Introduction to the Boundary Element Method
Salim Meddahi
University of Oviedo, Spain
University of Trento, Trento April 27 - May 15, 2015
1
Potential theory The Sobolev setting
2
Can we extend the technique to the usual Sobolev setting?
We want generalize the previous strategies in the following situation
I Ω ⊂ Rd(d = 2, 3) bounded polygonal/polyhedral and Lipschitz domain
I Jγ uK := ψ ∈ H
1/2(Γ) and J∂nuK := λ ∈ H
−1/2(Γ)
The first step consists in finding weak versions of the single layer potential Sλ(x) :=
Z
Γ
E(x, y)λ(y)dsy x ∈ Ω ∪ Ωe and the double layer potential
Dψ(x) :=
Z
Γ
∂nyE(x, y)ψ(y)dsy x ∈ Ω ∪ Ωe
3
The Sobolev space H
1(Ω) and the trace operator
I The norm in L2(Ω) is denoted k·k0,Ω
I H1(Ω) is the classical Sobolev space of order one
I The norm in H1(Ω) is denoted k·k1,Ω I H12(Γ) := {ψ ∈ L2(Γ); |ψ|1/2,Γ< ∞} with
|ψ|21/2,Γ:=
Z
Γ
Z
Γ
|ψ(x) − ψ(y)|2
|x − y|d dsydsx
I The restriction operator defined for v ∈ C0( ¯Ω) by γv := v|Γhas a unique extension to a bounded linear operator
γ : H1(Ω) → H12(Γ)
I For any ψ ∈ H12(Γ), there exists Eψ ∈ H1(Ω) such that γEψ = ψ and kEψk1,Ω≤ Ckψk1/2,Γ ∀ψ ∈ H12(Γ)
4
The normal trace operator
I Since the inclusion H12(Γ) ⊂ L2(Γ) is dense and continuous, we can consider the Gelfand triple
H12(Γ) ⊂ L2(Γ) ⊂ H−12(Γ) := (H12(Γ))0
I The duality product will be denoted hλ, ϕiΓand if λ ∈ L2(Γ) hλ, ϕiΓ=
Z
Γ
λ(x)ϕ(x)dsx ∀ϕ ∈ H12(Γ)
I The linear operator ∂n: {v ∈ H1(Ω); ∆v ∈ L2(Ω)} → H−12(Γ) given by h∂nv, ϕiΓ:=
Z
Ω
∇v · ∇Eϕ + ∆v Eϕ ∀ϕ ∈ H12(Γ) is well-defined and continuous
5
A weighted Sobolev space
For d = 2 and d = 3 we consider, W1(Ωe) :=n
u : Rd→ R; ρ u ∈ L2(Ωe), ∇u ∈ L2(Ωe)do
where ρ(x) :=
(1 + |x|2)−1/2 if d = 3 (1 + |x|2)−1/2log(2 + |x|2)−1 if d = 2.
and introduce W01(Ωe) :=D(Ωe).
I W1(Ωe) is a Hilbert space when endowed with kvk2W1(Ωe):= kρvk20,Ωe+ k∇vk20,Ωe I 1 ∈ W1(Ωe) for d = 2 but 1 /∈ W1(Ωe) for d = 3
I W1(Ωe) ⊂ Hloc1 (Ωe) (the weight only affects the behaviour at infinity)
I H1(Ωe) ,→ W1(Ωe)
I There is a bounded right-inverse for the trace operator γe: W1(Ωe) → H12(Γ)
6
A norm equivalence
Proposition 1.
The norms v 7→ k∇vk0,Be1 and v 7→ kvkW1(Be1) are equivalent on W01(B1e).
Theorem 2.
The mapping v 7→ k∇vk0,Ωe+ (3 − d)|R
Γγev| defines a norm on W1(Ωe) and W01(Ωe) that is equivalent to v 7→ kvkW1(Ωe).
Proposition 3.
Assume that d = 3. For any λ ∈ H−12(Γ) and ψ ∈ H12(Γ), the problem
Find u ∈ H1(Ω) × W1(Ωe) such thatJγ uK = ψ and Z
Ω∪Ωe
∇u · ∇v = hλ, γviΓ ∀v ∈ W1(R3) (1) is well-posed.
