Page 1
HIERARCHICAL STRUCTURES
MAX PLANCK INSTITUTE
FOR DYNAMICS OF COMPLEX
TECHNICAL SYSTEMS
MAGDEBURG
November 5, 2010
Hierarchical Matrices:Concept, Applications and Eigenvalues
Thomas Mach
Max Planck Institute for Dynamics of Complex Technical SystemsComputational Methods in Systems and Control Theory
Max Planck Institute Magdeburg Thomas Mach, Hierarchical Matrices 1/28
Page 2
HIERARCHICAL STRUCTURES
MAX PLANCK INSTITUTE
FOR DYNAMICS OF COMPLEX
TECHNICAL SYSTEMS
MAGDEBURG
November 5, 2010
Hierarchical Matrices:Concept, Applications and Eigenvalues
Thomas Mach
Max Planck Institute for Dynamics of Complex Technical SystemsComputational Methods in Systems and Control Theory
Max Planck Institute Magdeburg Thomas Mach, Hierarchical Matrices 1/28
Page 3
HIERARCHICAL STRUCTURES
MAX PLANCK INSTITUTE
FOR DYNAMICS OF COMPLEX
TECHNICAL SYSTEMS
MAGDEBURG
November 5, 2010
Hierarchical Matrices:Concept, Applications and Eigenvalues
Thomas Mach
Max Planck Institute for Dynamics of Complex Technical SystemsComputational Methods in Systems and Control Theory
Max Planck Institute Magdeburg Thomas Mach, Hierarchical Matrices 1/28
Page 4
HIERARCHICAL STRUCTURES
MAX PLANCK INSTITUTE
FOR DYNAMICS OF COMPLEX
TECHNICAL SYSTEMS
MAGDEBURG
November 5, 2010
Hierarchical Matrices:Concept, Applications and Eigenvalues
Thomas Mach
Max Planck Institute for Dynamics of Complex Technical SystemsComputational Methods in Systems and Control Theory
Max Planck Institute Magdeburg Thomas Mach, Hierarchical Matrices 1/28
Page 5
Concept Application Eigenvalues
Householder Notation
λ ∈ R
Scalar
Max Planck Institute Magdeburg Thomas Mach, Hierarchical Matrices 2/28
Page 6
Concept Application Eigenvalues
Householder Notation
x =
x1
x2
x3...xn
∈ Rn
Vector
Max Planck Institute Magdeburg Thomas Mach, Hierarchical Matrices 3/28
Page 7
Concept Application Eigenvalues
Householder Notation
A =
a11 a12 a13 . . . a1n
a21 a22 a23 . . . a2n
a31 a32 a33 . . . a3n...
......
. . ....
an1 an2 an3 . . . ann
∈ Rn×n
Matrix
Max Planck Institute Magdeburg Thomas Mach, Hierarchical Matrices 4/28
Page 8
Concept Application Eigenvalues
Dense Matrix Format
Store n vectors (Fortran):
a11
a21
a31...
an1
a12
a22
a32...
an2
a13
a23
a33...
an3
· · ·a1n
a2n
a3n...
ann
n2 entries in the storage
Ax costs O(n2) flops
AB costs O(nδ) flops, δ ≥ 2.376 usually δ = 3
A−1 costs O(n3) flops
There is a constant c , so that Ax costs not more than cn2 flops
Landau Symbol
1 flop = (αβ + γ → γ)
Flops
expensive!
Max Planck Institute Magdeburg Thomas Mach, Hierarchical Matrices 5/28
Page 9
Concept Application Eigenvalues
Dense Matrix Format
Store n vectors (Fortran):
a11
a21
a31...
an1
a12
a22
a32...
an2
a13
a23
a33...
an3
· · ·a1n
a2n
a3n...
ann
n2 entries in the storage
Ax costs O(n2) flops
AB costs O(nδ) flops, δ ≥ 2.376 usually δ = 3
A−1 costs O(n3) flops
There is a constant c , so that Ax costs not more than cn2 flops
Landau Symbol
1 flop = (αβ + γ → γ)
Flops
expensive!
Max Planck Institute Magdeburg Thomas Mach, Hierarchical Matrices 5/28
Page 10
Concept Application Eigenvalues
Dense Matrix Format
Store n vectors (Fortran):
a11
a21
a31...
an1
a12
a22
a32...
an2
a13
a23
a33...
an3
· · ·a1n
a2n
a3n...
ann
n2 entries in the storage
Ax costs O(n2) flops
AB costs O(nδ) flops, δ ≥ 2.376 usually δ = 3
A−1 costs O(n3) flops
There is a constant c , so that Ax costs not more than cn2 flops
Landau Symbol
1 flop = (αβ + γ → γ)
Flops
expensive!
Max Planck Institute Magdeburg Thomas Mach, Hierarchical Matrices 5/28
Page 11
Concept Application Eigenvalues
Sparse Matrix Format
4 −1 0 −1 0 0 0 0 0−1 4 −1 0 −1 0 0 0 00 −1 4 0 0 −1 0 0 0−1 0 0 4 −1 0 −1 0 00 −1 0 −1 4 −1 0 −1 00 0 −1 0 −1 4 0 0 −10 0 0 −1 0 0 4 −1 00 0 0 0 −1 0 −1 4 −10 0 0 0 0 −1 0 −1 4
non-zero entries per row/column = c ≤ 5
Max Planck Institute Magdeburg Thomas Mach, Hierarchical Matrices 6/28
Page 12
Concept Application Eigenvalues
Sparse Matrix Format
4 −1 0 −1 0 0 0 0 0−1 4 −1 0 −1 0 0 0 00 −1 4 0 0 −1 0 0 0−1 0 0 4 −1 0 −1 0 00 −1 0 −1 4 −1 0 −1 00 0 −1 0 −1 4 0 0 −10 0 0 −1 0 0 4 −1 00 0 0 0 −1 0 −1 4 −10 0 0 0 0 −1 0 −1 4
non-zero entries per row/column = c ≤ 5
Max Planck Institute Magdeburg Thomas Mach, Hierarchical Matrices 6/28
Page 13
Concept Application Eigenvalues
Sparse Matrix Format
4 −1 −1 −1−1 −1 −1 4−1 4 −1 −1−1 −1 4 −14 −1 −1 −1−1 −1 −1 4−1 −1 4−1 4 −14 −1 −1
non-zero entries per row/column = c ≤ 5
Max Planck Institute Magdeburg Thomas Mach, Hierarchical Matrices 6/28
Page 14
Concept Application Eigenvalues
Sparse Matrix Format (Matlab)
Store vector of tripel: (1, 1, 4)(2, 1,−1)(4, 1,−1)(1, 2,−1)
...
O(n) entries in the storage (if #non-zeros/column < c n)
Ax costs O(n) flops
AB may be much denser ⇒ better use A(Bx)
A−1 not possible, only solve Ax = b
Max Planck Institute Magdeburg Thomas Mach, Hierarchical Matrices 7/28
Page 15
Concept Application Eigenvalues
Is there something in between?
Yes, e.g.
Low rank matrices and Tucker tensor format
Toeplitz/Hankel matrices
Semiseparable matrices [Gantmacher, Krein 1937]
Cauchy matrices
Fast Multipole Method (FMM) [Greengard, Rokhlin ’87]
Mosaic-skeleton matrices [Tyrtyshnikov ’96]
...
Max Planck Institute Magdeburg Thomas Mach, Hierarchical Matrices 8/28
Page 16
Concept Application Eigenvalues
Is there something in between?
Yes, e.g.
Low rank matrices and Tucker tensor format
Toeplitz/Hankel matrices
Semiseparable matrices [Gantmacher, Krein 1937]
Cauchy matrices
Fast Multipole Method (FMM) [Greengard, Rokhlin ’87]
Mosaic-skeleton matrices [Tyrtyshnikov ’96]
...
Max Planck Institute Magdeburg Thomas Mach, Hierarchical Matrices 8/28
Page 17
Concept Application Eigenvalues
Is there something in between?
Yes, e.g.
