Département du Premier Cycle

Filière SCAN
Première Année

LINEAR ALGEBRA

2020–2021 Romaric Pujol

 Linear Algebra

Romaric Pujol

E-mail address: [email protected]

Key words and phrases. Linear Algebra, Linear System, Vector Space, Linear Map, Matrix,
Determinant, Diagonalization

To the SCAN First students.

March 14, 2021

 Contents

Chapter 1. Linear Systems 1

1. Definition. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
2. Operations on Systems and Equivalent Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
3. Gaussian Elimination . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

Chapter 2. Vector Spaces 9

1. Commutative Fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2. Vector Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3. Linear Subspaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.1. Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.2. Intersection of Subspaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.3. Linear Span of a Subset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.4. Linear Combination of Vectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
4. Sum and Direct Sum of Subspaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
4.1. Sums . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
4.2. Independent Subspaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
4.3. Direct Sum . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
5. Bases and Dimension . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
5.1. Independent Family of Vectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
5.2. Generating Families of Vectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
5.3. Some Technical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
5.4. Basis of a Vector Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
5.5. Standard Bases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
5.6. Dimension of a Vector Space and Grassmann Formula . . . . . . . . . . . . . . . . . 22
5.7. Coordinates of a Vector in a Basis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
5.8. Inclusion–Equality Theorem and Other Results about Dimension . . . . . . 26
6. Rank of a Family of Vectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

Chapter 3. Linear Maps 35

1. Definitions, Vocabulary and Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
1.1. Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
1.2. Vocabulary and Notations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
1.3. Basic Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
1.4. Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
1.5. Kernel and Image. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
2. Rank of a Linear Map and the Rank–Nullity Theorem . . . . . . . . . . . . . . . . . . . . . . 41
3. Matrix of a Linear Map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
4. Powers and Polynomials of Endomorphisms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
5. Codimension of a Subspace. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
v
vi CONTENTS

6. Duality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
6.1. Dual of a Linear Map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
6.2. Dual Family and Dual Basis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
Chapter 4. Matrices 55
1. Computations with Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
1.1. Interpretation of a Matrix and Notations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
1.2. Sum of Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
1.3. Multiplication by a Scalar . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
1.4. Vector Space Structure on M mn (K) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
1.5. Product of Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
1.6. Identity Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
1.7. Invertible Matrix — Rank of a Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
1.8. Computing the Inverse of a Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
1.9. Transposition and Symmetric Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
2. Change of Basis Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
2.1. Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
2.2. Change of Bases of Linear Maps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
2.3. Similar Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
3. Link with Linear Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
Chapter 5. Determinants 73
1. Determinant of a Square Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
2. Determinant of an Endomorphism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
3. Geometric Interpretation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
4. Trace of a matrix — Trace of an Endomorphism . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
5. Proof of Theorem 5.3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
5.1. Permutations and Signatures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
5.2. Leibniz Formula for Determinants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
6. Cofactors, Comatrix and a Formula for the Inverse of a Matrix . . . . . . . . . . . . . . 83
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
Chapter 6. Diagonalization of Endomorphisms 87
1. Eigenvalues, Eigenvectors and Eigenspaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
2. Characteristic Polynomial . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
3. Diagonalization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
4. Hamilton–Cayley Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
Review Exercises 97
Bibliography 105
Index 107
 Linear Systems
1
1. Definition
Definition 1.1. A linear system with real or complex coefficients is a system of equations of
the form
a11 x1 + a12 x2 + · · · + a1 p x p = b1


a x + a x + ··· + a x = b

21 1 22 2 2p p 2
(S)
 · · ·
an1 x1 + an2 x2 + · · · + an p x p = bn .



• The ai j ’s are the coefficients of the system (real or complex numbers), the bi ’s are
the constant terms (real or complex numbers) and the xi ’s are the unknowns (or
the variables).
• The number n is the number of equations, and p is the number of unknowns.
We also say it is an n × p system (n equations and p unknowns).
• The linear system is said to be upper-triangular (or in row-echelon form) if
(1) all equations with all coefficients equal to zero (if any) are at the bottom of
the system,
(2) in each row, the first non-zero coefficient (from the left) is to the right of the
first non-zero coefficient in the row above it.
A lower-triangular linear system is defined analogously.
• The linear system is said to be diagonal if ∀i 6= j , ai j = 0.
• A linear system is said to be homogeneous if all the constant terms are 0.
• The homogeneous linear system associated with the system (S) is the following n × p
linear system:
a11 x1 + a12 x2 + · · · + a1 p x p = 0


a x + a x + ··· + a x = 0

21 1 22 2 2p p
(H)
 · · ·
an1 x1 + an2 x2 + · · · + an p x p = 0.



• The set of solutions of System (S) is the set of all n-tuples (x1 , . . . , x p ) of Rn or Cn
that satisfy simultaneously all n equations of System (S).
1
2 1. LINEAR SYSTEMS

For example, a 2 × 2 system can be seen as the intersection of two straight lines (except
when a11 = a12 = 0 or a21 = a22 = 0). In this case, you know that either the two straight-
lines intersect at one point only, or, if they are parallel they either never intersect or they
are the same.
The following two propositions are standard in the theory of linear systems:

Proposition 1.2. For a linear system over R or C, there are only three possible situations:
• The linear system may have no solutions.
• The linear system may have a unique solution.
• The linear system may have an infinite number of solutions.

Proof. See Exercise 1.1.

Proposition 1.3. In the case where a linear system over R or C has more unknowns than
equations, it cannot have a unique solution: It either has no solutions at all, or an infinite
number of solutions.

Proof. See Exercises 1.2

Remarks 1.4.
(1) A homogeneous linear system always has at least one solution: the null solution.
(2) Triangular linear systems are easy to solve (and diagonal systems are even easier).

2. Operations on Systems and Equivalent Systems

Definition 1.5. Two linear n × p systems are said to be equivalent if they have the same
solutions.

Consider an n × p linear system

a11 x1 + a12 x2 + · · · + a1 p x p = b1


 a x + a x + ··· + a x = b

21 1 22 2 2p p 2
(S)
 · · ·
an1 x1 + an2 x2 + · · · + an p x p = bn



There are three types of operations on (S) we can consider. Bear in mind that we want these
operations to yield an equivalent system:
• We can exchange two rows: the operation that exchanges Row i and Row j is
denoted by Ri ↔ R j . This operation is known as row switching.
• We can multiply a row by a non-zero number λ: this operation is denoted by
Ri ← λRi . This operation is known as row multiplication.
• We can add a multiple of a row to another row: the operation that adds to Row i
the Row j multiplied by λ (with i 6= j ) is denoted by Ri ← Ri + λR j . This
operation is known as row addition.
These three operations are known as elementary row operations.

Remarks 1.6.
(1) In row addition, the factor λ can be 0, but in that case it’s a useless operation.
 3. GAUSSIAN ELIMINATION 3

(2) Since we want to obtain an equivalent system when performing elementary row
operations, it is important that in row multiplication the factor λ is non-zero, and
it is important that in row addition we consider two distinct rows.
(3) Obviously, after performing any elementary row operation to an n × p linear
system, we obtain another n × p linear system.
(4) These row operations apply to both, the left-hand side and the right-hand side of
the system.

Example 1.7. We illustrate each elementary row operation on the following system:

 x + y − z =4
(S) 3x + y − 2z = 1
 −x + 2y =2
• The row switching operation R2 ↔ R3 performed on System (S) yields:

 x + y − z =4
−x + 2y =2
 3x + y − 2z = 1.

• The row multiplication operation R2 ← 3R2 performed on System (S) yields:


 x + y − z =4
9x + 3y − 6z = 3
 −x + 2y = 2.
• The row addition operation R2 ← R2 − 4R1 performed on System (S) yields:

 x + y − z = 4
−x − 3y + 2z = −15
 −x + 2y = 2.

The following result should be obvious:

Proposition 1.8. A linear system is equivalent to the system obtained by performing any
number of elementary row operations.

You can be convinced that a solution of a linear system will also be a solution of the linear
system obtained after performing an elementary row operation. The fact that we don’t
lose any solutions or create any new solutions is clear from the fact that each elementary
operation is an invertible transformation. Indeed:
• The row switching Ri ↔ R j is canceled when performing the very same row
switching Ri ↔ R j .
• If λ 6= 0, the row multiplication Ri ← λRi is canceled when performing the row
1
multiplication Ri ← Ri (that’s why it’s important that λ 6= 0).
λ
• If i 6= j and if λ is a number, the row addition Ri ← Ri + λR j is canceled when
performing the row addition Ri ← Ri −λR j (that’s why it’s important that i 6= j ).

3. Gaussian Elimination
The goal is to apply successively elementary row operations to obtain an upper-triangular
system, because upper-triangular systems are easy to solve.
4 1. LINEAR SYSTEMS

We illustrate the method on the following example:


 2x + y + 2z = 1
(S) 3x − 3y + z = 3
 x − 2y = 2.
The first step is to choose a pivot, i.e., a non-zero coefficient in the first column of the
system. Two things about this pivot (apart from the fact that it must be non-zero): use
a row switching operation to put it on the first row, and choose this pivot as simple as
possible, since the row addition operation will involve dividing by this pivot. In our case,
we have three possibilities: a 3, a 2 and a 1. We might as well choose the 1 on the last row.
Hence, the first step is to use the row switching operation R1 ↔ R3 :

 x − 2y =2
3x − 3y + z = 3
 2x + y + 2z = 1.

Then, with the pivot 1 on the first row, we perform two row addition operations to cancel
the other terms on the first column: R2 ← R2 − 3R1 :

 x − 2y = 2
3y + z = −3
 2x + y + 2z = 1.

Then, with the same pivot we cancel the first coefficient of the last row by performing the
row addition operation R3 ← R3 − 2R1 :

 x − 2y = 2
3y + z = −3
 5y + 2z = −3.
Now we only consider the remaining two last lines, and perform the same operations: we
have two choices for the pivot, either 3 or 5. None is simpler than the other, so we might
5
as well go for the 3. We hence perform the row addition operation R3 ← R3 − R2 :
3
= 2

 x − 2y
3y + z = −3
1
3 z = 2.


Now the linear system is upper-triangular, and we solve it using back substitution, that is
from the bottom row, upwards (this method can also be seen as successive row addition
and row multiplication operations to obtain a diagonal system):
  
 x − 2y = 2  x − 2y = 2  x − 2y = 2
3y + z = −3 ⇐⇒ 3y = −3 − z = −9 ⇐⇒ y = −3
 z= 6  z =6  z= 6

 x = 2 + 2y = −4
⇐⇒ y = −3
 z = 6.

Hence, the unique solution of System (S) is x = −4, y = −3 and z = 6.

We now investigate two other examples: one in which there are no solutions and another
one that possesses an infinite number of solutions:
 3. GAUSSIAN ELIMINATION 5

Examples 1.9.
(1) Consider the following 4 × 3 linear system:
x+ y + z = 1

x − 2y + 3z = −2


(S)
 x + 3y − z = 2

x + 2y + z = 1.
For the pivot, we choose the 1 in the first position of the first row, and we proceed
with the first step of the Gaussian elimination by performing successively the
elementary row operations R2 ← R2 − R1 , R3 ← R3 − R1 and R4 ← R4 − R1 , to
obtain the following equivalent system:
x+ y + z = 1

− 3y + 2z = −3


(S) ⇐⇒
 2y − 2z = 1
= 0.

y
We choose the 1 in the last row as a pivot, hence we first perform the row switching
R2 ↔ R4 :
x+ y + z = 1

= 0

y

(S) ⇐⇒
 − 3y + 2z = −3
2y − 2z = 1,


and now we successively perform the row additions R3 ← R3 + 3R2 and R4 ←

R4 − 2R2 :
x+y+ z = 1

= 0

y

(S) ⇐⇒
 2z = −3
− 2z = 1.


At this step, we see that the last two rows are incompatible, but let’s carry on the
Gaussian elimination to obtain a triangular system, by the row addition R4 ←
R4 + R3 :
x+y+ z = 1

= 0

y

(S) ⇐⇒
 2z = −3
0 = −2.


We observe that the last row is impossible, hence System (S) has no solutions.
(2) Consider the following 2 × 3 linear system:
x+ y + z = 1
§
(S)
x − 2y + 3z = −2.
We perform the row addition R2 ← R2 − R1 and we obtain:
x+ y + z = 1
§
(S) ⇐⇒
− 3y + 2z = −3.
To see that this linear system possesses an infinite number of solutions, we can
consider z as a parameter by writing all the z variables on the right-hand side of
the equalities (hence considering the variable z as a parameter):
x + y =1− z
§
(S) ⇐⇒
− 3y = −3 − 2z.
6 1. LINEAR SYSTEMS

We now consider this system as a 2 × 2 system, with z a parameter, and solve it

using the back substitution procedure:

5
x +y =1− z x = 1 − z − y = −3 z
 
(S) ⇐⇒ 2 ⇐⇒ 2
y =1+ 3z y = 1 + 3 z.

We see that, given any value of z, we can find values for x and y. Geometrically,
System (S) represents the intersection of two planes of R3 , and the parametrized
solution
 5
x = − z

3
y = 1 + 2 z

3

corresponds to a straight line of R3 .

Remark 1.10. It can sometimes be convenient to exchange columns of a system to consider

a simpler pivot. This column switching operation obviously yields an equivalent linear
system.

Definition 1.11. The rank of a linear system is the number of equations (that still contain
variables) after the triangularization procedure of the Gaussian elimination.

Example 1.12. The rank of System (S) of Example 1.9(1) is 3, and the rank of System (S)
of Example 1.9(2) is 2.

Remark 1.13. The solutions of an n × p linear system that possesses solutions can be
expressed in terms p − k parameters, where k is the rank of the system. In particular, if the
rank is equal to p (the number of unknowns), then the system has at most one solution.
From this remark, we understand why the rank of a linear system is more important that
its number of equations.

Exercises
Exercise 1.1. Prove that if an n × p linear system over R or C admits two distinct solutions,
then it admits an infinite number of solutions.

Exercise 1.2. Consider an n × p linear system (S) over R or C with p > n.

1. Prove that if the system (S) has a solution, it has an infinite number of solutions.
2. Deduce that the associated homogeneous system (H) possesses an infinite number of
solutions.
3. Prove that all the solutions of (S) can be deduced from a particular solution of (S) and
the solutions of (H).

Exercise 1.3. Consider an n ×n linear system (S). Prove that if the associated homogeneous
system has a unique solution, then (S) also has a unique solution.
 EXERCISES 7

Exercise 1.4. Let m, a, b , c ∈ R. Solve the following linear systems (with unknowns
x, y, z, t , u, . . .) using Gaussian elimination, and specify their rank:
 
 x − 3y + z = 3  x − 3y + z = 1
(1) y − 2z = 1 (2) 2x + y − z = −1
 z =2  x + 11y − 5z = 5

x + y + z =3
 
 x − 3y + z = 0
x + 2y − z = 2


(3) 2x + y − z = 0 (4)
 x + 11y − 5z = 0 x
 + 3z = 4
2x + y + 4z = 7
x + y − 5z = −7
 
 2x − 3y + z + 6t − 6u = 3
2x − y − z = 4


(5) 2x − 2y + 2z + 4t − 6u = 4 (6)
 −2x + 4y + z − 8t + 3u = 0  3x − 2y + 4z = 11

3x + 4y − 2z = 11
 
x + y =m  x + y + z = m +1
(7) y +z= 2 (8) m x + y + (m − 1)z = m
 x + 2y + z = 3  x + my + z =1
mx + y + z = 1
 
 3x − 5y + 2z + 4t = a
x + my + z = m


(9) 7x − 4y + z + 3t = b (10)
 5x + 7y − 4z − 6t = c  x + y + mz = 1

x + y + z =m
x + y + z =0
 
 2i x + (1 + i)y + z = 2
(11) −x +z=i (12) x + j y + j 2 z = 0
 2x + (1 − i)y + z = 6  x + j 2y + j z = 0

In the last system, j = e2iπ/3 .

Exercise 1.5 (From Term-Test, May 2010). We consider the following linear system:

 x + y + z =a
(S) 2x − 3y + 4z = b
 3x − 7y + 7z = c

with unknowns x, y, z and constant terms a, b , c.

1. Use Gauss’ method to obtain a triangular system that is equivalent to System (S).
2. Deduce a necessary and sufficient condition on the data a, b and c for System (S) to
admit solutions.
3. In this case, find the solutions of System (S).
4. We consider the following vectors of R3 : u = (1, 2, 3), v = (1, −3, −7) and w = (1, 4, 7).
Is the vector p = (6, 13, 20) a linear combination of the vectors u, v and w?

Exercise 1.6. Prove that the Gaussian elimination with back substitution applied on an
n3 n n 3 n 2 5n
n ×n system requires +n 2 − multiplications or divisions and + − additions
3 3 3 2 6
or subtractions. As n grows, the number of total operations required is of the order of
2n 3
. As of October 2012, the fastest supercomputer is the Titan, an American computer
3
manufactured by Cray at Oak Ridge National Laboratory, that is capable of an Rmax of
17.59 petaFLOPS, that is, it can perform 17.59 × 1015 floating point operations per second.
8 1. LINEAR SYSTEMS

It is not rare in engineering to have systems of size 107 × 107 , for example in weather
forecasting. Evaluate the time needed to the Titan computer to solve such a system, using
Gaussian elimination. Would you say it is a good strategy for weather forecasting?

Exercise 1.7. We can mix, under controlled conditions, toluene C7 H8 and nitric acid
HNO3 to produce trinitrotoluene1 C7 H5 O6 N3 along with the byproduct water. In what
proportion should we mix those components?

1TNT
 2
Vector Spaces

1. Commutative Fields
Definition 2.1. A commutative field is a set K together with two operations:
+ : K × K −→ K · : K × K −→ K
(λ, µ) 7−→ λ + µ, (λ, µ) 7−→ λ · µ,
(called addition and multiplication) that satisfy the following conditions:
• The addition is associative, that is, ∀(a, b , c) ∈ K3 , a + (b + c) = (a + b ) + c.
• There is an element 0 ∈ K such that ∀a ∈ K, 0 + a = a + 0 = a. It can be shown
that the element 0 is unique.
• ∀a ∈ K there is an element a 0 ∈ K such that a + a 0 = 0. It can be shown that the
element a 0 is unique; it is denoted by −a.
• The addition is commutative, that is, ∀(a, b ) ∈ K2 , a + b = b + a.
• The multiplication is associative, that is, ∀(a, b , c) ∈ K3 , a · (b · c) = (a · b ) · c.
• There is an element 1 ∈ K such that ∀a ∈ K, 1 · a = a · 1 = a. It can be shown that
1 is unique.
• For each a ∈ K, a 6= 0, there exists an element a 0 ∈ K such that a · a 0 = 1. It can
be shown that a 0 is unique; it is denoted by a −1 .
• The multiplication is commutative, that is, ∀(a, b ) ∈ K2 , a · b = b · a.
• The addition and the multiplication are compatible in the following sense:

∀(a, b , c) ∈ K3 , (a + b ) · c = a · c + b · c.

Remark 2.2. When writing the multiplication of two elements, we usually drop the sym-
bol ·.

Examples 2.3.
(1) Q, R, C are fields when equipped with the usual addition and multiplication.
(2) N and Z are not fields when equipped with the usual addition and multiplication.
(3) Let K be a field. The set of polynomials with coefficients in K, denoted by K[X ]
is not a field, but the set of rational functions, denoted by K(X ) is a field.
9
10 2. VECTOR SPACES

(4) There are finite fields, for example F2 , or more generally, if p is a prime number,
F p is a field.
p p
(5) Another example of field: Q[ 2] = r + r 0 2 (r, r 0 ) ∈ Q2 .


Remark 2.4. In the sequel, we’ll only consider the commutative fields R and C, even
though most results will still be valid for arbitrary commutative fields.

2. Vector Space
Definition 2.5. Let K be a commutative field. A vector space over K is a non-empty set E
together with two operations:

+ : E × E −→ E · : K × E −→ E
(u, v) 7−→ u + v, (λ, v) 7−→ λ · v,

(the addition and the scalar multiplication) that satisfy the following properties:
• The addition is commutative, that is, ∀u, v ∈ E, u + v = v + u.
• The addition is associative, that is, ∀u, v, w ∈ E, (u + v) + w = u + (v + w).
• There is an element 0E ∈ E, called the null vector of E such that ∀u ∈ E, 0E + u =
u + 0E = u.
• For all u ∈ E, there is an element u 0 ∈ E called the opposite of u such that u + u 0 =
u 0 + u = 0E . Such an element is denoted by −u.
• For all (λ, µ) ∈ K and all u ∈ E, λ · (µ · u) = (λµ) · u and (λ + µ) · u = λ · u + µ · u.
• For all λ ∈ K and (u, v) ∈ E 2 , λ · (u + v) = λu + λv.
• For all u ∈ E, 1 · u = u.
The elements of a vector space are called vectors. The elements of K are called scalars.

Remarks 2.6.
(1) When writing the scalar multiplication, we usually drop the symbol ·.
(2) If u, v ∈ E, we usually write v − u instead of v + (−u).
(3) When K = R we usually say real vector space and when K = C we say complex
vector space.
From this definition we prove the following result:

Proposition 2.7. Let E be a vector space over a commutative field K. Then:

• The null vector 0E of E is unique.
• For all u ∈ E, the opposite of u is unique.
• For all λ ∈ K and u ∈ E,
 
λu = 0E ⇐⇒ λ = 0 or u = 0E .

• For all λ ∈ K and u ∈ E, λ(−u) = (−λ)u = −(λu), and in particular, −u =

(−1) · u.

Proof. We show that the four statements are true:

• Let 0E and 00E be two null vectors of E, and let’s show that 0E = 00E : since 0E is a
null vector of E, we must have 0E + 00E = 00E and since 00E is a null vector of E we
must also have 0E + 00E = 0E , hence 0E + 00E = 0E = 00E .
 2. VECTOR SPACE 11

• Let u ∈ E and u 0 and u 00 be two opposites of u in E, and let’s show that u 0 = u 00 :

since the addition in E is associative and since u 00 + u = 0E = u + u 0 , we have:
u 0 = 0E + u 0 = (u 00 + u) + u 0 = u 00 + (u + u 0 ) = u 00 + 0E = u 00 ,
hence u 0 = u 00 .
• We first show that for all u ∈ E, 0u = 0E : indeed, 0u = (0 + 0)u = 0u + 0u,
hence 0E = 0u. Next, we show that for all λ ∈ E, λ0E = 0E : indeed,
λ0E = λ(0E + 0E ) = λ0E + λ0E ,
hence 0E = λ0E . Now, let λ ∈ K and u ∈ E such that λu = 0E and let’s show that
either λ = 0 or u = 0E , by contradiction: if λ 6= 0 and u 6= 0E , then
1 1 1
 ‹
0E = (0E ) = (λu) = λ u = 1u = u 6= 0E ,
λ λ λ
which is impossible.
• Let λ ∈ K and u ∈ E. Clearly,
0E = 0u = (λ − λ)u = λu + (−λ)u,
hence (−λ)u = −(λu). Similarly,
0E = λ0E = λ(u − u) = λu + λ(−u),
hence λ(−u) = −(λu).

The following examples are fundamental. All the vector spaces we are going to study will
be obtained from them. Let K be a commutative field:

Examples 2.8.
(1) The one element set {0} is a vector space over K (when endowed with the only
obvious operations). It is a boring vector space, and we call it the null vector space.
(2) Let n ∈ N∗ . We set:
Kn = (x1 , . . . , xn ) ∀1 ≤ i ≤ n, xi ∈ K .


The set Kn is a vector space over K when endowed with the following operations:
• the scalar multiplication is defined thus: if (x1 , . . . , xn ) ∈ Kn and λ ∈ K,
λ · (x1 , . . . , xn ) = (λx1 , . . . , λxn ),
• the addition of two vectors of Kn is defined thus: if (x1 , . . . , xn ) ∈ Kn and
(y1 , . . . , yn ) ∈ Kn , their sum is defined as:
(x1 , . . . , xn ) + (y1 , . . . , yn ) = (x1 + y1 , . . . , xn + yn ).
In particular, the field K = K1 is itself a vector space over K.
(3) The sets K[X ] and K(X ) are vector spaces over K, when equipped with the usual
sum and scalar multiplication.
(4) Let A be a set. The set of functions from A to K is a vector space over K, denoted
by KA, when endowed with the following operations:
• the scalar multiplication is defined thus: if f ∈ KA, that is, f is a function
from A to K, and if λ ∈ K, we define λ f as the function from A to K defined
by:
∀a ∈ A, (λ f )(a) = λ · f (a),
12 2. VECTOR SPACES

• the addition is defined thus: if f , g ∈ KA, that is, f and g are functions from A
to K, the addition of f and g , denoted by f + g is the function from A to K
defined by:
∀a ∈ A, ( f + g )(a) = f (a) + g (a).
(5) More generally, if E is a vector space over K and if A is any set, the set of functions
from A to E, denoted by E A is a vector space over K when endowed with the
following operations:
• the scalar multiplication is defined thus: if f ∈ E A, that is, f is a function
from A to E, and if λ ∈ K, we define λ f as the function from A to E defined
by:
∀a ∈ A, (λ f )(a) = λ · f (a),
• the addition is defined thus: if f , g ∈ E A, that is, f and g are functions from
A to E, the addition of f and g , denoted by f + g is the function from A
to E defined by:
∀a ∈ A, ( f + g )(a) = f (a) + g (a).
(6) Let a ∈ Z. The set KZ≥a consisting of all sequences (starting from index a) with
values in K forms a vector space over K, when endowed with the natural opera-
tions. This fact is clear since a sequence indexed by Z≥a with values in K is just
a function from Z≥a to K. More generally, if E is a vector space over K, then
the set E Z≥a that consists of all sequences (starting from index a) with values in E
forms a vector space over K, when endowed with the natural operations.
(7) If E1 , . . . En are vectors spaces over K, then the set
E1 × · · · × En = (u1 , . . . , un ) ∀1 ≤ i ≤ n, ui ∈ Ei


is a vector space over K when endowed with the following operations:

• the scalar multiplication is defined thus: if (u1 , . . . , un ) ∈ E1 × · · · × En and if
λ ∈ K we set
λ(u1 , . . . , un ) = (λu1 , . . . , λun ),
• the addition is defined thus: if (u1 , . . . , un ) ∈ E1 × · · · × En and (v1 , . . . , vn ) ∈
E1 × · · · × En , their sum is defined by:
(u1 , . . . , un ) + (v1 . . . , vn ) = (u1 + v1 , . . . , un + vn ).
This vector space is also denoted by
Y n
Ei
i=1
and is called the product space of E1 , . . . , En .
In the next example, we show that complex vector spaces can also be seen as real vector
spaces:

Examples 2.9.
(1) The set C is a vector space over R when endowed with the following operations:
• the addition of two vectors of C is the usual sum of complex numbers,
• the scalar multiplication is the natural one: if z ∈ C and if λ ∈ R, λz is just
the usual product of the real number λ with the complex number z.
(2) More generally, if E is a vector space over C, we can also consider it as a vector
space over R with the natural operations.
 3. LINEAR SUBSPACES 13

3. Linear Subspaces
3.1. Definition
Definition 2.10. Let E be a vector space over K. A non-empty subset F ⊂ E is called a
(linear) subspace of E if it satisfies the following property:

∀λ, µ ∈ K, ∀u, v ∈ F , λu + µv ∈ F .

Remarks 2.11.
/ F , then F is not a linear
(1) If F is a subspace of E then 0E ∈ F . In particular, if 0E ∈
subspace of E.
(2) If E is a vector space, then {0E } and E itself are subspaces of E.
The following result will be very useful in practice:

Theorem 2.12. Let E be a vector space over K and let F be a subspace of E. Then F is itself a
vector space over K.

We will use Theorem 2.12 a lot to prove that a set is a vector space: instead of checking that
some set F satisfies all the properties of Definition 2.5, we’ll prove that F is a subspace of
some well-known vector space (e.g., the vectors spaces of Example 2.8), by just checking
that F satisfies the property of Definition 2.10.
3.2. Intersection of Subspaces
Definition 2.13. Let (Ai )i∈I be a family of subsets of some set X . The intersection of the
family (Ai )i∈I is another subset of X defined by:
\
Ai = x ∈ X

∀i ∈ I , x ∈ Ai .
i∈I

If A and B are two subsets of X , we usually write A ∩ B for their intersection.

