1 COMPUTING THE JORDAN FORM OF A MATRIX 1

Computing the Jordan Form of a Matrix

1 Computing the Jordan Form of a Matrix
Let T be a linear operator on a finite dimensional vector space V over a field F . Suppose that
the characteristic polynomial of T is a product of linear polynomials in F [x]. We proved in the
last lecture that there exists a basis for V such that the matrix representing T with respect to
this basis is in Jordan form. This result has an immediate corollary for matrices:

Corollary 1. Let A ∈ Mn×n (F ) be a matrix whose characteristic polynomial is a product of

linear polynomials in F [x]. Then there exists a matrix J in Jordan form and an invertible matrix
P such that
A = P JP −1

Definition 1. The matrix J in the above corollary is often referred to as the Jordan canonical
form of the matrix A.

These results tell us in a sense what we can do in the cases when A (or T ) is not diagonali-
zable, therefore generalize our previous discussion. One important comment is that the condition
for the characteristic polynomial to be a product of linear factors is automatically satisfied in
an algebraically closed field, for instance when F = C.
One question that we did not address is the uniqueness of Jordan form: Uniqueness actually
does hold, except for an obvious possibility of changing the places of Jordan blocks. We will
skip the proof. There are no further technical obstacles to prove it though; it can be left as a
challenging exercise at this point.
The proof of existence of Jordan form was somewhat opaque, in the sense that it is tricky to
follow the details for an actual computation. Here, we will solve several examples in order to see
how Jordan canonical forms can actually be computed. Instead of giving a full algorithm that
covers all possibilities (which is possible but involves some elaborate combinatorics) we will try
to select enough examples which will allow the reader to guess the pattern to a certain extent.

Example: Find the Jordan canonical form of the matrix

 
0 −4
A=
1 4

in M2×2 (R).
Let us find the characteristic polynomial of A first:

∆A (x) = det(xI − A) = x2 − 4x + 4 = (x − 2)2 .

Therefore the characteristic polynomial is a product of linear factors in R[x], so the matrix has
a real Jordan canonical form. We see that the only eigenvalue is λ = 2. At this point, without
any further calculation, there are only two possibilities for the Jordan form of a 2 × 2 matrix
with only a single eigenvalue λ = 2:
   
2 0 2 1
J1 = , J2 =
0 2 0 2
1 COMPUTING THE JORDAN FORM OF A MATRIX 2

Which of these possibilities is actually the Jordan form of A? We claim that it must be J2 .
Indeed, suppose that it is J1 . But we see that J1 = 2I. So if A = P J1 P −1 , then A = P (2I)P −1 =
2I. But A 6= 2I, contradiction. Therefore this cannot be the case.
At this point, we decided that the Jordan form of A is J2 . But this doesn’t yet give us a
matrix P such that A = P J2 P −1 . To find such a P , suppose that
 
P = v1 |v2

where v1 and v2 are column vectors of P . Let us look at the equation AP = P J2 in detail:

AP = P J2
 
    2 1
A v1 |v2 = v1 |v2
0 2
   
Av1 |Av2 = 2v1 |v1 + 2v2

Therefore v1 and v2 must be solutions of the two linear systems

Av1 = 2v1 , Av2 = v1 + 2v2

The first equation tells us that v1 must be an eigenvector of A. The second vector v2 is not
an eigenvector, but since the equation is resembling an eigenvector equation, sometimes such a
vector is called a generalized eigenvector.
Find v1 by solving (A − 2I)v1 = 0:
     
−2 −4 0 −2k
v1 = ⇒ v1 = ,k ∈ R
1 2 0 k
Select a non-zero value of k in order to get a first column for P (it must be non-zero since P
needs to be invertible). For instance choose k = 1 so that
 
−2
v1 =
1
Next, find v2 by solving (A − 2I)v2 = v1 :
       
−2 −4 −2 1 −2
v2 = ⇒ v2 = +` , ` ∈ R.
1 2 1 0 1
Select a value of ` in order to get a second column for P . This time, ` does not have to be
nonzero since the resulting matrix P will be invertible regardless of the choice of ` (please check
this). Select for instance, ` = 0 (any other choice would be equally valid). So we get
 
1
v2 =
0
Now we have the matrix P :  
−2 1
P =
1 0
Finally, the equality A = P J2 P −1 looks like
     −1
0 −4 −2 1 2 1 −2 1
=
1 4 1 0 0 2 1 0
1 COMPUTING THE JORDAN FORM OF A MATRIX 3

We remark that the Jordan canonical form J2 is unique, although the matrix P is far from
being unique; we had certain choices for k and ` along the way.

