fourier analysis of iterative methods for the helmholtz...
TRANSCRIPT
Fourier Analysis of Iterative Methodsfor the Helmholtz problem
Hanzhi Diao
MSc Thesis Defencesupervisor: Prof. dr. ir. C. Vuik
December 20, 2012
Hanzhi Diao (TU Delft) Helmholtz problem December 20, 2012 1 / 32
Outline
1 Problem Formulation
2 Iterative MethodsIterative SolversPreconditioning TechniquesMultilevel Krylov Multigrid Method
3 Fourier AnalysisTheoryAnalysis of the PreconditioningMultigrid AnalysisMultigrid Convergence
4 Numerical Experiments
5 Future Work
Hanzhi Diao (TU Delft) Helmholtz problem December 20, 2012 2 / 32
Problem Formulation
Helmholtz Problem
Helmholtz equation
−∆u(x)− k2 u(x) = f(x) in Ω ∈ R3
Boundary condition
Dirichlet / Neumann / Sommerfeld
Discretization
finite difference method / finite element method
Linear system • sparse• symmetric but non-Hermitian
Hanzhi Diao (TU Delft) Helmholtz problem December 20, 2012 3 / 32
Problem Formulation
Thesis Work
Objective
spectral properties =⇒ convergence behaviour
Task1 Preconditioning techniques
– shifted Laplacian preconditioner M– deflation operator P and Q
2 Iterative solver– multigrid method for M−1
– Krylov subspace method for Ax = b / AM−1x = b / AM−1Qx = b
3 Fourier analysis– spectrum distribution– convergence factor
4 Numerical solution
Hanzhi Diao (TU Delft) Helmholtz problem December 20, 2012 4 / 32
Problem Formulation
Thesis Work
Objective
spectral properties =⇒ convergence behaviour
Task1 Preconditioning techniques
– shifted Laplacian preconditioner M– deflation operator P and Q
2 Iterative solver– multigrid method for M−1
– Krylov subspace method for Ax = b / AM−1x = b / AM−1Qx = b
3 Fourier analysis– spectrum distribution– convergence factor
4 Numerical solution
Hanzhi Diao (TU Delft) Helmholtz problem December 20, 2012 4 / 32
Problem Formulation
Model Problem
1D dimensionless Helmholtz problem with homogeneous Dirichletboundary condition
−∆u(x)− k2 u(x) = f(x) for x ∈ (0, 1),
u(0) = u(1) = 0.
The resulting linear system
Ax = b where A =1
h2
2 −1−1 2 −1
. . . . . . . . .−1 2 −1−1 2
− k2IWave resolution
gw · h =2π
k
Hanzhi Diao (TU Delft) Helmholtz problem December 20, 2012 5 / 32
Problem Formulation
Model Problem
Eigenvalue
λl =4
h2sin2(lπh/2)− k2 for l = 1, 2, · · · , n
Difficulty in solving Helmholtz problem
0 100 200 300 400 500
102
104
106
108
wavenumber k
extremitiesofλ
λmax
|λmin|
indefiniteness
0 100 200 300 400 500
102
104
106
108
wavenumber k
conditionnumberκ
condition number
Hanzhi Diao (TU Delft) Helmholtz problem December 20, 2012 6 / 32
Iterative Methods Iterative Solvers
Multigrid Method
The solver for inverting the shifted Laplacian preconditioner M• The coarsening strategy is done by doubling the mesh size,
i.e. Ωh → Ω2h.• The smoother is ω-Jacobi iteration operator.
– ω is chosen as the optimal one ωopt.• The intergrid transfer
– restriction by full weighting operator– prolongation by linear interpolation operator
The failure of MG in solving Ax = b
• The coarse grid cannot cope with high wavenumber problem.• The ω-Jacobi iteration does not converge.
