Computers and Mathematics with Applications 71 (2016) 85104

Contents lists available at ScienceDirect

Computers and Mathematics with Applications

journal homepage: www.elsevier.com/locate/camwa

On the approximation of electromagnetic fields by edge finite

elements. Part 1: Sharp interpolation results for
low-regularity fields
Patrick Ciarlet Jr.
POEMS, ENSTA ParisTech, CNRS, INRIA, Universit Paris-Saclay, 828 Bd des Marchaux, 91762 Palaiseau Cedex, France

article info abstract

Article history: We propose sharp results on the numerical approximation of low-regularity electromag-
Received 16 July 2015 netic fields by edge finite elements. We consider general geometrical settings, includ-
Received in revised form 1 October 2015 ing topologically non-trivial domains or domains with a non-connected boundary. In the
Accepted 18 October 2015
model, the electric permittivity and magnetic permeability are symmetric, tensor-valued,
Available online 28 November 2015
piecewise smooth coefficients. In all cases, the error can be bounded by h times a constant,
where h is the meshsize, for some exponent ]0, 1] that depends both on the geometry
Keywords:
Maxwells equations
and on the coefficients. It relies either on classical estimates when > 1/2, or on a new
Interface problem combined interpolation operator when < 1/2. The optimality of the value of is dis-
Edge elements cussed with respect to abstract shift theorems. In some simple configurations, typically for
Interpolation operators scalar-valued permittivity and permeability, the value of can be further characterized.
Error estimates This paper is the first one in a series dealing with the approximation of electromagnetic
fields by edge finite elements.
2015 Elsevier Ltd. All rights reserved.

1. Introduction

The aim of this paper is to study the numerical approximation of electromagnetic fields, governed by Maxwells equations
with volume sources in bounded regions of R3 . More precisely, we are interested in exhibiting the approximation capabilities
of those fields with the help of edge element interpolation operators. Typically, the domain under scrutiny is bounded and
enclosed in a perfect conductor, and it can be made of different materials. In particular, we shall provide interpolation results
that depend on the geometry of the domain, on the electric permittivity and the magnetic permeability that describe the
materials, and also on the regularity of the sources, that is the current and charge densities. Special attention will be devoted
to cases where the regularity of the fields is minimal.
In the next section, we begin by recalling a model problem in electromagnetic theory, namely the time-harmonic Maxwell
equations set in a bounded domain. We recall equivalent variational formulations and well-posedness results, and the
approximation by edge elements. To obtain a priori convergence estimates, we then study the minimal regularity of those
fields: this is the main topic of Section 3. The regularity results are derived thanks to a splitting of the fields and their curls
into a regular part and a gradient. In Section 4, we study in detail the approximability by edge finite elements of the fields. We
review the classical interpolation results, before we define a new, combined, interpolation operator which relies explicitly
on the splitting of the fields, and not only on the minimal regularity. We conclude this section by a comparison with the

E-mail address: patrick.ciarlet@ensta-paristech.fr.

URL: http://www.ensta.fr/ciarlet.

http://dx.doi.org/10.1016/j.camwa.2015.10.020
0898-1221/ 2015 Elsevier Ltd. All rights reserved.
86 P. Ciarlet Jr. / Computers and Mathematics with Applications 71 (2016) 85104

more recent quasi-interpolation theory. As a result of the approximability properties, we finally derive in Section 5 optimal
error estimates.
Throughout the paper, C is used to denote a generic positive constant which is independent of the meshsize, the
triangulation and the fields of interest. On the other hand, C may depend on the geometry of the domain, or on the
coefficients defining the model. We also use the shorthand notation A . B for the inequality A CB, where A and B
are two scalar fields, and C is a generic constant. Respectively, A h B for the inequalities A . B and B . A. We denote
constant fields by the symbol cst. Vector-valued (resp. tensor-valued) function spaces are written in boldface character (resp.
blackboard bold characters); for the latter, the index sym indicates symmetric fields. Given an open set O of R3 , we use the
notation (|)0,O (respectively 0,O ) for the L2 (O ) and the L 2 (O ) := (L2 (O ))3 hermitian scalar products (resp. norms).
More generally, (|)s,O and s,O (respectively | |s,O ) denote the hermitian scalar product and the norm (resp. semi-norm)
of the Sobolev spaces H s (O ) and H s (O ) := (H s (O ))3 for s R (resp. for s > 0). The index zmv indicates zero-mean-value
fields. If moreover the boundary O is Lipschitz, n denotes the unit outward normal vector field to O . Finally, it is assumed
that the reader is familiar with function spaces related to Maxwells equations, such as H (curl; O ), H0 (curl; O ), H (div ; O ),
and H0 (div ; O ). We refer to the monograph of Monk [1] for details. We will define more specialized function spaces later on.

2. Time-harmonic problems in electromagnetics

Let be a domain in R3 , i.e. an open, connected and bounded subset of R3 with a Lipschitz-continuous boundary .
For a given pulsation > 0, the time-harmonic Maxwells equations (with time-dependence exp(t )) write
curl h + e = j in , (1)
curl e h = 0 in , (2)
div e = in , (3)
div h = 0 in . (4)
Above, the real-valued coefficient is the electric permittivity tensor and the real-valued coefficient is the magnetic
permeability tensor, whereas (e, h) is the couple of electromagnetic fields, and the source terms j and are respectively the
current density and the charge density. The latter are related by the charge conservation equation
+ div j = 0 in . (5)
The other two electromagnetic fields are the electric displacement d and the magnetic induction b. They are related to e and
h by the constitutive relations
d = e, b = h in . (6)
In what follows, we focus mainly on the couple of fields (e, h). However the results are easily extended to the couple of
fields (d , b) thanks to the relations (6).
We assume that the coefficients , , together with their inverses 1 , 1 , belong to L sym ( ). Classically, to be able
1

to define the electromagnetic energy, they are such that min () > 0 and min () > 0 a.e. in where min stands for the
smallest eigenvalue, and the couple of electromagnetic fields belongs to L 2 ( ) L 2 ( ). We choose source terms j L 2 ( ),
and H 1 ( ).
We assume that the medium is surrounded by a perfect conductor, so that the boundary condition below holds:
en=0 on . (7)
Hence the couple of electromagnetic fields (e, h) belongs to H0 (curl; ) H (curl; ).

2.1. Variational formulations

The Maxwell problem can be formulated in the electric field e only, namely

Find e H0 (curl; ) such that

2 e + curl(1 curl e) = j in (8)

div e = in .

Note that in (8), the equation div e = is implied by the second-order equation 2 e + curl(1 curl e) = j, together
with the charge conservation Eq. (5), so it can be omitted. Furthermore, the magnetic field can be recovered using Faradays
law (2). Moreover, one can check that the equivalent variational formulation in H0 (curl; ) writes

Find e H0 (curl; ) such that

(9)
(1 curl e|curl v )0, 2 ( e|v )0, = (j |v )0, , v H0 (curl; ).

1 For more exotic configurations of Maxwells equations, in which or exhibit a sign-change of one or several eigenvalues across some interface, we
refer to [24].
 P. Ciarlet Jr. / Computers and Mathematics with Applications 71 (2016) 85104 87

On the other hand, one can also write the time-harmonic Maxwell problem in the magnetic field h only. Note that as e
belongs to H0 (curl; ), its curl belongs to H0 (div ; ): it follows from Faradays law (2) that h n| = 0. In addition, the
field 1 (curl h j ) actually belongs to H0 (curl; ) (cf. Ampres law (1)). So the Maxwell problem formulated in h only is
Find h H (curl; ) such that


2 h + curl(1 (curl h j )) = 0 in


div h = 0 in (10)
h1 n = 0 on



(curl h j ) n = 0 on .
Again, the equation div h = 0 in (10) is implied by the second-order equation 2 h + curl( 1 (curl h j )) = 0.
Likewise the boundary condition h n| = 0 is implied by the second-order equation and the boundary condition
1 (curl h j ) n| = 0. One now checks that the equivalent variational formulation in H (curl; ) writes
Find h H (curl; ) such that

(11)
(1 curl h|curl v )0, 2 (h|v )0, = ( 1 j |curl v )0, , v H (curl ; ).

2.2. Divergence conditions

We already remarked that the divergence conditions on the fields are consequences of the second-order equations. Also,
one notices that the magnetic field h is automatically div -free. A similar property can be exhibited for the electric field as
follows. Indeed, introduce the scalar field H01 ( ) such that div = in H 1 ( ), and write e = + e0 . Then, e0
belongs to H0 (curl; ), with div e0 = 0. Plugging this splitting of e in (8), one finds that the div -free field e0 is governed
by the equation 2 e0 + curl(1 curl e0 ) = j0 in , with j0 := j L 2 ( ) and div j0 = 0, plus the boundary
condition e0 n| = 0, or equivalently by the variational formulation

Find e0 H0 (curl; ) such that

(12)
(1 curle0 |curl v )0, 2 (e0 |v )0, = (j0 |v )0, , v H0 (curl; ).
Note that both splittings (of e and j) are completely characterized by the scalar field . By construction, one has the
orthogonality relation ( |j0 )0, = 0 so that j 0, h 0, + j0 0, .

2.3. Well-posedness of the time-harmonic Maxwell problems

We refer to [5, Section 5] or to [6] for the solution of the variational formulation (9), with the help of the Fredholm
alternative. Following [6, Section 3.1 and Section 4.2], one can provide a similar construction for the variational formulation
(11). In these references, an infsup condition is obtained, which relies on the definition of an appropriate bijective (one-to-
one and onto) operator that maps H0 (curl; ) into itself. This is possible as soon as 2 is not an eigenvalue of the Maxwell
eigenproblem. We recall that, expressed with the help of the electric field, the eigenproblem writes

Find (e, ) H0 (curl; ) R, e = 0 such that

(1 curl e|curl v ) = (e|v )0, , v H0 (curl; ) (13)

div e = 0 in . 0,

Denoting by ( )Z the sequence of (nonnegative) eigenvalues that goes to + (with finite multiplicity), well-posedness
holds if, and only if, 2 { : Z}.

