When Is A Linear Operator Diagonalizable?

Marco Abate
0. Introduction
As it often happens, everything began with a mistake. I was teaching for the third year
in a row a linear algebra course to engineering freshmen. One of the highlights of the
course was eigenvector theory, and in particular the diagonalization of linear operators on
nite-dimensional vector spaces (i.e., of square real or complex matrices). Toward the end
of the course I assigned a standard homework: prove that the matrix
A =

1 1 2
1 0 1
0 1 1

,
is diagonalizable. Easy enough, I thought. The characteristic polynomial is
p
A
() = det(AI
3
) =
3
+ ,
whose roots are evidently 0, 1, 1. We have three distinct eigenvalues in a three-dimen-
sional space, and a standard theorem ensures that A is diagonalizable.
To my surprise, the students came complaining that they were unable to solve the
exercise. Perplexed (some of the complaining students were very bright), I looked over the
exercise againand I understood. What happened was that, in the homework, I actually
gave them the matrix
B =

1 1 2
1 0 1
0 1 1

,
whose characteristic polynomial is
p
B
() =
3
2
2
+ 2,
which has no rational roots. The students were unable to compute the eigenvalues of B,
and they got stuck.
This accident started me wondering whether it might be possible to decide when a
linear operator T on a nite-dimensional real or complex vector space is diagonalizable
without computing the eigenvalues. If one is looking for an orthonormal basis of eigenvec-
tors, the answer is well known to be yes: the spectral theorem says that such a basis exists
in the complex case if and only if T is normal (i.e., it commutes with its adjoint), and if
and only if T is symmetric in the real case. The aim of this note is to give an explicit
procedure to decide whether a given linear operator on a nite-dimensional real or complex
vector space is diagonalizable. By explicit I mean that it can always be worked out with
pen and paper; it can be long, it can be tedious, but it can be done. Its ingredients (the
minimal polynomial and Sturms theorem) are not new; but putting them together yields
a result that can be useful as an aside in linear algebra classes.
2 Marco Abate
1. The minimal polynomial
The rst main ingredient in our procedure is the minimal polynomial. Let T: V V be a
linear operator on a nite-dimensional vector space over the eld K. We shall denote by T
k
the composition of T with itself k times, and for any polynomial p(t) = a
k
t
k
+ +a
0
K[t]
we put
p(T) = a
k
T
k
+ + a
1
T + id
V
,
and say that p is monic if a
k
= 1. A minimal polynomial
T
K[t] of the linear operator T
is a monic polynomial of minimal degree such that (T) = 0.
The theory of the minimal polynomial is standard. For completeness sake, I briey
recall the results we shall need. First of all:
Proposition 1.1: Let T: V V be a linear operator on a nite-dimensional vector
space V over the eld K. Then:
(i) the minimal polynomial
T
of T exists, has degree at most n = dimV , and is unique;
(ii) if p K[t] is such that p(T) = 0, then there is some q K[t] such that p = q
T
.
For our procedure it is important to show that the minimal polynomial can be explic-
itly computed. Take v V , and let d be the minimal non-negative integer such that the
vectors {v, T(v), . . . , T
d
(v)} are linearly dependent. Clearly d n always; d = 0 if and
only if v = 0, and d = 1 if and only if v is an eigenvector of T. Choose a
0
, . . . , a
d1
K
such that
T
d
(v) + a
d1
T
d1
(v) + + a
1
T(v) + a
0
v = 0
(note that we can assume the coecient of T
d
(v) to be 1 because of the minimality of d),
and then set

T,v
(t) = t
d
+ a
d1
t
d1
+ + a
1
t + a
0
K[t].
By denition, v Ker
T,v
(T); more precisely,
T,v
is the monic polynomial p K[t] of
least degree such that v Ker p(T).
Now, if p K[t] is any common multiple of
T,v
1
and
T,v
2
for any two vectors v
1
and v
2
, then both v
1
and v
2
belong to Ker p(T). More generally, if B = {v
1
, . . . , v
n
} is
a basis of V , and p is any common multiple of
T,v
1
, . . . ,
T,v
n
, then B Ker p(T), and
thus p(T) = 0. Hence the following result comes as no surprise:
Proposition 1.2: Let T: V V be a linear operator on a nite-dimensional vector
space V over the eld K. Let B = {v
1
, . . . , v
n
} be a basis of V . Then
T
is the least
common multiple of
T,v
1
, . . . ,
T,v
n
.
Proof : Let p K[t] be the least common multiple of
T,v
1
, . . . ,
T,v
n
. We have al-
ready remarked that p(T) = 0, and so
T
divides p. Conversely, for j = 1, . . . , n write