Proposition 2.14. Let E be a vector space over K and (Fi )i∈I be a family of subspaces of E.
Then intersection of the family (Fi )i∈I is a subspace of E. In particular, the intersection F ∩ G
of two subspaces F and G of E is a subspace of E.

Proof. Let E be a vector space over K and Let (Fi )i∈I be a family of subspaces E. We denote
by H the intersection of the family (Fi )i∈I , i.e.,
\
H= Fi ,
i∈I

and we prove that H is a subspace of E: clearly, H 6= ;, since 0E ∈ H , because for all i ∈ I ,

0E ∈ Fi . Now, let u and v be two elements of H and let λ and µ be two scalars. We need
to check that λu + µv ∈ H . Clearly, since for all i ∈ I , Fi is a subspace of E, we must
have λu + µv ∈ Fi , hence, since this is true for all i ∈ I , we conclude that λu+µv ∈ H .

Remark 2.15. In general, the union of subspaces is not a subspace. Actually, can you find
a condition for the union of two subspaces of E to be a subspace of E?
14 2. VECTOR SPACES

3.3. Linear Span of a Subset

Definition 2.16. Let E be a vector space over K and A a non-empty subset of E. The
intersection of all subspaces of E that contain A is called the linear span of A and is denoted
by Span A.

Proposition 2.17. Let E be a vector space over K and A a non-empty subset of E. Then Span A
is a subspace of E. Moreover, it is the smallest (for the inclusion) subspace of E that contains A.

Proof. We denote by F the family of all linear subspaces of E that contain A:

F = F F is a subspace of E and A ⊂ F .


Clearly, F =
6 ; (since E ∈ F ) and by definition,
\
Span A = F.
F ∈F
Hence, by Proposition 2.14, Span A is a subspace of E. Moreover, if G is another subspace
of E that contains A, we clearly have G ∈ F , and hence Span A ⊂ G.

Remark 2.18. Let E be a vector space over K. Clearly, Span{0E } = {0E } and Span E = E.

3.4. Linear Combination of Vectors

Definition 2.19. Let E be a vector space over K and let (u1 , . . . , un ) be a family of n vectors
of E. We call linear combination of the family (u1 , . . . , un ) any vector of E that can be written
as
λ1 u1 + · · · + λn un
for some scalars λ1 , . . . , λn in K.
More generally, if (ui )i∈I is a family of vectors of E, we call linear combination of the family
(ui )i∈I any vector of E that can be written as
X
λi ui
i∈I
where ∀i ∈ I , λi ∈ K and for all i ∈ I , except a finite number, λi = 0.

Proposition 2.20. Let A be a non-empty subset of the vector space E. The set of all linear
combinations of finite families of vectors of A is a subspace of E. In fact, it is the linear span
of A.

Proof. We denote be F the set of all linear combinations of finite families of vectors of A,
i.e.,
F = λ1 u1 + · · · + λn un ; n ≥ 1, ∀1 ≤ i ≤ n, λi ∈ K and ui ∈ A .


We first show that F is a subspace of E: clearly F is non-empty, since it contains, for

example, all the elements of A. Now, let u = λ1 u1 + · · · + λ m u m and v = µ1 v1 + · · · + µn vn
be two elements of F . We must show that if a and b are two scalars, then a u + b v is still
in F . Clearly,
a u + b v = aλ1 u1 + · · · + aλ m u m + b µ1 v1 + · · · + µn vn ,
and we see that it is a linear combination of elements of A, hence a u + b v ∈ F , and we
conclude that F is a subspace of E. We now show that F = Span A: let G be a subspace of E
that contains A. Clearly, G also contains all the linear combinations of finite families of A,
 4. SUM AND DIRECT SUM OF SUBSPACES 15

and hence F ⊂ G. We thus conclude that F is the smallest (for the inclusion) subspace of
E that contains A, that is, F = Span A.

4. Sum and Direct Sum of Subspaces

4.1. Sums
Definition 2.21. Let E be a vector space over K. Let F and G be two subspaces of E. The
sum of F and G is the set
F + G = u + v u ∈ F and v ∈ G .


For multiple spaces: let F1 , . . . , Fn be subspaces of E. The sum of the Fi ’s is the set of all
linear combinations of elements of the spaces F1 , . . . , Fn :
n
X
F1 + · · · + F n = Fi = u1 + · · · + un ; ∀1 ≤ i ≤ n, ui ∈ Fi .

i=1

More generally, if (Fi )i∈I is a family of subspaces of E, the sum of the Fi ’s is the set of all
finite linear combinations of elements of the family (Fi )i ∈I :
X
Fi = ui1 + · · · + uik ; k ≥ 1, i1 , . . . , ik ∈ I , ∀1 ≤ j ≤ k, ui j ∈ Fi j .

i∈I

Remark 2.22. It will often be convenient to think of the sum of the family (Fi )i∈I as the
set of all sums of the form:
X
ui
i∈I

with ∀i ∈ I , ui ∈ Fi , with all the ui ’s being equal to 0E except a finite number of them.

Proposition 2.23. Let E be a vector space over K.

• The sum of subspaces of E is a subspace of E.
• If A and B are two non-empty subsets of E, then
Span(A ∪ B) = Span A + Span B.
• More generally, if (Ai )i∈I is a family of non-empty subsets of E, then
‚ Œ
[ X
Span Ai = Span Ai .
i∈I i ∈I

Proof.
• Let (Fi )i ∈I be a family of subspaces of E and denote by G the sum of the fam-
ily (Fi )i∈I , i.e.,
X
G= Fi .
i ∈I

We prove that G is a subspace of E: clearly, G is non-empty since it contains all

the spaces Fi , i ∈ I . Let u and v be two elements of G, say,
X X
u= ui , v= vi ,
i∈I i∈I
16 2. VECTOR SPACES

with ∀i ∈ I , ui ∈ Fi and vi ∈ Fi , and all the ui ’s and vi ’s are zero, except a finite
number of them. Let λ and µ be two scalars. Then,
X X X
λu + µv = λ ui + µ vi = (λui + µvi ).
i∈I i ∈I i∈I

Since for all i ∈ I the set Fi is a subspace of i, the vector λui +µvi belongs to Fi , and
we have, for all i ∈ I except a finite number, (λui +µvi ) = 0E , hence λu + µv ∈ G,
and we hence conclude that G is a subspace of E.
• Clearly, Span A+Span B is a subspace of E that contains A and B, hence it contains
A∪ B. Thus, by minimality of Span(A∪ B), Span(A∪ B) ⊂ Span A+ Span B. Also
by minimality, we conclude that Span A ⊂ Span(A∪B) and Span B ⊂ Span(A∪B),
hence, by the previous result, Span A + Span B ⊂ Span(A ∪ B).
• The proof is the similar to that of the previous statement and is left as an exercise
to the reader.
4.2. Independent Subspaces
Definition 2.24. Let E be a vector space over K.
• Let F and G be two subspaces of E. We say that F and G are independent subspaces
of E if:
 
∀u ∈ F , ∀v ∈ G, u + v = 0E =⇒ u = v = 0E .

• Let F1 , . . . , Fn be n subspaces of E. We say that the family of subspaces (F1 , . . . , Fn )

is an independent family of subspaces of E if:
 
∀(u1 , . . . , uN ) ∈ F1 × · · · × FN , u1 + · · · + un = 0E =⇒ u1 = · · · = un = 0E .

• More generally, let (Fi )i∈I be a family of subspaces of E. We say that the fam-
ily (Fi )i∈I is an independent family of subspaces of E if given any distinct indices
i1 , . . . , in ∈ I , the following proposition holds true:
 
∀(ui1 , . . . , uin ) ∈ Fi1 × · · · × Fin , ui1 + · · · + uik = 0E =⇒ ui1 = · · · = uin = 0E .

Remark 2.25. We can express the last statement of the previous definition thus: let (Fi )i∈I
be a family of subspaces of E. We say that the family (Fi )i∈I is an independent family of
subspaces of E if for all family (ui )i∈I of vectors of E such that ∀i ∈ I , ui ∈ Fi and all the ui ’s
are zero except a finite number,
X
ui = 0E =⇒ ∀i ∈ I , ui = 0E .
i∈I

Proposition 2.26. Let (Fi )i∈I be a family of independent subspaces of E. Any vector
X
u∈ Fi
i ∈I

has a unique decomposition of the form

X
u= ui ,
i∈I

with ∀i ∈ I , ui ∈ Fi and all the ui ’s are zero except a finite number.

 4. SUM AND DIRECT SUM OF SUBSPACES 17

Proof. It is clear, by the very definition of the sum of subspaces, that any vector
X
u∈ Fi
i∈I

possesses such a decomposition. What remains to be shown is that the decomposition is

unique. We consider two such decompositions and we show that they must be the same:
write u as X X
u= ui = vi
i∈I i∈I
with ∀i ∈ I , ui , vi ∈ Fi and for all i ∈ I , except a finite number, ui = vi = 0E . We prove
that for all i ∈ I , ui = vi : indeed,
X X X
0E = u − u = ui − vi = (ui − vi ).
i∈I i∈I i∈I

Now, for all i ∈ I , (ui − vi ) ∈ Fi , and since the family (Fi )i ∈I is an independent family of
subspaces of E we must have, for all i ∈ I , (ui − vi ) = 0E , that is ui = vi .

Proposition 2.27. Let (Fi )i∈I be a family of subspaces of E. The family (Fi )i∈I is an indepen-
dent family of subspaces of E if and only if:
X
∀k ∈ I , Fk ∩ Fi = {0E }.
i∈I
i6=k

In particular, two subspaces F and G of E are independent subspaces if and only if F ∩G = {0E }.

Proof. Assume that the family (Fi )i∈I is an independent family of subspaces of E, and
let k ∈ I and X
u ∈ Fk ∩ Fi .
i∈I
i 6=k
This means that u ∈ Fk and that u can be written as
X
u= ui
i∈I

with ∀i ∈ I , ui ∈ Fi , and for all i ∈ I , expect a finite number, ui = 0E , and uk = 0E .

Since the family (Fi )i∈I and by the uniqueness of the decomposition of such an element
(Proposition 2.26), we must have u = 0E . This shows that
X
Fk ∩ Fi = {0E }.
i∈I
i6=k

Conversely, if X
∀k ∈ I , Fk ∩ Fi = {0E },
i∈I
i6=k
we must show that the family (Fi )i∈I is an independent family of subspaces
X of E: for all i ∈ I ,
let ui ∈ Fi such that for all i ∈ I , except a finite number, ui = 0E and ui = 0E . We must
i∈I
show that for all i ∈ I , ui = 0E : let k ∈ I . Clearly,
X
ui = −uk ,
i∈I
i6=k
18 2. VECTOR SPACES

hence X
−uk ∈ Fk ∩ Fi
i∈I
i6=k

and hence uk = 0E , by the assumption made on the family (Fi )i∈I .

4.3. Direct Sum
Definition 2.28. Let E be a vector space over K and (Fi )i∈I a family of subspacesX of E. If
the family (Fi )i ∈I is an independent family of subspaces of E we say that the sum Fi is
i∈I
direct, and we write: M X
Fi = Fi .
i∈I i∈I
For a finite number of independent subspaces F1 , . . . , Fn we write F1 ⊕ · · · ⊕ Fn for their
direct sum.

Definition 2.29. Let E be a vector space over K and F and G two subspaces of E. The
pair (F , G) is called a complementary pair of subspaces of E if E = F ⊕ G. We also say that F
and G are complementary subspaces of E.

The following Theorem relies on Zorn’s Lemma (and hence the axiom of choice) and will
not be proved. Anyway, the result is theoretical and not constructive.

Theorem 2.30. Let E be a vector space and F a subspace of E. There exists a subspace G of E
such that E = F ⊕ G.

5. Bases and Dimension

5.1. Independent Family of Vectors
Definition 2.31. Let E be a vector space over K and (u1 , . . . , un ) a finite family of vectors
of E. The family (u1 , . . . , un ) is said to be an independent family of vectors if
 
∀λ1 , . . . , λn ∈ K, λ1 u1 + · · · + λn un = 0E =⇒ λ1 = · · · = λn = 0 .

A family (ui )i∈I of vectors of E is said to be independent if all its finite subfamilies are
independent.
A family of vectors of E is said to be a dependent family if it is not independent.

The following results are straightforward:

Proposition 2.32. Let E be a vector space over K. Let (ui )i∈I be an independent family of
vectors of E. Then for all subset J ⊂ I the family (ui )i∈J is also an independent family of E.

Proposition 2.33. Let E be a vector space over K. A familyF = (ui )i∈I of vectors of E is
dependent if and only if there exists j ∈ I such that u j ∈ Span ui i ∈ I \ { j } .

Proposition 2.34. Let E be a vector space over K. Let F be an independent family of vectors
/ Span F , then the family F ∪ (u) is an independent
of E. If u is a vector of E such that u ∈
family.
 5. BASES AND DIMENSION 19

In the very special case where the family of vectors we consider only contains two vectors,
we have the following useful result:

Proposition 2.35. Let u and v be two vectors of a vector space E. The family (u, v) is inde-
pendent if and only if u and v are not collinear (i.e., if they are not proportional).

Remarks 2.36.
(1) If (ui )i∈I is an independent family of vectors of E, then ∀i ∈ I , ui 6= 0E .
(2) If u ∈ E \ {0E } then (u) is an independent family of vectors of E.
(3) Adding any number of vectors to a dependent family yields a dependent family.
(4) Removing any number of vectors to an independent family yields an independent
family.

Example 2.37. Let n ≥ 1 and E = Kn . It should be clear that any family of vectors of E
that contains more that n vectors is a dependent family. Indeed, let F = (u1 , . . . , un+1 ) be
a family of (n + 1) vectors of E. Let λ1 , . . . , λn+1 ∈ K such that λ1 u1 + · · · + λn+1 un+1 =
0E . Writing this equality componentwise yields a homogeneous system with n equations
and (n +1) unknowns. Hence this system is of rank at most n and, since it is homogeneous
and has (n + 1) unknowns, it possesses an infinite number of solutions.

5.2. Generating Families of Vectors

Definition 2.38. A family (ui )i∈I is said to be a generating family of the vector space E if

E = Span ui i ∈ I .


Remarks 2.39.
(1) Let E be a vector space. Adding any number of vectors of E to a generating family
of E yields a generating family of E.
(2) It is important to specify the space the family generates: the sentence “The family
of vectors (ui )i∈I is a generating family” is incomplete (and hence false) because
it misses the specification of the space it generates. In fact, any family can be
considered generating. . . it is a generating family of the space it generates!
5.3. Some Technical Results
The following two technical results will be used in the sequel:

Lemma 2.40. Let E be a vector space over K and let (v1 , . . . , v m ) be a generating family
of vectors of E. Let k ∈ {1, . . . , m} and for i ∈ {1, . . . , m}, λi ∈ K such that λk 6= 0. For
all i ∈ {1, . . . , m}, define the vectors wi of E as:

vmi
 if i 6= k
wi = X

 λ j v j if i = k.
j =1

Then the family (w1 , . . . , w m ) is again a generating family of E.

Proof. Let u ∈ E. We must prove that u can be written as a linear combination of the family
of vectors (w1 , . . . , w m ). We know that the family of vectors (v1 , . . . , v m ) is a generating
20 2. VECTOR SPACES

family of E, hence u can be written as a linear combination of the family (v1 , . . . , v m ), say
m
X
u= µi vi
i =1

with for all i ∈ {1, . . . , m}, µi ∈ K. Clearly, since λk 6= 0,

1 X λi
vk = wk − w.
λk λ i
i∈{1,...,m}\{k} k

Hence,
!
X 1 X λi
u= µi wi + µk wk − wi
i∈{1,...,m}\{k}
λk i∈{1,...,m}\{k}
λ k

λ µ
 
µi − i wi + k wk .
X
=
i∈{1,...,m}\{k}
λk λk

Hence u can indeed be written as a linear combination of the family of vectors (w1 , . . . , w m ).

Lemma 2.41. Let E be a vector space over K, let G = (v1 , . . . , v m ) be a generating family
of vectors of E and let u be a non-zero vector of E. Then there exists k ∈ {1, . . . , m} such
that the family of vectors (v1 , . . . , vk−1 , u, vk+1 , . . . , v m ) is a generating family of vectors of E.
Moreover, if ` ∈ {1, . . . , m − 1} is such that the family of vectors (v1 , . . . , v` , u) is independent,
then we may choose k > `.

Proof. Since the family G is a generating family of E, the vector u can be written as a linear
combination of the vectors of G , say
m
X
u= λi vi
i=1

with for all i ∈ {1, . . . , m}, λi ∈ K. Since u is not the zero vector of E, there exists
k ∈ {1, . . . , m} such that λk 6= 0. Hence, by Lemma 2.40, the family of vectors
(v1 , . . . , vk−1 , u, vk+1 , . . . , v m )
is a generating family of vectors of E. We now prove the last part: if ` ∈ {1, . . . , m − 1} is
such that the family of vectors (v1 , . . . , v` ) is independent, then since we have
X̀ m
X
u− λi vi = λi vi .
i=1 i=`+1

we conclude that it is impossible to have ∀i ∈ {` + 1, . . . , m}, λi = 0, because otherwise we

would have 1 = 0, which is impossible.

The following theorem will make the proof of Theorem 2.47 below straightforward. But
it is also an interesting result, in that it may shortcut some computations in some exercises.

Theorem 2.42. Let E be a vector space over K, let F = (u1 , . . . , un ) be an independent

family of vectors of E and let G = (v1 , . . . , v m ) be a generating family of vectors of E. Then
Card F = n ≤ Card G = m.
 5. BASES AND DIMENSION 21

Proof. Since the family F is an independent family of vectors, for all i ∈ {1, . . . , n}, ui 6=
0E . Hence, we proceed by steps: the first step is to apply Lemma 2.41 to the generating
family G and the non-zero vector u1 . Hence, and after reindexing if necessary, we obtain
the following generating family of E: (u1 , v2 , . . . , v m ). The second step is similar: since
the vector u2 is non-zero, we apply Lemma 2.41 again, and we obtain, after reindexing
if necessary, a generating family of E: (u1 , u2 , v3 , . . . , v m ). Note that we keep the vector
u1 because of the last part of the statement of Lemma 2.41. Now, if m < n, the process
eventually exhausts the vi ’s and leads to the generating family of E: (u1 , . . . , u m ). But
this is impossible since u m+1 cannot be written as a linear combination of the family of
vectors (u1 , . . . , u m ). Hence, n ≤ m.
5.4. Basis of a Vector Space
Definition 2.43. A family (ui )i∈I of vectors of E is said to be a basis of the vector space E if
it is an independent family and a generating family of E.

Theorem 2.44 (Characterization of bases). Let E be a vector space over K and let (ui )i∈I
be a basis of the vector space E. Then any vector of E can be written as a linear combination
of the family of vectors (ui )i∈I , and the linear combination is unique.

Proof. Since (ui )i∈I is a generating family of E, any vector of E can be written as a linear
combination of the ui ’s. Since the family (ui )i∈I is an independent family, this linear
combination must be unique.

We admit the following theorems, that can be proved using Zorn’s Lemma:

Theorem 2.45 (Existence of Basis). Any vector space E 6= {0E } over K possesses bases.

The following theorem is more specific, but non-constructive. We admit it as its proof also
relies on Zorn’s Lemma:

Theorem 2.46 (Incomplete Basis Theorem). Let F = (ui )i∈I be an independent family of
vectors of the vector space E (the family F can possibly be the empty family) and let G = (v j ) j ∈J
be a generating family of E. Then we can complete the family F with vectors of G to obtain a
basis of E.

The following theorem is a very important theoretical result, as it will enable us to define
the dimension of a vector space:

Theorem 2.47. Let E be a vector space over K and let B and B 0 be bases of E. Then B and
B 0 have the same number of elements (either a positive integer or the symbol +∞).

Proof. We split the proof in two parts:

• If both B and B 0 have a finite number of elements, say

Card B = m and Card B 0 = n.

Then, from Theorem 2.42, we clearly have m ≤ n and n ≤ m, hence m = n.

• If one of B or B 0 have a finite number of elements and the other has an infinite
number of elements, say Card B = m ∈ N∗ and Card B 0 = +∞: we show that
this situation is impossible. Since B 0 is an independent family that contains an
22 2. VECTOR SPACES

infinite number of elements, any subfamily of m + 1 vectors of B 0 is an indepen-

dent family of vectors of E, hence, by Theorem 2.42 we must have m + 1 ≤ m,
which is impossible.
5.5. Standard Bases
Some standard vector spaces have standard bases, known as standard bases (or canonical
basis or natural basis).

1) Standard basis of E = Kn . Let n ≥ 1. The standard basis of the vector space E = Kn

over K is the family B = (e1 , . . . , en ), where

ei = (0, . . . , 1, . . . , 0)

(the “1” is at the i-th position). We prove that B is indeed a basis of E, in two steps:
• The family B is an independent family of vectors of E: indeed, let λ1 , . . . , λn be n
scalars such that λ1 e1 +· · ·+λn en = 0E . If we compute λ1 e1 +· · ·+λn en , we obtain

λ1 e1 + · · · + λn en = (λ1 , . . . , λn ),

hence we see that λ1 e1 + · · · + λn en = 0E if and only if λ1 = · · · = λn = 0. Hence

the familyB is an independent family of vectors of E.
• The family B is a generating family of vectors of E: let u = (x1 , . . . , xn ) ∈ Kn ,
that is, for any i ∈ {1, . . . , n}, xi ∈ K. Clearly, u = x1 e1 + · · · + xn en , hence u
can be written as a linear combination of B, hence B is a generating family of
vectors of E.

2) Standard basis of E = Kn [X ]. Let n ≥ 0, and consider the vectors space E = Kn [X ]

over K. The standard basis of E is the family

B = (1, X , . . . , X n ).

It should be clear that B is a basis of E, since every polynomial of degree non-greater that n
is a linear combination of vectors of B, and the decomposition of every polynomial of E
on B is unique.

3) Standard basis of E = K[X ]. In E = K[X ], the standard basis of E is

B = (1, X , . . . , X k , . . .).

It is clear that B is a basis of E, since every polynomial of E can be uniquely written as a

linear combination of vectors of B.
5.6. Dimension of a Vector Space and Grassmann Formula
Theorem 2.47 justifies the following definition:

Definition 2.48. Let E be a vector space over K.

• The dimension of E, denoted by dim E (or dimK E), is the number of vectors of
any basis of E. If E is the null vector space we set dim E = 0.
• The vector space E is said to be finite-dimensional if its dimension is a finite number.
We sometimes write dim E < +∞ in this case.
• It is said to be infinite-dimensional if dim E = +∞.
 5. BASES AND DIMENSION 23

Example 2.49. The following equalities are obvious from the knowledge of the standard
bases:
dim Kn = n,
dim K[X ] = +∞,
dim Kn [X ] = n + 1.

Proposition 2.50. Let E be a vector space over K and let F be a subspace of E. Then
dim F ≤ dim E.

Proof. If F = {0E }, the result is clear. Otherwise, let (ui )i∈I be a basis of F . Since the
family (ui )i ∈I is an independent family of vectors of E, we can complete it in order to
obtain a basis of E, say (ui , i ∈ I ; v j , j ∈ J ). Clearly,
dim F = Card I ≤ Card I + Card J = dim E.

Theorem 2.51 (Grassmann Formula). Let F and G be two subspaces of the vector space E.
Then:
dim(F + G) + dim(F ∩ G) = dim F + dim G.
In particular, if F and G are independent subspaces:
dim(F ⊕ G) = dim F + dim G.

Proof. The equality obviously holds if either F = {0E } or G = {0E } or F ∩ G = {0E }

(and the latter equality corresponds to the case where F and G are independent, i.e., in
direct sum). Let’s now prove Grassmann Formula in the case F 6= {0E }, G 6= {0E } and
F ∩ G 6= {0E }. By Theorem 2.45, the vector space F ∩ G possesses a basis, say (ui )i∈I . By
Theorem 2.46, we can complete the independent family (ui )i∈I to obtain a basis of F , say
(ui , i ∈ I ; v j , j ∈ J )
and similarly we can complete the family (ui )i∈I to obtain a basis of G, say (ui , i ∈
I ; wk , k ∈ K). Clearly, the family
(ui , i ∈ I ; v j , j ∈ J ; wk , k ∈ K)
is a basis of F + G. Hence,
dim(F + G) + dim(F ∩ G) = Card(I ) + Card(J ) + Card(K) + Card(I ),
dim F + dim G = Card(I ) + Card(J ) + Card(I ) + Card(K),
hence the formula.

Remark 2.52. The equality in Grassmann Formula holds and makes sense in N ∪ {+∞}.
Hence, we avoid writing an equality like dim(F +G) = dim F +dim G −dim(F ∩G) when
we don’t know in advance whether dim(F ∩ G) is finite: remember that in N ∪ {+∞}, the
operation +∞ − (+∞) is not valid.

5.7. Coordinates of a Vector in a Basis

Definition 2.53. Let E be a finite-dimensional vector space over K, B = (u1 , . . . , un ) a
basis of E and v ∈ E. The vector v admits a unique decomposition of the form
v = λ1 u1 + · · · + λn un , λi ∈ K.
24 2. VECTOR SPACES

We say that (λ1 , . . . , λn ) are the coordinates of v in B and we write:

λ1
 
 .. 
[v]B =  .  .
λn

Examples 2.54.
(1) Consider the vector space E = R3 together with its standard basis B = (e1 , e2 , e3 ).
Consider the vector u = (−1, 3, 2) ∈ E. Then
 
−1
[u]B =  3  .
2

Now consider the following vectors of E: f1 = (1, 1, 1), f2 = (0, 1, 1) and f3 =

(0, 0, 1). We now prove that the family D = ( f1 , f2 , f3 ) is a basis of E and we
determine [u]D : By the characterization of a basis, in order to show that D is a
basis of E, we need to show that every vector of E can be uniquely decomposed
as a linear combination of vectors of D. Hence, let v = (a, b , c) ∈ E and let’s
try to write v as a linear combination of D, say v = λ1 f1 + λ2 f2 + λ3 f3 , i.e.,
(a, b , c) = (λ1 , λ1 + λ2 , λ1 + λ2 + λ3 ), i.e.,
 
 λ1 =a  λ1 = a
λ1 + λ2 = b ⇐⇒ λ2 = b − a
 λ + λ + λ = c. λ = c − b .
1 2 3 3

Since the system we solved does possess a solution and its solution is unique, we
conclude that the family D is a basis of E. But also, by solving the system, we
showed that for all v = (a, b , c) ∈ R3 ,
v = (a, b , c) = a f1 + (b − a) f2 + (c − b ) f3 ,
or equivalently,
 
a
[v]D = (a, b , c) D =  b − a  .
 
c−b

In particular,
 
−1
[u]D = (−1, 3, 2) D =  4  ,
 
−1
which exactly means that u = − f1 + 4 f2 − f3 . Let’s check this equality:
− f1 + 4 f2 − f3 = −(1, 1, 1) + 4(0, 1, 1) − (0, 0, 1)
= (−1, −1, −1) + (0, 4, 4) + (0, 0, −1)
= (−1, 3, 2).

We indeed recover the vector u.

(2) Consider the vector
space E = C4 [X ], together with its standard basis B =
1, X , X 2 , X 3 , X 4 . For example, the coordinates of the polynomial P = (1 +
 5. BASES AND DIMENSION 25

i)X 4 − 3X 2 + 2i X − 3 + 2i in the basis B are

−3 + 2i
 
 2i 
[P ]B = 
 
 −3  .