2.4. Discretization of electromagnetic fields

For the ease of exposition, we assume in this subsection that is a Lipschitz polyhedron. The case of a curved
Lipschitz polyhedron is easily addressed, but it involves more technicalities. To define finite dimensional subspaces (Vh+ )h of
H (curl; ), resp. (Vh )h of H0 (curl; ), we consider a family of simplicial meshes of , and we choose the so-called Ndlecs
first family of edge finite elements [7,1]. Note that, because we are dealing with electromagnetic fields with low regularity,
using the first-order finite elements will be sufficient for our purposes. However, all the analyses are valid for higher-order
element methods, with marginal modifications. We consider that is triangulated by a shape regular family of meshes (Th )h ,
made up of (closed) simplices, generically denoted by K . A mesh isindexed by h := maxK hK (the meshsize), wherehK is the
diameter of K . We use the notation Lp (K ), respectively H s (K ), and K dx in lieu of Lp (int (K )), resp. H s (int (K )), and int (K ) dx.
Let us introduce Ndlecs H (curl; )-conforming (first family, first-order) finite element spaces
Vh+ := {vh H (curl; ) : vh |K R1 (K ), K Th }, Vh := Vh+ H0 (curl; ),
where R1 (K ) is the six-dimensional vector space of polynomials on K defined by
R1 (K ) := {v P1 (K ) : v (x) = a + b x, a, b R3 }.
88 P. Ciarlet Jr. / Computers and Mathematics with Applications 71 (2016) 85104

It is shown in [7, Theorem 1] that any element v in R1 (K ) is uniquely determined by the degrees of freedom in the moment
set ME (v ):

ME (v ) := v t dl . (14)
e eAK

Above, AK is the set of edges of K , and t is a unit vector along the edge e.
One can then define the global set of moments on Vh+ , resp. on Vh , by taking one degree of freedom as above per edge
of Th , resp. per interior edge of Th . We recall that the basic approximability properties for the edge finite element write
(cf. [1, Lemma 7.10])

lim inf v vh H (curl; ) = 0, v H (curl; ),
h0 +
vh Vh
(15)
lim inf v vh H (curl; ) = 0, v H0 (curl; ).
h0 vh Vh

Assuming for simplicity that the integrals are computed exactly, the discrete electric problem writes

Find eh Vh such that
(16)
(1 curl eh |curl vh )0, 2 ( eh |vh )0, = (j |vh )0, , vh Vh .
According2 to [5, Section 5], [1, Section 7], or to [6], one has
h0 > 0, h < h0 , e eh H (curl; ) . inf e vh H (curl; ) . (17)
vh Vh

In particular, it follows from (15) that

lim e eh H (curl; ) = 0. (18)
h0

On the other hand, the discrete magnetic problem writes

Find hh Vh+ such that

(19)
( 1 curl hh |curl vh )0, 2 (hh |vh )0, = ( 1 j |curl vh )0, , vh Vh+ .
Again, one has
h0 > 0, h < h0 , h hh H (curl; ) . inf h vh H (curl; ) (20)
+
vh Vh

and as a consequence
lim h hh H (curl; ) = 0. (21)
h0

The aim of the rest of the paper is to refine the convergence estimates (18) and (21), under minimal regularity assumptions
on the data.

3. Building splittings of electromagnetic fields

In this section, we present some abstract tools, which then yield precise regularity results for the couple of electromag-
netic fields.
Let be a real-valued tensor field. We introduce
XN ( , ) := {v H0 (curl; ) : v H (div ; )},
XT ( , ) := {v H (curl; ) : v H0 (div ; )},
ZB ( , ) := {v XB ( , ) : curl v = 0, div v = 0 in }, B {N , T }.
The function spaces XN ( , ) and XT ( , ) are endowed with the graph norm v (v 2H (curl ; ) + v 2H (div ; ) )1/2 . Briefly,
we recall that the Maxwell electric, resp. the Maxwell magnetic, problems are well-posed within the Fredholm alternative
framework (cf. Section 2.3) thanks to the Weber compact embedding results [9] stated in the next theorem.

Theorem 1. Let be a real-valued tensor field such that , 1 L

sym ( ), min ( ) > 0 a.e. in . Then both XN ( , ) and
XT ( , ) are compactly embedded into L 2 ( ).
If we write e = + e0 where H01 ( ) is characterized by = div , then e0 XN ( , ); obviously,
1 curl e = 1 curl e0 XT ( , ). On the other hand, h XT ( , ); in addition, if we use now j0 = j ,
then 1 (curl h j0 ) XN ( , ).

2 We do not discuss the issue of the threshold value h , cf. for instance [8].
0
 P. Ciarlet Jr. / Computers and Mathematics with Applications 71 (2016) 85104 89

3.1. Abstract geometrical setting

The domain can be topologically non-trivial, or with a non-connected boundary. We recall some basic results
concerning these categories.
First, the notion of trivial topology: given a vector field v defined on such that curl v = 0 in , does there exist
a continuous, single-valued function p such that v = p? This question is addressed with the help of (co)homology
theory [10]:

either (Top)I =0 given any curl-free vector field v C 1 ( ), there exists p C 0 ( ) such that v = p in ;
or (Top)I >0 there exist I non-intersecting manifolds, 1 , . . . , I , with boundaries i , such that, if we let
=
\ Ii=1 i , given any curl-free vector field v, there exists p C 0 ( ) such that v = p in
.

When I = 0, = . For short, we write (Top)I to cover both cases.

Regarding the practical definition of the manifolds, or cuts, (i )i=1,...,I , finding them to enforce (Top)I >0 is inexpensive
in terms of algorithmic complexity, see [10, Chapter 6]. In particular, one can build cuts that are piecewise plane. We keep
this assumption from now on. Finally, we assume, that is a connected set.
The domain is said to be topologically trivial when I = 0. When I > 0, the set has pseudo-Lipschitz boundary in
the sense of [11], the continuation operator from L2 ( ) to L2 ( ) is denoted by, whereas the jump across i is denoted by
[]i , for i = 1, . . . , I. The definition of the jump depends on the (fixed) orientation of the normal vector field to i . For all
i, we let , i denote the duality pairing between H 1/2 (i ) and (H 1/2 (i )) .
One has the integration by parts formula [11, Lemma 3.10]:

(v | q )0, + (div v |q)0, = v n, [q]i i , v H0 (div ; ), q H 1 (
).
1i I

In this configuration, we also introduce the subspace P (

) of H 1 (
):

P (
) := {q H 1 (
) : [q]i = csti , 1 i I }.

Above, for i = i , csti and csti may be different. If q P ( q H 1 ( ).

) with vanishing jumps ([q]i )i=1,...,I , then
Second, when the boundary is not connected, let (k )k=0,...,K be its (maximal) connected components. Otherwise,
0 = . For all k, we let , k denote the duality pairing between H 1/2 (k ) and H 1/2 (k ) = (H 1/2 (k )) . We introduce
the subspace H1 ( ) of H 1 ( ):

H1 ( ) := {q H 1 ( ) : q|0 = 0, q|k = cstk , 1 k K }.

Above, for k = k , cstk and cstk may be different.

For domains that fit into the above categories, one can build scalar potentials for curl-free elements, and also vector
potentials for divergence-free elements, provided some compatibility conditions are fulfilled. Those results are recalled in
the next subsection (see [12,11] for details) since they are a crucial ingredient to derive the splittings of electromagnetic
fields.

3.2. Scalar and vector potentials

Let us begin by the extraction of scalar potentials.

Theorem 2. Let be a domain such that (Top)I is fulfilled. Then, given v L 2 ( ), there holds

curl v = 0 in p P (
), p in .
v=

The scalar potential p is unique up to a constant, and |p|1, = v 0, .

Remark 1. When I > 0, if p H 1 (

) \ P (
), then curl
p = 0 in .

Theorem 3. Let be a domain. Then, given v L 2 ( ), there holds

curl v = 0 in ,

p H1 ( ), v = p in .
v n| = 0

The scalar potential p is unique, and |p|1, = v 0, .

Let us continue by the extraction of vector potentials.

90 P. Ciarlet Jr. / Computers and Mathematics with Applications 71 (2016) 85104

Theorem 4. Let be a domain. Then, given v L 2 ( ), there holds

div v = 0 in , v ( ),
1
w Hzm
v = curl w in .
v n, 1k = 0, k div w = 0 in ,

Furthermore, one may choose the vector potential w such that w 1, . v 0, .

Theorem 5. Let be a domain such that (Top)I is fulfilled. Then, given v L 2 ( ), there holds

div v = 0 in , w H0 (curl; ),

v n| = 0, div w = 0 in , v = curl w in .
v n, 1i = 0, i w n, 1k = 0, k
The vector potential w is unique, and w H (curl; ) . v 0, .

Remark 2. As indicated in the statement of Theorems 25, the assumptions on the field v are necessary and sufficient to
guarantee the existence of the potential. Regarding the bounds, the constants hidden in . depend only on the geometry of
the domain .

As a consequence of these results, one may derive auxiliary results on the measure of the couple of electromagnetic fields.

