High-Order Shifted Laplace Preconditioners for Wave EquationsIbrahim.Zangre/files/ECCOMAS... ·...
Transcript of High-Order Shifted Laplace Preconditioners for Wave EquationsIbrahim.Zangre/files/ECCOMAS... ·...
High-Order Shifted Laplace Preconditionersfor Wave Equations
Ibrahim Zangre
Institut Elie Cartan Nancy (IECN), Universite de LorraineApplied and Computational Electromagnetics (ACE), University of Liege
PhD advisors: Xavier Antoine (IECN) & Christophe Geuzaine (ACE)
(funded by Fondation EADS and ANR Microwave)
ECCOMAS 2012, September 10th 2012, Vienna.
ÉLIEInst
itu
t
CARTANNancy
I. Zangre (IECN/ACE) Preconditioning Scattering Problems ECCOMAS, Sept. 10th 2012 1 / 19
Outline
1 The scattering problem
2 Preconditioners
3 Shifted Laplace Preconditioners (SLP)
4 High-Order Shifted Laplace Preconditioners (HSLP)
5 Numerical results
6 Conclusion
I. Zangre (IECN/ACE) Preconditioning Scattering Problems ECCOMAS, Sept. 10th 2012 2 / 19
The scattering problem
Truncated sound-soft acoustic scattering problem
−∆u − k2u = 0 in Ω,
u = f on Γ,
∂nu + Bu = 0 on Γ∞,
(1)
where Ω is the bounded domain enclosed by the fictitious boundary Γ∞ and Γand where the operator B represents an approximation of the DtN operator(for example B = −ık) on Γ∞. The vector n is the outwardly directed unitnormal to Γ∞.
Sound-hard acoustic scattering problem
The same but with the Neumann BC on Γ: does not fundamentally changethe talk
I. Zangre (IECN/ACE) Preconditioning Scattering Problems ECCOMAS, Sept. 10th 2012 3 / 19
The scattering problem: an example of 2D configuration
x = 3, y = 1, e = 10−1
(Γ∞)
y
xe
(Ω)
uinc
Typically: f = −uinc = −e−ıkα·x
x = (x1, x2, x3) ∈ R3, ı =√
−1.Incidence angle α: normalized on theunit sphere (|α| = 1)k: wavenumber related tothe wavelength λ by: k = 2π/λ.
I. Zangre (IECN/ACE) Preconditioning Scattering Problems ECCOMAS, Sept. 10th 2012 4 / 19
The scattering problem
The FEM discretization of the weak form leads to the following linear system
Hhuh :=(Sh − k2
Mh − ıkBh
)uh = bh
h: mesh size
nh : total number of DOFs
Sh : stiffness matrix
Mh : mass matrix
Bh : Γ∞-surface mass matrix
nλ = λ/h: density of discretization points per wavelength
uh , bh ∈ Cnh
I. Zangre (IECN/ACE) Preconditioning Scattering Problems ECCOMAS, Sept. 10th 2012 5 / 19
The scattering problem
Difficulties (among others)
high frequency regime: we want to consider small wavelength problems:λ ≪ size(scatterer): computation of a highly oscillating solution
related to the geometry: singularities, resonances, guiding structures...
Hh is large sized, complex-valued, highly indefinite
Goals: solve efficiently the linear system
limitation of direct solvers: nh is very large for large k
we use preconditioned iterative Krylov solvers:e.g. GMRES [Saad & Schultz 87]
P−1Hhuh := P−1
(Sh − k2
Mh − ıkBh
)uh = P−1bh
I. Zangre (IECN/ACE) Preconditioning Scattering Problems ECCOMAS, Sept. 10th 2012 6 / 19
A few preconditioning techniques
Classical algebraic approaches: discretize and approximate
ILU-like factorizations
pARMS (algebraic recursive multilevel)
Linear methods (diagonal, SOR, SSOR)
Multigrid...
