Professional Documents
Culture Documents
10 1137@S003614290037174X
10 1137@S003614290037174X
Abstract. We analyze three discontinuous Galerkin approximations for solving elliptic problems
in two or three dimensions. In each one, the basic bilinear form is nonsymmetric: the first one has a
penalty term on edges, the second has one constraint per edge, and the third is totally unconstrained.
For each of them we prove hp error estimates in the H 1 norm, optimal with respect to h, the mesh
size, and nearly optimal with respect to p, the degree of polynomial approximation. We establish
these results for general elements in two and three dimensions. For the unconstrained method, we
establish a new approximation result valid on simplicial elements. L2 estimates are also derived for
the three methods.
Key words. discontinuous Galerkin methods, elliptic equations, error estimates, penalties
PII. S003614290037174X
1. Introduction. Over the past five years, discontinuous Galerkin methods have
become very widely used for solving a large range of computational fluid problems.
They are preferred over more standard continuous methods because of their flexibility
in approximating globally rough solutions, their local mass conservation, their pos-
sible definition on unstructured meshes, their potential for error control and mesh
adaptation, and their straightforward applications to anisotropic materials. Another
interesting aspect of these methods is that they can employ polynomials of degree k
(Pk ) on quadrilaterals, whose dimension is substantially lower than that of Qk , the
standard space of functions on quadrilaterals. Similarly, they lead to smaller systems
of equations than the locally conservative mixed finite element methods and do not
require the solution of saddle-point problems.
In this paper, we discuss three numerical algorithms for elliptic problems which
employ discontinuous approximation spaces. The three methods are called the non-
symmetric interior penalty Galerkin (NIPG) method, the nonsymmetric constrained
Galerkin (NCG) method, and the discontinuous Galerkin (DG) method. The three
algorithms are closely related in that the underlying bilinear form for all three is the
same and is nonsymmetric.
In the NIPG method, we modify the bilinear form of the interior penalty Galerkin
method treated by Douglas and Dupont [9], Wheeler [18], Arnold [2], and Wheeler
and Darlow [19]. Here, we antisymmetrize the bilinear form by changing the sign of
one term and as a consequence we require only a positive penalty, whereas the proofs
in [18] assume that the penalty is bounded below by a problem-dependent constant.
We choose the penalty term in such a way that we can establish both H 1 and L2
optimal convergence rates.
∗ Received by the editors May 1, 2000; accepted for publication (in revised form) March 8, 2001;
In the NCG method, we impose that the jumps on each edge (or face) of the
subdivision have integral average zero. One constraint per edge (or face) is sufficient
for again proving optimal rates of convergence in the H 1 and L2 norms.
Downloaded 12/30/12 to 129.173.72.87. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php
The DG method is totally unconstrained and for that reason its error analysis
is more delicate. We recover full accuracy in the H 1 norm by introducing a special
interpolant, local to each element, whose average flux is orthogonal to constants on
each edge (or face). This interpolant is constructed first for constant tensors and then
extended to W 1,∞ tensors by an approximation result.
The results presented here extend significantly the work in [15]. To our knowledge
these results are new.
The DG method with this bilinear form was first introduced by Baumann [6], and
Oden, Babus̆ka, and Baumann in [13]. In [6], Baumann showed that the method is
elementwise conservative and he proved a stability result in one dimension for poly-
nomials of at least degree three. However, his proof is based on an inf-sup condition,
which seems problematic in higher dimensions. Our approach uses special approxi-
mation results and trace theorems but no isomorphisms.
This paper consists of five sections after this introduction. In section 2, we list the
notation, state the problem, and describe the formulation of the three methods. In
sections 3, 4, and 5, the proofs of the error estimates of the three methods described
in section 2 are respectively given. A numerical example is presented in section 6. In
the last section, we present some conclusions.
1 1
{φ} = (φ|Ei )|ek + (φ|Ej )|ek , [φ] = (φ|Ei )|ek − (φ|Ej )|ek .
2 2
904 B. RIVIÈRE, M. F. WHEELER, AND V. GIRAULT
The reader can refer to Adams [1] and to Lions and Magenes [12] for the properties
of Sobolev spaces.
Let r be a positive integer. The finite element subspace is taken to be
Nh
Dr (Eh ) = Pr (Ej ),
j=1
where by Pr (Ej ) we mean that the polynomials are defined on Ej and not on a
reference set Ê. This distinction is unnecessary when Ej is a triangle, a tetrahedron,
or a parallelogram because the transformation from Ê to Ej is affine. But it is
important when Ej is a quadrilateral: in this case, it is shown by [3] that Pr (Ê) does
not have optimal approximation properties, whereas it is shown by [10] that Pr (Ej )
has optimal approximation properties.
We use the following hp approximation properties, proven in [4, 5]. Let Ej ∈ Eh
and φ ∈ H s (Ej ). Then there exists a constant C depending on s, τ, ρ but independent
of φ, r, and h and a sequence zrh ∈ Pr (Ej ), r = 1, 2, . . . , such that for any 0 ≤ q ≤ s
hµ−q
j
(2.2) φ − zrh q,Ej ≤C φ s,Ej , s ≥ 0,
rs−q
µ− 12
hj 1
(2.3) φ− zrh 0,γi ≤C φ s,Ej , s> ,
1
rs− 2 2
where µ = min(r + 1, s) and γi is an edge or a face on ∂Ej . Using the same technique
as in [4], it can be shown that we have the additional approximation result
µ− 32
hj 3
(2.4) φ− zrh 1,γi ≤C φ s,Ej , s> .
3
rs− 2 2
Finally, we shall often use the following bounds with respect to h that extend to three-
dimensional formulae (2.4), (2.5) of [2]:
1
(2.5) ∀ φ ∈ H 1 (Ej ), φ 2
0,γi ≤C |φ|2 + hj |φ|21,Ej ,
hj 0,Ej
2
∂φ 1
(2.6) ∀ φ ∈ H 2 (Ej ), ≤C |φ|2 + hj |φ|22,Ej .