Corollary 1. Let be a domain. Let be a real-valued tensor field such that , 1 L

sym ( ), min ( ) > 0 a.e. in . There
holds:

v 0, . curl v 0, + div v 1, + PN v ZN ( , ) , v H0 (curl; );

v 0, . curl v 0, + div v 0, + PN v ZN ( , ) , v XN ( , ).
Above, PN is an idempotent operator acting from H0 (curl; ) onto ZN ( , ).

Corollary 2. Let be a domain such that (Top)I is fulfilled. Let be a real-valued tensor field such that , 1 L
sym ( ),
min ( ) > 0 a.e. in . There holds:
v 0, . curl v 0, + div v 0, + PT v ZT ( , ) , v XT ( , ).
Above, PT is an idempotent operator acting from H (curl; ) onto ZT ( , ).

3.3. Assumption on the coefficients

Let us proceed with the proper assumptions on the coefficients and that are needed later on.

J
Definition 1. P := {j }j=1,...,J is a partition of if (j )j=1,...,J are disjoint domains, and = j=1 j .

Given a partition P , define the interfaces Fjj := j j and Fint := {Fjj , 1 j = j J }; Fj = j and
Fbdry := {Fj , 1 j J }; F := Fint Fbdry . By convention, if the Hausdorff dimension of Fjj (resp. Fj ) is lower than 2, then
Fjj = (resp. Fj = ). For a field v defined on , we denote by vj its restriction to j , for all j. Define further:

PH t ( ) := {v L2 ( ) : vj H t (j ), 1 j J }, t > 0;
1, 1,
PW ( ) := { L ( ) : j W (j ), 1 j J };


(g |g )0,F :=

(g |g )0,F , g , g L2 (F ), F {Fint , Fbdry , F };
F F
1/2
PH (Fint ) := {g L2 (Fint ) : g|F H 1/2 (F ), F Fint }.
Above, the reference to P is omitted to simplify the notations.
Classically, in a domain , one has PH t ( ) = H t ( ) for all partitions P and for all t ]0, 1/2[. On the other hand, when
the partition is trivial, that is P = { }, one has PH t ( ) = H t ( ) for all t > 0, etc.

Definition 2. Let be a real-valued tensor field such that , 1 L

sym ( ), min ( ) > 0 a.e. in . fulfills the coefficient
assumption if there exists a partition P of such that PW1, ( ).

Remark 3. If fulfills the coefficient assumption on a partition, then 1 fulfills the coefficient assumption on the same
partition.
 P. Ciarlet Jr. / Computers and Mathematics with Applications 71 (2016) 85104 91

Let fulfill the coefficient assumption, and define

HB ( , ) := XB ( , ) PH 1 ( ), B {N , T }.
If is smooth on , then one may choose P = { }. In the particular case where is equal to the identity, one writes XN ( )
instead of XN ( , 1), etc., and one has obviously
HB ( ) H 1 ( ), B {N , T },
where refers to an algebraical and topological embedding.

3.4. Case of constant coefficients

To begin with, we recall the BirmanSolomyak splitting of elements of XN ( ), see [13, Theorem 4.1]. This fundamental
result complements those on the extraction of scalar and vector potentials.

Theorem 6. Let be a domain. Then, there exists a continuous splitting operator acting from XN ( ) to HN ( ) H01 ( ).
More precisely, given v XN ( ),

(vreg , q) HN ( ) H01 ( ), v = vreg + q in , (22)

and one has
vreg 1, + q1, + 1q0, . v XN ( ) . (23)

On the other hand, having at hand additional results for the function spaces ZB ( ), B {N , T } is very useful to tackle
the general case. Indeed since we consider fairly general geometrical settings (topologically non-trivial domains or non-
connected boundary), we must take into account the null spaces. Let us recall the characterizations of ZN ( ) and ZT ( )
provided by [11].
First, we remark that, for all 1 k K , there exists one, and only one qk H1 ( ) such that 1qk = 0 in and
qk |k = k k for 1 k K . Defining vk := qk L 2 ( ), one checks that
curl vk = 0, div vk = 0 in , vk n| = 0, 1 k K .
Let QN ( ) := Span1k K (qk ) be the vector space of potentials, of dimension K . Notice that an element q of QN ( ) is
obviously characterized by its boundary values (q|k )k =1,...,K . One has the characterization below for the null space ZN ( ).

Proposition 1. Let be a domain. One has ZN ( ) = Span1k K ( qk ). In addition, v ZN ( ) can be characterized by the
fluxes (v n, 1k )k=1,...,K .
As far as the regularity of elements of the null space ZN ( ) is concerned, the previous characterization allows one to derive
it easily.

Corollary 3. Let be a domain. There holds ZN ( ) H 1/2 ( ).

Proof. Let v ZN ( ): according to Proposition 1, there exists pz QN ( ) such that v = pz . By construction, 1pz = 0 in
, and moreover pz | H 1 ( ). Thanks to [14], one has pz H 3/2 ( ), hence pz H 1/2 ( ). This proves the claim.
In particular, 1/2, and all p -norms measuring the fluxes (v n, 1k )k=1,...,K in CK are equivalent norms over the
finite-dimensional vector space ZN ( ).
Regarding the null space ZT ( ), one considers, for all 1 i I, the scalar field p i defined on
as the solution to:
Find p i Pzmv ( ) such that

( p i | q )0, = [q]i , q Pzmv (

).

Then, vi =
p i L 2 ( ) is such that
curl vi = 0, div vi = 0 in , vi n| = 0, and vi n, 1i = ii , 1 i I .

Let us define a second vector space QT (

) := Span1i I (pi ) of scalar potentials, of dimension I. Its elements p may be
characterized by their jumps ([p]i )1i I . Next, if p QT (
) fulfills [p]i = 0 for all i, then p =
p belongs to H 1 ( ). On the
other hand, using q = p in the variational formulation that defines p yields p = 0 in
: it follows that p = 0 and p = 0.
Hence, (continuations to of) non-zero elements of QT ( ) do not belong to H 1 ( ).

Proposition 2. Let be a domain such that (Top)I is fulfilled. One has ZT ( ) = Span1i I (
p i ). In addition, an element v of
ZT ( ) can be characterized by the fluxes (v n, 1i )i=1,...,I .
92 P. Ciarlet Jr. / Computers and Mathematics with Applications 71 (2016) 85104

Regarding the regularity of elements of the null space ZT ( ), one may prove the result below.

Corollary 4. Let be a domain such that (Top)I is fulfilled. There holds ZT ( ) H 1/2 ( ).

Proof. Let v ZT ( ), v = 0: according to Proposition 2, there exists p Pzmv (

) such that v =
p . As noted above,
p H 1 ( ).

However, one may address this difficulty by using a partition of unity. Let (i )i=1,...,I be such that for all i: i
C ( , [0, 1]) with connected support, i = 1 in a neighborhood of i , and supp(i ) i = for i = i. One may
further define connected, open subsets (Oi )i=1,...,I of such that supp(i ) Oi and Oi i = , for i = i . Each subset
is split into two parts, Oi and Oi+ , according to the orientation of the normal vector to i , so that [z ]i = z| O + z| O . By
i i
defining 0 = 1 i , one gets a partition of unity ( )=0,...,I on .

1iI
Next, let p = p for all : by construction, p
0 H 1 ( ), whereas p i P (
) for 1 i I. Introduce, for 1 i I,
pi L2 (Oi ) defined as pi = p i in Oi and pi = p i [pi ]i in Oi+ . As [pi ]i = 0, it holds pi H 1 (Oi ), and in addition
1pi L2 (Oi ) and n pi | Oi L2 ( Oi ). So we obtain that pi H 3/2 (Oi ), cf. [14,15], which implies p i = pi H 1/2 (Oi ). It
follows that p i belongs to H 1/2 ( ) because p i vanishes in a neighborhood of Oi (and p i = 0 in \ Oi ). Likewise,

0 belongs to H 1/2 ( ). Using the definition of the partition of unity, one concludes that v =
p 0 = p p H 1/2 ( ).

If we let

P 3/2 (
) := {q P ( q H 1/2 ( )},
):
3/2
we have proven in passing that QT ( ) Pzmv ( ). Moreover, 1/2, and all p -norms measuring the fluxes (v
n, 1i )i=1,...,I in CI are equivalent norms over ZT ( ).
In the next subsection, we proceed with the splittings of elements of XN ( , ), resp. of elements of XT ( , ).

3.5. Splittings of fields

We provide now splittings into a regular part, and a gradient part, of elements of XN ( , ) (electric case), resp. of
elements of XT ( , ) (magnetic case), called regular/gradient splittings. Since we are dealing with general geometrical
settings, an additional part is present, which belongs to the null space with the same boundary condition.
Let us begin with the electric case.

Theorem 7. Let be a domain such that (Top)I is fulfilled, and assume that fulfills the coefficient assumption. Then, there
exists a continuous splitting operator acting from XN ( , ) to HN ( ) ZN ( ) H01 ( ).
More precisely, given v XN ( , ),

(vreg , z , p0 ) HN ( ) ZN ( ) H01 ( ), v = vreg + z + p0 in ; (24)

the scalar field p0 is governed by the variational formulation below, for some f L2 ( ) and gF PH 1/2 (Fint ):

Find p0 H01 ( ) such that

(25)
( p0 |)0, = ( z |)0, + (f |)0, + (gF |)0,Fint , H01 ( );
one has

vreg 1, + vreg XN ( ) + z 1/2, . v H (curl; ) ,

(26)
z PH 1/2 ( ) + f 0, + gF 1/2,Fint . v XN ( , ) .
In addition, one may choose pN : v z, acting from XN ( , ) to ZN ( ) to be an idempotent operator.

