Elliptic Partial Differential Equations

Leonardo Abbrescia
Advised by Daniela De Silva, PhD and Ovidiu Savin, PhD.
Contents
1 Introduction and Acknowledgments 2

2 Variational Methods and Sobolev Embedding Theorems 3

2.1 Laplacian . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.2 A non-linear PDE and Method of Continuity . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.2.1 Non-linear variations on Manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.2.2 Regularity Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

3 Harnack Inequality for Divergence Equations 23

3.1 Regularity Estimates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.2 Harnack Inequality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.2.1 Structural Inequalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.2.2 Moser Iteration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.2.3 Applications of the Harnack Inequality . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

4 Harnack Inequality for Non-Divergence Equations 38

4.1 ABP Estimate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
4.2 Measure theory and Harnack Inequality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40

5 Curved C 1,α Domains 44

5.1 Estimate for Laplacian . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
5.2 General Elliptic Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49

1
1 Introduction and Acknowledgments
Second order elliptic partial differential equations are fundamentally modeled by Laplace’s equation ∆u = 0.
This thesis begins with trying to prove existence of a solution u that solves ∆u = f using variational methods.
In doing so, we introduce the theory of Sobolev spaces and their embeddings into Lp and C k,α . We then
move on to applying our techniques to a non-linear elliptic equation on a compact Riemannian manifold. We
introduce the method of continuity along the way to provide another way of solving the equation. We move
onto proving Schauder estimates for general elliptic equations in divergence form: ∂i (aij ∂j u) + c(x)u = f
with various assumptions on a, c, and f . We conclude our study of equations in divergence form by proving
the Harnack Inequality using Moser iteration. Personally I would have liked to have proved the Harnack
inequality in my own flavor, but due to lack of time, I had to follow very closely the proof given in [1].
The second half of the thesis revolves around equations in non-divergence form: aij uij = 0. As a
disclaimer, I wrote the second half separately from the first and so my notations change heavily. We first
start with the proof of the ABP maximum principle which is used heavily in, not only the proof for the
Harnack inequality for non-divergence equations, but for the section on curved C 1,α domains. In that
section I go over a completely new way of getting regularity estimates: approximation by polynomials. We
first show how this can be used for ∆u = f and then for general elliptic operators. We conclude the paper by
introducing Krylov’s regularity results for flat domains and generalize it to curved domains whose boundaries
look locally like the graph of a C 1,α function.
As one more disclaimer, I was not able to prove every single detail in this book due to lack of time. I
leave tiny bits and pieces as exercises for the reader, but the overwhelming majority is proved in rigorous
detail.
The writing of this paper was a long and arduous process. It grew out of many discussions with Professor
Daniela De Silva and Professor Ovidiu Savin. I am very grateful for their insights and very very patient
guidance. I would not have been able to set my path as a mathematician without them. I’d like to give
thanks to Professor Michael Weinstein for being an exceptional guidance counselor and instructor whom I’ve
learned a lot from about singular integrals, methods of characteristics, and graduate schools. I’d like to give
special thanks to Professor Duong Phong, whose Analysis II class helped refine the details of this thesis.

2
2 Variational Methods and Sobolev Embedding Theorems
2.1 Laplacian
We begin with the simplest problem. Let Ω ⊂⊂ Rn be a bounded domain, and f a function on Ω. We wish
to find a u such that 
∆u = f in Ω
u = 0 on ∂Ω
The way we will find the solution to this problem is by finding the minimum for a specific functional. This
is the idea: let g(x) be a function on Ω ⊂ R such that ∃G(x) on Ω with

G0 (x) = g(x)

and we want to find an x0 with g(x0 ) = 0. Then one way to approach this problem is to find an x0 ∈ Ω
such that G(x) ≥ G(x0 ) ∀x ∈ Ω =⇒ G0 (x0 ) = 0 =⇒ g(x0 ) = 0. By comparisons, g(x0 ) = 0 would be the
equivalent of ∆u = 0. So now we try to find the equivalent of G. Define the functional
ˆ ˆ
1
I(u) = |Du|2 + f u.
2 Ω Ω

This functional is going to be defined for u ∈ W 1,2 (Ω). Now lets go over some definitions.

Definition 2.1. Let B be a normed Banach space. Then its completion B := {uk } ⊂ B, uk is cauchy .
Example 2.2. Q = R.
Example 2.3. {C0∞ (Ω), kukp < ∞} = Lp (Ω) (modulo the equivalence that f ≡ g if f 6= g on a set of
measure zero).
Example 2.4. W01,2 (Ω) = {uk ∈ C0∞ , kuk − ul k2 → 0, kDuk − Dul k2 → 0}. Also, W01,2 (Ω) = {u ∈
L2 (Ω), L2 (Ω) 3 u = limk→∞ uk , kDuk − Dul k2 → 0, Duk → v ∈ L2 .}.
Now we go back to our question: does there exist a u0 ∈ W01,2 (Ω) such that I(u) ≥ I(u0 ) for any
u ∈ W01,2 (Ω)? We will begin to show this in two steps. Our first is to show that our functional is bounded
below. We will show that ∃C > 0 such that I(u) ≥ −C for any u ∈ W01,2 (Ω). This will at least give us a
starting point to find a minimum because we no longer have the ambiguity of I(u) exploding to −∞. Next,
assuming this, if I(u) did have a minimum, then min I(u) < ∞. Now pick a minimizing sequence {uj }
such that I(uj ) → min I(u). The reason why this can be picked will be shown later. We wish to show that
uj → u0 and I(u0 ) = min I(u).
Claim. We claim that ∃C > 0 such that I(u) ≥ −C ∀u ∈ W01,2 (Ω).
First recall the definition of I:
ˆ ˆ
1
I(u) = |Du|2 dx + f udx.
2 Ω Ω

Then applying Hölder’s inequality and Cauchy’s inequality, we have

ˆ
1
I(u) ≥ |Du|2 dx − kf k2 kuk2
2 Ω
ˆ  
1  1
≥ |Du|2 dx − kuk22 + kf k22 .
2 Ω 2 2

Now note that we would be done with our claim if we show that 21 kDuk22 ≥ 2 kuk22 . This is saying that the
gradient controls our function u. But we are in luck because we choose u to have compact support, so it is
zero on the boundary. Then this implies that the gradient not only approximates values near u, but tells us
what they are.

3
Lemma 2.5. There exists  = (Ω) such that 2 kuk22 ≤ 14 kDuk22 .
Proof. First note that this will imply
ˆ  
1 1
I(u) ≥ |Du|2 − f 2 dx.
Ω 4 2
Now we move onto a claim:
Claim. Let Ω be convex, and bounded in Rn . Then I will show that ∀u ∈ C ∞ (Ω), ∀x ∈ Ω, then
ˆ ˆ
(diam Ω)n |Du(y)|
|u(x) − u(y)| dy ≤ dy.
Ω n Ω |x − y|n−1
Proof of Claim. Let x, y ∈ Ω be arbitrary, and denote r = |x − y|. Let ω = (y − x)/|x − y| be the unit vector
in this direction. Then we have that
ˆ r
d
u(y) − u(x) = u(x + tω) dt
0 dt
ˆ rX n
∂u
= (x + tω)ωi dt
0 i=1 ∂xi
ˆ r
|u(x) − u(y)| ≤ |Du(x + tω)||ω| dt
ˆ ˆ0 ˆ r
|u(x) − u(y)| dy ≤ |Du(x + tω)| dt dy.
Ω Ω 0

We transform the RHS of this integral inequality using the polar transformation centered at x: y 7→ (r, ω).
Let `ω denote the distance from x to ∂Ω.
ˆ ˆ ˆ `ω ˆ r 
|u(x) − u(y)| dy ≤ |Du(x + tω)| dt rn−1 dr dσ(ω)
Ω S n−1 0 0
ˆ ˆ `ω ˆ r 
= rn−1 dr |Du(x + tω)| dt dσ(ω)
S n−1 0 t
ˆ ˆ `ω
`nω
≤ |Du(x + tω)|dt dσ(ω)
S n−1 0 n
ˆ ˆ `ω
(diam Ω)n
≤ |Du(x + tω)| dt dσ(ω)
n S n−1 0
ˆ ˆ `ω
(diam Ω)n |Du(x + tω)| n−1
= t dt dσ(ω)
n S n−1 0 tn−1
ˆ
(diam Ω)n |Du(x + tω)|
= dy
n Ω tn−1
ˆ
(diam Ω)n |Du(y)|
= dy.
n Ω |x − y|n−1

With the proof of the claim out of the way, we can prove the following corollary:
Corollary 2.6. Let u ∈ C0∞ (Ω). Then for any x ∈ Rn , we have
ˆ
|Du|
|u(x)| ≤ cn dy.
R n |x − y|n−1

4
Proof of Corollary. We begin by introducing a little bit of notation:

ūΩ (y) := u(y) dy.

Then we have for Ω = BR (0)

ˆ
1
|u(x) − ūBR (0) | = (u(x) − u(y)) dy
|BR (0)| BR (0)
ˆ
1
≤ |u(x) − u(y)| dy
|BR (0)| BR (0)
n ˆ
1 (2R) |Du(y)|
≤ n−1
dy
|BR (0)| n BR (0) |x − y|
ˆ
|Du(y)|
= cn n−1
dy.
BR (0) |x − y|

Notice that we will be done with our corollary if we show that ūBR (0) → 0 as R → ∞. But this is where we
use the fact that u has compact support:
|Du(y)| |Du(y)|
ūBR (0) = dy = dy → 0 as R → ∞.
BR (0) |x − y|n−1 Rn |x − y|n−1

Now that we are done with this corollary, we will quote a lemma that will be proved later:
Lemma 2.7 (Estimates for Integral Operator). Assume
ˆ
|u(x)| ≤ K(x, y) · |v(y)| dy.

Then for any 1 ≤ p ≤ ∞, we have ˆ ˆ

p
|u(x)| dx ≤ A |v(y)|p dy

where  ˆ ˆ 
A = max sup K(x, y) dy, sup K(x, y) dx. .
x y

Applying this lemma to Corollary 2.6 gives us kukp ≤ CkDukp for u ∈ C0∞ (Ω) (and in fact for u ∈ W01,2 (Ω)
by approximations). We are finally done with our Lemma 2.5.
Now that we are done with the first part of the problem, we go back to the infimum question. Assume
we know that inf u∈W 1,2 (Ω) I(u) > −∞ and pick a minimizing sequence, i.e. uk ∈ W01,2 (Ω) such that
0
I(uk ) → inf I(u). By approximations, we may assume uk ∈ C0∞ (Ω). The question now is do the uk ’s
converge? To do this we have to go through a few things:
Claim. I claim that {uk } is a bounded sequence, i.e. ∃C > 0 that is independent of k such that kuk k2 ≤
C, kDuk2 ≤ C.
Proof of Claim. The fact that uk is a minimizing sequence of I(u) implies that ∀k,
1 1
C1 ≥ I(uk ) ≥ kDuk22 − kf k22 .
4| 2
We can bring the f term to the other side and get C2 ≥ kDuk2 . Poincaré’s inequality then implies that
C3 ≥ kuk22 .

5
 Ok now we have that {uk } is a bounded sequence. Recall that in a finite dimensional vector space,
boundedness implies pre-compactness. However, our functional space W01,2 (Ω) is infinite dimensional, so we
need to find a weaker substitute called “weak compactness.”
Definition 2.8. Let B be a Banach space and B ∗ be its dual space (space of bounded linear functionals),
i.e., ` ∈ B ∗ is linear and |hl, ui| ≤ Ckuk for any u ∈ B. Let {uk } ⊂ B. Then we say that uk * u weakly if
∀` ∈ B ∗ ,
h`, uk i → h`, ui.
It is easy to see that if uk → u in the usual sense, then uk * u weakly. Let ` ∈ B ∗ so we have
|h`, uk − ui| ≤ Ckuk − uk → 0. The converse is not true. For an √easy example, let uk be an orthonormal basis
in an infinite dimensional Hilbert Space. Then kuk − ul k = 2, so we obviously do not have convergence.
On the other hand, from Parseval’s formula, we have
X
|h`, uk i|2 = k`k2 → 0.
uk

Then this implies that hl, uk i → 0 ∀` =⇒ uk * 0 but uk 6→ 0. Now we go to a result from analysis:
Theorem 2.9 (Bamach-Alaoglu). Let B be a reflexive separable Banach space. Then for any bounded
sequence uk ⊂ B, there exists a subsequence ukl ⊂ B such that ukl * u weakly.

We will apply this to our problem. We have that I(uk ) → inf I(u) and kuk k2 + kDuk2 ≤ C. By passing
through our subsequence, we have a u∞ ∈ W01,2 (Ω) such that uk * u∞ and Duk * Du∞ . Now the
question that we have to answer is if u∞ is the minimum that we seek after. However, we can’t say that
uk * u∞ , Duk * Du∞ implies I(uk ) → I(u∞ ). The problem with this is that I(u) is not continuous with
respect to weak convergence. However, it is lower semi-continuous! i.e., uk * u weakly in L2 implies that
kuk2 ≤ lim inf k→∞ kuk k2 . Here is the proof of this:
Proof. ˆ ˆ
kuk22 = uū = lim uūk ≤ lim inf kuk2 kuk k2 .
k→∞ k→∞

In particular,
1
I(u∞ ) = kDu∞ k22 + hf, u∞ i
2  
1
≤ lim inf kDuk k22 + hf, uk i
k→∞ 2
= lim inf I(uk ).
k→∞

However, since the following inequality is automatic,

inf I(u) ≤ I(u∞ ) ≤ lim inf I(uk ),

k→∞

and we have that I(u∞ ) = inf I(u).

We used the Banach-Alaoglu Theorem: if B is a reflective and separable Banach space and {uj } is a
bounded sequence, then there exists a weakly precompact sequence. The missing steps we have are to show
the proof of this, and completely show that kuk2 ≤ CkDuk2 for u ∈ C0∞ (Ω). The proof of Theorem 2.9 will
be left as a black box. We move onto the proof of Lemma 2.7.

6
Proof of Lemma 2.7. Our main tool will be Hölder’s inequality. Choose p and p∗ such that 1/p + 1/p∗ = 1.
Then we have
ˆ
1 1
|u(x)| ≤ |K(x, y)| p∗ |K(x, y)| p |v(y)|dy
ˆ  p1∗ ˆ  p1
p
≤ |K(x, y)|dy |K(x, y)||v(y)| dy
ˆ ˆ ˆ  pp∗ ˆ !
p p
|u(x)| dx ≤ |K(x, y)|dy |K(x, y)||v(y)| dy dx
ˆ  pp∗ ˆ ˆ 
p
≤ sup |K(x, y)|dy |K(x, y)||v(y)| dy dx.
x
´
Since |v(y)|p dy is a constant in terms of x, we can take it out of the integral of the RHS. Then switching
the order of integration and applying the same bounding trick we have
ˆ ˆ  pp∗ ˆ ˆ
p
|u(x)| dx ≤ sup |K(x, y)|dy sup |K(x, y)|dx |v(y)|p dy
x y
p
+1
kukpp ≤A p∗ kvkpp
kukp ≤ Akvkp .

Notice that this lemma would be pointless if A = ∞ because then we learn nothing know with kukp ≤ ∞.
We are in luck because we can actually deduce that A is finite in our case! The reason for this is because
|x − y|n−1 yields a singularity of dimension strictly less than n when taking the supremum over the y’s, and
so it integrable. Finally since x is over a set of compact support, it is bounded by a constant. Hence for
u ∈ C0∞ (Ω) =⇒ kukp ≤ CkDukp . Now we are finally done with our result from earlier.
Now we propose a question: how do we sharpen our estimates? A better way to visualize this question
is by noticing that the Kernel is integral for any power less than n. We actually used the worst power in
our previous proof. The answer to our question comes from the following inequalities:

Theorem 2.10 (Sobolev Inequality). Let Ω ⊂⊂ Rn and u ∈ C0∞ (Ω). Then for any p < n we have

kuk n−p
np ≤ C
n,p kDukp .

Theorem 2.11 (Trudinger Ineqality). For p = n, there exists constants K, C > 0 such that
ˆ  n
 n−1
|u(x)|
exp ≤ C.
Ω KkDukn

Theorem 2.12 (Morrey’s Inequality). For any p > n, we have

kukC α (Ω) ≤ CkDukp

where α = 1 − n/p and

|u(x) − u(y)|
kukC α (Ω) = sup |u| + sup .
Ω x6=y |x − y|α
Lets first think of why the first inequality is a better estimate than the Poincaré inequality that we have.
It is better because np/(n − p) > p and we know that the Lp norms grow bigger as p increases. This means
that we have essentially closed the gap between u and Du that we got from Lemma 2.7. Ok now lets try to

7
prove this. The proof is going to follow the same theme from Lemma 2.7’s proof. From Hölder’s inequality
we have
ˆ
|u(x)| ≤ |K(x, y)|α |K(x, y)|1−α |v(y)|1−β |v(y)|β dy

ˆ 1/a ˆ 1/c ˆ 1/b

≤ |K(x, y)|αa dy |K(x, y)|(1−α)c |v(y)|(1−β)c dy |v(y)|βb dy

where 1/a + 1/b + 1/c = 1. We need to choose our parameters wisely so that we have our desired estimates.
One obvious constraint to put is βb = p and (1 − β)c = p because we want |v(y)| on the RHS to have powers
of p. Raising everything to the power q and integrating gives us
ˆ ˆ ˆ  ˆ q/a  ˆ  ! q/c βq
|u(x)|q dx ≤ |K(x, y)|αa dy |K(x, y)|(1−α)c |v(y)|p dy |v(y)|p dy dx

ˆ ˆ q/a ˆ q/c !
αa (1−α)c p
= |K(x, y)| dy |K(x, y)| |v(y)| dy kvkβq
p dx.

