MTH5112 Linear Algebra I, 2009-2010: Oscar Bandtlow 30 March 2010

MTH5112 Linear Algebra I, 2009–2010
Oscar Bandtlow
30 March 2010
ii
Contents
1 Systems of Linear Equations 1

1.1 Basic terminology and examples . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Gaussian elimination . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Special classes of linear systems . . . . . . . . . . . . . . . . . . . . . . . . . 7
2 Matrix Algebra 9
2.1 Revision from Geometry I . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2 Transpose of a matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.3 Special types of square matrices . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.4 Linear systems in matrix notation . . . . . . . . . . . . . . . . . . . . . . . . 14
2.5 Elementary matrices and the Invertible Matrix Theorem . . . . . . . . . . . . 16
2.6 Gauss-Jordan inversion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3 Determinants 23
3.1 General definition of determinants . . . . . . . . . . . . . . . . . . . . . . . . 23
3.2 Properties of determinants . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.3 Cramer’s Rule and a formula for A−1 . . . . . . . . . . . . . . . . . . . . . . 31
4 Vector Spaces 35
4.1 Definition and examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
4.2 Subspaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
4.3 The span of a set of vectors . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
4.4 Linear independence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.5 Basis and dimension . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.6 Coordinates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
4.7 Row space and column space . . . . . . . . . . . . . . . . . . . . . . . . . . 55
5 Linear Transformations 59
5.2 Linear transformations on general vector spaces . . . . . . . . . . . . . . . . . 61
5.3 Image and Kernel . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
5.4 Matrix representations of linear transformations . . . . . . . . . . . . . . . . . 65
5.5 Composition of linear transformations . . . . . . . . . . . . . . . . . . . . . . 68
5.6 Change of basis and similarity . . . . . . . . . . . . . . . . . . . . . . . . . . 70
6 Orthogonality 75
6.1 Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
6.2 Orthogonal complements . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
6.3 Orthogonal sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
6.4 Orthonormal sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
iii
iv CONTENTS
6.5 Orthogonal projections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82

6.6 Gram Schmidt process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
6.7 Least squares problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
7 Eigenvalues and Eigenvectors 87

7.2 Diagonalisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
7.3 Interlude: complex vector spaces and matrices . . . . . . . . . . . . . . . . . 96
7.4 Spectral Theorem for Symmetric Matrices . . . . . . . . . . . . . . . . . . . 98
Index 103
Chapter 1
Systems of Linear Equations
Systems of linear equations are the bread and butter of linear algebra. They arise in many
areas of the sciences, including physics, engineering, business, economics, and sociology. Their
systematic study also provided part of the motivation for the development of modern linear
algebra at the end of the 19th century.
The material in this chapter will be familiar from Geometry I, where systems of linear
equations have already been discussed in some detail. As this chapter is fundamental for what
is to follow, you want to make sure that the basic ideas are hardwired in your brain for the
rest of term!
1.1 Basic terminology and examples

A linear equation in n unknowns is an equation of the form
a1 x1 + a2 x2 + · · · + an xn = b ,
where a1 , . . . , an and b are given real numbers and x1 , . . . , xn are variables.
A system of m linear equations in n unknowns is a collection of equations of the form
a11 x1 + a12 x2 + · · · + a1n xn = b1
a21 x1 + a22 x2 + · · · + a2n xn = b2
..
.
am1 x1 + am2 x2 + · · · + amn xn = bm
where the aij ’s and bi ’s are all real numbers. We also call such systems m × n systems.
Example 1.1.
x1 − x2 = 0
2x1 + x2 = 4 x1 + x2 − x3 = 3
(a) (b) (c) x1 + x2 = 3 .
3x1 + 2x2 = 7 2x1 − x2 + x3 = 6
x2 = 1
(a) is a 2 × 2 system, (b) is a 2 × 3 system, and (c) is a 3 × 2 system.
A solution of an m × n system is an ordered n-tuple (x1 , x2 , . . . , xn ) that satisfies all
equations of the system.
Example 1.2. (1, 2) is a solution of Example 1.1 (a).
For each α ∈ R, the 3-tuple (3, α, α) is a solution of Example 1.1 (b) (CHECK!).
Example 1.1 (c) has no solution, since, on the one hand x2 = 1 by the last equation, but the
first equation implies x1 = 1, while the second equation implies x1 = 2, which is impossible.
1
2 CHAPTER 1. SYSTEMS OF LINEAR EQUATIONS
A system with no solution is called inconsistent, while a system with at least one solution
is called consistent.
The set of all solutions of a system is called its solution set, which may be empty if the
system is inconsistent.
The basic problem we want to address in this section is the following: given an arbitrary
m × n system, determine its solution set. Later on, we will discuss a procedure that provides
a complete and practical solution to this problem (the so-called ‘Gaussian algorithm’). Before
we encounter this procedure, we require a bit more terminology.
Definition 1.3. Two m × n systems are said to be equivalent, if they have the same solution
set.
Example 1.4. Consider the two systems
5x1 − x2 + 2x3 = −3 5x1 − x2 + 2x3 = −3

(a) x2 = 2 (b) −5x1 + 2x2 − 2x3 = 5 .
3x3 = 6 5x1 − x2 + 5x3 = 3
(a) is easy to solve: looking at the last equation we find first that x3 = 2; the second from
the bottom implies x2 = 2; and finally the first one yields x1 = (−3 + x2 − 2x3 )/5 = −1. So
the solution set of this system is {(−1, 2, 2)}.
To find the solution of (b), add the first and the second equation. Then x2 = 2, while
subtracting the first from the third equation gives 3x3 = 6, that is x3 = 2. Finally, the first
equation now gives x1 = (−3 + x2 − 2x3 )/5 = −1, so the solution set is again {(−1, 2, 2)}.
Thus the systems (a) and (b) are equivalent.
In solving system (b) above we have implicitly used the following important observation:
Lemma 1.5. The following operations do not change the solution set of a linear system:
(i) interchanging two equations;
(ii) multiplying an equation by a non-zero scalar;
(iii) adding a multiple of one equation to another.
Proof. (i) and (ii) are obvious. The proof of (iii) was an exercise in Geometry I (Coursework
4, Exercise 3) for the special case of two equations in three unknowns. The general case can
be proved in exactly the same way and is left as an exercise.
We shall see shortly how to use the above operations systematically to obtain the solution
set of any given linear system. Before doing so, however, we introduce a useful short-hand.
Given an m × n linear system
a11 x1 + · · · + a1n xn = b1
..
.
am1 x1 + · · · + amn xn = bm
we call the array  

a11 ··· a1n b1
 .. .. .. 
 . . . 
am1 · · · amn bm
1.2. GAUSSIAN ELIMINATION 3
the augmented matrix of the linear system, and the m × n matrix

 
a11 · · · a1n
 .. .. 
 . . 
am1 · · · amn
the coefficient matrix of the linear system.
Example 1.6.

3x1 + 2x2 − x3 = 5 3 2 −1 5
system: augmented matrix: .
2x1 + x3 = −1 2 0 1 −1
A system can be solved by performing operations on the augmented matrix. Corresponding
to the three operations given in Lemma 1.5 we have the following three operations that can
be applied to the augmented matrix, called elementary row operations.
Definition 1.7 (Elementary row operations).
Type I interchanging two rows;
Type II multiplying a row by a non-zero scalar;
Type III adding a multiple of one row to another row.
1.2 Gaussian elimination

Gaussian elimination is a systematic procedure to determine the solution set of a given lin-
ear systen. The basic idea is to perform elementary row operations on the corresponding
augmented matrix bringing it to a simpler form from which the solution set is readily obtained.
The simple form alluded to above is given in the following definition.
Definition 1.8. A matrix is said to be in row echelon form if it satisfies the following three
conditions:
(i) All zero rows (consisting entirely of zeros) are at the bottom.
(ii) The first non-zero entry from the left in each nonzero row is a 1, called the leading 1
for that row.
(iii) Each leading 1 is to the right of all leading 1’s in the rows above it.
A row echelon matrix is said to be in reduced row echelon form if, in addition it satisfies
the following condition:
(iv) Each leading 1 is the only nonzero entry in its column
Roughly speaking, a matrix is in row echelon form if the leading 1’s form an echelon (that
is, a ‘steplike’) pattern.
Example 1.9. Matrices in row echelon form:
   
1 4 2 1 3 1 0
0 0 1 2
1 3 , 0 0 1 3 , .
0 0 1
0 0 1 0 0 0 0
Matrices in reduced row echelon form:
     
1 2 0 1 0 1 5 0 2 1 0 3
0 0 1 2 0 , 0 0 1 −1 , 0 1 2 .
0 0 0 0 1 0 0 0 0 0 0 0
The variables corresponding to the leading 1’s of the augmented matrix in row echelon
form will be referred to as the leading variables, the remaining ones as the free variables.
Example 1.10.

1 2 3 −4 6
(a) .
0 0 1 2 3
Leading variables: x1 and x3 ; free variables: x2 and x4 .

1 0 5
(b) .
0 1 3
Leading variables: x1 and x2 ; no free variables.
Note that if the augmented matrix of a system is in row echelon form, the solution set is
easily obtained.
Example 1.11. Determine the solution set of the systems given by the following augmented
matrices in row echelon form:
 
1 −2 0 1 2
1 3 0 2
(a) , (b)  0 0 1 −2 1  .
0 0 0 1
0 0 0 0 0
Solution. (a) The corresponding system is
x1 + 3x2 = 2
0 = 1
so the system is inconsistent and the solution set is empty.

(b) The corresponding system is
x1 − 2x2 + x4 = 2
x3 − 2x4 = 1
0 = 0
We can express the leading variables in terms of the free variables x2 and x4 . So set x2 = α
and x4 = β, where α and β are arbitrary real numbers. The second line now tells us that
x3 = 1 + 2x4 = 1 + 2β, and then the first line that x1 = 2 + 2x2 − x4 = 2 + 2α − β. Thus
the solution set is { (2 + 2α − β, α, 1 + 2β, β) | α, β ∈ R }.
It turns out that every matrix can be brought into row echelon form using only elementary
row operations. The procedure is known as the
1.2. GAUSSIAN ELIMINATION 5
Gaussian algorithm:
Step 1 If the matrix consists entirely of zeros, stop — it is already in row echelon form.
Step 2 Otherwise, find the the first column from the left containing a non-zero entry (call it
a), and move the row containing that entry to the top position.
Step 3 Now multiply that row by 1/a to create a leading 1.
Step 4 By subtracting multiples of that row from rows below it, make each entry below the
leading 1 zero.
This completes the first row. All further operations are carried out on the other rows.
Step 5 Repeat steps 1-4 on the matrix consisting of the remaining rows
The process stops when either no rows remain at Step 5 or the remaining rows consist of
zeros.
Example 1.12. Solve the following system using the Gaussian algorithm:
x2 + 6x3 =4
3x1 − 3x2 + 9x3 = −3
2x1 + 2x2 + 18x3 = 8
Solution. Performing the Gaussian algorithm on the augmented matrix gives:

    1  
0 1 6 4 3 −3 9 −3 R 1 1 −1 3 −1
 3 −3 9 −3  ∼ R1 ↔ R2  0 1 6 4  ∼
3
0 1 6 4 
2 2 18 8 2 2 18 8 2 2 18 8
     
1 −1 3 −1 1 −1 3 −1 1 −1 3 −1
∼ 0 1 6 4  ∼ 0 1 6 4 ∼ 0 1 6 4  ,
R3 − 2R1 0 4 12 10 R3 − 4R2 0 0 −12 −6 − 12 R3 0 0 1 12
1
where the last matrix is now in row echelon form. The corresponding system reads:
x1 − x2 + 3x3 = −1
x2 + 6x3 = 4
x3 = 21
Leading variables are x1 , x2 and x3 ; there are no free variables. The last equation now implies
x3 = 12 ; the second equation from bottom yields x2 = 4−6x 33= 11and finally the first equation
3
yields x1 = −1 + x2 − 3x3 = − 2 . Thus the solution is (− 2 , 1, 2 ) .
A variant of the Gauss algorithm is the Gauss-Jordan algorithm, which brings a matrix to
reduced row echelon form:
Gauss-Jordan algorithm
Step 1 Bring matrix to row echelon form using the Gaussian algorithm.
Step 2 Find the row containing the first leading 1 from the right, and add suitable multiples
of this row to the rows above it to make each entry above the leading 1 zero.
This completes the first non-zero row from the bottom. All further operations are carried out
on the rows above it.
Step 3 Repeat steps 1-2 on the matrix consisting of the remaining rows.
Example 1.13. Solve the following system using the Gauss-Jordan algorithm:
x1 + x2 + x3 + x4 + x5 = 4
x1 + x2 + x3 + 2x4 + 2x5 = 5
x1 + x2 + x3 + 2x4 + 3x5 = 7
Solution. Performing the Gauss-Jordan algorithm on the augmented matrix gives:

     
1 1 1 1 1 4 1 1 1 1 1 4 1 1 1 1 1 4
1 1 1 2 2 5  ∼ R2 − R1  0 0 0 1 1 1 ∼ 0 0 0 1 1 1
1 1 1 2 3 7 R3 − R1 0 0 0 1 2 3 R3 − R2 0 0 0 0 1 2
   
R1 − R3 1 1 1 1 0 2 R1 − R2 1 1 1 0 0 3
∼ R2 − R3  0 0 0 1 0 −1  ∼  0 0 0 1 0 −1  ,
0 0 0 0 1 2 0 0 0 0 1 2
where the last matrix is now in reduced row echelon form. The corresponding system reads:
x1 + x2 + x3 = 3
x4 = −1
x5 = 2
Leading variables are x1 , x4 , and x5 ; free variables x2 and x3 . Now set x2 = α and
x3 = β, and solve for the leading variables starting from the last equation. This yields
x5 = 2, x4 = −1, and finally x1 = 3 − x2 − x3 = 3 − α − β. Thus the solution set is
{ (3 − α − β, α, β, −1, 2) | α, β ∈ R }.
We have just seen that any matrix can be brought to (reduced) row echelon form using
only elementary row operations, and moreover that there is an explicit procedure to achieve
this (namely the Gaussian and Gauss-Jordan algorithm). We record this important insight for
later use.
Theorem 1.14.
(a) Every matrix can be brought to row echelon form by a series of elementary row opera-
tions.
(b) Every matrix can be brought to reduced row echelon form by a series of elementary row
operations.
Proof. For (a):apply the Gaussian algorithm; for (b): apply the Gauss-Jordan algorithm.
1.3. SPECIAL CLASSES OF LINEAR SYSTEMS 7
Remark 1.15. It can be shown (but not in this module) that the reduced row echelon form
of a matrix is unique. A moment’s thought (to which you are warmly invited) shows that this
is not the case for the row echelon form.
The remark above implies that if a matrix is brought to reduced row echelon form by
any sequence of elementary row operations (that is, not necessarily by those prescribed by the
Gauss-Jordan algorithm) the leading ones will nevertheless always appear in the same positions.
As a consequence, the following definition makes sense.
Definition 1.16. A pivot position in a matrix A is a location that corresponds to a leading
1 in the reduced row echelon form of A. A pivot column is a column of A that contains a
pivot position.
Example 1.17. Let  
1 1 1 1 1 4
A = 1 1 1 2 2 5 .
1 1 1 2 3 7
By Example 1.13 the reduced row echelon form of A is
 
1 1 1 0 0 3
0 0 0 1 0 −1 ,
0 0 0 0 1 2
Thus the pivot positions of A are the (1, 1)-entry, the (2, 4)-entry, and the (3, 5)-entry and
the pivot columns of A are columns 1, 4, and 5.
The notion of a pivot position and a pivot column will come in handy later in the module.
1.3 Special classes of linear systems

In this last section of our first chapter we’ll have a look at a number of special types of linear
systems and derive the first important consequences of the fact that every matrix can be
brought to row echelon form by a series of elementary row operations.
We start with the following classification of linear systems:
Definition 1.18. An m × n linear system is said to be
• overdetermined if it has more equations than unknowns (i.e. m > n);
• underdetermined if it has fewer equations than unknowns (i.e. m < n).

Note that overdetermined systems are usually (but not necessarily) inconsistent. Under-
determined systems may or may not be consistent. However, if they are consistent, then they
necessarily have infinitely many solutions:
Theorem 1.19. If an underdetermined system is consistent, it must have infinitely many
solutions.
Proof. Note that the row echelon form of the augmented matrix of the system has r ≤ m
non-zero rows. Thus there are r leading variables, and consequently n − r ≥ n − m > 0 free
variables.
Another useful classification of linear systems is the following:
Definition 1.20. A linear system
a11 x1 + a12 x2 + · · · + a1n xn = b1

a21 x1 + a22 x2 + · · · + a2n xn = b2
.. (1.1)
.
am1 x1 + am2 x2 + · · · + amn xn = bm
is said to be homogeneous if bi = 0 for all i. Otherwise it is said to be inhomogeneous.

Given an inhomogeneous system (1.1), call the system obtained by setting all bi ’s to zero,
the associated homogeneous system .
Example 1.21.
3x1 + 2x2 + 5x3 = 2 3x1 + 2x2 + 5x3 = 0

2x1 − x2 + x3 = 5 2x1 − x2 + x3 = 0
| {z } | {z }
inhomogeneous system associated homogeneous system
The first observation about homogeneous systems is that they always have a solution, the
so-called trivial or zero solution: (0, 0, . . . , 0).
For later use we record the following useful consequence of the previous theorem on con-
sistent homogeneous systems:
Theorem 1.22. An underdetermined homogeneous system always has non-trivial solutions.
Proof. We just observed that a homogeneous systems is consistent. Thus, if the system is
underdetermined and homogeneous, it must have infinitely many solutions by Theorem 1.19,
hence, in particular, it must have a non-zero solution.
Our final result in this section is devoted to the special case of n × n systems. For such
systems there is a delightful characterisation of the existence and uniqueness of solutions of a
given system in terms of the associated homogeneous systems. At the same time, the proof of
this result serves as another illustration of the usefulness of the row echelon form for theoretical
purposes.
Theorem 1.23. An n × n system is consistent and has a unique solution, if and only if the
only solution of the associated homogeneous system is the zero solution.
Proof. Follows from the following two observations:
• The same sequence of elementary row operations that brings the augmented matrix
of a system to row echelon form, also brings the augmented matrix of the associated
homogeneous system to row echelon form, and vice versa.
• An n×n system in row echelon form has a unique solution precisely if there are n leading
variables.
Thus, if an n × n system is consistent and has a unique solution, the corresponding homoge-
neous system must have a unique solution, which is necessarily the zero solution.
Conversely, if the associated homogeneous system of a given system has the zero solution
as its unique solution, then the original inhomogeneous system must have a solution, and this
solution must be unique.
Chapter 2
Matrix Algebra
2.1 Revision from Geometry I

Recall that an m × n matrix A is a rectangular array of scalars (real numbers)
 
a11 · · · a1n
 .. ..  .
 . . 
am1 · · · amn
We write A = (aij )m×n or simply A = (aij ) to denote an m × n matrix whose (i, j)-entry is
aij , i.e. aij is the i-th row and in the j-th column.
If A = (aij )m×n we say that A has size m × n. An n × n matrix is said to be square.
Example 2.1. If
1 3 2
A= ,
−2 4 0
then A is a matrix of size 2 × 3. The (1, 2)-entry of A is 3 and the (2, 3)-entry of A is 0.
Definition 2.2 (Equality). Two matrices A and B are equal and we write A = B if they
have the same size and aij = bij where A = (aij ) and B = (bij ).
Definition 2.3 (Scalar multiplication). If A = (aij )m×n and α is a scalar, then αA (the scalar
product of α and A) is the m × n matrix whose (i, j)-entry is αaij .
Definition 2.4 (Addition). If A = (aij )m×n and B = (bij )m×n then the sum A + B of A
and B is the m × n matrix whose (i, j)-entry is aij + bij .
Example 2.5. Let    

2 3 0 1
A = −1 2 and B =  2 3 .
4 0 −2 1
Then      
6 9 0 2 6 11
3A + 2B = −3 6 +  4 6 = 1 12 .
12 0 −4 2 8 2
Definition 2.6 (Zero matrix). We write Om×n or simply O (if the size is clear from the
context) for the m × n matrix all of whose entries are zero, and call it a zero matrix.
9
10 CHAPTER 2. MATRIX ALGEBRA
Scalar multiplication and addition of matrices satisfy the following rules proved in Geom-
etry I:
Theorem 2.7. Let A, B and C be matrices of the same size, and let α and β be scalars.
Then:
(a) A + B = B + A;
(b) A + (B + C) = (A + B) + C;
(c) A + O = A;
(d) A + (−A) = O, where −A = (−1)A;
(e) α(A + B) = αA + αB;
(f) (α + β)A = αA + βA;
(g) (αβ)A = α(βA);
(h) 1A = A.
Example 2.8. Simplify 2(A + 3B) − 3(C + 2B), where A, B, and C are matrices with the
same size.
Solution.
2(A + 3B) − 3(C + 2B) = 2A + 2 · 3B − 3C − 3 · 2B = 2A + 6B − 3C − 6B = 2A − 3C .
Definition 2.9 (Matrix multplication). If A = (aij ) is an m × n matrix and B = (bij ) is an

n × p matrix then the product AB of A and B is the m × p matrix C = (cij ) with
n
X
cij = aik bkj .
k=1
Example 2.10. Compute the (1, 3)-entry and the (2, 4)-entry of AB, where
 
2 1 6 0
3 −1 2
A= and B =  0 2 3 4 .
0 1 4
−1 0 5 8
Solution.
(1, 3)-entry: 3 · 6 + (−1) · 3 + 2 · 5 = 25;
(2, 4)-entry: 0 · 0 + 1 · 4 + 4 · 8 = 36.
Definition 2.11 (Identity matrix). An identity matrix I is a square matrix with 1’s on the
diagonal and zeros elsewhere. If we want to emphasise its size we write In for the n × n
identity matrix.
2.1. REVISION FROM GEOMETRY I 11
Matrix multiplication satisfies the following rules proved in Geometry I:
Theorem 2.12. Assume that α is a scalar and that A, B, and C are matrices so that the
indicated operations can be performed. Then:
(a) IA = A and BI = B;
(b) A(BC) = (AB)C;
(c) A(B + C) = AB + AC;
(d) (B + C)A = BA + CA;
(e) α(AB) = (αA)B = A(αB).
Notation 2.13.
• Since A(BC) = (AB)C, we can omit the brackets and simply write ABC and similarly
for products of more than three factors.
• If A is a square matrix we write Ak = AA · · · A} for the k-th power of A.

| {z
k factors
Warning: In general AB 6= BA, even if AB and BA have the same size!
Example 2.14.
1 0 0 1 0 1
=
0 0 0 0 0 0
but
0 1 1 0 0 0
=
0 0 0 0 0 0
Definition 2.15. If A and B are two matrices with AB = BA, then A and B are said to
commute.
Finally we recall the notion of an inverse of a matrix.
Definition 2.16. If A is a square matrix, a matrix B is called an inverse of A if
AB = I and BA = I .
A matrix that has an inverse is called invertible.
Note that not every matrix is invertible. For example the matrix

1 0
A=
0 0
cannot have an inverse since for any 2 × 2 matrix B = (bij ) we have

1 0 b11 b12 b11 b12
AB = = 6= I2 .
0 0 b21 b22 0 0
Later on in this chapter we shall discuss an algorithm that lets us decide whether a matrix
is invertible and at the same furnishes an inverse if the matrix is invertible.
It turns out that if a matrix is invertible its inverse is uniquely determined:
Theorem 2.17. If B and C are both inverses of A, then B = C.
Proof. Since B and C are inverses of A we have AB = I and CA = I. Thus
B = IB = (CA)B = C(AB) = CI = C .
If A is an invertible matrix, the unique inverse of A is denoted by A−1 . Hence A−1 (if it
exists!) is a square matrix of the same size as A with the property that
AA−1 = A−1 A = I .
Note that the above equality implies that if A is invertible, then its inverse A−1 is also invertible
with inverse A, that is,
(A−1 )−1 = A .
Slightly deeper is the following result:
Theorem 2.18. If A and B are invertible matrices of the same size, then AB is invertible
and
(AB)−1 = B −1 A−1 .
Proof. Observe that
(AB)(B −1 A−1 ) = A(BB −1 )A−1 = AIA−1 = AA−1 = I ,
(B −1 A−1 )(AB) = B −1 (A−1 A)B = B −1 IB = B −1 B = I .

