Mixed Finite Element Methods
Mixed Finite Element Methods
In this chapter we utilize the RaviartThomas spaces to present and analyze specific
mixed finite element methods applied to some of the examples studied in Chap. 2.
The corresponding discussion follows mainly the presentations in [12, 39, 50, 52].
We begin with a preliminary section dealing with the approximation properties of
the finite element subspaces to be employed.
Pdiv,h
k : H(div; ) Hhk ,
Pk1,h : H 1 ( ) Xhk ,
Pkh : L2 ( ) Xhk ,
Phk : L2 ( ) Yhk ,
where Hhk is the global RaviartThomas space defined at the beginning of Sect. 3.4.1,
that is, for each k 0
Hhk := h H(div; ) : h |K RTk (K) K Th , (4.1)
Xhk is the usual Lagrange finite element space, that is, for each k 1
Xhk := vh C( ) : vh |K Pk (K) K Th , (4.2)
G.N. Gatica, A Simple Introduction to the Mixed Finite Element Method: Theory 93
and Applications, SpringerBriefs in Mathematics, DOI 10.1007/978-3-319-03695-3 4,
Gabriel N. Gatica 2014
94 4 M IXED FINITE ELEMENT M ETHODS
Pdiv,h
k
( )div, := inf h div, hk ( )div, ,
h Hhk
which, thanks to Theorem 3.6, implies that for each [H l+1 ( )]n , with div
H l+1 ( ), 0 l k, there also holds
Pdiv,h
k
( )div, C hl+1 | |l+1, + |div |l+1, . (4.4)
v hk (v)0, + |v hk (v)|1,
and
v Pkh (v)0, := inf v vh 0, v hk (v)0, ,
vh Xhk
which, thanks to the known estimates for hk (e.g., [20, 51]), yields for each v
H l+1 ( ), 1 l k, that
and
v Pkh (v)0, C hl+1 |v|l+1, . (4.6)
Now, applying the BrambleHilbert lemma (cf. Theorem 3.5) to S = and
:= Pk1,h (with m = 1 and k = 0), and noting that (p) = p p P0 ( ) Xhk ,
we deduce that
v Pk1,h (v)1, C |v|1, ,
which proves that (4.5) can be extended to l = 0, and therefore we can write
Lemma 4.1. Let be a convex domain, and let k 1. Then there exists C > 0,
independently of h, such that for each v H l+1 ( ), 0 l k, there holds
for each vh Xhk , where the orthogonality condition characterizing the orthogonal
projector Pk1,h is used in the last equality. In particular, taking vh := Pk1,h (T (r)) and
96 4 M IXED FINITE ELEMENT M ETHODS
This last estimate, together with (4.13), gives (4.8) and completes the proof.
It is important to remark here, as stated earlier, that starting from the preceding
lemma one can extend inequality (4.6) to l = 0. In fact, it is clear that
min hK c h h > 0.
KTh
4.1 Projection Operators 97
Then, applying estimate (3.15) (cf. Lemma 3.12) and the geometric bounds given
by Lemma 3.14, we obtain
hK
from which, using that c [cf. (3.32)] and hK c h K Th , it follows that
K
But, since the norms are certainly equivalent in the finite-dimensional space Pk (K),
there holds |vh |1, K C vh 0, K , and hence, using now estimate (3.14) (cf. Lemma
3.12), we deduce that
where we have used that Pk1,h (v) = Pkh (Pk1,h (v)). Then, applying the inverse inequal-
ity given by Lemma 4.2 and the estimate (4.12) (cf. proof of Lemma 4.1), we obtain
that
where PKk : L2 (K) Pk (K) is the local orthogonal projector. Now, from the first
term of (4.20) applied to = K Th , which is obviously convex, we find that
2(l+1)
C 2 hK |v|2l+1, K C h2(l+1) |v|2l+1, ,
KTh
that is,
v Phk (v)0, C hl+1 |v|l+1, . (4.21)
4.2 Poisson Problem 99
In this section we analyze a Galerkin scheme for the Poisson problem studied in
Sect. 2.4.1. For this purpose, we first recall that, given a bounded domain
Rn , n {2, 3}, with polyhedral boundary , and given f L2 ( ) and g H 1/2 ( ),
the corresponding mixed variational formulation reduces to [cf. (2.19)] the follow-
ing problem: find ( , u) H Q such that
a( , ) + b( , u) = F( ) H ,
b( , v) = G(v) v Q,
a( h , h ) = h 20, = h 2div, ,
Proof. Let [H 1 ( )]n . Then, applying the upper bound for the local interpolation
hK
error given by Lemma 3.17 (with m = 0 and l = 0), and using that c, we
K
obtain
h2
hk ( )0, K C K | |1, K C hK | |1, K K Th ,
K
and then
hk ( )0, hk ( )0, + 0,
b( h( ), vh ) = 0 H, v h Qh , h > 0.