Hanzhi Diao (TU Delft) Helmholtz problem December 20, 2012 7 / 32
Iterative Methods Iterative Solvers
Multigrid Method
The solver for inverting the shifted Laplacian preconditioner M• The coarsening strategy is done by doubling the mesh size,
i.e. Ωh → Ω2h.• The smoother is ω-Jacobi iteration operator.
– ω is chosen as the optimal one ωopt.• The intergrid transfer
– restriction by full weighting operator– prolongation by linear interpolation operator
The failure of MG in solving Ax = b
• The coarse grid cannot cope with high wavenumber problem.• The ω-Jacobi iteration does not converge.
Hanzhi Diao (TU Delft) Helmholtz problem December 20, 2012 7 / 32
Iterative Methods Iterative Solvers
Krylov Subspace Methods
The solver for solving the linear system Ax = b
• GMRES, used in the thesis work• CG• BiCGStab• GCR,IDR(s),. . .
Hanzhi Diao (TU Delft) Helmholtz problem December 20, 2012 8 / 32
Iterative Methods Iterative Solvers
Approximated Inversion
Given the iteration operator G, there is the approximated inversion
A−1 = (I −G)A−1.
For the stationary iteration, there is
A−1m = (I −Gm)A−1.
For the multigrid iteration, there is
A−1MG = (I − Tm1 )A−1.
Hanzhi Diao (TU Delft) Helmholtz problem December 20, 2012 9 / 32
Iterative Methods Preconditioning Techniques
Shifted Laplacian Preconditioner
M := −∆h − (β1 + ιβ2︸ ︷︷ ︸shift
)k2I
• preconditioned system
A := AM−1 = M−1A and σ(AM−1) = σ(M−1A)
• preservation of symmetry
(AM−1)T = AM−1
• circular spectrum distribution
(λr −1
2)2 + (λi −
β1 − 1
2β2)2 =
β22 + (1− β1)2(2β2)2
Hanzhi Diao (TU Delft) Helmholtz problem December 20, 2012 10 / 32
Iterative Methods Preconditioning Techniques
Shifted Laplacian Preconditioner
M := −∆h − (β1 + ιβ2︸ ︷︷ ︸shift
)k2I
• preconditioned system
A := AM−1 = M−1A and σ(AM−1) = σ(M−1A)
• preservation of symmetry
(AM−1)T = AM−1
• circular spectrum distribution
(λr −1
2)2 + (λi −
β1 − 1
2β2)2 =
β22 + (1− β1)2(2β2)2
Hanzhi Diao (TU Delft) Helmholtz problem December 20, 2012 10 / 32
Iterative Methods Preconditioning Techniques
Shifted Laplacian Preconditioner
β1 = 1 =⇒ the most compact distribution
−1 0 1−1
0
1
2
ℜ(λ)
ℑ(λ
)
shift = 0 − 0.5ι
−1 0 1−1
0
1
2
ℜ(λ)
ℑ(λ
)
shift = 0 − 1ι
−1 0 1−1
0
1
2
ℜ(λ)ℑ(λ
)
shift = 1 − 0.5ι
−1 0 1−1
0
1
2
ℜ(λ)
ℑ(λ
)
shift = 1 − 1ι
The spectrum distributions of the preconditioned matrix AM−1 with respect toseveral typical shifts when k = 100
Hanzhi Diao (TU Delft) Helmholtz problem December 20, 2012 11 / 32
Iterative Methods Preconditioning Techniques
Deflation Operator
For an invertible A, take any n× r full rank matrices Y and Zleft P := I − A ZE−1Y T + λd ZE
−1Y T ,
right Q := I − ZE−1Y T A+ λd ZE−1Y T ,
where E = Y T AZ.
