New Format

Elements of Linear Algebra
Peter Ioan Radu
L
aszl
o Szil
ard Csaba
Cluj-Napoca
2014
Viorel Adrian
Contents
Introduction
iii
1 Matrices
1.1 Basic denitions, operations and properties. . . . . . . . . . . . . . .
1.2 Determinants and systems of linear equations . . . . . . . . . . . . . 13

1.3 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2 Vector Spaces
28
2.1 Denition of Vector Space and basic properties . . . . . . . . . . . . . 28

2.2 Subspaces of a vector space . . . . . . . . . . . . . . . . . . . . . . . 29
2.3 Basis. Dimension. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.4 Local computations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
2.5 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3 Linear maps between vector spaces
53
3.1 Properties of LpV, W q . . . . . . . . . . . . . . . . . . . . . . . . . . . 57

3.2 Local form of a linear map . . . . . . . . . . . . . . . . . . . . . . . . 63
3.3 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
CONTENTS
4 Proper vectors and the Jordan canonical form
ii
72
4.1 Invariant subspaces. Proper vectors and values . . . . . . . . . . . . . 72

4.2 The minimal polynomial of an operator . . . . . . . . . . . . . . . . . 76
4.3 Diagonal matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
4.4 The Jordan canonical form . . . . . . . . . . . . . . . . . . . . . . . . 86
4.5 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
5 Inner product spaces
95
5.1 Basic denitions and results . . . . . . . . . . . . . . . . . . . . . . . 95

5.2 Orthonormal Bases . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
5.3 Orthogonal complement . . . . . . . . . . . . . . . . . . . . . . . . . 105
5.4 Linear manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
5.5 The Gram determinant. Distances. . . . . . . . . . . . . . . . . . . . 113
5.6 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
6 Operators on inner product spaces.
126
6.1 Linear functionals and adjoints . . . . . . . . . . . . . . . . . . . . . 126

6.2 Normal operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
6.3 Isometries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
6.4 Self adjoint operators . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
6.5 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
7 Elements of geometry
145
7.1 Quadratic forms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145

7.2 Quadrics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
7.3 Conics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
7.4 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
CONTENTS
Bibliography
iii
155
Introduction
The aim of this book is to give an introduction in linear algebra, and at the same
time to provide some applications that might be useful both in practice and theory.
Hopefully this book will be a real help for graduate level students to understand
the basics of this beautiful mathematical eld called linear algebra. Our scope is
twofold: one hand we give a theoretical introduction to this eld, which is more
than exhaustive for the need and understanding capability of graduate students, on
the other hand we present fully solved examples and problems, that might be helpful in preparing to exams and also show the techniques used in the art of problem
solving on this eld. At the end of every chapter, this work contains several proposed problems that can be solved using the theory and solved examples presented
previously. The book is structured on seven chapters, we begin with the study of
matrices and determinants and we end with the study of quadratic forms with some
useful applications in analytical geometry.
Our hope is that our students, and not just them, will enjoy studying this book
and they will nd very helpful in preparing for exams.
iv
1
Matrices
1.1
Basic denitions, operations and properties.
Denition 1.1. A matrix of dimension m n with elements in a eld F, (where

usually F R, or F C), is a function A : t1, . . . , mu t1, . . . , nu F,
Api, jq aij P F, @i P t1, 2, . . . , mu, j P t1, 2, . . . , nu.
Usually an m n matrix is represented
a12
a
11
a21 a22
A
..
..
.
.
am1 am2
as a table with m lines and n columns:
. . . a1n
. . . a2n
.. .
.
...
. . . amn
Hence, the elements of a matrix A are denoted by aij , where aij stands for the
number that appears in the ith row and the j th column of A (this is called the pi, jq
entry of A) and the matrix is represented as A paij qi1,m .
j1,n
We will denote the set of all m n matrices with entries in F by Mm,n pFq
respectively, when m n by Mn pFq. It is worth mentioning that the elements of
1
Mn pFq are called square matrices. In what follows, we provide some examples.
Example 1.2. Consider the matrices
1 2 3
i 2ì
0
,
A 4 5 6 , respectively B
?
2 1 ` 3i
3
7 8 9
where i is the imaginary unit. Then A P M3 pRq, or in other words, A is a real valued
square matrix, meanwhile B P M2,3 pCq, or in other words, B is a complex valued
matrix with two rows and three columns.
In what follows we present some special matrices.
Example 1.3. Consider the matrix In paij qi,j1,n P Mn pFq, aij 1, if i
j and aij 0 otherwise. Here 1 P F, respectively 0 P F are the multiplicative
identity respectively the zero element of the eld F.
Then
0
In
..
.
0 ... 0
1 . . . 0
..
..
. . . . .
0 ... 1
and is called the identity matrix (or unit matrix ) of order n.

Remark 1.4. Sometimes we denote the identity matrix simply by I.
Example 1.5. Consider the matrices O paij qi1,m P Mm,n pFq having all entries
the zero element of the eld F. Then
0
O
..
.
j1,n
0 ... 0
0 . . . 0
..
..
. . . . .
0 ... 0
and is called the null matrix of order m n.

Example 1.6. Consider the matrices A paij qi,j1,m P Mn pFq given by aij 0
whenever i j, respectively
a
a
...
11 12
0 a22 . . .
A
..
..
.
. ...
0
0 ...
aij 0 whenever i j. Then
a
a1n
0
11
a
a2n
a
, respectively A 21 22
..
..
.
..
.
.
an1 an2
ann
...
... 0
..
.
...
. . . ann
is called upper triangular, respectively lower triangular matrix.

If all entries outside the main diagonal are zero, A is called a diagonal matrix.
In this case we have
a
0
11
0 a22
A
..
..
.
.
0
0
...
...
...
0
0
..
.
. . . ann
Addition of Matrices.
If A and B are m n matrices, the sum of A and B is dened to be the m n
matrix A ` B obtained by adding corresponding entries. Hence, the addition
operation is a function
` : Mm,n pFq Mm,n pFq Mm,n pFq,
paij qi1,m ` pbij qi1,m paij ` bij qi1,m , @ paij qi1,m , pbij qi1,m P Mm,n pFq.
j1,n
j1,n
j1,n
j1,n
In other words, for A, B P Mm,n pFq their sum is dened as

C A ` B pcij qi1,m
j1,n
where cij aij ` bij for all i P t1, 2, . . . , mu, j P t1, 2, . . . , nu.
j1,n
Properties of Matrix Addition.

Let O P Mm,n pFq the null matrix of size m n. For a given matrix
X pxij qi1,m P Mm,n pFq we denote by X its additive inverse (opposite) that is
j1,n
X pxij qi1,m P Mm,n pFq. For every A, B, C P Mm,n pFq the following properties
hold:
j1,n
1. A ` B is again an m n matrix (closure property).

2. pA ` Bq ` C A ` pB ` Cq (associative property).
3. A ` B B ` A (commutative property).
4. A ` O O ` A A (additive identity).
5. A ` pAq pAq ` A O (the additive inverse).
It turns out that pMm,n pFq, `q is an Abelian group.
Scalar multiplication.
For A P Mm,n pFq and P F dene A paij qi1,m . Hence, the scalar
multiplication operation is a function
j1,n
: F Mm,n pFq Mm,n pFq,

paij qi1,m p aij qi1,m , @ P F, paij qi1,m , P Mm,n pFq.
j1,n
j1,n
j1,n
Properties of Scalar multiplication.

Obviously, for every A, B P Mm,n pFq and , P F the following properties hold:
1. A is again an m n matrix ( closure property).
2. pqA pAq (associative property).
3. pA ` Bq A ` B (distributive property).
4. p ` qA A ` A (distributive property).
5. 1A A, where 1 is the multiplicative identity of F (identity property).
Of course that we listed here only the left multiplication of matrices by scalars. By
dening A A we obtain the right multiplication of matrices by scalars.
1 0 2
1 1 1
Example 1.7. If A 0
2 1 and B 1 1 1 , then
0 1 2
2 2
0
3 2 0
1 2 4
2A B 1 5 3 and 2A ` B 1
3 1 .
4 5 2
4 3
2
Transpose.
The transpose of a matrix A P Mm,n pFq is dened to be a matrix AJ P Mn,m pFq
obtaining by interchanging rows and columns of A. Locally, if A paij qi1,m , then
j1,n
AJ paji q j1,n .
i1,m
It is clear that pAJ qJ A. A matrix, that has many columns, but only one row, is
called a row matrix. Thus, row matrix A with n columns is an 1 n matrix, i.e.
A pa1 a2 a3 . . . an q.
A matrix, that has many rows, but only one column, is called a column matrix.
Thus, a column matrix A with m rows is an m 1 matrix, i.e.
a1
a2
A .
.
..
am
Obviously, the transpose of a row matrix is a column matrix and viceversa, hence,
in inline text a column matrix A is represented as
A pa1 a2 . . . am qJ .
Conjugate Transpose.
Let A P Mm,n pCq. Dene the conjugate transpose of A paij qi1,m P Mm,n pCq by
j1,n
A paji q j1,n , where z denotes the complex conjugate of the number z P C. We

i1,m
have that pA q A and AJ A whenever A contains only real entries.

Properties of the Transpose.
For every A, B P Mm,n pFq and P K hold:
1. pA ` BqJ AJ ` B J .
2. pA ` Bq A ` B .
3. pAqJ AJ and pAq A .
Symmetries.
Let A paij q i1,n P Mn pFq be a square matrix. We recall that
j1,n
A is said to be a symmetric matrix whenever A AJ (locally aij aji for all

i, j P t1, 2, . . . nu).
A is said to be a skew-symmetric matrix whenever A AJ (locally
aij aji for all i, j P t1, 2, . . . nu).
A is said to be a hermitian matrix whenever A A (locally aij aji for all
i, j P t1, 2, . . . nu).
A is said to be a skew-hermitian matrix whenever A A (locally

aij aji for all i, j P t1, 2, . . . nu).
It can easily be observed, that every symmetric real matrix is hermitian,
respectively, every skew-symmetric real matrix is skew-hermitian.
1 2 4
Example 1.8. The matrix A 2 0 3 is a symmetric matrix,
4
3 2
0
1 3
meanwhile the matrix B 1 0

3 is a skew-symmetric matrix.
3 3 0
1
1ì
i
The matrix C 1 i
3
3 2i is a hermitian matrix, meanwhile the
i 3 ` 2i
2
i
2i
3i
matrix D 2 i
i
2 ` 3i is a skew-hermitian matrix.
3i 2 ` 3i
0
Matrix multiplication.
For a matrix X pxij qi1,m P Mm,n pFq we denote by Xi its ith row, i.e. the row
matrix
j1,n
Xi pxi1 xi2 . . . xin q.

Similarly, the j th column of X is the column matrix
Xj px1j x2j . . . xmj qJ .
It is obvious that
pX J qi pXi qJ ,
respectively
pX J qj pXj qJ .
We say that the matrices A and B are conformable for multiplication in the order
AB, whenever A has exactly as many columns as B has rows, that is A P Mm,p pFq
and B P Mp,n pFq.
For conformable matrices A paij qi1,m and B pbjk q j1,p the matrix product AB
j1,p
is dened to be the m n matrix C pcik qi1,m with
k1,n
k1,n
cik Ai Bk
aij bjk .
j1
In the case that A and B failed to be conformable, the product AB is not dened.
Remark 1.9. Note, the product is not commutative, that is, in general,
AB BA even if both products exists and have the same shape.
1 1
1 0 1
and B 0
Example 1.10. Let A
1 .
1 1 0
1 1
2 1 1
2 0
and BA
Then AB
1 1
0 .
1 2
0
1
1
Rows and columns of a product.
Suppose that A paij qi1,m P Mm,p pFq and B pbij q i1,p P Mp,n pFq.
j1,p
j1,n
There are various ways to express the individual rows and columns of a matrix
product. For example the ith row of AB is
Ci rABsi rAi B1 Ai B2 . . . Ai Bn s Ai B

B
1
B2
ai1 ai2 . . . aip

..
.

Bp
There are some similar representations for individual columns, i.e. the j th column
is
Cj rABsj rA1 Bj A2 Bj . . . Am Bj sJ ABj

b
1j
b2j
A1 A2 . . . Ap
..
.

bpj
Consequently, we have:
1. rABsi Ai B
pith
row of AB).
2. rABsj ABj
pj th
column of AB).
3. rABsi ai1 B1 ` ai2 B2 ` ` ai pBp
aik Bk .
4. rABsj A1 b1j ` A2 b2j ` ` Ap bpj
Ak bkj .
k1
k1
The last two equations has both theoretical and practical importance. They shows
that the rows of AB are combinations of rows of B, while columns of AB are
combinations of columns of A. So it is waisted time to compute the entire product
when only one row or column is needed.
10
Properties of matrix multiplication.

Distributive and associative laws.
For conformable matrices one has:
1. ApB ` Cq AB ` AC (left-hand distributive law).
2. pB ` CqA BA ` CA (right-hand distributive law).
3. ApBCq pABqC (associative law).
For a matrix A P Mn pFq, one has
AIn A and In A A ,
where In P Mn pFq is the identity matrix of order n.
Proposition 1.11. For conformable matrices A P Mm,p pFq and B P Mp,n pFq, on
has
pABqJ B J AJ .
The case of conjugate transposition is similar:
pABq B A .
Proof. Let C pcij q i1,n pABqJ . Then for every
j1,m
i P t1, 2, . . . , nu, j P t1, 2, . . . , mu one has cij rABsji Aj Bi . Let us consider

now the pi, jq entry of B J AJ .
J
rB A sij pB qi pA qj pBi q pAj q
k1
bki ajk
Aj Bi
k1
rB J sik rAJ skj
k1
ajk bki
11
Exercise. Prove, that for every matrix A paij qi1,m P Mm,n pFq the matrices
j1,n
AAJ and AJ A are symmetric matrices.
For a matrix A P Mn pFq, one can introduce its mth power by

A0 In , A1 A, Am Am1 A.
1 0
0 1
0 1
, A3
then A2
Example 1.12. If A
0 1
1 0
1 0
1 0
I2 . Hence Am Amp mod q4 .
and A4
0 1
Trace of a product. Let A be a square matrix of order n. The trace of A is the
sum of the elements of the main diagonal, that is
trace A
aii .
i1
Proposition 1.13. For A P Mm,n pCq and B P Mn,m pCq one has
trace AB trace BA.
Proof. We have
trace AB
rABsii
i1
m
n
i1 k1
bki aik
pAqi pBqi
i1
n
m
k1 i1
bki aik
m
n
aik bki
i1 k1
n
rBAskk trace BA.
k1
Block Matrix Multiplication.

Suppose that A and B are partitioned into submatrices-referred to as blocks- as
indicated below:
A
A12
11
A21 A22
A
..
..
.
.
As1 As2
B
A1r
B12
11
B
A2r
B22
and B 21
..
..
..
.
.
.
Br1 Br2
Asr
12
...
...
...
...
...
B1t
. . . B2r
..
.
...
. . . Brt
We say that the partitioned matrices are conformable partitioned if the pairs
pAik , Bkj q are conformable matrices, for every indices i, j, k. In this case the
product AB is formed by combining blocks exactly the same way as the scalars are
combined in ordinary matrix multiplication. That is, the pi, jq block in the
product AB is
Ai1 B1j ` Ai2 B2j ` . . . Air Brj .
Matrix Inversion.
For a square matrix A P Mn pFq, the matrix B P Mn pFq that satises
AB In and BA In
(if exists) is called the inverse of A and is denoted by B A1 . Not all square
matrices admits an inverse (are invertible). An invertible square matrix is called
nonsingular and a square matrix with no inverse is called singular matrix.
Although not all matrices are invertible, when an inverse exists, it is unique.
Indeed, suppose that X1 and X2 are both inverses for a nonsingular matrix A.
Then
X1 X1 In X1 pAX2 q pX1 AqX2 In X2 X2
which implies that only one inverse is possible.
Properties of Matrix Inversion. For nonsingular matrices A, B P Mn pFq, the
following statements hold.
Determinants and systems of linear equations
13
1. pA1 q1 A
2. The product AB is nonsingular.
3. pABq1 B 1 A1 .
4. pA1 qJ pAJ q1 and pA1 q pA q1 .
One can easily prove the following statements.
Products of nonsingular matrices are nonsingular.
If A P Mn pFq is nonsingular, then there is a unique solution X P Mn,p pFq for the
equation
AX B, where B P Mn,p pFq,
and the solution is X A1 B.
A system of n linear equations in n unknowns can be written in the form Ax b,
with x, b P Mn,1 pFq, so it follows when A is nonsingular, that the system has a
unique solution x A1 b.
1.2
Determinants and systems of linear

equations
Determinants.
For every square matrix A paij q i1,n P Mn pFq one can assign a scalar denoted
j1,n
detpAq called the determinant of A. In extended form we write
a11 a12 a1n
a21 a22 a2n
detpAq .
..
..
.. .
.
.
.
.
.
an1 an2 ann
14
In order to dene the determinant of a square matrix, we need the following

notations and notions. Recall that by a permutation of the integers t1, 2, ..., nu we
mean an arrangement of these integers in some denite order. In other words, a
permutation is a bijection : t1, 2, . . . , nu t1, 2, . . . , nu. It can easily be observed
that the number of permutations of the integers t1, 2, ..., nu equals
n! 1 2 . . . n. Let us denote by Sn the set of all permutations of the integers
t1, 2, ..., nu. A pair pi, jq is called an inversion of a permutation P Sn is i j and
piq pjq. A permutation P Sn is called even or odd according to whether the
number of inversions of is even or odd respectively. The sign of a permutation
P Sn , denoted by sgn pq, is `1 if the permutation is even and 1 if the
permutation is odd.
Denition 1.14. Let A P Mn pFq. The determinant of A is the scalar dened by
the equation
detpAq
sgn pq a1p1q a2p2q . . . anpnq .
PSn
It can easily be computed, that for A paij q i1,2 P M2 pFq, one has
j1,2
detpAq a11 a22 a12 a21 .

Similarly, if A paij q i1,3 P M3 pFq, then its determinant can be calculated by the
rule
j1,3
detpAq
a11 a22 a33 ` a13 a21 a32 ` a12 a23 a31 a13 a22 a31 a11 a23 a32 a12 a21 a33 .
1 2 3
Example 1.15. If A 4 5 6 then
7 8 9
detpAq 1 5 9 ` 3 4 8 ` 2 6 7 3 5 7 1 6 8 2 4 9 0.
15
Laplaces theorem.
Let A P Mn pFq and let k be an integer, 1 k n. Consider the rows i1 . . . ik and
the columns j1 . . . jk of A. By deleting the other rows and columns we obtain a
submatrix of A of order k, whose determinant is called a minor of A and is
...jk
denoted by Mij11...i
. Now let us delete the rows i1 . . . ik and the columns j1 . . . jk of
k
A.. We obtain a submatrix of A of order n k. Its determinant is called the

...jk
j1 ...jk . Finally let us denote
and it is denoted by M
complementary minor of Mij11...i
i1 ,...ik
k
(the so called cofactor)

...jk
j1 ...jk .
Aji11...i
p1qi1 `ìk `j1 ``jk M
i1 ...ik
k
The adjugate of A is the matrix adjpAq
pAji q i1,n
, that is
j1,n
adjpAq
A11
A12
A1n
A21
A22
A2n
..
.
..
.
An1 An2
..
.
n
An
The next result provides a computation method of the inverse of a nonsingular

matrix.
Theorem 1.16. A square matrix A P Mn pFq is invertible if and only if
detpAq 0. In this case its inverse can be obtained by the formula
A1
1
adjpAq.
detpAq
Corollary 1.17. A linear system Ax 0 with n equations in n unknowns has a

non-trivial solution if and only if detpAq 0.
We state, without proof, the Laplace expansion theorem:
16
Theorem 1.18.
detpAq
...jk j1 ...jk
Ai1 ...ik , where
Mij11...i
k
The indices i1 . . . ik are xed

The indices j1 . . . jk runs over all the possible values, such that
1 j1 jk n.
As immediate consequences we obtain the following methods of calculating
determinants called row expansion and column expansion.
Corollary 1.19. Let A P Mn pFq. Then
(i) detpAq
aik Aki , (expansion by row i)
(ii) detpAq
akj Ajk , (expansion by column j).
k1
k1
Properties of the determinant.

Let A, B P Mn pFq and let a P F. Then
(1) detpAJ q detpAq.
(2) A permutation of the rows, (respectively columns) of A multiplies the
determinant by the sign of the permutation.
(3) A determinant with two equal rows (or two equal columns) is zero.
(4) The determinant of A is not changed if a multiple of one row (or column) is
added to another row (or column).
(5) detpA1 q
1
.
detpAq
(6) detpABq detpAq detpBq.
17
(7) detpaAq an detpAq.

(8) If A is a triangular matrix, i.e. aij 0 whenever i j (aij 0 whenever
i j), then its determinant equals the product of the diagonal entries, that
is detpAq a11 a22 . . . ann ni1 aii .

Rank. Elementary transformations.
A natural number r is called the rank of the matrix A P Mm,n pFq if
1. There exists a square submatrix M P Mr pFq of A which is nonsingular (that
is detpMq 0).
2. If p r, for every submatrix N P Mp pFq of A one has detpNq 0.
We denote rank pAq r.
It can be proved that for A P Mm,n pFq and B P Mn,p pFq, then
rank pAq ` rank pBq m rank pABq mintrank pAq, rank pBqu .
Theorem 1.20. Let A, B P Mn pFq with detpAq 0. Then rank pABq rank pBq.
Proof. Since detpAq 0, we have rank pAq n. By using the above notation with
m p n we obtain rank pBq rank pABq rank pBq. Hence
rank pABq rank pBq.
Denition 1.21. The following operations are called elementary row
transformations on the matrix A P Mm,n pFq:
1. Interchanging of any two rows.
2. Multiplication of a row by any non-zero number.
3. The addition of one row to another.
18
Similarly one can dene the elementary column transformations.

Consider an arbitrary determinant. If it is nonzero it will be nonzero after
preforming elementary transformations. If it is zero it will remain zero. One can
conclude that the rank of a matrix does not change if we perform any elementary
transformation on the matrix. So we can use elementary transformation in order
to compute the rank.
Namely given a matrix A P Mm,n pFq we transform it by an appropriate succession
of elementary transformations- into a matrix B such that
the diagonal entries of B are either 0 or 1, all the 1s preceding all the 0s on
the diagonal.
all the other entries of B are 0.
Since the rank is invariant under elementary transformations, we have
rank pAq rank pBq, but it is clear that the rank of B is equal to the number of 1s
on the diagonal.
The next theorem oers a procedure to compute the inverse of a matrix:
Theorem 1.22. If a square matrix is reduced to the identity matrix by a sequence
of elementary row operations, the same sequence of elementary row transformations
performed on the identity matrix produces the inverse of the given matrix.
1 2 0
Example 1.23. Compute the inverse of the matrix A 0 2 1 by using
0 0 3
elementary row operations.