T
= q
j

T,v
j
+ r
j
, with deg r
j
< deg
T,v
j
. Then
0 =
T
(T)v
j
= q
j
(T)

T,v
j
(T)v
j

+ r
j
(T)v
j
= r
j
(T)v
j
,
and the minimality of the degree of
T,v
j
forces r
j
0. Since every
T,v
j
divides
T
, their
least common multiple p also divides
T
, and hence p =
T
.
When Is A Linear Operator Diagonalizable? 3
Thus one method to compute the minimal polynomial is to compute the polynomials

T,v
1
, . . . ,
T,v
n
and then their least common multiple. To avoid unnecessary calculations,
it could be useful to remember that deg
T
n.
Example 1. Let us compute the minimal polynomials of the matrices A and B of
the introduction. Let B = {e
1
, e
2
, e
3
} be the canonical basis of R
3
. We have
Ae
1
= Be
1
=

1
1
0

, A
2
e
1
= B
2
e
1
=

2
1
1

,
A
3
e
1
=

1
1
0

= Ae
1
, B
3
e
1
=

1
1
2

= 2B
2
e
1
Be
1
+ 2e
1
;
therefore

A,e
1
(t) = t
3
t,
B,e
1
(t) = t
3
+ 2t
2
+ t 2.
Since deg
A,e
1
= 3 and the minimal polynomial of A should be a monic multiple of
A,e
1
of
degree at most three, we can conclude that
A
=
A,e
1
without computing
A,e
2
and
A,e
3
(and it is easy to check that
A,e
2
(t) = t
2
t and
A,e
3
(t) = t
3
t). For the same reason
we have
B
=
B,e
1
.
Let
1
, . . . ,
k
K be the distinct eigenvalues of T. If T is diagonalizable, then
Proposition 1.2 immediately yields
T
(t) = (t
1
) (t
k
). This is the standard
characterization of diagonalizable linear operators:
Theorem 1.3: Let T: V V be a linear operator on a nite-dimensional vector space V
over the eld K. Then T is diagonalizable if and only if
T
is of the form

T
(t) = (t
1
) (t
k
), (1.1)
where
1
, . . . ,
k
are distinct elements of K.
Therefore to decide whether a given linear operator on a nite-dimensional vec-
tor space is diagonalizable it suces to check whether its minimal polynomial is of the
form (1.1).
2. The procedure
Our aim now is to nd an eective procedure to decide whether a given polynomial p K[t]
can be written in the form (1.1). To do so, we need to know when all the roots of p have
multiplicity one, and when they all belong to the eld K. The rst question has a standard
answer:
Proposition 2.1: Let p K[t] be a non-constant polynomial, and let p

K[t] denote its

have a common root in K;

) of p and p

has a root in K.
Recalling Theorem 1.3 we get the following
4 Marco Abate
Corollary 2.2: Let T: V V be a linear operator on a nite-dimensional vector space V
over the eld K. Then:
(i) If K is algebraically closed (e.g., K = C), then T is diagonalizable if and only if
g.c.d.(
T
,

T
) = 1;
(ii) If K is not algebraically closed (e.g., K = R), then T is diagonalizable if and only if
all the roots of
T
are in K and g.c.d.(
T
,

T
) = 1.
To decide whether a complex linear operator T is diagonalizable it then suces to
compute the greatest common divisor of
T
and

T
. On the other hand, if K = R this is
not enough; to complete the picture we need Sturms theoremand to state it we need a
few more denitions.
Let c = (c
0
, . . . , c
s
) R
s+1
be a nite sequence of real numbers. If c
0
c
s
= 0, the
number of variations in sign of c is the number of indices 1 j s such that c
j1
c
j
< 0
(that is, such that c
j1
and c
j
have opposite sign). If some element of c is zero, then the
number of variations in sign of c is the number of variations in sign of the sequence of
non-zero elements of c. We denote the number of variations in sign of c by V
c
.
Now let p R[t] be a non-constant polynomial. The standard sequence associated
with p is the sequence p
0
, . . . , p
s
R[t] dened by
p
0
= p, p
1
= p

