CHAPTER

5
Statue of Italian mathematician
Leonardo of Pisa (1170–1250,
approximate dates), also known as
Fibonacci. Exercise 16 in Section 5.C
shows how linear algebra can be used
to find an explicit formula for the
Fibonacci sequence.

Eigenvalues, Eigenvectors, and

Invariant Subspaces
Linear maps from one vector space to another vector space were the objects
of study in Chapter 3. Now we begin our investigation of linear maps from
a finite-dimensional vector space to itself. Their study constitutes the most
important part of linear algebra.
Our standing assumptions are as follows:

5.1 Notation F, V

 F denotes R or C.

 V denotes a vector space over F.

LEARNING OBJECTIVES FOR THIS CHAPTER

invariant subspaces
eigenvalues, eigenvectors, and eigenspaces
each operator on a finite-dimensional complex vector space has an
eigenvalue and an upper-triangular matrix with respect to some
basis
© Springer International Publishing 2015 131
S. Axler, Linear Algebra Done Right, Undergraduate Texts in Mathematics,
DOI 10.1007/978-3-319-11080-6__5
132 CHAPTER 5 Eigenvalues, Eigenvectors, and Invariant Subspaces

5.A Invariant Subspaces

In this chapter we develop the tools that will help us understand the structure
of operators. Recall that an operator is a linear map from a vector space to
itself. Recall also that we denote the set of operators on V by L.V /; in other
words, L.V / D L.V; V /.
Let’s see how we might better understand what an operator looks like.
Suppose T 2 L.V /. If we have a direct sum decomposition

V D U1 ˚    ˚ Um ;

where each Uj is a proper subspace of V, then to understand the behavior of

T, we need only understand the behavior of each T jUj ; here T jUj denotes
the restriction of T to the smaller domain Uj . Dealing with T jUj should be
easier than dealing with T because Uj is a smaller vector space than V.
However, if we intend to apply tools useful in the study of operators (such
as taking powers), then we have a problem: T jUj may not map Uj into itself;
in other words, T jUj may not be an operator on Uj . Thus we are led to
consider only decompositions of V of the form above where T maps each Uj
into itself.
The notion of a subspace that gets mapped into itself is sufficiently impor-
tant to deserve a name.

5.2 Definition invariant subspace

Suppose T 2 L.V /. A subspace U of V is called invariant under T if
u 2 U implies T u 2 U.

In other words, U is invariant under T if T jU is an operator on U.

5.3 Example Suppose T 2 L.V /. Show that each of the following

subspaces of V is invariant under T :

(a) f0g; The most famous unsolved problem

in functional analysis is called the
(b) V; invariant subspace problem. It
deals with invariant subspaces of
(c) null T ; operators on infinite-dimensional
vector spaces.
(d) range T.
 SECTION 5.A Invariant Subspaces 133

Solution

(a) If u 2 f0g, then u D 0 and hence T u D 0 2 f0g. Thus f0g is invariant

under T.

(b) If u 2 V, then T u 2 V. Thus V is invariant under T.

(c) If u 2 null T, then T u D 0, and hence T u 2 null T. Thus null T is

invariant under T.

(d) If u 2 range T, then T u 2 range T. Thus range T is invariant under T.

Must an operator T 2 L.V / have any invariant subspaces other than f0g
and V ? Later we will see that this question has an affirmative answer if V is
finite-dimensional and dim V > 1 (for F D C) or dim V > 2 (for F D R/;
see 5.21 and 9.8.
Although null T and range T are invariant under T, they do not necessarily
provide easy answers to the question about the existence of invariant subspaces
other than f0g and V , because null T may equal f0g and range T may equal
V (this happens when T is invertible).
 
5.4 Example Suppose that T 2 L P.R/ is defined by Tp D p 0.
Then P4 .R/, which is a subspace of P.R/, is invariant under T because
if p 2 P.R/ has degree at most 4, then p 0 also has degree at most 4.

Eigenvalues and Eigenvectors

We will return later to a deeper study of invariant subspaces. Now we turn to an
investigation of the simplest possible nontrivial invariant subspaces—invariant
subspaces with dimension 1.
Take any v 2 V with v ¤ 0 and let U equal the set of all scalar multiples
of v:
U D fv W  2 Fg D span.v/:
Then U is a 1-dimensional subspace of V (and every 1-dimensional subspace
of V is of this form for an appropriate choice of v). If U is invariant under an
operator T 2 L.V /, then T v 2 U, and hence there is a scalar  2 F such that

T v D v:

Conversely, if T v D v for some  2 F, then span.v/ is a 1-dimensional

subspace of V invariant under T.
134 CHAPTER 5 Eigenvalues, Eigenvectors, and Invariant Subspaces

The equation
T v D v;
which we have just seen is intimately connected with 1-dimensional invariant
subspaces, is important enough that the vectors v and scalars  satisfying it
are given special names.

5.5 Definition eigenvalue

Suppose T 2 L.V /. A number  2 F is called an eigenvalue of T if
there exists v 2 V such that v ¤ 0 and T v D v.

