HAL Id: hal-01957588
https://hal.archives-ouvertes.fr/hal-01957588
Submitted on 17 Dec 2018
HAL is a multi-disciplinary open access archive for the deposit and dissemination of sci- entific research documents, whether they are pub- lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés.
Adapted Numerical Methods for the Poisson Equation with L 2 Boundary Data in NonConvex Domains
Thomas Apel, Serge Nicaise, Johannes Pfefferer
To cite this version:
Thomas Apel, Serge Nicaise, Johannes Pfefferer. Adapted Numerical Methods for the Poisson Equa-
tion with L
2Boundary Data in NonConvex Domains. SIAM Journal on Numerical Analysis, Society
for Industrial and Applied Mathematics, 2017, 55 (4), pp.1937-1957. �10.1137/16m1062077�. �hal-
01957588�
ADAPTED NUMERICAL METHODS FOR THE POISSON EQUATION WITH L
2BOUNDARY DATA IN NONCONVEX
DOMAINS
∗THOMAS APEL†, SERGE NICAISE‡,AND JOHANNES PFEFFERER§
Abstract. The very weak solution of the Poisson equation withL2boundary data is defined by the method of transposition. The finite element solution with regularized boundary data converges in theL2(Ω)-norm with order 1/2 in convex domains but has a reduced convergence order in nonconvex domains although the solution remains to be contained inH1/2(Ω). The reason is a singularity in the dual problem. In this paper we propose and analyze, as a remedy, both a standard finite element method with mesh grading and a dual variant of the singular complement method. The error order 1/2 is retained in both cases, also with nonconvex domains. Numerical experiments confirm the theoretical results.
Key words. elliptic boundary value problem, very weak formulation, finite element method, mesh grading, singular complement method, discretization error estimate
AMS subject classifications. 65N30, 65N15 DOI. 10.1137/16M1062077
1. Introduction. In this paper we consider the boundary value problem
−∆y = f in Ω, y = u on Γ = ∂Ω,
(1)
with right-hand side f ∈ H
−1(Ω) and boundary data u ∈ L
2(Γ). We assume Ω ⊂ R
2to be a bounded polygonal domain with boundary Γ. Such problems arise in optimal control when the Dirichlet boundary control is considered in L
2(Γ); see for example [22, 24, 28].
For boundary data u ∈ L
2(Γ) we cannot expect a weak solution y ∈ H
1(Ω).
Therefore we define a very weak solution by the method of transposition which goes back at least to Lions and Magenes [27, Chapter 2, section 6]: Find
y ∈ L
2(Ω) : (y, ∆v)
Ω= (u, ∂
nv)
Γ− (f, v)
Ω∀v ∈ V (2)
with (w, v)
G:= R
G
wv denoting the L
2(G) scalar product or an appropriate duality product. In our previous paper [4] we showed that the appropriate space V for the test functions is
V := H
∆1(Ω) ∩ H
01(Ω) with H
∆1(Ω) := {v ∈ H
1(Ω) : ∆v ∈ L
2(Ω)}.
(3)
Note that from Theorems 4.4.3.7 and 1.4.5.3 of [25] the embedding V , → H
3/2+ε(Ω) for 0 < ε < ε
0follows with ε
0depending on the maximal interior angle of the domain Ω.
∗Received by the editors February 17, 2016; accepted for publication (in revised form) March 9, 2017; published electronically August 17, 2017. This paper is an extension of our previous technical report, arXiv:/505.00414 [math.NA], 2015 [2].
http://www.siam.org/journals/sinum/55-4/M106207.html
Funding:The work of the authors was partially supported by Deutsche Forschungsgemeinschaft, IGDK 1754.
†Institut f¨ur Mathematik und Bauinformatik, Universit¨at der Bundeswehr M¨unchen, D-85579 Neubiberg, Germany (thomas.apel@unibw.de).
‡LAMAV, Institut des Sciences et Techniques de Valenciennes, Universit´e de Valenciennes et du Hainaut Cambr´esis, B.P. 311, 59313 Valenciennes Cedex, France (snicaise@univ-valenciennes.fr).
§Lehrstuhl f¨ur Optimalsteuerung, Technische Universit¨at M¨unchen, D-85748 Garching bei M¨unchen, Germany (pfefferer@ma.tum.de).
1937
In particular this ensures ∂
nv ∈ L
2(Γ) for v ∈ V such that the formulation (2) is well defined. We proved the existence of a unique solution y ∈ L
2(Ω) for u ∈ L
2(Γ) and f ∈ H
−1(Ω), and that the solution is even in H
1/2(Ω). The method of transposition is used in different variants also in [24, 9, 15, 14, 22, 28].
Consider now the discretization of the boundary value problem. Let T
hbe a quasi-uniform family of conforming finite element meshes, and introduce the finite element spaces
Y
h:= {v
h∈ H
1(Ω) : v
h|
T∈ P
1∀T ∈ T
h}, Y
0h:= Y
h∩ H
01(Ω), Y
h∂:= Y
h|
∂Ω. Since the boundary datum u is in general not contained in Y
h∂we have to approximate it by u
h∈ Y
h∂, e. g., by using L
2(Γ)-projection or quasi-interpolation. In this way, the boundary datum is even regularized since u
h∈ H
1/2(Γ). Hence we can consider a regularized (weak) solution in Y
∗h:= {v ∈ H
1(Ω) : v|
Γ= u
h},
y
h∈ Y
∗h: (∇y
h, ∇v)
Ω= (f, v)
Ω∀v ∈ H
01(Ω).
