Lecture 4.2.

Jordan form

March 31, 2020

Today we finally address deficient matrices, that is those which are non-
diagonalizable, which do not have enough eigenvectors to span the space.
A typical example is the following m × m matrix

λ1 1 0 . . . 0 0
 

 0 λ1 1 . . . 0 0 

0 0 λ1 . . . 0 0
 
 
Jm (λ1 ) =  .

 ... ... ... ... ... ... 

0 0 0 . . . λ1 1
 
 
0 0 0 . . . 0 λ1

We have λ1 on the main diagonal, 1 on the diagonal above the main, all other
entries are 0. Such a matrix is called a Jordan block of size m with eigenvalue
λ1 . Its characteristic polynomial is (λ1 − λ)m , so the only eigenvalue is λ1 ,
and the eigenspace corresponding to this eigenvalue is 1-dimensional. It is
spanned by an eigenvector (1, 0, . . . , 0)T . Check all this!
Look how this matrix maps the vectors of the standard basis:

Ae1 = λ1 e1 , Ae2 = λe2 + e1 , Ae3 = λ1 e3 + e2 ,

and so on. So e1 is an eigenvector, and the rest of ej are called generalized

eigenvectors.
Definition. Let A be a square matrix, and λ be an eigenvalue and v 6= 0
an eigenvector, so that Av = λv. Then the first generalized eigenvector
attached to v is a solution v1 of the equation

Av1 = λv1 + v.

1
The second generalized eigenvector v2 attached to v is a solution of

Av2 = λv2 + v1 .

and so on. Generalized eigenvectors form a chain v1 , v2 , v2 , . . . , vk such that

Avj+1 = λvj+1 + vj .

Jordan’s Theorem. For every linear operator L in a (complex) finite-

dimensional space there is a basis consisting of eigenvectors and generalized
eigenvectors.
It is called the Jordan basis for L.
I recall that two matrices A and B are called similar if there is a non-
singular matrix C such that

A = CBC −1 .

Geometrically, this means they they represent the same linear operator, and
encode it with respect to two different bases.
The matrix of the operator with respect to its Jordan basis has a Jordan
form which consists of diagonal blocks, each block is a Jordan block.
Corollary. Every square matrix is similar to its Jordan form. Two matrices
are similar if and only if they have the same Jordan form (up to permutation
of Jordan blocks).
If A is a matrix, and J is its Jordan form, then

A = BJB −1 ,

where B is the matrix whose columns are eigenvectors and generalized eigen-
vectors.
Examples. Here are all possible Jordan forms for n = 2:
! !
λ1 0 λ1 1
, .
0 λ2 0 λ1

The first of these is diagonalizable, it has two Jordan blocks of size 1. The
second has one Jordan block of of size 2.

2
 Next we list all possible Jordan forms for n = 3:
     
λ1 0 0 λ1 1 0 λ1 1 0
 0 λ2 0  ,  0 λ1 0  ,  0 λ1 1  .
     

0 0 λ3 0 0 λ2 0 0 λ1

Jordan forms obtained by permutation of the blocks are considered the same.
First matrix has 3 blocks of size 1, second has 2 blocks of sizes 2 and 1, and
the third matrix has one block of size 3.
First matrix is diagonalizable (λ1 , λ2 , λ3 are not necessarily distinct!): it
has three linearly independent eigenvectors.
The second has two linearly independent eigenvectors, one with eigen-
value λ1 and one with eigenvalue λ2 (again λ1 and λ2 ) are not necessarily
distinct!). It also has one generalized eigenvector attached to the true eigen-
vector corresponding to λ1 .
The third matrix has only one eigenvector (up to proportionality) and to
it two generalized eigenvectors are attached.
In general, suppose we have some eigenvalue λ. The dimension of the
eigenspace corresponding to λ is the number of Jordan blocks with this λ on
the main diagonal. To each of these eigenvector, some number of generalized
eigenvectors can be attached. The sum of sizes of all Jordan blocks with
eigenvalue λ is the multiplicity of λ as a root of the characteristic equation.
So, for example, this matrix
 
2 1 0 0 0
0 2 1 0 0
 
 
 

 0 0 2 0 0 


 0 0 0 2 1 

0 0 0 0 2

has characteristic equation λ5 − 32 = 0. The only eigenvalue is λ = 2. The

eigenspace corresponding to this eigenvalue has dimension 2. So we have two
linearly independent eigenvectors, they are in fact e1 and e4 . In addition we
have generalized eigenvectors: to e1 correspond two of them: first e2 and
second e3 . To the eigenvector e4 corresponds a generalized eigenvector e5 .
To find the Jordan form and the Jordan basis for some matrix, you do
the following:
a) find eigenvalues.

