Graphics Programs Reference
In-Depth Information
=
/
|
|
and repeat the process.Each iterationwill
increase the dominance of the first term in Eq. (9.29)sothat the process converges to
In subsequentcycles we set
v
z
z
1
λ
1
v
1
x
1
=
1
λ
1
x
1
z
=
(atthisstage
v
0).
The inverse powermethodalso works with the nonstandard eigenvalue problem
=
x
1
,sothat
v
1
=
1,
v
2
=
v
3
=···=
Ax
=
λ
Bx
(9.30)
provided that Eq. (9.28) is replacedby
Az
=
Bv
(9.31)
The alternative is, of course,totransform the problem to standard formbefore apply-
ing the powermethod.
Eigenvalue Shifting
By inspection of Eq. (9.29) we see that the rate of convergence is determinedbythe
strength of the inequality
|
λ
1
/λ
2
|
<
1 (the second term in the equation). If
|
λ
2
|
is well
|
λ
1
|
separated from
, the inequalityisstrong and the convergence is rapid. On the other
hand, close proximity of these twoeigenvalues results in very slow convergence.
The rate of convergence can be improvedbyatechniquecalled
eigenvalue
shifting
. If we let
λ
=
λ
∗
+
s
(9.32)
where
s
is apredetermined“shift,” the eigenvalue probleminEq. (9.27) istrans-
formed to
λ
∗
+
=
Ax
(
s
)
x
or
A
∗
x
=
λ
∗
x
(9.33)
where
A
∗
=
A
−
s
I
(9.34)
λ
1
Solving the transformedprobleminEq. (9.33) by the inverse powermethodyields
λ
1
is the smallest eigenvalueof
A
∗
. The corresponding eigenvalueofthe
original problem,
and
x
1
, where
λ
=
λ
1
+
s
, isthus the
eigenvalue closest to s
.
Eigenvalueshifting hastwo applications.Anobviousone is the determination of
the eigenvalue closest to a certain value
s
.For example, if the working speed of a shaft
Search WWH ::
Custom Search