z = f in B, z = 0 on B ,
h ( ) := hk ( ) H(div; ). (4.29)
and then
which confirms the uniform boundedness of {h }h>0 . Next, using the estimate (3.8)
given by Lemma 3.7 and the fact that Phk is the orthogonal projector of L2 ( ) into
Qh := Yhk , we deduce that for each H and for each vh Qh there holds
b( h ( ), vh ) = vh (div div h ( ))
= vh (div div hk ( ))
= vh (div Phk (div ))
= vh (div Phk (div )) = 0,
which yields the second property required by {h }h>0 , and consequently the Fortin
lemma (cf. Lemma 2.6) guarantees that b satisfies the discrete inf-sup condition
on Hh Qh with a constant > 0, independently of h. It is important to observe
here that, proceeding similarly to the construction of the present Fortin operator,
102 4 M IXED FINITE ELEMENT M ETHODS
one can prove that in this case there also holds div Hh = Qh . Indeed, given vh Qh ,
it suffices to replace div by vh in the preceding definition of f , which leads to
div hk ( ) = Phk (div ) = Phk (vh ) = vh in .
Consequently, a straightforward application of Theorems 2.4 and 2.6 implies that
there exist a unique ( h , uh ) Hh Qh solution of (4.22) and constants C1 ,C2 > 0,
independent of h, such that
( h , uh )H Q C1 f 0, + g1/2,
and
h H + u uh Q C2 dist( , Hh ) + dist(u, Qh ) , (4.31)
On the other hand, if ( , u) is not sufficiently regular and estimates (4.32) and (4.33)
do not necessarily hold, the convergence of the Galerkin scheme (4.24), but without
any rate of convergence, can still be proved by employing suitable density argu-
ments. More precisely, we have the following result.
Proof. We use that [C ( )]n and C0 ( ) are dense in H(div; ) and L2 ( ), respec-
tively, whence there exist sequences { j } jN [C ( )]n and {u j } jN C0 ( )
j j
such that j div, 0 and u u j 0, 0. Then, since it is clear that
j [H 1 ( )]n , div j H 1 ( ), and u j H 1 ( ), it follows from (4.4) and (4.21)
that for each j N
4.3 Primal-Mixed Formulation of Poisson Problem 103
j Pdiv,h
k
( j )div, C h {| j |1, + |div j |1, } (4.35)
and
u j Phk (u j )0, C h |u j |1, . (4.36)
Now, given > 0, there exists N N such that
Next, for j = N we deduce from (4.35) and (4.36) that there exists h0 > 0 such
that
N Pdiv,h
k
( N )div, < /4 and uN Phk (uN )0, < /4 h h0 .
Therefore, from the Cea estimate (4.31) we conclude that for each h h0
h div, + u uh 0, C2 dist( , Hh ) + dist(u, Qh )
C2 Pdiv,h
k
( N )div, + u Phk (uN )0,
C2 N div, + N Pdiv,hk
( N )div,
+ u uN 0, + uN Phk (uN )0,
C2 ,
that is,
Vh V := v H : v=0 .
