Notes on Canonical Forms

2012-04-24 08:40
Jeffrey Diller
April 24, 2012

1 Introduction
Throughout these notes we take T : V → V to be a linear operator on a finite dimensional1
vector space V over a field F. Our goal is to relate the structure of T to that of its charac-
teristic polynomial pchar (λ) := det(λ id − T ). The model for this project, and the best case
scenario, occurs when T is diagonalizable: we have a basis B = {v1 , . . . vn } ⊂ V , in which
each vj is an eigenvector for some eigenvalue λj ∈ F of T . The matrix for T relative to B is
then diagonal, given by  
λ1
[T ]BB = 
 ... 

λn
In particular the characteristic polynomial of T factors completely: pchar (λ) = (λ−λ1 ) . . . (λ−
λn ).
We have seen, however, that not all linear operators are diagonalizable. There are essen-
tially two obstacles to diagonalization, which we illustrate by example. Consider the linear
operators on S, T : R2 → R2 with standard matrices
   
0 1 1 1
AS = , AT = ,
−1 0 0 1

respectively. The characteristic polynomial of S is λ2 + 1, which has no real roots and

therefore no (real) eigenvectors. This leaves us two independent eigenvectors short of a basis
for R2 . Note that the problem here is more with the field R than with the operator. If
we change our field to C, then the characteristic polynomial has two distinct complex roots
±i and therefore two independent eigenvectors, which suffice to diagonalize S as a complex
linear operator.
The characteristic polynomial of T , on the other hand, is (λ − 1)2 . So T has two real (but
equal) roots λ = 1, 1. The problem is that the eigenspace of λ = 1 is only one dimensional
spanned by (1, 0). So this time we are one eigenvector short of a basis. This problem is
much less common than the one we had with T , but it is more serious: we cannot make T
1
The assumption dim V < ∞ isn’t always needed, but ultimately it’s the case we care about here.

1
diagonalizable by simply changing the field to C. Indeed, any field F contains elements 0
and 1, so in fact the matrix AT defines a linear operator T : F2 → F2 over any field F, and
the missing eigenvector argument shows that T is not diagonalizable over F either.
All of this discussion is meant to suggest that the behavior of a linear operator T is
closely tied to the way in which its characteristic polynomial can or cannot be factored. To
make this idea more precise requires us to replace the notion of eigenvector with something
more general and flexible. Namely,

Definition 1.1 A subspace H ⊂ V is T -invariant if T (H) ⊂ H.

Exercise 1.1 Prove each of the following.

• A one dimensional subspace H is invariant if and only if it is generated by an eigen-

vector.

• The eigenspace of an eigenvalue is an invariant subspace (regardless of its dimension).

• If H1 , H2 ⊂ V are invariant, then so are H1 ∩ H2 and H1 + H2 .

• If F = C, then any invariant subspace H ⊂ V contains a one dimensional invariant

subspace.

Exercise 1.2 Find all invariant subspaces of the operators S and T above. Do the same for
the operators with standard matrices
   
1 0 1 0
, .
0 −1 0 1

These illustrate all the main possibilities for linear operators on two dimensional vector
spaces. Can you prove this? What further possibilities are there for linear operators on three
dimensional vector spaces?

Note that if H is a T -invariant subspace, then we obtain another a linear operator

T : H → H simply by limiting the domain of T to consist only of vectors in H. We call
this operator the restriction of T to H and denote it by T |H . Below, we will discuss two
important ways to locate and describe invariant subspaces of V . Before any of this, however,
we digress a bit and consider polynomials as objects of interest in their own right.

2 Polynomials
In this section, we present and discuss some useful facts concerning the set

F[x] := {ck xk + · · · + c1 x + c0 : cj ∈ F}

of all polynomials with coefficients in F. We have already seen one element of F[x], the
characteristic polynomial, associated with the linear operator T . In this section we will find
another polynomial, the minimal polynomial, that is canonically associated to T .

2
 The degree of a polynomial p(x) = ck xk + · · · + c0 is the largest power k of x for which
the coefficient ck is non-zero. We adopt the convention that deg 0 = −∞. We call a non-zero
polynomial monic if its leading coefficient ck = 1.
Polynomials in F[x] can be added, subtracted and multiplied in the usual way, and all the
relevant axioms for arithmetic hold. In contrast with F itself, however, there is no operation
of division2 on F[x]. If a, b, q ∈ F[x] are polynomials such that a = bq, then we will say that
a is divisible by b or, more commonly, that b divides a, signifying the relationship by writing
b|a. We will often take advantage of the fact that b|a implies deg b ≤ deg a.
Even when b does not divide a, we may still perform ‘division with remainder’. This is
arguably the most important basic fact concerning polynomials with field coefficients.

Theorem 2.1 (Division algorithm) For any polynomials a(x), b(x) ∈ F[x], there are
unique q(x), r(x) ∈ F[x] such that deg r < deg b and

a = bq + r.

Proof. Let S ⊂ F[x] be the set of all polynomials of the form a − bp for some p ∈ F[x]. Let
r = a − bq ∈ S be a polynomial (possibly zero) of minimal degree. Suppose deg r(x) = k
with leading coefficient ck 6= 0 and that deg b(x) = ` with leading coefficient c0` . If k ≥ `,
then
ck ck
r(x) − 0 xk−` b = a − (q + xk−` )b ∈ S
c` c`
is a polynomial with degree strictly smaller than k, because the leading terms in the difference
on the left cancel each other. This contradicts the minimality of deg r, so it must be instead
that k < `. We conclude that a = bq + r where deg r < deg b, as the theorem asserts.
To prove that r, q ∈ F[x] are unique, suppose that r̃, q̃ ∈ F[x] also satisfy the conclusion
of the theorem. Then bq + r = bq̃ + r̃. Rearranging, we find that

b(q − q̃) = r − r̃.

Comparing degrees then gives

deg b + deg(q − q̃) = deg(r − r̃) < deg b,

which implies that deg(q − q̃) < 0; i.e. q = q̃, and therefore r = r̃. So the polynomials
q, r ∈ F[x] are unique. 
All the other results in this section, whether we prove them or not, depend ultimately
on the division algorithm. The reader might note in all this that there is a very compelling
analogy between polynomials and integers, with the notion of ‘degree’ for polynomials playing
the role of ‘absolute value’ for integers. In particular, the notion of ‘prime number’ is replaced
by that of ‘irreducible polynomial’.

Definition 2.2 A non-constant polynomial p ∈ F[x] is called irreducible if the only the
polynomials in F[x] that divide p are constants and constant multiples of p.
2
in mathematical parlance, this state of affairs is summarized by saying that F[x] is not a field, but rather
a commutative ring.

3
 Any polynomial of degree one is irreducible. The fundamental theorem of algebra (‘every
complex polynomial of degree at least one has a complex root’) implies that when F = C,
the converse statement holds: any irreducible polynomial in C[x] has degree one.
For arbitrary fields, it is a tricky thing to determine whether a given polynomial of degree
two or higher is irreducible. For instance x2 + 1 is irreducible as a polynomial in R[x] but
not as a polynomial in C[x]. Likewise x2 − 2 is irreducible as a polynomial in Q[x] but not
as a polynomial in R[x]. Keeping this in mind might make the next two theorems seem a
little less ‘obvious’. The hard part of each theorem is the uniqueness.

Theorem 2.3 Given any two polynomials a, b ∈ F[x], not both equal to zero, there is a
unique monic d ∈ F[x] such that d|a, d|b and deg d ≥ deg d˜ for every other d˜ ∈ F[x] that
divides both a and b. In fact if d˜ ∈ F[x] divides both a and b, then d|d,
˜ too.