0 0.5 1
−0.5
0
0.5
ℜ(λ)
ℑ(λ
)
r = n/20
0 0.5 1
−0.5
0
0.5
ℜ(λ)
ℑ(λ
)
r = n/10
0 0.5 1
−0.5
0
0.5
ℜ(λ)
ℑ(λ
)
r = n/5
0 0.5 1
−0.5
0
0.5
ℜ(λ)
ℑ(λ
)
r = n/2
The spectrum distributions of the deflated matrix AQ towards λd = 0.2where k = 100, shift = 1− ι1,AZ = Z Λr and Y T A = Λr Y
T
Hanzhi Diao (TU Delft) Helmholtz problem December 20, 2012 12 / 32
Iterative Methods Preconditioning Techniques
Deflation Operatorσ(A) = λ1, · · · , λn with |λ1| 6 · · · 6 |λn|
• Projector in case of λd = 0
PD · PD = PD and QD ·QD = QD.
• Preservation of symmetry in case of λd = 0,• Spectrum distribution
σ(PA) = σ(AQ) = λd, · · · , λd, µr+1, · · · , µn.
• Condition number
κ(PA) =|µn|
min|λd|, |µr+1|in case of λd 6= 0,
Hanzhi Diao (TU Delft) Helmholtz problem December 20, 2012 13 / 32
Iterative Methods Preconditioning Techniques
Deflation OperatorInaccuracy in E−1
Assume AZ = Z Λr and Y T A = Λr YT , then E = Y T AZ = Λr.
E−1 = diag(1− ε1λ1
, · · · , 1− εrλr
)
Hanzhi Diao (TU Delft) Helmholtz problem December 20, 2012 14 / 32
Iterative Methods Preconditioning Techniques
Deflation OperatorInaccuracy in E−1
Assume AZ = Z Λr and Y T A = Λr YT , then E = Y T AZ = Λr.
E−1 = diag(1− ε1λ1
, · · · , 1− εrλr
)
⇓σ(PA) = (1− ε1)λd + λ1ε1, · · · , (1− εr)λd + λrεr, λr+1, · · · , λn
Hanzhi Diao (TU Delft) Helmholtz problem December 20, 2012 14 / 32
Iterative Methods Preconditioning Techniques
Deflation OperatorInaccuracy in E−1
Assume AZ = Z Λr and Y T A = Λr YT , then E = Y T AZ = Λr.
E−1 = diag(1− ε1λ1
, · · · , 1− εrλr
)
⇓σ(PA) = (1− ε1)λd + λ1ε1, · · · , (1− εr)λd + λrεr, λr+1, · · · , λn
⇓λd = 0 ⇒ σ(PA) = λ1ε1︸︷︷︸
6=0
, · · · , λrεr︸︷︷︸6=0
, λr+1, · · · , λn,
λd = λn ⇒ σ(PA) ≈ λn, · · · , λn, λr+1, · · · , λn.
Hanzhi Diao (TU Delft) Helmholtz problem December 20, 2012 14 / 32
Iterative Methods Multilevel Krylov Multigrid Method
Multilevel Krylov Multigrid Method
A recursive Krylov solution of E−1
1 Use the approximation
M−1 ≈ Z(Y TMZ)−1Y T .
2 Take the replacement
E :=Y T AZ = Y TAM−1Z ≈ Y TAZ︸ ︷︷ ︸A(2)
(Y TMZ︸ ︷︷ ︸M(2)
)−1 Y TZ︸ ︷︷ ︸B(2)
E−1 ≈(A(2)M
−1(2)B(2)
)−13 Solve A−1(2) in the same way as A−1
Hanzhi Diao (TU Delft) Helmholtz problem December 20, 2012 15 / 32
Iterative Methods Multilevel Krylov Multigrid Method
Multilevel Krylov Multigrid Method
1
2
3
4
5
k-th iteration (k + 1)-th iteration multigrid method• multilevel Krylov method exact inversion of M(m)
exact inversion of A(m)
The illustration of multilevel Krylov multigrid method in a five-level grid
Hanzhi Diao (TU Delft) Helmholtz problem December 20, 2012 16 / 32
Fourier Analysis Theory
Principles of Fourier Analysis
Find out a subspace E = spanφ1, · · · , φm such that
KE ⊂ E =⇒ K Φ = Φ K.
