Notes of the Lecture Course

Numerical Methods for Partial Differential Equations

(Mathematical Engineering)

Stefano Berrone, Claudio Canuto

Dipartimento di Scienze Matematiche
Politecnico di Torino
10129 Torino Italy
[email protected]

February 7, 2021
2
Introduction

Bla bla bla

3
4
Chapter 1

A Few Concepts
on Partial Differential Equations
(
F ind u ∈ V such that
[VP ] (1.1)
a(u, v) = F (v) ∀v ∈ V .

(u, v)a = a(u, v) ∀u, v ∈ V . (1.2)

p
with corresponding norm kvka = (v, v)a .
√ p
α kvkV ≤ kvka ≤ kak kvkV ∀v ∈ V . (1.3)

Lu = −∇ · (µ∇u) + β · ∇u + σu (1.4)

Semi-norms and norms

Let v be a function defined in an open set A ⊆ Rd . For any multi-index α = (α1 , α2 , . . . , αd ) ∈
Nd , with |α| = α1 + α2 + · · · αd , let us denote by
∂ |α| v
Dα v =
∂xα1 1 ∂xα2 2 . . . ∂xαd d
the partial derivative of v of index α; the integer |α| is the order of the derivative. For conve-
nience, the derivative of order 0 of v is the function itself.
Given an integer ` ≥ 0, assume that all the partial derivatives of v of order ` (in the classical
or distributional sense) are square-integrable in A. Then, we define the (Sobolev) semi-norm of
order ` of v in A by
 1/2
X Z
|v|`,A :=  |Dα v|2  . (1.5)
A
α∈Nd , |α|=`

Given an integer m ≥ 0, we define the (Sobolev) norm of order m of v in A by

m
!1/2
X
2
kvkm,A := |v|`,A . (1.6)
`=0

This is the norm of v in the Sobolev space H m (A). Note that H 0 (A) = L2 (A).

5
6 CHAPTER 1. PARTIAL DIFFERENTIAL EQUATIONS

Global vs local Sobolev regularity

Let A ⊆ Rd be an open set, which is partitioned into two open subsets A1 and A2 by a smooth
(e.g., Lipschitz continuous) (d − 1)-dimensional manifold Γ: precisely,

A = A1 ∪ Γ ∪ A2 with A1 ∩ A2 = ∅ , Γ = ∂A1 ∩ ∂A2 .

Given two functions v1 ∈ H 1 (A1 ) and v2 ∈ H 1 (A2 ), denote by v the function defined in A by
(
v1 in A1 ,
v=
v2 in A2 .

Property 1.0.1 v ∈ H 1 (A) if and only if v1|Γ = v2|Γ . In this case

kvk2H 1 (A) = kv1 k2H 1 (A1 ) + kv2 k2H 1 (A2 ) .

Proof. Obviously, v ∈ L2 (A) with kvk2L2 (A) = kv1 k2L2 (A1 ) + kv2 k2L2 (A2 ) . Then, v ∈ H 1 (A) if
∂v
and only if, for 1 ≤ i ≤ d, each partial derivative ∂x i
in the sense of distributions coincides
with a function gi ∈ L (A). Since by assumption ∂xi ∈ L2 (A1 ) and ∂v
2 ∂v1 2
∂xi ∈ L (A2 ), the natural
2

candidate for gi is the function 

 ∂v1 in A1 ,
∂xi
gi =
 ∂v2 in A .
∂xi 2

Let us compute the i-th distributional derivative of v: for any test function φ ∈ D(A), one has
Z
∂v ∂φ
h , φi = − v (definition of distributional derivative)
∂xi A ∂xi
Z Z
∂φ ∂φ
= − v − v (additivity of integrals)
A1 ∂xi A2 ∂xi
Z Z
∂φ ∂φ
= − v1 − v2 (definition of v)
A1 ∂xi A2 ∂xi
Z Z Z Z
∂v1 (1) ∂v1 (2)
= φ − v1 φ ni + φ − v2 φ ni (integration by parts)
A1 ∂x i Γ A1 ∂x i Γ
Z Z
(1) (2)
= gi φ − (v1 ni + v2 ni ) φ (definition of gi ) ,
A Γ

(1)
where ni denotes the i-th component of the unit normal vector to Γ, oriented outward with
(2)
respect to A1 and ni has the similar meaning with respect to A2 . Then,
Z
∂v
h , φi = gi φ ∀φ ∈ D(A)
∂xi A

if and only if Z
(1) (2)
(v1 ni + v2 ni ) φ = 0 ∀φ ∈ D(A) .
Γ
(1) (2)
Since φ is arbitrary, this is equivalent to v1 ni + v2 ni = 0 in Γ, which in turn is equivalent to
(2) (1)
v1 − v2 = 0 on Γ since ni = −ni .
Chapter 2

Galerkin Methods

The Galerkin approach is a conceptually simple, yet powerful strategy to discretize the varia-
tional problem [VP ] introduced in (1.1), i.e., to generate a similar problem in finite dimension
(hence, computationally tractable).

2.1 Definition of a Galerkin method

The idea is simple: let us choose a finite dimensional subspace Vδ ⊂ V , of dimension Nδ . Here,
δ > 0 is a characteristic parameter related to the definition of Vδ (such as, e.g., a mesh size, or
the reciprocal of a polynomial degree); later on, we will indeed consider sequences of subspaces
for δ → 0 and Nδ → ∞. Then, we restrict [VP ] to Vδ , i.e., we consider the discrete variational
problem
(
F ind uδ ∈ Vδ such that
[VP ]δ (2.1)
a(uδ , vδ ) = F (vδ ) ∀vδ ∈ Vδ .

The well-posedness of this problem directly stems from the assumptions made on the exact
(infinite dimensional) variational problem [VP ]. Indeed, Vδ is a closed subspace of V , being
finite dimensional; consequently, it inherits the Hilbert-space structure from V . Furthermore,
the forms a and F are continuous on Vδ since they are on the larger space V , and similarly a is
coercive on Vδ since it is on V . Thus, we can apply the Lax-Milgram theorem to [VP ]δ , getting
the following fundamental result.

Theorem 2.1.1 The discrete variational problem [VP ]δ stated in (2.1) admits one and only one
solution uδ , which satisfies
1
kuδ kV ≤ kF kV 0 . (2.2)
α

The bound (2.2) has an important interpretation: it is an instance of the property called
numerical stability. It means that a suitable norm of the discrete solution (here, the norm
of V ) can be bounded by a quantity independent of the discretization parameter δ. This is a
necessary condition to have the convergence of uδ to the exact solution u of [VP ], if we let δ
tend to 0; indeed, remember that a convergent sequence is necessarily bounded.

7
8 CHAPTER 2. GALERKIN METHODS

2.2 Reduction to an algebraic system

In order to compute uδ , we must introduce a basis in Vδ . This allows us to transform [VP ]δ into
an equivalent algebraic system of Nδ linear equations in Nδ unknowns, that can be solved by
one of the available techniques in Numerical Linear Algebra.
Let us choose any basis Φδ = {ϕ1 , ϕ2 , . . . , ϕNδ } in Vδ (to be precise, we should append the
index δ to each function, say ϕδ,1 , ϕδ,2 , . . . , but we retain from this to keep the notation simple).
Then, any vδ ∈ Vδ can be represented as
Nδ
X
vδ = vj ϕj (2.3)
j=1

for suitable coefficients vj ∈ R; consequently, we can associate to vδ the vector v = (vj )1≤j≤Nδ ∈
RNδ of its coefficients with respect to the chosen basis. By definition of basis, the mapping
vδ 7→ v is an isomorphism; hence, we can identify a function in Vδ with a vector in RNδ . This
is important, because computers and programming languages are built to handle vectors and
matrices in the most efficient way. Note that to apply correctly the rules of linear algebra, v
and the subsequent vectors must be considered as column vectors.
Once a basis is selected, we obtain Nδ equations by picking each basis function as a test
function in the variational equations (2.1), i.e.,

a(uδ , ϕj ) = F (ϕj ), 1 ≤ j ≤ Nδ . (2.4)

Amazingly, this finite set of equations is equivalent to the infinitely-many equations in (2.1)
(one for each choice of vδ ∈ Vδ ); this is a consequence of the linear dependence of both sides
upon vδ . Indeed, if vδ has the representation (2.3) and the equations (2.4) hold true, then by
multiplication of both sides by vj and summation over j we get

Nδ
X Nδ
X
vj a(uδ , ϕj ) = vj F (ϕj ) ,
j=1 j=1

which by linearity is the same as

Nδ
X XNδ
a(uδ , v j ϕj ) = F ( v j ϕj ) , i.e., a(uδ , vδ ) = F (vδ ) .
j=1 j=1

To complete our transformation, we represent uδ with respect to the selected basis, as

Nδ
X
uδ = uk ϕk (2.5)
k=1

for some unknown coefficients uk , collected in the column vector u = (uk )1≤k≤Nδ ∈ RNδ . Sub-
stituting this expression into (2.4) and using the linearity of a with respect to the first variable,
we obtain
XNδ
uk a(ϕk , ϕj ) = F (ϕj ), 1 ≤ j ≤ Nδ . (2.6)
k=1
2.3. FEATURES OF THE STIFFNESS MATRIX 9

This suggests to introduce the square matrix A = (ajk )1≤j,k≤Nδ ∈ RNδ ×Nδ and the column
vector f = (fj )1≤j≤Nδ ∈ RNδ whose entries are

ajk = a(ϕk , ϕj ), fj = F (ϕj ) , (2.7)

so that (2.6) can be written as

Nδ
X
ajk uk = fj , 1 ≤ j ≤ Nδ , or, in matrix form, Au = f .
k=1

Summarizing, we have established the following result.

Proposition 2.2.1 Let Φδ = {ϕ1 , ϕ2 , . . . , ϕNδ } be any basis in Vδ . Then, the Galerkin problem
[VP ]δ is equivalent to the algebraic system

Au = f , (2.8)

where the entries of the matrix A and the vector f are given in (2.7).

Borrowing from a terminology used in structural mechanics, the matrix A will be called the
stiffness matrix associated with the bilinear form a and the basis Φδ .

2.3 Features of the stiffness matrix

Since [VP ]δ admits existence and uniqueness of a solution, the same is true for the linear system
(2.8). Thus, we are guaranteed that the matrix A is non-singular. But we can say more, thanks
to the coercivity of the form a.

Property 2.3.1 The stiffness matrix A is positive-definite, i.e., it satisfies

v T Av > 0 for any vector v 6= 0 .

Proof. The jth component of the vector Av is

Nδ
X Nδ
X XNδ
(Av)j = ajk vk = vk a(ϕk , ϕj ) = a( vk ϕk , ϕj ) = a(vδ , ϕj ) ,
k=1 k=1 k=1
PNδ
where vδ = k=1 vk ϕk is the function in Vδ associated with the vector v. Then,
Nδ
X Nδ
X Nδ
X
v T Av = vj (Av)j = vj a(vδ , ϕj ) = a(vδ , vj ϕj ) = a(vδ , vδ )
j=1 j=1 j=1

since vδ = N
P δ
j=1 vj ϕj as well (j or k above are just dummy variables of summation). Thus,
invoking coercivity,
v T Av = a(vδ , vδ ) ≥ αkvδ k2V ≥ 0 . (2.9)
We can conclude once we have shown that v T Av = 0 implies v = 0. If v T Av = 0, then from
the previous inequality we deduce kvδ kV = 0, hence vδ = 0 ∈ Vδ . By definition of basis, this is
possible only if all its coefficients vk are zero, which is precisely equivalent to v = 0.
10 CHAPTER 2. GALERKIN METHODS

This important property has the following equivalent interpretation. If A is symmetric (which
happens if the bilinear form a(u, v) is symmetric), then positive definiteness is equivalent to the
fact that all the eigenvalues of A are strictly positive. If A is not symmetric, we can decompose
it in the sum of a symmetric matrix and an antisymmetric one,
A = As + Aas with As = 12 (A + AT ), Aas = 12 (A − AT ),
and observe that v T Aas v = 0 for any vector v, since
v T Aas v = v T (−ATas )v = −(Aas v)T v = −v T Aas v ,
which is possible only if this quantity is 0. We conclude that v T As v = v T Av > 0, i.e., As is
symmetric and positive-definite. Summarizing, we have proven the following result.
Corollary 2.3.2 All the eigenvalues of the symmetric part As of the stiffness matrix A are
strictly positive.
The previous property allows one to apply to the solution of the system (2.8) all the (direct
or iterative) techniques that exploit the positive definiteness of the matrix A in the symmetric
case (such as, e.g., the Choleski factorization or the Conjugate Gradient method), or of the
symmetric part of A in the general case (such as, e.g., the GMRES method).
Other features of the stiffness matrix are of interest. For instance, if the selected basis functions
have a localized support in the domain Ω of interest, i.e., they are different from 0 only in a
small portion of the domain, the stiffness matrix may be sparse, meaning that the percentage
of non-zero entries is “small”. Furthermore, if the basis functions are numbered convenably,
the stiffness matrix may be banded, meaning that the non-zero entries are located inside a
(often small) band around the main diagonal. All such features can be cleverly exploited to save
memory in the representation of the stiffness matrix in the computer, and to save computational
time in the solution of the linear system.
Finally, we should not forget that to every square non-singular matrix we can associate a
number ≥ 1, the condition number (with respect to a chosen norm), whose magnitude indi-
cates how “good” or “bad” is to operating with such matrix at the numerical level (in terms
of, e.g., propagation of the numerical errors in a direct method, or steps needed to match a
prescribed error tolerance in an iterative method). For a stiffness matrix, the condition number
depends on the choice of the basis used to represent functions in Vδ . There are indeed “good”
bases and “bad” bases. Many preconditioning techniques, aimed at transforming the linear
system into an equivalent one with a matrix having a better condition number, can be viewed
indeed as the application of a suitable change of basis in Vδ .

2.4 Error analysis. Convergence of the Galerkin solutions

It is important, for both theoretical and practical purposes, to be able to say something about
the error committed by replacing the exact solution u of Problem [VP ] by the Galerkin solution
of Problem [VP ]δ . With this goal in mind, we introduce the discretization error
eδ = u − uδ , (2.10)
which can be measured by any norm or seminorm in V . The most natural measure in our setting
is the quantity
εV,δ = keδ kV = ku − uδ kV . (2.11)
2.4. ERROR ANALYSIS 11

There exists two main strategies to estimate the discretization error, which are somehow
complementary to each other:

• A priori error estimation: one gives estimates of the error before even computing the
Galerkin solution uδ , by just knowing (or assuming) certain qualitative properties of the
exact solution u.

• A posteriori error estimation: one gives estimates of the error without any knowledge
of the exact solution, using only the information on the data of the problem and the
computed Galerkin solution uδ .

The former strategy is more theoretically oriented, aiming at predicting the qualitative behavior
of the discretization error, such as e.g. its asymptotic decay as the discretization parameter δ
tends to 0. The latter strategy may have a more direct and practical impact, for instance by
providing quantitative information that can be used to decide how to improve the quality of the
current approximation (adaptivity).

2.4.1 A priori error estimates

We start with a simple, yet fundamental property of any Galerkin discretization.

Property 2.4.1 The Galerkin solution uδ satisfies the following equations:

a(u − uδ , vδ ) = 0 , ∀vδ ∈ Vδ . (2.12)

Proof. The exact solution of Problem [VP ] satisfies, in particular for any vδ ∈ Vδ ,

a(u, vδ ) = F (vδ ) .

Subtracting from this the equation satisfied by uδ

a(uδ , vδ ) = F (vδ )

and using the linearity of a with respect to the first argument, we obtain the result.
This property is often known as the Galerkin orthogonality property. Indeed, when the form
a is symmetric, (2.12) can be rephrased as

(u − uδ , vδ )a = 0 , ∀vδ ∈ Vδ ,

where (·, ·)a is the inner product in V associated to a (see (1.2)). These relations show that
u − uδ is a-orthogonal to the subspace Vδ , or, equivalently, that uδ is the a-orthogonal projection
of u upon Vδ (see Fig. 2.4.1). This gives an immediate minimality property for the discretization
error: indeed, since the orthogonal projection is the closest element to u in Vδ , it holds

ku − uδ ka ≤ ku − vδ ka , ∀vδ ∈ Vδ ,

which shows that uδ is the best approximation of u in Vδ in the a-norm. In terms of the original
norm in V , using the norm equivalence (1.3), we deduce from the previous inequality
√ p
α ku − uδ kV ≤ kak ku − vδ kV , ∀vδ ∈ Vδ ,
12 CHAPTER 2. GALERKIN METHODS

Figure 2.1: Orthogonal projection of u on Vδ

whence r
kak
ku − uδ kV ≤ ku − vδ kV , ∀vδ ∈ Vδ . (2.13)
α
A similar estimate can be obtained when the form a is not symmetric, using again the Galerkin
orthogonality property. We proceed as follows:

α ku − uδ k2V ≤ a(u − uδ , u − uδ ) (by coercivity)

= a(u − uδ , u) − a(u − uδ , uδ ) (by linearity)
= a(u − uδ , u) − 0 (by (2.12) with vδ = uδ )
= a(u − uδ , u) − a(u − uδ , vδ ) (for any vδ ∈ Vδ , again by (2.12))
= a(u − uδ , u − vδ ) (again by linearity)
≤ kak ku − uδ kV ku − vδ kV (by continuity) .

Simplifying, we arrive at
kak
ku − uδ kV ≤ ku − vδ kV , ∀vδ ∈ Vδ . (2.14)
α
Together with (2.13), we have thereby established the following fundamental result, often referred
to as the Céa Lemma.
Theorem 2.4.2 The Galerkin discretization error satisfies the bound

ku − uδ kV ≤ Ca ku − vδ kV , ∀vδ ∈ Vδ , (2.15)

where q
 kak if a is symmetric ,
α
Ca = (2.16)
kak
otherwise .

α
2.4. ERROR ANALYSIS 13

Equivalently, one can write

ku − uδ kV ≤ Ca distV (u, Vδ ) , (2.17)
where
distV (u, Vδ ) = min ku − vδ kV (2.18)
vδ ∈Vδ

is the distance of u from the subspace Vδ , measured in the norm of V .

The importance of Céa’s Lemma comes from the fact that (2.17) gives a bound of the V -norm
of the discretization error (which depends in an intricate way upon both the bilinear form a and
the approximation subspace Vδ ) as the product of two factors, the one (Ca ) depending only on
a, the other (distV (u, Vδ ) ) depending only on Vδ .
Hence, assuming that we know Ca or an upper bound for it, the problem of estimating
the discretization error is shifted to the (simpler) problem of estimating the approximation
error associated with the subspace Vδ . This task will be accomplished in some of the following
chapters, particularly for finite element subspaces.
Finally, note that by definition of distance, we have distV (u, Vδ ) ≤ ku − uδ kV ; hence, (2.17)
can be rephrased as
distV (u, Vδ ) ≤ ku − uδ kV ≤ Ca distV (u, Vδ ) , (2.19)
which shows that discretization error and best approximation error are as much close to each
other as the constant Ca is close to 1 (recall that the ratio kak/α is always ≥ 1). Also note that
the constant Ca is smaller in the symmetric case than in the non-symmetric one.

2.4.2 Convergence results

Let us now assume that we have not just one subspace Vδ , but a family {Vδ }δ>0 of finite
dimensional subspaces of V , parametrized by some parameter δ that varies continuously or
discretely and becomes smaller and smaller. We are interested in seeing what happens to the
Galerkin approximations uδ in the limit δ → 0.
Let us assume that the following consistency assumption holds true:

distV (u, Vδ ) → 0 as δ → 0, (2.20)

i.e., the exact solution u of our Problem [VP ] can be approximated arbitrarily well by functions
in the subspaces Vδ . Then, eq. (2.17) and the comparison theorem for limits immediately imply
that
ku − uδ kV → 0 as δ → 0 , (2.21)
which expresses the convergence of the family of Galerkin solutions to the exact solution, in
the norm of V . In short, we say that the Galerkin discretization method (2.1) is convergent (to
the solution u).
At this point, we have to check the consistency assumption for our specific choice of discretiza-
tion spaces, and also to get information about the speed of convergence of the discretization,
i.e., the way the distance distV (u, Vδ ) decays when δ tends to 0. The following result provides a
general framework, which can be applied in many different situations.
Theorem 2.4.3 Assume that Z ⊂ V is a normed space, for which the following consistency
condition holds: there exists a function ψZ : R+ → R+ satisfying ψZ (δ) → 0 for δ → 0, such
that
distV (w, Vδ ) ≤ ψZ (δ) kwkZ , ∀w ∈ Z . (2.22)
14 CHAPTER 2. GALERKIN METHODS

Then, if u ∈ Z, the Galerkin solutions converge to u in V , with the error estimate

ku − uδ kV ≤ Ca ψZ (δ) kukZ . (2.23)
Furthermore, if Z is dense in V , the consistency assumption (2.20) is fulfilled for any u ∈ V ;
hence, the Galerkin discretization method is convergent to any u ∈ V .
Proof. Estimate (2.23) follows from (2.17) and (2.22) applied to u. To prove the last result,
let u be any element in V . By the density assumption, for any ε > 0 there exists uε ∈ Z such
that ku − uε kV ≤ 2ε . On the other hand, by the assumption on ψZ , there exists δε > 0 such that
ε
ψZ (δ) ≤ , ∀δ < δε .
2kuε kZ
Then, using ku − vδ kV ≤ ku − uε kV + kuε − vδ kV for any vδ ∈ Vδ , we obtain for any δ < δε ,
ε ε ε
distV (u, Vδ ) ≤ ku − uε kV + distV (uε , Vδ ) < + ψZ (δ) kuε kZ ≤ + = ε ,
2 2 2
which proves (2.20).