7
Canonical splitting of the solution
It is clear that if u solves Problem (1) then:
I u = (ui, ue) ∈ H1(Ω) × W1(Ωe)
I ∆u = 0 in Ω ∪ Ωe
I Jγ uK = ψ and J∂nuK = λ By superposition
u = ˜S(J∂nuK) −D(˜ Jγ uK) where
H−12(Γ) 3 λ → ˜S(λ) ∈ W1(R3) satisfies Z
R3
∇ ˜S(λ) · ∇v = hλ, γvi ∀v ∈ W1(R3) and
H12(Γ) 3 ψ → ˜D(ψ) ∈ H1(Ω) × W1(Ωe) is such thatJγD(ψ)˜ K = −ψ and Z
Ω∪Ωe
∇ ˜D(ψ) · ∇v = 0 ∀v ∈ W1(R3)
Corollary 4.
The operators ˜S : H−12(Γ) → W1(R3) and ˜D : H12(Γ) → H1(Ω) × W1(Ωe) are linear and bounded.
8
Weak form of the integral representation formula
Given λ ∈ H−12(Γ) and ψ ∈ H12(Γ), we consider the compactly supported distributions γtλ and ∂ntψ given by
hγtλ, ϕi = hλ, γϕiΓ and h∂ntψ, ϕi = hψ, ∂nϕiΓ ∀ϕ ∈ D(R3)
Theorem 5.
Assume that w ∈ H1(Ω) × W1(Ωe) and ∆w = 0 in Ω ∪ Ωe then
−∆w = γtJ∂nwK − ∂
t
nJγ wK in D0(R3).
Moreover, w is given by
w = F ∗ γtJ∂nwK − F ∗ ∂
t nJγ wK and it holds that
F ∗ γtλ = ˜Sλ ∀λ ∈ H−12(Γ) and
F ∗ ∂ntφ = ˜Dφ ∀φ ∈ H12(Γ)
We conclude from the theorem that the unique solution u of (1) is given by u = F ∗ γtλ − F ∗ ∂ntψ
9
Properties of S and D
As a consequence, we have the following result.
Theorem 6.
There holds the mapping properties (i) S : H−12(Γ) → W1(R3) (ii) γS : H−12(Γ) → H12(Γ) (iii) ∂nS, ∂neS : H−12(Γ) → H−12(Γ)
(iv) D : H12(Γ) → H1(Ω) × W1(Ωe) (v) γD, γeD : H12(Γ) → H12(Γ) (vi) ∂nD : H12(Γ) → H−12(Γ) and the jump conditions
(vii) Jγ S λK = 0 ∀λ ∈ H−12(Γ) (viii) Jγ DϕK = −ϕ ∀ϕ ∈ H12(Γ)
(ix) J∂nSλK = λ ∀λ ∈ H−12(Γ) (x) J∂nDϕK = 0 ∀ϕ ∈ H12(Γ)
10
Boundary integral operators
Definition 7.
We introduce the boundary integral operators
I V := γS : H−12(Γ) → H12(Γ) (Single-layer operator)
I W := −∂nD : H12(Γ) → H−12(Γ) (Hypersingular operator)
I K := {γD} : H12(Γ) → H12(Γ) (Double-layer operator)
I Kt:= {∂nS} : H−12(Γ) → H−12(Γ) (Dual of the double-layer operator)
With the above definitions, there holds, with I denoting the identity operator, (i) γS = V
(iv) ∂nS =12I + Kt, ∂neS = −12I + Kt
(ii) ∂nD = −W
(iii) γD = −12I + K, γeD = 12I + K
11
Ellipticity properties
Theorem 8.
The symmetric bilinear form aV(λ, µ) := hµ, V λiΓis bounded on H−12(Γ) and
∃α > 0; aV(λ, λ) ≥ α kλk2−1/2,Γ
for all λ ∈ H−12(Γ) if d = 3 and for all λ ∈ H−
1 2
0 (Γ) if d = 2.
Theorem 9.
The symmetric bilinear form aW(ψ, φ) := hW ψ, φiΓis bounded on H12(Γ) and
∃β > 0; aW(φ, φ) ≥ β kφk2
H12(Γ)/P0
∀ϕ ∈ H12(Γ)/P0.