Low rank matrices and Tucker tensor format
Toeplitz/Hankel matrices
Semiseparable matrices [Gantmacher, Krein 1937]
Cauchy matrices
Fast Multipole Method (FMM) [Greengard, Rokhlin ’87]
Mosaic-skeleton matrices [Tyrtyshnikov ’96]
Hierarchical (H-) Matrices [Hackbusch ’98]
Max Planck Institute Magdeburg Thomas Mach, Hierarchical Matrices 8/28
Page 18
Concept Application Eigenvalues
Integral Equation
Fredholm equation of first kind:∫Ω g(x , y)u(y)dy = f (x)
electrostatic problem:
g(x , y) = 14πε‖x−y‖
u(y) charge densityf (x) electrostatic potential
inverse of elliptic differential operators
population dynamics [Koch, Hackbusch, Sundmacher]
unknown
Max Planck Institute Magdeburg Thomas Mach, Hierarchical Matrices 9/28
Page 19
Concept Application Eigenvalues
Integral Equation
Fredholm equation of first kind:∫Ω g(x , y)u(y)dy = f (x)
electrostatic problem:
g(x , y) = 14πε‖x−y‖
u(y) charge densityf (x) electrostatic potential
inverse of elliptic differential operators
population dynamics [Koch, Hackbusch, Sundmacher]
unknown
Max Planck Institute Magdeburg Thomas Mach, Hierarchical Matrices 9/28
Page 20
Concept Application Eigenvalues
Integral Equation
Fredholm equation of first kind:∫Ω g(x , y)u(y)dy = f (x)
electrostatic problem:
g(x , y) = 14πε‖x−y‖
u(y) charge densityf (x) electrostatic potential
inverse of elliptic differential operators
population dynamics [Koch, Hackbusch, Sundmacher]
unknown
Max Planck Institute Magdeburg Thomas Mach, Hierarchical Matrices 9/28
Page 21
Concept Application Eigenvalues
H-Matrices [Hackbusch ’98]
∫Ω g(x , y)u(y)dy = f (x)
Ritz-Galerkin method:
u(y) =∑
i uiψi (y)
⇒∑i
∫Ω
∫Ωg(x , y)ψi (y)dyψj(x)dx︸ ︷︷ ︸
:=Mji
ui=
∫Ωf (x)ψj(x)dx︸ ︷︷ ︸
=fj
x ∈ Ωs , y ∈ Ωt : g(x , y) ≈∑kp=1 g
x ,sp (x)g y ,t
p (y)
Mji =
∫Ω
∫Ωg(x , y)ψj(x)ψi (y)dydx
≈k∑
p=1
∫Ωg xp (x)ψj(x)dx
∫Ωg yp (y)ψi (y)dy = Aj ·B
Ti ·
discretization error ε ∼ 1nκ , 0 < κ < 1
Mu = f
Max Planck Institute Magdeburg Thomas Mach, Hierarchical Matrices 10/28
Page 22
Concept Application Eigenvalues
H-Matrices [Hackbusch ’98]
∫Ω g(x , y)u(y)dy = f (x)
Ritz-Galerkin method:
u(y) =∑
i uiψi (y)
⇒∑i
∫Ω
∫Ωg(x , y)ψi (y)dyψj(x)dx︸ ︷︷ ︸
:=Mji
ui=
∫Ωf (x)ψj(x)dx︸ ︷︷ ︸
=fj
x ∈ Ωs , y ∈ Ωt : g(x , y) ≈∑kp=1 g
x ,sp (x)g y ,t
p (y)
Mji =
∫Ω
∫Ωg(x , y)ψj(x)ψi (y)dydx
≈k∑
p=1
∫Ωg xp (x)ψj(x)dx
∫Ωg yp (y)ψi (y)dy = Aj ·B
Ti ·
discretization error ε ∼ 1nκ , 0 < κ < 1
Mu = f
Max Planck Institute Magdeburg Thomas Mach, Hierarchical Matrices 10/28
Page 23
Concept Application Eigenvalues
H-Matrices [Hackbusch ’98]
∫Ω g(x , y)u(y)dy = f (x)
Ritz-Galerkin method:
u(y) =∑
i uiψi (y)
⇒∑i
∫Ω
∫Ωg(x , y)ψi (y)dyψj(x)dx︸ ︷︷ ︸
:=Mji
ui=
∫Ωf (x)ψj(x)dx︸ ︷︷ ︸
=fj
x ∈ Ωs , y ∈ Ωt : g(x , y) ≈∑kp=1 g
x ,sp (x)g y ,t
p (y)
Mji =
∫Ω
∫Ωg(x , y)ψj(x)ψi (y)dydx
≈k∑
p=1
∫Ωg xp (x)ψj(x)dx
∫Ωg yp (y)ψi (y)dy = Aj ·B
Ti ·
discretization error ε ∼ 1nκ , 0 < κ < 1
Mu = f
Max Planck Institute Magdeburg Thomas Mach, Hierarchical Matrices 10/28
Page 24
Concept Application Eigenvalues
H-Matrices [Hackbusch ’98]
∫Ω g(x , y)u(y)dy = f (x)
Ritz-Galerkin method:
u(y) =∑
i uiψi (y)
⇒∑i
∫Ω
∫Ωg(x , y)ψi (y)dyψj(x)dx︸ ︷︷ ︸
:=Mji
ui=
∫Ωf (x)ψj(x)dx︸ ︷︷ ︸
=fj
x ∈ Ωs , y ∈ Ωt : g(x , y) ≈∑kp=1 g
x ,sp (x)g y ,t
p (y)
Mji =
∫Ω
∫Ωg(x , y)ψj(x)ψi (y)dydx
≈k∑
p=1
∫Ωg xp (x)ψj(x)dx
∫Ωg yp (y)ψi (y)dy = Aj ·B
Ti ·
discretization error ε ∼ 1nκ , 0 < κ < 1
Mu = f
Max Planck Institute Magdeburg Thomas Mach, Hierarchical Matrices 10/28
Page 25
Concept Application Eigenvalues
H-Matrices [Hackbusch ’98]
∫Ω g(x , y)u(y)dy = f (x)
Ritz-Galerkin method:
u(y) =∑
i uiψi (y)
⇒∑i
∫Ω
∫Ωg(x , y)ψi (y)dyψj(x)dx︸ ︷︷ ︸
:=Mji
ui=
∫Ωf (x)ψj(x)dx︸ ︷︷ ︸
=fj
x ∈ Ωs , y ∈ Ωt : g(x , y) ≈∑kp=1 g
x ,sp (x)g y ,t
p (y)
Mji =
∫Ω
∫Ωg(x , y)ψj(x)ψi (y)dydx
≈k∑
p=1
∫Ωg xp (x)ψj(x)dx
∫Ωg yp (y)ψi (y)dy = Aj ·B
Ti ·
discretization error ε ∼ 1nκ , 0 < κ < 1
Mu = f
Max Planck Institute Magdeburg Thomas Mach, Hierarchical Matrices 10/28
Page 26
Concept Application Eigenvalues
H-Matrices [Hackbusch ’98]
22
3 3
7 10
3 7
3 10
19 10
10 31
14 8
11 11
14 11
8 11
19 10
10 31 11
11 31
11 9
9 16 12
11 1611 8
9 16 11
11 16
61
9 73 3
8
11 11
9 3
7 3
11
8 11
25 10
10 19 11
11 31
8 5
11 8
8 15 8 12
116 5
15 613
13
7 5
13 8
8 11 8 11
11
6 5
15 6
8
11 8
8 15
5 8 11
12
6 15
5 613
13
7
13 8
8 11
5 811
11
6 15
5 6
6110 103
6 14
10 3 6
10 14
25 10 6
10
6
19 10
10 31
15 9
9 1110
10 16
15 9
8 1110
10 16
5110 10
7 9 73 3
10 7
10
9 3
7 3
25 11
11
25 10
10 19
11 811 8
8 15 9
8 1512 13
13
10 7
13 8
8 11 9
9 15 1119
20
10 7
13 8
9
11 8
8 15 11 12
13
9 7
13 8
9
13 8
8 1111
10
11 8
8 15 8
9 15
7 12 13
13
10
13 8
8 11 8
9 15
7 1120
19
9
13 9
8
11 8
8 15
7 11 12
129
13 9
8
13 8
8 11
7 11
39 10
10 25
3 7
3 10 10
7 10 63 3
7 10 7
10
10
6
22
3 3
7 10
3 7
3 10
19 10
10 31
15 10
11
11 9
9 16
15 11
10
11 8
9 16
34 10
10 25
13 10
7 11
13 7
10 11 61
6 5
13 6 11
12
8 5
11 8
8 15 812
13
6 5
13 6 11
11 23
6 13
5 6 12
118
11 8
8 15
5 813
12
6 13
5 6 10
11 23
20 9
9 39
9 7
10
3 7
3 10
9 10
7
3 3
7 1061
15 10
1015 9
9 11
15 10
1015 9
8 11
20 9
9 34
9 7
10 13
9 10
7 1351
x ∈ Ωs , y ∈ Ωt : g(x, y) ≈k∑
p=1
gx,sp (x)gy,tp (y)
Mji =
∫Ω
∫Ωg(x, y)ψj (x)ψi (y)dydx
≈k∑
p=1
∫Ωgxp (x)ψj (x)dx
∫Ωgyp (y)ψi (y)dy = Aj·B
Ti·
rank (Ms×t) = 19store:
A ∈ R|s|×19,
B ∈ R|t|×19
with Ms×t = ABT
⇒ only 19(|s|+ |t|) instead of |s| |t| storage⇒ reduces the required storage to 15%
s
t
=
if x ≈ y then 1‖x−y‖ is large
small matrices on the diag-onal have no low rank ap-proximation⇒ use dense matrix format
non-admissible block
admissible block
Max Planck Institute Magdeburg Thomas Mach, Hierarchical Matrices 11/28
Page 27
Concept Application Eigenvalues
H-Matrices [Hackbusch ’98]
22
3 3
7 10
3 7
3 10
19 10
10 31
14 8
11 11
14 11
8 11
19 10
10 31 11
11 31
11 9
9 16 12
11 1611 8
9 16 11
11 16
61
9 73 3
8
11 11
9 3
7 3
11
8 11
25 10
10 19 11
11 31
8 5
11 8
8 15 8 12
116 5
15 613
13
7 5
13 8
8 11 8 11
11
6 5
15 6
8
11 8
8 15
5 8 11
12
6 15
5 613
13
7
13 8
8 11
5 811
11
6 15
5 6
6110 103
6 14
10 3 6
10 14
25 10 6
10
6
19 10
10 31
15 9
9 1110
10 16
15 9
8 1110
10 16
5110 10
7 9 73 3
10 7
10
9 3
7 3
25 11
11
25 10
10 19
11 811 8
8 15 9
8 1512 13
13
10 7
13 8
8 11 9
9 15 1119
20
10 7
13 8
9
11 8
8 15 11 12
13
9 7
13 8
9
13 8
8 1111
10
11 8
8 15 8
9 15
7 12 13
13
10
13 8
8 11 8
9 15
7 1120
19
9
13 9
8
11 8
8 15
7 11 12
129
13 9
8
13 8
8 11
7 11
39 10
10 25
3 7
3 10 10
7 10 63 3
7 10 7
10
10
6
22
3 3
7 10
3 7
3 10
19 10
10 31
15 10
11
11 9
9 16
15 11
10
11 8
9 16
34 10
10 25
13 10
7 11
13 7
10 11 61
6 5
13 6 11
12
8 5
11 8
8 15 812
13
6 5
13 6 11
11 23
6 13
5 6 12
118
11 8
8 15
5 813
12
6 13
5 6 10
11 23
20 9
9 39
9 7
10
3 7
3 10
9 10
7
3 3
7 1061
15 10
1015 9
9 11
15 10
1015 9
8 11
20 9
9 34
9 7
10 13
9 10
7 1351
x ∈ Ωs , y ∈ Ωt : g(x, y) ≈k∑
p=1
gx,sp (x)gy,tp (y)
