LECTURE NOTES

„LINEAR PARTIAL DIFFERENTIAL EQUATIONS“

MARTIN KELL
TÜBINGEN, JULY 25, 2017
MARTIN.KELL@MATH.UNI-TUEBINGEN.DE

Contents
1. Classes of PDEs 2
2. Harmonic functions on Rn 5
2.1. The mean value property, maximum principle and Harnack’s inequality 5
2.2. Poisson’s formula for solutions on the ball 9
2.3. Convergence theorems for harmonic functions 10
2.4. Gradient estimates 11
2.5. Constructing harmonic functions in general domains 12
3. Classical Maximum Principles 17
3.1. Elliptic maximum principles 17
3.2. Parabolic maximum principles 21
4. Sobolev theory in Rn 24
4.1. Banach spaces 24
4.2. Function spaces 24
4.3. Properties of Lp -spaces 25
4.4. Sobolev spaces and minimizers of quadratic functionals 27
4.5. Weak derivatives 35
4.6. Higher order Sobolev spaces 39
4.7. Poincaré inequality, and embeddings of Sobolev and Morrey. 40
4.8. Lax–Milgram Theorem 44
4.9. (to be written) Trace and Extension operators 46
4.10. (partially covered in exercise) A weak topology via convex sets 46
4.11. Difference quotients of Sobolev functions. 48
4.12. (not covered) Maximal Function Theorem 49
5. L2 -Estmates 52
6. Properties of weak (sub)solution 56
7. Schauder esimates 58
8. Method of Continuity 63
Appendix A. Topology 65
Appendix B. Measure and Integration Theory 67
Appendix C. Polar coordinates. 68
Appendix D. List of definition/symbols 68

1
 LECTURE NOTES „LINEAR PARTIAL DIFFERENTIAL EQUATIONS“ 2

Note: The following notes are written for the lecture “Linear Partial Differential
Equations” during the summer semester 2017 at the University of Tübingen.

1. Classes of PDEs
Definition 1.1 (Partial Differential Equation). Let Ω ⊂ Rn be an open domain
k k−1
and F a function on Rn × Rn × · · · Rn × R × Ω. Then an equation of the form
F (Dk u(x), Dk−1 u(x), . . . , Du(x), u(x), x) = 0 for all x ∈ Ω
for an unknown function u ∈ C k (Ω) is called a partial differential equation of k-th
order.
• The equation is called linear, if F is linear in all but the last entry. In this
case the equation transformed into the following equation
X
aγ (x)∂γ u(x) = f (x)
|γ|≤k

where γ is a multi-index and ∂γ = ∂γ1 · · · ∂γl where γ = (γ1 , . . . , γl ).

• The equation is called semi-linear, if F is lienar in the first entry. In this
case the equation is of the form
X
aγ (x)∂γ u(x) + F̃ (Dk−1 u(x), . . . , u(x), x) = 0 for all x ∈ Ω
|γ|=k
k−1
where F̃ is a function on Rn × · · · × R × Ω.
k−1
• The equation is called quasi-linear, if there are function aγ , F̃ : Rn ×· · ·×
R × Ω → R such that the equation can be transformed into the following
X
aγ (Dk−1 u(x), . . . , u(x), x)∂γ u(x)+F̃ (Dk−1 u(x), . . . , u(x), x) = 0 for all x ∈ Ω.
|γ|=k

In the course of the lecture we will focus mainly on linear partial differential
equations of second order. The general equation is then given as follows: Let
Ω ⊂ Rn be an open and connected domain, ai,j : Ω → R, bk : Ω → R and
c : Ω → R (continuous) functions on Ω where i, j, k ∈ {1, . . . , n}. Then for a
function u ∈ C 2 (Ω) we define an operator L : C 2 (Ω) → C 0 (Ω) by
Xn X
Lu(x) = ai,j (x)∂i ∂j u(x) + bk (x)∂k u(x) + c(x)u(x).
i,j=1 k=1

Remark. One may verify that Lu(x) = L̃u(x) for all u ∈ C 2 (Ω) if ai,j is replaced
1 i,j
by its symmetrization ai,j
sym = 2 (a + aj,i ). Thus without loss of generality we can
i,j n
assume that the matrix (a (x))i,j=1 is a symmetric matrix.
To get a partial differential equation one needs in addition a function f ∈ C 0 (Ω)
and asks whether u ∈ C 2 (Ω) satisfies the equation
Lu = f in Ω.
The most natural one is to look for (non-trivial) functions u such that Lu = 0. As
L is linear one can also ask whether there are (non-trivial) u and a λ ∈ R such that
Lu = λu in Ω
in which case we call u an eigenfunction of L. Instead of solving Lu = λu one may
equivalently solve the equation Lλ u = 0 where Lλ = L − λ id.
 LECTURE NOTES „LINEAR PARTIAL DIFFERENTIAL EQUATIONS“ 3

Example 1.2 (Classical Examples). Classical linear PDEs are given as follows
Pn
Laplace equation: Let L = ∆ := i=1 ∂ii then
∆u = 0 in Ω
is called the Laplace equation and solutions u are called harmonic function.
Poisson equation: More generally, one may look the Poisson equation
∆u = f in Ω
n
Heat equation: Assume ΩT ⊂ R × R where the first coordinate will be denoted
by t and ∆ is defined as above. Then the following is called the wave
equation
∂t u − ∆u = 0 in ΩT
Wave equation: As above ΩT ⊂ R × R:
∂tt u − ∆u = 0 in ΩT
A class of examples of non-linear PDEs a
Non-linear Laplace equation:
∆u = f (u) in Ω
where f : R → R is a function. This is an example of a quasi-linear PDE.
p-Laplace equation:
div(|∇u|p−2 ∇u) = 0 in Ω.
This is an example of a semi-linear PDE.
Minimal surface equation:
!
∇u
div p = 0 in Ω.
1 + |∇u|2
Monge-Ampère equation:
det D2 u = 0 in Ω.

Note that just asking for Lu = 0 gives in general an underdetermined system if

∂Ω is non-empty. This is best observed in the one-dimensional setting.
Example (One dimensional). In case n = 1, Ω = (a, b) for a < b ∈ R and the
operator L is of the form
Lu = au00 + bu0 + cu
for functions a, b, c : (a, b) → R. Hence the linear partial differential equation of
second order
Lu = f
is just a linear ordinary differential equation of second order. For a ≡ 1 and
b ≡ c ≡ f ≡ 0 the equation
u00 = 0
is solved by all affine functions, i.e. for a0 , b0 ∈ R
u = a0 x + b0
00
satisfies u = 0. Note that this is obviously underdetermined so that it is natural
to find solution with given boundary values, i.e. find u such that u(a) = g(a) and
u(b) = g(b) with for some g : {a, b} → R, or shorter written as u ∂Ω = g.
 LECTURE NOTES „LINEAR PARTIAL DIFFERENTIAL EQUATIONS“ 4

Thus also in the higher dimensional setting it is natural to ask for solution
u ∈ C 2 (Ω) ∩ C 0 (Ω̄) that satisfy
(
Lu = f in Ω
u ∂Ω = g.
A solution of a PDE with boundary data g is usually called a solution to the
Dirichlet problem (with boundary data g).
Remark. By linearity it is possible to focus only on case where either f ≡ 0 or g ≡ 0
(given that there are sufficiently many solutions u ∈ C 2 (Ω) ∩ C 0 (Ω̄) with Lu = f
and resp. u ∂Ω = 0:
(1) If ũ ∈ C 2 (Ω) ∩ C 0 (Ω̄) satisfies Lũ = f and v ∈ C 2 (Ω) ∩ C 0 (Ω̄) solves
Lu = 0 with boundary data g̃ = g − ũ ∂Ω then u = v + ũ solves Lu = f
with boundary data g..
(2) If û ∈ C 2 (Ω) ∩ C 0 (Ω̄) satisfies û ∂Ω = g and v ∈ C 2 (Ω) ∩ C 0 (Ω̄) solves
Lu = f˜ with boundary data g ≡ 0 where f˜ = f − Lû then u = v + ũ solves
Lu = f with boundary data g.
In general the equation Lu = 0 with given boundary data might still be unsolv-
able if the highest order coefficients are too general. As we can assume the matrix
(ai,j )ni,j=1 is symmetric, it can be diagonalized so that we can define the following
three main categories of PDEs:
Elliptic: The operator L is called elliptic if for all x ∈ Ω the matrices (ai,j (x))ni,j=1
have positive eigenvalues1. This can be expressed by assuming there are
functions λ, Λ : Ω → (0, ∞) such that
n
X n
X n
X
λ(x) ξi ξi ≤ ai,j (x)ξi ξj ≤ Λ(x) ξi ξi
i=1 i=1 i=1

for all ξ ∈ Rn and all x ∈ Ω. If λ and Λ can be chosen independently of

x ∈ Ω then we say the operator is uniformly elliptic.
Hyperbolic: The operator L is called hyperbolic if for all x ∈ Ω the matrices
(ai,j (x))ni,j=1 non-zero eigenvalues. Note that if x 7→ ai,j (x) is continuous
then the number of positive (and resp. negative) eigenvalues remains con-
stant along Ω by connectedness of Ω and continuity of the spectrum of the
matrices (ai,j (x))ni,j=1 . In a simpler setting, one only looks at the class
of hyperbolic PDEs where a1,1 (x) = −1, a1,j (x) = aj,1 (x) = 0 such that
(ai,j (x))ni,j=2 is elliptic and independent of the first coordinate. In that case
a PDE of the form ∂tt − Lu = f in a domain ΩT ⊂ R × Rn where L is an
elliptic operator L.
Similarly one could look at operators Lu on domains ΩT ⊂ R × Rn where ai,j , bk
and c depend on the first coordinate t. Note, however, that an equation
Lu = f
would be just a time-dependent elliptic equation. An alternative is to add a deriv-
ative ∂t u in the equation and obtain the following:

1resp. negative eigenvalues as we might switch signs and replace L by −L.

 LECTURE NOTES „LINEAR PARTIAL DIFFERENTIAL EQUATIONS“ 5

Parabolic: A parabolic PDE on the domain ΩT ⊂ R×Rn is of the form ∂t −Lu = 0

where L (resp. it’s coefficients ai,j , bk and c) are allowed to depend on t,
in which case it would be call a time-dependent parabolic PDE.
The focus of the lecture is on elliptic and parabolic PDEs as many techniques are
similar.

2. Harmonic functions on Rn
In this section we study harmonic functions on domain Ω ⊂ R.
Definition 2.1 (Harmonic function). A function u ∈ C 2 (Ω) is called subharmonic
if ∆u ≥ 0 and it is called superharmonic if ∆u ≤ 0. If it is both sub- and super-
harmonic then we say u is harmonic.
Observe that the sum of two subharmonic functions is itself subharmonic.
2.1. The mean value property, maximum principle ffland Harnack’s ´ in-
1
equality. In the following we frequently use the notation A f dµ = µ(A) A
f dµ
for
ffl a measure µ.
´ In case µ is the Lebesgue measuren then this is written as
A
f dx = λn1(A) A f dx.
Theorem 2.2 (Mean value property). Assume u ∈ C 2 (Ω) is subharmonic. Then
for all Br (x) ⊂⊂ Ω it holds

u(x) ≤ u(z)dz
∂Br (x)
and
u(x) ≤ u(y)dy.
Br (x)

Remark. Later on we show that the mean value property is actually equivalent to
being subharmonic for all C 2 -functions.
Proof. In polar coordinates it holds
ˆ
n−1
u(z)dz = r u(x + rω)dω.
∂Br (x) Sn−1

Also note that for z = rω ∈ ∂Br (x) unit outer normal ν is given by ω. Hence
d
∂ν u(z) = u(x + sω) .
ds s=r
Then polar coordinates show
ˆ ˆ
d
∂ν u(z)dz = rn−1 u(x + sω) dω.
∂Br (x) Sn−1 ds s=r

Because u ∈ C 1 (Ω) we can pull out derivative under the integral and obtain
ˆ ˆ 
n−1 d
∂ν u(z)dz = r u(x + sω)dω .
∂Br (x) ds Sn−1 s=r
!
d
= rn−1 nωn u(z)dz
ds ∂Bs (x) s=r

where we used again the polar coordinate transformation and the fact that |∂Br | =
nωn rn−1 .
 LECTURE NOTES „LINEAR PARTIAL DIFFERENTIAL EQUATIONS“ 6

Now Green’s formula applied yields

ˆ
0≤ ∆u(y)dy.
Br (x)
ˆ
= ∂ν u(z)dz
∂Br (x)
!
n−1 d
=r nωn u(z)dz
ds ∂Bs (x) s=r

implying that
r 7→ u(z)dz
∂Br (x)
is non-decreasing in r 7→ (0, r0 ) for some r0 > r with Br0 (x) ⊂ Ω. Since u is
continuous at x it holds
u(x) = lim u(z)dz.
r→0 ∂Br (x)
This proves the first claim. To obtain the second claim note that
ˆ r ˆ rˆ ˆ
n n−1
ωn r u(x) = nωn s u(x)ds = u(z)dzds = u(y)dz.
0 0 ∂Bs (x) Br (x)

Because |Br (x)| = ωn rn we obtain the second claim. 

Corollary 2.3 (Strong Maximum Principle). For all subharmonic functions u ∈
C 2 (Ω) ∩ C 0 (Ω̄) it holds
sup u = sup u
Ω ∂Ω
and if for some x0 ∈ Ω it holds
u(x0 ) = sup u
Ω

then u is constant on Ω. In particular, if u ≥ 0 and u ∂Ω

= 0 then u ≡ 0.
Proof. Let x0 ∈ Ω̄ be such that u(x0 ) = supΩ u. If x0 ∈ ∂Ω there is nothing to
prove. In case x0 ∈ Ω there is a ball Br (x0 ) ⊂⊂ Ω such that the mean value
property holds for Br (x0 ). It suffices to show that u is constant.
Now the choice of x0 yields the following

sup u = u(x0 ) ≤ u(y)dy ≤ sup u.

Ω Br (x0 ) Ω

This, however, can only hold if u(y) = u(x0 ) for all y ∈ Br (x0 ) implying that u is
locally constant. As Ω is connected we see that u must be constant. 
Because ±u is subharmonic for each harmonic function u ∈ C 2 (Ω) ∩ C 0 (Ω̄) we
obtain the following corollary which gives also a uniqueness result.
Corollary 2.4. For all harmonic functions u ∈ C 2 (Ω) ∩ C 0 (Ω̄) it holds
sup |u| = sup |u|.
Ω ∂Ω

In particular, if u1 , u2 ∈ C 2 (Ω) ∩ C 0 (Ω̄) are harmonic in Ω with u1 ∂Ω

= u2 ∂Ω
then u1 ≡ u2 .
Proposition 2.5. Let u ∈ C 0 (Ω). Then the following are equivalent.
 LECTURE NOTES „LINEAR PARTIAL DIFFERENTIAL EQUATIONS“ 7

ffl
• for all Br (x) ⊂⊂ Ω it holds u(x) ≤ ∂Br (x) u(z)dz
ffl
• for all Br (x) ⊂⊂ Ω it holds u(x) ≤ Br (x) u(y)dy
• for all Br (x) ⊂⊂ Ω it holds u(x) ≤ h(x) where h is a harmonic function
on Br (x) with boundary data g ≥ u ∂B (x) .
r

2
If, in addition, u ∈ C (Ω) then either of the condition above is equivalent to u being
subharmonic.

Proof. Exactly as in the proof of the mean value property, the first property implies
the second by integration.
Assume the second property holds: then the maximum principle holds on Br (x)
for u − h whenever h is a harmonic function on Br (x) with boundary data u ∂B (x) .
r
Thus
sup u − h = sup u − h ≤ 0
Br (x) ∂Br (x)

implying that u ≤ h.
Let2 h is a harmonic function on Br (x) with boundary data u ∂B (x) . Then the
r
mean value property holds for h. Thus assuming the third property we get

u(x) ≤ h(x) = h(z)dz = u(z)dz.

∂Br (x) ∂Br (x)

It remains to show that either of the first three properties implies that u is
subharmonic if u ∈ C 2 (Ω). Assume by contradiction ∆u(x) < 0 for some x ∈ Ω.
Then there is an open ball Br (x) ⊂⊂ Ω such that ∆u(x) ≤ − < 0 for all y ∈ Br (x).
In that case u is superharmonic in Br (x). However, this implies that

u(y) ≥ udy 0
Bs (y)

for all Bs (y) ⊂⊂ Br (x) (actually for all Bs (y) ⊂ Br (x) since u ∈ C 0 (B̄r (x))).
However, this means

u(y) = udy 0 .
Bs (y)

Pick y ∈ Br (x) and define a function v ∈ C 2 (B̄r (x)) by


v (y) := u(y) + ky − xk2Euclid
n
and note that v (y) = u(y). Since ∆u ≤ − on Br (x) we have

∆v ≤ − + ∆ky − xk2Euclid ≤ 0,
n
i.e. v is superharmonic on Br (x). In particular, it holds

v (y) ≥ v dy 0
Bs (y)

2Strictly speaking we assume the existence of harmonic functions in balls with given boundary
data, see next section.
 LECTURE NOTES „LINEAR PARTIAL DIFFERENTIAL EQUATIONS“ 8

for all Bs (y) ⊂⊂ Br (x). However, this leads to the following contradiction
 0
0< ky − xk2Euclid dy 0
Bs (y) n

= v − udy 0 ≤ v (y) − u(y) = 0.

Bs (y)

The proposition shows allows us to define a weak form of (sub-/super-)harmonicity.
Definition 2.6 (Mean-value harmonic). A (bounded) measurable function u ∈
C 0 (Ω) is called mean-value subharmonic if for all Br (x) ⊂⊂ Ω

u(x) ≤ u(y)dy.
Br (y)

It is mean-value superharmonic if −u is mean-value subharmonic. If it is both

mean-value sub- and superharmonic we call it mean-value harmonic.
Remark. Later we prove that a mean-value harmonic function is indeed a C 2 -
function and hence harmonic.
The following is now a consequence of monotonicity of the averaged integral over
balls.
Lemma 2.7. Let {ui }i∈I be mean-value subharmonic functions in Ω for some finite
index set I and define
u(x) := sup ui (x).
i∈I
Then u is mean-value subharmonic.
Theorem 2.8 (Harnack’s Inequality). Assume u ∈ C 2 (Ω) is a non-negative har-
monic function in Ω. Then for each Ω0 ⊂⊂ Ω there a constant C = C(n, Ω0 , Ω)
such that
sup u ≤ C inf0 u.
Ω0 Ω

Proof. Let B4r (y) ⊂⊂ Ω then for all x1 , x2 ∈ Br (y) it holds

ˆ
1
u(x1 ) ≤ u(y)dy ≤ u(y)dy
Br (x1 ) |Br (x1 )| B2r (y)
and ˆ
1
u(x2 ) ≥ u(y)dy ≥ u(y)dy.
B3r (x1 ) |B3r (x2 )| B2r (y)
Since |B3r (x2 )| = 3n |Br (x1 )| this shows
u(x1 ) ≤ 3n u(x2 ).
In particular,
sup u ≤ 3n inf u.
Br (y) Br (y)

Let Ω ⊂⊂ Ω then there is a radius r > 0 such that for all y ∈ cl Ω0 it holds
0

B4r (y) ⊂⊂ Ω. Since cl Ω0 is compact there are finitely many y1 , . . . , yN such that
N
[
Ω0 ⊂ Br (y).
i=1
 LECTURE NOTES „LINEAR PARTIAL DIFFERENTIAL EQUATIONS“ 9

Thus for all x1 , x2 ∈ Ω0 there is a sequence z0 = x1 , z1 , . . . , zN = x2 ∈ Ω0 such that

for each i = 1, . . . , N there is an index ji ∈ {1, . . . , N } such that zi−1 , zi ∈ Br (yji ).
Thus we obtain
u(z0 ) ≤ 3n u(z1 ) ≤ 32n u(z2 ) ≤ · · · ≤ 3N n u(zN )
showing that
sup u ≤ 3N n inf0 u.
Ω0 Ω

Remark. The Harnack inequality also implies the strong maximum principle, in-
deed, if u ∈ C 0 (Ω̄) and M = supΩ u = u(x0 ) for some x0 ∈ M then M − u ≥ 0 is a
non-negative harmonic function. Thus for all Ω0 ⊂⊂ Ω with x0 ∈ Ω0 it holds
sup(M − u) ≤ inf0 (M − u) = 0
Ω Ω

implying u ≡ M .
2.2. Poisson’s formula for solutions on the ball.
Theorem 2.9 (Poisson’s formula). Let g ∈ C 0 (∂B1 (0)) be a continuous function
and define a function u : B̄1 (0) → R as follows
1−|x|2 ´
(
g(z)
nωn ∂B1 (0) |x−z|2
dz x ∈ B1 (0)
u(x) =
g(x) x ∈ ∂B1 (0).
Then u ∈ C ∞ (B1 (0)) ∩ C 0 (B̄1 (0)) is a harmonic function in B1 (0) with boundary
data g.
Proof. Exercise. 
0
Corollary 2.10. Let g ∈ C (∂Br (a)) be a continuous function and define a func-
tion u : B̄r (a) → R as follows
r −|x−a|2 ´
( 2
g(z)
rnωn ∂Br (a) |x−z|2
dz x ∈ Br (a)
u(x) =
g(x) x ∈ ∂Br (a).
Then u ∈ C ∞ (Br (a)) ∩ C 0 (B̄r (a)) is a harmonic function in Br (a) with boundary
data g.
This gives us immediately a regularity theorem for (mean value) harmonic func-
tions in general domains.
0
ffl If u ∈ C (Ω) satisfies the
Corollary 2.11. mean value property on all ball Br (x) ⊂⊂
Ω, i.e. u(x) = Br (x) u(y)dy, then u ∈ C ∞ (Ω) and u is harmonic in Ω.
Proof. It suffices that u ∈ C ∞ (Br (x)) for all Br (x) ⊂⊂ Ω. Now let v be the har-
monic function on Br (x) given by Poisson’s formula with boundary data u ∂Br (x) .
Then u−v is a still satisfies the mean value property function on all balls Bs (x0 ) ⊂⊂
Br (x) and hence the maximum principle on Br (x). However, (u − v) ∂Br (x) ≡ 0
implying
sup |u − v| = sup |u − v| = 0,
Br (x) ∂Br (x)
i.e. u ≡ v on Br (x). As v ∈ C ∞ (Br (x)) this yields the claim. 
Combined with Proposition 2.5 applied to ±u we get t he following.
 LECTURE NOTES „LINEAR PARTIAL DIFFERENTIAL EQUATIONS“ 10

Corollary 2.12. Let u ∈ C 0 (Ω). Then the following are equivalent.

ffl
• for all Br (x) ⊂⊂ Ω it holds u(x) = ∂Br (x) u(z)dz
ffl
• for all Br (x) ⊂⊂ Ω it holds u(x) = Br (x) u(y)dy
• for all Br (x) ⊂⊂ Ω it holds u(x) = h(x) where h is a harmonic function
on Br (x) with boundary data u ∂Br (x) .
• u is twice differentiable in Ω (i.e. u ∈ C 2 (Ω)) and harmonic in Ω (i.e.
∆u = 0)
• u is infinitely many times differentiable in Ω and harmonic in Ω.
2.3. Convergence theorems for harmonic functions. The first lemma is just
a consequence that convergence of second derivative implies convergence of the
Laplacian hence being harmonic is preserved under this strong form of convergence.
Lemma 2.13. If un ∈ C 2 (Ω) is a sequence of subharmonic functions that converges
locally uniformly in C 2 to a function u ∈ C 2 (Ω) then u is subharmonic.
Using Poisson’s formula this can be improved as follows:
Proposition 2.14. Assume un ∈ C 2 (Ω) is a sequence of harmonic functions con-
verging locally uniformly (in C 0 ) to a function u ∈ C 0 (Ω) then u ∈ C ∞ (Ω) is
harmonic.
Proof. Note that the mean value property is preserved under uniform convergence.
Thus the claim follows from Corollary 2.11. 

