Linear Algebra I

Richard Earl

Michaelmas Term 2024

0. PRELIMINARY MATERIAL

0.1 Syllabus

Systems of linear equations. Matrices and the beginnings of matrix algebra. Use of matrices
to describe systems of linear equations. Elementary Row Operations (EROs) on matrices. Re-
duction of matrices to echelon form. Application to the solution of systems of linear equations.
[2.5]
Inverse of a square matrix. The use of EROs to compute inverses; computational efficiency of
the method. Transpose of a matrix; orthogonal matrices. [1]
Vector spaces: definition of a vector space over a field (such as R, Q, C). Subspaces. Many
explicit examples of vector spaces and subspaces. [1.5]
Span of a set of vectors. Examples such as row space and column space of a matrix. Linear
dependence and independence. Bases of vector spaces; examples. The Steinitz Exchange
Lemma; dimension. Application to matrices: row space and column space, row rank and
column rank. Coordinates associated with a basis of a vector space. [2]
Use of EROs to find bases of subspaces. Sums and intersections of subspaces; the dimension
formula. Direct sums of subspaces. [1.5]
Linear transformations: definition and examples (including projections associated with direct-
sum decompositions). Some algebra of linear transformations; inverses. Kernel and image,
Rank-Nullity Theorem. Applications including algebraic characterisation of projections (as
idempotent linear transformations). [2]
Matrix of a linear transformation with respect to bases. Change of Bases Theorem. Applications
including proof that row rank and column rank of a matrix are equal. [2]
Bilinear forms; real inner product spaces; examples. Mention of complex inner product spaces.
Cauchy–Schwarz inequality. Distance and angle. The importance of orthogonal matrices. [1.5]

PRELIMINARY MATERIAL 1
0.2 Reading list

(1) Gilbert Strang, Introduction to linear algebra (Fifth edition, Wellesley-Cambridge 2016).
http://math.mit.edu/˜gs/linearalgebra/
(2) T.S. Blyth and E.F. Robertson, Basic linear algebra (Springer, London, 1998).
Further Reading:
(3) Richard Kaye and Robert Wilson, Linear algebra (OUP, Oxford 1998), Chapters 1-5 and 8.
(4) Charles W. Curtis, Linear algebra - an introductory approach (Springer, London, Fourth
edition, reprinted 1994).
(5) R. B. J. T. Allenby, Linear algebra (Arnold, London, 1995).
(6) D. A. Towers, A guide to linear algebra (Macmillan, Basingstoke, 1988).
(7) Seymour Lipschutz and Marc Lipson, Schaum’s outline of linear algebra (McGraw Hill, New
York & London, Fifth edition, 2013).

READING LIST 2
1. LINEAR SYSTEMS AND MATRICES

1.1 Systems of linear equations

Definition 1 (a) By a linear system, or linear system of equations, we will mean a set
of m simultaneous equations in n real variables x1 , x2 , . . . , xn which are of the form

a11 x1 + a12 x2 + ··· + a1n xn = b1 ;

a21 x1 + a22 x2 + ··· + a2n xn = b2 ;
.. .. .. .. (1.1)
. . . .
am1 x1 + am2 x2 + · · · + amn xn = bm ,

where aij and bi are real constants.

(b) Any vector (x1 , x2 , . . . , xn ) which satisfies (1.1) is said to be a solution; if the linear
system has one or more solutions then it is said to be consistent. The general solution to
the system is any description of all the solutions of the system. We will see, in due course, that
such linear systems can have zero, one or infinitely many solutions.
(c) We will often write the linear system (1.1) as the augmented matrix (A | b) where
   
a11 a12 · · · a1n b1
 a21 a22 · · · a2n   b2 
A =  .. .. , b =  .. .
   
..
 . . .   . 
am1 am2 · · · amn bm

For now, we won’t consider a matrix (such as A) or vector (such as b) to be anything more
than an array of numbers.

Consider as a first example the following linear system of 3 equations in 3 variables.

Example 2 Determine the solutions (if any) to the following equations.

3x + y − 2z = −2; x + y + z = 2; 2x + 4y + z = 0.

Solution. We can substitute z = 2 − x − y from the second equation into the first and third
to find

3x + y − 2(2 − x − y) = 5x + 3y − 4 = −2 =⇒ 5x + 3y = 2;
2x + 4y + (2 − x − y) = x + 3y + 2 = 0 =⇒ x + 3y = −2.

Subtracting the second of these equations from the first gives 4x = 4 and so we see

x = 1, y = (−2 − x)/3 = −1, z = 2 − x − y = 2. (1.2)

LINEAR SYSTEMS AND MATRICES 3

Thus there is a unique solution (x, y, z) = (1, −1, 2). We can verify easily that this is indeed
a solution (just to check that the system contains no contradictory information elsewhere that
we haven’t used).
Whilst we solved the above rigorously – we showed of necessity (1, −1, 2) was the only
possible solution and then verified it is a solution – our approach was a little ad hoc; at least,
it’s not hard to appreciate that if we were presented with 1969 equations in 2021 variables then
we would need a much more systematic approach to treat them – or more likely we would need
to be more methodical while programming our computers to determine any solutions for us.
We introduce such a process called row-reduction here.
We first improve the notation, writing the system as an augmented matrix.
 
3 1 −2 −2
 1 1 1 2 . (1.3)
2 4 1 0
All that has been lost in this representation are the names of the variables, but these names are
unchanging and unimportant in the actual handling of the equations. The advantages, we shall
see, are that we will be able to progress systematically towards any solution and at each stage
we shall retain all the information that the system contains – any redundancies (superfluous,
unnecessary equations) or contradictions will naturally appear as part of the calculation.
This process is called row-reduction. It relies on three types of operation, called elementary
row operations or EROs, which importantly do not affect the set of solutions of a linear system
as we apply them.
Definition 3 Given a linear system of equations, an elementary row operation or ERO
is an operation of one of the following three kinds.
(a) The ordering of two equations (or rows) may be swapped – for example, one might
reorder the writing of the equations so that the first equation now appears third and vice versa.
(b) An equation may be multiplied by a non-zero scalar – for example, one might replace
2x − y + z = 3 by x − 12 y + 21 z = 23 from multiplying both sides of the equation by 12 .
(c) A multiple of one equation might be added to a different equation – for example, one
might replace the second equation by the second equation plus twice the third equation.
Notation 4 (a) Let Sij denote the ERO which swaps rows i and j (or equivalently the ith and
jth equations).
(b) Let Mi (λ) denote the ERO which multiplies row i by λ ̸= 0 (or equivalently both sides
of the ith equation).
(c) For i ̸= j, let Aij (λ) denote the ERO which adds λ times row i to row j (or does the
same to the equations).
Note this is not standard notation in any way, but I’ve introduced it here for
convenience.
All these operations may well seem uncontroversial (their validity will be shown in Corollary
40) but it is probably not yet clear that these three simple operations are powerful enough to
reduce any linear system to a point where any solutions can just be read off (Proposition 44,
Theorem 47). Before treating the general case, we will see how the three equations in (1.3) can
be solved using EROs to get an idea of the process.

SYSTEMS OF LINEAR EQUATIONS 4

Example 5 Find all solutions of the linear system (1.3).
Solution. If we use S12 to swap the first two rows the system becomes
   
3 1 −2 −2 1 1 1 2
S12
 1 1 1 2  −→  3 1 −2 −2 .
2 4 1 0 2 4 1 0
Now subtract three times the first row from the second, i.e. A12 (−3) and follow this by sub-
tracting twice the first row from the third, i.e. A13 (−2), so that
     
1 1 1 2 1 1 1 2 1 1 1 2
 3 1 −2 −2  A−→ 12 (−3)
 0 −2 −5 −8  A−→ 13 (−2)
 0 −2 −5 −8 . (1.4)
2 4 1 0 2 4 1 0 0 2 −1 −4
We can now divide the second row by −2, i.e. M2 (−1/2) to find
   
1 1 1 2 1 1 1 2
 0 −2 −5 −8  M2−→ (−1/2)
 0 1 2 12 4 .
0 2 −1 −4 0 2 −1 −4
We then subtract the second row from the first, i.e. A21 (−1), and follow this by subtracting
twice the second row from the third, i.e. A23 (−2), to obtain
1 0 −1 12 −2 0 −1 12
     
1 1 1 2 1 −2
A21 (−1) A23 (−2)
 0 1 21 4  −→  0 1 22 1
4  −→  0 1 2 12 4 . (1.5)
2
0 2 −1 −4 0 2 −1 −4 0 0 −6 −12
If we divide the third row by −6, i.e. M3 (−1/6), the system becomes
1 0 −1 12 −2 1 0 −1 12 −2
   
M3 (−1/6)
 0 1 21 4  −→  0 1 2 12 4 .
2
0 0 −6 −12 0 0 1 2
Finally, we subtract 2 12 times the third row from the second, i.e. A32 (−2 21 ), and follow this by
adding 1 21 times the third row to the first, i.e. A31 (1 21 ).
1 0 −1 12 -2 1 0 −1 21 −2
     
1 0 0 1
A32 (−5/2) A31 (3/2)
 0 1 21 4  −→  0 1 0 −1  −→  0 1 0 −1 .
2
0 0 1 2 0 0 1 2 0 0 1 2
The rows of the final matrix represent the equations x = 1, y = −1, z = 2 as expected from
(1.2).
Remark 6 In case the systematic nature of the previous example isn’t apparent, note that
the first three operations S12 , A12 (−3), A13 (−2) were chosen so that the first column became
(1, 0, 0)T in (1.4). There were many other ways to achieve this: for example, we could have
begun with M1 (1/3) to divide the first row by 3, then used A12 (−1) and A13 (−2) to clear the
rest of the column. Once done, we then produced a similar leading entry of 1 in the second
row with M2 (−1/2) and used A21 (−1) and A23 (−2) to turn the second column into (0, 1, 0)T
in (1.5). The final three EROs were chosen to transform the third column to (0, 0, 1)T at which
point we could simply read off the solutions.

SYSTEMS OF LINEAR EQUATIONS 5

 Here are two slightly different examples, the first where we find that there are infinitely
many solutions, whilst in the second example we see that there are no solutions.

Example 7 Find the general solution of the following systems of equations in variables x1 , x2 , x3 , x4 .

(a) x1 − x2 + x3 + 3x4 = 2; 2x1 − x2 + x3 + 2x4 = 4; 4x1 − 3x2 + 3x3 + 8x4 = 8.

(b) x1 + x2 + x3 + x4 = 4; 2x1 + 3x2 − 2x3 − 3x4 = 1; x1 + 5x3 + 6x4 = 1.

Solution. (a) This time we will not spell out at quite so much length which EROs are being
used. But we continue in a similar vein to the previous example and proceed by the method
outlined in Remark 6.
  A (−2)   A (1)  
1 −1 1 3 2 12 1 −1 1 3 2 21 1 0 0 −1 2
 2 −1 1 2 4  A−→ 13 (−4)
 0 1 −1 −4 0  A−→ 23 (−1)
 0 1 −1 −4 0 .
4 −3 3 8 8 0 1 −1 −4 0 0 0 0 0 0

We have manipulated our system of three equations to two equations equivalent to the original
system, namely
x1 − x4 = 2; x2 − x3 − 4x4 = 0. (1.6)
The presence of the zero row in the last matrix means that there was some redundancy in the
system. Note, for example that the third equation can be deduced from the first two (it’s the
second equation added to twice the first) and so it provides no new information. As there are
now only two equations in four variables, it’s impossible for each column to contain a row’s
leading entry. In this example, the third and fourth columns lack such an entry. To describe
all the solutions to a consistent system, we assign parameters to the columns/variables without
leading entries. In this case that’s x3 and x4 and we’ll assign parameters by setting x3 = s,
x4 = t, and then use the two equations in (1.6) to read off x1 and x2 . So

x1 = t + 2, x2 = s + 4t, x3 = s, x4 = t, (1.7)

or we could write

(x1 , x2 , x3 , x4 ) = (t + 2, s + 4t, s, t) = (2, 0, 0, 0) + s(0, 1, 1, 0) + t(1, 4, 0, 1). (1.8)

For each choice of s and t we have a solution as in (1.7) and this is one way of representing the
general solution. (1.8) makes more apparent that these solutions form a plane in R4 , a plane
which passes through (2, 0, 0, 0) is parallel to (0, 1, 1, 0) and (1, 4, 0, 1) with s, t parametrizing
the plane.
(b) Applying EROs again in a like manner, we find
  A (−2)  
1 1 1 1 4 12 1 1 1 1 4
2 3 −2 −3 1  A−→ 13 (−1)
0 1 −4 −5 −7 
1 0 5 6 1 0 −1 4 5 −3
  M (−1/10)  
1 1 1 1 4 3 1 0 −5 −6 −11
A23 (1) A21 (−1)
−→ 0 1 −4 −5 −7  −→ 0 1 −4 −5 −7 .
0 0 0 0 −10 0 0 0 0 1

SYSTEMS OF LINEAR EQUATIONS 6

Note that any x = (x1 , x2 , x3 , x4 ) which solves the final equation must satisfy
0x1 + 0x2 + 0x3 + 0x4 = 1.
There clearly are no such xi and so there are no solutions to this equation. Any solution to
the system has, in particular, to solve the third equation and so this system has no solutions.


In fact, this was all apparent once the third row had become 0 0 0 0 −10 as the
equation it represents is clearly insolvable also. The final two EROs were simply done to put
the matrix into what is called reduced row echelon form (see Definition 41).
Examples 5, 7(a) and 7(b) are specific examples of the following general cases.
• A linear system can have no, one or infinitely many solutions.
We shall prove this in due course (Proposition 44). We finish our examples though with
a linear system that involves a parameter – so really we have a family of linear systems, one
for each value of that parameter. What EROs may be permissible at a given stage may well
depend on the value of the parameter and so we may see (as below) that such a family can
exhibit all three of the possible scenarios just described.

Example 8 Consider the system of equations in x, y, z,

x + z = −5; 2x + αy + 3z = −9; −x − αy + αz = α2 ,
where α is a constant. For which values of α has the system one solution, none or infinitely
many?
Solution. Writing this system in matrix form and applying EROs we can argue as follows.
  A (−2)    
1 0 1 −5 12 1 0 1 −5 1 0 1 −5
A13 (1)
 2 α 3 −9  −→  0 α 1 1  A−→
23 (1)
 0 α 1 1 .
2 2 2
−1 −α α α 0 −α α + 1 α − 5 0 0 α+2 α −4
(1.9)
At this point, which EROs are permissible depends on the value of α. We would like to divide
the second equation by α and the third by α + 2. Both these are permissible provided that
α ̸= 0 and α ̸= −2. We will have to treat separately those particular cases but, assuming for
now that α ̸= 0, −2, we obtain
  M (1/α)   A (−1)  
1 0 1 −5 2 1 0 1 −5 31 1 0 0 −α − 3
 0 α 1 1  M3 (1/(α+2))
−→  0 1 1/α 1/α  A32−→ (−1/α)
 0 1 0 3/α − 1 
2
0 0 α+2 α −4 0 0 1 α−2 0 0 1 α−2
and we see that the system has a unique solution when α ̸= 0, −2. Returning though to the
last matrix of (1.9) for our two special cases, we would proceed as follows.
  A (−2)  
1 0 1 −5 23 1 0 1 −5
α = 0:  0 0 1 1  M3−→ (−1/6)
 0 0 1 1 .
0 0 2 −4 0 0 0 1
   
1 0 1 −5 1 0 1 −5
α = −2 :  0 −2 1 1  M2−→ (−1/2)
 0 1 −1/2 −1/2 .
0 0 0 0 0 0 0 0

SYSTEMS OF LINEAR EQUATIONS 7

We see then that the system is inconsistent when α = 0 (because of the insolvability of the
third equation) whilst there are infinitely many solutions x = −5 − t, y = (t − 1)/2, z = t, when
α = −2. We assign a parameter, here t, to the variable z as the third column has no leading
entry.
Before we treat linear systems more generally, we will first need to discuss matrices and
their algebra.

1.2 Matrices and matrix algebra

At its simplest, a matrix is just a two-dimensional array of numbers; for example

 
  1  
√1 2 −3 ,  −1.2 , 0 0
(1.10)
2 π 0 0 0
−1

are all matrices. The examples above are respectively a 2 × 3 matrix, a 3 × 1 matrix and a
2 × 2 matrix (read ‘2 by 3’ etc.); the first figure refers to the number of horizontal rows and the
second to the number of vertical columns in the matrix. Row vectors in Rn are 1 × n matrices
and columns vectors in Rncol are n × 1 matrices.

Definition 9 Let m, n be positive integers. An m × n matrix is an array of real numbers

arranged into m rows and n columns.
√

Example 10 Consider the first matrix above. Its second row is 2 π 0 and its third
 
−3
column is .
0

Definition 11 The numbers in a matrix are its entries. Given an m × n matrix A, we will
write aij for the entry in the ith row and jth column. Note that i can vary between 1 and m,
and that j can vary between 1 and n. So
 
a1j
ith row = (ai1 , . . . , ain ) and jth column =  ... .
 
amj

Notation 12 We shall denote the set of real m × n matrices as Mmn . Note that M1n = Rn
and that Mn1 = Rncol .

Example 13 If we write A for the first matrix in (1.10) then we have a23 = 0 and a12 = 2.

There are three important operations that can be performed with matrices: matrix addition,
scalar multiplication and matrix multiplication. As with vectors, not all pairs of matrices can
be meaningfully added or multiplied.

MATRICES AND MATRIX ALGEBRA 8

Definition 14 Addition Let A = (aij ) be an m×n matrix (recall: m rows and n columns) and
B = (bij ) be a p × q matrix. As with vectors, matrices are added by adding their corresponding
entries. So, as with vectors, to add two matrices they have to be the same size – that is, to add
A and B, we must have m = p and n = q. If we write C = A + B = (cij ) then
cij = aij + bij for 1 ⩽ i ⩽ m and 1 ⩽ j ⩽ n.
Example 15 Let
     
1 2 1 2 3 1 −1
A= , B= , C= . (1.11)
−1 0 3 2 1 1 −1
| {z } | {z } | {z }
2×2 2×3

Of the possible sums involving these matrices, only A + C and C + A make sense as B is a
different size. Note that  
2 1
A+C = = C + A.
0 −1
Remark 16 In general, matrix addition is commutative as for matrices M and N of the
same size we have
M + N = N + M.
Addition of matrices is also associative as
L + (M + N ) = (L + M ) + N
for any matrices of the same size.
Definition 17 The m × n zero matrix is the matrix with m rows and n columns whose every
entry is 0. This matrix is simply denoted as 0 unless we need to specify its size, in which case
it is written 0mn . For example,  
0 0 0
023 = .
0 0 0
A simple check shows that A + 0mn = A = 0mn + A for any m × n matrix A.
Definition 18 Scalar Multiplication Let A = (aij ) be an m × n matrix and k be a real
number (a scalar). Then the matrix kA is defined to be the m × n matrix with (i,j)th entry
equal to kaij .
Example 19 Show that 2(A + B) = 2A + 2B for the following matrices:
   
1 2 0 −2
A= ; B= .
3 4 5 1
Solution. Here we are checking the distributive law in a specific example. We note that
   
1 0 2 0
A+B = , and so 2(A + B) = ;
8 5 16 10
     
2 4 0 −4 2 0
2A = , and 2B = , so 2A + 2B = .
6 8 10 2 16 10

MATRICES AND MATRIX ALGEBRA 9

Remark 20 More generally the following identities hold. Let A, B, C be m × n matrices and
λ, µ be real numbers.

A + 0mn = A; A + B = B + A; 0A = 0mn ;
A + (−A) = 0mn ; (A + B) + C = A + (B + C); 1A = A;
(λ + µ)A = λA + µA; λ(A + B) = λA + λB; λ(µA) = (λµ)A.

These are readily verified and show that Mmn is a real vector space. ■

Based on how we added matrices then you might think that we multiply matrices in a similar
fashion, namely multiplying corresponding entries, but we do not. At first glance the rule for
multiplying matrices is going to seem rather odd but, in due course, we will see why matrix
multiplication is done as follows and that this is natural in the context of matrices representing
linear maps.

