Professional Documents
Culture Documents
FQG Apuntes EDCI
FQG Apuntes EDCI
FERNANDO QUIRÓS
We assume that the student is familiar with the notion of measurable function and
integrable function (we will work always with Lebesgue’s measure). If this is not the case,
you should review this material in a good book on measure theory, for example [3]. You
should at least understand and know how to use the results below.
Definition 1.1. Let Ω ⊂ Rd be an open set. We set
ˆ ˆ
1
L (Ω) = {f : Ω → R; f measurable , |f | < ∞}, kf kL1 (Ω) = kf k1 = |f |.
Ω Ω
Theorem 1.1. L1 (Ω) is a vector space and k · k1 is a norm. The space L1 (Ω) with this
norm is a Banach space.
Theorem 1.2 (Monotone Convergence Theorem, Beppo Levi). Let (fn ) be a se-
quence of functions in L1 (Ω) that satisfy
ˆ
(a) fn ≤ fn+1 a.e. on Ω for all n ∈ N; (b) sup fn < ∞.
n Ω
Then fn converges a.e. on Ω to a finite limit, which we denote by f ; the function f belongs
to L1 (Ω), and kfn − f k1 → 0.
Theorem 1.3 (Dominated Convergence Theorem, Lebesgue). Let (fn ) be a se-
quence of functions in L1 (Ω) that satisfy
(a) fn → f a.e. on Ω;
(b) there is a function g ∈ L1 (Ω) such that for al n, |fn | ≤ g a.e. on Ω.
Then f ∈ L1 (Ω), and kfn − f k1 → 0.
ˆ ˆ
1
Remark. If fn → f in L (Ω), then fn → f (we can pass to the limit inside the
Ω Ω
integral). Indeed,
ˆ ˆ ˆ
fn − f ≤ |fn − f | = kfn − f kL1 (Ω) → 0.
Ω Ω Ω
Lemma 1.4 (Fatou’s lemma). Let (fn ) be a sequence of functions in L1 (Ω) that satisfy
ˆ
(a) for all n, fn ≥ 0 a.e. on Ω for all n ∈ N; (b) sup fn < ∞.
n Ω
1
For a.e. x ∈ Ω we set f (x) = lim inf n→∞ fn (x) ≤ ∞. Then f ∈ L (Ω), and
ˆ ˆ
f ≤ lim inf fn .
Ω n→∞ Ω
Remark. A point x ∈ Ω for which the result holds is called a Lebesgue point for f .
2. A review of Lp spaces
You are expected to be familiar with some basic results on Lp spaces listed below. If
this is not the case, it may be a good idea to read chapters 4 and 5 of the book [1].
Definition 2.1. Let Ω ⊂ Rd be an open set and 1 ≤ p < ∞. We set
ˆ ˆ 1/p
p p p
L (Ω) = {f : Ω → R; f measurable , |f | < ∞}, kf kLp (Ω) = kf kp = |f | .
Ω Ω
d
Definition 2.2. Let Ω ⊂ R be an open set. A measurable function f : Ω → R is
essentially bounded if there is a constant C such that |f | ≤ C a.e. on Ω
Definition 2.3. Let Ω ⊂ Rd be an open set. We set
L∞ (Ω) = {f : Ω → R; f measurable and essentially bounded},
kf kL∞ (Ω) = kf k∞ = inf{C > 0 : |f | ≤ C a.e. on Ω}.
Theorem 2.5. Cc∞ (Ω) is dense in Lp (Ω) for all p ∈ [1, ∞).
Remark. The result is not true for p = ∞.
Theorem 2.6. Let (fn ) be a sequence in Lp , and let f ∈ Lp be such that kfn − f kp → 0.
Then, there exists a subsequence (fnk ) and a function h ∈ Lp such that
fnk → f a.e. on Ω, |fnk | ≤ h a.e. on Ω for all k.
Definition 2.4. Let U and V be open subsets of Rd . We say that U is strongly included
in V , and we write U ⊂⊂ V , if U is compact and U ⊂ V .
Notation. We wil denote the indicator function of a set A by 1A , that is,
(
1, x ∈ A,
1A (x) =
0, x 6∈ A.
Definition 2.5. We say that f ∈ Lploc (Ω), 1 ≤ p ≤ ∞, if f 1Ω0 ∈ Lp (Ω0 ) for all Ω0 ⊂⊂ Ω.
Definition 2.6. Let p, 1 ≤ p ≤ ∞. A sequence (fn ) in Lploc (Ω) converges in Lploc (Ω) to a
function f ∈ Lploc (Ω) if fn 1Ω0 → f 1Ω0 for all Ω0 ⊂⊂ Ω.
Definition 3.1. A function f ∈ L1loc (Ω) is weakly differentiable with respect to xi if there
exists a function gi ∈ L1loc (Ω) such that
ˆ ˆ
f ∂i φ = − gi φ for all φ ∈ Cc∞ (Ω).
Ω Ω
The function gi is called the weak ith partial derivative of f and is denoted by ∂i f .
Thus, for weak derivatives the integration by parts formula
ˆ ˆ
f ∂i φ = − ∂i f φ
Ω Ω
∞
holds by definition for all φ ∈ Cc (Ω).
Weak derivatives are unique. To show this we will need the following standard lemma.
Lemma 3.1. Let u ∈ L1loc (Ω) be such that
ˆ
uφ = 0 for all φ ∈ Cc∞ (Ω).
Ω
Then u = 0 almost everywhere in Ω.
Proof. This is problem 3 in worksheet 2.
Lemma 3.2 (uniqueness of weak derivatives). Let g, h ∈ L1loc (Ω) be weak derivatives
with respect to xi of f ∈ L1loc (Ω). Then g = h almost everywhere in Ω.
Remark. The weak derivative of a continuously differentiable function agrees with the
pointwise derivative. The existence of a weak derivative is, however, not equivalent to the
existence of a pointwise derivative almost everywhere.
Example. Let f ∈ C(R) given by
(
x if x ≥ 0,
f (x) = {x}+ =
0 if x < 0.
Then f is weakly differentiable with weak derivative given by the so-called Heaviside
function,
(1) f 0 (x) = H(x) := 1[0,∞) (x).
The choice of the value of f 0 (x) at x = 0 is irrelevant, since the weak derivative is only
defined up to pointwise almost everywhere equivalence. To prove (1), note that for any
φ ∈ Cc∞ (R) an integration by parts gives
ˆ ˆ ∞ ˆ ∞ ˆ
0 0
fφ = xφ = − φ = − Hφ.
R 0 0 R
Example. The Heaviside function is not weakly differentiable. Indeed, if it had a weak
derivative H 0 ∈ L1loc (R), then
ˆ ∞ ˆ ∞ ˆ ∞
0 0
Hφ=− Hφ = − φ0 = φ(0) for all φ ∈ Cc∞ (R).
−∞ −∞ 0
Proof. Let φ ∈ Cc∞ (Ω) be any test function. Then ψφ ∈ Cc∞ (Ω) and the weak differentia-
bility of f implies that ˆ ˆ
f ∂i (ψφ) = − (∂i f )ψφ.
Ω Ω
Expanding ∂i (ψφ) = ψ∂i (φ) + φ∂i (ψ), and rearranging the result we get
ˆ ˆ
ψf (∂i φ) = − f (∂i ψ) + ψ(∂i f ) φ.
Ω Ω
Definition 3.2 (Sobolev space). Let Ω ⊂ Rd be an open set and 1 ≤ p ≤ ∞. The
Sobolev space W 1,p (Ω) consists of all functions f ∈ Lp (Ω) such that for all i ∈ {1, 2, · · · , d}
the weak derivative ∂i f exists and belongs to Lp (Ω).
Proposition 3.4. The Sobolev space W 1,p (Ω) is a Banach space when equipped with the
norm
d
!1/p
p
X p
kf kW 1,p (Ω) = kf kLp (Ω) + k∂i f kLp (Ω) for p ∈ [1, ∞),
i=1
d
X
kf kW 1,∞ (Ω) = kf kL∞ (Ω) + k∂i f kL∞ (Ω) .
i=1
The Sobolev space H 1 (Ω) is a Hilbert space when equipped with the scalar product
d
X
(f, g)H 1 (Ω) = (f, g)L2 (Ω) + (∂i f, ∂i g)L2 (Ω) .
i=1
Proof. Proving that k · kW 1,p (Ω) are norms and that (·, ·)H 1 (Ω) is a scalar product is an easy
task that I leave as an exercise.
Let (fn ) be a Cauchy sequence in W 1,p (Ω). Then the sequences (fn ), (∂i fn ), i = 1, . . . , d,
are Cauchy sequences in Lp (Ω), which is a Banach space. Let us denote the corresponding
limits by f , gi , i = 1, . . . , d. In order to finish the proof it is enough to check that gi = ∂i f ,
i = 1, . . . , d.
Using Hölder’s inequality we get
ˆ ˆ ˆ
fn ∂i φ − f ∂i φ ≤ |fn − f ||∂i φ| ≤ kfn − f kp k∂i φkp0 → 0,
ˆΩ ˆΩ ˆ Ω
∂i fn φ − gi φ ≤ |∂i fn − gi ||φ| ≤ k∂i fn − gi kp kφkp0 → 0,
Ω Ω Ω
´ ´ ´ ´
which means that Ω fn ∂i φ → Ω f ∂i φ, Ω ∂i fn φ → Ω gi φ for all φ ∈ Cc∞ (Ω). Therefore,
since ˆ ˆ
fn ∂i φ = − ∂i fn φ,
Ω Ω
passing to the limit we get. ˆ ˆ
f ∂i φ = − gi φ,
Ω Ω
that is, ∂i f = gi .
Our goal is to approximate “bad” functions by good (smooth) ones. The idea is to
average, giving increasing importance to nearby points and decreasing importance to the
rest.
Definition 4.1. A mollifying family is any family of functions {ρδ }δ>0 such that
ˆ
∞ d
ρδ ∈ Cc (R ), supp ρδ ⊂ B(0, δ), ρδ = 1, ρδ ≥ 0.
Rd
6 F. QUIRÓS
It is easy to generate a family of mollifiers starting from a function ρ ∈ Cc∞ (Rd ) such
that supp ρ ⊂ B(0, 1), ρ ≥ 0, ρ 6≡ 0, for instance
( 2
e1/(|x| −1) si |x| < 1,
ρ(x) =
0 si |x| > 1.
´
To this aim just define ρδ (x) = Cδ −d ρ(x/δ), with C = 1/ Rd ρ.
Remark. The mollifying family with ρ as above is known as the standard mollifier.
Definition 4.2. Given δ > 0, let Ωδ = {x ∈ Ω : dist(x, ∂Ω) > δ}. The mollification of
f ∈ L1loc (Ω) is the function f δ : Ωδ → R given by
ˆ
δ
(2) f (x) = ρδ (x − y)f (y) dy .
Ω
| {z }
(ρδ ?f )(x)
Proof. (a) Let x ∈ Ωδ , i ∈ {1, ..., d}. Let h ∈ R be small enough, so that x + hei ∈ Ωδ .
There exists Ω0 ⊂⊂ Ω such that
ρδ (x + hei − y) − ρδ (x − y)
− ∂i ρδ (x − y) = 0 if y 6∈ Ω0 .
h
On the other hand, using Taylor’s expansion we get
ρδ (x + hei − y) − ρδ (x − y)
− ∂i ρδ (x − y) ≤ C|h| if y ∈ Ω0 .
h
We conclude that
δ
f (x + hei ) − f δ (x)
− (∂i ρδ ? f )(x)
h
ˆ
ρ δ (x + he i − y) − ρ δ (x − y)
= − ∂i ρδ (x − y) f (y) dy ≤ C|h|kf kL1 (Ω0 ) ,
Ω h
which implies that ∂i f δ (x) exists and is equal to (∂i ρδ ? f )(x).
The proof for higher derivatives proceeds by induction.
for all x ∈ K.
(b) Let x ∈ Ωδ . Then, applying Hölder’s inequality,
ˆ ˆ
δ
0
|f (x)| = ρδ (x − y)f (y)dy ≤ ρδ (x − y)1/p ρδ (x − y)1/p |f (y)| dy
Bδ (x) Bδ (x)
ˆ 1/p0 ˆ 1/p
≤ ρδ (x − y) dy ρδ (x − y)|f (y)|p dy .
Bδ (x) Bδ (x)
Now, thanks to (b) we know that f δ → f in Lploc (Ω). On the other hand, using this
result to ∂i f , which belongs to Lp (Ω), since f ∈ W 1,p (Ω), we get that
∂i f δ = (∂i f )δ → ∂i f in Lploc (Ω).
Partitions of unity allow us to piece together global results from local results.
Theorem 4.3 (partitions of unity). Let O be a family of open subsets of Rd . There is
a collection Φ of C ∞ functions ϕ defined in Ω = ∪ω∈O ω with the following properties:
8 F. QUIRÓS
(a) 0 ≤ ϕ ≤ 1 in Ω;
(b) for each compact subset K of Ω, the set {ϕ ∈ Φ : supp ϕ ∩ K 6= ∅} is finite;
P
(c) ϕ∈Φ ϕ = 1 on Ω (by (b), this sum is finite in every compact subset of Ω);
(d) for each ϕ ∈ Φ there is an open set U ∈ O such that supp ϕ ⊂ U .
Proof. I am taking the proof of this nice theorem from the book [4].
Take a family {Ωi }∞
i=1 of open subsets of Ω such that Ωi ⊂⊂ Ωi+1 and
Ω = ∪∞
i=1 Ωi .
We consider a smooth partition of unity F subordinated to the open cover {Ωi+1 \Ωi−1 }∞ i=1 ,
where Ω0 = ∅. For each i ∈ N let ψi be the sum of all the finitely many ψ ∈ F such
that supp ψ ⊂ Vi := Ωi+1 \ Ωi−1 and such that P they have not already been selected at
previous steps j < i. Then ψi ∈ Cc∞ (Vi ) and ∞ ψ
i=1 i = 1 in Ω. Hence, ψ i u ∈ W 1,p
(Ω)
and supp (ψi u) ⊂ Vi .
Let ε > 0. For all i ∈ N there is a value δi > 0 such that the mollification ui = ρδi ?(ψi u)
satisfies that supp ui ⊂ Vi and
ε
kui − ψi ukW 1,p (Vi ) ≤ i+1 .
2
P∞
We define v := i=0 ui ∈ C ∞ (Ω) (in each Ω0 ⊂⊂ Ω Pthe sum has only a finite number
∞
of terms different from 0). Let us also note that u = i=1 (ψi u).
For x ∈ Ω` ,
`
X `
X
u(x) = (ψi u)(x), v(x) = ui (x)
i=1 i=0
Therefore,
` ` ` ∞
X X
X X ε
kv − ukW 1,p (Ω` ) =
ui − ψi u
≤ kui − ψi ukW 1,p (Vi ) ≤ = ε.
2i+1
i=0 i=0 W 1,p (Ω` ) i=0 i=0
Letting ` → ∞, it follows from the monotone convergence theorem that kv−ukW 1,p (Ω) ≤ ε.
This also implies that u − v (and, in turn, v) belongs to the space W 1,p (Ω).
Corollary 4.5 (density of Cc∞ (Rd ) in W 1,p (Rd )). Let 1 ≤ p < ∞. Then Cc∞ (Rd ) is
dense in W 1,p (Rd ).
Proof. Given u ∈ W 1,p (Rd ) and ε > 0, we consider a function v ∈ C ∞ (Rd )∩W 1,p (Rd ) such
that ku − vkW 1,p (Rd ) ≤ ε/2. The existence of such function is guaranteed by Meyers-Serrin
theorem.
Let ϕ ∈ Cc∞ (Rd ) be such that 0 ≤ ϕ ≤ 1, ϕ(x) = 1 if |x| ≤ 1, ϕ(x) = 0 if |x| ≥ 2.
We define ϕR (x) = ϕ(x/R) (which is knwon as a cut-off function). We claim that
kϕR v − vkW 1,p (Rd ) ≤ ε/2 if R is large enough, from where the result follows immediately,
since ϕR v ∈ Cc∞ (Rd ).
SOBOLEV SPACES 9
Let now Ω ⊂ Rd be a bounded open set. Can we approximate a function in W 1,p (Ω)
by a function in C ∞ (Ω), instead of only by functions in C ∞ (Ω) ∩ W 1,p (Ω)? This is indeed
the case if ∂Ω is smooth.
Definition 4.4. Let Ω ⊂ Rd be a bounded open set. We say that its boundary, ∂Ω, is
C 1 if for all x0 ∈ ∂Ω there is a radius r > 0 and a C 1 function γ : Rd−1 → R such that
(relabeling and reorienting the axes if needed),
Ω ∩ Br (x0 ) = {x ∈ Br (x0 ) : xd > γ(x1 , · · · , xd−1 )}.