Remark 2.4.4 i) A typical example of consistency estimate (2.22) for finite element approxi-
mations, that will be proven in the forthcoming chapters, is
distH 1 (Ω) (w, Vh ) ≤ C h kwkH 2 (Ω) ∀w ∈ H 2 (Ω) ,
where (setting here δ = h) Vh is a space of continuous, piecewise polynomial functions on a
partition of the domain Ω and h is a characteristic size of the elements of the partition.
ii) Inequality (2.23) is an example of a priori error estimate for the Galerkin method.
iii) We have obtained the convergence results (2.21) or (2.23) by exploiting on the one side
the consistency condition (2.20) and on the other side the bound (2.17) in Céa’s theorem; this
bound relies on the property of coercivity of the bilinear form a, which – when applied on the
subspaces Vδ – yields the numerical stability property (2.1.1).
One can actually explicitly prove that convergence is implied by consistency and numerical sta-
bility, as an instance of a general theorem in Numerical Analysis, known as the Lax-Richmyer
Theorem.

2.4.3 A posteriori error estimates

Once the Galerkin solution uδ ∈ Vδ has been computed, we can define the residual associated
with it, as the mapping
rδ : V → R , rδ (v) = F (v) − a(uδ , v) ∀v ∈ V . (2.24)
This is a linear continuous form on V , as a consequence of the properties of a and F ; hence, rδ
is an element of the dual space V 0 .
We note that
rδ (vδ ) = 0 ∀vδ ∈ Vδ (2.25)
as a consequence of the definition of Galerkin solution; furthermore,
rδ = 0 ∈ V 0 (i.e., rδ (v) = 0 for all v ∈ V ) if and only if uδ = u .
It is therefore natural to try to relate the size of the residual rδ to the size of the error uδ − u.
This is expressed by the subsequent result.
2.4. ERROR ANALYSIS 15

Theorem 2.4.5 The following inequalities hold:

1 1
krδ kV 0 ≤ ku − uδ kV ≤ krδ kV 0 . (2.26)
kak α

Proof. One has

rδ (v) = F (v) − a(uδ , v) = a(u, v) − a(uδ , v) = a(u − uδ , v) ≤ kak ku − uδ kV kvkV ,

whence
rδ (v) kak ku − uδ kV kvkV
krδ kV 0 = sup ≤ sup = kak ku − uδ kV ,
v∈V \{0} kvkV v∈V \{0} kvkV

which gives the left-hand inequality in (2.26). On the other hand, from the coercivity inequality
αku − uδ k2V ≤ a(u − uδ , u − uδ ), we obtain

a(u − uδ , u − uδ ) a(u − uδ , v) rδ (v)

αku − uδ kV ≤ ≤ sup = sup = krδ kV 0 ,
ku − uδ kV v∈V \{0} kvk V v∈V \{0} kvkV

which gives the right-hand inequality in (2.26).

The theorem guarantees that the dual norm of the residual is equivalent to the V -norm of
the discretization error, up to constants depending on the bilinear form a. Thus, in theory, one
could use the quantity krδ kV 0 as an a-posteriori error estimator. However, a dual norm is hardly
computable in practice, as it is defined via a supremization procedure. Nonetheless, the result
suggests to approximate krδ kV 0 by a computable quantity ηδ , for which one can prove bounds
of the form
c C
ηδ ≤ ku − uδ kV ≤ ηδ , (2.27)
kak α
for suitable constants c, C > 0 independent of δ (and known exactly or approximately). Any
such ηδ defines an a posteriori error estimator. The right-hand inequality expresses the
reliability property, meaning that if the estimator is below a prescribed threshold, then we are
guaranteed that also the discretization error is below a multiple of that threshold. Conversely,
the left-hand inequality expresses the efficiency property, meaning that the actual discretization
error cannot be much smaller than the estimator; this avoids generating discretizations, based
on the estimator, that are much richer than needed to reach the target error threshold.

Remark 2.4.6 In the manipulation of the dual norm of the residual, with the aim of designing
a practical a posteriori error estimator, one can use the relation (2.25) to express the residual as

rδ (v) = rδ (v) − rδ (vδ ) = rδ (v − vδ ) = F (v − vδ ) − a(uδ , v − vδ ) .

This implies that, for any vδ ∈ Vδ , we have

F (v − vδ ) − a(uδ , v − vδ )
krδ kV 0 = sup . (2.28)
v∈V \{0} kvkV

Choosing conveniently vδ , one can start from this expression to derive a computable error esti-
mator ηδ ' krδ kV 0 . We will give some examples in Sects. 3.4.3 and ...
16 CHAPTER 2. GALERKIN METHODS

2.5 Extensions
Hereafter, we present three variants of the classical Galerkin method discussed above. They
may lead to a better accuracy or a better computational efficiency. A combination of variants
is also possible.

2.5.1 Petrov-Galerkin methods

In some applications (such as, e.g., convection-diffusion problems in the convection-dominated
regime), it may be convenient to allow the freedom of playing with two different subspaces of
V , say Wδ in which the discrete solution is sought and Vδ in which test functions are chosen.
The two subspaces must have the same dimension Nδ < ∞.
The discrete variational problem (2.1) is modified as follows:
(
F ind uδ ∈ Wδ such that
[VP ]δ (2.29)
a(uδ , vδ ) = F (vδ ) ∀vδ ∈ Vδ .
This is the general setting of a Petrov-Galerkin method.
The well-posedness of this problem cannot be obtained from the coercivity property of the
form a; indeed, uδ need not be in Vδ , so we are not allowed to choose vδ = uδ as a test function.
The new condition that makes things work is the following one: there exists a constant αδ > 0
(possibly depending on δ) such that
a(wδ , vδ )
∀wδ ∈ Wδ sup ≥ αδ kwδ kV . (2.30)
vδ ∈ Vδ kvδ kV
vδ 6= 0

The condition is called inf-sup condition, since it can be equivalently stated as

a(wδ , vδ )
inf sup ≥ αδ . (2.31)
wδ ∈ Wδ vδ ∈ Vδ kwδ kV kvδ kV
wδ 6= 0 vδ 6= 0

The inf-sup condition is a generalization of the coercivity condition, in the sense that if Wδ = Vδ
the inf-sup condition follows from the coercivity condition: indeed, for any wδ ∈ Vδ , wδ 6= 0, one
has
a(wδ , wδ ) a(wδ , vδ )
αkwδ kV ≤ ≤ sup .
kwδ kV v ∈V kvδ kV
δ δ
vδ 6= 0
Using the inf-sup condition, we can prove the well-posedness of the variational problem (2.29).
Indeed, we have
a(uδ , vδ ) F (vδ ) F (v)
αδ kuδ kV ≤ sup = sup ≤ sup = kF kV 0 ,
vδ ∈ Vδ kvδ kV vδ ∈ Vδ kvδ kV v ∈ Vδ kvkV
vδ 6= 0 vδ 6= 0 v 6= 0

i.e.,
1
kuδ kV ≤
kF kV 0 . (2.32)
αδ
Two important consequences stem from this inequality:
2.5. EXTENSIONS 17

• if F = 0, then uδ = 0, which implies existence and uniqueness of the solution, since [VP ]δ
is a linear problem in finite dimension;

• we assess the numerical stability of the discretization, provided the inf-sup condition is
uniform in δ, i.e., there exists a constant ᾱ > 0 independent of δ such that αδ ≥ ᾱ. Indeed,
in this case we have
1
kuδ kV ≤ kF kV 0
ᾱ
and the right-hand side is independent of δ.

Finally, in order to estimate the discretization error u − uδ by exploiting the inf-sup condition,
we write u − uδ = (u − wδ ) + (wδ − uδ ) for an arbitrary wδ ∈ Wδ , and use the triangle inequality

ku − uδ kV ≤ ku − wδ kV + kwδ − uδ kV . (2.33)

The second addend on the right-hand side can be controlled by the inf-sup condition

a(wδ − uδ , vδ )
αδ kwδ − uδ kV ≤ sup ,
vδ ∈ Vδ kvδ kV
vδ 6= 0

where the numerator can be transformed as follows:

a(wδ − uδ , vδ ) = a(wδ , vδ ) − a(uδ , vδ ) = a(wδ , vδ ) − F (vδ )

= a(wδ , vδ ) − a(u, vδ ) = a(wδ − u, vδ ) ≤ kak ku − wδ kV kvδ kV .

Hence,
kak
kwδ − uδ kV ≤ ku − wδ kV .
αδ
Inserting this bound in (2.33), and taking the infimum over all wδ ∈ Wδ , we conclude that
 
kak
ku − uδ kV ≤ 1 + distV (u, Wδ ) . (2.34)
αδ

This result is the counterpart of the Céa Lemma for Petrov-Galerkin methods. If the inf-sup
condition is uniform, then the constant in parenthesis can be bounded independently of δ, which
means that the discretization error behaves as the best approximation error of u in Wδ . On the
contrary, if αδ becomes smaller and smaller as δ → 0, then in practice the discretization error
may be significantly larger than the best approximation error.
We see here a clear example of the Lax-Richtmyer theorem: if the discretization is numerically
stable (i.e., αδ ≥ ᾱ > 0) and consistent (i.e., distV (u, Wδ ) → 0 as δ → 0), then it is convergent
(ku − uδ kV → 0 as δ → 0).

2.5.2 Approximate bilinear/linear forms

In the applications to partial differential equations, the bilinear form a and the linear form F are
defined by means of integrals. Often, it is not possible or convenient to compute these integrals
exactly, and numerical integration formulas have to be used. Since such formulas are not exact,
they lead to approximate bilinear and linear forms, here indicated by aδ and Fδ .
18 CHAPTER 2. GALERKIN METHODS

Thus, we assume that we are given a bilinear form aδ : Vδ × Vδ → R and a linear form
Fδ : Vδ → R, and we consider the approximate discrete variational problem
(
F ind uδ ∈ Vδ such that
[VP ]δ (2.35)
aδ (uδ , vδ ) = Fδ (vδ ) ∀vδ ∈ Vδ .
This is the natural setting for a Galerkin method with numerical integration (see Sect. ... ).
To guarantee the well-posedness of this problem, we have to make a suitable assumption on
the bilinear form aδ . A sufficient assumption is the discrete coercivity condition, i.e., there exists
αδ > 0 such that
aδ (vδ , vδ ) ≥ αδ kvδ k2V ∀vδ ∈ Vδ .
In order to estimate the error u − uδ , we use again (2.33) with arbitrary wδ ∈ Vδ . By the
coercivity of aδ in Vδ , we have
aδ (wδ − uδ , vδ )
αδ kwδ − uδ kV ≤ sup ,
vδ ∈ V δ kv δ k V
vδ 6= 0

with aδ (wδ − uδ , vδ ) = aδ (wδ , vδ ) − aδ (uδ , vδ ) = aδ (wδ , vδ ) − Fδ (vδ ). Adding and subtracting the
quantity a(wδ , vδ ) − F (vδ ) on the right-hand side, we obtain
aδ (wδ − uδ , vδ ) = a(wδ , vδ ) − F (vδ ) + [aδ (wδ , vδ ) − a(wδ , vδ )] − [Fδ (vδ ) − F (vδ )]
= a(wδ − u, vδ ) + (aδ − a)(wδ , vδ ) − (Fδ − F )(vδ ) .
Thus,
a(wδ − u, vδ ) (aδ − a)(wδ , vδ ) (Fδ − F )(vδ )
αδ kwδ − uδ kV ≤ sup + sup + sup
vδ ∈ V δ kvδ kV vδ ∈ Vδ kvδ kV vδ ∈ V δ kvδ kV
vδ 6= 0 vδ 6= 0 vδ 6= 0
≤ kak ku − wδ kV + Eδ (a; wδ ) + Eδ (F ) ,
where
(a − aδ )(wδ , vδ ) (F − Fδ )(vδ )
Eδ (a; wδ ) := sup , Eδ (F ) := sup (2.36)
vδ ∈ Vδ kvδ kV vδ ∈ V δ kvδ kV
vδ 6= 0 vδ 6= 0

are discretization errors of the forms a and F , respectively. Inserting this bound in (2.33), we
get  
kak
ku − uδ kV ≤ 1 + ku − wδ kV + Eδ (a; wδ ) + Eδ (F ) ;
αδ
since wδ ∈ Vδ is arbitrary, we conclude with the a-priori error bound
  
kak
ku − uδ kV ≤ inf 1+ ku − wδ kV + Eδ (a; wδ ) + Eδ (F ) . (2.37)
wδ ∈ Vδ αδ
wδ 6= 0
This result is known as first Strang Lemma. It implies that the convergence of the discretiza-
tion follows from the numerical stability (i.e., αδ ≥ ᾱ for some ᾱ > 0 independent of δ), and the
following consistency conditions:
- for all δ, there exists wδ ∈ Vδ such that ku − wδ kV → 0 and Eδ (a; wδ ) → 0 as δ → 0;
- Eδ (F ) → 0 as δ → 0.
2.5. EXTENSIONS 19

2.5.3 Non-conforming/external methods

Sometimes one gains in flexibility and efficiency by removing the requirement that the finite
dimensional space Vδ is contained in V ; Vδ is just required to be a subspace of a larger space H,
such that V ⊂ H. In this case, the bilinear form a and the linear form F must be extended in
such a way that they are defined on Vδ . Denoting by aδ and Fδ these extensions, we formally
end up with the same discrete variational problem [VP ]δ as in (2.35). When Vδ 6⊂ V , the
discretization is termed non-conforming, or external.
In the finite element method, an external approximation occurs if one relaxes the continuity
requirements between contiguous elements. A popular example is the so-called Discontinuous-
Galerkin Method, which uses discontinuous, piecewise polynomial functions over a triangulation
of the domain.
20 CHAPTER 2. GALERKIN METHODS
Chapter 3

Discretizations in One Space

Variable

Throughout this chapter, I will denote a closed bounded interval of the real line; up to a
translation, we may assume that I = [0, L] for some real L > 0. By a small abuse of notation,
we will denote by L2 (I), H 1 (I), ... the corresponding spaces defined on the open interval int(I) =
(0, L).

3.1 Approximation spaces

Our first purpose is to approximate a given function defined in I, using piecewise polynomial
functions on a partition T of I. Why polynomials? Because polynomials can be easily computed,
differentiated, integrated, and so on. Why piecewise polynomials? Because this gives the freedom
of choosing different polynomials in different parts of the domain I.
A partition T of I is defined by assigning distinct N + 1 points xj (0 ≤ j ≤ N ) in I, called
the nodes of the partition (or the mesh). We assume that nodes are ordered increasingly, i.e.,
xj−1 < xj for any j, with x0 = 0 and xN = L (see Fig. 4.4.1). The nodes define the elements of
the partition, which are the intervals

Ij = [xj−1 , xj ] ,1≤j≤N, with size hj = |Ij | = xj − xj−1 .

Thus, T = {Ij : 1 ≤ j ≤ N } and I = N

S
j=1 Ij . It is convenient to define h = max1≤j≤N hj and,
sometimes, to indicate the partition by Th .
L
A uniform or equally-spaced partition is such that all elements have the same size hj = h = N ;
L
in this case, the nodes are given by xj = jh = j N .
As a first step, we discuss how to represent polynomials in each element E of the partition.

3.1.1 Bases of polynomials in an interval

For any integer k ≥ 0, Pk will denote the vector space of all algebraic polynomials in one variable
of degree ≤ k (defined on the real line), whereas Pk (E) will denote the space of their restrictions
to the interval E. The dimension of this space is

dim Pk (E) = k + 1 .

Here, we list some bases in Pk (E).

21
22 CHAPTER 3. DISCRETIZATIONS IN ONE SPACE VARIABLE

Figure 3.1: An example of partition T of the interval [0, L]

• The basis of the monomials 1, x, x2 , . . . , xk . This is the classical basis, however for increas-
ing values of k it becomes numerically unstable, or ill-conditioned (the monomials tend to
become linearly dependent). Hence, it should never be used in computations, except for
very small values of k.

• The basis of Legendre orthogonal polynomials L0 (x), L1 (x), . . . , Lk (x), which satisfy
Z
deg Ln (x) = n and Ln (x) Lm (x) dx = δn,m , 0 ≤ n, m ≤ k .
E

Other inner products can be used in E, leading e.g. to Chebyshev or Jacobi polynomials.
All such bases are numerically stable. However, since these polynomials are non-zero at
both endpoints of the interval E, glueing them to form globally continuous functions on
the partition T leads to composite bases whose functions are supported over the whole
interval I.

• The interpolatory, or Lagrange basis Φ = {ϕ0 (x), ϕ1 (x), . . . , ϕk (x)} associated with k + 1
distinct point ξ0 < ξ1 < · · · < ξk in E. Each ϕ` ∈ Pk (E) is uniquely defined by the
conditions
ϕ` (ξi ) = δi,` , i = 0, 1, . . . , k .
Note that if E = [α, β] and we choose ξ0 = α, ξk = β, then only ϕ0 is non-zero at the left
endpoint of E, only ϕk is non-zero at the right endpoint of E, whereas all the other basis
functions ϕ` for 1 ≤ ` ≤ k − 1 vanish at both endpoints of E; a basis with such a property
is often called a boundary adapted basis. The property easily allows creating a composite
basis whose functions are supported on a single element, or at most on two contiguous
elements.
For k = 1, the basis functions are
β−x x−α
ϕ0 (x) = , ϕ1 (x) = , (3.1)
β−α β−α

whereas if k = 2, denoting by γ = 12 (α + β) the midpoint of E, we have

(β − x)(γ − x) (β − x)(x − α) (x − α)(x − γ)

ϕ0 (x) = , ϕ1 (x) = , ϕ2 (x) = .
(β − α)(γ − α) (β − γ)(γ − α) (β − α)(β − γ)
(3.2)
3.1. APPROXIMATION SPACES 23

Note that as soon as the polynomial degree k exceeds a few units, interpolation in E
should not be performed at equally spaced nodes (to avoid instabilities, namely intra-node
oscillations, as in the celebrated Runge phaenomenon), but rather at a family of nodes with
Gaussian-type spacing, such as e.g. the Gauss-Lobatto nodes which include the endpoints
of the element.

• The modal, or Babuška-Shen basis Ψ = {ψ0 (x), ψ1 (x), . . . , ψk (x)}, defined as follows:
ψ0 = ϕ0 and ψ1 = ϕ1 are the linear functions given in (3.1), whereas ψn for 2 ≤ n ≤ k are
primitives of Legendre polynomials, that vanish at both endpoints of E; precisely
Z β
ψn (x) = Ln−1 (s) ds = Ln (x) − Ln−2 (x) .
x

This is another example of a boundary adapted basis.

3.1.2 Spaces of piecewise polynomials

Consider again a partition Th in I. Let us introduce the following finite dimensional vector
spaces.

• For any k ≥ 0, set

V(Th ; k) = {v ∈ L2 (I) : v|Ij ∈ Pk (Ij ) for 1 ≤ j ≤ N } . (3.3)

This is a subspace of L2 (I), of dimension (k + 1)N . Functions in such space need not be
continuous at the nodes of the partition (see Fig. 3.1.2, upper plot).

• For any k ≥ 1, set

V (Th ; k) = {v ∈ C 0 (I) : v|Ij ∈ Pk (Ij ) for 1 ≤ j ≤ N } . (3.4)

This is a subspace of H 1 (I), formed by globally continuous functions. Its dimension is

(k + 1)N − (N − 1) = kN + 1, since this space is obtained from V(Th ; k) by enforcing one
continuity condition (i.e., v(x− +
j ) = v(xj )) at each of the N − 1 internal nodes.

• For any k ≥ 1, set

V0 (Th ; k) = {v ∈ V (Th ; k) : v(0) = 0, v(L) = 0} . (3.5)

This is a subspace of H01 (I), of dimension kN + 1 − 2 = kN − 1.

Let us now discuss how to create bases in these spaces, starting from bases in the polynomial
spaces on the elements.