Definition 21 Matrix Multiplication We can multiply an m × n matrix A = (aij ) with an

p × q matrix B = (bij ) if n = p. That is, A must have as many columns as B has rows. If this
is the case then the product C = AB is the m × q matrix with entries
n
X
cij = aik bkj for 1 ⩽ i ⩽ m and 1 ⩽ j ⩽ q. (1.12)
k=1

It may help to write the rows of A as r1 , . . . , rm and the columns of B as c1 , . . . , cq . Rule (1.12)
then states that

the (i, j)th entry of AB = ri · cj for 1 ⩽ i ⩽ m and 1 ⩽ j ⩽ q. (1.13)

We dot (i.e. take the scalar product of ) the rows of A with the columns of B; specifically to
find the (i, j)th entry of AB we dot the ith row of A with the jth column of B.

Remark 22 We shall give full details later as to why it makes sense (and, in fact, is quite
natural) to multiply matrices as in (1.12). For now, it is worth noting the following. Let A be
an m × n matrix and B be n × p so that AB is m × p. There is a map LA from Rncol to Rm col
associated with A, as given an n × 1 column vector v in Rncol then Av is a m × 1 column vector
in Rm
col . (Here the L denotes that we are multiplying on the left or premultiplying.) So we have
associated maps

LA from Rncol to Rm
col , LB from Rpcol to Rncol , LAB from Rpcol to Rm
col .

Multiplying matrices as we have, it turns out that

LAB = LA ◦ LB .

This is equivalent to (AB)v = A(Bv) which follows from the associativity of matrix multipli-
cation. So matrix multiplication is best thought of as composition: performing LAB is equal to
the performing LB then LA . ■

MATRICES AND MATRIX ALGEBRA 10

Example 23 Calculate the possible products of the pairs of matrices in (1.11).
Solution. Recall that a matrix product M N makes sense if M has the same number of columns
as N has rows. A, B, C are respectively 2 × 2, 2 × 3, 2 × 2 matrices. so the products we can
form are AA, AB, AC, CA, CB, CC. Let’s slowly go through the product AC.
      
1 2 1 −1 1 × 1 + 2 × 1 ?? 3 ??
= = .
−1 0 1 −1 ?? ?? ?? ??
This is how we calculate the (1, 1)th entry of AC. We take the first row of A and the first
column of C and dot them together. We complete the remainder of the product as follows:
      
1 2 1 −1 2 1 × (−1) + 2 × (−1) 3 −3
= = ;
−1 0 1 −1 ?? ?? ?? ??
      
1 2 1 −1 2 4 3 −3
= = ;
−1 0 1 −1 (−1) × 1 + 0 × 1 ?? −1 ??
      
1 2 1 −1 2 4 3 −3
= = .
−1 0 1 −1 0 (−1) × (−1) + 0 × (−1) −1 1
So finally     
1 2 1 −1 3 −3
= .
−1 0 1 −1 −1 1
We complete the remaining examples more quickly but still leaving a middle stage in the
calculation to help see the process.
      
1 2 1 2 1−2 2+0 −1 2
AA = = = ;
−1 0 −1 0 −1 + 0 −2 + 0 −1 −2
      
1 2 1 2 3 1+6 2+4 3+2 7 6 5
AB = = = ;
−1 0 3 2 1 −1 + 0 −2 + 0 −3 + 0 −1 −2 −3
      
1 −1 1 2 1+1 2−0 2 2
CA = = = ;
1 −1 −1 0 1+1 2−0 2 2
      
1 −1 1 2 3 1−3 2−2 3−1 −2 0 2
CB = = = ;
1 −1 3 2 1 1−3 2−2 3−1 −2 0 2
      
1 −1 1 −1 1 − 1 −1 + 1 0 0
CC = = = .
1 −1 1 −1 1 − 1 −1 + 1 0 0

Definition 24 The n × n identity matrix In is the n × n matrix with entries


1 if i = j,
δij =
0 if i ̸= j.
For example,  
  1 0 0
1 0
I2 = , I3 =  0 1 0 .
0 1
0 0 1
The identity matrix will be simply denoted as I unless we need to specify its size. The (i, j)th
entry of I is denoted as δij which is referred to as the Kronecker delta.

MATRICES AND MATRIX ALGEBRA 11

Remark 25 (Sifting Property of the Kronecker Delta) Let x1 , . . . , xn be n real numbers,
and 1 ⩽ k ⩽ n. Then
Xn
xi δik = xk .
i=1
This is because δik = 0 when i ̸= k and δkk = 1. Thus the above sum sifts out (i.e. selects) the
kth element xk . ■
There are certain important points to highlight from Example 23, some of which make
matrix algebra crucially different from the algebra of real numbers.
Proposition 26 (Properties of Matrix Multiplication) (a) For an m × n matrix A and
positive integers l, p,
A0np = 0mp ; 0lm A = 0ln ; AIn = A; Im A = A.
(b) Matrix multiplication is not commutative; AB ̸= BA in general, even if both products
meaningfully exist and have the same size.
(c) Matrix multiplication is associative; for matrices A, B, C, which are respectively m×n, n×
p and p × q we have
A(BC) = (AB)C.
(d) The distributive laws hold for matrix multiplication; whenever the following products and
sums make sense,
A(B + C) = AB + AC, and (A + B)C = AC + BC.
(e) In Example 23 we saw CC = 0 even though C ̸= 0 – so one cannot conclude from M N = 0
that either matrix M or N is zero.
Proof. (a) To find an entry of the product A0np we dot a row of A with a zero column of 0np
and likewise in the product 0lm A we are dotting with zero rows. Also, by the sifting property,
X
n
the (i, j)th entry of AIn = k=1 aik δkj = aij ;
X
n
the (i, j)th entry of In A = k=1 δik akj = aij .

(b) In Example 23, we saw that AC ̸= CA. More generally, if A is m × n and B is n × p then
the product AB exists but BA doesn’t even make sense as a matrix product unless m = p.
(c) Given i, j in the ranges 1 ⩽ i ⩽ m, 1 ⩽ j ⩽ q, we see
p n
!
X X
the (i, j)th entry of (AB)C = ais bsr crj ;
r=1 s=1
p
n
!
X X
the (i, j)th entry of A(BC) = ais bsr crj .
s=1 r=1

These are equal as the order of finite sums may be swapped.

(d) This is left as an exercise.
Because matrix multiplication is not commutative, we need to be clearer than usual in what
we might mean by a phrase like ‘multiply by the matrix A’; typically we need to give some
context as to whether we have multiplied on the left or on the right.

MATRICES AND MATRIX ALGEBRA 12

Definition 27 Let A and M be matrices.
(a) To premultiply M by A is to form the product AM – i.e. premultiplication is multiplication
on the left.
(b) To postmultiply M by A is to form the product M A – i.e. postmultiplication is multipli-
cation on the right.

Notation 28 We write A2 for the product AA and similarly, for n a positive integer, we write
An for the product
AA · · · A}.
| {z
n times

Note that A must be a square matrix for this to make sense. We also define A0 = I. Note that
Am An = Am+n for natural numbers m, n. Given a polynomial p(x) = ak xk + ak−1 xk−1 + · · · +
a1 x + a0 , then we define

p(A) = ak Ak + ak−1 Ak−1 + · · · + a1 A + a0 I.

Example 29 Let
   
cos α sin α 0 1
A= and B= . (1.14)
sin α − cos α 0 0

Then A2 = I2 for any choice of α. Also there is no matrix C (with real or complex entries)
such that C 2 = B. This shows that the idea of a square root is a much more complicated issue
for matrices than for real or complex numbers. A square matrix may have none or many, even
infinitely many, different square roots.

Solution. We note for any α that

cos2 α + sin2 α
   
2 cos α sin α − sin α cos α 1 0
A = = = I2 .
sin α cos α − cos α sin α sin2 α + (− cos α)2 0 1

To show B has no square roots, say a, b, c, d are real (or complex) numbers such that
   2  
0 1 a b a2 + bc b(a + d)
B= = = .
0 0 c d c(a + d) bc + d2

Looking at the (2, 1) entry, we see c = 0 or a + d = 0. But a + d = 0 contradicts b(a + d) = 1

from the (1, 2) entry and so c = 0. From the (1, 1) entry we see a = 0 and from the (2, 2) entry
we see d = 0, but these lead to the same contradiction.
Let’s look at a simple case of simultaneous equations: 2 linear equations in two variables,
such as
ax + by = e; cx + dy = f. (1.15)
Simple algebraic manipulations show that typically there is a unique solution (x, y) given by

de − bf af − ce
x= ; y= . (1.16)
ad − bc ad − bc

MATRICES AND MATRIX ALGEBRA 13

However if ad − bc = 0 then this solution is meaningless. It’s probably easiest to appreciate
geometrically why this is: the equations in (1.15) represent lines in the xy-plane with gradients
−a/b and −c/d respectively, and hence the two lines are parallel if ad − bc = 0. (Notice that
this is still the correct condition when b = d = 0 and the lines are parallel and vertical.) If the
lines are parallel then there cannot be a unique solution.
We can represent the two scalar equations in (1.15) and (1.16) by a single vector equation
in each case:
    
a b x e
= ; (1.17)
c d y f
    
x 1 d −b e
= . (1.18)
y ad − bc −c a f

Equation (1.17) is just a rewriting of the linear system (1.15). Equation (1.18) is a similar
rewriting of the unique solution found in (1.16) and something we typically can do. It also
introduces us to the notion of the inverse of a matrix. Note that
     
d −b a b a b d −b
= (ad − bc)I2 = . (1.19)
−c a c d c d −c a

So if ad − bc ̸= 0 and we set
   
a b 1 d −b
A= and B= ,
c d ad − bc −c a

then BA = I2 and AB = I2 .

Definition 30 Let A be a square matrix. We say that B is an inverse of A if BA = AB = I.

We refer to a matrix with an inverse as invertible and otherwise the matrix is said to be
singular.

Proposition 31 (Properties of Inverses)

(a) (Uniqueness) If a square matrix A has an inverse, then it is unique. We write A−1 for
this inverse.
(b) (Product Rule) If A, B are invertible n × n matrices then AB is invertible with (AB)−1 =
B −1 A−1 .
−1
(c) (Involution Rule) If A is invertible then so is A−1 with (A−1 ) = A.

Proof. (a) Suppose B and C were two inverses for an n × n matrix A then

C = In C = (BA)C = B(AC) = BIn = B

as matrix multiplication is associative. Part (b) is left as Sheet 1, Exercise S3. To verify (c)
note that
A−1 A = A(A−1 ) = I

−1
and so (A−1 ) = A by uniqueness.

MATRICES AND MATRIX ALGEBRA 14

Definition 32 If A is m × n and BA = In then B is said to be a left inverse; if C satisfies
AC = Im then C is said to be a right inverse.

• If A is m × n where m ̸= n then A cannot have both left and right inverses. (This is
non-trivial. We will prove this later.)

• If A, B are n × n matrices with BA = In then, in fact, AB = In (Proposition 168).

Inverses, in the 2 × 2 case, are a rather simple matter to deal with.

 
a b
Proposition 33 The matrix A = has an inverse if and only if ad − bc ̸= 0. If
c d
ad − bc ̸= 0 then  
−1 1 d −b
A = .
ad − bc −c a

Remark 34 The scalar quantity ad − bc is called the determinant of A, written det A. It is

a non-trivial fact to show that a square matrix is invertible if and only if its determinant is
non-zero. This will be proved in Linear Algebra II next term. ■

Proof. We have already seen in (1.19) that if ad − bc ̸= 0 then AA−1 = I2 = A−1 A. If however
ad − bc = 0 then  
d −b
B=
−c a
satisfies BA = 0. If an inverse C for A existed then, by associativity, 0 = 0C = (BA)C =
B(AC) = BI2 = B. So each of a, b, c and d would be zero and consequently A = 0 which
contradicts AC = I2 .
We conclude this section with the following theorem. The proof demonstrates the power
of the sigma-notation for matrix multiplication introduced in (1.12) and that of the Kronecker
delta. In this proof we will make use of the standard basis for matrices.

Notation 35 For I, J in the range 1 ⩽ I ⩽ m, 1 ⩽ J ⩽ n, we denote by EIJ the m × n matrix

with entry 1 in the Ith row and Jth column and 0s elsewhere. Then

the (i, j) th entry of EIJ = δIi δJj

as δIi δJj = 0 unless i = I and j = J in which case it is 1. These matrices form the standard
basis for Mmn .

Theorem 36 Let A be an n × n matrix such that AM = M A for all n × n matrices M . i.e.

A commutes with all n × n matrices. Then A = λIn for some real number λ.

MATRICES AND MATRIX ALGEBRA 15

Proof. As A commutes with every n × n matrix, then in particular it commutes with each of
the n2 basis matrices EIJ . So the (i, j)th entry of AEIJ equals that of EIJ A for every I, J, i, j.
Using the sifting property
n
X
the (i, j) th entry of AEIJ = aik δIk δJj = aiI δJj ;
k=1
n
X
the (i, j) th entry of EIJ A = δIi δJk akj = δIi aJj .
k=1

Hence for all I, J, i, j,

aiI δJj = δIi aJj . (1.20)
Let i ̸= j. If we set I = J = i, then (1.20) becomes 0 = aij showing that the non-diagonal
entries of A are zero. If we set I = i and J = j, then (1.20) becomes aii = ajj , which shows that
all the diagonal entries of A are equal – call this shared value λ and we have shown A = λIn .
This shows that any such M is necessarily of the form λIn , and conversely such matrices do
indeed commute with every other n × n matrix.

1.3 Reduced Row Echelon Form

Now looking to treat linear systems more generally, we will first show that the set of solutions
of a linear system does not change under the application of EROs. We shall see that applying
any ERO to a linear system (A|b) is equivalent to premultiplying by an invertible elementary
matrix E to obtain (EA|Eb), and it is the invertibility of elementary matrices that means the
set of solutions remains unchanged when we apply EROs.

Proposition 37 (Elementary Matrices) Let A be an m × n matrix. Applying any of the

EROs SIJ , MI (λ) and AIJ (λ) is equivalent to pre-multiplying A by certain matrices which we
also denote as SIJ , MI (λ) and AIJ (λ). Specifically these matrices have entries


 1 i = j ̸= I, J,
1 i = J, j = I,

the (i, j)th entry of SIJ =

 1 i = I, j = J,
0 otherwise.


 1 i = j ̸= I,
the (i, j)th entry of MI (λ) = λ i = j = I,
0 otherwise.


 1 i = j,
the (i, j)th entry of AIJ (λ) = λ i = J, j = I,
0 otherwise


The above matrices are known as elementary matrices.

REDUCED ROW ECHELON FORM 16

Proof. The proof is left as an exercise.

Example 38 When m = 3 we see

     
0 1 0 1 0 0 1 0 −2
S21 =  1 0 0 , M3 (7) =  0 1 0 , A31 (−2) =  0 1 0 .
0 0 1 0 0 7 0 0 1

Note that these elementary matrices are the results of performing the corresponding EROs
S21 , M3 (7), A31 (−2) on the identity matrix I3 . This is generally true of elementary matrices.

Proposition 39 Elementary matrices are invertible.

Proof. This follows from noting that

(Sij )−1 = Sji = Sij ; (Aij (λ))−1 = Aij (−λ); (Mi (λ))−1 = Mi (λ−1 ),

whether considered as EROs or their corresponding matrices.

Corollary 40 (Invariance of Solution Space under EROs) Let (A|b) be a linear system
of m equations and E an elementary m × m matrix. Then x is a solution of (A|b) if and only
if x is a solution of (EA|Eb).

Proof. The important point here is that E is invertible. So if Ax = b then EAx = Eb follows
by premultiplying by E. But likewise if EAx = Eb is true then it follows that Ax = b by
premultiplying by E −1 .
So applying an ERO, or any succession of EROs, won’t alter the set of solutions of a linear
system. The next key result is that, systematically using EROs, it is possible to reduce any
system (A|b) to reduced row echelon form. Once in this form it is simple to read off the system’s
solutions.

Definition 41 A matrix A is said to be in reduced row echelon form (or simply RRE
form) if
(a) the first non-zero entry of any non-zero row is 1;
(b) in a column that contains such a leading 1, all other entries are zero;
(c) the leading 1 of a non-zero row appears to the right of the leading 1s of the rows above
it;
(d) any zero rows appear below the non-zero rows.

Definition 42 The process of applying EROs to transform a matrix into RRE form is called
row-reduction, or just simply reduction. It is also commonly referred to as Gauss-Jordan
elimination.

REDUCED ROW ECHELON FORM 17

Example 43 Of the following matrices
   √ 
0 1 2 0 −4 1 0 2 0
 0 0 0 1 π ,  0 1 2 0 ,
0 0 0 0 0 0 0 0 1
     √ 
1 0 1 2 1 0 0 3
 0 1 ,  0 1 ,  0 1 0 0 ,
0 0 0 0 0 0 2 1

the first three are in RRE form. The fourth is not as the second column contains a leading
1 but not all other entries of that column are 0. The fifth matrix is not in RRE form as the
leading entry of the third row is not 1.

We have yet to show that any matrix can be uniquely put into RRE form using EROs
(Theorem 122) but – as we have already seen examples covering the range of possibilities – it
seems timely to prove the following result here.

Proposition 44 (Solving Systems in RRE Form) Let (A|b) be a matrix in RRE form
which represents a linear system Ax = b of m equations in n variables. Then
(a) the system has no solutions if and only if the last non-zero row of (A|b) is


0 0 ··· 0 1 .

(b) the system has a unique solution if and only if the non-zero rows of A form the identity
matrix In . In particular, this case is only possible if m ⩾ n.
(c) the system has infinitely many solutions if (A|b) has as many non-zero rows as A,
and not every column of A contains a leading 1. The set of solutions can be described with k
parameters where k is the number of columns not containing a leading 1.


Proof. If (A|b) contains the row 0 0 · · · 0 1 then the system is certainly inconsistent
as no x satisfies the equation
0x1 + 0x2 + · · · + 0xn = 1.
As (A|b) is in RRE form, then this is the only way in which (A|b) can have more non-zero rows
than A. We will show that whenever (A|b) has as many non-zero rows as A then the system
(A|b) is consistent.
Say, then, that both (A|b) and A have r non-zero rows, so there are r leading 1s within
these rows and we have k = n − r columns without leading 1s. By reordering the numbering
of the variables x1 , . . . , xn if necessary, we can assume that the leading 1s appear in the first r
columns. So, ignoring any zero rows, and remembering the system is in RRE form, the system
now reads as the r equations:

x1 + a1(r+1) xr+1 + · · · + a1n xn = b1 ; ··· xr + ar(r+1) xr+1 + · · · + arn xn = br .

We can see that if we assign xr+1 , . . . , xn the k parameters sr+1 , . . . , sn , then we can read off
from the r equations the values for x1 , . . . , xr . So for any values of the parameters we have a
solution x. Conversely though if x = (x1 , . . . , xn ) is a solution, then it appears amongst the

REDUCED ROW ECHELON FORM 18

solutions we’ve just found when we assign values sr+1 = xr+1 , . . . , sn = xn to the parameters.
We see that we have an infinite set of solutions associated with k = n−r independent parameters
when n > r and a unique solution when r = n, in which case the non-zero rows of A are the
matrix In .

Remark 45 Note we showed in this proof that

• a system (A|b) in RRE form is consistent if and only if (A|b) has as many non-zero rows
as A;
• all the solutions of a consistent system can be found by assigning parameters to the vari-
ables corresponding to the columns without leading 1s. ■

Example 46
 
 1 0 0 2 
1 −2 0 2 3
 0 0 1 1 −2 ,  0 1 0 −1 ,
 
 0 0 1 3 
0 0 0 0 1
no solutions
0 0 0 0
unique solution
   
1 2 0 0 3 1 −2 0 2 3
 0 0 1 0 2 ,  0 0 1 1 −2 .
0 0 0 1 1 0 0 0 0 0
one parameter family of solutions two parameter family of solutions
(3−2s,s,2,1) (3+2s−2t,s,−2−t,t)

Theorem 47 (Existence of RRE Form)

Every m × n matrix A can be reduced by EROs to a matrix in RRE form.