The word eigenvalue is half- The comments above show that T

German, half-English. The Ger- has a 1-dimensional invariant subspace
man adjective eigen means “own” if and only if T has an eigenvalue.
in the sense of characterizing an in- In the definition above, we require
trinsic property. Some mathemati- that v ¤ 0 because every scalar  2 F
cians use the term characteristic satisfies T 0 D 0.
value instead of eigenvalue.

5.6 Equivalent conditions to be an eigenvalue

Suppose V is finite-dimensional, T 2 L.V /, and  2 F . Then the
following are equivalent:

(a)  is an eigenvalue of T ;
Recall that I 2 L.V / is the iden-
(b) T  I is not injective; tity operator defined by I v D v for
all v 2 V.
(c) T  I is not surjective;

(d) T  I is not invertible.

Proof Conditions (a) and (b) are equivalent because the equation T v D v
is equivalent to the equation .T  I /v D 0. Conditions (b), (c), and (d) are
equivalent by 3.69.

5.7 Definition eigenvector

Suppose T 2 L.V / and  2 F is an eigenvalue of T. A vector v 2 V is
called an eigenvector of T corresponding to  if v ¤ 0 and T v D v.
 SECTION 5.A Invariant Subspaces 135

Because T v D v if and only if .T  I /v D 0, a vector v 2 V with v ¤ 0

is an eigenvector of T corresponding to  if and only if v 2 null.T  I /.

5.8 Example Suppose T 2 L.F2 / is defined by

T .w; z/ D .z; w/:

(a) Find the eigenvalues and eigenvectors of T if F D R.

(b) Find the eigenvalues and eigenvectors of T if F D C.

Solution

(a) If F D R, then T is a counterclockwise rotation by 90ı about the

origin in R2 . An operator has an eigenvalue if and only if there exists a
nonzero vector in its domain that gets sent by the operator to a scalar
multiple of itself. A 90ı counterclockwise rotation of a nonzero vector
in R2 obviously never equals a scalar multiple of itself. Conclusion: if
F D R, then T has no eigenvalues (and thus has no eigenvectors).
(b) To find eigenvalues of T, we must find the scalars  such that
T .w; z/ D .w; z/
has some solution other than w D z D 0. The equation above is
equivalent to the simultaneous equations
5.9 z D w; w D z:
Substituting the value for w given by the second equation into the first
equation gives
z D 2 z:
Now z cannot equal 0 [otherwise 5.9 implies that w D 0; we are
looking for solutions to 5.9 where .w; z/ is not the 0 vector], so the
equation above leads to the equation
1 D 2 :
The solutions to this equation are  D i and  D i . You should
be able to verify easily that i and i are eigenvalues of T. Indeed,
the eigenvectors corresponding to the eigenvalue i are the vectors of
the form .w; wi /, with w 2 C and w ¤ 0, and the eigenvectors
corresponding to the eigenvalue i are the vectors of the form .w; wi/,
with w 2 C and w ¤ 0.
136 CHAPTER 5 Eigenvalues, Eigenvectors, and Invariant Subspaces

Now we show that eigenvectors corresponding to distinct eigenvalues are

linearly independent.

5.10 Linearly independent eigenvectors

Let T 2 L.V /. Suppose 1 ; : : : ; m are distinct eigenvalues of T and
v1 ; : : : ; vm are corresponding eigenvectors. Then v1 ; : : : ; vm is linearly
independent.

Proof Suppose v1 ; : : : ; vm is linearly dependent. Let k be the smallest posi-

tive integer such that
5.11 vk 2 span.v1 ; : : : ; vk1 /I
the existence of k with this property follows from the Linear Dependence
Lemma (2.21). Thus there exist a1 ; : : : ; ak1 2 F such that
5.12 vk D a1 v1 C    C ak1 vk1 :
Apply T to both sides of this equation, getting
k vk D a1 1 v1 C    C ak1 k1 vk1 :
Multiply both sides of 5.12 by k and then subtract the equation above, getting
0 D a1 .k  1 /v1 C    C ak1 .k  k1 /vk1 :
Because we chose k to be the smallest positive integer satisfying 5.11,
v1 ; : : : ; vk1 is linearly independent. Thus the equation above implies that all
the a’s are 0 (recall that k is not equal to any of 1 ; : : : ; k1 ). However, this
means that vk equals 0 (see 5.12), contradicting our hypothesis that vk is an
eigenvector. Therefore our assumption that v1 ; : : : ; vm is linearly dependent
was false.

The corollary below states that an operator cannot have more distinct
eigenvalues than the dimension of the vector space on which it acts.

5.13 Number of eigenvalues

Suppose V is finite-dimensional. Then each operator on V has at most
dim V distinct eigenvalues.

Proof Let T 2 L.V /. Suppose 1 ; : : : ; m are distinct eigenvalues of T.

