Numer. Math.

(2008) 108:487–500 Numerische

DOI 10.1007/s00211-007-0121-y Mathematik

A posteriori error estimates for mixed finite element

approximations of elliptic problems

Mats G. Larson · Axel Målqvist

Received: 25 January 2005 / Revised: 25 September 2007 / Published online: 27 November 2007
© Springer-Verlag 2007

Abstract We derive residual based a posteriori error estimates of the flux in L 2 -norm
for a general class of mixed methods for elliptic problems. The estimate is applicable to
standard mixed methods such as the Raviart–Thomas–Nedelec and Brezzi–Douglas–
Marini elements, as well as stabilized methods such as the Galerkin-Least squares
method. The element residual in the estimate employs an elementwise computable
postprocessed approximation of the displacement which gives optimal order.

Mathematics Subject Classification (2000) 65N30 · 65N15 · 65N12

1 Introduction

The model problem We consider the mixed formulation of the Poisson equation
with Neumann boundary conditions:
⎧
⎨ σ − ∇u = 0 in
,
−∇ · σ = f in
, (1.1)
⎩
n · σ = 0 on ,

where

is a polygonal domain in Rn , n = 2 or 3 with boundary . Assuming

f d x = 0, we get a well posed problem with a solution u ∈ H (
)/R and
1

M. G. Larson (B)
Department of Mathematics,
Umeå University, 90187 Umeå, Sweden
e-mail: mats.larson@math.umu.se

A. Målqvist
Division of Scientific Computing, Department of Information Technology,
Uppsala University, 75105 Uppsala, Sweden
e-mail: axel.malqvist@it.uu.se

123
488 M. G. Larson, A. Målqvist

σ ∈ V = {v ∈ H (div,
) : n · v = 0 on }. See [9] for definitions of these function

spaces.

Previous work Several works present a posteriori error estimates for mixed methods.
In Carstensen [10] an error estimate in the H (div,
) norm of the flux is presented. The
H (div,
) norm may be dominated by the div-part which is directly computable. When
it comes to error estimates of the flux in L 2 norm of methods using richer spaces for the
flux σ than the displacement u, such as Raviart–Thomas–Nedelec (RTN) elements,
there are known difficulties. Braess and Verfürth presents a suboptimal estimate in
[4]. The reason for the suboptimality is that the natural residual that arises from the
first equation σ − ∇u = 0 in problem (1.1) may be large if the flux space is richer
than the displacement space. In a recent paper Lovadina and Stenberg [12] derive an
a posteriori error estimate of the L 2 -norm of the flux for the RTN based methods
which employs a particular postprocessed approximation U . The proof is based on
a posteriori error analysis of an equivalent method which involves the postprocessed
approximation U .

New contributions We derive a general a posteriori error estimate in the energy

norm which is applicable to most mixed methods including the classical inf-sup
stable elements, RTN elements and the Brezzi–Douglas–Marini (BDM) elements. Our
estimate is closely related to the estimate presented by Lovadina and Stenberg [12],
however, our proof is more general and also reveals the fact that one can use any pie-
cewise polynomial approximation of the displacement when computing the residual.
By a small adjustment of the argument we finally, derive an estimate for the stabili-
zed mixed method of Masud and Hughes [13]. The same technique applies to other
stabilized schemes, for instance the Galerkin least squares method.

Outline We start by presenting finite elements and the discrete version of equation
(1.1) in Sect. 2 then we present the a posteriori error estimates in Sect. 3.

2 Weak formulation and the finite element method

Weak formulation We multiply the first equation in (1.1) by a test function v ∈ V

and integrate by parts. The second equation in (1.1) is multiplied by a test function
w ∈ W = L 2 (
)/R. The weak form reads: find σ ∈ V and u ∈ W such that,

(σ , v) + (u, ∇ · v) = 0 for all v ∈ V ,
(2.1)
(−∇ · σ , w) = ( f, w) for all w ∈ W.

Our aim is to derive a posteriori error estimates of finite element approximations

{, U } of the exact solution {σ , u} in the energy norm σ − 0 , where  · 0 denotes
the L 2 (
) norm.

