See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/225418360

Fully discrete finite element approaches for time-dependent Maxwell

equations

Article in Numerische Mathematik · January 1999

DOI: 10.1007/s002110050417

CITATIONS READS

195 250

2 authors:

Patrick Ciarlet Jun Zou

ENSTA Chinese University of Hong Kong
177 PUBLICATIONS 3,993 CITATIONS 232 PUBLICATIONS 6,866 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Jun Zou on 31 May 2014.

The user has requested enhancement of the downloaded file.

Numer. Math. (1999) 82: 193–219
Numerische
Mathematik
c Springer-Verlag 1999
Electronic Edition

Fully discrete finite element approaches

for time-dependent Maxwell’s equations
P. Ciarlet, Jr1 , Jun Zou2,?
1
Ecole Nationale Supérieure des Techniques Avancées, 32, boulevard Victor,
F-75739 Paris Cedex 15, France; e-mail: [email protected]
2
Department of Mathematics, The Chinese University of Hong Kong, Shatin, N.T.,
Hong Kong; e-mail: [email protected]

Received February 3, 1997 / Revised version received February 27, 1998

Summary. A fully discrete finite element method is used to approximate

the electric field equation derived from time-dependent Maxwell’s equa-
tions in three dimensional polyhedral domains. Optimal energy-norm error
estimates are achieved for general Lipschitz polyhedral domains. Optimal
L2 -norm error estimates are obtained for convex polyhedral domains.
Résumé. On résout, dans un domaine polyédrique, les équations de Maxwell
temporelles. Une méthode par éléments finis discrète en temps et en espace
est proposée pour calculer le champ électrique. Une estimation d’ordre op-
timal est obtenue pour l’erreur en norme-énergie dans le cas général. Pour
la norme L2 , on obtient une estimation optimale dans le cas d’un polyèdre
convexe.
Mathematics Subject Classification (1991): 65N30, 35L15

1. Introduction

Many problems in sciences and industry involve the solutions of Maxwell’s

equations, for example, problems arising in plasma physics, microwave de-
vices, diffraction of electromagnetic waves. In this paper, we are interested in
the numerical solution of time-dependent Maxwell’s equations in a bounded
polyhedral domain in three dimensions. In the literature, one can find a great
deal of work on numerical approximations to time-dependent Maxwell’s
?
The work of this author was partially supported by Hong Kong RGC Grant No. CUHK
338/96E
Correspondence to: J. Zou

Numerische Mathematik Electronic Edition

page 193 of Numer. Math. (1999) 82: 193–219
194 P. Ciarlet, Jr, J. Zou

equations and also analyses on the convergence of numerical schemes for

stationary Maxwell’s equations and related models. We refer readers to
Raviart [21], Assous et al [5], Hewett-Nielson [16], Degond-Raviart [11],
Ambrosiano-Brandon-Sonnendrücker [2] and Ciarlet-Zou [8], etc. But to
our knowledge, it seems that there are few existing works on the convergence
analysis for semi-discrete or fully discrete numerical methods for the time-
dependent Maxwell systems. In [18], Monk obtained error estimates for a
semi-discrete finite element approximation to the time-dependent Maxwell’s
equations using Nédélec’s elements, from which our current paper was initi-
ated. Furthermore, in [17] Makridakis-Monk proposed a fully discrete finite
element scheme and obtained the error estimates under strong regularities
on the solutions. This scheme involves solving coupled non-symmetric and
indefinite linear algebraic systems of both electric and magnetic fields.
The purpose of the current paper is to analyse the convergence of a
simple fully discrete finite element scheme for the electric field equation
derived from Maxwell’s equations by eliminating the magnetic field. The
scheme is a fully discrete version of the semi-discrete scheme studied in
[18], and it is constructed in a way that involves only solving a symmetric
and positive definite linear algebraic system. One of our major interests here
is to investigate the convergence order of the fully discrete scheme without
making use of strong regularities on the solutions, which is certainly of
practical importance. Under appropriate assumptions on the regularity of
the continuous solutions, we derive for the concerned fully discrete scheme
the optimal energy-norm error estimates for general polyhedral domains and
optimal L2 -norm error estimates for convex polyhedral domains.
We now introduce the Maxwell’s equations to be considered in the paper.
Let Ω be a bounded Lipschitz continuous polyhedral domain in R3 , E(x, t)
and H(x, t) the electric and magnetic fields respectively. Then Maxwell’s
equations can be formulated as follows:
(1) εE t + σE − curl H = J in Ω × (0, T ),
(2) µH t + curl E = 0 in Ω × (0, T ),
where ε(x) and σ(x) are the dielectric constant and the conductivity of the
medium respectively, while µ(x) and J (x, t) are the magnetic permeability
of the material in Ω and the applied current density respectively. Here, the
subscript t denotes the time derivative. It is assumed that these coefficients
are piecewise smooth, real, bounded and positive, that is, there exist ε0 > 0
and µ0 > 0 such that, for all x ∈ Ω,
(3) ε0 ≤ ε(x), µ0 ≤ µ(x), and 0 ≤ σ(x).
Moreover, these coefficients ε(x), µ(x) and σ(x) may be discontinuous.
We assume that the boundary of the domain Ω is a perfect conductor, that

Numerische Mathematik Electronic Edition

page 194 of Numer. Math. (1999) 82: 193–219
Fully discrete schemes for Maxwell’s equations 195

is,
(4) E × n = 0 on ∂Ω × (0, T ).
We supplement Maxwell’s equations with the initial conditions

E(x, 0) = E 0 (x) and H(x, 0) = H 0 (x), x ∈ Ω.

Instead of solving the coupled system (1)-(2) with both the electric and
magnetic fields as unknowns, we eliminate the magnetic field H, by taking
the time derivative of (1) and using (2), to obtain the second order electric
field equation:
1
(5) εE tt + σE t + curl ( curl E) = J t , in Ω × (0, T ),
µ
with the boundary condition still being (4) but the previous initial conditions
being replaced by

(6) E(x, 0) = E 0 (x) and E t (x, 0) = E 1 (x),

where E 1 (x) = ε−1 (J (x, 0) + curl H 0 (x) − σ(x)E 0 (x)).

Remark 1.1 We have implicitly assumed that the electromagnetic field is

generated by a current with density J , without any charge density: i.e. the
medium is locally electrically neutral, and div J = 0. In the more general
case, the charge conservation equation reads:

ρt + div J = 0,

where ρ is the charge density.

Therefore, if σ = 0, we derive from (1) and (2) that

div (εE) = −ρ, and div (µH) = 0,

when these relations hold for the initial data. In this case, one may con-
sider a saddle-point approach, like in Raviart [21] or Ciarlet-Zou [8], where
Darwin’s model of approximation to Maxwell’s equations was studied.

We end this section with the introduction of some notations used in the
paper. We define

H(div ; Ω) = {v ∈ (L2 (Ω))3 ; div v ∈ L2 (Ω)},

H(div 0; Ω) = {v ∈ H(div ; Ω); div v = 0},
H(curl ; Ω) = {v ∈ (L2 (Ω))3 ; curl v ∈ (L2 (Ω))3 },
H α (curl ; Ω) = {v ∈ (H α (Ω))3 ; curl v ∈ (H α (Ω))3 },
H0 (curl ; Ω) = {v ∈ H(curl ; Ω); v × n = 0 on Γ },

Numerische Mathematik Electronic Edition

page 195 of Numer. Math. (1999) 82: 193–219
196 P. Ciarlet, Jr, J. Zou

where α is a nonnegative real number. H(div ; Ω), H(curl ; Ω) and

H α (curl ; Ω) are equipped with the norms
 1/2
||v||0,div = ||v||20 + ||div v||20 ,
 1/2
||v||0,curl = ||v||20 + ||curl v||20 ,
 1/2
||v||α,curl = ||v||2α + ||curl v||2α .

Here and in the sequel of the paper, k · k0 will always mean the (L2 (Ω) )3 -
norm (or L2 (Ω)-norm if only scalar functions are involved). And in general,
we will use k · kα and | · |α to denote the norm and semi-norm in the Sobolev
space (H α (Ω) )3 (or H α (Ω) if only scalar functions are involved). We refer
to Adams [1] and Grisvard [15] for more details on Sobolev spaces. C will
always denote a generic constant which is independent of both the time step
τ and the finite element mesh size h.

