Fixed Point Methods for Nonlinear PDEs

Andrea Mondino

HT 2021

1 Fixed Point Theorems 8

1.1 Banach’s FPT and applications to ODEs . . . . . . . . . . . . . . . . . . . 8
1.1.1 Banach’s Fixed point theorem (= CMT) . . . . . . . . . . . . . . . 8
1.1.2 Application to Initial value problems . . . . . . . . . . . . . . . . . 10
1.2 Brouwer’s FPT and Calculus of Variations . . . . . . . . . . . . . . . . . . 13
1.2.1 The Fixed point Theorem of Brouwer . . . . . . . . . . . . . . . . . 13
1.2.2 Calculus of Variations . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.2.3 Null-Lagrangians . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.2.4 Proof of Brouwer’s FPT and the retraction principle . . . . . . . . 18
1.3 Schauder’s Fixed Point Theorem . . . . . . . . . . . . . . . . . . . . . . . 22
1.3.1 Three versions of the theorem and their proof . . . . . . . . . . . . 22
1.4 Application: Peano’s Existence Theorem for ODEs . . . . . . . . . . . . . 25

2 Applications to semilinear PDEs 27

2.1 Some results for linear PDEs and Sobolev spaces . . . . . . . . . . . . . . . 27
2.2 Application I . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.3 Application II . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.4 A Glimpse into positive operators : the maximum principle . . . . . . . . . 34

3 Variational inequalities and monotone operators 38

3.1 Differentiation in Banach spaces . . . . . . . . . . . . . . . . . . . . . . . . 38
3.2 Monotone Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.3 Variational inequalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42

1
 3.4 Proof of Theorem 3.5 by the Galerkin Method . . . . . . . . . . . . . . . . 44

4 Applications to quasilinear PDEs and variational inequalities 48

4.1 Some Quasilinear problems in W01,p (Ω) . . . . . . . . . . . . . . . . . . . . 48
4.2 A remark on the growth bound . . . . . . . . . . . . . . . . . . . . . . . . 51
4.3 A (simple) variational inequality: an elastic beam . . . . . . . . . . . . . . 53
4.4 The Stationary Navier–Stokes equations . . . . . . . . . . . . . . . . . . . 56
4.5 Appendix: Convex hulls of precompact sets are precompact . . . . . . . . . 60
4.6 Appendix: Characterisation of compact sets in C and Lp . . . . . . . . . . 61
4.7 Appendix: Positive and negative parts of W 1,p (Ω) functions are in W 1,p (Ω) 61

2
Introduction

The goal of this part of the course is to introduce some basic methods to establish existence
of solutions to nonlinear equations in infinite-dimensional spaces, such as nonlinear partial
differential equations and variational inequalities. In the first part we introduce and prove
the major fixed point theorems by Picard, Brouwer and Schauder. In the second part we
apply them to solve some nonlinear ordinary and partial differential equations. In the third
chapter we then use the fixed point theorems to prove an abstract result on the existence of
solutions to variational inequalities. This result applies for example to equations given by
monotone operators, which often appear in connection with Euler-Lagrange equations of
convex functionals, but can also arise in a nonvariational context. In the final chapter we
apply the result on variational inequalities to quasilinear second order partial differential
equations and obstacle problems. Throughout this lecture we will restrict ourselves to
elliptic partial differential equations, but the methods can be extended without too much
additional effort to parabolic equations.

These lecture notes are based on Yves Capdeboscq’s and Melanie Rupflin’s lecture notes.
Recommended Literature
H.-W. Alt, Nonlinear Functional Analysis, Lecture course given at the University of Bonn,
1990.
L.C. Evans, Partial Differential Equations, Graduate Studies in Mathematics 19, AMS,
1998.
O. Kavian, Introduction à la théorie des points critiques et applications aux problèmes
elliptiques, Springer Paris, 1994.
H. Le Dret, Équations aux dérivées partielles elliptiques, Lecture course given at the Uni-
versity Pierre et Marie Curie (Paris VI), 2010.

Further Reading
M. Ruzicka, Nichtlineare Funktionalanalysis, Eine Einführung, Springer Berlin, 2004.
M.S. Berger, Nonlinearity and Functional Analysis, Academic Press, 1977.

3
K. Deimling, Nonlinear Functional Analysis, Springer-Verlag, 1985.
E. Zeidler, Nonlinear functional analysis and its applications I, Fixed Point theorems,
Springer New York, 1986
E. Zeidler, Nonlinear functional analysis and its applications II A+B, Monotone operators,
Springer New York, 1990
L. Nirenberg, Topics in Nonlinear Functional Analysis, Courant Institute Lecture Notes,
AMS, 2001.
R.E. Showalter, Monotone operators in Banach spaces and nonlinear partial differential
equations, Mathematical Surveys and Monographs, vol. 49, AMS, 1997.

Before embarking on the theory we start out with some typical examples.

Examples.

(1) Nonlinear ODE

We look for a function y : [a, b] ⊂ R → RN , (or y : [a, b] ⊂ R → X, X Banach space)
such that for some t0 ∈ (a, b)

y 0 (t) = f (t, y(t)), y(t0 ) = y0 .

This is equivalent to finding a solution of the integral equation

Z t
y(t) = y0 + f (s, y(s)) ds.
t0

This is again equivalent to finding a zero of the map F defined via

Z t
(F y)(t) := y0 − y(t) + f (s, y(s)) ds
t0

or a fixed point of the map T defined via T := F + Id, that is

Z t
(T y)(t) := y0 + f (s, y(s)) ds .
t0

(2) Nonlinear PDE

(a) Semilinear elliptic equations, e.g.

Look for a function u : Ω ⊂ Rn → Rm that solves
−∆u = f (u) in Ω
u = u0 on ∂Ω ,

4
 where f : Rm → Rm is a – typically nonlinear – function. Equivalently look for
a fixed point of
T u := (−∆u0 )−1 (f (u)).
The subscript u0 is intended to remind the reader of the boundary condition.
(b) Quasilinear elliptic equations, e.g. p-Laplacian.
Given a continuous function f : R → R, we look for a function u : Ω ⊂ Rn → R
such that 1
−div(|∇u|p−2 ∇u) = f (u) in Ω
u = u0 on ∂Ω .
(c) Stationary Navier–Stokes equation.
For Ω ⊂ R3 find u : Ω ⊂ R3 → R3 (the velocity field) and p : Ω → R (the
pressure) such that

−ν∆u + (u · ∇)u = −∇p + f in Ω

divu = 0 in Ω .
u = 0 on ∂Ω.

Here ν > 0 is the viscosity, f is an outer force (e.g. gravity) and ((u · ∇)u)j =
P3
i=1 ui ∂i uj is a convective term describing the transport of fluid particles with
the flow.

(3) Variational problems

(i) First example. We want to find the shortest curve between two points (0, a)
and (1, b) in R2 . In a graph formulation this corresponds to finding a function
u : [0, 1] → R, u(0) = a, u(1) = b which minimises the length functional
Z 1q
L(v) := 1 + |v 0 (x)|2 dx
0

under all curves with the same boundary conditions, that is, under all v ∈ M :=
{v ∈ C 1 ([0, 1]) | v(0) = a, v(1) = b}. If u is a minimiser of L on M , then it
satisfies L(u) 6 L(v) for all v ∈ M .
We are going to derive the corresponding Euler-Lagrange equations. Let u ∈ M
and φ ∈ M0 := {φ ∈ C 1 ([0, 1]) | φ(0) = 0, φ(1) = 0}. Then u + εφ ∈ M for all
ε ∈ R.
1 ∂
We use the notation ∇u = (∂1 u, . . . , ∂n u), where ∂j = ∂xj = ∂x j
, for the vector of partial derivatives
n m
of u. If u : R → R is given by the components u1 , · · · , um , then ∇u denotes the matrix (∂j ui ), i =
1, · · · m, j = 1, · · · n. Another notation is Du.

5
Define I (ε) := L(u + εφ). If u is a minimiser of L then I 0 (0) = 0.
We now compute I 0 (0):
1
d  u0 (x)φ0 (x)
Z
I (0) =
0
 I (ε) = q dx.
dε ε=0 0 0 2
1 + |u (x)|

In an abstract form we can write the conclusion I 0 (0) = 0 as

hA(u), φi = 0 for all φ ∈ M,

where we consider A to be the operator that assigns to each u ∈ M the linear

map A(u) on M0 that is defined by
Z 1
u0 (x)φ0 (x)
hA(u), φi = q dx.
0 0 2
1 + |u (x)|

Hence we are looking — in a weak form — for a zero of the operator A in the
set M .
If we know that u ∈ C 2 ([0, 1]) we can integrate by parts to obtain
 0
Z 1 0
u (x)
0= − q  φ(x) dx for all φ ∈ M0
0 0 2
1 + |u (x)|

which gives the corresponding Euler–Lagrange equation to L

 0
0
u
− q  = 0 in (0, 1), u(0) = a, u(1) = b.
0 2
1 + |u |

Remarks.
(a) Easy to solve: integrate and square to obtain that u0 ≡ const. Hence u
describes the line segment connecting (0, a) and (1, b).
(b) In general not clear that solution of Euler–Lagrange equation is also min-
imiser. Here it is, since L is convex.
(c) We can do the same for surfaces that are described as graphs of functions
v : Ω ⊂ R2 → R. Set
Z q
I(v) = Area(graph(v)) = 1 + |∇v|2 dx, v = v0 on ∂Ω.
Ω

6
 We obtain
∇u · ∇φ
Z
I (0) =
0
q dx
Ω 2
1 + |∇u|
and the Euler–Lagrange equation:
 
∇u
−div  q  = 0.
2
1 + |∇u|

For n > 1 this is in general not easy to solve!

(ii) Now we consider the corresponding obstacle problem: Find a curve of shortest
length which connects (0, a) and (1, b) and lies above an obstacle given by h ∈
C 1 ([0, 1]).

b
u

a h

0 1

Figure 1: An obstacle problem

Thus we want to find minimiser u of I(v) on

M := {v ∈ C 1 ([0, 1]) | v(0) = a, v(1) = b, v(x) > h(x) ∀x ∈ [0, 1]}.

To make sure that M is non-empty we require that h(0) 6 a and h(1) 6 b. M

is a convex set. Hence with u, v ∈ M we also have u + ε(v − u) ∈ M for all
ε ∈ [0, 1]. If u is minimiser then I(u) 6 I(u + ε(v − u)) for all ε ∈ [0, 1] and
v ∈ M . Hence I 0 (0) > 0 which gives the variational inequality hA(u), v−ui > 0
for all v ∈ M with A as above.

Remark. The above examples are given without specifying properties of the solutions, e.g.
continuity or differentiability. In fact, one of the main tasks to apply Fixed Point Theorems
and related results is to identify a suitable subset of an appropriate function space in which
we are looking for a solution.

7
Chapter 1

Fixed Point Theorems

This section will discuss three fixed point theorems: the Contraction Mapping Theorem,
Brouwer’s Theorem and Schauder’s Theorem.

Definition 1. Let (X, d) be a metric space and T : M ⊂ X → X be a map. A solution of

T x = x is called a fixed point of T .

We will see several fixed point theorems with different assumptions on the space X and
the map T respectively.

Example. Find a zero of a map F : M ⊂ X → X. This problem can be formulated in

different ways as a fixed-point problem, e.g.

(i) T x : = x − F (x)

(ii) T x : = x − wF (x), w ∈ R (linear relaxation)

−1
(iii) T x : = x − DF (x) F (x) (Newton’s method)

1.1 Banach’s FPT and applications to ODEs

1.1.1 Banach’s Fixed point theorem (= CMT)
Strategy. Find fixed point of map T as limit of the iteration defined by xn+1 = T xn .

Definition 2. Let (X, d) be a metric space and T : M ⊆ X → X.

T is a contraction if there exists k ∈ (0, 1) such that

d(T x, T y) 6 kd(x, y) ∀x, y ∈ M.

8
Warning: ”Contraction” is not used is a uniform way in the literature and the property
asked for in the above definition is called ’strongly contractive’ or ’κ-contractive’ in some
books.

Theorem 1.1 (Contraction Mapping Theorem). Let (X, d) be a complete metric space,
let M ⊆ X be non-empty and closed and T : M ⊂ X → M a contraction. Then

1. the equation T x = x has a unique solution x ∈ M ;

2. the sequence (xn ) defined via xn+1 = T xn converges for every x0 ∈ M to x.

Proof. Note that X is complete, and M ⊂ X is non-empty and closed, therefore (M, d) is
also a complete metric space. So it is sufficient to consider the case M = X.

• Uniqueness. Suppose that there are a, b ∈ E, a 6= b such that T a = a and T b = b.

Then d(T a, T b) = d(a, b), and since T is contractive d(T a, T b) < d(a, b), which is a
contradiction.

• Existence. Given x0 ∈ X, consider the iterated sequence xn+1 = T xn . By assumption,

d(xn+1 , xn ) = d(T xn , T xn−1 ) 6 kd(xn , xn−1 ),

and by induction
d(xn+1 , xn ) 6 k n d(x1 , x0 ).
The previous inequality and the triangle inequality yield, for any m > 0,
m−1
X
d(xn+m , xn ) 6 d(xn+p+1 , xn+p )
p=0
m−1
X
6 k n+p d(x0 , x1 )
p=0
1 − km
= kn d(x0 , x1 )
1−k
kn
6 d(x0 , x1 ).
1−k
Since k < 1, we find d(xn , xn+m ) → 0 as n → ∞. Hence (xn ) is a Cauchy sequence
in X and since X is complete it has a limit x. Next, note that a contractive map is
continuous, therefore

T x = T lim xn = lim T xn = lim xn+1 = x.

n→∞ n→∞ n→∞

9
Remarks.

(i) All assumptions in the Theorem are necessary (see Problem Sheet 0).

(ii) The most difficult part in applications is often to show that T maps M into itself.

(iii) We also have the following error estimates (see Problem Sheet 0):

kn
d(xn , x) 6 d(x1 , x0 ) (a-priori error estimate),
1−k
k
d(xn+1 , x) 6 d(xn+1 , xn ) (a-posteriori error estimate),
1−k
d(xn+1 , x) 6 kd(xn , x) (linear convergence rate).