Proof. Note that a 1 × n matrix is either zero or can be put into RRE form by dividing by its
leading entry. Suppose, as our inductive hypothesis, that any matrix with fewer than m rows
can be transformed with EROs into RRE form. Let A be an m × n matrix. If A is the zero
matrix, then it is already in RRE form. Otherwise there is a first column cj which contains a
non-zero element α. With an ERO we can swap the row containing α with the first row and
then divide the first row by α ̸= 0 so that the (1, j)th entry now equals 1. Our matrix now
takes the form
0 · · · 0 1 ã1(j+1) . . . ã1n
 
.. .. .. 
 0 · · · 0 ã2j . . . 

 . . . . . .. ,
 .. · · · .. .. .. .. . 
0 · · · 0 ãmj ãm(j+1) · · · ãmn
for some new entries ã1(j+1) , . . . , ãmn . Applying consecutively A12 (−ã2j ), A13 (−ã3j ), . . . , A1m (−ãmj )
leaves column cj = eT1 so that our matrix has become
 
0 · · · 0 1 ã1(j+1) . . . ã1n
 0 ··· 0 0 
.
 
 .. .. ..
 . ··· . . B 
0 ··· 0 0

REDUCED ROW ECHELON FORM 19

By induction, the (m − 1) × (n − j) matrix B can be put into some RRE form by means of
EROs. Applying these same EROs to the bottom m − 1 rows of the above matrix we would
have reduced A to  
0 · · · 0 1 ã1(j+1) . . . ã1n
 0 ··· 0 0 
.
 
 .. .. ..
 . ··· . . RRE(B) 
0 ··· 0 0
To get the above matrix into RRE form we need to make zero any of ã1(j+1) , . . . , ã1n which are
above a leading 1 in RRE(B); if ã1k is the first such entry to lie above a leading 1 in row l then
Al1 (−ã1k ) will make the required edit and in due course we will have transformed A into RRE
form. The result follows by induction.

REDUCED ROW ECHELON FORM 20

2. INVERSES AND TRANSPOSES

Definition 48 A square matrix is a matrix with an equal number of rows and columns. The
diagonal of an n × n matrix A comprises the entries a11 , a22 , . . . , ann – that is, the n entries
running diagonally from the top left to the bottom right. A diagonal matrix is a square matrix
whose non-diagonal entries are all zero. We shall write diag(c1 , c2 , . . . , cn ) for the n×n diagonal
matrix whose (i, i)th entry is ci .

Definition 49 Given an m × n matrix A, then its transpose AT is the n × m matrix such

that the (i, j)th entry of AT is the (j, i)th entry of A.

Proposition 50 (Properties of Transpose)

(a) (Addition and Scalar Multiplication Rules) Let A, B be m × n matrices and λ a real
number. Then
(A + B)T = AT + B T ; (λA)T = λAT .
(b) (Product Rule) Let A be an m×n matrix and B be an n×p matrix. Then (AB)T = B T AT .
(c) (Involution Rule) Let A be an m × n matrix. Then (AT )T = A.
(c) (Inverse Rule) A square matrix A is invertible if and only if AT is invertible. In this
case (AT )−1 = (A−1 )T .

Proof. These are left to Sheet 2, Exercise 3.

Definition 51 A square matrix A = (aij ) is said to be

• symmetric if AT = A.

• skew-symmetric (or antisymmetric) if AT = −A.

• upper triangular if aij = 0 when i > j. Entries below the diagonal are zero.

• strictly upper triangular if aij = 0 when i ⩾ j. Entries on or below the diagonal are
zero.

• lower triangular if aij = 0 when i < j. Entries above the diagonal are zero.

• strictly lower triangular if aij = 0 when i ⩽ j. Entries on or above the diagonal are
zero.

• triangular if it is either upper or lower triangular.

Example 52 Let
   
  1 0   1 0 0
1 2 0 1
A= , B =  2 −1 , C= , D =  0 2 0 .
0 3 −1 0
1 −1 0 0 3

INVERSES AND TRANSPOSES 21

Then
 
      1 0 0
1 0 1 2 1 0 −1
AT = , BT = , CT = , DT =  0 2 0 .
2 3 0 −1 −1 1 0
0 0 3

Note that A is upper triangular and so AT is lower triangular. Also C and C T are skew-
symmetric. And D is diagonal and so also symmetric, upper triangular and lower triangular.

We return now to the issue of determining the invertibility of a square matrix. There is no
neat expression for the inverse of an n × n matrix in general – we have seen that the n = 2 case
is easy enough (Proposition 33) though the n = 3 case is already messy – but the following
method shows how to determine efficiently, using EROs, whether an n × n matrix is invertible
and, in such a case, how to find the inverse.

Algorithm 53 (Determining Invertibility) Let A be an n×n matrix. Place A side-by-side

with In as an augmented n × 2n matrix (A | In ). There are EROs that will reduce A to a matrix
R in RRE form. We will simultaneously apply these EROs to both sides of (A | In ) until we
arrive at (R | P ).

• If R = In then A is invertible and P = A−1 .

• If R ̸= In then A is singular.

Proof. Denote the elementary matrices representing the EROs that reduce A as E1 , E2 , . . . , Ek ,
so that (A | In ) becomes

(Ek Ek−1 · · · E1 A|Ek Ek−1 · · · E1 ) = (R | P ) (2.1)

and we see that R = P A and Ek Ek−1 · · · E1 = P . If R = In then

(Ek Ek−1 · · · E1 )A = In =⇒ A−1 = Ek Ek−1 · · · E1 = P

as elementary matrices are (left and right) invertible. If R ̸= In then, as R is in RRE form
and square, R must have at least one zero row. It follows that (1, 0, . . . , 0)(P A) = 0. As P is
invertible, if A were also invertible, we could postmultiply by A−1 P −1 to conclude (1, 0, . . . , 0) =
0, a contradiction. Hence A is singular; indeed we can see from this proof that as soon as a
zero row appears when reducing A then we know that A is singular.

Example 54 Determine whether the following matrices are invertible, finding any inverses
that exist.  
  1 3 −1 0
1 2 1  0 2 1 1 
A =  2 1 0 , B=  3
.
1 2 1 
1 3 1
0 1 5 3

INVERSES AND TRANSPOSES 22

Solution. Quickly applying a sequence of EROs leads to
A (−2)
 
1 2
12 1 1 0 0
A13 (−1)
(A|I3 ) −→  0 −3 −2 −2 1 0 
0 1 0 −1 0 1
A31 (−2)
   
1 0 1 3 0 −2 1/2 1/2 −1/2
A32 (3) M3 (−1/2)
−→  0 1 0 −1 0 1  −→  I3 −1 0 1 .
S23 A31 (−1)
0 0 −2 −5 1 3 5/2 −1/2 −3/2

Hence  −1  
1 2 1 1/2 1/2 −1/2
 2 1 0  =  −1 0 1 .
1 3 1 5/2 −1/2 −3/2
For B we note
 
A13 (−3)
1 3 −1 0 1 0 0 0
S24  0 1 5 3 0 0 0 1 

(B|I4 ) −→  0 −8

5 1 −3 0 1 0 
0 2 1 1 0 1 0 0
 
A23 (8) 1 3 −1 0 1 0 0 0
A24 (−2)  0 1 5 3 0 0 0 1 
−→  .
A34 (1/5)  0 0 45 25 −3 0 1 8 
0 0 0 0 −3/5 1 1/5 −2/5

The left matrix is not yet in RRE form, but the presence of a zero row is sufficient to show
that B is singular.

Remark 55 We have defined matrix multiplication in such a way that we can see how to
implement it on a computer. But how long will it take for a computer to run such a calculation?
To multiply two n × n matrices in this way, for each of the n2 entries we must multiply n
pairs and carry out n−1 additions. So the process takes around n3 multiplications and n2 (n−1)
additions. When n is large, these are very large numbers!
In 1969, Strassen gave a faster algorithm, which has since been improved on. It is not known
whether these algorithms give the fastest possible calculations. Such research falls into the field of
computational complexity, drawing on ideas from both mathematics and computer science.

Finally we define the orthogonal matrices, matrices important to the geometry of Rn .

Definition 56 An n × n matrix is orthogonal if AT = A−1 .

Proposition 57 Let A and B be orthogonal n × n matrices. Then:

(a) AB and A−1 are othogonal. Consequently the n × n orthogonal matrices form a group
O(n).
(b) A is orthogonal if and only if its columns (or rows) are n unit length, mutually perpen-
dicular vectors.
(c) A preserves the dot product. If x,y ∈ Rncol then Ax · Ay = x · y.

INVERSES AND TRANSPOSES 23

Proof. (a) We have that

(AB)T = B T AT = B −1 A−1 = (AB)−1 ,

by the product rules for transposes and inverses, showing that AB is orthogonal. Similarly

T
−1
−1
A−1 = AT = A−1 ,

showing that A−1 is orthogonal.

The reason that orthogonal matrices are important in geometry is that the orthogonal
matrices are precisely those matrices that preserve the dot product.

Proposition 58 Let A be an n × n matrix. Then A is orthogonal if and only if Ax · Ay = x · y

for all x, y ∈ Rncol .

Proof. Note that xT y = x · y. So

Ax · Ay = x · y
⇐⇒ (Ax)T Ay = xT y
⇐⇒ xT AT Ay = xT y.

If AT A = In then the above is clearly true. Conversely, by choosing x = ei and y = ej , from

the standard basis of Rncol the above implies

the (i, j) th entry of AT A = δij = the (i, j) th entry of In .

As this is true for all i, j then this implies AT A = In .

INVERSES AND TRANSPOSES 24

3. VECTOR SPACES

Currently when you speak of vectors, you usually mean coordinate vectors represented either as
a row vector in some Rn or as a column vector in some Rncol . But vectors exist without reference
to coordinate systems. Wherever you are at the moment, look around you and choose some
−→
point near you and label it P, then pick a second point and label it Q. Then P Q is a vector. If
−→
you want to treat P as the origin then P Q is the position vector of Q. Or you might think of
−→
P Q as a movement and any parallel movement, with the same length and direction, equals the
−→ −→
vector P Q. Importantly though P Q has no coordinates, or at least doesn’t until you make a
choice of origin and axes. This is going to be an important aspect of the Linear Algebra I and
II courses, namely choosing coordinates sensibly. This will also be an important aspect of the
Geometry and Dynamics courses – in Geometry the change between two coordinate systems
will need to be an isometry so that the lengths, areas, angles are measured to be the same;
in Dynamics an inertial frame would be necessary for Newton’s laws to hold and otherwise
so-called ‘fictitious forces’ will arise.
But the vector spaces we will introduce are not just geometrical vectors like these coordinate
or coordinateless vectors. A vector space’s elements might contain functions, sequences, ma-
trices, equations or, of course, vectors. Importantly, these more abstract vector spaces do have
the same algebraic operations in common with the vectors familiar to you: namely, addition
and scalar multiplication.

3.1 What is a vector space?

A real vector space is a non-empty set with operations of addition and scalar multiplication.
Formally this means:

Definition 59 A real vector space is a non-empty set V together with a binary operation
V × V → V given by (u, v) 7→ u + v (called addition) and a map R × V → V given by
(λ, v) 7→ λv (called scalar multiplication) that satisfy the vector space axioms

• u + v = v + u for all u, v ∈ V (addition is commutative);

• u + (v + w) = (u + v) + w for all u, v, w ∈ V (addition is associative);

• there is 0V ∈ V such that v + 0V = v = 0V + v for all v ∈ V (existence of additive

identity);

• for all v ∈ V there exists w ∈ V such that v + w = 0V = w + v (existence of additive

inverses);

VECTOR SPACES 25
 • λ(u + v) = λu + λv for all u, v ∈ V , λ ∈ R (distributivity of scalar multiplication over
vector addition);

• (λ + µ)v = λv + µv for all v ∈ V , λ, µ ∈ R (distributivity of scalar multiplication over

field addition);

• (λµ)v = λ(µv) for all v ∈ V , λ, µ ∈ R (scalar multiplication interacts well with field
multiplication);

• 1v = v for all v ∈ V (identity for scalar multiplication).

R is referred to as the field of scalars or base field. Elements of V are called vectors
and elements of R are called scalars.

Remark 60 There are a lot of axioms on the above list, but the most important in practice are
those requiring:

• V has a zero vector 0V .

• V is closed under addition.

• V is closed under scalar multiplication.

If these three axioms hold, and addition and scalar multiplication are defined naturally, then
usually the remaining axioms will follow as a matter of routine checks.
The subsets of R3 that are real vector spaces are the origin, lines through the origin, planes
through the origin and all of R3 . It’s perhaps not surprising then that another term for a vector
space is a ‘linear space’.

Example 61 We write Rn for the set of n-tuples (v1 , . . . , vn ) with v1 , . . . , vn ∈ R. Then Rn

is a real vector space under componentwise addition and scalar multiplication:

(x1 , . . . , xn ) + (y1 , . . . , yn ) = (x1 + y1 , . . . , xn + yy ) (3.1)

and λ(x1 , . . . , xn ) = (λx1 , . . . , λxn ). (3.2)

These satisfy the vector space axioms. The zero vector is (0, 0, . . . , 0) and the additive inverse
of (v1 , . . . , vn ) is (−v1 , . . . , −vn ).
We think of R2 as the Cartesian plane, and R3 as three-dimensional space. We can also
consider n = 1: R1 is a real vector space, which we think of as the real line. We tend to write
it simply as R.

Notation 62 I will often denote a single coordinate vector (v1 , v2 , . . . , vn ) as v. I will use
this bold notation for coordinate vectors, but vectors, as elements of a vector space, will not be
written in bold.

Example 63 The field C is a real vector space, it is essentially the same as R2 as a vector
space. (The technical term for ‘essentially the same’ is ‘isomorphic’. More on this later.)

WHAT IS A VECTOR SPACE? 26

Example 64 For m, n ⩾ 1, the set Mm×n (R) is a real vector space as previously stated in
Remark 20.

Example 65 Let V = RR = {f : R → R}, the space of all real-valued functions on R, with

addition and scalar multiplication defined pointwise. That is,

(f + g)(r) := f (r) + g(r), (αf )(r) := αf (r),

for f, g ∈ V and α, r ∈ R.

Example 66 Let V = {f : R → R, f is differentiable}, the space of all differentiable real-valued

functions on R, with addition and scalar multiplication defined pointwise. This is a vector space
as, in particular,
(f + g)′ = f ′ + g ′ , (αf )′ = αf ′ ,
for f, g ∈ V and α ∈ R, so that the sum of two differentiable functions and the scalar multiple
of a differentiable function is differentiable. It’s these facts that mean V is closed under addition
and scalar multiplication.

Example 67 Let V = {f : R → R, f ′′ = f }. This is a vector space mainly as f ′′ = f is a

linear differential equation. That is, if f and g are solutions then so is αf + βg where α, β
are real scalars. The general solution can be written as

f (x) = A cosh x + B sinh x

or as
f (x) = Aex + Be−x .
In expressing the general vector in this way note that in each case we are ‘coordinatizing the
space’ and identifying V with R2 . But note that the coordinate vector (1, 0) corresponds to
different vectors as we are using different choices of coordinates; it corresponds to the vector
cosh x in the first case and to ex in the second.

Example 68 Let V = RN = {(x0 , x1 , x2 , ...) : xi ∈ R} . This is the space of all real sequences
with addition and scalar multiplication defined componentwise.
Other important sequence spaces are

l∞ = {(xn )∞n=0 : (xn ) is bounded} .

c = {(xn )∞
n=0 : (xn ) converges} .
c0 = {(xn )∞
n=0 : (xn ) converges to 0} .

As an exercise, what theorems of analysis (concerning convergence) need to hold for these last
three examples all to be vector spaces?

Our main focus in this course will be real vector spaces. However, vector spaces can be
defined over any field, as can simultaneous equations be considered over any field. Formally
a vector space V over a field F is a non-zero set V with addition V × V → V and scalar
multiplication F × V → V satisfying the vector space axioms in Definition 59. Common
examples of other fields that we will encounter are:

WHAT IS A VECTOR SPACE? 27

 • C – the field of complex numbers.

• Q – the field of rational numbers.

• Zp – the field of integers modulo a prime number p.

The theory of vector spaces applies equally well for all fields. There can be some differences
worth noting though depending on the choice of field.

• A non-zero real vector space is an infinite set. This need not be the case over a finite field
like Zp .

• When we consider C as a vector space over R, then every z can be uniquely written as
x1 + yi for two real scalars x and y. But when C is considered as a vector space over C,
then every z can be uniquely written as z1 for a single complex scalar z. (In due course we
will appreciate that C is a 2-dimensional real vector space and a 1-dimensional complex
vector space.

Lemma 69 Let V be a vector space over F. Then there is a unique additive identity element
0V .

Proof. Suppose that 0 and 0′ are two elements that have the properties of 0V . Then

0 = 0 + 0′ [as 0′ is a zero vector]

= 0′ [as 0 is a zero vector]

and so 0 = 0′ , thus showing 0V to be unique.

Remark 70 Where it will not be ambiguous, we often write 0 for 0V .

Lemma 71 Let V be a vector space over F. Take v ∈ V . Then there is a unique additive
inverse for v. That is, if there are w1 , w2 ∈ V with v + w1 = 0V = w1 + v and v + w2 = 0V =
w2 + v, then w1 = w2 .

Proof. With the notation introduced above we have

w2 = 0 + w2 [as 0 is a zero vector]

= (w1 + v) + w2 [by a hypothesis]
= w1 + (v + w2 ) [by associativity]
= w1 + 0 [by a hypothesis]
= w1 . [as 0 is a zero vector]

Remark 72 Using the notation of Lemma 71, we write −v for the unique additive inverse of
v.

WHAT IS A VECTOR SPACE? 28

Proposition 73 Let V be a vector space over a field F. Take v ∈ V , λ ∈ F. Then
(a) λ0V = 0V ;
(b) 0v = 0V ;
(c) (−λ)v = −(λv) = λ(−v);
(d) if λv = 0V then λ = 0 or v = 0V .
(e) −v = (−1)v.

Proof. (a) We have

λ0V = λ(0V + 0V ) [definition of additive identity]

= λ0V + λ0V [distributivity of scalar · over vector +].

Adding −(λ0V ) to both sides, we have

0V = λ0V .

(b) Exercise (hint: in F we have 0 + 0 = 0).

(c) We have

λv + λ(−v) = λ(v + (−v)) [distributivity of scalar · over vector +]

= λ0V [definition of additive inverse]
= 0V [by (b)].

So λ(−v) is the additive inverse of λv (by uniqueness), so λ(−v) = −(λv).

Similarly, we see that λv + (−λ)v = 0V and so (−λ)v = −(λv).
(d) Suppose that λv = 0V , and that λ ̸= 0. Then λ−1 exists in F, and

λ−1 (λv) = λ−1 0V = 0V [by (a)].

So
(λ−1 λ)v = 0V [scalar · interacts well with field ·],
showing
v = 1v = 0V [identity for scalar multiplication].
(e) Note that

v + (−1)v = 1v + (−1)v [by a vector space axiom]

= (1 + (−1)) v [distributivity]
= 0v [definition of additive inverse in field]
= 0V . [by (a)]

Hence by the uniqueness of the additive inverse (−1)v = −v.

WHAT IS A VECTOR SPACE? 29

3.2 Subspaces

Whenever we have a mathematical object with some structure, we want to consider subsets
that also have that same structure.

Definition 74 Let V be a vector space over F. A subspace of V is a non-empty subset of V

that is closed under addition and scalar multiplication, that is, a subset U ⊆ V such that
(i) U ̸= ∅ (U is non-empty); (this usually involves showing 0V ∈ U ).
(ii) u1 + u2 ∈ U for all u1 , u2 ∈ U (U is closed under addition);
(iii) λu ∈ U for all u ∈ U , λ ∈ F (U is closed under scalar multiplication).
Note that the operations of addition and scalar multiplication referred to are those of V , not
some separate, different operations of U.

Definition 75 The sets {0V } and V are always subspaces of V . The subspace {0V } is some-
times called the zero subspace or the trivial subspace. Subspaces other than V are called
proper subspaces.

Proposition 76 (Subspace test) Let V be a vector space over F, let U be a subset of V .

