Chapter 5 A Priori Error Estimates For Nonconfor-Ming Finite Element Approximations 5.1 Strang's First Lemma

Chapter 5 A priori error estimates for nonconfor-
ming nite element approximations

5.1 Strangs rst lemma
We consider the variational equation
(5.1) a(u, v) = (v) , v V H
1
() ,
and assume that the conditions of the Lax-Milgram lemma are satised
so that (5.1) admits a unique solution u V .
We assume that H is a null sequence of positive real numbers and
(V
h
)
hH
an associated family of conforming nite element spaces
V
h
V, h H. We approximate the bilinear form a(, ) : V V lR
and the functional () : V lR in (5.1) by bounded bilinear forms
and bounded linear functionals
(5.2) a
h
(, ) : V
h
V
h
lR ,
h
() : V
h
lR , h H
and consider the nite dimensional variational equations
(5.3) a
h
(u
h
, v
h
) =
h
(v
h
) , v
h
V
h
, h H .
Denition 5.1 Uniform V
h
-ellipticity
The sequence (a
h
(, ))
hH
of approximate bilinear forms a
h
(, ) : V
h
V
h
lR is called uniformly V
h
-elliptic, if there exists a positive
constant such that uniformly in h H
(5.4) a
h
(u
h
, u
h
) u
h
2
V
, u
h
V
h
.
Under the assumption of uniform V
h
-ellipticity, the variational equa-
tions (5.3) admit unique solutions u h V
h
. The following result,
known as Strangs rst lemma, can be viewed as a generalization of
Ceas lemma.
Theorem 5.1 Strangs rst lemma
Assume that (a
h
(, ))
hH
is a uniformly V
h
-elliptic family of bilinear
forms and denote by u V and u
h
V
h
, h H the unique solutions
of (5.1) and (5.3), respectively. then, there exists a constant C lR
+
,
independent of h H such that
u u
h
V
(5.5)
C
_
inf
v
h
V
h
_
u v
h
V
+ sup
w
h
V
h
|a(v
h
, w
h
) a
h
(v
h
, w
h
)|
w
h
V
_
+
+ sup
w
h
V
h
|(w
h
)
h
(w
h
)|
w
h
V
_
, h H .
1
2 Ronald H.W. Hoppe
Proof. Taking advantage of the uniform V
h
-ellipticity, for v
h
V
h
we
obtain
u
h
v
h
2
V
a
h
(u
h
v
h
, u
h
v
h
) a(u v
h
, u
h
v
h
) =
= a(u v
h
, u
h
v
h
) +
_
a(v
h
, u
h
v
h
) a
h
(v
h
, u
h
v
h
)
_
+
+
_
h
(u
h
v
h
) (u
h
v
h
)
_
.
Using the boundedness of the bilinear form, i.e., |a(u, v)| Mu
V
v
V
,
and dividing the previous inequality by u
h
v
h
V
, it follows that
u
h
v
h
2
V
Mu
h
v
h
V
+
|a(v
h
, u
h
v
h
) a
h
(v
h
, u
h
v
h
)|
u
h
v
h
V
+
+
|
h
(u
h
v
h
) (u
h
v
h
)|
u
h
v
h
M u
h
v
h
V
+ sup
w
h
V
h
|a(v
h
, w
h
) a
h
(v
h
, w
h
)|
w
h
V
+
+ sup
w
h
V
h
|
h
(w
h
) (w
h
)|
w
h
V
.
Using the triangle inequality
u u
h
V
u v
h
V
+ u
h
v
h
V
and the previous inequality, we deduce (5.5).
Strangs rst lemma shows that the upper bound for the global dis-
cretization error consists of two parts: the approximation error
(5.6) inf
v
h
V
h
u v
h
V
and the consistency errors
inf
v
h
V
h
sup
w
h
V
h
|a(v
h
, w
h
) a
h
(v
h
, w
h
)|
w
h
V
, (5.7)
sup
w
h
V
h
|(w
h
)
h
(w
h
)|
w
h
V
. (5.8)
As an application of Strangs rst lemma, in the subsequent section
we will study the eect of numerical integration in the nite element
approximation of second order elliptic boundary value problems.
Finite Element Methods 3
5.2 Numerical integration
5.2.1 Construction of quadrature formulas
We consider the model variational equation
(5.9) a(u, v) = (v) , v V := H
1
0
() ,
where lR
d
is supposed to be a polygonal resp. polyhedral domain
and the bilinear form a(, ) : V V lR and the functional () : V
lR are given according to
(5.10) a(u, v) :=
_
i,j=1
a
ij
u
x
i
v
x
j
dx ,
and
(5.11) (v) :=
_
fv dx .
Here, f L
2
() and a
ij
C(), 1 i, j d, satisfying
(5.12)
d
i,j=1
a
ij
(x)
i
j

d
i=1
2
i
, x , > 0 .
We approximate (5.9) using a nite element space V
h
generated by a
triangulation T
h
and nite elements (K, P
K
,
K
), K T
h
, satisfying
the assumptions (A1),(A2),(A3) of section 4.3 and