Theorem 4.6 (global approximation by functions smooth up to the boundary).
Let Ω ⊂ Rd be a bounded open set with C 1 boundary ∂Ω and let 1 ≤ p < ∞. Then C ∞ (Ω)
is dense in W 1,p (Ω).
Proof. Let x0 ∈ ∂Ω. Let r > 0 and a C 1 function γ : Rd−1 → R such that (relabeling and
reorienting the axes if needed),
Ω ∩ Br (x0 ) = {x ∈ Br (x0 ) : xd > γ(x1 , · · · , xd−1 )}.
Set V := Ω ∩ B r2 (x0 ). Given x ∈ V and δ > 0, we define the shifted point
xδ := x + λδen .
Observe that for some fixed, sufficiently large number λ > 0 the ball Bδ (xδ ) lies in
Ω ∩ Br (x0 ) for all x ∈ V and all small δ > 0.
For x ∈ V we define uδ (x) := u(xδ ). This is the function u translated a distance λδ in
the en direction. Next write vδ = ρδ ? uδ . The idea is that we have moved up enough so
that theire is room to mollify within Ω. Clearly vδ ∈ C ∞ (V ).
We now claim that vδ → u in W 1,p (V ). To confirm this, observe that
kvδ − ukLp (V ) ≤ kvδ − uδ kLp (V ) + kuδ − ukLp (V ) ,
k∂i vδ − ∂i ukLp (V ) ≤ k∂i vδ − ∂i uδ kLp (V ) + k∂i uδ − ∂i ukLp (V ) , i = 1, . . . , d.
Reasoning like in the proof of Theorem 4.2-(c) we can check that the first term on the
right-hand side of each line goes to 0 with δ. The second term also vanishes in the limit,
since translations are continuous in the Lp -norm (see problem 4 in worksheet 2), and the
claim is proved.
10 F. QUIRÓS
Choose ε > 0. Since ∂Ω is compact, we can find finitely many points x0j ∈ ∂Ω, radii
rj > 0, corresponding sets Vj = Ω ∩ Oj where Oj := B rj (x0j ), and functions vj ∈ C ∞ (Vj ),
2
j = 1, . . . , M , such that ∂Ω ⊂ ∪M
j=1 Oj , and
We consider a smooth partition of unity F subordinated to the open cover {Oj }Mj=0 . Let
ψj be the sum of all the finitely many ψ ∈ F such that supp ψ ∈ Oj , supp ψ ∩ Ω 6= ∅, and
such that they have not already been selected at previous steps k < j. Then ψj ∈ Cc∞ (Vj )
and ∞
P PM ∞
j=1 ψj = 1 in Ω. We define v := j=0 ψj vj . Then clearly v ∈ C (Ω). In addition,
PM
since u = j=0 ψj u in Ω, we finally obtain
M
X
kv − ukW 1,p (Ω) ≤ kψj (vj − u)kW 1,p (Oj ∩Ω) .
j=1
But kψj (vj − u)kLp (Oj ∩Ω) ≤ Ckvj − ukLp (Oj ∩Ω) , and
k∂i (ψj (vj − u))kLp (Oj ∩Ω) ≤ k(vj − u)∂i ψj kLp (Oj ∩Ω) + kψj ∂i (vj − u)kLp (Oj ∩Ω)
≤ C(kvj − ukLp (Oj ∩Ω) + k∂i (vj − u)kLp (Oj ∩Ω) ,
and we conclude that
M
X
kv − ukW 1,p (Ω) ≤ C kvj − ukW 1,p (Oj ∩Ω) ≤ C(M + 1)ε.
j=0
5. Extensions
Our next goal is to extend functions in the Sobolev space W 1,p (Ω) to become functions
in the Sobolev space W 1,p (Rd ).
Theorem 5.1 (Extension theorem). Let Ω ⊂ Rd be an open set with C 1 boundary and
1 ≤ p ≤ ∞. Let V be a bounded open set such that Ω ⊂⊂ V . Then there exists a bounded
linear operator
E : W 1,p (Ω) → W 1,p (Rd )
such that for all u ∈ W 1,p (Ω):
(a) Eu = u a.e. in Ω;
(b) Eu has support within V;
(c) kEukW 1,p (Rd ) ≤ CkukW 1,p (Ω) , the constant C depending only on p, Ω and V .
We call Eu an extension of u to Rd .
Proof. I will give a proof that uses global approximations by functions that are smooth up
to the boundary, Theorem 4.6, which hence is only valid for 1 ≤ p < ∞. I took it from [2].
You can find a (not so) different proof that is also valid for p = ∞ in [1, Theorem 9.7].
SOBOLEV SPACES 11
Fix x0 ∈ ∂Ω and suppose first that ∂Ω is flat near x0 , lying in the plane {xd = 0}.
Then we may assume that there exists an open ball B = Br (x0 ) for some radius r > 0,
such that
B + := B ∩ {xd ≥ 0} ⊂ Ω, B − := B ∩ {xd ≤ 0} ⊂ Rd \ Ω.
Let us also assume temporarily that u ∈ C 1 (Ω). Then we define
(
u(x) if x ∈ B + ,
ū(x) :=
−3u(x1 , . . . , xd−1 , −xd ) + 4u(x1 , . . . , xd−1 , −xd /2) if x ∈ B − .
This is called a higher-order reflection of u from B + to B − .
We claim that ū ∈ C 1 (B). We only have to check what happens at the plane {xd = 0},
since smoothness in the rest of the domain under consideration is obvious.
Let us write u− := uB − , u+ := uB + . A straightforward computation shows that
∂d u− (x) = 3∂d u(x1 , . . . , xd−1 , −xd ) − 2∂d u(x1 , . . . , xd−1 , −xd /2),
and hence
∂d u− x
= ∂d u+ x .
d =0 d =0
On the other hand, it is trivial to check that u− = u+ on {xd = 0}. This implies moreover
that
∂i u− x =0 = ∂i u+ x =0 , i = 1, . . . , d − 1,
d d
and the claim is proved.
The above computations yield easily (check it!) that
kūkW 1,p (B) ≤ CkukW 1,p (B + )
for some constant C that does not depend on u.
If ∂Ω is not flat near x0 , we straighten the boundary near x0 . To be more specific, we
set
yi = xi =: Φi (x), i =, . . . , d − 1, yd = xd − γ(x1 , . . . , xd−1 ) =: Φd (x),
where γ is the C function describing the boundary in a neighbourhood of x0 , and write
1
M
X
(3) kūkW 1,p (Rd ) ≤ kψj ūj kW 1,p (Wj ) ≤ CkukW 1,p (Ω)
j=0
for some constant C that does not depend on u. We write Eu := ū and observe that
the mapping u → Eu is linear.
We finally get rid of the assumption u ∈ C 1 (Ω) using the global approximations smooth
up to the boundary provided by Theorem 4.6. It is here that we are using that p is finite.
Let u ∈ W 1,p (Ω). We choose (uk ) ⊂ C ∞ (Ω) converging to u in W 1,p (Ω). Using (3) and
the linearity of E we get
Therefore, (Euk ) is a Cauchy sequence in W 1,p (Rd ) and so converges to ū =: Eu. This
extension, which does not depend on the particular choice of the approximating sequence
(check it!), satisfies the conclusions of the theorem.
6. Traces
Theorem 6.1 (Trace theorem). Let 1 ≤ p < ∞, and assume that Ω ⊂ Rd is a bounded
open set with smooth boundary. Then there exists a bounded linear operator
such that
(a) T u = u∂Ω if u ∈ W 1,p (Ω) ∩ C(Ω),
(b) kT ukLp (∂Ω) ≤ CkukW 1,p (Ω) for all u ∈ W 1,p (Ω),
Proof. We assume first that u ∈ C 1 (Ω), the general case following by approximation.
We also assume that the boundary is flat near x0 ∈ ∂Ω, lying in the plane {xd = 0}.
We choose an open ball B as in the proof of the previous theorem, and we denote by B̂
the concentric ball with radius r/2.
Select ζ ∈ Cc∞ (B) with ζ ≥ 0 in B, ζ ≡ 1 on B̂ (such a function is known as a cut-off
function). Denote by Γ the portion of ∂Ω within B̂. Set x0 = (x1 , . . . , xd−1 ) ∈ Rd−1 =
SOBOLEV SPACES 13
{xd = 0}. Then, applying the Fundamental Theorem of Calculus and Young’s inequality1
ˆ ˆ ˆ
p 0 p 0
|u| dx ≤ ζ|u| dx = − ∂d (ζ|u|p ) dx
Γ ˆd =0}
{x B +
ˆ
p p−1
= − |u| ∂d ζ + ζp|u| (sign u)∂d u dx ≤ C |u|p + |∂d u|p dx,
B+ B+
which yields
(4) kukLp (Γ) ≤ CkukW 1,p (Ω) .
In order to compute the weak derivative of |u|p we have used problems 3 and 4 in worksheet
3.
If ∂Ω is not flat around x0 ∈ ∂Ω, we straighten out the boundary, do the above compu-
tation, and then undo the change of variables, in the same fashion as in the proof of the
previous theorem, to obtain again (4), where now Γ is some open subset of ∂Ω containing
x0 .
Since ∂Ω is compact, there exist finitely many point x0j ∈ ∂Ω and open subsets Γj ⊂ ∂Ω
containing x0j , j = 1, . . . , M , such that ∂Ω = ∪M
j=1 Γj and
T u := lim T uk ,
k→∞
p
the limit taken in L (∂Ω). Acccording to (6) this definition does not depend on the
particular choice of smooth functions approximating u.
Finally, if u ∈ W 1,p (Ω) ∩ C(Ω), we note that the functions of the approximation
of u
constructed in Theorem 4.6 converge uniformly to u on Ω. Hence T u = u ∂Ω .
Definition 6.1. ( The space W01,p (Ω) is the closure of Cc∞ (Ω) in W 1,p (Ω). It is customary
to write H01 (Ω) to denote W01,2 (Ω).
Remark. Corollary 4.5 implies that W01,p (Rd ) = W 1,p (Rd ) if 1 ≤ p < ∞.
Theorem 6.2 (Trace-zero functions in W 1,p (Ω)). Let 1 ≤ p < ∞. Assume that Ω is a
bounded open subset of Rd with smooth boundary. Let u ∈ W 1,p (Ω). Then,
u ∈ W01,p (Ω) if and only if T u = 0 on ∂Ω.
Proof. We prove that u ∈ W01,p (Ω) ⇒ T u = 0 on ∂Ω. A proof of the other implication
can be found for example in [2, Section 5.5].
0
1ab ap bp
≤ p + p0 , a, b > 0, 1 < p, p0 < ∞, 1
p + 1
p0 = 1.
14 F. QUIRÓS
If u ∈ W01,p (Ω), by definition, there exist functions uk ∈ Cc∞ (Ω) such that uk converges
to u in W 1,p (Ω). Since T uk = 0 on ∂Ω for all k, and T : W 1,p (Ω) → Lp (∂Ω) is a bounded
linear operator, hence continuous, we deduce that T u = 0 on ∂Ω.
Thus, asking a function to belong to some space W01,p (Ω) is a reasonable (weak) way
to ask it to be zero at ∂Ω.
7. Sobolev embeddings
We often think of i as identifying elements of the smaller space X with elements of the
larger space Y ; if X is a subset of Y then i is the inclusion map. The boundedness of i
means that ther is a constant C such that kixkY ≤ CkxkX for all x ∈ X, so the weaker
Y -norm of ix is controlled by the stronger X-norm of X.
We write an embedding as X ,− → Y , or as X ⊂ Y when the boundedness is understood.
We wonder whether W 1,p (Rd ) is continuously embedded in some other spaces. The
answer will be “yes”, but in wich other spaces depends on the relation between p and d.
Proof. Step 1. We first prove the estimate for the case p = 1. The general case will
follow from it.
Since u has compact support, for each i ∈ {1, . . . , d} and x ∈ Rd we have
ˆ xi
u(x) = ∂i u(x1 , . . . , xi−1 , yi , xi+1 , . . . , xd ) dyi ,
−∞
and so
ˆ ∞
|u(x)| ≤ |Du(x1 , . . . , xi−1 , yi , xi+1 , . . . , xd )| dyi , i ∈ {1, . . . , d}.
−∞
Consequently
d ˆ ∞
1
d−1
d Y
|u(x)| d−1 ≤ |Du(x1 , . . . , xi−1 , yi , xi+1 , . . . , xd )| dyi .
i=1 −∞
Integrating this inequality with respect to x1 and using the general Hölder’s inequality2
with pi = d − 1, i = 1, . . . , d − 1, we obtain
ˆ ∞ ˆ ∞Y d ˆ ∞ 1
d−1
d
|u(x)| d−1 dx1 ≤ |Du| dyi dx1
−∞ −∞ i=1 −∞
ˆ∞
1 ˆ
d−1 ∞ d ˆ
Y ∞
1
d−1
= |Du| dy1 |Du| dyi dx1
−∞ −∞ i=2 −∞
ˆ ∞
1
d−1 Yd ˆ ∞ ˆ ∞
1
! d−1
= |Du| dy1 |Du| dx1 dyi .
−∞ i=2 −∞ −∞
where ˆ ∞ ˆ ∞ ˆ ∞
I1 := |Du| dy1 , Ii := |Du| dx1 dyi , i = 3, . . . , d.
−∞ −∞ −∞
2Given 1 < p1 , p2 , . . . , pn < ∞ with p11 + p12 + · · · + p1n = 1 and fi ∈ Lpi (Ω), i = 1, 2, . . . , n,
ˆ
f1 f2 . . . fn ≤ kf1 kp1 kf2 kp2 . . . kfn kpn .
Ω
16 F. QUIRÓS
= |Du| dx ,
Rd
Remark. We really do need u to have compact support for (9) to hold, as the example
u ≡ 1 shows. But remarkably the constant C does not depend at all upon the size of the
support of u.
Corollary 7.2. Let p ∈ [1, d) and p∗ the Sobolev conjugate of p. The Gagliardo-Nirenberg-
Sobolev inequality (10) holds for all u ∈ W 1,p (Rd ). Moreover, for every q ∈ [p, p∗ ] there
is a constant C = C(d, p, q) such that
(13) kukLq (Rd ) ≤ CkukW 1,p (Rd ) for all u ∈ W 1,p (Rd ),
so that W 1,p (Rd ) ,−
→ Lq (Rd ).
Corollary 7.3. Let p ∈ [1, d). Let Ω be a bounded, open subset of Rd with C 1 boundary.
Then there is a constant C = C(d, p, Ω) such that
(14) kukLp∗ (Ω) ≤ CkukW 1,p (Ω) for all u ∈ W 1,p (Ω).
Proof. Under our assumptions, there exists an extension Eu = ū ∈ W 1,p (Rd ) such that
ū = u in Ω, ū has compact support, and kūkW 1,p (Rd ) ≤ CkukW 1,p (Ω) .
Let (uk )∞ ∞ d
k=1 be a sequence of functions in Cc (R ) such that uk → ū in W
1,p
(Rd ). Accord-
ing to inequality (10),
kuk − uj kLp∗ (Rd ) ≤ CkDuk − Duj kLp (Rd ) for all k, j ≥ 1.
∗
Thus, uk → ū in Lp (Rd ) as well. Since we also have kuk kLp∗ (Rd ) ≤ CkDuk kLp (Rd ) , passing
to the limit we obtain the desired bound (14).
Corollary 7.4. Let p ∈ [1, d) and q ∈ [1, p∗ ]. Let Ω be a bounded, open subset of Rd with
C 1 boundary. Then there is a constant C = C(d, p, q, Ω) such that
kukLq (Ω) ≤ CkukW 1,p (Ω) for all u ∈ W 1,p (Ω).
Taking γ = d we obtain
kuk d2 ≤ CkukW 1,d (Rd ) ,
d−1
which implies, by interpolation, the desired inclusion for all q ∈ [d, d2 /(d − 1)].
Take now γ = d + 1. We obtain, using also the previous step, the estimate
kuk d(d+1) ≤ C(kuk d2 + kDukd ) ≤ CkukW 1,d (Rd ) ,
d−1 d−1
which implies, by interpolation, the desired inclusion for all q ∈ [d, d(d + 1)/(d − 1)].
Taking γ = d + 2, γ = d + 3, . . . , we eventually reach any finite exponent q > d.
3If f ∈ Lp (Ω) ∩ Lr (Ω), p < r, then f ∈ Lq (Ω) for all q ∈ [p, r], and kf kq ≤ kf kθp kf k1−θ , where θ ∈ [0, 1]
r
1 θ 1−θ
is given by q = p + r .
18 F. QUIRÓS
7.3. Morrey’s inequality. Let now p ∈ (d, ∞). We will show that if u ∈ W 1,p then u is
in fact Hölder continuous, after possibly being redefined on a set of measure zero.