A slightly weaker (though also different) convergence result was obtained by

Harnack using the Harnack inequality.
Theorem 2.15 (Harnack’s Convergence Theorem). Let un ∈ C 2 (Ω) be a sequence
of harmonic functions such that un ≤ un+1 . If for some y ∈ Ω the sequence
{un (y)}n∈N is bounded then (un )n∈N converges locally uniformly to a harmonic
function u ∈ C 2 (Ω).
Proof. Since (un (y))n∈N is a bounded, non-decreasing sequence it is, in particular,
convergence and hence a Cauchy sequence. Thus for all  > 0 there is an N ∈ N
such that for all m ≥ n ≥ N it holds
0 ≤ um (y) − un (y) ≤ .
Let Ω ⊂⊂ Ω and C = C(n, Ω0 , Ω) be the constant given in Theorem 2.8. Since
0

um − un ≥ 0 is harmonic for all m ≥ n ≥ N we get

sup |um − un | = sup(um − un ) ≤ C inf(um − un ) ≤ C(um (y) − um (y)) ≤ C
Ω Ω

showing that (un )n∈N is a Cauchy sequence in C 0 (Ω0 ). In particular, un → u

uniformly on Ω0 where u(x) := limn∈N un (x). 

Going back to Poisson’s formula we observe the following: for all x ∈ Bs (a) ⊂⊂
1
Br (a) the function z 7→ |x−z| is uniformly bounded (by Cs,r ). Thus it suffices to
assume g ∈ L (∂Br (a)) to obtain a function u ∈ C ∞ (Br (a)) which is harmonic in
1

Br (a). In that case we still say u is the (unique) harmonic function with boundary
data g. Furthermore, the uniform convergence on the boundary data can be replace
by
 LECTURE NOTES „LINEAR PARTIAL DIFFERENTIAL EQUATIONS“ 11

Proposition 2.16. Let (gn )n∈N be a sequence of functions in L∞ (∂Br (a)) converg-
ing in L1 to a function g. If un is the (unique) harmonic function with boundary
data gn and u the one corresponding to g. Then the sequence (un )n∈∞ converges
locally uniformly to u in Br (a).
Proof. Let x ∈ Bs (a) then
ˆ
|u(x) − un (x)| ≤ rCr,s |g(z) − gn (z)|dz → 0
∂Br (a)

where the convergence of the right hand side does not depend on x. Hence un → u
uniformly on Bs (a) which implies the claim as s < r can be chosen arbitrary. 

Corollary 2.17. If (gn )n∈N is a non-decreasing sequence of functions in L∞ (∂Br (a))

such that |gn | ≤ C. Then un converges locally uniformly to a harmonic function
u ∈ C ∞ (Br (a)).
Proof. Let g = lim gn then by the Monotone Convergence Theorem it holds
ˆ ˆ
|g − gn |dz = g − gn dz → 0,
∂Br (a) ∂Br (a)
1
i.e. gn → g in L (Ω). Then the previous proposition yields the claim. 

2.4. Gradient estimates. Note by linearity if u ∈ C 2 (Ω) is harmonic in Ω then

for Br (x) ⊂⊂ Ω it holds
!
∂i u(x) = ∂i udz = ∂i udz.
Br (x) Br (x)

In particular, ∂i u is harmonic in Ω for all i = 1, . . . , n. Furthermore, it holds

n n
∇u(x) = ∇u(z)dz = u·ν ≤ sup |u|.
Br (x) r ∂Br (x) r ∂Br (x)
Thus
n
|∇u|(x) ≤ sup |u|
dx Ω
where
dx = d(x, ∂Ω) = inf{d(x, y) | y ∈ ∂Ω}.
Lemma 2.18. If u is harmonic bounded from above and below by a constant D
then for each Ω0 ⊂⊂ Ω there is a constant C = C(n, D · d(Ω0 , Ω)) such that u is
Lipschitz continuous on Ω̄0 with Lipschitz constant bounded by C.
Combining Arzela–Ascoli and Proposition 2.14 gives the following corollary.
Corollary 2.19. Let (un )n∈N be a sequence of C 2 -harmonic functions in Ω that
is uniformly bounded above and below by a constant D then there is a subsequence
(unk )nk ∈N such that unk converges locally uniformly in C 0 (Ω) to a harmonic func-
tion u ∈ C ∞ (Ω).
Remark. If (un )n∈N is non-decreasing and bounded then the sequence converges
uniformly and gives yet another proof of Harnack’s Convergence Theorem (Theorem
2.15).
 LECTURE NOTES „LINEAR PARTIAL DIFFERENTIAL EQUATIONS“ 12

An iterated argument also shows that ∂γ u is harmonic for any multi-index γ.

Then as above we obtain
|Dk u|(x) ≤ C sup |u| x ∈ Ω0 ⊂⊂ Ω
Ω
0
where C = C(k, n, Ω , Ω). Looking at (∂γ u)|γ|<k would show that a bounded se-
quence of harmonic functions is actually C k -compact in any Ω0 ⊂⊂ Ω.

2.5. Constructing harmonic functions in general domains. The Poisson in-

tegral formula shows that it is possible to solve the Dirichlet problem on any ball
with any given continuous (resp. L1 ) data. In this section we use this information
to construct harmonic functions on more general domains Ω and show that they
satisfy certain boundary regularity if ∂Ω is not too bad.

Perron’s method of subharmonic functions. Let Ω be an open, bounded and con-

nected domain. Given g ∈ C 0 (∂Ω) we define the following two sets
S− = S− (g, Ω) = {u ∈ C 0 (Ω̄) | u is mean-value subharmonic in Ω and u ∂Ω
≤ g}
0
S+ = S+ (g, Ω) = {u ∈ C (Ω̄) | u is mean-value superharmonic in Ω and u ∂Ω
≥ g}.

Note that any function u ∈ S− ∩ S+ ⊂ C 0 (Ω̄) would be mean-value harmonic in Ω

(hence harmonic in Ω) and satisfies u ∂Ω = g and by uniqueness S− ∩ S+ contains
at most one element.
Also note that for any u± ∈ S± the function u− − u+ is mean-value subharmonic
with non-positive boundary data hence u− ≤ u+ . Since the function v− ≡ inf g is
in S− and the function v+ ≡ sup g is in S+ we also have
inf g ≤ u− ≤ u+ ≤ sup g
∂Ω ∂Ω

for all u± ∈ S± .
Thus it is natural to look at the following
u∗ (x) = sup u− (x) ∈ [inf g, sup g]
u− ∈S− ∂Ω ∂Ω
∗
u (x) = inf u+ (x) ∈ [inf g, sup g].
u+ ∈S+ ∂Ω ∂Ω

Perron’s Method is to show the following:

Step 1: Show that u∗ (and thus also u∗ ) is harmonic.
Step 2: Show u∗ ∂Ω = g (and thus u∗ = u∗ ).
Before we start with the Step 1 we shows that a function stays mean-value sub-
harmonic if we replace it locally a harmonic function of given boundary data.
Lemma 2.20 (Replacement Lemma). Let u ∈ C 0 (Ω̄) be a mean-value subharmonic
function in Ω and Br (x) ⊂⊂ Ω. If h is a harmonic function in Br (x) with boundary
data u ∂Ω then the following function
(
h(x) x ∈ B̄r (x)
ũ(x) =
u(x) x ∈ Ω\B̄r (x).

Is continuous in Ω̄ and mean-value subharmonic in Ω.

 LECTURE NOTES „LINEAR PARTIAL DIFFERENTIAL EQUATIONS“ 13

Proof. Let Bs (y) ⊂⊂ Ω be an arbitrary ball and h be a harmonic function with

ĥ ∂Bs (y) = ũ ∂Bs (y) . We need to show ĥ ≥ ũ on B̄s (y).
Since u is continuous and ũ ≥ u, it holds ĥ ≥ u on B̄s (y) by subharmonicity of
u. Thus it suffices to show ĥ ≥ h on Bs (x) ∩ Br (x).
Note that ∂(Bs (y)∩Br (x)) = ∂Bs (y)∩B̄r (x)∪∂Br (x)∩B̄s (y). On ∂Bs (y)∩B̄r (x)
it holds ĥ = ũ = h. And if ỹ ∈ ∂Br (x) ∩ B̄s (y) then h(ỹ) = u(ỹ). Since u − ĥ ≤ 0
on B̄s (y) we have by the maximum principle
sup h − ĥ = sup h − ĥ
B̄s (y)∩B̄r (x) ∂(Bs (y)∩Br (x))
( )
= max sup (h − ĥ), sup (h − ĥ)
∂Bs (y)∩B̄r (x) ∂Br (x)∩B̄s (y)
( )
= max sup (ĥ − ĥ), sup (u − ĥ) ≤0
∂Bs (y)∩B̄r (x) ∂Br (x)∩B̄s (y)

implying ĥ ≥ h on B̄s (y) ∩ B̄r (x) and thus ĥ ≥ ũ on B̄s (x). 

Theorem 2.21 (Perron’s Method). The functions u∗ and u∗ are harmonic in Ω.
Proof. It suffices to show that u∗ is harmonic on each Br (x) ⊂⊂ Ω. Let vn ∈ S−
be such that
u∗ (x) = lim vn .
n→∞
Then
0
vn = max{v1 , . . . , vn }
0
is also a sequence in S− with u∗ (x) = limn→∞ vn . Using Poisson’s integral for-
mula we find harmonic functions hn ∈ C ∞ (Br (x)) ∩ C 0 (B̄r (x)) with hn ∂Br (x) =
0
vn ∂Br (x)
.
The previous lemma shows that the functions ṽn : Ω → R defined by
(
hn (y) y ∈ B̄r (x)
ṽn (y) = 0
vn (y) y ∈ Ω\B̄r (x)
0
are subharmonic. By definition ṽn = vn on ∂Ω. Thus ṽn ∈ S− . In particular,
ṽn ≤ u∗ . Furthermore, (ṽn )n∈N is still non-decreasing and bounded by sup∂Ω g so
that Harnack’s Convergence Theorem (Theorem 2.15) implies3 that on Br (x) the
sequence ṽn Br (x) converges uniformly to some harmonic function h̃ on Br (x) with
h̃(x) = u∗ (x).
We claim h̃ = u∗ on Br (x). If this was not the case then there is an z ∈ Br (x)
and subharmonic function wn ∈ S− with u∗ (z) = limn→∞ wn (z) > h̃(z). As above
observe that
0 0 0
vn ≤ wn = max{vn , wn } ∈ S−
so that as above we obtain a harmonic function ĥ on Br (x) with h̃ ≤ ĥ and ĥ(z) =
u∗ (z). Note that this also shows ĥ(x) = h̃(x) = u∗ (x). Now the strong maximum
principle on Br (x) applied to h̃ − ĥ ≤ 0 with h̃(x) − ĥ(x) = 0 shows h̃ − ĥ ≡ 0 in
Br (x). However, this is a contradiction because by construction ĥ(z) > h̃(z).
3Alternatively, we may use Corollary 2.17 or 2.19.
 LECTURE NOTES „LINEAR PARTIAL DIFFERENTIAL EQUATIONS“ 14

To conclude just observe that u∗ agrees on Br (x) with a harmonic function

implying. Because Br (x) ⊂⊂ Ω is arbitrary we see that u is harmonic in Ω. An
analogous argument applies to u∗ . 
On arbitrary domain Ω one can show that u∗ 6= u∗ , in particular, u∗ does not
agree with g on the boundary ∂Ω.
Definition 2.22 (Regular point). A point x0 ∈ ∂Ω is called regular if for all
g ∈ C 0 (∂Ω) and all  > 0 there are functions u± ∈ S± such that |g(x0 )−u± (x0 )| ≤ .
Proposition 2.23. The Dirichlet problem is solvable for all g ∈ C 0 (∂Ω) if and
only if each point in ∂Ω is regular.
Proof. The “only if” direction follows by taking the Dirichlet solution with boundary
1
data g as u± . For the opposite direction let x0 ∈ ∂Ω and u− n
∈ S− , n ∈ N, as in
1
the definition of regularity of x0 ∈ ∂Ω. Then max{u−
n
, u∗ } ∈ S− for all n ∈ N. But
1
the definition shows u−
n
≤ u∗ showing that
1 1
0 ≤ g(x0 ) − u∗ (x0 ) ∂Ω
≤ g(x0 ) − u−
n
(x0 ) ≤
n
1
which shows g(x0 ) = u∗ (x0 ). Note that the same applies to u+n and u∗ .
To see that u∗ is continuous each x0 ∈ ∂Ω (and thus u∗ ∈ C 0 (Ω̄)) it suffices to
show that for xn ∈ Ω with xn → x0 it holds u∗ (xn ) → u∗ (x0 ) = g(x0 ). Observe
u− ≤ u∗ ≤ u∗ ≤ u+
where u± ∈ S± is as in definition of regularity of x0 . Because both u− and u+ are
continuous in Ω̄ there is a δ > 0 such that
|u± (x) − u± (x0 )| ≤  for all x ∈ Bδ (x0 ) ∩ Ω̄.
Assume chose now N > 0 such that xn ∈ Bδ (x0 ) for n ≥ N . Then we obtain
u∗ (xn ) − u∗ (x0 ) ≤ u+ (xn ) − g(x0 )
≤ |u+ (xn ) − u+ (x0 )| + |u+ (x0 ) − g(x0 )|
≤ 2
and similarly
u∗ (x0 ) − u∗ (xn ) ≤ g(x0 ) − u− (xn )
≤ |u− (xn ) − u− (x0 )| + |u− (x0 ) − g(x0 )|
≤ 2.
Thus u∗ (xn ) → u∗ (x0 ). 
The proof shows that it suffices to only look at u− ∈ S− . More generally we
want to show that it suffices to look at solution u∗ obtained from Perron’s Method
with g = −d(·, x0 ) ∂Ω .
Definition 2.24 (Barrier function). A subharmonic function b ∈ C 2 (Ω) ∩ C 0 (Ω̄)
is called a lower barrier at x0 ∈ ∂Ω if b(x) < 0 for all x ∈ ∂Ω\{x0 } and b(x0 ) = 0.
A superharmonic function b is an upper barrier at x0 ∈ ∂Ω if −b is a lower barrier.
Proposition 2.25. If there is a lower barrier function b at x0 ∈ ∂Ω then x0 is
regular.
 LECTURE NOTES „LINEAR PARTIAL DIFFERENTIAL EQUATIONS“ 15

Proof. By continuity there is a δ > 0 such that |g(x) − g(x0 )| ≤  whenever x ∈

Bδ (x0 ). Furthermore, there is a k > 0 such that
2 sup |g| + sup kb ≤ 0.
∂Ω ∂Ω\Bδ (x0 )

Such a k exists since b < 0 on ∂Ω\Bδ (x0 ).

Define
u := g(x0 ) −  + kb.
Then u is subharmonic. In order to show that u ∈ S− it suffices to show u ∂Ω
≤ g.
Since kb ≤ 0 we have for x ∈ Bδ (x0 ) ∩ ∂Ω
g(x) − u(x) = g(x) − g(x0 ) +  − kb ≥ 0
and
|u(x0 ) − g(x0 )| = .
If, however, x ∈ ∂Ω\Bδ (x0 ) then
u(x) − g(x) = − − g(x) + g(x0 ) + kb ≤ 2 sup |g| + sup kb ≤ 0.
∂Ω ∂Ω\Bδ (x0 )

An analogous argument shows that

v = g(x0 ) +  − kb ∈ S+
and
|v(x0 ) − g(x0 )| = .


The proof of Proposition 2.23 shows that a (lower) barrier exists at x0 ∈ ∂Ω if

and only if x0 ∈ ∂Ω. Indeed, choose g = −d(·, x0 ) ∂Ω ∈ C 0 (∂Ω) and apply the
first step of Perron’s Method. Then u∗ is harmonic in ∂Ω and u∗ ≤ g on ∂Ω. In
particular, u∗ < 0 on ∂Ω\{x0 }. Thus if x0 is regular then u∗ (x0 ) = g(x0 ) = 0 so
that u∗ is a lower barrier at x0 .
In Rn there is a sufficient condition for a boundary point to admit a (lower)
barrier.
Definition 2.26 (Exterior ball condition). The domain Ω satisfies the exterior ball
condition at x0 ∈ ∂Ω if there is a y ∈ Rn \Ω and an R > 0 such that
B̄R (y) ∩ Ω̄ = {x0 }.
Lemma 2.27. Assume Ω satisfies the exterior ball condition at x0 ∈ ∂Ω then x0
is regular.
Proof. Observe that the function
(
kx − yk2−n − R2−n n>2
u(x) =
log kx−yk
R n=2

then u is in C ∞ (Ω̄) and ∆u = 0 in Ω. Furthermore, u < 0 outside of B̄R (y) and

u(x0 ) = 0 showing that u is a barrier at x0 and thus x0 regular. 
 LECTURE NOTES „LINEAR PARTIAL DIFFERENTIAL EQUATIONS“ 16

Poincaré’s Method. In Perron’s Method we replaced locally vn by a harmonic func-

tion and then used this to conclude that locally the limit is harmonic and agrees
with u∗ . Instead of using local argument and define u∗ via the (pointwise) maxi-
mum of subharmonic functions, we could think of replacing successively an initial
function by a function that is locally harmonic. If we do it everywhere sufficiently
often this should give a non-decreasing sequence of uniformly bounded subharmonic
functions. The limit should then be harmonic. This method is called Poincaré’s
Method.
For this let Bn := Brn (xn ) ⊂⊂ Ω be balls such that Ω = ∪n∈N Bn . Define now a
sequence
nk = 1, 21, 2, 3, 1, 2, 3, 4, . . . , 1, . . . , n, 1, . . .
0
Let g ∈ C (Ω̄) and set u0 = g. For k ≥ 1 let hk be the harmonic function on Bnk
with boundary data uk−1 ∂Ω and define
(
uk−1 x ∈ / Ω̄\Bnk
uk =
hk x ∈ B nk .

Proposition 2.28 (Poincaré’s Method). Assume each point in ∂Ω is regular then

the sequence (uk )k∈N converges to a function u which is harmonic in Ω and satisfies
u ∂Ω = g ∂Ω .

Proof. The maximum principle on Bnk implies that

g = u0 ≤ u1 ≤ · · · ≤ uk ≤ · · · ≤ sup g.
Ω

Define
u(x) = sup uk (x) = lim uk (x)
k∈N k→∞

Assume first g ∈ C 2 (Ω) ∩ C 0 (Ω̄) is subharmonic in Ω. Then by the Replacement

Lemma (Lemma 2.20) uk is mean-value subharmonic. Furthermore, for each x ∈ Ω
there is a sequence (kl )n∈N such that ukl is harmonic on B = Bnkl with x ∈ B. Thus
by Harnack’s Convergence Theorem the sequence (ukl )l∈N converges uniformly on
B to a harmonic function h. However, this implies
u(x) = lim uk (x) = lim ukl (x) = h(x)
k→∞ l→∞

and thus u is harmonic on B. Because x ∈ Ω is arbitrary we see that u is harmonic.

The existence of barriers at x0 ∈ ∂Ω then implies as above that u(x0 ) = g(x0 ).
λ
Now suppose g ∈ C 2 (Ω) ∩ C 0 (Ω̄) satisfies ∆g ≥ −λ. Then g0 = g + 2n k · k2
2
satisfies ∆g0 = ∆g + λ ≥ 0, i.e. g0 is subharmonic. As ∆k · k ≥ 0 we can solve first
via Poincaré’s Method with g0 and obtain a harmonic function u0 with boundary
λ
data g0 ∂Ω . Then solve via Poincaré’s Method with 2n k · k2 to obtain a harmonic
λ 2
function ǔ with boundary data 2n k · k ∂Ω .
By linearity we see that u0 − ǔ is harmonic with boundary data g ∂Ω .
Finally if g ∈ C 0 (Ω̄) then there exists a sequence of functions gn ∈ C ∞ (Ω̄)
with gn ≤ gn+1 , n ∈ N such that g = limn→∞ gn . Via Poincaré’s Method we
obtain a sequence of harmonic functions un with boundary data gn ∂Ω converging
monotonically to a harmonic function u with boundary data g ∂Ω . 
 LECTURE NOTES „LINEAR PARTIAL DIFFERENTIAL EQUATIONS“ 17

1
´
Remark. If D(u) = 2 Ω
|∇u|2 dx denotes the Dirichlet energy of u then Poincaré’s
Method satisfies
D(g) ≥ D(u1 ) ≥ · · · ≥ D(uk ) ≥ D(u)
provided that uk has a well-defined Dirichlet energy.
Generalization of Perron’s Method to solution operators. Assume Ω ⊂ Rn is an
open set
Solution operator:
Given any ball Br (x) ⊂⊂ Ω there is a linear operator PB : C 0 (∂Br (x)) →
C 0 (B̄r (x)) such that given g ∈ C 0 (∂Br (x)) and h := PBr (x) g it holds
• if g ≡ c for some c ∈ R then
PBr (x) (g) ≡ c on B̄r (x).
• if g ≥ 0 then h ≥ 0 and h satisfies the Harnack inequality on Ω+ =
int{h ≥ 0}, i.e. for Ω0 ⊂⊂ Ω+ there is a constant C > 0, not depending
on h, such that
sup h ≤ C inf0 h.
Ω0 Ω
Weak subharmonicity:
A function u ∈ C 0 (Ω) is called weakly subharmonic if for all Br (x) ⊂⊂ Ω
it holds
u ≤ PB g on B̄r (x)
whenever u ∂B (x) ≤ g. If both u and −u are weakly subharmonic then u
r
is called weakly harmonic.
Using those ingredients it is possible to show the following generalized variant of
Perron’s Method.
Theorem (Perron’s Method). Let g ∈ C 0 (∂Ω) and define
S− = {u ∈ C 0 (Ω̄) | u is weakly subharmonic and u ∂Ω
≤ g}.
Then the function u∗ defined by
u∗ (x) = sup u(x)
u∈S−

is weakly harmonic in Ω.

3. Classical Maximum Principles

3.1. Elliptic maximum principles. Let L be an elliptic operator on function in
C 2 (Ω) such that
n
X X n
X
λ(x) ξi ξi ≤ aij (x)ξi ξj ≤ Λ(x) ξi ξi
i=1 i=1
and
n |bk (x)|
b := sup sup .
k=1 x∈Ω λ(x)

Lemma 3.1. Assume c = 0. If u ∈ C 2 (Ω) with Lu > 0 in Ω then u does not

assume a maximum in Ω, i.e. for all x ∈ Ω it holds u(x) < supΩ u. In particular,
u ∈ C 2 (Ω) ∩ C 0 (Ω̄) with Lu > 0 satisfy the strong maximum principle
u(x) < sup u = sup u.
Ω ∂Ω
 LECTURE NOTES „LINEAR PARTIAL DIFFERENTIAL EQUATIONS“ 18

Proof. Assume u(x) = supΩ u for some x ∈ Ω. Then the Hessian D2 u(x) is non-
positive. Furthermore, ∂i u(x) = 0 for all i = 1, . . . , n. Hence
X
Lu(x) = aij (x)∂ij u(x) = tr(A(x) · D2 u(x)) ≤ 0

since A(x) = (aij (x))ni,j=1 is symmetric positive definite. 