This result is also called Banach’s Fixed Point Theorem, and Picard Iteration Theorem
(depending on the country).

1.1.2 Application to Initial value problems

Let X be a Banach space endowed with the norm k·kX and let t0 ∈ R and y0 ∈ X be
given. We consider the following problem: for t0 ∈ R and a Banach space X we look for a
function y : [t0 − c, t0 + c] → X such that

y 0 (t) = f (t, y(t)), t ∈ (t0 − c, t0 + c), y(t0 ) = y0 , (C)

where f : R × X → X is continuous and y0 ∈ X. We will consider the equivalent integral

equation problem given by
Z t
y(t) = y0 + f (s, y(s)) ds, (C1)
t0

and view it as a fixed point problem in an appropriate space. In what follows, for any
T > 0 and r0 > 0, we write

Q(T, r0 ) = [t0 − T, t0 + T ] × B̄(y0 , r0 )

= {(t, y) ∈ R × X : |t − t0 | 6 T, ky − y0 kX 6 r0 }

10
Remark. Strictly speaking, we do not know yet how differentiation and integration of
Banach space valued functions are defined. You can in the following assume that X = RN ,
but statement and proof of the following theorem do not change in the general case of a
Banach space X.

Theorem 1.2 (Cauchy-Lipschitz). We assume that for some positive numbers a, b, L, K,

the following assumptions hold:

(a) f : Q(a, b) ⊂ R × X → X is continuous

(b) f is L-Lipschitz-continuous with respect to y in Q(a, b), that is

kf (t, y) − f (t, ỹ)kX 6 L ky − ỹkX ∀(t, y), (t, ỹ) ∈ Q(a, b),

(c) f is bounded by K on Q(a, b), that is,

sup kf (t, y)kX 6 K.

(t,y)∈Q

Then we have

(i) There exists a unique solution y to (C1) in [t0 − c, t0 + c] with c := min a, Kb ;

(ii) The sequence Z t

yn+1 (t) = y0 (t) + f (s, yn (s))ds
t0

converges uniformly on [t0 − c, t0 + c] to y.

Remark. Also called Picard-Lindelöf Theorem. If X is RN (finite dimensional), assumption

(c) follows from (a) by compactness of closed bounded sets in RN +1 . If f is not Lipschitz
continuous locally, the solution may not be unique (see Problem Sheet 1).

Proof. We introduce the space Z := C([t0 − c, t0 + c], X). Z is a Banach space endowed
with the norm
kyk0 := max ky(t)kX .
t∈[t0 −c,t0 +c]

We also introduce
kyk1 := max e−L|t−t0 | ky(t)kX .
t∈[t0 −c,t0 +c]

Then
e−Lc kyk0 6 kyk1 6 kyk0 ∀y ∈ Z

11
and hence k·k0 and k·k1 are equivalent and (Z, k·k1 ) is also a Banach space.
Let M := {y ∈ Z | ky − y0 k0 6 b} and define
T : M ⊂ (Z, k·k1 ) → (Z, k·k1 )
via Z t
(T y)(t) = y0 + f (s, y(s)) ds.
t0
We are going to show that T and M satisfy the conditions of the Contraction Mapping
Theorem.
(1) M is closed. Let (yn ) ⊂ M and yn → y in (Z, k·k1 ). Then due to the equivalence of
the norms yn → y in (Z, k·k0 ). Since (yn ) ⊂ M we have kyn − y0 k0 6 b for all n ∈ N
and passing to the limit n → ∞ in this inequality we find ky − y0 k0 6 b. Hence
y ∈ M and M is closed.

(2) T : M → M . We have for y ∈ M that

Z t

kT y − y0 k0 = max f (s, y(s)) ds
t∈[t0 −c,t0 +c]
t0

0
Z t
6 max kf (s, y(s))k0 ds
t∈[t0 −c,t0 +c] t0
6 cK
6 b.
Hence T y ∈ M .

(3) T is strongly contractive. Using the Lipschitz continuity of f in y we find

Z t
−L|t−t0 |

kT y − T ỹk1 = max e (f (s, y(s)) − f (s, ỹ(s))) ds

t∈[t0 −c,t0 +c] t0 X
Z t
6 max e−L|t−t0 | L ky(s) − ỹ(s)kX ds
t∈[t0 −c,t0 +c] t0

= max e−L|t−t0 |
t∈[t0 −c,t0 +c]
Z t
· L ky(s) − ỹ(s)kX e−L|s−t0 | eL|s−t0 | ds
t0 | {z }
6ky−ỹk1
Z t
6 L ky − ỹk1 max eL|s−t0 | e−L|t−t0 | ds
t∈[t0 −c,t0 +c] t0
1 L|t−t0 |
− 1 e−L|t−t0 |


= L ky − ỹk1 max e
t∈[t0 −c,t0 +c] L
−Lc


6 1−e ky − ỹk1

12
 Hence T is strongly contractive on M in (Z, k·k1 ) with k = 1 − e−Lc .

The Contraction Mapping Theorem implies existence and uniqueness of a solution and
convergence of yn to y in (Z, k·k1 ) and consequently also in (Z, k·k0 ).

Remark (Error estimates). We have

kn
(i) kyn − yk1 6 ky1 − y0 k1 .
1−k
k
(ii) kyn+1 − yk1 6 kyn+1 − yn k1 .
1−k
(iii) kyn+1 − yk1 6 k kyn − yk1 .

These estimates follow from Remark (iii) in 1.1.1.

1.2 Brouwer’s FPT and Calculus of Variations

1.2.1 The Fixed point Theorem of Brouwer
The Contraction Mapping Theorem only required few prerequisites on the space X but
strong conditions on the map T .
For Brouwer’s (and later Leray-Schauder’s) Fixed Point Theorem we only need continuity
of T but strong conditions on the space X. We will also lose uniqueness of a fixed point
and (xn ), defined via xn+1 = T xn does not necessarily converge.

Theorem 1.3 (Brouwer’s Fixed Point Theorem). Let K ⊂ Rn be homeomorphic to

BR (0) ⊂ Rn and let T : K → K be continuous. Then T has a fixed point.

An important special case of this Theorem is

Corollary 1.4. Let BR (0) ⊂ Rn and let T : BR (0) → BR (0) be continuous. Then T has a
fixed point.

Remarks.

(i) The proof is very simple for n = 1: Consider a continuous f : [a, b] → [a, b]. Then f
has a fixed point x.

Proof. Define g(x) := f (x) − x. Then g(a) > 0 and g(b) 6 0. Since g is continuous,
there exists a zero x of g in [a, b].

13
 (ii) Brouwer’s Fixed Point Theorem is false in infinite-dimensional spaces. The reason
behind this is that a closed ball is not compact. A counter-example is provided
Problem Sheet 2, but a Theorem due to Kakutani shows that counter-examples exists
in any infinite-dimensional separable Hilbert space. Note also that in an infinite
dimensional space, a continuous function may well be unbounded on closed and
bounded sets (and this in fact makes the study of nonlinear equations much more
difficult).

The proof of Brouwer’s Fixed Point Theorem in dimensions larger than 1 is not simple.
There are many different ways of proving it and you might have seen a proof already
in another lecture (e.g. Topology). We choose a relatively simple analytical proof which
connects to elements from the Calculus of Variations.

1.2.2 Calculus of Variations

We consider so-called energy functionals of the form
Z
I(v) := L(∇v(x), v(x), x) dx,
Ω

where Ω ⊂ Rn is a smooth domain,

v : Ω → Rm ,
L : Rm×n × Rm × Ω → R is smooth,
and we use the notation L = L(p, z, x) = L(p11 , . . . , pmn , z1 , . . . , zm , x1 , . . . , xn ) where p =
(pij ) ∈ Rm×n , z = (zi ) ∈ Rm and x = (xj ) ∈ Ω.
L is called the Lagrangian.

Example. If p = ∇v then pij = ∂j vi , i = 1, . . . , m, j = 1, . . . , n. We also use the notation

Lpij = ∂pij L, Lzi = ∂zi L.

Let now g : ∂Ω → Rm be a given smooth function. We are looking for the minimiser u of
I in the set
M = {v : Ω → Rm , v smooth, v = g on ∂Ω}.
Let u be a smooth minimiser of I in M (the existence of such a function is in general not
clear). Then u solves the corresponding Euler–Lagrange equations which we will compute
now.
Let φ ∈ C0∞ (Ω)m (where C0∞ (Ω)m = φ : Ω → Rm , φ ∈ C ∞ (Ω)m , suppφ ⊂⊂ Ω ). Then


u + εφ ∈ M for all ε ∈ R. We consider I (ε) := I(u + εφ) and since u is a minimiser, we

14
have I 0 (0) = 0. We compute I 0 (ε):
n X
Z X m
I (ε) =
0
Lpij (∇u + ε∇φ, u + εφ, x)∂j φi dx
Ω j=1 i=1
Z X m
+ Lzi (∇u + ε∇φ, u + εφ, x)φi dx.
Ω i=1

Hence I 0 (0) = 0 implies

n X
Z X m m
Z X
0 = Lpij (∇u, u, x)∂j φi dx + Lzi (∇u, u, x)φi dx
Ω j=1 i=1 Ω i=1
n
Z X m
! m
Z X
X
= − ∂j Lpij (∇u, u, x) φi dx + Lzi (∇u, u, x)φi dx.
Ω j=1 i=1 Ω i=1

Since this equation is true for all φ ∈ C0∞ (Ω)m we find that the minimiser u satisfies the
following system of PDE:
n
X

− ∂j Lpij (∇u, u, x) + Lzi (∇u, u, x) = 0 in Ω for all i = 1, . . . , m
j=1
u = g on ∂Ω.

These are the Euler–Lagrange equations for I.

Example.
1
Z
I(v) = |∇v|2 − F (v)
Ω 2
with v : Ω ⊂ R → R and a differentiable function F : Rm → R. Here
n m

1 1
|P |2 − F (z) = (p11 )2 + · · · + (pmn )2 − F (z).


L(P, z) =
2 2
Hence ∂pij L(P, z) = pij and thus ∂pij L(∇u) = ∂j ui . Furthermore Lzi = −∂zi F so that the
EL equations are given by
X
0=− ∂j (∂j ui ) − ∂zi F (u) for all i = 1, . . . , m
j

or
0 = −∆u − ∇F (u).

15
1.2.3 Null-Lagrangians
Definition 3. L is called a Null-Lagrangian if the corresponding Euler-Lagrange equations
are satisfied by all smooth functions (i.e. all C ∞ -functions).

Example. Suppose u : Ω → Rn . The Lagrangian L(∇u, u, x) = u·∂1 u is a Null-Lagrangian

since n
X

− ∂j ∂pj L(∇u, u, x) + Lz (∇u, u, x) = −∂1 (u) + ∂1 u = 0.
j=1

Proposition 1.5. Let L be a Null-Lagrangian, let u, v : Ω → Rm be two smooth functions

with u = v on ∂Ω. Then I(u) = I(v), i.e. for a Null-Lagrangian I depends only on the
boundary values.

Proof. Let I : [0, 1] → R, I (τ ) = I(τ u + (1 − τ )v). Then

n X
Z X m
I (τ ) =
0
Lpij (τ ∇u + (1 − τ )∇v, τ u + (1 − τ )v, x)(∂j ui − ∂j vi ) dx
Ω j=1 i=1
Z X m
+ Lzi (τ ∇u + (1 − τ )∇v, τ u + (1 − τ )v, x)(ui − vi ) dx
Ω i=1
m Z 
X n
X
= − ∂j Lpij (τ ∇u + (1 − τ )∇v, τ u + (1 − τ )v, x)
i=1 Ω j=1

+Lzi (τ ∇u + (1 − τ )∇v, τ u + (1 − τ )v, x) (ui − vi ) dx

since u − v = 0 on ∂Ω. As τ u + (1 − τ )v is smooth and hence a solution to the EL equation

we find that I 0 (τ ) = 0 and consequently I ≡ const on [0, 1], hence I (0) = I (1) and
thus I(u) = I(v).

We are going to show

Proposition 1.6. The determinant is a Null-Lagrangian.

To prove this we need a few technical Lemmas.

Lemma 1.7. Let g : Ω ⊂ Rn → Rn−1 be a C 2 function and let B (i) be the (n − 1) × (n − 1)

matrix obtained from dg by removing the i-th column (i.e. ∂i g).
Then, we have
Xn
(−1)i ∂i det(B (i) ) = 0. (1)
i=1

16
Proof. We can write
n

X
∂i det(B (i) ) = det(C (i,j) )
j=1
j6=i

where the matrix C (i,j) is obtained from dg by removing the i-th column and by replacing
the j-th column by ∂i ∂j g.
As ∂i ∂j g = ∂j ∂i g the matrices C (i,j) and C (j,i) have the same column vectors, though not
appearing in the same order and one easily checks that

det(C (i,j) ) = (−1)j+i−1 det(C (j,i) ).

Thus
n
X n
X n
X
i i
(i) (i,j)
(−1)2i+j+1 det(C (j,i) ) = −α


α := (−1) ∂i det(B = (−1) det(C )=
i=1 i,j=1 i,j=1
i6=j i6=j

so α = 0 as claimed.

Recall that the inverse of a matrix A can be expressed in terms of det(A) and the cofactor
matrix cof(A) where

(cofA)ij = (−1)i+j det Aij , i, j = 1, . . . , n,

Aij ∈ R(n−1)×(n−1) the matrix obtained from A by deleting the i-th row and the j-th
column.

Lemma 1.8. Let u : Rn → Rn be a C 2 function. Then

n
X
∂j (cof∇u)ij = 0 for i = 1, . . . , n.
j=1

Proof. We have
(cof∇u)ij = (−1)i+j det ∂p ul


p6=j,l6=i

For a given j, applying Lemma 1.7 to the function

g = u1 , . . . , ui−1 , ui+1 , . . . , un

gives the announced result.

17
Proof of Proposition 1.6. Let L(P ) = det P . From Cramer’s rule we know that

(det P )Id = P T cofP. (2)

In particular, for every k

n
X
T
(det P ) = (P cofP )kk = plk (cofP )lk . (3)
l=1

Given 1 6 i, j 6 n we choose k = j and obtain

n
∂ det P X ∂(cofP )lj
= δli (cofP )lj + plj = (cofP )ij . (4)
∂pij l=1
∂pij

We need to show that nj=1 ∂j (Lpij (∇u)) = 0 for all smooth functions. Since Lpij (∇u) =
P

(cof∇u)ij due to (4) the statement follows from the previous Lemma.

