TICAM REPORT 00-27

October 17, 2000

Review of A Priori Error Estimation for Discontinuous Galerkin Methods

S. Prudhomme, F. Pascaly , J.T. Oden , and A. Romkes

 Texas Institute for Computational and Applied Mathematics,

The University of Texas at Austin,
Austin, TX 78712, USA

y Laboratoire de Mathématique, Bât. 425,

CNRS et Université de Paris-Sud,
Orsay, 91405, France
ii

Contents
1 Introduction 1

2 Notations and Preliminaries 1

2.1 Finite Element Partition . . . . . . . . . . . . . . . . . . . . . . . . . . 2
2.2 Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

3 Formulations for the Poisson Model Problem 4

3.1 Model Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
3.2 Weak Formulations and Finite Element Discretizations . . . . . . . . 7
3.2.1 Global Element Method - GEM . . . . . . . . . . . . . . . . . . 8
3.2.2 Symmetric Interior Penalty Galerkin Method - SIPG . . . . . . 8
3.2.3 Discontinuous hp Galerkin FE Method - DGM . . . . . . . . . 10
3.2.4 Non-Symmetric Interior Penalty Galerkin Method - NIPG . . 10
3.3 Equivalence of Strong and Weak Problems . . . . . . . . . . . . . . . 11
3.4 Properties of the Bilinear Forms . . . . . . . . . . . . . . . . . . . . . . 12
3.4.1 Mesh-dependent norms . . . . . . . . . . . . . . . . . . . . . . 12
3.4.2 Continuity of the bilinear forms . . . . . . . . . . . . . . . . . 13
3.4.3 Coercivity of the bilinear forms in the discrete spaces . . . . . 15

4 A Priori Error Estimates 17

4.1 SIPG and NIPG Methods . . . . . . . . . . . . . . . . . . . . . . . . . . 17
4.1.1 Proof of Theorem 4.1 for SIPG and NIPG . . . . . . . . . . . . 18
4.1.2 Alternative Proof of Theorem 4.1 for NIPG . . . . . . . . . . . 20
4.2 DG Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
4.2.1 A priori error estimate when c is nonzero . . . . . . . . . . . . 22
4.2.2 Discussion of the case in which c is zero . . . . . . . . . . . . . 24
4.2.3 New Interpolants . . . . . . . . . . . . . . . . . . . . . . . . . . 26
4.2.4 A priori error estimate when c is zero . . . . . . . . . . . . . . . 30
4.2.5 Alternative estimate when c is nonzero . . . . . . . . . . . . . 34

5 Concluding Remarks 35
5.1 Remarks on the Discontinuous Formulations . . . . . . . . . . . . . . 35
5.2 Future Challenges . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35

A Appendix 38
A.1 Discrete Schwarz Inequality . . . . . . . . . . . . . . . . . . . . . . . . 38
A.2 Multiplicative Trace Inequalities . . . . . . . . . . . . . . . . . . . . . 38
A.3 Poincaré–Friedrich’s Inequalities . . . . . . . . . . . . . . . . . . . . . 40
A.4 Inverse Property . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
A.5 Interpolation Error Estimates . . . . . . . . . . . . . . . . . . . . . . . 41
 1

1. Introduction

There has been renewed interest in Discontinuous Galerkin Methods (DGM) re-
cently, primarily due to the discovery that variants of these methods could be used
effectively to solve diffusion problems as well as problems of pure convection. One
such DGM was presented in the dissertation of Baumann [7] and reported in the
paper of Oden, Babuška, and Baumann [18]; summary of other versions of DGMs
and a lengthy historical review of this subject can be found in the record volume
edited by Cockburn, Karniadakis, and Shu [10]. The DGM possesses a number
of important properties that set them apart from traditional conforming Galerkin-
finite element methods: they are elementwise conservative, can support high order
local approximations that can vary nonuniformly over the mesh, are readily paral-
lelizable, and, for time-dependent problems, lead to block-diagonal mass matrices,
even for high-order polynomial approximations. These properties make DGMs
attractive candidates for a broad collection of applications.
Several papers have been published in the mathematical literature on a priori error
estimates for various DGMs for diffusion problems. In particular, an analysis of
one-dimensional versions of the Baumann-Oden method was reported by Babuška,
Oden, and Baumann [2]. Error estimates for several types of DGMs and for the
related Internal Penalty Galerkin Methods were presented in the dissertation of
Rivière [19] and in the paper of Rivière, Wheeler, and Girault [20]. Several other
studies on a priori error estimates for DGMs have appeared recently; see, for exam-
ple, the report of Chen [9] and the analysis of Süli, Schwab, and Houston [22,15].
Convergence analysis of other variants of DGM can be found in [10].
In the present work, we present a detailed derivation of a priori error estimates
for several hp-versions of DG-finite element methods for linear diffusion prob-
lems (the Poisson problem) on two-dimensional domains. In some cases, important
steps in our analysis follows the approach of Rivière, Wheeler, and Girault [20], but
other steps differ in detail. We present a series of approaches in which different
versions of DGMs, including those with penalty terms, can be analyzed. Our final
estimates differ in predicted rates of convergence with respect to the polynomial
degree p obtained in [20,19] and reflect rates consistent with the computed results
of Baumann [7].

2. Notations and Preliminaries

In the present report, we shall choose the domain Ω as a bounded open set in R 2 ,
with Lipschitz continuous boundary @ Ω. We will denote Γ D the part of the bound-
ary @ Ω on which Dirichlet conditions are prescribed and Γ N the part on which
Neumann conditions are prescribed. Formally, the boundary @ Ω is decomposed
[ = \
into the parts Γ D and Γ N such that Γ̄ D Γ̄ N @ Ω, and Γ D Γ N ?.=
2

2.1. Finite Element Partition

Let Ph denote a partition of the domain Ω, i.e. P h is a finite collection of Ne open
=
subdomains (elements) Ki , i 1; 2; : : : ; Ne , such that:
[
Ω = Ki ; and Ki \ K j = ?; i 6= j:
2
Ki Ph

The size and shape of an element K i , or simply K, of Ph are measured in terms of

two quantities, h K and  K , defined as:

hK = diam(K);
 K = sup fdiam(B ); B is a ball contained in Kg:
We also introduce the parameter h associated with the partition P h :

h = max
K 2P
hK : (2.1)
h

f g
Definition A family Ph of partitions Ph is said to be shape regular as h tends to zero if
there exists a number % > 0, independent of h and K such that:

hK
K
 %; 8K 2 Ph : (2.2)

In this report, all partitions Ph are assumed to be shape-regular.

In addition, we shall associate with each element K the element boundary @ K. The
unit normal vector outward from K (resp. K i ) is denoted by n (resp. n i ). j
Given a partition Ph , we shall denote the collection of edges of P h (points in one
dimension, faces in three dimensions) by the set E h l , l =f g =
1; : : : ; N . Edges
represent here open subsets of either Ω or @ Ω. We thus introduce the set Γ int of
interior edges as:
N
!
[
Γint = l n @Ω (2.3)
l 1 =
so that:
N
[
¯l = Γ̄D [ Γ̄N [ Γ̄int :
=
l 1

In the same way, we shall decompose E h into three subsets as:

E h = E h; D [ E h; N [ E h;int :
 3

1
0
0
1
0
1
0
1
γ
0
1 ij
0
1
0
1 N 4,i N 3,i
ΓD1
0
0
1 1
0
0
1
0
1
11
00
00
11
00
11
Ki Kj
n
1
0
0
1
0
1
0
1
1
0
0
1 11
00
00
11 ΓN
0
1
0
1 0
1 00
11
0
1 N 1,i N 2,i
0
1
n 1
0
0
1
0
1
0
1
0
1
0
10
1
0
1
0
1
0
1 Ki
0
1
0
1
0
1 11
00 11
00
00
11 00
11
n

Figure 1. Element interface ij and unit nor-

mal vector n.

Then, 2
E h; D if it lies on Γ D , and 2
E h; N if it lies on ΓN . Moreover, as shown
2
in Fig. 1, ij E h;int denotes an edge (interface) between two adjacent elements K i
and K j , where by convention i > j. For each edge , we also associate a unit normal
vector n. In the case is an edge associated with an element K i adjacent to @ Ω, i.e.
2 [ = j
E h; D E h; N , the unit normal vector is simply defined as n n i . For an interior
2
edge ij E h;int , with the convention i > j, n is chosen as the unit normal vector
outward from Ki , so that n n i = j= j
n j (see Fig. 1). In subsequent analyses, C will
denote generic positive constants, not necessarily the same in different places.

Remark 1 Using simple geometrical properties, one can show that each edge in a shape-
regular partition satisfies:
1
%
hK  K  j j  hK ; (2.4)

where jj denotes the length of . In other words, hK and are equal within a constant.
Therefore, we will interchangeably use hK or (preferably hK ) in this report.

