Discontinuous Galerkin Finite Element Methods For Interface Problems: A Priori and A Posteriori Error Estimations

c 2011 Society for Industrial and Applied Mathematics
Downloaded 05/04/16 to 202.3.77.252. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php
SIAM J. NUMER. ANAL.

Vol. 49, No. 5, pp. 17611787
DISCONTINUOUS GALERKIN FINITE ELEMENT METHODS FOR

INTERFACE PROBLEMS: A PRIORI AND A POSTERIORI ERROR
ESTIMATIONS
ZHIQIANG CAI , XIU YE , AND SHUN ZHANG
Abstract. Discontinuous Galerkin (DG) nite element methods were studied by many researchers for second-order elliptic partial dierential equations, and a priori error estimates were
established when the solution of the underlying problem is piecewise H 3/2+ smooth with > 0.
However, elliptic interface problems with intersecting interfaces do not possess such a smoothness. In
this paper, we establish a quasi-optimal a priori error estimate for interface problems whose solutions
are only H 1+ smooth with (0, 1) and, hence, ll a theoretical gap of the DG method for elliptic
problems with low regularity. The second part of the paper deals with the design and analysis of
robust residual- and recovery-based a posteriori error estimators. Theoretically, we show that the
residual and recovery estimators studied in this paper are robust with respect to the DG norm, i.e.,
their reliability and eciency bounds do not depend on the jump, provided that the distribution of
coecients is locally quasi-monotone.
Key words. a priori error estimation, a posteriori error estimator, discontinuous Galerkin
methods, interface problems
AMS subject classifications. 65N30, 65N15
DOI. 10.1137/100805133
1. Introduction. Consider the following interface problem:

(k(x) u) = f
(1.1)
in
with boundary conditions

(1.2)
u = gD
on D
and n (k u) = gN
on N ,
where f , gD , and gN are given scalar-valued functions; is a bounded polygonal

D
N and D N = ; n = (n1 , n2 ) is the
domain in 2 with boundary =
outward unit vector normal to the boundary; and diusion coecient k(x) is positive
and piecewise constant on polygonal subdomains of with possible large jumps across
subdomain boundaries (interfaces):
k(x) = ki > 0
in i
for i = 1, . . . , n. Here, {i }ni=1 is a partition of the domain with i being an open

polygonal domain. Dene
Received
by the editors August 12, 2010; accepted for publication (in revised form) June 2,
2011; published electronically September 13, 2011. This work was supported in part by the National
Science Foundation under grant DMS-0810855.
http://www.siam.org/journals/sinum/49-5/80513.html
Department of Mathematics, Purdue University, West Lafayette, IN 47907-2067 (zcai@math.
purdue.edu).
Department of Mathematics, University of Arkansas at Little Rock, Little Rock, AK 72204
(xxye@ualr.edu). This authors work was supported in part by National Science Foundation grant
DMS-0813571.
Division of Applied Mathematics, Brown University, Providence, RI 02912 (Shun Zhang@brown.
edu).
1761
Copyright by SIAM. Unauthorized reproduction of this article is prohibited.
1762
ZHIQIANG CAI, XIU YE, AND SHUN ZHANG
kmin = min ki
1in
and kmax = max ki .

1in
Assume that f is square integrable over . For simplicity, we assume that gD and
gN are piecewise linear and constant, respectively, and that D is not empty (i.e.,
mes (D ) = 0).
Discontinuous Galerkin (DG) nite element methods for elliptic boundary value
problems have been studied since the late 1970s and now constitute an active research
area (see, e.g., [6, 7] and recent books [25, 33]). For problems with discontinuous
coecients such as interface problems, the stabilization (edge jump) term in the DG
nite element method needs special treatments in order to be robust. Robustness in
this paper means that constants in a priori error estimates or in the reliability and
eciency bounds of a posteriori error estimators are independent of the jump of the
coecients. Recently, Ern, Stephansen, and Zunino [23] developed a DG method using
general weighted averages and proper weights for the stabilization term for advectiondiusion equations and established a robust a priori error estimate, provided that the
solution is piecewise H 2 smooth.
A posteriori error estimation for continuous Galerkin nite element methods has
been extensively studied for the past three decades (see, e.g., books by Verf
urth [37],
Ainsworth and Oden [4], and Babuska and Strouboulis [8] and the references therein).
Recently, there has been increasing interest in a posteriori error estimation for the
DG nite element method (see, e.g., [27, 9, 26, 2, 20, 34, 35, 36, 18]).
For elliptic interface problems considered in this paper, robust a posteriori error
estimators have been investigated. For the continuous Galerkin method, Bernardi
and Verf
urth [11] and Petzoldt [31] studied a residual-based estimator, and we in
[14] studied recovery-based estimators that were further extended to mixed and nonconforming nite element methods in [15]. For a DG nite element method without
proper weights, Ainsworth [3] developed an a posteriori error estimator based on the
so-called numerical ux. This result is further extended to meshes with hanging nodes
in [5]. However, the error estimator and its analysis in [3, 5] depend on the jumps
of coecients. Recently, Ern, Nicaise, and Vohralik [19] and Ern, Stephansen, and
Vohralik [21] investigated an equilibrated error estimator using the numerical ux
for the DG method with the harmonic average weight; and Ern and Stephansen [22]
studied a residual-based error estimator for DG approximations to diusion and to
advection-diusion-reaction equations. Both the equilibrated and the residual-based
estimators studied in [21, 22] contain a so-called nonconforming error term which is
the energy norm of dierence between the DG approximation and its recovery through
the Oswald interpolation. Theoretically, they showed that their reliability constants
are independent of the jump of the diusion coecients, but their eciency constants
do depend on the jump. This dependency could be removed for the interface problem by using a modied Oswald interpolate (see, e.g., section 4) as remarked in [22],
provided that the diusion coecient is locally quasi-monotone. Such a hypothesis is
assumed in order to prove the robustness of estimators in [11, 31, 14, 15] and in this
paper.
For elliptic interface problems with intersecting interfaces, it is well known [28, 24]
that their solutions may have only H 1+ regularity with small > 0 and are not
piecewise H s , s > 3/2, smooth. Since the standard a priori error estimate of the
DG method requires that the exact solution be piecewise H 3/2+ smooth, there are
no known results for elliptic interface problems. One of the purposes of this paper
is to ll this theoretical gap. To this end, we introduce a nonstandard variational
formulation (see (2.14)) that uses general weighted averages. The formulation is
DG METHODS FOR INTERFACE PROBLEMS
1763
dened in an appropriate solution space V (see (2.9)) that permits discontinuity

across interior edges of a triangulation T and that does not require piecewise H s ,
s > 3/2, smoothness. The corresponding DG (or standard Galerkin) nite element
approximation is then the solution of this variational problem in a discontinuous (or
continuous) nite dimensional subspace. In this setting, the error equation is then
obtained in a straightforward manner and, most importantly, unnecessary smoothness
of the solution is not assumed.
As usual, this formulation involves a parameter {1, 0, 1} and a stabilization
parameter . For = 1, the formulation is stable for all 0. For = 1 or 0,
we show that there exists a positive constant 0 such that the variational problem in
the discontinuous piecewise polynomial space is stable, provided that 0 . The
0 is computable and depends only on the shape of elements but not on the mesh
size and the jump of the diusion coecient. With this discrete stability and the
error equation, we are then able to obtain a robust a priori error estimate of the
DG method for the interface problem with low regularity. Note that the DG method
corresponding to = 1 was rst introduced and analyzed in [23] for smooth solution.
The second part of the paper deals with the development and analysis of various
robust a posteriori error estimators including residual- and recovery-based a posteriori error estimators. The residual-based estimator studied in the paper is standard
and may be regarded as an extension of that by Bernardi and Verf
urth [11] and Petzoldt [31] to the DG method. The recovery-based a posteriori error estimators follow
ideas of our previous work in [14, 15] for conforming, nonconforming, and mixed nite
element methods. Theoretically, we show that the residual and recovery estimators
are robust with respect to the DG norm; i.e., their reliability and eciency bounds do
not depend on the jump, provided that the distribution of coecients is locally quasimonotone. Finally, we remark that there is numerical evidence showing that both the
residual and the recovery estimators are not subject to the locally quasi-monotone
assumption (see, e.g., [14, 15]).
This paper is organized as follows. DG nite element methods are introduced
and their well-posedness is established in section 2. In section 3, we obtain a robust
a priori error estimate in a norm which is stronger than the broken energy norm.
Modied Oswald and Clement types of interpolations are described in section 4. A
residual-based a posteriori error estimator is introduced and analyzed in section 5.
We introduce and analyze ux recoveries and the resulting recovery-based estimators
in sections 6 and 7, respectively. Numerical results for a test problem are reported in
section 8.
1.1. Notation. For a subdomain G , we use the standard notation and definitions for Sobolev spaces (see, e.g., Lions and Magenes [29], Adams [1], or Grisvard
[24]). For G being an open region, denote the Sobolev space by W s,r (B) on B = G or
G equipped with the standard Sobolev norm s,r,B , where s is a real number and
1 r . When r = 2, W s,2 (B) is a Hilbert space and is denoted by H s (B) with
the norm s,B . When s = 0, W 0,r (B) is the standard Lr (B) space. By the trace
theorem [24], the trace of any function in W s,r (G) lies in W s1/r,r (G), provided
that s 1/r > 0 is not an integer.
In two dimensions, for a vector-valued function = (1 , 2 )t , dene the divergence
by
=
1
2
+
.
x1
x2
1764
For a scalar-valued function v, dene the operator by