The mixed finite element method We let K = {K } be a partition of
into simplicial

elements of diameter h K and define the mesh function, h(x) :
→ R+ , by letting

123
A posteriori error estimates for mixed finite element approximations of elliptic problems 489

h(x) = h K for x ∈ K . We assume that the elements are shape regular, i.e., there is
a constant C such that h K /ρ K ≤ C for all K ∈ K, where ρ K is the diameter of the
largest ball that can be inscribed in K . We seek an approximate solution in discrete
spaces V h ⊂ V and Wh ⊂ W defined on the partition K. It is well known that for
finite element methods based on the standard weak form (2.1) the discrete spaces
must be chosen so that the inf-sup condition, see [9], is satisfied in order to guarantee
a stable method. Only rather special constructions of the discrete spaces yield stable
methods. In Sect. 3.4 we consider a stabilized mixed finite element method based on
a modified weak formulation which can be based on standard continuous piecewise
polynomials. We summarize some of the most well known choices of stable discrete
spaces on triangles and tetrahedra for a given integer k ≥ 1:
• Raviart–Thomas–Nedelec (RTN) elements, see [14,15],
V h = {v ∈ H (div,
) : v| K ∈ [Pk−1 (K )]n ⊕ x P̃k−1 (K ) for all K ∈ K},
Wh = {w ∈ L 2 (
)/R : w| K ∈ Pk−1 (K ) for all K ∈ K}.
• Brezzi–Douglas–Marini (BDM) elements, see [7,8],
V h = {v ∈ H (div,
) : v| K ∈ [Pk (K )]n for all K ∈ K},
Wh = {w ∈ L 2 (
)/R : w| K ∈ Pk−1 (K ) for all K ∈ K}.
Here C(
) denotes the space of continuous functions on
, Pk (K ) the space of
polynomials of degree k on element K , and P̃k (K ) the set of homogeneous polynomials
of degree k. For a more complete account of these spaces and the inf-sup condition
we refer to Brezzi–Fortin, [9]. Note, that our a posteriori error analysis does not use
the inf-sup condition explicitly.
The mixed finite element method reads: find  ∈ V h and U ∈ Wh such that:

(, v) + (U, ∇ · v) = 0 for all v ∈ V h ,
(2.2)
(−∇ · , w) = ( f, w) for all w ∈ Wh .

3 A posteriori error estimates

3.1 Preliminaries

We use the following notation for the standard Sobolev norms,  · s,ω =  ·  H s (ω) =
 · W2s (ω) , see [1], and we let (·, ·)ω denote the L 2 (ω) inner product. In the case ω =

we simplify the notation and write  · s,
=  · s and (·, ·)
= (·, ·). We shall also
need suitable norms, see [18], on the element Sobolev spaces H 1 (K ), H 1/2 (∂ K ), and
H −1/2 (∂ K ). For each element K ∈ K we define the following norms

|v|21,K = ∇v20,K + h −2
K v0,K for all v ∈ H (K ),
2 1
(3.1)

|v|21/2,∂ K = inf |η|21,K for all v ∈ H 1/2 (∂ K ), (3.2)

η ∈ H 1 (K )
η = v on ∂ K

v, w ∂ K
|v|−1/2,∂ K = sup for all v ∈ H −1/2 (∂ K ), (3.3)
w∈H 1/2 (∂ K ) |w|1/2,∂ K

123
490 M. G. Larson, A. Målqvist

where ·, · ∂ K denotes the duality pairing between spaces H −1/2 (∂ K ) and H 1/2 (∂ K ).
With these definitions we clearly have the inequality

v, w ≤ |v|−1/2,∂ K |w|1/2,∂ K for all v ∈ H −1/2 (∂ K ) and w ∈ H 1/2 (∂ K ).

∂K
(3.4)
Furthermore, we have the following elementwise normal trace inequality.
Lemma 3.1 The following trace inequality holds

|n · v|2−1/2,∂ K ≤ C(v20,K + h 2K ∇ · v20,K ), (3.5)

for all v ∈ H (div, K ) with constant C independent of h K .

Proof To prove (3.5) we first derive the estimate on the reference element and then
 be the
use a scaling argument to show uniformity in the size of the element h K . Let K
 and exterior unit normal 
reference element with boundary ∂ K n. To prove the estimate
on the reference element we employ Green’s formula, see [11], to get the identity


n ·
v, w
 
∂K= (v, ∇ w
) K + (∇ · ) K for all 
v, w ) and w
v ∈ H (div, K ).
 ∈ H 1(K
(3.6)
Estimating the right hand side using Cauchy-Schwartz inequality we have


n ·
v, w
 
∂K v2K + h 2K∇ · 
≤ ( v2K)1/2 |
w| K, (3.7)

where we divided and multiplied by a scaling factor of h −2

 . Since (3.7) holds for all
K
) we get, for any 
 ∈ H 1(K
w ),
µ ∈ H 1/2 (∂ K


n ·
v, 
µ 
∂K v2K + h 2K∇ · 
≤ ( v2K)1/2 inf |
w| K (3.8)
w )
 ∈ H 1(K
w
= 
µ on ∂ K

v2K + h 2K∇ · 
= ( v2K)1/2 |
µ|1/2,∂ K. (3.9)

