Numer. Math.

Numerische
DOI 10.1007/s00211-012-0451-2 Mathematik

A Nitsche-based domain decomposition method

for hypersingular integral equations

Franz Chouly · Norbert Heuer

Received: 1 February 2011 / Revised: 20 December 2011

© Springer-Verlag 2012

Abstract We introduce and analyze a Nitsche-based domain decomposition method

for the solution of hypersingular integral equations. This method allows for discreti-
zations with non-matching grids without the necessity of a Lagrangian multiplier, as
opposed to the traditional mortar method. We prove its almost quasi-optimal conver-
gence and underline the theory by a numerical experiment.

Mathematics Subject Classification (2000) 65N38 · 65N55

1 Introduction

We propose and analyze the Nitsche method as a simple domain decomposition method
for the solution of hypersingular boundary integral equations. In this context, simple
means that (i) its implementation is not more difficult than a conforming approach and
(ii) its numerical analysis avoids mathematical difficulties inherent to usual domain
decomposition approaches. Still, a thorough analysis of our method, given in this paper,
faces the problem of non-existence of a well-posed continuous counterpart for the
discrete formulation. This is due to the low regularity of the underlying energy space.

Supported by FONDECYT project 1080044.

F. Chouly
Laboratoire de Mathématiques de Besançon, CNRS UMR 6623,
Université de Franche-Comté, 16 route de Gray, 25030 Besançon Cedex, France
e-mail: franz.chouly@univ-fcomte.fr

N. Heuer (B)
Facultad de Matemáticas, Pontificia Universidad Católica de Chile,
Avenida Vicuña Mackenna 4860, Macul, Santiago, Chile
e-mail: nheuer@mat.puc.cl

123
 F. Chouly, N. Heuer

Main attraction of the Nitsche method, apart from its relative simplicity, is that it can
maintain ellipticity and symmetry of the original problem. Domain decomposition
approaches introduce a level of parallelism to discrete problems, and non-conform-
ing approximations greatly facilitate the construction and implementation of discrete
schemes. The Nitsche approach combines both advantages. Although parallelizing
boundary element techniques is not straightforward due to the non-locality of integral
operators, non-conforming discretizations are very desirable, e.g. in the presence of
complicated surfaces.
We study the hypersingular integral equation governing the Laplacian in R3 exte-
rior to an open surface, subject to a Neumann boundary condition. In principle, our
domain decomposition approach is applicable to more realistic problems like linear
elasticity and acoustics. A generalization to the Helmholtz equation with small wave
number k (exterior to an open surface) appears feasible without substantial prob-
lems (it is a compact perturbation of the Laplace case, see [36]) since in this case
the needed integration-by-parts formula is fully understood (cf. [29]). Of course, in
the case of closed surfaces in acoustics, problems of spurious modes appear which
require different integral representations. For the Lamé case there is an equivalent var-
iational formulation which holds in the energy space, cf. [18,29], and an application of
the Nitsche decomposition is possible by extending this equivalence to an integration-
by-parts formula. Such a formula is currently unknown.
For the solution of partial differential equations, domain decomposition is a clas-
sical strategy. It is used mainly for parallelization and the solution of linear sys-
tems. A variety of techniques exist, such as Schwarz methods, see, e.g., [32,37].
Of particular interest are methods that allow for non-matching meshes at the interface
between sub-domains. They facilitate to a great extent mesh generation for complicated
geometries. The so-called mortar method has been designed for this purpose [9,10].
It consists in introducing an unknown Lagrangian multiplier on the interface and add-
ing interface conditions in a weak sense. For an analysis of the Laplacian in two and
three space dimensions see [7]. This method transforms the original problem into a sad-
dle-point structure, so that any numerical scheme requires a discrete inf-sup condition,
i.e. compatibility between approximation spaces on sub-domains and the interface.
An alternative to the mortar method is Nitsche’s method, originally published in
[1,30], and adapted in [6] to a domain decomposition framework. The interface con-
dition is again treated weakly; not as an additional equation but like a penalization
term in the (discrete) variational formulation. Other terms are added to the formula-
tion to achieve consistency and ellipticity. Moreover, symmetry can be maintained for
symmetric problems. As a result, Nitsche’s method differs from classical penalization
methods where consistency is lost [6].
In conclusion, main advantages of Nitsche’s method are that

1. no additional unknown is needed on the interface,

2. no inf-sup condition must be satisfied among discrete spaces (except for the global
one, of course), and that
3. discrete problems are elliptic and can be symmetric for symmetric problems, so
that
4. standard linear solvers can be used.

123
A Nitsche-based domain decomposition method

Nitsche’s method is closely related to the stabilized method of Barbosa and Hughes
[3,4], which also circumvents the inf-sup compatibility condition that arises when a
Dirichlet boundary condition is imposed weakly through Lagrangian multipliers. The
connection between the two methods is established in [34].
In the context of partial differential equations, the Nitsche method has been applied
successfully to a variety of problems such as linear elasticity [5,15], two-phase flows
[33], and fluid-structure interaction [2,12,19].
In the context of boundary integral equations and the use of non-matching grids
or weakly imposed boundary (or interface) conditions, we only know of the results
[16,20]. Both analyze a setting based on Lagrangian multipliers. The former reference
provides the basic results like an integration-by-parts formula for the hypersingular
integral operator, and analyzes the implementation of Dirichlet boundary conditions
in a fractional order Sobolev space of order 1/2. The latter reference proposes and
analyzes the mortar domain decomposition approach for the hypersingular integral
equation. An extreme case, the use of discontinuous basis functions for hypersingular
operators, is studied in [22].
Let us also mention that there are several papers on domain decomposition involv-
ing boundary elements. Among early references, we may quote [23] where stan-
dard boundary elements are used for problems on sub-domains of the PDE problem.
Variants of the additive Schwarz method for boundary integral equations have been
analysed in, e.g., [28,38]. Recently, a combination of domain decomposition with
boundary element techniques has led to the BETI (Boundary Element Tearing and
Interconnecting) method [24,31], based on the finite element analogue FETI. All these
papers from domain decomposition (except the earlier cited [16,20,22]) are based on
conforming approximations of functions in (fractional order) Sobolev spaces.
In this paper we propose and analyze a Nitsche domain decomposition variant for
the hypersingular integral equation governing the Laplacian. Although this approach is
simpler than mortar strategies in important aspects, as explained before, there are some
non-trivial obstacles in its numerical analysis. Energy spaces of hypersingular opera-
tors are fractional order Sobolev spaces of order 1/2. These spaces form the natural
basis for variational formulations. Now, domain decomposition introduces interfaces
where discontinuities arise. In the variational setting, these discontinuities are not well
posed, simply because no well-defined trace operator exists. Therefore, we analyze
the discrete Nitsche method without using a corresponding variational formulation.
This is very much in the spirit of Strang’s second lemma for non-conforming methods.
The difficulty of non-existence of a well-posed trace operator reappears in the analysis
of the discrete problem. We deal with this problem by making use of a whole scale
of Sobolev spaces (of higher regularity than 1/2) and by using inverse properties of
discrete functions. The result is an almost quasi-optimal error estimate for the Nitsche
method. Here, “almost” refers to perturbations which are only logarithmic in the mesh
size.
The rest of this paper is organized as follows. In Sect. 2 we define some Sobolev
spaces and our model problem. We also briefly recall the standard boundary element
approximation. In Sect. 3 we introduce a domain decomposition (for simplicity only
into two sub-domains; but this generalizes to more sub-domains in a straightforward
way), the Nitsche-based discretization, and present our main result (Theorem 3.1).

