Lectures on Elliptic PDE

Tarek M. Elgindi

April 12, 2021

Abstract
These are notes on Elliptic PDE following the notes of Q. Han and F.H. Lin for the first
part and then Caffarelli and Cabre for the second part. I also benefited from and have taken
from notes of L. Silvestre, C. Mooney, and X. Ros Oton and X. Fernandez.

Contents
1 Introductory Remarks 3
1.1 Some contexts in which Elliptic PDEs arise . . . . . . . . . . . . . . . . . . . . . 3
1.2 What the course is about . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

2 Notation and a few Definitions 6

2.1 A few function spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

3 Minimization, the Maximum Principle, and properties of Harmonic functions 7

3.1 Minimzation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
3.2 The Maximum Principle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.3 Mean-Value Property . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.4 Smoothness of Harmonic Functions . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.5 Weyl Lemma . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.6 Strong Maximum Principle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.7 Liouville Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.8 Harnack Inequality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
3.9 Hopf Lemma . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

4 General Elliptic Equations 21

5 Maximum Principles for General Elliptic Problems 25

5.1 A-Priori Estimates Through the Maximum Principle . . . . . . . . . . . . . . . . 31
5.2 Boundary Estimates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
5.3 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38

1
6 The C α and C 1,α Theory 39
6.1 General results on Morrey-Campanato Spaces . . . . . . . . . . . . . . . . . . . . 40
6.2 C α bounds for u . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
6.3 C 1,α bounds for u . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
6.4 Exercises on the Newtonian Potential . . . . . . . . . . . . . . . . . . . . . . . . 50

7 De Giorgi’s Theorem 52
7.1 From L2 to L∞ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
7.2 From L∞ to C α . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
7.2.1 Harnack Inequality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
7.2.2 Oscillation Lemma . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
7.3 From C α to C 1,α to C ∞ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
7.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67

8 Preliminaries on Viscosity Solutions: Tangent Paraboloids and Second Order

Differentiability 68

9 Viscosity Solutions 73
9.1 Elliptic Equations and the Definition of Viscosity Solutions . . . . . . . . . . . . 74
9.1.1 Basic Properties of Viscosity Solutions . . . . . . . . . . . . . . . . . . . . 76
9.2 The Class of solutions of all uniformly Elliptic Equations and the Pucci Operators 79

10 Alexandroff Maximum Principle and Applications 84

10.1 The case of classical solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
10.1.1 Contact Set . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
10.2 Maximum Principle on Small Domains . . . . . . . . . . . . . . . . . . . . . . . . 86
10.3 ABP Estimate for Viscosity Solutions . . . . . . . . . . . . . . . . . . . . . . . . 87

11 Harnack Inequality and Oscillation Lemma 91

11.1 Barrier and Calderón-Zygmund Decomposition . . . . . . . . . . . . . . . . . . . 91
11.2 The Harnack Inequality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
11.2.1 Bound on the distribution function . . . . . . . . . . . . . . . . . . . . . . 94
11.2.2 Bounds on all scales . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97

12 Uniqueness and C 1,α Regularity 100

12.1 The upper  envelope and its properties . . . . . . . . . . . . . . . . . . . . . . . 101
12.2 C 1,α regularity of solutions to F (D2 u) . . . . . . . . . . . . . . . . . . . . . . . . 106
12.3 The case of Concave Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107

13 The Evans-Krylov Theorem, C 2,α regularity for solutions to Concave equa-

tions 109
2 2,α
13.1 Evans-Krylov Theorem (C to C ) . . . . . . . . . . . . . . . . . . . . . . . . . 109
13.2 L∞ to C 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113

2
1 Introductory Remarks

1.1 Some contexts in which Elliptic PDEs arise

Elliptic PDEs arise naturally in a number of different contexts. Two important examples are:

1. The Calculus of Variations.

2. Stationary/Equilibrium solutions to physical problems.

Indeed, an important branch of the calculus of variations is to study functionals of the form:

Z b
L(u) = F (x, u(x), u0 (x))dx. (1.1)
a

It is an exercise to check that whenever F (x, p, z) is convex in the variable z, the Euler-Lagrange

equation associated to the functional is “elliptic.” These variational problems also include a

number of important geometric problems such as the problems of minimal surfaces and obstacle

problems.

Regarding the second class of problems, solutions to so-called semilinear elliptic PDE such

as:

∆ψ = F (ψ) (1.2)

serve as stationary solutions to the incompressible Euler equation and the non-linear heat,

wave, and Schrodinger equations. Even in the context of evolution equations, the pressure

in the incompressible Euler and Navier-Stokes equations solves an elliptic problem. Solutions

to semilinear problems like (1.2) are also helpful in studying blow-up dynamics of non-linear

equations.

1.2 What the course is about

The course is about the qualitative properties of solutions to elliptic PDEs. We will be most

concerned with the following matters:

1. Maximum and Comparison Principles and Symmetry results,

3
 2. Classical and modern regularity theory.

Through homework assignments, I hope we will also learn about:

3. Spectral theory of elliptic operators

4. Existence results, relevant function spaces, and potential theory

1.3 Exercises

Remark 1.1. Exercises in this course are of varying degrees of complexity. Some will require one

line and some will require several pages (of course, those requiring more work are not necessarily

more difficult).

1. Derive the equation for the pressure in the incompressible Euler/Navier-Stokes equation

∂t u + u · ∇u + ∇p = ν∆u,

div(u) = 0.

2. Prove that there exists no smooth solution to the minimization problem:

Z 1
0
L(u) = e−u + u2 ,
0

where the admissible space is X = {u ∈ C 2 ([0, 1]) : u(0) = 0, u(1) = 1}. An interesting
1
fact is that if the boundary condition u(1) = 1 is replaced by u(1) = 2, then there is

existence (but you don’t need to show this!).

3. Consider Pn , the vector space of all (real) polynomials defined on Rn . Let’s define the

subspace Pn,k to be the space of homogeneous harmonic polynomials of degree k. Compute

the dimension of Pn,k .

4
4. Find ψ ∈ C 2 (B1 (0) \ {0}) with ∆ψ ∈ C(B1 (0)) while D2 ψ is not bounded at 0. Next, find

a ψ ∈ C 2 (B1 (0) \ {0}) with det(D2 ψ) ∈ C(B1 (0)) while D2 ψ is not bounded at 0. Could

a single function work for both questions?

5. (Functional Analysis) Let K ∈ L2 ([0, 1] × [0, 1]) and assume that K(x, y) = K(y, x) for all

x, y ∈ [0, 1]. Define the operator

Z 1
LK (f ) = K(x, y)f (y)dy.
0

• Prove that LK is bounded and kLK k ≤ |K|L2 ([0,1]2 ) .

• Show that LK is self adjoint. That is,

(LK f, g)L2 = (g, LK f )L2 ,

for all f, g ∈ L2 .

• Show that if K ∈ C 1 ([0, 1] × [0, 1]), then LK is compact.

• Show that if K ∈ L2 ([0, 1] × [0, 1]), then LK is compact.

• Does it matter that we used [0, 1]?

6. Consider the boundary value problem

u00 (x) = f (x),

for x ∈ [0, 1], with the boundary conditions

u(0) = 0, u(1) = 0.

Write the solution operator as an operator of the type in Exercise 4.

5
2 Notation and a few Definitions

Rd = {(x1 , x2 , ..., xd ) : xi ∈ R}.

For x, y ∈ Rd ,
d
X
(x, y) := xi yi ,
i=1

and
p
|x| := (x, x).

Br (x) = {y ∈ Rd : |x − y| < r}.

∂Br (x) = {y ∈ Rd : |x − y| = r}

Br (x) = {y ∈ Rd : |x − y| ≤ r}.

We write ∂xi as the partial derivative with respect to xi . We write ∂xki for k−partial derivatives

with respect to xi , ∂xi xj is the mixed partial derivative with respect to xi and xj , etc. One of

the most important operators we will encounter is the Laplacian:

d
X
∆ := ∂x2i .
i=1

Note that for any twice differentiable function f , ∆f = div(∇f ).

2.1 A few function spaces

Definition 2.1. A function f : Br (x) → R is said to be infinitely differentiable if it possesses

partial derivatives of all orders in all directions. In this case we write f ∈ C ∞ (Br (x)).

Definition 2.2. Let f : Br (x) → R be measurable. We say that f ∈ L∞ (Br (x)) if there exists

M ≥ 0 so that {y ∈ Br (x) : |f (y)| ≥ M } has measure-zero. If f ∈ L∞ (Br (x)), we define

kf kL∞ := inf{M ||f (y)| ≤ M almost everywhere}.

6
When f is continuous and bounded on Br (x), kf kL∞ is just the supremum of |f |.

Definition 2.3. Let 0 ≤ α ≤ 1. A function f : Br (x) → R is said to be Hölder continuous of

order α if
|f (y1 ) − f (y2 )|
sup |f (y)| + sup < ∞.
y∈Br (x) y1 ,y2 ∈Br (x) |y1 − y2 |α

In this case we write f ∈ C α (Br (x)) and

|f (y1 ) − f (y2 )|
kf kC α = sup |f (y)| + sup .
y∈Br (x) y1 ,y2 ∈Br (x) |y1 − y2 |α

Definition 2.4. A function is said to belong to C k,α (Br (x)) if it is k−times differentiable and

all its ith partial derivatives, 0 ≤ i ≤ k belong to C α (Br (x)).

3 Minimization, the Maximum Principle, and properties of Har-

monic functions

The first place we usually encounter elliptic equations is through the Dirichlet principle. Physi-

cally, this is the problem of finding the steady temperature distribution of a plate whose bound-

ary has a given temperature distribution. For simplicity, let’s take the plate to be B1 (0) ⊂ R2 .

Now let’s assume that the given temperature distribution on ∂B1 (0) is some function g. If the

temperature distribution is smooth, it satisfies Laplace’s equation:

∆u(y) = 0 , y ∈ B1 (0) (3.1)

u(y) = g(y), y ∈ ∂B1 (0). (3.2)

Definition 3.1. Let Ω ⊂ Rd be open. If u is twice differentiable on Ω and ∆u(y) = 0 for all

y ∈ Ω, u is called harmonic in Ω.

Upon reflecting upon this equation, there are several questions that come to mind:

1. Given a g, does there always exist a solution u to (3.1)-(3.2)? Is it unique? What about

7
 if g is rough?

2. Assuming that the first question is settled, what qualitative properties can we tell about

u given g?

3. Can we find an “exact” expression for u if we are given g?

4. How much of the above remains once we change the plate from B1 (0) to a general shape

which may be quite erratic?

We will touch upon some aspects of some of these questions during these lectures. First, I

would like to point to one important source of (3.1)-(3.2).

3.1 Minimzation

Let g ∈ C 2 (∂B1 (0)) be given. Define L : C 2 (B1 (0)) → R by

Z
L(u) = |∇u|2 .
B1 (0)

We let

X = {u ∈ C 2 (B1 (0)) : u(y) = g(y) ∀y ∈ ∂B1 (0)}.

Theorem 1. Assume that u ∈ X and

L(u) = inf L(w).

w∈X

Then, u satisfies (3.1)-(3.2).

Remark 3.2. This theorem does not show that such a u exists. However, it shows that if a

minimzer of L existed, it would have to satisfy (3.1)-(3.2), which is called the Euler-Lagrange

equation associated to minimization of L over X.

Proof. Assume u ∈ X and L(u) = inf w∈X L(w). Let

V := {v ∈ C 2 (B1 (0)) : v(y) = 0 ∀y ∈ ∂B1 (0)}.

8
V is called the set of admissible “variations.” Then, for all v ∈ V and t ∈ R, u + tv ∈ X. This

is because v ∈ C 2 (B1 (0)) vanishes on ∂B1 (0). This implies that L(u) = inf t∈R L(u + tv). It is

easy to check that, as a function of t, L(u + tv) is continuously differentiable (in fact, it is just

a polynomial in t). Thus, since L(u + tv) has a minimum at t = 0, we have:

d
L(u + tv) = 0.
dt t=0

Now upon plugging this into the expression for L we get:

Z Z
d 2 d
0= |∇u + t∇v| = |∇u|2 + 2t(∇u, ∇v) + t2 |∇v|2
dt B1 (0) t=0 dt B1 (0) t=0
Z
= 2(∇u, ∇v).
B1 (0)

Thus, for all v ∈ V , we have:

Z
(∇u, ∇v) = 0,
B1 (0)

now we write:
Z Z
div(v∇u) = (∇u, ∇v) + v∆u.
B1 (0) B1 (0)

Thus, by the divergence theorem,

Z Z
v∇u · n = v∆u.
∂B1 (0) B1 (0)

But since v(y) = 0 for all y ∈ ∂B1 (0) we see that

Z
v∆u = 0
B1 (0)

for all v ∈ V . Now we have a lemma.

9
Lemma 3.3. Assume f ∈ C(B1 (0)) and

Z
fv = 0
B1 (0)

for all v ∈ V = {v ∈ C 2 (B1 (0)) : v(y) = 0 ∀y ∈ ∂B1 (0)}. Then, f (y) = 0 for all y ∈ B1 (0).

The proof of this lemma is a part of Homework 1.

Using this lemma we see ∆u(y) = 0 for all y ∈ B1 (0). This concludes the proof of Theorem

(1).

Minimization problems are a great source of elliptic problems which are relevant both in

theoretical and applied mathematics. We now move to another important aspect of elliptic

equations: the maximum principle. Physically, the maximum principle says that if the steady

state temperature profile of a plate is maximized on the boundary of the plate if there are no

interior heat sources.

3.2 The Maximum Principle

We begin with the following simple lemma.

Lemma 3.4. Assume that f : Rd → R is twice differentiable. Assume f attains a local maximum

(resp. minimum) at x0 . Then, ∆f (x0 ) ≤ 0 (resp. ≥ 0).

Proof. ∂x2i f (x0 ) ≤ 0 for all i.