,
p
0
= q
1
p
1
p
2
, with deg p
2
< deg p
1
,
.
.
.
.
.
.
p
j1
= q
j
p
j
p
j+1
, with deg p
j+1
< deg p
j
,
.
.
.
.
.
.
p
s1
= q
s
p
s
, (that is, p
s+1
0).
In other words, the standard sequence is obtained changing the sign in the remain-
der term of the Euclidean algorithm for the computation of g.c.d.(p, p

). In particular,
g.c.d.(p, p

) = 1 if and only if p
s
is constant.
Sturms theorem then says:
Theorem 2.3: Let p R[t] be a polynomial such that g.c.d.(p, p

) = 1, and take a < b

and b =

p
0
(b), . . . , p
s
(b)

.
For a proof see [1, pp. 295299].
Now, for any polynomial p(t) = a
d
t
d
+ +a
0
R[t] there exists M > 0 such that p(t)
has the same sign as a
d
, the leading coecient of p, if t M and the same sign as (1)
d
a
d
if t M. In particular, all the roots of p are contained in [M, M], and Sturms theorem
implies the following:
Corollary 2.4: Let p R[t] be a non-constant polynomial such that g.c.d.(p, p

) = 1.
Let p
0
, . . . , p
s
R[t] be the standard sequence associated with p, and let d
j
be the de-
gree and c
j
R the leading coecient of p
j
for j = 0, . . . , s. Then the number of real
When Is A Linear Operator Diagonalizable? 5
roots of p is given by V

V
+
, where V

is the number of variations in sign of the se-

(1)
d
0
c
0
, . . . , (1)
d
s
c
s

, and V
+
is the number of variations in sign of the sequence
(c
0
, . . . , c
s
).
Proof : It suces to choose M > 0 large enough so that p
j
(t) has the same sign as c
j
when t M and the same sign as (1)
d
j
c
j
when t M, for each j = 0, . . . , s, and then
apply Sturms theorem with a = M and b = M.
We nally have all the ingredients necessary to state the desired procedure. Let
T: V V be a linear operator on a nite-dimensional vector space V over K = R or C.
Then:
(1) Compute the minimal polynomial
T
.
(2) Compute the standard sequence p
0
, . . . , p
s
associated with
T
. If p
s
is not constant,
then T is not diagonalizable. If p
s
is constant and K = C, then T is diagonalizable.
If p
s
is constant and K = R, go to Step (3).
(3) Compute V

and V
+
for
T
. Then T is diagonalizable if and only if V

V
+
= deg
T
.
Thus we are always able to decide whether a given linear operator on a nite-dimensional
real or complex vector space is diagonalizable or not. One feature that I nd particularly
interesting in this procedure is that the solution of a typical linear algebra problem is
reduced to an apparently totally unrelated manipulation of polynomials, showing in a
simple case how dierent parts of mathematics can be connected in unexpected ways.
We end this note with some examples of application of our procedure.
Example 2. First of all, we solve the original homework. We have already computed
the minimal polynomial
B
(t) = t
3
+2t
2
+t2. The standard sequence associated with
B
is
p
0
(t) = t
3
+ 2t
2
+ t 2, p
1
(t) = 3t
2
+ 4t + 1, p
2
(t) =
2
9
t +
20
9
, p
3
(t) = 261.
Since p
3
is constant, B is diagonalizable over C. To compute V

V
+
we count the number
of variations in sign of the sequences (1, 3,
2
9
, 261) and (1, 3,
2
9
, 261). We obtain
V

V
+
= 2 1 = 1 < 3 = deg
B
,
and so B is not diagonalizable over R. On the other hand, the standard sequence associated
with
A
is
p
0
(t) = t
3
t, p
1
(t) = 3t
2
1, p
2
(t) =
2
3
t, p
3
(t) = 1.
The number of variations in sign of (1, 3,
2
3
, 1) is V

= 3, and of (1, 3,
2
3
, 1) is V
+
= 0;
therefore V

V
+
= 3 0 = 3, and thus A is diagonalizable over R (as it should be).
Both these matrices were diagonalizable over C; since their minimal polynomials have
degree 3, necessarily their (complex) eigenvalues are all distinct. In the next example this
is not the case:
6 Marco Abate
Example 3. Let
C =

0 2 2 6
2 4 1 5
3 4 3 7
1 2 1 3

.
To compute the minimal polynomial of C we start, as in Example 1, by applying the
iterates of C to e
1
. We get
Ce
1
=

