Error analysis

Problem
Find u ∈ V := H10 (Ω), such that
! !
a(u, v) := ∇u · ∇v dx = fv dx =: !(v) ∀v ∈ V.
Ω Ω

Method
Find uh ∈ Vh , such that

a(uh , vh ) = !(vh ) ∀vh ∈ Vh

Recalling Friedrichs’ inequality, V = H10 (Ω) equipped with

$v$V := $∇v$L2 (Ω) ,

is a Hilbert space. Natural energy norm, on which we shall perform error analysis.

115
Error analysis
Subtracting (Method) from (Problem with v = vh ∈ Vh ), we get:

Galerkin Orthogonality: a(u − uh , vh ) = 0 ∀vh ∈ Vh

We have then, respectively,

$u − uh $2V = a(u − uh , u − uh )
= a(u − uh , u − uh ) + a(u − uh , vh − uh )
" #$ %
∈Vh

=a(u − uh , u − vh )
≤$u − uh $V $u − vh $V .

Therefore, $u − uh $V ≤ $u − vh $V for all vh ∈ Vh , and the following is proven.

Theorem (Cèa’s Lemma)

Let u ∈ V solve Problem and uh solve Method. Then,

$u − uh $V = min $u − vh $V .
vh ∈Vh

116
Interpolant

Definition
Let v ∈ C(Ω̄) and Ṽh a finite element space with respect to a mesh T with
Lagrange basis {φi }N total
i=1 , corresponding to the mesh nodes {x i } Ntotal
i=1 (internal and
boundary), so that dim(Ṽh ) = Ntotal . (If Ṽh ≡ Vh ⊂ H10 (Ω), we can simply ignore
the boundary nodes.) We define the interpolant vI ∈ Ṽh of v by
N
&total

vI := v(xi )φi .
i=1

Remark
In the study of Sobolev Spaces, one soon comes across the, somewhat surprising
at first sight, fact that H1 (Ω) is not a subset of C(Ω) when Ω ⊂ Rd domain, for
d ≥ 2. Hence the interpolant above is not well-defined for H1 functions.

117
Interpolation error

Lemma (Interpolation estimates, see, e.g., B&S)

Let Ω ⊂ Rd polygonal/polyhedral domain subdivided through the mesh T into
triangular/tetrahedral elements and let Ṽh the respective space of element-wise
linear functions subordinate to T .
Let v ∈ H2 (Ω) and vI ∈ Ṽh its interpolant. Then, we have
'& ( 12
$∇(v − vI )$L2 (Ω) ≤ Capp,H1 h2T |v|2H2 (T) ,
T∈T

and
'& ( 12
$v − vI $L2 (Ω) ≤ Capp,L2 h4T |v|2H2 (T) ,
T∈T

for Capp,H1 > 0 dependent on the maximum angle in T and Capp,L2 > 0
)* + 12
α 2
depending only on the geometry of Ω; here |v|H2 (T) := |a|=2 $D v$L2 (T) .

118
Error Analysis

Theorem (Convergence in the V-norm)

If u ∈ H10 (Ω) ∩ H2 (Ω) is the solution to Problem and uh ∈ Vh the solution to
Method, then
'& ( 12
$u − uh $V ≤ Capp,H1 h2T |u|2H2 (T) ≤ Capp,H1 hmax |u|H2 (Ω) ,
T∈T

for hmax := maxT∈T hT .

1st order convergence is reasonable, as the $·$V -norm involves first

derivatives

Can we do better in terms of convergence rate in a different norm?

119
Error analysis in the L2 -norm

Theorem (Convergence in the L2 -norm)

Let u ∈ H10 (Ω) ∩ H2 (Ω) be the solution to Problem. Moreover, let Ω ⊂ Rd be such
that the elliptic regularity estimate

$u$H2 (Ω) ≤ Cell−rec $∆u$L2 (Ω)

holds. (For instance Ω being polygonal/hedral and convex, or Ω having a

C2 -surface as boundary, suffices to ensure the last estimate.)
If uh ∈ Vh is the solution to Method, then we have the error bound
'& ( 12
$u − uh $L2 (Ω) ≤ Chmax h2T |u|2H2 (T) ≤ CL2 h2max |u|H2 (Ω) ,
T∈T

for hmax := maxT∈T hT and CL2 := C2app,H1 Cell−rec .

120
Error analysis in the L2 -norm
Proof. For the proof, we will employ the, so-called, Aubin-Nitsche duality
argument.

Consider the (dual) problem

−∆z = u − uh in Ω, z=0 on ∂Ω.

Since Ω such that the elliptic regularity estimate holds, we deduce

$z$H2 (Ω) ≤ Cell−rec $∆z$L2 (Ω) = Cell−rec $u − uh $L2 (Ω) ($)

Then, !
$u − uh $2L2 (Ω) = (u − uh )(−∆z) dx
!Ω
= ∇(u − uh ) · ∇z dx
!Ω
= ∇(u − uh ) · ∇(z − zI ) dx,
Ω

from Galerkin Orthogonality, since the interpolant zI ∈ Vh ; recall (u − uh )|∂Ω = 0.

121
Error analysis in the L2 -norm
Employing Cauchy-Schwarz inequalities, we get

$u − uh $2L2 (Ω) ≤ $∇(u − uh )$L2 (Ω) $∇(z − zI )$L2 (Ω) .