Let v1 ; : : : ; vm be corresponding eigenvectors. Then 5.10 implies that the list
v1 ; : : : ; vm is linearly independent. Thus m  dim V (see 2.23), as desired.
 SECTION 5.A Invariant Subspaces 137

Restriction and Quotient Operators

If T 2 L.V / and U is a subspace of V invariant under T, then U determines
two other operators T jU 2 L.U / and T =U 2 L.V =U / in a natural way, as
defined below.

5.14 Definition T jU and T =U

Suppose T 2 L.V / and U is a subspace of V invariant under T.

 The restriction operator T jU 2 L.U / is defined by

T jU .u/ D T u
for u 2 U.

 The quotient operator T =U 2 L.V =U / is defined by

.T =U /.v C U / D T v C U
for v 2 V.

For both the operators defined above, it is worthwhile to pay attention

to their domains and to spend a moment thinking about why they are well
defined as operators on their domains. First consider the restriction operator
T jU 2 L.U /, which is T with its domain restricted to U, thought of as
mapping into U instead of into V. The condition that U is invariant under T
is what allows us to think of T jU as an operator on U, meaning a linear map
into the same space as the domain, rather than as simply a linear map from
one vector space to another vector space.
To show that the definition above of the quotient operator makes sense,
we need to verify that if v C U D w C U, then T v C U D T w C U. Hence
suppose v C U D w C U. Thus v  w 2 U (see 3.85). Because U is invariant
under T, we also have T .v  w/ 2 U, which implies that T v  T w 2 U, which
implies that T v C U D T w C U, as desired.
Suppose T is an operator on a finite-dimensional vector space V and U is
a subspace of V invariant under T, with U ¤ f0g and U ¤ V. In some sense,
we can learn about T by studying the operators T jU and T =U, each of which
is an operator on a vector space with smaller dimension than V. For example,
proof 2 of 5.27 makes nice use of T =U.
However, sometimes T jU and T =U do not provide enough information
about T. In the next example, both T jU and T =U are 0 even though T is not
the 0 operator.
138 CHAPTER 5 Eigenvalues, Eigenvectors, and Invariant Subspaces

5.15 Example Define an operator T 2 L.F2 / by T .x; y/ D .y; 0/. Let

U D f.x; 0/ W x 2 Fg. Show that

(a) U is invariant under T and T jU is the 0 operator on U ;

(b) there does not exist a subspace W of F2 that is invariant under T and
such that F2 D U ˚ W ;

(c) T =U is the 0 operator on F2 =U.

Solution

(a) For .x; 0/ 2 U, we have T .x; 0/ D .0; 0/ 2 U. Thus U is invariant

under T and T jU is the 0 operator on U.

(b) Suppose W is a subspace of V such that F2 D U ˚ W. Because

dim F2 D 2 and dim U D 1, we have dim W D 1. If W were invariant
under T, then each nonzero vector in W would be an eigenvector of T.
However, it is easy to see that 0 is the only eigenvalue of T and that all
eigenvectors of T are in U. Thus W is not invariant under T.

(c) For .x; y/ 2 F2 , we have

 
.T =U / .x; y/ C U D T .x; y/ C U
D .y; 0/ C U
D 0 C U;

where the last equality holds because .y; 0/ 2 U. The equation above
shows that T =U is the 0 operator.

EXERCISES 5.A

1 Suppose T 2 L.V / and U is a subspace of V.

(a) Prove that if U  null T, then U is invariant under T.

(b) Prove that if range T  U, then U is invariant under T.

2 Suppose S; T 2 L.V / are such that S T D T S. Prove that null S is

invariant under T.
 SECTION 5.B Eigenvectors and Upper-Triangular Matrices 143

5.B Eigenvectors and Upper-Triangular

Matrices
Polynomials Applied to Operators
The main reason that a richer theory exists for operators (which map a vector
space into itself) than for more general linear maps is that operators can be
raised to powers. We begin this section by defining that notion and the key
concept of applying a polynomial to an operator.
If T 2 L.V /, then T T makes sense and is also in L.V /. We usually write
2
T instead of T T. More generally, we have the following definition.

5.16 Definition Tm
Suppose T 2 L.V / and m is a positive integer.

 T m is defined by
Tm D T    T… :
„ ƒ‚
m times

 T 0 is defined to be the identity operator I on V.

 If T is invertible with inverse T 1 , then T m is defined by

m
T m D .T 1 / :

You should verify that if T is an operator, then

T m T n D T mCn and .T m /n D T mn ;
where m and n are allowed to be arbitrary integers if T is invertible and
nonnegative integers if T is not invertible.

5.17 Definition p.T /

Suppose T 2 L.V / and p 2 P.F/ is a polynomial given by
p.z/ D a0 C a1 z C a2 z 2 C    C am z m
for z 2 F. Then p.T / is the operator defined by
p.T / D a0 I C a1 T C a2 T 2 C    C am T m :

This is a new use of the symbol p because we are applying it to operators,

not just elements of F.
144 CHAPTER 5 Eigenvalues, Eigenvectors, and Invariant Subspaces

 
5.18 Example Suppose D 2 L P.R/ is the differentiation operator
defined by Dq D q 0 and p is the polynomial defined by p.x/ D 73x C5x 2 .
Then p.D/ D 7I  3D C 5D 2 ; thus
 
p.D/ q D 7q  3q 0 C 5q 00
for every q 2 P.R/.

