Galerkin Boundary Element Methods for

Electromagnetic Scattering

A. Buffa1 and R. Hiptmair2

1
Istituto di Matematica applicate e tecnologie informatiche del CNR, Pavia, Italy,
annalisa@ian.pv.cnr.it
2
Seminar für Angewandte Mathematik, ETH Zürich, CH-8092 Zürich,
hiptmair@sam.math.ethz.ch

in Computational Methods in Wave Propagation, M. Ainsworth, ed., Springer,

New York, 2003, pp. 85-126. In print
Extended version with new appendix on Scattering from coated dielectric
objects. Draft version December 12, 2007.

Summary. Methods based on boundary integral equations are widely used in the numeri-
cal simulation of electromagnetic scattering in the frequency domain. This article examines
a particular class of these methods, namely the Galerkin boundary element approach, from
a theoretical point of view. Emphasis is put on the fundamental differences between acous-
tic and electromagnetic scattering. The derivation of various boundary integral equations is
presented, properties of their discretized counterparts are discussed, and a-priori convergence
estimates for the boundary element solutions are rigorously established.

Key words: Electromagnetic scattering, boundary integral equations, boundary element

methods

1 Introduction
The numerical simulation of electromagnetic scattering aims at computing the inter-
action of electromagnetic waves with a physical body, the so-called scatterer. The
scatterer occupies a bounded domain Ωs in three-dimensional affine space R3 . In
general, Ωs will have Lipschitz-continuous boundary Γ := ∂Ωs [41, Section 1.2],
which can be equipped with an exterior unit normal vectorfield n ∈ L∞ (Γ ). With
boundary element methods in mind, we do not lose generality by considering only
piecewise smooth Ωs , i.e., curvilinear Lipschitz polyhedra in the parlance of [35].
We only consider linear materials and time-harmonic electromagnetic fields of
angular frequency ω > 0. Excitation is provided by the fields ei , hi of an incident
(plane) wave. Under these circumstances we can derive the following transmission
problem from Maxwell’s equations [31, Ch. 6]:
2 A. Buffa and R. Hiptmair

curl e = −iωµh , curl h = iωϵe in Ωs ∪ Ω ′ ,

γt+ e − γt− e = −γt+ ei , γt+ h − γt− h = −γt+ hi on Γ ,
! (1)
2
|γt h × n + γe| dS → 0 for r → ∞ .
∂Br

Here and below, Br denotes a ball of radius r centered in the origin, γt u stands
for the tangential trace u × n, and superscripts − and + tag traces onto Γ from
Ωs and Ω ′ := R3 \ Ω, respectively. The vectorfields e = e(x), h = h(x) repre-
sent the unknown complex amplitudes (phasors) of the electric and magnetic field,
respectively. The material parameters µ = µ(x) (permeability tensor), ϵ = ϵ(x)
(dielectric tensor), x ∈ R3 , are uniformly positive definite and bounded. In fact,
information on the scatterer is completely contained in µ and ϵ: inside Ωs they may
vary, but in the “air region” Ω ′ both material parameters agree with the constants
µ0 > 0 and ϵ0 , respectively. At ∞ the so-called Silver-Müller radiation conditions
are imposed.
This system of equations can always be reduced to a second order wave equation
in terms either of the electric or the magnetic field, e.g., e satisfies the electric wave
equation

curl µ−1 curl e − ω 2 ϵe = 0 in Ωs ∪ Ω ′ . (2)

Note that the uniqueness of solutions of the system (1) is a direct consequence of
Rellich’s Lemma [24, 53].
Apart from generic dielectric and even lossy scatterers the following special
situations are of practical interest.
– The scatterer is assumed to be a “perfect conductor” in which no electric field can
exist. This leads to an exterior Dirichlet problem for the electric wave equations
in Ω ′ , because the transmission conditions in (1) are replaced by the boundary
condition γt+ e = −γt+ ei on Γ for the electric field.
– If the scatterer is a thin perfectly conducting sheet, we arrive at a screen problem.
In this case Ωs = ∅ and Γ becomes a compact piecewise smooth two-dimensional
surface with boundary. As before, we demand γt e = −γt ei on both sides of
Γ . For screen problems Ω ′ does not possess a Lipschitz boundary any more.
Moreover, the screen Γ itself might not even be a Lipschitz surface itself, in case
it branches. The resulting mathematical complications are treated in [16].
– If the scatterer is a good conductor with smooth surface, its impact on the fields
can be modelled by impedance boundary conditions (Leontovich boundary con-
ditions) [3, 10, 54]

γt+ e − η(γt+ h × n) = η(γt+ hi × n) − γt+ ei on Γ .

The surface impedance η is a complex tensor with uniformly positive definite real
part and non-zero imaginary part.
 Galerkin BEM in Electromagnetics 3

All these problems have in common that scattered fields on the unbounded domain
Ω ′ have to be determined. As the material coefficients are constant in Ω ′ , bound-
ary integral equation methods are perfectly suited for this job. In addition, they are
posed on the two-dimensional surface Γ , which relieves us from meshing (a part of)
Ω ′ . In the case of complicated geometries this is a strong point in favour of bound-
ary integral equation methods, compared to volume based schemes with absorbing
boundary conditions (cf. the contribution of T. Hagstrom on absorbing layers and
radiation boundary conditions in this collection) at an artificial cut-off boundary.

Transmission problem for second order PDE, Equ. (1)

Representation formula for solutions involving potentials that take

jumps of Cauchy data as arguments (Sect. 4)

✓ ✏
Trace operators (Sect. 2)
+ Jump relations for potentials (Sect. 4)
✒ ✑

Calderón projector (Sect. 5)

Indirect BIE (Sect. 7.2)
Direct BIE (Sect. 7.1)

(Generalized) Gårding inequality for variational form of BIE

⇒ Existence of continuous solutions by Fredholm argument

✓ ✏
Conforming boundary element (BEM) space
+ based on a triangulation of the surface (Sect. 8)
✒ ✑

Discrete inf-sup-condition
⇒ existence, uniqueness, and asymptotically optimal conver-
gence of the discrete solutions (Sect. 9)

Fig. 1. “Road map” for the derivation and analysis of Galerkin boundary element methods
for electromagnetic scattering (BIE = boundary integral equation)
4 A. Buffa and R. Hiptmair

In this article we will exclusively deal with the Galerkin method for the dis-
cretization of the boundary integral equations. It is based on variational formula-
tions in suitable trace spaces. This permits us to use powerful tools from functional
analysis. They pave the way for a rigorous and comprehensive convergence theory.
We acknowledge that several other numerical methods based on boundary integral
formulations exist and are widely used alternatives to Galerkin schemes:
– the collocation method, which can be regarded as a special Petrov-Galerkin ap-
proach [42, Sect. 4.4]
– the method of source potentials, which requires a second surface away from Γ ,
on which a source distribution is sought. An example of this method for electro-
magnetic scattering is analysed in [43].
– Nyström methods, which directly tackle the boundary integral equations by
means of a quadrature rule. For an exposition we refer to the contribution of O.
Bruno in this volume and to [49, Ch. 12].
Unfortunately the theoretical understanding of these methods is rudimentary in com-
parison with Galerkin schemes. For this reason we us restrict the presentation to
Galerkin methods.
As far as Galerkin boundary element methods are concerned, there is a fairly
canonical approach to their construction and theoretical examination. This standard
procedure is depicted in the flowchart of Fig. 1. The plan of this paper closely fol-
lows these lines.
We point out that issues of implementation and efficient solution of the resulting
linear systems of equations are not covered by this article. We will also skip quite
a few proofs, which the reader may look up in the research papers that underly
this survey. In particular, we mention [15, 20] as main references for Sect. 2.1, [45,
Ch. 5] for Sect. 3, [31, Ch. 6] and [16, 21] as regards Sects. 4-9, and [44] as source
for Sect. 10.

2 Function Spaces and Traces

In order to write problem (1) or equivalent formulations of it in a mathematically
rigorous way, we need a precise characterization of the function spaces, on which
the equations are posed. This section is devoted to definitions and main properties
of function spaces which are concerned with the rigorous formulation of the prob-
lem (1). The first section concerns spaces on the domain, either Ωs , Ω ′ = R3 \Ωs or
R3 , while in the second we define and characterize suitable spaces on the manifold
Γ , which will be of key importance for the definition of integral operators.

2.1 Function spaces in the domain

Let Ω ⊆ R3 be any of the sets Ωs , Ω ′ , R3 and define the Fréchet space L2loc (Ω) of
complex, vector valued, locally square integrable functions u : Ω → C3 . We also
 Galerkin BEM in Electromagnetics 5

make use of the Sobolev spaces H sloc (Ω), s ≥ 0 with the convention H 0 ≡ L2
(see, e.g., [1] for definitions). The sub-fix loc is systematically removed when Ω
is bounded: in this case, the H s (Ω) are Hilbert spaces endowed with the natural
graph-norm ∥u∥H s (Ωs ) and semi-norm |u|H s (Ωs ) , respectively [1]. Round brackets
will consistently be used to express inner products.
With d a first order differential operator, we define for any s ≥ 0

H sloc (d, Ω) := {u ∈ H sloc (Ω) : du ∈ H sloc (Ω)} , (3)

H sloc (d0, Ω) := {u ∈ H sloc (Ω) : du = 0} . (4)

When s = 0, we simplify the notation by setting H 0 = H. If Ω is bounded,

H sloc (d, Ω) is endowed with the graph norm ∥·∥2H s (d,Ω) := ∥·∥2H s (Ω) +∥d·∥2H s (Ω)
and seminorm | · |2H s (d,Ω) := | · |2H s (Ω) + |d · |2H s (Ω) . This defines the spaces
H s (curl, Ω), H s (div, Ω) and H s (curl 0, Ω), H s (div 0, Ω).
¿From Gauß’ theorem we obtain integration by parts formulae for the spaces
H(curl, Ωs ), H(div, Ωs ). If u , v ∈ C ∞ (Ωs )3 and p ∈ C ∞ (Ωs ), then we have
div(up) = div u p + u · ∇p and div(u × v) = curl u · v − curl v · u, and, finally,
(u × v) · n = −(u × n) · v on the boundary Γ . These imply the following formulae:
! ! !
div(up) = (div u p + u · ∇p) dx = p u · ndS , (5)
Ωs Ωs Γ
! !
(u · curl v − curl u · v) dx = (u × n) · v|Γ dS . (6)
Ωs Γ

These formulae suggest the definitions of the mappings γt : u -→ u|Γ × n and

γn : u -→ u|Γ · n, u ∈ C ∞ (Ωs )3 .
The trace theorem for H 1 (Ω) [40, Theorem 1.5.1.1] shows that the tangential
trace γt : C ∞ (Ω̄) -→ L∞ (Γ ) and the normal trace: γn : C ∞ (Ω̄) -→ L∞ (Γ ) are
1 1
continuous as mappings H(curl; Ω) -→ H − 2 (Γ ) and H(div; Ω) -→ H − 2 (Γ ),
1 1 1
respectively. Here, H − 2 (Γ ) and H − 2 (Γ ) are the dual space of H 2 (Γ ) and
1 1
H 2 (Γ ) := (H 2 (Γ ))3 , respectively, with respect to the pivot spaces L2 (Γ )/L2 (Γ ).
Consequently, the traces can be extended to H(curl; Ω) and H(div; Ω), respec-
tively.
In the sequel we will consider the electric wave equation (2). Now, since the fact
that the field is a locally square-integrable function satisfying curl curl u − u = 0,
we can conclude that curl curl u is locally square-integrable, too. Hence, the space

H loc (curl2 , Ω) := {u ∈ H loc (curl; Ω), curl curl u ∈ L2loc (Ω)}

comes into play as the natural space for the solutions of the electric/magnetic wave
equation with constant coefficients. It will be crucial for meaningful strong formu-
lations of electromagnetic transmission problems.
6 A. Buffa and R. Hiptmair

2.2 Function spaces on the manifold Γ

Recall that through local charts one defines standard Sobolev spaces on the manifold
Γ = ∂Ωs . We denote them as H s (Γ ), Hs (Γ ), s ∈ [−1, 1], for scalars and vectors,
respectively. We saw that the tangential trace operator γt possesses an interpretation
1
as a continuous mapping H(curl; Ω) -→ H − 2 (Γ ). This is actually sufficient for
the understanding of homogeneous boundary conditions for fields in the Hilbert
space context. However, in order to impose meaningful non-homogeneous boundary
conditions or, even more important, to lay the foundations for boundary integral
equations we need to identify a proper trace space “X(Γ )” of H(curl; Ω), Ω ⊂ R3
a “generic” domain. It has to meet two essential requirements:
1. The inner product on X(Γ ) has an intrinsic definition that does not rely on the
embedding of Γ into R3 , i.e, X(Γ ) should have an interpretation as sections of
the tangent bundle to T Γ of Γ .
2. We demand that γt : H(curl; Ω) -→ X(Γ ) is continuous and surjective.
1
Note that the same issue for the operator γn : H(div; Ω) → H − 2 (Γ ) was resolved
a long time ago [39, Sect. I.2.2].
We emphasize that for the discussion of traces it hardly matters, whether Ω
is bounded or not. We assume in this section that Ω is bounded (this allows for
integration on Ω), but with this slight change, the results of this section remain true
also for unbounded domains, in particular, the open complement of Ω.

Smooth boundaries. To illustrate ideas, we first consider a C ∞ -smooth Γ . Then

the Sobolev spaces H s (Γ ) of functions and H st (Γ ) of tangential vector-fields,
as well as differential surface operators (we shall use the self evident notation
divΓ , curlΓ , curlΓ , . . . ) can be defined for all s ∈ R using local charts and trans-
formations [24, Sect. 3.1, Appendix] [53, Sect. 2.5.2]. It is a classical result that
smooth functions on Γ are dense in all these spaces. Standard trace and tangential
1
trace generate continuous and surjective operators γ : H s+ 2 (Ω) -→ H s (Γ ) and
1
γt : H s+ 2 (Ω) -→ H st (Γ ) for all s > 0, where

H st (Γ ) ∼
= {φ ∈ H s (Γ ), φ · n = 0} ⊂ L2t (Γ ) (7)

are Sobolev spaces of tangential vector-fields. We denote by H −st (Γ ) the dual space
of H st (Γ ) with L2t (Γ ) as a pivot space. Angle brackets will designate the duality
pairings. Now, since for any u ∈ H(curl; Ω) we have γt u · n = 0, thus γt u ∈
−1/2
Ht (Γ ). Moreover, using both (5) and (6), we can easily see that

divΓ (γt u) = γn (curl u) ∀ u ∈ H(curl; Ω), (8)

which implies divΓ (γt u) ∈ H −1/2 (Γ ).

Now, it is natural to define the space

T H s (divΓ ; Γ ) := {µ ∈ H st (Γ ) , divΓ µ ∈ H s (Γ )}. (9)

 Galerkin BEM in Electromagnetics 7
− 21
The tangential trace γt : H(curl; Ω) -→ T H (divΓ ; Γ ) turns out to be con-
tinuous and surjective [53, Theorem 5.4.2]. For smooth surfaces the issue of tan-
gential traces in H(curl; Ω) was investigated in the papers of L. Paquet [55] and
Alonso/Valli [2]. A survey of the results is also given in the monographs by M.
Cessenat [24, Sect. 2.1] and J.-C. Nédélec [53, Sect. 5.4.1].
Moreover, if we define the anti-symmetric pairing
!
⟨µ, η⟩τ ,Γ := (µ × n) · η dS , µ, η ∈ L2t (Γ ) , (10)
Γ

then we can rewrite (6) as

!
(curl u · v − u · curl v) dx = ⟨γt v, γt u⟩τ ,Γ , (11)
Ω

1
which suggests that the space T H − 2 (divΓ ; Γ ) coincides with its dual when using
⟨·, ·⟩τ ,Γ as duality pairing. This statement will be clarified in the case of non-smooth
surfaces at the end of this section.