123
 F. Chouly, N. Heuer

Technical details and the proof of Theorem 3.1 are given in Sect. 4. In Sect. 5 we
present some numerical experiments that confirm the theoretical result.
Throughout the article, we will use the symbols “” and “” in the usual sense. In
short ah (v)  bh (v) when there exists a constant C > 0 independent of v, the mesh
size h and a fractional Sobolev index ε (if present), such that: ah (v) ≤ C bh (v). Also,
ah (v)  bh (v) means that ah (v)  bh (v) and ah (v)  bh (v).

2 Sobolev spaces and model problem

First let us briefly define the needed Sobolev spaces. We consider standard Sobolev
spaces where the following norms are used: For  ⊂ Rn and 0 < s < 1 we define

u2H s () := u2L 2 () + |u|2H s ()

with semi-norm
  |u(x) − u(y)|2 1/2
|u| H s () := d x d y .
|x − y|2s+n
 

For a Lipschitz domain  and 0 < s < 1, the space H̃ s () is defined as the completion
of C0∞ () under the norm

  1/2
u(x)2
u H̃ s () := |u|2H s () + dx .
(dist(x, ∂)) 2s


For s ∈ (0, 1/2), · H̃ s () and · H s () are equivalent norms whereas for s ∈ (1/2, 1)
there holds H̃ s () = H0s (), the latter space being the completion of C0∞ () with
norm in H s (). Also we note that functions from H̃ s () are continuously extendible
by zero onto a larger domain. For all these results we refer to [17,25]. For s > 0
the spaces H −s () and H̃ −s () are the dual spaces (with L 2 () as pivot space) of
H̃ s () and H s (), respectively.
Let  be a plane open surface with polygonal boundary. In the following we will
identify  with a domain in R2 , thus referring to sub-domains of  rather than sub-
surfaces. The boundary of  is denoted by ∂.
Our model problem is: For a given function f ∈ L 2 () find u ∈ H̃ 1/2 () such that

1 ∂ ∂ 1
W u(x) := − u(y) d S y = f (x), x ∈ . (2.1)
4π ∂n x ∂n y |x − y|


Here, n is a normal unit vector on , e.g. n = (0, 0, 1)T . Note that W maps H̃ 1/2 ()
continuously onto H −1/2 () (see [13]). The variational formulation of (2.1) is: Find
u ∈ H̃ 1/2 () such that

123
A Nitsche-based domain decomposition method

W u, v  =  f, v  ∀v ∈ H̃ 1/2 (). (2.2)

Here, ·, ·  denotes the duality pairing between H −1/2 () and H̃ 1/2 (). Throughout,
this generic notation will be used for the L 2 -inner product as well as for other dualities,
the domain being mentioned by the index.
A standard boundary element method for the approximate solution of (2.2) is to
select a piecewise polynomial subspace H̃h ⊂ H̃ 1/2 () and to define an approximant
ũ h ∈ H̃h by

W ũ h , v  =  f, v  ∀v ∈ H̃h . (2.3)

Such a scheme is known to converge quasi-optimally in the energy norm. In Sect. 5 we

will compare such a conforming approximation with our proposed Nitsche approach
and a Lagrangian multiplier variant.

3 Discrete variational formulation with Nitsche coupling

In this section, we introduce the Nitsche-based boundary element method for the
approximate solution of problem (2.2), and present the main result, Theorem 3.1.

3.1 Some preliminaries

We consider a decomposition of  into two non-intersecting polygonal sub-domains

1 and 2 . The extension to an arbitrary number of sub-domains is straightforward.
We will denote this partition of  as

T := {1 , 2 }.

The interface between the sub-domains is denoted by γ := ¯ 1 ∩ ¯ 2 . Throughout the

paper, we will use the notation vi for the restriction of a function v to a sub-domain i .
Also, as in [6], we will use the following notation for the jump on γ :

[v] := (v1 − v2 )|γ .

Corresponding to the decomposition of , we will need product Sobolev spaces, e.g.

H s (T ) := H s (1 ) × H s (2 ),

with usual product norm. This notation (putting the decomposition T instead of ) is
used generically, i.e. also for the spaces H̃ s (T ). We introduce the inner product

v, w T := v1 , w1 1 + v2 , w2 2

for v, w ∈ L 2 (T )(= L 2 ()) and its extension by duality to H̃ s (T ) × H −s (T ).

Throughout, and consistent with the previous notation vi := v|i , we will identify an

123
 F. Chouly, N. Heuer

element (v1 , v2 ) ∈ H s (T ) with v defined on  by v|i := vi (i = 1, 2) so that, for

instance, it makes sense to relate H s () ⊂ H s (T ).
For 1/2 ≤ s ≤ 1 and ν > 0, we introduce the following (semi-)norms, that are
needed for the error analysis:
⎧
⎪
⎪
2
⎪
⎪ |v|2
:= |v|2H s (i ) ,
⎪
⎨ H (T )
s
i=1
(3.1)
⎪
⎪
2
⎪
⎪
⎪
⎩ v2
Hν (T )
s := |v|2H s (i ) + ν[v]2L 2 (γ ) ,
i=1

where |v| H s (i ) is the Sobolev–Slobodeckij semi-norm as previously defined. The

case s = 1/2 will be used only for discrete functions where the jump across γ is
well defined. Furthermore, for s > 1/2 we will consider functions of H s (T ) with
vanishing trace on ∂ so that  ·  Hνs (T ) is indeed a norm.
To introduce the discrete scheme, let us define regular, quasi-uniform meshes Ti ,
i = 1, 2, of shape regular elements (quadrilaterals or triangles): ¯ i = ∪ K ∈Ti K̄ . The
maximum, respectively minimum, diameter of the elements of Ti is denoted by h i ,
respectively h i , and we define:

h := max {h 1 , h 2 }, h := min{h 1 , h 2 }.

Throughout this paper we assume that h < 1. This is no restriction of generality and is
just needed to simplify the writing of logarithmic terms. We introduce discrete spaces
on sub-domains consisting of piecewise (bi)linear functions:

X h,i := {v ∈ C 0 (i ); v| K is a polynomial of degree one for all K∈ Ti ; v|∂∩∂i = 0},

for i = 1, 2. We define a global discrete space on :

X h := X h,1 × X h,2 .

Note that functions v ∈ X h do satisfy the homogeneous boundary condition along

∂ but are in general discontinuous across the interface γ . Therefore X h ⊂ H̃ 1/2 (),
and this discrete space cannot be used directly for the discretization (2.3). Instead,
we reformulate (2.3) as a Nitsche variant so that X h can be used to approximate the
continuous problem (2.2).

3.2 Setting of the Nitsche-based domain decomposition

For the setup of the Nitsche method let us introduce the surface differential operators

T
curl ϕ := ∂x2 ϕ, −∂x1 ϕ, 0 , curl ϕ := ∂x1 ϕ2 − ∂x2 ϕ1 for ϕ = (ϕ1 , ϕ2 , ϕ3 )T .