(4)
The finite element solution y
his now searched for in Y
∗h:= Y
∗h∩ Y
h: Find y
h∈ Y
∗h: (∇y
h, ∇v
h)
Ω= (f, v
h)
Ω∀v
h∈ Y
0h.
(5)
The same discretization was derived previously by Berggren [9] from a different point of view. In [4] we showed that the discretization error estimate
ky − y
hk
L2(Ω)≤ Ch
sh
1/2kf k
H−1(Ω)+ kuk
L2(Γ)holds for s = 1/2 if the domain is convex; this is a slight improvement of the result of Berggren, and the convex case is completely treated. In the case of nonconvex domains this convergence order is reduced although the very weak solution y is also in H
1/2(Ω); the finite element method does not lead to the best approximation in L
2(Ω). In order to describe the result we assume for simplicity that Ω has only one corner with interior angle ω ∈ (π, 2π). We proved in [4] the convergence order s = λ− 1/2 − ε, where λ := π/ω and ε > 0 arbitrarily small, and showed by numerical experiments that the order of almost λ − 1/2 is sharp. Note that s → 0 for ω → 2π.
This is the state of the art for this kind of problem, and our aim is to devise methods to retain the convergence order s = 1/2 in the nonconvex case.
In order to explain the reduction in the convergence order and our first remedy, let us first mention that we have to modify the Aubin–Nitsche method to derive L
2(Ω)- error estimates. The first reason is that our problem has no weak solution, only the dual problem,
v
z∈ V : (ϕ, ∆v
z)
Ω= (z, ϕ)
Ω∀ϕ ∈ L
2(Ω), (6)
has. The second reason is that the solution y has inhomogeneous Dirichlet data such
that an estimate of the L
2(Γ)-interpolation error of ∂
nv
zis needed. The H
1(Ω)-error
of a standard finite element method is of order one in convex domains but reduces
to s = λ − ε in the case of nonconvex domains; moreover, the order of the L
2(Γ)-
interpolation error of ∂
nv
zreduces from 1/2 to λ − 1/2 − ε. It has been known for a
long time that locally refined (graded) meshes and augmenting of the finite element
space by singular functions are appropriate to retain the optimal convergence order
for such problems; see, e. g., [8, 11, 17, 29, 31, 33]. We use these strategies in this
paper.
The novelty is that the adapted methods act now implicitly and occur essentially in the analysis for the dual problem. This sounds particularly simple in the case of mesh grading. However, the convergence proof in [4] contains not only interpolation error estimates for the dual solution and its normal derivative (which are improved now) but also the application of an inverse inequality which gives a too pessimistic result if used unchanged in the case of graded meshes. We prove in section 2 a sharp result by using a weighted norm in intermediate steps. Note we suggest a strong mesh grading with grading parameter µ → 0 (the parameter is explained in section 2) for ω → 2π because of the interpolation error estimate of ∂
nv
z; the numerical tests show that weaker grading is not sufficient.
The basic idea of the dual singular function method (see [11]), or the singular complement method (see [17]), is to augment the approximation space for the solution by one (or more, if necessary) singular function of type r
λsin(λθ) and the space of test functions by a dual function of type r
−λsin(λθ), where r, θ are polar coordinates at the concave corner. In this paper we do it the other way round and compute an approximate solution
z
h∈ Y
h⊕ Span{r
−λsin(λθ)}
such that the error estimate
ky − z
hk
L2(Ω)≤ Ch
1/2h
1/2kf k
H−1(Ω)+ kuk
L2(Γ)can be shown. Note that the original singular complement method augments the stan- dard finite element space with a function which is part of the representation of the solution. Here, we complement the finite element space with r
−λsin(λθ) 6∈ H
1/2(Ω), and, although y ∈ H
1/2(Ω), this has an effect on the approximation order in the L
2(Ω)-norm. This makes the method different from the original singular comple- ment method, [17], and we call it the dual singular complement method. Numerical experiments in section 4 confirm the theoretical results.
Finally in this introduction, we would like to note that higher order finite elements are not useful here since the solution has low regularity. The extension of our methods to three-dimensional domains should be possible in the case of mesh grading (at considerable technical expenses in the analysis) but is not straightforward in the case of the dual singular complement method since the space V \ H
2(Ω) is in general not finite dimensional; see [18] for the Fourier singular complement method to treat special domains. Curved boundaries could be treated at the price of using nonaffine finite elements; see, e. g., [10, 12, 22].
2. Graded meshes. Recall from the introduction that Ω ⊂ R
2is a bounded polygonal domain with boundary Γ, and we consider here the case that Ω has exactly one corner (called singular corner ) with interior angle ω ∈ (π, 2π). The convex case was already treated in [4] and the case of more than one nonconvex corner can be treated similarly since corner singularities are local phenomena.
Without loss of generality we can assume that the singular corner is located at the origin of the coordinate system, and that one boundary edge is contained in the positive x
1-axis. We recall from [25, Theorem 4.4.3.7] or [26, sections 1.5, 2.3, and 2.4] that the weak solution of the boundary value problem (1) with f ∈ L
2(Ω) and u = 0 is not contained in H
2(Ω) but in
H
∆1(Ω) ∩ H
01(Ω) =
H
2(Ω) ∩ H
01(Ω)
⊕ Span{ξ(r) r
λsin(λθ)},
(7)
ξ being a cutoff function, while r and θ denote polar coordinates at the singular corner.