1 0 0
1 2 0
p 1 A3 À2 q
We write 0 2 1 0 1 0

0 0 1
0 0 3
19
2 0
1 0 0

pA À q
2
1

2 0 0 1 3

0 3
0 0 1

1 1 13
0 0

p 1 A2 , 1 A3 q

2
3
1
2 0
0 1 3

0 3
0 0
1
1
1 1 3
1
1
1
Hence A 0 2 6 .
1
0 0
3
Recall that a matrix is in row echelon form if
1 0 0
0 1 0
0 0 1
1 1
1
3
1
2
16
1
3
(1) All nonzero rows are above any rows of all zeroes.
(2) The rst nonzero element (leading coecient) of a nonzero row is always
strictly to the right of the rst nonzero element of the row above it.
If supplementary the condition
p3q Every leading coecient is 1 and is the only nonzero entry in its column, is
also satised, we say that the matrix is in reduced row echelon form.
An arbitrary matrix can be put in reduced row echelon form by applying a nite
sequence of elementary row operations. This procedure is called the Gauss-Jordan
elimination procedure.
Existence of an inverse. For a square matrix A P Mn pFq the following
statements are equivalent.
1. A1 exists (A is nonsingular).
2. rank pAq n.
3. A is transformed by Gauss Jordan in In .
20
4. Ax 0 implies that x 0.
Systems of linear equations.
Recall that system of m linear equations in n unknowns can be written as
$
a11 x1 ` a12 x2 ` a1n xn b1
& a x ` a x ` a x b
21 1
2n n
22 2
% a x ` a x ` a x b .
m1 1
m2 2
mn n
m
Here x1 , x2 , . . . , xn are the unknowns, a11 , a12 , . . . , amn are the coecients of the
system, and b1 , b2 , . . . , bm are the constant terms. Observe that a systems of linear
equations may be written as Ax b, with A paij qi1,m P Mm,n pFq, x P Mn,1 pFq
j1,n
and b P Mm,1 pFq. The matrix A is called the coecient matrix, while the matrix
rA|bs P Mm,n`1 pFq,
$
& a if j n ` 1
ij
rA|bsij
% b if j n ` 1
i
is called the augmented matrix of the system.

We say that x1 , x2 , ..., xn is a solution of a linear system if x1 , x2 , ..., xn satisfy each
equations of the system. A linear system is consistent if it has a solution, and
inconsistent otherwise. According to the Rouche-Capelli theorem, any system of
linear equations is inconsistent if the rank of the augmented matrix is greater than
the rank of the coecient matrix. If, on the other hand, the ranks of these two
matrices are equal, the system must have at least one solution. The solution is
unique if and only if the rank equals the number of variables. Otherwise the
general solution has k free parameters where k is the dierence between the
number of variables and the rank. Two linear systems are equivalent if and only if
they have the same solution set.
21
In row reduction, the linear system is represented as an augmented matrix rA|bs.

This matrix is then modied using elementary row operations until it reaches
reduced row echelon form. Because these operations are reversible, the augmented
matrix produced always represents a linear system that is equivalent to the
original. In this way one can easily read the solutions.
Example 1.24. By using Gauss-Jordan elimination procedure solve the following
systems of linear equations.
$
&
x1 x2 ` 2x4 2
2x1 ` x2 x3 4
x1 x2 2x3 ` x4 1
% x ` x ` x 1.
2
3
4
We
1 1
2 1 1 0
4 p2A1 À2 ,A1 À3 q
have rA|bs
1 1 2 1
1
0 1
1 1
1
2
1 1 0
2
8 pA2 A4 q
0 3 1 4
3
0 0 2 1
1
0 1
1
1
2
1 1 0
2
1 pA2 À1 ,3A2 À4 q

0 1
1
1
3
0 0 2 1
8
0 3 1 4
3
1 0 1
3
0 1 1
1 p 12 A3 À1 , 21 A3 À.21 ,2A3 À4 q
1
0 0 2 1
3
11
0 0 4 7
5
3
2
1 0 0
2
1
1
p 2 A4 À1 , 101 A4 À2 , 15 A4 À3 q
0 1 0
2
2
0 0 2 1
3
0 0 0 5
5
1 0 0 0
1 0 0
0
1
1
1
0 1 0
0
1 p 2 A3 , 5 A4 q 0 1 0 0
1
0 0 1 0
0 0 2 0
2
1
0 0 0 1
0 0 0 5
5
1
One can easily read the solution x1 1, x2 1, x3 1, x4 1.
Recall that a system of linear equations is called homogeneous if b p0 0 0qJ
that is
a11 x1 ` a12 x2 ` a1n xn 0
& a x ` a x ` a x 0
21 1
22 2
2n n
% a x ` a x ` a x 0.
m1 1
m2 2
mn n
A homogeneous system is equivalent to a matrix equation of the form

Ax O.
Obviously a homogeneous system is consistent, having the trivial solution
x1 x2 xn 0.
It can be easily realized that a homogeneous linear system has a non-trivial
solution if and only if the number of leading coecients in echelon form is less
than the number of unknowns, in other words, the coecient matrix is singular.
22
Problems
1.3
23
Problems
Problem
D1 0
1.3.1. By using Laplaces

theorem compute
the following determinants.
2 1 0 0 0 0
2 3 4 5
1
2
1
0
0
0
1 2 3 4
0 1 2 1 0 0
2 1 2 3 , D2
0 0 1 2 1 0
0 2 1 2
0 0 0 1 2 1
0 0 2 1
0 0 0 0 1 2
Problem 1.3.2. Compute the following determinants.
1 2 3
2
3

1
, where P C such that the relation 2 ` ` 1 0 holds.
a)
2
3

1
2

1
1 1
1
...
1
1
. . . n1

2
4
2pn1q
` i sin 2
.
b) 1 2
, where cos 2

...
n
n
..
..
..
..
..
.
.
.
.
.
2
1 n1 2pn1q . . . pn1q
Problem 1.3.3. Let A paij q i1,n P Mn pCq and let us denote
j1,n
A paij q i1,n P Mn pCq. Show that

j1,n
a) detpAq detpAq.
b) If aij aji , i, j P t1, 2, . . . , nu then detpAq P R.
Problem 1.3.4. Let a1 , a2 , . . . an P C. Compute the following determinants.
Problems
1
1
a1
a2
a) a21
a22
..
..
.
.
n1 n1
a1
a2
a1 a2 a3
an a1 a2
b) .
.. ..
..
. .
a2 a3 a4
24
...
a3
...
an
a23
..
.
...
..
.
a2n
..
.
an1
. . . an1
n
3
. . . an
. . . an1
.. .
..
.
.
. . . a1
Problem 1.3.5. Compute An , n 1 for the following matrices.
7
4
a b
, A
, a, b P R.
a) A
9 5
b a
1 3 5
a b b
b) A 0 1 3 , A b a b , a, b P R.
0 0 1
b b a
Problem 1.3.6. Compute the rank of the following matrices by using the
Gauss-Jordan elimination method.

1
2 2 3 2
0
1 2 3 5

6 1 1
2
3
3 1 1 3 4
a)
,
2 4
3
2
1 2 1
0
1 1

3 0
2
1
2
2
0
0 1 0
Problems
25
1 2 3 5 3 6
0 1 2 3 4 7
b) 2 1 3 3 2 5 .
5 0 9 11 7 16
2 4 9 12 10 26
Problem 1.3.7. Find the inverses of the following matrices by using the
Gauss-Jordan elimination method.
2 1 1
1 1
, B 1 2
a) A
3 .
1 3
3 1 1
b) A paij q i1,n
j1,n
$
& 1 if i j
P Mn pRq, where aij
% 0 otherwise.
Problem 1.3.8. Prove that if A and B are square matrices of the same size, both
invertible, then:
a) ApI ` Aq1 pI ` A1 q1 ,
b) pA ` BB J q1 B A1 BpI ` B J A1 Bq1 ,
c) pA1 ` B 1 q1 ApA ` Bq1 B,
d) A ApA ` Bq1 A B BpA ` Bq1 B,
e) A1 ` B 1 A1 pA ` BqB 1
f) pI ` ABq1 I ApI ` BAq1 B,
g) pI ` ABq1 A ApI ` BAq1 .
Problems
26
Problem 1.3.9. For every matrix A P Mm,n pCq prove that the product A A and
AA are hermitian matrices.
Problem 1.3.10. For a quadratic matrix A of order n explain why the equation
AX XA I
has no solution.
Problem 1.3.11. Solve the following systems of linear equations by using
Gauss-Jordan elimination procedure.
a)
$
&
2x1 3x2 ` x3 ` 4x4 13

3x1 ` x2 x3 ` 8x4 2
5x1 ` 3x2 4x3 ` 2x4 12
% x ` 4x 2x ` 2x 12.
1
2
3
4
b)
$
&
x1 x2 ` x3 x4 ` x5 x6 1
x1 ` x2 ` x3 ` x4 ` x5 ` x6 1
2x1 ` x3 x5 1
x2 3x3 ` 4x4 4
x1 ` 3x2 ` 5x3 x6 1
% x ` 2x ` 3x ` 4x ` 5x ` 6x 2
1
2
3
4
5
6
Problem 1.3.12. Find m, n, p P R such that the following systems be consistent,
and then solve the systems.
Problems
27
a)
$
2x y z 0
x ` 2y 3z 0
& 2x ` 3y ` mz 0
nx ` y ` z 0
x ` py ` 6z 0
% 2ex y ` z ` 2.
b)
$
2x y ` z 0
x ` 2y ` z 0
& mx y ` 2z 0
x ` ny 2z 0
3x ` y ` pz 0
% x2 ` y 2 ` x2 3.
2
Vector Spaces
2.1
Denition of Vector Space and basic

properties
Denition 2.1. A vector space V over a eld F (or F vector space) is a set V
with an addition ` (internal composition law) such that pV, `q is an abelian group
and a scalar multiplication : F V V, p, vq v v, satisfying the
following properties:
1. pv ` wq v ` w, @ P F, @v, w P F
2. p ` qv v ` v, @, P F, @v P V
3. pvq pqv
4. 1 v v, @v P V
The elements of V are called vectors and the elements of F are called scalars. The
scalar multiplication depends upon F. For this reason when we need to be exact
28
Subspaces of a vector space
29
we will say that V is a vector space over F, instead of simply saying that V is a
vector space. Usually a vector space over R is called a real vector space and a
vector space over C is called a complex vector space.
Remark. From the denition of a vector space V over F the following rules for
calculus are easily deduced:
0V 0
0F v 0V
v 0V 0F or v 0V .
Examples. We will list a number of simple examples, which appear frequently in
practice.
V Cn has a structure of R vector space, but it also has a structure of C
vector space.
V FrXs, the set of all polynomials with coecients in F with the usual
addition and scalar multiplication is an F vector space.
Mm,n pFq with the usual addition and scalar multiplication is a F vector space.
Cra,bs , the set of all continuous real valued functions dened on the interval
ra, bs, with the usual addition and scalar multiplication is an R vector space.
2.2
It is natural to ask about subsets of a vector space V which are conveniently closed
with respect to the operations in the vector space. For this reason we give the
following:
30
Denition 2.2. Let V a vector space over F. A subset U V is called subspace of

V over F if it is stable with respect to the composition laws, that is,
v ` u P U, @v, u P U, and v P U@ P F, v P U, and the induced operations verify the
properties form the denition of a vector space over F.
It is easy to prove the following propositions:
Proposition 2.3. Let V be a F vector space and U V a nonempty subset. U is
a vector subspace of V over F i the following conditions are met:
v u P U, @v, u P U
v P U, @ P F, @v P U
Proof. Obviously, the properties of multiplication with scalars, respectively the
associativity and commutativity of addition operation are inherited from V . Hence,
it remain to prove that 0 P U and for all u P U one has u P U. Since u P U for
every u P U and P F it follows that 0u 0 P U and 0 u u P U.
Proposition 2.4. Let V be a F vector space and U V a nonempty subset. U is
a vector subspace of V over F i
v ` u P U, @, P F, @v, u P V.
Proof. Let u, v P U. For 1, 1 we have v u P U. For 0 and P F we
obtain v P U. The conclusion follows from the previous proposition.
Example 2.5. Let S tpx, y, zq P R3 |x ` y ` z 0u. Show that S is a subspace
of R3 .
To see that S is a subspace we check that for all , P R and all
v1 px1 , y1 , z1 q , v2 px2 , y2 , z2 q P S
v1 ` v2 P S.
31
Indeed, since v1 , v2 P S we have

x1 ` y1 ` z1 0
x2 ` y2 ` z2 0,
and by multiplying the equations with and respectively, and adding the
resulting equations we obtain
px1 ` x2 q ` py1 ` y2q ` pz1 ` z2 q 0.
But this is nothing else than the fact that
v1 ` v2 px1 ` x2 , y1 ` y2 , z1 ` z2 q satises the equation that denes S.
The next propositions show how one can operate with vector subspaces (to obtain
a new vector subspace) and how one can obtain a subspace from a family of
vectors.
Proposition 2.6. Let V be a vector space and U, W V two vector subspaces.
The sets
U X W and U ` W tu ` w|u P U, w P W u
are subspaces of V .
Proof. We prove the statements by making use of the Proposition 2.4. Let , P F
and let u, v P U X W. Then u, v P U and u, v P W. Since U and W are vector spaces
it follows that v ` u P U, respectively v ` u P W. Hence v ` u P U X W.
Now consider , P F and let x, y P U ` W. Then x u1 ` w1 , y u2 ` w2 for
some vectors u1 , u2 P U, w1 , w2 P W. But then
x ` y pu1 ` u2q ` pw1 ` w2 q P U ` W.
32
The subspace U X W is called the intersection vector subspace, while the subspace
U ` W is called the sum vector subspace. Of course that these denitions can be
also given for nite intersections (respectively nite sums) of subspaces.
Proposition 2.7. Let V be a vector space over F and S V nonempty. The set
(
n
xSy
i1 i vi : i P F and vi P S, for all i 1, n, n P N is a vector subspace
over F of V .
Proof. The proof is straightforward in virtue of Proposition 2.4.
The above vector space is called the vector space generated by S, or the linear hull
of the set S and is often denoted by spanpSq. It is the smallest subspace of V
which contains S, in the sense that for every U subspace of V with S U it
follows that xSy U.
Now we specialize the notion of sum of subspaces, to direct sum of subspaces.
Denition 2.8. Let V be a vector space and Ui V subspaces, i 1, n. The sum
U1 ` ` Un is called direct sum if for every v P U1 ` ` Un , from
v u1 ` ` un w1 ` ` wn with ui , wi P Ui , i 1, n it follows that
ui wi , for every i 1, n.
The direct sum of the subspaces Ui , i 1, n will be denoted by U1 Un . The
previous denition can be reformulated as follows. Every u P U1 ` ` Un can be
written uniquely as u u1 ` u2 ` . . . ` un where ui P Ui , i 1, n.
The next proposition characterizes the direct sum of two subspaces.
Proposition 2.9. Let V be a vector space and U, W V be subspaces. The sum
U ` W is a direct sum i U X W t0V u.
Proof. Assume that U ` W is a direct sum and there exists s P U X W, s 0V .
But then every x P U ` W, x u ` w can be written as
33
x pu sq ` pw ` sq P U ` W. From the denition of the direct sum we have

u u s, w w ` s hence s 0V , contradiction.
Conversely, assume that U X W t0V u and U ` W is not a direct sum. Hence,
there exists x P U ` W such that x u1 ` w1 u2 ` w2 P U ` W and u1 u2 or
w1 w2 . But then u1 u2 w1 w2 , hence u1 u2 , w1 w2 P U X W. It follows
that u1 u2 and w1 w2 , contradiction.
Let V be a vector space over F and U be a subspace. On V one can dene the
following binary relation RU : let u, v P V , u RU v i u v P U.
It can easily be veried that the relation RU is an equivalence relation, that is
(r) v RU v, for all v P V . (reexivity)
(t) u RU v and v RU w u RU w, for all u, v, w P V. (transitivity)
(s) u RU v v RU u, for all u, v P V. (symmetry)
The equivalence class of a vector v P V is dened as
RU rvs tu P V : v RU uu v ` U.
The quotient set (or factor set) V {RU is denoted by V {U and consists of the set of
all equivalence classes, that is
V {U tRU rvs : v P V u.
Theorem 2.10. On the factor set V {U there is a natural structure of a vector
space over F.
Proof. Indeed, let us dene the sum of two equivalence class RU rvs and RU rws by
RU rvs ` RU rvs RU rv ` ws
Basis. Dimension.
34
and the multiplication with scalars by

RU rvs RU rvs.
Then, is an easy verication that with these operations V {U becomes an F
space.
The vector space from the previous theorem is called the factor vector space, or
the quotient vector space.
2.3
Basis. Dimension.
Up to now we have tried to explain some properties of vector spaces in the large.
Namely we have talked about vector spaces, subspaces, direct sums, factor space.
The Proposition 2.7 naturally raises some questions related to the structure of a
vector space V . Is there a set S which generates V (that is xSy V )? If the
answer is yes, how big should it be? Namely how big should a minimal one
(minimal in the sense of cardinal numbers) be? Is there a nite set which generates
V ? We will shed some light on these questions in the next part of this chapter.
Why are the answers to such questions important? The reason is quite simple. If
we control (in some way) a minimal system of generators, we control the whole
space.
Denition 2.11. Let V be a F vector space. A nonempty set S V is called
system of generators for V if for every v P V there exists a nite subset
tv1 , . . . , vn u V and the scalars 1 , . . . , n P F such that v 1 v1 ` ` n vn (it
is also said that V is a linear combination of v1 , . . . , vn with scalars in F). V is
called dimensionally nite, or nitely generated, if it has a nite system of
generators.
Basis. Dimension.
35
A nonempty set L V is called a linear independent system of vectors if for every

nite subset tv1 , . . . , vn u L of it 1 v1 ` . . . n vn 0 implies that ai 0 for all
i 1, n.
A nonempty set of vectors which is not linear independent is called linearly
dependent.
A subset B V is called basis of V if it is both a system of generators and linearly
independent. In this case every vector v P V can be uniquely written as a linear
combination of vectors from B.
Example 2.12. Check whether the vectors p0, 1, 2q , p1, 2, 0q , p2, 0, 1q are linearly
independent in R3 .
By denition, the three vectors are linearly independent if the implication
1 p0, 1, 2q ` 2 p1, 2, 0q ` 3 p2, 0, 1q 0R3 1 2 3 0
holds.
Checking the above implication actually amounts (after computing the right hand
side) to investigating whether the linear system
$
2 ` 22 0
&
1 ` 22
% 2
`
1
has only the trivial solution p1 , 2 , 3 q p0, 0, 0q or not. But we can easily
compute the rank of the matrix, which is 3 due to
0 1 2
1 2 0 9 0,
2 0 1
Basis. Dimension.
36
to see that, indeed, the system has only the trivial solution, and hence the three
vectors are linearly independent.
We have the following theorem.
Theorem 2.13. (Existence of basis) Every vector space V has a basis.
We will not prove this general theorem here, instead we will restrict to nite
dimensional vector spaces.
Theorem 2.14. Let V t0u be e nitely generated vector space over F. From
every nite system of generators one can extract a basis.
Proof. Let S tv1 , . . . , vr u be a nite generators system. It is clear that there are
nonzero vectors in S (otherwise V t0u). Let 0 v1 P S. The set tv1 u is linearly
independent (because v1 0 0 from v1 0). That means that S contains
linearly independent subsets. Now P pSq is nite (S being nite), and in a nite
number of steps we can extract a maximal linearly independent system, let say
B tv1 , . . . , vn u, 1 n r in the following way:
v2 P Szxv1 y,
v3 P Szxtv1 , v2 uy
..
.
vn P Szxtv1 , v2 , . . . , vn1 uy.
We prove that B is a basis for V . It is enough to show that B generates V ,
because B is linearly independent by the choice of it. Let v P V . S being a system
of generators it follows that it is enough to show that every vk P S, n k r is a
linear combination of vectors from B. Suppose, by contrary, that vk is not a linear
combination of vectors from B. It follows that the set B Y tvk u is linearly
independent, contradiction with the maximality of B.
Basis. Dimension.
37
Corollary 2.15. Let V be a F vector space and S a system of generators for V .

Every linearly independent set L S can be completed to a basis of V .
Proof. Let L S be a linearly independent set in S. If L is maximal by the
previous Theorem it follows that L is a basis. If L is not maximal, there exists a
linearly independent set L1 with L L1 S. If L1 is maximal it follows that L1 is
a basis. If it is not maximal, we repeat the previous step. Because S is a nite set,
after a nite number of steps we obtain a system of linearly independent vectors B
which is maximal, L B S, so B is a basis for V , again by the previous
Theorem.
Theorem 2.16. Let V be a nitely generated vector space over F. Every linearly
independent system of vectors L can be completed to a basis of V .
Proof. Let S be a nite system of generators. The intersection L X S is again a
system of generators and L L X S. We apply the previous corollary and we
obtain that L can be completed to a basis of V .
Theorem 2.17. (The cardinal of a basis). Let V be a nitely generated F vector
space. Every basis of V is nite and has the same number of elements.
Proof. Let B te1 , . . . .en u be a basis of V , and let B1 te11 , . . . , e1m u a system of
vectors with m n. We show that B1 can not be a basis for V .
Because B is a basis the vectors e1i can be uniquely written as e1i nj1 aij ej ,
1
1 i m. If B1 is linearly independent, then it follows that m
i1 i ei 0 implies
i 0, i 1, m, or, in other words, the system m

i1 aij i 0, j 1, n has only
the trivial solution, impossible.
Denition 2.18. Let V t0u be a F vector space nitely generated. The number
of elements in a basis of V is called the dimension of V (it does not depend on the
Basis. Dimension.
38
choice of the basis, and it is denoted by dim F V ). The vector space V is said to be
of nite dimension. For V t0u , dim F V 0.
Remark 2.19. According to the proof of Theorem 2.17, if dim F V n then any
set of m n vectors is linear dependent.
Corollary 2.20. Let V be a vector space over F of nite dimension, dim F V n.
1. Any linearly independent system of n vectors is a basis. Any system of m
vectors, m n is linearly dependent.
2. Any system of generators of V which consists of n vectors is a basis. Any
system of m vectors, m n is not a system of generators
Proof. a) Consider L tv1 , . . . , vn u a linearly independent system of n vectors.
From the completion theorem (Theorem 2.16) it follows that L can be completed
to a basis of V . It follows from the cardinal basis theorem (Theorem 2.17) that
there is no need to complete L, so L is a basis.
Let L1 be a system of m vectors, m n. If L1 is linearly independent it follows that
L1 can be completed to a basis (Theorem 2.16), so dim F V m n, contradiction.
b) Let S tv1 , . . . , vn u be a system of generators which consists of n vectors.
From the Theorem 2.14 it follows that a basis can be extracted from its n vectors.
Again from the basis Theorem 2.17 it follows that there is no need to extract any
vector, so S is a basis.
Let S 1 be a generators system which consists of m vectors, m n. From the
Theorem 2.14 it follows that from S 1 one can extract a basis, so dim F V m n,
contradiction.
Remark 2.21. The dimension of a nite dimensional vector space is equal to any
of the following:
Basis. Dimension.
39
The number of the vectors in a basis.