t
v v
0
,
with Q =
v = Q v =
1
x2 x1
1
0

.
We shall use the Hilbert space

H(div; G) = { L2 (G)2 : L2 (G)}
equipped with the norm
1/2

H(div; G) = 20,G + 20,G
.
Finally, let
1
Hg,D
() = {v H 1 () : v = g on D }
and
Hg,N (div; ) = { H(div; ) : n = gN on N }.
2. Discontinuous finite element approximation.
2.1. Finite element spaces. For simplicity of presentation, consider only triangular elements. Let T = {K} be a nite element partition of the domain . Assume
that the triangulation T is regular (see [17]); i.e., for all K T , there exists a positive
constant such that
h K K ,
where hK denotes the diameter of the element K and K the diameter of the largest
circle that may be inscribed in K. Note that the assumption of the regularity does
not exclude highly locally rened meshes. Furthermore, assume that interfaces
F = {i j : i, j = 1, . . . , n}
do not cut through any element K T .
Let Pk (K) be the space of polynomials of degree k on element K. Denote the
conforming continuous piecewise linear nite element space and the discontinuous
Galerkin linear nite element space associated with the triangulation T by
U = {v H 1 () : v|K P1 (K) K T }
and
U DG = {v L2 () : v|K P1 (K) K T },
and denote the subspaces of U by
Ug = {v U : v = gD on D }
and U0 = {v U : v = 0 on D }.
Denote the local lowest-order RaviartThomas (RT) [32, 13] and BrezziDouglas
Marini (BDM) spaces [12, 13] on element K T by
RT0 (K) = P0 (K)2 + x P0 (K) and BDM1 (K) = P1 (K)2 ,
1765
respectively, where x = (x1 , x2 ). Then the standard H(div; ) conforming RT and

BDM spaces are dened by
RT0 = { H(div; ) : |K RT0 (K) K T }
and
BDM1 = { H(div; ) : |K BDM1 (K) K T },
respectively. For convenience, denote RT0 or BDM1 by V.
2.2. Jumps and averages. Denote by EK the set of three edges of element
K T . Denote the set of all edges of the triangulation T by
E := EI ED EN ,
where EI is the set of all interior element edges, and ED and EN are the sets of all
boundary edges belonging to the respective D and N . For each e E, denote by he
the length of the edge e; denote by ne a unit vector normal to e. When e ED EN ,
e
denote by K+
the element with the edge e, and assume that ne is the unit outward
e
e
and K
be the two elements
normal vector. For each interior edge e EI , let K+
e
sharing the common edge e such that the unit outward normal vector of K+
coincides
with ne . For any e E, denote by v|e+ and v|e , respectively, the traces of a function
v over e.
Dene jumps over edges by
e
e EI ,
v|+ v|e ,
[[v]]e :=
e
v|+ ,
e ED EN .
e
e
Let w+
[0, 1] and w
[0, 1] be weights dened on e satisfying
e
e
w+
(x) + w
(x) = 1,
(2.1)
and dene the weighted averages

{v(x)}ew =
e e
e e
v + w+
v+ , e EI ,
w
v|e+ ,
e ED EN ,
e e
e e
v + w
v+ ,
w+
e EI ,
0,
e ED EN ,
and

{v(x)}w
e
for all e E. When there is no ambiguity, the subscript or superscript e in the designation of the jump and the weighted averages will be dropped. A simple calculation
leads to the following identity:
(2.2)
e
[[uv]]e = {v}w
e [[u]]e + {u}w [[v]]e .
e
e
and K
,
Let e be the interface of elements K+
e
e
e = K+
K
,
1766
e
e
e
e
and denote by k+
and k
the diusion coecients on K+
and K
, respectively. There
are several possible choices of the weights:
e
e
k
k
1
e
e
e
(2.3)
w,1 = , w,2 = e
, and w,3 = e e .
e
2
k+ + k
k+ + k
Denote by
We = {k}ew
the weighted average of k on edge e. For the above choices of the weights, We is then
the arithmetic, the harmonic, and the geometric averages:
(2.4)
We,1 =
e
e
+ k
k+
,
2
We,2 =
e e
k
2k+
e
e ,
k+ + k
and We,3 =

e ke ,
k
+
respectively. It is well known that these averages have the following relations:
We,2 We,3 We,1 .
Since
e 2 e
e 2 e
(w+
) k+
) k+
(we )2 k e
(w+
e
= e e+ +
w+
1,
e
e
e
e
We
w+ k+ + w k
w+ k+
the same argument shows that

e 2 e
) k+
(w+
1
We
(2.5)
and
e 2 e
) k
(w
1
We
for any weights satisfying (2.1). Letting

e
e
e
kmin
= min{k+
, k
}
e
e
e
and kmax
= max{k+
, k
},
it is easy to check that

(2.6)
1 e
e
k
We,1 kmax
,
2 max
e
e
kmin
We,2 2kmin
,
e
e
and kmin
We,3 kmax
,
which implies that the arithmetic and harmonic averages are equivalent to the maxie
, set
mum and the minimum, respectively. For boundary edge e K+
e
w+
=1
e
and We = k+
.
2.3. DG finite element approximation. To describe the DG nite element

method, we introduce a nonstandard variational formulation for (1.1)(1.2) dened in
a proper solution space which is a subspace of piecewise H 1+ functions for 0 < 1.
The corresponding DG nite element approximation is the solution of this variational
problem in a discontinuous nite dimensional subspace. Also, the standard Galerkin method is the variational problem restricted in a continuous nite dimensional
subspace.
To dene a proper solution space, let us start with the following Greens formula:

(w n) v ds := w n, vK = (w, v)K + (w, v)K K T
(2.7)
K
1767
holds for all w H 1+ (K) with w L2 (K) and for all v H 1 (K) with 0 < <
1/2. Let H (K) be the dual of H0 (K) which is the closure of C0 (K) in the H (K)
norm. Since H (K) is the same space as H0 (K) for (0, 1/2) (see, e.g., Theorem
1.4.2.4 in [24]), v is then in H (K)2 . That is, the term (w, v)K in (2.7) can
be viewed as a duality pairing between H (K)2 and H (K)2 . The validity of (2.7)
follows from the standard density argument and the fact that (2.7) holds for C (K)
functions.
By the trace theorem [24], v|K is in H 1/2 (K). Hence, the formal boundary
integral in the left-hand side of (2.7) may be regarded as the duality pairing between
H 1/2 (K) and H 1/2 (K), which is dened by the right-hand side of (2.7). Since,
for each edge e K, the trivial extension of functions in H 1/2 (e) by zero to all
of K belongs to H 1/2 (K) (see, e.g., Theorem 1.5.2.3 in [24]), this interpretation
enables us to dene the duality pairing on each edge e of K,

(w n) v ds := w n, ve ,
e
1/2
where (w n)|e H
(e) and v|e H 1/2 (e).
Lemma 2.1. Letting K T , e K, and 0 < < 1/2, for any H 1+ (K)
with L2 (K), there exists a positive constant C independent of such that

(2.8)
n1/2,e C ,K + h1
K 0,K .
Proof. Inequality (2.8) is contained in the proof of Corollary 3.3 on page 1384
of [10]. For the convenience of readers, we provide a proof here. For any g H 1/2 (e),
there exists a lifting vg of g such that vg H 1 (K), vg |e = g, vg |K\e = 0, and
vg ,K + h1
K vg 0,K cg1/2,e.
It then follows from the Greens formula in (2.7), the CauchySchwarz inequality, and
the denition of the dual norm that
n, ge = n, vg K = (, vg )K + (, vg )K
0,K vg 0,K + ,K vg ,K

C ,K + h1
K 0,K g1/2,e ,
which, combining with the denition of the dual norm
1/2,e =
n, ge
,
gH 1/2 (e) g1/2,e
sup
implies (2.8). This completes the proof of the lemma.