Piecewise smooth and Lipschitz boundaries. Only recently results have been ob-
tained for non-smooth boundaries. We owe it to the pioneering work of one of the
authors together with P. Ciarlet jr., who first examined piecewise smooth boundaries
in [14, 17, 18]. The issue of traces of H(curl; Ω) for general Lipschitz-domains
was finally settled jointly by one of the authors, M. Costabel and D. Sheen in [20].
These articles and Sect. 2 of [19] supply the main references for the current section.
The challenges faced in the case of piecewise smooth boundaries are highlighted
1
by simple consideration: even if u ∈ C ∞ (Ω̄) we do not have γt u ∈ H 2 (Γ ),
because the tangential trace is inevitably discontinuous across edges of Γ . The first
consequence of this fact is that γt H 1 (Ω)×n ̸⊆ γt H 1 (Ω), although the two objects
are both good candidates to be “tangential” vector fields of “regularity” 21 . Thus we
have to resort to the following definition:
Definition 1. We introduce the Hilbert space H s× (Γ ) := γt (H s+1/2 (Ω)), s ∈
(0, 1), equipped with an inner product that renders γt : H s+1/2 (Ω) -→ H s× (Γ )
continuous and surjective. Its dual space with respect to the pairing ⟨·, ·⟩τ ,Γ is de-
noted by H −s
× (Γ ).
1
The dual space is well defined due to the density of H × 2
(Γ ) ⊂ L2t (Γ ). The case of
smooth and non-smooth surfaces differ considerably, which we aim to highlight by
different notations: H st (Γ ) for smooth Γ and H s× (Γ ) for non-smooth Γ .
For curvilinear polyhedra this space can be given a more concrete meaning. To
that end, write"Γ 1 , . . . , Γ P , P ∈ N, for the finitely many curved polygonal faces of
P
Γ , i.e. Γ := j=1 Γ̄ j , meeting at non-degenerate edges. For any tangential vector
µ, we denote by µj the restriction of µ to Γ j . Then, according to [17, Proposi-
1
tion 1.6] an equivalent norm on H × 2
(Γ ) can be expressed as
8 A. Buffa and R. Hiptmair
P
# P #
# !
2 |µi · ν ij (x) − µj · ν ji (y)|2 2
∥µ∥ 1 := ∥µj ∥2 1 + dS .
2 (Γ )
H×
j=1
H t2 (Γ j )
j=1 i∈Ij
|x − y|3
Γ j ×Γ i

where Ij is the set of indices of smooth components abutting Γ j , and ν ij denotes

the tangential outer normal to Γ i restricted to the edge Γ j ∩ Γ i . Loosely speaking,
1 1
H× 2
(Γ ) contains vector-fields that are in H t2 (Γ j ) for each face Γ j and feature a
“weak normal continuity” enforced by the second term in the definition of the norm.
Using (6) and the same reasoning as for regular surfaces, we have that γt :
−1
H(curl; Ω) → H × 2 (Γ ) is linear and continuous. In view of (8), we also know that
this operator does not admit a right inverse. In order to repeat the argument sketched
above for regular domains, we need a theory of differential operators on non-smooth
manifolds. We do not want to delve into the details of these developments, and we
refer the reader to [14, 16, 20] for a discussion on the subject. We need only the
following definition: for u ∈ C ∞ (Ω) set
$
divj (γt u)j on Γ j ,
divΓ γt u := % & (12)
(γt u)j · ν ij + (γt u)i · ν ji δij on Γ j ∩ Γ i ;

where δij is the delta distribution (in local coordinates) whose support is the edge
Γ j ∩ Γ i and divj denotes the 2D-divergence computed on the face Γ j . By density,
this differential operator can be extended to less regular distributions and, in partic-
−1
ular, to functionals in H × 2 (Γ ). Moreover, (8) holds true in the appropriate sense.
Thus, we set
−1 −1 1
H × 2 (divΓ , Γ ) := {µ ∈ H × 2 (Γ ), divΓ µ ∈ H − 2 (Γ )} .

Finally, we denote by curlΓ the operator adjoint to divΓ with respect to the scalar
product ⟨·, ·⟩τ ,Γ , i.e.,

−1 1
⟨curlΓ q, p⟩τ ,Γ = ⟨divΓ p, q⟩ 21 ,Γ , p ∈ H × 2 (divΓ , Γ ) , q ∈ H 2 (Γ ) . (13)

−1
The following theorem proves that the space H × 2 (divΓ , Γ ) fits the criterion
announced at the beginning of this section:
−1
Theorem 1. The operator γt : H(curl; Ω) -→ H × 2 (divΓ , Γ ) is continuous, sur-
jective, and possesses a continuous right inverse.
Proof. See Theorem 4.4 in [18] for the case of Lipschitz polyhedra. The more gen-
eral assertion for Lipschitz domains is shown in [20, Sect. 4].
In the case of Maxwell’s equations the role of Cauchy data is played by γt e and
−1
γt h. By the fundamental symmetry of electric and magnetic field, H × 2 (divΓ , Γ )
is the right trace space for both fields. Everything is fitting, because this space is its
own dual, as is confirmed by the following theorem [20, Lemma 5.6]
 Galerkin BEM in Electromagnetics 9
− 21
Theorem 2 (Self-duality of H × (divΓ , Γ )). The pairing ⟨·, ·⟩τ ,Γ can be extended
−1
to a continuous bilinear form on H × 2 (divΓ , Γ ). With respect to ⟨·, ·⟩τ ,Γ the space
−1
H × 2 (divΓ , Γ ) becomes its own dual.
When we want to examine the convergence of boundary element methods quan-
titatively, extra smoothness of the functions to be approximated is indispensable.
For any s > 21 , we define Hs− (Γ ) := {u ∈ L2t (Γ ) : u|Γ j ∈ H st (Γ j )} and
1
H s× (Γ ) := H ×2
(Γ ) ∩ Hs− (Γ ). The corresponding space of scalar functions will
s
be denoted by H− (Γ ). To characterize the smoothness we resort to the family of
Hilbert spaces
⎧ −1
⎪
⎨H × (divΓ , Γ ) ,
2
if s = − 21 ,
s s
H × (divΓ , Γ ) := {µ ∈ H × (Γ ), divΓ µ ∈ H s (Γ )} , if − 12 < s < 21 ,
⎪
⎩
{µ ∈ H s× (Γ ), divΓ µ ∈ H−s
(Γ )} , if s > 12 .

As demonstrated in [16, Appendix 2], these spaces can be obtained through complex
interpolation for − 12 ≤ s < 12 . From this fact we conclude the following trace
theorem (see [16]).
Theorem 3. The tangential trace mapping γt can be extended to a continuous map-
s− 1
ping γt : H s (curl, Ω) -→ H × 2 (divΓ , Γ ) for all 0 ≤ s < 1.

3 Maxwell versus Helmholtz

There is a striking similarity between the electric wave equation (2) and the scalar
Helmholtz equation

− div(µ−1 grad p) − ω 2 ϵp = 0 , in Ωs ∪ Ω ′ . (14)

In fact, the relationship between (2) and (14) runs much deeper than mere appear-
ance: both equations emerge from a single equation for differential forms on R3 ,
where (14) involves 0-forms, whereas (2) is the version for 1-forms [45, Sect. 2].
Hardly surprising, the theories of boundary integral equation methods for the re-
lated boundary value problems largely rely on the same principles. Nevertheless,
the technical difficulties encountered in the treatment of the electric wave equation
and related boundary element methods are significantly bigger than in the case of
(14).
To appreciate what accounts for the fundamental difference between electro-
magnetism and acoustics, let us temporarily consider the variational source problem
in a bounded Lipschitz domain Ω, cf. [45, Sect. 5]. For (2) this reads: for j ∈ L2 (Ω)
find e ∈ H(curl; Ω) such that for all v ∈ H(curl; Ω)
% &
aM (e, v) := µ−1 curl e, curl v 0 − ω 2 (ϵe, v)0 = −iω (j, v)0 , (15)
10 A. Buffa and R. Hiptmair
+
where (u, v)0 := Ω uv dx. The related problem for the Helmholtz equation and
f ∈ L2 (Ω) seeks p ∈ H 1 (Ω) such that

aH (p, q) := (grad p, grad q)0 − ω 2 (p, q)0 = (f, q)0 ∀q ∈ H 1 (Ω) . (16)

Investigations of the convergence of Galerkin schemes for (16) usually centre on the
concept of coercivity of the underlying bilinear form aH (·, ·), that is, the fact that
the zero order term is a compact perturbation of the second order term, the principal
part, and that a Gårding inequality of the form
2
|aH (p, p) + cH (p, p)| ≥ C ∥p∥H 1 (Ω) ∀p ∈ H 1 (Ω) (17)

holds with C > 0 and a bilinear form cH (··), which is compact in H 1 (Ω). As has
been demonstrated by Schatz [58], cf.also [62], this is the key to a priori asymptotic
error estimates for Galerkin finite element methods. Evidently, we cannot expect an
analogue of (17) from aM . The blame lies with the infinite dimensional kernel of the
curl-operator, which foils compactness of the imbedding H(curl; Ω) (→ L2 (Ω).
The issue of coercivity can also be discussed from the point of view of “en-
ergies”: both acoustic and electromagnetic scattering are marked by an incessant
conversion of energies. In acoustics, potential and kinetic energy of the fluid are
converted into each other, in electromagnetism the same roles are played by the
electric and magnetic energy. In acoustics the potential energy (with respect to the
bounded control volume Ω) is a compact perturbation of the kinetic energy1. There-
fore we can clearly single out the Laplacian as the principal part of the Helmholtz
operator. Conversely, in electromagnetism the electric and magnetic energies of a
field are perfectly symmetric. Neither is a compact perturbation of the other. This
means that no part of the electric wave equation is “principal”. Formally speaking,
the operator of the electric wave equation lacks the essential property of strong el-
lipticity. A concise summary is given in Table 1.

Table 1. Acoustics vs. electromagnetics in terms of dominant energies

Acoustic wave equation Electric wave equation

−∆p − κ2 p = 0 curl curl e − κ2 e = 0

Energies entering the Lagrangian:

Kinetic “energy” Ω |Rgrad p|2 dx Magnetic “energy” ΩR | curl e|2 dx

R R

Potential “energy” Ω |p|2 dx Electric “energy” Ω |e|2 dx

Potential energy a is compact Symmetry between electric

perturbation of kinetic energy and magnetic quantities

1
Roles might be reversed depending on the formulation of the acoustic equations.
 Galerkin BEM in Electromagnetics 11

The lack of a principal part can be overcome by the splitting of the fields into
two components. One set of components, called the electric, will feature dominant
electric energy. With the other set, the magnetic quantities, the situation is reversed.
This will promote either curl curl or Id to the role of a principal part. As a con-
sequence, on each component the electric wave equation should be amenable to
the same treatment as the Helmholtz equation. In the context of electromagnetic
problems the splitting idea has been pioneered by Nédélec and was first applied to
integral operators in [37]. Since then it has emerged as a very powerful theoreti-
cal tool, see [6, 19, 26] and, in particular, the monograph [53]. Three features of a
splitting prove essential:
1. one subspace in the splitting agrees with the kernel of curl,
2. the compact embedding of the other subspace (complement space) into L2 (Ω),
3. the extra smoothness of vector-fields in the complement space.
This makes it possible to opt for the Helmholtz-type regular splitting provided by
the next lemma.
Lemma 1 (Regular decomposition lemma). There exists a continuous projector
R : H(curl; Ω) -→ H 1 (Ω) ∩ H(div 0; Ω) such that Ker(R) = H(curl 0; Ω).
The proof is given in [45, Sect. 2.4] and makes use of the existence of regular vector
potentials, cf. Lemma 3.5 in [7].
Evidently, the three requirements are satisfied by the decomposition

H(curl; Ω) = X (Ω) ⊕ N (Ω) , X (Ω) := R(H(curl; Ω)) ⊂ H 1 (Ω) , (18)

where we write N (Ω) := H(curl 0; Ω). The continuity of the projectors guaran-
tees the stability of this decomposition. Now, we can consider the variational prob-
lem (15) with respect to (18): thanks to the compact embedding of H 1 (Ω) into
L2 (Ω) we see that the second term of the bilinear form
% &
(e⊥ , v⊥ ) -→ µ−1 curl e⊥ , curl v⊥ 0 − ω 2 (ϵe⊥ , v⊥ )0 , e⊥ , v⊥ ∈ X (Ω) ,

is a compact perturbation of the first: aM is coercive on X (Ω). Coercivity on N (Ω)

is trivial. In addition, terms like

(e0 , v⊥ ) -→ (ϵe0 , v⊥ )0 , e0 ∈ N (Ω), v⊥ ∈ X (Ω) ,

which effect the coupling of X (Ω) and N (Ω) with respect to aM , can also be
dismissed as compact perturbations. Using the isomorphism XΩ : H(curl; Ω) -→
H(curl; Ω), defined by XΩ := R − Z, where Z := Id −R is the complementary
projector to R, to “flip signs”, we arrive at
2
|aM (u, XΩ u) − cM (u, u)| ≥ C ∥u∥H(curl;Ω) ∀u ∈ H(curl; Ω) , (19)

with some C > 0 and a compact bilinear form cM on H(curl; Ω). The general-
ized Gårding inequality (19) is the crucial assumption in the following fundamental
theorem:
12 A. Buffa and R. Hiptmair

Theorem 4. If a bilinear form a : V × V -→ C on a reflexive Banach space V

satisfies
2
|a(u, XΩ u) − c(u, u)| ≥ C ∥u∥V ∀u ∈ V ,

with C > 0, a compact bilinear form c : V × V -→ C, and an isomorphism

XΩ : V -→ V , then the associated operator A : V -→ V ′ is Fredholm with index 0.
In particular, for a bilinear form meeting the requirements of the theorem, injectivity
of the associated operator implies its surjectivity by the Fredholm alternative [11].
It is hardly surprising that the splitting idea also plays a pivotal role in the anal-
ysis of boundary integral equations arising from the electric wave equation. Here, it
−1
is applied to the trace space H × 2 (divΓ , Γ ):
−1 1
Lemma 2. There exists a projection RΓ : H × 2 (divΓ , Γ ) -→ H ×
2
(Γ ) such that
−1
Ker(RΓ ) = H × 2 (divΓ 0, Γ ) and

∥RΓ µ∥ 1 ≤ C∥divΓ µ∥ 1 . (20)

2 (Γ )
H× H − 2 (Γ )

−1 1
Proof. Pick λ ∈ H × 2 (divΓ , Γ ) and set µ := divΓ λ ∈ H − 2 (Γ ). Solve the Neu-
mann problem

w ∈ H 1 (Ωs )/R : ∆w = 0 in Ωs , γn− grad w = µ on Γ .

We find that v := grad w ∈ H(div 0; Ωs ). Using Lemma 3.5 in [7], there exists
w ∈ H 1 (Ωs ) such that v = curl w, div w = 0. This defines an operator J :
1
H − 2 (Γ ) -→ H 1 (Ωs ) by Jµ := w. Its continuity is elementary

∥J µ∥H 1 (Ωs ) ≤ C ∥v∥L2 (Ωs ) ≤ C ∥µ∥ 1 ,

H − 2 (Γ )

−1 1
and inherited by the mapping RΓ := γt ◦ J ◦ divΓ : H × 2 (divΓ , Γ ) -→ H ×
2
(Γ ).
Γ − −
Moreover, we see that divΓ R λ = γn curl Lv = γn v = divΓ λ . ⊓
9

As before, the projector complementary to RΓ will be denoted by ZΓ . We arrive at

a stable decomposition of the trace space
−1
H × 2 (divΓ , Γ ) := X (Γ ) ⊕ N (Γ ) , (21)

−1 −1
where X (Γ ) := RΓ (H × 2 (divΓ , Γ )) and N (Γ ) = H × 2 (divΓ 0, Γ ). Both com-
−1
ponents inherit the norm of H × 2 (divΓ , Γ ).

Corollary 1. The embedding X (Γ ) (→ L2t (Γ ) is compact.

 Galerkin BEM in Electromagnetics 13

It is illuminating to give a physical interpretation of the decomposition. First, view

−1
H × 2 (divΓ , Γ ) as a space of tangential components of electric fields. Then, we en-
counter traces of “static” irrotational fields in N (Γ ), whereas traces of “dynamic”
field components, whose curls do not vanish, are associated with X (Γ ). By Fara-
day’s law the latter are linked with magnetic fields. All in all, we can attribute an
“electric nature” to the space N (Γ ), and a “magnetic nature” to X (Γ ). The argu-
ments are simply reversed when considering magnetic traces γN e, because given the
absence of source currents the magnetic field is irrotational in the stationary case.
This means that components of γN e that belong to N (Γ ) are “magnetic”, whereas
components in X (Γ ) are “electric”, cf. Table 2.

Table 2. Physical nature of components occurring in the splitting of fields and traces

Magnetic Electric
Field Space
components components
e H (curl; Ω) X (Ω) N (Ω)
−1
γt e H × 2 (divΓ , Γ ) X (Γ ) N (Γ )
−1
γN e H × 2 (divΓ , Γ ) N (Γ ) X (Γ )

4 Representation Formulas
In this section we start from the electric wave equation (2) in the air region Ω ′ , where
µ and ϵ can be regarded as scalar constants µ0 and ϵ0 . Then, the partial differential
equation (2) can be recast as

curl curl e − κ2 e = 0 . (22)

√
The constant κ := ω ϵ0 µ0 > 0 is called the wave number, because κ/2π tells
us the number of wavelengths per unit length. Henceforth, κ will stand for a fixed
positive wave number2.
Definition 2. A distribution e ∈ H loc (curl2 , Ω) is called a Maxwell solution on
some generic domain Ω, if it satisfies (22) in Ω, and the Silver–Müller radiation
conditions at ∞, if Ω is not bounded.
It is our objective to derive a boundary integral representation formula for Maxwell
solutions. In order to handle transmission conditions in the calculus of distributions,
we introduce currents, that is, distributions supported on Γ . For a function ϕ ∈
1
H − 2 (Γ ), a tangential vector-field ξ ∈ H −1
× (Γ ), and test functions Φ ∈ D(R ),
3
3 3 3
Φ ∈ D(R ) := (D(R )) , we define
2
We point out that all considerations remain true if Im κ > 0 (lossy media)
14 A. Buffa and R. Hiptmair

(ϕδΓ )(Φ) := ⟨ϕ, γΦ⟩ 1 ,Γ , (ξδ Γ )(Φ) := ⟨ξ, γt Φ⟩τ ,Γ = ⟨ξ, γΦ⟩−1,Γ .
2

Recall the notation used in the introduction: the superscripts − and + tag traces onto
Γ from Ωs and Ω ′ respectively. Now, in the sense of distributions, integration by
parts yields, cf. [16, Section 2.3],

for u ∈ H loc (div; Ωs ∪ Ω ′ ) : div u = div u|Ωs ∪Ω ′ + [γn ]Γ (u) δΓ ,

for u ∈ H loc (curl; Ωs ∪ Ω ′ ) : curl u = curl u|Ωs ∪Ω ′ − [γt ]Γ (u) δ Γ ,
−1
for ξ ∈ H × 2 (divΓ , Γ ) : div(ξ δ Γ ) = (divΓ ξ) δΓ .