The minimal number of vectors in a system of generators.
The maximal number of vectors in a linearly independent system.
Example 2.22. Let S tpx, y, zq P R3 |x ` y ` z 0u. Give an example of a basis
of S.
In example 2.5 we have shown that S is a subspace of R3 . One can see that, from a
geometric point of view, S is a plane passing through the origin, so dim S 2.
This follows also from rewriting S as follows
S
px, y, zq P R4 |x ` y ` z 0
tpx, y, x yq |x, y P Ru
tx p1, 0, 1q ` y p0, 1, 1q |x, y P Ru
span tp1, 0, 1q , p0, 1, 1qu .
The vectors p1, 0, 1q and p0, 1, 1q are linearly independent so they form a basis
of S.
Theorem 2.23. Every linearly independent list of vectors in a nite dimensional
vector space can be extended to a basis of the vector space.
Proof. Suppose that V is nite dimensional and tv1 , . . . , vm u is linearly
independent. We want to extend this set to a basis of V . V being nite
dimensional, there exists a nite set tw1 , . . . , wn u, a list of vectors which spans V .
If w1 is in the span of tv1 , . . . , vm u, let B tv1 , . . . , vm u. If not, let
B tv1 , . . . , vm , w1 u.
Basis. Dimension.
40
If wj is in the span of B, let B unchanged. If wj is not in the span of B,

extend B by jointing wj to it.
After each step B is still linearly independent. After n steps at most, the span of
B includes all the ws. Thus B also spans V , and being linearly independent, it
follows that it is a basis.
As an application we show that every subspace of a nite dimensional vector space
can be paired with another subspace to form a direct sum which is the whole space.
Theorem 2.24. Let V be a nite dimensional vector space and U a subspace of V .
There exists a subspace W of V such that V U W .
Proof. Because V is nite dimensional, so is U. Choose tu1 , . . . , um u a basis of U.
This basis of U a linearly independent list of vectors, so it can be extended to a
basis tu1 , . . . , um , w1 , . . . , wn u of V . Let W xw1 , . . . , wn y.
We prove that V U W . For this we will show that
V U ` W, and U X W t0u
Let v P V , there exists pa1 , . . . , am , b1 , . . . , bn ) such that
v a1 u1 ` ` am um ` b1 w1 ` ` bn wm ,
because tu1 , . . . , um , w1 , . . . , wn u generates V . By denoting
a1 u1 ` ` am um u P U and b1 w1 ` ` bn wm w P W we have just proven
that V U ` W .
Suppose now that U X W t0u, so let 0 v P U X W . Then there exist scalars
a1 , . . . , am P F and b1 , . . . , bn P F not all zero, with
v a1 u1 ` ` am um b1 w1 ` ` bn wm ,
Basis. Dimension.
41
so
a1 u1 ` ` am um b1 w1 bn wm 0.
But this is a contradiction with the fact that tu1 , . . . , um , w1 , . . . , wn u is a basis of
V , so we obtain the contradiction, i.e. U X W t0u.
The next theorem relates the dimension of the sum and the intersection of two
subspaces with the dimension of the given subspaces:
Theorem 2.25. If U and W are two subspaces of a nite dimensional vector
space V , then
dim pU ` W q dim U ` dim W dim pU X W q .
Proof. Let tu1 , . . . , um u be a basis of U X W , so dim U X W m. This is a linearly
independent set of vectors in U and W respectively, so it can be extended to a
basis tu1 , . . . , um , v1 . . . vi u of U and a basis tu1 , . . . , um , w1 , . . . wj u of W , so
dim U m ` i and dim W m ` j. The proof will be complete if we show that
tu1 , . . . , um , v1 . . . , vi , w1 , . . . , wj u is a basis for U ` W , because in this case
dim pU ` W q m ` i ` j
pm ` iq ` pm ` jq m
dim U ` dim W dimpU X W q
The set spantu1 , . . . , um , v1 . . . , vi , w1 , . . . , wj u contains U and W , so it contains
U ` W . That means that to show that it is a basis for U ` W it is only needed to
show that it is linearly independent. Suppose that
a1 u1 ` ` am um ` b1 v1 ` ` bi vi ` c1 w1 ` ` cj wj 0 .
We have
c1 w1 ` ` cj wj a1 u1 am um b1 v1 bi vi
Basis. Dimension.
42
which shows that w c1 w1 ` ` cj wj P U . But this is also in W , so it lies in

U X W . Because u1 , . . . , um is a basis in U X W it follows that there exist the
scalars d1 , . . . , dm P F, not all zero, such that
c1 w1 ` ` cj wj pd1 u1 ` ` dm um q .
But tu1 , . . . , um , w1 , . . . , wj u is a basis in W , so it is linearly independent, that is
all ci s are zero.
The relation involving as, bs and cs becomes
a1 u1 ` ` am um ` b1 v1 ` ` bi vi 0 ,
so as and bs are zero because the vectors tu1 , . . . , um , v1 . . . , vi u form a basis in U.
So all the as, bs and cs are zero, that means that
tu1 , . . . , um , v1 , . . . , vi , w1 , . . . , wj u are linearly independent, and because that
generates U ` W , they form a basis of U ` W .
The previous theorem shows that the dimension ts well with the direct sum of
spaces. That is, if U X W t0u, the sum is the direct sum and we have
dim pU W q dim U ` dim W .
This is true for the direct sum of any nite number of spaces as it is shown in the
next theorem:
Theorem 2.26. Let V be a nite dimensional space, Ui subspaces of V , i 1, n,
such that
V U1 ` ` Un ,
and
dim V dim U1 ` ` dim Un .
Basis. Dimension.
43
Then
V U1 Un .
Proof. One can choose a basis for each Ui . By putting all these bases in one list,
we obtain a list of vectors which spans V (by the rst property in the theorem),
and it is also a basis, because by the second property, the number of vectors in this
list is dim V .
Suppose that we have ui P Ui , i 1, n, such that
0 u1 ` ` un .
Every ui is represented as the sum of the vectors of basis of Ui , and because all
these bases form a basis of V , it follows that we have a linear combination of the
vectors of a base of V which is zero. So all the scalars are zero, that is all ui are
zero, so the sum is direct.
We end the section with two important observations. Let V be a vector space over
F (not necessary nite dimensional). Consider a basis B pei qiPI of V .
We have the rst representation theorem:
Theorem 2.27. Let V be a vector space over F (not necessary nite dimensional).
Let us consider a basis B pei qiPI . For every v P V, v 0 there exist a unique
subset B1 B, B1 tei1 , . . . , eik u and the nonzero scalars ai1 , . . . , aik P F , such
that
v
aij eij ai1 ei1 ` ` aik eik .
j1
Proof. Obviously, by the denition of basis v is a nite linear combination of the

elements of the basis. We must show the uniqueness. Assume the contrary, that
n
m
ji eji
ki eki , ji 0, i 1, n, ki 0, i 1, m.
v
i1
i1
Local computations
44
Assume that there exists eks R tej1 , . . . , ejn u. Then, since

m
n
i1 ji eji
i1 ki eki 0 we obtain that ks 0, contradiction. Similarly,
ejs P tek1 , . . . , em u, for all s 1, n. Hence, m n and one may assume that
v
ji eji
i1
Using the relation
ki eki , ji 0, i 1, n, ki 0, i 1, n.
i1
i1
ji eji
i1
ki eki 0 again we obtain that
ji ki , i P t1, . . . , nu, contradiction.

Example 2.28. Show that B tp1, 1q , p1, 1qu.is a basis of R2 , and nd the
representation with respect to B of the vector v p3, 1q.
Our aim is to nd the representation of v p3, 1q with respect to B, that is, to
nd two scalars x, y P R such that
v x p1, 1q ` y p1, 1q .
Expressing the above equality component wise gives a system with two unknowns,
x and y
$
& x ` y 3
% x y 1.
Its unique solution, and the answer to our problem is x 1, y 2.
2.4
Local computations
In this section we deal with some computations related to nite dimensional vector
spaces.
Let V be an F nite dimensional vector space, with a basis B te1 , . . . , en u. Any
vector v P V can be uniquely represented as
v
i1
ai ei a1 e1 ` ` an en .
Local computations
45
The scalars pa1 , . . . , an q are called the coordinates of the vector v in the basis B. It
1
is obvious that if we have another basis B , the coordinates of the same vector in
the new basis change. How we can measure this change? Let us start with a
situation that is a bit more general.
Theorem 2.29. Let V be a nite dimensional vector space over F with a basis
1
B te1 , . . . , en u. Consider the vectors S te1 , . . . , em u V :
e1 a11 e1 ` ` a1n en
...
1
em am1 e1 ` ` amn en
Denote by A paij qi1,m the matrix formed by the coecients in the above
j1,n
equations. The dimension of the subspace xSy is eqaul to the rank of the matrix A,
i.e. dimxSy rankA.
Proof. Let us denote by Xi pai1 , . . . , ain q P Fn , i 1, m the coordinates of
1
1
ei , i 1, m in B. Then, the linear combination m
i1 i ei has its coordinates
m
i1 i Xi in B. Hence the set of all coordinate vectors of elements of xSy equals
1
the subspace of Fn generated by tX1 , ..., Xm u. Moreover e1 , . . . , em will be linearly

independent if and only if X1 , . . . , Xm are. Obviously, the dimension of the
subspace xX1 , ..., Xm y of Fn is equal with the rank of the matrix
X1
..
. A.
Xm
Local computations
46
1
Consider now the case of m n in the above discussion. The set S te1 , . . . , en u
is a basis i rankA n We have now
1
e1 a11 e1 ` ` a1n en
1
e2 a21 e1 ` ` a2n en
...
1
en an1 e1 ` ` ann en ,
1
representing the relations that change from the basis B to the new basis B S.
The matrix AJ is denoted by
a11 a21 . . .
P pe,e q
a12 a22 . . .
... ... ...
a1n a2n . . .
an1
an2
...
ann
.
The columns of this matrix are given by the coordinates of the vectors
1
of the new basis e with respect to the old basis e!

Remarks
In the matrix notations we have
1
e
e
1
1
1
e
e2
A 2
...
...
1
en
en
or pe1 q1,n pP pe,e q qJ peq1,n
Consider the change of the basis from B to B with the matrix P pe,e q and
1
the change of the basis from B to B with the matrix P pe ,e q . We can think
Local computations
47
at the composition of these two changes, i.e. the change of the basis from
2
B to B with the matrix P pe,e q . It is easy to see that one has

1
P pe,e q P pe ,e q P pe,e
If in the above discussion we consider B B one has

1
P pe,e q P pe ,eq In ,
that is
1
pP pe ,eq q1 P pe,e q
At this step we try to answer the next question, which is important in
applications. If we have two basis, a vector can be represented in both of them.
What is the relation between the coordinates in the two basis?
Let us x the setting rst. Consider the vector space V , with two basis
1
B te1 , . . . , en u and B te1 , . . . , en u and P pe,e q the matrix of the change of basis.
Let v P V . We have
1
v a1 e1 ` ` an en b1 e1 ` ` bn en ,
where pa1 , . . . an q and pb1 , . . . bn q are the coordinates of the same vector in the two
basis. We can write
pvq
a1 a2 . . .
e
e
1
1
e2
e1
2
b b ... b
an
1
2
n
...
...
1
en
en
Local computations
48
Denote
a
1
a2
pvqe
...
an
and
b
1
b2
.
pvqe1
...
bn
the matrices of the coordinates of v in the two basis.
Denote further the basis columns
peq1n
e
1
e2
...
en
the column matrix of the basis B and
pe q1n
e
1
1
e2
...
1
en
the matrix column of the basis B , we have

1
J
J
pe,e q J
v pvqJ
q peq1n
e peq1n pvqe1 pe q1n pvqe1 pP
Because v is uniquely represented in a basis it follows

1
pP pe,e q qJ pvqJ
pvqJ
e ,
e1
Problems
49
or
1
pvqe1 pP pe,e q q1 pvqe P pe ,eq pvqe .

Hence,
1
pvqe pP pe,e q qpvqe1 .
2.5
Problems
Problem 2.5.1. Show that for spanpv1 , . . . , vn q V one has

spanpv1 v2 , v2 v3 , . . . , vn1 vn , vn q V .
Problem 2.5.2. Find a basis for the subspace generated by the given vectors in
M3 pRq.

0 1 2
,
2 4 1 2 1 1 ,
0 1
1
3 1 1
1 2
2 2 1 .
1 2 1
Problem 2.5.3. Let V be a nite dimensional vector space dim V n. Show that
there exist one dimensional subspaces U1 , . . . , Un , such that
V U1 Un .
Problem 2.5.4. Find three distinct subspaces U, V, W of R2 such that
R2 U V V W W U.
Problem 2.5.5. Let U, W be subspaces of R8 , with dim U 3, dim W 5 and
dim U ` W 8. Show that U X W t0u.
Problem 2.5.6. Let U, W be subspaces of R9 with dim U dim W 5. Show
that U X W t0u.
Problems
50
Problem 2.5.7. Let U and W be subspaces of a vector space V and suppose that
each vector v P V has a unique expression of the form v u ` w where u belongs
to U and w to W. Prove that
V U W.
Problem 2.5.8. In Cra, bs nd the dimension of the subspaces generated by the
following sets of vectors:
a) t1, cos 2x, cos2 xu,
b) tea1 x , . . . , ean x u, where ai aj for i j
Problem 2.5.9. Find the dimension and a basis in the intersection and sum of
the following subspaces:
U spantp2, 3, 1q, p1, 2, 2, q, p1, 1, 3qu,
V spantp1, 2, 1q, p1, 1, 1q, p1, 3, 3qu.
U spantp1, 1, 2, 1q, p0, 1, 1, 2q, p1, 2, 1, 3u,
V spantp2, 1, 0, 1q, p2, 1, 1, 1q, p3, 0, 2, 3qu.
Problem 2.5.10. Let U, V, W be subspaces of some vector space and suppose that
U W. Prove that
pU ` V q X W U ` pV X W q.
Problem 2.5.11. In R4 we consider the following subspace
V spantp2, 1, 0, 1q, p2, 1, 1, 1q, p3, 0, 2, 3qu. Find a subspace W of R4 such
that R4 V W .
Problem 2.5.12. Let V, W be two vector spaces over the same eld F. Find the
dimension and a basis of V W.
Problems
51
Problem 2.5.13. Find a basis in the space of symmetric, respectively

skew-symmetric matrices of dimension n.
Problem 2.5.14. Let
V tpx1 , . . . , xn q P Rn |, x1 ` x2 ` . . . ` xn 0, x1 ` xn 0u. Find a basis in V .
Problem 2.5.15. Let Mn pRq be the set of the real square matrices of order n,
and An , respectively Sn the set of symmetric, respectively skew-symmetric
matrices of order n. Show that Mn pRq An Sn .
Problem 2.5.16. Let us denote by Rn rXs the set of all polynomials having degree
at most n with real coecients. Obviously Rn rXs is a subspace of RrXs with the
induced operations. Find the dimension of the quotient space Rn rXs{U where U is
the subspace of all real constant polynomials.
Problem 2.5.17. Let V be a nite-dimensional vector space and let U and W be
two subspaces of V. Prove that
dim ppU ` W q{W q dim pU{pU X W qq.
Problem 2.5.18. Let us consider the matrix
1 3 5 3 6
1 2 3 4 7
M 1 3 3 2 5 .
0 9 11 7 16
4 9 12 10 26
Let U and W be the subspaces of R5 generated by rows 1, 2 and 5 of M, and by
rows 3 and 4 of M respectively. Find the dimensions of U ` W and U X W.
Problems
52
Problem 2.5.19. Find bases for the sum and intersection of the subspaces U and
W of R4 rXs generated by the respective sets of polynomials
t1 ` 2x ` x3 , 1 x x2 u and tx ` x2 3x3 , 2 ` 2x 2x3 u.
3
Linear maps between vector spaces
Up to now we met with vector spaces. It is natural to ask about maps between
them, which are compatible with the linear structure of a vector space. These are
called linear maps, special maps which also transport the linear structure. They
are also called morphisms of vector spaces or linear transformations.
Denition 3.1. Let V and W be two vector spaces over the same eld F. A linear
map from V to W is a map f : V W which has the property that
f pv ` uq f pvq ` f puq for all v, u P V and , P F.
The class of linear maps between V and W will be denoted by LF pV, W q or
HomF pV, W q.
From the denition it follows that f p0V q 0W and
fp
i1
i vi q
i f pvi q, @ i P F, @vi P V, i 1, n.
i1
We shall dene now two important notions related to a linear map, the kernel and
the image.
Consider the sets:
ker f f 1 p0W q tv P V |f pvq 0w u, and
53
54
imf f pV q tw P W |D v P V, f pvq wu.

Denition 3.2. The sets ker f and f pV q are called the kernel (or the null space),
respectively the image of f .
An easy exercise will prove the following:
Proposition 3.3. The kernel and the image of a linear map f : V W are
subspaces of V and W respectively.
Example 3.4. Let T : R2 R2 be given by px, yq px ` y, x ` yq. Find ker T
and T pR2 q.
By denition
ker T
(
px, yq P R2 |T px, yq p0, 0q
(
px, yq P R2 | px ` y, x ` yq p0, 0q

(
px, yq P R2 |x ` y 0 .
Geometrically, this is the straight line with equation y x. Clearly

ker T span tp1, 1qu and dim ker T 1.
From the way T is dened we see that all vectors in the image T pR2 q of T , have
both components equal to each other, so
`
T R2 tp, q | P Ru
span tp1, 1qu .
For the nite dimensional case the dimension of ker and im of a linear map
between vector spaces are related by the following:
Theorem 3.5. Let f : V W be a linear map between vector spaces V and W
over the eld F, V being nite dimensional.
dim V dim ker f ` dim imf.
55
Proof. Let n and m be the dimensions of V and ker f , m n. Consider a basis

te1 , . . . , em u for ker f . The independent system of vectors e1 , . . . , em can be
completed to a basis te1 , . . . , em , em`1 , . . . , en u of V .
Our aim is to prove that the vectors f pem`1 q, . . . , f pen q form a basis for f pV q. It is
sucient to prove that the elements f pem`1 q, . . . , f pen q are linearly independent
since they generate f pV q.
Suppose the contrary, that f pem`1 q, . . . , f pen q are not linearly independent. There
exists m`1 , . . . , n P F such that
n
k f pek q 0W ,
km`1
and by the linearity of f ,

n
fp
k ek q 0W .
km`1
Hence
1
k ek P ker f
km`1
and v 1 can be written in terms of e1 , . . . , em . This is only compatible with the fact
that e1 , . . . , en form a basis of V if m`1 n 0, which implies the linear
independence of the vectors f pem`1 q, . . . , f pen q.
Theorem 3.6. Let f : V W be a linear mapping between vector spaces V and
W , and dim V dim W 8. Then, f pV q W i ker f t0V u. In particular f is
onto i f is one to one.
Proof. Suppose that ker f t0V u. Since f pV q is a subspace of W it follows that
dim V dim f pV q dim W , which forces dim f pV q dim W , and this implies
that f pV q W .
56
The fact that f pV q W implies that ker f t0V u follows by reversing the
arguments.
Proposition 3.7. Let f : V W be a linear map between vector spaces V, W over
F. If f is a bijection, it follows that its inverse f 1 : W V is a linear map.
Proof. Because f is a bijection @w1 , w2 P W , D! v1 , v2 P V , such that
f pvi q wi , i 1, 2. Because f is linear, it follows that
1 w1 ` 2 w2 1 f pv1 q ` 2 f pv2 q f p1 v1 ` 2 v2 q.
It follows that 1 v1 ` 2 v2 f 1 p1 w1 ` 2 w2 q, so
f 1 p1 w1 ` 2 w2 q 1 f 1 pw1 `q ` 2 f 1 pw2 q.
Denition 3.8. A linear bijective map f : V W between vector spaces V, W

over F is called an isomorphism of the vector space V over W , or isomorphism
between the vector spaces V and W .
A vector space V is called isomorphic with a vector space W if there exists an
isomorphism f : V W . The fact that the vector spaces V and W are isomorphic
will denote by V W .
Example 3.9. Let V be a F vector space and V1 , V2 two supplementary spaces,
that is V V1 V2 . It follows that @v P V we have the unique decomposition
v v1 ` v2 , with v1 P V1 and v2 P V2 . The map
p : V V1 , ppvq v1 , @v P V
is called the projection of V on V1 , parallel with V2 .
Properties of LpV, W q
57
The map s : V V, spvq v1 v2 , @v P V is called the symmetry of V with

respect to V1 , parallel with V2 .
It is easy to see that for v P V1 , v2 0, so ppvq v and spvq v, and for v P V2 ,
v1 0, so ppvq 0 and spvq v.
3.1
In this section we will prove some properties of linear maps and of LpV, W q.
Proposition 3.10. Let f : V W be a linear map between the linear spaces V, W
over F.
1. If V1 V is a subspace of V , then f pV1 q is a subspace of W .
2. If W1 W is a subspace of W , then f 1 pW1 q is a subspace of V .
Proof. 1. Let w1 , w2 be in f pV1 q. It follows that there exist v1 , v2 P V1 such that
f pvi q wi , i 1, 2. Then, for every , P F we have
w1 ` w2 f pv1 q ` f pv2 q f pv1 ` v2 q P f pV1 q.
2. For v1 , v2 P f 1 pW1 q we have that f pv1 q, f pv2 q P W1 , so
@ , P F, f pv1 q ` f pv2 q P W1 . Because f is linear
f pv1 q ` f pv2 q f pv1 ` v2 q v1 ` v2 P f 1 pW1 q.
The next proposition shows that the kernel and the image of a linear map
characterize the injectivity and surjectivity properties of the map.
Proposition 3.11. Let f : V W be a linear map between the linear spaces V, W .
58
1. f is one to one (injective) ker f t0u.

2. f is onto (surjective) f pV q W .
3. f is bijective ker f t0u and f pV q W .
Proof. 1 Suppose that f is one to one. Because f p0V q 0W it follows that
ker f t0V u V . For the converse, suppose that ker f t0V u. Let v1 , v2 P V with
f pv1 q f pv2 q. It follows that f pv1 v2 q 0 and because ker f t0u we have that
v1 v2 . The claims 2. and 3. can be proved in the same manner.
Next we shall study how special maps act on special systems of vectors.
Proposition 3.12. Let f : V W be a linear map between the linear spaces V, W
and S tvi |i P Iu a system of vectors in V .
1. If f is one to one and S is linear independent, then f pSq is linear
independent.
2. If f is onto and S is a system of generators, then f pSq is s system of
generators.
3. If f is bijective and S is a basis for V , then f pSq is a basis for W .
Proof. 1. Let tw1 , . . . , wn u be a nite subsystem from f pSq, and i P F with
n
i1 i wi 0. There exist the vectors vi P V such that f pvi q wi , for all
i P t1, . . . , nu. Then ni1 i wi ni1 i f pvi q f p ni1 i vi q 0, so

n
i1 i vi 0. Because S is linearly independent it follows that i 0 for all
i 1, n, so f pSq is linearly independent.
2. Let w P W . There exists v P V with f pvq w. Because S is a system of
generators, there exists a nite family of vectors in S, vi , and the scalars
59
i P F, i 1, n such that
i1
i vi v. It follows that
w f pvq f p
i1
i vi q
i f pvi q.
i1
3. Because f is bijective and S is a basis for V , it follows that both 1. and 2. hold,
that is f pSq is a basis for W .
Denition 3.13. Let f, g : V W be linear maps between the linear spaces V
and W over F, and P F. We dene
1. f ` g : V W by pf ` gqpvq f pvq ` gpvq, @ v P V , the sum of the linear
maps, and
2. f : V W by pf qpvq f pvq, @ v P V, @ P F, the scalar multiplication
of a linear map.
Proposition 3.14. With the operations dened above LpV, W q becomes a vector
space over F.
The proof of this statement is an easy verication.
In the next part we specialize in the study of the linear maps, namely we consider
the case V W .
Denition 3.15. The set of endomorphisms of a linear space V is:
EndpLq tf : V V | f linear u.
By the results from the previous section, EndpV q is an F linear space.
Let W, U be two other linear spaces over the same eld F, f P LpV, W q and
g P LpW, Uq. We dene the product (composition) of f and g by
h g f : V U,
hpvq gpf pvqq, @ v P V.
60
Proposition 3.16. The product of two linear maps is a linear map.