Lemma 3.2. The function
v : x = (x1 , . . . , xn ) 7→ e−γx1
satisfies Lv > 0 provided γb < 1 and c = 0.
Proof. Just note that
Lv(x) = γ 2 a11 (x) + γb1 (x)
|b1 (x)|
 
≥ γλ(x) 1 − γ > 0.
λ(x)

Theorem 3.3 (Weak Maximum Principle). Assume c = 0 and b < ∞. If u ∈
C 2 (Ω) ∩ C 0 (Ω̄) satisfies Lu ≥ 0 then u satisfies the weak maximum principle, i.e.
sup u = sup u.
Ω ∂Ω
1
Proof. Choose γ ∈ (0, b) and let v be as in the previous lemma. Then
L(u + v) > 0
so that
sup(u + v) = sup(u + v).
Ω ∂Ω
Letting  → 0 implies the result. 
Corollary 3.4 (Uniqueness of the Dirichlet problem). If c = 0 and b < ∞ then
Lu = Lv and u ∂Ω = v ∂Ω for functions u, v ∈ C 2 (Ω) ∩ C 0 (Ω̄) implies u = v on Ω.
Proof. Observe that
L [±(u − v)] = 0
so that the maximum principle implies
 
sup |u − v| = max sup [+(u − v)] , sup [−(u − v)]
Ω Ω Ω
 
= max sup [+(u − v)] , sup [−(u − v)] = 0
∂Ω ∂Ω

implying the result. 

Definition 3.5 (Interior Ball Condition). A point x0 ∈ ∂Ω satisfies the interior
ball condition if there is a BR (y) ⊂ Ω such that B̄R (y) ∩ ∂Ω = {x0 }.
We make the first observation.
Lemma 3.6. If x0 ∈ ∂Ω satisfies the interior ball condition and the (outer) unit
normal ν of ∂Ω at x0 exists then ν = x0R−y where BR (y) ⊂ Ω and B̄R (y) ∩ ∂Ω.
Furthermore, ν is also the unit normal of ∂BR (y) at x0 .
 LECTURE NOTES „LINEAR PARTIAL DIFFERENTIAL EQUATIONS“ 19

Remark. A point x0 ∈ ∂Ω satisfies the interior ball condition if x0 ∈ ∂(Rn \Ω̄)

saisfies the exterior ball condition. Since the unit normal ν of ∂Ω at x0 exists if
and only if the unit normal ν̄ of ∂(M \Ω̄) at x0 exists the exterior ball condition
also gives
Lemma 3.7. Assume c = 0 and b < ∞. For fixed 0 < ρ < R there is a sufficiently
large α  1 such that the function
2 2
v(x) = e−αkx−yk − e−αR
2
satisfies Lu ≥ 0 on BR (y)\Bρ (y) and ∂ν v(x0 ) = −2αRe−αR < 0 for all x0 ∈
∂BR (y).
Proof. For x ∈ BR (y)\Bρ (y) it holds
" n
#
−αkx−yk2
X X
2 ij ii i



Lv(x) = e 4α a (x)(xi − yi )(xj − yj ) − 2α a (x) + b (xi − yi )
i=1
" n  #
2 X X n
≥ e−αkx−yk 4α2 ij ii k


a (x)(xi − yi )(xj − yj ) − 2α a (x) + sup |b (x)|kxi − yi k
i=1
i=1
" n  #
−αkx−yk2 2 2
X
ii n k
≥e 4α λ(x)kx − yk − 2α a (x) + sup |b (x)|kxi − yi k
i=1
i=1
" n  #
−αkx−yk2 2 2
X
ii n k
≥e 4α λ(x)ρ − 2α a (x) + sup |b (x)|R
i=1
i=1
" n #
X  ii a (x)
−αkx−yk2 2
≥e 2αλ(x) 2αρ − + bR
i=1
λ(x)
Since
aii (x) ≥ λ(x)
we may choose α  1 so that
n  ii 
X a (x)
2αρ2 − + bR ≥0
i=1
λ(x)
implying
Lv ≥ 0.


Lemma 3.8. Assume c = 0 and b < ∞. Let u ∈ C 2 (Ω) ∩ C 1 (Ω̄) satisfy Lu ≥ 0.

In addition, assume the following holds for x0 ∈ ∂Ω
• the point x0 ∈ ∂Ω satisfies the interior ball condition
• u(x0 ) > u(x) for all x ∈ Ω.
Then ∂ν u(x0 ) > 0 provided the unit normal at x0 ∈ ∂Ω exists.
Proof. Let BR (y) be given by the interior ball condition of x0 and pick ρ > 0 and
choose α such that v as in the previous lemma satisfies Lv ≥ 0.
Set A := BR (y)\Bρ (y) and observe that
u − u(x0 ) < 0 on ∂Bρ (y).
 LECTURE NOTES „LINEAR PARTIAL DIFFERENTIAL EQUATIONS“ 20

Furthermore, v = 0 on ∂BR (y). Thus there is an  > 0 such that

u − u(x0 ) + v ≤ 0 on ∂A.
Since L(u − u(x0 ) + v) ≥ 0 the weak maximum principle implies
u − u(x0 ) + v ≤ 0 on A.
Assuming the unit normal ν at x0 ∈ ∂Ω exists this yield
∂ν u ≥ −∂v v > 0.

Theorem 3.9 (Strong Maximum Principle). Assume L is an elliptic operator with
c = 0 and b < ∞. If u ∈ C 2 (Ω) ∩ C 0 (Ω̄) satisfies Lu ≥ 0 and u(x∗ ) = supΩ u for
some x∗ ∈ Ω then u is constant in Ω.
Proof. It suffices to prove that
Ω+ = {x ∈ Ω | u(x) < u(x∗ )}
is empty.
Assume by contradiction Ω+ is non-empty. Note by continuity Ω+ is open.
We first claim there is a y ∈ Ω+ such that
d(y, ∂Ω+ ) < d(y, ∂Ω).
Indeed, by assumption ∂Ω+ \∂Ω 6= ∅. Taking a point x ∈ ∂Ω+ \∂Ω ⊂ Ω and using
compactness of ∂Ω we may take y ∈ B (x0 ) ∩ Ω+ for  < 21 d(x, ∂Ω).
Let x0 ∈ ∂Ω+ be such that R := d(y, ∂Ω+ ) = d(y, x0 ). Then u(x0 ) = u(x∗ ) =
supΩ u and u restricted to Ω̃ = BR (y) and x0 ∈ ∂ Ω̃ satisfies the assumptions of the
previous lemma. Since the unit normal at x0 ∈ ∂ Ω̃ = ∂BR (y) exists it holds
∂ν u(x0 ) > 0.
However, x0 is an interior point of Ω so that ∂i u(x0 ) = 0. This is a contradiction
and implies that Ω+ is empty and thus u must be constant. 
We complete this subsection with the case c ≤ 0.
Theorem 3.10 (Maximum principle for c ≥ 0). Assume Ω is open, bounded and
connected, and L is an elliptic operator with c ≤ 0 and b < ∞. Then for all
u ∈ C 2 (Ω) ∩ C 0 (Ω̄) with Lu ≥ 0 it holds
sup u ≤ sup u+
Ω ∂Ω
+
where u = max{0, u}. Furthermore, if u(x0 ) = supΩ u > 0 for some x0 ∈ Ω then
u is constant in Ω.
Proof. Let
Ω+ = {u > 0}.
+
If Ω = ∅ then u ≤ 0 so that
sup u ≤ 0 = max{0, u+ } = sup u+ .
Ω ∂Ω
+
If Ω 6= ∅ then
L̃u := Lu − cu ≥ 0 on Ω+
 LECTURE NOTES „LINEAR PARTIAL DIFFERENTIAL EQUATIONS“ 21

and L̃ = L − c is an elliptic operator with c = 0. Thus

0 < sup u = sup u.
Ω+ ∂Ω+

Since Ω is bounded, we see that ∂Ω is compact so that there is an x ∈ ∂Ω+ with

+

u(x0 ) = supΩ+ u > 0.

To see that x0 ∈ ∂Ω observe that every point y ∈ ∂(Ω\Ω+ ) satisfies u(y) ≤ 0.
Thus
x ∈ ∂Ω ∩ ∂Ω+
and
sup u = sup u = u(x)
∂Ω ∂Ω∩∂Ω+
which implies supΩ u = sup∂Ω u+ as u+ = u on ∂Ω ∩ ∂Ω+ .
If u(x0 ) = supΩ u > 0 for some x0 ∈ Ω then x0 ∈ Ω+ and u is constant on the
connected component Ω+ x0 containing x0 . Thus
−1
Ω+
x0 = u ((−∞, 0)) = u−1 ({u(x0 )})
is both closed and open in Ω. Hence by connectedness Ω = Ω+
x0 , i.e. u is constant
on Ω. 
Corollary 3.11. Assume c ≤ 0. If v, u ∈ C 2 (Ω) ∩ C 0 (Ω̄) with Lu = Lv and
u ∂Ω = v ∂Ω then u = v on Ω.
Proof. Since L(u − v) = 0 we see that
sup(u − v)+ ≤ sup(u − v)+ = 0
Ω ∂Ω
and
sup(v − u)+ ≤ sup(v − u)+ = 0
Ω ∂Ω
which shows supΩ |u − v| = 0 and thus u = v. 
3.2. Parabolic maximum principles. In the following we assume L is an elliptic
operator so that ∂t − L is a parabolic operator on Q = (0, T ) × Ω.
Lemma 3.12. Assume c = 0. If u ∈ C 2 (Q) with ∂t − Lu < 0 in Q then u does not
assume a maximum in Q, i.e. for all x ∈ Ω it holds u(x) < supΩ u. In particular,
u ∈ C 2 (Q) ∩ C 0 (Q̄) with Lu > 0 satisfy the strong maximum principle
u(t, x) < sup u = sup u for all (t, x) ∈ Q
Q ∂0Q

where ∂ 0 Q = (0, T ) × ∂Ω ∪ {0} × Ω is the parabolic boundary of Q.

Proof. Define Q0 = (0, T 0 ) × Ω for T 0 ∈ (0, T ). Assume u(t0 , x0 ) = supQ u for some
(t0 , x0 ) ∈ Q. Then
∂t u, ∂i u = 0
and
0 > ∂t u(t, x) − Lu(t, x) = −tr(A(t, x) · Dx2 u(t, x)) ≥ 0
which is a contradiction. Here Dx2 u(t, x) denotes the Hessian matrix (∂ij u(t, x))ni,j=1
in the space variables only.
Thus
sup u = sup u.
Q0 ∂Q0
 LECTURE NOTES „LINEAR PARTIAL DIFFERENTIAL EQUATIONS“ 22

By the same argument u(T 0 , x) < supQ u so that

sup u = sup u.
Q0 ∂ 0 Q0

The claim follows by letting T 0 tend to T . 

Theorem 3.13 (Weak Maximum Principle). Assume c = 0 and b < ∞. If u ∈

C 2 (Q) ∩ C 0 (Q̄) satisfies Lu ≥ 0 then u satisfies the weak maximum principle, i.e.

sup u = sup u.
Ω ∂0Q

Proof. Note that for v(t, x) = −t and  > 0 it holds

∂t (u + v) − L(u + v) < 0.

Thus

sup u = lim sup(u + v)

Q →0 Q

= lim sup(u + v) = sup u.

→0 ∂ 0 Q ∂0Q

By the same argument we also obtain the weak maximum principle if c ≤ 0.

Theorem 3.14. Assume c ≤ 0 and b < ∞. If u ∈ C 2 (Q) ∩ C 0 (Q̄) satisfies Lu ≥ 0

then u satisfies the weak maximum principle, i.e.

sup u ≤ sup u+ .
Ω ∂0Q

Corollary 3.15 (Uniqueness). Assume c ≤ 0 and b < ∞. If ∂t u − Lu = ∂t v − Lv

for function u, v ∈ C 2 (Q) ∩ C 0 (Q̄) with u ∂ 0 Q = v ∂ 0 Q then u = v on Q.

For the strong maximum principle we look at the following function: Given
(s, y) ∈ Q and R > 0 define
2
v(t, x) := e−αr(t,x) − e−αR

where in the set

 
R
Qy,s,R = (t, x) ∈ Q | kx − yk > , r(t, x)2 < R2 , t < s
2

1
where r(t, x) = kx − yk2 + η 2 (s − t) 2 with η 2 to be determined later on.
As in the elliptic case it is possible to force v to be a parabolic subsolution.

Lemma 3.16. Assume L is uniformly elliptic, c = 0 and b < ∞ (thus supnk=1 |bk | <
∞. Then it holds
(∂t − L)v < 0
for sufficiently large α.
 LECTURE NOTES „LINEAR PARTIAL DIFFERENTIAL EQUATIONS“ 23

Proof. It holds
" n
#
−αr 2
X X
2 ij
aii (x) + bi (xi − yi ) + 1



(∂t − L)v = e −4α a (x)(xi − yi )(xj − yj ) + 2α
i=1
" n  #
2 X n
≤ 2αe−αr −2αλkx − yk2 + aii (x) + sup |bk (x)|kxi − yi k + 1
i=1
i=1
" n  #
−αr 2 R2 X n
≤ 2αe −2αλ + aii (x) + sup |bk (x)|R + 1
2 i=1
i=1

which is negative for α sufficiently large. 

R2
Note that the set Q̄y,s,R is a cone with tip at (s − η 2 , y) and base {s} × B̄R (y).
Furhtermore, the construction shows
∂v v(x, s) < 0
x−y
for v = kx−ykwhere ∂v denotes the derivative in direction v. We want to this to
show that at boundary {s} × ∂ B̄R (y) of the base and maximum point must have
non-vanishing derivative.
Lemma 3.17. Assume L is uniformly elliptic, c = 0 and b < ∞. Let u ∈ C 2 (Ω)
satisfy ∂t − Lu ≤ 0 and assume the for (t0 , x0 ) ∈ Ω × (0, T ) there is a (y, t0 ) ∈ Q
and R > 0 such that for kx0 − yk = R it holds
Qy,t0 ,R ⊂ Q
(t0 , x0 ) ∈ ∂Qy,t0 ,R
and u(t0 , x0 ) > u(t, x) for all (t, x) ∈ Qx0 ,t0 ,R . Then ∂ν̃ u(t, x0 ) > 0 where ν̃ =
x0 −y
(0, kx−yk ).
Proof. Observe that
R
∂Qy,t0 ,R = {(t, x) ∈ Q | r(t, x) = R, t ≤ t0 } ∪ {(t, x) ∈ Q̄ | kx − yk = , t ≤ t0 }.
2
= S1 ∪ S2
and
v=0 on S1
2 2
v ≤ e−αρ − e−αR on S2 .
Since
u − u(t0 , x0 ) ≤ 0 on S1
u − u(t0 , x0 ) < 0 on S2
and S2 is compact
(∂t − L)u − u(t0 , x0 ) + v ≤ 0 in Qy,t0 ,R
u − u(t0 , x0 ) + v ≤ 0 on ∂ 0 Qy,t0 ,R .
for sufficiently small  > 0.
Then the weak maximum principle implies u − u(t0 , x0 ) + v ≤ 0 in Qy,t0 ,R so
that
∂ν̃ u ≥ −∂ν̃ v > 0.
 LECTURE NOTES „LINEAR PARTIAL DIFFERENTIAL EQUATIONS“ 24


Theorem 3.18 (Strong Parabolic Maximum Principle). Assume L is uniformly
elliptic with c = 0 and b < ∞. If u ∈ C 2 (Ω) ∩ C 0 (Ω̄) satisfies (∂t − L)u ≤ 0 and
u(t∗ , x∗ ) = supQ u for some (t∗ , x∗ ) ∈ Q then u is constant in Q.
Proof. Define an open subset of Q by
Q+ = {(t, x) ∈ Q | u(t, x) < u(t∗ , x∗ )}.
We need to show that Q+ is empty. Assume this is not the case then there is an
(y, s) ∈ Q+ and  > 0 such that (y, s0 ) ∈ Q+ for all s0 ∈ [s − , s]. √
By compactness of Q̄ there is a maximal η > 0 such that for R =  · η
Qy,s,R ⊂ Q+ .
As in the elliptic case we can find (y, s) ∈ Q+ such that
∂Qy,s, R ,R ∩ ∂Q = ∅.
2

In particular, there is a point (t0 , x0 ) ∈ ∂Qy,s, R ,R with u(t0 , x0 ) = u(t∗ , x∗ ). Note

√ 2
that the choice of R =  · η ensures t0 6= s −  and x0 6= y.
Then for R0 = ky − x0 k > 0 it holds
Qy,t0 , R0 ,R0 ⊂ Qy,s, R ,R ⊂ Q+
2 2

and (t0 , x0 ) ∈ ∂Qy,t0 , R2

0
,R0 . In this case the previous lemma yields
∂ν̃ u(t0 , x0 ) > 0.
However, (t0 , x0 ) ∈ Q is a maximum point of u so that
∂ν̃ u(t0 , x0 ) = 0
which is a contradiction. 

4. Sobolev theory in Rn
4.1. Banach spaces.
Definition 4.1 (Banach space). A complete normed space (X, k · k) is called a
Banach space, i.e. X is a vector space and k · k is a norm such that the induced
metric d defined by dk·k (v, w) = kv − wk makes (X, dk·k ) into a complete metric
space.
Remark (Construction of Banach spaces). Using completion we can obtain from a
general normed space (X, k · k) a (unique up to linear isomorphism) Banach space
0
(X̃, k · k ) such that X seen as a subset of X̃ is dense in X and the norms k · k and
0
k · k agree on X.
4.2. Function spaces. Let A ⊂ Rn , e.g. A = Ω or A = Ω̄, and define the following
spaces:
• C 0 (A) = {space of continuous functions on A}.
• C k (A) = {u ∈ C 0 (A) | ∂I u ∈ C 0 (A) for all multi indices I with |I| ≤ k}.
• for α ∈ (0, 1]: C 0,α (A) = {u ∈ C 0 (A) | supx,y∈A |u(x)−u(y)|
kx−ykα < ∞}. If α ∈
0,α α
(0, 1) the space C (A) = C (A) is called the space of Hölder function and
C 0,1 (A) is called the space of Lipschitz functions.
• C k,α (A) = {u ∈ C k (A) | ∂I u ∈ C 0,α (A) for all multi indices I with |I| = k}.
 LECTURE NOTES „LINEAR PARTIAL DIFFERENTIAL EQUATIONS“ 25

• L0 (A) = L0 (A, λn A ) = {space of equivalence classes of measurable functions on A}

(two function u, v ∈ L0 (A) are equivalent if λn A ({u 6= v}) = 0, i.e. u = v
almost everywhere. ´
• for p ∈ (0, ∞): Lp (A) = {u ∈ L0 (A) | A |u|p dλn < ∞}
• L∞ (A) = {u ∈ L0 (A) | ess sup |u| < ∞}.
On each of those spaces there is a natural norm making the subspace of functions
with finite norm into a Banach space4.
• kukC 0 = supx∈A |u(x)|.
• kukC k = supx∈A,|I|≤k |∂I u(x)|.
|u(x)−u(y)|
• kukC 0,α = max{supx∈A |u(x)|, supx,y∈A kx−ykα .
´
1
• kukp = A |u|p dλn p , p ∈ [1, ∞).
• kuk∞ = ess sup |u|.
4.3. Properties of Lp -spaces. The classical Lp -spaces are the space Lp (A) =
Lp (A, λn ) for a Lebesgue measurable set A ⊂ Rn and λn the Lebesgue measure.
Given two Lp -spaces we can define the Lp -product of two Lp -spaces Lp (A) and
Lp (B) (possibly A = B) as the product vector space Lp (A) × Lp (B) equipped with
the following norm
  p1
k(f, g)kp = kf kpLp (A) + kgkpLp (B) .