2.2. Spaces
Let s be a positive integer. For any given open set S (S may define the whole
domain Ω, an element K of Ph , or an edge of E h ), the spaces H s (S) will denote the
usual Sobolev spaces with norm k
k
s ; S . In the particular case in which S represents
Ω, the norm will simply be denoted k
k 1
s . Moreover, H0 (S) is the set of functions in
H (S) which vanish on the boundary @ S of S, i.e.
1

H01 (S) = fv 2 H1 (S); v = 0 on @ Sg;

4

and H(div; S) denotes the space:

H(div ; S) = fv 2 (L2(S))2; r
v 2 L2(S)g:

The so-called (mesh-dependent) broken space H s (Ph ) will be defined as:

H s (Ph ) = fv 2 L2(Ω); vjK 2 Hs(K); 8K 2 Ph g:

The norm associated with the space H s (Ph ) is given as:
!1 = 2
k v ks ; P = ∑ k v k 2
s; K
2
h
K Ph

where vk ks;K is the Sobolev norm on K.

We will consider finite element spaces V hp of polynomial functions, possibly dis-
continuous at the element interfaces, such as:

V hp = fv 2 L2(Ω); vjK = v̂ Æ FK 1 ; v̂ 2 Pp( K)

b ; 8 K 2 Ph g (2.5)

where FK is the affine mapping from the master element K b to the element K in the
b b
partition, and Pp ( K) is the space of polynomial functions of degree at most p on K.
In hp methods, the polynomial degree can actually vary from one element to the
other. Denoting p K the polynomial degree associated with the element K, we define
the global value p for the partition Ph as:

p = Kmin p :
2P K
(2.6)
h

One advantage of DGMs over conventional hp finite element methods is that the
polynomial degrees p K do not necessarily match at the interfaces of the elements.

3. Formulations for the Poisson Model Problem

3.1. Model Problem

In this report, we shall consider the following Poisson model problem: find the
scalar function u which is the solution of

∆u + cu = f ; in Ω; (3.1)

and which satisfies the boundary conditions:

= u0 ;
u on Γ D ;
(3.2)
n
ru = g ; on Γ N :
 5

2
Here f L2 (Ω) represents the load scalar and c is a positive constant over the do-
main Ω.
We now proceed with the derivation of weak formulations of the Poisson equation
on discontinuous spaces. Let u, for the moment, be a sufficiently smooth function.
The regularity of u shall be discussed later in the report, namely in Subsection 3.3.
Multiplying (3.1) by a function v in H2 (Ph ) and integrating over the domain Ω, we
obtain:
Z Z

Ω
( r
ru + cu) v dx = Ω
f v dx:

Unlike the classical continuous finite element approach, we shall first decompose
the integrals in the above equation into element contributions
Z Z Z
∑ ( r
ru) v dx + ∑ cuv dx =∑ f v dx;
2
K Ph K 2
K Ph K 2
K Ph K

and then integrate by parts, so that:

Z Z Z
∑ r
rv + cuv) dx ∑
( u (n
ru) v ds = ∑ f v dx: (3.3)
2
K Ph K 2
K Ph @ K 2
K Ph K

We observe that the boundary integrals are defined on each element boundary;
those are now splitted according to the type of boundary such as:
Z Z
∑ (n
ru) v ds = ∑ (n
ru) v ds
2
K Ph @ K 2E h;D
Z
+ ∑ (n
ru) v ds
2E h;N
Z
+ ∑ (n
ru)i vi + (n
ru) j v j ds:
ij 2E h;int ij

where v i and v j denote the restrictions of v on the elements K i and K j respectively.

In the same way, (n u)i and (n
r
r
u) j represents the restrictions of the flux n u
r
on Ki and K j.
In general, except occasionally to avoid confusion, we shall simplify the notation
of these boundary integrals, by rewriting them, for instance,
Z Z
∑ (n
ru) v ds = (n
ru) v ds;
2E h;D ΓD
Z Z
∑ (n
ru) v ds = (n
ru) v ds:
2E h;N ΓN
6

Moreover, the treatment of the interior boundary integrals is as follows. Given an

2
edge ij E h;int shared by two adjacent elements K i and K j , i > j, we first note that:

(n
ru)i vi + (n
ru) j v j = n
(ru)i vi
r

n ( u) j v j ;

where n is now the unit normal vector with respect to the edge ij as defined in the
previous section. By analogy with the formula below where a; b; c and d are real
numbers:

ac bd = 12 (a + b)(c d) + 12 (a b)(c + d); (3.4)

we can write the integrand as:

r
n ( u)i vi
r
n ( u) j v j
 
 
= 12 n
(ru)i + n
(ru) j vi vj + 12
r
n ( u)i
r
n ( u) j vi + v j

= hn
rui [v] + [n
ru] hvi :

hi
Here [v] and v respectively denote the jump and average of v on an interior edge
ij , i > j, of any function v 2 
H s (Ki ) H s (K j ), s > 1=2, i.e.

[v] = vi v j ;
hvi = 12 (vi + v j ):
hi
We conveniently extend the definition of [v] and v , following Chen [9], to an edge
lying on Γ D as:

[v] = v;
hvi = v:
It allows us to combine the interior and Dirichlet boundary terms in only one inte-
gral as:
Z Z
∑ (n
ru)i vi + (n
ru) j v j ds + ∑ (n
ru) v ds
ij 2E h;int ij 2E h;D
Z
= [
Γint ΓD
hn
rui [v] + [n
ru] hvi ds:

Remark 2 Note that when u 2 H2(Ω), the fluxes [n
ru] are continuous almost every-
where in Ω, which yields
Z

Γint
[n
ru] hvi ds = 0; 8v 2 H2(Ph ): (3.5)
 7

Consequently, (3.3) can now be reduced, when u 2 H 2(Ω) and applying the Neu-
mann boundary condition, to:
Z Z Z Z
∑ r
rv + cuv) dx
( u
[
hn
rui [v] ds = ∑ f v dx + gv ds:
2
K Ph K Γint ΓD 2
K Ph K ΓN

We introduce the following bilinear form B( ; ) defined on H 2 (Ph )

 H2(Ph ) and the


linear form L( ) defined on H 2 (Ph ) such as:
Z
B(u; v) =∑ r
rv + cuv) dx;
( u (3.6)
K Ph2 K
Z Z
F(v) =∑ f v dx + gv ds: (3.7)
K Ph2 K ΓN

We also consider the bilinear form J( ; ) on H 2 (Ph )

 H2(Ph ), which incorporates all
boundary integrals on Γint and Γ D , as:
Z
J(u; v) = Γint ΓD [
hn
rui [v] ds: (3.8)

Then, a general discontinuous weak formulation of the Poisson equation reads:

B(u; v) J(u; v) = F(v); 8v 2 H2(Ph ): (3.9)

This above variational form constitutes the starting point to derive formulations of
various Discontinuous Galerkin Finite Element Methods (concisely, DGMs.)

3.2. Weak Formulations and Finite Element Discretizations

All the formulations presented below use the observation that, for u 2 H 1(Ω) \
H 2 (Ph ), the jump [u] vanishes on each ij :
Z
v [u] ds = 0; 8v 2 L2( ij): (3.10)
ij

It follows that:
Z

Γint
hn
rvi [u] ds = 0; 8v 2 H2(Ph ): (3.11)

Moreover, the Dirichlet boundary condition can be applied in the following weak
manner:
Z Z

ΓD
(n
rv) u ds = ΓD
(n
rv) u0 ds; 8v 2 H2 (Ph ): (3.12)
8

Therefore, introducing the linear form J0 ( ) defined as:

Z
J0 (v) = ΓD
(n
rv) u0 ds; 8v 2 H2 (Ph ); (3.13)

we observe that, for u 2 H 1(Ω) \ H2 (Ph ) and u = u0 on Γ D,

J(v; u) = J0 (v); 8v 2 H2(Ph ): (3.14)

3.2.1. Global Element Method - GEM

Introducing the bilinear form B ( ; ), the subscript referring to the fact that we
substract the term J(v; u) to the left hand side of (3.9), and the linear form F ( )

B (u; v) = B(u; v) J(u; v) J(v; u);
(3.15)
F (v) = F(v) J0 (v);

the Global Element Method consists in finding u such that:

B (u; v) = F (v); 8v 2 H2(Ph ): (3.16)

One advantage of this method is that it defines a symmetric problem. On the other
hand, a significant disadvantage is that the bilinear form is not guaranteed to be
semi-positive definite. When dealing with time-dependent problems, this could
imply that some eigenvalues have negative real parts, causing the formulation to
be unconditionally unstable.
The corresponding finite element discretization of the above problem consists in
2
finding u h V hp such that:

B (uh ; v) = F (v); 8v 2 V hp : (3.17)

This method was introduced by Delves et al. [11–14] with the particular objective
of accelerating convergence of iterative schemes.