Let us start by explaining what a Hölder space is. Assume Ω ⊂ Rd is open. Let us
recall that a function u : Ω → R is said to be Lipschitz continuous in Ω if there exists
some constant C such that
|u(x) − u(y)| ≤ C|x − y| for all x, y ∈ Ω.
This estimate of course implies that u is continuous. More importantly, it allows to quan-
tify how close should be x and y in order for u(y) to be close to u(x). Such quantitative
estimates are important in the analysis of PDEs. Hölder continuous functions satisfy a
variant of the above estimate.
Definition 7.3. A function u : Ω → R is said to be Hölder continuous in Ω with exponent
γ ∈ (0, 1] if there exists some constant C such that
|u(x) − u(y)| ≤ C|x − y|γ for all x, y ∈ Ω.
Definition 7.4. (i) If u : Ω → R is bounded and continuous we write
kukC(Ω) := sup |u(x)|.
x∈Ω
Definition 7.5. The Hölder space C k,γ (Ω) consists of all functions u ∈ C k (Ω) for which
the norm X X
kDα ukC(Ω) + Dα u C (0,γ (Ω)
kukC k,γ (Ω) :=
|α|≤k |α|=k
is finite
So the space C k,γ (Ω) consists of those functions u that are k times continuously dif-
ferentiable and whose k th -partial derivatives are bounded and Hölder continuous with
exponent γ. Such functions are well behaved, and the space C k,γ (Ω) itself posesses a good
mathematical structure.
Theorem 7.6 (Hölder spaces as function spaces). The space of functions C k,γ (Ω)
is a Banach space.
Proof. Step 1. We claim that there exists a constant C, depending only on d, such that
ˆ
|Du(y)|
(15) |u(y) − u(x)| dy ≤ C d−1
dy.
Br (x) Br (x) |y − x|
To prove the claim, fix any point w ∈ ∂B1 (0). Then, if 0 < s < r,
ˆ s ˆ s ˆ s
d
|u(x+sw)−u(x)| = u(x + tw) dt = Du(x + tw) · w dt ≤ |Du(x + tw)| dt.
0 dt 0 0
Hence, integrating in ∂B1 (0),
ˆ ˆ sˆ
|u(x + sw) − u(x)| dS(w) ≤ |Du(x + tw)| dS(w)dt
∂B1 (0) ˆ sˆ
0 ∂B 1 (0)
|Du(y)|
= d−1
dS(y)dt
ˆ0 ∂Bt (x) t
|Du(y)|
= d−1
dy
ˆB s (x) |x − y|
|Du(y)|
≤ d−1
dy,
Br (x) |x − y|
Proof. We give the proof when p ∈ (d, ∞). The case p = ∞ is left as an exercise; see
problem 2 in worksheet 5.
Thanks to the properties of Ω, we know that there exists an extension Eu = ū ∈
W 1,p (Rd ) such that
ū = u in Ω, ū has compact support, kūkW 1,p (Rd ) ≤ CkukW 1,p (Ω) .
Meyer-Serrin’s theorem guaranties the existence of a sequence of functions uk ∈ Cc∞ (Rd )
such that uk → ū in W 1,p (Rd ). Now, thanks to Morrey’s inequality, we have
kuk − uj kC 0,γ (Rd ) ≤ Ckuk − uj kW 1,p (Rd ) for all k, j ≥ 1
whence there exists a function u∗ ∈ C 0,γ (Rd ) such that uk → u∗ in C 0,γ (R). The con-
vergence of uk in Lp (Rd ) implies that along some subsequence ukj → ū a.e. Hence,
u∗ = ū a.e., and u∗ = u a.e. in Ω, so that u∗ is a version of u. Morrey’s inequality
also implies kuk kC 0,γ (Rd ) ≤ Ckuk kW 1,p (Rd ). Passing to the limit we obtain kūkC 0,γ (Rd ) ≤
CkūkW 1,p (Rd ), from where the desired inequality follows easily.
Theorem 7.9. Let p ∈ (d, ∞) and γ = 1 − dp . There is a constant C = C(d, p) such that
every u ∈ W 1,p (Rd ) has a version u∗ ∈ C 0,γ (Ω) satisfying
ku∗ kC 0,γ (Rd ) ≤ CkukW 1,p (Rd ) .
SOBOLEV SPACES 21
Remark. If u ∈ W 1,p (Rd ), p ∈ (d, ∞), then lim|x|→∞ u(x) = 0. The proof is left as an
exercise for the student (Hint: Approximate u by functions in Cc∞ (Rd ) and use Morrey’s
inequality to pass to the limit in C 0,γ (Rd )). The result is not true for p = ∞.
8. Compact embeddings
Since
kukj − ukl kLq (V ) ≤ kukj − uδkj kLq (V ) + kuδkj − uδkl kLq (V ) + kuδkl − ukl kLq (V ) ,
we get the claim.
Step 6. The proof will now follow employing Step 5 with ε = 1, 21 , 13 , 14 , . . . plus a
standard diagonal argument. Indeed, we start with ε = 1 and consider a subsequence of
(1) ∞
(uk )∞
k=1 , that we denote by (uk )k=1 , such that
(1) (1)
lim sup kuk − uj kLq (V ) ≤ 1.
k,j→∞
(1) (2)
Next we consider a subsequence of (uk )∞ ∞
k=1 , that we denote by (uk )k=1 , such that
(2) (2) 1
lim sup kuk − uj kLq (V ) ≤ .
k,j→∞ 2
(n−1) ∞
We iterate this argument, considering a subsequence of (uk )k=1 , that we denote by
(n)
(uk )∞k=1 , such that
(n−1) (n−1) 1
lim sup kuk − uj kLq (V ) ≤ .
k,j→∞ n
(k)
We finally consider the subsequence of the original sequence given by (uk )∞
k=1 , which is
a Cauchy sequence in Lq (V ), and hence convergent.
Corollary 8.2. Let Ω be a bounded open subset of Rd with C 1 boundary. Then
W 1,d (Ω) ⊂⊂ Lq (Ω) for each q ∈ [1, ∞).
q
Proof. Given q ∈ [1, ∞), take p ∈ [1, d) such that p∗ > q (any p ∈ ( d+q , d) will do the job).
Since Ω is bounded, a bounded sequence in W (Ω) will also be bounded in W 1,p (Ω), and
1,d
because of the compact embedding W 1,p (Ω) ⊂⊂ Lr (Ω) for all r ∈ [1, p∗ ), it will have a
convergent subsequence in Lq (Ω), since q ∈ [1, p∗ ).
Lemma 8.3. Let Ω be a bounded, open set in Rd , and α ∈ (0, 1]. Then
(a) C 0,α (Ω) ⊂⊂ C 0,β (Ω) for all β ∈ (0, α),
(b) C 0,α (Ω) ⊂⊂ Lq (Ω) for all q ∈ [1, ∞].
Let (uk )∞
k=1 be a bounded sequence in C
0,α
(Ω). It is obviously uniformly bounded and
equicontinuous in Ω. Hence, thanks to Arzelà-Ascoli compactness criterion we know that
it has a subsequence (ukj )∞j=1 that converges uniformly in Ω. Using the above estimate
and the boundedness of the sequence we get, for some constant C,
kukj − ukl kC 0,β (Ω) = kukj − ukl kL∞ (Ω) + ukj − ukl C 0,β (Ω)
αβ 1− αβ
≤ kukj − ukl kL∞ (Ω) + ukj − ukl C 0,α (Ω) 2kukj − ukl kL∞ (Ω)
1− αβ
≤ kukj − ukl kL∞ (Ω) + C kukj − ukl kL∞ (Ω) .
Corollary 8.4. Let Ω be a bounded open subset of Rd with C 1 boundary, and p ∈ (d, ∞].
Then
1,p 0,β d
W (Ω) ⊂⊂ C (Ω) for each β ∈ 0, 1 − .
p
As a consequence, W 1,p (Ω) ⊂⊂ Lq (Ω) for all q ∈ [1, ∞].
d
→ C 0,1− p (Ω), a bounded sequence in W 1,p (Ω) is
Proof. Since in this range W 1,p (Ω) ,−
d
also boundedin C 0,1− p (Ω). But the latter space is compactly embedded in C 0,β (Ω) if
β ∈ 0, 1 − dp , and the result follows.
9. Poincaré inequalities
The result for q = ∞ follows after passing to the limit (recall that bounded sequences in
W 1,p (Ω), p ∈ (d, ∞], have a subsequence that converges uniformly). This then yields the
result for finite q by means of Hölder’s inequality.
Remark. In view of Poincaré-Friedrichs’ inequality, on W01,p (Ω) the norm kDukLp (Ω) is
equivalent to kukW 1,p (Ω) if Ω is bounded and smooth.
The compactness results in the previous section can be used to generate new inequali-
ties.
Theorem 9.2 (Poincaré-Wirtinger’s inequality). Let Ω be a bounded, connected,
open subset with C 1 boundary. Assume 1 ≤ p ≤ ∞. Then there exists a constant C,
depending only on d, p and Ω, such that
1,p
u − u
p ≤ CkDukLp (Ω) for all u ∈ W (Ω).
Ω L (Ω)
Proof. We argue by contradiction. Were the stated estimate false, there would exist for
each integer k = 1, . . . a function uk ∈ W 1,p (Ω) satisfying
uk − uk
(21)
p > kkDuk kLp (Ω) .
Ω L (Ω)
26 F. QUIRÓS
We normalize by defining
ffl
uk − Ω uk
vk :=
ffl
uk − uk
p , k = 1, . . . .
Ω L (Ω)
ffl
Then, Ω vk = 0, kvk kLp (Ω) = 1, and (21) implies kDvk kLp (Ω) < 1/k, k = 1, 2, . . . . In par-
ticular, the sequence (vk )∞
k=1 is bounded in W
1,p
(Ω). Therefore, thanks to the compactness
results of the previous section, we know that there exists a subsequence (vkj )∞ ∞
j=1 ⊂ (vk )k=1
p p
and a function v ∈ L (Ω) such that vkj → v in L (Ω). It is then immediate to chek (do
it!) that
v = 0, kvkLp (Ω) = 0.
Ω
Thanks to our results on traces, we know that W0k,p (Ω) consists of those functions in
W k,p (Ω) such that Dα u = 0 on ∂Ω for all |α| ≤ k − 1 in the sense of traces.
We also have several embeddings, which follow in a straightforward manner from the
ones for the first order derivatives.
1 1 k
Theorem 10.2. Let p ∈ [1, ∞), pk < d. Then, W k,p (Rd ) ,−
→ Lq (Rd ), where = − .
q p d
Remarks. (a) Linear operators are bounded if and only if they are continuous. Thus,
X ∗ is the collection of all continuous linear functionals on X. It is sometimes denoted as
the topological dual, to distinguish it from the algebraic dual.
(b) The dual of X is also denoted as X 0 .
Definition 11.2. (i) If u ∈ X, F ∈ X ∗ , we write hF, ui to denote the real number F (u).
The symbol h·, ·i denotes the pairing of X ∗ and X.
(ii) We define a norm in X ∗ by
kF kX ∗ := sup{hF, ui : kuk ≤ 1}.
(iii) A Banach space X is reflexive if (X ∗ )∗ = X. More precisely, this means that for each
u∗∗ ∈ (X ∗ )∗ there exists u ∈ X such that
hu∗∗ , u∗ i = hu∗ , ui for all u∗ ∈ X ∗ .
In a Hilbert space H it is very easy to write down continuous linear functionals. Pick
any f ∈ H. Then the map u 7→ (f, u) is a continous linear functional on H. It is a
remarkable fact that all continuous linear functionals on H are obtained in this fashion.
Theorem 11.1 (Riesz-Fréchet representation theorem). Given any F ∈ H ∗ there
exists a unique f ∈ H such that
hF, ui = (f, u) for all u ∈ H.
Morevorer, kf kH = kF kH ∗ .
You can find a proof of this important theorem for instance in [1, Chapter 6]. If you
are not acquainted with the basic theory of Hilbert spaces, I strongly recommend you to
read carefully that chapter.
The mapping F 7→ f , which is a linear surjective isometry, allows us to H ∗ with H. In
particular, any Hilbert space is reflexive.
Continuous linear functionals on Lp with p finite can be represented “concretely” as an
integral.
Theorem 11.2 (Riesz representation theorem). Let p ∈ [1, ∞), p0 its dual exponent,
0
and F ∈ (Lp )∗ . Then there exists a unique function f ∈ Lp such that
ˆ
hF, ui = f u for all u ∈ Lp .
then
sup kTi kL(X,Y ) < ∞.
i∈I
In other words, there exists a constant c such that
kTi xkY ≤ ckxkX for all x ∈ X and i ∈ I.
The conclusion of this theorem is quite remarkable and surprising. From pointwise
estimates one derives a global (uniform) estimate. You cand find a proof, which is based
in Baires’s Category Theorem, for example in [1, Theorem 2.2].
As a corollary we get the following result, that says that in order to prove that a set B
is bounded it suffices to “look” at B through the bounded linear functionals.
Proposition 11.4. Let X be a Banach space and let B be a subset of X. If for every
F ∈ X ∗ the set F (B) = {hF, ui : u ∈ B} is bounded (in R), then B is bounded.
Proof. For every b ∈ B we set Tb (F ) = hF, bi, F ∈ X ∗ . Since supb∈B |Tb (F )| < ∞ for all
F ∈ X ∗ , it follows from the Uniform Boundedness Principle that there exists a constant
c such that |hF, bi| ≤ ckF kX ∗ for all F ∈ X ∗ and b ∈ B. Therefore, using the well-known
formula
(22) kxkX = sup |hF, xi|,
F ∈X ∗ ,kF kX ∗ ≤1
Proof. Since uk * u, then hF, uk i → hF, ui in R for all F ∈ X ∗ , which implies that hF, uk i
is bounded in R for all F ∈ X ∗ . Therefore, Proposition 11.4 implies that the sequence is
bounded.
On the other hand, passing to the limit in the inequality
|hF, uk i| ≤ kF kX ∗ kuk kX ,
30 F. QUIRÓS
we obtain
|hF, ui| ≤ kF kX ∗ lim inf kuk kX ,
k→∞
Theorem 11.6 (Weak compactness). Let X be a reflexive Banach space and suppose
the sequence (uk )∞ ∞ ∞
k=1 in X is bounded. Then there exists a subsequence (ukj )j=1 ⊂ (uk )k=1
and u ∈ X such that ukj * u.
In other words, bounded sequences in a reflexive Banach space are weakly precompact.
Important example. If p ∈ [1, ∞), fk * f in Lp means
ˆ ˆ
0
gfk → gf as k → ∞ for all g ∈ Lp .
Definition 11.4. Let Ω be an open subset of Rd . We denote the dual of W01,p (Ω) by
0
W −1,p (Ω), and the dual of H01 (Ω) by H −1 (Ω).
Theorem 11.7 (Characterization of H −1 (Ω)). Let Ω be an open subset of Rd and
F ∈ H −1 (Ω).
(a) There exist functions f 0 , f 1 . . . , f d in L2 (Ω) such that
ˆ X d
0
(23) hF, ui = f u+ f i ∂i u for all u ∈ H01 (Ω).
Ω i=1
!1/2
d
X
(b) kF kH −1 (Ω) = inf kf i k2L2 (Ω) 0 1 d
: f , f , . . . , f satisfy (23) .
i=0
that is,
d
!1/2 d
!1/2
X X
kf i k2L2 (Ω) ≤ kg i k2L2 (Ω) .
i=0 i=0
References
[1] Brezis, H. “Functional analysis, Sobolev spaces and partial differential equations”. Universitext.
Springer, New York, 2011. ISBN: 978-0-387-70913-0.
[2] Evans, L. C. “Partial differential equations”. Second edition. Graduate Studies in Mathematics, 19.
American Mathematical Society, Providence, RI, 2010. ISBN: 978-0-8218-4974-3.
[3] Folland, G. B. “Real analysis. Modern techniques and their applications”. Second edition. Pure and
Applied Mathematics (New York). A Wiley-Interscience Publication. John Wiley & Sons, Inc., New
York, 1999. ISBN: 0-471-31716-0.
[4] Leoni, G. “A first course in Sobolev spaces. Second edition”. Graduate Studies in Mathematics, 181.
American Mathematical Society, Providence, RI, 2017. ISBN: 978-1-4704-2921-8.
[5] Spivak, M. “Calculus on manifolds. A modern approach to classical theorems of advanced calculus”.
W. A. Benjamin, Inc., New York-Amsterdam 1965.
LINEAR ELLIPTIC EQUATIONS
FERNANDO QUIRÓS
The aim of this chapter is to investigate the solvability of uniformly elliptic, second-
order partial differential equations, subject to prescribed boundary conditions. Some of
the tools that we will present are also valid for systems and for higher order equations.