It is not difficult to see that (Lp (A) ⊗ Lp (B), k(·, ·)k) is a Banach space and very
similar to the regular Lp -spaces. Furthermore, an iterated construction also shows
that
O n
Lp (Ai )
i=1
with corresponding norm is a Banach space.
Definition 4.2 (r-uniform convexity). A Banach space (X, k · k) is said to be
r-uniformly convex if there is a Cp > 0 such that for all v, w ∈ X it holds
p
v+w 1 1
+ Cp kv − wkp ≤ kvkp + kwkp .
2 2 2
Lemma 4.3. For every p ∈ (1, ∞) the Lp -spaces are r-uniformly convex for r =
max{2, p} and (
p−1
p≤2
Cp = 1 4
4 p > 2.
Proof. The statement for p ≥ 2 appeared in the exercises. An accessible proof can
be found in Sharp uniform convexity and smoothness inequalities for trace norms
by Ball–Carlen–Lieb in Inv.Math. (1994). 
Proposition 4.4. Let C be a bounded, closed and convex subset in an r-uniformly
convex Banach space. Then
4If A is not compact then C k (A), k ∈ N, might contain unbounded functions. The set
of bounded continuous functions is usually denoted by Cb0 (A). Furthermore, for C 0,α (A) one
also needs to either add the C 0 -norm (resp. take the maximum of the C 0 -norm and the C 0,α -
seminorm).
 LECTURE NOTES „LINEAR PARTIAL DIFFERENTIAL EQUATIONS“ 26

Corollary 4.5. Let (Ci )i∈I be a net of bounded, closed and convex subsets in an
r-uniformly convex Banach space X such that Ci ⊂ Cj whenever j ≥ i. Then
\
Ci 6= ∅.
i∈I

Proof. We only prove the result for I = N. We define a map r : X → [0, ∞) by

r(x) = sup inf kx − zk.
n∈N z∈Cn

This implies that there is a sequence (zn )n∈N with zn ∈ Cn and

1
|r(x) − rn | ≤
n
and
1
kx − zn k ≤ rn +
n
where rn = inf z∈Cn kx − zn k.
Since zm ∈ Cn whenever m ≥ n we see that r(x) is bounded. We claim that (zn )
is a Cauchy sequence. Indeed, since X is complete, Cn is closed and {zm }m≥n
T ⊂ Cn
we see that zn → z for some z ∈ X and z ∈ Cn for all n ∈ N. Thus z ∈ n∈N Cn
which yields the result.
It remains to show that (zn )n∈N is Cauchy. Observe first that
zn + zm
rn ≤ kx − k
2
and that (rn )n∈N (and hence (rnr )n∈N ) is Cauchy. Fix  > 0 and let n, m ≥ 1 such
r
that rm − rnr ≤ . Then uniform convexity implies
1 1 zn + zm r
Cr kzn − zm kr ≤ kx − zn kr + kx − zm kr − kx − k
2 2 2
1 r 2
≤ (rm − rnr ) + ≤ 3.
2 n
Thus we see that (zn )n∈N is Cauchy. 
Corollary 4.6. For every closed convex subset C of an r-uniformly convex Banach
space (X, k · k) and all u ∈ X there is a unique uC ∈ C such that
ku − uC k = inf ku − vk.
v∈C

Proof. Let rC (x) = inf v∈C ku − vk then

Cn = B̄rC (x)+ n1 (x) ∩ C
satisfies the assumption of the previous statement. In particular, C̃ = ∩n∈N Cn 6= ∅
which implies that there is a uC ∈ C with the required properties. To see that
uC is unique just observe that for any other v ∈ C̃ it holds v+u 2
C
∈ C̃. However,
uniform r-convexity (even weak, the strict convexity of the norm) implies that
Cr kv − uC kr ≤ 0
which means v = uC . 
Remark. The proof of the previous two result are “equivalent”, i.e. it is possible to
prove the latter without the former and then give a proof of the former using the
statement of the latter.
 LECTURE NOTES „LINEAR PARTIAL DIFFERENTIAL EQUATIONS“ 27

4.4. Sobolev spaces and minimizers of quadratic functionals. Using Green’s

formula it is possible to show that a function u ∈ C 2 (Ω) ∩ C 0 (Ω̄) is a harmonic
function in Ω with u ∂Ω = g if and only if it is the minimizer
ˆ
v 7→ |∇v|2 dx
Ω
2 0
among all function v ∈ C (Ω)∩C (Ω̄) satisfying v ∂Ω = u ∂Ω
= g. Since v −u ∂Ω
=
0 for any such v this is equivalent to saying that
ˆ ˆ
|∇u|2 dx ≤ |∇(u + ϕ)|2 dx
Ω Ω
2 0
for all functions ϕ ∈ C (Ω) ∩ C (Ω̄) with ϕ ∂Ω . By density argument one may
equally take ϕ ∈ Cc∞ (Ω), i.e. ϕ is smooth with support compactly contained in Ω.
A similar argument to so-called elliptic operators in divergence form with bk =
c = 0 where we say L is in divergence form if for u ∈ C 2 (Ω)
n
X
Lu := ∂j (aij ∂i u).
i,j=1

Now u ∈ C 2 (Ω) ∩ C 0 (Ω̄) satisfies Lu = 0 in Ω and u ∂Ω = g if and only if it is the

minimizer ˆ X n
E A,Ω : v 7→ aij (x)∂i u(x)∂i u(x)dx
Ω i,j=1

among all function v ∈ C 2 (Ω) ∩ C 0 (Ω̄) satisfying v ∂Ω = u ∂Ω = g. Again this is

equivalent to saying that
ˆ X n ˆ X n
ij
a (x)∂i u(x)∂i u(x)dx ≤ aij (x)∂i (u + ϕ)(x)∂i (u + ϕ)(x)dx
Ω i,j=1 Ω i,j=1

for all functions ϕ ∈ C 2 (Ω) ∩ C 0 (Ω̄) with ϕ ∂Ω

(resp. ϕ ∈ Cc∞ (Ω)).
Remark (Operators in divergence form). If aij ∈ C 1 (Ω) then via product rule a
divergence operator can be brought into non-divergence form. Note that in general
the bk -terms are non-zero and the arguements of this section require bk and c to
be zero, see however the section on the Lax–Milgram Theorem for more general
results.
Observe that for u1 , u2 ∈ C 1 (Ω) and λ ∈ (0, 1) it holds
E A,Ω ((1 − λ)u1 + λu2 ) = (1 − λ)E A,Ω (u1 ) + λE A,Ω (u2 ) − (1 − λ)λE A,Ω (u1 − u2 ).
Thus E A,Ω is a convex function. Furthermore, if λ : Ω → (0, ∞) is the ellipticity
constant of L then
ˆ
E A,Ω (u1 − u2 ) ≥ λ(x)|∇(u1 − u2 )(x)|2 dx > 0

unless u1 = u2 + c for a constant c ∈ R. Thus the convexity inequality above is

strict if the difference of u1 and u2 is non-constant.
The above argument shows that we should look at minimizers of E A,Ω . In
particular, we have to show that the set
{u ∈ C 2 (Ω) ∩ C 0 (Ω̄) | E(u) = inf E(u + ϕ)}
ϕ∈Cc∞ (Ω)
 LECTURE NOTES „LINEAR PARTIAL DIFFERENTIAL EQUATIONS“ 28

is non-empty. Given a function u0 with u0 ∂Ω = g we can always find a sequence

un with un ∂Ω such that limn→∞ E(un ) = inf ϕ∈Cc∞ (Ω) E(u0 + ϕ). However, the
sequence (un )n∈N might not contain any convergent subsequence (in C 2 (Ω)∩C 0 (Ω̄))
as the bounds of E(un ) do not imply any bounds on the first and second derivatives5.
To circumvent we may observe that E A,Ω satisfies the parallelogram inequality.
Hence it seems natural to look at
 21
ˆ X

n
Φ : u 7→  aij (x)∂i u(x)∂i u(x)dx
Ω i,j=1

this mapping from {u ∈ C 1 (Ω) | Φ(u) < ∞} to [0, ∞) satisfies all properties of a
norm but the definiteness. Indeed, as observed above if u1 = u2 + c for a constant
c then Φ(u1 ) = Φ(u2 ). A natural choice is to add the L2 -norm of u and define
 21
ˆ ˆ X

n
Ψ : u 7→  |u|2 dx + aij (x)∂i u(x)∂i u(x)dx .
Ω Ω i,j=1

This mapping gives a norm on the vector space Y = {u ∈ C 1 (Ω) | Ψ(u) < ∞}.
Thus we can take the completion of (Y, Ψ) to obtain a natural Banach spaces.
The only problem with this construction is that the completion might depend on
(aij : Ω → R)ni,j=1 . As we are only interested in existence of minimizer in the
Banach space we may replace the norm by a more suitable one.
Definition 4.7 (uniformly equivalent). A two norms k · k1 and k · k2 on a vector
space X are said to be uniformly equivalent if there is a constant C ≥ 1 such that
for all v ∈ X it holds
C −1 kvk1 ≤ kvk2 ≤ Ckvk1 .
The concept of uniformly equivalent shows that a Cauchy sequence with respect
to one of the norms is also a Cauchy sequence of the other one. Hence we obtain
the following lemma.
Lemma 4.8. Given any two uniformly equivalent norms k · k1 and k · k1 on a
vector space X the completion with respect to either of the norm will give the same
completion X̃ and the naturally associated norm k · k1 and k · k2 are still uniformly
equivalent norms on X̃. In particular, they induce the same complete metric topol-
ogy on X̃.
Note that for each element v ∈ X̃ and every vn → v with vn ∈ i(X) where i is
the natural embedding of the completion process it holds kvki = lim kvki .
Recall that the elliptic operator is uniformly elliptic if there are constant λ, Λ ∈
(0, ∞) such that
λhξ, ξi ≤ hξ, A(x)ξi ≤ Λhξ, ξi.
Thus for a uniformly elliptic operator we may show that
C −1 Ψ(u) ≤ kukW 1,2 ≤ CΨ(u)

5no obvious bound

 LECTURE NOTES „LINEAR PARTIAL DIFFERENTIAL EQUATIONS“ 29

where C 2 = max{λ−1 , Λ} and

ˆ ˆ X
n
! 21
2 2
kukW 1,2 = |u| dx + (∂i u) dx .
Ω Ω i=1

Thus the two norms are equivalent. We call k · kW 1,2 the (W 1,2 -)Sobolev norm
(sometimes also H 1 -norm).
Definition 4.9 (First Sobolev space). The completion of
{u ∈ C 1 (Ω) | kukW 1,2 < ∞}
with respect to the norm k · kW 1,2 is called the first (L2 -)Sobolev space and is
denoted by W 1,2 (Ω).
Since the Sobolev norm is equivalent to the norm Ψ we can also extend the
functional E A,Ω to all function6 in W 1,2 (Ω). Indeed, it holds un → u in W 1,2 (Ω)
then kun − uk2 → 0 and Ψ(un ) → Ψ(u). Thus observing that
E A,Ω (u) = Ψ(u)2 − kuk22
for u ∈ C 1 (Ω) ∩ W 1,2 (Ω) we can choose any un ∈ C 1 (Ω) ∩ W 1,2 (Ω) with un → u ∈
W 1,2 (Ω) and uniquely define
E A,Ω (u) = lim Ψ(un ) − kun k22 .
n→∞

Alternatively we may observe the following.

Lemma 4.10. There is an isometric7 embedding
n
O
i : W 1,2 (Ω) → L2 (Ω)
i=0

satisfying
i(u) = (u, ∂1 u, . . . , ∂n u)
1 1,2
for all u ∈ C (Ω) ∩ W (Ω).
Proof. Just observe that i on C 1 (Ω) ∩ W 1,2 (Ω) is an isometry which extends
uniquely to its closure which is by definition W 1,2 (Ω). 
Thus we see that for u ∈ C 1 (Ω) ∩ W 1,2 (Ω) it holds ∂i u ∈ C 0 (Ω) ∩ L2 (Ω) so that
for any u ∈ W 1,2 (Ω) there is a uniquely defined object ∂i u ∈ L2 (Ω) such that for
all un → u in W 1,2 (Ω) with un ∈ C 1 (Ω) ∩ W 1,2 (Ω) it holds ∂i un → ∂i u in L2 (Ω)
(IMPORTANT: Currently ∂i u for u ∈ W 1,2 (Ω) is just a suggestive notation. Later
we will show that ∂i u satisfies indeed the properties of a partial derivative hence
justifying the use of ∂i ). As uniform ellipticity implies that aij is bounded in Ω the
functional E A,Ω satisfies
ˆ X n
E A,Ω (u) = aij (x)∂i u(x)∂i u(x)dx.
Ω i,j=1

Another observation is the following.

6As W 1,2 (Ω) behaves similar to L2 (Ω) the statement u ∈ W 1,2 (Ω) is a function is meant to
say that the measurable function u represents an equivalence class (of functions) in W 1,2 (Ω).
7norm-preserving
 LECTURE NOTES „LINEAR PARTIAL DIFFERENTIAL EQUATIONS“ 30

Corollary 4.11. The space W 1,2 (Ω) is a Hilbert space. In particular, it is 2-

uniformly convex.

Now we have defined E A,Ω on an appropriate Banach space. The next step is to
find a minimizer given certain boundary data. As we currently do not have “trace
operator” which are maps from W 1,2 (Ω) to L2 (Ω), we try to find a u ∈ W 1,2 (Ω)
that is minimal among all perturbations ϕ with zero boundary data.

Definition 4.12 (Soblev space with zero boundary data). The space W01,2 (Ω) is
defined as the closure of Cc1 (Ω) ∩ W 1,2 (Ω) in W 1,2 (Ω).

It is easy to see that (W01,2 (Ω), k · kW 1,2 ) is a closed subspace of the Banach space
W 1,2 (Ω) and thus a Banach space as well. Furthermore, for each v ∈ W01,2 (Ω) there
is a sequence vn ∈ Cc1 (Ω) ∩ W 1,2 (Ω) with vn → v in W 1,2 (Ω). In particular, by
continuity of E A,Ω we see that

E A,Ω (u) ≤ E A,Ω (u + v) for all v ∈ W01,2 (Ω)

if and only if

E A,Ω (u) ≤ E A,Ω (u + ϕ) for all ϕ ∈ C 1 (Ω) ∩ W01,2 (Ω).

Thus given u0 ∈ W 1,2 (Ω) we may change the minimization problem to find v ∈
W01,2 (Ω) such that
Eu0 : v 7→ E A,Ω (u0 + v).

Lemma 4.13. For each  ≥ 0 the sets

C = {v ∈ W01,2 (Ω) | Eu0 (v) ≤ inf Eu0 + }

is closed and convex with C ⊂ C0 for 0 ≤ . Furthermore, C 6= ∅ for  > 0 and
each element in C0 is a minimizer of Eu0 .

Proof. Convexity follows by observing that

Eu0 ((1 − λ)v1 + λv2 ) = E A,Ω ((1 − λ)(u0 + v1 ) + λ(u0 + v2 ))

≤ (1 − λ)E A,Ω (u0 + v1 ) + λE A,Ω (u0 + v2 )
= (1 − λ)EuA,Ω
0
(v1 ) + λEuA,Ω
0
(v2 ).

Furthermore, if vn → v in W01,2 (Ω) then u0 + vn → u0 + v in W 1,2 (Ω). Thus

continuity of E A,Ω implies that Eu0 is continuous thence

C = Eu−1
0
((−∞, ])

is closed. 

If C is bounded then we would be allowed to use Corollary 4.5 to show that

\
C0 = C n1 6= ∅
n∈N

which yields the existence of a minimizer for Eu0 .

 LECTURE NOTES „LINEAR PARTIAL DIFFERENTIAL EQUATIONS“ 31

However, in general we cannot ensure that k · kW 1,2 is bounded on C as we have

only the following for each v ∈ C
kvk2W 1,2 ≤ 2ku0 + vk2W 1,2 + 2ku0 k2W 1,2
ˆ ˆ
2
≤4 |u0 | + 4 |v|2 dx + 2ΛEu0 (v) + 2ku0 k2W 1,2
Ω Ω
ˆ
2
≤ 6ku0 kW 1,2 + 2Λ(inf Eu0 + ) + |v|2 dx.
Ω

Thus in order to ensure that C is bounded in W01,2 (Ω) it suffices to bound the
L2 (Ω).
Surprisingly the following statement holds:
Theorem 4.14 (Weak version of Gagliardo–Nirenberg Sobolev inequality). For all
v ∈ W01,2 (Ω) there is a constant C = C(n) such that
ˆ ˆ X
n
|v|2 dx ≤ C |∂i v|2 dx.
Ω Ω i=1

Before we prove this theorem we want to show that C is bounded. Indeed it

holds
ˆ
kvk2W 1,2 ≤ 6ku0 k2W 1,2 + 2Λ(inf Eu0 + ) + |v|2 dx.
Ω
ˆ X
n
≤ 6ku0 k2W 1,2 + 2Λ(inf Eu0 + ) + C |∂i v|2 dx.
Ω i=1

≤ 6ku0 k2W 1,2 + 2Λ(inf Eu0 + ) + 2CEu0 (v) + 2Cku0 k2W 1,2
≤ (6 + 2C)ku0 k2W 1,2 + (2Λ + 2C)(inf Eu0 + )
which implies that C is bounded.
Corollary 4.15. If L is an uniformly elliptic operator then for any u0 ∈ W 1,2 (Ω)
there is a unique minimizer of the function Eu0 = E A,Ω (u0 + ·).
Proof. The existence follows from 4.5 using boundedness of C implied by Gagliardo–Nirenberg.
To observe uniqueness, assume Eu0 (v) = Eu0 (v 0 ) for v, v 0 ∈ C0 . It suffices to show
that kv − v 0 kW 1,2 = 0.
By convexity of C0 we see that 21 v + 12 v 0 ∈ C0 so that the parallelogram identity
for E A,Ω yields
ˆ X n ˆ
λ |∂i (v − v 0 )|2 dx ≤ aij ∂i (v − v 0 )∂j (v − v 0 )dx = 0.
Ω i=1 Ω

Since v − v 0 ∈ W01,2 (Ω) Gagliardo–Nirenberg implies

ˆ ˆ X
n
|v − v 0 |2 dx ≤ C |∂i (v − v 0 )|2 dx = 0
Ω Ω i=1

showing that kv − v 0 kL2 = 0. Thus kv − v 0 kW 1,2 = 0 proving the claim. 

Instead of proving the weak version of Gagliardo–Nirenberg let us prove the

following more general statement.
 LECTURE NOTES „LINEAR PARTIAL DIFFERENTIAL EQUATIONS“ 32

Definition 4.16 (First Lp -Sobolev spaces). Let p ∈ [1, ∞) and define a mapping
on C 1 (Ω) by
ˆ ˆ X
n
! p1
kukW 1,p = |u|p dx + |∂i u|p dx .
i=1

Then the first Lp -Sobolev space W 1,p (Ω) is defined as the completion of the vector
space X = {u ∈ C 1 (Ω) | kukW 1,2 (Ω) } equiped with the norm k · kW 1,p . Similarly,
W01,p (Ω) is the closure of Cc1 (Ω) ∩ W 1,p (Ω) in W 1,p (Ω).
Remark. Again it is possible to isometrically embed W 1,p (Ω) into ⊗ni=1 Lp (Ω) to
obtain objects ∂i u ∈ Lp (Ω). This embedding then shows that the norm k · kp is
p0 -uniformly convex if p ∈ (1, ∞) and p0 = max{2, p}.

Theorem 4.17 (Gagliardo–Nirenberg Sobolev inequality). Let v ∈ W01,p (Ω). For

all p ∈ [1, n) there is a constant Cp = C(n, p) such that
ˆ  p1∗ ˆ X n
! p1
p∗ p
|v| dx ≤ Cp |∂i v| dx
Ω Ω i=1

where
np
p∗ = if p ∈ [1, n).
n−p
In order to prove the theorem we need a couple of technical lemma. For nota-
tional purpose we denote by x̂i the vector obtained from x = (x1 , . . . , xn ) ∈ Rn via
remove the i-th coordinate, i.e. x̂i = (x1 , . . . , xi−1 , xi+1 , . . . , xn ).
Lemma 4.18. Assume Fi : Rn−1 → [0, ∞) for i = 1, . . . , n are bounded continuous8
functions with compact support. Then
ˆ ˆ 1
 n−1
1
n n
Πi=1 Fi (x̂i ) n−1 dx ≤ Πi=1 Fi (y)dy .
Rn Rn−1
´
Proof. In the following x ∈ Rn . Then we write Fi (ŷi )dy1 , i > 1 for
ˆ
Fi (y1 , x2 , . . . , xi−1 , xi+1 , . . . , xn )dy1 .

Similarly,
ˆ ˆ
Fi (ŷi )dy1 dy2 = Fi (y1 , y2 , x3 , . . . , xi−1 , xi+1 , . . . , xn )dy1 dy2 .

Using this notation we observe that integrating over the first coordinate and using
the generalized Hölder inequality applied to the last n − 1 terms yields
ˆ Y n ˆ Y
n
1 1 1
Fi (ŷi ) n−1 dy1 = F1 (x̂1 ) n−1 Fi (x̂i ) n−1 dy1
i=1 i=2
n ˆ 1
 n−1
1 Y
≤ F1 (x̂1 ) n−1 Fi (ŷi )dy1 .
i=2

8One may assume measurablilty using Fubini’s theorem.

 LECTURE NOTES „LINEAR PARTIAL DIFFERENTIAL EQUATIONS“ 33

Integrating this inequality over the second inequality yields

ˆ ˆ Y n ˆ 1
 n−1
1
Fi (ŷi ) n−1 dy1 dy2 = F2 (ŷ2 )dy1
i=1
ˆ " ˆ ˆ 1 #
 n−1
1
· F1 (ŷ1 ) n−1 · Πni≥3 Fi (ŷi )dy1 dy2

ˆ 1
 n−1
≤ F2 (ŷ2 )dy1 ·
ˆ 1
 n−1 n ˆ ˆ
Y
1
 n−1
· F1 (ŷ1 )dy2 · Fi (ŷi )dy1 dy2
i≥3

where we applied again the generalized Hölder inequality to the last n − 1 terms.
Then a similar argument with y3 gives
ˆ ˆ ˆ Y n ˆ ˆ 1
 n−1
1
Fi (x̂i ) n−1 dy1 dy2 dy3 ≤ F1 (ŷ1 )dy2 dy3
i=1
ˆ ˆ 1
 n−1
· F2 (ŷ2 )dy1 dy3
ˆ ˆ 1
 n−1
· F3 (ŷ3 )dy1 dy2
n ˆ ˆ ˆ
Y
1
 n−1
· Fi (ŷi )dy1 dy2 dy3
i≥4

where
Ik = {(i1 , . . . , ik ) ∈ {1, . . . , k} | il 6= ik }.
Continuing with y4 to yn yields
ˆ n ˆ 1
 n−1
1 Y
Πni=1 Fi (ŷi ) n−1 dy ≤ Fi (ŷi )dŷi
i=1

which is the claim. 

Lemma 4.19. For v ∈ Cc1 (Ω) ∩ W 1,1 (Ω) it holds

ˆ  n−1
n
ˆ X
n
n
|v| n−1 dx ≤ |∂i v|dx.
Ω Ω i=1

Proof. For x = (x1 , . . . , xn ) ∈ Ω and i = 1, . . . , n it holds

ˆ xi
|u(x)| ≤ |∂i u|(x1 , . . . , yi , . . . , xn )dyi
−∞

so that
 ˆ 1
 n−1
n
|u(x)| n−1 ≤ Πni=1 |∂i u|(ŷi )dyi .
 LECTURE NOTES „LINEAR PARTIAL DIFFERENTIAL EQUATIONS“ 34

´
Define now Fi (x̂i ) = |∂i u(ŷi )|dyi . Then the previous lemma shows
ˆ ˆ Y n n ˆ 1
 n−1
n 1 Y
|u(x)| n−1 dx ≤ Fi (x̂i ) n−1 dx ≤ Fi (y)dy .
i=1 i=1 Rn−1

Observing that
ˆ ˆ X
n
Fi (y)dy ≤ |∂j u|dx
Rn−1 Ω j=1

shows that n
ˆ ˆ X
  n−1
n
n
|u(x)| n−1 dx ≤  |∂j u|dx
Ω j=1

proving the result. 