The polynomial d is called the greatest common divisor of a and b and denoted gcd(a, b).
If gcd(a, b) = 1, then a and b are said to be relatively prime. It turns out, for reasons we
discuss below, that gcd(a, b) is not very sensitive to the underlying field. For instance

gcd(x4 − 1, 3x3 + 3x) = x2 + 1

regardless of whether x4 − 1 and x3 + x are though of as polynomials in Q[x], in R[x], or

in C[x]. This makes the concept of ‘relatively prime polynomials’ more straightforward in
many cases than that of ‘irreducible polynomial’.

Theorem 2.4 Every non-constant polynomial p(x) ∈ F[x] can be factored

p = q1 . . . qk

into irreducible polynomials qj ∈ F[x]. The factorization is unique except for the order and
leading coefficients of the polynomials qj .

The decomposition of p into irreducible polynomials is called the prime factorization of

p. Often the ambiguity concerning leading coefficients in prime factorizations is addressed
by requiring p and all the factors qj to be monic. Moreover, it is common to acknowledge re-
peated factors explicitly in prime factorizations by writing the factorization in the alternative
form
p = q1m1 . . . q`m` ,
and implicitly assuming that all the qj are distinct (i.e. i 6= j implies qi 6= cqj for any c ∈ F).
Here is the concept that links the division algorithm to the previous two results.

Definition 2.5 A non-empty set of polynomials S ⊂ F[x] is called an ideal if for any
a, b ∈ S and p ∈ F[x], we have that a + b ∈ S and ap ∈ S.

The resemblance between the notion of an ‘ideal’ of F[x] and that of a ‘subspace’ of a
vector space is not a coincidence. The main fact concerning ideals of F[x] is that they are
all ‘one dimensional.’

4
Theorem 2.6 Suppose that S ⊂ F[x] is an ideal containing at least one non-zero polynomial.
Then S contains a unique (up to constant multiple) non-zero polynomial of smallest possible
degree, and in fact
S = pF[x] := {pq : q ∈ F[x]}
is the set of all polynomial multiples of p.
The polynomial p in the statement of this theorem is called the generator of S. We can
(and usually do) assume with no loss of generality that p is monic.
Proof. Given p as in the theorem, we have by definition of ideal that S contains every poly-
nomial multiple pq, q ∈ F[x] of p; i.e. that pF[x] ⊂ S. Suppose now (to get a contradiction)
that S contains something that is not a multiple of p. That is, suppose there exists p̃ ∈ S
such that p does not divide p̃. Then by the division algorithm, we have r, q ∈ F[x] such that
deg r < deg p and p̃ = pq + r. Since p does not divide p̃, it follows that r 6= 0. Moreover,
since r = pq − p̃ we have from the definition of ideal that r ∈ S. That is, there is a non-zero
element of S whose degree is smaller than that of p—a contradiction. We conclude that p̃
does not exist and that S is precisely equal to pF[x].
To see that p is unique, suppose that p̃ ∈ S is another non-zero polynomial of smallest
degree. Then, as we have just shown, p̃ = pq for some q ∈ F[x]. Since deg p = deg p̃ =
deg p + deg q, it follows that deg q = 0. That is, q = c0 ∈ F is a constant. 
We illustrate the power of the ‘ideal’ concept as follows.
Proof of Theorem 2.3. Given a, b ∈ F[x] as in the theorem, we let
S = {ap + bq : p, q ∈ F[x]}
be the set of all polynomial combinations of a and b. The reader will (on pain of lightening
strike for failing to comply) verify that S is an ideal of F[x] and that S contains a non-zero
element. Hence S = dF[x], where d ∈ S is the unique non-zero and monic element of smallest
degree.
Then on the one hand, we have d|a and d|b, since a, b ∈ S. And on the other hand d
belongs to S, so we have by definition of S that
d = ap + bq
for some p, q ∈ F[x]. From this, one may (i.e. you will now pull out pencil and paper in
order to) deduce that any other common factor d˜ of a and b also divides d. In particular, if
deg d˜ ≤ deg d, and if deg d˜ = deg d, then d˜ and d are just constant multiples of one another.
Hence d = gcd(a, b) is unique. 
Incidentally, the same idea leads to a very efficient method for actually computing greatest
common divisors called the Euclidean algorithm. I’ll be happy to provide further details in
person. Beyond showing the usefulness of ideals, our discussion contains some facts that we
will need later. These I summarize as follows.
Theorem 2.7 For any non-zero polynomials a, b ∈ F[x], there are p, q ∈ F[x] such that
ap + bq = gcd(a, b).
In particular, if a and b are relatively prime, then there are p, q ∈ F[x] such that ap + bq = 1.

5
 Now let us return to the linear operator T : V → V which is the main object considered
in these notes. If p(x) = cn xn + · · · + c0 ∈ F[x] is any polynomial, then we can define a new
linear operator p(T ) : V → V by substituting T for the unknown x:

p(T ) := cn T n + · · · + c0 id.

Note here that T j means the j-fold composition T ◦ T ◦ · · · ◦ T .

We note several convenient features of this construction as follows, leaving the reader the
exercise of verifying them.

Proposition 2.8 If p, q ∈ F[x] are polynomials, then

• every T -invariant subspace of V is also p(T ) invariant;

• p(T ) ◦ q(T ) = q(T ) ◦ p(T ) = (pq)(T );

• ker p(T ) is a T -invariant subspace.

• If p|q, then ker p(T ) ⊂ ker q(T ).

The second assertion in Proposition 2.8 says among other things that for any p, q ∈ F[x],
the operators p(T ) and q(T ) commute. Typically in what follows, we will write p(T )q(T )
instead of p(T )◦q(T ). Besides emphasizing the connection between composition of operators
and multiplication of polynomials, this abbreviation accords well with our tendency to write
T v instead of T (v) when the parentheses start to pile up.
The third assertion in Proposition 2.8 affords us one of two basic means for finding
invariant subspaces of V . Usually, ker p(T ) = {0} is trivial, but we will see below that
when p is e.g. a factor of the characteristic polynomial of T , the subspace ker p(T ) is more
interesting. For instance, when p(x) = x − λ for some root λ of pchar , then ker p(T ) is just
the eigenspace for λ.
A particularly important case occurs when ker p(T ) = V .

Proposition 2.9 The set

IT = {p ∈ F[x] : p(T ) = 0}
of polynomials that ‘annihilate’ T is a non-trivial ideal.

Proof. We show only that T is non-trivial, leaving the reader to verify that IT is an ideal.
Note that the vector space L(V ) of linear operators on V is a finite dimensional vector space.
Hence when N = dim L(V ) = (dim V )2 , it follows that the N + 1 operators id, T, . . . , T N
are dependent: there exists a non-trivial combination

cN T N + · · · + c1 T + c0 id = 0

that vanishes. Hence cN xN + . . . c1 x + c0 ∈ I(x). 

Definition 2.10 The generator of IT is called the minimal polynomial pmin (x) of T .

6
 The proof of Proposition 2.9 shows that deg pmin ≤ (dim V )2 . We will see later that in
fact deg pmin ≤ dim V . For now, we content ourselves with the following exercise, which
suggests that the minimal and characteristic polynomials of T are closely related.

Exercise 2.1 Suppose that T is diagonalizable.

1. Show that if the roots of pchar are all distinct, then pmin = pchar .

2. Show that if the roots of pchar are all equal to λ (i.e. T = λid), then pmin (x) = x − λ.

3. Show most generally that if pchar (x) = (x − λ1 )m1 . . . (x − λ` )m` , where λj ∈ F are all
distinct, then
pmin (x) = (x − λ1 ) . . . (x − λ` ).