Another example for a Null-Lagrangian is

L(P ) = tr(P 2 ) − (tr(P ))2 ,

see Problem Sheet 2.

1.2.4 Proof of Brouwer’s FPT and the retraction principle

The proof of Brouwer’s FPT relies on the retraction principle:

Definition 4. A retraction from A ⊂ Rn to B ⊂ A is a continuous map r : A → B so

that r(x) = x for every x ∈ B.

Theorem 1.9 (Retraction Principle (for balls)). There exists no retraction from B =
B1 (0) ⊂ Rn to ∂B, i.e. there is no continuous map r : B → ∂B so that r|∂B = id.

We prove this result below and first explain how it leads to the proof of Brouwer’s FPT.

Proof of Brouwer’s Fixed Point Theorem. .

It is enough to consider the case that Ω = B := B1 (0) ⊂ Rn for if T : K → K is continuous
and h : B1 (0) → K is a homeomorphism (that is, h is bijective and both h and h−1 are
continuous) then T̃ := h−1 ◦ T ◦ h is a continuous map from B to B and any fixed point
x̃0 of T̃ yields a fixed point x0 = h(x̃0 ) of T .

18
 Tx

x
r(x)

We argue by contradiction:
Assume that T : B → B is so that for all x ∈ B, T x 6= x. We may therefore, for a given x
define the line lx passing through T x and x,

T x + λ(x − T x), λ ∈ R

This line cuts ∂B in two points, given by the two solutions of quadratic equation in λ

λ2 kx − T xk2 + 2λ(T x, x − T x) + kT xk2 − 1 = 0,

Choose the positive root, λ(x), a continuous function of x. Then

r : x → T x + λ(x)(x − T x)

is continuous, and contradicts Theorem 1.9.

Explicitly, introducing Qx = (x − T x)/kx − T xk, we have
p 
r(x) = T x + Qx 2 2
(Qx, T x) + 1 − kT xk − (Qx, T x) .

Remark. Theorem 1.9 (the retraction principle) is in fact equivalent to Brouwer’s Fixed
Point Theorem. See Problem Sheet 2 for the other implication.

Proof of the Retraction Principle (Thm. 1.9. Step 1. We first show that there is no smooth
function r : B → ∂B such that r(x) = x for all x ∈ ∂B.
Assume there exists such a r. Define r̃(x) = x. Then r̃ = r on ∂B and, due to Proposi-
tions 1.5 and 1.6 in 1.2.3,
Z Z Z
det ∇r dx = det ∇r̃ dx = 1 dx = |B| =
6 0. (∗)
B B B

By assumption |r(x)|2 = 1 for all x ∈ B and hence (∇r)T r = 0. Since |r| = 1 it follows
that (∇r)T has eigenvalue 0. But then det ∇r(x) = 0 for all x ∈ B which contradicts (∗).

19
Step 2. We show by approximation that there is no continuous function r : B → ∂B such
that r(x) = x for all x ∈ ∂B.
Assume there exists such a function r and extend it via r(x) = x for all x ∈ Rn \B. Then
r(x) 6= 0 for all x ∈ Rn .
1
Define rε (x) := (φε ∗ r)(x) where φε = ε−n φ( xε ) with φ(x) = Ce |x|2 −1 for |x| < 1 and C is
R
such that Rn φε (x) dx = 1. In particular, φε is radially symmetric.
We know that rε → r locally uniformly on Rn (as rε is a mollification of r see C5.1a Lecture
notes, Section 4.4) and hence rε (x) 6= 0 for all x ∈ Rn for sufficiently small ε > 0.
Let x ∈ ∂B2 (0) and ε be small (smaller that 1/2). Then, due to the properties of φ,
Z
rε (x) = φε (y)(x − y) dy = x.
Bε (0)

Define
rε (2x)
r̃ε (x) = .
|rε (2x)|
Then r̃ε : B → ∂B is smooth and r̃ε (x) = x for all x ∈ ∂B which is a contradiction, thanks
to step 1.

Corollary 1.10. If K ⊂ Rn is homeomorphic to B1 (0) then there exists no retraction from

K to ∂K.

Proof. If h : B1 (0) → K is a homeomorphism (that is, h is bijective and both h and h−1 are
continuous), then any retraction r from K to ∂K would yield a retraction r̃ := h−1 ◦ r ◦ h
of B to ∂B which would then contradict Theorem 1.9.

Proposition 1.11. Let g : Rn → Rn be a continuous vector field such that g(x) · x > 0 for
all x with |x| = R. Then there exists an x0 ∈ Rn with |x0 | 6 R and g(x0 ) = 0.

Proof. Assume that there exists no such x0 . Then we can define

g(x)
f (x) = −R .
|g(x)|
f is continuous and f : BR (0) → BR (0). Brouwer’s FPT implies that there exists x1 ∈
BR (0) such that f (x1 ) = x1 . Then |x1 | = |f (x1 )| = R such that the assumption on g
implies g(x1 ) · x1 > 0.
On the other hand
|g(x1 )| |x1 |2 |g(x1 )|
g(x1 ) · x1 = −f (x1 ) · x1 =− < 0,
R R
which is a contradiction.

20
This application is in fact equivalent to the contraction principle, and therefore to Brouwer’s
Fixed Point Theorem. Indeed, given r a continuous retraction on the closed ball BR (0),
define g by g = r on BR (0) and g(x) = x elsewhere. It is continuous, and when |x| = R
we have g(x) · x = R2 > 0. But g never cancels since |g| > R, which is a contradiction.

Corollary 1.12. Let K ⊂ Rn be convex, compact, and non-empty, and T : K → K be

continuous. Then T has a fixed point.

Proof. Step 1: If K has non-empty interior. We will prove that a convex compact subset
of Rn with non-empty interior is homeomorphic to the closed unit sphere. The conclusion
then follows. We may assume without loss of generality that the ball Br (0) ⊂ K ⊂ BR (0)
for some 0 < r < R < ∞ — true up to a translation (an homeomorphism), since K has
non empty interior. We then define the map (the gauge of K)
n x o
j(x) = inf t > 0 such that ∈ K
t
You will show the following properties on Problem Sheet 3,

(i) The map j : Rn → R is continuous,

(ii) For all λ > 0, j(λx) = λj(x),

(iii) For all x ∈ Rn , kxk/R 6 j(x) 6 kxk/r,

(iv) j(x) 6 1 if and only if x ∈ K.

and show that these properties allow you to define the desired homeomorphism g : K →
B1 (0) and its inverse h as
 
 x j(x) when x 6= 0,  kyk y when y 6= 0,
g(x) = kxk and h(y) = j(y)
0 if x = 0, 0 if y = 0.

Step 2: General case. We can assume again that 0 ∈ K. Either K = {0}, in which case the
result is trivial, or there exists a maximum of x1 , . . . , xm independent vectors in K, with
m 6 n. By convexity, the m-simplex (0, x1 , . . . , xm ), which contains an m dimensional
ball, is contained in K. If m = n, so K has non empty interior. If m < n, C is contained
in a space of dimension m, and we apply the previous step in that space.

We will use this last result to prove Schauder’s Fixed Point Theorem.

21
1.3 Schauder’s Fixed Point Theorem
1.3.1 Three versions of the theorem and their proof
We will introduce Schauder’s Theorem in three different formulation, together with the
closely related Leray-Schauder Theorem.

Theorem 1.13 (Schauder’s FPT, version I). Let X be a Banach space, K ⊂ X be non-
empty, convex and compact and T : K ⊂ X → K be continuous. Then T has a fixed point
in K.

For the proof we use

Lemma 1.14. Let K, T and X be as in Theorem 1.13. Then for any ε > 0 there exist a
finite dimensional subspace Lε of X and a continuous map Tε : K → K ∩ Lε so that

kT x − Tε xk < ε for all x ∈ K.

Proof of Lemma 1.14. Since T is continuous and K is compact, T is uniformly continuous.

Thus, given ε > 0, there exists δ > 0 such that, for all x, y ∈ K we have kT x − T yk 6 ε
provided kx − yk 6 δ. Furthermore, there exists a finite set of points {x1 , . . . , xN } ⊂ K
such that K ⊂ ∪ Bδ (xi ) , where Bδ (xi ) is the open ball of centre xi and radius δ. The
16i6N
vector space Lε = span {T xi , 1 6 i 6 N } is finite dimensional, and therefore K ∩ Lε is
non-empty, compact, convex and finite dimensional.
For i 6 j 6 N , we define the continuous function ψj : X → R by

0 if kx − xj k > δ,
ψj =
1 − 1 kx − xj k otherwise.
δ

We see that ψj is strictly positive on B(xj , δ) and vanishes elsewhere. Therefore we have
PN
j=1 ψi (x) > 0 for all x ∈ K, and this continuous function is therefore bounded below on
K. We can therefore define a partition of unity,
ψi (x)
φi (x) = PN for all 1 6 i 6 N,
j=1 ψi (x)

PN
which satisfy i=1 φi (x) = 1 for all x ∈ K. We now define an approximation of T on K
by
N
X
Tε x = φi (x)T xi for all x ∈ K.
i=1

22
We have, for every x ∈ K,
N
X
Tε x − T x = φi (x) (T xi − T x)
i=1

Whenever φi 6= 0, we have kx − xi k < δ and therefore kT x − T xi k < ε. Thus,

N
X N
X
kTε x − T xk 6 φi (x) kTε x − T xk 6 ε φi (x) = ε.
i=1 i=1

Finally note that the map Tε x is continuous, and takes its values in K ∩ Lε , since K is
convex and T̃ x is a weighted average of T xi , i = 1, . . . , N .

Proof of Theorem 1.13. Take any sequence εn → 0 and let Tεn and Lεn be as in Lemma
1.14. We can then apply Corollary 1.12 to conclude that

Tεn : Lεn ∩ K → Lεn ∩ K

has a fixed point xn . As K is compact a subsequence xnj of the xn converges to a point x∗

which is a fixed point of the original map T since

kT x∗ − x∗ k 6 T x∗ − T xnj + T xnj − Tεnj xnj + xnj − x∗ → 0

thanks to the continuity of T and Lemma 1.14.

Remark. For a set M its convex hull conv(M ) is defined as

n m
X m
X o
conv(M ) := y ∈ X | ∃yi ∈ M, αi ∈ [0, 1], αi = 1 s.t. y = αi yi .
i=1 i=1

We often need an alternative version to this Theorem,

Theorem 1.15 (Schauder’s FPT, version II). Let X be a Banach space, M ⊂ X be non-
empty, convex and closed and T : M ⊂ X → M be a continous operator such that T (M )
is precompact. Then T has a fixed point.

For the proof we need the following lemma whose proof can be found in the appendix.

Lemma 1.16 (Mazur). Let X be a Banach space and M ⊆ X be precompact. Then

conv(M ) is precompact.

23
Proof of Theorem 1.15. The previous Lemma implies that conv(T (M )) is precompact.
Hence conv(T (M )) is compact. Furthermore, it is non-empty and convex. Note that
T (M ) ⊂ M , and since M is convex, conv(T (M )) ⊂ M , and because M is closed conv(T (M )) ⊂
M . Thus, Schauder’s Fixed Point Theorem can be applied to the continuous map T : conv(T (M )) →
conv(T (M )), and T has a fixed point in conv(T (M )).

Remark. In applications to nonlinear problems, this result will be used as follows. First,
the problem is reformulated as a fixed-point problem, for some T . Then we select a space
X where T is continuous, and a closed convex set M such that T : M → M , where either
M is compact, or T (M ) is precompact. To show this last property, it is sufficient to show
that for any sequence xn ∈ M , there exists a sub-sequence T (xnk ) which converges in X
—in particular we need not show that xnk itself converges.
In its second form, Schauder’s Fixed Point Theorem implies a statement about compact
operators. The general definition of compact operators is as follows. Let X, Y be Banach
spaces.

Definition 5. Let T : M ⊆ X → Y , T is a compact operator if

(i) T is continuous;

(ii) T (B) is compact in Y for all bounded subsets B ⊆ M .

Compact operators on bounded sets can be equivalently characterised as being those op-
erators that can be approximated by continuous operators to finite dimensional spaces as
described in Lemma 1.14, namely
Remark (Approximation of compact operators by “finite-dimensional” operators). Let M ⊂
X be bounded and T : M ⊂ X → Y . Then the following are equivalent:

(1) T is a compact operator.

(2) For all n ∈ N there exists a compact operator Tn : M → Y such that Tn (M ) ⊂ Yn

with dim Yn < ∞ and Tn → T uniformly in M .

The proof, more or less contained in the proof of Schauder’s Theorem is left as an (optional)
exercise. In terms of compact operators, Schauder Theorem can be written as follows.

Theorem 1.17 (Schauder’s FPT, version III). Let X be a Banach space, M ⊂ X be non-
empty, convex, bounded and closed and T : M ⊂ X → M be a compact operator. Then T
has a fixed point.

24
To conclude this section, let us mention the following variant of Schauder’s Theorem,
known as Leray-Schauder Fixed Point Theorem, or Schaefer’s Theorem.

Theorem 1.18 (Leray-Schauder/Schaefer Theorem). Let X be a Banach space and T : X →

X be a compact map with the following property: there exists R > 0 such that the state-
ment (x = τ T x with τ ∈ [0, 1)) implies kxkX < R. Then T has a fixed point x∗ such that
kx∗ k 6 R.

Proof. For the proof, see Problem Sheet 3.

Theorem 1.18 says that a sequence of a priori estimates on fixed points of τ T , whether
they exist or not, implies in fact the existence of a fixed point, a surprising result.

Corollary 1.19. Let X be a Banach space and let T : X → X be compact so that there
exist b > 0 and a ∈ [0, 1) so that

kT xk 6 a kxk + b for all x ∈ X

then T has a fixed point.

1.4 Application: Peano’s Existence Theorem for ODEs

Consider again as in Section 1.1.2 the initial value problem

y : [t0 − a, t0 + a] → RN , y 0 (t) = f (t, y(t)), y(t0 ) = y0 .