Dividing by |µ|1/2,∂ K and taking the supremum over all  ) we obtain

µ ∈ H 1/2 (∂ K
the trace inequality (3.5) on the reference element K .
Given an element K ∈ K, let F : Rn → Rn be an affine mapping of the form
Fx = b+B  x, where B is an invertible n×n matrix with positive determinant, b ∈ Rn ,
and K = F K . Furthermore, let B = sup x∈Rn B x/x be the standard Euclidian
matrix norm. We then have the following bounds B ≤ h K /ρ K and B −1  ≤ h K/ρ K
and thus B B −1  ≤ C as a consequence of shape regularity.
Next we define a mapping F : H 1 ( K ) → H 1 (K ) by v = F v =v ◦ F −1 . We then
v |m,K ≤ C(det B)
have the estimate | −1/2 B |v|m,K , for m = 0, 1, see [9], and thus
m

we conclude that
v |1, K ≤ C(det B)−1/2 B |v|1,K .
| (3.10)
Taking the appropriate infimum on both sides of (3.10) we obtain

µ|1/2,∂ K ≤ C(det B)−1/2 B |µ|1/2,∂ K .

| (3.11)

123
A posteriori error estimates for mixed finite element approximations of elliptic problems 491

) →
To transform vector valued functions we use the Piola transform P : H (div, K
H (div, K ) defined by P v = (det B) B−1 −1
v ◦ F . We then have the identity
n · v, w ∂ K =  n ·v, w
 ∂ K for v = P v and w = F w . Furthermore, we have
the estimates

v K ≤ (det B)1/2 B −1  v K , ∇ · 

 v K ≤ (det B)1/2 ∇ · v K , (3.12)

see [9]. Using these results we obtain the following estimate

n · v, η ∂ K
|n · v|−1/2,∂ K = sup (3.13)
η∈H 1/2 (∂ K ) |η|1/2,∂ K

n ·
v, 
η ∂ K
≤ C(det B)−1/2 B sup (3.14)
 )
η∈H 1/2 (∂ K |
η|1/2,∂ K
≤ C(det B)−1/2 B|
n ·
v|−1/2,∂ K. (3.15)

Finally, we can prove the desired estimate (3.5) on element K as follows

|n · v|−1/2,∂ K ≤ C(det B)−1/2 B|

n ·
v|−1/2,∂ K (3.16)
−1/2
≤ C(det B) B(
v2K + h 2K∇ ·
v2K)1/2 (3.17)
−1/2 −1
≤ C(det B) B(det B) B  1/2
 1/2
× v2K + (h KB)2 ∇ · v2K (3.18)
≤ C(v2K + h 2K ∇ · v2K )1/2 , (3.19)

where we used: (3.15) in (3.16), the trace inequality (3.5) on the reference element
K in (3.17), the estimates for the Piola transform (3.12) in (3.18) together with the
observation that ∇·v ≤ B −1  B∇·v since 1 = I  ≤ B −1  B, and finally
in (3.19) we used the estimate B ≤ h K /ρ K to conclude that h K B ≤ (h K/ρ K)h K .

3.2 Estimate for standard mixed methods

Here we present a general a posteriori error estimate in the energy norm σ − 0
involving a piecewise polynomial function Q, which may be obtained by postproces-
sing U . The possibility to replace U by Q is important since it leads to a posteriori
error estimates of optimal order. We are not interested in tracking the constants in the
error estimates.

