El) A =

E2) A =
[
[
"
-15
J2
j2 J;]
4 3
2

10 -12 6 1
20 -4 2

E3lA= [-i 2 -3
-21 - 60 1

E4)A=[i
-1 -4
-i I]
In the examples considered so far, it was possible for us to find all the roots of the
characteristic equation exactly. But this may not always be possible. This is
particularly true for n > 3. In such cases some iterative method like Newton-Raphson
[I:; method may have to be used to find a particular eigenvalue or all the eigenvalues
from the characteristic equation. However, in many practical problems, we do not
:1
iY'
%
require all the eigenvalues but need only a selected eigenvalue. For example, when
we use iterative methods for solving a nonhomogeneous system of linear equations
Ax = b, we need to know only the largest eigenvalue in magnitude-qf the iteration
matrix H, to find out whethd the method converges or not. One iterative method;
which is frequently used to determine the largest eigenvalue in magnitude (also called
the dominant eigenvalue) and the corresponding eigenvector for a given square matrix
A is the power method. In this method d e do not find the characteristic equation.
This method is applicable only when all the eigenvalues are real and distinct. If the
.magnitude of two or more eigenvalues is the same then the method converges slowly.

8.3 THE POWER METHOD

Let us consider the eigenvalue problem
A x = Ax.
Let A,, A*, ...,An be the n real and distinct eigenvalues of A such that

Therefore, Al is the dominant eigenvalue of A.

 In this method, we start with an arbitrary nonzero vector y(O) (not an eigenvector),
I
solution Or Liwar A'gebmicEquntions
and form a sequence of vectors ( Y ( ~ ) )

In the limit as k + m, y(k) converges to the eigenvector corresponding to the

dominant eigenvalue of the matrix A. We can stop the iteration when the largest
element in magnitude in y'"+')-y'") is less than the predefined error tolerance. For
simplicity, we usually take the initial vector ycU)with all its elements equal to one.
I Note that in the process of multiplying the matrix A with the vector y(k), the
I

I
Vector for which scal~nghas been elements of the vector y(k+'J may becbme very large. To avoid this, we normalize'(or
done is called a scaled vector scale) the vector y(k) at each step by dividing y(k),by its largest element in magnitude.
1 otherw~se,it IS unsealed. This will make the largest element in magnitude in 'the vector y(k+') as one and the
I remaining elements less than one.
I

I
If y(k) represents the unscaled vector and y(k) the scaled vector then, we have the
I
I power method.

v(()) = y0)and
mk+,being the largest element in magnitude of y(k+') We then
obtain the dominant eigenvalue by taking the limit

(~(~+'))r
A, = lim
k-tm (~(~))r

where r represents the rth component of that vector. Obviously, there are n ratios of
numbers. As k-m all these ratios tend to the same value, which is thedargest
eigenvalue in magnitude i.e., A,. The iteration is stopped when the magnitude of the
difference of any two ratios is less than the prescribed tolerance.
The corresponding eigenvector is then dk+')
obtained at the end of t'he last iteration
performed.
We now illustrate the method through an example.

Example 4 : Find the dominant eigenvalue and the corresponding eigenvector correct
to two decimal places of the matrix

using the power method.

Solution : We take
y(O) = v(O) = (1 1 1)T
Using Eqn. (14), we obtain

[ -1 -: -:I 1.1 [-i ]

Again,
2 -1
y(') = AV(') = =
 v(7) = -y'7' = [0.7071 -1 0.70711~
3.4146

After 7 iterations, the r a t i o s B are given as 3.4138, 3.4146 and 3.4138. The
(v(@)r
maximum error in these ratios is 0.0008. Hence the dominant eigenvalue can be taken
as 3.414 and the corresponding eigenvector is [0.7071 -1 0.70711~

Note that the exact dominant eigenvalue of A as obtained in Example 3 was

2 + J2 = 3.4142 and the corresponding eigenvector was [I -& 11' which can also
1 1
be written as [-- -1 --lT = [0.7071 -1 0.70711~
J2 J2
You may now try the followingexercises.

Using four iterations of the power method and taking the initial vector y(") with all
its elements equal to one, find the dominant eigenvalue and the correspondingeigen-
vector for the following matrices.

You must have realised that an advantage of the power method is that the eigenvector
'corresponding to the dominant eigenvalue is also generated at the same time.
Usually, for most of the methods of determining eigenvalues, we need to do separate
computations to obtain the eigenvector.
In some problems, the most important eigenvalue is the eigenvalue of least
magnitude. We shall discuss now the inverse power method which gives the least
ejgenvalue in magnitude.
'1 Soktlon ot Llnenr Algebralc Equations
8.4 THE INVERSE POWER METHOD
1
We first note that if A is the smallest eigenvalue in magnitude of A, then - is the
X.-
largest eigenvalue in magnitude of A-'. The corresponding eigenvectors are same.
If we apply the power method to A-l, we obtain its largest eigenvalue and the
corresponding eigenvector. This eigenvalue is then the smallest eigenvalue in
magnitude of A and the eigenvector is same. Since power method is applied to A-l,
it is called the inverse power method.
Consider the method
y(k+l) = ~-1G(k) k=o ,1,2 ,.........
7 (17)

where y(O) is an arbitrary nonzero vector diffiient from the eigenvector of A.

