Anda di halaman 1dari 4

COMPUTATION OF EIGENVALUES AND EIGENVECTORS

Let A be an nn matrix. An eigenvalue and a corresponding eigenvector x will satisfy A x = x , where we


require that x 0 . After writing this relation as A x = I x , collecting the terms to the left hand side, and
factoring out x, it takes the form (A - I) x = 0. This is a linear system for x , which obviously has the trivial
solution x = 0. A square linear system can have more than one solution only if the determinant of its coefficient
matrix is zero. So, in order to find values of that permit a non-trivial (non-zero) x , we first search for values of
that make
det (A - I) = 0. This relation is a polynomial equation of degree n for . Its n solutions are the eigenvalues of A.
For each of the eigenvalues, we then look for a matching eigenvector x . With now known, we can apply
Gaussian elimination to the system (A - I) x = 0 in order to find such a vector (or vectors) x.
If all eigenvalues i, i = 1,2,...,n of an nn matrix A are different, there will be one eigenvector belonging to each
eigenvalue. If an eigenvalue has multiplicity k > 1, there may be any number from 1 to k of eigenvectors
belonging to it. The procedure below to find the eigenvectors will reveal how many there are.
If there turns out to be fewer than k eigenvectors belonging to an eigenvalue of multiplicity k, we can find
generalized eigenvectors in place of the missing regular eigenvectors.

Example 1:

3 2 4

Determine the eigenvalues and eigenvectors of the matrix A = 2 0 2 .

4 2 3

Eigenvalues: We start by finding these from the relation


3 2
4
0 = det (A I) =
2 2
4
2 3
= 2

2 4
2 3

3 4
4 3

it is probably easiest to

just expand along any row or


column; say, for ex, along row 2

3 2
4 2

= 2 (2 ) (7 6 + 2 ) 2 (2 2) = 3 + 6 2 + 15 + 8 .
Solving this cubic equation gives 1 = 8, 2= -1, 3 = -1.
At this point, it is a good idea to make a check against the theorem that, for all square matrices,

i = a ii . We find i = 6 and a ii = 6 (so the likelihood that we have done an algebraic error is
rather small).

Eigenvectors: We need to solve a separate linear system for each distinct eigenvalue:

1 = 8: The system to solve is

5 2 4

2 8 2

4 2 5

x1 0


x2 = 0


x3 0

A 8I

0
0
0

1 4 1

...

0 2 1

0 0 0

0
0
0

. In compact notation:

x 1 4x 2 + x 3 = 0

2x 2 x 3 = 0

0x 3 = 0

Back substitution gives x3 = s, x2 = s, x1 = s . This can be summarized as


1
x1
1

x2 = s 2

x3
1

The vector above is the eigenvector. It is undetermined with respect to a scalar multiplier. To
avoid working with fractions, we write
2

1 = 8, x 1 = 1 .

2
2 = 3 = -1:

The system (A (1)I) x = 0 becomes


4 2 4

2 1 2

4 2 4

0
0
0

2 1 2

... 0 0 0

0 0 0

0
0
0

2 x 1 + x 2 + 2 x3 = 0

0 x 2 + 0 x3 = 0

0 x3 = 0

Back substitution gives


0 x3 = 0
x3 = s
0 x2 + 0 s = 0
x2 = t
2 x 1 + t + 2 s = 0 x 1 = s 12 t

(a free parameter)
(another free parameter)

and the general solution becomes


x1
1 12

x2 = s 0 + t 1

x3
1 0

Hence, to the double eigenvalue 2 =3 = -1 corresponds two eigenvectors


1

x2 = 0 , x3 =

2 .

Although the matrix A has a multiple eigenvalue, it has a full set of eigenvectors.

Example 2:

Determine the eigenvalues, eigenvectors, and generalized eigenvectors (if any) of the matrix
1 1
A=
.
0 1

Eigenvalues:
The eigenvalues for an upper (or lower) triangular matrix are the same as the diagonal
elements. Therefore, we can immediately read off 1 = 2 = 1.

0 1 0
The system to solve is (A -1 I) x = 0 or, in compact form,
. This system
0
0 0
is immediately ready for back substitution. The last equation gives x2 = s . Then the first equation
becomes 0 x1 + s = 0. This is a conflict unless s = 0. After setting s = 0 , x1 becomes arbitrary: x1 = t. The
general solution to the system becomes
Eigenvectors:

x1
1

=t .
x
2
0
Although A is a 22 matrix, it has only one eigenvector.
It depends on the application, whether we consider this to be 'the end of the story' or not regarding finding
the eigenvalues and eigenvectors of A. If the reason we searched for these was that we needed to solve a
linear system of ODEs x = A x , we will need to continue the search for a second vector associated to =
1. This vector x2 would then be what is called a generalized eigenvector, and we require it to satisfy (A I) x2 = x1 (as opposed to (A - I) x2 = 0 , which would hold for a regular eigenvector). In compact
form, the system we need to solve for x2 is
0 1

0 0

1
.
0

This is again right away in upper triangular form. The last equation gives x2 = s. The first one,
0 x1 + 1 s = 1, forces s = 1. The unknown x1 becomes undetermined, and we set x1 = t. So the solution to
the linear system is
x1 0 1

= +t .
x2 1 0
1
0
The last vector above, , only re-confirms the eigenvector we found earlier. The first vector is
0
1
what is called a generalized eigenvector, belonging also to = 1.

We note that, while an eigenvector is undetermined with respect to an arbitrary factor, a generalized
eigenvector is not (it is however undetermined with respect to a multiple of the corresponding
eigenvector) .

In conclusion for this example: The matrix A has a double eigenvalue 1 = 2 = 1. However, it does not
have a 'full set' of eigenvectors. It has instead one eigenvector and one generalized eigenvector,
1
x1 =
0

and

0
x2 =
1

respectively.

Some more comments about generalized eigenvectors:


If we have an example with, say, an eigenvalue of order four and we find only one eigenvector x1 , we
need to find three generalized eigenvectors for . These vectors would then form what is called a chain and we find them by solving the sequence of linear systems
(A - I) x1 = 0,

(A - I) x2 = x1,

(A - I) x3 = x2,

(A - I) x4 = x3.

The only real complication with generalized eigenvectors occurs when a multiple eigenvalue has several
eigenvectors, but still not a full set of them. Then it is not clear which of the eigenvectors we should use to
initialize the one (or several) chains that can occur. Trial and error is the only procedure (within the scope
of this course) that we can suggest for use that rare case.

Anda mungkin juga menyukai