In order to make our calculations a little bit easier take q = c. Then we can do the following:
ˆ ˆ ˆ  ˆ q/a  !
|u(x)|q dx ≤ |K(x, y)|αa dy |K(x, y)|(1−α)c |v(y)|p dy kvkβq
p dx
ˆ q/a ˆ ˆ 
≤ sup |K(x, y)|αa dy |K(x, y)|(1−α)c |v(y)|p dy kvkβq
p dx
x
ˆ q/a ˆ 
≤ sup |K(x, y)|αa dy sup |K(x, y)|(1−α)c dx kvkp+βq
p
x y

Recall that so far 0 < α < 1 is arbitrary. Choose it so that αa = (1 − α)c. Now lets play around with these
parameters. Recall that we chose q = c and (1 − β)c = p =⇒ β = 1 − p/q. We can then plug this into
βb = p =⇒ b = p/β to get
p p pq
b= = = .
β 1 − pq q−p
Now that we have parameters b and c, we can plug this into 1/a + 1/b + 1/c = 1 to get the parameter a.
After skipping some steps we see that 1/a = 1 − 1/p. Finally recall that we have αa = (1 − α)c. plugging in
our value for a and c and solving for α gives us
q(p − 1) pq
α= , αa = .
p + pq − q p + pq − q
When applied to our gradient estimates, this means we need the integral
ˆ
|K(x, y)|αa dy

to be finite. Following our explanation after our proof of Lemma 2.7 tells us that we need the integral
ˆ  pq
 p+pq−q
1
dy
|x − y|n−1
to be finite. Comparing the powers would require
pq
(n − 1) < n.
p + pq − q

8
After playing around with this inequality we get
pn
q< .
n−p
What suffices to show the full proof of Theorem 2.10 is that our coefficient in front of kDukp must depend
only on n and p. Additionally, setting K(x, y) = |x − y|−n+1 in our generalization of Lemma 2.7 would mean
that A would only be finite when
ˆ  pq
 pq+p−q
1
sup dx < ∞.
x |x − y|n−1
pq 1 1 1
One can see that our kernel will be integrable ⇐⇒ (n − 1) pq+p−q < n ⇐⇒ p − q < n. We now prove the
following general lemma where u no longer has compact support:
1 1 1
Lemma 2.13. Assume p − q < n. If u satisfies
ˆ ˆ
(diam Ω)n |Du(y)|
|u(x) − u(y)|dy ≤ dy
Ω n Ω |x − y|n−1
then !1+ q1 − p1
1 1
1+ q − p (diam Ω)n
ku − uΩ kq ≤ cn 1 1 1 1 1 kDukp .
n + q − p |Ω|1− n + p
Proof. It suffices to show, after dropping some constants,
( ˆ  pq
 pq+p−q )1+ q1 − p1
1+ 1
− 1
!1+ q1 − p1
1 q p 1 1 1
sup dx ≤ 1 1 1 |Ω| n − p + q . (1)
x |x − y|n−1 n + q − p

The way to do this is from this simple fact: ∀0 ≤ µ < 1,

ˆ
dy 1
µn
≤ cn |Ω|1−µ .
Ω |x − y| 1−µ
The way to prove this fact is to show that it is true for balls and then for general Ω through “rearrangement
inequalities.” Then (1) easily follows.
np
We have finally finished the proof of the Sobolev inequality. Observe now that p1 − 1q < n1 ⇐⇒ q < n−p .
np
We can ask the question if this holds for q ≤ n−p ? We’d think that the answer is no! Before we do that we
give an alternative proof for the Sobolev Inequality:
Theorem 2.14. Let u ∈ C0∞ (Ω) and 1 ≤ p < n. Then
np ≤ C kDuk .
kuk n−p s p

Proof. Consider first p = 1 and u ≥ 0. Notice that we can write u(x) in the following way:
ˆ ∞
u(x) = χ{u>t} dt.
0

This makes sense because 

1 u>t
χ{u>t} =
0 u≤t
´u
and so the RHS becomes 0
dt. Since p = 1 our goal is to show that kuk n−1
n ≤ CkDuk1 . Then we see
ˆ ∞
kuk n−1
n ≤ kχ{u>t} (·)k n−1
n dt.

9
But the inside of this inequality is
ˆ  n  n−1
n
ˆ n
! n−1
n−1
 n−1
kχ{u>t} (·)k n−1
n = χ{u>t} (·)k n−1
n dx = dx = (Vol{u > t}) n

Ω {u>t}

Notice that this has dimension of surface area because we can interpret Vol as having n dimensions, then we
take the nth root of it leaving one spacial dimension, and then we raise it to the n + 1th dimension again.
Then we can use the following isoperimetric inequality:
n−1
(Vol{u > t}) n
≤ CS Area(∂{u > t}).
This gives us now the inequality
ˆ ∞
kuk n−1 ≤ Cs Area(∂{u > t})dt.
n
0

Now we will apply iterated integrals using the coarea formula: let u be a real valued function that isn’t
constant. Then
1
dx = dσt dt (2)
|Du|
is the coarea formula. What we are doing is integrating the {u = t} level set with the dσt measure and then
integrating with respect to dt. Applying this gives us
ˆ ∞
kuk n−1 ≤ Cs Area(∂{u > t})dt
n
0
ˆ ∞ ˆ 
= Cs dσt dt
ˆ0 u=t

= Cs |Du|dx.
Ω

We are done with the proof because for p 6= 1 we notice that u ∈ Lp ⇐⇒ up ∈ L1 .

With this being said we turn back to the Trudinger inequality: for p = n, there exists constants Kn , Cn >
0 such that ˆ   n
|Ω| |u(x) − uΩ | n−1
exp dx ≤ Cn |Ω|.
Ω Kn (diam Ω)n kDukn
Notice that this is stronger than what the Sobolev inequality says for p < n because exponentials are
always greater than polynomials, and so this is stronger than ku − uΩ kp < ∞. However it is weaker than
ku − uΩ k∞ ≤ C because this deals with the supremum, which is the best bound we can have.
Proof of the Trudinger Inequality. We will recall the integral inequality
!1+ q1 − p1
1 + 1q − p1
ku − uΩ kq ≤ cn 1 1 1 kDukp (3)
n + q − p

and drop terms involving Ω because in the end they’re just constants. We expand the exponential as a power
series to get
ˆ  n
 n−1 ∞ ˆ   kn
1 |u(x) − uΩ | X 1 |u(x) − uΩ | n−1
exp dx = dx
Ω Kn kDukn k! Ω Kn kDukn
k=0
kn
∞
! n−1
X 1 ku − uΩ k n−1
kn
= (4)
k! Kn kDukn
k=0

10
Now we go back to (3) and plug in p = n to get the following inequality:
  1+ q1 − n1
ku − uΩ kq 1 1
≤ cn q 1 + −
kDukn q n
1
≤ cn q 1− n .
kn
This was verified on paper by playing little tricks. Anyway, we plug this into (4) with q = n−1 :

ˆ  n
 n−1 ∞  kn
1− n1 ! n−1
1 |u(x) − uΩ | X 1 1 kn
exp dx ≤ cn
Ω Kn kDukn k! Knn−1
k=0
∞ kn  k
X 1 1 n−1 kn
= cn
k! Kn n−1
k=0
∞
!k
X kk n
= n .
k! Knn−1 (n − 1)
k=0

Then something called Stirling’s formula makes this converge.

Recall that we have shown this basic inequality before: Consider p, q and assume 1 ≤ p ≤ q ≤ ∞ and
1
p − 1q < n1 . Then
!1+ q1 − p1
1 + 1q − p1 (diam Ω)n
ku − uΩ kq ≤ cn 1 1 1 1 1 kDukp
n + q − p |Ω|1− n + p
for Ω convex and u satisfying one of our integral inequalities. Then certainly we have
!1+ q1 − p1
1 + 1q − p1 (diam Ω)n
|u − uΩ | ≤ cn 1 1 1 1 1 kDukp (5)
n + q − p |Ω|1− n + p
We have used this to show the Sobolev Inequality and Trudinger inequality. Now we prove Morrey’s inequal-
ity.
Theorem 2.15 (Morrey’s Inequality). Let p > n, α = 1 − np , Ω = S n−1 . Then

kukC α ≤ CkDukp .
Proof. Recall that
u(x) − u(y)
kukC α = kuk∞ + sup
x6=y |x − y|α
where 0 < α < 1. It the suffices to show that each term is bounded by kDukp i.e. kuk∞ ≤ CkDukp and
[u]C α ≤ CkDukp . Clearly the first part follows from (5) with q = ∞. Now we prove the second part. Fix
x, y, x 6= y and let δ = |x − y|. Now define Ω̃ := Bδ (x) ∩ Bδ (y). Clearly this is convex. Then we have the
following inequalities
|u(x) − u(y)| ≤ |u(x) − uΩ̃ | + |uΩ̃ − u(y)|
ˆ
(diam Ω̃)n
ˆ 
|Du(z)| |Du(z)|
≤ n−1
dz + n−1
dz
|Ω̃| Ω̃ |x − z| Ω̃ |y − z|

Now notice that diam(Ω̃) ≤ 2δ and |Ω̃| ≤ |Bδ (x)|. Finally since all the terms inside the integral are positive
then we can write
ˆ ˆ !
Du(z) |Du(z)|
|u(x) − u(y)| ≤ C n−1
dz + n−1
dz .
Bδ(x) |x − z| Bδ (y) |y − z|

11
Now at this point we don’t have to reinvent the wheel so we see now that our function satisfies the correct
requirements of (5) so we can write

diam(Bδ (x))n
|u(x) − u(y)| ≤ CkDuk 1 1
|Bδ (x)|1− n + p
1 1
≤ CkDukp (2δ)n · (ωn δ n ) n − p −1
n
≤ CkDukp δ 1− p −n δ n
≤ CkDukp δ α

And we are done with the proof.

Theorem 2.16 (Sobolev Embedding Theorem). Let u ∈ W k,p (Ω). Then kukL∞ (Ω) ≤ kukW k,p (Ω) for
1/p < k/n.
Proof. Lets recall Morrey’s inequality. This says that if 1/p > 1/n, then

|u(x) − u(y)|
kuk∞ + sup = kukC α ≤ CkDukp
x6=y |x − y|α

for u ∈ C0∞ and α = 1 − n/p.

Let k = 1. Then the theorem follows from Morrey’s inequality because 1/p < 1/n and kuk∞ is obviously
bounded by kukC α . Now assume the theorem holds for k and ∀B, consider Dβ u where |β| ≤ k. Set v = Dβ u.
Then by definition we have kDvkp ≤ kukW k+1,p . Since by we are assuming p1 < k+1n pick some  > 0 such
that p = n − . Now recall the Sobolev inequality: kvkq ≤ kDvkp for p − q < n1 ⇐⇒ p1 − n1 ≤ 1q . We
1 k+1 1 1

will choose a q satisfying (for some 0 < )

1 1 1
= − + 0
q p n
k+1 1
= − + 0 − 
n n
k k
= + 0 −  < .
n n
1 k 1 k
Now we have kvkq ≤ kukW k+1,p for q < n and by definition kukW k,q ≤ kukW k+1 ,p for q < n. And so our
induction hypothesis tells us
kuk∞ ≤ kukW k,q ≤ kukW k+1,p .

As a consequence we have then that if u ∈ W0k,p (Ω) then u ∈ C 0 (Ω) if 1/p < k/n. The way to see this
is the following: Let {uj } ⊂ C0∞ (Ω) and uj → u with respect to k · kW k,p (Ω) . This of course exists because
W0k,p (Ω) = {C0∞ (Ω)|k · kW k,p (Ω) < ∞}. Then we apply the Sobolev embedding theorem to uj − um . Then
we have that kuj − um kC 0 ≤ Ckuj − um kW k,p (Ω) → 0. This implies that uj converges uniformly and so
lim uj is continuous. In fact, uj → u uniformly in the usual sense because uniform convergence implies that
uj → u in Lp . However, since W k,p (Ω) convergence is stronger, we have uj → u in the usual sense.
Lets summarize what we’ve done. Let Ω ⊂⊂ Rn , f ∈ L2 (Ω). Define
ˆ  
1 2
I(u) = |Du| + f u dx
Ω 2

for u ∈ W01,2 (Ω). Then we showed that ∃u∞ ∈ W01,2 (Ω) with I(u∞ ) = inf I(u). We will observe that the
minimum of a functional I(u) is going to be a generalized solution of the Euler-Lagrange equation for I(u).
The basis on which we will set our ground on is that if x∞ = minx∈Ω f (x) =⇒ f 0 (x∞ ) = 0.

12
 Let ϕ ∈ C0∞ (Ω) and consider for t << 1 the function R 3 t 7→ A(t) = I(u∞ + tϕ). Then we see that
t = 0 is going to be a minimum for A(t) =⇒
ˆ  
dA d 1
0= = |D(u∞ + tϕ)|2 + f · (u∞ + tϕ) dx
dt t=0 dt 2
ˆ  t=0 
d 1 2 2 2
= (|Du∞ | + 2tDu∞ Dϕ + t |Dϕ| ) + f · (u∞ + tϕ) dx
dt 2
ˆ t=0
= (Du∞ Dϕ + f ϕ) dx
ˆ X
 
n
∂u ∞ ∂ϕ
=  + f ϕ dx.
j=1
∂xj ∂x j

Now if we assume temporarily that u∞ ∈ C 2 (Ω) then we are allowed to integrate by parts and get
ˆ X ∂2u
!
0= − + f ϕdx.
∂x2j

Since this is true for any ϕ then we have −∆u∞ + f = 0, which is the Laplace equation that we wanted to
solve from the first page.

2.2 A non-linear PDE and Method of Continuity

2.2.1 Non-linear variations on Manifolds
Before we can extend our current groundwork to general manifolds, we need to expand our theory to more
general boundary values. Let Rn+1
+ = {x ∈ Rn+1 |xn+1 ≥ 0} and let Ω ⊂⊂ Rn+1 ∞
+ . Assume u ∈ C0 (Ω). Then
we will attempt to show

ku(·, 0)kLp (Rn ) ≤ CΩ k∂n+1 ukLp (Rn+1 ) . (6)

+

Pick d > diam Ω. Since u has compact support we have

ˆ d
∂u
|u(x, 0)| = |u(x, 0) − u(x, d)| = − (x, xn+1 )dxn+1
0 ∂xn+1
ˆ d p
!1/p ˆ
d
!1/q
∂u
≤ (x, xn+1 ) dxn+1 1q dxn+1
0 ∂xn+1 0
ˆ d p
∂u
|u(x, 0)|p ≤ C (x, xn+1 ) dxn+1
0 ∂x n+1
ˆ ˆ ˆ d p
p ∂u
|u(x, 0)| dx ≤ CΩ (x, xn+1 ) dxn+1 dx.
0 ∂xn+1
Taking the pth root gives us the desired claim.
We will now attempt to show that this inequality will allow us to define u|∂Ω for an arbitrary u ∈ W 1,p (Ω).
Take uj ∈ C ∞ (Ω) where uj → u in W 1,p . We know this sequence converges because W 1,p is the completion
of C ∞ with respect to the p norm. Then by our inequality (6), we can see

k(uj − uk )|∂Ω kp ≤ Ckuj − uk kW 1,p → 0

by definition and so uj is cauchy. But since Lp is complete we can then formally define

u|∂Ω := lim uj |∂Ω

13
where the limit is taken over the Lp norm. Recall that we still had Ω being some some simple semi-circle in
Rn+1
+ .
Lets extend the notion of boundary values in more general Ω with Ω ∈ C ∞ . Let Ω̃ be a small subset of
Ω such that Ω̃ ∩ ∂Ω 6= ∅. Let v ∈ C0∞ (Ω̃). We will define (using norms) v|∂ Ω̃ . Let y ∈ Ω̃. Since the boundary
is smooth we can map Ω̃ into an upper half sphere as before with y 7→ x. Of course, we can go backwards.
So then we can say v(y) = v(y(x)) =: u(x) and note that via our definitions, v|∂ Ω̃ = u(x, 0) and from our
previous observations we have
∂u
ku(·, 0)kLp (Rn ) ≤ C
∂xn+1 Lp (Rn+1 )
∂
=C v(y(x))
∂xn+1 Lp (Rn+1 )
n+1
X ∂v
≤C (7)
∂y l Lp (Rn+1 )
l=1
≤ CkvkW 1,p (Rn+1 ) .
Noticed that (7) we did something very fishy that I will now justify. We did the change of variables from
integrating with respect to the x coordinates to integrating with respect to the y coordinates. The problem
with this is that the integrals might not be bounded in the correct way. Recall
∂y l
dy = det dx
∂xj
∂y l
and c ≤ det ∂xj ≤ C. And so we have for a general function f
ˆ
kf (·)kpLpy = |f (y)|p dy
ˆ
∂y l
= |(f ◦ y)(x)|P det dx
∂xj

and so we have
ckf ◦ ykpLpx ≤ kf kpLpy ≤ Ckf ◦ ykpLpx
and (7) is valid.
Claim. The following inequality holds for any v (not just those supported in a boundary neighborhood):
v ∂Ω Lp (∂Ω)
≤ CkvkW 1,p (Ω) .