Theorem 10.
aW(ψ, φ) = aV(curlΓψ, curlΓφ) ∀ψ, ϕ ∈ H12(Γ)
12
Kernels
Proposition 11.
ker(1
2I + K) = P0
Proposition 12.
ker(1
2I + Kt) = span{∂neu0} where u0 is the solution of
−∆u0 = 0 in Ωe (2)
γeu = 1 (3)
13
The Calder´ on Projectors
We have shown that the interior and exterior Cauchy data of a harmonic function u in Ω ∪ Ωe are related by
γu
∂nu
= 1
2I + H γu
∂nu
and γeu
∂neu
= 1
2I − H γeu
∂enu
where
H =−K V
W Kt
It holds that
1 2I + H
2
= 1 2I + H
, 1
2I − H
2
= 1 2I − H
Notice that we have ∂nu = DtN γu with DtN := V−1(1
2I + K) : H12(Γ) → H−12(Γ)
is the Dirichlet-to-Neumann operator also known as the Steklov-Poincar´e map.
Using the second equation of the Calder´on projector we deduce that DtN = W + (1
2I + Kt)V−1(1 2I + K)
which proves that DtN is H1/2(Γ)/P0-elliptic 14
Direct and indirect boundary integral formulations for the Laplace problem
15
The interior Dirichlet problem
Given g ∈ H12(Γ), the
−∆u = 0 in Ω
u = g on Γ (4)
admits a unique weak solution u ∈ H1(Ω).
I The indirect method: (d = 3) The problem
Find λ ∈ H−12(Γ); aV(λ, µ) = hµ, giΓ, ∀µ ∈ H−12(Γ)
has a unique solution and u = Sλ|Ω∈ H1(Ω) is the unique (weak) solution of (4).
I The direct method: (d = 3) The problem
Find λ ∈ H−12(Γ); aV(λ, µ) = hµ, (1
2I + K)giΓ, ∀µ ∈ H−12(Γ) is well-posed and u = (Sλ − Dg)|Ω∈ H1(Ω) is the unique (weak) solution of (4).
Moreover, λ = ∂nu.
16
The interior Neumann problem
Given g ∈ H−012(Γ), the problem
−∆u = 0 in Ω
∂nu = g on Γ (5)
admits a unique weak solution u ∈ H1(Ω)/P0.
I The indirect method The problem
Find ψ ∈ H12(Γ)/P0; aW(ψ, φ) = hg, φiΓ, ∀µ ∈ H12(Γ)/P0
has a unique solution and u = −D(ψ)|Ωis the unique weak solution of (5) in H1(Ω)/P0.
I The direct method The problem
Find ψ ∈ H12(Γ)/P0; aW(ψ, φ) = h(1
2I − Kt)g, φiΓ, ∀µ ∈ H12(Γ)/P0
is well-posed. Notice that the right hand-side is compatible since K1 = −12 yields h(1
2I − Kt)g, 1iΓ= hg, (1
2I − K)1iΓ= hg, 1iΓ= 0.
The function given in Ω by u = Sg − Dψ is the unique weak solution of (5) in
H1(Ω)/P0. 17
The exterior Dirichlet problem
Given g ∈ H12(Γ), the problem
−∆u = 0 in Ω
u = g on Γ (6)
admits a unique weak solution u ∈ W1(Ωe).
I The indirect method: (d = 3) The problem
Find λ ∈ H−12(Γ); aV(λ, µ) = hµ, giΓ, ∀µ ∈ H−12(Γ)
has a unique solution and u = Sλ|Ωe ∈ W1(Ωe) is the unique weak solution of (6).
I The direct method: (d = 3) The problem
Find λ ∈ H−12(Γ); aV(λ, µ) = −hµ, (1
2I − K)giΓ, ∀µ ∈ H−12(Γ) is well-posed and u = (−Sλ + Dg)|Ωe∈ W1(Ωe) is the unique weak solution of (6).
Moreover, λ = ∂neu.
18
The exterior Neumann problem
Given g ∈ H−12(Γ), the problem
−∆u = 0 in Ω
∂neu = g on Γ (7)
admits a unique weak solution u ∈ W1(Ωe).
I The indirect method:
The problem
Find ψ ∈ H12(Γ)/R; aW(ψ, φ) = h¯g, φiΓ, ∀µ ∈ H12(Γ)/R, with ¯g = g −hg,1i|Γ|Γ, has a unique solution and u = −Dψ|Ωe∈ W1(Ωe) is the unique weak solution of (7).
