You are on page 1of 4

10.

2 Generalized Eigenvectors
In the previous section we looked at the case where each eigenvalue of a square matrix A
has as many linearly independent eigenvectors as its multiplicity. In that case we could
diagonalize A and use this to compute its powers as in chapter six. Now we turn to
situations where some eigenvalues do not have as many linearly independent
eigenvectors as their multiplicities. In such cases we shall consider what are called
generalized eigenvectors. These turn out to be a substitute for regular eigenvectors. In
particular, they can be used to express A as A = TJT-1 where J, called the Jordan canonical
form of A, is "almost" diagonal. This is a generalization of diagonalization and can be
used to compute the powers in a fashion similar to chapter six. Let's look at an example.
Example 1. Let A = . We find the eigenvalues.
0 =

= (1 - )(3 - ) + 1 = 2 - 4 + 4 = ( - 2)2

The only eigenvalue is = 2 which is of multiplicity two. An eigenvector v = satisfies


= (A - 2I)v =
Both equations are - x + y = 0 or y = x so the eigenvectors are multiples of v1 = . So even
though the eigenvalue = 2 is of multiplicity two it has only one linearly independent
eigenvector. In situations like this we turn to generalized eigenvectors as a substitute for
eigenvectors. In order to appreciate the definition of a generalized eigenvector, note that
an eigenvector v for an eigenvalue satisfies (A - I)v = 0. If we replace A - I by (A I)m we get the definition of a generalized eigenvector.
Definition 1. Let A be a square matrix and be an eigenvalue of A. A vector v is a
generalized eigenvector for if
1.
v 0
2.
(A - I)mv = 0 for some positive integer m
The smallest positive integer m such that (A - I)mv = 0 is called the degree of the
generalized eigenvector.
Note that any eigenvector of A is a generalized eigenvector of degree one since it satisfies
(A - I)mv = 0 with m = 1. In fact an eigenvector v of A satisfies (A - I)mv for any
positive integer m since (A - I)mv = (A - I)m-1(A - I)v = (A - I)m-10 = 0.
Example 2. Find the generalized eigenvectors of A = .
In Example 1 we saw that = 2 was the only eigenvalue and the only eigenvectors were
multiples of v1 = . These are the generalized eigenvectors of degree one. The generalized
eigenvectors of degree two are the solutions of (A - 2I)2v = 0. One has A - 2I = so (A

10.2 - 1

- 2I)2 =
= = 0. So the equation (A - 2I)2v = 0 is 0v = 0. Every vector v = satisfies
this equation, so every non-zero vector v is a generalized eigenvector of A. The ones that
do not lie on the line through v1 = are of degree two.
Now we look at how the generalized eigenvectors can be used to create a substitute for
the diagonalization of the matrix. For simplicity we consider a 22 matrix A with a
single eigenvalue that has only one linearly independent eigenvector v1 as in Example
1. It turns out that A must have a generalized eigenvector v of degree two. So (A - I)v is
a regular eigenvector. So (A - I)v = cv1. Let v2 = v/c. Then (A - I)v2 = v1 or
Av2 = v1 + v2. Let
T = matrix whose columns are v1 and v2
J =
Since Av1 = v1 and Av2 = v1 + v2 the matrix AT has columns equal to v1 and v1 + v2.
Consider TJ. Recall that the kth column of TJ is a linear combination of the columns of T
using the entries in the kth column of J as coeffients. So TJ has columns equal to v1 and
v1 + v2. So AT = TJ or
(1)

A = TJT-1 = T T-1

J = is called the Jordan canonical form of A and we shall call formula (1) the
Jordanization of A. It turns out to be a convenient substitute for the diagonalization of A
when it comes to computing An and other computations with A.
Example 3. Let A = . Find the Jordanization (1) of A.
In Example 1 we saw that = 2 was the only eigenvalue and the only eigenvectors are
multiples of v1 = . We need to find v = such that (A - I)v2 = v1 = . Since A - 2I = so
= . Both equations are x + y = 1 or y = 1 + x so v = = = + x where x can be any
number. Let's take x = 0 so v2 = . Then (1) becomes
=

-1

Now let's consider how to compute the powers of a matrix A using its Jordanization (1).
As in chapter six one has
An = TJnT-1 = T
n

It turns out that


2

T-1

is quite simple. One has


=

10.2 - 2

and in general

So
An = T T-1
Example 4. Let A = . Find An and use it to solve the difference equations
xn+1 =

xn +

yn

x0 = 4

xn+1 = - xn + 3yn

y0 = 5

From Example 3 one has


n

-1

-1

, so

= 2n

= 2n
The difference equations can be written as
=

The solution is
=

= 2n

= 2n

So
xn = (4 + n/2) 2n

yn = (5 + n/2) 2n

Here are some elementary properties of generalized eigenvectors.


Definition 2. Let A be a square matrix and be an eigenvalue of A. Let N,m = {v: (A I)mv = 0} be the set of generalized eigenvectors for of degree at most m along with the
zero vector. N,m is called the generalized eigenspace of degree m for .
The generalized eigenspaces are subspaces since they are the null spaces of the matrices
(A - I)m.
Proposition 1. Let A be a square matrix and be an eigenvalue of A.
(a) If (A - I)mv = 0 for some positive integer m then (A - I)m+kv = 0 for any
positive integer k, i.e. (A - I)pv = 0 for some positive integer p that is larger
than m.

10.2 - 3

(b)
(c)
(d)
(e)

N,1 N,2 N,m N,m+1


If N,m = N,m+1 for some m then N,m = N,m+k for all positive integers k, i.e. N,m
= N,p for all positive integers p larger than m.
v is a generalized eigenvector of degree m if and only if (A - I)v is a
generalized eigenvector of degree m 1.
v is a generalized eigenvector of degree m if and only if Av = v + w where w is
a generalized eigenvector of degree m 1.

Proof. (a) If (A - I)mv = 0 then (A - I)m+kv = (A - I)k(A - I)mv = (A - I)k0 = 0. (b)


follows from (a). (c) We shall show that N,m = N,m+1 implies N,m+1 = N,m+2. Applying
this over and over again will show N,m = N,m+k for all positive integers k. Note that
N,m = N,m+1 is equivalent to (A - I)m+1v = 0 (A - I)mv = 0. Suppose (A - I)m+2v = 0.
We write (A - I)m+2v = (A - I)m+1(A - I)v. So (A - I)m+1(A - I)v = 0. So (A - I)m(A
- I)v = 0. So (A - I)m+1v = 0. Thus (A - I)m+2v = 0 (A - I)m+1v = 0. So N,m+1 =
N,m+2. (d) This follows from the fact that (A - I)mv = (A - I)m-1(A - I)v. (e) follows
from (d). /

10.2 - 4

You might also like