P
K
H
1
(

K)
which implies conformity, i.e., V
h
V .
We approximate the bilinear form (5.10) and the functional (5.11) by
evaluating the integrals by means of appropriate quadrature formu-
las with respect to the individual elements K T
h
of the triangulation.
Taking into account the ane equivalence of the nite elements with
respect to a reference element

K
F
K
:

K K
x F
K
( x) = B
K
x +b
K
, K T
h
,
due to the transformation of variables formula
(5.13)
_
K
(x) dx = det(B
K
)
_
K
( x) d x , := F
K
,
a quadrature formula for an integral over the reference element

K in-
duces a suitable quadrature formula for the associated integral over the
4 Ronald H.W. Hoppe
actual element K = F
K
(

K) of T
h
.
In particular, we consider the reference quadrature formula
(5.14)

Q
K
( ) :=
L
=1

,

K
(
b
,

K
) ,
with weights
,

K
, 1 L, and nodes

b
,

K

K, 1 L.
We refer to
(5.15)

E
K
( ) :=
_
K
( x) d x

Q
K
( )
as the associated quadrature error functional.
Lemma 5.1 Quadrature formulas in the FEM
Assume that

Q
K
is a quadrature formula with respect to the reference
element

K as given by (5.14). Then, for the ane equivalent element
K := F
K
(

K) we obtain a quadrature formula according to
Q
K
() :=
L
i=1
i,K
(b
i,K
) , (5.16)
,K
:= det(B
K
)
,

K
, b
,K
:= F
K
(
b
,

K
) , 1 L .
Moreover, referring to
(5.17) E
K
() :=
_
K
(x) dx Q
K
()
as the related quadrature error functional, we have
(5.18) E
K
() = det(B
K
)

E
K
( ) .
Proof. The proof is a direct consequence of (5.13).
Based on the quadrature formula (5.16), we dene the approximate
bilinear forms a
h
(, ) : V
h
V
h
lR by means of
(5.19) a
h
(u
h
, v
h
) :=
KT
h
L
=1
,K
d
u,j=1
(a
ij
u
h
x
i
v
h
x
j
)(b
,K
) ,
and the approximate linear functionals
h
() : V
h
lR according
to
(5.20)
h
(v
h
) :=
KT
h
L
=1
,K
(fv
h
)(b
,K
) .
Then, the nite element approximation of the model problem (5.9)
amounts to the computation of u
h
V
h
, h H, as the solution of the
approximate variational equation
(5.21) a
h
(u
h
, v
h
) =
h
(v
h
) , v
h
V
h
.
We note that according to the Lax-Milgram lemma, (5.21) admits a
unique solution, if the bilinear form (a
h
(, ))
hH
is bounded and V
h
-
elliptic.
We rst derive sucient conditions for the uniform V
h
-ellipticity
of the family (a
h
(, ))
hH
of approximate bilinear forms.
Theorem 5.2 Sucient conditions for uniform V
h
-ellipticity
Let

Q
K
be a quadrature formula with respect to the reference element
K with positive weights

,

K
, 1 L, and the property that there
exists q lN such that
K
P
q
(

K) ,
(i) the quadrature formula is exact for polynomials p P
2q2
(

K),
or
(ii) the union of nodes
L
=1
{
b
,

K
} contains a P
q1
(

K)-unisolvent
subset.
Then, there exists a positive constant , independent of h H, such
that
(5.22) a
h
(v
h
, v
h
) |v
h
|
2
1,
, v
h
V
h
, h H .
Proof. Observing v
h
|
K
= p
K
P
K
and using the ellipticity condition
(5.12), we nd
L
=1
,K
d
i,j=1
(a
ij
v
h
x
i
v
h
x
j
)(b
,K
) = (5.23)
=
L
=1
,K
d
i,j=1
(a
ij
p
K
x
i
p
K
x
j
)(b
,K
)

L
=1
,K
d
i=1
(
p
K
x
i
(b
,K
))
2
. .
= Dp
K
(b
,K
)
2
.
Now, observing
D p
K
(
b
,

K
) = Dp(b
,K
)(B
K
) , 1 L ,
6 Ronald H.W. Hoppe
we have
D p
K
(
b
,

K
) B
K
Dp(b
,K
) , 1 L ,
and hence, using Theorem 4.3, we get
(5.24)
L
=1
,K
d
i=1
(
p
K
x
i
(b
,K
))
2
. .
= Dp
K
(b
,K
)
2
B
K
2
L
=1
,K
d
i=1
(
p
K
x
i
(
b
,

K
))
2
= det(B
K
) B
K
2
L
=1

,

K
d
i=1
(
p
K
x
i
(
b
,

K
))
2
.
(i) Let us rst assume that the quadrature formula

Q
K
is exact for
polynomials p P
2q2
(

K). Since
d
i=1
(
p
K
x
i
)
2
P
2q2
(

K) ,
we then have
(5.25) | p
K
|
2
1,

K
=
_
K
d
i=1
(
p
K
x
i
)
2
d x =
d
i=1
(
p
K
x
i
(
b
,

K
))
2
.
Inserting (5.25) into (5.24) and using again Theorem 4.3, it follows that
L
=1
,K
d
i=1
(
p
K
x
i
(b
,K
))
2
(5.26)
det(B
K
) B
K
2
L
=1