For the sake of brevity, we have used the jump operator [·]Γ defined by [γ]Γ :=
γ + − γ − for some trace γ onto Γ . For notational simplicity it is also useful to resort
to the average {γ}Γ = 12 (γ + +γ − ). Both operators can only be applied to functions
±
defined in Ωs ∪ Ω ′ . Moreover, we set γN := κ−1 γt± ◦ curl.
Now, let u be a Maxwell solution in Ωs ∪ Ω ′ , which, of course, satisfies
div u = 0 in Ωs ∪ Ω ′ . Then the following identity holds in the sense of distribu-
tions,

− ∆u − κ2 u = curl curl u − grad div u − κ2 u

% &
= curl curl u|Ωs ∪Ω ′ − [γt ]Γ (u) δ Γ − grad ([γn ]Γ (u) δΓ ) − κ2 u
= curl curl u|Ωs ∪Ω ′ − κ [γN ]Γ (u) δ Γ − curl([γt ]Γ (u) δ Γ )−
− grad([γn ]Γ (u) δΓ ) − κ2 u
= −κ [γN ]Γ (u) δ Γ − curl([γt ]Γ (u) δ Γ ) − grad([γn ]Γ (u) δΓ ) .

As far as the differential operator curl curl −κ2 Id is concerned, the integration by
parts formula (11) suggests the distinction between Dirichlet trace γt and Neumann
trace γN . The trace γN can be labelled “magnetic”, because it actually retrieves
the tangential trace of the magnetic field solution. It has much in common with the
Neumann trace operator γn ◦ grad for the Helmholtz equation: for instance, it fails
to be defined on H loc (curl; Ωs ∪ Ω ′ ), but the weak definition
!
1
− curl u · curl v − curl curl u · v dx = ⟨γN u, γt v⟩τ ,Γ , (23)
κ Ω

v ∈ D(R3 ), renders it meaningful on H loc (curl2 , Ωs ∪ Ω ′ ) [46, Lemma 3.3]:

Lemma 3. The trace γN furnishes a continuous and surjective mapping γN :
−1
H loc (curl2 , Ω ′ ∪ Ωs ) -→ H × 2 (divΓ , Γ ).
−1 −1
Definition 3. Pairs (ζ, µ) ∈ H × 2 (divΓ , Γ ) × H × 2 (divΓ , Γ ) are called inte-
rior/exterior Maxwell Cauchy data, if there is a Maxwell solution u in Ωs and Ω ′ ,
respectively, such that ζ = γt± u, µ = γN
±
u.

We know from [31, Theorem 6.7] that the Cartesian components of Maxwell
solutions will satisfy the Sommerfeld radiation condition and the scalar Helmholtz
 Galerkin BEM in Electromagnetics 15

equation in Ωs ∪ Ω ′ . Using the results from [52, Ch. 9], we can apply component-
wise convolution with the outgoing fundamental solution of the Helmholtz equation
Eκ (x) := exp(iκ|x|)/4π|x|, x ̸= 0, and we find that almost everywhere in R3 the
components of u = (u1 , u2 , u3 )T satisfy

uj (x) = − κ([γN ]Γ (u) δ Γ )(Eκ (x − ·)ej ) − ([γt ]Γ (u) δ Γ )(curl(Eκ (x − ·)ej ))+
+ ([γn ]Γ (u) δΓ )(div(Eκ (x − ·)ej )) , j = 1, 2, 3 .

Using gradx Eκ (x − y) = − grady Eκ (x − y), we arrive at the famous Stratton–

Chu representation formula for the electric field in Ωs ∪ Ω ′ [60], cf.[31, Sect. 6.2],
[53, Sect. 5.5], [24, Ch. 3, Sect. 1.3.2],

u = −κΨ κV ([γN ]Γ (u)) − curl Ψ κV ([γt ]Γ (u)) − grad ΨVκ ([γn ]Γ (u)) . (24)

Here, Ψ κV (·) and ΨVκ (·) are potentials, that is, mappings of boundary data to ana-
lytic functions defined everywhere off the boundary. In detail, ΨVκ and Ψ κV are the
scalar and vectorial single layer potential, whose integral representation is given by
(x ̸∈ Γ )
! !
ΨVκ (φ)(x) := φ(y)Eκ (x − y) dS(y), Ψ κV (µ)(x) := µ(y)Eκ (x − y) dS(y).
Γ Γ

A simplification of (24) is possible by observing that, by (8)

1
±
divΓ (γN u) = κ−1 γn± (curl curl u) = κ(γn± u) in H − 2 (Γ ) . (25)

This enables us to get rid of the normal components trace in (24). We end up with
the pointwise identity

u(x) = −Ψ κDL ([γt ]Γ (u))(x) − Ψ κSL ([γN ]Γ (u))(x) , x ∈ Ωs ∪ Ω ′ , (26)

where we have introduced the (electric) Maxwell single layer potential according to
1
Ψ κSL (µ)(x) := κΨ κV (µ)(x) + gradx ΨVκ (divΓ µ)(x) , x ̸∈ Γ , (27)
κ
and the (electric) Maxwell double layer potential

Ψ κDL (µ)(x) := curlx Ψ κV (µ)(x) , x ̸∈ Γ . (28)

We have chosen these names in order to underscore the similarity of (26) with the
representation formula for solutions of the Helmholtz equation [31, Sect. 3.1], [52,
Ch. 9].
Next, we aim to fit the potentials into the functional framework devised in
Sect. 2. To this end we have to show that the potentials Ψ κDL and Ψ κSL are continu-
ous operators between the canonical function spaces for traces and the appropriate
function spaces for Maxwell solutions. An important result from [19] and [33] will
be useful
16 A. Buffa and R. Hiptmair

Lemma 4. The single layer potentials ΨVκ and Ψ κV give rise to continuous mappings
1 − 1 +s
1+s
ΨVκ : H − 2 +s (Γ ) -→ Hloc (R3 ) , Ψ κV : H × 2 (Γ ) -→ H 1+s 3
loc (R ) ,

for any s, − 21 < s ≤ 1

2 .
¿From this lemma we conclude that both Ψ κSL and Ψ κDL are well defined for
−1
arguments in the trace space H × 2 (divΓ , Γ ). To gain deeper insights into the con-
tinuity property of the Maxwell single layer and double layer potentials, we have to
make some preparations, cf. Lemma 2.3 in [51].
−1
Lemma 5. For µ ∈ H × 2 (divΓ , Γ ) we have div Ψ κV (µ) = ΨVκ (divΓ µ) in L2 (R3 ).
By definition and curl ◦ grad = 0, it is immediate that curl ◦Ψ κSL = κΨ κDL
−1
on H × 2 (divΓ , Γ ). On the other hand, using the previous lemma, we get for µ ∈
−1
H × 2 (divΓ , Γ ),

curl Ψ κDL (µ) = curl curl Ψ κV (µ) = (−∆ + grad div)Ψ κV (µ)
= κ2 Ψ κV (µ) + grad ΨVκ (divΓ µ) = κΨ κSL (µ) .

Here, we have used −∆Ψ κV (µ) = κ2 Ψ κV (µ). Altogether, both potentials are
−1
Maxwell solutions, that is, for µ ∈ H × 2 (divΓ , Γ ) they fulfil

(curl curl −κ2 Id)Ψ κSL (µ) = 0 , (curl curl −κ2 Id)Ψ κDL (µ) = 0 , (29)

off the boundary Γ in a pointwise sense, and, globally, in L2loc (R3 ). In addition,
they comply with the Silver–Müller radiation conditions. From these relationships
and Lemma 4 we infer the desired continuity properties.
Theorem 5. The following mappings are continuous
−1
Ψ κSL : H × 2 (divΓ , Γ ) -→ H loc (curl2 , Ωs ∪ Ω ′ ) ∩ H loc (div 0; Ωs ∪ Ω ′ ) ,
−1
Ψ κDL : H × 2 (divΓ , Γ ) -→ H loc (curl2 , Ωs ∪ Ω ′ ) ∩ H loc (div 0; Ωs ∪ Ω ′ ) .

Now, we are in a position to extract the desired identities from (26).

Theorem 6 (Stratton-Chu representation formula). Any Maxwell solution u in
Ωs possesses the representation

u = Ψ κDL (γt− u) + Ψ κSL (γN

−
u) in H(curl2 , Ωs ) .

If u is a Maxwell solution in Ω ′ that satisfies the Silver–Müller radiation conditions,

it can be written as

u = −Ψ κDL (γt+ u) − Ψ κSL (γN

+
u) in H loc (curl2 , Ω ′ ) .
 Galerkin BEM in Electromagnetics 17

5 Boundary Integral Operators

By Lemma 3, Theorem 5 provides the foundation for applying both the Dirichlet
trace γt and the Neumann trace γN to the potentials Ψ κSL and Ψ κDL . This is the
canonical way of constructing boundary integral operators [52, Chapter 7]. In the
case of second order elliptic problems, four different boundary integral operators
arise. Yet, due to the fact that curl ◦Ψ κSL = κ Ψ κDL , curl ◦Ψ κDL = κ Ψ κSL implies
± κ
γN Ψ DL = γt± Ψ κSL ,
± κ
Ψ SL = γt± Ψ κDL , γN (30)
two different boundary integral operators are sufficient for electromagnetic scatter-
ing: we obtain the boundary integral operators
Sκ := {γt }Γ ◦ Ψ κSL = {γN }Γ ◦ Ψ κDL , Cκ := {γt }Γ ◦ Ψ κDL = {γN }Γ ◦ Ψ κSL .
The continuity of Sκ and Cκ is immediate from Theorem 5, in conjunction with
Lemma 3 and Theorem 1.
−1 −1
Corollary 2. The operators Sκ , Cκ : H × 2 (divΓ , Γ ) -→ H × 2 (divΓ , Γ ) are con-
tinuous.
As auxiliary boundary integral operators, which supply building blocks for Sκ and
Cκ , we introduce the two single layer boundary integral operators
Vκ := {γ}Γ ◦ ΨVκ , Vκ := {γt }Γ ◦ Ψ κV .
By combining Lemma 4 with continuity properties of the traces, we obtain the fol-
lowing result
1 1
Corollary 3. The boundary integral operators Vκ : H − 2 (Γ ) -→ H 2 (Γ ) and Vκ :
− 12 1
H × (Γ ) -→ H × (Γ ) are continuous.
2

By inspecting the potential Ψ κSL , and recalling γt ◦ grad = curlΓ ◦ γ, it is clear

that we can write
Sκ = κVκ + κ−1 curlΓ ◦ Vκ ◦ divΓ . (31)
For the sake of implementation, more concrete boundary integral representations of
the boundary integral operators are indispensable. It takes subtle theory to establish
them, but here we only cite the result. A comprehensive treatment for second-order
elliptic operators is given in [52, Section 7.2]. As variational formulations are our
primary concern, expressions for the bilinear forms associated with Sκ and Cκ will
be given: for tangential vectorfields µ, ξ ∈ L∞ (Γ ) we obtain
! !
⟨Sκ µ, ξ⟩τ ,Γ = − κ Eκ (x − y) µ(y) · ξ(x) dS(y, x) + (32)
Γ Γ
! !
1
+ Eκ (x − y) divΓ µ(y) divΓ ξ(x) dS(y, x) ,
κ
! Γ! Γ
⟨Cκ µ, ξ⟩τ ,Γ =− gradx Eκ (x − y) · (µ(y) × ξ(x)) dS(y, x) (33)
Γ Γ
18 A. Buffa and R. Hiptmair

The first integral arises from (31) through integration by parts. Its kernel Eκ (x − y)
is weakly singular, because Eκ (x − y) = O(|x − x|−1 ) for y → x. Thus, the
integral makes sense as an improper integral. The second integral has a strongly
singular kernel behaving like O(|x − y|−2 ) for y → x, and has to be read as a
Cauchy principal value.
A fundamental tool for deriving boundary integral equations are jump relations
describing the behavior of the potentials when crossing Γ . For the Maxwell sin-
gle and double layer potential they closely resemble those for conventional single
and double layer potentials for second order elliptic operators [52, Chapter 6]. For
smooth domains these results are contained in [31, Thm. 6.11], [53, Thm. 5.5.1],
and [57].
Theorem 7 (Jump relations). The interior and exterior Dirichlet- and Neumann-
−1
traces of the potentials Ψ κSL and Ψ κDL are well defined and, on H × 2 (divΓ , Γ ),
satisfy
[γt ]Γ ◦ Ψ κSL = [γN ]Γ ◦ Ψ κDL = 0 , [γN ]Γ ◦ Ψ κSL = [γt ]Γ ◦ Ψ κDL = − Id .
Proof. The jump condition for the Dirichlet trace of the single layer potential is
immediate from its regularity asserted in Lemma 4. By (30) we get the continuity
of the Neumann trace Ψ κDL . Then, the jump of the Neumann trace of Ψ κSL can
be determined from (26). Finally, by (30), this also settles the contention for the
Dirichlet trace of the double layer potential. ⊓
9
Now, with the jump relations in mind, let us apply the exterior and interior trace
operators to the representation formulae of Theorem 6:
γt− u = 21 γt− u+Cκ(γt− u)+Sκ (γN
−
u) , γt+ u = 12 γt+ u−Cκ (γt+ u)−Sκ (γN
+
u) ,
−
γN −
u = Sκ (γt− u)+ 21 γN −
u+Cκ (γN +
u), γN u = −Sκ (γt+ u)+ 21 γN
+ +
u−Cκ (γN u).
A concise way to write these formulae relies on the Calderon projectors, cf. [21,
Section 3.3], [37, Formula (29)], and [53, Sect. 5.5],
,1 - ,1 -
− 2 Id +Cκ Sκ + 2 Id −Cκ −Sκ
Pκ := 1 , Pκ := 1 . (34)
Sκ 2 Id +Cκ −Sκ 2 Id −Cκ

−1 −1
By Theorem 6 the operators P− + 2 2 2 2
κ , Pκ : H × (divΓ , Γ ) -→ H × (divΓ , Γ ) are
projectors, that is,
P− − −
κ ◦ Pκ = Pκ , P+ + +
κ ◦ Pκ = Pκ . (35)
Also note that P− + +
κ + Pκ = Id and that the range of Pκ coincides with the kernel of
−
Pκ and vice versa. The next result promotes Calderon projectors to a pivotal role in
the derivation of boundary integral equations, cf. [61, Thm. 3.7].
−1 −1
Theorem 8. The pair of functions (ζ, µ) ∈ H × 2 (divΓ , Γ ) × H × 2 (divΓ , Γ ) are
suitable interior or exterior Maxwell Cauchy data, if and only if they lie in the kernel
of P+ −
κ or Pκ , respectively.
 Galerkin BEM in Electromagnetics 19

For the subsequent analysis it is convenient to examine the operator

, -
Cκ Sκ −1 −1
Aκ := : H × 2 (divΓ , Γ )2 -→ H × 2 (divΓ , Γ )2 .
Sκ Cκ
1 1
It is linked with the Calderon projectors by P− +
κ = 2 Id +Aκ , Pκ = 2 Id −Aκ .
The operators Cκ enjoy a hidden symmetry, made precise in the next lemma,
see [21, Thm. 3.9].
−1
Lemma 6. We have ⟨Cκ ζ, µ⟩τ ,Γ = ⟨Cκ µ, ζ⟩τ ,Γ for all ζ, µ ∈ H × 2 (divΓ , Γ ).