A consequence of this is the maximum principle which is derived from physical intuition: if

we are heating a plate (no matter what shape it is) on its boundaries, the temperature inside will

be lower than the highest temperature on the boundary and higher than the lowest temperature

on the boundary unless the temperature on the boundary is constant. If the temperature on the

boundary were not constant, the temperature should attain its maximum and minimum values

only on the boundary (since the inside values should be some kind of average of the boundary

10
values). What was just described in words is called the strong maximum principle. We will now

give the weak maximum principle which says the maximum and the minimum must be attained

on the boundary.

Theorem 2. (Weak maximum/minimum principle) Assume f ∈ C 2 (Ω) with Ω a bounded open

set. Assume that ∆f = 0 on Ω. Then,

sup f (x) = sup f (x),

x∈Ω x∈∂Ω

and

inf f (x) = inf f (x).

x∈Ω x∈∂Ω

Proof. Since Ω is bounded, Ω ⊂ BK (0) for some K ≥ 0. Let

M = sup f (x)
x∈∂Ω

and

m = inf f (x).
x∈∂Ω

Define f = f − |x|2 . Observe that

∆f = ∆f − 2d = −2d < 0

(since ∆|x|2 = 2d). This means that f has no interior minimum, since at an interior minimum

of f , we must have ∆f non-negative by the preceding lemma. This implies that

inf f (x) = inf f (x).

x∈Ω x∈∂Ω

However,

inf f (x) ≤ inf f (x) = inf f (x) ≤ inf f (x) + K 2

x∈∂Ω x∈∂Ω x∈Ω x∈Ω

for all  > 0. This implies inf x∈∂Ω f (x) ≤ inf x∈Ω f (x) and so they must be equal (since

11
we trivially have the opposite inequality). Now applying the same reasoning to −f , we get

supx∈∂Ω f (x) = supx∈Ω f (x)

Remark 3.5. Note that if ∆u ≤ 0 in Ω, then u will satisfy the maximum principle on Ω while

if ∆u ≥ 0 in Ω then u will satisfy the minimum principle on Ω.

Corollary 3.6. (Uniqueness and Stability)

Assume u, v are harmonic on an open bounded set Ω. Then we have:

sup |u(y) − v(y)| ≤ sup |u(y) − v(y)|.

y∈Ω y∈∂Ω

What this means is that there is at most one solution of (3.1)-(??) for a given g (uniqueness).

Moreover, we have that if the boundary values of two harmonic functions are close, then they

are close inside as well (stability). The proof follows simply from the fact that u − v is also

harmonic and satisfies the maximum principle.

3.3 Mean-Value Property

Theorem 3. Assume u is twice differentiable and harmonic on Br (x) for some x ∈ Rd and

r > 0. Then,
Z Z
1 1
u(x) = u(y)dy = u(y)dS(y)
|∂Bδ (x)| ∂Bδ (x) |Br (x)| Br (x)

for all δ < r.

Proof. First note that, by translation invariance, we may assume x = 0. (By replacing u with

ũ(·) = u(· + x), which is then harmonic on Br (0)). Next, notice that

Z
1
lim u(y)dS(y) = u(0).
δ→0 |∂Bδ (0)| ∂Bδ (0)

Indeed,

Z Z Z
1 1 1
u(y)dS(y)−u(0) = (u(y)−u(0))dS(y) ≤ |u(y)−u(0)|dS(y)
|∂Bδ (0)| ∂Bδ (0) |∂Bδ (0)| ∂Bδ (0) |∂Bδ (0)| ∂Bδ (0)

12
 Z
1
≤ |∇u|L∞ δdS(y) = δ|∇u|L∞ → 0
|∂Bδ (0)| ∂Bδ (0)

as δ → 0. Next, let us define, for 0 < δ < r,

Z
1
I(δ) := u(y)dS(y).
|∂Bδ (0)| ∂Bδ (0)

We will show that I 0 (δ) = 0 for all 0 < δ < r. This will imply that I(δ) = u(x) for all δ < r

since limδ→0 I(δ) = u(x). Observe that

Z Z Z
d−1
∆u(y)dy = ∂n u(y)dS(y) = δ ∂δ u(yδ)dS(y).
Bδ (0) ∂Bδ (0) ∂B1 (0)

The first equality uses the divergence theorem and the second is just using the change of variables

y → yδ. Now note:

Z Z Z
d−1 d−1 d d−1 d 1 
δ ∂δ u(yδ)dS(y) = δ u(yδ)dS(y) = δ u(y)dS(y) ,
∂B1 (0) dδ ∂B1 (0) dδ δ d−1 ∂Bδ (0)

where in the last equality we just the changed variables back again yδ → y. Thus we see:

Z Z
d−1 d 1 
∆u(y)dy = δ u(y)dS(y) .
Bδ (0) dδ δ d−1 ∂Bδ (0)

This implies that if ∆u ≡ 0, I 0 (δ) = 0 for all δ > 0. This concludes the proof of the surface

mean value property. Now we have that

Z
1
u(0) = u(y)ds(ω)
|∂Bs (0)| ∂Bs (0)

for all s ≤ δ < r. Now if we integrate this expression from 0 to δ we get:

Z δ Z δ Z
u(0) |∂Bs (0)| = u(ω)dS(ω)ds
0 0 ∂Bs (0)

13
which then implies
Z
u(0)|Bδ (0)| = u(y)dy.
Bδ (0)

This concludes the proof.

The mean value property has a number of important consequences: the strong maximum

principle, the Harnack inequality, and the Liouville theorem.

Remark 3.7. Since we will make great use of the mean value property, it will be useful to use

the following notation:

Z Z
1
− f (y)dy := f (y)dy.
Bδ (x) |Bδ (x)| Bδ (x)

3.4 Smoothness of Harmonic Functions

Theorem 4. Let Ω be an open set. Assume u ∈ C(Ω) satisfies the mean value property on Ω.

Then, u is infinitely differentiable and harmonic.

Remark 3.8. A continuous function f on Ω satisfies the mean value property on Ω if whenever
R
Br (x) ⊂ Ω implies that then u(x) = −Bδ (x) u(y)dy for all δ < r. Note that the end of the

preceding proof gives that the “surface” and “volume” mean value properties are equivalent.

Proof. We will follow the proof in the book of Han and Lin. Let φ ∈ Cc∞ (B1 (0)) with
R
B1 (0) φ =1

and φ(x) = ψ(|x|). In other words,

Z 1
ωd rd−1 ψ(r)dr = 1.
0

1 z
Now define φ (z) = n φ(  ) for  > 0. Now, for all x ∈ Ω with  < dist(x, ∂Ω) we have:

Z Z Z
1
u(y)φ (y − x)dy = u(x + y)φ(y)dy = n u(x + y)φ(y/)dy
Ω  |y|<

Z
= u(x + y)φ(y)dy,
|y|<1

14
where we have just changed variables twice. Now we observe:

Z Z 1 Z
u(x + y)φ(y)dy = rn−1 u(x + rω)φ(rω)dSω dr
|y|<1 0 ∂B1 (0)

Z 1 Z
= ψ(r)rn−1 u(x + rω)dSω dr.
0 |ω=1|
R
Now notice that |ω=1| u(x + rω)dSω is independent of r using the mean-value property. Thus,

Z Z 1
u(y)φ (y − x)dy = u(x)ωn ψ(r)rn−1 dr = u(x).
Ω 0

is a C ∞
R
But notice that since u is continuous and φ is compactly supported, Ω u(y)φ (y − x)dy

function of x. Thus, u ∈ C ∞ (Ω). Now we use the equality from the previous proof:

Z Z
d−1 d 1 
∆u(y)dy = r u(y)dS(y)
Br (x) dr rd−1 ∂Br (x)

and by the mean-value property the quantity on the left hand side is identically zero for all r

and x. Thus we get:

Z
∆u(y)dy = 0
Br (x)

for all r and x. This implies that ∆u ≡ 0 and u is harmonic.

3.5 Weyl Lemma

We now move to prove the so-called Weyl Lemma.

Lemma 3.9. Suppose that f ∈ L1 (B1 (0)) and that = 0 for all φ ∈ Cc∞ (B1 (0)).
R
B1 (0) f ∆φ

Then, f ∈ C ∞ and f is harmonic on B1 (0).

Remark 3.10. This means that any “weakly harmonic” function is smooth and harmonic.

Proof. The idea of the proof is to convolve f with an approximation of the identity. We will

then show that f satisfies the local mean value property. That is, we will show that for all

15
x ∈ B1 (0) there exists rx > 0 so that

Z
− f (y) = f (x),
Br (x)

for all r < rx . It then follows from the preceding proof that f is harmonic and smooth. Let

Φ ∈ Cc∞ and Rd Φ = 1. It is easy to construct a radial function with these two properties.
R

Define
1 x
Φ (x) = d
Φ( ).
 

Define
Z
f (x) = Φ (x − y)f (y)dy.
B1 (0)

Observe that f ∈ C ∞ (B1 (0)). Now let us compute:

Z
∆f (x) = ∆Φ (x − y)f (y)dy,
B1 (0)

for all x ∈ B1 (0). If  is sufficiently small, we then see that ∆f (x) = 0. It follows that f

satisfies the mean value property:

Z
f (x) = − f .
Br (x)

Now we just send  → 0 and observe that f must satisfy the (local) mean value property almost

everywhere. It is then easy to show that f is actually continuous. Indeed, by the mean value

property we have that for x, y ∈ B1 (0)

Z
1
|f (x) − f (y)| ≤ |f |
|Br | Br (x)∆Br (y)

and we have that

|Br (x)∆Br (y)| ≤ C|x − y|.

16
Now we just observe that for an L1 function, given an  > 0, there is a δ > 0 so that if |A| < δ,

Z
|f | < .
A

We now move to discuss a few more consequences of the mean-value property.

3.6 Strong Maximum Principle

Theorem 5. Let Ω ⊂ Rd be open and connected. Assume u ∈ C(Ω̄) satisfies the mean-value

property. Then, u assumes its maximum and minimum only on ∂Ω unless u is constant.

Proof. We will only do the case of maximum (just multiply by −1 to get the other case!). Let

Σ := {x ∈ Ω̄ : u(x) = sup u(y)}

y∈Ω̄

Since u is continuous, Σ is closed. Now let’s show that Σ is also open. Indeed, let x ∈ Σ ∩ Ω.

Then there exists r > 0 such that Br (x) ⊂ Σ. Thus,

Z
M = u(x) = − u(y)dy ≤ M
Br (x)

by the mean value property. Thus, u(y) = M for all y ∈ Br (x). This means that Σ is open and

closed. Thus Σ = Ω or Σ = φ.

3.7 Liouville Theorem

We next move to the Liouville Theorem:

Theorem 6. Any harmonic function on Rd which is bounded from above or below is identically

constant

To prove this theorem, we will need to establish gradient estimates for the Laplacian.

17
Theorem 7. Suppose u ∈ C 1 (BR (x0 )) is harmonic in BR (x0 ). Then,

√
d d
|D(u)(x0 )| ≤ sup |u|.
R BR (x0 )

Proof. Observe that ∂i u is harmonic and so we may apply to it the mean-value formula:

Z Z Z
d
∂i u(x0 ) = − ∂i u(y)dy = − div(0, .., u, 0, ...)dy = − u(y)νi dS(y).
BR (x0 ) BR (x0 ) R ∂BR (x0 )

Thus,
d
|∂i u(x0 )| ≤ sup |u|.
R BR (x0 )

Note that using the above proof, we also have that if u ≥ 0 is harmonic on BR , then

d
|∂i u(x0 )| ≤ u(x0 )
R

using the surface mean-value property. In particular, if u is a non-negative harmonic function

on Rd , then ∂i u(x) = 0 for all x ∈ Rd . This implies Liouville’s theorem.

3.8 Harnack Inequality

The Harnack inequality is contained in the following theorem:

Theorem 8. Suppose u is harmonic in Ω (an open and connected subset of Rd ). Then, for any

compact subset K ⊂ Ω, there exists C = C(Ω, K) such that if u ≥ 0 in Ω, then

1
u(y) ≤ u(x) ≤ Cu(y)
C

for all x, y ∈ K.

18
Proof. Let x0 ∈ K and R > 0. Assume B2R (x0 ) ⊂ Ω. Now let y ∈ BR (x0 ). On the one hand,

we have
Z Z
u(y) = − u(z)dz ≤ 2n − u(z)dz = 4n u(x0 ).
BR (y) B2R (x0 )

On the other hand,

Z Z
1 1
u(y) = − u(z)dz ≥ − u(z)dz = n u(x0 ).
B2R (y) 2n BR (x0 ) 2

Thus,
1
u(x0 ) ≤ u(y) ≤ 2n u(x0 ).
2n

Thus, for all x, y ∈ BR (x0 ), we have:

1
u(y) ≤ u(x) ≤ 4n u(y).
4n

Now take x1 , ..., xN ∈ K with the property that BR (xi ) covers K with 4R < dist(K, ∂Ω). Then
 N
we get the claimed inequality for C = 41n .

3.9 Hopf Lemma

Next we will prove the Hopf Lemma.

Theorem 9. Suppose u ∈ C 1 (B1 (0)) is a harmonic function in B1 (0). If u(x) < u(x0 ) for any

x ∈ B1 (0) and some x0 ∈ ∂B1 (0), then we have:

∂n u(x0 ) ≥ C(u(x0 ) − u(0))

where C is a constant depending only on n.

Proof. We follow the proof in Han-Lin. Define v : B1 (0) → R by

2
v(x) = e−α|x| − e−α ≥ 0

19
on B1 (0). Notice that

∆v(x) = div(−2αx exp(−α|x|2 )) = exp(−α|x|2 )(−2αn + 4α2 |x|2 ).

Thus,

∆v(x) > 0

1
if |x| ≥ 2 and α > 2n. Thus, v is subharmonic in B1 \ B1/2 and it is positive in |x| < 1 while it

vanishes on |x| = 1. Now let  > 0 and define

h (x) = u(x) − u(x0 ) + v(x).

Since u is harmonic,

∆h > 0

in B1 \ B 1 . Now, we have h (x0 ) = 0 while h ≤ 0. Now, we have u(x) < u(x0 ) for |x| = 21 . Thus,
2

we may take  very small so that h < 0 on |x| = 12 . In particular, h achieves its maximum at

the point x0 by the weak maximum principle (since h is maximum at x0 in ∂(B1 \ B 1 )). Thus,
2

∂n h (x0 ) ≥ 0.