For any v = Φc ∈ E, there is
K v = K Φc = ΦK c where K amplifies c.
Assume E is the union of several disjoint subspaces. Then, there is adiagonal block matrix
K :∧= [K l] with l as the block index.
Hanzhi Diao (TU Delft) Helmholtz problem December 20, 2012 17 / 32
Fourier Analysis Theory
Principles of Fourier Analysis
Find out a subspace E = spanφ1, · · · , φm such that
KE ⊂ E =⇒ K Φ = Φ K.
For any v = Φc ∈ E, there is
K v = K Φc = ΦK c where K amplifies c.
Assume E is the union of several disjoint subspaces. Then, there is adiagonal block matrix
K :∧= [K l] with l as the block index.
Hanzhi Diao (TU Delft) Helmholtz problem December 20, 2012 17 / 32
Fourier Analysis Theory
Fourier Analysis for Multigrid Analysis
In a two-level grid, the invariance subspace is given by
Elh := spanφlh, φn−lh in Ωh =⇒ El2h := spanφl2h in Ω2h.
A1,M1, S :Elh → Elh
A2,M2 :El2h → El2h
R21 :Elh → El2h
P 12 :El2h → Elh
=⇒ T 21 : Elh → Elh
In a multilevel grid, there is
Tmk = Sν2k
(I − P kk+1 (I − Tmk+1)M
−1k+1 R
k+1k Mk
)Sν1k with Tmm = 0,
Hanzhi Diao (TU Delft) Helmholtz problem December 20, 2012 18 / 32
Fourier Analysis Theory
Fourier Analysis for Multigrid Analysis
In a two-level grid, the invariance subspace is given by
Elh := spanφlh, φn−lh in Ωh =⇒ El2h := spanφl2h in Ω2h.
A1,M1, S :Elh → Elh
A2,M2 :El2h → El2h
R21 :Elh → El2h
P 12 :El2h → Elh
=⇒ T 21 : Elh → Elh
In a multilevel grid, there is
Tmk = Sν2k
(I − P kk+1 (I − Tmk+1)M
−1k+1 R
k+1k Mk
)Sν1k with Tmm = 0,
Hanzhi Diao (TU Delft) Helmholtz problem December 20, 2012 18 / 32
Fourier Analysis Theory
Application to Preconditioning
Shifted Laplacian preconditioner
˜A = AM−1
Deflation operator
Q = I − P 12
˜E2R
12(λnI − ˜
A) with ˜E2 = R2
1˜AP 1
2
Advantage of Fourier analysis• computational time• memory requirement• accuracy
Hanzhi Diao (TU Delft) Helmholtz problem December 20, 2012 19 / 32
Fourier Analysis Theory
Application to Preconditioning
Shifted Laplacian preconditioner
˜A = AM−1
Deflation operator
Q = I − P 12
˜E2R
12(λnI − ˜
A) with ˜E2 = R2
1˜AP 1
2
Advantage of Fourier analysis• computational time• memory requirement• accuracy
Hanzhi Diao (TU Delft) Helmholtz problem December 20, 2012 19 / 32
Fourier Analysis Analysis of the Preconditioning
Basic Preconditioning Effect
• The spectrum of AM−1 is restricted to a circular distribution.
−0.5 0 0.5 1 1.5−1
−0.5
0
0.5
1
ℜ(λ)
ℑ(λ
)
k =10
−0.5 0 0.5 1 1.5−1
−0.5
0
0.5
1
ℜ(λ)ℑ(λ
)
k =50
−0.5 0 0.5 1 1.5−1
−0.5
0
0.5
1
ℜ(λ)
ℑ(λ
)
k =100
−0.5 0 0.5 1 1.5−1
−0.5
0
0.5
1
ℜ(λ)
ℑ(λ
)
k =500
• The spectrum of AM−1Q is clustered around (1, 0).