123
A Nitsche-based domain decomposition method

The definitions of the surface curl operators are appropriate just for flat surfaces (as in
our case) but can be extended to open and closed Lipschitz surfaces (see e.g. [11,16]).
We define corresponding piecewise differential operators curlT and curlT as follows:

2
2
curlT ϕ := (curli ϕi ) , curlT ϕ :=
0
(curli ϕ i )0
i=1 i=1

for sufficiently smooth scalar function ϕ = (ϕ1 , ϕ2 ) and vector function ϕ = (ϕ 1 , ϕ 2 )

where, e.g., ϕi denotes restriction ϕ|i . Moreover, curli and curli refer to the restric-
tions of curl and curl, respectively, to i , and (·)0 indicates extension by zero to .
We made use of the notation introduced before ϕi = ϕ|i , ϕ i = ϕ|i . Furthermore,
we need the single layer potential operator V defined by:

1 ϕ(y)
V ϕ(x) := d S y , ϕ ∈ ( H̃ −1/2 ())3 , x ∈ .
4π |x − y|


The relation between the operators W and V is given by

W = curl (V curl ) (3.2)

as a linear continuous mapping from H̃ 1/2 () to H −1/2 () (see [26,29]), so that

W u, v  = curl u, V curl v  ∀u, v ∈ H̃ 1/2 (),

cf. also [16, Lemma 2.3]. As in [16,20], these formulas give rise to our discrete for-
mulation.
We define the following bilinear form on X h × X h :

AT (u h , vh ) := V curlT u h , curlT vh T
1 σ
+ T1 u h − T2 u h , [vh ] γ + [u h ], T1 vh − T2 vh γ
2 2
+ν[u h ], [vh ] γ , (3.3)

where ν > 0 and σ ∈ {−1, 1} are numerical parameters. The operators Ti are defined
as
 
Ti v = (V curlT v)|i · ti |γ (i = 1, 2). (3.4)

Here, ti is the unit tangential vector on ∂i (in mathematically positive orientation
when identifying i with a subset of R2 which is compatible with the identification
of  as a subset of R2 ). Note that Ti v is not well defined for v ∈ H̃ 1/2 () in general
since there is no well-defined trace from H 1/2 (i ) to ∂i .
The Nitsche-based boundary element method associated to problem (2.2) then
reads:

123
 F. Chouly, N. Heuer

Find u h ∈ X h such that

AT (u h , vh ) =  f, vh  (3.5)

for all vh ∈ X h .

Remark 3.1 For any function u ∈ H s () (s > 1/2), in particular for the solu-
tion of (2.1), there holds [u], [v] γ = [u], T1 v − T2 v γ = 0 for sufficiently
smooth v. Therefore the terms [u h ], [vh ] γ and [u h ], T1 vh − T2 vh γ are not required
for consistency of (3.5). However, the additional term σ2 [u h ], T1 vh − T2 vh γ in the
Nitsche-based formulation is of interest for two reasons [6, Remark 2.11]. First, for
σ = 1, the bilinear form AT (·, ·) becomes symmetric, as in the standard case (2.2).
This allows in particular to make use of fast linear solvers for symmetric matrices.
Also, for σ = −1, symmetry is lost, but we recover ellipticity of AT (·, ·) for any value
of the parameter ν > 0 [see Lemma 4.4 (i)]. In fact, any value of σ (including σ = 0)
can be chosen though only values −1 and 1 lead to interesting particular cases.

The main result of this paper is:

Theorem 3.1 Let u ∈ H r () with r ∈ (1/2, 1) be the solution of (2.1). In the case
σ = −1 let ν ≥ ρ for a fixed constant ρ > 0, and in the case σ = 1 let ν ≥ C1 | log h|3
for a sufficiently large constant C1 > 0. Then, the discrete problem (3.5) is uniquely
solvable and there exists a constant C2 > 0, depending on r , but not on ν, u and the
actual mesh, such that there holds the error estimate

u − u h  H 1/2 (T ) ≤ C2 max{ν −1/2 | log h|3/2 , ν 1/2 | log h|1/2 }h r −1/2 u H r () .
ν

A proof of this result will be given at the end of Sect. 4.

Remark 3.2 It is known that u ∈ H r () for any r < 1, see, e.g., [35]. Using this
regularity, Theorem 3.1 proves a convergence which is close to O(h 1/2 ), the optimal
one. The reduction to h r −1/2 for any r < 1 is due to the assumed regularity in standard
Sobolev spaces, and not a sub-optimality of the method. On the other hand, the log-
arithmic perturbation (of the order | log h| for σ = −1, ν = O(| log h|) and | log h|2
for σ = 1, ν = O(| log h|3 )) is due to the Nitsche coupling, and is also present in
non-conforming approaches (| log h|3/2 appears in the Lagrangian multiplier approach
[16], and in the mortar coupling [20] the exponent is 2). It is unknown whether these
logarithmic terms in the upper bounds are optimal. Note that ν appears in the upper
bound since it is also included in the definition of the norm  ·  H 1/2 (T ) .
ν

4 Technical results and the proof of the main theorem

In Sect. 4.1, we present some preliminary results and lemmas. In Sect. 4.2 we then
prove the consistency of the method, using an integration-by-parts formula coming
from [16,20]. Discrete continuity and discrete ellipticity are studied in Sect. 4.3. We
conclude with a proof of the main theorem in Sect. 4.4.

123
A Nitsche-based domain decomposition method

The steps followed in the error analysis are quite similar to those of the analysis of
a Nitsche-based method for finite elements (see e.g. [6]). The main difficulty in the
case of boundary elements consists in the non-existence of a well-posed variational
Nitsche formulation. Error estimates are wanted in spaces related to H 1/2 () where
no well-defined trace operator exists. Therefore, the numerical analysis of (3.5) makes
use of a whole family of Sobolev spaces H r with r close to 1/2. Additional difficulty
in our case is that operators are non-local in contrast to the finite element setting.
In opposition to the mortar boundary element method [20], no inf-sup condition needs
to be checked since no Lagrangian multipliers are introduced.

4.1 Preliminary results

We first introduce the following spaces for the definition of the single layer potential
operator V (see [16]):
s−1
3
H̃ t () := {ψ ∈ H̃ s−1 () ; ψ · n = 0},

3 (4.1)
H st () := {ψ ∈ H s () ; ψ · n = 0},

where 0 ≤ s ≤ 1 and the normal vector n has been defined previously. We will make
use of the continuity (see [13]):
s−1
V : H̃ t () → H st (), 0 ≤ s ≤ 1. (4.2)

Let us recall the following technical results.

Lemma 4.1 Let R ⊂ R2 be a Lipschitz domain with boundary ∂ R.
(i) For any s ∈ (−1/2, 1/2), and any v ∈ H s (R), there holds

1
v H̃ s (R)  v H s (R) .
1/2 − |s|

(ii) For any s ∈ (1/2, 1], and any v ∈ H s (R), there holds

1
v L 2 (∂ R)  √ v H s (R) .
s − 1/2

(iii) Let ρ ∈ (0, 1/2) be fixed. Then for i = 1, 2, curli is a continuous mapping
from H s (i ) to H s−1
t (i ) uniformly for s ∈ [1/2, 1 − ρ].
Proof Parts (i) and (ii) are proved in [21, Lemma 5] and [16, Lemma 4.3], respec-
tively. Assertion (iii) follows by interpolation from standard continuity properties of
curli , cf. [20, Lemma 3.4]. For details on interpolation properties and uniform norm
equivalences we refer to standard references [8,25,27], see also [21] for further details.