3.2.2. Symmetric Interior Penalty Galerkin Method - SIPG

To enforce stability of the discontinuous method, i.e. continuity of the solution at
the interface of the elements, penalty terms have been added to the formulation by
Arnold [1] and Wheeler [23]. Let us introduce the following penalty terms:
Z Z Z
J (u; v)

= ∑  [u] [v] ds + ∑  uv ds =
[
 [u] [v] ds;
ij 2E h;int ij 2E h;D Γint ΓD

and
Z Z

J0 (v) = ∑  u0 v ds =  u0 v ds;
2E h;D ΓD
 9

where  represents the penalty parameter which depends on the length of the
edges ij and and the polynomial degree used in the elements; namely  =
 (h; p). Then the SIPG method is similar to the GEM except for the penalty terms.
Indeed, introducing the forms:

B  (u; v) = B(u; v) J(u; v) J(v; u) + J (u; v);

(3.18)
F  (v) = F(v) J0 (v) + J0 (v);
the Symmetric Interior Penalty Galerkin problem is to find u such that:

B  (u; v) = F  (v); 8v 2 H2(Ph ): (3.19)

Note that when  takes on the value zero, we naturally retrieve the GE method.
The finite element analogue of problem (3.19) is to find u h 2 V hp such that:
B  (uh ; v) = F  (v); 8v 2 V hp : (3.20)

Remark 3 Following Baker and Karakashian [5,6,16], we consider a variant of the SIPG
method. Instead of using the formula (3.4), one may use:

ac bd = ac ad + ad bd = a(c d) + (a b)d (3.21)

so that, by analogy:

r
n ( u)i vi
r
n ( u) j v j = n
(ru)i [v] + [n
ru] v j
and, since the fluxes, for u 2 H2 (Ω), are continuous across the interelement boundaries, we
have:
Z Z
(n
ru)i vi + (n
ru) j v j ds =
r
n ( u)i [v] ds:
ij ij

The new bilinear form for the boundary terms is now defined as:
Z
I(u; v) = [
Γint ΓD

r
n ( u)i [v] ds

so that the new formulation reads: Find u 2 H1(Ω) \ H2(Ph ) such that, for all v 2 H2 (Ph ),
B(u; v) I(u; v) I(v; u) + J  (u; v) = F(v) J0 (v) + J0 (v): (3.22)

We now see that we recover the SIPG method from the Baker-Karakashian formulation by
replacing the term n ( u)i by n
r h
r i
u . It follows that all the properties associated with
the SIPG method will also apply to the Baker-Karakashian formulation.
10

3.2.3. Discontinuous hp Galerkin FE Method - DGM

The discontinuous Galerkin method by Baumann et al. [7,18] differs from the Global
Element Method by just a sign. Indeed, by introducing the forms:

B+ (u; v) = B(u; v) J(u; v) + J(v; u);

(3.23)
F+ (v) = F(v) + J0 (v);

the DG formulation reads: Find u such that

B+ (u; v) = F+ (v); 8v 2 H2(Ph ): (3.24)

It is straightforward to show that the bilinear form is positive semidefinite.

The associated finite element version of the DG method consists then in finding
2
uh V hp such that

B+ (uh ; v) = F+ (v); 8v 2 V hp : (3.25)

3.2.4. Non-Symmetric Interior Penalty Galerkin Method - NIPG

This method was introduced by Rivière [19] and Süli, Schwab and Houston [22,15]
and is inspired from the DG method with the addition of penalty terms. The new
bilinear and linear forms read:

B+ (u; v) = B(u; v) J(u; v) + J(v; u) + J (u; v);

(3.26)
F+ (v) = F(v) + J0 (v) + J0 (v);

so that the problem to solve by the NIPG method becomes: Find u such that

B+ (u; v) = F+ (v); 8v 2 H2(Ph ): (3.27)

Once again, we may consider DG as a special case of NIPG with  = 0.

The finite element problem corresponding to the NIPG formulation (3.27) is to find
2
uh V hp such that

B+ (uh ; v) = F+ (v); 8v 2 V hp : (3.28)

The four methods presented thus far are all very similar, except for a plus or minus
sign in front of the term J(v; u) and the addition of a penalty term J  (u; v) or not. We
shall see in the remainder of this report how these changes modify the properties
of the respective formulations.
 11

3.3. Equivalence of Strong and Weak Problems

We shall show the equivalence of the strong and weak formulations only with re-
spect to the Global Element method. The results are identical for the other formu-
lations, namely the SIPG, DG and NIPG methods. Existence of solutions of the
discontinuous formulations is then somewhat guaranteed. However, we empha-
size here that Theorem 3.1 does not infer anything about the uniqueness of the
solutions. This question still remains an open issue.

2
Theorem 3.1 (GE Method) Let u C2 (Ω) be the solution of Problem (3.1)-(3.2). Then
u satisfies the weak formulation (3.16). Conversely, if u H1 (Ω) H2 (Ph ) is a solu- 2 \
tion of (3.16) then u satisfies the partial differential equation (3.1) and boundary condi-
tions (3.2).

Proof: The first part of the theorem has been proved along with the derivation of
the Global Element formulation, since (3.9) is satisfied when u C 2 (Ω). 2
The converse follows the proof given in Rivière [19]. Let D (K) H 2 (K) be the 
space of infinitely differentiable functions with compact support on element K and
2
let v D (K). Then (3.16) gives:
Z Z

K
r
rv + cuv) dx =
( u
K
f v dx

which implies, after integration by parts and since v is arbitrary in D (K), that

∆u + cu = f ; a.e. in K : (3.29)

Next, we consider an interior edge ij shared by the elements K i and K j . Let v be

[  
a function in H02 (Ki K j ) H 2 (Ki ) H 2 (K j ), extended by zero outside. Then the
boundary terms J(u; v) and J(v; u) vanish, because [u] [v] 0 on ij , and the weak = =
formulation (3.16) reduces to
Z Z

[
Ki K j
r
rv + cuv) dx =
( u
[
Ki K j
f v dx (3.30)

On the other hand, multiplying (3.29) by v, integrating on Ki and K j and using

Green’s formula, we have:
Z Z Z

Ki
( u r
rv + cuv) dx (n
ru)i v ds = Ki
f v dx;
ij
Z Z Z

Kj
( u r
rv + cuv) dx (n
ru) j v ds = Kj
f v dx;
ij

so that
Z Z Z

[
Ki K j
r
rv + cuv) dx
( u [n
ru] v ds = [
Ki K j
f v dx: (3.31)
ij
12

Comparing (3.30) and (3.31), one observes that:

Z
[n
ru] v ds = 0; 8v 2 H02(Ki [ K j):
ij

Then, [n
r =
u] 0 for all element edges ij , which implies u H(div; Ω). This r 2
allows us to conclude that u satisfies Poisson Equation globally on Ω, i.e.

∆u + cu = f ; a.e. in Ω: (3.32)

To recover the Dirichlet boundary conditions, we now consider a function v 2

\
H01 (Ω) H2 (Ω), so that integrating (3.32) provides:
Z Z

Ω
r
rv + cuv) dx =
( u
Ω
f v dx;

whereas (3.16) yields:

Z Z Z Z

Ω
r
rv + cuv) dx
( u
ΓD
(n
rv) u ds = Ω
f v dx
ΓD
(n
rv) u0 ds:

Substracting both equations, we obtain:

Z

ΓD
(n
rv) (u u0 ) ds = 0; 8v 2 H01(Ω) \ H2(Ω);
and conclude that u = u0 on Γ D.
In the same way, choosing v 2 H 2 (Ω)  H2 (Ph ) such that v = 0 on Γ D , we get:
Z

ΓN
(n
ru g) v ds = 0;
so that n
ru = g on Γ N . 
Remark 4 When c is zero, C2 (Ω) can be replaced in Theorem 3.1 by H1 (Ω) \ H2 (Ph ) since
r 2
u H(div; Ω).

3.4. Properties of the Bilinear Forms

3.4.1. Mesh-dependent norms
We now introduce norms associated with the bilinear forms:

1. Energy Norm:
 
kvk2e;P = B(v; v) = ∑ kvk2e;K = ∑ krvk20;K + ckvk20;K (3.33)
2 2
h
K Ph K Ph
 13

2. Norm proposed by Süli et al. in [22,15]:

Z
kvkP = B(v; v) + J (v; v) = kvk +
2
h
 2
e ;Ph
[
Γint ΓD
 [v]2 ds (3.34)

3. Norm proposed by Baumann et [Link] [7,17,18] and by Baker and Karakashian

in [6]:
Z
jjvjjP = kvkP +
2
h
2
h
[
Γint ΓD 
1
hn
rvi2 ds (3.35)

We note that the energy norm becomes a seminorm when c is zero.

3.4.2. Continuity of the bilinear forms

We shall show now that the bilinear forms B ( ; ) and B ( ; ) are continuous on




H 2 (Ph ) with respect to the norm jj
jj
Ph defined in (3.35). Unfortunately, we are
unable to show continuity with respect to the other two norms (3.33) and (3.34).

Theorem 3.2 (GEM and DGM) Let B ( ; ) be the bilinear form defined either in (3.15)
or in (3.23) . Then,
jB(u; v)j  jjujjP jjvjjP ; h h
8u; 8v 2 H2(Ph ): (3.36)

Proof:
First note that:
jB(u; v)j = j B(u; v) J(u; v)  J(v; u)j
 j B(u; v)j + j J(u; v)j + j J(v; u)j
It is clear that
Z
j B(u; v)j  ∑ jru
rv + cuvj dx  kuke;P kvke;P
2
h h
K Ph K

The first boundary term gives:

Z
j J(u; v)j  [
Γint ΓD
jhn
rui [v]j ds
rZ rZ
 [
Γint ΓD
 1 hn
rui 2
ds
[
Γint ΓD
 [v]2 ds:

Likewise,
Z
j J(v; u)j  [
Γint ΓD
jhn
rvi [u]j ds
rZ rZ
 [
Γint ΓD
 2
[u] ds
[
Γint ΓD
 1 hn
rvi2 ds:
14

In consequence, we have, using the discrete Schwarz inequality (A.1):

jB(u; v)j  kuke;P kvke;P h h
rZ rZ
+ Γint ΓD [
 1 hn
rui 2
ds
[
Γint ΓD
 [v]2 ds
rZ rZ
+ Γint ΓD [
 2
[u] ds
[
Γint ΓD
 1 hn
rvi2 ds:
r Z Z
 kuk2e;P + h
[
Γint ΓD
 [u]2 ds + [
Γint ΓD
 1 hn
rui2 ds
r Z Z
 kvk2e;P + h
[
Γint ΓD
 [v]2 ds + [
Γint ΓD
 1 hn
rvi2 ds
 jjujjP jjvjjP ;
h h

which completes the proof. 