Then U is a subspace if and only if

(i) 0V ∈ U ; and
(ii) λu1 + u2 ∈ U for all u1 , u2 ∈ U and λ ∈ F.

Proof. (⇒) Assume that U is a subspace of V .

0V ∈ U : Since U is a subspace, it is non-empty, so there exists u ∈ U . Since U is closed
under scalar multiplication, 0u = 0V ∈ U .
λu1 + u2 ∈ U for all u1 , u2 ∈ U and all λ ∈ F: Take u1 , u2 ∈ U , and λ ∈ F. Then λu1 ∈ U
because U is closed under scalar multiplication, so λu1 + u2 ∈ U because U is also closed under
addition.
(⇐) Assume that 0V ∈ U and that λu1 + u2 ∈ U for all u1 , u2 ∈ U and λ ∈ F.
U is non-empty: we note 0V ∈ U .
U is closed under addition: for u1 , u2 ∈ U have u1 + u2 = 1u1 + u2 ∈ U .
U is closed under scalar multiplication: for u ∈ U and λ ∈ F, have λu = λu + 0V ∈ U .
So U is a subspace of V .

Notation 77 If U is a subspace of the vector space V , then we write U ⩽ V .

Proposition 78 Let V be a vector space over F, and let U ⩽ V . Then

(a) U is a vector space over F. In fact, the only subsets of V that are vector spaces over F
are the subspaces;
(b) if W ⩽ U then W ⩽ V (“a subspace of a subspace is a subspace”).

Proof. (a) We need to check the vector space axioms, but first we need to check that we have
legitimate operations. Since U is closed under addition, the operation + restricted to U gives

SUBSPACES 30
a map U × U → U . Likewise since U is closed under scalar multiplication, that operation
restricted to U gives a map F × U → U .
Now for the axioms.
Commutativity and associativity of addition are inherited from V .
There is an additive identity (by the subspace test).
There are additive inverses: if u ∈ U then multiplying by −1 ∈ F and shows that −u =
(−1)u ∈ U .
The remaining four properties are all inherited from V . That is, they apply to general
vectors of V and vectors in U are vectors in V.
(b) This is immediate from the definition of a subspace.

Proposition 79 Let V be a vector space. Take U , W ⩽ V . Then U +W ⩽ V and U ∩W ⩽ V ,

where
U + W = {u + w | u ∈ U, w ∈ W } .
Indeed U + W is the smallest subspace of V which contains U and W and U ∩ W is the largest
subspace of V which is contained in both U and W.

Proof. (a) As U ⩽ V and W ⩽ V then 0V ∈ U and 0V ∈ W so that 0V = 0V + 0V ∈ U + W.

Say that v1 , v2 ∈ U + W and λ ∈ F. By definition there exist u1 , u2 ∈ U and w1 , w2 ∈ W
such that
v1 = u1 + w1 , v2 = u2 + w2 ,
and then

λv1 + v2 = λ(u1 + w1 ) + u2 + w2 = (λu1 + u2 ) + (λw1 + w2 ) ∈ U + W

as λu1 + u2 ∈ U and λw1 + w2 ∈ W because U ⩽ V and W ⩽ V.

(b) The statements concerning the intersection are left as exercises.

3.3 Further examples

Example 80 Consider a system of homogeneous linear equations with real coefficients aij :

a11 x1 + a12 x2 + · · · + a1n xn = 0

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

(We say this is homogeneous because all the real numbers on the right are 0.)
Let V be the set of real solutions of the this linear system. Then V is a real vector space.
This becomes more apparent if we write the equations in matrix form. We see the system
corresponds to Ax = 0, where A = (aij ) ∈ Mm×n (R), x = (x1 , x2 , . . . , xn )T is an n × 1 column

FURTHER EXAMPLES 31
vector of variables, and 0 is shorthand for 0n×1 . Each element of V can be thought of as an
n × 1 column vector of real numbers.
To show that V is a vector space, we show that it is a subspace of Rncol .
Clearly V is non-empty, because 0 ∈ V .
For v1 , v2 ∈ V , we have Av1 = 0 and Av2 = 0, so A(v1 + v2 ) = Av1 + Av2 = 0 + 0 = 0,
so v1 + v2 ∈ V . So V is closed under addition.
For v ∈ V and λ ∈ F, we have A(λv) = λ(Av) = λ0 = 0, so λv ∈ V . So V is closed under
scalar multiplication.
So V ⩽ Rncol , and so V is a vector space.

Example 81 The set R[x] of all real polynomials in a variable x is a real vector space. We
will show that it is a subspace of RR . Addition and scalar multiplication are defined by
X  X  X X  X
an x n + bn x n = (an + bn ) xn , λ an x n = (λan )xn .

As the sums are finite, then the addition and scalar multiple are also finite and hence polyno-
mials. Finally teh zero function is a polynomial.

Example 82 Let n be a non-negative integer. The set of polynomials cn xn + · · · + c1 x + c0

with c0 , c1 , . . . , cn ∈ R (that is, real polynomials with degree ⩽ n) is a real vector space, and a
subspace of R[x].

Example 83 Let X be a set. Define RX := {functions f with f : X → R}, the set of real-
valued functions on X. This is a real vector space with operations of pointwise addition and
pointwise multiplication by a real number: for x ∈ X, we define

(f + g)(x) = f (x) + g(x) and (λf )(x) = λf (x).

Example 84 We can study the solutions of a homogeneous linear second-order differential

equation. These are twice-differentiable real functions y that satisfy an equation of the form

y ′′ + a(x)y ′ + b(x)y = 0.

This equation is linear because y and its derivatives occur only to the first power and are not
multiplied together. And it is homogeneous because of the 0 on the right-hand side. Such
equations are important in many applications of mathematics.
The set S of solutions of this homogeneous linear second-order differential equation is a vec-
tor space, a subspace of RR . Note S is clearly non-empty (the 0 function satisfies the differential
equation), and if w = u + λv where u, v ∈ S and λ ∈ R, then

w′′ + a(x)w′ + b(x)w = (u′′ + λv ′′ ) + a(x)(u′ + λv ′ ) + b(x)(u + λv)

= (u′′ + a(x)u′ + b(x)u) + λ(v ′′ + a(x)v ′ + b(x)v)
= 0,

so w ∈ S. So, by the Subspace Test, S ⩽ RR .

This generalises to homogeneous linear differential equations of any order.

FURTHER EXAMPLES 32
Example 85 What are the subspaces of R?
Let V = R, let U be a non-trivial subspace of V . Then there exists u ∈ U with u ̸= 0. Take
x ∈ R. Let λ = ux . Then x = λu ∈ U , because U is closed under scalar multiplication. So
U =V.
So the only subspaces of R are {0} and R.

Example 86 What are the subspaces of R2 ?

Let V = R2 , let U be a non-trivial subspace of V . Then there exists u ∈ U with u ̸= (0, 0),
say u = (a, b). We have ⟨u⟩ = {λu : λ ∈ R} ⊆ U . (Such ‘spans’ will more generally be defined
in the next chapter.)
Case 1: ⟨u⟩ = U .
If a ̸= 0, then let m = ab . Then ⟨u⟩ = {(x, y) ∈ R2 | y = mx}.
If a = 0, then ⟨u⟩ = {(0, y) | y ∈ R}.
So if U is, geometrically, a line in R2 through the origin, and every such line in R2 through
the origin corresponds to a subspace.
Case 2: ⟨u⟩ = ̸ U.
Then there is some v = (c, d) ∈ U \⟨u⟩.
 
a b
Consider the matrix . Applying any sequence of EROs to this matrix gives a matrix
c d  
1 0
whose rows are in U . The matrix must have RRE form . [Why?] So U contains the
0 1
vectors (1, 0) and (0, 1), and hence U = R2 . [Why?]
So the subspaces of R2 are {0}, lines in R2 through the origin and R2 .

FURTHER EXAMPLES 33
4. BASES

One key goal of this section is to develop a sensible notion of the ‘dimension’ of a vector
space. In order to do this, we need to develop some theory that is in itself both important and
interesting.

4.1 Spans and Spanning sets

Lemma 87 Let V be a vector space over a field F, and take S = {u1 , u2 , . . . , um } ⊆ V .

Define
U := {α1 u1 + · · · + αm um : α1 , . . . , αm ∈ F}.
Then U ⩽ V .

Proof. Applying the subspace test, we note.

• 0V ∈ U : have 0V = 0u1 + · · · + 0um ∈ U .

• λv1 + v2 ∈ U : take v1 , v2 ∈ U , say v1 = α1 u1 + · · · + αm um and v2 = β1 u1 + · · · + βm um ,

where α1 , . . . , αm , β1 , . . . , βm ∈ F. Take λ ∈ F. Then

λv1 + v2 = (λα1 + β1 )u1 + · · · + (λαm + βm )um ∈ U.

So, by the subspace test, U ⩽ V .

Definition 88 Let V be a vector space over F, take u1 , u2 , . . . , um ∈ V . A linear combi-

nation of u1 , . . . , um is a vector α1 u1 + · · · + αm um for some α1 , . . . , αm ∈ F. We define the
span of u1 , . . . , um to be

⟨u1 , . . . , um ⟩ := {α1 u1 + · · · + αm um : α1 , . . . , αm ∈ F}.

This is the smallest subspace of V that contains u1 , . . . , um .

More generally, we can define the span of any set S ⊆ V (even a potentially infinite set S)
as
⟨S⟩ := {α1 s1 + · · · + αm sm : m ⩾ 0, s1 , . . . , sm ∈ S, α1 , . . . , αm ∈ F}.
Note that a linear combination only ever involves finitely many elements of S, even if S is
infinite. There isn’t enough structure in a vector space to be able to define infinite sums. By
convention the span of the empty set in {0V }.

Definition 89 Let V be a vector space over F. If S ⊆ V is such that V = ⟨S⟩, then we say
that S spans V , and that S is a spanning set for V .

BASES 34
Example 90 {(1, 1), (2, −1)} spans R2 as every (x, y) can be written
   
x + 2y x−y
(x, y) = (1, 1) + (2, −1) .
3 3

Whilst the span of {(2, 2), (−1, −1)} is the line y = x in R2 .

Example 91 {(1, 1, 2), (2, −1, 3)} spans the plane given parametrically as

r = α(1, 1, 2) + β(2, −1, 3) α, β ∈ R.

By eliminating α, β from the expressions

x = α + 2β, y = α − β, z = 2α + 3β,

then we can see this is the plane with equation

5x + y − 3z = 0.

Example 92 The three vectors {(1, 1, 2), (2, −1, 3), (3, 0, 5)} span the same plane 5x+y −3z =
0. This is because
(3, 0, 5) = (1, 1, 2) + (2, −1, 3)
and so the third vector is itself a linear combination of the first two. Note that any point in the
plane can be written in many different ways as a linear combination of the three vectors. For
example

(0, 3, 1) = 2(1, 1, 2) − 1(2, −1, 3) + 0(3, 0, 5)

= 1(1, 1, 2) − 2(2, −1, 3) + 1(3, 0, 5)
= 3(1, 1, 2) + 0(2, −1, 3) − 1(3, 0, 5).

This third vector means there is redundancy in the set. Any two of the three vectors are sufficient
to span the plan. The issue here is that the three vectors are not linearly independent.

Definition 93 Given a matrix, its row space is the span of its rows and its column space
is the span of its column. For an m × n matrix A, we write Row(A) ⩽ Rn for its row space
and Col(A) ⩽ Rm col for its column space.

Example 94 In Example 7 we met the matrix on the left below, and the matrix on the right
is its RRE form.
   
1 −1 1 3 2 1 0 0 −1 2
 2 −1 1 2 4 ,  0 1 −1 −4 0  .
4 −3 3 8 8 0 0 0 0 0

A check will show that these two matrices have the same row space – we will see in Proposition
117 that EROs don’t change row space. However it is clear that (1, 2, 4)T is in the column space
of the first matrix and not of the second – so EROs do change column space.

SPANS AND SPANNING SETS 35

4.2 Linear independence

Definition 95 Let V be a vector space over F. We say that v1 , . . . , vm ∈ V are linearly

independent if the only solution to the equation

α1 v1 + · · · + αm vm = 0V where α1 , . . . , αm ∈ F

is
α1 = α2 = · · · = αm = 0.
Otherwise v1 , . . . , vm are said to be linearly dependent, which means there is a non-trivial linear
combination of v1 , . . . , vm which adds to 0V .
We say that S ⊆ V is linearly independent if every finite subset of S is linearly independent.

Example 96 {(1, 1, 2), (2, −1, 3)} ⊆ R2 is linearly independent. To check this, we see that
comparing the x- and y-coordinates in

α(1, 1, 2) + β(2, −1, 3) = (0, 0, 0),

implies
α + β = 0, 2α − β = 0.
These equations alone are enough to show α = β = 0. Note though that these two vectors do
not span R3 .

Example 97 {(1, 1, 2), (2, −1, 3), (3, 0, 5)} is linearly dependent. We previously noted that

(1, 1, 2) + (2, −1, 3) = (3, 0, 5)

so that
1(1, 1, 2) + 1(2, −1, 3) + (−1)(3, 0, 5) = (0, 0, 0).
This is a non-trivial linear combination which adds up to 0.

Example 98 Let V denote the vector space of differentiable functions f : R → R. Then the set

S = {sin x, cos x, sin 2x}

is linearly independent. Say that

α sin x + β cos x + γ sin 2x = 0V ,

noting 0V denotes the zero function, so that the above is an identity of functions. If we set
x = 0 then this gives β = 0. If we set x = π/2 then α = 0. Hence γ = 0 also.

Proposition 99 Let S = {v1 , . . . , vm } be a linearly independent subset of a vector space V.

Then
α1 v1 + · · · + αm vm = β1 v1 + · · · + βm vm
if and only if αi = βi for all 1 ⩽ i ⩽ m. Hence we may ‘compare coefficients’.

LINEAR INDEPENDENCE 36
Proof. If αi = βi for all 1 ⩽ i ⩽ m then the result clearly follows. Conversely, we can rearrange
the above equation as
(α1 − β1 )v1 + · · · + (αm − βm )vm = 0V .
As S is linearly independent then αi − βi = 0 for all i as required.

Example 100 Let V = C, considered as a real vector space. Then {1, i} is linearly independent
for if
x + yi = 0C
then x = Re 0C = 0 and y = Im 0C = 0. Hence by the previous proposition ‘comparing real and
imaginary parts’ is valid.

Example 101 Let V = R[x], the vector space of polynomials with real coefficients. Then the
set S = {1, x, x2 , . . .} is linearly independent. Recall that an infinite set is linearly independent
if every finite subset is linearly independent. So say that

a0 1 + a1 x + a2 x2 + · · · + an xn = 0R[x]

for some coefficients a0 , a1 , a2 , . . . , an . Recall that the above is an identity of functions. We can
see that a0 = 0 by setting x = 0. We can then see that a1 = 0 by differentiating and setting
x = 0. In a similar fashion we can see that all the coefficients are zero and that S is linearly
independent.

Lemma 102 Let v1 , . . . , vm be linearly independent elements of a vector space V . Let vm+1 ∈
V . Then v1 , v2 , . . . , vm , vm+1 are linearly independent if and only if

vm+1 ̸∈ ⟨v1 , . . . , vm ⟩.

Proof. ( ⇐= ) Suppose that vm+1 ̸∈ ⟨v1 , . . . , vm ⟩. Take α1 , . . . , αm+1 ∈ F such that

α1 v1 + · · · + αm+1 vm+1 = 0V .

We aim to show that the αi are all 0. If αm+1 ̸= 0, then we have

1
vm+1 = − (α1 v1 + · · · + αm vm ) ∈ ⟨v1 , . . . , vm ⟩,
αm+1
which is a contradiction. So αm+1 = 0 and hence α1 v1 + · · · + αm vm = 0V . But v1 , . . . , vm are
linearly independent, so this means that α1 = · · · = αm = 0 as well.
( ⇐= ) Conversely say that v1 , v2 , . . . , vm , vm+1 are linearly independent. If vm+1 ∈
⟨v1 , . . . , vm ⟩ then there exist α1 , . . . , αm ∈ F such that

vm+1 = α1 v1 + · · · + αm vm

so that α1 v1 + · · · + αm vm − vm+1 = 0V which contradicts the linear independence of v1 , v2 , . . . ,

vm , vm+1 . Hence vm+1 ̸∈ ⟨v1 , . . . , vm ⟩.

LINEAR INDEPENDENCE 37
4.3 Bases

Definition 103 Let V be a vector space. A basis of V is a linearly independent, spanning set.
(The plural is ‘bases’, pronounced ‘bay-seas’.)
If V has a finite basis, then we say that V is finite-dimensional.
Remark 104 It is important to note the language here. We can talk about ‘a’ basis of a vector
space. Typically, vector spaces have many bases so we should not talk about ‘the’ basis. Some
vector spaces have a ‘standard’ or ‘canonical’ basis though.
Remark 105 Not every vector space is finite-dimensional. For example, the space of real
polynomials or the space of real sequences do not have finite bases. But in this course we’ll
generally study finite-dimensional vector spaces. The courses on Functional Analysis in Parts
B and C (third and fourth year) explore the theory of infinite-dimensional vector spaces which
have further analytical structure. Note in a vector space that only finite sums are well-defined.
To meaningfully form an infinite sum, a notion of convergence is needed which is why further
structure is needed.
Where possible, we will work with general vector spaces, but sometimes we’ll need to spe-
cialise to the finite-dimensional case.
Example 106 In Rn , for 1 ⩽ i ⩽ n, let ei be the row vector with coordinate 1 in the ith entry
and 0 elsewhere. Then e1 , . . . , en are linearly independent: if
α1 e1 + · · · + αn en = 0
then by looking at the ith entry we see that αi = 0 for all i. Also, e1 , . . . , en span Rn , because
(a1 , . . . , an ) = a1 e1 + · · · + an en .
So e1 , . . . , en is a basis of Rn . We call it the standard basis or canonical basis of Rn .
Example 107 Let V = Mm×n (F) denote the vector space of m × n matrices over a field F.
Then the standard basis of V is the set
{Eij | 1 ⩽ i ⩽ m, 1 ⩽ j ⩽ n}
which has entry of 1 for the (i, j)th entry, and all other entries are zero. Note that a matrix
A = (aij ) can be written
Xm X n
A= aij Eij
i=1 j=1

and this is the unique expression of A as a linear combination of the standard basis.
Example 108 Let V = {(x.y.z) ∈ R3 | x + 2y + z = 0} ⩽ R3 . Then a basis for V is
{(1, 0, −1), (0, 2, −1)} .
To see this note that x and y can be used to parameterize V and a general vector can be written
uniquely as
(x, y, −x − 2y) = x(1, 0, −1) + y(0, 2, −1).

BASES 38
Example 109 Let V ⩽ R5 be the space of vectors (x1 , x2 , x3 , x4 , x5 ) satisfying the three equa-
tions

x1 + x2 − x3 + x5 = 0;
x1 + 2x2 + x4 + 3x5 = 0;
x2 + x3 + x4 + 2x5 = 0.

We can represent these equations as

   
1 1 −1 0 1 1 0 −2 −1 −1
RRE
 1 2 0 1 3  →  0 1 1 1 2 .
0 1 1 1 2 0 0 0 0 0

If we assign parameters to the last three columns (as there are no leading 1s in these columns)
by setting x3 = α, x4 = β, x5 = γ then

x1 = 2α + β + γ, x2 = −α − β − 2γ

and hence

(x1 , x2 , x3 , x4 , x5 ) = (2α + β + γ, −α − β − 2γ, α, β, γ)

= α(2, −1, 1, 0, 0) + β(1, −1, 0, 1, 0) + γ(1, −2, 0, 0, 1).

So a basis for V is
{(2, −1, 1, 0, 0), (1, −1, 0, 1, 0), (1, −2, 0, 0, 1)} .

Example 110 The space F[x] of polynomials over a field F (that is, with coefficients from the
field F) has standard basis
1, x, x2 , x3 , . . . .


Every polynomial can be uniquely written as a finite linear combination of this basis.

Proposition 111 Let V be a vector space over F, let S = {v1 , . . . , vn } ⊆ V . Then S is a

basis of V if and only if every vector in V has a unique expression as a linear combination of
elements of S.