This implies that ∂n u(x0 ) ≥ −∂n v(x0 ) = 2α exp(−α). Now we need to bound  from below.

Let w(x) = u(x0 ) − u(x) > 0 in B1 . Since w is harmonic in B1 , we may apply to it the Harnack

inequality to get:

inf w ≥ c(n)w(0).
B1/2

Thus,

u(x0 ) − max u ≥ c(n)(u(x0 ) − u(0)).

B1/2

20
Now since v ≤ 1, we may take  = 12 c(n)(u(x0 ) − u(0)). Thus,

∂n u(x0 ) ≥ αc(n)(u(x0 ) − u(0)).

α can be taken to be 2n + 1.

4 General Elliptic Equations

Remark 4.1. In the following, we will use the “summation convention” that repeated indices

are summed. So, ai bi means di=1 ai bi .

P

The Laplacian is:

∆ = ∂ii = δi,j ∂i,j ,

where δi,j = 1 if i = j and δi,j = 0, i 6= j. In other words, {δi,j }i,j is just the identity matrix.

Another way to write this is:

∆f = div(A∇f ),

where A = Id. Elliptic equations are just those where A is a symmetric positive definite matrix

with eigenvalues uniformly bounded above and below.

Definition 4.2. aij : Ω → R, for 1 ≤ i, j ≤ d is said to be uniformly elliptic if there exist

constants λ, Λ > 0 so that

X
λ|ξ|2 ≤ aij ξi ξj ≤ Λ|ξ|2 ,
i,j

for all x ∈ Ω and all ξ ∈ Rn .

A nice fact about uniformly elliptic operators is the following:

If for example,

∂i (ai,j ∂j u) = f,

21
u, f ∈ Cc∞ then we have:
Z Z
∂i (ai,j ∂j u)u = uf.

Thus,
Z Z
ai,j ∂i u∂j u = uf ,

which implies that

Z Z Z Z
2
λ |∇u| ≤ ai,j ∂i u∂j u = uf ≤ | uf |.

It follows that ∇u can be controlled by u itself! This is what is known as the “energy method.”

Let us now give a localized version which is known as the Cacciopolli inequality.

Lemma 4.3. Suppose u ∈ C 1 (B1 ) satisfies

Z
ai,j Di uDj φ = 0, for any φ ∈ C01 (B1 )

In other words, u is a weak solution of ∂j (aij Di u) = 0. Then, for any function η ∈ C01 (B1 ), we

have
Z  Λ 2 Z
η 2 |Du|2 ≤ 4 |Dη|2 u2 ,
B1 λ B1

where C depends only on λ and Λ

Λ
Remark 4.4. We will find that many of the results we will prove rely on the ratio λ, which

may be important when studying degenerate elliptic problems when λ vanishes at some points

in the domain.

Proof. For any η ∈ C01 (B1 ) set φ = η 2 u. Then,

Z
aij Di uDj (η 2 u) = 0.

Therefore,
Z Z
2
η aij Di uDj u = −2 uηaij Di uDj η.

22
It follows that

Z Z sZ sZ
λ η 2 |Du|2 ≤ 2Λ |Du||Dη||u||η| ≤ 2Λ |Du|2 η 2 |Dη|2 u2 .

It follows that
Z  Λ 2 Z
2 2
|Du| η ≤ 4 |u|2 |Dη|2 .
λ

Corollary 4.5. Let u be as in the preceding Lemma. Then For any 0 ≤ r < R ≤ 1 we have that

Z Z
16  Λ 2
2
|Du| ≤ u2 .
Br (R − r)2 λ BR \Br

2
Proof. Take η such that η = 1 in Br and η = 0 outside BR and |Dη| ≤ R−r . Why can we do

this?

Let us first recall the Poincare inequalities:

Lemma 4.6. Let u ∈ C 1 (Ω) and assume that Ω is bounded. If u = 0 on ∂Ω then we have

Z Z
|u|2 ≤ CΩ |Du|2 .
Ω Ω

In general,
Z Z
2
|u − ū| ≤ C2,Ω |Du|2 .
Ω Ω

These inequalities will be explored in the homework.

Next, we have an “oscillation lemma.”

Corollary 4.7. Let u be as in the last two results. Then, for any 0 < R ≤ 1, there holds

Z Z
2
u ≤ θ1 u2
BR/2 BR

23
and
Z Z
2
|Du| ≤ θ2 |Du|2 ,
BR/2 BR

where θ1 , θ2 = C( Λλ , n) ∈ (0, 1) are explicitly given in the proof.

3
Proof. Let η ∈ C01 (BR ) with η = 1 on BR/2 and |Dη| ≤ R. Then,

Z Z Z  Λ 2 Z
2 2 2 2 2 2
|D(ηu)| = |ηDu + uDη| ≤ 2 |Du| η + u |Dη| ≤ 2(1 + 4 ) |Dη|2 u2 ,
BR BR λ BR

by Lemma 4.3. Thus,

Z Z
C1 2
|D(ηu)| ≤ 2 u2 ,
BR R BR \BR/2

where
 Λ 2
C1 = 18(1 + 4 ).
λ

Now, by the Poincaré inequality, since η vanishes on ∂BR ,

Z Z
2 2
(ηu) ≤ c(n)R |D(ηu)|2 ,
BR BR

where c(n) is a constant depending only on the dimension. How can we remember that we get

the R2 factor in the Poincare inequality? It then follows that

Z Z
(ηu)2 ≤ c(n)C1 u2 .
BR BR \BR/2

Since η = 1 in BR/2 we get that

Z Z
u2 ≤ c(n)C1 u2 .
BR/2 BR \BR/2

Thus,
Z Z
C1 c(n)
2
u ≤ u2 .
BR/2 C1 c(n) + 1 BR

24
Therefore, the first part of the theorem works with

C1 c(n)
θ1 = < 1.
C1 c(n) + 1

Now we move to the second inequality. To achieve this, we apply Corollary (4.5) with r = R/2
R
to u − a, where a = −BR \B u. Then we have:
R/2

Z Z  Λ 2 Z
64  Λ 2
2 2
|Du| ≤ 2 |u − a| ≤ 64c2 (n) |Du|2 .
BR/2 R λ BR \BR/2 λ BR \BR/2

 2
Λ
64c2 (n) λ
Thus, we can take θ2 =  2 , where c2 (n) is the second Poincaré constant.
Λ
64c(n) λ
+1

5 Maximum Principles for General Elliptic Problems

Now we consider general elliptic operators:

Lu := aij (x)Dij u + bi (x)Di u + c(x)u.

We only assume that aij , bi , c are bounded and continuous in Ω̄ and that aij is elliptic. One of the

purposes of this section is to use maximum principles to prove a-priori estimates and establish

symmetry results. Another nice application will be the Alexandroff Maximum Principle.

Remark 5.1. A consequence of ellipticity is that aii ≥ λ for each i, just by choosing ξ = ei in

the definition of ellipticity.

Lemma 5.2. If u ∈ C 2 (Ω) ∩ C(Ω̄) and Lu > 0 in Ω and c ≤ 0, then u cannot attain a

non-negative maximum in Ω. In other words,

sup u ≤ sup u+
Ω ∂Ω

Proof. Assume it attains a non-negative maximum at x = x∗ . Then, aij (x∗ )∂ij u(x)∗ ≤ 0. Why

is this? Moreover, bi ∂i u(x∗ ) = 0 and c(x∗ )u(x∗ ) ≤ 0.

25
Remark 5.3. Is the condition c ≤ 0 necessary?

Next we have the weak maximum principle:

Theorem 10. If u ∈ C 2 (Ω) ∩ C(Ω̄) and Lu ≥ 0 in Ω and c ≤ 0, then u cannot attain a

non-negative maximum in Ω. Then, u attains its non-negative maximum on ∂Ω.

Proof. The proof is somewhat similar to the case of the Laplacian done in the previous section.

Note that a11 ≥ λ. Define hα (x) = exp(αx1 ). We are going to choose α very large to magnify

the elliptic term. Indeed,

Lhα = exp(αx1 )(α2 a11 (x) + αb1 (x) + c(x)) ≥ exp(αx1 )(α2 λ + αb1 (x) + c(x))

≥ exp(αx1 )(α2 λ + αB + C),

where B = |b|L∞ , C = |c|L∞ . Now if α is large, we get that Lhα > 0. Next, define

w = u + hα .

We have that

Lw > 0,

for all  > 0. Thus, by the preceding Lemma, we have

sup w ≤ sup w+
Ω ∂Ω

Consequently,

sup u ≤ sup w ≤ sup w+ ≤ sup u+ +  sup exp(αx1 ),

Ω Ω ∂Ω ∂Ω x∈∂Ω

for all  > 0.

Remark 5.4. This implies uniqueness for the Dirichlet problem:

Lu = f, in Ω

26
 u = g, on ∂Ω.

What happens when c ≥ 0?

To get uniqueness for the Neumann problem, we need the Hopf Lemma.

Theorem 11. Let B be an open ball in Rn and assume that u ∈ C 2 (B) ∩ C(B \ {x0 }) satisfies

Lu ≥ 0 in B and c(x) ≤ 0 in B. Assume also that

u(x) < u(x0 )

for any x ∈ B and u(x0 ) ≥ 0. Then, for any outward direction ν at x0 with ν · n > 0, we have:

1
lim inf (u(x0 ) − u(x0 − tν)) > 0.
t→0 t

Remark 5.5. That the limit is ≥ 0 is obvious from the assumptions but all the ≥ 0 statement

requires is that u(x0 ) ≥ u(x) for all x rather than the strict inequality. The assumption of

having a strict inequality is a key part of the proof.

Proof. We may assume that B = Br (0), u ∈ C(B̄), and that u(x) < u(x0 ) for all x ∈ B̄ \ {x0 }.

Why? Just as before, our goal will be to perturb u a little bit so that we keep the maximum

at x0 but so that the normal derivative changes. Note that since u(x) < u(x0 ) throughtout

B̄ \ {x0 },

M := sup u(x) < u(x0 ).

x∈Br/2

As in the proof of the Hopf Lemma for the Laplacian, let us define:

h(x) = exp(−α|x|2 ) − exp(−αr2 ).

Note that h vanishes on ∂Ω. Let us Claim that Lh > 0 in Br \ Br/2 so long as α is sufficiently

large. Moreover, h(∂B) = 0. Hence, if we take

v = u + h,

27
we will have that v(x) ≤ v(x0 ), for all x ∈ ∂Br/2 , so long as  < u(x0 ) − M . Thus, by the weak

maximum principle, we have that

v(x) ≤ v(x0 ),

for all x ∈ Br \ Br/2 . Therefore,

v(x0 ) − v(x0 − tν)

lim inf ≥ 0.
t→0 t

Thus,
v(x0 ) − v(x0 − tν)
lim inf ≥ −∂ν h(x0 ).
t→0 t

However, ∂ν h(x0 ) < 0 by construction since ν · n(x0 ) > 0.

It remains now to prove the Claim:

 
Lh = exp(−α|x|2 ) 4α2 aij (x)xi xj − 2αaii (x) − 2αbi (x)xi + c − c exp(−αr2 ) − c(x) exp(−αr2 )

 
≥ exp(−α|x|2 ) 4α2 λ|x|2 − 2αaii (x) − 2αbi (x)xi + c ≥ exp(−α|x|2 )(α2 λr2 − (α + 1)K),

for some fixed K depending only on r, c, bi , Λ. Thus, taking α large, we get that Lh > 0.

A nice corollary of the Hopf Lemma is the Strong Maximum Principle (which then justifies

the assumptions in the Hopf Lemma).

Theorem 12 (Strong Maximum Principle). Assume u ∈ C 2 (Ω) ∩ C(Ω̄) satisfies Lu ≥ 0 with

c ≤ 0 in Ω. Then, the nonnegative maximum of u in Ω̄ is assumed only on ∂Ω unless u is

constant.

Proof. As in the proof for harmonic functions, let us define the set

Σ = {x ∈ Ω : u(x) = M },

where M is the non-negative maximum of u on Ω̄. Our goal is to show that Σ is open and closed

in Ω. How did we conclude this in the case of harmonic functions?

28
 In this general case, assume toward a contradiction that Σ is a non-empty proper subset of

Ω. Note that since Ω is open and connected, it is also path connected. Note that Ω \ Σ is open.

Now take x1 ∈ Ω \ Σ and x2 ∈ Σ. Draw a continuous path inside of Ω from x1 to x2 and let x∗

be the first time this path hits Σ. Since x∗ ∈ Ω, there is a small ball B1 around x∗ which lies in

Ω. This ball contains members of Σ and Ω \ Σ by construction. Thus, we may take y ∈ B1 for

which δ := d(y, Σ) < d(y, ∂Ω). Thus, there is an x ∈ Σ ∩ B̄δ satisfying the conditions of the Hopf

Lemma. It follows that ∂n u(x) > 0. But ∇u(x) = 0 since x∇Σ. This is a contradiction.

Next, we will establish a comparison principle due to Serrin.

Theorem 13. Suppose u ∈ C 2 (Ω) ∩ C(Ω̄) satisfies Lu ≥ 0 If u ≤ 0 in Ω, then either u < 0 in

Ω or u ≡ 0 in Ω.

Proof. Suppose u(x0 ) = 0 for some x0 ∈ Ω. We will prove that u ≡ 0 in Ω. Write c = c+ − c− .

Then, if Lu ≥ 0, we have

aij Dij u + bi u − c− u ≥ −c+ u ≥ 0.

Now we apply the strong maximum principle since −c− ≤ 0.

Now we prove a general maximum principle in the case of general c.

