implicit a posteriori error estimation for the time ...web.cs.elte.hu/~izsakf/talk060614.pdf ·...
TRANSCRIPT
Implicit a posteriori error estimation for the
time-harmonic Maxwell equations
Davit Harutyunyan, Ferenc Izsak∗ and Jaap van der Vegt
Numerical Analysis and Computational Mechanics Group
Department of Applied Mathematics
University of Twente, The Netherlands
MAFELAP Conference, Brunel University, London
14 June, 2006
∗ Also at Department of Applied Analysis and Computational Mathematics, ELTE, Budapest
University of Twente - Chair Numerical Analysis and Computational Mechanics 1
Outline of Presentation
• Time harmonic Maxwell equations
• Definition of the local error equations
• Numerical solution of the local error equations
• Theoretical results for the local error equations and implicit error estimates
• Numerical examples
University of Twente - Chair Numerical Analysis and Computational Mechanics 2
Time Harmonic Maxwell Equations
• Time harmonic Maxwell equations - function spaces with perfectly conducting
boundary conditions:
curl curl E − k2E = J , in Ω ⊂ R
3
E × ν = 0, on ∂Ω,(1)
with J ∈ [L2(Ω)]3 a given source term and ν the outward normal on ∂Ω.
• The Hilbert space corresponding to the Maxwell equation is:
H(curl, Ω) = u ∈ [L2(Ω)]3: curl u ∈ [L2(Ω)]
3,
H0(curl, Ω) = u ∈ H(curl, Ω) : ν × u|∂Ω = 0
equipped with the curl norm:
‖u‖curl,Ω = (‖u‖2[L2(Ω)]3 + ‖curl u‖2
[L2(Ω)]3)1/2.
University of Twente - Chair Numerical Analysis and Computational Mechanics 3
Definitions
• Define the bilinear form BK on H(curl, K) × H(curl, K) and the trace
operators γτ and πτ with
BK(u, v) = (curl u, curl v)K − k2(u, v)K
γτv = ν × v|∂K and πτv = (ν × v|∂K) × ν.
• Weak formulation of (1): Find an E ∈ H0(curl, Ω) such that
BΩ(E, v) = (J , v), ∀ v ∈ H0(curl, Ω).
• Discrete weak form of (1): Find an Eh ∈ Hh0 (curl, Ω) such that
BΩ(Eh, vh) = (J , vh), ∀ vh ∈ Hh0 (curl, Ω),
where Hh0 is the Nedelec edge element space (e.g. of order 1).
University of Twente - Chair Numerical Analysis and Computational Mechanics 4
Error Estimation: Basic Setting
• Define the computational error on an element K as:
eh = (E − Eh)|K.
• Overall aim: estimate ‖eh‖curl,K.
• Informations at hand:
⊲ J , Eh with boundary conditions, the subdomain K.
⊲ The element residual on K
rK := J − curl curl Eh + k2Eh in K.
⊲ The tangential jump of the curl at the element face lj between K and Kj
Rlj=
1
2(νj ×
[
curl Eh|K − curl Eh|Kj
]
).
University of Twente - Chair Numerical Analysis and Computational Mechanics 5
A Posteriori Error Estimates
• Objective: estimation of the error for an existing approximation.
• Main tool: weak form for the error, localization to the subdomains.
• Two types of a posteriori error estimates are frequently used:
Explicit error estimates, which use data provided by the numerical solution;
practically a “simple formula” depending on known data.
Implicit error estimates, which solve local error equations with the data provided
by the numerical solution.
University of Twente - Chair Numerical Analysis and Computational Mechanics 6
Local Error Equation
• Variational formulation for the local error: Find an eh ∈ H(curl, K) such that:
BK(eh, v) = (curl eh, curl v)K − k2(eh, v)K
= (curl E, curl v)K − k2(E, v)K − ((curl Eh, curl v)K − k
2(Eh, v)K)
= (J, v)K − (γτcurl E, πτv)∂K − BK(Eh, v), ∀v ∈ H(curl, K).
• At the faces lj of element K we use the approximation:
γτcurl E|lj ≈ γτcurl E|lj =1
2(νj ×
[
curl Eh|K + curl Eh|Kj
]
).
University of Twente - Chair Numerical Analysis and Computational Mechanics 7
Numerical Solution of Local Problems
• Local error formulation:
On each element K, find eh ∈ Vh,K, the implicit error estimate such that
∀ w ∈ Vh,K the following relation is satisfied:
(curl eh, curl w)K − k2(eh, w)K = (J , w)K − (curl Eh, curl w)K
+k2(Eh, w)K − 1
2(νj ×
(
curl Eh|K + curl Eh|Kj
)
, w)∂K.