Limitations1
These classes of preconditioning techniques are not sufficient in practice:indeed, the solver diverges at some point for large enough values of k
But: work much better for dissipative media:k has a sufficiently large positive imaginary part
ReasonThinking only in terms of matrix is not sufficient because we loose theinformation related to the true physics and included into the operator (6=positive definite matrices which are much easier to solve)
1O. Ernst and M.J. Gander, Why it is Difficult to Solve Helmholtz Problems
with Classical Iterative Methods, 2011.
I. Zangre (IECN/ACE) Preconditioning Scattering Problems ECCOMAS, Sept. 10th 2012 7 / 19
A few preconditioning techniques
Analytic approaches: approximate and discretize
The idea based on an operator point of view: build a preconditioner based onthe operator approximation, and possibly hybridize with algebraic approaches
A popular approach is
SLP: Shifted Laplace Preconditioners [Erlangga & al. 2004]:
idea: introduce artificial dissipation
A(α) := −∆ − αk2 = −∆ − (a + ıb)k2
the discrete SLP (by FEM) is:
P = Aık(α),h := Sh − αk2
Mh − ıkBh
then SLP can be solved by ILU-like, ARMS, Multigrid, ...we typically use ILUT here to solve SLP
I. Zangre (IECN/ACE) Preconditioning Scattering Problems ECCOMAS, Sept. 10th 2012 8 / 19
Shifted Laplace Preconditioners (SLP)
1D analysis of the spectral properties:
discrete eigenvalues of H = −∆ − k2: µn = k2n − k2 (kn = nπ, n ≥ 1)
discrete eigenvalues of A(α) = −∆ − αk2: νn = k2n − αk2
discrete eigenvalues of A−1(α)H:
σn =k2
n − k2
k2n − αk2
=1 + zn
α + zn
, zn = −k2n/k2
→ σn lie on a circular arc in the right half plane.
Choice of α = a + ıb:
in general: a = 1
two conflicting arguments for b:
b is too large: solvers (ILUT, Multigrid ...) behave well, but SLP is nomore a good approximationb is too small: SLP is a good candidate, but: solvers (ILUT, Multigrid ...)encounter the same difficulties as for the original operator!
I. Zangre (IECN/ACE) Preconditioning Scattering Problems ECCOMAS, Sept. 10th 2012 9 / 19
SLP: eigenvalues clustering - 1D example
1D Example:
Ω = (0, 1), Γ = 0, Γ∞ = 1, k = 20π, nλ = 80, α = 1 + 0.5ı
Real part
Imagin
ary
part
−800 −600 −400 −200 0 200 400 600 800
−600
−400
−200
0
200
400
600
H
I. Zangre (IECN/ACE) Preconditioning Scattering Problems ECCOMAS, Sept. 10th 2012 10 / 19
SLP: eigenvalues clustering - 1D example
Real part
Imagin
ary
part
−0.4 −0.2 0 0.2 0.4 0.6 0.8 1 1.2 1.4
−0.6
−0.4
−0.2
0
0.2
0.4
0.6
0.81D - k = 20π, nλ = 80, α = 1 + 0.5ı
A−1(α)H (analytic)
zone E
zone G
zone H
[A0(α),h]
−1[Hh]
[Aik
(α),h]−1[Hh]
[Ai√
αk
(α),h ]−1[Hh]
I. Zangre (IECN/ACE) Preconditioning Scattering Problems ECCOMAS, Sept. 10th 2012 11 / 19
SLP: Gmres(50) convergence - 1D example
Here an ILUT(10−3) is used to invert SLP / nλ = 200, k ∈ [100π, 8000π]
0 0.5 1 1.5 2 2.5 3
x 104
0
1000
2000
3000
4000
5000
6000
7000
8000
9000
Wavenumber
GM
RE
SIt
erati
ons
1D - Acoustic
SLP[1+1i]SLP[1+0.17i]SLP[1+0.039i]SLP[1+0.009i]
I. Zangre (IECN/ACE) Preconditioning Scattering Problems ECCOMAS, Sept. 10th 2012 12 / 19
HSLP: High-Order Shifted Laplace Preconditioners
+ SLP introduces dissipation and positive definiteness in the problem
+ SLP can be coupled with ILU-like, Multigrid, ARMS...