Proof. (⇒) Let S be a basis of V . Take v ∈ V . Since S spans V , there exist α1 , . . . , αn ∈ F

such that v = α1 v1 + · · · + αn vn . Further as S is linearly independent, then by Proposition 99
these scalars α1 , . . . , αn are unique.

(⇐) Conversely, suppose that every vector in V has a unique expression as a linear combination
of elements of S.

• S spanning set: for any v ∈ V we can write v as a linear combination of elements of S.

So span(S) = V .

• S linearly independent: for α1 , . . . , αn ∈ F, if α1 v1 + · · · + αn vn = 0 = 0v1 + · · · + 0vn ,

then by uniqueness we have αi = 0 for all i.

BASES 39
 So S is a basis for V .

Propostion 111 allows us to define:

Definition 112 Given a basis {v1 , . . . , vn } of V then every v ∈ V can be uniquely written

v = α1 v1 + · · · + αn vn

and the scalars α1 , . . . , αn are known as the coordinates of v with respect to the basis {v1 , . . . , vn }.

Remark 113 Thus choosing a basis {v1 , . . . , vn } for a finite-dimensional vector space V iden-
tifies V with Rn . To a vector v can be associated a coordinate vector v = (α1 , . . . , αn ) .
A vector space has an origin, but no axes. Choosing a basis of V introduces αi -axes into V
and identifies a vector v with a coordinate vector v. I will denote coordinate vectors in bold,
or underline them when writing by hand. It is important to note that a coordinate vector is
meaningless without the context of a basis as we can see in the following example.

Example 114 Let V = {f : R → R, f ′′ (x) = 4f (x)} . Then the general solution of the differ-
ential equation can be written uniquely as

f (x) = Ae2x + Be−2x

or as
f (x) = C sinh 2x + D cosh 2x.
So {e2x , e−2x } is a basis of V as is {sinh 2x, cosh 2x} . Note that the same vector e2x has coor-
dinates (A, B) = (1, 0) using the first basis and has coordinates (C, D) = (1, 1) with respect to
the second basis as
e2x = sinh 2x + cosh 2x,
Similarly the same coordinate vector (1, 0) represents the vector e2x with respect to the first
basis, but a different vector sinh 2x with respect to the second basis.

Remark 115 The above, of course, raises the question of whether there is a best way to coor-
dinatize a vector space – or equivalently a best way to choose a basis.

Remark 116 The question of whether all vector spaces have a basis is an important founda-
tional one. Every vector space does have a basis provided we assume the so-called ‘axiom of
choice’, which is not a standard axiom of set theory. However, it can be shown that a basis of a
space like l∞ , the space of bounded real sequences, is necessarily uncountable. So the structure of
vector spaces, solely, is not well suited to working with some infinite-dimensional vector spaces
which explains why the topic of infinite-dimensional space is more one of ‘functional analysis’
where infinite linear combinations can be well-defined.

We now turn to the question of how we determine whether a set of vectors is linearly
independent or spanning. Recall that we write Row(M ) for the row space of a matrix M, that
is the span of the rows of M.

BASES 40
Proposition 117 Let A = (aij ) be an m × n matrix and let B = (bij ) be a k × m matrix. Let
R = (rij ) be a matrix in RRE form which can be obtained by EROs from A.
(a) The non-zero rows of R are independent.
(b) The rows of R are linear combinations of the rows of A.
(c) Row(BA) is contained in Row(A).
(d) If k = m and B is invertible then Row(BA) = Row(A).
(e) Row(R) = Row(A).

Proof. (a) Denote the non-zero rows of R as r1 , . . . , rr and suppose that c1 r1 + · · · + cr rr = 0.

Say the leading 1 of r1 appears in the jth column. Then

c1 + c2 r2j + c3 r3j + · · · + cr rrj = 0.

But as R is in RRE form each of r2j , r3j , . . . , rrj is zero, being entries under a leading 1. It
follows that c1 = 0. By focusing on the column which contains the leading 1 of r2 we can
likewise show that c2 = 0 and so on. As ci = 0 for each i then the non-zero rows ri are
independent.
We shall prove (c) first and then (b) follows from it. Recall that
m
X
(i, j) th entry of BA = bis asj (1 ⩽ i ⩽ k, 1 ⩽ j ⩽ n).
s=1

Thus the ith row of BA is the row vector

m m
! m
X X X
bis as1 , . . . , bis asn = bis (as1 , as2 , · · · , asn ), (4.1)
| {z }
s=1 s=1 s=1
sth row of A

which is a linear combination of the rows of A. So every row of BA is in Row(A). A vector

in the row space Row(BA) is a linear combination of BA’s rows which, in turn, are linear
combinations of A’s rows. Hence Row(BA) is contained in Row(A). Because R = Ek · · · E1 A
for some elementary matrices E1 , E2 , . . . , Ek then (b) follows from (c) with B = Ek · · · E1 . Now
(d) also follows from (c). We know Row(BA) is contained in Row(A) and likewise Row(A) =
Row(B −1 (BA)) is contained in Row(BA). Finally (e) follows from (d) by taking B = Ek · · · E1
which is invertible as elementary matrices are invertible.

Corollary 118 (Test for Independence) Let A be an m×n matrix. Then RRE(A) contains
a zero row if and only if the rows of A are dependent.

Proof. We have that RRE(A) = BA where B is a product of elementary matrices and so

invertible. Say the ith row of BA is 0. By (4.1)
m
X
0= bis (sth row of A).
s=1

Now bis are the entries of the ith row of B which, as B is invertible, cannot all be zero. The
above then shows the rows of A are linearly dependent.

BASES 41
 Conversely suppose that the rows of A are linearly dependent. Let r1 , r2 , . . . , rm denote the
rows of A and, without any loss of generality, assume that rm = c1 r1 + · · · + cm−1 rm−1 for real
numbers c1 , . . . , cm−1 . By performing the EROs A1m (−c1 ), . . . , A(m−1)m (−cm−1 ) we arrive at
a matrix whose mth row is zero. We can continue to perform EROs on the top m − 1 rows,
leaving the bottom row untouched, until we arrive at a matrix in RRE form. Once we have
shown RRE form is unique (to follow) then we have that RRE(A) has a zero row.

Corollary 119 (Test for a Spanning Set) Let A be an m × n matrix. Then the rows of A
span Rn if and only if  
In
RRE (A) = .
0(m−n)n

Proof. Let r1 , r2 , . . . , rm be the rows of A in Rn and suppose they span Rn . Now row space is
invariant under EROs. If it were the case that the ith column of RRE (A) does not contain a
leading 1 then ei would not be in the row space. Consequently every column contains a leading
1 and so  
In
RRE (A) = .
0(k−n)n
Conversely if RRE (A) has the above form then the rows of RRE (A) are spanning and hence
so are the original rows r1 , r2 , . . . , rm .

Remark 120 The above corollaries show that if vectors v1 , . . . , vk are linearly independent in
Rn then k ⩽ n. (For if k > n then the RRE form will necessarily have a zero row.) They
further show that if v1 , . . . , vk are spanning then there must be n leading 1s and hence we must
have k ⩾ n. This then shows that a basis, any basis, of Rn contains n vectors.
The above takes a coordinate approach, and relies on some results we are yet to prove –
especially uniqueness of the RRE form. We will shortly prove this result more formally, without
making use of coordinates, but we will see that this is generally true of finite-dimensional vector
spaces. This common cardinality of all bases is called the dimension of the vector space.

Example 121 (a) The vectors v1 = (1, 2, −1, 0), v2 = (2, 1, 0, 3), v3 = (0, 1, 1, 1) in R4 are
linearly independent. If we row reduce the matrix with rows v1 , v2 , v3 we get
   
1 2 −1 0 1 0 0 1.6
 2 1 0 3  RRE →  0 1 0 −0.2 
0 1 1 1 0 0 1 1.2

and hence the three vectors are independent because there is no zero row.
(b) A vector x = (x1 , x2 , x3 , x4 ) is a linear combination of v1 , v2 , v3 if and only if 8x1 +6x3 =
x2 + 5x4 . One way to see this is to row reduce the matrix
 
1 2 −1 0
 2 1 0 3 
 ,
 0 1 1 1 
x1 x2 x 3 x4

BASES 42
which reduces to  
1 0 0 1.6
 0
 1 0 −0.2 

 0 0 1 1.2 
0 0 0 x4 − 1.6x1 + 0.2x2 − 1.2x3
which has a zero row if and only 8x1 + 6x3 = x2 + 5x4 .

BASES 43
4.4 Addendum

In this addendum we will show that the RRE form a matrix is unique. (Recall we’ve already
shown existence.) The proof is somewhat technical and the proof (but not knowledge of the
result) is off-syllabus, but is included for completeness. This also allows us to define row rank.

Theorem 122 (Uniqueness of RRE Form) The reduced row echelon form of an m × n
matrix A is unique.

Proof. The proof below follows by fixing the number of rows m and arguing by induction on
the number of columns n. The only m × 1 matrices which are in RRE form are 0 and eT1 . The
zero m × 1 matrix will reduce to the former and non-zero m × 1 matrices to the latter. In
particular, the RRE form of an m × 1 matrix is unique.
Suppose, as our inductive hypothesis, that all m×(n−1) matrices M have a unique reduced
row echelon form RRE(M ). Let A be an m × n matrix and let à denote the m × (n − 1) matrix
comprising the first n − 1 columns of A. Given any EROs which reduce A to RRE form, these
EROs also reduce à to RRE(Ã) which is unique by hypothesis. Say RRE(Ã) has r non-zero
rows.
There are two cases to consider: (i) any RRE form of A has one more non-zero row than
RRE(Ã); (ii) any RRE form of A has the same number of non-zero rows as RRE(Ã). These
can be the only cases as the first n − 1 columns of an RRE form of A are those of RRE(Ã) and
both matrices are in RRE form; note further that an extra non-zero row in any RRE form of
A, if it exists, must equal en . Case (i) occurs if en is in the row space of A and case (ii) if not.
In particular, it is impossible that different sets of EROs might reduce a given A to both cases
(i) and (ii).
So RRE(A) has one of the following two forms:
 
non-zero 0    
 ..  non-zero * r1 (R)
 RRE(Ã) .   ..   .. 
 RRE(Ã) .  .
(i)  , (ii) = .
   
rows 0  
 rows
 

 0 ··· 0
 *   r r (R) 
1 
m − r zero rows m − r zero rows
m − r − 1 zero rows

In case (i) the last column of any RRE form of A is eTr+1 and so we see that RRE(A) is uniquely
determined as we also know the first n − 1 columns to be RRE(Ã) by our inductive hypothesis.
In case (ii), then any RRE form of A and RRE(Ã) both have r non-zero rows. Let R1 and
R2 be RRE forms of A. By hypothesis, their first n − 1 columns agree and equal RRE(Ã). By
Proposition 117(e),
Row(R1 ) = Row(A) = Row(R2 ).
In particular, this means that the rows rk (R1 ) of R1 are linear combinations of the rows rk (R2 )
of R2 . So, for any 1 ⩽ i ⩽ r, there exist real numbers λ1 , . . . , λr such that
X X
r
ri (R1 ) = k=1 λk rk (R2 ) and hence r i (RRE( Ã)) = k=1 λk rk (RRE(Ã))
r

ADDENDUM 44
by focusing on the first n − 1 columns. RRE(Ã) is in RRE form and so its non-zero rows are
independent; it follows that λi = 1 and λj = 0 for j ̸= i. In particular ri (R1 ) = ri (R2 ) for each
i and hence R1 = R2 as required.
We may now define:

Definition 123 The row rank, or simply rank, of a matrix A is the number of non-zero
rows in RRE(A). We write this as rank(A). The uniqueness of RRE(A) means row rank is
well-defined.

Corollary 124 Let (A|b) be the matrix representing the linear system Ax = b. Then the
system is consistent (i.e. has at least one solution) if and only if rank(A|b) = rank(A).

Proof. Note this result was already demonstrated for systems in RRE form during the proof
of Proposition 44. Say that RRE(A) = P A where P is a product of elementary matrices that
reduce A.
Now if E is an elementary matrix then RRE(EA) = RRE(A) by the uniqueness of RRE
form and so rank(EA) = rank(A). We then have

rank(A|b) = rank(A) ⇐⇒ rank(P A|P b) = rank(P A)

⇐⇒ the system P Ax = P b is consistent
⇐⇒ the system Ax = b is consistent

as the set of solutions to Ax = b is unaffected by EROs.

Proposition 125 Let A be an m × n matrix and b in Rm col .

(a) the system Ax = b has no solutions if and only if 0 0 · · · 0 1 is in Row(A|b).
If the system Ax = b is consistent then
(b) there is a unique solution if and only if rank(A) = n. It follows that m ⩾ n.
(c) there are infinitely many solutions if rank(A) < n. The set of solutions is an n−rank(A)
parameter family.

Proof. As we know that (A|b) can be put into RRE form, and that EROs affect neither the
row space nor the set of solutions, the above is just a rephrasing of Proposition 44.

Remark 126 One might rightly guess that there is the equivalent notion of column rank.
Namely the number of non-zero columns remaining when a matrix is similarly reduced us-
ing ECOs (elementary column operations). It is the case, in fact, that column rank and row
rank are equal and we will prove this later. So we may refer to the rank of a matrix without
ambiguity. ■

ADDENDUM 45
5. DIMENSION

We are now in a position to define the dimension of a vector space with a basis, and to show
that dimension is well-defined. Implicitly we have already seen this result in the tests for linear
independent sets and for spanning sets. We showed in those tests that a linear independent
subset of Rn cannot have more than n elements and that a spanning set of Rn cannot have
fewer than n. The proof below has the merit of not relying on coordinates.

Theorem 127 (Steinitz Exchange Lemma) Let V be a vector space over a field F. Take
X = {v1 , v2 , . . . , vn } ⊆ V . Suppose that u ∈ ⟨X⟩ but that u ̸∈ ⟨X\{vi }⟩ for some i. Let

Y = (X\{vi }) ∪ {u}

(that is, we “exchange u for vi ”). Then ⟨Y ⟩ = ⟨X⟩.

Proof. Since u ∈ ⟨X⟩, there are α1 , . . . , αn ∈ F such that

u = α1 v1 + · · · + αn vn .

There is vi ∈ X such that u ̸∈ ⟨X\{vi }⟩. Without loss of generality, we may assume that i = n.
Since u ̸∈ ⟨X\{vn }⟩, we see that αn ̸= 0. So we can divide by αn and rearrange, to obtain

1
vn = (u − α1 v1 − · · · − αn−1 vn−1 ).
αn
Now if w ∈ ⟨Y ⟩ then we have an expression of w as a linear combination of elements of Y .
We can replace u by α1 v1 + · · · + αn vn to express w as a linear combination of elements of X.
So ⟨Y ⟩ ⊆ ⟨X⟩. And if w ∈ ⟨X⟩ then we have an expression of w as a linear combination of
elements of X. We can replace vn by
1
(u − α1 v1 − · · · − αn−1 vn−1 )
αn
to express w as a linear combination of elements of Y . So ⟨Y ⟩ ⊇ ⟨X⟩.
The Steinitz Exchange Lemma is called a lemma, which sounds unimportant, and it looks
a bit like a niche technical result. But in fact it is fundamental to defining the dimension of a
vector space.

Theorem 128 Let V be a vector space. Let S, T be finite subsets of V . If S is linearly

independent and T spans V , then |S| ⩽ |T |.

Proof. Assume that S is linearly independent and that T spans V . List the elements of S as
u1 , . . . , um and the elements of T as v1 , . . . , vn . We will use the Steinitz Exchange Lemma to
swap out the elements of T with those of S, one at a time, ultimately exhausting S.

DIMENSION 46
 Let T0 = {v1 , . . . , vn }. Since ⟨T0 ⟩ = V , then u1 ∈ ⟨v1 , . . . , vi ⟩ for some 1 ⩽ i ⩽ n and choose
i to be minimal in this regard. Note then that u1 ∈ ⟨v1 , . . . , vi ⟩ but that u1 ∈ / ⟨v1 , . . . , vi−1 ⟩.
The Steinitz Exchange Lemma then shows that

⟨v1 , . . . , vi ⟩ = ⟨u1 , v1 , . . . , vi−1 ⟩

and hence

V = ⟨v1 , . . . , vn ⟩
= ⟨v1 , . . . , vi ⟩ + ⟨vi+1 , . . . , vn ⟩
= ⟨u1 , v1 , . . . , vi−1 ⟩ + ⟨vi+1 , . . . , vn ⟩
= ⟨u1 , v1 , . . . , vi−1 , vi+1 , . . . , vn ⟩.

Now, by relabelling the elements of T, we can assume without loss of generality assume that
u1 has been exchanged for v1 and we set

T1 = {u1 , v2 , . . . , vn } noting that ⟨T1 ⟩ = V.

We proceed inductively in this manner creating sets

Tk = {u1 , . . . , uk , vk+1 , . . . , vn } such that ⟨Tk ⟩ = V..

Note that at each stage uk+1 ∈ ⟨Tk ⟩ but that uk+1 ̸∈ ⟨u1 , . . . , uk ⟩ as the set S is independent.
Hence we can keep continuing to replace elements of T with elements of S. The process can
only terminate when S is exhausted which means that m ⩽ n.

Corollary 129 Let V be a finite-dimensional vector space. All bases of V are finite and of the
same size.

Proof. Since V is finite-dimensional then V has a finite basis B. By Theorem 128 any finite
linearly independent subset of V has size at most |B| . Given another basis S of V , it is linearly
independent, so every finite subset of S is linearly independent. So in fact S must be finite, and
|S| ⩽ |B|. But B is linearly independent and S is spanning and so by Theorem 128 |B| ⩽ |S|.

Definition 130 Let V be a finite-dimensional vector space. The dimension of V , written

dim V , is the size of any basis of V .

Definition 131 We can now redefine row rank using this notion of dimension. The row rank
of a matrix is the dimension of its row space. When in RRE form, the non-zero rows of the
matrix are linearly independent. Further EROs do not affect the row space. So the non-zero
rows of a matrix in RRE form are a basis of the row space.

DIMENSION 47
5.1 Subspaces and Dimension

We include the following result here as it fits in naturally with some of the subsequent results;
in what follows we will show:

• A spanning set contains a basis.

• A linearly independent set can be extended to a basis. (This result requires the notion of
dimension.)

• A basis is a maximal linearly independent set.

• A basis is a minimal spanning set.

Proposition 132 Let V be a vector space over F and let S be a finite spanning set. Then S
contains a basis.

Remark 133 That is, if V has a finite spanning set, then V has a basis. We say nothing here
about what happens if V does not have a finite spanning set. This question is addressed in the
Part B course on Set Theory (using the Axiom of Choice).

Proof. Let S be a finite spanning set for V . Take T ⊆ S such that T is linearly independent,
and T is a largest such set (that is, no linearly independent subset of S strictly contains T ).
Suppose, for a contradiction, that ⟨T ⟩ =
̸ V . Then, since ⟨S⟩ = V , there must exist v ∈ S\⟨T ⟩.
Now by Lemma 102 we see that T ∪ {v} is linearly independent, and T ∪ {v} ⊆ S, and
|T ∪ {v}| > |T |, which contradicts the maximality of T . So T spans V , is linearly independent,
and thus a basis.

Proposition 134 Let U be a subspace of a finite-dimensional vector space V . Then U is

finite-dimensional, and dim U ⩽ dim V . Further if dim U = dim V then U = V.

Proof. Let n = dim V . Then, by Theorem 128, every linearly independent subset of V has
size at most n. Let S be a largest linearly independent set contained in U (and so in V ), so
|S| ⩽ n.
Suppose, for a contradiction, that ⟨S⟩ ̸= U . Then there exists u ∈ U \⟨S⟩. Now by Lemma
102 S ∪ {u} is linearly independent, and |S ∪ {u}| > |S|, which contradicts our choice of S. So
U = ⟨S⟩ and S is linearly independent, so S is a basis of U , and as we noted earlier |S| ⩽ n.
Say now that dim U = dim V and U ̸= V. Then there exists v ∈ V \U. This v may then be
added to a basis of U to create a linearly independent subset of V with

dim U + 1 = dim V + 1

vectors, which is a contradiction. Hence dim U = dim V implies U = V.

