Partial Differential Equations (Week 6+7)

Elliptic Regularity and Weak Solutions

Gustav Holzegel
March 16, 2019

1 Elliptic Regularity
We will now entertain the following questions. Say u is C 2 and satisfies

∆u = f (1)

in Ω. How regular is u depending on f ? What is the minimum regularity on f

so that we can actually find a C 2 solution of (1)? What about weak solutions?
Our formula (??) is already suggestive as far as the relation of the regularity of
f and the regularity of u is concerned. Let’s take for simplicity the case of the
ball for which we found a nice Green’s function for the Dirichlet problem:
Z
u (ξ) = G (x, ξ) f (x) dx (2)
Ω

Now since the Green’s function is symmetric in x and ξ what is suggested by

(2) is that we can gain (almost) two derivatives (explain!).1
We will now try to quantify this regularity gain but not at the level of the
Hoelder spaces C k,α (which would be an option leading to Schauder regularity
theory) but at the level of the Sobolev spaces H k .
Recall that H k (Ω) is defined as the space of L1loc (Ω) functions, which have
k weak derivatives and for which
X Z
2
kukH k (Ω) = |Dα u|2 dx < ∞ . (3)
|α|≤k Ω

We also define the homogeneous Sobolev “semi”-norm

X Z
2
kukḢ k (Ω) = |Dα u|2 dx , (4)
|α|=k Ω

1 There are counter examples of solutions of ∆u = f with with f continuous but u not C 2 .

The right spaces to quantify the vague statement made above are the Hoelder spaces C k,α
(which are Banach spaces), for which one can derive the estimates of the type kukC 2,α ≤
Ckf kC 0,α . from (2). Such an estimate is a regularity estimate. Search for Schauder estimates
in the literature.

1
which unlike (3) behaves well under scaling of the coordinates (in case that
Ω = Rn ). The spaces H k (Ω) are Hilbert spaces with the inner-product
X Z
< u, v >H k = Dα uDα v dx .
|α|≤k Ω

Remark 1.1. There are more general Sobolev spaces W k,p (Ω) (which become
the H k (Ω) for p = 2). Those are still Banach spaces. We won’t need them in
this course.

The associated notion of orthogonality for the H k and the fact that these
spaces are complete make them useful to do analysis. Moreover, these spaces
are actually suggested by the equation themselves, more precisely, the estimates
that one can derive from them.
Indeed, suppose for a moment that both f and u in (1) were smooth and of
compact support in Ω. Then the elementary computation
Z Z XZ XZ
f2 = |∆u|2 dx = ∂i2 u∂j2 u = (∂i ∂j u) (∂i ∂j u) dx = kuk2Ḣ 2 (Ω)
Ω Ω i,j Ω i,j Ω

with the boundary terms vanishing in the integration by parts by the assumption
of compact support, we would have

kuk2Ḣ 2 (Ω) = kf k2H 0 (Ω) := kf k2L2 (Ω) , (5)

expressing the “gain” of two-derivatives at the level of Sobolev spaces. Now

we can commute the Laplace equation by ∂xi for i = 1, ..., d and by the same
computation obtain

kuk2Ḣ 2+k (Ω) = kf k2H k (Ω) (6)

for all integers k > 0. I claim that these identities will be extremely useful in
helping us to prove

Proposition 1.2. Let Ω ⊂ Rn be open, u ∈ C 2 (Ω) and ∆u = f in Ω with

f ∈ C ∞ (Ω). Then u is also C ∞ in Ω.
Unfortunately we used the smoothness (and assumed compact support) of
u to derive the identity (6). The idea of the following proof is to turn things
around and to prove the smoothness of u using the identity (6).

Preliminaries
Let χ : Rd → R, d ≥ 2 be a smooth radial cut-off function which satisfies
• supp χ ⊂ B (0, 1)
R
• Rd χdx = 1

2
 • χ = 1 for |x| ≤ 12 .
Then weR know that χ (x) := −d χ x it supported in the ball B (0, ) and

satisfies Rd χ dx = 1
For a given x0 ∈ Ω ⊂ Rd we pick a ball around x0 which is compactly
contained in Ω. Let this ball have radius 2ρ. We define
 
x − x0
ũ = χ u
ρ

which localizes the u to the ball of radius ρ around x0 , and the associated

f˜ := ∆ũ = χf + 2∇χ∇u + ∆χ u . (7)

We note that f˜ is supported in B (x0 , ρ) and agrees with f in B x0 , ρ2 .

In addition, we define the sequence of mollifications

ũm = ũ ? χ mρ for any m > 1 (8)

Note that ũm is smooth and supported in B (x0 , 2ρ). Also, the mollifications

f˜m = f˜ ? χ mρ for any m > 1 (9)

are smooth and supported in B (x0 , 2ρ), hence in H k = H k (Rn ) for all k.