We assume now that f : [t0 − a, t0 + a] × RN → RN is continuous but not necessarily

Lipschitz-continuous w.r.t. y. We know that under these conditions solutions are in general
not unique. The following theorem, however, shows that solutions always exist.

Theorem 1.20 (Peano). Let (t0 , y0 ) ∈ R × RN , a, b > 0 and

Q := {(t, y) ∈ R × RN | |t − t0 | 6 a, |y − y0 | 6 b}.

Let f : Q → RN be continuous. Let K = maxQ |f (t, y)|. Then there exists a continuous
solution y : [t0 − c, t0 + c] → RN , with c = min(a, b/K) of the integral equation
Z t
y(t) = y0 + f (s, y(s)) ds.
t0

25
Proof. We define as in Section 1.1.2 the map T via
Z t
(T y)(t) = y0 + f (s, y(s)) ds
t0

and with I := [t0 − c, t0 + c]

n o
0 N
M = y ∈ C (I, R ) | kyk = max |y(t)| , ky − y0 k 6 b .
t∈I

Then M is non-empty, convex, closed (and bounded). We have

Z t
kT y − y0 k 6 max | |f (s, y(s))| ds| 6 cK 6 b,
t∈I t0

therefore T (M ) ⊂ M . We show that T is continuous: let yn → y in M . Then

Z t0 +c
max |(T yn )(t) − (T y)(t)| 6 |f (s, yn (s)) − f (s, y(s))| ds → 0
t∈I t0 −c

as n → ∞, since f is uniformly continuous on Q.

We show that T (M ) is pre-compact using Arzéla–Ascoli’s Theorem (see Appendix). Using
the assumptions we find supy∈M |(T y)(t)| 6 |y0 | + aK and

sup |(T y)(t1 ) − (T y)(t2 )| 6 K |t1 − t2 | → 0 as |t1 − t2 | → 0.

y∈M

Hence Schauder’s FPT (version II) implies that T has a fixed point.

26
Chapter 2

Applications to semilinear PDEs

Our goal in this section is to use Schauder’s FPT to prove existence of solutions of equations
of the form
−∆u = f (u) in Ω
u = 0 on ∂Ω.
Remark. For simplicity we restrict ourselves to scalar equations and zero boundary condi-
tions.
In the following we always assume that Ω ⊂ Rn is open, bounded and smooth.

2.1 Some results for linear PDEs and Sobolev spaces

Here we recall some results from the course C4.3: Functional Analytic Methods for PDEs
(and from A.04 Integration) that are needed later on in this section.
Proposition 2.1. Let g ∈ H −1 (Ω) and µ ∈ [0, ∞). There exists a unique weak solution
vg of
−∆v + µv = g in Ω
,
v = 0 on ∂Ω.
i.e. a unique function v = vg ∈ H01 (Ω) that solves the variational problem
Z Z
1
∀ w ∈ H0 (Ω), ∇v · ∇w dx + µ vw dx = hg, wi. (1)
Ω Ω

Furthermore, the map

(−∆ + µId )−1 : H −1 (Ω) → H01 (Ω)
g 7→ vg
is continuous.

27
This is a consequence of Riesz Representation Theorem, and of Poincaré’s inequality.
Notation. Note that by (−∆ + µId )−1 g we always mean the solution of the equation with
zero boundary data.
The version of Poincaré inequality that we will use in these notes is as follows

Theorem 2.2 (Poincaré). For p ∈ [1, ∞), there exists a constant C = C(Ω, p) such that

∀u ∈ W01,p (Ω) kukLp (Ω) 6 Ck∇ukLp (Ω:Rn )

A key tool to obtain the compactness of the fixed point maps we will consider is the
following theorem.

Theorem 2.3 (Rellich-Kondrachov). For p ∈ [1, ∞),

np
• If 1 6 p < n, W 1,p (Ω) ,→ Lq (Ω) for 1 6 q < n−p
,

• If p = n, W 1,p (Ω) ,→ Lq (Ω) for 1 6 q < ∞,

• If p > n, W 1,p (Ω) ,→ C 0,γ (Ω) for 0 6 γ < 1 − np ,

and all these embbedings are compact.

An important consequence for us is

Corollary 2.4. Let µ > 0. Then the map g 7→ (−∆ + µId )−1 g is

• continuous as map from L2 (Ω) to H01 (Ω) in other words,

kvkH01 (Ω) 6 C(Ω)kgkL2 (Ω)

• compact as map from L2 (Ω) to L2 (Ω)

Proof. The first part is due to the fact that L2 (Ω) is continuously embedded in H −1 (Ω).
The second part follows as (−∆ + µId )−1 : L2 (Ω) → L2 (Ω) can be viewed as composi-
tion of the continuous map (−∆ + µId )−1 : L2 (Ω) → H01 (Ω) and the compact embedding
H01 (Ω) ,→ L2 (Ω) and as the composition of a compact linear operator and a continuous
linear operator is again compact.

Remark. As we shall see later also the first embedding is compact, though the proof of this
is more involved.
Finally, let us clarify what f (u) means when u is not a continuous function.

28
Lemma 2.5. For f ∈ C(R), and any two measurable functions u1 and u2 on Ω also f (u1 )
and f (u2 ) are measurable and if furthermore if u1 = u2 almost everywhere in Ω, then
f (u1 ) = f (u2 ) almost everywhere.

Proof. If u1 is measurable, so is f (u1 ) since the preimage of open sets under u is measurable
and since f is continuous so the the preimage of any open set under f is again open. If
u1 = u2 almost everywhere, there exists a set of zero measure N such that if x ∈ Ω \ N
u1 (x) = u2 (x). Then, certainly f (u1 (x)) = f (u2 (x)), therefore f (u1 ) = f (u2 ) almost
everywhere.

Lemma 2.6. Given f ∈ C(R) such that |f (t)| 6 a + b|t|r , where a > 0, b > 0 and r > 0
are positive constants. Then the map u 7→ f (u) is continuous from Lp (Ω) to Lp/r (Ω) for
p > max(1, r) and maps bounded subsets of Lp (Ω) to bounded subsets of Lp/r (Ω)

Proof. Thanks to Jensen’s inequality,

(a + b|t|r )p/r 6 2p/r−1 ap/r + 2p/r−1 bp/r |t|p 6 C(1 + |t|p ),

where C is a positive constant depending on a, b, p and r only. Since u ∈ Lp (Ω), we have

Z  Z 
p/r p
|f (u)| dx 6 C(a, b, p, r) |Ω| + u dx < ∞,
Ω Ω

therefore f (u) ∈ Lp/r (Ω). Let un be a sequence converging to u in Lp (Ω). There exists a
subsequence un0 and a function g ∈ Lp (Ω) such that un0 converges almost everywhere to
u, that is, for all x ∈ Ω \ N where N is a negligible set, un0 (x) → u(x), and |un0 (x)| 6 g(x)
almost everywhere. This is sometimes called the generalized DCT, or the partial converse of
the DCT, or the Riesz-Fisher Theorem. From the continuity of f , |f (u(x))−f (un0 (x))| → 0
on Ω \ N , and
|f (u(x)) − f (un0 (x))|p/r 6 C(1 + g(x)p + |f (u)|p ),
where C is another positive constant depending on a, b, p and r only. The left-hand-side is
independent of n0 , and is in L1 (Ω). We can apply the Dominated Convergence Theorem
R
to conclude that Ω |f (u(x)) − f (un0 (x))|p/r dx → 0, or in other words,

kf (u(x)) − f (un0 (x))kLp/r (Ω) → 0.

Since the limit does not depend on the subsequence, this convergence holds for un .

We now turn to applications.

29
2.2 Application I
We look for a weak solution u : Ω → R of

−∆u = f (u) in Ω
(1)
u = 0 on ∂Ω

under suitable conditions on f : R → R.

Theorem 2.7. Let f ∈ C(R) and supx∈R |f (x)| = a < ∞. Then (1) has a weak solution
u ∈ H01 (Ω), i.e. Z Z
∇u · ∇φ dx = f (u)φ dx ∀φ ∈ C0∞ (Ω).
Ω Ω

Proof. Our strategy is to apply Schauder’s Fixed Point Theorem (version III) to the map

T : L2 (Ω) → L2 (Ω)
u 7→ (−∆)−1 (f (u))

T is continuous. Lemma 2.6 shows that u → f (u) is continuous from L2 (Ω) into itself.
Corollary 2.4 shows that (−∆)−1 is continuous from L2 (Ω) into H01 (Ω), which is continu-
ously embedded in L2 (Ω).
Find a closed, non-empty bounded convex set such that T : M → M . Given u ∈ L2 (Ω),
T u satisfies
Z Z
∇T u · ∇T u dx = f (u)T u dx 6 a|Ω|kT ukL2 (Ω) Cauchy-Schwarz (2)
Ω Ω

Therefore, using Poincaré’s inequality

kT uk2L2 (Ω) 6 C(Ω)k∇T uk2L2 (Ω) 6 a|Ω|C(Ω)kT ukL2 (Ω)

Thus if we set R = a|Ω|C(Ω) and choose

M = {u s.t. kukL2 (Ω) 6 R},

we have established that T : M → M .

T is compact. Using Poincaré’s inequality on the right-hand-side in (2), we obtain

k∇T uk2L2 (Ω) 6 Rk∇T ukL2 (Ω)

Thus T (M ) ⊂ {u s.t. kukH 1 (Ω) 6 R}, and since the embedding of H 1 (Ω) into L2 (Ω) is
compact, T is compact.

The proof could also have been articulated differently, if we had defined T : H01 (Ω) →
H01 (Ω) instead.

30
2.3 Application II
We look for a weak solution u : Ω → R of
)
−∆u + g(u, ∇u) + µu = h in Ω
(1)
u = 0 on ∂Ω
where µ > 0, b ∈ C(R × Rn : R) grows at most linearly at infinity, i.e. there exists
M1 > 0,M2 > 0 such that |g(z, p)| 6 M1 + M2 (|z| + |p|) for all z ∈ R and p ∈ Rn , and
h ∈ L2 (Ω). The following generalization of Lemma 2.6 shows that B : u 7→ b(u, ∇) is
continuous from H01 (Ω) into L2 (Ω).
Lemma 2.8. Let g ∈ C(R × Rn ) be such that |g(z, p)| 6 a + b |z|α + c |p|, where a,b and c
are non negative constants, and 0 < 2α < 2∗ , where 2∗ = 2n/(n − 2) if n > 3, and 2∗ = ∞
if n = 1, 2. Then the map u 7→ g(u, ∇u) is continuous from H01 (Ω) to L2 (Ω) and maps
bounded subsets of H01 (Ω) to bounded subsets of L2 (Ω).
Proof. See Problem Sheet 4.
Theorem 2.9. If M2 = 0, that is, if g is bounded, there exists a weak solution u ∈ H01 (Ω)
of (1), i.e.
Z Z
∇u · ∇φ + g(u, ∇u)φ + µuφ dx = hφ dx ∀φ ∈ C0∞ (Ω).
Ω Ω

If M2 6= 0 the result holds provided µ is larger than a constant depending on Ω and M2

only.
Remark. If g(u, ∇u) is not bounded, there might not be a solution for an arbitrary µ.
Take for example g(u, ∇) = −(λ + µ)u, where λ is a simple eigenvalue of the ∆, with a
R
corresponding eigenvector ψ and choose h such that Ω hψ dx 6= 0. Then the Fredholm
alternative shows that there is no solution, and it can be checked directly: if u is a solution,
then we have
Z Z Z Z
∇u · ∇ψ dx = λuψ + hψ dx and ∇ψ · ∇u dx = λψu dx
Ω Ω Ω Ω
R
which implies Ω
hψ dx = 0.

Proof of Theorem 2.9. We are going to use the Leray-Schauder-Schaefer Theorem, in the
form detailed in Problem Sheet 3, namely: find a continous, compact operator T on X =
H01 (Ω), a Banach space endowed the norm
sZ
kukX = |∇u|2 + |u|2 dx,
Ω

31
which satisfies for all u ∈ X the a priori estimate

kT kX 6 a kukX + b,

for some b ∈ R and 0 6 a < 1 independent of u. This is sufficient to obtain the existence
of a fixed point (at it incidentally gives a bound on its norm). If T is chosen such that a
fixed point corresponds to a weak solution of (1), we have a proof.
Choice for T . As before, place the linear part on the left, and non linear and source part
on the right, and write

T : X = H01 (Ω) → X
u → (−∆ + µId )−1 (−g(u, ∇u) + h)

Clearly, T is well defined and continuous. Lemma 2.8 shows that u 7→ −g(u, ∇u) is
continuous from X to L2 (Ω), and Corollary 2.4 says that v → (−∆ + µId )−1 v is continuous
from L2 (Ω) to X.
A priori estimates Given u ∈ X, we write v = T u. We have, for any φ ∈ H01 (Ω),
Z Z
∇v · ∇φ + µvφ dx = −g(u, ∇u)φ + hφ dx ∀φ ∈ C0∞ (Ω).
Ω Ω

Choosing φ = v, and bounding the right-hand side with Cauchy-Schwarz, we obtain

Z
|∇v|2 + µ |v|2 dx 6 kg(u, ∇u)kL2 (Ω) kvkL2 (Ω) + khkL2 (Ω) kvkL2 (Ω)
Ω
6 (M1 + khkL2 (Ω) ) kvkL2 (Ω) + M2 kukX kvkL2 (Ω) ,

where in the second line we used the bound on g. Thanks to Poincaré inequality, we have
Z
2
C(Ω) kvkX 6 |∇v|2 .
Ω

For some positive constant C(Ω) > 0 depending on Ω only. On the right-hand side, noting
that for all κ, a, b, > 0 there holds ab 6 κ2 a2 + 2κ
1 2
b , we have

1 M22
M2 kukX kvkL2 (Ω) 6 C(Ω) kuk2X + kvk2L2 (Ω) ,
2 2C(Ω)
C(Ω) (M1 + khkL2 (Ω) )2
(M1 + khkL2 (Ω) ) kvkL2 (Ω) 6 kvk2L2 (Ω) +
4 C(Ω)
Since kvkL2 (Ω) 6 kvkX , We have therefore obtained, dividing both sides by C(Ω),

3 2 µ 2 1 2
(M1 + khkL2 (Ω) )2 M22
kvkX + kvkL2 (Ω) 6 kukX + + kvk2L2 (Ω) .
4 C(Ω) 2 C(Ω)2 2C(Ω)2

32
if µ > M22 /(2C(Ω)2 ), this implies,
r
2 (M1 + khkL2 (Ω) )
kvkX 6 kukX +
3 C(Ω)

which is the desired estimate.

