Deflation describes a technique from numerical mathematics with which a matrix is formed into a block triangular shape so that the spectrum of is just the union of the spectra of the diagonal blocks.
A.
∈
C.
n
×
n
{\ displaystyle A \ in \ mathbb {C} ^ {n \ times n}}
A.
{\ displaystyle A}
Deflation principle
Let be an endomorphism and the associated mapping matrix . By changing the base , this matrix can be converted into a matrix of the form
F.
∈
End
(
V
)
{\ displaystyle F \ in \ operatorname {End} (V)}
A.
∈
C.
n
×
n
{\ displaystyle A \ in \ mathbb {C} ^ {n \ times n}}
B.
{\ displaystyle B}
B.
:
=
(
B.
11
B.
12
0
B.
22nd
)
{\ displaystyle B \ colon {=} {\ begin {pmatrix} B_ {11} & B_ {12} \\ 0 & B_ {22} \ end {pmatrix}}}
to be transformed with for and . The following applies to
the spectra
B.
i
i
∈
C.
k
i
×
k
i
{\ displaystyle B_ {ii} \ in \ mathbb {C} ^ {k_ {i} \ times k_ {i}}}
i
∈
{
1
,
2
}
{\ displaystyle i \ in \ {1,2 \}}
k
1
+
k
2
=
n
{\ displaystyle k_ {1} + k_ {2} = n}
σ
(
B.
i
i
)
{\ displaystyle \ sigma (B_ {ii})}
σ
(
A.
)
=
σ
(
B.
11
)
∪
σ
(
B.
22nd
)
.
{\ displaystyle \ sigma (A) = \ sigma (B_ {11}) \ cup \ sigma (B_ {22}).}
Instead of the - eigenvalue problem , one can use the two smaller eigenvalue problems
(
n
×
n
)
{\ displaystyle (n \ times n)}
A.
x
=
λ
x
{\ displaystyle Ax = \ lambda x}
B.
i
i
y
=
λ
y
,
i
=
1
,
2
{\ displaystyle B_ {ii} y = \ lambda y, \ quad i = 1,2}
to solve. This method can be continued iteratively.
Deflation through Similarity Transformation
Theoretical basis
Let be a square matrix and an eigenpair of consisting of the eigenvalue and an associated eigenvector . This own pair can be obtained, for example, using the power method . The matrix is now using the similarity transformation
A.
∈
C.
n
×
n
{\ displaystyle A \ in \ mathbb {C} ^ {n \ times n}}
(
λ
,
v
)
{\ displaystyle (\ lambda, v)}
A.
{\ displaystyle A}
λ
∈
C.
{\ displaystyle \ lambda \ in \ mathbb {C}}
v
∈
C.
n
{\ displaystyle v \ in \ mathbb {C} ^ {n}}
A.
{\ displaystyle A}
B.
:
=
T
-
1
A.
T
{\ displaystyle B \ colon {=} T ^ {- 1} AT}
transferred to a matrix . The transformation matrix is given by with where the identity matrix is and . This particular base transformation is a household transformation . Hence, and the matrix has the form
B.
{\ displaystyle B}
T
{\ displaystyle T}
T
:
=
I.
-
2
w
w
T
w
T
w
{\ displaystyle T \ colon {=} I-2 {\ tfrac {ww ^ {T}} {w ^ {T} w}}}
w
=
v
+
‖
v
‖
2
e
1
,
{\ displaystyle w = v + \ | v \ | _ {2} e_ {1},}
I.
{\ displaystyle I}
e
1
:
=
(
1
0
...
0
)
T
{\ displaystyle e_ {1} \ colon {=} {\ begin {pmatrix} 1 & 0 & \ ldots & 0 \ end {pmatrix}} ^ {T}}
T
=
T
-
1
{\ displaystyle T = T ^ {- 1}}
B.
{\ displaystyle B}
B.
=
(
λ
b
t
0
B.
1
)
.