Thus, by definition of invertibility, AB is invertible with inverse B −1 A−1 .
2.2 Transpose of a matrix

The first new concept we encounter is the following:
Definition 2.19. The transpose of an m × n matrix A = (aij ) is the n × m matrix B = (bij )

given by
bij = aji
The transpose of A is denoted by AT .
Example 2.20.  
1 4
1 2 3
(a) A = ⇒ AT = 2 5
4 5 6
3 6

1 2 T 1 3
(b) B = ⇒B =
3 −1 2 −1
2.3. SPECIAL TYPES OF SQUARE MATRICES 13
Matrix transposition satisfies the following rules:
Theorem 2.21. Assume that α is a scalar and that A, B, and C are matrices so that the
indicated operations can be performed. Then:
(a) (AT )T = A;
(b) (αA)T = α(AT );
(c) (A + B)T = AT + B T ;
(d) (AB)T = B T AT .
Proof. (a) is obvious while (b) and (c) are proved as Exercise 6 in Coursework 2. For the proof
of (d) assume A = (aij )m×n and B = (bij )n×p and write AT = (ãij )n×m and B T = (b̃ij )p×n
where
ãij = aji and b̃ij = bji .
Notice that (AB)T and B T AT have the same size, so it suffices to show that they have the
same entries. Now, the (i, j)-entry of B T AT is
n
X n
X n
X
b̃ik ãkj = bki ajk = ajk bki ,
k=1 k=1 k=1
which is the (j, i)-entry of AB, that is, the (i, j)-entry of (AB)T . Thus B T AT = (AB)T .
Transposition ties in nicely with invertibility:
Theorem 2.22. Let A be invertible. Then AT is invertible and
(AT )−1 = (A−1 )T .
Proof. See Exercise 8 in Coursework 2.
2.3 Special types of square matrices

In this section we briefly introduce a number of special classes of matrices which will be studied
in more detail later in this course.
Definition 2.23. A matrix is said to be symmetric if AT = A.
Note that a symmetric matrix is necessarily square.
1 2 4
5 2
symmetric:  2 −1 3 , .
2 −1
4 3 0
 
2 2 4
1 1 1
not symmetric:  2 2 3 .
1 1 1
1 3 5
Innocent as this definition may seem, symmetric matrices play an important role in many
parts of pure and applied Mathematics as well as in some of the ‘hard’ sciences. Some of
the reasons for this will become clearer towards the end of this course, when we shall study
symmetric matrices in much more detail.
Some other useful classes of square matrices are the triangular ones, which will also play
a role later on in the course.
Definition 2.25. A square matrix A = (aij ) is said to be

upper triangular if aij = 0 for i > j;
strictly upper triangular if aij = 0 for i ≥ j;
lower triangular if aij = 0 for i < j;
strictly lower triangular if aij = 0 for i ≤ j;
diagonal if aij = 0 for i 6= j.
If A = (aij ) is a square matrix of size n × n, we call a11 , a22 , . . . , ann the diagonal entries
of A. So, informally speaking, a matrix is upper triangular if all the entries below the diagonal
entries are zero, and it is strictly upper triangular if all entries below the diagonal entries and
the diagonal entries itself are zero. Similarly for (strictly) lower triangular matrices.
1 0 0 0
1 2 0 3 0 0
upper triangular: , diagonal:  
0 3 0 0 5 0
0 0 0 3
 
0 0 0
strictly lower triangular: −1 0 0 .
2 3 0
We close this section with the following two observations:
Theorem 2.27. The sum and product of two upper triangular matrices of the same size is
upper triangular.
Proof. See Exercise 5, Coursework 2.
2.4 Linear systems in matrix notation

We shall now have another look at systems of linear equations. The added spice in this
discussion will be that we now use the language of matrices to study them. More precisely,
we shall now introduce two equivalent ways of writing systems of linear equations. Both
reformulations will in their own way shed some light on both linear systems and matrices.
Before discussing these reformulations let us recall from Geometry I that an n × 1 matrix
 
a1
 a2 
 
 .. 
.
an
2.4. LINEAR SYSTEMS IN MATRIX NOTATION 15
is called a column vector of dimension n, or simply an n-vector. The collection of all

n-vectors is denoted by Rn . Thus:
  

 a1 

  a2 
 

n
R =  ..  a1 , a2 , · · · , an ∈ R .
 

  .  


 a 

n
Suppose now that we are given an m × n linear system
a11 x1 + a12 x2 + · · · + a1n xn = b1

a21 x1 + a22 x2 + · · · + a2n xn = b2
.. . (2.1)
.
am1 x1 + am2 x2 + · · · + amn xn = bm
The first reformulation is based on the observation that we can write this system more suc-
cinctly as a single matrix equation
Ax = b , (2.2)
where
     
a11 ··· a1n x1 b1
 .. ..  ,  ..   .. 
A= . .  x =  .  ∈ R , and b =  .  ∈ Rm ,
n
am1 · · · amn xn bm
and where Ax is interpreted as the matrix product of A and x.
Example 2.28. Using matrix notation the system
2x1 − 3x2 + x3 = 2
3x1 − x3 = −1
can be written  
x
2 −3 1  1 

2
x2 = .
3 0 −1 −1
| {z } x3 | {z }
=A | {z } =b
=x
Apart from obvious notational economy, writing (2.1) in the form (2.2) has a number of
other advantages which will become clearer shortly.
The other useful way of writing (2.1) is the following: with A and x as before we have
     
a11 x1 + · · · + a1n xn a11 a1n
..  .   . 
Ax =   = x1  ..  + · · · + xn  ..  ,
 
.
am1 x1 + · · · + amn xn am1 amn
| {z } | {z }
=a1 =an
where aj ∈ Rm is the j-th column of A.

Thus the linear system (2.1) can also be represented as a matrix (or vector) equation of
the form
x1 a1 + · · · + xn an = b . (2.3)
Example 2.29. The linear system
2x1 − 3x2 − 2x3 = 5

5x1 − 4x2 + 2x3 = 6
can also be written as

2 3 −2 5
x1 + x2 + x3 = . (2.4)
5 −4 2 6
Sums such as the left-hand side of (2.3) or (2.4) will turn up time and again in this course,
so it will be convenient to introduce the following terminology
Definition 2.30. If a1 , . . . , an are vectors in in Rm and α1 , . . . , αn are scalars, a sum of the

form
α1 a1 + · · · + αn an
is called a linear combination of the vectors a1 , . . . , an with weights α1 , . . . , αn .
Summarising the previous discussion, we now have the following characterisation of con-
sistency:
Theorem 2.31 (Consistency Theorem for Linear Systems). A linear system Ax = b is con-
sistent if and only if b can be written as a linear combination of the column vectors of A.
2.5 Elementary matrices and the Invertible Matrix The-

orem
Using the reformulation of linear systems discussed in the previous section we shall now have
another look at the process of solving them. Instead of performing elementary row operations
we shall now view this process in terms of matrix multiplication. This will shed some light
on both matrices and linear systems and will be useful for formulating and proving the main
result of this chapter, the Invertible Matrix Theorem, which will be presented towards the end
of this section. Before doing so, however, we shall consider the effect of multiplying both sides
of a linear system in matrix form by an invertible matrix.
Lemma 2.32. Let A be an m × n matrix and let b ∈ Rm . Suppose that M is an invertible

m × m matrix. The following two systems are equivalent:
Ax = b (2.5)
M Ax = M b (2.6)
Proof. Note that if x satisfies (2.5), then it clearly satisfies (2.6). Conversely, suppose that x
satisfies (2.6), that is,
M Ax = M b .
Since M is invertible, we may multiply both sides of the above equation by M −1 from the left
to obtain
M −1 M Ax = M −1 M b ,
so IAx = Ib, and hence Ax = b, that is, x satisfies (2.5).
2.5. ELEMENTARY MATRICES AND THE INVERTIBLE MATRIX THEOREM 17
We now come back to the idea outlined at the beginning of this section. It turns out that
we can ‘algebraize’ the process of applying an elementary row operation to a matrix A by
left-multiplying A by a certain type of matrix, defined as follows:
Definition 2.33. An elementary matrix of type I (respectively, type II, type III) is a
matrix obtained by applying an elementary row operation of type I (respectively, type II, type
III) to an identity matrix.
Example 2.34.
 
0 1 0
type I: E1 = 1
 0 0 (take I3 and swap rows 1 and 2)
0 0 1
 
1 0 0
type II: E2 = 0
 1 0 (take I3 and multiply row 3 by 4)
0 0 4
 
1 0 2
type III: E3 = 0
 1 0 (take I3 and add 2 times row 3 to row 1)
0 0 1
Let us now consider the effect of left-multiplying an arbitrary 3 × 3 matrix A in turn by

each of the three elementary matrices given in the previous example.
Example 2.35. Let A = (aij )3×3 and let El (l = 1, 2, 3) be defined as in the previous example.
Then
    
0 1 0 a11 a12 a13 a21 a22 a23
E1 A = 1 0 0 a21 a22 a23  = a11 a12 a13  ,
0 0 1 a31 a32 a33 a31 a32 a33
    
1 0 0 a11 a12 a13 a11 a12 a13
E2 A = 0 1 0 a21 a22 a23  =  a21 a22 a23  ,
0 0 4 a31 a32 a33 4a31 4a32 4a33
    
1 0 2 a11 a12 a13 a11 + 2a31 a12 + 2a32 a13 + 2a33
E3 A = 0 1 0 a21 a22 a23  =  a21 a22 a23 .
0 0 1 a31 a32 a33 a31 a32 a33
You should now pause and marvel at the following observation: interchanging rows 1 and 2
of A produces E1 A, multiplying row 3 of A by 4 produces E2 A, and adding 2 times row 3 to
row 1 of A produces E3 A.
This example should convince you of the truth of the following theorem, the proof of which
will be omitted as it is straightforward, slightly lengthy and not particularly instructive.
Theorem 2.36. If E is an m × m elementary matrix obtained from I by an elementary row

operation, then left-multiplying an m × n matrix A by E has the effect of performing that
same row operation on A.
Slightly deeper is the following:
Theorem 2.37. If E is an elementary matrix, then E is invertible and E −1 is an elementary

matrix of the same type.
Proof. The assertion follows from the previous theorem and the observation that an elementary
row operation can be reversed by an elementary row operation of the same type. More precisely,
• if two rows of a matrix are interchanged, then interchanging them again restores the
original matrix;
• if a row is multiplied by α 6= 0, then multiplying the same row by 1/α restores the
original matrix;
• if α times row q has been added to row r, then adding −α times row q to row r restores
the original matrix.
Now, suppose that E was obtained from I by a certain row operation. Then, as we just
observed, there is another row operation of the same type that changes E back to I. Thus
there is an elementary matrix F of the same type as E such that F E = I. A moment’s
thought shows that EF = I as well, since E and F correspond to reverse operations. All in
all, we have now shown that E is invertible and its inverse E −1 = F is an elementary matrix
of the same type.
Example 2.38. Determine the inverses of the elementary matrices E1 , E2 , and E3 in Exam-
ple 2.34.
Solution. In order to transform E1 into I we need to swap rows 1 and 2 of E1 . The elementary
matrix that performs this feat is
 
0 1 0
E1−1 = 1 0 0 .
0 0 1
Similarly, in order to transform E2 into I we need to multiply row 3 of E2 by 14 . Thus

 
1 0 0
E2−1 = 0 1 0  .
0 0 41
Finally, in order to transform E3 into I we need to add −2 times row 3 to row 1, and so
 
1 0 −2
E3−1 = 0 1 0  .
0 0 1
Before we come to the main result of this chapter we need some more terminology:
Definition 2.39. A matrix B is row equivalent to a matrix A if there exists a finite sequence
E1 , E2 , . . . , Ek of elementary matrices such that
B = Ek Ek−1 · · · E1 A .
In other words, B is row equivalent to A if and only if B can be obtained from A by a

finite number of row operations. In particular, two augmented matrices (A|b) and (B|c) are
row equivalent if and only if Ax = b and Bx = c are equivalent systems.
The following properties of row equivalent matrices are easily established:
2.5. ELEMENTARY MATRICES AND THE INVERTIBLE MATRIX THEOREM 19
1
Fact 2.40.
(a) A is row equivalent to itself;
(b) if A is row equivalent to B, then B is row equivalent to A;
(c) if A is row equivalent to B, and B is row equivalent to C, then A is row equivalent to

C.
Property (b) follows from Theorem 2.37. Details of the proof of (a), (b), and (c) are left
as an exercise.
We are now able to formulate and prove the first highlight of this module, a truly delightful
characterisation of invertibility of matrices. More precisely, the following theorem provides
three equivalent conditions for a matrix to be invertible. Later on in this course, we will
encounter further equivalent conditions.
Before stating the theorem we recall that the zero vector, denoted by 0, is the column
vector all of whose entries are zero.
Theorem 2.41 (Invertible Matrix Theorem). Let A be a square n × n matrix. The following
are equivalent:
(a) A is invertible;
(b) Ax = 0 has only the trivial solution;
(c) A is row equivalent to I;
(d) A is a product of elementary matrices.
Proof. We shall prove this theorem using a cyclic argument: we shall first show that (a)
implies (b), then (b) implies (c), then (c) implies (d), and finally that (d) implies (a). This is
a frequently used trick to show the logical equivalence of a list of assertions.
(a) ⇒ (b): Suppose that A is invertible. If x satisfies Ax = 0, then
x = Ix = (A−1 A)x = A−1 0 = 0 ,
so the only solution of Ax = 0 is the trivial solution.

(b) ⇒ (c): Use elementary row operations to bring the system Ax = 0 to the form
U x = 0, where U is in row echelon form. Since, by hypothesis, the solution of Ax = 0 and
hence the solution of U x = 0 is unique, there must be exactly n leading variables. Thus U is
upper triangular with 1’s on the diagonal, and hence, the reduced row echelon form of U is I.
Thus A is row equivalent to I.
(c) ⇒ (d): If A is row equivalent to I, then there is a sequence E1 , . . . , Ek of elementary
matrices such that
A = Ek Ek−1 · · · E1 I = Ek Ek−1 · · · E1 ,
that is, A is a product of elementary matrices.
(d) ⇒ (a). If A is a product of elementary matrices, then A must be invertible, since
elementary matrices are invertible by Theorem 2.37 and since the product of invertible matrices
is invertible by Theorem 2.18.
1
In the language of MTH4104 (Introduction to Algebra) which some of you will have taken, these statements
mean that ‘row equivalence’ is an equivalence relation.
An immediate consequence of the previous theorem is the following perhaps surprising

result:
Corollary 2.42. Let A and C be square matrices such that CA = I, then also AC = I; in
particular both A and C are invertible with C = A−1 and A = C −1 .
Proof. By Exercise 1 on Coursework 3 it follows that A is invertible. But then C is invertible
since
C = CI = CAA−1 = IA−1 = A−1 .
Furthermore, C −1 = (A−1 )−1 = A and AC = C −1 C = I.
What is surprising about this result is the following: suppose we are given a square matrix
A. If we want to check that A is invertible, then, by the definition of invertibility, we need
to produce a matrix B such that AB = I and BA = I. The above corollary tells us that
if we have a candidate C for an inverse of A it is enough to check that either AC = I or
CA = I in order to guarantee that A is invertible with inverse C. This is a non-trivial fact
about matrices, which is often useful.
2.6 Gauss-Jordan inversion

The Invertible Matrix Theorem provides a simple method for inverting matrices. Recall that
the theorem states (amongst other things) that if A is invertible, then A is row equivalent to
I. Thus there is a sequence E1 , . . . Ek of elementary matrices such that
Ek Ek−1 · · · E1 A = I .
Multiplying both sides of the above equation by A−1 from the right yields
Ek Ek−1 · · · E1 = A−1 ,
that is,
Ek Ek−1 · · · E1 I = A−1 .
Thus, the same sequence of elementary row operations that brings an invertible matrix to
I, will bring I to A−1 . This gives a practical algorithm for inverting matrices, known as
Gauss-Jordan inversion.
Note that in the following we use a slight generalisation of the augmented matrix notation.
Given an m × n matrix A and an m-vector b we currently use (A|b) to denote the m × (n + 1)
matrix consisting of A with b attached as an extra column to the right of A, and a vertical
line in between them. Suppose now that B is an m × r matrix then we write (A|B) for the
m × (n + r) matrix consisting of A with B attached to the right of A, and a vertical line
separating them.
Gauss-Jordan inversion
Bring the augmented matrix (A|I) to reduced row echelon form. If A is row equivalent to I,
then (A|I) is row equivalent to (I|A−1 ). Otherwise, A does not have an inverse.
Example 2.43. Show that  
1 2 0
A = 2 5 3
0 3 8
is invertible and compute A−1 .
2.6. GAUSS-JORDAN INVERSION 21
Solution. Using Gauss-Jordan inversion we find

   
1 2 0 1 0 0 1 2 0 1 0 0
 2 5 3 0 1 0  ∼ R2 − 2R1  0 1 3 −2 1 0
0 3 8 0 0 1 0 3 8 0 0 1
   
1 2 0 1 0 0 1 2 0 1 0 0
∼  0 1 3 −2 1 0 ∼   0 1 3 −2 1 0 
R3 − 3R2 0 0 −1 6 −3 1 (−1)R3 0 0 1 −6 3 −1
   
1 2 0 1 0 0 R1 − 2R2 1 0 0 −31 16 −6
∼ R2 − 3R3  0 1 0 16 −8 3  ∼  0 1 0 16 −8 3  .
0 0 1 −6 3 −1 0 0 1 −6 3 −1
Thus A is invertible (because it is row equivalent to I3 ) and

 
−31 16 −6
A−1 =  16 −8 3  .
−6 3 −1
Chapter 3
Determinants
3.1 General definition of determinants

Let A = (aij ) be a 2 × 2 matrix. Recall that the determinant of A was defined by

a11 a12
det(A) = = a11 a22 − a21 a12 . (3.1)
a21 a22
In other words, with every 2 × 2 matrix A it is possible to associate a scalar, called the
determinant of A, which is given by a certain sum of products of the entries of A. The
following fact was proved in Geometry I by a brute force calculation:
Fact 3.1. If A and B are 2 × 2 matrices then
(a) det(A) 6= 0 if and only if A is invertible;
(b) det(AB) = det(A) det(B).
This fact reveals one of the main motivations to introduce this somewhat non-intuitive
object: the determinant of a matrix allows us to decide whether a matrix is invertible or not.
In this chapter we introduce determinants for arbitrary square matrices, study some of their
properties, and then prove the generalisation of the above fact for arbitrary square matrices.
Before giving the general definition of the determinant of an n × n matrix, let us recall the
definition of 3 × 3 determinants given in Geometry I:
If A = (aij ) is a 3 × 3 matrix, then its determinant is defined by

a11 a12 a13
a22 a23 a12 a13 a12 a13
det(A) = a21 a22 a23 = a11 − a21 a32 a33 + a31 a22 a23 (3.2)

a31 a32 a33 a32 a 33
= a11 a22 a33 − a11 a32 a23 − a21 a12 a33 + a21 a32 a13 + a31 a12 a23 − a31 a22 a13 .
Notice that the determinant of a 3 × 3 matrix A is given in terms of the determinants of

certain 2 × 2 submatrices of A. In general, we shall see that the determinant of a 4 × 4 matrix
is given in terms of the determinants of 3 × 3 submatrices, and so forth. Before stating the
general definition we introduce a convenient short-hand:
23
24 CHAPTER 3. DETERMINANTS
Notation 3.2. For any square matrix A, let Aij denote the submatrix formed by deleting the
i-th row and the j-th column of A.
Warning: This notation differs from the one used in the course text
Example 3.3. If  
3 2 5 −1
−2 9 0 6
 7 −2 −3 1  ,
A= 
4 −5 8 −4
then  
3 2 −1
A23 = 7 −2 1  .
4 −5 −4
If we now define the determinant of a 1 × 1 matrix A = (aij ) by det(A) = a11 , we can
recast (3.1) and (3.2) as follows:
• if A = (aij )2×2 then
det(A) = a11 det(A11 ) − a21 det(A21 ) ;
• if A = (aij )3×3 then

det(A) = a11 det(A11 ) − a21 det(A21 ) + a31 det(A31 ) .
This observation motivates the following recursive definition:

Definition 3.4. Let A = (aij ) be an n × n matrix. The determinant of A, written det(A),
is defined as follows:
• If n = 1, then det(A) = a11 .
• If n > 1 then det(A) is the sum of n terms of the form ±ai1 det(Ai1 ), with plus and
minus signs alternating, and where the entries a11 , a21 , . . . , an1 are from the first column
of A. In symbols:
det(A) = a11 det(A11 ) − a21 det(A21 ) + · · · + (−1)n+1 an1 det(An1 )
Xn
= (−1)i+1 ai1 det(Ai1 ) .
i=1
Example 3.5. Compute the determinant of

 
0 0 7 −5
−2 9 6 −8
A= .
0 0 −3 2 
0 3 −1 4
Solution.

0 0 7 −5
0 7 −5
−2 9 6 −8 7 −5
0 0 −3 2 = −(−2) 0 −3 2 = 2 · 3 −3 2 = 2 · 3 · [7 · 2 − (−3) · (−5)] = −6 .

3 −1 4
0 3 −1 4
3.1. GENERAL DEFINITION OF DETERMINANTS 25
To state the next theorem, it will be convenient to write the definition of det(A) in a
slightly different form.
Definition 3.6. Given a square matrix A = (aij ), the (i, j)-cofactor of A is the number Cij
defined by
Cij = (−1)i+j det(Aij ) .
Thus, the definition of det(A) reads
det(A) = a11 C11 + a21 C21 + · · · + an1 Cn1 .
This is called the cofactor expansion down the first column of A. There is nothing special
about the first column, as the next theorem shows:
Theorem 3.7 (Cofactor Expansion Theorem). The determinant of an n × n matrix A can be

computed by a cofactor expansion across any column or row. The expansion down the j-th
column is
det(A) = a1j C1j + a2j C2j + · · · + anj Cnj
and the cofactor expansion across the i-th row is
det(A) = ai1 Ci1 + ai2 Ci2 + · · · + ain Cin .
Although this theorem is fundamental for the development of determinants, we shall not
prove it here, as it would lead to a rather lengthy digression.1
Before moving on, notice that the plus or minus sign in the (i, j)-cofactor depends on
the position of aij in the matrix, regardless of aij itself. The factor (−1)i+j determines the
following checkerboard pattern of signs
 
+ − + ···
− + − 
.
 
+ − +
..
 
..
. .
Example 3.8. Use a cofactor expansion across the second row to compute det(A), where
 
4 −1 3
A = 0 0 2 .
1 0 7
Solution.
det(A) = a21 C21 + a22 C22 + a23 C23

= (−1)2+1 a21 det(A21 ) + (−1)2+2 a22 det(A22 ) + (−1)2+3 a23 det(A23 )

−1 3 4 3 4 −1
= −0 + 0 − 2
0 7 1 7 1 0
= −2[4 · 0 − 1 · (−1)] = −2 .
1
If not knowing the proof causes you sleepless nights, ask me and I will point you towards one.
Example 3.9. Compute det(A), where

 
3 0 0 0 0
−2 5 0 0 0
 
A=  9 −6 4 −1 3 .

2 4 0 0 2
8 3 1 0 7
Solution. Notice that all entries but the first of row 1 are 0. Thus it will shorten our labours
if we expand across the first row:

5 0 0 0

−6 4 −1 3
det(A) = 3 .
4 0 0 2
3 1 0 7
Again it is advantageous to expand this 4 × 4 determinant across the first row:

4 −1 3

det(A) = 3 · 5 · 0 0 2 .
1 0 7
We have already computed the value of the above 3 × 3 determinant in the previous example
and found it to be equal to −2. Thus det(A) = 3 · 5 · (−2) = −30.
Notice that the matrix in the previous example was almost lower triangular. The method
of this example is easily generalised to prove the following theorem:
Theorem 3.10. If A is either an upper or a lower triangular matrix, then det(A) is the product
of the diagonal entries of A.
3.2 Properties of determinants

We saw time and again in this module that elementary row operations play a fundamental role
in matrix theory. It is only natural to enquire how det(A) behaves when an elementary row
operation is applied to A.
Theorem 3.11. Let A be a square matrix.
(a) If two rows of A are interchanged to produce B, then det(B) = − det(A).
(b) If one row of A is multiplied by α to produce B, then det(B) = α det(A).
(c) If a multiple of one row of A is added to another row to produce a matrix B then
det(B) = det(A).
Proof. These assertions follow from a slightly stronger result to be proved later in this chapter
(see Theorem 3.21).
Example 3.12.

1 2 3 4 5 6

(a) 4 5 6 = − 1 2 3 by (a) of the previous theorem.
7 8 9 7 8 9
3.2. PROPERTIES OF DETERMINANTS 27

0 1 2 0 1 2

(b) 3 12 9 = 3 1 4 3 by (b) of the previous theorem.

1 2 1 1 2 1

3 1 0 3 1 0

(c) 4 2 9 = 7 3 9 by (c) of the previous theorem.

0 −2 1 0 −2 1
The following examples show how to use the previous theorem for the effective computation
of determinants:
Example 3.13. Compute

3 −1 2 −5

0 5 −3 −6
−6 7 −7 4 .