I The direct method:
The problem
Find ψ ∈ H12(Γ)/R; aW(ψ, φ) = −h(1
2I + Kt)g, φiΓ, ∀φ ∈ H12(Γ)/R is well-posed and u = (−Sg + Dψ)|Ωe∈ W1(Ωe) is the unique weak solution of (7).
Moreover, ψ = γeu.
19
The BEM for the Laplacian
20
Surface meshes
We assume that Ω is polyhedral and Lipschitz. Hence Γ = ∪Jj=1Γj, where each Γj is a plane panel. A surface mesh Thof Γ is a decomposition of Γ into finitely disjoint and non-degenerate triangles T ⊂ Γ,
Γ =∪T ∈ThT .
Let ˆT := {(ˆx1, ˆx2) ∈ R2; 0 < ˆx2< ˆx1< 1} be the reference element. Given T ∈ Th
with vertices P0, P1, P2, the affine mapping FT: ˆT → R3
ˆ
x = (ˆx1, ˆx2) → FT(ˆx) := P0+ ˆx1(P1− P0) + ˆx2(P2− P0) satisfies FT( ˆT ) = T . We denote
BT= DFT= (P1− P0, P2− P0) ∈ R3×2 Notice that, according to our hypothesis on Th, if GT := BtTBT then
detGT > 0 ∀T ∈ Th
21
Shape regular families of triangulations
The diameter of T ∈ This given by
hT= max
x,y∈T|x − y|
and the inner width ρT is the in-circle diameter of T . We set h = max
T ∈ThhT. Definition 13.
I {Th}his shape regular if there exists c1> 0 independent of h such that
T ∈Tmaxh hT
ρT
≤ c1 ∀h.
I {Th}his quasi-uniform if there exists c2> 0 independent of h such that 1 < h
minT ∈ThhT
≤ c2 ∀h.
I The mesh This said to be conforming if T 6= T0∈ Th ⇒ T ∩ T0=
∅
common side common vertex
22
Galerkin approximation of the exterior Dirichlet Laplace problem
We recall that the indirect boundary integral approach to solve the exterior Dirichlet Laplace problem (6) is
Find λ ∈ H−12(Γ); aV(λ, µ) = hµ, giΓ, ∀µ ∈ H−12(Γ) (8) Given a triangulation Th(not necessarily conforming) we consider the discrete problem
Find λh∈ Yh; aV(λh, µ) = hµ, giΓ, ∀µ ∈ Yh (9) where Yh⊂ H−12(Γ) is given by
Yh:= {µ ∈ L2(Γ); µ|T ∈ P0, ∀T ∈ Th}
Theorem 14 (C´ea).
There exists a constant C > 0 independent of h such that kλ − λhk−1/2,Γ≤ C inf
µh∈Yhkλ − µhk−1/2,Γ
23
Galerkin approximation of the exterior Neumann Laplace problem
The indirect boundary integral approach to solve the exterior Neumann Laplace problem (7) is
Find ψ ∈ H12(Γ); aW(ψ, φ) + h1, ψiΓh1, φiΓ= hφ, ¯giΓ, ∀φ ∈ H12(Γ) (10) Given a conforming triangulation Th, we consider the discrete problem
Find ψh∈ Xh; aW(ψh, φ) + h1, ψhiΓh1, φiΓ= hφ, ¯giΓ, ∀φ ∈ Xh (11) where Xh⊂ H12(Γ) is given by
Xh:= {φ ∈ C0(Γ); φ|T ∈ P1, ∀T ∈ Th}
Theorem 15 (C´ea).
There exists a constant C > 0 independent of h such that kψ − ψhk1/2,Γ≤ C inf
φh∈Xhkψ − φhk1/2,Γ
24
Approximation properties
25
Interpolation of operators
We assume here that Γj is one of the plane panels constituting Γ.
Theorem 16.
Assume that T ∈ L(L2(Γj), L2(Γj)) and T ∈ L(L2(Γj), H1(Γj)). Then T ∈ L(L2(Γj), H1/2(Γj)) and
kT kL(L2(Γj),H1/2(Γj))≤q
kT kL(L2(Γj),L2(Γj))
q
kT kL(L2(Γj),H1(Γj))
Theorem 17.