Finally, give an example of an operator T (evidently not diagonalizable) for which pmin (x) =
(x − 1)2 .

3 Quotient Spaces
Let H ⊂ V be any subspace of our vector space V . In this section, we show how to define
a new vector space V /H whose role is similar to that of a complementary subspace. Recall
that since V is finite dimensional, we always have subspaces H 0 ⊂ V complementing H.
However, these are generally far from being unique. The quotient space V /H, on the other
hand, is uniquely defined, and will therefore serve as a canonical replacement for the choice
of a complement.

Definition 3.1 Two vectors v1 , v2 ∈ V are equivalent modulo H if their difference v1 − v2

belongs to H. In this case we write v1 ∼H v2 . To each v ∈ V we associate the subset

ṽ := {w ∈ V : w ∼ v},

which we call the equivalence class of v; conversely, we call any vector w ∈ ṽ a representative
of ṽ. We let
V /H := {ṽ : v ∈ V }
denote the set of all possible equivalence classes of vectors in V . We call V /H the quotient
space of V modulo H.

Exercise 3.1 Do each of the following.

1. Show that ∼ is an equivalence relation. More precisely, for any vectors u, v, w, show
that ∼ is

• reflexive: v ∼ v;
• symmetric: v ∼ w implies w ∼ v;
• transitive: u ∼ v and v ∼ w implies that u ∼ w.

7
 2. Show that, consequently, V /H is a partition of V . That is,
• If two equivalence classes ṽ, w̃ intersect, then ṽ = w̃ ;
• V is the union of all equivalence classes ṽ ∈ V /H.

One might visualize V /H as a deck of cards, where H = 0̃ is the card through the origin,
and any other ‘card’ ṽ in the deck is obtained by translating H away from 0 by the vector
v. Note that V /H is not a subset of V but rather a set of subsets of V . The great thing
is that we can add these sets to one another and multiply them by scalars in a well-defined
way. The idea is deceptively simple. For any vectors v, w ∈ V and any scalar λ ∈ F, we
declare
• ṽ + w̃ = v^
+ w;
• λṽ = λv;
f

That is, in order to e.g. add equivalence classes, we first choose vectors representing each
class, add these representatives and then take the equivalence class of the sum. The problem
is one of ‘well-definedness’; we need to know that the final result does not depend on which
vectors we choose to represent our equivalence classes.

Theorem 3.2 The operations + and · are well-defined on V /H, and with these operations
V /H becomes a vector space over F. In particular,
• the additive identity in V /H is 0̃.
• the additive inverse of any vector ṽ ∈ V /H is given by −ṽ = −v
f for any

• the function π(v) = ṽ is a surjective linear transformation π : V → V /H.

Finally, dim V /H = dim V − dim H.

Proof. To see that addition on V /H is well-defined, suppose that v1 , v2 , w1 , w2 ∈ V satisfy

v1 ∼ v2 and w1 ∼ w2 ; i.e. suppose that ṽ1 = ṽ2 and w̃1 = w̃2 . Then by definition
v1 − v2 , w1 − w2 ∈ H. Hence

(v1 + w1 ) − (v2 + w2 ) = (v1 − v2 ) + (w1 − w2 ) ∈ H,

since H is closed with respect to addition. Thus v1 + w1 ∼ v2 + w2 , and addition is well-

defined. We leave it to the reader to verify that scalar multiplication is also well-defined.
The vector space axioms for V /H and linearity of the quotient map π : V → V /H now
follow from chasing definitions and applying the vector space axioms on V . Let us verify, for
instance, that scalar multiplication is distributive over addition: by definition of the vector
space operations on V /H, we have

λ(ṽ + ũ) := λ(v

^ ^
+ u) := λ(v ^
+ u) = λv + λu := λv
f + λu
f := λṽ + λũ.

To give another example, let us also verify that π respects scalar multiplication:

π(λv) := λv
f := λe
v := λπ(v).

8
We leave the other verifications to the reader.
Concerning the formula for the dimension of V /H, observe that by definition of V /H,
the quotient map π is surjective and has kernel equal to H. So applying the rank theorem
to π gives us that
dim V /H + dim H = dim V.


Theorem 3.3 Let H 0 ⊂ V be any subspace complementary to H. Then the restriction π :

H 0 → V /H is an isomorphism. V /H is isomorphic to any subspace H 0 ⊂ V complementing
H. More precisely, if B 0 = {v1 , . . . , vk } is a basis for H 0 , then B̃ = {ṽ1 , . . . , ṽk } is a basis
for V /H.

Proof. Since ker π = H and H ∩ H 0 = {0}, we have that ker π|H 0 is trivial; i.e. π|H 0 is
injective. Since dim H 0 = dim V /H, we conclude that π is also surjective and therefore and
isomorphism. The final assertion proceeds from the fact that isomorphisms carry bases to
bases. 
Now let us bring our linear operator T : V → V back into the picture.

Theorem 3.4 If H ⊂ V is a T -invariant subspace, then T ‘induces’ a linear operator

T̃ : V /H → V /H given by
T̃ (ṽ) = T]
(v).
If, moreover, B = {v1 , . . . , vn } is a basis for V obtained by extending a basis BH = {v1 , . . . , vk }
for H, then the matrix for T relative to B has block upper triangular form
 
[T |H ]BH ∗
[T ]BB = ,
0 [T̃ ]B̃

where B̃ = {ṽk+1 , . . . , ṽn } is the induced basis for V /H.

Proof. Note that the proposed definition of T̃ (ṽ) depends which v we choose to represent
ṽ. However, if u ∼ v is another vector representing ṽ, then u − v ∈ H and by invariance
T (u − v) ∈ H, too. Therefore, linearity of the projection map and of T give us that

T]
(u) − T] ^
(v) = T (u − v) = 0.

Hence T̃ is well-defined. Linearity of T̃ is inherited from T ; e.g. given ṽ ∈ V /H and λ ∈ F,

we have
f := T^
T̃ (λṽ) := T (λv) ^
(λv) = λT (v) := λT] (v) := λT̃ (ṽ).
Now we turn to the assertion relating the matrices for T , T |H and T̃ . For 1 ≤ j ≤ k, we
have bj ∈ H and therefore T bj ∈ H. Hence
 
[T bj ]BH
[T bj ]B = .
0

9
  
[T |H ]BH
So we see that the first k columns of [T ]BB are
0
Turning to the remaining columns k + 1 ≤ j ≤ n of A, we note that if

T bj = c1 b1 + . . . cn bn ,

then because each vector in H is equivalent to 0 modulo H,

T̃ b̃j = ck+1 b̃k+1 + · · · + cn b̃n .

So the last n − k coordinates of T bj relative

! to B are equal to the coordinates of T̃ b̃j relative
h ∗i
to B̃. That is, [T bj ]B = . This means that taken together, the last n − k
T̃ b̃j
B̃  
∗
columns of A comprise a matrix of the form . So A has the block form asserted in
[T̃ ]B̃
the theorem.


Corollary 3.5 If H ⊂ V is a T -invariant subspace, then pchar (λ) = pH (λ)p̃(λ) where pH is

the characteristic polynomial of T |H and p̃ is the characteristic polynomial of T̃ .

In the case where F = C, we can use Theorem 3.4 to ‘upper triangularize’ the operator
T . We don’t really need the result in what follows, but it costs us very little effort to state
and prove it now.

Corollary 3.6 (Schur Representation) If V is a complex vector space, then there is a

basis B ⊂ V such that [T ]BB is upper triangular with diagonal entries equal to the roots of
pchar .