Denote by H s (T ) the broken Sobolev space of degree s > 0 with respect to T ,
H s (T ) = {v L2 () : v|K H s (K) K T },
and denote its subspace by
(2.9)
V s (T ) = {v H s (T ) : (kv) L2 (K) K T }.
Let u be the solution of problem (1.1)(1.2); then it is well known from the regularity
estimate [28, 24] that u H 1+ () for some positive which could be very small.
1768
Since f L2 (), it is then easy to see that u V 1+ (T ) for any 0 < < , and the
ux = ku is in H(div; ) H (T )2 .
Denote the discrete gradient and divergence operators by
(h v)|K = (v|K ) and (h )|K = ( |K )
for all K T , respectively. Multiplying (1.1) by a test function v V 1+ (T ),
integrating by parts, and using boundary conditions (1.2), we have
(ku, v)K
(ku ne ) v ds
(f, v) =
KT
KT
= (kh u, h v)
eEI
[[(ku ne ) v]] ds
eED

(ku ne ) v ds
gN v ds.
N
Together with (2.2) and the continuity of the ux, for all v V 1+ (T ),

[[ku ne ]] {v}w ds = 0 e EI
(2.10)
e
implies that
(2.11)
(kh u, h v)
eEI ED
{ku ne }w [[v]] ds = (f, v) +
gN v ds
N
for all v V 1+ (T ). By the continuity of u H 1+ () and the Dirichlet boundary

condition, we have that

(2.12)
{kv ne }w [[u]] ds =
gD (kv ne )ds
e
eEI ED
and
(2.13)
eEI ED
eED
h1
e We [[u]][[v]] ds =
h1
e We
ED
gD vds
for all v V 1+ (T ), where > 0 is a parameter to be determined.

For {1, 0, 1}, dene a bilinear form for u, v V 1+ (T ) by

a (u, v) = (kh u, h v) +
h1
e We [[u]][[v]] ds
eEI ED
eEI ED
f (v) =
(f, v)K +
KT

EN
eEI ED
and a linear form for v V 1+ (T ) by
{ku ne }w [[v]]ds +
h1
e We
ED
gN v ds +
{kv ne }w [[u]]ds

eED
gD v ds
gD (kv ne ) ds.
1769
Note that the bilinear form is symmetric if = 1 and nonsymmetric otherwise. Using
(2.11), (2.12), and (2.13), the weak solution of (1.1)(1.2) u H 1+ () V 1+ (T )
satises the following variational equation:
v V 1+ (T ).
a (u, v) = f (v)
(2.14)
Note that (2.13) is used in (2.14) for stabilizing the formulation or, equivalently,
enforcing weakly the continuity of the solution. For = 1, (2.12) is used for symmetrizing and stabilizing the bilinear form. Note also that the Dirichlet boundary
condition is enforced weakly in (2.14). One could enforce it strongly in the solution
space by removing all boundary integrals over edge e ED .
Remark 2.2. It is obvious that the bilinear form a1 (, ) is coercive in V 1+ (T )
with respect to the energy/DG norm if 1 > 0. But it is dicult, if not impossible, to
show that the bilinear form a (, ) is coercive in V 1+ (T ) for = 1, 0. Nonetheless,
the fact that the weak solution of (1.1)(1.2) satises problem (2.14) implies existence
of solutions of problem (2.14). Uniqueness of problem (2.14) follows from that of
(1.1)(1.2) and the fact that solution of problem (2.14) satises (1.1)(1.2) in the
weak sense.
Now, the corresponding DG nite element method is to nd uT U DG
1+
(T ) such that
V
a (uT , v) = f (v) v U DG .
(2.15)
Methods corresponding to = 1, 0, or 1 are the so-called symmetric interior penalty

Galerkin (SIPG), incomplete interior penalty Galerkin (IIPG), or nonsymmetric interior penalty Galerkin (NIPG) methods, respectively. A special and interesting case
of the NIPG method is 1 = 0, which was studied in [30] by Oden, Bab
uska, and
Baumann. With the special choices of weights in (2.3), the corresponding DG methods in (2.15) are called the arithmetic, the harmonic, and the geometric weighted DG
methods, respectively. The SIPG method dened in (2.15) with general weights but a
slightly dierent stabilization term was introduced and analyzed recently in [23] and
a robust a priori error bound was obtained, provided that the solution is piecewise
H 2 smooth and that is large enough.
2.4. Well-posedness of the DG finite element formulation. For any element K T , let {i (x)}3i=1 be the standard barycentric coordinates of K. Following
the idea of Lemma 1 in [3], we introduce an element stiness matrix of the Laplace
operator (instead of the diusion operator in [3]):
K
K
SK = Sij
and Sij
= (i , j )K .
33
The matrix SK is positive semidenite, and its largest eigenvalue, (SK ), depends only
on the shape of the element K but not on its size hK . Obviously, (SK ) is independent
of the coecient k. If shapes of the elements in T are reasonably regular, (SK ) is of
similar size for all K T . Denote the maximum of (SK ) over T by
T = max (SK ).
KT
For any v U DG , let
|||v|||DG,K = k 2 v20,K +
1
1/2
2
h1
e We [[v]]0,e
eEK \EN
1770
for any K T . The so-called DG norm is dened as follows:

1/2
1
2
1
2
|||v|||DG = k 2 h v0, +
he We [[v]]0,e
.
eEI ED
Lemma 2.3 (uniqueness and coercivity).

(i) The bilinear form a1 (, ) is coercive in U DG with the coercivity constant
min{1, 1 }, provided that 1 > 0, i.e,
a1 (v, v) min{1, 1 }|||v|||2DG
(2.16)
v U DG .
Thus the NIPG problem in (2.15) has a unique solution, provided that 1 > 0.
e
e
(ii) Let w+
and w
be weights satisfying (2.1). Then SIPG and IIPG problems
(2.15) have a unique solution, provided that > 2(1 + )2 T . Moreover, the symmetric/incomplete bilinear form a (, ) for = 1 or 0 is coercive in U DG with a
coercivity constant 0 (0, 1) independent of the mesh size and the ratio kmax /kmin ,
i.e.,
a (v, v) 0 |||v|||2DG
(2.17)
v U DG ,
2(1+)2
for = 1 and 0, provided that > 10 T + 0 .

Proof.Let be a positive constant to be determined. For any v U DG and
for e EI ED , the CauchySchwarz inequality and the inequality of arithmetic and
geometric means give
(2.18)
he
We
2
{kv ne }w [[v]] ds
{kv ne }w 20,e +
[[v]]20,e .
We
he
e
eEI ED
eEI ED
eEI ED
e
e
For e EI , let e = K+
K
. Since kv is constant on each element,
he
{kv ne }w 20,e
We
e 2

(we )2 k
(w+ ) k+
2h2e
k+ (v + ne )2 +
k (v ne )2
We
We
e 2 e

e 2 e

) k
(w+ ) k+ (w
,
2h2e max
k+ (v + ne )2 + k (v ne )2 .
We
We
Similarly, for e ED and e K, we have
he
1 2
{kv ne }w 20,e = kK
he (kK v K ne )2 = h2e kK (vK ne )2 .
We
Summing up over all edges in EI ED and using (2.5) imply that
(2.19)
eEI ED
he
{kv n}w 2e 2
h2e kK (vK ne )2 .
We
KT eEK
It is proved in [3] that
T 2
h2e (v nK )2 = 4 vK
SK vK ,
eEK
1771
T
where vK is the vector of values of v at vertices of K. Since kK vK
SK vK = (kv, v)K ,
thus
T
h2e kK (vK ne )2 4kK vK S2K vK 4(SK )kK vK
SK vK = 4(SK ) (kv, v)K ,
eEK
which, together with (2.19), leads to
eEI ED
he
{kv n}w 2e 8T (kh v, h v).
We
Using (2.18), we now have

{kv ne }w [[v]] ds 4T (kh v, h v) +
eEI ED
eEI ED
We
[[v]]20,e .
2 he
Hence,
a (v, v) (1 4(1 + )T )(kh v, h v) +
eEI ED
For any constant 0 [0, 1), assume that >

> 0 such that
2(10 )
1+
>
>
4(1+)T
10
1 4(1 + )T > 0

1 + We
[[v]]20,e .