Let the finite element mesh T
h= {T } be graded with the mesh grading parameter µ ∈ (0, 1], i. e., the element size h
T= diam T and the distance r
Tof the element T to the singular corner are related by
c
1h
1/µ≤ h
T≤ c
2h
1/µfor r
T= 0, c
1hr
T1−µ≤ h
T≤ c
2hr
T1−µfor r
T> 0.
(8)
This type of graded mesh was investigated before in [8, 29, 31, 32]; see also the overview and background information in [5, section 2.3] and [1, section 7]. Define the finite element spaces
Y
h= {v
h∈ H
1(Ω) : v
h|
T∈ P
1∀T ∈ T
h}, Y
0h= Y
h∩ H
01(Ω), Y
h∂= Y
h|
∂Ω, (9)
and let the regularized boundary datum u
h∈ Y
h∂⊂ H
1/2(Γ) be defined by the L
2(Γ)- projection Π
hu or by the Carstensen interpolant C
hu; see [13]. To define the latter let N
Γbe the set of nodes of the triangulation on the boundary, and set
C
hu = X
x∈NΓ
π
x(u)λ
xwith π
x(u) = R
ωx
uλ
xR
ωx
λ
x= (u, λ
x)
ωx(1, λ
x)
ωx,
where λ
xis the standard hat function related to x and ω
x= supp λ
x⊂ Γ. As already outlined in [4], the advantages of the interpolant in comparison to the L
2-projection are its local definition and the property
u ∈ [a, b] ⇒ C
hu ∈ [a, b];
see [21]; a disadvantage may be that C
hu
h6= u
hfor piecewise linear u
h. With these regularized boundary data we then define the regularized weak solution y
h∈ Y
∗h:=
{v ∈ H
1(Ω) : v|
Γ= u
h} by (4).
Lemma 2.1. If the mesh is graded with parameter µ < 2λ − 1 the effect of the regularization of the boundary datum can be estimated by
ky − y
hk
L2(Ω)≤ ch
1/2kuk
L2(Γ). Proof. In view of
ky − y
hk
L2(Ω)= sup
z∈L2(Ω),z6=0
(y − y
h, z)
Ωkzk
L2(Ω)(10)
we have to estimate (y − y
h, z)
Ω. To this end, let z ∈ L
2(Ω) be an arbitrary function and let v
z∈ V be defined by (6). Since the weak regularized solution y
h∈ Y
∗h:=
{v ∈ H
1(Ω) : v|
Γ= u
h} defined by (4) is also a very weak solution, (y
h, ∆v)
Ω= (u
h, ∂
nv)
Γ− (f, v)
Ω∀v ∈ V, (11)
we get with (2) and (6)
(y − y
h, z)
Ω= (u − u
h, ∂
nv
z)
Γ.
(12)
If u
his the L
2(Γ)-projection Π
hu of u we can continue with
(u − u
h, ∂
nv
z)
Γ= (u − u
h, ∂
nv
z− Π
h(∂
nv
z))
Γ= (u, ∂
nv
z− Π
h(∂
nv
z))
Γ≤ kuk
L2(Γ)k∂
nv
z− Π
h(∂
nv
z)k
L2(Γ)≤ kuk
L2(Γ)k∂
nv
z− C
h(∂
nv
z)k
L2(Γ)= kuk
L2(Γ)X
x∈NΓ
∂
nv
z− π
x(∂
nv
z) λ
xL2(Γ)
≤ ckuk
L2(Γ)
X
x∈NΓ
k∂
nv
z− π
x(∂
nv
z)k
2L2(ωx)
1/2
.
If u
his the Carstensen interpolant of u, there holds (u − C
hu, ∂
nv
z)
Γ=
X
x∈NΓ
(u − π
xu)λ
x, ∂
nv
z
Γ
= X
x∈NΓ
(u − π
x(u), (∂
nv
z)λ
x)
Γ= X
x∈NΓ
(u − π
x(u), (∂
nv
z− π
x(∂
nv
z))λ
x)
Γ≤ X
x∈NΓ
kuk
L2(ωx)k∂
nv
z− π
x(∂
nv
z)k
L2(ωx)≤ ckuk
L2(Γ)
X
x∈NΓ
k∂
nv
z− π
x(∂
nv
z)k
2L2(ωx)
1/2
,
i. e., in both cases we have to estimate P
x∈NΓ
k∂
nv
z− π
x(∂
nv
z)k
2L2(ωx). To this end we notice that
v
z∈ V =
H
2(Ω) ∩ H
01(Ω)
⊕ Span{ξ(r) r
λsin(λθ)}
and, consequently,
∂
nv
z∈ V
Γ=
N
Y
j=1
H
001/2(Γ
j)
⊕ Span{ξ(r) r
λ−1};
see [4, Remark 2.2] or [25, Theorem 1.5.2.8]. This means that we can split ∂
nv
z= αξ(r) r
λ−1+ P
Nj=1
w
jwith w
j∈ H
001/2(Γ
j) and
|α| +
N
X
j=1
kw
jk
H1/200 (Γj)
=: k∂
nv
zk
VΓ≤ ckv
zk
V:= k∆v
zk
L2(Ω)= kzk
L2(Ω). In the remaining part of the proof we show for j = 1, . . . , N,
X
x∈NΓ
kw
j− π
xw
jk
2L2(ωx)
1/2
≤ ch
1/2kw
jk
H1/2 00 (Γj), (13)
X
x∈NΓ
kξ(r) r
λ−1− π
x(ξ(r) r
λ−1)k
2L2(ωx)
1/2
≤ ch
1/2,
(14)
to conclude P
x∈NΓ
k∂
nv
z− π
x(∂
nv
z)k
2L2(ωx) 1/2≤ ch
1/2kzk
L2(Ω)and, hence, (u − u
h, ∂
nv
z)
Γ≤ ch
1/2kuk
L2(Γ)kzk
L2(Ω)which, together with (10) and (12), finishes the proof.