Proof. We work by induction on dim V . The case dim V = 1 is immediate. Supposing the
assertions are true when dim V = n − 1, we consider the case dim V = n. Since F = C, the
fundamental theorem of algebra gives us a root λ ∈ C of pchar . Let v1 ∈ V be an eigenvector
with eigenvalue λ. Then H := span{v} is a one-dimensional invariant subspace of V , so
Theorem 3.4 gives us an induced operator T̃ : V /H → V /H. Since dim V /H = n − 1, our
inductive hypothesis gives us a basis B̃ := {ṽ2 , . . . , ṽn } ⊂ V /H such that [T̃ ]BB is upper
triangular.
We claim that B := {v1 , v2 , . . . , vn } is a basis for V . To see this, note that independence
of ṽ2 , . . . , ṽn means that no non-trivial combination of v2 , . . . , vn lies in H. In particular, no
non-trivial combination can vanish or be equal to v1 ; i.e. v2 , . . . , vn are independent and v1
is not in their span. This implies that v1 , v2 , . . . , vn are independent and, since n = dim V ,
form a basis for V as claimed.
Finally, Theorem 3.4 gives us that
 
λ ∗
[T ]BB == ,
0 [T̃ ]B̃B̃
which, since the bottom right block is upper triangular, is itself upper triangular. 

10
4 The Cayley-Hamilton Theorem
We saw above that the kernel of p(T ) for any p ∈ F[x] is a T -invariant subspace. We begin
this section by describing a second way to construct invariant subspaces. Let v ∈ V be any
vector, and consider the forward orbit v, T v, T 2 v, . . . of v under T . This is typically an
infinite set of vectors, but since V is finite dimensional, the finite segment v, . . . , T k v will
be linearly dependent for k ∈ N large enough. Taking k to be the first such integer, we have
that {v, . . . , T k−1 v} is independent and therefore a basis for the subspace Hv that it spans.
We call Hv the cyclic subspace generated by T and v.
Theorem 4.1 The cyclic subspace Hv generated by v is T -invariant, and the minimal and
characteristic polynomials of T |Hv are the same.
Proof. By definition of Hv , any vector w ∈ Hv can be written w = c0 v + · · · + ck−1 T k−1 v =
p(T )v where p(x) = c0 + c1 x + . . . ck−1 xk−1 ∈ F[x] is a polynomial of degree less than k. Our
choice of k implies that T k v ∈ Hv . Hence,
T w = T ◦ p(T )v = p(T )(T v) = c0 T v + · · · + ck−1 T k v ∈ Hv ,
too, since Hv is closed with respect to linear combinations. This shows that Hv is T -invariant.
Moreover, taking w = T k v, and setting pv (x) = xk − p(x), we see that pv (T )v =
T k v − p(T )v = 0. We will show that p is both the minimal and the characteristic polynomial
of the restricted operator T |Hv .
To see that pv is the minimal polynomial, first observe that if p ∈ F[x] is any non-zero
polynomial with deg p < deg pv = k, then p(T )v is a non-trivial linear combination of the
basis vectors v, T v, . . . , T k−1 v for Hv . Hence p(T )v 6= 0 and in particular p(T )|Hv does not
vanish (i.e. is not the zero operator). On the other hand, if w = p(T )v is any vector in Hv ,
then
pv (T )w = pv (T )p(T )v = p(T )pv (T )v = p(T )0 = 0.
So pv is a monic polynomial with minimal degree among polynomials p ∈ F[x] such that
p(T )|Hv = 0; i.e. pv is the minimal polynomial of T |Hv
Let us now rewrite pv (x) = xk + ck−1 xk−1 + · · · + c0 (this amounts to reversing the
signs of the scalars cj at the beginning of the proof). It remains to show that pv (x) =
det(xid − T ). To this end, we ask the reader to verify that the matrix of T relative to the
basis {v, T v, . . . , T k−1 v} is
 
0 0 0 . . . 0 −c0
1 0 0 . . . 0 −c1 
 
0 1 0 . . . 0 −c2 
A := 0 0 1 . . . 0 −c 
 
 3 
 .. 
 . 
0 0 0 . . . 1 −ck−1
We further ask that the reader use e.g. cofactor expansion about the last column of the
matrix xI − A to compute (this takes some concentration but probably less concentration
than following someone elses version of the computation) that
det(xI − A) = pv

11
as asserted. 
We will call any subspace H ⊂ V cyclic if it is the cyclic subspace associated to some
vector v ∈ V . We call the operator T cyclic if V = Hv is itself a cyclic subspace. Though
we will not prove it here, most linear operators are cyclic.
Exercise 4.1 Verify the following.
1. For any v ∈ V , the set Iv := {p ∈ F[x] : p(T )v = 0} is an ideal generated by pv .
2. pv |pmin for any v ∈ V .
3. A one dimensional subspace is cyclic if and only if it is spanned by an eigenvector.
4. A diagonalizable linear operator is cyclic if and only if it has no repeated eigenvalues
(in the ’if ’ direction, it helps to use a fact about Vandermonde determinants).
5. If the minimal polynomial of T is irreducible, then T is cyclic.
The fact that pmin = pchar when T is cyclic suggests that there might be a close rela-
tionship between the minimal and characteristic polynomials for more general T . The next
result describes this relationship in general.
Theorem 4.2 For any linear operator T : V → V on a finite dimenionsal vector space V
we have pmin |pchar . Conversely, any irreducible factor p of pchar must also divide pmin .
To put it slightly differently, the prime decompositions of pmin and pchar have the same
irreducible factors, but the multiplicities of the factors of pchar can be larger. As an immediate
consequence of the first assertion in Theorem 4.2 and the fourth item in Propostion 2.8 we
obtain the well-known Cayley-Hamilton Theorem.
Corollary 4.3 pchar (T ) = 0.
Proof. Given any v ∈ V , Theorems 3.4 and 4.1 together imply that that pchar = pv q for
some q ∈ F[x]. Thus
pchar (T )v = q(T )pv (T )v = q(T )0 = 0.
This proves for any non-zero v ∈ V that pchar (T )v = 0; i.e. pchar (T ) is the zero operator.
Hence pchar ∈ IT which means that pmin |pchar .
We prove the second assertion in the theorem by induction on dim V . If dim V = 1, then
T = λid for some λ ∈ F, and one verifies the assertion readily. So assume the assertion is
true whenever dim V < n, and consider the case dim V = n. Let p ∈ F[x] be an irreducible
factor of pchar and let v ∈ V be a non-zero vector. Since pchar = pv p̃char where p̃char is the
characteristic polynomial for the induced operator T̃ on V /Hv , we have that p|pv or p|p̃char .
In the first case, we have from the exercise above that pv |pmin . Since divisibility is
transitive, it follows that p|pmin as desired. In the second case, the fact that v 6= 0 implies
that dim V /Hv = dim V − dim Hv ≤ n − 1. So our inductive hypothesis and p|p̃ implies
that p divides the minimal polynomial p̃min of T̃ . But we saw in ???? that p̃min |pmin , so
transitivity again gives p|pmin . This completes the induction step and the proof 
In closing this section we note an interesting consequence of the Theorem 4.2, whose
proof we leave as an exercise for the reader.

12
Corollary 4.4 Let p ∈ F[x] be any polynomial. Then p(T ) is invertible if and only p and
pchar are relatively prime.

5 Direct Sums
Definition 5.1 We say that the subspaces H1 , . . . , Hk ⊂ V are independent if the only
vectors v1 ∈ H1 , . . . , vk ∈ Hk satisfying

v1 + · · · + vk = 0

are v1 = · · · = vk = 0.