2. Fully discrete finite element schemes

We consider discretizing the electric field Cauchy problem (5)-(6) by the

implicit backward difference scheme in time together with Nédélec’s finite
elements in space.
Let us first triangulate the space domain Ω and assume that T h is a shape
regular triangulation of Ω with a mesh size h made of tetrahedra. An element
of T h is denoted by K, and the diameters of K and its inscribed ball are
denoted by hK and ρK respectively. As usual, we let h = maxK∈T h hK .
As the triangulation is shape regular, we have hK /ρK ≤ C (cf. Ciarlet [6]).
We then introduce the following Nédélec’s H(curl ; Ω)-conforming finite
element space
Vh = {vh ∈ H(curl ; Ω); vh |K ∈ (P1 )3 , ∀ K ∈ T h}
where P1 is the space of linear polynomials. It was proved in Nédélec [20]
that any function v in Vh can be uniquely determined by the degrees of
freedom in the moment set ME (v) on each element K ∈ T h . Here ME (v)
is defined as follows:
Z
ME (v) = { (v · τ ) q ds; ∀ q ∈ P1 (e) on any edge e of K},
e
where τ is the unit vector along the edge e.
From [3], Lemma 4.7, we know that the integrals required in the definition
of ME (v) make sense for any v ∈ Xp (K), with p > 2, where
Xp (K) = {v ∈ (Lp (K))3 ; curl v ∈ (Lp (K))3 ; v × n ∈ (Lp (∂K))3 }.

Numerische Mathematik Electronic Edition

page 196 of Numer. Math. (1999) 82: 193–219
Fully discrete schemes for Maxwell’s equations 197

Thus we can define, for any v ∈ H 1/2+δ (curl ; Ω)3 with δ > 0 (which
implies that curl v ∈ (Lpδ (K))3 and v ∈ (Lpδ (∂K))3 for some pδ > 2
which depends on δ), an interpolation Πh v of v such that Πh v ∈ Vh and
Πh v has the same degrees of freedom (defined by ME (v)) as v on each
K ∈ T h.
In order to take the boundary condition E × n = 0 on ∂Ω into account,
we define a subspace of Vh :

Vh0 = {vh ∈ Vh ; vh × n = 0 on ∂Ω}.

This can be done simply by zeroing the degrees of freedom which correspond
to the boundary edges.
Next we divide the time interval (0, T ) into M equally-spaced subinter-
vals by using nodal points

0 = t0 < t1 < · · · < tM = T

with tn = nτ , and denote the n-th subinterval by I n = (tn−1 , tn ]. For a

given sequence {un }M 2 2 3
n=0 ⊂ L (Ω) or (L (Ω)) , we introduce the first and
second order backward finite differences:
un − un−1 ∂τ un − ∂τ un−1
∂τ un = , ∂τ2 un = .
τ τ
For a continuous mapping u : [0, T ] → L2 (Ω) or (L2 (Ω))3 , written as u ∈
C(0, T ; (L2 (Ω))3 ) subsequently, we define un = u(·, nτ ) for 0 ≤ n ≤ M .
Using the above notation, our fully discrete finite element approximation
to the electric field equations (5)-(6) is formulated as follows:

(7) E 0h = Πh E 0 , E 0h − E −1
h = τ Πh E 1 ,

and for n = 1, 2, · · · , M , find E nh ∈ Vh0 such that

1
(ε∂τ2 E nh , v) + (σ∂τ E nh , v) + ( curl E nh , curl v) = (∂τ J n , v),
(8) µ
∀v ∈ Vh0 .

Obviously, for each n = 1, 2, · · · , M , it is clear that, by Lax-Milgram

theorem, the system (8) has a unique solution E nh as its left-hand side defines
a symmetric positive definite bilinear form in H(curl ; Ω) with respect to
E nh . In addition, as (8) is symmetric and positive definite, it can be solved
by the well-known conjugate gradient method.

Remark 2.1 Instead of the first order backward difference in time used in the
fully discrete scheme (7)-(8), one can also use some second order difference

Numerische Mathematik Electronic Edition

page 197 of Numer. Math. (1999) 82: 193–219
198 P. Ciarlet, Jr, J. Zou

approximation in time, e.g. the Crank-Nicolson scheme. In this case, the

whole discrete system can be taken as the following:

(9) E 0h = Πh E 0 , E 1h − E −1
h = 2τ Πh E 1 ,

and for n = 0, 1, · · · , M − 1, find E n+1

h ∈ Vh0 such that

1 n
(ε δτ2 E nh , v) + (σ δ2τ E nh , v) + ( curl Ē h , curl v) = (δ2τ J n , v),
(10) µ
∀v ∈ Vh0 .

where δτ2 un = (un+1 −2un +un−1 )/τ 2 , ūn = (un+1 +un−1 )/2, δ2τ un =
(un+1 − un−1 )/(2τ ). Note that the scheme preserves the symmetry and the
positive definiteness. The first unknown E 1h can be solved by using the initial
approximation in (9) and (10) for n = 0, and the resultant linear system
is also symmetric and positive definite. With this scheme we can achieve
similar convergence results as obtained in the paper, see Remark 4.3.

3. Interpolation properties

This section is devoted to some basic approximation properties of the finite

element interpolant Πh defined in Sect. 2, which will be needed in the later
error estimates for the finite element scheme (7)-(8). First of all, we know
the following properties of Πh : for any u ∈ (H 2 (Ω))3 ,

(11) ku − Πh uk0 ≤ C h2 |u|2 ,

while for any u ∈ H 1 (curl ; Ω), we have

(12) kcurl (u − Πh u)k0 ≤ C h kcurl uk1 .

The estimate (11) can be found in Girault [13] (Theorem 3.1) and Nédélec
[20] (Proposition 3). The estimate (12) was proved by Monk [18] (Lemma
2.3).
The estimates (11) and (12) stand for functions which are appropriately
smooth, i.e. for functions in (H 2 (Ω))3 or H 1 (curl ; Ω). But usually the
solutions of the Maxwell system considered in the paper may not have such
kind of regularity, especially when the domain Ω is not convex and only
Lipschitz continuous. Next we are going to present some approximation
properties of the interpolant Πh under weak assumptions on regularity. We
first show a similar result to (12) but for the L2 -norm. Comparing with the
estimate (12) for the curl operator, we lose one error order. Similar results
were obtained in [12] for a different finite element (see Remark 3.3).

Numerische Mathematik Electronic Edition

page 198 of Numer. Math. (1999) 82: 193–219
Fully discrete schemes for Maxwell’s equations 199

Lemma 3.1 We have

ku − Πh uk0 ≤ C h ||u||1,curl , ∀ u ∈ H 1 (curl ; Ω).
The proof of the lemma is omitted, since it can be inferred from that of
Lemmas 3.2 and 3.3 (see [9] for a detailed proof).
Lemma 3.2 We have, for 1/2 < α ≤ 1,
ku − Πh uk0 ≤ C hα ||u||α,curl , ∀ u ∈ H α (curl ; Ω).
Remark 3.1 α > 1/2 is needed for the definition of the moments in
ME (v, φ).
Proof. For any element K ∈ T h , let x = BK x̂ + bK be the affine mapping
between K and the reference element K̂, and we define (cf. Nédélec [20]),
∗ −1 ∗
(13) u(x) = (BK ) û(x̂) or û(x̂) = BK u(x),
where BK ∗ is the transpose of the matrix B . Let Π̂ be the interpolant on
K
the reference element K̂, then
(14) ku − Πh uk2L2 (K) ≤ |BK | k(BK
∗ −1 2
) k kû − Π̂ ûk2L2 (K̂) .