Theorem 3.3 (SIPG and NIPG Methods) Let B ( ; ) be the bilinear form defined ei-


ther in (3.18) or in (3.26). Then,
jB (u; v)j  C jjujjP jjvjjP ; h h
8u; 8v 2 H2(Ph ): (3.37)
where C is a constant, C  2.

Proof:
As before we have:
jB (u; v)j = j B(u; v) J(u; v)  J(v; u) + J (u; v)j
 j B(u; v)j + j J(u; v)j + j J(v; u)j + j J  (u; v)j
 jjujjP jjvjjP + j J  (u; v)j :
h h

And
Z rZ rZ
j j
J  (u; v)
[
Γint ΓD
j [u] [v]j ds  [
Γint ΓD
 [u]2 ds
[
Γint ΓD
 [v]2 ds:
Therefore, making use again of the discrete Schwarz inequality (A.1), we obtain:
rZ rZ
jB(u; v)j  jjujjP jjvjjP +

h h
Γint ΓD [
 [u] 2
ds
[
Γint ΓD
 [v]2 ds:
r Z r Z
 jjujj2P + h
[
Γint ΓD
 [u]2 ds jjvjj2P + h
[
Γint ΓD
 [v]2 ds
q q
 2 jjujjP 2 jjvjj2P 2
h h

 2 jjujjP jjvjjP ;
h h

and we see that C is at most equal to 2. 

 15

3.4.3. Coercivity of the bilinear forms in the discrete spaces

Here we wish to show that the bilinear forms B  ( ; ) and B ( ; ) are coercive in


H 2 (Ph ) with respect to the norm jj
jj
Ph in order to be able to apply classical theorems
for existence and uniqueness of solutions of the discontinuous methods. Unfortu-
nately, to date, we are able to prove coercivity only in the discrete discontinuous
spaces V hp , and then, only for the SIPG and NIPG formulations .

Theorem 3.4 (NIPG Method) Let  =  p2 =h,  being a positive number. Then, for all
 > 0, there exists a positive constant, > 0, such that:
B+ (z; z)  jjzjj2P ; h
8z 2 V hp : (3.38)

Here is independent of h and p.

Proof: Let be an arbitrary real number and choose a z 2 V hp. Then

B+ (z; z) jjzjj2P
h
Z
= (1 ) B(z; z) + (1 ) J (z; z)

[
Γint ΓD
1

hn
r zi2 ds

Since nh
r i z is the average of the flux at the interface of two elements K i and
K j , the corresponding integral can be split into two integrals with integrands (n

r z)i = and (n
r
z) j = , each one associated with the elements K i or K j respectively.
Therefore, let  [
Γint Γ D and consider the integral associated with the element
K. Using the trace inequality (A.3) and the inverse property (A.7), we have
Z
1


r
(nz)2 ds  kr zk20;
1

 
  h krzk0;K + krzk0;K kr zk0;K
C 1 2 2
K
 2 
 C 1
hK
+ C0 hpK krzk20;K
K

p2K
 C h krzk20;K ;
K

so that, selecting  to be equal to  p 2K =h K , we obtain:

Z

Γ 
1

r z)2 ds 
(n
C

krzk20;K :
Note that, when the mesh size h Ki and h K j and the polynomial degrees p Ki and p K j
are different from each other in the two elements K i and K j sharing the edge ij , we
16

actually choose  as

max(p2Ki ; p2K j )
= ;
min(hKi ; h K j )

so that:
Z p2 i
1


r
(nz)2i ds 
C K
 hK
krzk20;K i
ij i

C min(hKi ; h K j ) p Ki
2
  max(p2Ki ; p2K j ) hKi
krzk20;K i

 C krzk20;K : i

It then follows that:

B+ (z; z) jjzjj2P  (1

h
C=) B(z; z) + (1 ) J (z; z):

Therefore, we certainly can pick a value of such that

0<  1 +1C=



for which the bilinear form B+ ( ; ) is coercive in V hp , for all  > 0. 
Theorem 3.5 (SIPG Method) Let  =  p2 =h,  being a positive number. Then, for
 > 0 , there exists a positive constant independent of h and p, > 0, such that:
B  (z; z)  jjzjj2P ;
h
8z 2 V hp : (3.39)

Proof: Let be an arbitrary real number and choose z 2 V hp . Then

B  (z; z) jjzjj2P = (1
h
) B(z; z) + (1 ) J (z; z)
Z Z
2
[
Γint ΓD
hn
rzi [z] ds 1
Γint [ΓD 
h n
rzi2 ds
There exists a positive number " such that for every edge 2 Γint [ ΓD:
Z rZ rZ
2 hn
rzi [z] ds  2  1 hn
rzi 2
ds  [z]2 ds
Z Z
 "  hn
rzi 1 2
ds + "  [z]2 ds
1
 17

which yields, using the result in the previous proof:

   
B (z; z)

jjzjjP 
2
h
1 ( + ") 
C
B(z; z) + 1
1
"
J  (z; z):

In order to prove coercivity, we want to find > 0 such that both factors in the
inequality are positive, in other words:
   
1 ( + ")  > 0
C
and 1
1
"
> 0:
The second inequality requires that:

0<  1 1"
which means that
" > 1:
On the other hand the first inequality requires that:

0<
= 1 C=  C
 11 +"CC=  1 + C=   + C
This completes the proof by taking  sufficiently large, namely    0 (where for
instance 0 > C.) 



Remark 5 We note that B+ ( ; ) (for NIPG Method) is coercive in H2 (Ph ) with respect to
the norm k
k 2
Ph . Indeed, for all v H (Ph ), 2
B+ (v; v) = B(v; v) + J(v; v) + J (v; v) = kvk2P :
J(v; v) h
(3.40)

It is also straightforward to show that B+ (
;
) (for DGM) is coercive in H2 (Ph ) with respect
to the energy norm k
ke;P : h

B+ (v; v) = B(v; v) J(v; v) + J(v; v) = B(v; v) = kvk2e;P : (3.41)

h

These results will be crucial in deriving a priori error estimates in the next section.

4. A Priori Error Estimates

4.1. SIPG and NIPG Methods

2 \ 
Theorem 4.1 Let u H 1 (Ω) H s (Ph ), s 2, be a solution of (3.18) (SIPG) or (3.26)
(NIPG) and uh be the discrete discontinuous solution of
B (uh ; v) = F (v); 8v 2 V hp : (4.1)
18

=
Then, choosing   p2 =h, ( > 0 for NIPG and   0 for SIPG), the numerical error
=
e u uh satisfies:

 1
keke;P  C phs
h 3=2
kuk s (4.2)

where  = min(p + 1; s) and p  1.

4.1.1. Proof of Theorem 4.1 for SIPG and NIPG
First, by definition of the norms, we note that e e;Ph kk  j j jj
e Ph . In other words, it
suffices here to estimate the error with respect to the norm Ph . The proof is jj
jj
inspired by [5,6,16] where the authors have derived the rate of convergence in h
only for the SIPG method of the (3.22) form. Here we extend their results to the
NIPG formulation as well and also show for both methods the rate of convergence
in p.
Proof: Let z p be an interpolant of u in V hp . We shall use the notation  u z p =
= = =
and  uh z p so that e u u h   . Applying the triangle inequality, we
have:

j j e j jP = j j u
h
jj = jj  jjP  jjjjP + jj jjP :
u h Ph h h h

From the coercivity of the bilinear form B ( ; ), since 

2 V hp, we have
jj jj2P  CB (;  );
h

and from the “orthogonality” property B  (u uh ; v) = 0, 8v 2 V hp, we get

B (;  ) = B (;  ); 8v 2 V hp :



Using the continuity of B ( ; ), we know that

B (;  )  C jj jjP jj jjP ;

h h

which implies

jj jjP  C jjjjP :

h h

Finally, we have

jjejjP  jjjjP + jj jjP  C jjjjP :

h h h h

We recall here that C is a generic constant independent of h and p which takes

different values at different places.
 19

We now choose the interpolant z p as defined in Lemma A.7. Then:

Z   Z Z
jjjj2P = ∑ jrj2 + c2 dx + [
1

hn
r i2 ds + [
 [ ]2 ds
h
2
K Ph K Γint ΓD Γint ΓD

(4.3)

The integrals in the leading term are estimated as, using (A.8):
Z
!2
hK 1

K
jrj 2
dx C psK 1
kuk2s;K ; s  1;
Z   2

K
c dx2
 cC hK
psK
kuk2s;K ; s  0;
so that
Z   2 2

K
jrj2 + c2 dx  C hpK2s 2 kuk2s;K ; s  1:
K

Let ij denote an interior edge shared by the elements K i and K j . Then, using the
+  +
inequality (a b)2 2a2 2b2 , we observe that
Z Z Z
1 2
h n
r i ds  (n
(r )i ) ds + n
(r ) j ds:
1 2 1 1 2 1
ij  2  ij2  ij