Mji =
∫Ω
∫Ωg(x, y)ψj (x)ψi (y)dydx
≈k∑
p=1
∫Ωgxp (x)ψj (x)dx
∫Ωgyp (y)ψi (y)dy = Aj·B
Ti·
rank (Ms×t) = 19store:
A ∈ R|s|×19,
B ∈ R|t|×19
with Ms×t = ABT
⇒ only 19(|s|+ |t|) instead of |s| |t| storage⇒ reduces the required storage to 15%
s
t
=if x ≈ y then 1
‖x−y‖ is largesmall matrices on the diag-onal have no low rank ap-proximation⇒ use dense matrix format
non-admissible block
admissible block
Max Planck Institute Magdeburg Thomas Mach, Hierarchical Matrices 11/28
Page 28
Concept Application Eigenvalues
H-Matrices [Hackbusch ’98]
22
3 3
7 10
3 7
3 10
19 10
10 31
14 8
11 11
14 11
8 11
19 10
10 31 11
11 31
11 9
9 16 12
11 1611 8
9 16 11
11 16
61
9 73 3
8
11 11
9 3
7 3
11
8 11
25 10
10 19 11
11 31
8 5
11 8
8 15 8 12
116 5
15 613
13
7 5
13 8
8 11 8 11
11
6 5
15 6
8
11 8
8 15
5 8 11
12
6 15
5 613
13
7
13 8
8 11
5 811
11
6 15
5 6
6110 103
6 14
10 3 6
10 14
25 10 6
10
6
19 10
10 31
15 9
9 1110
10 16
15 9
8 1110
10 16
5110 10
7 9 73 3
10 7
10
9 3
7 3
25 11
11
25 10
10 19
11 811 8
8 15 9
8 1512 13
13
10 7
13 8
8 11 9
9 15 1119
20
10 7
13 8
9
11 8
8 15 11 12
13
9 7
13 8
9
13 8
8 1111
10
11 8
8 15 8
9 15
7 12 13
13
10
13 8
8 11 8
9 15
7 1120
19
9
13 9
8
11 8
8 15
7 11 12
129
13 9
8
13 8
8 11
7 11
39 10
10 25
3 7
3 10 10
7 10 63 3
7 10 7
10
10
6
22
3 3
7 10
3 7
3 10
19 10
10 31
15 10
11
11 9
9 16
15 11
10
11 8
9 16
34 10
10 25
13 10
7 11
13 7
10 11 61
6 5
13 6 11
12
8 5
11 8
8 15 812
13
6 5
13 6 11
11 23
6 13
5 6 12
118
11 8
8 15
5 813
12
6 13
5 6 10
11 23
20 9
9 39
9 7
10
3 7
3 10
9 10
7
3 3
7 1061
15 10
1015 9
9 11
15 10
1015 9
8 11
20 9
9 34
9 7
10 13
9 10
7 1351
x ∈ Ωs , y ∈ Ωt : g(x, y) ≈k∑
p=1
gx,sp (x)gy,tp (y)
Mji =
∫Ω
∫Ωg(x, y)ψj (x)ψi (y)dydx
≈k∑
p=1
∫Ωgxp (x)ψj (x)dx
∫Ωgyp (y)ψi (y)dy = Aj·B
Ti·
rank (Ms×t) = 19store:
A ∈ R|s|×19,
B ∈ R|t|×19
with Ms×t = ABT
⇒ only 19(|s|+ |t|) instead of |s| |t| storage⇒ reduces the required storage to 15%
s
t
=
if x ≈ y then 1‖x−y‖ is large
small matrices on the diag-onal have no low rank ap-proximation⇒ use dense matrix format
non-admissible block
admissible block
Max Planck Institute Magdeburg Thomas Mach, Hierarchical Matrices 11/28
Page 29
Concept Application Eigenvalues
H-Matrices [Hackbusch ’98]
22
3 3
7 10
3 7
3 10
19 10
10 31
14 8
11 11
14 11
8 11
19 10
10 31 11
11 31
11 9
9 16 12
11 1611 8
9 16 11
11 16
61
9 73 3
8
11 11
9 3
7 3
11
8 11
25 10
10 19 11
11 31
8 5
11 8
8 15 8 12
116 5
15 613
13
7 5
13 8
8 11 8 11
11
6 5
15 6
8
11 8
8 15
5 8 11
12
6 15
5 613
13
7
13 8
8 11
5 811
11
6 15
5 6
6110 103
6 14
10 3 6
10 14
25 10 6
10
6
19 10
10 31
15 9
9 1110
10 16
15 9
8 1110
10 16
5110 10
7 9 73 3
10 7
10
9 3
7 3
25 11
11
25 10
10 19
11 811 8
8 15 9
8 1512 13
13
10 7
13 8
8 11 9
9 15 1119
20
10 7
13 8
9
11 8
8 15 11 12
13
9 7
13 8
9
13 8
8 1111
10
11 8
8 15 8
9 15
7 12 13
13
10
13 8
8 11 8
9 15
7 1120
19
9
13 9
8
11 8
8 15
7 11 12
129
13 9
8
13 8
8 11
7 11
39 10
10 25
3 7
3 10 10
7 10 63 3
7 10 7
10
10
6
22
3 3
7 10
3 7
3 10
19 10
10 31
15 10
11
11 9
9 16
15 11
10
11 8
9 16
34 10
10 25
13 10
7 11
13 7
10 11 61
6 5
13 6 11
12
8 5
11 8
8 15 812
13
6 5
13 6 11
11 23
6 13
5 6 12
118
11 8
8 15
5 813
12
6 13
5 6 10
11 23
20 9
9 39
9 7
10
3 7
3 10
9 10
7
3 3
7 1061
15 10
1015 9
9 11
15 10
1015 9
8 11
20 9
9 34
9 7
10 13
9 10
7 1351
x ∈ Ωs , y ∈ Ωt : g(x, y) ≈k∑
p=1
gx,sp (x)gy,tp (y)
Mji =
∫Ω
∫Ωg(x, y)ψj (x)ψi (y)dydx
≈k∑
p=1
∫Ωgxp (x)ψj (x)dx
∫Ωgyp (y)ψi (y)dy = Aj·B
Ti·
rank (Ms×t) = 19store:
A ∈ R|s|×19,
B ∈ R|t|×19
with Ms×t = ABT
⇒ only 19(|s|+ |t|) instead of |s| |t| storage⇒ reduces the required storage to 15%
s
t
=
if x ≈ y then 1‖x−y‖ is large
small matrices on the diag-onal have no low rank ap-proximation⇒ use dense matrix format
non-admissible block
admissible block
Max Planck Institute Magdeburg Thomas Mach, Hierarchical Matrices 11/28
Page 30
Concept Application Eigenvalues
H-Matrices [Hackbusch ’98]
22
3 3
7 10
3 7
3 10
19 10
10 31
14 8
11 11
14 11
8 11
19 10
10 31 11
11 31
11 9
9 16 12
11 1611 8
9 16 11
11 16
61
9 73 3
8
11 11
9 3
7 3
11
8 11
25 10
10 19 11
11 31
8 5
11 8
8 15 8 12
116 5
15 613
13
7 5
13 8
8 11 8 11
11
6 5
15 6
8
11 8
8 15
5 8 11
12
6 15
5 613
13
7
13 8
8 11
5 811
11
6 15
5 6
6110 103
6 14
10 3 6
10 14
25 10 6
10
6
19 10
10 31
15 9
9 1110
10 16
15 9
8 1110
10 16
5110 10
7 9 73 3
10 7
10
9 3
7 3
25 11
11
25 10
10 19
11 811 8
8 15 9
8 1512 13
13
10 7
13 8
8 11 9
9 15 1119
20
10 7
13 8
9
11 8
8 15 11 12
13
9 7
13 8
9
13 8
8 1111
10
11 8
8 15 8
9 15
7 12 13
13
10
13 8
8 11 8
9 15
7 1120
19
9
13 9
8
11 8
8 15
7 11 12
129
13 9
8
13 8
8 11
7 11
39 10
10 25
3 7
3 10 10
7 10 63 3
7 10 7
10
10
6
22
3 3
7 10
3 7
3 10
19 10
10 31
15 10
11
11 9
9 16
15 11
10
11 8
9 16
34 10
10 25
13 10
7 11
13 7
10 11 61
6 5
13 6 11
12
8 5
11 8
8 15 812
13
6 5
13 6 11
11 23
6 13
5 6 12
118
11 8
8 15
5 813
12
6 13
5 6 10
11 23
20 9
9 39
9 7
10
3 7
3 10
9 10
7
3 3
7 1061
15 10
1015 9
9 11
15 10
1015 9
8 11
20 9
9 34
9 7
10 13
9 10
7 1351
x ∈ Ωs , y ∈ Ωt : g(x, y) ≈k∑
p=1
gx,sp (x)gy,tp (y)
Mji =
∫Ω
∫Ωg(x, y)ψj (x)ψi (y)dydx
≈k∑
p=1
∫Ωgxp (x)ψj (x)dx
∫Ωgyp (y)ψi (y)dy = Aj·B
Ti·
rank (Ms×t) = 19store:
A ∈ R|s|×19,
B ∈ R|t|×19
with Ms×t = ABT
⇒ only 19(|s|+ |t|) instead of |s| |t| storage⇒ reduces the required storage to 15%
s
t
=
if x ≈ y then 1‖x−y‖ is large
small matrices on the diag-onal have no low rank ap-proximation⇒ use dense matrix format
non-admissible block
admissible block
Max Planck Institute Magdeburg Thomas Mach, Hierarchical Matrices 11/28
Page 31
Concept Application Eigenvalues
H-Matrices [Hackbusch ’98]
rank-k-matrix: Ms×t = ABT , A ∈ Rn×k ,B ∈ Rm×k (k n,m)
hierarchical tree TI block H-tree TI× I
I = 1, 2, 3, 4, 5, 6, 7, 8
1, 2, 3, 4 5, 6, 7, 8
1, 2 3, 4 5, 6 7, 8
12345678
1 1 1 12 2 2 23 3 3 34 4 4 45 5 5 56 6 6 67 7 7 78 8 8 8
1 1 1 12 2 2 23 3 3 34 4 4 45 5 5 56 6 6 67 7 7 78 8 8 8
dense matrices, rank-k-matrices
Max Planck Institute Magdeburg Thomas Mach, Hierarchical Matrices 12/28
Page 32
Concept Application Eigenvalues
H-Matrices [Hackbusch ’98]
Hierarchical matrices
H(TI× I, k) =M ∈ RI× I
∣∣ rank (Ms×t) ≤ k ∀s × t admissible
22
3 3
7 10
3 7
3 10
19 10
10 31
14 8
11 11
14 11
8 11
19 10
10 31 11
11 31
11 9
9 16 12
11 1611 8
9 16 11
11 16
61
9 73 3
8
11 11
9 3
7 3
11
8 11
25 10
10 19 11
11 31
8 5
11 8
8 15 8 12
116 5
15 613
13
7 5
13 8
8 11 8 11
11
6 5
15 6
8
11 8
8 15
5 8 11
12
6 15
5 613
13
7
13 8
8 11
5 811
11
6 15
5 6
6110 103
6 14
10 3 6
10 14
25 10 6
10
6
19 10
10 31
15 9
9 1110
10 16
15 9
8 1110
10 16
5110 10
7 9 73 3
10 7
10
9 3
7 3
25 11
11
25 10
10 19
11 811 8
8 15 9
8 1512 13
13
10 7
13 8
8 11 9
9 15 1119
20
10 7
13 8
9
11 8
8 15 11 12
13
9 7
13 8
9
13 8
8 1111
10
11 8
8 15 8
9 15
7 12 13
13
10
13 8
8 11 8
9 15
7 1120
19
9
13 9
8
11 8
8 15
7 11 12
129
13 9
8
13 8
8 11
7 11
39 10
10 25
3 7