∂ν 0,γi hj 1,Ej
ERROR ESTIMATES 905
Proof. Let us briefly recall the proof for the reader’s convenience. In all the
proofs, C will be a generic constant with different values on different places, that is,
independent of h and r. We consider the case where n = 2; the proof extends easily
−1
to three dimensions. The factor hE 2 results from a straightforward scaling argument,
and we have only to exhibit the factor r in a reference setting. First, let us consider
the case where E is the reference square [−1, 1]×[−1, 1], e is the segment [−1, 1]×{1},
and p is a polynomial of Qr , i.e., a polynomial of degree at most r in each variable.
Then, expanding p in terms of Legendre polynomials
p(x1 , x2 ) = αm,n Lm (x1 )Ln (x2 ),
m,n
Therefore,
1
2 2
p 0,e = αm,n 1
m n
m+ 2
2.2. Problem and nonsymmetric bilinear form. Let the boundary of the
domain ∂Ω be the union of two disjoint sets ΓD and ΓN . We denote ν the unit normal
1
vector to each edge of ∂Ω exterior to Ω. For f given in L2 (Ω), p0 given in H 2 (ΓD ),
Downloaded 12/30/12 to 129.173.72.87. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php
(2.10) p = p0 on ΓD ,
K∇p · ν = g on ΓN .
With the above assumptions on K and α, problem (2.10) has a unique solution in
H 1 (Ω) when |ΓD | > 0 or when α = 0. On the other hand, when ∂Ω = ΓN and
α = 0, problem (2.10)
has a solution in H 1 (Ω) which is unique up to an additive
constant, provided Ω f = − ∂Ω g (see [11]) . For K in W 1,4 (Eh ) and ψ, φ ∈ H 2 (Eh ),
we consider the nonsymmetric bilinear form
Nh
aN S (ψ, φ) = (K∇ψ∇φ + αψφ)
j=1 Ej
Ph
Ph
− {K∇ψ · ν k } [φ] + {K∇φ · ν k } [ψ]
k=1 ek k=1 ek
− (K∇ψ · ν k )φ + (K∇φ · ν k )ψ.
ek ∈ΓD ek ek ∈ΓD ek
2.3. Finite element schemes. First we introduce the following interior and
boundary penalty term:
Ph
rσk
rσk
J0σ,β (φ, ψ) = [φ][ψ] + φψ,
|ek |β ek |ek |β ek
k=1 ek ∈ΓD
where σ is a discrete positive function that takes the constant value σk on the edge
or face ek and is bounded below by σ0 > 0, above by σm , |ek | denotes the measure of
ek , and β ≥ 12 is a real number. The Galerkin approximation P NIPG in Dr (Eh ) solves
the following discrete problem:
that (2.12) has at most one solution p in H 2 (Eh ) (or H 2 (Eh )/R if |ΓD | = 0 and α ≡ 0).
Indeed, if p solves (2.12) with zero data, then p is constant in each Ej and [p] = 0 on
each interior edge ek . Hence p is constant in Ω and if |ΓD | > 0 or α = 0, then this
constant is zero. Otherwise, p is unique in H 2 (Eh )/R. Thus, it suffices to prove that
if the unique solution p of (2.10) belongs to H 2 (Eh ), then it also solves (2.12). For
this, let v be an element in H 2 (Eh ). We multiply the first equation of (2.10) by v,
integrate on Ej , and sum all over j.
Nh
Ph
(K∇p∇v + αpv) − {K∇p · ν k }[v] − (K∇p · ν)v = (f, v).
j=1 Ej k=1 ek ∂Ω
We add ek ∈ΓD ek (K∇v ·ν k )p and ek ∈ΓD |erσk |kβ ek pv to both sides, use the Dirich-
let boundary condition, and note that [p] = 0. We clearly have (2.12).
We note that on each element, the mass conservation for the NIPG method can
be written as
rσk
αP NIPG − {K∇P NIPG · ν ∂Ej } + [P NIPG ][1]
Ej ∂Ej \ΓN |ek |β ek
ek ∈∂Ej \∂Ω
rσk NIPG rσk
+ P = f + g + p0 .
|ek |β ek Ej ∂Ej ∩ΓN |ek |β ek
ek ∈∂Ej ∩ΓD ek ∈∂Ej ∩ΓD
The fact that this scheme is consistent with problem (2.10) has been shown by
Baumann [6].
908 B. RIVIÈRE, M. F. WHEELER, AND V. GIRAULT
Lemma 2.3. If r ≥ 1, the discrete solution of each of the three methods exists
and is unique, with one exception: if α is not bounded away from zero, then in the
DG scheme, we assume that r ≥ 2 and the elements are not quadrilaterals.
Downloaded 12/30/12 to 129.173.72.87. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php
1 1 hµ−1
|||K 2 ∇(P NIPG − p)|||0 ≤ C ,K |||p|||s .
σ0 rs−1−δ
If α ≥ α0 > 0, then
1 hµ−1
|||P NIPG − p|||1 ≤ C , K, α ∞ |||p|||s ,
σ0 rs−1−δ
h2µ−2
(3.1) J0σ,β (P NIPG − p, P NIPG − p) ≤ C |||p|||2s ,
r2s−2−2δ
ERROR ESTIMATES 909
Ph
Ph
− {K∇(p − p̃) · ν k }[χ] + {K∇χ · ν k }[p − p̃]
k=1 ek k=1 ek
− (K∇(p − p̃) · ν k )χ + (K∇χ · ν k )(p − p̃)
ek ∈ΓD ek ek ∈ΓD ek
Ph
rσk
rσk
(3.2) + [p − p̃][χ] + (p − p̃)χ.
|ek |β ek |ek |β ek
k=1 ek ∈ΓD
The first two terms in (3.2) can be bounded in the following way:
Nh
K∇(p − p̃) · ∇χ ≤ C|||∇(p − p̃)|||0 |||K 2 ∇χ|||0 ,
1
j=1 Ej
1 1
≤ |||K 2 ∇χ|||20 + C|||∇(p − p̃)|||20 .