,

K
d
i=1
(
p
K
x
i
(
b
,

K
))
2
=
= det(B
K
) B
K
2
| p
K
|
2
1,

K

_
B
K
B
1
K

_
2
|p
K
|
2
1,K
.
Due to the shape regularity of T
h
, h H,
B
K
B
1
K

K
h
k
K
C .
Combining (5.23),(5.24) and (5.26), we deduce the existence of a posi-
tive constant , independent of h H, such that
L
=1
,K
d
i,j=1
(a
ij
v
h
x
i
v
h
x
j
)(b
,K
) |v
h
|
2
1,K
, v
h
V
h
. (5.27)
Forming the sum over all K T
h
in (5.27), gives the assertion.
(ii) What remains to be shown is that the assertion also holds true,
if we assume that the union
L
=1
{
b
,

K
} contains a P
q1
(

K)-unisolvent
subset. We claim that in this case (5.25) has to be replaced by
(5.28)
d
i=1
(
p
K
x
i
(
b
,

K
))
2

C | p
K
|
2
1,

K
,
where

C is a positive constant. The proof then proceeds in the same
way as before. In order to verify (5.28), it suces to show that
_
L
=1

,

K
d
i=1
(
p
K
x
i
(
b
,

K
))
2
_
1/2
provides a norm on the quotient space

P
K
/P
0
(

K), since so does | |
1,

K
,
and we may conclude taking advantage of the equivalence of norms on
nite dimensional spaces.
For that purpose, we assume
L
=1

,

K
d
i=1
(
p
K
x
i
(
b
,

K
))
2
= 0 .
Then, the positivity of the weights
,

K
, 1 L, yields
p
K
x
i
(
b
,

K
) = 0 , 1 i d , 1 L .
But for each i {1, ..., d}
p
K
x
i
P
q1
(

K) ,
and hence,
p
K
x
i
0 ,
since it vanishes on a P
q1
(

K)-unisolvent subset.
With Theorem 5.2 at hand, we may apply Strangs rst lemma. If
the solution u V of (5.9) satises u V H
k+1
(), k lN,, for the
approximation error we get
(5.29) inf
v
h
V
h
u v
h
1,
u
h
u
1,
C h
k
|u|
k+1,
.
Therefore, we have to provide sucient conditions ensuring that the
consistency error does not deteriorate this order of convergence, i.e.,
8 Ronald H.W. Hoppe
we are looking for consistency error estimates of the form
(5.30) sup
w
h
V
h
|a(v
h
, w
h
) a
h
(v
h
, w
h
)|
w
h
V
C(a
ij
, u) h
k
,
(5.31) sup
w
h
V
h
|(w
h
)
h
(w
h
)|
w
h
V
C() h
k
,
where C(a
ij
, u) and C() are positive constants, independent of h H.
The following auxiliary result will be used to establish such consistency
error estimates.
Lemma 5.2 Generalized Leibniz formula
Assume that lR
d
is a Lipschitz domain and u H
m
() and
v W
m,
() for some m lN
0
. Then, there exists a positive constant
C(m, d) such that
(5.32) |uv|
m,
C(m, d)
m
j=0
|u|
mj,
|v|
j,,
.
Proof. The proof is left as an exercise.
Theorem 5.3 Consistency error estimate I
Let

Q
K

,

K
exists k lN such that
(5.33)

P
K
= P
k
(

K) ,
(5.34)

E
K
( ) = 0 , P
2k2
(

K) .
Then, there exists a positive constant C, independent of h H, such
that for all a W
k,
() and all p, q P
k
(K)
(5.35) |E
K
(a
q
x
i
p
x
j
)| C h
k
K
a
k,,K
|p|
1,K
q
k,K
.
Proof. Since
q
x
i
,
p
x
j
P
k1
(K), it suces to prove (5.35) for
E
K
(avw) , a W
k,
(K) , v, w P
k1
(K) .
In view of (5.18) in Lemma 5.1, we have
E
K
(avw) = det(B
K
)

E
K
( a v w)
where
a W
k,
(

K) , v, w P
k1
(

K) .
(i) Since a v W
k,
(

K), we rst provide an estimate for
K
( w) , W
k,
(

K) , w P
k1
(

K) .
By direct estimation and using the equivalence of norms on P
k1
(

K),
we obtain
|
K
( w)| = |
_
K
w d x
L
=1

,

K
( w)(
b
,

K
)|

C | w|
0,,

K

C | |
0,,

K
| w|
0,,

K

C
k,,

K
| w|
0,

K
,
which proves the continuity of the functional

E
K
( w) : W
k,
(

K) lR
given by (

E
K
( w))( ) :=

E
K
( w) with
(5.36)
K
( w)