In other words, in splitting the second integrals on the right hand side of (4.3) as
above, we actually associate with each E h;int E h; D an element K, such that 2 [
Z
1


r
(n )2 ds  1 kr k20;

 
  h kr k0;K + kr k0;K kr k0;K
C 1 2 2
K
 
 C 1
hK
kk + kk1;K kk2;K
2
1; K
!
2 2
hK 1
hK 2
 C 1 hK
h K p2s 2
+ ps 1
psK 2
kuk2s;K
K K
!
h2K 3
h2K 3
 C
p2s 2
+ p2s 3
kuk2s;K
K K
2 3
 C hpK2s 3 kuk2s;K
K

h2K 2
 C p2s 1
kuk2s;K ; s  2:
K
20

Again, for an interior edge ij shared by Ki and K j , using (a b)2  2a2 + 2b2 , we
have:
Z Z

2 Z Z
 [ ]2 ds =  i  j ds  2  (i )2 ds + 2   j
2 ds
ij ij ij ij

This means that the edge integrals making the third term of (4.3) are bounded by:
Z
h2K 1 2 2
 ( )2 ds  C kuk2s;K  C hpK2s 3 kuk2s;K
p2s
K
1
K

In combining the above results, we thus obtain

( )1=2
h2K 2
h2K 2
h2K 2
jjejjP  C jjjjP  C ∑ p2s 2
+ p2s 1 + p2s 3
kuk s ; K
h h
K Ph2 K K K

hK 1
C ∑ 3=2
kuk s ; K
K Ph2 psK
 1
 C phs 3=2
kuk s
which is the expected a priori error estimate. 
4.1.2. Alternative Proof of Theorem 4.1 for NIPG
Alternatively, we present a second proof of Theorem 4.1 for the NIPG method only
as it is based on the nonsymmetry of the formulation. The proof is inspired by
the one found in [22]. However, our rate of convergence with respect to p was
improved from (s 2) to (s 3=2) using the interpolation estimates of Lemma A.7.
Later, the same authors proposed in [15] a comparable version of the proof with
(s 3=2) as the rate of convergence.
Proof: Once again, z p is the interpolant of u in V hp as defined in Lemma A.7. and
= =
we denote  u z p and  uh z p as before. Then,

keke;P  kekP = ku uh kP = k  kP  kkP + k kP :

h h h h h h

Moreover, from the definition of B+ (
;
) and the norm k
ke;P (see (3.40)) and the h
“orthogonality” relation, we have:

k k2P = B+ (;  ) = B+ (;  ):

h

The goal is now to bound B+ (;  ) in terms of k kP . Recall that:

h

B+ (;  ) = B(;  ) + J (;  )

 
+ J(; )
J(;  )
 j B(;  )j + j J  (;  )j + j J(;  )j + j J(; )j
 21

The first term on the right hand side of the equation above gives:
Z
j B(;  )j  ∑ jr
r + c j dx  kke;P k ke;P  kkP k kP :
2
h h h h
K Ph K

j j
The term J  (;  ) is bounded by:
Z
j J  (;  )j  [
Γint ΓD
j [] [ ]j ds
rZ rZ
 Γ [Γ  [] 2
ds
Γint ΓD [
 [ ]2 ds
int D

 kkP k kP ; h h

whereas we have for the third term:

Z
j J(;  )j  Γint ΓD [
jhn
r i [ ]j ds
rZ rZ
 Γint ΓD [
 1 hn
r  i 2
ds
[
Γint ΓD
 [ ]2 ds
rZ
 k kP h
[
Γint ΓD
 1 hn
r i2 ds:
Likewise, J(;  ) is bounded by:
rZ
j J(; )j  kkP h
[
Γint ΓD
 1 hn
r i2 ds
Using again the trace inequality (A.3) and the inverse property (A.7), it is shown
that:
Z 2


1


(n r  )2 ds  C K kr k20; K
p
 hK
In other words, using  =  p 2K =h K

j J(; )j  CkkP k kP h h

Combining the above results, we have:

s !
Z
k kP  C kkP +
h h
[
Γint ΓD h
1
n
r i2 ds  C jjjjP h

so that:

keke;P  kekP  kkP + k kP  kkP + C jjjjP  C jjjjP :

h h h h h h h

We conclude the proof by employing the estimate on jj jjP shown in the previous
proof.  h
22

4.2. DG Method
We recall that the DG formulation proposed in [7,18] is deduced from the NIPG
method by simply setting the penalty parameter  to zero. However, unlike NIPG,



continuity and coercivity of the bilinear form B+ ( ; ) cannot be proved simultane-
ously using the same norm. At best it is shown that:

B(v; v) = kvk2e;P ; h
8v 2 H2(Ph );
and that:

B+ (u; v)  jjujjPh jjvjjPh ; 8u; v 2 H2(Ph ):

The main issue in finding a priori error estimates for the error e u u h in the nu- =
merical approximation uh of the DG problem consists in deriving an upper bound
on:
rZ
[ ]2 ds
[
Γint ΓD

kk
with respect to the norm  e;Ph when c = 0. This integral does indeed appear when
bounding the term J(;  ), i.e.
Z rZ rZ
j J(;  )j  [
Γint ΓD
jhn
r i [ ]j ds  [
Γint ΓD
hn
r  i 2
ds
[
Γint ΓD
[ ]2 ds:

We present below two approaches, by treating separately the case when c is zero
and the case when c is nonzero.

4.2.1. A priori error estimate when c is nonzero

We find it instructive to analyze the special case in which c is strictly greater than
zero. In this case, we still can use the methodology presented earlier for the NIPG
method. However, we shall see that the rate of convergence with respect to the
mesh size becomes suboptimal as stated in the following theorem.

2 \ 
Theorem 4.2 Let u H 1 (Ω) H s (Ph ), s 2, be a solution of (3.23) with c > 0 and uh be
the discrete discontinuous solution of (3.24). Then, the numerical error e u uh satisfies: =
 2
keke;P  C phs
h 3=2
kuk s (4.4)

where  = min(p + 1; s) and p  1.

Proof: Using the same procedure and notation as before, we have:

k e ke ; P = k u
h
uh ke ;P = k  ke ;P  k ke ;P + k ke ;P :
h h h h
 23

Moreover, from the definition of B+ ( ; ) (see (3.41)) and the “orthogonality” rela-
tion, we further show that:

k k2e;P = B+(;  )
h

= B+(;  )
= B(;  ) J(;  ) + J(; )
 j B(;  )j + j J(;  )j + j J(; )j
We now consider each term one at a time. The first term B(;  ) is straightforwardly
bounded by:
 1
j B(;  )j  kke;P k ke;P  C hps 1 kuksk ke;P
h h h
(4.5)

We expect that the third term J(;  ) can be treated as before and should not pose
any problems. Indeed, applying the Cauchy-Schwartz inequality, we have:
rZ rZ
j J(; )j  [
Γint ΓD
hn
r i 2
ds
[
Γint ΓD
[ ]2 ds

When  Γint [ ΓD and  2 V hp (K), we have already shown that:

Z 2
(n
r )2 ds  C hpK kr k20;K :
K

Next, we obtain from the approximation property (A.9)

Z 2 1
 2 ds = k k20;  C hpK2s 1 kuk2s;K :
K

Therefore the term J(;  ) is bounded by:

 1
j J(; )j  C phs 3=2
kuk s k ke ;P :
h

Finally we need to consider the term J(;  ), which is held responsible for deterio-
rating the convergence rate of the solution. By the Cauchy-Schwarz inequality, we
have:
rZ rZ
j J(;  )j  [
Γint ΓD
hn
r i 2
ds
[
Γint ΓD
[ ]2 ds

Once again, the approximation property gives

Z 2 3
(n
r)2 ds  C hpK2s 3 kuk2s;K
K
24

while from the trace inequality (A.3), we have:

 
k k  C
2
0;
1
hK
k k + k k0;K kr k0;K
2
0; K
 
C 1
hK
k k + k k + hK kr k
2
0; K
1
hK
2
0; K
2
0; K
 
C 1
hK
k k + hK kr k
2
0; K
2
0; K (4.6)
 
C 1
ch K
k k2e;K + hK k k2e;K
 chC k k2e;K :
K

It is important to point out here that the norm k k 0; is bounded as long as c > 0.
Then we have:
 2
j J(;  )j  C phs 3=2
kuk s k ke ;P :
h

In conclusion,
  
1 h 1 h 2  2
k ke ;P  C
h
h
ps 1
+ p s 3=2
+ p s 3=2
kuks  C phs 3=2
kuk s
which completes the proof. 

Remark 6 Note that C is inversely proportional to c. Therefore the error is expected to

grow as c gets smaller.

4.2.2. Discussion of the case in which c is zero

The operator, when c is zero, reduces to the pure Laplacian. In this case, the energy
norm k
k
e ;Ph becomes the seminorm kr
k
0;Ph . Following the same procedure as
before, we would have:

kr k20;P = B+(;  ) = B+(;  )

h
(4.7)

= =
where  u z p ,  u h z p and z p defines an arbitrary interpolant of u on V hp .
However, from (4.6), we can see right now that the term B + (;  ) would then be
kk
bounded by  0;Ph . In turn, it is impossible to bound  0;Ph with respect to kk
kr k
 0;Ph . Therefore, the previous methodology to obtain error estimates cannot
be applied in the present case.
 25

Suppose that we introduce an elementwise constant function ¯ to be defined later.