• Chief question: how to choose Vh,K?
University of Twente - Chair Numerical Analysis and Computational Mechanics 8
Basis Functions for the Local Error Equation
On the reference element K = (0, 1)3 define Vh,K = span(φ0
k9k=1), where
φ01 = ((1 − ξ)(1 − η)η(1 − ζ)ζ, 0, 0)
T, φ
02 = (ξ(1 − η)η(1 − ζ)ζ, 0, 0)
T,
φ07 = ((1 − ξ)ξ(1 − η)η(1 − ζ)ζ, 0, 0)
T, φ
03 = (0, (1 − ξ)ξ(1 − η)(1 − ζ)ζ, 0)
T,
φ04 = (0, (1 − ξ)ξη(1 − ζ)ζ, 0)
T, φ
08 = (0, (1 − ξ)ξ(1 − η)η(1 − ζ)ζ, 0)
T,
φ05 = (0, 0, (1 − ξ)ξ(1 − η)η(1 − ζ))
T, φ
06 = (0, 0, (1 − ξ)ξ(1 − η)ηζ)
T,
φ09 = (0, 0, (1 − ξ)ξ(1 − η)η(1 − ζ)ζ)
T.
The space Vh,K is obtained with an affine transformation; K is rectangular.
University of Twente - Chair Numerical Analysis and Computational Mechanics 9
Well Posedness of Local Problems
• Theorem 1 Assume that k is not a Maxwell eigenvalue in the sense that the
problem: Find u ∈ H(curl, K) such that
B(u, v) = 0, ∀ v ∈ H(curl, K)
has only the trivial (u = 0) solution. Then the variational problem:
Find eh ∈ H(curl, K) such that
B(eh, v) = (J , v), ∀ v ∈ H(curl, K) (2)
has a unique solution for all J ∈ [L2(K)]3.
• Using lifting, the variational problem for the error can be rewritten into (2).
• This is also the case for any reasonable choice of γτcurl E.
University of Twente - Chair Numerical Analysis and Computational Mechanics 10
Eigenvalues for the Local Error Equation on a Rectangular
Domain
• Theorem 2 The eigenfunctions in the Maxwell eigenvalue problem
(corresponding to (2)):
Find v ∈ H(curl, K) and k ∈ R such that
curl curl v − k2v = 0 in K,
ν × curl v = 0 on ∂K
with K = (0, a) × (0, b) × (0, c) are:
v(x, y, z) =
C1 sink1π
a x cosk2π
b y cosk3π
c z
C2 cosk1π
a x sink2π
b y cosk3π
c z
C3 cosk1π
a x cosk2π
b y sink3π
c z
,
University of Twente - Chair Numerical Analysis and Computational Mechanics 11
for any k1, k2, k3 ∈ N with
k1π
aC1 +
k2π
bC2 +
k3π
cC3 = 0 and C1, C2, C3 ∈ R.
The appropriate eigenvalues are k2 =(
k1π
a
)2
+(
k2π
b
)2
+(
k3π
c
)2
with
k1, k2, k3 ∈ N arbitrary.
⋄ Example: a = b = c = 1 (unit cube).
Smallest eigenvalue: π2(12 + 12 + 12) ≈ 29.6.
Corollary 1 (2) is well posed whenever k is not a Maxwell eigenvalue in the sense
of Theorem 2.
University of Twente - Chair Numerical Analysis and Computational Mechanics 12
Implicit Error Estimate as Lower Bound of the Error
• Introduce ηK as error indicator - an explicit error estimate on K:
ηK = (h2‖r‖2[L2(K)]3 + h‖R‖2
[L2(∂K)]3)12 .
• Theorem 3 The implicit a posteriori error estimate eh can be used as a lower
bound for the exact error with respect to the curl norm as follows:
‖eh‖curl,K ≤ CηK ≤√
DC(1 + k2)‖eh‖curl,K + Res(r, R)
with C, D constants independent of the mesh size h and frequency k and
Res(r, R) = (h2‖r − r‖[L2(K)]3 + h‖R − R‖[L2(∂K)]3)1/2
arising from the interpolation errors of r and R.
University of Twente - Chair Numerical Analysis and Computational Mechanics 13
Solution of the local problems
K: cubic subdomain with edge length h.
The bilinear form for the bubble function space span(φk9k=1) on K:
• B1,K - stiffness matrix of size 9 × 9.
• B1,K = B1,curl − k2B1,0, where
B1,curl[i][j] =
∫
K
curl φi · curl φj, B1,0[i][j] =
∫
K
φi · φj.