 0 
1+i
Consider the polynomials P0 = 1, P1 = X + 1, P2 = X 2 + 1, P3 = X 3 + 1, P4 =
X 4 + 1 of E, and the family B 0 = (P0 , P1 , P2 , P3 , P4 ) of vectors of E. We show
that B 0 is a basis of E, by using the characterization of a basis: let Q = aX 4 +
b X 3 + cX 2 + d X + e be a vector of E, and try to write it as a linear combination
X4
of vectors of B 0 , say Q = λi Pi . This leads to the following linear system:
i=0

λ4 = a λ4 = a
 
λ3 =b  λ3 = b

 


 
λ2 = c ⇐⇒ λ2 = c
λ1 =d  λ1 = d

 

 
λ0 + λ1 + λ2 + λ3 + λ4 = e λ0 = e − a − b − c − d


Since this system always possesses a unique solution we conclude that the family
of vectors B 0 is a basis of E, and moreover,
 
e −a − b −c −d
 4
 d 
3 2
[Q]B 0 = aX + b X + cX + d X + e B 0 = 
  
 c .

 b 
a
In particular,
[P ]B 0 = (1 + i)X 4 − 3X 2 + 2i X − 3 + 2i B 0
 

−3 + 2i − (1 + i) + 3 − 2i
 
 2i 
=
 
 −3 

 0 
1+i
 
−1 − i
 2i 
=
 
 −3  ,

 0 
1+i
which exactly means that P = (−1 − i)P0 + 2i P1 − 3P2 + 0P3 + (1 + i)P4 . Let’s
check this identity:
(−1 − i)P0 + 2i P1 −3P2 + 0P3 + (1 + i)P4
= −1 − i + 2i(X + 1) − 3 X 2 + 1 + (1 + i) X 4 + 1

= (1 + i)X 4 − 3X 2 + 2i X − 3 + 2i.
We indeed recover the vector P .
26 2. VECTOR SPACES

5.8. Inclusion–Equality Theorem and Other Results about Dimension

Theorem 2.55 (Inclusion–Equality Theorem). Let E be a finite dimensional vector space
over K, and let F be a subspace of E. If dim F = dim E, then F = E.

Proof. If E = {0E }, the result is obvious, since F can only be the null subspace. Let’s
prove the theorem in the case E 6= {0E }, say dim E = n ∈ N∗ : By Theorem 2.45, the
vector space F possesses a basis, and this basis must have dim F = dim E = n elements,
say (u1 , . . . , un ). If F =
6 E, since F ⊂ E, by the Incomplete Basis Theorem 2.46, we can
complete the independent family (u1 , . . . , un ) to a basis of E, say (u1 , . . . , un , v1 , . . . , v m ),
which means that dim E = n + m > n, which is impossible.
We will often use the Inclusion–Equality Theorem in the following case, the proof of which
is straightforward from Theorem 2.55.

Corollary 2.56. Let E be a vector space over K and let F and G be two subspaces of E. If the
following two conditions are fulfilled,
• dim F = dim G < +∞,
• F ⊂ G,
then F = G.

Remark 2.57. The statements F ⊂ G and dim F = dim G < +∞ are very important in
Corollary 2.56.

In the following theorem, we have gathered some very simple results, but it’s important
to keep them in mind:

Theorem 2.58. Let E be a vector space over K of dimension n ∈ N∗ . Then:

(1) n independent vectors of E are also generating vectors of E, hence they form a basis
of E.
(2) n generating vectors of E are also independent vectors of E, hence they form a basis
of E.
(3) If (u1 , . . . , uk ) is an independent family of vectors of E, then k ≤ n.
(4) n + 1 vectors of E are always dependent.
(5) Every generating family of E possesses at least n elements.

Proof.
(1) Let (u1 , . . . , un ) be n independent vectors of E. If they don’t form a generating
family of E, it means we can complete it to form a basis of E, say
(u1 , . . . , un , v1 , . . . , v m ),
and hence dim E = m + n, which is impossible.
(2) Let (u1 , . . . , un ) be a family of generating vectors of E with n elements. If this
family of vectors is not independent, it means we can extract a subfamily that
will be a basis of E, say, after reindexing if necessary, (u1 , . . . , u m ) with m < n, in
which case dim E = m < n, which is impossible.
(3) Let F = (u1 , . . . , uk ) be an independent family of k vectors of E. If F is a basis
of E, then we know that k = n, hence k ≤ n, otherwise, we know that we can
complete F in order to obtain a basis of E, say (u1 , . . . , uk , v1 , . . . , v m ). Hence
n = dim E = k + m, which shows that k = n − m ≤ n.
 6. RANK OF A FAMILY OF VECTORS 27

(4) Let (u1 , . . . , un+1 ) be a family of n + 1 vectors of E. If they are independent, it

would mean, by the previous result, that n + 1 ≤ dim E = n, which is impossible.
(5) We know that from a generating family of vectors of E we can extract a subfamily
that will be a basis of E, hence this family must contain at least n vectors.

Definition 2.59 (Lines, Planes). Let E be a vector space over K.

• A line of E is a subspace of E of dimension 1.
• A plane of E is a subspace of E of dimension 2.

6. Rank of a Family of Vectors

Definition 2.60. Let F be a family of vectors of the vector space E. The rank of F , denoted
by rk F , is the dimension of Span F .

Proposition 2.61. Let F = (u1 , . . . , un ) be a family of n vectors of the vector space E, n ∈ N∗ .

Then:
rk F ≤ n.

Moreover,
rk F = n ⇐⇒ the family F is an independent family.

Proof. Clearly, the family F = (u1 , . . . , un ) is a generating family of the linear subspace
Span{u1 , . . . , un }, hence, by Theorem 2.58, we must have

n ≥ dim Span{u1 , . . . , un } = rk F .

The equality holds if and only if F is also a basis of Span F , hence if and only if the
family F is independent.

There is a nice algorithm to determine the rank of a family of vectors in a finite-dimensional

vector space: let E be a finite-dimensional vector space over K, and let B be a basis of E.
Consider a family F = (v1 , . . . , vk ) of vectors of E. The goal is to find the rank of F , that
is, the dimension of Span{v1 , . . . , vk }. For i ∈ {1, . . . , k} we set Vi = [vi ]B , and we write
the Vi ’s next to each other in a table. We then perform elementary column operations on
this table, to obtain a triangular table. At the end of the triangularization procedure, the
rank of F is the number of non-zero columns we’re left with. The elementary column
operations are similar to the elementary row operations used for the Gaussian elimination
procedure:
• We can exchange two columns.
• We can multiply any column by any non-zero number.
• We can add to any column any multiple of any other column.
• We can exchange two rows.
We illustrate the procedure on the following example: consider the real vector space E = R4
with its standard basis B, and consider the following vectors of E:

v1 = (1, 4, −1, 2), v2 = (2, 3, −1, 2), v3 = (3, 2, −1, 2),

v4 = (4, 1, 5, −2), v5 = (5, 10, 3, 2).
28 2. VECTOR SPACES

We determine the rank of the family F = (v1 , v2 , v3 , v4 , v5 ): we first write a table that
consists of the coordinates of the vi ’s in the basis B:

 
1 2 3 4 5
4 3 2 1 10
 
−1 −1 −1 5 3
2 2 2 −2 2

We now perform the elementary operations to get a triangular table:

   
1 2 3 4 5 1 0 0 0 0
4 3 2 1 10  4 −5 −10 −15 −10
rk 
−1 −1 −1 5
 = rk 
−1 1

3 C2 ←C2 −2C1 2 9 8 
2 2 2 −2 2 C 3←C3 −3C1 2 −2 −4 −10 −8
C4 ←C4 −4C1
C5 ←C5 −5C1
 
1 0 0 0 0
 4 −5 0 0 0
= rk 
−1 1 0 6

C3 ←C3 −2C2 6
C4 ←C4 −3C2 2 −2 0 −4 −4
C5 ←C5 −2C2
 
1 0 0 0 0
4 −5 0 0 0
= rk 
−1

C5 ←C5 −C4 1 0 6 0
2 −2 0 −4 0
 
1 0 0 0 0
4 −5 0 0 0
= rk 
−1

C2 ↔C3 1 6 0 0
2 −2 −4 0 0

We obtain three non-zero columns after the procedure, hence rk F = 3.

Exercises
Exercise 2.1 (Subspaces of R2 ). We consider the real vector space E = R2 .

1. Let D be a straight line in E that passes through 0E = (0, 0). Show that D is a subspace
of E.
2. Let F be a subspace of E such that F 6= {0E } and F 6= E. Show that F is a straight line
that passes through 0E .

Exercise 2.2 (Subspaces of R3 ). We consider the real vector space R3 .

1. Let D be a straight line in E that passes through 0E = (0, 0, 0). Show that D is a subspace
of E.
2. Let P be a plane in E that passes through 0E = (0, 0, 0). Show that P is a subspace of E.
3. Let F be a subspace of E such that F 6= {0E } and F 6= E. Show that F is either a straight
line that passes through 0E or a plane that passes through 0E .
 EXERCISES 29

Exercise 2.3. We consider the real vector space E = R[X ]. Which of the following subsets
of E are subspaces of E?

P1 = P ∈ E P (0) = 1 , P2 = P ∈ E P (3) = 0 ,
 

P3 = P ∈ E P (1) = P (2) , P4 = P ∈ E P + P 0 + P 00 = 0 ,
 

P5 = P ∈ E P (0) + 5P (4) = 0 , P6 = P ∈ E deg(P ) ≥ 8 ,

 

for n ∈ N, Rn [X ] = P ∈ E deg(P ) ≤ n .


Exercise 2.4 (Examples of subspaces of CN ). We consider the complex vector space E = CN .

Which of the following subsets of E are subspaces of E?

F1 = (un )n∈N ∈ E u42 = 1 ,



F2 = (un )n∈N ∈ E u42 = 0 ,


n o
F3 = (un )n∈N ∈ E lim un = 1 ,
n→∞
1
§  ‹ª
F4 = (un )n∈N ∈ E un = o ,
n→+∞ n
n o
F5 = (un )n∈N ∈ E lim un = u0 ,
n→+∞
n o
F6 = (un )n∈N ∈ E lim un = 0 ,
n→+∞

F7 = (un )n∈N ∈ E (un )n∈N converges ,



F8 = (un )n∈N ∈ E ∀n ∈ N, un+2 = un+1 − 5un .



Exercise 2.5 (Examples of subspaces of RR ). We consider the real vector space E = RR .

Which of the following subsets of E are subspaces of E?

F1 = f ∈ E f (0) = a ,


F2 = f

∈ E f is continuous ,
F3 = f

∈ E f is odd ,
F4 = f

∈ E f is even ,
F5 = f

∈ E f is not continuous at 0 ,
F6 = f

∈ E f is monotone ,
n o
for a ∈ R, F7 = f ∈ E lim f (x) = a ,
x→+∞
n o
F8 = f ∈E lim f (x) = lim f (x) = 0 ,
x→∞ x→−∞

for T ∈ R, F9 = f

∈ E f is periodic of period T ,
F10 = f ∈ E ∃T > 0, f is periodic of period T ,


for k ∈ N, C k (R) = f ∈ E f is of class C k on R .


30 2. VECTOR SPACES

Exercise 2.6 (Subspaces in R4 and intersection). We consider the vector space E = R4 .

Which of the following subsets of E are subspaces of E?
F1 = (x, y, z, t ) ∈ R4 x 2 + y − z = 0 ,


F2 = (x, y, z, t ) ∈ R4 x + 2y + 3z − t = 0 and 2x − t = 0 ,


F3 = (x, y, z, t ) ∈ R4 ∃a ∈ R, x = 3a, y = 2a, z = a ,



F4 = (x, y, z, t ) ∈ R4 x = y .


What is the intersection F2 ∩ F4 ?

Exercise 2.7 (Linear Combination). Let E = R4 and

u = (1, 0, 2, −3), v = (0, 4, 2, 2), w = (2, −4, 2, −8), a = (1, 3, 4, 2).
Show that w ∈ Span{u, v}. Does a belong to Span{u, v}? Does a belong to Span{u, v, w}?
Does w belong to Span{a, u}?

Exercise 2.8. Let E = R3 , u = (0, 1, 2), v = (2, 1, 0), a = (1, 1, 1) and b = (−1, 2, −3). Let
P = Span{u, v} and Q = Span{a, b }.
1. Show that P + Q = E.
2. Is the sum P + Q a direct sum?
3. Is it possible to find a subspace F of E such that F ⊕ P = F ⊕ Q = E?

Exercise 2.9. Let E = R[X ]. For all n ∈ N we define the following subsets of E:
Fn = λX n λ ∈ R .


1. Show that, for all n ∈ N, Fn is a subspace of E.

M (Fn )n≥0 is an independent family of subspaces of E.
2. Show that the family
3. Do we have E = Fn ?
n≥0

4. Is the family Rn [X ] n≥0 an independent family of E?
X
5. Show that E = Rn [X ].
n≥0

Exercise 2.10. Let E = CN . For n ∈ N we define the following subsets of E:

Fn = (u p ) p∈N ∈ E ∀ p 6= n, u p = 0 , and Gn = (u p ) p∈N ∈ E ∀ p > n, u p = 0 .
 

1. Show that for all n ∈ N, Fn and Gn are subspaces of E.

2. Show that (Fn )n≥0 is an independent family of subspaces of E. Is (Gn )n≥0 an independent
family of E? M X
3. Do we have E = Fn ? Do we have E = Gn ?
M n≥0 X n≥0
4. Show that Fn = Gn . Describe this subspace.
n≥0 n≥0

Exercise 2.11. Let E = C (R, R), the vector space of continuous functions on R with values
in R. We consider the following vectors of E:
u = sin, v = cos, w = 1, a = sin2 , b = cos2 , c = (x 7→ x), d = exp .
 EXERCISES 31

We also consider the following subspaces of E:

A = Span{u, v}, B = Span{a, b }, C = A + B, D = B + Span{c, d },
E0 = f ∈ E f is even , E1 = f ∈ E f is odd .
 

1. Show that w ∈ B.
2. Is the sum A + B a direct sum?
3. Show that D = B ⊕ Span{c, d }.
4. Show that E = E0 ⊕ E1 . Give the decomposition of the vectors u, v, w, a, b , c, d with
respect to the direct sum E0 ⊕ E1 .
5. Do we have D = (D ∩ E0 ) ⊕ (D ∩ E1 )?

Exercise 2.12. We consider the real vector space E = RR . Let a ∈ R.

1. Show that Fa = f ∈ E f (a) = 0 is a subspace of E.


2. Determine two different subspaces Ga and Ha such that E = Fa ⊕ Ga = Fa ⊕ Ha .

3. What is the dimension of Fa ? of Ga ? of Ha ?
4. We are now going to describe all the possible complementary subspaces of Fa in E: let K
be a subspace of E such that E = Fa ⊕ K.
a) Explain why there exists k ∈ K such that k(a) = 1.
b) Show that for all f ∈ K, f = f (a)k.
c) Deduce that K = Span{k}. This shows that every complementary subspace of Fa in
E is the linear span of a function that doesn’t vanish at a.

Exercise 2.13.
1. Show that the family B = (0, 1, 2), (−1, 0, 1), (3, 2, 0) is a basis of R3 .

2. What are the coordinates of the vector u = (0, 2, 1) in the standard basis of R3 ? In the
basis B ?

Exercise 2.14. We consider the vector space E = R4 over R and the vectors x1 = (2, −2, 3, 1)
and x2 = (−1, 4, −6, −2) of E. Let F = Span{x1 , x2 }.
1. Compute dim F .
2. Show that there exists a basis of F containing the vector e1 = (1, 0, 0, 0). Determine such
a basis B of F .
3. Does there exist a basis of R4 containing B ?

Exercise 2.15. Determine the rank of the following families of vectors of E, and specify if
they are independent, generating, and bases.
1. E = R3 :
a) A = (1, 0, 1), (−1, 1, 2), (−2, 1, 2) .

b) B = (1, 0, 1), (2, 0, 3), (−1, 1, 1), (0, 0, 1) .

2. E = R3 [X ]: C = (X 2 + 3X − 1, 3X 2 − 5X +
5, −7X 2 + 9X − 17).
3. E = R4 : D = (a, 1, 1, 0), (1, a, 1, 0), (1, 1, a, 0) (Specify according to the value of a ∈ R).
4. E = C (R), F = ( f1 : x 7→ x 2 , f2 : x 7→ e x , f3 : x 7→ sin x).

Exercise 2.16. We consider the following vectors of R4 :

a = (1, 2, 3, 4), b = (2, 2, 2, 6), c = (0, 2, 4, 2), d = (1, 0, −1, 2), e = (2, 3, 0, 1).
We define two subspaces of R4 by F = Span{a, b , c} and G = Span{d , e}. Determine the
dimension and a basis of each of the following subspaces: F , G, F + G and F ∩ G.
32 2. VECTOR SPACES

Exercise 2.17.
1. Determine a basis of the subspace of R4 defined by:
E1 = (a, b , c, d ) ∈ R4 a = 2b − c and d = a + b + c .


2. We set E2 = Span (3, 1, 0, 3), (−1, 1, 1, 0) . Show that E1 and E2 are complementary


subspaces in R4 .

Exercise 2.18. Let A be a polynomial of degree 2. We consider the set

F = P ∈ R5 [X ] P is divisible by A .


1. Prove that F is a subspace of R5 [X ], and determine a basis of F .

2. Determine a subspace G of R5 [X ] such that F and G are complementary subspaces in
R5 [X ].

Exercise 2.19. We consider the subset E of RN that consists of all real sequences (un )n∈N
satisfying:
∀n ∈ N, un+2 = 5un+1 − 6un .
1. Show that E is a real vector space.
2. Determine the geometric sequences that belong to E . Deduce that for all (λ, µ) ∈ R2 ,
the sequence (un )n∈N defined by ∀n ∈ N, un = λ2n + µ3n belongs to E .
3. Let (a, b ) ∈ R2 . Show that there exists at most one sequence (un )n≥0 of E satisfying
u0 = a, u1 = b .
4. Deduce that for each sequence (un )n≥0 of E we can associate a unique pair of real num-
bers (λ, µ) such that for all n ∈ N one has un = λ2n + µ3n .
5. What can you deduce about the vector space E ?

Exercise 2.20. In the real vector space E = RR we consider the functions f1 , f2 , f3 , g1 , g2

and h defined as
f1 (x) = 1, f2 (x) = cos(2x), f3 (x) = cos(4x),
2 2
g1 (x) = sin x, g2 (x) = cos (2x), h(x) = cos2 x
We set F = Span{ f1 , f2 , f3 }, G = Span{ g1 , g2 } and H = Span{h}.
1. Determine the dimensions of F , G and H .
2. Show that G ⊂ F and H ⊂ F .
3. Show that F = G ⊕ H .

Exercise 2.21 (From Term-Test, April 2010). We consider the following subset of R4 :
F = (x, y, z, t ) ∈ R4 2x − 3y = 0 and y + z + 4t = 0 .


We also consider the following vectors of R4 :

e1 = (1, 0, 2, −1), e2 = (3, 2, 2, −1), e3 = (0, 0, 1, 0),
and we set G = Span(e1 , e2 , e3 ).
1. Show that F is a subspace of R4 . Determine a basis of F and its dimension.
2. What is the dimension of G?
3. Are the subspaces F and G complementary subspaces of R4 ?
4. Do we have F + G = R4 ?
5. What is the dimension of F ∩ G? Determine a basis of F ∩ G.
 EXERCISES 33

Exercise 2.22 (Complex Structure on a Real Vector Space). Let E be a real vector space,
and let j ∈ L(E) such that j 2 = − IdE . Such an endomorphism j is called a complex structure
on E.
1. Let z = x + i y ∈ C, with x, y ∈ R, let u ∈ E, and consider the following operation:
z u = x u + y j (v).
Prove that E, together with its addition and with the scalar multiplication defined by
is a complex vector space.
2. Let f ∈ LR (E). Under what condition is f an endomorphism of the complex vector
space E, with the complex structure inherited from j ? Does j itself satisfy this condi-
tion?
3. If E is a finite-dimensional vector space, prove that such a j exists if and only if dimR E
1
is even, and that in this case dimC E = dimR E.
2
 3
Linear Maps

Throughout this chapter, E and F are two vector spaces over K = R or C.

1. Definitions, Vocabulary and Properties

1.1. Definitions
Definition 3.1. A map f from E to F is said to be a linear map (or linear transformation,
linear function, linear operator, or a homomorphism) if:
• It is additive, that is:
∀u, v ∈ E, f (u + v) = f (u) + f (v).
• It is homogeneous (of degree 1), that is:
∀λ ∈ K, ∀u ∈ E, f (λu) = λ f (u).
The set of linear maps from E to F is denoted by L(E, F ).

Remark 3.2. If f ∈ L(E, F ), then f (0E ) = 0F . Indeed, since f is linear, it is homogeneous

of degree 1, hence,
f (0E ) = f (0 · 0E ) = 0 · f (0E ) = 0F .
In particular, if a map f : E −→ F satisfies f (0E ) 6= 0F , then f cannot be linear.

The following result is obvious:

Proposition 3.3. Let f : E −→ F . The map f is linear if and only if

∀λ, µ ∈ K, ∀u, v ∈ E, f (λu + µv) = λ f (u) + µ f (v).

Remarks 3.4.
(1) Let f ∈ L(E, F ) and let u1 , . . . , un be n vectors of E, n ∈ N∗ , and let λ1 , . . . , λn be
n scalars. Then:
f (λ1 u1 + · · · + λn un ) = λ1 f (u1 ) + · · · + λn f (un ).
35
36 3. LINEAR MAPS

(2) More generally if (ui )i ∈I is a family of vectors of E such that for all i ∈ I except a
finite number, ui = 0E , and if (λi )i∈I is a family of scalars, then
‚ Œ
X X
f λi ui = λi f (ui ).
i∈I i∈I

(3) Similarly, if (ui )i∈I is a family of vectors of E, and if (λi )i∈I is a family of scalars
such that for all i ∈ I except a finite number, λi = 0, then
‚ Œ
X X
f λi ui = λi f (ui ).
i∈I i∈I

In practice, we’ll use the following characterization of linear maps:

Proposition 3.5. Let f : E −→ F . The map f is linear if and only if it satisfies the following
condition:
(∗) ∀λ ∈ K, ∀u, v ∈ E, f (u + λv) = f (u) + λ f (v).

Proof. If f is linear, then it clearly satisfies Condition (∗). Conversely, let’s assume that f
satisfies Condition (∗), and let’s prove that f is linear: we show that f is additive: let
u, v ∈ E, then
f (u + v) = f (u + 1 · v) = f (u) + 1 · f (v) = f (u) + f (v).
We now show that f (0E ) = 0F :
f (0E ) = f (0E + 0E ) = f (0E ) + f (0E ),
hence f (0E ) = 0F . We now show that f is homogeneous of degree 1: let u ∈ E and λ ∈ K,
then
f (λu) = f (0E + λu) = f (0E ) + λ f (u) = 0F + λ f (u) = λ f (u).
Hence f is a linear map.
1.2. Vocabulary and Notations
1.2.1. Vocabulary
Definition 3.6. Let f ∈ L(E, F ).
• The map f is said to be 1 − 1 (or injective or a monomorphism) if:
 
∀u, v ∈ E, f (u) = f (v) =⇒ u = v .

• The map f is said to be onto (or surjective or an epimorphism) if:

∀w ∈ F , ∃u ∈ E, f (u) = w.
• The map f is said to be a bijection from E to F (or bijective or an isomorphism
from E to F ) if it is both 1 − 1 and onto.
• In the special case when E = F , we call f an endomorphism of E. If, moreover, f
is an isomorphism, we say that f is an automorphism of E.
• In the special case when F = K, we call f a linear form on E.
• The vector spaces E and F are said to be isomorphic if there exists an isomorphism
between E and F .
 1. DEFINITIONS, VOCABULARY AND PROPERTIES 37

1.2.2. Notations
• L(E, F ) is the set of linear maps from E to F .
• L(E) = L(E, E) is the set of endomorphisms of E.
• E ∗ = L(E, K) is the set of linear forms on E. It is called the dual space of E. See
Section 6.
• GL(E) is the set of automorphisms of E.
1.2.3. The Identity Map
There is a very special map in L(E), called the identity map of E and denoted by id or idE .
It is defined by:
id : E −→ E
u 7−→ u.
1.3. Basic Results
If a mapping is linear, injectivity can be checked by only checking whether the nil vector
(of its codomain) is only attained once:

Proposition 3.7. A linear map f ∈ L(E, F ) is injective if and only if:

 
∀u ∈ E, f (u) = 0F =⇒ u = 0E .

Proof. Suppose f injective. Since f (0E ) = 0F one has:

f (u) = 0F =⇒ f (u) = f (0E ) =⇒ u = 0E .
The first implication follows from the fact that f is linear. Now suppose that f satisfies
the property of the statement. Let u, v ∈ E. Then:
f (u) = f (v) =⇒ f (u) − f (v) = 0F
=⇒ f (u) + f (−v) = 0F
=⇒ f (u − v) = 0F
=⇒ u − v = 0E
=⇒ u = v.

Proposition 3.8. Let f ∈ L(E, F ).

(1) f is injective if and only if f maps every independent family of vectors of E to an
independent family of vectors of F .
(2) Let (ui )i∈I be a basis of E. Then f is injective if and only if the family f (ui ) i∈I is

an independent family of vectors of F .

(3) Let (ui )i∈I be a generating family of vectors of E. Then f is surjective if and only if
the family f (ui ) i∈I is a generating family of vectors of F .

(4) Let (ui )i∈I be a basis of E. Then f is bijective if and only if f (ui ) i∈I is a basis of F .

Proof.
(1) Let (ui )i∈I be an independent family of vectors of E. If the family f (ui ) i∈I is

dependent, it means that there exists distinct elements i1 , . . . , ik ∈ I and non-zero

scalars λi1 , . . . , λik ∈ K such that
λi1 f ui1 + · · · + λik f uik = 0F ,



38 3. LINEAR MAPS

hence
f λi1 ui1 + · · · + λik uik = 0F .

Now since λi1 ui1 + · · · + λik uik 6= 0E , the map f cannot be injective.
Conversely, assume that f maps every independent family of vectors of E to
an independent family of vectors of F . Let u ∈ E be a non-zero vector. Since
u 6= 0E , the family (u) is an independent family of vectors of
E, and hence is
mapped to an independent family of vectors of F , namely f (u) , and as such, we
must have f (u) 6= 0F , hence f is injective.
(2) Assume that f is injective. Since the family (ui )i∈I is a basis of E, it is also an
independent family of vectors of E, and hence, by the previous statement, the
family f (ui ) i ∈I is an independent family of vectors of F . Conversely, assume

that the family f (ui ) i ∈I is an independent family of vectors of F , and let u ∈ E be

a vector such that f (u) = 0F . We must show that u = 0E : since the family (ui )i∈I
is a basis of E, there exists distinct elements i1 , . . . , ik ∈ I and scalars λi1 , . . . , λik
such that u = λi1 ui1 + · · · + λik uik . Now since f (u) = 0F , and since f is linear, we
must have f (u) = λi1 f ui1 +· · ·+λik f uik = 0F , and since the family f (ui ) i∈I

is independent, we must have λi1 = · · · = λik = 0. Hence u = 0E .

(3) We first show that if f is surjective, then the family f (ui ) i∈I is a generating

family of F : Let v ∈
F ; we must show that v can be written as a linear combination

of the family f (ui ) i∈I . Since f is surjective, there exists u ∈ E such that f (u) =
v. Now, since the family (ui )i ∈I is a generating family of E, there exists elements
i1 , . . . , ik ∈ I and scalars λi1 , . . . , λik such that u = λi1 ui1 + · · · + λik uik , hence
v = f (u) = f λi1 ui1 + · · · + λik uik = λi1 f ui1 + · · · + λik f uik .

Conversely, we show that if the family f (ui ) i ∈I is a generating family of F ,

then f is a surjective map: let v ∈ F ; we
must show that there exists u ∈ E

such that f (u) = v. Since the family f (ui ) i∈I is a generating family of F , there
exists elements i1 , . . . , ik ∈ I and scalars λi1 , . . . , λik such that v = λi1 f ui1 + · · · +

λik f uik . If we set u = λi1 ui1 + · · · + λik uik , we clearly have f (u) = v, since f is

linear.
(4) Since f
is injective and since (ui )i∈I is an independent family of E, the family
f (ui ) i∈I is also an independent family. Since f is surjective and since (ui )i∈I
is a generating family of E, the family f (ui ) i∈I is also a generating family of F .

Hence the family f (ui ) i∈I is a basis of F .

This simple proposition has the very important (but very simple) consequence:

Proposition 3.9.
• If there exists an injective linear map from E to F then dim E ≤ dim F .
• If there exists a surjective linear map from E to F then dim E ≥ dim F .
• In particular, if there exists an isomorphism between E and F then dim E = dim F .