2
he
2(1+)2 T
10
+ 0 ; then there exists
, which is equivalent to
and
1+
> 0 .
2
This implies the coercivity of a (v, v) in (2.17) for any 0 (0, 1). When 0 = 0, it
yields that a (v, v) is positive and denite in U DG and, hence, problem (2.15) has a
unique solution. This completes the proof of the lemma.
Remark 2.4. The constant that appears in [5] is chosen to be greater than
(1 + )2 maxKT kK (SK ), which depends on k for = 1, and, hence, it is not
optimal.
3. A priori error estimate. Let e = uuT , where u and uT are the solutions of
(2.14) and (2.15), respectively. The dierence of (2.14) and (2.15) yields the following
error equation:
a (e, v) = 0 v U DG .
(3.1)
Let > 0 be a very small constant, and dene

vk,, = k 1/2 v, .
1+
1+
Let PT : Hg,D
() Ug be the orthogonal projection operator from Hg,D
() onto
Ug with respect to the inner product associated with the norm k,, . Then the
standard interpolation argument and an analysis similar to that for Proposition 2.4
1+
() H 1+s (T ) with s 1,
in [11] give that for Hg,D

(3.2)
1/2
( PT ), C
1/2
2(s)
hK
1/2
2s,K
KT
1772
where C is a positive constant independent of the mesh size and the ratio kmax /kmin .
For any v H 1+s (T ), 0 < s 1, denote
1/2
1/2

2(s)
1/2
2
2 1
2
Bs (h, v) =
hK
k vs,K
+
hK kK f 0,K
.
KT
KT
Lemma 3.1. Assume that the solution u V 1+ (T ) of problem (2.14) belongs to
H 1+s (T ) with 0 < s 1. Then

(3.3)
{k(PT u u) ne }w [[PT u uT ]] ds C Bs (h, u)|||PT u uT |||DG ,
eEI ED
Proof. Let z = PT u u and zT = PT u uT . By using the denition of the dual
norm, the triangle inequality, the inverse inequality, (2.5), Lemma 2.1, and (3.2), we
have

{k(PT u u) ne }w [[PT u uT ]] ds =
{kz ne }w [[zT ]] ds
eEI ED
eEI ED
{kz ne }w 1/2,e [[zT ]]1/2,e
eEI ED

1
e
e
w
k z|K ne 12 ,e + w+
k+ z|K+ ne 12 ,e he 2 [[zT ]]0,e

1
1/2
1/2
k z|K ne 12 ,e + k+ z|K+ ne 12 ,e he 2 We1/2 [[zT ]]0,e
eEI ED
eEI ED
eEI ED

1
2
1
2
h k z,K + hK k z0,K We2 he 2 [[zT ]]0,e
Ke
CBs (h, u)|||PT u uT |||DG .

This completes the proof of the lemma.
Theorem 3.2. Assume that the solution u V 1+ (T ) of problem (2.14) belongs
2(1+)2 T
to H 1+s (T ) H 1+ () with 0 < s 1 and that > 10

0, and 1. Then we have the following a priori error bound:
(3.4)
+ 0 for = +1,
|||u uT |||DG C Bs (h, u),
Proof. The triangle inequality gives
|||e|||DG |||u PT u|||DG + |||PT u uT |||DG .
Since u PT u is continuous and vanishes on D , thus
|||u PT u|||DG = k 1/2 (u PT u)0, k 1/2 (u PT u), .
Now, by (3.2) with = u it suces to show that
(3.5)
|||PT u uT |||DG C (|||u PT u|||DG + Bs (h, u)) .
1773
To this end, using the coercivity of a (, ) in (2.17), the error equation in (3.1),
the CauchySchwarz inequality, and the fact that [[PT u u]]EI ED = 0, we have
0 |||PT u uT |||2DG a (PT u uT , PT u uT ) = a (PT u u, PT u uT )

= (kh (PT u u), h (PT u uT )) +
h1
e We [[PT u u]][[PT u uT ]] ds
eEI ED
eEI ED
({k(PT u u) ne }w [[PT u uT ]]
{k(PT u uT ) ne }w [[PT u u]]) ds

C |||PT u u|||DG |||PT u uT |||DG +
{k(PT u u) ne }w [[PT u uT ]] ds,
eEI ED
which, together with Lemma 3.1, implies (3.5) and, hence, (3.4). This completes the
proof of the theorem.
4. Oswald- and Cl
ement-type interpolations. Denote by N , ND , and NK
the sets of all vertices of the triangulation T , on the D , and of element K T ,
respectively. For any z N , denote by z the nodal basis function of U, and let

z = K z : kK = max
kK .
z = {K T : K supp (z )} and

K z
The number of elements in

z is denoted by cd(z). Also, denote by EK the set of
edges that share at least a vertex with K.
In this section and sections 5 and 7, assume that the distribution of the coecients kK for all K T is locally quasi-monotone [31], which is slightly weaker than
Hypothesis 2.7 in [11]. For the convenience of the readers, we restate it here.
Definition 4.1. Given a vertex z N , the distribution of coecients kK ,
K z , is said to be quasi-monotone with respect to the vertex z if there exists a
subset
K,z,qm of z such that the union of elements in
K,z,qm is a Lipschitz domain
and that the following hold:
z
K,z,qm and kK maxK z kK ;
if z N \ND , then {K}
K,z,qm ,
K,z,qm D = , and kK maxK z kK .
if z ND , then K
The distribution of coecients kK , K T , is said to be locally quasi-monotone if it
is quasi-monotone with respect to every vertex z N .
For a given function v U DG , dene the Oswald interpolation operator I :
DG
U
Ug by
Iv =
Iv(z)z (x),
zN
where the nodal value of the interpolant Iv at z is dened by
gD (z)
if z ND ,
Iv(z) =
1
vK (z) if z N \ND
cd(z)
K
z
1774
with vK = v|K the restriction of v on K.

Lemma 4.2. Assume that the triangulation T is quasi-uniform. Then for any
v U DG , there exists a positive constant C independent of the ratio kmax /kmin such
that
he We [[v]]20,e +
he We v gD 20,e
(4.1) kK v Iv20,K C
eEK \ED
and
(4.2)
kK (v Iv)20,K C
1/2
eEK ED
eEK \ED
We
[[v]]20,e +
he
eEK ED
We
v gD 20,e
he
for all K T .
Proof. For any v U DG and any K T , the inverse inequality implies that
1/2
h2K kK (v Iv)20,K C kK v Iv20,K .