C w
0,

K
.
Since P
k1
(

K) Ker(

E
K
( w)), using (5.36) in the Bramble-Hilbert
lemma, we get
(5.37) |
K
( w)|

C | |
k,,

K
w
0,

K
.
(ii) We now consider the case = a v where a W
k,
(

K) and v
P
k1
(

K). In view of the generalized Leibniz formula (cf. Lemma
5.2) and taking | v|
k,,

K
= 0 into account, we obtain
(5.38) |
K
( a v w)|

C
_
k1
j=1
| a|
kj,,

K
| v|
j,

K
_
w
0,

K
.
Now, Theorem 4.3 tells us
(5.39) | a|
kj,,

K

C h
kj
K
|a|
kj,,K
, 0 j k 1 ,
(5.40) | v|
j,

K

C |det(B
K
)|
1/2
h
j
K
|v|
j,K
, 0 j k 1 ,
(5.41) w
0,

K

C |det(B
K
)|
1/2
w
0,K
.
Using (5.38)-(5.41), we nally obtain
|E
K
(avw)| C h
k
K
_
k1
j=0
|a|
kj,,K
|v|
j,K
_
w
0,K

C h
k
K
a
k,,K
v
k1,K
w
0,K
.
Theorem 5.4 Consistency error estimate II
10 Ronald H.W. Hoppe
Let

Q
K

,

K
exists k lN such that
(5.42)

P
K
= P
k
(

K) ,
(5.43)

E
K
( ) = 0 , P
2k2
(

K) .
Assume that g W
k,p
() and p P
k
(K) and suppose that k lN
satises
(5.44) k >
d
p
.
Then, there exists a constant C lR
+
, independent of h H, such
that for all K T
h
(5.45) |E
K
(gp)| C (meas(K))
(
1
2
1
p
)
h
k
K
g
k,p,K
p
1,K
.
Proof. In view of (5.18) in Lemma 5.1, we have
E
K
(gp) = det(B
K
)

E
K
( g p)
where
g W
k,p
(

K) , p P
k
(

K) .
Denoting by

Q : L
2
(

K) P
1
(

K) the orthogonal projection of
L
2
(

K) onto P
k
(

K), we split

E
K
( g p) according to
(5.46)

E
K
( g p) =

E
K
( g

Q p) +

E
K
( g( p

Q p))
and estimate both parts separately. (Note that a direct estimation
(without such a splitting) would result in a non optimal estimate).
(i) Estimation of the rst term in (5.46)
In view of (5.44) and the Sobolev imbedding theorem, the space
W
k,p
(

K) is continuously imbedded in C
0
(

K). Hence, for

W
k,p
(

K)
|
K
(
)|

C |
|
0,,

K

C
k,p,

K
.
Since
K
(
) = 0 ,

P
k1
(

K) ,
the Bramble-Hilbert lemma infers
(5.47) |
K
(
)|

C |
|
k,p,

K
.
Now, for

= g

Q p, observing that
|
Q p|
,,

K
= 0 , 2 ,
the generalized Leibniz formula implies
| g

Q p|
k,p,

K

C
_
| g|
k,p,

K
|
Q p|
0,,

K
+ | g|
k1,p,

K
|
Q p|
1,,

K
_
.
Taking advantage of the equivalence of norms on P
1
(

K), we get
(5.48) | g

Q p|
k,p,

K

C
_
| g|
k,p,

K
|
Q p|
0,

K
+ | g|
k1,p,

K
|
Q p|
1,

K
_
.
Since

Q is an orthogonal projection, we have
(5.49) |
Q p|
0,

K
| p|
0,

K
.
Moreover, since

Q p = p , p P
0
(

K), the Bramble-Hilbert lemma
gives
| p

Q p|
1,

K

C | p|
1,

K
,
whence
(5.50) |
Q p|
1,

K
| p

Q p|
1,

K
+| p|
1,

K
(1 +

C) | p|
1,

K
.
Combining (5.47)-(5.50), we obtain
(5.51)

E
K
( g

Q p)|

C
_
| g|
k,p,

K
| p|
0,

K
+| g|
k1,p,

K
| p|
1,

K
_
.
(ii) Estimation of the second term in (5.46)
Since p

Q p = 0 for p P
1
(

K), we may assume k 2.
We further choose r [1, ) large enough so that the mappings
W
k,p
(

K) W
k1,r
(

K) , W
k1,r
(

K) C
0
(

K)
represent continuous imbeddings.
Case 1 1 p < d
We may choose
1
r
=
1
p

1
d
so that by the Sobolev imbedding the-
orem the mapping W
1,p
(

K) L
r
(

K) and hence, also the mapping
W
k,p
(

K) W
k1,r
(

K) is a continuous imbedding.
Moreover, in view of (5.44), we have
k 1
d
r
= k
d
p
> 0 .
Then, the Sobolev imbedding theorem also guarantees that the
mapping W
k1,r
(

K) C
0
(

K) is a continuous imbedding.
Case 2 d < p
In this case, due to the Sobolev imbedding theorem the mapping
W
1,p
(