6 Compactness and Coercivity

The ultimate goal is to establish the coercivity of bilinear forms occuring in weak
formulations of boundary integral equations. To achieve this we need to identify
compact perturbations, cf. Lemma 3.2 of [47] and the proof of Thm. 3.12 in [21].
1 1
Lemma 7. The integral operators δVκ := Vκ − V0 : H − 2 (Γ ) -→ H 2 (Γ ) and
− 21 1
δVκ := Vκ − V0 : H × (Γ ) -→ H × (Γ ) are compact.
2

Slightly abusing notation, we define

S0 := κV0 + κ−1 curlΓ ◦ V0 ◦ divΓ . (36)

¿From Lemma 7 and (31) we find that switching from Sκ to S0 amounts to a com-
pact perturbation.
−1 −1
Corollary 4. The operator Sκ − S0 : H × 2 (divΓ , Γ ) -→ H × 2 (divΓ , Γ ) is com-
pact.
The significance of this can be appreciated in light of the following result, cf. Thm. 3
in [36, Vol. IV, Ch. XI, § 2], and Thm. 6.2 in [46].
Lemma 8 (Ellipticity of single layer potentials). The operators V0 and V0 are
continuous, selfadjoint with respect to the bilinear pairings ⟨·, ·⟩ 1 ,Γ and ⟨·, ·⟩τ ,Γ ,
2
respectively, and satisfy
2 1
⟨µ, V0 µ⟩ 1 ,Γ ≥ C ∥µ∥ 1 ∀µ ∈ H − 2 (Γ ) ,
2 H − 2 (Γ )
2 −1
⟨µ, V0 µ⟩τ ,Γ ≥ C ∥µ∥ −1 ∀µ ∈ H × 2 (divΓ 0, Γ ) .
H× 2 (Γ )

with constants C > 0 only depending on Γ .

Again, it proves highly instructive to remember facts about boundary integral
operators related to the Helmholtz equation: Table 3 lists similarities and differ-
ences of the situations faced in the case of the Helmholtz equation and Maxwell’s
equations, respectively. The lack of ellipticity of the off-diagonal operators in Aκ
20 A. Buffa and R. Hiptmair

Table 3. Comparison of analytical aspects of the acoustic and electromagnetic boundary inte-
gral operators, supplementing Table 1. The symbols Kκ and Dκ denote the double layer inte-
gral operator, and the hypersingular integral operator for the Helmholtz operator −∆−κ2 Id,
respectively. Details can be found in [52, Ch. 9].

Helmholtz equation Maxwell equations

−∆p − κ2 p = 0 curl curl e − κ2 e = 0

Boundary integral operators:

1 −1
Dirichlet trace γp ∈ H 2 (Γ ) Dirichlet trace γt e ∈ H × 2 (divΓ , Γ )
1 1 1 −1
Neumann trace γ (grad p)
κ n
∈ H − 2 (Γ ) Neumann trace γ (curl e)
κ t
∈ H × 2 (divΓ , Γ )

„ « „ «
Kκ Vκ Cκ Sκ
Aκ = Aκ =
Dκ K˜κ Sκ Cκ

The issue of coercivity

„ « „ «
Kκ V0 Cκ S0
Aκ = + compact pert. Aκ = + compact pert.
D0 Kκ∗ S0 Cκ

Ellipticity on trace spaces: No ellipticity, because S0 indefinite:

⟨ϕ, V0 ϕ⟩ 1 ,Γ ≥ C ∥ϕ∥2 −1 , S0 = V0 + κ1 curlΓ ◦ V0 ◦ divΓ .

2 H 2 (Γ )

⟨D0 ϕ, ϕ⟩ 1 ,Γ ≥ C ∥ϕ∥2 1 . Yet, individual terms are (semi)-definite.

2 H 2 (Γ )/C

has the same roots as the absence of a direct compact embedding in the case of the
Maxwell source problem, cf. Section 3.
The roots of the difficulties being the same as for the Maxwell source problem,
the same ideas should provide remedies: we have to employ stable splittings that
−1
target the trace space H × 2 (divΓ , Γ ) and decompose it into the kernel of divΓ and
a suitable more regular complement. The decomposition (21) introduced in Sect. 3
meets all the requirements and will be used below.
Please recall the discussion in Sect. 3 of the coercivity of the bilinear form as-
sociated with the Maxwell source problem. The same considerations will now be
applied to the bilinear form spawned by the boundary integral operator Sκ through
(ξ, µ) -→ ⟨Sκ µ, ξ⟩τ ,Γ . To begin with, the “lower order” term Vκ in the operator
Sκ , cf. Formula (31), becomes compact on the “regular component” X (Γ ) of the
decomposition (21).
 Galerkin BEM in Electromagnetics 21
− 12
Lemma 9. The bilinear forms ⟨Vκ ·, ·⟩τ ,Γ : X (Γ )×H × (Γ ) -→ C and ⟨Vκ ·, ·⟩τ ,Γ :
−1
H × 2 (Γ ) × X (Γ ) -→ C are compact.
−1 1
Proof. Since Vκ : H × 2 (Γ ) -→ H ×
2
(Γ ) is continuous according to Theorem 3, the
−1
compact embedding X (Γ ) (→ H × 2 (Γ ) (Corollary 1) gives the result. ⊓
9

We can even establish a generalized Gårding inequality for S0 on Lipschitz bound-

aries: looking at the formula (31) and, in particular, the bilinear form
1
⟨S0 µ, ξ⟩τ ,Γ = ⟨divΓ µ, V0 divΓ µ⟩ 1 ,Γ − κ ⟨µ, V0 ξ⟩τ ,Γ , (37)
κ 2

we realize a striking similarity to the bilinear form of the Maxwell source problem
(15). Thus, it is natural to employ the splitting idea of Sect. 3 based on (21) and the
isomorphism
−1 −1
XΓ = RΓ − ZΓ : H × 2 (divΓ , Γ ) -→ H × 2 (divΓ , Γ ) . (38)

Lemma 10 (Generalized Gårding inequality for Sκ ). There is a compact bilinear

−1 −1
form cΓ : H × 2 (divΓ , Γ ) × H × 2 (divΓ , Γ ) -→ C and a constant CG > 0 such that

−1
| (Sκ µ, XΓ µ)τ + cΓ (µ, µ)| ≥ CG ∥µ∥2 −1 ∀µ ∈ H × 2 (divΓ , Γ ) .
H × 2 (divΓ ,Γ )

Proof. We set
. / . / . /
cΓ (µ, ξ) := − Vκ RΓ µ, RΓ ξ τ ,Γ + Vκ RΓ µ, ZΓ ξ τ ,Γ − Vκ ZΓ µ, RΓ ξ τ ,Γ ,

which is compact by Lemma 9. Noting that

. / 1. / . /
S0 µ, XΓ ξ τ ,Γ = V0 divΓ RΓ µ, divΓ RΓ ξ 0;Γ+κ ZΓ µ, V0 ZΓ ξ τ ,Γ−cΓ (µ, ξ),
κ
we invoke Lemma 8 and the stability of the decomposition to finish the proof. ⊓
9

In the case of smooth domains this result is sufficient to obtain coercivity of Aκ ,

because for smooth boundaries the singularity of the kernel of Cκ partly cancels.
This is a well-known effect in the case of double layer potentials for second order
elliptic operators. For Cκ the observation was made by Nédélec [53, Section 5.5].
Lemma 11. If Γ is smooth, that is, of class C ∞ , then Cκ is continuous as an opera-
1 1
tor Cκ : H st (Γ ) -→ H s+1
t (Γ ) and Cκ : T H s− 2 (divΓ ; Γ ) -→ T H s+ 2 (divΓ ; Γ )
for all s ∈ R.

Proof. The first part of the proof boils down to manipulations of (33) using the
product rule for curlx and the identity (b × c) × a = c(a · b) − b(a · c). For
µ, ξ ∈ L∞ (Γ ) ∩ T H (divΓ ; Γ ) we end up with
22 A. Buffa and R. Hiptmair

⟨Cκ µ, ξ⟩τ ,Γ =
! !
=− ((µ(y) × gradx Eκ (x − y)) × n(x)) · (ξ(x) × n(x)) dS(y, x)
!Γ !Γ
=− µ(y)(gradx Eκ (x − y) · n(x))· (ξ(x)×n(x)) dS(y, x)+
Γ Γ
! !
+ gradx Eκ (x − y) (µ(y) · (n(x)−n(y)))·(ξ(x)×n(x)) dS(y, x) .
Γ Γ

According to [25, Section 6.4] we have |n(x) − n(y)| = O(|x − y|) for smooth
surfaces. Thus, a closer scrutiny of the formulae shows that

gradx Eκ (x − y) · n(x) ≃ gradx Eκ (x − y)(n(x) − n(y))T ≃ O(|x − y|−1 ) ,

for x → y. Both kernels are weakly singular, as is the kernel of Sκ . So the theory
of pseudo-differential operators [25, Ch. 4.4] shows that Cκ is continuous as an
operator from H st (Γ ) -→ H s+1
t (Γ ), s ∈ R (Note that on a smooth boundary the
infinite scale of Sobolev spaces is available).
Next, pick a smooth tangential vector-field µ, use Lemma 5 and apply simple
manipulations based on vector identities
!
1
divΓ Cκ (µ)(x) = curl curl Eκ (x − y)µ(y) dS · n(x)
κ
Γ
! !
1 ∂
= Eκ (x − y)divΓ µ(y) dS + κ Eκ (x − y)µ(y) dS · n(x) .
κ ∂n(x)
Γ Γ

s− 21 1
By density, we conclude that divΓ ◦ Cκ : T H (divΓ ; Γ ) -→ H s+ 2 (Γ ) is con-
tinuous. This can be combined with the previous results and confirms the second
assertion of the theorem. ⊓
9

The crucial message sent by this lemma and Lemma 7 is that on smooth bound-
−1 −1
aries the operator Aκ : H × 2 (divΓ , Γ )2 -→ H × 2 (divΓ , Γ )2 can be converted into
, -
0 S0
Aκ ≃ .
S0 0

by dropping “compact perturbations”3. In other words, Dirichlet and Neumann

traces are coupled by compact terms only. On smooth boundaries we merely have
to examine Sκ , if we are interested in coercivity.
Unfortunately, the coupling terms Cκ in Aκ cannot be discarded in the case of
non-smooth boundaries for want of a result like Lemma 11: in general, we have to
3
Here and below we use the symbol ≃ to express equality of operators and bilinear forms
up to addition of compact terms
 Galerkin BEM in Electromagnetics 23

deal with the two different traces γt e and γN e together. This is a completely new
aspect of boundary integral operators that we have not encountered in the case of
the Maxwell source problem. Thus, splitting alone is not enough, but has to be ac-
companied by an appropriate grouping of the components. This is where the “phys-
ical meaning” of the splitting (21) that we discussed in Table 2 offers an important
hint: it suggests that we distinguish between trace components of electric and mag-
netic nature. The ultimate justification for this idea is the profound result that it is
merely compact terms, by which electric and magnetic components are coupled in
the boundary integral operator Aκ . The next lemma rigorously expresses this in-
sight, cf. Prop. 3.13 in [21].
Lemma 12. The bilinear form ⟨Cκ ·, ·⟩τ ,Γ is compact both on N (Γ ) × N (Γ ) and
X (Γ ) × X (Γ ).
Proof. We restrict ourselves to the proof of the second assertion. We choose some
ζ, µ ∈ X (Γ ) and recall the definition of Cκ along with the jump relations. It is
important to note that, by virtue of the definition of X (Γ ), µ can be extended by
v := J(divΓ µ) ∈ H 1 (Ωs ), J defined in the proof of Lemma 2, such that γt v = µ
and ∥v∥H 1 (Ωs ) ≤ C ∥µ∥ − 21 . Also, exploiting div v = 0, we get
H × (divΓ ,Γ )
. − κ / 1
⟨Cκ ζ, µ⟩τ ,Γ = γN Ψ V (ζ), µ τ ,Γ − 2 ⟨ζ, µ⟩τ ,Γ .

Using the identity curl curl Ψ κV (ζ) = grad ΨVκ (divΓ ζ) + κ2 Ψ κV (ζ) and the inte-
gration by parts (5), we obtain:
!
. − κ /
γN Ψ V (ζ), µ τ ,Γ = − curl Ψ κV · curl v + κ2 Ψ κV (ζ) · v dx+
Ωs
. /
+ γ − ΨVκ (divΓ µ), γn− v 1 ,Γ .
2

This means that

| (Cκ (ζ), µ)τ | ≤ |Ψ κV (ζ)|H 1 (Ωs ) ∥curl v∥L2 (Ωs ) + κ2 ∥Ψ κV (ζ)∥L2 (Ωs ) ∥v∥L2 (Ωs )
0 0
+ ∥Vκ (divΓ ζ)∥ 2 0γn− v0 2
L (Γ ) + ∥ζ∥L2 (Γ ) ∥γ − v∥L2 (Γ )
L (Γ ) t t t

≤ C(∥ζ∥ −1 + ∥Vκ (divΓ ζ)∥L2 (Γ ) + ∥ζ∥L2t (Γ ) ) ∥v∥H 1 (Ωs ) ,

H× 2 (Γ )

with some C = C(Ωs ) > 0. It goes without saying that the operator Vκ :
1
H − 2 (Γ ) -→ L2 (Γ ) is compact. Then, the compact embedding of X (Γ ) in L2t (Γ )
according to Corollary 1 finishes the proof. ⊓
9
To understand the meaning of these results, we consider the combined boundary
−1 −1
integral operator Aκ : H × 2 (divΓ , Γ )2 -→ H × 2 (divΓ , Γ )2 with respect to the
splitting (21). As usual, we adopt a variational perspective and study the bilinear
form asscociated with Aκ . It will be based on the following anti-symmetric pairing
−1 −1
on the product space H × 2 (divΓ , Γ ) × H × 2 (divΓ , Γ ),
24 A. Buffa and R. Hiptmair
1, - , -2
ζ ξ
, := ⟨ζ, λ⟩τ ,Γ + ⟨µ, ξ⟩τ ,Γ .
µ λ τ ×τ

−1
Now, pick ζ, µ, ξ, λ ∈ H × 2 (divΓ , Γ ) and use superscripts ⊥ and 0 to tag their
components in X (Γ ) and N (Γ ), respectively.
1 , - , -2 3 , - , -4 3 , - , -4
ζ ξ ζ⊥ ξ⊥ ζ0 ξ0
Aκ , = Aκ , + A κ , +
µ λ τ ×τ µ0 λ0 µ⊥ λ⊥
τ ×τ τ ×τ
3 , - , -4 3 , - , -4
ζ⊥ ξ0 ζ0 ξ⊥
+ Aκ , + A κ , .
µ0 λ⊥ µ⊥ λ0
τ ×τ τ ×τ
(39)

Let us take a look at the bilinear forms in the second line:

3 , - , -4 ⎧5 6 5 6
⊥ ⊥ 0 ⊥
ζ ⊥
ξ 0 ⎨ C κ ζ , λ + κ Vκ µ , λ +
Aκ , ⊥ = 5 6
τ ,Γ
. / τ ,Γ
0 ⎩ +κ Vκ ζ ⊥ , ξ0
µ λ
τ ×τ
+ Cκ µ0 , ξ 0 τ ,Γ .
τ ,Γ

Lemmas 9 and 12 show that this is a compact bilinear form! The same applies to
the other term in the second line of (39). Harking back to the discussion in Sect. 3,
we emphasize that both Dirichlet and Neumann trace involve electric and magnetic
components, which are isolated by the splitting:

Electric components: ζ 0 , µ⊥ ←→ Magnetic components: ζ ⊥ , µ0 .

The bottom line is that up to compact terms electric and magnetic components of
the traces are decoupled in Aκ . It has turned out that the decoupling observed in the
case of smooth boundaries does not reflect the “physics of the fields”.
Using the appropriate splitting and decoupling, we can proceed as in the case of
−1 −1
S0 : we introduce the isomorphism XΓ : H × 2 (divΓ , Γ )2 -→ H × 2 (divΓ , Γ )2 by
, - , -
ζ XΓ ζ −1
XΓ := , ζ, µ ∈ H × 2 (divΓ , Γ ) . (40)
µ XΓ µ

Then we get the following generalization of Lemma 10.

Theorem 9 (Generalized Gårding inequality for Aκ ). There is a constant CG > 0
−1 −1
and a compact bilinear form cΓ on H × 2 (divΓ , Γ ) × H × 2 (divΓ , Γ ) such that
71 , - , -2 , - , - 77 0, -02
7 ζ ζ ζ ζ 7 0 ζ 0
7
7 Aκ , XΓ − cΓ ( , )7 ≥ CG 0 0
0 µ 0 − 12
7 µ µ τ ×τ µ µ 7 H (divΓ ,Γ )
×

−1
for all ζ, µ ∈ H × 2 (divΓ , Γ ).
 Galerkin BEM in Electromagnetics 25

Proof. We have already found that up to compact perturbations

1 , - , -2 3 , - , -4 3 , - , -4
ζ ξ ζ⊥ ξ⊥ ζ0 ξ0
Aκ , ≃ Aκ , 0 + A κ , .
µ λ τ ,Γ µ0 λ µ⊥ λ⊥
τ ,Γ τ ,Γ

What comes next amounts to reusing arguments from the proof of Lemma 10. We
inspect the first summand and find, using Lemmas 6 and 9,
3 , - , ⊥ -4
ζ⊥ ζ 5 6 5 6
⊥ 0 0 0
Aκ , = − C κ ζ , µ − κ Vκ µ , µ +
µ0 −µ0 5 6 τ ,Γ 5 6 τ ,Γ
τ ×τ ⊥ ⊥
+ Sκ ζ ⊥ , ζ + Cκ µ 0 , ζ
τ ,Γ τ ,Γ
85 6 9 5 6 1 . /
≃ −2i Im Cκ ζ ⊥ , µ 0 −κ V0 µ0 , µ0 + V0 (divΓ ζ, divΓ ζ τ ,Γ .
τ ,Γ τ ,Γ κ
Appealing to Lemmas 2 and 8, we conclude that
73 , - , ⊥ -4 7 , -
7 ζ⊥ ζ 7 0 0 02 0 02
7 7 0 0 0 ⊥0
7 Aκ , + comp. 7 ≥ C µ 1 + 0ζ 0 − 21 .
−µ0
−
7 µ0 7 H × 2 (Γ ) H × (divΓ ,Γ )
τ ×τ

The same manipulations can be carried out for the second summand. Together with
the stability of (21) this gives the assertion. ⊓
9

7 Boundary Integral Equations

Boundary integral equations (BIE) can be obtained in two ways, either by the di-
rect method or the indirect method. The distinct feature of the direct method is that
traces of the solution of the transmission problem/boundary value problem occur
as unknowns in the formulation. Its integral equations immediately arise from the
Calderon projectors P− +
κ and Pκ via Theorem 8. Conversely, the unknowns of the in-
direct methods are jumps of traces across Γ . It can be motivated by the fact that the
potentials Ψ κSL and Ψ κDL already provide solutions to the homogeneous equations,
cf. (29). An excellent presentation of the main ideas of indirect methods is given in
[36, Vol. IV, Ch. XI].