The proof
We first note that we have by definition for any  > 0
 
x−y
Z Z y
∆x ũm (x) = ∆ −d χ ũ (y) dy = ∆ −d χ ũ (x − y) dy .
Rd  Rd 

Since ũ is compactly supported, the Laplacian goes through the integral, which
– in view of its translation invariance – produces

∆ũm = f˜m (10)

1. Step 1. Since f ∈ C ∞ and u is in C 2 , the identity (7) gives that f˜ is C 1 .

Since f˜ is also of compact support in Ω, f˜ ∈ H 1 .
2. Step 2. We now recall an exercise about mollifiers from the previous
week, which implies that the mollifications f˜m defined in (9) converge to
f˜m → f˜ in H 1 as m → ∞. In particular, f˜m is Cauchy in H 1 . By the
same argument, we know that ũk → ũ in H 2 and hence ũk Cauchy in H 2 .
3. Step 3. Since ∆ (ũk − ũl ) = f˜k − f˜l and both differences are smooth and
of compact support, the identity (6) is now valid and takes the form

kũk − ũl k2Ḣ 2+k = kf˜k − f˜l k2Ḣ k . (11)

3
 Since we already know that ũk is Cauchy in H 2 we can apply the above
with k = 1 to conclude that ũk is actually Cauchy in H 3 . Since we already
2 3

is ũ ∈ H we infer ũ ∈ H . Noting that ũ agrees with u
know the limit
ρ
on B x0 , 2 and that we can perform the above steps around any point
x0 , we conclude u ∈ H 3 (no tilde!) on any compact subset of Ω.

4. Step 4. We now repeat the procedure. We pick x0 and the balls as above,
and revisiting (7), we see that now f˜ is actually H 2 . Now go to Step 2
replacing H s by H s+1 everywhere.
k
By induction, this shows that u ∈ Hloc (Ω) for all positive integers k. By
n
Sobolev embedding (see Exercises) we conclude that u ∈ C k− 2 −δ (Ω).

Reflecting on this proof, we see that it required the translation invariance of

the Laplacian (true for all constant coefficient linear operators) and the avail-
ability of the fundamental estimate for the H k -norm arising from the equation.
We already note at this stage that only the second ingredient is fundamental,
the first is technical.
Exercise 1.3. Below we will show existence of weak solutions of ∆u = f with
f ∈ L2 (Ω) and u ∈ H01 (Ω). It’s worth revisiting the above proof after that.

2 Existence of weak solutions for ∆u = f

2.1 The Energy Estimate


From Green’s identities we know that for u ∈ C 2 (Ω) ∩ C 1 Ω̄ we have the
identity Z Z
∇u · ∇u = − uf
Ω Ω
for solutions ∆u = f of Poisson’s equation which vanish on the boundary of Ω
(which we assume to be sufficiently regular): u = 0 on ∂Ω. Applying Cauchy-
Schwarz yields
Z Z Z
2 1
∇u · ∇u ≤  u + f2 (12)
Ω Ω  Ω

We have the following fundamental Poincare inequality

Proposition

2.1. Let Ω ⊂ Rn be bounded and open with ∂Ω smooth. Let
u ∈ C 1 Ω̄ vanish on ∂Ω. Then
Z Z
u2 dx ≤ CΩ |∇u|2 dx (13)
Ω Ω

holds for a constant CΩ depending only on Ω.

4
Proof. We first enclose Ω in a big cube Γ := {x ∈ Rn | |xi | ≤ a for i = 1, ..., n}.
Continue u to be zero in Γ outside Ω. Then, for any (x1 , ..., xn ) ∈ Γ we have
Z x 2 Z a
2 2
u (x) = ∂x1 u (ξ, x2 , ...xn ) dξ1 ≤ 2a dξ1 [∂x1 u (ξ1 , x2 , ...xn )] (14)
−a −a

Hence
Z a Z a
2
dx1 u2 (x1 , ..., xn ) ≤ 4a2 dξ1 [∂x1 u (ξ1 , x2 , ...xn )] (15)
−a −a

Integrating over x2 , ..., xn yields the desired inequality with CΩ = 4a2 .

Remark 2.2. The proof shows that Ω ⊂ Rn just needs to be bounded in one
direction and u ∈ H01 (Ω) (see below).
As a consequence of Poincare’s inequality, we obtain from (12) that
Z Z
|∇u|2 dx ≤ CΩ f 2 dx (16)
Ω Ω


holds for all u ∈ C 2 (Ω) ∩ C 1 Ω̄ which vanish on the boundary and satisfy
Poisson’s equation ∆u = f in Ω. In view of Poincare’s inequality,
Z
kuk2H 1 (Ω) ≤ CΩ f 2 dx (17)
Ω

2.2 Construction of Weak solutions

We now address the problem of constructing weak solutions to Dirichlet’s prob-
lem. We shall work with the Sobolev space H01 (Ω), which can be characterized
as the closure of the space of smooth functions of compact support in Ω with
respect to the H 1 (Ω) norm. By density, (13) holds for u ∈ H01 (Ω) and implies
that the H 1 (Ω) norm restricted to u ∈ H01 (Ω) is equivalent to the Ḣ 1 (Ω) norm.
It will be favorable to continue working with the Ḣ 1 (Ω)-inner product on
1
H0 (Ω) functions
Z
hu, vi = ∇u · ∇v dx (18)
Ω

We will call a u ∈ H01 (Ω) such that

hu, vi = − < f, v >L2 (19)
holds for all v ∈ H01
(Ω) a weak solution. Equivalently, we could formulate this
as u ∈ H01 (Ω) satisfying
< u, ∆v >L2 =< f, v >L2 (20)
for all test functions v ∈ C0∞ (Ω). (Recall that the latter are dense in H01 (Ω).
This is the distributional formulation we’ve seen earlier.
Remarkably, the existence of a weak solution of the Dirichlet problem follows
directly from the Riesz Representation theorem:

5
Theorem 2.3. Let (H, h·i) be a Hilbert space. Every bounded linear functional
φ on H can be represented uniquely in the form φ [u] = hu, vi for some v ∈ H.
Proof. The kernel of φ is a closed linear subspace of H. We have the decompo-
⊥
sition H = kerφ ⊕ (kerφ) .2 If φ (u) = 0 for all u, then v = 0 will work. If not,
⊥
there will be a w ∈ (kerφ) with kwk = 1 satisfying hw, ui = 0 for all u with
φ (u) = 0. It’s not hard to check that for any u the vector φ (u) w − φ (w) u is
in the kernel of φ. Hence hw, φ (u) w − φ (w) ui = 0 which yields