1.4. Properties
Proposition 3.10.
(1) Let f ∈ L(E, F ) and λ ∈ K. Then λ f ∈ L(E, F ).
(2) Let f , f 0 ∈ L(E, F ). Then f + f 0 ∈ L(E, F ).
 1. DEFINITIONS, VOCABULARY AND PROPERTIES 39

(3) In particular, L(E, F ) is a vector space over K. In particular, E ∗ is also a vector space
over K.
(4) Let G be a third vector space over K, and f ∈ L(E, F ) and g ∈ L(F , G). Then
g ◦ f ∈ L(E, G).
(5) The composition of two injective linear maps is again an injective linear map.
(6) The composition of two surjective linear maps is again a surjective linear map.
(7) In particular, the composition of two bijective maps is again a bijective map.
(8) In particular, if f , g ∈ GL(E) then f ◦ g ∈ GL(E).
(9) If f ∈ L(E, F ) is an isomorphism, then f −1 ∈ L(F , E) and f −1 is an isomorphism.
In particular, if f ∈ GL(E) then f −1 ∈ GL(E).
(10) Let f ∈ E ∗ . Then f is surjective if and only if f 6= 0E ∗ .

Proof.
(1) Obvious.
(2) Obvious.
(3) The set L(E, F ) is a non-empty subset of F E , the functions from E to F , which is
a vector space over K. Hence, to prove that L(E, F ) is a vector space over K, we
only need to prove that it is a subspace of F E , but this is a direct consequence of
the two previous statements.
(4) Let u, v ∈ E and let λ ∈ K. Then:

g ◦ f (u + λv) = g f (u + λv)

= g f (u) + λ f (v)

= g f (u) + λg f (v)

= g ◦ f (u) + λg ◦ f (v).

Hence, g ◦ f is a linear map from E to G.

(5) Let f ∈
L(E, F ) and g ∈ L(F , G) be two injective maps, and let u ∈ E such that
g f (u) = 0G . Since g is injective, we must have f (u) = 0F , and since f is
injective, we must have u = 0E , hence g ◦ f is injective.
(6) Let f ∈ L(E, f ) and g ∈ L(F , G) be two surjective
maps, and let w ∈ G. We must
show that there exists u ∈ E such that g f (u) = w. Since g is surjective, there
exists v ∈ F such that g (v)
= w. Since f is surjective, there exists u ∈ E such that
f (u) = v. Hence, g f (u) = g (v) = w.
(7) This is a direct consequence of the two previous statements.
(8) This is a direct consequence of the previous statement.
(9) Let u, v ∈ F and let λ ∈ K. We must show that f −1 (u + λv) = f −1 (u) + λ f −1 (v):
let w = f −1 (u) + λ f −1 (v), and let’s show that w = f −1

(u + λv). Clearly, since
f is linear, f (w) = f f (u) + λ f (v) = f f (u) + λ f f (v) = u + λv,
−1 −1 −1 −1

hence, w = f −1 f (w) = f −1 (u + λv).

(10) If f is surjective, there exists u ∈ E such that f (u) = 1, hence f =

6 0E ∗ . Conversely,
if f 6= 0E ∗ then there exists u ∈ E such that f (u) 6= 0. Let λ ∈ K, then, since f is
linear,
λ λ
 
f u = f (u) = λ.
f (u) f (u)

Hence f is surjective.
40 3. LINEAR MAPS

1.5. Kernel and Image

Definition 3.11. Let f ∈ L(E, F ). We define the kernel of f as:
Ker f = f [−1] {0F } = u ∈ E f (u) = 0F .



We define the image of f as:

Im f = f (E) = f (u); u ∈ E .


Proposition 3.12. Let f ∈ L(E, F ). Then Ker f is a subspace of E and Im f is a subspace of F .

Proof.
• Ker f : let u, v ∈ Ker f and λ, µ ∈ K. Then f (λu + µv) = λ f (u) + µ f (v) = 0F ,
hence λu + µv ∈ Ker f .
• Im f : let u, v ∈ Im f and λ, µ ∈ K. There exists x, y ∈ E such that f (x) = u
and f (y) = v. Now
λu + µv = λ f (x) + µ f (y) = f (λx + µy),
hence λu + µv ∈ Im f .
The following proposition is a paraphrase of the definition of injectivity and surjectivity:

Proposition 3.13. Let f ∈ L(E, F ).

• f is injective if and only if Ker f = {0E }.
• f is surjective if and only if Im f = F .

Proof. We only prove the first point, since the second point is obvious: let f ∈ L(E, F ) be
an injective mapping. Since f is linear, f (0E ) = 0F , and since f is injective,
 
∀x ∈ E, f (x) = 0F =⇒ x = 0E ,

hence Ker f = {0E }. Conversely, let f ∈ L(E, F ) be such that Ker f = {0E }. We now
prove that f is injective: let x, y ∈ E such that f (x) = f (y). Since f is linear, we must
have f (x − y) = 0F , i.e., x − y ∈ Ker f , hence x − y = 0E hence x = y.

Remarks 3.14.
(1) Assume that f ∈ L(E, F ) is an injective map. If we consider f ∈ L E, Im f , then

f is a bijection.
(2) Assume that f ∈ L(E, F ) is a surjective map. By Theorem 2.30 there exists a
subspace G of E such that E = Ker f ⊕ G. Then the restriction of f to G is a
linear bijection from G to F .

Proposition 3.15. Let f ∈ L(E, F ). Then:

• f is injective if and only if there exists g ∈ L(F , E) such that g ◦ f = idE .
• f is surjective if and only if there exists g ∈ L(F , E) such that f ◦ g = idF .

Proof.
• Let g ∈ L(F , E) such that g ◦ f = idE . We must show that f is injective: let u ∈ E
such that f (u) = 0F . Then, u = idE (u) = g ◦ f (u) = g (0F ) = 0E . Conversely, as-
sume that f is injective, and let’s construct a map g ∈ L(F , E) such that g ◦ f = idE :
by Theorem 2.30, there exists a subspace G of F such that F = Im f ⊕ G. Define
 2. RANK OF A LINEAR MAP AND THE RANK–NULLITY THEOREM 41

the mapping g : F −→ E as follows: if v ∈ F , v admits a unique decompo-

sition as v = vF + vG , with vF ∈ Im f and vG ∈ G. Since f is injective, by Re-
mark 3.14(1), there exists a unique vector u ∈ E such that vF = f (u). Set g (v) = u.
We now prove that g is a linear map from F to E: let v, v 0 ∈ F and λ ∈ K, and
decompose v and v 0 on the direct sum Im f ⊕ G: v = vF + vG , v 0 = vF0 + vG0 ,
with vF , vF0 ∈ Im f and vG , vG0 ∈ G, and let u, u 0 ∈ E such that f (u) = vF and
f (u 0 ) = vF0 . Now v + λv 0 = vF + λvF0 + vG + λvG0 , with vF + λvF0 ∈ Im f ,

and f (u + λu 0 ) = vF + λvF0 . Hence, g v + λv 0 = u + λu 0 = g (v) + λg v 0 .

• Let g ∈ L(F , E) such that f ◦ g
= idF . We must show that f is surjective: let v ∈ F .
Clearly, v = idF (v) = f g (v) , hence v is in the image of f . Hence f is surjective.
Conversely, assume that f is surjective, and let’s construct a linear map g ∈ L(F , E)
such that f ◦ g = idF : by Theorem 2.30, there exists
a subspace G of E such that
E = Ker f ⊕ G. Consider the linear map h ∈ L G, F defined by: ∀u ∈ G, h(u) =
f (u). By Remark 3.14(2), the map h is a linear bijection, hence h −1 ∈ L(F , G),
hence, since G is a subspace of E we can consider h −1 as a linear map from F to
E; we denote by g ∈ L(F , E) this linear map. Clearly, for all v ∈ F one has:
f ◦ g (v) = f h −1 (v) = h h −1 (v) = v,

hence f ◦ g = idF .

Here’s a useful fact to determine the image of a linear map:

Proposition 3.16. Let
f ∈ L(E, F ) and let F = (ui )i∈I be a generating family of E. Then the
family f (F ) = f (ui ) i ∈I is a generating family of Im f .

Proof. Let w ∈ Im f , say w = f (v) for some v ∈ E. Since F is a generating family of E,

there exists a family (λi )i ∈I of scalars, with all the λi ’s nil except possibly a finite number
such that
X
v= λi ui .
i∈I
Then, since f is linear,
X
w = f (v) = λi f (ui ) ∈ Span F .
i∈I

2. Rank of a Linear Map and the Rank–Nullity Theorem

Definition 3.17. Let f ∈ L(E, F ). We define the rank of f as the dimension of its image
(in N ∪ {+∞}):
rk f = dim Im f .

Theorem 3.18 (Rank–Nullity Theorem). Let f ∈ L(E, F ). Then

dim E = rk f + dim Ker f .

Proof. By Theorem 2.30 there exists a subspace G of E such that E = G ⊕ Ker f . We

set V = Im f and we consider the linear map
g : G −→ V
u 7−→ f (u).
42 3. LINEAR MAPS

Clearly, g well-defined, linear and surjective. We show that g is an isomorphism:

• g is injective: let u ∈ G such that g (u) = 0V . Then, by the very definition of g ,
f (u) = 0F , hence u ∈ Ker f . Hence u ∈ G ∩ Ker f = {0F }, hence u = 0F = 0V .
• g is surjective: let v ∈ V = Im f . Hence there exists u ∈ E such that v = f (u).
Now u can be uniquely decomposed as the sum of an element in G and an element
in Ker f , say u = a + b with a ∈ G and b ∈ Ker f . Now:
f (u) = f (a + b ) = f (a) + f (b ) = f (a) + 0F = f (a) = g (a) = v.
Hence there exists a ∈ G such that g (a) = u which shows that g is surjective.
Now, by Proposition 3.9, since g is a bijection from G to V , we must have
dim G = dimV = dim Im f = rk f .
Hence, by Grassmann Formula:
dim E = dim(G ⊕ Ker f ) = dim G + dim Ker f = rk f + dim Ker f .
The Rank–Nullity Theorem has a very important (but very simple) consequence for en-
domorphisms:

Theorem 3.19. Let E be a finite-dimensional vector space and f ∈ L(E).

• If f is injective then it is also surjective (hence it is bijective), hence it is an automor-
phism.
• If f is surjective then it is also injective (hence it is bijective), hence it is an automor-
phism.

Proof.
• If f ∈ L(E) is injective, then, since dim Ker f = 0 we have, by the Rank–Nullity
Theorem, dim E = rk f = dim Im f , hence, since dim E < +∞, by the Inclusion–
Equality Theorem 2.55 we must have Im f = E, i.e., f is surjective.
• If f ∈ L(E) is surjective, then, since dim Im f = rk f = dim E we have, by the
Rank–Nullity Theorem, dim E = rk f + dim Ker f , hence dim Ker f = 0, hence
Ker f = {0E }, i.e., f is injective.
It is very important to apply Theorem 3.19 only for finite-dimensional vector spaces, as
shown by the following two counterexamples:

Counterexamples 3.20.
(1) Consider the real vector space E = RN , and consider the unilateral shift operator
T : E → E defined thus: if u = (un )n≥0 ∈ E, the sequence T (u) is defined by:
¨
un−1 if n ≥ 1
T (u)n =
0 if n = 0.
The operator T is clearly an endomorphism of E, and Ker T = {0E } since if
u = (un )n≥0 ∈ E is such that T (u) = 0E , we must have ∀n ∈ N∗ , un−1 = 0, i.e.,
∀n ∈ N, un = 0, i.e., u = 0E . But T is not surjective, as Im T doesn’t contain any
sequence that has a non-zero first term.
(2) Consider the real vector space E = RN , and consider the operator S : E → E
defined thus: if u = (un )n≥0 ∈ E, the sequence S(u) is defined by:
S(u)n = un+1 .
 3. MATRIX OF A LINEAR MAP 43

Clearly, S is an endomorphism of E, and T ◦ S = idE , hence S is surjective. But S

is not injective since, e.g., the sequence u = (un )n≥0 defined by:
¨
1 if n = 0
un =
0 if n ≥ 0
satisfies u 6= 0E and S(u) = 0E .

3. Matrix of a Linear Map

Let f ∈ L(E, F ) and let (ui )i∈I be a basis of E. Every vector x ∈ E can be uniquely
decomposed on the ui ’s:
x = λ1 ui1 + · · · + λn uin .
Now,
f (x) = λ1 f (ui1 ) + · · · + λn f (uin ).
We state the conclusion of this fact as a theorem:

Theorem 3.21. A linear map is uniquely determined by the image of a basis of its domain.

In the finite dimensional case: Let BE = (u1 , . . . , un ) be a basis of E and BF = (v1 , . . . , v m )

be a basis of F . The vectors f (u j ) can be uniquely decomposed on the basis BF as:
f (u j ) = a1 j v1 + · · · + a m j .
The coefficients ai j ’s uniquely determine the linear map f in the bases BE and BF . We
store this information in a big table as
 
a11 · · · a1n
 a21 · · · a2n 
[ f ]BE ,BF = 
 ··· ··· ··· .


a m1 · · · a mn
This table is called the matrix of the linear map f in the bases BE and BF . It is an m × n
matrix, that is, it has m rows and n columns.
Another way of writing this matrix:
[ f ]BE ,BF = [ f (u1 )]BF · · · [ f (un )]BF .

In this chapter we don’t do much with matrices of linear maps, except the following two
procedures: be able to write the matrix of a linear map in bases, and conversely, from a
matrix in bases, be able to recover the associated linear map.

Examples 3.22.
(1) Consider the real vector spaces E = R4 [X ] and F = R2 [X ], and consider the
map ∆ defined by
∆ : E −→ F
P 7−→ P 00 .
Clearly, ∆ is well-defined and linear. Let BE be the standard basis of E and BF be
the standard basis of F . We now compute [∆]BE ,BF . For this, we need to express
the coordinates of ∆(1), . . . , ∆(X 4 ) in the basis BF :
∆(1) = 0, ∆(X ) = 0, ∆ X 2 = 2, ∆ X 3 = 6X , ∆ X 4 = 12X 2 ,




44 3. LINEAR MAPS

hence,
     
0 0 2
∆(1) B = 0 , ∆(X ) B = 0 , ∆ X 2 B = 0 ,
    

F F F
0 0 0
   
0 0
∆ X 3 B = 6 , ∆ X 4 B =  0  ,

 

F F
0 12
hence,  
0 0 2 0 0
[∆]BE ,BF = 0 0 0 6 0  .
0 0 0 0 12
(2) Consider the real vector spaces E = R3 and F = R4 , and consider the linear map
f ∈ L(E, F ), the matrix of which in the standard bases BE and BF is
 
1 2 3
4 5 6
[ f ]BE ,BF = 
 7 8 9 .


10 11 12
Let’s find an expression of f (x, y, z), for (x, y, z) ∈ E: from the matrix [ f ]BE ,BF
we read:
f (1, 0, 0) = (1, 4, 7, 10), f (0, 1, 0) = (2, 5, 8, 11), f (0, 0, 1) = (3, 6, 9, 12).
Now, by linearity of f ,
f (x, y, z) = f x(1, 0, 0) + y(0, 1, 0) + z(0, 0, 1) = x f (1, 0, 0) + y f (0, 1, 0) + z f (0, 0, 1)

= x(1, 4, 7, 10) + y(2, 5, 8, 11) + z(3, 6, 9, 12)

= (x + 2y + 3z, 4x + 5y + 6z, 7x + 8y + 9z, 10x + 11y + 12z).
The following result is straightforward:

Proposition 3.23. Let E and F be two finite dimensional vector spaces, let B be a basis of E
and C be a basis of F . Let f , g ∈ L(E, F ). Then:
[ f ]B,C = [g ]B,C ⇐⇒ f = g .

Proposition 3.24. If E and F are two non null vector spaces over K, then
dim L(E, F ) = dim E dim F .

Proof. Let B = (ui )i∈I be a basis of E and let C = (v j ) j ∈J be a basis of F . For all (i, j ) ∈
I × J , define the linear map fi j ∈ L(E, F ) on the basis B as
¨
v j if k = j
fi j (uk ) =
0F if k 6= i.
Clearly, the family F = ( fi j )(i, j )∈I ×J is an independent family of L(E, F ), which shows
that if either dim E = +∞ or dim F = +∞, then dim L(E, F ) = +∞ = dim E dim F . In
the case dim E < +∞ it follows from the fact that a linear map is uniquely determined
by the image of the vectors of B (and there’s a finite number of such vectors), that F is a
basis of L(E, F ), hence dim L(E, F ) = Card I × Card J = dim E dim F .
 4. POWERS AND POLYNOMIALS OF ENDOMORPHISMS 45

4. Powers and Polynomials of Endomorphisms

Definition 3.25. Let f ∈ L(E). We define the powers of f as:
f 0 = idE , if n ∈ N, f n+1 = f ◦ f n = f n ◦ f .
If f ∈ GL(E), we moreover define the negative powers of f as

n
n ∈ N∗ , f −n = f −1 .
If P ∈ K[X ], say P = an X n + · · · + a1 X + a0 , we define the evaluation of the polynomial P
at f as
P ( f ) = an f n + · · · + a1 f + a0 idE .

Remark 3.26. It should be clear that for all m, n ∈ Z for which f m and f n are defined, we
have
f m+n = f m ◦ f n .
Also, whenever f m and ( f m )n are defined,
( f m )n = f mn .

Proposition 3.27. Let P ∈ K[X ] be such that P (0) 6= 0. If f ∈ L(E) is such that P ( f ) = 0L(E) ,
then f is an automorphism of E. Moreover, f −1 can be expressed as polynomial in f .

Proof. Let P = an X n + · · · + a0 . Since P (0) 6= 0, we must have a0 6= 0. Since P ( f ) = 0L(E) ,

we have:
an f n + · · · + a1 f + a0 idE = 0L(E) ,
hence
f ◦ an f n−1 + · · · + a1 idE + a0 idE = an f n−1 + · · · + a1 idE ◦ f + a0 idE = 0,

hence
   
an n−1 a1 an n−1 a1
f ◦ − f − · · · − idE = − f − · · · − idE ◦ f = idE ,
a0 a0 a0 a0
a a
hence, by Proposition 3.15, f is invertible and f −1 = − n f n−1 − · · · − 1 idE , which is
a0 a0
indeed a polynomial in f .

Example 3.28. Let θ ∈ R. Consider the real vector space E = R2 and the endomorphism
f of E, the matrix of which in the standard basis can = (e1 , e2 ) of E is
cos(θ) − sin(θ)
 ‹
[ f ]can = .
sin(θ) cos(θ)
Let P = X 2 − 2 cos(θ)X + 1. We now prove that P ( f ) = 0L(E) : From the matrix of f in
the basis can of E, we read:
f (e1 ) = cos(θ)e1 + sin(θ)e2 , f (e2 ) = − sin(θ)e1 + cos(θ)e2 .
Hence,
f 2 (e1 ) = f cos(θ)e1 + sin(θ)e2

= cos(θ) f (e1 ) + sin(θ) f (e2 )

= cos(θ) cos(θ)e1 + sin(θ)e2 + sin(θ) − sin(θ)e1 + cos(θ)e2



46 3. LINEAR MAPS

= cos2 θ − sin2 θ e1 + 2 cos θ sin θe2 ,

f 2 (e2 ) = f − sin(θ)e1 + cos(θ)e2

= − sin(θ) f (e1 ) + cos(θ) f (e2 )

= − sin(θ) cos(θ)e1 + sin(θ)e2 + cos(θ) − sin(θ)e1 + cos(θ)e2

= −2 sin θ cos θe1 + cos2 θ − sin2 θ e2 .

Hence:
P ( f )(e1 ) = f 2 (e1 ) − 2 cos(θ) f (e1 ) + e1
= cos2 θ − sin2 θ e1 + 2 cos θ sin θe2 − 2 cos(θ) cos(θ)e1 + sin(θ)e2 + e1

= 0E ,
P ( f )(e2 ) = f 2 (e2 ) − 2 cos(θ) f (e2 ) + e2
= −2 sin θ cos θe1 + cos2 θ − sin2 θ e2 − 2 cos(θ) − sin(θ)e1 + cos(θ)e2 + e2

= 0E .
The endomorphism P ( f ) of E vanishes on a basis of E, hence P ( f ) = 0L(E) . Since P (0) =
1 6= 0, we conclude, by Proposition 3.27, that f is an automorphism. Also:
f −1 = − f + 2 cos(θ) idE ,
hence we conclude:
cos(θ) sin(θ)
 ‹
[ f −1 ]can = .
− sin(θ) cos(θ)

The following result is an analogous of Newton’s formula for real numbers:

Proposition 3.29. If f , g ∈ L(E) such that f ◦ g = g ◦ f (we say that f and g commute),
then for all n ∈ N,
n  
n
X n k n−k
(f + g) = f g .
k=0
k

Proof. By induction on n: for n = 0 the result is obviously true. We assume the result true
for some n ≥ 0 and prove it for n + 1 (using the fact that f and g commute and that the
composition of endomorphisms is associative):
( f + g )n+1 = ( f + g ) ◦ ( f + g )n
‚ n   Œ
X n
k n−k
= (f + g) ◦ f g
k=0
k
‚ n   Œ ‚ n   Œ
X n X n
k n−k k n−k
=f ◦ f g +g f g
k=0
k k=0
k
n   n  
X n k+1 n−k X n k n+1−k
= f g + f g
k=0
k k=0
k
n+1   n  
X n k n+1−k
X n k n+1−k
= f g + f g
k=1
k −1 k=0
k
  n      
n n+1 X n n k n+1−k n n+1
= f + + f g + g
n k=1
k −1 k 0
 5. CODIMENSION OF A SUBSPACE 47

n 
n + 1 k n+1−k
    
n n+1 n n+1 X
= f + g + f g
n 0 k=1
k
n+1 
n + 1 k n+1−k

X
= f g ,
k=0
k

since
n +1
         
n n n n
= = 1, ∀1 ≤ k ≤ n, + = .
0 n k −1 k k

5. Codimension of a Subspace
The following lemma will justify Definition 3.31 below.

Lemma 3.30. Let E be a vector space over K and let F be a subspace of E. All complementary
subspaces of F in E have the same dimension (in N ∪ {+∞}).

Proof. Let G and G 0 be two complementary subspaces of F in E, i.e., E = F ⊕G = F ⊕G 0 .

We explicitly construct an isomorphism g from G to G 0 : let u ∈ G. Since G is a subspace
of E and since the sum F ⊕ G 0 is direct, the vector u possesses a unique decomposition
u = uF + uG 0 with uF ∈ F and uG 0 ∈ G 0 . We set g (u) = uG 0 . Clearly, g is a well-defined
mapping from G to G 0 . We now prove that g is linear: let u, v ∈ G and λ ∈ K, and write

u = uF + uG 0 v = vF + vG 0

with uF , vF ∈ F and uG 0 , vG 0 ∈ G 0 . Clearly, u + λv = (uF + λvF ) + (uG 0 + λvG 0 ) with

uF + λvF ∈ F and uG 0 + λvG 0 ∈ G 0 . Hence, g (u + λv) = uG 0 + λvG 0 = g (u) + λg (v),
and hence g is indeed a linear map from G to G 0 . We now prove that g is injective: let
u ∈ G such that g (u) = 0. With the previous notations, g (u) = uG 0 , hence uG 0 = 0E , hence
u = uF ∈ F . Since the sum F ⊕ G is direct, F ∩ G = {0E }, and since u ∈ F ∩ G we must
have u = 0E , hence g is injective. We now prove that g is onto: let v ∈ G 0 . Decompose v
on F ⊕ G, say v = vF + vG with vF ∈ F and vG ∈ G, and decompose vG on F ⊕ G 0 , say
vG = a + b with a ∈ F and b ∈ G 0 . Clearly, v = vF + vG = vF + a + b , with vF + a ∈ F
and b ∈ G 0 , hence, since v ∈ G 0 and F ⊕ G 0 is a direct sum, vF + a = 0E and b = v. Now,
by definition of g , g (vG ) = b = v. Hence g is an isomorphism from G to G 0 and hence,
by Proposition 3.9, dim G = dim G 0 .

By Theorem 2.30, we know that every subspace F of E possesses a complementary, and

we just showed that they all have the same dimension, hence it makes sense to define the
codimension of a subspace as:

Definition 3.31. Let E be a vector space over K and F a subspace of E. The codimension
of F in E is the dimension of any subspace G of E such that E = F ⊕ G. It is denoted by
codimE F .

The following result is a direct consequence of Grassmann Formula:

Proposition 3.32. Let F be a subspace of E. Then:

dim E = dim F + codimE F .

48 3. LINEAR MAPS

6. Duality
Now that we know enough about linear maps, we develop in more extent the notion of
duality. We recall that E ∗ is the vector space over K of linear forms on E, that is,
E ∗ = L(E, K).
A funny result is the following:

Proposition 3.33. There exists a natural injective linear map

inj : E → (E ∗ )∗
defined thus: if u ∈ E, we set
inj(u) : E ∗ −→ K
ϕ 7−→ ϕ(u).
In particular,
dim E ≤ dim(E ∗ )∗ .

In order to prove Proposition 3.33, there are a few things to check. The details are very
easy and are left to the reader.
6.1. Dual of a Linear Map
Definition 3.34 (Dual of a linear map). Let f ∈ L(E, F ). Then the dual map to f is the
following mapping:
f ∗ : F ∗ −→ E ∗
ϕ 7−→ ϕ ◦ f .

Remark 3.35. Clearly, in the previous definition, the mapping f ∗ is well-defined, since,
according to Proposition 3.10, if ϕ ∈ F ∗ = L(F , K) then ϕ ◦ f ∈ L(E, K) = E ∗ .

Proposition 3.36. Let f ∈ L(E, F ). Then f ∗ ∈ L(F ∗ , E ∗ ).

Proof. Let λ ∈ K and ϕ, ϕ 0 ∈ F ∗ . Then

f ∗ (ϕ + λϕ 0 ) = (ϕ + λϕ 0 ) ◦ f = ϕ ◦ f + λϕ ◦ f = f ∗ (ϕ) + λ f ∗ (ϕ 0 ).
6.2. Dual Family and Dual Basis
Definition 3.37 (Dual family). Let B = (ei )i∈I be a basis of E. We define the dual family
of B as the family B ∗ = (ei∗ )i∈I of vectors of E ∗ with, if i ∈ I , ei∗ is the unique linear map
from E to K defined on the basis B of E as
¨
∗ 1 if i = j
i, j ∈ I , ei (e j ) =
0 if i 6= j .

Remark 3.38. It is also common to use the notation e i instead of ei∗ .

Proposition 3.39. Let B be a basis of E. Then the dual family of B is an independent family
of vectors of E ∗ .

Proof. Let B = (ei )i∈I be a basis of E and let B ∗ = (ei∗ )i∈I be the dual basis of B. Let
(λi )i∈I be a family of elements of K that are all nil except possibly a finite number and such
 6. DUALITY 49

that X
λi ei∗ = 0E ∗ .
i∈I
Let j ∈ J and let’s show that λ j = 0: clearly,
‚ Œ
X X
λj = λi ei∗ (e j ) = λi ei∗ (e j ) = 0.
i ∈I i∈I
∗
Hence B is an independent family of vectors.
As a direct corollary, we have

Corollary 3.40.
dim E ≤ dim E ∗ .

If E is infinite-dimensional and if B = (ei )i∈I is a basis of E, we cannot expect the dual

family of B to be a basis of E ∗ , since if we consider ϕ ∈ E ∗ defined on B as
∀i ∈ I , ϕ(ei ) = 1,
then ϕ cannot be decomposed on B ∗ . Indeed, assume that ϕ is a linear combination of
elements of B ∗ , say
X
ϕ= λi ei∗
i ∈I
with (λi )i ∈I a family of elements of K that are all nil except a finite number. Then there
exists j ∈ I such that λ j = 0, hence we would have
‚ Œ
X
1 = ϕ(e j ) = λi ei∗ (e j ) = λ j e ∗j (e j ) = 0
i∈I

which is impossible.
In the finite-dimensional case, we do have

Proposition 3.41. Let E be a finite-dimensional vector space and let B be a basis of E. Then
B ∗ is a basis of E ∗ . In particular,
dim E = dim E ∗ .