Compactness of T In this case, it is not a direct application of Rellich-Kondrachov em-
bedding Theorem: g is only continuous, it has no reason to be compact (it could be, for
example, kukX ). We show below that v 7→ (−∆ + µId )−1 v is in fact compact from L2 (Ω)
into X. With this result, T is the composition of a continuous map for which the images
of bounded subsets are again bounded and a compact map; it is therefore compact, and
the proof is complete.

Lemma 2.10. The map v 7→ (−∆ + µId )−1 v is compact from L2 (Ω) into H01 (Ω).

Proof. Take a bounded sequence vn ∈ L2 (Ω). We can extract a weakly converging sub-
sequence vn0 * v. Let us show that the sequence of solution of weak solutions in H01 (Ω)
of
−∆wn0 + µwn0 = vn0 in Ω
converges strongly in H01 (Ω). Then, we will have shown that from every bounded sequence
in L2 (Ω), we can extract a sequence which converges in H01 (Ω), which is what compactness
means. The variational formulation is
Z Z
∇wn0 · ∇φ + µwn0 φ dx = vn0 φ dx
Ω Ω

Choose φ = wn0 to obtain that

Z Z
2
|∇wn0 | dx 6 vn wn0 dx 6 kvn0 kL2 (Ω) kwn0 kL2 (Ω) .
Ω Ω

Using Poincaré inequality, this gives

Z
2
kwn0 kH 1 (Ω) 6 vn wn0 dx 6 kvn0 kL2 (Ω) kwn0 kH 1 (Ω) ,
0 0
Ω

Therefore the sequence wn0 is bounded in H01 (Ω). Thus there exists another subsequence,
wn00 which converges weakly in H01 (Ω) to a limit w00 . Passing to the limit in the variational
formulation along that subsequence, we obtain
Z Z
00 00
∇w · ∇φ + µw φ dx = vφ dx
Ω Ω

33
Since Corollary 2.4 shows that this has a unique solution, w00 = w does not depend on
the extraction of a subsequence, therefore wn0 * w in H01 (Ω). The Rellich-Kondrachov
Theorem shows that wn0 → w in L2 (Ω). We therefore have the following two variational
formulations: for all φ ∈ H01 (Ω),
Z Z
∇w · ∇φ + µwφ dx = vφ dx,
Ω Ω

and Z Z
∇wn0 · ∇φ + µwn0 φ dx = vn0 φ dx.
Ω Ω
Substracting these two identities, and choosing φ = w − wn0 , we have
Z Z
2 2
|∇(w − wn0 )| + µ |(w − wn0 )| dx = (v − vn0 )(w − wn0 ) dx.
Ω Ω

The right-hand-side is the product of a weakly converging sequence by a strongly converging

one, so Z
lim (v − vn0 )(w − wn0 ) dx = 0.
n0 →∞ Ω

Since wn0 − w → 0 in L2 (Ω),

Z
lim
0
µ |(w − wn0 )|2 dx = 0.
n →∞ Ω

Therefore the last term on the right hand side has a limit, and
Z
lim
0
|∇(w − wn0 )|2 dx = 0.
n →∞ Ω

In other words, wn0 → w in H01 (Ω).

The compactness of (−∆)−1 could have also been used in the first application, instead of
the Sobolev embeddings.

2.4 A Glimpse into positive operators : the maximum

principle
It often happens in applications that finding a non-negative solution is particularly inter-
esting — for instance, when the unknown u models a density. Working in the positive
cone (a convex, closed set) of a given function space is thus sometimes a natural choice. It

34
is particularly judicious whenever the non linearity is unbounded in general, but bounded
when limited to positive functions.
A key tool, to ensure that we indeed stay inside the positive cone is a weak form of the
maximum principle.

Proposition 2.11. Let Ω ⊂ Rn be open, smooth and bounded, let u ∈ H 1 (Ω) and assume
that u > 0 on ∂Ω. Furthermore assume that −∆u > 0 in Ω in a weak sense, i.e.
Z
∇u · ∇v dx > 0 for all v ∈ H01 (Ω), v > 0 a.e. in Ω.
Ω

Then u > 0 a.e. in Ω.

One way of making sense of saying that u > 0 on ∂Ω is to say that min(u, 0) ∈ H01 (Ω).
In this section we will focus on the generic example

−∆u = f (u) in Ω u ∈ H01 (Ω), (1)

An application. Consider the equation (1) with f (x) = e−x . The function f is continu-
ous, but grows very rapidly as u → −∞ and thus we cannot directly apply the previous
technique. However, f is bounded on

C = {u ∈ H01 (Ω) : u > 0} ⊂ H01 (Ω).

Therefore we may consider the map T : C ⊂ H01 (Ω) → H01 (Ω) given by T u = (−∆)−1 f (u).
The set C is convex and closed. Thanks to the maximum principle, T (C) ⊂ C. We now
check that T (C) is precompact. Note that for all f is continuous on H01 (Ω) and when
u ∈ C, f (u) 6 1, therefore
 p 
T (C) ⊂ (−∆)−1 {v ∈ L2 (Ω) s. t. kvkL2 (Ω) 6 |Ω|} .

Thanks to Lemma 2.10 T (C), is therefore contained in a (bounded) closed compact set, and
is therefore precompact, and the existence of a positive solution is obtained by Schauder’s
Fixed Point Theorem (version 2).
We also notice that a solution of (1) is unique, c.f. problem sheet 3.

Definition 6. We say that u ∈ H 1 (Ω) is a weak-subsolution of (1) when

Z Z
∇u · ∇v 6 f (u)v dx ∀v ∈ H01 (Ω), v > 0.
Ω Ω

We say that ū ∈ H 1 (Ω) is a weak-supersolution of (1) when

Z Z
∇ū · ∇v > f (ū)v dx ∀v ∈ H01 (Ω), v > 0.
Ω Ω

35
When sub and super solutions to the problem exists, the weak maximum principle provides
a constructive method to find a solution, as we will see.

Theorem 2.12. Assume there exist a subsolution and a supersolution in the weak sense
to (1), satisfying

u 6 0 6 ū in the sense of traces on ∂Ω and u 6 ū a.e. in Ω.

Suppose that f ∈ C(R), that there exists λ > 0 such that x → f (x) + λx is non-decreasing,
and that |f (t)| 6 C(1 + |t|) for all t ∈ R . Then, there exists a solution of (1) satisfying
u 6 u 6 ū.

Proof. Without loss of generality, we may assume that f is non decreasing, since we can
consider otherwise the problem

−∆u + λu = f (u) + λu, u ∈ H01 (Ω),

and replace in everything that follows ∆ by ∆+λ. Define u0 := u and for all k ∈ {0, 1, 2, . . .}
define inductively uk+1 = T uk where

T (u) := (−∆)−1 f (u).

We have already established that T is a compact map from H01 (Ω) to itself — however, it
is not clear that T takes a bounded closed convex set in H01 (Ω) into itself so one cannot
directly apply one of the FPTs.
The sequence uk is monotone increasing. We claim that u0 6 u1 6 . . . 6 uk 6 . . . a.e. in
Ω. We argue by induction. When k = 0 u1 and −u0 satisfies (in the weak sense)

−∆u1 = f (u0 ), −∆(−u0 ) > (−f (u0 )), in Ω

And u1 − u0 > 0 on ∂Ω therefore the weak maximum principle shows that u1 − u0 > 0 a.e.
in Ω. For an arbitrary k > 1, if uk > uk−1 , since f is non decreasing we have, in the weak
sense,
−∆(uk+1 − uk ) = f (uk ) − f (uk−1 ) > 0
and uk+1 − uk = 0 > 0 on ∂Ω, so the weak maximum principle shows that uk+1 > uk .
The sequence uk is bounded from above by ū. We argue again by induction. It is true for
u0 = u 6 ū. If it is true for uk , then we have

−∆ū > f (ū), −∆(−uk+1 ) = (−f (uk )), in Ω,

36
thus −∆(ū − uk+1 ) > f (ū) − (−f (uk )) > 0 since f is non decreasing and ū − uk+1 = ū > 0
on ∂Ω. Thus, ū > uk+1 almost everywhere.
The sequence uk → u solution of (1) in H01 (Ω). Since u0 6 u1 6 . . . 6 uk 6 . . . 6 ū a.e ,
the limit
u(x) := lim uk (x)
k→∞

exists for almost every x, and from the DCT, uk → u in L2 (Ω). Next, note that for all
k, |f (uk )| 6 C(1 + |ū| + |u0 |), thus f maps the sequence uk into the L2 (Ω) bounded set
A := {v s.t. kvkL2 (Ω) 6 kC(1 + |ū| + |u0 |+)kL2 (Ω) }. Therefore T (uk ) ∈ ∆−1 (A), a bounded
set in H01 (Ω). Therefore, there exists a weakly converging subsequence ukn in H01 (Ω).
Necessarily, ukn * u in H10 (Ω). Since T is continuous, and uk+1 = T uk , uk → u in H01 (Ω),
and passing to the limit in the relationship uk+1 = T uk we conclude the proof.

37
Chapter 3

Variational inequalities and

monotone operators

So far, we have considered semi-linear problems: the nonlinearity only appears in terms
for which the number of derivatives is strictly less than the maximal order of derivatives
appearing in the equation. We used to a great extent that the principal part of the
operator was a linear operator. The second part of this course is devoted to more general
situations, the typical case under consideration being the case of quasi-linear problems. In
this chapter we will prove a general existence theorem for variational inequalities while the
following chapter will then be devoted to applications of this result to quasilinear PDEs
and variational problems.
Let X be a Banach space and X ∗ its dual. If x ∈ X and x∗ ∈ X ∗ we use the notation
hx∗ , xi := x∗ (x).

3.1 Differentiation in Banach spaces

Definition 7. Let X, Y be Banach spaces, M ⊆ X, x ∈ M and F : M ⊆ X → Y . We
say

(1) F has a directional derivative at x in direction e ∈ X if

F (x + te) − F (x)
∂e F (x) := lim ∈Y exists.
t→0,t>0 t

(2) F is Gâteaux differentiable at x, if ∂e F (x) exists for all e ∈ X and is continuous,

i.e. there exists L ∈ L(X, Y ) such that ∂e F (x) = Le.

38
 (3) F is Fréchet differentiable at x if there exists a continuous linear map DF (x) : X → Y
such that
F (x + h) = F (x) + DF (x)h + o(khkX ) as khkX → 0.

(4) F is continuously differentiable at x, if y →

7 DF (y) is continuous in x. If DF is
continuous for all x ∈ M , then we write F ∈ C 1 (M, Y ).

Remarks.

(i) If F is Fréchet differentiable at x then F is Gâteaux differentiable at x.

[h = th̃, h̃ fixed, t → 0.]

(ii) Gâteaux differentiability does not imply Fréchet differentiability.

Theorem 3.1 (Fundamental Theorem of Calculus). For t ∈ [0, 1] let F (x + te) be Gâteaux
differentiable and t 7→ ∂e F (x + te) continuous. Then
Z 1
F (x + e) − F (x) = ∂e F (x + te) dt.
0

Proof. Let y ∗ ∈ Y ∗ be arbitrary, g : [0, 1] → Y , g(t) := F (x + te), h : [0, 1] → R, h(t) :=

hy ∗ , g(t)i.
By assumption g 0 (t) = ∂e F (x + te), h0 (t) = hy ∗ , g 0 (t)i exist. The Fundamental Theorem of
R1
Calculus in R gives h(1) − h(0) = 0 h0 (t) dt, that is
Z 1
∗ ∗
hy , F (x + e)i − hy , F (x)i = hy ∗ , ∂e F (x + te)i dt
0
Z 1
∗
= hy , ∂e F (x + te) dti .
0

Since y ∗ ∈ Y ∗ is arbitrary the claim follows.

Examples.

|∇u|2 dx. Expanding, we find

R
(1) F : H 1 (Ω) → R, u 7→ Ω
Z Z Z
2
F (u + v) − F (v) = 2 ∇u∇v dx + |∇v| dx = 2 ∇u∇v dx + o(kvkH 1 )
Ω Ω Ω
R
and thus F is Fréchet differentiable with DF (u)v = 2 Ω
∇u∇v dx (and even in
C 1 (H 1 (Ω)).

39
 Z q
1
(2) F : H (Ω) → R, u 7→ 1 + |∇u|2 dx, Ω bounded.
Ω
The directional derivative is given by
∇u · ∇e
Z
∂e F (u) = q dx.
Ω 2
1 + |∇u|

Is F ZGâteaux differentiable on H 1 (Ω)? We need to show that L : H 1 (Ω) → R,

∇u · ∇e
e 7→ √ 2 dx is continuous. Indeed,
Ω 1 + ∇u
|∇u|
Z Z
|Le| 6 q |∇e| dx 6 |∇e| dx 6 |Ω|1/2 k∇ekL2 (Ω)
Ω 1 + |∇u|2 Ω
| {z }
61

and thus F is Gâteaux differentiable on H 1 (Ω).

Is F Fréchet differentiable on H 1 (Ω)? Let
∇u · ∇h
Z
DF (u)h := q dx.
Ω 2
1 + |∇u|

Then

|F (u + h) − F (u) − DF (u)h|
 
Z q 
∇u · ∇h 
q
2 2

6  1 + |∇u + ∇h| − 1 + |∇u| − q  dx

Ω 2
1 + |∇u|
 
 
2 |∇u| |∇u|
Z
 
6 |∇h|  q q −q  dx
 1 + |∇u + ∇h|2 + 1 + |∇u|2 2

Ω 1 + |∇u|


2 |∇u| |∇u|
6 khkH 1 (Ω) q q − q
2 2 2

1 + |∇u + ∇h| + 1 + |∇u| 1 + |∇u| 2
L (Ω)
q q q
Since 2 |∇u| /( 1 + |∇u + ∇h| + 1 + |∇u| ) − |∇u| / 1 + |∇u|2 → 0 a.e. (for a
2 2

subsequence) when h → 0 in H 1 (Ω), and this quantity is bounded by 3|∇u| ∈ L2 (Ω),

it follows from the DCT that

|F (u + h) − F (u) − DF (u)h| = o (khkH 1 ) ,

and F is Fréchet differentiable on H 1 (Ω).