Remark 4. In the splitting (24) of v XN ( , ), all three terms vreg , z , p0 have vanishing tangential components on the
boundary . Regarding regularity in (24), one has vreg H 1 ( ), resp. z = pz H 1/2 ( ) with pz H 1 ( ), resp.
p0 L 2 ( ). Since fulfills the coefficient assumption, the variational formulation (25) is well-posed. In the first bound
in (26), the constant hidden in . depends only on the geometry, whereas, in the last bound in (26), the constant hidden in
. also depends on PW1, ( ) . The idea of the proof follows closely [16, Section 3].

Proof. Let y = curl v H0 (div ; ). By construction div y = 0 in , and one checks that y n, 1i = 0 for all i.
According to Theorem 5 on vector potentials, there exists w XN ( ) with div w = 0 in , w n, 1k = 0 for all k,
such that y = curl w in and w XN ( ) . y 0, . Next, we know that there exists a BirmanSolomyak splitting of w, see
Theorem 6:

vreg HN ( ), q H01 ( ), w = vreg + q in ,

 P. Ciarlet Jr. / Computers and Mathematics with Applications 71 (2016) 85104 93

with continuous dependence (23). By construction, curl(v vreg ) = 0 in , with (v vreg ) H0 (curl; ). According to
Theorem 3 on scalar potentials, there exists p H1 ( ) such that v = vreg + p in . Using the definition of the space of
scalar potentials QN ( ), one may further split p as p = p0 + pz in , with p0 H01 ( ), and pz QN ( ). Introducing finally
z = pz ZN ( ), we have proved that there holds

v = vreg + z + p0 in ,

with vreg HN ( ), z ZN ( ), p0 H01 ( ), which is precisely (24).

Let us proceed with the definition of p0 as the solution to (25). Let H01 ( ), then

( p0 |)0, = ( z |)0, + ( v |)0, ( vreg |)0, .

Below, we study the last two terms separately.
Consider first v XN ( , ). One has in particular v H (div , ), so by integration by parts on one gets

( v |)0, = (div v |)0, .

Consider next vreg HN ( ). If is only piecewise smooth3 on , vreg n has jumps across faces of Fint . On the other hand,
one has j vreg ,j H 1 (j ) for all j. Therefore, one can integrate by parts over each subdomain to find

( vreg |)0, = (j vreg ,j |j )0,j
j

= (div j vreg ,j |j )0,j ([ vreg n]|)0,F
j F Fint

= (div
vreg |)0, ([ vreg n]|)0,F .
F Fint

If we introduce

f = div v + div
vreg L2 ( ), gF = [ vreg n] PH 1/2 (Fint ),

we obtain that p0 is governed by (25). We derive next the (uniform) estimates (26) to prove that the splitting operator is
continuous. By construction

vreg 1, . w XN ( ) . y 0, v H (curl; ) ;
vreg XN ( ) w XN ( ) + qXN ( ) . w XN ( ) v H (curl; ) .
For instance, z ZN ( ) can be measured by the 1 -norm of (z n, 1k )k=1,...,K :

|z n, 1k | = |z n, qk H 1/2 ( ) | = |(z | qk )0, |

= |(z + p0 | qk )0, | = |(v vreg | qk )0, |
(v 0, + vreg 0, ) qk 0, . v H (curl; ) .
Above, we used first the definition of (qk )k given in Section 3.4, and then the fact that p0 and qk are orthogonal with
respect to (|)0, (integrate by parts).
For a given j,

z 1/2,j . v H (curl; ) ,
f 0,j div v 0,j + div vreg 0,j . v XN ( , ) .
And for a given F = j j Fint , we find, thanks to the continuity of the trace mapping,

gF 1/2,F = [ vreg n]1/2,F [ vreg ]1/2,F . vreg 1, . v XN ( , ) .
=j,j

In the last three bounds, the constant hidden in . depends on PW1, ( ) .

Finally, to prove that the operator pN is idempotent, let us build the splitting (24) for v ZN ( ): as y = curl v = 0, it
follows that w = 0, so that vreg = 0 and q = 0 in its BirmanSolomyak splitting. Next, let p H1 ( ) be such that v = p:
according to Proposition 1, p actually belongs to QN ( ). By uniqueness, one obtains that p0 = 0, hence the splitting (24)
writes v = 0 + v + 0 in , so that pN v = v. To conclude that pN is an idempotent operator, one simply remarks that for all
v XN ( ), pN v belongs to ZN ( ) by definition, so it holds pN (pN v ) = pN v, or p2N = pN in operator form.

3 If is globally smooth on , P = { } and F = .

int
94 P. Ciarlet Jr. / Computers and Mathematics with Applications 71 (2016) 85104

To carry on, one needs regularity results regarding p0 , where p0 is governed by the variational formulation (25). For that,
we use an abstract shift theorem, proven in [17, Theorem 3.1], that deals with second order elliptic PDEs complemented with
Dirichlet boundary conditions. This result provides a lower bound on the a priori regularity of p0 in all the configurations
that we consider in this paper.4

Theorem 8. Let be a domain, and assume that fulfills the coefficient assumption. There exists Dir := Dir ( ) ]0, 1/2[
depending only on the geometry and the coefficient such that: for all s [0, Dir [, for all H s1 ( ), the solution to

Find u H01 ( ) such that

( u|)0, = , H 1 ( ) , H01 ( ),
0

belongs to H s+1 ( ), and moreover 5 us+1, .s s1, .

Combining the two theorems yields the result regarding the regular/gradient splitting of elements of XN ( , ). Below,
[H s+1 ( ) H1 ( )] denotes the range of the gradient operator from H s+1 ( ) H1 ( ) to H s ( ).

Corollary 5. Let be a domain such that (Top)I is fulfilled, and assume that fulfills the coefficient assumption. There holds

XN ( , ) HN ( ) + [H s+1 ( ) H1 ( )], s [0, Dir [. (27)

Proof. Let s [0, Dir [. Given v XN ( , ), we use the splitting (24), namely

vreg HN ( ), z ZN ( ), p0 H01 ( ), v = vreg + z + p0 in ,

where p0 is governed by (25), with the uniform bounds (26). Hence, vreg 1, . v XN ( , ) . Furthermore, thanks to
Corollary 3, one can write z = pz , with pz H 3/2 ( ) H1 ( ) H s+1 ( ) H1 ( ), so it holds pz 1+s, . v XN ( , ) .
Then, p0 is characterized by (25), with a right-hand side

: ( z |)0, + (f |)0, + (gF |)0,Fint

that belongs to (H01s ( )) = H s1 ( ). Indeed, if H01s ( ), then:

H s ( ) = (H s ( )) (recall that s [0, 1/2[), and z PH 1/2 ( ) H s ( ), so one may write the first term as
, z H s ( ) ;
for all F Fint , |F L2 (F ).
Hence, according to the shift Theorem 8, it follows that p0 H s+1 ( ), with continuous dependence. So we get:

p0 1+s, . s1, . z PH 1/2 ( ) + f 0, + gF 0,Fint . v XN ( , ) .

This proves the claim.

Let us continue with the magnetic case.

Theorem 9. Let be a domain such that (Top)I is fulfilled, and assume that fulfills the coefficient assumption. Then, there
exists a continuous splitting operator acting from XT ( , ) to Hzm
1
v ( ) ZT ( ) Hzmv ( ).
1

More precisely, given v XT ( , ),

(wreg , z , q0 ) Hzm
1
v ( ) ZT ( ) Hzmv ( ),
1
v = wreg + z + q0 in ; (28)
1/2
the scalar field q0 is governed by the variational formulation below, for some f L ( ) and gF PH
2
(F ):

v ( ) such that
1

Find q0 Hzm
(29)
( q0 |)0, = ( z |)0, + (f |)0, + (gF |)0,F , 1
Hzmv ( );

one has

wreg 1, + z 1/2, . v H (curl; ) ,

(30)
z PH 1/2 ( ) + f 0, + gF 1/2,F . v XT ( , ) .

4 In some configurations, it can happen that the limit exponent is larger than 1/2. However, we are interested here only in the existence of such an
Dir
exponent. More precise results may be derived for fairly general subclasses of the configurations, we refer to Section 5.2.
5 The symbol . means that the value of the given constant appearing in the inequality depends on s: s, C , H s1 ( ), u C .
s s s+1, s s1,
 P. Ciarlet Jr. / Computers and Mathematics with Applications 71 (2016) 85104 95

Remark 5. In the splitting (28) of v XT ( , ), wreg does not fulfill any boundary condition in general. Regarding regularity,
one has wreg H 1 ( ), resp. z = p z H 1/2 ( ) with p z Pzmv ( ), resp. q0 L 2 ( ). Since fulfills the coefficient
assumption, the variational formulation (29) is well-posed. In the first bound in (30), the constant hidden in . depends only
on the geometry, whereas in the last bound the constant hidden in . also depends on PW1, ( ) . Again, the idea of the
proof follows closely [16, Section 3].

Proof. Let y = curl v H (div ; ). One has div y = 0 in , and y n, 1k = 0 for all k. Thanks to Theorem 4 on vector
v ( ) with div wreg = 0 in such that y = curl wreg in and
1
potentials, there exists wreg Hzm

wreg 1, . y 0, v H (curl; ) .
By construction, curl(v wreg ) = 0 in , with (v wreg ) H (curl; ). According to Theorem 2 on scalar potentials, there
exists q Pzmv (
) such that v = wreg +
q in . And |q|1, v 0, + wreg 0, . v H (curl; ) .
Since elements of QT (
) can be characterized by their jumps, we introduce next p QT (
) such that [p]i = [q]i for
all i, and then z = p ZT ( ). The norm pQ ( ) is bounded by ([q] )i (CI ) , which is itself bounded by |q|1, , so one
T i 1
gets z 1/2, . v H (curl; ) .
p , one has q0 H 1 ( ), and in addition there holds
If one lets q0 = q zmv

v = wreg + z + q0 in ,

v ( ), z ZT ( ), q0 Hzmv ( ), i.e. (28).

