Standard Error Estimates for the Finite Element

Method
[H] 4.1, 4.I
2
 Basic properties of linear spaces
 Sobolev norms
 “Best approximation” property
 Error estimates

3
 To understand why the FEM works well
 To find ways to improve it when it doesn’t

4
5
 Real linear space: A collection of objects for
which the operations of addition and scalar
multiplication are defined and behave as
follows: If 𝑥 and 𝑦 are members of the linear
space 𝔸 and 𝛼, 𝛽 ∈ ℝ, then 𝛼𝑥 + 𝛽𝑦 ∈ 𝔸.
 Example: ℝ𝑛 is a linear space

6
 An inner product ⋅,⋅ on a real linear space 𝔸
is a map ⋅,⋅ : 𝔸 × 𝔸 → ℝ with the follow
properties for any 𝑥, 𝑦, 𝑧 ∈ 𝔸 and 𝛼 ∈ ℝ:
1. Symmetry: 𝑥, 𝑦 = 𝑦, 𝑥
2. Linearity: 𝛼𝑥, 𝑦 = 𝛼 𝑥, 𝑦 , 𝑥 + 𝑦, 𝑧 = 𝑥, 𝑧 +
𝑦, 𝑧
3. Positive-definiteness: 𝑥, 𝑥 ≥ 0 and 𝑥, 𝑥 = 0
iff 𝑥 = 0
 Note: iff means “if and only if”

7
 1
 Is 𝑤, 𝑣 = an inner product?
‫׬‬0 𝑤,𝑥 𝑣,𝑥 𝑑𝑥
 It depends on which linear space we talk
about!
◦ If 𝔸 = 𝐻1 0,1 , then?
◦ If 𝔸 = 𝐻01 0,1 ≔ 𝑤 ∈ 𝐻1 0,1 𝑤 0 = 𝑤 1 = 0 , then?

8
 Let 𝔸, ⋅,⋅ be an inner product space. Then
𝑥, 𝑦 ∈ 𝔸 are said to be orthogonal or
perpendicular (with respect to ⋅,⋅ ) if 𝑥, 𝑦 = 0.
 Example: Let 𝔸 = 𝐿2 −1,1 , 𝑓: 𝑥 ↦ 1, 𝑔: 𝑥 ↦ 𝑥,
then 𝑓 and 𝑔 are orthogonal with respect to
1
the inner product 𝑓, 𝑔 = ‫׬‬−1 𝑓𝑔 𝑑𝑥.

9
 Let ⋅,⋅ : 𝔸 × 𝔸 → ℝ be an inner product, then
for any 𝑥, 𝑦 ∈ 𝔸,
𝑥, 𝑦 2 ≤ 𝑥, 𝑥 𝑦, 𝑦
 Proof. Let 𝛼 ∈ ℝ. Then by definition
0 ≤ 𝑥 + 𝛼𝑦, 𝑥 + 𝛼𝑦 = 𝑥, 𝑥 + 2𝛼 𝑥, 𝑦 + 𝛼 2 𝑦, 𝑦
This holds for any 𝛼 ∈ ℝ. Thus its discriminant
is nonpositive, i.e.,
2 𝑥, 𝑦 2 − 4 𝑥, 𝑥 𝑦, 𝑦 ≤ 0
∎
 Q: When does equality hold?

10
 A norm ⋅ on a linear space 𝔸 is a map ⋅ : 𝔸 → ℝ,
with the follow properties for any 𝑥, 𝑦 ∈ 𝔸 and 𝛼 ∈
ℝ:
1. Positive-definiteness: 𝑥 ≥ 0 and 𝑥 = 0 iff
𝑥=0
2. 𝛼𝑥 = 𝛼 𝑥
3. Triangle inequality: 𝑥 + 𝑦 ≤ 𝑥 + 𝑦
 Note: An inner product space 𝔸, ⋅,⋅ possesses a
natural norm defined by 𝑥 = 𝑥, 𝑥 1Τ2
 Note: Another form to write the Cauchy-Schwarz
inequality: 𝑥, 𝑦 ≤ 𝑥 𝑦