Then, we can rewrite (4.7) as:

kr k20;P = B+(;  ) = B+(;  ¯ + ¯ ) = B+(;  ¯ ) + B+(; ¯ ):

h
(4.8)

Suppose now we can construct a new interpolant such that:

B+ (; ¯ ) = 0: (4.9)

Then we would have

kr k20;P = B+(;  ) = B+(;  ¯ )

h

= B(;  ¯ ) J(;  ¯ ) + J( ;¯ ) (4.10)

= B(;  ) J(;  ¯ ) + J(; )
We have seen that the terms B(;  ) and J(;  ) are easily bounded in terms of
kr k0;P . The other term reads:
h

Z  
J(;  ¯ ) = hn
ri  ¯ ds:
[
Γint ΓD

According to Lemma A.5, this integral can be bounded with respect to kr k 0;Ph
under the condition that ¯ is chosen as the average of  on each element.
This approach has been followed in principle by Rivière, Wheeler and Girault
in [20,19] where they construct special interpolants  u which satisfied (4.9) and

ku  uk0;K  C phs K 2 kuks;K ;
K
 1
kr(u  u)k0;K  C hpKs 2 kuks;K ;
K

hK 2
kr (u  u)k0;K  C ps 2 kuks;K ;
2

= +  
where  min(p K 1; s), s 2, pK 2. Using these interpolants, they were able
to derive an a priori error estimate of the form:
 1
krek0;P  C hps 4 kuks :
h
(4.11)

Although the rate of convergence is optimal in h, we show next that the rate of
convergence in p is in reality better than (s 4). We improve this result by con-
structing better approximation properties for the new interpolant and by refining
the analysis.
26

4.2.3. New Interpolants

Lemma 4.1 Let K be a triangular element of the partition Ph and u a function in H s (K),

s 2. There exists a positive constant C depending on s and % but independent of u, pK ,
2
and h K , and a polynomial  u Pp K (K), p K 2, such that 
Z
n
r(u  u) ds = 0; 8  @ K; (4.12)

and
hK
k u  uk 0 ; K  C s 3=2
kuk s ; K ;
pK
 1
kr(u  u)k0;K  C hsK 3=2 kuks;K ; (4.13)
pK
 2
kr2 (u  u)k0;K  C hpKs 2 kuks;K ;
K

where  = min(pK + 1; s).

We present the proof of this theorem for triangular elements only. The proof is
similar for quadrilaterals.
2
Proof: Let the triangle K Ω be the image of the master element K̂ by the affine
mapping FK as shown in Figure 2. The mapping FK is often rewritten as:

FK (x̂) = Bx̂ + b (4.14)

where B represents a two-by-two matrix whose components are independent of x̂

and b is a two-dimensional vector. Here, will refer to the edge between node
N2 and N3 , unless stated otherwise, and ˆ on K̂ will denote its image by FK 1 . We
associate with ˆ and the unit normal vector n̂ and n, respectively.
2 =
Given  H2 (K), namely  u z p , where z p is the interpolant of u as defined
in Lemma A.7, the objective here is to construct a polynomial function q in V hp (K)
such that:
Z Z
n
r ds = n
rq ds: (4.15)

Indeed we would have:

Z Z Z
n
r(  q) ds = n
r(u zp q) ds = n
r(u (z p + q)) ds = 0;

and the new interpolant could be derived as  u = z p + q.

 27

N3
FK
1.0
n

n y γ
K
γ N2
K N1
x

0.0 1.0

Figure 2. Reference element K̂ and mapping FK from K̂ to the ele-

ment K in the physical domain.

Following [19], and assuming p K  2, we introduce the polynomial function q̂

associated with the edge ˆ on K̂:

q̂ = C (1 x̂ ŷ)(x̂ + ŷ); 8 x̂ = (x̂; ŷ) 2 K̂: (4.16)

where C is a constant to be defined. We observe in Fig. 3 that such a polynomial

function satisfies:

ˆ 12
n̂
r̂q̂ ds =0
Z

ˆ 23 ˆ =
n̂
r̂q̂ ds = 2C
Z

ˆ 31
n̂
r̂q̂ ds =0

with ˆ i j defining an edge on K̂ joining the nodes N i and N j .

The constant C is found so that (4.15) is satisfied in the physical space. Further-
28

qγ

N3

γ
K
N1 N2

Figure 3. Polynomial function q̂ on the reference element K̂.

more, we obtain an upper bound for C (see Rivière [19]) as:

jC j  Ck Bk2 k B 1k2 kr̂ˆ k0; ˆ

!2 !2
C hK
ˆ
ĥ
K
1=2
hK krk0;
!2
 C 2 ĥ
ˆ
1=2
hK krk0;
 C h1K=2 krk0;
n o1=2
 C kr k20;K + hK krk0;K kr2 k0;K
where we make use of the Trace Inequality (A.3). We also observe that:

kq k0;K  C jdetBj1=2 kqˆ k0;K̂

 C hK jC j K̂
 C h K jC j :
Likewise, we have:

krq k0;K  C jC j
kr2 q k0;K  C hK 1 jC j
So far, we have carried out the analysis for the edge between node N 2 and N3 .
We point out that the same results are obtained for the other two edges. We then
associate with each edge 12 , 23 , 31 , a polynomial q12 , q23 , q31 respectively such
 29

that

q̂12 = C12 ŷ(1 ŷ)

q̂23 = C23 (1 x̂ ŷ)(x̂ + ŷ)
q̂31 = C31 x̂(1 x̂)

Adding these polynomial functions together, we construct on the element K a new

2
function q P2 (K)

q(x) = q12 (x) + q23(x) + q31(x); 8x 2 K ;

which satisfies
Z Z Z Z
n
rq ds = n
r(q12 + q23 + q31) ds = n
rq12 ds = n
r ds;
12 12 12 12
Z Z Z Z
n
rq ds = n
r(q12 + q23 + q31) ds = n
rq23 ds = n
r ds;
23 23 23 23
Z Z Z Z
n
rq ds = n
r(q12 + q23 + q31) ds = n
rq31 ds = n
r ds:
31 31 31 31

In other words, there exists a function  u 2 PP(K),  u = z p + q such that

Z
n
r(u  u) ds = 0; 8  @ K:

Now, by the triangle inequality,

ku  uk0;K  ku z pk0;K + kqk0;K

 kk0;K + kq12 k0;K + kq23 k0;K + kq31 k0;K
n o1 = 2
 kk0;K + C hK krk20;K + hK krk0;K kr2 k0;K
8 !1=2 9
< h h2K 2  1  2 =
 C : pKs + hK p2s 2
+ hK hpKs 1
hK
psK 2 ;
kuk s ; K
K K K
( )
hK hK 1
C psK
+ hK s 3=2
kuk s ; K
pK
hK
C s 3=2
kuk s ; K :
pK
30

In the same manner, we find:

kr(u  u)k0;K  kr(u z p)k0;K + krqk0;K

n o1=2
 krk0;K + C krk20;K + hK kr k0;K kr2 k0;K
( )
hK 1
hK 1
C psK 1
+ s 3=2
kuk s ; K
pK
 1
 C phs 3=2
kuk s ; K ;
and

kr2 (u  u)k0;K  kr2(u z p)k0;K + kr2 qk0;K

n o1=2
 kr2 k0;K + ChK 1 kr k20;K + hK krk0;K kr2 k0;K
( )
hK 2
hK 1
C psK 2
+ hK 1 s 3=2
kuk s ; K
pK
 2
 C hps 2 kuks;K :

We observe that the first two estimates are governed by the rate of convergence of
kk
q 0; K and kr k
q 0; K respectively, while the last estimate is governed by the rate of
convergence of kr k
2
0; K . 
4.2.4. A priori error estimate when c is zero
2 \ 
Theorem 4.3 Let u H 1 (Ω) H s (Ph ), s 2 be a solution of (3.23) and uh be the discrete
=
discontinuous solution of (3.24) with c 0 and p 2. Then, the numerical error e  =
u u h satisfies:
 1
krek0;P  C phs
h 5=2
kuk s (4.17)

where  = min(p + 1; s).

Proof: Let  u be the interpolant of u in V hp , defined on each element K of P h as
=
in Lemma 4.1. We also introduce  u  u and  u h  u. Using the triangle =
inequality, we have:

krek0;P = kr(  )k0;P  krk0;P + kr k0;P

h h h h
 31

and from (4.7) and (4.8), we recall that:

kr k20;P = B+(;  ) = B+(;  ¯ ) + B+(; ¯ ):

h

Here ¯ is chosen as the average of  over each K, i.e.