Theorem 3.1 For arbitrary Q ∈ K ∈K Pl (K ), with l ≥ 0 and f ∈ L 2 (
) it holds,

 
σ − 20 ≤ C h 2K  f + ∇ · 20,K +  − ∇ Q20,K + h −1
K  [Q] 0,∂ K ,
2

K ∈K
(3.20)

123
492 M. G. Larson, A. Målqvist

where the jump denoted by [·] is the difference in function value over a face in the
mesh.
Proof Starting with the left hand side we have

σ − 20 = (σ − , σ − ) (3.21)
= (σ , σ − ) − (, σ − ) (3.22)
= −(u, ∇ · (σ − )) − (, σ − ) (3.23)
= −(u − Q, ∇ · (σ − )) + (Q, −∇ · (σ − )) − (, σ − ) (3.24)

= (u − Q, f + ∇ · ) + ((Q, −∇ · (σ − )) K − (, σ − ) K )

K ∈K
(3.25)
= I + I I. (3.26)

We treat the two terms in equation (3.26) separately, beginning with I . From the second
part of equation (2.2) we have the Galerkin orthogonality property ( f + ∇ · , w) = 0
for all w ∈ Wh . Let Ph denote the L 2 projection onto Wh . Using Galerkin orthogonality
(2.2) to subtract the projection Ph (u − Q) ∈ Wh of (u − Q) followed by the projection
error estimate v − Ph v0,K ≤ Ch K ∇v0,K we obtain

I ≤ |( f + ∇ · , u − Q)| (3.27)
−1
≤ h( f + ∇ · )0 h (u − Q − Ph (u − Q))0 (3.28)
≤ Ch( f + ∇ · )0 ∇(u − Q)0,K (3.29)
= Ch( f + ∇ · )0 σ −  +  − ∇ Q0,K (3.30)
3C 2 1 1
≤ h( f + ∇ · )20 + σ − 20 +  − ∇ Q20,K . (3.31)
2 4 2

Here, and below,  · 0,K denotes the broken L 2 -norm defined by v20,K =
K ∈K v0,K . We now turn to the second term I I in equation (3.26) and start with
2

integration by parts,

II = ((Q, −∇ · (σ − )) K − (, σ − ) K ) (3.32)

K ∈K

= ((∇ Q, σ − ) K − (Q, n · (σ − ))∂ K − (, σ − ) K ) (3.33)

K ∈K

= (∇ Q − , σ − ) − Q, n · (σ − ) ∂K (3.34)
K ∈K

1
≤ ∇ Q − 20,K + σ − 20 + Q, n · (σ − ) ∂K . (3.35)
4
K ∈K

123
A posteriori error estimates for mixed finite element approximations of elliptic problems 493

Next we note that K ∈K v, n · (σ − ) ∂ K = v, n · (σ − ) ∂
= 0 for all

v ∈ H 1 (
), since σ −  ∈ V . This identity follows from using Green’s formula and
decomposing the integral over
into a sum of integrals over the elements and then
using Green’s formula again. Thus we can subtract an arbitrary function v ∈ H 1 (
)
in the term K ∈K Q, n · (σ − ) ∂ K = K ∈K Q − v, n · (σ − ) ∂ K . We then
have the estimate

1
I I ≤ ∇ Q −20,K + σ −20 + inf Q − v, n · (σ − ) ∂K . (3.36)
4 v∈H 1 (
)
K ∈K

We now use inequality (3.4) followed by the trace inequality (3.5), to estimate the sum
in equation (3.36) as follows

inf Q − v, n · (σ − ) ∂K (3.37)
v∈H 1 (
)
K ∈K

≤ inf |Q − v|1/2,∂ K |n · (σ − )|−1/2,∂ K (3.38)

v∈H 1 (
)
K ∈K
1/2 1/2
≤ inf |Q − v|21/2,∂ K |n · (σ − )|2−1/2,∂ K (3.39)
v∈H 1 (
)
K ∈K K ∈K
1/2
≤C inf |Q − v|21/2,∂ K
v∈H 1 (
)
K ∈K
1/2
 
× σ − 20,K + h 2K ∇ · (σ − )20,K (3.40)
K ∈K
1/2
 1/2
≤C inf |Q − v|21/2,∂ K σ − 20 + h( f + ∇ · )20
v∈H 1 (
)
K ∈K
(3.41)
3C 2 1 1
≤ inf |Q − v|21/2,∂ K + σ − 20 + h( f + ∇ · )20 .
2 v∈H 1 (
) 4 2
K ∈K
(3.42)

Together equation (3.36) and equations (3.37–3.42) give a bound of the second term,
I I , in equation (3.26),

1 1
I I ≤ ∇ Q − 20,K + σ − 20 + h( f + ∇ · )20
2 2
3C 2
+ inf |Q − v|21/2,∂ K . (3.43)
2 v∈H 1 (
)
K ∈K

123
494 M. G. Larson, A. Målqvist

We combine equation (3.31) and equation (3.43) to get,

3 3
I + II ≤ ∇ Q − 20,K + σ − 20 + Ch( f + ∇ · )20
2 4
3C 2
+ inf |Q − v|21/2,∂ K . (3.44)
2 v∈H 1 (
)
K ∈K