Moreover if f and g as above are isomorphisms, then the product h g f is an
isomorphism.
Proof. We check that for all v1 , v2 P V and all , P F
hpv1 ` v2 q gpf pv1 ` v2 qq
gpf pv1q ` f pv2 qq
gpf pv1qq ` gpf pv2qq
hpv1 q ` hpv2 q.
The last statement follows from the fact that h is a linear bijection.
It can be shown that the composition is distributive with respect to the sum of
linear maps, so EndpV q becomes an unitary ring.
It can easily be realized that:
Proposition 3.17. The isomorphism between two linear spaces is an equivalence
relation.
Denition 3.18. Let V be an F linear space. The set
AutpV q tf P EndpV q| f isomorphism u
is called the set of automorphisms of the vector space V .
Proposition 3.19. AutpV q is a group with respect to the composition of linear
maps.
Proof. It is only needed to list the properties.
1. the identity map IV is the unit element.
61
2. g f is an automorphism for f and g automorphisms.

3. the inverse of an automorphism is an automorphism.
The group of automorphisms of a linear space is called the general linear group
and is denoted by GLpV q.
Example 3.20.
Projectors endomorphisms. An endomorphism
p : V V is called projector of the linear space V i

p2 p,
where p2 p p. If p is a projector, then:
1. ker p ppV q V
2. the endomorphism q IV p is again a projector.
Denote v1 ppvq and v2 v v1 , it follows that
ppv2 q ppvq ppv1 q ppvq p2 pvq 0V , so v2 P ker f . Hence
v v1 ` v2 , @ v P V,
where v1 , v2 P f pV q and, moreover, the decomposition is unique, so we have
the direct sum decomposition ker p ppV q V . For the last assertion simply
compute q 2 pIV pq pIV pq IV p p ` p2 IV p q, because p is
a projector. It can be seen that qpV q ker p and ker q qpV q. Denote by
V1 ppV q and V2 ker p. It follows that p is the projection of V on V1 ,
parallel with V2 , and q is the projection of V on V2 parallel with V1 .
Involutive automorphisms. An operator s : V V is called involutive i
s2 IV . From the denition and the previous example one has:
62
1. an involutive operator is an automorphism

2. for every involutive automorphism, the linear operators:
1
ps : V V, ps pvq pv ` spvqq
2
1
qs : V V, qs pvq pv spvqq
2
are projectors and satisfy the relation ps ` qs 1V .
3. reciprocally, for a projector p : V V , the operator sp : V V , given
by sp pvq 2ppvq v is an involutive automorphism.
From the previous facts it follows that ps s s ps p, sp p p sp p. An
involutive automorphism s is a symmetry of V with respect to the subspace ps pV q,
parallel with the subspace ker ps .
Example 3.21. Let V be a vector space and f : V V a linear map such that
ker f imf . Determine the set imf 2 , where f 2 denotes the composition of f with
itself, f 2 f f .
We start by writing down explicitly
imf 2 imf f
f f pV q
f pf pV qq .
But, f pV q imf ker f is the set of all vectors which are mapped by f to zero, so
imf 2 f pker f q
0.
Local form of a linear map
3.2
63
Let V and W be two vector spaces over the same led F, dim V m, dim W n,
and e te1 , . . . , em u and f tf1 , . . . , fn u be bases in V and W respectively. A
linear map T P LpV, W q is uniquely determined by the values of the basis e.
We have
T pe1 q a11 f1 ` ` a1n fn ,
T pe2 q a21 f1 ` ` a2n fn ,
..
.
T pem q am1 f1 ` ` amn fn ,
or, in the matrix notation
T pe1 q
T pe2 q
..
T pem q
f1
f2
..
.
where A paij qi1,m .
j1,n
fn
pf,eq
The transposed of A is denoted by MT
and is called the matrix of the linear
map T is the basis e and f .

From the denition of the matrix of a linear map it follows that:
Theorem 3.22.
For T1 , T2 P LpV, W q and a1 , a2 P F

Ma1 T1 à2 T2 a1 MT1 ` a2 MT2
The vector space LpV, W q is isomporphic to Mm,n pFq by the map

T P LpV, W q MT pFq P Mm,n pFq.
Particularly EndpV q is isomorphic to Mn pFq.
64
Now we want to see how the image of a vector by a linear map can we expressed.
J
Let v P V, v m
i1 vi ei , or in the matrix notation pvqe peq1m , where, as usual
v1
pvqe
v2
..
.
vn
and
e1
peq1m
e2
..
.
em
Now denote T pvq w
j1 wj ej
P W , we have
T pvq pwqJ
f pf q1n .
T being linear, we have T pvq
i1
vi T pei q, or, again in matrix notation:
T pvq pvqJ
e pT peqq1m .
pf,eq
From the denition of MT
it follows that
pf,eq J
pT peqq1m pMT
q pf q1n .
So nally we have
pf,eq J
J
pwqJ
f pf q1n pvqe pMT
q pf q1n .
By the uniqueness of the coordinates of a vector in a basis it follows that

pf,eq J
J
pwqJ
f pvqe pMT
q .
Taking the transposed of the above relation we get

pf,eq
pwqf pMT
qpvqe .
65
3 0 2
3
3
Example 3.23. Let T : R R , T 1 1 0 . Find a basis in ker T and
2 1 2
3
nd the dimension of T pR q .
Observe that the kernel of T ,
$
&
ker T px, y, zq P |T
,
/
/
/
.

y 0 ,
/
/
/
z
0
is the set of solutions of the linear homogeneous system

$
3x
` 2z 0
&
x ` y
0
% 2x ` y ` 2z 0,
the matrix of the system being exactly T . To solve this system we need to
compute the rank of the matrix T . We
get that
0 2
1 0 0
1 2
and that rank A 2. To solve the system we chose x as a parameter and

express y and z in terms of x from the rst two equations to get
x ,
y x,
3
z x.
2
The set of solutions is
*
"
"
*
3
3
1, 1,
, , | P R span
2
2
(3.1)
66
and hence, a basis in ker T consists only of 1, 1, 32 , dim ker T 1.

Based on the dimension formula
`
dim ker T ` dim T R3 dim R3
we infer that dim T pR3 q 2.
Proposition 3.24. Let V, W, U be vector spaces over F, of dimensions m, n, p, and
T P LpV, W q, S P LpW, Uq, with matrices MT and MS , in some basis. Consider the
composition map S T : V U with the matrix MST . Then
MST Ms MT .
Proof. Indeed, one can easily see that for v P V we have pT pvqq MT pvq where
pT pvqq, respectively pvq stand for the coordinate of T pvq, respectively v in the
appropriate bases. Similarly, for w P W one has pSpwqq MS pwq.
Hence, pS T pvqq pSpT pvqqq MS pT pvqq MS MT pvq, or, equivalently
MST MS MT .
Let V and W be vector spaces and T P LpV, W q be a linear map. In V and W we

consider the bases e te1 , . . . , em u and f tf1 , . . . , fn u, with respect to these
pf,eq
bases the linear map has the matrix MT
. If we consider two other bases
e1 te11 , . . . , e1m u and f 1 tf11 , . . . , fn1 u the matrix of T with respect to these bases
pf 1 ,e1 q
will be MT
. What relation do we have between the matrixes of the same linear
map in these two bases?

pf 1 ,e1 q
Theorem 3.25. In the above conditions MT
pf,eq
P pf ,f q MT
P pe,e q .
Problems
67
Proof. Let us consider v P V and let w T pvq. We have

pf 1 ,e1 q
pwqf 1 MT
pf 1 ,e1 q
pvqe1 MT
P pe ,eq pvqe .
On the other hand

1
pf,eq
pwqf 1 P pf ,f q pwqf P pf ,f q pT pvqqf P pf ,f q MT

1
pvqe .
Taking into account that pP pe ,eq q1 P pe,e q we get

pf 1 ,e1 q
MT
pf,eq
P pf ,f q MT
pf,eq
pP pe ,eq q1 P pf ,f q MT
P pe,e q .
Corollary 3.26. Let e and e1 be two bases of a nite-dimensional vector space V

pe,eq
and let T : V V be a linear mapping. If T is represented by matrices A MT

pe1 ,e1 q
and A1 MT
with respect to e and e1 respectively, then A1 P AP 1 where P is
the matrix representing the change of basis e to e1 .
3.3
Problems
Problem 3.3.1. Consider the following mappings T : R3 R3 . Study which one

of them is a linear mapping.
a) T px1 , x2 , x3 q px21 , x2 , x23 q.
b) T px1 , x2 , x3 q px3 , x1 , x2 q.
c) T px1 , x2 , x3 q px1 1, x2 , x3 q.
d) T px1 , x2 , x3 q px1 ` x2 , x2 x3 , x1 ` x2 ` x3 q.
e) T px1 , x2 , x3 q px3 , 0, 0q.
Problems
68
f) T px1 , x2 , x3 q px1 , 2x2 , 3x3 q.

Problem 3.3.2. Let T P EndpV q and let tei : i 1, nu be a basis in V . Prove
that the following statements are equivalent.
1. The matrix of T , with respect to the basis tei : i 1, nu is upper triangular.
2. T pek q P spante1 , . . . , ek u for all k 1, n.
3. T pspante1 , . . . , ek uq spante1 , . . . , ek u for all k 1, n.
Problem 3.3.3. Let T1 , T2 : R3 R3 having
3 1
MT1 0 2
1 2
the matrices
1 ,
respectively
MT2
1 4 2
0 4 1
0 0 5
in the canonical basis of R3 .

a) Find the image of p0, 1, 1q through T1 , T11 , T2 , T21 .
b) Find the image of p1, 3, 2q through T1 ` T2 , pT1 ` T2 q1 .
c) Find the image of p1, 2, 0q through T1 T2 , T2 T1 .
Problem 3.3.4. Let V be a complex vector space and let T P EndpV q. Show that
there exists a basis in V such that the matrix of T relative to this basis is upper
triangular.
Problems
69
Problem 3.3.5. Let T : R4 R3 be a
M 1
linear mapping represented by the matrix
0
1
2
1
0
1 .
1 1 3
Find a basis in ker T, im T and the dimension of the spaces V, W, ker T and im T .
Problem 3.3.6. Show that a linear transformation T : V W is injective if and
only if it has the property of mapping linearly independent subsets of V to linearly
independent subsets of W.
Problem 3.3.7. Show that a linear transformation T : V W is surjective if and
only if it has the property of mapping any set of generators of V to a set of
generators of W.
Problem 3.3.8. Let T : V W be a linear mapping represented by the matrix
1
1
1
2
M 1 1
1
1 .
0 2 2 3
Compute dim V, dim W and nd a basis in im T and ker T .
Problem 3.3.9. Find all the linear mappings T : R R with the property
im T ker T .
Find all n P N such that there exists a linear mapping T : Rn Rn with the
property im T ker T .
Problem 3.3.10. Let V , respectively Vi , i 1, n be vector spaces over C. Show
that, if T : V1 V2 Vn V is a linear mapping then there exist and they
are unique the linear mappings Ti : Vi V , i 1, n such that
T pv1 , . . . , vn q T1 pv1 q ` T2 pv2 q ` ` Tn pvn q.
Problems
70
Problem 3.3.11. (The rst isomorphism theorem). If T : V W is a linear

transformation between vector spaces V and W , then
V { ker T im T.
[Hint: show that the mapping S : V { ker T im T, Spv ` ker T q T pvq is a
bijective linear mapping.]
Problem 3.3.12. (The second isomorphism theorem). If U and W are subspaces
of a vector space V, then
pU ` W q{W U{pU X W q.
[Hint: dene the mapping T : U pU ` W q{W by the rule T puq u ` W , show
that T is a linear mapping and use the previous problem.]
Problem 3.3.13. (The third isomorphism theorem). Let U and W be subspaces
of a vector space V such that W U. Prove that U{W is a subspace of V {W and
that pV {W q{pU{W q V {U.
[Hint: dene a mapping T : V {W V {U by the rule T pv ` W q v ` U, show
that T is a linear mapping and use the rs isomorphism theorem.]
Problem 3.3.14. Show that every subspace U of a nite-dimensional vector space
V is the kernel and the image of suitable linear operators on V.
Problem 3.3.15. Let T : R4 R4 having the matrix
1 2 0 1
3 0 1 2
MT
2 5 3 1
1 2 1 3
in the canonical basis te1 , e2 , e3 , e4 u of R4 .
Find the matrix of T with respect to the following basis.
Problems
a) te1 , e3 , e2 , e4 u.
b) te1 , e1 ` e2 , e1 ` e2 ` e3 , e1 ` e2 ` e3 ` e4 u.
c) te4 e1 , e3 ` e4 , e2 e4 , e4 u.
Problem 3.3.16. A linear transformation T : R3 R2 is dened by
T px, x2 , x3 q px1 x2 x3 , x1 ` x3 q. Let e tp2, 0, 0q, p1, 2, 0q, p1, 1, 1qu and
f tp0, 1q, p1, 2qu be bases in R3 and R2 respectively. Find the matrix that
represents T with respect to these bases.
71
4
Proper vectors and the Jordan canonical
form
4.1
Invariant subspaces. Proper vectors and

values
In this part we shall further develop the theory of linear maps. Namely we are
interested in the structure of an operator.
Let us begin with a short description of what we expect to obtain.
Suppose that we have a vector space V over a eld F and a linear operator
T P EndpV q. Suppose further that we have the direct sum decomposition:
V
i1
Ui ,
where each Ui is a direct subspace of V. To understand the behavior of T it is only

needed to understand the behavior of each restriction T |Uj . Studying T |Uj should
be easier than dealing with T because Uj is a smaller vector space than V .
72
Invariant subspaces. Proper vectors and values
73
However we have a problem: if we want to apply tools which are commonly used in
the theory of linear maps (such as taking powers for example) the problem is that
generally T may not map Uj into itself, in other words T |Uj may not be an
operator on Uj . For this reason it is natural to consider only that kind of
decomposition for which T maps every Uj into itself.
Denition 4.1. Let V be an operator on the vector space V over F and U a
subspace of V . The subspace U is called invariant under T if T pUq U, in other
words T |U is an operator on U.
Of course that another natural question arises when dealing with invariant
subspaces. How does an operator behave on an invariant subspace of dimension
one? Every one dimensional subspace is of the form U tu| P Fu. If U is
invariant by T it follows that T puq should be in U, and hence there should exist a
scalar P F such that T puq u. Conversely if a nonzero vector u exists in V
such that T puq u, for some P F, then the subspace U spanned by u is
invariant under T and for every vector v in U one has T pvq v. It seems
reasonable to give the following denition:
Denition 4.2. Let T P EndpV q be an operator on a vector space over the eld F.
A scalar P F is called eigenvalue (or proper value) for T if there exists a nonzero
vector v P V such that T pvq v. A corresponding vector satisfying the above
equality is called eigenvector (or proper vector) associated to the eigenvalue .
The set of eigenvectors of T corresponding to an eigenvalue forms a vector space,
denoted by Epq, the proper subspace corresponding to the proper value . It is
clear that Epq kerpT IV q
For the nite dimensional case let MT be the matrix of T in some basis. The
equality T pvq v is equivalent with MT v v or, with pMT In qv 0, which
74
is a linear system. Obviously this homogeneous system of linear equations has a

nontrivial solution if and only if
detpMT In q 0.
Observe that detpMT In q is a polynomial of degree n in , where n dim V.
This polynomial is called the characteristic polynomial of the operator T. Hence,
the eigenvalues of T are the roots of its characteristic polynomial.
Notice that the characteristic polynomial does not depend on the choice of the
basis B that we choose when computing the matrix MT of the transformation T .
Indeed, let B 1 be another basis and MT1 the matrix of T with respect to this new
basis. Further, let P be transition matrix from B to B 1 . So MT1 P 1 MT P and
detpP q 0. We have
detpP 1MT P Iq detpP 1MT P P 1 pIqP q
detpP 1 pMT IqP q
1
detpMT Iq detpP q detpMT Iq,
detpP q
which proves our claim.

Theorem 4.3. Let T P EndpV q. Suppose that i , i 1, m are distinct eigenvalues
of T , and vi , i 1, m are the corresponding proper vectors. The set tv1 , . . . , vm u is
linearly independent.
Proof. Suppose, by contrary, that the set tv1 , . . . , vm u is linearly dependent. It
follows that a smallest index k exists such that
vk P spantv1 , . . . , vk1 u.
Thus the scalars a1 , . . . ak1 exist such that
vk a1 v1 ` . . . ak1 vk1 .
75
Applying T to the above equality, we get

k vk a1 1 v1 ` . . . ak1 k1 vk1 .
It follows that
0 a1 pk 1 qv1 ` ` ak1 pk k1 qvk1 .
Because we choose k to be the smallest index such that vk a1 v1 ` ` ak1 vk1 ,
it follows that the set tv1 , vk1 u is linearly independent. It follows that all the
as are zero.
Corollary 4.4. An operator T on a nite dimensional vector space V has at most
dim V distinct eigenvalues.
Proof. This is an obvious consequence of the fact that in a nite dimensional
vector space we have at most dim V linearly independent vectors.
The linear maps which have exactly n dim V linearly independent eigenvectors
have very nice and simple properties. This is the happiest case we can meet with
in the class of linear maps.
Denition 4.5. A linear map T : V V is said to be diagonalizable if there
exists a basis of V consisting of n independent eigenvectors, n dim V .
Recall that matrices A and B are similar if there is an invertible matrix P such
that B P AP 1. Hence, a matrix A is diagonalizable if it is similar to a diagonal
matrix D.
The minimal polynomial of an operator
4.2
76
The main reason for which there exists a richer theory on operators than for linear
maps is that operators can be raised to powers (we can consider the composition of
an operator with itself).
Let V be an n-dimensional vector space over a eld F and T : V V be a linear
operator.
Now, LpV, V q EndpV q is an n2 dimensional vector space. We can consider
T 2 T T and of course we obtain T n T n1 T inductively. We dene T 0 as
being the identity operator I IV on V . If T is invertible (bijective), then there
exists T 1 , so we dene T m pT 1 qm . Of course that
T m T n T m`n , for m, n P Z.
For T P EndpV q and p P FrXs a polynomial given by
ppzq a0 ` a1 z ` . . . am z m , z P F
we dene the operator ppT q given by
ppT q a0 I ` a1 T ` . . . am T m .
This is a new use of the same symbol p, because we are applying it to operators
not only to elements in F. If we x the operator T we obtain a function dened on
FrXs with values in EndpV q, given by p ppT q which is linear. For p, q P FrXs we
dene the operator pq given by ppqqpT q ppT qqpT q.
Now we begin the study of the existence of eigenvalues and of their properties.
Theorem 4.6. Every operator over a nite dimensional, nonzero, complex vector
space has an eigenvalue.
77
Proof. Suppose V is a nite dimensional complex vector space and T P EndpV q.

Choose v P V , v 0. Consider the set
pv, T pvq, T 2pvq, . . . T n pvqq.
This set is a linearly dependent system of vectors (they are n ` 1) vectors and
dim V n. Then there exist complex numbers, a0 , . . . an , not all 0, such that
0 a0 v ` a1 T pvq ` ` an T n pvq .
Let m be the largest index such that am 0. Then we have the decomposition
a0 ` a1 z ` ` am z m a0 pz 1 q . . . pz m q .
It follows that
0 a0 v ` a1 T pvq ` . . . an T n pvq
pa0 I ` a1 T ` . . . an T n qpvq
a0 pT 1 Iq . . . pT m Iqpvq .
which means that T j I is not injective for at least one j, or equivalently T has
an eigenvalue.
Remark 4.7. The analogous statement is not true for real vector spaces. But on
real vector spaces there are always invariant subspaces of dimension 1 or 2.
Example 4.8. Let T : F2 F2 given by T px, yq py, xq. It has no eigenvalues
and eigenvectors if F R. Find them for F C.
Obviously, T px, yq px, yq leads to py, xq px, yq, or equivalently
$
& x ` y 0
% y x 0.
78
The previous system is equivalent to x y, p2 ` 1qy 0.

If P R then the solution is x y 0, but note that p0, 0q is excluded from
eigenvectors by denition.
If P C we obtain the eigenvalues 1 i, 2 i and the corresponding
eigenvectors pi, 1q P C2 , respectively pi, 1q P C2 .
Theorem 4.9. Every operator on an odd dimensional real vector space has an
eigenvalue.
Proof. Let T P EndpV q and n dim V odd. The eigenvalues of T are the roots of
the characteristic polynomial that is detpMT In q. This polynomial is a
polynomial of degree n in , hence, since n is odd, the equation detpMT In q 0
has at least one real solution.
A central goal of linear algebra is to show that a given operator T P EndpV q has a
reasonably simple matrix in a given basis. It is natural to think that reasonably
simple means that the matrix has as many 01 s as possible.
Recall that for a basis tek , k 1, nu,
T pek q
aik ei ,
i1
where MT paij qi1,m is the matrix of the operator.

j1,n
Theorem 4.10. Suppose T P EndpV q and tei , i 1, nu is a basis of V . Then the

following statements are equivalent:
1. The matrix of T with respect to the basis tei , i 1, nu is upper triangular.
2. T pek q P spante1 , . . . , ek u for k 1, n.
3. spante1 , . . . , ek u is invariant under T for each k 1, n.
79
Proof. 12 obviously follows from a moments tought and the denition. Again
32. It remains only to prove that 23.
So, suppose 2 holds. Fix k P t1, . . . , nu. From 2 we have
T pe1 q P
spante1 u
spante1 , . . . , ek u
T pe2 q P
..
.
spante1 , e2 u
T pek q P
spante1 , . . . , ek u.
So, for v a linear combination of te1 , . . . , ek u one has that

T pvq P spante1 , . . . , ek u,
consequently 3. holds.
Theorem 4.11. Suppose that V is a complex vector space and T P EndpV q. Then
there exists a basis of V such that T is an upper-triangular matrix with respect to
this basis.
Proof. Induction on the dim V . Clearly this holds for dim V 1.
Suppose that dim V 1 and the result holds for all complex vector spaces of
dimension smaller then the dimension of V . Let be an eigenvalue of T (it exists)
and
U impT Iq.
Because T I is not surjective, dim U dim V . Furthermore U is invariant
under T , since for u P U there exists v P V such that u T pvq v, hence
T puq T pT pvqq T pvq pT Iqpwq P U where w T pvq.
80
So, T |U is an operator on U. By the induction hypothesis there is a basis

tu1 , . . . , um u of U with respect to which T |U has an upper-triangular matrix. So,
for each j P t1, . . . , mu we have
T puj q T |U puj q P spantu1 , . . . , um u.
Extend the basis tu1 , . . . , um u of U to a basis tu1 , . . . um , v1 , . . . vn u of V . For each
k 1, n
T pvk q pT Iqpvk q ` vk .
By the denition of U, pT Iqpvk q P U spantu1 , . . . , um u. Thus the equation
above shows that
T pvk q P spantu1 , . . . , um , vk u.
From this, in virtue of the previous theorem, it follows that T has an
upper-triangular matrix with respect to this basis.
One of the good points of this theorem is that, if we have this kind of basis, we can
decide if the operator is invertible by analysing the matrix of the operator.
Theorem 4.12. Suppose T P EndpV q has an upper triangular matrix with respect
to some basis of V . Then T is invertible if and only if all the entries on the
diagonal are non zero.
Proof. Let te1 , . . . , en u be a basis of V with respect to which T has the matrix
1 . . .
0
MT
...
...
81
We will prove that T is not invertible i one of the k s equals zero. If 1 0, then
T pv1 q 0, so T is not invertible, as desired.
Suppose k 0, 1 k n. The operator T maps the vectors e1 , . . . , ek1 in
spante1 , . . . , ek1 u and, because k 0, T pek q P te1 , . . . , ek1 u. So, the vectors
T pe1 q, . . . , T pek q are linearly dependent (they are k vectors in a k 1 dimensional
vector space, spante1 , . . . , ek1 u. Consequently T is not injective, and not
invertible.
Suppose that T is not invertible. Then ker T t0u, so v P V, v 0 exists such
that T pvq 0. Let
v a1 e1 ` ` an en
and let k be the largest integer with ak 0. Then
v a1 e1 ` ` ak ek ,
and
0 T pvq,
0 T pa1 e1 ` ` ak ek q,
0 pa1 T pe1 q ` ` ak1 T pek1 qq ` ak T pek q.
The term pa1 T pe1 q ` ` ak1 T pek1 qq is in spante1 , . . . , ek1 u, because of the
form of MT . Finally T pek q P spante1 . . . , ek1 u. Thus when T pek q is written as a
linear combination of the basis te1 , . . . , en u, the coecient of ek will be zero. In
other words, k 0.
Theorem 4.13. Suppose that T P EndpV q has an upper triangular matrix with
respect to some basis of V . Then the eigenvalues of T are exactly of the entries on
the diagonal of the upper triangular matrix.
Proof. Suppose that we have a basis te1 , . . . , en u such that the matrix of T is
upper triangular in this basis. Let P F, and consider the operator T I. It has
Diagonal matrices
82
the same matrix, except that on the diagonal the entries are i if those in the
matrix of T are j . It follows that T I is not invertible i is equal with some
j . So is proper value as desired.
4.3
Diagonal matrices
A diagonal matrix is a matrix which is zero except possibly the diagonal.