• A basis in V(Th ; k) can be obtained by selecting a basis in each Pk (Ij ) and then extending it
by 0 outside Ij . For instance, a basis in the N -dimensional space V(Th ; 0) of the piecewise
constant functions on the partition is given by
(
1 if x ∈ Ij ,
ϕj (x) = 1≤j≤N.
0 otherwise ,
24 CHAPTER 3. DISCRETIZATIONS IN ONE SPACE VARIABLE

Figure 3.2: Examples of functions in V(Th ; 2) (upper plot), in V (Th ; 2) (middle plot), in V0 (Th ; 2)
(lower plot)

• A basis in V (Th ; k) can be obtained as follows:

– select a boundary adapted basis in each Pk (Ij ), 1 ≤ j ≤ N ;
– for each pair of contiguous elements, consider the two basis functions (one per ele-
ment) that do not vanish at the common node, glue them together by continuity and
extend the resulting function by zero in the remaining portion of the domain;
– extend by zero any other basis function supported in one element.
For instance, consider the (N + 1)-dimensional space V (Th ; 1) of the continuous, piecewise
linear functions on T ; note that each function v in this space is uniquely identified by its
values v(xj ) at the N + 1 nodes of the partition. Glueing and extending the interpolatory
basis functions given by (3.1) in each space P1 (Ij ), we obtain the composite Lagrange basis
Φ formed by the so-called hat functions
x − xj−1


 if x ∈ Ij ,


 hj
ϕj (x) = xj+1 − x if x ∈ Ij+1 , 0≤j≤N (3.6)


 hj+1

0 otherwise ,
(see Fig. 3.1.2, left plot), with the obvious changes when j = 0 and j = N (see Fig. 3.1.2,
3.1. APPROXIMATION SPACES 25

Figure 3.3: Basis functions in V (Th ; q): internal function ϕj (left plot), boundary functions ϕ0
and ϕN (center and right plot)

center and right plots). For 1 ≤ j ≤ N − 1, these functions are supported (i.e., different
from 0) only on the union Ij ∪ Ij+1 of two continguous elements, whereas ϕ0 is supported
only in I1 and ϕN is supported only in IN . The Lagrange representation of any function
v ∈ V (Th ; 1) is
XN
v(x) = v(xj )ϕj (x) , ∀x ∈ I .
j=0

Next, consider the (2N + 1)-dimensional space V (Th ; 2) of the continuous, piecewise
quadratic functions on T ; now each function v in this space is uniquely identified by
its values v(xj ) at the N + 1 nodes of the partition and v(xj−1/2 ) at the N midpoints
xj−1/2 = 21 (xj−1 + xj ) of the elements of the partition. Glueing and extending the in-
terpolatory basis functions given by (3.2) in each space P2 (Ij ), we obtain the composite
Lagrange basis Φ, formed by two families of functions: the so-called pagoda functions
(x − xj )(x − xj−1/2 )

2 if x ∈ Ij ,


h2j




ϕj (x) = (xj+1 − x)(xj+1/2 − x) 0≤j≤N (3.7)
2 2 if x ∈ Ij+1 ,
hj+1





0 otherwise ,


(with the obvious changes when j = 0 and j = N ), and the so-called bubble functions

 (xj+1 − x)(x − xj ) if x ∈ Ij ,

ϕj−1/2 (x) = h2j 1≤j≤N (3.8)

0 otherwise ,
(see Fig. 3.1.2). With these basis functions, the Lagrange representation of any function
v ∈ V (Th ; 2) is
N
X N
X
v(x) = v(xj )ϕj (x) + v(xj−1/2 )ϕj−1/2 (x) , ∀x ∈ I .
j=0 j=1
26 CHAPTER 3. DISCRETIZATIONS IN ONE SPACE VARIABLE

Figure 3.4: The basis functions ϕj (upper plot) and ϕj−1/2 (lower plot) in V (Th ; 2)

• At last, a basis in V0 (Th ; k) can be obtained from the basis Φ in V (Th ; k) defined above,
simply by discarding the two basis functions ϕ0 and ϕN , which are the only ones that do
not vanish at both endpoints of I.

3.2 Approximation errors

Let us start with an auxiliary result. Let φ : I → R be any bounded and continuously differen-
tiable function. Let E = [α, β] ⊆ I be a subinterval of I, of length hE = β − α. Assume that
φ(x0 ) = 0 for some x0 ∈ E. Then, using the fundamental theorem of Calculus, we can write,
for any x ∈ E, Z x Z x
φ(x) = φ(x0 ) + φ0 (s) ds = φ0 (s) ds ,
x0 x0
whence Z x
 Z
0
|φ0 (s)| ds ,
 
|φ(x)| =  φ (s) ds ≤
x0 I[x0 ,x]

where I[x0 , x] ⊆ E is the interval whose endpoints are x0 and x. Using Hölder inequalities on
the right-hand side, we obtain
Z 1/2
|φ(x)| ≤ |x − x0 | max |φ0 (s)| or |φ(x)| ≤ |x − x0 |1/2 |φ0 (s)|2 ds .
s∈E E

From these inequalities, we can derive bounds of a norm of φ in terms of the length of the
interval E and a norm of φ0 . For instance,
Z 1/2
0 1/2 0 2
max |φ(x)| ≤ hE max |φ (s)| or max |φ(x)| ≤ hE |φ (s)| ds ,
x∈E s∈E x∈E E

whereas integration over E immediately gives

Z 1/2 Z 1/2 Z 1/2
2 1 3/2 0 2 1 0 2
|φ(x)| ds ≤ √ hE max |φ (s)| or |φ(x)| ds ≤ √ hE |φ (s)| ds .
E 3 s∈E E 2 E
3.2. APPROXIMATION ERRORS 27

Such inequalities extend to continuous functions with distributional first derivative in L2 (E)
or in L∞ (E), since they can be approximated arbitrarily well by smooth functions (classical
functions are dense in Sobolev spaces). In this way, one can obtain bounds as stated below.
Property 3.2.1 Let φ vanish in E and let φ0 ∈ L2 (E). Then,
1/2
kφkL2 (E) ≤ √1
2
hE kφ0 kL2 (E) and kφkL∞ (E) ≤ hE kφ0 kL2 (E) .

Furthermore, if ϕ0 ∈ L∞ (E), then

3/2
kφkL2 (E) ≤ √1
3
hE kφ0 kL∞ (E) and kφkL∞ (E) ≤ hE kφ0 kL∞ (E) .

3.2.1 Local interpolation errors

Consider now a function v ∈ H 1 (E) (hence, continuous in E). We aim at approximating it by
a polynomial vI ∈ Pk (E) of degree k that interpolates v at certain points in E.

• Case k = 0
Let vI = IE0 v ∈ P0 (E) be the constant function that satisfies vI = v(x0 ) for some x0 ∈ E.
Then, applying Property 3.2.1 to the function φ = v − vI and noticing that vI0 = 0, we get
kv − vI kL2 (E) ≤ √12 hE kv 0 − vI0 kL2 (E) = √12 hE kv 0 kL2 (E) , which we can write as

kv − IE0 vkL2 (E) ≤ √1

2
hE |v|H 1 (E) .

• Case k = 1
Let vI = IE1 v ∈ P1 (E) be the linear function that interpolates v at the endpoints of E, i.e.,
vI (α) = v(α) and vI (β) = v(β) Then, assuming that v ∈ H 2 (E) and applying Property
3.2.1 to the function ϕ = v − vI , we get

kv − vI kL2 (E) ≤ √1
2
hE kv 0 − vI0 kL2 (E) . (3.9)

On the other hand, by Rolle’s theorem, the function ψ = v 0 − vI0 vanishes at some point s0
in E; applying once more Property 3.2.1 to this function, and noticing that now vI00 = 0,
we get kv 0 − vI0 kL2 (E) ≤ √12 hE kv 00 kL2 (E) . Concatenating the two previous inequalities, we
get

kv − IE1 vkL2 (E) ≤ 1

2 h2E |v|H 2 (E) and |v − IE1 v|H 1 (E) ≤ √1
2
hE |v|H 2 (E) . (3.10)

• Case k = 2
Let vI = IE2 v ∈ P2 (E) be the quadratic function that interpolates v at the endpoints of
E and at the baricenter γ = 12 (α + β) of E. Then, assuming that v ∈ H 3 (E), applying
Property 3.2.1 to the functions φ = v − vI , ψ = v 0 − vI0 and η = v 00 − vI00 and noticing that
now vI000 = 0, we arrive at the inequalities

kv − IE2 vkL2 (E) ≤ 1

√
2 2
h3E |v|H 3 (E) ,

|v − IE2 v|H 1 (E) ≤ 1

2 h2E |v|H 3 (E) , (3.11)

|v − IE2 v|H 2 (E) ≤ √1

2
hE |v|H 3 (E) .
28 CHAPTER 3. DISCRETIZATIONS IN ONE SPACE VARIABLE

• Case k > 2
Let vI = IEk v ∈ Pk (E) be the polynomial that interpolates v at k + 1 distinct points in
E. Iterating the argument above, one can prove the following error estimates: for any `, r
satisfying 0 ≤ ` < r ≤ k + 1, if v ∈ H r (E) one has

|v − IEk v|H ` (E) ≤ c`,r,k hr−`

E |v|H r (E) ,

where c`,r,k > 0 are constants independent of v and hE .

Remark 3.2.2 The order of decay of the interpolation error, i.e., the power of hE in the
previous estimates, depends not only upon the polynomial degree k, but also upon the (Sobolev)
smoothness of the function v. For instance, if v only belongs to H 1 (E) but not to H 2 (E), then
the interpolation error using a linear polynomial can only be bounded as the interpolation error
using a constant polynomial, i.e.,

kv − IE1 vkL2 (E) ≤ c1 hE |v|H 1 (E) (3.12)

for some constant c1 > 0. This bound easily follows from (3.9), observing that vI = IE1 v satisfies
vI0 = v(β)−v(α) = h1E E v 0 (s) ds, which immediately gives kvI0 kL2 (E) ≤ kv 0 kL2 (E) .
R
β−α
Similarly, for the interpolation that uses a quadratic polynomial, we have the error bounds

kv − IE2 vkL2 (E) ≤ c2 hE |v|H 1 (E)

if v only belongs to H 1 (E) but not to H 2 (E), and

kv − IE2 vkL2 (E) ≤ c3 h2E |v|H 2 (E) and |v − IE2 v|H 1 (E) ≤ c4 hE |v|H 2 (E) .

if v is in H 2 (E) but not in H 3 (E).

3.2.2 Global interpolation errors

We now consider the approximation of a function v ∈ H 1 (I) by a piecewise polynomial function
on a partition Th of I, obtained by “glueing” local interpolating polynomials. We only consider
the case of polynomial degree k ≥ 1, since for k = 0 we obtain a piecewise constant function,
which does not belong to H 1 (I).

• Let us start with the linear case (k = 1). For any subinterval Ij (1 ≤ j ≤ N ) of the
partition Th , let vIj = II1j v ∈ P1 (Ij ) be the linear interpolant of v at the endpoints of Ij ,
i.e., vIj (xj−1 ) = v(xj−1 ) and vIj (xj ) = v(xj ). Then, we define

Ih1 v ∈ V (Th ; 1) by the conditions (Ih1 v)|Ij = vIj , 1 ≤ j ≤ N.

Equivalently, Ih1 v is identified by the interpolatory conditions

(Ih1 v)(xj ) = v(xj ), 0 ≤ j ≤ N,

which lead to the following representation in terms of the Lagrangean basis defined in (3.6)
N
X
(Ih1 v)(x) = v(xj )ϕj (x), ∀x ∈ I.
j=0
3.2. APPROXIMATION ERRORS 29

The error v − Ih1 v can be estimated using the bounds (3.10) elementwise. Precisely, as-
suming that v|Ij ∈ H 2 (Ij ) for all j, one has

N
X N
X N
X
kv − Ih1 vk2L2 (I) = k(v − Ih1 v)|Ij k2L2 (Ij ) = kv − vIj k2L2 (Ij ) ≤ 1
4 h4j |v|2H 2 (Ij ) ,
j=1 j=1 j=1

whence, recalling that h = maxj hj ,

 1/2
N
X
kv − Ih1 vkL2 (I) ≤ 1
2
 h4j |v|2H 2 (Ij )  (3.13)
j=1
 1/2
N
X
≤ 1
2 h2  |v|2H 2 (Ij ) 
j=1

1 2
= 2 h |v|H 2 (I) if v ∈ H 2 (I) . (3.14)

On the other hand,

 1/2
N
X
|v − Ih1 v|H 1 (I) = kv 0 − (Ih1 v)0 kL2 (I) ≤ √1
2
 h2j |v|2H 2 (Ij )  (3.15)
j=1
 1/2
XN
≤ √1
2
h |v|2 H 2 (Ij )

j=1

= √1 h |v|H 2 (I)
2
if v ∈ H 2 (I) . (3.16)

Remark 3.2.3 On the right-hand sides of these estimates, terms like (3.13) or (3.15)
provide a more accurate information than terms like (3.14) or (3.16). Indeed, the former
terms show the interplay between the size hj of the elements and the regularity of v inside
the elements. Such expressions suggest to use smaller elements where derivative of v are
larger, thus providing a rationale for mesh adaptivity.
For the sake of simplicity, in the sequel we will mostly present error estimates with terms of
type (3.14) or (3.16) on the right-hand side, but the reader should always keep in mind that
more accurate expressions could be written instead.

• Considering now the quadratic case (k = 2), we define

Ih2 v ∈ V (Th ; 2) by the conditions (Ih2 v)|Ij = II2j v, 1 ≤ j ≤ N,

which is equivalent to the interpolatory conditions

(Ih2 v)(xj ) = v(xj ), 0 ≤ j ≤ N, (Ih2 v)(xj−1/2 ) = v(xj−1/2 ) 1≤j≤N,

where xj−1/2 = 21 (xj−1 + xj ) is the midpoint of the interval Ij .

30 CHAPTER 3. DISCRETIZATIONS IN ONE SPACE VARIABLE

Assuming that v ∈ H 3 (I) and using the bounds (3.11) elementwise, we obtain as above
the following error estimates

kv − Ih2 vkL2 (I) ≤ 1

√
2 2
h3 |v|H 3 (I) , |v − Ih2 v|H 1 (I) ≤ 1
2 h2 |v|H 3 (I) .

Note that it would be meaningless to estimate the error in the H 2 (I)-seminorm, since the
piecewise interpolant Ih2 v is continuous but in general exhibits jumps in the first derivative
at the nodes of the partition, i.e., it does not belong to H 2 (I). Also note that if v does not
belong to H 3 (I) but only to H 2 (I), then recalling Remark 3.2.2 the previous estimates
must be replaced by

kv − Ih2 vkL2 (I) ≤ c3 h2 |v|H 2 (I) , |v − Ih2 v|H 1 (I) ≤ c4 h |v|H 1 (I) .

• In general, if we consider the piecewise polynomial interpolation Ihk v by polynomials of

degree k, defined by

Ihk v ∈ V (Th ; k) such that (Ihk v)|Ij = IIkj v, 1 ≤ j ≤ N, (3.17)

and we assume that v ∈ H r (I) for some r satisfying 1 ≤ r ≤ k + 1, we have the following
general error estimates

kv − Ihk vkL2 (I) ≤ c0,r,k hr |v|H r (I) , |v − Ihk v|H 1 (I) ≤ c1,r,k hr−1 |v|H r (I) , (3.18)

for suitable constants c0,r,k , c1,r,k > 0 independent of v and hE .

Remark 3.2.4 The previous estimates concern a so-called h-type approximation method, mean-
ing that the quality of the approximation is dictated by the fineness of the partition Th , whereas
the polynomial degree is kept fixed. One can also consider a p-type method (sometimes called
a spectral method), in which the partition is kept fixed, whereas the polynomial degree, often
denoted by pj , is allowed to increase in each element Ij .
Combining the two stategies leads to the so-called hp-type methods, in which a reduction of
the approximation error is achieved by simultaneously refining the partition and increasing the
polynomial degrees. In this case, the typical error estimates (assuming for simplicity to have
the same polynomial degree p in each element) are

hmin(p+1,r) hmin(p+1,r)−1
kv − Ihp vkL2 (I) ≤ c0,r kvkH r (I) , |v − Ihp v|H 1 (I) ≤ c1,r kvkH r (I) ,
pr pr−1
for v ∈ H r (I) with r ≥ 1 and p ≥ 1.
Remarkably, when the solution is analytic in I, the error can be shown to decay exponentially
in p.

3.3 A first example of Galerkin discretization

As a first example of a Galerkin discretization, we consider the one dimensional homogeneous
Dirichlet boundary value problem

− (µ u0 )0 + β u0 + σu = f in (0, L) ,
(3.19)
u(0) = u(L) = 0 .
3.3. A FIRST EXAMPLE OF GALERKIN DISCRETIZATION 31

Note that L = −(µ u0 )0 + β u0 + σu is the one-dimensional version of the general linear second-
order operator introduced in (1.4). We assume that µ, β, σ ∈ L∞ (I), f ∈ L2 (I) and, in addition,
µ ≥ µ0 in I, β 0 ∈ L∞ (I) and − 21 β 0 + σ ≥ σ0 ≥ σ0 in I, for certain constants µ0 > 0 and σ0 ≥ 0.
Under these conditions, the bilinear form
Z
a(u, v) = (µ u0 v 0 + β u0 v + σ u v) (3.20)
I

is continuous and coercive on V = H01 (I), and the linear form

Z
F (v) = f v (3.21)
I

is continuous on V . Hence, Problem (3.19) can be given the variational formulation

u ∈ V : a(u, v) = F (v) ∀v ∈ V . (3.22)

To discretize it, we choose the subspace Vh = V0 (Th ; 1) of the continuous, piecewise linear
functions on a partition Th of the domain I, which vanish at the endpoints of I (see (3.5)), and
we formulate the Galerkin problem

uh ∈ Vh : a(uh , vh ) = F (vh ) ∀vh ∈ Vh . (3.23)

Using the Lagrangean basis Φ = {ϕj : 1 ≤ j ≤ N − 1} in V0 (Th ; 1), we transform this problem
into the algebraic problem
Au = f ,
where, according to Sect. 2.2,
Z
A = (ajk )1≤j,k≤N −1 , ajk = (µ ϕ0k ϕ0j + β ϕ0k ϕj + σ ϕk ϕj ) ,
I
Z
f = (fj )1≤j≤N −1 , fj = f ϕj ,
I
and u = (uk )1≤k≤N −1 with uk = uh (xk ). Observe that the stiffness matrix A is the sum of three
matrices, the diffusion matrix D, the convection matrix C, and the reaction matrix R, whose
entries are Z Z Z
0 0 0
djk = µ ϕk ϕj , cjk = β ϕk ϕj , σ ϕk ϕj ,
I I I
namely,
A = D + C + R.
Recalling that supp ϕj = Ij ∪ Ij+1 for 1 ≤ j ≤ N − 1, each integral is 0 when the supports of
ϕj and ϕk are disjoint or intersect at a point only; these situations occur when |j − k| > 1. In
other words,
djk = cjk = rjk = 0 whenever k 6= j − 1, j, j + 1,
which gives the following
Property 3.3.1 The matrices D, C, R, and consequently A, are tridiagonal.
We now proceed to the computation of the entries of these matrices, as well as the right-hand
side f .
32 CHAPTER 3. DISCRETIZATIONS IN ONE SPACE VARIABLE

• Computation of D.
At first, let us compute dj,j−1 (for j > 1). The intersection of the supports of ϕj−1 and
ϕj is the interval Ij , hence Z
dj,j−1 = µ ϕ0j−1 ϕ0j .
Ij
On Ij the basis functions are linear, so their derivatives are constant; to be precise, by
(3.6) we have
1 1
ϕ0j−1 = − , ϕ0j = ,
hj hj
whence Z
1 1
dj,j−1 = − 2 µ = − µ̄j−1/2 ,
hj Ij hj
where µ̄j−1/2 = h1 j Ij µ dx is the mean value over Ij of the elastic coefficient µ. When this
R

number cannot be computed exactly, we can approximate it: let µj−1/2 denote the com-
puted approximation. For instance, using the mid-point quadrature rule to approximate
the integral, we obtain
µj−1/2 = µ(xj−1/2 ) where xj−1/2 = 21 (xj−1 + xj ) is the mid-point of Ij ;
using the trapezoidal rule instead, we get
µj−1/2 = 12 (µ(xj−1 ) + µ(xj )) .
In conclusion, we set
1
dj,j−1 = − µ .
hj j−1/2
Next, let us compute the diagonal entry djj . Since
Z Z
0 2
djj = µ (ϕj ) + µ (ϕ0j )2
Ij Ij+1

with
1 1
ϕ0j = on Ij , ϕ0j = − on Ij+1 ,
hj hj+1
the same approximation for the coefficient µ used above gives
1 1
djj = µj−1/2 + µ .
hj h j+1 j+1/2
Finally, by the symmetry of the matrix we get
1
dj,j+1 = dj+1,j = dj+1,(j+1)−1 = − µj+1/2 .
hj+1

To sum up, the entries of the diffusion matrix D are

 µ µj+1/2
j−1/2
 + if k = j ,
hj hj+1



 µ
− j−1/2

if k = j − 1 ,

djk = hj (3.24)
µj+1/2
−


 if k = j + 1 ,



 hj+1
0 otherwise .

3.3. A FIRST EXAMPLE OF GALERKIN DISCRETIZATION 33

Remark 3.3.2 It is worth noticing the property that the sum of the entries in each row
j, except the first and last one, is zero:

N
X −1
djk = 0 , 2 ≤ j ≤ N − 2.
k=1

Indeed,
N −1 N −1
!0
X Z xj+1 X
djk = µ ϕk ϕ0j ;
k=1 xj−1 k=1
PN −1
the linear function k=1 ϕk takes the value 1 at each internal node, hence it is identically
equal to 1 in any interval Ij , except the first and last one. Therefore, its derivative is
identically zero in the integration interval [xj−1 , xj+1 ].