7.1 The direct method

We start with the discussion of direct methods for scattering at a perfect conductor,
that is, the exterior Dirichlet problem for the homogeneous electric wave equation
curl curl e − κ2 e = 0 in Ω ′ , γt+ e = γt+ ei , (41)
plus Silver–Müller radiation conditions. We know that we can always find a unique
solution of (41) [31, Thm. 6.10]. However, it is a bewildering feature of many
boundary integral equations connected with (41) that they fail to have unique so-
lutions, if κ coincides with “forbidden wave numbers” [26, 38]. Those are related to
interior eigenvalues of the operator of (41).
26 A. Buffa and R. Hiptmair

Definition 4. λ ∈ R is called an interior electric/magnetic Maxwell eigenvalue, if

there is a non-zero e ∈ H 0 (curl; Ωs ) or e ∈ H(curl; Ωs ), respectively, such that

(curl e, curl v)0;Ωs = λ (e, v)0;Ωs ∀v ∈ H 0 (curl; Ωs ) or v ∈ H(curl; Ωs ) .

Note that these eigenvalues form a discrete sequence accumulating at ∞.

The first direct method relies on Theorem 8, which tells us that (γt+ ei , λ)
+
are exterior Cauchy data according to Definition 3, if P−κ (γt ei , λ) = 0. From
the first row of this equation we obtain the integral equation of the first kind
Sκ λ = −( 21 Id +Cκ )(γt+ ei ) for the unknown Neumann data λ := γN +
e of a solu-
−1
tion e of (41). In weak form it reads: seek λ ∈ H × 2 (divΓ , Γ ) such that
. / −1
⟨Sκ λ, µ⟩τ ,Γ = − ( 12 Id +Cκ )(γt+ ei ), µ τ ,Γ ∀µ ∈ H × 2 (divΓ , Γ ) . (42)

Conversely, if (γt+ ei , λ) satisfies (42), we find

, + - , -
γt ei 0 −1
P−κ = for some ξ ∈ H × 2 (divΓ , Γ ) .
λ ξ

Hence, by Theorem 8, ξ is the Neumann trace of an electric eigenmode of Ωs . If κ2

does not coincide with an interior electric eigenvalue, this eigenmode can only be
trivial, which means ξ = 0. The next lemma summarizes our findings.
Lemma 13. Assume that κ2 is not an interior electric eigenvalue. Then λ ∈
−1
H × 2 (divΓ , Γ ) is a solution of (42) if and only if (γt+ ei , λ) are Cauchy data for
(41).
Remark 1. If κ2 is an interior electric eigenvalue, then λ is unique up to Neumann
traces ξ of the corresponding eigenmodes. Thanks to the representation formula
(26), we find that Ψ κSL (ξ) vanishes in Ω ′ . In other words, the representation

e = −Ψ κDL (γt+ ei ) − Ψ κSL (λ) (43)

will produce the unique field solution in Ω ′ .

Now, a standard Fredholm alternative argument can be applied:
Theorem 10. Provided that κ satisfies the assumptions of Lemma 13, there exists a
unique solution of (42) for any ei .
1 +
Using the second row of P−
κ we obtain the B.I.E. ( 2 Id +Cκ )λ = −Sκ (γt ei ),
−1
whose associated variational problem can be stated as: seek λ ∈ H × 2 (divΓ , Γ )
such that
. 1 / . / −1
( 2 Id +Cκ )λ, µ τ ,Γ = − Sκ (γt+ ei ), µ τ ,Γ ∀µ ∈ H × 2 (divΓ , Γ ) . (44)

In contrast to (42), in order to show unique solvability of (44) we need not only
avoid “forbidden wave numbers”, but have to assume smooth boundaries, too.
 Galerkin BEM in Electromagnetics 27

Theorem 11. Let κ2 be different from a interior magnetic eigenvalue. If Γ is C ∞ -

1
smooth, then (44) has a unique solution λ ∈ T H − 2 (divΓ ; Γ ) and the pair
+ i
(γt e , λ) supplies Cauchy data for the electric wave equation in Ω ′ .

Proof. As in the justification of Lemma 13 it turns out that

, + - , -
− γt ei ξ 1
Pκ = for some ξ ∈ T H − 2 (divΓ ; Γ ) ,
λ 0

if λ satisfies (44). By the assumption on κ we have ξ = 0 and, by Theorem 8,

(γt+ ei , λ) is identified as valid Maxwell Cauchy data for the exterior problem. Re-
calling the uniqueness result for (41), this means that solutions of (44) are unique.
1
Next, use Lemma 11, which asserts the compactness of Cκ : T H − 2 (divΓ ; Γ ) →
1
T H − 2 (divΓ ; Γ ). This confirms that the operator in (44) is Fredholm of index zero.
⊓
9

If Γ ∈ C ∞ and γt+ ei ∈ H 1t (Γ ) (which, e.g., is fulfilled for exciting plane

waves), the lifting properties of the operators Cκ according to Lemma 11 and the
fact that divΓ ◦ Sκ = κ divΓ ◦ Vκ bear out that the solution of (44) will be-
long to T H(divΓ , Γ ). Hence, a completely equivalent variational formulation in
T H(divΓ , Γ ) is possible: find λ ∈ T H(divΓ , Γ ) such that ∀µ ∈ T H(divΓ , Γ )
% 1 & % &
( 2 Id +Cκ )λ, µ T H(divΓ ,Γ ) = − Sκ (γt+ ei ), µ T H(divΓ ,Γ ) . (45)

Given a sufficiently smooth γt+ ei , the right hand side is a continuous functional
on T H(divΓ , Γ ). In addition, Lemma 11 shows that Cκ : T H(divΓ , Γ ) -→
T H 1 (divΓ , Γ ) and, hence, the sesqui-linear form in (45) turns out to be T H(divΓ , Γ )-
coercive. Thus, Thm. 11 will remain valid for (45). The real rational behind the
lifting of (44) into T H(divΓ , Γ ) will be elaborated in Sect. 9.
smallskip
Next, we tackle scattering at an isotropic, homogeneous dielectric object occu-
− −
pying Ωs . Inside Ωs material
: parameters ϵ > 0 and µ > 0 prevail, leading to
−
a wave number κ := ω ϵ µ . Outside we face ϵ0 , µ0 and wave number κ+ .
− −
− +
These wave numbers underlie the definition of γN and γN . The transmission condi-
tions from (1) become
κ− − κ+ +
γt− e = γt+ e , µ− γN e = µ0 γN e .

Taking our cue from the approach to acoustic scattering in [61], we introduce scaled
boundary integral operators
< = , - < = , -
Id 0 Id 0 Id 0 Id 0
;
Aκ − = Aκ − ;
, Aκ + = + Aκ + .
0 κµ+0
− −
0 µκ− 0 µκ− 0 κµ0

The following scaled traces match the scaled operators

28 A. Buffa and R. Hiptmair
+ −
(ζ + , λ+ ) = (γt+ e, κµ0 γN
+
e) , (ζ − , λ− ) = (γt− e, µκ− γN
−
e) .

For them the transmission condition takes the simple form

, −- , +- , + -
ζ ζ γt ei
− = . (46)
λ− λ+ γt+ hi

A scaled version of Theorem 8 bears out that (ζ − , λ− ) and (ζ + , λ+ ) are inte-

rior/exterior Cauchy data for the electric wave equation with wave numbers κ− and
κ+ , respectively, if and only if
, −- , +-
; κ− ) ζ − = 0 , ( 1 Id +A
( 21 Id −A ; κ+ ) ζ + = 0 . (47)
λ 2 λ

Using (46), this immediately implies that

> ? ,ζ+ - , + -
γt ei
; κ− + A
A ; κ+ + = ( 1
Id − ;
A κ − ) . (48)
λ 2
γt+ hi

These are the boundary integral equations of the direct method for the transmission
% − & % + & %γ + e &
problem. Conversely, if (ζ + , λ+ ) is a solution of (48), set λζ − = λζ + + γ t+ hi ,
t i
and consider
, @− - , −- , @+ - , +-
ζ ζ
− ; κ− ) ζ
:= ( 12 Id −A , + ; κ+ ) ζ
:= ( 12 Id +A .
λ@ λ− λ@ λ+
−
@ − ) and (ζ@+ , λ
Owing to Thm. 8, the pairs (ζ@ , λ @ + ) are Maxwell Cauchy data for Ω ′
−
and Ωs (and κ− , κ+ ), respectively. From equation (48) we infer that (ζ@ , λ @−) =
+ +
(ζ@ , λ
@ ). Thus, the interior and exterior Dirichlet and Neumann traces of the re-
lated Maxwell solutions agree. A combination of these Maxwell solutions solves
the homogeneous electric wave equation (with κ+ inside Ωs and κ− outside) in R3
and satisfies the Silver–Müller radiation conditions. Thanks to the uniqueness of so-
−
@−) = 0
lutions of the exterior Maxwell problem, it has to vanish. This implies (ζ@ , λ
+
@ + ) = 0, so that we recover (48). This confirms the following result.
and (ζ@ , λ
Lemma 14. Any solution (ζ + , λ+ ) of (48) provides (scaled) exterior Cauchy data
for the transmission problem with excitation by an incident wave (ei , hi ).
−1
Using the pairing ⟨·, ·⟩τ ×τ , the variational formulation of (48) in H × 2 (divΓ , Γ ) ×
−1
H × 2 (divΓ , Γ ) is straightforward. So is the next theorem that arises from Thm. 9,
the previous Lemma, and a Fredholm argument.
Theorem 12. The boundary integral equation (48) has a unique solution for any
excitation.
 Galerkin BEM in Electromagnetics 29

7.2 The indirect method

We will only discuss the exterior Dirichlet problem for the electric wave equation.
Let e+ denote the unique solution of the exterior Dirichlet boundary value problem,
satisfying γt+ e+ = γt+ ei and the Silver–Müller radiation conditions at ∞. Write
e− for the solution of an interior Dirichlet problem for the electric wave equation,
such that γt− e− = γt+ e+ . Again it is crucial to stay away from “forbidden wave
numbers”: let us assume that κ2 does not coincide with an interior electric eigen-
value. Therefore, such an e− exists and is unique. Call e the Maxwell solution in
Ωs ∪Ω ′ that emerges by combining e+ and e− . As [γt ]Γ (e) = 0, the representation
formula (26) becomes

e = −Ψ κSL ([γN ]Γ (e)) in H loc (curl2 , Ωs ∪ Ω ′ ) .

Applying the exterior Dirichlet trace γt+ gives us the final integral equation in weak
−1
form: seek the unknown jump λ := [γN ]Γ (e) ∈ H × 2 (divΓ , Γ ), which satisfies
. / −1
⟨Sκ λ, µ⟩τ ,Γ = − γt+ ei , µ τ ,Γ ∀µ ∈ H × 2 (divΓ , Γ ) . (49)

This integral equation is also known as electric field integral equation (EFIE) or
Rumsey’s principle. Theorem 10 applies, because (42) and (49) feature the same
bilinear form.
Parallel to the case of direct methods for the exterior Dirichlet problem, we
have a second option also in the case of the indirect approach. We assume that κ2
does not agree with an interior magnetic eigenvalue. Then, we may choose e− as
a Neumann extension of e+ , that is, e− is the solution of the interior Neumann
− − + +
problem for the electric wave equation with Neumann data γN e = γN e . Com-
+ − κ
bining e and e to form e, we conclude from (26) that e = −Ψ DL ([γt ]Γ (e))
in H loc (curl2 , Ωs ∪ Ω ′ ). Applying the exterior Dirichlet trace to this equation, we
get the so-called magnetic field integral equation (MFIE), an integral equation of
−1
the second kind: find ζ ∈ H × 2 (divΓ , Γ ) with
. 1 / . / −1
( 2 Id −Cκ )ζ, µ τ ,Γ = γt+ ei , µ τ ,Γ ∀µ ∈ H × 2 (divΓ , Γ ) . (50)

Its theoretical analysis on smooth surfaces is already covered by Thm. 11.

A serious drawback of the integral equations stated so far is their vulnerability to
the presence of forbidden wave numbers, though the related boundary value problem
always possesses a unique solution. Only one class of indirect BIE, the so-called
combined field integral equations (CFIE), enjoys immunity. They owe their name
to the fact that both Ψ κDL and Ψ κSL enter the trial expression for e. A crucial prere-
−1 −1
quisite is a compact “smoothing operator” M : H × 2 (divΓ , Γ ) -→ H × 2 (divΓ , Γ )
that satisfies
−1
µ ∈ H × 2 (divΓ , Γ ) : ⟨Mµ, µ⟩τ ,Γ > 0 ⇔ µ ̸= 0 .
30 A. Buffa and R. Hiptmair

It is an important building block of the trial representation formula

e = −iηΨ κSL (ζ) − Ψ κDL (Mζ) , (51)
− 21
where ζ ∈ H × (divΓ , Γ ), η > 0. By (29), this field is a Maxwell solution in
Ωs ∪ Ω ′ . The exterior Dirichlet trace applied to (51) results in the combined field
−1 −1
integral equation: find ζ ∈ H × 2 (divΓ , Γ ) such that ∀µ ∈ H × 2 (divΓ , Γ )
. / . /
−i ⟨ηSκ (ζ), µ⟩τ ,Γ + ( 12 Id −Cκ )(Mζ), µ τ ,Γ = γt+ ei , µ τ ,Γ . (52)
The idea to use a regularizing operator to state a combined field integral equation is
due to Kress [48].
Theorem 13. The boundary integral equation (52) has a unique solution ζ ∈
−1
H × 2 (divΓ , Γ ) for all η > 0, κ > 0.
−1
Proof. To demonstrate uniqueness, we assume that ζ ∈ H × 2 (divΓ , Γ ) solves
−iηSκ (ζ) + ( 12 Id −Cκ )(Mζ) = 0 . (53)
It is immediate from the jump relations that e given by (51) is an exterior Maxwell
solution with γt+ e = 0. By uniqueness we infer that e = 0 in Ω ′ . Appealing to the
jump relations from Theorem 7 once more, we find
γt− e = −Mζ , −
γN e = −iηζ .
Next, we use (11) and see that
5 6 !
. / − 1
iη ζ, Mζ τ ,Γ = γN e, γt− e = | curl e|2 dx − κ|e|2 dx ∈ R .
τ ,Γ Ωs κ
% &
Necessarily, ζ, Mζ τ = 0, so that the requirements on M imply ζ = 0.
Knowing that M is compact, we conclude from Lemma 10 that the bilinear form
of (52) satisfies a generalized Gårding inequality. Thus, Thm. 4 gives existence from
uniqueness. ⊓
9
A possible candidate for M can be introduced through a variational definition:
−1
for ζ ∈ H × 2 (divΓ , Γ ) and all q ∈ H × (divΓ , Γ ), Mζ ∈ H × (divΓ , Γ ) is to
satisfy
⟨Mζ, q⟩0;Γ + ⟨divΓ Mζ, divΓ q⟩0;Γ = ⟨q, ζ⟩τ ,Γ . (54)
−1
Obviously, M : H × 2 (divΓ , Γ ) -→ H × (divΓ , Γ ) is a continuous linear operator.
−1
By density of H × (divΓ , Γ ) in H × 2 (divΓ , Γ ), M must be injective, which also
means
. / 2
Mζ, ζ τ ,Γ = ∥Mζ∥H × (divΓ ,Γ ) > 0 ⇔ ζ = ̸ 0.