Proof. Let B = (e1 , . . . , en ) be a basis of the finite-dimensional vector space E. We already

know that B ∗ is an independent family of vectors of E. We show that B ∗ is a generating
family of E ∗ : let ϕ ∈ E ∗ . Clearly,
n
X
ϕ= ϕ(ei )ei∗
i=1

Remarks 3.42.
(1) In fact, the previous proof relies on the following fact: if B = (ei )i∈I is a basis of
E, and if B ∗ = (ei∗ )i∈I is the dual family of B, then for all u ∈ E we have:
X
u= ei∗ (u)ei .
i∈I
50 3. LINEAR MAPS

(2) If E is a finite-dimensional vector space and if B is a basis of E, then B ∗ is called

the dual basis of B, since it is a basis of E ∗ .

Exercises
Exercise 3.1. In each of the following cases, is the map f : E → F a linear map? If it is
the case, specify if it is injective, surjective, bijective, and determine (if possible) its image
and kernel (and, if possible, a basis of its kernel and image). In the finite dimensional case,
write its matrix in the standard bases.
(1) E = R2 , F = R, f (x, y) = 3x y,
(2) E = C4 , F = C, f (x, y, z, t ) = −y,
2 2
(3) E = C , F =C , f (x, y) = 2i x, (1 + i)y + 3x ,

(4) E = RN , F = R, f (un ) = u0 + 4u5 ,

(5) E = R3 , F = R2 , f (x, y, z) = (2x + y − z, x 2 ),

(6) E = C[X ], F = C, f (P ) = P (5) + 5i,
(7) E = F = R[X ], f (P ) = X 2 P 00 (X ) + 3P (2),
Z1
2
(8) C [0, 1], R , F = R, f (g ) = e t g (t ) dt .


0

Exercise 3.2. We consider the following linear maps:

f : R2 −→ R2 g : R2 −→ R3
(x, y) 7−→ (3x + 6y, x + 2y), (x, y) 7−→ (x − 3y, 2x + y, 4x + 2y),
h : R2 −→ R2 k : R3 −→ R2
(x, y) 7−→ (3x − 4y, x − y), (x, y, z) 7−→ (x + y − z, 2x − y),
` : C3 −→ C3 m : R3 [X ] −→ R2
(x, y, z) 7−→ z, −i x, y + (2 + i)z , 7−→ P (0), P 0 (0) .



P
For each of them:
1. Determine its kernel and image (and give a basis of each), and write its matrix in the
standard bases.
2. Is the map injective? Surjective? What is its rank?
3. If the map is bijective, determine its inverse.

Exercise 3.3.
1. Justify the existence of a unique linear map f : R3 → R2 such that:
f (1, 0, 0) = (0, 1), f (1, 1, 0) = (2, 0), f (1, 1, 1) = (1, 1).
2. For all (x, y, z) ∈ R3 , determine f (x, y, z) and deduce its matrix in the standard bases.
3. Determine the image and the kernel of f .

Exercise 3.4. We consider the real vector space E = C ∞ (R). We define the map ∆:
∆ : E −→ E
f 7−→ f 00 .
1. Show that ∆ is well-defined and that it is an endomorphism of E.
2. Determine Ker ∆, Im ∆ and their dimension.
 EXERCISES 51

Exercise 3.5. Let E be a vector space over K and f ∈ L(E). Prove the following equivalence:
Im f = Im f 2 ⇐⇒ E = Ker f + Im f .

Exercise 3.6 (Projections). Let E be a vector space over K and p ∈ L(E) such that p 2 = p
(we say that p is a projection of E).
1. Show that q = id − p is also a projection of E and that Im p = Ker q and Im q = Ker p.
2. Show that E = Im p ⊕ Ker p. Is this consistent with the Rank–Nullity Theorem?
3. In E = R2 we consider the map
p : R2 −→ R2
(x, y) 7−→ (4x − 6y, 2x − 3y).
a) Show that p is a projection of R2 .
b) Explicit q = id − p and check that it is indeed a projection of E.
c) Determine a basis of Ker p and a basis of Im p; check on this example that Im p =
Ker q and R2 = Im p ⊕ Ker p.

Exercise 3.7. Let n ∈ N. We consider the linear map

u : Rn [X ] −→ Rn [X ]
P 7−→ P (1).
1. Is the map u a projection of Rn [X ] ?
2. Compute the rank of u. Describe Ker u, and determine a complementary subspace of
Ker u in Rn [X ].

Exercise 3.8. We consider the map f : R4 [X ] → R4 [X ] defined as f (P ) = P + (1 − X )P 0 .

1. Show that f is well defined and that it is a linear map. Determine the kernel and the
image of f .
2. Show that R4 [X ] = Ker f ⊕ Im f .
3. Is f a projection of R4 [X ] ?

Exercise 3.9 (Symmetries). Let E be a vector space over R and s ∈ L(E) such that s 2 = id
(we say that s is a symmetry of E). We set
E+ = x ∈ E s(x) = x , E− = x ∈ E s(x) = −x .
 

Show that E = E+ ⊕ E− . Express E+ and E− as a kernel of a linear map.

Exercise 3.10. We consider the real vector space E = R2 . Let u = (1, −2). We define the
following subspaces of E:
F = (x, y) ∈ R2 x − y = 0 , G = Span(u).


1. Show that F and G are complementary subspaces in E.

2. Let v = (x, y). Determine the projection p(v) of v on F along G, as well as the sym-
metric s(v) of v with respect to F along G.
3. Same questions, this time in E = R3 with u = (0, 1, 0) and
F = (x, y, z) ∈ R3 x + y − z = 0 .


Exercise 3.11. Let E be a vector space over the commutative field K and let u ∈ L(E). Show
that Im u 2 ⊂ Im u and that Ker u ⊂ Ker u 2 . Find an example where the equality doesn’t
hold.
52 3. LINEAR MAPS

Exercise 3.12. Let E be a vector space over K and u ∈ L(E). For k ∈ K we set:
Fk = x ∈ E u(x) = k x .


1. Show that Fk is a subspace of E. What can you say about F0 ?

2. Show that if k 6= k 0 then Fk ∩ Fk 0 = {0}.
3. We consider the following endomorphism u of the real vector space E = R2 :
u : R2 −→ R2
(x, y) 7−→ (x − 2y, −2x + y).
a) Show that F3 6= {0} and determine a basis of F3 .
b) Show that there exists a unique k ∈ R∗ , k 6= 3, such that Fk 6= {0} and determine a
basis of Fk .
c) Are the subspaces Fk and F3 complementary subspaces of R2 ?

Exercise 3.13. In this exercise, C is considered as a real vector space.

1. Let f be the function from C to C defined as f (z) = z + i z. Show that f is a linear map
and determine its kernel and its image.
2. Show that all (real) endomorphism of C has the form f : z 7→ c z + d z for some c, d ∈ C.
What is dimR L(C)?

Exercise 3.14. Let n ∈ N∗ and let a1 , . . . , an be distinct real numbers. We consider the map:
f : Rn−1 [X ] −→ Rn
7−→ P (a1 ), . . . , P (an ) .


P
1. Show that f is an injective linear map. What can you deduce?
2. Give an interpretation of this result in terms of interpolation.

Exercise 3.15 (From Term-Test, May 2011). Let a, b , c and d be four real numbers, not all
zero, and consider the mapping ϕ defined by
ϕ : R4 −→ R
(x1 , x2 , x3 , x4 ) 7−→ a x1 + b x2 + c x3 + d x4 .
1. Prove that ϕ is a linear form on R4 .
2. What is the rank of ϕ?
3. Deduce the dimension of the kernel of ϕ.
4. Let
H1 = (x1 , x2 , x3 , x4 ) ∈ R4 x1 − x2 = 0 , H2 = (x1 , x2 , x3 , x4 ) ∈ R4 x3 − x4 = 0 .
 

a) Prove that H1 and H2 are two subspaces of R4 , and give their dimensions. (Hint: You
may use the previous questions, or proceed directly)
b) Give a basis of H1 ∩ H2 and give the dimension of H1 ∩ H2 .
c) Find a complementary subspace of H1 ∩ H2 in R4 . What is its dimension?

Exercise 3.16 (From Test #5, May 2012). Let a ∈ R. We consider the mapping
fa : R3 [X ] −→ R3 [X ]
P 7−→ P (X − 1) + aP (X ).
1. Show that fa is a well-defined endomorphism of R3 [X ].
2. Write the matrix [ fa ]B of fa in the standard basis B = 1, X , X 2 , X 3 of R3 [X ].

3. Determine the rank of fa in terms of a.

 EXERCISES 53

4. We now consider the special case a = −1. We set g = f−1 .

a) Determine a basis of Im(g ) and of Ker(g ).
b) Are the spaces Im(g ) and Ker(g ) complementary subspaces of R3 [X ]?
 4
Matrices

1. Computations with Matrices

1.1. Interpretation of a Matrix and Notations
Recall that an m × n matrix  
a11 ... a1n
 .. .. .. 
A=  . . . 
a m1 . . . a mn
has the following interpretation:
Let E be a vector space of dimension n and F a vector space of dimension m. Given a
basis B = (u1 , . . . , un ) of E and a basis C = (v1 , . . . , v m ) of F we can construct a unique
linear map f ∈ L(E, F ) such that
∀i ∈ {1, . . . , n}, f (ui ) = a1i v1 + · · · + a mi v m .
We wrote this fact as: [ f ]B,C = A. Observe that the i -th column of A,
 
a1i
 .. 
 . 
a mi
represents the coordinates of the vector f (ui ) in the basis C :
 
a1i
 .. 
[ f (ui )]C =  . 
a mi
The coefficients ai j ’s are called the coefficients of the matrix A.
Notation: the set of m × n matrices with coefficients in K is denoted by M mn (K). In the
particular case when m = n we just write M n (K), and we say that an n × n matrix is a
square matrix of size n.
Special notation for endomorphisms: let f ∈ L(E) and B a basis of E. We write [ f ]B
instead of [ f ]B,B .
55
56 4. MATRICES

We also write A = (ai j )i=1..m , or just A = (ai j ) when no confusion is possible.

j =1..n

1.2. Sum of Matrices

Definition 4.1. Let A = (ai j )i =1..m and B = (bi j )i =1..m be two m × n matrices. Their sum,
j =1..n j =1..n
denoted by C = A + B is the matrix C = (ai j + bi j )i=1..m , or more explicitly:
j =i ..n

a11 + b11 . . . a1n + b1n

     
a11 ... a1n b11 ... b1n
 .. .. ..  +  .. .. ..  = C =  .. .. ..
A+ B =  . .

. .   . . .   . . .
a m1 . . . a mn b m1 . . . b mn a m1 + b m1 . . . a mn + b mn

Remark 4.2. It only makes sense to compute the sum of two matrices that have the same
size. It doesn’t make any sense whatsoever to add a 3 × 4 matrix with a 4 × 3 matrix.

Proposition 4.3. Let E and F be two finite-dimensional vector spaces over K, and let B be a
basis of E and C be a basis of F . Let f and g in L(E, F ). Then:
[ f + g ]B,C = [ f ]B,C + [g ]B,C .

Proof. Set B = (u1 , . . . , un ) and C = (v1 , . . . , v m ). Let A = (ai j ) = [ f ]B,C and B = (bi j ) =
[g ]B,C . Let i ∈ {1, . . . , n}. The i-th column of [ f + g ]B,C is, by definition, ( f + g )(ui ) C .
 

Now, since
f (ui ) = a1i v1 + · · · + a mi v m , g (ui ) = b1i v1 + · · · + b mi v m ,

we have
( f + g )(ui ) = f (ui ) + g (ui ) = (a1i + b1i )v1 + · · · + (a mi + b mi )v m ,
hence,
a1i + b1i
 

( f + g )(ui ) C = 
 ..
.
 
.
a mi + b mi
Hence, the i-th column of [ f + g ]B,C exactly corresponds to the i-th column of [ f ]B,C +
[g ]B,C as we defined it.
1.3. Multiplication by a Scalar
Definition 4.4. Let A = (ai j )i=1..m be an m × n matrix with coefficients in K. Let λ ∈ K.
j =1..n
The scalar multiplication of A by λ is the following m × n matrix:

λa11 . . . λa1n
 

λA = (λai j ) =  ... .. ..  .

. . 
λa m1 . . . λa mn

Proposition 4.5. Let E and F be two finite-dimensional vector spaces over K, and let B be a
basis of E and C be a basis of F . Then
[λ f ]B,C = λ[ f ]B,C .
 1. COMPUTATIONS WITH MATRICES 57

Proof. Set B = (u1 , . . . , un ) and C = (v1 , . . . , v m ). Let A = (ai j ) = [ f ]B,C , and let i ∈
{1, . . . , n}. The i-th column of [λ f ]B,C is, by definition, λ f (ui ) C . Now, since
 

f (ui ) = a1i v1 + · · · + a mi v m ,
we have
(λ f )(ui ) = λ · f (ui ) = λa1i v1 + · · · + λa mi v m ,
hence
λa1i
 

(λ f )(ui ) C =  ...  .
   
λa mi
Hence, the i-th column of [λ f ]B,C exactly corresponds to the i-th column of λ[ f ]B,C
as we defined it.
1.4. Vector Space Structure on M mn (K)
From the previous two results, we conclude that we have constructed a vector space over
K structure on M mn (K), that is compatible with the vector space structure on the set of
the linear maps. More precisely:

Proposition 4.6. With the usual addition and multiplication by a scalar, the set M mn (K) is
a vector space over K. Moreover, if E and F are vector spaces over K with dim E = n and
dim F = m, if B is a basis of E and C is a basis of F , then the map
Φ : L(E, F ) −→ M mn (K)
f 7−→ [ f ]B,C
is an isomorphism.

The standard basis of M mn (K) is the basis B = Ei j , i ∈ {1, . . . , m}, j ∈ {1, . . . , n} where

Ei j ∈ M mn (K) is the matrix that has null entries, except its i j entry that is 1:
j
 ↓ 
0
. ··· 0 ··· 0
 .. .. .. 
··· . ··· .
Ei j = .
 
i → 0 ··· 1 ··· 0

. .. .. 
 .. ··· . ··· .
0 ··· 0 ··· 0
It should be clear that B us a basis of M mn (K). As a consequence, we obtain:
dim M mn (K) = m × n.
(Compare Proposition 3.24 and Proposition 4.6).
1.5. Product of Matrices
Definition 4.7. Let A = (ai j ) ∈ M mn (K) and B = (bi j ) ∈ M n` (K). The product of A and B
is the matrix C = (ci j ) ∈ M m` (K) defined by:
n
X
∀i ∈ {1, . . . , m}, ∀ j ∈ {1, . . . , `}, ci j = ai k bk j .
k=1
58 4. MATRICES

Proposition 4.8. Let E, F and G be three vector spaces of dimension n, m, ` respectively and
let
B = (u1 , . . . , un ), C = (v1 , . . . , v m ), D = (w1 , . . . , w` )
be bases of E, F and G respectively. Let g ∈ L(E, F ), f ∈ L(F , G) such that [ f ]C ,D = A = (ai j )
and [g ]B,C = B = (bi j ). Then
[ f ◦ g ]B,D = AB = [ f ]C ,D [g ]B,C .

Proof. Let j ∈ {1, . . . , n}. The j -th column of the matrix [ f ◦ g ]B,D is
( f ◦ g )(u j ) D .
 

Now,
g (u j ) = b1 j v1 + · · · + b m j v m
and for all k ∈ {1, . . . , m},
X̀
f (vk ) = a1k w1 + · · · + a`k w` = ai k wi .
i=1
Hence:
m
X
( f ◦ g )(u j ) = b1 j f (v1 ) + · · · + b m j f (v m ) = bk j f (vk )
k=1
m
X X̀
= bk j ai k wi
k=1 i=1
‚ m Œ
X̀ X
= ai k bk j wi .
i=1 k=1

Hence the j -th column of the matrix [ f ◦ g ]B,D is

X m 
 a1k bk j 
 k=1 
[ f ◦ g ]B,D = 
 .
.

,
.


Xm 
a mk bk j
 
k=1

which exactly corresponds to the j -th column of [ f ]C ,D [g ]B,C = AB as we defined it.

The proof of the following proposition is left to the reader:

Proposition 4.9. Let E and F be two finite-dimensional vector spaces. Let f ∈ L(E, F ) and
let x ∈ E. Let B be a basis of E and let C be a basis of F . Then:
f (x) C = [ f ]B,C [x]B .
 

Remarks 4.10.
(1) In order to multiply two matrices, the number of columns of the first has to be
equal to the number of rows of the second. Then, the number of rows of the
product is the number of rows of the first, and the number of columns of the
product is the number of columns of the second.
 1. COMPUTATIONS WITH MATRICES 59

(2) The product of matrices is not commutative: in general, AB 6= BA. Actually, it

can happen that, e.g., the product AB is defined, but not the product BA.
(3) If A, B ∈ M n (K),
(A + B)2 = A2 + AB + BA + B 2 .
If the matrices A and B commute (that is, AB = BA) then (and only then), we have
(A + B)2 = A2 + 2AB + B 2 .
(4) The product of matrices is associative: A(BC ) = (AB)C .
(5) The result of Remark 4.10(3) can be generalized to obtain the Binomial Theorem
for commuting matrices: if A, B ∈ M n (K) commute, i.e., AB = BA, then for all
n ∈ N:
n  
n
X n k n−k
(A + B) = A B .
k=0
k
The proof of this fact is the same as the proof of the Binomial Theorem for real
or complex numbers.
In the remaining part of this section we discuss some very interesting properties to keep
in mind about products of matrices.
1.5.1. Extracting a Column from a Matrix
Let A ∈ M mn (K). For all i ∈ {1, . . . , n}, denote by Ei the following column matrix in
M n1 (K):
 
0
 .. 
.
 
 0
Ei = 
 
1

 0
 
.
 .. 
0
where the “1” is at the i-th position. It follows from the very definition of the product of
matrices that the i-th column of A is equal to the product:
AEi .
Notice how this is consistent with Proposition 4.9: let B = (u1 , . . . , un ) be a basis of a
vector space E of dimension n and let C = (v1 , . . . , v m ) be a basis of a vector space F of
dimension m. Let f ∈ L(E, F ) be the linear mapping from E to F such that [ f ]B,C = A.
Observe that for all i ∈ {1, . . . , n}, [ui ]B = Ei . Then, we know that the i-th column of A
is f (ui ) C , and from Proposition 4.9 we indeed have, for all i ∈ {1, . . . , n},


f (ui ) C = [ f ]B,C [ui ]B = AEi .

 

1.5.2. Extracting a Row from a Matrix

Let A ∈ M mn (K). For all j ∈ {1, . . . , m}, denote by R j the following row matrix in M1m (K):
Rj = 0 · · · 0 1 0 · · · 0 .

From the very definition of the product of matrices, we observe that the j -th row of A is
obtained by the following product:
R j A.
60 4. MATRICES

1.5.3. Product of Matrices as a Column-Wise Operation

Let A ∈ M mn (K) and let B ∈ M n p (K) (so that the product AB is well-defined and is an
element of M m p (K)). Denote by B1 , . . . , B p the columns of B:
 
| | ··· |
B = B1 B2 · · · B p  .
| | ··· |
Then the product AB can be understood as the elements AB1 , . . . , AB p stacked next to each
other:    
| | ··· | | | ··· |
AB = A B1 B2 · · · B p  = AB1 AB2 · · · AB p  .
| | ··· | | | ··· |
In Chapter 6 or when we cover change of basis matrices below, this point of view can be
enlightening.
1.6. Identity Matrix
Definition 4.11. The identity matrix in M n (K) is the matrix
 
1 0 0 ... 0
0 1 0 . . . 0
 
I n = 0 0 1 . . . 0 .
 
 .. .. .. .. .. 
. . . . .
0 0 0 ... 1

The following result is straightforward:

Proposition 4.12. If A ∈ M mn (K), then AIn = I m A = A.

Remark 4.13. The identity matrix is always a square matrix. It corresponds to the matrix
of the identity endomorphism in any basis (i.e., when you take the same basis for the source
space and the target space).

Caveat 4.14. Let E = R2 , B = (1, 0), (0, 1) and C = (1, 1), (1, −1) . Let’s compute

A = [idE ]B,C : its first column is

1/2
 ‹
idE (1, 0) C = (1, 0) C =
   
.
1/2
Its second column is
1/2
 ‹
idE (0, 1) C = (0, 1) C =
   
.
−1/2
Hence,
1/2 1/2
 ‹
A = [idE ]B,C = .
1/2 −1/2
The matrix A is not equal to the identity matrix I2 , since we didn’t choose the same bases
for the source and the target. But of course, we have:
1 0
 ‹
[idE ]B = [idE ]C = I2 = .
0 1
 1. COMPUTATIONS WITH MATRICES 61

1.7. Invertible Matrix — Rank of a Matrix

Proposition 4.15. Let A ∈ M n (K). Then the following two conditions are equivalent:
• There exists B ∈ M n (K) such that AB = In .
• There exists C ∈ M n (K) such that C A = In .
Moreover, if there exists matrices B, C ∈ M n (K) such that AB = C A = In , then B = C , and
such a matrix B is unique.

Proof. Let E = Kn and let B be the standard basis of E, and let f be the endomorphism
of E such that [ f ]B = A. If there exists B ∈ M n (K) such that AB = In , let g be the
endomorphism of E such that [g ]B = B. Clearly, [ f ◦ g ]B = AB = In = [idE ]B , hence
f ◦ g = idE , hence g = f −1 , hence g ◦ f = idE , hence BA = In .

Definition 4.16. Let A ∈ M n (K). We say that A is invertible if there exists a matrix B ∈
M n (K) such that AB = BA = I . This matrix B is called the inverse of A, and it is denoted
by A−1 .

Remarks 4.17.
(1) It only makes sense to talk about the inverse of a square matrix.
(2) If A is invertible, then so is A−1 , and A−1 )−1 = A.
The following proposition is now straightforward:

Proposition 4.18. Let E and F be two finite-dimensional vector spaces over K, let B be a
basis of E and C be a basis fo F . Let f ∈ L(E, F ) be an isomorphism from E to F . Then:
 −1 
f C ,B
= [ f ]−1
B,C
.

Definition 4.19. The rank of a matrix, denoted by rk A, is the rank of the associated linear
map in any bases.

Remark 4.20. A way to compute the rank of A is to compute the rank of the columns of A
(using the algorithm we used for computing the rank of a family of vectors).

Proposition 4.21. Let A ∈ M n (K). Then A is invertible if and only if its rank is n.

Proof. Let E = Kn and let B be any basis of E and let f ∈ L(E) such that [ f ]B = A. Then
A is invertible if and only if f is invertible, if and only if rk f = rk A = n.

Caveat 4.22. Let A, B, C ∈ M n (K). Then

AC = BC =⇒
6 A = B.

The result is the following: let A, B, C ∈ M n (K) such that C is invertible. Then

AC = BC =⇒ A = B.

The proof is easy: just multiply by C −1 :

AC = BC =⇒ AC C −1 = BC C −1 =⇒ AI = B I =⇒ A = C .
62 4. MATRICES

1.8. Computing the Inverse of a Matrix

The best way to compute the inverse of a square matrix is to solve a linear system. We
illustrate the procedure on a simple example:

Example 4.23. Compute the inverse of

2 1
 ‹
A= .
3 2
(1) Write the associated system with a general right-hand side:
2x + y = a
§
3x + 2y = b .
(2) Solve the system (there should be a unique solution):
y + 2x = a y + 2x = a y = a + 2b − 4a = −3a + 2b
§ § §
⇐⇒ ⇐⇒
2y + 3x = b − x = b − 2a x = −b + 2a
(3) Write this solution as a system (putting all the variables in the correct order):
2a − b = x
§
−3a + 2b = y.
(4) Since the system possesses a unique solution, we conclude that A is invertible and
that the inverse of A is the matrix:
2 −1
 ‹
−1
A =
−3 2

1.9. Transposition and Symmetric Matrices

Definition 4.24. Let A = (ai j ) ∈ M mn (K). The transpose of the matrix A is the matrix
denoted by t A ∈ M n m (K) obtained by exchanging the rows and the columns of A:
t
A = (a j i ) ∈ M n m (K).
A matrix A ∈ M n (K) is said to be symmetric if t A = A (its entries are symmetric with
respect to the diagonal).

Proposition 4.25. Let A ∈ M mn (K) and B ∈ M n` (K). Then:

t
(AB) = t B t A.

Proof. Let A = (ai j ), B = (bi j ) and C = (ci j ) = t B t A ∈ M`m (K). For all i ∈ {1, . . . , `} and
j ∈ {1, . . . , m},
n
X Xn
ci j = bk i a j k = a j k bk i ,
k=0 k=0
hence C = t (AB).

2. Change of Basis Matrix

2.1. Definition
Definition 4.26. Let E be a vector space over K and B = (u1 , . . . , un ) be a basis of E.
 2. CHANGE OF BASIS MATRIX 63

• If F = (v1 , . . . , v p ) is a family of p vectors of E, the coordinate matrix of F in the

basis B is the matrix denoted by [F ]B ∈ M n p (K) that consists of the coordinates
of the vectors of F in the basis B, i.e., the i-th column of [F ]B is [vi ]B .
• If C = (v1 , . . . , vn ) is another basis of E, the change of basis matrix from the basis B
to the basis C is the matrix [C ]B ∈ M n (K).

Remark 4.27. The change of basis matrix [C ]B is also written as PB→C .

The following result result is straightforward:

Proposition 4.28. Let B be a basis of the finite-dimensional vector space E over K and let C
be a finite family of vectors of E. Then the family C is a basis of E if and only if [C ]B is square
matrix that is invertible. In this case, [C ]B is the change of basis matrix from the basis B to
the basis C and
[C ]B = [idE ]C ,B

Proposition 4.29. A change of basis matrix is always invertible, and:

−1
[C ]B = [B]C .

Proof. Clearly,

−1
−1
[C ]B = [idE ]C ,B = id−1
E B,C = [idE ]B,C = [B]C .
 

Proposition 4.30. Let E be a finite-dimensional vector space over K and let B and C be two
bases of E. Then for all x ∈ E,
[x]C = [B]C [x]B .

Proof. Let x ∈ E. Then:

[x]C = idE (x) C = [idE ]B,C [x]B = [B]C [x]B .
 

2.2. Change of Bases of Linear Maps

Proposition 4.31. Let E and F be two finite-dimensional vector spaces, let B and B 0 be two
bases of E and let C and C 0 be two bases of F . Let f ∈ L(E, F ). Then:
[ f ]B 0 ,C 0 = [C ]C 0 [ f ]B,C [B 0 ]B .

Proof. Clearly,
[ f ]B 0 ,C 0 = [idE ◦ f ◦ idE ]B 0 ,C 0 = [idE ]C ,C 0 [ f ]B,C [idE ]B 0 ,B = [C ]C 0 [ f ]B,C [B 0 ]B .

Remark 4.32. An alternate way of writing the result of Proposition 4.31 is:
[ f ]B 0 ,C 0 = PC 0 →C [ f ]B,C PB→B 0 .

The following proposition is a special case of Proposition 4.31 for endomorphisms:

Proposition 4.33. Let E be a finite-dimensional vector space over K and let B and C be
two bases of E. Let f ∈ L(E). Then:

−1
[ f ]C = [B]C [ f ]B [C ]B = [C ]B [ f ]B [C ]B .
64 4. MATRICES

Remark 4.34. The result of Proposition 4.33 is sometimes written thus: let
P = PB→C = [C ]B .
Then
[ f ]C = P −1 [ f ]B P.

2.3. Similar Matrices

Definition 4.35. Let A and B be matrices in M n (K). We say that A and B are similar
matrices if there exists an invertible matrix P ∈ M n (K) such that B = P −1 AP .

In other words, A and B are similar if they represent the same endomorphism in (possibly)
different bases.
The notion of similar matrices will be important in Chapter 6.

3. Link with Linear Systems

Consider the following m × n linear system with coefficients in K
a11 x1 + a12 x2 + · · · + a1n xn = b1

 a21 x1 + a22 x2 + · · · + a2n xn = b2


(S) .. .. .. .
 .

 . ··· . = ..
a m1 x1 + a m2 x2 + · · · + a mn xn = b m ,
denote by A = (ai j ) ∈ M mn (K) the associated matrix and set
   
b1 x1
 ..   .. 
B =  .  ∈ M m1 (K), X =  .  ∈ M n1 (K).
bm xn
Clearly, System (S) is equivalent to AX = B. The rank of System (S) is exactly the rank
of A.
We now exhibit the elementary row operations using matrix notations. We first introduce
the elementary matrices:

Definition 4.36 (Elementary matrices). Let n ∈ N∗ .