3 10 10
7 10 63 3
7 10 7
10
10
6
22
3 3
7 10
3 7
3 10
19 10
10 31
15 10
11
11 9
9 16
15 11
10
11 8
9 16
34 10
10 25
13 10
7 11
13 7
10 11 61
6 5
13 6 11
12
8 5
11 8
8 15 812
13
6 5
13 6 11
11 23
6 13
5 6 12
118
11 8
8 15
5 813
12
6 13
5 6 10
11 23
20 9
9 39
9 7
10
3 7
3 10
9 10
7
3 3
7 1061
15 10
1015 9
9 11
15 10
1015 9
8 11
20 9
9 34
9 7
10 13
9 10
7 1351
adaptive rank k(ε)
storage NSt,H(T , k) = O(n log n k(ε))
complexity of approximate arithmetic
MHv O(n log n k(ε))+H,−H O(n log n k(ε)2)
∗H,HLU(·), (·)−1H O(n (log n)2 k(ε)2)
Max Planck Institute Magdeburg Thomas Mach, Hierarchical Matrices 13/28
Page 33
Concept Application Eigenvalues
H-Matrices [Hackbusch ’98]
Hierarchical matrices
H(TI× I, k) =M ∈ RI× I
∣∣ rank (Ms×t) ≤ k ∀s × t admissible
22
3 3
7 10
3 7
3 10
19 10
10 31
14 8
11 11
14 11
8 11
19 10
10 31 11
11 31
11 9
9 16 12
11 1611 8
9 16 11
11 16
61
9 73 3
8
11 11
9 3
7 3
11
8 11
25 10
10 19 11
11 31
8 5
11 8
8 15 8 12
116 5
15 613
13
7 5
13 8
8 11 8 11
11
6 5
15 6
8
11 8
8 15
5 8 11
12
6 15
5 613
13
7
13 8
8 11
5 811
11
6 15
5 6
6110 103
6 14
10 3 6
10 14
25 10 6
10
6
19 10
10 31
15 9
9 1110
10 16
15 9
8 1110
10 16
5110 10
7 9 73 3
10 7
10
9 3
7 3
25 11
11
25 10
10 19
11 811 8
8 15 9
8 1512 13
13
10 7
13 8
8 11 9
9 15 1119
20
10 7
13 8
9
11 8
8 15 11 12
13
9 7
13 8
9
13 8
8 1111
10
11 8
8 15 8
9 15
7 12 13
13
10
13 8
8 11 8
9 15
7 1120
19
9
13 9
8
11 8
8 15
7 11 12
129
13 9
8
13 8
8 11
7 11
39 10
10 25
3 7
3 10 10
7 10 63 3
7 10 7
10
10
6
22
3 3
7 10
3 7
3 10
19 10
10 31
15 10
11
11 9
9 16
15 11
10
11 8
9 16
34 10
10 25
13 10
7 11
13 7
10 11 61
6 5
13 6 11
12
8 5
11 8
8 15 812
13
6 5
13 6 11
11 23
6 13
5 6 12
118
11 8
8 15
5 813
12
6 13
5 6 10
11 23
20 9
9 39
9 7
10
3 7
3 10
9 10
7
3 3
7 1061
15 10
1015 9
9 11
15 10
1015 9
8 11
20 9
9 34
9 7
10 13
9 10
7 1351
adaptive rank k(ε)
storage NSt,H(T , k) = O(n log n k(ε))
complexity of approximate arithmetic
MHv O(n log n k(ε))+H,−H O(n log n k(ε)2)
∗H,HLU(·), (·)−1H O(n (log n)2 k(ε)2)
A1
BT1
+ A2
BT2
= A1 A2
BT1
BT2
Max Planck Institute Magdeburg Thomas Mach, Hierarchical Matrices 13/28
Page 34
Concept Application Eigenvalues
Special Case: H`-Matrices
A8BT8
B8AT8
A4BT4
B4AT4
A12BT12
B12AT12
A2BT2
A6BT6
A10BT10
A14BT14
B2AT2
B6AT6
B10AT10
B14AT14
F1
F3
F5
F7
F9
F11
F13
F15
Figure: Structure of an H3(k)-matrix
Max Planck Institute Magdeburg Thomas Mach, Hierarchical Matrices 14/28
Page 35
Concept Application Eigenvalues
Applications
Max Planck Institute Magdeburg Thomas Mach, Hierarchical Matrices 15/28
Page 36
Concept Application Eigenvalues
Applications
Finite-Element-Method (FEM)
Inverse of FEM matrices
Integral equations
Solving matrix equations for model order reduction
AX + XAT + BBT = 0, A ∈ H,B ∈ Rn×k
Sign function iteration[Baur ’07, Grasedyck ’03, et al.]
Boundary-Element-Method (BEM)
Max Planck Institute Magdeburg Thomas Mach, Hierarchical Matrices 16/28
Page 37
Concept Application Eigenvalues
BEM for Electrostatic Computations
−4 −20
24
·103
−20
2·103
−4
−2
0
2
4
·103
Max Planck Institute Magdeburg Thomas Mach, Hierarchical Matrices 17/28
Page 38
Concept Application Eigenvalues
BEM for Electrostatic Computations
−4 −20
24
·103
−20
2·103
−4
−2
0
2
4
·103Electrostatic problem:∫
Ω g(x , y)u(y)dy = f (x)
g(x , y) = 14πε‖x−y‖
unknown: u(y) charge density
given: f (x) electrostatic potential
Max Planck Institute Magdeburg Thomas Mach, Hierarchical Matrices 17/28
Page 39
Concept Application Eigenvalues
BEM for Electrostatic Computations
−4 −20
24
·103
−20
2·103
−4
−2
0
2
4
·103
Max Planck Institute Magdeburg Thomas Mach, Hierarchical Matrices 17/28
Page 40
Concept Application Eigenvalues
BEM for Electrostatic Computations
Max Planck Institute Magdeburg Thomas Mach, Hierarchical Matrices 17/28
Page 41
Concept Application Eigenvalues
BEM for Electrostatic Computations
Max Planck Institute Magdeburg Thomas Mach, Hierarchical Matrices 18/28
Page 42
Concept Application Eigenvalues
BEM for Electrostatic Computations
−4 −20
24
·103
−20
2·103
−4
−2
0
2
4
·103
Max Planck Institute Magdeburg Thomas Mach, Hierarchical Matrices 18/28
Page 43
Concept Application Eigenvalues
BEM for Electrostatic Computations26 14
14 20 19
1925 14
14 20
23 16
17 23
23 17
16 23
22 20
20 22 22
22 36
13
15
17
18
13 15 17 18
19 15
15 21
20 16
19 25 8
11 18 8
20 19
16 25
11
188 8
25 14
14 20 16
16 26
22 15
13 15
22 13
15 15
36 19
19 288
8 8
17 12
27 19
19 26 20 2513 23 10 11 23 9
25 14
16 10 13
3016 10
26 16
16 18 17 16
8 8
1727 19
19 26
12 20
13
23
10
11
2523
9
30
25
16
10
14 13
1626 16
16 18
10 17 8
16 8
22 20
20 22 22
22 36
28 20
20 28
28 20
20 28
21 11
11 12 20
20 22 29
2928 18
18 18
18
36 22
32
21
28 22
17 18
18 36 32
22 21
28 17
22 18
19 13
13 15 23
2323 13
13 1525
25
20 12
12 15 23
2320 12
12 15
50
19 12
15 20
24 20
17 25
20 14
13 20
50
24 17
20 2519 15
12 20
20 13
14 20
36 19
19 28 40
40
26 18
18 18 23
2321 13
13 14
20 14
17 20
20 17
14 20
20 12
12 20
22
22 11
6 18 10 8
20
13 22
13
17
18 12 21
17 2224
8 6
22
19 8
23 9
17 10 9 7
20 43
19
23 10
6 16 10
6 11 9
15
10
27 19
16 19
6 19 15
17 8
15
34 17
13
25
16
20 19
16 17
20 18
17 24 19
20
20 17
18 20
48 12
27 13 18 12
18 12 27 13 10
23 14
19 24
20 13
14 20
36 26
20 24 33
4644 29
23 3024
20 16
17 20 11
22 20
22
6
18
10
11 8
13 13
22
17 18
12
17
21 228
24 6
19 15
23
6 6
16
10 11
10 9
10 6
27 16
19 1919 17
15 8
2219
8 20
23
17
10
9
9 7
43
15 34 13 25
17 1620 16
19 17
20 17
18 24 20
1920 18
17 20
23 19
14 24
20 14
13 20
48
27
13
18
12
18
12
27
13
12 10
36 20
26 2446
33
44 23
29 30
20 17
16 20
24 11
22 18
18
19 13
13 18 9 17
9
17
31 19
19 27 12
12 12
27
27
18 18
18
24 18
18
21 17
17
24 19
19 19
19 12
13 17 14
9
16
11
12
18 13
16 15
19 13
12 17 9 16
14 11 12
18 16
13 15
28 14
14 17 14
1418 12
12 15
15 15
18
8
10 6 6
35 25
24
13 16
17
10 10
12 13
12 27 17 19 19
8 9 6 6
11 19 9 14
15 18 8
15
10
35 24
25
13 17
16
10
12 12
13 27
10 17
6
619
8
9
11
19
6
69
19 14
20 17
17 24 19
1920 17
17 20
29 12 17 10
16 10 26 11 11
29
12
16
10
17
10
26
11
11
20 15
15 15 27
2723 15
15 15
15 17
15 18
15 15
17 18
15 14
14 18
49
49
31 31
31
15 15
15
27 20
20 2037
37
18 17
17
15 15
15 24
14 14
14 21
14 14
14 21
14 13
13 21
39 27
25 49
39 25
27 49
20 15
15 15 28
2820 15
15 1530
30 3045
45
17 16
16 18 35
35
15 15
15
28 16
16 16
13 13
12 2913 12
13 29
15 15
15
27 18
18 18
26 17
18 26
26 18
17 26
26 13
13 26
35
12 138 6 7
7
27 8
10 117 6 7 7
9
19 11
22
12
11 10
116 7
8 9 6
12 12 10
32 13
49
8 6
17 11
11 17 15
8
15
11
40
18 12
12 17 15 9
8
15
12 713
20 15
15 25 12
18 12
12 16 15
1318 14
12 15
24 13
13 17 15 9
1318 14
12 15 10
10 339 5
9
13
24
6
6
9
10
8
926
6
614
17 18
8
8
8
11 7
15
45
12
19
12
22
12
36
15 11
46 16 23
25
12 11
38 1726 15
14 26
26 16
16 26
14
35 4912 8
13
6
727
7 8
22
10
7
6
11
7
7
19
9 11
12 12
11 11