Proof. Note that cdx ≤ dσ ≤ Cdx and we are able to apply the same argument as above (i.e. norms are
equivalent under change of variable). So now the problem is to deal with
P the full Ω. Since Ω is compact we
∞
cover it Ω = ∪N
α=1 Ωα and pick a partition of unity χα ∈ C0 (Ωα ) and χ = 1. Then
N
X N
X
v ∂Ω Lp (∂Ω)
≤ kχα vkLP (∂Ω) ≤ kχα vkW 1,p (Ω)
α=1 α=1
N
X

≤ kχα vkLp (Ω) + kD(χα v)kLp (Ω)
α=1
≤ CkvkW 1,p (Ω)
Where we have bounded kχα vkp ≤ kvk and expanded the second term using the Leibniz rule.

14
 With this done we can now extend our previous work to more general boundary conditions. Say you
wanted to solve ∆u = f with u|∂Ω = g. Then if g ∈ Lp (∂Ω) then choose a G ∈ W 1,p (Ω) whose restriction is
equal to g and then consider v = u − G and the problem ∆v = ∆u − ∆G and v|∂Ω .
We will begin by analyzing a Non-Linear PDE. Let M be a compact Riemannian manifold of dimension
2. Let
X 2
ds2 = gij dxi dxj
i,j=1
∞
be the Riemannian metric. Let v ∈ C (M ) and let R be a given negative constant. We want to solve

∆u + λeu+v − R = 0. (8)
√
Notice that the exponential of u makes this a very non-linear equation. Let g := det gij . Then we will
define the laplacian as follows
2
1 X √ ij

∆u = √ ∂k gg ∂j u
g
j,k=1
2
X 1 √ jk

= √ gg ∂k ∂j u + first order terms
g
j,k=1
2
X
= g jk ∂k ∂j u + first order terms.
j,k=1

Example 2.17. In Euclidean space, we have ds2 = (dxi )2 =⇒ gij = g ij = δij and so
P

X ∂2u
∆u = + first order terms.
∂x2j

In order to solve this, we would have to consider the following functional:

ˆ ˆ
√ √
I(u) = |Du|2 gdx + R u gdx
M M

subject to the constraint ˆ ˆ

1 u+v √ √
e gdx = 1, V = gdx
V M M
and attempt to mimic our work for ∆u = f (show that the function is bounded from below and find a
minimum).
Note that Du = ∂j u and |Du|2 = ∂j u∂k u then we need the metric to contract the indices so we introduce
jk
g . Note that we can assume ˆ
1 √
v gdx = 0 (9)
V
because we can shift everything about constants. We will prove boundedness from below for our functional,
but first lets compare with Ω ⊂⊂ Rn and
ˆ ˆ
1
I(u) = |Du|2 dx + f udx
2 Ω Ω

for any u ∈ W01,2 (Ω) and f ∈ L2 (Ω). The way we proved this was by showing that
ˆ ˆ
|u|2 dx ≤ C |Du|2 dx,
Ω Ω

15
but can we do this on a general compact manifold? No! In general we have the following Poincaré inequality
ˆ
2 2 1 √
ku − ūkL2 (M ) ≤ CkDukL2 (M ) , ū = u gdx.
V M
Now we can write our functional in the following way
ˆ ˆ ˆ
1 √ √ √
I(u) = |Du|2 gdx + R (u − ū) gdx + Rū gdx .
2 M M M
| {z } | {z }
bounded as before can blow up

In order to deal with this difficulty we exploit our constraint

ˆ
1 √
1= eu+v gdx
V M
 ˆ 
1 u+v √
0 = log e gdx
V M
√ ´
and note that V1 gdx has total measure 1. Now recall consider Jensen’s Inequality: If dµ = 1 then
ˆ ˆ 
log(f )dµ ≤ log f dµ .

We will also use the fact that  

a+b 1 1
log ≥ log a + log b
2 2 2
in the following way
 ˆ  ˆ
1 √ 1 √
0 = log eu+v gdx ≥ log(eu+v ) gdx
V M V M
ˆ ˆ
1 √ 1 √
= (u + v) gdx = u gdx
V M V M
where we have used (9) in the last step. So now we have that our constraint implies ū ≤ 0 and so we do
indeed have our bound ˆ ˆ
1 2√ √
I(u) = |Du| gdx + R u gdx ≥ 0.
2
Now as before lets pick a minimizing sequence uj ∈ C ∞ (M ) such that I(uj ) → min I(u) and each uj satisfies
the constraint. Clearly kDuk2L2 (M ) ≤ C and so
1
kuj k2 ≤ kuj − ūj k2 + kūj k ≤ CkDuj k2 + I(uj ) ≤ C.
R
This means that our minimizing sequence is bounded and so we can apply the Banach-Alaoglu Theorem to
show that ∃u∞ ∈ W 1,2 (M ) such that Duj * Du weakly and uj * u weakly. Now the question is if u∞ is
the minimum or not? We do have
ˆ ˆ
1 2√ √
I(u∞ ) ≤ lim inf |Duj | gdx + Ruj gdx
2
= lim inf I(uj ) = lim I(uj ).
j→∞

The real question is if u∞ satisfies the constraint i.e. we wish to show

ˆ
1 √
eu∞ +v gdx = 1 (10)
V M
The bottom-line question here is if we can pass through the limit. We need to strengthen our converging
hypothesis as much as possible (i.e. from weak convergence to point-wise convergence). In fact, we claim
that for the minimizing sequence we have, Duj * Du∞ weakly but uj → u in L2 . For this we will need

16
Lemma 2.18 (Rellich’s Lemma). Let M be a compact manifold and {uj } ⊂ W 1,p (M ) with kuj kW 1p (M ) ≤ C
with C independent of j, then there exists u∞ ∈ W 1,p (M ) and a subsequent {ujk } such that ujk → u∞ in
Lp (M ).
Note. Note that this also holds for W 1,p (Rn ) if supp uj ⊂ k ⊂⊂ Rn ∀j.
Now recall a theorem from measure theory:
Theorem 2.19. If uj → u∞ in Lp for 1 ≤ p < ∞, then there exists ujk → u∞ point wise a.e.
Note that all together, we can take a subsequence of our subsequence to find a sequence kDuj k ≤ C, uj →
u∞ almost everywhere. We will see that in two dimensions these properties imply that euj +v → eu∞ +v in
L1 . Since these values are always positive, the L1 norm is just convergence of the integral, which is what we
needed. To see this,
ˆ 1
d h tu∞ +(1−t)uj +v i
euj +v − eu∞ +v = − e dt
0 dt
ˆ 1
=− (u∞ − uj )etu∞ +(1−t)uj +v dt.
0
1
Taking the L norm will give us
ˆ ˆ 1 ˆ 
√ √
|euj +v − eu∞ +v | gdx ≤ |u∞ − uj |etu∞ +(1−t)uj +v gdx dt
M 0 M
ˆ 1 ˆ 1/2 ˆ 1/2
√ √
≤ |u∞ − uj |2 gdx e2(tu∞ +(1−t)uj +v) gdx dt
0 M M
ˆ 1 ˆ 1/2
2(tu∞ +(1−t)uj +v) √
= ku∞ − uj kL2 e gdx dt
0 M

and now note that we need the second integral to be uniformly bounded for the RHS of the inequality to go
to zero. ´ √
We claim that ewj gdx ≤ C (independent of j) if kwj kL2 ≤ C and kDuj kL2 ≤ C. In this case we have
wj = 2(tu∞ +(1−t)uj +v) satisfying the condition because kDuj kL2 ≤ C and kDu∞ kL2 ≤ lim inf kDuj kL2 ≤
C and similarly the L2 norm of uj and u∞ is bounded. Now lets see why this claim is true.
Recall the Trudinger inequality: u ∈ C0∞ (B) implies that
ˆ  n
 n−1
|u(x)|
exp ≤C
KkDukLn
and so for n = 2 we have ˆ  2
|u(x)|
exp ≤ C.
KkDukL2
Now note that we can write
 2
|u(x)| 1 |u(x)| 1
|u(x)| = KkDukL2 ≤ + (KkDukL2 )2 .
KkDukL2 2 KkDukL2 2
Since the exponential function is increasing we can write
"  2 #
1 |u(x)| 1 2
exp |u(x)| ≤ exp + (KkDukL2 )
2 KkDukL2 2
ˆ ˆ  2
1 2 1 |u(x)|
exp |u(x)| ≤ e 2 (KkDukL2 ) exp
2 KkDukL2
1 2
≤ e 2 (KkDukL2 ) C

17
And this concludes the proof. A great exercise is as follows. Let M be a compact Riemannian n-manifold
and show that ˆ
√
ew gdx ≤ C exp (CkDuknLn + kwknLn ) .
´ √
As a summary, we have shown that for´a sequence satisfying I(uj ) → min I(u) and V1 euj +v gdx = 1 then
1 u∞ +v √
uj → u∞ (as explained above) and V e gdx = 1 so we have
inf I(u) ≤ I(u∞ ) ≤ inf I(u).
Now we claim that u∞ satisfies our partial differential equation in the Euler-Langrange sense. Fix
ϕ ∈ C ∞ (M ) and consider u∞ + tϕ + ct . We add the constant ct so that this function still satisfies the
constraint. To see what ct has to be note
ˆ
1 √
1= eu∞ +tϕ+ct +v gdx
V M
 ˆ −1
ct 1 u∞ +tϕ+v √
e = e gdx
V M
 ˆ 
1 u∞ +tϕ+v √
ct = − log e gdx
V M
and thus I(u∞ + tϕ + ct ) ≥ I(u∞ ) for any t and so we leave it as an exercise to show in detail
d
I(u∞ + tϕ + ct ) = 0.
dt t=0

Lets recap our problem. Our functional was

ˆ ˆ
1 √ √
I(u) = |Du|2 gdx + R u gdx
2 M M

subject to the constraint ˆ

1 √
eu+v gdx = 1 (11)
V M
where v is some smooth function and V is the volume over this compact manifold. We obtained a u∞ subject
u∞ +v
to the same constraint

that satisfies ∆u∞ + λe − R = 0. Recall that this general laplacian has the form
1 √
∆u = √g ∂i g∂j u . Where does this come from? If we assume ut is smooth then we can check
ˆ ˆ 
d 2√ d ij √
|Dut | gdx = g ∂i ut ∂j ut gdx
dt M dt
ˆ 
d ij

√
=− ut ∂i g ∂j ut gdx
dt
ˆ 
d √
=− ut ∆ut gdx
dt
So we see that this laplacian is the perfect analogue of the one taken on Rn .
We want to derive the generalized Euler-Lagrange equation for our PDE, so let ϕ ∈ C0∞ (M ) and ut =
u∞ + tϕ + ct , where ct is picked so that ut also satisfies the constraint (11). Lets figure out what λ has to
be in order to for our equation to be solved correctly. Integrate on both sides of our PDE to get
ˆ ˆ ˆ
√ u∞ +v √ √
0= ∆u∞ g + λ e g−R gdx
ˆM M
ˆ M
ˆ
1 √ ij
u∞ +v √ √
= √ i ∂ gg ∂ u
j ∞ + λ e g − R gdx
M g M M
= λV − RV

18
so we see that λ has to equal R. The reason the first term vanishes is because we are integrating an exact
form over a compact manifold. So then we have that our PDE is

∆u + Reu+v − R = 0 (12)

and u satisfies (12) in the general sense if ∀ϕ ∈ c∞ 0 we have

ˆ

√
0= ∆u + Reu+v − R ϕ g
ˆ

√
= −g ij ∂i ∂j ϕ + Reu+v − Rϕ gdx (13)

and (13) is the generalized Euler-Lagrange equation.

2.2.2 Regularity Theory

Lets finally start some regularity theory. Let gij be a smooth Riemannian metric and assume Λ ≥ gij ≥ λ > 0.
1,2
If u ∈ Wloc satisfies ∆u = f in the weak sense then
1. If f ∈ W k,p and 1 < p < ∞ the u ∈ W k+2,p and for any Ω ⊂⊂ Ω0


kukW k+2,p (Ω) ≤ CΩ,Ω0 kf kW k,p (Ω0 ) + kukW k,p (Ω0 )

2. If f ∈ C k,α and 1 < p < ∞ the u ∈ C k+2,α and for any Ω ⊂⊂ Ω0

kukC k+2,α (Ω) ≤ CΩ,Ω0 kf kC k,p (Ω0 ) + kukW k,p (Ω0 )

Claim. It follows easily that if u satisfies (13) in the generalized sense, then u ∈ C ∞ and actually satisfies
(12) in the standard sense.
Proof. Indeed we can let −Reu+v + R = f and f ∈ L2 = W 0,2 by the Trudinger’s inequality (it tells us that
this exponential is bounded by the L2 norm of the weak derivative) then by regularity we will get u ∈ W 2,2 .
Now recall that Morrey’s inequality tells us that if p1 < nk then W k,p is embedded in a Hölder space, but
since p = n = 2 we have u ∈ C α . But then since ∆u = f this implies f ∈ C α and we can apply regularity
to get u ∈ C 2,α . We can iterate this and find of course that u ∈ C ∞ .
We will prove these regularity statements by the method of continuity and a priori estimates. Consider

∆u + Reu+v − R = 0

on (M, gij (x)) where v is a smooth function. Note that ∆u + Reu − R = 0 admits a solution u ≡ 0 so we
see that the difficulty comes from the v term. Let t ∈ [0, 1] and introduce the family of equations

∆u + Reu+tv − R = 0 (14)

And consider the set I = {t ∈ [0, 1]|(14) admits a solution ut }. Note that I 6= ∅ because 0 ∈ I is a solution.
Then we obviously want to show I = [0, 1], so we need to show that I is open and closed. Lets discuss this
very briefly (to be made concise later).
Say we want to show I is open. We will do this by an analogue of the implicit function theorem. Recall
that it says given f (x0 , y0 ) = 0 and ∂f ∂y (x0 , y0 ) 6= 0 then ∃ > 0 such that for |x − x0 | < , then ∃!y such
that f (x, y) = 0. Now let f (t, u) = ∆u + Reu+tv − R. Then we want to solve f (x, u) = 0 where we know
f (t, u0 ) = 0. So our goal is going to be an implicit function theorem for Banach spaces and we need to check
∂f
that ∂u (t0 , u0 ) 6= 0 in a way we will define later.
Now lets briefly discuss how we will show that I is closed. Take tj ∈ I such that (14) will admit a solution
called uj , and assume tj → T . Then closeness of I is equivalent to showing that (14) will admit a solution
for T . It will suffice to have a subsequence converge in C 2 .

19
Proof. Lets start by showing that I is closed. Let tj → T with tj ∈ I and let uj be the corresponding
solution of (15). Observe that if suffices to show that ∃C independent of j so that kuj kC 3 ≤ C where in
general X
kukC 3 (Ω) = k∂ α ukC 0 (Ω) .
|α|≤3

The reason why this will help is that this would imply that ∀β ≤ 2, then ∂ β uj is an equicontinuous family.
Then the Arzela-Ascolati Theorem tells us that if we have an equicontinuous family, then by going through
a subsequence ∂ β uj we have uniform convergence to Dα uT where uT ∈ C 2 . However we don’t necessarily
have that T ∈ I because we need uT to be smooth and so far it is only C 2 . This is fixed however by our
regularity observations.
Before we can apply our regularity conditions we must show that our equation is uniformly elliptic. Recall
that a second order PDE is said to be uniformly elliptic if the leading coefficient satisfies

X
λ|ξ|2 ≤ aα ξ α ≤ Λ|ξ|2 .
|α|=2

This we need to show that our Laplacian is elliptic.

1 √
∆u = √ ∂i ( g∂j u)
g
1 √
= √ gg ij ∂i ∂j u + first order terms.
g

So the symbol (the middle term of the ellipticity inequality requirement) of our Laplacian is going to be

σ∆ (x, ξ) = g ij ξi ξj

and since g is positive definite we definitely have the ellipticity requirement. So then we can apply our
regularity theorems by viewing ∆u = f ∈ C 2 ⊂ C 1,α .
So now we have to prove the a priori estimate kuj kC 3 ≤ C. We will use the maximum principle. Since
there are so many different formulations of maximum principles, it is a good idea to simply examine what
happens near a maximum. Let u ∈ C ∞ satisfy ∆u + Rtw+u − R = 0 (by denoting ut = u) and then I claim
that kukC 0 ≤ C where C is independent of t. Let x0 ∈ Ω such that u(x0 ) is a maximum. Then ∆u(x0 ) ≤ 0
and since R < 0 we have

0 ≤ −∆u = Retw+u − R
R ≤ Retw+u
−|R| ≤ −|R|etw+u
1 ≥ etw+u
0 ≥ tw + u
u ≤ −tw
≤ kwkC 0 .