Proof of the theorem. By density it suffices to show the result for functions Cc1 (Ω)∩
W 1,p (Ω). The previous lemma shows that it is true for p = 1 and with C(n, 1) = 1.
For p ∈ (1, n) and γ > 0 observe that |u|γ ∈ Cc1 (Ω) whenever u ∈ Cc1 (Ω).
Furthermore, it holds |∂i |u|γ | = γ|u|γ−1 |∂i u|.
Thus using the result for p = 1 applied to |u|γ gives
ˆ 1
 n−1 ˆ X n
γn
|u| n−1 ≤ |∂j |u|γ |dx
j=1
ˆ n
X
= γ|u|γ−1 |∂j u|dx
j=1
 p1
ˆ X

ˆ  p−1
p
n
p
(γ−1) p−1
≤γ |u| dx  |∂j u|p dx .
j=1

γn
If we now choose γ such that n−1 = (γ − 1) p−1
p then γ = n−1
n−p p > 0 and p∗ =
np γn
n−p = n−p so that
 p1
ˆ X

ˆ  n−1
n
ˆ  p−1
p
n
∗ ∗
|u|p ≤γ |u|p  |∂j u|p dx .
j=1

To conclude observe that

n−1 p−1 (n − 1)p − n(p − 1) n−p 1
− = = = ∗
n p pn np p
proving
 p1
ˆ X
p−1

ˆ  p1∗ ˆ  n−1
n − p
n
∗ ∗
|u|p = |u|p ≤γ |∂j u|p dx .
j=1


Below we will provide a proof via embedding theorems of the Riesz potential
operator whose proof depends on the Hardy–Littlewood Maximal Functional The-
orem.
 LECTURE NOTES „LINEAR PARTIAL DIFFERENTIAL EQUATIONS“ 35

4.5. Weak derivatives. From theNdefinition of ∂i u as the i-th coordinate of the

n
natural embedding i : W 1,p (Ω) ,→ i=0 Lp (Ω) we see that ∂i (αu + βv) = α∂i u +
β∂i v and k∂i ukp ≤ kukW 1,p . Thus ∂i : u 7→ ∂i u is a bounded linear operator from
W 1,p (Ω) into Lp (Ω).
Lemma 4.20. For all u ∈ W 1,p (Ω) and all ϕ ∈ Cc1 (Ω) it holds
ˆ ˆ
u · ∂i ϕdx = − ∂i u · ϕdx.
Ω
Proof. Observe that the result holds for all u ∈ W 1,p (Ω) ∩ C 1 (Ω). If un → u in
W 1,p (Ω) for un ∈ C 1 (Ω) ∩ W 1,p (Ω) then un → u and ∂i un → ∂i u in L2 (Ω) which
implies
ˆ ˆ
u · ∂i ϕdx = − lim un · ∂i ϕdx
n→∞
Ω
ˆ ˆ
= − lim ∂i un · ϕdx = − ∂i u · ϕdx.
n→∞


Definition 4.21 (Weak derivative). A function gi ∈ L1loc (Ω) is called a weak i-th
coordinate derivative of u ∈ L1loc (Ω) for some i ∈ {1, . . . , n} if for all ϕ ∈ Cc1 (Ω) it
holds ˆ ˆ
gi · ϕdx = u · ∂i ϕdx.
Ω
Remark. By abuse of notation the index i of gi denotes also “which” derivative is
to be considered.
It is not difficult to see that weak derivatives are unique. Furthermore, if u ∈
W 1,p (Ω) then ∂i u is a weak coordinate derivative of u. Furthermore, if u = v on
some subset Ω0 ⊂⊂ Ω and gi and hi are weak derivatives of u and resp. v then
gi = hi on Ω0 . In particular, if u has compact support then any weak derivatives
has compact support as well.
Lemma 4.22 (Chain rule). For all u ∈ W 1,p (Ω) and α ∈ C 1 (R) with α(0) = 0
and |α0 | ≤ M for some M > 0 it holds α(u) ∈ Lp (Ω) and α0 (u) · ∂i u is a weak
derivative of α(u).
Proof. The result holds for function u ∈ C 1 (Ω). If un → u in W 1,p (Ω) for un ∈
C 1 (Ω) ∩ W 1,p (Ω) then un → u and ∂i un → ∂i u in L2 (Ω). In particular, there is a
subsequence (nk )k∈N such that unk → u and ∂i unk → ∂i u almost everywhere.
Now
kα(un ) − α(u)kp ≤ M kun − ukp
implies that (α(unk ))n∈N is a Cauchy in Lp (Ω) converging to α(u). Note that
α0 (u) · ∂i u is clearly in Lp (Ω). Thus we only need to show α0 (u) · ∂i u is a weak
derivative. For this let ϕ ∈ Cc1 (Ω) then by the Dominated Convergence Theorem
it holds
ˆ ˆ
0
α (u) · ∂i u · ϕdx = lim α0 (unk ) · ∂i unk · ϕdx
k→∞
ˆ
= lim ∂i α(unk ) · ϕdx
k→∞
ˆ ˆ
= − lim α(unk ) · ∂i ϕdx = − α(u) · ∂i ϕdx.
k→∞
 LECTURE NOTES „LINEAR PARTIAL DIFFERENTIAL EQUATIONS“ 36

Lemma 4.23 (Product rule). If ζ ∈ C 1 (Ω) with kζkC 1 < ∞ and g ∈ Lp (Ω) is a
weak derivative of u ∈ Lp (Ω) then hi = ∂i ζ · u + ζ · g ∈ Lp (Ω) is a weak derivative
of ζ · u.
Remark. The result holds more generally for u · v for general u, v ∈ L∞ (Ω) ∩ Lp (Ω)
admitting weak derivatives.
Proof. Let ϕ ∈ Cc1 (Ω) then ζ · ϕ ∈ Cc1 (Ω) so that
ˆ ˆ ˆ
− ζ · u · ∂i ϕdx = − u · ∂i (ζ · ϕ)dx + ∂i ζ · u · ϕdx
ˆ ˆ
= ζ · g · ϕdx + ∂i ζ · u · ϕdx.

Using the weak topology of W 1,p (Ω) and the fact that kα(u)kW 1,p ≤ M kα(u)k
for all u ∈ C 1 (Ω) we may even show that α(u) ∈ W 1,p (Ω). Instead of this we show
the following theorem.
Theorem 4.24. A measurable function u : Ω → R is in W 1,p (Ω) if and only if
u ∈ Lp (Ω) and for each i ∈ {1, · · · , n} there is a weak derivative gi ∈ Lp (Ω).
Note that u ∈ W 1,p (Ω) if and only if it can be k · kW 1,p -approximated by C 1 -
functions. The theorem is saying that Lp -functions with weak derivative in Lp
always admit C 1 -approximation. This was an open question until Meyers and
Serrin proved in a paper titled “H = W ” in 1964.
In order to prove the theorem we need the concepts of cut-off functions, mollifiers
and smooth partitions of unit. For this define the following function
1
(
1 − 1−kxk2
c e kxk < 1
ϕ(x) = n

0 kxk ≥ 1
where ˆ
1
− 1−kxk 2
cn = e dx,
B1 (0)
´
i.e. Rn ϕdx = 1. Also define ϕ (x) = −n ϕ(x/).
Given a function u ∈ L1loc (Rn ) define
ˆ
u (x) = ϕ (x − y)u(y)dy.
Rn

Proposition 4.25. For all  > 0 the functions u are C ∞ (Rn ) and
supp u ⊂ supp u + supp ϕ = supp u + B (0).
Furthermore,
ku kp ≤ kukp .
p
Finally if gi ∈ L (Ω) is a weak derivative then (gi ) = ∂i u .
Proof. The first two statements follows by exchanging integration and differentia-
tion.
 LECTURE NOTES „LINEAR PARTIAL DIFFERENTIAL EQUATIONS“ 37

The next claim follows from the fact that

ˆ p
|u (x)|p = ϕ (x − y)u(y)dx
ˆ
≤ ϕ (x − y)|u(y)|p dy.

Thus
ˆ ˆ ˆ
|u |p dx ≤ ϕ (y)|u(x − y)|p dydx
ˆ ˆ
≤ ϕ (y) |u(x − y)|p dxdy
ˆ ˆ ˆ
= ϕ (y) |u|p dxdy = |u|p dx.

Furthermore, the last claim let ψ ∈ Cc1 (Rn ). Then again using Fubini
ˆ ˆ ˆ
(gi ) · ψdx = ϕ (y) gi (x − y)ψ(x)dxdy
ˆ ˆ
= ϕ (y) u(x − y)∂i ψ(x)dxdy
ˆ
= u · ∂i ψdx.


Lemma 4.26. If u ∈ Lp (Rn ) has compact support then
ku − ukp → 0 as  → 0.
Proof. Let un → u in Lp (R) such that un ∈ Cc0 (Ω) where Ω is a bounded open
subset with supp u ⊂ Ω. It is easy to see that (un ) → un uniformly in Ω. For a
given δ > 0 choose n 3 N such that
ku − un kp ≤ δ.
Now it holds
lim ku − ukp ≤ lim [ku − (un ) kp + k(un ) − un kp + kun − ukp ]
→0 →0
≤ lim [k(un ) − un kp + 2kun − ukp ] ≤ 2δ.
→0

Since δ > 0 is arbitrary we see that ku − ukp → 0 as  → 0. 

Corollary 4.27. For all u ∈ Lp (Rn ) with compact support admitting weak deriva-
tives gi ∈ Lp (Rn ) for i = 1, . . . , n there is a sequence un ∈ Cc1 (Rn ) such that un → u
and ∂i un → gi in Lp (Rn ). In particular, u ∈ W 1,p (Rn ).
In order to prove the theorem we need to split a given Sobolev function into
summands with compact support in Ω.
Definition 4.28 (Partition of Unitity). Let Ω be open and {Ωn }n∈N be an locally
finite covering of bounded sets. A family of function ηn ∈ Cc∞ (Ωi ) is called a
smooth partition of unity subordinate to {Ωn }n∈N if
X
ηn ≡ 1.
n∈N
 LECTURE NOTES „LINEAR PARTIAL DIFFERENTIAL EQUATIONS“ 38

Remark. Note that local finiteness of the covering implies that for each xP∈ Ω there
are finitely many nx1 , . . . , nxm(x) ∈ N such that x ∈ Ωnxk . In particular, ηn (x) =
Pm(x)
k=1 ηnk (x).
x

Proposition 4.29. For each locally finite bounded covering {Ωn }n∈N of an open
set Ω with Ωn ⊂⊂ Ω there is a smooth partition of unity subordinate to the covering.
Proof. Given a locally finite covering we can finite open set Vn ⊂⊂ Ωn such that
{Vn }n∈N is still a covering of Ω. Just observe that given any covering {Un }n∈N of
Ω with Un ⊂⊂ Ω and n0 ∈ N there is an  > 0 such that
{Un−
0
} ∪ {Un }n6=n0
is still a locally finite covering where
Un−
0
= {x ∈ Un0 | d(x, ∂Un0 ) > }.
Thus the sets Vn = Ω−
n
n
can be constructed inductively.
Define now
un := (χVn ) 2n
and
un
ηn = P .
m∈N um
By construction, un ∈ Cc∞ (Ωn ) and the summand in the definition of ηn consists
of only finitely many um . Furthermore,
P
X un (x)
ηn (x) = P n∈N = 1.
n∈N m∈N m (x)
u

Proof of the Theorem ??. It suffices to show that any function u ∈ Lp (Ω) with weak
derivatives gi can be approximated by C 1 -functions un such that ∂i un converges to
gi .
Define
1 1
Ωn = {x ∈ Ω | d(x, ∂Ω) ∈ ( , )}.
n−1 n+1
Then {Ωn }n∈N is a locally finite covering of Ω and each Ωn is bounded. Thus there
is a partition of unit ηn . Observe now that for sufficiently small n the function
(ηn · u)n is in Cc∞ (∪n+1
k=1 Ωk ) and
δ
k(ηn · u)n − ηn · ukp , k(∂i (ηn · u))n − hni kp ≤ .
2n
wherePhni = ∂i ηn · u + ηn · gi ∈ Lp (Ω) is the weak derivative of ηn · u. Observe that
gi = n∈N hni .P
Thus uδ = n∈N (ηn · u)n is well-defined and in C ∞ (Ω) and it holds
X X
kuδ − ukp ≤ k(ηn · u)n − ηn · ukp ≤ δ 2−n = δ
n∈N n∈N
X X
k∂i uδ − gi kp ≤ k∂i (ηn · u)n − hi kp ≤ δ 2−n = δ
n∈N n∈N

Letting δ → 0 proves the result. 

 LECTURE NOTES „LINEAR PARTIAL DIFFERENTIAL EQUATIONS“ 39

4.6. Higher order Sobolev spaces.

Definition 4.30 (Higher order Sobolev spaces). Let u ∈ C k (Ω) then we can define

X ˆ
 

kukW k,p =  |∂I u|p dx

|I|≤k Ω

where I = (i1 , · · · , ik ) is a multi-index and ∂I u = ∂i1 ···in u and if |I| = 0 then

∂I u = u.
The k-th order Sobolev space W k,p (Ω) is the completion of {u ∈ C k (Ω) | kukW k,p <
∞} with respect to k · kW k,p .
Similarly, the space W0k,p (Ω) is the closure of Cck (Ω) ∩ W k,p (Ω) in W k,p (Ω).
Remark. Again there is a natural embedding of C k (Ω) ∩ W k,p (Ω) into
Nk
O
Lp (Ω)
i=0
Pk i
where Nk = i=1 n which is defined by
u 7→ (∂I u)|I|≤k .
This extends naturally to u ∈ W k,p (Ω) with well-defined objects ∂I u ∈ Lp (Ω),
|I| ≤ k.
Lemma 4.31. Let u ∈ W k,p (Ω). Then for all |I| ≤ k and ϕ ∈ Cck (Ω) it holds
ˆ ˆ
|I|
∂I u · ϕdx = (−1) u · ∂I ϕdx.

Definition 4.32 (Weak k-th derivatives). Given a multi-index I with k = |I| > 0
we say gI ∈ L1loc (Ω) is a weak k-th order derivative of u ∈ L1loc (Ω) if for all ϕ ∈
Cck (Ω) it holds
ˆ ˆ
gI · ϕdx = (−1)k u · ∂I ϕdx.
Ω Ω

Remark. By abuse of notation, the index I of gI denotes also “which” derivative is

to be considered.
The following is an easy observation which follows from commutativity of the
partial derivatives of C k -functions.
Lemma 4.33. Let u ∈ W k,p (Ω) and I1 and I2 be two multi-indices such that the
composed multi-index I = I1 t I2 = (i11 , . . . , i1l , i21 , . . . i2m ) satisfies |I| ≤ k. Then gI
is a weak m-th order derivative of gI1 and a weak l-th order derivative of gI2 .
The following theorem can be proved almost exactly as Theorem 4.24.
Theorem 4.34. Let u ∈ Lp (Ω) then the following are equivalent:
• u ∈ W k,p (Ω), i.e. there is a sequence un ∈ C k (Ω) ∩ W k,p (Ω) with un → u
in W k,p (Ω).
• For each multi-index I with |I| ≤ k there is a weak derivative gI ∈ Lp (Ω)
of u.
• u ∈ W 1,p (Ω) and gi ∈ W k−1,p (Ω) for each i ∈ {1, . . . , n}.
 LECTURE NOTES „LINEAR PARTIAL DIFFERENTIAL EQUATIONS“ 40

k−|I|
Note that if u ∈ Cck (Ω) then ∂I u ∈ Cc (Ω) whenever |I| ≤ k. Thus if u ∈
k−|I|,p
W0k,p (Ω) then for each |I| < k it holds ∂I u ∈ W0 (Ω). Furthermore, W0l,p (Ω) ⊂
W01,p (Ω) Thus we can apply Theorem 4.17 to each ∂I u ∈ W k−|I| (Ω) and obtain the
following
Theorem 4.35 (Higher order Gagliardo–Nirenberg). If p ∈ [1, n) and u ∈ W0k,p (Ω)
∗
then u ∈ W0k−l,p (Ω) and
kukW k−1,p∗ ≤ Cp kukW k,p
where
np
p∗ = .
n−p
n
Corollary 4.36. Let l ∈ {0, . . . , k − 1}. If p ∈ [1, k−l ) and u ∈ W0k,p (Ω) then for
l,q
it holds u ∈ W0 (Ω) with
kukW l,q ≤ Cp,(k−l) kukW k,p
where
np
q= .
n − (k − l)p
n−2 n
Remark. For p = 2 and 2 < (k − l) < 2 it holds
2n
> n.
n − (k − l)2
4.7. Poincaré inequality, and embeddings of Sobolev and Morrey.
Proposition 4.37. Let Ω be a convex, bounded and open subset of Rn . Then for
each p ∈ [1, ∞) there for all u ∈ W 1,p (Ω) it holds
n
X 1
|u − uΩ |p dx ≤ 2n (diam Ω)p |∂i u|p dx p
Ω Ω i=1

where
uΩ = udx.
Ω
Pn 1
Proof. Set gu = ( i=1 |∂i u|p ) p and ` = diam Ω. Since C 1 -functions are dense in
W 1,p (Ω) it suffices to assume u ∈ C 1 (Ω) ∩ W 1,p (Ω). By Jensen’s inequality we have
ˆ ˆ p
p
|u − uΩ | dx = u(x) − udy dx
Ω
ˆΩ Ω
p
≤ |u(x) − u(y)| dydx.
Ω Ω

Let given x, y ∈ Ω define γxy (t) = (1 − t)x + ty and note by convexity of Ω we have
γxy (t) ∈ Ω. Thus
ˆ 1
|u(x) − u(y)| = kx − yk ∂t u((1 − t)x + ty)dt
0
ˆ 1
≤` gu (γxy (t))dt.
0
 LECTURE NOTES „LINEAR PARTIAL DIFFERENTIAL EQUATIONS“ 41

Again using Jensen’s inequality we get

ˆ ˆ
p
|u − uΩ | dx ≤ |u(x) − u(y)|p dx
Ω Ω
ˆ ˆ 1 p
p
≤` gu (γxy (t))dt dydx
Ω Ω 0
ˆ ˆ 1
≤ `p gu (γxy (t))p dtp dydx
Ω Ω 0
ˆ ˆ ˆ 1
2`p
= gu (γxy (t))p dtdydx
|Ω| Ω Ω 1
2

where we use the symmetry γxy (t) = γyx (1 − t).

Using change of coordinates we also have
ˆ ˆ ˆ
1 1
gu (γxy (t))p dy = n gu (y)p dy ≤ n gu (y)p dy
Ω t Ωt,x t Ω

where
Ωt,x = {γxy (t) | y ∈ Ω} ⊂ Ω.
We estimate ˆ 1
t−n dt ≤ 2n−1
1
2

so that
ˆ ˆ ˆ 1 ˆ
p 2`p 1
|u − uΩ | dx ≤ gu (y)p dydtdx
|Ω| Ω 1 tn Ω
n p
2
ˆ ˆ ˆ
2 `
≤ gu (y)p dydx = 2n `p gu (y)p dy.
(n − 1)|Ω| Ω Ω Ω

Dividing each side by |Ω| gives the result. 

Observe that the Hölder inequality implies

n
! n
! p1
X X
p
|∂i u|dx ≤ |∂i u| dx
Ω i=1 Ω i=1

which yields the following corollary.

Corollary 4.38. For all p ∈ [1, ∞) it holds
n
! p1
X
n p
|u − uΩ |dx ≤ 2 diam Ω |∂i u| dx .
Ω Ω i=1

Lemma 4.39. If Ω is convex then there are constants cΩ and CΩ such that for all
x ∈ Ω and r > 0 it holds
cΩ · rn ≤ |Br (x) ∩ Ω| ≤ CΩ · rn .
Proposition 4.40 (Morrey’s Embedding). Assume Ω is bounded and convex. If
n
u ∈ C 1 (Ω) ∩ W 1,p (Ω) for p ∈ (n, ∞) then u ∈ C 0,1− p (Ω) such that
n
|u(x) − u(y)| ≤ Cn kDukp d(x, y)1− p .
 LECTURE NOTES „LINEAR PARTIAL DIFFERENTIAL EQUATIONS“ 42

Proof. Let x, y ∈ Ω and for i ∈ N>0 define balls

Bi = B2−(i−1) d(x,y) (x) ∩ Ω
B−i = B2−(i−1) d(x,y) (y) ∩ Ω
and
B0 = B2d(x,y) (x) ∩ Ω.
Then each ball Bi , i ∈ Z is convex and for i > 0 it holds
diam Bi = 2−i d(x, y)
and
|Bi | ≥ cΩ · 2−(i−1)n d(x, y)n ≥ cΩ · 2−in d(x, y)n .
Pn 1
Let gu = ( i=1 |∂i u|p ) p . Then

|uBi − uBi+1 | = | udx − udy|

Bi Bi+1

≤ |u − ( udy)|dx
Bi+1 Bi

≤ |u − udy|dx
Bi+1 Bi

≤ |u − uBi |dx
Bi
ˆ  p1
n 1
≤ 2 (diam Bi ) 1 gup dx
|Bi | p Bi
−i(1− n
p) 1− n
≤ cp,n · 2 d(x, y) p kgu kp
1
−p
for cn,p = 2n · cΩ . Similarly,
n n
|uB−i − uB−(i+1) | ≤ cp,n · 2−i(1− p ) d(x, y)1− p kgu kp
and n
|uB±1 − uB0 | ≤ 4 · cn,p · d(x, y)1− p kgu kp .
Since u is continuous we have
uBi → u(x)
and
uB−i → u(y).
Thus we get the telescope sum
X
|u(x) − u(y)| ≤ |uBi − uBi+1 |
i∈Z
" n
#
−i(1− n n
X
≤ 8 · cp,n · 2 p) · kgu kp · d(x, y)1− p
i=0
n
= Cp,n · kgu kp · d(x, y)1− p
where " n #
−i(1− n
X
Cp,n = 8 · cp,n · 2 p) < ∞.
i=0
Since kgu kp = kDukp we conclude. 
 LECTURE NOTES „LINEAR PARTIAL DIFFERENTIAL EQUATIONS“ 43

Note that for any Ω there is a ball B of radius 12 diam Ω containing Ω. Since
trivially any function u ∈ W0k,p (Ω) is also the W 1,p -limit of function Cck (Ω) ⊂ C k (B)
we have u ∈ W k,p (B).
k−1,1− n
Corollary 4.41. If u ∈ W0k,p (Ω) with p > n then u ∈ C0 p
(Ω) with
kuk k−1,1− n
p
≤ Cp,n kukW k,p .
C

In combination with the Gagliardo–Nirgenberg we also obtain the following.

Corollary 4.42. If u ∈ W0k,p (Ω) and kp > n then u ∈ C k−l,γ where
 
n
l= +1
p
and
(
l − np n
p ∈
/N
γ= n
any number in (0, 1) p ∈ N.
k,p
Remark. Observe that for n−2
2 < (k − (l + 1)) any Sobolev function u ∈ W0 (Ω)
l l
is in C (Ω) with bounded C -norm.
For completeness we also give a Sobolev inequality for general functions u ∈
W 1,p (Ω). We only sketch its proof as it suffices to have bounds for functions
u ∈ W01,p (Ω).
Proposition 4.43 (Sobolev Embedding). Assume Ω is bounded and convex. Then
there is a C = C(n, p) > 0 such that for all u ∈ W 1,p (Ω) with p ∈ [1, n) it holds
ku − uΩ kp∗ ≤ CkDukp .
Note that this is a variant of the Sobolev inequality of Gagliardo–Nirenberg.
Indeed, if u ∈ W01,p (Ω) for some bounded Ω then there is a (convex) ball Br (x0 )
containing Ω so that u ∈ W 1,p (Br (x0 )).
The proof of this inequality relies on the following bound.
Lemma 4.44 (Bound for the Riesz Potential). For all p ∈ [1, n) there is an C =
C(n, p) such that for all f ∈ Lp (Rn ) it holds
kV1 f kp∗ ≤ Ckf kp
where ˆ
f (y)
Vα f (x) = dy
kx − ykn−α
for α > 0.
The lemma provides an easy proof of Gagliardo–Nirenberg’s Sobolev inequality:
For this note that for all u ∈ Cc1 (Rn ) it holds
ˆ Pn
1 i=1 (xi − yi )∂i u(y)
u(x) = dy.
nωn Rn kx − ykn
In particular,
ˆ Pn 1
1 ( i=1 |∂i u|p (y)) p
|u(x)| ≤ dy.
nωn Rn kx − ykn−1
 LECTURE NOTES „LINEAR PARTIAL DIFFERENTIAL EQUATIONS“ 44

Pn 1
Thus for f = ( i=1 |∂i u|p (y)) p it holds
ˆ n
X
! p1
kukp∗ ≤ kV1 f kp∗ ≤ Ckf kp = C |∂i u|p (y)dy .
Rn i=1

We will show that a similar argument also proves the general Sobolev inequality.
Proof of the Sobolev embedding. Note that it suffices to show the inequality for u ∈
C 1 (Ω) ∩ W 1,p (Ω). Let ` = diam Ω we claim
ˆ
`n f
|u(x) − uΩ | ≤ dy
n|Ω| Ω kx − ykn−1
Pn 1
where f = ( i=1 |∂i u|p (y)) p . Then the claim follows as above.
To see the claim note that
ˆ kx−yk
u(x) − u(y) = − ∂r u(x + rωy,x )dr
0
y−x
for ωy,x = ky−xk . Thus

ˆ ˆ kx−yk
1
|u(x) − uΩ | = ∂r u(x + rωy,x )drdy
|Ω| Ω 0
ˆ ˆ ∞
1
≤ f (x + rωy,x )drdy
|Ω| kx−yk<d 0
ˆ ∞ ˆ ˆ d
1
≤ f (x + rωy,x )ρn−1 dρdωdr
|Ω| 0 kωk=1 0
ˆ ∞ ˆ
`n
= f (x + rωy,x )dωdr
n|Ω| 0 kωk=1
ˆ ∞ ˆ
`n f (y)
= dy
n|Ω| 0 kωk=1 kx − ykn−1
where we extended f to a function on Rn by setting f = 0 on Rn \Ω. 
4.8. Lax–Milgram Theorem.
Proposition 4.45 (Riesz respresentation). Let (H, h·, ·i. Then for every bounded
linear map α : H → R there is a unique u ∈ H such that for all v ∈ H it holds
α(v) = hu, vi.
Proof. See Exercise Sheet 8. 
Definition 4.46. A bi-linear map B : X × X → Ris called bounded if for some C
and all u, v ∈ X it holds
|B(u, v)| ≤ C · kuk · kvk
it is called coercive if for some c > 0 and all u ∈ X it holds
B(u, u) ≥ Ckuk2 .
Definition 4.47. Let X be a vector space. A symmetric bi-linear map h·, ·i :
X × X → R such that hu, ui > 0 is called a scalar product on X. For u ∈ X define
1
kuk = hu, ui 2 . If (X, k · k) is a Banach space then (X, h·, ·i) will be called a Hilbert
space and k · k its induced norm.
 LECTURE NOTES „LINEAR PARTIAL DIFFERENTIAL EQUATIONS“ 45

Lemma 4.48. If (X, k · k) admits a bounded, coercive bi-linear map B : X × X →

Rthen
1
hu, viB = (B(u, v) + B(v, u))
2
makes X into a Hilbert space such that k · k and the norm k · kB induced by this
scalar product are uniformly equivalent.
Theorem 4.49 (Lax-Milgram). For all bounded bi-linear, coercive map B : X ×
X → R and all bounded linear maps α : X → R there is a unique u ∈ X such that
B(u, v) = α(v)
for all v ∈ X.
Proof. Since B is bounded and coercive there is a scalar product h·, ·i making X
into a Hilbert space such that the norms k · kB and k · k are uniformly equivalent.
In particular, B is still bounded and coercive with respect to k · kB .
By the Riesz Representation Theorem, there is a linear map T : X → X such
that
B(u, v) = hT u, viB .
Note that by the assumptions on B it holds
c(kukB )2 ≤ B(u, u)
= hT u, uiB
≤ kukB · kT ukB
and
kT uk2B ≤ B(u, T u) ≤ CkukB · kT ukB .
which implies that T is bounded and injective and T (X) is closed. We claim T is
also onto. Indeed, if this was false then there is a z 0 ∈ T (X)⊥ \{0} such that
hz 0 , T (u)iB = 0
for all u ∈ X. However, by coercivity, we would have
hz 0 , T (z 0 )iB ≥ c(kz 0 kB )2 > 0
which is contradiction.
Let g be the representative of α. Then there is a unique u = T −1 (g) such that
B(u, v) = hg, viB = α(v).