We extend w
jto the whole boundary Γ by zero on Γ \ Γ
jand start with the estimate
kw
j− π
sw
jk
L2(ωx)≤ ch
sxks
jk
Hs(ωx), s = 0, 1, x ∈ N
Γ. (15)
This estimate follows for s = 0 from the definition of π
x. For s = 1 it follows from a Bramble–Hilbert-type argument if x is not a corner of Ω. In the case of a corner point x we use instead the zero boundary condition of w
jon one end of ω
x. Adding these estimates and using that H
001/2(Γ
j) is an interpolation space of L
2(Γ
j) and H
01(Γ
j) we obtain (13). Note that the local element size h
xis bounded by h from above.
Denote by N
Γ,reg⊂ N
Γthe set of nodes where ω
xdoes not contain the singular corner. Let r
xbe the distance of x ∈ N
Γ,regto the set of corners of Ω, and note that the local mesh size satisfies both h
x≤ chr
1−µxand h
x≤ cr
x. One can estimate by using (15) with s = 1,
X
x∈NΓ,reg
kξ(r) r
λ−1− π
x(ξ(r) r
λ−1)k
2L2(ωx)≤ c X
x∈NΓ,reg
h
2xkr
λ−2k
2L2(ωx)≤ ch X
x∈NΓ,reg
r
1−µxr
xkr
λ−2k
2L2(ωx)≤ ch
Z
diamΩ 0r
2−µ+2(λ−2)dr = ch
for µ < 2λ−1. For the three nodes x ∈ N
Γ\N
Γ,regwe cannot use the H
1(ω
x)-regularity of r
λ−1but, by using the stability of π
x, the properties of ξ(·), and h
x∼ h
1/µthere holds
kξ(r) r
λ−1− π
x(ξ(r) r
λ−1)k
L2(ωx)≤ ckr
λ−1k
L2(ωx)∼ h
λ−1/2x∼ h
(λ−1/2)/µ≤ ch
1/2for µ < 2λ − 1. Note that we computed the norm in the middle step. This finishes the proof.
We consider now a lifting ˜ B
hu
h∈ Y
∗h:= Y
∗h∩ Y
hdefined by the nodal values as follows:
( ˜ B
hu
h)(x) =
( u
h(x) for all nodes x ∈ Γ, 0 for all nodes x ∈ Ω.
(16)
The function y
hand its finite element approximation y
h∈ Y
∗hare now defined by y
h= y
f+ ˜ B
hu
h+ ˜ y
h0as well as y
h= y
f h+ ˜ B
hu
h+ ˜ y
0h,
(17)
where y
f, y ˜
0h∈ H
01(Ω) and y
f h, y ˜
0h∈ Y
0hsatisfy (∇y
f, ∇v)
Ω= (f, v)
Ω∀v ∈ H
01(Ω), (18)
(∇y
f h, ∇v
h)
Ω= (f, v
h)
Ω∀v
h∈ Y
0h, (19)
(∇˜ y
h0, ∇v)
Ω= −(∇( ˜ B
hu
h), ∇v)
Ω∀v ∈ H
01(Ω), (20)
(∇ y ˜
0h, ∇v
h)
Ω= −(∇( ˜ B
hu
h), ∇v
h)
Ω∀v
h∈ Y
0h. (21)
In order to estimate ky
h− y
hk
L2(Ω)we estimate ky
f− y
f hk
L2(Ω)and k˜ y
0h− y ˜
0hk
L2(Ω).
Lemma 2.2. If the mesh is graded with parameter µ < λ the error in approximat- ing y
fsatisfies
ky
f− y
f hk
L2(Ω)≤ chkf k
H−1(Ω).
Note that the condition µ < λ is weaker than the condition µ < 2λ−1 from Lemma 2.1 since λ < 1.
Proof. As in the proof of Lemma 2.1, let z ∈ L
2(Ω) be an arbitrary function, let v
z∈ V be defined via (6), and let v
zh∈ Y
0hbe the Ritz projection of v
z. By the definitions (18) and (19) and using the Galerkin orthogonality we get
(y
f− y
f h, z)
Ω= (∇(y
f− y
f h), ∇v
z)
Ω= (∇(y
f− y
f h), ∇(v
z− v
zh))
Ω= (∇y
f, ∇(v
z− v
zh))
Ω≤ k∇y
fk
L2(Ω)k∇(v
z− v
zh)k
L2(Ω). By using standard a priori estimates (see, e.g., [7, Theorem 3.2]), we obtain with grading µ < λ the bounds k∇y
fk
L2(Ω)≤ kf k
H−1(Ω), k∇(v
z−v
zh)k
L2(Ω)≤ chkzk
L2(Ω), and, hence, with
ky
f− y
f hk
L2(Ω)= sup
z∈L2(Ω),z6=0
(y
f− y
f h, z)
Ωkzk
L2(Ω),
the assertion of the lemma.
In the proof of Lemma 2.4 we will employ a regularity result which is proved in [4, section II.C]. Reducing notation for the price of a slightly weaker statement we have the folllowing lemma.