Example: Find the Jordan canonical form of the matrix

 
2 1 2
A = 0 0 2
0 1 1

in M3×3 (F3 ) where F3 is the finite field with 3 elements. Let us first find the characteristic
polynomial of A.
x−2 2 1
∆A (x) = 0 x 1 = (x − 2)3 .
0 2 x−1
(Remember that we are working over the field F3 .) There are three possibilities for the Jordan
form (up to the ordering of Jordan blocks):
     
2 0 0 2 1 0 2 1 0
J1 = 0 2 0 ,
  J2 = 0 2
 0 , J3 = 0 2 1
0 0 2 0 0 2 0 0 2

Which of these is actually the Jordan form? Let us try to find the minimal polynomial δA (x)
of A. We know that the minimal polynomial must be one of the three polynomials

(x − 2), (x − 2)2 , (x − 2)3 .

Similar matrices have the same minimal polynomial, therefore A and its Jordan canonical form
should have the same minimal polynomial. This should give us a clue. Now, first of all it is clear
that A − 2I 6= 0 therefore δA (x) 6= (x − 2). Let us try (x − 2)2 :
 2  
0 1 2 0 0 0
(A − 2I)2 = 0 1 2 = 0 0 0
0 1 2 0 0 0

therefore we deduce that δA (x) = (x−2)2 . It can be quickly checked that the minimal polynomial
of J1 is (x − 2), the minimal polynomial of J2 is (x − 2)2 and the minimal polynomial of J3 is
(x − 2)3 . We deduce that the Jordan canonical form of A is J2 .  
Let us now try to find an invertible matrix P such that AP = P J2 . Say P = v1 |v2 |v3
where v1 , v2 , v3 are the columns of P . Writing the equation AP = P J2 , we get

AP = P J2
 
    2 1 0
A v1 |v2 |v3 = v1 |v2 |v3 0 2 0
0 0 2
   
Av1 |Av2 |Av3 = 2v1 |v1 + 2v2 |2v3

Therefore, v1 , v2 , v3 must be the solutions of

Av1 = 2v1 , Av2 = v1 + 2v2 , Av3 = 2v3

1 COMPUTING THE JORDAN FORM OF A MATRIX 4

So, v1 and v3 must be eigenvectors of A and v2 must be a generalized eigenvector. The eigenvector
equation (A − 2I)v = 0 gives us
       
0 1 2 0 1 0
0 1 2 v = 0 ⇒ v = s 0 + t 1 , s, t ∈ F3
  
0 1 2 0 0 1
where s and t are not both 0. We have to be careful now: In order to find v2 , we would like to
solve the equation (A − 2I)v2 = v1 . But we have some freedom for choosing v1 and it is unclear
which choices would lead to a solution for v2 and not leave us with an inconsistent linear system.
So, at this point, let us write the form of the most general eigenvector in place of v1 :
   
0 1 2 s
0 1 2 v2 =  t  .
0 1 2 t
It is easy to see that this system has a solution if and only if s = t, therefore we must select the
eigenvector v1 accordingly. Set  
1
v1 = 1

1
Then select a solution v2 of the linear system above. For instance, we can take
 
0
v2 = 1
0
What about v3 ? All that we want is it should be an eigenvector and {v1 , v3 } should be a basis
for the eigenspace W2 . There are many choices. Take for instance by setting s = 1 and t = 0:
 
1
v3 = 0
0
This now gives us the matrix P . We have
 
1 0 1
P = 1 1 0 .
1 0 0
The equation A = P JP −1 , explicitly written, looks like
   −1
1 0 1 2 1 0 1 0 1
A = 1 1 0 0 2 0 1 1 0
1 0 0 0 0 2 1 0 0

Example: Find the Jordan canonical form of the following matrix in M6×6 (R):
 
1 0 0 0 1 0
0 1
 1 1 0 −1 
0 0 1 0 0 0
A= 0 0

 0 1 0 0 
0 0 1 0 1 −1
0 0 0 0 0 1
1 COMPUTING THE JORDAN FORM OF A MATRIX 5

Let us start by finding its characteristic polynomial:

x−1 0 0 0 −1 0
0 x − 1 −1 −1 0 1
0 0 x−1 0 0 0
∆A (x) = = (x − 1)6
0 0 0 x−1 0 0
0 0 −1 0 x−1 1
0 0 0 0 0 x−1

(easy way to compute this determinant: Observe that only the identity permutation contributes
a non-zero term.) Hence we have a six-fold repeated eigenvalue, λ = 1. We have quite a lot of
possibilities for the Jordan form (up to ordering of Jordan blocks):
     
1 0 0 0 0 0 1 1 0 0 0 0 1 1 0 0 0 0
0 1 0 0 0 0 0 1 0 0 0 0 0 1 0 0 0 0
     
0 0 1 0 0 0 0 0 1 0 0 0 0 0 1 1 0 0
J1 = 
0
, J2 =  , J3 =  
 0 0 1 0 0
0
 0 0 1 0 0
0
 0 0 1 0 0
0 0 0 0 1 0 0 0 0 0 1 0 0 0 0 0 1 0
0 0 0 0 0 1 0 0 0 0 0 1 0 0 0 0 0 1
     