Proposition 135 Let V be a finite-dimensional vector space over F and let S be a linearly
independent set. Then there exists a basis B such that S ⊆ B.

SUBSPACES AND DIMENSION 48

Proof. If ⟨S⟩ = V then we are done as S is linearly independent and spanning, and so a basis.
If ⟨S⟩ =
̸ V then by Lemma 102 we can extend S to S1 = S ∪ {u1 } where u1 ∈ U \⟨S⟩ to create a
larger linearly independent set. If ⟨S1 ⟩ = V, then we are done as S1 is a basis. This process can
continue and only terminates at some Sk if Sk is a basis. However this process must terminate
as we know every linearly independent subset of V must contain at most dim V elements.

Corollary 136 A maximal linearly independent subset of a finite-dimensional vector space is

a basis.

Proof. Let S be a maximal linearly independent subset of a finite-dimensional vector space V.

If ⟨S⟩ =
̸ V then by Lemma 102 we can extend S1 = S ∪ {u1 } which is still linearly independent,
but contradicts the maximality of S. So ⟨S⟩ = V.

Corollary 137 A minimal spanning subset of a finite-dimensional vector space is a basis.

Proof. Let S be a minimal spanning subset of a finite-dimensional vector space V. If S is not

linearly independent, then there exists v ∈ S which can be written as a linear combination of
elements of S\{v}. Then S\{v} is linearly independent, and as shown in Lemma 102 S\{v} is
still spanning, which contradicts the minimality of S.

Question Let S be a finite set of vectors in Rn . How can we (efficiently) find a basis of ⟨S⟩?

Example 138 Let S = {(0, 1, 2, 3), (1, 2, 3, 4), (2, 3, 4, 5)} ⊆ R4 . Define
 
0 1 2 3
A =  1 2 3 4 .
2 3 4 5

So ⟨S⟩ = Row(A). Applying EROs to A does not change the row space. Now
 
1 0 −1 −2
RRE(A) =  0 1 2 3 .
0 0 0 0

As has been commented before, the non-zero rows are a basis for the row space, or equivalently
for ⟨S⟩.

5.2 The dimension formula

We previously saw that the sum U + W and intersection U ∩ W of two subspaces are subspaces.
We now prove a useful theorem connecting their dimensions. Recall that we can extend bases
of subspaces to bases of larger spaces, but in general a basis of a vector space won’t contain a
basis of a subspace (or possibly even any elements from the subspace). Thus it makes sense to
begin with U ∩ W, the smallest of the relevant spaces.
The next result is particularly useful.

THE DIMENSION FORMULA 49

Theorem 139 (Dimension Formula) Let U , W be subspaces of a finite-dimensional vector
space V over F. Then

dim(U + W ) + dim(U ∩ W ) = dim U + dim W.

Proof. Take a basis v1 , . . . , vm of U ∩ W . Now U ∩ W ⩽ U and U ∩ W ⩽ W , so by Theorem

135 we can separately extend this set to a basis v1 , . . . , vm , u1 , . . . , up of U , and a basis v1 ,
. . . , vm , w1 , . . . , wq of W . With this notation, we see that

dim(U ∩ W ) = m, dim U = m + p, dim W = m + q.

We aim to show that S = {v1 , . . . , vm , u1 , . . . , up , w1 , . . . , wq } is a basis of U + W. It contains

m + p + q = (m + p) + (m + q) − m
= dim U + dim W − dim(U ∩ W )

elements. So if we can verify this aim then the result follows.

S is spanning: Take x ∈ U + W, so that x = u + w for some u ∈ U , w ∈ W . Then

u = α1 v1 + · · · + αm vm + α1′ u1 + · · · + αp′ up ,
w = β1 v1 + · · · + βm vm + β1′ w1 + · · · + βq′ wq

for some scalars αi , αi′ , βi , βi′ ∈ F. Then

x = u + w = (α1 + β1 )v1 + · · · + (αm + βm )vm + α1′ u1 + · · · + αp′ up + β1′ w1 + · · · + βq′ wq ∈ ⟨S⟩,

showing ⟨S⟩ = U + W .
S is linearly independent: Take α1 , . . . , αm , β1 , . . . , βp , γ1 , . . . , γq ∈ F such that

α1 v1 + · · · + αm vm + β1 u1 + · · · + βp up + γ1 w1 + · · · + γq wq = 0.

Then
α1 v1 + · · · + αm vm + β1 u1 + · · · + βp up = −(γ1 w1 + · · · + γq wq ).
The vector on the LHS is in U , and the vector on the RHS is in W . So they are both in U ∩ W .
As v1 , . . . , vm form a basis of U ∩ W , there are λ1 , . . . , λm ∈ F such that

−(γ1 w1 + · · · + γq wq ) = λ1 v1 + · · · + λm vm ,

which rearranges to
γ1 w1 + · · · + γq wq + λ1 v1 + · · · + λm vm = 0.
But {v1 , . . . , vm , w1 , . . . , wq } is linearly independent (it’s a basis for W ), and so each γi is 0.
This then implies that

α1 v1 + · · · + αm vm + β1 u1 + · · · + βp up = 0.

But {v1 , . . . , vm , u1 , . . . , up } is linearly independent (it’s a basis for U ), so each αi and βi equals
0. So S is linearly independent and the result follows.

THE DIMENSION FORMULA 50

Example 140 Let V be a vector space of dimension 10. Let X, Y be subspaces of dimension
6. Then X + Y ⩽ V so dim(X + Y ) ⩽ dim V = 10. So, by the dimension formula,

dim(X ∩ Y ) = dim(X) + dim(Y ) − dim(X + Y ) ⩾ 6 + 6 − 10 = 2.

It is not hard to see that the possibilities

2 ⩽ dim(X ∩ Y ) ⩽ 6,

are all possible. This is left as an exercise.

Definition 141 Let U , W be subspaces of a vector space V . If U ∩W = {0V } and U +W = V ,

then we say that V is the direct sum of U and W , and we write V = U ⊕ W .

Proposition 142 Let U , W be subspaces of a finite-dimensional vector space V . The following

are equivalent:
(a) V = U ⊕ W ;
(b) every v ∈ V has a unique expression as u + w where u ∈ U and w ∈ W ;
(c) dim V = dim U + dim W and V = U + W ;
(d) dim V = dim U + dim W and U ∩ W = {0V };
(e) if u1 , . . . , um is a basis for U and w1 , . . . , wn is a basis for W , then u1 , . . . , um , w1 ,
. . . , wn is a basis for V .

Proof. Exercise. Hint: (a) ⇔ (b) follows from the definition of direct sum.
Try using the dimension formula to prove that (a)/(b) are equivalent to (c)/(d)/(e).

Remark 143 To conclude, two more general comments on direct sums.

• A vector space V is said to be the direct sum of subspaces X1 , . . . , Xk ⩽ V if every v ∈ V

can be uniquely written

v = x1 + x2 + · · · + xk where xi ∈ X i for all i.

Thus it is statement (b) in the proposition above which naturally generalizes.

• Writing a vector space as a sum of subspaces is called an internal direct sum. Given
vectors spaces V1 , . . . , Vk then the external direct sum

V1 ⊕ V2 ⊕ · · · ⊕ Vk

has the Cartesian product V1 ×V2 ×· · ·×Vk as the underlying set, with addition and scalar
multiplication defined componentwise. That is

(v1 , . . . , vk ) + (w1 , . . . , wk ) = (v1 + w1 , . . . , vk + wk );

α.(v1 , . . . , vk ) = (α.v1 , . . . , α.vk )

THE DIMENSION FORMULA 51

6. LINEAR TRANSFORMATIONS

We have objects with some structure (vector spaces). This section is about structure-preserving
maps between these objects. You will see a similar phenomenon in lots of other contexts too –
whenever we have objects with some kind of structure, we can ask about structure-preserving
maps between objects. (This can lead to further abstraction, which is explored in Category
Theory, an interesting part of mathematics and a Part C course.)

6.1 What is a linear transformation?

Definition 144 Let V , W be vector spaces over F. We say that a map T : V → W is linear
if
(i) T (v1 + v2 ) = T (v1 ) + T (v2 ) for all v1 , v2 ∈ V (preserves additive stucture);
and
(ii) T (λv) = λT (v) for all v ∈ V and λ ∈ F (preserves scalar multiplication).
We call T a linear transformation or a linear map.

Proposition 145 Let V , W be vector spaces over F, let T : V → W .be a linear map. Then
T (0V ) = 0W .

Proof. Note that T (0V ) + T (0V ) = T (0V + 0V ) = T (0V ), and hence T (0V ) = 0W .

Proposition 146 Let V , W be vector spaces over F, let T : V → W . The following are
equivalent:

(a) T is linear;

(b) T (αv1 + βv2 ) = αT (v1 ) + βT (v2 ) for all v1 , v2 ∈ V and α, β ∈ F;

(c) for any n ⩾ 1, if v1 , . . . , vn ∈ V and α1 , . . . , αn ∈ F then

T (α1 v1 + · · · + αn vn ) = α1 T (v1 ) + · · · + αn T (vn ).

Proof. Exercise.

Example 147 • Let V be a vector space. Then the identity map idV : V → V given by
idV (v) = v for all v ∈ V is a linear map.

• Let V , W be vector spaces. The zero map 0 : V → W that sends every v ∈ V to 0W is

a linear map.

LINEAR TRANSFORMATIONS 52
 • For m, n ⩾ 1, with A ∈ Mm×n (R). Then we define the left multiplication map
LA : Rncol → Rm
col by LA (v) = Av for v ∈ Rncol .
This is a linear map. Similarly, we have a right multiplication map
RA : Rm → Rn by RA (v) = vA for v ∈ Rm .

• Take m, n, p ⩾ 1 with A ∈ Mm×n (R). The left multiplication map Mn×p (R) →
Mm×p (R) sending X to AX is a linear map.
• Let V be a vector space over F with subspaces U , W such that V = U ⊕ W . For v ∈ V
there are unique u ∈ U , w ∈ W such that v = u + w. Define P : V → V by P (v) = w.

Proposition 148 P is a linear map. P is called projection of V onto W along U .

Proof. Take v1 , v2 ∈ V and α1 , α2 ∈ F. Then there are u1 , u2 ∈ U , w1 , w2 ∈ W such

that v1 = u1 + w1 and v2 = u2 + w2 . Now
α1 v1 + α2 v2 = α1 (u1 + w1 ) + α2 (u2 + w2 )
= (α1 u1 + α2 u2 ) + (α1 w1 + α2 w2 )
where α1 u1 + α2 u2 ∈ U and α1 w1 + α2 w2 ∈ W . So by uniqueness
P (α1 v1 + α2 v2 ) = α1 w1 + α2 w2 = α1 P (v1 ) + α2 P (v2 ).

• For A = (aij ) ∈ Mn×n (R), we define the trace of A to be

trace(A) : = a11 + a22 + · · · + ann ,
(the sum of the entries on the main diagonal of A). The map trace : Mn×n (R) → R is a
linear map.
• Let Rn [x] be the vector space of polynomials of degree at most n. Define D : Rn [x] → Rn [x]
by p(x) 7→ p′ (x), that is,
D(an xn + · · · + a1 x + a0 ) = nan xn−1 + · · · + a1 .
This is a linear map from Rn [x] to Rn [x]. We could also think of it as a linear map Rn [x]
to Rn−1 [x].
• Let C 1 (R) be the subspace of RR consisting of differentiable functions f : R → R. The
differential operator D : C 1 (R) → RR sending f to f ′ is a linear map.
• Let C ∞ (R) be the subspace of RR consisting of functions f : R → R that are infinitely
differentiable. The differential operator D : C ∞ (R) → C ∞ (R) sending f to f ′ is a linear
map.
• Let X be a set, let V = RX be the space of real valued functions on X. For a ∈ X, the
evaluation map Ea : V → R sending f to f (a) is a linear map.

WHAT IS A LINEAR TRANSFORMATION? 53

6.2 Combining linear transformations

We can add linear transformations (pointwise), and we can multiply a linear transformation by
a scalar (pointwise).

Proposition 149 Let V , W be vector spaces over a field F. For S, T : V → W and λ ∈ F, we

may define linear maps S + T and λS by
S+T: V →W by (S + T )(v) = S(v) + T (v) for v ∈V;
λS : V → W by (λS)(v) = λS(v) for v ∈ V.
With these operations (and the zero map 0 : V → W ), the set of linear transformations V → W
forms a vector space denoted Hom(V, W ).

Proof. Firstly S + T is a linear map as

(S + T )(α1 v1 + α2 v2 ) = S(α1 v1 + α2 v2 ) + T (α1 v1 + α2 v2 ) [by definition]
= α1 S(v1 ) + α2 S(v2 ) + α1 T (v1 ) + α2 T (v2 ) [by linearity]
= α1 (S(v1 ) + T (v1 )) + α2 (S(v2 ) + T (v2 )) [rearranging]
= α1 (S + T )(v1 ) + α2 (S + T )(v2 ) [by definition]
showing linearity. That λS is linear is left as an exercise. Verifying the vector space axioms for
Hom(V, W ) involves showing:
(i) S + T = T + S: this follows from commutativity of + in W.
(ii) S + (T + U ) = (S + T ) + U : this follows from associativity of + in W.
(iii) S + 0 = S: this follows from properties of 0W .
def
(iv) S has an additive inverse (−S)(v) = −(S(v)).
(v)
λ(S + T ) = (λS) + (λT ); (λ + µ)S = (λS) + (µS); (λµ)S = λ(µS); 1S = S.
These properties follow from the same properties for vectors in W as addition and scalar
multiplication of linear maps is defined pointwise.

We can also compose linear transformations.

Proposition 150 Let U , V , W be vector spaces over F. Let S : U → V and T : V → W be

linear. Then T ◦ S : U → W is linear.

Proof. Take u1 , u2 ∈ U and λ1 , λ2 ∈ F. Then

(T ◦ S)(λ1 u1 + λ2 u2 ) = T (S(λ1 u1 + λ2 u2 )) [definition of composition]
= T (λ1 S(u1 ) + λ2 S(u2 )) [S is linear]
= λ1 T (S(u1 )) + λ2 T (S(u2 )) [T is linear]
= λ1 (T ◦ S)(u1 ) + λ2 (T ◦ S)(u2 ) [definition of composition]
so T ◦ S is linear.

COMBINING LINEAR TRANSFORMATIONS 54

Remark 151 We often write T ◦ S as T S. The notation T ◦ S removes any possible ambiguity
about the order of the functions.
Definition 152 Let V , W be vector spaces and let T : V → W be linear. We say that T is
invertible if there is a linear transformation S : W → V such that ST = idV and T S = idW
(where idV and idW are the identity maps on V and W respectively). In this case, we call S
the inverse of T , and write it as T −1 . An invertible linear map is called an isomorphism.
Remark 153 T is a function, so if it is invertible then it has a unique inverse (you saw this
in the Introduction to University Maths course), so there is no ambiguity in writing T −1 .
Proposition 154 Let V , W be vector spaces. Let T : V → W be linear. Then T is invertible
if and only if T is bijective.
Proof. (⇒) If T is invertible, then it is certainly bijective (see the Introduction to University
Maths course).
(⇐) Assume that T is bijective.
Then T has an inverse function S : W → V , but it remains to show that S is linear. Take
w1 , w2 ∈ W and λ1 , λ2 ∈ F. Let v1 = S(w1 ), v2 = S(w2 ). Then
T (v1 ) = T S(w1 ) = w1 and T (v2 ) = T S(w2 ) = w2 .
Now
S(λ1 w1 + λ2 w2 ) = S(λ1 T (v1 ) + λ2 T (v2 ))
= S(T (λ1 v1 + λ2 v2 )) [since T is linear]
= λ1 v1 + λ2 v [as S is inverse to T ]
= λ1 S(w1 ) + λ2 S(w2 ).
So S is linear.
Proposition 155 Let U , V , W be vector spaces. Let S : U → V and T : V → W be invertible
linear transformations. Then T S : U → W is invertible, and (T S)−1 = S −1 T −1 .
Proof. Exercise.
Proposition 156 (a) Let V, W be vector spaces with V finite-dimensional. If there is an
invertible linear map T : V → W then dim V = dim W.
(b) Let V, W be finite-dimensional vector spaces with dim V = dim W. Then there is an
invertible linear map T : V → W.
Consequently V and W are isomorphic if and only if dim V = dim W.
Proof. (a) Let v1 , . . . , vn be a basis for V. It is left as an exercise to show that T v1 , . . . , T vn is
a basis for W. Then n = dim V = dim W.
(b) Let n = dim V = dim W . Let v1 , . . . , vn be a basis for V and w1 , . . . , wn be a basis for
W. It is left as an exercise to show that
X  X
T: V →W given by T αi v i = αi w i

is a well-defined, invertible linear map.

COMBINING LINEAR TRANSFORMATIONS 55

Example 157 Let V = R[x] denote the vector space of polynomials in x with real coefficients,
and let W denote the vector space of real sequences (an )∞
n=0 . Then V and W are both infinite-
dimensional but are not isomorphic.

Solution. The set B = {1, x, x2 , x3 , . . .} is a basis for V . That it is linearly independent shows
that V is not finite-dimensional. The set of sequences {(δin )∞ n=0 | i ⩾ 0} is linearly independent
and so W is also infinite-dimensional.
However the set S = {(tn )∞n=0 | t ∈ R} is an uncountable linearly independent subset of W
and hence W does not have a countable basis. We prove that S is linearly independent below.
Suppose that
α1 (tn1 ) + · · · + αk (tnk ) = (0)
for real numbrs α1 , . . . , αk , t1 , . . . , tk with the ti distinct. Then for all n ⩾ 0 we have

α1 tn1 + · · · + αk tnk = 0.

These equations for 0 ⩽ n < k can be rewritten as the single matrix equation
    
1 1 1 ··· 1 α1 0
 t1
 t 2 t 3 · · · t 
k 
 α 
2   0 
 
 t2 t22 t23 · · · t2k 
  α3  =  0  .
   
 1
 .. .. .. . . . ..   .   . 
 . . . .   ..   .. 
tk−1
1 tk−1
2 tk−1
3 · · · tkk αk 0

As the ti are distinct, then the above k ×k matrix is invertible – this is proved in Linear Algebra
II next term. Hence S is an uncountable linearly independent set. No such set exists in V and
hence W is not isomorphic to V.

6.3 Rank and nullity

Definition 158 Let V , W be vector spaces. Let T : V → W be linear. We define the kernel
(or null space) of T to be
ker T := {v ∈ V | T (v) = 0W }.
We define the image of T to be

Im T := {T (v) | v ∈ V }.

Here are some useful properties of kernels and images.

Proposition 159 Let V , W be vector spaces over F. Let T : V → W be linear. Then

(a) ker T is a subspace of V and Im T is a subspace of W ;
(b) T is injective if and only if ker T = {0V }.
(c) if A is a spanning set for V , then T (A) is a spanning set for Im T ;
(d) if V is finite-dimensional, then ker T and Im T are finite-dimensional.

RANK AND NULLITY 56

Proof. (a) Note that ker T ⊆ V and Im T ⊆ W .
ker T Note that T (0V ) = 0W so 0V ∈ ker T .
Take v1 , v2 ∈ ker T and λ ∈ F, so T (v1 ) = T (v2 ) = 0W .
Then T (λv1 + v2 ) = λT (v1 ) + T (v2 ) = λ0W + 0W = 0W , so λv1 + v2 ∈ ker T .
So, by the Subspace Test, ker T ⩽ V .
Im T We have T (0V ) = 0W so 0W ∈ Im T .
Take w1 , w2 ∈ Im T and λ ∈ F. Then there are v1 , v2 ∈ V such that T (v1 ) = w1 and
T (v2 ) = w2 . Then λw1 + w2 = λT (v1 ) + T (v2 ) = T (λv1 + v2 ) ∈ Im T .
So, by the Subspace Test, Im T ⩽ W .
(b) Say that T is 1–1. If v ∈ ker T then T v = 0W = T (0V ). By injectivity we have v = 0V
and so ker T = {0V }.
Conversely say that ker T = {0V } and that T v1 = T v2 . Then T (v1 − v2 ) = T v1 − T v2 = 0W .
Then v1 − v2 = 0V and so T is 1–1.
(c) Let A be a spanning set for V .
Take w ∈ Im T . Then w = T (v) for some v ∈ V .
Now there are v1 , . . . , vn ∈ A and α1 , . . . , αn ∈ F such that v = α1 v1 + · · · + αn vn . So

w = T (α1 v1 + · · · + αn vn ) = α1 T (v1 ) + · · · + αn T (vn ) ∈ ⟨T (A)⟩.