But since x0 is a maximum we have that ∀x, u(x) ≤ u(x0 ) ≤ kwkC0 and applying the same argument for
the minimum we have kukC0 ≤ kwkC0 . In order to get higher derivatives we write ∆u = f ∈ C 0 and so for
f ∈ Lp ∀1 ≤ p < ∞ we have u ∈ W 2,p ⊂ C α for some α when n < kp.
Now lets show that I is open. Let t0 ∈ I i.e. there exists u0 which is a smooth solution of (14). Let
(u, t) 7→ F (u, t) = ∆u + Retw+u − R. Then we want to show that ∃δ > 0 so that |t − t0 | < δ implies
∃ut ∈ C ∞ |F (ut , t) = 0. The main tool will be the Implicit Function Theorem for Banach Spaces, which goes
as follows.

20
 Let B1 and B2 be Banach spaces. Let F ∈ C 1 and consider B1 × R ⊃ Ω 3 (u, t) 7→ F (u, t) ∈ B2 . Assume
that F (u0 , t0 ) = 0. Let ∂F
∂u (u0 , t0 ) be the derivative of F at (t0 , u0 ) viewed as a linear operator B1 → B2 .
Note that if
∂F
khkB1 ≤ C (u0 , t0 )h ∀h ∈ B1 (15)
∂u B2

then ∂F
∂u (u0 , t0 ) is injective and surjective. So assume that (15) holds. Then the Implicit Function Theorem
for Banach Spaces says that ∃δ > 0, ∃V that is a neighborhood of u0 such that ∃!ut ∈ V with F (ut , t) = 0.
Before we can even apply this theorem, we need to mae sense of derivatives in terms of Banach spaces.
Let B1 3 u 7→ F (u) ∈ B2 . Then F is differentiable at u0 if ∃L : B1 → B2 that is a bounded linear operator
satisfying
F (t, u + h) = F (t, u) + Lh + E(t, u, h)
with
kE(u, h)kB2
lim = 0.
h→0 khkB1
In order to apply our theorem we need so specify our Banach spaces. We will want to have (t, u) ∈ R×C 2,α →
F (u, t) ∈ C 0,α . What we will then need to check are the assumptions of the implicit function theorem. Let
t0 , u0 ∈ R × C 2,α satisfy the conditions of the IFT. Lets determine L. Consider the expression F (t0 , u0 + h):

F (t0 , u0 + h) = ∆(u0 + h) + Ret0 w+u0 +h − R

= F (t0 , u0 ) + ∆h + Ret0 w+u0 +h − Ret0 w+u
= F (t0 , u0 ) + ∆h + Ret0 w+u eh − 1

= F (t0 , u0 ) + ∆ + Ret0 +u h + Ret0 w+u (eh − 1 − h).

So we have figured out our L. Now we leave it as an exercise to show that

kRet0 w+u0 (eh − 1 − h)kC 0,a ≤ Ckhk2C 2,α .

The main tool for this is the integral form of Taylor’s Remainder Theorem, which starts as follows
ˆ 1 ˆ 1 1
d
h (1 − t) [f 0 (u + th)] dt = f 0 (u + th)dt + h(1 − t)f 0 (u + th)
0 dt 0 0
ˆ 1
d
= f (u + th)dt − hf 0 (u)
0 dt
= f (u + h) − f (u) − hf 0 (u).

However we can write the LHS as

ˆ 1 ˆ 1
d
h (1 − t) [f 0 (u + th)] dt = (1 − t)f 00 (u + th)dt
0 dt 0

and so we have the final part as

ˆ t
0
f (u + h) = f (u) + f (u)h + h 2
(1 − t)f 00 (u + th)dt.
0

Thus we have that L = ∂F ∂u (t0 , u0 ) : C

2,α
3 h 7→ ∆h + Ret0 w+u h ∈ C 0,α , which is clearly a bounded operator.
It is trickier to show that it is injective and bijective.
Assume 0 = Lh = ∆h + Ret0 w+u h. Let x0 me a maximum (again this means ∆h ≤ 0). Then we have

−Ret0 w+u h = ∆h(x0 ) ≤ 0

21
and since the coefficients are both strictly positive, it means that h(x0 ) ≤ 0. Since this is a maximum it
means that ∀x, h(x) ≤ h(x0 ) ≤ 0. Applying the same process to the minimum we get that h(x) ≥ 0 and so
h ≡ 0. This implies that the kernel of L is zero, and so it is injective.
In order to finally complete the proof we will show that L is onto i.e. ∀f ∈ C 0,α , we want to show
∃h ∈ C 2,α so that Lh = f . We will show this by variational methods again. Set
ˆ
√ ij √
I(h) = gg ∂i h∂j h − Ret0 w+u0 h2 + f h gdx

for some h ∈ W 1,2 (M ). We leave it as an exercise to show that I(h) attains its minimum for some h∞ . One
way to do this is by showing that
1
I(h) ≥ kDhk22 + khk22 − kf k22

and using our tricks. Assuming the exercise, we invoke the black box to make h smooth. Recall that it says
if f ∈ C k,α then
khkC k+2,α ≤ C (kLhkC k,α + khkC k,α ) .
We now improve the black box by saying that if ker L = 0, then

khkC k+2 ,α ≤ CkLhkC k,α . (16)

Now we will prove this little lemma. We will use weak compactness. If {ul } ⊂ C k,α with kul kC k,α ≤ C
with C independent of l, then either ∀k < k, ∀β or k 0 = k with 0 < β < α then there exists a convergent
0
subsequence in C k ,β by the routine application of the Arzela-Ascolti Theorem. Now assume (17) does not
hold. Then for any N, ∃hN ∈ C k+2,α such that khN kC k+2,α > N kLhn kC k,α . Set

hN
h̃N =
khN kC k+2,α

and notice that kh̃N kC k+2,α = 1 and so this implies that kLh̃kC k,α < N1 → 0. By the weak compactness,
going through a subsequence, we can assume that h̃N → h∞ in C k,α . Applying the black box gives
 
kh̃N − h̃M kC k+2,α ≤ C kLh̃N − h̃M kC k,α + kh̃N − h̃M kC k,α

which implies that h̃N → h∞ in C k+2,α . Thus Lh̃N → Lh∞ in C k,α and so h∞ ∈ ker L. Since

kh∞ kC k+2,α = lim kh̃N kC k+2,α = 1

this contradicts the fact that ker L = 0.

22
3 Harnack Inequality for Divergence Equations
3.1 Regularity Estimates
Suppose that u ∈ W 1,2 (Ω) solves ∂i (aij ∂j u) + cu = f in the generalized sense, where aij is uniformly elliptic
i.e. 0 < λ ≤ aij ≤ Λ. The question we want to answer is: when is u “regular” i.e. u ∈ C α , W k,2 , C ∞ , etc?
Lets look at the simplest case where aij is constant and c = f = 0. In this case u solving aij ∂i ∂j u = 0
in the generalized sense. This means that ∀v ∈ W01,2 (Ω),
ˆ
aij ∂i u∂j v = 0. (17)

Then we will show that u ∈ C ∞ (Ω) and for |α| = k, 0 < r < R, we have
ˆ ˆ
α 2 Cλ,Λ,k
|D u| ≤ 2k
|u|2 . (18)
Br (x0 ) (R − r) BR (x0 )

Note that this inequality is very powerful because we have that the derivative is being bounded by the
function, where w usually have it the other way around.
Proof. We apply (17) with v = χ2 u with 0 ≤ χ ≤ 1, χ ≡ 1 on Br (x0 ) and χ ∈ C01 (BR (x0 )). Additionally
assume
2
Dχ(x0 ) ≤ .
R−r
Applying (17) gives us ˆ ˆ
χ2 aij ∂i u∂j u = −2 aij (χ∂j χ)u∂i u
1 1
and note that ellipticity gives us |aij ui vj | ≤ (aij ui uj ) 2 (aij vi vj ) 2 . Putting absolute values gives us
ˆ ˆ  21 ˆ  12
2 2 2
χ aij ∂i u∂j u ≤ 2 aij ∂i u∂j uχ aij ∂i χ∂j χ|u|
ˆ ˆ
1
≤ aij ∂i u∂j uχ2 + 4 aij ∂i χ∂j χ|u|2
2
ˆ ˆ
1 2
aij ∂i u∂j uχ ≤ 4 aij ∂i χ∂j χ|u|2
2

where we have used the standard fact that ab ≤ 2 a2 + 21 2

b where we chose  to give us the right coefficients.
Then from the ellipticity condition we have
ˆ ˆ ˆ ˆ ˆ
λ 1 8Λ
|Du|2 ≤ aij ∂i u∂j uχ2 ≤ 4 aij ∂i χ∂j χ|u|2 ≤ 16 Λ|Dχ|2 |u|2 ≤ |u|2
2 Br (x0 ) 2 (R − r)2 BR (x0 )

which proves the inequality for k = 1. To prove |α| = k ∈ Z, we proceed by induction. Assume u ∈ C ∞ .
Then Dα u ∈ C ∞ and satisfies the same equation aij ∂i ∂j (Dα ) = 0. Applying the previous case with k = 1
gives (by induction)
ˆ ˆ
α 2 C
|D(D u)| ≤
2 |Dα u|2
r+R
Br (x0 ) − r BR−r (x0 )
2
ˆ
C
≤ |u|2
R+r R+r 2k B (x )



2 −r R− 2 R 0
ˆ
C
= 2(k+1)
|u|2 .
(R − r) BR (x0 )

23
As one can expect, for the non-smooth case, we use mollifiers. Take η ∈ C0∞ (|x| < 1) with
ˆ
η=1
Rn

1 x
and define η = n η(  ). Then define
ˆ ˆ
u (x) = u(x − y)η (y)dy = u(y)η (x − y)dy

which is well defined for dist(x, ∂Ω) > . By an exercise we leave to the reader, note that if u ∈ Lp (Ω) for
1 < p < ∞, then for any K ⊂⊂ Ω and  < K , then u → u in Lp (K). Using this, and by a single exercise
that shows ˆ ˆ
aij ∂i u ∂j v = aij ∂i u∂j v

we can conclude that if u satisfies aij ∂i ∂j u = 0 in the generalized sense then so does u . Thus we have that
u ∈ C ∞ and aij ∂i ∂j u = 0 and thus we can apply our estimate (18) to find
ˆ ˆ
α 1
|D u | ≤ |u |2 .
Br (x0 ) (R − r)2k BR (x0 )

Lets try to generalize this. We will do this with the following theorem.
Theorem 3.1. Let u ∈ W 1,2 (Ω) be a weak solution to

∂i (aij ∂j u) + c(x)u = f (19)

in Ω ⊂ Rn . Assume that
i) 0 < λ ≤ aij ≤ Λ
ii) aij are continuous with modulus of continuity τ i.e. |aij (x) − aij (y)| ≤ τ (|x − y|)
n
iii) c ∈ Ln , f ∈ Lq where 2 < q < n.
n
Then for any BR (x) ⊂⊂ Ω, u ∈ C α (BR (x)) with α = 2 − q and 0 < α < 1 with the estimate


kukC α (B) ≤ Cn,λ,Λ,τ,kckp kf kLq (Ω) + kukW 1,2 (Ω) .

In order to prove this we will need a few lemmas. They are as follows.
Lemma 3.2. Assume u ∈ W 1,2 (B) and
ˆ
|u − ūx0 |2 dx ≤ M 2 rn+2α .
Br (x0 )

Then u ∈ C α and kukC α (B) ≤ C(M + kukL1 ).

Proof. First note that what we are doing makes some sort of sense. In the assumption, we have the difference
between u and its average at x0 being squared, so that explains the 2α. Then since we are integrating over
the ball, we have the factor of rn .
Let 0 < r < R and look at the difference between averages

|ūr (x0 ) − ūR (x0 )| ≤ |ūr (x0 ) − u(x)| + |u(x) − ūR |

|ūr (x0 ) − ūR (x0 )|2 ≤ 2 |ūr − u(x)|2 + |u(x) − ūR |2 .

24
We now integrate over the ball of radius r and using the fact that the LHS is a constant and the assumption
of the lemma we have
ˆ ˆ !
rn |ūr (x0 ) − ūR (x0 )|2 ≤ 2 |ūr − u(x)|2 + |u(x) − ūR |2
Br (x0 ) Br (x0 )
ˆ !
2 n+2α 2
≤2 M r + |u(x) − ūR |
BR (x0 )

≤ CM 2 rn+2α + Rn+2α


  n 
2 2 2α R 2α
|ūr (x0 ) − ūR (x0 )| ≤ CM r + R
r
  n 
R
≤ CM 2 1 + R2α
r
  n2 !
R
|ūr (x0 ) − ūR (x0 )| ≤ CM 1 + Rα . (20)
r

Let r = 2−l−1 L, R = 2−l L and plug this into (20) to see that
 n2 !
2−l L

|ūr − ūR | ≤ CM 1+ (2−l L)α
2−l−1 L
n
= CM (1 + 2 2 )(2−l L)α
= CM (2−l L)α . (21)

Now use (21) to see that

m
X
|ū2−m−1 L − ū2−l L | ≤ |ū2−i−1 L − ū−i |
i=l
m
X
≤ CM (2−i L)α
i=l
≤ CM (2−l L)α (22)

where we have used the fact that we had a telescoping sequence and that we have a geometric sequence that
we have bounded by the highest term. This implies that {ū2−l L } is a cauchy sequence and so we can define

u∗ := lim ū2−l L (x0 ).

l→∞

We will now show that this is independent of L. Take L < L0 . Then (22) implies that
  0 n 
L
|ū2−l L (x0 ) − ū2−l L0 | ≤ CM 1 + (2−l L0 )α
L
and taking l → ∞ shows that u∗ is independent of L.
From all of our hard work, we can say that u∗ = u almost everywhere by the Lebesgue Differentiation
Theorem which says that if u ∈ L1 , then limr→0 ūr (x0 ) = u(x0 ) for almost every x0 . Now taking m → ∞ in
(22) we are able to get

|ū2−m−1 L (x0 ) − ū2−l L (x0 )| ≤ CM (2−l L)α

|u∗ (x0 ) − ū2−l L (x0 )| ≤ CM (2−l L)α
|u∗ (x0 ) − ūr (x0 )| ≤ CM rα . (23)

25
Letting r = 1 implies that ku∗ (x0 )kL∞ (B1 ) ≤ C(M + kukL1 (B1 ) )
Now in order to complete this theorem we need to estimate [u∗ ]C α . Let x, y ∈ Ω such that Br (x), Br (y) ⊂⊂
Ω and Br (x) ∩ Br (y) 6= ∅. Denote δ = |x − y| and let z be the point midway between x and y. By convexity
we can see that Bδ (z) ⊂⊂ Br (x) ∩ Br (y). Then we write

|u∗ (x) − u∗ (y)| ≤ |u∗ (x) − ūδ (x)| + |u∗ (y) − ūδ (y)| + |ūδ (x) − u(z)| + |ūδ (y) − u(z)|
|u∗ (x) − u∗ (y)|2 ≤ C |u∗ (x) − ūδ (x)|2 + |u∗ (y) − ūδ (y)|2 + |ūδ (x) − u(z)|2 + |ūδ (y) − u(z)|2

≤ CM 2 δ 2α + C|ūδ (x) − u(z)|2 + C|ūδ (y) − u(z)|2

where we have used (23) on the first two terms. Integrating the inequality with respect to z and using the
assumptions of the lemma yields the required results.
Lemma 3.3. Assume that ˆ
|Du|2 ≤ M 2 rn−2+2α .
Br (x0 )

Then u ∈ C α .
Proof. The details of the proof will be left as an exercise. Here is a sketch of it. Recall the Poincaré
inequality: ˆ ˆ
2
|u − ūS | ≤ λS |Du|2 .
S S
Then you need to show that ˆ ˆ
2 2
|u − ūr (x0 )| ≤ Cr |Du|2
Br (x0 ) Br (x0 )

where c is independent of r. The way to do this is to apply the Poincaré Inequality with r = 1. Then consider
the rescaling ũ = u(rx). After doing this, apply the assumption of the lemma and you’ll find yourself in the
position of Lemma 3.2.
Lets recall our goals. We were proving Schauder estimates to get regularity. Assume u ∈ W 1,2 is a weak
solution of ∂i (aij ∂j u) = 0 i.e. ˆ
aij ∂j u∂i v = 0 ∀v ∈ W01,2 .

