Jordan Normal Form

Alastair Fletcher
January 5th 2003

Introduction

Any matrix over C (or any algebraically closed field, if that means anything
to you!) is similar to an upper triangular matrix, but not necessarily similar
to a diagonal matrix. Despite this we can still demand that it be similar to a
matrix which is as nice as possible, which is the Jordan Normal Form. This
has applications to systems of difference or differential equations, which can be
represented by matrices - putting the matrix in Jordan Normal Form makes it
easier to find solutions to the system of difference or differential equations. For
more on this, see [1].
This article gives the general method for finding the Jordan Normal Form
of 2x2 and 3x3 matrices with a view to answering examination questions. This
means that the article will discuss the cases for R2 and R3 and even though not
all such matrices can be put in upper triangular form (and hence not in Jordan
Normal Form), we will assume that the matrices we work with can be put in
upper triangular form. This condition is equivalent to being able to factorise the
characteristic polynomial into linear factors over R, and in examination questions this should always be the case. If this causes concern, then whenever an
R appears in the text, replace it with a C and work over the complex numbers
- then the arguments run through whatever the matrix is. Some of the justification and technical detail will be omitted from the article, but can be found in
[1] which also gives the general method for finding the Jordan Normal Form of
nxn matrices.

Definitions

Let f be a linear map from a vector space V to itself, and A the matrix representing f with respect to the standard basis.
(i) Then the characteristic polynomial is A (t) = det(A tI).
(ii) The minimal polynomial is the monomial (which means a polynomial
with the coefficient of the highest power being equal to one) of least degree such
that mA (A) = 0. The minimal polynomial is unique (which you may like to
prove), and divides the characteristic polynomial. Also, if P is any invertible
matrix, then mA (P 1 AP ) = 0 (try to prove this).
1

(iii) is an eigenvalue of A if A () = 0.
(iv) v is an eigenvector for the eigenvalue if Av = v.
(v) The eigenspace for the eigenvalue is ker(A I). This may also be
called the -eigenspace.
(vi) If the factor (t ) appears in the polynomial with multiplicity n (ie.
mA (t) = (t )n g(t) for some polynomial g(t) which does not have (t ) as a
factor), then the generalised -eigenspace is ker[(A I)n ].
Here is an important theorem:
Theorem 2.1 (Primary Decomposition Theorem). Let f : V V be
linear and represented by matrix A with minimal polynomial mA (t) = (t
1 )e1 ...(t r )er where 1 , ..., r are distinct eigenvalues of A and e1 , ..., er are
positive integers. Let V1 , ..., Vr denote the generalised eigenspaces ie. Vi =
ker[(A i I)ei ]. Then V = V1 ... Vr .
This theorem says that two (or more) eigenvectors with distinct eigenvalues
are linearly independent (among other things). The proof can be found in [1]
p216.
Theorem 2.2 (Cayley-Hamilton). A matrix satisfies its own characteristic
equation, ie. A (A) = 0.
Proposition 2.3. Using the notation from Theorem 2.1,
ker(A i I) ker[(A i I)2 ] ... ker[(A i I)ei ].
I will use these three results to give a little explanation of the reasoning
behind some of the arguments below.

The 2x2 case, V = R2 .

The first thing to do is always find the characteristic polynomial (which will
be a quadratic polynomial), and then the eigenvalues (which are the roots of
the characteristic polynomial, and you should be able to find them without too
much difficulty). There are three different cases - (i) distinct eigenvalues, (ii)
one eigenvalue and A is just I, (iii) one eigenvalue and A is not I.
Case(i) A (t) = (t 1 )(t 2 ) and 1 6= 2 . Find eigenvectors
v1 and v2


v1 v2 , which means that
such that
Av
=

v
for
i
=
1,
2.
Form
matrix
P
=
i
i
i
 
 
 


a
b
ab
1 0
if v1 =
and v2 =
then P =
. Then P 1 AP =
.
c
d
cd
0 2
P will be invertible, because we have distinct eigenvalues and the Primary
Decomposition Theorem says eigenvectors with different eigenvalues are linearly
independent, so v1 and v2 are linearly independent and thus P is invertible.
In this case, the minimal polynomial is the same as the characteristic polynomial, mA (t) = (t 1 )(t 2 ).
2

Case(ii) A (t) = (t )2 . So we have only one eigenvalue. If the eigenspace is 2-dimensional, this means Av = v for all vectors v R2 , and
this can only happen if A = I, which means A was in Jordan Normal Form
already.
In this case, the minimal polynomial is just mA (t) = (t ).
Case(iii) A (t) = (t )2 . Again we have only one eigenvalue, but A is
not I, which means when we try to find the -eigenspace, we only get one
vector, call it v1 . Now, the Cayley-Hamilton theorem (Theorem 2.2) says that
(A I)2 = 0, which means that ker[(A I)2 ] = R2 . Thus we can find a
vector v2 which is in ker[(A I)2 ] but not in ker(A I), so v1 and v2 provide
a basis for R2 .
Now modify this basis by replacing v1 with v10 = (A I)v2 . You can check
that since v2 is in ker[(A I)2 ], this means that v10 is in ker(A I), and so
v10 and v2 still form a basis of R2 .
 