Hence, it suces to establish the validity of (4.1).
To this end, for any K T and any z NK , denote by z,K (x) = z (x)|K the
restriction of z in K. Then
vK (x) = v|K (x) =

vK (z)z,K (x) v U DG .
zNK
Since z,K 0,K C hK and

vK (z) Iv(z) = vK (z)
1
1
vK (z) =
(vK (z) vK (z))
cd(z)
cd(z)
K
z
K
z
for all z N \ND , we then have
(v Iv)(z)z,K 0,K
v Iv0,K =
zNK
(v Iv)(z)z,K 0,K C
zNK
hK |vK (z) Iv(z)|
zNK
zNK \ND
K
(4.3) C
hK |vK (z) vK (z)| +
hK |gD (z) vK (z)| .
zNK ND
For z NK \ND and any K

z , by the fact that the distribution of the
coecients of K is quasi-monotone with respect to z, there is a connected path,
{K = K0 , K1 , . . . , Kl = K }
with Ki z ,
from K to K such that the diusion coecient is monotonically increasing. Denote

by ei the common edge of Ki1 and Ki ; then we have
kK We1 Wel .
1775
Now, it follows from the triangle inequality and the inverse inequality that for any
K
z ,
1/2
1/2
kK |vK (z) vK (z)| kK
|vKi (z) vKi+1 (z)|
i=0
l
l
Wei [[v]],ei C
(4.4)
i=0

Wei
[[v]]0,ei .
hei
i=0
For z NK ND , by Denition 4.1, there exists a K

z such that EK ED = .
Let eD = EK ED = ; then the triangle inequality and an argument similar to that
above yield
1/2
1/2
1/2
kK |gD vK (z)| kK |gD vK (z)| + kK |vK (z) vK (z)|

l

WeD gD v,ei +
Wei [[v]],ei
i=0

(4.5)
WeD
gD v0,ei + C
heD
i=0

Wei
[[v]]0,ei .
hei
Now, (4.1) is a direct consequence of (4.3), (4.4), (4.5), and the CauchySchwarz
inequality. This completes the proof of the lemma.
Clement-type interpolation operators (see, e.g., [11, 31]) are often used for establishing the reliability bound of a posteriori error estimators. We dene a weighted
Clement-type interpolation operator and state its approximation and stability properties. For more details, see [14].
For a given function v, dene its weighted average over
z by

v z dx
.
(4.6)
v dx = z
dx
z z
Following [11, 31], dene the Clement-type interpolation operator J : L2 () U0
by
Jv =
(z v)z (x),
zN
where the nodal value at z is dened by

(J v)(z) = z v =
z N \ D ,
v dx,
z N D .
0,
1
(), the estimates
Lemma 4.3 (see [14]). For any K T and v H0,D
(4.7)
1/2
v J v0,K C hK kK
k 1/2 v0,K
and
(4.8)
1/2
(v J v)0,K C kK
k 1/2 v0,K
1776
hold, where K is the union of all elements that share at least one vertex with K.
1
For any e EI EN and v H0,D
(), the estimate
1/2
v J v0,e C h1/2
k 1/2 v0,e
e (We,1 )
(4.9)
holds, where e is the union of all elements that share at least one vertex with edge e.
1
Lemma 4.4 (see [14]). For any v H0,D
(), there exists a positive constant C
independent of the mesh size and the ratio kmax /kmin such that
|(f, v J v)| C Hf k 1/2 v0,
(4.10)
with
Hf =
1 2
kK
hK f 20,K +
zN (F ) Kz
1/2

1 2
kK
hK f f dx20,K .
z
zN \(F ) Kz
Remark 4.5 (see [16]). If f L2 (), the second term in Hf is o(maxKT hK ).

If f Lp () with p > 2, the same holds for the rst term.
5. Residual-based a posteriori error estimators. In this section, we study
the following residual-based a posteriori error estimator:

R =
1/2
2
R,K
KT
where the local indicator on K T is dened by

R,K = R2
f ,K
+ J2 ,K + J2u ,K + R2
D ,K
+ R2
1/2
N ,K
with
R2
R2
f ,K
D ,K
h2K f 20,K
, R2 ,K =
N
kK
eEK EN
eEK ED
J2 ,K =
1
2
he
g kuT n20,e ,
ke N
We
g uT 20,e ,
he D
eEK EI
he [[kuT n]]20,e

,
We,1
and J2u ,K =
eEK EI
We
[[uT ]]20,e .
he
The Rf ,K is the element residual, the RD ,K and RN ,K are the boundary residuals,
and the J ,K and Ju ,K are associated with edge jumps of the ux and the solution,
respectively. For k = 1, i.e., the Poisson equation, this residual-based estimator is
identical to that of [27]. For k being a tensor, Ern and Stephansen in [22] recently
studied a residual-based error estimator, which can be made robust for scalar k and
under the assumption of local monotonicity. Their estimator diers from the estimator
R,K in both the ux jump term and the solution jump term. Instead of Ju ,K , they
1777
use the so-called nonconforming error which is the energy norm of dierence between
the DG approximation and its continuous recovery through the Oswald interpolation.
To analyze the estimator R and recovery-based estimators to be introduced in
section 7, we employ a standard technique that uses the Helmholtz decomposition.
To this end, we cite the following decomposition (see, e.g., [3]): for any given vector1
() and q H such that
valued function L2 ()2 , there exist p H0,D
= k(x)p + q,
(5.1)
where H is a subspace of H 1 () having zero mean value and homogeneous tangential

derivatives on N :

v
= 0 on N .
H = v H 1 () :
v dx = 0 and
t
Integrating by parts gives

(p, q) = 0
(5.2)
1
for all p H0,D
() and all q H, which, in turn, implies that the decomposition is
orthogonal with respect to the weighted L2 inner product (k 1 , ):
(5.3)
(k 1 , ) = (kp, p) + (k 1 q, q).
1
Let e = u uT ; then there exist p H0,D
() and q H such that
kh e = kp + q
(5.4)
and
(5.5)
k 1/2 h e20, = k 1/2 p20, + k 1/2 q20, .
Denote the weighted oscillations of the data f over the collection T of elements by
h2
K
osc(f, T )2 =
f fK 20,K ,
kK

KT
where fK is the average of f over K.

Theorem 5.1 (reliability). Assume that u V 1+ (T ) is the solution of problem
(2.14) and that the triangulation is quasi-uniform. Then the residual-based a posteriori
error estimator R satises the following global reliability bound:
|||u uT |||DG C R ,
(5.6)
Proof. It follows from (5.2), the error equation in (3.1) with v = J p, and the
continuity of J p U0 that
k 1/2 p20, = (kh e, p)
= (kh e, (p J p)) +

{ke n}w [[J p]] {kJ p n}w [[e]]
eEI ED e
(5.7)
= (kh e, (p J p))
eEI ED

We

[[J p]][[e]] ds
he

{kJ p n}w [[e]] ds.
e
1778
For any e EI , using the CauchySchwarz inequality, (2.5), the trace theorem,
and the inverse inequality, we have

{kJ p n}w [[e]] ds = w+ (k+ J p n) [[e]] ds + w (k J p n) [[e]] ds
e
1/2
1/2
he w+ k+
1/2
We
1/2
k+ J p+
1/2
n0,e +
1/2
he w k
1/2
We
1/2
k J p
n0,e
1/2
We
1/2
he
[[uT ]]0,e
W 1/2

e
1/2
1/2
k+ J p0,K+ + k J p0,K
[[uT ]]0,e .
1/2
he
Thus, summing over all e EI and using the CauchySchwarz inequality lead to

(5.8)
eEI
{kJ p n}w [[e]] ds C
We
[[uT ]]20,e
he
1/2
k 1/2 p0, .
eEI
Similarly, we have

(5.9)
eED
{kJ p n}w [[e]] ds C
We
g uT 20,e
he D
1/2
k 1/2 p0, .
eED
Denote eJ = p J p. Using integration by parts, (1.1), (2.10), the facts that

(kuT ) = 0 on K T
1
and eJ H0,D
(),
(1.2), the CauchySchwarz inequality, and Lemma 4.3, we have

(5.10) (kh e, (p J p)) =
( (kh e), eJ ) +
(ke n) eJ ds
KT
KT
(f, eJ )
eEI
[[kuT n]] eJ ds +
f 0,K eJ 0,K +
KT
KT

eEN
(gN kuT n) eJ ds
[[kuT n]]0,e eJ 0,e
eEI
gN kuT n0,e eJ 0,e
eEN

C
hK f 0,K
1/2
kK
KT
eEN

C
KT
k 1/2 p0,K +
eEI

he
[[kuT n]]0,e k 1/2 p0,e
We,1
he
g kuT n0,e k 1/2 p0,e
ke N
1/2
2
(R
f ,K
+ J ,K + R
N ,K
k 1/2 p0, .
1779
Combining (5.7), (5.8), (5.9), and (5.11) yields

k 1/2 p0, C R .
Together with (5.5), then it suces to show
k 1/2 q0, C R .
(5.11)
To this end, since u H 1 (), IuT H 1 (), u IuT = 0 on D , and q H,

integrating by parts gives
((u IuT ), q) = 0.
This, together with the CauchySchwarz inequality and (4.2), implies that
k 1/2 q20, = (h e, q) =
((u uT ), q)K
KT
((IuT uT ), q)K
KT
k 1/2 (IuT uT )0,K k 1/2 q0,K
KT
12
k
1/2
(IuT uT
)20,K
k 1/2 q0,
KT

KT
J2u ,K + R2
1/2
,
D K
k 1/2 q0, .
Thus
k 1/2 q0, C
(5.12)

KT
J2u ,K + R2
1/2
,
D K
which proves the validity of (5.11) and, hence, the theorem.