Z
¯ = 2 Ph :
1
j j K  dx;
K
K

We note here that the authors in [20,19] chose ¯ as the average of  over each edge
and their proof is thus slightly different from ours.
This particular choice of the interpolant  u and piecewise constant function ¯ does
indeed yield:

B+ (; ¯ ) = B(; ¯ ) + J(;¯ ) =

J(; ¯ ) J(; ¯ )
Z  
= Γint ΓD [
hn
r i ¯ ds
Z
= ¯  [
Γint ΓD
hn
r i ds
=0
since the last integral is zero according to the property (4.12) of the interpolant  u.
Therefore

kr k20;P = B+(;  ¯ )

h
(4.18)

We now show how B+ (;  ¯ ) can be bounded with respect to kr k0;Ph . We nat-
urally have from (4.10)

B+(;  ¯ )  j B(;  )j + J(;  ¯ ) + j J(; )j

The first term gives, using the approximation properties of Lemma 4.1 and the
discrete Schwarz inequality:
Z
j B(;  )j  ∑ jr
r j dx  ∑ krk0;K kr k0;K
2
K Ph K 2
K Ph

hK 1
∑ C s 3=2
kuks;Kkr k0;K
2
K Ph pK
 1
 C phs 3=2
kukskr k0;P h
32

The third term is treated as usual. We have

Z
j J(; )j  [
Γint ΓD
jhn
r i []j ds  ∑ khn
r ik0; k[]k0;
C ∑ ∑ kn
r k0; k k0;
2
K Ph 2@ KnΓN
C ∑ ∑ kr k0; k k0;
2
K Ph 2@ KnΓN
From the trace inequality (A.3) and the inverse property (A.7), we show that:
 1=2
kr k0;  C 1
hK
kr k + kr k0;K kr  k0;K
2
0; K
2

 1=2
p2K
C 1
hK
kr k + kr k
2
0; K C
0; K 0
hK
kr k0;K
 C p1K=2 kr k0;K
hK

and, from the approximation properties of Lemma 4.1:

 1=2
kk0;  C 1
hK
kk + kk0;K krk0;K
2
0; K

( )1=2
2
hK hK 1
C 1 hK
h K p2s 3
u kk + 2
s; K s 3=2 s 3=2
kuk 2
s; K
K pK pK
( )1=2
h2K 1
h2K 1
C p2s 3
+ p2s 3
kuk s ; K
K K
 1=2
 C hKs 3=2
kuk s ; K
pK

In conclusion, we find that:

 1=2
j J(; )j  C ∑ pK
1=2
kr k0;K hKs 3=2
kuk s ; K
2 hK
K Ph pK
hK 1
C ∑ 5=2
kr k0;K kuks;K
2 psK
K Ph
 1
 C phs 5=2
kukskr k0;P : h
 33

In the same manner as before, we obtain for the term J(;  ¯ ):

J(;  ¯ ) C ∑ ∑  0;  ¯ 0; kr k k k
K Ph 2 2@ KnΓN
In this case, we have using also the approximation properties of the interpolant:
 1=2
krk0;  C 1
hK
krk + krk0;K kr k0;K
2
0; K
2

( )1=2
2 2
hK 1
hK 2
C 1 hK
h K p2s 3
kuk2s;K + s 3=2 psK 2
kuk2s;K
K pK
( )1=2
h2K 3
h2K 2
C p2s 3
+ 2s 7=2
kuk s ; K
K pK
 3=2
C hK
s 7=4
kuk s ; K
pK
However, for the other term, we have, using Lemma A.5:
 1=2
k ¯ k0;  C 1
hK
k ¯ k + k ¯ k0;K kr( ¯ )k0;K
2
0; K

 1=2
C 1
hK
k ¯ k + k ¯ k0;K kr k0;K
2
0; K

 1=2
C 1 2
h
hK K
kr k + hK kr k0;K kr k0;K
2
0; K

 Ch1K=2kr k0;K
It follows that:
 3=2
J(;  ¯ ) C ∑ s 7=4
hK
kuks;K h1K=2kr k0;K
2
K Ph p K

hK 1
C ∑ 7=4
kuks;Kkr k0;K
K Ph 2 psK
 1
 C phs 7=4
kuks kr k0;P h

Combining the previous results, we finally get:

 
h 1 h 1 h 1  1
kr k0;P  C
h
+
p s 1=2
+ p s 7=4 p s 5=2
kuks  C phs 5=2
kuk s
and this completes the proof since kr k 0;P converges with a greater rate of con-
vergence than kr k0;P . 
h

h
34

4.2.5. Alternative estimate when c is nonzero

We now use the previous results to review the error estimate when c is nonzero.
The new estimate is given in the following theorem:

2 \ 
Theorem 4.4 Let u H 1 (Ω) H s (Ph ), s 2, be a solution of (3.23) with c > 0 and uh be
the discrete discontinuous solution of (3.24). Then, the numerical error e u uh satisfies: =
 1
keke;P  C phs
h 5=2
kuk s (4.19)

where  = min(p + 1; s) and p  2.

Proof: In this case, we have:

k k2e;P = B+(;  )
h

= B(;  ) J(;  ) + J(; )

= B(;  ) J(;  ¯ ) J(; ˆ ) + J(; )
= B(;  ) J(;  ¯ ) + J(; )
 j B(;  )j + J(;  ¯ ) + j J(; )j
if the interpolant is chosen as in Lemma 4.1.
Moreover, results from the previous theorem provide us with:
Z  1
j B(;  )j  ∑ jr 
r + c j dx  C hs 3=2
kuk s k ke ;P ;
2
h
K Ph K p
 1  1
j J(; )j  C phs 5=2
kuks kr k0;P  C phs
h 5=2
kuk s k ke ;P ;
h

 1  1
j J(;  )j  C phs 7=4
kuks kr k0;P  C phs
h 7=4
kuk s k ke ;P ;
h

so that
 1
k ke;P  C phs
h 5=2
kuk s ;
and this completes the proof. 
This time, the rate of convergence is optimal with respect to h but the rate of con-
vergence in p is worse than in the previous estimate. This makes us believe that the
error estimates for the DG method can still be improved with respect to p. Maybe
better interpolants are yet to be found.
 35

5. Concluding Remarks

5.1. Remarks on the Discontinuous Formulations

In this report, we studied four different formulations of the so-called Discontinu-
ous Galerkin Method (DGM). These formulations simply vary by one sign (plus or
minus) and by the addition of a penalty term (or not). However, they greatly differ
in nature from a mathematical point of view. We now review each formulation one
by one and recount our findings in the case of linear diffusion problems.
Global Element Method. Little can be proved for this method. We were able to
derive the continuity of the associated bilinear form, but failed to even obtain a
priori error estimates. This is because the bilinear form is not guaranteed to be
semi-positive definite.
Symmetric Interior Penalty Galerkin Method. The SIPG Method is similar to the
GEM except for the addition of the penalty term. However, it allows us to prove
non only continuity of the bilinear form, but also coercivity in the discrete discon-
tinuous space (for sufficiently large values of the penalty parameter), and thus a
priori error estimates optimal with respect to h ( 1) and slighty suboptimal with
respect to p (s 3=2). One major drawback of this method is that its behavior de-
pends on the selection of the penalty parameter. If not chosen carefully, the method
can fail.
Non-Symmetric Interior Penalty Galerkin Method. The limitation of the SIPG
method is remedied by changing one minus sign by a plus sign. Indeed, although
the NIPG formulation results in a non-symmetric system of equations, all the prop-
erties and error estimates are shown to be independent of the choice of the penalty
parameter. We also find the same rates of convergence with respect to h and p as
SIPG.
Discontinuous Galerkin Method. DGM is deduced from the NIPG method by
setting the penalty parameter to zero. We then observe that the rate of convergence
with respect to h or p deteriorates. Also, in the case of the pure Laplacian operator,
when c is set to zero in the Poisson problem, we obtain a priori error estimates only
by defining some new interpolants whose fluxes are weakly equal to the fluxes of
the exact solution over each edge of the elements. Although the rate of convergence
in h remains optimal, the one in p is then estimated to be s 5=2. We believe that
it might be possible to improve this rate of convergence by considering other types
of interpolants. At this point, detailed numerical experiments would be helpful to
understand how the penalty term affects the quality of the approximations.

5.2. Future Challenges

The great challenges for DGMs are to 1) prove uniqueness of the solutions of the
continuous formulations, 2) perform more numerical experiments to understand
36

the role played by the penalty terms, 3) still improve the a priori error estimates for
the Discontinuous Galerkin Method of Baumann and Oden, 4) derive rigorous a
posteriori error estimates for the various formulations.

Acknowledgement. The first author (S.P.) acknowledges the support of the Laboratoire de
Mathématique at the University of Paris-Sud at Orsay where part of the research on the subject of
this paper was performed. The second author (F.P.) was a TICAM visiting Faculty Fellow in 1999. We
wish to thank John Osborn of the University of Maryland for several useful suggestions regarding the
proofs of some lemmas. The TICAM authors gratefully acknowledge the support of their research by
ONR under Grant N00014-95-1-0401, by ARO under Contract DAAH04-96-1-0062 and by a project
supported by ERDC through Contract NRC-CR-97-0002 with Nichols Research Corporation.