1
0
1
Form the matrix P = v1 v2 , and then P AP =
.
0
In this case the minimal polynomial is mA (t) = (t )2 .

The 3x3 case, V = R3

As with the 2x2 case, start by finding the characteristic equation (usually a little
bit more work than for the 2x2 case), and then find the eigenvalues by looking
for roots of the characteristic equation (in an exam question, you should always
be able to spot one root and then factorise from there, so start by checking
if 1,-1,2,-2 are roots). There are three main cases, some of which split up into
subcases, but all follow the same general method : (i) three distinct eigenvalues,
(ii) two distinct eigenvalues, ie. the characteristic equation has one repeated
root, and one other root, (iii) one eigenvalue, ie. the characteristic equation has
one thrice repeated root.
Case(i) A (t) = (t 1 )(t 2 )(t 3 ) with i distinct. This case is very
similar to case(i) from the 2x2 section. Find eigenvectors v1 , v2 and v3 such
that Avi = i vi for i = 1, 2, 3. Then by Theorem 2.1, v1 , v2 and v3 form a
linearly independent set which span R3 .

1 0 0


Define the matrix P = v1 v2 v3 and then P 1 AP = 0 2 0.
0 0 3
In this case the minimal polynomial is the same as the characteristic polynomial, mA (t) = (t 1 )(t 2 )(t 3 ).
Case(ii) A (t) = (t 1 )(t 2 )2 . Firstly find an eigenvector v1 such that
Av1 = 1 v1 . Next find eigenvectors for 2 , ie. find ker(A2 I) - this splits into
two subcases, where dim(ker(A 2 I)) = 2 and where dim(ker(A 2 I)) = 1.
Subcase(a) If dim(ker(A 2 I)) = 2, then pick two linearly independent


vectors in ker(A 2 I), say v2 and v3 . Form the matrix P = v1 v2 v3 and
3

1 0 0
then P 1 AP = 0 2 0. In this case the minimal polynomial is mA (t) =
0 0 2
(t 1 )(t 2 ).
Subcase(b) If dim(ker(A 2 I)) = 1, then let v2 be in ker(A 2 I). Theorem 2.1 and Proposition 2.3 together give that the generalised 2 -eigenspace
is ker[(A 2 I)2 ], and is 2-dimensional since R3 is a direct sum of generalised eigenspaces. So pick a vector v3 which is in ker[(A 2 I)2 ] but not in
ker(A 2 I), so that v1 , v2 and v3 form a basis of R3 . Modify the basis by
replacing v2 with v20 = (A 2 I)v3 . Then since v3 is in ker[(A 2 I)2 ], this
means v20 is in ker(A 2 I) (This modification is exactly the same modification
we did in Case (iii) of the 2x2 case).

1 0 0


Now form the matrix P = v1 v20 v3 , then P 1 AP = 0 2 1. In this
0 0 2
case, mA (t) = (t 1 )(t 2 )2 .
Case(iii) A (t) = (t )3 . Arguably the most likely to come up in an exam,
because more can happen in this case. Conversely, if you can understand this
case, you should be in a position to understand everything else in this article!
As always, start by finding the eigenvectors which satisfy (A I)v = 0.
There are subcases according to whether the dimension of ker(A I) is 1, 2
or 3.
Subcase(a) dim(ker(A I)) = 3. This subcase is analagous with Case(ii)
of the 2x2 case, since Av = v for all vectors v R3 , which means A = I, and
A is already in Jordan Normal Form. In this case, the minimal polynomial is
mA (t) = (t ).
Subcase(b) dim(ker(AI)) = 2. Pick linearly independent vectors v1 and
v2 which are span ker(AI). Proposition 2.3 implies that ker[(AI)2 ] = R3 ,
so pick vector v3 which is in ker[(A I)2 ] but is not in ker(A I) so that
v1 , v2 and v3 form a basis of R3 . Now modify this basis by replacing v1 with
v10 = (A I)v3 . As in previous cases, v3 in ker[(A I)2 ] implies that v10 is in
ker(A I), so v10 , v2 and v3 still form a basis of R3 (for a proof of this, see [1]
p211).


v10 v3 v2, and note the order that the vectors are
Now form matrix P =
10
taken in. Then P 1 AP = 0 0. In this case the minimal polynomial is
00
mA (t) = (t )2 .
Subcase(c) dim(ker(A I)) = 1. Pick vector v1 which spans ker(A I).
Now Proposition 2.3 says that ker(A I) ker[(A I)2 ] ker[(A I)3 ]
and we cannot have dim(ker[(A I)2 ]) = 3 in this case (try and prove this), so
extend v1 to a basis of ker[(A I)2 ] given by v1 , v2 (ie. find such a v2 ). Then
extend again to give a basis of ker[(A I)3 ] (which is R3 by Cayley-Hamilton)
given by v1 , v2 and v3 (ie. find such a v3 ).