However, algorithm (17) is not in suitable form, as one has find A-'. Alternately,
we write Eqn. (17) as
~,,(k+') = v(k)

We now need to solve a system of equations for Y ( ~ + 'which

), can be obtained using
any of the method discussed in the previous units. The ,largest eigenvalue of A-' ks
again given by
(y(k+l))r
p = lim
L - + ~ (dk))r

The corresponding eigenvector is d k + ' ) .

%e now illustrate the method thr$ugh an example.

Example 5 : Find the smallest eigenvalue in magnitude and the. corresponding

eigenvector of the matrix. .

using four iterations of the inverse power method.

.-" ~

Solution : Taking v(O) = [l 1 llT, we write

First iteration

..
For solving the system of Eqns. (19), we use the LU decomposition method. We write
. -

comparing the coefficients on both sides bf Eqns. (20): we obtain

and then uy(')= z
we obtain
y(') = [$ 2 $1 T
= [ 1.5 2.0 1.5 ]
T

Second iteration
~ ~ ( =2 v(l)
)

Solving LZ = v(')
and uy(')= z
we obtain

Third iteration

Fourth iteration

A p r 4 iterations, the ratios @?!

(d3))r
are given as 1.7059, 1.7083, 1.7059. The
maximum error in these ratios is 0.0024. Hence the dominant eigenvalue of A-' can
be taken as 1.70. Therefore, -- 0.5882 is the smallest eigenvalue of A in
1.70
magnitude and the corresponding eigenvector is given by [0.7073 1 0.70731~.
Note that the smallest eigenvalue in magnitude of A as calculated in Example 3 was
2-J2 = 0.5858 and the corresponding eigenvector was [I J2 llTor [0.7071 1 0.70711~.

You may now try the following exercise :

E7) Find the smallest eigenvalue in magnitude and the corresponding eigenvector of
the matrix

with do)= [-1 1lT, using four iterations of the inverse powe; method.

The inyersc! power method can be further generalized to find some other selected
eigenvalues of A. For instance, one may be interested to find the eigenvalue of A
which is nearest to some chosen number q. You know from P6 of Sec. 8.2 that the
matrices A and A-qI have the same set of eigenvectors. Further, for each eigenvalue
hi of A, hi-q is the eigenvalue of A-qI.
 -

wutlon of Ll-r Al~ebraic

. Equations
. We can therefore use the iteration

with scaling as described in Eqns. (14) - (16). We determine the dominant

eigenvalue p of ( A - ~ I ) - ' using the procedure given in Eqn. (18), i.e.

pt-1) - y(k+l)
mk+l
Using P6, we have the relation
p= -
1
where A is an eigenvalue of A.
7

A-q

1
Now since p i s the largest eigenvalue in magnitude of (A-~I)-', - must be the
wr
smallest eigenvalue in magnitude of A-qI. Hence, the eigenvalue - + q of A is
I.L
closest to q.

~xamp'le6 : Find the eigenvalue of the matrix A, nearest to 3 and also the
corresponding eigenvector using four iterations of the inverse power method where,
2 -1

-1
Solution : In this case q = 3. Thus we have

A-31 = [ -i
-1 -1

To find Y(~+'),
we need to solve the system

L J
and normalise y(k+l) as given in Eqn. (22).
First iteration
Starting with v(O) = [l 1 llT and using the Gauss elimination method to solve the
system (24), we obtain

Second iteration

Third iteration
After four iterations, the ratios are given as 2.5, 2.333, 2.5. The maximum
(d3))r
error in these ratios is 0.1667. Hence the dominant eigenvalue of (A-31)-' can
be taken as 2. Thus the eigenvalue A of A closest to 3 as given by Eqn. (23) is

and the corresponding eigenvector is d4)=

[7
-1 -:IT = [0.7143 -1 0.7143
IT
Note that the eigenvalue of A closest to 3 as obtained in Example 3 was 2+J2 = 3.4142.
The eigenvector corresponding to this eigenvalue was 0.7071 -1 0.7071
IT
And now a few exercises for you.

E8) Find the eigenvalue which is nearest to - 1 and the corresponding eigenvector for
the matrix

with do)= [ -1 ;IT, using four iterations of the inverse power method.
E9) Using four iterations of the inverse power method, find the eigenvalue which is
nearest to 5 and the corresponding eigenvector for the matrix

A = [: :] (exact eigenvalues are = 1 and 6)

with do)= [1 l r

The eigenvalues of a given matrix can alst be estimated. That is, for a given matrix
A , we can find the region in which all its eigenvalues lie. This can be done as follows:
Let Xi be an e'igenvalue of A and xi be the corresponding eigenvector, i.e..
Axi = hixi (25)
or