Throughout the paper, |A| means |det(A)| for any square matrix A.
Let us now bound kû− Π̂ ûkL2 (K̂) . For that, let ê (respectively F̂ ) be any
edge (respectively face) of K̂. For p > 2 and p0 such that 1/p + 1/p0 = 1,
on any edge ê of K̂ we define
|Mê (v̂, φ̂)|
(15) ||v̂||Mê = sup
φ̂∈P1 (ê)3 ||φ̂||W 1−1/p0 ,p0 (ê)
R
where Mê (v̂, φ̂) = ê (v̂ · τ̂ )φ̂ds. Using the norm equivalence in finite
dimensional spaces, we have
X X
kΠ̂ ûkL2 (K̂) ≤ C kΠ̂ ûkMê = C kûkMê
ê⊂K̂ ê⊂K̂
n X o
d ûk p
≤ C kcurl + kû × n̂k
L (K̂) Lp (F̂ ) ,
F̂ ⊂K̂

where the last inequality is obtained by integration by parts and the standard
extension and lifting techniques (cf. Lemma 4.7 of [3]). This implies
n X o
d
kû − Π̂ ûkL2 (K̂) ≤ C kcurl ûkLp (K̂) + kûkL2 (K̂) + kû × n̂kLp (F̂ )
F̂ ⊂K̂
n o
d ûk α
≤ C kcurl + kûk
(16) H (K̂) H (K̂) .
α

Numerische Mathematik Electronic Edition

page 199 of Numer. Math. (1999) 82: 193–219
200 P. Ciarlet, Jr, J. Zou

As the left hand side does not change when replacing û by û plus any
constant, we have
n o
d ûk α
kû − Π̂ ûkL2 (K̂) ≤ C kcurl + inf kû + p̂k
H (K̂) H α (K̂) .
p̂∈P0 (K̂)3

Note Z Z 1/2
||ŵ(x̂) − ŵ(ŷ)||2
|ŵ|H α (K̂) = dx̂dŷ ,
K̂ K̂ ||x̂ − ŷ||3+2α
it is clear that |ŵ + p̂|H α (K̂) = |ŵ|H α (K̂) for all p̂ ∈ P0 (K̂)3 . From this
point, one can easily adapt the proof of Theorem 14.1 in [6] to obtain the
norm equivalence in the quotient space H α /P0 . Then one has
n o
d ûk α
kû − Π̂ ûkL2 (K̂) ≤ C kcurl + |û|
(17) H (K̂) H (K̂) .
α

There remains to bound the right-hand side in (17). Noting that ||x−y|| ≤
−1
||BK || ||BK (x − y)||, we deduce
−1 2
|û|2H α (K̂) ≤ ||BK ||5+2α |BK | |u|2H α (K) .

Similarly we have (see [9] for details)

d û||2 2
||curl −1
≤ C||BK ||4 |BK | ||curl u||2L2 (K) , and
L (K̂)
d û|2 α
|curl −1 2
≤ C ||BK ||7+2α |BK | |curl u|2H α (K) .
H (K̂)

This with (17) shows (for ||BK || small)

kû − Π̂ ûk2L2 (K̂)

−1 2 −1
≤ C max(||BK ||5+2α |BK | , ||BK ||4 |BK |)||u||2H α (curl ;K) .

Using the bounds on BK and the shape regularity of T h , we get from (14)
that
ku − Πh uk2L2 (K) ≤ C h2α 2
K ||u||H α (curl ;K) .
2

Remark 3.2 The following lemma is an improvement on the results obtained

in Nédélec [20] (Propositions 1 and 2) and Monk [18] (Lemma 2.3), where
only integers α ≥ 1 were considered.

Lemma 3.3 For 1/2 < α ≤ 1, we have

kcurl (u − Πh u)k0 ≤ Chα |curl u|H α (Ω) , ∀ u ∈ H α (curl ; Ω).

Numerische Mathematik Electronic Edition

page 200 of Numer. Math. (1999) 82: 193–219
Fully discrete schemes for Maxwell’s equations 201

Proof. We follow Nédélec [20] for the notation used below. Let Θh be the
H(div 0; Ω)-conforming space of degree 0:
Θh = {v ∈ H(div 0; Ω); v|K ∈ (P0 )3 , ∀K ∈ T h },
and let rh be the corresponding interpolant to Θh with the moment set:
Z
0
MF (v) = { (v · n) q dσ; ∀ q ∈ P0 (F ) on any face F of K}.
F

There are four degrees of freedom attached to this finite element, but as
div v = 0 by definition, their sum (with q = 1) is equal to 0.
We can show
(18) rh (curl u) = curl (Πh u).
Indeed, curl (Πh u) ∈ (P0 )3 , div (curl (Πh u)) = 0 and

Z Z Z
curl (Πh u) · n dσ = curl F (Πh u) dσ = Πh u · τ ds
F F ∂F
Z Z
= u · τ ds = curl F (u) dσ
Z∂F F

= curl u · n dσ.
F

Hence
(19) kcurl (u − Πh u)k0 = k(curl u) − rh (curl u)k0 ,
this with the following result (20) gives the lemma.
Next, we show that for any element K and 1/2 < α ≤ 1,
kw − rh wkL2 (K) ≤ Chα |w|H α (K) ,
(20)
∀ w ∈ H(div 0; Ω) ∩ H α (Ω)3 .

To prove this, we replace the degrees of freedom in MF0 (v) (cf. [19] or [20])
by Z
2
{ (v · n)qdσ; ∀ q|K ∈ (P0 )3 , K ∈ T h }.
|BK | F
Then the following transformation
x = BK x̂ + bK , w(x) = BK ŵ(x̂),
preserves the interpolation and divergence, i.e.

r̂ŵ(x̂) = rh w(x), d ŵ(x̂) = div w(x),

div

where r̂ is the reference interpolant on K̂.

Numerische Mathematik Electronic Edition

page 201 of Numer. Math. (1999) 82: 193–219
202 P. Ciarlet, Jr, J. Zou

R
Now consider the moment MF̂0 (ŵ, φ̂) = F̂ (ŵ · n̂)φ̂ dσ̂. Noting that
d ŵ = 0, we have by integration by parts
div w = 0 implies div
Z Z
0
|MF̂ (ŵ, φ̂)| = d
div ŵ φ̂ dx̂ + d φ̂ dx̂
ŵ · grad
ZK̂ K̂

= d φ̂ dx̂
ŵ · grad
K̂
(21) ≤ C ||ŵ||Lp (K̂) ||φ̂||W 1−1/p0 ,p0 (F̂ )

where p0 is again the conjugate number of p and φ̂ the extension by zero

0 0 0 0
from W 1−1/p ,p (F̂ ) into W 1−1/p ,p (∂ K̂) combined with a lifting operator
0 0 0
from W 1−1/p ,p (∂ K̂) onto W 1,p (K̂).
Using (21), we can bound

|MF̂0 (ŵ, φ̂)|

||ŵ||M 0 = sup
F̂
φ̂∈(P0 (F̂ ))3 ||φ̂||W 1−1/p0 ,p0 (F̂ )

by kŵkLp (K̂) . This with the norm equivalence in finite dimensional spaces
gives
X X
kr̂ŵkL2 (K̂) ≤ C kr̂ŵkM 0 = C kŵkM 0
F̂ F̂
F̂ ∈K̂ F̂ ∈K̂
≤ C kŵkLp (K̂)3 ≤ C kŵkH α (K̂) ,

that implies
(22) kŵ − r̂ŵkL2 (K̂) ≤ CkŵkH α (K̂) .
As replacing ŵ by ŵ plus any constant does not change the left hand side,
we come to
kŵ − r̂ŵkL2 (K̂) ≤ C|ŵ|H α (K̂) .
Thus we finally derive
kw − rh wk2L2 (K) ≤ kBK k2 |BK | kŵ − r̂ŵk2L2 (K̂) ≤ Ch2α 2
K |w|H α (K) ,

this proves (20). 2

Remark 3.3 All the results of this paper are also valid for certain other
first order, H(curl ; Ω)-conforming, Nédélec’s elements, e.g. the element
defined in [19], i.e. each element function has the form vh|K ∈ R1 (K) =
{aK + bK × x, (aK , bK ) ∈ R6 }, with the related degrees of freedom. The
crucial step for the validity is to establish Lemmas 3.1-3.3 for this element.
In fact, Lemma 3.1 was established in ([12], Theorem 3.2). Lemmas 3.2-
3.3 can be extended to include this first order element by means of similar

Numerische Mathematik Electronic Edition

page 202 of Numer. Math. (1999) 82: 193–219
Fully discrete schemes for Maxwell’s equations 203

techniques as used in the present paper. Note, in addition, that higher order
finite elements are not considered here as we do not assume, for the solutions,
a stronger regularity than H α (curl ; Ω) with 1/2 < α ≤ 1 in Subsect. 4.1
(energy-norm error estimates), or than H 2 (Ω) in Subsect. 4.2 (L2 -norm
error estimates).