Now modify this basis by replacing v2 by v20 = (A I)v3 and by replacing

v1 with v10 = (A I)v20 = (A I)2 v3 . v10 still gives a basis of ker(A I) and
v10 , v20 together still give a basis of ker[(A I)2 ]. So v10 , v20 and v3 give a basis
of R3 .

10


Form the matrix P = v10 v20 v3 and then P 1 AP = 0 1. The minimal
00
3
polynomial is mA (t) = (t ) in this case.

The minimal polynomial

By looking at all the cases above, we see that the knowing the characteristic
polynomial and the minimal polynomial actually determines what the Jordan
Normal Form will be (NB this is not true in higher dimensions). There is usually
less work in finding what the minimal polynomial is than finding a base-change
matrix which will give Jordan Normal Form. This is because we know the
minimal polynomial divides the characteristic polynomial, so we just have to
check a few cases to find what the minimal polynomial is, which amounts to
multiplying a few matrices together. So if in an exam, youre just asked to find
the Jordan Normal Form without finding a base change matrix, you could use
this. Even if you are asked for a base-change matrix, it at least provides a good
check.
For 2x2 matrices:
mA (t)
Jordan Normal
 Form
1 0
(t 1 )(t 2 )
0
 2
0
(t )
0
 
1
(t )2
0
For 3x3 matrices:
A (t)
(t 1 )(t 2 )(t 3 )

(t 1 )(t 2 )2

(t 1 )(t 2 )2

(t )3

mA (t)

Jordan
Normal
Form
1 0 0
0 2 0
(t 1 )(t 2 )(t 3 )
0 0 3
1 0 0
0 2 0
(t 1 )(t 2 )
0 0 2
1 0 0
0 2 1
(t 1 )(t 2 )2
0 0 2
00
0 0
(t )
00
5

(t )3

(t )2

(t )3

(t )3

10
0 0
00

10
0 1
00

Worked Examples

You should now have an idea how to go about finding the Jordan Normal Form,
and just to show that it really works, here are two examples from the University
of Warwicks AlgebraI Assignment
7 by Dr. Keilen.

41
. The characteristic equation is given by A (t) =
Example 1) A =
1 6
(4t)(6t)+1 = (t5)2 , so then 5 is the only eigenvalue. We have to be in case
(iii) in the 2x2 section, since A is not a multiple
  of the identity, which means
51
the Jordan Normal Form must be given by
, so now we just have to find a
05

   
41
x
5x
base change matrix to get to it. So, find the eigenvector:
=
.
1 6
y
5y


1
So 4x + y = 5x and x + 6y = 5y which reduce to y = x, so define v1 =
.
1
 
1
Now extend to basis of R2 by, for example, v2 =
(there are different vectors
0
0
we could have taken for v2 ).
 Now
 modify the basis by taking v1 = (A 5I)v
2


1
1 1
0
which if you work it out is
. So form the matrix P = v1 v2 =
.
1
1 0


 
0 1
51
P 1 =
, and Ill leave it to you to check that P 1 AP =
.
1 1
05

3 1 1
Example 2) A = 1 3 1. Working out the characteristic equation
2 2 0
3
gives
A (t) = (t+2)
. So
-2 is the only eigenvalue. Now find eigenvec3 1 1
x
2x
tors, 1 3 1 y = 2y . These equations reduce to -x-y+z=0, so
2 2 0
z
2z
ker(A + 2I) is 2 dimensional and
(iii)b)
weare in case
of the 3x3 section. Pick a
1
0
basis of ker(A + 2I), eg. v1 = 1 and v2 = 1. Now, ker[(A + 2I)2 ] must
0
1

1
be R3 , so extend v1 and v2 to a basis of R3 by, for example, v3 = 0. Now
0


1
modify the basis by replacing v1 with v10 = (A + 2I)v3 = 1. Form the ma2

0 1 1
1
1
0


trix P = v10 v3 v2 = 1 0 1. You can now check that P 1 = 1 1 1
2 0 1
0 2 1

2 1 0
and that P 1 AP = 0 2 0. We can verify this is the expected Jordan
0 0 2
Normal Form by checking that (A + 2I)2 is the zero matrix and comparing with
the table in the previous section.
c
Alastair
Fletcher, 2003

References
[1] R.Kaye and R.Wilson Linear Algebra (OUP 1998)