40
 R
(3) F : X = L1 (Ω) → R, u 7→ Ω
|u(x)| dx. For t > 0 we find
Z ( )
F (u + tv) − F (u) |u + tv| − |u| (signu)v u 6= 0
Z
= dx → dx
t Ω t Ω |v| u=0

and thus the directional derivative exists and is given by

Z Z
∂v F (u) = (signu)v + |v| .
Ω\{u=0} Ω∩{u=0}

Hence F is only Gâteaux differentiable if u(x) 6= 0 for a.a. x, since otherwise ∂v F (u) 6∈
L(X).

See Problem Sheet 3 for more examples.

We remark that the differentials DF in the examples (i) and (ii) are monotone operators.

3.2 Monotone Operators

In the following, unless specified otherwise, X will always be a reflexive separable Banach
space, X ∗ its dual and M ⊆ X non-empty, convex and closed.

Definition 8. A : M → X ∗ is a monotone operator if

(i) A is monotone, i.e. hA(u) − A(v), u − vi > 0 for all u, v ∈ M ;

(ii) A is hemicontinuous, i.e. for all u, v ∈ M and w ∈ X the map

t 7→ hA((1 − t)u + tv), wi is continuous on [0, 1].

A monotone operator is called strictly monotone if hA(u) − A(v), u − vi > 0, and hA(u) −
A(v), u − vi = 0 if and only if u = v.

Note that if A is continuous from M (strong) to X ∗ weak, then A is hemicontinuous.

Lemma 3.2 (Minty). Let A : M → X ∗ be a monotone operator. Then the following are
equivalent:

(i) hA(u) − ξ, u − vi 6 0 ∀v ∈ M .

(ii) hA(v) − ξ, u − vi 6 0 ∀v ∈ M .

41
This lemma is useful for example when A(u) and u converge weakly to a limit: for a
fixed v, while it is not clear what the limit in (1) will be, because it involves the duality
bracket between two weakly convergent sequences, in (2) only one appears, and the limit
is straightforward.

Proof. Let us first show that (1) implies (2).

hA(v) − ξ, u − vi = hA(v) − A(u), u − vi + hA(u) − ξ, u − vi 6 0.

| {z } | {z }
60 by monotonicity 60 by assumption

Let us now show the other implication. Given v ∈ M , let vε := (1 − ε)u + εv ∈ M . We

have, by assumption

hA(vε ) − ξ, ε(u − v)i = hA(vε ) − ξ, u − vε i 6 0.

ε→0
On the other hand, thanks to the hemicontinuity, hA(vε ) − ξ, u − vi −→ hA(u) − ξ, u − vi,
and we have proved (1).

Problem Sheet 6 is devoted to the study of interesting properties of monotone operators.

In particular, you will show:

• If A is monotone, then A satisfies condition (H3) encountered later on in Theorem

3.5

• Suppose that F : X → R is Gâteaux differentiable in every u in X, with derivative

F 0 (u). Then F is convex if and only if F 0 : X → X ∗ is a monotone operator.

• If A is strictly monotone, then the solution of the variational inequality (1) (with
F = A) is unique.

On Problem Sheet 3 you also find several necessary and sufficient conditions for a function
to be monotone.

3.3 Variational inequalities

Consider F : M → X ∗ . In this section, we wish to address the following problem
Model Problem. Find u ∈ M such that

hF (u), u − vi 6 0 ∀v ∈ M. (1)

Remarks.

42
 • When M = X then (1) is equivalent to hF (u), ηi = 0 ∀η ∈ X —choose v = u ± η.
Thus, F (u) = 0 in X ∗ , thus variational equalities are included in this settings.

• F could come from a variational problem, and correspond to a critical point of a

functional I — F (u) = I 0 (u), with I : X → R and I 0 is its Gâteaux derivative.

We will assume that F satisfies the following three assumptions

(H1) F maps bounded sets into bounded sets.

No pre-compactness is required for example.

(H2) F is coercive with respect to some u0 ∈ M , that is, there exists u0 ∈ M such that

hF (u), u − u0 i
→ ∞ as kuk → ∞, u ∈ M.
ku − u0 k

This hypothesis is only relevant when M is unbounded.

If F is additionally a monotone operator then we find

Theorem 3.3. Let A : M → X ∗ be a monotone operator (where as usual M is convex,

non-empty and closed and X is a seperable reflexive Banach space) so that conditions (H1)
and (H2) are satisfied. Then the variational inequality (1) has a solution.

We will obtain this result as a consequence of a more general theorem in which the assump-
tion of F being a monotone operator is replaced with the weaker condition (c.f. Problem
sheet 3)

(H3) F is satisfies the following weak sequential lower semi-continuity condition: Given a
sequence (un ) n ∈ N, un ∈ M , such that un * u in X with u ∈ M , and F (un ) * ξ
in X ∗ , then
hξ, ui 6 lim hF (un ), un i.
n→∞

Furthermore, if hξ, ui = lim hF (un ), un i then

n→∞

hF (u) − ξ, u − vi 6 0 ∀v ∈ M.

As above this implies F (u) = ξ if M = X. This last assumption seems more mysterious
that the other two, but as we will see is very natural in the context of calculus of variations.
It seems we do not require F to be continuous, but it is partially an illusion as the following
proposition shows.

43
Proposition 3.4. Suppose that F satisfies (H1) and (H3) and that M has a non-empty
interior. Then, if un → u in Int (M ), then F (un ) * F (u) in X ∗ .

Proof. Let un be a sequence such that un → u in X ∩ M . Then, kun kX < ∞ and since
F maps bounded sets into bounded sets, kF (un )kX ∗ is bounded and X ∗ is reflexive, so
we may extract a subsequence un0 such that un0 → u in M and F (un0 ) * ξ in X 0 . Thus
hF (un0 ), un0 i → hξ, ui (weak-strong duality bracket..), and (H3) implies that

hF (u) − ξ, u − vi 6 0 ∀v ∈ M.

Since u ∈ Int (M ), there exists a ball B(u, δ) ⊂ M with δ > 0. Therefore, for any
w ∈ X \ {0}, choose v = u − λw w, with λw = δ/(2 kwkX ) > 0, and divide by λw to obtain

hF (u) − ξ, wi 6 0 ∀w ∈ X,

thus F (u) = ξ. Since the limit does not depend on the subsequence, the convergence holds
globally.

The motivation of the introduction of Assumptions (H1), (H2) and (H3) is the following
result.

Theorem 3.5. If F satisfies conditions (H1), (H2) and (H3) there exists u ∈ M such that
hF (u), u − vi 6 0 for all v ∈ M .

Remark. In condition (H2) it is essential that u0 ∈ M !

2
Consider e.g. X = R2 , M = {x2 = 1}, I(x) = ex1 + ex1 (x2 −1) .
Thus I(x) = ex1 + 1 on M and I does not attain its minimum on M .
 2 2

F (x) = ∇I(x) = ex1 + 2x1 (x2 − 1)ex1 (x2 −1) , x21 ex1 (x2 −1) .

For x ∈ M :
hF (x), xi x1 ex1 + x21
= p 2 → ∞ as |x1 | → ∞.
|x| x1 + 1
Hence F is coercive with respect to 0 6∈ M , but F does not have a zero on M .

3.4 Proof of Theorem 3.5 by the Galerkin Method

The Galerkin method is very general and robust. The idea is as follows. To tackle a
problem posed in an infinite dimensional space, start with a studying its approximation

44
on a nested sequence of finite dimensional sub-spaces. Solving the approximate problem
is generally simpler than solving the infinite dimensional one. Passing to the limit, we
construct a solution of the original problem. This method is very popular for numerical
methods, because it is constructive. It is also interesting from a theoretical point of view.
We proceed in several steps. We first recall a result concerning Banach spaces

Lemma 3.6. Let X be a separable Banach space of infinite dimension. There exists a
countable linearly independent family (vi )i∈N , vi ∈ X, such that the linear combinations of
vi are dense in X.

Another way of writing this is that if Xi = span(v1 , . . . , vi ) then the ∪∞

i=0 Xi is dense in X.
For problem, we will apply this result to both X and M and choose X1 so that u0 , the
coercivity point from (H2), is in X1 . To construct a finite dimensional approximation,
simply restrict the variational problem (1) to Xi and Mi = M ∩ Xi .
Let us first address the approximated problem.

Proposition 3.7. If F satisfies the conditions (H1), (H2), (H3) and if dim X < ∞, then
there exists u ∈ M such that hF (u), u − vi 6 0 for all v ∈ M .

Proof. Without loss of generality we may assume that M does not have an empty interior
(otherwise, reduce dimensions).
1. The interior case. Let us first prove the result for the restriction of F to a smaller closed
convex set M̃ such that M̃ ⊂ Int (M ). Proposition 3.4 shows that F is continuous on M̃
— since in finite dimensional spaces, weak and strong convergences are equivalent. Using
Riesz Theorem, there exists F̃ : M̃ → X such that hF (u), vi = (F̃ (u), v) for all v ∈ X.
Recall the Projection Theorem:

Let X be real Hilbert space, M̃ ⊂ X non-empty, closed, convex. Then there

exists a unique map P : X → M̃ such that

kx − P (x)kX = dist(x, M̃ ),

which is equivalent to

(x − P (x), a − P (x)) 6 0 for all a ∈ M̃ .

Using the orthogonal projection P on M̃ we define a map G : M̃ → M̃ via G := P ◦(Id− F̃ ).

The map G is continuous, since it is the composition of continuous maps. Brouwer’s Fixed

45
 x

P (x)
a

Point Theorem implies that that there exists u ∈ M̃ such that G(u) = u. Hence P (w) = u
with w = u − F̃ (u). Due to the properties of P we find

hF (u), u − vi = −(v − u, F̃ (u)) = (v − P (w), w − P (w)) 6 0 for all v ∈ M̃ ,

2. The bounded case. Given n > 0, consider now M fn = M ∩{x ∈ M s.t. dist(x, ∂M ) > 1 },
n
a closed, convex set, inside Int (M ) non-empty for n big enough (the convexity follows from
that of M ). From the previous step, there exists un ∈ M fn such that

hF (un ), un − vi 6 0 for all v ∈ M

fn . (1)

Since M is bounded, there exists a subsequence such that un0 → u and F (un0 ) → ξ as
n0 → ∞. Passing to the the limit on the left-hand side in (1), and noting that the space
M
fn1 ⊂ M fn2 if n1 < n2 , we obtain hξ, u − vi 6 0 for all v ∈ M
fn for any n. Passing now
to the limit in n, we have

hξ, u − vi 6 0 for all v ∈ ∪n M

fn = M.

On the other hand, Condition (H3) (and strong convergence) implies that

hF (u) − ξ, u − vi 6 0 for all v ∈ M.

Therefore hF (u), u − vi 6 hξ, u − vi 6 0.

3. The unbounded case. Now let M be unbounded and define now M
fR := M ∩ BR (0).
We have obtained that there exists uR such that

hF (uR ), uR − vi 6 0 for all v ∈ M

fR .

The coercivity condition (H2) implies that there exists u0 ∈ M such that hF ku−u
(u),u−u0 i
0k
→∞
as kuk → ∞. Hence there exists C0 > 0 such that hF (u), u − u0 i > 0 for all u with
kuk > C0 . Choose now R := max(ku0 k , C0 ) + 1. Then, u0 ∈ MfR , and therefore

hF (uR ), uR − u0 i 6 0

46
and consequently kuR kX 6 C0 . Then, for any v ∈ M introduce vR = (1 − θ)uR + θv, with
θ = (C0 + kvkX )/(C0 + kvkX + 1) (or anything else small enough). Then, kvR kX 6 R, that
is, vR ∈ MR and therefore

hF (uR ), uR − vi = θ−1 hF (uR ), uR − vR i 6 0,

Thus uR is a solution.

As the bound on kuR k depends only on the constant C0 obtained from the coercivity
condition (H2) we thus conclude

Corollary 3.8. Let M, X, F be as in Theorem 3.5 and let Xi = span(v1 , . . . , vi ) the finite
dimensional subspaces obtained in Lemma 3.6.
Then there exist solutions ui of the variational inequalities

hF (ui ), ui − vi 6 0, for all v ∈ Mi

which are uniformly bounded, i.e. so that sup kui k < ∞.

Conclusion of the proof of Theorem 3.5. Since the sequence ui is bounded it follows from
(H1) that the sequence F (ui ) is also bounded. Therefore we may extract a subsequence
un such that un * u weakly in M (since M is closed), and F (un ) * ξ weakly in X ∗ .
Since the sets Mi and Xi are nested, that is, Mi ⊂ Mi+1 and Xi ⊂ Xi+1 , we have, for all
n > i,
hF (un ), un i 6 hF (un ), vi for all v ∈ Mi .
Now, for any given v ∈ Mi hF (un ), vi → hξ, vi, and in turn

lim infhF (un ), un i 6 hξ, vi for all v ∈ ∪∞

i=0 Mi = M.

Since condition (H3) guarantees that hξ, ui 6 lim infhF (un ), un i, choosing v = u ∈ M , we
have in fact lim infhF (un ), un i = limhF (un ), un i = hξ, vi. The second part of condition
(H3) then shows that

hF (u), u − vi 6 hξ, u − vi for all v ∈ M.

Finally, notice that for any from condition (H3) again, since un − v * u − v for a given i
and v ∈ Mi , and by construction hF (un ), un − vi 6 0 for all n > i,

hξ, u − vi 6 lim infhF (un ), un − vi 6 0 for all v ∈ ∪∞

i=0 Mi = M,

which is what we wished to show.

47
Chapter 4

Applications to quasilinear PDEs and

variational inequalities

We shall now detail various applications of Theorem 3.5.