6
Nh
1
α(p − p̃)χ ≤ α 2
p − p̃
1
α 2 χ 0,
∞ 0
j=1 Ej
1 1 2 1
≤ α2 χ 0 + α ∞ p − p̃ 20 .
2 2
The third term is bounded by
1 1
|ek |β 2
rσk 2
{K∇(p − p̃) · ν k }[χ] ≤ {K∇(p − p̃) · ν k } [χ] 0,ek ,
rσk
0,ek
|ek |β
ek
P
h 1 Ph
|ek |β
{K∇(p − p̃) · ν k }[χ] ≤ J0σ,β (χ, χ) + C {K∇(p − p̃) · ν k } 2
0,ek .
ek 8 rσk
k=1 k=1
h
(3.3) {K∇(p − p̃) · ν k }[χ] ≤ J0σ,β (χ, χ) + |ek |β k2s−2 |||p|||2s,E 12 .
ek 8 σ0 r k
k=1 k=1
In the case of triangles or tetrahedra, we can choose a continuous p̃ and for the
pure Neumann problem, there are no other terms; thus we can conclude by choosing
β ≥ (n − 1)−1 . In the general case, to bound the fourth term in (3.2), we deduce
similarly by using the approximation result (2.3) and the inverse inequality (2.8)
{K∇χ · ν k }[p − p̃] ≤ {K∇χ · ν k } 0,e [p − p̃] 0,e
k k
ek
µ− 12
r 1 h̄k
≤C 1 |||K ∇χ|||0,Ek12
2
1 |||p|||s,Ek12 ,
h̄k2 rs− 2
P
h 1 h2µ−2
1
(3.4) {K∇χ · ν k }[p − p̃] ≤ |||K 2 ∇χ|||20 + C 2s−3 |||p|||2s .
ek 6 r
k=1
1 1 h2µ−2 h2µ
|||K 2 ∇χ|||20 + α 2 χ 2
0 + J0σ,β (χ, χ) ≤ C 2s−2
|||p|||2s + C α ∞ 2s |||p|||2s ,
r r
Ph
2µ−3
h̄ h̄2µ−1
+C |ek |β k2s−3 + |ek |−β k2s−2 |||p|||2s,E 12
r r k
k=1
h2µ−3 2µ−1
−β hk
(3.6) +C |ek |β k
+ |ek | |||p|||2s,Ek .
r2s−3 r2s−2
ek ∈ΓD
−∇ · (K∇φ) + αφ = f in Ω,
K∇φ · ν = 0 on ∂Ω
−∇ · (K∇φ) + αφ = P NIPG − p in Ω,
K∇φ · ν = 0 on ∂Ω.
Note that if α ≡ 0, then since both P NIPG and p are defined up to an additive
constant, this constant can be chosen so that P NIPG − p has zero mean value. This
compatibility condition guarantees that the dual problem has a unique solution.
Denote χ = P NIPG − p. Then
2
χ 0,Ω = (−∇ · (K∇φ) + αφ, χ).
Nh
Ph
= (K∇φ∇χ + αφχ) − {K∇φ · ν k }[χ]
j=1 Ej k=1 ek
because the regularity of φ and K implies that the jumps [K∇φ · ν k ]|ek = 0. By
subtracting the orthogonality equation for any φ∗ ∈ Dr (Eh )
and using the symmetry of K, the regularity of φ and K, and choosing for φ∗ a
continuous interpolant of φ satisfying (2.2)–(2.4), we obtain
Nh
2
χ 0,Ω = (K∇(φ − φ∗ )∇χ + α(φ − φ∗ )χ)
j=1 Ej
Ph
Ph
+ {K∇(φ − φ∗ ) · ν k }[χ] − 2 {K∇φ · ν k }[χ].
k=1 ek k=1 ek
912 B. RIVIÈRE, M. F. WHEELER, AND V. GIRAULT
The first two terms are bounded in the following fashion by using (2.2) and (3.7):
Nh Nh
Downloaded 12/30/12 to 129.173.72.87. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php
K∇(φ − φ∗
)∇χ ≤C 1
φ − φ∗ 1,Ej K 2 ∇χ 0,Ej
j=1 Ej j=1
h 1
φ 2,Ω |||K 2 ∇χ|||0
≤C
r
h 1
≤ C χ 0,Ω |||K 2 ∇χ|||0 ;
r
Nh
2
α(φ − φ∗
)χ ≤ C α ∞ h φ 2,Ω χ 0,Ω
r2
j=1 Ej
h2
≤C α ∞ 2 χ 0,Ω |||χ|||1 .
r
By the Cauchy–Schwarz inequality, by (3.1), for any β ≥ (n − 1)−1 , and by properties
(2.4) and (3.7), we have
P
h Ph
|ek |β
1 1
2−2
∗
{K∇(φ − φ ) · ν k }[χ] ≤ {K∇(φ − φ∗ ) · ν k } 0,ek [χ] 0,ek
ek rσk
k=1 k=1
1
Ph
2
1 |ek |β
≤ J0σ,β (χ, χ) 2 {K∇(φ − φ∗ ) · ν k } 2
0,ek
rσk
k=1
β
µ− 12 2 (n−1)
h h
≤C |||p|||s χ 0,Ω .
rs
Similarly, by Cauchy–Schwarz, (3.1), and (2.6),
P
h Ph
|ek |β
1
2
rσk
1
2
On the other hand, for the pure Dirichlet problem, even on triangles or tetrahedra,
we must choose β = (n − 1)−1 in order to establish Theorem 3.1, and these are
incompatible with the values of β in Theorem 3.2. However, if the Dirichlet datum
Downloaded 12/30/12 to 129.173.72.87. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php
We can show [14] that there is an interpolant p̃ ∈ Dr∗ (Eh ) that satisfies the approxi-
mation properties (2.2)–(2.4). Let χ = P NCG − p̃; then
aN S (χ, χ) = aN S (p − p̃, χ)
Nh
= (K∇(p − p̃)∇χ + αχ(p − p̃))
j=1 Ej
Ph
Ph
(4.1) − {K∇(p − p̃) · ν k }[χ] + {K∇χ · ν k }[p − p̃].
k=1 ek k=1 ek
The first two terms are easily bounded by Cauchy–Schwarz and by using (2.2):
Nh
K∇(p − p̃)∇χ ≤ |||K 2 ∇χ|||0 |||K 2 ∇(p − p̃)|||0
1 1
j=1 Ej
12
Nh
h2µ−2
j
≤ C|||K 2 ∇χ|||0 p 2s,Ej .