A sum of independent subspaces is called a direct sum, and one generally signifies a
direct sum by writing H1 ⊕ · · · ⊕ Hk instead of H1 + · · · + Hk . Independent subspaces are
like independent vectors, except that a collection of independent subspaces can (repeatedly!)
include the trivial subspace, whereas an independent set of vectors cannot include the zero
vector.

Exercise 5.1 Verify the following:

1. H1 , . . . , Hk are independent if and only if any v ∈ H1 + · · · + Hk can be uniquely
expressed v = v1 + · · · + vk with vj ∈ Hj .

2. A set of non-zero vectors {v1 , . . . , vk } ⊂ V is linearly independent if and only if

span v1 , . . . , span vk are independent subspaces.

3. Subspaces H, H 0 ⊂ V are complementary if and only if V = H ⊕ H 0 .

It will be useful to us below to be able to verify independence of a collection of subspaces

inductively. So to this end we prove

Proposition 5.2 Suppose that H1 , . . . , Hk−1 ⊂ V are independent subspaces and that Hk ⊂
V is another subspace that intersects H1 ⊕ · · · ⊕ Hk−1 trivially. Then H1 , . . . , Hk are inde-
pendent subspaces.

Proof. Suppose that vj ∈ Hj , 1 ≤ j ≤ k are given and

v1 + · · · + vk = 0.

Then
v1 + · · · + vk−1 = −vk
is a vector in (H1 ⊕ · · · ⊕ Hk−1 ) ∩ Hk . But this intersection is trivial by hypothesis, so both
sides of the last equation must be zero. In particular, independence of H1 , . . . , Hk−1 and the
vanishing of the left side imply that v1 = · · · = vk−1 = 0. Hence H1 , . . . , Hk are independent
subspaces. 
When k = 3, Proposition 5.2 says that (H1 ⊕H2 )⊕H3 = H1 ⊕H2 ⊕H3 , so ⊕ is associative
(and clearly also commutative).

13
Exercise 5.2 Let H1 , . . . , Hk ⊂ V be subspaces satisfying Hi ∩ Hj = {0} when i 6= j. Does
it follow that the subspaces are independent? Prove or give a counterexample.

A collection of subspaces whose direct sum is V is analogous to a basis for V . The next
proposition amplifies this analogy.

Proposition 5.3 Suppose that V = H1 ⊕ · · · ⊕ Hk and that for each 1 ≤ j ≤ k, we are

given a basis Bj for Hj (hence Hj is non-trivial). Then these bases are pairwise disjoint and
B := B1 ∪ · · · ∪ Bk is a basis for V .

Let us say that a basis B := B1 ∪ · · · ∪ Bk as in the statement of this proposition is

compatible with the decomposition V = H1 ⊕ · · · ⊕ Hk .
Proof. If two of the bases Bj and Bk have a vector v in common, then v ∈ Hj and −v ∈ Hk .
Since v + (−v) = 0, it follows that Hj and Hk are not independent, contrary to assumption.
So the bases are pairwise disjoint.
Clearly B spans each of the subspaces Hj , and since any vector in V can be written as a
sum of vectors in H1 , . . . , Hk , it follows that B spans V .
To see that B is independent, suppose that some S linear combination of vectors in B
vanishes. Taking advantage of the fact that B = Bj , we may express this assumption as
follows:
Xk X
cb b = 0.
j=1 b∈Bj
P P
Since b∈Bj cb b ∈ Vj , it follows from independence of the subspaces Vj that b∈Bj cb b = 0
for each j separately. Since Bj is independent, it then follows further that cb = 0 for each
b ∈ Bj . Thus all coefficients in the linear combination vanish, and we conclude that B is
independent. 

Corollary 5.4 If V = H1 ⊕ · · · ⊕ Hk , then dim V = dim H1 + · · · + dim Hk .

Proof. For each j ∈ {1, . . . , k}, let Bj be a basis for Hj . The previous proposition gives
X X
dim V = # ∪ Bj = #Bj = dim Hj .

The first equality follows from the fact that ∪Bj is a basis for V ; the second equality follows
from the fact that the bases Bj are mutually disjoint. 
Now we return to the linear transformation T : V → V introduced at the

Definition 5.5 A T -invariant decomposition of V is a collection of invariant subspaces Hj ,

1 ≤ j ≤ k such that V = kj=1 Hj . The decomposition is proper if Hj 6= V for any j.
L

Theorem 5.6 (Block diagonalization) Let V = H1 ⊕· · ·⊕Hk be a T -invariant decompo-

sition of V into non-trivial subspaces Hj , and let B = B1 ∪· · ·∪Bk be a compatible basis for V .
For each j, let Ajj be the matrix relative to Bj of the restricted transformation T : Hj → Hj ,

14
and let pj be its characteristic polynomial. Then the matrix of T relative to B has block
diagonal form  
A11 0 ... 0
 0 A22 ... 0 
.
 
 ..
 . 
0 0 ... Akk
In particular the characteristic polynomial of T : V → V is p1 . . . pk .
Proof. This is most easily done by induction on the number k of subspaces in the decom-
position. If k = 1 there is nothing to prove. We treat the case k = 2 separately because
the induction step relies on it implicitly. In this case we have B = {v1 , . . . , vn } where
B1 = {v1 , . . . , vk } and B2 = {vk+1 , . . . , vn }. Hence for any v ∈ V , we have
 
[v]B1
[v]B = ,
[v]B2
where the top or bottom component vanishes if v ∈ H1 or v ∈ H2 , respectively. So since
H1 = span B1 and H2 = span B2 are T -invariant, the jth column of [T ]B is given by
   
[T vj ]B1 0
[T vj ]B = if 1 ≤ j ≤ k, and [T vj ]B = if k + 1 ≤ j ≤ n.
0 [T vj ]B2
Putting all the columns together gives
 
A1 0
[T ]B = ,
0 A2
so the assertion is proved when k = 2.
Supposing now that the assertion is proved when k = K − 1, I consider the case k = K.
I have V = H1 ⊕ H 0 where H 0 = H2 ⊕ · · · ⊕ HK is also an invariant subspace. So by the case
k = 2,  
A1 0
[T ]B = ,
0 A0
where A0 is the matrix for T : H 0 → H 0 relative to the basis B 0 = B2 ∪ · · · ∪ BK . And by the
induction hypothesis we further have
 
A22 0 ... 0
 0 A33 ... 0 
A0 =  ,
 
..
 . 
0 0 ... Akk
so the assertion for k = K follows immediately. 
With this theorem we can now better state the goal of these notes: find a proper invariant
decomposition V = H1 ⊕ · · · ⊕ Hk into non-trivial subspaces Hj that are as small as possible.
As the theorem indicates, this will allow us to find a matrix representing T that is as
‘diagonal’ as possible. Before moving on, though, it is worth recasting the content of this
section in more ‘functorial’ terms. We leave the reader to verify the following fact, which is
a more or less immediate consequence of definitons.

15
Proposition 5.7 Let H1 , . . . , Hk ⊂ V be subspaces. Then σ(v1 , . . . , vk ) = v1 + · · · + vk
defines a linear transformation σ : H1 × · · · × Hk → V . In addition,

• σ is injective if and only if H1 , . . . , Hk are independent.

• σ is surjective if and only if H1 + · · · + Hk = V .

Hence V = H1 ⊕ · · · ⊕ Hk if and only if σ is an isomorphism.

This proposition allows one to infer Corollary 5.4 and Proposition 5.3 from corresponding
facts about product vector spaces. Chasing definitions also allows one to give a functorial
(or ‘coordinate-free’) version of Theorem 5.6.

