operator from Theorem 4.2(ii).

We estimate
h∂n v, wiΓ h∂n v, wiΓ
k∂n vkH −1/2 (Γ) = sup ≤C sup
w∈H 1/2 (Γ)\{0} kwkH 1/2 (Γ) w∈H 1/2 (Γ)\{0} kEwkH 1 (Ω)
R R R R
Ω ∇v · ∇Ew + Ω ∆v Ew Ω ∇v · ∇W + Ω ∆v W
= C sup ≤C sup
w∈H 1/2 (Γ)\{0} kEwkH 1 (Ω) W ∈H 1 (Ω)\{0} kW kH 1 (Ω)
kvkH 1 (Ω) kW kH 1 (Ω) + k∆vkH̃ −1 (Ω) kW kH 1 (Ω)
≤ C sup
W ∈H 1 (Ω)\{0} kW kH 1 (Ω)


= C kvkH 1 (Ω) + k∆vkH̃ −1 (Ω) .

Remark 4.3 ∆v ∈ L2 (Ω) implies ∆v ∈ H̃ −1 (Ω).

5 Finite element error analysis for elliptic problems

In this section we deal with the error analysis of the finite element method. Key steps in the
error analysis are the Lax-Milgram lemma (Theorem 2.1), which proves the unique existence of
uh and its stability, and Céa’s lemma (Theorem 3.2) proving
Ca
ku − uh k ≤ ku − vk ∀v ∈ Vh .
α
Here, several assumptions are needed, in particular the boundedness of a (with bound Ca ) and
its V -ellipticity (with ellipticity constant α). Therefore, to bound the error in the energy norm
(or the norm of V ) we only need to select an appropriate function v ∈ Vh for which we are
able to further estimate ku − vk. If Vh consists of continuous, piecewise linear functions then a
standard candidate is the piecewise linear interpolant Ih u ∈ Vh (defined below). First, in §5.1,
we deal with approximation theory in a more general and abstract form. Then, in §5.2, we apply
the approximation results to the finite element method.

5.1 Approximation theory

Definition 5.1 Let (X, k · kX ), (Y, k · kY ) be normed linear spaces, and A ∈ L(X, Y ), where
L(X, Y ) denotes the space of bounded linear operators X → Y . Then, A is compact if and only
if (Axn )n∈IN ⊂ Y has a convergent subsequence for any bounded sequence (xn )n∈IN ⊂ X.

This can be equivalently formulated as: A is compact if and only if every bounded subset of
X is mapped to a relatively compact subset of Y .

Proposition 5.1 (Rellich’s embedding theorem) Let Ω be a Lipschitz domain. Then for any
t > s, the injection i : H t (Ω) → H s (Ω) is compact.

23
Proposition 5.2 (Sobolev’s embedding theorem) Let Ω be a Lipschitz domain in IRn . Then,
the injection i : H n/2+ε (Ω) → C 0 (Ω̄) is continuous for all ε > 0, that is,

sup |u(x)| ≤ Cε kukH n/2+ε (Ω) for all u ∈ H n/2+ε (Ω).

x∈Ω

Remark 5.1 A simple argument for the influence of the dimension n is the following: Using
the Fourier transform, we see for u ∈ C0∞ (IRn ) that
Z Z
|u(x)| ≤ |û(ξ)| dξ = (1 + |ξ|2 )−s/2 (1 + |ξ|)s/2 |û(ξ)| dξ
In
R In
R
Z 1/2
2 −s
≤ (1 + |ξ| ) dξ kukH s (IRn ) ,
In
R
2 −s dξ <
R
where the last inequality follows from the Cauchy-Schwarz inequality. Thus, R I n (1 + |ξ| )
∞ will be sufficient. As the integrand is bounded on every bounded set, we only need to study
the behaviour as |ξ| → ∞. Transforming to polar coordinates and choosing some r ∗ > 0,
Z Z ∞ Z ∞
2 −s −2s n−1
(1 + |ξ| ) dξ ∼ r r dr = r n−2s−1 dr.
In
R r∗ r∗

The last integral is finite if and only if n−2s−1 < −1. This corresponds exactly to the condition
s > n2 .

Lemma 5.1 Let Ω ⊂ IR2 be a Lipschitz domain, t ≥ 2 integer, s = t(t+1)2 , and {z1 , z2 , . . . , zs } ⊂
Ω be given points such that the interpolation operator I : H t (Ω) → Pt−1 is well-defined. Here,
Pt−1 are the polynomials of degree up to t − 1. Then, there exists C ≥ 0 such that

ku − IukH t (Ω) ≤ C|u|H t (Ω) for all u ∈ H t (Ω).

and |||v||| := |v|H t (Ω) + si=1 |v(zi )| are equivalent norms.

P
Proof. We first prove that kvkH t (Ω)
Then it follows that
s
!
X
ku − IukH t (Ω) ≤ C |||u − Iu||| = C |u − Iu|H t (Ω) + |u(zi ) − Iu(zi )|
i=1
= C|u − Iu|H t (Ω) = C|u|H t (Ω) ,

since the tth derivatives of Iu ∈ Pt−1 vanish.