Proposition 4.14. If T P EndpV q has dim V distinct eigenvalues, then T has a
diagonal matrix
2
..
with respect to some basis.

Proof. Suppose that T has dim V distinct eigenvalues, 1 , . . . , n , where
n dim V . Choose corresponding eigenvectors e1 , . . . , en . Because nonzero vectors
corresponding to distinct eigenvalues are linearly independent, we obtain a set of
vectors with the cardinal equal to dim V , that is a basis, and in this basis the
matrix of T is diagonal.
The next proposition imposes several conditions on an operator that are equivalent
to having a diagonal matrix.
Proposition 4.15. Suppose T P EndV . Denote 1 , . . . , n the distinct eigenvalues
of T . The following conditions are equivalent.
1. T has a diagonal matrix with respect to some basis of V .
Diagonal matrices
83
2. V has a basis consisting of proper vectors.

3. There exists one dimensional subspaces U1 , . . . , Um of V , each invariant
under T such that
V U1 Um .
4. V kerpT 1 Iq kerpT n Iq.
5. dim V dim kerpT 1 Iq ` ` dim kerpT n Iq.
Proof. We saw that 1 2. Suppose 2 holds. Choose te1 , . . . , em u a basis
consisting of proper vectors, and Ui spantei u, for i 1, m. Hence 2 3.
Suppose 3 holds. Choose a basis ej P Uj , j 1, m. It follows that ej , j 1, m is a
proper vector, so they are linearly independent, and because they are m vectors,
they form a basis. Thus 3 implies 2.
Now we know that 1, 2, 3 are equivalent. Next we will prove the following chain of
implications
2452
Suppose 2 holds, then V has a basis consisting of eigenvectors. Then every vector
in V is a linear combination of eigenvectors of T , that is
V kerpT 1 Iq ` ` kerpT n Iq.
We show that it is a direct sum. Suppose that
0 u1 ` ` un ,
with uj P kerpT j Iq, j 1, n. They are linearly independent, so all are 0.
Finally 4 5 is clear because in 4 we have a direct sum.
5 2. dim V dim kerpT 1 Iq ` ` dim kerpT n Iq. According to a
precedent result, distinct eigenvalues give rise to linear independent eigenvectors.
Diagonal matrices
84
Let te11 , . . . , e1i1 u, . . . , ten1 , . . . , enin u bases in kerpT 1 Iq, . . . , kerpT n Iq. Then
dim V i1 ` ` in , and te11 , . . . , e1i1 , . . . , en1 , . . . , enin u are linearly independent.
Hence V spante11 , . . . , e1i1 , . . . , en1 , . . . , enin u which shows that 2 holds.
2 1 1
Example 4.16. Consider the matrix A 1 2 1 . Show that A is
1 1 0
diagonalizable and nd the diagonal matrix similar to A.
The characteristic polynomial of A is
detpA Iq 3 ` 42 6 p ` 1qp 2qp 3q.
Hence, the eigenvalues of A are 1 1, 2 2 and 3 3. To nd the
corresponding eigenvectors, we have to solve the three linear systems pA ` Iqv 0,
pA 2Iqv 0 and pA 3Iqv 0. On solving these systems, we nd that the
solution spaces are
tp, , 2q : P Ru,
tp, , q : P Ru,
respectively
tp, , 0q : P Ru.
Hence, the corresponding eigenvectors associated to 1 , 2 and 3 respectively, are
v1 p1, 1, 2q, v2 p1, 1, 1q and v3 p1, 1, 0q respectively. There exists 3 linear
independent eigenvectors, thus A is diagonalizable.
1 1
1
Our transition matrix P rv1 |v2 |v3 s 1 1 1 .
2 1 0
Diagonal matrices
85
1 1
We have P 1 16 2 2
3 3
Hence, the diagonal matrix
2 .
0
similar to A is
1 0 0
D P 1AP 0 2 0 .
0 0 3
Obviously one may directly compute D, by knowing, that D is the diagonal matrix
having the eigenvalues of A on its main diagonal.
Proposition 4.17. If is a proper value for an operator (endomorphism) T , and
v 0, v P V is a proper vector then one has:
1. @k P N, k is a proper value for T k T T (k times) and v is a proper
vector of T k .
2. If p P FrXs is a polynomial with coecients in F, then ppq is an eigenvalue
for ppT q and v is a proper vector of ppT q.
3. For T automorphism (bijective endomorphism), 1 is a proper value for T 1
and v is an eigenvector for T 1 .
Proof. 1. We have T pvq v, hence T T pvq T pvq T pvq 2 v. Assume,
that T k1pvq k1 v. Then T k pvq
T T k1 pvq T pT k1 pvqq T pk1 vq k1T pvq k1 v k v.
2. Let p a0 ` a1 x ` ` an xn P FrXs. Then ppT qpvq
a0 Ipvq ` a1 T pvq ` ` an T n pvq a0 v ` a1 pvq ` ` an pn vq ppqv.
The Jordan canonical form
86
3. T 1 pvq u such that T puq v. But v 1 T pvq T p1 vq, hence

T puq T p1 vq. Since T is injective we have u 1 v, or equivalently
T 1 pvq 1 v.
Example 4.18. Let T : V V be a linear map. Prove that if 1 is an eigenvalue
of T 2 ` T then 1 is an eigenvalue of T 3 . Here I is the identity map and
T 2 T T , etc.
From the fact that 1 is an eigenvalue of T 2 ` T there exists v 0 such that
` 2
T ` T v v,
or, equivalently
`
T 2 ` T ` I v 0.
Now, we apply the linear map T I (recall that the linear maps form a vector
space, so the sum or dierence of two linear maps is still linear) to the above
relation to get
`
pT Iq T 2 ` T ` I v 0.
Here we have used that, by linearity, pT Iq 0 0.
Finally, simple algebra yields pT Iq pT 2 ` T ` Iq T 3 I, so the above equation
shows that
T 3 v v,
as desired.
4.4
In a previous sections we have seen the endomorphisms which are diagonalizable.
87
Let V be a vector space of nite dimension n over a eld F. Let T : V V and let
0 be an eigenvalue of T . Consider the matrix form of the endomorphism in a
given basis, T pvq MT v. The eigenvalues are the roots of the characteristic
polynomial detpMt In q 0. It can be proved that this polynomial does not
depend on the basis and of the matrix MT . So, it will be called the characteristic
polynomial of the endomorphism T , and it will be denoted by P pq, and of course
deg P n. Sometimes it is called the characteristic polynomial of the matrix, but
we understand that is the matrix associated to an operator.
Denote by mp0 q the multiplicity of 0 as a root of this polynomial. Associated to
the proper value 0 we consider the proper subspace corresponding to 0 :
Ep0 q tv P V |T pvq 0 vu.
Consider a basis of V and let MT be the matrix of T with respect to this basis. We
havev that:
Theorem 4.19. With the above notations, the following holds
dim Ep0 q n rank pMT 0 Iq mp0 q.
Proof. Obviously is enough to prove the claim in V Rn . Let x1 , x2 , . . . , xr be
linearly independent eigenvectors associated to 0 , so that dim Ep0 q r.
Complete this set with xr`1 , . . . xn to a basis of Rn . Let P be the matrix whose
columns are xi , i 1, n. We have MT P r0 x1 | . . . |0 xr | . . . s. We get that the
rst r columns of P 1MT P are diagonal with 0 on the diagonal, but that the rest
of the columns are indeterminable. We prove next that P 1 MT P has the same
characteristic polynomial as MT . Indeed
detpP 1MT P Iq detpP 1MT P P 1 pIqP q
detpP 1 pMT IqP q
88
1
detpMT Iq detpP q detpMT Iq.
detpP q
But since the rst few columns of P 1 MT P are diagonal with 0 on the diagonal
we have that the characteristic polynomial of P 1MT P has a factor of at least
p0 qr , so the algebraic multiplicity of 0 is at least r.
The value dim Ep0 q is called the geometric multiplicity of the eigenvalue 0 .
Let T P EndpV q, and suppose that the roots of the characteristic polynomial are in
F. Let be a root of the characteristic polynomial, i.e. an eigenvalue of T .
Consider m the algebraic multiplicity of and q dim Epq, the geometric
multiplicity of .
It is possible to nd q eigenvectors and m q principal vectors (also called
generalized eigenvectors), all of them linearly independent, and an eigenvector v
and the corresponding principal vectors u1 , . . . , ur satisfy
T pvq v, T pu1q u1 ` v, . . . , T pur q ur ` ur1
The precedent denition can equivalently be stated as
A nonzero vector u is called a generalized eigenvector of rank r associated with the
eigenvalue if and only if pT Iqr puq 0 and pT Iqr1 puq 0. We note that
a generalized eigenvector of rank 1 is an ordinary eigenvector. The previously
dened principal vectors u1 , . . . , ur are generalized eigenvectors of rank 2, . . . , r ` 1.
It is known that if is an eigenvalue of algebraic multiplicity m, then there are m
linearly independent generalized eigenvectors associated with .
These eigenvectors and principal vectors associated to T by considering all the
eigenvalues of T form a basis of V , called the Jordan basis with respect to T . The
89
matrix of T relative to a Jordan basis is called a Jordan matrix, and it has the form
J2
..
.
Jp
The Js are matrices, called Jordan cells. Each cell represents the contribution of
an eigenvector v, and the corresponding principal vectors, u1 , . . . ur , and it has the
form
1
..
. 1
P Mr`1 pFq
It is easy to see that the Jordan matrix is a diagonal matrix i there are no
principal vectors i mpq dim Epq for each eigenvalue .
Let MT be the matrix of T with respect to a given basis B, and J be the Jordan
matrix with respect to a Jordan basis B 1 . Late P be the transition matrix from B
to B 1 , hence it have columns consisting of either eigenvectors or generalized
eigenvectors. Then J P 1MT P , hence MT P JP 1 .
Example 4.20. (algebraic multiplicity
0 1
operator with the matrix A 4 4
2 1
transition matrix of A.
3,geometric multiplicity 2) Consider the

0
0 . Find the Jordan matrix and the
90
The characteristic polynomial of A is detpA Iq p2 q3 , hence 2 is an

eigenvalue with algebraic multiplicity 3. By solving the homogenous system
pA 2Iqv 0 we obtain the solution space
Ep2q kerpA 2Iq tp, 2, q : , P Ru. Hence the dimension of Ep2q is 2,
consequently the eigenvalue 2 has geometric multiplicity 2. Therefore we can
take the linear independent eigenvectors v1 p1, 2, 1q respectively v2 p0, 0, 1q.
Note that we need a generalized eigenvector, which can be obtained as a solution
of the system
pA 2Iqu v1 .
The solutions of this system lie in the set tp, 2 ` 1, q : , P Ru, hence a
generalized eigenvector, is u1 p1, 3, 0q.
hence we take
Note that v1 , u1 , v
2 are linear
independent,
the transition matrix
1 1 0
3 1 0
P rv1 |u1 |v2 s 2 3 0 . Then P 1 2 1 0 , hence
1 0 1
3 1 1
2 1 0
J P 1 AP 0 2 0 .
0 0 2
Example 4.21. (algebraic multiplicity
3, geometric
multiplicity 1) Consider the
1 18 7
operator with the matrix A 1 13 4 . Find the Jordan matrix and the
1 25
8
transition matrix of A.
The characteristic polynomial of A is detpA Iq p ` 2q3 , hence 2 is an
eigenvalue with algebraic multiplicity 3. By solving the homogenous system
pA ` 2Iqv 0 we obtain the solution space
Problems
91
Ep2q kerpA ` 2Iq tp5, 3, 7q : P Ru. Hence the dimension of Ep2q is 1,

consequently the eigenvalue 2 has geometric multiplicity 1. Therefore we can
take the linear independent eigenvector v p5, 3, 7q. Note that we need two
generalized eigenvectors, which can be obtained as a solution of the system
pA ` 2Iqu1 v,
respectively
pA ` 2Iqu2 u1 .
, 2`3
, q : P Ru, hence a
The solutions of the rst system lie in the set tp 1`5
7
7
generalized eigenvector, for 4 is u1 p3, 2, 4q.
The solutions of the system pA ` 2Iqu2 u1 with u1 p3, 2, 4q lie in the set
tp 3`5
, 13
, q : P Ru, hence a generalized eigenvector, for 5 is
7
7
hence we take the
u1 p4, 2, 5q. Note that v, u1 , u2are linear independent,
5 3 4
transition matrix P rv1 |u1 |u2 s 3 2 2 . Then
7 4
5
2 1 2
P 1 1 3 2 , hence
2 1 1
2 1
0
1
J P AP 0 2 1 .
0
0 2
4.5
Problems
Problem 4.5.1. Find the eigenvalues and eigenvectors of the operator

T : C 8 p1, bq C 8 p1, bq, T pf qpxq
f 1 pxq
.
xex2
Problems
92
Problem 4.5.2. Find matrices which diagonalize the following: aq
1 5
3 3
1 2 1
bq 1 0
1 .
4 4 5
Problem 4.5.3. Find the Jordan canonical form and the transition matrix for the
matrix
2 1 1
3 2 3 .
2 2 3
Problem 4.5.4. Prove that a square matrix and its transpose have the same
eigenvalues.
matrix
6 6 15
1 5 5 .
1 2 2
T : Cr, s Cr, s,
T pf qpxq
px cos y ` sin x sin yqf pyqdy.
matrix
2 2 2 .
1 1 4
Problems
93

T : Cr, s Cr, s,
T pf qpxq
pcos3 px yq ` 1qf pyqdy.
matrix
7 12 6
10 19 10 .
12 24 13
f 1 pxq
.
sin2 x
T : C 8 p1, 2q C 8 p1, 2q, T pf qpxq
1 1
.
Problem 4.5.11. Triangularize the matrix A
1 3
Problem 4.5.12. Find the Jordan canonical form and the transition matrix for
the matrix
4 5 2
5 7 3 .
6 9 4

T : C 8 p1, bq C 8 p1, bq, T pf qpxq
f 1 pxq
.
tan2 x
the matrix
4 2 1 .
4
1 2
Problems
94
Problem 4.5.15. Prove that a complex 2

2 matrix
is not diagonalizable if and
a b
, where b 0.
only if it is similar to a matrix of the form
0 a
the matrices
1 3 3
4 6 15
2 6 13 , 1 3 5 .
1 4 8
1 2 4
Problem 4.5.17. Prove that if A and B are n n matrices, then AB and BA

have the same eigenvalues.
the matrix
2 6 15
1 1 5 .
1 2 6
5
Inner product spaces
5.1
Basic denitions and results
Up to now we have studied vector spaces, linear maps, special linear maps.
We can measure if two vectors are equal, but we do not have something like
length, so we cannot compare two vectors. Moreover we cannot say anything
about the position of two vectors.
In a vector space one can dene the norm of a vector and the inner product of two
vectors. The notion of the norm permits us to measure the length of the vectors,
and compare two vectors. The inner product of two vectors, on one hand induces a
norm, so the length can be measured, and on the other hand (at least in the case
of real vector spaces), lets us measure the angle between two vectors, so a full
geometry can be constructed there. Nevertheless in the case of complex vector
spaces, the angle of two vectors is not clearly dened, but the orthogonality is.
Denition 5.1. An inner product on a vector space V over the eld F is a
function (bilinear form) x, y : V V R with the properties:
(positivity and deniteness) xv, vy 0 and xv, vy 0 i v 0.
95
96
(additivity in the rst slot) xu ` v, wy xu, wy ` xv, wy, for all u, v, w P V.

(homogeneity in the rst slot) xv, wy xv, wy for all P F and v, w P V.
(conjugate symmetry) xv, wy xw, vy for all v, w P V .
An inner product space is a pair pV, x, yq, where V is vector space and x, y is an
inner product on V .
The most important example of an inner product space is Fn . Let v pv1 , . . . , vn q
and w pw1 , . . . wn q and dene the inner product by
xv, wy v1 w 1 ` ` vn w n .
This is the typical example of an inner product, called the Euclidean inner
product, and when Fn is referred to as an inner product space, one should assume
that the inner product is the Euclidean one, unless explicitly stated otherwise.
a b
be a positive denite matrix, that
Example 5.2. Let A P M2 pRq, A
b c
2
is a 0, detpAq
for every u pu1 , u2q, v pv1 , v2 q P R we dene
0. Then
xu, vy pv1 v2 qA
u1
u2
It can easily be veried that x, y is an inner product on the real linear space R2 .
If A I2 we obtain the usual inner product xu, vy u1 v1 ` u2 v2 .

From the denition one can easily deduce the following properties of an inner
product:
xv, 0y x0, vy 0,
xu, v ` wy xu, vy ` xu, wy,
xu, vy xu, vy,
for all u, v, w P V and P F
97
Denition 5.3. Let V be a vector space over F. A function

}}:V R
is called a norm on V if:
(positivity) }v} 0, v P V, }v} 0 v 0 ;
(homogeneity) }v} || }v}, @ P F, @v P V ;
(triangle inequality) }u ` v} }u} ` }v}, @u, v P V.
A normed space is a pair pV, } }q, where V is a vector space and } } is a norm on
V.
Example 5.4. On the real linear space Rn one can dene a norm in several ways.
Indeed, for any x px1 , x2 , . . . , xn q P Rn dene its norm as
a
}x} x21 ` x22 ` ` x2n . One can easily verify that the axioms in the denition
of norm are satised. This norm is called the euclidian norm.
More generally, for any p P R, p 1 we can dene
1
}x} p|x1 |p ` |x2 |p ` ` |xn |p q p , the so called pnorm on Rn .

Another way to dene a norm on Rn is }x} maxt|x1 |, |x2 |, . . . , |xn |u. This is the
so called maximum norm.
Denition 5.5. Let X be a nonempty set. A function d : X X R satisfying
the following properties:
(positivity) dpx, yq 0, @x, y P X and dpx, yq 0 x y;
(symmetry) dpx, yq dpy, xq, @x, y P X;
(triangle inequality) dpx, yq dpx, zq ` dpz, yq, @x, y, z P X;
98
is called a metric or distance on X. A set X with a metric dened on it is called a

metric space.
Example 5.6. Let X be an arbitrary set. One can dene a distance on X by
$
& 0, if x y
dpx, yq
% 1, otherwise.
This metric is called the discrete metric on X. On Rn the Chebyshev distance is
dened as
dpx, yq max |xi yi |, x px1 , x2 , . . . , xn q, y py1 , y2 , . . . , yn q P Rn .
1in
In this course we are mainly interested in the inner product spaces. But we should
a
point out that an inner product on V denes a norm, by }v} xv, vy for v P V ,
and a norm on V denes a metric by dpv, wq }w v}, for v, w P V .
On the other hand, from their generality point of view the metrics are the most
general ones (can be dened on any set), followed by norms (which assumes the
linearity of the space where is dened) and on the last position is the inner
product. It should be pointed that every inner product generates a norm, but not
every norm comes from an inner product, as is the case for the max norm dened
above.
For an inner product space pV, x, yq the following identity is true:
C
i1
i vi ,
j1
G
j wj
m
n
i j xvi , wj y.
i1 j1
Denition 5.7. Two vectors u, v P V are said to be orthogonal (uKv) if xu, vy 0.

In a real inner product space we can dene the angle of two vectors as
xv, wy
{
pv,
wq arccos
}v} }w}
99
We have
{
vKw xv, wy 0 pv,
wq .
2
Theorem 5.8. (Parallelogram law) Let V be an inner product space and
u, v P V . Then
}u ` v}2 ` }u v}2 2p}u}2 ` }v}2 q.
Proof.
}u ` v}2 ` }u v}2 xu ` v, u ` vy ` xu v, u vy xu, uy ` xu, vy ` xv, uy ` xv, vy
`xu, uy xu, vy xv, uy ` xv, vy
2p}u}2 ` }v}2q.
Theorem 5.9. (Pythagorean Theorem) Let V be an inner product space, and

u, v P V orthogonal vectors. Then
}u ` v}2 }u}2 ` }v}2 .
Proof.
}u ` v}2 xu ` v, u ` vy
xu, uy ` xu, vy ` xv, uy ` xv, vy
}u}2 ` }v}2 .
Now we are going to prove one of the most important inequalities in mathematics,
namely the Cauchy-Schwartz inequality. There are several methods of proof for
this, we will give one related to our aims.
100
Consider u, v P V . We want to write u as a sum between a vector collinear to v

and a vector orthogonal to v. Let P F and write u as u v ` pu vq.
Imposing now the condition that v is orthogonal to pu vq, one obtains
0 xu v, vy xu, vy }v}2,
so one has to choose
xu,vy
,
}v}2
and the decomposition is
xu, vy
xu, vy
u
v` u
v .
}v}2
}v}2
Theorem 5.10. Cauchy-Schwartz Inequality Let V be an inner product space
and u, v P V . Then
|xu, vy| }u} }v}.
The equality holds i one of u, v is a scalar multiple of the other (u and v are
collinear).
Proof. Let u, v P V . If v 0 both sides of the inequality are 0 and the desired
xu,vy
xu,vy
result holds. Suppose that v 0. Write u }v}2 v ` u }v}2 v . Taking into
account that the vectors
xu,vy
v
}v}2
and u
xu,vy
v
}v}2
are orthogonal, by the Pythagorean
theorem we obtain
}u}
2
xu, vy 2
xu,
vy
` u

v
v
}v}2
}v}2
2
|xu, vy|2
xu, vy
` u
v
}v}2
}v}2
|xu, vy|2
,
}v}2
inequality equivalent with the one in the theorem.

