebra? 1

2 Introduction to Complex Numbers 11

5 Span and Bases 48

7 Eigenvalues and Eigenvectors 82

8 Permutations and the Determinant of a Square Matrix 99

9 Inner Product Spaces 117

10 Change of Bases 136

11 The Spectral Theorem for Normal Linear Maps 144

Supplementary Notes on Matrices and Linear Systems 161

A The language of Sets and Functions 214

B Summary of Algebraic Structures Encountered 227

C Some Common Math Symbols & Abbreviations 237

D Summary of Notation Used 244

What is Linear Algebra?

1.1 Introduction to MAT 67

3. In the setting of Linear Algebra, you will be introduced to abstraction. We will

1.2 What is Linear Algebra?

• Characterization of solutions: Are there solutions to a given system of linear

This system has a unique solution for x1 , x2 ∈ R, namely x1 = 1

From this, x2 = −2/3. Then, by further substitution,

Alternatively, we can a more systematic approach in eliminating variables. Here, for

In general a system of m linear equations in n unknowns x1 , x2 , . . . , xn is a collection

1.3 Systems of linear equations

1.3.1 Linear equations

Example 1.3.1. Let f : R → R be the function f (x) = x3 − x. Then f (x) = x3 − x = 1 is

In this setting, a system of equations is just another kind of equation.

Example 1.3.2. Let X = Y = R2 = R × R be the Cartesian product of the set of real

f (x1 , x2 ) = (2x1 + x2 , x1 − x2 ), (1.4)

two “linear” operations defined on R2 , namely addition and scalar multiplication:

x + y := (x1 + y1 , x2 + y2 ) (vector addition) (1.5)

f (cx) = cf (x) (1.7)

1.3.2 Non-linear equations

1.3.3 Linear transformations

and are called linear transformations.

1.3.4 Applications of linear equations

• Differential Equations (as in a course like MAT 22B or MAT 118AB);

• Fourier Analysis (as in a course like MAT 129);

Exercises for Chapter 1

(a) System of 3 equations in the unknowns x, y, z, w:

(b) System of 4 equations in the unknowns x, y, z:

(c) System of 3 equations in the unknowns x, y, z:

ax1 + bx2 = 0, (1.14)

Note that x1 = x2 = 0 is a solution for any choice of a, b, c, and d. Prove that if

Introduction to Complex Numbers

2.1 Definition of complex numbers

Definition 2.1.1. The set of complex numbers C is defined as

Note that if we write 1 = (1, 0), then we can express z = (x, y) ∈ C as

z = (x, y) = x(1, 0) + y(0, 1) = x1 + yi = x + yi.

It is often significantly easier to perform arithmetic operations on complex numbers when

2.2 Operations on complex numbers

2.2.1 Addition and subtraction of complex numbers

Example 2.2.2. (3, 2) + (17, −4.5) = (3 + 17, 2 − 4.5) = (20, −2.5).

1. (Associativity) (z1 + z2 ) + z3 = z1 + (z2 + z3 ).

2.2.2 Multiplication and division of complex numbers

(x1 , y1)(x2 , y2 ) = (x1 x2 − y1 y2 , x1 y2 + x2 y1 ).

According to this definition, i2 = −1. In other words, i is a solution of the polynomial

1. (Associativity) (z1 z2 )z3 = z1 (z2 z3 ).

4. (Distributivity of Multiplication over Addition) z1 (z2 + z3 ) = z1 z2 + z1 z3 .

Theorem 2.2.6 (continued).

5. (Multiplicative Inverses) Given z ∈ C with z 6= 0, there is a unique complex number,

Proof. (Uniqueness.) A complex number w is an inverse of z if zw = 1 (by the commutativity

and one can check that zw = 1.

2.2.3 Complex conjugation

Definition 2.2.8. Given a complex number z = (x, y) ∈ C with x, y ∈ R, we define the

Theorem 2.2.9. Given two complex numbers z1 , z2 ∈ C,

3. 1/z1 = 1/z1 , for all z1 6= 0.

4. z1 = z1 if and only if Im(z1 ) = 0.

6. the real and imaginary parts of z1 can be expressed as

2.2.4 The modulus (a.k.a. norm, length, or magnitude)

Definition 2.2.10. Given a complex number z = (x, y) ∈ C with x, y ∈ R, the modulus

In particular, given x ∈ R, note that

To see this geometrically, construct a figure in the Euclidean plane, such as

Theorem 2.2.12. Given two complex numbers z1 , z2 ∈ C,