Lemma 2.3. If ω > π then the very weak solution y from (2) satisfies kr
−βyk
L2(Ω)≤ c
kf k
H−1(Ω)+ kuk
L2(Γ)for all β ∈ 1 − λ, 1
2 i .
Proof. The statement is proved in [4, Lemma 2.8]. Concerning the assumptions on the regularity of the data, note that f and u are from bigger spaces there if β ≤
12; see [4, Remark 2.7]. Concerning the definition of the solution y in [4, (2.15)] note that the test space there contains V , which is seen by using the splitting (7), and since the solutions of both formulations are unique they must be equal.
In order to estimate k˜ y
0h− y ˜
0hk
L2(Ω), we divide the domain Ω into subsets Ω
J, i.e.,
Ω =
I
[
J=0
Ω
J,
where Ω
J:= {x ∈ Ω : d
J+1≤ |x| ≤ d
J} for J = 1, . . . , I − 1, Ω
I:= {x ∈ Ω : |x| ≤ d
I}, and Ω
0:= Ω\ S
IJ=1
Ω
J. The radii d
Jare set to 2
−Jand the index I is chosen such that
d
I= 2
−I= c
Ih
1/µ(22)
with a constant c
I> 1 exactly specified later on. In addition we define the extended domains Ω
0Jand Ω
00Jby
Ω
0J:= Ω
J−1∪ Ω
J∪ Ω
J+1and Ω
00J:= Ω
0J−1∪ Ω
0J∪ Ω
0J+1,
respectively, with the obvious modifications for J = 0, 1 and J = I − 1, I .
Lemma 2.4. With σ := r + d
Ithere holds the estimate
kσ
(1−µ)/2∇˜ y
0hk
L2(Ω)+ kσ
(1−µ)/2∇( ˜ B
hu
h)k
L2(Ω)≤ ch
−1/2kuk
L2(Γ). Proof. We start by rearranging terms, i.e.,
kσ
(1−µ)/2∇˜ y
h0k
2L2(Ω)= Z
Ω
σ
1−µ∇˜ y
0h· ∇˜ y
0h= Z
Ω
∇ y ˜
h0· ∇(˜ y
h0σ
1−µ) − Z
Ω
˜
y
h0∇˜ y
h0· ∇σ
1−µ. (23)
For the first term in (23) we conclude according to (20) Z
Ω
∇ y ˜
h0· ∇(˜ y
h0σ
1−µ) = − Z
Ω
∇( ˜ B
hu
h) · ∇(˜ y
0hσ
1−µ)
= − Z
Ω
σ
1−µ∇( ˜ B
hu
h) · ∇˜ y
h0− Z
Ω
˜
y
h0∇( ˜ B
hu
h) · ∇σ
1−µ≤ kσ
(1−µ)/2∇( ˜ B
hu
h)k
L2(Ω)kσ
(1−µ)/2∇˜ y
0hk
L2(Ω)+ kσ
(−1−µ)/2y ˜
0hk
L2(Ω), (24)
where we used the Cauchy–Schwarz inequality and
∇σ
1−µ= (1 − µ)σ
−µ(cos θ, sin θ)
T. (25)
Having in mind the decomposition of the domain in subdomains Ω
J, an application of the Poincar´ e inequality yields for the latter term in (24)
kσ
(−1−µ)/2y ˜
0hk
2L2(Ω)=
I
X
J=0
kσ
(−1−µ)/2y ˜
0hk
L2(ΩJ)kσ
(−1−µ)/2y ˜
0hk
L2(ΩJ)≤
I
X
J=0
d
(−1−µ)/2Jk y ˜
0hk
L2(ΩJ)kσ
(−1−µ)/2y ˜
h0k
L2(ΩJ)≤ c
I
X
J=0
d
(1−µ)/2Jk∇˜ y
0hk
L2(Ω0J)kσ
(−1−µ)/2y ˜
h0k
L2(ΩJ)≤ ckσ
(1−µ)/2∇ y ˜
h0k
L2(Ω)kσ
(−1−µ)/2y ˜
0hk
L2(Ω),
where we used d
J∼ σ for x ∈ Ω
0Jtwice and the discrete Cauchy–Schwarz inequality.
Consequently, we get from (24) Z
Ω
∇ y ˜
h0· ∇(˜ y
h0σ
1−µ) ≤ ckσ
(1−µ)/2∇( ˜ B
hu
h)k
L2(Ω)kσ
(1−µ)/2∇˜ y
0hk
L2(Ω). (26)
Similarly to the above steps, we get for the second term in (23) by means of (25)
Z
Ω
˜
y
h0∇˜ y
h0· ∇σ
1−µ≤ kσ
(1−µ)/2∇˜ y
h0k
L2(Ω)kσ
(−1−µ)/2y ˜
0hk
L2(Ω)(27)
≤ kσ
(1−µ)/2∇˜ y
0hk
L2(Ω)kσ
(−1−µ)/2(˜ y
h0+ ˜ B
hu
h)k
L2(Ω)+ kσ
(−1−µ)/2B ˜
hu
hk
L2(Ω), such that we infer from (23), (26), and (27) that
kσ
(1−µ)/2∇˜ y
h0k
L2(Ω)≤ c
kσ
(−1−µ)/2B ˜
hu
hk
L2(Ω)+ kσ
(1−µ)/2∇( ˜ B
hu
h)k
L2(Ω)+ kσ
(−1−µ)/2(˜ y
h0+ ˜ B
hu
h)k
L2(Ω).
