MA106 Linear Algebra lecture notes

Lecturers: Martin Bright and Daan Krammer

Warwick, January 2011

Contents
1 Number systems and fields 3
1.1 Axioms for number systems . . . . . . . . . . . . . . . . . . . . . . . . . 3

2 Vector spaces 4
2.1 Examples of vector spaces . . . . . . . . . . . . . . . . . . . . . . . . . . 5

3 Linear independence, spanning and bases of vector spaces 6

3.1 Linear dependence and independence . . . . . . . . . . . . . . . . . . . . 6
3.2 Spanning vectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
3.3 Bases of vector spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
3.4 Existence of a basis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

4 Subspaces 10

5 Linear transformations 13
5.1 Definition and examples . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
5.2 Kernels and images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
5.3 Rank and nullity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
5.4 Operations on linear maps . . . . . . . . . . . . . . . . . . . . . . . . . . 17

6 Matrices 18

7 Linear transformations and matrices 20

7.1 Setting up the correspondence . . . . . . . . . . . . . . . . . . . . . . . . 20
7.2 The correspondence between operations on linear maps and matrices . . 22
7.3 Linear equations and the inverse image problem . . . . . . . . . . . . . . 24

8 Elementary operations and the rank of a matrix 25

8.1 Gauss transformations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
8.2 Elementary row operations . . . . . . . . . . . . . . . . . . . . . . . . . 26
8.3 The augmented matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
8.4 Row reducing a matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
8.5 Elementary column operations . . . . . . . . . . . . . . . . . . . . . . . 29
8.6 The rank of a matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
8.7 The rank criterion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33

9 The inverse of a linear transformation and of a matrix 33

9.1 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
9.2 Matrix inversion by row reduction . . . . . . . . . . . . . . . . . . . . . 34
9.3 Elementary matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
9.4 Application to linear equations . . . . . . . . . . . . . . . . . . . . . . . 36

1
10 The determinant of a matrix 37
10.1 Definition of the determinant . . . . . . . . . . . . . . . . . . . . . . . . 37
10.2 The effect of matrix operations on the determinant . . . . . . . . . . . . 39
10.3 The determinant of a product . . . . . . . . . . . . . . . . . . . . . . . . 42
10.4 Minors and cofactors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
10.5 The inverse of a matrix using determinants . . . . . . . . . . . . . . . . 44
10.6 Cramer’s rule for solving simultaneous equations . . . . . . . . . . . . . 45

11 Change of basis and equivalent matrices 45

12 Similar matrices, eigenvectors and eigenvalues 48

12.1 Similar matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
12.2 Eigenvectors and eigenvalues . . . . . . . . . . . . . . . . . . . . . . . . 49
12.3 The scalar product – symmetric and orthogonal matrices . . . . . . . . . 54

2
1 Number systems and fields
We introduce the number systems most commonly used in mathematics.

1. The natural numbers N = {1, 2, 3, 4, . . .}.

In N, addition is possible but not subtraction; e.g. 2 − 3 6∈ N.

2. The integers Z = {. . . , −2, −1, 0, 1, 2, 3, . . .}.

In Z, addition, subtraction and multiplication are always possible, but not division;
e.g. 2/3 6∈ Z.

3. The rational numbers Q = {p/q | p, q ∈ Z, q 6= 0}.

In Q, addition, subtraction,
√ multiplication and division (except by zero) are all
possible. However, 2 6∈ Q.

4. The real numbers R. These are the numbers which can be expressed as decimals.
The rational numbers are those with finite or recurring decimals.
In R, addition, subtraction, multiplication and division (except
√ by zero) are still
possible, and all positive numbers have square roots, but −1 6∈ R.

5. The complex numbers C = {x + iy | x, y ∈ R}, where i2 = −1.

In C, addition, subtraction, multiplication and division (except by zero) are still
possible, and all, numbers have square roots. In fact all polynomial equations with
coefficients in C have solutions in C.

1.1 Axioms for number systems

Laws governing the way numbers combine together are called axioms. Any particular
axiom might be true in some number systems but not in others.

Axioms for addition. Let S be a number system.

A1. α + β = β + α for all α, β ∈ S.

A2. (α + β) + γ = α + (β + γ) for all α, β, γ ∈ S.

A3. There is a number 0 ∈ S such that α + 0 = 0 + α = α for all α ∈ S.

A4. For each number α ∈ S there exists a number −α ∈ S such that α + (−α) =
(−α) + α = 0.

These axioms may or may not be satisfied by a given number system S. For example,
in N, A1 and A2 hold but A3 and A4 do not hold. A1–A4 all hold in Z, Q, R and C.

Axioms for multiplication.

M1. α.β = β.α for all α, β ∈ S.

M2. (α.β).γ = α.(β.γ) for all α, β, γ ∈ S.

M3. There is a number 1 ∈ S such that α.1 = 1.α = α for all α ∈ S.

M4. For each number α ∈ S with α 6= 0, there exists a number α−1 ∈ S such that
α.α−1 = α−1 .α = 1.

In N and Z, M1–M3 hold but M4 does not hold. M1–M4 all hold in Q, R and C.

3
Axiom relating addition and multiplication.

D. (α + β).γ = α.γ + β.γ for all α, β, γ ∈ S.

Definition. A set S on which addition and multiplication are defined is called a field if
it satisfies each of the axioms A1, A2, A3, A4, M1, M2, M3, M4, D, and if, in addition,
1 6= 0.

Roughly speaking, S is a field if addition, subtraction, multiplication and division

(except by zero) are all possible in S. We shall always use the letter K for a general
field.

Example. N and Z are not fields, but Q, R and C are all fields.

There are many other fields, including some finite fields. For example, for each prime
number p, there is a field Fp = {0, 1, 2, . . . , p − 1} with p elements, where addition and
multiplication are carried out modulo p. Thus, in F7 , we have 5 + 4 = 2, 5 × 4 = 6 and
5−1 = 3 because 5 × 3 = 1. The smallest such field F2 has just two elements 0 and 1,
where 1+1 = 0. This field is extremely important in Computer Science since an element
of F2 represents a bit of information.
Various other familiar properties of numbers, such as 0α = 0, (−α)β = −(αβ) =
α(−β), (−α)(−β) = αβ, (−1)α = −α, for all α, β ∈ S, can be proved from the axioms.
Why would we want to do this, when we can see they’re true anyway? The point is
that, when we meet a new number system, it is enough to check whether the axioms
hold; if they do, then all these properties follow automatically.
However, occasionally you need to be careful. For example, in F2 we have 1 + 1 = 0,
and so it is not possible to divide by 2 in this field.

2 Vector spaces
Definition. A vector space over a field K is a set V which has two basic operations,
addition and scalar multiplication, satisfying certain requirements. Thus for every pair
u, v ∈ V , u + v ∈ V is defined, and for every α ∈ K, αv ∈ V is defined. For V to
be called a vector space, the following axioms must be satisfied for all α, β ∈ K and all
u, v ∈ V .

(i) Vector addition satisfies axioms A1, A2, A3 and A4.

(ii) α(u + v) = αu + αv;

(iii) (α + β)v = αv + βv;

(iv) (αβ)v = α(βv);

(v) 1v = v.

Elements of the field K will be called scalars. Note that we will use boldface letters
like v to denote vectors. The zero vector in V will be written as 0V , or usually just 0.
This is different from the zero scalar 0 = 0K ∈ K.

For nearly all results in this course, there is no loss in assuming that K is the field R
of real numbers. So you may assume this if you find it helpful to do so. Just occasionally,
we will need to assume K = C the field of complex numbers.
However, it is important to note that nearly all arguments in Linear Algebra use
only the axioms for a field and so are valid for any field, which is why shall use a general
field K for most of the course.

4
2.1 Examples of vector spaces
1. K n = {(α1 , α2 , . . . , αn ) | αi ∈ K}. This is the space of row vectors. Addition and
scalar multiplication are defined by the obvious rules:

(α1 , α2 , . . . , αn ) + (β1 , β2 , . . . , βn ) = (α1 + β1 , α2 + β2 , . . . , αn + βn );

λ(α1 , α2 , . . . , αn ) = (λα1 , λα2 , . . . , λαn ).

The most familiar examples are

R2 = {(x, y) | x, y ∈ R} and R3 = {(x, y, z) | x, y, z ∈ R},

which we can think of geometrically as the points in ordinary 2- and 3-dimensional

space, equipped with a coordinate system.
Vectors in R2 and R3 can also be thought of as directed lines joining the origin to
the points with coordinates (x, y) or (x, y, z).

: (x, y)

 


0 =(0,0) 

Addition of vectors is then given by the parallelogram law.


  v1 + v2
*
 
  
  
v 2  
   :
  v1
  

  
 
0



Note that K 1 is essentially the same as K itself.

2. Let K[x] be the set of polynomials in an indeterminate x with coefficients in the

field K. That is,

K[x] = {α0 + α1 x + · · · + αn xn | n ≥ 0, αi ∈ K}.

Then K[x] is a vector space over K.

3. Let K[x]≤n be the set of polynomials over K of degree at most n, for some n ≥ 0.
Then K[x]≤n is also a vector space over K; in fact it is a subspace of K[x].
Note that the polynomials of degree exactly n do not form a vector space. (Why
not?)

4. Let K = R and let V be the set of n-times differentiable functions f : R → R

which are solutions of the differential equation

dn f dn−1 f df
λ0 + λ 1 + · · · + λn−1 + λn f = 0.
dxn dxn−1 dx
for fixed λ0 , λ1 , . . . , λn ∈ R. Then V is a vector space over R, for if f (x) and
g(x) are both solutions of this equation, then so are f (x) + g(x) and αf (x) for all
α ∈ R.

5. The previous example is a space of functions. There are many such examples that
are important in Analysis. For example, the set C k ((0, 1), R), consisting of all
functions f : (0, 1) → R such that the kth derivative f (k) exists and is continuous,
is a vector space over R with the usual pointwise definitions of addition and scalar
multiplication of functions.

5
 6. Any n bits of information can be thought of as a vector in Fn2 .
Facing such a variety of vector spaces, a mathematician wants to derive useful meth-
ods of handling all these vector spaces. If work out techniques for dealing with a single
example, say R3 , how can we be certain that our methods will also work for R8 or even
C8 ? That is why we use the axiomatic approach to developing mathematics. We must
use only arguments based on the vector space axioms. We have to avoid making any
other assumptions. This ensures that everything we prove is valid for all vector spaces,
not just the familiar ones like R3 .
We shall be assuming the following additional simple properties of vectors and scalars
from now on. They can all be deduced from the axioms (and it is a useful exercise to
do so).
(i) α0 = 0 for all α ∈ K
(ii) 0v = 0 for all v ∈ V
(iii) −(αv) = (−α)v = α(−v), for all α ∈ K and v ∈ V .
(iv) if αv = 0 then α = 0 or v = 0.

3 Linear independence, spanning and bases of vector

spaces
3.1 Linear dependence and independence
Definition. Let V be a vector space over the field K. The vectors v1 , v2 , . . . vn are
said to be linearly dependent if there exist scalars α1 , α2 , . . . , αn ∈ K, not all zero, such
that
α1 v1 + α2 v2 + · · · + αn vn = 0.
v1 , v2 , . . . vn are said to be linearly independent if they are not linearly dependent. In
other words, they are linearly independent if the only scalars α1 , α2 , . . . , αn ∈ K that
satisfy the above equation are α1 = 0, α2 = 0, . . . , αn = 0.
Definition. Vectors of the form α1 v1 + α2 v2 + · · · + αn vn for α1 , α2 , . . . , αn ∈ K are
called linear combinations of v1 , v2 , . . . vn .
Example. Let V = R2 , v1 = (1, 3), v2 = (2, 5).
Then α1 v1 + α2 v2 = (α1 + 2α2 , 3α1 + 5α2 ), which is equal to 0 = (0, 0) if and only
if α1 + 2α2 = 0 and 3α1 + 5α2 = 0. Thus we have a pair of simultaneous equations in
α1 , α2 and the only solution is α1 = α2 = 0, so v1 , v2 are linearly independent.
Example. Let V = Q2 , v1 = (1, 3), v2 = (2, 6).
This time the equations are α1 + 2α2 = 0 and 3α1 + 6α2 = 0, and there are non-zero
solutions, such as α1 = −2, α2 = 1, and so v1 , v2 are linearly dependent.
Lemma 3.1. v1 , v2 , . . . , vn ∈ V are linearly dependent if and only if either v1 = 0 or,
for some r, vr is a linear combination of v1 , . . . , vr−1 .
Proof. If v1 = 0 then by putting α1 = 1 and αi = 0 for i > 1 we get α1 v1 +· · ·+αn vn =
0, so v1 , v2 , . . . , vn ∈ V are linearly dependent.
If vr is a linear combination of v1 , . . . , vr−1 , then vr = α1 v1 + · · · + αr−1 vr−1 for
some α1 , . . . , αr−1 ∈ K and so we get α1 v1 + · · · + αr−1 vr−1 − 1 vr = 0 and again
v1 , v2 , . . . , vn ∈ V are linearly dependent.
Conversely, suppose that v1 , v2 , . . . , vn ∈ V are linearly dependent, and αi are
scalars, not all zero, satisfying α1 v1 + α2 v2 + · · · + αn vn = 0. Let r be maximal
with αr 6= 0; then α1 v1 + α2 v2 + · · · + αr vr = 0. If r = 1 then α1 v1 = 0 which, by (iv)
above, is only possible if v1 = 0. Otherwise, we get
α1 αr−1
vr = − v1 − · · · − vr−1 .
αr αr
In other words, vr is a linear combination of v1 , . . . , vr−1 .

6
3.2 Spanning vectors
Definition. The vectors v1 , . . . , vn in V span V if every vector v ∈ V is a linear
combination α1 v1 + α2 v2 + · · · + αn vn of v1 , . . . , vn .

3.3 Bases of vector spaces

Definition. The vectors v1 , . . . , vn in V form a basis of V if they are linearly indepen-
dent and span V .
Proposition 3.2. The vectors v1 , . . . , vn form a basis of V if and only if every v ∈ V
can be written uniquely as v = α1 v1 + α2 v2 + · · · + αn vn ; that is, the coefficients
α1 , . . . , αn are uniquely determined by the vector v.
Proof. Suppose that v1 , . . . , vn form a basis of V . Then they span V , so certainly every
v ∈ V can be written as v = α1 v1 + α2 v2 + · · · + αn vn . Suppose that we also had
v = β1 v1 + β2 v2 + · · · + βn vn for some other scalars βi ∈ K. Then we have
0 = v − v = (α1 − β1 )v1 + (α2 − β2 )v2 + · · · + (αn − βn )vn
and so
(α1 − β1 ) = (α2 − β2 ) = · · · = (αn − βn ) = 0
by linear independence of v1 , . . . , vn . Hence αi = βi for all i, which means that the αi
are uniquely determined.
Conversely, suppose that every v ∈ V can be written uniquely as v = α1 v1 + α2 v2 +
· · · + αn vn . Then v1 , . . . , vn certainly span V . If α1 v1 + α2 v2 + · · · + αn vn = 0, then
α1 v1 + α2 v2 + · · · + αn vn = 0v1 + 0v2 + · · · + 0vn
and so the uniqueness assumption implies that α1 = α2 = · · · = αn = 0, and v1 , . . . , vn
are linearly independent. Hence they form a basis of V .
Definition. The scalars α1 , . . . , αn in the statement of the proposition are called the
coordinates of v with respect to the basis v1 , . . . , vn .
With respect to a different basis, v will have different coordinates. Indeed, a basis
for a vector space can be thought of as a choice of a system of coordinates.

Examples Here are some examples of bases of vector spaces.

1. (1, 0) and (0, 1) form a basis of K 2 . This follows from Proposition 3.2, because each
element (α1 , α2 ) ∈ K 2 can be written as α1 (1, 0) + α2 (0, 1), and this expression is
clearly unique.
2. More generally, (1, 0, 0), (0, 1, 0), (0, 0, 1) form a basis of K 3 , (1, 0, 0, 0), (0, 1, 0, 0),
(0, 0, 1, 0), (0, 0, 0, 1) form a basis of K 4 and so on. This is called the standard
basis of K n for n ∈ N.
(To be precise, the standard basis of K n is v1 , . . . , vn , where vi is the vector with
a 1 in the ith position and a 0 in all other positions.)
3. There are many other bases of K n . For example (1, 0), (1, 1) form a basis of K 2 ,
because (α1 , α2 ) = (α1 − α2 )(1, 0) + α2 (1, 1), and this expression is unique. In
fact, any two non-zero vectors such that one is not a scalar multiple of the other
form a basis for K 2 .
4. As we have defined a basis, it has to consist of a finite number of vectors. Not every
vector space has a finite basis. For example, let K[x] be the space of polynomials
in x with coefficients in K. Let p1 (x), p2 (x), . . . , pn (x) be any finite collection of
polynomials in K[x]. Then, if d is the maximum degree of p1 (x), p2 (x), . . . , pn (x),
any linear combination of p1 (x), p2 (x), . . . , pn (x) has degree at most d, and so
p1 (x), p2 (x), . . . , pn (x) cannot span K[x]. On the other hand, it is possible (with
a little care) to define what it means for an infinite set of vectors to be a basis of
a vector space; in fact the infinite sequence of vectors 1, x, x2 , x3 , . . . , xn , . . . is a
basis of K[x].

7
 A vector space with a finite basis is called finite-dimensional. In fact, nearly all of
this course will be about finite-dimensional spaces, but it is important to remember that
these are not the only examples. The spaces of functions mentioned in Example 5. of
Section 2 typically have uncountably infinite dimension.
Theorem 3.3 (The basis theorem). Suppose that v1 , . . . , vm and w1 , . . . , wn are both
bases of the vector space V . Then m = n. In other words, all finite bases of V contain
the same number of vectors.
The proof of this theorem is quite tricky and uses the concept of sifting which we
introduce after the next lemma.
Definition. The number n of vectors in a basis of the finite-dimensional vector space
V is called the dimension of V and we write dim(V ) = n.
Thus, as we might expect, K n has dimension n. K[x] is infinite-dimensional, but
the space K[x]≤n of polynomials of degree at most n has basis 1, x, x2 , . . . , xn , so its
dimension is n + 1 (not n).
Note that the dimension of V depends on the field K. Thus the complex numbers
C can be considered as
• a vector space of dimension 1 over C, with one possible basis being the single
element 1;
• a vector space of dimension 2 over R, with one possible basis given by the two
elements 1, i;
• a vector space of infinite dimension over Q.
The first step towards proving the basis theorem is to be able to remove unnecessary
vectors from a spanning set of vectors.
Lemma 3.4. Suppose that the vectors v1 , v2 , . . . , vn , w span V and that w is a linear
combination of v1 , . . . , vn . Then v1 , . . . , vn span V .
Proof. Since v1 , v2 , . . . , vn , w span V , any vector v ∈ V can be written as

v = α1 v1 + · · · + αn vn + βw,

But w is a linear combination of v1 , . . . , vn , so w = γ1 v1 + · · · + γn vn for some scalars

γi , and hence
v = (α1 + βγ1 )v1 + · · · + (αn + βγn )vn
is a linear combination of v1 , . . . , vn , which therefore span V .
There is an important process, which we shall call sifting, which can be applied to
any sequence of vectors v1 , v2 , . . . , vn in a vector space V , as follows. We consider
each vector vi in turn. If it is zero, or a linear combination of the preceding vectors
v1 , . . . , vi−1 , then we remove it from the list.
Example. Let us sift the following sequence of vectors in R3 .

v1 = (0, 0, 0) v2 = (1, 1, 1) v3 = (2, 2, 2) v4 = (1, 0, 0)

v5 = (3, 2, 2) v6 = (0, 0, 0) v7 = (1, 1, 0) v8 = (0, 0, 1)
v1 = 0, so we remove it. v2 is non-zero so it stays. v3 = 2v2 so it is removed. v4 is
clearly not a linear combination of v2 , so it stays.
We have to decide next whether v5 is a linear combination of v2 , v4 . If so, then
(3, 2, 2) = α1 (1, 1, 1) + α2 (1, 0, 0), which (fairly obviously) has the solution α1 = 2,
α2 = 1, so remove v5 . Then v6 = 0 so that is removed too.
Next we try v7 = (1, 1, 0) = α1 (1, 1, 1) + α2 (1, 0, 0), and looking at the three compo-
nents, this reduces to the three equations

1 = α1 + α2 ; 1 = α1 ; 0 = α1 .