It is easy to see that M inherits compactness from the embedding H × (divΓ , Γ ) (→

−1
H × 2 (divΓ , Γ ).
 Galerkin BEM in Electromagnetics 31

8 Boundary Element Spaces

We equip the piecewise smooth compact two-dimensional surface Γ with an ori-
ented triangulation Γh . This means that all its edges are endowed with a direc-
tion. We assume a perfect resolution of Γ , that is Γ = K̄1 ∪ . . . ∪ K̄N , where
Kh := {K1 , . . . , KN } is the set of mutually disjoint open cells of Γh . Moreover,
no cell may straddle boundaries of the smooth faces Γ j of Γ . We will admit tri-
angular and quadrilateral cells only: for each K ∈ Kh there is a diffeomorphism
ΦK : K ; -→ K̄, where K ; is the “unit triangle” or unit square in R2 , depending on
the shape of K [27, Sect. 5].
This paves the way for a parametric construction of boundary elements: to begin
with, choose finite-dimensional local spaces W(K) ; ⊂ (C ∞ (K)) ; 2 of polynomial
vectorfields together with a dual basis of so-called local degrees of freedom (d.o.f.).
Possible choices for W(K) ; and related d.o.f. abound: we may use the classical
triangular Raviart-Thomas (RTp ) elements of polynomial order p ∈ N0 [56],

; := {x -→ p1 (x) + p2 (x) · x, x ∈ K,
W(K) ; p1 ∈ (Pp (K))
; 2 , p2 ∈ Pp (K)}
; ,

where Pp (K); is the space of two-variable polynomials of total degree ≤ p. An al-

ternative are the triangular BDMp elements of degree p [12], p ∈ N0 , which rely on
W(K) ; := (Pp+1 (K)) ; 2 . In both cases, the usual d.o.f. involve certain polynomial
moments of normal components on edges, together with interior vectorial moments
for p > 0. For instance, in the case of RT0 , edge fluxes are the appropriate degrees
of freedom:
!
µh ∈ W(K) ; -→ µh · n ; dS , e; edge of K ; .
e
b

Similar local spaces and degrees of freedom are available for the unit square.
Using the pull-back of 1-forms the local spaces can be lifted to the cells of Γh .
In terms of vectorfields this is equivalent to the Piola transformation
:
(FK µ)(x) := det(G) G−1 DΦTK (; x)µ(;x) , (55)

where G := DΦ(; x)T DΦ(;x), x = ΦK (;

x), x ; Thus, we can introduce the
; ∈ K.
global boundary element space
; ∀K ∈ Kh } .
W h := {µ ∈ H × (divΓ , Γ ) : µ|K ∈ FK (W(K)) (56)

In practice, W h ⊂ H × (divΓ , Γ ) is ensured by a suitable choice of d.o.f. Remem-

ber that d.o.f. have to be associated with individual edges of K ; or the interior of K.
;
It is crucial that the normal component of any µ ; h ∈ W(K) ; on any edge e; of K ;
vanishes if and only if µ; h belongs to the kernel of all local d.o.f. associated with e;.
In light of (12), this ensures W ⊂ H × (divΓ , Γ ). In the sequel W h will designate
a generic H × (divΓ , Γ )-conforming boundary element space. It may arise from the
RTp family of elements, p ∈ N0 , the BDMp family, or a combination of both.
32 A. Buffa and R. Hiptmair

Based on the degrees of freedom we can introduce local interpolation operators

Πh : Dom(Πh ) -→ W h . It is a projector onto W h and enjoys the fundamental
commuting diagram property

divΓ ◦ Πh = Qh ◦ divΓ on H × (divΓ , Γ ) ∩ Dom(Πh ) . (57)

Here, Qh is the L2 (Γ )-orthogonal projection onto a suitable space Qh of Γh -

piecewise polynomial discontinuous functions. It must be emphasized that the inter-
polation operators Πh fail to be bounded on H × (divΓ , Γ ); slightly more regularity
of tangential vectorfields in Dom(Πh ) is required [7, Lemma 4.7].
Next, we turn our attention to asymptotic properties of the boundary element
spaces, in particular to estimates of interpolation errors and best approximation er-
rors. We restrict ourselves to the h-version of boundary elements, which relies on
shape-regular families {Γh }h∈H of triangulations of Γ [30, Ch. 3,§ 3.1]. Here, H
stands for a decreasing sequence of meshwidths, and H is assumed to converge to
zero.
By means of transformation to reference elements, the commuting diagram
property, and Bramble-Hilbert arguments, interpolation error estimates can easily
be obtained [13, III.3.3].
Lemma 15 (Interpolation error estimate). For 0 < s ≤ p + 1 we find constants
C > 0, depending only on the shape regularity of the meshes and s, such that for
all µ ∈ H s× (Γ ) ∩ H × (divΓ , Γ ), h ∈ H,
> ?
∥µ − Πh µ∥L2 (Γ ) ≤ Chs ∥µ∥H s× (Γ ) + ∥divΓ µ∥L2 (Γ ) ,

s
and such that for all µ ∈ H × (divΓ , Γ ), divΓ µ ∈ H− (Γ )

∥divΓ (µ − Πh µ)∥L2 (Γ ) ≤ Chs ∥divΓ µ∥H−s (Γ ) .

Corollary 5. The union of all boundary element spaces W h , h ∈ H, is dense in

−1
H × 2 (divΓ , Γ ).

A particular variant of the above interpolation error estimate addresses vector fields
with discrete surface divergence:
Lemma 16. If µ ∈ H s× (Γ ), 0 < s ≤ 1, and divΓ µ ∈ Qh , then there is a constant
C > 0, depending on the shape-regularity of the meshes only, such that

∥µ − Πh µ∥L2t (Γ ) ≤ Chs ∥µ∥H s× (Γ ) .

¿From the interpolation error estimates we instantly get best approximation es-
timates in terms of the H × (divΓ , Γ )-norm. Yet, what we actually need is a result
about approximation in the “energy norm” ∥·∥ − 12 of the form
H × (divΓ ,Γ )

1
inf ∥µh − ξ h ∥ −1 ≤ Chs+ 2 ∥µ∥H s (divΓ ,Γ ) . (58)
ξh H× 2 (divΓ ,Γ ) ×
 Galerkin BEM in Electromagnetics 33

The estimate in H × (divΓ , Γ ) does not directly provide (58). Even worse, stan-
dard duality arguments cannot be applied. Recall their main idea: we set out from
a Hilbertian triple V ⊂ H ⊂ V ′ , have a finite dimensional subspace of H, say
Vh , and we want to estimate the best approximation error in V ′ . Then it is crucial
that we know how to use the difference in regularity between H and V through an
estimate of the type ∃ vh ∈ Vh : ∥u − vh ∥H ≤ C(h)∥u∥V , with C(h) optimal in
a suitable sense.
Here, we have an estimate between H × (divΓ , Γ ) and H s× (divΓ , Γ ), s > 0.
Thus, we should use H × (divΓ , Γ ) as self dual space, i.e., the standard inner
product in H × (divΓ , Γ ). But, in order to conclude, we should be able to prove
that H −s s 1
× (divΓ , Γ ) is dual of H × (divΓ , Γ ) for 0 < s ≤ 2 with respect to the
H × (divΓ , Γ ) inner product. Unfortunately this is the case for regular surfaces but
not for non-regular ones [29].
The question of obtaining (58) has been addressed in [16] and the idea is to
use the duality argument face by face (which are seen as regular open manifolds),
exploiting continuity of the normal components of vector-fields in H × (divΓ , Γ ).
At the end of a technical procedure we obtain the following result:
−1
Theorem 14. Let Ph : H × 2 (divΓ , Γ ) → W h be the orthogonal projection with
−1
respect to the H × 2 (divΓ , Γ ) inner product. Then, for any − 21 ≤ s ≤ p + 1 we
have
1
∥µ − Ph µ∥ −1 ≤ Chs+ 2 ∥µ∥H s (divΓ ,Γ ) ∀µ ∈ H s× (divΓ , Γ ) .
H× 2 (divΓ ,Γ ) ×

(59)
This theorem tells us that we can expect good approximation properties, but these
cannot be obtained using local interpolation operators.

9 Galerkin Discretization
−1
The Galerkin approach simply consists of replacing the Hilbert spaces H × 2 (divΓ , Γ )
and H × (divΓ , Γ ) in the variational formulations by finite dimensional subspaces
W h.

9.1 Integral equations of the first kind

First, we study the simplest BIE of the first kind, namely the electric field integral
equations (42) and (49), that is, we examine variational problems like: seek λ ∈
−1
H × 2 (divΓ , Γ ) such that
−1
a(λ, µ) := ⟨Sκ λ, µ⟩τ ,Γ = r.h.s.(µ) ∀µ ∈ H × 2 (divΓ , Γ ) , (60)
for a suitable continuous functional on the right hand side. If κ stays away from inte-
−1
rior electric Maxwell eigenvalues, we saw that the operator Sκ : H × 2 (divΓ , Γ ) -→
34 A. Buffa and R. Hiptmair
− 12
H × (divΓ , Γ ) defines an isomorphism. This is equivalent to the existence of a
constant CS > 0 such that the following continuous inf-sup condition holds true:
|a(µ, η)| −1
sup ≥ CS ∥µ∥ −1 ∀µ ∈ H × 2 (divΓ , Γ ) .
−1 ∥η∥ − 12 H× 2 (divΓ ,Γ )
η∈H × 2 (divΓ ,Γ ) H × (divΓ ,Γ )
(61)

We aim at establishing a uniform discrete inf-sup-condition of the form: there exists

CD > 0 such that ∀µh ∈ W h ,
|a(µh , ηh )|
sup ≥ CD ∥µh ∥ −1 , h∈H. (62)
η h ∈W h ∥η h ∥ − 12 H × 2 (divΓ ,Γ )
H × (divΓ ,Γ )

According to Babuška’s theory [8] refined in [63] this guarantees existence of dis-
crete solutions λh ∈ W h and translates into their quasi-optimal behaviour:
−1
∥λ − λh ∥ −1 ≤ CD CA inf ∥λ − η h ∥ −1 ∀h ∈ H ,
H × 2 (divΓ ,Γ ) η h ∈W H × 2 (divΓ ,Γ )
(63)

where CA > 0 is the operator norm of a(·, ·). As a first step towards a discrete
inf-sup condition, we have to find a suitable candidate for η in (61). To that end,
−1 −1
introduce the operator T : H × 2 (divΓ , Γ ) -→ H × 2 (divΓ , Γ ) through
−1 −1
a(η, Tµ) = cΓ (µ, η) ∀η ∈ H × 2 (divΓ , Γ ), µ ∈ H × 2 (divΓ , Γ ) ,

where cΓ is the compact bilinear form of Lemma 10. Owing to (61) this is a valid
definition of a compact operator T. It is immediate from (61) and Lemma 10 that

|a(µ, (XΓ + T)µ)| = |a(µ, XΓ µ) + cΓ (µ, µ)| ≥ CG ∥µ∥2 −1 (64)

H × 2 (divΓ ,Γ )

−1
for all µ ∈ H × 2 (divΓ , Γ ). The choice η := (XΓ + T)µ will make (61) hold
with CS = CG . The challenge is that (XΓ + T)µh will not be a boundary element
function even for µh ∈ W h . This is clear because neither XΓ nor T may leave
the boundary element spaces invariant. It will be necessary to project XΓ µh and
Tµh back to W h . This can be achieved by applying suitable continuous projection
−1
operators PX T
h : XΓ (W h ) -→ W h , Ph : H × (div Γ , Γ ) -→ W h . Then, for an
2

arbitrary µh ∈ W h we can hope that ηh := (PX T

h ◦XΓ +Ph ◦T)µh is an appropriate
choice for η h in (62). Making use of (64) we see that

|a(µh , η h )| = |a(µh , (XΓ +T)µh )−a(µh , ((Id−PX T

h )XΓ +(Id−Ph )T)µh )|. (65)

2
We know that |a(µh , (XΓ + T)µh )| ≥ CG ∥µh ∥H(curl;Ω) and we need to estimate
the second term in the left hand side by the triangle inequality. Obviously, the pro-
jectors PX T T
h , Ph have to guarantee uniform convergence (Id −Ph ) ◦ T|W h → 0 and
 Galerkin BEM in Electromagnetics 35
− 12
(Id −PX T
h ) ◦ XΓ |W h → 0 in H × (div Γ , Γ ) as h → 0. For Ph this is easy: we
−1
choose PTh as the H × 2 (divΓ , Γ )-orthogonal projection. Due to the compactness of
T, we know [49, Corollary 10.4] that there exists a decreasing function ϵ = ϵ(h)
−1
such that limh→0 ϵ(h) = 0 and, for all µh ∈ H × 2 (divΓ , Γ ),
0 0
0(Id −PTh ) ◦ Tµh 0 − 1 < ϵ(h) ∥µh ∥ − 21 . (66)
2
H × (divΓ ,Γ ) H × (divΓ ,Γ )

As regards PX X
h , a crucial hint lies in the observation that Ph acts on functions in
Γ
XΓ (W h ). From divΓ R µ = divΓ µ we conclude that divΓ (XΓ (W h )) ⊂ Qh . We
see that PXh has to be applied to functions with discrete divΓ only. We remind of
Lemma 16, which bears out that XΓ (W h ) is contained in the domain of the local
interpolation operators Πh . We discover that a perfectly valid candidate for PXh is
the local interpolation operator: PXh := Πh . Then, Lemma 16 is the key to uniform
convergence (Id −PX h )XΓ |W h → 0.

Lemma 17. There is a C∗ = C∗ (Ω, p, shape regularity) > 0 such that for all µh ∈
Wh
0 0
0(Id −PX 0
h )XΓ µh H × (divΓ ,Γ ) ≤ C∗ h
1/2
∥divΓ µh ∥ − 12 . (67)
H (Γ )

Proof. Note that (Id −Πh )XΓ µh = (Id −Πh )(2RΓ − Id)µh = 2(Id −Πh )RΓ µh ,
and that divΓ RΓ µh = divΓ µh . Thus, for the estimate of the L2t norm, we need
only combine Lemma 16 (applied to RΓ µh ) with (20). The observation, based on
the commuting diagram property (57), that

divΓ ((Id −Πh )XΓ µh ) = (Id −Qh )divΓ (XΓ µh ) = (Id −Qh )divΓ µh = 0

finishes the proof. ⊓

9

Using (66) and (67) in (65), we obtain:

1 2
|a(µh , η h )| ≥ (CG − CA (ϵ(h) + C∗ h 2 )) ∥µh ∥ −1 .
H × 2 (divΓ ,Γ )

1 1
This means that for h small enough to ensure 1 − CA (ϵ(h) + C∗ h 2 )/CG > 2 we
have the discrete inf-sup condition (62). This yields the main result:
Theorem 15. Provided that κ2 is not an interior electric eigenvalue, there is a
h∗ > 0, depending on the parameters of the continuous problem and the shape-
regularity of the triangulation, such that a unique solution λh ∈ W h of the dis-
cretized problem (60) exists, provided that h < h∗ . It supplies an asymptotically
optimal approximation to the continuous solution λ of (60) in the sense of (63).
Exactly the same arguments apply to (48) and give us an analogue of Thm. 15 for
the Galerkin BEM discretization in the case of the transmission problem.
36 A. Buffa and R. Hiptmair

Remark 2. Asymptotic quasi-optimality alone does not provide information about

the actual speed of convergence as h → 0, unless we have information about the
smoothness of λ. To assess the regularity of λ it is necessary to recall its meaning
as a boundary value or the jump of a trace of Maxwell solutions. Then the results
on the regularity of Maxwell solutions given in [34] can be used. Ultimately we will
always have λ ∈ H s× (divΓ , Γ ) for some s > 0 depending on the excitation and
the geometry of Γ . In combination with Thm. 14 we can predict asymptotic rates of
convergence for the h-version of the Galerkin boundary element schemes.

Remark 3. A striking difference between (42) and (49) is the choice of unknowns.
In the indirect method λ is a jump. Hence, when solving the boundary integral
equations on a polyhedron, the unknown of the indirect method will be affected by
the corner and edge singularities of both interior and exterior Maxwell solutions
[34]. As any edge is re-entrant when seen from either Ωs or Ω ′ , the jump [γN ]Γ (e)
will invariably possess a very low regularity. As a consequence, it might be much
harder to approximate by boundary elements than the unknown of the direct method.

In terms of Galerkin discretization the CFIE from Sect. 7.2 poses an extra dif-
fculty, because of the composition of the integral operator Cκ and the smooth-
ing operator M. The usual trick to avoid such operator products is to switch to a
mixed formulation introducing the new unknown p := Mζ. If we use the particular
smoothing operator from (54), we get p ∈ H × (divΓ , Γ ) and may simply incor-
−1
porate (54) into the eventual mixed variational problem: find ζ ∈ H × 2 (divΓ , Γ ),
−1
p ∈ H × (divΓ , Γ ) such that for all µ ∈ H × 2 (divΓ , Γ ), q ∈ H × (divΓ , Γ ),
. 1 / . /
−iη ⟨Sκ ζ, µ⟩τ ,Γ + ( 2 Id −Cκ )p, µ τ ,Γ = γt+ ei , µ τ ,Γ ,
(68)
⟨q, ζ⟩τ ,Γ − ⟨p, q⟩0;Γ − ⟨divΓ p, divΓ q⟩0;Γ = 0 .