− The Gmres convergence remains wavenumber-dependent
Can we still improve SLP?
→ We propose a generalization of SLP concept to provide more robustnessand efficiency through high order approximations in the microlocalanalysis framework
IDEA: compute a pseudo-inverse P2 of H such that
P2H = Iwhere P is the nonlocal pseudodifferential operator of order −1
P := H− 12 =
[−∆ − k2
]− 12
with symbol (continuous eigenvalue-like)
σ[P] =(|ξ|2 − k2
)− 12 =
1
ık(1 + z)− 1
2 , z = −|ξ|2k2
∈ R−
I. Zangre (IECN/ACE) Preconditioning Scattering Problems ECCOMAS, Sept. 10th 2012 13 / 19
HSLP: High-Order Shifted Laplace Preconditioners
Then (1 + z)−
12 is localized by Pade rational approximants:
Pℓ1−1(z)
Qℓ1(z),
Pℓ2 (z)
Qℓ2(z)
Using poles decomposition (for possible parallel implementation),
Pℓ1−1(z)/Qℓ1 (z) =
ℓ1∑
j=1
fj
z − αj
, Pℓ2 (z)/Qℓ2 (z) = g0 +
ℓ2∑
j=1
gj
z − βj
Indeed: HSLP takes the discrete form P−1 := P1P2 where
P1 :=
ℓ1∑
j=1
fj
(Sh + αjk
2Mh + ıkBh
)−1
,P2 := g0I−k2Mh
ℓ2∑
j=1
gj
(Sh + βjk
2Mh + ıkBh
)−1
Finally: code implementation of HSLP is straightforward from that of SLP.
1D example ofeigenvalue clustering:
ℓ1 = ℓ2 = 2 (left)
ℓ1 = ℓ2 = 4 (right)
−0.5 0 0.5 1 1.5−0.8
−0.6
−0.4
−0.2
0
0.2
0.4
0.6
0.8HSLP[2|2|π
2 ]
Real part
Imag
inar
ypar
t
analytic arc
HSLP[2|2|π2 ]
−0.5 0 0.5 1 1.5
−0.6
−0.4
−0.2
0
0.2
0.4
0.6
HSLP[4|4|π2 ]
Real part
Imag
inar
ypar
t
analytic arc
HSLP[4|4|π2 ]
I. Zangre (IECN/ACE) Preconditioning Scattering Problems ECCOMAS, Sept. 10th 2012 14 / 19
1D High frequencies examples: k ∈ [100π, 8000π]
Preconditioners are solved using an ILUT(10−3) / nλ = 200
0 2000 4000 6000 8000 10000 12000 140000
500
1000
1500
2000
2500
3000
3500
4000
4500
5000
Wavenumber
GM
RE
SIt
erati
ons
1D - Acoustic – Npade = 1
SLP[1+1i]
HSLP[1]
0 0.5 1 1.5 2 2.5 3
x 104
0
500
1000
1500
2000
2500
3000
3500
4000
4500
Wavenumber
GM
RE
SIt
erati
ons
1D - Acoustic – Npade = 2
SLP[1+0.17i]
HSLP[2]
I. Zangre (IECN/ACE) Preconditioning Scattering Problems ECCOMAS, Sept. 10th 2012 15 / 19
1D High frequencies examples: k ∈ [100π, 8000π]
Preconditioners are solved using an ILUT(10−3) / nλ = 200
0 0.5 1 1.5 2 2.5 3
x 104
0
500
1000
1500
Wavenumber
GM
RE
SIt
erati
ons
1D - Acoustic – Npade = 4
SLP[1+0.039i]
HSLP[4]
0 1 2 3 4 5 6 7
x 104
0
200
400
600
800
1000
1200
Wavenumber
GM
RE
SIt
erati
ons
1D - Acoustic – Npade = 8
SLP[1+0.009i]
HSLP[8]