φ (w)w
φ (u) = hu, i = hu, φ (w)wi
< w, w >
The uniqueness is a simple exercise.
Theorem 2.4. Let Ω ⊂ Rn be open, bounded with smooth boundary. Let f ∈
L2 (Ω). Then there exists a unique u ∈ H01 (Ω) such that u is a weak solution of
∆u = f .
Proof. Note that φ [v] := −hf, viL2 defines a bounded linear functional on
H01 (Ω) equipped with inner product (18). (Why?) Then apply Riesz-Theorem
to conclude that (19) holds.
This proof almost seems too easy and it’s good to recall what goes into the
Riesz Representation theorem. The variational character can also be understood
from the following proposition which is known as Dirichlet’s principle


Proposition 2.5. Suppose that u ∈ C 1 Ω̄ , u = 0 on ∂Ω and f ∈ L2 (Ω).
Then the following are equivalent
1. ∆u = f in D0 (Ω).


2. J [u] ≤ J [w] for all w ∈ C 1 Ω with w = 0 on ∂Ω for J being the
functional Z
|∇w (x) |2 + 2w (x) f (x) dx
 
J [w] =
Ω

3. u is a critical point of J in the sense that

d 
= 0 holds for all φ ∈ C 1 Ω̄ with φ = 0 on ∂Ω


J [u + tφ] 

dt t=0

Proof. Exercise.
The proposition suggests to find a distributional solution of ∆u = f via
proving the existence of a minimizer of a variational problem. This is a very
important technique, as many PDEs can be shown to admit a variational for-
mulation.
2 This uses the variational principle: Given a closed subspace B ⊂ H and a v ∈ H, there is

an element ṽ in B which is closest to v in the sense that kx − ṽk = inf x∈B kx − vk for x ∈ B.
Moreover, this v − ṽ is orthogonal to B in that hv − ṽ, xi holds for any x ∈ B. Once this is
established (Exercise. Hint: Parallelogram identity!), write any x ∈ H as x = x0 + (x − x0 )
with x0 ∈ B and, by the previous result, x − x0 in B ⊥ .

6
2.3 More Regularity
With Theorem 2.4 you should now revisit the proof of Proposition 1.2 (cf. Ex-
ercise 1.3) and try to prove that the weak solution is more regular in Ω (in fact,
classical), given more regularity for f .

3 General Second Order Elliptic Equations

How much does all this depend on the fact that we were dealing with Laplace’s
equation? As we will mentioned, the crucial ingredients were the a-priori es-
timates that we derived on the solution. For this section I will follow closely
Evans, Chapter 6.2.
We consider the following general linear operator
n
X n
X
aij (x)uxi bi (x) uxi + c (x) u


Lu = − xj
+ (21)
i,j i=1

written in divergence form, with suitable regularity assumptions on the coeffi-

cients (made precise below) and aij = aji symmetric. We consider the Dirichlet
problem
Lu = f in U
(22)
u=0 on ∂U
Definition 3.1. L is called uniformly elliptic if there exists a constant θ > 0
such that
Xn
aij (x) ξi ξj ≥ θ|ξ|2 (23)
i,k=1

for almost every x ∈ U and all ξ ∈ Rn .

In other words, the symmetric matrix aij is uniformly positive definite at
(almost) every point of U with the smallest eigenvalue being larger or equal to
θ.
The next two definitions are obvious generalizations of our treatment of the
Laplace equation:
Definition 3.2. The bilinear form B [·] associated with L is defined to be
 
Z Xn n
X
B [u, v] :=  aij uxi vxj + bi uxi v + cu · v  dx (24)
U i,j i=1

for u, v ∈ H01 (U ).
Definition 3.3. We say that u ∈ H01 (U ) is a weak solution of the boundary
value problem (22) if
B [u, v] = (f, v)L2 (U ) (25)
holds for all v ∈ H01 (U ).

7
3.1 Existence
Let (H, < · >) be a Hilbert space. We need a generalization of the Riesz repre-
sentation theorem for the general bilinear form (24).
Theorem 3.4. Assume that

B :H ×H →R

is a bilinear map satisfying the following two conditions:

|B [u, v] | ≤ αkukkvk for some α > 0 and all u, v ∈ H (boundedness)

βkuk2 ≤ B [u, u] for some β > 0 and all u ∈ H (coercivity)

Assume further that f : H → R is a bounded linear function on H. Then there

exists a unique element u ∈ H such that

B [u, v] = f [v] holds for all v ∈ H

This is the famous Lax-Milgram theorem. Note that if B was in addition

symmetric, the Riesz representation theorem would provide again a two line
proof. The general case is more difficult and the content of
Exercise 3.5. Prove this theorem using the following outline (or otherwise).
For fixed u, apply Riesz to v → B [u, v], which is a bounded linear functional.
Get B [u, v] =< Au, v > where A maps u to the element promised by Riesz.
Show that A is a bounded linear operator, which is one-to-one and has range all
of H. Then apply Riesz again.
With this at hand, all we need to do to conclude the existence of a weak
solution of (22), is to check whether the linear elliptic operator L (or rather its
associated bilinear form (24)) satisfies the two conditions of Theorem 3.4. Let
us assume that the coefficients a, b, c of L are in L∞ (U ).
The first (boundedness) condition is very easily checked (Exercise). For the
second, we start with the ellipticity condition
Z Z X n Z "X n
#
2 ij i 2
θ |Du| dx ≤ a uxi uxj = B [u, u] − b uxi u + cu dx . (26)
U U i,j=1 U i=1