Note that (5.4) indicates that the true error e = u uT comes from two kinds of
sources: discontinuous approximations of the normal component of the ux and the
tangential component of the gradient. As shown in the following lemma, the element
residual, the Neumann boundary residual, and the edge jump of the ux may be
bounded by the energy norm of p plus higher-order terms.
Lemma 5.2. There exists a positive constant C independent of the mesh size and
the ratio kmax /kmin such that

Rf ,K C k 1/2 p0,K + osc(f, K)
(5.13)
K T,

(5.14)

he
[[kuT n]]0,e C k 1/2 p0,e + osc(f, e )
We,1

(5.15)

he
gN kuT 0,e C k 1/2 p0,e + osc(f, e )
k
e EI ,
e EN ,
1780
where e is the collection of all elements that share the common edge e.
Proof. For any element K T , let bK be the standard cubic bubble function
whose support is K. Then (see, e.g., [37])
(5.16)
1/2
1/2
C hK bK 0,K ,
bK ,K C,
and bK ,K C h1
K .
Using (5.4) and (5.2), we have

(kp, v) = (kh e, v)
(5.17)
1
v H0,D
().
Choosing v = fK bK in (5.17) and integrating by parts lead to

(kp, (fK bK ))K = (kh e, (fK bK ))K = (f fK , fK bK )K + (fK , fK bK )K .
It follows from the CauchySchwarz inequality and (5.16) that
CfK 20,K (fK , fK bK )K = (kp, (fK bK ))K (f fK , fK bK )K
k 1/2 p0,K k 1/2 fK bK 0,K + f fK 0,K fK bK 0,K

1/2
kK
1/2
C
k p0,K fK 0,K + f fK 0,K fK 0,K .
hK
Hence,
hK
fK 0,K C
1/2
kK

k
1/2
p0,K
hK
+ 1/2 f fK 0,K
kK
Now, (5.13) is a direct consequence of the triangle inequality.

For any edge e EI , let be be the standard piecewise quadratic edge bubble
function corresponding to the edge e whose support is e . Then (see, e.g., [37])
(5.18)
C h1/2
b1/2
e
e 0,e ,
be ,e C,
and be ,e C h1
e .
Choosing v = [[kuT n]]e be in (5.17) and integrating by parts lead to

(kp, ([[kuT n]]e be ))e = (kh e, ([[kuT n]]e be ))e
2
= (f, [[kuT n]]e be )e + [[kuT n]] b1/2
e 0,e .
It follows from (5.18), the CauchySchwarz inequality, and (2.6) that

C [[kuT n]]20,e
2
[[kuT n]] b1/2
e 0,e
kp0,e [[kuT n]]e be 0,e + f 0,e [[kuT n]] be 0,e

1/2
1/2
1/2
k
,
C h1/2
[[ku
n]]
k
p
+
h
k
f

0,e
0,K
K
0,K
T
e
K
K
Ke
which, together with (2.6) and the triangle inequality, implies that

hK (fK 0,K + f fK 0,K )
he
1/2
[[kuT n]]0,e C k p0,e +
.
1/2
We,1
k
Ke
1781
Now, (5.14) follows from (5.13). For e EN , (5.15) may be proved in a similar fashion
by choosing v = (gN kuT n) be . This completes the proof of the lemma.
Theorem 5.3 (eciency). Assume that the diusion coecient is locally quasimonotone. Then there exists a positive constant C independent of the mesh size and
R,K C (|||e|||DG,K + osc(f, K ))
K T,
where K is the collection of elements in T that share a common edge with K.

Proof. The local eciency bound is a straightforward consequence of Lemma 5.2
and the denition of the DG norm.
6. Flux recovery. The ux dened by
= k(x)u
(6.1)
in
is an important physical quantity which is often the primary concern in practice. In

this section, we describe both implicit and explicit recoveries of the ux that are used
to design robust a posteriori error estimators in the subsequent section. In addition,
we show that the implicitly recovered ux is a good approximation to the ux.
6.1. Implicit approximation. The implicitly recovered ux is dened as follows: nd T VN such that
(k 1 T , ) = (h uT , ) VN ,
(6.2)
where VN is the subspace of V (RT0 or BDM1 ) satisfying the Neumann boundary

conditions
VN = { V : n = gN on N }.
Theorem
6.1. There exists a positive constant C independent of the ratio

kmax kmin such that the a priori error bound
(6.3) k
1/2

( T )0, C
inf k
1/2
VN

( )0, + k
1/2
(u h uT )0,
holds.
Proof. By using the error equation
(k 1 ( T ), ) = (h uT u, ) VN ,
(6.3) may be proved in a fashion similar to that of Theorem 3.1 in [14].
6.2. Explicit approximations. Let e e denote the Kronecker delta:

1 if e = e ,
e e =
0 if e = e .
For RT0 , its nodal basis function, e , corresponding to the edge e E is uniquely
determined by
e ne = e e
e E.
1782
T in RT0 = span{e : e E} by
Dene the explicit approximation
T =
e e ,
(6.4)
eE
T on the edge e E dened as a weighted

where
e is the normal component of
average of the normal components of the approximated ux kh uT ; i.e.,
e := {kh uT ne }ewi
(6.5)
with weights wi dened in (2.3). To ensure the eciency bound independent of the
size of jumps, we choose i = 2 or i = 3, i.e., the harmonic or the geometric weights.
Note that these weights satisfy the following inequality:

e
e
)2 (w,i
)2
(w+,i
1
(6.6)
max
,
i = 2, 3.
e
e
e
e ,
k
k+
k+ + k
7. Recovery-based a posteriori error estimators. For any element K T ,
based on the implicitly and explicitly recovered uxes, dene
,K = k 1/2 T + k 1/2 uT 0,K
T + k 1/2 uT 0,K .
and ,K = k 1/2
Obviously,
,K ,K .
(7.1)
Let
2 =
KT
2
,K
,
J2u =
KT
J2u ,K ,
and R2
KT
R2
D ,K
It is easy to see that

= k 1/2 T + k 1/2 h uT 0, = min k 1/2 + k 1/2 h uT 0, .
VN
Now, we dene the recovery-based a posteriori error estimators as follows:

1/2
= 2 + J2u + R2
D

1/2
and = 2 + J2u + R2
.
D
Theorem 7.1 (reliability). There exists a positive constant C independent of the

mesh size and the ratio kmax /kmin such that

(7.2)
|||e|||DG C ( + Hf ) C + Hf .
Proof. The second inequality in (7.2) follows directly from (7.1). To establish the
validity of the rst inequality in (7.2), by (5.5), (5.12), and the denition of the DG
norm, it suces to show that
(7.3)
k 1/2 p0, C ( + Hf ) .
Letting eJ = p J p, where J is the Clement-type interpolation operator dened in

section 4, it then follows from integration by parts, homogeneous boundary conditions
1783
of eJ on D and (ku + T ) n on N , the CauchySchwarz inequality, (4.8), the

fact that h (kh uT ) = 0, the inverse inequality, (4.7), and (4.10) that
(kh e, eJ ) = (ku + T , eJ ) ( T + kh uT , eJ )
(f T , eJ ) + C k 1/2 p0,
= (f, eJ ) (h ( T + kh uT ), eJ ) + C k 1/2 p0,
C ( + Hf ) k 1/2 p0, .
Combining with (5.7), (5.8), and (5.9) yields the rst inequality in (7.3). This completes the proof of the theorem.
Remark 7.2. A dierent explicit recovery-based estimator is derived in [21].
Its ux recovery achieves a tighter local connection between and f . Its reliability bound similar to (7.2) is established with C = 1 and Hf being replaced by a
superconvergent term if f H 1 ().
Lemma 7.3. For any element K T , let K be the union of elements sharing
a common edge with K. There exists a positive constant C independent of the mesh
size and the ratio kmax /kmin such that
1/2