10
6 8
7
9
6
12 32
10 13
8
17 11
11 17 8 15
15 11 40
18 12
12 16 13
15
18 12
14 15
18 12
12 17 8 15
15
9
12 13
20 15
15 25
7 12
24 13
13 17 13
15
9
18 12
14 15
10
6 10
9
9 6 6
13 24 9 10 17
8 9 6 6
26 14 18
5 8 8 8
33
11
15 45 12 22
12 19 12 36
26 14
15 26
15 46
11 16 25
23
12 38
11 17
26 16
16 26
7 14
26 16
16
20 14
14 25
15 12
15 23
15 15
12 23
28 19
19 3640
40
20 14
14 25 18
18
20 14
14 26
22 11 11
16 24
22 16
11
11 24
36 22
22
22 20
20 22
16 1118 16
16 26 17 24
25
18 1126 19
19 27 20
16
18 16
16 26
11 17 25
24
18
26 19
19 27
11 20
28 19
19 36 40
40
21 12
12 14 23
23
28 18
18 18
50
50
36 22
22
22 20
20 22
28 20
20 29
28 20
20 29
26 18
18 18 29
29
12 12
12 21 21
21 22
34
13 20
13
20 16
18 27
11
17
2413
26
23 36 21
13 23 26
41 18
23
23
14
34 20
26 38
23 19
24 36 31
41
27 22
24 4245
19 47
36 23
30 40
34
13 13
20
20 18
16 27
24
11 17 13 26
41 23
18 23 14
23
36
13
23
21 26
34 26
20 38
23 24
19 3641
31
27 24
22 42
19
47
45
36 30
23 40
19 12
12 17 16
16
21 16
16
31 13
13 13
19 13
13 17 17 9
16
21 16
16
27 15
13 1318
19 13
13 17 16
17
9
21 16
16
27 13
15 13
18
28 15
15 18 16 9
16
9
21 16
16
27 16
16 2624
24
32 17
17 21 25
25 25
9
6
11
27
9 6 11 27
10 10
10
29 17
17 23
9 5
30 10
9 5 6
30 18 10 10
8 5
29
17 12
10 5 6
32 19
21 41 18 2610 12 23 11 34 11
9 30
5 10
8 29 17
5 12
9
30
18
5
6
10
10
10
32 21
19 41
5
618
10
12
23
11
2634
11
32 13
13 13 29
29
16 16
16
12 12
12 21 14
14 14
31
31
30 13
13 15 29
2930 14
14 15
27 15
13 13 29
27
16
12 12
14 20 16
14 14
32 29 19
32 18
30 13
14 15 27
2830 13
14 15
3127 13
15 13 27
2916
12 14
12 20 14
16 14
32
32
18
29
19
30 14
13 15 28
2730 14
13 15
31
27 15
15 17 29
29
17 17
17 23 16
16 1937
37
20 13
13 15
17 12
12 13
17 12
12 13
21 13
13 15
25 23
20 30
25 20
23 30
25 20
20 30
31 22
19 45
31 19
22 45
30 14
14 15 30
3030 14
14 15
31
31
20 13
13 15 28
2820 13
13 15
25 21
22 25
25 22
21 25
25 20
20 25
18
15
48 25
11
27
12
45
18
15 48 11 27
25 12 45
20 20
20
20 17
17 24 29
29
23 19
19 2937
37
20 20
20
20 17
17 25 29
29
20 17
17 25
19 1018 11
30 21
22 29 19 29
2910 8
25 1032
33
15 11
42 16 18
18 14
19
15 9
25
22
25 10
10 9 8 7
18 12 10 9 8
21 46 14
14 924 22 21
26
10
9
8
7
18
12
10
9
9 8
19 11
41 1519
14 8
15
14 921 19
19
20 1339 26
26
17 21
21
17 1413 13
13 29 1720 19
19 14
15
21
13 9
11 14
96 7 7
9 7 710
21 11 13
13 98 8
19
7 7
6
9 78
11 10
8
13
13 99 8
11 9
11
19 1227
16 19
7
8 7
11 10
9
13
21 23
13 9
11 9
14
9 7
9
6
7
20
10 1113
12 9
9
12 9
11 20 11 16
27 19
8
7 8 7
11 7 8 9
13
18 1016 11
42 15 26
27
19 11
29 20
22 30 1818
14 10
57 15
15 10
25
38 24
22 2814
13 107 6 9 7 8 6
1011 8
32 11
13 9
30 18 12
12 9
30
18 13
20 1332 18
18 46 20
13
14 11
10
25
8
12
13
8 12 11
10 25 14
12 10
38 12
13
7
7
6
8
106
8
19
18
30 22
21 29
11 19 29
2910 25
8 1033
32
15 42
11 16 18
18 14
10 19
14
24
22
26 9
10 9 18 12
8 7 10 9 8
21
19 41
11 15 14
19 8
9 15
15
25 22 21
25
10
9
18
12
8
7
10
9
10 8
46
9 14
14
21 19
19
20
39 26
26
17 21
21
17
13 13
13 29
14 17
13 20 19
19 14
9 15
21
13
11 14
99 9
7 6 710
20 11
9 13
2312
9 8
12
11
20
9 1127
16 19
7
8 11
7
7
8
9
9 13
2113
8 19
8
7 6 9
7 7 11
8 10 8
9 13
13
11 9
14
6
7 9
7
7
7
21
10 11
9 13
13
9
11 11 19
9 12 16
27 19
8
7 8 11
7 10 9
9 13
18
16 42
11 15 27
26
19
29 22
20 30
11 18
10 18
15
25 38 22
24 28
10 14
14 57
10 15
13
7
6
9
7
10
8
6
11 32
8 11
12 30 18
9 13
13
30
18
9 12
20
32 18
18 46
13 20
13
8
12
10
25
11 14
13
14 10 25
11 8 12
12 38
10 12
10 13 107 7 6 8 6 8
22 21
21
21 12
12 1229
29
18 18
18 26
29 20
20 28
29 20
20 28
22 20
20 22 22
22 36
2118 17
21 2630
23
19
18 18
17 18
21
18 21
17 26 23
30 19
18 17
18 18
20 12
12 15 23
2320 12
12 1525
25
23 13
13 15 23
2319 13
13 15
50
20 15
13 19
27 16
21 24
20 13
14 20
5027 21
16 2420 13
15 19
20 14
13 20
18 18
18 28 23
23
14 13
13 2140
4036 19
19 28
20 17
14 20
20 14
17 20
20 12
12 20
20 11
27 19
19 26 18
10
11
13
23
269
23
29
2511
16
15 13
17 11
26 16
16 18 16 8
15 8
2027 19
19 26
11 18 2610 11 13 23 9 23
25 15
11 16 13
29
1726 16
16 18
11 16 15
8 8
22 20
20 22 22
22 36
2411
11
16 22
24 16
11 11 22
26 14
14 20 18
1825 14
14 20
15
18
16
13
15 18 16 13
19 15
15 21
40 8
17 11 8
40
17
118 8
36 19
19 28
23 12
15 15
23 15
12 15
25 14
14 20 16
16 268
8 8
16
17 16
18 20
13
24
18
15
34
20 18
17 20 20
19
24 17
18 20
47 12
27 13 18 12
18 12 27 13 10
24 20
15 22
20 14
13 20
46 30
24 30 48
3436 26
20 2424
20 17
16 20 11
19
22 10
11 6
16 10 69
16
19 6
27 19
16 19 1015
17 8
22
8 19
24 10
9 17 10 7
20 43
22
23 11
18 10 6 8
20
30 21
17 22 13
21 13 238 6
1617 18
16 20 18
13 24 15 34
20 17
18 20 19
2024 18
17 20
24 15
20 22
20 13
14 20
47
27
13
18
12
18
12
27
13
12 10
46 24
30 3034
48
36 20
26 24
20 16
17 20
24 11
22
8
19 20
24
9
17
10
10 7
43
19 16
2211
16
10
6 6
10 9
19
27 16
19 19
6 10
17
15 8
22 20
23
18
10
6
11 8
30 17
21 2221
13 13
8
23 6
20 17
17 20 19
1924 17
17 20
26 11 16 10
17 10 29 12 11
26
11
17
10
16
10
29
12
11
20 15
15 15 28
2820 15
15 1530
30 3045
45
18 16
16 17 35
35
15 15
15
27 18
18 18
13 12
13 2913 13
12 29
15 15
15
28 16
16 16
39 25
27 49
39 27
25 49
23 15
15 15 27
2720 15
15 15
18 15
17 15
18 17
15 15
18 14
14 15
49
49
31 31
31
18 17
17
24 15
15 15
21 14
14 14
21 14
14 14
21 13
13 14
37
37
15 15
15
27 20
20 20
26 18
17 26
26 17
18 26
26 13
13 26
15 17 8
15
19
6
6
35 23
25
13 17
16
17 10
12 12
13 27 10
10
9
6
6
11
19
8
9
19 14
15 15
17
8
19
35 25
23
13 16
17
17
12 13
12 27
10 10
6 6 10
19
9 11 19
6 6 8 9 14
22 18
18
31 19
19 27 12
12 12
27
27
18 18
18
21 17
17
24 19
19 1918
18 24
9
17
9 17
18 13
13 19
11
12
15 16
12 18
9
16
14
17 14
12 19
11 12
15 12
16 18 14
9 16
17 12
14 19
15 12
12 18 14
14
17 14
14 28
35
34 21
10 126 7 7 7
9
19 11
11 138 7 6
7
26 8
49
11 7
12
36
12
22
12
19
15
45
12 11
38 17 25
23
15 11
46 1626 14
15 26
26 16
16 26
14
9 5
9
10
6
6
13
24 9
14
6
6
26
8
9
17 18
8
8
8
33
8 6
11
40
8
15
15
17 11
11 17
12 713
25 15
15 2012
8
15
15 9
17 12
12 18
15 12
14 18 13
15
16 12
12 18
15 12
14 18 10 13
15 9
17 13
13 24
10
35 49
34
10
6
7
12
7
7
19
9 11
21
11 8
137
6
26
7 8
11
12 36 12 19
12 22 15 45
26 15
14 26
12 38
11 17 23
25
15 46
11 16
26 16
16 26
7 1433
9
9 10 13 24
6 6 9 17
14 266 6 8 9
18
5 8 8 8
8
11 40 15
8 15
17 11
11 17
15 14
12 1815
13
16 12
12 18
12 13
25 15
15 20
7 12
15
9
8 15
17 12
12 18
15 14
12 18
10
15
9
13
17 13
13 24
6 10
14 13
13 21 23
23
18 18
18 2640
40
28 19
19 36
31 22
21 25
31 21
22 25
18 18
18 28 29
29
22 21
21
12 11
11 21
28 20
20 28
28 20
20 28
36 22
22
22 20
20 22
17 10
18 16
16 26 16 25
25
20 12
26 19
19 27 17
17
18 16
16 26
10 16 25
25
20
26 19
19 27
12 17
28 19
19 36
15 15
13 22
15 13
15 22
26 16
16
20 14
14 2540
40
36 22
22
22 20
20 22
23 16
17 23
23 17
16 23
20 14
14 25 19
19
20 14
14 26
38 26
19 34
27 23
24 40 42
31
23 19
24 3644
19 4740 31
23 37
41 18
2314
24
23 36 21
23 26
34
21 16
18 27 13
20 13
18
11
24
25