Using now that b is in L∞ and applying Cauchy’s inequality with  to the mixed
term we easily obtain
Z Z
θ 2
|Du| dx ≤ B [u, u] + C u2 dx (27)
2 U U

for a constant C. By the Poincare inequality for u ∈ H01 , the left hand side also
controls the L2 norm of u and we are done, obtaining

βkuk2H 1 (U ) ≤ B [u, u] + γkuk2L2 (U ) (28)

0

8
for a (possibly small) constant β and a (possibly large) constant γ, which can be
computed from the L∞ bounds on the coefficients (in particular, the ellipticity
constant θ) and the geometry of U only. The estimate (28) is not quite the
desired coercivity condition unless γ = 0.
Exercise 3.6. State conditions on the coefficients a, b, c which would allow one
to derive the estimate (28) with γ = 0.
The following example shows that γ > 0 can be a true obstruction to finding
a unique weak solution:
Example 3.7. Let L = −∂x2 − ∂y2 − 2π 2 and U the unit square. Then u = 0
and u (x, y) = sin (πx) sin (πy) both solve Lu = 0.
We will understand this better in due course. In any case, if we add a large
enough zeroth order term to L we always have:
Theorem 3.8. There is a number γ ≥ 0 such that for each µ ≥ γ and each
f ∈ L2 (U ) the following statement holds. There exists a unique weak solution
u ∈ H01 (U ) of the boundary value problem
Lu + µu = f in U
(29)
u=0 on ∂U
Proof. Observe that Bµ [u, v] = B [u, v] + µ (u, v)L2 (U ) satisfies the assumptions
of the Lax-Milgram theorem.

3.2 Interior Regularity: From H 1 to Hloc

2

2
We begin by showing that a weak solution in U in actually in Hloc . As we will
derive regularity statements only in the interior, it will suffice to only assume
u ∈ H 1 (and not H01 ). Note that for the Laplacian you could repeat the proof
of Proposition 1.2 to arrive at the same conclusion.
Theorem 3.9. Assume aij ∈ C 1 (U ), bi , c ∈ L∞ (U ), f ∈ L2 (U ). Suppose u ∈
H 1 (U ) is a weak solution of the elliptic PDE Lu = f in U . Then u ∈ Hloc
2
(U )
and for each open V ⊂⊂ U we have the estimate


kukH 2 (V ) ≤ C kf kL2 (U ) + kukL2 (U )
for a constant depending only on V, U and the coefficients a, b, c of L.
For the proof we will need the notion of a finite difference quotient. For
u ∈ L1loc (U ) and V ⊂⊂ U we define
u (x + hek ) − u (x)
Dkh u (x) = (30)
h
for x ∈ V ⊂⊂ U and h a real number satisfying 0 < |h| < dist (V, ∂U ) and ek
the k th unit vector in Rn . The Dkh u is called the k th difference quotient of size
h. We also define
Dh u := D1h u, ..., Dnh u


(31)

9
Exercise 3.10. Verify the following properties of difference quotients. Let
u, v, w ∈ H01 (U ) be compactly supported in V ⊂⊂ U . Then
1. Dkh u is in H01 (U ) for sufficiently small h (regularity)
2. U Dkh v · w = − U vDk−h w for h sufficiently small
R R

(integration by parts formula)

3. Dkh (vw) = v h Dkh w + Dkh v w with v h = v (x + hek ) (product rule)
More difficult are the following crucial properties
Exercise 3.11. If u ∈ H 1 (U ), then
1
kDh ukL2 (V ) ≤ CkDukL2 (U ) holds for all 0 < |h| < dist (V, ∂U ) (32)
2
(which you can prove for smooth u and then argue by density). Very important
is the converse property: If
1
kDh ukL2 (V ) ≤ C for all 0 < |h| < dist (V, ∂U ) (33)
2
then u ∈ H 1 (V ) with kDukL2 (V ) ≤ C. Hence if we can bound the difference
quotients uniformly in h, then we can conclude the existence of a weak derivative!
Proof of Theorem 3.9. Fix a V ⊂⊂ U open and choose a W open with V ⊂⊂
W ⊂⊂ U . Select a smooth cut-off function ζ which is equal to 1 in V and equal
to zero in Rn \ W . Since u is a weak solution, we have
n Z Z !
X X
ij i
A := a uxi vxj dx = f− b uxi − cu v dx =: B (34)
i,j=1 U U i

If we knew u was smooth we would set, v = −ζ 2 ukk and integrate by parts.

However, at this point we only know u ∈ H 1 (U )! Therefore we set
v = −Dk−h ζ 2 Dkh u


(35)
and use the integration by parts formula for difference quotients. The goal will
be to bound the difference quotient kDkh Duk in L2 on V and then to apply the
second property of Exercise 3.11, which will guarantee that u ∈ H 2 (V ).
Noting that difference quotients commute with weak derivatives, we have
Z
aij uxi Dk−h ζ 2 Dkh u x dx


A =−
j
Z U
Dkh aij uxi 2ζζxj Dkh u + ζ 2 Dkh uxj dx

 
=
ZU
aij,h Dkh uxi + Dkh aij uxi 2ζζxj Dkh u + ζ 2 Dkh uxj dx


 
=
ZU
aij,h Dkh uxi Dkh uxj ζ 2 dx + error-terms


= (36)
U

10
Therefore,
Z
A ≥θ ζ 2 |Dkh Du|2 dx − |error-terms| (37)
U

Observe now that (note this uses aij ∈ C 1 (U ) – where?)