We have equality i u
xu,vy
v
}v}2
0, that is i u is a scalar multiple of v.
Orthonormal Bases
5.2
101
Orthonormal Bases
Denition 5.11. Let pV, x, yq an inner product space and let I be an arbitrary
index set. A family of vectors A tei P V |i P Iu is called an orthogonal family, if
xei , ej y 0 for every i, j P I, i j. The family A is called orthonormal if it is
orthogonal and }ei } 1 for every i P I.
One of the reason that one studies orthonormal families is that in such special
bases the computations are much more simple.
Proposition 5.12. If pe1 , e2 , . . . , em q is an orthonormal family of vectors in V ,
then
}1 e1 ` 2 e2 ` ` m em }2 |1 |2 ` |2 |2 ` ` |m |2
for all 1 , 2 , . . . , m P F.
Proof. Apply Pythagorean Theorem, that is
}1 e1 ` 2 e2 ` ` m em }2 |1 |2 }e1 }2 ` |2 |2 }e2 }2 ` ` |m |2 }en }2 .
The conclusion follows taking into account that }ei } 1, i 1, n.
Corollary 5.13. Every orthonormal list of vectors is linearly independent.
Proof. Let pe1 , e2 , . . . , em q be an orthonormal list of vectors in V and
1 , 2 , . . . , m P F with
1 e1 ` 2 e2 ` ` m em 0.
It follows that }1 e1 ` 2 e2 ` ` m em }2 |1 |2 ` |2 |2 ` ` |m |2 0, that is
j 0, j 1, m.
Orthonormal Bases
102
An orthonornal basis of an inner product vector space V is a basis of V which is

also an orthonormal list of V . It is clear that every orthonormal list of vectors of
length dim V is an orthonormal basis (because it is linearly independent, being
orthonormal).
Theorem 5.14. Let pe1 , e2 , . . . , en q be an orthonormal basis of an inner product
space V . If v 1 e1 ` 2 e2 ` ` n en P V , then
i xv, ei y, for all i P t1, 2, . . . , nu and
}v}
2
|xv, ei y|2
i1
Proof. Since v 1 e1 ` 2 e2 ` ` n en , by taking the inner product in both

sides with ei we have
xv, ei y 1 xe1 , ei y ` 2 xe2 , ei y ` ` i xei , ei y ` ` n xen , ei y i .
The second assertion comes from applying the previous proposition. Indeed,
}v}2 }1 e1 ` ` n en }2 |1 |2 ` ` |n |2
|xv, ei y|2 .
i1
Up to now we have an image about the usefulness of orthonormal basis. The

advantage is that in an orthonormal basis the computations are easy, as in the
euclidean two or three dimensional spaces. But how does one go to nd them?
The next result gives an answer to the question. The following result is a well
known algorithm in linear algebra, called the Gram-Schmidt procedure. The
procedure is pointed here, giving a method for turning a linearly independent list
into an orthonormal one, with the same span as the original one.
Theorem 5.15. Gram-Schmidt If pv1 , v2 , . . . , vm q is a linearly independent set
of vectors in V , then there exists an orthonormal set of vectors pe1 , . . . em q in V,
Orthonormal Bases
103
such that
spanpv1 , v2 , . . . , vk q spanpe1 , e2 . . . , ek q
for every k P t1, 2, . . . , mu.
Proof. Let pv1 , v2 , . . . , vm q be a linearly independent set of vectors. The family of
orthonormal vectors pe1 , e2 . . . , em q will be constructed inductively. Start with
e1
v1
.
}v1 }
Suppose now that j 1 and an orthonormal family pe1 , e2 , . . . , ej1 q has
been constructed such that

spanpv1 , v2 , . . . , vj1q spanpe1 , e2 , . . . , ej1 q
Consider
ej
vj xvj , e1 ye1 xvj , ej1yej1

}vj xvj , e1 ye1 xvj , ej1yej1 }
Since the list pv1 , v2 , . . . , vm q is linearly independent, it follows that vj is not in

spanpv1 , v2 , . . . , vj1q, and thus is not in spanpe1 , e2 , . . . , ej1q. Hence ej is well
dened, and }ej } 1. By direct computations it follows that for 1 k j one has
vj xvj , e1 ye1 xvj , ej1yej1

xej , ek y
, ek
}vj xvj , e1 ye1 xvj , ej1yej1 }
xvj , ek y xvj , ek y
}vj xvj , e1 ye1 xvj , ej1 yej1}

0,
B
thus pe1 , e2 , . . . ek q is an orthonormal family. By the denition of ej one can see

that vj P spanpe1 , e2 , . . . , ej q, which gives (together with our hypothesis of
induction), that
spanpv1 , v2 , . . . , vj q spanpe1 , e2 , . . . , ej q
Orthonormal Bases
104
Both lists being linearly independent (the rst one by hypothesis and the second
one by orthonormality), it follows that the generated subspaces above have the
same dimension j, so they are equal.
Remark 5.16. If in the Gram-Schmidt process we do not normalize the vectors
we obtain an orthogonal basis instead of an orthonormal one.
Example 5.17. Orthonormalize the following list of vectors in R4 :
tv1 p0, 1, 1, 0q, v2 p0, 4, 0, 1q, v3 p1, 1, 1, 0q, v4 p1, 3, 0, 1qu.
First we will orthogonalize by using the Gram-Schmidt procedure.
Let u1 v1 p0, 1, 1, 0q.
xv2 , u1y
4
u1 p0, 4, 0, 1q p0, 1, 1, 0q p0, 2, 2, 1q.
xu1 , u1 y
2
xv3 , u1 y
xv3 , u2 y
1 1 4
.
u3 v3
u1
u2 1, , ,
xu1, u1 y
xu2, u2 y
9 9 9
1
2
xv4 , u1 y
xv4 , u2y
xv4 , u3y
1 1
u4 v4
u1
u2
u3
, , ,
.
xu1, u1 y
xu2 , u2 y
xu3 , u3 y
11 22 22 11
u2 v2
It can easily be veried that the list tu1 , u2, u3 , u4 u is orthogonal. Take now
wi
ui
,
}ui }
i 1, 4. We obtain
1 1
w1 0, ? , ? , 0 ,
2 2
2 2 1
,
w2 0, , ,
3 3 3
1
1
4
3
w3 ? , ? , ? , ?
,
11 3 11 3 11 3 11
?
?
?
?
22 22
22 2 22
w4
,
,
,
.
11 22
22
11
Obviously the list tw1 , w2 , w3 , w4 u is orthonormal.

Now we can state the main results in this section.
Orthogonal complement
105
Corollary 5.18. Every nitely dimensional inner product space has an

orhtonormal basis.
Proof. Choose a basis of V , apply the Gram-Schmidt procedure to it and obtain
an orthonormal list of length equal to dim V . It follows that the list is a basis,
being linearly independent.
The next proposition shows that any orthonormal list can be extended to an
orthonormal basis.
Proposition 5.19. Every orhtonormal family of vectors can be extended to an
orthonormal basis of V .
Proof. Suppose pe1 , e2 , . . . , em q is an orthonormal family of vectors.. Being linearly
independent, it can be extended to a basis, pe1 , e2 , . . . , em , vm`1 , . . . , vn q. Applying
now the Gram-Schmidt procedure to pe1 , e2 , . . . , em , vm`1 , . . . , vn q, we obtain the
list pe1 , e2 , . . . , em , fm`1 , . . . , fn q, (note that the Gram Schmidt procedure leaves
the rst m entries unchanged, being already orthonormal). Hence we have an
extension to an orthonormal basis.
Corollary 5.20. Suppose that T P EndpV q. If T has an upper triangular form
with respect to some basis of V , then T has an upper triangular form with respect
to some orthonormal basis of V .
Corollary 5.21. Suppose that V is a complex vector space and T P EndpV q. Then
T has an upper triangular form with respect to some orthonormal basis of V .
5.3
Let U V be a subset of an inner product space V . The orthogonal complement

of U, denoted by U K is the set of all vectors in V which are orthogonal to every
106
vector in U i.e.:
U K tv P V |xv, uy 0, @u P Uu.
It can easily be veried that U K is a subspace of V , V K t0u and t0uK V , as
well that U1 U2 U2K U1K .
Theorem 5.22. If U is a subspace of V , then
V U UK
Proof. Suppose that U is a subspace of V . We will show that
V U ` UK
Let te1 , . . . , em u be an orthonormal basis of U and v P V . We have
v pxv, e1ye1 ` ` xv, em yem q ` pv xv, e1 ye1 xv, em yem q
Denote the rst vector by u and the second by w. Clearly u P U. For each
j P t1, 2, . . . , mu one has
xw, ej y xv, ej y xv, ej y
0
Thus w is orthogonal to every vector in the basis of U, that is w P U K ,
consequently
V U ` U K.
We will show now that U X U K t0u. Suppose that v P U X U K . Then v is
orthogonal to every vector in U, hence xv, vy 0, that is v 0. The relations
V U ` U K and U X U K t0u imply the conclusion of the theorem.
Proposition 5.23. If U1 , U2 are subspaces of V then
107
a) U1 pU1K qK .
b) pU1 ` U2 qK U1K X U2K .
c) pU1 X U2 qK U1K ` U2K .
Proof. a) We show rst that U1 pU1K qK . Let u1 P U1 . Then for all v P U1K one has
vKu1 . In other words xu1 , vy 0 for all v P U1K . Hence u1 P pU1K qK .
Assume now that pU1K qK U1 . Hence, there exists u2 P pU1K qK zU1 . Since
V U1 U1K we obtain that there exists u1 P U1 such that u2 u1 P U1K pq.
On the other hand, according to the rst part of proof u1 P pU1K qK and pU1K qK is a
linear subspace, hence u2 u1 P pU1K qK . Hence, for all v P U1K we have
pu2 u1 qKv pq.
pq and pq implies that pu2 u1 qKpu2 u1 q that is xu2 u1 , u2 u1 y 0, which
leads to u1 u2 contradiction.
b) For v P pU1 ` U2 qK one has xv, u1 ` u2 y 0 for all u1 ` u2 P U1 ` U2 . By taking
u2 0 we obtain that v P U1K and by taking u1 0 we obtain that v P U2K . Hence
pU1 ` U2 qK U1K X U2K .
Conversely, let v P U1K X U2K . Then xv, u1y 0 for all u1 P U1 and xv, u2y 0 for all
u2 P U2 . Hence xv, u1 ` u2y 0 for all u1 P U1 and u2 P U2 , that is v P pU1 ` U2 qK .
c) According to a) ppU1 X U2 qK qK U1 X U2 .
According to b) and a) pU1K ` U2K qK pU1K qK X pU2K qK U1 X U2 .
Hence, ppU1 X U2 qK qK pU1K ` U2K qK which leads to pU1 X U2 qK U1K ` U2K .
Example 5.24. Let U tpx1 , x2 , x3 , x4 q P R4 |x1 x2 ` x3 x4 0u. Knowing
that U is a subspace of R4 , compute dim U and U K .
Linear manifolds
108
It is easy to see that

U
px1 , x2 , x3 , x4 q P R4 |x1 x2 ` x3 x4 0
tpx1 , x2 , x3 , x1 x2 ` x3 q |x1 , x2 , x3 P Ru
tx1 p1, 0, 0, 1q ` x2 p0, 1, 0, 1q ` x3 p0, 0, 1, 1q |x1 , x2 , x3 P Ru
span tp1, 0, 0, 1q , p0, 1, 0, 1q , p0, 0, 1, 1qu .
The three vectors p1, 0, 0, 1q , p0, 1, 0, 1q , p0, 0, 1, 1q are linearly independent (the
rank of the matrix they form is 3), so they form a basis of U and dim U 3.
The dimension formula
dim U ` dim U K dim R4
tells us that dim U K 1, so U K is generated by a single vector. A vector that
generates U K is p1, 1, 1, 1q, the vector formed by the coecients that appear in
the linear equation that denes U. This is true because the right hand side of the
equation is exactly the scalar product between uK p1, 1, 1, 1q and a vector
v px1 , x2 , x3 , x4 q P U.
5.4
Linear manifolds
Let V be a vector space over the eld F.

Denition 5.25. A set L v0 ` VL tv0 ` v|v P VL u , where v0 P V is a vector
and VL V is a subspace of V is called a linear manifold (or linear variety). The
subspace VL is called the director subspace of the linear variety.
Remark 5.26. One can easily verify the following.
A linear manifold is a translated subspace, that is L f pVL q where
f : V V , f pvq v0 ` v.
Linear manifolds
109
if v0 P VL then L VL .
v0 P L because v0 v0 ` 0 P v0 ` VL .
for v1 , v2 P L we have v1 v2 P VL .
for every v1 P L we have L v1 ` VL .
L1 L2 , where L1 v0 ` VL1 and L2 v01 ` VL2 i VL1 VL2 and
v0 v01 P VL1 .
Denition 5.27. We would like to emphasize that:
1. The dimension of a linear manifold is the dimension of its director subspace.
2. Two linear manifolds L1 and L2 are called orthogonal if VL1 KVL2 .
3. Two linear manifolds L1 and L2 are called parallel if VL1 VL2 or VL2 VL1 .
Let L v0 ` VL be a linear manifold in a nitely dimensional vector space V . For
dim L k n dim V one can choose in the director subspace VL a basis of nite
dimension tv1 , . . . , vk u. We have
L tv v0 ` 1 v1 ` ` k vk |i P F, i 1, ku
We can consider an arbitrary basis (xed) in V , lets say E te1 , . . . , en u and if we
use the column vectors for the coordinates in this basis, i.e.
vrEs px1 , . . . , xn qJ , v0rEs px01 , . . . , x0n qJ , vjrEs px1j , . . . , xnj qJ , j 1, k, one has
the parametric equations of the linear manifold
Linear manifolds
110
x1 x01 ` 1 x11 ` ` k x1k
&
..
.
%xn x0n ` 1 xn1 ` ` k xnk

The rank of the matrix pxij qi1,n is k because the vectors v1 , . . . , vk are linearly
j1,k
independent.
It is worthwhile to mention that:

1. a linear manifold of dimension one is called line.
2. a linear manifold of dimension two is called plane.
3. a linear manifold of dimension k is called k plane.
4. a linear manifold of dimension n 1 in an n dimensional vector space is
called hyperplane.
Theorem 5.28. Let us consider V an n-dimensional vector space over the eld F.
Then any subspace of V is the kernel of a surjective linear map.
Proof. Suppose VL is a subspace of V of dimension k. Choose a basis te1 , . . . , ek u
in VL and complete it to a basis te1 , . . . , ek , ek`1 , . . . , en u of V . Consider
U spantek`1 , . . . , en u. Let T : V U given by
T pe1 q 0, . . . T pek q 0, T pek`1q ek`1 , . . . , T pen q en .
Obviously, T p1 e1 ` ` n en q 1 T pe1 q ` ` n T pen q k`1 ek`1 ` ` n en
denes a linear map. It is also clear that ker T VL as well that T is surjective,
i.e. im T U.
Linear manifolds
111
Remark 5.29. In fact the map constructed in the previous theorem is nothing
but the projection on U parallel to the space spante1 , . . . , ek u.
Theorem 5.30. Let V, U two linear spaces over the same eld F. If T : V U is
a surjective linear map, then for every u0 P U, the set L tv P V |T pvq u0 u is a
linear manifold.
Proof. T being surjective, there exists v0 P V with T pv0 q u0 . We will show that
tv v0 |v P Lu ker T .
Let v P L. We have T pv v0 q T pvq T pv0 q 0, so tv v0 |v P Lu ker T .
Let v1 P ker T , i.e. T pv1 q 0. Write v1 pv1 ` v0 q v0 . T pv1 ` v0 q u0, so
pv1 ` v0 q P L. Hence, v1 P tv v0 |v P Lu or, in other words ker T tv v0 |v P Lu.
Consequently L v0 ` ker T, which shows that L is a linear manifold.
The previous theorems give rise to the next:
Theorem 5.31. Let V a linear space of dimension n. Then, for every linear
manifold L V of dimension dim L k n, there exists an n k-dimensional
vector space U, a surjective linear map T : V U and a vector u P U such that
L tv P V |T pvq uu.
Proof. Indeed, consider L v0 ` VL , where the dimension of the director subspace
VL k. Choose a basis te1 , . . . , ek u in VL and complete it to a basis
te1 , . . . , ek , ek`1, . . . , en u of V . Consider U spantek`1 , . . . , en u. Obviously
dim U n k. According to a previous theorem the linear map
T : V U, T p1 e1 ` ` k ek ` k`1 ek`1 ` ` n en q k`1 ek`1 ` ` n en is
surjective and ker T VL . Let T pv0 q u. Then, according to the proof of the
previous theorem L tv P V |T pvq uu.
Linear manifolds
112
Remark 5.32. If we choose in V and U two bases and we write the linear map by
matrix notation MT v u we have the implicit equations of the linear manifold L,
$
a11 v1 ` a12 v2 ` ` a1n vn u1
&
..
.
%ap1 v1 ` ap2 v2 ` ` apn vn up

where p n k dim U rank paij q i1,p .
j1,n
A hyperplane has only one equation

a1 v1 ` ` an vn u0
The director subspace can be seen as
VL tv v1 e1 ` ` vn en |f pvq 0u ker f,
where f is the linear map (linear functional) f : V R with
f pe1 q a1 , . . . , f pen q an .
If we think of the hyperplane as a linear manifold in the euclidean space Rn , the
equation can be written as
xv, ay u0 , where a a1 e1 ` ` an en , u0 P R.
The vector a is called the normal vector to the hyperplane.
Generally in a euclidean space the equations of a linear manifold are
xv, v1 y u1
&
..
.
%xv, vp y up
The Gram determinant. Distances.
113
where the vectors v1 , . . . vp are linearly independent. The director subspace is given
by
$
xv, v1 y 0
&
..
.
%xv, vp y 0
so, the vectors v1 , . . . , vp are orthogonal to the director subspace VL .
5.5
In this section we will explain how we can measure the distance between some
linear sets, which are linear manifolds.
Let pV, x, yq be an inner product space and consider the vectors vi P V , i 1, k.
The determinant

xv1 , v1 y xv1 , v2 y

xv2 , v1 y xv2 , v2 y
Gpv1 , . . . , vk q
......
...

xvk , v1 y xvk , v2 y
...
...
...
...

xv1 , vk y

xv2 , vk y

xvk , vk y
is called the Gram determinant of the vectors v1 . . . vk .

Proposition 5.33. In an inner product space the vectors v1 , . . . , vk are linearly
independent i Gpv1 , . . . , vk q 0.
Proof. Let us consider the homogenous system

x
0
1

x2 0

G
.. .. .
. .

xk
0
This system can be written as

$
xv1 , vy 0
&
..
.
%xvk , vy 0
114
where v x1 v1 ` . . . xk vk .
The following statements are equivalent.

The vectors v1 , . . . , vk are linearly dependent. There exist x1 , . . . , xk P F, not
all zero such that v 0. The homogenous system has a nontrivial solution.
det G 0.
Proposition 5.34. If te1 , . . . , en u are linearly independent vectors and tf1 , . . . , fn u
are vectors obtained by Gram Schmidt orthogonalization process, one has:
Gpe1 , . . . , en q Gpf1 , . . . , fn q }f1 }2 . . . }fn }2
Proof. In Gpf1 , . . . , fn q replace fn by en a1 f1 an1 fn1 and we obtain
Gpf1 , . . . , fn q Gpf1 , . . . , fn1 , en q.
By an inductive process the relation in the theorem follows. Obviously
Gpf1 , . . . , fn q }f1 }2 . . . }fn }2 because in the determinant we have only on the
diagonal xf1 , f1 y, . . . , xfn , fn y.
Remark 5.35. Observe that:
d
Gpe1 , . . . ek q
}fk }
Gpe1 , . . . , ek1 q
fk ek a1 f1 . . . ak1 fk1 ek vk one obtains ek fk ` vk ,
vk P spante1 , . . . , ek1 u and fk P spante1 , . . . , ek1uK , so fk is the orthogonal
complement of ek with respect to the space generated by te1 . . . , ek1 u.
115
The distance between a vector and a subspace

Let U be a subspace of the inner product space V . The distance between a vector
v and the subspace U is
dpv, Uq inf dpv, wq inf }v w}.
wPU
wPU
Remark 5.36. The linear structure implies a very simple but useful fact:
dpv, Uq dpv ` w, w ` Uq
for every v, w P V and U V , that is the linear structure implies that the distance
is invariant by translations
We are interested in the special case when U is a subspace.
Proposition 5.37. The distance between a vector v P V and a subspace U is given
by
d
dpv, Uq }v K }
Gpe1 , . . . , ek , vq
,
Gpe1 , . . . , ek q
where v v1 ` v K , v1 P U, v K P U K and e1 , . . . , ek is a basis in U.

Proof. First we prove that }v K } }v v1 } }v u},
@u P U. We have
}v K } }v u}
xv K , v K y xv K ` v1 u, v K ` v1 uy
xv K , v K y xv K , v K y ` xv1 u, v1 uy.
b
1 ,...,ek ,vq
K
, follows from the previous
The second part of the equality, i.e. }v } Gpe
Gpe1 ,...,ek q
remark.
Denition 5.38. If e1 , . . . , ek are vectors in V the volume of the k- parallelepiped
a
constructed on the vectors e1 , . . . , ek is dened by Vk pe1 , . . . , ek q Gpe1 , . . . , ek q.
116
We have the following inductive relation

Vk`1 pe1 , . . . , ek , ek`1 q Vk pe1 , . . . , ek qdpek`1 , spante1 , . . . , ek uq.
The distance between a vector and a linear manifold
Let L v0 ` VL be a linear manifold, and let v be a vector in a nitely
dimensional inner product space V . The distance induced by the norm is invariant
by translations, that is, for all v1 , v2 P V one has
dpv1 , v2 q dpv1 ` v0 , v1 ` v0 q }v1 v2 } }v1 ` v0 pv2 ` v0 q}
That means that we have
dpv, Lq inf dpv, wq inf dpv, v0 ` vL q
vL PVL
wPL
inf dpv v0 , vL q
vL PVL
dpv v0 , VL q.
Finally,
d
dpv, Lq dpv v0 , VL q
Gpe1 , . . . , ek , v v0 q
,
Gpe1 , . . . , ek q
where e1 , . . . , ek is a basis in VL .
Example 5.39. Consider the linear manifolds
L tpx, y, z, tq P R4 |x ` y ` t 2, x 2y ` z ` t 3u,
K tpx, y, z, tq P R4 |x ` y ` z t 1, x ` y ` z ` t 3u. Find the director
subspaces VL , VK and a basis in VL X VK . Find the distance of v p1, 0, 2, 2q from
L, respectively K, and show that the distance between L and K is 0.
Since L v0 ` VL and K u0 ` VK it follows that VL L v0 and VK K u0
for some v0 P L, u0 P K. By taking x y 0 in the equations that describe L we
117
obtain t 2, z 1, hence v0 p0, 0, 1, 2q P L. Analogously u0 p0, 0, 2, 1q P K.