Using now the interpolation estimates, we deduce

'& ( 12 ' & ( 12
$u − uh $2L2 (Ω) ≤ C2app,H1 h2T |u|2H2 (T) h2T |z|2H2 (T)
T∈T T∈T
'& ( 12
≤ C2app,H1 h2T |u|2H2 (T) hmax |z|H2 (Ω) .
T∈T

Using now ($), we arrive at

'& ( 12
$u − uh $2L2 (Ω) ≤ C2app,H1 h2T |u|2H2 (T) hmax Cell−rec $u − uh $L2 (Ω)
T∈T

and the result follows.

122
Going high order...
The error analysis for high order elements is completely analogous.
Indeed, we only need a suitable best approximation result for high order elements.

Theorem (“Bramble-Hilbert Lemma” / best approximation estimatea )

a see, eg, B&S Thm (4.4.4)

Let T̂ the reference simplex and a function v̂ ∈ Hk (Ω), k ∈ N, k ≥ 2. Then, the

nodal interpolant v̂I ∈ Pp (T̂) to v̂ at the reference element nodes x̂i ,
i = 1, . . . , dim(Pp (T̂)), satisfies the estimate

$v̂ − v̂I $Hs (T̂) ≤ C(s, r)|v̂|Hr (T̂) ,

for any s < r ≤ min{p + 1, k}, for constant C(s, r) > 0, depending only on s and r.

) & + 12
Here |v̂|Hm (ω) := $Dα v̂$2L2 (ω)
α:|α|=m

123
Going high order...
We need to extend this to a general element T ∈ T . Again, we employ the affine
mappings FT : T̂ → T with FT (x̂) = BT x̂ + c.
Consider a function w ∈ Hk (T) and change variables:
!
2 2
, ,2
|w|H1 (T) = $∇w$L2 (T) = , ∇w(x), dx
T
!
, −# ,2
= ,B ∇w(F
ˆ T (x̂)) det BT dx̂
, (recall JFT = BT
#
)
T
T̂
!
−1 2
, ,2
≤ $BT $2 | det BT | ,∇w(Fˆ T (x̂)) dx̂
,
T̂
2 2
= $B−1
T 2$ | det BT ||w ◦ F T | H1 (T̂)
.
Inductively, for s ∈ N, we can show
1
s
T 2 | det BT | |w ◦ FT |Hs (T̂)
|w|Hs (T) ≤ $B−1 $ 2

Completely analogously, we also have

− 12
|w ◦ FT |Hs (T̂) ≤ $BT $s2 | det BT | |w|Hs (T)

124
Going high order...

Noting vI ◦ FT = (v ◦ FT )I , the best approximation estimate gives for v ∈ Hk (T):

1
s
$v − vI $Hs (T) ≤ $B−1
T $2 | det BT | $(v − vI ) ◦ FT $Hs (T̂)
2

1
s
= $B−1
T $2 | det BT | $v ◦ FT − (v ◦ FT )I $Hs (T̂)
2

1
s
≤ $B−1
T $2 | det BT | C(s, r)|v ◦ FT |Hr (T̂)
2

1
s − 12
≤ $B−1
T $2 | det BT | 2 C(s, r)$BT $r2 | det BT | |v|Hr (T)
or
s r
$v − vI $Hs (T) ≤ C(s, r)$B−1
T $2 $BT $2 |v|Hr (T)

125
Going high order...

s r
$v − vI $Hs (T) ≤ C(s, r)$B−1
T $2 $BT $2 |v|Hr (T)

It is an elementary geometric exercise to show:

$BT $2 ≤ ChT
hT = diam(T) hT = diam(T)
hT = diam(T)

$B−1
T $ 2 ≤ Cρ −1
T
ρT
ρT

for C > 0 independent of T ρT

giving
hrT
$v − vI $Hs (T) ≤ C s |v|Hr (T)
ρT
for any s < r ≤ min{p + 1, k}, for some C > 0 depending only on r, s.
126
Shape-regular mesh families

Definition
A family of meshes {Th }h∈H , for some index set H, is called shape-regular if there
exists a constant Csh > 1, independent of h, such that

hT ≤ Csh ρT ∀T ∈ Th and ∀h ∈ H.

For a shape-regular family, therefore, we have

$v − vI $Hs (T) ≤ Cap hr−s

T |v|Hr (T)

for any s < r ≤ min{p + 1, k}, with Cap := CCssh .

127
Error Analysis

Theorem (Convergence in the V-norm)

If u ∈ H10 (Ω) ∩ Hk (Ω) is the solution to Problem and uh ∈ Vhp the solution to
Method, then
'& ( 12
2(r−1)
$u − uh $V ≤ Cap hT |u|2Hr (T) ≤ Cap hr−1
max |u|Hr (Ω) ,
T∈T

for 2 ≤ r ≤ min{p + 1, k}, and hmax := maxT∈T hT .

If r = k = p + 1, we have p-th order convergence rate in the $·$V -norm.

128
Error analysis in the L2 -norm

Theorem (Convergence in the L2 -norm)

Let u ∈ H10 (Ω) ∩ Hk (Ω) be the solution to Problem. Moreover, let Ω ⊂ Rd be such
that the elliptic regularity estimate

$u$H2 (Ω) ≤ Cell−rec $∆u$L2 (Ω)

holds.
If uh ∈ Vhp is the solution to Method, then we have the error bound
'& ( 12
2(r−1)
$u − uh $L2 (Ω) ≤ Chmax hT |u|2Hr (T) ≤ CL2 hrmax |u|Hr (Ω) ,
T∈T

for 2 ≤ r ≤ min{p + 1, k}, and hmax := maxT∈T hT and CL2 := Cap Cell−rec .

The proof is left as an exercise.

Completely analogous considerations (and error bounds) can be proven for
tensor-product elements.
129