The next lemma studies continuity properties of the operators Ti defined in (3.4).

123
 F. Chouly, N. Heuer

Lemma 4.2 For i = 1, 2 there holds

Ti v L 2 (γ )  (s −1/2)−1/2  curlT v s−1 ∀v ∈ H s (T ) 1/2 < s ≤ 1, (4.3)

H̃ t ()

Ti v L 2 (γ )  (s −1/2)−3/2 |v| H s (T ) ∀v ∈ H s (T ) 1/2 < s ≤ 1, (4.4)

Ti vh  L 2 (γ )  | log h |3/2 |vh | H 1/2 (T ) ∀vh ∈ X h . (4.5)

Proof Let v ∈ H s (T ), with 1/2 < s ≤ 1. We use Lemma 4.1(ii) and the continuity
of V (4.2) to bound

1 1
Ti v2L 2 (γ )  V curlT v2H s ()   curlT v2 s−1 .
s − 1/2 t s − 1/2 H̃ t ()

This proves (4.3).

By a standard norm estimate from domain decomposition (see, e.g. [21, Lemma 1])
and (i), (iii) of Lemma 4.1 we obtain

2
 curlT v2 s−1   curli v2 s−1
H̃ t () H̃ t (i )
i=1

1
2
1
2
  curl  v2
s−1  v2H s (i ) .
(s −1/2)2 i H t (i ) (s −1/2)2
i=1 i=1

(This, by the way, proves that the upper bound in (4.3) is finite.) Combining these two
estimates and using a quotient space argument proves (4.4).
Now we pick vh ∈ X h , apply (4.4) and the inverse property

|vh | H s (i )  h 1/2−s |vh | H 1/2 (i ) (4.6)

which follows from the standard inverse property [21, Lemma 4] together with a quo-
tient space argument. For s bounded away from 0 and 1, the Sobolev–Slobodeckij
norm and the norm defined by interpolation (K-method) are uniformly equivalent
(cf. the references given in the proof of Lemma 4.1) so that the constant in (4.6) is
independent of h for s in a neighborhood of 1/2. This yields

Ti vh  L 2 (γ )  (s − 1/2)−3/2 |vh | H s (T )  h 1/2−s (s − 1/2)−3/2 |vh | H 1/2 (T ) .

The choice s = 1/2 + | log h|−1 gives h 1/2−s = e (provided that h < 1) and
proves (4.5). 


4.2 Consistency of the Nitsche formulation

In this part, we show that the Nitsche formulation (3.5) for the hypersingular operator is
consistent, a classical result for the Nitsche method in the standard case (see e.g. [6]).
Usual difficulty in the (continuous) boundary element setting is that the boundary

123
A Nitsche-based domain decomposition method

operator Ti is not well defined for v ∈ H̃ 1/2 (). In our particular case we avoid this
difficulty thanks to the increased regularity of the solution u of (2.1).
First, we need to start from an appropriate integration-by-parts formula for the
hypersingular operator. For the convenience of the reader we recall the setting from
[16,20]. For a smooth scalar function v and a smooth tangential vector field ϕ, inte-
gration by parts on i gives

ϕ · ti , vi ∂i = curli ϕ, vi i − curli vi , ϕ i ,

for i = 1, 2. We apply this formula to ϕ = (V curl u)|i , so that:

(V curl u)|i · ti , vi ∂i = curli (V curl u), vi i − curli vi , V curl u i .

Recalling the definition of Ti , and using a function v that vanishes on ∂, we obtain

Ti u, vi γ = curli (V curl u), vi i − curli vi , V curl u i . (4.7)

We formulate the consistency of the Nitsche method.

Lemma 4.3 Let ν > 0 and |σ | = 1. Then, for 1/2 < s < 1, the bilinear form AT (·, ·)
is well defined on H̃ s () × X h . Moreover, the Nitsche formulation is consistent, i.e.
the solution u of (2.1) (W u = f ∈ L 2 ()) solves the discrete setting (3.5),

AT (u, vh ) =  f, vh  ∀vh ∈ X h .

Proof For v ∈ H̃ s () (1/2 < s < 1), the trace of v on γ is well defined and belongs to
L 2 (γ ). Furthermore, Ti v (i = 1, 2) is also well defined and belongs to L 2 (γ ). There-
fore, the bilinear form AT (·, ·) introduced in (3.3) is well defined on H̃ s () × X h . It
is well known that the solution u of (2.1) belongs to H s () for any s < 1, see, e.g.,
[35]. Therefore, AT (u, vh ) is well defined for all vh ∈ X h and, furthermore, [u] = 0.
Since [(V curl u)|1 − (V curl u)|2 ]|γ = 0 and t1 = −t2 on γ , there holds

T1 u + T2 u = 0 on γ . (4.8)

We obtain for vh ∈ X h

1 σ
AT (u, vh ) = V curlT u, curlT vh T + T1 u −T2 u, [vh ] γ + [u], T1 vh −T2 vh γ
2 2
+ν[u], [vh ] γ
1
= V curl u, curlT vh T + T1 u − T2 u, [vh ] γ .
2

Using (4.8) to write Ti u = 21 Ti u − 21 T j u (i = j), and rearranging terms, we obtain

1 2
T1 u − T2 u, [vh ] γ = Ti u, vh,i γ
2
i=1

123
 F. Chouly, N. Heuer

so that, together with the integration-by-parts formula (4.7) and relation (3.2),
2 

AT (u, vh ) = V curl u, curli vh,i i + Ti u, vh,i γ
i=1
2
= curli (V curl u), vh,i i
i=1
= curl (V curl u), vh  = W u, vh  =  f, vh  .
This proves the lemma. 


4.3 Discrete ellipticity and continuity

Main advantage of the Nitsche method is that it yields elliptic bilinear forms in the case
of elliptic problems. In the boundary element setting, we do not have an appropriate
variational formulation. Nevertheless, discrete ellipticity is still achievable. This is
contents of the first lemma. Afterwards, we briefly state discrete continuity without
giving a bound for the continuity constant. This bound is studied in more detail in the
proof of the main theorem in Sect. 4.4.

Lemma 4.4 (i) Let σ = −1. There exists a constant C > 0 independent of ν and
h such that, for all ν > 0, there holds
AT (vh , vh ) ≥ C vh 2 1/2 ∀vh ∈ X h
Hν (T )

and
 √ 
AT (vh , vh ) ≥ C  curlT vh  H̃ −1/2 () + ν[vh ] L 2 (γ ) vh  H 1/2 (T ) ∀vh ∈ X h .
ν

(ii) Let σ = 1. There exist two constants C1 , C2 > 0 which are independent of ν
and h such that, if ν ≥ C1 | log h|3 , then
AT (vh , vh ) ≥ C2 vh 2 1/2 ∀vh ∈ X h
Hν (T )

and
 √ 
AT (vh , vh ) ≥ C2  curlT vh  H̃ −1/2 () + ν[vh ] L 2 (γ ) vh  H 1/2 (T ) ∀vh ∈ X h .
ν

Proof (i) Case σ = −1.