−5 −8 0 9
Solution. Perhaps the easiest way to compute this determinant is to spot that when adding
two times row 1 to row 3 we get two identical rows, which, by another application of the
previous theorem, implies that the determinant is zero:

3 −1 2 −5 3 −1 2 −5

0 5 −3 −6 0 5 −3 −6
−6 7 −7 4 = R3 + 2R1 0

5 −3 −6
−5 −8 0 9 −5 −8 0 9

3 −1 2 −5

0 5 −3 −6
= = 0,
R3 − R2 0
0 0 0
−5 −8 0 9
by a cofactor expansion across the third row.
Example 3.14. Compute det(A), where

 
0 1 2 −1
2 5 −7 3 
A= 0
.
3 6 2
−2 −5 4 −2
Solution. Here we see that the first column already has two zero entries. Using the previous
theorem we can introduce another zero in this column by adding row 2 to row 4. Thus

0
1 2 −1 0 1 2 −1
2 5 −7 3 2 5 −7 3
det(A) = = .
0 3 6 2 0 3 6 2
−2 −5 4 −2 0 0 −3 1
If we now expand down the first column we see that

1 2 −1

det(A) = −2 3 6 2 .
0 −3 1
The 3 × 3 determinant above can be further simplified by subtracting 3 times row 1 from row
2. Thus
1 2
−1
det(A) = −2 0 0 5 .
0 −3 1
Finally we notice that the above determinant can be brought to triangular form by swapping
row 2 and row 3, which changes the sign of the determinant by the previous theorem. Thus

1 2 −1

det(A) = (−2) · (−1) 0 −3 1 = (−2) · (−1) · 1 · (−3) · 5 = −30 ,
0 0 5
by Theorem 3.10.
We are now able to prove the first important result about determinants. It allows us to
decide whether a matrix is invertible or not by computing its determinant. It will play an
important role in later chapters.
Theorem 3.15. A matrix A is invertible if and only if det(A) 6= 0.
Proof. Bring A to row echelon form U (which is then necessarily upper triangular). Since we
can achieve this using elementary row operations, and since, in the process we only ever multiply
a row by a non-zero scalar det(A) = γ det(U ) for some γ with γ 6= 0, by Theorem 3.11. If A
is invertible, then det(U ) = 1, since U is upper triangular with 1’s on the diagonal, and hence
det(A) = γ det(U ) 6= 0. Otherwise, at least one diagonal entry of U is zero, so det(U ) = 0,
and hence det(A) = γ det(U ) = 0.
Definition 3.16. A square matrix A is called singular if det(A) = 0. Otherwise it is said to

be nonsingular.
Corollary 3.17. A matrix is invertible if and only if it is nonsingular
Our next result shows what effect transposing a matrix has on its determinant:
Theorem 3.18. If A is an n × n matrix, then det(A) = det(AT ).
Proof. The proof is by induction on n (that is, the size of A).2 The theorem is obvious for
n = 1. Suppose now that it has already been proved for k × k matrices for some integer k.
Our aim now is to show that the assertion of the theorem is true for (k + 1) × (k + 1) matrices
as well. Let A be a (k + 1) × (k + 1) matrix. Note that the (i, j)-cofactor of A equals the
(i, j)-cofactor of AT , because the cofactors involve k × k determinants only, for which we
assumed that the assertion of the theorem holds. Hence
cofactor expansion of det(A) across first row

=cofactor expansion of det(AT ) down first column
so det(A) = det(AT ).
Let’s summarise: the theorem is true for 1 × 1 matrices, and the truth of the theorem for
k × k matrices for some k implies the truth of the theorem for (k + 1) × (k + 1) matrices.
2
If you have never encountered this method of proof, don’t despair! Simply read through the following
argument. The last paragraph explains the underlying idea of this method.
3.2. PROPERTIES OF DETERMINANTS 29
Thus, the theorem must be true for 2 × 2 matrices (choose k = 1); but since we now know
that it is true for 2 × 2 matrices, it must be true for 3 × 3 matrices as well (choose k = 2);
continuing with this process, we see that the theorem must be true for matrices of arbitrary
size.
By the previous theorem, each statement of the theorem on the behaviour of determinants
under row operations (Theorem 3.11) is also true if the word ‘row’ is replaced by ‘column’,
since a row operation on AT amounts to a column operation on A.
Theorem 3.19. Let A be a square matrix.
(a) If two columns of A are interchanged to produce B, then det(B) = − det(A).
(b) If one column of A is multiplied by α to produce B, then det(B) = α det(A).
(c) If a multiple of one column of A is added to another column to produce a matrix B
then det(B) = det(A).
Example 3.20. Find det(A) where
 
1 3 4 8
−1 2 1 9
A=
2
.
5 7 0
3 −4 −1 5
Solution. Adding column 1 to column 2 gives

1 3 4 8 1 4 4 8

−1 2 1 9 −1 1
1 9
det(A) = = .
2 5 7 0 2 7 7 0

3 −4 −1 5 3 −1 −1 5
Now subtracting column 3 from column 2 the determinant is seen to vanish by a cofactor
expansion down column 2.

1 0 4 8

−1 0 1 9
det(A) = = 0.
2 0 7 0

3 0 −1 5
Our next aim is to prove that determinants are multiplicative, that is, det(AB) = det(A) det(B)
for any two square matrices A and B of the same size. We start by establishing a baby-version
of this result, which, at the same time, proves the theorem on the behaviour of determinants
under row operations stated earlier (see Theorem 3.11).
Theorem 3.21. If A is an n × n matrix and E an elementary n × n matrix, then
det(EA) = det(E) det(A)
with

−1
 if E is of type I (interchanging two rows)
det(E) = α if E is of type II (multiplying a row by α) .

1 if E is of type III (adding a multiple of one row to another)

Proof. By induction on the size of A. The case where A is a 2×2 matrix is an easy exercise (see
Exercise 1, Coursework 4). Suppose now that the theorem has been verified for determinants
of k ×k matrices for some k with k ≥ 2. Let A be (k +1)×(k +1) matrix and write B = EA.
Expand det(EA) across a row that is unaffected by the action of E on A, say, row i. Note
that Bij is obtained from Aij by the same type of elementary row operation that E performs
on A. But since these matrices are only k × k, our hypothesis implies that
det(Bij ) = r det(Aij ) ,
where r = −1, α, 1 depending on the nature of E.

Now by a cofactor expansion across row i
k+1
X
det(EA) = det(B) = aij (−1)i+j det(Bij )
j=1
k+1
X
= aij (−1)i+j r det(Aij )
j=1
= r det(A) .
In particular, taking A = Ik+1 we see that det(E) = −1, α, 1 depending on the nature of E.
To summarise: the theorem is true for 2 × 2 matrices and the truth of the theorem for
k × k matrices for some k ≥ 2 implies the truth of the theorem for (k + 1) × (k + 1) matrices.
By the principle of induction the theorem is true for matrices of any size.
Using the previous theorem we are now able to prove the second important result of this
chapter:
Theorem 3.22. If A and B are square matrices of the same size, then
det(AB) = det(A) det(B) .
Proof. Case I: If A is not invertible, then neither is AB (for otherwise A(B(AB)−1 ) = I,

which by the corollary to the Invertible Matrix Theorem (Corollary 2.42), would force A to be
invertible). Thus, by Theorem 3.15,
det(AB) = 0 = 0 · det(B) = det(A) det(B) .
Case II: If A is invertible, then by the Invertible Matrix Theorem A is a product of elementary
matrices, that is, there exist elementary matrices E1 , . . . , Ek , such that
A = Ek Ek−1 · · · E1 .
For brevity, write |A| for det(A). Then, by the previous theorem,
|AB| = |Ek · · · E1 B| = |Ek ||Ek−1 · · · E1 B| = . . .

= |Ek | · · · |E1 ||B| = . . . = |Ek · · · E1 ||B|
= |A||B| .
3.3. CRAMER’S RULE AND A FORMULA FOR A−1 31
3.3 Cramer’s Rule and a formula for A−1

In the following, we use a shorthand to identify a matrix by its columns. Let A be an m × n
matrix. If aj ∈ Rm is the j-th column vector of A, then we write
A = (a1 . . . an ) .
Note that if B is an l × m matrix then, by the definition of matrix multiplication,
BA = (Ba1 . . . Ban ) ,
that is, the j-th column of BA is Baj .

Cramer’s Rule is a curious formula that allows us to write down the solution for certain
n × n systems in terms of quotients of two determinants. Before stating it we need some more
notation.
For any n × n matrix A and any b ∈ Rn write Ai (b) for the matrix obtained from A by
replacing column i by b, that is,
Ai (b) = (a1 . . . b . . . an ) .
col i
Theorem 3.23 (Cramer’s Rule). Let A be an invertible n × n matrix. For any b ∈ Rn , the
unique solution x of Ax = b has entries given by
det(Ai (b))
xi = for i = 1, . . . , n .
det(A)
Proof. Let a1 . . . , an be the columns of A, and e1 , . . . , en the columns of the n × n identity
matrix I. Then
AIi (x) = A(e1 . . . x . . . en )

= (Ae1 . . . Ax . . . Aen )
= (a1 . . . b . . . an )
= Ai (b) ,
and hence
det(A) det(Ii (x)) = det(Ai (b)) .
But det(Ii (x)) = xi by a cofactor expansion across row i, so
det(Ai (b))
xi = ,
det(A)
since det(A) 6= 0.
Example 3.24. Use Cramer’s Rule to solve the system
3x1 − 2x2 = 6
.
−5x1 + 4x2 = 8
Solution. Write the system as Ax = b, where

3 −2 6
A= , b= .
−5 4 8
Then
6 −2 3 6
A1 (b) = , A2 (b) = ,
8 4 −5 8
and Cramer’s Rule gives
det(A1 (b)) 40
x1 = = = 20 ,
det(A) 2
det(A2 (b)) 54
x2 = = = 27 .
det(A) 2
Cramer’s Rule is not really useful for practical purposes (except for very small systems),
since evaluation of determinants is time consuming when the system is large. For 3×3 systems
and larger, you are better off using Gaussian elimination. Apart from its intrinsic beauty, its
main strength is as a theoretical tool. For example, it allows you to study how sensitive the
solution of Ax = b is to a change in an entry in A or b.
As an application of Cramer’s Rule, we shall now derive an explicit formula for the inverse
of a matrix. Before doing so we shall have another look at the process of inverting a matrix.
Again, denote the columns of In by e1 , . . . , en . The Gauss-Jordan inversion process bringing
(A|I) to (I|A−1 ) can be viewed as solving the n systems
Ax = e1 , Ax = e2 , ... Ax = en .
Thus the j-th column of A−1 is the solution of
Ax = ej ,
and the i-th entry of x is the (i, j)-entry of A−1 . By Cramer’s rule
det(Ai (ej ))
(i, j)-entry of A−1 = xi = . (3.3)
det(A)
A cofactor expansion down column i of Ai (ej )) shows that
det(Ai (ej )) = (−1)i+j det(Aji ) = Cji ,
where Cji is the (j, i)-cofactor of A. Thus, by (3.3), the (i, j)-entry of A−1 is the cofactor
Cji divided by det(A) (note that the order of the indices is reversed!). Thus
 
C11 C21 · · · Cn1
1   C12 C22 · · · Cn2 

−1
A =  . .. ... ..  . (3.4)
det(A)  .. . . 
C1n C2n · · · Cn
| {z }
= adj (A)
The matrix of cofactors on the right of (3.4) is called the adjugate of A, and is denoted by
adj (A). The following theorem is simply a restatement of (3.4):
Theorem 3.25 (Inverse Formula). Let A be an invertible matrix. Then
1
A−1 = adj (A) .
det(A)
3.3. CRAMER’S RULE AND A FORMULA FOR A−1 33
Example 3.26. Find the inverse of the following matrix using the Inverse Formula
 
1 3 −1
A = −2 −6 0  .
1 4 −3
Proof. First we need to calculate the 9 cofactors of A:

−6 0 −2 0 −2 −6
C11 = + = 18 , C12 = − = −6 , C13 = + = −2 ,
4 −3 1 −3 1 4

3 −1 1 −1 1 3
C21 = − = 5, C22 = + = −2 , C23 = − = −1 ,
4 −3 1 −3 1 4

3 −1 1 −1 1 3
C31 = + = −6 , C32 = −
−2 0 = 2 , C33 = +
= 0.
−6 0 −2 −6
Thus  
18 5 −6
adj (A) = −6 −2 2  ,
−2 −1 0
and since det(A) = 2, we have
5
 
9 2
−3
A−1 = −3 −1 1  .
−1 − 12 0
Note that the above calculations are just as laborious as if we had used the Gauss-Jordan
inversion process to compute A−1 . As with Cramer’s Rule, the deceptively neat formula for
the inverse is not useful if you want to invert larger matrices. As a rule, for matrices larger
than 3 × 3 the Gauss-Jordan inversion algorithm is much faster.
Chapter 4
Vector Spaces
4.1 Definition and examples

In this chapter, we will study abstract vector spaces. Roughly speaking a vector space is a
mathematical structure on which an operation of addition and an operation of scalar multi-
plication is defined, and we require these operations to obey a number of algebraic rules. We
have already encountered examples of vector spaces in this module. Recall that Rn is the
collection of all n-vectors. On Rn two operations were defined:
• addition: if    
x1 y1
x =  ...  ∈ Rn ,
 .. 
and y =  .  ∈ Rn ,
 
xn yn
then x + y is the n-vector given by
 
x1 + y1
x + y =  ...  .
 
xn + yn
• scalar multiplication: if
 
x1
x =  ...  ∈ Rn , and α is a scalar
 
xn
then αx is the n-vector given by

 
αx1
αx =  ...  .
 
αxn
After these operations were defined, it turned out that they satisfy a number of rules (see
Theorem 2.7).1 We are now going to turn this process on its head. That is, we start from
a set on which two operations are defined, we postulate that these operations satisfy certain
rules, and we call the resulting structure a ‘vector space’:
1
Go back and look them up!
35
36 CHAPTER 4. VECTOR SPACES
Definition 4.1. A vector space is a non-empty set V on which are defined two operations,
called addition and scalar multiplication, such that the following axioms hold for all u, v, w
in V and all scalars α, β:
(C1) the sum of u and v, denoted by u + v, is in V ;
(C2) the scalar multiple of u by α, denoted by αu, is in V;
(A1) u + v = v + u;
(A2) u + (v + w) = (u + v) + w;
(A3) there is an element 0 in V such that u + 0 = u;
(A4) for each u in V there is an element −u in V such that u + (−u) = 0;
(A5) α(u + v) = αu + αv;
(A6) (α + β)u = αu + βu;
(A7) (αβ)u = α(βu);
(A8) 1u = u.
We will refer to V as the universal set for the vector space. Its elements are called vectors,
and we usually write them using bold letters u, v, w, etc.
The term ‘scalar’ will usually refer to a real number, although later on we will sometimes
allow scalars to be complex numbers. To distinguish these cases we will use the term real
vector space (if the scalars are real numbers) or complex vector space (if the scalars are
complex numbers). For the moment, however, we will only consider real vector spaces.
Note that in the above definition the axioms (C1) and (C2), known as closure axioms,
simply state that the two operations produce values in V , that is, they are bona fide operations
on V . The other eight axioms, also known as the classical vector space axioms, stipulate how
the two operations interact.
Let’s have a look at some examples:
Example 4.2. Let Rm×n denote the set of all m × n matrices. Define addition and scalar
multiplication of matrices in the usual way. Then Rn×m is a vector space by Theorem 2.7.
Example 4.3. Let Pn denote the set of all polynomials with real coefficients of degree less
than n. Thus, an element p in Pn is of the form
p(t) = a0 + a1 t + a2 t2 + · · · + an tn ,
where the coefficients a0 , . . . , an and the variable t are real numbers.

Define addition and scalar multiplication on Pn as follows: if q ∈ Pn is given by
q(t) = b0 + b1 t + b2 t2 + · · · + bn tn ,
p is as above and α a scalar, then
• p + q is the polynomial
(p + q)(t) = (a0 + b0 ) + (a1 + b1 )t + · · · + (an + bn )tn

4.1. DEFINITION AND EXAMPLES 37
• αp is the polynomial
(αp)(t) = (αa0 ) + (αa1 )t + · · · + (αan )tn .
Note that (C1) and (C2) clearly hold, since if p, q ∈ Pn and α is a scalar, then p + q and αp
are again polynomials of degree less than n. Axiom (A1) holds since if p and q are as above,
then
(p + q)(t) = (a0 + b0 ) + (a1 + b1 )t + · · · + (an + bn )tn

= (b0 + a0 ) + (b1 + a1 )t + · · · + (bn + an )tn
= (q + p)(t)
so p + q = q + p. A similar calculation shows that (A2) holds. Axiom (A3) holds if we let 0
be the zero polynomial, that is
0(t) = 0 + 0 · t + · · · + 0 · tn ,
since then (p + 0)(t) = p(t), that is, p + 0 = p. Axiom (A4) holds if, given p ∈ Pn we set
−p = (−1)p, since then
(p + (−p))(t) = (a0 − a0 ) + (a1 − a1 )t + · · · + (an − an )tn = 0(t) ,
that is p+(−p) = 0. The remaining axioms are easily verified as well, using familiar properties
of real numbers.
Example 4.4. Let C[a, b] denote the set of all real-valued functions that are defined and
continuous on the closed interval [a, b]. For f , g ∈ C[a, b] and α a scalar, define f + g and αf
pointwise, that is, by
(f + g)(t) = f (t) + g(t) for all t ∈ [a, b]
(αf )(t) = αf (t) for all t ∈ [a, b]

Equipped with these operations, C[a, b] is a vector space. The closure axiom (C1) holds
because the sum of two continuous functions on [a, b] is continuous on [a, b], and (C2) holds
because a constant times a continuous function on [a, b] is again continuous on [a, b]. Axiom
(A1) holds as well, since for all t ∈ [a, b]
(f + g)(t) = f (t) + g(t) = g(t) + f (t) = (g + f )(t) ,
so f + g = g + f . Axiom (A3) is satisfied if we let 0 be the zero function,
0(t) = 0 for all t ∈ [a, b] ,
since then
(f + 0)(t) = f (t) + 0(t) = f (t) + 0 = f (t) ,
so f + 0 = f . Axiom (A4) holds if, given f ∈ C[a, b], we let −f be the function
(−f )(t) = −f (t) for all t ∈ [a, b],
since then
(f + (−f ))(t) = f (t) + (−f )(t) = f (t) − f (t) = 0 = 0(t) ,
that is, f + (−f ) = 0. We leave it as an exercise to verify the remaining axioms.
We shall now derive a number of elementary properties of vector spaces.

Theorem 4.5. If V is a vector space and u and v are elements in V , then
(a) 0u = 0;
(b) if u + v = 0 then v = −u;2
(c) (−1)u = −u.

Proof. (a) We start by observing that
(A8) (A6) (A8)
u = 1u = (0 + 1)u = 0u + 1u = 0u + u . (4.1)
Now, by (A4), there is an element −u ∈ V such that
u + (−u) = 0 . (4.2)
Thus
(4.2) (4.1) (A2) (4.2) (A3)
0 = u + (−u) = (0u + u) + (−u) = 0u + (u + (−u)) = 0u + 0 = 0u .
(b) Suppose that u + v = 0. Then
(A3) (A2)
− u = −u + 0 = −u + (u + v) = (−u + u) + v
(A1) (A4) (A1) (A3)
= (u + (−u)) + v = 0 + v = v + 0 = v .
(c) Notice that

(a) (A6) (A8)
0 = 0u = (1 + (−1))u = 1u + (−1)u = u + (−1)u ,
so, by (b), we conclude that (−1)u = −u.
4.2 Subspaces
Given a vector space V , a ‘subspace’ of V is, roughly speaking, a subset of V that inherits
the vector space structure from V , and can thus be considered as a vector space in its own
right. One of the main motivations to consider such ‘substructures’ of vector spaces, is the
following. As you might have noticed, it can be frightfully tedious to check whether a given
set, call it H, is a vector space. Suppose that we know that H is a subset of a larger set V
equipped with two operations (addition and scalar multiplication), for which we have already
checked that the vector space axioms are satisfied. Now, in order for H to be the universal set
of a vector space equipped with the operations of addition and scalar multiplication inherited
from V , the set H should certainly be closed under addition and scalar multiplication (so that
(C1) and (C2) are satisfied). Somewhat surprisingly, checking these two axioms is enough in
order for H to be a vector space in its own right, as we shall see shortly. To summarise: if H
is a subset of a vector space V , and if H is closed under addition and scalar multiplication,
then H is a vector space in its own right. So instead of having to check 10 axioms, we only
need to check two in this case. Let’s cast these observations into the following definition:
2
In the language of MTH4104 (Introduction to Algebra) this statement says that the additive inverse is
unique.
4.2. SUBSPACES 39
Definition 4.6. Let H be a nonempty subset of a vector space V . Suppose that H satisfies
the following two conditions:
(i) if u, v ∈ H, then u + v ∈ H;
(ii) if u ∈ H and α is a scalar, then αu ∈ H.
Then H is said to be a subspace of V .
Theorem 4.7. Let H be a subspace of a vector space V . Then H with addition and scalar
multiplication inherited from V is a vector space in its own right.
Proof. Clearly, by definition of a subspace, (C1) and (C2) are satisfied. Axioms (A3) and (A4)
follow from Theorem 4.5 and condition (ii) of the definition of a subspace. The remaining
axioms are valid for any elements in V , so, in particular, they are valid for any elements in H
as well.
Remark 4.8. If V is a vector space, then {0} and V are clearly subspaces of V . All other
subspaces are said to be proper subspaces of V . We call {0} the zero subspace of V .
Let’s have a look at some more concrete examples:
Example 4.9. Show that the following are subspaces of R3 :

(a) L = (r, s, t, )T r, s, t ∈ R and r = s = t ;3

(b) P = (r, s, t, )T r, s, t ∈ R and r − s + 3t = 0 .
Solution. (a) Notice that an arbitrary element in L is of the form r(1, 1, 1)T for some real
number r. Thus, in particular, L is not empty, since (0, 0, 0)T ∈ L. In order to check that L
is a subspace of R3 we need to check that conditions (i) and (ii) of Definition 4.6 are satisfied.
We start with condition (i). Let x1 and x2 belong to L. Then x1 = r1 (1, 1, 1)T and
x2 = r2 (1, 1, 1)T for some real numbers r1 and r2 , so
     
1 1 1
x1 + x2 = r1 1 + r2 1 = (r1 + r2 ) 1 ∈ L .
    
1 1 1
Thus condition (i) holds.

We now check condition (ii). Let x ∈ L and let α be a real number. Then x = r(1, 1, 1)T
for some real number r ∈ R, so
 
1
αx = αr 1 ∈ L .
1
Thus condition (ii) holds.

3
In order to save paper, hence trees and thus do our bit to prevent climate change, we shall sometimes
write n-vectors x ∈ Rn in the form (x1 , . . . , xn )T . So, for example,
 
2
(2, 3, 1)T = 3 .
1
Let’s summarise: the non-empty set L satisfies conditions (i) and (ii), that is, it is closed
under addition and scalar multiplication, hence L is a subspace of R3 as claimed.
(b) In order to see that P is a subspace of R3 we first note that (0, 0, 0)T ∈ P , so P is not
empty.
Next we check condition (i). Let x1 = (r1 , s1 , t1 )T ∈ P and x2 = (r2 , s2 , t2 )T ∈ P . Then
r1 − s1 + 3t1 = 0 and r2 − s2 + 3t2 = 0, so
 
r1 + r2
x1 + x2 = s1 + s2  ∈ P ,
t1 + t2
since (r1 + r2 ) − (s1 + s2 ) + 3(t1 + t2 ) = (r1 − s1 + 3t1 ) + (r2 − s2 + 3t2 ) = 0 + 0 = 0. Thus

condition (i) holds.
We now check condition (ii). Let x = (r, s, t)T ∈ P and let α be a scalar. Then
r − s + 3t = 0 and  
αr
αx = αs ∈ P
αt
since αr − αs + 3αt = α(r − s + 3t) = 0. Thus condition (ii) holds as well.
As P is closed under addition and scalar multiplication, P is a subspace of R3 as claimed.
Remark 4.10. In the example above the two subspaces L and P of R3 can also be thought
of as geometric objects. More precisely, L can be interpreted geometrically as a line through
the origin with direction vector (1, 1, 1)T , while P can be interpreted as a plane through the
origin with normal vector (1, −1, 3)T .
More generally, all proper subspaces of R3 can be interpreted geometrically as either lines or
planes through the origin. Similarly, all proper subspaces of R2 can be interpreted geometrically
as lines through the origin.
If this is not crystal clear to you, think about it!

Example 4.11. H = (r2 , s, r)T r, s ∈ R is not a subspace of R3 , since
     
1 1 2
0 ∈ H, but 2 0 = 0 6∈ H .
1 1 2
Example 4.12. The set

a b
2×2
H= ∈ R a, b ∈ R ,
0 1
is not a subspace of R2×2 . In order to see this, note that every subspace must contain the
zero vector. However,
O2×2 6∈ H .
Example 4.13. Let H = { f ∈ C[−2, 2] | f (1) = 0 }. Then H is a subspace of C[−2, 2]. First
observe that the zero function is in H, so H is not empty. Next we check that the closure
properties are satisfied.
Let f , g ∈ H . Then f (1) = 0 and g(1) = 0, so
(f + g)(1) = f (1) + g(1) = 0 + 0 = 0 ,

4.2. SUBSPACES 41
so f + g ∈ H. Thus H is closed under addition.

Let f ∈ H and α be a a real number. Then f (1) = 0 and
(αf )(1) = αf (1) = α · 0 = 0 ,
so αf ∈ H. Thus H is closed under scalar multiplication.

Since H is closed under addition and scalar multiplication it is a subspace of C[−2, 2] as
claimed.
A class of subspaces we have already encountered (but didn’t think about them in this way)
are the solution sets of homogeneous systems. More precisely, if A ∈ Rm×n is the coefficient
matrix of such a system, then the solution set can be thought of as the collection of all x ∈ Rn
with Ax = 0, and the collection of all such x is a subspace of Rn . Before convincing us of
this fact, we introduce some convenient terminology:
Definition 4.14. Let A ∈ Rm×n . Then
N (A) = { x ∈ Rn | Ax = 0 }
is called the nullspace of A.
Theorem 4.15. If A ∈ Rm×n , then N (A) is a subspace of Rn .
Proof. Clearly 0 ∈ N (A), so N (A) is not empty.

If x, y ∈ N (A) then Ax = 0 and Ay = 0, so
A(x + y) = Ax + Ay = 0 + 0 = 0 ,
and hence x + y ∈ N (A).

Furthermore, if x ∈ N (A) and α is a real number then Ax = 0 and
A(αx) = α(Ax) = α0 = 0 ,
so αx ∈ N (A).
Thus N (A) is a subspace of Rn as claimed.
Example 4.16. Determine N (A) for

 
−3 6 −1 1 −7
A =  1 −2 2 3 −1 .
2 −4 5 8 −4
Solution. We need to find the solution set of Ax = 0. To do this you can use your favourite
method to solve linear systems. Perhaps the fastest one is to bring the augmented matrix (A|0)
to reduced row echelon form and write the leading variables in terms of the free variables. In
our case, we have
   
−3 6 −1 1 −7 0 1 −2 0 −1 3 0
 1 −2 2 3 −1 0  ∼ · · · ∼  0 0 1 2 −2 0  .
2 −4 5 8 −4 0 0 0 0 0 0 0
The leading variables are x1 and x3 , and the free variables are x2 , x4 and x5 . Now setting
x2 = α, x4 = β and x5 = γ we find x3 = −2x4 + 2x5 = −2β + 2γ and x1 = 2x2 + x4 − 3x5 =
2α + β − 3γ. Thus
         
x1 2α + β − 3γ 2 1 −3
x2   α 1 0 0
      
   
x3  =  −2β + 2γ  = α 0 + β −2 + γ  2  ,
         
x4   β  0 1 0
x5 γ 0 0 1
hence        

 2 1 −3


1 0  0 

 
  

   


N (A) = α 0 + β −2 + γ  2  α, β, γ ∈ R .
     