Then, utilizing the generalized Poincare inequality (cf. [46, Theorem 5.11.2]), one
can prove that 1, and | |1, are equivalent in V and, hence, in Vh . It follows that
b(vh , h )
sup h 1/2, h Qh ,
vh Hh vh 1,
vh =0
4.3 Primal-Mixed Formulation of Poisson Problem 105
which is
h , vh
sup h 1/2, h Qh .
vh Hh vh 1,
vh = 0
For this purpose we need an inverse inequality for Qh , which is proved by the
following lemma for a generic space Qh .
Lemma 4.6. Let {1 , 2 , , m } be a partition of , denote h j := |j | j
{1, , m}, assume that there exists c > 0 such that
and define
Qh := h L2 ( ) : h |j P0 (j ) j {1, , m} .
h 1/2, h0 h 1/2, h Qh ,
and then conclude by interpolation estimates (cf. [49, Appendix B]). In fact, given
h Qh , we let j := h |j P0 (j ) and observe that
m m m
h 20, = h 20,j = h j j2 = h j j 20, ,
j=1 j=1 j=1
1/2 1/2
where 1/2,00, is the norm of H00 ( ), which is the dual of H00 ( ). On the
other hand, applying the inequality [cf. (3.15), Lemma 3.12]
|v|m,K c B1
K |det BK |
m 1/2
|v|m,K
106 4 M IXED FINITE ELEMENT M ETHODS
to K = j and K = , we obtain
1/2
|j | 1/2
v0,j c v0, = c h j v0, v L2 (j )
| |
and
1
1/2
h |j | 1/2
|v|1,j c |v|1, c h j |v|1, v H 1 (j ).
j | |
Then, according to the interpolation estimates for Sobolev spaces (cf. [49, Appendix
1/2
B]) and using that H00 (j ) = (H01 (j ), L2 (j ))1/2 , we find that
1/2
v1/2,00,j c v1/2,00, v H00 (j ),
1/2
where 1/2,00,S denotes the norm of H00 (S) for S {j , }.
Analogously, applying now [cf. (3.14), Lemma 3.12]
1/2
to K = j and K = , using interpolation estimates again, and noting that H00 ( )
is given by (H01 ( ), L2 ( ))1/2 , we deduce that
1/2
v1/2,00, c v1/2,00,j v H00 ( ) ,
and therefore
v1/2,00,j =
v1/2,00, .
1/2
Consequently, given H00 (j ), we obtain by a duality argument that
, v
1/2,00, = sup
1/2 v1/2,00,
v H00 ( )
v=0
h1
j , vj
= sup
1/2 v1/2,00,
v H00 ( )
v= 0
h1
j 1/2,00,j v1/2,00,j
C sup
1/2 v1/2,00,j
v H00 ( )
v= 0
= C h1
j 1/2,00,j .
h , vh
sup h 1/2, h Qh .
vh Hh v h 1,
vh = 0
z + z = 0 in , z n = h in .
z1, c h 1/2, .
z1+ , C h 1/2+ , .
and clearly
v P11,h (v)1, h0 v1, v H 1 ( ),
the estimates for the interpolation of Sobolev spaces (cf. [49, Appendix B]) imply
It follows that
and using the inverse inequality for Qh (cf. Lemma 4.6), that is,
h 1/2+ , C h h 1/2, ,
we arrive at
h
z P11,h (z)1, C h 1/2, . (4.40)
h
Next, it is clear that
Then, using the Green identity in H(div; ) (cf. Lemma 1.4) we see that
h , z = z n, z = n ( z), 0 (z)
= z div z + z z = z21, ,
which yields
h , z h 21/2, . (4.42)
h
Ch 1/2, C h 1/2,
h
h
= C C h 1/2, ,
h
1/
C
where, choosing h C0 h, with C0 = , we deduce the existence of > 0
2C
such that
4.3 Primal-Mixed Formulation of Poisson Problem 109
h , vh
sup h 1/2, h Qh ,
vh Hh vh 1,
vh = 0
Note that the first term on the right-hand side of the preceding equation reduces to
u P11,h(u)1, , which can be bounded by means of (4.7). To estimate the second
term we need the following previous result.