Proposition 5.8 Let H1 , . . . , Hk ⊂ V and σ be as in Proposition 5.7. Suppose additionally

that T : V → V is a linear operator. If each of the subspaces Hj is T -invariant, then there
is a well-defined linear operator S : H1 × · · · × Hk → H1 × · · · × Hk given by S(v1 , . . . , vk ) =
(T v1 , . . . , T vk ) and satisfying σ ◦ S = T ◦ σ. If V = H1 ⊕ · · · ⊕ Hk , then T = σ ◦ S ◦ σ −1 is
similar to S.

6 Primary Decomposition
Having dwelt on cyclic subspaces in order to prove Theorem 4.2, we return to our other
means of identifying invariant subspaces.

Definition 6.1 Let p ∈ F[x] be any irreducible polynomial. The primary subspace associ-
ated to p and T is
[
Hp := {v ∈ V : p(T )k v = 0 for some k ∈ N} = ker p(T )k
k∈N

We leave the reader to verify that

Proposition 6.2 Hp is a T -invariant subspace. It is non-trivial if and only if p|pmin .

The main goal of this section is to establish the following important connection between
factors of the minimal/characteristic polynomials and decompositions of V into invariant
subspaces.

Theorem 6.3 (Primary Decomposition Theorem) Let p1 , . . . , p` be the distinct irre-

ducible factors of pmin . Then we have a T -invariant decomposition

V = Hp1 ⊕ · · · ⊕ Hp`

of V into the corresponding primary subspaces. The minimal polynomial of the restriction
r
T |Hpj is the largest power of pj j dividing pmin , and the characteristic polynomial is the largest
m
power pj j dividing pchar . Hence dim Hpj = mj deg pj .

16
 In contrast with the cyclic decomposition theorem to be stated and proven later, the
primary decomposition of a linear operator is canonical, completely determined by the field
F, the vector space V and the linear operator T . Proving Theorem 6.3 requires a preliminary
result that is interesting all by itself.

Lemma 6.4 Suppose that p, q ∈ F[x] are relatively prime polynomials. Then

ker p(T )q(T ) = ker p(T ) ⊕ ker q(T ).

Proof. The hypothesis implies that there exist a, b ∈ F[x] such that ap + bq = 1. Hence if
v ∈ ker p(T ) ∩ ker q(T ), then

v = id(v) = (a(T )p(T ) + b(T )q(T ))v = a(T )p(T )v + b(T )q(T )v = a(T )0 + b(T )0 = 0.

This proves that ker p(T ) ∩ ker q(T ) = 0, i.e. that ker p(T ) and ker q(T ) are independent
subspaces.
If, moreover, v = u + w where u ∈ ker p(T ) and w ∈ ker q(T ), then

p(T )q(T )v = q(T )p(T )v + p(T )q(T )u = q(T )0 + p(T )0 = 0.

Hence ker p(T ) ⊕ ker q(T ) ⊂ ker p(T )q(T ).

Finally, suppose v ∈ ker p(T )q(T ). From the first paragraph, we have

v = id(v) = (a(T )p(T ) + b(T )q(T ))v = w + u

where w = a(T )p(T )v and u = b(T )q(T )v. Thus

q(T )w = q(T )a(T )p(T )v = a(T )p(T )q(T )v = a(T )0 = 0,

so w ∈ ker q(T ). Similarly, u ∈ ker p(T ). Hence v = w + u ∈ ker p(T ) ⊕ ker q(T ). We
conclude that ker p(T )q(T ) = ker p(T ) ⊕ ker q(T ), as desired. 
Proof of Theorem 6.3. The unique factorization theorem for polynomials allows us to
write pmin = pr11 . . . pr` ` , where the factors pj are distinct irreducible polynomials and the
m1 m`
multiplicities rj are
P positive integers. Theorem 4.2 tells us that pchar = p1 . . . p` , where
mj ≥ rj satisfy mj deg pj = dim V . We will prove our assertions by induction on the
number ` of distinct irreducible factors.
If ` = 1, there is nothing to do except invoke Theorem 4.2 to see that when pmin = pr11 ,
one also has pchar = pm 1 for some m1 ≥ r1 .
1

Assuming the theorem is true when ` = k − 1, we proceed to the case ` = k. The

polynomials p = pr11 . . . pr`−11
and q = pr` ` are relatively prime, so Lemma 6.4 tells us that

V = ker pmin (T ) = ker p(T ) ⊕ Hp` .

The minimal polynomial for T |Hp` divides pmin and is therefore equal to pr` for some r < r` .
To see that r = r` , Note that if v ∈ V is any vector, then v = u + w where u ∈ ker p(T )
and w ∈ ker pr` (T ). Thus

p(T )pr` (T )v = p(T )pr` (T )u + p(T )pr` (T )w = pr` (T )0 + p(T )0 = 0.

17
Hence r deg p` + deg p = deg pr` p ≥ deg pmin = r` deg p` + deg p, which means r ≥ r` , too. We
conclude that r = r` as desired.
One shows similarly, that p is the minimal polynomial of T |ker p(T ) . Our inductive hy-
pothesis therefore implies that
ker p(T ) = Hp1 ⊕ · · · ⊕ Hp`−1 ,
with the minimal and characteristic polynomials of T |Hpj as described in the Theorem.
Putting together the previous two paragraphs completes the inductive step and the proof.

It is instructive to consider the implications of Theorem 6.3 in the case where the under-
lying field F is C. Then the primary decomposition of the characteristic polynomial is given
by
pchar (x) = (x − λ1 )m1 . . . (x − λ` )m` .
Hence
V = ker(T − λ1 id)m1 ⊕ · · · ⊕ ker(T − λ` id)m` .
Now suppose that B = B1 ∪ · · · ∪ B` is a basis for V obtained by concatenating bases for each
the T -invariant subspaces ker(T − λj id)mj . Then by Corollary 5.6, we see that the matrix
for T relative to B has block diagonal form
 
A1 0 . . . 0
 0 A2 . . . 0 
A=
 
.. 
 . 
0 0 . . . Ak
where Aj is the matrix for T |ker(T −λj id)mj relative to Bj . In particular (Aj − λj I)mj = 0.
That is, Aj = λj I + Nj , where Nj is nilpotent (of order mj ). Reassembling we see that
   
λ1 I 0 . . . 0 N1 0 . . . 0
 0 λ2 I . . . 0   0 N2 . . . 0 
A= + ,
   
.. ..
 .   . 
0 0 . . . λ` I 0 0 . . . Nk
where corresponding blocks in the two matrices each have the same sizes. Hence A = S + N
where S is diagonal, N is nilpotent and S and N commute. If we also use S and N to denote
the linear operators on V given by these matrices, we arrive at
Theorem 6.5 (SN Decomposition) If T : V → V is a linear operator on a finite dimen-
sional complex vector space, the T = S + N , where S is diagonalizable, N is nilpotent, and
S and N commute.
This theorem is very useful for computing eA where A is a matrix with complex entries.
The theorem tells us that eA = eS P (N ) where eS is easily computed for diagonal S and P is
the Taylor polynomial for ex with degree one less than the order of the nilpotent matrix N .
The interested reader can verify a couple of further facts describing the relationship
between the factors in the primary decomposition and more general invariant subspaces of
V.

18
Exercise 6.1 Let pj be the irreducible factors of pmin and H ⊂ V be any invariant subspace.
Then

1. The primary decomposition of the restricted operator T |H is H = `j=1 H ∩ Hpj , where

L
some of the intersections might be trivial.

2. The primary decomposition of the induced operator T̃ on V /H is V /H = `j=1 H

L g
pj ,
where H̃pj := {ṽ : v ∈ Hpj } might be trivial for some j.