11
 Example: ℝ𝑛 equipped with the dot product
𝑥, 𝑦 = 𝑥 ⋅ 𝑦 is an inner product space. The natural
norm 𝑥 = 𝑥 ⋅ 𝑥 1Τ2 is the length of vector 𝑥
 Example: 𝐿2 0,1 equipped with the inner product
1
𝑢, 𝑣 = ‫׬‬0 𝑢𝑣 𝑑𝑥 is an inner product space. The
1 2 1Τ2
natural norm is 𝑢 𝐿2 = ‫׬‬0 𝑢 𝑑𝑥 .
 Example: 𝐻1 0,1 equipped with the inner product
1
𝑢, 𝑣 = ‫׬‬0 𝑢𝑣 + 𝑢,𝑥 𝑣,𝑥 𝑑𝑥 is an inner product space.
1 1Τ2
The natural norm is 𝑢 𝐻1 = ‫׬‬0 𝑢2 + 𝑢,𝑥
2 𝑑𝑥
.

12
 Two norms ⋅ 𝑎 and ⋅ 𝑏 are equivalent if
there exist 𝑐1 , 𝑐2 ∈ ℝ+ such that for all 𝑥 ∈ 𝔸,
𝑐1 𝑥 𝑎 ≤ 𝑥 𝑏 ≤ 𝑐2 𝑥 𝑎
 Example: In ℝ𝑛 , the following two norms are
equivalent:
𝑥 2 = 𝑥12 + ⋯ + 𝑥𝑛2 1Τ2
𝑥 ∞ = max |𝑥𝑖 |
1≤𝑖≤𝑛
 In finite dimensions, all norms are equivalent.

13
 Completeness: A space is complete if the
limit of a convergent sequence of elements in
the space is also in the space.
 Hilbert space: An inner product space which
is complete in its natural norm is called a
Hilbert space.

14
 A seminorm ⋅ on a linear space 𝔸 is a map
⋅ : 𝔸 → ℝ with the following properties for any
𝑥, 𝑦 ∈ 𝔸 and 𝛼 ∈ ℝ:
1. Positive-semidefiniteness: 𝑥 ≥ 0
2. 𝛼𝑥 = 𝛼 𝑥
3. Triangle inequality: 𝑥 + 𝑦 ≤ 𝑥 + 𝑦

1 2 1Τ2
 Example: 𝑢 = ‫𝑢 ׬‬,𝑥 𝑑𝑥 defines a
0
seminorm on 𝐻1 0,1 but not a norm.

15
16
Consider a domain Ω ⊆ ℝ𝑛 , 𝑛 ≥ 1, and let
𝑢, 𝑣: Ω → ℝ
 The 𝐿2 Ω inner product and norm are defined
by
𝑢, 𝑣 = න 𝑢𝑣 𝑑Ω
Ω
and
𝑢 = 𝑢, 𝑢 1Τ 2
𝐿2

17
 The 𝐻1 Ω inner product and norm are
defined by
𝑢, 𝑣 1 = න 𝑢𝑣 + 𝑢,𝑖 𝑣,𝑖 𝑑Ω
Ω
with summation over 1 ≤ 𝑖 ≤ 𝑛 implied, and
Τ
𝑢 𝐻 1 = 𝑢, 𝑢 11 2

18
 The 𝐻 𝑠 Ω inner product and norm are
defined by
𝑢, 𝑣 𝑠
= න 𝑢𝑣 + 𝑢,𝑖 𝑣,𝑖 + 𝑢,𝑖𝑗 𝑣,𝑖𝑗 + ⋯
Ω
ቀ
+ 𝑢, 𝑖𝑗…𝑘 𝑣, 𝑖𝑗…𝑘 𝑑Ω
ቁ
𝑠 indices 𝑠 indices
with summation over 1 ≤ 𝑖, 𝑗, … , 𝑘 ≤ 𝑛 implied,
and
1Τ 2
𝑢 𝐻 = 𝑢, 𝑢 𝑠
𝑠