0.96 1 1.04
−0.04
0
0.04
ℜ(λ)
ℑ(λ
)
k =10
0.96 1 1.04
−0.04
0
0.04
ℜ(λ)
ℑ(λ
)
k =50
0.96 1 1.04
−0.04
0
0.04
ℜ(λ)
ℑ(λ
)
k =100
0.96 1 1.04
−0.04
0
0.04
ℜ(λ)
ℑ(λ
)
k =500
Hanzhi Diao (TU Delft) Helmholtz problem December 20, 2012 20 / 32
Fourier Analysis Analysis of the Preconditioning
Choice of Shift β1 + ιβ2
1 β2 determines the length of arc on which the eigenvalues ofAM−1 are located.
−0.5 0 0.5 1 1.5−1
−0.5
0
0.5
1
ℜ(λ)
ℑ(λ
)
β =1-0.2i
−0.5 0 0.5 1 1.5−1
−0.5
0
0.5
1
ℜ(λ)
ℑ(λ
)
β =1-1i
−0.5 0 0.5 1 1.5−1
−0.5
0
0.5
1
ℜ(λ)
ℑ(λ
)
β =1-5i
−0.5 0 0.5 1 1.5−1
−0.5
0
0.5
1
ℜ(λ)
ℑ(λ
)
β =1-10i
2 β2 has the indirect influence on the tightness of spectrumdistribution of AM−1Q.
0.8 0.9 1 1.1
−0.2
−0.1
0
ℜ(λ)
ℑ(λ
)
β =1-0.2i
0.8 0.9 1 1.1
−0.2
−0.1
0
ℜ(λ)
ℑ(λ
)
β =1-1i
0.8 0.9 1 1.1
−0.2
−0.1
0
ℜ(λ)
ℑ(λ
)β =1-5i
0.8 0.9 1 1.1
−0.2
−0.1
0
ℜ(λ)
ℑ(λ
)
β =1-10i
Hanzhi Diao (TU Delft) Helmholtz problem December 20, 2012 21 / 32
Fourier Analysis Analysis of the Preconditioning
Influence of Wave resolution gw
1 High resolution exerts little negative influence on the spectrumdistribution of AM−1.
−0.5 0 0.5 1 1.5−1
−0.5
0
0.5
1
ℜ(λ)
ℑ(λ
)
gw =10
−0.5 0 0.5 1 1.5−1
−0.5
0
0.5
1
ℜ(λ)
ℑ(λ
)
gw =30
−0.5 0 0.5 1 1.5−1
−0.5
0
0.5
1
ℜ(λ)
ℑ(λ
)
gw =60
−0.5 0 0.5 1 1.5−1
−0.5
0
0.5
1
ℜ(λ)
ℑ(λ
)
gw =120
2 High resolution results in a more favourable spectrum distributionof AM−1Q.
0.8 1 1.2
−0.2
0
0.2
ℜ(λ)
ℑ(λ
)
gw = 10
0.96 1 1.04
−0.04
0
0.04
ℜ(λ)
ℑ(λ
)
gw =30
0.96 1 1.04
−0.04
0
0.04
ℜ(λ)
ℑ(λ
)gw =60
0.96 1 1.04
−0.04
0
0.04
ℜ(λ)
ℑ(λ
)
gw =120
Hanzhi Diao (TU Delft) Helmholtz problem December 20, 2012 22 / 32
Fourier Analysis Multigrid Analysis
Approximated Shifted Laplacian PreconditioningAM−1 = A(I − Tm
1 )M−1
• The multigrid introduces disturbance to the preconditioning effect.
−0.5 0 0.5 1 1.5−1
−0.5
0
0.5
1
ℜ(λ)
ℑ(λ
)
1-grid analysis
−0.5 0 0.5 1 1.5−1
−0.5
0
0.5
1
ℜ(λ)
ℑ(λ
)
2-grid analysis
−0.5 0 0.5 1 1.5−1
−0.5
0
0.5
1
ℜ(λ)
ℑ(λ
)
3-grid analysis
−0.5 0 0.5 1 1.5−1
−0.5
0
0.5
1
ℜ(λ)
ℑ(λ
)
4-grid analysis
• The disturbance can be easily corrected by several iterations at acheap cost.