1
(4.2) 2s−2
j=1
r
Now we estimate the third term in (4.1), which we denote by A. Let ck be any
constant:
Downloaded 12/30/12 to 129.173.72.87. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php
Ph
Ph
A≡ {K∇(p − p̃) · ν k }[χ] = {K∇(p − p̃) · ν k }[χ − ck ].
k=1 ek k=1 ek
The other factor is bounded in the following way; here again, we assume that ek is
the interface between two elements E 1 and E 2 of Eh :
χ̂ − ck 0,ê
ˆ ê χ̂
≤ Ĉ ∇ 0,ê ,
χ̂ − ck ˆ ê χ̂
≤ Ĉr ∇
0,ê 0,Ê .
Therefore,
1 1 1
(χ − ck )E 1 0,ek ≤ Ĉ|ek | 2 h1 |E 1 |− 2 r ∇χ 0,E 1 ≤ Ĉh12 r ∇χ 0,E 1 .
hµ−1 1
(4.6) |A| ≤ C 5 |||p|||s |||K 2 ∇χ|||0 .
rs− 2
ERROR ESTIMATES 915
1
(4.7) {K∇χ · ν k }[p − p̃] ≤ C|||K 2 ∇χ|||0 s− 3 |||p|||s .
ek r 2
k=1
and therefore the error terms arising from the Dirichlet boundary condition have the
same estimates as the interior jump terms. Hence, if we impose (4.8), the statement
of Theorem 4.1 is also valid for the problem with mixed boundary conditions.
The next theorem shows that the NCG scheme gives optimal results with respect
to h in the L2 norm. For simplicity, we do not specify the dependence with respect
to r.
Theorem 4.2. Under the assumptions of Theorem 3.2, we have
P NCG − p 0,Ω ≤ Chµ |||p|||s
for r ≥ 1, s ≥ 2, and C independent of h and p.
Proof. As in the proof of Theorem 3.2, we consider the following dual problem:
−∇ · (K∇ψ) + αψ = P NCG − p in Ω,
K∇ψ · ν = 0 on ∂Ω.
Denote χ = P NCG − p. Thus, we have
2
χ 0,Ω = (−∇ · (K∇ψ) + αψ, χ)
Nh
Ph
(4.9) = (K∇ψ∇χ + αψχ) − {K∇ψ · ν k }[χ].
j=1 Ej k=1 ek
Now, we subtract (4.10) from (4.9) and use the regularity of K∇ψ · ν k on ek :
Nh
χ 20,Ω = (K∇(ψ − ψ ∗ )∇χ + α(ψ − ψ ∗ )χ)
j=1 Ej
Ph
Ph
(4.11) −2 (K∇ψ · ν k )[χ] + {K∇(ψ − ψ ∗ ) · ν k }[χ].
k=1 ek k=1 ek
916 B. RIVIÈRE, M. F. WHEELER, AND V. GIRAULT
As in the proof of Theorem 4.1, the first two terms are easily bounded by using
Cauchy–Schwarz, Hölder inequalities, and the approximation property (2.2):
Downloaded 12/30/12 to 129.173.72.87. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php
Nh
K∇(ψ − ψ )∇χ ≤ Ch χ 0,Ω |||K 2 ∇χ|||0 ,
∗ 1
j=1 Ej
Nh
α(ψ − ψ )χ ≤ C α ∞ h2 χ 0,Ω |||χ|||1 ,
∗
j=1 Ej
1
and we apply Theorem 4.1 to bound the factor |||K 2 ∇χ|||0 in the first term and |||χ|||1,Ω
in the second term. To bound the third term in (4.11), let p̃ be an element of Dr∗ (Eh )∩
C 0 (Ω̄) and 1c be any constant vector; we can write
Ph
Ph
2 (K∇ψ · ν k )[χ] = 2 (K∇ψ · ν k )[P NCG − p̃]
k=1 ek k=1 ek
Ph
=2 ((K∇ψ − 1c) · ν k )[P NCG − p̃].
k=1 ek
As it was proved in Theorem 4.1, (4.5), and from the approximation result (2.2), we
have
Ph
Nh
[P NCG − p̃] 2
0,ek ≤C hj ∇(P NCG − p̃) 2
0,Ej
k=1 j=1
Nh
Nh
NCG 2 2
≤C hj ∇(P − p) 0,Ej +C hj ∇(p − p̃) 0,Ej
j=1 j=1
≤ Ch2µ−1 |||p|||2s .
On the other hand, by choosing
1
1c = 1 K∇ψ,
|E | E1
then
P
h
2 (K∇ψ · ν k )[χ] ≤ Chµ χ 0,Ω |||p|||s .
ek
k=1
|A| ≤ C {∇(ψ − ψ ∗ ) · ν k } 2
0,ek [χ] 2
0,ek .
k=1 k=1
ERROR ESTIMATES 917
{∇(ψ − ψ ∗ ) · ν k } 2
0,ek ≤ Ch|||ψ|||22 ≤ Ch χ 2
0,Ω ,
k=1
Thus, we obtain
|A| ≤ hµ χ 0,Ω |||p|||s .