Theorem 14. Suppose there exists w ∈ C 2 (Ω) ∩ C 1 (Ω̄) with w > 0 in Ω̄ and Lw ≤ 0 in Ω. If
u
u ∈ C 2 (Ω) ∩ C 1 (Ω̄) satisfies Lu ≥ 0 in Ω, then w cannot assume its non-negative maximum in
u u
Ω unless w ≡ constant. Additionally, if w assumes its non-negative maximum at x0 ∈ ∂Ω and
u
w 6≡ constant, then
u
∂ν (x0 ) > 0.
w
u
Proof. Set v = w. Then, v satisfies

 Lw 
aij Dij v + Bi Di v + v ≥ 0,
w

29
for some Bi . Indeed,

u 1 2 1
aij Dij = aij Dij u − 2 aij Di uDj w + uaij Dij
w w w w

1 2 Dj w
= aij Dij u − 2 aij Di uDj w − uaij Di 2
w w w
1 u Di Dj w 2 1
= aij Dij u − aij − 2 aij Di uDj w + 2uaij 3 Di wDj w
w w w w w

Moreover,
u 1 u Di w
Di = Di u − .
w w w w

Let’s define
1
bˆi = aij Dj w
w

It follows that
u u Lw
aij Dij + Bi Di + v=
w w w
1 u Di Dj w ˆ 2 2u  u Di w  u 
 1 
aij Dij u− aij +bi − Di u+ 2 Di w +Bi Di u− + 2 aij Dij w+bi Di w +cv,
w w w w w w w w w

Now we define:

Bi := 2b̂i + bi .

Then we see:
 Lw  1
aij Dij v + Bi Di v + v= Lu ≥ 0.
w w
Lw
Note that w ≤ 0. This means we can apply the strong maximum principle and Hopf Lemma

to v.

Remark 5.6. This implies that L has a comparison principle. If Lu ≥ 0 and u ≤ 0 on ∂Ω then

u ≤ 0 on Ω.

Next, we establish the maximum principle in “thin-domains.”

Proposition 5.7. Assume that |(x − y) · e| < d for all x, y ∈ Ω, where e is a unit vector and

30
d > 0. There exists d0 > 0 depending on λ, |bi |L∞ , |c+ |L∞ so that if d ≤ d0 , the assumptions of

the preceding Theorem are satisfied.

Proof. Without a loss of generality, let us assume that e = (1, 0, ..., 0) and that Ω̄ ⊂ {0 < x1 <

d}. Assume that sup |bi |L∞ , |c+ |L∞ ≤ N . Define w = eαd − eαx1 > 0 in Ω̄. Observe that

Lw = eαx1 (−a11 α2 − b1 α) + c(eαd − eαx1 ) ≤ −α2 λ + αN + N eαd .

Now we choose α large, depending only on N and λ, so that −α2 λ + αN < −4N . It follows

that

Lw < −4N + N eαd .

1
Now we just take d0 = α and we are done.

An example of this is the following fact:

∂xx + k 2 : H 2 → L2

2π
is an isomorphism on spatially periodic functions of period strictly less than k .

Remark 5.8. Later we will show that if the diameter of Ω is just bounded (not necessarily

small) but if the volume |Ω| is small, then we again have a comparison principle.

5.1 A-Priori Estimates Through the Maximum Principle

We first have the L∞ bound.

Proposition 5.9. Suppose u ∈ C 2 (Ω) ∩ C(Ω̄) satisfies

Lu = f in Ω

u=φ on ∂Ω,

31
for some f ∈ C(Ω̄) and φ ∈ C(∂Ω). If c ≤ 0, then we have:

|u|L∞ ≤ |φ|L∞ + C|f |L∞ ,

where C = C(λ, Λ, diam(Ω)).

Proof. Define the positive constants F, Φ by F = |f |L∞ and Φ = |φ|L∞ . Now, suppose that we

have a w such that

Lw ≤ −F in Ω

w≥Φ on ∂Ω.

It would then follow by the comparison principle that L(w + u) ≤ 0 and w + u ≥ 0 on ∂Ω. From

which it implies that w ≥ −u everywhere. Similarly for −w + u so that w ≥ u everywhere. It

follows then that

w ≥ |u|

. Consequently, it suffices to find a w with the advertised bound.

Suppose that Ω ⊂ {0 < x1 < d}. Note that we don’t assume that d is small. Set

w = Φ + (eαd − eαx1 )F,

where we will again choose α large later. then,

−Lw = (a11 α2 + b1 α)F eαx1 − cΦ − c(eαd − eαx1 )F

≥ (λα2 + b1 α)F ≥ F

if α is sufficiently large. Now note that

|w| ≤ Φ + eαd F,

32
from which the result follows by the comparison principle as we discussed before.

Now let us move to the Gradient Estimates. Let

L = aij Dij + bi Di .

Before that, let us prove three useful Lemmas.

Lemma 5.10.

L(u2 ) ≥ 2λ|Du|2 + 2uLu.

Proof.

L(u2 ) = 2uLu + aij Di uDj u ≥ 2uLu + λ|Du|2 .

Lemma 5.11.
1
L(|Du|2 ) ≥ λ|D2 u|2 − C|Du|2 − |D(Lu)|2
2

Proof.

Di (|Du|2 ) = 2Dk uDki u

So that

Dij (|Du|2 ) = 2Dkj uDki u + 2Dk uDkij u.

Thus,

L(|Du|2 ) = aij Dij |Du |2 + bi Di |Du|2

= 2aij Dkj uDki u + 2Dk uaij Dijk u + 2bi Dk uDki u

≥ λ|D2 u|2 + 2Dk u(aij Dijk u + bi Dik u))

33
 ≥ λ|D2 u|2 + 2Dk u(Dk (aij Dij u) − Dk aij Dij u) − C|Du||D2 u|

≥ λ|D2 u|2 Dk u(Dk (Lu − bi Di u)) − C|Du||Du|2

≥ λ|D2 u|2 − C|Du||D2 u| − C|Du|2 − |Du||D(Lu)|

1
≥ λ|D2 u|2 − C|Du|2 − |D(Lu)|2 .
2

Proposition 5.12. Suppose u ∈ C 3 (Ω) ∩ C 1 (Ω̄) satisfies

aij Dij u + bi Di u = f (x, u)

with aij , bi ∈ C 1 (Ω̄), f ∈ C 1 (Ω̄ × R). Then,

|Du|L∞ ≤ |Du|L∞ (∂Ω) + C,

with C a constant depending on λ, diam(Ω), |aij , bi |C 1 , |u|L∞ , |f |C 1 .

Proof. Let

L = aij Dij + bi Di .

Our goal will be to show that

L(|Du|2 + αu2 + eβx1 ) > 0

for properly chosen α, β.

34
 Next, we have the interior gradient estimate.

Proposition 5.13. Under the same assumptions as the previous proposition, if Ω0 ⊂⊂ Ω then

sup |Du| ≤ C,
Ω0

where C is as in the last proposition.

Proof. Let γ ∈ C0∞ (Ω) and γ ≥ 0. Consider the function

w = γ|Du|2 + α|u|2 + eβx1 .

We will again show that Lw ≥ 0 and apply the maximum principle.

Set v = γ|Du|2 . Observe that

Lv = Lγ|Du|2 + γL(|Du|2 ) + 2aij Di γDj |Du|2 .

As before, we have that

L(|Du|2 ) ≥ λ|D2 u|2 − C|Du|2 − C.

Thus,

Lv ≥ λγ|D2 u|2 − Λ|Dγ|D2 u| − C|Du|2 − C.

Thus,
1 λγ
Lv ≥ λγ |D2 u|2 + ( − Λ2 |Dγ|2 )|D2 u|2 − C|Du|2 − C .
2 2

Now, assuming that

|Dγ|2 ≤ Cγ

in Ω. How do we know that we can assume this? Now we proceed as before.

35
5.2 Boundary Estimates

The preceding results were on interior estimates (or global estimates assuming boundary esti-

mates). Now we will show how to establish boundary estimates. We will assume that u satisfies

the following problem:

Lu = f (x, u) in Ω,

u = φ on ∂Ω,

where we assume that

• f ∈C

• Ω satisfies the uniform exterior sphere condition

• φ ∈ C 2 (Ω̄).

• L is uniformly elliptic

Proposition 5.14. Under the above assumptions, u ∈ C 2 (Ω) ∩ C(Ω̄) satisfies:

|u(x) − u(x0 )| ≤ C|x − x0 |,

for all x ∈ Ω and x0 ∈ ∂Ω, where C depends only on the sup norm of f, aij , b, u and λ and |φ|C 2 .

Proof. By replacing u → u + φ, we may assume that u = 0 on ∂Ω. We may also assume that

L = aij Dij + bi Di .

Let F = |f |L∞ and note that

L(±u) ≥ −F.

Fix x0 ∈ ∂Ω. We will construct a function w such that

Lw ≤ −F

36
in Ω while w(x0 ) = 0 and w|∂Ω ≥ 0. Once we can show that the Lipschitz norm of w is bounded

by C as in the statement, we will be done.

Consider the exterior ball BR (y) with BR (y) ∩ Ω̄ = {x0 }. Define d(x) = |x − y| − R for any

x ∈ Ω. Consider w = ψ(d), where ψ will be chosen so that

ψ(0) = 0

ψ(d) > 0, d > 0

|ψ 0 | ≤ C∗ .

Let us also require that ψ 0 > 0 and ψ 00 < 0. Now let’s compute Lw for x ∈ Ω:

Lw = ψ 00 aij Di dDj d + ψ 0 aaij Dij d + ψ 0 bi Di d.

Now,
xi − yi
Di d(x) = ,
|x − y|

δij (xi − yi )(xj − yj )

Dij d(x) = − .
|x − y| |x − y|3

Thus,
aii aij (xi − yi )(xj − yj ) nΛ λ
aij Dij d ≤ − 3
≤ −
|x − y| |x − y| |x − y| |x − y|
nΛ − λ
≤ .
R

Thus,
 nΛ − λ 
Lw ≤ ψ 00 λ + ψ 0 + |b|L∞ .
R

Thus, in order to ensure that

Lw ≤ −F,

37
we need
nΛ − λ
λψ 00 + ψ 0 ( + |b|) + F ≤ 0.
R

Now let us take ψ to satisfy

ψ 00 + Aψ 0 + B = 0.

The solution is given by

B C1 C2 −Ad
ψ(d) = − d+ − e ,
A A A

where we are allowed to choose C1 and C2 as we please. Since ψ(0) = 0, we need C1 = C2 .

Hence,
B C
ψ(d) = − d + (1 − e−Ad ).
A A

Observe that
B
ψ 0 = Ce−Ad −
A

and

ψ 00 = −ACe−Ad .

In order that ψ 0 (d) > 0, we take C = B aD

Ae , where D is the diameter of Ω. We are now done.

5.3 Exercises

1. Assume that K is a self-adjoint compact operator on L2 (like the operator we discussed

in the first homework assignment).

• Show that eigenvalues of K must be real and that eigenvectors with different eigen-

values are orthogonal.

• Let M := sup|v|=1 |(Kv, v)|. Show that there exists v∗ with |v∗ | = 1 satisfying so that

|(Kv∗ , v∗ )| = M .

• Show that v∗ is an eigenvector of K.

38
 • Conclude that K has a complete basis of eigenvectors.

• Given q ∈ C([0, 1]), prove that there exists an orthonormal basis en of L2 ([0, 1])

satisfying:

e00n + qen = λn en

en (0) = en (1) = 0.

2. Find a function ψ ∈ C ∞ (R3 ) with lim|x|→∞ ψ(x) = 0 and ∆ψ > 0 on R3 . Is this possible

in R2 ?

3. Assume that ∆ψ = 0 on B1 (0) and that ψ and ∂n ψ both vanish on a segment of ∂B1 (0).

Show that ψ ≡ 0.

4. Find a function φ ∈ Cc1 (R2 ) so any solution to

∆ψ = 0 in y > 0,

ψ = φ on y = 0,

satisfies |∂y φ(0, y)| → ∞ as y → 0. Extra credit: show that if φ ∈ C 1,α for some α > 0,

then this is not possible.

5. Find a solution to ∆ψ = 1 on Ω and ψ = 0 on ∂Ω, where Ω is any sector in R2 (including

the slit domain R2 \ {(x, 0) : x > 0}.

6 The C α and C 1,α Theory

We will now study the basics of the Schauder theory. Here, we will assume that the ellipticity

coefficients aij are continuous. We will only consider equations in the following form:

−Dj (aij (x)Di u) + c(x)u = f.

39
Recall that this equation has nice energy properties and we will use these to derive regularity

results on solutions u of the above equation. We will start by showing how local energy control

leads to Hölder regularity in general.

6.1 General results on Morrey-Campanato Spaces

Theorem 15. Suppose u ∈ L2 (Ω) satisfies:

Z
|u − ux,r |2 ≤ M 2 rn+2α ,
Br (x)

for any Br (x) ⊂ Ω, for some α ∈ (0, 1). Then, u ∈ C α (Ω) and for any Ω0 ⊂⊂ Ω there holds

|u(x) − u(y)|  
sup |u| + sup ≤ c M + |u| L2 ,
Ω0 x,y∈Ω0 |x − y|α

where c = c(n, α, Ω, Ω0 ).

Proof. Let R0 = d(∂Ω, Ω0 ). For x0 ∈ Ω0 and r1 < r2 ≤ R0 we have:

 
|ux0 ,r1 − ux0 ,r2 |2 ≤ 2 |u(x) − ux0 ,r1 |2 + |u(x) − ux0 ,r2 − u(x)|2

Averaging over Br1 (x0 ) gives:

|ux0 ,r1 − ux0 ,r2 |2 ≤ c(n)M 2 r1−n (r1n+2α + r2n+2α ).

R R
Thus, if r1 = 2i+1
and r2 = 2i
, we see that

|ux0 ,2−(1+i) R − ux0 ,2−i R | ≤ c(n)2−iα M Rα .

Thus, for h < k, we have

c(n, α)
|ux0 ,2−h R − ux0 ,2−k R | ≤ M Rα .
2hα

40
Note that c(n, α) blows up as α → 0. This means that {ux0 ,2−i R } is Cauchy and its limit is

independent of R. Call the limit û(x0 ) It follows that for every r we have that

|ux0 ,r − û(x0 )| ≤ c(n, α)M rα .