19
Let 𝑢, 𝑣: Ω → ℝ𝑚 (i.e., vector-valued)
 The 𝐻 Ω, ℝ inner product and norm are
𝑠 𝑚
defined by
𝑢, 𝑣 𝑠
= න 𝑢𝑖 𝑣𝑖 + 𝑢𝑖,𝑗 𝑣𝑖,𝑗 + 𝑢𝑖,𝑗𝑘 𝑣𝑖,𝑗𝑘 + ⋯
Ω
ቀ
+ 𝑢𝑖, 𝑗𝑘…𝑙 𝑣𝑖, 𝑗𝑘…𝑙 𝑑Ω
ቁ
𝑠 indices 𝑠 indices
with summation over 1 ≤ 𝑖 ≤ 𝑚, 1 ≤ 𝑗, 𝑘, … , 𝑙 ≤ 𝑛
implied, and
1Τ 2
𝑢 𝐻 = 𝑢, 𝑢 𝑠
𝑠

20
 𝐻 0 = 𝐿2
 The 𝑠th Sobolev space, denoted by 𝐻 𝑠 , is the
collection of functions 𝑢: Ω → ℝ𝑚 with finite 𝐻 𝑠
norm.
 The Sobolev spaces 𝐻 𝑠 Ω are Hilbert spaces
 𝐻 𝑟 ⊂ 𝐻 𝑠 , if 𝑟 ≥ 𝑠

21
 Q: Are 𝐻 𝑠 functions smooth in the classical
sense?
 Definition: Let 𝐶𝑏𝑘 , 𝑘 ≥ 0, be the space of
functions 𝑢: Ω → ℝ𝑚 which are
◦ Bounded
◦ Continuous and have continuous and bounded
derivatives of order 𝑗, 1 ≤ 𝑗 ≤ 𝑘.
 Theorem. If Ω is an open set in ℝ𝑛 and 𝑠 >
𝑛Τ2 + 𝑘, then
𝐻 𝑠 ⊂ 𝐶𝑏𝑘

22
 Examples:
◦ In 1D: 𝐻1 ⊂ 𝐶𝑏0 , 𝐻 2 ⊂ 𝐶𝑏1 , etc.
◦ In 2D: 𝐻 2 ⊂ 𝐶𝑏0
 Note that the result is “sharp”. For example,
in 2D, 𝐻1 ⊄ 𝐶𝑏0 . For example, take Ω =
1
2
𝑥, 𝑦 ∈ ℝ 𝑟 = 𝑥 + 𝑦 2
≤ . Let 𝑢: Ω → ℝ
2 1Τ 2
2
be 𝑢 = log − log 𝑟 . Show that 𝑢 ∈ 𝐻1 Ω but 𝑢 ∉
𝐶𝑏0 Ω .

23
 Solution. It is obvious that 𝑢 ∉ 𝐶𝑏0 Ω . Now we
compute
𝜕𝑢 𝜕𝑢 𝜕𝑟 1 𝑥 cos 𝜃 𝜕𝑢 sin 𝜃
= = = , =
𝜕𝑥 𝜕𝑟 𝜕𝑥 𝑟 log 𝑟 𝑟 𝑟 log 𝑟 𝜕𝑦 𝑟 log 𝑟
where 𝜃 is a polar coordinate such that 𝑥 = 𝑟 cos 𝜃. As a
result,
𝑢 2𝐻 1
1Τ2 2𝜋 2 2
2
cos 𝜃 sin 𝜃
=න න log − log 𝑟 + + 𝑑𝜃 𝑟𝑑𝑟
0 0 𝑟 log 𝑟 𝑟 log 𝑟
1Τ2
2
1
= 2𝜋 න log − log 𝑟 + 2 𝑟𝑑𝑟
0 𝑟 log 𝑟

24
 Clearly
1Τ2
න log − log 𝑟 2𝑟 𝑑𝑟 < ∞
0
because the integrand is finite. In fact, it can be
shown with L’Hospital’s rule that
lim+ log − log 𝑟 2 𝑟 = 0
𝑟→0
 For the other term, we check
1Τ2 1Τ2 1Τ2
𝑟 𝑑𝑟 𝑑 log 𝑟 1 1
න 2
=න 2
= − = <∞
0 𝑟 log 𝑟 0 log 𝑟 log 𝑟 0
log 2
1
since lim+ = 0.
𝑟→0 log 𝑟

25
26
 In what sense is the finite element method
optimal?