8
The second and third of these equations contradict each other, and so there is no solution.
Hence v7 is not a linear combination of v2 , v4 , and it stays.
Finally, we need to try

v8 = (0, 0, 1) = α1 (1, 1, 1) + α2 (1, 0, 0) + α3 (1, 1, 0)

leading to the three equations

0 = α1 + α2 + α3 0 = α1 + α3 ; 1 = α1

and solving these in the normal way, we find a solution α1 = 1, α2 = 0, α3 = −1. Thus
we delete v8 and we are left with just v2 , v4 , v7 .
Of course, the vectors that are removed during the sifting process depends very much
on the order of the list of vectors. For example, if v8 had come at the beginning of the
list rather than at the end, then we would have kept it.
The idea of sifting allows us to prove the following theorem, stating that every finite
sequence of vectors which spans a vector space V actually contains a basis for V .
Theorem 3.5. Suppose that the vectors v1 , . . . , vr span the vector space V . Then there
is a subsequence of v1 , . . . , vr which forms a basis of V .
Proof. We sift the vectors v1 , . . . , vr . The vectors that we remove are linear combina-
tions of the preceding vectors, and so by Lemma 3.4, the remaining vectors still span
V . After sifting, no vector is zero or a linear combination of the preceding vectors
(or it would have been removed), so by Lemma 3.1, the remaining vectors are linearly
independent. Hence they form a basis of V .
The theorem tells us that any vector space with a finite spanning set is finite-
dimensional, and indeed the spanning set contains a basis. We now prove the dual
result: any linearly independent set is contained in a basis.
Theorem 3.6. Let V be a vector space over K which has a finite spanning set, and
suppose that the vectors v1 , . . . , vr are linearly independent in V . Then we can extend
the sequence to a basis v1 , . . . , vn of V , where n ≥ r.
Proof. Suppose that w1 , . . . , wq is a spanning set for V . We sift the combined sequence

v1 , . . . , vr , w1 , . . . , wq .

Since w1 , . . . , wq span V , the whole sequence spans V . Sifting results in a basis for V as
in the proof of Theorem 3.5. Since v1 , . . . , vr are linearly independent, none of them can
be a linear combination of the preceding vectors, and hence none of the vi are deleted
in the sifting process. Thus the resulting basis contains v1 , . . . , vr .
Example. The vectors v1 = (1, 2, 0, 2), v2 = (0, 1, 0, 2) are linearly independent in R4 .
Let us extend them to a basis of R4 . The easiest thing is to append the standard basis
of R4 , giving the combined list of vectors

v1 = (1, 2, 0, 2), v2 = (0, 1, 0, 2), w1 = (1, 0, 0, 0),

w2 = (0, 1, 0, 0), w3 = (0, 0, 1, 0), w4 = (0, 0, 0, 1),
which we shall sift. We find that (1, 0, 0, 0) = α1 (1, 2, 0, 2)+α2 (0, 1, 0, 2) has no solution,
so w1 stays. However, w2 = v1 − v2 − w1 so w2 is deleted. It is clear that w3 is not a
linear combination of v1 , v2 , w1 , because all of those have a 0 in their third component.
Hence w3 remains. Now we have four linearly independent vectors, so must have a basis
at this stage, and we can stop the sifting early. The resulting basis is

v1 = (1, 2, 0, 2), v2 = (0, 1, 0, 2), w1 = (1, 0, 0, 0), w3 = (0, 0, 1, 0).

We are now ready to prove Theorem 3.3. Since bases of V are both linearly inde-
pendent and span V , the following proposition implies that any two bases contain the
same number of vectors.

9
Proposition 3.7 (The exchange lemma). Suppose that vectors v1 , . . . , vn span V and
that vectors w1 , . . . , wm ∈ V are linearly independent. Then m ≤ n.
Proof. The idea is to place the wi one by one in front of the sequence v1 , . . . , vn , sifting
each time.
Since v1 , . . . , vn span V , w1 , v1 , . . . , vn are linearly dependent, so when we sift, at
least one vj is deleted. We then place w2 in front of the resulting sequence and sift
again. Then we put w3 in from of the result, and sift again, and carry on doing this
for each wi in turn. Since w1 , . . . , wm are linearly independent none of them are ever
deleted. Each time we place a vector in front of a sequence which spans V , and so the
extended sequence is linearly dependent, and hence at least one vj gets eliminated each
time.
But in total, we append m vectors wi , and each time at least one vj is eliminated,
so we must have m ≤ n.
Corollary 3.8. Let V be a vector space of dimension n over K. Then any n vectors
which span V form a basis of V , and no n − 1 vectors can span V .
Proof. After sifting a spanning sequence as in the proof of Theorem 3.5, the remaining
vectors form a basis, so by Theorem 3.3, there must be precisely n = dim(V ) vectors
remaining. The result is now clear.
Corollary 3.9. Let V be a vector space of dimension n over K. Then any n linearly
independent vectors form a basis of V and no n + 1 vectors can be linearly independent.
Proof. By Theorem 3.6 any linearly independent set is contained in a basis but by
Theorem 3.3, there must be precisely n = dim(V ) vectors in the extended set. The
result is now clear.

3.4 Existence of a basis

Although we have studied bases quite carefully in the previous section, we have not
addressed the following fundamental question. Let V be a vector space. Does it contain
a basis?
Theorem 3.5 gives a partial answer that is good for many practical purposes. Let us
formulate it as a corollary.
Corollary 3.10. If a non-trivial vector space V is spanned by a finite number of vectors,
then it has a basis.
In fact, if we define the idea of an infinite basis carefully, then it can be proved that
any vector space has a basis. That result will not be proved in this course. Its proof,
which necessarily deals with infinite sets, requires a subtle result in axiomatic set theory
called Zorn’s lemma.

4 Subspaces
Let V be a vector space over the field K. Certain subsets of V have the nice property
of being closed under addition and scalar multiplication; that is, adding or taking scalar
multiples of vectors in the subset gives vectors which are again in the subset. We call
such a subset a subspace:
Definition. A subspace of V is a non-empty subset W ⊆ V such that
(i) W is closed under addition: u, v ∈ W ⇒ u + v ∈ W ;
(ii) W is closed under scalar multiplication: v ∈ W, α ∈ K ⇒ αv ∈ W .
These two conditions can be replaced with a single condition
u, v ∈ W, α, β ∈ K ⇒ αu + βv ∈ W.
A subspace W is itself a vector space over K under the operations of vector ad-
dition and scalar multiplication in V . Notice that all vector space axioms of W hold
automatically. (They are inherited from V .)

10
Example. The subset of R2 given by

W = {(α, β) ∈ R2 | β = 2α},

that is, the subset consisting of all row vectors whose second entry is twice their first
entry, is a subspace of R2 . You can check that adding two vectors of this form always
gives another vector of this form; and multiplying a vector of this form by a scalar always
gives another vector of this form.
For any vector space V , V is always a subspace of itself. Subspaces other than V are
sometimes called proper subspaces. We also always have a subspace {0} consisting of
the zero vector alone. This is called the trivial subspace, and its dimension is 0, because
it has no linearly independent sets of vectors at all.
Intersecting two subspaces gives a third subspace:
Proposition 4.1. If W1 and W2 are subspaces of V then so is W1 ∩ W2 .
Proof. Let u, v ∈ W1 ∩ W2 and α ∈ K. Then u + v ∈ W1 (because W1 is a subspace)
and u + v ∈ W2 (because W2 is a subspace). Hence u + v ∈ W1 ∩ W2 . Similarly, we get
αv ∈ W1 ∩ W2 , so W1 ∩ W2 is a subspace of V .
Warning! It is not necessarily true that W1 ∪ W2 is a subspace, as the following
example shows.
Example. Let V = R2 , let W1 = {(α, 0) | α ∈ R} and W2 = {(0, α) | α ∈ R}. Then
W1 , W2 are subspaces of V , but W1 ∪ W2 is not a subspace, because (1, 0), (0, 1) ∈
W1 ∪ W2 , but (1, 0) + (0, 1) = (1, 1) 6∈ W1 ∪ W2 .
Note that any subspace of V that contains W1 and W2 has to contain all vectors of
the form u + v for u ∈ W1 , v ∈ W2 . This motivates the following definition.
Definition. Let W1 , W2 be subspaces of the vector space V . Then W1 + W2 is defined
to be the set of vectors v ∈ V such that v = w1 + w2 for some w1 ∈ W1 , w2 ∈ W2 .
Or, if you prefer, W1 + W2 = {w1 + w2 | w1 ∈ W1 , w2 ∈ W2 }.
Do not confuse W1 + W2 with W1 ∪ W2 .
Proposition 4.2. If W1 , W2 are subspaces of V then so is W1 + W2 . In fact, it is the
smallest subspace that contains both W1 and W2 .
Proof. Let u, v ∈ W1 + W2 . Then u = u1 + u2 for some u1 ∈ W1 , u2 ∈ W2 and
v = v1 +v2 for some v1 ∈ W1 , v2 ∈ W2 . Then u+v = (u1 +v1 )+(u2 +v2 ) ∈ W1 +W2 .
Similarly, if α ∈ K then αv = αv1 + αv2 ∈ W1 + W2 . Thus W1 + W2 is a subspace of
V.
Any subspace of V that contains both W1 and W2 must contain W1 + W2 , so it is
the smallest such subspace.
Theorem 4.3. Let V be a finite-dimensional vector space, and let W1 , W2 be subspaces
of V . Then

dim(W1 + W2 ) = dim(W1 ) + dim(W2 ) − dim(W1 ∩ W2 ).

Proof. First note that any subspace W of V is finite-dimensional. This follows from
Corollary 3.9, because a largest linearly independent subset of W contains at most
dim(V ) vectors, and such a subset must be a basis of W .
Let dim(W1 ∩ W2 ) = r and let e1 , . . . , er be a basis of W1 ∩ W2 . Then e1 , . . . , er is
a linearly independent set of vectors, so by Theorem 3.6 it can be extended to a basis
e1 , . . . , er ,f1 , . . . , fs of W1 where dim(W1 ) = r + s, and it can also be extended to a basis
e1 , . . . , er , g1 , . . . , gt of W2 , where dim(W2 ) = r + t.
To prove the theorem, we need to show that dim(W1 + W2 ) = r + s + t, and to do
this, we shall show that
e1 , . . . , er , f1 , . . . , fs , g1 , . . . , gt
is a basis of W1 + W2 . Certainly they all lie in W1 + W2 .

11
 First we show that they span W1 + W2 . Any v ∈ W1 + W2 is equal to w1 + w2 for
some w1 ∈ W1 , w2 ∈ W2 . So we can write

w1 = α1 e1 + · · · + αr er + β1 f1 + · · · + βs fs

for some scalars αi , βj ∈ K, and

w2 = γ1 e1 + · · · + γr er + δ1 g1 + · · · + δt gt

for some scalars γi , δj ∈ K. Then

v = (α1 + γ1 )e1 + · · · + (αr + γr )er + β1 f1 + · · · + βs fs + δ1 g1 + · · · + δt gt

and so e1 , . . . , er , f1 , . . . , fs , g1 , . . . , gt span W1 + W2 .
Finally we have to show that e1 , . . . , er , f1 , . . . , fs , g1 , . . . , gt are linearly independent.
Suppose that

α1 e1 + · · · + αr er + β1 f1 + · · · + βs fs + δ1 g1 + · · · + δt gt = 0

for some scalars αi , βj , δk ∈ K. Then

α1 e1 + · · · + αr er + β1 f1 + · · · + βs fs = −δ1 g1 − · · · − δt gt (∗)

The left-hand side of this equation lies in W1 and the right-hand side of this equation lies
in W2 . Since the two sides are equal, both must in fact lie in W1 ∩ W2 . Since e1 , . . . , er
is a basis of W1 ∩ W2 , we can write

−δ1 g1 − · · · − δt gt = γ1 e1 + · · · + γr er

for some γi ∈ K, and so

γ1 e1 + · · · + γr er + δ1 g1 + · · · + δt gt = 0.

But, e1 , . . . , er , g1 , . . . , gt form a basis of W2 , so they are linearly independent, and

hence γi = 0 for 1 ≤ i ≤ r and δi = 0 for 1 ≤ i ≤ t. But now, from the equation (∗)
above, we get
α1 e1 + · · · + αr er + β1 f1 + · · · + βs fs = 0.
Now e1 , . . . , er , f1 , . . . , fs form a basis of W1 , so they are linearly independent, and hence
αi = 0 for 1 ≤ i ≤ r and βi = 0 for 1 ≤ i ≤ s. Thus e1 , . . . , er , f1 , . . . , fs , g1 , . . . , gt are
linearly independent, which completes the proof that they form a basis of W1 + W2 .
Hence

dim(W1 + W2 ) = r + s + t = (r + s) + (r + t) − r = dim(W1 ) + dim(W2 ) − dim(W1 ∩ W2 ).

Another way to form subspaces is to take linear combinations of some given vectors:

Proposition 4.4. Let v1 , . . . , vn be vectors in the vector space V . Then the set of all
linear combinations α1 v1 + α2 v2 + · · · + αn vn of v1 , . . . , vn forms a subspace of V .

The proof of this is completely routine and will be omitted. The subspace in this
proposition is known as the subspace spanned by v1 , . . . , vn .

Definition. Two subspaces W1 , W2 of V are called complementary if W1 ∩ W2 = {0}

and W1 + W2 = V .

Proposition 4.5. Let W1 , W2 be subspaces of V . Then W1 , W2 are complementary

subspaces if and only if each vector in v ∈ V can be written uniquely as v = w1 + w2
with w1 ∈ W1 and w2 ∈ W2 .

12
Proof. Suppose first that W1 , W2 are complementary subspaces and let v ∈ V . Then
W1 + W2 = V , so we can find w1 ∈ W1 and w2 ∈ W2 with v = w1 + w2 . If we also had
v = w10 + w20 with w10 ∈ W1 , w20 ∈ W2 , then we would have w1 − w10 = w20 − w2 . The
left-hand side lies in W1 and the right-hand side lies in W2 , and so both sides (being
equal) must lie in W1 ∩ W2 = {0}. Hence both sides are zero, which means w1 = w10
and w2 = w20 , so the expression is unique.
Conversely, suppose that every v ∈ V can be written uniquely as v = w1 + w2 with
w1 ∈ W1 and w2 ∈ W2 . Then certainly W1 + W2 = V . If v was a non-zero vector in
W1 ∩ W2 , then in fact v would have two distinct expressions as w1 + w2 with w1 ∈ W1
and w2 ∈ W2 , one with w1 = v, w2 = 0 and the other with w1 = 0, w2 = v. Hence
W1 ∩ W2 = {0}, and W1 and W2 are complementary.

Examples We give some examples of complementary subspaces.

1. As in the previous example, let V = R2 , W1 = {(α, 0) | α ∈ R} and W2 = {(0, α) |
α ∈ R}. Then W1 and W2 are complementary subspaces.
2. Let V = R3 , W1 = {(α, 0, 0) | α ∈ R} and W2 = {(0, α, β) | α, β ∈ R}. Then W1
and W2 are complementary subspaces.
3. Let V = R2 , W1 = {(α, α) | α ∈ R} and W2 = {(−α, α) | α ∈ R}. Then W1 and
W2 are complementary subspaces.

5 Linear transformations
When you study sets, the notion of function is extremely important. There is little to
say about a single isolated set, while functions allow you to link different sets. Similarly,
in Linear Algebra, a single isolated vector space is not the end of the story. We have to
connect different vector spaces by functions. However, a function having little regard to
the vector space operations may be of little value.

5.1 Definition and examples

Often in mathematics, it is as important to study special classes of functions as it is
to study special classes of objects. Often these are functions which preserve certain
properties or structures. For example, continuous functions preserve which points are
close to which other points. In linear algebra, the functions which preserve the vector
space structure are called linear transformations.
Definition. Let U, V be two vector spaces over the same field K. A linear transforma-
tion or linear map T from U to V is a function T : U → V such that
(i) T (u1 + u2 ) = T (u1 ) + T (u2 ) for all u1 , u2 ∈ U ;
(ii) T (αu) = αT (u) for all α ∈ K and u ∈ U .
Notice that the two conditions for linearity are equivalent to a single condition

T (αu1 + βu2 ) = αT (u1 ) + βT (u2 ) for all u1 , u2 ∈ U, α, β ∈ K.

First let us state a couple of easy consequences of the definition:

Lemma 5.1. Let T : U → V be a linear map. Then
(i) T (0U ) = 0V ;
(ii) T (−u) = −T (u) for all u ∈ U .
Proof. For (i), the definition of linear map gives

T (0U ) = T (0U + 0U ) = T (0U ) + T (0U ),

and therefore T (0U ) = 0V . For (ii), just put α = −1 in the definition of linear map.

13
Examples Many familiar geometrical transformations, such as projections, rotations,
reflections and magnifications are linear maps, and the first three examples below are
of this kind. Note, however, that a nontrivial translation is not a linear map, because it
does not satisfy T (0U ) = 0V .

1. Let U = R3 , V = R2 and define T : U → V by T ((α, β, γ)) = (α, β). Then T is a

linear map. This type of map is known as a projection, because of the geometrical
interpretation.
z
v

 y



 - T(v)
0 X

XXX
XX x

(Note: In future we shall just write T (α, β, γ) instead of T ((α, β, γ)).)

2. Let U = V = R2 . We interpret v in R2 as a directed line vector from 0 to v

(see the examples in Section 2), and let T (v) be the vector obtained by rotating
v through an angle θ anti-clockwise about the origin.

T (v)



1 v
 θ
 

 

0

It is easy to see geometrically that T (u1 + u2 ) = T (u1 ) + T (u2 ) and T (αu) =

αT (u) (because everything is simply rotated about the origin), and so T is a
linear map. By considering the unit vectors, we have T (1, 0) = (cos θ, sin θ) and
T (0, 1) = (− sin θ, cos θ), and hence

T (α, β) = αT (1, 0) + βT (0, 1) = (α cos θ − β sin θ, α sin θ + β cos θ).

(Exercise: Show this directly.)

3. Let U = V = R2 again. Now let T (v) be the vector resulting from reflecting v
through a line through the origin that makes an angle θ/2 with the x-axis.