4. Finite element error estimates

We are now going to derive the error estimates for the fully discrete finite
element scheme (7)-(8) both in the energy-norm and the L2 -norm. Through-
out this section, E n and E nh will denote the solutions of the electric field
equations (5)-(6) and the finite element approximation (8) at time t = tn .
For the error analysis, we need the solution function E to be defined also
in the interval [−2τ, T ] in terms of the time variable t. This can be done by
extending E with some regularity from the time interval [0, T ] to the interval
[−2τ, T ]. So we shall always implicitly assume that E is well defined in
terms of time variable t on the interval [−2τ, T ]. Furthermore, to achieve
the optimal energy-norm error estimates for the concerned fully discrete
finite element scheme, we introduce an important projection operator Ph :
H0 (curl ; Ω) → Vh0 defined by
(23) a(Ph u, v) = a(u, v), ∀ v ∈ Vh0
where a(u, v) is the scalar product associated with k · k0,curl . Obviously,
Ph is well-defined in H0 (curl ; Ω).
By the definition of the projection Ph in (23), we easily see that, for
α > 1/2,
ku − Ph uk0,curl ≤ ku − Πh uk0,curl ,
(24)
∀ u ∈ H0 (curl ; Ω) ∩ H α (curl ; Ω).
Later on, we will need the following identity
k
X k
X
(25) (am − am−1 )bm = ak bk − a0 b0 − am−1 (bm − bm−1 )
m=1 m=1

and the following estimates for B = H 1 (curl ; Ω)

or B = (H α (Ω))3 with
α ≥ 0,
Z n
1 t
(26) k∂τ un k2B ≤ kut (t)k2B dt, ∀ u ∈ H 1 (0, T ; B),
τ tn−1
Z n
2 n 2 1 t
(27) k∂τ u kB ≤ kutt (t)k2B dt, ∀ u ∈ H 2 (0, T ; B),
τ tn−2
Z tn
n 2 n 2
(28) k∂τ ut − ∂τ u kB ≤ Cτ kuttt (t)k2B dt, ∀ u ∈ H 3 (0, T ; B).
tn−2

Numerische Mathematik Electronic Edition

page 203 of Numer. Math. (1999) 82: 193–219
204 P. Ciarlet, Jr, J. Zou

4.1. Energy-norm error estimates

This subsection is devoted to the estimate on the energy-norm error for

E n − E nh . For the purpose, we first analyse the error ηhk = E kh − Ph E k ,
for 1 ≤ k ≤ n. Once we have estimates for ηhn , we can easily get the error
estimates for E n − E nh by the triangle inequality, the projection properties
(24) and the interpolation properties discussed in Sect. 3.
We conduct our analysis only for the constant coefficients case, i.e., we
assume ε(x), µ(x) and σ(x) are all constants. It is straightforward to extend
the analysis to the non-constant or elementwise constant case by simply
keeping these coefficients inside the integrals or norms and bounding them
by taking their maximum or minimum values if necessary.
To analyse ηhk , we multiply the equation (5) by v/τ ∈ Vh0 and integrate
then the resultant over Ω in space and over I k in time to obtain
Z
1  
ε (∂τ E kt , v) + σ(∂τ E k , v) + curl E dt, curl v = (∂τ J k , v),
τ µ Ik
0
(29) ∀ v ∈ Vh .
Now subtracting (29) from (8) and making some rearrangements, we have
1
ε (∂τ2 ηhk , v) + (curl ηhk , curl v) + σ(∂τ ηhk , v)
µ
   
= ε ∂τ (E kt − ∂τ Ph E k ), v + σ ∂τ (E k − Ph E k ), v
Z
1  
+ curl (E − Ph E k ) dt, curl v , ∀ v ∈ Vh0 .
τ µ Ik

Then taking v = τ ∂τ ηhk = ηhk − ηhk−1 above and using a(a − b) ≥ a2 /2 −

b2 /2, for any real numbers a and b, yield
ε ε 
στ k∂τ ηhk k20 + k∂τ ηhk k20 − k∂τ ηhk−1 k20
2 2
 1 1 
+ kcurl ηhk k20 − kcurl ηhk−1 k20
2µ 2µ
 
≤ τ σ (∂τ E − Ph ∂τ E k ), ∂τ ηhk
k

 
+τ ε ∂τ (E kt − ∂τ E k ), ∂τ ηhk
 
+τ ε ∂τ2 E k − Ph ∂τ2 E k ), ∂τ ηhk
Z
1 
+ curl (E − E k ) dt, curl ∂τ ηhk
µ Ik
τ 
+ curl (E k − Ph E k ), curl ∂τ ηhk
µ

Numerische Mathematik Electronic Edition

page 204 of Numer. Math. (1999) 82: 193–219
Fully discrete schemes for Maxwell’s equations 205

5
X
(30) ≡: (I)i .
i=1

Next, we will estimate (I)i for i = 1, 2, 3, 4, 5 one by one.

First for (I)1 , using Cauchy-Schwarz inequality, we have

1 1
(I)1 ≤ στ k∂τ ηhk k20 + στ k∂τ E k − Ph ∂τ E k k20
2 2
1  
k 2 k k 2
≤ στ k∂τ ηh k0 + C σ τ k∂τ E − Πh ∂τ E k0,curl (by (24))
2
1
≤ στ k∂τ ηhk k20 + C σ τ h2 k∂τ E k k21,curl (by (12) and Lemma 3.1)
2 Z
1 k 2 2
≤ στ k∂τ ηh k0 + C σ h kE t k21,curl dt (by (26)).
2 I k

k
R For the estimation of (I)2 , by writing ∂τ (·) into the integral of form
I k (·)t dt and using Cauchy-Schwarz inequality, we easily come to

Z tk
1
(I)2 ≤ τ εk∂τ ηhk k20 + C ετ 2 kE ttt k20 dt.
2 tk−2

The estimate for (I)3 is achieved using the same technique as used for
(I)1 ,
Z
1
(I)3 ≤ ετ k∂τ ηhk k20 + C ε h2 kE tt k21,curl dt.
2 I k

To analyse (I)4 , we use Green’s formula and the boundary condition to

derive
Z  
1 k k
(I)4 = curl curl (E − E ), ∂τ ηh dt
µ Ik
Z Z tk  
1
=− curl curl E t , ∂τ ηhk dt0 dt,
µ Ik t

then by Cauchy-Schwarz inequality we obtain

Z
1 τ2
(I)4 ≤ τ εk∂τ ηhk k20 + kcurl curl E t k20 dt.
2 2εµ2 Ik

Finally, we estimate (I)5 . By the definition of Ph in (23), we have

τ  τ 
(I)5 = curl (E k − Ph E k ), curl ∂τ ηhk = − E k − Ph E k , ∂τ ηhk ,
µ µ

Numerische Mathematik Electronic Edition

page 205 of Numer. Math. (1999) 82: 193–219
206 P. Ciarlet, Jr, J. Zou

then applying the Cauchy-Schwarz inequality and (12), (24) and Lemma 3.1,
we come to
τ
(I)5 ≤ kE k − Ph E k k0 k∂τ ηhk k0
µ
τ
≤ 2 k∂τ ηhk k20 + C τ h2 kE k k21,curl .
µ
This completes all the estimates for (I)i in (30). Now summing both sides
in (30) over k = 1, 2, · · · , n and making use of the previous estimates for
(I)i (1 ≤ i ≤ 5), lead to
ε 1
k∂τ ηhn k20 + kcurl ηhn k20 ≤ C m0 (E)(τ 2 + h2 ) + d(ηh0 )
2 2µ
X n  
(31) +C τ k∂τ ηhk k20 + kcurl ηhk k20 ,
k=1

where C is a constant depending on the coefficients ε, σ and µ, and m0 (E)

is an a priori bound of E of the following form
Z T
2
m0 (E) = max kE(t)k1,curl + (kE tt k21,curl + kcurl curl E t k20 ) dt
0≤t≤T 0
Z T
+ kE ttt k20 dt,
−τ

while d(ηh0 ) is the initial error

ε 1
d(ηh0 ) = k∂τ ηh0 k20 + kcurl ηh0 k20 ,
2 2µ
which can be analysed as follows:
First by the definitions of E 0h and the projection Ph , we have
kcurl ηh0 k0 = kcurl (Πh E 0 − Ph E 0 )k0 = kcurl Ph (Πh E 0 − E 0 )k0
≤ kΠh E 0 − E 0 k0,curl ≤ C h kE(0)k1,curl .
Then for the first term in d(ηh0 ), by definition of ηh0 , E 0h and E −1
h , we
know
∂τ ηh0 = τ −1 (τ Πh E t (0) − Ph E(0) + Ph E(−τ ))
= τ −1 Ph (E(−τ ) − E(0) + τ E t (0)) + Ph (Πh E t (0) − E t (0)),
using the property of the projection Ph , we derive
k∂τ ηh0 k0 ≤ τ −1 kE(−τ ) − E(0) + τ E t (0)k0,curl
+kΠh E t (0) − E t (0)k0,curl
≤ C τ sup kE tt k1 + C h kE t (0)k1,curl .
(−τ,0)