4.1 Some Quasilinear problems in W01,p(Ω)

Let Ω be a smooth bounded domain in Rn , p ∈ (1, ∞), X = W 1,p (Ω) and F : Rn → Rn
continuous and monotonous, which satisfies the growth condition

F (λ) 6 C(1 + |λ|p−1 ) for all λ ∈ Rn .

Suppose furthermore that there exists α > 0 such that for all

F (λ) · λ > α|λ|p for all λ ∈ Rn .

We denote by p0 the conjugate exponent of p, i.e. p−1 + (p0 )−1 = 1.

0
Theorem 4.1. For all f ∈ W −1,p (Ω), there exists u ∈ W01,p (Ω) such that

−div (F (∇u)) = f in the sense of D0 (Ω).

If F is strictly monotonous, this solution is unique.

Proof. Let us check that A := −div (F (∇u)) − f is a map from X to X ∗ which satisfies
(H1), (H2) and (H3). Let us first verify that −f satisfies (H1) and (H3). By the definition
0
of X ∗ = W −1,p (Ω), there exist a constant C < ∞ (the smallest such constant is kf kX ∗ )
such that for all v ∈ X we have |hf, v, i | 6 C kvkX , so f maps bounded sets into bounded

48
sets, so (H1) is satisfied. Hypothesis (H3) is obviously true: since f does not depend on u,
since ξ = f . Let us now turn to B(u) := −div (F (∇u)). By assumption
  p
p0  p−1
p−1
F (λ) 6 C 1 + |λ| 6 C 0 (1 + |λ|p ) .

 
0
Therefore, if ∇u ∈ Lp (Ω), then F (∇u)p  ∈ L1 (Ω), and for any w ∈ X = W01,p (Ω) we have
 

Z
p−1

hB(u), wi = F (∇u) · ∇w dx 6 C 1 + kukX kwkX
Ω

Therefore B : X → X ∗ maps bounded sets into bounded sets. Let us now verify that
B is monotone and hemicontinuous: this will imply that (H3) is also satisfied. From the
monotonicity of F , we have
Z
hB(u) − B(v), u − vi = (F (∇u) − F (∇v)) · (∇u − ∇v) dx > 0,
Ω

and Z
hB(tu + (1 − t)v), wi = F (t∇u + (1 − t)∇v) · ∇w dx.
Ω
Since F is continuous, the hemicontinuity follows from the Dominated Convergence The-
orem.
Finally, let us verify that A is coercive at 0. We have
Z
hA(u), ui > F (∇u) · ∇u dx − kf kX ∗ kukX > α k∇ukpLp (Ω) − kf kX ∗ kukX .
Ω

Thanks to Poincaré’s inequality, α k∇ukpLp (Ω) > α0 kukpX , thus we have

|hA(u), ui|
> α0 kukp−1
X − kf kX ∗ ,
kukX

and this lower bound tends to infinity with kukX , since p > 1.

Example. Find a solution to the p-Laplacian, that is

−div |∇u|p−2 ∇u + µu = f


in Ω, with u = 0 on ∂Ω.

Where µ ∈ R, and the domain is smooth and bounded. For this equation to make sense,
a natural choice looking at the principal terms seems to be u ∈ W01,p (Ω). For the right-
hand-side to make sense, we need f φ dx to be well defined for φ ∈ W01,p (Ω). So choose
R
0
f ∈ Lp (Ω), with p1 + p10 = 1, or better, using Sobolev embeddings choose f ∈ Lp̃ (Ω) with

49
1
+ p1 = 1 + n1 if p < n. We also need uφ dx to make sense. Since u and φ lie in the same
R
p̃
space, the best we can do is
 Z 
 
µ uφ dx 6 |µ| kuk 2 kφk 2 ,
  L (Ω) L (Ω)

that is, require that u ∈ L2 (Ω). Therefore a good space to work in is X = W01,p (Ω) ∩ L2 (Ω)
when µ 6= 0, and X = W01,p (Ω) if µ = 0.
When µ = 0. We can apply theorem 4.1 provided we check that F (λ) := |λ|p−2 ζ satisfies
the required properties. The bound |F (λ)| 6 |λ|p−1 comes immediately, and the lower
bound F (λ)·λ > |λ|p as well. We simply need to check that F is monotone and continuous,
but that is clear since p > 1, F is the differential of p1 |λ|p a convex function since p > 1.
When µ > 0. No change for the principal part. Let us check directly that the second term
hµu, φi satisfies the (H1) (H3) requirements:

hµu, φi 6 |µ| kukL2 (Ω) kφkL2 (Ω)

so a bounded set in X = W01,p (Ω) ∩ L2 (Ω) is mapped into a bounded set. Next, suppose
un * u. We write
Z Z
hµun , un i = µ (un − u) − u + +2uun dx > µ −u2 + +2uun dx,
2 2
Ω Ω

and the right-hand side term has a limit, hµu, ui. Thus lim infhµun , un i > hµu, ui. The
second identity is obvious since u → µu is linear.
Finally, let us verify that (H2) is holds with 0 as a coercivity point. Let us choose the norm
on X = W01,p (Ω) ∩ L2 (Ω) to be kukX = k∇ukLp (Ω) + kukL2 (Ω) . This norm is equivalent to
the canonical norm, X = k∇ukLp (Ω) + k∇ukL2 (Ω) + kukL2 (Ω) by Poincaré’s inequality. We
have

h−div |∇u|p−2 ∇u + µu, ui

Z Z
> |∇u| dx + µ |u|2 dx
p

Ω Ω
k∇ukLp (Ω)
= θ k∇ukp−1
Lp (Ω) + (1 − θ) kµukL2 (Ω) with θ = .
kukX

If kukL2 (Ω) → ∞ while k∇ukLp (Ω) stays bounded, θ → 0 and the lower bound tends to
infinity. Symmetrically, kukL2 (Ω) stays bounded while k∇ukLp (Ω) → ∞, θ → 1 and the
lower bound tends to infinity. Finally, when both tend to infinity, θa + (1 − θ)b > min(a, b)
therefore the lower bound also tends to infinity.

50
 When µ < 0. No change for the principal part. We cannot just apply the same argument
as before (the inequalities are in the wrong direction!). The coercivity will also be an issue,
if we cannot dominate the L2 (Ω) part by the W01,p (Ω). We therefore wish to use some
compactness to bypass these difficulties. So we only consider the case when p > 2. Then,
the Rellich-Kondrachov embeddings show that W01,p (Ω) ,→ Lp (Ω), and L2 (Ω) ⊂ Lp (Ω)
since Ω is bounded. Indeed, using Hölder,
Z Z Z
r−1
2 2r r1
u dx 6 ( u ) ( 1) r
Ω W W

and thus for r = p/2 > 1, we obtain kukL2 (Ω) 6 C kukLp (Ω) .
Now, given a sequence un * u in X = W01,p (Ω)(= W01,p (Ω) ∩ L2 (Ω)), we have un → u
in L2 (Ω) by compactness, therefore hµun , un i → hµu, ui, so (H3) is satisfied. Assumption
(H1) is satisfied as before.
Finally, let us verify that (H2) holds, with 0 as coercivity point. We have

h−div |∇u|p−2 ∇u + µu, ui

Z Z
> |∇u| dx + µ |u|2 dx
p

Ω Ω
= k∇ukLp (Ω) − |µ| kuk2L2 (Ω)
p

> k∇ukpLp (Ω) − C |µ| kuk2Lp (Ω)

> kukpX − C |µ| kuk2X

and since p > 2, this tends to infinity with kukpX .

The conclusion is that when µ = 0, the solution is unique in W01,p (Ω) for p > 1, when
µ > 0, the solution is unique in W01,p (Ω) ∩ L2 (Ω) for p > 1, and when µ < 0 the solution
exists for p > 2.

4.2 A remark on the growth bound

In the theorem on the existence of solution to the variational inequality problem we proved,
we did not impose a specific growth constraint, but we required that F maps bounded sets
in M into bounded sets in X ∗ . In the two applications we just considered, we satisfied
this hypothesis by imposing a growth bound on F , namely that F (λ) grows at most
polynomially in λ. In the case of exponential growth, the situation is much more delicate,
as the following example shows.

51
Consider the following one-dimensional boundary value problem

−u00 (x) = Ceu in (0, 1), with u(0) = u(1) = 0.

where C > 0 is a constant. Let us show that there is no u ∈ C([0, 1]) solution to this
problem for some values of C. Let us first show that there is no solution in H01 (Ω). Note
that the function ψ(x) = sin(πx) satisfies −ψ 00 (x) = π 2 ψ(x), and ψ(0) = ψ(1). Writing
the weak form of the equation, we find
Z 1 Z 1 Z 1
0 0 00
u w dx = − uw dx = C eu wdx,
0 0 0

for all w ∈ H 2 (0, 1) ∩ H01 (Ω). Choosing w = ψ, and placing all u dependent terms on the
right hand side, we have Z 1
(π 2 u − Ceu )ψdx = 0.
0
2
The function x → π 2 x − Cex is negative for all for C > πe . Thus, for any such C this
implies that (π 2 u − Ceu )ψ < c < 0, which leads to a contradiction. To show that there is
no continuous solution to the weak formulation
Z 1 Z 1
00
− uw dx = C eu wdx,
0 0

with w ∈ Cc∞ (0, 1), we can argue similarly, since ψ can be obtained as the limit of such
functions.
On the other hand for C small enough, one can prove existence of a solution using e.g.
Schauder’s FPT, and indeed it is possible to find a closed form solution, and the threshold
x

2
value of non-existence by hand, C = 8 maxx>0 cosh x
≈ 3.5138307. Up to that critical
value, the solution (represented below for C = 3.5138306) is perfectly smooth.
This non-existence proof relies crucially on the existence of a non-negative function ψ ∈
H01 (Ω) such that −∆ψ = λψ with λ > 0. This fact is true in any dimension for smooth
domains but this is out of the scope of this course.

52
4.3 A (simple) variational inequality: an elastic beam
u

0 l

We consider an elastic beam whose deviation from a flat state is described by a function
v : [0, l] → R. We fix the beam at the end points x = 0, l such that v satisfies the boundary
conditions v(0) = v(l) = 0.
A simple model for the energy of the beam is
Z ln
a0 00 2 o
I(v) = |v | + f v dx
0 2
where f : [0, l] → R represents an outer force and a0 > 0 is an elasticity constant.

53
Furthermore g : [0, l] → R, g smooth, g(0), g(l) < 0 represents a rigid obstacle.
Let f ∈ L2 (0, l) and X = H 2 (0, l) ∩ H01 (0, l). Then X is a reflexive separable Banach space
and kuk := ku00 kL2 (0,l) + ku0 kL2 (0,l) is a norm on X due to Poincaré’s inequality.
We are interested in finding a minimiser u of I in the set

M := v ∈ H 2 (0, l) ∩ H01 (0, l) | v(x) > g(x) ∀x ∈ [0, l]



Remark. Since H01 (0, l) ,→ C([0, l]) it makes sense to say that v(x) > g(x) for all x.

Claim. M is a non-empty, closed and convex subset of H 2 (0, l) ∩ H01 (0, l)

Proof.

(a) Convexity: if u, v ∈ M then tu(x) + (1 − t)v(x) > g(x) for all t ∈ [0, 1] and hence
tu + (1 − t)v ∈ M ∀t ∈ [0, 1].

(b) Closedness: if (un ) ⊂ M such that un → u in H 2 (0, l)∩H01 (0, l) then, since H01 (0, l) ,→
C([0, l]), we find that un → u uniformly in C([0, l]) Thus u(x) > g(x) for all x ∈ [0, 1]
and consequently u ∈ M .

(c) M non-empty: Since g(0), g(l) < 0 and g is smooth we can easily construct u ∈ M ,
e.g. by a parabola with sufficiently negative curvature.

Next we derive the variational inequality which is satisfied by a minimiser u of I on M . If

u is minimiser, then I(u) 6 I(v) for all v ∈ M . If v ∈ M then u + ε(v − u) ∈ M for all
ε ∈ (0, 1) and thus I(u) 6 I(u + ε(v − u)) for all v ∈ M and ε ∈ (0, 1) and consequently
I(u + ε(v − u)) − I(u)
06 for all v ∈ M and ε ∈ (0, 1)
ε
Passing to the limit ε → 0 we find ∂v−u I(u) > 0 for all v ∈ M such that
Z l
0 6 ∂v−u I(u) = a0 u00 (v 00 − u00 ) + f (v − u) dx =: hF (u), v − ui.
0

We now introduce the operator

Z l
∗
2
A : H (0, l) ∩ H01 (0, l) 2
→ (H (0, l)) , hA(u), vi := a0 u00 v 00 dx.
0

On Problem Sheet 4 you show that A is a strongly monotone operator on H 2 (0, l)∩H01 (0, l)
and hence coercive and satisfies the continuity condition (H3). To apply the existence
Theorem 3.5 we still need to show that
A maps bounded into bounded sets: Let u ∈ H 2 ∩ H01 and kuk 6 K Then

|hA(u), vi| 6 a0 ku00 kL2 kv 00 kL2 6 a0 kuk kvk 6 a0 K kvk

54
and thus sup kA(u)k(H 2 )∗ 6 a0 K.
kuk6K
Z l
2 ∗
Now consider F : M → (H ) defined via hF (u), vi := hA(u), vi + f v dx, i.e. we add a
0
constant linear operator to A. This operator is well-defined and continuous, since
Z l 
 
 f v dx 6 kf k 2 kvk 2 6 C kf k 2 kvk ,
  L L L
0

and as a constant operator it is also hemicontinuous.

We conclude that F is also a coercive monotone operator, (notice that the constant part
drops out in the monotonicity condition: hF (u) − F (v), u − vi = hA(u) − A(v), u − vi >
c0 ku − vk2 ), which maps bounded into bounded sets.
Hence, due to the Theorem 3.5 and the uniqueness result for strictly monotone operators
there exists a unique u ∈ M such that hF (u), u − vi 6 0 for all v ∈ M .

What else can we say about u?