Z h i
|error-terms| ≤ C |Dkh Du||Dkh u|ζ|Dζ| + |Du||Dkh Du|ζ 2 + |Du||Dhk u|ζ|Dζ| dx
ZU Z
θ
≤ ζ 2 |Dkh Du|2 dx + C |Du|2 dx (38)
2 U U

Combining (37) and (38) and using that A = B we have

Z Z
θ
ζ 2 |Dkh Du|2 dx ≤ | B | + C |Du|2 dx (39)
2 U U

Finally,
C2
Z Z Z
2
|B |≤C (|f | + |Du| + |u|) |v|dx ≤ (|f | + |Du| + |u|) dx +  |v|2 dx
U  U U

where v is as defined in (35). We easily compute

Z Z Z
|v|2 ≤ C |D ζ 2 Dkh u |2 dx ≤ C ζ 2 |Dkh Du|2 + |Dkh u|2 dx



U U W
Z
 2 h
ζ |Dk Du|2 + |Du|2 dx

≤C (40)
U

and obtain, choosing the  above appropriately,

Z Z
2 θ
|B |≤C (|f | + |Du| + |u|) dx + |v|2 dx (41)
U 4 U
Combining (39) with (41) we finally obtain
Z Z Z
|Dkh Du|2 dx ≤ ζ 2 |Dkh Du|2 dx ≤ C |f |2 + |Du|2 + |u|2 dx


(42)
V U U

for h sufficiently small and k = 1, 2, ..., n.

By the properties of the difference quotients and the fact that u ∈ H 1 (U )
and f ∈ L2 (U ) we already conclude u ∈ Hloc 2
(U ). We can refine the above
estimate further. To do this, we observe that the same argument as above
shows that for any V ⊂⊂ W ⊂⊂ U we have


kukH 2 (V ) ≤ C kf kL2 (W ) + kukH 1 (W ) . (43)
Now choose a new cut-off function χ which is equal to 1 in W and equal to zero
in a neighborhood of the boundary of U . Setting v = ζ 2 u in the formula for the
generalized solution produces (Exercise)
Z  
ζ 2 |Du|2 dx ≤ C kf k2L2 (U ) + kuk2L2 (U ) . (44)
U

Combining this with (43) we obtain the estimate claimed in the Theorem.

11
 It is clear that this argument can be iterated in the sense that if f ∈ H k (U )
and appropriate assumptions hold on the coefficients, then every weak solution
k+2
u ∈ H01 (U ) is actually in u ∈ Hloc (U ). You may want to formulate this
statement precisely and do a proof by induction or at least look it up in the
literature, e.g. in the book of Evans.

3.3 Boundary Regularity: From H 1 to H 2

In the previous section we proved (in particular) that if u ∈ H01 (U ) is a weak
2
solution of Lu = f in U , then u is in Hloc (U ) under very mild conditions on the
coefficients. Note that this already implies that the equation Lu = f is satisfied
almost everywhere (Exercise) or, if f is sufficiently regular, that u is a classical
solution in the interior.
We now turn to the issue of whether a weak solution u ∈ H01 (U ) remains
smooth up to the boundary. For the following theorem we hence assume that
the boundary of U is smooth.
Theorem 3.12. Assume aij ∈ C 1 Ū , bi , c ∈ L∞ (U ), f ∈ L2 (U ). Suppose

u ∈ H01 (U ) is a weak solution of the elliptic PDE Lu = f in U , u = 0 on ∂U .

Then u ∈ H 2 (U ) with the estimate


kukH 2 (U ) ≤ C kf kL2 (U ) + kukL2 (U )
for a constant depending only on U and the coefficients a, b, c of L.
The proof has two steps, of which we are only going to sketch the first.
Namely, we first prove the statement for U being a half-ball in Rn , i.e. U =
B (0, 1)∩{xn > 0} with straight boundary {xn = 0}∩B (0, 1). The second step is
to construct – for arbitrary U – a diffeomorphism which maps the neighborhood
of a point on ∂U to the half-ball and to carry the estimates over. (This step
requires certain regularity of the boundary.)
To carry out the first step, as mentioned we let U = B (0, 1) ∩ Rn+ . We
set V := B (0, 1/2) ∩ Rn+ and choose a cut-off function which is equal to 1 on
B (0, 1/2) and vanishes identically on Rn \ B (0, 1). In particular, it vanishes on
the curved part of ∂U .
The idea is to repeat the “interior” computation with
v = −Dk−h ζ 2 Dkh u

where k = 1, ....n − 1 involves only difference quotients in the tangential di-

rections. This v is well-defined and actually in H01 (U ) (why?). Repeating the
“interior” computation we obtain the analogue of (43)
n
X

kuxk xl kL2 (V ) ≤ C kf kL2 (U ) + kukH 1 (U ) (45)
k,l=1
k+l<2n

We are cheating here slightly. The reason is that what we proved for finite
difference quotients (the second property of Exercise 3.11) is slightly weaker
than what is need here.