Lemma 4.8. Let Ph0 : L2 ( ) Qh be the orthogonal projector with respect to the
L2 ( )-inner product. Then there holds
and
Ph0 ( )0, C h 1, H 1 ( ),
Next, using a duality argument and the preceding estimate, we have that for each
H 1/2 ( ) there holds
Ph0 ( ),
Ph0 ( )1/2, = sup
H 1/2 ( )
1/2,
=
Ph0 ( ), 0,
= sup
H 1/2 ( )
1/2,
=
110 4 M IXED FINITE ELEMENT M ETHODS
= C h 1/2, ,
which completes the proof.
and that
that is,
which, together with (4.44) and thanks to the interpolation estimates for Sobolev
spaces, gives
1
P1/2,h( )1/2, C hr+ 2 r, H r ( ), r [1/2, 1/2].
(4.45)
Therefore, recalling that
In this section we analyze a Galerkin scheme for the two-dimensional version of the
Poisson problem studied in Sect. 2.4.2 with Neumann boundary conditions, that is,
when N = . In other words, given a bounded domain of R2 , f L2 ( ), and
g H 1/2 ( ), we are interested in the boundary value problem
u
u = f in , = g in , u = 0,
n
for which we need to assume that the data satisfy the compatibility condition
f + g, 1 = 0.
:= u in and := 0 (u) in
and proceeding as in Sects. 2.4.1 and 2.4.2, one arrives at the mixed variational
formulation: find ( , (u, )) H Q such that
a( , ) + b( , (u, )) = 0 H,
b( , (v, )) = f v + g, (v, ) Q,
and
b( , (v, )) := v div + n, H, (v, ) Q . (4.48)
a( h , h ) + b( h , (uh , h )) = 0 h Hh ,
(4.49)
b( h , (vh , h )) = f vh + g, h (vh , h ) Qh .
112 4 M IXED FINITE ELEMENT M ETHODS
For the analysis of (4.49) we first focus on the discrete inf-sup condition for b, that
is, on the eventual existence of > 0, independent of h, such that
b( h , (vh , h ))
sup (vh , h )Q (vh , h ) Qh . (4.50)
h Hh h H
h = 0
Since b can be decomposed as the sum of two bilinear forms b1 and b2 , that is [cf.
(4.48)],
b( , (v, )) = b1 ( , v) + b2 ( , ) H, (v, ) Q , (4.51)
we could certainly utilize the corresponding characterization result established in
[40, Theorem 7] to prove (4.50). Alternatively, and due to the same decomposition,
one could also employ the slightly different equivalence given in [45, Theorem 3.1].
However, and to provide further points of view to this analysis, in what follows we
apply another procedure that can be seen as a combination of the aforementioned
approaches. Indeed, we first use the boundedness of the normal trace of vectors in
H(div; ) [see (1.44) in the proof of Theorem 1.7] to deduce that
b( h , (vh , h )) b1 ( h , vh )
sup sup h 1/2,
h Hh h H h Hh h div,
h = 0 h = 0
vh div h
= sup h 1/2, (vh , h ) Qh .
h Hh h div,
h = 0
Next, considering the particular subspaces [cf. (4.1) and (4.3)]
Hh := Hh0 := h H(div; ) : h |K RT0 (K) K Th ,
Quh := Yh0 L20 ( ) := vh L20 ( ) : vh |K P0 (K) K Th
b( h , (vh , h ))
sup vh 0, h 1/2, (vh , h ) Qh , (4.52)
h Hh h H
h = 0
b( h , (vh , h )) b2 ( h , h )
sup sup
h Hh h H h V1,h h div,
h = 0 h = 0
(4.53)
h n, h
= sup (vh , h ) Qh ,
h V1,h h div,
h = 0
4.4 Poisson Problem with Neumann Boundary Conditions 113
Hence, it is easy to see from (4.52) and (4.53) that, in order to conclude (4.50), it
suffices to show that there exists a constant > 0, independent of h, such that
h n, h
sup h 1/2, h Qh . (4.55)
h V1,h h div,
h = 0
Throughout the rest of this section we aim to prove (4.55). To this end, we now
introduce the following definition.