• Scaling: B1,K = 1hB1,curl − k2hB1,0.
University of Twente - Chair Numerical Analysis and Computational Mechanics 14
Properties of the local problems
Ill conditioned matrices arise:
cond(B1,K) ∼ 1
h2.
The solution of the local problem seems to become less precise as h → 0.
Theorem 4 The following inf-sup condition holds:
There is a C ∈ R such that for any u ∈ span(φj9j=1)
sup‖v‖curl,K=1
v∈span(φj9j=1)
|B1,K(u, v)| ≥ C‖u‖curl,K,
where C does not depend on u and the finiteness parameter h.
University of Twente - Chair Numerical Analysis and Computational Mechanics 15
Main difficulties and tasks
• Challenges in the analysis and computations
Non-coercive bilinear forms in 3D.
Non-smooth solutions arise.
Non-convex domains with edges, corners arise.
Presence of big or even small wave numbers (k).
• The performance of the implicit error estimator is evaluated by:
checking if the estimator provides the right type of error distribution
comparing the magnitude of the global error estimate and its convergence under
mesh refinement with the exact error.
University of Twente - Chair Numerical Analysis and Computational Mechanics 16
Definition of Error Contributions
• The exact error δK and the implicit local error estimate δK on element K are
defined as:
δK = ‖E − Eh‖curl,K, δK = ‖eh‖curl,K.
• The exact global error δ and the implicit global error estimate δh are obtained by
summing the local contributions
δ =
∑
K∈Th
δ2K
1/2
, δh =
∑
K∈Th
δ2K
1/2
.
• In all cases: comparison of
analytic error : ‖E − Eh‖curl,K = ‖eh‖curl,K and implicit error : ‖eh‖curl,K.
University of Twente - Chair Numerical Analysis and Computational Mechanics 17
Test case 1: Smooth Solution
• The Maxwell equations on Ω = (0, 1)3 with given source term
J(x, y, z) = (π2(p2 + m2) − 1)
sin(πpy) sin(πmz)
sin(πpz) sin(πmx)
sin(πpx) sin(πmy)
with p, m ∈ N, which has a smooth solution:
E(x, y, z) =
sin(πpy) sin(πmz)
sin(πpz) sin(πmx)
sin(πpx) sin(πmy)
.
University of Twente - Chair Numerical Analysis and Computational Mechanics 18
Test case 1: Locations for Local Error Computations
0 0.2 0.4 0.6 0.8 1
0
0.5
10
0.2
0.4
0.6
0.8
1
2⋅
7⋅16⋅
4⋅
9⋅
14⋅5⋅
1⋅
6⋅
12⋅3⋅
8⋅
Location of some elements where the implicit error estimation was conducted.
University of Twente - Chair Numerical Analysis and Computational Mechanics 19
Test case 1: Error Distribution for Smooth Solution
1 2 3 4 5 6 7 8 9 12 14 16
3
6
8x 10
−3
loca
l err
or e
stim
ator
implicit error
Error distribution in H(curl)-norm (p = m = 1, h = 116).
University of Twente - Chair Numerical Analysis and Computational Mechanics 20
Test case 1: Error distribution for Smooth Solution
1 2 3 4 5 6 7 8 9 12 14 16
0.02
0.04
0.06
0.08
0.1
0.13
loca
l err
or e
stim
ator
implicit error
Error distribution in H(curl) norm (p = 5, m = 1, h = 116).
University of Twente - Chair Numerical Analysis and Computational Mechanics 21
Test case 1: Global Error for Smooth Solution
Variation of the global error estimate and exact error in the
H(curl)-norm (p = m = 1).
University of Twente - Chair Numerical Analysis and Computational Mechanics 22
Test case 2: Non-Smooth Solution
• Consider the domain Ω = (−1, 1)3 and the function:
f : Ω → R with f = max|x|, |y|, |z|.
• Define E : Ω → R as E := −∇f(x, y, z).
Then E solves the following Maxwell problem:
curl curl E − E = ∇f in Ω,
E × ν = 0 on ∂Ω.
• The right hand side is in [L2(Ω)]3, but the exact solution is not smooth,
E 6∈ [H12(Ω)]3, but E ∈ [H
12−ǫ(Ω)]3 for any ǫ ∈ R
+.
University of Twente - Chair Numerical Analysis and Computational Mechanics 23
Test case 2: Locations for Local Error Computations
−1 −0.5 0 0.5 1
−1
0
1−1
−0.5
0
0.5
1
2⋅
7⋅
4⋅
9⋅
5⋅
14⋅
1⋅
6⋅
12⋅
17⋅
3⋅
8⋅
Location of some elements where the implicit error estimation was conducted.