1 1
with wreg Hzm
About the definition of q0 as the solution to (29), let Hzm
1
v ( ):

( q0 |)0, = ( z |)0, + ( v |)0, ( wreg |)0, .

As v H0 (div ; ) one finds by integration by parts ( v |)0, = (div v |)0, .
For the third term, one proceeds as in the proof of Theorem 7, the only difference being that there are additional boundary
terms:

( wreg |)0, = (div
wreg |)0, ([ wreg n]|)0,F ( wreg n|)0, .
F Fint

Define next,

f = div v + div
wreg L2 ( ), gF = [ wreg n] PH 1/2 (F ),

where, for all F Fbdry and z L2 (F ), the jump [z ] is simply equal to z. It follows that q0 is characterized by (29).
Finally, the first bound in (30) has already been derived, and the second one is obtained exactly as in the proof of
Theorem 7, hence continuity of the splitting operator is obtained.

In the magnetic case, one needs regularity results regarding q0 , where q0 is now governed by (29). We use the abstract
shift theorem [17, Theorem 3.1] for PDEs with Neumann boundary conditions (see footnote 4, page 15, for some comments
on the optimality of the limit exponent, here Neu ).

Theorem 10. Let be a domain, and assume that fulfills the coefficient assumption. There exists Neu := Neu ( ) ]0, 1/2[
depending only on the geometry and the coefficient such that: for all s [0, Neu [, for all (Hzm
1s
v ( )) , the solution to

v ( ) such that
1

Find u Hzm
( u|)0, = , H 1 ( ) ,
zmv
1
Hzmv ( ),

belongs to H s+1 ( ), and moreover us+1, .s (H 1s ( )) .

zmv

More precise results may be derived for subclasses of the configurations, cf. Section 5.2. Combining the two Theorems 9 and
10 yields the result for the regular/gradient splitting of elements of XT ( , ). The proof is omitted, as it is very close to the
one of Corollary 5.

Corollary 6. Let be a domain such that (Top)I is fulfilled, and assume that fulfills the coefficient assumption. There holds
3/2
XT ( , ) H 1 ( ) + [
Pzmv ( )] + [Hzmv ( )],
s+1
s [0, Neu [. (31)

For the sake of completeness and because this result will be used later on, we mention that it is also possible to derive a
splitting of XT ( , ) which preserves the homogeneous boundary condition on the normal trace, under some moderate
restrictions on the domain .
96 P. Ciarlet Jr. / Computers and Mathematics with Applications 71 (2016) 85104

Theorem 11. Let be a (curved) Lipschitz polyhedron such that (Top)I is fulfilled, and assume that fulfills the coefficient
assumption. Then, there exists a continuous splitting operator acting from XT ( , ) to HT ( ) ZT ( ) Hzm
1
v ( ). Given
v XT ( , ),

(vreg , z , p0 ) HT ( ) ZT ( ) Hzm
1
v ( ), v = vreg + z + p0 in , (32)

and one may choose the operator pT : v z, acting from XT ( , ) to ZT ( ), to be idempotent.

v ( ) such that curl wreg = curl v in and

1
Proof (Outlined). Let us begin as for Theorem 9 to derive wreg Hzm
wreg 1, . curl v 0, v H (curl; ) . A priori, wreg n| = 0. But, in the (curved) Lipschitz polyhedron one has [18],
qreg
qreg H 2 ( ), | = wreg n| ; qreg 2, . wreg 1, .
n
It follows that vreg = wreg qreg XT ( ) H 1 ( ) = HT ( ), curl vreg = curl v in and vreg 1, . v H (curl; ) .
Because curl(v vreg ) = 0 in , with (v vreg ) H (curl; ), there exists p Pzmv ( ) such that v = vreg + p in
p = z + p0 , with z ZT ( ) and
p as
and |p|1, . v H (curl; ) (cf. Theorem 2 on scalar potentials). One then writes
v ( ).
1
p0 Hzm
One then follows the proof of Theorem 9 to prove that the splitting operator v (vreg , z , p0 ) is continuous from XT ( , )
to HT ( ) ZT ( ) Hzm 1
v ( ).
Finally, one checks step-by-step that the construction of the splitting (32) for v ZT ( ) yields the decomposition
v = 0 + v + 0 in , so that pT v = v. Hence p2T = pT , i.e. the operator pT is idempotent.

Remark 6. In the particular case where is equal to the identity, (32) may be viewed as a second BirmanSolomyak equality.

4. Interpolation and quasi-interpolation

We assume that , fulfill the coefficient assumption on the same partition P := {j }j=1,...,J of and, for the ease of
exposition, we also assume that the domain and the subdomains {j }j=1,...,J are Lipschitz polyhedra. A triangulation Th is
compatible with the partition P if, for all K Th , there exists j {1, . . . , J } such that K j . On the other hand, if the domain
is such that (Top)I >0 is fulfilled, we have at hand some piecewise plane cuts (i )i=1,...,I , cf. Section 3.1. A triangulation Th is
compatible with the cuts (i )i=1,...,I if, for all 1 i I and for all K Th , int (K ) i = .

Definition 3. A triangulation is compatible if it is both compatible with the partition and with the cuts.

From now on, we assume that (Th )h is a shape regular family of compatible meshes. We are interested in the interpolation
or quasi-interpolation of v such that v XN ( , ) and 1 curl v XT ( , ) (electric case) or, vice versa, such that
v XT ( , ) and 1 curl v XN ( , ) (magnetic case).

4.1. Classical Ndlec interpolation

Introduce the local interpolation operator

K : X (K ) R1 (K ),
where X (K ) is some function space defined on K and given v X (K ), K v is by definition the only element of R1 (K ) with
moments equal to ME (v ), cf. (14). Then, one defines the global interpolation operator h+ with values in Vh+ for all elements
v H (curl; ), resp. h with values in Vh for all elements v H0 (curl; ), such that v|K X (K ) for all K Th , by

(h+ v )|K := K v , resp. (h v )|K := K v , K Th .

One uses vh := h v in the electric case (resp. vh := h+ v in the magnetic case), provided that the action of the operator
h (resp. of the operator h+ ) on v is well-defined. This yields local, simplex-by-simplex estimates with respect to hK , and
then global estimates with respect to h.
Several choices of X (K ) have been proposed over the years. We list some of them below. Let H (curl; K ) := {v
H (K ) : curl v H (K )} for > 0.

Case 1 X2 (K ) := H 2 (K ) [7, Theorem 2]:

v K v H (curl;K ) . hK |v |2,K .
Case 2 X1,1 (K ) := H 1 (curl; K ) [19, Lemma 2.3]:
v K v H (curl;K ) . hK (|v |1,K + |curl v |1,K ).
 P. Ciarlet Jr. / Computers and Mathematics with Applications 71 (2016) 85104 97

Case 3 X 1 +, 1 + (K ) := H (curl; K ) for some ]1/2, 1[ [20, Section 5] or [21, Section 3]:
2 2

v K v H (curl;K ) . (hK ) (|v |,K + |curl v |,K ).

The fact that the action of K is well-defined for elements of H (curl; K ) when ]1/2, 1[ stems from the result
[11, Lemma 4.7].

Proposition 3. Let K be a simplex and p > 2. Then the operator K is well-defined on X (K ) := {v L p (K ) : curl v
L p (K ), v n| K L p ( K )}.

For ]1/2, 1[, v H (K ) implies that v| K H 1/2 ( K ). Then, due to classical embedding theorems (cf. [22, Theorem
7.57]), there exists p := p() > 2 such that v L p (K ) and v| K L p ( K ).
Case 4 X 1 +,0+ (K ) := {v H (K ) : curl v H (K )}, for some ]1/2, 1], ]0, 1] [23, Lemma 5.1]:

v K v H (curl;K ) . (hK )min(, ) (|v |,K + |curl v | ,K ).

Indeed, the authors of [23] first note that one may still use Proposition 3 for elements of X 1 +,0+ (K ) thanks to the same
2
embedding theorems (i.e. [22, Theorem 7.57]), so that the action of K is actually well-defined for elements of X 1 +,0+ (K ).
2
Then, they conclude by applying the same theory as the one developed for Case 3 (cf. [20,21]) to find the desired local
estimate.
Global estimates can be derived easily, starting from the local ones of Cases 14. For instance, for ]1/2, 1] and ]0, 1],
given v PH ( ) such that curl v PH ( ), one finds

inf v vh H (curl; ) . hmin(, ) (|v |PH ( ) + |curl v |PH ( ) );

+
vh Vh

inf v vh H (curl; ) . hmin(, ) (|v |PH ( ) + |curl v |PH ( ) ) if v n| = 0.

vh Vh

4.2. Combined interpolation in the general case

We consider now the case where the simplex-by-simplex regularity of the fields is minimal but provable, cf. Section 3,
namely v H (K ), curl v H (K ), for orders 0 < , < 1/2 that can be arbitrarily small. Precisely, we use the regularity

and the splitting results of Corollaries 5 and 6, in the sense that we study the local, simplex-by-simplex interpolation of

v {v H 1 (K ) + [H 1+ (K )] : curl v H (K )}.

(33)

Remark 7. This condition is the one that holds for conforming triangulations, in the sense that the jumps of the scalar
potential, if any, can only occur at the boundary of its simplices.

The dissymmetry appearing in (33) (fields are split, but not their curl) is used as follows. Noting that curl() = 0, we
actually have to study the approximability of the fields in two cases:

Case a v X1,0+ (K ) := {v H 1 (K ) : curl v H (K )}, for > 0;

Case b v [H 1+ (K )], for > 0.