Definition 4.1. Let h ( ) := { h n| : h V1,h }. We say that a linear operator
Lh : h ( ) V1,h is a STABLE DISCRETE LIFTING if
(i) Lh (h ) n = h on h h ( );
(ii) c > 0, independently of h such that
Lemma 4.9. Assume that there exists a stable discrete lifting Lh : h ( ) V1,h .
Then the discrete inf-sup condition (4.55) is equivalent to the existence of C > 0,
independent of h, such that
h , h
sup C h 1/2, h Qh . (4.56)
h h ( ) h 1/2,
h = 0
|h , h | c |h , h | | h n, h |
c sup
h 1/2, Lh(h )div, h V1,h h div,
h = 0
and, according to the bound given by (1.44) (cf. Theorem 1.7), that
| h n, h | | h n, h | |h , h |
sup ,
h div, h n1/2, h h ( ) h 1/2,
h = 0
and 5.3] and assume that Th is quasi-uniform around . This means that there exist
a neighborhood of and a constant c > 0, independent of h, such that, denoting
Th, := {K Th : K = 0}
/ , (4.57)
there holds
max hK c min hK .
KTh, KTh,
hK
c K Th , h > 0
K
or, equivalently, that {Th }h>0 satisfies the minimum angle condition, the quasi-
uniformity assumption implies that the partition on inherited from Th , say h , is
also quasi-uniform, that is, there exists c > 0, independently of h, such that
h := max |e| : e h c min |e| : e h .
We now define
h ( ) := h L2 ( ) : h |e P0 (e) e h
Theorem 4.1. Under the previously stated assumptions, there exists a stable discrete
lifting Lh : h ( ) V1,h .
v1+ , C h 1/2+ , .
It follows that v [H ( )]2 H(div; ) [note that div (v) = v = 1
| | h
R], and hence we can define (see remark immediately preceding Lemma 3.19)
Lh (h ) := h0 (v) .
According to the preceding equation and (3.8) (cf. Lemma 3.7), we obtain
1
div Lh (h ) = div h0 (v) = Ph0 (div v) = Ph0 ( v) = h P0 ( ) ,
| |
which confirms that Lh (h ) V1,h , and, thanks to (3.36) (see proof of Lemma 3.18),
we find that
Next, we recall from (4.57) the definition of Th, and introduce the sets
h1 := K Th : K Th, \
and
h2 := \h1 = K Th, .
Since \ is strictly contained in , the interior elliptic regularity result (cf. [49,
Theorem 4.16]) implies that v| \ H 2 ( \ ) and
v2, \ C2 h 1/2, .
116 4 M IXED FINITE ELEMENT M ETHODS
It follows that
Lh (h )0, Lh (h )0, 1 + Lh (h )0, 2
h h
On the other hand, applying estimate (3.37) (cf. Lemma 3.19) and inverse
inequality (4.58), we obtain
v h0 (v)20, 2 =
h
v K0 (v)20,K
KTh,
1 2
C hK |v| ,K +
2 2
h
| | 0,K
KTh,
C max h2K v21+ , 2 + h 21/2,
KTh, h
2
C max hK v1+ , + h 21/2,
2
KTh,
C max h2K h 21/2+ , + h 21/2,
KTh,
C max h2K h2 h 21/2, + h 21/2,
KTh,
C h 21/2, ,
where the fact that hK C h K Th, has been used in the last inequality.