−0.5 0 0.5 1 1.5−1
−0.5
0
0.5
1
ℜ(λ)
ℑ(λ
)
1 iteration
−0.5 0 0.5 1 1.5−1
−0.5
0
0.5
1
ℜ(λ)
ℑ(λ
)
2 iterations
−0.5 0 0.5 1 1.5−1
−0.5
0
0.5
1
ℜ(λ)
ℑ(λ
)3 iterations
−0.5 0 0.5 1 1.5−1
−0.5
0
0.5
1
ℜ(λ)
ℑ(λ
)
4 iterations
Hanzhi Diao (TU Delft) Helmholtz problem December 20, 2012 23 / 32
Fourier Analysis Multigrid Analysis
Approximated Deflation PreconditioningAM−1Q where the construction of Q is based on the M−1
• The preconditioning AM−1Q is much more sensitive to theaccuracy in the approximation of M−1.
0.8 0.9 1 1.1−0.2
−0.1
0
0.1
ℜ(λ)
ℑ(λ
)
1 iteration
0.8 0.9 1 1.1−0.2
−0.1
0
0.1
ℜ(λ)
ℑ(λ
)
2 iterations
0.8 0.9 1 1.1−0.2
−0.1
0
0.1
ℜ(λ)
ℑ(λ
)
3 iterations
0.8 0.9 1 1.1−0.2
−0.1
0
0.1
ℜ(λ)
ℑ(λ
)
4 iterations
Hanzhi Diao (TU Delft) Helmholtz problem December 20, 2012 24 / 32
Fourier Analysis Multigrid Convergence
Multigrid Convergence Factor
• independence of k• independence of the sign of β2
0.2
0.2
0.2
0.3
0.3
0.3
0.3
0.4
0.4
0.4
0.4
0.4
0.5
0.5
0.5
0.5
0.5
0.5
0.6
0.6
0.6
0.6
0.6
0.6
0.7
0.7
0.7
0.7
0.7
0.7
0.8
0.8
0.8
0.8
0.8
0.8
0.9
0.9
0.9
0.90.9
0.9
1
1
1
1
1
11
β1
β2
0 2 4 6 8 10
−5
−3
−1
0
1
3
5
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
k = 30 and gw = 30
0.2
0.2
0.2
0.2
0.2
0.3
0.3
0.3
0.3
0.3
0.4
0.4
0.4
0.4
0.4
0.4
0.5
0.5
0.5
0.5
0.5
0.6
0.6
0.6
0.6
0.6
0.7
0.7
0.70.7
0.7
0.8
0.8
0.80.8
0.8
0.9
0.9
0.9
0.9
11
1
1
0.9
1
1
11 11
β1
β2
0 2 4 6 8 10
−5
−3
−1
0
1
3
5
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
k = 30 and gw = 120
• High resolution is favourable for the convergence.
Hanzhi Diao (TU Delft) Helmholtz problem December 20, 2012 25 / 32
Fourier Analysis Multigrid Convergence
Optimal Shift for the Preconditioner
• A small shift is favourable for the Krylov convergence of AM−1.• A large shift is favourable for the multigrid convergence of M−1.
To find out
(β1 + ιβ2)opt := arg min|β1 + ιβ2| : max16l6n−1
G(l, β1, β2) 6 c < 1.