Then we will show that u ∈ C α where 0 < λ ≤ aij ≤ Λ and |aij (x) − aij (y)| ≤ τ (|x − y|) with τ ↓ 0 as R ↓ 0.
We will be using the following key estimate for 0 < r < R
ˆ h r n iˆ
2
|Du| ≤ C + τ (R) |Du|2 . (24)
Br (x0 ) R BR (x0 )

Proof. We first prove the case where we have a constant coefficient i.e. τ ≡ 0. Then we will show that
ˆ  r n ˆ
|Du|2 ≤ C |Du|2 (25)
Br (x0 ) R BR (x0 )

and use a the Lemma of De Giorgi (seen later). By rescaling v(x) = u(Rx) what we need to show becomes
ˆ  r n ˆ
R−2 |Dv|2 ≤ C |Dv|2 · R−2
B r (x0 ) R B1 (x)
R

m
ˆ ˆ
|Dv|2 ≤ Csn |Dv|2 .
Bs (x0 ) B1 (x0 )

26
This will follow from previous work. Take s small (s < 12 ). Then By the Sobolev Embedding Theorem and
by the first theorem proved in this section we are able to say
ˆ
|Dv|2 ≤ sup |Dv|2 · sn
Bs (x0 ) Bs

X ˆ
 

≤ Dα (Dv)2  sn
|α|≤k B 3 (x0 )
ˆ 4

n
≤ Cs |Dv|2
B1 (x0 )

and we are done showing (25), which corresponds to the case where aij are constants. Since this case will
follow when we prove the Lemma of De Giorgi, we move on to the non-constant case.
We will prove the non-constant case as a perturbation of the equation aij (x0 )∂i ∂j w = 0, which we have
already finished. To carry this out, consider the following Dirichlet problem

aij (x0 )∂i ∂j w = 0 weak sense
w − u ∈ W01,2 (BR (x0 )).
Set v = u − w and u = v + w. Then
ˆ "ˆ ˆ #
2 2 2
|Du| ≤ 2 |Dv| + |Dw|
Br (x0 ) Br (x0 ) Br (x0 )
"ˆ
 r n ˆ
#
2 2
≤2 |Dv| + c |Dw|
Br (x0 ) R BR (x0 )
"ˆ
 r n ˆ
#
2 2 2
≤2 |Dv| + c |Du| + |Dv|
Br (x0 ) R BR (x0 )
 r n ˆ ˆ
" #
2 2
≤C |Du| + |Dv| (26)
R Br (x0 ) BR (x0 )

where we have used the fact that w solves our PDE with constant coefficients, saw that the integral over r
n
is ≤ than the integral over R, and saw max{1, Rr n } = 1. Now we need to control the integral of |Dv|2 , so we
use the fact that v = u − w, where u solves ∂i (aij ∂j u) = 0 and w solves aij (x0 )∂i ∂j w = 0 in the weak sense.
Now since v ∈ W01,2 , we are able to use it as a test function in the definition of weak derivatives to get
ˆ ˆ
aij (x0 )Dj vDi v = aij (x0 )(Dj u + Dj w)Di v
ˆ
= aij (x0 )Dj uDi v
ˆ ˆ
= (aij (x0 ) − aij (x))Dj uDi v + aij (x)Dj uDi v
ˆ
= (aij (x0 ) − aij (x))Dj uDi v.

We now use the fact that aij is elliptic and use its modulus of continuity to see
ˆ ˆ
λ |Dv|2 ≤ τ (R) |Du||Dv|
BR (x0 ) BR (x0 )
ˆ
1 1
≤ τ (R) |Du|2 + |Dv|2
BR (x0 ) 2 2
ˆ ˆ
|Dv|2 ≤ Cτ (R) |Du|2
BR (x0 ) BR (x0 )

27
which we can plug into (26) to see that we have finally proved (24) for the case of non-constant aij .
We still need to relate everything to get that u ∈ C α . Now we finally state the Lemma of De Girogi:
Lemma 3.4 (Lemma of De Giorgi). Let ϕ(r) ≥ 0 with ϕ(r) ↓ as r ↓, and A, B ≥ 0. Assume that 0 < r < R
and ∃α > β > 0 such that h r α i
ϕ(r) ≤ A +  ϕ(R) + BRβ . (27)
R
Then ∀0 < β < γ < α, ∃0 such that  < 0 implies
h r γ i
ϕ(r) ≤ C ϕ(R) + Brβ . (28)
R
We will apply the Lemma with ˆ
ϕ(r) = |Du|2
Br (x0 )

because (24) implies that ϕ(r) satisfies (27) with B = 0 and α = n. Then by the Lemma of De Giorgi we
have  r γ
ϕ(r) ≤ C ϕ(R)
R
and by taking R = 1 we imply that ϕ(r) ≤ Crγ after absorbing ϕ(1) into the constant. We are now in the
case of Lemma 3.3 because we are free to choose any γ < α = n and
ˆ
|u − ūr (x0 )|2 dx ≤ Crγ+2 .

Hence u ∈ C α .
Recall that we consider weak solutions of

∂i (aij ∂j u) + c(x)u = f

where 0 < λ ≤ aij ≤ Λ and |aij (x) − aij (y)| ≤ τ (|x − y|) where τ ↓ 0 as R ↓ 0, c ∈ Ln and f ∈ Lq for
n α n
2 < q < n. We want to show that u ∈ C for α = 2 − q .

Proof. We already got the case for c = f = 0, where for 0 < r < R the main tool was showing
ˆ (
 r n ˆ ˆ )
2 2 2
|Du| ≤ C |Du| + τ (R) |Dv| . (29)
Br (x0 ) R BR (x0 ) BR (x0 )

Recall that we defined v = u − w where w solves


aij (x0 )∂i ∂j w = 0 on BR (x0 )
u − w ∈ W01,2 (BR (x0 ))

in the generalized sense. By the same argument as before, (29) can also be established for non-trivial lower
ordered terms. Lets now estimate |Dv| in (29):
ˆ ˆ
aij (x0 )Di vDj v = aij (x0 )(Di u − Di w)Dj v
ˆ
= aij (x0 )Di uDj v
ˆ ˆ
= aij (x)Di uDj v + (aij (x0 ) − aij (x))Di uDj v
ˆ ˆ ˆ
= − c(x)uv + f v + (aij (x0 ) − aij (x))Di uDj v. (30)

28
These calculations show two new terms that were not there before in the proof of the previous theorem. We
handle these terms by bounding their derivatives. Recall the Sobolev inequality:

kvk 2n ≤ CkDvkL2
L n−2

and use the Holder’s and the Sobolev inequality

ˆ ˆ !1/2
2
c(x)uv ≤ kukL2 |cv|
BR (x0 ) BR (x0 )
"ˆ ? ˆ ? #1/2
2n
2
≤ kukL2 |v| n−2 |c |

1 1
where we have to figure out which powers to use. We need Lesbegue conjugates p + q = 1 and so we let
2n n
2p = n−2 =⇒ p = n−2 , q = n2 . This implies
ˆ ˆ  n−2
n
ˆ 2/n
2n n
2 2
|cv| ≤ |v| n−2 (|c| ) 2 = kvk2 2n kck2Ln
L n−2

Putting this together yields

ˆ
c(x)uv ≤ kukL2 kvk 2n kckLn
L n−2

≤ kukL2 kDvkL2 kckLn

 1
≤ kDvk2L2 + kuk2L2 kck2Ln
2 2
and now the first term gets absorbed in (29). The second term in (30) can be bounded as follows
ˆ
f v ≤ kf k n+2
2n kvk 2n
n−2
BR L L

≤ kf k 2n kDvkL2
L n+2
 1
≤ kDvk2L2 + kf k2 2n .
2 2 L n+2

Once again the Dv term gets absorbed in (29). Putting everything together gives us the following estimate
ˆ (
 r n ˆ
)
2
|Du| ≤ C + τ (R) |Du| + Ckuk2L2 (BR ) kck2Ln (BR ) + Ckf k2 2n .
2
(31)
Br (x0 ) R BR (x0 ) L n+2

Since we want to show u ∈ C α , we look for things of the form

ˆ
|Du|2 ≤ M rn+2−2α .
Br (x0 )

Now recall the Lemma of De Giorgi: ϕ(r) ↓ 0 as r ↓ 0, ϕ(r) ≥ 0. Assume that ∀0 < r < R, β < α and
 r α 
ϕ(r) ≤ A +  ϕ(R) + BRβ .
R
Then ∀0 < β < γ < α, ∃0 > 0 such that  < 0 implies
 r γ
ϕ(r) ≤ C ϕ(R) + Brβ .
R

29
In particular we can let R = 1 above and have

ϕ(r) ≤ C̃rβ .

The analysis tells us that in order to apply the lemma, we need to estimate the additional terms in (31) by
Rγ for γ = n − 2 + 2α. Lets begin with the f term

ˆ ! n+2
n ˆ ? ˆ ? ! n+2
n
2n
2 q
kf k 2n = |f | n+2 ≤ |f | 1 .
L n+2 (BR (x0 )) BR (x0 )

2n q(n+2) 1 1 2n
In order to figure out the exponents, we need n+2 p = q =⇒ p = 2n =⇒ m =1− p =1− (n+2)q . Thus

ˆ 1/p ˆ 1/m ! n+2

n
2 q
kf k 2n ≤ |f | 1
L n+2 (BR (x0 ))
  n+2
n
q/p
= kf kLq Rn/m
n+2
= kf k2Lq R m = kf k2Lq Rn−2+2α

where α = 2 − nq . Now consider the other term in (31):

ˆ
kuk2L2 (BR (x0 )) = |u|2
BR (x0 )
ˆ  n−2
n
ˆ 2/n
2n
≤ |u| n−2 1
BR BR
2
≤ kuk 2n (Rn )2/n
L n−2
≤ kuk2W 1,2 (BR ) R2 .

In the end the lower ordered terms are all bounded by

Rmin{n−2+2α,2} (kuk2W 1,2 (BR ) + kf k2Lq )

| {z }
B

Then (31) and the De Giorgi Lemma tell us

ˆ  r γ ˆ
|Du|2 ≤ C |Du|2 + Brmin{n−2+2α,2}
Br (x0 ) R BR (x0 )

≤ Crmin{n−2+2α,2}

where again we have used the trick of letting R = 1.

Now there are two cases. If min{n − 2 + 2α, 2} = n − 2 + 2α, we are done by Lemma 3.3. If it is not,
consider the integral of |u|2 over Br (x0 ). I claim that
ˆ  r n ˆ
2
|u| ≤ |u|2 + CR2+min{n−2+2α,2} .
Br (x0 ) R Br (x0 )

If I manage to show this, observe that we are done by the De Giorgi Lemma. Consider the triangle type

30
inequality |u|2 ≤ |u − ūR (x0 )|2 + ū2R and integrate over the ball of radius r to see
ˆ ˆ ˆ
|u|2 ≤ |u − ūR (x0 )|2 + ū2R
Br (x0 ) B (x ) B (x )
ˆ r 0 ˆ r 0
2
≤ |u − ūR (x0 )| + ū2R (r ≤ R)
BR (x0 ) Br (x0 )
ˆ ˆ
≤ R2 |Du|2 + ū2R (Poincaré)
BR (x0 ) Br (x0 )
ˆ
2+min{n−2+2α,2}
≤R + ū2R
Br (x0 )
ˆ ˆ !2
2+min{n−2+2α,2} 1
=R + u
Br (x0 ) Rn BR (x0 )

ˆ ˆ ˆ !1/2 2
 !1/2
≤ R2+min{n−2+2α,2} +  1 u2 1 
n
Br (x0 ) R BR (x0 ) BR (x0 )

ˆ ˆ !1/2 2


= R2+min{n−2+2α,2} +  1 u2 
Br (x0 ) Rn/2 BR (x0 )
ˆ
2+min{n−2+2α,2} rn
=R + n u2 .
R BR (x0 )

3.2 Harnack Inequality

The Harnack Inequality is essential in proving Regularity estimates in a different way as the previous section.
We follow the proof of presented in Gilbarg and Trudinger very closely. We begin by defining our differential
operator
Lu = Di (aij Dj u + bi u) + ci Di u + du
and force boundedness of the coefficients on L:
X X 
|aij (x)|2 ≤ Λ2 , λ−2 |bi (x)|2 + |ci (x)|2 + λ−1 |d(x)| ≤ ν 2 .

We also make this an elliptic equation by enforcing

aij (x)ξi ξj ≥ λ|ξ|2 , ∀x ∈ Ω, ξ ∈ Rn .

Then u is a weak solution to Lu = 0(≥ 0, ≤ 0) in the domain Ω if

ˆ
 ij
a Dj u + bi u Di v − (ci Di u + du)v dx = 0(≤ 0, ≥ 0)


Ω

for all non-negative functions v ∈ C01 (Ω). Let f i , g be locally integrable functions in Ω. Then u is a weak
solution of the inhomogeneous equation
Lu = g + Di f i
in Ω if it satisfies
ˆ ˆ
aij Dj u + bi u Di v − (ci Di u + du)v dx = f i Di v − gv dx



Ω Ω

for all v ∈ C01 (Ω).

31
3.2.1 Structural Inequalities
We rewrite Lu = g + Di f i as
Di Ai (x, u, Du) + B(x, u, Du) = 0, (32)
where

Ai (x, z, p) = aij pj + bi z − f i
B(x, z, p) = ci pi + dz − g

for (x, z, p) ∈ Ω × R × Rn . Then we say that u is a weak subsolution (supersolution, solution) of (32) in Ω
if Ai (x, u, Du) and B(x, u, Du) are locally integrable and
ˆ
Di vAi (x, u, Du) − vB(x, u, Du) dx ≤ (≥, =)0


(33)
Ω

for all non-negative v ∈ C01 (Ω). Writing b = (b1 , . . . , bn ), c = (c1 , . . . , cn ), f = (f 1 , . . . , f n ) and using the
Schwarz inequality, we have the inequalities
λ 2 1
pi Ai (x, z, p) ≥ |p| − (|bz|2 + |f |2 )
2 λ
|B(x, z, p)| ≤ |c||p| + |dz| + |g|.

We can consolidate this further by writing

z̄ = |z| + k, b̄ = λ−2 (|b|2 + |c|2 + k −2 |f |2 ) + λ−1 (|d| + k −1 |g|)

for some k > 0. Then for an 0 <  < 1, we finally have the following inequalities
λ
pi Ai (x, z, p) ≥ (|p|2 − 2b̄z̄ 2 )
2 
λ 2 b̄ 2
|z̄B(x, z, p)| ≤ |p| + z̄ .
2 

If we denote aij by a, then we can write |A(x, z, p)| ≤ |a||p| + |bz| + |f |. Also note that we can divide (32)
by λ2 to finally get the structural inequalities

|A(x, z, p)| ≤ |a||p| + 2b̄1/2 z̄

p · A ≥ |p|2 − 2b̄z̄ 2 (34)

1
|z̄B(x, z, p)| ≤ |p|2 + b̄z̄ 2

for some  ∈ (0, 1], z̄ = |z| + k, b̄ = 41 (|b|2 + |c|2 + k −2 |f |) + 12 (|d| + k −1 |g|). For the purposes of the following
proof, we will let k = k(R) = 12 (Rδ kf kq + R2δ kgkq/2 ), for δ = 1 − n/q.

3.2.2 Moser Iteration

Before proving The Harnack Inequality, we must prove two theorems first.
Theorem 3.5. Let L be uniformly elliptic with bounded coefficients. f i ∈ Lq (Ω), g ∈ Lq/2 (Ω) for q > n. Let
u ∈ W 1,2 (Ω) be a subsolution in Ω. Then for any B2R (y) ⊂ Ω, and p > 1, we have
 
sup u ≤ C R−n/p ku+ kLp (B2R (y)) + k(R) . (35)
BR (y)

32
Theorem 3.6. Let L be uniformly elliptic with bounded coefficients and suppose that f i ∈ Lq (Ω), g ∈ Lq/2 (Ω)
for q > n. Let u ∈ W 1,2 (Ω) be a supersolution in Ω. If u is non-negative in B4R (y) ⊂ Ω and 1 ≤ p <
n/(n − 2), we have
 
−n/p
R kukLp (B2R (y)) ≤ C inf u + k(R) . (36)
BR (y)

Proof. It is convenient to prove these two theorems conjointly in the case where u is a bounded non-negative
subsolution. We begin by assuming that R = 1, k > 0. The general case will be obtained from transforming
x 7→ x/R and letting k → 0. Let β 6= 0, η ∈ C01 (B4 ) be non-negative. We define v := η 2 ūβ . Recall that
ū = u + k. Then we have that
Dv = 2ηDηūβ + βη 2 ūβ−1 Du.
Note that v is a valid test function. Then we plug this v into our definition of subsolutions
ˆ
Di vAi (x, u, Du) − vB(x, u, Du) dx ≤ 0


Ω

and obtain
ˆ ˆ
β 2 β−1
η 2 ūβ B(x, u, Du) dx.


2ηDηū A(x, u, Du) + βη ū DuA(x, u, Du) dx ≤ (37)
Ω Ω

We will now attempt to apply our structural inequalities (34) into (37):

η 2 ūβ−1 DuA(x, u, Du) ≥ η 2 ūβ−1 |Du|2 − 2η 2 b̄ūβ+1

|η 2 ūβ B(x, u, Du)| = η 2 ūβ−1 |ūB(x, u, du)|
 
2 β−1 2 1 2
≤ η ū |Du| + b̄ū

1
= η 2 |Du|2 uβ−1 + b̄η 2 ūβ+1 .