• The row switching transformation matrix is the square matrix obtained by exchang-
ing row i and j (i 6= j ) of the identity matrix:
 
1
 ..
.

 
 

 0 1 

Ti j = 
 . ..  ∈ M n (K).

 

 1 0 


 . .. 


1
In terms of the standard basis (Ei j ) of M n (K),
Ti j = In − Ei i − E j j + Ei j + E j i .
Such matrices are also known as transposition matrices.
 3. LINK WITH LINEAR SYSTEMS 65

• The row multiplication transformation matrix is the matrix obtained from the
identity matrix by changing the i-th diagonal term to λ ∈ K, λ 6= 0:
 
1
 ..
.

 
Di (λ) =  λ  ∈ M (K).
 

 n
 .. 
 . 
1
In terms of the standard basis,
Di (λ) = In + (λ − 1)Ei i .
Such matrices are also known as dilation matrices.
• The row addition transformation matrix is the matrix (with i 6= j ):
 
1
 ..
.

 
 

 1 

Si j (λ) = 
 . ..  ∈ M n (K).

λ
 

 1 

 .. 
 . 
1
In terms of the standard basis,
Si j (λ) = In + λEi j .
Such matrices are also known as transvections or shear transformations.
An elementary matrix is a square matrix that is either a transposition matrix, a dilation
matrix or a transvection matrix.

The following result is straightforward:

Proposition 4.37. Every elementary matrix is invertible.

Proof. Clearly,
Ti−1
j
= Ti j
1
 ‹
−1
∀λ 6= 0, Di (λ) = Di
λ
−1
∀i 6= j , Si j (λ) = Si j (−λ).

The elementary operations on System (S) have their counterpart on the matrices A and B:
• The linear system obtained from (S) after performing the row switching operation
Ri ↔ R j on (S) is the linear system associated with the matrix Ti j A, and with
right-hand side Ti j B.
• The linear system obtained from (S) after performing the row multiplication
operation Ri ← λRi on (S) (with λ = 6 0) is the linear system associated with the
matrix Di (λ)A, and with right-hand side Di (λ)B.
66 4. MATRICES

• The linear system obtained from (S) after performing the row addition operation
Ri ← Ri + λR j on (S) (with λ ∈ K and i 6= j ) is the linear system associated with
the matrix Si j (λ)A, and with right-hand side Si j (λ)B.
The following theorem is a corollary of the observations above, and of the Gaussian elimi-
nation. It will be helpful in Chapter 5 to prove multiplicativity of the determinant.

Theorem 4.38. Let A ∈ M n (K) be a matrix of rank k. Then there exists a matrix B that is
a product of elementary matrices and there exists a matrix R with all its coefficients equal to
0, except k terms equal to 1 on its diagonal such that A = B R. In particular, if A is invertible
(i.e., rk A = n), then A can be written as a product of elementary matrices.

Exercises
 
1 −2 −6
Exercise 4.1. We consider the matrix A = −3 2 9 .
2 0 −3
1. Compute A2 and A3 .
2. Let n ≥ 1. Deduce an expression for An .
3. Is the matrix A invertible?
 
1 1 1 1
1 1 −1 −1
Exercise 4.2. We consider the matrix B =  2
1 −1 1 −1. Compute B and deduce


1 −1 −1 1
−1 n
that B is invertible. Determine B and B for n ∈ Z.
 
3 1 −1
Exercise 4.3. We consider the matrix A = 1 1 1 . Show that A is invertible and
2 0 2
determine its inverse.
 
2 4 6
Exercise 4.4. We set A = 0 2 3 and N = A − 2I3 . Compute the powers of N and
0 0 2
n
deduce the value of A for n ∈ N.

Exercise 4.5. Let E be a real vector space with basis B = (e1 , e2 ) and F a real vector space
with basis C = ( f1 , f2 , f3 ).
1. Show that there exists a unique linear map ϕ ∈ L(E, F ) such that:

ϕ(e1 ) = 2 f2 − 3 f3 , ϕ(e2 ) = f1 − f2 + f3 .

2. Determine [ϕ]B,C .
x
 ‹
3. Let x ∈ E such that its coordinates in the basis B are 1 . Determine the coordinates
x2
 
y1
y2  of ϕ(x) in the basis C .
y3
4. Determine Ker ϕ and Im ϕ. Is the map ϕ injective? Surjective? Bijective?
 EXERCISES 67

Exercise 4.6. Let f and g be the two linear maps defined by:
f : R2 −→ R3 g : R3 −→ R2
(x, y) 7−→ (x + y, x − y, 2x), (x, y, z) 7−→ (2y − z, x + y + z).
1. Determine the matrices [ f ]B,C and [g ]C ,B in the following cases:
a) B and C are respectively the standard bases of R2 and R3 .
b) B = (1, 1), (1, −1) and C is the standard basis of R3 .

c) B = (1, 1), (1, −1) and C = (1, 1, 1), (0, 1, 1), (0, 0, 1) .

2. Determine the expression of f ◦ g using the composition of the maps f and g . Recover
this result using the product of matrices.
3. Same question for g ◦ f .

Exercise 4.7. We consider the real vector space E = R3 [X ] and the map f : E → E defined
by f (P ) = P 0 − P .
1. Show that f is a well-defined endomorphism of E.
2. Determine Ker f and Im f .
3. What is the matrix M of f in the standard basis of E? Show that M is invertible.

Exercise 4.8. Let E be a real vector space of dimension 3 and B = (e1 , e2 , e3) a basis of E. Let

1 −1 2
f be the endomorphism of E the matrix of which in the basis B is M = −2 1 −3.
−1 1 −2
1. Determine a basis of Ker f and Im f .
2. We set u = e1 + e2 .
a) Show that B 0 = u, f (u), f 2 (u) is a basis of E.

b) What is the matrix T of f in the basis B 0 ?

3. Compute T 2 and T 3 . Deduce M n for all positive integer n.

Exercise 4.9. We consider the functions f1 and f2 defined by:

∀x ∈ R, f1 (x) = e2x and f2 (x) = xe2x .
Let E = α f1 + β f2 (α, β) ∈ R2 .


1. Show that E is a real vector space with basis B = ( f1 , f2 ).

2. Let φ : E → E be the map defined by: ∀ f ∈ E, φ( f ) = f 0 . Show that φ is a well-defined
endomorphism of E. and determine the matrix A of φ in the basis B.
3. Write A = 2I2 + B and compute An for all n ∈ N∗ .
4. Deduce the expression of the n-th derivative of the function f : x 7→ (3x + 1)e2x .

Exercise 4.10. We consider a real vector space E of dimension 3 together with a basis
B = (e1 , e2 , e3) of E. We consider
 the endomorphism f of E, the matrix in the basis B of
0 2 3
which is A = 1 1 3 .
0 0 −1
1. Show that f is a bijection.
2. a) Find the real numbers α and β such that f 2 = α f + β idE .
b) Deduce that f −1 is a linear combination of f and idE . Explicit the matrix A−1 .
c) Show that for all n ∈ N, there exists real numbers αn and βn such that f n = αn f +
βn idE , but don’t try to determine these numbers.
68 4. MATRICES

3. We set v1 = 3e2 − 2e3 , v2 = e1 + e2 − e3 and v3 = e1 + e2 .

a) Show that B 0 = (v1 , v2 , v3 ) is a basis of E.
b) Give the coordinates of f (v1 ), f (v2 ) and f (v3 ) in the basis B. Deduce the matrix
A0 of f in the basis B 0 .
c) Write the change of basis matrix P , from the basis B to the basis B 0 . What is the
relation between A and A0 ?
d) Compute P −1 .
e) Decompose the vector u = 2e1 + 3e2 − e3 in the basis B 0 and u 0 = −v1 + 2v2 + v3 in
the basis B.
4. Use the matrices A0 and (A0 )n to find the expression of αn and βn .

Exercise 4.11 (From Term-Test, May 2010). Let E = R3 and B0 = (e1 , e2 , e3 ) denote the
standard basis of E. Let f be the endomorphism of E the matrix of which, in the basis B0
is  
1 1 1
A = −1 1 −1 .
−2 0 −2
1. a) What is the rank of f ?
b) Let u1 = (1, 0, −1). Compute f (u1 ).
c) Deduce a basis of the kernel of f .
2. Let u2 = (−2, 1, 3) and u3 = (−3, 1, 5).
a) Show that the family B = (u1 , u2 , u3 ) is a basis of E.  
0 2 1
b) Let T be the matrix of f in the basis B. Show that T = 0 0 −1.
0 0 0
c) Let n ∈ N. Compute T n .
3. Let P be the change of basis matrix from the basis B0 to the basis B.
a) Write P explicitly.
b) Give a relation between A, P and T .
c) Let n ∈ N. Compute An .
d) Let I be the identity matrix in M3 (R) and set B = I + A. Compute B n .
e) Let u = (x, y, z) ∈ E. Determine, in terms of x, y and z, the coordinates of u in the
basis B.

Exercise 4.12 (From Term-Test, May 2010). We consider the mapping

f : R3 −→ R3
(x, y, z) 7−→ (y + z, x + z, x + y).
We denote by id the identity map on R3 .
1. a) Show that f is an endomorphism of R3 .
b) Write explicitly the matrix A of f in the standard basis of R3 .
2. a) Show that f is a bijective map.
b) Compute A−1 and f −1 (a, b , c) for (a, b , c) ∈ R3 .
3. a) Show that f 2 can be written as a linear combination of f and id.
b) Deduce that f −1 is a linear combination of f and id.
c) Show that, for all n ∈ N, there exists real numbers an and bn such that
f n = an f + bn id
(do not try to find an exact expression of an and bn ).
 EXERCISES 69

Exercise 4.13 (Matrix Construction of Complex Numbers). Let C be the following subset
of M2 (R):
a −b
§ ‹ ª
C= ∈ M2 (R) a, b ∈ R .
b a
Prove that C is a vector space over R of dimension 2. Consider the mapping
Φ : C −→  C ‹
a −b
z = a + i b 7−→ .
b a
Prove that Φ is an isomorphism (of real vector spaces), and that moreover, for all z, z 0 ,
Φ(z z 0 ) = Φ(z)Φ(z 0 ), Φ (z) = t Φ(z).

Exercise 4.14 (Quaternions). By definition, the set of quaternions1 is the following subset
of M2 (C):
a − i d −b + i c
§ ‹ ª
H= ∈ M2 (C) a, b , c, d ∈ R .
b + ic a + id
Prove that H is a real subspace of M2 (C) of dimension 4, and that for all A, B ∈ H, AB ∈ H.
Prove that an element A ∈ H is invertible if and only if A 6= 0H . Let I be the matrix
0 −1
 ‹
I=
1 0
and define
c : H −→ H
A 7−→ I A
Prove that c is well-defined and that it defines a complex structure on H, i.e., that c 2 = −IH .
Hence H can be seen as a complex vector space of dimension 2.

Exercise 4.15. Consider the matrix

 
1 1 1
A =  1 2 1 .
−3 2 1
Prove that rk A = 3 and write A as a product of elementary matrices.

Exercise 4.16 (From Term-Test, May 2011). We consider f the endomorphism of R3 the
matrix of which in the standard basis B0 = (e1 , e2 , e3 ) of R3 is
 
1 0 0
A = 1 2 1 .
2 −2 −1

We denote by id the identity endomorphism of R3 .

1. Determine the rank of f .
2. Let u = e2 − 2e3 . Prove that Ker f = Span{u}.

1H is an example of a non-commutative field, and was introduced by Hamilton in 1843. It has applications
in computer images, signal theory and theoretical physics.
70 4. MATRICES

3. Let E1 = Ker( f −id). Prove that an element x ∈ R3 belongs to E1 if and only if f (x) = x.

We admit that E1 = Span{v}, where v = e2 − e3 , and you may use this result without
any justification in the sequel.
4. a) Show that the family B = (u, v, e1 ) is a basis of R3 .  
0 0 −3
b) Let T = [ f ]B be the matrix of f in the basis B. Prove that T = 0 1 4 .
0 0 1
c) Let P = [B]B0 be the change of basis matrix from the basis B0 to the basis B.
i) Explicit the matrix P and compute P −1 .
ii) Give a relation between A, P and T .
5. In this question, our goal is to find all the matrices M ∈ M3 (R) such that M 2 = A. Let
M ∈ M3 (R) such that M 2 = A. We denote by g the endomorphism of R3 such that
[g ]B0 = M .
a) Prove that g ◦ g = f .
b) Deduce that f ◦ g = g ◦ f .
c) Show that ( f ◦ g )(u) = 0. Deduce that g (u) ∈ Ker f . What can you conclude from
question 2? Deduce the value of (g ◦ g )(u) and deduce that g (u) = 0.
d) Compute ( f ◦ g )(v) and show that there exists a real number x such that g (v) = x v.
e) Prove that the matrix S of g in the basis B = (u, v, e1 ) is of the form
 
0 0 z
S = 0 x t  ,
0 0 y

where x is the real number found in the previous question and where z, t and y are
real numbers.
f) Determine all the matrices S ∈ M3 (R) such that S 2 = T .
g) Let S ∈ M3 (R) such that S 2 = T and set M = P S P −1 . Prove that M 2 = A.
h) Explain how we can obtain all matrices M ∈ M3 (R) such that M 2 = A.

Exercise 4.17 (From Test #5, May 2012). Let E be a real vector space of dimension 3 and let
B = (e1 , e2 , e3 ) be a basis of E. Consider the endomorphism u of E, the matrix of which
in the basis B is
 
1 1 −1
[u]B = A = 2 0 −1 .
4 2 −3
1. Determine a basis of Ker(u).
2. Compute the rank of u + I .
3. We consider the following vectors of E:

e10 = −e1 + 2e2 , e20 = e1 − e2 + e3 , e30 = e1 + e2 + 2e3 .

a) Prove that (e10 , e20 ) is a basis of Ker(u + I ).

b) Prove that B 0 = (e10 , e20 , e30 ) is a basis of E.
c) Determine the matrix D = [u]B 0 of u in the basis B 0 .
d) Explicit P = [B 0 ]B , the change of basis matrix from the basis B to the basis B 0 ,
and give a relation between A and D.
 EXERCISES 71

e) Compute D 2 and deduce that p = u 2 is a projection of E. Give the characteristic

elements of p, i.e., the subspaces F and G of E such that p is the projection on F
along G.

Exercise 4.18 (From Test #5, May 2012). We consider the matrix
 
1 1 −2
A = 0 1 2  .
0 0 1

Part I
We set N = A − I3 .
1. Prove that A is an invertible matrix and determine A−1 .
2. For k ∈ N, compute N k .
3. For n ∈ N, deduce a formula for An . Is this formula still valid for n ∈ Z? (Justify)
Part II
We consider the real vector space E = RR and the following vectors of E:
f1 : R −→ R f2 : R −→ R f3 : R −→ R

x 7−→ e x , x 7−→ (x + 1)e x , x 7−→ x 2 − 1 e x .
We set F = Span{ f1 , f2 , f3 }.
1. Show that B = ( f1 , f2 , f3 ) is a basis of F .
2. We consider the vector g ∈ E defined by
g : R −→ R
x 7−→ 3x 2 − 2x + 5 e x .

Show that g ∈ F and determine the coordinates of g in the basis B.

3. We consider the mapping
u : F −→ F
f 7−→ f 0 .
Prove that u is a well-defined endomorphism of F .
4. Determine the matrix [u]B of u in the basis B.
5. For n ∈ N, determine g (n) , the n-th derivative of the function g defined in question 2.
 5
Determinants

Throughout this chapter, K = R or C. We present an effective algorithm to compute

determinants (Laplace expansions), and we’ll prove in a later section that it is a well-defined
procedure.

1. Determinant of a Square Matrix

If A is a square matrix in M n (K), the determinant of A is a scalar denoted by det A or |A|. We
introduce an iterative procedure to compute determinants, and we study some properties.

Definition 5.1.

• If A = (a) ∈ M1 (K), then det A = a.

a b
 ‹
• If A = ∈ M2 (K), then det A = ad − b c.
c d

How to compute determinants of higher order matrices? Choose a row or a column of

A (typically we choose the row or column that has the most zeros, if any). Then “expand
the determinant” along that row or column, considering the following signs, as shown in
Example 5.2 below.

(+) (−) (−)n+1

 
a11 a12 ··· a1n
(−) (+) (−)n
a21 a22 ··· a2n
 
 
.. .. .. ..
 
 
 . . . . 
 (−)i+1 (−)i+2 (−)n+i  .
 
 ai 1 ai2 ··· ai n 
 . .. .. .. 
 .
 . . . . 

(−)n+1 (−)n+2 (−)2n
an1 an2 ··· ann

73
74 5. DETERMINANTS
 
4 1 −1 2
3 −1 2 0
Example 5.2. Let A = 
1 1 −1 1. We expand the determinant along the last col-


2 −2 1 1
umn (we could also expand it along the second row):
4 1 −1 2
3 −1 2 0
det A =
1 1 −1 1
2 −2 1 1
3 −1 2 4 1 −1 4 1 −1
= −2 1 1 −1 − 3 −1 2 + 3 −1 2 .
2 −2 1 2 −2 1 1 1 −1
Then start again with each smaller determinant until you get the result:
3 −1 2
1 −1 −1 2 −1 2
1 1 −1 = 3 − +2
−2 1 −2 1 1 −1
2 −2 1
= 3(1 − 2) − (−1 + 4) + 2(1 − 2) = −3 − 3 − 2 = −8,
4 1 −1
−1 2 1 −1 1 −1
3 −1 2 = 4 −3 +2
−2 1 −2 1 −1 2
2 −2 1
= 4(−1 + 4) − 3(1 − 2) + 2(2 − 1) = 12 + 3 + 2 = 17,
4 1 −1
−1 2 1 −1 1 −1
3 −1 2 = 4 −3 +
1 −1 1 −1 −1 2
1 1 −1
= 4(1 − 2) − 3 × 0 + (2 − 1) = −4 + 1 = −3.
Hence det A = −2 × (−8) − 17 − 3 = −4.

The following technical result needs some rather long computations and definitions, and
will be proved in a separate section (Section 5) at the end of the chapter.

Theorem 5.3 (Laplace). The procedure thus described to compute determinants is well-defined,
that is, no matter which column or row you choose to expand the determinant, the result will
always be the same.

The following proposition is straightforward:

Proposition 5.4. Let A ∈ M n (K) and λ ∈ K.

(1) det Idn = 1.
(2) det t A = det A.
(3) The determinant of a triangular matrix is the product of its diagonal coefficients.
(4) For all λ ∈ K, det(λA) = λn det A.

We now state the behavior of the determinant under multiplication by an elementary

matrix. The proof is not given, but it is elementary:

Proposition 5.5. Let A ∈ M n (K). Then:

 1. DETERMINANT OF A SQUARE MATRIX 75

(1) For all i =

6 j , det(Ti j A) = − det A, i.e., exchanging two rows (or two columns) of A
changes the sign of its determinant.
(2) For all λ ∈ K, det Di (λ)A = λ det A, i.e., if you multiply a row (or a column) of A

by λ, then its determinant is multiplied by
λ (this is also valid if λ = 0).

(3) For all i 6= j and ∀λ ∈ K, det Si j (λ)A = det A, i.e., the determinant of A is un-
changed when adding a multiple of any row (or columns) of A to another row (or
column) of A.

Remark 5.6. The following determinants are straightforward:

• For all i 6= j , det Ti j = −1.
• For all λ ∈ K, det Di (λ) = λ.
• For all i 6= j and λ ∈ K, det Si j (λ) = 1.
Hence, Proposition 5.5 states that for all A ∈ M n (K) and for all elementary matrix Q ∈
M n (K), then
det(QA) = det Q det A.
Also, since the transposed of any elementary matrix is also an elementary matrix, one has:
det(AQ) = det t (AQ) = det t Q t A = det t Q det t A = det Q det A = det Adet Q.

Proposition 5.7. Let A, B ∈ M n (K) and λ ∈ K.

(1) A is invertible if and only if det A 6= 0.
(2) det(AB) = det Adet B.
1
(3) If A is invertible, then det A−1 = .
det A
(4) In particular, if A has two proportional rows (or columns) then det A = 0.
(5) If you add a multiple of a row (or column) of A to another row (or column) of A the
determinant doesn’t change.
(6) The determinants of two similar matrices are equal.

Proof.
(1) By Theorem 4.38, there exists elementary matrices A1 , . . . , A p , and a matrix R
with all its coefficients equal to 0, except that R has rk A ones on its diagonal, such
that A = A1 · · · A p R. By Proposition 5.5 (see Remark 5.6),
det A = det(A1 · · · A p R) = det A1 · · · det A p det R.
Now since det R = 1 if and only if rk A = n if and only if A is invertible, we
conclude that det A 6= 0 if and only if A is invertible.
(2) If A is not invertible or if B is not invertible, then AB is not invertible and hence
det(AB) = 0 = det Adet B.
Now, if both matrices are invertible, then by Proposition 5.5 there exists elemen-
tary matrices A1 , . . . , A p such that A = A1 · · · A p . Now by Proposition 5.5 (see
Remark 5.6), we have
det(AB) = det(A1 · · · A p B) = det A1 · · · det A p det B = det(A1 · · · A p ) det B = det Adet B.
(3) If A is invertible, then
1 = det Idn = det AA−1 = det Adet A−1 ,


76 5. DETERMINANTS

1
hence det A−1 = .
det A
(4) If A has two proportional rows or columns, then A is not invertible, hence det A =
0.
(5) This is a particular case of Proposition 5.5.
(6) Let A and B be two similar matrices in M n (K), i.e., there exists an invertible matrix
P ∈ M n (K) such that B = P −1 AP . Then:
1
det B = det P −1 AP = det P −1 det Adet P = det Adet P = det A.


det P
As shown by statement (5) of Proposition 5.7, we can use Gaussian elimination to compute
determinants, as shown in the following example. For generic matrices, this procedure
uses less operations (i.e., additions, subtractions, multiplications and divisions) than row-
column expansions.

Example 5.8. We use Gaussian elimination to compute the determinant of a matrix A, that
is, we perform elementary row (or column) operations to A in order to get a triangular
matrix (or at least a simpler matrix), to simplify the computation of the determinant. Let
 
4 1 −1 2
3 −1 2 0
A=  1 1 −1 1 .


2 −2 1 1
Then:
4 1 −1 2
3 −1 2 0
det A =
1 1 −1 1
2 −2 1 1
1 1 −1 1
3 −1 2 0
=− , R1 ↔ R3
4 1 −1 2
2 −2 1 1
1 1 −1 1
0 −4 5 −3
=− , R2 ← R2 − 3R1 , R3 ← R3 − 4R1 , R4 ← R4 − 2R1
0 −3 3 −2
0 −4 3 −1
1 1 −1 1
0 −3 5 −4
= , C2 ↔ C4
0 −2 3 −3
0 −1 3 −4
1 1 −1 1
0 −1 3 −4
=− , R2 ↔ R4
0 −2 3 −3
0 −3 5 −4
1 1 −1 1
0 −1 3 −4
=− , R3 ← R3 − 2R2 , R4 ← R4 − 3R2
0 0 −3 5
0 0 −4 8
 3. GEOMETRIC INTERPRETATION 77

−1 3 −4
= − 0 −3 5 , expanding with respect to the first column
0 −4 8
−3 5
= , expanding with respect to the first column
−4 8
= −24 + 20 = −4.

2. Determinant of an Endomorphism
Proposition 5.9. In a finite dimensional space, the determinant of the matrix of an endomor-
phism doesn’t depend on the choice of the basis.

Proof. Let B and C be two bases of E. We want to show that det[ f ]B = det[ f ]C .
We know that
[ f ]C = [C ]−1B
[ f ]B [C ]B .
Hence:
det[ f ]B [C ]B
det[ f ]C = = det[ f ]B .
det[C ]B
Proposition 5.9 justifies the following definition:

Definition 5.10. Let E be a finite-dimensional vector space over K and f ∈ L(E). Let B
be a basis of E. By definition, det f = det[ f ]B .

Proposition 5.11. Let E be a finite dimensional vector space over K and let f ∈ L(E). Then:
f is injective ⇐⇒ f is surjective ⇐⇒ f is bijective ⇐⇒ det f 6= 0.

3. Geometric Interpretation
The determinant has a nice geometric interpretation in terms of volumes.
Let E = R2 and let u = (a, b ), v = (c, d ) ∈ E. Is is well-known that the following expression
a c
 ‹
det = |ad − b c|
b d
is the surface area of the parallelogram generated by the vectors u and v. In fact, the
determinant ad − b c is exactly the third component of the cross product
     
a c 0
 b  × d  =  0  ,
0 0 ad − b c
the norm of which is the area of the parallelogram generated by (a, b ) and (c, d ).
Let E = R3 and let u = (u1 , u2 , u3 ), v = (v1 , v2 , v3 ), w = (w1 , w2 , w3 ) ∈ R3 . It is well-known
that the following expression
 
u1 v1 w1
det  u2 v2 w2  = ((u, v, w)) = u · (v × w)
u3 v3 w3
is the volume of the parallelepiped generated by the vectors u, v and w. Indeed, the term
on the right-hand side of the previous equality is the absolute value of the triple scalar
product of the vectors u, v and w.
78 5. DETERMINANTS

The sign of the determinant has the following interpretation: if E = R2 , the determinant
a b
 ‹
det
c d
is positive if and only if the family (a, b ), (c, d ) is positively oriented with respect to the

standard basis of E. Similarly, if E = R3 , the determinant

 
u1 v1 w1
det  u2 v2 w2 
u3 v3 w3
is positive if and only if the family (u, v, w) is positively oriented with respect to the stan-
dard basis of E.

Definition 5.12. Let E be vector space over K of dimension n, let B be a basis of E.

• Let F = (u1 , . . . , un ) be a family of n vectors of E. The determinant of F with
respect to the basis B, denoted by detB F = detB (u1 , . . . , un ) is the following
determinant:
detB (u1 , . . . , un ) = det[F ]B ,
where [F ]B ∈ M n (K) is the coordinates matrix of the family F in the basis B
• Let C be another basis of E. We say that B and C have the same orientation
if detB C > 0. Otherwise, if detB C < 0, we say that B and C have opposite
orientations.
• Let F = (u1 , . . . , un ) be a family of n vectors of E. The algebraic volume with re-
spect to B of the parallelepiped generated by the family F and is denoted by VolB F .

The determinant of an endomorphism has a nice interpretation in terms of volumes:

Proposition 5.13. Let E be a finite dimensional vector space over K, let B = (u1 , . . . , un ) be
a basis of E. Then
det f = detB f (B),
where f (B) = f (u1 ), . . . , f (un ) .

Hence, the determinant of f is positive if and only if the families B and f (B) have the
same orientation. Moreover, | det f | measures how much f “expands volumes”. Clearly,
det f = 0 means that f (B) is not a basis of E, i.e., it is not independent, i.e., the volume
of the parallelepiped generated by f (B) is zero.

4. Trace of a matrix — Trace of an Endomorphism

Definition 5.14. Let A = ai j ∈ M n (K). The trace of A is the scalar:

tr A = a11 + a22 + · · · + ann .

Proposition 5.15. Let A and B two matrices in M n (K). Then

tr(AB) = tr(BA).
Moreover, if P ∈ M n (K) is an invertible matrix, then
tr P −1 AP = tr A.


 5. PROOF OF THEOREM 5.3 79

Proof. Let A = (ai j ), B = (bi j ), C = (ci j ) = AB and D = (di j ) = BA. Then:

n
X n X
X n
tr C = ci i = ai k bk i
i=1 i=1 k=1
Xn X n Xn
= bk i ai k = dk k
k=1 i=1 k=1
= tr D.
If moreover P ∈ M n (K) is an invertible matrix, then
tr P −1 AP = tr P −1 (AP ) = tr (AP )P −1 = tr AP P −1 = tr(AIdn ) = tr A.

Proposition 5.15 justifies the following definition:

Definition 5.16. Let E be a finite dimensional vector space, f ∈ L(E), and B be any basis
of E. Then the trace of f is:
tr f = tr[ f ]B .

Proposition 5.17.
• The trace of matrices is a linear form on M n (K), i.e., it is a linear map from M n (K)
to K.
• Let E be a finite-dimensional space over K. The trace of endomorphisms of E is a
linear form on L(E), i.e., it is a linear map from L(E) to K.