Let vh ∈ X h . By the ellipticity of V and [16, Lemma 4.1] there holds

2
V curlT vh , curlT vh T   curlT vh 2 −1/2   curli vh 2 −1/2
H̃ () t Ht (i )
i=1

2
 |vh |2H 1/2 ( ) = |vh |2H 1/2 (T ) . (4.9)
i
i=1

123
A Nitsche-based domain decomposition method

This proves that

AT (vh , vh ) = V curlT vh , curlT vh T + ν[vh ], [vh ] γ

 |vh |2H 1/2 (T ) + ν[vh ]2L 2 (γ ) = vh 2 1/2 (4.10)
Hν (T )

(which is the first assertion) and also

AT (vh , vh )   curlT vh 2 −1/2 + ν[vh ]2L 2 (γ ) .

H̃ t ()

Both estimates together prove the second assertion.

(ii) In the case σ = 1 we obtain for vh ∈ X h

AT (vh , vh ) = V curlT vh , curlT vh T +ν[vh ], [vh ] γ +T1 vh −T2 vh , [vh ] γ .

By (4.9) there holds

AT (vh , vh )  |vh |2H 1/2 (T ) +ν[vh ]2L 2 (γ ) −|T1 vh −T2 vh , [vh ] γ |. (4.11)

Using the Cauchy–Schwarz and Young’s inequalities, and (4.4), we bound

1
|T1 vh − T2 vh , [vh ] γ |  δT1 vh − T2 vh 2L 2 (γ ) + [vh ]2L 2 (γ )
δ
δ 1
 |vh | H s (T ) + [vh ]2L 2 (γ ) ∀δ > 0,
2
(s − 1/2)3 δ

so that, together with (4.11) and the inverse property (4.6),

 h 1−2s 
AT (vh , vh )  1 − δc1 |vh |2H 1/2 (T )
(s − 1/2)3
 c2 
+ ν− [vh ]2L 2 (γ ) ∀δ > 0 (4.12)
δ
for two unknown constants c1 , c2 > 0. First selecting

1 c3
s= (1 + | log h|−1 ) and δ = | log h|−3 for c3 ∈ (0, 1)
2 8c1 e

this yields

h 1−2s
1 − δc1 = 1 − c3 ,
(s − 1/2)3

and then selecting c4 > 1 we obtain for ν ≥ c4 cδ2  | log h|3 the bound

c2
ν−  ν.
δ

123
 F. Chouly, N. Heuer

Therefore, (4.12) proves that

AT (vh , vh )  |vh |2H 1/2 (T ) + ν[vh ]2L 2 (γ ) = vh 2 1/2

Hν (T )

for ν ≥ c4 cδ2 = 8c1 cc32 c4 e | log h|3 . This is the first estimate in (ii). As in the case
σ = −1, and using (4.3) in addition to (4.4), one proves the second assertion
under the same condition on ν. 


Lemma 4.5 Let ν > 0 and |σ | = 1. The bilinear form AT is continuous:

AT (vh , wh )  C(h, ν)vh  H 1/2 (T ) wh  H 1/2 (T ) ∀vh , wh ∈ X h ,

ν ν

with C(h, ν) > 0 a number that depends on the mesh (indicated by h) and ν.

Proof This estimate follows by using the mapping properties of the involved operators
V, curlT , Ti and inverse properties of discrete functions. 


4.4 Proof of the main theorem

By Lemma 4.4, and under the stated assumptions, the bilinear form AT is elliptic.
Moreover, by Lemma 4.5, this bilinear form is also continuous on X h (with bound
depending on the mesh) so that problem (3.5) has a unique solution. It remains to bound
the error. To this end we follow the lines of a Strang estimate for non-conforming
methods. By Lemma 4.4 there holds for any vh ∈ X h

u − u h  H 1/2 (T ) ≤ u − vh  H 1/2 (T ) + u h − vh  H 1/2 (T )

ν ν ν
 u − vh  H 1/2 (T )
ν
AT (u h − vh , wh )
+ sup √ . (4.13)
wh ∈X h \{0}  curlT wh  −1/2 + ν[wh ] L 2 (γ )
H̃ t ()

Now, by the consistency (see Lemma 4.3) we obtain AT (u h − vh , wh ) = AT (u − vh ,

wh ) so that we continue bounding [using duality estimates and the continuity of V
(4.2)]

AT (u h − vh , wh ) = V curlT (u − vh ), curlT wh T
1
+ T1 (u − vh ) − T2 (u − vh ), [wh ] γ
2
σ
+ [u − vh ], T1 wh − T2 wh γ + ν[u − vh ], [wh ] γ
2
  curlT (u − vh ) −1/2  curlT wh  −1/2
H̃ t () H̃ t ()
2 

+ Ti (u − vh ) L 2 (γ ) [wh ] L 2 (γ )
i=1

123
A Nitsche-based domain decomposition method


+Ti wh  L 2 (γ ) [u − vh ] L 2 (γ )
+ν[u − vh ] L 2 (γ ) [wh ] L 2 (γ ) . (4.14)

We bound the terms on the right-hand side.

In the following let s be a small positive number. Exactly as in the proof of (4.4) we
obtain, by using a domain decomposition estimate and parts (i), (iii) of Lemma 4.1,

2
 curlT (u − vh )2 −1/2   curlT (u − vh )2 −1/2
H̃ t () H̃ t (i )
i=1

1
2
  curlT (u − vh )2 s−1/2
s2 Ht (i )
i=1

1 2
 |u − vh |2H s+1/2 ( ) , (4.15)
s2 i
i=1

and estimate (4.4) proves

1
Ti (u − vh ) L 2 (γ )  |u − vh | H s+1/2 (T ) . (4.16)
s 3/2

Eventually, by (4.3) and the inverse property (cf. (4.6) and [21, Lemma 4]),

1 1
Ti wh 2L 2 (γ )   curlT wh 2 s−1/2  2s  curlT wh 2 −1/2 . (4.17)
s H̃ t () sh H̃ t ()

Combination of (4.14)–(4.17), and using that ν  1 by assumption, proves that for

any wh ∈ X h \{0} and any small s > 0 there holds

AT (u h − vh , wh )
√
 curlT wh  −1/2 + ν[wh ] L 2 (γ )
H̃ t ()
   
 s −1 + ν −1/2 s −3/2 |u − vh | H s+1/2 (T ) + h −s s −1/2 + ν 1/2 [u − vh ] L 2 (γ ) .

By Lemma 4.1(ii) we bound

[u − vh ] L 2 (γ )  s −1/2 u − vh  H s+1/2 (T )

where the latter norm is the standard one:

u − vh 2H s+1/2 (T ) := u − vh 2L 2 () + |u − vh |2H s+1/2 (T ) .

123
 F. Chouly, N. Heuer

Combination with the previous estimate yields

AT (u h − vh , wh )
 curlT wh  −1/2 + ν 1/2 [wh ] L 2 (γ )
H̃ t ()
 
−1 −1/2 −3/2
 s +ν s + h −s s −1 + ν 1/2 s −1/2 u − vh  H s+1/2 (T ) . (4.18)

By a standard approximation result we have that for r ∈ (1/2, 1) there exists vh ∈ X h

such that

u − vh  H s+1/2 (T )  h r −s−1/2 u H r () .