T (v)
7J

 
  J
  JJ v
 :





θ/2  0



This is again a linear map. We find that T (1, 0) = (cos θ, sin θ) and T (0, 1) =
(sin θ, − cos θ), and so

T (α, β) = αT (1, 0) + βT (0, 1) = (α cos θ + β sin θ, α sin θ − β cos θ).

4. Let U = V = R[x], the set of polynomials over R, and let T be differentiation; i.e.
T (p(x)) = p0 (x) for p ∈ R[x]. This is easily seen to be a linear map.

14
 5. Let U = K[x], the set of polynomials over K. Every α ∈ K gives rise to two
linear maps, shift Sα : U → U, Sα (f (x)) = f (x − α) and evaluation Eα : U →
K, Eα (f (x)) = f (α).

The next two examples seem dull but are important!

6. For any vector space V , we define the identity map IV : V → V by IV (v) = v for
all v ∈ V . This is a linear map.
7. For any vector spaces U, V over the field K, we define the zero map 0U,V : U → V
by 0U,V (u) = 0V for all u ∈ U . This is also a linear map.

One of the most useful properties of linear maps is that, if we know how a linear
map U → V acts on a basis of U , then we know how it acts on the whole of U .
Proposition 5.2 (Linear maps are uniquely determined by their action on a basis).
Let U, V be vector spaces over K, let u1 , . . . , un be a basis of U and let v1 , . . . , vn be
any sequence of n vectors in V . Then there is a unique linear map T : U → V with
T (ui ) = vi for 1 ≤ i ≤ n.
Proof. Let u ∈ U . Then, since u1 , . . . , un is a basis of U , by Proposition 3.2, there exist
uniquely determined α1 , . . . , αn ∈ K with u = α1 u1 + · · · + αn un . Hence, if T exists at
all, then we must have

T (u) = T (α1 u1 + · · · + αn un ) = α1 v1 + · · · + αn vn ,

and so T is uniquely determined.

On the other hand, it is routine to check that the map T : U → V defined by the
above equation is indeed a linear map, so T does exist and is unique.

5.2 Kernels and images

To any linear map U → V , we can associate a subspace of U and a subspace of V .
Definition. Let T : U → V be a linear map. The image of T , written as im(T ), is the
set of vectors v ∈ V such that v = T (u) for some u ∈ U .
Definition. The kernel of T , written as ker(T ), is the set of vectors u ∈ U such that
T (u) = 0V .
If you prefer:

im(T ) = {T (u) | u ∈ U }; ker(T ) = {u ∈ U | T (u) = 0V }.

Examples Let us consider the examples 1–7 above.

• In example 1, ker(T ) = {(0, 0, γ) | γ ∈ R}, and im(T ) = R2 .
• In example 2 and 3, ker(T ) = {0} and im(T ) = R2 .
• In example 4, ker(T ) is the set of all constant polynomials (i.e. those of degree 0),
and im(T ) = R[x].
• In example 5, ker(Sα ) = {0}, and im(Sα ) = K[x], while ker(Eα ) is the set of all
polynomials divisible by x − α, and im(Eα ) = K.
• In example 6, ker(IV ) = {0} and im(T ) = V .
• In example 7, ker(0U,V ) = U and im(0U,V ) = {0}.
Proposition 5.3. Let T : U → V be a linear map. Then
(i) im(T ) is a subspace of V ;
(ii) ker(T ) is a subspace of U .

15
Proof. For i, we must show that im(T ) is closed under addition and scalar multiplication.
Let v1 , v2 ∈ im(T ). Then v1 = T (u1 ), v2 = T (u2 ) for some u1 , u2 ∈ U . Then

v1 + v2 = T (u1 ) + T (u2 ) = T (u1 + u2 ) ∈ im(T )

and
αv1 = αT (u1 ) = T (αu1 ) ∈ im(T ),
so im(T ) is a subspace of V .
Let us now prove ii. Similarly,we must show that ker(T ) is closed under addition
and scalar multiplication. Let u1 , u2 ∈ ker(T ). Then

T (u1 + u2 ) = T (0U + 0U ) = T (0U ) = 0V

and
T (αu1 ) = αT (u1 ) = α0V = 0V ,
so u1 + u2 , αu1 ∈ ker(T ) and ker(T ) is a subspace of U .

5.3 Rank and nullity

The dimensions of the kernel and image of a linear map contain important information
about it, and are related to each other.

Definition. let T : U → V be a linear map.

(i) dim(im(T )) is called the rank of T ;

(ii) dim(ker(T )) is called the nullity of T .

Theorem 5.4 (The rank-nullity theorem). Let U, V be vector spaces over K with U
finite-dimensional, and let T : U → V be a linear map. Then

rank(T ) + nullity(T ) = dim(U ).

Proof. Since U is finite-dimensional and ker(T ) is a subspace of U , ker(T ) is finite-

dimensional. Let nullity(T ) = s and let e1 , . . . , es be a basis of ker(T ). By Theorem 3.6,
we can extend e1 , . . . , es to a basis e1 , . . . , es , f1 , . . . , fr of U . Then dim(U ) = s + r, so
to prove the theorem we have to prove that dim(im(T )) = r.
Clearly T (e1 ), . . . , T (es ), T (f1 ), . . . , T (fr ) span im(T ), and since

T (e1 ) = · · · = T (es ) = 0V

this implies that T (f1 ), . . . , T (fr ) span im(T ). We shall show that T (f1 ), . . . , T (fr ) are
linearly independent.
Suppose that, for some scalars αi , we have

α1 T (f1 ) + · · · + αr T (fr ) = 0V .

Then T (α1 f1 + · · · + αr fr ) = 0V , so α1 f1 + · · · + αr fr ∈ ker(T ). But e1 , . . . , es is a basis

of ker(T ), so there exist scalars βi with

α1 f1 + · · · + αr fr = β1 e1 + · · · + βs es =⇒ α1 f1 + · · · + αr fr − β1 e1 − · · · − βs es = 0U .

But we know that e1 , . . . , es , f1 , . . . , fr form a basis of U , so they are linearly independent,

and hence
α1 = · · · = αr = β1 = · · · = βs = 0,
and we have proved that T (f1 ), . . . , T (fr ) are linearly independent.
Since T (f1 ), . . . , T (fr ) both span im(T ) and are linearly independent, they form a
basis of im(U ), and hence dim(im(T )) = r, which completes the proof.

16
Examples Once again, we consider examples 1–7 above. Since we only want to deal
with finite-dimensional spaces, we restrict to an (n + 1)-dimensional space K[x]≤n in
examples 4 and 5, that is, we consider T : R[x]≤n → R[x]≤n , Sα : K[x]≤n → K[x]≤n ,
and Eα : K[x]≤n → K correspondingly. Let n = dim(U ) = dim(V ) in 5 and 6.

Example rank(T ) nullity(T ) dim(U )

1 2 1 3
2 2 0 2
3 2 0 2
4 n 1 n+1
5 Sα n+1 0 n+1
4 Eα 1 n n+1
6 n 0 n
7 0 n n

Corollary 5.5. Let T : U → V be a linear map, and suppose that dim(U ) = dim(V ) =
n. Then the following properties of T are equivalent:
(i) T is surjective;
(ii) rank(T ) = n;
(iii) nullity(T ) = 0;
(iv) T is injective;
(v) T is bijective;
Proof. That T is surjective means precisely that im(T ) = V , so (i) ⇒ (ii). But if
rank(T ) = n, then dim(im(T )) = dim(V ) so (by Corollary 3.9) a basis of im(T ) is a
basis of V , and hence im(T ) = V . Thus (ii) ⇔ (i).
That (ii) ⇔ (iii) follows directly from Theorem 5.4.
Now nullity(T ) = 0 means that ker(T ) = {0} so clearly (iv) ⇒ (iii). On the other
hand, if ker(T ) = {0} and T (u1 ) = T (u2 ) then T (u1 − u2 ) = 0, so u1 − u2 ∈ ker(T ) =
{0}, which implies u1 = u2 and T is injective. Thus (iii) ⇔ (iv). (In fact, this argument
shows that (iii) ⇔ (iv) is true for any linear map T .)
Finally, (v) is equivalent to (i) and (iv), which we have shown are equivalent to each
other.
Definition. If the conditions in the above corollary are met, then T is called a non-
singular linear map. Otherwise, T is called singular. Notice that the terms singular and
non-singular are only used for linear maps T : U → V for which U and V have the same
dimension.

5.4 Operations on linear maps

We can define the operations of addition, scalar multiplication and composition on linear
maps.
Let T1 : U → V and T2 : U → V be two linear maps, and let α ∈ K be a scalar.
Definition (Addition of linear maps). We define a map

T1 + T2 : U → V

by the rule (T1 + T2 )(u) = T1 (u) + T2 (u) for u ∈ U .

Definition (Scalar multiplication of linear maps). We define a map

αT1 : U → V

by the rule (αT1 )(u) = αT1 (u) for u ∈ U .

Now let T1 : U → V and T2 : V → W be two linear maps.

17
Definition (Composition of linear maps). We define a map

T2 T1 : U → W

by (T2 T1 )(u) = T2 (T1 (u)) for u ∈ U .

In particular, we define T 2 = T T and T i+1 = T i T for i > 2.

It is routine to check that T1 + T2 , αT1 and T2 T1 are themselves all linear maps (but
you should do it!).
Furthermore, for fixed vector spaces U and V over K, the operations of addition
and scalar multiplication on the set HomK (U, V ) of all linear maps from U to V makes
HomK (U, V ) into a vector space over K.
Given a vector space U over a field K, the vector space U ∗ = HomK (U, K) plays a
special role. It is often called the dual space or the space of covectors of U . One can
think of coordinates as elements of U ∗ . Indeed, let ei be a basis of U . Every x ∈ U can
be uniquely written as
x = α1 e1 + . . . αn en , αi ∈ K.

The elements αi depend on x as well as on a choice of the basis, so for each i one can
write the coordinate function

ei : U → K, ei (x) = αi .

It is routine to check that ei is a linear map, and indeed the functions ei form a basis
of the dual space U ∗ .

6 Matrices
The material in this section will be familiar to many of you already, at least when K is
the field of real numbers.

Definition. Let K be a field and m, n ∈ N. An m × n matrix A over K is an m × n

rectangular array of numbers (i.e. scalars) in K. The entry in row i and column j is
often written αij . (We use the corresponding Greek letter.) We write A = (αij ) to make
things clear.

For example, we could take

 
2 −1 −π 0
K = R, m = 3, n = 4, A = (αij ) =  3 −3/2 0 6 ,
−1.23 0 1010 0

and then α13 = −π, α33 = 1010 , α34 = 0, and so on.

Having defined what matrices are, we want to be able add them, multiply them by
scalars, and multiply them by each other. You probably already know how to do this,
but we will define these operations anyway.

Definition (Addition of matrices). Let A = (αij ) and B = (βij ) be two m × n matrices

over K. We define A + B to be the m × n matrix C = (γij ), where γij = αij + βij for
all i, j.

Example.
     
1 3 −2 −3 −1 −0
+ = .
0 2 1 −4 1 −2

Definition (Scalar multiplication of matrices). Let A = (αij ) be an m × n matrix over

K and let β ∈ K be a scalar. We define the scalar multiple βA to be the m × n matrix
C = (γij ), where γij = βαij for all i, j.

18
Definition (Multiplication of matrices). Let A = (αij ) be an l × m matrix over K and
let B = (βij ) be an m × n matrix over K. The product AB is an l × n matrix C = (γij )
where, for 1 ≤ i ≤ l and 1 ≤ j ≤ n,
m
X
γij = αik βkj = αi1 β1j + αi2 β2j + · · · + αim βmj .
k=1

It is essential that the number m of columns of A is equal to the number of rows of B;

otherwise AB makes no sense.

If you are familiar with scalar products of vectors, note also that γij is the scalar
product of the ith row of A with the jth column of B.

Example. Let  
  2 6
2 3 4
A= , B = 3 2 .
1 6 2
1 9
Then    
2×2+3×3+4×1 2×6+3×2+4×9 17 54
AB = = ,
1×2+6×3+2×1 1×6+6×2+2×9 22 36
 
10 42 20
BA =  8 21 16 .
11 57 22
 
2 3 1
Let C = . Then AC and CA are not defined.
6 2 9
   
1 2 4 15 8
Let D = . Then AD is not defined, but DA = .
0 1 1 6 2
Proposition 6.1. Matrices satisfy the following laws whenever the sums and products
involved are defined:

(i) A + B = B + A;

(ii) (A + B)C = AC + BC;

(iii) C(A + B) = CA + CB;

(iv) (λA)B = λ(AB) = A(λB);

(v) (AB)C = A(BC).

Proof. These are all routine checks that the entries of the left-hand sides are equal to
the corresponding entries on the right-hand side. Let us do (v) as an example.
Let A, B and C be l×m, m×n
Pm and n×p matrices, respectively. Then AB = D = (δij )
× n matrix with δij = s=1 αis βsj , and BC = E = (εij ) is an m × p matrix with
is an lP
n
εij = t=1 βit γtj . Then (AB)C = DC and A(BC) = AE are both l × p matrices, and
we have to show that their coefficients are equal. The (i, j)-coefficient of DC is
n
X n X
X m m
X n
X m
X
δit γtj = ( αis βst )γtj = αis ( βst γtj ) = αis εsj
t=1 t=1 s=1 s=1 t=1 s=1

which is the (i, j)-coefficient of AE. Hence (AB)C = A(BC).

There are some useful matrices to which we give names.

Definition. The m × n zero matrix 0mn over any field K has all of its entries equal to
0.

Definition. The n × n identity matrix In = (αij ) over any field K has αii = 1 for
1 ≤ i ≤ n, but αij = 0 when i 6= j.

19
Example.  
  1 0 0
1 0
I1 = (1), I2 = , I3 = 0 1 0 .
0 1
0 0 1
Note that In A = A for any n × m matrix A and AIn = A for any m × n matrix A.

The set of all m × n matrices over K will be denoted by K m,n . Note that K m,n is
itself a vector space over K using the operations of addition and scalar multiplication
defined above, and it has dimension mn. (This should be obvious – is it?)
A 1 × n matrix is called a row vector. We will regard K 1,n as being the same as K n .
A n × 1 matrix is called a column vector. We will denote the the space K n,1 of all
column vectors by K n,1 . In matrix calculations, we will use K n,1 more often than K n .

7 Linear transformations and matrices

We shall see in this section that, for fixed choice of bases, there is a very natural one-one
correspondence between linear maps and matrices, such that the operations on linear
maps and matrices defined in Chapters 5 and 6 also correspond to each other. This
is perhaps the most important idea in linear algebra, because it enables us to deduce
properties of matrices from those of linear maps, and vice-versa. It also explains why
we multiply matrices in the way we do.

7.1 Setting up the correspondence

Let T : U → V be a linear map, where dim(U ) = n, dim(V ) = m. Suppose that we are
given a basis e1 , . . . , en of U and a basis f1 , . . . , fm of V .
Now, for 1 ≤ j ≤ n, the vector T (ej ) lies in V , so T (ej ) can be written uniquely as
a linear combination of f1 , . . . , fm . Let

T (e1 ) = α11 f1 + α21 f2 + · · · + αm1 fm

T (e2 ) = α12 f1 + α22 f2 + · · · + αm2 fm
···
T (en ) = α1n f1 + α2n f2 + · · · + αmn fm

where the coefficients αij ∈ K (for 1 ≤ i ≤ m, 1 ≤ j ≤ n) are uniquely determined.

Putting it more compactly, we define scalars αij by
m
X
T (ej ) = αij fi for 1 ≤ j ≤ n.
i=1

The coefficients αij form an m × n matrix

 
α11 α12 · · · α1n
 α21 α22 · · · α2n 
A= .
 
.. .. . 
 .. . . .. 
αm1 αm2 · · · αmn

over K. Then A is called the matrix of the linear map T with respect to the chosen
bases of U and V . In general, different choice of bases gives different matrices. We shall
address this issue later in the course, in Section 11.
Notice the role of individual columns in A The jth column of A consists of the
coordinates of T (ej ) with respect to the basis vf1 , . . . , fm of V .

Theorem 7.1. Let U, V be vector spaces over K of dimensions n, m, respectively. Then,

for a given choice of bases of U and V , there is a one-one correspondence between the
set HomK (U, V ) of linear maps U → V and the set K m,n of m × n matrices over K.

20
Proof. As we saw above, any linear map T : U → V determines an m × n matrix A over
K.
Conversely, let A = (αij ) be an m × n matrix Pmover K. Then, by Proposition 5.2,
there is just one linear T : U → V with T (ej ) = i=1 αij fi for 1 ≤ j ≤ n, so we have a
one-one correspondence.

Examples Once again, we consider our examples from Section 5.

1. T : R3 → R2 , T (α, β, γ) = (α, β). Usually, we choose the standard bases of K m

and K n , which in this case are e1 = (1, 0, 0), e2 = (0, 1, 0), e3 = (0, 0, 1) and
f1 = (1, 0), f2 = (0, 1). We have T (e1 ) = f1 , T (e2 ) = f2 ), T (e3 ) = 0, and the
matrix is  
1 0 0
.
0 1 0

But suppose we chose different bases, say e1 = (1, 1, 1), e2 = (0, 1, 1), e3 = (1, 0, 1),
and f1 = (0, 1), f2 = (1, 0). Then we have T (e1 ) = (1, 1) = f1 +f2 , T (e2 ) = (0, 1) =
f1 , T (e3 ) = (1, 0) = f2 , and the matrix is
 
1 1 0
.
1 0 1

2. T : R2 → R2 , T is a rotation through θ anti-clockwise about the origin. We saw

that T (1, 0) = (cos θ, sin θ) and T (0, 1) = (− sin θ, cos θ), so the matrix using the
standard bases is  
cos θ − sin θ
.
sin θ cos θ

3. T : R2 → R2 , T is a reflection through the line through the origin making an

angle θ/2 with the x-axis. We saw that T (1, 0) = (cos θ, sin θ) and T (0, 1) =
(sin θ, − cos θ), so the matrix using the standard bases is
 
cos θ sin θ
.
sin θ − cos θ

4. This time we take the differentiation map T from R[x]≤n to R[x]≤n−1 . Then, with
respect to the bases 1, x, x2 , . . . , xn and 1, x, x2 , . . . , xn−1 of R[x]≤n and R[x]≤n−1 ,
respectively, the matrix of T is
 
0 1 0 0 ··· 0 0
0 0 2 0 · · · 0 0
 
0 0 0 3 · · · 0 0
..  .
 
 .. .. .. .. . . ..
. . . .
 . . .
0 0 0 0 · · · n − 1 0 
0 0 0 0 ··· 0 n

5. Let Sα : K[x]≤n → K[x]≤n be the shift. With respect to the basis 1, x, x2 , . . . , xn

of K[x]≤n , we calculate Sα (xn ) = (x−α)n . The binomial formula gives the matrix
of Sα ,
1 −α α2 · · · (−1)n αn
 
0 1 −2α · · · (−1)n−1 nαn−1 
1 · · · (−1)n−2 n(n−1)
 
0 0
2 αn−2 
.
 
. . .. . . ..
 .. .. . . . 
 
0 0 0 ··· −nα 
0 0 0 ··· 1
In the same basis of K[x]≤n and the basis 1 of K, Eα (xn ) = αn . The matrix of
Eα is
1 α α2 · · · αn−1 αn .