1. As t ≥ 2 we see that the embedding H t (Ω) → H 2 (Ω) is continuous and by Proposition 5.2
we have that H 2 (Ω) → C 0 (Ω̄) is continuous. Therefore, the injection H t (Ω) → C 0 (Ω̄) is
continuous. Thus, |v(zi )| ≤ CkvkH t (Ω) , i = 1, . . . , s, and
s
X
|||v||| = |v|H t (Ω) + |v(zi )| ≤ (1 + sC)kvkH t (Ω) for all v ∈ H t (Ω).
i=1

24
 2. Assume that kvkH t (Ω) ≤ C |||v||| for every v ∈ H t (Ω) is false for any constant C > 0.
Then, there exists a sequence (vn )n∈IN ⊂ H t (Ω) such that kvn kH t (Ω) = 1 and |||vn ||| ≤ n1
for all n ∈ IN. We see that (vn ) is bounded in H t (Ω), so by Proposition 5.1 there is a
subsequence of (vn ) which converges in H t−1 (Ω). We assume without loss of generality
that this subsequence is (vn ). In particular, it follows that (vn ) is a Cauchy sequence in
H t−1 (Ω) and thus, since |vn |H t (Ω) ≤ |||vn ||| → 0 for n → ∞,

kvk − vl k2H t (Ω) ≤ kvk − vl k2H t−1 (Ω) + (|vk |H t (Ω) + |vl |H t (Ω) )2 → 0 for k, l → ∞.

Therefore, (vn ) is a Cauchy sequence in H t (Ω) and by completeness there exists v ∗ ∈ H t (Ω)
such that vn → v ∗ in H t (Ω) for n → ∞. By the continuity of the norms it follows from
kvn kH t (Ω) = 1 that kv ∗ kH t (Ω) = 1, and from |||vn ||| ≤ n1 that |||v ∗ ||| = 0 since, by the first
part,

|||v ∗ ||| ≤ |||v ∗ − vn ||| + |||vn ||| ≤ Ckv ∗ − vn kH t (Ω) + |||vn ||| → 0 as n → ∞.

By definition of |||·||| it follows that |v ∗ |H t (Ω) = 0, that is, v ∗ ∈ Pt−1 , and |v ∗ (zi )| = 0,
i = 1, . . . , s. Therefore, v ∗ vanishes at t(t+1)
2 distinct points. It follows that v ∗ = 0 and
this is a contradiction to kv ∗ kH t (Ω) = 1.
Therefore, there exists a constant C such that kvkH t (Ω) ≤ C |||v|||.

Theorem 5.1 (Bramble-Hilbert Lemma) Let Ω ⊂ IR2 be a Lipschitz domain, and t ≥ 2 integer.
For a normed linear space Y let L ∈ L(H t (Ω), Y ).
If Pt−1 ⊂ ker L then there exists a constant C ≥ 0 such that

kLvkY ≤ C|v|H t (Ω) for all v ∈ H t (Ω).

Proof. As L is bounded and linear there exists D ≥ 0 such that kLvkY ≤ DkvkH t (Ω) . Let
I : H t (Ω) → Pt−1 be an interpolation operator as in Lemma 5.1. Then, Iv ∈ Pt−1 ⊂ ker L for
all v ∈ H t (Ω) and

kLvkY = kL(v − Iv)kY ≤ Dkv − IvkH t (Ω) ≤ CD|v|H t (Ω)

by Lemma 5.1. 2

25
5.2 Finite element error estimate for elliptic problems
We deal with the case V = H 1 (Ω) and Vh = {v ∈ V : v|K ∈ Pt−1 (K) ∀K ∈ Th } where
Th = {K} is a triangulation of Ω, which is assumed to be polygonal (so that it can be discretised
by triangular meshes). Here, Pt−1 (K) denotes the space of polynomials of degree t − 1 on K.
The mesh needs to satisfy certain conditions. We define
hk = diameter of K = length of longest side of K,
ρk = diameter of the largest circle in K,
h = max hk
K∈Th

and require that there exists β > 0 which is independent of h such that
ρk
≥β ∀K ∈ Th . (5.1)
hk
This means that the elements K ∈ Th are not too thin, i.e. the interior angles of K are not too
small (they are bounded from below by a positive constant). One also says that the elements of
Th , or Th , are shape regular. Since we are interested in a sequence of meshes {Th } such that we
can study the behaviour of the finite element error ku − uh k for a sequence of mesh sizes {h},
the constant β in (5.1) must be independent of h.
We now apply the Bramble-Hilbert Lemma to prove a piecewise polynomial approximation
result.

Theorem 5.2 For a Lipschitz domain Ω ⊂ IR2 with polygonal boundary and a given integer
t ≥ 2 let {T : T ∈ Th } be a shape regular triangulation of Ω.
Then, for a piecewise polynomial interpolation operator Ih of degree t − 1 (piecewise with
respect to Th ) there holds
 1/2
X
 ku − Ih uk2H m (T )  ≤ Cht−m |u|H t (Ω) for all u ∈ H t (Ω) and all 0 ≤ m ≤ t.
T ∈Th

Here, the constant C is independent of h and u.