Combining (4.13) and (4.18), this proves that

 
u − u h  H 1/2 (T )  ν −1/2 s −3/2 h −s + s −1 h −2s + ν 1/2 s −1/2 h −s h r −1/2 u H r () .
ν

 
Selecting s = | log h|−1 and bounding | log h| ≤ 1
2 ν −1/2 | log h|3/2 + ν 1/2 | log h|1/2
this estimate finishes the proof of Theorem 3.1.

5 Numerical results

We consider the model problem (2.1) with  = (0, 1) × (0, 1) and f = 1. Previ-
ously, for ease of presentation, we have restricted our analysis to two sub-domains. It
generalizes in a straightforward way to an arbitrary number of sub-domains. Also, the
homogeneous boundary condition of discrete functions on ∂ can be included in the
Nitsche formulation. This is done in our numerical experiments, with a decomposition
of  into three sub-domains, as indicated in Fig. 1 below.
In this case, the bilinear form is

1
AT (u h , vh ) := V curlT u h , curlT vh T + Ti u h − T j u h , [vh ] γi j
2
0≤i< j
σ
+ [u h ], Ti vh − T j vh γi j + ν [u h ], [vh ] γi j .
2
0≤i< j 0≤i< j

Here, the operators Ti are defined as before and γi j is the interface between i and  j .
The index i = 0 appears due to the boundary condition on ∂, i.e. 0 is the exterior
domain and, formally, u h = vh = 0 on 0 .
We use meshes Th on  which are uniform on sub-domains and consist of rectangles.
The initial mesh is shown in the left-upper part of Fig. 1 and is refined by introducing,
in each refinement step, another “slide” in x1 - and x2 -directions on each sub-domain.
The next three meshes are also shown in Fig. 1. The discrete spaces X h are made of
piecewise bilinear polynomials on the mesh Th which are continuous on sub-domains.

123
A Nitsche-based domain decomposition method

Fig. 1 Decomposition of  into three sub-domains and initial mesh sequence

The Nitsche method reads as before (using the updated bilinear form): Find u h ∈ X h
such that

AT (u h , vh ) =  f, vh  for all vh ∈ X h . (5.1)

Since the exact solution u of (2.1) is unknown, the error

u − u h 2 1/2 = |u − u h |2H 1/2 (T ) + ν[u h ]2L 2 (∪ ∂ )

Hν (T ) i i

cannot be computed directly. Instead, we approximate an upper bound to the semi-

norm |u − u h | H 1/2 (T ) as follows.
First, note that there holds

|u − u h |2H 1/2 (T )  V curlT (u − u h ), curlT (u − u h )  ,

due to ellipticity of V , cf. (4.9) which also holds for u − u h . Taking into account that
u is solution of (2.1) and u h is solution of (5.1), we find that

|u − u h |2H 1/2 (T )  V curl u, curl u  + V curlT u h , curlT u h 

−2V curl u, curlT u h 

123
 F. Chouly, N. Heuer

1
= W u, u  +  f, u h  − Ti u h − T j u h , [u h ] γi j
2
0≤i< j
σ
− [u h ], Ti u h − T j u h γi j
2
0≤i< j

−ν [u h ], [u h ] γi j −2W u, u h  + Ti u −T j u, [u h ] γi j
0≤i< j 0≤i< j
1+σ
= W u, u  −  f, u h  − Ti u h − T j u h , [u h ] γi j
2
0≤i< j

−ν [u h ], [u h ] γi j + Ti u − T j u, [u h ] γi j .
0≤i< j 0≤i< j

Now denoting by γ the union of all interfaces including the boundary of , using a
generic notation T for the sub-domain related operators Ti , bounding [u h ], [u h ] γi j ≥0,
applying the Cauchy–Schwarz inequality and (4.5), we obtain

|u − u h |2H 1/2 (T )  W u, u  −  f, u h  + T u h  L 2 (γ ) [u h ] L 2 (γ )

+T u L 2 (γ ) [u h ] L 2 (γ )  W u, u  −  f, u h 
 
+ | log h|3/2 |u h | H 1/2 (T ) + T u L 2 (γ ) [u h ] L 2 (γ ) .

Note that for this specific problem, T u ∈ L 2 (γ ). Furthermore, since the method is
stable, |u h | H 1/2 (T ) is bounded independently of h. This proves that

1/2
|u − u h | H 1/2 (T )  |W u, u  −  f, u h  |1/2 + | log h|3/4 [u h ] L 2 (γ ) .

1/2
Moreover, since [u h ] L 2 (γ )  [u h ] L 2 (γ ) there also holds (assuming that ν  1)

√ 1/2
u − u h  H 1/2 (T )  |W u, u  −  f, u h  |1/2 + | log h|3/4 ν[u h ] L 2 (γ ) .
ν

1/2
The terms  f, u h  and [u h ] L 2 (γ ) are easy to compute. The energy norm W u, u 
of u can be approximated through extrapolation, denoted by uex in the following,
see [14]. Therefore,
  √ 
u2 −  f, u h  1/2 + | log h|3/4 ν[u h ]1/2 uex
ex L (γ )
2

is a computable and reasonable measure for an upper bound of the error u−u h  H 1/2 (T )
ν
normalized by u H̃ 1/2 () . Below we present numerical results for the two contribu-
tions

123
A Nitsche-based domain decomposition method

relative errors

0.1

|log h|*h^{1/2}
conforming BEM
Nitsche H^{1/2}, asym, nu=1
Nitsche H^{1/2}, asym, nu=2
Nitsche H^{1/2}, asym, nu=5
Nitsche H^{1/2}, asym, nu=10
0.01
10 100 1000 10000
dim(X_h)

Fig. 2 Skew-symmetric Nitsche method (σ = −1): relative error curves (“H 1/2 ” error). Comparison with
conforming BEM

 
u2 −  f, u h  1/2 /uex (5.2)
ex

(referred to as “H 1/2 ” error in the figures) and

1/2
[u h ] L 2 (γ ) /uex (5.3)

(referred to as “L 2 ” error).
We first consider some tests in the skew-symmetric case (σ = −1), for different
values of ν. The corresponding results are given in Figs. 2 and 3, using a double log-
arithmic scale and plotting versus the dimension of the discrete space X h . The first
figure shows also the errors of the conforming approximation on uniform meshes,
plotted versus the degrees of freedom of the conforming spaces. Figure 2 presents
results for the term (5.2) of the error and indicates that the Nitsche-based method
converges for all the tested values of ν = 1, 2, 5, 10, with a logarithmic perturbation
of the convergence, as expected by the theory (Theorem 3.1). As a consequence, the
convergence is asymptotically a bit slower (by a factor of | log h|) than in the case of the
conforming BEM. The latter method converges like O(h 1/2 ), and for comparison we
have given the curve | log h|h 1/2 as well (with a constant factor for adjustment). For all
studied values of ν the curves exhibit the same asymptotic convergence order, though
their initial behavior differ. In particular, for ν ≥ 2, a minimum is reached quickly,
after which the asymptotic behavior is recovered. Apparently, for any particular mesh,
there is simply an optimal value of ν for term (5.2). Figure 3 shows that the other part
of the error [given by (5.3)] also behaves as predicted. All the curves are parallel and
the parameter ν does not seem to have a great influence, except for shifting the curves