27
 We assume the exact problem can be written as:
Find 𝑢 ∈ 𝕍 such that for all 𝑢 ∈ ℤ,
𝑎 𝑢, 𝑢 = ℓ 𝑢
 The approximate finite element problem can be
stated as: Find 𝑢ℎ ∈ 𝕍ℎ such that for all 𝑢 ∈ ℤℎ ,
𝑎 𝑢, 𝑢ℎ = ℓ 𝑢
 Recall our assumptions:
◦ 𝕍ℎ ⊂ 𝕍, ℤℎ ⊂ ℤ
◦ 𝑎 ⋅,⋅ is symmetric and bilinear, and ℓ ⋅ is linear
◦ 𝑎 ⋅,⋅ 1Τ2 and ⋅ 𝐻 𝑚 define equivalent norms on ℤ, where
for heat conduction and linear elasticity, 𝑚 = 1; for the
Euler-Bernoulli beam theory, 𝑚 = 2.

28
 Let 𝑒 = 𝑢ℎ − 𝑢 denote the error in the finite
element approximation.
 Theorem.
a. 𝑎 𝑢, 𝑒 = 0 for all 𝑢 ∈ ℤℎ
b. 𝑎 𝑒, 𝑒 ≤ 𝑎 𝑈 ℎ − 𝑢, 𝑈 ℎ − 𝑢 for all 𝑈 ℎ ∈ 𝕍ℎ
 Part (a) means that the error is orthogonal to
the subspace ℤℎ
 Part (b) states the best approximation property
of the method: Among all possible linear
combinations of the chosen shape functions,
the finite element method gives us the best one
in the sense that it minimizes the error in the
norm 𝑎 ⋅,⋅ 1Τ2 .

29
Proof. (a) Let 𝑢 ∈ ℤℎ , then
𝑎 𝑢, 𝑢ℎ = ℓ 𝑢
But ℤℎ ⊂ ℤ, therefore 𝑢 ∈ ℤ, as a result,
𝑎 𝑢, 𝑢 = ℓ 𝑢
Subtracting these equations, we get
𝑎 𝑢, 𝑢ℎ − 𝑎 𝑢, 𝑢 = 0
Or,
𝑎 𝑢, 𝑒 = 0

30
 (b) Let 𝑢 = 𝑈 ℎ − 𝑢ℎ , then 𝑢 ∈ ℤℎ , and 𝑈 ℎ − 𝑢 =
𝑈 ℎ − 𝑢ℎ + 𝑢ℎ − 𝑢 = 𝑢 + 𝑒.
𝑎 𝑈 ℎ − 𝑢, 𝑈 ℎ − 𝑢 = 𝑎 𝑢 + 𝑒, 𝑢 + 𝑒
= 𝑎 𝑢, 𝑢 + 2𝑎 𝑢, 𝑒 + 𝑎 𝑒, 𝑒
Since 𝑎 𝑢, 𝑢 ≥ 0, 𝑎 𝑢, 𝑒 = 0, we have
𝑎 𝑈 ℎ − 𝑢, 𝑈 ℎ − 𝑢 ≥ 𝑎 𝑒, 𝑒
∎