Consequently, gathering together the preceding estimates, we conclude that
We now aim to prove the following result, which, according to Lemma 4.9, will
suffice to conclude the required discrete inf-sup condition for the term on .
Lemma 4.10. There exists > 0, independent of h, such that
h , h
sup h 1/2, h Qh .
h h ( ) h 1/2,
h = 0
For the aforementioned purpose we proceed as in [39, Sect. 5.3] and describe in
what follows two different procedures under which Lemma 4.10 is proved.
4.4 Poisson Problem with Neumann Boundary Conditions 117
Proof. Given h Qh , we let z H 1 ( ) be the unique solution of the problem
z + z = 0 in , z = h on .
Since Qh H 1 ( ), we obtain by elliptic regularity (cf. [42]) that z H 1+ ( )
and z1+ , C h 1/2+ , [0, 0 ], where 0 := min 12 , and is
the largest interior angle of . In what follows, we fix (0, 0 ], < 1/2, and
observe that z n H 1/2+ ( ) and
z n C z1+ , .
1/2+ ,
1/2
where Ph : H 1/2 ( ) h ( ) is the orthogonal projector with respect to the
H 1/2 ( )-inner product. Hence, applying the inverse inequality for Qh (see remark
following proof), we deduce that
118 4 M IXED FINITE ELEMENT M ETHODS
1/2 h
z n Ph (z n) C h 1/2, .
1/2, h
We notice now that, to have the inverse inequality for Qh , one needs to see that
h 1/2, h0 h 1/2, h Qh
for which it suffices to demonstrate that |h |1, C h1/2 h 1/2, h Qh .
PROCEDURE 2. Let h ( ) := h L2 ( ) : h |e P0 (e) e h , define h :=
max |e| : e h , and assume that the number of edges e of h is even. Then we set
Qh := h C( ) : h |e P1 (e) e 2h ,
4.4 Poisson Problem with Neumann Boundary Conditions 119
where 2h is the partition of that arises by joining adjacent edges (certainly lying
on the same line). Note that dim h ( ) = 2 dim Qh . As was the case earlier, the
goal here is to show the existence of > 0 such that
h , h
sup h 1/2, h Qh . (4.59)
h h ( ) h 1/2,
h = 0
To this end, we assume henceforth (see details in [39]) that there exist h ( )
h ( ) and constants 0 , 1 > 0, such that dim h ( ) = dim Qh ,
h , h
sup 0 h 0, h Qh , (4.60)
h h ( )
h 0,
h = 0
and
h , h
sup 1 h 1, h Qh , (4.61)
h h ( )
h 1,
h = 0
so that (4.59) is proved next by interpolating (4.60) and (4.61). In fact, we have
the following result (cf. F.J. Sayas 2012, private communication).
Lemma 4.12. There exists > 0, independent of h, such that (4.59) holds.
Proof. Let Gh : L2 ( ) Qh be the operator defined by Gh ( ) := h for each
L2 ( ), where h is the unique element in Qh such that
h , h = h , h h ( ) .
Note that the inf-sup condition (4.60) and the fact that dim h ( ) = dim Qh guar-
antee the existence and uniqueness of h . Observe, in addition, that h is a Petrov
Galerkin type approximation of . Now it is easy to see that (4.60) and (4.61) can
be put together in the form
h , h
sup s h s, h Qh , s {0, 1}.