(28)
Due to the definition of ˜ B
hand the definition of the element size h
Tin the case of graded meshes we easily obtain by means of the norm equivalence in finite dimensional spaces that
kσ
(−1−µ)/2B ˜
hu
hk
L2(Ω)+ kσ
(1−µ)/2∇( ˜ B
hu
h)k
L2(Ω)≤ ch
−1/2ku
hk
L2(Γ)≤ ch
−1/2kuk
L2(Γ), (29)
where we employed the stability of u
hin L
2(Γ) in the last step. Having in mind the definition (22) of d
Iand applying Lemma 2.3 with β =
12for the solution ˜ y
h0+ ˜ B
hu
hof the homogeneous equation with boundary datum u
hwe conclude that
kσ
(−1−µ)/2(˜ y
h0+ ˜ B
hu
h)k
L2(Ω)≤ d
−µ/2Ikσ
−1/2(˜ y
h0+ ˜ B
hu
h)k
L2(Ω)≤ ch
−1/2kr
−1/2(˜ y
h0+ ˜ B
hu
h)k
L2(Ω)≤ ch
−1/2ku
hk
L2(Γ)≤ ch
−1/2kuk
L2(Γ), (30)
where we used again the stability of u
h. The estimates (28), (29), and (30) end the proof.
Lemma 2.5. Let σ := r + d
Iand µ ∈ (0, 2λ − 1). Then there is the estimate kσ
−(1−µ)/2(˜ y
0h− y ˜
0h)k
L2(Ω)≤ ch
1/2kuk
L2(Γ).
Proof. Let v ∈ H
01(Ω) be the weak solution of
−∆v = σ
−(1−µ)(˜ y
0h− y ˜
0h) in Ω, v = 0 on Γ,
which, according to Theorem 2.15 of [20], has the regularity v ∈ V
(1−µ)/22,2(Ω) (as µ < 2λ − 1) and hence
12(1 − µ) > 1 − λ) and satisfies the a priori estimate
|v|
V2,2(1−µ)/2(Ω)
≤ ckσ
−(1−µ)(˜ y
0h− y ˜
0h)k
V0,2(1−µ)/2(Ω)
≤ ckσ
−(1−µ)/2(˜ y
0h− y ˜
0h)k
L2(Ω), (31)
where we use the weighted Sobolev space V
βk,2(Ω) := {v ∈ D
0: kvk
Vk,2β (Ω)
< ∞} with kvk
2Vβk,2(Ω)
:=
k
X
j=1
|v|
2Vj,2β−k+j(Ω)
, |v|
Vj,2β (Ω)
:= kr
β∇
jvk
L2(Ω). Then we obtain by using integration by parts and the Galerkin orthogonality
kσ
−(1−µ)/2(˜ y
0h− y ˜
0h)k
2L2(Ω)= (˜ y
h0− y ˜
0h, −∆v)
Ω= (∇(˜ y
0h− y ˜
0h), ∇(v − I
hv))
Ω≤
I
X
J=0
k∇(˜ y
h0− y ˜
0h)k
L2(ΩJ)k∇(v − I
hv)k
L2(ΩJ), (32)
where I
his the Lagrange interpolant.
By employing standard interpolation error estimates on graded meshes we obtain for any µ ∈ (0, 1]
k∇(v − I
hv)k
L2(ΩJ)≤ chd
(1−µ)/2J|v|
V2,2(1−µ)/2(Ω0J)
, (33)
where the constant c is independent of c
I; see, e.g., [6, Lemma 3.7] or [30, Lemma
3.58]. In fact, the constant is essentially the one appearing in the local, elementwise
interpolation error estimate. Note that this kind of independence will be crucial when applying a kickback argument further below.
Local finite element error estimates from [23, Theorem 3.4] yield k∇(˜ y
h0− y ˜
0h)k
L2(ΩJ)≤ c min
vh∈Y0h
k∇(˜ y
h0− v
h)k
L2(Ω0J)+ 1
d
Jk y ˜
0h− v
hk
L2(Ω0J)+ c 1
d
Jk˜ y
0h− y ˜
0hk
L2(Ω0J).
By choosing v
h≡ 0 and by applying the Poincar´ e inequality, we conclude k∇(˜ y
0h− y ˜
0h)k
L2(ΩJ)≤ c
k∇˜ y
0hk
L2(Ω0J)+ 1
d
Jk˜ y
h0− y ˜
0hk
L2(Ω0J)≤ c
k∇˜ y
0hk
L2(Ω00J)+ d
(−1−µ)/2Jkσ
−(1−µ)/2(˜ y
h0− y ˜
0h)k
L2(Ω0J), (34)
where we used d
J∼ σ for x ∈ Ω
0J. Consequently, we get from (32)–(34) kσ
−(1−µ)/2(˜ y
0h− y ˜
0h)k
2L2(Ω)≤ c
I
X
J=0
hkσ
(1−µ)/2∇˜ y
0hk
L2(Ω00J)+ hd
−µJkσ
−(1−µ)/2(˜ y
h0− y ˜
0h)k
L2(Ω0J)|v|
V2,2 (1−µ)/2(Ω0J)≤ c
hkσ
(1−µ)/2∇ y ˜
0hk
L2(Ω)+ c
−µIkσ
−(1−µ)/2(˜ y
0h− y ˜
0h)k
L2(Ω)|v|
V2,2 (1−µ)/2(Ω),
where we again employed d
J∼ σ for x ∈ Ω
00J, hd
−µJ≤ c
−µI, which holds due to the definition (22) of d
I, and the discrete Cauchy–Schwarz inequality. For µ ∈ (0, 2λ − 1) we infer by the a priori estimate (31) that
kσ
−(1−µ)/2(˜ y
h0− y ˜
0h)k
L2(Ω)≤ c
hkσ
(1−µ)/2∇ y ˜
0hk
L2(Ω)+ c
−µIkσ
−(1−µ)/2(˜ y
0h− y ˜
0h)k
L2(Ω).