Proof.
• Let A = (ai j ) ∈ M n (K), B = (bi j ) ∈ M n (K) and λ ∈ K. Then:
n
X
tr(A + λB) = (ai i + λbi i )
i =1
X n n
X
= ai i + λ bi i
i =1 i=1
= tr A + λ tr B.
• Let f , g ∈ L(E) and λ ∈ K. Let B be a basis of E. Then:
tr( f + λg ) = tr[ f + λg ]B = tr [ f ]B + λ[g ]B = tr[ f ]B + λ tr[g ]B = tr f + λ tr g .

5. Proof of Theorem 5.3

5.1. Permutations and Signatures
5.1.1. Permutations
Definition 5.18. Let n ∈ N∗ and let Jn be the set with n elements, Jn = {1, . . . , n}. A
permutation of Jn is a bijection σ of Jn . The set of all permutations of Jn is denoted by Sn
and is called the symmetric groups of Jn .

Remarks 5.19.
(1) If σ, σ 0 ∈ Sn , then σ ◦σ 0 ∈ Sn , as the composition of two bijections is a bijection.
The composition is associative in the sense that if σ, σ 0 , σ 00 ∈ Sn , then σ ◦ σ 0 ◦
σ 00 = σ ◦σ 0 ◦σ 00 . Also, the identity map on Jn , denoted by IdJn , is an element of



80 5. DETERMINANTS

Sn . Finally, every permutation σ ∈ Sn possesses an inverse, that is, there exists

an element σ −1 ∈ Sn such that σ ◦ σ −1 = σ −1 ◦ σ = Idn . Sn is hence what is
called a group.
(2) Clearly, Card Sn = n!.
5.1.2. Signature
Definition 5.20. Let n ∈ N, n ≥ 2 and let σ ∈ Sn . The signature of σ is the number
sign(σ) ∈ {−1, 1} defined thus:
Y σ(i) − σ( j )
sign(σ) = .
1≤i< j ≤n
i−j

Remark 5.21. Let n ∈ N, n ≥ 2. It should be clear that for any permutation σ ∈ Sn , the
number sign(σ) is well-defined, since the denominator of sign(σ) never vanishes (as the
product is taken for all numbers i < j ). Also, sign(σ) is indeed in {−1, 1} since every factor
in the denominator of sign(σ) appears in its numerator, with possibly the opposite sign, as
σ is a bijection from Jn to Jn .

5.1.3. A Technical Lemma and Notations

We fix some notation: Let n ∈ N, n ≥ 2. For all i ∈ Jn , we define a one-to-one (but not
onto!) mapping Bi : Jn → Jn+1 by
¨
k if k < i
Bi (k) =
k + 1 if k ≥ i.

Notice that the only element that is not in the range of Bi is i. Let i, j ∈ Jn and σ ∈ Sn+1
be such that σ(i) = j . We define σi j ∈ Sn as
¨
σ Bi (k) if σ Bi (k) < j



σi j (k) =
σ Bi (k) − 1 if σ Bi (k) > j .

Clearly, since σ(i) = j , σi j is a well-defined element of Sn , and

B j ◦ σi j = σ ◦ Bi .

We set:
Sn+1,i, j = σ ∈ Sn+1 σ(i) = j .


Clearly, the mapping

Bi j : Sn+1,i, j −→ Sn
σ 7−→ σi j
is onto. In fact, it is also one-to-one, since we can always recover σ from σi j : If σ, σ 0 ∈
Sn+1,i, j are such that σi j = σi0 j , then B j ◦ σi j = σ ◦ Bi = B j ◦ σi0 j = σ 0 ◦ Bi , hence σ and σ 0
are equal on Jn+1 \ {i}, but since σ(i) = j = σ 0 (i), σ and σ 0 are equal on Jn+1 .
We will also use the very simple observation that if i ∈ Jn+1 ,
n+1
[
Sn+1 = Sn+1,i, j ,
j =1
 5. PROOF OF THEOREM 5.3 81

and that this union is a disjoint union. Similarly, if j ∈ Jn+1 , we also have the disjoint union
n+1
[
Sn+1 = Sn+1,i , j .
i =1

Remark 5.22. The notation σi j is only defined if σ(i ) = j , i.e., if σ ∈ Sn+1,i, j . Otherwise,
it doesn’t make sense to write σi j .

Lemma 5.23. Let n ≥ 2, i, j ∈ Jn+1 and σ ∈ Sn+1,i, j . Then

sign(σi j ) = (−1)i+ j sign(σ).

Proof.
Y σi j (k) − σi j (`)
sign(σi j ) =
1≤k<`≤n
k −`
B j σi j (k) − B j σi j (`)



Y
=
1≤k<`≤n
B j (k) − B j (`)
σ Bi (k) − σ Bi (`)



Y
=
1≤k<`≤n
B j (k) − B j (`)
Y 
 Y 1
= σ Bi (k) − σ Bi (`)

1≤k<`≤n
B (k) − B j (`)
1≤k<`≤n j
Y Y 1
= σ(k) − σ(`)

1≤k<`≤n+1 1≤k<`≤n+1
k −`
k6=i, `6=i k6= j , `6= j
Y 1 Y
= σ(k) − σ(`)

1≤k<`≤n+1
σ(k) − σ(`) 1≤k<`≤n+1
k = i or ` = i
Y Y 1
(k − `)
1≤k<`≤n+1 1≤k<`≤n+1
k −`
k = j or ` = j
n+1 i −1 n+1 j −1
Y 1 Y 1 Y Y
= sign(σ) ( j − `) (k − j )
`=i+1
σ(i) − σ(`) k=1 σ(k) − σ(i) `= j +1 k=1
n+1 n+1
Y 1 Y
= (−1)i+ j −2 sign(σ) ( j − `)
`=1
σ(i) − σ(`) `=1
`6=i `6= j
i+j
= (−1) sign(σ).

Finally, we need another notation that will be useful later (especially when we introduce
cofactors): let A ∈ M n+1 (K) and i, j ∈ {1, . . . , n + 1}. We denote by Ai j ∈ M n (K) the
matrix obtained from A by removing the i-th row and the j -th column. More specifically,
if B = (bk` )k=1..n = Ai j ∈ M n (K), then the coefficients bk` of B are given by
`=1..n

bk` = aBi (k),B j (`) .

82 5. DETERMINANTS

5.2. Leibniz Formula for Determinants

Theorem 5.24 (Leibniz Formula). Let A = (ai, j ) ∈ M n (K), n ≥ 2. Then
X n
Y
det A = sign(σ) ai,σ(i) .
σ∈Sn i=1

Proof. By induction on n: if n = 2, Leibniz formula reads as

det A = a11 a22 − a12 a21
which is exactly how we defined the determinant of a 2 × 2 matrix. We now assume
that Leibniz formula is valid for some n ≥ 2 and we prove that it is valid for n + 1: let
A = (ai j ) ∈ M n+1 (K).
• Let i ∈ Jn+1 = {1, . . . , n + 1}. We prove that if we expand the determinant of A
with respect to the i-th row, we get Leibniz formula. Since the expansion only
involves determinants of n × n matrices, we can use our hypothesis, i.e., that
Leibniz formula holds for n × n matrices. Denote by Di· this expansion. It reads
exactly as:
n+1
X
Di· = (−1)i+ j ai j det(Ai j )
j =1
!
n+1
X X n
Y
= (−1)i+ j ai j sign(σ) aBi (k),B j ◦σ(k)
j =1 σ∈Sn k=1
 
n+1
X X n
Y
= (−1)i+ j ai j  sign(σi j ) aBi (k),B j ◦σi j (k) 
j =1 σ∈Sn+1,i, j k=1
 
n+1
X X n
Y
= ai j  sign(σ) aBi (k),σ◦Bi (k) 
j =1 σ∈Sn+1,i, j k=1
 
n+1
X X n+1
Y
=
 
ai j 
 sign(σ) ak,σ(k) 

j =1 σ∈Sn+1,i, j k=1
k6=i
 
n+1
X X n+1
Y
=  sign(σ) ak,σ(k) 
j =1 σ∈Sn+1,i, j k=1

X n+1
Y
= sign(σ) ak,σ(k) .
σ∈Sn+1 k=1

This proves that if we expand the determinant along any row, we obtain the same
value, given by Leibniz formula.
• Let j ∈ Jn+1 = {1, . . . , n + 1}. We prove that if we expand the determinant of
A with respect to the j -th column, we get Leibniz formula. Denote by D· j this
expansion. It reads exactly as:
n+1
X
D· j = (−1)i+ j ai j det(Ai j )
i =1
 6. COFACTORS, COMATRIX AND A FORMULA FOR THE INVERSE OF A MATRIX 83
!
n+1
X X n
Y
= (−1)i+ j ai j sign(σ) aBi (k),B j ◦σ(k)
i=1 σ∈Sn k=1
 
n+1
X X n
Y
= (−1)i+ j ai j  sign(σi j ) aBi (k),B j ◦σi j (k) 
i=1 σ∈Sn+1,i , j k=1
 
n+1
X X n
Y
= ai j  sign(σ) aBi (k),σ◦Bi (k) 
i=1 σ∈Sn+1,i, j k=1
 
n+1
X X n+1
Y
=
 
ai j 
 sign(σ) ak,σ(k) 

i=1 σ∈Sn+1,i , j k=1
k6=i
 
n+1
X X n+1
Y
=  sign(σ) ak,σ(k) 
i=1 σ∈Sn+1,i, j k=1

X n+1
Y
= sign(σ) ak,σ(k) ,
σ∈Sn+1 k=1

and we recover once again Leibniz formula.

Clearly, the proof of Theorem 5.24 is also a proof for Theorem 5.3.

6. Cofactors, Comatrix and a Formula for the Inverse of a Ma-

trix
In this section, we still use the notations of Section 5, in particular the notation Ai j , which
stands for the matrix obtained from A when erasing the i-th row and the j -th column.

Definition 5.25. Let A ∈ M n (K) and 1 ≤ i, j ≤ n. The (i, j )-cofactor of A is the scalar
ci j = (−1)i+ j det(Ai j ).
The comatrix of A is the matrix C = (ci j ) ∈ M n (K), that consists of the cofactors of A. The
adjugate of A is the transposed of the comatrix: adj(A) = t C .

In fact, the definition of determinants we gave is known as Laplace expansion or Laplace

formula for determinants. And the result can be formalized thus (using cofactors), the proof
of which being a direct consequence of the computations of the proof of Theorem 5.24:

Theorem 5.26. Let A ∈ M n (K) and let C = (ci j ) be its comatrix. Then for all 1 ≤ i ≤ n,
n
X
det A = ai j ci j
j =1

(expansion along the i -th row), and for all 1 ≤ j ≤ n,

Xn
det A = ai j ci j
i =1
(expansion along the j -th row).
84 5. DETERMINANTS

The use of the comatrix is mainly theoretical in the following formula (we will also use it
to prove Hamilton–Cayley Theorem in Chapter 6):

Theorem 5.27. Let n ≥ 2 and A ∈ M n (K) with comatrix C = (ci j ). Then:

t
C A = A t C = (det A) Idn .
In particular, if A is an invertible matrix then
1 t 1
A−1 = C= adj(A).
det A det A

Proof. We prove that A t C = (det A) Idn ; the proof of t C A = (det A) Idn is similar. Set
M = (mi j ) = A t C . Let i, j ∈ {1, . . . , n} and let’s compute the (i , j )-th coefficient mi j of
M = AtC:
Xn
mi j = ai k c j k .
k=1
Clearly, from Theorem 5.26, mi i = det(A). We now compute mi j if i 6= j (and we must
get 0):
n
X
mi j = ai` c j `
`=1
n n
ai` (−1) j +`
X X Y
= sign(σ) aB j (k),B` ◦σ(k)
`=1 σ∈Sn−1 k=1
n n
ai` (−1) j +`
X X Y
= sign(σ j ` ) aB j (k),B` ◦σ j ` (k)
`=1 σ∈Sn, j ,` k=1
n
X X n
Y
= ai` sign(σ) aB j (k),σ◦B j (k)
`=1 σ∈Sn, j ,` k=1
n
X X n
Y
= ai` sign(σ) ak,σ(k) .
`=1 σ∈Sn, j ,` k=1
k6= j

Define a matrix B = (bk` ) ∈ M n (K) as

¨
ak` if k 6= j
bk` =
ai` if k = j .
With this notation,
Xn X n
Y X n
Y
mi j = sign(σ) bk,σ(k) = sign(σ) bk,σ(k) = det B = 0,
`=1 σ∈Sn, j ,` k=1 σ∈Sn k=1

since B has two distinct rows that are equal, namely its i-th and j -th rows are equal (i 6=
j ).
a b
 ‹
In dimension 2, we obtain a very useful formula for the inverse of a matrix: Let A =
c d
be such that det A = ad − b c 6= 0. Then
1 d −b
 ‹
A−1 = .
ad − b c −c a
 EXERCISES 85

Corollary 5.28. Let A ∈ M n (R) be a matrix such that

• det A = ±1.
• All the coefficients of A are integers.
Then the coefficients of A−1 are also integers.

Proof. Clearly A is invertible (since det A 6= 0) and all the coefficients of adj(A) are integers,
hence
1
A−1 = adj(A) = ± adj(A)
det A
has only integer coefficients.

Exercises
 
3 6 −3
1 4
 ‹
Exercise 5.1. Let A = and B = 4 1 6 .
2 −3
5 −7 0
1. Compute the determinant of A.
2. Compute the determinant of B using five different methods:
a) By expanding with respect to the first row.
b) By expanding with respect to the second row.
c) By expanding with respect to the third column.
d) By using the Gaussian elimination.

m 2 −9 −4
 

Exercise 5.2. Let m ∈ R and B = 2m −3 −4. For which values of m is the matrix
8 3m −8
B invertible?

Exercise 5.3. For each of the following matrices, say if it is invertible, and if it is invertible,
compute its inverse.
   
1 −5 7 1 3 5
A=  7 1 −5 , B = 2 7 1  ,
−5 7 1 0 1 −9
 
  1 −2 1 0
1 1 1  1 −2 2 −3
C = 1 j j 2  , D = ,
2
0 1 −1 1 
1 j j
−2 3 −2 3

where j is the third root of unity: j = e2iπ/3 .

Exercise 5.4. Let a, b and x be real numbers. Solve in R the following equations in x:

x a a2
1 x x x
1
x 1 x x x a = 0,
(1) = 0, (2) a
x x 1 x
x x x 1 1 1
x
a2 a
86 5. DETERMINANTS

1 1 1 1 x a b x
1 1 cos x cos b a x x b
(3) = 0, (4) = 0.
1 cos x 1 cos a b x x a
1 cos b cos a 1 x b a x
(In Equation (2) we assume that a 6= 0).

Exercise 5.5. Compute the following determinants:

1 2 3 4
a−b −c 2a 2a
1 3 6 10
D1 = , D2 = 2b b −c −a 2b ,
1 4 10 20
2c 2c c −a − b
1 5 15 35
1 a a2 a3 m 1 1 ··· 1
1 b b2 b3 1 m 1 ··· 1
D3 = , D4 = .. .. .. .. .. .
1 c c2 c3 . . . . .
1 d d2 d3 1 1 1 ··· m

Exercise 5.6. Let E be a finite dimensional vector space over K. What is the determinant
of a projection of E? What is the determinant of a symmetry of E?

Exercise 5.7. We consider the vector space E = R2 .

1. Let θ ∈ R. We denote by rθ the rotation of angle θ. We assume that rθ ∈ L(E). Write
the matrix of rθ in the canonical basis of E. What is det rθ ?
2. Let (x, y) ∈ E. Deduce from the previous question the form of rθ (x, y).
3. Let θ and θ0 be two real numbers. Considering the fact that rθ ◦ rθ0 = rθ+θ0 , recover
the usual addition formulas for sin and cos.
4. Let λ ∈ R. The hyperbolic rotation of angle λ is the endomorphism hλ of E such that its
cosh λ sinh λ
‹
matrix in the canonical basis of E is Hλ = . What is hλ ◦ hλ0 ?
sinh λ cosh λ
5. Compute the determinant of hλ .

Exercise 5.8 (From Test #5, May 2012). Find all real numbers m that satisfy the following
equation:
m 1 1 1
1 m 1 1
= 0.
1 1 m 1
1 1 1 m
 Diagonalization of Endomorphisms
6
1. Eigenvalues, Eigenvectors and Eigenspaces
Definition 6.1. Let E be a vector space over K (possibly infinite-dimensional) and f ∈ L(E).
A scalar λ ∈ K is called an eigenvalue of f if the following subspace of E

Eλ = x ∈ E f (x) = λx


is different from {0E }. In this case, the subspace Eλ is called the eigenspace of f corresponding
to the eigenvalue λ. Any non-zero vector of Eλ is called an eigenvector of f corresponding to
the eigenvalue λ. The set of all eigenvalues of f is called the spectrum of f and is denoted
by spec f .

Remark 6.2. Let f ∈ L(E) and λ ∈ K. Another way to say that λ is an eigenvalue is the
following:
λ ∈ spec f ⇐⇒ Ker( f − λ idE ) 6= {0E }.
This is true since the space Eλ is exactly Ker( f − λ idE ).

Eigenspaces of an endomorphisms are very useful to decompose an endomorphism into

simpler endomorphisms (i.e., multiplications by a scalar) on smaller spaces.
Similarly, for square matrices:

Definition 6.3. Let A ∈ M n (K). We say that a scalar λ ∈ K is an eigenvalue of A if there

exists a non-zero n × 1 matrix X ∈ M n1 (K) such that AX = λX . Such a column matrix X
is called an eigenvector of A corresponding to the eigenvalue λ. The set of all eigenvalues of A
is called the spectrum of A and is denoted by spec A.

Proposition 6.4. Let E be a finite dimensional vector space over K, let B be any basis of E
and f ∈ L(E). Then
spec f = spec[ f ]B .

Proof. We prove the statement by double inclusion:

87
88 6. DIAGONALIZATION OF ENDOMORPHISMS

If λ ∈ spec f then there exists a non-zero vector x ∈ E such that f (x) = λx. We transform
this equality into an equality of coordinates in the basis B:
f (x) B = [λx]B .
 

Now, f (x) B = [ f ]B [x]B , [λx]B = λ[x]B , and [x]B is not zero since x 6= 0E . Hence:
 

[ f ]B [x]B = λ[x]B
and this shows that λ ∈ spec[ f ]B . Hence spec f ⊂ spec[ f ]B .
If λ ∈ spec[ f ]B , there exists a non-zero n × 1 matrix X such that
[ f ]B X = λX .
Let x ∈ E be such that [x]B = X . Obviously x 6= 0E and
[ f ]B X = [ f ]B [x]B = f (x) B = λ[x]B .
 

Hence f (x) = λx which shows that λ ∈ spec f . Hence spec[ f ]B ⊂ spec f .

Proposition 6.5. Let f ∈ L(E) and λ and µ two distinct eigenvalues of f . Then the eigenspaces
of f corresponding to λ and µ are independent, i.e., Eλ ∩ Eµ = {0E }.


More generally, the family Eλ λ∈spec f is an independent family of subspaces of E.

Proof. For n ≥ 2, consider the proposition

(Pn ) For all distinct λ1 , . . . , λn ∈ spec f ,

the family (Eλi )i=1...n is an independent family of subspaces of E.
We now prove by induction that for all n ≥ 2 the property (Pn ) is true:
• For n = 2: let λ1 , λ2 be two distinct eigenvalues of f and let x ∈ Eλ1 ∩ Eλ2 . Clearly,
f (x) = λ1 x = λ2 x,
hence (λ1 − λ2 )x = 0E , and since λ1 6= λ2 we must have x = 0E . Hence Eλ1 and
Eλ2 are independent subspaces of E.
• We assume that the property (Pn ) is true for some n ≥ 2 and we prove that Pn+1
is true: Let λ1 , . . . , λn+1 be n + 1 distinct eigenvalues of f and let
n
X
x ∈ Eλn+1 ∩ Eλi ,
i=1
say
n
X
x = xn+1 = xi
i=1
with xi ∈ Eλi . Let’s prove that x = 0E : clearly,
f (x) = f (xn+1 ) = λn+1 xn+1
and ‚ n Œ
X n
X n
X
f (x) = f xi = f (xi ) = λi xi .
i =1 i=1 i =1
Hence
n
X
λn+1 xn+1 = λi xi .
i=1
 2. CHARACTERISTIC POLYNOMIAL 89

n
X
Since xn+1 = xi we also have:
i=1
n
X
(λn+1 − λi )xi = 0E .
i=1

From the property (Pn ) we deduce that for all i ∈ {1, . . . , n}, (λn+1 − λi )xi = 0E .
Since the λi ’s are distinct, we deduce that for all i ∈ {1, . . . , n}, xi = 0E , hence
x = 0E .
Since this is true for any distinct eigenvalues λ1 , . . . , λn+1 of f , we conclude that
(Pn+1 ) is true.
Hence, the family (Eλ )λ∈spec f is an independent family of subspaces of E.

2. Characteristic Polynomial
This is the recipe to find eigenvalues of a matrix (and hence of an endomorphism in a finite
dimensional space).
Let A ∈ M n (K). We want to find the λ’s in K such that there is a non-zero (column) vector
X ∈ M n1 (K) satisfying
AX = λX .
This equation also reads as:
(A − λ Id)X = 0.
Hence, the condition:
There exists X ∈ M n1 (K) such that AX = λX
is equivalent to the condition:
(A − λ Id) is not invertible
which is also equivalent to:
det(A − λ Id) = 0.

Definition 6.6. The characteristic polynomial of a matrix A ∈ M n (K) is the polynomial χA

defined as:
χA(λ) = det(A − λ Id).
Let f be an endomorphism of a finite-dimensional vector space E. The characteristic poly-
nomial of f is the polynomial χ f defined as:
χ f (λ) = det( f − λ idE ).

Remark 6.7. Clearly, the characteristic polynomial of an endomorphism is the characteris-

tic polynomial of its matrix in any basis: if B is any basis of the finite-dimensional vector
space E, then
χ f = χ[ f ]B .

The following proposition justifies the name characteristic polynomial (or at least the name
“polynomial”):

Proposition 6.8. Let A ∈ M n (K). Then:

• χA is a polynomial of degree n.
90 6. DIAGONALIZATION OF ENDOMORPHISMS

• If n ≥ 2, then:
χA(λ) = (−1)n λn + (−1)n−1 tr Aλn−1 + · · · + det A.
• The set of all eigenvalues of A (i.e., the spectrum of A) is the set of all roots of χA:
 
∀λ ∈ K, λ ∈ spec A ⇐⇒ χA(λ) = 0 .

Proof. Let A = (ai j ) and In = (δi j ). By Leibniz formula,

X n
Y
χA(X ) = det(A − X Idn ) = ai,σ(i ) − X δi,σ(i) .


sign(σ)
σ∈Sn i=1

Hence χA is a sum of polynomials of degrees at most n, hence χA is a polynomial of degrees

n
Y
ai,σ(i) − X δi,σ(i) is of degree n, and this occurs


at most of degree n. It is possible that
i=1
only when σ is the identity permutation, in which case the dominant coefficient is (−1)n .
Clearly, χA(0) = det(A), so, in order to finish the proof of the second point, we need to find
the coefficient of the degree n − 1 term in χA. The only way we can get a term of degree
Yn
ai,σ(i) − X δi,σ(i) is also when σ is the identity permutation. In this case, the


n − 1 in
i =1
product reads as
n
Y

ai i − X
i=1
and it can be easily seen that the coefficient in front of X n−1 is exactly (−1)n−1 tr A. The
third point is obvious from the previous discussion.
Once we know the spectrum of a matrix, we can compute the eigenvectors.

3. Diagonalization
Definition 6.9. An endomorphism of E is said to be diagonalizable if there exists a basis
of E consisting of eigenvectors of f .
A matrix A ∈ M n (K) is said to be diagonalizable if there exists an invertible matrix P ∈
M n (K) and a diagonal matrix D ∈ M n (K) such that
A = P D P −1 .

Remark 6.10. If E is a finite-dimensional vector space, B a basis of E and f ∈ L(E), then:

f is diagonalizable ⇐⇒ [ f ]B is diagonalizable.

The following proposition is straightforward, from the very definition of diagonalization:

Proposition 6.11. Let E be a finite-dimensional vector space over K, f ∈ L(E) and spec f =
{λ1 , . . . , λk }. Then:
f is diagonalizable ⇐⇒ E = Eλ1 ⊕ · · · ⊕ Eλk .

Proposition 6.12. Let E be a finite-dimensional vector space and f ∈ L(E). If χ f can be

factored in K:
χ f (X ) = (−1)dim E (X − λ1 )n1 · · · (X − λk )nk
 3. DIAGONALIZATION 91

then, for all 1 ≤ i ≤ k:

1 ≤ dim Eλi ≤ ni .

Proof. Let i ∈ {1, . . . , k} and let m = dim Eλi . Let (u1 , . . . , u m ) be a basis of Eλi . We can
complete this finite family into a basis B = (u1 , . . . , u p , v1 , . . . , vn− p ) of E. In the basis B,
the matrix of f has the form
λi · · · 0
 
a1, p+1 · · · a1n
. .
 .. . . ... ..
.
..
.
.. 
. 
 
 0 ··· λ a p, p+1 · · · apn 
 i 
A = [ f ]C =  .
 
 0 ··· 0 a p+1, p+1 · · · a p+1,n 
 
 .. .. .. .. .. .. 
 
. . . . . . 
0 ··· 0 an, p+1 · · · an,n
Let B ∈ M n− p (K) be the matrix
 
a p+1, p+1 · · · a p+1,n
B =
 .. .. ..  .
. . . 
an1 ··· an,n
Clearly,
χ f (X ) = (−1)n (X − λ1 )n1 · · · (X − λk )nk = χA(X ) = (λi − X ) p χB (X ).
Since all the λi ’s are distinct, we must have p ≤ n.

Proposition 6.13. Let E be a finite-dimensional vector space over K, f ∈ L(E). Then f is

diagonalizable if and only if the following two conditions are fulfilled:
• The characteristic polynomial of f can be factored in K:
χ f (X ) = (−1)dim E (X − λ1 )n1 · · · (X − λk )nk
• For all 1 ≤ i ≤ k,
dim Eλi = ni .


Proof. If the two conditions are fulfilled, then, since the family of subspaces Eλi 1≤i ≤k is
independent and since n1 + · · · + nk = dim E, we clearly have (by the inclusion–equality
theorem and by Grassmann formula)
E = Eλ1 ⊕ · · · ⊕ Eλk
which exactly means that f is diagonalizable. Conversely, if f is diagonalizable, then
its eigenvalues are in K, and hence its characteristic polynomial can be factors in K[X ].
Moreover, since
E = Eλ1 ⊕ · · · ⊕ Eλk
and since, by Proposition 6.12 dim Eλi ≤ ni , we must have dim Ei = ni by Grassmann


formula and the fact that the family of subspaces Eλi 1≤i≤k is independent.

The following result is important as it helps to justify the fact that a matrix is diagonalizable
in some usual cases:
92 6. DIAGONALIZATION OF ENDOMORPHISMS

Proposition 6.14. Let E be a vector space over K of dimension n. If f ∈ L(E) has n distinct
eigenvalues in K, then it is diagonalizable.

Proof. By Proposition 6.12, for all λ ∈ spec( f ), dim Eλ = 1, hence, by Grassmann formula:
!
M
dim Eλ = n,
λ∈spec( f )

hence, by the inclusion–equality theorem,

M
Eλ = E.
λ∈spec( f )

Hence f is diagonalizable.

Proposition 6.15. If a matrix A ∈ M n (K) has n distinct eigenvalues, then it is diagonalizable.

Proof. Let E = Kn together with its standard basis B and let f ∈ L(E) such that [ f ]B = A.
Clearly, by Proposition 6.14, f is diagonalizable, and hence A is also diagonalizable.

Proposition 6.16. Let A ∈ M n (K) be a matrix that possesses a unique eigenvalue λ of multi-
plicity n. Then A is diagonalizable if and only if A is diagonal with A = λ Idn .