Numerische Mathematik Electronic Edition

page 206 of Numer. Math. (1999) 82: 193–219
Fully discrete schemes for Maxwell’s equations 207

Therefore, we get the estimates for the initial error d(ηh0 ):

d(ηh0 ) ≤ C τ 2 sup kE tt k21 + C h2 kE(0)k21,curl + kE t (0)k21,curl .
(−τ,0)

Then substituting this into (31) and applying the well-known discrete Gron-
wall’s inequality, we conclude that
 
max k∂τ (E nh − Ph E n )k20 + kcurl (E nh − Ph E n )k20
1≤n≤M

≤ C m0 (E) (τ 2 + h2 ).
Finally, applying the triangle inequality to
∂τ E nh − E nt = (∂τ E nh − Ph ∂τ E n ) + (Ph ∂τ E n − ∂τ E n ) + (∂τ E n − E nt )
and
E nh − E n = (E nh − Ph E n ) + (Ph E n − E n ),
we have proved the following energy-norm error estimates
Theorem 4.1 Let E and E nh be the solutions of the electric field equa-
tions (5)-(6) and the finite element approximation (7)-(8) at time t = tn ,
respectively. Assume that
E ∈ H 2 (0, T ; H0 (curl ; Ω) ∩ H 1 (curl ; Ω) ) ∩ H 3 (0, T ; (L2 (Ω))3 ).
Then we have
 
max k∂τ E nh − E nt k20 + kcurl (E nh − E n )k20 ≤ C (τ 2 + h2 )
1≤n≤M

where C is a constant independent of both the time step τ and the meshsize
h.

Remark 4.1 The error estimate in Theorem 4.1 is optimal both in terms of
time step size τ and mesh size h as we have used only the H 1 (curl ; Ω)-
regularity in space and H 3 (0, T )-regularity in time.
If the solution E has no so much regularity in space as in Theorem 4.1 (cf.
Costabel [10], Assous et al [4]), we then have the following weaker error
estimates
Theorem 4.2 Let E and E nh be the solutions of the electric field equa-
tions (5)-(6) and the finite element approximation (7)-(8) at time t = tn ,
respectively. Assume that for some 1/2 < α < 1,
E ∈ H 2 (0, T ; H0 (curl ; Ω) ∩ H α (curl ; Ω) ) ∩ H 3 (0, T ; (L2 (Ω))3 ).
Then we have
 
max k∂τ E nh −E nt k20 +kcurl (E nh −E n )k20 ≤ C (τ 2 +τ 2 h2(α−1) +h2α ).
1≤n≤M

Numerische Mathematik Electronic Edition

page 207 of Numer. Math. (1999) 82: 193–219
208 P. Ciarlet, Jr, J. Zou

Proof. The proof is almost identical to the one for Theorem 4.1 by replacing
H 1 (curl ; Ω) by H α (curl ; Ω), Lemma 3.1 by Lemmas 3.2-3.3, (12) by
Lemma 3.3. The only remaining term we have to re-estimate is the term
(I)4 in (30) as we now have no regularity curl (curl E t ) in (L2 (Ω))3 as
used in that proof. Instead we can bound the term as follows: by assumption
we have E ∈ H 1 (0, T ; H α (curl ; Ω) ), so curl E ∈ H 1 (0, T ; (H α (Ω))3 )
and then curl (curl E) ∈ H 1 (0, T ; (H α−1 (Ω))3 ), where H α−1 (Ω) is the
dual space of H 1−α (Ω) (note that this is true as, when 0 < 1 − α < 1/2,
H 1−α (Ω) = H01−α (Ω)). Therefore, by Green’s formula and the inverse
inequality we have
Z D E
1
(I)4 = curl curl (E − E k ), ∂τ ηhk α−1 1−α dt
µ Ik H ,H
Z Z tk D E
1
=− curl (curl E t ), ∂τ ηhk α−1 1−α dt0 dt
µ Ik t H ,H
Z
τ
≤ kcurl (curl E t )kH α−1 (Ω) k∂τ ηhk kH 1−α (Ω) dt
µ Ik
Z
α−1 k
≤ Cτ h k∂τ ηh k0 kcurl (curl E t )kH α−1 (Ω) dt
Ik
Z
1 k 2 2 2(α−1)
≤ τ εk∂τ ηh k0 + Cτ h kcurl (curl E t )k2H α−1 (Ω) dt.
4 Ik

That completes the proof of Theorem 4.2. 2

Remark 4.2 It is easy to see that if we take τ to be the same magnitude as

h, then the error estimate in Theorem 4.2 is of optimal order, i.e. O(h2α ),
in the sense of the space regularity we have used.

Remark 4.3 Theorems 4.1-4.2 can be extended to the Crank-Nicolson

scheme (9)-(10). The major difference in proving the related theorems is to
derive a similar equation as (29) and to choose an appropriate test function v.
For the former, we may multiply equation (5) at time t = tn by v/τ ∈ Vh0
and then integrate the resulting equation over Ω. For the latter, we may
choose the test function v = ηhn+1 − ηhn−1 = (ηhn+1 − ηhn ) + (ηhn − ηhn−1 ).

4.2. L2 -norm error estimates

This subsection is dedicated to the derivation of the L2 -norm error estimates.

The basic technique used here is borrowed from Girault [13] and Monk [18],
where Nédélec’s finite element methods were applied for stationary Navier-
Stokes equations and semi-discrete schemes with Nédélec’s finite elements
for time-dependent Maxwell’s equations, respectively.

Numerische Mathematik Electronic Edition

page 208 of Numer. Math. (1999) 82: 193–219
Fully discrete schemes for Maxwell’s equations 209

As usual, we assume the domain Ω is a convex polyhedron in order to

achieve the optimal L2 -norm error estimates. If the domain Ω is a general
Lipschitz polyhedron, we have the following L2 -norm error estimates: for
E ∈ H 3 (0, T ; (L2 (Ω))3 ) and E ∈ H 2 (0, T ; H0 (curl ; Ω)∩H 1 (curl ; Ω) ),
max kE nh − E n k0 ≤ C (τ + h),
1≤n≤M

which can be obtained immediately by applying the triangle inequality to

the relation
n
X
E nh − E n = (Πh E 0 − E 0 ) + τ (∂τ E kh − ∂τ E k )
k=1
Xn n
X
= (Πh E 0 − E 0 ) + τ (∂τ E kh − E kt ) + τ (E kt − ∂τ E k )
k=1 k=1

and the results in Theorem 4.1.

A similar result, when E is only in H 2 (0, T ; H0 (curl ; Ω)∩
H (curl ; Ω) ) and H 3 (0, T ; (L2 (Ω))3 ), for 1/2 < α < 1, is
α

max kE nh − E n k0 ≤ C hα ,
1≤n≤M

by using Theorem 4.2 and taking τ to be of order O(h).

But if the domain Ω is convex and the solution E is more regular, we
can expect higher convergence order. We now show this is possible. We will
consider only the case that the coefficients ε and µ are constants (we take 1
for simplicity) and σ = 0. We need the following decomposition (cf. Ciarlet
[7])
(32) L2 (Ω)3 = M ⊕ M ⊥
where M and M ⊥ are two spaces defined by
M = {v = ∇z; z ∈ H01 (Ω)}
and
M ⊥ = {v ∈ L2 (Ω)3 ; (v, ∇z) = 0 ∀ z ∈ H01 (Ω)} = H(div 0; Ω).
The last equality can be proved directly from definition. As Ω is convex, we
have H0 (curl ; Ω) ∩ M ⊥ ⊂ H 1 (Ω)3 (see [14]).
To introduce a discrete version of (32), we define a finite element space
Sh = {zh ∈ H01 (Ω); zh |K ∈ P2 ∀ K ∈ T h }.
An L2 (Ω) projection Qh onto Sh will be needed in our later analysis, i.e.
we define Qh : L2 (Ω) 7→ Sh as follows
(Qh z, vh ) = (z, vh ), ∀ z ∈ L2 (Ω), vh ∈ Sh .