The last structural inequality gives us this:

|ηDη · A(x, u, Du)ūβ | ≤ |a|η|Dη|ūβ |Du| + 2b̄1/2 η|Dη|ūβ+1

β−1 β+1
= |a|η|Dη|ū 2 ū 2 |Du| + 2b̄1/2 η|Dη|ūβ+1
 2 β−1 |a|2
≤ η ū |Du|2 + |Dη|2 ūβ+1 + |Dη|2 ūβ+1 + b̄η 2 ūβ+1
2 2
|a|2
 
 2 β−1 2
= η ū |Du| + 1 + |Dη|2 ūβ+1 + b̄η 2 ūβ+1 .
2 2

Applying this to (37) yields

ˆ ˆ 
2 β−1 2 β+1 1
η 2 |Du|2 ūβ−1 + b̄η 2 ūβ+1 + η 2 ūβ−1 |Du|2


βη ū − 2b̄βη ū dx ≤
Ω Ω 
|a|2
  
2 β+1 2 β+1
+ 2+ |Dη| ū + 2b̄η ū dx


Combining like terms yields

ˆ ˆ 
|a|2
   
1
(β − 2)η 2 |Du|2 ūβ−1 dx ≤ 2β + 2 + b̄η 2 + 2 + |Dη|2 ūβ+1 dx.
Ω Ω  

33
 n o
Let  = min 1, β4 . Then we can further consolidate this into
ˆ ˆ
η 2 |Du|2 ūβ−1 dx ≤ C(β) b̄η 2 + 1 + |a|2 |Dη|2 ūβ+1 dx.



(38)
Ω Ω

We introduce the function 

ū(β+1)/2 if β 6= −1
w :=
log ū if β = −1
so that
η β+1
 (β−1)/2
2 ū |Du| if β =
6 −1
|ηDw| =
η|Du|ū−1 if β = −1
and let γ = β + 1. Then we can rewrite (38) as
ˆ ( ´


2
2 ´
C(|β|)γ b̄η 2 + 1 + |a|2 |Dη|2 w2 dx if β 6= −1
|ηDw| dx ≤ Ω

(39)
Ω C Ω b̄η 2 + 1 + |a|2 |Dη|2 dx if β = 1.

Before we move any further, we need to introduce a few results from analysis of Sobolev spaces.
Lemma 3.7 (Interpolation Inequality). Let p ≤ q ≤ r. Then for u ∈ Lr (Ω), we have

kukq ≤ kukr + −µ kukp ,

where
1 1
p − q
µ= 1 1 .
q − r

From the Sobolev Inequality, we have

ˆ
kηwk22n̂/(n̂−2) ≤ C |Dηw|2 + |ηDw|2 dx,


Ω

for n̂ = n for n > 2, and 2 < 2̂ < q. Now we apply Hölder’s Inequality and the Interpolation Inequality to
get
ˆ
b̄(ηw)2 dx ≤ kb̄kq/2 kηwk22q/(q−2)
Ω

2
≤ kb̄kq/2 kηwk2n̂/(n̂−2) + −σ kηwk2 ,

where σ = n̂/(q
´ − n̂). Now we attempt to plug in these estimates into (39). Let ξ = n̂/(n̂ − 2). Now we add
a factor of Ω |wDη|2 dx and carry out some computations to get the following

34
 ˆ
kηwk2χ ≤ Cγ 2 b̄(ηw)2 + |wDη|2 + |a|2 |wDη|2 dx


Ω
ˆ
2 −σ

2 2
≤ Cγ kb̄kq/2 kηwk2ξ +  kηwk2 + Cγ (1 + |a|2 )|wDη|2 dx
ˆ Ω

2
= Cγ 2 kηwk2χ + −σ kηwk2 + Cγ 2 (1 + |a|2 )|wDη|2 dx
Ω
≤ Cγ 2 2 kηwk22χ + 1−σ kηwk2χ kηwk2 + −2σ kηwk22 + kwDηk22

kηwk22χ (1 − Cγ 2 2 ) ≤ Cγ 2 1−σ kηwk2χ kηwk2 + −2σ kηwk22 + kwDηk22

 1−σ
1−σ

2  2 2 −2σ 2 2
≤ Cγ kηwk2χ + kηwk2 +  kηwk2 + kwDηk2
2 2
1−σ 1−σ
    
kηwk22χ 1 − Cγ 2  + ≤ Cγ 2 + −2σ kηwk22 + kwDηk22


2 2
 1−σ 
Cγ 2  2 + −2σ
kηwk22χ ≤  kηwk22 + kwDηk22



1−σ
1 − Cγ 2  2 + −2σ
≤ C(1 + |γ|)σ+1 k(η + |Dη|)wk22

Then we finally get

kηwk2χ ≤ C(1 + |γ|)σ+1 k(η + |Dη|)wk2 , (40)

where C = C(n̂, Λ, ν, q, |β|) is bounded when |β| is bounded away from zero. We will now get a better cutoff
function η. Let r1 , r2 be such that 1 ≤ r1 < r2 ≤ 3, η ≡ 1 in Br1 , β ≡ 0 in Ω\Br2 , with
2
|Dη| ≤ .
r2 − r2
Then we have from (40)

C(1 + |γ|)σ+1
kwkL2χ (Br1 ) ≤ kwkL2 (Br2 ) . (41)
r2 − r1
Before we move on, lets make a quick backtrack to functional spaces. Let Ω ⊂ Rn be a bounded domain.
Then I claim that if u is a measurable function on Ω such that |u|p ∈ L1 (Ω) for some p ∈ R, and define
 ˆ 1/p
1
φp (u) := |u|p dx ,
|Ω| Ω

then we have

lim φp (u) = sup |u|. (42)

p→∞ Ω

To start, note that

 ˆ 1/p  ˆ  p 1/p
1 p 1
φp (u) = |u| dx ≤ sup |u| dx = sup |u|.
|Ω| Ω |Ω| Ω Ω Ω

Taking the lim sup yields

lim sup φp (u) ≤ sup |u|.
p→∞ Ω

35
Now fix  and define A = {x ∈ Ω||u| ≥ supΩ |u| − }. Then we see that
 ˆ 1/p  ˆ 1/p
1 1
φp (u) = |u|p dx ≥ |u|p dx
|Ω| Ω |Ω| A
 ˆ  p 1/p
1
≥ sup |u| −  dx
|Ω| A Ω
 
|A|
= sup |u| −  .
|Ω| Ω

Taking  → 0 and then the lim inf yields

lim inf φp (u) ≥ sup |u|.
p→∞ Ω

This and our first inequality yield the claim (42). We go back to our proof. For r < 4, we define the function
ˆ 1/p
φ(p, r) := |ū|p dx .
Br

From what we just showed, we have that

φ(∞, r) = lim φ(p, r) = sup ū.
p→∞ Br

Now we can rewrite inequality (42) as

2/|γ|
C(1 + |γ|)σ+1

φ(χγ, r1 ) ≤ φ(γ, r2 ) if γ > 0 (43)
r2 − r1
2/|γ|
C(1 + |γ|)σ+1

φ(γ, r2 ) ≤ φ(χγ, r1 ) if γ < 0 (44)
r2 − r1
Now we are in the right position to start our iteration. Recall that when u is a subsolution, we have β > 0
and γ > 1. Taking p > 1 and setting γ = γm = χm p and rm = 1 + 2−m . Then plugging this into our
inequality (43) gives us
2/(χm−1 p)
C(1 + |χm−1 p|)1+σ

m
φ(χ p, 1) ≤ φ(χm−1 p, r2 )
r2 − 1
2/(χm−2 p)  2/(χm−1 p)
C(1 + |χm−2 p|)1+σ C(1 + |χm−1 p|)1+σ

≤ φ(χm−2 , r4 ) ≤ · · ·
r3 − r2 r2 − 1
mχ−m
P
≤ (Cχ)2(1+σ) φ(p, 2) = Cφ(p, 2).
Letting m → ∞ in the above yields
sup |u| ≤ CkūkLp (Br2 ) , C = C(n̂, Λ, ν, q, p).
B1

Transforming x 7→ x/R we have the desired estimate (35) if u is a subsolution. Now in the cases when u is
a supersolution, we need to approach the problem a bit differently. Recall that when u is a supersolution,
β < 0 and γ < 1. Then for any p, p0 such that 0 < p0 < p < χ, we have
φ(p, 2) ≤ Cφ(p0 , 3)
φ(−p0 , 3) ≤ Cφ(−∞, 1).
Then we will finish proving our theorem if we can show that φ(p0 , 3) ≤ C(−p0 , 3). This is done in an intricate
way and is left as an exercise.

36
 Putting our two theorems together give us the full Harnack Inequality:
Theorem 3.8. Let L be uniformly elliptic and have bounded coefficients. Let u ∈ W 1,2 (Ω) satisfy u ≥ 0 in
Ω and Lu = 0 in Ω. Then for any ball B4R (y) ⊂ Ω, we have

sup u ≤ C inf u,
BR (y) BR (y)

where C = C(n, Λ/λ, νR).

3.2.3 Applications of the Harnack Inequality

We can give a new proof of the strong maximum principle (instead of using Hopf’s Lemma) now:
Theorem 3.9. L uniformly elliptic, bounded coeff, u ∈ W 1,2 (Ω), Lu ≥ 0 in Ω. Then if for some ball B ⊂⊂ Ω
we have
sup u = sup u ≥ 0,
B Ω

then the function u must be constant in Ω.

Proof. We apply the Harnack inequality with p = 1 to the function v = M − u and show that M − u = 0.

Finally, we can use the Harnack Inequality to imply Hölder continuity.

Theorem 3.10. Let L be uniformly elliptic and have bounded coefficients. Then if u ∈ W 1,2 (Ω) is a solution
of
Lu = g + Di f i ,
then u is locally Hölder continuous in Ω, and for any ball B0 = BR0 (y) ⊂ Ω and R ≤ R0 we have

oscBR (y) u ≤ CRα (R0−α sup |u| + k).

B0

Proof. Start out by defining M0 , M1 , M4 , m1 , m4 . Then apply Harnack inequality with p = 1 to M4 − u and
u − m4 to end up with
ω(R) ≤ γω(4R) + k(R).
Here ω(R) = oscBR u. Before we go on, we have to prove something else first. Suppose ω is non-decreasing
on (0, R0 ], R ≤ R0 satisfies
ω(τ R≤ γω(R) + σ(R),
where σ is also non-decreasing, 0 < γ, τ, < 1. Then for any µ ∈ (0, 1), we have
 α 
R µ 1−µ
ω(R) ≤ C ω(R)0 + σ(R R0 ) .
R0

Of course the De Giorgi Lemma concludes the proof.

37
4 Harnack Inequality for Non-Divergence Equations
The formulation of weak solutions to divergence equations relied heavily on the fact that the operator L was
in divergence form. This allowed us to integrate by parts and so a weak solution u needs to be once weakly
differentiable (W 1,2 ). A classical solution u must be at least second order continuously differentiable. In this
section we will concern ourselves with the intermediate situation of strong solutions.
Definition 4.1. For operators of the form

Lu = aij uij + bi ui + c(x)u (45)

with coefficients aij , bi , c(x) defined on a domain Ω ⊂ Rn and a function f on Ω, a strong solution of Lu = f
is a function u ∈ W 2,p (Ω) that satisfies (45) almost everywhere. See [1][p. 185] for existence and uniqueness
of equations of this type.

4.1 ABP Estimate

We begin by proving a maximum principle for strong solutions analogous to the classical one for classical
solutions. It is called the Alexandrov-Bakelmann-Pucci Maximum Principle. In particular, we will prove this
2,n
maximum principle for solutions in the space Wloc (Ω) as it is the natural environment for such equations.
Recall that an operator L in the form (45) is said to be elliptic in the domain Ω ⊂ Rn if the matrix (aij ) is
positive definite everywhere in Ω. Let D := det(aij ) and D ? = D 1/n so that

0 < λ ≤ D? ≤ Λ

where λ, Λ are the minimum and maximum eigenvalues of (aij ). Assume b = c = 0 so that we have
Lu = aij uij . Our condition on aij and f are now

f /D ? ∈ Ln (Ω).
2,n
Theorem 4.2 (ABP Maximum Principle). Let Lu ≥ f in a bounded domain Ω and u ∈ C 0 (Ω) ∩ Wloc (Ω).
Then
d f
sup u ≤ sup + 1/n
Ω ∂Ω nωn D ? Ln (Γ+ )
where d = diam Ω.
2,n
It is important to note that the Morrey’s embedding theorem guarantees that u ∈ Wloc (Ω) will be at
least continuous in Ω because whenever kp > n, W k,p
is embedded in C for 0 ≤ α < k − np . Before we
α

begin the proof, we must first go through notions of contact sets and normal mappings.
Definition 4.3. Suppose u is an arbitrary function on Ω. The upper contact set Γ+ or Γ+
u is defined to be
the subset of Ω such that the graph of u is below a support hyperplane in Rn+1 , i.e.

Γ+ = {y ∈ Ω|u(x) ≤ p(x − y) + u(y) ∀x ∈ Ω, for some p ∈ Rn }.

From the definition, u will be a concave function on Ω if and only if Γ+ = Ω. It is clear that p = Du(y) if
u ∈ C 1 (Ω). Finally, if u ∈ C 2 (Ω), the Hessian D2 u ≤ 0 on Γ+ . This means that we can essentially think of
Γ+ as the subset of Ω where u is concave down. The upper contact set of u is closed in Ω.
Definition 4.4. Suppose u ∈ C 0 (Ω) is arbitrary. We define the normal mapping χ(y) = χu (y) for a point
y ∈ Ω to be the set of slopes of supporting hyperplanes at y lying above the graph of u, i.e.

χ(y) = {p ∈ Rn |u(x) ≤ p(x − y) + u(y) ∀x ∈ Ω}.

Clearly χ(y) is nonempty if and only if y ∈ Γ+ . When u ∈ C 1 (Ω), then χy = Du(y).

38
Proof of Theorem 4.2. Assume that u ∈ C 0 (Ω) ∩ C 2 (Ω). Subtracting sup∂Ω u from u yields the same differ-
ential inequality L(u − sup∂Ω ) ≥ f and so we can assume u ≤ 0 on the boundary. Note that we can assume
supΩ u ≥ 0 because if it was negative, there would be nothing to prove. This assumption implies that if
u(y) := supΩ u, then y is in the interior of Ω because u ≤ 0 on the boundary. Now I show that

Bu(y)/d ⊂ χ(Γ) = Du(Γ). (46)

This can be seen by sliding hyperplanes onto the graph of u. Consider the cone with vertex u(y) with base
∂Ω. Then the slope of the cone is u(y)/d and u ≤ 0 on ∂Ω implies that dropping down hyperplanes of slope
u(y)/d will eventually be tangent to u(x) at a point x ∈ Γ. This can be seen very clearly in pictures:

It is a bit difficult to see, but in the picture, we have u|∂Ω≤0 . Now we compute
ˆ
|Du(Γ)| = 1
Du(Γ)
ˆ
≤ | det D2 u|. (47)
Γ

Lets do a bit of linear algebra. I claim that given two positive matrices A, B, then
 n
Tr AB
det A det B ≤ .
n

To see this it suffices to show that  n

Tr A
det A ≤ .
n
Let λ1 , . . . , λn be A’s eigenvalues. Then
n
Y n
X
det A = λi , Tr A = λi .
i=1 i=1

Thus what we wish to show is equivalent to

Y 1/n 1X
λi ≤ λi ,
n

39
which is exactly the inequality of arithmetic and geometric means. Taking A = D2 u,B = (aij ) then on Γ
1
| det D2 u| = det −D2 u = det(aij ) det −D2 u
 n D
1 f
≤ − . (48)
D n

Combining (46),(47), and (48) yields

 n
u(y) 1
ωn ≤ kf − knLn (Γ)
d nn D
d f−
u(y) ≤ .
nω 1/n D? Ln (Γ)

which is precisely the ABP estimate because we had replaced u with u − sup∂Ω u.
The ABP maximum principle can be naturally extended to functions u ∈ C 0 (Ω) ∩ Wloc 2.n
(Ω) by approxi-
mating with smooth functions. It can also be extended for coefficients satisfying |b|/D ∈ Ln (Ω) and c ≤ 0
?

in Ω, but these will be taken as a black box. Now we have the correct tools to begin proving the Harnack
inequality for non-divergence equations. We say an operator of the form (45) is strictly elliptic when
 2
Λ |b| |c|
≤ γ, , ≤ ν.
λ λ λ

We will assume throughout the rest of this section that the operator L given by (45) is uniformly elliptic.
Note that have the same ABP estimate from the other side: if Lu ≤ f and u|∂B1 ≥ 0, then
ˆ 1/n
n
| inf u| ≤ Cn f
B1 Γ

where Γ is the convex envelop of u (contrary to the upper contact set Γ+ )

4.2 Measure theory and Harnack Inequality

In order to prove the Harnack inequality for non-divergence equations with measurable coefficients, we need
a basic measure theoretic estimate that is true at “all scales.” Our first main theorem is as follows

Theorem 4.5. Let u ∈ W 2,n and assume aij uij ≤ 0 for some bounded measurable and uniformly elliptic
aij in B2 . If u ≥ 0 and u(x) ≤ 1 for some x ∈ ∂B1 , then

|{u ≤ M } ∩ B1/2 | ≥ µ (49)

for some M large and µ small universal.