12
Exercise 3.13. State and prove the property of finite differences needed to
obtain the estimate (45).
In view of (45), the only thing missing is an H 2 estimate for the second
derivative uxn xn . For this we recall that we can write the equation Lu = f
(satisfied almost everywhere in U !) as
 
X n
X Xn
− aij uxi xj + bi − aij
xj
 uxi + cu = f (46)
i,j=1 i=1 j=1

and moreover, as
 
n
X n
X n
X
ann uxn xn = akl uxk xl + bi − aij
xj
 uxi + cu − f (47)
k,l=1 i=1 j=1
k+l<2n

The uniform ellipticity condition yields ann ≥ θ > 0 and therefore

 
Xn
|uxn xn | ≤ C  |uxk xl | + |Du| + |u| + |f | (48)
k,l=1,k+l<2n

The right hand side is in L2 (V ) and hence finally

kukH 2 (V ) ≤ C kf kL2 (U ) + kukH 1 (U ) ≤ C kf kL2 (U ) + kukL2 (U ) (49)

with the last inequality following from the basic energy estimate
h i
kukH01 (U ) ≤ C kf k2L2 (U ) + kuk2L2 (U )

derived earlier for u ∈ H01 (U ).

Once the estimate is established for the half ball one constructs an appro-
priate diffeomorphism, i.e. introduces coordinates near a point on the boundary
in which the boundary locally corresponds to xn = 0. It is easy to see that the
crucial ellipticity condition is preserved.
Again, this method can be iterated and used to establish that u is actually
smooth up to the boundary provided f and all the coefficients are smooth in Ū .
We refer again to the book of Evans.

3.4 The Fredholm Alternative

Let us understand better the problem of non-uniqueness (and non-existence) of
weak solutions for elliptic equation which was touched upon in Example 3.7.
For this we need a brief digression reminding you of the properties of compact
operators.

13
3.4.1 Compact Operators
Definition 3.14. A bounded linear operator

K:X→Y

between two real Banach spaces X and Y is called compact provided for each
bounded sequence {uk }∞ ∞
k=1 ⊂ X, the sequence {Kuk }k=1 is precompact in Y ,
i.e. there exists a subsequence {ukj }j=1 such that {Kukj }∞
∞
j=1 converges in Y .

Exercise 3.15. Find examples and non-examples. Prove that if K is compact,

then it is bounded (hence continuous). Prove that K is compact if and only if
the adjoint K ? is compact.
We will mostly deal with compact operators between Hilbert spaces. That
is a good occasion to do
Exercise 3.16. Remind yourself of the spectral theorem for symmetric (self-
adjoint) compact operators K : H → H in a Hilbert space.
The property of compact operators most relevant to us goes by the name of
the Fredholm alternative and concerns compact perturbations of the identity.
Theorem 3.17. Let K : H → H be a compact linear operator. Then
1. N (I − K) is finite dimensional

2. R (I − K) is closed
⊥
3. R (I − K) = N (I − K ? )
4. N (I − K) = {0} iff R (I − K) = H
5. dim N (I − K) = dim N (I − K ? )

Proof. See appendix of Evans.

Corollary 3.18. Precisely one of the following statements holds:
1. for each f ∈ H the equation u − Ku = f has a unique solution
2. the homogeneous equation u − Ku = 0 has non-trivial solutions.

Moreover, in the case of 2. then, the space of solutions of the homogenous

equation is finite dimensional and u − Ku = f has a solution if and only if
⊥
f ∈ N (I − K ? ) .

14
3.4.2 Application to Lu = f
Let us connect the theory of compact operators with our problem Lu = f and
assume for convenience that the coefficients a, b, c of L are actually in C ∞ Ū .

Definition 3.19. The formal adjoint (or transpose) associated with L is the
operator L? defined by
!
X X X
? ij i i


L v := − a vx j x i − b vx i + c − bxi v .
i,j i i

The adjoint bilinear form associated with L is defined to be

B ? : H01 (U ) × H01 (U ) → R B ? [v, u] = B [u, v] .

Finally, we call v ∈ H01 (U ) a solution of the adjoint problem
L? u = f in U , v = 0 on ∂U
provided that B ? [v, u] = (f, u)L2 (U ) holds for all u ∈ H01 (U ).
Theorem 3.20. Precisely one of the following statements holds:
1. For each f ∈ L2 (U ) there exists a unique weak solution u of the inhomo-
geneous problem 
Lu = f in U
(IP ) (50)
u = 0 on ∂U
2. There exists a weak solution u 6= 0 of the homogeneous problem

Lu = 0 in U
(HP ) (51)
u = 0 on ∂U

If the second alternative holds, then the dimension of the subspace N ⊂

H01 (U ) of weak solutions to (HP ) is finite and equals the dimension of the
subspace N ? ⊂ H01 (U ) of weak solutions of the homogeneous adjoint problem
 ?
L v = 0 in U
(AP ) (52)
v=0 on ∂U
Finally, the problem (IP ) has a weak solution if and only if (f, v) = 0 holds
for all v ∈ N ? .
Proof. Step 1: Relating the problem to Fredholm theory. We let
Bγ [u, v] = B [u, v] + γ (u, v)L2
with γ defined such that the associated operator Lγ u = Lu+γu has the property
that Lγ u = g has a unique weak solution u ∈ H01 (U ) for each g ∈ L2 (U )
(cf. Theorem 3.8), i.e.3
Bγ [u, v] = (g, v) for all v ∈ H01 (U ) (53)
3 We can assume γ > 0 wlog since for γ = 0 we already proved that the first statement and

not the second applies.