Numerische Mathematik Electronic Edition

page 209 of Numer. Math. (1999) 82: 193–219
210 P. Ciarlet, Jr, J. Zou

Now we can introduce the following discrete decomposition

(33) Vh0 = Mh ⊕ Mh⊥
where Mh and Mh⊥ are two spaces defined by
Mh = {v = ∇zh ; zh ∈ Sh }
and
Mh⊥ = {v ∈ Vh0 ; (v, ∇zh ) = 0 ∀ zh ∈ Sh }.
Note that, by definition, M (resp. Mh ) is the kernel of the curl operator
in H0 (curl ; Ω) (resp. Vh0 ).
Before discussing our main results in this section, we introduce some
operators which are very useful in the later L2 -norm error estimates. We
first define an operator T . For any u ∈ M ⊥ , T u ∈ H0 (curl ; Ω) ∩ M ⊥ and
satisfies
(curl (T u), curl v) = (u, v), ∀ v ∈ H0 (curl ; Ω) ∩ M ⊥ .
For a given u, T u can be regarded also as the solution, in variable w, to the
following system (cf. Girault [13]):
(34) curl (curl w) = u in Ω,
(35) div w = 0 in Ω,
(36) w × n = 0 on Γ.
The problem (34)-(36) is a particular case of the system in (w0 , p):
(37) curl (curl w0 ) + ∇ p = g in Ω,
(38) div w0 = 0 in Ω,
(39) w0 × n = 0 on Γ,
(40) p=0 on Γ,
for a given g in (L2 (Ω))3 . According to Lemma 4.1 in [13], the system (37)-
(40) has a unique solution w0 ∈ H0 (curl ; Ω)∩H(div 0; Ω) and p ∈ H01 (Ω).
In addition, the following properties hold:
– w0 ∈ H 1 (Ω)3 and ||w0 ||1 ≤ C ||curl w0 ||0 ≤ C ||g||0 .
– curl w0 ∈ H 1 (Ω)3 and ||curl w0 ||1 ≤ C ||g||0 .
– |p|1 ≤ C ||g||0 .
Clearly, if g ∈ M ⊥ , then p = 0: this is the case with g = u.
The discrete version Th : M ⊥ → Mh⊥ of the operator T can be defined
by
(curl (Th u), curl vh ) = (u, vh ), ∀ u ∈ M ⊥ , vh ∈ Mh⊥ .
We now state some properties of the operators T and Th . The results (i)
and (ii) were established by Monk in [18], so we focuse on the proof of (iii)
and (iv). The property (iii) is an improvement over the result in [18].

Numerische Mathematik Electronic Edition

page 210 of Numer. Math. (1999) 82: 193–219
Fully discrete schemes for Maxwell’s equations 211

Lemma 4.1 Let Ω be a convex polyhedron and suppose T u ∈ H 2 (Ω)3 , for

u ∈ H0 (curl ; Ω)∩H(div 0; Ω). Then there exists a constant C independent
of h and u such that
(i) kcurl ((T − Th )u)k0 ≤ C h kcurl (T u)k1 ,
(ii) kcurl (Th u)k0 ≤ C kuk0 ,
(iii) k(T − Th )uk0 ≤ C h2 kT uk2 ,
1/2
(iv) kTh uk0 ≤ C kuk0 .

Proof. Notice that by using Weber’s result [22], one easily gets that
||curl w||0 defines a norm in H0 (curl ; Ω)∩H(div 0; Ω) which is equivalent
to its canonical norm.
Now, as p = 0, w = T u is the solution to:
(curl w, curl v) = (u, v), ∀ v ∈ H0 (curl ; Ω),
and wh = Th u ∈ Mh⊥ also satisfies the discrete variational formulation:

(curl wh , curl vh ) = (u, vh ), ∀ vh ∈ Vh0 ⊂ H0 (curl ; Ω).

Therefore, kcurl (w − wh )k0 ≤ ||curl w||0 ≤ C ||u||0 .
We first prove the fourth inequality. We see
1/2
||Th u||20 = (Th u, u) = (wh , u) = (wh − w, u) + (w, u).
Define q ∈ H01 (Ω) by
(∇q, ∇µ) = (wh , ∇µ), ∀µ ∈ H01 (Ω).
Note that ∇q and u belong to orthogonal subspaces of L2 (Ω)3 . Set v =
wh − ∇q: curl v = curl wh , div v = 0, v × n = 0 on Γ . Therefore
wh − ∇q − w ∈ H0 (curl ; Ω) ∩ H(div 0; Ω) and
||wh − ∇q − w||0 ≤ C ||curl (wh − w)||0 ≤ C ||curl w||0 ≤ C ||u||0 .
Also, ||w||0 ≤ C ||curl w||0 ≤ C ||u||0 , which proves (iv).

Now we turn to the third inequality: take any g in L2 (Ω)3 and solve
(37)-(40). Then
(g, wh − w)
= (curl w0 , curl (wh − w)) + (∇p, wh − w)
= (curl (w0 − wh∗ ), curl (wh − w)) + (∇(p − ph ), wh − w)
= (curl (w0 − wh∗ ), curl (wh − w)) + (∇(p − ph ), wh − yh )
(41) +(∇(p − ph ), yh − w), ∀ wh∗ , yh ∈ Vh0 , ∀ ph ∈ Sh .

Numerische Mathematik Electronic Edition

page 211 of Numer. Math. (1999) 82: 193–219
212 P. Ciarlet, Jr, J. Zou

Set zh = wh − yh and split up zh into zh = ∇q + v, with q ∈ H01 (Ω)

defined by
(∇q, ∇µ) = (zh , ∇µ), ∀µ ∈ H01 (Ω).
As usual v = zh − ∇q is an element of H0 (curl ; Ω) ∩ H(div 0; Ω). As zh
belongs to Vh0 , it has sufficient regularity for defining Πh zh = zh . Therefore,
it can also be split up into
zh = Πh v + ∇qh , with qh ∈ Sh .
Remark 4.4 To prove the splitting one can write Πh (∇q) as ∇qh , then use
(18) which leads to curl (Πh (∇q)) = 0: thus there exists q 0 ∈ H 1 (Ω)
such that Πh (∇q) = ∇q 0 . Moreover, by definition, Πh (∇q) is in Vh0 , so
its restriction to any tetrahedron K belongs to (P1 )3 . Putting these together
shows that q 0 is an element of Sh .
Now, the estimate of (∇(p − ph ), zh ) proceeds as in Girault [13]: choose
for ph the H01 -projection of p into Sh , hence
(42) |(∇(p − ph ), zh )| ≤ |p|1 ||Πh v − v||0 .
We are going to estimate ||Πh v − v||0 . First, following [3],
n X o
d v̂k p
kv̂ − Π̂ v̂kL2 (K̂) ≤ C kcurl L (K̂) + kv̂kL2 (K̂) + kv̂ × n̂kLp (F̂ ) .
F̂ ⊂K̂

d v̂ =
But curl v = curl zh and (13) preserves the curl . Therefore, curl
d ẑh which belongs to a finite dimensional space. Thus,
curl
d v̂k p
kcurl d
L (K̂) ≤ Ĉ kcurl v̂kL2 (K̂) .

Using the norm equivalence in the quotient space H 1 /P0 , we come to

n o
d v̂k 2
kv̂ − Π̂ v̂kL2 (K̂) ≤ C kcurl + |v̂|
L (K̂) H (K̂) .
1

Thanks to the estimates which bound the right-hand side (cf. [9] and [6]), it
follows from (14) that
kv − Πh vkL2 (K) ≤ C hK |v|H 1 (K) .
Thus
kv − Πh vk0 ≤ C h |v|1 = C h ||curl v||0
≤ C h ||curl zh ||0 = C h ||curl (wh − yh )||0 .
We then have from (42) that
|(∇(p − ph ), zh )| ≤ C h |p|1 {||curl (wh − w)||0 + ||curl (w − yh )||0 } .