Proof. Let λ be the unique eigenvalue of A of multiplicity n. Then A is diagonalizable if

and only if E = Eλ , which exactly means that A = λ Idn .
We will prove the following result next year, but it may shortcut some computations in
some problems:

Proposition 6.17. If a real matrix A ∈ M n (R) is symmetric (i.e., t A = A), then it is diagonal-
izable in R. Moreover there exists a basis of eigenvectors of A that is orthogonal with respect to
the standard dot product of Rn .

4. Hamilton–Cayley Theorem
Theorem 6.18 (Hamilton–Cayley Theorem for matrices). Let n ≥ 2 and A ∈ M n (K),
with characteristic polynomial χA. Then χA(A) = 0.

As a direct corollary, we get:

Theorem 6.19 (Hamilton–Cayley Theorem for endomorphisms). Let E be a finite vec-

tor space over K and let f ∈ L(E) with characteristic polynomial χ f . Then χ f ( f ) = 0L(E) .

Before we give a (valid) proof of Theorem 6.18, we give a wrong proof
: Since χA(X ) =
det(A − X Idn ), then χA(A) = det(A − AIdn ) = det(A − A) = det 0M n (K) = 0. Can you see
what’s wrong with this proof?
Proof of Hamilton–Cayley Theorem for matrices. Consider the matrix adj(A − X Idn ).
Clearly, all its coefficients are elements of Kn−1 [X ]. Hence, we can write
n−1
X
adj(A − X Idn ) = Bk X k
k=0
 4. HAMILTON–CAYLEY THEOREM 93

for some Bk ∈ M n (K). Now from Theorem 5.27 we have:

(A − X Idn ) adj(A − X Idn ) = det(A − X Idn ) Idn = χA(X ) Idn .

Hence,
‚ n−1 Œ n−1 n−1
X X X
k
χA(X ) Idn = (A − X Idn ) Bk X = ABk X k − Bk X k+1
k=0 k=0 k=0
‚ n−1 Œ
X
ABk − Bk−1 X − Bn−1 X n .

k
= AB0 +
k=1

Now, if we call a0 , . . . , an the coefficients of χA, we have

n
X
χA(X ) = ak X k ,
k=0

hence
n
X
χA(X ) Idn = ak X k Idn ,
k=0

and by identification we get:

a Id = AB0

 0 n

ak Idn = ABk − Bk−1 , k ∈ {1, . . . , n − 1}
a Id = −B .

n n n−1

Now,
n
X
χA(A) = ak Ak
k=0
n
X
= Ak (ak Idn )
k=0
‚ n−1 Œ
X
= AB0 + A (ABk − Bk−1 ) − An Bn−1
k

k=1
‚ n−1 Œ ‚ n−1 Œ
X X
k+1
= AB0 + A Bk − k
A Bk−1 − An Bn−1
k=1 k=1
‚ n−1 Œ ‚ n Œ
X X
k+1 k
= A Bk − A Bk−1 = 0M n (K) .
k=0 k=1

Remark 6.20. If a matrix A (or an endomorphism) is diagonalizable, Hamilton–Cayley

Theorem is much easier to prove. See Exercise 6.3.

As an application of Hamilton–Cayley Theorem, we have the following result, that shows

how the inverse of an invertible matrix is a polynomial of that matrix: Let n ≥ 2 and
A ∈ M n (R) be an invertible matrix with characteristic polynomial χA, say
n
X
χA(X ) = ak X k .
k=0
94 6. DIAGONALIZATION OF ENDOMORPHISMS

Since det(A) = χA(0) = a0 6= 0, and since, by Hamilton–Cayley Theorem, χA(A) = 0M n (K) ,

we have:
n−1
1 X
A−1 = − a Ak .
a0 k=0 k+1

Exercises
Exercise 6.1. Are the following matrices diagonalizable in R? in C?
     
0 −1 0 0 1 1 1 −1 2
A1 = −1 0 −1 , A2 = 1 0 1 , A3 = −2 1 −3 ,
0 2 0 1 1 0 −1 1 −2
   
1 1 0 −1 1 −1
A4 = −1 2 1 , A5 =  0 1 0  .
1 0 1 1 0 1

Determine a basis of eigenvectors for A2 .

Exercise 6.2. Let E be a vector space over R of dimension 3 and let B = (e1 , e2 , e3 ) be a
basis of E. We consider the endomorphism f of E such that its matrix in the basis B is
 
1 1 0
A = −1 3 1 .
1 −1 4

1. Is the endomorphism f diagonalizable? Determine a basis of the eigenspaces of f .

2. We set v1 = e1 + e2 , v2 = e1 + 2e2 + e3 .
a) Determine a vector v3 ∈ E such that f (v3 ) = v2 + 3v3 .
b) Show that B 0 = (v1 , v2 , v3 ) is a basis of E and explicit A0 = [ f ]B 0 .

Exercise 6.3 (Hamilton–Cayley Theorem for a diagonalizable endomorphism). Let E be

a finite-dimensional vector space over K and f ∈ L(E). Given a polynomial P ∈ K[X ], say
P (X ) = an X n + an−1 X n−1 + · · · + a1 X + a0 , we associate the endomorphism P ( f ) of E
defined as
P ( f ) = an f n + an−1 f n−1 + · · · + a1 f + a0 idE .
1. Show that every eigenvector u of f is also an eigenvector of f k (k ∈ N).
2. Show that if x is an eigenvector of f associated with the eigenvalue λ, then x is an
eigenvector of P ( f ) associated with the eigenvalue P (λ).
3. Show that if f is diagonalizable, then P ( f ) is also diagonalizable.
4. Use the previous questions to show that if f is a diagonalizable endomorphism of E
and if χ f is its characteristic polynomial, then χ f ( f ) = 0.

Exercise 6.4. We consider the matrices

   
4 0 0 4 0 0
A =  2 3 0 , D = 0 3 0 .
−1 0 1 0 0 1

1. Show that A is diagonalizable on R, and determine an invertible matrix P such that

A = P D P −1 .
 EXERCISES 95
 
a b c
2. Let Y = d e f  be a matrix in M3 (R) such that Y 2 = D.

g h i
a) Show that Y D = DY .
b) Compute the products Y D and DY , and deduce that the matrix Y is a diagonal
matrix.
c) Deduce the form of all the matrices Y ∈ M3 (R) such that Y 2 = D.
3. Determine all the matrices X ∈ M3 (R) such that X 2 = A.

Exercise 6.5 (Linear Systems of Sequences). Let u0 , v0 and w0 be three real numbers. We
consider three sequences of real numbers (un )n≥0 , (vn )n≥0 and (wn )n≥0 such that for all
n ∈ N:

 un+1 = vn + wn
v =w + u
 wn+1 = u n + v n.
n+1 n n
 
un
1. We set Un =  vn . Determine the matrix A such that for all n ∈ N, Un+1 = AUn . How
wn
can you express Un in terms of U0 ?
2. a) Diagonalize the matrix A and write the change of basis matrix P from the canonical
basis to a basis of eigenvectors of A.
b) Compute P −1 and deduce the expression of An in terms of n.
c) Deduce the expressions of un , vn and wn in terms of n, u0 , v0 and w0 .
3. Use the same method to study the system of sequences

sn+1 = −10sn − 28tn

§
tn+1 = 6sn + 16tn .

Exercise 6.6 (System of Linear Differential Equations). We are looking for three differen-
tiable functions x, y and z such that

 x 0 = 3x + y − z
(S) y 0 = x + 3y − z
 z 0 = x + y + z.

   0
x x
In the sequel we consider the matrices X =  y  and X 0 =  y 0 .
z z0
1. Show that System (S) can be written in the form X 0 = AX where A is a matrix you will
determine.
2. Diagonalize the matrix A and specify a change of basis matrix from the canonical basis
to a basis of eigenvectors of A. Denote by B the diagonal matrix you
 obtain.

u
3. We set U = P −1 X and we denote the components of U thus: U =  v . Explain why
w
U = P X . Show that System (S) is equivalent to U = B U .
0 −1 0 0

4. Determine the functions u, v and w, and deduce the solutions x, y and z of System (S).
96 6. DIAGONALIZATION OF ENDOMORPHISMS

5. Use the same method to solve the systems

x = −x + y + z
  0
 x 0 = x − 3y − 2z
y = x −y+z
 0

(1) y 0 = −5x − y + 2z (2)
z = x + y + z
0
 z 0 = −5x − 3y + 4z, 
x(0) = 1, y(0) = z(0) = 2.

Exercise 6.7. Let E be a vector space over R of dimension n. Let u and v be two endomor-
phisms of E. Show that u ◦ v and v ◦ u have the same spectrum. (Hint: Distinguish two
cases: λ = 0 and λ 6= 0)

Exercise 6.8. Determine the spectrum of a projection, of a symmetry, of a rotation in R2

and R3 and of a hyperbolic rotation in R2 . Are they diagonalizable in R? In C?

Exercise 6.9. Let E be the vector space of functions of class C ∞ on R and periodic of
period 2π. We consider the following endomorphism of E:
∆ : E −→ E
f 7−→ f 00 .
1. Briefly explain why ∆ is indeed an endomorphism of E.
2. Determine the eigenvalues (in R) of ∆ and the multiplicity of each eigenvalue.
 Review Exercises

Exercise (From DS June 2007). We consider the subset E of M3 (R) that consists of matrices
of the form  
a b c
0 a b 
0 0 a
with a, b , c ∈ R. We also consider the following matrix:
 
0 1 0
N = 0 0 1  .
0 0 0
1. For n ∈ N, compute N n . For M ∈ E, determine a general formula for M n .
2. Prove that E is a vector space, and determine a basis of E and its dimension.
3. Determine all the matrices of E that are diagonalizable.
4. Let f be the endomorphism of R3 the matrix of which in the standard basis B =
(e1 , e2 , e3 ) of R3 is  
1 −1 1
[ f ]B = M = 0 3 −1 .
0 4 −1
a) Is the endomorphism f diagonalizable?
b) We consider the following vectors of R3 :
e10 = (1, 0, 0), e20 = (−1, 1, 2), e30 = (1, 0, −1).
Prove that C = (e10 , e20 , e30 ) is a basis of R3 and determine M 0 = [ f ]C .
c) Prove that f is an automorphism.
d) For n ∈ N, determine f n in terms of idE , f and f 2 . Is this formula still valid for
n ∈ Z?

Exercise (From DS June 2007). Let n ≥ 2.

1. Give (without any justification) the dimension of M n (R).
2. We consider the subspace F of M n (R) spanned by Idn . Explicit (without any justifica-
tion) the elements of F and the dimension of F .
3. We recall that tr is the trace of an n ×n matrix, that is the sum of its diagonal coefficients.
It is well-known that tr is a linear form on M n (R).
a) Determine Im tr, Ker tr and their dimension.
b) Prove that M n (R) = Ker tr ⊕F .
4. We consider the mapping f defined thus:
f : M n (R) −→ M n (R)
M 7−→ M + tr(M ) Idn .
a) Prove that f is a well-defined endomorphism of M n (R).
b) Determine the eigenvalues and the eigenspaces of f .
97
98 REVIEW EXERCISES

c) Deduce that f is an automorphism and that f is diagonalizable.

5. Let J ∈ M n (R) be a non-zero matrix such that tr(J ) = 0. Let g be the mapping defined
thus:
g : M n (R) −→ M n (R)
M 7−→ M + tr(M )J .
You are given that g is an endomorphism of M n (R).
a) Determine g 2 and prove that g 2 = 2 g − id.
b) Prove that if λ ∈ spec g , then λ = 1.
c) Prove that 1 is indeed an eigenvalue of g , and describe the associated eigenspace.
d) Is the endomorphism g diagonalizable?

Problem (From  Final,

 EurINSA 2004). In this exercise, we are interested in the column
un
vectors Xn =  vn  such that Xn+1 = AXn + B, where A ∈ M3 (R) and B ∈ M3,1 (R)
wn
are given. We consider the standard basis B = (e1 , e2 , e3 ) of E, and f ∈ L(E) such that
[ f ]B = A. We denote by I the identity matrix in M3 (R).
Part I
‚ n−1 Œ
X
1. Prove that for all n ∈ N∗ , Xn = An X0 + Ak B.
k=0
n−1
X
2. Show that if the matrix (I − A) is invertible, then ∀n ∈ N∗ , Ak = I − An (I − A)−1 .

k=0

Part II  
4 −1 −1
In this Part only we consider the special case A = 2 1 −1.
4 −2 0
1. Compute and factor the characteristic polynomial p of A. Deduce that I − A is not
invertible.
2. Show that A is diagonalizable in M3 (R) and give its diagonal form D, as well as a relation
between A and D.
n−1
X
3. For n ∈ N compute D n and Dk.
k=0
n−1
X
4. For n ∈ N∗ , determine an explicit expression of An and Ak .
k=0
5. We consider the sequences (un )n∈N , (vn )n∈N and (wn )n∈N satisfying:

 un+1 = 4un − vn − wn + 1

∀n ∈ N, vn+1 = 2un + vn − wn
n+1 = 4un − 2vn + 1

w

and subject to the initial conditions

u0 = 1, v0 = 0, w0 = 0.
Determine an explicit expression of un , vn and wn in terms of n only. Check your
answer on the first terms of these sequences.
 REVIEW EXERCISES 99

Part III  
5 −6 3
In this Part only we consider the special case A = 2 −2 2.
1 −2 3
1. Show that A is not diagonalizable.  
2 0 0
2. Find a basis B 0 = (e10 , e20 , e30 ) of E such that [ f ]B 0 = T = 0 2 1, and give a relation
0 0 2
between A and T .
3. Show that
∀n ∈ N∗ , f n = 2n idE +2n−1 n( f − 2 idE ).
Deduce that for all n ∈ N∗ , An = 2n I + 2n−1 n(A − 2I ).
n−1
X
4. Show that A is invertible and deduce an explicit expression of Ak .
k=0
5. We consider the sequences (un )n∈N , (vn )n∈N and (wn )n∈N satisfying:

 un+1 = 5un − 6vn + 3wn + 1

∀n ∈ N, vn+1 = 2un − 2vn + 2wn + 1
n+1 = un − 2vn + 3wn + 1

w

and subject to the initial conditions

u0 = 1, v0 = 1, w0 = 1.
Determine an explicit expression of un , vn and wn in terms of n only. Check your
answer on the first terms of these sequences.

Exercise (Fibonacci Sequence). Determine a closed formula for the Fibonacci Sequence
(Fn )n∈N defined thus:
F0 = 0, F1 = 1, ∀n ∈ N, Fn+2 = Fn+1 + Fn .

Exercise (From Final, EurINSA 2005). We consider the matrices

   
1 0 1 2 1 0
A = 2 4 −2 and T = 0 2 0 .
3 4 −1 0 0 0
1. Show that A is not diagonalizable.
2. Show that A and T are similar matrices, and find a relation between A and T (you will
explicit the change of basis matrix P ).
3. For n ∈ N, compute T n and determine An .
4. We consider the sequences (un )n∈N , (vn )n∈N and (wn )n∈N satisfying:

 un+1 = un − +wn

∀n ∈ N, vn+1 = 2un + 4vn − 2wn
n+1 = 3un + 4vn − wn

w

and subject to the initial conditions

u0 = 1, v0 = 0, w0 = 0.
Determine an explicit expression of un , vn and wn in terms of n only. Check your
answer on the first terms of these sequences.
100 REVIEW EXERCISES

Problem (From Final, June 2010).

In this problem we only consider matrices with real coefficients.

Part I
Let n ≥ 2 be an integer, and A ∈ M n (R). We say that a matrix R ∈ M n (R) is a square root
of A if R2 = A. We denote by Sqrt(A) the set of square roots of A.

1. a) Let R ∈ Sqrt(A). What relation exists between det R and det A?

b) What can you deduce about Sqrt(A) when det A < 0?
c) If the matrix A is invertible, what can be said about its square roots, if they exist?
2. Show that if R is a square root of A, then AR = RA.
3. Let c1 , c2 , . . . , cn be distinct real numbers, and let C be the following diagonal n × n
matrix:  
c1 0 0 · · · 0
 0 c2 0 · · · 0 
 
C =  0 0 c3 · · · 0  .
 
 .. .. .. . . .. 
. . . . .
0 0 0 · · · cn
a) Let S ∈ M n (R), of general term si j . Let V = C S and W = SC .
i) Compute the general term vi j of V and the general term wi j of W .
ii) Deduce using Question 2 that if S is a square root of C then S is a diagonal
matrix.
iii) What is the relation between the diagonal coefficients of S and the diagonal
coefficients of C ?
b) Deduce that if at least one of the ci is negative, then C doesn’t admit a square root.
c) From now on we suppose that all the ci ’s are non-negative. Determine all the square
roots of C . What is the maximal number of square roots of C ?
4. Let A and B be two similar matrices, say A = QBQ −1 with Q ∈ M n (R) invertible.
a) Show that a matrix R is a square root of A if and only if Q −1 RQ is a square root of
B.
b) Deduce that Sqrt(A) and Sqrt(B) have the same number of elements (possibly an
infinite number of elements).
5. Let A ∈ M n (R) be a matrix that admits n distinct real eigenvalues, say λ1 < λ2 < · · · < λn .
a) Show that A and D are similar matrices, where D is the following matrix:
λ1 0 0 · · · 0
 
 0 λ2 0 · · · 0 
 0 0 λ ··· 0 
 
3 .
 .. .. .. . . .. 

. . . . . 
0 0 0 · · · λn
b) Deduce a necessary and sufficient condition on the λi ’s for A to admit at least one
square root.
Part II
For a real number a, we set Ba to be the following matrix of M3 (R):
1 − 4a −1 + 4a
 
1
Ba = −3a −1 + 2a 2 + a .
−3a −2 − a 3 + 4a
 REVIEW EXERCISES 101
 
1
1. Show that Y = 1 is an eigenvector of Ba and specify to which eigenvalue λ1 it is

1
associated.
2. In this question only we assume that a = 0.
a) Determine the spectrum of B0 .
b) Is the matrix B0 diagonalizable?
3. In this question only we assume a 6= 0.
a) Let Na = Ba − (1 + 3a)I3 . Determine the rank of the matrix Na (you will distinguish
two cases depending on the value of a).
b) What eigenvalue λ2 is hence discovered? Give the dimension of the eigenspace asso-
ciated to the eigenvalue λ2 (depending on the value of a).
c) What is the sum of the eigenvalues of Ba , counted with their multiplicities? Deduce
that the multiplicity of the eigenvalue λ2 is 2.
d) For what value(s) of a is the matrix Ba diagonalizable?
e) For what value(s) of a is the matrix Ba invertible?
4. In this question only we assume that a = 1.
a) Determine an invertible matrix P ∈ M3 (R) and a diagonal matrix D1 ∈ M3 (R) such
that B1 = P D1 P −1 (choose P so that its last column consists of 1’s only).
0 α
 ‹
b) For (α, β) ∈ R2 compute the square of the matrix K = .
β 0
c) Deduce that the matrix Z = 4I2 possesses an infinite number of square roots in M2 (R).
d) Deduce that the matrix B1 possesses an infinite number of square roots in M3 (R).

Problem (From Final, June 2011).

Part I
We consider the vector space E =R3 . Let f be the endomorphism of E, the matrix of
0 −2 2
which in the standard basis is A = 1 −2 1 .
0 −2 2
1. Compute A2 and A3 .
2. We consider the following vectors of E:
w1 = (0, 0, 1), w2 = (2, 1, 2), w3 = (2, 2, 2).
Prove that the family B = (w1 , w2 , w3 ) is a basis of E.
3. Determine [ f ]B , the matrix of f in the basis B.
4. Is f diagonalizable?
Part II
Let E be a real vector space of dimension 3, together with a basis B0 . The null endomor-
phism of E is simply denoted by 0. We consider an endomorphism f of E that satisfies
f 2 6= 0 and f 3 = 0.
1. Give an example of such an endomorphism.
2. Let a ∈ E such that f 2 (a) 6= 0E . Show that the family B1 = a, f (a), f 2 (a) is an

independent family. Deduce that B1 is a basis of E.

3. Determine [ f ]B1 , the matrix of f in the basis B1 .
4. Is f diagonalizable?
Part III
102 REVIEW EXERCISES

Let Q ∈ R3 [X ] of the form Q = X 3 + a2 X 2 + a1 X + a0 . We define the matrix CQ ∈ M3 (R)

by
 
0 1 0
CQ =  0 0 1 .
−a0 −a1 −a2
1. Determine the characteristic polynomial of CQ .
2. Give a necessary and sufficient condition on a0 , a1 and a2 for the matrix CQ to be
invertible.
3. Let λ be an eigenvalue of CQ and let Eλ be the associated eigenspace. Evaluate dim Eλ .
4. Deduce that CQ is diagonalizable in M3 (R) if and only if Q admits three distinct real
roots.
5. We assume, in this question only, that Q has three distinct real roots λ1 , λ2 and λ3 .
a) Construct a basis of eigenvectors (v1 , v2 , v3 ) such that the first component of each of
these vectors is 1.
b) Deduce that the determinant
1 1 1
λ1 λ2 λ3
λ21 λ22 λ23
is not zero.
λ1 0 0
 

c) We consider the matrix DQ =  0 λ2 0 . Give a relation between DQ and CQ .

0 0 λ3
d) If n ∈ N, how would you compute CQn ? (don’t give the explicit computation).

Part IV
We now consider the real vector space RN that consists of real-valued sequences indexed
by N. We denote by F the set of all sequences (un )n∈N ∈ RN such that
∀n ∈ N, un+3 = 2un+2 + un+1 − 2un .
We denote by T ∈ R3 [X ] the polynomial T = X 3 − 2X 2 − X + 2.
1. Show that F is a linear
 subspace of R .
N

un
2. We set Xn =  un+1 . Give a relation between Xn+1 and Xn . Deduce a relation between
un+2
Xn and X0 .
3. Write the matrix CT (see the previous part). Compute CTn explicitly.
4. Deduce an expression of un in terms of n, u0 , u1 and u2 only.
5. What is the dimension of F ?

Problem (From Final, June 2014). Let n ∈ N, n ≥ 2.

Definition. We say that two matrices A and B in M n (R) are similar if there exists an
invertible matrix P ∈ M n (R) such that A = P B P −1 , and in this case we’ll write A ∼ B.

Note that if A ∼ B then B ∼ A.

We denote by E the set of all invertible matrices of M n (R) that are similar to their inverse,
i.e.,
E = A ∈ M n (R) A is invertible and A ∼ A−1 .

 REVIEW EXERCISES 103

The identity matrix of M n (R) is denoted by I .

Part I – A non-diagonalizable case
In this part only we assume that A = I + N where
 
0 1 0
N = 0 0 1  .
0 0 0
1. Explain why A is not diagonalizable.
2. For k ∈ N∗ , compute N k .
3. Show that A is invertible and that A−1 = I − N + N 2 .
4. Show that A ∈ E if and only if there exists an invertible matrix P such that
(∗) N P + P N = P N 2.
5. Show that A ∈ E by explicitly determining all the invertible matrices P ∈ M3 (R) such
that P −1 AP = A−1 . You may consider the columns C1 , C2 and C3 of P , or any other
method (e.g., a linear system involving the coefficients of P ).
Part II
Let α ∈ R and
   
1/2 0 0 3 0 −4
A1 =  4 1 −1 , A2 = 0 α 0  .
2 −2 0 0 0 −1
1. Show that A1 is diagonalizable and invertible.
2. Show that A2 is diagonalizable (you will distinguish 3 cases).
Part III – The diagonalizable case
In this part, A ∈ M n (R) is an invertible matrix and f is the endomorphism of Rn such that
[ f ]std = A. For an eigenvalue λ ∈ R of f , we denote by Eλ the eigenspace of f associated
with λ. The set of all eigenvalues of f is denoted by spec f .
1. Explain why 0 6∈ spec f .
2. Show that A ∈ E if and only if there exists a bijective endomorphism g of Rn such that
f ◦ g ◦ f = g.
3. In this question only we assume that A is an element of E . We denote by g a bijective
endomorphism of Rn such that f ◦ g ◦ f = g .
a) Let λ ∈ spec f and let u be an eigenvector of f associated with λ. Show that 1/λ is
an eigenvalue of f , and that g (u) is an eigenvector of f associated with 1/λ.
b) Deduce that: ∀λ ∈ spec f , dim Eλ = dim E1/λ .
4. In this question only we assume that the endomorphism f is diagonalizable and that:
 
∀λ ∈ spec f , 1/λ ∈ spec f and dim Eλ = dim E1/λ = 1 .

Moreover, for all λ ∈ spec f we choose an eigenvector uλ of f associated with λ.

a) Explain why there exists a unique endomorphism h of Rn such that
∀λ ∈ spec f , h(uλ ) = u1/λ .

b) Explain why h is invertible (hint: determine h 2 ).

c) Show that f ◦ h ◦ f = h.
d) Deduce that A ∈ E .
104 REVIEW EXERCISES

5. Let  
2 1 1
A3 = 1 2 1 .
1 1 2
Show that A3 6∈ E .
6. Show that A1 ∈ E and determine the values of α for which A2 ∈ E . Note: the matrices
A1 and A2 are defined in Part II. You’ll get a bonus credit if you can explicitly determine
an invertible matrix P such that A−11
= P −1 A1 P .
7. Bonus question: determine a necessary and sufficient condition for a diagonalizable
matrix A ∈ M n (R) to belong to E . (No justification required).
 Bibliography

[Artin, 1991] Artin, M. (1991). Algebra. Englewood Cliffs, NJ: Prentice Hall Inc.
[Grifone, 1990] Grifone, J. (1990). Algèbre linéaire. Toulouse: Cépaduès Éditions.
[Lang, 1987] Lang, S. (1987). Linear algebra. Undergraduate Texts in Mathematics. New York: Springer-
Verlag, third edition.
[Roman, 2008] Roman, S. (2008). Advanced linear algebra, volume 135 of Graduate Texts in Mathematics.
New York: Springer, third edition.

105
 Index

1 − 1 linear map, 36 Generating family of vectors, 19

Grassmann Formula, 23
Adjugate, 83
Automorphism, 36 Hamilton–Cayley Theorem, 92
Homomorphism, 35
Basis of a vector space, 21
Bijective linear map, 36 Identity
map, 37
Canonical basis, 22 matrix, 60
Cayley–Hamilton Theorem, 92 Image of a linear map, 40
Change of bases of linear maps, 63 Inclusion–Equality Theorem, 26
Change of basis matrix, 62 Independent
Characteristic polynomial, 89 family of vectors, 18
Codimension of a subspace, 47 subspaces, 16
Cofactor, 83 Injective linear map, 36
Comatrix, 83 Inverse of a matrix (formula), 84
Commutative field, 9 Invertible matrix, 61
Complementary pair of subspaces, 18 Isomorphism, 36
Complex structure on a real vector space, 33
Coordinates matrix, 62 Kernel of a linear map, 40
Coordinates of a vector, 23
Laplace formula, 83
Dependent family of vectors, 18 Leibniz formula, 82
Determinant Linear combination, 14
of a family of vectors, 78 Linear form, 36
of a square matrix, 73 Linear map, 35
of an endomorphism, 77 Linear span, 14
Diagonalization, 90 Linear subspace, 13
Dilation matrix, 65 Linear system, 1, 64
Dimension of a vector space, 22 of differential equations, 95
Direct sum of subspaces, 18 of sequences, 95
Dual
Matrices, 55
basis, 50
Matrix of a linear map, 43
family, 48
Monomorphism, 36
of a linear map, 48
space, 37, 48 Natural basis, 22
Eigenspace, 87 One-to-one linear map, 36
Eigenvalue, 87 Onto linear map, 36
Eigenvector, 87
Elementary Permutation, 79
matrix, 64 Polynomial of an endomorphism, 45
row operations, 2, 65 Powers of an endomorphism, 45
Endomorphism, 36 Product of matrices, 57
Epimorphism, 36 Projection, 51

Gaussian elimination, 3 Quaternions, 69

108 INDEX

Rank
of a family of vectors, 27
of a linear map, 41
of a linear system, 6
of a matrix, 61
Rank–Nullity Theorem, 41

Shear transformation matrix, 65

Signature of a permutation, 80
Similar matrices, 64
Span, 14
Square root of a matrix, 100
Standard basis, 22
Subspace, 13
Sum
of matrices, 56
of subspaces, 15
Surjective linear map, 36
Symmetric matrix, 62
Symmetry, 51

Trace
of a matrix, 78
of an endomorphism, 78
Transposed of a matrix, 62
Transposition matrix, 64
Transvection matrix, 65

Vector space, 9