As an application:
Corollary 4.50. Given a uniformly elliptic operator L on a bounded domain Ω
in divergence form. If bk and c are bounded by a constant only depending on Ω
and the lower ellipticity constant λ of L then for all f ∈ L2 (Ω) there is a unique
u ∈ W01,2 (Ω) such that for all v ∈ W01,2 (Ω) if holds
ˆ Xn n
X ˆ
ij k
− a · ∂i u · ∂j v + b · ∂k u · v + c · u · vdx = f vdx.
i,j=1 k=1

In particular, if u ∈ W 2,2 (Ω) then

Xn n

X
∂j aij · ∂i u + bk · ∂k u + cu = f.
i,j=1 k=1
 LECTURE NOTES „LINEAR PARTIAL DIFFERENTIAL EQUATIONS“ 46

4.9. (to be written) Trace and Extension operators.

4.10. (partially covered in exercise) A weak topology via convex sets.
Definition 4.51 (co-convex topology). Given a Banach space first a base for the
topology by
σco = {U ∈ 2X | ∃k ∈ N, C1 , . . . Ck closed and convex such that U = X\ ∪ki=1 Ci }.
Then the following defines a topology
[
τco = {U ∈ 2X | ∃Ui ∈ σco , i ∈ I : U = Ui }.
i∈I

Remark. This can be shorten to saying that τco is the topology generated by the
subbase of sets which are complements of closed and convex subsets. Hence the
name co-convex topology.
From the definition we see that each set in σco (hence also each set in τco ) is
open with respect to the norm topology. Hence the co-convex topology is weaker
than the norm topology which is usually called the strong convergence. Whereas we
write vn → v for the norm convergence we use vn * v for the co-convex topology9.
In particular, if vn → v in X then vn * v.
Definition 4.52 (Convex hull). Let A ⊂ X be any subset. Then the closed convex
hull of A is defined as
X X
convA := cl{v ∈ X | ∃λm m
n ∈ [0, 1], vn ∈ A : λm
n = 1, v = lim λm m
n vn }.
m→∞
n∈N

Note that we may equally define by requiring that the convex combinations on the
right hand side are finite, i.e. for each m there is an Nm ∈ N such that λm
n = 0
whenever n ≥ Nn .
As a direct corollary of the definition of τco we obtain the following theorem.
Theorem 4.53 (Mazur’s Lemma). If (vn )n∈N converges to v with respect to τco
then \
v∈ conv{vn }n≥m .
m∈N
In particular, there is sequences (λm )n∈N in [0, 1] and Nm > m in N with Nm → ∞
nP
such that λnm = 0 for n ∈ / [m, Nm ], n∈N λm n = 1 and
∞
X
v = lim λm m
n vn .
m→∞
n=1

Proof. The subsequence (vn )n≥m also converges to v in τco . Furthermore, the sets
Cm = conv{vn }n≥m are closed and convex. Since by definition Cm is closed in τco
its τco -limit v must be in Cm as well. Thus by definition of conv{vn }n≥m there are
λm m
n ∈ [0, 1] and a Nm ∈ N such that λn = 0 for n ∈ / [m, Nm ] and
X 1
kv − λm
n vn k ≤ .
m
n≥m


9below it will be shown that the co-convex topology is the weak topology.
 LECTURE NOTES „LINEAR PARTIAL DIFFERENTIAL EQUATIONS“ 47

Lemma 4.54. If f : X → R is a convex lower-semicontinuous function then f is

lower semi-continuous with respect to the co-convex topology.
Proof. The proof is shown for sequences but by change of notation also works for
general nets.
Let choose any sequence (vn )n∈N with vn * v. Then there is a subsequence
(vnk )k∈N such that
C = lim f (vnk ) = lim inf f (vn ).
k→∞ n→∞
Because f is convex and lower semi-continuous the sets {f ≤ r}, r ∈ R, are closed
and convex. Thus it suffices to show that v ∈ {f ≤ C + } for all  > 0.
Now for every  > 0 there is an K ∈ N such that for all k ≥ K it holds
f (vnk ) ≤ C +  implying that vnk ∈ {f ≤ C + }. However, the set {f ≤ C + } is
closed and convex which imply that the limit v of (vnk )k∈N is in v ∈ {f ≤ C +}. 
Let X ∗ be the set of bounded linear functions, i.e. α : X → R is linear and there
is a C such that
|α(v)| ≤ Ckvk.
On can show that
kαk∗ = sup |α(v)|
kvk=1
is a norm on X ∗ making it into a Banach space.
Lemma 4.55. A linear function α : X → R is continuous if and only if it is
bounded.
Proof. Assume first α ∈ X ∗ . Let vn → v then kvn − vk → 0. Thus by linearity and
boundedness of α it holds
lim |α(v) − α(vn )| = lim |α(v − vn )| ≤ C lim kvn − vk = 0.
n→∞ n→∞ n→∞
Assume α is unbounded. We will show that α is not continuous at the origin
0 ∈ X. Indeed, unboundedness implies there is a sequence vn with kvn k = 1 such
that 0 < an := α(vn ) → ∞. Observe wn := a1n vn → 0 so that
lim α(wn ) = 1 6= α(0) = 0.
n→∞

Definition 4.56 (Weak topology). The weak topology τw on X is the weak-
est/smallest topology on X such that each bounded linear function is continuous.
Corollary 4.57. Any bounded linear function is continuous with respect to the co-
convex topology. In particular, the co-convex topology is stronger10 than the weak
topology, i.e. if (vn )n∈N converges with respect to τco then it also converges with
respect to τw .
Using the Hahn–Banach Separation Theorem we can also show the following.
As we don’t need the result in the general setting, we only sketch its proof. The
Hahn–Banach Theorem also implies that the τw (and thus τco ) are Hausdorff topolo-
gies, i.e. a sequence/net can converge to at most one point. We only state the
version needed here.
10More precisely, “not weaker” as the two topologies are equivalent/the same.
 LECTURE NOTES „LINEAR PARTIAL DIFFERENTIAL EQUATIONS“ 48

Theorem (Hahn–Banach Separation Theorem). For any Banach space (X, k · k)

and every v ∈ X\{0} and every convex closed set C with v ∈
/ C there is a linear
function α with α(v) ∈
/ [inf w∈C α(w), supw∈C α(w)].
Proof. Without proof. See any textbook on functional analysis. 

Corollary. For any Banach space it holds τco = τw , i.e. vi → v in τco if and only
if vi → v in τw for any net (vi )i∈I .
Proof. By the corollary above it suffices to show that vi → v in τw implies that
vi → v in τco . If C is a convex set then the Hahn–Banach Separation Theorem
implies that there is an α such that
α(v) ∈
/ [ inf α(w), sup α(w)].
w∈C w∈C

In particular, by τw -continuity of α there is an i0 ∈ I such that

α(vi ) ∈
/ [ inf α(w), sup α(w)].
w∈C w∈C

Thus vi ∈ X\C for i ≥ i0 . Now let U ∈ σco with v ∈ U . Then there are closed
convex sets Ck and αk for k = 1, . . . , n such that
U = X\ ∪nk=1 Ck
and by continuity of αk there is an i0 ∈ I such that
αk (vi ) ∈
/ [ inf αk (w), sup αk (w)] for all i ≥ i0 .
w∈Ck w∈Ck

But then vi ∈ / Ck for all i ≥ i0 and k = 1, . . . , n showing that vi ∈ U for i ≥ i0 .

Finally, let V ∈ τco be an arbitrary neighborhood of v then there is a U ∈ σco
such that U ⊂ V . As above there is an i0 ∈ I such that vi ∈ U ⊂ V for all i ≥ i0 .
As V was arbitrary we have shown that vi → v in τco . 

Remark. The proof indicates that it suffices to look at the sets {X\C | C is closed and convex} ⊂
τw which is a subbase for τw . A similar argument below will show that a set is com-
pact if and only if every cover formed by elements of this subbase admits a finite
subcover, this result – called the Alexander Subbase Theorem – holds more gener-
ally for topologies obtained from a subbase.
4.11. Difference quotients of Sobolev functions. Given i ∈ {1, . . . , n}, a real
number h and a function u ∈ Lp (Ω) we define the following
u(x + hei ) − u(x)
∂ih u(x) = .
h
It is easy to see that ∂ih u ∈ Lp (Ω) if u ∈ Lp (Ω). In general, however, the Lp -norm
of ∂ih u is unbounded as can be seen from the following proposition.
Proposition 4.58. Assume u ∈ Lp (Ω). Then u ∈ W 1,p (Ω) if and only if there is
a K > 0 such that for all Ω0 ⊂⊂ Ω, 0 < h < d(Ω0 , ∂Ω) and i ∈ {1, . . . , n} it holds
k∂ih ukLp (Ω0 ) ≤ K.
Remark. If u ∈ W 1,p (Ω) then one can show that K = kDukp sastisfies the assump-
tion.
 LECTURE NOTES „LINEAR PARTIAL DIFFERENTIAL EQUATIONS“ 49

Proof. We first show that

k∂ih ukp ≤ k∂i ukp .
Indeed, if u ∈ C 1 (Ω) then for fixed x ∈ Ω0 and h < d(Ω0 , ∂Ω) it holds
ˆ
u(x + hei ) − u(x) p 1 h
| 1 | ≤ |∂i u(x + tei )|p dt
h
h 0
which yields
ˆ ˆ ˆ h
1
|∂ih u(x p
+ hei )| dx ≤ |∂i u(x + tei )|p dtdx
Ω0 Ω0 h 0
ˆ h ˆ
1
= |∂i u(x + tei )|p dxdt
h 0 Ω0
ˆ h ˆ
1
≤ |∂i u(x + tei )|p dxdt
h
ˆ 0 Ω

= |∂i u(x + tei )|p dxdt

Ω

Assume now that k∂ih ukLp (Ω0 ) ≤ K for all Ω0 ⊂⊂ Ω. Fix Ω0 and observe that
the bound implies that ∂ihn u * gi weakly in Lp (Ω0 ) for a sequence hn → 0.
Let ϕ ∈ Cc1 (Ω0 ) and hn < d(Ω0 , ∂Ω) for all large n. Then
ˆ ˆ
∂ihn u · ϕdx = − u · ∂i−hn ϕdx.
Ω
By the weak convergence we see that the left hand side converges to
ˆ
gi · ϕdx.
Ω0

Since ϕ has compact support and is differentiable we have ∂i−hn ϕ → ∂i ϕ uniformly

implying that ˆ ˆ
gi · ϕdx = − u · ∂i ϕ
Ω0 Ω0
which shows that gi is a weak derivative of u restricted to Ω0 . Furtermore, gi is
bounded by K. Since weak derivatives are unique on Ω1 ∩ Ω2 for two domains
Ω1 , Ω2 ⊂⊂ Ω we see that gi can be defined on all of Ω with bound only depending
on K. In particular, u has weak derivatives gi ∈ Lp (Ω), i = 1, . . . , n, implying that
u ∈ W 1,p (Ω). 

4.12. (not covered) Maximal Function Theorem. The following can be de-
rived using the fact that any open set is given by
[
U= Bq (xn )
n∈N,q∈Q,Bq (xn )⊂U.

where {xn }n∈N is a fixed dense subset of Rn .

Lemma 4.59. Given any open cover {Ui }i∈I of a subset E of Rn there is an at
most countable I 0 ⊂ I such that {Ui }i∈I 0 still covers E.
In the following given a ball B = Br (x) for some x ∈ Rn and r > 0 we denote
by kB the ball Bkr (x).
 LECTURE NOTES „LINEAR PARTIAL DIFFERENTIAL EQUATIONS“ 50

Lemma 4.60 (5r-Covering Lemma). Assume I ⊂ N and {Bi }i∈I is a collection of

balls with sup{rad Bi } < ∞. Then there is a subset of indices J ⊂ I such that the
subcollection {Bi }i∈J of balls is disjoint and
[ [
Bi ⊂ 5Bi .
i∈I i∈J

Exercise (Vitali Covering Lemma for finitely many balls). Show that ifSI is finite
S J ⊂ I can be chosen such that the balls {Bi }i∈J are disjoint and i∈I Bi ⊂
then
i∈J 3Bi .

Proof of the lemma. Let R = sup{rad Bi } and we define

In = {i ∈ I | Bi ∈ (2−(n+1) R, 2−n ].
We define inductively sequences Jn and Kn as follows: Let K0 = I0 and J0 ⊂
K0 such that {Bi }i∈J0 is disjoint and for any i ∈ K0 \J0 there is a i0 ∈ J0 with
Bi ∩ Bi0 = ∅, i.e. {Bi }i∈J0 is maximal disjoint subcollection of {Bi }i∈K0 . Assume
now Kn and Jn are constructed. Define
Kn+1 = {i ∈ In+1 | Bi ∩ Bj = ∅ for all j ∈ J0 ∪ . . . ∪ Jn }.
As above choose Jn+1 such that {Bi }i∈Jn+1 is a maximal disjoint subcollection of
{Bi }i∈Kn+1 . S
We claim that J = n∈N Jn sastisfies the assumptions of the claim. Indeed, by
construction the collection {Bi }i∈N is disjoint. Furthermore, for i ∈ I\J there is
an n ∈ N such that i ∈ In . By triangle inequality it suffices to show that there is
an j ∈ J such that Bi ∩ Bj 6= ∅.
Since i ∈ In either Bi ∩ Bj 6= ∅ for some j ∈ J0 ∪ . . . ∪ Jn−1 or i ∈ Kn . In the
latter case, the choice of Jn implies that the collection
{Bi } ∪ {Bj }j∈Jn
is not disjoint, i.e. Bi ∩ Bj = ∅ for some j ∈ Jn . 
Definition 4.61 (Hardy–Littlewood maximal function). Given a measurable func-
tion u : Rn → R and R ∈ (0, ∞], the (Euclidean) maximal functions MR f are
defined as
MR u(x) = sup |u|dx.
r∈(0,R) Br (x)

Remark. If u : Ω → R is measurble we may assume u = 0 outside of Ω . This allows

us to apply the maximal operator M also functions defined only on subdomains of
Rn .
It is not difficult to see that for u, v : Rn → R measurable and λ ∈ R it holds
MR (u + v) ≤ MR u + MR v
and
MR (λu) = |λ|MR u.
Theorem 4.62 (Hardy–Littlewood). Let u : Rn → R be a measurable function.
Then the following statements hold:
(1) There is a C1 ∈ (0, ∞) such that for all t > 0 it holds
ˆ
n C1
λ ({MR u > t}) ≤ |u|dx.
t
 LECTURE NOTES „LINEAR PARTIAL DIFFERENTIAL EQUATIONS“ 51

(2) For all p > 1 there is a Cp ∈ (0, ∞) such that

kMukp ≤ Cp kukp .

Proof. Since the bounds are independent of R ∈ (0, ∞] and M∞ u = limR→∞ MR u

it suffices to assume R ∈ (0, ∞).
/ L1 (Rn ) then there is nothing to prove. Assume u ∈ L1 (Rn ).
(1): If u ∈
By definition, for all x ∈ {MR u > t} there is an rx ∈ (0, R) such that

|u| ≥ t.
Brx (x)

Since {BrX }x∈Et is a cover of Et we first pick a countable subcover and then use
the 5r-Covering Lemma to obtain disjoint ball {Bi }i∈I such that

[
Et ⊂ 5Bi .
i∈I

Then

X
λn ({MR u > t}) ≤ λn (5Bi )
i∈I
X
≤ 5n λn (Bi )
i∈I
ˆ ˆ
5n X 5n
≤ |u|dx ≤ |u|dx.
t Bi t
i∈I

/ Lp (Rn ) then there is nothing to prove. So assume u ∈ Lp (Rn ).

(2): Again if u ∈
Furthermore, since Mu = M|u| we may assume u ≥ 0.
For t > 0 note that

t t
u ≤ (u − )+ +
2 2

where (u − 2t )+ = max{u − 2t , 0}. Then

t t
{MR u > t} ⊂ {MR (u − )+ > }.
2 2
 LECTURE NOTES „LINEAR PARTIAL DIFFERENTIAL EQUATIONS“ 52

Using Cavalleri’s Principle, the bound obtained for p = 1 appied to (u − 2t )we

obtain
ˆ ˆ ∞
|MR u|p dµ = p tp−1 λn ({Mr u > t})dt
ˆ ∞
0
t t
≤p tp−1 λn ({Mr (u − )+ > })dt
2 2
0
ˆ ∞
= p2p tp−1 λn ({Mr (u − t)+ > t})dt
0
ˆ ∞ ˆ
p p−2
≤ C1 p2 t uχ{u>t} dxdt
n
ˆ0 ˆ ∞ R
= C1 p2p tp−2 uχ{u>t} dtdx
Rn 0
ˆ ˆ u(x)
p
= C1 p2 u(x) tp−2 dtdx
ˆ Rn 0
p2p
= C1 up dx.
p−1
 1
p2p p
so that the result holds for Cp = C1 p−1 . 

5. L2 -Estmates
In the following let L be a uniformly elliptic operator such that aij , bk , c : Ω̄ → R
are Lipschitz continuous. Furthermore, let L0 with aij ij k
0 = a and b0 = c0 = 0. Note
1,2
that in this case if u ∈ W (Ω) is a weak solution to Lu = α for a bounded linear
map on W01,2 (Ω), i.e. it satisfies
ˆ X n n
X
BL (u, v) = aij · ∂i u · ∂i v + bk · ∂k u · v + c · u · vdx = α(v) ∀v ∈ W01,2 (Ω)
i,j=1 k=1

then u is also a weak solution to L0 u = αu where

ˆ X n
αu (v) = α(v) − bk · ∂k u · v + c · u · vdx
k=1
ˆ X
n
= α(v) + (bk · u · ∂k v + ∂k bk · u · v) − c · u · vdx
k=1

Note that
n
X
|αu (v)| ≤ |α(v)| + ( kbk kC 0,1 + kckk ) · kuk2 · kvkW 1,2
k=1

which implies that αu is also a bounded linear map on W01,2 (Ω) with norm de-
pending only on kuk2 and the bounds on bk and c. Note that a natural norm for
bounded linear maps α : W01,2 (Ω) is given by
kαk = sup |α(v)|.
v∈W01,2 (Ω)\{0}

The argument show that it suffices to look at weak solutions L0 u = α.