This implies that u is uniformly bounded by CM Rα + ux,R for any x ∈ Ω0 , while the latter can

be estimated by the L2 norm of u. Next, let us give the bound on the Hölder semi-norm of u.
R0
Let x, y ∈ Ω0 and let R = |x − y| < 2 . Then,

|u(x) − u(y)| ≤ |u(x) − ux,2R | + |u(y) − uy,2R | + |ux,2R − uy,2R |.

The first two are bounded by CRα as before. To estimate the last one we write

|ux,2R − uy,2R |2 ≤ 2|ux,2R − u(ζ)|2 + 2|uy,2R − u(ζ)|2

and we average over ζ ∈ B2R (x) ∩ B2R (y). Noting that BR (x) ⊂ B2R (x) ∩ B2R (y), we get

|ux,2R − uy,2R |2 ≤ c(n, α)M 2 R2α .

Now the result follows.

Next we have Morrey’s inequality

Corollary 6.1. Suppose u ∈ H 1 (Ω) and that

Z
|Du|2 ≤ M 2 rn−2+2α
Br (x)

for some α ∈ (0, 1) and for all Br (x) ⊂ Ω. Then u ∈ C α and for any Ω0 ⊂⊂ Ω we have the

bound as in the previous theorem.

41
Proof. By the Poincaré inequality,

Z Z
|u − ux,r |2 ≤ c(n)r2 |Du|2 ≤ c(n)M 2 rn+2α .
Br (x) Br (x)

Next, we will need the following result.

Lemma 6.2. Suppose that u ∈ H 1 (Ω) satisfies

Z
|Du|2 ≤ M rµ ,
Br (x0 )

for any Br (x0 ) ⊂ Ω, for some µ ∈ [0, n). Then, for any Ω0 ⊂⊂ Ω there holds

Z
|u|2 ≤ c(n, λ, µ, Ω, Ω0 )(M + |u|2L2 )rλ ,
Br

where λ = µ + 2 if µ < n − 2 and λ ∈ [0, n) if n − 2 ≤ µ < n.

Proof. Note that λ < n in all cases. By the Poincaré inequality, we have that

Z
|u − ux0 ,r |2 ≤ c(n)M rµ+2 ≤ c(n)M rλ .
Br (x0 )

Now, for ρ < r ≤ R0 we have

Z  ρ n Z
2 λ n 2 λ
u ≤ 2c(n)M r + c(n)ρ |ux0 ,r | ≤ 2c(n)M r + c(n) u2 ,
Bρ r Br

where we used the Cauchy-Schawrz inequality in the second inequality and the triangle inequality

in the first. Thus, the function

Z
φ(r) = u2
Br (x0 )

satisfies:
  ρ n 
φ(ρ) ≤ C M rλ + φ(r) ,
r

42
for any ρ < r ≤ R0 and for some λ ∈ (0, n). If we could replace rλ on the right by ρλ we would

be done by taking r = R0 . Now we use the following Lemma.

Lemma 6.3. Let φ(t) be nonnegative and nondecreasing and assume that

 ρ 
φ(ρ) ≤ A ( )α +  φ(r) + Brβ ,
r

whenever ρ ≤ r ≤ R. Then, for any γ ∈ (β, α) there is a constant 0 so that if  < 0 we have

 ρ 
φ(ρ) ≤ c ( )γ φ(r) + Bρβ .
r

Proof. Fix τ ∈ (0, 1) and r < R. We have

φ(τ r) ≤ Aτ α (1 + τ −α )φ(r) + Brβ .

Choose τ < 1 so that 2Aτ α = τ γ and assume 0 τ −α < 1. Then, for r < R

φ(τ r) ≤ τ γ φ(r) + Brβ

and therefore,

φ(τ k+1 r) ≤ τ γ φ(τ k r) + Bτ kβ rβ

k
X
≤ τ (k+1)γ φ(r) + Bτ kβ rβ τ j(γ−β)
j=0

Bτ kβ rβ
≤ τ (k+1)γ φ(r) + .
1 − τ γ−β

Now we choose k so that τ k+2 r < ρ ≤ τ k+1 r so that

1  ρ γ Bρβ
φ(ρ) ≤ φ(r) + .
τγ r τ 2β (1 − τ γ−β )

43
6.2 C α bounds for u

As we explained before in class, the key idea is that when the coefficients are continuous we

can compare the solution to the solution to the constant-coefficient problem. Let us start by

recalling the basic estimate for harmonic functions. We have already established this (which

follows from the interior regularity results, using the surface mean-value property), but let us

recall the basic estimate:

Lemma 6.4. Suppose {aij } is a constant positive definite matrix with

λ|ξ|2 ≤ aij ξi ξj ≤ Λ|ξ|2 ,

for all ξ ∈ Rn . Suppose w ∈ H 1 (Br (x0 )) is a weak solution of

aij Dij w = 0

in Br (x0 ). Then, if ρ ≤ r, we have:

Z  ρ n Z
2
|Dw| ≤ c |Dw|2
Bρ (x0 ) r Br (x0 )

Z  ρ n+2 Z
|Dw − (Dw)x0 ,ρ |2 ≤ c |Dw − (Dw)x0 ,r |2
Bρ (x0 ) r Br (x0 )

Proof. Exercise.

Corollary 6.5. Suppose w is as in the preceding Lemma. Then, for any u ∈ H 1 (Br (x0 )) there

holds for 0 < ρ ≤ r,

Z  ρ
Z Z 
n
|Du|2 ≤ c |Du|2 + |D(u − w)|2
Bρ (x0 ) r Br (x0 ) Br (x0 )

and

Z  ρ
Z Z 
2 n+2 2
|Du − (Du)x0 ,ρ | ≤ c |Du − (Du)x0 ,r | + |D(u − w)|2
Bρ (x0 ) r Br (x0 ) Br (x0 )

44
Proof.

Now we will use the preceding two results to establish the following Theorem.

Theorem 16. Let u ∈ H 1 (B1 ) be a weak solution to

Dj (aij Di u) + cu = f

Assume that aij ∈ C(B̄1 ) with modulus of continuity τ and that c ∈ Ln (B1 ), f ∈ L1 (B1 ) for some

q ∈ (n/2, n). Then, u ∈ C α (B1 ) with α = 2− nq . Moreover, there exists an R0 = R0 (λ, Λ, τ, |c|Ln )

such that for any x ∈ B1/2 and r ≤ R0 there holds

Z  
|Du|2 ≤ Crn−2+2α |f |2Lq + |u|2H 1 ,
Br (x)

where C = C(λ, Λ, τ, |c|Ln ).

Proof. We will focus on the case n > 2. The case n = 2 should be checked separately (and is

easier). Let us fix Br (x0 ) ⊂ B1 and write:

Z Z Z
aij (x0 )Di wDj φ = f φ − cuφ + (aij (x0 ) − aij (x))Di uDj φ.
B1

Through usage of the Poisson kernel*, the problem

aij (x0 )Dij w = 0,

w = u on ∂Br (x0 ) has a unique solution. Thus we may choose φ = u − w in the following:

Z Z
Di vDj φ = f φ − cuφ(aij (x0 ) − aij (x))Di uDj φ.
Br (x0 )

Note that we have that

Z 2n
(n−2)/2n Z 1/2
v n−2 ≤ c(n) |Dv|2 .
Br Br (x0 )

45
 2n 2n
Observe that the conjugate exponent to n−2 is n+2 . Note that

1 1 n−2
+ + = 1.
n 2 2n

It follows using the Cauchy-Schwarz and Hölder inequalities,

Z  Z 
|Dv|2 ≤ c τ (r)2 |Du|2 + |c|2Ln |u|2L2 + |f |2L2n/(n+2) .
Br (x0 ) Br (x0 )

Thus, it follows that

Z Z

2 2 ρ
n 
|Du| ≤ c (τ (r) + ) |Du|2 + |c|2Ln |u|2L2 + |f |2L2n/(n+2) .
Bρ (x0 ) r Br (x0 )

Now, by the Hölder inequality,

|f |L2n/(n+2) ≤ |f |2Lq rn−2+2α ,

where α = 2 − n/q (again q ∈ (n/2, n)). Thus,

Z Z
2

2 ρ
n 
|Du| ≤ c (τ (r) + ) |Du|2 + rn−2+2α |f |2Lq + |c|2Ln |u|2L2 .
Bρ (x0 ) r Br (x0 )

Now let us complete the proof first in

Case 1: c ≡ 0.

We have for any Br (x0 ) ⊂ B1 and for any 0 < ρ ≤ r,

Z Z
2

2 ρ
n 2 n−2+2α 2

|Du| ≤ c (τ (r) + ) |Du| + r |f |Lq .
Bρ (x0 ) r Br (x0 )

Now we apply our technical Lemma and note that there is a R0 for which τ (r) as small as we

want for r < R0 and arrive at the result that

Z  ρ n−2+2α Z 
2
|Du| ≤ C |Du|2 + ρn−2+2α |f |2Lq .
Bρ (x0 ) r Br (x0 )

46
Applied at r = R0 , we arrive at the result.

Next, we consider the general case.

Case 2: General Case.

We have that

Z  ρ Z Z 
2 n 2 2 n−2+2α 2
|Du| ≤ C ( ) + τ (r) |Du| + r |f |Lq + u2 .
Bρ (x0 ) r Br (x0 ) Br

We will show that if x0 ∈ B1/2 we have for ρ < r ≤ 1/2 that

Z  ρ Z h
2 n 2

|Du| ≤ C ( ) + τ (r) |Du|2 + rn−2+2α |f |2Lq + |u|2H 1 .
Bρ (x0 ) r Br (x0 )

The proof will follow a simple bootstrap argument.

Indeed, if we apply Lemma (6.2) (with µ = 0), then we get:

Z Z Z
u2 ≤ Crδ1 ( u2 + |Du|2 ),
Br (x0 ) B1 B1

with δ1 = 2 if n > 2 and δ1 arbitrary in (0, 2) if n = 2. It follows that

Z  ρ Z h
2 n 2

|Du| ≤ C ( ) + τ (r) |Du|2 + max{rn−2+2α , rδ1 } |f |2Lq + |u|2H 1 .
Bρ (x0 ) r Br (x0 )

If n > 2 and n − 2 + 2α > 2, we arrive at the fact that

Z
|Du|2 ≤ Cr2 (|f |2Lq + |u|2H 1 ),
Br

from which we gain another power of 2 on u2 and so on.

Remark 6.6. A consequence of the result is that if ∆u = f in B1 and f ∈ Lq (B1 ), with

n/2 < q < n, then u ∈ C α with α = 2 − nq . Is this sharp? Check homogeneous examples.

47
6.3 C 1,α bounds for u

We again study weak solutions to the equation:

−Dj (aij Di u) + cu = f.

We now move to prove the Hölder continuity of Du, under the right assumptions on f and c.

This is the content of the following theorem.

Theorem 17. Let u ∈ H 1 (B1 ) be a weak solution to the above elliptic problem. Assume that
n
aij ∈ C α (B̄1 ) and c, f ∈ Lq (B1 ) with q > n. Let α = 1 − q ∈ (0, 1). Then, Du ∈ C α (B1 ).

Moreover, there exists R0 = R0 (λ, |aij |C α , |c|Lq ) so that for any x ∈ B1/2 and r ≤ R0 , there

holds
Z
|Du − (Du)x,r |2 ≤ Crn+2α (|f |2Lq + |u|2H 1 ),
Br (x)

C = C(λ, |aij |C α , |c|Lq ).

Remark 6.7. Check what happens when aij ∈ C.

Proof. As before, we decompose u into v + w where w satisfies the homogeneous equation with

constant coefficients. That is,

Z Z
aij (x0 )Di vDj φ = f φ − cuφ + (aij (x0 ) − aij (x))Di uDj φ.
Br Br (x0 )

As before, we take φ = v as a test function and obtain:

Z  Z 
|Dv|2 ≤ C τ (r)2 |Du|2 + |c|2Ln |u|2L2 (Br ) + |f |2L2n/(n+2) .
Br (x0 ) Br (x0 )

Now, remember that u = v − w, where w satisfies the constant coefficient homogeneous problem.

It follows that for 0 < ρ ≤ r we have

Z  ρ
Z 
2 n 2
|Du| ≤ C ( + τ (r) ) |Du|2 + |c|2Ln |u|2L2 (Br ) + |f |2L2n/(n+2)
Bρ r Br

48
and

Z  ρ
Z Z 
n+2
|Du−(Du)x0 ,ρ |2 ≤ C |Du−(Du)x0 ,r |2 +τ (r)2 |Du|2 +|c|2Ln |u|2L2 (Br ) +|f |2L2n/(n+2) .
Bρ r Br Br

Now, by the Holder inequality, because of the choice α = 1 − nq , we have:

|f |2L2n/(n+2) ≤ c(n)|f |2Lq rn+2α ,

while

|c|2Ln ≤ c(n)r2α |c|2Lq .

In the case where c ≡ 0 and aij (x) ≡ aij (x0 ), the result is clear since we then get that

Z  ρ
Z 
2 n+2
|Du − (Du)x0 ,ρ | ≤ C |Du − (Du)x0 ,r |2 + rn+2α |f |2Lq ,
Bρ r Br

from which the result follows using the technical Lemma. Next, let us look at the case

where aij are variable but c ≡ 0.

In this case, we have

Z  ρ
Z 
n
|Du|2 ≤ C ( + τ (r)2 ) |Du|2 + rn+2α |f |2Lq .
Bρ r Br

By the technical Lemma, if r is sufficiently small, we have:

Z
|Du|2 ≤ Crn−2δ (|f |2Lq + |u|2H 1 ).
Br

Now it follows that

Z  ρ
Z 
2 n+2
|Du − (Du)x0 ,ρ | ≤ C |Du − (Du)x0 ,r |2 + rn+2α−2δ (|f |2Lq + |u|2H 1 ) .
Bρ r Br

49
Now using the technical Lemma again gives us that Du ∈ C α−δ . It follows that

Z
|Du|2 ≤ c(n)rn ,
Br

from which it follows that Du ∈ C α since now

Z  ρ
Z 
n+2
|Du − (Du)x0 ,ρ |2 ≤ C |Du − (Du)x0 ,r |2 + rn+2α (|f |2Lq + |u|2H 1 ) .
Bρ r Br

We leave the general case as an exercise.