Ph
Ph
−2 (K∇ψ · ν k )[χ] + {K∇(ψ − ψ ∗ ) · ν k }[χ]
k=1 ek k=1 ek
−2 (K∇ψ · ν k )χ + {K∇(ψ − ψ ∗ ) · ν k }χ.
ek ∈ΓD ek ek ∈ΓD ek
Since p̃ can be constructed so that ek χ = 0 for all ek in ΓD , the above boundary
terms have the same estimates as the interior jump terms.
5. A priori error estimate for the DG method. In this section, we derive
an a priori optimal error estimate in two or three dimensions for the problem with
Neumann, or Dirichlet, or mixed boundary conditions, on triangles, parallelograms,
or tetrahedra. We make several additional assumptions:
• K ∈ [W 1,∞ (Ej )]n×n ∀ j = 1, . . . , Nh , uniformly in Ej , i.e., ∃ γ2 > 0,
12
n
max ∇kim 2∞,Ej ≤ γ2 .
Ej
i,m=1
It is easy to prove that K̄ is also symmetric positive definite and that the
largest eigenvalue of K̄ is ≤ γ1 and its smallest eigenvalue is ≥ γ0 , where γ1
and γ0 are the constants of (2.1).
918 B. RIVIÈRE, M. F. WHEELER, AND V. GIRAULT
a3
e1
e2
Downloaded 12/30/12 to 129.173.72.87. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php
E
a2
e3
a1
Fig. 5.1. Triangular element.
hµ−i
(5.3) ∇i (p̃I − p) 0,E ≤C E
3 p s,E for i = 0, 1, 2,
rs− 2 −δ
µ− 32
I hE
(5.4) |K̄∇(p̃ − p) · ν k | ≤ C 3 p s,E ,
ek rs− 2
where a12 is the midpoint of e3 = [a1 , a2 ]. It is easy to check that each component of
∇q1 = 4q1 (a12 )∇λ1 (1 − 2λ1 ) has zero mean-value on e2 and e3 , and (∇λ1 )λ1 vanishes
on e1 . Therefore, q1 (a12 ) is determined by the conditions
4q1 (a12 ) K̄∇λ1 · ν 1 = K̄∇f · ν 1 .
e1 e1
However,
ν 1 |e1 |
(5.5) ∇λ1 = − .
2 |E|
Therefore,
1 |E| 1
q1 (a21 ) = − 2
K̄∇f · ν 1 .
2 |e1 | (K̄ν 1 , ν 1 ) e1
ERROR ESTIMATES 919
a3
a4
Downloaded 12/30/12 to 129.173.72.87. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php
a2
a1 E
Hence,
C γ1 1 ∂f
|q1 (a12 )| ≤ K̄∇f · ν 1 ≤ C hE2 .
γ0 e1 γ0 ∂n 0,e1
Therefore, for i = 0, 1, 2,
1 1 1
−i ∂f
(5.6) ∇i q1 0,E ≤ C|E| 2 h−i
E |q1 (a12 )| ≤ C|E| hE
2 2
.
∂n 0,e1
and such that the inequalities of (5.6) hold for q2 and q3 , with respect to e2 and e3 .
Let q = q1 + q2 + q3 , constructed with f = p − pI , where pI is an approximation of p
satisfying (2.2)–(2.4), and set p̃I = q + pI . Then p̃I satisfies (5.2), and applying (2.4)
to (5.6) we derive, for i = 0, 1, 2,
|p̃I − p|i,E ≤ |q|i,E + |pI − p|i,E
hµ−i
≤C E
3 p s,E + |pI − p|i,E ,
rs− 2
which has the same order of approximation as |pI − p|i,E . Similarly,
|K̄∇(p̃I − p) · ν k | ≤ |K̄∇q · ν k | + |K̄∇(pI − p) · ν k |,
ek ek ek
and applying (2.4) to the second term, (2.6) to the first term, and the above estimate
for |q|i,E , we obtain (5.4).
The case of tetrahedra. The situation is much the same for tetrahedra. Let E be
a tetrahedron with vertices a1 , a2 , a3 , a4 , opposite faces e1 , e2 , e3 , e4 , and unit exterior
normal vectors ν 1 , ν 2 , ν 3 , ν 4 as in Figure 5.2. Let λ1 , λ2 , λ3 , λ4 be the barycentric
s
coordinates of a1 , a2 , a3 , and a4 in E. Again, we will show that given f in H (E)
with s ≥ 2, there is a polynomial q1 in P2 (E) such that e1 K∇(q1 − f ) · ν 1 = 0,
eδ
K∇q1 · ν δ = 0 for δ = 2, 3, 4. For this, consider the polynomial
υ2 a3
a4 e2
υ1
Downloaded 12/30/12 to 129.173.72.87. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php
e1
υ3
E
e3
e4 a2
υ4
a1
Fig. 5.3. Parallelogram element.
1 1 ∂f
∇i q1 0,E ≤ C|E| 2 h−i −i
E |q1 (a1 )| ≤ C|E| hE
2 .
∂n 0,e1
and we want to adjust the numbers qi so that q satisfies the above conditions. Con-
sidering that E is a parallelogram, it is easy to check that
|ei |
∇φi = − ν i.
|E|
ERROR ESTIMATES 921
Set c3 = |e3 |(K̄ν 3 , ν 3 ), c4 = |e4 |(K̄ν 4 , ν 4 ), c34 = |e3 |(K̄ν 3 , ν 4 ), c43 = |e4 |(K̄ν 3 , ν 4 ).
Then the numbers qi must satisfy the following system of equations:
Downloaded 12/30/12 to 129.173.72.87. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php
|e3 |
(3c3 q1 + c43 q2 + c3 q3 − c43 q4 ) = K̄∇f · ν 3 ,
|E| e3
c3 q1 − c43 q2 + 3c3 q3 + c43 q4 = 0,
−c34 q1 + c4 q2 + c34 q3 + 3c4 q4 = 0,
c3 q1 + 3c4 q2 − c34 q3 + c4 q4 = 0.