I. Zangre (IECN/ACE) Preconditioning Scattering Problems ECCOMAS, Sept. 10th 2012 16 / 19
2D case: scattering by cavity: k = 15π
x = 3, y = 1, e = 10−1
(Γ∞)
y
xe
(Ω)
uinc
Preconditioners are solved using anILUT(10−3) / nλ = 10
GMRES(50) iterations
GM
RE
SR
esid
ual
2D Cavity, nλ = 10, k = 15π
0 50 100 150 200 250 30010−10
10−9
10−8
10−7
10−6
10−5
10−4
10−3
10−2
10−1
100
SLP[1+0.009i]HSLP[8]
I. Zangre (IECN/ACE) Preconditioning Scattering Problems ECCOMAS, Sept. 10th 2012 17 / 19
Conclusion
HSLPIs a Generalization of the concept of SLP
Based idea: Square-Root operator + Pade Approximation + ILU
The limitations of HSLP are similar to the limitations of the pole’simaginary part of SLP
Hence, one important perspective is to find good rational approximationswith the possible largest imaginary parts
Extension to Maxwell equations is straightforward and already done(computational tests under progress)
I. Zangre (IECN/ACE) Preconditioning Scattering Problems ECCOMAS, Sept. 10th 2012 18 / 19
Conclusion
Merci pour votre attention!
I. Zangre (IECN/ACE) Preconditioning Scattering Problems ECCOMAS, Sept. 10th 2012 19 / 19
I. Zangre (IECN/ACE) Preconditioning Scattering Problems ECCOMAS, Sept. 10th 2012 20 / 19
I. Zangre (IECN/ACE) Preconditioning Scattering Problems ECCOMAS, Sept. 10th 2012 21 / 19
Shifted Laplace Preconditioners (SLP)
Let us consider the principal symbol of H and A(α) (ξ is the Fourier variable)
σ[H](ξ) = |ξ|2 − k2 and σ[A(α)](ξ) = |ξ|2 − αk2, ∀ξ ∈ Rd .
Then the symbol of the preconditioned operator A−1(α)H is
σ[A−1(α)
H](z) =|ξ|2 − k2
|ξ|2 − αk2=
1 + z
α + z, ∀z = −|ξ|2
k2∈ R
−.
→ The “continuous eigenvalues” of A−1(α)H ly on a circular arc with endpoints:
α−1 related to the propagative modes:H :=
(k, ξ) ∈ R × Rd/|ξ| < k
(z → 0)
(1, 0) related to the evanescent modes:E :=
(k, ξ) ∈ R × Rd/|ξ| > k
(z → −∞)
(0, 0) related to the grazing modes:G :=
(k, ξ) ∈ R × Rd/|ξ| ≈ k
(z → −1)
I. Zangre (IECN/ACE) Preconditioning Scattering Problems ECCOMAS, Sept. 10th 2012 22 / 19
SLP: FEM formulation
The discrete Shifted Laplace operator reads
If one considers the classic ABC ∂nu − ıku = 0 on Γ∞:
Aık(α),h := Sh − αk2
Mh − ıkBh .
or the well-suited ABC ∂nu − ı√
αku = 0 on Γ∞:
Aı√
αk
(α),h := Sh − αk2Mh − ı
√αkBh .
1D Example:
uinc = e−ıkx1 , Ω = (0, 1), Γ = 0, Γ∞ = 1k = 20π, nλ = 80, α = 1 + 0.5ıSLP preconditioners are solved by an exact LU.