To fix ideas, we will start by considering Dirichlet boundary conditions, prescribing the
value of the unknown u at the boundary of the domain. To make things simpler, we will
prescribe the value 0, so that we have homogeneous Dirichlet boundary conditions.
Given Ω ⊂ Rd open and bounded, and f : Ω −→ R, we consider the homogeneous
Dirichlet problem
(HDP) Lu = f in Ω, u = 0 in ∂Ω,
where L denotes a second-order partial differential operator in divergence form,
d
X d
X
ij
(1) Lu := − ∂j (a (x)∂i u) + bi (x)∂i u + c(x)u.
i,j=1 i=1
Remark. If the coefficients aij ∈ C 1 (Ω), then L can be written in the non-divergence form
d
X d
X
Lu = − aij (x)∂ij u + bei (x)∂i u + c(x)u,
i,j=1 i=1
Pd
with ebi = bi − j ∂j aij . Though this form may offer sometimes advantages, in this chapter
we will always deal with operators L written in divergence form.
In what follows we will always assume the symmetry condition aij = aji , plus the
boundedness of the coefficients, aij , bi , c ∈ L∞ (Ω).
Definition 1.1. We say the partial differential operator L is (uniformly) elliptic if there
exists a constant θ > 0 such that
Xd
aij (x)ξi ξj ≥ θ|ξ|2 , for almost every x ∈ Ω and all ξ ∈ Rd .
i,j=1
Ellipticity thus means that for each point x ∈ Ω the symmetric d × d matrix A(x) =
(aij (x))di,j=1 is positive definite, with smallest eigenvalue greater or equal than θ.
Remark. We also have, using Cauchy-Schwarz’s inequality, i,j aij (x)ξi ξj ≤ Λ|ξ|2 , with
P
Example. If aij (x) ≡ δij , bi ≡ 0 and c ≡ 0, then L = −∆. Solutions for the general
second-order elliptic partial differential equation Lu = 0 are similar in many ways to
harmonic functions. However, for these partial differential equations we do not have
available the various explicit formulas developed for harmonic functions.
1
2 F. QUIRÓS
Even when all the data of the problem (the domain, the coefficients and the right-
hand side) are smooth, proving the existence of a classical solution may be a hard task.
Moreover, if any of the data is not smooth, it may be the case that a classical solution
does not exist. Hence, it may be convenient to follow a different approach. One tries to
weaken the notion of solution so that is is easier to prove the existence of one. But we
have to be careful not weaken it too much, since we would like to have only one solution.
In a second step, one tries to show that if the data are smooth, the solution is smooth,
and hence a classical solution. Thus we considering separately two main issues: existence
(and uniqueness) and regularity. The second step is usually the hardest.
Any reasonable notion of weak solution should satisfy the following two requirements:
• A classical solution should be a weak solution.
• A smooth weak solution should be a classical solution.
A third requirement, which is not always achievable, is the existence of a unique weak
solution if the data are in a reasonable class.
When the operator L is in divergence form, a possibility to weaken the notion of solution
is to multiply the equation satisfied by classical solutions by a test function ϕ ∈ Cc∞ (Ω)
and integrate by parts, so that the highest derivative in the formulation is now a first
order one, instead of a second order one, and appears in an integrated form. Hence, in
order for the weak notion of solution to make sense we will only need the solution to be
in some first order Sobolev space. Let us apply this idea to (HDP).
Let u be a classical solution to (HDP) (all the data are assumed to be smooth). We
multiply the equation by v ∈ Cc∞ (Ω) and integrate by parts in the first term to obtain
ˆ X d X ˆ
ij i
(2) a ∂i u∂j v + b ∂i uv + cuv = f v.
Ω i,j=1 i Ω
There are no boundary terms since v = 0 on ∂Ω. By density, the same identity holds
with the smooth function v replaced by any v ∈ H01 (Ω). The resulting identity makes
sense if only u ∈ H01 (Ω). We choose the space H01 (Ω) instead of H 1 (Ω) to incorporate
LINEAR ELLIPTIC EQUATIONS 3
the boundary data. Notice that if u, v ∈ H01 (Ω) the identity makes sense requiring only
aij , b and c to be in L∞ (Ω), f ∈ L2 (Ω), and without any smoothness assumption on the
domain.
Definition 1.3. Let Ω be an open, bounded subset of Rd , ai,j , bi , c ∈ L∞ (Ω), and f ∈
L2 (Ω).
(i) The bilinear form B : H01 (Ω)×H01 (Ω) → R associated with the divergence form elliptic
operator L given in (1) is defined by
ˆ X d X
B[u, v] := aij ∂i u∂j v + bi ∂i uv + cuv .
Ω i,j=1 i
Indentity (3) is sometimes called the variational formulation of (HDP). The origin of
this terminology will become clear later in the course.
Remark. More generally, we may consider a right-hand side f ∈ H −1 (Ω) ≡ (H01 (Ω))∗ .
We say then that u ∈ H01 (Ω) is a weak solution of (HDP) if
(4) B[u, v] = hf, vi for all v ∈ H01 (Ω).
Remember that any f ∈ H −1 (Ω) can be characterized as f = f 0 − di=1 ∂i f i for some
P
functions f 0 , f 1 , . . . , f d ∈ L2 (Ω), meaning that
ˆ d
X
!
hf, vi = f 0v + f i ∂i v for all v ∈ H01 (Ω).
Ω i=1
It should be clear from the way we reached the notion of weak solution that classical
solutions are weak solutions. On the other hand, if a weak solution belongs to C 2 (Ω) ∩
C(Ω), then, since u ∈ H01 (Ω), we have that u(x) = 0 for all x ∈ ∂Ω. If the data of the
problem are moreover smooth, integrating by parts in (2) we obtain
ˆ
(Lu − f )v = 0 for all v ∈ H01 (Ω).
Ω
Since Lu − f ∈ C(Ω), we conclude that (Lu − f )(x) = 0 for all x ∈ Ω, and hence that u
is a classical solution.
2. Lax–Milgram’s Theorem
We now introduce a fairly simple abstract principle from linear functional analysis which
will provide under suitable hypotheses the existence and uniqueness of weak solutions to
several boundary-value problems.
Along this section H is a real Hilbert space, with inner product (·, ·) and norm k · k.
By h·, ·i we will denote the pairing of H with its dual space H ∗ .
Theorem 2.1 (Lax-Milgram’s Theorem). Assume that B : H × H −→ R is a bilinear
mapping for which there exist constants α, β > 0 such that
• |B[u, v]| ≤ αkukkvk (continuity);
4 F. QUIRÓS
a bijection, we have that there exists a unique u ∈ H such that Au = w. For this u we
have
B[u, v] = (Au, v) = (w, v) = hf, vi for all v ∈ H,
which is the desired existence result.
Let u1 , u2 ∈ H such that
B[u1 , v] = hf, vi = B[u2 , v] for all v ∈ H.
Then B[u1 − u2 , v] = 0 for all v ∈ H. Taking v = u1 − u2 and using the coerciveness we
get
0 = B[u1 − u2 , u1 − u2 ] ≥ βku1 − u2 k.
Hence u1 − u2 = 0, and we have uniqueness.
It should be clear from the previous steps that the application T : H ∗ → H is a linear
bijection. It is also bounded, hence continuous. Indeed, if u = T f ,
βkuk2 ≤ B[u, u] = hf, ui ≤ kf kH ∗ kuk,
hence kT f k ≤ β1 kf kH ∗ . By the Open Mapping Theorem2 it is an isomorphism.
Assume now that B is symmetric. Let u be the unique solution to (5) and v ∈ H,
v 6= u. Then
0 < β2 ku − vk2 ≤ 1
2
B[u − v, u − v] = 21 B[v, v] + 12 B[u, u] − B[u, v]
= J(v) − J(u) + B[u, u − v] − hf, u − vi = J(v) − J(u).
Therefore, J(u) < J(v) for all v ∈ H, v 6= u.
Remark. If the bilinear form B is symmetric we can fashion a much simpler proof by
noting that((u, v)) := B[u, v] is a new inner product on H, to which Riesz’s Representation
Theorem directly applies. Consequently, Lax-Milgram’s Theorem is primarily significant
in that it does bit require the symmetry of B.
We now go back to the Dirichlet problem (HDP), with f ∈ H −1 (Ω). Our aim is to give
conditions guaranteing that the bilinear form associated to an elliptic operator L satisfies
the hypotheses of Lax-Milgram’s theorem. We remark that B is symmetric if and only if
bi = 0 for all i = 1, . . . d.
Proposition 3.1 (Energy estimates). Let Ω be an open subset of Rd .
(i)) There exists a constant α > 0 such that
|B(u, v)| ≤ αkukH 1 (Ω) kvkH 1 (Ω) for all u, v ∈ H 1 (Ω).
(ii) If Ω is moreover bounded, there exist constants β > 0 and γ ≥ 0 such that
βkuk2H 1 (Ω) ≤ B[u, u] + γkuk2L2 (Ω) for all u ∈ H01 (Ω).
0
2Open Mapping Theorem. If X and Y are Banach spaces and A : X → Y is a surjective continuous
linear operator, then A is an open map (i.e. if U is an open set in X, then A(U ) is open in Y ).
As a corollary, if A : X → Y is a bijective continuous linear operator between the Banach spaces X
and Y , then the inverse operator A−1 : Y → X is continuous as well. This result, known as the Bounded
Inverse Theorem, is the one we are using here.
6 F. QUIRÓS
for α = di,j=1 kaij kL∞ (Ω) + di=1 kbi kL∞ (Ω) + kckL∞ (Ω) .
P P
Proof. Take γ from Proposition 3.1. Let µ ≥ γ and define then the bilinear form
Bµ [u, v] = B[u, v] + µ(u, v), u, v ∈ H01 (Ω),
which corresponds to the operator Lµ u := Lu + µu. Then Bµ satisfies the hypotheses of
Lax-Migram’s Theorem, hence the result.
3Young’s inequality with ε, also known as Peter-Paul inequality, valid for every ε > 0, says that
b2
ab ≤ εa2 + 4ε . The second name refers to the fact that tighter control of the first term is achieved at the
cost of losing some control of the second one – one must “rob Peter to pay Paul”. This inequality follows
immediately from Young’s inequality with p = q = 2.
LINEAR ELLIPTIC EQUATIONS 7
We already´ know from Proposition 3.1 that B is continuous. On the other hand, since
B[u, u] = Ω (|Du|2 + cu2 ), if c0 := essinf c > 0, then B is also coercive, and by Lax-
Milgram’s theorem the problem (7) has a unique weak solution.
If Ω is moreover bounded, we can relax the hypothesis on c. Indeed, thanks to Poincaré’s
´ 1/2
inequality, we know that Ω |Du|2 is equivalent to the H01 (Ω) norm when Ω is bounded.
Therefore, if c ≥ 0, ˆ
B[u, u] ≥ |Du|2 ≥ βkuk2H 1 (Ω)
0
Ω
for some β > 0, and B is coercive.
We can go even further. Let µ = µ(Ω) > 0 be the best constant in Poincaré’s inequality,
so that
kukL2 (Ω) ≤ µkDukL2 (Ω) .
Assume that 0 > c0 > −1/µ2 . Then
ˆ
B[u, u] ≥ (|Du|2 + c0 u2 ) ≥ kDuk2L2 (Ω) + c0 µ2 kDuk2L2 (Ω) = (1 + c0 µ2 )kDuk2L2 (Ω) ,
Ω
and B is again coercive.
Note that B is in this case symmetric. Hence, the unique weak solution of problem (7)
given by Lax-Milgram’s method under the conditions on c that guarantee that B is coer-
cive is the unique minimizer of the functional J : H01 (Ω) → R given by
ˆ
1
J(u) = (|Du|2 + cu2 ) − hf, ui.
2 Ω
(b) We consider now a bounded open set Ω, and an elliptic operator with associated
bilinear form
ˆ d
X
i
B[u, v] = Du · Dv + b ∂i uv .
Ω i=1
Continuity is once more provided by Proposition 3.1. On the other hand,
d
X
B[u, u] ≥ kDuk2L2 (Ω) − kbi kL∞ (Ω) k∂i ukL2 (Ω) kukL2 (Ω) .
i=1
i
Let ν := max1≤i≤d kb kL∞ . Since, by Peter-Paul inequality,
k∂i uk2L2 (Ω)
kukL2 (Ω) k∂i ukL2 (Ω) ≤ εkuk2L2 (Ω) +
4ε
for any ε > 0, then
d d
X
i
X k∂i uk2L2 (Ω)
− kb kL∞ (Ω) k∂i ukL2 (Ω) kukL2 (Ω) ≥ −ν εkuk2L2 (Ω) +
i=1 i=1
4ε
ν
= −νdεkuk2L2 (Ω) − kDuk2L2 (Ω) .
4ε
Therefore, using Poincaré’s inequality,
ν ν
B[u, u] ≥ 1 − kDuk2L2 (Ω) − νdεkuk2L2 (Ω) ≥ 1 − − νdεµ2 kDuk2L2 (Ω) .
4ε 4ε
We take now ε > 0 maximizing 1 − 4ε − νdεµ2 , that is, ε = 2√1dµ , which gives a value
ν
√
1 − ν dµ. Therefore, if ν < √1dµ we obtain the desired coercivity.
8 F. QUIRÓS
(c) We consider now the case in which the matrix aij is not the identity. Let
ˆ X
B[u, v] = aij ∂i u∂j v + cuv .
Ω i,j
d ˆ d ˆ 2
X
i
X
i
u
(B − B)[u, u] = b ∂i uu = b ∂i = 0,
i=1 Ω i=1 Ω 2
so that B[u, u] = B[u, u]. The result for general u ∈ H01 (Ω) follows by approximation.
(d) We consider now the system
−∆u1 + u2 = f1 in Ω, −∆u2 − u1 = f2 in Ω, u1 = u2 = 0 in ∂Ω.
We will work in the Hilbert space H = H01 (Ω) × H01 (Ω). We multiply the first equation by
v1 ∈ H01 (Ω) and the second one by v2 ∈ H01 (Ω), integrate by parts and add the equations
to obtain h i
u1 v f1 v
B , 1 = , 1 ,
u2 v2 f2 v2
where ˆ
u1 v
B , 1 = (Du1 · Dv1 + u2 v1 + Du2 · Dv2 − u1 v2 ),
u2 v2
ˆ Ω
f1 v1
, = f1 v1 + f2 v2 .
f2 v2 Ω
Since, thanks to Poincaré’s inequality,
ˆ
u1 u1
B , = (|Du1 |2 + |Du2 |2 ) = kDu1 k2L2 (Ω) + kDu2 k2L2 (Ω)
u2 u2 Ω
2
2 2
u1
≥ βku1 kH 1 (Ω) + βku2 kH 1 (Ω) = β
u2
0 0
H
for some β > 0, the bilinear form is coercive. It is also continuous, since
B u1 , v1 ≤kDu1 kL2 (Ω) kDv1 kL2 (Ω) + ku2 kL2 (Ω) kv1 kL2 (Ω)
u2 v2
+ kDu2 kL2 (Ω) kDv2 kL2 (Ω) + ku1 kL2 (Ω) kv2 kL2 |(Ω)
u1
v1
≤4
u2
v2
.
H H
Coercivity then follows from Poincaré’s inequality, since kukL2 (Ω) ≤ ckDukL2 (Ω) and
kDukL2 (Ω) ≤ C1 kD2 ukL2 (Ω) .
4. Fredholm’s alternative
there is a unique element zv ∈ H such that hv (u) = (u, zv ) for all u ∈ H. If we denote
A∗ v = zv , then we have (Au, v) = (u, A∗ v) for all u, v ∈ H, as desired.
(b) As can be easily checked (do it!), A∗ is linear and bounded.
Theorem 4.2 (Fredholm’s alternative, abstract version). Let H be a Hilbert space
and K : H → H a compact operator. Then:
(a) N(I − K) is finite dimensional.
(b) Im(I − K) is closed.
(c) N(I − K) = Im(I − K ∗ )⊥ .
(d) N(I − K) = {0} ⇐⇒ Im(I − K) = H.
(e) dim N(I − K) = dim N(I − K ∗ ).
Remark. Property (d) implies the following alternative: either (I − K)x = b has a
unique solution x ∈ H for all b ∈ H, or there is a nontrivial solution x ∈ H \ {0} of
the homogeneous problem (I − K)x = 0. This is an infinite-dimensional version of the
well-known result of Linear Algebra that says that either the system of equations Ax = b
(here A is a d × d matrix) has a unique solution x ∈ Rd for all b ∈ Rd of the homogeneous
system Ax = 0 has a nontrivial solution.
Definition 4.3. (a) The operator L∗ , the formal adjoint of L, is
d d d
!
X X X
L∗ v := − ∂i (aij ∂j v) − bi ∂ i v + c− ∂i bi v,
i,j=1 i=1 i=1
1
provided bi ∈ C (Ω), i = 1, . . . , d.