K) L
r
(

K) is a continuous imbedding for all r [1, ], if
d < p, and for all r [1, ), if d = p. It follows that the same holds
true for the mapping W
k,p
(

K) W
k1,r
(

K).
Further, choosing r so large that
k 1
d
r
> 0 ,
the Sobolev imbedding theorem implies that W
k1,r
(

K) C
0
(

K)
is a continuous imbedding.
Now, in either case we obtain
|
K
( g( p

Q p))|

C | g( p

Q p)|
0,,

K

C | g|
0,,

K
| p

Q p|
0,,

K

C g
k1,r,

K
| p

Q p|
0,,

K
.
Consequently, the linear functional

E
K
( p) : W
k1,r
(

K) lR given by
(

E
K
( p))( g) :=

E
K
( g( p

Q p)) is continuous with
K
( p)

C p

Q p
0,,

K
.
Since
K
( p) = 0 , p P
k2
(

K) ,
by the Bramble-Hilbert lemma
(5.52) |
K
( g( p

Q p))|

C | g|
k1,r,

K
| p

Q p|
0,

K
.
Moreover,

Q p = p, p P
0
(

K), and hence,
(5.53) | p

Q p|
o,

K

C | p|
1,

K
.
In order to estimate g|
k1,r,

K
in (5.52), we note that in view of the
Sobolev imbedding theorem the mapping
W
1,p
(
k) L
r
(
k)
is a continuous imbedding, whence
|
f|
0,r,

K

C
_
|
f|
0,p,

K
+ |
f|
1,p,

K
_
,
and hence,
(5.54) | g|
k,r,

K

C
_
| g|
k1,p,

K
+ | g|
k,p,

K
_
.
Using (5.53) and (5.54) in (5.52), we get
(5.55) |
K
( g( p

Q p))|

C
_
| g|
k1,p,

K
+| g|
k,p,

K
_
| p|
1,

K
.
Finally, combining (5.46),(5.51),(5.55) and observing that Theorem 4.3
infers
| g|
kj,p,

K

C (det(B
K
))
1/p
h
kj
K
|g|
kj,p,K
, 0 j 1 ,
| p|
j,

K

C (det(B
K
))
1/2
h
j
K
|p|
j,K
, 0 j 1 ,
we arrive at the assertion.
After having estimated both the approximation and the consistency
errors, we are now in a position to establish an a priori estimate of
the global discretization error in the H
1
()-norm.
For this purpose, we provide the following auxiliary result.
Lemma 5.3 Generalized Schwarz inequality for sums
Assume a
i
, b
i
, c
i
lR, 1 i m, and r
1
, r
2
, r
3
lN satisfying
1
r
1
+
1
r
2
+
1
r
3
= 1. Then, there holds
(5.56)
m
i=1
|a
i
b
i
c
i
|
_
m
i=1
|a
i
|
r
1
_
1/r
1
_
m
i=1
|b
i
|
r
2
_
1/r
2
_
m
i=1
|c
i
|
r
3
_
1/r
3
.
Theorem 5.5 A priori estimate in the H
1
()-norm
In addition to the assumptions (A1),(A2),(A3), suppose that for
some integer k lN we have
(5.57)

P = P
k
(

K) ,
and that
(5.58) H
k+1
(

K) C
s
(

K)
is a continuous imbedding, where s lN
0
is the maximal order of partial
derivatives in the denition of the set

K
of degrees of freedom.
Moreover, let

Q
K
be a quadrature formula with respect to the reference
element

,

K
, 1 L, satisfying
(5.59)

E
K
( ) = 0 , P
2k2
(

K) .
As far as the data of the model problem (5.9) are concerned, we re-
quire the existence of some p 2 with k >
d
p
such that the following
regularity assumptions are met
(5.60) a
ij
W
k,
() , 1 i, j d , f W
k,p
() .
We denote by u V := H
1
0
() the solution of (5.9) and by u
h

V
h
, h H the solution of its nite element approximation (5.21), and
we assume
(5.61) u V H
k+1
() .
Then, there exists a positive constant C lR, independent of h
H, such that the following a priori error estimate for the global
discretization error holds true
(5.62)
uu
h
1,
C h
k
_
|u|
k+1,
+
d
i,j=1
a
ij
k,,
u
k+1,
+ f
k,p,
_
.
Proof. The assumptions apply that the family (a
h
(, ))
hH
of ap-
proximate bilinear forms a
h
(, ) : V
h
V
h
lR, h H, is uniformly
V
h
-elliptic. We may thus apply Strangs rst lemma and provide
appropriate estimates of the approximation error and the consis-
tency error.
(i) Approximation error
In view of the regularity assumption (5.61), it follows from Theorem
4.5 that
(5.63) u
h
u
1,
C h
k
|u|
k+1,
.
(ii) Consistency error, Part I
By Theorem 5.3 and the Cauchy-Schwarz inequality, for w
h
V
h
we
obtain
|a(
h
u, w
h
) a
h
(
h
u, w
h
)| (5.64)
KT
h
d
i,j=1
|E
K
(a
ij
K
u
x
i
w
h
x
j
)|
C
KT
h
h
k
K
_
d
i,j=1
a
ij
k,,K
_