So T (A) spans Im T .
(d) Assume that V is finite-dimensional. Then ker T ⩽ V so ker T is finite-dimensional.
Also, Im T is finite-dimensional by (iii).

Corollary 160 Given a matrix A the image of LA is Col(A), the column space of A and the
image of RA is Row(A), the row space of A.

Proof. The canonical basis e1 , . . . , em spans Rm and hence the rows of A, r1 = e1 A, . . . , rm =

em A span Im RA . Hence Im RA ⊆ Row(A). Conversely if v ∈ Row(A) then

v = α1 r1 + · · · + αm rm = (α1 e1 + · · · + αm em )A ∈ Im RA .

Likewise Im LA = Col(A).

Definition 161 Let V , W be vector spaces with V finite-dimensional. Let T : V → W be

linear. We define the nullity of T to be nullity(T ) := dim(ker T ), and the rank of T to be
rank(T ) := dim(Im T ).
Given the previous corollary, the rank of LA equals the column rank of A and the rank of
RA equals the row rank of A.

The next theorem is very important!

Theorem 162 (Rank-Nullity Theorem) Let V , W be vector spaces with V finite-dimensional.

Let T : V → W be linear. Then

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

RANK AND NULLITY 57

Proof. Take a basis v1 , . . . , vn for ker T , where n = nullity(T ) and extend this to a basis v1 ,
. . . , vn , v1′ , . . . , vr′ of V so that dim V = n + r. We claim that B = {T v1′ , . . . , T vr′ } is a basis
for Im T .
B spans Im T :
By Proposition 159(c), T (v1 ), . . . , T (vn ), T (v1′ ), . . . , T (vr′ ) span Im T, being the images of
a basis. But v1 , . . . , vn ∈ ker T , meaning T (v1 ) = · · · = T (vn ) = 0W , so in fact T v1′ , . . . , T vr′
span Im T .
B is linearly independent:
Take α1 , . . . , αr ∈ F such that

α1 T (v1′ ) + · · · + αr T (vr′ ) = 0W .

As T is linear, we can rewrite this as T (α1 v1′ + · · · + αr vr′ ) = 0W . So α1 v1′ + · · · + αr vr′ ∈ ker T.
As v1 , . . . , vn is a basis for ker T , there are β1 , . . . , βn ∈ F such that

α1 v1′ + · · · + αr vr′ = β1 v1 + · · · + βn vn ,

But v1 , . . . , vn , v1′ , . . . , vr′ are linearly independent (being a basis for V ), so

β1 = · · · = βn = α1 = · · · = αr = 0.

This shows w1 , . . . , wr are linearly independent and the claim follows.

Now using the claim we have rankT = r, and so dim(V ) = n + r = nullity(T ) + rank(T ).

Here are a couple of useful results in their own right that also illustrate the usefulness of
the Rank-Nullity Theorem.

Corollary 163 Let V be a finite-dimensional vector space. Let T : V → V be linear. The

following are equivalent:
(a) T is invertible;
(b) rankT = dim V ;
(c) nullityT = 0.

Proof. (a) ⇒ (b):

Assume that T is invertible. Then T is bijective, so is surjective, so Im T = V , meaning
rankT = dim V .
(b) ⇒ (c):
Assume that rankT = dim V . Then, by Rank-Nullity, nullityT = 0.
(c) ⇒ (a):
Assume that nullityT = 0 so that ker T = {0V }. Then T is injective.
Also, by Rank-Nullity, rankT = dim V and Im T ⩽ V , so Im T = V , so T is surjective.
So T is bijective, so T is invertible

The next result is important, and we’ll use it again later in the course.

Corollary 164 Let V be a finite-dimensional vector space. Let T : V → V be linear. Then

any one-sided inverse of T is a two-sided inverse, and so is unique.

RANK AND NULLITY 58

Proof. Suppose that T has a right inverse S : V → V , so T ◦ S = idV . Since idV is surjective,
T is surjective, so rankT = dim V .
So, by the previous corollary T is invertible, say with two-sided inverse S ′ .
Then S ′ = S ′ ◦ idV = S ′ ◦ (T ◦ S) = (S ′ ◦ T ) ◦ S = idV ◦ S = S.
Hence S is the (unique) two-sided inverse.
If instead we suppose that T has a left inverse S : V → V , so S ◦ T = idV , then idV is
injective so that T is injective and hence nullityT = 0, and the argument is similar to the
previous one.
Which, when reworded in terms of matrices, is the following result.

Corollary 165 Let A, B be square matrices of the same size. If AB is invertible then A and
B are invertible.

Lemma 166 Let V and W be vector spaces, with V finite-dimensional. Let T : V → W be

linear and U ⩽ V . Then

dim U − nullityT ⩽ dim T (U ) ⩽ dim U.

In particular, if T is injective then dim T (U ) = dim U .

Proof. Let S : U → W be the restriction of T to U (that is, S(u) = T (u) for all u ∈ U ). Then
S is linear, and ker S ⩽ ker T so nullityS ⩽ nullityT . Also, Im S = T (U ). By Rank-Nullity,

dim T (U ) = dim Im S = dim U − nullityS ⩽ dim U ;

dim T (U ) = dim U − nullityS ⩾ dim U − nullityT.

If T is injective, then nullityT = 0, so dim T (U ) = dim U .

Remark 167 A take on the Rank-Nullity Theorem using matrices.

Let A be an m × n matrix and consider its reduced form RRE(A). We know that there are
as many leading 1s as there are non-zero rows and this is the row rank of A. We also know that
the kernel of A can be parameterized by assigning parameters to every column which does not
have a leading 1. Hence

row rank of A = number of leading 1s;

nullity of A = number of columns without leading 1s.

Hence

n = dim Rncol
= (number of leading 1s) + (number of columns without leading 1s)
= (row rank of A) + (nullity of A) .

Corollary 168 (Criteria for Invertibility) Let A be an n × n matrix. The following state-
ments are equivalent:
(a) A is invertible.

RANK AND NULLITY 59

 (b) A has a left inverse.
(c) A has a right inverse.
(d) Row(A) = Rn .
(e) The columns of A are linearly independent.
(f ) The rows of A are linearly independent.
(g) The only solution x in Rncol to the system Ax = 0 is x = 0.
(h) The row rank of A is n.
(i) RRE(A) = In .

Proof. These are separately left as exercises. Some of the equivalencies have already been
demonstrated.

RANK AND NULLITY 60

7. LINEAR MAPS AND MATRICES

7.1 Representing linear maps with matrices

We saw examples of linear maps arising from multiplying by a matrix: for A ∈ Mm×n (R), we
defined LA : Rncol → Rm m n
col by LA (v) = Av, and we defined RA : R → R by RA (v) = vA.
We shall see that linear maps are to matrices, as vectors are to coordinate vectors. Impor-
tantly recall that a vector has different coordinates in different coordinate systems and that
each choice of coordinates (or basis) associates coordinates with a vector. Similarly given a
linear map T : V → W for each choice of coordinates (or bases) for V and W we will see that
T is represented by a matrix; change your choice of bases and that matrix will change too!

Definition 169 Let V be an n-dimensional vector space over F with an ordered basis V of
vectors v1 , . . . , vn . Let W be an m-dimensional vector space over F with an ordered basis W
of vectors w1 , . . . , wm . So every vector in V and W is represented by a coordinate vector in
Fn and Fm respectively.
Let T : V → W be a linear transformation. We define the matrix for T with respect to
the bases V and W to be the matrix which takes the coordinate vector of v to the coordinate
vector of T v.
More explicitly, this is the m × n matrix A = (aij ) where
m
X
T (vi ) = aki wk .
k=1

We will write W TV for this matrix A.

Remark 170 Firstly note that this matrix is well-defined. For each 1 ⩽ i ⩽ n then T (vi ) can
be uniquely expressed as a linear combination of W.

Remark 171 Further a1i , . . . , ami are the coordinates of T (vi ). These are the entries in the ith
column of A. The coordinate column vector of vi is eTi and the ith column of A is AeTi . So we
can see that the entries of A are the coordinates of the images of the basis V as claimed.
Note that this is what matrices normally do! Given an m × n matrix A then the first column
of A equals Ae1 where e1 = (1, 0, . . . , 0), and more generally the ith column is AeTi .

Remark 172 So if we use V and W to identify V and W with Fncol and Fm

col then we identify
T with LA .

Remark 173 Importantly in this the bases are listed in an order. If the order of either basis
changed then the matrix will change too.

LINEAR MAPS AND MATRICES 61

Remark 174 If V = W and we use the same ordered basis for both domain and codomain of
T : V → V , then we talk about the matrix for T with respect to this basis.

Example 175 Let T : R3 → R3 be defined by T (x, y, z) = (0, x, y). This is linear (check!). If
we take
i = (1, 0, 0), j = (0, 1, 0), k = (0, 0, 1),
as the basis E for both the domain and codomain then we see that

T (1, 0, 0) = (0, 1, 0), T (0, 1, 0) = (0, 0, 1), T (0, 0, 1) = (0, 0, 0).

Hence  
0 0 0
E TE =
 1 0 0 .
0 1 0
Note that (E TE )2 = E TE2 and (E TE )3 = 0 = E TE3 (again check!).

Example 176 Let T : R3 → R3 be defined by T (x, y, z) = (0, x, y) as in the previous example.

Now let F be the ordered basis

f1 = (1, 1, 1), f 2 = (1, 1, 0), f 3 = (0, 1, 1).

Find (i) E TF , (ii) F TE and (iii) F TF .

Solution. (i) Note that

T (1, 1, 1) = (0, 1, 1), T (1, 1, 0) = (0, 1, 1), T (0, 1, 1) = (0, 0, 1).

Hence  
0 0 0
E TF =
 1 1 0 .
1 1 1
(ii) Note that

T (1, 0, 0) = (0, 1, 0) = −f1 + f2 + f3 , T (0, 1, 0) = (0, 0, 1) = f1 − f2 , T (0, 0, 1) = (0, 0, 0).

Hence  
−1 1 0
F TE =
 1 −1 0  .
1 0 0
(iii) Note that

T (1, 1, 1) = (0, 1, 1) = f3 , T (1, 1, 0) = (0, 1, 1) = f3 , T (0, 1, 1) = (0, 0, 1) = f1 − f2 .

Hence  
0 0 1
F TF =
 0 0 −1  .
1 1 0

REPRESENTING LINEAR MAPS WITH MATRICES 62

Proposition 177 Let V be an n-dimensional vector space over F with ordered basis V. Let W
be an m-dimensional vector space over F with ordered basis W. Then
(i) the matrix W 0V of the zero map is 0m×n for any choice of V and W;
(ii) the matrix V idV of identity map In for any choice of V;
(iii) if S : V → W , T : V → W are linear and α, β ∈ F, then

W (αS + βT )V = α (W SV ) + β (W TV ) .

Proof. These are left as exercises.

Theorem 178 Let U , V , W be finite-dimensional vector spaces over F, of dimensions m, n, p,

with ordered bases U, V, W respectively. Let S : U → V and T : V → W be linear. Let

A = V SU and B= W TV .

Then
BA = W T SU .

Proof. Note that A is n × m and B is p × n, so the product matrix BA is p × m. Let U be

u1 , . . . , um , V be v1 , . . . , vn and W be w1 , . . . , wp .
As usual, we write A = (aij ) and B = (bij ). By definition of A and B, we have
n
X
S(ui ) = aji vj for 1 ⩽ i ⩽ m;
j=1
p
X
T (vj ) = bkj wk for 1 ⩽ j ⩽ n.
k=1

Now for 1 ⩽ i ⩽ m we have

n
!
X
(T ◦ S)(ui ) = T (S(ui )) = T aji vj
j=1
n
X
= aji T (vj ) since T is linear
j=1
n p
X X
= aji bkj wk
j=1 k=1
p n
!
X X
= bkj aji wk
k=1 j=1
p
X
= (BA)ki wk .
k=1

Thus, by definition, W T SU = BA.

Remark 179 This is why we define multiplication of matrices in the way that we do!

REPRESENTING LINEAR MAPS WITH MATRICES 63

Remark 180 As we are about to see, this gives a relatively clear and painless proof that matrix
multiplication is associative as composition is associative.
Corollary 181 Take A ∈ Mm×n (F), take B ∈ Mn×p (F), take C ∈ Mp×q (F). Then A(BC) =
(AB)C.
Proof. We consider the left multiplication maps
LA : Fncol → Fm
col and LB : Fpcol → Fncol and LC : Fqcol → Fpcol .
With respect to the standard bases of these spaces, the matrices of LA , LB , LC are A, B, C
respectively.Hence, by the previous theorem A(BC) and (AB)C are the matrices of
LA ◦ (LB ◦ LC ) : Fqcol → Fm
col , and (LA ◦ LB ) ◦ LC : Fqcol → Fm
col

respectively. But composition of functions is associative, so

LA ◦ (LB ◦ LC ) = (LA ◦ LB ) ◦ LC
and hence A(BC) = (AB)C.
Corollary 182 Let V be a finite-dimensional vector space and let T : V → V be an invertible
linear transformation. Let A be the matrix of T with respect to an ordered basis (for both
domain and codomain). Then A is invertible, and A−1 is the matrix of T −1 with respect to the
same basis.
Proof. Exercise.

7.2 Change of basis

Question Take two matrices for the same linear transformation with respect to different bases.
How are the matrices related?
Example 183 Define T : R2 → R2 by T (x, y) = (2x + y, 3x − 2y). To find the matrix of T
with respect to the standard ordered basis E, note that
T (1, 0) = (2, 3) and T (0, 1) = (1, −2)
so the matrix for T with respect to this basis is
 
2 1
E TE = .
3 −2
That is T = LA . Let f1 = (1, −2) and f2 = (−2, 5). Then f1 , f2 is an ordered basis of R2 which
we will denote as F. Note that
T (f1 ) = (0, 7) = 14f1 + 7f2
T (f2 ) = (1, −16) = −27f1 − 14f2

CHANGE OF BASIS 64
so the matrix for T with respect to this basis is
 
14 −27
F TF = .
7 −14
How are these two matrices related? Well, by Theorem 178, we can see that

F TF = (F IE ) (E TE ) (E IF ) .
The matrix E IF represents the identity transformation, so it is does not change vectors; however
it changes the coordinate vector for a vector with respect to some basis F to the coordinate vector
for the same vector with respect to a different basis E. Note that the inverse of this matrix
is F IE .
We can take f1 , f2 and write them with respect to e1 , e2 : we have
f1 = e1 − 2e2 , f2 = −2e1 + 5e2
so we get a ‘change of basis matrix’
 
1 −2
E IF = .
−2 5

If this matrix is applied to (1, 0)T then this coordinate vector represents f1 . The image of the
coordinate vector (1, 0)T is (1, −2)T which represents e1 − 2e2 . But this is of course the same
vector! This vector just has different coordinates with respect to the bases E and F.
It is then the case that
 −1  
1 −2 5 2
F IE = = ,
−2 5 2 1
which represents that
e1 = 5f1 + 2f2 , e2 = 2f1 + f2 .
And we can verify that
(F IE ) (E TE ) (E IF )
   
5 2 2 1 1 −2
=
2 1 3 −2 −2 5
  
16 1 1 −2
=
7 0 −2 5
 
14 −27
= =F TF
7 −14
as expected.

Corollary 184 (Change of basis theorem) Let V be a finite-dimensional vector space over
F with ordered bases V, V ′ . Let W be a finite-dimensional vector space over F with ordered
bases W, W ′ . Let T : V → W be a linear map. Then

W ′ TV ′ = (W ′ IW ) (W TV ) (V IV ′ ) .

CHANGE OF BASIS 65
Proof. This is an immediate corollary to Theorem 178.

Corollary 185 (Change of basis theorem version 2) Let V be a finite-dimensional vector

space over F with ordered bases V, V ′ and let T : V → V be a linear map. Then

V ′ TV ′ = (V ′ IV ) (V TV ) (V IV ′ ) .

If we set A = V ′ TV ′ , B = V TV and P = V IV ′ then note

A = P −1 BP.

Proof. This is a special case of the previous corollary.

Definition 186 Take A, B ∈ Mn×n (F). If there is an invertible n × n matrix P such that
A = P −1 BP , then we say that A and B are similar. Similarity is then an equivalence relation.

Remark 187 So two matrices representing the same linear transformation from a finite-dimensional
vector space to itself, but with respect to different bases, are similar.

Remark 188 Properties of Linear Maps As many different matrices can represent the
same linear transformation T : V → V it would be disturbing if different conclusions about the
properties of T could be determined by using different matrix representatives. For example, if
we said a linear map T is invertible if a matrix representative of it is invertible, could T end
up being invertible and not invertible? Reassuringly the answer is no.
Let A = V TV and B = W TW be matrices representing T with respect to two bases, so that
A = P −1 BP for some invertible P. Then

• A is invertible if and only if B is invertible.

• The trace of A equals the trace of B. [This follows from the identity trace(M N ) =
trace(N M ).]

• A functional identity satisfied by A, such as A2 = A, is also satisfied by B.

• The determinant of A equals the determinant of B. [Determinants will be formally defined

in Linear Algebra II next term.]

• The eigenvalues of A equal the eigenvalues of B. [Eigenvalues will be formally defined in

Linear Algebra II next term.]

Thus we may, in a well-defined fashion, refer to the invertibility, trace, determinant of a

linear map.
Note we cannot, in a well-defined manner, refer to the transpose of a linear map. If A =
P BP then it need not be the case that AT = P −1 B T P. Though you might note that this is
−1

true if P is orthogonal! (This is something that will be addressed when you meet adjoints in
the second year.)

CHANGE OF BASIS 66
7.3 Matrices and rank

For a matrix A ∈ Mm×n (F), we have defined the row space and row rank, and analogously the
column space and column rank. It makes sense to ask if rowrank(A) and colrank(A) related?

Remark 189 From the definitions, we see that Col(A) = Row(AT ) and so colrank(A) =
rowrank(AT ). Similarly, Row(A) = Col(AT ) and so rowrank(A) = colrank(AT ).

We first prove the following:

Lemma 190 The linear system (A|b) is consistent if and only if Col(A|b) = Col(A).

Proof. Say that A is m × n and denote the columns of A as c1 , c2 , . . . , cn . Then

(A|b) is consistent ⇐⇒ Ax = b for some x ∈ Fncol

⇐⇒ x1 c1 + · · · xn cn = b for some x ∈ Fncol
⇐⇒ b ∈ Col(A)
⇐⇒ Col(A|b) = Col(A).

Theorem 191 The column rank of a matrix equals its row rank.

Proof. We prove this by induction on the number of columns in the matrix. A non-zero m × 1
matrix has column rank 1 and also row rank 1 as the matrix reduces to eT1 ; the column rank and
row rank of 0m×1 are both 0. So the n = 1 case is true. Suppose, as our inductive hypothesis,
that column rank and row rank are equal for m × n matrices. Any m × (n + 1) matrix (A|b)
can be considered as an m × n matrix A alongside b in Fm col . If the system (A|b) is consistent
then
colrank(A|b) = colrank(A) [by previous lemma]
= rowrank(A) [by inductive hypothesis]
= rowrank(A|b) [see Remark 45].
On the other hand, if the system (A|b) has no solutions then

colrank(A|b) = (colrankA) + 1 as b ∈/ Col(A)

= (rowrankA) + 1 [by inductive hypothesis]
= rowrank(A|b) [see Remark 45].

So if the system is consistent the row rank and column rank maintain their common value. If
inconsistent, then b adds a further dimension to the column space and (0 0 · · · 0 | 1) adds an
extra dimension to the row space. Either way the column rank and row rank of (A|b) still
agree and the proof follows by induction.