In order to solve this system of equations and derive an upper bound for qi , set
α34 = (K̄ν 3 , ν 3 )(K̄ν 4 , ν 4 ) − (K̄ν 3 , ν 4 )2 ,
and observe that
1 1
α34 = K̄ 2 ν 3 2
K̄ 2 ν 4 2
sin2 θ,
1 1
where θ is the angle between K̄ 2 ν 3 and K̄ 2 ν 4 . Owing to the regularity of Eh , this
angle is bounded below for all E ∈ Eh by a fixed angle θ0 > 0:
1 1
∀ E ∈ Eh , α34 ≥ K̄ 2 ν 3 2
K̄ 2 ν 4 2
sin2 θ0 ≥ γ02 sin2 θ0 .
Then, the solution of the above system is
b |E| (K̄ν 3 , ν 4 )
q4 = , q2 = −q4 ,
2 |e3 ||e4 | α34
b |E| 1 (K̄ν 3 , ν 4 )2 1
q3 = − 2
+ ,
2 |e3 | (K̄ν 3 , ν 3 ) α34 4
b |E| 1 (K̄ν 3 , ν 4 )2 3
q1 = 2
+ ,
2 |e3 | (K̄ν 3 , ν 3 ) α34 4
where b = e3
K̄∇f · ν 3 . We can easily check that
C γ12 1 1 ∂f
|q2 | = |q4 | ≤ 2 2 hE2 ,
2 γ0 sin θ0 ∂n 0,e3
C γ1 12 γ12 1 ∂f
|q3 | ≤ h + ,
2 γ0 E 2 2
γ0 sin θ0 4 ∂n 0,e3
C γ1 12 γ12 3 ∂f
|q1 | ≤ h + .
2 γ0 E γ02 sin2 θ0 4 ∂n 0,e3
From these bounds, we immediately deduce that ∇i q satisfies (5.6) and the proof
finishes as in the triangular case.
The next corollary extends the result of Theorem 5.1 to nonconstant tensors.
Corollary 5.2. We retain the notation of Theorem 5.1. Let K be a tensor-
valued function in [W 1,∞ (E)]n×n . There exists h0 > 0 independent of E such that
for all hE ≤ h0 , there exists an interpolant of p, p̃I ∈ Dr (Eh ) satisfying
K∇(p̃I − p) · ν k = 0 ∀ ek ∈ ∂E,
ek
hµ−i
∇i (p̃I − p) 0,E ≤C E
3 p s,E , i = 0, 1, 2,
rs− 2 −δ
µ− 32
I hE
|K∇(p̃ − p) · ν k | ≤ C 3 p s,E ,
ek rs− 2
922 B. RIVIÈRE, M. F. WHEELER, AND V. GIRAULT
where · denotes the matrix norm subordinated to the Euclidean norm. It is easy
to check that
Therefore,
In the course of the proof of Theorem 5.1, for any f ∈ H s (E), we have constructed a
polynomial q ∈ P2 (E) such that
∀ e ∈ ∂E, K̄∇q · ν = K̄∇f · ν.
e e
We propose to achieve the same result, with K instead of K̄, by means of the following
iterative procedure. Starting with q 0 = 0, construct q m ∈ P2 (E), as in Theorem 5.1,
the solution of
(5.8) ∀ e ∈ ∂E, K̄∇q · ν = K∇f · ν − (K − K̄)∇q m−1 · ν.
m
e e e
m
Theorem 5.1 guarantees the existence of q ; moreover, combining (5.6), (5.7), and
(5.8) we obtain in the two-dimensional (2D) case
1 −1
(5.9) ∇q m 0,E ≤ C2 |E| 2 hE 2 ( ∇f · ν 0,∂E + C∞ hE ∇K ∞,E ∇q m−1 · ν 0,∂E )
(5.10) m
∇q 0,E ≤ 2C 2 |E| 2 h 2 ∇f · ν
E 0,∂E in two dimensions,
1 −1
2C3 |E| hE ∇f · ν 0,∂E in three dimensions.
2
This is trivially true for m = 0. Assume that (5.10) holds for m − 1. Applying (2.8),
we have in the 2D case
1 −1 1
∇q m 0,E ≤ C2 |E| 2 hE 2 ( ∇f · ν 0,∂E + CC∞ hE2 ∇K ∞,E ∇q m−1 0,E )
ERROR ESTIMATES 923
with a similar inequality in the 3D case. Therefore, in both cases, there exists a
constant C0 that depends on the above constants but is independent of h, m, and E,
such that if
Downloaded 12/30/12 to 129.173.72.87. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php
1
(5.11) h0 =
2C0 ∇K ∞,E
and hE ≤ h0 , then the perturbation is bounded by ∇q m−1 0,E /2. But either
∇q m−1 0,E ≤ ∇q m 0,E , in which case (5.10) is established for ∇q m 0,E , or
and the induction hypothesis implies (5.10). Therefore, in both cases, (5.10) is proven
by induction. Furthermore, substituting (5.10) and (2.8) into the analogue of (5.9)
for ∇i q m , we easily derive that, for i = 0, 1, 2,
1
1 −i
i m
∇ q 0,E ≤ C|E| 2 h2
E ∇f · ν 0,∂E in two dimensions,
1 −i
C|E| hE ∇f · ν 0,∂E in three dimensions.
2
The estimates for q follow from the previous bounds and the argument of Theorem
5.1.
Corollary 5.3. Let h ≤ h0 , where h0 is defined by (5.11), let Eh consist of
triangles, parallelograms, or tetrahedra (in three dimensions), and let p ∈ H s (Eh ), for
s ≥ 2, and let r ≥ 2. There exists an interpolant of p, p̃I ∈ Dr (Eh ) satisfying
(5.12) {K∇(p̃I − p) · ν k } = 0 ∀ k = 1, . . . , Ph ,
ek
(5.13) K∇(p̃I − p) · ν k = 0 ∀ ek ∈ ΓD ,
ek
hµ−i
(5.14) |||∇i (p̃I − p)|||0 ≤ C 3 |||p|||s , i = 0, 1, 2,
rs− 2 −δ
Ph
Mh
3
hµ− 2
(5.15) |{K∇(p̃I − p) · ν k }| + |K∇(p̃I − p) · ν k | ≤ C(K) 3 ,
k=1 ek k=Ph +1 ek rs− 2
1 hµ−1
|||K 2 ∇(p − P DG )|||0 ≤ C(K) 5 |||p|||s .
rs− 2
924 B. RIVIÈRE, M. F. WHEELER, AND V. GIRAULT
hµ−1
Downloaded 12/30/12 to 129.173.72.87. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php
aN S (P DG − p, v) = 0 ∀ v ∈ Dr (Eh ).