6.4 Exercises on the Newtonian Potential

The following problems give a different point of view on the regularity theory for the Laplacian

in Rd by studying properties of the Newtonian potential. It turns out that for the scenarios

we studied in the preceding section, all the results can be established perturbatively from the

results we can get using the Newtonian potential. This approach is adopted, for example, in the

section on the Schauder Estimates in the book of Gilbarg and Trudinger.

1. Define:
cd
K(x) = ,
|x|d−2

when d 6= 2 and K(x) = c2 log |x| when d = 2. If f ∈ Cc∞ (Rd ), we define ψ by

Z
ψ(x) = K(x − y)f (y)dy.
Rd

Show that there are constants cd , c2 so that

∆ψ = f

on Rd .

• Observe that ∆K = 0 for x 6= 0.

50
 • Change variables in the integral so that all derivatives fall on f . By translation

invariance, observe that it suffices to show that ∆ψ(0) = f (0).

• Remove a ball of radius  around y = 0 in the integral (after differentiating twice in

x and evaluating at x = 0) and then proceed to integrate by parts twice in y.

• Send  → 0 and observe that all terms tend to zero except one boundary term which

tends to a (dimension-dependent) multiple of f .

• Find a formula for a solution to the problem:

∆ψ = f,

ψ(0) = 0, |∇ψ(x)| → 0, as |x| → ∞.

Is the solution unique?

• For d ≥ 3, find the formula for K using the Fourier transform (Hint: what must K̂(ξ)

be?)

• Use the preceding idea to find the kernels of the fractional Laplacian operators.

2. We now move to look at a-priori estimates for solutions. In the following, we will take

f ∈ Cc∞ (Rd ) and define

Z
ψ(x) = K(x − y)f (y)dy,
Rd

where K is as in the previous problem.

• Fix p < d < q. Prove that there is a constant Cp,q (independent of f ) so that

|∇ψ|L∞ ≤ Cp,q (|f |Lp + |f |Lq ).

• Show that the above cannot hold if p = q = d.

51
 • Show there is a universal C (independent of f ) so that

|D2 ψ|L2 ≤ C|f |L2 .

Hint: Deduce from the formula that ∇ψ, D2 ψ ∈ L2 and decay. Next, compute

(∆ψ)2 and use that ∆ψ = f . If you integrate by parts, make sure to justify why
R

this is allowed. Note that ψ, ∇ψ, D2 ψ do decay, but not very fast.

• Fix α > 0. Show that there is a constant Cα so that

|D2 ψ|L∞ ≤ Cα (|f |L1 + |f |C α ).

• Bonus: If 0 < α < 1, show that we can replace |D2 ψ|L∞ by |D2 ψ|C α in the preceding

estimate.

7 De Giorgi’s Theorem

Hilbert’s 19th problem was to establish the regularity of minimizers to functionals of the form:

Z
F (∇u),
Ω

when F is a smooth convex function. The problem is that the “Direct Method” of the Calculus

of variations (a compactness argument) shows only the existence of an H 1 “weak solution” to

the associated fully-nonlinear elliptic problem:

div(∇F (∇u)) = 0,

which is just

∂ij F (∇u)∂ij u = 0.

52
By differentiating the equations and using the Schauder theory, it is possible to deduce C ∞

smoothness of u once we know, for example, that u ∈ C 1,α . All what we know from the energy

is that u ∈ H 1 . The De Giorgi-Nash theorem consists of establishing first that any H 1 weak

solution to

aij Dij u = 0

is uniformly bounded by the L2 norm when the coefficients are merely bounded and measurable.

After that, we will establish C α regularity of u and C 1,α regularity in the case that the functional

F is smooth. Higher smoothness follows from the Schauder Theory. We will follow De Giorgi’s

proof.

7.1 From L2 to L∞

Theorem 18. Assume that aij ∈ L∞ (B1 ) and c ∈ Lq (B1 ) for some q > n/2 :

|aij |L∞ + |c|Lq ≤ Λ,

while

aij ξi ξj ≥ λ|ξ|2 .

Suppose that u ∈ H 1 is a weak subsolution in the sense that

Z
aij Di uDj φ + cuφ ≤ fφ

for any φ ∈ H01 (B1 ) and φ ≥ 0 in B1 . If f ∈ Lq (B1 ), then u+ ∈ L∞

loc (B1 ). Moreover, we have,

for any p > 0:

 1 
sup u+ ≤ C |u+
|Lp (B ) + |f |Lq (B ) ,
1 1
Bθ (1 − θ)n/p

with

C = C(n, λ, Λ, p, q).

53
Proof. The proof is somewhat complicated and involves a number of steps. First, define:

v = (u − k)+ ,

where k ≥ 0. Let ζ ∈ C01 (B1 ). We will take 0 ≤ ζ ≤ 1. Take as a test function

φ = vζ 2 .

Observe that Dv = Du almost everywhere in {u > k} and v = 0, Dv = 0 in the set {u ≤ k}.

Hencewhen we take this choice of test function, the domain of integration will be the set {u > k}:

Z
aij Di uDj φ = aij Di uDj vζ 2 + 2aij Di uDj ζvζ

Z Z
≥λ |Dv|2 ζ 2 − 2Λ |Dv||Dζ|vζ

Λ2
Z Z
λ 2 2
≥ |Dv| ζ − 2 |Dζ|2 v 2 ,
2 λ

where we just used the Cauchy-Schwartz inequality. Note that this is just the same trick we

did in the proof of the Cacciopolli inequality. Thus, using the definition of subsolution, we have

that
Z Z Z Z Z
2 2 Λ  2 2 2 2 2

|Dv| ζ ≤ C( ) v |Dζ| + |c|v ζ + k |c|ζ + |f |vζ 2 ,
2
λ

where we just used that |u| ≤ k + v. Again we recall that vζ ∈ H01 (B1 ) so that we can apply

the Sobolev inequality:

|vζ|L2n/(n−2) ≤ c(n)|D(vζ)|L2 .

Thus, by the Hölder inequality

Z
1− n−2 − 1q
|f |vζ 2 ≤ |f |Lq |vζ|L2n/(n−2) |{}| 2n

1 q
+n − 1q
≤ c(n)|f |Lq |D(vζ)f |L2 |{vζ 6= 0}| 2

54
 c(n) 2 1− 1
≤ δ|D(vζ)|2L2 + |f |Lq |{vζ 6= 0}| q ,
δ

where we used that q > n/2. Consequently, we see:

Z Z Z
2 Λ 
2 2 2 2 2 1− 1q

|D(vζ)| ≤ C( , n) v |Dζ| + |c|v ζ + k |c|ζ 2 + F 2 |{vζ 6= 0}| ,
λ

with F = |f |Lq (B1 ). Now, we next observe using the same estimate as above on the f term

Z
1− n−2 − 1q
|c|v 2 ζ 2 ≤ c(n)|c|Lq |D(vζ)|2L2 |{vζ 6= 0}| n

2
− 1q
≤ c(n)|c|Lq |D(vζ)|2L2 |{vζ 6= 0}| n .

Similarly, by the Hölder inequality

Z Z
2
|c|ζ ≤ |c| ≤ |c|Lq |{vζ 6= 0}|1−1/q .

Consequently, if |{|vζ 6= 0}| is small enough (depending on n and |c|Lq ), we have that

Z Z
Λ2 2 2 2 2 1− 1q

|D(vζ)| ≤ C( , n) v |Dζ| + (k + F )|{vζ 6= 0}| . (7.1)
λ

Note that if k is large, we must have that |{vζ = 0}| is quite small since v ∈ L2 . Next, with

(7.1) in hand, we observe that by the Sobolev inequality again we have:

Z
(vζ)2 ≤ |vζ|2L2n/(n−2) |{vζ 6= 0}|2/n ≤ c(n)|D(vζ)|L2 |{vζ 6= 0}|2/n .

Now using (7.1), we have

Z Z 2
− 1q

2 1+ n
(vζ) ≤ C v 2 |Dζ|2 |{vζ 6= 0}|2/n + (k + F )2 |{vζ 6= 0}| .

55
In particular, if |{vζ 6= 0}| is small, we have that there is an  > 0 so that

Z Z 
(vζ)2 ≤ C v 2 |Dζ|2 |{vζ 6= 0}| + (k + F )2 |{vζ 6= 0}|1+ .

Choice of the Cut-off.

We now choose the cut-ff function. For fixed 0 < r < R ≤ 1 we choose ζ ∈ C0∞ (BR ) such
2
that ζ ≡ 1 in Br and 0 ≤ ζ ≤ 1 and |Dζ| ≤ R−r in B1 . Now set

A(k, r) = {x ∈ Br : u ≥ k}.

We thus have that

Z Z
2
1 
(u − k) ≤ C 2
|A(k, R)| (u − k)2 + (k + F )2 |A(k, R)|1+ .
A(k,r) (R − r) A(k,R)

The key here is that we have a freedom of choosing two scales when using this inequality, the

relation between r and R and k and h. Now observe that

Z
1 1 +
|A(k, r)| ≤ u+ ≤ |u |L2 . (7.2)
k A(k,R) k

1
In fact, we have that is bounded by k2
on the right hand side, but we do not need this extra

smallness I guess. Thus (7.1) holds so long as k ≥ k0 = C|u+ |L2 where C = C( Λλ , |c|Lq , n). Our

goal is to now show that there is a k large, precisely when k ≥ k∗ = C(k0 + F ) so that

Z
(u − k)2 = 0.
A(k, 12 )

This will imply that u+ ≤ k on B 1 . Now, let us take h > k ≥ k0 0 < r < 1. Then,
2

Z Z Z
2 2
(u − h) ≤ (u − k) ≤ (u − k)2 ,
A(h,r) A(h,r) A(k,r)

since if u ≥ h ≥ k, we have that (u − h)2 ≤ (u − k)2 gives the first inequality and the A(h, r) ⊂

56
A(k, r) gives the second. Moreover,

Z
1
|A(h, r)| ≤ (u − k)2 .
(h − k)2 A(k,r)

Thus, applying (7.2), we see that

Z Z
 1 
(u − h)2 ≤ C|A(R, h)| (u − h)2 + (h + F )2 |A(h, R)|
A(h,r) (R − r)2 A(h,R)

(h + F )2 
Z
 1 1  1+
≤C + (u − k)2 .
(R − r)2 (h − k)2 (h − k)2 A(k,R)

Hence, if we define φ(k, r) = |(u − k)+ |L2 (Br ) , for h > k ≥ k0 and R > r we have

 1 h+F 1
φ(h, r) ≤ C + φ(k, R)1+ . (7.3)
R−r h − k (h − k)

Note that φ is increasing in R and decreasing in k as we established before. Our goal is to now

show that
1
φ(k∗ , ) = 0
2

if k∗ is sufficiently large. We will do this be defining a sequence

1
kl = k0 + k(1 − )
2l

1 1
rl = + l+1 ,
2 2

where k will be chosen large. Observe that kl increases to k + k0 =: k∗ and rl decreases to 12 .

Let us thus apply (7.3) for (h, r) = (kl , rl ) and (k, R) = (kl−1 , rl−1 ). Observe that

1 k
rl−1 − rl = kl − kl−1 = .
2l+1 2l

Then we see:
C l(1+)  k0 + k + F 
φ(kl , rl ) ≤ 2 1 + φ(kl−1 , rl−1 )1+ .
k k

57
Thus,
C l(1+)
φ(kl , rl ) ≤ 2 φ(kl−1 , rl−1 )1+ .
k

Claim: There exists (a large) γ > 1 so that

φ(k0 , r0 )
φ(kl , rl ) ≤ ,
γl

for all l ≥ 0. Once the claim is established, we send l → ∞ and see that φ(k∗ , 12 ) = 0. We

establish this claim by induction on l. Clearly it is true for l = 0. Assume it is true for l − 1.

Now we have:

C l(1+) φ(k0 , r0 )1+ φ(k0 , r0 ) l(1+) γ 1+ φ(k0 , r0 )

φ(kl , rl ) ≤ 2 = C 2 .
k γ (l−1)(1+) k γ l γl

Remember that we still have to choose k and γ. We are done if we make a choice of k, γ so that

φ(k0 , r0 ) l(1+) γ 1+

C 2 ≤ 1.
k γ l

Choose γ first so that

γ  = 21+ .

Then choose k so that

φ(k0 , r0 ) 1+
C γ ≤ 1.
k
1
This completes the theorem in the case θ = 2 and p = 2.

Now we move to the general case 0 < θ < 1 and p > 0, which will follow essentially by scaling

and interpolation. The first thing to observe is that scaling down does not affect the ellipticity

constants. We thus arrive at the following:

 1 
sup u+ ≤ C |u+
| L2 (B ) + R
2−n/q
|f |Lq (B )
BR/2 Rn/2 R R

58
Now let us get an idea of how this works. By iterating this bound, it is thus easy to see that

 
sup u+ ≤ C 2nj/2 |u+ |L2 (B1 ) + |f |Lq (B1 ) .
B1−2−j

Thus means that for any θ ∈ (0, 1), we have:

 1 
sup u+ ≤ C |u+
| L2 (B ) + |f |Lq (B ) .
1
Bθ (1 − θ)n/2 1

In fact, by scaling in R again we have

 1 
sup u+ ≤ C |u+
| 2
L (BR ) + |f |L (B1 ) .
q
BθR ((1 − θ)R)n/2

Now, for p ∈ (0, 2) we have that:

2−p + p
|u+ |2L2 (BR ) ≤ |u+ |L∞ |u |Lp .

It follows by Young’s inequality that

1
|u+ |L2 ≤ δ 2/(2−p) |u+ |L∞ + |u+ |Lp ,
δ 2/p

since
p 2−p
+ = 1.
2 2

Thus,
1  1 
|u+ |L∞ (BθR ) ≤ |u+ |L∞ (BR ) + C |u+
| p
L (BR ) + |f |L (B1 ) .
q
2 ((1 − θ)R)n/p

Now let’s define f (t) = |f |L∞ (Bt ) . What we have shown is that

1 C
f (r) ≤ f (R) + |u+ |Lp (B1 ) + C|f |Lq (B1 ) .
2 (R − r)n/p

We will now show that we can in fact remove the first term on the right side.