31
32
 Is there a way for us to predict the convergence rate of the
error, i.e., of 𝑒 𝐻𝑚 ?
 Yes. First, from the norm equivalence, we know that there
exist 𝑐1 , 𝑐2 ∈ ℝ+ such that for all 𝑢 ∈ 𝕍,
𝑐1 𝑢 𝐻 𝑚 ≤ 𝑎 𝑢, 𝑢 1Τ2 ≤ 𝑐2 𝑢 𝐻 𝑚
Invoking also the best approximation property,
1 1Τ2 1 ℎ ℎ 1Τ2 𝑐2 ℎ
𝑒 𝐻 ≤ 𝑎 𝑒, 𝑒
𝑚 ≤ 𝑎 𝑈 − 𝑢, 𝑈 − 𝑢 ≤ 𝑈 − 𝑢 𝐻𝑚
𝑐1 𝑐1 𝑐1
where 𝑈 ℎ is any member of 𝕍ℎ . Since 𝑈 ℎ is arbitrary, we can
even write
𝑐2 ℎ−𝑢
𝑒 𝐻 ≤
𝑚 min 𝑈 𝐻𝑚
𝑐1 𝑈 ℎ ∈𝑉 ℎ
which is remarkable, since this means, given 𝑢, we are free
to construct 𝑈 ℎ to our advantage. What we need is just an
upper bound for the right hand side above.

33
 Given 𝑢 ∈ 𝐻 𝑟 , we can construct 𝑈 ℎ , called the
interpolant (interpolate in [H]), such that
there exists 𝑐 > 0 independent of 𝑢 such that
𝑢 − 𝑈 ℎ 𝐻 𝑚 ≤ 𝑐ℎ𝛼 𝑢 𝐻 𝑟
where 𝛼 = min(𝑘 + 1 − 𝑚, 𝑟 − 𝑚), 𝑘 is the degree
of complete polynomials appearing in the
shape functions, and ℎ is the smallest circle
that contains the biggest element

34
 With the above, we have
𝑒 𝐻 𝑚 ≤ 𝑐ℎ𝛼 𝑢 𝐻 𝑟
where 𝑐 = 𝑐𝑐1 Τ𝑐2 .
 How about error estimates in 𝐻 , 0 ≤ 𝑠 ≤ 𝑚?
𝑠

If 𝑟 ≥ 𝑘 + 1 (i.e., 𝑢 ∈ 𝐻𝑘+1 ), then

𝑒 𝐻 𝑠 ≤ 𝑐ℎ𝛽 𝑢 𝐻 𝑘+1
where 𝑐 is a constant independent of 𝑢 and ℎ, and
𝛽 = min 𝑘 + 1 − 𝑠, 2 𝑘 + 1 − 𝑚
Example: If 𝑚 = 1, 𝑘 = 1, 𝑠 = 0, what do we get?
Example: If 𝑚 = 2, 𝑘 = 3, 𝑠 = 0,1,2, what do we get?

35
36
 In the case of plane stress or plane strain, if the
domain has a crack, then the sharp result for 𝑢 ∈
3
𝐻𝑟 Ω, ℝ2 is 𝑟 = − 𝜖 for 𝜖 > 0. (Check with 𝑢 =
1 2
𝑟 2 𝑓(𝜃) with 𝑓 a smooth enough function).
 Now with 𝑚 = 1,
𝑢 − 𝑈 ℎ 𝐻 1 ≤ 𝑐ℎ𝛼 𝑢 𝐻 𝑟
1 1
where 𝛼 = min(𝑘 + 1 − 𝑚, 𝑟 − 𝑚) = min 𝑘, −
𝜖 = −
2 2
𝜖. That is, regardless of 𝑘!
 In plane words: the standard finite element
method is unable to get the optimal convergence
rate for fracture mechanics problems

37
 When 𝜈 → 0.5, we may see strange results,
especially with low-order elements
𝑢 − 𝑈 ℎ 𝐻 1 ≤ 𝑐ℎ𝛼 𝑢 𝐻 𝑟
The problem is with 𝑐. It can be shown that 𝑐 is
proportional to
𝜇
1 − 2𝜈
(equivalently, proportional to the bulk modulus)
Hence the error becomes unbounded.

38