K
u
k,K
|w
h
|
1,K

C h
k
_
d
i,j=1
a
ij
k,,K
_ _

KT
h
K
u
2
k,K
_
1/2
|w
h
|
1,
.
Applying Theorem 4.5 once more, we get
_

KT
h
K
u
2
k,K
_
1/2
(5.65)
u
k,K
+
_

KT
h
u
K
u
2
k,K
_
1/2
u
k,K
+ C h |u|
k+1,
C u
k+1,
.
Using (5.65) in (5.64) yields
sup
w
h
V
h
|a(
h
u, w
h
) a
h
(
h
u, w
h
)|
w
h
1,
(5.66)
C h
k
d
i,j=1
a
ij
k,,
u
k+1,
. (5.67)
(ii) Consistency error, Part II
Using Theorem 5.4, for w
h
V
h
it follows that
|(w
h
)
h
(w
h
)|
KT
h
|E
K
(fw
h
)| (5.68)
C
KT
h
(meas(K))
(
1
2
1
p
)
h
k
K
f
k,p,K
w
h
1,K
.
Applying the generalized Schwarz inequality (Lemma 5.3) to the
last sum in (5.68) with
m := card(T
h
) ,
1
r
1
:=
1
2

1
p
, r
2
:= p , r
3
:= 2 ,
results in
|(w
h
)
h
(w
h
)|
KT
h
|E
K
(fw
h
)| (5.69)
C h
k
(meas())
(
1
2
1
p
)
f
k,p,
w
h
1,
.
We thus get
(5.70) sup
w
h
V
h
|(w
h
)
h
(w
h
)|
w
h
1,
C h
k
(meas())
(
1
2
1
p
)
f
k,p,
.
Finally, combining (5.63),(5.66) and (5.70), the assertion is a direct
consequence of Strangs rst lemma.
5.2.2 Special quadrature formulas
We rst consider quadrature formulas for simplicial triangula-
tions.
Lemma 5.4 Quadrature formulas for simplicial triangulations
Let

K be a non degenerate 2-simplex with vertices a
i
, 1lei 3, and
denote by a
ij
, 1 i < j 3, the midpoints of the edges and by a
123
the center of gravity. then, there holds:
(i) The quadrature formula
(5.71)
_
K
( x) d x meas(

K) ( a
123
)
is exact for polynomials p P
1
(

K), i.e.,
(5.72)

E( p) = 0 , p P
1
(

K) .
(ii) The quadrature formula
(5.73)
_
K
( x) d x
1
3
meas(

K)
1i<j3
( a
ij
)
2
(

K), i.e.,
(5.74)

E( p) = 0 , p P
2
(

K) .
(iii) The quadrature formula
(5.75)
_
K
( x) d x
1
60
meas(

K)
_
3
3
i=1
( a
i
) + 8
1i<j3
( a
ij
) + 27 ( a
123
)
_
3
(

K), i.e.,
(5.76)

E( p) = 0 , p P
3
(

K) .
Proof. Let
i
( x), 1 i 3, be the barycentric coordinates of x

K.
Then, for
i
lN
0
, 1 i 3, there holds
(5.77)
_
K
3
i=1
i
i
( x) d x = meas(

K)
d!
3
i=1
i
!
(
3
i=1
i
+d)!
.
The proof of (5.77) is left as an exercise.
We next consider quadrature formulas for rectangular triangula-
tions.
Lemma 5.5 Quadrature formulas for rectangular triangula-
tions
Let

K := [0, 1]
2
with vertices
a
1
:= (0, 0) , a
2
:= (1, 0) , a
3
:= (1, 1) , a
4
:= (0, 1)
and center a
s
:= (1/2, 1/2). Then, there holds (i) The quadrature
formula
(5.78)
_
K
( x) d x
1
4
4
i=1
( a
i
)
is exact for polynomials p Q
1
(

K), i.e.,
(5.79)

E( p) = 0 , p Q
1
(

K) .
(ii) The quadrature formula
(5.80)
_
K
( x) d x ( a
s
)
1
(

K), i.e.,
(5.81)

E( p) = 0 , p Q
1
(

K) .
(iii) Let k lN
0
and denote by
b
i
(0, 1), 1 i k+1, the nodes and
by
i
, 1 i k+1, the weights of the Gauss-Legendre quadrature
formula
1
_
0
( x) d x
k+1
i=1

i
) (
b
i
) ,
which is exact for polynomials p P
2k+1
(
k.
Then, the quadrature formula
(5.82)
_
K
( x) d x
i
1
,i
2
{1,...,k+1}

i
1

i
2
(
b
i
1
,
b
i
2
)
2k+1
(

K), i.e.,
(5.83)