To estimate the last term on the right hand side in equation (3.44) we employ the tech-
nique of Lemma 4 in [3]. For completeness we include the details of the proof. We let N
be the set of nodes in the mesh, {φ}i∈N be the lowest order Lagrange basis functions,
ωi = supp(φi ), C Pl,i = {v ∈ C(ωi ) : v| K ∈ Pl (K ), for all K ∈ K with K ⊂ ωi },
i.e., continuous piecewise polynomials of degree l on ωi , and C Pl = ⊕i∈N φi C Pl,i ⊂
H 1 (
).
Using that C Pl ⊂ H 1 (
) followed by the inverse inequality |Q − v|21/2,∂ K ≤
Ch −1
K Q − v0,∂ K , which holds since both v and Q are piecewise polynomials, we
2

get

inf |Q − v|21/2,∂ K ≤ inf |Q − v|21/2,∂ K

v∈H 1 (
) v∈C Pl
K ∈K K ∈K

≤ C inf h −1
K Q − v0,∂ K .
2
(3.45)
v∈C Pl
K ∈K

We write v = i∈N φi vi ∈ C Pl and proceed with the estimate as follows

Q − v20,∂ K = (Q − v, φi (Q − vi ))∂ K , (3.46)

i∈N
1/2 1/2
≤ φi (Q − v)0,∂ K φi (Q − vi )0,∂ K , (3.47)
i∈N
1/2
1/2
≤ Q − v0,∂ K φi (Q − vi )20,∂ K , (3.48)
i∈N

where we used that {φi }i∈N is a partition of unity. Dividing inequality (3.48) by
Q − v0,∂ K and combining with estimate (3.45) we arrive at

h −1
1/2
inf |Q − v|21/2,∂ K ≤ inf K φi (Q − vi )20,∂ K . (3.49)
v∈H 1 (
) vi ∈C Pl,i
K ∈K i∈N K ∈K

Next we employ the following inequality

h −1 h −1
1/2 1/2
inf
vi ∈C Pl,i K φi (Q − vi )20,∂ K ≤ C K φi [Q] 20,∂ K . (3.50)
K ∈K K ∈K

123
A posteriori error estimates for mixed finite element approximations of elliptic problems 495

We return to the proof of this inequality below. Together (3.49) and (3.50) give

h −1
1/2
inf |Q − v|21/2,∂ K ≤ C K φi [Q] 20,∂ K
v∈H 1 (
)
K ∈K K ∈K i∈N

=C h −1
K  [Q] 0,∂ K ,
2
(3.51)
K ∈K

again we used that {φi }i∈N is a partition of unity.

Combining equation (3.43) and (3.51) we get

3 3
I +I I ≤ ∇ Q −20,K + σ −20 +Ch( f +∇ ·)20 +C h −1
K  [Q] 0,∂ K .
2
2 4
K ∈K
(3.52)
Since I + I I = σ − 20 from equations (3.21–3.26) we just need to subtract
3/4σ − 20 from both sides of equation (3.52) to prove the theorem.
It remains to prove inequality (3.50). We first note that it follows from shape regu-
larity that there are constants c and C such that c ≤ h K / h K  ≤ C for all elements
K , K  ∈ K such that K , K  ⊂ ωi . We may therefore conclude that there is an h i
and constants c and C such that ch i ≤ h K ≤ Ch i for all K ∈ K with K ⊂ ωi .
Next we observe that it also follows from shape regularity that there is only a finite
number, say Nc , of possible element configurations in ωi . Let  ω j , j = 1, . . . , Nc
be the corresponding reference configurations and let j (i) denote the index of the
reference configuration corresponding to patch ωi . Let Fωi :  ω j (i) → ωi be a
C 0 -diffeomorphism such that Fωi | K is affine. Let E denote the set of edges or faces in
the mesh. For each E ∈ E, E ⊂ ∂ K , K ⊂ ωi we let Fωi ,E = Fωi | E with correspon-
ding Jacobian Bωi ,E . Using shape regularity it follows that there are constants c and
C such that ch in−1 ≤ detBωi ,E ≤ Ch in−1 (where n is the spatial dimension).
With these preparations accomplished we begin with the estimate as follows

h −1 h −1
1/2 1/2
K φi (Q − vi )20,∂ K = K φi (Q − vi )20,E (3.53)
K ∈K K ∈K E∈E ,E⊂∂ K

= h −1 detBωi ,E φ −
1/2 ( Q vi )20, E
K i
K ∈K E∈E ,E⊂∂ K
(3.54)
≤ Ch in−2 φ −
1/2 ( Q vi )20, E, (3.55)
i
K ∈K E∈E ,E⊂∂ K

where we splitted the integral over the element boundaries to edge or face contributions
in (3.53), mapped to the reference configuration in (3.54), and used the estimates
h −1 −1
K ≤ Ch i and detBωi ,E ≤ Ch i
n−1
in (3.55).

