ES 361 - Computing

Methods in Engineering
Lecture 5
Eigenvalue Problems
The equation

‫ݕ = ݔ ܣ‬

can be viewed as a linear transformation that transforms a given vector {x} into a new vector {y}. Vectors that
are transformed into multiples of themselves are important in many applications (i.e. engineering problems
involving vibrations, elasticity, and other oscillating systems), including finding solutions to systems of first-
order linear differential equations with constant coefficients.

To find such vectors, we set {y} = λ{x}, where λ is a scalar proportionality factor, and seek solutions of the
equation:

‫= ݔ ܣ‬λ ‫ݔ‬

which can be rewritten as:

( ‫ ܣ‬−λ ‫ = ݔ ) ܫ‬0 Equation 1

The latter equation has nonzero solutions if and only if λ is chosen so that
det( ‫ ܣ‬− λ ‫ = ) ܫ‬0 Equation 2
Eigenvalue Problems
Equation 2 is a polynomial equation of degree n and is called the characteristic equation of the matrix [A].
Values of λ that satisfy Equation 2 may be either real- or complex-valued and are called eigenvalues of [A].
The nonzero solutions {x} of Equation 1 that are obtained by using such a value of λ are called the
eigenvectors corresponding to that eigenvalue.

Eigenvectors can be determined only up to an arbitrary nonzero multiplicative constant; if this constant is
specified in some way, then the eigenvectors are said to be normalized. Sometimes it is useful to normalize
an eigenvector {x} by choosing a constant so that its length ‫ݔ = ݔ‬, ‫ ݔ‬ଵ/ଶ = 1.

A matrix of size n x n has at least one and at most n eigenvalues, because the determinant yields a
polynomial of size n for λ. However, this approach to the solution of eigenvalue problem is not feasible, so
alternative numerical methods, such as the power method is usually employed.
Power Method
The power method is an iterative approach that can be employed to determine the largest eigenvalue. With
slight modification, it can also be employed to determine the smallest and the intermediate values. It has the
additional benefit that the corresponding eigenvector is obtained as a by-product of the method.

For the recursive formula, Equation 1 can be expressed as:

‫ݔ ܫ ߣ = ݔ ܣ‬

Since the identity matrix [I] has no effect on the product:

‫ݔ ߣ = ݔ ܣ‬

So a recursive relationship similar to that of fixed-point iteration is built such that:

‫ݔ ܣ‬ ௜ = ߣ௜ାଵ ‫ݔ‬ ௜ାଵ Equation 3

Note that, ‫ ݔ‬is dependent on λ, so a consistent normalization of elements of ‫ ݔ‬is necessary. Hence λ is
taken as the maximum element in the result of the product ‫ ݔ ܣ‬௜ .
Power Method
The power method is an iterative approach that can be employed to determine the largest eigenvalue. With
slight modification, it can also be employed to determine the smallest and the intermediate values. It has the
additional benefit that the corresponding eigenvector is obtained as a by-product of the method.

The power method can yield the minimum eigenvalue of [A] as well, after the following steps.

1
‫ݔ = ݔ ܣ‬
λ
λത ‫ܣ = ݔ‬ ିଵ ‫ݔ‬
1
‫ܣ‬ ିଵ ‫ݔ‬ ௜ = λത௜ାଵ ‫ݔ‬ ௜ାଵ where λത =
λ

In this case the power method will give the maximum λത which is equal to the inverse of minimum λ of [A].
The corresponding eigenvector stays unaffected.