University of Twente - Chair Numerical Analysis and Computational Mechanics 24
Test case 2: Error Distribution for Non-Smooth Solution
1 2 3 4 5 6 7 8 9 12 14 170
0.006
0.009
0.011
0.014
loca
l err
or e
stim
ator
implicit error
Error distribution in H(curl)-norm (h = 116).
University of Twente - Chair Numerical Analysis and Computational Mechanics 25
Test case 2: Global Error for Non-Smooth Solution
Variation of the global error estimate and exact error in the H(curl)-norm.
University of Twente - Chair Numerical Analysis and Computational Mechanics 26
Test case 3: Fichera Cube
• Consider a Fichera cube Ω = (−1, 1)3 \ [−1, 0]3.
• The solution on this domain has corner and edge singularities and can serve as an
interesting test case.
• Boundary conditions and source term in the Maxwell equations are chosen such
that the exact solution is E = grad(r2/3 sin(23φ)) in spherical coordinates, with
r =√
x2 + y2 + z2, φ = arccosr
z.
• Note, E does not belong to [H1(Ω)]3.
University of Twente - Chair Numerical Analysis and Computational Mechanics 27
Test case 3: Locations for Local Error Computations
−1−0.5
00.5
1
−1
0
1−1
−0.5
0
0.5
1
25⋅
17⋅8⋅
29⋅
10⋅
1234567
⋅⋅⋅⋅⋅⋅⋅⋅15⋅14
31⋅
27⋅
Location of some elements where the implicit error estimation was conducted.
University of Twente - Chair Numerical Analysis and Computational Mechanics 28
Test case 3: Error Distribution for Fichera Cube Solution
1 2 3 4 5 6 7 8 10 1415 17 25 27 29 31
0.002
0.006
0.01
0.016
0.02
loca
l err
or e
stim
ator
implicit error
Error distribution in H(curl)-norm (h = 18)
University of Twente - Chair Numerical Analysis and Computational Mechanics 29
Test case 3: Fichera Cube
Variation of the global error estimate and exact error in the H(curl)-norm.
University of Twente - Chair Numerical Analysis and Computational Mechanics 30
Comparisons with some existing schemes
• Beck, Hiptmair et al. (M2AN 2000) consider the following elliptic boundary value
problem
curl (curl E) + βE = J in Ω = (0, 1)3,
E × ν = 0 on ∂Ω,
where β is a given positive function on Ω.
• The exact solution is given by E = (0, 0, sin(πx)).
• Note, the bilinear form for this problem is coercive, contrary to the bilinear form
discussed for the Maxwell equations.
• We make comparisons in terms of the effectivity index εh :=δh
δ. This quantity
merely reflects the quality of the global estimate.
University of Twente - Chair Numerical Analysis and Computational Mechanics 31
Effectivity Index
β
Level0 1 2 3 4 5
10−4 4.05 8.05 8.18 8.24 8.27 8.29
10−2 4.05 8.05 8.17 8.23 8.27 8.29
1 3.01 7.64 7.78 7.84 7.87 7.89
102 2.29 4.27 4.70 4.95 5.20 5.26
104 2.33 4.23 4.66 4.86 4.95 5.00
Effectivity index εh for the error estimator given by Beck et. al.
University of Twente - Chair Numerical Analysis and Computational Mechanics 32
Effectivity Index
β
h 1
4
1
8
1
16
1
32
1
6410−4 0.67 0.67 0.67 0.67 0.67
10−2 0.67 0.67 0.67 0.67 0.67
1 0.67 0.67 0.67 0.67 0.67
102 0.63 0.65 0.67 0.67 0.67
104 0.44 0.37 0.40 0.53 0.63
Effectivity index εh for the new implicit error estimator.
University of Twente - Chair Numerical Analysis and Computational Mechanics 33
Conclusions
• We have developed and analyzed an implicit a posteriori error estimation technique
for the time harmonic Maxwell equations.
• The algorithm is well suited both for cases where the bilinear form is coercive and
the more complicated indefinite case.
• The algorithm is tested on a number of increasingly complicated test cases and the
results show that it gives an accurate prediction of the error distribution and also
the local and global error.
• Also, in comparison with other a posteriori estimation techniques, it gives for the
cases considered a sharper estimate of the error and the error distribution.
University of Twente - Chair Numerical Analysis and Computational Mechanics 34
Further works
• The same procedure on a tetrahedral mesh (in progress).
• Chief question: how to use the local basis?
The choice span(Φj9j=1) does the job, but what else could be chosen?
• A meaningful adaptation strategy using the error estimates.