Let D = {x ∈ (0, l) | u(x) > g(x)}. Notice that, since u and g are continuous, the set D is
open. Let us also assume that f is smooth and that u is smooth in D – not that difficult
to show, but outside the scope of this lecture where regularity is not mentioned. Then for
any x0 ∈ D there exists δ > 0 such that Bδ (x0 ) ⊂ D. Now choose η ∈ C0∞ (Bδ (x0 )) such
that we can take v = u ± εη ∈ M for sufficiently small ε. We conclude that
Z l
a0 u00 η 00 + f η dx = 0
0

and after an integration by parts that

Z l
(a0 u00 )00 + f η dx = 0.
0

Since x0 ∈ D is arbitrary and η ∈ C0∞ (Bδ (x0 )) is arbitrary, it follows that

(a0 u00 )00 + f = 0 in D.

Thus, whenever the beam is strictly above the obstacle, the function v fulfils the Euler–
Lagrange equation for I. Where the beam is sitting on obstacle, v is obviously equal to
the obstacle and we can also conclude that (a0 u00 )00 + f > 0.
An important question in pratice is how smooth u is at the points where it touches the
obstacle (less regular implies more prone to break. It is indeed quite smooth, but a proof
is (far) beyond the scope of this lecture.

55
Remark. We could also consider a domain Ω ⊂ Rn , smooth and bounded, and a smooth
function g : Ω → R with g 6 0 on ∂Ω. The higher dimensional analogue of the energy I is
a0
Z
I(v) = |∆v|2 + f v dx .
Ω 2
We are interested in finding a minimiser u of I on the set
M := v ∈ H 2 (Ω) ∩ H01 (Ω) | v > g a.e. on Ω .


Notice that now we can in general not request that v(x) > g(x) for all x ∈ Ω, since a
function in M is not necessarily continuous (even though in the case of a plate, i.e. in case
n = 2, it is, due to the embedding theorems).
We can now proceed analogously to the one-dimensional case (see Problem Sheet 7) to find
that there exists a unique u ∈ M that minimizes I on M . This u satisfies
Z
a0 ∆u∆(v − u) + f (v − u) dx > 0 for all v ∈ M .
Ω
Define D := {x ∈ Ω | u(x) > g(x)}. One cannot immediately say that D is open, however,
one can prove (which is again not possible within this lecture) that the minimiser u is
sufficiently smooth, such that indeed D is open. Proceeding then as above, we find that u
solves a0 ∆2 u + f = 0 in D. The set ∂D ∩ Ω is an unknown in this problem, a so called
free boundary. The advantage of variational inequalities is that one can study this problem
without having to decide what this free boundary is.
In the next section we consider another physically motivated problem.

4.4 The Stationary Navier–Stokes equations

In this chapter we consider a container, whose interior is represented by a smooth, open
and bounded domain Ω ⊂ R3 , which is filled with a fluid. We denote by
u : Ω → R3 the velocity field of the fluid,
p : Ω → R the pressure,
f : Ω → R3 an outer force density, e.g. gravity,
Re > 0 the Reynolds number.
Our goal is to find a solution of the stationary Navier-Stokes equations, which are given
by

1
− ∆u + (u · ∇)u + ∇p = u in Ω (1)
Re
divu = 0 in Ω (2)
u = 0 on ∂Ω (3)

56
where
3
X
(u · ∇)ui = uj ∂j ui .
j=1
P3
(Recall also: ∆u ∈ R3 , (∆u)i = j=1 ∂jj 2
ui ; divu ∈ R, divu = 3j=1 ∂j uj .)
P

Equation (1) comes from conservation of momentum, (2) means that the fluid is incom-
pressible, and (3) is the so-called no-slip boundary condition.
Our first goal is to find an appropriate weak formulation of (1)-(3). First note the following
algebraic identity
3
X 3
X
div(u ⊗ u)i = ∂j (ui uj ) = uj ∂j ui + uj div(u).
j=1 j=1

Therefore, when div(u) = 0, we have

 
1 1
− ∆u + (u · ∇)u + ∇p = −div ∇u − u ⊗ u + ∇p = −div (F (∇u, u)) + ∇p
Re Re
1
with F (∇u, u) := Re ∇u − u ⊗ u.
Now we define the space X = H01 (Ω, R3 ) endowed with the norm kukX = k∇ukL2 (Ω)3×3
and consider the incompressible subset

M = {v ∈ X | divv = 0 a.e. in Ω},

that is, be the subset of divergence-free vector-fields. We have incorporated all the con-
straints coming from the equations.
Given w ∈ M , we find
Z Z
−div (F (u)) · w + ∇p · w dx = f · w dx
Ω Ω

An integration by parts gives

Z Z
F (u) : ∇w + pdivw = f · w dx,
Ω Ω

where for A, B ⊂ Rn×n we denote A : B = ni,j=1 aij bij . Note that the second term on the
P

left-hand side is nought, since w ∈ M .

For the right-hand side to make sense, we choose f ∈ L2̃ , with 2̃1 + 12 = 1 + 13 , that is, 2̃ = 65 .
We now have a clearly defined problem : find u ∈ M such that

hF (u), wi = hf , wi for all w ∈ M. (4)

57
Theorem 4.2. Given f ∈ L6/5 (Ω, R3 ) there exists a solution to (4). If kf kL6/5 (Ω) is small
enough, the solution is unique.

Proof. We need to show:

(a) M is a closed subspace of X.

Let vn → v in M . Then there exists a subsequence (vnk ) such that ∇vnk → ∇v
a.e. in Ω Thus 0 = divvnk → divv a.e. in Ω. Thus divv = 0 a.e. in Ω, that is, v ∈ M .

(b) If f ∈ L6/5 (Ω) then hf , wi is well-defined.

The embedding theorems give H01 (Ω, R3 ) ,→ L6 (Ω, R3 ). If p = 6 then p0 = 6/5 and
0
thus f ∈ Lp ' (L6 )∗ implies that f ∈ X ∗ . Hence hf , wi is indeed the representation
of a map in X ∗ .

(c) F : X → X ∗ is well-defined and maps bounded into bounded sets. We estimate

1
Z Z
|hF (∇u, u), wi| = ∇u : ∇w dx − u ⊗ u : ∇w dx
Re Ω Ω
1 2
6 k∇ukL2 k∇wkL2 + |u| L2 k∇wkL2
Re
1
6 kukX kwkX + kuk2L4 k∇wkL2 .
Re
Thanks to the Sobolev embeddings, H01 (Ω, R3 ) ,→ L4 (Ω, R3 ), thus
 
1 2
|hF (∇u, u), wi| 6 kukX + C kukX k∇wkL2 .
Re
which implies that F is well-defined and maps bounded into bounded sets.

(d) F (∇u, u) − f is coercive.

For any w ∈ Cc∞ (Ω) ∩ M , we have
Z 3
Z X
w ⊗ w : ∇w dx = wj ∂j wi wi dx
Ω Ω i,j=1

3
1X
Z
= wj ∂j (wi2 ) dx
Ω 2 i,j=1

3
1X
Z
= − (∂j wj )wi2 dx
Ω 2 i,j=1
1
Z
= − |w|2 divw dx = 0.
Ω 2

58
 By approximation, this also holds for all w ∈ M . Therefore
1
Z
hF (∇u, u) − f , ui > ∇u : ∇u dx + 0 − kf kL6/5 kukX = kukX (kukX − kf kL6/5 ) .
Re Ω
and the coercivity follows.

(e) F (∇u, u) − f satisfies (H3).

Suppose un * u weakly in X with u ∈ M . Then, ∇un * ∇u in L2 . Since the
embedding X ,→ L4 is compact, un → u in L4 , and un ⊗ un → u ⊗ u in L2 . In
particular, F (∇un , un ) − f * F (∇u, u) − f . We now compute

1
Z Z
hF (∇un , un ) − f , un i = ∇un : ∇un dx − 0 − f · un dx
Re Ω ΩZ
1 1
Z
> |∇(un − u)|2 dx − ∇u : ∇u dx
Re Ω Re Ω
2
Z Z
+ ∇u : ∇un dx − f · un dx
Re Ω ΩZ
1 2
Z Z
> − ∇u : ∇u dx + ∇u : ∇un dx − f · un dx
Re Ω Re Ω Ω
1
Z Z
→ ∇u : ∇u dx − f · u dx
Re Ω Ω
= hF (∇u, u) − f , ui,

and (H3) is verified.

We have obtained that there exists u ∈ M such that for all v ∈ M ,

hdiv(F (u)) − f , u − wi 6 0 for all w ∈ M.

Now, note that if η = u − w ∈ M , then so is −η, so the equality holds.

Let us now turn to uniqueness. Assume that u1 and u2 are two solutions of (4). Let us
first bound their norm. Using w = u1 in (4), we obtain
Z Z Z Z
∇u1 : u1 = ∇u1 : u1 − u1 ⊗ u1 : ∇u1 = f · u1 6 C kf kL6/5 (Ω) ku1 kX ,
Ω Ω Ω Ω

using Solobev embeddings, therefore ku1 k+ ku2 k6 C kf kL6/5 (Ω) . Next, we note that

1
|u1 ⊗ u1 − u2 ⊗ u2 | = |(u1 + u2 ) ⊗ (u1 − u2 ) + (u1 − u2 ) ⊗ (u1 + u2 )|
2
6 |u1 + u2 | |u1 − u2 | .

59
Subtracting (4) written for u1 and u2 we have
Z Z
∇(u1 − u2 ) : ∇w = (u1 ⊗ u1 − u2 ⊗ u2 ) : ∇w 6 ku1 + u2 kL4 ku1 − u2 kL4 kwkX .
Ω Ω

Applying this inequality to w = (u1 − u2 ) we obtain

ku1 − u2 k2X 6 ku1 + u2 kL4 ku1 − u2 kL4 ku1 − u2 kX .

Thanks to the embedding X ,→ L4 , we obtain

ku1 − u2 k2X 6 C ku1 + u2 kX ku1 − u2 k2X 6 C kf kL6/5 (Ω) ku1 − u2 k2X .

This implies u1 = u2 if C kf kL6/5 (Ω) < 1.

Remark. To solve the full problem one also needs to show that there exists p ∈ L2 (Ω) such
that
1
Z Z
{∇u : ∇φ + u · ∇φu + pdivφ} dx = f · φ dx ∀φ ∈ X.
Re Ω Ω
This can be done, but is outside the scope of this lecture.

4.5 Appendix: Convex hulls of precompact sets are

precompact
Lemma 4.3 (Mazur). Let X be a Banach space and M ⊆ X be precompact. Then conv(M )
is precompact.

Proof. Let ε > 0. Since M is precompact, there exist x1 , . . . , xN ∈ M such that for all
x ∈ M there exists i ∈ {1, . . . , N } with
ε
kx − xi kX < . (∗)
2
Define v(x) := j, where j is the smallest index such that (∗) is satisfied. If y ∈ conv(M )
then y = m
P Pm
i=1 αi yi for some yi ∈ M and αi ∈ [0, 1] with i=1 αi = 1.
Then

m m m
X X X ε
y − αi xv(yi ) = αi (yi − xv(yi ) ) 6 αi yi − xv(yi ) X < .


i=1

i=1

i=1
2
X X
Pm
Since i=1 αi xv(yi ) ∈ K =: conv(x1 , . . . xN ) we thus showed that
[
conv(M ) ⊆ B 2ε (x).
x∈K

60
But K is also the image of a compact set under the following continuous map:
X
ψ : [0, 1]n → X : (α1 , . . . , αn ) 7→ α i xi .

With A := {(α1 , . . . , αn ) ∈ [0, 1]n :

P
αi = 1} we have ψ(A) = K. Hence K is compact
and thus there exist k1 , . . . , km ∈ K such that K ⊆ M
S
i=1 B 2 (Ki ). As a consequence
ε
SM
conv(M ) ⊆ i=1 Bε (Ki ).

4.6 Appendix: Characterisation of compact sets in C

and Lp
Theorem 4.4 (Arzela–Ascoli). Let Ω ⊆ Rn be bounded and F ⊂ C(Ω)N be a family
of continuous functions on Ω. Then F is precompact if and only if F is bounded and
equicontinuous, i.e.

(i) supf ∈F supx∈Ω |f (x)| 6 C;

(ii) supf ∈F |f (x + h) − f (x)| → 0 as |h| → 0 (∀x ∈ Ω, x + h ∈ Ω).

Theorem 4.5 (Riesz–Kolmogorov). Let 1 6 p < ∞ and F ⊂ Lp (Rn ). Then F is precom-

pact if and only if

(i) supf ∈F kf kLp (Rn ) 6 C;

(ii) supf ∈F kf (· + h) − f kLp (Rn ) → 0 as |h| → 0;

(iii) supf ∈F kf kLp (Rn \BR (0)) → 0 as R → ∞.

4.7 Appendix: Positive and negative parts of W 1,p(Ω)

functions are in W 1,p(Ω)
Given u ∈ W 1,p (Ω), define u+ = max(0, u) and u− = max(0, −u).

Lemma 4.6. Assume u ∈ W 1,p (Ω) and let H be the Heaviside function given by

H(x) = 1 if x > 0 and 0 otherwise.

61
Then, u+ , u− , |u| ∈ W 1,p (Ω) and

∇ u+


= H (u) ∇u
∇ u− = −H (−u) ∇u

∇|u| = ∇u (H (u) − H (−u))

Proof. It is sufficient to prove it for p = 1. Indeed, if p > 1 then in any open Ω1 such that
Ω̄1 is compact in Ω, we have ∇u ∈ L1 (Ω1 ) thus ∇ (u+ ) = H(u)∇u. Since Ω1 is arbitrary,
this shows that ∇u+ in a distributional sense, is in Lp (Ω).
Let j : R → R be such that
√
j (t) = t2 + 2 −  for t > 0 and j (t) = 0 for t 6 0.

It is easy to see that j converges uniformly towards j(t) = t+ and that j0 (t) converges for
all t towards H(t). Let u ∈ L1loc (Ω). Thanks to the Dominated Convergence Theorem,
j (u) converges u+ in L1loc (Ω).
−1/2
Furthermore, ∇ (j (u)) = (u2 + 2 ) (u+ ∇u) converges, for almost every x towards
H (u) ∇u and is dominated by |∇u|. We deduce that in L1 (Ω), there holds

lim ∇ (j (u)) = H (u) ∇u.

→0

Altogether this shows that u+ ∈ W 1,p (Ω) and ∇ (u+ ) = H (u) ∇u. The other results follow
from the fact that u− = (−u)+ .

62