gw = 10 gw = 30 gw = 60 gw = 120 gw = 240
m = 2 0.1096 0.0126 0 0 0m = 3 0.3228 0.0616 0.0150 0 0m = 4 0.3931 0.2002 0.0632 0.0155 0m = 5 0.3931 0.2886 0.2012 0.0636 0.0156
The optimal β2 in the shift 1 + ιβ2 for ρ(Tm1 ) 6 c = 0.9
Hanzhi Diao (TU Delft) Helmholtz problem December 20, 2012 26 / 32
Numerical Experiments
Basic Convergence Behaviour
0 100 200 300 400 500
10−14
10−12
10−10
10−8
10−6
10−4
10−2
100
iteration step m
‖r m
‖2/‖
r 0‖2
k = 100 and n = 479
A
AM−1
AM−1Q
Overview of the convergence behaviour by different preconditioning
• verify– the different preconditioning effect– the advantage of AM−1Q over AM−1
– the influence of wavenumber and wave resolution
Hanzhi Diao (TU Delft) Helmholtz problem December 20, 2012 27 / 32
Numerical Experiments
Influence of Orthogonalization
• Householder reflection outperforms modified Gram-Schmidt inconvergence behaviour.• GMRES using modified Gram-Schmidt fails to converge in the
very small system.
0 5 10 15
10−14
10−12
10−10
10−8
10−6
10−4
10−2
100
iteration step m
‖r m
‖2/‖
r 0‖2
k = 10 and n = 15
modified Gram-SchmidtHouseholder reflection
very small system
0 100 200 300 400 500
10−14
10−12
10−10
10−8
10−6
10−4
10−2
100
iteration step m
‖r m
‖2/‖
r 0‖2
k = 100 and n = 479
modified Gram-SchmidtHouseholder reflection
normal size system
Hanzhi Diao (TU Delft) Helmholtz problem December 20, 2012 28 / 32
Numerical Experiments
Influence of Approximated Preconditioning
• The inaccuracy in M−1 has little influence on the convergencebehaviour
gw k = 10 k = 50 k = 100 k = 200 k = 300 k = 400 k = 500
10 11/11 36/36 60/60 108/105 153/149 193/188 265/25830 12/12 36/36 60/58 114/108 161/152 209/196 255/24060 12/12 36/36 63/62 113/111 161/158 207/204 255/250
Number of iterations with respect to different degrees of approximations i.e. AM−1 / AM−1
• The inaccuracy in Q slows down the convergence.
gw k = 10 k = 50 k = 100 k = 200 k = 300 k = 400 k = 500
10 6/9/9 11/18/18 14/26/27 21/43/44 28/59/61 33/71/70 39/98/11130 4/5/5 6/13/13 6/15/17 8/36/37 9/54/55 10/73/75 11/92/ 9460 3/4/4 4/ 5/ 8 4/ 7/ 9 5/12/16 6/18/22 6/24/27 6/32/ 34
Number of iterations with respect to different degrees of approximations i.e. AM−1Q / AM−1Q / AM−1Q
Hanzhi Diao (TU Delft) Helmholtz problem December 20, 2012 29 / 32
Numerical Experiments
Internal Iteration in MKMG(,#2, · · · ,#m−1, )
• It is worth doing more iterations on the higher levels.• The convergence behaviour will be slowed by more iterations on
the lower levels.
0 100 200 300 400 5000
10
20
30
40
50
60
wavenumber k
number
ofiterations
gw = 30
(⋄,2,2,2,2,)(⋄,4,2,2,2,)(⋄,2,4,2,2,)(⋄,2,2,4,2,)(⋄,2,2,2,4,)
0 100 200 300 400 5000
10
20
30
40
50
60
wavenumber k
number
ofiterations
gw = 60
(⋄,2,2,2,2,)(⋄,4,2,2,2,)(⋄,2,4,2,2,)(⋄,2,2,4,2,)(⋄,2,2,2,4,)
Number of iterations with respect to different MKMG setup in a six-level grid
Hanzhi Diao (TU Delft) Helmholtz problem December 20, 2012 30 / 32
Future Work
Suggestion on Future Work
• higher dimensional problems• local Fourier analysis• different Krylov solvers
Hanzhi Diao (TU Delft) Helmholtz problem December 20, 2012 31 / 32
Thank you for watching !
Hanzhi Diao (TU Delft) Helmholtz problem December 20, 2012 32 / 32