Remark 3.3.3 In the relevant case in which the coefficient µ is constant and the mesh is
L
equally spaced (i.e. hj = h = N for any j), the matrix D has the simple structure

µ
D= tridiag [ −1 2 − 1] ,
h
where the symbol tridiag [a b c] denotes the tridiagonal matrix having constant value a
on the first lower diagonal, b on the main diagonal and c on the first upper diagonal.

• Computation of C.

We have Z
cj,j−1 = β ϕ0j−1 ϕj
Ij

which, after replacing β by a constant approximation βj−1/2 in Ij as we did above for

µj−1/2 , is replaced by

βj−1/2
Z  Z  
1 1 hj
cj,j−1 = βj−1/2 ϕ0j−1 ϕj = βj−1/2 − ϕj = βj−1/2 − =− ,
Ij hj Ij hj 2 2

since the integral of ϕj on Ij is the area of a triangle of basis hj and height 1.

Diagonal entries are defined as
Z Z
cjj = βj−1/2 ϕ0j ϕj + βj+1/2 ϕ0j ϕj dx
Ij Ij+1
βj−1/2 βj+1/2
   
1 hj 1 hj+1
= βj−1/2 + βj+1/2 − = − .
hj 2 hj+1 2 2 2

Finally,
βj+1/2
Z  
1 hj+1
cj,j+1 = βj+1/2 ϕ0j+1 ϕj = βj+1/2 = .
Ij+1 hj+1 2 2
34 CHAPTER 3. DISCRETIZATIONS IN ONE SPACE VARIABLE

Summarising, the entries of the convection matrix C are:

βj−1/2 βj+1/2


 − if k = j ,
2 2



 βj−1/2


− if k = j − 1 ,
cjk = 2 (3.25)

 β j+1/2

 if k = j + 1 ,



 2
0 otherwise .


Note that, again, sum of the entries in each row j, except the first and last one, is zero.
If β is constant then C reduces to
β
C= tridiag [−1 0 1] ,
2
which is a skew-symmetric matrix (i.e., C T = −C).

• Computation of R.
We have Z
rj,j−1 = σ ϕj−1 ϕj ,
Ij

which, after replacing σ by a constant approximation σj−1/2 in Ij , is replaced by

Z
rj,j−1 = σj−1/2 ϕj−1 ϕj .
Ij

We use (3.6) for the maps ϕj−1 and ϕj on the interval Ij , and we introduce the change
the integration variable x = xj−1 + shj , 0 ≤ s ≤ 1, so that dx = hj ds. Then,
Z Z xj
1
ϕj−1 ϕj dx = (xj − x)(x − xj−1 ) dx
Ij h2j xj−1
1 3 1 1 2 1 3 1 1
Z  
= h (1 − s)s ds = hj s − s = hj
h2j j 0 2 3 0 6

and so
1
rj,j−1 = σj−1/2 hj .
6
For the diagonal term Z Z
rjj = σϕ2j dx + σϕ2j dx ,
Ij Ij+1

we proceed as above to get

1 1
rjj = σj−1/2 hj + σj+1/2 hj+1 .
3 3
Finally, by symmetry we have
1
rj,j+1 = rj+1,j = σj+1/2 hj+1 .
6
3.4. ERROR ANALYSIS 35

The entries of the matrix R are therefore


1
3 (σj−1/2 hj + σj+1/2 hj+1 ) if k = j ,




1
6 σj−1/2 hj if k = j − 1 ,


rjk = (3.26)
1
6 σj+1/2 hj+1 if k = j + 1 ,





0 otherwise .


The special case in which σ is constant on [0, L] and the subdivision is equidistant with
stepsize h yields the simplified form
1 2 1

R = σh tridiag 6 3 6 . (3.27)

• Computation of f .
We have Z Z
fj = f ϕj + f ϕj .
Ij Ij+1

For a generic f we cannot calculate the integrals exactly. Approximating f by constant

values fj−1/2 in Ij and fj+1/2 in Ij+1 , leads us to define
Z Z
1

fj = fj−1/2 ϕj + fj+1/2 ϕj = fj−1/2 hj + fj+1/2 hj+1 .
Ij Ij+1 2

Rb
A popular alternative to this formula is obtained by applying the trapezoidal rule a g '
1
2 (g(a) + g(b))(b − a) to the function g = f ϕj in each interval; in this case one sets

1 1
fj = f (xj−1 )ϕj (xj−1 ) + f (xj )ϕj (xj ) + f (xj+1 )ϕj (xj+1 ) ;
2 2

recalling the ϕj vanishes at xj−1 and xj+1 , and equals 1 at xj , we obtain the formula

hj + hj+1
fj = f (xj ) . (3.28)
2

3.4 Error analysis

In this final section, we particularize the general error estimates given in Sect. 2.4 for the one-
dimensional Galerkin setting considered in this chapter, using the approximation error results
given in Sect. 3.2.
On the interval I, let u be the solution of a variational problem [VP ] as in (3.22) and let uh be
the solution of the Galerkin problem [VP ]h as in (3.22), where V is a closed subspace of H 1 (I)
(such as H 1 (I) itself, or H01 (I) as considered in the example of Sect. 3.3) and Vh = V (Th ; k) ∩ V
is a subspace of V of continuous, piecewise polynomial functions on some partition Th in I.
We start with the a priori analysis.
36 CHAPTER 3. DISCRETIZATIONS IN ONE SPACE VARIABLE

3.4.1 A priori error estimates

To get an a priori error estimate in the H 1 -norm, we recall the general estimate given in Theorem
2.4.2
ku − uh kH 1 (I) ≤ Ca distH 1 (I) (u, Vh ) .
The distance between u and Vh can be bounded by the distance between u and the interpolant
Ihk u ∈ Vh introduced in (3.17),

distH 1 (I) (u, Vh ) ≤ ku − Ihk ukH 1 (I) .

Assuming that u ∈ H r (I) for some r satisfying 1 ≤ r ≤ k + 1, we have from the interpolation
error bounds (3.18)

ku − Ihk uk2H 1 (I) = ku − Ihk uk2L2 (I) + |u − Ihk u|2H 1 (I)

≤ c20,r,k h2r |v|2H r (I) + c21,r,k h2r−2 |v|2H r (I)
= (c20,r,k h2 + c21,r,k ) h2r−2 |v|2H r (I) .

Since h ≤ |I| = L, we deduce the existence of a constant Cr,k > 0 such that

ku − Ihk ukH 1 (I) ≤ Cr,k hr−1 |v|H r (I) .

Concatenating the previous inequalities, we arrive at the following result.

Proposition 3.4.1 Under the previous assumptions, the Galerkin discretization error satisfies
the a priori bound
ku − uh kH 1 (I) ≤ C hr−1 |v|H r (I) , (3.29)
where C = Ca Cr,s and 1 ≤ r ≤ k + 1.

In particular, for a piecewise linear discretization (k = 1) and u ∈ H 2 (I), we have

ku − uh kH 1 (I) ≤ C h |v|H 2 (I) , (3.30)

whereas for a piecewise quadratic discretization (k = 2) and u ∈ H 3 (I), we have

ku − uh kH 1 (I) ≤ C h2 |v|H 3 (I) . (3.31)

3.4.2 Practical estimation of the rate of decay of the error

The previous proposition suggests a law of decay of the discretization error of the form

ku − uh kH 1 (I) ∼ K hq (3.32)

where K and q depend upon the regularity of the exact solution and the polynomial degree used
in the discretization. Indeed, this is the kind of observed behavior for a generic solution, when
the mesh-size is successively reduced.
A simple interpretation of this law is as follows. Let εh = ku − uh kH 1 (I) be the error produced
L
by a partition of mesh-size h (to fix the ideas, a uniform partition with h = N ), and let
3.4. ERROR ANALYSIS 37

h
εh/2 = ku − uh/2 kH 1 (I) be the error produced by the partition of mesh-size 2 obtained by
halving each element. Applying (formally) the law (3.32), we get

εh K hq
∼ = 2q ,
εh/2 K ( h2 )q

i.e., the effect of halving h is to divide the error roughly by 2q (e.g., by 2 in the linear case
(3.30) and by 4 in the quadratic case (3.31)). If a known solution is available, and one is
able to compute the discretization errors εh , εh/2 , εh/4 , . . . for successive refinements, one can
experimentally produce an approximation of the exponent q, and check the validity of the law
(3.32).

A more sophisticated strategy leads to identifying both K and q in (3.32). Taking the loga-
rithm of both sides, one formally obtains

log εh ∼ q log h + log K ,

which shows that in a log-log scale, h and εh are placed along a straight line with slope equal
to q. A linear regression, based on the computation of εh for several values of h, yields an
approximation of K and q.

3.4.3 A posteriori error estimates. Adaptivity

Combining (2.26) with (2.28), we have

1 F (v − vh ) − a(uh , v − vh )
ku − uh kH 1 (I) ≤ sup , ∀vh ∈ Vh . (3.33)
α v∈V \{0} kvkH 1 (I)

Let us choose vh = Ih1 v ∈ Vh , the piecewise linear interpolant of v at the nodes of the partition,
and let us set w = v − Ih1 v for convenience. Then, recalling (3.20)-(3.21), we can write
Z
F (w) − a(uh , w) = f w − (µ u0h w0 + β u0h w + σ uh w)
I
N Z !
X
= f w − (µ u0h w0 + β u0h w + σ uh w)
j=1 Ij

N Z !
X
= f w − (−(µ u0h )0 w + β u0h w + σ uh w)
j=1 Ij

N Z
X
= (f − Luh ) w
j=1 Ij

where the performed integration by parts does not produce boundary terms since w vanishes at
all nodes of the partition. Recalling (3.12), we can bound each integral as
Z
(f − Luh ) w ≤ k(f − Luh )kL2 (Ij ) kwkL2 (Ij ) ≤ c1 hj k(f − Luh )kL2 (Ij ) |v|H 1 (Ij ) ,
Ij
38 CHAPTER 3. DISCRETIZATIONS IN ONE SPACE VARIABLE

whence,
 1/2 1/2
XN XN
F (w) − a(uh , w) ≤ c1  h2j k(f − Luh )k2L2 (Ij )   |v|2H 1 (Ij )
j=1 j=1
 1/2
XN
= c1  h2j k(f − Luh )k2L2 (Ij )  |v|H 1 (I) .
j=1

Substituting in (3.33) and simplifying, we arrive at the following result.

Proposition 3.4.2 Define the error estimator

 1/2
XN
2 
ηh =  ηh,j , with ηh,j = hj k(f − Luh )kL2 (Ij ) . (3.34)
j=1

Then, the following a posteriori upper bound for the Galerkin error holds true:
c1
ku − uh kH 1 (I) ≤ ηh . (3.35)
α
The quantity ηh is computable in practice, to the extent that the integrals expressing the L2 -
norms in (3.34) are computable. This occurs, e.g., if the forcing term f and the coefficients µ, β,
σ of the operator are piecewise polynomial on the partition; otherwise, one should approximate
them by piecewise polynomials, introducing an extra term in the bound above, called data
oscillation. For simplicity, we assume here that there is no data oscillation.
The information provided by the estimator ηh can be used to define a strategy of mesh
adaptation in which, starting from an initial partition T (0) = Th0 , one generates a sequence of
partitions T (`) = Th` , ` = 1, 2, . . . by subsequent refinements based on the estimator, until a
target accuracy is reached. Often, one step of adaptation is expressed by the chain

SOLVE → ESTIMATE → MARK → REFINE

Indeed, suppose that, for some chosen tol > 0, we want to be guaranteed that

ku − uh kH 1 (I) ≤ tol . (3.36)

By (3.35), this happens if our estimator ηh satisfies

α
ηh ≤ tol .
c1
So, on the current partition Th , we compute the Galerkin solution (SOLVE), and then the es-
timator (ESTIMATE). If the desired inequality is satisfies, then we stop; otherwise, we must
identify the elements of the partition which carry the largest errors (MARK), and generate a
new partition e.g. by halving each marked element (REFINE). The marking procedure can be
accomplished by a so-called greedy algorithm, such as the following one (known as Dörfler mark-
ing and very popular):
3.4. ERROR ANALYSIS 39

i) the local estimators ηh,j are ordered by decreasing size

ηh,j1 ≥ ηh,j2 ≥ · · · ≥ ηh,jN ;

ii) the marked elements are Ij1 , Ij2 , . . . , IjM , where M ≤ N is the smallest integer for which
M
X N
X
2 2
ηh,j m
≥ϑ ηh,j
m=1 j=1

for some fixed ϑ ∈ (0, 1].

One can prove that the target condition (3.36) is fulfilled after at most a number of refinements
proportional to | log TOL|. In addition, the resulting partition is near-optimal, i.e., it compares
in terms of number of elements with the smallest partition that realizes (3.36).
40 CHAPTER 3. DISCRETIZATIONS IN ONE SPACE VARIABLE
Chapter 4

General Finite Elements

bla bla bla

4.1 Triangulations of the domain

bla bla bla

hT = max hE ; (4.1)
E∈T
we will set h = hT when no confusion is possible.

4.2 Basic concepts of finite elements

The domain Ω ⊂ Rd (d = 1, 2, 3) in which we want to solve a given boundary-value problem is
partitioned into geometric elements, which will be denoted by E. Typical examples are intervals
in 1D, triangles or quadrangles in 2D, tetrahedra, esahedra, prisms or pyramids in 3D; one may
also consider images of such objects under smooth mappings, e.g., triangle with curvilinear edges
to fit the boundary of Ω.
In each E, we choose a finite dimensional space of functions, denoted by VE with dimension
NE . These functions will contribute to define the shape of the discrete solution of our problem
(in this sense, they are called elemental shape functions, or trial functions), and will allow us
to enforce the differential equation in a weak sense (for this, they are called elemental test
functions).
Typically, the space VE is made of algebraic polynomials, or images of such polynomials under
a smooth mapping; sometimes, VE may just contain a subspace of (images of) polynomials. In
dimension d > 1, different spaces of polynomials can be considered for the same degree k. In
2D,
Pk = span {xi y j : i, j ≥ 0, i + j ≤ k}
is the space of polynomials of total degree ≤ k, whose dimension is
 
k+2 (k + 1)(k + 2)
dim Pk = = .
2 2
On the other hand,
Qk = span {xi y j : 0 ≤ i, j ≤ k}

41
42 CHAPTER 4. GENERAL FINITE ELEMENTS

is the space of polynomials of individual degree ≤ k in each variable, whose dimension is

dim Qk = (k + 1)2 .
Note that each pair (i, j) of exponents defines the coordinates of a point in the lattice N × N
in the plane. For Pk such points are contained in the triangle of vertices (0, 0), (k, 0), (0, k),
whereas for Qk such points are contained in the square of vertices (0, 0), (k, 0), (0, k), (k, k). This
is a quite heuristic motivation for the popular choice of using polynomials in Pk over triangles,
and polynomials of Qk over quadrangles.
In 3D, we have
Pk = span {xi y j z l : i, j, l ≥ 0, i + j + l ≤ k} ,
with  
k+3 (k + 1)(k + 2)(k + 3)
dim Pk = = ,
3 6
and
Qk = span {xi y j z l : 0 ≤ i, j, l ≤ k} , with dim Qk = (k + 1)3 .
In this case, Pk is naturally associated with a tetrahedron, Qk with an exahedron. By varying
the bounds of the exponents i, j, l, one can easily define intermediate spaces between Pk and Qk ,
which are naturally associated with a prism or a pyramid.
In any space dimension, we will deal with restrictions of polynomials to our geometric element
E. The spaces of such restrictions will be denoted by Pk (E), Qk (E), and so on.
Since VE has dimension NE , functions in this space depend upon NE parameters. Hence, to
identify any such function v, we must assign NE independent conditions. A natural way to do
this, is to prescribe the values of a set of selected functions ` depending on v, i.e., ` : VE → R.
The most convenient choice is to use linear functions, i.e., linear forms ` on VE . These are
typically defined by the value of v or some of its derivatives at same points in E, or by the
integral of v over some subset in E; for instance,
• `(v) = v(x? ), for some chosen x? ∈ E;
∂v
• `(v) = ∂xi (x? ), for some i ∈ {1, . . . , d} ;
R
• `(v) = E v;
R
• `(v) = Γ v, for some Γ ⊆ ∂E .
Summarizing, we want to identify a function v ∈ VE by the values
αj = `j (v), j = 1, . . . , NE
of NE linear forms on VE . This is possible if and only if the mapping
` = (`j ) : VE → RNE
(4.2)
v 7→ (`j (v))1≤j≤NE
is invertible. Since both spaces have the same dimension, an equivalent condition is that ` is
injective, i.e., for any v ∈ VE
`j (v) = 0 for all j = 1, . . . , NE =⇒ v = 0. (4.3)
If this happens, we say that the set LE = {`j : VE → R, 1 ≤ j ≤ NE } is unisolvent for VE ,
and the linear forms `j are called the degrees of freedom of the functions in VE .
4.3. FINITE ELEMENTS FROM A MATHEMATICAL PERSPECTIVE 43

4.3 Finite elements from a mathematical perspective

We are ready to introduce an axiomatic definition of finite element, due to Ciarlet (1975).
Definition 4.3.1 A finite element in Rd is a triple (E, VE , LE ), where:

1. E is a non-empty compact and connected set in Rd , such that E = E̊ and the boundary
∂E is Lipschitz-continuous.

2. VE is a linear space of functions defined in E, of finite dimension NE .

3. LE = {`j : 1 ≤ j ≤ NE } is a set of linear forms `j : VE → R, which is unisolvent for VE .

Associated with any finite element, there is a canonical basis in VE which allows us to represent
any function in terms of its degrees of freedom, as well as a canonical projection operator, which
allows us to approximate functions outside VE by functions in VE .
The canonical basis ΦE is formed by the functions ϕk ∈ VE , 1 ≤ k ≤ NE , uniquely defined by
the following conditions
`j (ϕk ) = δjk , 1 ≤ j ≤ NE , (4.4)
where δjk is the Kroneker symbol. Note that ϕk corresponds to the element ek = (δjk )j of the
canonical basis in RNE , via the inverse of the mapping ` introduced in (4.2). Thus, any v ∈ VE
can be canonically represented as
NE
X
v= `k (v)ϕk . (4.5)
k=1
In order to define the canonical projection operator ΠE , we assume that the degrees of freedom
`j are indeed defined on a larger (infinite dimensional) space VE containing VE . Then, we set
NE
X
ΠE : VE → VE such that ΠE v = `k (v)ϕk ∀v ∈ VE . (4.6)
k=1

Note that by (4.5) ΠE v = v for all v ∈ VE , which proves that ΠE is indeed a projection upon
VE .
A finite element in which each degree of freedom is the value of v at a selected point in E is
termed a Lagrangian finite element. The projection operator ΠE is the Lagrange interpolation
operator at the set of selected nodes.
A finite element in which each degree of freedom is the value of v or of some partial derivative
of v at a selected point in E is termed a Hermitian finite element.

4.3.1 Finite elements defined by affine mappings

Starting from a finite element (Ê, V̂ , L̂) on some “reference” element Ê, one can generate a
finite element (E, VE , LE ) on some region E, by means of some smooth and invertible mapping
F : Rd → Rd . This is a rather common procedure, especially when F is affine.
(..... citare esempio Courant visto nella Sez. 1.....)
The following notation will be helpful in the sequel:

i) If x̂ and x are points in Rd , the symbol x ↔ x̂ means x = F(x̂) or, equivalently, x̂ =

F−1 (x).
44 CHAPTER 4. GENERAL FINITE ELEMENTS

ii) If v̂ and v are real functions defined on subsets of Rd , the symbol v ↔ v̂ means v = v̂ ◦ F−1
or, equivalently, v̂ = v ◦ F; in other words, it holds v(x) = v̂(x̂) whenever x ↔ x̂.

iii) At last, if `ˆ and ` are linear forms defined on spaces of real functions, the symbol ` ↔ `ˆ
means `(v) = `(v ˆ ◦ F) for any v or, equivalently, `(v̂)
ˆ = `(v̂ ◦ F−1 ) for any v̂; in other
ˆ
words, it holds `(v) = `(v̂) whenever v ↔ v̂.

Given the mapping F, the finite element (E, VE , LE ) is defined as follows:

E = F(Ê) ,
VE = {v : v ↔ v̂ for some v̂ ∈ V̂ } , (4.7)
LE = {`j : `j ↔ `ˆj with `ˆj ∈ L̂, 1 ≤ j ≤ N̂ } ,

where N̂ = dim V̂ . It is easily checked that dim VE = dim V̂ and that LE is unisolvent for VE .
Thus, the axioms of finite element are satisfied.
Furthermore, it is immediate to check that the canonical bases in VE and V̂ are related by

ϕj ↔ ϕ̂j , 1 ≤ j ≤ N̂ , (4.8)

and the canonical projectors on VE and V̂ are related by

ΠE v ↔ Π̂ v̂ whenever v ↔ v̂ , (4.9)

which we will also write as

(ΠE v)ˆ = Π̂ v̂ whenever v ↔ v̂ , (4.10)

If the degrees of freedom in L̂ are defined on a larger space V̂ ⊃ V̂ , then the degrees of freedom
in LE are defined on VE = {v : v ↔ v̂ for some v̂ ∈ V̂} ⊃ VE .