 LECTURE NOTES „LINEAR PARTIAL DIFFERENTIAL EQUATIONS“ 53

Proposition 5.1 (Cacciopoly inequality). Let u ∈ W 1,2 (Ω) be L0 u = α weakly,

i.e.
ˆ X n
− aij · ∂i u · ∂j vdx = α(v)
i,j=1

for all v ∈ W01,2 (Ω). Then for all Ω0 ⊂⊂ Ω it holds

kukW 1,2 (Ω0 ) ≤ C(Ω, Ω0 , λ, Λ, n) kαk + kukL2 (Ω)

where
|α(v)|
kαk = sup .
v∈W01,2 (Ω)\{0} kvkW 1,2

Proof. Let η : Ω → [0, 1] be a smooth map with compact support in Ω such that
η ≡ 1 on Ω0 . It is not difficult to see that the derivatives of η are bounded by a
constant C only depending on the distance of Ω0 and ∂Ω.
Now it is easy to see that v = η 2 u is a function in W01,2 (Ω) so that
BL0 (u, v) = α(v).
From Gagliardo–Nirenberg we obtain
kvkW 1,2 ≤ C(Ω, n) · kD(η 2 u)k2
≤ C(Ω, n) · kη 2 |Du|k2 + kηkC 1 kuk2

≤ C(Ω, Ω0 , n)kη|Du|k2 .
In particular, ˆ
1
|α(v)| ≤ kαk · kvk ≤ kαk +  η 2 |Du|2 dx.
2
4
Since η ≡ 1 on Ω0 and L0 is uniformly elliptic we obtain
ˆ ˆ X n
λ η 2 · |Du|2 dx ≤ η 2 · aij · ∂i u · ∂j udx
Ω Ω i,j=1
ˆ Xn ˆ X
ij
= a · ∂i u · ∂j vdx − aij · ∂i u · ∂j η · 2η · udx
Ω i,j=1 i,j=1
ˆ
= α(v) + Λ η · |Du| · |Dη| · |u|dx

≤ kαkkvkW 1,2 + Λkη · |Du|k2 · k2 · |Dη| · |u|k2

ˆ
1
≤ (kαk2 + ΛkηkC 1 kuk22 ) + (1 + Λ) η 2 · |Du|2 dx.
4
λ
Now choosing  small (only depending on λ and Λ) such that (1 + Λ) < 2 we
obtain
ˆ
λ 1
η 2 · |Du|2 ≤ (kαk2 + ΛkηkC 1 kuk22 ).
2 4
Since  and kηkC 1 only depend on the data and η ≡ 1 on Ω0 we obtain
ˆ ˆ  
u2 dx + |Du|2 dx ≤ C(Ω, Ω0 , λ, Λ, n) kαk2 + kuk2L2 (Ω) .
Ω0 Ω0

 LECTURE NOTES „LINEAR PARTIAL DIFFERENTIAL EQUATIONS“ 54

Theorem 5.2 (L2 -estimate). Assume u is a weak solution to L0 u = f in Ω. If aij

2,2
are Lipschitz functions in Ω̄ then u ∈ Wloc (Ω) and for all Ω0 ⊂⊂ Ω it holds
kukW 2,2 (Ω0 ) ≤ C(Ω, Ω0 , λ, Λ, n)(kαkW 1,2 (Ω)∗ + (1 + kaij kC 0,1 ) · kukL2 (Ω) ).
0

k
In particular, if Lu = f and b is also Lipschitz on Ω̄ then
n
X n
X
− ∂j (aij ∂i u) + bk ∂k u + cu = f almost everywhere in Ω.
i,j=1 k=1

2,2
Proof. Note that the last result follows by observing that u ∈ Wloc (Ω) and
ˆ
BL (u, ϕ) = f ϕdx

for all ϕ ∈ Cc∞ (Ω) implies that we can apply partial integration to the quadratic
term in BL which yields
ˆ ˆ
 
Xn n
X
− ij k
∂j (a ∂i u) + b ∂k u + cu · ϕdx = f ϕdx.

i,j=1 k=1

Back to the equation L0 u = α. Let Ω0 ⊂⊂ Ω and choosing in the following

0 < h < d(Ω0 , ∂Ω). Then for all v ∈ W01,2 (Ω0 ) it holds v(x ± hek ) ∈ W01,2 (Ω).
Note that for v ∈ W01,2 (Ω0 ) it holds
ˆ X n ˆ X n
aij ∂i (∂kh u)∂j vdx = − ∂i u∂k−h aij · ∂j v dx


i,j=1 i,j=1
ˆ X
n ˆ X
n
=− a ij
∂i u∂j (∂k−h v)dx − (∂k−h aij )∂i u · ∂j v(· − hek )dx
i,j=1 i,j=1
ˆ ˆ X
n
= f (∂k−h v)dx − (∂kh aij )∂i u(· + hek ) · ∂j vdx =: αh (v),
i,j=1

i.e. L0 ∂kh u = −αh weakly.

Let now Ω00 such that Ω0 ⊂⊂ Ω00 ⊂⊂ Ω and 0 < h < d(Ω0 , ∂Ω00 ) then from the
Cacciopoli inequality we obtain
αh (v) ≤ kf k2 k∂k−h vkL2 (Ω00 ) + kaij kC 0,1 (Ω00 ) · kukW 1,2 (Ω00 ) · kvkW 1,2 (Ω)
0

≤ (kf k2 + kaij kC 0,1 (Ω00 ) · kukL2 (Ω) ) · kvkW 1,2 (Ω) .

0

Thus Cacciopoli gives

k∂kh ukW 1,2 (Ω0 ) ≤ C(Ω, Ω00 , Ω0 , Λ, λ, n) kf k2 + (1 + kaij kC 0,1 (Ω00 ) ) · kukL2 (Ω) .

Finally, we may choose Ω00 such that the constant does not depend on Ω00 anymore.


Corollary 5.3 (Higher L2 -regularity). Assume f ∈ W k,2 (Ω) and Lu = f weakly.

k+2,2
If aij , bk ∈ C k−1,1 (Ω) and c ∈ C k−1 (Ω) then u ∈ Wloc (Ω) and for all Ω0 there is
a constant C depending the data such that
kukW k+2,2 (Ω0 ) ≤ C(kf kW k,2 (Ω) + kukL2 (Ω) ).
 LECTURE NOTES „LINEAR PARTIAL DIFFERENTIAL EQUATIONS“ 55

Proof. We only give the idea for the W 3,2 -regularity: From the proof above we have
ˆ X n ˆ ˆ X n
aij ∂i (∂kh u)∂j vdx = − ∂kh f vdx − (∂kh aij )∂i u(· + hek ) · ∂j vdx
i,j=1 i,j=1

so that
ˆ X n ˆ ˆ X
n
aij ∂i (∂k u)∂j vdx = − ∂kh f vdx − (∂k aij ∂i u) · ∂j vdx
i,j=1 i,j=1
ˆ n
X ˆ
=− (∂k f − ∂j (∂k aij ∂i u)) · vdx = fˆi · vdx.
i,j=1

If v ∈ W01,2 (Ω0 ) for Ω0 ⊂⊂ Ω then

n
X
kfˆk2 ≤ kaij kC 1,1 kukW 2,2 .
i,j=1

2,2
This shows that ∂k u ∈ Wloc (Ω). Using nested domains we can therefore bound the
0
W -norm of u on Ω ⊂ Ω by the C 1,1 -norm of aij . and the L2 -norms of f and u
3,2

as well. 

Corollary 5.4 (Inner C ∞ -regularity). Assume f ∈ C ∞ and Lu = f weakly. If

aij , bk , c ∈ C ∞ (Ω) then u ∈ C ∞ (Ω).
k,2
Proof. Just observe that u ∈ Wloc (Ω) for all k ∈ N. Thus using a smooth cut-off
function η which is constant equal to one on Ω0 ⊂⊂ Ω we see ηu ∈ W0k,2 (Ω) for
all k ∈ N so that via Sobolev and Morrey embedding we obtain ηu ∈ C l (Ω) for all
l ∈ N. Since u = ηu on Ω0 , we see that u is differentiable of all orders in Ω0 (more
precisely u agrees with an infinitely differentiable function on Ω0 ). 

Lemma 5.5 (boundary Cacciopoli inequality). Let Ω = B2 (0)+ = {x ∈ B2 (0) | xn ≥

0} and L0 be as above and Then for any bounded linear map α : W01,2 (B2 (0)+ ) → R
and ϕ ∈ W 1,2 (B2 (0)+ ). Then there is a constant C = C(λ, Λ, n) such that for any
weak solution L0 u = α with u − ϕ ∈ W01,2 (B2 (0)+ ) it holds
kukW 1,2 (B1 (0)+ ) ≤ C(kαk + kϕkW 1,2 (B2 (0)+ ) + kukL2 (B2 (0)+ ) ).
Proof. Note that because
kukW 1,2 (B1 (0)+ ) ≤ ku − ϕkW 1,2 (B1 (0)+ + kϕkW 1,2 (B2 (0)+ )
and
BL (u − ϕ, v) = α(v) − BL (ϕ, v) =: αϕ (v)
where kαkϕ ≤ kαk + kϕkW 1,2 (B2 (0)+ ) , it suffices to show the estimate for ϕ ≡ 0.
In this case observe that for a smooth cut-off function η : B2 (0) → [0, 1] with
η ≡ 1 on B1 (0) the function v := η 2 u is in W01,2 (B2 (0)+ ). Thus the same argument
as in proof of the Cacciopoli inequality applies. 

Theorem 5.6 (Global regularity). Assume Ω is a bounded open domain with

C k+1,1 -boundary. If L is uniformly elliptic and aij , bk ∈ W k+1,∞ (Ω) and c ∈
W k−1,∞ (Ω) then for any f ∈ W k,2 (Ω) and ϕ ∈ W k+2,2 (Ω̄) there is an u ∈
 LECTURE NOTES „LINEAR PARTIAL DIFFERENTIAL EQUATIONS“ 56

W k+2,2 (Ω) solving the PDE Lu = f (weakly) in Ω and having trace T u ∈ W k,2 (∂Ω)
equal to T ϕ. Furthermore, it holds
kukW k+2,2 ≤ C(Ω, λ, Λ, n)(kf kW k,2 + kϕkW k+2,2 + kukL2 ).
Sketch of the Proof. If Ω = B2 (0) the proof follows exactly as the interior regularity
proof by using the boundary Cacciopoli inequality. For the general case note that
locally near every x0 ∈ ∂Ω we can use C 1,1 -diffeomorphism Ψ : Ux+0 → B2 (0)+ such
that ũ = u ◦ Ψ solves a uniformly elliptic PDE. More precisely, there is a uniformly
elliptic operator L̃ with
n
X
ãkl = (aij ∂i Ψl ∂j Ψk ) ◦ Ψ−1 · | det(DΨ−1 )|
i,j=1
X n
b̃p = (bk ∂k Ψp ) ◦ Ψ−1 · | det(DΨ−1 )|
k=1

and
c̃ = c ◦ Ψ−1 · | det(DΨ−1 )|
f˜ = f ◦ Ψ−1 · | det(DΨ−1 )|
and it holds L̃ũ = f˜ weakly.
Since Ψ ∈ C 1,1 and aij , bk ∈ C 0,1 it holds ãkl , b̃p ∈ C 0,1 with norms depending
only on Ψ. Similarly, the uniform ellipticity constants of L̃ only depend on Ψ and
the constants of L.
Now for k = 1, . . . , n − 1 we can apply the argument of the inner regularity to
conclude that ∂k ũ ∈ W 1,2 (B1 (0)) with corresponding bounds depending on the
L2 -bounds of ũ and f˜ .
For k = n observe that
 
X n
X
nn −1 
∂nn ũ = (ã ) − kl
ã ∂kl ũ − (∂k ã )∂l ũ + f˜ in B2 (0)+ .
kl

(k,l)6=(n,n) k,l=1

nn −1
Since (ã ) is bounded by a constant depending only on L and Ψ, we see that
the Cacciopoli inequality gives
k∂nn ũkL2 (B (0)+ ) ≤ C(Ψ, λ, Λ)(kf˜kL2 (B (0)+ ) + (1 + kãkl kC 0,1 )kukL2 (B (0)+ ) )
1 2 1

2,2 +
which gives the desired W -bound of ũ on B1 (0) .
Using the uniformy bounds on the derivatives of Ψ we then obtain W 2,2 -bounds
for u on Ux0 .
Since Ω̄ is compact we may find finitely many Ui so that the W 2,2 -norm is
bounded which gives the desired global W 2,2 -bound. 

6. Properties of weak (sub)solution

Definition 6.1. Let L be a uniformly elliptic operator on a bounded domain Ω
with bk = c = 0 then u ∈ W 1,2 (Ω) is called a weak subsolution of the equation
Lu = f (short Lu ≥ f weakly) if for all v ∈ W01,2 (Ω) it holds
ˆ Xn ˆ
BL (u, v) := aij · ∂i u · ∂j udx ≤ − f · vdx.
i,j=1
 LECTURE NOTES „LINEAR PARTIAL DIFFERENTIAL EQUATIONS“ 57

Note that if u ∈ W 1,2 (Ω) then also u+ = max{u, 0}, u− = max{−u, 0} and
|u| = u+ + u− are in W 1,2 (Ω). Now if Ω has a Lipschitz boundary then the trace
T : W 1,2 (Ω) → L2 (Ω) is well-defined and u ∈ W01,2 (Ω) if and only if T u = 0. Thus
by linearity we have
T u = T (u+ ) − T (u− )
and T u± ≥ 0 which implies
ess sup∂Ω T u = kT (u+ )kL∞ (∂Ω) = inf{M ∈ R | (u − M )+ ∈ W01,2 (Ω)}.
Note finally that the right hand expression makes sense in general.
Proposition 6.2 (Weak maximum principle). If Lu ≥ 0 weakly for u ∈ C 0 (Ω) ∩
W 1,2 (Ω) and (u − M )+ ∈ W01,2 (Ω) then u ≤ M in Ω.
Proof. The assumptions imply v = (u − M )+ can be used as test function. Thus
ˆ X n ˆ X
n
ij
a · ∂i u · ∂j udx = aij · ∂i u · ∂j vdx ≤ 0.
{u>M } i,j=1 i,j=1

Since L is uniformly elliptic we see that ∂i u = 0 on {u > M }. This, however implies

that u is locally constant in the open set {u > M }. If {u > M } were non-empty
this would give a contradiction as Ω is connected. 
Remark. The statement also holds for u ∈ W 1,2 (Ω) with u ≤ M a.e. in Ω. For this
observe if ∂i v = ∂i u almost everywhere on {v = u} whenever u, v ∈ W 1,2 (Ω).
Theorem 6.3 (Strong Maximum Principle). If Lu ≥ 0 weakly for u ∈ C 0 (Ω) ∩
W 1,2 (Ω) and (u − M )+ ∈ W01,2 (Ω) then either u ≡ M or u < M in Ω.
Proof. Assume C = {x ∈ Ω | u(x) = M } is non-empty and not equal to Ω. Then
there is a x0 ∈ Ω and y ∈ C such that
d(x0 , y) = inf
0
d(x, y 0 ).
y ∈C
x0 +y
Furthermore, let x1 = then Br (x) ⊂ {u < M } and ∂Br (x) ∩ C = {y}.
2
As in Lemma 3.7, choose α  1 such that
Lv(x) ≥ 0
2 2
for v(x) = e−αkx−yk − e−αr .
Choose s  1 such that Ω0 := Bs (y) ⊂ Ω and observe that v < 0 on ∂Ω0 \B̄r (x1 )
and u < M = u(y) on ∂Ω0 ∩ B̄r (x1 ). Now choose  > 0 such that
M − inf ∂Ω0 ∩B̄r (x1 ) u
<
sup v
then
u := u + v ∂Ω0
< M = u(y) + v(y).
In particular, there is an M̃ < M such that (u − M̃ )+ ∈ W01,2 (Ω0 ). However, the
weak maximum principle implies
u ≤ M̃
which contradicts the fact that y ∈ Ω0 and the u (y) = M > M̃ . 
We also obtain the following result. The proof is similar to the one below using
the weak maximum principle.
 LECTURE NOTES „LINEAR PARTIAL DIFFERENTIAL EQUATIONS“ 58

Corollary 6.4 (A priori bounds). Let L be a uniformly elliptic operator with bk =

k
c = 0 on a bounded domain Ω such that β := sup |bλ | < ∞ and M ≥ 0. Then for
all u ∈ W 1,2 (Ω) ∩ C 0 (Ω) with Lu ≥ f weakly and (u − M )+ ∈ W01,2 (Ω) it holds
C
sup u ≤ M + sup |f − |
Ω λ Ω
(β+1) diam Ω
where C = e − 1.
Proposition 6.5 (Limits of subsolutions). Assume (un )n∈N is a sequence in W 1,2 (Ω)
converging weakly to u ∈ W 1,2 (Ω). If Lun ≥ f weakly then Lu ≥ f weakly.
Proof. Note that un * u weakly implies that for all bounded linear functionals
α : W 1,2 (Ω) → R it holds
lim α(un ) = α(u).
n→∞
Observe that uniform ellipticity of L implies
u 7→ BL (u, v)
is a bounded linear functional for every fixed v ∈ W01,2 (Ω). In parrticular, we have
ˆ ˆ
BL (u, v) = lim BL (un , v) ≥ lim − f · vdx = − f · vdx
n→∞ n→∞

proving that Lu ≥ f weakly. 

7. Schauder esimates
The Sobolev regularity theory shows that provided the coefficients behave nicely
and f is sufficiently many times weakly differentiable (and thus Hölder continuous)
then Lu = f implies u is a classical solution with Hölder continuous second deriva-
tives. Via so-called Lp -estimates it is possible to obtain W 2,p -bounds which would
yield C 1,α -solutions provided f is a bounded function.
More classical it is possible to directly obtain C 2,α -bounds if f is a C α -function.
Those estimates are called Schauder estimates. In order to simplify the notation we
define the following semi-norm for α ∈ (0, 1] and u : Ω → R (we omit the domain if
it is clear from the context)
|u(x) − u(y)|
[u]α,Ω = sup
x6=y∈Ω d(x, y)α
and X X
kukk,α,Ω = k∂I uk0 + [∂I u]∗α,Ω
|I|≤k |I|=k

Note that it is easy to see that [u]∗α,Ω

= 0 implies u ≡ const.
If u ∈ C k,α (Ω) then we use the short hand notation
X
[Dk u]α,Ω = [∂I u]∗α,Ω .
|I|=k

Lemma 7.1 (Ehrling’s Lemma). Assume X, Y, Z are three Banach spaces and there
are two bounded linear maps K : X → Y and I : Y → Z such that K is compact
and I is injective. Then for all  > 0 there is a C = C() > 0 such that
kKxkY ≤ kxkX + C · kI(Kx)kZ .
 LECTURE NOTES „LINEAR PARTIAL DIFFERENTIAL EQUATIONS“ 59

Proof. If the claim was wrong then fixed  > 0 there is a sequence (xm )m∈N in X
such that
kxm kX + m · kI(Kxm )kZ ≤ kKxm kY = 1.
Since K compact there is a subsequence such that Kxml → y. Note that 1 =
limn→∞ kKxml k = kyk which implies y 6= 0. By continuity it holds
1
kIyk ≤ lim kI(Kxm )k ≤ lim = 0.
m→∞ m
m→∞

Then injectivity of I shows y = 0 which is a contradiction. 

By Arzela–Ascoli we obtain the following.

Corollary 7.2. For all bounded open sets Ω and all  > 0 there is a constant
C = C() > 0 such
kukC 2 ≤ kukC 2,α + C()kukC 0
and
kukC 2 ≤ kukC 2,α + C()kukL2 .
Lemma 7.3. Assume u is a functions satisfying the following: for all  > 0 is an
R > 0 such that
|u(x)| ≤ kxk
whenever x ∈ / BR (0) (in short |u(x)| = o(kxk) as kxk → ∞). If u is harmonic
then u is constant.
Proof. Let x, y ∈ Rn then
|BR (x)∆BR (y)| = |BR (x)\BR (y) ∪ BR (y)\BR (x)|
≤ d(x, y)|∂BR (0) = Cn d(x, y)Rn−1 .
Let  > 0 and R > (R + kxk + kyk). Then BR (x)∆BR (y) ⊂ Rn \BR and the
mean-value property shows
ˆ
0 1
|u(x) − u(x )| ≤ |u(z)|dz
|BR (0)| BR (x)∆BR (y)
≤ 2C̃n d(x, y).
As  is arbitrary we see that u(x) = u(y) which implies the claim. 

Corollary 7.4. Assume u is harmonic on Rn and [Dk u]α < ∞ for k ∈ N and
α ∈ (0, 1). Then u is a polynomial of order at most k, i.e. Dk u ≡ const.
Proof. Since ∂I u is harmonic if u is harmonic, it suffices to show that a harmonic
function with [u]α ≤ C < ∞ is constant. Now the bound implies
|u(x)| ≤ |u(x) − u(0)| + |u(0)|
≤ C(1 + kxkα )
which shows that u satisfies the conditions of the previous lemma. In particular, u
must be constant. 

Proposition 7.5. For all u ∈ C 2,α (Rn ) then it holds

[D2 u]α,Ω ≤ C(n, α)[∆u]α,Ω .
 LECTURE NOTES „LINEAR PARTIAL DIFFERENTIAL EQUATIONS“ 60

2,α
Proof. If the claim was wrong then there is a sequence (um )m∈N in Cloc (Rn ) such
that
m[∆um ]α,Ω ≤ [D2 um ]α,Ω < ∞.
Via rescaling and shifting we can assume [D2 um ]α,Ω = 1 and
|∂I um (y) − ∂I um (0)| 1
sup sup ≥ .
|I|=2 y∈Rn kykα 2
Furthermore, extracting a subsequence (uml )l∈N we find a multi-index I with |I| =
2, a direction ei and sequence (hl )l∈N such that
|∂I uml (hl · ei ) − ∂I uml (0)| 1
≥ 2 > 0.
hαl 4n
Set
ũl (x) = h2−α
l uml (hl · x)
then still
m[∆ũl ]α,Ω ≤ [D2 ũm ]α,Ω < ∞
and
[D2 um ]α,Ω = 1.
Furthermore, |∂I ũl (ei ) − ∂I ũl (0)| ≥ 4n1 2 . In addition, we can find a quadratic
polynomial vl = hx, b + Axi + c such that
ûl (0) = 0
Dûl (0) = 0
D2 ûl (0) = 0
and choosing vl appropriately we also have
û(ei ) 6= û(0)
where ûl = ũl − vl . Note that [D2 ûl ]α = [D2 ũl ]α = 1 so that we can extract
a subsequence (ûlk )k∈N converging locally uniformly in C 2 (Rn ) to a function u
satisfying
∆u = const
[D2 u]α ≤ 1
∂I u(ei ) 6= ∂I u(0)
Du(0) = 0
D2 u(0) = 0.
Thus u is harmonic and previous result D2 u ≡ const. However, this means that u
would be constant, contradicting the fact that ∂I u(ei ) 6= ∂I u(0). 
Corollary 7.6. If A is a symmetric positive matrix with λIn ≤ A ≤ ΛIn then
[D2 u]α ≤ C(n, α, λ, Λ)[LA u]α .
where LA is the elliptic operator in non-divergence form obtained from A.
Proposition 7.7. Let L be an uniformly elliptic opertor in non-divergence form
such that aij ∈ C α (B2 (0)) then for all u ∈ C 2,α (B2 (0)) it holds
kD2 ukC 2,α (B1 (0)) ≤ C(n, α, λ, Λ, [aij ]α )(kLukC 0,α (B2 (0)) + kukC 2 (B2 (0)) ).
 LECTURE NOTES „LINEAR PARTIAL DIFFERENTIAL EQUATIONS“ 61

Proof. Let η be a smooth cut-off function for B1 (0) ⊂ B2 (0) and define a new
cut-off function by
 
x − x0
ηx0 ,ρ = η .
ρ
Then
v = u · ηx0 ,ρ ∈ C02,α (B2ρ (x0 ))
for x0 ∈ B1 (0) and B2ρ (x0 ) ⊂ B2 (0). Furthermore, v = u on Bρ (x0 ).
Let L0 be the elliptic operator obtained via A0 = (aij (x0 ))ni,j=1 . Let
X
L0 v = Lv − (aij − aij (x0 ))∂ij v.

Thus
[L0 v]α ≤ [Lv]α + kD2 vkC 0 · [aij − aij (x0 )]α + kaij − aij (x0 )kC 0 (B2ρ (x0 )) · [D2 v]α .
≤ [Lv]α + [aij ]α · kvkC 2 + 2ρα · [aij ]α · [D2 v]C α .
Also observe that
[D2 v]α ≤ C1 (n, α, λ, Λ)[L0 v]α .
Thus choosing ρ small (depending only on gives (n, α, λ, Λ, [aij ]α )) gives
[D2 v]α ≤ C1 (n, α, λ, Λ, [aij ]α )([Lv]α + kvkC 2 ).
Now observe
kvkC 2 ≤ C(n) · kukC 2 · kηx0 ,ρ k ≤ C2 (n, α, λ, Λ, [aij ]α )kukC 2 (B2 (0)
and
X
Lv = Lu · ηx0 ,ρ + u · Lηx0 ,ρ + 2 aij ∂i u∂j ηx0 ,ρ
which implies
[Lv]α ≤ C3 (n, α, λ, Λ, [aij ]α )(kLukC 0,α (B2 (0)) + kukC 1,α (B2 (0)) )
≤ C3 (n, α, λ, Λ, [aij ]α )(kLukC 0,α (B2 (0)) + kukC 2 (B2 (0)) )
Finally
[D2 u]α,B1 (0) ≤ sup [D2 v]α,B1 (0)
x0 ∈B1 (0)

≤ C4 (n, α, λ, Λ, [aij ]α )(kLukC 0,α (B2 (0)) + kukC 2 (B2 (0)) ).