Numerische Mathematik Electronic Edition

page 212 of Numer. Math. (1999) 82: 193–219
Fully discrete schemes for Maxwell’s equations 213

Therefore, the last two terms in (41) are bounded by

|p|1 {C h (||curl (wh − w)||0 + ||curl (w − yh )||0 ) + ||w − yh ||0 } ,
∀yh ∈ Vh0 .
By assumption, w belongs to H 2 (Ω): choosing yh = Πh w gives
kw − yh k0 ≤ C h2 |w|2 ,
(43) kcurl (w − yh )k0 ≤ C h kcurl wk1 .
And, with the help of (i), we obtain the final bound for the last two terms in
(41), that is
C h2 |p|1 kwk2 ≤ C h2 ||g||0 kwk2 .
To conclude, as w0 belongs to H 1 (curl ; Ω) ∩ H0 (curl ; Ω), we can choose
wh∗ = Πh w0 , and using (43) and (i), we have the final bound for the first
term in (41),
(curl (w0 − wh∗ ), curl (wh − w)) ≤ C h2 ||g||0 ||w||2 .
This proves (iii). 2

Now we are in a position to derive the main results in this section, i.e.
the L2 -norm error estimate. By means of the decomposition (32), for any
J (t) ∈ L2 (Ω), t ∈ (0, T ), we can write
(44) J = J 1 + ∇z, J 1 ∈ M ⊥ , z ∈ H01 (Ω).
This decomposition enables us to separate our error estimates into two parts,
i.e. Theorems 4.3 and 4.4.
The general principles of the proofs for these two theorems are similar
to those of the proofs for Theorems 4.1 and 4.2 in [18]. But as our scheme
is fully discrete, a lot of technical details need to be treated newly. We will
give only an outline for each proof but refer to [9] for details.
First we show
Theorem 4.3 Let E and E nh be the solutions to (5)-(6) and (7)-(8). Assume
that J (t) ∈ M ⊥ for t ∈ (0, T ), and that E 0 and E 1 belong to M ⊥ ∩
(H 2 (Ω))3 . Moreover we assume that
E ∈ C 1 (0, T ; (H 2 (Ω))3 ) ∩ C 2 (0, T ; (H 1 (Ω))3 ) ∩ H 3 (0, T ; (L2 (Ω))3 ).
Then we have
kE n − E nh k0 ≤ C (τ + h2 ) for n = 1, 2, · · · , M
where the constant C is independent of τ and h but may depend on the a
priori bounds on E.

Numerische Mathematik Electronic Edition

page 213 of Numer. Math. (1999) 82: 193–219
214 P. Ciarlet, Jr, J. Zou

Proof. By the assumption, it is easy to derive E(t) ∈ M ⊥ . Using (33), we

k k
can write E kh = E h + ∇zhk with E h ∈ Mh⊥ and zhk ∈ Sh , for 1 ≤ k ≤ n.
Then
k
(45)E k − E kh = (E k − E h ) − ∇zhk ≡ ηhk − ∇zhk , for 1 ≤ k ≤ n.
It suffices to estimate the L2 -norms for both ηhn and ∇zhn .
We first analyse ∇zhn . Taking v = ∇∂τ zhk ∈ Mh for any zhk ∈ Sh in (8)
and the Cauchy-Schwarz inequality, we get
k∂τ ∇zhn k20 ≤ k∂τ ∇zhn−1 k20 ≤ · · · ≤ k∂τ ∇zh0 k20 .
But by orthogonality, we can write
k∂τ ∇zh0 k20 = (∂τ E 0h , ∂τ ∇zh0 ) = (∂τ E 0h − ∂τ E 0 , ∂τ ∇zh0 ),
from which and the initial condition (7) we can show
(46) k∂τ ∇zh0 k0 ≤ C(h2 kE 1 k2 + τ sup kE tt (t)k1 ).
t∈(−τ,0)

Similarly, we have k∇zh0 k0 ≤ C h2 |E 0 |2 . Using this, (46) and the identity

n
X
∇zhn = ∇zh0 + τ ∂τ ∇zhk ,
k=1

we derive that
(47) k∇zhn k0 ≤ C h2 (kE 1 k2 + kE 0 k2 ) + Cτ sup kE tt (t)k1 .
t∈(−τ,0)

Next we estimate ηhk in (45). Taking v = Th φh in (8) with φh ∈ Mh⊥ ,

we have, for 1 ≤ k ≤ n,
k k
(48) (∂τ2 Th E h , φh ) + (E h , φh ) = (∂τ Th J k , φh ), ∀ φh ∈ Mh⊥ .
A similar relation to (48) for the continuous solution E can be derived
by multiplying (5) by v = τ −1 T φ with φ ∈ M ⊥ and then integrating over
I k in time and over Ω in space:
k
(49) (∂τ T E kt , φ) + (Ẽ , φ) = (∂τ T J k , φ), ∀ φ ∈ M ⊥ ,
k R
where Ẽ = τ −1 I k E(t) dt.
Note for any φh ∈ Mh⊥ , we can write φh = φ + ∇z1 with φ ∈ M ⊥ and
z1 ∈ H01 (Ω). Using this relation we obtain from (49) that
k
(50) (∂τ T E kt , φh ) + (Ẽ , φh ) = (∂τ T J k , φh ), ∀ φh ∈ Mh⊥ .

Numerische Mathematik Electronic Edition

page 214 of Numer. Math. (1999) 82: 193–219
Fully discrete schemes for Maxwell’s equations 215

Subtracting (48) from (50) and making some arrangements, we derive for
any φh ∈ Mh⊥ that

(51) (∂τ2 Th ηhk , φh ) + (ηhk , φh )

k
= ( (T − Th )∂τ J k , φh ) + (∂τ2 Th E k − T ∂τ E kt , φh ) + (E k − Ẽ , φh ),
k
then taking φh = ∂τ E h gives
(∂τ2 Th ηhk , ∂τ ηhk ) + (ηhk , ∂τ ηhk )
k
= ( Th ∂τ2 ηhk , ∂τ E k ) + (ηhk , ∂τ E k ) + ( (Th − T )∂τ J k , ∂τ E h )
k k k
+(Ẽ − E k , ∂τ E h ) + (∂τ T E kt − ∂τ2 Th E k , ∂τ E h ).
Note that, (52) holds actually for any φh ∈ Vh0 by orthogonality. Then,
adding the previous equation to the equation (52) with φh = Πh ∂τ E k ∈ Vh0 ,
we get
(∂τ2 Th ηhk , ∂τ ηhk ) + (ηhk , ∂τ ηhk )
= ( Th ∂τ2 ηhk , ∂τ E k − Πh ∂τ E k ) + (ηhk , ∂τ E k − Πh ∂τ E k )
k k k
+( (Th − T )∂τ J k , ∂τ E h − Πh ∂τ E k ) + (Ẽ − E k , ∂τ E h − Πh ∂τ E k )
k
+(∂τ T E kt − ∂τ2 Th E k , ∂τ E h − Πh ∂τ E k )
X5
≡: (I)i ,
i=1

which implies, using a(a − b) ≥ 12 a2 − 12 b2 , for any a, b ∈ R,

 
1 1/2 k 2 1 1/2 k−1 2
kT ∂τ ηh k0 − kTh ∂τ ηh k0
2 h 2
  X 5
1 k 2 1 k−1 2
(52) + kηh k0 − kηh k0 ≤ τ (I)i .
2 2
i=1

The terms (I)i , i = 1, · · · , 5 can be estimated one by one (cf. [9] for details).
With these estimates, we then sum up (52) over k = 1, · · · , n and finally
obtain
1/2 1/2
kTh ∂τ ηhn k20 + kηhn k20 ≤ kTh ∂τ ηh0 k20 + kηh0 k20 + C(E)(τ 2 + h4 )
≤ C(E)(τ 2 + h4 ).
This with (45) and (47) implies the desired estimate of Theorem 4.3. 2

Now we turn to the second part of the L2 -norm error estimate, i.e. we
are going to prove

Numerische Mathematik Electronic Edition

page 215 of Numer. Math. (1999) 82: 193–219
216 P. Ciarlet, Jr, J. Zou

Theorem 4.4 Let E and E nh be the solutions to (5)-(6) and (7)-(8). Assume
that J is in H 1 (0, T ; M ∩ (H 2 (Ω))2 ), and that E 0 and E 1 belong to
M ∩ (H 2 (Ω))3 . Moreover we assume that
E ∈ C 2 (0, T ; (H 2 (Ω))3 ) ∩ H 3 (0, T ; (L2 (Ω))3 ).
Then we have
kE n − E nh k0 ≤ C (τ + h2 ) for n = 1, 2, · · · , M
where the constant C is independent of τ and h but may depend on the a
priori bounds on E.
Proof. By the assumption, it is easy to get E(t) ∈ M . So we have E(t) =
k
∇z(t) with z ∈ H01 (Ω). Using (33), we can write E kh = E h + ∇zhk with
k
E h ∈ Mh⊥ and zhk ∈ Sh , for 1 ≤ k ≤ n. Note E 0 ∈ M , so E 0 = ∇z 0 for
some z 0 ∈ H01 (Ω), thus we have E 0h = Πh E 0 = Πh ∇z 0 = ∇z 0h for some
0 −1
z 0h ∈ Sh , i.e. E h = 0. Similarly we have E h = 0.
By means of these decompositions, we may write
Xn Z tk
n n 0 0
E − E h = (E − E h ) + (E t − ∂τ E kh )dt
k=1 tk−1
n Z
X tk
= (E 0 − Πh E 0 ) + ∇(zt − ∂τ zhk )dt
k=1 tk−1
n Z
X tk
k
− ∂τ E h dt.
k=1 tk−1

which, with the fact E(t) = ∇z(t), thus

Z T 1/2
k
k∇(zt − ∂τ zh )k0 ≤ τ kE tt k20 dt + k∇(ztk − ∂τ zhk )k0 ,
0
implies
n
X n
X k
kE n − E nh k0 ≤ τ k∇(ztk − ∂τ zhk )k0 + τ k∂τ E h k0
k=1 k=1
Z T 1/2
(53) +Ch2 |E 0 |2 + τ kE tt k20 dt .
0
It remains to estimate the first and second terms in the right side of (53). Let
k k
us first estimate the term k∂τ E h k0 . Taking v = ∂τ E h ∈ Mh⊥ in (8), and
using orthogonality and J (t) = ∇g(t) for some g ∈ H01 (Ω), we get
k k k k k
(∂τ2 E h , ∂τ E h ) + (curl E h , curl ∂τ E h ) = (∂τ ∇g k , ∂τ E h ),