h h ( )
h s,
h = 0
Applying the preceding inequality to Gh ( ) Qh , we obtain
1 h , Gh ( )
Gh ( )s, sup
s ( ) h s,
h h
h = 0
1 h , 1
= sup s, H s ( ) ,
s h h ( )
h s, s
h = 0
120 4 M IXED FINITE ELEMENT M ETHODS
that is,
and then, thanks to the interpolation estimates for Sobolev spaces (cf. [49, Appendix
B]), we obtain
1/2
We now let Ph : H 1/2( ) h ( ) be the orthogonal projector. Then, given
H ( ), we consider the functional f : H 1/2 ( ) R defined by
1/2
1/2
f ( ) := Ph ( ), H 1/2 ( )
H 1/2 ( )
is
the usual isometry characterizing reflexive spaces, that is, J(v)(F) = F(v) F
H 1/2 ( ) H 1/2 ( )
, v H 1/2 ( ). It follows that h h ( ) there holds
h , Gh (v ) = h , v = J(v )(h ) = f (h )
1/2
= Ph (h ), = h , = h , Gh ( ) ,
Finally, applying the preceding expression to h Qh , noting in this case that
Gh (h ) = h , and recalling that h ( ) h ( ), we conclude that
h , h h , h
sup sup h Qh ,
h h ( )
h 1/2, h h ( ) h 1/2,
h = 0 h = 0
where P1/2,h : H 1/2 ( ) Qh is the orthogonal projector with respect to the inner
product of H 1/2 ( ).
In this section we analyze the Galerkin scheme for the two-dimensional version
of the linear elasticity problem with Dirichlet boundary conditions studied in
Sect. 2.4.3.1. To this end, we recall that, given a bounded domain R2 with
Lipschitz-continuous boundary , and given f L2 ( ), the corresponding mixed
formulation reduces to [cf. (2.50)] the following problem: find ( , (u, )) H0 Q
such that
122 4 M IXED FINITE ELEMENT M ETHODS
a( , ) + b( , (u, )) = F( ) H0 ,
(4.63)
b( , (v, )) = G(v, ) (v, ) Q ,
where (cf. Sect. 2.4.3.1)
H0 := H0 (div; ) : tr( ) = 0 , Q := L2 ( ) L2skew ( ),
We now let {Hh }h>0 , {Q1,h }h>0 , and {Q2,h }h>0 be families of arbitrary finite
element subspaces of H0 , Q1 := L2 ( ), and Q2 := L2skew ( ), respectively. Then,
denoting Qh := Q1,h Q2,h , we consider the associated Galerkin scheme: find
( h , (uh , h )) Hh Qh such that
a( h , h ) + b( h , (uh , h )) = F( h ) h Hh ,
(4.67)
b( h , (vh , h )) = G(vh , h ) (vh , h ) Qh .
Our goal is to apply the theory from Sect. 2.5 to find specific subspaces Hh , Q1,h
and Q2,h ensuring the unique solvability and stability of (4.67). More precisely,
assuming in advance that a is going to be elliptic on the discrete kernel Vh of b
(which actually will be shown later on), we concentrate in what follows on proving
the discrete inf-sup condition for b, that is, the existence of > 0, independent of
h, such that
b( h , (vh , h ))
sup (vh , h )Q (vh , h ) Qh . (4.68)
h Hh h H
h = 0
To accomplish this, we know from Fortins lemma (cf. Lemma 2.6) that it suffices to
build a sequence of uniformly bounded operators {h }h>0 L (H, Hh ) such that
b( , (v, )) = b1 ( , v) + b2 ( , ) ( , (v, )) H0 Q,
that is,
b1 ( , v) := v div and b2 ( , ) := : .
b1 ( h ( ), vh ) + b2 ( h ( ), h ) = 0 (vh , h ) Qh . (4.70)
then we aim to find a second sequence of uniformly bounded operators {2,h }h>0
L (H, Hh ) such that
(i) b1 (2,h ( ), vh ) = 0 vh Q1,h and
(ii) b2 ( 1,h ( ) 2,h ( ), h ) = 0 h Q2,h ,
so that defining h := 1,h + 2,h we satisfy the condition (4.70) [equivalently,
(4.69)]. Indeed, it is easy to see from the analysis in Sect. 4.2 that, given a triangu-
larization Th of and an integer k 0, and defining
and
Q1,h := {vh L2 ( ) : vh |K [Pk (K)]2 K Th } , (4.73)
where h,i stands for the ith row of h , one can proceed by rows as in (4.29) to define
a uniformly bounded explicit family {1,h }h>0 satisfying (4.71).