On the one hand, one addresses the Case a for v X1,0+ (K ) using the operator K and the last interpolation estimate of
Section 4.1 (see Case 4).
On the other hand, handling the Case b is very classical: let v = for some H 1+ (K ). Globally we introduce Lagranges
H 1 ( )-conforming (P1 family) finite element space
Vh+ := {h H 1 ( ) : h |K P1 (K ), K Th }, Vh := Vh+ H1 ( ).

By construction, Vh+ Vh+ and Vh Vh .

One can use the (modified) Clment, or the ScottZhang, interpolation operators from H 1 ( ) to Vh+ , resp. from H1 ( )
to Vh (cf. [24,25]). Denoting by K the local operator, we know that
| K |1,K . hK ||1+,SK
where SK is the neighborhood of the simplex K that is defined by SK := int (Ki , Ki K = Ki ). In other words, the local estimate
on K depends on the regularity on the whole neighborhood SK . Let h , h+ denote the associated global operators.
In the electric case, the field defined on writes v = where H 1+ ( ) H1 ( ) (see Corollary 5). Introducing
vh Vh equal to vh = (h ) on , one aggregates the local estimates to obtain v vh H (curl; ) . h |v |, .
98 P. Ciarlet Jr. / Computers and Mathematics with Applications 71 (2016) 85104

In the magnetic case, there is no global regularity result because of the constant, non-zero jumps across the cuts:
v = where H 1+ ( ) Pzmv (
) (see Corollary 6). However, one may still aggregate the local estimates as follows.
If the simplex K does not intersect any cut, then H 1+ (SK ) and there is no difficulty. On the other hand, if there exists6
i {1, . . . , I } such that K i = , then one chooses an H 1 -conforming continuation of |K to SK , called Kcont , replacing
by []i for the simplices of SK that lie opposite to K with respect to the cut. One has Kcont H 1+ (SK ) (see again
Corollary 6), with = cont on SK . Applying K yields a local estimate on K , namely
K

|Kcont K Kcont |1,K . hK |Kcont |1+,SK = hK |Kcont |,SK = hK ||

,S .
K

If one defines now vh Vh+ by vh = (K ) on K , resp. vh = (K Kcont ) on K , this yields v vh H (curl; ) . h |v |, .

As a consequence, one derives
inf v vh H (curl; ) . h |v |, for v [H 1+ ( ) H1 ( )],
vh Vh

inf v vh H (curl; ) . h |v |, for v [ 3/2

Pzm v ( )] + [Hzmv ( )].
1+
+
vh Vh

A combination of the two Cases a-b then leads to the desired result.

Proposition 4. Let be a domain such that (Top)I is fulfilled, and assume that , fulfill the coefficient assumption on the
same partition. Let := min(Dir (), Neu ()) ]0, 1/2[. Let (Th )h be a shape regular family of compatible meshes.
In the electric case, let v XN ( , ) such that 1 curl v XT ( , ). There holds, for all s [0, [,

inf v vh H (curl; ) .s hs {v XN ( ,) + 1 curl v XT ( ,) }. (34)

vh Vh

In the magnetic case, let v XT ( , ) such that 1 curl v XN ( , ). There holds, for all s [0, [,

inf v vh H (curl; ) .s hs {v XT ( ,) + 1 curl v XN ( ,) }. (35)

+
vh Vh

Proof. Let us outline the proof in the electric case. Let v XN ( , ) such that 1 curl v XT ( , ). With the help of
the embedding result of Corollaries 5 and 6, v and 1 curl v may be split continuously into a regular part and a gradient,
or the continuation of a gradient. Because , fulfill the coefficient assumption, it follows that for all simplices K Th , the
local assumption (33) holds, for all ]0, Dir ()[ and for all ]0, Neu ()[. Then, one may apply the previous construction
(Cases a-b) to find the result.
The proof in the magnetic case is similar.
To conclude this subsection on the interpolation of electromagnetic fields with minimal regularity, note that we have built
combined interpolation operators in the process, that rest on both the classical operators h+ , h , and also on (h+ ), (h ).
Under the assumptions of Proposition 4, these combined operators comb +
,h , comb,h are well-defined according respectively
to the maps
comb,h : {v XN ( , ) : 1 curl v XT ( , )} Vh ;
comb ,h : {v XT ( , ) : curl v XN ( , )} Vh .
+ 1 +

4.3. Quasi-interpolation and commuting diagrams

g ,+ g
Let us recall some classical results. Denote by h , resp. h , the standard Lagrange interpolation operator for (sufficiently
d,+
smooth) elements of H 1 ( ), resp. H1 ( ), and h , resp. hd , the RaviartThomas interpolation operator for (sufficiently
d,+
smooth) elements of H (div ; ), resp. H0 (div ; ), cf. [12,24,25]. Defining the finite dimensional subspaces (Vh )h of
H (div ; ), resp. (Vhd )h of H0 (div ; ), based on the so-called first order RaviartThomas finite elements [12,24,25], we
d,+
have that curl[Vh+ ] Vh , resp. curl[Vh ] Vhd . In addition, the classical commuting diagram properties write, provided
the degrees of freedom exist:
g ,+
hd,+ (curl v ) = curl(h+ v ), and h+ ( q) = (h q);

g (36)
hd (curl v ) = curl(h v ), and h ( q) = (h q).
On the other hand, when one has only v H (curl; ), the moments (14) may not exist on some simplices. To address
this difficulty, the idea developed in [26,27] is to apply, on the simplex K , a (local) smoothing operator sK to v, and then

6 The cuts do not intersect one another so, for sufficiently small h, a simplex intersects at most with one cut.
 P. Ciarlet Jr. / Computers and Mathematics with Applications 71 (2016) 85104 99

the interpolation operator K itself: this results in the so-called quasi-interpolation operator K sK . With this approach,
it is possible to obtain estimates as soon as v H ( ) for some > 0. Briefly, if one introduces SK a local neighborhood
of K (different from SK , regardless still local in the sense of connectivity and neighboring simplices in Th ), then it holds
[26, Theorem 5]:

v (K sK )v 0,K . (hK ) |v |,SK .

The minor drawback of this approach is that, due to its generality, the commuting diagram properties now require the
introduction of other local smoothing operators. For short, we consider only the case with boundary conditions next.
In the general situation considered in [26,27], one introduces a smoothing operator s for elements of H0 (curl; ) (see
above), together with smoothing operators sg for elements of H1 ( ), resp. sd for elements of H0 (div ; ). Given v
H0 (curl; ) H ( ) and q H1 ( ) H ( ) for some > 0, it holds now [27, Section 3]:

(hd sd )(curl v ) = curl((h s)v ) and (h s)( q) = ((hg sg )q).

Whereas, for the combined interpolation operator comb,h defined on {v XN ( , ) : 1 curl v XT ( , )}, by using
[28, Section 3], one checks easily that

hd (curl v ) = curl(comb,h v ).

Remark 8. We refer to [29] for the abstract theory on quasi-interpolation operators within the framework of exterior
calculus.

5. Error estimates

We apply to the electromagnetic fields the results of Sections 4.14.2 with the classical and combined interpolation
operators.

5.1. Minimal regularity assumptions

To obtain error estimates that are better than (18) or (21), one must impose extra regularity on the source terms
j L 2 ( ), H 1 ( ) even in the low-regularity case. Let = min(Dir (), Neu ()) ]0, 1/2[ as in Proposition 4.
Recall that we introduced in Section 2.2 the scalar charge potential H01 ( ) such that div = in H 1 ( ), and
that we split the current density as j := j0 + j , with j0 L 2 ( ), div j0 = 0 and j = L 2 ( ). Below, we may
impose that j H ( ) or that j0 H ( ).7

Theorem 12. Let be a domain such that (Top)I is fulfilled, and assume that , fulfill the coefficient assumption on the same
partition. Let := min(Dir (), Neu ()) ]0, 1/2[. Let (Th )h be a shape regular family of compatible meshes.
In the electric case, assume that j H ( ). There holds, for all s [0, [,

inf e vh H (curl; ) .s hs {j 0, + |j |s, }. (37)

vh Vh

In the magnetic case, assume that j0 H ( ). There holds, for all s [0, [,

inf h vh H (curl; ) .s hs {j 0, + |j0 |s, }. (38)

+
vh Vh

Remark 9. Note that because < 1/2, we only give a lower optimal bound on the rate of convergence. In the case where
the source terms are more regular then, depending on the smoothness of the coefficients and , the convergence rate may
be as fast as . h for order one finite elements, cf. Case 2 of Section 4.1. We highlight these situations with some examples
in Section 5.2.