123
496 M. G. Larson, A. Målqvist

Next we have the following inequality

inf 
φ
1/2 −
(Q vi )20, E ≤ C 
φ
1/2  2 ,
[ Q]
 i i 0, E

vi ∈C P l,i
K ∈K E∈E ,E⊂∂ K K ∈K E∈E ,E⊂∂ K
(3.56)
∈
for all Q K ∈K,K ⊂ωi P l ( 
K ) and i ∈ N . To prove (3.56) we observe that if the
right hand side is zero then Q  is continuous on ωi and thus we can take 
vi = Q  and
therefore the left hand side is also zero and then the inequality follows by using finite
dimensionality of the discrete space K ∈K,K ⊂ωi Pl ( K ) together with the fact that
there are only a finite number Nc of reference configurations.
Using (3.53–3.55) together with (3.56) and mapping back to the actual configuration
we obtain

h −1
1/2
(Q − vi )20,∂ K ≤ Ch in−2  1/2  2
K φi φi [ Q]0, E
(3.57)
K ∈K K ∈K E∈E ,E⊂∂ K

(detBωi ,E )−1 φi

1/2
≤ Ch in−2 [Q]20,E
K ∈K E∈E ,E⊂∂ K
(3.58)
h −1
1/2
≤C K φi [Q]20,E , (3.59)
K ∈K E∈E ,E⊂∂ K

where we finally used the estimates (detBωi ,E )−1 ≤ Ch i−(n−1) and h i ≤ Ch K in

(3.59). This estimate concludes the proof of (3.50).

3.3 Estimate based on postprocessing

We now turn to the question of how to choose Q in Theorem 3.1. We know that
choosing Q = U results in a suboptimal estimate of the energy norm error, [4]. A
natural idea is to choose Q to be a postprocessed version of U . There have been several
works [5,8,12,16,17] following Arnold and Brezzi [2], published in the mid eighties,
on postprocessing methods where information from the calculated flux  is used to
compute an improved approximation of u.
We focus on the method considered in Lovadina and Stenberg [12] and show that
Theorem 3.1 directly gives the estimate presented in [12]. We denote the postprocessed
version of U by U ∗ . To define U ∗ we introduce some notations. For all K ∈ K we let
Ph,K : L 2 (
) → Wh,K be the L 2 projection onto Wh,K , where Wh,K is the restriction
of Wh onto K . Furthermore, we let Wh,K∗ denote the following spaces: W ∗ = P (K )
h,K k
∗
for RTN elements and Wh,K = Pk+1 (K ) for BDM elements.
Definition 3.1 (Postprocessing method) Find U ∗ such that U ∗ | K = U K∗ ∈ Wh,K
∗
∗
where U K is defined by
Ph,K U K∗ = U | K , (3.60)
and
(∇U ∗ , ∇v) K = (, ∇v) K for all v ∈ (I − Ph,K )Wh,K
∗
. (3.61)

123
A posteriori error estimates for mixed finite element approximations of elliptic problems 497

To compute U ∗ we need to solve one small problem for each element and thus the
total cost is very low.

Corollary 3.1 Given f ∈ L 2 (
) it holds,

  ∗ 2 
σ − 20 ≤ C h 2K  f + ∇ · 20,K +  − ∇U ∗ 20,K + h −1
K  U 0,∂ K ,
K ∈K
(3.62)
where U ∗ is taken from Definition 3.1.

Proof The proof follows directly from Theorem 3.1 with Q = U ∗ .

Remark 3.1 In Corollary 2.8 on page 1667 in [12] the following a priori estimate of
the error is presented for BDM and RTN elements,

σ −  + u − U ∗ 1,K ≤ Ch k+1 |u|k+2 for BDM,

(3.63)
σ −  + u − U ∗ 1,K ≤ Ch k |u|k+1 for RTN,

where |·|k is the H k (
) semi norm, see [1], and ·1,K is the broken H 1 norm defined
by v21,K = K ∈K v21,K . These estimates show that the postprocessed function
U ∗ gives optimal order estimates. Further in Theorem 3.1. on the same page in [12]
the critical term in the error estimate we present in Corollary 3.1,  − ∇U ∗ 2K , is
also proven to be of optimal order.