7 Cyclic Decomposition
In the wake of the Primary Decomposition Theorem, one might reasonably ask if we can do
better. That is, can we further decompose the primary subspaces Hpj into smaller invariant
subspaces. Certainly this is possible in some cases. Suppose for instance pj (x) = x − λ
appears as factor of pmin with multiplicity rj = 1. Then Hpj is just the eigenspace for the
eigenvalue λ. If, moreover, m := dim Hpj > 1, then any basis {v1 , . . . , vm } for Hpj gives us
a decomposition
Hpj = Hv1 ⊕ · · · ⊕ Hvm
into the one dimensional cyclic subspaces generated by the eigenvectors vi . Observe that
this decomposition is not uniquely determined, since a different choice of basis results in a
different decomposition. Nor are the subspaces Hvj realizable as ker p(T ) for some p ∈ F[x].
After all, p would have to divide the minimal polynomial for T |Hpj which is x − λ, which
implies p(x) = x − λ.

Definition 7.1 Suppose that v1 , . . . , vk ∈ V are vectors such that

• V = kj=1 Hvj
L

s
• for each j, there is a positive integer sj and an irreducible pj ∈ F[x] such that pvj = pj j .

Then we say that the subspaces Hvj give a cyclic decomposition of V relative to T . We call
the subspaces Hvj the factors in the decomposition.

The requirement that pvj is a prime power implies that Hvj ⊂ Hpj is contained in a
primary subspace. This turns out to guarantee that the decomposition in the definition is
as fine as possible.

Definition 7.2 An invariant subspace H ⊂ V is irreducible if there is no proper invariant

decomposition H = H1 ⊕ H2 .

Proposition 7.3 For any v ∈ V , the subspace Hv is irreducible if and only if pv = ps for
p ∈ F[x] irreducible and s ≥ 0.

19
Proof. If pv is not a prime power, then we can write pv = pq, where p and q are non-constant
relatively prime polynomials. Lemma 6.4 then gives us an invariant decomposition

Hv = (ker p(T ) ∩ Hv ) ⊕ (ker q(T ) ∩ Hv ).

Since pv is the minimal polynomial associated to T |Hv , it follows that neither factor in the
decomposition is trivial, and therefore neither equals H. So H is reducible.
Suppose instead that pv = ps is a prime power. If H1 , H2 ⊂ Hv are invariant subspaces,
then the minimal polynomials pj of T |Hj must divide ps ; i.e. pj = psj for sj ≤ s. Since pv
is also the characteristic polynomial of T |Hv , we have dim Hv = s deg p. So if H1 , H2 6= Hv ,
we see that sj deg p ≤ dim Hj < dim Hv = s deg p; i.e. s0 = max{s1 , s2 } < s. Now if
H = H1 + H2 and w ∈ H is any vector, we write w = w1 + w2 where wj ∈ Hj and see that
0 0 0
ps (T )w = ps (T )w1 + ps (T )w2 = 0 + 0.
0
That is, ps (T ) = 0 on Hv , contradicting the fact that ps is the minimal polynomial of T |H .
We conclude that Hv is irreducible. 

Theorem 7.4 (Cyclic decomposition theorem) Any linear operator T : V → V on a

finite dimensional vector space V admits a cyclic decomposition. Any two cyclic decomposi-
tions
Mk Mk
Hvj = V = Hwj
j=1 j=1

of V relative to T are equivalent (after reordering factors if necessary) in the sense that
each has the same number of factors and there is an isomorphism φ : V → V such that
φ(vj ) = wj and φ ◦ T = T ◦ φ. Hence φ(Hvj ) = Hwj and pvj = pwj .

The rest of this section will be devoted to the proof of Theorem 7.4. The proof requires
some warm-up discussion. Since each subspace Hvj in a cyclic decomposition lies inside some
primary subspace Hpj , and since Theorem 6.3 tells us that V is a direct sum of primary
subspaces, it suffices to work with the restriction of T to a single primary subspace. That is,
we assume henceforth that the minimal polynomial for T has the form pmin = pr for some
r ≥ 1, and therefore V is equal to the primary subspace associated to p. We call r the order
of T .
From here, there are two central ideas behind our arguments. The first is that since
p(T ) is nilpotent on V , it is easier to work as much as possible in terms of p(T ) rather than
T . The second is to work inductively, reducing the order of T by considering the induced
operator T̃ : Ṽ → Ṽ where Ṽ := V / ker p(T ). The reader should check definitions to verify
that the order of T̃ is r − 1. The remainder of our preliminary discussion is aimed at making
it possible for us to use the two main ideas laid out here.
Let v ∈ V be any vector. The minimal polynomial pv of T |Hv must divide pmin ; hence
pv = ps for some s ≤ r. That is, s is the smallest non-negative integer such that p(T )s v = 0.
We call s the order of v. Observe that s is positive if and only if v 6= 0, that s ≥ 1 if and
only if v 6∈ ker p(T ), and that the order of p(T )k s is s − k for all 0 ≤ k ≤ s.

20
Lemma 7.5 For any v ∈ V , the cyclic subspace Hv is invariant under p(T ). If v has order
s ≥ 1, then
p(T )s−1 Hv = Hv ∩ ker p(T ) = Hp(T )s−1 v .
In particular, if s ≥ 2, then any w ∈ Hv ∩ ker p(T ) can be written w = p(T )u for some
u ∈ Hv − ker p(T ).

Proof. The first item in Proposition 2.8 tells us that Hv is p(T )-invariant. Since Hv ⊂
ker p(T )s and p(T )s−1 Hv is spanned by vectors p(T )s−1 T j v, j ≥ 0, we have Hp(T )s−1 v =
p(T )s−1 Hv ⊂ ker p(T ) ∩ Hv . To establish the remaining inclusion, let w ∈ Hv ∩ ker p(T ) be
any vector. By defintion of Hv , we have w = q(T )v for some q ∈ F[x]. We can therefore
use p(T )w = p(T )q(T )v = 0 to infer that ps |qp, i.e. q = aps−1 for some a ∈ F[x]. Finally,
taking u = a(T )v ∈ Hv , we see that w = p(T )s−1 u ∈ p(T )s−1 Hv . We conclude that
Hv ∩ ker p(T )s−1 = p(T )Hv as desired. 
The next three lemmas will allow us to prove the existence of a cyclic decomposition for
V . Given a subspace H ⊂ V , we set H̃ := {ṽ ∈ V / ker p(T ) : v ∈ H} be the image of H in
the Ṽ under the quotient map. Note in particular that H fv = Hṽ for any v ∈ V .

Lemma 7.6 Given vectors v1 , . . . , vk ∈ V − ker p(T ), the associated cyclic subspaces Hvj
are independent if and only if their images Hvj ⊂ Ṽ under the quotient map are independent.
g

Proof. We leave the reader to check that if Hvj are independent in V , then H g vj are
independent in Ṽ . For this, one can even allow ṽj ∈ ker p(T ).
Assume instead then that the subspaces H vj are independent in Ṽ . Given wj ∈ Hvj
g
P P
satisfying wj = 0, it follows that w̃j = 0̃. Thus w̃j = 0̃ for each j, i.e. wj ∈ ker p(T ).
By Lemma 7.5 we may write wj = p(T )uj where uj ∈ Hvj − ker p(T ). The fact that
P P
wj = 0 may then be recast as ũj = 0̃. Independence of the subspaces H gvj then gives
uj = ker p(T ) for each j, too; hence wj = p(T )uj = 0. This proves that the subspaces Hvj
are independent in V . 