The iterations for power method can be stopped when the change in λ is less than a reasonable percentage.
Example:
Find the maximum and minimum eigenvalues of [A] by using the power method. Use the error criteria
of 2 % .
8 3 0
‫= ܣ‬ 1 4 2
0 2 6
Solution for λmax
A reasonable choice for x is 1 1 1 T
଴

First Iteration:
8 3 0 1 11 1.000
1 4 2 1 = 7 = 11.0 0.636
0 2 6 1 8 0.727

1.000
λଵ = 11 ‫ݔ‬ ଵ = 0.636
0.727
Example:

Second Iteration:
8 3 0 1.000 9.91 1.000 9.91 − 11.0
1 4 2 0.636 = 5.00 = 9.91 0.505 ߝ௔ = = 11.0%
9.91
0 2 6 0.727 5.64 0.569

Third Iteration:

8 3 0 1.000 9.51 1.000 9.51 − 9.91

ߝ௔ = = 4.2%
1 4 2 0.505 = 4.16 = 9.51 0.437 9.51
0 2 6 0.569 4.42 0.465

Fourth Iteration:

8 3 0 1.000 9.31 1.000 9.31 − 9.51

1 4 2 0.437 = 3.68 = 9.31 0.395 ߝ௔ = = 2.4%
9.31
0 2 6 0.465 3.66 0.393
Example:
Fifth Iteration:

8 3 0 1.000 9.19 1.000

1 4 2 0.395 = 3.37 = 9.19 0.367 9.19 − 9.31
ߝ௔ = = 1.4%
0 2 6 0.393 3.15 0.343 9.19

So the maximum eigenvalue is approximately 9.19 with the eigenvector x = 1.000 0.367 0.343 T
Solution for λmin

0.1408 −0.1268 0.0423

‫ܣ‬ ିଵ = −0.0423 0.3380 −0.1127
00.0141 −0.1127 0.2042
ିଵ
With the recursive equation: ‫ܣ‬ ‫ݔ‬ ௜ = ߣ௜ାଵ ‫ݔ‬ ௜ାଵ

First Iteration:

0.1408 −0.1268 0.0423 1 0.308

−0.0423 0.3380 −0.1127 1 = 0.183 1.000
00.0141 −0.1127 0.2042 1 0.577

Second Iteration:
0.1408 −0.1268 0.0423 0.308 −0.227
−0.0423 0.3380 −0.1127 1.000 = 0.260 1.000 ߝ௔ = 29.6%
00.0141 −0.1127 0.2042 0.577 0.036

Third Iteration:
0.1408 −0.1268 0.0423 −0.227 −0.458
−0.0423 0.3380 −0.1127 1.000 = 0.344 1.000 ߝ௔ = 24.3%
00.0141 −0.1127 0.2042 0.036 −0.316
Fourth Iteration:
0.1408 −0.1268 0.0423 −0.458 −0.521
−0.0423 0.3380 −0.1127 1.000 = 0.393 1.000 ߝ௔ = 12.6%
00.0141 −0.1127 0.2042 −0.316 −0.467

Fifth Iteration:
0.1408 −0.1268 0.0423 −0.521 −0.533
−0.0423 0.3380 −0.1127 1.000 = 0.413 1.000 ߝ௔ = 4.8%
00.0141 −0.1127 0.2042 −0.467 −0.522

Sixth Iteration:
0.1408 −0.1268 0.0423 −0.533 −0.534
−0.0423 0.3380 −0.1127 1.000 = 0.419 1.000 ߝ௔ = 1.6%
00.0141 −0.1127 0.2042 −0.522 −0.541

So the minimum eigenvalue is approximately 1 / 0.419 = 2.39 with the eigenvector

x = −0.523 1.000 − 0.541 T
The power method can determine the intermediate λs if λmax is removed from A . A method was
proposed for this purpose by Hoteling (1933) for symmetric matrices.

Suppose that λ1 is the maximum eigenvalue of A 1 that has the eigenvector x ଵ . Then:
୘
xଵ x ଵ
A ଶ = A ଵ − λଵ
x ଵ୘ x ଵ
The power method yields the maximum eigenvalue of A ଶ which is equal to the second maximum
eigenvalue of A ଵ

The third eigenvalue is calculated by:

୘
x ଶ x ଶ
A ଷ = A ଶ − λଶ ୘
x ଶ x ଶ

Because of the accumulating error among successive eigenvalues, this method is preferred in case only
a number of predominant eigenvalues are necessary.