 0 1  0  


 
0 0 1
 
4.3 The span of a set of vectors

In this section we shall have a look at a way to construct subspaces from a collection of vectors.
Definition 4.17. Let v1 , . . . , vn be vectors in a vector space. The set of all linear combi-
nations4 of v1 , . . . , vn is called the span of v1 , . . . , vn and is denoted by Span (v1 , . . . , vn ),
that is,
Span (v1 , . . . , vn ) = { α1 v1 + · · · + αn vn | α1 , . . . , αn ∈ R } .
Example 4.18. Let e1 , e2 , e3 ∈ R3 be given by
     
1 0 0
e1 = 0 , e2 = 1 ,
   e3 = 0 .

0 0 1
Determine Span (e1 , e2 ) and Span (e1 , e2 , e3 ).

Solution. Since
  
α1 α1
α1 e1 + α2 e2 = α2  , while α1 e1 + α2 e2 + α3 e3 = α2  ,
0 α3
we see that
  
 x1

3
Span (e1 , e2 ) =  x2 ∈ R x3 = 0 ,
 while Span (e1 , e2 , e3 ) = R3 .
x3
 
Notice that in the above example Span (e1 , e2 ) can be interpreted geometrically as the
x1 , x2 plane, that is, the plane containing the x1 - and the x2 -axis. In particular, Span (e1 , e2 )
is a subspace of R3 . This is true more generally:
4
In case you forgot what this means go back to Definition 2.30.
4.3. THE SPAN OF A SET OF VECTORS 43
Example 4.19. Given vectors v1 and v2 in a vector space V , show that H = Span (v1 , v2 )
is a subspace of V .
Solution. Notice that 0 ∈ H (since 0 = 0v1 + 0v2 ), so H is not empty. In order to show that
H is closed under addition, let u and w be arbitrary vectors in H. Then there are scalars α1 ,
α2 and β1 , β2 , such that
u = α1 v1 + α2 v2 ,
w = β1 v1 + β2 v2 .
Now by axioms (A1), (A2) and (A6)
u + w = (α1 v1 + α2 v2 ) + (β1 v1 + β2 v2 ) = (α1 + β1 )v1 + (α2 + β2 )v2 ,
so u + w ∈ H.
In order to show that H is closed under scalar multiplication, let u ∈ H, say, u =
α1 v1 + α2 v2 , and let γ be a scalar. Then, by axioms (A5) and (A7)
γu = γ(α1 v1 + α2 v2 ) = (γα1 )v1 + (γα2 )v2 ,
so γu ∈ H.
More generally, using exactly the same method of proof, it is possible to show the following:
Theorem 4.20. Let v1 , . . . , vn be vectors in a vector space V . Then Span (v1 , . . . , vn ) is a

subspace of V .
We have just seen that the span of a collection of vectors in a vector space V is a subspace
of V . As we saw in Example 4.18, the span may be a proper subspace of V , or it may be
equal to all of V . The latter is sufficiently interesting a case to merit its own definition:
Definition 4.21. Let V be a vector space, and let v1 , . . . , vn ∈ V . We say that the set
{v1 , . . . , vn } is a spanning set for V if
Span (v1 , . . . , vn ) = V .
If {v1 , . . . , vn } is a spanning set for V , we shall also say that {v1 , . . . , vn } spans V , that
v1 , . . . , vn span V or that V is spanned by v1 , . . . , vn .
Notice that the above definition can be rephrased as follows. A set {v1 , . . . , vn } is a
spanning set for V , if and only if every vector in V can be written as a linear combination of
v1 , . . . , vn .
Example 4.22. Which of the following sets are spanning sets for R3 ?
(a) (1, 0, 0)T , (0, 1, 0)T , (0, 0, 1)T , (1, 2, 4)T (b) (1, 1, 1)T , (1, 1, 0)T , (1, 0, 0)T

(c) (1, 0, 1)T , (0, 1, 0)T (d) (1, 2, 4)T , (2, 1, 3)T , (4, −1, 1)T

Solution. (a) Let (a, b, c)T be an arbitrary vector in R3 . Clearly

         
a 1 0 0 1
 b  = a 0 + b 1 + c 0 + 0 2 ,
c 0 0 1 4
so the set is a spanning set for R3 .

(b) Let (a, b, c)T be an arbitrary vector in R3 . We need to determine whether it is possible to
find constants α1 , α2 , α3 such that
       
1 1 1 a
α1 1 + α2 1 + α3 0 = b  .
      
1 0 0 c
This means that the weights α1 , α2 and α3 have to satisfy the system
α1 + α2 + α3 = a
α1 + α2 = b
α1 = c
Since the coefficient matrix of the system is nonsingular, the system has a unique solution. In
fact, using back substitution we find
   
α1 c
α2  =  b − c  .
α3 a−b
Thus        
a 1 1 1
 b  = c 1 + (b − c) 1 + (a − b) 0 ,
c 1 0 0
so the set is a spanning set for R3 .
(c) Noting that      
1 0 α1
α1 0 + α2 1 = α2  ,
    
1 0 α1
we see that a vector of the form (a, b, c)T with a 6= c cannot be in the span of the two vectors.
Thus the set is not a spanning set for R3 .
(d) Proceeding as in (b), we let (a, b, c)T be an arbitrary vector in R3 . Again, we need to
determine whether it is possible to find constants α1 , α2 , α3 such that
       
1 2 4 a
α1 2 + α2 1 + α3 −1 = b  .
      
4 3 1 c
This means that the weights α1 , α2 and α3 have to satisfy the system
α1 + 2α2 + 4α3 = a
2α1 + α2 − α3 = b
4α1 + 3α2 + α3 = c
A short calculation shows that the coefficient matrix of the system is singular, from which we
could conclude that the system cannot have a solution for all a, b, c ∈ R. In other words, the
vectors cannot span R3 . It is however instructive to reach the same conclusion by a slightly
different route: using Gaussian elimination we see that the system is equivalent to the following
α1 + 2α2 + 4α3 = a
2a−b
α2 + 3α3 = 3
0 = 2a + 5b − 3c
4.4. LINEAR INDEPENDENCE 45
It follows that the system is consistent if and only if

2a + 5b − 3c = 0 .
Thus a vector (a, b, c)T in R3 belongs to the span of the vectors (1, 2, 4)T , (2, 1, 3)T , and
(4, −1, 1)T if and only if 2a + 5b − 3c = 0. In other words, not every vector in R3 can
be written as a linear combination of the vectors (1, 2, 4)T , (2, 1, 3)T , and (4, −1, 1)T , so in
particular these vectors cannot span R3 .
Example 4.23. Show that {p1 , p2 , p3 } is a spanning set for P2 , where
p1 (x) = 2 + 3x + x2 , p2 (x) = 4 − x , p3 (x) = −1 .
Solution. Let p be an arbitrary polynomial in P2 , say, p(x) = a + bx + cx2 . We need to show
that it is possible to find weights α1 , α2 and α3 such that
α1 p1 + α2 p2 + α3 p3 = p ,
that is
α1 (2 + 3x + x2 ) + α2 (4 − x) − α3 = a + bx + cx2 .
Comparing coefficients we find that the weights have to satisfy the system
2α1 + 4α2 − α3 = a
3α1 − α2 = b
α1 = c
The coefficient matrix is nonsingular, so the system must have a unique solution for all choices
of a, b, c. In fact, using back substitution yields α1 = c, α2 = 3c − b, α3 = 14c − 4b − a. Thus
{p1 , p2 , p3 } is a spanning set for P2 .
Example 4.24. Find a spanning set for N (A), where
 
−3 6 −1 1 −7
A =  1 −2 2 3 −1 .
2 −4 5 8 −4
Proof. We have already calculated N (A) for this matrix in Example 4.16, and found that
       

 2 1 −3


1 0 0

       

       
N (A) = α 0 + β −2 + γ  2  α, β, γ ∈ R .
     


 0 1  0  


 
0 0 1
 

Thus, (2, 1, 0, 0, 0)T , (1, 0, −2, 1, 0)T , (−3, 0, 2, 0, 1)T is a spanning set for N (A).
4.4 Linear independence

The notion of linear independence plays a fundamental role in the theory of vector spaces.
Roughly speaking, it is a certain minimality property a collection of vectors in a vector space
may or may not have. To motivate it consider the following vectors in R3 :
     
−3 2 −5
x1 =  1  , x2 = −1 , x3 =  1  . (4.3)
2 1 8
Let’s ask the question: what is Span (x1 , x2 , x3 )?

Notice that
x3 = 3x1 + 2x2 . (4.4)
Thus any linear combination of x1 , x2 , x3 can be written as a linear combination of x1 and x2
alone, because
α1 x1 + α2 x2 + α3 x3 = α1 x1 + α2 x2 + α3 (3x1 + 2x2 )
= (α1 + 3α2 )x1 + (α2 + 2α3 )x2 .
Hence
Span (x1 , x2 , x3 ) = Span (x1 , x2 ) .
Observing that equation (4.4) can be written as
3x1 + 2x2 − x3 = 0 , (4.5)
we see that any of the three vectors can be expressed as a linear combination of the other
two, so
Span (x1 , x2 , x3 ) = Span (x1 , x2 ) = Span (x1 , x3 ) = Span (x2 , x3 ) .
In other words, because of the dependence relation (4.5), the span of x1 , x2 , x3 can be written
as the span of only two of the given vectors. Or, put yet differently, we can throw away one of
the three vectors without changing their span. So the three vectors are not the most economic
way to express their span, because two of them suffice.
On the other hand, no dependency of the form (4.5) exists between x1 and x2 . Indeed,
suppose that there were scalars c1 and c2 , not both 0, such that
c1 x1 + c2 x2 = 0 , (4.6)
then we could solve for one of the vectors in terms of the other, that is,
c2 c1
x1 = − x2 (if c1 6= 0) or x2 = − x1 (if c2 6= 0) .
c1 c2
However, neither vector is a multiple of the other, so (4.6) can hold only if c1 = c2 = 0. In
particular, Span (x1 ) and Span (x2 ) are both proper subspaces of Span (x1 , x2 ), so we cannot
reduce the number of vectors further to express Span (x1 , x2 , x3 ) = Span (x1 , x2 ).
This discussion motivates the following definitions:
Definition 4.25. The vectors v1 , . . . , vn in a vector space V are said to be linearly depen-
dent if there exist scalars c1 , . . . , cn , not all zero, such that
c1 v1 + · · · + cn vn = 0 .
Example 4.26. The three vectors x1 , x2 , x3 defined in (4.3) are linearly dependent.
Definition 4.27. The vectors v1 , . . . , vn in a vector space V are said to be linearly inde-
pendent if they are not linearly dependent, that is, if
c1 v1 + · · · + cn vn = 0 ,
forces all scalars c1 , . . . , cn to be 0.

4.4. LINEAR INDEPENDENCE 47

2 1
Example 4.28. The vectors , ∈ R2 are linearly independent. In order to see this,
1 1
suppose that
2 1 0
c1 + c2 = .
1 1 0
Then c1 and c2 must satisfy the 2 × 2 system
2c1 + c2 = 0
c1 + c2 = 0
However, as is easily seen, the only solution of this system is c1 = c2 = 0. Thus, the two
vectors are indeed linearly independent as claimed.
Example 4.29. Let p1 , p2 ∈ P1 be given by
p1 (t) = 2 + t , p2 (t) = 1 + t .
Then p1 and p2 are linearly independent. In order to see this, suppose that
c1 p1 + c2 p2 = 0 .
Then, for all t

c1 (2 + t) + c2 (1 + t) = 0 ,
so, for all t
(2c1 + c2 ) + (c1 + c2 )t = 0 .
Notice that the polynomial on the left-hand side of the above equation will be the zero
polynomial if and only if its coefficients vanish, so c1 and c2 must satisfy the 2 × 2 system
2c1 + c2 = 0
c1 + c2 = 0
However, as in the previous example, the only solution of this system is c1 = c2 = 0. Thus p1
and p2 are indeed linearly independent as claimed.
Example 4.30 (Geometric interpretation of linear independence in R2 and R3 ). In case you

missed the hands-on demonstration with audience participation of linear independence given
in the lectures, make sure you hard-wire the following mental images of linear independence
in R2 and R3 to your brain for the rest of this module:
(a) If x and y are linearly dependent in R2 then
c1 x + c2 y = 0 ,
where c1 and c2 are not both 0. If, say c1 6= 0, then

c2
x=− y.
c1
Thus one of the vectors must be a scalar multiple of the other, or, put differently, the
two vectors must be collinear.
Conversely, if two vectors in R2 are not collinear, they are linearly independent.
(b) Just as in R2 , two vectors in R3 are linearly dependent if and only if they are collinear.
Suppose now that x and y are two linearly independent vectors in R3 . Since they are
not collinear, they will span a plane (through the origin). If z is another vector lying in
this plane, then 0 can be written as a linear combination of x and y, hence x, y, z are
linearly dependent. Conversely, if z does not lie in the plane spanned by x and y, then
x, y, z are linearly independent.
In other words, three vectors in R3 are linearly independent if and only if they are not
coplanar.
We finish this section with some more theoretical observations:
Theorem 4.31. Let x1 , . . . , xn be n vectors in Rn and let X ∈ Rn×n be the matrix whose
j-th column is xj . Then the vectors x1 , . . . , xn are linearly dependent if and only if X is
singular.
Proof. The equation

c1 x1 + · · · + cn xn = 0
can be written as  
c1
 .. 
Xc = 0 , where c =  .  .
cn
This system has a non-trivial solution c 6= 0 if and only X is singular.
Example 4.32. Determine whether the following three vectors in R3 are linearly independent:
     
−1 5 4
 3  , 2 , 5 .
1 5 6
Solution. Since

−1 5 4 −1 3 1 R1 + R2 4 5 6 R1 − R3 0 0 0

3 2 5 = 5 2 5 = 5 2 5 = 5 2 5 = 0 ,

1 5 6 4 5 6 4 5 6 4 5 6
the vectors are linearly dependent.
The following result will become important later in this chapter, when we discuss coordinate
systems.
Theorem 4.33. Let v1 , . . . , vn be vectors in a vector space V . A vector v ∈ Span (v1 , . . . , vn )

can be written uniquely as a linear combination of v1 , . . . , vn if and only if v1 , . . . , vn are lin-
early independent.
Proof. If v ∈ Span (v1 , . . . , vn ) then v can be written
v = α1 v1 + · · · + αn vn , (4.7)
for some scalars α1 , . . . , αn . Suppose that v can also be written in the form
v = β1 v1 + · · · + βn vn , (4.8)
4.5. BASIS AND DIMENSION 49
for some scalars β1 , . . . , βn . We start by showing that if v1 , . . . , vn are linearly independent,

then αi = βi for every i = 1, . . . , n (that is, the representation (4.7) is unique). To see this,
suppose that v1 , . . . , vn are linearly independent. Then subtracting (4.8) from (4.7) gives
(α1 − β1 )v1 + · · · + (αn − βn )vn = 0 , (4.9)
which forces αi = βi for every i = 1, . . . , n as desired.
Conversely, if the representation (4.7) is not unique, then there must be a representation
of the form (4.8) where αi 6= βi for some i between 1 and n. But then (4.9) means that
there exists a non-trivial linear dependence between v1 , . . . , vn , so these vectors are linearly
dependent.
4.5 Basis and dimension

The concept of a basis and the related notion of dimension are among the key ideas in the
theory vector of spaces, of immense practical and theoretical importance. Let’s start with the
definition of a basis, delaying the discussion of its interpretation for a bit:
Definition 4.34. A set {v1 , . . . , vn } of vectors forms a basis for a vector space V if
(i) v1 , . . . , vn are linearly independent;
(ii) Span (v1 , . . . , vn ) = V .
In other words, a basis for a vector space is a ‘minimal’ spanning set, in the sense that it
contains no superfluous vectors: every vector in V can be written as a linear combination of
the basis vectors (because of property (ii)), and there is no redundancy in the sense that no
basis vector can be expressed as a linear combination of the other basis vectors (by property
(i)). Let’s look at some examples:
Example 4.35. Let      
1 0 0
e1 = 0 ,
 e2 = 1 ,
 e3 = 0 .

0 0 1
Then {e1 , e2 , e3 } is a basis for R3 , called the standard basis.
Indeed, as is easily seen, every vector in R3 can be written as a linear combination of
e1 , e2 , e3 and, moreover, the vectors e1 , e2 , e3 are linearly independent.
Example 4.36.      
 1 1 1 
0 , 1 , 1
0 0 1
 
is a basis for R3 .
To see this note that the vectors are linearly independent, because

1 1 1

0
1 1 = 1 6= 0 .
0 0 1
Moreover, the vectors span R3 since, if (a, b, c)T is an arbitrary vector in R3 , then
       
a 1 1 1
 b  = (a − b) 0 + (b − c) 1 + c 1 .
c 0 0 1
The previous two examples show that a vector space may have more than one basis. This
is not a nuisance, but, quite to the contrary, a blessing, as we shall see later in this module. For
the moment, you should only note that both bases consist of exactly three elements. We will
revisit and expand this seemingly innocent observation shortly, when we discuss the dimension
of a vector space.
Example 4.37. Let

1 0 0 1 0 0 0 0
E11 = , E12 = , E21 = , E22 = .
0 0 0 0 1 0 0 1
Then {E11 , E12 , E21 , E22 } is a basis for R2×2 , because the four vectors span R2×2 (as was
shown in Coursework 5, Exercise 7(b)) and they are linearly independent. To see this, suppose
that
c1 E11 + c2 E12 + c3 E21 + c4 E22 = O2×2 .
Then
c1 c2 0 0
= ,
c3 c4 0 0
so c1 = c2 = c3 = c4 = 0.
Most of the vector spaces we have encountered so far have particularly simple bases, termed
‘standard bases’:
Example 4.38 (Standard bases for Rn , Rm×n and Pn ).
Rn : The n columns of In form the standard basis of Rn , usually denoted by {e1 , e2 , . . . , en }.
Rm×n : A canonical basis can be constructed as follows. For i = 1, . . . , m and j = 1, . . . , n
let Eij ∈ Rm×n be the matrix whose (i, j)-entry is 1, and all other entries are 0. Then
{ Eij | i = 1, . . . , m , j = 1, . . . , n } is the standard basis for Rm×n .
Pn : The standard basis is the collection {p0 , . . . , pn } of all monomials of degree less than
n, that is,
pk (t) = tk , for k = 0, . . . , n.
If this is not clear to you, you should check that it really is a basis!
Going back to Examples 4.35 and 4.36, recall the observation that both bases of R3
contained exactly three elements. This is not pure coincidence, but has a deeper reason. In
fact, as we shall see shortly, any basis of a vector space must contain the same number of
vectors. Before proving this important result we need the following:
Theorem 4.39. Let v1 , . . . , vn be vectors in a vector space V . If Span (v1 , . . . , vn ) = V ,
then any collection of m vectors in V where m > n is linearly dependent.
Proof. Let u1 , . . . , um be m vectors in V where m > n. Then, since v1 , . . . , vn span V , we
can write
ui = αi1 v1 + · · · + αin vn for i = 1, . . . , m .
Thus, a linear combination of the vectors u1 , . . . , um can be written as
X m
c1 u1 + · · · + cm um = ci ui
i=1
m n
!
X X
= ci αij vj
i=1 j=1
n m
!
X X
= αij ci vj .
j=1 i=1
4.5. BASIS AND DIMENSION 51
Now consider the system of n equations for the m unknowns c1 , . . . , cm

m
X
αij ci = 0 for j = 1, . . . , n.
i=1
This is a homogeneous system with more unknowns than equations, so by Theorem 1.22 it
must have a non-trivial solution (ĉ1 , . . . , ĉm )T . But then
n
X
ĉ1 u1 + · · · + ĉm um = 0vj = 0 ,
j=1
so u1 , . . . , um are linearly dependent.

We are now able to prove the observation alluded to earlier:
Corollary 4.40. If a vector space V has a basis of n vectors, then every basis of V must have
exactly n vectors.
Proof. Suppose that {v1 , . . . , vn } and {u1 , . . . , um } are both bases for V . We shall show
that m = n. In order to see this, notice that, since Span (v1 , . . . , vn ) = V and u1 , . . . , um are
linearly independent it follows by the previous theorem that m ≤ n. By the same reasoning,
since Span (u1 , . . . , um ) = V and v1 , . . . , vn are linearly independent, we must have n ≤ m.
So, all in all, we have n = m, that is, the two bases have the same number of elements.
In view of this corollary it now makes sense to talk about the number of elements of a
basis, and give it a special name:
Definition 4.41. Let V be a vector space. If V has a basis consisting of n vectors, we say
that V has dimension n, and write dim V = n.
The vector space {0} is said to have dimension 0. The vector space V is said to be finite
dimensional if there is a finite set of vectors spanning V ; otherwise it is said to be infinite
dimensional .
Example 4.42. By Example 4.38 the vector spaces Rn , Rm×n and Pn are finite dimensional
with dimensions
dim Rn = n , dim Rm×n = mn , dim Pn = n + 1 .
As an example of an infinite dimensional vector space, consider the set of all polynomials with
real coefficients, and call it P . In order to see that P is a vector space when equipped with the
usual addition and scalar multiplication, notice that P is a subset of the vector space C[−1, 1]
of continuous functions on [−1, 1] (in fact, it is a subset of C[a, b] for any a, b ∈ R with
a < b), which is closed under addition and scalar multiplication. Thus P is a vector space.
Note that any finite collection of monomials is linearly independent, so P must be infinite
dimensional. For the same reason, C[a, b] and C 1 [a, b] are infinite dimensional vector spaces.
While infinite dimensional vector spaces play an important role in many parts of contemporary
applied and pure mathematics, we shall be mainly concerned with finite dimensional vector
spaces for the rest of this module.
Example 4.43. We are now able to classify the subspaces of R3 discussed in Remark 4.10:
• 0-dimensional subspaces. Only the zero subspace {0}.

• 1-dimensional subspaces. Any subspace spanned by a nonzero vector, that is all lines
through the origin.
• 2-dimensional subspaces. Any subspace spanned by two linearly independent vectors,
that is all planes through the origin.
• 3-dimensional subspaces. Only R3 .
We close this section with the following result, which is often useful when trying to decide
whether a collection of vectors forms a basis of a vector space:
Theorem 4.44. If V is a vector space with dim V = n, then:
(a) any set consisting of n linearly independent vectors spans V ;
(b) any n vectors that span V are linearly independent.
Proof. (a) Let v1 , . . . , vn be a linearly independent. Pick v ∈ V . Since dim V = n, the n + 1
vectors v, v1 , . . . , vn must be linearly dependent by Theorem 4.39. Thus
c0 v + c1 v1 + · · · + cn vn = 0 , (4.10)
where c0 , c1 , . . . , cn are not all 0. But c0 6= 0 (for otherwise (4.10) would imply that the
vectors v1 , . . . , vn are linearly dependent), hence

c1 cn
v= − v1 + · · · + − vn ,
c0 c0
so v ∈ Span (v1 , . . . , vn ). But v was arbitrary, so Span (v1 , . . . , vn ) = V .
(b) Suppose Span (v1 , . . . , vn ) = V . In order to show that v1 , . . . , vn are linearly in-
dependent we argue by contradiction: suppose to the contrary that v1 , . . . , vn are linearly
dependent. Then one of the vi ’s, say vn can be written as a linear combination of the other
vectors. So v1 , . . . , vn−1 also span V . If v1 , . . . , vn−1 are still linearly dependent we can
eliminate another vector and still have a spanning set. We can continue this process until we
have found a spanning set with k linearly independent vectors where k < n. This, however,
contradicts the fact that dim V = n.
Remark 4.45. The above theorem provides a convenient tool to check whether a set of vectors
forms a basis. The theorem tells us that n linearly independent vectors in an n-dimensional
vector space are automatically spanning, so these vectors are a basis for the vector space. This
is often useful in situations where linear independence is easier to check than the spanning
property.
Remark 4.46. The above theorem also provides two perspectives on a basis of a vector space:
(
a spanning set that is as small as possible;
a basis is
a linearly independent collection of vectors that is as large as possible.
So, for example:

                 
 1 2   1 2 4   1 2 4 7 
0 , 3 , 0 , 3 , 5 , 0 , 3 , 5 , 8 .
0 0 0 0 6 0 0 6 9
     
linearly independent, basis for R3 spans R3 ,
but doesn’t span R3 but not linearly independent
4.6. COORDINATES 53
4.6 Coordinates
In this short section we shall discuss an important application of the notion of a basis. In
essence, a basis allows us to view a vector space of dimension n as if it were Rn . This is a
tremendously useful idea, with many practical and theoretical applications, many of which you
will see in the following chapters.
The basic idea is the following. Suppose that {b1 , . . . , bn } is a basis for a vector space
V . Since the basis vectors are spanning, given v ∈ V , there are scalars c1 , . . . , cn such that
v = c1 b1 + · · · + cn bn .
Moreover, since the basis vectors b1 , . . . , bn are linearly independent, the scalars c1 , . . . , cn
are uniquely determined by Theorem 4.33. Thus, the vector v in the vector space V , can
be uniquely represented as an n-vector (c1 , . . . , cn )T in Rn . This motivates the following
definition:
Definition 4.47. Suppose that B = {b1 , . . . , bn } is a basis for a vector space V . If v ∈ V

then the uniquely determined scalars c1 , . . . , cn such that
v = c1 b1 + · · · + cn bn ,
are called the coordinates of v relative to B . The n-vector (c1 , . . . , cn )T ∈ Rn is called the
B-coordinate vector of v, or the coordinate vector of v relative to B, and is denoted
by [v]B .
Example 4.48. Consider the basis B = {b1 , b2 } for R2 , where

1 1
b1 = , b2 = .
0 2
Suppose that x ∈ R2 has B-coordinate vector [x]B = (−2, 3)T . Find x.
Solution.
1 1 1
x = −2b1 + 3b2 = (−2) +3 = .
0 2 6

1
Example 4.49. The entries of x = are the coordinates of x relative to the standard
6
basis E = {e1 , e2 }, since

1 1 0
=1 +6 = 1e1 + 6e2 .
6 0 1
Thus, x = [x]E .
Remark 4.50. If you missed

the lecture where I gave a graphical demonstration of the repre-
1
sentation of the vector ∈ R2 with respect to the two bases B and E given in the previous
6
two examples, you should spend two minutes pondering these two representations. I’ll put two
pictures up at this spot, once I have figured out how to do this.
The following problems arise naturally in this context:

• Given a basis B for a vector space V and the B-coordinate vector [v]B for some v ∈ V ,
find v.5
• Given a basis B for a vector space V and a vector v ∈ V , find the B-coordinate vector
[v]B .
• Given two bases B and D for a vector space V and [v]D , find [v]B .
We shall now address these problems in the case where V is Rn , deferring the general discussion
to the next chapter, when we will have better machinery to deal with these problems.
Theorem 4.51. Let B = {b1 , . . . , bn } be a basis for Rn . There is an invertible n × n matrix
PB such that for any x ∈ Rn
x = PB [x]B .
In fact, the matrix PB is the matrix whose j-th column is bj .
Proof. Let [x]B = (c1 , . . . , cn )T . Then
x = c1 b1 + · · · + cn bn ,
so  
c1
 .. 
x = (b1 · · · bn )  .  .
| {z }
=PB cn
Moreover, by Theorem 4.31, the matrix PB is invertible since its columns are linearly indepen-
dent.
Since a vector x ∈ Rn is equal to its coordinate vector relative to the standard basis,
the matrix PB given in the theorem above is called the transition matrix from B to the
standard basis.
Corollary 4.52. Let B = {b1 , . . . , bn } be a basis for Rn . For any x ∈ Rn
[x]B = PB−1 x .

