NOTES ON JORDAN FORM

SANTIAGO CAÑEZ

These notes are meant to clarify the notion of a Jordan form which the book talks about
in the last section of Chapter 8. In particular, we give the actual definition of a Jordan form,
which the book never quite gets to explicitly (although they do it in disguise!).
We will work with a finite-dimensional complex vector space V .
Definition 1. A Jordan block of size k is a k × k matrix of the form
λ 1
 
 λ 1 
 .. .. 

 . . 

 .. 
 . 1
λ
where the missing entries are all zero. In other words, a Jordan block is almost a multiple of
the identity, except for 1’s above the main diagonal.
Definition 2. A square matrix is said to be in Jordan form if it is block diagonal and each
block is a Jordan block.
This is exactly the type of matrix described in the book near the top of page 186. With
this terminology, Theorem 8.47 can then be restated as follows:
Theorem 1. Any operator T on V can be represented by a matrix in Jordan form. This
matrix is unique up to a rearrangement of the order of the Jordan blocks, and is called the
Jordan form of T .
A basis of V which puts M(T ) in Jordan form is called a Jordan basis for T . This last
section of Chapter 8 is all about proving the above theorem. The key step is Lemma 8.40,
which shows that Jordan bases exist for nilpotent operators. The above theorem then follows
by applying the lemma to the nilpotent operators which are the restrictions of T − λi I to the
generalized eigenspaces of T corresponding to the eigenvalues λi .
As you can see when reading this last section, the proof of this lemma is quite difficult.
Indeed, most people probably never actually see a proof, which is okay because the most
important thing is the existence of the Jordan form and what it tells you about the operator.
This is what we will clarify in these notes.
If v is a nonzero generalized eigenvector of T corresponding to the eigenvalue λ, then there
exists a smallest positive integer k so that
(T − λI)k v = 0.
The list (v, (T − λI)v, . . . , (T − λI)k−1 v) is then linearly independent by the problem on the
last quiz. Now we see what the point of that problem was!
Definition 3. For a nonzero generalized eigenvector of T corresponding to an eigenvalue λ,
the list
(v, (T − λI)v, . . . , (T − λI)k−1 v),
2 SANTIAGO CAÑEZ

where k is defined as above, is called the Jordan chain corresponding to v.

Note that for such a Jordan chain, the last vector is a usual eigenvector of T since applying
T − λI to it gives zero.
A Jordan basis is then exactly a basis of V which is composed of Jordan chains. Lemma
8.40 exactly says that such a basis exists for nilpotent operators, which then implies that
such a basis exists for any T as in Theorem 8.47. Each Jordan block in the Jordan form of
T corresponds to exactly one such Jordan chain. Indeed, the point of Theorem 8.47 is that
the matrix of an operator corresponding to a specific Jordan chain written in reverse order is
exactly a Jordan block. This is how we get a matrix M(T ) which is block diagonal and where
each block is a Jordan block.
Let λ be a specific eigenvalue of T . Then we know there is a basis for the generalized
eigenspace of T corresponding to λ consisting of Jordan chains. How many such Jordan chains
are there? The last vectors in these Jordan chains are eigenvectors corresponding to λ, and
they in fact give a basis for the eigenspace corresponding to λ. In other words:
Proposition 1. The number of Jordan blocks in the Jordan form of T corresponding to λ is
the dimension of the (usual) eigenspace of T corresponding to λ.
The problem we are interested in now is finding the Jordan form of an operator. The
characteristic polynomial tells us how many times a certain eigenvalue will appear in this
Jordan form, and the dimension of each eigenspace tells us how many Jordan blocks there will
be for a specific eigenvalue. What about the size of the blocks themselves? The important
fact here is the following:
Proposition 2. The size of the largest Jordan block corresponding to an eigenvalue λ of T is
exactly the degree of the (z − λ) term in the minimal polynomial of T .
We have essentially already seen this in class: to kill off a Jordan block of size k, say of the
form given in the first definition in these notes, we need the polynomial (z − λ)k . Any smaller
Jordan block with the same eigenvalue will also be send to zero under this polynomial, and
this implies the result of the proposition.
Let us summarize what we have learned so far about the relation between the Jordan form
of an operator, its characteristic polynomial, and its minimal polynomial:
• The characteristic polynomial tells us the eigenvalues and the dimension of each gen-
eralized eigenspace, i.e. the number of times the eigenvalue appears along the diagonal
of the Jordan form.
• The dimension of each eigenspace tells us how many Jordan blocks there are in the
Jordan form corresponding to each eigenvalue.
• The exponents of the different terms in the minimal polynomial tell us the size of the
largest Jordan blocks corresponding to each eigenvalue.
In the best situation, these facts give us all the information we need to explicitly find the
Jordan form of an operator. We will look at some examples in class — also check problems
12 and 14 on the practice final.
Finally, we say a word about the importance of the Jordan form and look at an application.
We know that an operator on a finite-dimensional complex vector space can sometimes be
represented by a diagonal matrix — i.e. exactly when the operator is diagonalizable. Note
that for a diagonalizable operator, the Jordan form itself only consists of 1 × 1 Jordan blocks
and is diagonal; this follows from the fact that for a diagonalizable operator, each generalized
eigenvectors is a usual eigenvector. For a general operator, we can always at least represent it
by an upper-triangular matrix. So the question is: what is the “best” possible matrix we can
 NOTES ON JORDAN FORM 3