E( p) = 0 , p Q
2k+1
(

K) .
We note that the quadrature formulas (5.71),(5.73),(5.75) and (5.78),
(5.80),(5.82) can be easily generalized to 3D.
5.3 Strangs second lemma
In this section, we will derive an abstract error estimate, known as
Strangs second lemma, that is applicable for nite element approx-
imations of variational equations
(5.84) a(u, v) = (v) , v V ,
where the nite element spaces V
h
, h H, are no longer subspaces of
the underlying innite dimensional function space V .
We assume that (V
h
)
hH
is a family of nite dimensional linear spaces
and suppose that
h
, h T
h
, provides a mesh-dependent norm on
V
h
+V . We further suppose that the linear functional is well dened
on V
h
+V .
We consider an associated family of approximate bilinear forms
(5.85) a
h
(, ) : (V
h
+V ) (V
h
+V ) lR .
We suppose that the family (a
h
(, ))
hH
is uniformly V
h
-elliptic in
the sense that there exists a positive constant , independent of h H,
such that
(5.86) a
h
(v
h
, v
h
) v
h
2
V
h
, v
h
V
h
, h H .
We further suppose that the family (a
h
(, ))
hH
is uniformly bounded
on V
h
+ V in the sense that there exists a positive constant

M, inde-
pendent of h H, such that
(5.87) |a
h
(u, v)|

M u
h
v
h
, u, v V
h
+V , h H .
We approximate (5.84) by the nite dimensional variational equations
(5.88) a
h
(u
h
, v
h
) = (v
h
) , v
h
V
h
, h H .
Strangs second lemma may be viewed as another generalization of
Ceas lemma.
Theorem 5.6 Strangs second lemma
Assume that (a
h
(, ))
hH
is a uniformly bounded and uniformly
V
h
-elliptic family of approximate bilinear forms and suppose that u
V and u
h
V
h
, h H, are the unique solutions of the variational
equations (5.84) and (5.88), respectively.
Then, there exists a positive constant C lR, independent of h H,
such that
(5.89) uu
h
h
C
_
inf
v
h
V
h
uv
h
h
+ sup
w
h
V
h
|a
h
(u, w
h
) (w
h
)|
w
h
h
_
.
Proof. Taking advantage of the uniform V
h
-ellipticity and observ-
ing (5.88), for an arbitrary v
h
V
h
we obtain
u
h
v
h
2
h
a
h
(u
h
v
h
, u
h
v
h
) =
= a
h
(u v
h
, u
h
v
h
) +
_
(u
h
v
h
) a
h
(u, u
h
v
h
)
_
.
By the uniform boundedness (5.87) we get
u
h
v
h
h
a
h
(u
h
v
h
, u
h
v
h
)