−1
Thanks to the compact embedding H × (divΓ , Γ ) (→ H × 2 (divΓ , Γ ) the off-
diagonal terms in (68) are compact. Thus, a generalized Gårding inequality is im-
mediate from Lemma 10. As far as the analysis of the Galerkin discretization in
W h ×W h is concerned, we only need to deal with the diagonal terms in (68): using
exactly the same arguments as above we conclude the quasi-optimality of Galerkin
solutions on sufficiently fine meshes. Please note that the estimates now employ the
−1
norm of the product space H × 2 (divΓ , Γ ) × H × (divΓ , Γ ). Thus, asymptotic rates
of convergence will depend on the smoothness of both ζ and p.

9.2 Integral equations of the second kind

Pitfalls have to be avoided when performing a Galerkin boundary element dis-

cretization of the Fredholm integral equation of the second kind (44). A straight-
forward Galerkin discretization would lead to: seek λh ∈ W h such that
. 1 /
( 2 Id +Cκ )λh , µh τ ,Γ = r.h.s.(µh ) ∀µh ∈ W h . (69)
 Galerkin BEM in Electromagnetics 37

If Γ ∈ C ∞ , then existence and uniqueness of solutions of the continuous variational

problem are clear from Theorem 11. However, this does not necessarily remain true
for (69). The cause of the difficulties is the failure of Theorem 2 to hold in the
discrete setting. In other words, W h may not be dual to itself with respect to the
pairing ⟨·, ·⟩τ ,Γ . More precisely, in [29, Section 3.1] it has been shown by means of
Hodge decompositions that for RT0 boundary elements and quasiuniform families
of surface meshes Γh there is α > 0 and spaces Kh ⊂ W h such that, for all h ∈ H,
dim Kh ≥ α dim W h and
| ⟨µh , ξ h ⟩τ ,Γ | 1
∀µh ∈ Kh sup ≤ Ch 2 ∥µh ∥ −1 .
ξh ∈W h ∥ξh ∥ −1 H × 2 (divΓ ,Γ )
H × 2 (divΓ ,Γ )

The discretization of ⟨·, ·⟩τ ,Γ on W h is not stable! This bars us from deriving a dis-
crete inf-sup condition, though the continuous bilinear form satisfies a generalized
Gårding inequality.

Remark 4. The instability of ⟨·, ·⟩τ ,Γ in W also thwarts the straightforward applica-
tion of an otherwise effective preconditioning strategy for boundary integral equa-
tions of the first kind, which is based on Calderón projectors [59]. The gist of the
remedy, devised in [28], is to express ⟨·, ·⟩τ ,Γ via an approximate discrete Hodge
decomposition of W.

This instability forces us to switch from (44) to (45), before a Galerkin dis-
cretization by means of W h becomes feasible: the stability of the T H(divΓ , Γ ) in-
ner product in the discrete setting is a moot point. Hence, provided that the assump-
tions of Theorem 11 hold, the Galerkin discretization in W h will produce asymptot-
ically optimally convergent solutions on sufficiently fine meshes. The proof follows
the standard approach to coercive variational problems [62]. However, note that all
estimates will be based on the H × (divΓ , Γ )-norm, that is

∥λ − λh ∥H × (divΓ ,Γ ) ≤ C inf ∥λ − ηh ∥H × (divΓ ,Γ ) .

η∈W h

We point out that the bilinear expressions (λh , µh ) -→ ⟨divΓ Cκ λh , divΓ µh ⟩0;Γ
that have to be evaluated for basis functions of W h in order to get the system matrix
can be converted into sums of two integrals over Γ × Γ featuring weakly singular
kernels. Details can be found in the proof of Lemma 11.

10 Coupling of Finite Elements and Boundary Elements

The solution of the transmission problem of electromagnetic scattering by means of
direct boundary integral equations is confined to the case of homogeneous scatter-
ers, because the simple representation formula (26) for Maxwell solutions cannot
accommodate variable material coefficients ϵ = ϵ(x), µ = µ(x), x ∈ Ωs . This sit-
uation poses no problems for a Galerkin finite element discretization of the spatial
38 A. Buffa and R. Hiptmair

variational problem inside Ωs . On the other hand, the field problem in the air region
Ω ′ is not amenable to a treatment by classical finite elements, but offers a perfect
setting for the boundary element methods discussed in the previous sections. Thus,
it is natural to tackle scattering at an inhomogeneous body by a combined Galerkin
discretization involving both finite elements and boundary elements. In this section
the focus will be on a method based on the Calderon projector P+ κ from (34).
Using aM (·, ·) defined in (15), the electric field in Ωs satisfies
. /
aM (e, v) − µ−1 γt− curl e, γt− v τ ,Γ = 0 (70)

for all v ∈ H(curl; Ωs ). The gist of coupling is to employ an operator representa-

−1 −1
tion of the Dirichlet-to-Neumann map DtN+ κ : H × (div Γ , Γ ) -→ H × (div Γ , Γ ),
2 2

which is a linear operator returning γN e for a Maxwell solution e in Ω ′ if γt+ e is

+

prescribed. If this was available, we could use the transmission conditions

γt− e = γt+ e + γt+ ei , µ−1 γt− curl e = κ +

µ0 γN e + γt+ hi , (71)

to cast the scattering problem in the variational form: seek e ∈ H(curl; Ωs ) such
that for all v ∈ H(curl; Ωs )
. /
aM (e, v) − κµ−1
0 DtN+ − −
κ γt e, γt v τ ,Γ = r.h.s(v) .

By Thm. 8 either row of the interior Calderon projector P−

κ immediately supplies a
realization of DtN+
κ :

DtN+ 1
κ = −( 2 Id +Cκ )
−1
Sκ , DtN+ −1 1
κ = −Sκ ( 2 Id +Cκ ) . (72)

Both formulae describe the same operator, but appear vastly different. The reason is
that they both break the inherent symmetry of magnetic and electric fields. Symme-
try can be preserved by combining both rows of P+κ in a clever manner:

DtN+ 1 −1 1
κ = −Sκ − ( 2 Id −Cκ )Sκ ( 2 Id +Cκ ) . (73)

This discovery was first presented in [32] and is the foundation for the so-called
symmetric approach to marrying finite elements and boundary elements. It has been
applied to a wide range of transmission problems, see, for instance [22, 23, 50]. In
the case of electromagnetism the idea was examined theoretically in [4–6], and in
[9] for a related problem involving impedance boundary conditions.
Of course, a variational formulation suited for Galerkin discretization has to
dispense with the explicit inverse S−1
κ . Instead another equation is added, which
−1
leads to: seek e ∈ H(curl; Ωs ), λ ∈ H × 2 (divΓ , Γ ) with
5 6 5 6
aM (e, e′ ) + µκ0 Sκ γt− e, γt− e′ − µκ0 ( 12 Id −Cκ )λ, γt− e′ = ...,
τ ,Γ τ ,Γ (74)
. 1 / . /
( 2 Id +Cκ )γt− e, λ′ τ ,Γ + Sκ λ, λ′ τ ,Γ = ...,
 Galerkin BEM in Electromagnetics 39
− 12
for all e′ ∈ H(curl; Ωs ), λ′ ∈ H × (divΓ , Γ ). The new unknown λ will provide
+
the exterior Neumann trace γN e.
Note that the symmetric version of DtN+κ involves the inverse of Sκ . This sug-
gests that “forbidden wave numbers” will also haunt the coupled formulations, cf.
Sect. 7. Similar to Lemma 14 one proves the following theorem, see [44].
Theorem 16. If κ2 is not an interior electric eigenvalue, a solution (e, λ) of
(74) provides a solution of the transmission problem (1) by retaining e in Ωs
and using the exterior Stratton-Chu representation formula (26) with the data
(γt− e − γt+ ei , λ).

Corollary 6. If κ2 is not an interior electric eigenvalue, the solution (e, λ) of (74)

is unique.

We point out that even if κ violates the assumption of the theorem, the solution for
e will remain unique. This will no longer be true for λ, which is unique only up
to Neumann traces of interior electric eigenmodes. This can be seen by refining the
arguments in the proof of Theorem 16.
−1
We denote by dκ the bilinear form on H(curl; Ωs ) × H × 2 (divΓ , Γ ) that is
associated with the the variational problem (74). Pursuing the same policy as in
Sect. 3 and 7, we aim to establish a generalized Gå rding inequality for dκ . Of
course, the splitting idea will pave the way. More precisely, the crucial “sign flipping
isomorphism” XV will involve both splittings (18) and (21) employed in Sects. 3
−1
and 7. Writing, V := H(curl; Ωs ) × H × 2 (divΓ , Γ ), it reads
, - , -
u (R − Z)u
XV := : V -→ V .
ξ (RΓ − ZΓ )ξ

We make the important observation that the trace γt− maps curl-free vectorfields
into N (Γ ). In addition we can use the symmetry of Cκ stated in lemma 6 and
proceed as in the proof of Thm. 9. This will give us the desired strengthened Gå
rding inequality:
Theorem 17. There exists a compact bilinear form c : V × V -→ C and a constant
CG > 0 such that
7 ,, - , -- ,, - , --7
7 u u u u 7
7dκ , X − c , 7≥
7 µ
V
µ µ µ 7
, -
≥ CG ∥u∥2H(curl;Ωs ) + ∥µ∥2 −1
H× 2 (divΓ ,Γ )

−1
holds for all u ∈ H(curl; Ωs ), µ ∈ H × 2 (divΓ , Γ ).
Hence, in conjuction with Cor. 6, a Fredholm alternative argument confirms the
existence of solutions of the variational problem (74).
40 A. Buffa and R. Hiptmair

Besides Γh the Galerkin discretization of (74) requires a triangulation Ωh of Ωs .

In principle, both can be independent of each other, but implementation is greatly fa-
cilitated if Γh = Ωh|Γ . Then, we can rely on the H × (divΓ , Γ )-conforming bound-
ary element spaces W h to approximate λ, and special H(curl; Ωs )-conforming
finite elements for e. The latter are thoroughly discussed in [45, Ch. 3]. They enjoy
all the properties that permit us to prove a discrete inf-sup-condition as in Sect. 9.1.
Thus we can get asymptotic quasi-optimality of discrete solutions obtained by the
symmetric coupling of finite elements and boundary elements for the electromag-
netic scattering problem.

References
1. R. A DAMS, Sobolev Spaces, Academic Press, New York, 1975.
2. A. A LONSO AND A. VALLI, Some remarks on the characterization of the space of tan-
gential traces of H(rot; Ω) and the construction of an extension operator, Manuscripta
mathematica, 89 (1996), pp. 159–178.
3. H. A MMARI AND S. H E , Effective impedance boundary conditions for an inhomoge-
neous thin layer on a curved metallic surface, IEEE Trans. Antennas and Propagation,
46 (1998), pp. 710–715.
4. H. A MMARI AND J. N ÉD ÉLEC, Couplage éléments finis-équations intégrales pour la
résolution des équations de Maxwell en milieu hétérogene, in Equations aux derivees
partielles et applications. Articles dedies a Jacques-Louis Lions, Gauthier-Villars, Paris,
1998, pp. 19–33.
5. H. A MMARI AND J.-C. N ÉD ÉLEC, Coupling of finite and boundary element methods
for the time-harmonic Maxwell equations. II: A symmetric formulation, in The Maz’ya
anniversary collection. Vol. 2, J. Rossmann, ed., vol. 110 of Oper. Theory, Adv. Appl.,
Birkhäuser, Basel, 1999, pp. 23–32.
6. , Coupling integral equations method and finite volume elements for the resolution
of the Leontovich boundary value problem for the time-harmonic Maxwell equations
in three dimensional heterogeneous media, in Mathematical aspects of boundary ele-
ment methods. Minisymposium during the IABEM 98 conference, dedicated to Vladimir
Maz’ya on the occasion of his 60th birthday on 31st December 1997, M. Bonnet, ed.,
vol. 41 of CRC Research Notes in Mathematics, CRC Press, Boca Raton, FL, 2000,
pp. 11–22.
7. C. A MROUCHE , C. B ERNARDI , M. DAUGE , AND V. G IRAULT , Vector potentials in
three–dimensional nonsmooth domains, Math. Meth. Appl. Sci., 21 (1998), pp. 823–864.
8. I. BABU ŠKA, Error bounds for the finite element method, Numer. Math., 16 (1971),
pp. 322–333.
9. A. B ENDALI, Boundary element solution of scattering problems relative to a generalized
impedance boundary condition, in Partial differential equations: Theory and numerical
solution. Proceedings of the ICM’98 satellite conference, Prague, Czech Republic, Au-
gust 10-16, 1998., W. Jäger, ed., vol. 406 of CRC Res. Notes Math., Boca Raton, FL,
2000, Chapman & Hall/CRC, pp. 10–24.
10. A. B ENDALI AND L. V ERNHET , The Leontovich bondary value problem and its bound-
ary integral equations solution, Preprint, CNRS-UPS-INSA, Department de Genie Math-
ematique, Toulouse, France, 2001.
11. H. B REZIS, Analyse fonctionnelle. Théorie et applications, Masson, Paris, 1983.
 Galerkin BEM in Electromagnetics 41

12. F. B REZZI , J. D OUGLAS , AND D. M ARINI, Two families of mixed finite elements for
2nd order elliptic problems, Numer. Math., 47 (1985), pp. 217–235.
13. F. B REZZI AND M. F ORTIN, Mixed and hybrid finite element methods, Springer, 1991.
14. A. B UFFA, Hodge decompositions on the boundary of a polyhedron: The multiconnected
case, Math. Mod. Meth. Appl. Sci., 11 (2001), pp. 1491–1504.
15. , Traces theorems for functional spaces related to Maxwell equations: An over-
wiew. To appear in Proceedings of the GAMM Workshop on Computational Electro-
magnetics, Kiel, January 26th - 28th, 2001.
16. A. B UFFA AND S. C HRISTIANSEN, The electric field integral equation on Lips-
chitz screens: Definition and numerical approximation, Numer. Mathem. (2002), DOI
10.1007/s00211-002-0422-0.
17. A. B UFFA AND P. C IARLET, J R ., On traces for functional spaces related to Maxwell’s
equations. Part I: An integration by parts formula in Lipschitz polyhedra., Math. Meth.
Appl. Sci., 24 (2001), pp. 9–30.
18. , On traces for functional spaces related to Maxwell’s equations. Part II: Hodge
decompositions on the boundary of Lipschitz polyhedra and applications, Math. Meth.
Appl. Sci., 21 (2001), pp. 31–48.
19. A. B UFFA , M. C OSTABEL , AND C. S CHWAB, Boundary element methods for Maxwell’s
equations on non-smooth domains, Numer. Mathem. 92 (2002) 4, pp. 679-710.
20. A. B UFFA , M. C OSTABEL , AND D. S HEEN, On traces for H(curl, Ω) in Lipschitz
domains, J. Math. Anal. Appl., 276/2 (2002), pp. 845-876.
21. A. B UFFA , R. H IPTMAIR , T. VON P ETERSDORFF , AND C. S CHWAB, Boundary element
methods for Maxwell equations on Lipschitz domains, Numer. Math., (2002). To appear.
22. C. C ARSTENSEN, A posteriori error estimate for the symmetric coupling of finite ele-
ments and boundary elements, Computing, 57 (1996), pp. 301–322.
23. C. C ARSTENSEN AND P. W RIGGERS, On the symmetric boundary element method and
the symmetric coupling of boundary elements and finite elements, IMA J. Numer. Anal.,
17 (1997), pp. 201–238.
24. M. C ESSENAT , Mathematical Methods in Electromagnetism, vol. 41 of Advances in
Mathematics for Applied Sciences, World Scientific, Singapore, 1996.
25. G. C HEN AND J. Z HOU, Boundary Element Methods, Academic Press, New York, 1992.
26. S. C HRISTIANSEN, Discrete Fredholm properties and convergence estimates for the
EFIE, Technical Report 453, CMAP, Ecole Polytechique, Paris, France, 2000.
27. , Mixed boundary element method for eddy current problems, Research Report
2002-16, SAM, ETH Zürich, Zürich, Switzerland, 2002.
28. S. C HRISTIANSEN AND J.-C. N ÉD ÉLEC, Des prćonditionneurs pour la résolution
numérique des équations intégrales de frontiére de l’electromagnétisme, C.R. Acad. Sci.
Paris, Ser. I Math, 31 (2000), pp. 617–622.
29. , A preconditioner for the electric field integral equation based on Calderón for-
mulas. To appear in SIAM J. Numer. Anal.
30. P. C IARLET , The Finite Element Method for Elliptic Problems, vol. 4 of Studies in Math-
ematics and its Applications, North-Holland, Amsterdam, 1978.
31. D. C OLTON AND R. K RESS, Inverse Acoustic and Electromagnetic Scattering Theory,
vol. 93 of Applied Mathematical Sciences, Springer, Heidelberg, 2nd ed., 1998.
32. M. C OSTABEL , Symmetric methods for the coupling of finite elements and boundary ele-
ments, in Boundary Elements IX, C. Brebbia, W. Wendland, and G. Kuhn, eds., Springer,
Berlin, 1987, pp. 411–420.
33. , Boundary integral operators on Lipschitz domains: Elementary results, SIAM J.
Math. Anal., 19 (1988), pp. 613–626.
42 A. Buffa and R. Hiptmair