14
38 19
26 34
27 24
23 4031
42
23 24
19 36
19
47
44
40 23
31 37
23
36 23
21 26
41 23
18 14 24
34
21 18
16 2720
13 13
24
18 11 25 14
30 14
14 15 29
2930 13
13 15
31
31
16 16
16
21 12
12 1214
14 1429
2932 13
13 13
30 13
14 15 28
2730 13
14 15
31 32 18
29 19
16
20 14
12 1216
14 1428
2927 15
13 13
3230 14
13 15 27
2830 14
13 15
31
29
19
32
18
16
20 12
14 12 14
16 1429
2827 13
15 13
32
30 14
14 15 30
3030 14
14 15
31
31
20 13
13 15 28
2820 13
13 15
25 22
21 25
25 21
22 25
25 20
20 25
31 19
22 45
31 22
19 45
23 17
17 17 16
16 1929
2927 15
15 17
37
37
21 13
13 15
17 12
12 13
17 12
12 13
20 13
13 15
30 20
23 25
30 23
20 25
30 20
20 25
18
12
45
11
27
25
15
48
18
12 45 25
11 27 15 48
20 20
20
20 17
17 25 29
29
20 17
17 2537
37
20 20
20
23 19
19 29 29
29
20 17
17 24
10 5
30 9
12 5
29 17 8
10
10
5
6
30
18 9
18
5
6
32 21
20 41 10
10
11
12
23
26
11
34
10 30
5 9
10 10 30 18
5 6 9
12
29
17
5 8
18
32 20
21 41
5 6 10
2610 11 12 23 11 34
31 13
13 13 17
17 2116
16
17 12
12 19
27 15
13 13 16
16 2118 16
17 9
17 13
13 19
27 13
15 13 16
16 21
18
17
9
16
17 13
13 19
27 16
16 26 16
16 2124
24
32 17
17 21 25
25 25
16
9
16 9
18 15
15 28
9
11
27
6
9 11 27 6
10 10
10
29 18
18 23
∈ R7416×7416
Max Planck Institute Magdeburg Thomas Mach, Hierarchical Matrices 19/28
Page 44
Concept Application Eigenvalues
BEM for Electrostatic Computations26 14
14 20 19
1925 14
14 20
23 16
17 23
23 17
16 23
22 20
20 22 22
22 36
13
15
17
18
13 15 17 18
19 15
15 21
20 16
19 25 8
11 18 8
20 19
16 25
11
188 8
25 14
14 20 16
16 26
22 15
13 15
22 13
15 15
36 19
19 288
8 8
17 12
27 19
19 26 20 2513 23 10 11 23 9
25 14
16 10 13
3016 10
26 16
16 18 17 16
8 8
1727 19
19 26
12 20
13
23
10
11
2523
9
30
25
16
10
14 13
1626 16
16 18
10 17 8
16 8
22 20
20 22 22
22 36
28 20
20 28
28 20
20 28
21 11
11 12 20
20 22 29
2928 18
18 18
18
36 22
32
21
28 22
17 18
18 36 32
22 21
28 17
22 18
19 13
13 15 23
2323 13
13 1525
25
20 12
12 15 23
2320 12
12 15
50
19 12
15 20
24 20
17 25
20 14
13 20
50
24 17
20 2519 15
12 20
20 13
14 20
36 19
19 28 40
40
26 18
18 18 23
2321 13
13 14
20 14
17 20
20 17
14 20
20 12
12 20
22
22 11
6 18 10 8
20
13 22
13
17
18 12 21
17 2224
8 6
22
19 8
23 9
17 10 9 7
20 43
19
23 10
6 16 10
6 11 9
15
10
27 19
16 19
6 19 15
17 8
15
34 17
13
25
16
20 19
16 17
20 18
17 24 19
20
20 17
18 20
48 12
27 13 18 12
18 12 27 13 10
23 14
19 24
20 13
14 20
36 26
20 24 33
4644 29
23 3024
20 16
17 20 11
22 20
22
6
18
10
11 8
13 13
22
17 18
12
17
21 228
24 6
19 15
23
6 6
16
10 11
10 9
10 6
27 16
19 1919 17
15 8
2219
8 20
23
17
10
9
9 7
43
15 34 13 25
17 1620 16
19 17
20 17
18 24 20
1920 18
17 20
23 19
14 24
20 14
13 20
48
27
13
18
12
18
12
27
13
12 10
36 20
26 2446
33
44 23
29 30
20 17
16 20
24 11
22 18
18
19 13
13 18 9 17
9
17
31 19
19 27 12
12 12
27
27
18 18
18
24 18
18
21 17
17
24 19
19 19
19 12
13 17 14
9
16
11
12
18 13
16 15
19 13
12 17 9 16
14 11 12
18 16
13 15
28 14
14 17 14
1418 12
12 15
15 15
18
8
10 6 6
35 25
24
13 16
17
10 10
12 13
12 27 17 19 19
8 9 6 6
11 19 9 14
15 18 8
15
10
35 24
25
13 17
16
10
12 12
13 27
10 17
6
619
8
9
11
19
6
69
19 14
20 17
17 24 19
1920 17
17 20
29 12 17 10
16 10 26 11 11
29
12
16
10
17
10
26
11
11
20 15
15 15 27
2723 15
15 15
15 17
15 18
15 15
17 18
15 14
14 18
49
49
31 31
31
15 15
15
27 20
20 2037
37
18 17
17
15 15
15 24
14 14
14 21
14 14
14 21
14 13
13 21
39 27
25 49
39 25
27 49
20 15
15 15 28
2820 15
15 1530
30 3045
45
17 16
16 18 35
35
15 15
15
28 16
16 16
13 13
12 2913 12
13 29
15 15
15
27 18
18 18
26 17
18 26
26 18
17 26
26 13
13 26
35
12 138 6 7
7
27 8
10 117 6 7 7
9
19 11
22
12
11 10
116 7
8 9 6
12 12 10
32 13
49
8 6
17 11
11 17 15
8
15
11
40
18 12
12 17 15 9
8
15
12 713
20 15
15 25 12
18 12
12 16 15
1318 14
12 15
24 13
13 17 15 9
1318 14
12 15 10
10 339 5
9
13
24
6
6
9
10
8
926
6
614
17 18
8
8
8
11 7
15
45
12
19
12
22
12
36
15 11
46 16 23
25
12 11
38 1726 15
14 26
26 16
16 26
14
35 4912 8
13
6
727
7 8
22
10
7
6
11
7
7
19
9 11
12 12
11 11
10
6 8
7
9
6
12 32
10 13
8
17 11
11 17 8 15
15 11 40
18 12
12 16 13
15
18 12
14 15
18 12
12 17 8 15
15
9
12 13
20 15
15 25
7 12
24 13
13 17 13
15
9
18 12
14 15
10
6 10
9
9 6 6
13 24 9 10 17
8 9 6 6
26 14 18
5 8 8 8
33
11
15 45 12 22
12 19 12 36
26 14
15 26
15 46
11 16 25
23
12 38
11 17
26 16
16 26
7 14
26 16
16
20 14
14 25
15 12
15 23
15 15
12 23
28 19
19 3640
40
20 14
14 25 18
18
20 14
14 26
22 11 11
16 24
22 16
11
11 24
36 22
22
22 20
20 22
16 1118 16
16 26 17 24
25
18 1126 19
19 27 20
16
18 16
16 26
11 17 25
24
18
26 19
19 27
11 20
28 19
19 36 40
40
21 12
12 14 23
23
28 18
18 18
50
50
36 22
22
22 20
20 22
28 20
20 29
28 20
20 29
26 18
18 18 29
29
12 12
12 21 21
21 22
34
13 20
13
20 16
18 27
11
17
2413
26
23 36 21
13 23 26
41 18
23
23
14
34 20
26 38
23 19
24 36 31
41
27 22
24 4245
19 47
36 23
30 40
34
13 13
20
20 18
16 27
24
11 17 13 26
41 23
18 23 14
23
36
13
23
21 26
34 26
20 38
23 24
19 3641
31
27 24
22 42
19
47
45
36 30
23 40
19 12
12 17 16
16
21 16
16
31 13
13 13
19 13
13 17 17 9
16
21 16
16
27 15
13 1318
19 13
13 17 16
17
9
21 16
16
27 13
15 13
18
28 15
15 18 16 9
16
9
21 16
16
27 16
16 2624
24
32 17
17 21 25
25 25
9
6
11
27
9 6 11 27
10 10
10
29 17
17 23
9 5
30 10
9 5 6
30 18 10 10
8 5
29
17 12
10 5 6
32 19
21 41 18 2610 12 23 11 34 11
9 30
5 10
8 29 17
5 12
9
30
18
5
6
10
10
10
32 21
19 41
5
618
10
12
23
11
2634
11
32 13
13 13 29
29
16 16
16
12 12
12 21 14
14 14
31
31
30 13
13 15 29
2930 14
14 15
27 15
13 13 29
27
16
12 12
14 20 16
14 14
32 29 19
32 18
30 13
14 15 27
2830 13
14 15
3127 13
15 13 27
2916
12 14
12 20 14
16 14
32
32
18
29
19
30 14
13 15 28
2730 14
13 15
31
27 15
15 17 29
29
17 17
17 23 16
16 1937
37
20 13
13 15
17 12
12 13
17 12
12 13
21 13
13 15
25 23
20 30
25 20
23 30
25 20
20 30
31 22
19 45
31 19
22 45
30 14
14 15 30
3030 14
14 15
31
31
20 13
13 15 28
2820 13
13 15
25 21
22 25
25 22
21 25
25 20
20 25
18
15
48 25
11
27
12
45
18
15 48 11 27
25 12 45
20 20
20
20 17
17 24 29
29
23 19
19 2937
37
20 20
20
20 17
17 25 29
29
20 17
17 25
19 1018 11
30 21
22 29 19 29
2910 8
25 1032
33
15 11
42 16 18
18 14
19
15 9
25
22
25 10
10 9 8 7
18 12 10 9 8
21 46 14
14 924 22 21
26
10
9
8
7
18
12
10
9
9 8
19 11
41 1519
14 8
15
14 921 19
19
20 1339 26
26
17 21
21
17 1413 13
13 29 1720 19
19 14
15
21
13 9
11 14
96 7 7
9 7 710
21 11 13
13 98 8
19
7 7
6
9 78
11 10
8
13
13 99 8
11 9
11
19 1227
16 19
7
8 7
11 10
9
13
21 23
13 9
11 9
14
9 7
9
6
7
20
10 1113
12 9
9
12 9
11 20 11 16
27 19
8
7 8 7
11 7 8 9
13
18 1016 11
42 15 26
27
19 11
29 20
22 30 1818
14 10
57 15
15 10
25
38 24
22 2814
13 107 6 9 7 8 6
1011 8
32 11
13 9
30 18 12
12 9
30
18 13
20 1332 18
18 46 20
13
14 11
10
25
8
12
13
8 12 11
10 25 14
12 10
38 12
13
7
7
6
8
106
8
19
18
30 22
21 29
11 19 29
2910 25
8 1033
32
15 42
11 16 18
18 14
10 19
14
24
22
26 9
10 9 18 12
8 7 10 9 8
21
19 41
11 15 14