I. Zangre (IECN/ACE) Preconditioning Scattering Problems ECCOMAS, Sept. 10th 2012 23 / 19
SLP: continuous eigenvalues
Real part
Imagin
ary
part
−800 −600 −400 −200 0 200 400 600 800
−600
−400
−200
0
200
400
600
H
I. Zangre (IECN/ACE) Preconditioning Scattering Problems ECCOMAS, Sept. 10th 2012 24 / 19
SLP: eigenvalues clustering
Real part
Imagin
ary
part
−0.4 −0.2 0 0.2 0.4 0.6 0.8 1 1.2 1.4
−0.6
−0.4
−0.2
0
0.2
0.4
0.6
0.8
A−1(α)H (analytic)
zone E
zone G
zone H
[A0(α),h]−1[Hh]
[Aik
(α),h]−1[Hh]
[Ai√
αk
(α),h ]−1[Hh]
I. Zangre (IECN/ACE) Preconditioning Scattering Problems ECCOMAS, Sept. 10th 2012 25 / 19
SLP: limitations
+ SLP can be solved efficiently by multigrids methodsIt is likewise well-suited for ILU-like preconditioning(provides factorization stability because of the dissipative term)It provides positive definiteness to the Helmholtz problem
− Many eigenvalues remain close to zero (grazing modes)Convergence remains dependent of the wavenumber (linearly)
→ We propose a generalization of SLP concept to provide robustness andefficiency through high order approximations in the microlocal analysisframework
I. Zangre (IECN/ACE) Preconditioning Scattering Problems ECCOMAS, Sept. 10th 2012 26 / 19
High-Order Shifted Laplace Preconditioners (HSLP)
We propose to compute a pseudo-inverse P2 of H such that
P2H = I
where P is the nonlocal pseudodifferential operator of order −1
P := H− 12 =
[−∆ − k2
]− 12
with symbol
σ[P] =(|ξ|2 − k2
)− 12 =
1
ık(1 + z)− 1
2 , z = −|ξ|2k2
∈ R−
Then P can be localized by using Pade rational approximants of
fτ (z) = (1 + z)τ , τ ∈ R.
I. Zangre (IECN/ACE) Preconditioning Scattering Problems ECCOMAS, Sept. 10th 2012 27 / 19
HSLP: Pade rational approximants
They are of the form
fτ (z) ≈ f[ℓ/m]τ (z) =
Pℓ(z)
Qm(z)
where Pℓ and Qm are polynomials of degrees ℓ and m; and provide an(ℓ + m + 1)-order approximation.
The coefficients of Pℓ and Qm can be computed by solving a certain linearsystem.
For numerical purposes, we use the following decomposition
f[ℓ/m]τ (z) =
Pℓ(z)
Qm(z)= r
[ℓ/m]0 (z) +
m∑
j=1
r[ℓ/m]j
z − q[ℓ/m]j
q[ℓ/m]j : j = 1, · · · , m are the zeros of Qm(z)
r[ℓ/m]j : j = 0, · · · , m are given from the zeros of Pℓ(z) and Qm(z).
I. Zangre (IECN/ACE) Preconditioning Scattering Problems ECCOMAS, Sept. 10th 2012 28 / 19
HSLP: Pade rational approximants (τ = −1/2)
Applying a θ-rotation of the branch-cut:
f− 12(z) =
1√1 + z
=1
eı θ
2
√e−ıθ(1 + z)
=e−ı θ
2
√1 + [e−ıθ(1 + z) − 1]
≃ e−ı θ
2 f[ℓ/m]
− 12
(e−ıθ(1 + z) − 1
)= r
[ℓ/m]0 (z) +
m∑
j=1
r[ℓ/m]j
z − q[ℓ/m]j
This formulation is more accurate for both regions z > −1 and z < −1
I. Zangre (IECN/ACE) Preconditioning Scattering Problems ECCOMAS, Sept. 10th 2012 29 / 19
HSLP: Implicit localization of HSLP via Pade
The exact pseudo-inverse of H can then be represented by the approximateprincipal symbol
σ[H−1] = σ[P2]
= − 1
k2
1
(1 + z)
≃ − 1
k2
[r[ℓ/m]0 +
m∑
j=1
r[ℓ/m]j
z − q[ℓ/m]j
] [r[ℓ/m]0 +
m∑
j=1
r[ℓ/m]j
z − q[ℓ/m]j
]
P2 is of order −2, and so must be its symbol. This is not the case here:2(ℓ − m) + 2(ℓ − m)→ thus we split the symmetric form above
I. Zangre (IECN/ACE) Preconditioning Scattering Problems ECCOMAS, Sept. 10th 2012 30 / 19
HSLP: Implicit localization of HSLP via Pade
We introduce P1 and P2 such that
σ[P1]
=1
ıke−ı θ
2 f[ℓ1−1/ℓ1]
− 12
(e−ıθ(1 + z) − 1
)=
1
ık
[ℓ1∑
j=1
r[ℓ1−1/ℓ1]j
z − q[ℓ1−1/ℓ1]j
]
and
σ[P2]
=1
ıke−ı θ
2 f[ℓ2/ℓ2]
− 12
(e−ıθ(1 + z) − 1
)=
1
ık
[r[ℓ2/ℓ2]0 +
ℓ2∑
j=1
r[ℓ2/ℓ2]j
z − q[ℓ2/ℓ2]j
].