21
 6. T : V → V is the identity map. Notice that U = V in this example. Provided that
we choose the same basis for U and V , then the matrix of T is the n × n identity
matrix In . We shall be considering the situation where we use different bases for
the domain and range of the identity map in Section 11.

7. T : U → V is the zero map. The matrix of T is the m × n zero matrix 0mn ,

regardless of what bases we choose. (The coordinates of the zero vector are all
zero in any basis.)

We now connect how a linear transformation acts on elements of a vector space to

how its matrix acts on their coordinates.
For the given basis e1 , . . . , en of U and a vector u = λ1 e1 + · · · + λn en ∈ U , let u
denote the column vector  
λ1
 λ2 
u =  .  ∈ K n,1 ,
 
 .. 
λn
whose entries are the coordinates of u with respect to that basis. Similarly, for the given
basis f1 , . . . , fm of V and a vector v = µ1 f1 + · · · + µm fm ∈ V , let v denote the column
vector  
µ1
 µ2 
v  .  ∈ K m,1
 
 .. 
µm
whose entries are the coordinates of v with respect to that basis.

Proposition 7.2. Let T : U → V be a linear map. Let the matrix A = (αij ) represent
T with respect to chosen bases of U and V , and let u and vv be the column vectors of
coordinates of two vectors u ∈ U and v ∈ V , again with respect to the same bases. Then
T (u) = v if and only if Au = v.

Proof. We have

Xn n
X
T (u) = T ( λj ej ) = λj T (ej )
j=1 j=1
n
X Xm m X
X n m
X
= λj ( αij fi ) = ( αij λj )fi = µi fi ,
j=1 i=1 i=1 j=1 i=1

Pn
where µi = j=1 αij λj is the entry in the ith row of the column vector Au. This proves
the result.

What is this theorem really telling us? One way of looking at it is this. Choosing
a basis for U gives every vector in U a unique set of coordinates. Choosing a basis
for V gives every vector in V a unique set of coordinates. Now applying the linear
transformation T to u ∈ U is “the same” as multiplying its column vector of coordinates
by the matrix representing T , as long as we interpret the resulting column vector as
coordinates in V with respect to our chosen basis.
Of course, choosing different bases will change the matrix A representing T , and will
change the coordinates of both u and v. But it will change all of these quantities in
exactly the right way that the theorem still holds.

7.2 The correspondence between operations on linear maps and

matrices
Let U, V and W be vector spaces over the same field K, let dim(U ) = n, dim(V ) = m,
dim(W ) = l, and choose fixed bases e1 , . . . , en of U and f1 , . . . , fm of V , and g1 , . . . , gl

22
of W . All matrices of linear maps between these spaces will be written with respect to
these bases.
We have defined addition and scalar multiplication of linear maps, and we have
defined addition and scalar multiplication of matrices. We have also defined a way to
associate a matrix to a linear map. It turns out that all these operations behave together
in the way we might hope.

Proposition 7.3. 1. Let T1 , T2 : U → V be linear maps with m × n matrices A, B

respectively. Then the matrix of T1 + T2 is A + B.

2. Let T : U → V be a linear map with m × n matrices A and let λ ∈ K be a scalar.

Then the matrix of λT is λA.

Proof. These are both straightforward to check, using the definitions, as long as you
keep your wits about you. Checking them is a useful exercise, and you should do it.

Note that the above two properties imply that the natural correspondence between
linear maps and matrices is actually itself a linear map from HomK (U, V ) to K m,n .
Composition of linear maps corresponds to matrix multiplication. This time the
correspondence is less obvious, and we state it as a theorem.

Theorem 7.4. Let T1 : V → W be a linear map with l × m matrix A = (αij ) and let
T2 : U → V be a linear map with m × n matrix B = (βij ). Then the matrix of the
composite map T1 T2 : U → W is AB.

Proof. Let AB Pbe the l×n matrix (γij ). Then by the definition of matrix multiplication,
m
we have γik = j=1 αij βjk for 1 ≤ i ≤ l, 1 ≤ k ≤ n.
Let us calculate the matrix of T1 T2 . We have

Xm m
X m
X l
X
T1 T2 (ek ) = T1 ( βjk fj ) = βjk T1 (fj ) = βjk αij gi
j=1 j=1 j=1 i=1
l X
X m l
X
= ( αij βjk )gi = γik gi ,
i=1 j=1 i=1

so the matrix of T1 T2 is (γik ) = AB as claimed.

Examples Let us look at some examples of matrices corresponding to the composition

of two linear maps.

1. Let Rθ : R2 → R2 be a rotation through an angle θ anti-clockwise  about the origin.


cos θ − sin θ
We have seen that the matrix of Rθ (using the standard basis) is .
sin θ cos θ
Now clearly Rθ followed by Rφ is equal to Rθ+φ . We can check the corresponding
result for matrices:
  
cos φ − sin φ cos θ − sin θ
sin φ cos φ sin θ cos θ
 
cos φ cos θ − sin φ sin θ − cos φ sin θ − sin φ cos θ
=
sin φ cos θ + cos φ sin θ − sin φ sin θ + cos φ cos θ
 
cos(φ + θ) − sin(φ + θ)
= .
sin(φ + θ) cos(φ + θ)

Note that in this case T1 T2 = T2 T1 . This actually gives an alternative way of

deriving the addition formulae for sin and cos.

2. Let Rθ be as in Example 1, and let Mθ : R2 → R2 be a reflection through a line

through the
 origin at an  angle θ/2 to the x-axis. We have seen that the matrix
cos θ sin θ
of Mθ is . What is the effect of doing first Rθ and then Mφ ? In
sin θ − cos θ

23
 this case, it might be easier (for some people) to work it out using the matrix
multiplication! We have
  
cos φ sin φ cos θ − sin θ
sin φ − cos φ sin θ cos θ
 
cos φ cos θ + sin φ sin θ − cos φ sin θ + sin φ cos θ
=
sin φ cos θ − cos φ sin θ − sin φ sin θ − cos φ cos θ
 
cos(φ − θ) sin(φ − θ)
= ,
sin(φ − θ) − cos(φ − θ)
which is the matrix of Mφ−θ .
We get a different result if we do first Mφ and then Rθ . What do we get then?

7.3 Linear equations and the inverse image problem

The study and solution of systems of simultaneous linear equations is the main motiva-
tion behind the development of the theory of linear algebra and of matrix operations.
Let us consider a system of m equations in n unknowns x1 , x2 . . . xn , where m, n ≥ 1.


 α11 x1 + α12 x2 + · · · + α1n xn = β1
 α21 x1 + α22 x2 + · · · + α2n xn = β2

.. (1)


 .
αm1 x1 + αm2 x2 + · · · + αmn xn = βm


All coefficients αij and βi belong to K. Solving this system means finding all collections
x1 , x2 . . . xn ∈ K such that the equations (1) hold.
Let A = (αij ) ∈ K m,n be the m × n matrix of coefficients. The crucial step is to
introduce the column vectors
   
x1 β1
 x2   β2 
x =  .  ∈ K n,1 and b =  .  ∈ K m,1 .
   
 ..   .. 
xn βm
This allows us to rewrite system (1) as a single equation
Ax = b (2)
where the coefficient A is a matrix, the right hand side b is a vector in K m,1 and the
unknown x is a vector K n,1 .
Using the notation of linear maps, we have just reduced solving a system of linear
equations to the inverse image problem. That is, given a linear map T : U → V , and a
fixed vector v ∈ V , find all u ∈ U such that T (u) = v.
In fact, these two problems are equivalent! In the opposite direction, let us first
forget all about A, x and b, and suppose that we are given an inverse image problem
to solve. Then we choose bases in U and V and denote a matrix of T in these bases
by A, the row vector of coordinates u by x and the row vector of coordinates v by b.
Proposition 7.2 says that T (u) = v if and only if Ax = b. This reduces the inverse
image problem to solving a system of linear equations.
Let us make several easy observations about the inverse image problem.
The case when v = 0 or, equivalently when βi = 0 for 1 ≤ i ≤ m, is called the
homogeneous case. Here the set of solutions is {u ∈ U | T (u) = 0}, which is precisely
the kernel ker(T ) of T . The corresponding set of column vectors x ∈ K n,1 with Ax = 0
is called the nullspace of the matrix A. These column vectors are the coordinates of the
vectors in the kernel of T , with respect to our chosen basis for U . So the nullity of A is
the dimension of its nullspace.
In general, it is easy to see (and you should work out the details) that if x is one
solution to a system of equations, then the complete set of solutions is equal to
x + nullspace(A) = {x + y | y ∈ nullspace(A)}.

24
It is possible that there are no solutions at all; this occurs when v 6∈ im(T ). If there are
solutions, then there is a unique solution precisely when ker(T ) = {0}, or equivalently
when nullspace(A) = {0}. If the field K is infinite and there are solutions but ker(T ) 6=
{0}, then there are infinitely many solutions.
Now we would like to develop methods for solving the inverse image problem.

8 Elementary operations and the rank of a matrix

8.1 Gauss transformations
There are two standard high school methods for solving linear systems: the substitution
method (where you express variables in terms of the other variables and substitute the
result in the remaining equations) and the elimination method (sometimes called the
Gauss method ). The latter is usually more effective, so we would like to contemplate its
nature. Let us recall how it is done.

Examples Here are some examples of solving systems of linear equations by the elim-
ination method.
1.
2x + y = 1 (1)
4x + 2y = 1 (2)
Replacing (2) by (2)−2×(1) gives 0 = −1. This means that there are no solutions.
2.
2x + y = 1 (1)
4x + y = 1 (2)
Replacing (2) by (2) − (1) gives 2x = 0, and so x = 0. Replacing (1) by (1) − 2 ×
(new 2) gives y = 1. Thus, (0, 1) is a unique solution.
3.
2x + y = 1 (1)
4x + 2y = 2 (2)

This time (2) − 2 × (1) gives 0 = 0, so (2) is redundant.

After reduction, there is no equation with leading term y, which means that y can
take on any value, say y = α. The first equation determines x in terms of y, giving
x = (1 − α)/2. So the general solution is (x, y) = ((1 − α)/2, α), meaning that for
each α ∈ R we find one solution (x, y). There are infinitely many solutions.
Notice also that one solution is (x, y) = (1/2, 0), and the general solution can be
written as (x, y) = (1/2, 0) + α(−1/2, 1), where α(−1/2, 1) is the solution of the
corresponding homogeneous system 2x + y = 0; 4x + 2y = 0.
4.
x + y + z = 1 (1)
x + z = 2 (2)
x − y + z = 3 (3)
3x + y + 3z = 5 (4)
Now replacing (2) by (2)−(1) and then multiplying by −1 gives y = −1. Replacing
(3) by (3)−(1) gives −2y = 2, and replacing (4) by (4)−3×(1) also gives −2y = 2.
So (3) and (4) both then reduce to 0 = 0, and they are redundant.
z does not occur as a leading term, so it can take any value, say α, and then (2)
gives y = −1 and (1) gives x = 1 − y − z = 2 − α, so the general solution is
(x, y, z) = (2 − α, −1, α) = (2, −1, 0) + α(−1, 0, 1).

25
8.2 Elementary row operations
Many types of calculations with matrices can be carried out in a computationally efficient
manner by the use of certain types of operations on rows and columns. We shall see
a little later that these are really the same as the operations used in solving sets of
simultaneous linear equations.
Let A be an m × n matrix over K with rows r1 , r2 , . . . , rm ∈ K 1,n . The three types
of elementary row operations on A are defined as follows.

(R1) For some i 6= j, add a multiple of rj to ri .

   
3 1 9 3 1 9
r3 →r3 −3r1
Example: 4 6 7 −−−−−−−→  4 6 7 
2 5 8 −7 2 −19

(R2) Interchange two rows

(R3) Multiply a row by a non-zero scalar.
   
2 0 5 2 0 5
r2 →4r2
Example: 1 −2 3 −− −−−→ 4 −8 12
5 1 2 5 1 2

8.3 The augmented matrix

We would like to make the process of solving a system of linear equations more me-
chanical by forgetting about the variable names w, x, y, z, etc. and doing the whole
operation as a matrix calculation. For this, we use the augmented matrix of the system
of equations, which is constructed by “glueing” an extra column on the right-hand side
of the matrix representing the linear transformation, as follows. For the system Ax = β
of m equations in n unknowns, where A is the m × n matrix (αij ) is defined to be the
m × (n + 1) matrix  
α11 α12 · · · α1n β1
 α21 α22 · · · α2n β2 
A= . ..  .
 
.. . . ..
 .. . . . . 
αm1 αm2 · · · αmn βm
The vertical line in the matrix is put there just to remind us that the rightmost column
is different from the others, and arises from the constants on the right hand side of the
equations.
Let us look at the following system of linear equations over R: suppose that we want
to find all w, x, y, z ∈ R satisfying the equations.


 2w − x + 4y − z = 1
w + 2x + y + z = 2


 w − 3x + 3y − 2z = −1
−3w − x − 5y = −3 .


Elementary row operations on A are precisely Gauss transformations of the correspond-

ing linear system. Thus, the solution can be carried out mechanically as follows:

Matrix Operation(s)
 
2 −1 4 −1 1
 1 2 1 1 2

 1 −3 3 −2 −1 
 r1 → r1 /2
−3 −1 −5 0 −3
 
1 −1/2 2 −1/2 1/2
 1 2 1 1 2
  r2 → r2 − r1 , r3 → r3 − r1 , r4 → r4 + 3r1
 1 −3 3 −2 −1 
−3 −1 −5 0 −3
26
 Matrix Operation(s)

 
1 −1/2 2 −1/2 1/2
 0 5/2 −1 3/2 3/2 
 
 0 −5/2 1 −3/2 −3/2  r3 → r3 + r2 , r4 → r4 + r2
0 −5/2 1 −3/2 −3/2
 
1 −1/2 2 −1/2 1/2
 0 5/2 −1 3/2 3/2 

0
 r2 → 2r2 /5
0 0 0 0
0 0 0 0 0
 
1 −1/2 2 −1/2 1/2
0 1 −2/5 3/5 3/5 

0
 r1 → r1 + r2 /2
0 0 0 0
0 0 0 0 0
 
1 0 9/5 −1/5 4/5
0
 1 −2/5 3/5 3/5 

0 0 0 0 0
0 0 0 0 0

The original system has been transformed to the following equivalent system, that
is, both systems have the same solutions.

w + 9y/5 − z/5 = 4/5
x − 2y/5 + 3z/5 = 3/5
In a solution to the latter system, variables y and z can take arbitrary values in
R; say y = α, z = β. Then the equations tell us that w = −9α/5 + β/5 + 4/5 and
x = 2α/5 − 3β/5 + 3/5 (be careful to get the signs right!), and so the complete set of
solutions is
(w, x, y, z) = (−9α/5 + β/5 + 4/5, 2α/5 − 3β/5 + 3/5, α, β)
= (4/5, 3/5, 0, 0) + α(−9/5, 2/5, 1, 0) + β(1/5, −3/5, 0, 1).

8.4 Row reducing a matrix

Let A = (αij ) be an m × n matrix over the field K. For the ith row, let c(i) denote the
position of the first (leftmost) non-zero entry in that row. In other words, αi,c(i) 6= 0
while αij = 0 for all j < c(i). It will make things a little easier to write if we use the
convention that c(i) = ∞ if the ith row is entirely zero.
We will describe a procedure, analogous to solving systems of linear equations by
elimination, which starts with a matrix, performs certain row operations, and finishes
with a new matrix in a special form. After applying this procedure, the resulting matrix
A = (αij ) will have the following properties.
(i) All zero rows are below all non-zero rows.
(ii) Let r1 , . . . , rs be the non-zero rows. Then each ri with 1 ≤ i ≤ s has 1 as its first
non-zero entry. In other words, αi,c(i) = 1 for all i ≤ s.
(iii) The first non-zero entry of each row is strictly to the right of the first non-zero
entry of the row above: that is, c(1) < c(2) < · · · < c(s).
(iv) If row i is non-zero, then all entries below the first non-zero entry of row i are zero:
αk,c(i) = 0 for all k > i.
Definition. A matrix satisfying properties (i)–(iv) above is said to be in upper echelon
form.

27
Example. The matrix we came to at the end of the previous example was in upper
echelon form.
There is a stronger version of the last property:
(v) If row i is non-zero, then all entries both above and below the first non-zero entry
of row i are zero: αk,c(i) = 0 for all k 6= i.
Definition. A matrix satisfying properties (i)–(v) is said to be in row reduced form.
An upper echelon form of a matrix will be used later to calculate the rank of a matrix.
The row reduced form (the use of the definite article is intended: this form is, indeed,
unique, though we shall not prove this) is used to solve systems of linear equations. In
this light, the following theorem says that every system of linear equations can be solved
by the Gauss (Elimination) method.
Theorem 8.1. Every matrix can be brought to row reduced form by elementary row
transformations.
Proof. We describe an algorithm to achieve this. For a formal proof, we have to show:
(i) after termination the resulting matrix has a row reduced form;
(ii) the algorithm terminates after finitely many steps.
Both of these statements are clear from the nature of the algorithm. Make sure that
you understand why they are clear!
At any stage in the procedure we will be looking at the entry αij in a particular
position (i, j) of the matrix. We will call (i, j) the pivot position, and αij the pivot
entry. We start with (i, j) = (1, 1) and proceed as follows.
1. If αij and all entries below it in its columns are zero (i.e. if αkj = 0 for all k ≥ i), then
move the pivot one place to the right, to (i, j + 1) and repeat Step 1, or terminate if
j = n.
2. If αij = 0 but αkj 6= 0 for some k > i then apply row operation (R2) to interchange
ri and rk .
3. At this stage αij 6= 0. If αij 6= 1, then apply row operation (R3) to multiply ri by
−1
αij .
4. At this stage αij = 1. If, for any k 6= i, αkj 6= 0, then apply row operation (R1), and
subtract αkj times ri from rk .
5. At this stage, αkj = 0 for all k 6= i. If i = m or j = n then terminate. Otherwise,
move the pivot diagonally down to the right to (i + 1, j + 1), and go back to Step 1.

If one needs only an upper echelon form, this can done faster by replacing steps 4
and 5 with weaker and faster steps as follows.
4a. At this stage αij = 1. If, for any k > i, αkj 6= 0, then apply (R1), and subtract αkj
times ri from rk .
5a. At this stage, αkj = 0 for all k > i. If i = m or j = n then terminate. Otherwise,
move the pivot diagonally down to the right to (i + 1, j + 1), and go back to Step 1.
In the example below, we find an upper echelon form of a matrix by applying the
faster algorithm. The number in the ‘Step’ column refers to the number of the step
applied in the description of the procedure above.
 