Proof. Regarding the electric case, we saw in Section 2.2 that one may split the electric field as e = + e0 with
e0 H0 (curl; ), div e0 = 0 and the same scalar potential as in the splitting of j. Because of the extra regularity on j ,

7 Other conditions can be considered when fulfills the coefficient assumption. On the one hand, one may impose the condition H 1 ( ), which
implies that H 1+t ( ), or that 1 j H t ( ), for all t ]0, [ (Theorem 8), hence j H t ( ), for all t ]0, [. So, with a slight abuse, one can consider
that H 1 ( ) implies the condition j H ( ). On the other hand, one may impose the global condition j H ( ). In this case, div j H 1 ( ),
hence H 1 ( ) thanks to the charge conservation Eq. (5). Then one has j H t ( ), for all t ]0, [, and it follows that j0 = j j H t ( ), also for
t ]0, [.
100 P. Ciarlet Jr. / Computers and Mathematics with Applications 71 (2016) 85104

one has H 1+ ( ) and so using the (modified) Clment, or the ScottZhang, interpolation operators (cf. Section 4.2)
one obtains
inf vh H (curl; ) . hs | |s, .
vh Vh

On the other hand, one has e0 XN ( , ) and 1 curl e0 XT ( , ). So we derive from Proposition 4 (electric case)
that, for all s [0, [,
inf e0 vh H (curl; ) .s hs {e0 XN ( ,) + 1 curl e0 XT ( ,) }.
vh Vh

Adding up the two estimates and noting finally that, due to the well-posedness of the variational formulation (12) in e0
(Section 2), one has
e0 XN ( ,) + 1 curl e0 XT ( ,) . j 0, ,
we find (37).
Regarding the magnetic case, one has h XT ( , ) and 1 (curl h j0 ) XN ( , ). Thanks to the extra regularity
on j0 , one may follow the proof of Proposition 4 (magnetic case). First, as an element of XT ( , ), the magnetic field
is decomposed as usual (31). Second, its curl exhibits the same regularity as before. Indeed, introduce the auxiliary field
x := 1 (curl h j0 ). Then given s [0, [, x XN ( , ) H s ( ), so that curl h j0 = x H s ( ) and curl h H s ( )
because j0 H s ( ), with the bound
curl hs, .s xXN ( ,) + j0 s, .
It follows that
inf h vh H (curl; ) .s hs {hXT ( ,) + xXN ( ,) + j0 s, }.
+
vh Vh

Due to the well-posedness of the variational formulation (11) in h (Section 2), one gets
hXT ( ,) + xXN ( ,) . j 0, .
Finally, we recover (38) with the help of j0 0, . j 0, .

Corollary 7. Let the assumptions of Theorem 12 on the geometry and the coefficients hold.
Assume that j H ( ) and let be := curl e H0 (div ; ). Then be is a divergence-free field and there holds, for all
s [0, [,

inf be wh H (div ; ) .s hs {j 0, + |j |s, }. (39)

wh Vhd

Assume that j0 H ( ) and let d0,h := curl h j0 H (div ; ). Then d0,h is a divergence-free field and there holds, for all
s [0, [,

inf d0,h wh H (div ; ) .s hs {j 0, + |j0 |s, }. (40)

d,+
wh Vh

Proof. Thanks to (36), one notices that

inf be wh H (div ; ) inf curl(e vh )H (curl; ) ,
wh Vhd vh Vh

hence (39) is a straightforward consequence of (37).

The second result is derived in a similar fashion if one recalls that, since j0 is a divergence-free field that belongs to H ( ),
one has j0 hd j0 H (div ; ) . hs |j0 |s, (cf. [28, Lemma 3.3]).

5.2. More on regularity in a polyhedral domain

We assume here that the domain and the subdomains of its partition are Lipschitz polyhedra (for short, a subdomain
is called subpolyhedron). In this setting, we would like to guarantee when the regularity of the fields is a priori sufficient
to apply the classical estimates of Section 4.1 or if one has to apply instead those of Sections 4.2 and 5.1. When the electric
permittivity or the magnetic permeability are scalar-valued coefficients, we write instead of I3 for {, } by abuse of
notation.
To start with, one has an improved regularity result for elements of XB ( ), for B {N , T }. It is based on the regularity
results for scalar fields written next (cf. [30, Corollary 23.5]), and on BirmanSolomyak splittings of elements of XB ( ) from
Theorems 6 and 11 (with = 1).
 P. Ciarlet Jr. / Computers and Mathematics with Applications 71 (2016) 85104 101

Theorem 13. Let be a Lipschitz polyhedron.

If is convex, then
{z H01 ( ) : z XN ( )} H 2 ( );
{z H 1 ( ) : z XT ( )} H 2 ( ).
If is non-convex, then there exist Dir , Neu ]1/2, 1[ that can be explicitly characterized such that,
{z H01 ( ) : z XN ( )} H 1+s ( ), s ]1/2, Dir [;
{z H 1 ( ) : z XT ( )} H 1+s ( ), s ]1/2, Neu [.

In the convex case, we use Dir = Neu = 1.

Remark 10. The characterizations of the regularity exponents Dir , Neu allow one to compute them numerically in principle.
Using a partition of unity as in the proof of Corollary 4 with functions (i )i=1,...,I that fulfill in addition the homogeneous
boundary condition n i | = 0, the embeddings ZT ( ) H 1 ( ) ( convex) and ZT ( ) H s ( ), s ]1/2, Neu [ (
non-convex) follow as direct consequences of Theorem 13. Likewise, using a partition of unity (k )1kK such that k = 1
in a neighborhood of k for all 1 k K , etc., ZN ( ) H 1 ( ) ( convex) and ZN ( ) H s ( ), s ]1/2, Dir [ (
non-convex) are again direct consequences of Theorem 13. Finally, with the help of the BirmanSolomyak splittings, one
gets the by-product below. Note that it includes elements of ZN ( ) and ZT ( ) as particular cases.

Corollary 8. Let be a Lipschitz polyhedron.

If is convex, then for B {N , T }, XB ( ) H 1 ( ).
If is non-convex, then there exist Dir , Neu ]1/2, 1[ such that,
XN ( ) H s ( ), s ]1/2, Dir [;
XT ( ) H s ( ), s ]1/2, Neu [.

Next, one remarks that if the scalar-valued coefficient is smooth, that is W 1, ( ) and 1 L ( ), then given
v L 2 ( ) such that div v L2 ( ), one may write div v = v + div v in , so div v = 1 (div v v ) belongs
to L2 ( ). Hence, one has XB ( , ) XB ( ), for B {N , T }, and one can use Corollary 8 to derive the regularity results that
are needed for obtaining the convergence estimates: Case 2 (convex domain) or Case 3 (non-convex domain) of Section 4.1
now apply.

Theorem 14. Let be a Lipschitz polyhedron such that (Top)I is fulfilled, and assume that the scalar coefficients , fulfill the
coefficient assumption on the trivial partition P = { }. Let := min(Dir , Neu ) ]1/2, 1]. Let (Th )h be a shape regular family
of compatible meshes.
In the electric case, assume that j H ( ). There holds, for s = (convex domain), or for all s [0, [ (non-convex
domain),

inf e vh H (curl; ) .s hs {j 0, + |j |s, }. (41)

vh Vh

In the magnetic case, assume that j0 H ( ). There holds, for s = (convex domain), or for all s [0, [ (non-convex
domain),

inf h vh H (curl; ) .s hs {j 0, + |j0 |s, }. (42)

+
vh Vh

Remark 11. According to Theorem 13, a sufficient condition for 1 j H ( ) to hold is that L2 ( ). Indeed,
j = , with H01 ( ) and div = . In particular, {z H 1 ( ) : z XN ( )}. This observation is
similar in spirit to the one made in footnote 7, page 24.
The next case we consider is when the scalar-valued coefficient is piecewise constant. The regularity results for elements
of XB ( , ) for B {N , T } rely on the regular/singular splitting of scalar fields below (cf. [16,31]).

Theorem 15. Let be a Lipschitz polyhedron, and assume that the scalar, piecewise constant coefficient fulfills the coefficient
assumption.
For all f L2 ( ) and gF PH 1/2 (Fint ), let p be the solution to
Find p H01 ( ) such that

BC = Dir
( p|)0, = (f |)0, + (gF |)0,Fint , H01 ( );
v ( ) such that
1

Find p Hzm
BC = Neu
( p|)0, = (f |)0, + (gF |)0,Fint , 1
Hzmv ( ).
102 P. Ciarlet Jr. / Computers and Mathematics with Applications 71 (2016) 85104

Then for BC {Dir , Neu}, p admits a continuous splitting p = preg + psing , with a regular part preg PH 2 ( ) and a singular
part psing PBC ( ); there exists BC ( ) ]0, 1] that can be explicitly characterized such that PBC ( ) PH 1+s ( ) for all
s ]0, BC ( )[, and PBC ( ) PH 2 ( ) if BC ( ) = 1.

Remark 12. In principle, the characterization of the regularity exponents Dir ( ), Neu ( ) allows one to compute them
numerically.
For B {N , T }, one can then derive embedding results for XB ( , ). To that aim, one uses Theorem 15, which requires to
reformulate the right-hand-sides defining8 p0 and q0 resp. in Theorems 7 and 11, without the part of the decomposition
that belongs to the null spaces ZB ( ). This is achieved with the help of the idempotent operators pB introduced in those
Theorems.
For example, in the electric case, let v ker(pN ) XN ( ). The splitting (24) writes v = vreg + 0 + p0 in , with p0
governed by (25) without the first term in the right-hand side. Hence Theorem 15 may be applied to p0 and one concludes
that
ker(pN ) PH s ( ), s ]0, Dir ( )[; ker(pN ) PH 1 ( ) if Dir ( ) = 1.
Next, let v XN ( ): it is split continuously as v = pN v + (v pN v ). The first part, pN v, belongs to ZN ( ) and as such
its regularity is governed by Corollary 8, with regularity exponent Dir . On the other hand, the operator pN is idempotent so
(v pN v ) ker(pN ) is governed by the regularity exponent Dir ( ).
Similarly for the magnetic case.

Corollary 9. Let the assumptions of Theorem 15 hold. Then

XN ( , ) PH s ( ), s ]0, min(Dir ( ), Dir )[;
XT ( , ) PH ( ), s
s ]0, min(Neu ( ), Neu )[.
Moreover if min(Dir ( ), Dir ) or min(Neu ( ), Neu ) is equal to 1, the corresponding inclusion holds for s = 1.