We provide here a second proof of the result, as it takes a somewhat different approach.

MATRICES AND RANK 67

Theorem 192 Let P = QR where P, Q, R are respectively k × l, k × m, m × l matrices over
the same field.
(a) Then the row rank of P is at most m.
(b) Let p = rowrankP . Then P may be written as the product of a k × p matrix and a p × l
matrix.
(c) The row rank and column rank of a matrix are equal.

Solution. (a) By Proposition 117(c) we have

Row(P ) = Row(QR) ⩽ Row(R).

Hence
rowrank(P ) = dim Row(P ) ⩽ dim Row(R) = rowrank(R) ⩽ m.
(b) There is a k × k invertible matrix E such that EP = RRE(P ) and hence

P = E −1 RRE(P ).

Now let Ẽ denote the first p columns of E −1 and P̃ denote the first p rows of RRE(P ). As the
last k − p rows of RRE(P ) are zero rows, we still have

P = Ẽ P̃ ,

where Ẽ is a k × p and P̃ is a p × l matrix.

(c) From (a) and (b) we know that the row rank of a k × l matrix P is the minimal value p
such that P can be written as the product QR of a k × p matrix and a p × l matrix. Whenever
P = QR then
P T = R T QT .
l×k l×p p×k

So the row rank of P T is similarly p. But the rowrankP T = colrankP as required.

MATRICES AND RANK 68

8. INNER PRODUCT SPACES

On first meeting vector spaces, it is quite natural to think of Rn as a typical example. However,
as has already been commented, it is important to appreciate that Rn has a lot of structure
beyond being just a real vector space. It has coordinates already assigned (and so a canonical
basis) and distances and angles can be measured, for example using the dot (or scalar) product.
Vector spaces, in general, have none of this extra structure.
The dot product is an example of an inner product; an inner product is a means of measuring
distance and angles within a vector space. A vector space together with an inner product is
called an inner product space. Initially we will consider inner products only on real vector
spaces, but we will later discuss complex inner product spaces. Inner products appear in many
areas of mathematics and they have particular importance in Fourier series and in quantum
theory.

8.1 Bilinear forms

Definition 193 Let V be a vector space over F. A bilinear form B on V is a function

B : V × V → F, such that

(a) B(α1 v1 + α2 v2 , v3 ) = α1 (v1 , v3 ) + α2 (v2 , v3 ) for all v1 , v2 , v3 ∈ V and α1 , α2 ∈ F;

(b) B(v1 , α2 v2 + α3 v3 ) = α2 (v1 , v2 ) + α3 (v1 , v3 ) for all v1 , v2 , v3 ∈ V and α2 , α3 ∈ F.

(a) says that B is linear in the first variable (when we fix the second variable), and (b) says
the same for the second variable.

Example 194 For x = (x1 , . . . , xn ) ∈ Fn and y = (y1 , . . . , yn ) ∈ Fn , we define

B(x, y) = x1 y1 + · · · + xn yn .

This gives a bilinear form. In Rn this is the familiar dot product, or scalar product, often
written x · y.

Example 195 Take A ∈ Mn×n (F). Note for x, y ∈ Fn that B(x, y) = xAyT defines a bilinear
form on Fn .
Note that the usual scalar product is an example of this in the special case that A = In ,
because x · y = xyT . (Officially, xAyT is a 1 × 1 matrix, not an element of F, but it is
completely natural to identify 1 × 1 matrices with scalars).
And for x, y ∈ Fncol then B(x, y) = xT Ay defines a bilinear form on Fncol .

INNER PRODUCT SPACES 69

Definition 196 Let V be a vector space over F and let B be a bilinear form on V . Take
v1 , . . . , vn ∈ V . The Gram matrix of B with respect to v1 , . . . , vn is the n × n matrix
(B(vi , vj )) ∈ Mn×n (F).
Proposition 197 Let V be a finite-dimensional vector space over F and let v1 , . . . , vn be a
basis for V. Let B be a bilinear form on V and let A ∈ Mn×n (F) be the associated Gram matrix.
For X, Y ∈ V , let x = (x1 , . . . , xn ) ∈ Fn and y = (y1 , . . . , yn ) ∈ Fn be the unique coordinate
vectors such that
X = x1 v1 + · · · + xn vn and Y = y1 v1 + · · · + yn vn .
Then B(X, Y ) = xAyT .
Remark 198 Consequently the bilinear form in Example 195 essentially describes all bilinear
forms on V . Note that if A is the Gram matrix of a bilinear form, then any other Gram matrix
of B (that is, a Gram matrix with respect to a different basis) equals P T AP where P ∈ Mn×n (F)
is invertible.
Proof. We have
n n
!
X X
B(X, Y ) = B xi v i , yj vj
i=1 j=1
n n
!
X X
= xi B vi , yj vj [using linearity in the first entry]
i=1 j=1
Xn n
X
= xi yj B(vi , vj ) [using linearity in the second entry]
i=1 j=1
n
XX n
= xi aij yj
i=1 j=1

= xAy T .

Definition 199 We say that a bilinear form B : V × V → F is symmetric if

B(v1 , v2 ) = B(v2 , v1 ) for all v1 , v2 ∈ V.
Note that a bilinear form is symmetric if and only if any Gram matrix of the bilinear form is
symmetric.

8.2 Inner product spaces

Definition 200 Let V be a real vector space. We say that a bilinear form B : V × V → R is
positive definite if B(v, v) ⩾ 0 for all v ∈ V , with B(v, v) = 0 if and only if v = 0. N.B.
we are defining real inner product spaces here; the requirement that B(v, v) ⩾ 0 does not make
sense in a general field.

INNER PRODUCT SPACES 70

Definition 201 An inner product on a real vector space V is a positive definite, symmetric,
bilinear form on V . Inner products are usually denoted ⟨x, y⟩ rather than B(x, y).
We say that a real vector space is an inner product space if it is equipped with an inner
product. Unless otherwise specified, ⟨−, −⟩ will denote an inner product, rather than a general
bilinear form.

Example 202 The dot product on Rn is an inner product. We noted earlier that it is a bilinear
form, and it is clearly symmetric. If x = (x1 , . . . , xn ) ∈ Rn and x ̸= 0, then

x · x = x21 + · · · + x2n > 0,

so the dot product is also positive definite. The inner product space consisting of Rn equipped
with the dot product is known as n-dimensional Euclidean space. The dot product also turns
Rncol into an inner product space.

Example 203 Let V = Rn [x], the vector space of polynomials of degree ⩽ n. For f , g ∈ V ,
define Z b
⟨f, g⟩ = f (x)g(x) dx
a
where a < b. Then ⟨−, −⟩ is bilinear – as integration is linear – and symmetric – as the integrand
is symmetric in f and g.
If f ∈ V and f ̸= 0, then f (x) = 0 for only finitely many x in [a, b], and (f (x))2 > 0 at
other x, and we find that Z b
⟨f, f ⟩ = f (x)2 dx > 0.
a
So ⟨−, −⟩ is positive definite.
Hence ⟨−, −⟩ is an inner product on V . In fact, more generally, ⟨−, −⟩ defines an inner
product on the space C[a, b] of continuous real-valued functions on the interval [a, b] .

Importantly, inner products allow us to define length and angle, something which is not
possible with the structure of a vector space alone.

Definition 204 Let V be an inner product space. For v ∈ V , we define the norm (or mag-
nitude or length) of v to be p
∥v∥ := ⟨v, v⟩.
The distance between two vectors v, w ∈ V is defined to be

d(v, w) = ∥v − w∥.

Proposition 205 The norm ∥ − ∥ has the following properties; for v, w ∈ V and α ∈ R,
(a) ∥v∥ ⩾ 0 and ∥v∥ = 0 if and only if v = 0V .
(b) ∥αv∥ = |α| ∥v∥ .
(c) ∥v + w∥ ⩽ ∥v∥ + ∥w∥ . This is known as the triangle inequality.

Proof. (a) and (b) are straightforward. To prove (c) we will first prove the Cauchy-Schwarz
inequality.

INNER PRODUCT SPACES 71

Proposition 206 (Cauchy-Schwarz Inequality) For v, w in an inner product space V,
then
|⟨v, w⟩| ⩽ ∥v∥ ∥w∥ .
Equality holds if and only if v and w are linearly dependent.

Proof. If w = 0 then the result is immediate, so assume that w ̸= 0. For t ∈ R, note that

0 ⩽ ∥v + tw∥2
= ⟨v + tw, v + tw⟩
= ⟨v, v⟩ + 2t⟨v, w⟩ + t2 ⟨w, w⟩ [by linearity and symmetry]
2 2
= ∥v∥ + 2t⟨v, w⟩ + t2 ∥w∥ .

As ∥w∥ =
̸ 0, the last line is a quadratic in t which is always non-negative. So it either has
complex roots or a repeated real root, meaning its discriminant is non-positive. So

discriminant = 4⟨v, w⟩2 − 4 ∥w∥2 ∥v∥2 ⩽ 0

and the Cauchy-Schwarz inequality follows. For equality, the discriminant has to be zero which
means there is a repeated real root t = t0 . But then ∥v + t0 w∥ = 0 and hence v + t0 w = 0V
showing that v and w are linearly dependent. The converse is immediate.

Proof. Continuing the proof of Proposition 205 (c): we now see

∥v + w∥2 = ⟨v + w, v + w⟩
= ⟨v, v⟩ + 2⟨v, w⟩ + ⟨w, w⟩
⩽ ∥v∥2 + 2 |⟨v, w⟩| + ∥w∥2
⩽ ∥v∥2 + 2 ∥v∥ ∥w∥ + ∥w∥2 [by the Cauchy-Schwarz inequality]
= (∥v∥ + ∥w∥)2

and the triangle inequality follows.

Proposition 207 The distance function d(v, w) = ∥v − w∥ satisfies the following properties:
for u, v, w ∈ V we have
(a) d(v, w) ⩾ 0 and d(v, w) = 0 if and only if v = w.
(b) d(v, w) = d(w, v).
(c) d(u, w) ⩽ d(u, v) + d(v, w).
Here (a), (b), (c) show d has the properties of a metric.

Proof. These properties follow straightforwardly from the properties of the norm.

You may recall that in R2 or R3 we have

x · y = ∥x∥∥y∥ cos θ,

INNER PRODUCT SPACES 72

where θ is the angle between the vectors x and y. In general, we can use this idea to define
a notion of angle in an abstract inner product space V : we define the angle between nonzero
vectors x, y ∈ V to be  
−1 ⟨x, y⟩
cos .
∥x∥∥y∥
Note by the Cauchy-Schwarz inequality that this is well-defined as

⟨x, y⟩
⩽1
∥x∥∥y∥

for any nonzero vectors x, y in an inner product space.

Example 208 Let m, n be integers. Show that sin mx and cos nx are perpendicular in C[−π, π]
with the inner product from Example 203. Show also that cos mx perpendicular to cos nx when
m ̸= n and find ∥cos mx∥ .

Solution. Note that Z π

⟨sin mx, cos nx⟩ = sin mx cos nx dx = 0
−π

as the integrand is odd. For the second part, recall the trigonometric identity
1
cos mx cos nx = [cos(m + n)x + cos(m − n)x].
2
So if m ̸= n then
1 π
Z
⟨cos mx, cos nx⟩ = cos(m + n)x + cos(m − n)x dx
2 −π
 π
1 sin(m + n)x sin(m − n)x
= +
2 m+n m−n −π
= 0.

If m = n ̸= 0 then
Z π
2 1
∥cos mx∥ = ⟨cos mx, cos mx⟩ = (cos 2mx + 1) dx = π,
2 −π

and if m = n = 0 then Z π
2
∥1∥ = ⟨1, 1⟩ = dx = 2π.
−π

Remark 209 The above orthogonality relations are crucial in the study of Fourier series.
If we can represent a function on −π < x < π as a Fourier series
∞
1 X
f (x) = a0 + (ak cos kx + bk sin kx) ,
2 k=1

INNER PRODUCT SPACES 73

then, provided the integration and infinite sum can be interchanged, we would have
∞
Z π Z π !
1 X
f (x) sin lx dx = a0 sin lx + (ak cos kx sin lx + bk sin kx sin lx) dx
−π −π 2 k=1
Z π  X ∞ Z π  X ∞ Z π 
1
= a0 sin lx dx + ak cos kx sin lx dx + bk sin kx sin lx dx
−π 2 k=1 −π k=1 −π
∞
X ∞
X
=0+ ak × 0 + bk πδkl
k=1 k=1
= πbl .

Hence Z π
1
bl = f (x) sin lx dx for l ⩾ 1.
π −π

Similar calculations show that

Z π
1
al = f (x) cos lx dx for l ⩾ 0.
π −π

These are the Fourier coefficients of f (x). Validating convergency issues and interchanging the
integration and infinite sum are difficult matters of analysis.

8.3 Orthogonal Maps

Definition 210 A linear map α : V → V of an inner product space V is said to be orthogonal

if
⟨α(v), α(w)⟩ = ⟨v, w⟩
for all v, w ∈ V.

Proposition 211 Let α : Rncol → Rncol be a linear map and let A denote the matrix of α with
respect to the standard basis. Then α is orthogonal with respect to the dot product if and only
if A is an orthogonal matrix.

Proof. Suppose that α is orthogonal with respect to the dot product. Denote the standard
basis as e1 , . . . , en . Then

δij = ei · ej = α(ei ) · α(ej ) = (Aei )T (Aej ) = eTi AT Aej .

Now eTi AT Aej is the (i, j)th entry of AT A, and this equals δij which is the (i, j)th entry of In .
Hence AT A = In as this is true for all i, j.

ORTHOGONAL MAPS 74
Reversing the implications of the above argument takes us from AT A = In to α(ei )·α(ej ) = δij .
But then by linearity
! !
X X XX
α ui ei · α vj ej = ui vj α(ei ) · α(ej )
i j i j
XX
= ui vj δij
i j
X
= ui vi
i
! !
X X
= ui ei · vj ej
i j

and so α is orthogonal.

Proposition 212 An orthogonal map is an isometry of an inner product space.

Proof. Say α is orthogonal. Then, by linearity,

d(α(v), α(w))2 = ∥α(v − w)∥2 = ⟨α(v − w), α(v − w)⟩ = ⟨v − w, v − w⟩ = ∥v − w∥2 = d(v, w)2 .

Hence α is an isometry. In fact, it can be shown that any linear isometry of a finite-dimensional
vector space is orthogonal. (This is proven in the Geometry course for Rn .)

Definition 213 Let V be an inner product space. We say that {v1 , . . . , vk } ⊆ V is an or-
thonormal set if for all i, j we have
(
1 if i = j
⟨vi , vj ⟩ = δij =
0 if i ̸= j.

So the vectors are of unit length and are mutually perpendicular.

Lemma 214 In an inner product space V , an orthonormal set is linearly independent.

Proof. Say {v1 , . . . , vk } is orthonormal and α1 , . . . , αk ∈ R such that α1 v1 + · · · + αk vk = 0V .

Then for 1 ⩽ i ⩽ k we have

0 = ⟨0V , vi ⟩ = ⟨α1 v1 + · · · + αk vk , vi ⟩
= α1 ⟨v1 , vi ⟩ + · · · + αk ⟨vk , vi ⟩
= αi

so α1 = · · · = αk = 0.

Remark 215 Note then that n orthonormal vectors in an n-dimensional inner product space
is an orthonormal basis. It is the case that every finite-dimensional inner product space has an
orthonormal basis, but this result will be proved in Linear Algebra II.

ORTHOGONAL MAPS 75
 Recall that a matrix X ∈ Mn×n (R) is orthogonal if XX T = In = X T X. Equivalently, X
is orthogonal if X is invertible and X −1 = X T .

Proposition 216 Take X ∈ Mn×n (R). Consider Rn (or Rncol ) equipped with the usual inner
product ⟨x, y⟩ = x · y. The following are equivalent:
(a) XX T = In ;
(b) X T X = In ;
(c) the rows of X form an orthonormal basis of Rn ;
(d) the columns of X form an orthonormal basis of Rncol ;
(e) for all x, y ∈ Rncol , we have Xx · Xy = x · y.

Proof. (a) ⇔ (b): For any A, B ∈ Mn×n (R), we have AB = In if and only if BA = In .
(a) ⇔ (c): Say the rows of X are x1 , . . . , xn . Note that the (i, j)th entry of XX T is xi · xj .
But XX T = In if and only if the (i, j) entry of XX T is δij , i.e. if and only if the rows are
orthonormal. As there are n rows then the rows further form an orthonormal basis.
(b) ⇔ (d): Say the columns of X are y1 , . . . , yn . We see that the (i, j)th entry of X T X is
yi · yj . The remainder of the argument is as given above.
(b) ⇒ (e): Recall that we can identify x · y with xT y. Assume that X T X = In and take x,
y ∈ Rncol . Then

(Xx) · (Xy) = (Xx)T (Xy)

= (xT X T )(Xy)
= xT (X T X)y
= xT In y
= xT y
= x · y.

(e) ⇒ (d): Assume that Xx · Xy = x · y for all x, y ∈ Rncol . Let e1 , . . . , en be the standard
basis of Rncol . But then
δij = ei · ej = Xei · Xej .
And so Xe1 , . . . , Xen , which are the columns of X, form an orthonormal basis.

Remark 217 Condition (e) says that the map RX : Rn → Rn sending x to xX preserves the
inner product, and hence preserves length and angle. Such a map is called an isometry of the
Euclidean space Rn . So the previous proposition says that X is orthogonal if and only if the
map RX is an isometry.

8.4 Complex inner product spaces

Whilst the above theory of inner products applies very much to real vector spaces, rather
than to vector spaces over a general field, the theory can be adapted and extended to vector

COMPLEX INNER PRODUCT SPACES 76

spaces over C. However, the usual dot product on Cn isn’t an inner product: it is bilinear and
symmetric but we’d find in C2 that

∥(1, i)∥2 = 12 + i2 = 0

even though (1, i) ̸= (0, 0) . We can avoid this problem by defining the standard inner product
on Cn to be n
X
(z1 , . . . , zn ) · (w1 , . . . , wn ) = zi wi .
i=1

We then have that n n

X X
(z1 , . . . , zn ) · (z1 , . . . , zn ) = zi zi = |zi |2
i=1 i=1

which is non-negative and zero if and only if (z1 , . . . , zn ) = 0.

Note that this form is linear in the first variable, positive definite, but ”conjugate symmetric”
in the sense that
⟨v1 , v2 ⟩ = ⟨v2 , v1 ⟩.

Definition 218 Let V be a complex vector space. A function ⟨−, −⟩ : V × V → C is a

sesquilinear form if

(a) ⟨α1 v1 + α2 v2 , v3 ⟩ = α1 ⟨v1 , v3 ⟩ + α2 ⟨v2 , v3 ⟩ for all v1 , v2 , v3 ∈ V and α1 , α2 ∈ C; and∥∥

(b) ⟨v1 , v2 ⟩ = ⟨v2 , v1 ⟩ for all v1 , v2 ∈ V .

In particular, we have ⟨v, v⟩ ∈ R for all v ∈ V . We say that a sesquilinear form is positive
definite if ⟨v, v⟩ ⩾ 0 for all v ∈ V , with ⟨v, v⟩ = 0 if and only if v = 0.
A positive definite sesquilinear form is an inner product on a complex vector space, thus
a complex inner product space is a complex vector space equipped with a positive definite
sesquilinear form.

Remark 219 The prefix sesqui- relates to ”1 12 times”; for example a sesquicentennary is 150
years.

Remark 220 The equivalent of orthogonal maps for real inner product spaces are the unitary
maps. That is, a linear map U : V → V of a complex inner product space V is unitary if
⟨U x, U y⟩ = ⟨x, y⟩ for all x, y ∈ V. A unitary matrix is a square matrix such that U U ∗ = I =
T
U ∗ U where U ∗ = U .

Should you study quantum theory later, then you will see that the theory is generally set
within complex inner product spaces. The wave function ψ of a particle is complex-valued and
its norm-squared ∥ψ∥2 = ψψ is a probability density function.

You will explore inner product spaces further in Linear Algebra II and Part A Linear
Algebra.

COMPLEX INNER PRODUCT SPACES 77