If we fix an operator T 2 L.V /, then the function from P.F/ to L.V /

given by p 7! p.T / is linear, as you should verify.

5.19 Definition product of polynomials

If p; q 2 P.F/, then pq 2 P.F/ is the polynomial defined by
.pq/.z/ D p.z/q.z/
for z 2 F.

Any two polynomials of an operator commute, as shown below.

5.20 Multiplicative properties

Part (a) holds because when ex-
Suppose p; q 2 P.F/ and T 2 L.V /. panding a product of polynomials
Then using the distributive property, it
(a) .pq/.T / D p.T /q.T /; does not matter whether the sym-
bol is z or T.
(b) p.T /q.T / D q.T /p.T /.

Proof
Pm j
Pn k
(a) Suppose p.z/ D j D0 aj z and q.z/ D kD0 bk z for z 2 F.
Then
m X
X n
.pq/.z/ D aj bk z j Ck :
j D0 kD0
Thus
m X
X n
.pq/.T / D aj bk T j Ck
j D0 kD0
X m  X
n 
j k
D aj T bk T
j D0 kD0
D p.T /q.T /:

(b) Part (a) implies p.T /q.T / D .pq/.T / D .qp/.T / D q.T /p.T /.
 SECTION 5.B Eigenvectors and Upper-Triangular Matrices 145

Existence of Eigenvalues
Now we come to one of the central results about operators on complex vector
spaces.

5.21 Operators on complex vector spaces have an eigenvalue

Every operator on a finite-dimensional, nonzero, complex vector space
has an eigenvalue.

Proof Suppose V is a complex vector space with dimension n > 0 and

T 2 L.V /. Choose v 2 V with v ¤ 0. Then

v; T v; T 2 v; : : : ; T n v

is not linearly independent, because V has dimension n and we have n C 1

vectors. Thus there exist complex numbers a0 ; : : : ; an , not all 0, such that

0 D a0 v C a1 T v C    C an T n v:

Note that a1 ; : : : ; an cannot all be 0, because otherwise the equation above

would become 0 D a0 v, which would force a0 also to be 0.
Make the a’s the coefficients of a polynomial, which by the Fundamental
Theorem of Algebra (4.14) has a factorization

a0 C a1 z C    C an z n D c.z  1 /    .z  m /;

where c is a nonzero complex number, each j is in C, and the equation holds

for all z 2 C (here m is not necessarily equal to n, because an may equal 0).
We then have

0 D a 0 v C a1 T v C    C a n T n v
D .a0 I C a1 T C    C an T n /v
D c.T  1 I /    .T  m I /v:

Thus T  j I is not injective for at least one j . In other words, T has an

eigenvalue.

The proof above depends on the Fundamental Theorem of Algebra, which

is typical of proofs of this result. See Exercises 16 and 17 for possible ways to
rewrite the proof above using the idea of the proof in a slightly different form.
146 CHAPTER 5 Eigenvalues, Eigenvectors, and Invariant Subspaces

Upper-Triangular Matrices
In Chapter 3 we discussed the matrix of a linear map from one vector space
to another vector space. That matrix depended on a choice of a basis of each
of the two vector spaces. Now that we are studying operators, which map a
vector space to itself, the emphasis is on using only one basis.

5.22 Definition matrix of an operator, M.T /

Suppose T 2 L.V / and v1 ; : : : ; vn is a basis of V. The matrix of T with
respect to this basis is the n-by-n matrix
0 1
A1;1 : : : A1;n
B :: :: C
M.T / D @ : : A
An;1 : : : An;n
whose entries Aj;k are defined by
T vk D A1;k v1 C    C An;k vn :
If the basis is not clear from the context, then the notation
M T; .v1 ; : : : ; vn / is used.

Note that the matrices of operators are square arrays, rather than the more
general rectangular arrays that we considered earlier for linear maps.
The k th column of the matrix If T is an operator on Fn and no
M.T / is formed from the coeffi- basis is specified, assume that the basis
cients used to write T vk as a linear in question is the standard one (where
combination of v1 ; : : : ; vn . the j th basis vector is 1 in the j th slot
and 0 in all the other slots). You can
then think of the j th column of M.T / as T applied to the j th basis vector.

5.23 Example Define T 2 L.F3 / by T .x; y; z/ D .2xCy; 5yC3z; 8z/.

Then 0 1
2 1 0
M.T / D @ 0 5 3 A :
0 0 8

A central goal of linear algebra is to show that given an operator T 2 L.V /,

there exists a basis of V with respect to which T has a reasonably simple
matrix. To make this vague formulation a bit more precise, we might try to
choose a basis of V such that M.T / has many 0’s.
 SECTION 5.B Eigenvectors and Upper-Triangular Matrices 147

If V is a finite-dimensional complex vector space, then we already know

enough to show that there is a basis of V with respect to which the matrix of
T has 0’s everywhere in the first column, except possibly the first entry. In
other words, there is a basis of V with respect to which the matrix of T looks
like 0 1

B 0  C
B C
B :: CI
@ : A
0
here the  denotes the entries in all the columns other than the first column.
To prove this, let  be an eigenvalue of T (one exists by 5.21) and let v be a
corresponding eigenvector. Extend v to a basis of V. Then the matrix of T
with respect to this basis has the form above.
Soon we will see that we can choose a basis of V with respect to which
the matrix of T has even more 0’s.