REFERENCES
1. D. N. Arnold. An interior penalty finite element method with discontinuous
elements. SIAM J. Numer. Anal., 19:742–, 1978.
2. I. Babuška, C. E. Baumann, and J. T. Oden. A discontinuous hp finite element
method for diffusion problems: 1-D analysis. Computers and Mathematics with
Applications, 37:103–122, 1999.
3. I. Babuška and M. Suri. The h-p version of the finite element method with
quasiuniform meshes. Mathematical Modeling and Numerical Analysis, 21(2):199–
238, 1987.
4. I. Babuška and M. Suri. The optimal convergence rate of the p-version of the
finite element method. SIAM J. Numer. Anal., 24(4), 1987.
5. G. A. Baker. Finite element methods for elliptic equations using nonconform-
ing elements. Mathematics of Computations, 31(137):45–59, 1977.
6. G. A. Baker, W. N. Jureidini, and O. A. Karakashian. Piecewise solenoidal vec-
tor fields and the Stokes problem. SIAM J. Numer. Anal., 27(6):1466–1485, 1990.
7. C. E. Baumann. An h-p adaptive discontinuous finite element method for computa-
tional fluid dynamics. PhD thesis, The University of Texas at Austin, 1997.
8. S. C. Brenner and L. R. Scott. The Mathematical Theory of Finite Element Methods.
Springer-Verlag, New-York, 1994.
9. Z. Chen. On the relationship of various discontinuous finite element meth-
ods for second-order elliptic equations. SMU Math Report 2000-02, Southern
Methodist University, 2000.
10. B. Cockburn, G. E. Karniadakis, and C. W. Shu, editors. Discontinuous Galerkin
Methods - Theory, Computation and Applications, volume 11 of Lectures Notes in
Computational Science and Engineering. Springer, Berlin, 2000.
11. L. M. Delves and C. A. Hall. An implicit matching principle for global element
calculations. J. Inst. Math. Appl., 23(2):223–234, 1979.
12. L. M. Delves and C. Phillips. A fast implementation of the global element
method. J. Inst. Math. Appl., 25(2):177–197, 1980.
 37

13. J. A. Hendry and L. M. Delves. The global element method applied to a har-
monic mixed boundary value problem. J. Comput. Phys., 33:33–, 1979.
14. J. A. Hendry, L. M. Delves, and C. Phillips. Numerical experience with the
global element method. In Mathematics of finite elements and applications, III,
pages 341–348. Academic Press, London-New York, 1979. (Proc. Third MAFE-
LAP Conf., Brunel Univ., Uxbridge, 1978).
15. P. Houston, C. Schwab, and E. Süli. Discontinuous hp-finite element methods
for advection-diffusion problems. Technical Report no. 00/15, Oxford Univer-
sity Computing Laboratory, 2000.
16. O. A. Karakashian and W. N. Jureidini. A nonconforming finite element
method for the stationary Navier-Stokes equations. SIAM J. Numer. Anal.,
35(1):93–120, 1998.
17. J. T. Oden, I. Babuška, and C. E. Baumann. A discontinuous hp finite element
method for diffusion problems. TICAM Report 97-21, The University of Texas
at Austin, 1997.
18. J. T. Oden, I. Babuška, and C. E. Baumann. A discontinuous hp finite element
method for diffusion problems. J. Comput. Phys., 146:491–519, 1998.
19. B. Rivière. Discontinuous Galerkin Methods for Solving the Miscible Displacement
Problem in Porous Media. PhD thesis, The University of Texas at Austin, May
2000.
20. B. Rivière, M. F. Wheeler, and V. Girault. Improved energy estimates for interior
penalty, constrained and discontinuous Galerkin methods for elliptic problems.
Part I. Computational Geosciences, 3-4:337–360, 1999.
21. C. Schwab. p- and hp- Finite Element Methods - Theory and Applications in Solid
and Fluid Mechanics. Oxford University Press, New York, 1998.
22. E. Süli, C. Schwab, and P. Houston. hp-DGFEM for partial differential equa-
tions with nonnegative characteristic form. In B. Cockburn, G. E. Karniadakis,
and C. W. Shu, editors, Discontinuous Galerkin Methods, volume 11 of Lectures
Notes in Computational Science and Engineering, pages 221–230. Springer, Berlin,
2000.
23. M. F. Wheeler. An elliptic collocation-finite element method with interior
penalties. SIAM J. Numer. Anal., 15:152–, 1978.
38

A. Appendix

A.1. Discrete Schwarz Inequality

f g f g
Lemma A.1 Let ai and bi define two sequences of N real numbers. Then
!1=2 !1=2
N N N
∑ ai bi  ∑ a2i ∑ b2i (A.1)
=
i 1 =
i 1 =
i 1

Proof: We shall show the discrete Schwarz inequality for N = 2 first. We have:
(a1 b1 + a2 b2 )2 = a21 b21 + a22 b22 + 2a1 b1 a2 b2
= (a21 + a22)(b21 + b22) a21 b22 a22 b21 + 2a1 b1 a2 b2
= (a21 + a22)(b21 + b22) (a1 b2 a2 b1 )2
 (a21 + a22)(b21 + b22)
so that:
q q
a1 b 1 + a2 b 2  a21 + a22 b21 + b22 :
The result is easily extended to N > 2 by recursivity. 

A.2. Multiplicative Trace Inequalities

Lemma A.2 Let Ω define a star-shaped domain with a smooth boundary @ Ω as shown in
Fig. 4. Then, for all v H1 (Ω) 2
!
kvk 2
0;@ Ω  2
j j kvk + sup jxjkvk0;Ωkrvk0;Ω :
2
0;Ω (A.2)
inf x x2Ω
x2@ Ω

2
Proof: Let O Ω be the origin and let n denote the unit normal outward vector on
@ Ω. From the definition of a star-shaped domain, there exists a positive constant
such that
jxj  x
n:
Applying Green’s Theorem for the vector field u 2 x, we have:
Z Z

@Ω
2


u x n ds = Ω r
(u2x) dx:
By the property of star-shaped domains, the first integral is shown to be bounded
below:
Z Z

@Ω


u2 x n ds  jj
inf x
x2@ Ω @Ω
u2 ds  inf x u
2
x @Ω
j jk k20;@ Ω:
 39

δΩ
Ω

x β |x| < x.n

0
n

Figure 4. Star-shaped domain.

On the other hand, the second integral is bounded above:

Z Z

Ω
r
(u x) dx =
2
Ω
u2 r
x + x
ru2 dx
Z Z
= Ω
2u dx2
+ Ω
2ux
ru dx
Z
 2kuk20;Ω + Ω jux
ruj dx
Z
 2kuk20;Ω + 2 sup jxj Ω jujjruj dx
2
x Ω

 2kuk + 2 sup jxjkuk0;Ωkruk0;Ω

2
0;Ω
2
x Ω

Using both bounds, we arrive at:

!
kuk20;@ Ω  2
j j kuk20;Ω + sup jxjkuk0;Ωkruk0;Ω
inf x x2Ω
x2@ Ω

which completes the proof. 

Lemma A.3 Let K be a triangle or a quadrilateral such that hK  %K (shape regular).
2
Then, for all v H1 (K),
 
kvk 2
0;@ K  C
1
hK
v k k + kvk0;K krvk0;K :
2
0; K (A.3)

where C is a positive constant.

40

Proof: Let the origin O be the center of the inscribed circle in K with radius  K =2.
We therefore have:

sup x j j  hK
x K 2
inf x
2
x @K
j j  K  hK =%
so that from (A.2)
 
kuk20;@ K  h2% kuk20;K + hK kuk0;K kruk0;K
K
 
 2% 1
hK
u k k + kuk0;K kruk0;K
2
0; K

The proof is complete when choosing C = 2%. 

A.3. Poincaré–Friedrich’s Inequalities

Lemma A.4 Let Ω be an open, bounded, connected domain of R2 with Lipschitz boundary
2
@ Ω. Let v H1 Ω such that
Z

Ω
v dx = 0: (A.4)

Then

kvk0;Ω  Ckrvk0;Ω (A.5)

where C = C(Ω) is a positive constant.

Proof: See Schwab [21, p.350] and Brenner and Scott [8, p.102]. 
Lemma A.5 Let z 2 Pp (K) and z̄ be the average of z on K, z̄ = (RK z dx)=jKj. Then
K

kz z̄k0; K  Ch K kr zk0; K (A.6)

where C is a positive constant independent of K and z.

Proof: Let z 2 Pp (K) and v = z

K
z̄. Then
Z Z Z Z

Ω
v dx = Ω
z z̄ dx = Ω
z dx
Ω
z̄ dx = jKj z̄ j j = 0:
z̄ K

By a scaling argument and Lemma A.4,

kvk0;K  ChK kv̂k0;K̂  C(K̂)hK kr̂v̂k0;K̂  ChK krvk0;K

Substituting z z̄ for v, it follows that k z z̄k0; K  Ch K kr(z k0;K , in other
z̄)
words, since z̄ is constant, k z z̄k0; K  Ch K kr zk0; K . 
 41

A.4. Inverse Property

Lemma A.6 Let z 2 Pp (K). Then
K

2
krzk0;K  C hpK kzk0;K (A.7)
K

Proof: See Schwab [21, p.208]. 

A.5. Interpolation Error Estimates

Lemma A.7 Let K be a triangle or parallelogram element of the partition Ph and u a
function in H s (K). There exists a positive constant C depending on s and % but independent
2 =  
of u, p K , and h K , and a sequence z p Pp K (K), p K 1; 2; : : : , such that for any q, 0 q s
 q
ku zp kq ; K  C hK
s q kuk s ; K ; s 0 (A.8)
pK
 1=2
ku zp k0;  C hKs 1=2
kuk s ; K ; s>
1
2
(A.9)
pK

where  = min(pK + 1; s), hK = diam (K) and  @ K.

Proof: See Babuška and Suri [3,4].