use to represent a general operator? The answer is the Jordan form, which is the “best” such
matrix in the sense that it is very close to being diagonal except for a few ones above the main
diagonal. The point is that this special form still allows us to do many of the nice things we
can do with diagonal matrices.
To see an application, recall that for an operator T , we define the operator eT by
∞
T
X Tk
e =I+ .
k!
k=1

The motivation for this of course comes from the power series expansion of ex . In general this
may not be so easy to actually compute, but we will see the Jordan forms give us a nice way
of describing eT via a matrix.
If A is a square matrix, we have the same definition. If A is block diagonal:
 
A1
A=
 .. ,

.
Am

it turns out that Ak is also block-diagonal and

 
Ak1
Ak = 
 .. .

.
Akm

This implies that

 A 
e 1
A
e =
 .. .

.
e Am

So, if we can represent T by a block diagonal matrix, we can describe eT by computing the
exponential of each block. We know that if a block is diagonal, computing its exponential is
easy. The point is that it is also easy for Jordan blocks!
For simplicity, consider the 2 × 2 Jordan block
 
λ 1
J= .
0 λ

Then
 2  3
λ 3λ2
 
2 λ 2λ 3
J = , J = ,
0 λ2 0 λ3

and in general
 k
λ kλk−1

k
J = .
0 λk
4 SANTIAGO CAÑEZ

Thus
∞
J
X Jk
e =I+
k!
k=1
∞
1 λk kλk−1
  X  
1 0
= +
0 1 k! 0 λk
k=1
∞
!
  X λk λk−1
1 0 k! (k−1)!
= + λk
0 1 0
k=1 k!
P∞ λk P∞ λk−1 !
k=0 k! k=1
= P∞ (k−1)!
λk
0 k=0 k!
 λ λ
e e
= ,
0 eλ
Thus we know exactly what the expoential of a 2 × 2 Jordan block is. Similarly, one can
show that for a size k Jordan block, its exponential will be the upper-triangular matrix which
as eλ down the main diagonal, eλ /1! down the diagonal above that, eλ /2! down the diagonal
above that, eλ /3! down the diagonal above that, and so on. Thus, we know exactly what the
exponential of any Jordan block is, and hence exactly what the exponential of any Jordan
form is. So, we can always find a way to explicitly express eT via a matrix representation.
As one final comment, here is another use of Jordan forms. We know that two matrices
represent the same operator with respect to different bases if and only if they are similar. The
still unanswered question is: Is there a quick way to determine if two matrices are similar?
The answer is yes:
Theorem 2. Two square complex matrices are similar if and only if they have the same
Jordan form, up to a rearrangement of the Jordan blocks. In particular then, two square
complex matrices represent the same operator if and only if they have the same Jordan form.