It remains therefore to build a family {2,h}h>0 L (H, Hh ) verifying (i) and
(ii). To this end, we now follow the approach introduced in [23] (see also [12] for
further extensions). More precisely, we let Xh and Yh be stable finite element sub-
spaces for the usual primal formulation of the Stokes problem and, given H0 ,
consider the associated Galerkin scheme: find (zh , ph ) Xh Yh such that
zh : h + ph div h = 0 h Xh ,
(4.74)
qh div zh = ( 1,h ( )) : S(qh ) qh Yh ,
0q
where S(q) := Q2 := L2skew ( ) q L2 ( ). Note that the stability
q 0
of (4.74) guarantees the existence of a constant C > 0, independent of h, such that
124 4 M IXED FINITE ELEMENT M ETHODS
Thus, a comparison between this identity and the required condition (ii) for 2,h
suggests defining 2,h ( ) := curlzh under the assumptions that curl(Xh ) Hh and
Q2,h S(Yh ). Next, since div 2,h ( ) = div curl zh = 0, it follows that
b1 (2,h ( ), vh ) = vh div 2,h ( ) = 0 vh Q1,h ,
which shows that (i) is also satisfied. In addition, thanks to the uniform boundedness
of {1,h }h>0 [cf. (4.30)] and the stability result given by (4.75), we find that for each
H0 there holds
In particular, if we consider the mini finite element (cf. [41, Chap. II, Sect. 4.1])
given by
Xh := h [C( )]2 : h |K [P1 (K) bK ]2 K Th
and
Yh := qh C( ) : qh |K P1 (K) K Th ,
where bK is the bubble function on the triangle K, then (4.76) (with k = 0) and
(4.77) become
Hh := h H0 : h,i |K RT0 (K) + curl bK K Th
and
0 qh
Q2,h := h := : qh C( ) and qh |K P1 (K) K Th ,
qh 0
which, together with Q1,h given by (4.73) (with k = 0), constitutes the well-known
PEERS finite element subspace of order 0 for linear elasticity (cf. [2]).
Finally, it is easy to see from these definitions of Hh , Q1,h , and Q2,h that the
discrete kernel of b becomes
Vh := h Hh : div h = 0 in and h : h = 0 h Q2,h .
Hence, according to the inequalities given by (2.52) and Lemma 2.3, we conclude
that a is Vh -elliptic, which completes the hypotheses required by the discrete
BabuskaBrezzi theory (cf. Theorem 2.4) for the well-posedness of (4.67).
We end this monograph by mentioning that certainly many interesting topics have
been left out of the discussion, including, to name just a few, nonlinear boundary
value problems, time-dependent problems, a posteriori error analysis, and further
applications in continuum mechanics and electromagnetism (see, e.g., the recent
book [13] and the extensive list of references therein for a thorough discussion of
them). In particular, it is worth mentioning that in the case of the linear elastic-
ity problem, new stable mixed finite element methods in two and three dimensions
with either strong symmetry or weakly imposed symmetry for the stresses have
been derived over the last decade using the finite element exterior calculus, a quite
abstract framework involving several sophisticated mathematical tools (e.g., [36]).
In addition, concerning a posteriori error estimates for mixed finite element meth-
ods, we refer to the key contributions in [1, 17] and, within the context of the lin-
ear elasticity and Stokes problems, to [18, 19, 47]. Furthermore, and as comple-
mentary bibliographic material addressing some of the related contributions by the
author, together with his main collaborators and former students, we may also refer
126 4 M IXED FINITE ELEMENT M ETHODS
to [10, 11, 21, 24, 25, 2730, 3234, 3638], which deal mainly with augmented
mixed methods for linear and nonlinear problems in elasticity and fluid mechanics,
twofold saddle point variational formulations, fluidsolid interaction problems, and
the corresponding a posteriori error analyses. We hope to write an extended version
of the present book in the near future that will incorporate the contents of most of
the aforementioned references.