1 1 0 0 0 0 1 1 0 0 0 0 1 1 0 0 0 0
0 1 0 0 0 0 0 1 1 0 0 0 0 1 1 0 0 0
     
0 0 1 1 0 0 0 0 1 0 0 0 0 0 1 0 0 0
J4 = 
0
, J5 =  , J6 =  
 0 0 1 0 0
0
 0 0 1 0 0
0
 0 0 1 1 0
0 0 0 0 1 1 0 0 0 0 1 0 0 0 0 0 1 0
0 0 0 0 0 1 0 0 0 0 0 1 0 0 0 0 0 1
     
1 1 0 0 0 0 1 1 0 0 0 0 1 1 0 0 0 0
0 1 1 0 0 0 0 1 1 0 0 0 0 1 1 0 0 0
     
0 0 1 0 0 0 0 0 1 1 0 0 0 0 1 1 0 0
J7 = 
0
, J8 =  , J9 =  ,
 0 0 1 1 0
0
 0 0 1 0 0
0
 0 0 1 0 0
0 0 0 0 1 1 0 0 0 0 1 0 0 0 0 0 1 1
0 0 0 0 0 1 0 0 0 0 0 1 0 0 0 0 0 1
   
1 1 0 0 0 0 1 1 0 0 0 0
0 1 1 0 0 0 0 1 1 0 0 0
   
0 0 1 1 0 0 0 0 1 1 0 0
J10 = 
0
, J11 =  
 0 0 1 1 0
0
 0 0 1 1 0
0 0 0 0 1 0 0 0 0 0 1 1
0 0 0 0 0 1 0 0 0 0 0 1

We notice that there is a bijection between the possible Jordan forms for a given eigenvalue
and all integer partitions of the matrix size. In particular, the matrices above correspond to
the partitions of the number 6 in the following order: 1+1+1+1+1+1, 2+1+1+1+1, 2+2+1+1,
2+2+2, 3+1+1+1, 3+2+1, 3+3, 4+1+1, 4+2, 5+1, 6.
Which of these is the actual Jordan canonical form? Let us try to find the minimal polynomial
δA (x). It must be one of the polynomials

(x − 1), (x − 1)2 , (x − 1)3 , (x − 1)4 , (x − 1)5 , (x − 1)6

1 COMPUTING THE JORDAN FORM OF A MATRIX 6

It is certainly not x − 1, since otherwise we would have A = I, but this is not true.
2 
0 0 1 0 0 −1
 
0 0 0 0 1 0
0 0 1 1 0 −1 0 0 0 0 0 0 
   
2
0 0 0 0 0 0  0 0 0 0 0 0 
(A − I) =  
 = 0 0 0 0 0 0 
  
0 0 0 0 0 0   
0 0 1 0 0 −1 0 0 0 0 0 0 
0 0 0 0 0 0 0 0 0 0 0 0

In particular (A − I)2 6= 0 hence the minimal polynomial of A cannot be (x − 1)2 . But by one
more similar computation we immediately see that (A − I)3 = 0. Therefore,

δA (x) = (x − 1)3 .

Let us now check which Ji ’s have this minimal polynomial, knowing that A and its Jordan
canonical form must have the same minimal polynomial. The minimal polynomial of J1 is x − 1,
the minimal polynomials of J2 , J3 and J4 are (x − 1)2 , the minimal polynomial of J5 , J6 and J7
are (x − 1)3 , the minimal polynomials of J8 and J9 are (x − 1)4 , the minimal polynomial of J10
is (x − 1)5 and the minimal polynomial of J11 is (x − 1)6 . One can compute all of these directly,
but the general rule is that the power of (x − 1) in the minimal polynomial will be equal to
the the largest integer in the partition corresponding to the Jordan form. This is not difficult
to generalize and prove.
We deduce that the Jordan canonical form of A must be one of J5 , J6 or J7 . How can we
distinguish between these three forms? Notice that if J is the Jordan form of A, then the matrix
J − I will be similar to the matrix A − I, therefore the vector spaces ker(J − I) and ker(A − I)
will have the same dimension, namely this tells us that we could distinguish between these
possibilities by looking at the dimension of the eigenspace for λ = 1. Looking at the equation
(A − I)v = 0, we get
   
0 0 0 0 1 0 0
0 0 1 1 0 −1 0
   
0 0 0 0 0 0 
 v = 0
 

0 0 0 0 0 0  0
   
0 0 1 0 0 −1 0
0 0 0 0 0 0 0
The coefficient matrix has rank 3, hence it also has nullity equal to 3 by the rank-nullity theorem.
Therefore the eigenspace W1 has dimension 3. Now, observe directly that

rank(J5 − I) = 2, rank(J6 − I) = 3, rank(J7 − I) = 4.