K T.
(7.4)
,K C k 1/2 p20,K + osc (f, K )2
Proof. To show the validity of (7.4), by (5.14) and (5.15), it suces to prove that
for any element K T ,
(7.5)

he
he
2
2
2
g kuT ne 0,e .
[[kuT ne ]]0,e +
,K C
We,1
k N
eEK EI
eEK EN
We provide the proof of (7.5) only in the case that EK EN = because it can be
proved in a similar fashion in the case that EK EN = . To this end, for any edge
e EK , without loss of generality, let ne be the outward unit vector normal to K,
and denote by Ke the adjacent element with the common edge e. Since = kh uT
is piecewise constant, |K may be represented in terms of the nodal basis function of
RT0 , {e }eK , as follows:
e,K e .
|K =
eEK
For any x K, (6.4) and (6.5) give
e
T =
(
e e,K ) e =
(w+,i
1) (e,K e,Ke ) e
eEK
eEK
we, [[ ne ]] e .
eEK
Now, it follows from the triangle inequality, the fact that

that
1
2
T ) 20,K C kK
T 20,K
,K
= k 1/2 (

1 2
2
C
kK we, he [[ ne ]] ds C
eEK
eEK

K
|e |2 dx C, and (6.6)
he
[[kuT ne ]]20,e
We,1

,
1784
which proves (7.5) and, hence, (7.4). This completes the proof of the lemma.
Theorem 7.4 (eciency). Assume that the diusion coecient is locally quasimonotone. Then there exists a positive constant C independent of the mesh size and
(7.6)
C (|||e|||DG + osc (f, T )) .
Proof. Inequality (7.6) is a direct consequence of (7.1), (7.4), and the denition
of the DG norm.
8. Numerical experiment. In this section, we report some numerical results
for an interface problem with intersecting interfaces used by many authors, which
is considered as a benchmark test problem. For this test problem, we numerically
illustrate the discretization error of the DG method and demonstrate the robustness
of our error estimators.
To this end, let = (1, 1)2 and
u(r, ) = r ()
in the polar coordinates at the origin with () being a smooth function of (see,
e.g., [14]). The function u(r, ) satises the interface equation with A = kI, N = ,
f = 0, and

R
in (0, 1)2 (1, 0)2 ,
k(x) =
1
in \ ([0, 1]2 [1, 0]2 ).
Note that the solution u(r, ) is only in H 1+ () for any > 0 and, hence, it is very
singular for small at the origin. This suggests that renement is centered around
the origin. The depends on the size of the jump. For the test problem, we choose
= 0.1 which is corresponding to R 161.
For simplicity of presentation, the harmonic weighted incomplete interior penalty
Galerkin (IIPG) method is used. Let uT U DG be the discontinuous nite element
approximation of the solution. Denote by N the number of unknowns. We start
with the coarsest triangulation T0 obtained from halving 16 congruent squares by
connecting the bottom left and upper right corners.
Numerical results on uniform meshes are reported in Figure 1. The a priori error
estimate in (3.4) is of order h0.1 . This indicates that the slope of the log(dof)-log(error)
should be 0.05. Figure 1 shows that the asymptotic convergence rate for this test
problem is slightly better than the theoretical prediction.
Starting with the coarse triangulation T0 , a sequence of meshes is generated by
using a standard adaptive meshing algorithm that adopts the maximum marking
strategy: (1) mark those elements such that K 0.5 maxK T K and (2) rene the
marked triangles by bisection. The stopping criterion
rel-err :=
|||u uT |||DG
k 1/2 u0,
tol
is used, and numerical results with tol = 0.1 are reported.

Meshes generated by R and are depicted in Figures 2 and 4, respectively.
Renements are centered at origin as expected for ecient estimators. Meshes generated by various a posteriori error estimators for various types of discretizations of
1785

1
10
error
0.05
error and number of unknowns
10
10
10
10
10
10
10
10
10
number of unknowns
Fig. 1. Error and number of unknowns for uniform renements.

1
10
relerr and R / ||k1/2 u||0
relerr
1/2
R / ||k u||0
10
10
10
10
Fig. 2. Mesh generated by R .
10
10
number of unknowns
10
10
Fig. 3.
Relative error and estimator
R /k 1/2 u0, .
1
10
relerr
1/2
/ ||k u||0
relerr and / ||k1/2 u||0
10
10
10
10
10
10
10
10
number of unknowns
Fig. 4. Mesh generated by .
Fig. 5.
Relative error and estimator
/k 1/2 u0, .
this test problem can be found in [14, 15]. As shown in Figures 3 and 5, the slopes
of the log(dof)-log(relative error) for the estimators are close to 1/2 when there are
enough grid points (about several hundreds of unknowns). This implies the optimal
decay of the error with respect to the number of unknowns. While the eectivity
index, e-index := |||uu ||| , of R is about 4, that of is close to 1. This means
T
DG
that the estimator is very accurate.
1786
REFERENCES
[1] R. A. Adams, Sobolev Spaces, Academic Press, New York, 1975.
[2] M. Ainsworth, A synthesis of a posteriori error estimation technique for conforming, nonconforming and discontinuous Galerkin nite element methods, in Recent Advances in
Adaptive Computation, Contemp. Math. 383, AMS, Providence, RI, 2005, pp. 114.
[3] M. Ainsworth, A posteriori error estimation for discontinuous Galerkin nite element approximation, SIAM J. Numer. Anal., 45 (2007), pp. 17771798.
[4] M. Ainsworth and J. T. Oden, A Posteriori Error Estimation in Finite Element Analysis,
Pure Appl. Math. (N. Y.), Wiley-Interscience, New York, 2000.
[5] M. Ainsworth and R. Rankin, Fully computable error bounds for discontinuous Galerkin
nite element approximations on meshes with an arbitrary number of levels of hanging
nodes, SIAM J. Numer. Anal., 47 (2010), pp. 41124141.
[6] D. N. Arnold, An interior penalty nite element method with discontinuous elements, SIAM
J. Numer. Anal., 19 (1982), pp. 742760.
[7] D. N. Arnold, F. Brezzi, B. Cockburn, and L. D. Marini, Unied analysis of discontinuous
Galerkin methods for elliptic problems, SIAM J. Numer. Anal., 39 (2002), pp. 17491779.
[8] I. Babu
ska and T. Strouboulis, The Finite Element Method and Its Reliability, Numer.
Math. Sci. Comput., Oxford University Press, New York, 2001.
[9] R. Becker, P. Hansbo, and M. G. Larson, Energy norm a posteriori error estimation for
discontinuous Galerkin methods, Comput. Methods Appl. Mech. Engrg., 192 (2003), pp.
723733.
[10] C. Bernardi and F. Hecht, Error indicators for the mortar nite element discretization of
the Laplace equation, Math. Comp., 71 (2001), pp. 13711403.
rth, Adaptive nite element methods for elliptic equations with
[11] C. Bernardi and R. Verfu
non-smooth coecients, Numer. Math., 85 (2000), pp. 579608.
[12] F. Brezzi, J. Douglas, and L. D. Marini, Two families of mixed nite elements for second
order elliptic problems, Numer. Math., 47 (1985), pp. 217235.
[13] F. Brezzi and M. Fortin, Mixed and Hybrid Finite Element Methods, Springer-Verlag, New
York, 1991.
[14] Z. Cai and S. Zhang, Recovery-based error estimator for interface problems: Conforming
linear elements, SIAM J. Numer. Anal., 47 (2009), pp. 21322156.
[15] Z. Cai and S. Zhang, Recovery-based error estimator for interface problems: Mixed and
nonconforming elements, SIAM J. Numer. Anal., 48 (2010), pp. 3052.
rth, Edge residuals dominate a posteriori error estimates for
[16] C. Carstensen and R. Verfu
low order nite element methods, SIAM J. Numer. Anal., 36 (1999), pp. 15711587.
[17] P. G. Ciarlet, The Finite Element Method for Elliptic Problems, NorthHolland, New York,
1978.
[18] S. Cochez-Dhondt and S. Nicaise, Equilibrated error estimators for discontinuous Galerkin
methods, Numer. Methods Partial Dierential Equations, 24 (2008), pp. 12361252.
[19] A. Ern, S. Nicaise, and M. Vohralik, An accurate H(div) ux reconstruction for discontinuous Galerkin approximations of elliptic problems, C. R. Math. Acad. Sci. Paris, 345
(2007), pp. 709712.
[20] A. Ern and J. Proft, A posteriori discontinuous Galerkin error estimates for transient
convection-diusion equations, Appl. Math. Lett., 18 (2005), pp. 833841.
[21] A. Ern, A. Stephansen, and M. Vohralik, Improved Energy Norm A Posteriori Error
Estimation Based on Flux Reconstruction for Discontinuous Galerkin Methods, Technical
report, Ecole nationale des ponts et chaussees and Universite Pierre et Marie Curie, Paris,
2007.
[22] A. Ern and A. Stephansen, A posteriori energy-norm error estimates for advection-diusion
equations approximated by weighted interior penalty methods, J. Comput. Math., 26 (2008),
pp. 488510.
[23] A. Ern, A. Stephansen, and P. Zunino, A discontinuous Galerkin method with weighted
averages for advection-diusion equations with locally small and anisotropic diusivity,
IMA J. Numer. Anal., 29 (2009), pp. 235256.
[24] P. Grisvard, Elliptic Problems in Nonsmooth Domains, Monogr. Stud. Math. 24, Pitman,
Boston, 1985.
[25] J. S. Hesthaven and T. Warburton, Nodal Discontinuous Galerkin Methods: Algorithms,
Analysis, and Applications, Springer-Verlag, New York, 2008.
tzau, and T. P. Wihler, Energy norm a posteriori error estimation of
[26] P. Houston, D. Scho
hp-adaptive discontinuous Galerkin methods for elliptic problems, Math. Models Methods
Appl. Sci., 17 (2007), pp. 3362.
1787
[27] O. A. Karakashian and F. Pascal, A posteriori error estimates for a discontinuous Galerkin
approximation of second-order elliptic problems, SIAM J. Numer. Anal., 41 (2003), pp.
23742399.
[28] R. B. Kellogg, On the Poisson equation with intersecting interfaces, Appl. Anal., 4 (1975),
pp. 101129.
[29] J.-L. Lions and E. Magenes, Non-homogeneous Boundary Value Problems and Applications,
Vol. I, Springer-Verlag, New York, 1972.
ska, and C. Baumann, A discontinuous hp nite element method for diusion
[30] J. Oden, I. Babu
problems, J. Comput. Phys., 146 (1998), pp. 491519.
[31] M. Petzoldt, A posteriori error estimators for elliptic equations with discontinuous coecients, Adv. Comput. Math., 16 (2002), pp. 4775.
[32] P. A. Raviart and I. M. Thomas, A mixed nite element method for 2nd order elliptic
problems, in Mathematical Aspects of Finite Element Methods, Lecture Notes in Math.
606, Springer-Verlag, Berlin, New York, 1977, pp. 292315.
[33] B. Rivi`
ere, Discontinuous Galerkin Methods for Solving Elliptic and Parabolic Equations:
Theory and Implementation, SIAM, Philadelphia, 2008.
[34] B. Rivi`
ere and M. F. Wheeler, A posteriori error estimates for a discontinuous Galerkin
method applied to elliptic problems, Comput. Math. Appl., 46 (2003), pp. 141163.
[35] A. Romkes, S. Prudhomme, and J. T. Oden, A posteriori error estimation for a new stabilized
discontinuous Galerkin method, Appl. Math. Lett., 16 (2003), pp. 447452.
[36] S. Sun and M. F. Wheeler, L2 (H 1 )-norm a posteriori error estimation for discontinuous
Galerkin approximations of reactive transport problems, J. Sci. Comput., 22 (2005), pp.
501530.
rth, A Review of A Posteriori Error Estimation and Adaptive Mesh-Renement
[37] R. Verfu
Techniques, Wiley-Teubner, Stuttgart, 1996.