M u v
h
h
+
|(u
h
v
h
) a
h
(u, u
h
v
h
)|
u
h
v
h

M u v
h
h
+ sup
w
h
V
h
|(w
h
) a
h
(u, w
h
)|
w
h
h
.
In conjunction with the triangle inequality
u u
h
h
u v
h
h
+ u
h
v
h
h
,
the previous inequality gives the assertion.
As in the case of Strangs rst lemma, we see that the upper bound for
the global discretization error consists of an approximation error
(5.90) inf
v
h
V
h
u v
h
h
and the consistency error
(5.91) sup
w
h
V
h
|(w
h
) a
h
(u, w
h
)|
w
h
h
,
which can be estimated separately.
5.4 A priori error estimate for nonconforming nite
element approximations
As an example for a nonconforming nite element in the discretiza-
tion of second order elliptic boundary value problems with respect to a
family of shape-regular simplicial triangulations T
h
, h H, of the com-
putational domain lR
d
we consider the lowest order Crouzeix-
Raviart element CR
1
(K) := (K, P
K
,
K
), K T
h
, also known as
the nonconforming P1-element:
P
K
:= P
1
(K) , (5.92)
K
:= { p(m
i
(K)) , 1 i d + 1 } , p P
K
,
where m
i
(K), 1 i d + 1, are the midpoints of the edges (d = 2)
and the centers of gravity of the faces (d = 3), respectively (cf.
Figure 5.1).
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
u
u u
u
u
u u
Fig. 5.1: Nonconforming Crouzeix-Raviart element
The associated Crouzeix-Raviart nite element space CR
1
(, T
h
),
composed by the Crouzeix-Raviart elements CR
1
(K), K T
h
, is then
given by
CR
1
(, T
h
) := (5.93)
{ v
h
L
2
() | v
h
|
K
P
1
(K) , K T
h
,
v
h
is continuous in m(f), f F
h
() } ,
where m(f), f F
h
() are the midpoints (centers of gravity) of interior
edges (faces).
Obviously, in general
(5.94) CR
1
(, T
h
) H
1
() .
The subspace CR
1,0
(, T
h
) is dened by means of
CR
1,0
(, T
h
) := { v
h
CR
1
(, T
h
) | v
h
(m(f)) = 0 , f F
h
() } .
We consider the approximation of Poissons equation under homo-
geneous Dirichlet boundary conditions
(5.95) a(u, v) :=
_
u vdx = (v) :=
_
fv dx , v H
1
0
()
by
(5.96) a
h
(u
h
, v
h
) = (v
h
) , v
h
CR
1,0
(, T
h
) , h H ,
where the mesh dependent bilinear form
a
h
(, ) : H
1
0
() CR
1,0
(, T
h
) H
1
0
() CR
1,0
(, T
h
) lR
is given by
(5.97) a
h
(u, v) :=
KT
h
()
_
K
uv dx , u, v H
1
0
()CR
1,0
(, T
h
) .
We associate with (5.97) the mesh dependent norm
(5.98) v
h
:=
_
a
h
(v, v) , v CR
1,0
(, T
h
) H
1
0
() .
We will derive a quasi-optimal a priori error estimate in the
h
-
norm by an application of Strangs second lemma:
Theorem 5.7 A priori error estimate in the
h
-norm
Let u H
1
0
() H
s
() , s = 2 (d = 2) , s = 3 (d = 3), be the
solution of (5.95) and u
h
CR
1,0
(, T
h
), h H, its nonconforming
P1-approximations. Then, there exists a constant C > 0, depending
only on the shape regularity of the triangulations T
h
(), h H, such
that
(5.99) u u
h
h
C h |u|
s,
.
Proof. It is easily veried that the family (a
h
(, ))
hH
of bilinear
forms is uniformly V
h
-elliptic (V
h
:= CR
1,0
(, T
h
)). According to
Strangs second lemma we have to estimate the approximation
error
(5.100) inf
v
h
CR
1,0
(,T
h
)
u v
h
h
and the consistency error
(5.101) sup
w
h
CR
1,0
(,T
h
)
a
h
(u, w
h
) (w
h
)
w
h
h
.
(i) Estimation of the approximation error
Since S
1,0
(, T
h
) CR
1,0
(, T
h
), in view of Theorem 4.7 we have
(5.102) inf
v
h
CR
1,0
(,T
h
)
uv
h
h
inf
v
h
S
1,0
(,T
h
)
uv
h
h
C h |u|
s,
.
(ii) Estimation of the consistency error
Observing that
u(x) = f(x) f.a.a. x ,
by Greens formula we obtain for w
h
CR
1,0
(, T
h
)
L
u
(w
h
) := a
h
(u, w
h
) (w
h
) = (5.103)
=
KT
h
_
_
K
u w
h
dx
_
K
fw
h
dx
_
=
=
KT
h
_
_
K
_
u f
_
w
h
dx +
_
K
u
n
w
h
d
_
=
=
KT
h
fK
_
f
u
n
w
h
d .
Denoting by w
h
|
f
the integral mean
w
h
|
f
:=
1
meas(f)
_
f
w
h
d , f F
h
() ,
it follows that
(5.104) L
u
(w
h
) =
KT
h
fK
_
f
u
n
(w
h
w
h
|
f
) d .
Moreover, since
I
h
u
n
P
0
(K
), 1 2, we have
_
fK
I
h
u
n
(w
h
w
h
|
f
) d = 0 ,
and hence, (5.104) gives rise to
L
u
(w
h
) =
KT
h
fK
_
f
(u I
h
u)
n
(w
h
w
h
|
f
) d .
The Cauchy-Schwarz inequality yields
|L
u
(w
h
)|
KT
h
fK
_
_
f
|
(u I
h
u)
n
|
2
d
. .
:= I
1
_
f
|w
h
w
h
|
f
|
2
d
. .
:= I
2
_
1/2
.
We will estimate I
1
and I
2
separately.
(ii)
1
Estimation of I
1
Taking advantage of the ane equivalence of the Crouzeix-Raviart
elements, for the reference element

K we nd by the trace theo-
rem and the Bramble-Hilbert lemma
_

K
|
(u I
h
u)
n
|
2
d Cu I
h
u
2
2,

K
C |u|
2
s,

K
.
By a standard scaling argument
(5.105)
_
K
|
(u I
h
u)
n
|
2
d C h |u|
2
s,K
.
(ii)
2
Estimation of I
2
Again, for the reference element

K we nd by the Bramble-Hilbert
lemma
_
f
| w
h
w
h
|
f
|
2
d C| w
h
|
2
1,

K
, w
h
P
1
(

K) ,

f F(

K) ,
and by a standard scaling argument, for w
h
CR
1,0
(, T
h
) and
f F(K), K T
h
we obtain
(5.106)
_
f
|w
h
w
h
|
f
|
2
d C h |w
h
|
2
1,K
.
Using (5.105) and (5.106), we nally obtain
|L
u
(w
h
)| 3 C h
KT
h
|u|
s,K
|w
h
|
1,K

C h
_

KT
h
|u|
2
s,K
KT
h
|w
h
|
2
1,K
_
1/2
= C h |u|
s,
w
h
h
.

Chapter 5 A Priori Error Estimates For Nonconfor-Ming Finite Element Approximations 5.1 Strang's First Lemma

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Chapter 5 A Priori Error Estimates For Nonconfor-Ming Finite Element Approximations 5.1 Strang's First Lemma

Uploaded by

Copyright:

Available Formats

Chapter 5 A priori error estimates for nonconfor-

ming nite element approximations

K with positive weights

K with positive weights

K with positive weights

You might also like