3.4 Estimate for stabilized methods

Here we extend our estimate to stabilized mixed methods, in particular, we consider

the recent method presented in Masud and Hughes [13]. Stabilized methods are based
on a modified weak formulation which yields a stable method for standard conti-
nuous piecewise polynomial approximations, e.g. piecewise linear functions for both
displacement and flux.
The method reads: find  ∈ V h = {v ∈ [C(
)]n : v| K ∈ [Pk (K )]n for allK ∈ K}
and U ∈ Wh = {v ∈ C(
) : v| K ∈ Pl (K ) for all K ∈ K}, with k, l ≥ 1 such that,

1
(−∇ · , w) + (, v) + (U, ∇ · v) − ( − ∇U, v + ∇w) = ( f, w), (3.64)
2

for all v ∈ V h and w ∈ Wh . Note, in particular, that in this method the order of
polynomials in Wh may be higher than in V h and thus in that case we do not expect
post processing of the pressure to be necessary. Applying the same ideas as in Theorem
3.1 to this stabilized method we obtain the following a posteriori error estimate. The
argument may be modified to cover other stabilized methods such as the Galerkin least
squares method.

Proposition 3.1 For the stabilized Galerkin method defined by (3.64) the following a
posteriori error estimate holds,

123
498 M. G. Larson, A. Målqvist

 
σ − 20 ≤ C h 2K  f + ∇ · 20,K +  − ∇U 20,K . (3.65)
K ∈K

Proof Using the same arguments as in equations (3.21–3.25) in the proof of Theorem
3.1, we obtain the following error representation formula,

σ − 20 = (u − Q, f + ∇ · ) + (Q, −∇ · (σ − )) − (, σ − ) (3.66)

= (u − Q, f + ∇ · ) + (Q, −∇ · (σ − )) K − (, σ − ) K
K ∈K
(3.67)
= I + I I. (3.68)

Using the same technique as in the proof of Theorem 3.1, i.e. by combining equations
(3.43) and (3.51), we can estimate the second term I I as follows

1 1
I I ≤ ∇ Q − 20,K + σ − 20 + h( f + ∇ · )20 + C h −1
K [Q]0,∂ K .
2
2 2
K ∈K
(3.69)
We turn to the first term I in (3.68). We let πh : L 2 (
) → Wh be the Scott–Zhang
interpolant, see [6], and note, by letting v = 0 in (3.64), that we have the Galerkin
orthogonality property

1
( f + ∇ · , w) = − ( − ∇U, ∇w), (3.70)
2

for all w ∈ Wh . By choosing w = πh (u − Q) in equation (3.70) we get the following

identity,
1
( f + ∇ · , πh (u − Q)) = − ( − ∇U, ∇πh (u − Q)). (3.71)
2
We then have,

I = (u − Q, f + ∇ · ) (3.72)
1
= (u − Q − πh (u − Q), f + ∇ · ) − ( − ∇U, ∇πh (u − Q)), (3.73)
2

using equation (3.71). In the first term we split the contributions over the elements
and use the Cauchy–Schwarz inequality. In the second term we only use the Cauchy–
Schwarz inequality,

I ≤ h −1 (u − Q − πh (u − Q))0,K h( f + ∇ · )0,K

K ∈K
1
+  − ∇U 0 ∇πh (u − Q)0 . (3.74)
2

123
A posteriori error estimates for mixed finite element approximations of elliptic problems 499

By using a standard interpolation error estimate for the Scott-Zhang interpolant we

get,
1
I ≤C ∇(u−Q)0,K h( f +∇·)0,K + −∇U 0 ∇πh (u−Q)0 . (3.75)
2
K ∈K

We proceed by using the Cauchy-Schwarz inequality for sums and the identification
σ = ∇u,
1/2 1/2
I ≤C σ − ∇ Q20,K h( f + ∇ · )20,K
K ∈K K ∈K
1
+  − ∇U 0 ∇πh (u − Q)0 . (3.76)
2
Subtracting and adding  to the first term and using the inequality ab ≤ a 2 /(2 ) +
b2 /2 we may choose such that,
1 1
I ≤  − ∇ Q20,K + σ − 20 + C h( f + ∇ · )20,K
4 4
K ∈K
1
+  − ∇U 0 ∇πh (u − Q)0 . (3.77)
2
At this point we collect all terms from the estimates (3.77) of I and (3.69) of I I ,
 