Hence the director subspaces are
VL tpx, y, z 1, t 2q P R4 |x ` y ` t 2, x 2y ` z ` t 3u
tpx, y, z, tq P R4 |x ` y ` t 0, x 2y ` z ` t 0u,
respectively
VK tpx, y, z 2, t 1q P R4 |x ` y ` z t 1, x ` y ` z ` t 3u
tpx, y, z, tq P R4 |x ` y ` z t 0, x ` y ` z ` t 0u.
$
&
x`y`t 0
By solving the homogenous systems
, respectively
% x 2y ` z ` t 0
$
& x`y`zt0
we obtain that
% x`y`z`t0
VL spamte1 p1, 1, 3, 0q, e2 p1, 0, 0, 1qu,
respectively
VK spamte3 p1, 1, 0, 0q, e4 p1, 0, 1, 0qu.
Since detre1 |e2 |e3 |e4 s 3 0 the vectors e1 , e2 , e3 , e4 are linearly independent,
hence VL X VK t0u. The distance of v from L is
d
c
Gpe1 , e2 , v v0 q
19
,
dpv, Lq dpv v0 , VL q
Gpe1 , e2 q
21
meanwhile
d
dpv, Kq dpv v0 , VK q
Gpe3 , e4 , v v0 q
Gpe3 , e4 q
4
.
3
118
x`y`t2
& x 2y ` z ` t 3
is
It is obvious that K X L H, since the system
x`y`zt1
% x`y`z`t3
consistent, having solution p1, 0, 1, 1q, hence we must have
dpL, Kq 0.
Let us consider now the hyperplane H of equation
xv v0 , ny 0 .
The director subspace is VH xv, ny 0 and the distance
dpv, Hq dpv v0 , VH q.
One can decompose v v0 n ` vH , where vH is the orthogonal projection of
v v0 on VH and n is the normal component of v v0 with respect to VH . It
means that
dpv, Hq }n}
Let us compute a little now, taking into account the previous observations about
the tangential and normal part:
xv v0 , ny xn ` vH , ny
xn, ny ` xvH , ny
}n}2 ` 0
So, we obtained
|xv v0 , ny|
||}n} }n}
}n}
that is
dpv, Hq
119
|xv v0 , ny|
}n}
In the case that we have an orthonormal basis at hand, the equation of the
hyperplane H is
a1 x1 ` ` ak xk ` b 0 ,
so the relation is now
dpv, Hq
|a1 v1 ` ` ak vk ` b|
a
.
a21 ` ` a2k
The distance between two linear manifolds

For A and B sets in a metric space, the distance between them is dened as
dpA, Bq inftdpa, bq|a P A , b P Bu.
For two linear manifolds L1 v1 ` V1 and L2 v2 ` V2 it easily follows:
dpL1 , L2 q dpv1 ` V1 , v2 ` V2 q dpv1 v2 , V1 V2 q
dpv1 v2 , V1 ` V2 q.
(5.1)
(5.2)
This gives us the next proposition.

Proposition 5.40. The distance between the linear manifolds L1 v1 ` V1 and
L2 v2 ` V2 is equal to the distance between the vector v1 v2 and the sum space
V1 ` V2 .
If we choose a basis in V1 ` V2 , lets say e1 , . . . , ek , then this formula follows:
d
Gpe1 , . . . , ek , v1 v2 q
dpL1 , L2 q
.
Gpe1 , . . . , ek q
Some analytic geometry
120
In this section we are going to apply distance problems in euclidean spaces.

Consider the vector space Rn with the canonical inner product, that is: for
x px1 , . . . , xn q, y py1 , . . . , yn q P Rn the inner product is given by
xx, yy
xk yk .
i1
Consider D1 , D2 two lines (one dimensional linear manifolds), M a point (zero

dimensional linear manifold, we assimilate with the vector xM 0M ), P a two
dimensional linear manifold (a plane), and H an n 1 dimensional linear manifold
(hyperplane). The equations of these linear manifolds are:
D1 : x x1 ` sd1 ,
D2 : x x2 ` td2 ,
M : x xM ,
P : x xP ` v1 ` v2 ,
respectively
H : xx, ny ` b 0,
where s, t, , , b P R. Recall that two linear manifolds are parallel if the director
space of one of them is included in the director space of the other.
Now we can write down several formulas for distances between linear manifolds.
GpxM x1 , d1 q
;
Gpd1 q
GpxM xP , v1 , v2 q
;
Gpv 1 , v 2 q
Gpx1 x2 , d1 , d2 q
if D1 D2
Gpd1 , d2 q
dpM, D1 q
dpM, P q
dpD1 , D2 q
121
Gpx1 x2 , d1 q
if D1 D2
Gpd1 , q
|xxM , ny ` b|
dpM, Hq
}n}
d
Gpx1 xP , d1 , v1 , v2 q
dpD1 , P q
if D1 P
Gpd1 , v 1 , v 2 q
dpD1 , D2 q
Example 5.41. Find the distance between the hyperplane

H tpx, y, z, tq P R4 : x ` y ` z ` t 1u and the line
D tpx, y, z, tq P R4 : x ` y ` z ` t 3, x y 3z t 1, 2x 2y ` 3z ` t 1u.
Since v0 p0, 0, 0, 1q P H its director subspace is VH tpx, y, z, tq P R4 :
x ` y ` z ` t 0u spamte1 p1, 0, 0, 1q, e2 p0, 1, 0, 1q, e3 p0, 0, 1, 1qu.
Since u0 p1, 1, 0, 1q P D its director subspace is VD tpx, y, z, tq P R4 :
x ` y ` z ` t 0, x y 3z t 0, 2x 2y ` 3z ` t 0u spamte4 p1, 1, 1, 3qu.
We have e4 e1 ` e2 ` e3 hence VD VH that is D and H are parallel. Obviously
one can compute their distance by the formula
d
Gpe1 , e2 , e3 , v0 u0 q
.
dpD, Hq
Gpe1 , e2 , e3 q
But, observe that the distance between these manifolds is actually the distance
between a point M P D and H, hence is more simple to compute from the formula
Problems
dpM, Hq
122
|xxM ,ny`b|
,
}n}
with xM u0 . Indeed the equation of H is x ` y ` z ` t 1,
thus n p1, 1, 1, 1q and b 1, hence

2
|xp1, 1, 0, 1q, p1, 1, 1, 1qy 1|
1.
}p1, 1, 1, 1q}
2
dpD, Hq
5.6
Problems
Problem 5.6.1. Prove that for the nonzero vectors x, y P R2 , it holds

xx, yy }x}}y} cos ,
where is the angle between x and y.
Problem 5.6.2. Find the angle between the vectors p2, 4, 3q and p1, 2, 3q.
Problem 5.6.3. Find the two unit vectors which are orthogonal to both of the
vectors p2, 3, 1q and p1, 1, 1q.
Problem 5.6.4. Let u, v P V , V inner product space. Show that
}u} }u ` av}, @a P F xu, vy 0.
Problem 5.6.5. Prove that
p
i1
ai bi q p
2
i1
ia2i qp
1
i1
b2i q,
for all ai , bi P R , , i 1, n.
Problem 5.6.6. Let S be the subspace of the inner product space R3 rXs, the
space of polynomials of degree at most 3, generated by the polynomials 1 x2 and
1
2 x ` x2 , where xf, gy 0 f pxqgpxqdx. Find a basis for the orthogonal
complement of S.
Problems
123
Problem 5.6.7. Let u, v P V , V inner product space. If

}u} 3 , }u ` v} 4 , }u v} 6,
nd }v}.
Problem 5.6.8. Prove or inrm the following statement: There exists an inner
product on R2 such that the norm induced by this scalar product satises
}px1 , x2 } |x1 | ` |x2 |,
for all px1 , x2 q P R2 .
Problem 5.6.9. Show that the planes P : x 3y ` 4z 12 and
P2 : 2x 6y ` 8z 6 are parallel and then nd the distance between them.
Problem 5.6.10. Let V be an inner product space. Then it holds:
xu, vy
}u ` v}2 }u v}2
, @u, v P V.
4
Problem 5.6.11. If V is a complex vector space with an inner product on it,

show that
xu, vy
}u ` v}2 }u v}2 ` i}u ` iv}2 i}u iv}2

, @ u, v P V.
4
Problem 5.6.12. Prove that the following set

*
"
sin nx cos x
cos nx
1 sin x
? , ? ,..., ? , ? ,..., ?
2
is orthonormal in Cr, s, endowed with the scalar product
xf, gy
f pxqgpxqdx.
Problems
124
Problem 5.6.13. Show that the set of all vectors in Rn which are orthogonal to a
given vector v P Rn is a subspace of Rn . What will its dimension be?
Problem 5.6.14. If S is a subspace of a nite dimensional real inner product
space V, prove that S K V {S.
Problem 5.6.15. Let V be an inner product space and let tv1 , . . . , vm u a list of
linearly independent vectors from V. How many orthonormal families te1 , . . . , em uq
can be constructed by using the Gram-Schmidt procedure, such that
spantv1 , . . . , vi u spante1 , . . . , ei u, @ i 1, m.
Problem 5.6.16. Orthonormalize the following list of vectors in R4
tp1, 11, 0, 1q, p1, 2, 1, 1q, p1, 1, 1, 0q, p1, 1, 1, 1qu.
Problem 5.6.17. Let V be an inner product space and let U V subspace. Show
that
dim U K dim V dim U.
Problem 5.6.18. Let te1 , . . . , em u be an orthonormal list in the inner product
space V . Show that
}v}2 |xv, e1 y|2 ` ` |xv, em y|2
if and only if v P spante1 , . . . , em u.
Problem 5.6.19. Let V be a nite-dimensional real inner product space with a
basis te1 , . . . , en u. Show that for any u, w P V it holds
xu, wy rusJ Gpe1 , . . . , en qrws where rus is the coordinate vector (represented as a
column matrix) of u with respect to the given basis and Gpe1 , . . . , en q is the matrix
having the same entries as the Gram determinant of te1 , . . . , en u..
Problem 5.6.20. Find the distance between the following linear manifolds.
Problems
a) L tpx, y, z, tq P R4 |x ` y ` t 1, x 2y ` z 1u, K tpx, y, z, tq P

R4 |y ` 2z t 1, x ` y ` z ` t 2, x y 2z 4u.
b) L tpx, y, z, tq P R4 |x ` y ` t 2, x 2y ` z 3u, K tpx, y, z, tq P
R4 |y ` z t 1, 2x y ` z ` t 3u.
c) L tpx, y, z, tq P R4 |x ` z ` t 1, x ` y ` z 2u, K tpx, y, z, tq P
R4 |y ` t 3, x ` t 4u.
d) L tpx, y, z, tq P R4 |x ` z ` t 1, x ` y ` z 2, x y ` t 2u, K
tpx, y, z, tq P R4 |2x ` y ` 2z ` t 4u.
Problem 5.6.21. Let V be an inner product space, let U V be an arbitrary
subset and let U1 , U2 V be subspaces. Show that U K is a subspace of V , and
V K 0,respectively 0K V . Further the following implication holds:
U1 U2 U1K U2K .
125
6
Operators on inner product spaces.
6.1
Linear functionals and adjoints
A linear functional on a vector space V over the eld F is a linear map f : V F.

Example 6.1. f : F3 F given by f pv1 , v2 , v3 q 3v1 ` 4v2 5v3 is a linear
functional on F3 .
Assume now that V is an inner product space. For xed v P V , the map f : V F
given by f puq xu, vy is a linear functional. The next fundamental theorem shows
that in case V is a Hilbert space, then every linear continuous functional on V is of
this form. Recall that an inner product space is a Hilbert space if is complete, that
is, every Cauchy sequence is convergent. In other words, if the sequence pxn q V
satises, the condition:
@ 0 Dn P N s.t. @n, m n }xn xm }V
then pxn q is convergent.
Theorem 6.2. Suppose f is a linear continuous functional on the Hilbert space V .
Then there is a unique vector v P V such that
126
127
f puq xu, vy .
Proof. We will present the proof only in the nite dimensional case. We show rst
that there is a vector v P V such that f puq xu, vy. Let te1 , . . . , en u be an
orthonormal basis of V . One has
f puq f pxu, e1ye1 ` ` xu, en yen q xu, e1 yf pe1 q ` . . . xu, en yf pen q

xu, f pe1 qe1 ` ` f pen qen y ,
for every u P V . It follows that the vector
v f pe1 qe1 ` ` f pen qen
satises f puq xu, vy for every u P V .
It remains to prove the unicity of v. Suppose that there are v1 , v2 P V such that
f puq xu, v1 y xu, v2y,
for every u P V . It follows that
0 xu, v1y xu, v2 y xu, v1 v2 y @ u P V
Taking u v1 v2 it follows that v1 v2 , so v is unique.
Remark 6.3. Note that every linear functional on a nite dimensional Hilbert
space H is continuous. Even more, on every nite dimensional inner product space
H, the inner product denes a norm (metric) such that with the topology induced
by this metric H becomes a Hilbert space.
128
Let us consider another vector space W over F, and an inner product on it, such
that pW, x, yq becomes a Hilbert space.
Let T P LpV, W q a continuous operator in the topologies induced by the norms
a
a
}v}V xv, vyV , respectively }w}W xw, wyW , (as a continuous function in
analysis). We dene now the adjoint of T , as follows.
Fix w P W . Consider the linear functional on V which maps v in xT pvq, wyW . It
follows that there exists a unique vector T pwq P V such that
xv, T pwqyV xT pvq, wyW
@v P V.
The operator T : W V constructed above is called the adjoint of T .

Example 6.4. Let T : R3 R2 given by T px, y, zq py ` 3z, 2xq.
Its adjoint operator is T : R2 R3 . Fix pu, vq P R2 . It follows
xpx, y, zq, T pu, vqy xT px, y, zq, pu, vqy

xpy ` 3z, 2xq, pu, vqy
yu ` 3zu ` 2xv
xpx, y, zq, p2v, u, 3uqy
forall px, y, zq P R3 . This shows that
T pu, vq p2v, u, 3uq.
Note that in the example above T is not only a map from R2 to R3 , but also a
linear map.
We shall prove this in general. Let T P LpV, W q, so we want to prove that
T P LpW, V q.
Let w1 , w1 P W . One has, by denition:
129
xT pvq, w1 ` w2 y xT pvq, w1y ` xT v, w2y

xv, T pw1 qy ` xv, T pw2 qy
xv, T pw1 q ` T pw2 qy,
which shows that T pw1 q ` T pw2 q plays the role of T pw1 ` w2 q. By the
uniqueness proved before, we have that
T pw1 q ` T pw2 q T pw1 ` w2 q .
Remains to check the homogeneity of T . For a P F one has
xT pvq, awy axT pvq, wy

axv, T pwqy
xv, aT pwqy .
This shows that aT pwq plays the role of T pawq, and again by the uniqueness we
have that
aT pwq T pawq .
Thus T is a linear map, as claimed.
One can easily verify that we have the following properties:
a) additivity pS ` T q S ` T for all S, T P LpV, W q.
b) conjugate homogeneity paT q aT for all a P F and T P LpV, W q.
c) adjoint of adjoint pT q T for all T P LpV, W q.
d) identity I I, if I IV , V W .
130
e) products pST q T S for all T P LpV, W q and S P LpW, Uq.

For the sake of completeness we prove the above statements. Let v P V and w P W.
aq Let S, T P LpU, W q. Then, xpS ` T qpvq, wy xv, pS ` T q pwqy. On the other
hand
xpS`T qpvq, wy xSpvq, wy`xT pvq, wy xv, S pwqy`xv, T pwqy xv, pS `T qpwqy.
Hence, pS ` T q S ` T .
bq Let a P F and T P LpU, W q. We have xpaT qpvq, wy xv, paT q pwqy. But
xpaT qpvq, wy axT pvq, wy axv, T pwq xv, aT pwqy.
Hence, paT q aT pwq.
cq Let T P LpU, W q. Then
xw, T pvqy xT pvq, wy xv, T pwqy xT pwq, vy xw, pT q pvqy.
Hence, pT q T.
dq Let V W. We have xv, Ipwqy xv, wy xIpvq, wy xv, I pwqy.
Hence I I .
eq Let T P LpV, W q and S P LpW, Uq. Then for all u P U and v P V it holds:
xT S puq, vy xS puq, ppT qq pvq xS puq, T pvqy xu, pS q T pvqy xu, ST pvqy
xST pvq, uy xv, pST qpuqy xpST qpuq, vy.
Hence, T S pST q.
Proposition 6.5. Suppose that T P LpV, W q is continuous. Then
1. ker T pim T qK .
2. im T pker T qK .
3. ker T pim T qK .
4. im T pker T qK .
131
Proof. 1. Let w P W . Then
w P ker T T pwq 0
xv, T pwqy 0
xT pvq, wy 0
@vPV
@v P V
w P pim T qK ,
that is ker T pim T qK . If we take the orthogonal complement in both sides we

get 4. Replacing T by T in 1 and 4 gives 3 and 2.
The conjugate transpose of a type pm, nq- matrix is an pn, mq matrix obtained by
interchanging the rows and columns and taking the complex conjugate of each
entry. The adjoint of a matrix (which is a linear transform between two nite
dimensional spaces in the appropriate bases) is the conjugate transpose of that
matrix as the next result shows.
Proposition 6.6. Suppose that T P LpV, W q. If te1 , . . . , en u, and tf1 , . . . , fm u are
orthonormal bases for V and W respectively, and we denote by MT and MT the
matrices of T and T in these bases, then MT is the conjugate transpose of MT .
Proof. The k th column of MT is obtained by writing T pek q as linear combination
of f s, the scalars used became the k th column of M . Being the basis with f s
j
orthonormal, it follows that

T pek q xT pek q, f1 yf1 ` ` xT pek q, fm yfm .
So on the position pk, jq of MT we have xT pek q, fj y. Replacing T with T and
interchanging the roles played by es andf s, we see that the entry on the position
Normal operators
132
pj, kq of MT the entry is xT pfk q, ej y, which equals to xfk , T pej qy, which equals to
xT pej q, fk y. In others words, MT equals to the complex conjugate of MT .
6.2
Normal operators
An operator on a Hilbert space is called normal if it commutes with its adjoint,

that is
T T T T .
Remark 6.7. We will call a complex square matrix normal if commute with its
conjugate transpose, that is A P Mn pCq is normal i
AA A A,
J
where A is the conjugate transpose of A, that is A A . It can easily be

observed that the matrix of a normal operator is a normal matrix.
Example 6.8. On F2 consider the operator which in the canonical basis has the
matrix
2 3
3
This is a normal operator.

Indeed let T : F2 F2 be the operator whose matrix is A. Then
T px, yq p2x 3y, 3x ` 2yq, thus xT px, yq, pu, vqy p2x 3yqu ` p3x ` 2yqv
p2v 3uqy ` p3v ` 2uqx xpx, yq, p3v ` 2u, 2v 3uqy. Hence, the adjoint of T is
T pu, vq p2u ` 3v, 3u ` 2vq. It can easily be computed that
T T pu, vq T T pu, vq p13u, 13vq, hence T T T T.
Proposition 6.9. An operator T P LpV q is normal operator i
}T pvq} }T pvq} for all v P V.
Normal operators
133
Proof. Let T P LpV q.

T is normal
T T T T 0
xpT T T T qpvq, vy 0 for all v P V
xT T pvq, vy xT T pvq, vy for all v P V
}T pvq}2 }T pvq}2 for all v P V.
Theorem 6.10. Let T be a normal operator on V and 0 be an eigenvalue of T .

1. The proper subspace Ep0 q is T invariant.
2. If v0 is an eigenvector of T corresponding to the eigenvalue 0 , then v0 is an
eigenvector of T corresponding to the eigenvalue 0 .
3. Let v, w be two eigenvectors corresponding to distinct eigenvalues , . Then
v, w are orthogonal.
Proof. Let v P Ep0 q. We have to prove that T pvq P Ep0 q.
Since T pvq 0 v, we have
T pT pvqq pT T qpvq pT T qpvq T pT pvqq T p0 vq 0 T pvq.
which show that T pvq P Ep0 q.
For the second statement in the theorem we have that T pv0 q 0 v0 . Let
w P Ep0 q. Then
xT pv0 q, wy xv0 , T pwqy
xv0 , 0 wy 0 xv0 , wy x0 v0 , wy.
That means that
xT pv0 q 0 v0 , wy 0 ,
Normal operators
134
for all w P Ep0 q. The rst term in the inner product lives in Ep0 q by the
previous statement. Take w T pv0 q 0 v0 and it follows that T pv0 q 0 v0 , i.e.
the second assertion in the theorem.
Now follows the last statement. One has T pvq v and T pq w. By the
previous point T pwq w, so
xT pvq, wy xv, T pwq
(def. of adjoint), which implies xv, wy xv, wy. Since , it follows that
xv, wy 0.
Proposition 6.11. If U is a T invariant subspace of V then U K is a T invariant
subspace of V .
Proof.
w P U K , v P V w P U K , T pvq P U xv, T pwqy xT pvq, wy.
That is T pwq P U K .
A unitary space is an inner product space over C.

Theorem 6.12. Suppose that V is a nite dimensional unitary space, and
T P LpV q is an operator. Then T is normal i there exists an orthonormal basis B
of V relative to which the matrix of T is diagonal.
Proof. First suppose that T has a diagonal matrix. The matrix of T is the
complex transpose, so it is again diagonal. Any two diagonal matrices commutes,
that means that T is normal.
Normal operators
135
To prove the other direction suppose that T is normal. Then, there is a basis
te1 , . . . , en u of V with respect to which the
a
a
1,1 1,2
0 a2,2
A
..
..
.
.
0
0
matrix of T is upper triangular, that is
. . . an,n
. . . a2,n
.. .
.
. . . an,n
We will show that the matrix A is actually a diagonal one.

We have
}T pe1 q}
|a1,1 |2
and
}T pe1 q}
|a1,1 |2 ` ` |a1,n |2 .
Because T is normal the norms are equal, so a1,2 a1,n 0.

b
b
2
2
}T pe2 q} |a1,2 | ` |a2,2 | |a2,2 |2
and
}T pe2 q}
|a2,2 |2 ` ` |a2,n |2 .
Because T is normal the norms are equal, so a2,3 a2,n 0.