To conclude note
kukC 2,α (B1 (0)) = [D2 u]α,B1 (0) + kukC 2 (B2 (0))
≤ (1 + C4 (n, α, λ, Λ, [aij ]α ))(kLukC 0,α (B2 (0)) + kukC 2 (B2 (0)) ).


Remark. Note that the cut-off function only depends on the distance d(B1 (0), ∂B2 (0))
so that we can prove a similar estimate for
kukC 2 (Ω0 ) ≤ C(n, α, λ, Λ, [aij ]α , d(Ω0 , ∂Ω)) · kLukC 0,α (Ω) + kukC 2 (Ω) .


 LECTURE NOTES „LINEAR PARTIAL DIFFERENTIAL EQUATIONS“ 62

Theorem 7.8. For any uniformly elliptic operator with bk = c = 0 there is a

constant C = C(n, α, λ, Λ, [aij ]α ) > 0 such that for all u ∈ C 2,α (Ω) it holds
sup min{d(x, ∂Ω), 1}2 |D2 u(x)| ≤ C kLukC 0,α (Ω) + kukC 0 (Ω)


x∈Ω
and
n
sup min{d(x, ∂Ω), 1} 2 +2 |D2 u(x)| ≤ C kLukC 0,α (Ω) + kukL2 (Ω) .


x∈Ω

In particular, for all Ω0 ⊂⊂ Ω there is a constant C̃ = C(n, α, λ, Λ, [aij ]α , d(Ω0 , ∂Ω))

such that (

kLukC 0,α (Ω) + kukC 0 (Ω)
kukC 2,α (Ω0 ) ≤ C̃ ·

kLukC 0,α (Ω) + kukL2 (Ω) .

Proof. The L2 -estimate only depends on an application of Ehrlings Lemma for the
inclusion C 2,α → C 2 → L2 and the scaling property of the L2 -norm, more precisely
n
if ũ(x) = u(x0 + ρx) with B2ρ (x0 ) ⊂ Ω we get kũkL1 (B1 (0)) = ρ 2 kukL1 (Bρ (x0 )) . The
n n
factor ρ 2 is carried along and induces the addition d(x, ∂Ω) 2 -factor in estimate
2
containing the L -norm of u. We leave the details to the reader and only focus on
the C 0 -estimate.
For this define
S = sup min{d(x, ∂Ω), 1}2 |D2 u(x)|.
x∈Ω
Pick x0 ∈ Ω and set ρ = min{ 13 d(x0 , ∂Ω), 31 }
< 1. Then B2ρ (x0 ) ⊂ Ω and for all
x ∈ B2ρ (x0 ) it holds min{d(x, ∂Ω), 1} ≥ ρ. Define a function ũ ∈ C 2,α (B2 (0)) by
ũ(x) = u(x0 + ρx) and observe
kũkC 0 (B2 (0)) ≤ kukC 0 (Ω)
kD2 ũkC 0 (B1 (0)) = ρ2 kD2 ukC 0 (Bρ (x0 ))
and
kL̃ũkC α (B2 (0)) = ρ2 kLukC 0 (B2ρ (x0 )) + ρ2+α [Lu]α,B2ρ (x0 )
≤ kLukC α (Ω)
where L̃ is the rescaled elliptic operator which is still uniformly elliptic with the
same constants as L and [ãij ]α ≤ [aij ]α thus the previous C 2,α - estimate holds for
C = C(n, α, λ, Λ, [aij ]α ).
Then
min{d(x0 , ∂Ω), 1}2 |D2 u(x0 )| = 9ρ2 |D2 u(x0 )|
≤ 9kD2 ũkC 0 (B1 (0))
≤ 9C(kL̃ũkC 0,α (B2 (0)) + kũkC 2 (B2 (0)) )
≤ 9C(kLukC 0,α (Ω) + ρ2 kD2 ukC 0 (B2ρ (x0 )) + Cn,α, kukC 0 (Ω) )
where we use Ehrlings Lemma to estimate
kũkC 2 (B2 (0) ≤ kD2 ũkC 0 (B2 (0)) + Cn,α, kũkC 0 (B2 (0)) .
Note that
ρ2 kD2 ukC 0 (B2ρ (0)) ≤ sup min{d(x, ∂Ω), 1}2 |D2 u(x)|
x∈B2ρ (x0 )
≤ S.
 LECTURE NOTES „LINEAR PARTIAL DIFFERENTIAL EQUATIONS“ 63

1
Thus choosing  = 18C then
1
min{d(x0 , ∂Ω), 1}2 |D2 u(x0 )| ≤ S + C2 (n, α, λ, Λ, [aij ]α )(kLukC 0,α (Ω) + kukC 0 (Ω) )
2
which implies the claim
S ≤ 2C2 (n, α, λ, Λ, [aij ]α )(kLukC 0,α (Ω) + kukC 0 (Ω) ).


Remark (Boundary/Global Regularity). Global regularity is obtained by doing the

same analysis on Rn,+ == {x ∈ Rn | xn ≥ 0} and B2+ (0) = B2 (0) ∩ Rn,+ . If the
boundary of Ω is sufficiently smooth then we may cover Ω by charts and obtain
estimates for a finite cover of Ω which yields the global regularity
kukC 2,α (Ω) ≤ C(n, α, λ, Λ, [aij ]α , Ω)(kLukC 0,α (Ω) + kukC 0 (Ω) )
for u ∈ C02,α (Ω).
If u = ϕ on ∂Ω for some ϕ ∈ C 2,α (Ω) then we also get
kukC 2,α (Ω) ≤ C(n, α, λ, Λ, [aij ]α , Ω)(kLukC 0,α (Ω) + kukC 0 (Ω) + kϕkC 2,α (Ω) ).

8. Method of Continuity
Lemma 8.1 (A priori bounds). Let L be an elliptic operator with c ≤ 0 on a
k
bounded domain Ω such that β := sup |bλ | < ∞. Then for all u ∈ C 2 (Ω) ∩ C 0 (Ω̄)
with Lu ≥ f it holds
C
sup u ≤ sup u+ + sup |f − |
Ω ∂Ω λ Ω
where C = e(β+1) diam Ω − 1.
P ij
a ∂ij + bk ∂k then for α ≥ β + 1 and d = inf x∈Ω x1 it holds
P
Proof. Set L0 =
L0 eα(x1 −d) = (α2 a11 + αb1 )eα(x1 −d) ≥ λα(α − β)eα(x1 −d) ≥ λ in Ω.
Set
|f − |
v = sup u+ + (eα diam Ω − eα(x1 −d) ) sup ≥0
∂Ω Ω λ
Then
|f − |
Lv = L0 v + cv ≥ L0 v ≤ −λ sup
Ω λ
so that
|f − | f
 
L(v − u) ≤ −λ sup + ≤ 0 in Ω.
Ω λ λ
Since (v − u) ∂Ω ≥ sup∂Ω u+ − u|∂Ω ≥ 0, the minimum principle applied to v − u
implies v − u ≥ 0 in Ω. Thus we have
  |f − |
sup u ≤ sup v ≤ sup u+ + e(β+1) diam Ω − 1 sup .
Ω Ω ∂Ω Ω λ

C
Corollary 8.2. If Lu = f then supΩ |u| ≤ sup∂Ω |u| + λ supΩ |f |.
 LECTURE NOTES „LINEAR PARTIAL DIFFERENTIAL EQUATIONS“ 64

Lemma 8.3 (Banach Fixed Point Theorem). Let Ψ : X → X be a map on a

Banach space (X, k · k) to itself. If
kΨ(x) − Ψ(y)k ≤ Kkx − yk
for some K < 1 then there is a unique x0 ∈ X such that Ψ(x0 ) = x0 .
Proof. The condition implies that Ψ is K-Lipschitz. Since K < 1 there can be
at most one fixed point. To obtain existence, observe that for a fixed y0 ∈ X the
sequence (Ψn (y0 ))n∈N is Cauchy and by continuity of Ψ converges a fixed points. 
Proposition 8.4 (Continuity Method). Assume L0 , L1 : X → Y are two bounded
linear maps between two Banach spaces such that for some c > 0 it holds
kLt xkY ≥ ckxkX
where Lt = (1 − t)L0 + tL1 and t ∈ [0, 1]. Then L0 is surjective if and only if L1 is
surjective.
Proof. Assume for some s ∈ [0, 1] the operator Ls is surjective. The condition
implies that Ls must be also injective.
Let t ∈ [0, 1]. Then the equation Lt x = y is equivalent to
Ls x + (s − t)(L1 − L0 )x = y.
Since Ls is bijective this is equivalent to
x = (s − t)L−1 −1
s (L1 − L0 )x + Ls y = Ψs,t,y (x).
Now
kΨs,t,y (x) − Ψs,t,y (y 0 )k ≤ |s − t|kL−1
s k · kL1 − L0 k

≤ c−1 |s − t|(kL0 k + kL0 k)

which implies that Ψs,t,y has a unique fixed point if
c
|s − t| < = c̃.
kL0 k + kL1 k
In particular, if t ∈ (s − c̃, s + c̃) ∩ [0, 1] then Lt is surjective as this bound is
independent of y ∈ Y and the fixed point satifies Lt x = y. Inductively we can show
that Lt is surjective if t ∈ (s − (n + 21 )c̃, s + (n + 21 )c̃) ∩ [0, 1] and Ls is surjective.
Since [0, 1] is bounded we immediately get the result. 
If L is uniformly elliptic then for u ∈ W01,2 (Ω) we have
ˆ ˆ
λ u2 dx ≤ λkuk2W 1,2 ≤ BL (u, u) = f udx ≤ kf k · kuk2

so that the global W 2,2 -estimate implies

kukW 2,2 ≤ Ckf k2
where C depends on the ellipticity constants and the Lipschitz norm of the coeffi-
cients. Then
Lt = (1 − t)∆ + tL
2,2
would satisfy the W -bounds with the same constant. In particular, we would
know that L is surjective between W 2,2 (Ω) ∩ W01,2 (Ω) and L2 (Ω) if and only if the
Laplace operator ∆ is surjective between those spaces.
Note that such a result already follows from the general existence theory of
uniformly elliptic operators. In the following we want to obtain a similar existence
 LECTURE NOTES „LINEAR PARTIAL DIFFERENTIAL EQUATIONS“ 65

result for uniformly elliptic operators satisfying the Schauder estimates. Assume Ω
has C ∞ -boundary. Using the a priori estimate and the Schauder estimate we also
obtain for u ∈ C02,α (Ω)
kukC 2,α (Ω) ≤ CkLukC α (Ω)
which shows that L : C02,α (Ω) → C0α (Ω) satisfies the assumption of the Method of
Continuity. Note that the coefficients of L need only to be Hölder continuous so
that we cannot use the W 2,2 -estimate which need Lipschitz coefficients.
To see that ∆ is surjective let f ∈ C α (Ω) be given and fn → f in C0α (Ω) with
fn ∈ Cc∞ (Ω). The Sobolev theory gives a sequence un ∈ W01,2 (Ω) with ∆un = fn
and the regularity theory shows un ∈ C ∞ (Ω). Now the bound above shows
kun kC 2,α (Ω) ≤ Ckfn kC α (Ω)
which shows by Arzela–Ascoli that un0 → u in C 2 -norm with u ∈ C 2,α (Ω). In
particular, ∆u = f . By the Method of Continuity there is also a ũ ∈ C02,α (Ω) with
Lũ = f .

Appendix A. Topology
Recall that a topological space is a tuple (X, τ ) (e.g. Rn with its open sets) such
that τ ⊂ 2X , the set of open subsets, satisfies ∅, X ∈ τ , U ∩ V ∈ τ for all U, V ∈ τ
and whenever Ui ∈ τ for an index set i ∈ I then also ∪i∈I Ui ∈ τ . Any set C ⊂ X
such that X\C is open, will be called closed. Note that ∩i∈I C is closed whenever
each Ci , i ∈ I, is closed.
A sequence (xn )n∈N is said to converge to x, denoted by x = limn→∞ xn , if all
open neighborhood U of x, i.e. U ∈ τ with x ∈ U , there is an n0 ∈ N such that
xn ∈ U for all n ≥ n0 .
In general, convergence of sequences is not enough to describe the topology
completely. For this one needs the concepts of nets: A net (xi )i∈I in X is “subset”
{xi }i∈I of X that is indexed by a directed set (I, ≥). A directed set (I, ≥) is a
partially order set (≥ is reflexive and transitive) such that any two elements have
an upper bound, i.e. for each a, b ∈ I there is a c ∈ I such that c ≥ a and c ≥ b.
Now we say that the net (xi )i∈I converges to x, written x = limi∈I xi if for open
neighborhoods U of x there is an i0 ∈ I such that xi ∈ U for all i ≥ i0 .
If A ⊂ X is any subset then we define the interior int A of A and the closure
cl A of A as follows
[
int A = U = {x ∈ A | ∃U open : x ∈ U ⊂ A}
U ⊂A,U ∈τ
\
cl A = C = {x ∈ X | ∃net (xi )i∈I : x = lim xi }.
i∈I
A⊂C,X\C∈τ

In case of brevity we sometimes11 write cl A = Ā and int A = Å.

A subset A ⊂ B be dense in B if for each x ∈ B there is a net (xi )i∈I in A
converging to B. This is equivalent to saying that the closure of A and B in X is
the same. The topological space (X, τ ) is called separable if it admits a countable
subset D which is dense in X.
11Be aware that for metric spaces the closed ball B̄ (x) is not necessarily equivalent to the
r
closure cl(Br (x)), hence the notation Ā will be avoided if possible!
 LECTURE NOTES „LINEAR PARTIAL DIFFERENTIAL EQUATIONS“ 66

Also a set A ⊂ X is compact 12 if whenever A ⊂ ∪i∈I Ui for open sets {Ui }i∈I
then there is a finite subset I 0 ⊂ I such that A ⊂ ∪i∈I 0 Ui . One can show that any
compact set has to be closed. Using the concept of nets one can show that a set
A is compact if every net in A admits a convergent subnet with limit in A. Note
however that a set B where every sequences in B admits a convergent subsequence
with limit in A is, in general, not compact. Sets satisfying this condition will be
called sequentially compact.
Let Ω be an open set of a topological space X. Then for a subset A ⊂ X we say
A is compactly contained in Ω, written as A ⊂⊂ Ω, if
cl(A) ⊂ Ω
and cl A is compact.
A map f : X → Y between topological spaces (X, τ ) and (Y, τ 0 ) is said to be
continuous if for all V ∈ τ 0 it holds f −1 (V ) ∈ τ . The function f has compact
support in Ω if supp f ⊂⊂ Ω.
A metric 13 d on X is a symmetric, positive definite function on X × X that
satisfies the triangle inequality, i.e. for all x, y, z ∈ X it holds d(x, z) ≤ d(x, y) +
d(y, z).
Given a metric d on a set X there is a natural topology τd on X induced by d:
τd = {U ∈ 2x | ∀x ∈ U ∃r > 0 : Br (x) ⊂ U }.
We call the tuple (X, d) a metric space.
Note that if A ⊂⊂ Ω then the function
x 7→ d(x, ∂Ω) = inf{d(x, y) | y ∈ ∂Ω}
is uniformly bounded away from zero on the set A.
The convergence with respect to metric topology τd is equivalent to the following:
xi → x in τd for a net (xi )i∈I iff
∀ > 0∃i0 ∈ I∀i ≥ i0 : d(x, xi ) < .
This also show that the topology τd can descripted entirely by sequences instead of
nets.
The metric also allows one to define the concept of Cauchy sequence, i.e. (xn )n∈N
is Cauchy if
∀ > 0∃N ∈ N∀n, m ≥ N : d(xn , xm ) < .
Using the triangle inequality we can show that a converging sequence (xn )n∈N is
Cauchy. The converse is in general not true. A metric space (X, d) for which every
Cauchy sequence is convergent will be called complete. Hence in a complete metric
space the concept of Cauchy sequences and convergent sequences is equivalent.
If a metric space (X, d) is not complete then there is a (unique up to isomorphism)
completion (X̃, d)˜ such that (X, d) embeds isometrically in (X̃, d) ˜ such that the
image of X in X̃ is dense. For that reason we often regard X as a subset of the
completion X̃.

12In a complete metric space (e.g. (Rn , k · − · k)) this is equivalent to “every sequence in A
has a subsequence converging to a point in A”.
13e.g. on Rn take any norm k · k and define d(x, y) = kx − yk.
 LECTURE NOTES „LINEAR PARTIAL DIFFERENTIAL EQUATIONS“ 67

Appendix B. Measure and Integration Theory

A measurable space (X, B) is a set X equipped with a σ-algebra B ⊂ 2X , i.e. B
contains the empty set and is closed under countable unions and complements. If
(X, τ ) is a topological space then the Borel σ-algebra B(X) is the smallest σ-algebra
containing all open set U ∈ τ . A set in A ∈ B(X) will be call (Borel) measurable.
A map µ : B(X) → [0, ∞] is called a Borel measure if it is monotone, σ-additive
and µ(∅) = 0. The statement “a property P(x) holds for µ-almost all x ∈ X”
is a short description for the following: there is a measurable set Ω such that
µ(X\Ω) = 0 and property P(x) holds for all x ∈ Ω.
A function f : X → R is called measurable if for all Borel set A ∈ B(R) the set
f −1 (A) is measurable. A measurable function is simple if f is of the form
X
f= ai χAi
i∈N

for ai ∈ Rn and disjoint Borel set {Ai }i∈N .

For simple functions f : X → [0, ∞) define
ˆ X
f dµ := ai µ(Ai ).
i∈N

If f : X → [0, ∞) is measurable then define

ˆ ˆ
f dµ = sup{ f˜dµ | f˜ : X → [0, ∞) is a simple measurable function with f˜ ≤ f }.

It is know that there is a sequence fn : X → [0, ∞) of simple functions such that

fn ≤ fn+1 and for µ-almost all x ∈ X it holds f (x) = limn→∞ fn (x), i.e. fn
converges µ-almost every to f . In that case it can be shown
ˆ ˆ
f dµ = lim fn dµ.
n→∞

More generally, the Theorem of Monotone Convergence says that whenever fn :

X → [0, ∞) is a non-decreasing ´ sequence measurable
´ functions converging almost
everywhere to f then it holds f dµ = limn→∞ fn dµ. Note that f is measurable
as it is the µ-almost everywhere limit of measurable functions.
Fatou’s Lemma says that if for a sequence fn : X → [0, ∞) of measurable func-
tion it holds f (x) = lim inf n→∞ fn (x) for µ-almost all x ∈ X then f is measurable
and ˆ ˆ
f dµ ≤ lim inf fn dµ.
n→∞
´
We say a ´measurable function f : X → R is µ-integrable if f± dµ < ∞, or
1
equivalently |f |dµ <
´ ∞. Denote the space of 1µ-integrable functions by L (µ).
The assignment f 7→ |f |dµ induces a norm on L (µ) and makes it into a complete
Banach space. Furthermore, we define
ˆ ˆ ˆ
f dµ = f+ dµ − f− dµ.

The Theorem of Dominated Convergence says if a sequence of µ-integrable func-

tions fn : X → R converges µ-almost everywhere´ to f : X → R and
´ |fn | ≤ g for a
µ-integrable function then f is µ-integrable and f dµ = limn→∞ fn dµ.
 LECTURE NOTES „LINEAR PARTIAL DIFFERENTIAL EQUATIONS“ 68

On Rn equipped with a norm k · k there is a unique Borel measure λn , called

the Lebesgue measure (associated to (Rn , k · k)) that is translation invariant14 and
satisfies
n
n π2
λ (B1 (0)) = ωn =
.
Γ n2 + 1
Note that λn depends on the norm k·k. However, the Lebesgue measures associated
to two given norm different by a (multiplicative constant). If one chooses the
norm induced by the standard scalar product on Rn the measure λn is the “usual”
Lebesgue measure.
For each open subset Ω ⊂ Rn the Lebesgue measure induces a natural measure
λn Ω (usually also denoted by λn ) by restricting λn to subsets of Ω. We denote the
space of λn -integrable functions on Ω by L1 (Ω).
If Ω ⊂ Rn is open and bounded and ∂Ω is “sufficiently” then there is a natural
measure λn−1 on ∂Ω such that for all continuous functions f : Rn → R it holds
ˆ ˆ ˆ
f dz = f dλn−1 = lim f dλn .
→0 (∂Ω)
n−1
Again the measure λ depends on the norm (resp. metric) chosen on Rn .

Appendix C. Polar coordinates.

n
Let f : R → R an integrable function with compact support (e.g. f is a bounded
function) then
ˆ ˆ rˆ
f (x)dx = f (z)dzdr.
0 ∂Br (0)

If we observe that for each z ∈ ∂Br (0) there is a unique ω ∈ ∂B1 (0) = Sn−1 such
that z = rω and that
|∂Br (0)| = rn−1 |∂B1 (0)|
we immediately the formula for polar coordinates, i.e.
ˆ ˆ rˆ
f (x)dx = f (rω)rn−1 dωdr.
0 Sn−1

Furthermore, we can translate x 7→ x − a to also obtain the following

ˆ ˆ rˆ
f (x)dx = f (a + rω)rn−1 dωdr.
0 Sn−1

Appendix D. List of definition/symbols

For a topological space (X, τ ) and A ⊂ X:

[
int A = Å = U
A⊃U open
\
cl A = Ā = C.
A⊂C closed

14λn (A) = λn (A + x) for all measurable set A and x ∈ Rn

 LECTURE NOTES „LINEAR PARTIAL DIFFERENTIAL EQUATIONS“ 69

If (X, d) is a metric space and A ⊂ X then

Br (x) = {y ∈ X | d(x, y) < r}
B̄r (x) = {y ∈ X | d(x, y) < r}
[
A = ” − neighrborhood of A” = B (x).
x∈A
The function χA : X → R is defined as
(
0 x∈/A
χA (x) =
1 x ∈ A.
For functions f : X → R and measures µ (on a given σ-algebra):
ˆ ´
1 f dµ
f dµ = f dµ = ´A
A µ(A) A A
dµ
{f > c} = {x ∈ X | f (x) > c} for > ∈ {<, >, ≥, ≤, =, 6=}
supp f = cl{f (x) 6= 0}
f+ = χ{f >0} · f
f− = −χ{f <0} · f
f = f+ − f−
|f | = f+ + f−
ˆ
Lp (µ) = {f : X → R | |f |p dµ < ∞}

ess supµ f = sup{r ∈ R | µ({f > r}) > 0}

ess inf µ f = − ess sup f.
For functions f : Rn → R and the n-dimensional Lebesgue measure λn and
measurable sets A ⊂ R
ˆ ˆ
f dx = f dλn
A A
ess sup f = ess supλn f
ess supΩ f = ess sup f.
λn
Ω
( ´
p p n {f : Ω → R | Ω |f |p dx < ∞} p ∈ (0, ∞)
L (Ω) = L (λ Ω ) =
ess supΩ |f | < ∞ p=∞
If Ω ⊂ Rn is open and bounded with ∂Ω “nice” then
ˆ ˆ ˆ !
n−1 1
f dz = f dλ = ” lim f dx” .
∂Ω ∂Ω →0 2 (∂Ω)