19 8
9 15
15
25 22 21
25
10
9
18
12
8
7
10
9
10 8
46
9 14
14
21 19
19
20
39 26
26
17 21
21
17
13 13
13 29
14 17
13 20 19
19 14
9 15
21
13
11 14
99 9
7 6 710
20 11
9 13
2312
9 8
12
11
20
9 1127
16 19
7
8 11
7
7
8
9
9 13
2113
8 19
8
7 6 9
7 7 11
8 10 8
9 13
13
11 9
14
6
7 9
7
7
7
21
10 11
9 13
13
9
11 11 19
9 12 16
27 19
8
7 8 11
7 10 9
9 13
18
16 42
11 15 27
26
19
29 22
20 30
11 18
10 18
15
25 38 22
24 28
10 14
14 57
10 15
13
7
6
9
7
10
8
6
11 32
8 11
12 30 18
9 13
13
30
18
9 12
20
32 18
18 46
13 20
13
8
12
10
25
11 14
13
14 10 25
11 8 12
12 38
10 12
10 13 107 7 6 8 6 8
22 21
21
21 12
12 1229
29
18 18
18 26
29 20
20 28
29 20
20 28
22 20
20 22 22
22 36
2118 17
21 2630
23
19
18 18
17 18
21
18 21
17 26 23
30 19
18 17
18 18
20 12
12 15 23
2320 12
12 1525
25
23 13
13 15 23
2319 13
13 15
50
20 15
13 19
27 16
21 24
20 13
14 20
5027 21
16 2420 13
15 19
20 14
13 20
18 18
18 28 23
23
14 13
13 2140
4036 19
19 28
20 17
14 20
20 14
17 20
20 12
12 20
20 11
27 19
19 26 18
10
11
13
23
269
23
29
2511
16
15 13
17 11
26 16
16 18 16 8
15 8
2027 19
19 26
11 18 2610 11 13 23 9 23
25 15
11 16 13
29
1726 16
16 18
11 16 15
8 8
22 20
20 22 22
22 36
2411
11
16 22
24 16
11 11 22
26 14
14 20 18
1825 14
14 20
15
18
16
13
15 18 16 13
19 15
15 21
40 8
17 11 8
40
17
118 8
36 19
19 28
23 12
15 15
23 15
12 15
25 14
14 20 16
16 268
8 8
16
17 16
18 20
13
24
18
15
34
20 18
17 20 20
19
24 17
18 20
47 12
27 13 18 12
18 12 27 13 10
24 20
15 22
20 14
13 20
46 30
24 30 48
3436 26
20 2424
20 17
16 20 11
19
22 10
11 6
16 10 69
16
19 6
27 19
16 19 1015
17 8
22
8 19
24 10
9 17 10 7
20 43
22
23 11
18 10 6 8
20
30 21
17 22 13
21 13 238 6
1617 18
16 20 18
13 24 15 34
20 17
18 20 19
2024 18
17 20
24 15
20 22
20 13
14 20
47
27
13
18
12
18
12
27
13
12 10
46 24
30 3034
48
36 20
26 24
20 16
17 20
24 11
22
8
19 20
24
9
17
10
10 7
43
19 16
2211
16
10
6 6
10 9
19
27 16
19 19
6 10
17
15 8
22 20
23
18
10
6
11 8
30 17
21 2221
13 13
8
23 6
20 17
17 20 19
1924 17
17 20
26 11 16 10
17 10 29 12 11
26
11
17
10
16
10
29
12
11
20 15
15 15 28
2820 15
15 1530
30 3045
45
18 16
16 17 35
35
15 15
15
27 18
18 18
13 12
13 2913 13
12 29
15 15
15
28 16
16 16
39 25
27 49
39 27
25 49
23 15
15 15 27
2720 15
15 15
18 15
17 15
18 17
15 15
18 14
14 15
49
49
31 31
31
18 17
17
24 15
15 15
21 14
14 14
21 14
14 14
21 13
13 14
37
37
15 15
15
27 20
20 20
26 18
17 26
26 17
18 26
26 13
13 26
15 17 8
15
19
6
6
35 23
25
13 17
16
17 10
12 12
13 27 10
10
9
6
6
11
19
8
9
19 14
15 15
17
8
19
35 25
23
13 16
17
17
12 13
12 27
10 10
6 6 10
19
9 11 19
6 6 8 9 14
22 18
18
31 19
19 27 12
12 12
27
27
18 18
18
21 17
17
24 19
19 1918
18 24
9
17
9 17
18 13
13 19
11
12
15 16
12 18
9
16
14
17 14
12 19
11 12
15 12
16 18 14
9 16
17 12
14 19
15 12
12 18 14
14
17 14
14 28
35
34 21
10 126 7 7 7
9
19 11
11 138 7 6
7
26 8
49
11 7
12
36
12
22
12
19
15
45
12 11
38 17 25
23
15 11
46 1626 14
15 26
26 16
16 26
14
9 5
9
10
6
6
13
24 9
14
6
6
26
8
9
17 18
8
8
8
33
8 6
11
40
8
15
15
17 11
11 17
12 713
25 15
15 2012
8
15
15 9
17 12
12 18
15 12
14 18 13
15
16 12
12 18
15 12
14 18 10 13
15 9
17 13
13 24
10
35 49
34
10
6
7
12
7
7
19
9 11
21
11 8
137
6
26
7 8
11
12 36 12 19
12 22 15 45
26 15
14 26
12 38
11 17 23
25
15 46
11 16
26 16
16 26
7 1433
9
9 10 13 24
6 6 9 17
14 266 6 8 9
18
5 8 8 8
8
11 40 15
8 15
17 11
11 17
15 14
12 1815
13
16 12
12 18
12 13
25 15
15 20
7 12
15
9
8 15
17 12
12 18
15 14
12 18
10
15
9
13
17 13
13 24
6 10
14 13
13 21 23
23
18 18
18 2640
40
28 19
19 36
31 22
21 25
31 21
22 25
18 18
18 28 29
29
22 21
21
12 11
11 21
28 20
20 28
28 20
20 28
36 22
22
22 20
20 22
17 10
18 16
16 26 16 25
25
20 12
26 19
19 27 17
17
18 16
16 26
10 16 25
25
20
26 19
19 27
12 17
28 19
19 36
15 15
13 22
15 13
15 22
26 16
16
20 14
14 2540
40
36 22
22
22 20
20 22
23 16
17 23
23 17
16 23
20 14
14 25 19
19
20 14
14 26
38 26
19 34
27 23
24 40 42
31
23 19
24 3644
19 4740 31
23 37
41 18
2314
24
23 36 21
23 26
34
21 16
18 27 13
20 13
18
11
24
25
14
38 19
26 34
27 24
23 4031
42
23 24
19 36
19
47
44
40 23
31 37
23
36 23
21 26
41 23
18 14 24
34
21 18
16 2720
13 13
24
18 11 25 14
30 14
14 15 29
2930 13
13 15
31
31
16 16
16
21 12
12 1214
14 1429
2932 13
13 13
30 13
14 15 28
2730 13
14 15
31 32 18
29 19
16
20 14
12 1216
14 1428
2927 15
13 13
3230 14
13 15 27
2830 14
13 15
31
29
19
32
18
16
20 12
14 12 14
16 1429
2827 13
15 13
32
30 14
14 15 30
3030 14
14 15
31
31
20 13
13 15 28
2820 13
13 15
25 22
21 25
25 21
22 25
25 20
20 25
31 19
22 45
31 22
19 45
23 17
17 17 16
16 1929
2927 15
15 17
37
37
21 13
13 15
17 12
12 13
17 12
12 13
20 13
13 15
30 20
23 25
30 23
20 25
30 20
20 25
18
12
45
11
27
25
15
48
18
12 45 25
11 27 15 48
20 20
20
20 17
17 25 29
29
20 17
17 2537
37
20 20
20
23 19
19 29 29
29
20 17
17 24
10 5
30 9
12 5
29 17 8
10
10
5
6
30
18 9
18
5
6
32 21
20 41 10
10
11
12
23
26
11
34
10 30
5 9
10 10 30 18
5 6 9
12
29
17
5 8
18
32 20
21 41
5 6 10
2610 11 12 23 11 34
31 13
13 13 17
17 2116
16
17 12
12 19
27 15
13 13 16
16 2118 16
17 9
17 13
13 19
27 13
15 13 16
16 21
18
17
9
16
17 13
13 19
27 16
16 26 16
16 2124
24
32 17
17 21 25
25 25
16
9
16 9
18 15
15 28
9
11
27
6
9 11 27 6
10 10
10
29 18
18 23
==
fuM
Max Planck Institute Magdeburg Thomas Mach, Hierarchical Matrices 19/28
Page 45
Concept Application Eigenvalues
BEM for Electrostatic Computations
−4 −20
24
·103
−20
2·103
−4
−2
0
2
4
·103
−5V
+5V+2V
−1
−0.5
0
0.5
1
Max Planck Institute Magdeburg Thomas Mach, Hierarchical Matrices 20/28
Page 46
Concept Application Eigenvalues
Eigenvalues
Max Planck Institute Magdeburg Thomas Mach, Hierarchical Matrices 21/28
Page 47
Concept Application Eigenvalues
Definition
Mx = λx
M ∈ Rn×n
eigenvectorx ∈ Cn, x 6= 0
eigenvalueλ ∈ C
Max Planck Institute Magdeburg Thomas Mach, Hierarchical Matrices 22/28
Page 48
Concept Application Eigenvalues
Definition
Mx = λx
M ∈ Rn×n
eigenvectorx ∈ Cn, x 6= 0
eigenvalueλ ∈ C
Max Planck Institute Magdeburg Thomas Mach, Hierarchical Matrices 22/28
Page 49
Concept Application Eigenvalues
Definition
Mx = λx
M ∈ Rn×n
eigenvectorx ∈ Cn, x 6= 0
eigenvalueλ ∈ C
Max Planck Institute Magdeburg Thomas Mach, Hierarchical Matrices 22/28
Page 50
Concept Application Eigenvalues
Definition
Mx = λx
M ∈ Rn×n
eigenvectorx ∈ Cn, x 6= 0
eigenvalueλ ∈ C
Max Planck Institute Magdeburg Thomas Mach, Hierarchical Matrices 22/28
Page 51
Concept Application Eigenvalues
Definition
Mx = λx
M = MT ∈ Rn×n
eigenvectorx ∈ Rn, x 6= 0
eigenvalueλ ∈ R
Max Planck Institute Magdeburg Thomas Mach, Hierarchical Matrices 22/28
Page 52
Concept Application Eigenvalues
Application
Eigenvalue problems occur in many different applications:
vibration analysis,
molecular and quantum dynamics(e.g. Schrodinger or Kohn-Sham equations),
finite state of a Markov chain,
numerical mathematics (e.g. MOR, eM , convergence theory),
...