59
Lemma 7.1. Let f ≥ 0 be bounded in [τ0 , τ1 ] with τ0 ≥ 0. Suppose for any τ0 < t < s ≤ τ1 we

have:
A
f (t) ≤ θf (s) + + B,
(s − t)α

for some θ ∈ [0, 1). Then we have that

 A 
f (t) ≤ c(α, θ) + B .
(s − t)α

Proof. Exercise.

This completes the proof of the Theorem.

7.2 From L∞ to C α

Consider the operator

Lu ≡ −Di (aij (x)Dj u).

Let us first define subsolutions and super solutions.

Definition 7.2. The function u ∈ H 1 (B1 ) is called a subsolution (supersolution) of the equation

Lu = 0

if for all φ ∈ H01 (B1 ) we have

Z
aij Di uDj φ ≤ 0 (≥ 0)
B1

for all φ ∈ H01 (B1 )d with φ ≥ 0.

We now observe the following nice Lemma (that we saw before in the case of the square

function)

0,1
Lemma 7.3. Let Φ ∈ Cloc (R) be convex. Then,

1. If u is a subsolution of Lu = 0 and Φ0 ≥ 0 then v = Φ(u) is also a subsolution provided

1 (B ).
v ∈ Hloc 1

60
 2. If u is a super solution and Φ0 ≤ 0 then v = Φ(u) is a subsolution provided v ∈ Hloc
1 (B ).
1

Remark 7.4. In both cases, v is a subsolution. It follows that for any k ∈ Z we have that

(u − k)+ is a subsolution when u is a subsolution.

2 (R). Then,
Proof. Assume Φ ∈ Cloc

Φ0 (s) ≥ 0, Φ00 (s) ≥ 0.

Let φ ∈ C01 (B1 ) with φ ≥ 0. Then,

Z Z Z
0 0
aij Di vDj φ = aij Φ (u)Di uDj φ = aij Di uDj (Φ (u)φ) − aij Di uDj uφΦ00 (u) ≤ 0,
B1 B1

since Φ0 (u)φ ≥ 0 and where we used ellipticity on the second term. The general case (Φ ∈ C 1 )

follows by approximation.

Next we have the following “Poincaré-Sobolev” inequality.

Lemma 7.5. For any  > 0, there exists C = C(, n) so that for any u ∈ H 1 with

|{x ∈ B1 : u = 0}| ≥ |B1 |,

we have
Z Z
u2 ≤ C(, n) |Du|2 .
B1 B1

Proof. Suppose not. Then, there exists  > 0 and a sequence un ∈ H 1 with

|{x ∈ B1 : un = 0}| ≥ |B1 |,

while
|Dun |L2 (B1 )
→ 0.
|un |L2 (B1 )

61
We may thus normalize un so that we have

|un |L2 = 1 |Dun |L2 → 0.

By the Poincaré inequality, we have that

|un − ūn |L2 → 0.

But we know that along a subsequence, un → c. Thus we have

|un − c|L2 → 0.

This contradicts the assumption.

Remark 7.6. Find a direct way to prove this result that gives us the dependence of C on  and

n.

7.2.1 Harnack Inequality

Next, we establish the following Lemma, which has the flavor of a Harnack inequality.

Lemma 7.7. Suppose u is a positive super solution in B2 with

|{x ∈ B1 : u ≥ 1}| ≥ |B1 |.

Then, there exists a constant C depending only on , n, Λλ so that

infB1/2 u ≥ C.

Proof. By adding a constant to u, we may assume that u ≥ δ > 0. Then we will send δ → 0.

By Lemma 4.6, if we define v = (logu)− , it is a subsolution and is bounded by log δ −1 . By the

62
L2 → L∞ bound we proved,

sup v ≤ C|v|L2 (B1 ) .

B1/2

By the preceding Lemma, we have

sup v ≤ C|Dv|L2 .
B1/2

We will now show that the right hand side is bounded (independent of δ). Choose as a test
ζ2
function u where ζ ∈ C01 (B2 ). Then we have:

ζ2
Z Z Z
aij Di uDj u ζaij Di uDj ζ
0≤ aij Di uDj =− ζ2 +2 .
u u2 u

This implies that

Z Z
2 2
ζ |D log u| ≤ C |Dζ|2 .

Thus, for fixed ζ ∈ C01 (B2 ) with ζ ≡ 1 in B1 , we have:

Z Z
2
|Dv| = |D log u|2 ≤ C.
B1 B1

Λ
Here, C only depends on λ and n. It now follows that

sup v ≤ C.
B1/2

This implies that

u ≥ exp(−C).

7.2.2 Oscillation Lemma

Finally, we will establish the Oscillation Theorem. First let us give a definition.

63
Definition 7.8. Fix u ∈ L∞ (Ω). Define the oscillation of u on Ω by:

oscΩ u = sup u − inf u.

Ω Ω

Remark 7.9. These are both defined in the “essential” supremum and infimum sense.

Theorem 19. Suppose that u is a bounded solution of Lu = 0 in B2 . Then, there exists a

γ = γ(n, Λλ ) ∈ (0, 1) so that

oscB1/2 u ≤ γ oscB1 u.

Proof. Let

α1 = sup u and β1 = inf u,

B1 B1

α2 = sup u and β2 = inf u.

B1/2 B1 /2

Observe that
1 u − β1 1
u ≥ (α1 + β1 ) ⇐⇒ ≥ ,
2 α1 − β 1 2
1 α1 − u 1
u ≤ (α1 + β1 ) ⇐⇒ ≥ .
2 α1 − β 1 2
α1 −β1
This might be confusing, but note that we could add or subtract 1 = α1 −β1 and it is clear. Now,

we know that for half of B1 one of the above two inequalities on the left must hold.

Case 1:
n 2(u − β1 ) o 1
x ∈ B1 : ≥ 1 ≥ |B1 |.
α1 − β 1 2

Now note that

u − β1
≥0
α1 − β1

on B1 is a solution. Thus, there is a C > 1 so that

u − β1 1
inf ≥ .
B1/2 α1 − β1 C

64
It follows that
1
β2 ≥ β1 + (α1 − β1 ).
C

Note now that

α2 ≤ α1 .

Thus we have:
1
(α2 − β2 ) ≤ (1 − )(α1 − β1 ).
C

In Case 2, we have
n 2(α1 − u) o 1
x ∈ B1 : ≥ 1 ≥ |B1 |.
α1 − β 1 2

In this case, we have

1
sup u ≤ α1 − (α1 − β1 )
B1/2 C

and we again get

1
(α2 − β2 ) ≤ (1 − )(α1 − β1 ).
C

Thus, we may take in the Theorem

1
γ =1− .
C

Corollary 7.10. Suppose u is an H 1 weak solution of Lu = 0 in B1 . Then, there holds

|u(x) − u(y)| Λ
sup |u| + sup α
≤ c(n, )|u|L2 (B1 ) .
B1/2 x6=y in B1/2 |x − y| λ

Proof. Exercise. Use the Oscillation Theorem and rescaling.

7.3 From C α to C 1,α to C ∞

We now show how the above Theorem of De Giorgi implies a positive answer to Hilbert’s 19th

problem on the regularity of minimizers of smooth convex functionals. Again, let F be a convex

65
functional. Then, any local minimizer satisfies

Z
∇F (∇u) · ∇φ = 0.

Now let’s assume that φ is compactly supported in B1 . Then, if h is very small, we have:

Z
∇F (∇u(· + h)) · ∇φ = 0.

It follows that
Z
[∇F (∇u(· + h)) − ∇F (∇u(·))] · ∇φ = 0,

which can be written as:

Z Z 1
D2 F (t(∇u(· + h)) + (1 − t)∇u(·))(∇u(· + h) − ∇u(·)) · ∇φ = 0.
0

R1
Now define à = 0 D2 F (t(∇u(· + h)) + (1 − t)∇u(·))dt. Then we see:

∇u(· + h) − ∇u(·)
Z
Ã( ) · ∇φ = 0.
h

Thus, the difference quotient is a weak solution for each h small (say we now restrict to B3/4 ).

It follows now from the De Giorgi Theorem that for each h small,

u(· + h) − u(·) u(· + h) − u(·)

| |C α (B1/2 ) ≤ C| |L2 ≤ C|u|H 1 .
h h

Since this is true for all h and C is independent of h, it follows in a not too difficult way that

|u|C 1,α (B1/2 ) ≤ C|u|H 1 .

From here, using the Schauder theory, it is easy to show that u is in fact C ∞ if F is.

66
7.4 Exercises

1. Fill in the proof of Lemma 7.1

2. Find the dependence of c on  and n in the Poincaré Sobolev inequality given in Lemma

7.5 (as alluded to Remark 7.6).

3. Show how the Oscillation Theorem implies the local C α regularity in De Giorgi’s Theorem.

4. Take a look at the paper of James Serrin “Pathological solutions of elliptic differential

equations.” Explain his examples and show how they relate to the De Giorgi theorem.

5. Let T be a compact self adjoint operator on a Hilbert space H and fix λ 6= 0. Show that

λI − T is injective if and only if λI − T is surjective. Discuss the solvability of the problem:

(λI − T )u = f,

for f ∈ H in light of this. Remember to consider carefully the case when the operator is

not injective. It is helpful to show first that the kernel must be finite dimensional.

6. (Calderón-Zygmund Estimates, I (reading)) Derivatives of Lp functions can be defined

using the notion of weak derivative, from integration by parts. For simplicity we just

define it formally in one dimension (in higher dimensions you just do the same for each

index). If f ∈ Lp , we say that g is the weak derivative of f if for all φ ∈ C01 we have

Z Z
gφ = − vφ0 .

In this case we write g = Df . It is not difficult to show that when f ∈ Lp ∩ C 1 , its weak

derivative is the same as its usual derivative. We let W k,p be the space of all functions

f ∈ Lp with Dα f ∈ Lp for all |α| ≤ k. Obviously Cc∞ ⊂ W k,p ⊂ Lp , where the inclusion

67
 is as sets. We may endow W k,p with the W k,p norm:

X
|f |W k,p = |Dα f |Lp .
|α|≤k

It is a standard result that the set W k,p endowed with this norm is actually a Banach

space for any k ∈ N and any 1 ≤ p ≤ ∞ and that this space is separable when p < ∞, the

class Cc∞ being dense in those cases. There are various embedding Theorems that follow

from the Sobolev and Morrey Inequalities that we will not discuss now. The goal of this

problem as well as the corresponding one on the next homework assignment will be to

establish the so-called Calderón-Zygmund estimates, namely that if ∆w = f and f ∈ Lp ,

then |D2 w|Lp ≤ Cp |f |Lp , when 1 < p < ∞. In a previous assignment we already saw the

case p = 2. This was just an integration by parts argument. In the next assignment we

will show it for p = 1 in a weak sense. Then, an interpolation argument will allow us to

conclude the result for all other values of p in the range. Read Sections 9.3 and 9.4

of Gilbarg and Trudinger’s book. A part of the next assignment will be devoted to

discussing some of the issues there in a slightly simpler setting.

8 Preliminaries on Viscosity Solutions: Tangent Paraboloids

and Second Order Differentiability

We now move to an explanation of the book of Cafarelli and Cabré. The general framework is

that regularity is derived by locally comparing functions to linear or quadratic functions.

Definition 8.1. A function L : Rn → R will be called affine if

L(x) = l0 + l · x,

with l0 ∈ R and l ∈ Rn . A paraboloid is any polynomial P : Rn → R of degree 2. Thus, P can

68
be written as:
1
P (x) = L(x) + (Ax, x),
2

with A = D2 P a constant symmetric matrix and L affine.

Definition 8.2. We say that P is a paraboloid of opening M whenever:

M 2
P (x) = L(x) ± |x| .
2

In the + case, P is convex and in the − case, P is concave.

Definition 8.3. Given two continuous functions u, v : A → R and x0 ∈ A, we say that v touches

u from above at x0 in A whenever

u(x) ≤ v(x), ∀x ∈ A,

u(x0 ) = v(x0 ).

Next,

Definition 8.4.

P+ (x0 , M ) := {P (x) = L(x) + M |x|2 } P− (x0 , M ) := {P (x) = L(x) − M |x|2 }

Definition 8.5. Let u ∈ C(Ω) and assume that A ⊂ Ω. For x0 ∈ A, we define

Θ̄(u, A)(x0 ) = inf{M : ∃P ∈ P+ (x0 , M ) touching v from above}. (8.1)

Θ(u, A)(x0 ) = inf{M : ∃P ∈ P− (x0 , M ) touching v from below}. (8.2)

Θ(u, A)(x0 ) = max{Θ̄(u, A)(x0 ), Θ(u, A)(x0 )}. (8.3)

Finally, let us define the second order difference quotient of u at x0 :

69
Definition 8.6. For u : Ω → R and h ∈ Rn

u(x0 + h) + u(x0 − h) − 2u(x0 )

∆2h u(x0 ) = ,
|h|2

for x ∈ Ω with x ± h ∈ Ω.

Remark 8.7. •
u(x0 +h)−u(x0 ) u(x0 )−u(x0 −h)
|h| − |h|
∆2h u(x0 ) =
|h|

• If L is affine, ∆2h L = 0.

• If P is a paraboloid P (x) = L(x) + 21 (Ax, x), then

1
∆2h P (x0 ) = (Ah, h),
|h|2

is independent of x0 . In particular, if P is of opening M , then

∆2h (P ) = M.

• If u is twice differentiable at x0 ,

∂ii u(x0 ) = lim ∆2δei u(x0 ).

δ→0

This follows from Taylor’s theorem and the preceding remarks.

We have the following simple Lemma.

Lemma 8.8. Assume that a paraboloid of opening M touches u from above. Then,

∆2h u(x0 ) ≤ M.