4.4 Examples
In this section, we will present several examples of finite elements and discuss their properties.
In preparation for that, we begin by establishing a useful result, which extends a well-known
property of polynomials in one space variable, namely: if v is a polynomial of degree k ≥ 1 in one
space variable satisfying v(x0 ) = 0 for some x0 ∈ R, then it factorizes as v(x) = (x − x0 )w(x),
where w is a polynomial of degree k − 1; if in addition v 0 (x0 ) = 0, then v factorizes as v(x) =
(x − x0 )2 u(x), where u is a polynomial of degree k − 2.

Lemma 4.4.1 Let L be a segment in the plane, not reduced to a point. If v belongs to Pk (R2 )
(for some k ≥ 1) and vanishes on L, then v factorizes as

v = eL w , with eL ∈ P1 (R2 ) and w ∈ Pk−1 (R2 ) ,

where eL (x, y) = 0 is the equation of the line containing L. If in addition the normal derivative
∂v
to the edge ∂n vanishes on L, then v factorizes as

v = (eL )2 u , with u ∈ Pk−2 (R2 ) .

A similar result holds in dimension 3, if now L is a polygon not reduced to a segment.

4.4. EXAMPLES 45

Proof. Assume first that the segment L lies on the y-axis. Since v(0, y) is a polynomial in y
that vanishes at the infinite points of L, it vanishes identically on the y-axis. Thus, for any fixed
y, x 7→ v(x, y) is a polynomial in one variable that vanishes at x = 0, hence, it can be factorized
as v(x, y) = x w(x, y), where w is a polynomial of degree ≤ k − 1. Since x = 0 is precisely the
equation of the y-axis, the first result is proven.
If L is arbitrarily placed in the plane, let us fix an orthogonal coordinate system ξ-η such
that L lies on the η-axis, and let (ξ, η) = Φ(x, y) be the affine change of variable between the
two systems of coordinates. Then, ṽ(ξ, η) = v(Φ−1 (ξ, η)) is a polynomial of degree ≤ k in the
variables ξ, η that vanishes on a segment on the axis ξ = 0, hence it factorizes as ṽ(ξ, η) =
ξ w̃(ξ, η) with w̃ polynomial of degree ≤ k − 1. Going back to the original coordinates, we get
v(x, y) = ξ(x, y) w(x, y), where ξ(x, y) = 0 is precisely the equation of the η-axis expressed in
the x, y-coordinates.
The proof of the second result can be done by similar arguments. The extension to the
three-dimensional case is trivial.

4.4.1 Triangular elements

Let E = T be a non-degenerate triangle in the plane, with vertices ai = (ai1 , ai2 ), i = 1, 2, 3,
that we suppose to be ordered counterclockwise. Let us denote its barycenter by

xB = 13 (a1 + a2 + a3 )

and its area by  

1 a11 − a31 a12 − a32
|T | = 2 det .
a21 − a31 a22 − a32
We also introduce the edges Li = [ai , ai+1 ] and the midpoints of the edges ai,i+1 = 12 (ai + ai+1 ),
i = 1, 2, 3. Note that, here and in the sequel, we identify cyclically the index 4 = 3 + 1 with the
index 1.
The reference element Ê is chosen as the standard symplex in the plane, i.e.,

T̂ = {(x, y) ∈ R2 : 0 ≤ x, y, x + y ≤ 1} .

A. Constant polynomial space

We choose the space VE = VT = P0 (T ), whose dimension is 1. As unique degree of freedom for
a polynomial v ∈ P0 (T ), we can choose – for instance – its value at the barycenter,

`(v) = v(xB )

or its integral average Z

1
`(v) = v.
|T | T
In both cases, it is trivially seen that the triple (T, P0 (T ), LT = {`}) satisfies the axioms of finite
element.
Note that for the same polynomial v the value of the two degrees of freedom is the same;
however, the space VE = VT on which these degrees of freedom can be extended is different: the
value at the barycenter requires the function v to be defined pointwise, so a natural choice is
VT = C 0 (T ), whereas the integral value only requires v to be integrable, so for instance we can
choose VT = L1 (T ) or VT = L2 (T ).
46 CHAPTER 4. GENERAL FINITE ELEMENTS

B. Linear polynomial space

We choose the space VE = VT = P1 (T ), whose dimension is 3. Several choices of a set LT of
three degrees of freedom are possible.
(1)
• The set LT of the values at the vertices:

v(a1 ), v(a2 ), v(a3 ) .

(2)
• The set LT of the values at the midpoints of the edges:

v(a1,2 ), v(a2,3 ), v(a3,1 ) .

(3)
• The set LT of the integral averages on the edges:
Z Z Z
1 1 1
v, v, v.
|L1 | L1 |L2 | L2 |L3 | L3

(4)
• The set LT of the moments of order ≤ 1 on the triangle:
Z Z Z
1 1 1
v, vx , vy .
|T | T |T | T |T | T

With any of these choices for LT , the triple (T, P1 (T ), LT ) satisfies the axioms of finite element.
(1)
Let us check that the set LT of degrees of freedom is unisolvent for P1 (T ). There are
many ways to do this; here, let us apply Lemma 4.4.1. Suppose that v ∈ P1 (T ) satisfies
v(a1 ) = v(a2 ) = v(a3 ) = 0. Then, v vanishes identically on each edge of the triangle. Indeed,
let us consider for instance the edge L1 = [a1 , a2 ] and let

x(t) = a1 + t(a2 − a1 ) , 0 ≤ t ≤ 1, (4.11)

be the affine parametrization of this edge. The function η(t) = v(x(t)) is a polynomial of degree
≤ 1 in the variable t, and satisfies η(0) = v(a1 ) = 0 and η(1) = v(a2 ) = 0; hence, it vanishes
identically, proving that v|L1 vanishes identically. By Lemma 4.4.1, we can write v = eL1 w with
w a polynomial of degree 0, i.e., a constant. If we compute v at a3 , we have

0 = v(a3 ) = eL1 (a3 ) w

and eL1 (a3 ) 6= 0 since a3 does not lie on the line containing L1 . Hence, necessarily w = 0, which
proves that v is identically zero.
(1)
The finite element (T, P1 (T ), LT ) is called the Courant element.
(2)
A similar argument proves that the set LT is unisolvent for P1 (T ). In general, so is the set
of degrees of freedom given by the values of v at any three points in T which are not aligned.
The corresponding finite elements are of Lagrangian type. Such degrees of freedom are defined
on VT = C 0 (T ).

(3) (4)
Example 4.4.2 Prove that the sets LT and LT are unisolvent for P1 (T ).
4.4. EXAMPLES 47

Figure 4.1: Various examples of finite elements with their degrees of freedom

C. Quadratic polynomial space

We choose the space VE = VT = P2 (T ), whose dimension is 6. A natural set LT of degrees of
freedom is given by the values of v at the vertices and the midpoints of the edges, i.e.,

v(a1 ), v(a2 ), v(a3 ), v(a1,2 ), v(a2,3 ), v(a3,1 ) .

This set is unisolvent for P2 (T ), thus (T, P2 (T ), LT ) is a Lagrangian finite element.

To show this, suppose that v ∈ P2 (T ) vanishes at the selected points. Consider again the
parametrization (4.11) of the edge L1 = [a1 , a2 ]. In this case, the function η(t) = v(x(t)) is a
polynomial of degree ≤ 2 in the variable t, and satisfies η(t) = 0 for t = 0, 12 , 1; hence it vanishes
identically. Therefore, by Lemma 4.4.1, we can factorize v as v = eL1 w with w a polynomial
of degree ≤ 1. Necessarily w vanishes at the three points v(a3 ), v(a2,3 ), v(a3,1 ), which do not
belong to the line containing L1 and are not aligned. Applying the argument above for linear
polynomials, we conclude that w, hence v vanishes identically.
48 CHAPTER 4. GENERAL FINITE ELEMENTS

D. Cubic polynomial space

We choose the space VE = VT = P3 (T ), whose dimension is 10. We propose two sets of degrees
(1)
of freedom. The first one, LT , is given by the values

v(aij ) , 0 ≤ i, j, i + j ≤ 3 ,

at the ten nodes in T defined as follows: divide each edge into three segments of equal length,
then draw the three family of lines, each one parallel to an edge, passing through the subdivision
points (including the vertices) (see Figure...). The desired nodes are the points in T where three
lines intersect.
Extending the arguments given above for the quadratic case to the cubic case, one can show
(1) (1)
that LT is unisolvent for P3 (T ), hence (T, P3 (T ), LT ) is a Lagrangian finite element.

Remark 4.4.3 More generally, the previous construction of nodes can be done starting from
the partition of each edge into k > 3 segments of equal length. One obtains a total of (k+1)(k+2)
2
nodes aij with 0 ≤ i, j and i + j ≤ k. The values of v at these nodes form a unisolvent set of
Lagrangian degrees of freedom in Pk (T ).  
i j
In the reference element Ê = T̂ , the coordinates of these nodes are âij = k, k .
The use of such finite elements is limited to small values of k. Indeed, since the nodes are
uniformly spaced in the element, the corresponding interpolation operator suffers from numerical
instability for increasing values of k.

(2)
The second set, LT , of degrees of freedom for P3 (T ) leads to our first example of a Hermitian
finite element. The ten degrees of freedom are the values of v and its gradient at the three
vertices, plus the value of v at the barycenter, i.e.,

v(aj ) , ∇v(aj ) , for 1 ≤ j ≤ 3 , and v(xB ) .

Let us check that this set is unisolvent. Suppose that v ∈ P3 (T ) has all such degrees of freedom
equal to 0. Consider again the parametrization (4.11) of the edge L1 = [a1 , a2 ]. In this case,
the function η(t) = v(x(t)) is a polynomial of degree ≤ 3 in the variable t; its derivative with
respect to t is given by η 0 (t) = ∇v(x(t)) · (a2 − a1 ). One has

η(0) = η 0 (0) = 0 , η(1) = η 0 (1) = 0 ,

which implies that η factorizes as η(t) = t2 (1 − t)2 u(t); but since the degree of η is ≤ 3,
necessarily u, hence η, is identically 0. This proves that v vanishes on L1 , and similarly on L2
and L3 . Applying Lemma 4.4.1 repeatedly, we obtain the factorization v = eL1 eL2 eL3 w, where
w is a constant. Finally, we use the last condition,

0 = v(xB ) = eL1 (xB ) eL2 (xB ) eL3 (xB )w

to deduce that necessarily w = 0 since xB does not lie on any edge; hence, v is identically 0.
(2)
For the Hermitian element (T, P3 (T ), LT ), it is natural to choose VE = VT = C 1 (T ).
4.4. EXAMPLES 49

E. Piecewise cubic polynomial space

We present here an important example of finite element on a triangle T , in which the function
space VT is not made of polynomials, although it contains a polynomial space.
Consider the sub-partition of T into the three triangles Ti , i = 1, 2, 3 obtained by joining each
vertex with the barycenter. Define the following space of piecewise polynomial functions

VT = {v ∈ C 1 (T ) : v|Ti ∈ P3 (Ti ), i = 1, 2, 3} .

Obviously, P3 (T ) ⊂ VT . But the latter space contains functions that are not global polynomials,
since
dim VT = 12 .
Indeed, to build a function in this space, one may start from three independent polynomials
vi ∈ P3 (Ti ), i = 1, 2, 3, which globally depend on 30 coefficients, and enforce the continuity
of these functions and their gradients across the interfaces between triangles; one can show
that such a ‘glueing’ process amounts to enforcing 18 independent linear conditions, whence the
dimension 12 = 30 − 18 of VT .
As a set LT of degrees of freedom, one chooses the values of the function v and its gradient
at the vertices of the triangle, plus the values of the normal derivative at the midpoints of the
edges, i.e.,
∂v
v(aj ) , ∇v(aj ) , (aj,j+1 ) for 1 ≤ j ≤ 3 .
∂n
One can prove that this set of degrees of freedom in unisolvent for VT , which makes the triple
(T, VT , LT ) a Hermitian finite element. It is called the Hsieh-Clough-Tocher element, or
simply the HCT element. We will come back to this element later on.

F. Quintic polynomial space

Let us now choose the space VE = VT = P5 (T ), whose dimension is 21. As a set LT of degrees
of freedom, we may chose the values of v, its gradient and its Hessian matrix at the vertices,
plus the values of the normal derivative at the midpoint of the edges; thus,

∂v
v(aj ) , ∇v(aj ) , Hv(aj ) , (aj,j+1 ) for 1 ≤ j ≤ 3 .
∂n
n 2 o
Note that the Hessian matrix Hv = ∂x∂p ∂x v
q
is symmetric by Schwarz’ theorem, thus prescrib-
ing it at a point amounts to prescribe 3 linear conditions. It follows that LT contains exactly
21 degrees of freedom.
In order to prove that it is unisolvent, consider a function v ∈ P5 (T ) that has all such degrees
of freedom equal to 0. Let us use again the parametrization (4.11) of the edge L1 = [a1 , a2 ]. In
this case, the function η(t) = v(x(t)) is a polynomial of degree ≤ 5 in the variable t; its first
derivative with respect to t is given by η 0 (t) = ∇v(x(t)) · (a2 − a1 ), whereas its second derivative
is given by η 00 (t) = (a2 − a1 ) · Hv(x(t))(a2 − a1 ). Hence, one has

η(0) = η 0 (0) = η 00 (0) = 0 , η(1) = η 0 (1) = η 00 (1) = 0 ,

which implies that η factorizes as η(t) = t3 (1 − t)3 %(t) for some %; but since the degree of η is
≤ 5, necessarily %, hence η, is identically 0. This proves that v vanishes on L1 .
50 CHAPTER 4. GENERAL FINITE ELEMENTS

Let us now consider the normal derivative of v on L1 , which we parametrize by ζ(t) =

∂v
= ∇v(x(t)) · n, where n denotes a unit vector perpendicular to L1 . The function ζ(t) is
∂n (x(t))
a polynomial of degree ≤ 4 in the variable t, whose first derivative is η 0 (t) = (a2 −a1 )·Hv(x(t))n.
One has
ζ(0) = ζ( 21 ) = ζ(1) = 0 , ζ 0 (0) = ζ 0 (1) = 0 ,
which implies that ζ factorizes as ζ(t) = t2 (t − 12 )(1 − t)2 σ(t) for some σ; but since the degree
∂v
of ζ is ≤ 4, necessarily σ, hence ζ, is identically 0. This proves that ∂n vanishes on L1 .
What we have proven for L1 holds for the two other edges as well. Hence, applying Lemma
4.4.1, we deduce that v factorizes as

v = (eL1 )2 (eL2 )2 (eL2 )2 u .

Since v is a polynomial of degree ≤ 5, necessarily u, hence v, vanish identically in T . This

concludes our proof.
The Hermitian finite element (T, P5 (T ), LT ) is called the Argyris element.

4.4.2 Quadrangular elements

Let E = R be a non-degenerate parallelogram in the plane, with vertices ai = (ai1 , ai2 ), i =
1, 2, 3, 4, which we suppose to be ordered counterclockwise. Let us denote its barycenter by

xB = 41 (a1 + a2 + a3 + a4 )

and its area by  

a11 − a41 a12 − a42
|R| = det .
a31 − a41 a32 − a42
We also introduce the edges Li = [ai , ai+1 ] and the midpoints of the edges ai,i+1 = 12 (ai + ai+1 ),
i = 1, 2, 3, 4. Note that now we identify cyclically the index 5 = 4 + 1 with the index 1.
The reference element Ê is chosen as the standard square in the plane, i.e.,

R̂ = {(x, y) ∈ R2 : 0 ≤ x, y ≤ 1} .

Let us just highlight three examples, without proofs.

G. Bilinear polynomial space

We choose the space VE = VR = Q1 (R), whose dimension is 4. The set LR of the Lagrangian
degrees of freedom
v(ai ) , 1 ≤ i ≤ 4,
is unisolvent for Q1 (R).

H. Biquadratic polynomial space

We choose the space VE = VR = Q2 (R), whose dimension is 9. The set LR of the Lagrangian
degrees of freedom

v(ai ) , 1 ≤ i ≤ 4, v(ai,i+1 ) , 1 ≤ i ≤ 4, and v(xB )

is unisolvent for Q2 (R).

4.5. FINITE ELEMENTS OF CLASS C 0 AND C 1 51

K. Bicubic polynomial space

We choose the space VE = VR = Q3 (R), whose dimension is 16. A set LR of Hermitian degrees
of freedom, unisolvent for Q3 (R), is given by

v(ai ) , 1 ≤ i ≤ 4, ∇v(ai ) , 1 ≤ i ≤ 4, ∂i2 v(ai ) , 1 ≤ i ≤ 4,

where ∂i2 v(ai ) denotes the second partial derivative of v taken along the two directions of the
edges that meet at ai .

4.4.3 3D elements
... ...

4.5 Finite elements of class C 0 and C 1

So far, we have focussed on one single finite element. Now, we wish to put together several
elements on a given partition of a domain Ω, and generate a finite dimensional space of functions
associated with the partition.
Let T = {E} be an admissible partition of Ω. Let us choose the same type of finite element
(e.g., the Courant element, or the Argyris element) in each E ∈ T ; let us denote it by (E, VE , LE ).
The largest space of functions defined in Ω and associated with this choice can be chosen as

V(T ) = {v ∈ L2 (Ω) : v|E ∈ VE ∀E ∈ T } . (4.12)

In many cases, we will need subspaces of this space, formed by functions with some kind of
global continuity (in the classical sense), or regularity (in the modern functional sense). So,
subspaces of interest could be for instance

V(T ) ∩ C 0 (Ω) , V(T ) ∩ C 1 (Ω) , or V(T ) ∩ H 1 (Ω) , V(T ) ∩ H 2 (Ω) .

Assuming that functions in each VE have the requested continuity or regularity, we are faced with
the problem of ‘glueing together’ functions across the interfaces between contiguous elements.
To be precise, suppose that E1 and E2 are two elements of the partition, which share a full
edge (in 2D) or face (in 3D); let L = ∂E1 ∩ ∂E2 denote this common part. Given two functions
v1 ∈ VE1 and v2 ∈ VE2 , let us denote by v the function defined in E1 ∪ E2 as
(
v1 on E1 ,
v= (4.13)
v2 on E2 .

If VEi ⊂ C 0 (Ei ) for i = 1, 2, then obviously

v ∈ C 0 (E1 ∪ E2 ) if and only if v1 |L = v2 |L .

The same equivalence holds if we replace C 0 by H 1 everywhere, according to Property 1.0.1.

On the other hand, if VEi ⊂ C 1 (Ei ) for i = 1, 2, then

v ∈ C 1 (E1 ∪ E2 ) if and only if v1 |L = v2 |L and (∇v1 )|L = (∇v2 )|L .

52 CHAPTER 4. GENERAL FINITE ELEMENTS

Note that such continuity conditions are equivalent to the conditions

∂v1 ∂v2
v1 |L = v2 |L and = ,
∂n |L ∂n |L
where both normal derivatives are taken with respect to the same unit vector n perpendicular to
L; indeed, denoting by τ a unit vector aligned with L and recalling the orthogonal decomposition
of a gradient
∂v ∂v
∇v = (∇v · τ ) τ + (∇v · n) n = τ+ n,
∂τ ∂n
one observes that the continuity of v across L implies the continuity of the tangential derivative,
whence the gradient is continuous across L if and only if the normal derivative is.
As above, the same equivalence holds if we replace C 1 by H 2 everywhere.
These considerations induce us to pose the following fundamental questions:
• Is the set L of degrees of freedom in our finite element such that the process of ‘glueing’
functions across element interfaces is particularly simple?

• In particular, can we identify degrees of freedom that are local on each edge/face, so that
if we give them the same value on both sides of an interface between two elements, we
enforce automatically the desired continuity?