2 −1 4
Example 4.53. Let b1 = , b2 = and x = . Let B = {b1 , b2 } be the
1 1 5
corresponding basis for R2 . Find the B-coordinates of x.
Solution. By the previous corollary
[x]B = PB−1 x .
Now
2 −1
PB = ,
1 1
so
−1 1 1 1
PB = .
3 −1 2
Thus
1 1 1 4 3
[x]B = = .
3 −1 2 5 2
5
We have already done a particular case in Example 4.48.
4.7. ROW SPACE AND COLUMN SPACE 55
Theorem 4.54. Let B and D be two bases for Rn . If x is any vector in Rn , then
[x]B = PB−1 PD [x]D .
Proof. Clear, since PB is invertible and
PB [x]B = x = PD [x]D .
The n × n matrix PB−1 PD given in the theorem above is called the transition matrix from
D to B.
Example 4.55. Let B = {b1 , b2 } be the basis given in Example 4.53, let D = {d1 , d2 },
where
1 1
d1 = , d2 = ,
0 2
and let x ∈ R2 . If the D-coordinates of x are (−3, 2)T , what are the B-coordinates of x?
Solution.
1 1 1 1 1 −3 1
[x]B = PB−1 PD [x]D = = .
3 −1 2 0 2 2 3
4.7 Row space and column space

In this final section of this rather long chapter on vector spaces, we shall briefly discuss a number
of naturally arising vector spaces associated with matrices. We have already encountered one
such space, the nullspace of a matrix.
Definition 4.56. Let A ∈ Rm×n .
• The subspace of R1×n spanned by the row vectors of A is called the row space of A
and is denoted by row(A).
• The subspace of Rm×1 spanned by the column vectors of A is called the column space
of A and is denoted by col(A).

1 0 0
Example 4.57. Let A = .
0 1 0
• Since
α 1 0 0 +β 0 1 0 = α β 0
row(A) is a 2-dimensional subspace of R1×3 .
• Since
1 0 0 α
α +β +γ =
0 1 0 β
col(A) is a 2-dimensional subspace of R2×1 .
Notice that the row space and column space of a matrix are generally distinct objects.
Indeed, one is a subspace of R1×n the other a subspace of Rm×1 . However, in the example
above, both spaces have the same dimension (namely 2). We shall see shortly, that, rather
surprisingly, this is always the case. Before exploring this topic further we introduce the
following important concept:
Definition 4.58. The rank of a matrix, denoted by rank A, is the dimension of the row space.
How does one calculate the rank of a matrix? The next result provides the clue:
Theorem 4.59. Two row equivalent matrices have the same row space, so, in particular, have
the same rank.
Proof. Let A and B be two row equivalent matrices. Since B is row equivalent to A, the
matrix B can be obtained from A by a finite sequence of elementary row operations. Thus the
rows of B are a linear combination of the rows of A. Consequently, row(B) is a subspace of
row(A). Exchanging the roles of A and B it follows, using the same argument, that row(A)
is also a subspace of row(B), so row(A) = row(B).
Combining the previous theorem with the observation that the nonzero rows of a matrix
in row echelon form are linearly independent, we obtain the following recipe for calculating a
basis for the row space and the rank of a matrix:
In order to calculate a basis for the row space and the rank of a matrix A:
• bring matrix to row echelon form U ;
• the nonzero rows of U will form a basis for row(A);
• the number of nonzero rows of U equals rank A.

1 −3 2
A = 1 −2 1 .
2 −5 3
Then      
1 −3 2 1 −3 2 1 −3 2
1 −2 1 ∼ 0 1 −1 ∼ 0 1 −1 .
2 −5 3 0 1 −1 0 0 0
Thus
1 −3 2 , 0 1 −1
is a basis for row(A), and rank A = 2.
It turns out that the rank of a matrix A is intimately connected with the dimension of its
nullspace N (A). Before formulating this relation, we require some more terminology:
Definition 4.61. If A ∈ Rm×n , then dim N (A) is called the nullity of A, and is denoted by
nul A.
Example 4.62. Find the nullity of the matrix
 
−3 6 −1 1 −7
A =  1 −2 2 3 −1 .
2 −4 5 8 −4
4.7. ROW SPACE AND COLUMN SPACE 57
Solution. We have already calculated the nullspace N (A) of this matrix in Example 4.16 by
bringing A to row echelon form U and then using back substitution to solve U x = 0, giving
N (A) = { αx1 + βx2 + γx3 | α, β, γ ∈ R } ,
where      
2 1 −3
1 0 0
     
0 ,
x1 =  −2 ,
x2 =  2.
x3 = 
  
0 1 0
0 0 1
It is not difficult to see that x1 , x2 , x3 are linearly independent, so {x1 , x2 , x3 } is a basis for
N (A). Thus, nul A = 3.
Notice that in the above example the nullity of A is equal to the number of free variables
of the system Ax = 0. This is no coincidence, but true in general! If this is not obvious to
you, you should mull about this for a little while!
The connection between the rank and nullity of a matrix, alluded to above, is the content
of the following beautiful theorem with an ugly name:
Theorem 4.63 (Rank-Nullity Theorem). If A ∈ Rm×n , then
rank A + nul A = n .
Proof. Bring A to row echelon form U . Write r = rank A. Now observe that U has r non-zero
rows, hence U x = 0 has n − r free variables, so nul A = n − r.
We now return to the perhaps rather surprising connection between the dimensions of the
row space and the column space of a matrix.
Theorem 4.64. Let A ∈ Rm×n . Then
dim col(A) = dim row(A) .
Proof. If A has rank r, then the row echelon form U of A will have r leading 1’s. The columns
of U corresponding to the leading 1’s will be linearly independent. They do not, however, form
a basis of col(A), since in general A and U will have different column spaces. Let UL denote
the matrix obtained from U by deleting all the columns corresponding to the free variables.
Delete the same columns from A and denote the new matrix by AL . The matrices AL and UL
are row equivalent. Thus, if x is a solution to AL x = 0, then UL x = 0. Since the columns
of UL are linearly independent, x must equal 0. It follows that the columns of AL must be
linearly independent as well. Since AL has r columns, the dimension of row(A) must be at
least r, that is,
dim col(A) ≥ r = dim row(A) .
Applying the above inequality to AT yields
dim row(A) = dim col(AT ) ≥ dim row(AT ) = dim col(A) ,
so dim col(A) = dim row(A) as claimed.

The above proof shows how to find a basis for the column space of a matrix:
In order to find a basis for the column space of a matrix A:
• bring A to row echelon form and identify the leading variables;
• the columns of A containing the leading variables form a basis for col(A).

1 −1 3 2 1
A = 1 0 1 4 1 .
2 −1 4 7 4
Then the row echelon form of A is
 
1 −1 3 2 1
0 1 −2 2 0 .
0 0 0 1 2
The leading variables are in columns 1,2, and 4. Thus a basis for col(A) is given by
     
 1 −1 2 
1 ,  0  , 4 .
2 −1 7
 
Chapter 5
Linear Transformations

Linear transformations are the bread and butter of Linear Algebra. You have already encoun-
tered them in Geometry I. Roughly speaking a linear transformation is a mapping between two
vector spaces that preserves the linear structure of the underlying spaces. To be precise:
Definition 5.1. Let V and W be two vector spaces. A mapping L : V → W (that is, a
mapping from V to W ) is said to be a linear transformation or a linear mapping if it
satisfies the following two conditions:
(i) L(v + w) = L(v) + L(w) for all v and w in V ;
(ii) L(αv) = αL(v) for all v in V and all scalars α.

Example 5.2. Let L : R2 → R2 be defined by
L(x) = 2x .
Then L is linear since, if x and y are arbitrary vectors in R2 and α is an arbitrary real number,
then
(i) L(x + y) = 2(x + y) = 2x + 2y = L(x) + L(y);
(ii) L(αx) = 2(αx) = α(2x) = αL(x).


x1
L(x) = x1 e1 , where x= .
x2
Then L is linear. In order to see this suppose that x and y are arbitrary vectors in R2 with

x1 y
x= , y= 1 .
x2 y2
Notice that, if α is an arbitrary real number, then

x1 + y1 αx1
x+y = and αx = .
x2 + y2 αx2
Thus
59
60 CHAPTER 5. LINEAR TRANSFORMATIONS
(i) L(x + y) = (x1 + y1 )e1 = x1 e1 + y1 e1 = L(x) + L(y);
(ii) L(αx) = (αx1 )e1 = α(x1 e1 ) = αL(x).
Hence L is linear, as claimed.
In order to shorten statements of theorems and examples let us introduce the following
convention:
If x is a vector in Rn , we shall henceforth denote its i-th entry by xi , and similarly for vectors
in Rn denoted by other bold symbols. So, for example, if y = (1, 4, 2, 7)T ∈ R4 , then y3 = 2.
Example 5.4. Let L : R2 → R2 be given by

−x2
L(x) = .
x1
L is linear, since, if x, y ∈ R2 and α ∈ R, then

−(x2 + y2 ) −x2 −y2
(i) L(x + y) = = + = L(x) + L(y);
x1 + y2 x1 y1

−αx2 −x2
(ii) L(αx) = =α = αL(x).
αx1 x1
Example 5.5. The mapping M : R2 → R1 defined by

q
M (x) = x21 + x22
√
order to see this note that M ((1, 0)T ) =
is not linear. Inp 12 = 1 while M (−(1, 0)T ) =
M ((−1, 0)T ) = (−1)2 = 1. Thus
M (−(1, 0)T ) = 1 6= −1 = −M ((1, 0)T ) .
Important Observation. Any m×n matrix A induces a linear transformation LA : Rn → Rm

given by
LA (x) = Ax for each x ∈ Rn .
The transformation LA is linear, since, if x, y ∈ Rn and α ∈ R, then
(i) LA (x + y) = A(x + y) = Ax + Ay = LA (x) + LA (y);
(ii) LA (αx) = A(αx) = αAx = αLA (x).
In other words, every m × n matrix gives rise to a linear transformation from Rn to Rm . We

shall see shortly that, conversely, every linear transformation from Rn to Rm arises from an
m × n matrix.
5.2. LINEAR TRANSFORMATIONS ON GENERAL VECTOR SPACES 61
5.2 Linear transformations on general vector spaces

So far we have only considered linear transformations from Rn to Rm . In this short section,
we shall have a look at linear transformations on abstract vector spaces. We start with some
general observations concerning linear transformations on abstract vector spaces and finish
with two examples.
Theorem 5.6. If V and W are vector spaces and L : V → W is a linear transformation then:
(a) L(0) = 0;
(b) L(−v) = −L(v) for any v ∈ V ;
(c) L( ni=1 αi vi ) = ni=1 αi L(vi ) for any vi ∈ V and any scalars αi where i = 1, . . . , n.
P P
Proof.
(a) L(0) = L(00) = 0L(0) = 0;
(b) L(−v) = L((−1)v) = (−1)L(v) = −L(v);
(c) follows by repeated application of the defining properties (i) and (ii) of linear transfor-
mations.
Note that we have used Theorem 4.5 for the proof of (a) and (b).
Let’s look at some examples, which should convince you that linear transformations arise
naturally in other areas of Mathematics.
Example 5.7. Let L : C[a, b] → R1 be defined by
Z b
L(f ) = f (t) dt .
a
L is linear since, if f , g ∈ C[a, b] and α ∈ R, then

Z b Z b Z b
(i) L(f + g) = (f (t) + g(t)) dt = f (t) dt + g(t) dt = L(f ) + L(g);
a a a
Z b Z b
(ii) L(αf ) = (αf (t)) dt = α f (t) dt = αL(f ).
a a
In other words, integration is a linear transformation.

Example 5.8. Let D : C 1 [a, b] → C[a, b] be defined to be the transformation that sends an
f ∈ C 1 [a, b] to its derivative f 0 ∈ C[a, b], that is,
D(f ) = f 0 .
Then D is linear since, if f , g ∈ C 1 [a, b] and α ∈ R, then

(i) D(f + g) = (f + g)0 = f 0 + g0 = D(f ) + D(g);
(ii) D(αf ) = (αf )0 = αf 0 = αD(f ).

In other words, differentiation is a linear transformation.
Example 5.9. Let V be a vector space and let Id : V → V denote the identity transfor-
mation (or identity for short) on V , that is,
Id(v) = v for all v ∈ V .
The transformation Id is linear, since, if v, w ∈ V and α is a scalar, then
(i) Id(v + w) = v + w = Id(v) + Id(w);
(ii) Id(αv) = αv = αId(v).
5.3 Image and Kernel

We shall now discuss two useful notions, namely that of the ‘image’ and that of the ‘kernel’ of
a linear transformation, that help to generalise two notions that you have already encountered
in connection with matrices.
Definition 5.10. Let V and W be vector spaces, and let L : V → W be a linear transfor-
mation. The kernel of L, denoted by ker(L), is the subset of V given by
ker(L) = { v ∈ V | L(v) = 0 } .
Example 5.11. If A ∈ Rm×n and LA is the corresponding linear transformation from Rn to

Rm , then
ker(LA ) = N (A) ,
that is, the kernel of LA is the nullspace of A.
The previous example shows that the kernel of a linear transformation is the natural gen-
eralisation of the nullspace of a matrix.
Definition 5.12. Let V and W be vector spaces. Let L : V → W be a linear transformation

and let H be a subspace of V . The image of H (under L), denoted by L(H), is the subset
of W given by
L(H) = { w ∈ W | w = L(v) for some v ∈ H } .
The image L(V ) of the entire vector space V under L is called the range of L.
Example 5.13. If A ∈ Rm×n and LA is the corresponding linear transformation from Rn to

Rm , then
LA (Rn ) = col(A) ,
that is, the range of LA is the column space of A.
The previous example shows that the range of a linear transformation is the natural gen-
eralisation of the column space of a matrix.
We saw previously that the nullspace and the column space of an m × n matrix are
subspaces of Rn and Rm respectively. The same is true for the abstract analogues introduced
above.
Theorem 5.14. Let V and W and be vector spaces. If L : V → W is a linear transformation

and H is a subspace of V , then
(a) ker(L) is a subspace of V ;

5.3. IMAGE AND KERNEL 63
(b) L(H) is a subspace of W .

Proof.
(a) First observe that ker(L) is not empty since 0 ∈ ker(L) by Theorem 5.6. Suppose now
that v1 , v2 ∈ ker(L). Then
L(v1 + v2 ) = L(v1 ) + L(v2 ) = 0 + 0 = 0 ,
so v1 + v2 ∈ ker(L). Moreover, if v ∈ ker(L) and α is a scalar, then
L(αv) = αL(v) = α0 = 0 ,
so αv ∈ ker(L). Thus, as ker(L) is closed under addition and scalar multiplication, it

is a subspace of V as claimed.
(b) First observe that L(H) is not empty since 0 ∈ L(H) by Theorem 5.6. Suppose now that
w1 , w2 ∈ L(H). Then there are v1 , v2 ∈ H such that L(v1 ) = w1 and L(v2 ) = w2
and so
w1 + w2 = L(v1 ) + L(v2 ) = L(v1 + v2 ) .
But v1 +v2 ∈ H, because H is a subspace, so w1 +w2 ∈ L(H). Moreover, if w ∈ L(H)
and α is a scalar, then there is v ∈ H such that L(v) = w and so
αw = αL(v) = L(αv) .
But αv ∈ H, because H is a subspace, so αw ∈ L(H). Thus, as L(H) is closed under

addition and scalar multiplication, it is a subspace of W as claimed.
Example 5.15. Let D : P3 → P3 be the differentiation transformation given by
D(p) = p0 .
Find ker(D) and D(P3 ).

Solution. The derivative of a polynomial p ∈ P3 is the zero polynomial if and only if p is a
constant. Thus
ker(D) = P0 .
Since differentiation lowers the degree of a polynomial by 1, we see that D(P3 ) is a subspace
of P2 . However, any polynomial in P2 has an antiderivative in P3 , so every polynomial in P2
will be the image of a polynomial in P3 under D. Thus
D(P3 ) = P2 .
Example 5.16 (Optional; for those of you who have taken MTH4102 ‘Differential Equations’).
Let L : C 1 [−1, 1] → C[−1, 1] be the linear transformation given by
L(f ) = f + f 0 .
Find the kernel and range of L.

Solution. In order to determine the kernel of L we need to find all f ∈ C 1 [−1, 1] such that
f + f0 = 0 . (5.1)
This is a first order homogeneous differential equation with integrating factor et . Thus, a
function f satisfies (5.1) if and only if
d t
(e f (t)) = 0 ,
dt
so
et f (t) = α ,
for some α ∈ R, and hence
f (t) = αe−t .
Thus
ker(L) = Span (h) ,
where h(t) = e−t .
The moral of this part of the example is that whenever you are solving a linear homogeneous
differential equation, you are in fact calculating the kernel of a certain linear transformation!
In order to determine the range of L, notice that L clearly sends continuously differentiable
functions to continuous functions. The question is whether every continuous function arises
as an image of some f ∈ C 1 [−1, 1] under L. The answer is yes! To see this, fix g ∈ C[−1, 1].
We need to show that there is an f ∈ C 1 [−1, 1] such that L(f ) = g, that is,
f0 + f = g . (5.2)
This is a first order linear inhomogeneous differential equation for f . Using the integrating
factor et we find that (5.2) is equivalent to
d t
(e f (t)) = et g(t) .
dt
But the right hand side of the equation above has an antiderivative, say H, that is,
d
H(t) = et g(t)
dt
so
et f (t) = H(t) + α ,
for some α ∈ R, hence
f (t) = e−t H(t) + αe−t .
Notice that the f just found is clearly continuously differentiable. To summarise, we have
shown that, given g ∈ C[−1, 1], we can find f ∈ C 1 [−1, 1] such that L(f ) = g. In other
words, the range of L is C[−1, 1].
The moral of this part of the example is the following: the statement that the range of
L is C[−1, 1] means that the differential equation (5.2) has a solution for any g ∈ C[−1, 1].
So statements about ranges of certain linear transformations are in fact statements about the
existence of solutions for certain classes of differential equations!
5.4. MATRIX REPRESENTATIONS OF LINEAR TRANSFORMATIONS 65
5.4 Matrix representations of linear transformations

This subsection contains the central result of this chapter, if not of the whole module. In
essence, this result says that every linear transformation between finite dimensional vector
spaces can be viewed as a matrix. Or in other words, matrices provide concrete realisations of
abstract linear transformations. This is one of the main reasons why matrices underpin large
parts of contemporary Mathematics.
Let’s start with a baby-version of this miraculous result. To appreciate it, recall the impor-
tant observation that any matrix A ∈ Rm×n gives rise to a linear transformation LA : Rn → Rm
by defining LA (x) = Ax. The following theorem says that every linear transformation from
Rn to Rm arises in this way:
Theorem 5.17. If L is a linear transformation from Rn to Rm , then there is an m × n matrix
A such that
L(x) = Ax for each x ∈ Rn .
In fact, the j-th column aj of A is given by
aj = L(ej ) for j = 1, . . . , n ,
where {e1 , . . . , en } denotes the standard basis of Rn .

Proof. For j = 1, . . . , n let aj = L(ej ) and let A ∈ Rm×n be the matrix whose j-th column
is aj . Now, if x ∈ Rn , then
x = x 1 e1 + · · · + x n en ,
so
L(x) = L(x1 e1 + · · · + xn en ) = x1 L(e1 ) + · · · + xn L(en ) = x1 a1 + · · · + xn an = Ax .
To paraphrase the theorem above yet again: given a linear transformation L between Rn
and Rm we can find a matrix that represents it. For this reason the matrix A constructed
above is called a matrix representation of the linear transformation L. For the construction
of A we have used the standard basis. We shall see shortly that there is nothing special
about the standard basis, that is, we could have chosen any other basis to represent a given
transformation. This will be the content of the ‘Matrix Representation Theorem’ we will
encounter shortly.
Example 5.18. Let L : R3 → R2 be given by

x1 − x2
L(x) = .
x2 + 2x3
The transformation L is easily seen to be linear. Now

1 − 0 1
L(e1 ) = =
0 + 2·0 0

0 − 1 −1
L(e2 ) = =
1 + 2·0 1

0 − 0 0
L(e3 ) = =
0 + 2·1 2
so if we set
1 −1 0
A= ,
0 1 2
then indeed  
x1
1 −1 0   x1 − x2
Ax = x2 = = L(x) .
0 1 2 x2 + 2x3
x3
We are now ready for the main result alluded to at the beginning of this section. This truly
marvellous result is an extension of Theorem 5.17. It states that any linear transformation
between arbitrary finite dimensional vector spaces can be represented by a matrix.
In order to appreciate this statement, let’s suppose that V is an n-dimensional vector space
with basis V and that W is an m-dimensional vector space with basis W. If L : V → W is a
linear transformation we can ask the following question: given an arbitrary vector v ∈ V with
V-coordinates [v]V , what are the W-coordinates of L(v)? It turns out that there is an m × n
matrix A, not depending on v, such that
[L(v)]W = A[v]V .
In other words, the matrix A represents the change of coordinates effected by the linear
transformation L. This important fact is the content of the following theorem:
Theorem 5.19 (Matrix Representation Theorem). If V = {v1 , . . . , vn } and W = {w1 , . . . , wm }
are bases for vector spaces V and W respectively, then corresponding to each linear transfor-
mation L : V → W there is an m × n matrix A such that
[L(v)]W = A[v]V for each v ∈ V .
In fact, the j-th column aj of A is given by
aj = [L(vj )]W .
Proof. For each j = 1, . . . n let aj = (a1j , . . . , amj )T be the W-coordinate vector of L(vj ).
Thus m
X
L(vj ) = a1j w1 + · · · + amj wm = aij wi .
i=1
Fix v ∈ V . Let x be the V-coordinate vector of v, that is,
n
X
v = x1 v1 + · · · + xn vn = xj vj .
j=1
Then
Xn
L(v) = L( xj vj )
j=1
n
X
= xj L(vj )
j=1
n m
!
X X
= xj aij wi
j=1 i=1
m n
!
X X
= aij xj wi .
i=1 j=1
5.4. MATRIX REPRESENTATIONS OF LINEAR TRANSFORMATIONS 67
If we now call
n
X
yi = aij xj for i = 1, . . . , m ,
j=1
then the above calculation shows that

 
y1
y =  ...  = Ax
 
ym
is the W-coordinate vector of L(v), that is
[L(v)]W = A[v]V .
Definition 5.20. Given vector spaces V and W with corresponding bases V and W, and a
linear transformation L : V → W , we call the matrix A constructed in the theorem
V above the
matrix representation of L with respect to V and W, and denote it by L W . Thus, for
any v ∈ V we have
V
[L(v)]W = L W [v]V .
L(x) = x1 b1 + (x2 + x3 )b2 ,
where
1 1
b1 = , b2 = .
0 1
Find the matrix representation of L with respect to the standard basis E = {e1 , e2 , e3 } and
the basis B = {b1 , b2 }.
Solution. Since
L(e1 ) = 1b1 + 0b2
L(e2 ) = 0b1 + 1b2
L(e3 ) = 0b1 + 1b2
we see that
1 0 0
[L(e1 )]B = , [L(e2 )]B = , [L(e3 )]B = ,
0 1 1
so
E 1 0 0
L B = .
0 1 1
Example 5.22. Consider the linear transformation D : P2 → P1 given by
(D(p))(t) = p0 (t) .
Define
p1 (t) = 1, p2 (t) = t, p3 (t) = t2 ,
and let P2 = {p1 , p2 , p3 } and P1 = {p1 , p2 } be bases for P2 and P1 respectively. Since
(D(p1 ))(t) = p01 (t) = 0

(D(p2 ))(t) = p02 (t) = 1
(D(p3 ))(t) = p03 (t) = 2t
we have
D(p1 ) = 0p1 + 0p2
D(p2 ) = 1p1 + 0p2
D(p3 ) = 0p1 + 2p2
so
P2 0 1 0
D P1 = .
0 0 2
Suppose now that p ∈ P2 is given by
p(t) = a + bt + ct2 .
We want to find D(p). Of course we could do this working directly from the definition of D,
but we can also use the Matrix Representation Theorem: since
p = ap1 + bp2 + cp3 ,
we have, by the Matrix Representation Theorem,

 
a
P 0 1 0   b
[D(p)]P1 = D P21 [p]P2 = b = ,
0 0 2 2c
c
so
D(p) = bp1 + 2cp2 ,
that is,
p0 (t) = b + 2ct ,
as expected.
Example 5.23. Let A ∈ Rm×n , and let LA be the corresponding linear transformation from
Rn to Rm . Since LA (ej ) is just the j-th column of A, we see that the matrix representation
of LA with respect to the standard bases of Rn and Rm is just A itself.
5.5 Composition of linear transformations

In this short section we shall have look at composition of linear transformations and its effect
on the corresponding matrix representations. Suppose that U , V and W are vector spaces
and that we are given two linear transformations
T :U →V ,
S:V →W.
We can then form a new transformation S ◦ T : U → W by defining
(S ◦ T )(u) = S(T (u)) for each u ∈ U .