Proof. The idea of the proof is to transform to the reference element T̂ , make a transition
from H m to H t , and transform back. The transformations give the required powers of h since
the Bramble-Hilbert Lemma gives the transition to a semi-norm on T̂ .
By the assumption of shape regularity it is enough to consider the case that Th is congruent
to T̂ . Then we can assume without loss of generality that Th = hT̂ := {(x1 , x2 ) : 0 ≤ x1 , x2 ≤
h, x1 + x2 ≤ h}. For v ∈ H t (Th ) we define v̂ ∈ H t (T ) by v̂(x1 , x2 ) := v(hx1 , hx2 ). For a
multi-index α = (α1 , α2 ) of order |α| = α1 + α2 with non-negative integers α1 , α2 let D α denote
the partial derivative operator defined by
∂ |α|
D α v(x1 , x2 ) := v(x1 , x2 ).
∂xα1 1 ∂xα2 2

26
We see that D α v = h−α Dα v̂ for all multi-indices α with |α| ≤ t. Thus,
XZ X Z
2 α 2 −2l 2
|v|H l (Th ) = (D v) dx = h h (D α v̂)2 dξ = h2−2l |v̂|2H l (T̂ ) .
|α|=l Th |α|=l T̂

Since the transform Iˆ of the interpolation operator Ih is again an interpolation operator, we can
transform to the reference element, apply the Bramble-Hilbert Lemma and transform back to
obtain

kv − Ih vk2H m (Th ) ≤ Ch2−2m kv̂ − Iˆv̂k2H m (T̂ ) ≤ Ch2−2m kv̂ − Iˆv̂k2H t (T̂ )
≤ Ch2−2m |v̂|2H t (T̂ ) = Ch2−2m h2t−2 |v|2H t (Th ) = Ch2(t−m) |v|2H t (Th ) .

Summing up this yields

X X
kv − Ih vk2H m (Th ) ≤ Ch2(t−m) |v|2H t (Th ) = Ch2(t−m) |v|2H t (Th ) .
T ∈Th T ∈Th

Remark 5.2 Note that in Theorem 5.2 we cannot write ku− Ih ukH m (Ω) in general since u− Ih u
might not be in H m (Ω). The operator Ih represents only a piecewise interpolation from which,
in general, no global regularity properties follow.

Let Ni , i = 1, . . . , M , be the nodes of Th . For a continuous function u ∈ C 0 (Ω̄) we now

consider the piecewise linear interpolant (again using the same operator symbol Ih ) Ih u ∈ Vh by

Ih u(Ni ) = u(Ni ), i = 1, . . . , M. (5.2)

Note that on K ∈ Th , Ih u is the linear interpolant of u.

Corollary 5.1 Let Ω ⊂ IR2 be a polygon with a quasi-uniform, regular and shape-regular mesh,
and Ih be the piecewise linear interpolation operator (piecewise with respect to the triangulation)
with respect to the vertices of the mesh.
Then,
ku − Ih ukH 1 (Ω) ≤ Ch|u|H 2 (Ω) for all u ∈ H 2 (Ω).

Proof. As Ih interpolates at the vertices of the mesh, Ih u is continuous and piecewise linear,
i.e. Ih u ∈ H 1 (Ω). An application of Theorem 5.2 proves
X
ku − Ih uk2H 1 (Ω) = ku − Ih uk2H 1 (T ) ≤ Ch2 |u|2H 2 (Ω) .
T ∈Th

27
 2

Now we are in a position to present an a priori error estimate for the finite element method
dealing with elliptic problems of second order. Assume that we are solving a variational problem
in V = H01 (Ω) (Ω ⊂ IR2 is a Lipschitz continuous polygonal domain),

u∈V : a(u, v) = L(v) ∀v ∈ V, (5.3)

where a is a continuous, V -elliptic bilinear form, and L is a continuous linear form on V . We

then consider the finite element approximation uh to u defined by

uh ∈ Vh : a(uh , v) = L(v) ∀v ∈ Vh . (5.4)

Selecting any finite-dimensional subspace Vh ⊂ V there holds Céa’s lemma. In particular,

selecting Vh to be the space of continuous, piecewise linear functions defined on a mesh Th
satisfying the shape-regularity condition (5.1) there holds (applying Céa’s lemma)

Ca
ku − uh kH 1 (Ω) ≤ ku − Ih ukH 1 (Ω) . (5.5)
α
Here, Ih is the interpolation operator defined in (5.2).
Therefore, applying the results from §5.1, in particular Corollary 5.1, we conclude that there
holds the following a priori error estimate.

Theorem 5.3 (a priori error estimate)

Assume that the solution u of (5.3) satisfies u ∈ H 2 (Ω) and that uh ∈ Vh is the finite element
approximation defined by (5.4) (using piecewise linear functions on a shape regular mesh). Then
there exists a constant C > 0 which is independent of h such that

ku − uh kH 1 (Ω) ≤ C h |u|H 2 (Ω) . (5.6)

This means that uh converges linearly in h to u in the H 1 (Ω)-norm.

28