Assume that T ∈ L(L2(Γj), H2(Γj) and T ∈ L(H1(Γj), H2(Γj)). Then T ∈ L(H12(Γj), H2(Γj)) and
kT kL(H1/2(Γj),H2(Γj))≤q
kT kL(L2(Γj),H2(Γj))
q
kT kL(H1(Γj),H2(Γj))
26
Change of variable formulas
Proposition 18.
kGTk ≤ 2h2T kG−1T k ≤ 2 π2(hT
ρT
)4h−2T
Proposition 19.
Given u : T → C, we consider ˆu(ˆx) := u(FT(ˆx)). Then, v ∈ Hm(T ) ⇔ ˆv ∈ Hm( ˆT ) and, for all 0 ≤ ` ≤ m,
|v|`,T ≤ C1h1−`T |ˆv|`, ˆT
|ˆv|`, ˆT ≤ C2h`−1T |v|`,T
with C1 and C2 independent of T .
27
Interpolation error estimates in Y
hWe introduce for s = 1/2, 1, 2 the Sobolev spaces
Hspw(Γ) := {µ ∈ L2(Γ); v|Γj ∈ Hs(Γj), j = 1, · · · , J } Notice that H1/2(Γ) ,→ H1/2pw(Γ).
Theorem 20.
Let Πhbe the L2(Γ)-projection onto Yh. If {Th}his shape regular then there exists a constant C > 0 independent of h such that
kµ − Πhµk0,Γ≤ C htkµkHtpw(Γ) ∀µ ∈ Htpw(Γ) (t = 1, 1/2)
Theorem 21.
Let Πhbe the L2(Γ)-projection onto Yh. If {Th}his shape regular then there exists a constant C > 0 independent of h such that
kµ − Πhµk−1
2,Γ≤ C h3/2kµkH1
pw(Γ) ∀µ ∈ H1pw(Γ)
28
Interpolation error estimates in X
hWe notice that H2pw(Γ) ∩ H1(Γ) ,→ C0(Γ). Consequently, if This conforming, we can then define the usual Lagrange interpolation operator Ih: H2pw(Γ) ∩ H1(Γ) → Xh. Theorem 22.
If {Th}his shape regular then there exists a constant C > 0 independent of h such that kφ − Ihφk1/2,Γ≤ C h3/2kφkH2
pw(Γ) ∀φ ∈ H2pw(Γ) ∩ H1(Γ)
29
Inverse estimates
Proposition 23.
Assume that {Th}his shape regular and quasi-uniform. Then, there exists a constant C > 0 independent of h such that,
kµhk1
2,Γ≤ Ch−12kµhk0,Γ ∀µh∈ Yh
Proposition 24.
Assume that {Th}his shape regular and quasi-uniform. Then, there exists a constant C > 0 independent of h such that,
kµhk0,Γ≤ Ch−12kµhk−1
2,Γ ∀µh∈ Yh
30
Error estimates for the exterior Dirichlet Laplace problems
Theorem 25.Assume that the solution λ ∈ H−12(Γ) of problem (6) belongs to H1pw(Γ). Then, kλ − λhk−1
2,Γ≤ C1h3/2kλkH1pw(Γ)
with C1> 0 independent of h. Moreover, if {Th}his quasi-uniform, kλ − λhk0,Γ≤ C2h kλkH1
pw(Γ)
Let us consider the function
˜ u =
Z
Γ
E(x, y)λh(y) x ∈ Ωe.
As y 7→ E(x, y) ∈ H12(Γ), if λ ∈ H1pw(Γ) we have that
|u(x) − ˜u(x)| ≤ kE(x, ·)k1
2,Γkλ − λhk−1
2,Γ≤ C h3/2kE(x, ·)k1 2,ΓkλkH1
pw(Γ)
for all x ∈ Ωe.
The last estimate can be improved to
|u(x) − ˜u(x)| = O(h3)
31
Error estimates for the exterior Neumann Laplace/Helmholtz problems
Theorem 26.
Assume that the solution ψ ∈ H12(Γ) of problem (7) belongs to H2pw(Γ) ∩ H1(Γ). Then, kψ − ψhk1
2,Γ≤ Ch3/2kψkH2pw(Γ)
with C > 0 independent of h.
Here again, if we consider
˜
u(x) = −Dψh(x) x ∈ Ωe then
|u(x) − ˜u(x)| ≤ kD1k−1
2,Γkψ − ψhk1
2,Γ≤ C h3/2kD1k−1
2,ΓkψkH2 pw(Γ)
for all x ∈ Ωe.
32