Heuristically, this is saying that if u goes below 1 at a certain point on the boundary of a ball of radius
one, then it can’t get much bigger in a ball of radius two. That is, the ABP estimate will tell us that the
size of u can be controlled.
Proof. Define α = max{1, (n − 1)Λ/λ − 1} and take a barrier function ϕ of the form ϕ(x) = M1 − M2 |x|α
in B2 − B1/2 with M1 and M2 chosen such that ϕ∂B2 = 0 and ϕ|∂B1 = −2. Note that we can cap off in
B1/2 smoothly so that the constants still depend only on n, λ, Λ (we say cap off to smooth out the inner

40
radius with a paraboloid of radius 2; it’ll be made clear with the picture). Suppose that r ≥ 1/2 at the point
(r, 0, . . . , 0). Then we compute some derivatives at this point:

Dij ϕ = 0 for i 6= j
D11 ϕ = −M2 α(1 + α)r−α−2 r−α−2
Dii ϕ = M2 αr−α−2 .

By rotational symmetry of our function and uniform ellipticity, we have for |x| ≥ 1/4,

aij ϕij |x = M2 (Λ(n − 1)α|x|−α−2 − λα(1 + α)|x|−α−2 )

= M2 α|x|−α−2 (Λ(n − 1) − λ(1 + α)) ≤ 0 (50)

by our choice of α. However, for |x| ≤ 1/4, aij uij ≤ C = C(n, λ, Λ). This observation along with (50) shows
that aij ϕij ≤ Cη for some C universal with η ≡ 1 in B1/4 and η ≡ 0 outside of B1/2 , i.e. η ∈ C0∞ (it is
smooth because of the smoothness of |x|α in this domain). Finally note that |ϕ| ≤ M for some M universal
given by the distance of the vertex of the parabola that we capped off by and the origin.

By our assumption that u ≤ 1 somewhere on ∂B1 we have that w = u + ϕ ≤ −1 somewhere on ∂B1 .

Then ABP given gives us
ˆ
1≤C η n ≤ C|{Γw } ∩ B1/2 |.
Γw ∩B1 }

Now the crucial part of the proof is to notice that the convex envelop of w is in the set that w ≤ 0 =⇒
u + ϕ ≤ 0 =⇒ u ≤ −ϕ ≤ M . Hence our estimate just gave us 1 ≤ C|{u ≤ M } ∩ B1/2 |. This concludes the
proof with µ = 1/C.
Corollary 4.6. We can show that this is scaling invariant. That is, if u is defined in B2r and u ≤ α
somewhere on ∂Br , then |{u ≤ M α} ∩ Br/2 | ≥ µ.

41
Lemma 4.7. Suppose we have the same u as in Theorem 3.5. Then

|{u ≥ M k } ∩ B1/2 | ≤ (1 − µ)k

for k = 1, 2, . . . , and M, µ are as in Theorem 3.5.

This implies that
|{u ≥ t} ∩ Q1 | ≤ dt− (51)
for any t > 0 where d,  are positive universal constants.

Proof. This proof will follow from Calderon-Zygmund decomposition and induction. For k = 1, this is the
statement of Theorem 4.5. Now suppose it holds for k − 1. We introduce the classical decomposition: if Q is
a dyadic cube different from Q1 , we say Q̃ is the predecessor of Q if Q is one of the 2n cubes obtained from
dividing Q̃. Then the decomposition states the following: if A ⊂ B ⊂ Q1 are measurable sets and 0 < δ < 1
such that

(1) |A| ≤ δ, and

(2) if Q is a dyadic cube such that |A ∩ Q| > δ|Q|, then Q̃ ⊂ B,
then |A| ≤ δ|B|.
Recall that we assume the Lemma holds for k − 1. For convenience, we use cubes because Theorem 4.5
still holds for them. Let A = {u > M k } ∩ Q1 and B = {u > M } ∩ Q1 . We’ll be done with our lemma if we
show that |A| ≤ (1 − µ)|B| by our inductive hypothesis. This is where we’ll use the decomposition. Clearly
A ⊂ B ⊂ Q1 and |a| ≤ |{u > M } ∩ Q1 | ≤ 1 − µ by Theorem 4.5. Now we need to show (2) also holds.
(2) will follow if we show that if Q = Q1/2i (x0 ) is a dyadic cube such that |A ∩ Q| > (1 − µ)|B|, then
Q̃ ⊂ B. Suppose it isn’t and take x̃ ∈ Q̃ such that u(x̃) ≤ M k−1 . Now consider the transformation
x = x0 + 1/2i y for y ∈ Q1 and x ∈ Q = Q1/2i (x0 ) and the function

ũ(y) = u(x)/M k−1 .

Then I claim that ũ is under the hypothesis of Theorem 4.5. By (49), it follows that

µ < |{ũ(y) ≤ M } ∩ Q1 | = 2in |{u(x) ≤ M k } ∩ Q|.

We see that |Q − A| > µ|Q|, which contradicts our assumption that |A ∩ Q| > (1 − µ)|Q|. Showing that ũ
satisfies the conditions of Theorem 4.5 will be left as an exercise. As a hint, the following property of dyadic
cubes: if Q is a dyadic cube Q = Q1/2i (x0 ) for some i ≥ 0 and x0 ∈ Q1 , then

Q4√n/2i (x0 ) ⊂ Q4√n ,

i ≥ 1 =⇒ Q̃ ⊂ Q3/2i (x0 ).
(51) follows from the results above by taking d = (1 − µ)−1 and  such that 1 − µ = M − .

Now we need yet another measure theoretic lemma. Let f be a measurable function on a domain Ω in
Rn . Define the distribution function µ(t) = |{f ≥ t}| for t > 0. This measures the “relative” size of f . Note
that µ is a decreasing function on the positive real line.
Lemma 4.8. For any p > 0 and |f |p ∈ L1 (Ω),
ˆ ˆ ∞
p
|f | = p tp−1 µ(t)dt.
Ω 0

42
Proof. This is just computations. Suppose f ∈ L1 . Then
ˆ ˆ ˆ |f (x)|
|f | = dtdx
Ω
ˆΩ∞ 0
= µ(t)
0

and the lemma will hold for general p after change of variables.

Now we are finally able to prove the weak Harnack inequality:

Theorem 4.9 (Weak Harnack Inequality). Suppose u ∈ W 2,n (Q1 ) satisfies aij uij ≤ 0 in Q1 and u is
non-negative. Then
kukLp (Q1/4 ) ≤ C inf u
Q1/2

where p > 0 and C universal.

Proof. We already have most of the machinery for the proof. This follows from (51) in Lemma 4.7 because
if we suppose u is under the assumptions of the emma, |{u ≥ t} ∩ Q1 | ≤ dt− and Lemma 4.8 implies for
p = /2
ˆ ˆ ∞
p
u =p tp−1 |{u ≥ t} ∩ Q1 |dt
Q1 0
ˆ ∞
≤p tp−1 · d · t− dt
0
ˆ ∞
=d·p tp−1 t−2p dt (inf u ≥ 1)
ˆ1 ∞
=d·p t−p−1 dt
1
∞
t−p
=d·p
p 0
= d.

hence kukLp (Q1 ) ≤ C and we re-scale u away from the assumptions of 4.7 to get the result of the weak HI.
The second piece of HI is called the local maximum principle

Theorem 4.10. Let u ∈ W 2,n (Q1 ) and let aij uij ≥ 0. Then for any 0 < p ≤ n, we have

sup u ≤ Cp ku+ kLp (Q3/4 ) .

Q1/2

for C universal depending on p.

Proof. See [1][p. 244].

Corollary 4.11. Taking p = 1 in Theorem 4.10, we obtain an extension of the mean value inequality for
non-negative subharmonic functions: ˆ
C
u(y) ≤ n u
R BR
when Lu ≥ 0, C = C(n, γ, νR2 ).

43
5 Curved C 1,α Domains
5.1 Estimate for Laplacian
We begin with a simple definition.

Definition 5.1. A continuous function u on Rn is said to be C 2,α at x0 if there exists a quadratic polynomial
P and constants C, ρ such that
ku − P kL∞ (Br ) ≤ Cr2+α , ∀r ≤ ρ. (52)
We say u ∈ C 2,α (B1 ) if u is C 2,α at all x ∈ B1 .
Now we develop a tool that helps us determine when a function is C 2,α .
Proposition 5.2. Let Ω ⊂ Rn be open and bounded. Suppose we can find a sequence of paraboloids Pk =
ak + bk · x + 21 xT ck x and an r < 1 such that

ku − Pk kL∞ (Brk ) ≤ Crk(2+α)

where Brk ⊂⊂ Ω. Then u ∈ C 2,α (0) where 0 < α < 1.

1
Proof. Consider the quadratic scaling of a function fr (x) = r 2 f (rx). By our hypothesis we have

kPk+1 − Pk kL∞ (Brk+1 ) ≤ ku − Pk+1 kL∞ (Brk+1 ) + ku − Pk kL∞ (Brk+1 )

≤ Kr(k+1)(2+α) + Krk(2+α)
≤ 2Krk(2+α) .

Quadratic rescaling gives us

kPk+1,rk − Pk,rk kL∞ (B1 ) ≤ 2Krkα .
Since the coefficients of polynomials on B1 are controlled by the L∞ norm, we have

|ak+1 − ak | ≤ Krk(2+α) , |bk+1 − bk | ≤ Krk(α+1) , |ck+1 − ck | ≤ Krkα . (53)

Thus we have that the Pk converges to a polynomial P = a + bx + xT cx. Putting these together yield

ku − P kL∞ (Brk ) ≤ ku − Pk kL∞ (Brk ) + |ak − a| + rk |br − b| + r2k |ck − c|

≤ CKrk(2+α)

and so we have that u ∈ C 2,α .

It is important to realize why we can’t have α = 0, 1. Notice that for α = 0, the proof breaks down in
(53) because we’d get |ck+1 − ck | ≤ K, so the sequence of polynomials doesn’t converge. The case for α = 1
is left as an exercise. We utilize this proposition to prove the following theorem.

Theorem 5.3. Suppose ∆u = f in B1 where f ∈ C α (B1 ). Then u ∈ C 2,α (B1/2 ) and

kukC 2,α (B1/2 ) ≤ Cn,α kukL∞ (B1 ) + kf kC α (B1 ) .
1
Proof. By subtracting 2n f (0)|x|2 we assume f (0) = 0. Dividing by kukL∞ (B1 ) + 1 kf kC α (B1 ) for some  > 0
to be chosen later, we can assume that kf kC α (B1 ) ≤  and |u| ≤ 1. Let w be the harmonic function that
agrees with u on ∂B1 . Then ∆(u − w) = f and again kf kC α (B1 ) ≤ . By the maximum principle,

ku − wkL∞ (B1 ) ≤ C.

44
Let P1 be the harmonic quadratic approximation to w that satisfies

kw − P1 kL∞ (Br ) ≤ Cr3 .

The constant C that is above is in terms of n because |w| ≤ 1. Putting these together we have

ku − P1 kL∞ (Br ) ≤ C(r3 + ).

Now here we use the fact that 0 < α < 1 so choose r so small that 2Cr3 ≤ r2+α and choose  so small that
2C ≤ r2+α . This gives
ku − P1 kL∞ (Br ) ≤ r2+α .
1
Now take the 2 + α rescaling v(x) = r2+α (u − P1 )(rx). Notice that |v| ≤ 1 and since P1 is harmonic we
1 1
have that ∆v = rα ∆u(rx) = rα f (rx) = g(x). The right side again satisfies |g| ≤  so we repeat our process
again to find a harmonic polynomial P2 such that

kv − P2 kL∞ (Br ) ≤ r2+α

and so scaling back this will become

ku(x) − P1 (x) − r2+α P2 (x/r)kL∞ (Br2 ) ≤ r2(2+α) .

Iterating this will give us a sequence of polynomials approximating u. Thus we are in the position of
Proposition 1.2, and we are finished.
Lets shift gears a little bit and develop tools for boundary estimates. We say that a Ω is a Schauder
domain, or a C k,α – domain if it is a domain in Euclidean space with sufficiently regular boundary i.e., ∂Ω
can locally be viewed as the graph of a C k,α function. For this reason, we will assume that ∂Ω = {(x0 , g(x0 ))}
where g ∈ C 2,α , x0 ∈ Rn−1 . Recall that the first step in the iteration for Theorem 5.3 required us to obtain
the estimate |u − w| ≤ C in B1 . Since this sort of iteration process is all we can do at the moment, I will
now attempt to show the first step:
Lemma 5.4. Let B1 be the ball centered at the origin, f ∈ L∞ , and suppose that u solves

∆u = f in Ω ∩ B1
u = 0 on ∂Ω

with |u| ≤ 1 and |f | ≤ δ. Suppose the domain is as flat as

B1 ∩ {xn ≥ } ⊂ B1 ∩ Ω.

Then there is a harmonic function w in B1/4 with w(x0 , 0) = 0 so that

|u − w| ≤ C( + δ)

on B 14 (0) ∩ Ω.

Proof. We first consider the case where f = 0, so u is harmonic on Ω ∩ B1 . Let Γ(x) be the fundamental
solution to Laplace’s equation, and define the following barrier function as follows
Γ(x) − Γ(r)
G(x) :=
Γ(R) − Γ(r)

for 0 < r < R. Notice that 0 ≤ G, G(R) = 1, G(r) = 0, and that G is harmonic in the annulus of little
radius r and big radius R. Consider the lines xn = ± and let x0 = (x00 , −) ∈ Rn be so small that it is in
B1 . Lets look at G the barrier function with smaller radius tangent to xn = −, and see that it is centered
at (x00 , − − r) =: x̃0 .

45
 Now I will show that |u| ≤ G on the common domain of influence. First note that G ≥ |u| on ∂Ω because
u = 0 on ∂Ω. Recall that we assumed |u| ≤ 1 in Ω, and since G = 1 on ∂B(x̃0 , R), we see that |u| ≤ G on
Ω ∩ ∂B(x̃0 , R). Putting these two together yields |u| ≤ G on ∂ (Ω ∩ B(x̃0 , R)), and the maximum principle
(which applies because these two are harmonic on Ω ∩ B(x̃0 , R)) tells us that |u| ≤ G on the entire common
domain of influence.
Lets restrict ourselves for a moment on rays starting from x̃0 and moving in the xn direction. Notice
that since G is Lipschitz, we have

|G(0, . . . , 0, t) − G(0, . . . , 0, s)| ≤ C|t − s|.

Let t ∈ R be such that x̃0 + ten ∈ Ω. We now use the fact that G is zero on the smaller ball of radius r:

|u(x̃0 + ten )| ≤ G(x̃0 + ten )

= G(x̃0 + ten ) − G(x̃0 + ren )
≤ C|t − r| = C|xn + |.

Observe that x̃0 can be moved horizontally as long as the outer ball of radius R is in B1 , and the inner ball
of r is below −, and so,
|u| ≤ C|xn + | on all of B 41 (0) ∩ Ω (54)
Now that we have this estimate, lets find the corresponding harmonic polynomial that gives us the desired
result. Let w be the harmonic polynomial such that
+
 w = u on ∂(B1/4 ) ∩ Ω ∩ {xn ≥ }

+
w = 0 on ∂(B1/4 ) ∩ {0 ≤ xn < }

w = 0 on {xn = 0}.
+
There is the possibility that w is not continuous on ∂(B1/4 ) ∩ Ω ∩ {xn = } because on one side, w = u and
on the other w = 0. The way to fix this is the following: move δ > 0 above  and consider a C ∞ cutoff
function vδ that satisfies
 +
0 ≤ vδ ≤ 1 W ⊂⊂ ∂(B1/4 ) ∩ Ω ∩ { ≤ xn ≤  + δ}
vδ ≡ 1 V ⊂⊂ W

where vδ decreases monotonically to zero. Now define the new harmonic function
 +

 w=u on ∂(B1/4 ) ∩ Ω ∩ {xn ≥  + δ}
 w = v 2 u on ∂(B + ) ∩ Ω ∩ { ≤ x ≤  + δ}

δ 1/4 n
wδ := +

 w = 0 on ∂(B 1/4 ) ∩ {0 ≤ x n < }

w=0 on {xn = 0}.


Then since

46
 +
(i) |wδ | = |u| ≤ G on ∂(B1/4 ) ∩ Ω ∩ {xn ≥  + δ}
+
(ii) |wδ | = |vδ2 u| ≤ |u| ≤ G on ∂(B1/4 ) ∩ Ω ∩ { ≤ xn ≤  + δ}
+
(iii) G ≥ 0 = wδ on ∂(B1/4 ) ∩ {0 ≤ xn < }

(iv) G ≥ 0 = wδ on {xn = 0},

 
+
putting (i) - (iv) together yields |wδ | ≤ G on ∂ Ω ∩ B1/4 . Thus the maximum principle then implies

+
|wδ | ≤ C|xn + | on all of B1/4 ∩ Ω. (55)

− −
Similarly on B1/4 ∩ Ω, one could extend oddly by w̃δ (x0 , xn ) = −wδ (x0 , −xn ) on B1/4 , and comparing w̃ with
−
−u, one achieves the same bound as (55) on B1/4 ∩ Ω. Define the new oddly reflected function
(
+
wδ on B1/4
wodd = −
w̃δ on B1/4

which satisfies

|wodd | ≤ C|xn + | on all of B1/4 ∩ Ω. (56)

Putting (54) and (56) together finally proves the Lemma for u harmonic:

|u − wodd | ≤ C

on B1/4 ∩ Ω. Now lets consider the general case ∆u = f where f ∈ L∞ and |f | ≤ δ. Here, apply the above
δ
to ũ = u ± 2n |x|2 and comparison principles allow us to say

|u − wodd | ≤ C( + δ).