By continuing the procedure we obtain that for every k P t1, . . . , n 1u we have
ak,k`1 ak,n 0, hence A is diagonal.
Theorem 6.13. (Complex spectral theorem) Suppose that V is a unitary

space. Then T has an orthonormal basis consisting of eigenvectors i T is normal.
Proof. Induction on n dim V . The statement is obvious for n 1. Suppose that
this is true for all dimensions less than n. Let T P LpV q. Then T has at least one
Isometries
136
eigenvalue . If dim Epq n it is enough to construct an orthonormal basis of

Epq. For dim Epq n, choose E K pq, and we have 0 dim E K pq n.
Now Epq is T invariant, so E K pq is T invariant. By the induction hypothesis,
E K pq has an orthonormal basis consisting of eigenvectors of T . Add this to the
orthonormal basis of Epq. The result is an orthonormal basis of V consisting of
eigenvectors.
6.3
Isometries
An operator T P LpV q is called an isometry if

}T pvq} }v} , for all v P V.
Example 6.14. Let I be the identity map of V (V complex v.s.), and P C with
|| 1. The map I is an isometry, since }Ipvq} }v} ||}v} }v}.
If T is an isometry it easily follows that T is injective.
Indeed, assume the contrary, that is, there exists u, v P V, u v such that
T puq T pvq. Hence, 0 }T puq T pvq} }T pu vq} }u v}, contradiction with
u v.
Theorem 6.15. Suppose T P LpV q. The following are equivalent:
1. T is an isometry.
2. xT puq, T pvqy xu, vy for every u, v P V .
3. T T I.
4. tT pe1 q, . . . , T pem qu is an orthonormal list for every orthonormal list
te1 , . . . , em u.
Isometries
137
5. There exists an orthonormal basis te1 , . . . , en u of V such that

pT te1 q, . . . , T pen qu is an orthonormal basis.
6. T is an isometry.
7. xT puq, T pvqy xu, vy for all u, v P V .
8. T T I
9. tT pe1 q, . . . , T pem qu is an orthonormal list for every orthonormal list
pe1 , . . . , em q .
10. There exists an orthonormal basis te1 , . . . , en u of V such that
tT pe1 q, . . . , T pen qu is an orthonormal basis.
Proof. Suppose that 1 holds. Let u, v P V. Then
}u v}2 }T pu vq}2 xT puq T pvq, T puq T pvqy
}T puq}2 ` }T pvq}2 2xT puq, T pvqy }u}2 ` }v}2 2xT puq, T pvqy. On the other
hand }u v}2 }u}2 ` }v}2 2xu, vy.
Suppose now that 2 holds. Then
xpT T Iqpuq, vy xT puq, T pvqy xu, vy 0 .
for every u, v P V . Take v pT T Iqpuq and it follows that T T I 0, i.e. 3.
Suppose 3 holds. Let pe1 . . . em q be an orthonormal list of vectors in V . Then
xT pej q, T pek qy xT T pej q, ek y xej , ek y,
i.e. 4 holds. Obviously 4 implies 5.
Isometries
138
Suppose 5 holds. Let pe1 , . . . en q be an orthonormal basis of V such that

pT pe1 q, . . . , T pen qq is orthonormal basis. For v P V
}T pvq}2 }T pxv, e1ye1 ` ` xv, en yen q}2
}xv, e1yT pe1 q ` ` xv, en yT pen q}2
|xv, e1 y|2 ` ` |xv, en y|2
}v}2 .
Taking square roots we see that T is an isometry. We have now
1 2 3 4 5 1. Replacing T by T we see that 6 through 10 are
equivalent. We need only to prove the equivalence of one assertion in the rst
group with one in the second group.
3 8 which is easy to see since T T I
T T puq u, @u P V pT T qpT puqq T puq, @u P V, or equivalently
T ppT T qpuqq T puq, @u P V , T is injective, hence T T I.
Conversely, T T I T T puq u, @u P V pT T qpT puqq T puq, @u P V, or
equivalently T ppT T qpuqq T puq, @u P V , T is injective, hence T T I.
Remark 6.16. Recall that a real square matrix A is called orthogonal i
AAJ AJ A I. A complex square matrix B is called unitary if BB B B I,
J
where B is the conjugate transpose of B, that is B B . It can easily be

observed that the matrix of an isometry on a real (complex) nite dimensional
inner product space is an orthogonal (unitary) matrix.
The last theorem shows that every isometry is a normal operator. So, the
characterizations of normal operators can be used to give a complete description of
isometries.
Theorem 6.17. Suppose that V is a complex inner product space and T P LpV q.
Isometries
139
Then T is an isometry i there is an orthonormal basis of T consisting of

eigenvectors of T whose corresponding eigenvalues have modulus 1.
Proof. Suppose that there is an othonormal basis te1 , . . . , en u consisting of
eigenvectors whose corresponding eigenvalues t1 , . . . , n u have absolute value 1. It
follows that for every v P V
T pvq xv, e1 yT pe1 q ` ` xv, en yT pen q
1 xv, e1 ye1 ` ` n xv, en yen .
Thus }T pvq}2 |xv, e1 y|2 ` ` |xv, en y|2 }v}2 that is
}T pvq} }v}.
Now we are going to prove the other direction. Suppose T is an isometry. By the
complex spectral theorem there is an orthonormal basis of V consisting of
eigenvectors te1 , . . . , en u.. Let ej , j P t1, . . . , nu be such a vector, associated to an
eigenvalue j . It follows that
|j |}ej } }j ej } }T pej q} }ej },
hence |j | 1, for all j P t1, . . . , nu.
Finally we state the following important theorem concerning on the form of the
matrix of an isometry.
Theorem 6.18. Suppose that V is a real inner product space and T P LpV q. Then
T is an isometry i there exist an orthonormal basis of V with respect to which T
has a block diagonal matrix where each block on the diagonal matrix is a p1, 1q
matrix containing 1 or 1, or a p2, 2q matrix of the form
cos sin
sin cos
Self adjoint operators
140
with P p0, q.
Proof. The eigenvalues of T have modulus 1, hence are the form 1, 1 or
cos sin . On the other hand, the matrix of T is similar to a diagonal matrix
whose diagonal entries are the eigenvalues.
6.4
An operator T P LpV q is called self-adjoint if T T that is xT pvq, wy xv, T pwqy

for all v, w P V .
Remark 6.19. Obviously a self adjoint operator T P LpV q is normal since in this
case holds
T T T T T T.
Example 6.20. Let T be an operator on F2 whose matrix with respect to the
standard basis is
2 b
3 5
Then T is self-adjoint i b 3.
Indeed, for px, yq P F2 one has T px, yq p2x ` by, 3x ` 5yq, hence for pu, vq P F2 it
holds
xT px, yq, pu, vqy p2x ` byqu ` p3x ` 5yqv xpx, yq, p2u ` 3v, bu ` 5vqy.
Thus T px, yq p2x ` 3y, bx ` 5yq.
In conclusion T is self adjoint, i.e. T T iif b 3.
It can easily be veried that the sum of two self adjoint operators and the product
of an self adjoint operator by a real scalar is an self-adjoint operator.
141
Indeed, let S, T P LpV q be two self adjoint operators. Then

pS ` T q S ` T S ` T, hence S ` T is self adjoint. On the other hand for
their product we have pST q T S T S. Hence T S is self adjoint i ST T S.
Let now a P R. Then paT q aT aT, hence aT is self adjoint.
Remark 6.21. When F C the adjoint on LpV q plays a similar role to complex
conjugation on C. A complex number is real i z z. Thus for a self adjoint
operator T the sum T ` T is analogous to a real number. The analogy is reected
in some important properties of a self-adjoint operator, beginning with its
eigenvalues.
Remark 6.22. Recall that a complex square matrix A is called hermitian i
J
A A , where A is the conjugate transpose of A, that is A A . If A is a

square matrix with real entries, then A is called symmetric i A AJ . It can
easily be observed that matrix of a self adjoint operator on a complex (real) inner
product space is hermitian (symmetric).
Proposition 6.23. The following statements hold.
Every eigenvalue of a self-adjoint operator is real.
Let v, w eigenvectors corresponding to distinct eigenvalues. Then xv, wy 0.
Proof. Suppose that T is a self-adjoint operator on V . Let be an eigenvalue of T ,
and v be an eigenvector, that is T pvq v. Then
}v}2 xv, vy
xT pvq, vy
xv, T pvqy (because T is self-adjoint)
xv, vy
}v}2 .
Problems
142
Thus , i.e. is real.

The next assertion cames from the fact that a self-adjoint operator is normal.
Theorem 6.24. Let T P LpV q, where V is an inner product space. The following
statements are equivalent.
1. T is self-adjoint.
2. There exists an orthonormal basis of V relative to which the matrix of T is
diagonal with real entries.
Proof. Assume that T is self adjoint. Since T is normal there exists exists an
orthonormal basis of V relative to which the matrix of T MT is upper triangular.
But the matrix of T in this basis is MT MT , and from T T one has
MT MT , hence MT is diagonal, and also the diagonal are formed byy real entries.
Conversely, let MT a diagonal matrix of T , with real entries in some orthonormal
basis. Then MT MTJ , hence MT MT or equivalently T T .
6.5
Problems
Problem 6.5.1. Suppose that A is a complex matrix with real eigenvalues which
can be diagonalized by a unitary matrix. Prove that A must be hermitian.
Problem 6.5.2. Prove or give a counter example: the product of any two self
adjoint operators on a nite dimensional inner product space is self adjoint.
Problem 6.5.3. Show that an upper triangular matrix is normal if and only if it
is diagonal.
Problem 6.5.4. Suppose p P LpV q is such that p2 p. Prove that p is an
orthogonal projection if and only if p is self adjoint.
Problems
143
Problem 6.5.5. Show that if V is a real inner product space, then the set of self
adjoint operators on V is a subspace of LpV q. Show that if V is a complex inner
product space, then the set of self-adjoint operators on V is not a subspace of
LpV q.
Problem 6.5.6. Show that if dim V 2 then the set of normal operators on V is
not a subspace of LpV q.
Problem 6.5.7. Let A be a normal matrix. Prove that A is unitary if and only if
all its eigenvalues satisfy || 1.
Problem 6.5.8. Let v be any unit vector in Cn and put A In 2XX . Prove
that A is both hermitian and unitary. Deduce that A A1 .
Problem 6.5.9. Suppose V is a complex inner product space and T P LpV q is a
normal operator such that T 9 T 8 . Prove that T is self adjoint and T 2 T.
Problem 6.5.10. Let A be a normal matrix. Show that A is hermitian if and
only if all its eigenvalues are real.
Problem 6.5.11. Prove that if T P LpV q is normal, then
im T im T .
and
ker T k ker T
im T k im T
for every positive integer k.
Problem 6.5.12. A complex matrix A is called skew-hermitian if A A.
Prove the following statements.
Problems
144
a) A skew-hermitian matrix is normal.

b) The eigenvalues of a skew-hermitian matrix are purely imaginary, that is,
have the real part 0.
c) A normal matrix is skew-hermitian if all its eigenvalues are purely imaginary.
Problem 6.5.13. Suppose V is a complex inner product space. An operator
?
S P LpV q is called a square root of T P LpV q if S 2 T. We denote S T . Prove
that every normal operator on V has a square root.
Problem 6.5.14. Prove or disprove: e identity operator on F2 has innitely many
self adjoint square roots.
Problem 6.5.15. Let T, S P LpV q be isometries and R P LpV q a positive operator,
?
(that is xRpvq, vy 0 for all v P V ), such that T SR. Prove that R T T .
Problem 6.5.16. Let R2 rXs be the inner product space of polynomials with
degree at most 2, with the scalar product
1
pptqqptqdt.
xp, qy
0
Let T P LpR2 rXsq, T pax2 ` bx ` cq bx.

a) Show that the matrix of T with respect to the given basis is hermitian.
b) Show that T is not self-adjoint.
(Note that there is no contradiction between these statements because the basis in
the rst statement is not orthonormal.)
Problem 6.5.17. Prove that a normal operator on a complex inner-product space
is self-adjoint if and only if all its eigenvalues are real.
7
Elements of geometry
following joint notes with I. Rasa and D. Inoan
7.1
Quadratic forms
Consider the n-dimensional space Rn and denote by x px1 , . . . , xn q the

coordinates of a vector x P Rn with respect to the canonical basis E te1 , . . . , en u .
A quadratic form is a map Q : Rn R
Qpxq a11 x21 ` . . . ann x2n ` 2a12 x1 x2 ` ` 2aij xi xj ` . . . 2an1,n xn1 xn ,
where the coecients aij are all real.
Thus, quadratic forms are homogenous polynomials of degree two in a number of
variables.
Using matrix multiplication, we can write Q in a compact form as
Qpxq X t AX,
where
145
Quadratic forms
146
x1
x2
..
.
and
xn
a11 a12 . . .
a1n
a12 a22 . . .
..
..
.
.
a2n
..
.
a1n a2n . . . ann
The symmetric matrix A (notice that aij aji ) is be called the matrix of the
quadratic form. Being symmetric (and real), A it is the matrix of a self-adjoint
operator with respect to the basis E. This operator, that we call T , is
diagonalizable and there exists a basis B tb1 , . . . , bn u formed by eigenvectors
with respect to which T has a diagonal matrix consisting of eigenvalues (also
denoted by T )
T diagt1 . . . . , n u.
Let C be the transition matrix from E to B and
x11
1
x2
X1
..
.
1
xn
the coordinates of the initial vector written in B. We have that
X CX 1
Knowing that T C 1 AC, and that C 1 C J we can compute that
Qpxq X J AX
J
pCX 1 q A pCX 1 q
X 1J C J ACX 1
X 1J T X 1
1 x1 1 ` ` n x1 n ,
2
Quadrics
147
and we say that we have reduced Q to its canonical form

Qpxq 1 x1 1 ` ` n x1 n .
2
This is called the geometric method.

The quadratic form is called
positive denite if Qpxq 0 for every x P Rn zt0u
negative denite if Qpxq 0 for every x P Rn zt0u.
We can characterize the positive deniteness of a quadratic form in terms of the
diagonal minors of its matrix
a11 a12
D1 a11 , D2
a12 a22
, . . . , Dn det A.
We have the following criteria:

Q is positive denite i Di 0 for every i 1, n
Q is negative denite i p1qi Di 0 for every i 1, n.
7.2
Quadrics
The general equation of a quadric is

a11 x2 ` a22 y 2 ` a33 z 2 ` 2a12 xy ` 2a13 xz ` 2a23 yz
2a14 x ` 2a24 y ` 2a34 z ` a44 0.
From a geometric point of view, quadrics, which are also called quadric surfaces,
are two-dimensional surfaces dened as the locus of zeros of a second degree
Quadrics
148
polynomial in x, y and z. Maybe the most prominent example of a quadric is the

sphere (the spherical surface).
The type is determined by the quadratic form that contains all terms of degree two
Q a11 x2 ` a22 y 2 ` a33 z 2 ` 2a12 xy ` 2a13 xz ` 2a23 yz.
We distinguish, based on the sign of the eigenvalues of the matrix of Q, between:
ellipsoids, elliptic or hyperbolic paraboloids, hyperboloids with one or two sheets,
cones and cylinders.
We study how to reduce the general equations of a quadric to a canonical form.
We reduce Q to a canonical form using the geometric method.
Consider the matrix A associated to Q. Being symmetric, A has real eigenvalues
1 , 2 , 3 . If they are distinct, the corresponding eigenvectors are orthogonal (if
not we apply the Gram-Schmidt algorithm). Thus, we obtain three orthogonal
unit vectors tb1 , b2 , b3 u, a basis in R3 .
Let R be the transition matrix from ti, j, ku to the new basis tb1 , b2 , b3 u. We recall
from previous chapters that R has the three vectors b1 , b2 , b3 as its columns
R rb1 |b2 |b3 s .
Now, we compute det R and check whether
det R 1.
If necessary, i.e., if .det R 1, we must change one of the vectors by its opposite
(for example take R rb1 |b2 |b3 s). This assures that the matrix R denes a
rotation, the new basis being obtained from the original one by this rotation. Let
px, y, zq and px1 , y 1, z 1 q be the coordinates of the same point in the original basis
Quadrics
149
and in the new one, we have
x
x

1
y R y .
1
z
z
We know that with respect to the new coordinates
Q 1 x1 ` 1 y 1 ` n z 1 ,
2
and thus, the equation of the quadric reduces to the simpler form
1 x1 ` 1 y 1 ` n z 1 ` 2a1 14 x1 ` 2a1 24 y 1 ` 2a1 34 z 1 ` a44 0.
2
To obtain the canonical form of the quadric we still have to perform another
transformation, namely a translation. To complete this step we investigate three
cases: (A) when A has three nonzero eigenvalues, (B) when one eigenvalue is zero
and (C) when two eigenvalues are equal to zero.
(A) For i 0 we obtain
1 px1 x0 q2 ` 2 py 1 y0 q2 ` 3 pz 1 z0 q2 ` a1 44 0
Consider the translation dened by
x2 x1 x0 ,
y 2 y 1 y0 ,
z 2 z 1 z0 .
In the new coordinates the equation of the quadric reduces to the canonical form
1 x22 ` 2 y 22 ` 3 z 22 ` a1 44 0.
The cases (B) and (C) can be treated similarly.
Conics
7.3
150
Conics
Studied since the time of ancient greek geometers, conic sections (or just conics)
are obtained, as their name shows, by intersecting a cone with a sectioning plane.
They have played a crucial role in the development of modern science, especially in
astronomy, the motion of earth round the sun taking place on a particular conic
called ellipse. Also, we point out the fact that the circle is a conic section, a special
case of ellipse.
The general equation of a conic is
a11 x2 ` 2a12 xy ` a22 y 2 ` 2a13 x ` 2a23 y ` a33 0.
The following two determinants obtained from the coecients of the conic play a
crucial role in the classication
a11 a12
a12 a22
a13 a23
of conics
a13
and
a23
a33
a11 a12
D2
a12 a22
Notice that the second determinant corresponds to the quadratic form dened by
the rst three terms.
Conical sections can be classied as follows:
Degenerate conics, for which 0. These include:two intersecting lines (when
D2 0), two parallel lines or one line (when D2 0) and one point (when D2 0).
Nondegenerate conics, for which 0. Depending on D2 we distinguish
between the
Ellipse pD2 0q whose canonical equation is
x2 y 2
` 2 1,
a2
b
Conics
151
P arabola pD2 0q whose canonical equation is y 2 2ax 0,

Hyperbola pD2 0q whose canonical equation is
Figure 7.1: Ellipse
x2
a2
x2 y 2
2 1.
a2
b
y2
b2
The reduction of a conic section to its canonical form is very similar with the
procedure that we have presented in the last section when dealing with quadrics.
Again, we must perform a rotation and a translation. We show how the reduction
can be performed by means of an example.
Example 7.1. Find the canonical form of 5x2 ` 4xy ` 8y 2 32x 56y ` 80 0.
The matrix of the quadratic form of this conic is
5 2
2 8
and its eigenvalues are the roots of 2 13 ` 36 0. So 1 9 and 2 4, while
two normed eigenvectors are v1
rotation matrix is thus
?1
5
p1, 2q and v2
?1
5
?25
?2
5
?1
5
?1
5
p2, 1q respectively. The
Conics
152
and we can check that det R 1.

Now, the relation between the old and the new coordinates is given by
1
x
x
R
y1
y
that is
1
x ? px1 2y 1q
5
1
y ? p2x1 ` y 1q .
5
By substituting these expressions in the initial equation we get
?
?
144 5 1 8 5 1
12
12
x `
y ` 80 0.
9x ` 4y
5
5
To see the translation that we need to perform we rewrite the above equation as
follows
?
?
? 2
? 2
8
5
5
5
5
8
9 x12 2
x1 `
y1 `
` 4 y 12 ` 2
5
5
5
5
? 2
? 2
5
8 5
4
` 80 0.
9
5
5
Finally, we obtain
? 2
? 2
5
8 5
1
1
9 x
`4 y `
30 0.
5
5
Thus, the translation x2 x1
?
8 5
, y2
5
y1 `
5
5
reduces the conic to the canonical
form
3 22
2
x ` y 22 1.
10
15
Problems
7.4
153
Problems
Problem 7.4.1. Find the canonical form of the following quadric surfaces:
a) 2y 2 ` 4xy 8xz 6x ` 8y ` 8 0,
b) 3x2 ` y 2 ` z 2 2x 4z 4 0,
c) xz y,
d) x2 ` y 2 ` 5z 2 6xy ` 2xz 2yz 4x ` 8y 12z ` 14 0
Problem 7.4.2. Find the canonical form of the following conics:
a) x2 6xy ` 9y 2 ` 20 0,
b) 3y 2 4xy 2y ` 4x 3 0,
c) 5x2 ` 6xy ` 2y 2 ` 2y 1 0,
d) xy 1.
Problem 7.4.3. Given the ellipsoid
pEq :
x2 y 2
`
` z 2 1 0,
4
3
nd the value of the parameter p for which the straight line

$
& xz`p
% y z`2
is tangent to pEq.
Problem 7.4.4. Find the equation of the plane that is tangent to the sphere
x2 ` y 2 ` z 2 1
at the point M
?
1 1
2
,
,
.
2 2 2
Problems
154
Problem 7.4.5. Find the equations of the planes that contain the line
$
& x1
% y0
and are tangent to the sphere
px 1q2 ` y 2 ` z 2 1.
Problem 7.4.6. Consider the hyperboloid of one sheet
x2 y 2 z 2
`
1.
16
9
4
Determine the straight lines that belong to this surface and pass through the point
P p4, 3, 2q.
Problem 7.4.7. Find the equation of the circle whose center is located in C p1, 2q
and whose radius is R 2.
Problem 7.4.8. Determine the center and the radius of the circle of equation
x2 ` y 2 ` 2x 4y 4 0.
Problem 7.4.9. Write the equation of the circle that passes through the points
A p1, 1q , B p1, 5q , C p4, 1q.
Bibliography
[1] Sheldon Axler: Linear Algebra Done Right, Springer, 2-nd edition, 2004.
[2] Derek Robinson: A Course in Linear Algebra with Applications, World
Scientic Publishing, 2-nd Edition, 2006.
155

New Format

Enviado por

Dados do documento

Direitos autorais

Formatos disponíveis

Compartilhar este documento

Compartilhar ou incorporar documento

Opções de compartilhamento

Você considera este documento útil?

Este conteúdo é inapropriado?

Direitos autorais:

Formatos disponíveis

New Format

Enviado por

Direitos autorais:

Formatos disponíveis

Elements of Linear Algebra

Peter Ioan Radu

1.1 Basic denitions, operations and properties. . . . . . . . . . . . . . .

1.2 Determinants and systems of linear equations . . . . . . . . . . . . . 13

2.1 Denition of Vector Space and basic properties . . . . . . . . . . . . . 28

3.1 Properties of LpV, W q . . . . . . . . . . . . . . . . . . . . . . . . . . . 57

4 Proper vectors and the Jordan canonical form

4.1 Invariant subspaces. Proper vectors and values . . . . . . . . . . . . . 72

5.1 Basic denitions and results . . . . . . . . . . . . . . . . . . . . . . . 95

6.1 Linear functionals and adjoints . . . . . . . . . . . . . . . . . . . . . 126

7.1 Quadratic forms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145

Basic denitions, operations and properties.

Denition 1.1. A matrix of dimension m n with elements in a eld F, (where

as a table with m lines and n columns:

Basic denitions, operations and properties.

and is called the identity matrix (or unit matrix ) of order n.

Basic denitions, operations and properties.

and is called the null matrix of order m n.

aij 0 whenever i j. Then

is called upper triangular, respectively lower triangular matrix.

In other words, for A, B P Mm,n pFq their sum is dened as

Basic denitions, operations and properties.

Properties of Matrix Addition.

1. A ` B is again an m n matrix (closure property).

: F Mm,n pFq Mm,n pFq,

Properties of Scalar multiplication.

Basic denitions, operations and properties.

Basic denitions, operations and properties.

A paji q j1,n , where z denotes the complex conjugate of the number z P C. We

have that pA q A and AJ A whenever A contains only real entries.

A is said to be a symmetric matrix whenever A AJ (locally aij aji for all

Basic denitions, operations and properties.

A is said to be a skew-hermitian matrix whenever A A (locally

Example 1.8. The matrix A 2 0 3 is a symmetric matrix,

meanwhile the matrix B 1 0

Xi pxi1 xi2 . . . xin q.

Basic denitions, operations and properties.

is dened to be the m n matrix C pcik qi1,m with

Basic denitions, operations and properties.

ai1 ai2 . . . aip

3. rABsi ai1 B1 ` ai2 B2 ` ` ai pBp

4. rABsj A1 b1j ` A2 b2j ` ` Ap bpj

Basic denitions, operations and properties.

Properties of matrix multiplication.

i P t1, 2, . . . , nu, j P t1, 2, . . . , mu one has cij rABsji Aj Bi . Let us consider

rB A sij pB qi pA qj pBi q pAj q

rB J sik rAJ skj

Basic denitions, operations and properties.

AAJ and AJ A are symmetric matrices.

For a matrix A P Mn pFq, one can introduce its mth power by

rBAskk trace BA.

Block Matrix Multiplication.

Basic denitions, operations and properties.

Determinants and systems of linear equations

Determinants and systems of linear

detpAq called the determinant of A. In extended form we write

a11 a12 a1n

a21 a22 a2n

an1 an2 ann

Determinants and systems of linear equations

In order to dene the determinant of a square matrix, we need the following

sgn pq a1p1q a2p2q . . . anpnq .

detpAq a11 a22 a12 a21 .

Example 1.15. If A 4 5 6 then