Max Planck Institute Magdeburg Thomas Mach, Hierarchical Matrices 23/28
Page 53
Concept Application Eigenvalues
Definition
Mx = λx
M = MT ∈ H(T , k)
eigenvectorx ∈ Rn, x 6= 0
eigenvalueλ ∈ R
Max Planck Institute Magdeburg Thomas Mach, Hierarchical Matrices 24/28
Page 54
Concept Application Eigenvalues
Preconditioned Inverse Iteration[Knyazev, Neymeyr, et al.]
Definition
The function
µ(x) = µ(x ,M) =xTMx
xT x
is called the Rayleigh quotient.
Minimize the Rayleigh quotient by a gradient method:
xi+1 := xi − α∇µ(xi ), ∇µ(x) =2
xT x(Mx − xµ(x)),
+ preconditioning ⇒ update equation:
xi+1 := xi − T−1 (Mxi − xiµ(xi )).
Residual r(x) = Mx − xµ(x).
Max Planck Institute Magdeburg Thomas Mach, Hierarchical Matrices 25/28
Page 55
Concept Application Eigenvalues
Preconditioned Inverse Iteration[Knyazev, Neymeyr, et al.]
Definition
The function
µ(x) = µ(x ,M) =xTMx
xT x
is called the Rayleigh quotient.
Minimize the Rayleigh quotient by a gradient method:
xi+1 := xi − α∇µ(xi ), ∇µ(x) =2
xT x(Mx − xµ(x)),
+ preconditioning ⇒ update equation:
xi+1 := xi − T−1 (Mxi − xiµ(xi )).
Residual r(x) = Mx − xµ(x).
Max Planck Institute Magdeburg Thomas Mach, Hierarchical Matrices 25/28
Page 56
Concept Application Eigenvalues
Preconditioned Inverse Iteration[Knyazev, Neymeyr, et al.]
Definition
The function
µ(x) = µ(x ,M) =xTMx
xT x
is called the Rayleigh quotient.
Minimize the Rayleigh quotient by a gradient method:
xi+1 := xi − α∇µ(xi ), ∇µ(x) =2
xT x(Mx − xµ(x)),
+ preconditioning ⇒ update equation:
xi+1 := xi − T−1 (Mxi − xiµ(xi )).
Residual r(x) = Mx − xµ(x).
Max Planck Institute Magdeburg Thomas Mach, Hierarchical Matrices 25/28
Page 57
Concept Application Eigenvalues
Preconditioned Inverse Iteration[Knyazev, Neymeyr, et al.]
Definition
The function
µ(x) = µ(x ,M) =xTMx
xT x
is called the Rayleigh quotient.
Minimize the Rayleigh quotient by a gradient method:
xi+1 := xi − α∇µ(xi ), ∇µ(x) =2
xT x(Mx − xµ(x)),
+ preconditioning ⇒ update equation:
xi+1 := xi − T−1 (Mxi − xiµ(xi )).
Residual r(x) = Mx − xµ(x).
Max Planck Institute Magdeburg Thomas Mach, Hierarchical Matrices 25/28
Page 58
Concept Application Eigenvalues
Preconditioned Inverse Iteration[Knyazev, Neymeyr ’09]
xi+1 := xi − T−1 (Mxi − xiµ(xi ))
If
M symmetric positive definite and
T−1 approximates the inverse of M, so that∥∥I − T−1M∥∥M≤ c < 1,
then Preconditioned INVerse ITeration (PINVIT) converges.
Max Planck Institute Magdeburg Thomas Mach, Hierarchical Matrices 26/28
Page 59
Concept Application Eigenvalues
Preconditioned Inverse Iteration[Knyazev, Neymeyr ’09]
xi+1 := xi − T−1 (Mxi − xiµ(xi ))
If
M symmetric positive definite and
T−1 approximates the inverse of M, so that∥∥I − T−1M∥∥M≤ c < 1,
then Preconditioned INVerse ITeration (PINVIT) converges.
Max Planck Institute Magdeburg Thomas Mach, Hierarchical Matrices 26/28
Page 60
Concept Application Eigenvalues
BEM, smallest eigenvalues
dense matrix, but approximable by an H-matrix
d = 4, c = 0.2, εeig = 10−4
258 1026 4098 16386 6553810−3
10−2
10−1
100
101
102
103
104
CP
Uti
me
ins
PINVIT
O(ni (log2 ni )CH)H-Cholesky factor.
O(ni (log2 ni )2 CH)
Max Planck Institute Magdeburg Thomas Mach, Hierarchical Matrices 27/28
Page 61
Concept Application Eigenvalues
BEM, smallest eigenvalues
dense matrix, but approximable by an H-matrix
ARPACK / eigs
d = 4, c = 0.2, εeig = 10−4
258 1026 4098 16386 6553810−3
10−2
10−1
100
101
102
103
104
CP
Uti
me
ins
PINVIT
O(ni (log2 ni )CH)H-Cholesky factor.
O(ni (log2 ni )2 CH)
Max Planck Institute Magdeburg Thomas Mach, Hierarchical Matrices 27/28
Page 62
Concept Application Eigenvalues
Concluding Remarks
H-arithmetic is cheap and applicable too many problems.
Computing eigenvalues of H-matrices in linear-polylogarithmiccomplexity is possible.
See poster: “Slicing the Spectrum of Symmetric H`-Matrices”.
Thank you for your attention.
Max Planck Institute Magdeburg Thomas Mach, Hierarchical Matrices 28/28
Page 63
Concept Application Eigenvalues
Concluding Remarks
H-arithmetic is cheap and applicable too many problems.
Computing eigenvalues of H-matrices in linear-polylogarithmiccomplexity is possible.
See poster: “Slicing the Spectrum of Symmetric H`-Matrices”.
Thank you for your attention.
Max Planck Institute Magdeburg Thomas Mach, Hierarchical Matrices 28/28
Page 64
Concept Application Eigenvalues
Concluding Remarks
H-arithmetic is cheap and applicable too many problems.
Computing eigenvalues of H-matrices in linear-polylogarithmiccomplexity is possible.
See poster: “Slicing the Spectrum of Symmetric H`-Matrices”.
Thank you for your attention.
Max Planck Institute Magdeburg Thomas Mach, Hierarchical Matrices 28/28
Page 65
Concept Application Eigenvalues
Concluding Remarks
H-arithmetic is cheap and applicable too many problems.
Computing eigenvalues of H-matrices in linear-polylogarithmiccomplexity is possible.
See poster: “Slicing the Spectrum of Symmetric H`-Matrices”.
Thank you for your attention.
Max Planck Institute Magdeburg Thomas Mach, Hierarchical Matrices 28/28
Page 66
Appendix
BEM, smallest eigenvalues
dense matrix, but approximable by an H-matrix
d = 4, c = 0.2, εeig = 10−4
258 1026 4098 16386 6553810−3
10−2
10−1
100
101
102
103
104
CP
Uti
me
ins
PINVIT
O(ni (log2 ni )CH)H-Cholesky factor.
O(ni (log2 ni )2 CH)
Max Planck Institute Magdeburg Thomas Mach, Hierarchical Matrices 28/28
Page 67
Appendix
BEM, smallest eigenvalues
dense matrix, but approximable by an H-matrix
eig ∈ O(n3)
d = 4, c = 0.2, εeig = 10−4
258 1026 4098 16386 6553810−3
10−2
10−1
100
101
102
103
104
CP
Uti
me
ins
PINVIT
O(ni (log2 ni )CH)H-Cholesky factor.
O(ni (log2 ni )2 CH)
Max Planck Institute Magdeburg Thomas Mach, Hierarchical Matrices 28/28