0
2
3
1

, C
2
e
1
=

4
6
10
4

, C
3
e
1
=

8
6
14
6

, C
4
e
1
=

4
8
0
0

.
It is easy to check that {e
1
, Ce
1
, C
2
e
1
, C
3
e
1
} are linearly independent and that
C
4
e
1
4C
3
e
1
+ 8C
2
e
1
8Ce
1
+ 4e
1
= 0;
therefore deg
C,e
1
= 4 and, as in Example 1, we can conclude that

C
(t) =
C,e
1
(t) = t
4
4t
3
+ 8t
2
8t + 4.
The standard sequence associated with
C
starts with
p
0
(t) = t
4
4t
3
+ 8t
2
8t + 4, p
1
(t) = 4t
3
12t
2
+ 16t 8, p
2
(t) = t
2
+ 2t 2.
Since p
2
divides p
1
, it is the last polynomial in the sequence; since it is not constant, we
conclude that C is not diagonalizable even over C (and in particular it cannot have four
distinct eigenvalues).
Our nal example involves a minimal polynomial of degree strictly less than the di-
mension of the space:
Example 4. Let
D =

2 2 2 8
2 4 1 9
1 4 3 11
1 2 1 5

.
To compute the minimal polynomial of D we start again by applying the iterates of D
to e
1
. We get
De
1
=

2
2
1
1

, D
2
e
1
=

2
4
2
2

= 2De
1
2e
1
;
When Is A Linear Operator Diagonalizable? 7
therefore
D,e
1
(t) = t
2
2t + 2, and we cannot conclude right now that
D,e
1
=
D
.
Proceeding with the computations we get
De
2
=

2
4
4
2

, D
2
e
2
=

4
6
8
4

= 2De
2
2e
2
; De
3
=

2
1
3
1

, D
2
e
3
=

4
2
4
2

= 2De
3
2e
3
;
De
4
=

8
9
11
5

, D
2
e
4
=

16
18
22
12

= 2De
4
2e
4
;
hence we have
D,e
2
=
D,e
3
=
D,e
4
=
D,e
1
and
D
(t) = t
2
2t + 2. In particular, D
has (at most) two distinct complex eigenvalues.
The standard sequence associated with
D
is
p
0
(t) = t
2
2t + 2, p
1
(t) = 2t 2, p
2
(t) = 1.
Since p
2
is constant, D is diagonalizable over C. The number of variations in sign of the
sequences (1, 2, 1) and (1, 2, 1) is
V

V
+
= 1 1 = 0 < 2 = deg
D
,
and so D is not diagonalizable over R.
References
[1] N. Jacobson, Basic Algebra, I. Freeman, San Francisco, 1974.
Marco Abate
Dipartimento di Matematica
Universit` a di Ancona
Via Brecce Bianche
60131 Ancona
Italy
[email protected]
When Is A Linear Operator Diagonalizable?
Marco Abate
Brief Descriptive Summary
Have you ever wondered whether that mystifying 1010 matrix were diagonalizable?
Computing the characteristic polynomial is useless; there are no self-evident eigenvalues
in view. And you dont know how to write a program to make the computer do the work.
And you are losing your sleep about it (well, almost). Grieve no more! We are proud to
present an explicit pen-and-paper procedure to let you decide whether any given square
matrix is diagonalizable, both over the complex and over the real numbers! Read and try
yourself; your sleep wont be troubled anymore.
[If this is too much of a joke, the following summary can be used instead.]
To decide whether a given square matrix is diagonalizable the usual techniques depend
on precisely computing the eigenvalues, which often is a hopeless task. This paper describes
instead an explicit pen-and-paper procedure to check diagonalizability without using any
information about the eigenvalues; the most dicult operation needed is just the Euclidean
algorithm for computing the greatest common divisor of two polynomials.
E-mail: [email protected]
When Is A Linear Operator Diagonalizable? 9
When Is A Linear Operator Diagonalizable?
Marco Abate
Biographical Sketch
MARCO ABATE was born in Milan in 1962, got his Ph.D. from Scuola Normale
Superiore, Pisa in 1988, and now is full professor of Geometry at the University of Ancona
(placing him among the ten youngest italian full professors). His interests include geomet-
rical function theory, holomorphic dynamical systems, complex dierential geometry and
traveling around the world. In his spare time, he writes comics, sometimes even about
mathematics.