5.24 Definition diagonal of a matrix

The diagonal of a square matrix consists of the entries along the line from
the upper left corner to the bottom right corner.

For example, the diagonal of the matrix in 5.23 consists of the entries
2; 5; 8.

5.25 Definition upper-triangular matrix

A matrix is called upper triangular if all the entries below the diagonal
equal 0.

For example, the matrix in 5.23 is upper triangular.

Typically we represent an upper-triangular matrix in the form
0 1
1 
B :: C
@ : AI
0 n
the 0 in the matrix above indicates We often use  to denote matrix en-
that all entries below the diagonal in tries that we do not know about or
this n-by-n matrix equal 0. Upper- that are irrelevant to the questions
triangular matrices can be considered being discussed.
reasonably simple—for n large, almost
half its entries in an n-by-n upper-
triangular matrix are 0.
148 CHAPTER 5 Eigenvalues, Eigenvectors, and Invariant Subspaces

The following proposition demonstrates a useful connection between

upper-triangular matrices and invariant subspaces.

5.26 Conditions for upper-triangular matrix

Suppose T 2 L.V / and v1 ; : : : ; vn is a basis of V. Then the following are
equivalent:

(a) the matrix of T with respect to v1 ; : : : ; vn is upper triangular;

(b) T vj 2 span.v1 ; : : : ; vj / for each j D 1; : : : ; n;

(c) span.v1 ; : : : ; vj / is invariant under T for each j D 1; : : : ; n.

Proof The equivalence of (a) and (b) follows easily from the definitions and
a moment’s thought. Obviously (c) implies (b). Hence to complete the proof,
we need only prove that (b) implies (c).
Thus suppose (b) holds. Fix j 2 f1; : : : ; ng. From (b), we know that

T v1 2 span.v1 /  span.v1 ; : : : ; vj /I
T v2 2 span.v1 ; v2 /  span.v1 ; : : : ; vj /I
::
:
T vj 2 span.v1 ; : : : ; vj /:

Thus if v is a linear combination of v1 ; : : : ; vj , then

T v 2 span.v1 ; : : : ; vj /:

In other words, span.v1 ; : : : ; vj / is invariant under T, completing the proof.

The next result does not hold on Now we can prove that for each
real vector spaces, because the first operator on a finite-dimensional com-
vector in a basis with respect to plex vector space, there is a basis of the
which an operator has an upper- vector space with respect to which the
triangular matrix is an eigenvector matrix of the operator has only 0’s be-
of the operator. Thus if an opera- low the diagonal. In Chapter 8 we will
tor on a real vector space has no
improve even this result.
eigenvalues [see 5.8(a) for an ex-
ample], then there is no basis with Sometimes more insight comes from
respect to which the operator has seeing more than one proof of a theo-
an upper-triangular matrix. rem. Thus two proofs are presented of
the next result. Use whichever appeals
more to you.
 SECTION 5.B Eigenvectors and Upper-Triangular Matrices 149

5.27 Over C, every operator has an upper-triangular matrix

Suppose V is a finite-dimensional complex vector space and T 2 L.V /.
Then T has an upper-triangular matrix with respect to some basis of V.

Proof 1 We will use induction on the dimension of V. Clearly the desired

result holds if dim V D 1.
Suppose now that dim V > 1 and the desired result holds for all complex
vector spaces whose dimension is less than the dimension of V. Let  be any
eigenvalue of T (5.21 guarantees that T has an eigenvalue). Let

U D range.T  I /:

Because T  I is not surjective (see 3.69), dim U < dim V. Furthermore,

U is invariant under T. To prove this, suppose u 2 U. Then

T u D .T  I /u C u:

Obviously .T  I /u 2 U (because U equals the range of T  I ) and

u 2 U. Thus the equation above shows that T u 2 U. Hence U is invariant
under T, as claimed.
Thus T jU is an operator on U. By our induction hypothesis, there is a
basis u1 ; : : : ; um of U with respect to which T jU has an upper-triangular
matrix. Thus for each j we have (using 5.26)

5.28 T uj D .T jU /.uj / 2 span.u1 ; : : : ; uj /:

Extend u1 ; : : : ; um to a basis u1 ; : : : ; um ; v1 ; : : : ; vn of V. For each k, we

have
T vk D .T  I /vk C vk :
The definition of U shows that .T  I /vk 2 U D span.u1 ; : : : ; um /. Thus
the equation above shows that