Ph
Ph
− {K∇(p − p̃I ) · ν k }[χ] + {K∇χ · ν k }[p − p̃I ]
k=1 ek k=1 ek
− (K∇(p − p̃I ) · ν k )χ + (K∇χ · ν k )(p − p̃I ).
ek ∈ΓD ek ek ∈ΓD ek
Ph
Ph
− {K∇(p − p̃I ) · ν k }([χ] − ck ) + {K∇χ · ν k }[p − p̃I ]
k=1 ek k=1 ek
I
(5.16) − (K∇(p − p̃ ) · ν k )(χ − ck ) + (K∇χ · ν k )(p − p̃I ),
ek ∈ΓD ek ek ∈ΓD ek
Hence,
1 1
[χ] − ck 0,ek ≤ C(h12 |χ|1,E 1 + h22 |χ|1,E 2 ).
The first two terms and the fourth term are estimated as in Theorem 4.1 (the same
result as in (3.4) except for the dependence with respect to r that becomes r2s−5 ).
Finally, the boundary terms are estimated as the interior terms, and the theorem is
Downloaded 12/30/12 to 129.173.72.87. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php
where ||| · |||k,m denotes the Sobolev broken norm on W k,m (Eh ). We observe that the
4
arguments in the proof of Lemma 2.2 are valid for φ in H 1 (Ω) ∩ W 2, 3 (Eh ):
aN S (φ, v) = χv ∀ v ∈ H 2 (Eh ).
Ω
Therefore, we obtain
2
χ 0,Ω = aN S (φ, χ).
We note that by symmetry of K, the following property holds for the nonsymmetric
bilinear form aN S :
Nh
Nh
aN S (φ, χ) + aN S (χ, φ) = 2 K∇χ∇φ + 2 αχφ.
j=1 Ej j=1 Ej
Thus, we obtain
Nh
2
(5.19) χ 0,Ω =2 (K∇χ∇φ + αχφ) − aN S (χ, φ).
j=1 Ej
926 B. RIVIÈRE, M. F. WHEELER, AND V. GIRAULT
Nh
2
χ 0,Ω =2 (K∇φ∇χ + αχφ) − aN S (χ, φ − φ∗ ) − aN S (p − p̃, φ∗ )
j=1 Ej
Nh
Nh
Nh
∗
=2 (K∇φ∇χ + αχφ) − K∇χ∇(φ − φ ) − αχ(φ − φ∗ )
j=1 Ej j=1 Ej j=1 Ej
Nh
Nh
− K∇(p − p̃)∇φ∗ − α(p − p̃)φ∗
j=1 Ej j=1 Ej
Ph
Ph
+ {K∇χ · ν k }[φ − φ∗ ] − {K∇(φ − φ∗ ) · ν k }[χ]
k=1 ek k=1 ek
Ph
(5.20) + {K∇(p − p̃) · ν k }[φ∗ ].
k=1 ek
4 3
Using the fact that φ belongs to W 2, 3 (Eh ) 7→ H 2 (Eh ), we bound the volume terms
as follows:
Nh
2 (K∇φ∇χ + αφχ) ≤ C(|||K 12 ∇χ|||0 + |||α 12 χ|||0 ) χ 0 ,
j=1 Ej
Nh
(K∇χ∇(φ − φ ∗
) + αχ(φ − φ ∗
))
1 1 1
The first edge term in (5.20) is bounded by the inverse estimate (2.8), the trace
theorem (2.5) combined with Corollary 5.3, with the same notation as in Theorem 3.1:
P
h Ph
−1 1
{K∇χ · ν k }[φ − φ∗ ] ≤ C h̄k 2 r K 2 ∇χ 0,Ek12 h̄k |||φ||| 32 ,Ek12
ek
k=1 k=1
1 1
(5.21) ≤ Crh 2 |||K 2 ∇χ|||0 χ 0,Ω .
P
h Ph
∗
{K∇(φ − φ ) · ν k }[χ] ≤ {K∇(φ − φ∗ ) · ν k } 0,ek [χ] − ck 0,ek
ek
k=1 k=1
Ph
1 1
≤C |||φ|||2, 43 ,E 12 h̄k2 |||K 2 ∇χ|||0,Ek12
k
k=1
1
Ph
2
1 1
≤ Ch |||K ∇χ|||0
2 2 |||φ|||22, 4 ,E 12
3 k
k=1
1 1
≤ Ch 2 |||K 2 ∇χ|||0 |||φ|||2, 43
1 1
≤ Ch 2 |||K 2 ∇χ|||0 χ 0,Ω .
(5.22)
Using the approximation properties, we can bound the last term in (5.20)
P
h Ph
∗
{K∇(p − p̃) · ν k }[φ ] ≤ {K∇(p − p̃) · ν k } 0,ek [φ∗ − φ] 0,ek
ek
k=1 k=1
3
Ph
µ−
h̄ 2 k
≤C 3 |||p|||s,Ek12 h̄k |||φ||| 32 ,E 12
k=1
rs− 2 k
µ− 12
h
(5.23) ≤C 3 |||p|||s χ 0,Ω .
rs− 2
The theorem is obtained by combining all the bounds above.