123
 F. Chouly, N. Heuer

relative errors

0.1

Nitsche H^{1/2}, asym, nu=1

Nitsche L^2, asym, nu=1
Nitsche H^{1/2}, asym, nu=2
Nitsche L^2, asym, nu=2
0.01 Nitsche H^{1/2}, asym, nu=5
Nitsche L^2, asym, nu=5
Nitsche H^{1/2}, asym, nu=10
Nitsche L^2, asym, nu=10
10 100 1000 10000
dim(X_h)

Fig. 3 Skew-symmetric Nitsche method (σ = −1): relative error curves (“H 1/2 ” error and “L 2 ” error)

1
relative errors

|log h|*h^{1/2}
conforming BEM
Nitsche H^{1/2}, sym, nu=1
0.1 Nitsche H^{1/2}, sym, nu=10
Nitsche H^{1/2}, sym, nu=|log h|
Nitsche H^{1/2}, sym, nu=|log h|^2
10 100 1000 10000
dim(X_h)

Fig. 4 Symmetric Nitsche method (σ = 1): relative error curves (“H 1/2 ” error). Comparison with
conforming BEM

which corresponds to multiplication of √ the error by a constant. This is compatible with

the theory since we ignored the factor ν in the plots of the L 2 -errors.
Next we study the symmetric case (σ = 1). The corresponding results are given in
Figs. 4 and 5. As expected, if the value of ν is not sufficiently large, the method does
not converge (see the curve for ν = 1 in Fig. 4). Indeed, if ν is too small, discrete
ellipticity of AT (·, ·) cannot be guaranteed (see Lemma 4.4). Taking higher values of

123
A Nitsche-based domain decomposition method

relative errors

0.1

0.01 Nitsche H^{1/2}, sym, nu=10

Nitsche L^2, sym, nu=10
Nitsche H^{1/2}, sym, nu=|log h|
Nitsche L^2, sym, nu=|log h|
Nitsche H^{1/2}, sym, nu=|log h|^2
Nitsche L^2, sym, nu=|log h|^2
10 100 1000 10000
dim(X_h)

Fig. 5 Symmetric Nitsche method (σ = 1): relative error curves (“H 1/2 ” error and “L 2 ” error)

Fig. 6 Conforming
approximation, obtained
0.45
from (2.3)
0.4
0.35
0.3
0.25
0.2
0.15
0.1
0.05
0
0.4
0.2
0
-0.2 0.4
0.2
-0.4 0
-0.2
-0.4

ν ensures convergence of the method. In particular, if ν is not taken as a constant but a

power of | log h| the asymptotic behavior improves. For ν = | log h|2 , the behavior of
the conforming BEM method is recovered (in the observed region), with quasi-optimal
convergence. Note that theoretically, a sufficient condition to guarantee discrete ellip-
ticity and convergence (in the symmetric case) is ν  | log h|3 (cf. Lemma 4.4 and
Theorem 3.1). Figure 5 presents the L 2 (γ )-errors (5.3) as well.
For illustration, below we present approximations to u [the solution of (2.1)].
Figure 6 shows a conforming approximation (including homogeneous boundary con-
dition) whereas Figs. 7 and 8 show the Nitsche results, for σ = −1 (skew-symmetric)
and σ = 1 (symmetric), respectively. The Nitsche approximations are based on the
last mesh displayed in Fig. 1. For all the Nitsche results, discontinuities of the solution

123
 F. Chouly, N. Heuer

σ=-1, ν=1 σ=-1, ν=10

0.45 0.45
0.4 0.4
0.35 0.35
0.3 0.3
0.25 0.25
0.2 0.2
0.15 0.15
0.1 0.1
0.05 0.05
0 0
0.4 0.4
0.2 0.2
0 0
-0.2 0.4 -0.2 0.4
0 0.2 0 0.2
-0.4 -0.2 -0.4 -0.2
-0.4 -0.4

Fig. 7 Skew-symmetric Nitsche approximation, ν = 1 (left) and ν = 10 (right)

σ=1, ν=1 σ=1, ν=10

0.45 0.45
0.4 0.4
0.35 0.35
0.3 0.3
0.25 0.25
0.2 0.2
0.15 0.15
0.1 0.1
0.05 0.05
0 0
0.4 0.4
0.2 0.2
0 0
-0.2 0.4 -0.2 0.4
0 0.2 0 0.2
-0.4 -0.2 -0.4 -0.2
-0.4 -0.4

Fig. 8 Symmetric Nitsche approximation, ν = 1 (left) and ν = 10 (right)

on the interfaces and non-zero values on the boundary ∂ can be observed, as expected
due to the non-conforming character of the method. The magnitude of these disconti-
nuities is greater for ν = 1 than for ν = 10 as expected since ν controls the penalization
of the jump on the interfaces and the boundary. For ν = 1, a strong difference between
the symmetric and skew-symmetric cases appears: the approximate solution from the
symmetric case is far from satisfying the homogeneous boundary condition whereas
the approximation from the skew-symmetric case does well. This seems to be related
with the non-ellipticity of the symmetric bilinear form for small values of ν whereas
the skew-symmetric formulation is robust in ν.
Finally, in Fig. 9 we compare the Nitsche method (symmetric and skew-symmetric
versions) with the Lagrangian multiplier-based method presented in [20], in the case
of only one sub-domain. One observes that the symmetric Nitsche method (with ν =
| log h|2 ) and the Lagrangian multiplier method have the same asymptotic conver-
gence, which appears to be quasi-optimal (without logarithmic perturbation) in this
example. The skew-symmetric method (with ν = 5) on the other hand, is almost
quasi-optimal, i.e. with logarithmic perturbation.

123
A Nitsche-based domain decomposition method

relative errors

0.1

Nitsche H^{1/2}, sigma=-1, nu=5

Nitsche L^2, sigma=-1, nu=5
Nitsche H^{1/2}, sigma=1, nu=|log h|^2
Nitsche L^2, sigma=1, nu=|log h|^2
Lagrange H^{1/2}
Lagrange L^2

10 100 1000 10000

dim(X_h)

Fig. 9 Comparing Nitsche and Lagrangian multiplier methods: relative error curves [upper bounds (5.2)
and (5.3) for all cases]

Concluding, the numerical experiments are in good agreement with the theory,
and illustrate the applicability of the Nitsche-based domain decomposition method
for hypersingular integral equations, e.g. as a possible alternative to a Lagrangian
multiplier approach which requires an additional unknown and destroys ellipticity.
In particular, the symmetric case seems to be more appealing due to its competitive
convergence for large values of ν, and since it maintains symmetry.

Acknowledgments Part of this work has been carried out during research stays of Franz Chouly at the
Universidad Técnica Federico Santa María, Valparaíso, and the Pontificia Universidad Católica de Chile,
Santiago. The authors thank two anonymous referees for their critical remarks which helped to improve the
presentation.