4.5.1 Finite elements of class C 0

Let us discuss a simple, yet instructive example. Consider a triangular element (T, P1 (T ), LT ),
where LT is one of the sets of degrees of freedom introduced in Case B of the previous section.
Assume that the triangles T1 and T2 share a common edge L = [a1 , a2 ]. If vi ∈ P1 (Ti ), i = 1, 2,
their restrictions to L belong to P1 (L), i.e., they are polynomials of degree ≤ 1 in the variable t
that parametrizes the edge.
(1)
Let us choose the Courant element, i.e., let us choose LT = LT , the set of three degrees of
freedom given by the values of a function at the vertices of T . Note that exactly two degrees of
freedom depend only on the values on an edge. In order to glue the functions v1 and v2 across
L, it is enough to equate the values of the degrees of freedom associated with the endpoints a1
and a2 , i.e., to enforce

v1 (a1 ) = v2 (a1 ) and v1 (a2 ) = v2 (a2 ) ;

this automatically implies v1 |L = v2 |L . Equivalently, if we consider the difference function

φ = v1 |L − v2 |L ∈ P1 (L), then by equating the degrees of freedom at the endpoints of L we
enforce φ(a1 ) = 0 and φ(a2 ) = 0, and this implies φ = 0 identically on L. Here, we use the fact
that the two degrees of freedom given by λ1 (φ) = φ(a1 ) and λ2 (φ) = φ(a2 ) form a unisolvent
set for P1 (L).
(1)
We say that the Courant element (T, P1 (T ), LT ) is a finite element of class C 0 .
(2)
If we choose LT = LT , the set of three degrees of freedom given by the values of a function
at the midpoints of the edges of T , then there is only one degree of freedom that solely depends
on the values of a function on an edge. We can glue the functions v1 and v2 at the midpoint
a1,2 of L by enforcing v1 (a1,2 ) = v2 (a1,2 ), but this does not guarantee that v1 |L = v2 |L . In order
to enforce full continuity on L, we should impose a condition on the other degrees of freedom in
T1 and T2 , which is in principle feasible but not efficient.
4.5. FINITE ELEMENTS OF CLASS C 0 AND C 1 53

(3) (4)
Similar considerations apply to the other sets of degrees of freedom LT and LT discussed in
(i)
Case B of the previous section. Thus, none of the finite elements (T, P1 (T ), LT ) with i = 2, 3, 4
is of class C 0 .
We are ready for the following definition.

Definition 4.5.1 Suppose that VE ⊂ C 0 (E). We say that the finite element (E, VE , LE ) is of
class C 0 if for any edge/face L of E there exists a subset LE,L ⊂ LE of degrees of freedom with
the following properties:

i) each form ` ∈ LE,L depends only on the values taken on L by the argument v;

ii) if v ∈ VE satisfies `(v) = 0 for all ` ∈ LE,L , then v|L is identically 0.

Item i) means that any ` ∈ LE,L can be written as

`(v) = λ(v|L ) ∀v ∈ VE , (4.14)

where λ is a linear form on the space VE,L = {v|L : v ∈ VE } of the restrictions to L of the
functions in VE .
Item ii) means that the set ΛE,L collecting all such forms λ is unisolvent for VE,L .

Examples 4.5.2 Each of the finite elements presented in Cases from C to K of the previous
section is of class C 0 .
Indeed, for the quadratic element presented in Case C, the degrees of freedom relative to the
edge L are
v(a1 ), v(a2 ), v(a1,2 ) ,

which form a unisolvent set for the space P2 (L) = (P2 (T ))|L . Similar considerations apply to
the cubic Lagrangian element presented in Case D, and to their higher-order extensions.
For the cubic Hermitian element presented in Case D and the composite Hsieh-Clough-Tocher
element presented in Case E, the degrees of freedom relative to the edge L provide the values
of the forms
∂v ∂v
v(a1 ), v(a2 ), (a1 ), (a2 ) ,
∂τ ∂τ
which form a unisolvent set for the space P3 (L) = (P3 (T ))|L (for the HCT element, remember
that functions in VT are polynomials of degree ≤ 3 near any edge of T ). Note that the value of
∂v
the tangential derivative ∂τ can be obtained from the assigned value of the gradient ∇v by the
∂v
formula ∂τ = ∇v · τ .
For the Argyris element presented in Case F, the degrees of freedom relative to the edge L
provide the values of the forms

∂v ∂v ∂2v ∂2v
v(a1 ), v(a2 ), (a1 ), (a2 ) (a1 ), (a2 ) ,
∂τ ∂τ ∂τ 2 ∂τ 2

which form a unisolvent set for the space P5 (L) = (P5 (T ))|L . Here, the second tangential
∂v 2
derivative is given by ∂τ 2
= τ · (Hv)τ .
54 CHAPTER 4. GENERAL FINITE ELEMENTS

With such definition at hand, let us go back to the two elements E1 and E2 of the partition
of Ω, which share a common edge/face L. Assume that the finite elements (E1 , VE1 , LE1 ) and
(E2 , VE2 , LE2 ) are of class C 0 . Furthermore, assume that the space of the restrictions to L of
the functions in VE1 and in VE2 coincide, i.e.,

VE1 ,L = VE2 ,L (= VL , say)

and the sets of linear forms on VL coincide as well, i.e,

ΛE1 ,L = ΛE2 ,L (= ΛL , say)

(remember (4.14) after Definition 4.5.3). Given two functions v1 ∈ VE1 and v2 ∈ VE2 , let us
equate all the degrees of freedom on L, namely let us enforce

λ(v1 |L ) = λ(v2 |L ) ∀λ ∈ ΛL . (4.15)

By linearity this is equivalent to

λ(v1 |L − v2 |L ) = 0 ∀λ ∈ ΛL ,

and since by assumption ΛL is unisolvent for VL , we deduce that

v1 |L − v2 |L = 0 , i.e., v1 |L = v2 |L .

We conclude that the matched function v defined in (4.13) is continuous across L. This argument
shows that the matching (4.15) of the local degrees of freedom on L guarantees C 0 -continuity in
E1 ∪ E2 .

4.5.2 Finite elements of class C 1

Such elements allow us to build finite dimensional subspaces of H 2 (Ω), which can be used, for
instance, to define Galerkin discretizations of fourth-order problems, such as the plate bending
problem presented in Sect. ??
Let us discuss the enforcement of continuity between normal derivatives.

Definition 4.5.3 Suppose that VE ⊂ C 1 (E). We say that the finite element (E, VE , LE ) is of
class C 1 if it is of class C 0 and if for any edge/face L of E there exists a subset L0E,L ⊂ LE of
degrees of freedom with the following properties:

i) each form ` ∈ L0E,L depends only on the values taken on L by the normal derivative ∂v
∂n of
the argument v;

ii) if v ∈ VE satisfies `(v) = 0 for all ` ∈ L0E,L , then ∂v

∂n |L is identically 0.

Examples 4.5.4 Obviously, a Lagrangian finite element cannot be of class C 1 , since one cannot
infer the values of the normal derivative on an edge by the assigned degrees of freedom.
Consider at first the cubic Hermitian element presented in Case D. The degrees of freedom
relative to the edge L provide the values of the two forms
∂v ∂v
(a1 ) and (a2 ) ;
∂n ∂n
4.5. FINITE ELEMENTS OF CLASS C 0 AND C 1 55

however, if v ∈ P3 (T ), its normal derivative on L belongs to P2 (L), which has dimension 3. The
∂v
two forms do not allow us to uniquely identify ∂n on L. Consequently, the cubic Hermitian
element presented in Case D is not of class C . 1

It is precisely this pitfall that suggested to enrich the space P3 (T ), leading to the introduc-
tion of the composite Hsieh-Clough-Tocher element presented in Case E. Indeed, its degrees of
freedom relative to the edge L provide the values of the forms
∂v ∂v ∂v
(a1 ), (a1,2 ), (a2 ) ,
∂n ∂n ∂n
which form a unisolvent set for P2 (L). Hence, the Hsieh-Clough-Tocher element if of class C 1 .
At last, consider the quintic Argyris element presented in Case F, and note that if v ∈ P5 (T ),
its normal derivative on L belongs to P4 (L). The degrees of freedom relative to the edge L
provide the values of the forms
∂v ∂v ∂v ∂2v ∂2v
(a1 ), (a1,2 ), (a2 ) , (a1 ), (a2 ) ,
∂n ∂n ∂n ∂τ ∂n ∂τ ∂n
which form a unisolvent set for P4 (L). Hence, the Argyris element if of class C 1 .

The argument given above about the C 0 -matching of two functions v1 ∈ VE1 and v2 ∈ VE2
across a common interface L can be easily extended to the case of C 1 -matching. Assuming that
the same finite element of class C 1 is used in E1 and E2 , it is enough to equate all the local
degrees of freedom on L to satisfy the matching conditions
∂v1 ∂v2
v1 |L = v2 |L and = .
∂n |L ∂n |L

We conclude that the matched function v defined in (4.13) satisfies v ∈ C 1 (E1 ∪ E2 ).

4.5.3 Building global spaces of finite element functions

The matching between two contiguous elements, described in the previous subsections, can be
repeated for all elements of a triangulation T in Ω. Thus, by using the same type of finite
element of class C 0 in all the triangulation, one can easily build the global spaces V(T ) ∩ C 0 (Ω)
or V(T ) ∩ H 1 (Ω), where the space V(T ) is defined in (4.12). Similarly, if one uses the same type
of C 1 element everywhere, one can build the global spaces V(T ) ∩ C 1 (Ω) or V(T ) ∩ H 2 (Ω).
Functions in any of these spaces can be uniquely identified by assigning the values of all
degrees of freedom in each element, with the constraint that degrees of freedom associated with
an interface between two elements should receive the same value in both elements.
A canonical basis can be obtained by picking one degree of freedom, say `j , and giving it
the value 1, while giving the value 0 to all other degrees of freedom `k . Thus, the global basis
function ϕj associated with the degree of freedom `j satisfies

`j (ϕj ) = 1 , `k (ϕj ) = 0 ∀k 6= j .

The construction of the basis functions can be done locally, element by element. Local basis
functions on contiguous elements, associated with the same degree of freedom, are automatically
matched across the common interface.
For instance, if one uses the Courant element in the definition of the space V (T ) := V(T ) ∩
C 0 (Ω) = V(T ) ∩ H 1 (Ω), the global pyramidal basis function ϕj associated with a vertex aj of
56 CHAPTER 4. GENERAL FINITE ELEMENTS

the triangulation is obtained by glueing together the local linear basis functions associated with
aj in each triangle having aj as a vertex.
The global basis functions obtained in this way have local support: if the function ϕj is
associated as above with the degree of freedom `j , it will be identically 0 in all elements, except
in those which include `j in their set L of degrees of freedom.
With similar arguments, one can see that a global canonical projection operator ΠT can be
defined by glueing together the local projections operators in each element; precisely, ΠT v is
such that
(ΠT v)|E = ΠE v|E ∀E ∈ T .
Interelement continuity is automatically guaranteed by the assumption that the used finite
elements are of class C 0 or C 1 .
For instance, if the Courant element is used, ΠT is the global interpolation operator by
piecewise affine functions: in each T ∈ T , (ΠT v)|T is the affine function which interpolates v at
the three vertices of the triangle.
Chapter 5

Finite Element Approximation

Given an admissible triangulation T in the domain Ω, let (E, VE , LE ) be a finite element asso-
ciated with any E ∈ T , and let ΠE the canonical projection upon VE . As already mentioned in
Sect. 4.5.3, we can glue together the local projections to define a global projection ΠT over the
space VT defined in (4.12), by setting

(ΠT v)|E = ΠE v|E ∀E ∈ T . (5.1)

Assuming that both v and ΠT v belong to the space H ` (Ω) for some ` ≥ 0, we want to produce
a measure of the approximation error v − ΠT v, for instance by estimating the semi-norm

|v − ΠT v|`,Ω .

In general, such an error will depend upon many features, namely

i) the size of the elements of the partition, globally represented by the meshsize h;

ii) the properties of the finite elements;

iii) the smoothness of the function v.

An estimate on the global semi-norm of the error can be derived from the estimates of the
semi-norms of the local errors, namely

|v|E − ΠE v|E |`,E ∀E ∈ T ,

using the additivity property of the integral with respect to the domain of integration. We will
accomplish this task under the (common) assumption that each finite element is the affine image
of a reference element (Ê, V̂ , L̂), as defined in Sect. 4.3.1. Then, our strategy will consist of
three steps:
1. transform the error on the generic element E into an error on the reference element Ê;

2. estimate each of these errors on the reference element;

3. transform back the result from the reference element to the generic element E.
We start by performing Step 2; next, we discuss how semi-norms transform when going from
a generic element to the reference element and back (useful in Steps 1 and 3); then, we combine
the ingredients by getting estimates on the local errors, and finally on the global error.

57
58 CHAPTER 5. FINITE ELEMENT APPROXIMATION

5.1 Error bounds on the reference element

We are going to make several assumptions on the reference element (Ê, V̂ , L̂).

Assumption 5.1.1 The space V̂ contains a subspace of polynomial functions. Precisely, there
exists an integer k ≥ 0 such that
Pk (Ê) ⊆ V̂ .

Assumption 5.1.2 Functions in V̂ have a certain Sobolev regularity. Precisely, there exists an
integer m ≥ 0 such that
V̂ ⊂ H m (Ê) .

Assumption 5.1.3 The degrees of freedom in L̂ are defined and continuous in a Sobolev space.
Precisely, there exists an integer q ≥ 0 such that

H q (Ê) ⊆ V̂

and there exists a constant Ĉ1 > 0 such that for all `ˆ ∈ L̂ it holds
ˆ
|`(v̂)| ≤ Ĉ1 kv̂kq,Ê ∀v̂ ∈ H q (Ê) .

Let us illustrate these assumptions with some examples.

Examples 5.1.4 For the Courant element, one has k = 1, m is arbitrary since any polynomial
is a C ∞ function, whereas q = 2 since the degrees of freedom are well defined for functions in
C 0 (Ê) and in dimension d = 2 it holds H 2 (Ê) ⊂ C 0 (Ê) with continuous injection.
For the Hsieh-Clough-Tocher element, one has k = 3, m = 2 since functions in V̂ are made of
three pieces of polynomials which are glued in a C 1 manner, whereas q = 3 since the degrees of
freedom are well defined for functions in C 1 (Ê) and in dimension d = 2 it holds H 3 (Ê) ⊂ C 1 (Ê)
with continuous injection.
For the Argyris element, one has k = 5, m is arbitrary since any polynomial is a C ∞ function,
whereas q = 4 since the degrees of freedom are well defined for functions in C 2 (Ê) and in
dimension d = 2 it holds H 4 (Ê) ⊂ C 2 (Ê) with continuous injection.

Under the previous assumptions, let us estimate the projection error v̂ − Π̂v̂. To this end,
recalling that Π̂v̂ = N̂ ˆ
P
j=1 `j (v̂)ϕ̂j and using Assumption 5.1.3, we have
 
N̂
X XN̂
kΠ̂v̂km,Ê ≤ |`ˆj (v̂)| kϕ̂j km,Ê ≤ Ĉ1  kϕ̂j k m,Ê
 kv̂k
q,Ê =: Ĉ2 kv̂kq,Ê ∀v̂ ∈ H q (Ê) .
j=1 j=1

This shows that Π̂ is a linear continuous operator between the spaces H q (Ê) and H m (Ê), i.e.,
Π̂ ∈ L(H q (Ê), H m (Ê)).
On the other hand, the identity mapping I satisfies I ∈ L(H q (Ê), H m (Ê)) if and only if
kv̂km,Ê ≤ Ckv̂kq,Ê for all v̂ ∈ H q (Ê), which is possible if and only if m ≤ q. This motivates the
following assumption.

Assumption 5.1.5 The integers m and q satisfy the inequality

m ≤ q.
5.1. ERROR BOUNDS ON THE REFERENCE ELEMENT 59

Note that for the Courant element or the Argyris element, and in general for all finite elements
in which V̂ is a space of polynomials, the fulfillment of this assumption suggests the definition
of m, i.e., one chooses m := q (hence, m = 2 for the Courant element or m = 4 for the Argyris
element). On the other hand, for the Hsieh-Clough-Tocher element, we have already observed
that m = 2 while q = 3, which shows that the assumption is fulfilled in this case.
Under the previous assumption, we have I − Π̂ ∈ L(H q (Ê), H m (Ê)), which means that there
exists a constant Ĉ3 > 0 such that

kv̂ − Π̂v̂km,Ê ≤ Ĉ3 kv̂kq,Ê ∀v̂ ∈ H q (Ê) . (5.2)

Finally, we formulate our last assumption.

Assumption 5.1.6 The integers k and q satisfy the inequality

q ≤ k + 1.

Note that this assumption is fulfilled for the three elements considered in the Examples above.
Under the previous assumption, recalling inequality (5.2) and the definition (1.6) of Sobolev
norm which implies the property kv̂kq,Ê ≤ kv̂kk+1,Ê , one has for 0 ≤ ` ≤ m

|v̂ − Π̂v̂|`,Ê ≤ Ĉ3 kv̂kk+1,Ê ∀v̂ ∈ H k+1 (Ê) . (5.3)

At this point, we recall the property that Π̂ is a projection operator, i.e., one has Π̂v̂ = v̂
for all v̂ ∈ V̂ . In particular, recalling Assumption 5.1.1, one has Π̂p̂ = p̂ for all polynomials
p̂ ∈ Pk (Ê). This property can be exploited to write the identity

v̂ − Π̂v̂ = (v̂ − p̂) − Π̂(v̂ − p̂)

and to apply the bound (5.3) to the function v̂ − p̂, getting the bound

|v̂ − Π̂v̂|`,Ê ≤ Ĉ3 kv̂ − p̂kk+1,Ê ∀p̂ ∈ Pk (Ê) ,

or, equivalently,
|v̂ − Π̂v̂|`,Ê ≤ Ĉ3 min kv̂ − p̂kk+1,Ê . (5.4)
p̂∈Pk (Ê)

Note that the quantity σk+1 (v̂) := minp̂∈Pk (Ê) kv̂ − p̂kk+1,Ê is equal to 0 if and only if v̂ = p̂
for some p̂ ∈ Pk (Ê), i.e., if and only if v̂ is a polynomial of degree ≤ k. There is another
quantity that has the same property, namely the semi-norm |v̂|k+1,Ê : indeed, it is easily seen
(by considering e.g. the Taylor expansion of v̂) that all the derivatives of v̂ of order k + 1 are
identically 0 if and only if v̂ is a polynomial of degree ≤ k. Thus, it is natural to ask ourselves
whether there is some relation between the two quantities. The following result, known as
Denis-Lions Lemma, answers this question showing that the two quantities are equivalent.

Lemma 5.1.7 There exists a constant ĈDL > 0 (depending only on k) such that

|v̂|k+1,Ê ≤ σk+1 (v̂) ≤ ĈDL |v̂|k+1,Ê ∀v̂ ∈ H k+1 (Ê) .

60 CHAPTER 5. FINITE ELEMENT APPROXIMATION

Note that the first inequality is trivial, since

|v̂|k+1,Ê = |v̂ − p̂|k+1,Ê ≤ kv̂ − p̂kk+1,Ê ∀p̂ ∈ Pk (Ê) .

On the contrary, the second inequality is a deep result, which can be proven by subtle arguments
in Functional Analysis.
Using this inequality in (5.3), we obtain the existence of a constant Ĉ (= Ĉ3 ĈDL ) such that

|v̂ − Π̂v̂|`,Ê ≤ Ĉ|v̂|k+1,Ê ∀v̂ ∈ H k+1 (Ê) . (5.5)

Finally, we extend this inequality by observing that if we just assume v̂ ∈ H r (Ê) for some
integer r satisfying q ≤ r ≤ k + 1, then one has the inclusion

Pr−1 (Ê) ⊆ Pk (Ê) ⊆ V̂ ;

hence, all the deductions above can be repeated with k replaced by r − 1, yielding the inequality

|v̂ − Π̂v̂|`,Ê ≤ Ĉ|v̂|r,Ê ∀v̂ ∈ H r (Ê) .

Let us summarize our current results as follows.

Proposition 5.1.8 Under Assumptions (5.1.1)-(5.1.3) and (5.1.5)-(5.1.6) (which imply the in-
equalities m ≤ q ≤ k + 1), let the integers ` and r satisfy the inequalities

0≤`≤m and q ≤ r ≤ k + 1.

Then, there exists a constant Ĉ > 0 (depending on m and k) such that the following inequality
holds:
|v̂ − Π̂v̂|`,Ê ≤ Ĉ|v̂|r,Ê ∀v̂ ∈ H r (Ê) . (5.6)

5.2 Semi-norm transformations

In this section, we analyze how a Sobolev semi-norm is transformed under the mapping x ↔ x̂
between the generic element E and the reference element Ê, introduced in Section 4.3.1. From
now on, we assume that the transformation is affine, i.e.,

x = F(x̂) = Bx̂ + b , x̂ = F−1 (x) = B−1 (x − b) (5.7)

for a suitable invertible matrix B and vector b.

At first, we establish a few technical results, starting with some definitions.
Given an element E of a triangulation T , let us introduce, together with the diameter hE of
the element, the diameter bE of the largest ball contained in E; obviously, it holds 0 < bE ≤ hE ,
hE
or equivalently 1 ≤ < ∞. The following assumption prevents this ratio to blow-up to infinity.
bE

Definition 5.2.1 A family F = {T } of admissible triangulations of Ω is said to be a family of

regular triangulations if there exists a constant γ ≥ 1 such that
hE
≤γ ∀E ∈ T , ∀T ∈ F . (5.8)
bE
5.2. SEMI-NORM TRANSFORMATIONS 61

An equivalent condition for polygonal/polyhedral elements is that the minimum angle of each
element does not degenerate to zero, when varying the element in any trangulation of the family.
We make the following assumption.
Assumption 5.2.2 All the triangulations T considered in the sequel belong to a family F of
regular triangulations of Ω.
We are going to use this assumption in the next properties. Herafter, we denote by c1 , c2 , . . .
positive constants independent of the element E.