0 0 1 2 1
2 4 2 −4 2
Example Let A =  .
3 6 3 −6 3
1 2 3 3 3

28
 Matrix Pivot Step Operation
 
0 0 1 2 1
2 4 2 −4 2
  (1, 1) 2 r1 ↔ r2
3 6 3 −6 3
1 2 3 3 3
 
2 4 2 −4 2
0 0 1 2 1
  (1, 1) 3 r1 → r1 /2
3 6 3 −6 3
1 2 3 3 3
 
1 2 1 −2 1
0 0 1 2 1 r3 → r3 − 3r1
  (1, 1) 4
3 6 3 −6 3 r4 → r4 − r1
1 2 3 3 3
 
1 2 1 −2 1
0 0 1 2 1

0
 (1, 1) → (2, 2) → (2, 3) 5, 1
0 0 0 0
0 0 2 5 2
 
1 2 1 −2 1
0 0 1 2 1

0
 (2, 3) 4 r4 → r4 − 2r2
0 0 0 0
0 0 2 5 2
 
1 2 1 −2 1
0 0 1 2 1

0
 (2, 3) → (3, 4) 5, 2 r3 ↔ r4
0 0 0 0
0 0 0 1 0
 
1 2 1 −2 1
0 0 1 2 1

0
 (3, 4) → (4, 5) → stop 5, 1
0 0 1 0
0 0 0 0 0

This matrix is now in upper echelon form.

8.5 Elementary column operations

In analogy to elementary row operations, one can introduce elementary column oper-
ations. Let A be an m × n matrix over K with columns c1 , c2 , . . . , cn as above. The
three types of elementary column operations on A are defined as follows.

(C1) For some i 6= j, add a multiple of cj to ci .

   
3 1 9 3 1 0
c3 →c3 −3c1
Example: 4 6 7 −− −−−−−→ 4 6 −5
2 5 8 2 5 2

(C2) Interchange two columns.

(C3) Multiply a column by a non-zero scalar.

   
2 0 5 2 0 5
c2 →4c2
Example: 1 −2 3 −− −−−→ 1 −8 3
5 1 2 5 4 2

29
 Elementary column operations change a linear system and cannot be applied to solve
a system of linear equations. However, they are useful for reducing a matrix to a very
nice form.
Theorem 8.2. By applying elementary row and column operations, a matrix can be
brought into the block form  
Is 0s,n−s
,
0m−s,s 0m−s,n−s
where, as in Section 6, Is denotes the s×s identity matrix, and 0kl the k ×l zero matrix.

Proof. First, use elementary row operations to reduce A to row reduced form.
Now all αi,c(i) = 1. We can use these leading entries in each row to make all the
other entries zero: for each αij 6= 0 with j 6= c(i), replace cj with cj − αij cc(i) .
Finally the only nonzero entries of our matrix are αi,c(i) = 1. Now for each number i
starting from i = 1, exchange ci and cc(i) , putting all the zero columns at the right-hand
side.

Definition. The matrix in Theorem 8.2 is said to be in row and column reduced form,
which is also called Smith normal form.
Let us look at an example of the second stage of procedure, that is, after reducing
the matrix to the row reduced form.

Matrix Operation
 
1 2 0 0 1
0
 0 1 0 2 c2 → c2 − 2c1
0 0 0 1 3 c5 → c5 − c1
0 0 0 0 0
 
1 0 0 0 0
0
 0 1 0 2 c2 ↔ c3
0 0 0 1 3 c5 → c5 − 3c4
0 0 0 0 0
 
1 0 0 0 0
0
 1 0 0 2 c3 ↔ c4
0 0 0 1 0 c5 → c5 − 2c2
0 0 0 0 0
 
1 0 0 0 0
0 1 0 0 0
 
0 0 1 0 0
0 0 0 0 0

Now we would like to discuss the number s that appears in Theorem 8.2, that is, the
number of non-zero entries in the Smith normal form. Does the initial matrix uniquely
determine this number? Although we have an algorithm for reducing a matrix to Smith
normal form, there will be other sequences of row and column operations which also put
the matrix into Smith normal form. Could we maybe end up with a different number
of non-zero entries depending on the row and column operations used?

8.6 The rank of a matrix

Let T : U → V be a linear map, where dim(U ) = n, dim(V ) = m. Let e1 , . . . , en be a
basis of U and let f1 , . . . , fm be a basis of V .
Recall from Section 5.3 that rank(T ) = dim(im(T )).
Now im(T ) is spanned by the vectors T (e1 ), . . . , T (en ), and by Theorem 3.5, some
subset of these vectors forms a basis of im(T ). By definition of basis, this subset has

30
size dim(im(T )) = rank(T ), and by Corollary 3.9 no larger subset of T (e1 ), . . . , T (en )
can be linearly independent. We have therefore proved:

Lemma 8.3. Let T : U → V be a linear transformation, and let e1 , . . . , en be a basis of

U . Then the rank of T is equal to the size of the largest linearly independent subset of
T (e1 ), . . . , T (en ).

Now let A be an m × n matrix over K. We shall denote the m rows of A, which are
row vectors in K n by r1 , r2 , . . . , rm , and similarly, we denote the n columns of A, which
are column vectors in K m,1 by c1 , c2 , . . . , cn .

Definition. 1. The row space of A is the subspace of K n spanned by the rows

r1 , . . . , rm of A. The row rank of A is equal to the dimension of the row space of
A. Equivalently, by the argument above, the row rank of A is equal to the size of
the largest linearly independent subset of r1 , . . . , rm .

2. The column space of A is the subspace of K m,1 spanned by the columns c1 , . . . , cn

of A. The column rank of A is equal to the dimension of the column space of A.
Equivalently, by the argument above, the column rank of A is equal to the size of
the largest linearly independent subset of c1 , . . . , cn .

There is no obvious reason why there should be any particular relationship between
the row and column ranks, but in fact it will turn out that they are always equal. First
we show that the column rank is the same as the rank of the associated linear map.

Theorem 8.4. Suppose that the linear map T has matrix A. Then rank(T ) is equal to
the column rank of A.

Proof. As we saw in Section 7.1, the columns c1 , . . . , cn of A are precisely the column
vectors of coordinates of the vectors T (e1 ), . . . , T (en ), with respect to our chosen basis
of V . The result now follows directly from Lemma 8.3.

 
1 2 0 1 1 r1
A = 2 4 1 3 0  r2
4 8 0 4 4 r3
c1 c2 c3 c4 c5

We can calculate the row and column ranks by applying the sifting process (described
in Section 3) to the row and column vectors, respectively.
Doing rows first, r1 and r2 are linearly independent, but r3 = 4r1 , so the row rank
is 2.
Now doing columns, c2 = 2c1 , c4 = c1 + c3 and c5 = c1 − 2c3 , so the column rank
is also 2.

Theorem 8.5. Applying elementary row operations (R1), (R2) or (R3) to a matrix does
not change the row or column rank. The same is true for elementary column operations
(C1), (C2) and (C3).

Proof. We will prove first that the elementary row operations do not change either the
row rank or column rank.
The row rank of a matrix A is the dimension of the row space of A, which is the
space of linear combinations λ1 r1 + · · · + λm rm of the rows of A. It is easy to see that
(R1), (R2) and (R3) do not change this space, so they do not change the row-rank.
(But notice that the scalar in (R3) must be non-zero for this to be true!)
The column rank of A = (αij ) is the size of the largest linearly independent subset
of c1 , . . . , cn . Let {c1 , . . . , cs } be some subset of the set {c1 , . . . , cn } of columns of A.
(We have written this as though the subset consisted of the first s columns, but this is
just to keep the notation simple; it could be any subset of the columns.)
Then c1 , . . . , cs are linearly dependent if and only if there exist scalars x1 , . . . , xs ∈
K, not all zero, such that x1 c1 +x2 c2 +· · ·+xs cs = 0. If we write out the m components

31
of this vector equation, we get a system of m simultaneous linear equations in the scalars
xi (which is why we have suddenly decided to call the scalars xi rather than λi ).

α11 x1 + α12 x2 + · · · + α1s xs = 0

α21 x1 + α22 x2 + · · · + α2s xs = 0
..
.
αm1 x1 + αm2 x2 + · · · + αms xs = 0

Now if we perform (R1), (R2) or (R3) on A, then we perform the corresponding opera-
tion on this system of equations. That is, we add a multiple of one equation to another,
we interchange two equations, or we multiply one equation by a non-zero scalar. None
of these operations change the set of solutions of the equations. Hence if they have some
solution with the xi not all zero before the operation, then they have the same solution
after the operation. In other words, the elementary row operations do not change the
linear dependence or independence of the set of columns {c1 , . . . , cs }. Thus they do not
change the size of the largest linearly independent subset of c1 , . . . , cn , so they do not
change the column rank of A.
The proof for the column operations (C1), (C2) and (C3) is the same with rows and
columns interchanged.
Corollary 8.6. Let s be the number of non-zero rows in the Smith normal form of a
matrix A (see Theorem 8.2). Then both row rank of A and column rank of A are equal
to s.
Proof. Since elementary operations preserve ranks, it suffices to find both ranks of a
matrix in Smith normal form. But it is easy to see that the row space is precisely the
space spanned by the first s standard vectors and hence has dimension s. Similarly the
column space has dimension s.
In particular, Corollary 8.6 establishes that the row rank is always equal to the
column rank. This allows us to forget this distinction. From now we shall just talk
about the rank of a matrix.
Corollary 8.7. The rank of a matrix A is equal to the number of non-zero rows after
reducing A to upper echelon form.
Proof. The corollary follows from the fact that non-zero rows of a matrix in upper
echelon form are linearly independent.
To see this, let r1 , . . . , rs be the non-zero rows, and suppose that λ1 r1 +· · ·+λs rs = 0.
Now r1 is the only row with a non-zero entry in column c(1), so the entry in column
c(1) of the vector λ1 r1 + · · · + λs rs is λ1 , and hence λ1 = 0.
But then r2 is the only row rk with k ≥ 2 with a non-zero entry in column c(2) and
so the entry in column c(2) of the vector λ2 r2 + · · · + λs rs is λ2 , and hence λ2 = 0.
Continuing in this way (by induction), we find that λ1 = λ2 = · · · = λs = 0, and so
r1 , . . . , rs are linearly independent, as claimed.

Corollary 8.7 gives the most

 efficient 
way of computing the rank of a matrix. For
1 2 0 1 1
instance, let us look at A = 2 4 1 3 0.
4 8 1 5 2

Matrix Operation
 
1 2 0 1 1
 2 4 1 3 0 r2 → r2 − 2r1
r3 → r3 − 4r1
4 8 1 5 2
 
1 2 0 1 1
0 0 1 1 −2 r3 → r3 − r2
0 0 1 1 −2
32
 Matrix Operation

 
1 2 0 1 1
0 0 1 1 −2
0 0 0 0 0

Since the resulting matrix in upper echelon form has 2 nonzero rows, rank(A) = 2.

8.7 The rank criterion

The following theorem is proved in Assignment Sheet 6.
Theorem 8.8. Let A be the augmented n × (m + 1) matrix of a linear system. Let B
be the n × m matrix obtained from A by removing the last column. The system of linear
equations has a solution if and only if rank(A) = rank(B).

9 The inverse of a linear transformation and of a ma-

trix
9.1 Definitions
As usual, let T : U → V be a linear map with corresponding m × n matrix A. If there
is a map T −1 : V → U with T T −1 = IV and T −1 T = IU then T is said to be invertible,
and T −1 is called the inverse of T .
If this is the case, and A−1 is the (n × m) matrix of T −1 , then we have AA−1 = Im
and A−1 A = In . We call A−1 the inverse of the matrix A, and say that A is invertible.
Conversely, if A−1 is an n × m matrix satisfying AA−1 = Im and A−1 A = In , then the
corresponding linear map T −1 satisfies T T −1 = IV and T −1 T = IU , so it is the inverse
of T .
Lemma 9.1. Let A be a matrix of a linear map T . A linear map T is invertible if and
only if its matrix A is invertible. The inverses T −1 and A−1 are unique.
Proof. Recall that, under the bijection between matrices and linear maps, multiplication
of matrices corresponds to composition of linear maps. It now follows immediately from
the definitions above that invertible matrices correspond to invertible linear maps. This
establishes the first statement.
Since the inverse map of a bijection is unique, T −1 is unique. Under the bijection
between matrices and linear maps, A−1 must be the matrix of T −1 . Thus, A−1 is unique
as well.
Theorem 9.2. A linear map T is invertible if and only if T is non-singular. In partic-
ular, if T is invertible then m = n, so only square matrices can be invertible.
See Corollary 5.5 for the definition of non-singular linear maps. We may also say
that the matrix A is non-singular if T is; but by this theorem, this is equivalent to A
being invertible.
Proof. If any function T has a left and right inverse, then it must be a bijection. Hence
ker(T ) = {0} and im(T ) = V , so nullity(T ) = 0 and rank(T ) = dim(V ) = m. But by
Theorem 5.4, we have

n = dim(U ) = rank(T ) + nullity(T ) = m + 0 = m

and we see from the definition that T is non-singular.

Conversely, if n = m and T is non-singular, then by Corollary 5.5 T is a bijection,
and so it has an inverse T −1 : V → U as a function. However, we still have to show
that T −1 is a linear map. Let v1 , v2 ∈ V . Then there exist u1 , u2 ∈ U with T (u1 ) =

33
v1 , T (u2 ) = v2 . So T (u1 + u2 ) = v1 + v2 and hence T −1 (v1 + v2 ) = u1 + u2 . If α ∈ K,
then
T −1 (αv1 ) = T −1 (T (αu1 )) = αu1 = αT −1 (v1 ),
so T −1 is linear, which completes the proof.
 
  1 −2
1 2 0
Example. Let A = and B =  0 1 . Then AB = I2 , but BA 6= I3 , so a
2 0 1
−2 5
non-square matrix can have a right inverse which is not a left inverse. However, it can
be deduced from Corollary 5.5 that if A is a square n × n matrix and AB = In then
A is non-singular, and then by multiplying AB = In on the left by A−1 , we see that
B = A−1 and so BA = In .
This technique of multiplying on the left or right by A−1 is often used for trans-
forming matrix equations. If A is invertible, then AX = B ⇐⇒ X = A−1 B and
XA = B ⇐⇒ X = BA−1 .

Lemma 9.3. If A and B are invertible n × n matrices, then AB is invertible, and

(AB)−1 = B −1 A−1 .
Proof. This is clear, because ABB −1 A−1 = B −1 A−1 AB = In .

9.2 Matrix inversion by row reduction

Two methods for finding the inverse of a matrix will be studied in this course. The
first, using row reduction, which we shall look at now, is an efficient practical method
similar to that used by computer packages. The second, using determinants, is of more
theoretical interest, and will be done later in Section 10.
First note that if an n × n matrix A is invertible, then it has rank n. Consider the
row reduced form B = (βij ) of A. As we saw in Section 8.6, we have βic(i) = 1 for
1 ≤ i ≤ n (since rank(A) = rank(B) = n), where c(1) < c(2) < · · · < c(n), and this is
only possible without any zero columns if c(i) = i for 1 ≤ i ≤ n. Then, since all other
entries in column c(i) are zero, we have B = In . We have therefore proved:
Proposition 9.4. The row reduced form of an invertible n × n matrix A is In .

To compute A−1 , we reduce A to its row reduced form In , using elementary row
operations, while simultaneously applying the same row operations, but starting with
the identity matrix In . It turns out that these operations transform In to A−1 .
In practice, we might not know whether or not A is invertible before we start, but
we will find out while carrying out this procedure because, if A is not invertible, then
its rank will be less than n, and it will not row reduce to In .
First we will do an example to demonstrate the method, and then we will explain
why it works. In the table below, the row operations applied are given in the middle
column. The results of applying them to the matrix
 
3 2 1
A = 4 1 3
2 1 6

are given in the left column, and the results of applying them to I3 in the right column.
So A−1 should be the final matrix in the right column.

Matrix 1 Operation(s) Matrix 2

   
3 2 1 1 0 0
4 1 3 0 1 0
2 1 6 0 0 1
 ↓  r1 → r1 /3  ↓ 
1 2/3 1/3 1/3 0 0
4 1 3   0 1 0
2 1 6 0 0 1
34
 Matrix 1 Operation(s) Matrix 2

.. ..
. r2 → r2 − 4r1 .
 ↓  r3 → r3 − 2r1  ↓ 
1 2/3 1/3 1/3 0 0
0 −5/3 5/3  −4/3 1 0
0 −1/3 16/3 −2/3 0 1
 ↓  r2 → −3r2 /5  ↓ 
1 2/3 1/3 1/3 0 0
0 1 −1   4/5 −3/5 0
0 −1/3 16/3 −2/3 0 1
.. ..
. r1 → r1 − 2r2 /3 .
 ↓  r3 → r3 + r2 /3  ↓ 
1 0 1 −1/5 2/5 0
0 1 −1  4/5 −3/5 0
0 0 5 −2/5 −1/5 1
 ↓  r3 → r3 /5  ↓ 
1 0 1 −1/5 2/5 0
0 1 −1  4/5 −3/5 0 
0 0 1 −2/25 −1/25 1/5
.. ..
. r1 → r1 − r3 .
 ↓  r2 → r2 + r3  ↓ 
1 0 0 −3/25 11/25 −1/5
0 1 0  18/25 −16/25 1/5 
0 0 1 −2/25 −1/25 1/5

So  
−3/25 11/25 −1/5
A−1 =  18/25 −16/25 1/5  .
−2/25 −1/25 1/5
It is always a good idea to check the result afterwards. This is easier if we remove the
common denominator 25, and we can then easily check that
       
3 2 1 −3 11 −5 −3 11 −5 3 2 1 25 0 0
4 1 3  18 −16 5  =  18 −16 5  4 1 3 =  0 25 0 
2 1 6 −2 −1 5 −2 −1 5 2 1 6 0 0 25
which confirms the result!

9.3 Elementary matrices

We shall now explain why the above method of calculating the inverse of a matrix works.
Each elementary row operation on a matrix can be achieved by multiplying the matrix
on the left by a corresponding matrix known as an elementary matrix. There are three
types of these, all being slightly different from the identity matrix.
1. E(n)1λ,i,j (where i 6= j) is the an n × n matrix equal to the identity, but with an
additional non-zero entry λ in the (i, j) position.
2. E(n)2i,j is the n × n identity matrix with its ith and jth rows interchanged.

3. E(n)3λ,i (where λ 6= 0) is the n × n identity matrix with its (i, i) entry replaced by
λ.
Example. Some 3 × 3 elementary matrices:
 
 1
 1 0 0 0  
1 0 3 1 0 0
0 0 0 1
E(3)11 ,1,3 = 0 1 0  , E(4)22,4 =   , E(3)3−4,3 = 0 1 0  .

3 0 0 1 0
0 0 1 0 0 −4
0 1 0 0

35
 Let A be any m × n matrix. Then E(m)1λ,i,j A is the result we get by adding λ times
the jth row of A to the ith row of A. Similarly E(m)2i,j A is equal to A with its ith and
jth rows interchanged, and E(m)3λ,i is equal to A with its ith row multiplied by λ. You
need to work out a few examples to convince yourself that this is true. For example
      
1 1 1 1 1 0 0 0 1 1 1 1 1 1 1 1
2 2 2 2 0 1 0 0 2 2 2 2 2 2 2 2
E(4)1−2,4,2 

=  = .
3 3 3 3  0 0 1 0 3 3 3 3 3 3 3 3
4 4 4 4 0 −2 0 1 4 4 4 4 0 0 0 0

So, in the matrix inversion procedure, the effect of applying elementary row opera-
tions to reduce A to the identity matrix In is equivalent to multiplying A on the left by a
sequence of elementary matrices. In other words, we have Er Er−1 . . . E1 A = In , for cer-
tain elementary n × n matrices E1 , . . . , Er . Hence Er Er−1 . . . E1 = A−1 . But when we
apply the same elementary row operations to In , then we end up with Er Er−1 . . . E1 In =
A−1 . This explains why the method works.
Notice also that the inverse of an elementary row matrix is another one of the same
type. In fact it is easily checked that the inverses of E(n)1λ,i,j , E(n)2i,j and E(n)3λ,i are
respectively E(n)1−λ,i,j , E(n)2i,j and E(n)3λ−1 ,i . Hence, if Er Er−1 . . . E1 A = In as in the
preceding paragraph, then by using Lemma 9.3 we find that

A = (Er Er−1 . . . E1 )−1 = E1−1 E2−1 . . . Er−1 ,

which is itself a product of elementary matrices. We have proved:

Theorem 9.5. An invertible matrix is a product of elementary matrices.