Remark 13. It can happen that Dir ( ) > Dir , or Neu ( ) > Neu .
Let us now highlight four practical situations, denoted by [c1], [c2], [c3], [c4] below. Two subpolyhedra are adjacent if their
boundaries intersect.

Theorem 16. Let be a Lipschitz polyhedron, and assume that the scalar coefficient is piecewise constant on the partition
P := {j }j=1,...,J . In addition, assume that:
either: [c1] is convex and the maximal number of adjacent subpolyhedra is equal to two;
or: [c2] there exists some j such that j and the maximal number of adjacent subpolyhedra is equal to two.
Then Dir ( ), Neu ( ) ]1/2, 1].

Remark 14. The situation where Dir ( ) or Neu ( ) is equal to 1 can occur only in a convex domain with special geometry,
such as a partition with flat interfaces intersecting the boundary at right angles. The second situation [c2] covers the case of
isolated inclusions of media in an otherwise homogeneous material.
Regarding the convergence rates for the situations [c1] and [c2], one gets results similar to those of Theorem 14 with
:= min(Dir ( ), Neu ( ), Dir , Neu ) ]1/2, 1]. In particular, one may again apply the classical results of Section 4.1 when
the assumptions of Theorem 16 hold.
On the other hand, as soon as there are three adjacent subpolyhedra or more, the regularity exponents Dir ( ) or Neu ( ) can
become arbitrarily close to 0.
[c3] Let us give an illustration: let be the unit cube, P := {j }j=1,2,3,4 where

1 1 1 1
1 := 0, 0, ]0, 1[; 2 := , 1 0, ]0, 1[;
2 2 2 2

1 1 1 1
3 := ,1 , 1 ]0, 1[; 4 := 0, , 1 ]0, 1[.
2 2 2 2
Define such that 1 = 3 = 1, and 2 = 4 = , for some parameter 1. We call this configuration the checkerboard
case. In this case, one may compute directly the regularity exponents by studying singular solutions of the problem Find
z H 1 ( ) such that div z = 0 in some neighborhood of the line x1 = x2 = 21 plus homogeneous boundary condition.

8 Note that g PH 1/2 (F ) in the statement of Theorem 15. So, in the magnetic case, one can use the splitting with boundary condition of Theorem 11,
F int
but not the one of Theorem 9.
 P. Ciarlet Jr. / Computers and Mathematics with Applications 71 (2016) 85104 103

One checks first that 3 z PH 1 ( ) and then that Dir ( ) and Neu ( ) are equal and, in addition, that their common value
is related to the parameter by the relation (see for instance [16, Section 8]):

1/2
= tan .
4
Hence is a decreasing function of , with = 1 if and only if = 1, lim + = 0, and in particular < 1/2 as soon
as > 5.8284. So in the situation [c3] when > 5.8284, one must use the results of Section 4.2 and Section 5.1 to derive
convergence rates, cf. Theorem 12.
[c4] The last case we consider is when the coefficient tensor- or scalar-valued coefficient is piecewise smooth, i.e. when
it fulfills the coefficient assumption on a non-trivial partition. One can apply the frozen coefficients technique, developed in
[32, Section 5.2]. Briefly, it is proven there that one can derive regular/singular splittings, where the singular part is governed
by equations with constant coefficients. The results are derived rigorously in 2D configurations; they can be extended for
instance to the checkerboard case. What is more, the constant coefficients are simply the limit of the value of the coefficients
at the corners of the interface. In principle, one may still compute the regularity exponents in these configurations. We refer
to [33, Section 5] for similar results.

6. Conclusion

We have presented some results on the numerical approximation of low-regularity electromagnetic fields by edge finite
elements. In particular, we addressed the case of general geometrical settings, including topologically non-trivial domains
or domains with a non-connected boundary, and tensor-valued, piecewise smooth electric permittivity and magnetic
permeability. In all cases, a convergence rate in h is recovered, where h is the meshsize, for some exponent ]0, 1]. It relies
either on classical estimates, cf. [20,21,1,23] when > 1/2, or on the combined interpolation operator when < 1/2. The
optimality of the value of has first been discussed with respect to abstract shift theorems. In some simple configurations,
typically for scalar-valued permittivity and permeability, the value of has been further characterized.

Acknowledgments

The author expresses his thanks to Serge Nicaise for mentioning the result of Remark 13, and also to Julia Charrier, Simon
Labrunie and Christian Stohrer for many helpful comments.

References

[1] P. Monk, Finite Element Methods for Maxwells Equations, Oxford University Press, 2003.
[2] A.-S. Bonnet-Ben Dhia, L. Chesnel, P. Ciarlet Jr., Two-dimensional Maxwells equations with sign-changing coefficients, Appl. Numer. Math. 79 (2014)
2941.
[3] A.-S. Bonnet-Ben Dhia, L. Chesnel, P. Ciarlet Jr., T-coercivity for the Maxwell problem with sign-changing coefficients, Comm. Partial Differential
Equations 39 (2014) 10071031.
[4] B. Desprs, L.-M. Imbert-Grard, R. Weder, Hybrid resonance of Maxwells equations in slab geometry, J. Math. Pures Appl. 101 (2014) 623659.
[5] R. Hiptmair, Finite elements in computational electromagnetics, Acta Numer. (2002) 237339.
[6] P. Ciarlet Jr., T-coercivity: application to the discretization of Helmholtz-like problems, Comput. Math. Appl. 64 (2012) 2234.
[7] J.C. Ndlec, Mixed finite elements in R3 , Numer. Math. 35 (1980) 315341.
[8] L. Demkowicz, Asymptotic convergence in finite and boundary element methods. Part I: theoretical results, Comput. Math. Appl. 27 (1994) 6984.
[9] C. Weber, A local compactness theorem for Maxwells equations, Math. Methods Appl. Sci. 2 (1980) 1225.
[10] P. Gross, P. Kotiuga, Electromagnetic Theory and Computation: A Topological Approach, in: MSRI Publications Series, vol. 48, Cambridge University
Press, 2004.
[11] C. Amrouche, C. Bernardi, M. Dauge, V. Girault, Vector potentials in three-dimensional non-smooth domains, Math. Methods Appl. Sci. 21 (1998)
823864.
[12] V. Girault, P.-A. Raviart, Finite Element Methods for NavierStokes Equations, in: Springer Series in Computational Mathematics, vol. 5, Springer
Verlag, 1986.
[13] M.Sh. Birman, M.Z. Solomyak, Maxwell operator in regions with nonsmooth boundaries, Sib. Math. J. 28 (1987) 1224.
[14] D.S. Jerison, C.E. Kenig, The Neumann problem on Lipschitz domains, Bull. Amer. Math. Soc. 4 (1981) 203207.
[15] M. Costabel, A remark on the regularity of solutions of Maxwells equations on Lipschitz domains, Math. Methods Appl. Sci. 12 (1990) 365368.
[16] M. Costabel, M. Dauge, S. Nicaise, Singularities of Maxwell interface problems, Math. Modelling Numer. Anal. 33 (1999) 627649.
[17] A. Bonito, J.-L. Guermond, F. Luddens, Regularity of the Maxwell equations in heterogeneous media and Lipschitz domains, J. Math. Anal. Appl. 408
(2013) 498512.
[18] M.Sh. Birman, M.Z. Solomyak, Construction in a piecewise smooth domain of a function of the class H 2 from the value of the conormal derivative,
J. Sov. Math. 49 (1990) 11281136.
[19] P. Monk, Analysis of a finite element method for Maxwells equations, SIAM J. Numer. Anal. 29 (3) (1992) 714729.
[20] A. Alonso, A. Valli, An optimal domain decomposition preconditioner for low-frequency time harmonic Maxwell equations, Math. Comp. 68 (1999)
607631.
[21] P. Ciarlet Jr., J. Zou, Fully discrete finite element approaches for time-dependent Maxwells equations, Numer. Math. 82 (1999) 193219.
[22] R.A. Adams, Sobolev Spaces, Academic Press, 1975.
[23] A. Bermdez, R. Rodrguez, P. Salgado, Numerical treatment of realistic boundary conditions for the eddy current problem in an electrode via Lagrange
multipliers, Math. Comp. 74 (2005) 123151.
[24] A. Ern, J.-L. Guermond, Theory and Practice of Finite Elements, Springer Verlag, 2004.
[25] S. Brenner, L.R. Scott, The Mathematical Theory of Finite Element Methods, third ed., Springer Verlag, 2008.
[26] J. Schberl, Commuting quasi-interpolation operators for mixed finite elements. Tech. Rep. ISC-01-10-MATH, Texas A&M University, 2001.
[27] J. Schberl, A posteriori error estimates for Maxwell equations, Math. Comp. 77 (2008) 633649.
104 P. Ciarlet Jr. / Computers and Mathematics with Applications 71 (2016) 85104

[28] A. Bermudez, P. Gamallo, M.R. Nogueiras, R. Rodriguez, Approximation of a structural acoustic vibration problem by hexahedral finite elements, IMA
J. Numer. Anal. 26 (2006) 391421.
[29] S.H. Christiansen, R. Winther, Smoothed projections in finite element exterior calculus, Math. Comp. 77 (2008) 813829.
[30] M. Dauge, Elliptic Boundary Value Problems on Corner Domains, in: Lecture Notes in Mathematics, vol. 1341, Springer-Verlag, 1988.
[31] S. Nicaise, A.M. Sndig, Transmission problems for the Laplace and elasticity operators: regularity and boundary integral formulation, Math. Models
Methods Appl. Sci. 9 (1999) 855898.
[32] P. Grisvard, Elliptic Problems in Nonsmooth Domains, Pitman, 1985.
[33] M. Dauge, Neumann and mixed problems on curvilinear polyhedra, Integral Equations Operator Theory 15 (1992) 227261.