Property 5.2.3 The determinants of the Jacobian matrices JF = B and JF−1 = B−1 satisfy
the inequalities
c1 hdE ≤ |det B| ≤ c2 hdE ∀E ∈ T , (5.9)
and
c3 h−d −1
E ≤ |det B | = |det B|
−1
≤ c4 h−d
E ∀E ∈ T . (5.10)

Proof. We apply the change-of-variable rule in integrals

Z Z
g = |det B| ĝ (5.11)
E Ê

to the constant function g = 1, getting

|E|
|E| = |det B||Ê| , i.e., |det B| = .
|Ê|

By definition of hE and bE , there exists a constant λd depending on the dimension (precisely,

λ2 = π4 in 2D, λ3 = π6 in 3D) such that

λd bdE ≤ |E| ≤ hdE .

Using Assumption 5.2.2, we obtain

λd d
h ≤ |E| ≤ hdE ,
γd E
which proves (5.9). Taking the reciprocals in this inequality we get (5.10).

Property 5.2.4 The Euclidean norms of the Jacobian matrices JF = B and JF−1 = B−1
satisfy the inequalities
kBk ≤ c5 hE ∀E ∈ T , (5.12)
and
kB−1 k ≤ c6 h−1
E ∀E ∈ T . (5.13)

Proof. By definition of Euclidean norm of B, we have

kBk = sup kBzk .

z∈RN̂ ,kzk=1

If b̂ = bÊ denotes the diameter of the largest ball contained in the reference element Ê, for any z
with kzk = 1 there exist two points x̂1 and x̂2 on the boundary of this ball, hence in Ê, such that
62 CHAPTER 5. FINITE ELEMENT APPROXIMATION

x̂1 − x̂2 = b̂ z. Their images x1 = F(x̂1 ) and x2 = F(x̂2 ) belong to E, hence, kx1 − x2 k ≤ hE .
On the other hand,

x1 − x2 = F(x̂1 ) − F(x̂2 ) = B(x̂1 − x̂2 ) = b̂ Bz ,

hE
showing that kBzk ≤ . Since z is arbitrary, we arrive at (5.12). Inequality (5.13) can be
b̂
obtained similarly, by reversing the roles of E and Ê and applying again Assumption 5.2.2.
We are ready to establish the main result of this section, namely we are going to learn how to
bound a semi-norm of a function by means of the corresponding semi-norm of the transformed
function under the mapping v ↔ v̂.

Proposition 5.2.5 Let the transformation v ↔ v̂ between functions in E and functions in Ê

be defined as in Sect. 4.3.1. Under Assumption 5.2.2, for any integer ` ≥ 0 there exist positive
constants C[`] and C [`] such that

d/2−`
|v|`,E ≤ C[`] hE |v̂|`,Ê ∀v ∈ H ` (E) , (5.14)

and
`−d/2
|v̂|`,Ê ≤ C [`] hE |v|`,E ∀v̂ ∈ H ` (Ê) . (5.15)

Proof. Let us begin by proving (5.15) for ` = 0; recalling (5.11) and (5.10), we have
Z Z
2
|v̂|0,Ê = 2
v̂ = |det B|−1
v 2 ≤ c4 h−d 2
E |v|0,E ,
Ê E

which gives the result.

In order to prove the inequality for ` = 1, let us first observe that the chain rule gives
∂v̂ X ∂v ∂xj X ∂v
(x̂) = (x) = bji (x) , (5.16)
∂ x̂i ∂xj ∂ x̂i ∂xj
j j

where B = (bij ) and indices i, j run between 1 and d; equivalently, it holds

ˆ = BT ∇v ,
∇v̂ (5.17)

whence
ˆ ≤ kBT k k∇vk = kBk k∇vk .
k∇v̂k
Thus, again by (5.11), (5.10) and now (5.12), we get
Z Z
ˆ −1
2
|v̂|1,Ê = 2
k∇v̂k ≤ |det B| kBk 2
k∇vk2 ≤ c4 c25 h2−d 2
E |v|1,E ,
Ê E

which gives the result.

The inequality for ` > 1 can be proven by the Induction Principle, assuming that it holds for
` − 1. So, we can write
X  ∂v̂ 2 X  ∂v̂ 2
2 [`−1] 2 2(`−1)−d −1 

|v̂|`,Ê ≤  
 ∂ x̂i  ≤ (C ) hE  ∂ x̂i ◦ F 
 . (5.18)
i `−1,Ê i `−1,E
5.3. LOCAL ERROR BOUNDS 63

Using (5.16), we have

 
∂v̂ −1 ∂v̂ X ∂v
◦F (x) = (x̂) = bji (x) ,
∂ x̂i ∂ x̂i ∂xj
j

whence
 1/2
X  ∂v 2
     
 ∂v̂ −1 
X  ∂v  X X
|bji |2 )1/2  |bji |2 )1/2 |v|`,E .

 ∂ x̂i ◦ F  ≤ |bji |  ≤(
      =(
`−1,E ∂xj 
`−1,E
 ∂xj 
`−1,E
j j j j

Thus, using the property |bji | ≤ kBk for all i, j, and (5.2.4), we obtain

X  ∂v̂ 2 X
−1
|bji |2 |v|2`,E ≤ d2 kBk|v|2`,E ≤ c5 d2 h2E |v|2`,E .

 ∂ x̂i ◦ F  ≤
 
i `−1,E i,j

This together with (5.18) yields the desired result. The proof of (5.14) follows the same path,
with the roles of E and Ê exchanged.

5.3 Local error bounds

We have all the ingredients to estimate the error semi-norm |v − ΠE v|`,E on an element E ∈ T .

Proposition 5.3.1 Under Assumptions 5.1.1–5.1.3, 5.1.5–5.1.6 and 5.2.2, let the integers `
and r satisfy the inequalities
0 ≤ ` ≤ m ≤ q ≤ r ≤ k + 1.
Then, there exists a constant C`,r,k > 0 independent of the element E ∈ T such that the following
inequality holds:
|v − ΠE v|`,E ≤ C`,r,k hr−`
E |v|r,E ∀v ∈ H r (E) . (5.19)

Proof. Recalling the identity (4.10), one has (v − ΠE v)ˆ = v̂ − Π̂ v̂. Then, concatenating
(5.14), (5.6) and (5.15), one obtains
d/2−`
|v − ΠE v|`,E ≤ C[`] hE |v̂ − Π̂v̂|`,Ê
d/2−`
≤ Ĉ C[`] hE |v̂|r,Ê
d/2−` r−d/2
≤ C [r] Ĉ C[`] hE hE |v|r,E ,

whence the result with C`,r,k = C [r] Ĉ C[`] .

Examples 5.3.2 For the Courant element (T, P1 (T ), LT ), one has k = 1 and m = q = k+1 = 2.
Thus, only the choice r = 2 is possible, which gives the estimates

kv − ΠT vk0,T ≤ Ch2T |v|2,T , |v − ΠT v|1,T ≤ ChT |v|2,T , ∀v ∈ H 2 (T ) . (5.20)

Note that the choice ` = 2 yields a trivial inequality, since the second derivatives of ΠT v are
identically zero.
64 CHAPTER 5. FINITE ELEMENT APPROXIMATION

For the quadratic Lagrangian element (T, P2 (T ), LT ), one has k = 2, m = q = 2 and k +1 = 3.

Thus, the case of maximal regularity r = 3 gives the estimates, valid for all v ∈ H 3 (T ),

kv − ΠT vk0,T ≤ Ch3T |v|3,T , |v − ΠT v|1,T ≤ Ch2T |v|3,T , |v − ΠT v|2,T ≤ ChT |v|3,T , (5.21)

whereas the choice r = 2 gives estimates similar to (5.20).

For the Hsieh-Clough-Tocher element, one has m = 2 and q = k = 3. In the case of maximal
regularity r = 4, one has the estimates, valid for all v ∈ H 4 (T ),

kv−ΠT vk0,T ≤ Ch4T |v|4,T , |v−ΠT v|1,T ≤ Ch3T |v|4,T , |v−ΠT v|2,T ≤ Ch2T |v|4,T . (5.22)

Finally, for the Argyris element, one has k = 5, m = q = 4 and k = 5. In the case of maximal
regularity r = 6, one has the estimates, valid for all v ∈ H 6 (T ),

kv−ΠT vk0,T ≤ Ch6T |v|6,T , |v−ΠT v|1,T ≤ Ch5T |v|6,T , |v−ΠT v|2,T ≤ Ch4T |v|6,T . (5.23)

Obviously, if the function v has lower regularity r = 4 or r = 5, the exponent of hT in the

previous estimates is reduced accordingly.

5.4 Global error bounds

Using the previous local estimates, we can easily derive bounds for the semi-norms |v − ΠT v|`,Ω
of the global error. Note that ΠT v is well defined if v ∈ H q (Ω), since in this case each local
projection ΠE v|E ∈ VE is well defined for E ∈ T . Furthermore, we have already observed in Sect.
4.5.3 that if the chosen finite element is of class C 0 , then ΠT v belongs to V (T ) = V(T ) ∩ H 1 (Ω).
Similarly, if the finite element is of class C 1 , then ΠT v belongs to V (T ) = V(T ) ∩ H 2 (Ω). In
general, we may assume that there exists an integer m̄ satisfying 0 ≤ m̄ ≤ m such that

ΠT v ∈ H m̄ (Ω) ∀v ∈ H q (Ω) .

Proposition 5.4.1 Under the same assumptions as in Proposition 5.3.1, the following inequal-
ities hold for any ` and r satisfying 0 ≤ ` ≤ m̄ ≤ q ≤ r ≤ k + 1:

|v − ΠT v|`,Ω ≤ C`,r,k hr−` |v|r,Ω ∀v ∈ H r (Ω) , (5.24)

where h is the meshsize defined in (4.1), and the constants C`,r,k are the same as those in (5.19).

Proof. Using (5.19) to estimate the local errors, and recalling that h = maxE∈T hE , we have
X
|v − ΠT v|2`,Ω = |(v − ΠT v)|E |2`,E
E∈T
X
= |v|E − ΠE v|E |2`,E
E∈T
2(r−`)
X
2
≤ C`,r,k hE |v|E |2r,E
E∈T
X
2
≤ C`,r,k h2(r−`) |v|E |2r,E
E∈T
2 2(r−`)
≤ C`,r,k h |v|2r,Ω ,

whence the result

5.4. GLOBAL ERROR BOUNDS 65

Remark 5.4.2 The most accurate estimate is

!1/2
2(r−`)
X
|v − ΠT v|`,Ω ≤ C`,r,k hE |v|E |2r,E ,
E∈T

which highlights that the squared global error is a sum of contributions coming from each
element, where the local measure of smoothness of v is multiplied by a power of the local
meshsize. This ideally suggests that, in order to have balanced contributions to the error, one
should take smaller meshsizes where the derivatives of v are larger, and allow larger meshsizes
where the derivatives of v are smaller.

Recalling that a Sobolev norm is obtained by summing up Sobolev semi-norms, we immedi-

ately obtain the following corollary.
Corollary 5.4.3 Under the same assumptions as in Proposition 5.3.1, for any ` and r satisfying
0 ≤ ` ≤ m̄ ≤ q ≤ r ≤ k + 1, there exists a constant C̄`,r,k > 0 such that the following inequality
holds:
kv − ΠT vk`,Ω ≤ C̄`,r,k hr−` |v|r,Ω ∀v ∈ H r (Ω) , (5.25)
Proof. One has
`
X
kv − ΠT vk2`,Ω = |v − ΠT v|2j,Ω
j=0
`
X
2
≤ Cj,r,k h2(r−j) |v|2r,Ω
j=0
 
`
X
2
=  Cj,r,k h2(`−j)  h2(r−`) |v|2r,Ω .
j=0

Since h ≤ diam Ω for P` any 2possible triangulation of Ω, the quantity in parenthesis can be
2
bounded by C̄`,r,k =: j=0 Cj,r,k (diam Ω)2(`−j) , whence the result.

Examples 5.4.4 For the Courant element, one has m̄ = 1, and the global estimates
kv − ΠT vk0,Ω ≤ Ch2 |v|2,Ω , kv − ΠT vk1,Ω ≤ Ch|v|2,Ω , ∀v ∈ H 2 (Ω) . (5.26)

For the quadratic Lagrangian element, one has again m̄ = 1 and, for the maximal regularity,
the global estimates
kv − ΠT vk0,Ω ≤ Ch3 |v|3,Ω , kv − ΠT vk1,Ω ≤ Ch2 |v|3,Ω , ∀v ∈ H 3 (Ω) . (5.27)

For the Hsieh-Clough-Tocher element, one has m̄ = 1 and, for the maximal regularity, the
global estimates
kv − ΠT vk0,Ω ≤ Ch4 |v|4,Ω , kv − ΠT vk2,Ω ≤ Ch2 |v|4,Ω , ∀v ∈ H 4 (Ω) . (5.28)

Finally, for the Argyris element, one has again m̄ = 1 and, for the maximal regularity, the
global estimates
kv − ΠT vk0,Ω ≤ Ch6 |v|6,Ω , kv − ΠT vk2,Ω ≤ Ch4 |v|6,Ω , ∀v ∈ H 6 (Ω) . (5.29)
66 CHAPTER 5. FINITE ELEMENT APPROXIMATION
Chapter 6

Error Analysis of Finite Element

Discretizations of Elliptic Problems

In this chapter, we apply the approximation results established in Chapter 5 in order to estimate
the discretization errors produced by a Galerkin method.
We will also provide inverse estimates, which prove useful in several applications.

6.1 Error estimates in the energy norm

Let u be the solution of the variational problem [VP ] given in (1.1), i.e.,

u ∈ V : a(u, v) = F (v) ∀v ∈ V . (6.1)

We suppose that this formulation encodes a second-order elliptic boundary-value problem, so

that the assumptions of the Lax-Milgram theorem are satisfied with V ⊆ H 1 (Ω) and kvkV =
kvk1,Ω .
Consider an admissible and regular triangulation T of Ω, and choose a reference finite element
(Ê, V̂ , L̂) of class C 0 , which generates local finite elements (E, VE , LE ) by affine mappings as
described in Sect. 4.3.1. Denoting by h the meshsize of T , consider the finite-dimensional
subspace of V
Vh = V(T ) ∩ V , (6.2)

where V(T ) is defined in (4.12); note that functions in Vh are globally continuous and inherit
the boundary conditions possibly prescribed by the definition of V . Once Vh is defined, we can
introduce the Galerkin solution of the discrete variational problem [VP ]δ given in (2.1) for δ = h,
namely,
uh ∈ Vh : a(uh , vh ) = F (vh ) ∀vh ∈ Vh . (6.3)

The error ku − uh k1,Ω can be estimated by Céa’s Lemma (Theorem 2.4.2). To this end, we
assume that the chosen finite element satisfies Assumptions 5.1.1–5.1.3 and 5.1.5–5.1.6, and we
consider the projection operator Πh := ΠT defined in (5.1), for which the approximation error
is estimated in (5.25). Note that

Πh v ∈ V(T ) ∩ H 1 (Ω)

67
68 CHAPTER 6. FEM DISCRETIZATIONS OF ELLIPTIC PROBLEMS

if v ∈ H r (Ω) with r > 1, since our reference finite element is of class C 0 . We further assume
that Πh v satisfies the same boundary conditions as v, namely it vanishes where v vanishes; this
condition, always satisfied in practice, means that
Πh : H r (Ω) ∩ V → Vh .
Hence, if u ∈ H r (Ω) we can pick vh = Πh u in (2.15) and, using the bound (5.25) with ` = 1, we
get
ku − uh k1,Ω ≤ Ca ku − Πh uk1,Ω ≤ Ca C̄1,r,k hr−1 |u|r,Ω .
Summarizing, we have obtained the following important result.
Proposition 6.1.1 Let u be the solution of (6.1) above, with V ⊆ H 1 (Ω). Let uh ∈ Vh be the
solution of the Galerkin problem (6.3). If the chosen finite element is of class C 0 and satisfies
the other assumptions stated above, there exists a constant C > 0 such that if u ∈ H r (Ω) for
some r satisfying q ≤ r ≤ k + 1, one has the error bound
ku − uh k1,Ω ≤ C hr−1 |u|r,Ω . (6.4)
In particular, in the case of maximal regularity u ∈ H k+1 (Ω), one has
ku − uh k1,Ω ≤ C hk |u|k+1,Ω . (6.5)
For instance, if we use the Courant element (linear Lagrangian element), we have a linear
decay of the error:
ku − uh k1,Ω ≤ C h|u|2,Ω if u ∈ H 2 (Ω) , (6.6)
whereas if we use the quadratic Lagrangian element (T, P2 (T ), LT ), we have a quadratic decay:
ku − uh k1,Ω ≤ C h2 |u|3,Ω if u ∈ H 3 (Ω) . (6.7)
Remark 6.1.2 Going back to the general Theorem 2.4.3, its assumptions are satisfied with the
choice Z = H r (Ω) ∩ V , which is dense in V since H r (Ω) is dense in H 1 (Ω). The function ψZ in
the theorem is ψZ (h) = C̄1,r,k hr−1 from (5.25), and the abstract bound (2.23) is precisely (6.4).
Furthermore, the theorem assures that the Galerkin solution uh converges to u for any solution
u ∈ V of the variational problem [VP ].

Let us now suppose that the variational problem (6.1) corresponds to a fourth-order elliptic
problem, as the Timoshenko plate bending problem considered in Sect. ....; in this case, V ⊆
H 2 (Ω) with kvkV = kvk2,Ω . If we use a reference element of class C 1 , then the operator Πh = ΠT
satisfies
Πh ∈ V(T ) ∩ H 2 (Ω)
if v ∈ H r (Ω) with r > 2. Therefore, the previous considerations apply with the obvious changes,
and one gets the error bound
ku − uh k2,Ω ≤ C hr−2 |u|r,Ω , (6.8)
which holds for any u ∈ H r (Ω) with 2 < q ≤ r ≤ k + 1. In particular, in the case of maximal
regularity u ∈ H k+1 (Ω), one gets
ku − uh k2,Ω ≤ C h2 |u|4,Ω
for the Hsieh-Clough-Tocher element, and
ku − uh k2,Ω ≤ C h4 |u|6,Ω
for the Argyris element.
6.2. ERROR ESTIMATES IN THE L2 -NORM 69

6.2 Error estimates in the L2 -norm

The energy norm, i.e., the norm of the space V for which the variational problem [VP ] is
well-posed, is the natural norm in which to estimate the Galerkin error u − uh . However, in
applications, one may be interested in the behavior of the error in other norms, such as the
L2 -norm or the L∞ -norm (the maximum norm). We begin with the L2 -norm.
Let us first observe that in order to estimate ku − uh k0,Ω , we cannot repeat the path followed
for the energy norm (namely, concatenating Céa’s Lemma with the approximation estimate for
the projection operator ΠT ); indeed, while the latter estimate is available also in the L2 -norm
(see (5.25) with ` = 0), no Céa’s Lemma may hold with V replaced by L2 (Ω), since the bilinear
form a associated with a differential operator will not be continuous with respect to this norm,
weaker than the H 1 -norm or the H 2 -norm. A different strategy is required.
The idea, due to Aubin and Nitsche (indeed, the procedure we are going to describe is known
as the Aubin-Nitsche trick), consists of considering the L2 -norm as a dual norm. Precisely,
it holds
(v, g)Ω
kvk0,Ω = sup ∀v ∈ L2 (Ω) . (6.9)
2
g∈L (Ω), g6=0 kgk 0,Ω

This follows from the double inequalities

(v, v)Ω (v, g)Ω kvk0,Ω kgk0,Ω
kvk0,Ω = ≤ sup ≤ sup = kvk0,Ω .
kvk0,Ω g∈L2 (Ω), g6=0 kgk0,Ω g∈L2 (Ω), g6=0 kgk0,Ω

where the Cauchy-Schwarz inequality has been used. Identity (6.9) means that the L2 -norm
of the function v coincides with the norm of the linear continuous form g 7→ (v, g)Ω on L2 (Ω),
i.e., the norm of this form in the dual space L2 (Ω)0 ; this is coherent with the identification
L2 (Ω)0 = L2 (Ω) via the Riesz Representation Theorem in Hilbert spaces.
We apply (6.9) to v = u − uh , getting
(u − uh , g)Ω
ku − uh k0,Ω = sup . (6.10)
g∈L2 (Ω), g6=0 kgk0,Ω

For any g ∈ L2 (Ω), we solve the auxiliary problem: find w ∈ V such that

a(v, w) = (v, g)Ω ∀v ∈ V . (6.11)

This problem is uniquely solvable, since the adjoint bilinear form aT , defined by aT (w, v) =
a(v, w) for all w, v ∈ V , fulfills the same assumptions of the Lax-Milgram theorem as the form
a does. Let us denote by wg the solution of problem (6.11). Then, we have

(u − uh , g)Ω = a(u − uh , wg ) = a(u − uh , wg − vh ) ≤ kak ku − uh k1,Ω kwg − vh k1,Ω ∀vh ∈ Vh ,

where we have used the Galerkin orthogonality property (2.12). Substituting in (6.10), we obtain
distH 1 (Ω) (wg , Vh )
ku − uh k0,Ω ≤ kak ku − uh k1,Ω sup . (6.12)
g∈L2 (Ω), g6=0 kgk0,Ω

We now give conditions which guarantee that the sup quantity is O(h). To this end, let us
assume that
g ∈ L2 (Ω) =⇒ wg ∈ H 2 (Ω) , with kwg k2,Ω ≤ Creg kgk0,Ω (6.13)
70 CHAPTER 6. FEM DISCRETIZATIONS OF ELLIPTIC PROBLEMS

for some Creg > 0. This is a regularity property of the adjoint problem (6.11), which holds, e.g,
if the coefficients of the elliptic operator are smooth and the boundary ∂Ω is smooth (or the
domain Ω is convex). Furthermore, let us assume that

the projection operator Πh = ΠT is defined in H 2 (Ω) (i.e., q = 2 in Corollary 5.4.3).