{\ displaystyle B = {\ begin {pmatrix} \ lambda & b ^ {t} \\ 0 & B_ {1} \ end {pmatrix}}.}
This matrix has the same eigenvalues as the matrix . Now you can apply the power method to the matrix again and get all eigenvalues iteratively.
A.
{\ displaystyle A}
B.
1
{\ displaystyle B_ {1}}
Numerical example
Be
A.
=
(
1
-
1
3
4th
2
1
3
1
9
)
{\ displaystyle A = {\ begin {pmatrix} 1 & -1 & 3 \\ 4 & 2 & 1 \\ 3 & 1 & 9 \ end {pmatrix}}}
With the power method one obtains as an own pair of . The transformation matrix is now calculated . It is
(
λ
1
,
v
)
=
(
10.22459
,
(
0.2585012
0.3343480
0.9063049
)
T
)
{\ displaystyle (\ lambda _ {1}, v) = \ left (10.22459, {\ begin {pmatrix} 0.2585012 & 0.3343480 & 0.9063049 \ end {pmatrix}} ^ {T} \ right)}
A.
{\ displaystyle A}
T
{\ displaystyle T}
T
=
I.
-
2
w
w
T
w
T
w
{\ displaystyle T = I-2 {\ frac {ww ^ {T}} {w ^ {T} w}}}
,
where is.
w
=
v
+
‖
v
‖
2
e
1
{\ displaystyle w = v + \ | v \ | _ {2} e_ {1}}
You get
T
=
(
-
0.258501
-
0.3343480
-
0.9063049
-
0.3343480
0.9111732
-
0.2407795
-
0.9063049
-
0.2407795
0.3473280
)
{\ displaystyle T = {\ begin {pmatrix} -0.258501 & -0.3343480 & -0.9063049 \\ - 0.3343480 & 0.9111732 & -0.2407795 \\ - 0.9063049 & -0.2407795 & 0.3473280 \ end {pmatrix}}}
and thus
T
A.
T
=
(
10.22459
3.5492494
0.5352000
0
-
1.5051646
-
2.3002829
0
1.7142389
0.2805751
)
{\ displaystyle TAT = {\ begin {pmatrix} 10.22459 & 3.5492494 & 0.5352000 \\ 0 & -1.5051646 & -2.3002829 \\ 0 & 1.7142389 & 0.2805751 \ end {pmatrix}}}
The eigenvalues of the matrix
C.
=
(
-
1.5051646
-
2.3002829
1.7142389
0.2805751
)
{\ displaystyle C = {\ begin {pmatrix} -1.5051646 & -2.3002829 \\ 1.7142389 & 0.2805751 \ end {pmatrix}}}
are and thus is
λ
2
=
-
0.6122947
+
1.7737021
i
{\ displaystyle \ lambda _ {2} = - 0.6122947 + 1.7737021i}
λ
3
=
-
0.6122947
-
1.7737021
i
{\ displaystyle \ lambda _ {3} = - 0.6122947-1.7737021i}
σ
(
A.
)
=
{
10.22459
,
-
0.6122947
+
1.7737021
i
,
-
0.6122947
-
1.7737021
i
}
{\ displaystyle \ sigma (A) = \ {10.22459, -0.6122947 + 1.7737021i, -0.6122947-1.7737021i \}}
literature
Martin Hanke-Bourgeois: Fundamentals of Numerical Mathematics and Scientific Computing. 1st edition, BG Teubner, Stuttgart 2002, ISBN 978-3-519-00356-4 .
Robert Schaback, Helmut Werner: Numerical Mathematics. Fourth completely revised edition, Springer Verlag Berlin Heidelberg GmbH, Berlin Heidelberg 1992, ISBN 978-3-540-54738-9 .
Willi Törnig: Numerical mathematics for engineers and physicists. Volume 1, Springer Verlag Berlin Heidelberg, Berlin Heidelberg 1979.
See also
Web links
<img src="https://de.wikipedia.org/wiki/Special:CentralAutoLogin/start?type=1x1" alt="" title="" width="1" height="1" style="border: none; position: absolute;">