5.5. COMPOSITION OF LINEAR TRANSFORMATIONS 69
The transformation S ◦ T is called the composite of S and T . Observe that S ◦ T is linear

as well. In order to see this, let u1 , u2 ∈ U and α1 , α2 be scalars. Then
(S ◦ T )(α1 u1 + α2 u2 ) = S(T (α1 u1 + α2 u2 ))

= S(α1 T (u1 ) + α2 T (u2 ))
= α1 S(T (u1 )) + α2 S(T (u2 ))
= α1 (S ◦ T )(u1 ) + α2 (S ◦ T )(u2 ) .
Choosing α1 = α2 = 1 in the above equality gives
(S ◦ T )(u1 + u2 ) = (S ◦ T )(u1 ) + (S ◦ T )(u2 ) ,
while choosing α1 = 1 and α2 = 0 gives
(S ◦ T )(α1 u1 ) = α1 (S ◦ T )(u1 ) ,
so S ◦ T is linear, as claimed.
Example 5.24. Suppose that A ∈ Rm×n and B ∈ Rn×r . Let LA : Rn → Rm and LB :
Rr → Rn be the corresponding linear transformations. Then LA ◦ LB : Rr → Rm is the linear
transformation given by
(LA ◦ LB )(x) = LA (LB (x)) = LA (Bx) = ABx ,
so
LA ◦ LB = LAB .
In other words, the composite of LA and LB is the linear transformation arising from the
product AB. Ultimately, this is the deeper reason why matrix multiplication is defined as it
is.
We shall shortly discuss a more abstract analogue of the above example for two arbitrary
linear transformations between finite dimensional vector spaces. Before doing so we digress to
introduce some more notation:
Notation 5.25. If V is a vector space and L : V → V is linear, we can compose L with itself
and call the resulting linear transformation L2 , that is,
L2 = L ◦ L .
More generally, we write

Ln = L
| ◦ ·{z
· · ◦ L}
n compositions
for the n-fold composite of L with itself.

Example 5.26. Let D : P5 → P5 be the linear transformation given by
D(p) = p0 .
Then
D2 (p) = p00
D3 (p) = p000
and so forth.
Suppose now that U , V and W are finite dimensional vector spaces with corresponding
bases U, V and W. If we are given two linear transformations T : U → V and S : V → W
we may want to ask the question, how the matrix representations of S and T are related to
that of S ◦ T . The following theorem provides a neat answer:
Theorem 5.27 (Composition Formula). Let U , V and W be finite dimensional vector spaces
with corresponding bases U, V and W. Suppose that T : U → V and S : V → W are linear
transformations. Then U V U
S◦T W = S W T V ,
that is, the matrix representation of S ◦ T with respect to U and W is the matrix product of
the matrix representation of S with respect V and W and the matrix representation of T with
respect to U and V.
V U
Proof. Write A = S W and B = T V . Let u ∈ U and let v = T (u) ∈ V . Thus, by the
Matrix Representation Theorem,
[S(v)]W = A[v]V ,
[T (u)]V = B[u]U .
Hence
[(S ◦ T )(u)]W = [S(T (u))]W = [S(v)]W = A[v]V = A[T (u)]V = AB[u]U .
Since u ∈ U was arbitrary, the above equation means that AB must be the matrix represen-
tation of S ◦ T with respect to U and W, that is,
U
AB = S ◦ T W .
In essence, the Composition Formula states that composition of linear transformations

corresponds to matrix multiplication of the corresponding matrix representations. We shall
see some applications of this formula in the next section. For the moment we shall be content
with the following simple consequence:
Corollary 5.28. Let V be a finite dimensional vector space with basis V and let L : V → V .
If A is the matrix representation of L with respect to V, then An is the matrix representation
of Ln with respect to V, that is
n V
L V = An .
5.6 Change of basis and similarity

In this last section of the chapter we shall consider the problem of how the matrix representation
of a given linear transformation changes when the bases of the underlying vector spaces are
changed. The solution to this problem will come in handy in the very last chapter of these
notes.
Before turning our attention to this problem we shall first consider the slightly simpler one
of how to describe coordinate changes in abstract vector spaces, thus extending the discussion
of coordinate changes in Rn detailed in Chapter 4.
5.6. CHANGE OF BASIS AND SIMILARITY 71
Question. Let V be a finite dimensional vector space with bases U and W.

Given v ∈ V and its U-coordinate vector [v]U how do we find the W-coordinate vector of
v?
The answer is the content of the following theorem.
Theorem 5.29. Let V be an n-dimensional vector space with bases U = {u1 , . . . , un } and
W = {w1 , . . . , wn }. Then there is an n × n matrix S such that
[v]W = S[v]U for any v ∈ V .
In fact, S is the matrix representation of the identity Id : V → V with respect to U and W.

In particular, the j-th column of S is given by [uj ]W , where j = 1, . . . , n.
Proof. Let S be the matrix representation of the identity Id : V → V with respect to U and
W. Then, by the Matrix Representation Theorem
U
[v]W = [Id(v)]W = Id W [v]U = S[v]W .
The n × n matrix S constructed in the theorem above is called the transition matrix
from U to W. It allows you to calculate the W-coordinates of a vector, given that you know
its U-coordinates. It shouldn’t come as a surprise that a transition matrix is invertible:
Theorem 5.30. Let V be a finite dimensional vector space with bases U and W. Then the
transition matrix from U to W is invertible, and its inverse is the transition matrix from W
to U.
Proof. By the Composition Formula we have
U W W W
Id W Id U = Id ◦ Id W = Id W = I ,
W U U U
Id U Id W = Id ◦ Id U = Id U = I ,
U W
so Id W is invertible with inverse Id U .
Example 5.31. Let P = {p1 , p2 , p3 } and Q = {q1 , q2 , q3 } be the bases of P2 given by
p1 (t) = 1, p2 (t) = t, p3 (t) = t2 ;

q1 (t) = 1 + t2 , q2 (t) = t, q3 (t) = 2 + 3t2 .
Now
q1 = 1p1 + 0p2 + 1p3
q2 = 0p1 + 1p2 + 0p3
q3 = 2p1 + 0p2 + 3p3
so      
1 0 2
[q1 ]P = 0 ,
 [q2 ]P = 1 ,
 [q3 ]P = 0 ,

1 0 3
hence the transition matrix from Q to P is
 
Q 1 0 2
Id P = 0 1 0 .
1 0 3
Thus, the transition matrix from P to Q is

 
P −1 3 0 −2
Q
Id Q = Id P = 0 1 0 .
−1 0 1
Now if p is given by
p(t) = 2 − 3t + t2 ,
then [p]P = (2, −3, 1)T , so
    
P 3 0 −2 2 4
[p]Q = Id Q [p]P =  0 1 0   −3 = −3 .
 
−1 0 1 1 −1
Thus
p = 4q1 − 3q2 − q3 .
We now turn to the problem mentioned at the beginning of this section.
Question. Let V be a finite dimensional vector space with bases U and W. Suppose that
L : V → V is a linear transformation. U W
How are the two matrix representations L U and L W related?
The answer is contained in the following theorem:
Theorem 5.32. Let V be a finite dimensional vector space with bases U and W, and let
L : V → V be a linear transformation. If S is the transition matrix from U to W, then
U W
L U = S −1 L W S .
U W
Proof. Note that S = Id W and S −1 = Id U . Thus, by the composition formula,
W W W U W U U U
S −1 L W S = Id U L W Id W = Id U L ◦ Id W = Id ◦ L ◦ Id U = L U .

Example 5.33. Let L : R3 → R3 be the linear transformation defined by L(x) = Ax, where
 
2 2 0
A = 1 1 2 .
1 1 2
(a) Find the matrix representation of L with respect to the basis B = {b1 , b2 , b3 }, where
     
1 −2 1
b1 = −1 , b2 =
   1 , b3 = 1 .
 
0 1 1
(b) Find Ln (x), where x = (−1, 2, 2)T .

5.6. CHANGE OF BASIS AND SIMILARITY 73
Solution.
(a) First observe that the matrix representation of L with respect to the standard basis E
of R3 is A itself. Now, the transition matrix S from B to E is simply
 
B 1 −2 1
S = Id E = PB = −1 1 1 ,
0 1 1
so the transition matrix from E to B is
 
0 −3 3
1
S −1 = −1 −1 2 .
3
1 1 1
Thus  
B 0 0 0
L B = S −1 AS = 0 1 0 .
0 0 4
(b) Notice that it would be extremely silly to try and work out An in order to obtain
Ln (x) = An x. What you should observe instead is that the matrix representation of L
with respect to B, call it B, is extremely simple: it is diagonal. Thus
 
n B 0 0 0
L B = B n = 0 1 0  .
0 0 4n
Hence, in order to work out Ln (x) we only need to find the coordinates of x with respect
to B and we are done. Now
    
0 −3 3 −1 0
E −1 1
[x]B = Id B [x]E = S x = −1 −1 2   2  = 1 ,

3
1 1 1 2 1
so    
0 0
n n B n 

[L (x)]B = L B [x]B = B 1 = 1 ,

1 4n
and thus  
−2 + 4n
Ln (x) = b2 + 4n b3 =  1 + 4n  .
1 + 4n
Let’s have another look at the previous theorem: is states that if V is an n-dimensional
vector space with two bases U and W and L is a linear transformation from V to itself, then
the matrix representation of L with respect to W, call it A, and the matrix representation of
L with respect to U, call it B, satisfy
B = S −1 AS ,
where S is an invertible n × n matrix. This type of relation is of sufficient importance to merit

a name of its own:
Definition 5.34. Let A and B be two n × n matrices. The matrix B is said to be similar
to A if there is an invertible S ∈ Rn×n such that
B = S −1 AS .
Notice that if B is similar to A, then A is similar to B, because if R = S −1 , then
A = SBS −1 = R−1 BR .
Thus we may simply say that A and B are similar matrices.

The content of Theorem 5.32 can now be rephrased as follows: if A and B are two matrix
representations of the same linear transformation on a vector space, then A and B are similar.
Chapter 6
Orthogonality
In this chapter we will return to the concrete vector space Rn and add a new concept that will
reveal new aspects of it. The added spice in the discussion is the notion of ‘orthogonality’.
This concept extends our intuitive notion of perpendicularity in R2 and R3 to Rn . Innocent
as it may seem, this new concept turns out to be a rather powerful device, as we shall see
shortly.
6.1 Definition
We start by revisiting a concept that you have already encountered in Geometry I. Before
stating it recall that a vector x in Rn is, by definition, an n × 1 matrix. Given another vector
y in Rn , we may then form the matrix product xT y of the 1 × n matrix xT and the n × 1
matrix y. Notice that by the rules of matrix multiplication xT y is a 1 × 1 matrix, which we
can simply think of as a real number.
Definition 6.1. Let x and y be two vectors in Rn . The scalar xT y is called the scalar
product or dot product of x and y, and is often written x·y. Thus, if
   
x1 y1
 x2   y2 
x =  ..  , y =  ..  ,
   
. .
xn yn
then
 
y1
 y2 

x·y = xT y = x1 x2 · · · xn  ..  = x1 y1 + x2 y2 + · · · + xn yn .

.
yn
Example 6.2. If
   
2 4
x = −3 and y = 5 ,
1 6
75
76 CHAPTER 6. ORTHOGONALITY
then
 
T
4
x·y = x y = 2 −3 1 5 = 2 · 4 + (−3) · 5 + 1 · 6 = 8 − 15 + 6 = −1 ,
6
 
T
2
y·x = y x = 4 5 6 −3 = 4 · 2 + 5 · (−3) + 6 · 1 = 8 − 15 + 6 = −1 .
1
Having had a second look at the example above it should be clear why x·y = y·x. In fact,
this is true in general. The following further properties of the dot product follow easily from
properties of the transpose operation:
Theorem 6.3. Let x, y and z be vectors in Rn , and let α be a scalar. Then
(a) x·y = y·x;
(b) (x + y)·z = x·z + y·z;
(c) (αx)·y = α(x·y) = x·(αy);
(d) x·x ≥ 0, and x·x = 0 if and only if x = 0.
Definition 6.4. If x = (x1 , . . . , xn )T ∈ Rn , the length or norm of x is the nonnegative

scalar kxk defined by
√ q
kxk = x·x = x21 + · · · + x2n .
A vector whose length is 1 is called a unit vector .
Example 6.5. If x = (a, b)T ∈ R2 , then

√
kxk = a2 + b 2 .
The above example should convince you that in R2 and R3 the definition of the length of
a vector x coincides with the standard notion of the length of the line segment from the origin
to x.
Note that if x ∈ Rn and α ∈ R then
kαxk = |α| kxk ,
because kαxk2 = (αx)·(αx) = α2 (x·x) = α2 kxk2 . Thus, if x 6= 0, we can always find a unit
vector y in the same direction as x by setting
1
y= x.
kxk
The process of creating a unit vector y from x is called normalising x.
Definition 6.6. For x and y in Rn , the distance between x and y, written dist(x, y), is
the length of x − y, that is,
dist(x, y) = kx − yk .
6.2. ORTHOGONAL COMPLEMENTS 77
Again, you might want to mull for a second about this definition and convince yourself
that in R2 and R3 the above definition of the distance between two vectors x and y coincides
with the standard notion of the Euclidean distance between the two points represented by x
and y.
We now come to the main part of this introductory section. In essence, we will extend
the concept of perpendicularity familiar in R2 and R3 to vectors in Rn . In order to motivate
it consider two lines through the origin in R2 or R3 determined by two vectors x and y. By
drawing a picture (which I will try to supply before long) of the triangle formed by the points
representing x, y and −y, convince yourself that the lines in the direction of x and y are
geometrically perpendicular if and only if the distance from x to y equals the distance from x
to −y. Now, using Theorem 6.3 repeatedly,
dist(x, −y)2 = kx + yk2
= (x + y)·(x + y)
= x·(x + y) + y·(x + y)
= x·x + x·y + y·x + y·y
= kxk2 + kyk2 + 2x·y .
Using the same calculation with −y in place of y we see that
dist(x, y)2 = kxk2 + k − yk2 + 2x·(−y)
= kxk2 + kyk2 − 2x·y .
Thus
dist(x, −y) = dist(x, y) if and only if x · y = 0.
In other words, the lines in the direction of x and y are geometrically perpendicular if and only
if x·y = 0. We now take this observation as our starting point for the following definition,
which extends the concept of perpendicularity to Rn .
Definition 6.7. Two vectors x and y in Rn are orthogonal (to each other) if x·y = 0.
Note that the zero vector is orthogonal to every other vector in Rn .
The following theorem is an old acquaintance in new clothing, and, at the same time,
contains a key fact about orthogonal vectors:
Theorem 6.8 (Pythagorean Theorem). Two vectors x and y in Rn are orthogonal if and only
if
kx + yk2 = kxk2 + kyk2 .
Proof. See Exercise 3 on Coursework 9.
6.2 Orthogonal complements

In this short section we introduce an important concept that will form the basis of subsequent
developments.
Definition 6.9. Let Y be a subspace of Rn . A vector x ∈ Rn is said to be orthogonal to
Y if x is orthogonal to every vector in Y . The set of all vectors in Rn that are orthogonal
to Y is called the orthogonal complement of Y and is denoted by Y ⊥ (pronounced ‘Y
perpendicular’ or ‘Y perp’ for short). Thus
Y ⊥ = { x ∈ Rn | x·y = 0 for all y ∈ Y } .
Example 6.10. Let W be a plane through the origin in R3 and let L be the line through
the origin and perpendicular to W . By construction, each vector in W is orthogonal to every
vector in L, and each vector in L is orthogonal to every vector in W . Hence
L⊥ = W and W⊥ = L.
The following theorem collects some useful facts about orthogonal complements.
Theorem 6.11. Let Y be a subspace of Rn . Then:
(a) Y ⊥ is a subspace of Rn .
(b) A vector x belongs to Y ⊥ if and only if x is orthogonal to every vector in a set that
spans Y .
Proof. See Exercises 4 and 6 on Coursework 9.
We finish this section with an application of the concepts introduced so far to the column
space and the nullspace of a matrix. These subspaces are sometimes called the fundamental
subspaces of a matrix. The next theorem, a veritable gem of Linear Algebra, shows that
the fundamental subspaces of a matrix and that of its transpose are intimately related via
orthogonality:
Theorem 6.12 (Fundamental Subspace Theorem). Let A ∈ Rm×n . Then:
(a) N (A) = col(AT )⊥ .
(b) N (AT ) = col(A)⊥ .
Proof. In this proof we shall identify the rows of A (which are strictly speaking 1 × n matrices)
with vectors in Rn .
(a) Let x ∈ Rn . Then
x ∈ N (A) ⇐⇒ Ax = 0
⇐⇒ x is orthogonal to every row of A
⇐⇒ x is orthogonal to every column of AT
⇐⇒ x ∈ col(AT )⊥ ,
so N (A) = col(AT )⊥ .
(b) Apply (a) to AT .
6.3 Orthogonal sets

In this section we shall investigate collections of vectors with the property that each vector
is orthogonal to every other vector in the set. As we shall see, these sets have a number of
astonishing properties.
Definition 6.13. A set of vectors {u1 , . . . , ur } in Rn is said to be an orthogonal set if each
pair of distinct vectors is orthogonal, that is, if
ui ·uj = 0 whenever i 6= j .
6.3. ORTHOGONAL SETS 79
Example 6.14. If     
3 −1 −1
u1 = 1 , u2 =  2  , u3 = −4 ,
1 1 7
then {u1 , u2 , u3 } is an orthogonal set since
u1 ·u2 = 3 · (−1) + 1 · 2 + 1 · 1 = 0
u1 ·u3 = 3 · (−1) + 1 · (−4) + 1 · 7 = 0
u2 ·u3 = (−1) · (−1) + 2 · (−4) + 1 · 7 = 0
The next theorem contains the first, perhaps surprising, property of orthogonal sets:
Theorem 6.15. If {u1 , . . . , ur } is an orthogonal set of nonzero vectors, then the vectors
u1 , . . . , ur are linearly independent.
Proof. Suppose that
c1 u1 + c2 u2 + · · · + cr ur = 0 .
Then
0 = 0·u1
= (c1 u1 + c2 u2 + · · · + cr ur )·u1
= c1 (u1 ·u1 ) + c2 (u2 · u1 ) + · · · + cr (ur ·u1 )
= c1 (u1 ·u1 ) ,
since u1 is orthogonal to u2 , . . . , ur . But since u1 is nonzero, u1 ·u1 is nonzero, so c1 = 0.

Similarly, c2 , . . . , cr must be zero, and the assertion follows.
Definition 6.16. An orthogonal basis for a subspace H of Rn is a basis of H that is also
an orthogonal set.
The following theorem reveals why orthogonal bases are much ‘nicer’ than other bases in
that the coordinates of a vector with respect to an orthogonal basis are spectacularly easy to
compute:
Theorem 6.17. Let {u1 , . . . , ur } be an orthogonal basis for a subspace H of Rn and let
y ∈ H. If c1 , . . . , cr are the coordinates of y with respect to {u1 , . . . , ur }, that is,
y = c1 u1 + · · · + cr ur ,
then
y·uj
cj = for each j = 1, . . . , r.
uj ·uj
Proof. As in the proof of the preceding theorem, the orthogonality of {u1 , . . . , ur } implies
that
y·u1 = (c1 u1 + · · · + cr ur )·u1 = c1 (u1 ·u1 ) .
Since u1 ·u1 is not zero (why?), we can solve for c1 in the above equation and find the stated
expression. In order to find cj for j = 2, . . . , r, compute y·uj and solve for cj .
Example 6.18. Show that the set {u1 , u2 , u3 } in Example 6.14 is an orthogonal basis for R3
and express the vector y = (6, 1, −8)T as a linear combination of the basis vectors.
Solution. Note that by Theorem 6.15 the vectors in the orthogonal set {u1 , u2 , u3 } are linearly
independent, so must form a basis for R3 , since dim R3 = 3.
Now
y·u1 = 11 , y·u2 = −12 , y·u3 = −66 ,
u1 ·u1 = 11 , u2 ·u2 = 6 , u3 ·u3 = 66 ,
so
y·u1 y·u2 y·u3 11 −12 −66
y= u1 + u2 + u3 = u1 + u2 + u3 = u1 − 2u2 − u3 .
u1 ·u1 u2 ·u2 u3 ·u3 11 6 66
6.4 Orthonormal sets

Definition 6.19. A set {u1 , . . . , ur } of vectors in Rn is said to be an orthonormal set if it
is an orthogonal set of unit vectors. Thus {u1 , . . . , ur } is an orthonormal set if and only if
ui ·uj = δij for i, j = 1, . . . , r ,
where (
1 if i = j
δij = .
0 if i 6= j
If H is a subspace of Rn spanned by {u1 , . . . , ur }, then {u1 , . . . , ur } is said to be an or-
thonormal basis for H.
Example 6.20. The standard basis {e1 , . . . , en } of Rn is an orthonormal set (and also an
orthonormal basis for Rn ). Moreover, any nonempty subset of {e1 , . . . , en } is an orthonormal
set.
Here is a less trivial example:
Example 6.21. If
 √  √ 
 
2/√6 −1/√ 3 0√
u1 = 1/√6 , u2 =  1/√3  , u3 = −1/√ 2 ,
1/ 6 1/ 3 1/ 2
then {u1 , u2 , u3 } is an orthonormal set, since
√ √ √
u1 ·u2 = −2/ 18 + 1/ 18 + 1/ 18 = 0
√ √ √
u1 ·u3 = 0/ 12 − 1/ 12 + 1/ 12 = 0
√ √ √
u2 ·u3 = 0/ 6 − 1/ 6 + 1/ 6 = 0
and
u1 ·u1 = 4/6 + 1/6 + 1/6 = 1
u2 ·u2 = 1/3 + 1/3 + 1/3 = 1
u3 ·u3 = 0/2 + 1/2 + 1/2 = 1
Moreover, since by Theorem 6.15 the vectors u1 , u2 , u3 are linearly independent and dim R3 = 3,
the set {u1 , u2 , u3 } is a basis for R3 . Thus {u1 , u2 , u3 } is an orthonormal basis for R3 .
6.4. ORTHONORMAL SETS 81
Matrices whose columns form an orthonormal set are important in applications, in particular
in computational algorithms. We are now going to explore some of their properties.
Theorem 6.22. An m × n matrix U has orthonormal columns if and only if U T U = I.
Proof. As an illustration of the general idea, suppose for the moment that U has only three
columns, each a vector in Rm . Write

U = u1 u2 u3 .
Then    T 
uT1 u 1 u1 u T
1 u 2 u T
1 u 3
U T U = uT2  u1 u2 u3 = uT2 u1 uT2 u2 uT2 u3 

uT3 uT3 u1 uT3 u2 uT3 u3

so the (i, j)-entry of U T U is just ui ·uj and the assertion follows from the definition of or-
thonormality.
The proof for the general case is exactly the same, once you have convinced yourself that
the (i, j)-entry of U T U is the dot product of the i-th column of U with the j-th column of
U.
The following theorem is a simple consequence:
Theorem 6.23. Let U ∈ Rm×n have orthonormal columns, and let x and y in Rn . Then:
(a) (U x)·(U y) = x·y;
(b) kU xk = kxk;
(c) (U x)·(U y) = 0 if and only if x·y = 0.

Here is a rephrasing of the theorem above in the language of linear transformations. Let
U ∈ Rm×n be a matrix with orthonormal columns and let LU be the corresponding linear
transformation from Rn to Rm . Property (b) says that the mapping LU preserves the lengths
of vectors, and (c) says that LU preserves orthogonality. These properties are important for
many computer algorithms.
Before concluding this section we mention a class of matrices that fits naturally in the
present context and which will play an important role in the next chapter:
Definition 6.24. A square matrix Q is said to be orthogonal if QT Q = I.
The above considerations show that every square matrix with orthonormal columns is an
orthogonal matrix. Two other interesting properties of orthogonal matrices are contained in
the following theorem.
Theorem 6.25. Let Q ∈ Rn×n be an orthogonal matrix. Then:
(a) Q is invertible and Q−1 = QT ;
(b) if {v1 , . . . , vn } is an orthonormal basis for Rn , then {Qv1 , . . . , Qvn } is an orthonormal

basis for Rn .