5.29 T vk 2 span.u1 ; : : : ; um ; v1 ; : : : ; vk /:

From 5.28 and 5.29, we conclude (using 5.26) that T has an upper-
triangular matrix with respect to the basis u1 ; : : : ; um ; v1 ; : : : ; vn of V, as
desired.
150 CHAPTER 5 Eigenvalues, Eigenvectors, and Invariant Subspaces

Proof 2 We will use induction on the dimension of V. Clearly the desired

result holds if dim V D 1.
Suppose now that dim V D n > 1 and the desired result holds for all
complex vector spaces whose dimension is n  1. Let v1 be any eigenvector
of T (5.21 guarantees that T has an eigenvector). Let U D span.v1 /. Then U
is an invariant subspace of T and dim U D 1.
Because dim V =U D n  1 (see 3.89), we can apply our induction hy-
pothesis to T =U 2 L.V =U /. Thus there is a basis v2 C U; : : : ; vn C U of
V =U such that T =U has an upper-triangular matrix with respect to this basis.
Hence by 5.26,

.T =U /.vj C U / 2 span.v2 C U; : : : ; vj C U /

for each j D 2; : : : ; n. Unraveling the meaning of the inclusion above, we

see that
T vj 2 span.v1 ; : : : ; vj /
for each j D 1; : : : ; n. Thus by 5.26, T has an upper-triangular matrix
with respect to the basis v1 ; : : : ; vn of V, as desired (it is easy to verify that
v1 ; : : : ; vn is a basis of V ; see Exercise 13 in Section 3.E for a more general
result).

How does one determine from looking at the matrix of an operator whether
the operator is invertible? If we are fortunate enough to have a basis with
respect to which the matrix of the operator is upper triangular, then this
problem becomes easy, as the following proposition shows.

5.30 Determination of invertibility from upper-triangular matrix

Suppose T 2 L.V / 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
of that upper-triangular matrix are nonzero.

Proof Suppose v1 ; : : : ; vn is a basis of V with respect to which T has an

upper-triangular matrix
0 1
1 
B 2 C
B C
5.31 M.T / D B : C:
@ :: A
0 n

We need to prove that T is invertible if and only if all the j ’s are nonzero.
 SECTION 5.B Eigenvectors and Upper-Triangular Matrices 151

First suppose the diagonal entries 1 ; : : : ; n are all nonzero. The upper-
triangular matrix in 5.31 implies that T v1 D 1 v1 . Because 1 ¤ 0, we have
T .v1 =1 / D v1 ; thus v1 2 range T.
Now
T .v2 =2 / D av1 C v2
for some a 2 F. The left side of the equation above and av1 are both in
range T ; thus v2 2 range T.
Similarly, we see that

T .v3 =3 / D bv1 C cv2 C v3

for some b; c 2 F. The left side of the equation above and bv1 ; cv2 are all in
range T ; thus v3 2 range T.
Continuing in this fashion, we conclude that v1 ; : : : ; vn 2 range T. Be-
cause v1 ; : : : ; vn is a basis of V, this implies that range T D V. In other words,
T is surjective. Hence T is invertible (by 3.69), as desired.
To prove the other direction, now suppose that T is invertible. This implies
that 1 ¤ 0, because otherwise we would have T v1 D 0.
Let 1 < j  n, and suppose j D 0. Then 5.31 implies that T maps
span.v1 ; : : : ; vj / into span.v1 ; : : : ; vj 1 /. Because

dim span.v1 ; : : : ; vj / D j and dim span.v1 ; : : : ; vj 1 / D j  1;

this implies that T restricted to dim span.v1 ; : : : ; vj / is not injective (by 3.23).
Thus there exists v 2 span.v1 ; : : : ; vj / such that v ¤ 0 and T v D 0. Thus T
is not injective, which contradicts our hypothesis (for this direction) that T is
invertible. This contradiction means that our assumption that j D 0 must be
false. Hence j ¤ 0, as desired.

As an example of the result above, we see that the operator in Example 5.23
is invertible.
Unfortunately no method exists for Powerful numeric techniques exist
exactly computing the eigenvalues of for finding good approximations to
an operator from its matrix. However, the eigenvalues of an operator from
if we are fortunate enough to find a ba- its matrix.
sis with respect to which the matrix of
the operator is upper triangular, then the
problem of computing the eigenvalues
becomes trivial, as the following propo-
sition shows.
152 CHAPTER 5 Eigenvalues, Eigenvectors, and Invariant Subspaces

5.32 Determination of eigenvalues from upper-triangular matrix

Suppose T 2 L.V / has an upper-triangular matrix with respect to some
basis of V. Then the eigenvalues of T are precisely the entries on the
diagonal of that upper-triangular matrix.

Proof Suppose v1 ; : : : ; vn is a basis of V with respect to which T has an

upper-triangular matrix
0 1
1 
B 2 C
B C
M.T / D B : C:
@ : : A
0 n

Let  2 F. Then
0 1
1   
B 2   C
B C
M.T  I / D B :: C:
@ : A
0 n  

Hence T  I is not invertible if and only if  equals one of the numbers

1 ; : : : ; n (by 5.30). Thus  is an eigenvalue of T if and only if  equals one
of the numbers 1 ; : : : ; n .