34. M. C OSTABEL AND M. DAUGE , Singularities of Maxwell’s equations on polyhedral

domains, in Analysis, Numerics and Applications of Differential and Integral Equations,
M. Bach, ed., vol. 379 of Longman Pitman Res. Notes Math. Ser., Addison Wesley,
Harlow, 1998, pp. 69–76.
35. , Maxwell and Lamé eigenvalues on polyhedra, Math. Methods Appl. Sci., 22
(1999), pp. 243–258.
36. R. DAUTRAY AND J.-L. L IONS, Mathematical Analysis and Numerical Methods for
Science and Technology, vol. 4, Springer, Berlin, 1990.
37. A. DE L A B OURDONNAYE , Some formulations coupling finite element amd integral
equation method for Helmholtz equation and electromagnetism, Numer. Math., 69
(1995), pp. 257–268.
38. L. D EMKOWICZ , Asymptotic convergence in finite and boundary element methods: Part
1, Theoretical results, Comput. Math. Appl., 27 (1994), pp. 69–84.
39. V. G IRAULT AND P. R AVIART , Finite element methods for Navier–Stokes equations,
Springer, Berlin, 1986.
40. P. G RISVARD, Elliptic Problems in Nonsmooth Domains, Pitman, Boston, 1985.
41. , Singularities in boundary value problems, vol. 22 of Research Notes in Applied
Mathematics, Springer, New York, 1992.
42. W. H ACKBUSCH, Integral equations. Theory and numerical treatment., vol. 120 of In-
ternational Series of Numerical Mathematics, Birkhäuser, Basel, 1995.
43. C. H AZARD AND M. L ENOIR, On the solution of time-harmonic scattering problems
for Maxwell’s equations, SIAM J. Math. Anal., 27 (1996), pp. 1597–1630.
44. R. H IPTMAIR, Coupling of finite elements and boundary elements in electromagnetic
scattering, Report 164, SFB 382, Universität Tübingen, Tübingen, Germany, July 2001.
Submitted to SIAM J. Numer. Anal.
45. , Finite elements in computational electromagnetism, Acta Numerica, (2002),
pp. 237–339.
46. , Symmetric coupling for eddy current problems, SIAM J. Numer. Anal., 40 (2002),
pp. 41–65.
47. R. H IPTMAIR AND C. S CHWAB, Natural boundary element methods for the electric
field integral equation on polyhedra, SIAM J. Numer. Anal., 40 (2002), pp. 66–86.
48. R. K RESS, On the boundary operator in electromagnetic scattering, Proc. Royal Soc.
Edinburgh, 103A (1986), pp. 91–98.
49. , Linear Integral Equations, vol. 82 of Applied Mathematical Sciences, Springer,
Berlin, 1989.
50. M. K UHN AND O. S TEINBACH, FEM-BEM coupling for 3d exterior magnetic field prob-
lems, Math. Meth. Appl. Sci., (2002). To appear.
51. R. M C C AMY AND E. S TEPHAN, Solution procedures for three-dimensional eddy-
current problems, J. Math. Anal. Appl., 101 (1984), pp. 348–379.
52. W. M C L EAN, Strongly Elliptic Systems and Boundary Integral Equations, Cambridge
University Press, Cambridge, UK, 2000.
53. J.-C. N ÉD ÉLEC, Acoustic and Electromagnetic Equations: Integral Representations for
Harmonic Problems, vol. 44 of Applied Mathematical Sciences, Springer, Berlin, 2001.
54. A. N ETHE , R.Q UAST, AND H. S TAHLMANN, Boundary conditions for high frequency
eddy current problems, IEEE Trans. Mag., 34 (1998), pp. 3331–3334.
55. L. PAQUET , Problemes mixtes pour le systeme de Maxwell, Ann. Fac. Sci. Toulouse, V.
Ser., 4 (1982), pp. 103–141.
56. P. A. R AVIART AND J. M. T HOMAS, A Mixed Finite Element Method for Second Order
Elliptic Problems, vol. 606 of Springer Lecture Notes in Mathematics, Springer, Ney
York, 1977, pp. 292–315.
 Galerkin BEM in Electromagnetics 43

57. M. R EISSEL , On a transmission boundary-value problem for the time-harmonic

Maxwell equations without displacement currents, SIAM J. Math. Anal., 24 (1993),
pp. 1440–1457.
58. A. S CHATZ , An observation concerning Ritz-Galerkin methods with indefinite bilinear
forms, Math. Comp., 28 (1974), pp. 959–962.
59. O. S TEINBACH AND W. W ENDLAND, The construction of some efficient precondition-
ers in the boundary element method, Adv. Comput. Math., 9 (1998), pp. 191–216.
60. J. S TRATTON AND L. C HU, Diffraction theory of electromagnetic waves, Phys. Rev., 56
(1939), pp. 99–107.
61. T. VON P ETERSDORFF, Boundary integral equations for mixed Dirichlet, Neumann and
transmission problems, Math. Meth. Appl. Sci., 11 (1989), pp. 185–213.
62. W. W ENDLAND, Boundary element methods for elliptic problems, in Mathematical The-
ory of Finite and Boundary Element Methods, A. Schatz, V. Thomée, and W. Wendland,
eds., vol. 15 of DMV-Seminar, Birkhäuser, Basel, 1990, pp. 219–276.
63. J. X U AND L. Z IKATANOV, Some oberservations on Babuška and Brezzi theories, Re-
port AM222, PennState Department of Mathematics, College Park, PA, September 2000.
To appear in Numer. Math.

A Scattering from coated dielectric objects

Now, we consider a partition of Γ := ∂Ωs into a part ΓPEC covered by a perfectly
conducting coating, and a part Γo that can be penetrated by electromagnetic fields.
Further, we assume that Ωs is filled with an isotropic, homogeneous dielectric ma-
terial with:material parameters ϵ− and µ− . Hence we end up with a wave number
√
κ− := ω ϵ− µ− inside Ωs , whereas κ+ := ω ϵ0 µ0 has to be used in Ω ′ . On Γ
we find the following boundary and transmission conditions:
γt− e = γt+ e = 0 on ΓPEC , γt− e = γt+ e on Γo ,
κ− − κ+ +
µ− γN e = µ0 γN e on Γo .
We aim to establish a variational direct boundary integral equation formulation for
this transmission problem. We closely follow the approach in Sect. 7.1, p. 28, and
will reuse the notations introduced there.
First, we recall some results about traces of functions in H(curl; Ω) onto parts
of the boundary, cf. [18]. To this end, let Ω ⊂ R3 be a generic domain and Γo ⊂ ∂Ω
an open subset of its piecewise smooth Lipschitz boundary Γ . We take for granted
that Γo is a union of faces of Γ . We introduce the space
H Γ \Γo (curl; Ω) := {u ∈ H(curl; Ω), γt u = 0 on Γ \ Γo } ,
and the two trace spaces
−1 −1 1
H × 2 (divΓ , Γo ) :={µ ∈ H × 2 (Γo ), divΓ µ ∈ H − 2 (Γo )} ,
−1 −1 −1
H ×,00
2
(divΓ , Γo ) :={µ ∈ H ×,00
2
(Γo ), divΓ µ ∈ H00 2 (Γo )} .
From [18, Thm. 5.3] we get the following fundamental trace theorem. Its statement
makes use of the restriction operator ro µ := µ|Γo in the sense of distributions.
44 A. Buffa and R. Hiptmair
−1
Theorem 18. The restricted trace mapping ro ◦γt : H(curl; Ω) -→ H ×,00 2
(divΓ , Γo )
is continuous and surjective.
−1
The trace mapping γt : H Γ \Γo (curl; Ω) -→ H × 2 (divΓ , Γo ) is continuous and
surjective.
−1
We immediately infer that H × 2 (divΓ , Γo ) is the space of tangential vectorfields
−1
on Γo that yield a function in H × 2 (divΓ , Γ ) when extended by zero to Γ .
The duality for the partial trace spaces is explained in [18, Sect. 5].
−1 −1
Theorem 19. The spaces H × 2 (divΓ , Γo ) and H ×,00
2
(divΓ , Γo ) are dual to each
other with respect to the pairing ⟨·, ·⟩τ ,Γ .
As in Sect. 7.1 we introduce the scaled traces
+ −
(ζ + , λ+ ) = (γt+ e, κµ0 γN
+
e) , (ζ − , λ− ) = (γt− e, µκ− γN
−
e) .

With this notation the transmission conditions on Γo read

−1
ζ − − ζ + = γt+ ei in H × 2 (divΓ , Γo ) , (75)
− + − 21
λ −λ = γt+ hi in H ×,00 (divΓ , Γo ) . (76)

These transmission conditions are due to the fact that e denotes the total field in
Ωs , whereas in Ω ′ it refers only to the scattered field that satisfies the Silver-Müller
radiation condition at ∞.
For the sake of completeness we note that

ζ− = 0 , ζ + = −γt+ ei on ΓPEC .

Let us rewrite (47) as

< =, -
µ−
− 21 Id +Cκ− κ − Sκ − ζ−
=0, (77)
κ−
µ− Sκ−
1
− 2 Id +Cκ− λ−
< =, -
− 12 Id −Cκ+ − κµ+0 Sκ+ ζ+
κ+ =0. (78)
1
− µ0 Sκ+ − 2 Id −Cκ+ λ+

From these equations we aim to derive the crucial electric to magnetic maps, which
provide the Poincaré-Stekhlov operators for electromagnetic scattering. First, we
use the second equations in (77) and (78) and get

λ− = ( µκ− Sκ− )ζ − + ( 21 Id +Cκ− )λ− ,

−
(79)
+
λ+ = (− κµ0 Sκ+ )ζ + + ( 12 Id −Cκ+ )λ+ . (80)

Then, we rely on the first equations to eliminate the magnetic traces remaining on
the right hand side:
 Galerkin BEM in Electromagnetics 45
−
> −
?
λ− = µκ− Sκ− − ( 12 Id +Cκ− )( µκ− Sκ− )−1 (− 12 Id +Cκ− ) ζ − , (81)
> + ?
λ+ = − κµ0 Sκ+ + ( 12 Id −Cκ+ )( κµ+0 Sκ+ )−1 (− 12 Id −Cκ+ ) ζ + . (82)

Strictly speaking, this formal manipulation is only valid, if there are no interior
resonances, because invertibility of both Sκ− and Sκ+ has to be guaranteed, see
Sect. 7.1. For the sake of brevity let us introduce the operators
−
P− := κ−
µ− Sκ− − ( 21 Id +Cκ− )( µκ− Sκ− )−1 (− 12 Id +Cκ− ) ,
+
P+ := − κµ0 Sκ+ + ( 12 Id −Cκ+ )( κµ+0 Sκ+ )−1 (− 12 Id −Cκ+ ) .
By Cor. 2 they map continuously
−1 −1
P− , P+ : H × 2 (divΓ , Γ ) -→ H × 2 (divΓ , Γ ) .
Moreover, the derivation of P− and P+ confirms that barring interior resonances,
(ζ − , P− ζ − ) and (ζ + , P+ ζ + ) will be interior/exterior electromagnetic Cauchy data.
Using these operators, the transmission condition (76) can be stated as
−1
P− ζ − − P+ ζ + = γt+ hi in H ×,00
2
(divΓ , Γo ) .
Plugging in (75) and retaining ζ := ζ − as unknown, this can be cast into the equiv-
alent variational form
. − / . / −1
(P ζ − P+ (ζ − γt+ ei ), µ τ ,Γ = γt+ hi , µ τ ,Γ ∀µ ∈ H × 2 (divΓ , Γo ) .
(83)
The presence of inverse operators in the definitions of P− and P+ makes (83) unsuit-
able for a direct Galerkin discretization. The usual trick to avoid these undesirable
inverses is to use (79) and (80) and switch to a mixed formulation:
−
P− ζ = ( κµ− Sκ− )ζ + ( 12 Id +Cκ− )λ− , λ− := −( µκ− Sκ− )−1 (− 12 Id +Cκ− )ζ ,
−

+
P+ ζ + = (− κµ0 Sκ+ )ζ + + ( 12 Id −Cκ+ )λ+ , λ+ := ( κµ+0 Sκ+ )−1 (− 12 Id −Cκ+ )ζ + ,
−1
where λ− , λ+ ∈ H × 2 (divΓ , Γ ) can be regarded as auxiliary unknowns de-
−1
fined on all of Γ . Hence, (83) is equivalent to: seek ζ ∈ H ×,00
2
(divΓ , Γo ),
−1 −1
λ− ∈ H × 2 (divΓ , Γ ), λ+ ∈ H × 2 (divΓ , Γ ) such that
5 − +
6
( µκ− Sκ− )ζ + ( κµ0 Sκ+ )(ζ −.γt+ ei ), µ + /
τ ,Γ
+ ( 12 Id +Cκ− )λ− , µ τ ,Γ +
. / . /
− ( 21 Id −Cκ+ )λ+ , µ τ ,Γ = γt+ hi , µ τ ,Γ ,
. 1 / 5 −
6
(− 2 Id +Cκ− )ζ, τ τ ,Γ + ( µκ− Sκ− )λ− , τ =0,
. 1 +
/ . µ0 +
/ τ ,Γ
( 2 Id +Cκ+ )(ζ − γt ei ), θ τ ,Γ + ( κ+ Sκ+ )λ , θ τ ,Γ =0.
(84)
−1 −1 −1
for all µ ∈ H × 2 (divΓ , Γo ), τ ∈ H × 2 (divΓ , Γ ), θ ∈ H × 2 (divΓ , Γ ).
46 A. Buffa and R. Hiptmair

Lemma 18. The variational problem (84) has a unique solution (ζ − , λ+ , λ− ) ∈

−1 −1 −1
H × 2 (divΓ , Γo ) × H × 2 (divΓ , Γ ) × H × 2 (divΓ , Γ ), provided that κ+ does not
coincide with an interior electric Maxwell eigenvalue of Ωs .
Proof. We study a solution (ζ, λ+ , λ− ) of the homogeneous system with γt+ ei =
and γt+ hi = 0. Then set
, @+ - < =, -
ζ 1 µ−
Id −C κ − − Sκ − ζ
:= 2 κ− ; κ −
, (85)
λ@+ − µ− Sκ− 12 Id −Cκ− λ−
, @− - < =, -
ζ 1
2 Id +Cκ+ κµ+0 Sκ+ ζ
:= , (86)
λ@− κ− ;
µ0 Sκ +
1
2 Id +C κ + λ +

Please note that the operators in (85) and (86) are the (scaled) exterior Calderón
projector for the interior wave number κ− and the interior Calderón projector for
% eζ + &
the exterior wave number κ+ . This means that λe + are exterior Maxwell Cauchy
% eζ − &
data, whereas λe − turn out to be interior Maxwell Cauchy data.
From the second and third equation of (84) with zero r.h.s. it is immediate that
−
ζ@ = ζA
+
=0.

@ + = 0. If κ+
Thus, the unique solvability of the exterior scattering problem yields λ
is different from an interior electric Maxwell eigenvalue, then we can also conclude
λ− = 0.
Hence, we have shown
< −
=, - , -
1
2 Id +Cκ− µκ− Sκ− ζ ζ
− = ,
κ− ;
µ− Sκ−
1
2 Id +Cκ−
λ λ−
< =, - , -
1
−Cκ+ − κµ+0 Sκ+
2 Id − ζ ζ
; κ+ 1 Id −Cκ+ = .
− κµ0 S 2 λ +
λ +

% &
This means that λζ− are Cauchy data for the the interior scattering problem with
% &
wave number κ− and λζ+ play the same role for an exterior scattering problem
with wave number κ+ .
Moreover, from the first equation of (84) we can infer that

; κ− +
λ− − λ+ = ( µκ− S
−
κ− ;
µ0 Sκ+ )ζ+

+ ( 12 Id +Cκ− )λ− − ( 12 Id −Cκ+ )λ+ = 0 on Γo .

Summing up, the boundary data (ζ, λ− , λ+ ) are the traces of the electric field and
the magnetic field, respectively, that solve the scattering problem for the coated
dielectric object Ω. Since we considered the case of zero excitation, the unique
solvability of the scattering problem enforces ζ = λ− = λ+ = 0.
 Galerkin BEM in Electromagnetics 47

The bilinear form a associated with the variational problem (84) reads
⎛⎛ ⎞ ⎛ ⎞⎞
ζ µ 5 − 6
+
a ⎝⎝λ− ⎠ , ⎝τ ⎠⎠ := ( µκ− Sκ− + κµ0 Sκ+ )ζ, µ −
+ τ ,Γ
λ θ . /
− ( 21 Id +Cκ− )λ− , µ τ ,Γ +
. /
+ ( 21 Id −Cκ+ )λ+ , µ τ ,Γ +
. /
+ (− 21 Id +Cκ− )ζ, τ τ ,Γ +
. /
+ ( 21 Id +Cκ+ )ζ, θ τ ,Γ +
5 − 6
+ ( µκ− Sκ− )λ− , τ +
τ ,Γ
. /
+ ( κµ+0 Sκ+ )λ+ , θ τ ,Γ ,

Lemma 19. The bilinear form a is satisfies a generalized Gårding inequality on

−1
(H × 2 (divΓ , Γ ))3 .

Wrong sign somewhere

Proof.