References

1. Arnold, D.: An interior penalty finite element method with discontinuous elements. SIAM J. Numer.
Anal. 19, 742–760 (1982)
2. Astorino, M., Chouly, F., Fernández, M.A.: Robin based semi-implicit coupling in fluid-structure
interaction: stability analysis and numerics. SIAM J. Sci. Comput. 31, 4041–4065 (2009/10)
3. Barbosa, H.J.C., Hughes, T.J.R.: The finite element method with Lagrange multipliers on the boundary:
circumventing the Babuška-Brezzi condition. Comput. Methods. Appl. Mech. Eng. 85, 109–128 (1991)
4. Barbosa, H.J.C., Hughes, T.J.R.: Boundary Lagrange multipliers in finite element methods: error
analysis in natural norms. Numer. Math. 62, 1–15 (1992)
5. Becker, R., Burman, E., Hansbo, P.: A Nitsche extended finite element method for incompress-
ible elasticity with discontinuous modulus of elasticity. Comput. Methods Appl. Mech. Eng. 198,
3352–3360 (2009)
6. Becker, R., Hansbo, P., Stenberg, R.: A finite element method for domain decomposition with
non-matching grids. M2AN Math. Model. Numer. Anal. 37, 209–225 (2003)

123
 F. Chouly, N. Heuer

7. Ben Belgacem, F.: The mortar finite element method with Lagrange multipliers. Numer. Math. 84,
173–197 (1999)
8. Bergh, J., Löfström, J.: Interpolation spaces. In: Grundlehren der mathematischen Wissenschaften,
vol. 223. Springer, Berlin (1976)
9. Bernardi, C., Maday, Y., Patera, A.T.: Domain decomposition by the mortar element method.
In: Asymptotic and Numerical Methods for Partial Differential Equations with Critical Parameters
(Beaune, 1992). NATO Advanced Science Institutes Series C: Mathematical and Physical Sciences,
vol. 384, pp. 269–286. Kluwer Academic Publishers, Dordrecht (1993)
10. Bernardi, C., Maday, Y., Patera, A.T.: A new nonconforming approach to domain decomposition: the
mortar element method. In: Nonlinear Partial Differential Equations and Their Applications. Collège de
France Seminar, vol. XI (Paris, 1989–1991). Pitman Research Notes in Mathematics Series, vol. 299,
pp. 13–51. Longman, Harlow (1994)
11. Buffa, A., Costabel, M., Sheen, D.: On traces for H(curl,) in Lipschitz domains. J. Math. Anal.
Appl. 276, 845–867 (2002)
12. Burman, E., Fernández, M.A.: Stabilization of explicit coupling in fluid-structure interaction involving
fluid incompressibility. Comput. Methods Appl. Mech. Eng. 198, 766–784 (2009)
13. Costabel, M.: Boundary integral operators on Lipschitz domains: elementary results. SIAM J. Math.
Anal. 19, 613–626 (1988)
14. Ervin, V.J., Heuer, N., Stephan, E.P.: On the h- p version of the boundary element method for Symm’s
integral equation on polygons. Comput. Methods Appl. Mech. Eng. 110, 25–38 (1999)
15. Fritz, A., Hüeber, S., Wohlmuth, B.I.: A comparison of mortar and Nitsche techniques for linear
elasticity. Calcolo 41, 115–137 (2004)
16. Gatica, G.N., Healey, M., Heuer, N.: The boundary element method with Lagrangian multipliers.
Numer. Methods Partial Differ. Equ. 25, 1303–1319 (2009)
17. Grisvard, P.: Elliptic Problems in Nonsmooth Domains. Pitman Advanced Pub. Program,
Boston (1985)
18. Han, H.: The boundary integro-differential equations of three-dimensional Neumann problem in linear
elasticity. Numer. Math. 68, 269–281 (1994)
19. Hansbo, P., Hermansson, J., Svedberg, T.: Nitsche’s method combined with space-time finite ele-
ments for ALE fluid-structure interaction problems. Comput. Methods Appl. Mech. Eng. 193,
4195–4206 (2004)
20. Healey, M., Heuer, N.: Mortar boundary elements. SIAM J. Numer. Anal. 48, 1395–1418 (2010)
21. Heuer, N.: Additive Schwarz method for the p-version of the boundary element method for the single
layer potential operator on a plane screen. Numer. Math. 88, 485–511 (2001)
22. Heuer, N., Sayas, F.-J.: Crouzeix–Raviart boundary elements. Numer. Math. 112, 381–401 (2009)
23. Hsiao, G.C., Wendland, W.L. : Domain decomposition in boundary element methods. In: Glowinski,
R., Kuznetsov, Y.A., Meurant, G.A., Périaux, J., Widlund, O.B. (eds.) Domain Decomposition Methods
for Partial Differential Equations, pp. 41–49. SIAM, Philadelphia (1991)
24. Langer, U., Of, G., Steinbach, O., Zulehner, W.: Inexact fast multipole boundary element tearing
and interconnecting methods. In: Domain Decomposition Methods in Science and Engineering XVI.
Lecture Notes in Engineering and Computer Science, vol. 55, pp. 405–412. Springer, Berlin (2007)
25. Lions, J.L., Magenes, E.: Non-Homogeneous Boundary Value Problems and Applications.
Springer, Berlin (1972)
26. Maue, A.-W.: Zur Formulierung eines allgemeinen Beugungsproblems durch eine Integralglei-
chung. Zeitschrift für Physik 126, 601–618 (1949)
27. McLean, W.: Strongly Elliptic Systems and Boundary Integral Equations. Cambridge University
Press, London (2000)
28. Mund, P., Stephan, E.P., Weiße, J.: Two-level methods for the single layer potential in R3 .
Computing 60, 243–266 (1998)
29. Nédélec, J.-C.: Integral equations with nonintegrable kernels. Integr. Equ. Oper. Theory 5,
562–572 (1982)
30. Nitsche, J.: Über ein Variationsprinzip zur Lösung von Dirichlet-Problemen bei Verwendung von
Teilräumen, die keinen Randbedingungen unterworfen sind. Abhandlungen aus dem Mathematischen
Seminar der Universität Hamburg 36, 9–15 (1971)
31. Pechstein, C.: Boundary element tearing and interconnecting methods in unbounded domains. Appl.
Numer. Math. 59, 2824–2842 (2009)

123
A Nitsche-based domain decomposition method

32. Quarteroni, A., Valli, A.: Domain decomposition methods for partial differential equations. In: Numer-
ical Mathematics and Scientific Computation. The Clarendon Press, Oxford University Press, Oxford
Science Publications, Oxford (1999)
33. Reusken, A., Nguyen, T.H.: Nitsche’s method for a transport problem in two-phase incompressible
flows. J. Fourier Anal. Appl. 15, 663–683 (2009)
34. Stenberg, R.: On some techniques for approximating boundary conditions in the finite element
method. J. Comput. Appl. Math. 63, 139–148 (1995)
35. Stephan, E.P.: A boundary integral equation method for three-dimensional crack problems in
elasticity. Math. Methods. Appl. Sci. 8, 236–257 (1986)
36. Stephan, E.P.: Boundary integral equations for screen problems in R3 . Integr. Equ. Oper. Theory
10, 257–263 (1987)
37. Toselli, A., Widlund, O.: Domain decomposition methods—algorithms and theory. In: Springer Series
in Computational Mathematics, vol. 34. Springer, Berlin (2005)
38. Tran, T., Stephan, E.P.: Additive Schwarz method for the h-version boundary element method. Appl.
Anal. 60, 63–84 (1996)

123