6.5 Orthogonal projections

In this section we shall study a particularly nice way of decomposing an arbitrary vector in
Rn . More precisely, if H is a subspace of Rn and y any vector in Rn then, as we shall see,
we can write y = ŷ + z where ŷ is in H, and z is orthogonal to H. This is a very useful
technique which has a number of interesting consequences, some of which you will see later
in this chapter.
Theorem 6.26 (Orthogonal Decomposition Theorem). Let H be a subspace of Rn . Then

each y in Rn can be written uniquely in the form
y = ŷ + z , (6.1)
where ŷ ∈ H and z ∈ H ⊥ . In fact, if {u1 , . . . , ur } is an orthogonal basis for H, then

y·u1 y·ur
ŷ = u1 + · · · + ur , (6.2)
u1 ·u1 ur ·ur
and z = y − ŷ.
Proof. Let ŷ be given by (6.2). Since ŷ is a linear combination of the vectors u1 , . . . , ur , the
vector ŷ must belong to H. Let z = y − ŷ. Then
z·u1 = (y − ŷ)·u1

u·u1
= y·u1 − (u1 ·u1 ) − 0 − · · · − 0
u1 ·u1
= y·u1 − y·u1
= 0,
so z is orthogonal to u1 . Similarly, we see that z is orthogonal to uj for j = 2, . . . , r, so

z ∈ H ⊥ by Theorem 6.11 (b).
In order to see that the decomposition (6.1) is unique, suppose that y can also be written
as y = yˆ1 + z1 , where ŷ1 ∈ H and z1 ∈ H ⊥ . Thus ŷ + z = ŷ1 + z1 , so
ŷ − ŷ1 = z1 − z .
The above equality shows that the vector v = ŷ − ŷ1 belongs to both H and H ⊥ . Thus
v·v = 0, which implies v = 0. Therefore ŷ = ŷ1 and z = z1 , so the decomposition (6.1) is
unique.
The vector ŷ in (6.1) is called the orthogonal projection of y onto H , and is written
projH y, that is,
ŷ = projH y .
One of the reasons why orthogonal projections play an important role in Linear Algebra,
and indeed in other branches of Mathematics, is made plain in the following theorem:
Theorem 6.27 (Best Approximation Theorem). Let H be a subspace of Rn , y any vector in

Rn , and ŷ = projH y. Then ŷ is the closest point in H to y, in the sense that
ky − ŷk < ky − vk (6.3)
for all v ∈ H distinct from ŷ.

6.6. GRAM SCHMIDT PROCESS 83
Proof. Take v ∈ H distinct from ŷ. Then ŷ − v ∈ H. By the Orthogonal Decomposition

Theorem, y − ŷ is orthogonal to H, so y − ŷ is orthogonal to ŷ − v.
Since
y − v = (y − ŷ) + (ŷ − v) ,
the Pythagorean Theorem (Theorem 6.8) gives
ky − vk2 = ky − ŷk2 + kŷ − vk2 .
But kŷ − vk2 > 0, since ŷ 6= v, so the desired inequality (6.3) holds.
The theorem above is the reason why the orthogonal projection of y onto H is often called
the best approximation of y by elements in H.
We conclude this section with the following consequence of the Orthogonal Decomposition
Theorem:
Theorem 6.28. Let H be a subspace of Rn . Then
(a) (H ⊥ )⊥ = H;
(b) dim H + dim H ⊥ = n.
6.6 Gram Schmidt process

In the previous sections we saw on a number of occasions how useful orthogonal bases of sub-
spaces can be. Witness, for example, the explicit expression for the orthogonal projection onto
a subspace given in the Orthogonal Decomposition Theorem. So far we have not addressed
the problem of how to manufacture an orthogonal basis. It turns out that there is a simple
algorithm that does just that, namely producing an orthogonal basis for any nonzero subspace
of Rn :
Theorem 6.29 (Gram Schmidt process). Given a basis {x1 , . . . , xr } of a subspace H of Rn

define
v1 = x1
x2 ·v1
v2 = x2 − v1
v1 ·v1
x3 ·v1 x3 ·v2
v3 = x3 − v1 − v2
v1 ·v1 v2 ·v2
..
.
xr ·v1 xr ·v2 xr ·vr−1
vr = xr − v1 − v2 − · · · − vr−1
v1 ·v1 v2 ·v2 vr−1 ·vr−1
Then {v1 , . . . , vr } is an orthogonal basis for H. In addition
Span (v1 , . . . , vk ) = Span (x1 , . . . , xk ) for 1 ≤ k ≤ r .

Proof. Write Hk = Span (x1 , . . . , xk ). Set v1 = x1 , so that Span (v1 ) = Span (x1 ). Suppose
that for some k < r we have already constructed v1 , . . . , vk so that {v1 , . . . , vk } is an
orthogonal basis for Hk . Define
vk+1 = xk+1 − projHk xk+1 .
By the Orthogonal Decomposition Theorem, vk+1 is orthogonal to Hk . Now the orthogonal

projection projHk xk+1 belongs to Hk , which in turn is a subset of Hk+1 , so vk+1 ∈ Hk+1 .
Moreover, vk+1 6= 0, since xk+1 6∈ Hk . Thus {v1 , . . . , vk+1 } is an orthogonal set of nonzero
vectors in Hk+1 . But dim Hk+1 = k + 1, so Hk+1 = Span (v1 , . . . , vk+1 ).
Remark 6.30. As with almost all the other results and techniques presented in this module,
the best way to remember the Gram Schmidt process is to understand the proof. Here is the
idea in a nut-shell: the Gram Schmidt process is an iterative procedure; if, at some stage, the
orthogonal vectors v1 , . . . , vk have already been constructed, the next vector vk+1 is obtained
by subtracting the orthogonal projection of xk+1 onto Hk = Span (v1 , . . . vk ) from xk+1 , that
is,
vk+1 = xk+1 − projHk xk+1 ,
as this makes the vector vk+1 orthogonal Hk , and thus in particular, orthogonal to all previously
constructed vectors v1 , . . . , vk .
Example 6.31. Let H = Span (x1 , x2 , x3 ) where

     
1 0 0
1 1 0
x1 = 
1 , x2 = 1 ,
   x3 = 
2 .

1 2 6
Clearly {x1 , x2 , x3 } is a basis of H. Construct an orthogonal basis of H.
Solution. We start by setting  

1
1
1 .
v1 = x1 =  
1
The vector v2 is constructed by subtracting the orthogonal projection of x2 onto Span (v1 )
from x2 , that is,  
−1
x2 ·v1 4  0
v2 = x2 − v1 = x2 − v1 =  0.

v1 ·v1 4
1
The vector v3 is constructed by subtracting the orthogonal projection of x3 onto Span (v1 , v2 )
from x3 , that is,
 
1
x3 ·v1 x3 ·v2 8 6  −2
v3 = x3 − v1 − v2 = x3 − v1 − v2 =  0,

v1 ·v1 v2 ·v2 4 2
1
producing the orthogonal basis {v1 , v2 , v3 } for H.

6.7. LEAST SQUARES PROBLEMS 85
6.7 Least squares problems

A type of problem that often arises in applications of Linear Algebra is to make sense of an
overdetermined system
Ax = b , (6.4)
where A ∈ Rm×n with m > n and b ∈ Rm , for example the one I showed in class, which
involved a 280 × 2 matrix. Clearly, the system (6.4) will not have a solution for every b ∈ Rm ;
in fact, as you will recall, the system has a solution if and only if b ∈ col(A).
What do we do if we still demand a solution? The idea is to find an x ∈ Rn that makes
Ax as close as possible to b. In other words, in cases where no exact solution exists, we
think of Ax as an approximation to b. The smaller the distance between b and Ax, given by
kb − Axk, the better the approximation.
The general least squares problem is to find x ∈ Rn that makes kb − Axk as small as
possible. Here, ‘least squares’ refers to the fact that kb − Axk is the square root of a sum of
squares.
Definition 6.32. Let A ∈ Rm×n and b ∈ Rm . A least squares solution of Ax = b is an

x̂ ∈ Rn such that
kb − Ax̂k ≤ kb − Axk for all x ∈ Rn .
How do we find least squares solutions of a given system Ax = b? To motivate the result
to follow, let
b̂ = projcol(A) b .
Since b̂ is in col(A), the equation Ax = b̂ is consistent, and there is an x̂ ∈ Rn such that
Ax̂ = b̂ .
By the Orthogonal Decomposition Theorem, b − b̂ is orthogonal to col(A), so
b − Ax̂ ∈ col(A)⊥ .
But by the Fundamental Subspace Theorem col(A)⊥ = N (AT ), so
b − Ax̂ ∈ N (AT ) .
Thus
AT (b − Ax̂) = 0 ,
and hence
AT Ax̂ = AT b .
To summarise what we have just said: a least squares solution of Ax = b satisfies
AT Ax = AT b . (6.5)
The matrix equation (6.5) represents a system of equations called the normal equations for
Ax = b.
Theorem 6.33. Let A ∈ Rm×n and b ∈ Rm . The set of least squares solutions of Ax = b
coincides with the non-empty solution set of the normal equations
AT Ax = AT b . (6.6)
Proof. We have just seen that a least squares solution x̂ must satisfy the normal equations.
It turns out that the argument outlined also works in the reverse direction. To be precise,
suppose that x̂ satisfies the normal equations, that is
AT Ax̂ = AT b .
Then AT (b − Ax̂) = 0, so b − Ax̂ ∈ N (AT ) = col(A)⊥ . Thus b − Ax̂ is orthogonal to
col(A). Hence the equation
b = Ax̂ + (b − Ax̂)
is a decomposition of b into a sum of a vector in col(A) and a vector orthogonal to it. By
the uniqueness of the orthogonal decomposition, Ax̂ must be the orthogonal projection of b
onto col(A). Thus, Ax̂ = b̂, and x̂ is a least squares solution.
Example 6.34. Find the least squares solution of the inconsistent system Ax = b, where
   
4 0 2
A = 0 2 b =  0  .
1 1 11
Solution. Compute  
4 0
4 0 1  17 1
AT A = 0 2 =
 ,
0 2 1 1 5
1 1
 
2
T 4 0 1   19
A b= 0 = .
0 2 1 11
11
Thus the normal equations AT Ax = AT b are

17 1 x1 19
= .
1 5 x2 11
This system can (and should, in general!) be solved by Gaussian elimination. In our case,
however, it is quicker to spot that the coefficient matrix is invertible with inverse

T −1 1 5 −1
(A A) = ,
84 −1 17
so AT Ax = AT b can be solved my multiplying both sides with (AT A)−1 from the left, giving
the least squares solution

T −1 T 1 5 −1 19 1
x̂ = (A A) A b = = .
84 −1 17 11 2
Often (but not always!) the matrix AT A is invertible, and the method shown in the
example above can be used. In general, the least squares solution need not be unique, and
Gaussian elimination has to used to solve the normal equations. The following theorem gives
necessary and sufficient conditions for AT A to be invertible.
Theorem 6.35. Let A ∈ Rm×n and b ∈ Rm . The matrix AT A is invertible if and only if
the columns of A are linearly independent. In this case, Ax = b has only one least squares
solution x̂, given by
x̂ = (AT A)−1 AT b .
Proof. See Exercise 1 on Coursework 11 for the first part. The remaining assertion follows as
in the previous example.
Chapter 7
Eigenvalues and Eigenvectors
In this last chapter of our exploration of Linear Algebra we will revisit eigenvalues and eigen-
vectors of matrices, two concepts you have already encountered in Geometry I and, in various
guises, in other modules. Let me first try to refresh your memory.

If A is a square n × n matrix, we may regard it as a linear transformation from Rn to Rn .
This transformation sends a vector x ∈ Rn to the vector Ax. For certain vectors, this action
can be very simple.
Example 7.1. Let

1 2 1 2
A= , u= , w= .
−1 4 1 1
Then

1 2 1 3
Au = = = 3u
−1 4 1 3

1 2 2 4
Aw = = = 2w
−1 4 1 2
so the action of A on u and w is very easy to picture: it simply amounts to a stretching by 3

and 2, respectively.
Definition 7.2. An eigenvector of an n × n matrix A is a nonzero vector x such that
Ax = λx ,
for some scalar λ. A scalar λ is called an eigenvalue of A if there is a non-trivial solution x to

Ax = λx, in which case we say that x is an eigenvector corresponding to the eigenvalue
λ.
Remark 7.3. Note that if x is an eigenvector of a matrix A with eigenvalue λ, then any
nonzero multiple of x is also an eigenvector corresponding to λ, since
A(αx) = αAx = αλx = λ(αx) .
87
88 CHAPTER 7. EIGENVALUES AND EIGENVECTORS
In the example above, u is an eigenvector of A corresponding to the eigenvalue 3, and w

is an eigenvector of A corresponding to the eigenvalue 2. One of the reasons why eigenvalues
and eigenvectors play a prominent role in Linear Algebra is that knowledge of the eigenvalues
and eigenvectors of a matrix sometimes goes a long way to understand the action of A on any
vector. In order to see this, suppose that x is an arbitrary vector in R2 . Since u and w are
linearly independent, they form a basis of R2 , and we can write
x = c1 u + c2 w
for some scalars c1 and c2 . But then
Ax = c1 Au + c2 Aw = 3c1 u + 2c2 w ,
so if we define the basis B = {u, w} of R2 , then

3 0
[Ax]B = [x]B .
0 2
In other words, relative to the basis B, the action of A on an arbitrary vector is very easy to
understand. We shall explore this observation in more detail shortly, and we shall see that it
holds the key to a new and powerful perspective on matrices.
We shall now investigate how to determine all the eigenvalues and eigenvectors of an n × n
matrix A. We start by observing that the defining equation Ax = λx can be written
(A − λI)x = 0 . (7.1)
Thus λ is an eigenvalue of A if and only if (7.1) has a non-trivial solution. The set of solutions
of (7.1) is N (A − λI), that is, the nullspace of A − λI, which is a subspace of Rn .Thus, λ is
an eigenvalue of A if and only if
N (A − λI) 6= {0} ,
and any nonzero vector in N (A − λI) is an eigenvector belonging to λ. Moreover, by the

Invertible Matrix Theorem, (7.1) has a non-trivial solution if and only if the matrix A − λI is
singular, or equivalently
det(A − λI) = 0 . (7.2)
Notice now that if the determinant in (7.2) is expanded we obtain a polynomial of degree n
in the variable λ,
p(λ) = det(A − λI) ,
called the characteristic polynomial of A , and equation (7.2) is called the characteristic
equation of A. So, in other words, the roots of the characteristic polynomial of A are exactly
the eigenvalues of A. The following theorem summarises our findings so far:
Theorem 7.4. Let A be an n × n matrix and λ a scalar. The following statements are
equivalent:
(a) λ is an eigenvalue of A;
(b) (A − λI)x = 0 has a non-trivial solution;
(c) N (A − λI) 6= {0};

(d) A − λI is singular;
(e) det(A − λI) = 0.

In view of the above theorem the following concept arises naturally:
Definition 7.5. If A is a square matrix and λ an eigenvalue of A, then N (A − λI) is called
the eigenspace corresponding to λ.
Note that if λ is an eigenvalue of a matrix A, then every nonzero vector in the corre-
sponding eigenspace N (A − λI) is an eigenvector corresponding to λ, and conversely, the set
of all eigenvectors corresponding to λ together with the zero vector forms the corresponding
eigenspace N (A − λI).
We shall now see how to use (e) in the theorem above to determine the eigenvalues and
eigenvectors of a given matrix.
Example 7.6. Find all eigenvalues and eigenvectors of the matrix

−7 −6
A= .
9 8
Proof. First we calculate the characteristic polynomial of A

−7 − λ −6
det(A − λI) = = (−7 − λ)(8 − λ) − (−6) · 9
9 8 − λ
= −56 + 7λ − 8λ + λ2 + 54 = λ2 − λ − 2 = (λ + 1)(λ − 2) .
Thus the characteristic equation is
(λ + 1)(λ − 2) = 0 ,
so the eigenvalues of the matrix are λ1 = −1 and λ2 = 2.

In order to find the eigenvectors belonging to λ1 = −1 we must determine the nullspace
of A − λ1 I = A + I. Or, put differently, we need to determine all solutions of the system
(A + I)x = 0. This can be done using your favourite method, but, for reasons which will
become clear in the next example, I strongly recommend Gaussian elimination, that is, we
bring the augmented matrix (A + I|0) to row echelon form:

−7 + 1 −6 0 −6 −6 0 1 1 0 1 1 0
(A + I|0) = = ∼ ∼ ,
9 8+1 0 9 9 0 9 9 0 0 0 0
so setting x2 = α we find x1 = −x2 = −α. Thus every vector in N (A + I) is of the form

−α −1
=α ,
α 1
so the eigenspace corresponding to the eigenvalue −1 is

−1
α α∈R ,
1

−1
and any nonzero multiple of is an eigenvector corresponding to the eigenvalue −1.
1
Similarly, in order to find the eigenvectors belonging to λ2 = 2 we bring (A − λ2 I|0) =

(A − 2I|0) to row echelon form:
1 23 0 1 23 0

−7 − 2 −6 0 −9 −6 0
(A − 2I|0) = = ∼ ∼ ,
9 8−2 0 9 6 0 9 6 0 0 0 0
so setting x2 = α we find x1 = − 23 x2 = − 23 α. Thus every vector in N (A − 2I) is of the form

2
−3α α −2
= ,
α 3 3
so the eigenspace corresponding to the eigenvalue 2 is

−2
α α∈R ,
3

−2
and any nonzero multiple of is an eigenvector corresponding to the eigenvalue 2.
3
Before we continue with another example you might want to have another look at the
above calculations of eigenspaces. Observe that since we need to solve a homogeneous linear
system there is no need to write down the right-most column of the augmented matrix (since
it consists only of zeros); we simply perform elementary row operations on the coefficient
matrix, keeping in mind that the right-most column of the augmented matrix will remain the
zero column. We shall use this short-cut in all the following calculations of eigenspaces.

2 −3 1
A = 1 −2 1 .
1 −3 2
Find the eigenvalues and corresponding eigenspaces.
Solution. A slightly tedious calculation using repeated cofactor expansions shows that the
characteristic polynomial of A is

2 − λ −3 1

det(A − λI) = 1
−2 − λ 1 = −λ(λ − 1)2 ,
1 −3 2 − λ
so the eigenvalues of A are λ1 = 0 and λ2 = 1.

In order to find the eigenspace corresponding to λ1 we find the nullspace of A − λ1 I = A
using Gaussian elimination:
   
2 −3 1 1 0 −1
A = 1 −2 1 ∼ · · · ∼ 0 1 −1 ,
1 −3 2 0 0 0
so setting x3 = α we find x2 = 0 − (−1)x3 = α and x1 = 0 − (−1)x3 = α. Thus, every

vector in N (A) is of the form    
α 1
α = α 1 ,
α 1
so the eigenspace corresponding to the eigenvalue 0 is

   
 1 
α 1 α ∈ R .
1
 
In order to find the eigenspace corresponding to λ2 we find the nullspace of A−λ2 I = A−I,
again using Gaussian elimination:
     
2−1 −3 1 1 −3 1 1 −3 1
A−I = 1 −2 − 1 1  = 1 −3 1 ∼ 0 0 0 ,
1 −3 2−1 1 −3 1 0 0 0
so setting x2 = α and x3 = β we find x1 = 3x2 −x3 = 3α −β. Thus every vector in N (A−I)
is of the form      
3α − β 3 −1
 α  = α 1 + β  0  ,
β 0 1
and so the eigenspace corresponding to the eigenvalue 1 is
     
 3 −1 
α 1 + β  0  α, β ∈ R .
0 1
 
Example 7.8. Find the eigenvalues of the matrix

 
1 2 3
A = 0 4 5 .
0 0 6
Solution. Using the fact that the determinant of a triangular matrix is the product of the
diagonal entries we find

1 − λ 2 3

det(A − λI) = 0
4−λ 5 = (1 − λ)(4 − λ)(6 − λ) ,
0 0 6 − λ
so the eigenvalues of A are 1, 4, and 6.

The above example and its method of solution are easily generalised:
Theorem 7.9. The eigenvalues of a triangular matrix are precisely the diagonal entries of the
matrix.
The next theorem gives an important sufficient (but not necessary) condition for two
matrices to have the same eigenvalues. It also serves as the foundation for many numerical
procedures to approximate eigenvalues of matrices, some of which you will encounter if you
take the module MTH5110, Introduction to Numerical Computing.
Theorem 7.10. Let A and B be two n × n matrices and suppose that A and B are similar,
that is, there is an invertible matrix S ∈ Rn×n such that B = S −1 AS. Then A and B have
the same characteristic polynomial, and, consequently, have the same eigenvalues.
Proof. If B = S −1 AS, then
B − λI = S −1 AS − λI = S −1 AS − λS −1 S = S −1 (AS − λS) = S −1 (A − λI)S .
Thus, using the multiplicativity of determinants,
det(B − λI) = det(S −1 ) det(A − λI) det(S) = det(A − λI) ,
because det(S −1 ) det(S) = det(S −1 S) = det(I) = 1.
We will revisit this theorem from a different perspective in the next section.
7.2 Diagonalisation
In many applications of Linear Algebra one is faced with the following problem: given a square
matrix A, find the k-th power Ak of A for large values of k. In general, this can be a very
onerous task. For certain matrices, however, evaluating powers is spectacularly easy:
Example 7.11. Let D ∈ R2×2 be given by

2 0
D= .
0 3
Then 2
2 2 0 2 0 2 0
D = =
0 3 0 3 0 32
and 2 3
3 2 2 0 2 0 2 0
D = DD = = .
0 3 0 32 0 33
In general,
k 2k 0
D = ,
0 3k
for k ≥ 1.
After having had another look at the example above, you should convince yourself that
if D is a diagonal n × n matrix with diagonal entries d1 , . . . , dn , then Dk is the diagonal
matrix whose diagonal entries are dk1 , . . . , dkn . The moral of this is that calculating powers for
diagonal matrices is easy. What if the matrix is not diagonal? The next best situation arises
if the matrix is similar to a diagonal matrix. In this case, calculating powers is almost as easy
as calculating powers of diagonal matrices, as we shall see shortly. We shall now single out
matrices with this property and give them a special name:
Definition 7.12. An n × n matrix A is said to be diagonalisable if it is similar to a diagonal

matrix, that is, if there is an invertible matrix P ∈ Rn×n such that
P −1 AP = D ,
where D is a diagonal matrix. In this case we say that P diagonalises A.