Lemma 7.7 Let H ⊂ V be any invariant subspace. Then there exist vectors vj ∈ V ,
1 ≤ j ≤ k such that

ker p(T ) = (H ∩ ker p(T )) ⊕ Hv1 ⊕ · · · ⊕ Hvk .

Proof. If ker p(T ) ⊂ H, there is nothing to prove. So suppose we have a non-zero v ∈

ker p(T ) − H. We claim that Hv ∩ H = {0} is trivial. Hence Hv and H are independent
subspaces, and either H ⊕ Hv = ker p(T ), or we can repeat this process with H ⊕ Hv in
place of H. Since dim(H ⊕ Hv ) ∩ ker p(T ) = (dim H ⊕ ker p(T )) + deg p increases with every
repetition, we will eventually arrive at the desired sequence of vectors v1 , . . . , vk ∈ ker p(T ).
To establish the claim, let W := H ∩ Hv . As the intersection of invariant subspaces
W is itself an invariant subspace of Hv . Let S denote the operator on Hv /W induced
by the restriction T |Hv . The characteristic polynomial of T |Hv is p, and the characteristic
polynomial q of S must therefore divide p. Since p is irreducible, it follows that q = 1 or
q = p. In the first case, we infer that dim Hv /W = 0 which means that Hv = W ⊂ H. This

21
contradicts the fact that v ∈ Hv − W . In the second case, we infer dim Hv /W = deg p =
dim Hv and therefore W = Hv ∩ H is trivial as claimed. 
The next lemma really doesn’t have anything to do with operators at all, but we state it
only for the current context.

Lemma 7.8 Suppose that H ⊂ V is a subspace and H̃ = Ṽ and that ker p(T ) = (H ∩
ker p(T )) ⊕ W for some W ⊂ ker p(T ). Then V = H ⊕ W .

Proof. First note that since W ⊂ ker p(T ), we have that if H ∩ W = (ker p(T ) ∩ H) ∩ W is
trivial. So H and W are independent.
Given v ∈ V , the hypothesis that ṽ ∈ H̃ means that v = v0 + u where v0 ∈ H and
u ∈ ker p(T ). The hypothesis that W complements H ∩ ker p(T ) in ker p(T ) allows us to
further decompose u = v00 +w where v00 ∈ H and w ∈ W . Hence v = (v0 +v00 )+w ∈ H +W .
So V = H ⊕ W . 
The next (and last) lemma will help us prove that different cyclic decompositions of V
are equivalent.

Lemma 7.9 Let v, w ∈ V be vectors with the same order s. Then linear transformation
φ : Hv → Hw determined by φ(T j v) = T j w, 0 ≤ j < s deg p is an isomorphism satisfying
φ ◦ T = T ◦ φ.

Proof. The requirement that φ(T j v) = T j w for 0 ≤ j < s deg p amounts to saying that φ
carries the usual basis for Hv to the corresponding basis for Hw . There is exactly one linear
transformation accomplishing this task, and it must be an automorphism. So φ is a well-
defined isomorphism from Hv to Hw . We have by construction that φ ◦ T (T j v) = T ◦ φ(T j v)
for 0 ≤ j ≤ s − 2. The same holds for j = s − 1 because p(T )s v = p(T )s w = 0—i.e. pv = pw
means that the final basis vectors for Hv and Hw can be written as linear combinations of
their predecessors in exactly the same way. At any rate, since φ ◦ T = T ◦ φ for all vectors
in a basis for Hv , the same holds for all vectors in Hv . 
Proof of Theorem 7.4. As mentioned above, we will work by induction on the order s
of T . L
When s = 1, we have ker p(T ) = V . Hence the existence of the cyclic decomposition
k
V = j=1 Hvj follows immediately from Lemma 7.7, and the characteristic polynomials
pvj for T |Hvj are all equal to p. Hence the number k of factors in the decomposition is
dim V / deg p which is in particular independent of the choice of vj .
If V = kj=1 Hwj is another cyclic decomposition of V , Lemma 7.9 gives us isomorphisms
L
φj : Hvj → Hwj between corresponding factors. We assemble these into an isomorphism
φ : V → V by using that any v ∈ V decomposes uniquely as v = u1 + · · · + uk , uj ∈ Hvj
and setting
φ(u1 + · · · + uk ) = φ1 (u1 ) + · · · + φk (uk ).
This completes the case s = 1.
Assuming inductively that the theorem is true when the order of T is s − 1, we consider
the case when the order of T is s. By our inductive hypothesis, we ṽ1 , . . . , ṽi ∈ V such that

22
Ṽ = V / ker p(T ) = ij=1 H̃vj . We can apply Lemmas 7.7 and 7.8 to H = Hv1 + · · · + Hvk
L
to get vectors vi+1 , . . . , vk such that

V = H ⊕ Vi+1 ⊕ · · · ⊕ Vk .

By Lemma 7.6, the subspaces Hvj remain independent in V , so we have in fact that

V = ⊕kj=1 Hvj .

This completes the proof of existence of a cyclic decomposition.

Now if V = ⊕kj=1 Hwj is another cyclic decomposition, we order the factors in both
decompositions from largest to smallest. Then if i0 is the largest integer such that order of
wi0 exceeds 1, then we get a cyclic decomposition
0
Ṽ = ⊕ij=1 H̃wj

of the quotient space, independence of the factors proceeding from Lemma 7.6. Our inductive
hypothesis then implies that i0 = i and that the order of vj equals that of wj for each
1 ≤ j ≤ i. In fact the orders remain the same for i < j ≤ k, too since vj and wj both have
order 1 for all j > i. So finally we can repeat the argument for equivalence from the case
s = 1 to construct an equivalence isomorphism φ : V → V carrying Hvj to Hwj for all j. 

8 Jordan Canonical Form

Let us apply the cyclic decomposition theorem to the case of a primary subspace H = Hx−λ
assoicated to a polynomial p(x) = x − λ of degree one. Replacing T with S := T − λid, we
have that S|H is nilpotent of some order r. We may therefore decompose H into a direct sum
of finitely many S-cyclic subspaces W = span B, where B = {v, Sv, . . . , S k−1 v} for some
v ∈ V , and S k v = 0. In matrix terms, this becomes
 
0 0 0 ... 0 0
1 0 0 . . . 0 0
 
0 1 0 . . . 0 0
[S|W ]BB = 0 0 1 . . . 0 0 .
 
 
 .. 
 . 
0 0 0 ... 1 0

The subspaces W will also be invariant (in fact cyclic) by T = S + λid, and we see that
 
λ 0 0 ... 0 0
1 λ 0 . . . 0 0
 
0 1 λ . . . 0 0
[T |W ]BB =  0 .
 
 0 1 . . . 0 0 
 .. 
 . 
0 0 0 ... 1 λ

23
Note that if we reverse the order of the vectors in B, then the matrix becomes upper trian-
gular, with ‘1’s above rather than below the main diagonal. This is the more conventional
way to present the matrix, but for the sake of internal consistency, we stick with the lower
triangular version here.
In the case F = C, when all irreducible factors of pmin have degree one, we can apply
these observations to every primary subspace and conclude as follows.

Theorem 8.1 (Jordan Canonical Form) Let T : V → V be a linear operator on a com-

plex vector space V . Then there is a basis B ⊂ V such that [T ]BB has block diagonal form
 
A1

 A2 


 . . .


Ak

where for each i, there exists an eigenvalue λi of T such that

 
λi 0 0 ... 0 0
 1 λi 0 ... 0 0
 
0 1 λi . . . 0 0 
Ai =  0 0 .
 
 1 ... 0 0 
 .. 
 . 
0 0 0 . . . 1 λi

Any two such bases for V give rise to similar such matrices, having the same blocks, albeit
possibly in different order.

24