3
 b) for each eigenvalue, find a basis of the eigenspace. If the sum of
the dimensions of eigenspaces is n, the matrix is diagonalizable, and your
eigenvectors make a basis of the whole space.
c) if not, try to find generalized eigenvectors v1 , v2 , . . . by solving (A −
λI)v1 = v, for an eigenvector v, then, if not enough, (A − λI)v2 = v1 , and
so on. You need only one solution of each of these equations.
These generalized eigenvectors will be attached to those λ for which the
number of true linearly independent eigenvectors is less than the multiplicity
of λ as the root of the characteristic equation.
d) Jordan’s Theorem guarantees you that eventually you will find suf-
ficient number of generalized eigenvectors, so that they, together with true
eigenvectors make a basis of the whole space. This is the Jordan basis.
Example 1. (Probl. 40, p. 305) Which pairs of these matrices are always
similar (for all a, b, c, d) and which are not (for some a, b, c, d)?
! ! ! !
a b b a c d d c
A= , B= , C= , D= .
c d d c a b b a
Answer: A and D are similar. B and C are similar. All other pairs may not
be similar for some a, b, c, d.
Explanation: A and D have the same characteristic polynomial. Also B
and C have the same characteristic polynomial. But these two polynomials
are different, and it is easy to find a, b, c, d for which they are numerically
different.
Now A is similar to D because
! ! ! !
a b 0 1 d c 0 1
= .
c d 1 0 b a 1 0
Multiplication on !
0 1
,
1 0
which is self-inverse, from the left interchanges rows and multiplication on
the same from the right interchanges columns. Similarly B is similar to C.
Example 2. Find the Jordan form and the Jordan basis for the following
matrix  
0 1 0
 −4

4 0 
.
−2 1 2

4
 Solution. The characteristic polynomial is (2 − λ)3 , so the only eigenvalue
is 2. By the standard procedure we find two linearly independent eigenvec-
tors: (0, 0, 1)T and (1, 2, 0)T . None of them has a generalized eigenvector, so
one has to try some linear combination of them, for example (1, 2, 1)T . To
this eigenvector, there is a generalized eigenvector (0, 1, 0)T . So the Jordan
form and a Jordan basis can be taken as
   
2 0 0 0 1 0
J = 0 2 1 

, and B =  0 2 1 

.
0 0 2 1 1 0
The main use of the Jordan form is for solving differential and difference
equations with deficient matrices. One can compute the powers and expo-
nential of the Jordan block (see the book, p. 300-301). The k-th power is
upper triangular and has the following form:
on the main diagonal stand λk .
on the diagonal above the main stand kλk−1 ,
on the next diagonal stand k(k−1)
2
λk−2 ,
and so on. On the diagonal number m ≤ k above the main diagonal stand
k!
λk−m ,
m!(k − m)!
the terms of the binomial formula.
The exponential eJt of the Jordan block is also upper triangular:
on the main diagonal: eλt ,
on the next above: teλt ,
then on the next: (t2 /2)eλt , and so on.
on the m-th diagonal above the main: (tm /m!)eλt .
Exercise. Prove these facts.
In practice, one does the following. Consider a differential equation
x′ = Ax.
Let λ be an eigenvalue of A and v an eigenvector. To it corresponds a solution
x0 (t) = eλt v.

5
let v1 be the first generalized eigenvector. To it corresponds the solution
x1 (t) = eλt (tv + v1 ).
Indeed
x′1 = λteλt v + eλt (v + λv1 ).
Plugging this to the equation and dividing by eλt we obtain
λtv + (v + λv1 ) = tAv + Av1 ,
and this is true since Av = λv and Av1 = λv1 + v. Similarly, if there is a
second generalized eigenvector, the form of the third solution is
x2 (t) = (t2 /2)eλt v + teλt v1 + eλt v2 .
And in general, for the (m + 1)st generalized eigenvector,
xm (t) = (tm /m!)eλt v + (tm−1 /(m − 1)!)eλt v1 + . . . + vm .
Exercise. Verify this.
Effect of the Jordan form on stability. When Re λ < 0, eλt → 0 as
t → ∞. This does not change if we multiply the exponent on any power of t.
Therefore when all eigenvalues have negative real part, the system x′ = Ax
has stable equilibrium at 0, no matter whether A is diagonalizable or not.
However, when Re λ = 0, then eλt is bounded, but it oscillates. Multipli-
cation on t gives an unbounded function. So when all eigenvalues have non-
positive real part, this is not sufficient for stability. If those eigenvalues with
Re λ = 0 are not deficient (have as many linearly independent eigenvectors
as their multiplicities as roots of the characteristic equation) then the system
is stable. However if there is at least one eigenvalue with Re λ = 0 which is
deficient (has fewer eigenvectors than its multiplicity) then the system is not
stable: each generalized eigenvector will give an unbounded solution.
Remark. If a matrix is known only approximately, for example, if its
entries are results of some measurements, and computation indicates that
the characteristic equation has a multiple root, one cannot conclude whether
this root is indeed multiple, and if it is, whether the root is deficient or not
(whether there are as many eigenvectors as its multiplicity, or less). Many
computer systems have a command called Jordan, which is supposed to find
the Jordan form, but one has to be very careful with them: it is easy to
deceive these programs, even with 2 × 2 matrices!