(b) The adjoint bilinear form B ∗ : H01 (Ω) × H01 (Ω) → R is defined by B ∗ [v, u] := B[u, v]
for all u, v ∈ H01 (Ω).
(c) We say that v ∈ H01 (Ω) is a weak solution of the adjoint problem
L∗ v = f in Ω, v = 0 in ∂Ω
∗
provided B [v, u] = (f, u) for all u ∈ H01 (Ω).
Theorem 4.3 (Second existence theorem for weak solutions).
(a) Precisely one of the two following statements holds:
either
(α) for each f ∈ L2 (Ω) there exists a unique weak solution u of
(8) Lu = f in Ω, u = 0 in ∂Ω,
or else
(β) there exists a weak solution u 6≡ 0 of the homogeneous problem
(9) Lu = 0 en Ω, u = 0 in ∂Ω.
The dichotomy (α) versus (β) is known as Fredholm’s alternative.
(b) Should assertion (β) hold, the dimension of the subspace N ⊂ H01 (Ω) of weak solutions
of (9) is finite and equals the dimension of the subspace N ∗ ⊂ H01 (Ω) of weak solutions of
(10) L∗ v = 0 in Ω, v = 0 in ∂Ω.
(c) Finally, the boundary-value problem (8) has a weak solution if and only if (f, v) = 0
for all v ∈ N ∗ .
LINEAR ELLIPTIC EQUATIONS 11
We observe next that u ∈ H01 (Ω) is a weak solution of (8) if and only if
Bγ [u, v] = (γu + f, v) for all v ∈ H01 (Ω),
that is, if and only if
u = L−1
γ (γu + f ).
We rewrite this equality to read
(12) u − Ku = h, for Ku := γL−1 −1
γ u and h := Lγ f.
In this section we will give through a series of examples an idea of how to deal with
boundary conditions different from Dirichlet homogeneous conditions.
Non-homogeneous Dirichlet conditions. We consider the problem
Lu = f in Ω, u = g in ∂Ω,
where L is a second order elliptic operator in divergence form. If there exists a function
u ∈ H 1 (Ω) such that T u = g, then the change of variables v = u − u transforms the
problem into one with homogeneous Dirichlet boundary conditions.
(14) Lv = f − Lu in Ω, v = 0 in ∂Ω.
Notice that Lu ∈ H −1 (Ω). Hence, if the bilinear form associated with L is continuous and
coercive, and f ∈ H −1 (Ω), Lax-Milgram’s theorem guarantees the existence of a unique
weak solution v to (14), which in turn provides us with a unique weak solution u = u + v
to our original problem (this is in fact a way to define the notion of weak solution for the
original problem).
Thus, we have to face the problem of finding u ∈ H 1 (Ω) such that T u = g. Let us
remind that the trace operator T : H 1 (Ω) → L2 (∂Ω) is linear and bounded. However,
it is not onto. Its image is the Sobolev space of fractional order H 1/2 (∂Ω), that can be
characterized as
ˆ ˆ
|v(x) − v(y)|2
1/2 2
H (∂Ω) = v ∈ L (∂Ω) : dσ(x)dσ(y) < ∞ .
∂Ω ∂Ω |x − y|d
By density, this equality will hold for all v ∈ H 1 (Ω). In order for this expression to make
sense we only need u ∈ H 1 (Ω). Therefore, we define a weak solution to (15) as a function
u ∈ H 1 (Ω) such that B[u, v] = (f, v) for´ all v ∈ H 1 (Ω) where B : H 1 (Ω) × H 1 (Ω) → R
is the bilinear form given by B[u, v] = Ω (Du · Dv + cuv). Let us remark that we have
proved that a classical solution is a weak solution.
If we deal with weak solutions we can relax the conditions on the data f and c. Indeed,
in order for the weak formulation to make sense we only need c ∈ L∞ (Ω) and f ∈ L2 (Ω).
In fact, we may even take f ∈ (H 1 (Ω))∗ , in which case the weak formulation consists in
finding u ∈ H 1 (Ω) such that B[u, v] = hf, vi for all v ∈ H 1 (Ω).
LINEAR ELLIPTIC EQUATIONS 13
We have already seen in Section 3 that if essinf Ω c > 0, then B is continuous and
coercive. Hence, by Lax-Milgram’s Theorem, the problem has a unique weak solution
for any data f ∈ (H 1 (Ω))∗ and c ∈ L∞ (Ω) satisfying the above mentioned positivity
condition. The unique weak solution is the unique minimizer in H 1 (Ω) of the functional
ˆ
1
J(v) = (|Dv|2 + cv 2 − f v).
2 Ω
Let us remark a big difference between homogeneous Dirichlet and Neumann boundary
conditions. The first ones are essential and have to be forced by asking the solutions to
belong to the functional space H01 (Ω)), while Neumann conditions are natural : they do
not have to be imposed in the functional space since minimization in H 1 (Ω) already tries
to make derivatives small,
Non-homogeneous Neumann conditions. Let Ω ⊂ Rd be again a bounded domain with
smooth boundary. We consider the non-homogeneous Neumann problem
∂u
(16) −∆u + cu = f in Ω, = g in ∂Ω.
∂ν
If f , g and c are smooth, then a classical solution to (16) is a function u ∈ C 2 (Ω)∩C 1 (Ω)
satisfying −∆u + cu = f everywhere in Ω and ∂ν u = g everywhere in ∂Ω.
Given a classical solution, if we multiply the equation by a test function v ∈ C 1 (Ω) and
integrate by parts we obtain
ˆ ˆ ˆ
f v = (Du · Dv + cuv) − gv.
Ω Ω ∂Ω
1
By density, this equality will hold for all v ∈ H (Ω). In order for this expression to make
sense we only need u ∈ H 1 (Ω). Therefore, we define a weak solution to (15) as a function
u ∈ H 1 (Ω) such that B[u, v] = F (v) ´for all v ∈ H 1 (Ω) where B : H 1 (Ω) × H 1 (Ω) → R is
1
´ v] = ´Ω (Du · Dv + cuv) and F : H (Ω) → R is the linear
the bilinear form given by B[u,
functional given by F (v) = Ω f v + ∂Ω gv. Let us remark that we have proved that a
classical solution is a weak solution.
If we deal with weak solutions we can relax the conditions on the data f , c and g. Indeed,
in order for the weak formulation to make sense we only need c ∈ L∞ (Ω), f ∈ L2 (Ω) (or
even f ∈ (H 1 (Ω))∗ ), and g ∈ L2 (∂Ω).
If essinf Ω c > 0, then B is continuous and coercive. Hence, to have existence and
uniqueness it is enough to check that F is continuous. This is an easy consequence of the
trace inequality,
|F (v)| ≤ kf kL2 (Ω) kvkL2 (Ω) + kgkL2 (∂Ω) kvkL2 (∂Ω) ≤ (kf kL2 (Ω) + kgkL2 (∂Ω) )kvkH 1 (Ω) .
Neumann condition for general elliptic operators. Once more Ω is a bounded open subset
of Rd with smooth boundary. Let L be a general elliptic operator in divergence form
d
X d
X
ij
Lu = − ∂j (a (x)∂i u) + bi (x)∂i u + c(x)u.
i,j=1 i=1
As we have already mentioned, Neumann conditions consist in prescribing the flux at the
boundary. The flux has to do with the operator, and in this case is given by ∂νA u =
Pd ij
i,j=1 a ∂i uνj . Thus, we consider the problem
Lu = f in Ω, ∂νA u = 0 in ∂Ω.
14 F. QUIRÓS
To check this, multiply the equation by v ∈ C 1 (Ω), integrate by parts and use the bound-
ary condition, and then use a density argument. A weak solution is a function u ∈ H 1 (Ω)
satisfying (17).
Neumann conditions for a non-coercive bilinear form. Let us now consider the Neumann
problem
−∆u = f in Ω, ∂ν u = 0 in ∂Ω,
where Ω ⊂ Rd is bounded, connected 2
´ and smooth, and f ∈ L (Ω). The bilinear form
associated to the operator, B[u, v] = Ω Du · Dv is not coercive in this case.
Remarks. (a) Weak solutions, if they exist, are not unique. Indeed, if we add a constant
to a solution we get another solution. The same is true for classical solutions.
(b) Any two weak solutions differ by a constant. Indeed, if u1 , u2 ∈ H 1 (Ω) solve B[u, v] =
F (v) for all v ∈ H 1 (Ω), then B[u1 − u2 , v] = 0 for all v ∈ H 1 (Ω). Taking v = u1 − u2 we
get that u1 − u2 is a constant.
(c) If u is a classical solution, then, using Gauss’ Divergence Theorem,
ˆ ˆ ˆ
f = − ∆u = − ∂ν u = 0.
Ω Ω ∂Ω
Thus,
´ there is no hope to obtain a classical solution unless the compatibility condition
Ω
f = 0 is satisfied.
In order to eliminate the lack of uniqueness due to´the possibility of adding constants,
we will work in the subspace H = {u ∈ H 1 (Ω) : Ω u = 0}. Observe that adding a
constant to an element of H takes us out of H.
The set H ´is the kernel of the linear and bounded application G : H 1 (Ω) → R given
by G(u) = Ω u. Hence H is a closed subspace of H 1 (Ω), and therefore a Hilbert
space. Is B[u, v] coervice in H? ffl The answer is yes, thanks to Poincaré-Wirtinger’s
inequality. Indeed, since ku − Ω ukL2 (Ω) ≤ CkDukL2 (Ω) for every u ∈ H 1 (Ω), then
kukL2 (Ω) ≤ CkDukL2 (Ω) for all u ∈ H, and we conclude that kDukL2 (Ω) is a norm in
H equivalent to the H 1 (Ω)-norm. Hence, B is coercive in H.
´
On the other hand, given f ∈ L2 (Ω), the functional F : H → R given by F (v) : Ω f v
is linear and bounded. Hence, by Lax-Milgram’s theorem, there exists a unique function
u ∈ H such that B[u, v] = F (v) for all v ∈ H.
Exercise. If f ∈ L2 (Ω) there is a weak solution even if f does not satisfy the compatibility
condition. What problem are we solving in this case?
Robin conditions. In this case we prescribe a value at the boundary for a certain
combination of u and its normal derivative,
−∆u = f in Ω, u + ∂ν u = 0 in ∂Ω.
LINEAR ELLIPTIC EQUATIONS 15
(with p = 2, and Γ = ∂Ω) valid for functions in W 1,p (Ω), if Γ ⊂ ∂Ω has nonzero ((d − 1)-
dimensional) measure. This inequality was proved in problem 2b of worksheet number
4.
Mixed conditions. In this case we prescribe Dirichlet conditions in part of the boundary
and Neumann conditions in the complement,
∂u
−∆u = f in Ω, u = 0 in Γ1 , = 0 in Γ2
∂ν
where ∂Ω = Γ1 ∪ Γ2 , Γ1 ∩ Γ2 = ∅, Γ1 open in ∂Ω and Γ1 6= ∅.
Dirichlet conditions are essential, and have to be imposed through the functional space.
Thus, we consider H = {u ∈ H 1 (Ω) : u = 0 a.e. in Γ1 }, which is a closed subspace of
H 1 (Ω), and hence a Hilbert space. Let us check this assertion more carefully.
Let (uk )∞ 1
k=1 be a sequence in H converging in H (Ω) to some function u. From the
trace inequality we have
kT ukL2 (Γ1 ) ≤ kT uk − T ukL2 (Γ1 ) ≤ kT (uk − u)kL2 (∂Ω) ≤ kuk − ukH 1 (Ω) → 0.
Let (uk )∞ 1
k=1 be a sequence in H converging in H (Ω) to some function u. Then
|u(1) − u(0)| = |u(1) − uk (1) + uk (1) − uk (0) + uk (0) − u(0)|
ˆ 1
0
= |u(1) − uk (1) + uk (0) − u(0)| = (u − uk )
0
ˆ 1
≤ |u − uk |0 ≤ k(u − uk )0 kL2 (Ω) ≤ ku − uk kH 1 (Ω) → 0.
0
Hence u(1) = u(0), which proves the claim.
Let us remark that the conditions on the function had to be included in the Hilbert
space (they are essential), in contrast with the conditions on the derivatives (which are
natural).
6. Regularity
Let us go back again to the homogeneous Dirichlet problem (HDP). Along this section
we always assume that Ω is Rd , Rd+ = Rd−1 × R+ or a bounded domain. We assume
moreover that L is a uniformly elliptic operator having the divergence form (1), and that
its associated bilinear form is continuous and coercive, so that we have a unique weak
solution u ∈ H01 (Ω) of problem (HDP) if f ∈ L2 (Ω). Is it true that if the data of the
problem are smooth, then u is smooth? The answer is yes. To make this statement more
precise, we will as necessary make various additional assumptions about the smoothness
of the coefficients aij , bi , and c, the right-hand side function f , and the domain Ω.
Let us first give conditions to have H 2 regularity.
Theorem 6.1 (H 2 regularity). Assume
aij ∈ C 1 (Ω), bi , c ∈ L∞ (Ω), f ∈ L2 (Ω), ∂Ω ∈ C 2 .
If u0 ∈ H01 (Ω) is the unique weak solution to (HDP), then u ∈ H 2 (Ω) and we have the
estimate kukH 2 (Ω) ≤ Ckf kL2 (Ω) , the constant C depending only on Ω and the coefficients
of L.
Let us explain the heuristics of the proof in the somewhat simpler case of problem
−∆u + u = f in Rd .
Remember that a weak solution of this problem is a function u ∈ H 1 (Rd ) such that
ˆ ˆ
(Du · Dv + uv) = f v for all v ∈ H 1 (Rd ).
Rd Rd
Take v = ∂ii u as a test function. Then, assuming that all the computations are justified,
ˆ ˆ ˆ
f ∂ii u = (Du · D∂ii u + u∂ii u) = − (|D∂i u|2 + (∂i u)2 ) = −k∂i uk2H 1 (Rd ) .
Rd Rd
Therefore
k∂i uk2H 1 (Rd ) ≤ kf kL2 (Rd ) k∂ii ukL2 (Rd ) ≤ kf kL2 (Rd ) k∂i ukH 1 (Rd ) ,
and we conclude that k∂i ukH 1 (Rd ) ≤ kf kL2 (Rd ) .
The problem with the above argument is that in order to use ∂ii u as a test function,
we would need u ∈ H 3 , a regularity that is not available. To circunvent this difficulty, we
follow the method of translations of L. Nirenberg, which takes as test functions difference
quotients instead of derivatives.
LINEAR ELLIPTIC EQUATIONS 17
The key point is that Sobolev spaces can be characterized in terms of incremental
quotients.
Theorem 6.2 (Difference quotients and weak derivatives).
(i) Let p ∈ [1, ∞). For each V ⊂⊂ Ω there exists a constant C such that
1
kDh ukLp (V ) ≤ CkDukLp (Ω) for all u ∈ W 1,p (Ω) and h ∈ Rd , 0 < |h| < dist(V, ∂Ω).
2
(ii) Let u ∈ Lp (V ), p ∈ (1, ∞). If there exists a constant C such that
1
(18) kDh ukLp (V ) ≤ C for all h ∈ R, 0 < |h| < dist(V, ∂Ω),
2
then u ∈ W 1,p (V ), and kDukLp (V ) ≤ C.
Proof. (i) Assume u is smooth. For each x ∈ V , i = 1, . . . , d, and 0 < |h| < 12 dist(V, ∂Ω)
we have
ˆ 1 ˆ 1
d
u(x + hei ) − u(x) = (u(x + thei )) dt = h ∂i u(x + thei ) dt,
0 dt 0
so that ˆ 1
|u(x + hei ) − u(x)|
≤ |Du(x + thei )| dt.
|h| 0
Therefore,
ˆ d ˆ ˆ
X 1 d ˆ
X 1 ˆ
h p p
|D u| ≤ C |Du(x + thei )| dtdx = C |Du(x + thei )|p dxdt
V i=1 V 0 i=1 0 V
d ˆ 1
X
≤C kDukpLp (Ω) ≤ CdkDukpLp (Ω) .
i=1 0
This estimate holds for all smooth u, and also by an approximation argument (check it!),
for all u ∈ W 1,p (Ω).
(ii) Assume now that (18) holds for some constant C. Choose i ∈ {1, . . . , d}, φ ∈ Cc∞ (V )
and note for small enough h that
ˆ ˆ
φ(x + hei ) − φ(x) u(x) − u(x − hei )
u(x) dx = − φ(x) dx.
V h V h
This equality, that follows just from a change of variables, can be written as
ˆ ˆ
u(Di ϕ) = − (Di−h u)φ,
h
V V
18 F. QUIRÓS
Di−hk u * vi weakly in Lp (V ).