15
We write u = L−1 2
γ g to denote the operator that maps a given g ∈ L (U ) to the
1
unique u ∈ H0 (U ) satisfying (53).
By definition, u ∈ H01 (U ) is a weak solution of (IP ) if and only if

Bγ [u, v] = (γu + f, v) for all v ∈ H01 (U ) (54)

and hence if and only if u = L−1

γ (γu + f ) holds.
We define Kw = γL−1 γ w for w ∈ L2 (U ). The point of this is that we can
write
(I − K) u = h := L−1
γ f

which is of the form in the Fredholm alternative, if we can show that K is a

linear, bounded, compact operator. Indeed, we have K : L2 (U ) → H01 (U ) ⊂⊂
L2 (U ). The first part is bounded since in view of

βkuk2H 1 (U ) ≤ Bγ [u, u] = (g, u) ≤ kgkL2 (U ) kukL2 (U ) ≤ kgkL2 (U ) kukH01 (U )

0

we have
kukH01 (U ) = kKgkH01 (U ) ≤ kgkL2 (U ) .
The second part (the inclusion into L2 (U )) is compact by Rellich’s theorem.4

Step 2: Applying the Fredholm alternative. By the latter precisely one of

the following statements hold:
1. for each h ∈ L2 (U ) the equation u − Ku = h has a unique solution
u ∈ L2 (U ).
2. the equation u − Ku = 0 has non-trivial solutions in L2 (U ).
If the first statement applies then – by the equivalence (54) – there is a unique
solution to (IP ). If the second statement applies then necessarily γ 6= 0 (why?)
and the space of solutions to u − Ku is finite and equals the dimension of the
space N ? of solutions to v −K ? v = 0. It is easy to see the following equivalences

u − Ku = 0 if and only if u ∈ H01 (U ) is a weak solution of (HP ).

and

v − K ? v = 0 if and only if v ∈ H01 (U ) is a weak solution of (AP ).

To prove the last claim of the theorem, observe that by the Fredholm al-
ternative u − Ku = h has a solution if and only if (h, v) = 0 holds for all
v ∈ N (I − K ? ), i.e. for all v satisfying v − K ? v = 0. Finally,

1 1 1
(h, v) = L−1
γ f, v = (Kf, v) = (f, K ? v) = (f, v)
γ γ γ

4 Recall that X ⊂⊂ Y means that kxk ≤ Ckxk

Y (continuity) holds and that each bounded
sequence in X is precompact in Y (compactness). See the exercises.

16
 We conclude with the following
Theorem 3.21. There exists an at most countable set Σ ⊂ R such that the
boundary value problem

Lu = λu + f in U
(55)
u=0 on ∂U

has a unique solution for each f ∈ L2 (U ) if and only if λ 6= Σ. If Σ is infinite,

∞
then Σ = (λk )k=1 form a non-decreasing sequence with λk → ∞.
The set Σ is called the spectrum of the operator L.
Corollary 3.22. 
Lu = λu in U
(56)
u=0 on ∂U
has a non-trivial solution w 6= 0 if and only if λ ∈ Σ. In this case λ is an
eigenvalue of L with associated eigenfunction w.
Proof of Theorem 3.21. Recall the constant γ ≥ 0 which is always such that
Lγ = L + γ is invertible. Wlog we can assume γ > 0. We can also assume that
λ > −γ since if λ + γ ≤ 0, the homogenous problem Lu + γu = (λ + γ) u the
homogeneous problem is easily seen not to admit non-trivial solutions.
We then have the following equivalent statements:
• Problem (55) has a unique weak solution for a given λ
• The associated homogeneous problem has only the trivial solution.
• Lu + γu = (γ + λ) u has only the trivial solution.
γ+λ
• u = L−1
γ (γ + λ) u = γ Ku has only the trivial solution
γ
• γ+λ is not an eigenvalue of K
As the collection of eigenvalues of a compact operator K is either finite or forms
a sequence converging to zero (see exercises below) we obtain that (55) has a
unique solution unless λ is the member of a sequence λk → ∞.

4 Exercises
1. Do the exercises in the text if you haven’t already done so. Exercises 3.5
and 3.11 are particularly recommended.
Pn
2. Let u be a smooth solution of Lu = − i,j=1 aij uxi xj = 0 in U , where L


is uniformly elliptic and the coefficients aij are in C 1 U .
Set v := |Du|2 + λu2 . Show that

Lv ≤ 0 holds in U for λ sufficiently large

17
 Deduce

kDukL∞ (U ) ≤ C kDukL∞ (∂U ) + kukL∞ (∂U ) .
Remark: The “deduce” step requires the maximum principle for general
elliptic operators which we haven’t shown in general, except in the 2-
dimensional case (Exercise 5 of Week 5). You can assume it here or con-
sider the proof another exercise. [Hint: For the maximum principle all
consideration are local. Suppose you have an interior maximum at x̃ ∈ U .
You want to change coordinates at x̃ by a rotation such that diagonalises
aij . For this you first need to understand howPthe PDE changes under lin-
ear coordinate transformations: yi = (x̃)i + j Rij (x − x̃)j for a matrix
Rij which diagonalises aij at x̃.]
3. (Best constant in Poincare’s inequality;

F. John Chapter 5) Show that if
there exists a function u ∈ C 2 Ω vanishing for ∂Ω for which the quotient

|∇u|2
R
ΩR

Ω
u2

reaches its smallest value λ, then u is an eigenfunction to the eigenvalue λ,

i.e. ∆u + λu = 0 in Ω. In fact
λ must be the smallest eigenvalue belonging
to an eigenfunction in C 2 Ω .
4. Let n = 3 and Ω be the ball |x| < π. Show that a solution u of ∆u + u =
w (x) with vanishing boundary values can only exist if

sin |x|
Z
w (x) dx = 0
Ω |x|

5. Dirichlet boundary condition for weak solutions. Suppose we are given

a weak solution u ∈ H01 (Ω) of the second order elliptic boundary value
problem Lu = f in Ω, u = 0 on ∂Ω where Ω is a bounded open domain
in Rn with smooth boundary. It is natural to ask in what sense is “u = 0
on ∂Ω”? Remember that functions in H01 (Ω) are only defined a.e. so it
does not make sense to talk about the value of the function on the bound-
ary unless the solution is more regular. The correct way to do this is via
traces of functions in Sobolev spaces. Without introducing this technical
machinery here, the following exercise describes a satisfactory notion of
the solution vanishing on the boundary.