Discontinuous Galerkin Finite Element Methods For Interface Problems: A Priori and A Posteriori Error Estimations

Încărcat de

Informații document

Descriere originală:

Titlu original

Drepturi de autor

Formate disponibile

Partajați acest document

Partajați sau inserați document

Opțiuni de partajare

Vi se pare util acest document?

Este necorespunzător acest conținut?

Drepturi de autor:

Formate disponibile

Discontinuous Galerkin Finite Element Methods For Interface Problems: A Priori and A Posteriori Error Estimations

Încărcat de

Drepturi de autor:

Formate disponibile

c 2011 Society for Industrial and Applied Mathematics

SIAM J. NUMER. ANAL.

DISCONTINUOUS GALERKIN FINITE ELEMENT METHODS FOR

1. Introduction. Consider the following interface problem:

with boundary conditions

where f , gD , and gN are given scalar-valued functions; is a bounded polygonal

for i = 1, . . . , n. Here, {i }ni=1 is a partition of the domain with i being an open

Copyright by SIAM. Unauthorized reproduction of this article is prohibited.

ZHIQIANG CAI, XIU YE, AND SHUN ZHANG

and kmax = max ki .

Copyright by SIAM. Unauthorized reproduction of this article is prohibited.

DG METHODS FOR INTERFACE PROBLEMS

dened in an appropriate solution space V (see (2.9)) that permits discontinuity

Copyright by SIAM. Unauthorized reproduction of this article is prohibited.

ZHIQIANG CAI, XIU YE, AND SHUN ZHANG

For a scalar-valued function v, dene the operator by

We shall use the Hilbert space

Copyright by SIAM. Unauthorized reproduction of this article is prohibited.

DG METHODS FOR INTERFACE PROBLEMS

respectively, where x = (x1 , x2 ). Then the standard H(div; ) conforming RT and

and dene the weighted averages

Copyright by SIAM. Unauthorized reproduction of this article is prohibited.

ZHIQIANG CAI, XIU YE, AND SHUN ZHANG

the same argument shows that

for any weights satisfying (2.1). Letting

it is easy to check that

2.3. DG finite element approximation. To describe the DG nite element

Copyright by SIAM. Unauthorized reproduction of this article is prohibited.

DG METHODS FOR INTERFACE PROBLEMS

implies (2.8). This completes the proof of the lemma.

Copyright by SIAM. Unauthorized reproduction of this article is prohibited.

ZHIQIANG CAI, XIU YE, AND SHUN ZHANG

{ku ne }w [[v]] ds = (f, v) +

for all v V 1+ (T ). By the continuity of u H 1+ () and the Dirichlet boundary

for all v V 1+ (T ), where > 0 is a parameter to be determined.

and a linear form for v V 1+ (T ) by

Copyright by SIAM. Unauthorized reproduction of this article is prohibited.

DG METHODS FOR INTERFACE PROBLEMS

Methods corresponding to = 1, 0, or 1 are the so-called symmetric interior penalty

For any v U DG , let

Copyright by SIAM. Unauthorized reproduction of this article is prohibited.

ZHIQIANG CAI, XIU YE, AND SHUN ZHANG

for any K T . The so-called DG norm is dened as follows:

Lemma 2.3 (uniqueness and coercivity).

for = 1 and 0, provided that > 10 T + 0 .

It is proved in [3] that

Copyright by SIAM. Unauthorized reproduction of this article is prohibited.

DG METHODS FOR INTERFACE PROBLEMS

which, together with (2.19), leads to

Using (2.18), we now have

For any constant 0 [0, 1), assume that >

+ 0 ; then there exists

Let > 0 be a very small constant, and dene

Copyright by SIAM. Unauthorized reproduction of this article is prohibited.

ZHIQIANG CAI, XIU YE, AND SHUN ZHANG

{kz ne }w 1/2,e [[zT ]] 1/2,e

h k z ,K + hK k z 0,K We2 he 2 [[zT ]] 0,e

CBs (h, u)|||PT u uT |||DG .

to H 1+s (T ) H 1+ () with 0 < s 1 and that > 10

|||u uT |||DG C Bs (h, u),

|||PT u uT |||DG C (|||u PT u|||DG + Bs (h, u)) .

Copyright by SIAM. Unauthorized reproduction of this article is prohibited.

{kz ne }w 1/2,e [[zT ]]1/2,e

h k z,K + hK k z0,K We2 he 2 [[zT ]]0,e

h2K kK (v Iv)20,K C kK v Iv20,K .

Wei [[v]],ei C

he [[kuT n]]20,e

k 1/2 h e20, = k 1/2 p20, + k 1/2 q20, .

[[kuT n]]0,e eJ 0,e

gN kuT n0,e eJ 0,e

k 1/2 (IuT uT )0,K k 1/2 q0,K

kp0,e [[kuT n]]e be 0,e + f 0,e [[kuT n]] be 0,e