But then
ˆ ˆ ˆ ˆ ˆ ˆ
hk −hk
u∂i φ = u∂i φ = lim uDi φ = − lim (Di u)φ = − vi φ = − vi φ.
V Ω hk →0 Ω hk →0 V V Ω
p
Thus, vi = ∂i u, i = 1, . . . , d, in the weak sense, and so ∂i u ∈ L (V ), i = 1, . . . , d. As
u ∈ Lp (V ), we deduce therefore that u ∈ W 1,p (V ).
Proof of Theorem 6.1. In order to simplify the exposition, we will only give the proof for
the equation −∆u + u = f , since this case already contains the main ideas, while keeping
technicalities to a minimum. On the other hand we will only consider the domains Ω = Rd
and Ω = Rd+ . Once these two cases are well understood, other domains can be treated
using a partition of unity and straightening the boundary.
The idea is to use as test function in the weak formulation the second difference quotient
u(x + hei ) + u(x − hei ) − 2u(x)
Di−h Dih u(x) = , h 6= 0,
h2
which is an approximation of ∂ii u(x) if u is smooth.
Les us start by considering the case Ω = Rd . Taking v = Di−h Dih u in the weak formu-
lation we get
ˆ ˆ
−h h −h h −h h
f Di Di u = DuD(Di Di u) + uDi Di u .
Rd Rd
Integrating by parts the difference quotients, we obtain
ˆ ˆ
−h h
f Di Di u = − |D(Dih u)|2 + |Dih |2 .
Rd Rd
Thus,
ˆ
kDih uk2H 1 (Rd ) =− f Di−h Dih u ≤ kf kL2 (Rd ) kDi−h Dih ukL2 (Rd )
Ω
≤ Ckf kL2 (Rd ) kD(Dih u)kL2 (Rd ) ≤ Ckf kL2 (Rd ) kDih ukH 1 (Rd ) ,
and hence kDih ukH 1 (Rd ) ≤ kf kL2 (Rd ) . Therefore, kDih DukL2 (Rd ) ≤ kf kL2 (Rd ) , and hence,
thanks to Proposition 6.2, k∂i DukL2 (Rd ) ≤ kf kL2 (Rd ) , i = 1, ..., d; in particular, u ∈
H 2 (Rd ), and kD2 ukL2 (Rd ) ≤ kf kL2 (Rd ) . On the other hand, if we take u as test function
in the definition of weak solution we get
ˆ
2 2 1 1
kDukL2 (Rd ) + kukL2 (Rd ) = f u ≤ kf kL2 (Rd ) kukL2 (Rd ) ≤ kf k2L2 (Rd ) + kukL2 (Rd ) ,
Rd 2 2
and hence kukH01 (Ω) ≤ kf k2L2 (Rd ) . Therefore kukH 2 (Ω) ≤ Ckf kL2 (Ω) .
LINEAR ELLIPTIC EQUATIONS 19
−h h
We consider next Ω = Rd−1 + . We take again as test function Di Di u, but only for i =
1, . . . , d−1. Argueing as in the case of the whole space we get k∂i DukL2 (Rd+ ) ≤ Ckf kL2 (Rd+ ) .
We still have to control ∂ii u. To this aim we use the equation. We have
ˆ ˆ X d−1 ˆ X d−1
∂d u∂d ϕ = − ∂i u∂i ϕ−uϕ+f ϕ = ∂ii u−u+f ϕ for all ϕ ∈ Cc∞ (Ω).
Rd+ Rd+ i=1 Rd+ i=1
Pd−1
Therefore, by the definition of weak derivative, ∂dd u = − i=1 ∂ii u + u − f ∈ L2 (Rd+ ) in
the weak sense, and hence
k∂dd ukL2 (Rd ) ≤ (d − 1)kf kL2 (Rd ) + kukL2 (Rd ) + kf kL2 (Rd ) ≤ Ckf kL2 (Rd ) .
The estimate kukH 2 L2 (Rd ) ≤ Ckf kL2 follows easily.
We observe that under the hypotheses of Theorem 6.1 the weak solution u of (HDP)
satisfies Lu = f almost everywhere in Ω. It is hence a strong solution: all the deriva-
tives appearing in the equation are L1loc functions, and the equation is satisfied almost
everywhere.
Theorem 6.3 (Higher regularity). Let m be a nonnegative integer, and assume
aij ∈ C m+1 (Ω), bi , c ∈ C m (Ω), f ∈ H m (Ω), ∂Ω ∈ C m+2 .
If u0 ∈ H01 (Ω) is the unique weak solution to (HDP), then u ∈ H m+2 (Ω) and we have
the estimate kukH m+2 (Ω) ≤ Ckf kH m (Ω) , the constant C depending only on Ω and the
coefficients of L.
The proof is by induction, and follows the same lines as the proof of the previous
theorem, hence we omit it. You can find a detailed proof in [2, Section 6.3].
Theorem 6.4 (Infinite differentiability). Assume
aij , bi , c ∈ C ∞ (Ω), f ∈ C ∞ (Ω), ∂Ω ∈ C ∞ .
If u0 ∈ H01 (Ω) is the unique weak solution to (HDP), then u ∈ C ∞ (Ω).
Proof. According to the previous theorem, u ∈ H m (Ω) for each m ∈ N. Thus, Theorem
10.3 in Chapter 1 implies that u ∈ C k (Ω) for each k ∈ N.
Remark. If u ∈ H 1 (Ω) satisfies the equation, but not necessarily the boundary conditions,
the above mentioned global regularity results (valid in the whole Ω) fail in general, but
we still have local regularity results, valid in any Ω0 ⊂⊂ Ω. These results do not require
smoothness of the boundary, or regularity of the coefficients up to the boundary. See [2,
Section 6.3] for the details.
Thanks to these regularity results we will be able to show that if the data of the problem
are smooth, then weak solutions are smooth, and hence classical. Let us check this in a
couple of examples.
Examples. (a) Dirichlet problem. Let u ∈ H01 (Ω) be the classical solution to (HDP).
Assume that the coefficients aij , bi and c, and the domain Ω are smooth. Let f ∈ H m (Ω),
m > d/2. The regularity result shows that u ∈ H m+2 (Ω), hence ∂ij u ∈ H m (Ω). Thus,
from the Sobolev embeddings, since 2m > d, we get that
∂ij u ∈ C m−[d/2]−1,γ (Ω) ⊂ C 0,γ (Ω), γ = 1 − {d/2}.
20 F. QUIRÓS
Therefore, u ∈ C 2,γ (Ω). In particular, u ∈ C 2 (Ω) ∩ C(Ω). Since u ∈ H01 (Ω), we moreover
have u = 0 at ∂Ω. We can now integrate by parts, and we get
ˆ
(Lu − f )ϕ = 0 for all ϕ ∈ Cc∞ ,
Ω
hence Lu = f everywhere in Ω, and we conclude that u is a classical solution.
(b) Neumann problem. Let u be a weak solution to the Neumann problem
−∆u + u = f in Ω, ∂ν u = 0 in ∂Ω.
That is, u ∈ H 1 (Ω) satisfies
ˆ ˆ
(Du · Dv + uv) = f vfor all v ∈ H 1 (Ω).
Ω Ω
It can be easily checked that the same proof that we gave for the Dirichlet problem
works in this case, and we get that f ∈ L2 (Ω) implies H 2 (Ω), while f ∈ H m (Ω) implies
u ∈ H m+2 (Ω).
As in example (a), if f ∈ H m (Ω), m > d/2, then u ∈ C 2,γ (Ω). In particular u ∈
C 2 (Ω) ∩ C 1 (Ω). We can now integrate by parts, and we obtain
ˆ ˆ
(−∆u + u − f )ϕ + ∂ν uϕ = 0 for all ϕ ∈ C 1 (Ω).
Ω ∂Ω
∞
´
In particular, for all v ∈ Cc (Ω), Ω (−∆u + u − f )ϕ = 0, and hence −∆u + u = f
´
everywhere in Ω. Once we know this, we have that ∂Ω ∂ν uϕ = 0 for all ϕ ∈ C 1 (Ω), which
implies that ∂ν u = 0 in ∂Ω.
Summarizing, u is a classical solution of the Neumann problem.
7. Maximum Principle
The Maximum Principle is a very useful tool that admits a number of formulations. We
present here a simple form, and only for the Dirichlet problem for a particular equation
in a bounded domain. But the technique that we will use can be adapted to deal with
other equations and boundary conditions.
Theorem 7.1 (Maximum Principle for the Dirichlet problem). Let Ω ⊂ Rd be a
bounded domain. Let f ∈ L2 (Ω) and u ∈ H 1 (Ω) ∩ C(Ω) satisfy4
ˆ ˆ
(Du · Dv + uv) = f v for all v ∈ H01 (Ω).
Ω Ω
Then for all x ∈ Ω
min{inf u, inf f } ≤ u(x) ≤ max{sup u, sup f }.
∂Ω Ω ∂Ω Ω
Proof. We use Stampacchias’s truncation method. Fix a function G ∈ C 1 (R) such that:
• |G0 (s)| ≤ M for all s ∈ R for some M ∈ R+ ;
• G is strictly increasing on R+ ;
4If 1
C one can remove the assumption u ∈ C(Ω) by1invoking the theory of traces to give
Ω is of class
a meaning to u ∂Ω . One can also remove this assumption if u ∈ H0 (Ω). It is a good exercise to remove
this assumption in these two cases.
LINEAR ELLIPTIC EQUATIONS 21
References
[1] Brezis, H. “Functional analysis, Sobolev spaces and partial differential equations”. Universitext.
Springer, New York, 2011. ISBN: 978-0-387-70913-0.
[2] Evans, L. C. “Partial differential equations”. Second edition. Graduate Studies in Mathematics, 19.
American Mathematical Society, Providence, RI, 2010. ISBN: 978-0-8218-4974-3.
SEMILINEAR ELLIPTIC EQUATIONS
FERNANDO QUIRÓS
The aim of this chapter is to introduce the student to variational techniques which
allow to study semilinear elliptic equations. The model problem that we have in mind is
−∆u = f (u) in Ω, u = 0 in ∂Ω.
The idea is to do differential calculus in Banach spaces.
= (Du · Dv − f v).
Ω
2. Variational approach
As we have seen in the previous chapter, one can easily prove the existence of a unique
weak solution for problem (2) by means, for example, of Lax-Milgram’s theorem. However,
this is a linear technique, and will not work to deal with the nonlinear problems we have
in mind. Hence, we follow another approach: the so-called variational method. The idea
is to identify a functional having as Euler-Lagrange equation the one that we are trying
to analyze. If we are able to find a critical point of the functional, we will have found a
weak solution to our problem.
How will we find critical points? A first idea is to find extremal points, minimizing or
maximizing the functional.
Proposition 2.1. Let X be a Banach space, and let I : X → R be Gâteaux differentiable
in u ∈ X. If I[u] = minv∈X I[v] then I 0 [u] = 0.
Remark. The argument is local. The result is also true if u is a minimum point in Bδ (u)
for some δ > 0.
Let us consider now a nonlinear example.
Proposition 2.2. Let Ω ⊂ Rd be bounded and smooth. We consider the functional
I : H01 (Ω) → R given by
ˆ ˆ u
I[u] = F (x, u(x)) dx, F (x, u) = f (x, s) ds,
Ω 0
Thanks to Sobolev’s embedding we know that v and |v|p+1 belong to L1 (Ω). Hence, it
is enough to check that |u|p |v| ∈ L1 (Ω). This follows easily from Hölder’s inequality and
Sobolev’s embedding. Indeed,
ˆ ˆ p ˆ
p+1 1
p+1
p p+1 p+1
|u| |v| ≤ |u| |v| < ∞.
Ω Ω Ω
´
We can therefore pass to the limit to obtain ∂v I[u] = Ω
f (x, u(x))v(x) dx.
Step 3: I is Gâteaux differentiable. It is trivially checked that v → ∂v I[u] is linear
in v. It is enough then to check that this application is bounded. Indeed, using Hölder’s
inequality and Sobolev’s embedding
ˆ ˆ d+2
2d
2d
|∂v I[u]| ≤ |f (x, u(x))||v(x)| dx ≤ |f (x, u(x))| d+2 dx kvk2∗
Ω Ω
ˆ d+2
2d
2d
≤C (1 + |u|p ) d+2 dx kvkH01 (Ω)
Ω
ˆ d+2
2d
2dp
≤C 1 + |u| d+2 dx kvkH01 (Ω) .
Ω
2dp
2dp 2d
Hence, it is enough to prove that |u| d+2 ∈ L1 (Ω), which is the case, since d+2
≤ d−2
(this
is equivalent to p ≤ d+2
d−2
).
Remark. One can find in the literature other definitions of coercivity, for example,
I[u] ≥ αkuk2X − β for all u ∈ X for some α > 0, β ∈ R,
which is more restrictive than the one we have given, or
I[uk ] → ∞ if kuk k → ∞,
which is less restrictive. All of them serve for our purposes.
Let (uk )∞
k=1 be a minimizing sequence, that is, a sequence of elements of X such that
I[uk ] → m. If k is big enough, then I[uk ] ≤ m + 1. Hence, using the coercivity,
αkuk kX − β ≤ m + 1,
which implies that kuk kX ≤ C < ∞. Now, if X is reflexive, then there exists a subsequence
that converges weakly in X, ukj * u in X. However, even if I is continuous, we do not
have in general I[ukj ] → I[u] (weak convergence is not enough). Hence, we will need to
require something else from I.
Definition 3.2. A functional I : X → R is sequentially weakly lower semicontinuous if
I[u] ≤ lim inf I[uk ] if uk * u in X.
k→∞
Thus, if (ukj )∞
j=1 is a weakly convergent subsequence of the minimizing sequence, and I
is sequentially weakly lower semicontinuous, then
m ≤ I[u] ≤ lim inf I[ukj ] = m;
j→∞
that is, I[u] = m, and hence I achieves its minimum at u ∈ X. We have therefore proved
the following result.
Theorem 3.1. Let X be a reflexive real Banach space, and I : X → R a coercive and
sequentially weakly lower semicontinuous functional. Then, there exists u ∈ X such that
I[u] = minv∈X I[v].
Proposition 3.2. Norms in a Hilbert space are sequentially weakly lower semicontinuous.
Remark. Let M ⊂ X be weakly closed. The proof of the theorem can be repeated to
show that there exists u ∈ M such that I[u] = inf{I[u] : u ∈ M}.
Theorem 3.3 (Mazur’s theorem). Let X be a reflexive Banach space. A closed convex
subset of X is weakly closed.
Corollary 3.4. Let X be a reflexive Banach space, and M ⊂ X a closed convex set. If
I is coercive and sequentially weakly lower semicontinuous, then there exists u ∈ M such
that I[u] = inf{I[u] : u ∈ M}.
Definition 3.3. The functional I : X → R is convex if I[λx+(1−λ)y] ≤ λI[x]+(1−λ)I[y]
for all x, y ∈ X and λ ∈ [0, 1]. If equality holds only for λ = 0 and λ = 1, then I is said
to be strictly convex.
Proposition 3.5. If I is strictly convex, then I has at most one minimum.
Integrating the inequality φ0 (λ) ≥ φ0 (0) we get φ(λ) ≥ φ(0) + φ0 (0)λ, that is,
I[x + λ(y − x)] ≥ I[x] + I 0 [x](y − x)λ.
Defining z := x+λ(y−x) we have (y−x)λ = z −x, and hence I[z] ≥ I[x]+I 0 [x](z −x).
Proposition 3.7. If I : X → R is sequentially lower semicontinuous and convex then it
is sequentially weakly lower semicontinuous. In particular, if I is continuous and convex
it is sequentially weakly lower semicontinuous.
You can find a proof, which is based in a lemma due to Mazur, for example in [3].
Proposition 3.8 (Generalized Dominated Convergence Theorem). Let Ω be a
measurable sets. Let (fj )∞ ∞ 1
j=1 and (hj )j=1 be sequences in L (Ω) such that fj → f almost
everywhere
´ in´ Ω, |fj | ≤ hj , hj → h almost everywhere in Ω and in L1 (Ω). Then f ∈ L1 (Ω)
and Ω fj → Ω f .
Proof. The sequence (hj − fj )∞j=1 converges almost everywhere in Ω to h − f . Since the
functions in the sequence are nonnegative, we may apply Fatou’s Lemma to obtain
ˆ ˆ ˆ
h − f = (h − f ) ≤ lim inf (hj − fj )
j→∞
Ω Ω Ω
ˆ ˆ ˆ ˆ
= lim hj − lim sup fj = h − lim sup fj .
j→∞ Ω j→∞ Ω Ω j→∞ Ω
Therefore, ˆ ˆ
lim sup fj ≤ f.
j→∞ Ω Ω
On the other hand, the sequence (hj + fj )∞ j=1 converges almost everywhere in Ω to
h + f . Since the functions in the sequence are nonnegative, we may apply Fatou’s Lemma
to obtain ˆ ˆ ˆ
h + f = (h + f ) ≤ lim inf (hj + fj )
j→∞
Ω Ω Ω
ˆ ˆ ˆ ˆ
= lim hj + lim inf fj = h + lim inf fj .
j→∞ Ω j→∞ Ω Ω j→∞ Ω
Therefore, ˆ ˆ
f ≤ lim inf fj .