9.4 Application to linear equations

The most familiar examples of simultaneous equations are those where we have the same
number n of equations as unknowns. However, even in that case, there is no guarantee
that there is a unique solution; there can still be zero, one or many solutions (for instance,
see examples in section 8.1). occurs exactly when the matrix A is non-singular.

Theorem 9.6. Let A be an n × n matrix. Then

(i) the homogeneous system of equations Ax = 0 has a non-zero solution if and only
if A is singular;

(ii) the equation system Ax = β has a unique solution if and only if A is non-singular.

Proof. We first prove (i). The solution set of the equations is exactly nullspace(A). If
T is the linear map corresponding to A then, by Corollary 5.5,

nullspace(T ) = ker(T ) = {0} ⇐⇒ nullity(T ) = 0 ⇐⇒ T is non-singular,

and so there are non-zero solutions if and only if T and hence A is singular.
Now (ii). If A is singular then its nullity is greater than 0 and so its nullspace is not
equal to {0}, and contains more than one vector. Either there are no solutions, or the
solution set is x + nullspace(A) for some specific solution x, in which case there is more
than one solution. Hence there cannot be a unique solution when A is singular.
Conversely, if A is non-singular, then it is invertible by Theorem 9.2, and one so-
lution is x = A−1 β. Since the complete solution set is then x + nullspace(A), and
nullspace(A) = {0} in this case, the solution is unique.

In general, it is more efficient to solve the equations Ax = β by elementary row

operations rather than by first computing A−1 and then A−1 β. However, if A−1 is
already known for some reason, then this is a useful method.

36
Example. Consider the system of linear equations
3x + 2y + z = 0 (1)
4x + y + 3z = 2 (2)
2x + y + 6z = 6. (3)
   
3 2 1 −3/25 11/25 −1/5
Here A = 4 1 3, and we computed A−1 =  18/25 −16/25 1/5  in Sec-
2 1 6   −2/25 −1/25 1/5
0
tion 9. Computing A−1 β with β = 2 yields the solution x = −8/25, y = −2/25,
6
z = 28/25. If we had not already known A−1 , then it would have been quicker to solve
the linear equations directly rather than computing A−1 first.

10 The determinant of a matrix

10.1 Definition of the determinant
Let A be an n × n matrix over the field K. The determinant of A, which is written
as det(A) or sometimes as |A|, is a certain scalar that is defined from A in a rather
complicated way. The definition for small values of n might already be familiar to you.
n=1 A = (α) det(A) = α
 
α11 α12
n=2 A= det(A) = α11 α22 − α12 α21
α21 α22
And, for n = 3, we have  
α11 α12 α13
A = α21 α22 α23 
α31 α32 α33
and
     
α α
det(A) = α11  22 23  − α12  α21 α23  + α13  α21 α22
    

α32 α33   α31 α33   α31 α32 
= α11 α22 α33 − α11 α23 α32 − α12 α21 α33
+ α12 α23 α31 + α13 α21 α32 − α13 α22 α31 .
Where do these formulae come from, and why are they useful?
The geometrical motivation for the determinant is that it represents area or volume.
For n = 2, consider the position vectors of two points (x1 , y1 ), (x2 , y2 ) in the plane.
Then, in the diagram below, the area of the parallelogram OABC enclosed by these two
vectors is
 
 x1 x2 
r1 r2 sin(θ2 − θ1 ) = r1 r2 (sin θ2 cos θ1 − sin θ1 cos θ2 ) = x1 y2 − x2 y1 = 
 
y1 y2 

 C
 
B = (x2 , y2 )   
 
 
 
 

r2  
 
 
 1
r1  

 θ2  A = (x1 , y1 )
  θ1

O

37
Similarly, when n = 3 the volume of the parallelepiped enclosed by the three position
vectors in space is equal to (plus or minus) the determinant of the 3 × 3 matrix defined
by the co-ordinates of the three points.
Now we turn to the general definition for n × n matrices. Suppose that we take the
product of n entries from the matrix, where we take exactly one entry from each row
and one from each column. Such a product is called an elementary product. There are
n! such products altogether (we shall see why shortly) and the determinant is the sum
of n! terms, each of which is plus or minus one of these elementary products. We say
that it is a sum of n! signed elementary products. You should check that this holds in
the 2 and 3-dimensional cases written out above.
Before we can be more precise about this, and determine which signs we choose for
which elementary products, we need to make a short digression to study permutations
of finite sets. A permutation of a set, which we shall take here to be the set Xn =
{1, 2, 3, . . . , n}, is simply a bijection from Xn to itself. The set of all such permutations
of Xn is called the symmetric group Sn . There are n! permutations altogether, so
|Sn | = n!.
(A group is a set of objects, any two of which can be multiplied or composed to-
gether, and such that there is an identity element, and all elements have inverses. Other
examples of groups that we have met in this course are the n × n invertible matrices
over K, for any fixed n, and any field K. The study of groups, which is known as Group
Theory, is an important branch of mathematics, but it is not the main topic of this
course!)
Now an elementary product contains one entry from each row of A, so let the entry
in the product from the ith row be αiφ(i) , where φ is some as-yet unknown function from
Xn to Xn . Since the product also contains exactly one entry from each column, each
integer j ∈ Xn must occur exactly once as φ(i). But this is just saying that φ : Xn → Xn
is a bijection; that is φ ∈ Sn . Conversely, any φ ∈ Sn defines an elementary product in
this way.
So an elementary product has the general form α1φ(1) α2φ(2) . . . αnφ(n) for some φ ∈
Sn , and there are n! elementary products altogether. We want to define
X
det(A) = ±α1φ(1) α2φ(2) . . . αnφ(n) ,
φ∈Sn

but we still have to decide which of the elementary products has a plus sign and which
has a minus sign. In fact this depends on the sign of the permutation φ, which we must
now define.
A transposition is a permutation of Xn that interchanges two numbers i and j in
Xn and leaves all other numbers fixed. It is written as (i, j). There is a theorem, which
is quite easy, but we will not prove it here because it is a theorem in Group Theory,
that says that every permutation can be written as a composite of transpositions. For
example, if n = 5, then the permutation φ defined by
φ(1) = 4, φ(2) = 5, φ(3) = 3, φ(4) = 2, φ(5) = 1
is equal to the composite (1, 4)◦(2, 4)◦(2, 5). (Remember that permutations are functions
Xn → Xn , so this means first apply the function (2, 5) (which interchanges 2 and 5)
then apply (2, 4) and finally apply (1, 4).)
Definition. Now a permutation φ is said to be even, and to have sign +1, if φ is a
composite of an even number of permutations; and φ is said to be odd, and to have sign
−1, if φ is a composite of an odd number of permutations.
For example, the permutation φ defined on Xn above is a composite of 3 transposi-
tions, so φ is odd and sign(φ) = −1. The identity permutation, which leaves all points
fixed, is even (because it is a composite of 0 transpositions).
Now at last we can give the general definition of the determinant.
Definition. The determinant of a n × n matrix A = (αij ) is the scalar quantity
X
det(A) = sign(φ)α1φ(1) α2φ(2) . . . αnφ(n) .
φ∈Sn

38
 (Note: You might be worrying about whether the same permutation could be both
even and odd. Well, there is a moderately difficult theorem in Group Theory, which we
shall not prove here, that says that this cannot happen; in other words, the concepts of
even and odd permutation are well-defined.)

10.2 The effect of matrix operations on the determinant

Theorem 10.1. Elementary row operations affect the determinant of a matrix as fol-
lows.

(i) det(In ) = 1.

(ii) Let B result from A by applying (R2) (interchanging two rows). Then det(B) =
− det(A).

(iii) If A has two equal rows then det(A) = 0.

(iv) Let B result from A by applying (R1) (adding a multiple of one row to another).
Then det(B) = det(A).

(v) Let B result from A by applying (R3) (multiplying a row by a scalar λ). Then
det(B) = λ det(A).

Proof. (i) If A = In then αij = 0 when i 6= j. So the only non-zero elementary

product in the sum occurs when φ is the identity permutation. Hence det(A) =
α11 α22 . . . αnn = 1.

(ii) To keep the notation simple, we shall suppose that we interchange the first two
rows, but the same argument works for interchanging any pair of rows. Then if
B = (βij ), we have β1j = α2j and β2j = α1j for all j. Hence

X
det(B) = sign(φ)β1φ(1) β2φ(2) . . . βnφ(n)
φ∈Sn
X
= sign(φ)α1φ(2) α2φ(1) α3φ(3) . . . αnφ(n) .
φ∈Sn

For φ ∈ Sn , let ψ = φ ◦ (1, 2), so φ(1) = ψ(2) and φ(2) = ψ(1), and sign(ψ) =
− sign(φ). Now, as φ runs through all permutations in Sn , so does ψ (but in a
different order), so summing over all φ ∈ Sn is the same as summing over all
ψ ∈ Sn . Hence
X
det(B) = − sign(ψ)α1ψ(1) α2ψ(2) . . . αnψ(n)
φ∈Sn
X
= − sign(ψ)α1ψ(1) α2ψ(2) . . . αnψ(n) = − det(A).
ψ∈Sn

(iii) Again to keep notation simple, assume that the equal rows are the first two. Using
the same notation as in (ii), namely ψ = φ ◦ (1, 2), the two elementary products:

α1ψ(1) α2ψ(2) . . . αnψ(n) and α1φ(1) α2φ(2) . . . αnφ(n)

are equal. This is because α1ψ(1) = α2ψ(1) (first two rows equal) and α2ψ(1) =
α2φ(2) (because φ(2) = ψ(1)); hence α1ψ(1) = α2φ(2) . Similarly α2ψ(2) = α1φ(1)
and the two products differ by interchanging their first two terms. But sign(ψ) =
− sign(φ) so the two corrresponding signed products cancel each other out. Thus
each signed product in det(A) cancels with another and the sum is zero.

39
(iv) Again, to simplify notation, suppose that we replace the second row r2 by r2 + λr1
for some λ ∈ K. Then
X
det(B) = sign(φ)α1φ(1) (α2φ(2) + λα1φ(2) )α3φ(3) . . . αnφ(n)
φ∈Sn
X
= sign(φ)α1φ(1) α2φ(2) . . . αnφ(n)
φ∈Sn
X
+λ sign(φ)α1φ(1) α1φ(2) . . . αnφ(n) .
φ∈Sn

Now the first term in this sum is det(A), and the second is λ det(C), where C is
a matrix in which the first two rows are equal. Hence det(C) = 0 by (iii), and
det(B) = det(A).
(v) Easy. Note that this holds even when the scalar λ = 0.

Definition. A matrix is called upper triangular if all of its entries below the main
diagonal are zero; that is, (αij ) is upper triangular if αij = 0 for all i > j.
The matrix is called diagonal if all entries not on the main diagonal are zero; that
is, αij = 0 for i 6= j.
   
3 0 −1/2 0 0 0
Example. 0 −1 −11  is upper triangular, and 0 17 0  is diagonal.
0 0 −2/5 0 0 −3
Corollary 10.2. If A = (αij ) is upper triangular, then det(A) = α11 α22 . . . αnn is the
product of the entries on the main diagonal of A.
Proof. This is not hard to prove directly from the definition of the determinant. Alter-
natively, we can apply row operations (R1) to reduce the matrix to the diagonal matrix
with the same entries αii on the main diagonal, and then the result follows from parts (i)
and (v) of the theorem.
The above theorem and corollary provide the most efficient way of computing det(A),
at least for n ≥ 3. (For n = 2, it is easiest to do it straight from the definition.) Use
row operations (R1) and (R2) to reduce A to upper triangular form, keeping track of
changes of sign in the determinant resulting from applications of (R2), and then use
Corollary 10.2.
Example.
     
0 1 1 2 1 2 1 1 1 2 1 1 
  
1 2 1 1 r2 ↔r1 0 1 1 2 r3 →r3 −2r1 0 1 1 2 
 = − = −
0 −3 1 −1
2 1 3 1
 2 1 3 1
1 2 4 2 1 2 4 2 1 2 4 2
     
1 2 1 1  1 2 1 1 1 2 1 1 
  
r4 →r4 −r1 0 1 1 2  r3 →r3 +3r2 0 1 1 2 r4 →r4 −3r3 /4 0 1 1 2 
= −  = − = −
0 −3 1 −1 0 0 4 5 0 0 4 5 
0 0 3 1 0 0 3 1 0 0 0 − 11
4

= 11

We could have been a little more clever, and stopped the row reduction one step before
the end, noticing that the determinant was equal to | 43 51 | = 11.
Definition. Let A = (αij ) be an m × n matrix. We define the transpose AT of A to be
the n × m matrix (βij ), where βij = αji for 1 ≤ i ≤ n, 1 ≤ j ≤ m.
 
 T 1 −2
1 3 5
For example, = 3 0  .
−2 0 6
5 6

40
Theorem 10.3. Let A = (αij ) be an n × n matrix. Then det(AT ) = det(A).

Proof. Let AT = (βij ) where βij = αji . Then

X
det(AT ) = sign(φ)β1φ(1) β2φ(2) . . . βnφ(n)
φ∈Sn
X
= sign(φ)αφ(1)1 αφ(2)2 . . . αφ(n)n .
φ∈Sn

Now, by rearranging the terms in the elementary product, we have

αφ(1)1 αφ(2)2 . . . αφ(n)n = α1φ−1 (1) α2φ−1 (2) . . . αnφ−1 (n) ,

where φ−1 is the inverse permutation to φ. Notice also that if φ is a composite τ1 ◦

τ2 ◦ · · · ◦ τr of transpositions τi , then φ−1 = τr ◦ · · · ◦ τ2 ◦ τ1 (because each τi ◦ τi is the
identity permutation). Hence sign(φ) = sign(φ−1 ). Also, summing over all φ ∈ Sn is
the same as summing over all φ−1 ∈ Sn , so we have
X
det(AT ) = sign(φ)α1φ−1 (1) α2φ−1 (2) . . . αnφ−1 (n)
φ∈Sn
X
= sign(φ−1 )α1φ−1 (1) α2φ−1 (2) . . . αnφ−1 (n) = det(A).
φ−1 ∈Sn

If you find proofs like the above, where we manipulate sums of products, hard to
follow, then it might be helpful to write it out in full in a small case, such as n = 3.
Then

det(AT ) = β11 β22 β33 − β11 β23 β32 − β12 β21 β33
+ β12 β23 β31 + β13 β21 β32 − β13 β22 β31
= α11 α22 α33 − α11 α32 α23 − α21 α12 α33
+ α21 α32 α13 + α31 α12 α23 − α31 α22 α13
= α11 α22 α33 − α11 α23 α32 − α12 α21 α33
+ α12 α23 α31 + α13 α21 α32 − α13 α22 α31
= det(A).

Corollary 10.4. All of Theorem 10.1 remains true if we replace rows by columns.

Proof. This follows from Theorems 10.1 and 10.3, because we can apply column op-
erations to A by transposing it, applying the corresponding row operations, and then
re-transposing it.

We are now ready to prove one of the most important properties of the determinant.

Theorem 10.5. For an n × n matrix A, det(A) = 0 if and only if A is singular.

Proof. A can be reduced to row reduced echelon form by using row operations (R1),
(R2) and (R3). By Theorem 8.5, none of these operations affect the rank of A, and so
they do not affect whether or not A is singular (remember ‘singular’ means rank(A) < n;
see definition after Corollary 5.5). By Theorem 10.1, they do not affect whether or not
det(A) = 0. So we can assume that A is in row reduced echelon form.
Then rank(A) is the number of non-zero rows of A, so if A is singular then it has
some zero rows. But then det(A) = 0. On the other hand, if A is nonsingular then,
as we saw in Section 9.2, the fact that A is in row reduced echelon form implies that
A = In , so det(A) = 1 6= 0.

41
10.3 The determinant of a product
   
1 2 −1 −1
Example. Let A = and B = . Then det(A) = −4 and det(B) = 2.
 3 2 2 0
0 1
We have A + B = and det(A + B) = −5 6= det(A) + det(B). In fact, in general
5 2
there is no simple relationship
  between det(A + B) and det(A), det(B).
3 −1
However, AB = , and det(AB) = −8 = det(A) det(B).
1 −3
In this subsection, we shall prove that this simple relationship holds in general.
Recall from Section 9.3 the definition of an elementary matrix E, and the prop-
erty that if we multiply a matrix B on the left by E, then the effect is to apply the
corresponding elementary row operation to B. This enables us to prove:
Lemma 10.6. If E is an n × n elementary matrix, and B is any n × n matrix, then
det(EB) = det(E) det(B).
Proof. E is one of the three types E(n)1λ,ij , E(n)2ij or E(n)3λ,i , and multiplying B on
the left by E has the effect of applying (R1), (R2) or (R3) to B, respectively. Hence,
by Theorem 10.1, det(EB) = det(B), − det(B), or λ det(B), respectively. But by con-
sidering the special case B = In , we see that det(E) = 1, −1 or λ, respectively, and so
det(EB) = det(E) det(B) in all three cases.
Theorem 10.7. For any two n × n matrices A and B, we have
det(AB) = det(A) det(B).
Proof. We first dispose of the case when det(A) = 0. Then we have rank(A) < n by
Theorem 10.5. Let T1 , T2 : V → V be linear maps corresponding to A and B, where
dim(V ) = n. Then AB corresponds to T1 T2 (by Theorem 7.4). By Corollary 5.5,
rank(A) = rank(T1 ) < n implies that T1 is not surjective. But then T1 T2 cannot
be surjective, so rank(T1 T2 ) = rank(AB) < n. Hence det(AB) = 0 so det(AB) =
det(A) det(B).
On the other hand, if det(A) 6= 0, then A is nonsingular, and hence invertible, so by
Theorem 9.5 A is a product E1 E2 . . . Er of elementary matrices Ei . Hence det(AB) =
det(E1 E2 . . . Er B). Now the result follows from the above lemma, because
det(AB) = det(E1 ) det(E2 · · · Er B)
= det(E1 ) det(E2 ) det(E3 · · · Er B)
= det(E1 ) det(E2 ) · · · det(Er ) det(B)
= det(E1 E2 · · · Er ) det(B)
= det(A) det(B).

10.4 Minors and cofactors

Definition. Let A = (αij ) be an n × n matrix. Let Aij be the (n − 1) × (n − 1)
matrix obtained from A by deleting the ith row and the jth column of A. Now let
Mij = det(Aij ). Then Mij is called the (i, j)th minor of A.
 
2 1 0    
3 2 1 0
Example. If A = 3 −1 2 , then A12 =
  and A31 = , and so M12 =
5 0 −1 2
5 −2 0
−10 and M31 = 2.
Definition. We define cij to be equal to Mij if i + j is even, and to −Mij if i + j is
odd. Or, more concisely,
cij = (−1)i+j Mij = (−1)i+j det(Aij ).
Then cij is called the (i, j)th cofactor of A.