By using the rank-nullity theorem for each of these matrices, we get

nullity(J5 − I) = 4, nullity(J6 − I) = 3, nullity(J7 − I) = 2.

Since the correct number is given only by J6 , we deduce that the Jordan canonical form of A
must be J6 .
The diligent reader will be probably worried about the following possibility: What if the
matrices were larger (or there were more eigenvalues) and there were still more possibilities
remaining after comparing these dimensions? One can continue as follows: Look at the dimensi-
ons of ker((A − λI)2 ), ker((A − λI)3 ), . . . for each eigenvalue until only one candidate remains.
1 COMPUTING THE JORDAN FORM OF A MATRIX 7

It can be proven that these dimensions eventually characterize the whole partition and we can
fully retrieve the Jordan form.
Let us alsofind an invertible  matrix P such that AP = P J6 . Let us write P in terms of its
columns: P = v1 |v2 |v3 |v4 |v5 |v6 . Then,

AP = P J6
 
1 1 0 0 0 0
0 1 1 0 0 0
 
    0 0 1 0 0 0
A v1 |v2 |v3 |v4 |v5 |v6 = v1 |v2 |v3 |v4 |v5 |v6 0 0 0 1 1

 0
0 0 0 0 1 0
0 0 0 0 0 1
   
Av1 |Av2 |Av3 |Av4 |Av5 |Av6 = v1 |v1 + v2 |v2 + v3 |v4 |v4 + v5 |v6

Therefore, we have the following linear systems:

Av1 = v1 , Av2 = v1 + v2 , Av3 = v2 + v3 , Av4 = v4 , Av5 = v4 + v5 , Av6 = v6

Therefore, v1 , v4 and v6 will be eigenvectors, the others will be certain generalized eigenvectors.
Let us find all eigenvectors of A by solving (A − I)v = 0:
         
0 0 0 0 1 0 0 1 0 0
0 0 1 1 0 −1 0 0 1 0
         
0 0 0 0 0 0 
 v = 0
  0 0 1

0 0 0 0 0 0  0 ⇒ v = r  + s  + t
    
    0
  0
 
0
 
0 0 1 0 0 −1 0 0 0 0
0 0 0 0 0 0 0 0 0 1
Now, in order to find v2 and v5 we need to solve the equation (A − I)w = v where v is an
eigenvector. This linear system looks like
   
0 0 0 0 1 0 r
0 0 1
 1 0 −1

s
 
0 0 0 0 0 0 w = t
 

0 0 0 0 0 0 0
  
0 0 1 0 0 −1 0
0 0 0 0 0 0 t

This system is consistent if and only if t = 0, and its solutions are of the form
 
a
b
 
c
w= s .

 
r 
c
1 COMPUTING THE JORDAN FORM OF A MATRIX 8

Finally, in order to find v3 we need to solve (A − I)u = w where w is as above. This reads
   
0 0 0 0 1 0 a
0
 0 1 1 0 −1

b
 
0 0 0 0 0 0 u = c .
 

0 0 0 0 0 0 s
  
0 0 1 0 0 −1 r 
0 0 0 0 0 0 c

This system is consistent if and only if c = s = 0.

To sum up: The “deepest” vector is v3 . In order to make all steps along the way solvable, we
need to select the eigenvector v1 such that s = t = 0. So, for instance we can take r = 1, s = t = 0
and get  
1
0
 
0
v1 = 
0

 
0
0
The vector v2 must satisfy (A − I)v2 = v1 and furthermore we must have c = 0 in the resulting
solution. One solution is (by taking a = b = c = 0)
 
0
0
 
0
 
0
v2 =  
0
 
1
0

Then, we find v3 by solving (A − I)v3 = v2 . One solution is

 
0
0
 
1
v3 = 
−1

 
0
0

Next, let us select v4 . The restrictions on it are: It must be an eigenvector independent of v1 ,

and t = 0 must be satisfied so that later we can solve for v5 . So we could take r = 0, s = 1 and
t = 0 in order to get  
0
1
 
0
v4 = 0

 
0
0
1 COMPUTING THE JORDAN FORM OF A MATRIX 9

Next, v5 must be a solution of (A − I)v5 = v4 . One solution is (by taking a = b = c = 0)

 
0
0
 
0
v5 = 
1 .

 
0
0

Finally, v6 must be an eigenvector independent of v1 , v4 . A simple way to get it is to set r = s = 0

and t = 1:  
0
0
 
1
v6 = 
0

 
0
1
At this point, we are ready to write P and finish the problem:
 
1 0 0 0 0 0
0 0 0 1 0 0
 
0 0 1 0 0 1
P = 0 0 −1 0

 1 0
0 1 0 0 0 0
0 0 0 0 0 1