Ω j→∞ Ω
Summarizing, ˆ ˆ ˆ
lim sup fj ≤ f ≤ lim inf fj ,
j→∞ Ω Ω j→∞ Ω
and hence ˆ ˆ
lim fj = f.
j→∞ Ω Ω
which is Gâteaux differentiable if |f (x, s)| ≤ A(1 + |s|p ), p ∈ [1, (d + 2)/(d − 2)] if d ≥ 3,
p ∈ [1, ∞] if d = 2, no growth condition if d = 1.
Is this functional sequentially weakly lower semicontinuous? This is clear for I1 [u] =
1
2
kukH01 (Ω) ,
since it is a multiple of the norm of a Hilbert space. As for I2 , it is also
sequentially weakly lower semicontinuous under the additional restriction p ∈ [1, (d +
2)/(d − 2)) if d ≥ 3, p ∈ [1, ∞) if d = 2. Let us prove for the case d ≥ 3. It is a good
exercise to do the cases d = 1, 2.
Let uk * u in H01 (Ω). Then kuk kH01 (Ω) ≤ C. Therefore, using Rellich-Kondrachov’s
Compactness Theorem, there is a subsequence (ukj )∞ j=1 such that ukj → u in L
p+1
(Ω) (the
∗
restriction on p is used here to guarantee that p + 1 ≤ 2 ) and ukj → u almost everywhere
in Ω. Hence, F (x, ukj (x)) → F (x, u(x)) for almost every x ∈ Ω. On the other hand,
|F (x, ukj (x))| ≤ A(|ukj (x) + |ukj (x)|p+1 ) → A(|u(x)| + |u(x)|p+1 ) ∈ L1 (Ω)
both almost everywhere in Ω and in L1 (Ω). Hence, we can apply the Generalized Domi-
nated Convergence Theorem to obtain that
ˆ ˆ
I2 [ukj ] = F (x, ukj (x)) dx → F (x, u(x)) dx = I2 [u].
Ω Ω
By using a contradiction argument (do it!) it is easy that this convergence is not restricted
to a subsequence. Hence I2 , and therefore I, is sequentially weakly lower semicontinuous.
What about coercivity? We will give several sufficient conditions guaranteing it. Let
us recall that ˆ
1 2
I[u] = kukH 1 (Ω) − F (x, u(x)) dx.
2 0
Ω
Condition 2: F (x, u(x)) ≤ γ|u(x)| + a(x) for some a ∈ L1 (Ω), γ ≥ 0 and q ∈ (1, 2).
q
where we are using that kukLq (Ω) ≤ CkukL2 (Ω) ≤ CkukH01 (Ω) , since Ω is bounded. There-
fore, I[u] = αkuk2H 1 (Ω) − β for some α > 0, because
0
! ( 1
1 2
1 C 4
kukH 1 (Ω) if kukH 1 (Ω) ≥ (4C) 2−q ,
kuk2H 1 (Ω) − ≥ 1 0
q
0
1
0 2 kuk2−q H 1 (Ω) 2
kuk2
H 1 (Ω)
− C(4C) 2−q if kukH01 (Ω) ≤ (4C)
2−q .
0 0
Condition 3: F (x, u(x)) ≤ γ|u(x)|2 + a(x) for some a ∈ L1 (Ω), and γ ≥ 0 small.
Indeed, under these conditions
1 1
I[u] ≥ kuk2H 1 (Ω) − γCkukqLq (Ω) − kakL1 (Ω) ≥ α kuk2H 1 (Ω) − kakL1 (Ω)
2 0 2 0
if γ < C/2.
SEMILINEAR ELLIPTIC EQUATIONS 9
We start by reminding a well-known result valid for finite dimensional spaces. Let
F, G ∈ C 1 (Rd ; R), and C = {x ∈ Rd : G(x) = 0}. If x0 ∈ C is such that F (x0 ) = minC F ,
then, either G0 (x0 ) = 0, or there exists a value µ ∈ R such that F 0 (x0 ) = µG0 (x0 ). The
value µ is said to be a Lagrange multiplier.
What happens in a general Banach space? As we will see, if we state it properly, the
result is still true.
We start by proving a technical lemma.
Lemma 4.1. Let X be a real Banach space and F, G ∈ C 1 (X; R). Let x0 ∈ X. If there
exist v, w ∈ X such that
(F 0 [x0 ]v)(G0 [x0 ]w) 6= (F 0 [x0 ]w)(G0 [x0 ]v),
then F cannot have an extremal point at x0 even if we restrict to the level set C = {x ∈
X : G[x] = G[x0 ]}.
Proof. If (a) does not hold, there exists w ∈ X such that G0 [x0 ]w 6= 0. By the previous
lemma, necessarily
(F 0 [x0 ]v)(G0 [x0 ]w) = (F 0 [x0 ]w)(G0 [x0 ]v).
Hence, for all v ∈ X we have F 0 [x0 ]v = µG0 [x0 ]v, with µ = (F 0 [x0 ]w)/(G0 [x0 ]w).
As an application of this theorem we will now study the eigenvalue problem for the
Laplacian with homogeneous Dirichlet boundary conditions.
Let Ω ⊂ Rd be bounded, connected, and smooth. We are looking for non-trivial (weak)
solutions u to
−∆u = λu in Ω, u = 0 in ∂Ω.
As we will see, such solutions only exist for certain values λ, the eigenvalues. The solution
u is said to be an eigenfunction associated to the eigenvalue λ. Thus, we are looking for
10 F. QUIRÓS
the Lagrange Multiplier’s Theorem will imply that there exists a value λ such that
F 0 [ū]v = λG0 [ū]v for all v ∈ H01 (Ω). Thus, the pair (λ, ū) will be a solution pair to
the eigenvalue problem.
Let us prove then that F achieves its minimum in C. Let I := inf{F [u] : u ∈ C} ≥ 0.
Let (uk )∞ k=1 , uk ∈ C, be a minimizing sequence in C, F [uk ] → I. We have kDuk k2 ≤ K.
Hence, thanks to Poincaré’s inequality, kuk kH01 (Ω) ≤ K. We can now apply Rellich-
Kondrachov’s compactness theorem to obtain a subsequence (ukj )∞ j=1 such that ukj → ū
2 1 2
in L (Ω), ukj * ū in H0 (Ω). The strong convergence in L (Ω) yields
Hence, ū ∈ C.
On the other hand, since the norm in a Banach space is sequentially weakly lower
semicontinuous,
I ≤ F [ū] = kūk2H 1 (Ω) − kūk2L2 (Ω) ≤ lim inf kukj k2H 1 (Ω) − lim kukj k2L2 (Ω)
0 j→∞ 0 j→∞
≤ lim inf kukj k2H 1 (Ω) − kukj k2L2 (Ω) = lim inf F [ukj ] = I,
j→∞ 0 j→∞
We recognize here Poincaré’s inequality. We conclude that the best constant in Poincaré’s
inequality is given by the first 1 eigenvalue. Moreover, equality in this inequality is achieved
by ū.
Notation. We denote the pair (λ, u) that we have just obtained by (λ1 , u1 ).
´
Observe that λ1 > 0. Indeed, since λ1 = Ω |Dū|2 , if λ1 = 0 we have that ū is a
constant, which has to be 0, since ū ∈ H01 (Ω). Therefore, ū ≡ 0, a contradiction, since
kūk1 = 1.
Are there other essentially different (eigenfunctions which are not multiples of u1 ) so-
lution pairs?
Assume that there is another solution pair, (λ2 , u2 ), ku2 k2 = 1. We have
ˆ ˆ ˆ ˆ ˆ
(λ2 − λ1 ) u1 u2 = λ2 u1 u2 − λ1 u1 u2 = Du1 · Du2 − Du1 · Du2 = 0,
Ω Ω Ω Ω Ω
where´we have used that both (λ1 , u1 ) and (λ2 , u2 ) are solution pairs. Therefore, if λ1 6= λ2 ,
then Ω u1 u2 = 0. In other words, (u1 , u2 )L2 (Ω) = 0. We have thus proved the following
result.
Proposition 4.3. The eigenfunctions of the Laplacian corresponding to different eigen-
values are orthogonal in L2 (Ω).
Since X1 ⊂ H01 (Ω), λ2 ≥ λ1 . Repeating what we did to obtain the pair (λ1 , u1 ), it is easily
checked that λ2 is attained at some u2 ∈ X1 .
Recursively, given the solution pairs (λj , uj ), j = 1, . . . , n, let
Xn = v ∈ H01 (Ω) : (v, uj )L2 (Ω) = 0, j = 1, . . . , n , and
ˆ ˆ
2 2
λn+1 = inf {F [v] : v ∈ Xn , kvk2 = 1} = inf |Dv| / v : v ∈ Xn .
Ω Ω
Then λn+1 ≥ λn , and λn+1 is achieved at some un+1 ∈ Xn+1 . Notice that (uk , uj )L2 (Ω) =
δkj , so that we have an orthonormal set in L2 (Ω). Moreover, (Duk , Duj )L2 (Ω) = λk δkj , so
that (uk , uj )H01 (Ω) = (λk + 1)δkj .
Remark. We do not know yet whether this procedure generates all the eigenvalues and
eigenfunctions.
We claim now that λk → ∞. Otherwise, λk is a bojnded sequence, and hence, since
kuk k2H 1 (Ω) )λk+1 , the sequence (uk )∞
k=1 would be bounded. Therefore, Rellich-Kondrachov’s
0
1In the next pages it will be clear why we denote the eigenvalue that we have just obtained as the
firstone
12 F. QUIRÓS
compactness theorem implies that the sequence has a convergent in L2 (Ω) subsequence
(ukj )∞
j=1 . It is then a Cauchy sequence, kukj − ukl kL2 (Ω) → 0 as j, l → ∞. However,
since the sequence is orthonormal in L2 (Ω), we have kukj − ukl kL2 (Ω) = 2 if k 6= l, a
contradiction.
Corollary 4.4. Each λk can appear in the list generated by this procedure only a finite
number of times.
Theorem 4.5. Let (uk )∞ k=1 be the orthonormal family of eigenfunctions of the Laplacian
generated in the procedure described above. Given f ∈ L2 (Ω), there exists a sequence
∞
(αk )∞ 2
P
k=1 of real numbers such that f = k=1 αk uk in L (Ω), that is,
N
X
f − αk uk
→ 0 as N → ∞.
2
k=1 L (Ω)
we conclude that
N
X N
X
kf − αk uk k2 ≤ 2kf − fε k2 + kfε − αk,ε uk k2 .
k=1 k=1
We conclude by letting ε → 0.
This theorem guarantees that all eigenvalues and eigenfunctions are generated in the
recursive minimization process described above. Indeed, if (λ, u) u 6≡ 0, is a solution pair
with λ 6= λn for all n ∈ N, then, since u ∈ H01 (Ω) ⊂ L2 (Ω),
X∞
u= (u, uk )L2 (Ω) uk in L2 (Ω).
k=1
Proof. The system is clearly orthornomal. Hence it is enough to check that it is maximal 3.
1/2
Let f ∈ H01 (Ω) be such that (f, un /λn ) = 0 for all n ∈ N. Since
ˆ ˆ
un 1/2
0= Df · D 1/2 = λn f un for all n ∈ N,
Ω λn Ω
and λn > 0, we conclude that (f, un )L2 (Ω) = 0 for all n ∈ N. But (un )∞ n=1 is a Hilbert
2 2
orthormal basis in L (Ω), hence a maximal system in L (Ω), and therefore f = 0. We
1/2
conclude that (un /λn )∞ n=1 is maximal in H0 (Ω).
1
Corollary 4.7. If f ∈ H01 (Ω), then f = ∞ 1
P
k=1 (f, uk )L2 (Ω) uk in the sense of H0 (Ω), that
is,
XN
f − (f, uk )L2 (Ω) uk
→ 0 as N → ∞.
1
k=1 H0 (Ω)
3An orthonormal system (vn )∞n=1 in a Hilber space H is said to be maximal if the only function f ∈ H
satisfying (f, vn )H = 0 is f = 0. It is well known that an orthonormal system is a Hilbert basis if and
only if it is maximal.
14 F. QUIRÓS
´
Proof. We are using the scalar product (u, v) = Ω
Du · Dv.
1/2 P∞ uk
Let µk = (f, uk /λk ), k ∈ N. From the previous proposition, f = k=1 µk 1/2 in the
λk
sense of H01 (Ω). But
ˆ ˆ
1/2 1 1/2 1/2
µk = (f, uk /λk ) = 1/2
Df · Duk = λk f uk = λk (f, uk )L2 (Ω) ,
λk Ω Ω
Proof. Since −∆u = λu, and λu ∈ H01 , the regularity theorem that we saw in Chapter 2
yields u ∈ H 3 (Ω). But then, the same theorem implies u ∈ H 5 (Ω). Iterating the argument
we get that u ∈ H m (Ω) for all m ∈ N, and Morrey’s theorem yields the result.
Remark. If the boundary is not smooth, we still get u ∈ C ∞ (Ω). But regularity up to
the boundary requires some smoothness of the boundary.
We now show that eigenfunctions associated to the first eigenvalue have a strict sign.
Theorem 4.9. Let Ω ⊂ Rd be bounded and connected. Let u 6≡ 0 be a weak solution to
(4) −∆u = λ1 u in Ω, u = 0 in ∂Ω,
´ 2
´ 2 1
with λ1 = inf{ Ω |Du| / Ω u : u ∈ H0 (Ω), u 6≡ 0}. Then, either u > 0 or u < 0 in Ω.
Proof. We assume without loss of generality that kuk2 = 1. Let u+ = max(u, 0), u− =
min(u, 0). Let ˆ ˆ
α = (u ) , β = (u− )2 .
+ 2
Ω Ω
Then, α + β = 1, and u± ∈ H01 (Ω), with
( (
Du, a.e. in {u ≥ 0}, 0, a.e. in {u ≥ 0},
Du+ = Du− =
0, a.e. in {u ≤ 0}, −Du, a.e. in {u ≤ 0}.
We have
ˆ ˆ ˆ ˆ ˆ
λ1 = 2
|Du| = |Du | + |Du | ≥ λ1 (u ) + λ1 (u− )2 = (α + β)λ1 = λ1 .
+ 2 − 2 + 2
Ω Ω Ω Ω Ω
Hence, the inequality is an equality, and therefore
ˆ ˆ ˆ ˆ
+ 2
|Du | = λ1 (u ) , + 2
|Du | = λ1 (u− )2 .
− 2
Ω Ω Ω Ω
Therefore, both u+ and u− solve (4) in the weak sense. Hence, u± ∈ C ∞ (ω). We conclude
that |u| = u+ + u− ∈ C ∞ (Ω), and −∆|u| = λ1 |u| in the classical sense.
ffl Since λ1 |u| ≥ 0,
|u| ≥ 0 is superharmonic in Ω, −∆|u| ≥ 0 in Ω. Therefore, |u|(x) ≥ Br (x) |u| if Br (x) ⊂ Ω.
ffl
Hence, if there is some x̄ ∈ Ω such that |u|(x̄) = 0, then 0 = |u|(x̄) ≥ Br (x̄) |u| ≥ 0, which
SEMILINEAR ELLIPTIC EQUATIONS 15
implies that |u| ≡ 0 in Br (x̄). We have thus proved that the set A = {x ∈ Ω : u(x) = 0}
is open. Since u is continuous, it is also closed. Therefore, since Ω is connected, either
A = Ω, or A = ∅. The first option is not possible, since u 6≡ 0, hence the result.
Remark. In the course of the proof we have proved the Strong Maximum Principle for
superharmonic functions: if u ∈ C 2 (Ω) is nonnegative and superharmonic in a connected
set Ω, then it is strictly positive in Ω, unless it is identically 0.
To conclude this section we prove that the first eigenvalue is simple.
d
Theorem 4.10 (The first eigenvalue is simple). Let ´ Ω ⊂ 2R ´ be 2bounded 1and con-
nected. Let u 6≡ 0 be a weak solution to (4) with λ1 = inf{ Ω |Du| / Ω u : u ∈ H0 (Ω), u 6≡
0}. Then, u = αu1 for some α ∈ R, α 6= 0.
´
Proof. Since u1 has a strict sign´in Ω, we´know that Ω u1 6= 0. Therefore, there exists a
value χ´ ∈ R, χ 6= 0, such that Ω u = χ Ω u1 (remember that also u has a strict sign).