42
Example. In the example above,
     
−1 2 3 2 3 −1
c11 = 
  = 4, c12 = −   = 10, c13 =  = −1,
−2 0 5 0 5 −2
     
 1 0 2 0  2 1 
c21 = −   = 0, c22 = 
  = 0, c23 = −   = 9,
−2 0 5 0 5 −2
     
 1 0 2 0 2 1 
c31 = 
  = 2, c32 = −   = −4, c33 =  = −5.
−1 2 3 2 3 −1

The cofactors give us a useful way of expressing the determinant of a matrix in terms
of determinants of smaller matrices.
Theorem 10.8. Let A be an n × n matrix.
(i) (Expansion of a determinant by the ith row.) For any i with 1 ≤ i ≤ n, we have
n
X
det(A) = αi1 ci1 + αi2 ci2 + · · · + αin cin = αij cij .
j=1

(ii) (Expansion of a determinant by the jth column.) For any j with 1 ≤ j ≤ n, we

have
n
X
det(A) = α1j c1j + α2j c2j + · · · αnj cnj = αij cij .
i=1

For example, expanding the determinant of the matrix A above by the first row, the
third row, and the second column give respectively:

det(A) = 2 × 4 + 1 × 10 + 0 × −1 = 18
det(A) = 5 × 2 + −2 × −4 + 0 × −5 = 18
det(A) = 1 × 10 + −1 × 0 + −2 × −4 = 18.

Proof of Theorem 10.8. By definition, we have

X
det(A) = sign(φ)α1φ(1) α2φ(2) . . . αnφ(n) (∗)
φ∈Sn

Step 1. We first find the sum of all of those signed elementary products in the sum (∗)
that contain αnn . These arise from those permutations φ with φ(n) = n; so the required
sum is
X
sign(φ)α1φ(1) α2φ(2) . . . αnφ(n)
φ∈Sn
φ(n)=n
X
= αnn sign(φ)α1φ(1) α2φ(2) . . . αn−1φ(n−1)
φ∈Sn−1

= αnn Mnn = αnn cnn .

Step 2. Next we fix any i and j with 1 ≤ i, j ≤ n, and find the sum of all of those
signed elementary products in the sum (∗) that contain αij . We move row ri of A to
rn by interchanging ri with ri+1 , ri+2 , . . . , rn in turn. This involves n − i applications
of (R2), and leaves the rows of A other than ri in their original order. We then move
column cj to cn in the same way, by applying (C2) n − j times. Let the resulting
matrix be B = (βij ) and denote its minors by Nij . Then βnn = αij , and Nnn = Mij .
Furthermore,
det(B) = (−1)2n−i−j det(A) = (−1)i+j det(A),
because (2n − i − j) − (i + j) is even.
Now, by the result of Step 1, the sum of terms in det(B) involving βnn is

βnn Nnn = αij Mij = (−1)i+j αij cij ,

43
and hence, since det(B) = (−1)i+j det(A), the sum of terms involving αij in det(A) is
αij cij .
Step 3. The result follows from Step 2, because every signed elementary product in
the sum (∗) involves exactly one array element αij from each row and from each column.
Hence, for any given row or column, we get the full sum (∗) by adding up the total of
those products involving each individual element in that row or column.
Example. Expanding by a row and column can sometimes be a quick method of eval-
uating the determinant of matrices containing a lot of zeros. For example, let
 
9 0 2 6
 1 2 9 −3
A=  0 0 −2 0  .


−1 0 −5 2
 9 0 6 
Then, expanding by the third row, we get det(A) = −2  1 2 −3 , and then expanding
 
 9 6 −1 0 2
by the second column, det(A) = −2 × 2  −1 2 = −96.


10.5 The inverse of a matrix using determinants

Definition. Let A be an n × n matrix. We define the adjugate matrix adj(A) of A to
be the n × n matrix of which the (i, j)th element is the cofactor cji . In other words, it
is the transpose of the matrix of cofactors.
The adjugate is also sometimes called the adjoint. However, the word “adjoint” is
also used with other meanings, so to avoid confusion we will use the word “adjugate”.
Example. In the example above,
   
2 1 0 4 0 2
A = 3 −1 2 , adj(A) =  10 0 −4 .
5 −2 0 −1 9 −5
The adjugate is almost an inverse to A, as the following theorem shows.
Theorem 10.9. A adj(A) = det(A)In = adj(A) A
Pn
Proof. Let B = A adj(A) = (βij ). Then βii = P k=1 αik cik = det(A) by Theorem 10.8
n
(expansion by the ith row of A). For i 6= j, βij = k=1 αik cjk , which is the determinant
of a matrix C obtained from A by substituting the ith row of A for the jth row. But
then C has two equal rows, so βij = det(C) = 0 by Theorem 10.1(iii). (This is
sometimes called an expansion by alien cofactors.) Hence A adj(A) = det(A)In . A
similar argument using columns instead of rows gives adj(A) A = det(A)In .
Example. In the example above, check that A adj(A) = adj(A) A = 18I3 .
Corollary 10.10. If det(A) 6= 0, then A−1 = 1
det(A) adj(A).
(Theorems 9.2 and 10.5 imply that A is invertible if and only if det(A) 6= 0.)
Example. In the example above,
 −1  
2 1 0 4 0 2
3 −1 2 = 1  10 0 −4 ,
18
5 −2 0 −1 9 −5
and in the example in Section 9,
   
3 2 1 3 −11 5
A = 4 1 3 , adj(A) = −18 16 −5 , det(A) = −25,
2 1 6 2 1 −5
−1
and so A−1 = 25 adj(A).
For 2 × 2 and (possibly) 3 × 3 matrices, the cofactor method of computing the
inverse is often the quickest. For larger matrices, the row reduction method described
in Section 9 is quicker.

44
10.6 Cramer’s rule for solving simultaneous equations
Given a system Ax = β of n equations in n unknowns, where A = (αij ) is non-singular,
the solution is x = A−1 β. So the ith component xi of this column vector is the ith
row of A−1 β. Now, by Corollary 10.10, A−1 = det(A)1
adj(A), and its (i, j)th entry is
cji / det(A). Hence
n
1 X
xi = cji βj .
det(A) j=1

Now let Ai be P the matrix obtained from A by substituting β for the ith column of A.
n
Then the sum j=1 cji βj is precisely the expansion of det(Ai ) by its ith column (see
Theorem 10.8). Hence we have xi = det(Ai )/ det(A). This is Cramer’s rule.
This is more of a curiosity than a practical method of solving simultaneous equations,
although it can be quite quick in the 2 × 2 case. Even in the 3 × 3 case it is rather slow.

Example. Let us solve the following system of linear equations:

2x + z = 1
y − 2z = 0
x + y + z = −1

Cramer’s rule gives

   
2 0 1  1 0 1
  
det(A) = 0 1 −2 = 5, det(A1 ) =  0 1 −2 = 4
1 1 1 −1 1 1 
   
2 1 1  2 0 1 
  
det(A2 ) = 0 0 −2 = −6, det(A3 ) = 0 1 0  = −3
1 −1 1  1 1 −1

so the solution is x = 4/5, y = −6/5, z = −3/5.

11 Change of basis and equivalent matrices

We have been thinking of matrices as representing linear maps between vector spaces.
But don’t forget that, when we defined the matrix corresponding to a linear map be-
tween vector spaces U and V , the matrix depended on a particular choice of bases for
both U and V . In this section, we investigate the relationship between the matrices cor-
responding to the same linear map T : U → V , but using different bases for the vector
spaces U and V . We first discuss the relationship between two different bases of the
same space. Assume throughout the section that all vector spaces are over the same
field K.
Let U be a vector space of dimension n, and let e1 , . . . , en and e01 , . . . , e0n be two
bases of U . The matrix P of the identity map IU : U → U using the basis e1 , . . . , en
in the domain and e01 , . . . , e0n in the range is called the change of basis matrix from the
basis of ei s to the basis of e0i s.
Let us look carefully what this definition says. Taking P = (σij ), we obtain from
Section 7.1
Xn
IU (ej ) = ej = σij e0i for 1 ≤ j ≤ n. (∗)
i=1

In other words, the columns of P are the coordinates of the “old” basis vectors ei with
respect to the “new” basis e0i .

Proposition 11.1. The change of basis matrix is invertible. More precisely, if P is the
change of basis matrix from the basis of ei s to the basis of e0i s and Q is the change of
basis matrix from the basis of e0i s to the basis of ei s then P = Q−1 .

45
 U U I I
Proof. Consider the composition of linear maps IU : U −−→ U −−→ U using the basis of
0
ei s for the first and the third copy of U and the basis of ei s for the middle copy of U .
The composition has matrix In because the same basis is used for both domain and
range. But the first IU has matrix Q (change of basis from e0i s to ei s) and the second
IU similarly has matrix P . Therefore by Theorem 7.4, In = P Q.
Similarly, In = QP . Consequently, P = Q−1 .

Example. Let U = R3 , e01 = (1, 0, 0), e02 = (0, 1, 0), e03 = (0, 0, 1) (the standard basis)
and e1 = (0, 2, 1), e2 = (1, 1, 0), e3 = (1, 0, 0). Then
 
0 1 1
P = 2 1 0 .
1 0 0

The columns of P are the coordinates of the “old” basis vectors e1 , e2 , e3 with respect
to the “new” basis e01 , e02 , e03 .

As with any matrix, we can take a column vector of coordinates, multiply it by the
change of basis matrix P , and get a new column vector of coordinates. What does this
actually mean?

Proposition 11.2. With the above notation, let v ∈ U , and let v and v0 denote the
column vectors associated with v when we use the bases e1 , . . . , en and e01 , . . . , e0n , re-
spectively. Then P v = v0 .

Proof. This follows immediately from Proposition 7.2 applied to the identity map IU .

This gives a useful way to think about the change of basis matrix: it is the matrix
which turns a vector’s coordinates with respect to the “old” basis into the same vector’s
coordinates with respect to the “new” basis.
Now we will turn to the effect of change of basis on linear maps. let T : U → V be
a linear map, where dim(U ) = n, dim(V ) = m. Choose a basis e1 , . . . , en of U and a
basis f1 , . . . , fm of V . Then, from Section 7.1, we have
m
X
T (ej ) = αij fi for 1 ≤ j ≤ n
i=1

where A = (αij ) is the m × n matrix of T with respect to the bases {ei } and {fi } of U
and V .
Now choose new bases e01 , . . . , e0n of U and f10 , . . . , fm
0
of V . There is now a new
matrix representing the linear transformation T :
m
X
T (e0j ) = βij fi0 for 1 ≤ j ≤ n,
i=1

where B = (βij ) is the m × n matrix of T with respect to the bases {e0i } and {fi0 } of
U and V . Our objective is to find the relationship between A and B in terms of the
change of basis matrices.
Let the n × n matrix P = (σij ) be the change of basis matrix from {ei } to {e0i }, and
let the m × m matrix Q = (τij ) be the change of basis matrix from {fi } to {fi0 }.
U V
e1 , e2 , . . . , en Matrix A f1 , f2 , . . . , fm

P ↓ T - ↓Q
e01 , e02 , . . . , e0n Matrix B f10 , f20 , . . . , fm
0

46
Theorem 11.3. With the above notation, we have BP = QA, or equivalently B =
QAP −1 .

Proof. By Theorem 7.4, BP represents the composite of the linear maps IU using bases
{ei } and {e0i } and T using bases {e0i } and {fi0 }. So BP represents T using bases {ei }
and {fi0 }. Similarly, QA represents the composite of T using bases {ei } and {fi } and
IV using bases {fi } and {fi0 }, so QA also represents T using bases {ei } and {fi0 }. Hence
BP = QA.

Another way to think of this is the following. The matrix B should be the matrix
which, given the coordinates of a vector u ∈ U with respect to the basis {e0i }, produces
the coordinates of T (u) ∈ V with respect to the basis {fi0 }. On the other hand, suppose
we already know the matrix A, which performs the corresponding task with the “old”
bases {ei } and {ff }. Now, given the coordinates of some vector u with respect to the
“new” basis, we need to:

(i) Find the coordinates of u with respect to the “old” basis of U : this is done by
multiplying by the change of basis matrix from {e0i } to {ei }, which is P −1 ;

(ii) find the coordinates of T (u) with respect to the “old” basis of V : this is what
multiplying by A does;

(iii) translate the result into coordinates with respect to the “new” basis for V ; this is
done by multiplying by the change of basis matrix Q.

Putting these three steps together, we again see that B = QAP −1 .

Corollary 11.4. Two m × n matrices A and B represent the same linear map from an
n-dimensional vector space to an m-dimensional vector space (with respect to different
bases) if and only if there exist invertible n × n and m × m matrices P and Q with
B = QAP .

Proof. It follows from the Theorem 11.3 that A and B represent the same linear map if
there exist change of basis matrices P and Q with B = QAP −1 , and by Proposition 11.1
the change of basis matrices are precisely invertible matrices of the correct size. By
replacing P by P −1 , we see that this is equivalent to saying that there exist invertible
Q, P with B = QAP .

Definition. Two m × n matrices A and B are said to be equivalent if there exist

invertible P and Q with B = QAP .

It is easy to check that being equivalent is an equivalence relation on the set K m,n
of m × n matrices over K. We shall show now that equivalence of matrices has other
characterisations.

Theorem 11.5. Let A and B be m × n matrices over K. Then the following conditions
on A and B are equivalent.

(i) A and B are equivalent.

(ii) A and B represent the same linear map with respect to different bases.

(iii) A and B have the same rank.

(iv) B can be obtained from A by application of elementary row and column operations.

Proof. (i) ⇔ (ii): This is true by Corollary 11.4.

(ii) ⇒ (iii): Since A and be both represent the same linear map T , we have rank(A) =
rank(B) = rank(T ).
(iii) ⇒ (iv): By Theorem 8.2, if A and B both have rank s, then they can both be
brought into the form  
Is 0s,n−s
Es =
0m−s,s 0m−s,n−s

47
by elementary row and column operations. Since these operations are invertible, we can
first transform A to Es and then transform Es to B.
(iv) ⇒ (i): We saw in Section 9.2 that applying an elementary row operation to A can
be achieved by multiplying A on the left by an elementary row matrix, and similarly ap-
plying an elementary column operation can be done by multiplying A on the right by an
elementary column matrix. Hence (iv) implies that there exist elementary row matrices
R1 , . . . , Rr and elementary column matrices C1 , . . . , Cs with B = Rr · · · R1 AC1 · · · Cs .
Since elementary matrices are invertible, Q = Rr · · · R1 and P = C1 · · · Cs are invertible
and B = QAP .
In the above proof, we also showed the following:
Proposition 11.6. Any m × n matrix is equivalent to the matrix Es defined above,
where s = rank(A).
The form Es is known as a canonical form for m × n matrices under equivalence.
This means that it is an easily recognizable representative of its equivalence class.

12 Similar matrices, eigenvectors and eigenvalues

12.1 Similar matrices
In Section 11 we studied what happens to the matrix of a linear map T : U → V when
we change bases of U and V . Now we look at the case when U = V , where we only have
a single vector space V , and a single change of basis. Surprisingly, this turns out to be
more complicated than the situation with two different spaces.
Let V be a vector space of dimension n over the field K, and let T : V → V be a
linear map. Now, given any basis for V , there will be a matrix representing T with
respect to that basis.
Let e1 , . . . , en and e01 , . . . , e0n be two bases of V , and let A = (αij ) and B = (βij )
be the matrices of T with respect to {ei } and {e0i } respectively. Let P = (σij ) be the
change of basis matrix from {e0i } to {ei }. Note that this is the opposite change of basis
to the one considered in the last section. Different textbooks adopt different conventions
on which way round to do this; this is how we’ll do it in this course.
Then Theorem 11.3 applies, and with both Q and P replaced by P −1 we find:
Theorem 12.1. With the notation above, B = P −1 AP .
Definition. Two n × n matrices over K are said to be similar if there exists an n × n
invertible matrix P with B = P −1 AP .
So two matrices are similar if and only if they represent the same linear map T : V →
V with respect to different bases of V . It is easily checked that similarity is an equivalence
relation on the set of n × n matrices over K.
We saw in Theorem 11.5 that two matrices of the same size are equivalent if and
only if they have the same rank. It is more difficult to decide whether two matrices are
similar, because we have much less flexibility - there is only one basis to choose, not two.
Similar matrices are certainly equivalent, so they have the same rank, but equivalent
matrices need not be similar.
   
1 0 1 1
Example. Let A = and B = .
0 1 0 1
Then A and B both have rank 2, so they are equivalent. However, since A = I2 , for
any invertible 2 × 2 matrix P we have P −1 AP = A, so A is similar only to itself. Hence
A and B are not similar.
To decide whether matrices are similar, it would be helpful to have a canonical form,
just like we had the canonical form Es in Section 11 for equivalence. Then we could test
for similarity by reducing A and B to canonical form and checking whether we get the
same result. But this turns out to be quite difficult, and depends on the field K. For
the case K = C (the complex numbers), we have the Jordan Canonical Form, which
Maths students learn about in the Second Year.

48
12.2 Eigenvectors and eigenvalues
In this course, we shall only consider the question of which matrices are similar to a
diagonal matrix.
Definition. A matrix which is similar to a diagonal matrix is said to be diagonalisable.
(Recall that A = (αij ) is diagonal if αij = 0 for i 6= j.) We shall see, for example,
that the matrix B in the example above is not diagonalisable.
It turns out that the possible entries on the diagonal of a matrix similar to A can be
calculated directly from A. They are called eigenvalues of A and depend only on the
linear map to which A corresponds, and not on the particular choice of basis.
Definition. Let T : V → V be a linear map, where V is a vector space over K. Suppose
that for some non-zero vector v ∈ V and some scalar λ ∈ K, we have T (v) = λv. Then
v is called an eigenvector of T , and λ is called the eigenvalue of T corresponding to v.
Note that the zero vector is not an eigenvector. (This would not be a good idea,
because T 0 = λ0 for all λ.) However, the zero scalar 0K may sometimes be an eigenvalue
(corresponding to some non-zero eigenvector).
Example. Let T : R2 → R2 be defined by T (α1 , α2 ) = (2α1 , 0). Then T (1, 0) = 2(1, 0),
so 2 is an eigenvalue and (1, 0) an eigenvector. Also T (0, 1) = (0, 0) = 0(0, 1), so 0 is an
eigenvalue and (0, 1) an eigenvector.
In this example, notice that in fact (α, 0) and (0, α) are eigenvectors for any α 6= 0.
In general, it is easy to see that if v is an eigenvector of T , then so is αv for any non-zero
scalar α.
In some books, eigenvectors and eigenvalues are called characteristic vectors and
characteristic roots, respectively.
Let e1 , . . . , en be a basis of V , and let A = (αij ) be the matrix of T with respect to
this basis. As in Section 7.1, to each vector v = λ1 e1 + · · · + λn en ∈ V , we associate its
column vector of coordinates  
λ1
 λ2 
v =  .  ∈ K n,1 .
 