Remark. In three dimensions, the duality argument used in proving Proposi-
tion 5.5 is more delicate. The regularity of the dual problem (5.17)–(5.18) should be
3
as follows: for any f ∈ L2 (Ω), the solution φ of (5.17)–(5.18) belongs to W 2, 2 (Eh )
with continuous dependence on f . But, as it is shown in [8], for K sufficiently smooth,
φ belongs to W 2,p (Ω) for any p < p0 and p0 ≥ 43 , and this is substantially smaller
4
than 32 . To simplify the discussion, let us assume that φ belongs to W 2, 3 (Eh ). Then
4 5
on one hand W 2, 3 (E) 7→ H 4 (E) and on the other hand, the trace of ∇φ on a face
1 4 3
e belongs to W 4 , 3 (e) 7→ L 5 (e). Thus, by using Hölder’s inequality, we can modify
the proof of Proposition 5.5 and derive in three dimensions that, under the very mild
4
regularity assumption φ ∈ W 2, 3 (Eh ), we have
hµ−1 1
P DG − p 0,Ω ≤C 5 (C1 + C2 h 2 r).
rs− 2
6. Numerical example. The reader will find in [6, 13] several examples of
numerical computations with the DG method that confirm our theoretical results in
the case of smooth solutions. In contrast, the experiment in this section investigates
the robustness of the DG method when solving problem (2.10) with discontinuous
coefficients and a nonsmooth solution. The domain Ω = (−1, 1)2 is divided into
four subdomains Ωi as shown in Figure 6.1. The tensor K is a constant tensor Ki
over each subdomain and we assume that K1 = K3 and K2 = K4 . We consider
928 B. RIVIÈRE, M. F. WHEELER, AND V. GIRAULT
111111
000000
Downloaded 12/30/12 to 129.173.72.87. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php
000000
111111
000000
111111
Ω2Ω
111111
000000 1
111111
000000
00000
11111
00000
11111
00000
11111
Ω Ω
00000
11111
3 4
00000
11111
Fig. 6.1. Initial domain for the problem with discontinuous coefficients.
the case where K1 = 5Id and K2 = Id. We seek an analytical solution of the form
rα (ai sin(αθ) + bi cos(αθ)), where (r, θ) are the polar coordinates of a given point in
Ω and ai , bi are constants that depend on the subdomain Ωi .
We can easily show that with the choice α = 0.53544094560, the analytical so-
lution pex satisfies the usual interface conditions: pex and K∇pex · ν are continuous
across the interfaces and −∇ · K∇p = 0 on Ω. In addition, the solution has a singu-
larity at the origin, so that it belongs to H 1 (Ω) but not to H 2 (Ω). We first analyze
the convergence of the DG solution on a sequence of uniformly refined meshes. The
coarse mesh consists of the four subdomains Ωi . The relative error in the L2 norm,
DG
defined as p−Pp 0 0 , is plotted against the number of degrees of freedom in Figure 6.2.
First, we choose finite elements of degree two, and then of degree three. We see, for
these elements and in this example, that the convergence rate does not depend on the
degree of the approximation, but that it depends on the regularity of the solution. In
that case, the convergence rate in the L2 norm is O(h2α ) = O(h1.07 ), where h is the
0
10
r=2
r=3
−1
10
Relative L2 error
−2
10
0.53
1
−3
10
1 2 3 4 5
10 10 10 10 10
Number of degrees of freedom
Fig. 6.2. Relative error in the L2 norm versus the number of degrees of freedom for uniform
meshes in the case r = 2, 3.
ERROR ESTIMATES 929
0
10
Downloaded 12/30/12 to 129.173.72.87. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php
r=2
r=3
Relative H error
1
0
−1
10
0.26
1
−2
10
1 2 3 4 5
10 10 10 10 10
Number of degrees of freedom
Fig. 6.3. Relative error in the H 1 norm versus the number of degrees of freedom for uniform
meshes in the case r = 2, 3.
mesh size, or O(N α ), where N is the number of degrees of freedom. We also study
the numerical error in the energy norm and we see in Figure 6.3 that the convergence
α
rate of the error in the energy norm is O(hα ) = O(N 2 ), which is a little bit better
than expected.
We now study the influence of adaptivity on the numerical DG solution. We refer
to [16] for more details on the mesh adaptation strategy that we use here. Here,
the error indicators are the norms of solutions of a discrete problem, for which the
approximation order is r + q. Therefore, we denote by q the degree of enrichment.
We begin with a uniform mesh of size h = 1 and we adaptively refine the mesh.
We fix r = 2 and q = 5 and we see that adaptive mesh refinements lead to optimal
convergence rates (Figure 6.4).
0
10
Downloaded 12/30/12 to 129.173.72.87. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php
unif r=2
adap r=2 q=5
Relative H error
1
0
−1
10
0.26
1
−2
10
1 2 3 4 5
10 10 10 10 10
Number of degrees of freedom
Fig. 6.4. Relative error in the H 1 norm versus the number of degrees of freedom for uniformly
and adaptively refined meshes in the case r = 2, q = 5.
REFERENCES
[15] B. Rivière, M. F. Wheeler, and V. Girault, Improved energy estimates for interior
penalty, constrained and discontinuous Galerkin methods for elliptic problems. Part I,
Comput. Geosci., 3 (1999), pp. 337–360.
Downloaded 12/30/12 to 129.173.72.87. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php
[16] B. Rivière and M. F. Wheeler, A posteriori error estimates and mesh adaptation strategy
for discontinuous Galerkin methods applied to diffusion problems, Comput. Math. Appl.,
to appear.
[17] B. Rivière Discontinuous Galerkin Methods for Solving the Miscible Displacement Problem
in Porous Media, Ph.D. Thesis, The University of Texas at Austin, Austin, 2000.
[18] M. F. Wheeler, An elliptic collocation-finite element method with interior penalties, SIAM
J. Numer. Anal., 15 (1978), pp. 152–161.
[19] M. F. Wheeler and B. L. Darlow, Interior penalty Galerkin procedures for miscible dis-
placement problems in porous media, in Computational Methods in Nonlinear Mechanics,
J. Tinsley Oden, ed., North-Holland, Amsterdam, New York, 1980, pp. 485–506.