I. Zangre (IECN/ACE) Preconditioning Scattering Problems ECCOMAS, Sept. 10th 2012 31 / 19
HSLP: Implicit localization of HSLP via Pade
Then the exact pseudo-inverse of H is approximated by H−1 ≃ P1P2 and itsprincipal symbol reads
σ[H−1] ≃ σ[P1P2]
= − 1
k2
[ℓ1∑
j=1
r[ℓ1−1/ℓ1]j
z − q[ℓ1−1/ℓ1]j
] [r[ℓ2/ℓ2]0 +
ℓ2∑
j=1
r[ℓ2/ℓ2]j
z − q[ℓ2/ℓ2]j
]
Then the global order of the approximation is
−2 := 2(ℓ1 − 1 − ℓ1) + 2(ℓ2 − ℓ2)
and the preconditioned operator P1P2H is a perturbation of the identityoperator. We designate this approximation as HSLP[ℓ1|ℓ2|θ]; then theHSLP[1|0|θ] case reads
σ[P1P2H]
= 2eı θ
21 + z
(1 − eıθ) + z
→ HSLP[1|0|θ] is exactly the same as SLP given in with α = 1 − eıθ.For example, for θ = π/2, one finds α = 1 + ı.
I. Zangre (IECN/ACE) Preconditioning Scattering Problems ECCOMAS, Sept. 10th 2012 32 / 19
HSLP: algebraic formulation
Let us come back to the Helmholtz matrix
[Hh ] = [Sh ] − k2[Mh ] − ık[Bh ]
Setting [Sh ] = [Sh ] − ık[Bh ] . It follows that
[Hh ] = [Sh ] − k2[Mh ]
= −k2[Mh ]
([I] − 1
k2[Mh ]−1[Sh ]
).
Therefore, the inverse of [Hh ] writes
[Hh ]−1 = − 1
k2
([I] − 1
k2[Mh ]−1[Sh ]
)−1
[Mh ]−1.
([I] − 1
k2 [Mh ]−1[Sh ])−1
is evaluated using pseudo-representation of (1 + z)−1,
and one finds
[Hh ]−1 ≃ P1P2
I. Zangre (IECN/ACE) Preconditioning Scattering Problems ECCOMAS, Sept. 10th 2012 33 / 19
HSLP: algebraic formulation
where P1 and P2 are given by:
P1 :=
[ℓ1∑
j=1
r[ℓ1−1/ℓ1]j
([Sh ] + q
[ℓ1−1/ℓ1]j k2[Mh ]
)−1]
P2 :=
[r[ℓ2/ℓ2]0 [I] − k2[Mh ]
ℓ2∑
j=1
r[ℓ2/ℓ2]j
([Sh ] + q
[ℓ2/ℓ2]j k2[Mh ]
)−1]
Then Helmholtz matrix is then preconditioned as follows:
P1P2[Hh ]
I. Zangre (IECN/ACE) Preconditioning Scattering Problems ECCOMAS, Sept. 10th 2012 34 / 19