 .. 
λn
Then, by Proposition 7.2, for u, v ∈ V , we have T (u) = v if and only if Au = v, and in
particular
T (v) = λv ⇐⇒ Av = λv.
So it will be useful to define the eigenvalues and eigenvectors of a matrix, as well as
of a linear map.
Definition. Let A be an n × n matrix over K. Suppose that, for some non-zero column
vector v ∈ K n,1 and some scalar λ ∈ K, we have Av = λv. Then v is called an
eigenvector of A, and λ is called the eigenvalue of A corresponding to v.
It follows from Proposition 7.2 that if the matrix A corresponds to the linear map
T , then λ is an eigenvalue of T if and only if it is an eigenvalue of A. It follows
immediately that similar matrices have the same eigenvalues, because they represent
the same linear map with respect to different bases. We shall give another proof of this
fact in Theorem 12.3 below.
Given a matrix, how can we compute its eigenvalues? Certainly trying every vector
to see whether it is an eigenvector is not a practical approach.
Theorem 12.2. Let A be an n × n matrix. Then λ is an eigenvalue of A if and only if
det(A − λIn ) = 0.
Proof. Suppose that λ is an eigenvalue of A. Then Av = λv for some non-zero v ∈ K n,1 .
This is equivalent to Av = λIn v, or (A − λIn )v = 0. But this says exactly that v is a
non-zero solution to the homogeneous system of simultaneous equations defined by the

49
matrix A−λIn , and then by Theorem 9.6(i), A−λIn is singular, and so det(A−λIn ) = 0
by Theorem 10.5.
Conversely, if det(A − λIn ) = 0 then A − λIn is singular, and so by Theorem 9.6(i)
the system of simultaneous equations defined by A − λIn has nonzero solutions. Hence
there exists a non-zero v ∈ K n,1 with (A−λIn )v = 0, which is equivalent to Av = λIn v,
and so λ is an eigenvalue of A.

If we treat λ as an unknown, we get a polynomial equation which we can solve to

find all the eigenvalues of A:

Definition. For an n × n matrix A, the equation det(A − xIn ) = 0 is called the

characteristic equation of A, and det(A − xIn ) is called the characteristic polynomial of
A.

Note that the characteristic polynomial of an n × n matrix is a polynomial of degree

n in x.
The above theorem says that the eigenvalues of A are the roots of the characteristic
equation, which means that we have a method of calculating them. Once the eigenvalues
are known, it is then straightforward to compute the corresponding eigenvectors.
 
1 2
Example. Let A = . Then
5 4
 
1 − x 2 
det(A − xI2 ) = 
  = (1 − x)(4 − x) − 10 = x2 − 5x − 6 = (x − 6)(x + 1).
5 4 − x

Hence the eigenvalues of A are the roots of (x − 6)(x + 1) = 0; that is, 6 and −1.
Let us now find the eigenvectors corresponding to the eigenvalue 6. We seek a non-
zero column vector ( xx12 ) such that
         
1 2 x1 x −5 2 x1 0
= 6 1 ; that is, = .
5 4 x2 x2 5 −2 x2 0
   
x1 2
Solving this easy system of linear equations, we can take = to be our
  x 2 5
2
eigenvector; or indeed any non-zero multiple of .
5
Similarly, for the eigenvalue −1, we want a non-zero column vector ( xx12 ) such that
         
1 2 x1 x1 2 2 x1 0
= −1 ; that is, = ,
5 4 x2 x2 5 5 x2 0
   
x1 1
and we can take = to be our eigenvector.
x2 −1

Example. This example shows that the eigenvalues can depend on the field K. Let
   
0 −1 −x −1 
 = x2 + 1,
A= . Then det(A − xI2 ) = 
1 0 1 −x

so the characteristic equation is x2 + 1 = 0. If K = R (the real numbers) then this

equation has no solutions, so there are no eigenvalues or eigenvectors. However, if
K = C (the complex numbers), then there are two eigenvalues
i and −i, and by a similar
calculation to the one in the last example, we find that −1
i and ( 1i ) are eigenvectors
corresponding to i and −i respectively.

Theorem 12.3. Similar matrices have the same characteristic equation and hence the
same eigenvalues.

50
Proof. Let A and B be similar matrices. Then there exists an invertible matrix P with
B = P −1 AP . Then

det(B − xIn ) = det(P −1 AP − xIn )

= det(P −1 (A − xIn )P )
= det(P −1 ) det(A − xIn ) det(P ) (by Theorem 10.3)
= det(P −1 ) det(P ) det(A − xIn )
= det(A − xIn ).

Hence A and B have the same characteristic equation. Since the eigenvalues are the
roots of the characteristic equation, they have the same eigenvalues.

Since the different matrices corresponding to a linear map T are all equivalent, they
all have the same characteristic equation, so we can unambiguously refer to it also as
the characteristic equation of T if we want to.
There is one case where the eigenvalues can be written down immediately.

Proposition 12.4. Suppose that the matrix A is upper triangular. Then the eigenvalues
of A are just the diagonal entries αii of A.

Proof. We saw in Corollary 10.2 that the determinant of A is the product

Qn of the diagonal
entries αii . Hence the characteristic polynomial of such a matrix is i=1 (αii − x), and
so the eigenvalues are the αii .
 
1 1
Example. Let A = . Then A is upper triangular, so its only eigenvalue is 1. We
0 1
can now see that A cannot be similar to any diagonal matrix B. Such a B would also
have just 1 as an eigenvalue, and then, by Proposition 10.2 again, this would force B to
be the identity matrix I2 . But P −1 I2 P = I2 for any invertible matrix P , so I2 is not
similar to any matrix other than itself! So A cannot be similar to I2 , and hence A is
not diagonalisable.

The next theorem describes the connection between diagonalisable matrices and
eigenvectors. If you have understood everything so far then its proof should be almost
obvious.

Theorem 12.5. Let T : V → V be a linear map. Then the matrix of T is diagonal with
respect to some basis of V if and only if V has a basis consisting of eigenvectors of T .
Equivalently, let A be an n × n matrix over K. Then A is similar to a diagonal
matrix if and only if the space K n,1 has a basis of eigenvectors of A.

Proof. The equivalence of the two statements follows directly from the correspondence
between linear maps and matrices, and the corresponding definitions of eigenvectors and
eigenvalues.
Suppose that the matrix A = (αij ) of T is diagonal with respect to the basis
e1 , . . . , en of V . Recall from Section 7.1 that the images of the ith basis vector of
V is represented by the ith column of A. But since A is diagonal, this column has
the single non-zero entry αii . Hence T (ei ) = αii ei , and so each basis vector ei is an
eigenvector of A.
Conversely, suppose that e1 , . . . , en is a basis of V consisting entirely of eigenvectors
of T . Then, for each i, we have T (ei ) = λi ei for some λi ∈ K. But then the matrix of A
with respect to this basis is the diagonal matrix A = (αij ) with αii = λi for each i.

We now show that A is diagonalisable in the case when there are n distinct eigen-
values.

Theorem 12.6. Let λ1 , . . . , λr be distinct eigenvalues of T : V → V , and let v1 , . . . , vr

be corresponding eigenvectors. (So T (vi ) = λi vi for 1 ≤ i ≤ r.) Then v1 , . . . , vr are
linearly independent.

51
Proof. We prove this by induction on r. It is true for r = 1, because eigenvectors are
non-zero by definition. For r > 1, suppose that for some α1 , . . . , αr ∈ K we have

α1 v1 + α2 v2 + · · · + αr vr = 0.

Then, applying T to this equation gives

α1 λ1 v1 + α2 λ2 v2 + · · · + αr λr vr = 0.

Now, subtracting λ1 times the first equation from the second gives

α2 (λ2 − λ1 )v2 + · · · + αr (λr − λ1 )vr = 0.

By inductive hypothesis, v2 , . . . , vr are linearly independent, so αi (λi − λ1 ) = 0 for

2 ≤ i ≤ r. But, by assumption, λi − λ1 6= 0 for i > 1, so we must have αi = 0 for
i > 1. But then α1 v1 = 0, so α1 is also zero. Thus αi = 0 for all i, which proves that
v1 , . . . , vr are linearly independent.

Corollary 12.7. If the linear map T : V → V (or equivalently the n × n matrix A) has
n distinct eigenvalues, where n = dim(V ), then T (or A) is diagonalisable.

Proof. Under the hypothesis, there are n linearly independent eigenvectors, which form
a basis of V by Corollary 3.9. The result follows from Theorem 12.5.

Example.
   
4 5 2 4 − x 5 2 

A = −6 −9 −4 . Then |A − xI3 | =  −6 −9 − x −4  .
6 9 4  6 9 4 − x

To help evaluate this determinant, apply first the row operation r3 → r3 + r2 and then
the column operation c2 → c2 − c3 , giving
   
4 − x 5 2  4 − x 3 2 

|A − xI3 | =  −6 −9 − x −4  =  −6 −5 − x −4  ,
 0 −x −x  0 0 −x

and then expanding by the third row we get

|A − xI3 | = −x (4 − x)(−5 − x) + 18 = −x(x2 + x − 2) = −x(x + 2)(x − 1)

so the eigenvalues are 0, 1 and −2. Since these are distinct, we know from the above
corollary that A can be diagonalised. In fact, the eigenvectors will be the new basis with
respect to which the matrix is diagonal, so we will calculate these.  x1 
In the following calculations, we will denote eigenvectors v1 , etc. by xx2 , where
3
x1 , x2 , x3 need to be calculated by solving simultaneous equations.
For the eigenvalue λ = 0, an eigenvector v1 satisfies Av1 = 0, which gives the three
equations:

4x1 + 5x2 + 2x3 = 0; −6x1 − 9x2 − 4x3 = 0; 6x1 + 9x2 + 4x3 = 0.

The third is clearly redundant, and adding twice

  the first to the second gives 2x1 +x2 = 0
1
and then we see that one solution is v1 = −2.
3
For λ = 1, we want an eigenvector v2 with Av2 = v2 , which gives the equations

4x1 + 5x2 + 2x3 = x1 ; −6x1 − 9x2 − 4x3 = x2 ; 6x1 + 9x2 + 4x3 = x3 ;

or equivalently

3x1 + 5x2 + 2x3 = 0; −6x1 − 10x2 − 4x3 = 0; 6x1 + 9x2 + 3x3 = 0.

52
Adding the second
 and third equations gives x2 + x3 = 0 and then we see that a solution
1
is v2 = −1.
1
Finally, for λ = −2, Av3 = −2v3 gives the equations
6x1 + 5x2 + 2x3 = 0; −6x1 − 7x2 − 4x3 = 0; 6x1 + 9x2 + 6x3 = 0,
 
1
of which one solution is v3 = −2.
2
Now, if we change basis to v1 , v2 , v3 , we should get the diagonal matrix with the
eigenvalues 0, 1, −2 on the diagonal. We can check this by direct calculation. Remember
that P is the change of basis matrix from the new basis to the old one and has columns
the new basis vectors expressed in terms of the old. But the old basis is the standard
basis, so the columns of P are the new basis vectors. Hence
 
1 1 1
P = −2 −1 −2
3 1 2
 
0 0 0
and, according to Theorem 12.1, we should have P −1 AP = 0 1 0  .
0 0 −2
To check this, we first need to calculate P −1 , either by row reduction or by the
cofactor method. The answer turns out to be
 
0 1 1
P −1 =  2 1 0  ,
−1 −2 −1
and now we can check that the above equation really does hold.
Warning! The converse of Corollary 12.7 is not true. If it turns out that there do
not exist n distinct eigenvalues, then you cannot conclude from this that the matrix is
not be diagonalisable. This is really rather obvious, because the identity matrix has only
a single eigenvalue, but it is diagonal already. Even so, this is one of the most common
mistakes that students make.
If there are fewer than n distinct eigenvalues, then the matrix may or may not
be diagonalisable, and you have to test directly to see whether there are n linearly
independent eigenvectors. Let us consider two rather similar looking examples:
   
1 1 1 1 2 −2
A1 = 0 −1 1 , A2 = 0 −1 2  .
0 0 1 0 0 1
Both matrices are upper triangular, so we know from Proposition 12.4 that both
have eigenvalues 1 and −1, with 1 repeated. Since −1 occurs only once, it can only have
a single associated linearly independent eigenvector. (Can you prove
 that?)
 Solving
 
1 1
the equations as usual, we find that A1 and A2 have eigenvectors −2 and −1,
0 0
respectively, associated with eigenvalue −1.
The repeated eigenvalue 1 is more interesting, because there could be one or two
associated linearly independent eigenvectors. The equation A1 x = x gives the equations
x1 + x2 + x3 = x1 ; −x2 + x3 = x2 ; x3 = x3 ,
so x2 +x3 = −2x 2 +x 3 = 0, which implies that x2 = x3 = 0. Hence the only eigenvectors
1
are multiples of 0. Hence A1 has only two linearly independent eigenvectors in total,
0
and so it cannot be diagonalised.

53
 On the other hand, A2 x = x gives the equations

x1 + 2x2 − 2x3 = x1 ; −x2 + 2x3 = x2 ; x3 = x3 ,

which reduce to the single equation x2 − x 3 = 0. This

 time there are two linearly
1 0
independent solutions, giving eigenvectors 0 and 1. So A2 has three linearly
0 1
independent eigenvectors in total, and it can be diagonalised. In fact, using the eigen-
vectors as columns of the change of basis matrix P as before gives
   
1 0 1 1 1 −1
P = 0 1 −1 and we compute P −1 = 0 0 1  .
0 1 0 0 −1 1
 
1 0 0
We can now check that P −1 A2 P = 0 1 0 , as expected.
0 0 −1

12.3 The scalar product – symmetric and orthogonal matrices

Definition. The (standard) scalar product of two vectors v = (α1 , . . . , αn ) and w =
(β1 , . . . , βn ) in Rn is defined to be
n
X
v·w = αi βi .
i=1

Definition. A basis b1 , . . . , bn of Rn is called orthonormal if

(i) bi · bi = 1 for 1 ≤ i ≤ n, and
(ii) bi · bj = 0 for 1 ≤ i, j ≤ n and i 6= j.
In other words, an orthonormal basis consists of mutually orthogonal vectors of
length 1. For example, the standard basis is orthonormal.
Definition. An n × n matrix A is said to symmetric if AT = A.
Definition. An n × n matrix A is said to orthogonal if AT = A−1 or, equivalently, if
AAT = AT A = In .
Example.
√ √ 
 
 1/3 2/3 2/3
2/√13 −3/√ 13 2/3 −2/3 1/3 
and
3/ 13 2/ 13
2/3 1/3 −2/3
are both orthogonal matrices.
The main result of this section is that we can diagonalise any real symmetric matrix
A by a real orthogonal matrix. We shall prove this only in the case when A has distinct
eigenvalues; the complete proof will be given in Year 2.
Proposition 12.8. An n × n matrix A over R is orthogonal if and only if the rows
r1 , . . . , rn of A form an orthonormal basis of Rn , if and only if the columns c1 , . . . , cn
of A form an orthonormal basis of Rn,1 .
Proof. Note that an orthogonal matrix A is invertible, which by Theorem 9.2 implies
that its row and column ranks are equal to n, and hence that the rows of A form a basis
of Rn and the columns form a basis of Rn,1 . By the definition of matrix multiplication,
AAT = In implies that ri · ri = 1 and ri · rj = 0 for i 6= j, and hence that the rows
form an orthonormal basis of Rn . Similarly, AT A = In implies that the columns of A
form an orthonormal basis of Rn,1 . Conversely, if the rows or columns of A form an
orthonormal basis of Rn or Rn,1 , then we get AAT = In or AT A = In , both of which
imply that AT = A−1 ; that is, that A is orthogonal.

54
Proposition 12.9. Let A be a real symmetric matrix. Then A has an eigenvalue in R,
and all complex eigenvalues of A lie in R.

Proof. (To simplify the notation, we will write just v for a column vector v in this
proof.)
The characteristic equation det(A − xIn ) = 0 is a polynomial equation of degree n
in x, and since C is an algebraically closed field, it certainly has a root λ ∈ C, which is
an eigenvalue for A if we regard A as a matrix over C. We shall prove that any such λ
lies in R, which will prove the proposition.
For a column vector v or matrix B over C, we denote by v or B the result of replacing
all entries of v or B by their complex conjugates. Since the entries of A lie in R, we
have A = A.
Let v be a complex eigenvector associated with λ. Then

Av = λv (1)

so, taking complex conjugates and using A = A, we get

Av = λv. (2)

Transposing (1) and using AT = A gives

vT A = λvT , (3)

so by (2) and (3) we have

λvT v = vT Av = λvT v.

But if v = (α1 , α2 , . . . , αn )T , then vT v = α1 α1 + · · · + αn αn , which is a nonzero real

number (eigenvectors are nonzero by definition). Thus λ = λ, so λ ∈ R.

Proposition 12.10. Let A be a real symmetric matrix, and let λ1 , λ2 be two distinct
eigenvalues of A, with corresponding eigenvectors v1 , v2 . Then v1 · v2 = 0.

Proof. (As in Proposition 12.9, we will write v rather than v for a column vector in this
proof. So v1 · v2 is the same as v1T v2 .) We have

Av1 = λ1 v1 (1) and Av2 = λ2 v2 (2).

Transposing (1) and using A = AT gives v1T A = λ1 v1T , and so

v1T Av2 = λ1 v1T v2 (3) and similarly v2T Av1 = λ2 v2T v1 (4).

Transposing (4) gives v1T Av2 = λ2 v1T v2 and subtracting (3) from this gives (λ2 −
λ1 )v1T v2 = 0. Since λ2 − λ1 6= 0 by assumption, we have v1T v2 = 0.

Combining these results, we obtain the following theorem.

Theorem 12.11. Let A be a real symmetric n × n matrix. Then there exists a real
orthogonal matrix P with P −1 AP (= P T AP ) diagonal.

Proof. We shall prove this only in the case when the eigenvalues λ1 , . . . , λn of A are all
distinct. By Proposition 12.9 we have λi ∈ R for all i, and so there exist associated
eigenvectors vi ∈ Rn,1 . By Proposition 12.10, we have vi · vj = viT vj = 0 for i 6=
j. Since each vi is non-zero, we have vi · vi = αi > 0. By replacing each vi by
√
vi / αi (which is also an eigenvector for λi ), we can assume that vi · vi = 1 for all i.
Since, by Theorem 12.6, the vi are linearly independent, they form a basis and hence
an orthonormal basis of Rn,1 . So, by Proposition 12.8, the matrix P with columns
v1 , . . . , vn is orthogonal. But P −1 AP is the diagonal matrix with entries λ1 , . . . , λn ,
which proves the result.

55
Example. Let  
1 3
A= .
3 1
Then
det(A − λI2 ) = (1 − λ)2 − 9 = λ2 − 2λ − 8 = (λ − 4)(λ + 2),
so the eigenvalues of A are 4 and −2.
Solving Av = λv for λ = 4 and −2, we find
1
corresponding eigenvectors ( 11 ) and −1 . Proposition 12.10 tells us that these vectors
are orthogonal
√ to each other (which we can of course check directly!). Their lengths are
both 2, so so we divide by them by their lengths to give eigenvectors
 √   √ 
1/√2 1/ √2
and
1/ 2 −1/ 2

of length 1.
The basis change matrix P has these vectors as columns, so
 √ √ 
1/√2 1/ √2
P = ,
1/ 2 −1/ 2

and we can check that P T P = I2 (i.e. P is orthogonal) and that

 
−1 T 4 0
P AP = P AP = .
0 −2

56