I + II ≤ C h( f + ∇ · )20,K + h −1K [Q] 2
0,∂ K (3.78)
K ∈K

5 3 1
+  − ∇ Q20,K + σ − 20 + +  − ∇U 0 ∇πh (u − Q)0 .
4 4 2
(3.79)
Next we choose Q = U and observe that the jump terms vanish since U is continuous.
We also take advantage of the fact that πh is stable in H 1 ,

I + II ≤ C h( f + ∇ · )20,K (3.80)

K ∈K
5 3
+  − ∇U 20,K + σ − 20 + C − ∇U 0 σ − ∇U 0 .
4 4
(3.81)
Again we add and subtract  in the last term and use the same trick as above to get,

7
σ −20 = I +I I ≤ C h( f +∇·)20,K +C−∇U 20 + σ −20 , (3.82)
8
K ∈K

and thus the proposition follows immediately by subtracting 7/8σ − 20 from both
sides.

123
500 M. G. Larson, A. Målqvist

Acknowledgments The authors wish to thank the referees for careful reading of the manuscript and
helpful advice that resulted in significant improvements of this paper.

References

1. Adams, R.A.: Sobolev spaces. Pure and Applied Mathematics, vol. 65. Academic, New York (1975)
2. Arnold, D.N., Brezzi, F.: Mixed and nonconforming finite element methods: implementation, postpro-
cessing and error estimates. RAIRO Modél. Math. Anal. Numér. 19, 7–32 (1985)
3. Becker, R., Hansbo, P., Larson, M.G.: Energy norm a posteriori error estimation for discontinuous
Galerkin methods. Comput. Methods Appl. Mech. Eng. 192, 723–733 (2003)
4. Braess, D., Verfürt, R.: A posteriori error estimators for the Raviart–Thomas element. SIAM J. Numer.
Anal. 33, 2431–2444 (1996)
5. Bramble, J.H., Xu, J.: A local post-processing technique for improving the accuracy in mixed finite-
element approximations. SIAM J. Numer. Anal. 26, 1267–1275 (1989)
6. Brenner, S.C., Scott, L.R.: The Mathematical Theory of Finite Element Methods. Springer, Heidelberg
(1994)
7. Brezzi, F., Douglas, J. Jr.., Durán, R., Fortin, M.: Mixed finite elements for second order elliptic
problems in three variables. Numer. Math. 51, 237–250 (1987)
8. Brezzi, F., Douglas, J. Jr.., Marini, D.L.: Two families of mixed finite elements for second order elliptic
problems. Numer. Math. 47, 217–235 (1985)
9. Brezzi, F., Fortin, M.: Mixed and Hybrid Finite Element Methods. Springer, Heidelberg (1991)
10. Carstensen, C.: A posteriori error estimate for the mixed finite element method. Math. Comp. 66, 465–
476 (1997)
11. Girault, V., Raviart, P.-A.: Finite Element Approximation of the Navier–Stokes Equation. Springer,
Berlin (1979)
12. Lovadina, C., Stenberg, R.: Energy norm a posteriori error estimates for mixed finite element methods.
Math. Comp. 75, 1659–1674 (2006)
13. Masud, A., Hughes, T.J.R.: A stabilized finite element method for Darcy flow. Comput. Methods Appl.
Mech. Eng. 191, 4341–4370 (2002)
14. Nédélec, J.-C.: A new family of mixed finite elements in R3 . Numer. Math. 50, 57–81 (1986)
15. Raviart, P., Thomas, J.: A mixed finite element method for second order elliptic problems. In: Mathe-
matical aspects of the finite element method. Lecture Notes in Math, vol. 606, pp. 292–315. Springer,
Heidelberg (1977)
16. Stenberg, R.: Some new families of finite elements for the Stokes equations. Numer. Math. 56, 827–
838 (1990)
17. Stenberg, R.: Postprocessing schemes for some mixed finite elements. RAIRO Modél. Math. Anal.
Numér. 25, 151–167 (1991)
18. Roberts, J.E., Thomas, J.-M.: Mixed and hybrid methods. Ciarlet, P.G., Lions, J.L. (eds) Handbook
of Numerical Analysis. Finite Element Methods (Part 1), vol. II, pp. 523–639. Elsevier, Amsterdam
(1991)

123