By choosing c
Ilarge enough we can kick back the second term in the above inequality such that Lemma 2.4 yields the desired result.
Theorem 2.6. For µ ∈ (0, 2λ − 1) we get ky − y
hk
L2(Ω)≤ ch
1/2kuk
L2(Ω)+ h
1/2kf k
H−1(Ω). (35)
Proof. Due to the boundedness of σ
(1−µ)/2independent of h for all µ ∈ (0, 1] we obtain from Lemma 2.5
k y ˜
0h− y ˜
0hk
L2(Ω)≤ kσ
(1−µ)/2k
L∞(Ω)kσ
−(1−µ)/2(˜ y
0h− y ˜
0h)k
L2(Ω)≤ ch
1/2kuk
L2(Γ). (36)
In view of (17) we get by using the triangle inequality
ky − y
hk
L2(Ω)≤ ky − y
hk
L2(Ω)+ ky
f− y
f hk
L2(Ω)+ k y ˜
0h− y ˜
0hk
L2(Ω).
Using Lemmas 2.1 and 2.2 as well as (36) we get (35).
3. The dual singular complement method.
3.1. Analytical background and regularization. Using the notation of the previous section, we recall that the splitting (7) implies that
R := {∆v : v ∈ H
2(Ω) ∩ H
01(Ω)}
(37)
is a closed subspace of L
2(Ω). It is shown in [26, sect. 2.3] that L
2(Ω) = R ⊕
⊥Span{p
s}
(38)
with the dual singular function
p
s= r
−λsin(λθ) + ˜ p
s, (39)
where ˜ p
s∈ H
1(Ω) is chosen such that the decomposition (38) is orthogonal for the L
2(Ω) inner product. Therefore, the dual singular function p
sis a solution of
w ∈ L
2(Ω) : (∆v, w) = 0 ∀v ∈ H
2(Ω) ∩ H
01(Ω), (40)
which proves the nonuniqueness of the solution of (40).
Due to (38) we can split any L
2(Ω)-function into L
2(Ω)-orthogonal parts. To this end denote by Π
Rand Π
psthe orthogonal projections on R and on Span{p
s}, respectively, i.e., for g ∈ L
2(Ω), it is g = Π
Rg + Π
psg, where
Π
psg = α(g) p
swith α(g) = (g, p
s)
Ωkp
sk
2L2(Ω), and Π
Rg = g − Π
psg.
Since p
s∈ L
2(Ω) there exists
φ
s∈ H
∆1(Ω) ∩ H
01(Ω) : −∆φ
s= p
s; (41)
see also section 3.3 for more details on φ
s. For the moment we assume that p
sand φ
sare explicitly known; the decomposition g = Π
Rg + α(g) p
scan be computed once g is given. Computable approximations of p
sand φ
sare discussed in section 3.3.
Now we come back to problem (2) and decompose its solution y in the form y = Π
Ry + α(y) p
s.
(42)
From the decomposition (38) we see that problem (2) is equivalent to (y, p
s)
Ω= −(u, ∂
nφ
s)
Γ+ (f, φ
s)
Ω,
(y, ∆v)
Ω= (u, ∂
nv)
Γ− (f, v)
Ω∀v ∈ H
2(Ω) ∩ H
01(Ω), and with the orthogonal splitting (42) to
α(y) (p
s, p
s)
Ω= −(u, ∂
nφ
s)
Γ+ (f, φ
s)
Ω,
(Π
Ry, ∆v)
Ω= (u, ∂
nv)
Γ− (f, v)
Ω∀v ∈ H
2(Ω) ∩ H
01(Ω).
The first equation directly yields α(y), namely,
α(y) = −(u, ∂
nφ
s)
Γ+ (f, φ
s)
Ω(p
s, p
s)
Ω, (43)
hence the projection of y on p
sis known. It remains to find an approximation of Π
Ry.
At this point we recall the regularization approach from [4] which we summarized already in the introduction. Let u
h∈ H
1/2(Γ) be a regularized boundary datum (this can be any, for example, Π
hu or C
hu from section 2, but we do not assume graded meshes here) such that we can define the regularized (weak) solution in Y
∗h:=
{v ∈ H
1(Ω) : v|
Γ= u
h},
y
h∈ Y
∗h: (∇y
h, ∇v)
Ω= (f, v)
Ω∀v ∈ H
01(Ω).
(44)
In [4, Remark 2.13] we showed that the regularization error can be estimated by ky − y
hk
L2(Ω)≤ cku − u
hk
H−s(Γ),
where 0 < s < λ −
12(if Ω was convex we would get s =
12, that means the regulariza- tion error is in general bigger in the nonconvex case). With the next lemma we show that Π
R(y − y
h) is not affected by nonconvex corners.
Lemma 3.1. There holds the estimate
kΠ
R(y − y
h)k
L2(Ω)≤ Cku − u
hk
H−1/2(Γ).
Proof. Recall V = H
∆1(Ω) ∩ H
01(Ω) from (3). From (44) and the Green formula, we have for any v ∈ V
(f, v)
Ω= (∇y
h, ∇v)
Ω= −(y
h, ∆v)
Ω+ (y
h, ∂
nv)
Γ.