Then Ω (u − χu1 ) = 0. But this is not possible, since u − χu1 is also a weak solution
to (4) and hence has a strict sign in Ω.
Let us now apply the above theory to the study of a nonlinear problem.
Let Ω ⊂ Rd be bounded and smooth, and a ∈ C ∞ (Ω). We consider the Dirichlet
problem
(5) −∆u = a|u|p−1 u in Ω, u = 0 in ∂Ω.
This problem has a classical solution, namely u = 0. Are there others (maybe weak)? If
a ≤ 0 this is not possible. Indeed, using the solution itself as test function we get
ˆ
(|Du|2 − a|u|p+1 ) = 0.
Ω
Therefore, if a ≤ 0, then Du = 0 almost everywhere in Ω. Hence, u is constant in each
connected component of Ω. Since u ∈ H01 (Ω), then it has to be identically 0. What
happens if a > 0 in Ω? (We do not consider here the interesting case in which a changes
sign.)
´ ´
Let F, G : H01 (Ω) → R be given by F [u] = 21 Ω |Du|2 , G[u] = Ω a|u|p+1 − 1. We
want to minimize F in the 0-level set of G, C = {u ∈ H01 (Ω) : G[u] = 0}. Notice that the
elements of C are nontrivial.
From now on we assume that d ≥ 3 (it is a good exercise to analyze the cases d = 1, 2).
∗ 2d
2d
Since H01 (Ω) ⊂ L2 (Ω) = L d−2 (Ω), then G is well defined if 2 ≤ p + 1 ≤ d−2 , that is, if
1
p ∈ [1, (d + 2)/(d − 2)]. Under this restriction we alreay know that F, G ∈ C (X; R).
It is easy to check that C 6= ∅. Indeed, take any ū ∈ H01 (Ω), ū 6≡ 0. Then
ˆ
p+1
G[λū] = λ a|ū|p+1 − 1 = 0
Ω
´ p+1 −1/(p+1)
if we take λ = ( Ω a|ū| ) . Therefore, I = inf{F [u] : u ∈ C} ≥ 0. Let (uk )∞ k=1 be
a minimizing sequence in C, uk ∈ C, F [uk ] → I. Then, I ≤ F [uk ] ≤ I + 1 if k is large,
from where we deduce that kuk kH01 (Ω) ≤ K < ∞. Since H01 (Ω) is reflexive, there are a
subsequence (ukj )∞ 1 1
j=1 and a function u ∈ H0 (Ω) such that ukj * u in H0 (Ω). Since F
is sequentially weakly lower semicontinuous (it is a norm in a Hilbert space), the weak
convergence implies that F [u] = I. On the other hand, if p ∈ [1, (d + 2)/(d − 2)), then
p + 1 ∈ [2, 2d/(d − 2)). Hence, Rellich-Kondrachov’s compactness theorem implies that
16 F. QUIRÓS
Taking τ = (µ(p + 1))1/(p−1) (this requires p > 1), we have µ(p+1)τ p−1
= 1, and hence w ∈
1
H0 (Ω) is a weak solution to (5). We still have to check that it is nontrivial which will
´be immediate
2
if µ >´ 0. Top+1
show that we take v = u as test function and we obtain
Ω
|Du| = µ(p + 1) Ω
a|u| . This implies that µ ≥ 0. Morevoer, if µ = 0, then Du = 0
almost everywhere in Ω and hence u is constant in each connected component of Ω. Since
u ∈ H01 (Ω), then u ≡ 0, a contradiction with u ∈ C.
Summarizing, we have proved that problem (5) with a > 0 in Ω has a nontrivial weak
solution if p ∈ (1, (d + 2)/(d − 2)).
Remark. The upper restriction is not technical. Indeed, by means of the so-called
Derrick-Pohozaev’s identity one can prove that the problem does not admit a nontrivial
solution if a ≡ 1, p > (d + 2)/(d − 2) and the domain is a ball.
Besides local or global minima and maxima, there are other types of critical points:
saddle-points.
Example. Let f : R2 → R be given by f (x, y) = x2 − y 2 . The point (0, 0) is a critical
point of f , that is, Df (0, 0) = 0. However, it is neither a local maximum nor a local
minimum. The graph of f looks like a horse saddle, hence (0, 0) is called a saddle-point.
If we move along the graph in the direction of the x-axis, (0, 0) is the minimum, while if
we move along the graph in the direction of the y-axis, (0, 0) is the maximum. For this
reason, we also call (0, 0) a mini-max critical point of f .
Idea to find saddle-points. Given two points separated by a mountain range, there is a
mountain pass that allows to go from one of the points to the other “climbing the less
possible”.
Let us formulate this idea mathematically. Let X be a real Banach space and let
I : X → R be a C 1 functional such that
(i) I[0] = 0;
(ii) there exist constants ρ, α > 0 such that I ≥ α;
∂Bρ (0)
(iii) there exists x̄ ∈ X \ Bρ (0) such that I[x̄] ≤ 0.
We consider a path connecting 0 and x̄,
γ : [0, 1] → X continuous, such that γ(0) = 0, γ(1) = x̄.
Since γ has to cross ∂Bρ (0),
max I[g(t)] ≥ α.
t∈[0,1]
SEMILINEAR ELLIPTIC EQUATIONS 17
In finite dimensional spaces, the critical value c is achieved for some γ0 ∈ Γ and t0 ∈ [0, 1].
What is more important, x0 := γ0 (t0 ) is a critical point of I, I 0 [x0 ] = 0 (the tangent plane
at a mountain pass is horizontal). The proof uses strongly that bounded sequences have
a convergent subsequence. This is no longer true in infinite dimensional Banach spaces.
In order to be able to follow this approach, we need to require that the functional satisfies
some compactness condition.
Definition 5.1. We say that the functional I satisfies the Palais-Smale condition (in
short, (PS)), if any sequence (uk )∞ 0
k=1 in X for which I[uk ] is bounded and I [uk ] → 0 in
∗
X as k → ∞ possesses a subsequence converging strongly in X.
Notation. Sequences like the ones appearing in the definition are known as Palais-Smale
sequences (in short, PS sequence).
Example. Let Ω ⊂ Rd be bounded and smooth. ´ Given
´ f ∈ L2 (Ω), we consider the
functional I : H01 (Ω) → R given by I[u] = 12 Ω |Du| 2
´ − Ω f u. We want to check that I
1
satisfies the condition (PS). We will take (u, v) = Ω Du · Dv as inner product in H0 (Ω).
´
Let (uk )∞ 1
k=1 be a PS sequence. On the one hand, C ≥ I[uk ] = 2 kuk kH01 (Ω) − Ω f uk . On
the other hand, using the inequalities of Hölder, Poincaré, and Peter-Paul,
ˆ
f uk ≤ kf kL( Ω) kuk kL( Ω) ≤ Ckuk kH01 (Ω) ≤ εkuk k2H 1 + Cε .
0
Ω
kuk kH01 (Ω) ≤ K < ∞. Using Rellich-Kondrachov’s compactness theorem and the reflex-
ivity of the space, we get that (uk )∞ ∞
k=1 has a subsequence (ukj )j=1 such that ukj → u in
L2 (Ω), ukj * u in H01 (Ω). This is not enough for our purposes. But we have not yet used
that I 0 [uk ] → 0 in H −1 (Ω).
Let v ∈ H01 (Ω). Then
ˆ
(Duk · Dv − f v) = |I 0 [uk ]v| ≤ kI 0 [uk ]kH −1 (Ω) kvkH 1 (Ω) .
j j j 0
Ω
the above information plus the strong convergence of ukj in L2 (Ω) yields kukj kH01 (Ω) →
kukH01 (Ω) . Weak convergence plus convergence of the norms finally yields the desired
strong convergence in H01 (Ω). Indeed,
kukj − uk2H 1 (Ω) = (ukj − u, ukj − u)H01 (Ω) = kukj k2H 1 (Ω) − 2(ukj , u)H01 (Ω) + kuk2H 1 (Ω)
0 0 0
We don’t give here the proof of the theorem. The interested student can find it, for
instance, in [2].
Remark. I[x0 ] ≥ α > 0 implies that x0 6≡ 0. So, this method gives us nontrivial solutions.
Example. We consider once more the Dirichlet problem (5) with a ∈ C ( Ω), a > 0 in Ω.
We know that u ≡ 0 is a classical solution. We want to find a second weak solution, that
is, a nontrivial function u ∈ H01 (Ω) such that
ˆ
(Du · Dv − a|u|p−1 uv) = 0 for all v ∈ H01 (Ω).
Ω
We will do this by means of the Mountain Pass Theorem. As we will check later, this is
the Euler-Lagrange equation of the functional I : H01 (Ω) → R, I[u] [u] − I2 [u], where
´ = I1p+1
1 1 2 1
I1 , I2 : H0 (Ω) → R are given by I1 [u] = 2 kDukH 1 (Ω) , I2 [u] = p+1 Ω a|u| .
0
We will do the analysis for d ≥ 3. It is good exercise to do the computations for the
cases d = 1, 2.
Step 1: The functional is well defined. Let p ≥ 1 be such that p+1 ≤ 2∗ = 2d
d−2
.
Then, applying Hölder’s inequality and the GNS Sobolev embedding,
ˆ ˆ
a|u|p+1
≤ kakL∞ (Ω) |u|p+1 ≤ Ckukp+1 p+1
2∗ ≤ CkukH 1 (Ω) .
0
Ω Ω
2d d+2
Therefore, the functional is well defined if 1 ≤ p ≤ d−2
−1= d−2
.
Step 2: Computation of directional derivatives. Let u, v ∈ H01 (Ω). Then,
∂v I[u] = limε→0 I[u+εv]−I[u]
ε
if the limit exists. We have on the one hand
1
´ ´ 2 ´ ´ ˆ
2 Ω
|Du|2 + ε Ω Du · Dv + ε2 Ω |Dv|2 − 21 Ω |Du|2
∂v I1 [u] = lim = Du · Dv.
ε→0 ε Ω
On the other hand ´
Ω
a(|u + εv|p+1 − |u|p+1 )
∂v I2 [u] = lim .
ε→0 (p + 1)ε
p+1
Let g(t) = |t|p+1 with p ≥ 1. Then g 0 (t) = |t|p−1 t and g 00 (t) = p|t|p−1 . Therefore,
performing a Taylor expansion around u,
|u + εv|p+1 |u|p+1 p|u + θεv|p−1 ε2 v 2
= + |u|p−1 uεv +
p+1 p+1 2
SEMILINEAR ELLIPTIC EQUATIONS 19
We conclude that ´v 7→ ∂v I[u] is bounded, and hence that I is Gâteaux differentiable, with
I 0 [u]v = ∂v I[u] = Ω (Du · Dv − a|u|p−1 uv). Then, the Euler-Lagrange equation associated
to I is ˆ
(Du · Dv − a|u|p−1 uv) = 0 for all v ∈ H01 (Ω),
Ω
(u ∈ H01 (Ω)) which is nothing but the weak formulation of problem (5).
Step 4: I is Frechet differentiable. A straightforward computation shows that
1 1
I1 [u + v] − I1 [u] − I10 [u]v = kDvk22 = kvk2H 1 (Ω) = o(kvkH01 (Ω) ) as kvkH01 (Ω) → 0.
2 2 0
as kvkH01 (Ω) → 0.
Since both I1 and I2 is Frechet differentiable, the same is true for I = I1 − I2 .
20 F. QUIRÓS
Therefore, kI10 [u] − I10 [v]kH −1 (Ω) ≤ ku − vkH01 (Ω) , which shows that I1 ∈ C 1 (H01 (Ω); R).
As for I2 ,
ˆ ˆ
|(I20 [u] I20 [v])w|
p−1 p−1
p−1
− = a(|u| u − |v| v)w ≤ kak∞ |u| u − |v|p−1 v |w| .
Ω Ω
Then, using the Mean Value Theorem, we have for some θ ∈ [0, 1]
ˆ ˆ
p−1 p−1
|θu + (1 − θ)v|p−1 |u − v||w|
|u| u − |v| v |w| = p
Ω
ˆΩ
(|u|p−1 + |v|p−1 )|u − v||w|
≤C
Ω
≤ Ck(|u|p−1 + |v|p−1 )|u − v|k 2d kwk2∗
d+2
≤ C(kukp−1 p−1
2∗ + kvk2∗ )ku − vk2∗ kwk2∗
≤ C(kukp−1
H01 (Ω)
p−1
+ kvkH 1 )ku − vkH01 (Ω) kwkH01 (Ω) .
0 (Ω)
Let f (ρ) = 12 ρ2 − Cρp+1 . If p > 1, the cuadratic term dominates for small values of ρ,
f (ρ) ≥ 41 ρ2 if ρ ∈ [0, ρ0] for some small ρ0 > 0. We conclude that there exists a constant
ρ̄ ∈ (0, ρ0 ) such that I ≥ α for some constant α > 0.
∂Bρ̄ (0)
if λ is positive and large enough. We can moreover take λ large so that kλūkH01 (Ω) > ρ̄.
Step 7: (PS) condition. Let (uk )∞ k=1 be a PS sequence, that is, I[uk ] is bounded and
0 −1
I [uk ] → 0 in H (Ω). We have to prove that there is a strongly convergent subsequence.
7.1. The sequence is bounded. We have
ˆ
1 1
C ≥ I[uk ] = kuk k2H 1 (Ω) − a|uk |p+1 .
2 0 p+1 Ω
SEMILINEAR ELLIPTIC EQUATIONS 21
Thus,
ˆ
1 1
kuk k2H 1 (Ω) ≤ a|uk |p+1 + C.
(6)
2 0 p+1 Ω
0 −1
On the other hand, since I [uk ] → 0 in H (Ω),
ˆ
Duk · Dv − a|uk | uk v = |I 0 [uk ]v| ≤ kI 0 [uk ]kH −1 (Ω) kvkH01 (Ω) ≤ kvkH01 (Ω)
p−1
Ω
and hence ˆ
a|uk |p+1 ≤ kuk k2H 1 (Ω) + kuk kH01 (Ω) .
0
Ω
Combining this estimate with (6) we obtain
1 1 1
− kuk k2H 1 (Ω) − kuk kH01 (Ω) ≤ C.
2 p+1 0 p+1
Therefore, if p > 1, necessarily kuk kH01 (Ω) ≤ K < ∞.
7.2. Strongly convergent subsequence. Since kuk kH01 (Ω) ≤ K, there are a subsequence
(ukj )∞ 1 1
j=1 and a function u ∈ H0 (Ω) such that ukj * u in H0 (Ω), ukj → u in L
p+1
(Ω). We
∗
are assuming that p < (d + 2)/(d − 2). Hence p + 1 < 2 .
We claim that (a|ukj |p−1 ukj converges in H −1 (Ω) towards a|u|p−1 u. Indeed, using the
Mean Value Theorem we get
ˆ ˆ
p−1 p−1
p−1 p−1
a(|u k | uk j
− |u| u)v ≤ kak∞ p
|v|(|uk j
| + |v| )|uk j
− u|
Ω Ω
p−1
≤ Ckv(|ukj | + |v|p−1 )k p+1 kukj − ukp+1
p
p−1 p−1
≤ Ck|ukj | + |v| k p+1 kvkp+1 kukj − ukp+1
p−1
≤ C(kukj kp−1
H01 (Ω)
p−1
+ kvkH 1 )kvkH01 (Ω) kukj − ukH01 (Ω)
0 (Ω)
p−1
≤ Ckukj − ukH01 (Ω) kvkH 1 (Ω) ,
0
that is, −∆ukj = I 0 [ukj ]+a|ukj |p−1 ukj in H −1 (Ω). But we know that ∆ : H01 (Ω) → H −1 (Ω)
is an isometry. Therefore, using the claim,
ukj = −∆−1 (I 0 [ukj ]) − ∆−1 (a|ukj |p−1 ukj ) → −∆−1 (a|u|p−1 u) in H01 (Ω),
so that the subsequence is strongly convergent.
Conclusion. The hypotheses of the Mountain Pass Theorem are fulfilled, and hence
the functional has a nontrivial critical point.
22 F. QUIRÓS
References
[1] Ambrosetti, A.; Rabinowitz, P. H. Dual variational methods in critical point theory and applications.
J. Functional Analysis 14 (1973), 349–381.
[2] Chen, W.; Li, C. “Methods on nonlinear elliptic equations”. AIMS Series on Differential Equations
& Dynamical Systems, 4. American Institute of Mathematical Sciences (AIMS), Springfield, MO,
2010. ISBN: 978-1-60133-006-2; 1-60133-006-5.
[3] Peral, I. Curso “Métodos variacionales y ecuaciones en derivadas parciales”. Universidad de Almerı́a,
1998. Disponible en http://matematicas.uam.es/ ireneo.peral/ALMERIA1.pdf.