5.33 Example Define T 2 L.F3 / by T .x; y; z/ D .2x C y; 5y C 3z; 8z/.

What are the eigenvalues of T ?

Solution The matrix of T with respect to the standard basis is

0 1
2 1 0
M.T / D @ 0 5 3 A :
0 0 8

Thus M.T / is an upper-triangular matrix. Now 5.32 implies that the eigen-
values of T are 2, 5, and 8.

Once the eigenvalues of an operator on Fn are known, the eigenvectors

can be found easily using Gaussian elimination.
 SECTION 5.C Eigenspaces and Diagonal Matrices 155

5.C Eigenspaces and Diagonal Matrices

5.34 Definition diagonal matrix
A diagonal matrix is a square matrix that is 0 everywhere except possibly
along the diagonal.

5.35 Example 0 1
8 0 0
@ 0 5 0 A
0 0 5
is a diagonal matrix.

Obviously every diagonal matrix is upper triangular. In general, a diagonal

matrix has many more 0’s than an upper-triangular matrix.
If an operator has a diagonal matrix with respect to some basis, then the
entries along the diagonal are precisely the eigenvalues of the operator; this
follows from 5.32 (or find an easier proof for diagonal matrices).

5.36 Definition eigenspace, E.; T /

Suppose T 2 L.V / and  2 F. The eigenspace of T corresponding to ,
denoted E.; T /, is defined by
E.; T / D null.T  I /:
In other words, E.; T / is the set of all eigenvectors of T corresponding
to , along with the 0 vector.

For T 2 L.V / and  2 F, the eigenspace E.; T / is a subspace of V

(because the null space of each linear map on V is a subspace of V ). The
definitions imply that  is an eigenvalue of T if and only if E.; T / ¤ f0g.

5.37 Example Suppose the matrix of an operator T 2 L.V / with respect

to a basis v1 ; v2 ; v3 of V is the matrix in Example 5.35 above. Then
E.8; T / D span.v1 /; E.5; T / D span.v2 ; v3 /:

If  is an eigenvalue of an operator T 2 L.V /, then T restricted to

E.; T / is just the operator of multiplication by .
156 CHAPTER 5 Eigenvalues, Eigenvectors, and Invariant Subspaces

5.38 Sum of eigenspaces is a direct sum

Suppose V is finite-dimensional and T 2 L.V /. Suppose also that
1 ; : : : ; m are distinct eigenvalues of T. Then

E.1 ; T / C    C E.m ; T /

is a direct sum. Furthermore,

dim E.1 ; T / C    C dim E.m ; T /  dim V:

Proof To show that E.1 ; T / C    C E.m ; T / is a direct sum, suppose

u1 C    C um D 0;
where each uj is in E.; T /. Because eigenvectors corresponding to distinct
eigenvalues are linearly independent (see 5.10), this implies that each uj
equals 0. This implies (using 1.44) that E.1 ; T / C    C E.m ; T / is a direct
sum, as desired.
Now
 
dim E.1 ; T / C    C dim E.m ; T / D dim E.1 ; T / ˚    ˚ E.m ; T /
 dim V;
where the equality above follows from Exercise 16 in Section 2.C.

5.39 Definition diagonalizable

An operator T 2 L.V / is called diagonalizable if the operator has a
diagonal matrix with respect to some basis of V.

5.40 Example Define T 2 L.R2 / by

T .x; y/ D .41x C 7y; 20x C 74y/:
The matrix of T with respect to the standard basis of R2 is
 
41 7
;
20 74
which is not a diagonal matrix. However, T is diagonalizable, because the
matrix of T with respect to the basis .1; 4/; .7; 5/ is
 
69 0
;
0 46
as you should verify.
 SECTION 5.C Eigenspaces and Diagonal Matrices 157

5.41 Conditions equivalent to diagonalizability

Suppose V is finite-dimensional and T 2 L.V /. Let 1 ; : : : ; m denote
the distinct eigenvalues of T. Then the following are equivalent:

(a) T is diagonalizable;

(b) V has a basis consisting of eigenvectors of T ;

(c) there exist 1-dimensional subspaces U1 ; : : : ; Un of V, each invariant

under T, such that

V D U1 ˚    ˚ Un I

(d) V D E.1 ; T / ˚    ˚ E.m ; T /;

(e) dim V D dim E.1 ; T / C    C dim E.m ; T /.

Proof An operator T 2 L.V / has a diagonal matrix

0 1
1 0
B :: C
@ : A
0 n

with respect to a basis v1 ; : : : ; vn of V if and only if T vj D j vj for each j .