Numerische Mathematik Electronic Edition

page 216 of Numer. Math. (1999) 82: 193–219
Fully discrete schemes for Maxwell’s equations 217

which implies
 
1 k 2 1 k−1 2
k∂τ E h k0 − k∂τ E h k0
2 2
 
1 k 2 1 k−1 2 k
+ kcurl E h k0 − k∇E h k0 ≤ τ (∂τ ∇g k , ∂τ E h ).
2 2

Summing up the above equations over k = 1, · · · , n, we obtain

n
X
n n k
k∂τ E h k20 + kcurl E h k20 ≤ 2τ (∂τ ∇g k , ∂τ E h )
k=1
Xn  
k
≤ 2τ ∇(∂τ g k − Qh ∂τ g k ), ∂τ E h
k=1
Xn Z T
k
≤ τ2 k∂τ E h k20 + Ch4 kgt k23 dt.
k=1 0

That means by Gronwall’s inequality and J (t) = ∇g that

Z T
n n
(54) k∂τ E h k20 + kcurl E h k20 ≤ Ch 4
kJ t k22 dt.
0

Next, we are going to estimate the term k∇(ztk −∂τ zhk )k0 in (53). Taking
v = ∇yhk for any yhk ∈ Sh in (29) and then subtracting it from (8) yields

(∂τ ∇(ztk − ∂τ zhk ), ∇yhk ) = 0, ∀ yhk ∈ Sh ,

using this, we obtain

1 1
k∇(ztk − ∂τ zhk )k20 − k∇(ztk−1 − ∂τ zhk−1 )k20
2 2
≤ τ (∂τ ∇(zt − ∂τ zh ), ∇(ztk − yhk ) ).
k k

summing the equations over k = 1, · · · , n and using (25) gives for any
yhk ∈ Sh ,

k∇(ztn − ∂τ zhn )k20

n
X
≤ k∇(zt0 − ∂τ zh0 )k20 + τ (∇(ztk − ∂τ zhk ), ∇∂τ (ztk − yhk ) )
k=1
+(∇(ztn − ∂τ zhn ), ∇(ztn − yhn ) ) − (∇(zt0 − ∂τ zh0 ), ∇(zt0 − yh0 ) ).
(55)

Numerische Mathematik Electronic Edition

page 217 of Numer. Math. (1999) 82: 193–219
218 P. Ciarlet, Jr, J. Zou

Now taking yhk = Qh ∂τ z k , the terms in (55) can be bounded as follows:

Z tk
k k 2
k∇(zt − yh )k0 ≤ Cτ kE tt k20 dt + Ch4 max |E t |22 dt,
tk−1 0≤t≤T
Z tk
k∇∂τ (ztk − yhk )k20 ≤ Cτ kE ttt k20 dt + Ch4 max |E tt |22 ,
tk−1 0≤t≤T

and
k∇(zt0 − ∂τ zh0 )k0 ≤ Ch2 |E 1 |2 .
Substituting these estimates into (55), we have
(56) k∇(ztn − ∂τ zhn )k20 ≤ C(E)(τ 2 + h4 ).
Theorem 4.4 then follows from (53), (54) and (56). 2

Acknowledgements. The authors wish to thank the two anonymous referees for many con-
structive comments. The authors also thank Prof. Vivette Girault for many helpful discussions
and for providing the improved error estimate (iii) of Lemma 4.1.

References
1. Adams, R.A. (1975): Sobolev spaces. Academic Press, New York
2. Ambrosiano, J.J., Brandon, S.T., Sonnendrücker, E. (1995): A finite element formulation
of the Darwin PIC model for use on unstructured grids. J. Comput. Physics 121, 281–297
3. Amrouche, C., Bernardi, C., Dauge, M., Girault, V.: Vector potentials in three-
dimensional nonsmooth domains. To appear in Math. Meth. Appl. Sci.
4. Assous, F., Ciarlet, Jr, P., Sonnendrücker, E.: Resolution of the Maxwell equations in a
domain with reentrant corners. To appear in RAIRO Math. Modelling Numer. Anal.
5. Assous, F., Degond, P., Heintzé, E., Raviart, P.-A., Segré, J. (1993): On a finite-element
method for solving the three-dimensional Maxwell equations. J. Comput. Physics 109,
222–237
6. Ciarlet, P. (1991): Basic error estimates for elliptic problems. In: P. Ciarlet and J.-L.
Lions, eds., Handbook of Numerical Analysis, Volume II, pp. 17–352. North Holland
7. Ciarlet, Jr, P. (1993): A decomposition of L2 (Ω)3 and an application to magnetostatic
equations. Math. Models Meth. App. Sci. 3, 289–301
8. Ciarlet, Jr, P., Zou, J. (1997): Finite element convergence for the Darwin model to
Maxwell’s equations. RAIRO Math. Modelling Numer. Anal. 31, 213–250
9. Ciarlet, Jr, P., Zou, J. (1996): Fully discrete finite element approaches for time-dependent
Maxwell equations. Research Report MATH-96-31 (105), Department of Mathematics,
The Chinese University of Hong Kong
10. Costabel, M. (1990): A remark on the regularity of solutions of Maxwell’s equations
on Lipschitz domains. Math. Meth. Appl. Sci. 12, 365–368
11. Degond, P., Raviart, P.-A. (1992): An analysis of the Darwin model of approximation
to Maxwell’s equations. Forum Math. 4, 13–44
12. Dubois, F. (1990): Discrete vector potential representation of a divergence-free vector
field in three-dimensional domains: Numerical analysis of a model problem. SIAM J.
Numer. Anal. 27, 1103–1141

Numerische Mathematik Electronic Edition

page 218 of Numer. Math. (1999) 82: 193–219
 Fully discrete schemes for Maxwell’s equations 219

13. Girault, V. (1988): Incompressible finite element methods for Navier-Stokes equations
with nonstandard boundary conditions in IR3 . Math. Comp. 51, 55–74
14. Girault, V., Raviart, P.-A. (1986): Finite element methods for Navier–Stokes equations.
Springer–Verlag, Berlin
15. Grisvard, P. (1985): Elliptic problems in nonsmooth domains. Pitman Advanced
Publishing Program, Boston
16. Hewett, D.W., Nielson, C. (1978): A multidimensional quasineutral plasma simulation
model. J. Comput. Phys. 29, 219–236
17. Makridakis, Ch.G., Monk, P. (1995): Time-discrete finite element schemes for
Maxwell’s equations. RAIRO Math. Modelling Numer. Anal. 29, 171–197
18. Monk, P. (1992): Analysis of a finite element method for Maxwell’s equations. SIAM
J. Numer. Anal. 29, 714–729
19. Nédélec, J.-C. (1980): Mixed finite elements in IR3 . Numer. Math. 35, 315–341
20. Nédélec, J.-C. (1986): A new family of mixed finite elements in IR3 . Numer. Math. 50,
57–81
21. Raviart, P.-A. (1993): Finite element approximation of the time-dependent Maxwell
equations. Technical Report GdR SPARCH #6, Ecole Polytechnique, France
22. Weber C. (1980): A local compactness theorem for Maxwell’s equations. Math. Meth.
Appl. Sci. 2, 12–25

Numerische Mathematik Electronic Edition

page 219 of Numer. Math. (1999) 82: 193–219

View publication stats