Corollary 5.5. Suppose u satisfies 

∆u = f in Ω
(57)
u ≤  on ∂Ω
with again the same flatness assumption on Ω (B1 ∩ {xn ≥ } ⊂ B1 ∩ Ω). Then the same estimate as Lemma
5.4 holds.

Proof. In order to see why this is true, we must examine where we used u = 0 on ∂Ω in the proof of Lemma
5.4. We were comparing trying to show that |u| ≤ G on the common domain of influence and used the fact
that u = 0 ≤ G on ∂Ω. In order to get the same estimates of the Lemma for (57), we simply move the point
x̃0 = (x00 , − − r) to x̃0 = x00 , − − r − 20 . The reason for this is that the barrier function G grows at the
rate
|DG(r)| ≤ Cr1−n ,
and so by moving down a sufficiently small amount, G will be bigger than , and in turn bigger than u on
∂Ω

We are now ready to prove C 1,α estimates at the boundary:

47
Theorem 5.6. Suppose u satisfies 
∆u = f in Ω
u = 0 on ∂Ω
with f (0) = g(0) = ∇g(0) = 0, |u| ≤ 1, |f | ≤ δ. Suppose Ω is a C 1,α domain, that is ∂Ω = {(x0 , g(x0 ))} for
g ∈ C 1,α that satisfies |g| ≤ δr1+α . Then u ∈ C 1,α (Br/2 ∩ Ω) with the estimate

kukC 1,α (Br/2 ∩Ω) ≤ Cn,α (kukL∞ + kf kL∞ + kgkC 1,α (|x0 |) )

Proof. Consider Br ∩ Ω we want to show that there exists a linear function l such that

ku − l|L∞ (Br ∩Ω) ≤ r1+α

and WLOG assume r ≤ 1. Then since u = 0 along the boundary, all the tangential derivatives are then zero
and hence we may reduce this to

ku − axn kL∞ (Br ∩Ω) ≤ r1+α (58)

and we may also assume |a| ≤ 1. In order to achieve this by some sort of iteration used in Theorem 5.3, we
need to conclude
ku − ãxn kL∞ (Bρr ∩Ω) ≤ (ρr)1+α

for some ã. The first thing we do is to re-scale the ball Br into B1 (and hence diluting Ω to Ω̃) and it is
clear that we need to define some ũ that satisfies

u(x) − axn = r1+α ũ(x/r)

because ũ’s domain of influence is B1 . Our hypothesis then becomes |ũ| ≤ 1 and note that we can rewrite it
as
u(rx) − a(rxn )
ũ(x) = .
r1+α
Observe that diluting Ω to Ω̃ makes the latter the graph of g̃(x0 ) = 1r g(rx̃0 ) which satisfies |g̃| ≤ δrα . Now we
2
need to figure out what equation ũ satisfies: ∆ũ = r ∆u(rx)
r 1+α = r1−α f (rx) =: f˜. The bound that f˜ satisfies
is |f˜| ≤ δr 1−α
≤ δ, where we used that r ≤ 1. Now lets figure out the boundary conditions that ũ satisfies:

u(rx) − a(rxn )
ũ|∂ Ω̃ =
r1+α ∂ Ω̃
|axn r| |axn |
= α+1 = α
r r
≤δ

where we used the fact that u(x) vanishes on ∂Ω =⇒ u(rx) vanishes on ∂ Ω̃. Putting everything together
gives us

|ũ| ≤ 1
|g̃| ≤ δrα (59)
∆ũ = f˜, |f˜| ≤ δ
|ũ| ≤ δ on ∂ Ω̃.

Now we are exactly in the position of Corollary 5.5 because the flatness condition is precisely met by (59)
(after replacing the  in the Lemma with δ because ∂ Ω̃ is the graph of g̃, and so its height in B1 is at most δ
by (59)). Applying the Lemma gives us |ũ − w| ≤ Cδ in B 14 ∩ Ω̃. Recall that w is the solution that vanishes

48
on x0 =⇒ |w − bxn |Bρ ≤ Cρ2 . Now noticing that the radius of B 41 doesn’t have to be 1
4 (as long as it is small
enough for the proof of Lemma 5.4 to hold) we reach
|ũ − bxn |Bρ∩Ω̃ ≤ C(δ + ρ2 ).

By picking δ and ρ small enough, and using the fact that 0 < α < 1 yields
|ũ − bxn |Bρ∩Ω̃ ≤ Cρ1+α .

Rescaling back gives us

|u − ãxn |Bρr∩Ω ≤ C(ρr)1+α
with ã = a + rα b. We can now iterate just like in the last step of Theorem 5.3 to find that u ∈ C 1,α .

5.2 General Elliptic Operators

The goal of this section is to generalize Krylov’s estimate for flat domains to general C 1,α domains. First
we state and prove his estimate
Theorem 5.7. Assume u ∈ W 2,n (B + )∩C 0 (B̄ + ) is non-negative and u(en /2) ≥ 1. If u satisfies the equation
aij uij = 0 in the half ball B + = B1 ∩ Rn+ with u = 0 on T = B1 ∩ ∂Rn+ , and the aij measurable bounded and
+
uniformly elliptic. Then u ∈ C 1,α (B1/2 ).
Proof. First we note that if u ≥ 1 at en /2, we can apply the Harnack inequality to find that u ≥ c on
B1/10 (en /2). Now look at the barrier function φ(x) = C1 |x − e2n |α − C1 with C1 and C2 chosen such that
φ ≡ 0 on ∂B1/2 (en /2) and φ = c on ∂B1/10 (en /2). Here α = α(n, Λ, λ). By an almost identical computation
to the barrier function constructed in the HI for non-divergence equations, one can see that aij φij ≥ 0 on
B1/2 (en /2) − B1/10 (en /2). Now we can use the ABP maximum principle to deduce that u ≥ φ on the
common domain of influence. Then by Hopf lemma we know that this barrier is going to start at an angle,
and so we can actually conclude that u(x) ≥ δ 0 xn .

49
 Note that this was just in this particular domain B1/2 (en ) but we want to establish the proof in all of
+
B1/2 . In order to remedy, we have to slide our barrier to the left and right by x0 such that outer ball where φ
vanishes is tangent to B + . The center of these two extreme balls is given by 21 (1, . . . , 1) and 12 (−1, . . . , −1, 1).
So how does this help us? We can actually apply the above estimate at each of the translated barriers because
we can iterate HI to get a thin strip of domain where we know u ≥ C and so we can construct similar barriers
on every point along this strip. Then at the end of these iterations take the smallest angle of the planes
generated by each barrier function and trap all of u ≥ cxn in B1/2 . Since u has this flatness associated to
is, it is C 1,α , as will be seen in the proof of the next theorem.
We now introduce the final theorem of this thesis. We first state it and prove an important lemma that
will be useful.
Theorem 5.8. Let u ∈ W 2,n (Ω) ∩ C 0 (Ω̄) where Ω is a C 1,α domain such that 0 ∈ ∂Ω and near zero
∂Ω = {(x0 , g(x0 ))} with |g(x0 )| ≤ δ|x0 |1+α . Suppose Lu = aij (x)uij = f (x) strongly on Ω with λI ≤ aij ≤
ΛI, |f | ≤ δ in B1 ∩ Ω and u ≤ δ|x0 |1+α on ∂Ω. Then u ∈ C 1,α (0).
Lemma 5.9. Assume u satisfies
a− +
0 xn − M δ ≤ u ≤ a0 xn + M δ
− + −
in B1 with |a+
0 − a0 | = 1 and a0 , a0 ∈ [−10, 10]. Then, in Br0 with r0 ≤ 1, u satisfies

a− 1+α
1 xn − M δr0 ≤ u ≤ a+ 1+α
1 xn + M δr0
− − −
with a+ + +
1 ≤ a0 + Cδ, a0 − Cδ ≤ a1 , and a1 − a1 ≤ (1 − η) for small η.

Proof. First we look at u(1/2en ) and notice that

1 − 1
a − M δ ≤ u(1/2en ) ≤ a+ + M δ.
2 0 2 0
−
Now notice that u(1/2en ) is either closer to the plane on top or the plane on the bottom: since 1/2|a+
0 −a0 | =
1/2 and the M δ is common in both hands of the inequality, we have either
1 1 1 + 1
u(1/2en ) − a− ≥ or a − u(1/2en ) ≥ .
2 0 4 2 0 4
Define ū := u − a−
0 xn . The hypothesis of the lemma becomes
−
−M δ ≤ ū ≤ (a+
0 − a0 )xn + M δ = xn + M δ.
−
Notice that a+
0 , a0 ∈ [−10, 10] implies

|ū| ≤ |u| + |a− 0 1+δ

0 xn | ≤ δ|x | + 10|xn | = δ|x0 |1+α + 10|g(x0 )| ≤ 11δ|x0 |1+α
on ∂Ω near zero. It is also important to notice that ū solves the same equation as u. Since we assumed that
u was closer to the lower plane, we have that ū(1/2en ) ≥ 14 − M δ. The Harnack Inequality tells us that
ū ≥ C0 − M δ (with C0 independent of δ) on B1/10 (1/2en ).
Now consider the barrier w(x) = M1 |x|−γ − M2 with M1 , M2 chosen such that w ≡ 1 on ∂B1/10 (1/2en )
and w ≡ 0 on ∂B1/2 (1/2en ). We compute the derivatives of w as

M1 γ(x21 + · · · + x2n )γ/2−1 + M1 γ(γ/2 − 1)(x21 + · · · + x2n )(2xi )(xi ) i = j


wij (x) =
M1 γ(γ/2 − 1)(x21 + · · · + x2n )γ/2−2 xi (2xj ) i 6= j.
Thus, looking in the radial direction (0, . . . , 0, r) we have
1 0 
r w (r)
 .. 
D2 w = 
 . 

1 0
 w (r)
r

w00 (r)

50
and so w00 > 0, w0 < 0 implies that
n−1 0
aij wij ≥ λw00 (r) + Λ w = γM1 rγ−2 (λ(γ − 1) + Λ(n − 1)) ≥ C1
r
by choosing γ. Now define
C0 w
W := − 100δ.
2
We want to do our estimates on Ω ∩ B1 − B1/10 (1/2en ) and so we examine first ∂B1 ∩ Ω. Here we know
W is very negative from its construction (more so than ū ≥ −M δ) and if it isn’t, we simply move our
computations from 1/2en 7→ 1/4en . This gives us ū ≥ W on Ω ∩ ∂B1 . From the bounds on g, we know that
the first points of contact between ∂B1 and ∂Ω will be at most size δ. But from our estimates of ū near
zero, we have
ū ≥ −11δ ≥ max W ≥ max W.
|xn |≤δ ∂Ω∩{|xn |≤δ}

Note that W ≡ C0 /2 − 100δ on ∂B1/10 (1/2en ) and ū ≥ C0 − M δ on B1/10 (1/2) by the harnack inequality.
Hence ū ≥ W on ∂B1/10 (1/2en ). If δ is small enough, which can be done by multiplying f by a constant
since it doesn’t change any of our inequalities, we have LW = C0 C1 ≥ Lū ∼ δ and so the maximum principle
dictates that W ≤ ū on Ω ∩ B1 − B1/10 (1/2en ).
Observe that in the radial direction {x0 = 0}, W satisfies
C0
W (0, xn ) = w(0, xn ) − 100δ
2
≥ C0 C3 xn − 100δ

because of how |x|−γ grows. Also notice that x0 = 0 was not specific and we could have taken any x0 ∈
{|xn | ≤ δ} ∩ ∂Ω as we look for planes in the radial xn direction starting from ∂Ω. Taking the maximum C3
of all such planes we obtain

u − a−
0 xn = ū ≥ W
≥ C0 C3 xn − 100δ
u ≥ (C0 C3 + a−
0 ) − 100δ

and C0 C3 + a− −
0 =: a1 .

Now assume that u is closer to the upper plane than the lower plane. In this case then u satisfies
1 +
a − u(1/2en ) ≥ 1/4.
2 0

51
We define u = a+
0 xn − u and clearly we are looking for a lower bound for u. The hypothesis of the Lemma
becomes

−a− +
0 xn + M δ ≥ −u ≥ −a0 xn − M δ
−
(a+
0 − a0 )xn + M δ ≥ u ≥ −M δ.
xn + M δ ≥ u ≥ −M δ.

Near zero u satisfies the estimate

0 1+α
|u| ≤ |a+
0 xn | + |u| ≤ 10|xn | + δ|x | = 10|g(x0 )| + δ|x0 |1+α ≤ 11δ|x0 |1+α .

Also since u was closer to the upper plane, u(1/2en ) ≥ 1/4 − M δ. The Harnack Inequality tells us that
u ≥ C0 − M δ. As one can probably tell, the proof is going to be identical to the one for ū.
Now consider the barrier w(x) = M1 |x|−γ − M2 with M1 , M2 chosen such that w ≡ 1 on ∂B1/10 (1/2en )
and w ≡ 0 on ∂B1/2 (1/2en ). In a very similar sense as the lower bound by the barrier in the previous case,
we have aij wij ≥ C1 = C1 (n, λ, Λ, γ). Define W by

C0 w
W := − 100δ.
2
We want to do our estimates on Ω ∩ B1 − B1/10 (1/2en ) and so we examine first ∂B1 ∩ Ω. Here we know
W is very negative from its construction (more so than u ≥ −M δ) and if it isn’t, we simply move our
computations from 1/2en 7→ 1/4en . This gives us u ≥ W on Ω ∩ ∂B1 . From the bounds on g, we know that
the first points of contact between ∂B1 and ∂Ω will be at most size δ. But from our estimates of u near zero,
we have
u ≥ −11δ ≥ max W ≥ max W.
|xn |≤δ ∂Ω∩{|xn |≤δ}

Note that W ≡ C0 /2 − 100δ on ∂B1/10 (1/2en ) and u ≥ C0 − M δ on B1/10 (1/2) by the harnack inequality.
Hence u ≥ W on ∂B1/10 (1/2en ). If δ is small enough, which can be done by multiplying f by a constant
since it doesn’t change any of our inequalities, we have LW = C0 C1 ≥ Lu ∼ δ and so the maximum principle
dictates that W ≤ u on Ω ∩ B1 − B1/10 (1/2en ).
Observe that in the radial direction {x0 = 0}, W satisfies

C0
W (0, xn ) = w(0, xn ) − 100δ
2
≥ C0 C3 xn − 100δ

because of how |x|−γ grows. Also notice that x0 = 0 was not specific and we could have taken any x0 ∈
{|xn | ≤ δ} ∩ ∂Ω as we look for planes in the radial xn direction starting from ∂Ω. Taking the maximum C3
of all such planes we obtain

a+
0 xn − u = u ≥ W
≥ C0 C3 xn − 100δ
(a+
0 − C0 C3 )xn + 100δ ≥ u

and a+ +
0 + C0 C3 =: a1 . Thus we can see that even though the distance from the origin increases a little bit,
the slope decreases a lot more.
Proof of Theorem. We rescale u by defining

u(r0 x)
ũ(x) = .
r0

52
Then by Lemma 5.9, in Br0 we have

a− α + α
1 xn − M δr0 ≤ ũ ≤ a1 xn + M δr0

−
with |a+ 0 0
1 − a1 | = 1 − η by direct computations. We also rescale the boundary by g̃(x ) = g(r0 x )/r0 =⇒
α 0 1+α
|g̃| ≤ δr0 |x | . We can also compute

aij (r0 x) r02 uij (r0 x)

L̃ũ = ãij (x)ũij (x) = = f (r0 x).
r0 r0
This implies that we can iterate the conclusion of the Lemma by rescaling and so iterating k times with
r = r0k , we have
a−
k xn − M δr
1+α
≤ u ≤ a+
k xn + M δr
1+α

− ±
in Br = Br0k with |a+ k β
k − ak | = (1 − η) ≈ r for some β. This implies convergence of the ak to an a∞ in the
−
sense that |a+ β β
k − a∞ | ≤ Cr , |ak − a∞ | ≤ Cr . Then

a−
k xn − M δr
1+α
≤ u ≤ a+
k xn + M δr
1+α

(a−
k − a∞ )xn − M δr
1+α
≤ u − a∞ xn ≤ (a+
k − a∞ )xn + M δr
1+α
.

Our analysis finally gives

|u − a∞ xn | ≤ Cr1+β + M δr1+α .
This proves that if β < α, then u ∈ C 1,β (0). If β ≥ α then u ∈ C 1,α (0).

53
References
[1] Gilbarg, D., Trudinger, N.S. Elliptic Partial Differential Equations of Second Order, 2nd ed., Springer-
Verlag, NY, 1983.
[2] Caffarelli, L., Cabré, X. Fully Nonlinear Elliptic Equations, AMS,1995

54