Thus (a) and (b) are equivalent.
Suppose (b) holds; thus V has a basis v1 ; : : : ; vn consisting of eigenvectors
of T. For each j , let Uj D span.vj /. Obviously each Uj is a 1-dimensional
subspace of V that is invariant under T. Because v1 ; : : : ; vn is a basis of V,
each vector in V can be written uniquely as a linear combination of v1 ; : : : ; vn .
In other words, each vector in V can be written uniquely as a sum u1 C  Cun ,
where each uj is in Uj . Thus V D U1 ˚    ˚ Un . Hence (b) implies (c).
Suppose now that (c) holds; thus there are 1-dimensional subspaces
U1 ; : : : ; Un of V, each invariant under T, such that V D U1 ˚    ˚ Un .
For each j , let vj be a nonzero vector in Uj . Then each vj is an eigenvector
of T. Because each vector in V can be written uniquely as a sum u1 C  Cun ,
where each uj is in Uj (so each uj is a scalar multiple of vj ), we see that
v1 ; : : : ; vn is a basis of V. Thus (c) implies (b).
At this stage of the proof we know that (a), (b), and (c) are all equivalent.
We will finish the proof by showing that (b) implies (d), that (d) implies (e),
and that (e) implies (b).
158 CHAPTER 5 Eigenvalues, Eigenvectors, and Invariant Subspaces

Suppose (b) holds; thus V has a basis consisting of eigenvectors of T.

Hence every vector in V is a linear combination of eigenvectors of T, which
implies that
V D E.1 ; T / C    C E.m ; T /:
Now 5.38 shows that (d) holds.
That (d) implies (e) follows immediately from Exercise 16 in Section 2.C.
Finally, suppose (e) holds; thus

5.42 dim V D dim E.1 ; T / C    C dim E.m ; T /:

Choose a basis of each E.j ; T /; put all these bases together to form a list
v1 ; : : : ; vn of eigenvectors of T, where n D dim V (by 5.42). To show that
this list is linearly independent, suppose

a1 v1 C    C an vn D 0;

where a1 ; : : : ; an 2 F. For each j D 1; : : : ; m, let uj denote the sum of all

the terms ak vk such that vk 2 E.j ; T /. Thus each uj is in E.j ; T /, and

u1 C    C um D 0:

Because eigenvectors corresponding to distinct eigenvalues are linearly inde-

pendent (see 5.10), this implies that each uj equals 0. Because each uj is a
sum of terms ak vk , where the vk ’s were chosen to be a basis of E.j ; T /, this
implies that all the ak ’s equal 0. Thus v1 ; : : : ; vn is linearly independent and
hence is a basis of V (by 2.39). Thus (e) implies (b), completing the proof.

Unfortunately not every operator is diagonalizable. This sad state of affairs

can arise even on complex vector spaces, as shown by the next example.

5.43 Example Show that the operator T 2 L.C2 / defined by

T .w; z/ D .z; 0/

is not diagonalizable.

Solution As you should verify, 0 is the only eigenvalue of T and furthermore

E.0; T / D f.w; 0/ 2 C2 W w 2 Cg.
Thus conditions (b), (c), (d), and (e) of 5.41 are easily seen to fail (of
course, because these conditions are equivalent, it is only necessary to check
that one of them fails). Thus condition (a) of 5.41 also fails, and hence T is
not diagonalizable.
 SECTION 5.C Eigenspaces and Diagonal Matrices 159

The next result shows that if an operator has as many distinct eigenvalues
as the dimension of its domain, then the operator is diagonalizable.

5.44 Enough eigenvalues implies diagonalizability

If T 2 L.V / has dim V distinct eigenvalues, then T is diagonalizable.

Proof Suppose T 2 L.V / has dim V distinct eigenvalues 1 ; : : : ; dim V .

For each j , let vj 2 V be an eigenvector corresponding to the eigenvalue j .
Because eigenvectors corresponding to distinct eigenvalues are linearly inde-
pendent (see 5.10), v1 ; : : : ; vdim V is linearly independent. A linearly indepen-
dent list of dim V vectors in V is a basis of V (see 2.39); thus v1 ; : : : ; vdim V
is a basis of V. With respect to this basis consisting of eigenvectors, T has a
diagonal matrix.

5.45 Example Define T 2 L.F3 / by T .x; y; z/ D .2x C y; 5y C 3z; 8z/.

Find a basis of F3 with respect to which T has a diagonal matrix.

Solution With respect to the standard basis, the matrix of T is

0 1
2 1 0
@ 0 5 3 A:
0 0 8
The matrix above is upper triangular. Thus by 5.32, the eigenvalues of T are
2, 5, and 8. Because T is an operator on a vector space with dimension 3 and
T has three distinct eigenvalues, 5.44 assures us that there exists a basis of F3
with respect to which T has a diagonal matrix.
To find this basis, we only have to find an eigenvector for each eigenvalue.
In other words, we have to find a nonzero solution to the equation

T .x; y; z/ D .x; y; z/

for  D 2, then for  D 5, and then for  D 8. These simple equations are
easy to solve: for  D 2 we have the eigenvector .1; 0; 0/; for  D 5 we have
the eigenvector .1; 3; 0/; for  D 8 we have the eigenvector .1; 6; 6/.
Thus .1; 0; 0/; .1; 3; 0/; .1; 6; 6/ is a basis of F3 , and with respect to this
basis the matrix of T is 0 1
2 0 0
@ 0 5 0 A:
0 0 8