7.2. DIAGONALISATION 93
Note that if A is a matrix which is diagonalised by P , that is, P −1 AP = D with D

diagonal, then
A = P DP −1 ,
A2 = P DP −1 P DP −1 = P D2 P −1 ,
A3 = AA2 = P DP −1 P D2 P −1 = P D3 P −1 ,
and in general
Ak = P Dk P −1 ,
for any k ≥ 1. Thus powers of A are easily computed, as claimed.
Before giving a characterisation of diagonalisable matrices we require an auxiliary result of
independent interest:
Theorem 7.13. If v1 , . . . , vr are eigenvectors that correspond to distinct eigenvalues λ1 , . . . , λr

of an n × n matrix A, then the vectors v1 , . . . , vr are linearly independent.
Proof. By contradiction. Suppose to the contrary that the vectors v1 , . . . , vr are linearly
dependent. We may then assume (after reordering the v’s and the λ’s if necessary), that
there is an index p < r such that vp+1 is a linear combination of the preceding linearly
independent vectors. Thus there exist scalars c1 , . . . , cp such that
c1 v1 + · · · + cp vp = vp+1 . (7.3)
Multiplying both sides of the above equation by A and using the fact that Avk = λk vk for
each k, we obtain
c1 λ1 v1 + · · · + cp λp vp = λp+1 vp+1 . (7.4)
Multiplying both sides of (7.3) by λp+1 and subtracting the result from (7.4), we see that
c1 (λ1 − λp+1 )v1 + · · · + cp (λp − λp+1 )vp = 0 . (7.5)
Since the vectors v1 , . . . , vp are linearly independent, the coefficients in (7.5) must all be zero.
But none of the factors λk − λp+1 is zero, because the eigenvalues are distinct, so we must
have c1 = . . . = cp = 0. But then (7.3) implies that vp+1 = 0, which is impossible, because
vp+1 is an eigenvector. Thus our assumption that the vectors v1 , . . . , vr are linearly dependent
must be false, that is, the vectors v1 , . . . , vr are linearly independent.
We are now able to prove the main the result of this section:
Theorem 7.14 (Diagonalisation Theorem). An n × n matrix A is diagonalisable if and only

if A has n linearly independent eigenvectors.
In fact, P −1 AP = D, with D a diagonal matrix, if and only if the columns of P are n
linearly independent eigenvectors of A. In this case, the diagonal entries of D are eigenvalues
of A that correspond, respectively, to the eigenvectors in P .
Proof. First observe that if P is any n × n matrix with columns v1 , . . . , vn , and if D is any
diagonal matrix with diagonal entries λ1 , . . . , λn , then

AP = A v1 · · · vn = Av1 · · · Avn , (7.6)
while
P D = λv1 · · · λn vn . (7.7)
First we prove the ‘only if’ part of the theorem. Suppose that A is diagonalisable and that
−1
P AP = D. Then AP = P D, and equations (7.6) and (7.7) imply that

Av1 · · · Avn = λ1 v1 · · · λn Avn . (7.8)
Thus
Av1 = λ1 v1 , Av2 = λ2 v2 , ..., Avn = λn vn . (7.9)
Since P is invertible its columns v1 , . . . , vn must be linearly independent. Moreover, none of its
columns can be zero, so (7.9) implies that λ1 , . . . , λn are eigenvalues of A with corresponding
eigenvectors v1 , . . . , vn . This finishes the proof of the ‘only if’ part.
In order to prove the ‘if’ part, suppose that A has n linearly independent eigenvectors
v1 , . . . , vn with corresponding eigenvalues λ1 , . . . , λn . Let P be the matrix whose columns
are v1 , . . . , vn and let D be the diagonal matrix whose diagonal entries are λ1 , . . . , λn . Then
equations (7.6), (7.7), and (7.8) imply that AP = P D. But since the columns of P are
linearly independent, P is invertible, so P −1 AP = D as claimed.
Example 7.15. Diagonalise the following matrix, if possible:
 
−7 3 −3
A = −9 5 −3 .
9 −3 5
Solution. A slightly tedious calculation shows that the characteristic polynomial is given by

−7 − λ 3 −3

p(λ) = det(A − λI) = −9 5 − λ −3 = −λ3 + 3λ2 − 4 .
9 −3 5 − λ
The cubic p above can be factored by spotting that −1 is a root. Polynomial division then
yields
p(λ) = −(λ + 1)(λ2 − 4λ + 4) = −(λ + 1)(λ − 2)2 ,
so the distinct eigenvalues of A are 2 and −1.
The usual methods (see Examples 7.6 and 7.7) now produce a basis for each of the two
eigenspaces and it turns out that
   
1 −1
N (A − 2I) = Span (v1 , v2 ), where v1 = 3 , v2 =
   0,
0 3
 
−1
N (A + I) = Span (v3 ), where v3 = −1 .
1
You may now want to confirm, using your favourite method, that the three vectors
v1 , v2 , v3 are linearly independent. As we shall see shortly, this is not really necessary: the
union of basis vectors for eigenspaces always produces linearly independent vectors (see the
proof of Corollary 7.17 below).
Thus, A is diagonalisable, since it has 3 linearly independent eigenvectors. In order to find
the diagonalising matrix P we recall that defining
 
1 −1 −1
P = v1 v2 v3 = 3 0 −1
0 3 1
7.2. DIAGONALISATION 95
does the trick, that is, P −1 AP = D, where D is the diagonal matrix whose entries are the
eigenvalues of A and where the order of the eigenvalues matches the order chosen for the
eigenvectors in P , that is,  
2 0 0
D = 0 2 0  .
0 0 −1
It is good practice to check that P and D really do the job they are supposed to do:
    
−7 3 −3 1 −1 −1 2 −2 1
AP = −9 5 −3 3 0 −1 = 6 0 1,
9 −3 5 0 3 1 0 6 −1
    
1 −1 −1 2 0 0 2 −2 1
P D = 3 0 −1 0 2 0  = 6 0 1,
0 3 1 0 0 −1 0 6 −1
so AP = P D, and hence P −1 AP = D as required.
Example 7.16. Diagonalise the following matrix, if possible:

 
−6 3 −2
A = −7 5 −1 .
8 −3 4
Solution. The characteristic polynomial of A turns out to be exactly the same as in the previous
example:
det(A − λI) = −λ3 + 3λ2 − 4 = −(λ + 1)(λ − 2)2 .
Thus the eigenvalues of A are 2 and −1. However, in this case it turns out that both
eigenspaces are 1-dimensional:
 
1
N (A − 2I) = Span (v1 ) where v1 =  2,
−1
 
−1
N (A + I) = Span (v2 ) where v1 = −1 .
1
Since A has only 2 linearly independent eigenvectors, the Diagonalisation Theorem implies
that A is not diagonalisable.
Put differently, the Diagonalisation Theorem states that a matrix A ∈ Rn×n is diagonalis-
able if and only if A has enough eigenvectors to form a basis of Rn . The following corollary
makes this restatement even more precise:
Corollary 7.17. Let A ∈ Rn×n and let λ1 , . . . , λr be the (distinct) eigenvalues of A. Then
A is diagonalisable if and only if
dim N (A − λ1 ) + · · · + dim N (A − λr I) = n .
Proof. For simplicity we will only consider the case where A has two distinct eigenvalues; the
general case can be proved along the same lines. Suppose that A has two distinct eigenvalues
λ1 and λ2 . Let {v1 , . . . , vp } be a basis for N (A − λ1 I) and let {w1 , . . . , wq } be a basis
for N (A − λ2 I). First we show that the p + q vectors v1 , . . . , vp , w1 , . . . , wq are linearly
independent. In order to see this suppose that there are scalars α1 , . . . , αp , β1 , . . . , βq such
that
α1 v1 + · · · + αp vp + β1 w1 + · · · + βq wq = 0 . (7.10)
Theorem 7.13 now implies that
α1 v1 + · · · + αp vp = 0 and β1 w1 + · · · + βq wq = 0 , (7.11)
because otherwise (7.10) would state that 0 can be written as a nontrivial linear combination
of eigenvectors belonging to distinct eigenvalues of A. But since v1 , . . . , vp and w1 , . . . wq
are linearly independent, it follows that α1 = · · · = αp = β1 = · · · = βq = 0. Thus the p + q
vectors v1 , . . . , vp , w1 , . . . , wq are linearly independent as claimed.
By the Diagonalisation Theorem A is diagonalisable if and only if A has n linearly inde-
pendent eigenvectors. By what we have just proved, this is the case if and only if p + q = n,
and the corollary follows because dim N (A − λ1 ) = p and dim N (A − λ2 I) = q.
A very useful special case of the Diagonalisation Theorem is the following:
Theorem 7.18. An n × n matrix with n distinct eigenvalues is diagonalisable.
Proof. Let v1 , . . . , vn be eigenvectors corresponding to the n distinct eigenvalues of A. Then
the n vectors v1 , . . . , vn are linearly independent by Theorem 7.13. Hence A is diagonalisable
by the Diagonalisation Theorem.
Remark 7.19. Note that the above condition for diagonalisability is sufficient but not nec-
essary : an n × n matrix which does not have n distinct eigenvalues may or may not be
diagonalisable (see Examples 7.15 and 7.16).
Example 7.20. The matrix  
1 −1 5
A = 0 2 6
0 0 3
is diagonalisable, since it has three distinct eigenvalues 1, 2, and 3.
7.3 Interlude: complex vector spaces and matrices

Consider the matrix
0 −1
A= .
1 0
What are the eigenvalues of A? Notice that

−λ −1
det(A − λI) = = λ2 + 1 ,
1 −λ
so the characteristic polynomial does not have any real roots, and hence A does not have any
real eigenvalues. However, since
λ2 + 1 = λ2 − (−1) = λ − i2 = (λ − i)(λ + i) ,
7.3. INTERLUDE: COMPLEX VECTOR SPACES AND MATRICES 97
the characteristic polynomial has two complex roots, namely i and −i. Thus it makes sense to
say that A has two complex eigenvalues i and −i. What are the corresponding eigenvectors?
Solving
(A − iI)x = 0
leads to the system
−ix1 − x2 = 0
x1 + ix2 = 0

1
Both equations yield the condition x2 = −ix1 , so is an eigenvector corresponding to
−i
the eigenvalue i. Indeed

0 −1 1 i i 1
= = 2 =i .
1 0 −i 1 −i −i

1
Similarly, we see that is an eigenvector corresponding to the eigenvalue −i. Indeed
i

0 −1 1 −i −i 1
= = 2 = −i .
1 0 i 1 −i i
The moral of this example is the following: on the one hand, we could just say that the matrix
A has no real eigenvalues and stop the discussion right here. On the other hand, we just saw
that it makes sense to say that A has two complex eigenvalues with corresponding complex
eigenvectors.
This leads to the idea of leaving our current real set-up, and enter a complex realm instead.
As it turns out, this is an immensely powerful idea. However, as our time is limited, we shall
only cover the bare necessities, allowing us to prove the main result of the next section.
Let Cn denote the set of all n-vectors with complex entries, that is,
  
 z1
 

n  .. 
C =  .  z1 , . . . , zn ∈ C .

 z 

n
Just as in Rn , we add vectors in Cn by adding their entries, and we can multiply a vector in
Cn by a complex number, by multiplying each entry.
Example 7.21. Let z, w ∈ C3 and α ∈ C, with

   
1+i −2 + 3i
z =  2i  , w =  1  , α = (1 + 2i) .
3 2+i
Then    
(1 + i) + (−2 + 3i) −1 + 4i
z+w = 2i + 1  =  1 + 2i 
3 + (2 + i) 5+i
     
(1 + 2i)(1 + i) 1 + 2i + i + 2i2 −1 + 3i
αz =  (1 + 2i)(2i)  =  2i + (2i)2  = −4 + 2i
(1 + 2i) · 3 3 + 6i 3 + 6i
If addition and scalar multiplication is defined in this way (now allowing scalars to be in
C), then Cn satisfies all the axioms of a vector space. Similarly, we can introduce the set
of all m × n matrices with complex entries, call it Cm×n , and define addition and scalar
multiplication (again allowing complex scalars) entry-wise just as in Rm×n . Again, Cm×n
satisfies all the axioms of a vector space.
Fact 7.22. All the results in Chapters 1–5, and all the results from the beginning of this
chapter hold verbatim, if ‘scalar’ is taken to mean ‘complex number’.
Since ‘scalars’ are now allowed to be complex numbers, Cn and Cm×n are known as
complex vector spaces.
The reason for allowing this more general set-up is that, in a certain sense, complex
numbers are much nicer than real numbers. More precisely, we have the following result:
Theorem 7.23 (Fundamental Theorem of Algebra). If p is a complex polynomial of degree
n ≥ 1, that is,
p(z) = cn z n + · · · + c1 z + c0 ,
where c0 , c1 , . . . , cn ∈ C, then p has at least one (possibly complex) root.
Corollary 7.24. Every matrix A ∈ Cn×n has at least one (possibly complex) eigenvalue and
a corresponding eigenvector z ∈ Cn .
Proof. Since λ is an eigenvalue of A if and only if det(A − λI) = 0 and since p(λ) =
det(A − λI) is a polynomial with complex coefficients of degree n, the assertion follows from
the Fundamental Theorem of Algebra.
The corollary above is the main reason why complex vector spaces are considered. We are
guaranteed that every matrix has at least one eigenvalue, and we may then use the powerful
tools developed in the earlier parts of this chapter to analyse matrices through their eigenvalues
and eigenvectors.
7.4 Spectral Theorem for Symmetric Matrices

This last section of the last chapter is devoted to one of the gems of Linear Algebra: the
Spectral Theorem. This result, which has many applications, a number of which you will see
in other modules, is concerned with the diagonalisability of symmetric matrices. Recall that a
matrix A ∈ Rn×n is said to be diagonalisable if there is an invertible matrix P ∈ Rn×n such
that P −1 AP is diagonal. We already know that A is diagonalisable if and only if A has n
linearly independent eigenvectors. However, this condition is difficult to check in practice. It
may thus come as a surprise that there is a sufficiently rich class of matrices that are always
diagonalisable, and moreover, that the diagonalising matrix P is of a special form. This is the
content of the Spectral Theorem for Symmetric Matrices, or Spectral Theorem for short:1
Theorem 7.25 (Spectral Theorem for Symmetric Matrices). Let A ∈ Rn×n be symmetric.
Then there is an orthogonal matrix Q ∈ Rn×n such that
QT AQ = D,
where D ∈ Rn×n is diagonal.
Or put differently: every symmetric matrix can be diagonalised by an orthogonal matrix.
1
There are other, more general versions of the Spectral Theorem. In this course, we will only consider the
symmetric case.
7.4. SPECTRAL THEOREM FOR SYMMETRIC MATRICES 99
The proof of this theorem requires a number of auxiliary results which we shall first state
and prove.
Lemma 7.26. The product of two orthogonal matrices of the same size is orthogonal, that
is, if Q0 ∈ Rn×n and Q1 ∈ Rn×n are orthogonal, then Q0 Q1 is also orthogonal.
Proof. Since, by definition, QT0 Q0 = I and QT1 Q1 = I, we have
(Q0 Q1 )T Q0 Q1 = QT1 QT0 Q0 Q1 = QT1 Q1 = I ,
so Q0 Q1 is orthogonal.
Lemma 7.27. Let A ∈ Rn×n be symmetric. Then, for any vectors x, y ∈ Rn ,
(Ax)·y = x·(Ay) .
Proof. Since AT = A, we have
(Ax)·y = (Ax)T y = xT AT y = xT Ay = x·(Ay) .
Lemma 7.28. Let A ∈ Rn×n be a symmetric matrix. If v ∈ Rn is an eigenvector of A and

H = Span (v), then
LA (H ⊥ ) ⊂ H ⊥ ,
that is,
w ∈ H ⊥ ⇒ Aw ∈ H ⊥ .
Proof. Let λ be the eigenvalue corresponding to the eigenvector v. Suppose that w ∈ H ⊥ ,

that is, w·v = 0. Then, by Lemma 7.27,
(Aw)·v = w·(Av) = w·(λv) = λ(w·v) = 0 ,
so Aw ∈ H ⊥ as claimed.
The following lemma contains the key result that will allow us to prove the Spectral
Theorem. We already know from the discussion in the previous section that every matrix has
a complex eigenvalue and a corresponding complex eigenvector. The following lemma shows
that, rather unexpectedly, symmetric matrices always have at least one real eigenvalue with a
corresponding real eigenvector:
Lemma 7.29. Every symmetric matrix A ∈ Rn×n has at least one real eigenvalue with
corresponding real eigenvector v ∈ Rn .
Proof. By Corollary 7.24 we know that A has at least one complex eigenvalue λ with corre-
sponding eigenvector z ∈ Cn , that is
Az = λz . (7.12)
Write
λ = a + ib where a, b ∈ R (7.13)
z = v + iw where v, w ∈ Rn
Thus, using (7.12) we have
Av + iAw = A(v + iw) = Az = λz = (a + ib)(v + iw) = (av − bw) + i(aw + bv) ,
which, by comparing real and imaginary parts, yields
Av = av − bw , (7.14)
Aw = aw + bv .
Now, by Lemma 7.27, we have
(av − bw)·w = (Av)·w = v·(Aw) = v·(aw + bv) ,
so
a(v·w) − bkwk2 = a(v·w) + bkvk2
and hence
b(kvk2 + kwk2 ) = 0 .
But z = v + iw is an eigenvector, so the vectors v and w cannot both be the zero vector.
Therefore kvk2 + kwk2 > 0, and hence
b = 0.
Thus, using (7.13) and (7.14), we see that λ = a ∈ R and Av = av = λv, that is, A has a
real eigenvalue with corresponding real eigenvector.
Proof of the Spectral Theorem. By induction on n. The theorem is trivial for n = 1. Suppose
now that the theorem has been shown to be true for some k. Let A ∈ R(k+1)×(k+1) be
symmetric. By Lemma 7.29 there is λ ∈ R and v ∈ Rk+1 such that
Av = λv .
Since multiplying an eigenvector by a nonzero scalar does not change its status as an eigen-
vector, we may assume that v is normalised, that is, kvk = 1. Next choose an orthonormal
basis {w1 , . . . , wk } for Span (v)⊥ . Then {v1 , w1 , . . . , wk } is an orthonormal basis for Rk+1
and it follows that the matrix Q0 ∈ R(k+1)×(k+1) given by

Q 0 = v w1 · · · w k
is an orthogonal matrix. Now

   
vT vT Av vT Aw1 · · · vT Awk
wT  T
w1 Av

T  1 T T
Q0 AQ0 =  ..  A v w1 · · · wk =  .. ,

 .   . B 
T
wk wk Av
where B = (bij ) is a real k × k matrix with
bij = wiT Awj . (7.15)
Moreover, since
vT Av = vT (λv) = λ(vT v) = λ ,
7.4. SPECTRAL THEOREM FOR SYMMETRIC MATRICES 101
and since, by Lemma 7.27 and Lemma 7.28,
vT Awj = v·(Awj ) = 0
wiT Av = wi ·(Av) = (Awi )·v = 0 ,
it follows that  
λ 0 ··· 0
0 
T
Q0 AQ0 =  .. .
 
. B 
0
Observe now that by Lemma 7.27 the k × k matrix B is symmetric, because
bij = wiT Awj = wi ·(Awj ) = (Awi )·wj = wj ·(Awi ) = wjT Awi = bji .
Thus, by the inductive hypothesis, there is an orthogonal k × k matrix R such that
RT BR = D ,
with D ∈ Rk×k diagonal. Define Q1 ∈ R(k+1)×(k+1) by

 
1 0 ··· 0
0 
Q1 =  .. .
 
. R 
0
A short calculation using the fact that RT R = I shows that QT1 Q1 = I, so Q1 is orthogonal.
If we now let
Q = Q0 Q1 ,
then Q is orthogonal by Lemma 7.26, and
     
λ 0 ··· 0 λ 0 ··· 0 λ 0 ··· 0
0  0  0 
QT AQ = QT1 QT0 AQ0 Q1 = QT1  .. Q = = ,
     
 1  .. T
  ..
. B  . R BR   . D 
0 0 0
so QT AQ is diagonal. To summarise: we have just shown that if a k × k symmetric matrix

can be diagonalised by an orthogonal matrix, then so can a (k +1) × (k +1) symmetric matrix.
Combining this with the initial observation that, trivially, any 1 × 1 matrix can be diagonalised,
the truth of the theorem now follows by the principle of induction.
Let us record the following consequence of the Spectral Theorem, which is of independent
interest:
Corollary 7.30. The eigenvalues of a symmetric matrix A are real, and eigenvectors corre-
sponding to distinct eigenvalues are orthogonal.
Proof. Combine the Diagonalisation Theorem and the Spectral Theorem.

Example 7.31. Consider the symmetric matrix

 
0 2 −1
A= 2 3 −2 .
−1 −2 0
Find an orthogonal matrix Q that diagonalises A.
Solution. The characteristic polynomial of A is
det(A − λI) = −λ3 + 3λ2 + 9λ + 5 = (1 + λ)2 (5 − λ) ,
so the eigenvalues of A are −1 and 5. Computing N (A + I) in the usual way shows that
{x1 , x1 } is a basis for N (A + I) where
   
1 −2
x1 = 0 , x2 =  1  .
1 0
Similarly, we find that the eigenspace N (A − 5I) corresponding to the eigenvalue 5 is 1-
dimensional with basis  
−1
x3 = −2 .
1
In order to construct the diagonalising orthogonal matrix for A it suffices to find orthonormal
bases for each of the eigenspaces, since, by the previous corollary, eigenvectors corresponding
to distinct eigenvalues are orthogonal.
In order to find an orthonormal basis for N (A + I) we apply the Gram Schmidt process to
the basis {x1 , x2 } to produce the orthogonal set {v1 , v2 }:
 
1
v1 = x1 = 0 ,
1
 
−1
x2 ·v1
v2 = x2 − v1 =  1  .
v1 ·v1
1
Now {v1 , v2 , x3 } is an orthogonal basis of R3 consisting of eigenvectors of A, so normalising
them to produce
 √   √   √ 
1/ 2 −1/√ 3 −1/√6
1 1 1
u1 = v1 =  0√  , u2 = v2 =  1/√3  , u3 = x3 = −2/√ 6 ,
kv1 k kv2 k kx3 k
1/ 2 1/ 3 1/ 6
allows us to write down the orthogonal matrix
 √ √ √ 
1/ 2 −1/
√ 3 −1/√6
Q= 1 u u 2 u 3 =
 0√ 1/√3 −2/√ 6
1/ 2 1/ 3 1/ 6
which diagonalises A, that is,
 
−1 0 0
QT AQ =  0 −1 0 .
0 0 5
Index
adjugate, 32 inverse, 11
Inverse Formula, 32
basis, 49 Invertible Matrix Theorem, 19
best approximation, 83
Best Approximation Theorem, 82 kernel, 62
characteristic equation, 88 leading 1, 3
characteristic polynomial, 88 least squares problem, 85
cofactor, 25 least squares solution, 85
cofactor expansion, 25 linear combination, 16
Cofactor Expansion Theorem, 25 linear equation, 1
column space, 55 linear mapping, 59
commuting matrices, 11 linear transformation, 59
complex vector space, 36, 98 linearly dependent vectors, 46
composite, 69 linearly independent vectors, 46
Composition Formula, 70
coordinate vector, 53 matrix, 9
coordinates, 53 augmented, 3
Cramer’s Rule, 31 coefficient, 3
diagonal, 14
determinant, 24 invertible, 11
diagonalisable matrix, 92 lower triangular, 14
Diagonalisation Theorem, 93 nonsingular, 28
dimension, 51 singular, 28
distance, 76 symmetric, 13
dot product, 75 upper triangular, 14
matrix representation, 67
eigenspace, 89
Matrix Representation Theorem, 66
eigenvalue, 87
eigenvector, 87 norm, 76
elementary row operation, 3 normal equations, 85
equivalent systems, 2 normalising, 76
Fundamental Subspace Theorem, 78 nullspace, 41
fundamental subspaces, 78 orthogonal basis, 79
Gauss-Jordan algorithm, 6 orthogonal complement, 77
Gauss-Jordan inversion, 20 Orthogonal Decomposition Theorem, 82
Gaussian algorithm, 5 orthogonal matrix, 81
Gaussian elimination, 3 orthogonal projection, 82
orthogonal set, 78
identity, 62 orthogonal vectors, 77
identity transformation, 62 orthonormal basis, 80
image, 62 orthonormal set, 80
103
104 INDEX
pivot column, 7
pivot position, 7
Pythagorean Theorem, 77
range, 62
rank, 56
real vector space, 36
reduced row echelon form, 3
row echelon form, 3
row equivalent matrices, 18
row space, 55
scalar product, 75
solution, 1
trivial, 8
zero, 8
span, 42
spanning set, 43
Spectral Theorem, 98
standard basis, 49, 50
subspace, 39
proper, 39
zero, 39
system
associated homogeneous, 8
consistent, 2
homogeneous, 8
inconsistent, 2
inhomogeneous, 8
overdetermined, 7
solution set of a, 2
underdetermined, 7
transition matrix, 54
transpose, 12
unit vector, 76
variable
free, 4
leading, 4
vector space, 36
finite dimensional, 51
infinite dimensional, 51
zero vector, 19

MTH5112 Linear Algebra I, 2009-2010: Oscar Bandtlow 30 March 2010

Uploaded by

Copyright:

Available Formats

MTH5112 Linear Algebra I, 2009-2010: Oscar Bandtlow 30 March 2010

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

MTH5112 Linear Algebra I, 2009-2010: Oscar Bandtlow 30 March 2010

Uploaded by

Copyright:

Available Formats

MTH5112 Linear Algebra I, 2009–2010

1 Systems of Linear Equations 1

6.5 Orthogonal projections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82

7 Eigenvalues and Eigenvectors 87

Systems of Linear Equations

1.1 Basic terminology and examples

Example 1.4. Consider the two systems

5x1 − x2 + 2x3 = −3 5x1 − x2 + 2x3 = −3

(i) interchanging two equations;

(ii) multiplying an equation by a non-zero scalar;

(iii) adding a multiple of one equation to another.

we call the array  

the augmented matrix of the linear system, and the m × n matrix

1.2 Gaussian elimination

Solution. (a) The corresponding system is

so the system is inconsistent and the solution set is empty.

Step 3 Now multiply that row by 1/a to create a leading 1.

Solution. Performing the Gaussian algorithm on the augmented matrix gives:

Solution. Performing the Gauss-Jordan algorithm on the augmented matrix gives:

1.3 Special classes of linear systems

• underdetermined if it has fewer equations than unknowns (i.e. m < n).

Definition 1.20. A linear system

a11 x1 + a12 x2 + · · · + a1n xn = b1

is said to be homogeneous if bi = 0 for all i. Otherwise it is said to be inhomogeneous.

3x1 + 2x2 + 5x3 = 2 3x1 + 2x2 + 5x3 = 0

Theorem 1.22. An underdetermined homogeneous system always has non-trivial solutions.

Proof. Follows from the following two observations:

2.1 Revision from Geometry I

Example 2.5. Let    

(d) A + (−A) = O, where −A = (−1)A;

(e) α(A + B) = αA + αB;

(f) (α + β)A = αA + βA;

(g) (αβ)A = α(βA);

2(A + 3B) − 3(C + 2B) = 2A + 2 · 3B − 3C − 3 · 2B = 2A + 6B − 3C − 6B = 2A − 3C .

Definition 2.9 (Matrix multplication). If A = (aij ) is an m × n matrix and B = (bij ) is an

Matrix multiplication satisfies the following rules proved in Geometry I:

(b) A(BC) = (AB)C;

(c) A(B + C) = AB + AC;

(d) (B + C)A = BA + CA;

(e) α(AB) = (αA)B = A(αB).

• If A is a square matrix we write Ak = AA · · · A} for the k-th power of A.

Warning: In general AB 6= BA, even if AB and BA have the same size!

Finally we recall the notion of an inverse of a matrix.

Definition 2.16. If A is a square matrix, a matrix B is called an inverse of A if

A matrix that has an inverse is called invertible.

cannot have an inverse since for any 2 × 2 matrix B = (bij ) we have

It turns out that if a matrix is invertible its inverse is uniquely determined:

Theorem 2.17. If B and C are both inverses of A, then B = C.

Proof. Since B and C are inverses of A we have AB = I and CA = I. Thus

Proof. Observe that

(AB)(B −1 A−1 ) = A(BB −1 )A−1 = AIA−1 = AA−1 = I ,

(B −1 A−1 )(AB) = B −1 (A−1 A)B = B −1 IB = B −1 B = I .

2.2 Transpose of a matrix

Definition 2.19. The transpose of an m × n matrix A = (aij ) is the n × m matrix B = (bij )

Matrix transposition satisfies the following rules:

(b) (αA)T = α(AT );

Transposition ties in nicely with invertibility:

Theorem 2.22. Let A be invertible. Then AT is invertible and