Note that v ∈ V is sufficient for the Green formula, and v ∈ H
2(Ω) is not required;
see [19, Lemma 3.4]. Subtracting this expression from the very weak formulation (2), we get
(y − y
h, ∆v)
Ω= (u − u
h, ∂
nv)
Γ∀v ∈ V.
Restricting this identity to v ∈ H
2(Ω) ∩ H
01(Ω), we have
(Π
R(y − y
h), ∆v)
Ω= (u − u
h, ∂
nv)
Γ∀v ∈ H
2(Ω) ∩ H
01(Ω).
(45)
Now for any z ∈ R, we let v
z∈ H
2(Ω) ∩ H
01(Ω) be the unique solution of
∆v
z= z, (46)
which satisfies
k∂
nv
zk
H1/2(Γ)≤ ckv
zk
H2(Ω)≤ ckzk
L2(Ω). (47)
Since for any g ∈ L
2(Ω) the equality
(Π
R(y − y
h), g)
Ω= (Π
R(y − y
h), Π
Rg)
Ω= (y − y
h, Π
Rg)
Ωholds, we get with (45)–(47) kΠ
R(y − y
h)k
L2(Ω)= sup
z∈R,z6=0
(y − y
h, z)
Ωkzk
L2(Ω)= sup
z∈R,z6=0
(u − u
h, ∂
nv
z)
Γkzk
L2(Ω)≤ ku − u
hk
H−1/2(Γ)sup
z∈R,z6=0
k∂
nv
zk
H1/2(Γ)kzk
L2(Ω)≤ cku − u
hk
H−1/2(Γ)which is the estimate to be proved.
3.2. Motivation for the dual singular complement method. As already discussed in the introduction, the adapted methods are motivated by the suboptimal convergence rate of the finite element solution on a family of quasi-uniform meshes. In this subsection, we recall these results and extend them by proving an estimate for the projection of the error into the space R from (37) which yields a better convergence rate. The insight into this structure of the discretization error motivates the new method which we call the dual singular complement method.
Recall from (9) the finite element spaces
Y
h= {v
h∈ H
1(Ω) : v
h|
T∈ P
1∀T ∈ T
h}, Y
0h= Y
h∩ H
01(Ω), Y
h∂= Y
h|
∂Ω, defined now on a quasi-uniform family T
hof conforming finite element meshes. Assume that the regularized boundary datum u
his contained in Y
h∂such that the estimates
ku
hk
L2(Γ)≤ ckuk
L2(Γ), (48)
ku − u
hk
H−1/2(Γ)≤ Ch
1/2kuk
L2(Γ), (49)
hold. It can be derived from [4, Lemma 2.14] that this can be accomplished by using the L
2(Γ)-projection or by quasi-interpolation: The stability (48) is explicitly stated there, and the error estimate (49) follows from the definition of the H
−1/2(Γ)-norm and the third estimate in [4, Lemma 2.14]. A consequence of Lemma 3.1 and (49) is the estimate
kΠ
R(y − y
h)k
L2(Ω)≤ Ch
1/2kuk
L2(Γ). (50)
(In the case of a convex domain the operator Π
Ris the identity, and the corresponding error estimates were already proven in [4].)
As already done in the introduction, define further the finite element solution y
h∈ Y
∗h:= Y
∗h∩ Y
hvia
y
h∈ Y
∗h: (∇y
h, ∇v
h)
Ω= (f, v
h)
Ω∀v
h∈ Y
0h. (51)
We proved in [4] that in the case of a quasi-uniform family of meshes T
hky − y
hk
L2(Ω)≤ Ch
sh
1/2kf k
H−1(Ω)+ kuk
L2(Γ)(52)
holds for s ∈ (0, λ −
12) (again s =
12for convex domains). As before, in the next lemma we show that Π
R(y − y
h) is not affected by the nonconvex corners.
Lemma 3.2. The following discretization error estimate holds:
kΠ
R(y − y
h)k
L2(Ω)≤ Ch
1/2h
1/2kf k
H−1(Ω)+ kuk
L2(Γ). Proof. By the triangle inequality we have
kΠ
R(y − y
h)k
L2(Ω)≤ kΠ
R(y − y
h)k
L2(Ω)+ kΠ
R(y
h− y
h)k
L2(Ω). (53)
The first term is estimated in (50). For the second term we first notice that y
h−y
h∈ H
01(Ω) satisfies the Galerkin orthogonality
(∇(y
h− y
h), ∇v
h)
Ω= 0 ∀v
h∈ Y
0h;
(54)
see (4) and (5). With that, we estimate kΠ
R(y
h− y
h)k
L2(Ω)by similar arguments as kΠ
R(y − y
h)k
L2(Ω)in the proof of Lemma 3.1. Recall from (46) and (47) that v
z∈ H
2(Ω) ∩ H
01(Ω) is the weak solution of ∆v
z= z ∈ R. It can be approximated by the Lagrange interpolant I
hv
zsatisfying
k∇(v
z− I
hv
z)k
L2(Ω)≤ chkv
zk
H2(Ω)≤ chkzk
L2(Ω). We get
kΠ
R(y
h− y
h)k
L2(Ω)= sup
z∈R,z6=0
(y
h− y
h, z)
Ωkzk
L2(Ω)= sup
z∈R,z6=0
(∇(y
h− y
h), ∇v
z)
Ωkzk
L2(Ω)= sup
z∈R,z6=0