Let Ω be as above and Ω = {x ∈ Ω | dist (x, ∂Ω) < }. Then there is a
constant C and an 0 > 0 > 0 such that for all u ∈ H01 (Ω) and  < 0 we
have the estimate Z Z
2 2
|u| dx ≤ C |∇u|2
Ω Ω

Hint: Prove this estimate for functions in C0∞ (Ω) and argue by density.
To do the former, construct suitable coordinates near the boundary and

18
 mimic the proof of the Poincare inequality.

Deduce that Z
1
|u|2 dx = o()
vol(Ω ) Ω
holds as  → 0+ and interpret your result.
6. Prove the following basic version of the Banach Alaoglu theorem (which
we used in connection with the difference quotients): Let (uk ) be a bounded
sequence in a separable Hilbert space H, i.e. kuk kH ≤ C. Then there ex-
ists a subsequence which converges weakly in H. Hint: Use the following
outline
(a) Pick an ONB (ek ) and use a diagonal argument to show that for
(n)
a subsequence of the (uk ), denoted (un ) (arising from a Cantor
diagonal argument) we have that
hun(n) , ek i → vk ∈ R
holds for all ek .
P∞ 2
P
(b) Show that k=1 |vk | < ∞ and hence v = k vk ek ∈ H.
(n)
(c) Show that un * v.
7. (Rellich’s theorem, simple version). Let (uk ) be a bounded sequence in
H01 (Ω) with Ω as in lectures. Then there exists a subsequence converging
strongly in L2 (Ω).
Outline: Extend each uk by zero outside Ω to get a sequence (uk ) in
H 1 (Rn ). By Banach-Alaoglu, there exists a subsequence with unk * u.
For fixed ξ the map Fξ : H01 (Ω) → R mapping v ∈ H01 (Ω) to v̂ (ξ) (Fourier
transform evaluated at ξ) is a bounded linear functional on H01 (Ω). Now
use the Fourier definition of L2 (Ω) and H 1 (Ω) to conclude.
8. Sobolev inequality on bounded domains [F. John, Chapter 5]
Definition: A conical sector Γ ⊂ Rn is the intersection of a ball with a
cone from its centre:
Γ = {y | y = x + tξ ; 0 ≤ t ≤ h ; ξ ∈ σ}
where σ is a relatively open subset of the unit sphere in Rn . We call x
the vertex and h the radius of Γ. The solid angle ω of Γ is the (n − 1)-
dimensional measure of σ. An open set Ω ⊂ Rn has the cone property
if there exist positive numbers h, ω such that each x ∈ Ω is vertex of a
conical sector Γ ⊂ Ω of radius h and solid angle ω.

Show that for any Ω ⊂ Rn with the cone property there exists a C (de-
pending on Ω) such that for any u ∈ C s (Ω) with s = n2 + 1 if n is even
and s = n2 + 12 if n is odd, and any x ∈ Ω we have
|u(x)| ≤ CkukH s (Ω) . (57)

19
 Outline: Show that for φ ∈ C s (R) with φ(t) = 0 for t > h we have
Z h
(−1)s
φ(0) = ts−1 φ(s) (t)dt .
(s − 1)! 0
Apply Cauchy-Schwarz to get
Z h
φ2 (0) = C tn−1 (φ(s) (t))2 dt . (58)
0
Finally, let ζ(t) ∈ C ∞ (R) with ζ(0) = 0 and ζ(t) = 0 for t > h and
consider
φ(t) = ζ(t)u(x + tξ) .
Integrate (58) over the set σ.
9. (Spectral properties of compact operators on Hilbert spaces) Let K : H →
H be a compact operator with H a separable Hilbert space. We say that
λ is an eigenvalue of K if Ku = λu holds for some u 6= 0. Show that
K can have at most countably many eigenvalues and that they can only
accumulate at 0.

Outline: Show that for any k > 0 the number of distinct eigenvalues
with |λ| ≥ k is finite. Assuming this is false there is a sequence (λn )
of distinct eigenvalues and a corresponding sequence of eigenvectors (vn ).
Let Yn := span(v1 , ..., vn ). Choose a sequence (yn ) with kyn k = 1, yn ∈ Yn
⊥
and yn ∈ Yn−1 . Show that kKym − Kyn k ≥ k for all m > n and conclude.

5 Non-examinable exercises and further reading

1. Prove Theorem 3.17. (See the book of Evans or my notes on Functional
Analysis.)
2. (See the book of Evans.) Formulate and prove a statement generalizing
the maximum principle we discussed for the Laplacian to general elliptic
operators
Lu = −aij uxi xj + bi uxi + cu
For the strong maximum principle, you will have to prove the following
Lemma, due to E. Hopf:


Lemma 5.1. Let u ∈ C 2 (U ) ∩ C 1 Ū , c ≡ 0 in U and Lu ≤ 0 for L as
above. Suppose there exists a point x0 ∈ ∂U such that u (x0 ) > u (x) for
all x ∈ U . Assume that U satisfies the interior ball condition at x0 (there
is an open ball B ⊂ U with x0 ∈ ∂B; this follows for instance if ∂U is
C 2 ). Then the strict inequality
∂u 0

x >0
∂ν
holds, where ν is the outer unit normal to B at x0 .

20