(6.14)
Under these assumptions, using (5.25) we have

distH 1 (Ω) (wg , Vh ) ≤ kwg − Πh wg k1,Ω ≤ C̄1,2,k h|v|2,Ω ≤ C̄1,2,k Creg h kgk0,Ω ,

which implies by (6.12)

ku − uh k0,Ω ≤ kak C̄1,2,k Creg h ku − uh k1,Ω .

Concatenating with (6.4), we arrive at the final result.

Proposition 6.2.1 Under the assumptions of Proposition 6.1.1 and the assumptions (6.13)–
(6.14) above, there exists a constant C > 0 such that if u ∈ H r (Ω) with 2 ≤ r ≤ k + 1, one has
the error bound
ku − uh k0,Ω ≤ C hr |u|r,Ω . (6.15)

The result we have obtained means that the Galerkin discretization error u − uh behaves
asymptotically (for h → 0) as the projection error u − Πh u not only in the energy norm, but
also in the norm of L2 (Ω).
For instance, with the Courant element we have a quadratic decay of the error:

ku − uh k0,Ω ≤ C h2 |u|2,Ω if u ∈ H 2 (Ω) , (6.16)

whereas with the quadratic Lagrangian element, we have a cubic decay:

ku − uh k0,Ω ≤ C h3 |u|3,Ω if u ∈ H 3 (Ω) . (6.17)

6.3 Error estimates in the maximum norm

Information on the point-wise error, i.e., the error between the exact solution u and the Galerkin
solution uh at a point in the domain, is important in many applications. Obtaining estimates
of this error is a non-trivial task, since in a Galerkin method equations are integrated over the
whole domain, rather than being enforced pointwise; thus, specific techniques of analysis, often
quite elaborate, have been devised.
Here, we content ourselves with an estimate of the error in the norm of L∞ (Ω), which for
continuous functions, such as u and uh , is nothing but the maximum norm. In other words, we
estimate the maximal pointwise error

ku − uh kL∞ (Ω) = max |u(x) − uh (x)| .

x∈Ω̄

To this end, we require regularity of the exact solution in Sobolev spaces based on L∞ (Ω),
namely for r ≥ 0 we consider the spaces

W r,∞ (Ω) = {v ∈ L∞ (Ω) : Dα v ∈ L∞ (Ω) ∀α ∈ Nd , |α| ≤ r} ,

6.4. INVERSE INEQUALITIES 71

equipped with the norm

kvkr,∞,Ω = max |v|m,∞,Ω , where |v|m,∞,Ω = max kDα vkL∞ (Ω) .

0≤m≤r |α|=m

Denoting as usual by k the polynomial degree of the chosen finite element, one can prove the
following error estimate:



h2 |u|2,∞,Ω if k = 1, d = 1 ,

ku − uh kL∞ (Ω) ≤ C h2 | log h| |u|2,∞,Ω if k = 1, d ≥ 2 , (6.18)


 k+1
h |u|k+1,∞,Ω if k ≥ 2, d ≥ 1 ,

provided the indicated semi-norms of u are finite.

Note the presence of the logarithmic factor for linear finite elements in dimension > 1. This
term has very little practical impact, since we know that | log h| tends to ∞ slower than any
power of h as h → 0+ ; hence, for the values of h used in practical computations its presence is
not really felt, and the decay of the error in the L∞ -norm is essentially quadratic as in L2 -norm.

6.4 Inverse inequalities

Let X and Y be infinite-dimensional normed spaces, with norms k · kX and k · kX ; assume that
Y ⊂ X with continuous injection, namely there exists a constant C > 0 such that

kykX ≤ C kykY ∀y ∈ Y .

Now, suppose that Z be a finite-dimensional subspace of Y , thus Z ⊂ Y ⊂ X. In Z, we have

two norms, inherited from X and from Y . An important result states that

in a finite-dimensional normed space, all norms are equivalent ,

i.e., any norm can be bounded by a constant times any other norm. This statement (which is
not true in infinite dimension!) is a simple consequence of Weierstrass’ theorem for continuous
functions in Euclidean spaces. Thus, there exists a constant CZ > 0 depending on Z such that

kzkY ≤ CZ kzkX ∀z ∈ Z . (6.19)

This is an instance of inverse inequality, which expresses the possibility of controlling a stronger
norm (the Y -norm) by a weaker norm (the X-norm), in finite dimension. In general, the constant
CZ depends on the dimension of Z, and blows-up when this dimension tends to infinity. In
applications, one is interested in estimating the size of the constant CZ , highlighting such a
dependence.

Let us discuss the following important example. Let X = L2 (Ω) and Y = H 1 (Ω). Consider
the finite-element subspace
Vh = V(T ) ∩ H 1 (Ω)
introduced in (6.2), under the same assumptions. We want to find a constant CVh satisfying

kvh k1,Ω ≤ CVh kvh k0,Ω ∀vh ∈ Vh .

72 CHAPTER 6. FEM DISCRETIZATIONS OF ELLIPTIC PROBLEMS

The idea is to go from a generic element to the reference element and back, using the equivalence
of norms on the reference element as well as the semi-norm transformations presented in Sect.
5.2. Hereafter, we will set vE = vh|E and we denote by v̂E ∈ V̂ the function on the reference
element associated to vE by the transformation ↔ introduced in Sect. 4.3.1. Using (5.14), one
gets X X X
|vh |21,Ω = |vE |21,E ≤ C[1]
2
hd−2
E |v̂ E |2
1,Ê
≤ C 2
[1] hd−2 2
E kv̂E k1,Ê .
E∈T E∈T E∈T

Since all norms are equivalent on the finite dimensional space V̂ , there exists a constant Ĉ such
that kv̂k1,Ê ≤ Ĉkv̂k0,Ê for all v̂ ∈ V̂ . Hence, using now (5.15), we have
X X
−d
|vh |21,Ω ≤ C[1]
2
Ĉ 2 hd−2 2 2 2 [0] 2
E kv̂E k0,Ê ≤ C[1] Ĉ (C ) hd−2 2
E hE kvE k0,E .
E∈T E∈T

Now, h−1 −1
E can be bounded by hmin , where hmin := minE∈T hE (recall that h = hT = maxE∈T hE ).
Hence,
X
|vh |21,Ω ≤ C[1]
2
Ĉ 2 (C [0] )2 h−2
min kvE k20,E = C[1]
2
Ĉ 2 (C [0] )2 h−2 2
min kvh k0,Ω .
E∈T

Adding kvh k20,Ω to both sides, we obtain the following inverse inequality.
Proposition 6.4.1 There exists a constant C > 0 independent of the triangulation such that

kvh k1,Ω ≤ C h−1

min kvh k0,Ω ∀vh ∈ Vh . (6.20)

We can make the meshsize h, rather than hmin , appear in the inequality, if we assume that
hmin is comparable to h. This is expressed by the following definition.
Definition 6.4.2 A family F = {T } of admissible triangulations of Ω is said to be a family of
quasi-uniform triangulations if there exists a constant τ ≥ 1 such that
h
≤τ ∀T ∈ F . (6.21)
hmin
1
In a quasi-uniform triangulation, one has hmin ≥ τ h, which rules out the possibility of having
elements in T exceedingly smaller than others.

Corollary 6.4.3 If T belongs to a quasi-uniform family of triangulations, one can find a con-
stant C > 0 independent of the triangulation such that

kvh k1,Ω ≤ C h−1 kvh k0,Ω ∀vh ∈ Vh . (6.22)

The inverse inequalities we have just established are sharp, in the sense that they would not
hold with an exponent of h larger than −1. Indeed, consider a uniform mesh on which the
Courant elements are used. If ϕj denotes as usual the Lagrange basis function associated with
a node xj , it easily checked that

kϕj k1,Ω ' hd/2−1 , kϕj k0,Ω ' hd/2 .

Finally, we mention that inverse inequalities in other norms can be obtained with the same
technique described above. Here is an example, which extends (6.22).
6.4. INVERSE INEQUALITIES 73

Proposition 6.4.4 Assume that the chosen reference finite element is of class C 1 , and let Vh =
V(T ) ∩ H 2 (Ω). Then, there exists a constant C > 0 independent of the triangulation such that

kvh k2,Ω ≤ C h−1

min kvh k1,Ω and kvh k2,Ω ≤ C h−2
min kvh k0,Ω , ∀vh ∈ Vh . (6.23)

We close this section by presenting two applications of the inverse inequalities, where we
provide estimates of the eigenvalue of certain matrices.

Application I: estimate of the condition number of the stiffness matrix

Let V ⊆ H 1 (Ω) and let a : V × V → R be a symmetric, continuous and coercive bilinear form
in V . Consider a triangulation T of Ω, belonging to a family of regular and quasi-uniform
triangulations, and let Vh = V(T ) ∩ V be a subspace of finite-element functions on T ; each
finite element is the affine image of a reference finite element (Ê, V̂ , L̂), which is of class C 0 and
Lagrangian, i.e., the degrees of freedom are the values of the functions in V̂ at a set of nodes x̂i ,
1 ≤ i ≤ N̂ , in Ê. The Courant element is the simplest example of such element.
Let {ϕj : 1 ≤ j ≤ Nh } be the associated Lagrangian basis in Vh , satisfying ϕj (xk ) = δjk for
the set of nodes xk in Ω̄ which in each element are images of the nodes in Ê. The associated
stiffness matrix
A = ajk = a(ϕk , ϕj ) 1≤j,k≤N ∈ RNh ×Nh ,


h

introduced in (2.7), appears in the linear system that expresses a Galerkin method in algebraic
terms. We aim at deriving an upper bound for the condition number cond2 (A) of A in the
Euclidean norm.
To this end, we need an auxiliary result. Recall that each function vh ∈ Vh is uniquely
identified by its degrees of freedom, which are the values vj = vh (xj ) at the nodes xj , since
vh = N Nh and let kvk denote its Euclidean norm.
P h
j=1 vj ϕj ; let v = (vj ) ∈ R

Property 6.4.5 There exist constants c1 , c2 > 0 independent of the triangulation T , such that

c1 hd/2 kvk ≤ kvh k0,Ω ≤ c2 hd/2 kvk ∀vh ∈ Vh . (6.24)

Proof. Let us denote again by vE ∈ VE the restriction of vh to the element E ∈ T , and by

v̂E ∈ V̂ the image of vE under the trasformation ↔; let vE denote the vector whose entries are
the values of the degrees of freedom of vh in E, namely the values vh (xj ) at the nodes in E.
One has X
kvh k20,Ω = kvE k20,E
E∈T

with
(C [0] )−2 hdE kv̂E k20,E ≤ kvE k20,E ≤ C[0]
2
hdE kv̂E k20,E
by Proposition 5.2.5 with ` = 0. Since the triangulation is quasi-uniform by assumption, we
have τ −1 h ≤ hE ≤ h for any E ∈ T (recall (6.21)). Hence,

(C [0] )−2 τ −2 hd kv̂E k20,E ≤ kvE k20,E ≤ C[0]

2
hd kv̂E k20,E .

Now observe that v̂E = vE , since by definition of image of the reference finite element, the
values of the degrees of freedom of a function v ∈ VE are the same as the values of the degrees
of freedom of the image v̂ ∈ V̂ (recall (4.7)). Furthermore, observe that kv̂k is a norm in V̂ ,
74 CHAPTER 6. FEM DISCRETIZATIONS OF ELLIPTIC PROBLEMS

since v̂ = 0 implies v̂ = 0 by the property of unisolvence of the set L̂ of degrees of freedom.

Consequently, by the equivalence of norms in finite dimension, there exist constants ĉ, Ĉ > 0
independent of E such that
ĉkvE k ≤ kv̂E k0,E ≤ ĈkvE k .
Combining the previous inequalities, we get
X X
(C [0] )−2 τ −2 ĉ2 hd kvE k2 ≤ kvh k20,Ω ≤ C[0]
2
Ĉ 2 hd kvE k2 .
E∈T E∈T

We conclude by observing that by the regularity of the triangulation, there exists a constant
C? > 0 independent of the triangulation, such that
X
kvk2 ≤ kvE k2 ≤ C? kvk2 ;
E∈T

indeed, each entry of v (say, vh (xj ) for some j) is also an entry of one or more vectors vE (those
corresponding to the elements E containing xj ), but the number of such vectors is uniformly
bounded, since the number of elements which have non-empty intersections with a given element
is bounded by the constant γ appearing in (5.8).

In order to estimate the Euclidean condition number of A, consider the eigenvalues µ of A,

which are real and strictly positive since the matrix is symmetric and positive-definite. Denoting
by µmin and µmax the smallest and the largest eigenvalue, one has
µmax
cond2 (A) = ; (6.25)
µmin
hence, we look for upper and lower bounds on the eigenvalues of A. By definition of eigenvalue
µ, there exists an eigenvector w 6= 0 such that

Aw = µ w .

Multiplying by wT , we obtain
wT Aw = µ wT w .
PNh
We have already observed that wT Aw = a(wh , wh ), where wh = j=1 wj ϕj (see (2.9)). On the
other hand, wT w = kwk2 ; hence, µ satisfies

a(wh , wh ) = µ kwk2 .

Let us denote by α and kak the coercivity and continuity constants of the form a. By Property
6.4.5, we have

αkwh k20,Ω ≤ αkwh k21,Ω ≤ a(wh , wh ) = µ kwk2 ≤ µ c−2 −d 2

1 h kwh k0,Ω ;

simplifying the common factor kwh k20,Ω at the endpoints, we obtain the lower bound

α c21 hd ≤ µ . (6.26)

On the other hand, using the inverse inequality (6.22), we have

µ kwk2 = a(wh , wh ) ≤ kak kwh k21,Ω ≤ kak C 2 h−2 kwh k20,Ω ≤ kak C 2 h−2 c22 hd kwk2 ,
6.4. INVERSE INEQUALITIES 75

which implies the upper bound

µ ≤ kak C 2 h−2 c22 hd . (6.27)

Combining (6.25), (6.26) and (6.27), we get

2
kak C 2 h−2 c22 hd

kak Cc2
cond2 (A) ≤ 2 d
= h−2 .
α c1 h α c1

Summarizing, we have obtained the following result.

Proposition 6.4.6 Under the previous assumptions, the condition number of the stiffness ma-
trix A satisfies the bound
cond2 (A) ≤ C h−2 , (6.28)

where C > 0 is a constant independent of the triangulation.

Using additional arguments, it is possible to prove that the bound is sharp, namely, cond2 (A)
does grow proportionally to h−2 as h → 0. This shows that the stiffness matrix associated with
a Lagrangian basis becomes more and more ill-conditioned as the mesh is refined.

Remark 6.4.7 If the stiffness matrix A corresponds to the discretization of a fourth order
problem, i.e., if the bilinear form a is coercive and continuous with respect to the H 2 (Ω) norm
and the Hsieh-Clough-Tocher element or the Argyris element is used to generate the subspace
Vh , then (6.28) is replaced by
cond2 (A) ≤ C h−4 . (6.29)

Application II: estimates of the approximate eigenvalues of an elliptic operator

Let H and V separable Hilbert spaces, such that V ⊂ H with compact injection (i.e., if {vi }
is a sequence bounded in V , one can extract a subsequence {vik } which converges to some v in
H). This occurs, e.g., if H = L2 (Ω) and V = H01 (Ω).
Let a : V × V → R be a symmetric, continuous and coercive bilinear form in V ; similarly, let
b : H × H → R be a symmetric, continuous and coercive bilinear form in H.
Consider the spectral problem:
(
F ind λ ∈ R and w ∈ V, w 6= 0, such that
(6.30)
a(w, v) = λ b(w, v) ∀v ∈ V .

One can prove by the Spectral Theorem for compact self-adjoint operators, that the problem
admits a non-decreasing sequence of strictly positive eigenvalues λn and corresponding eigen-
functions wn , satisfying

• 0 < λ1 ≤ · · · ≤ λn ≤ λn+1 ≤ . . . , with λn → +∞ as n → ∞;

• a(wn , wm ) = b(wn , wm ) = 0 if n 6= m.
76 CHAPTER 6. FEM DISCRETIZATIONS OF ELLIPTIC PROBLEMS

Example 6.4.8 An important example is the spectral problem for the Laplace operator with
Dirichlet boundary conditions: (
−∆w = λw in Ω ,
(6.31)
w=0 on ∂Ω ,
which describes the free vibrations of a homogeneous, elastic thin membrane. The problem can
be formulated variationally as

w ∈ H01 (Ω) : (∇w, ∇v)Ω = λ (w, v)Ω ∀v ∈ H01 (Ω) .

Problem (6.30) can be discretized by the Galerkin method. Let Vh a finite dimensional
subspace of V , whose dimension is Nh . Consider the problem
(
F ind λh ∈ R and wh ∈ Vh , wh 6= 0, such that
(6.32)
a(wh , vh ) = λh b(wh , vh ) ∀vh ∈ Vh .

Introducing a basis in Vh and denoting by A the stiffness matrix associated with the bilinear
form a, and by B the stiffness matrix associated with the bilinear form b, the discrete spectral
problem can be formulated at the algebraic level as

Aw = λh Bw , (6.33)

where w is the vector collecting the coefficients of the representation of wh in the chosen basis.
This is a generalized eigenvalue problem, which in Matlab can be solved by the command
eig(A, B).
The discrete eigenvalues and eigenfunctions satisfy properties similar to those of the exct
problem, namely:

• 0 < λh,1 ≤ · · · ≤ λh,n ≤ λh,n+1 ≤ λh,Nh ;

• a(wh,n , wh,m ) = b(wh,n , wh,m ) = 0 if n 6= m.

We are interested in deriving lower and upper bounds for the discrete eigenvalues.To this end,
let us assume that the following inverse inequality holds:

kvh kV ≤ C h−1 kvh kH ∀vh ∈ Vh . (6.34)

This is precisely (6.22) if H is endowed by the L2 (Ω)-norm and V is endowed by the H 1 (Ω)-
norm. Let α, kak denote the coercivity and continuity constants of the form a with respect to
the norm in V ; similarly, let β, kbk denote the coercivity and continuity constants of the form b
with respect to the norm in H. Furthermore, let c > 0 the constant of the continuity inequality
kvkH ≤ c kvkV for all v ∈ V .
Picking vh = wh in (6.32), we get

a(wh , wh ) = λh b(wh , wh ) .

Then, on the one hand,

α
kwh k2H ≤ αkwh k2V ≤ a(wh , wh ) = λh b(wh , wh ) ≤ λh kbk kwh k2H ,
c2
6.4. INVERSE INEQUALITIES 77

which – after cancellation of the term kwh k2H 6= 0 at the endpoints, yield
α
≤ λh .
c2 kbk

On the other hand, using (6.34), we have

λh βkwh k2H ≤ λh b(wh , wh ) = a(wh , wh ) ≤ kak kwh k2V ≤ kak C h−2 kwh k2H ,

which, after cancellation, yields

C kak −2
λh ≤ h .
β
Summarizing, we have obtained the following result.

Proposition 6.4.9 If the inverse inequality (6.34) holds, then any eigenvalue of the discrete
spectral problem (6.32) satisfies the bounds

α C kak −2
≤ λh ≤ h . (6.35)
c2 kbk β

For this application, too, a more sophisticated analysis shows that the two bounds are indeed
sharp. Precisely, there exists constants c? , C? > 0 independent of h such that

λh,min = λh,1 ∼ c? λh,max = λh,Nh ∼ C? h−2 as h → 0 . (6.36)

Remark 6.4.10 One can prove that each eigenvalue λh,n converges to the corresponding exact
eigenvalue λn , at a rate proportional to the square of the distance of the eigenfunction wn from
the subspace Vh , i.e.,
|λn − λh,n | ≤ C dist2V (wn , Vh ) .
For instance, if the Courant element is used and the eigenfunctions belong to H 2 (Ω) (which
occurs e.g. if Ω is convex), then the convergence of the eigenvalues is quadratic in h. However,
the constant depends on n, and gets larger and larger as n increases; in other words, on the
same mesh smaller eigenvalues are approximated better than larger eigenvalues.
78 CHAPTER 6. FEM DISCRETIZATIONS OF ELLIPTIC PROBLEMS
Chapter 7

Discretizations
of Parabolic Problems

bla bla bla

7.1

79