70
Proof. Since ∆2h L ≡ 0 for any affine L, we may assume that

1
u(x) ≤ M |x − x0 |2
2

and u(x0 ) = 0. Thus,

1
2 M |h|
2 + 12 M |h|2
∆2h u(x0 ) ≤ = M.
|h|2

Corollary 8.9. For any x0 ∈ Ω with B||h| (x0 ) ⊂ Ω, we have:

−Θ(u, B|h| (x0 ))(x0 ) ≤ ∆2h u(x0 ) ≤ Θ̄(u, B|h| (x0 ))(x0 ).

In particular,

|∆2h u(x0 )| ≤ Θ(u, B|h| (x0 ))(x0 ).

We have already seen above that control of Θ thus gives pointwise control of ∂vv u, for any

fixed vector v ∈ Rn . Now we want to see how control of Θ in Lp actually gives Lp control of

D2 u.

Proposition 8.10. Let 1 < p ≤ ∞ and u ∈ C(Ω). Let  > 0 and define

Θ(u, )(x) := Θ(u, Ω ∩ B (x))(x),

for any x ∈ Ω. If Θ(u, ) ∈ Lp (Ω), then D2 u ∈ Lp (Ω). In fact,

|D2 u|Lp (Ω) ≤ 2|Θ(u, )|Lp (Ω) .

Proof. Since 1 < p ≤ ∞, we know that

Z
|f |Lp = sup f g.
|g| 0 =1
Lp

71
Thus1 , it suffices to show that

Z
| u∂ij φ| ≤ 2|Θ(u, )|Lp |φ|Lp0 ,
Ω

for all φ ∈ Cc∞ (Ω). Let us start by observing (as we noted above) that if v ∈ Rn is a fixed unit

vector.

∂vv φ(x0 ) = lim ∆2δv φ(x0 ).

δ→0

Since φ is Cc∞ , the limit is actually uniform. Thus,

Z Z Z
u∂vv φ = lim u∆2δv φ = lim φ∆2δv u ≤ |Θ(u, )|Lp |φ|Lp0 ,
δ→0 δ→0

by Corollary 8.9. Note that in the second equality we used change of variables and the compact

support of φ. Next, let us see that mixed derivatives ∂ij can be bounded by a sum of derivatives

in a single direction. This is just the following:

1
∂ij = (2∂vv − ∂ii − ∂jj ),
2

where
ei + ej
v= √ .
2

The result follows.

Remark 8.11. The reason for the 2 on the right hand side of the inequality in the proposition

is just how we wrote ∂ij . My impression is that this is not really necessary.

A consequence of the p = ∞ case is that we have Lipschitz bounds on Du once we have L∞

bounds on Θ.

Proposition 8.12. Let u ∈ C(Ω) and assume that B ⊂⊂ Ω is convex. For  sufficiently small,
1
Note that for this estimate, we are just going to assume that u ∈ C 2 (Ω̄)). The result is in fact true, but we
should know a little more about Sobolev spaces to jump to the next point. Another way to think of this is just
as the definition of the norm of |D2 u|Lp .

72
define

Θ(u, )(x) := Θ(u, Ω ∩ B (x))(x),

for x ∈ B̄. Assume that Θ(x, ) ≤ K for all x ∈ B̄. Then, u ∈ C 1,1 (¯). Moreover,

|Du(x) − Du(y)| ≤ 2n|Θ(u, )|L∞ (B) |x − y|,

for all x, y ∈ B̄.

Proof. By the previous result, for p = ∞, we have that

|D2 u|L∞ ≤ 2|Θ(u, )|L∞ .

Now we just take any two points x, y and use the mean value theorem for each component of

Du(x) − Du(y).

9 Viscosity Solutions

Our goal now is to give the definition of viscosity subsolutions and supersolutions for fully non-

linear elliptic PDE, which can be viewed as “weak” solutions. Normally, when we define weak

solutions, we relax the notion of solution by using test functions. This is a property that strong

solutions have that allows us to broaden the notion of solutions to all objects enjoying that

property. This, along with the statement that weak + smooth implies strong, is what allows us

to say that the notion of weak solutions is a reasonable one (at least at first glance). Viscosity

solutions are a different type of relaxation where the property that we extend is based on the

comparison principle, rather than integration by parts. This has many uses. One use is that this

can be used to define solutions to fully nonlinear problems and non-divergence form equations.

A second one is that viscosity solutions do not have to be H 1 from the beginning but can merely

be continuous.

73
9.1 Elliptic Equations and the Definition of Viscosity Solutions

We are considering equations of the form

F (D2 u(x), x) = f (x), (9.1)

where x ∈ Ω, u, f : Ω → R. F : S × Ω → R, where S is the set of all real n × n symmetric

matrices.

Definition 9.1. F is said to be uniformly elliptic if there are two positive constants λ ≤ Λ such

that for all M ∈ S and x ∈ Ω,

λ|N | ≤ F (M + N, x) − F (M, x) ≤ Λ|N |,

for all positive semi-definite N ∈ S.

Remark 9.2. Note that in the above, by the properties of symmetric matrices, |N | is just the

size of the largest eigenvalue of N .

Example 9.3. If F (M, x) = tr(M ), this corresponds to the Laplacian. The Laplacian is elliptic

with ellipticity constants (1, n) (using this definition) since

|N | ≤ tr(N ) ≤ n|N |,

for all N ≥ 0.

Note that for any N ∈ S, we can write:

N = N + − N −,

where N + , N − ≥ 0 and N + N − = 0. This is clear by diagonalizing N . Note that the decom-

position is not unique since there could be many eigenvectors with zero eigenvalue. From the

definition, we have the following Lemma (but these do not affect the sizes of |N − | and |N + |.

74
Lemma 9.4. F is uniformly elliptic if and only if

F (M + N, x) ≤ F (M, x) + Λ|N + | − λ|N − |,

for all M, N ∈ S, x ∈ Ω.

Remark 9.5. It follows that any uniformly elliptic F is monotone increasing and Lipschitz

continuous in M .

Proof. First assume that F is uniformly elliptic.

F (M + N ) = F (M − N − + N + ).

Hence,

λ|N + | ≤ F (M + N ) − F (M − N − ) ≤ Λ|N + |

Moreover,

λ|N − | ≤ F (M ) − F (M − N − ) ≤ Λ|N − |

Thus,

−Λ|N − | ≤ F (M − N − ) − F (M ) ≤ −λ|N − |.

Adding the second and fourth inequalities thus yields

λ|N + | − Λ|N − | ≤ F (M + N ) − F (M ) ≤ Λ|N + | − λ|N − |.

Keeping the right inequality yields he result. Now let us assume the conclusion is true. if

N = N + , we have the upper bound in the definition of uniform ellipticity for free. Now if we

replace M by M − N and take N + = 0, we have the lower bound.

We now give the definition of viscosity solution.

Definition 9.6. u ∈ C(Ω) is called a viscosity subsolution [resp. viscosity supersolution] of

75
(9.1):

F (D2 u(x), x) = f

in Ω, when the following condition holds: if x0 ∈ Ω and φ ∈ C 2 (Ω) is such that u − φ has a local

maximum at x0 , then

F (D2 φ(x0 ), x0 ) ≥ f (x0 ).

In other words, if φ touches u at x0 from above, then the inequality holds. If u is a supersolution

and if φ touches u at x0 from below, then the opposite inequality holds. A viscosity solution of

the equation is a continuous function that is both a subsolution and a supersolution.

9.1.1 Basic Properties of Viscosity Solutions

We now have a simple characterization of viscosity subsolutions that will be useful.

Proposition 9.7. The following are equivalent.

1. u is a viscosity subsolution of (9.1) in Ω.

2. If x0 ∈ Ω and if A is an open neighborhood of x0 and φ ∈ C 2 (A),

u≤φ in A and u(x0 ) = φ(x0 ),

then F (D2 φ(x0 ), x0 ) ≥ f (x0 ).

3. Same as the previous point with φ ∈ C 2 replaced by φ is a paraboloid (same condition and

same conclusion).

Proof. (1) =⇒ (2) =⇒ (3), by the previous definition. To show that (3) =⇒ (1), assume

that φ ∈ C 2 (Ω) is such that u − φ has a local maximum at x0 . Then, for  > 0, define

1 
P (x) = u(x0 ) + Dφ(x0 ) · (x − x0 ) + (D2 φ(x0 )(x − x0 ), (x − x0 )) + |x − x0 |2 .
2 2

Then, P is a paraboloid that touches u from above at x0 on a small neighborhood A of x0 .

76
Thus, since we have (3), it follows that

F (D2 φ(x0 ) + I, x0 ) ≥ f (x0 ),

for all  > 0. Recall F is Lipschitz continuous in M by the definition of uniform ellipticity. In

fact, by the definition of uniform ellipticity

λ ≤ F (D2 φ0 (x0 ) + I, x0 ) − F (D2 φ0 (x0 ), x0 ) ≤ Λ.

Sending  → 0 shows that

F (D2 φ(x0 ) + I, x0 ) ≥ f (x0 ).

Thus, u is a subsolution of (9.1).

Lemma 9.8. If u is a viscosity super solution of F (D2 u, x) = f (x) in Ω, then v = −u is a

viscosity subsolution of G(D2 v, x) = −f (x), where

G(M, x) = −F (−M, x).

Note that G and F have the same ellipticity constants.

Proof. Exercise.

Next, we have that viscosity subsolutions that are also C 2 are true subsolutions.

Lemma 9.9 (Weak =⇒ Strong). Assume that u ∈ C 2 (Ω). Then, u is a viscosity subsolution of

(9.1) if and only if F (D2 u(x), x) ≥ f (x) for every x ∈ Ω.

Proof. First, in the forward case, we can take φ = u and we get the conclusion. In the reverse

case, if φ ∈ C 2 and u − φ has a local maximum at x0 , we have from calculus that

D2 (u − φ)(x0 ) ≤ 0.

77
Thus,

D2 u(x0 ) ≤ D2 φ(x0 ).

Since F is increasing in its first variable, if F (D2 u(x0 ), x0 ) ≥ f (x0 ) (u is a subsolution in the

classical sense), we have that

f (x0 ) ≤ F (D2 u(x0 ), x0 ) ≤ F (D2 φ(x0 ), x0 ).

Thus, u is a viscosity subsolution.

Next, we have that the maximum of subsolutions is a subsolution.

Proposition 9.10. Assume that u and v are viscosity subsolutions of (9.1) in Ω. Then, max u, v

is also a viscosity subsolution of (9.1) in Ω.

Proof. Any φ touching max u, v from above at a point x0 ∈ Ω touches either u or v from above

at x0 . Since both u and v are viscosity subsolutions, we must have F (D2 φ(x0 ), x0 ) ≥ f (x0 ).

Finally, we show that the family of viscosity subsolutions is closed.

Proposition 9.11. Let {Fk }k≥1 be a sequence of continuous uniformly elliptic operators with

ellipticity constants λ, Λ. Let uk ∈ C(Ω) be a sequence of viscosity subsolutions of Fk (D2 uk , x) ≥

f in Ω. Assume that Fk converges uniformly on compact sets of S × Ω to a continuous function

F and that uk converges uniformly on compact subsets of Ω to u. Assume f is continuous.

Then,

F (D2 u, x) ≥ f (x),

in the viscosity sense on Ω.

Proof. Let P be a paraboloid that touches u from above at x0 . Take a small r > 0 so that

Br (x0 ) ⊂ Ω. We know that uk → u uniformly on Br (x0 ). We want to estimate from below

F (D2 P (x0 ), x0 ) = (F − Fk )(D2 P (x0 ), x0 ) + Fk (D2 P (x0 ), x0 ).

78
Note that P is fixed, so the first term is arbitrarily small for k large. Now we just want to

estimate from below Fk (D2 P (x0 ), x0 ). Now, we have that P ≥ u on Br (x0 ). Thus, for  > 0,

consider the function

1
Qk, (x) = P (x) + |x − x0 |2 + uk (x0 ) − u(x0 ).
2

Observe that

D2 Qk, = D2 P + I.

Since uk → u uniformly, if we take k sufficiently large we have that

Qk, > uk , for x ∈ ∂Br (x0 ) and Qk, (x0 ) = uk (x0 ).

It follows from continuity that there exists xk ∈ Br (x0 ) so that Qk, touches uk from above at

xk . Thus,

Fk (D2 P + I, xk ) = D2 Fk (D2 Qk, (xk ), xk ) ≥ f (xk ).

The result now follows from continuity.

9.2 The Class of solutions of all uniformly Elliptic Equations and the Pucci
Operators

We begin by introducing Pucci’s extremal operators. Let 0 < λ ≤ Λ. For M ∈ S, we define

X X
M− (M, λ, Λ) = M− (M ) = λ ei + Λ ei ,
ei >0 ei <0

X X
M+ (M, λ, Λ) = M+ (M ) = Λ ei + λ ei ,
ei >0 ei <0

where ei are the eigenvalues of M .

Remark 9.12. The Pucci extremal operators are uniformly elliptic.

79
 Let A be a symmetric matrix such that

λ|ξ|2 ≤ Aij ξi ξj ≤ Λ|ξ|2 .

In this case, we write:

A ∈ Aλ,Λ .

It is easy to see that

X X
M− (M, λ, Λ) = λ ei (M ) + Λ ei (M ) = min tr(AM ).
A∈AλΛ
ei (M )>0 ei (M )<0

The minimum is achieved by an A which is diagonalizable in the same basis as M and has

eigenvalue λ when M has a positive eigenvalue and Λ when it has a negative one. Similarly,

X X
M+ (M, λ, Λ) = λ ei (M ) + Λ ei (M ) = max tr(AM ).
A∈AλΛ
ei (M )<0 ei (M )>0

Both equalities are seen by noting that the trace of a symmetric matrix is the sum of its

eigenvalues (with multiplicity).

We now enumerate a number of important properties of M− and M+ .

Lemma 9.13. M− and M+ satisfy the following properties

1. M− (M ) ≤ M+ (M ).

2. λ0 ≤ λ ≤ Λ ≤ Λ0 =⇒ M− (M, λ0 , Λ0 ) ≤ M− (M, λ, Λ) and M+ (M, λ0 , Λ0 ) ≥ M+ (M, λ, Λ).

3. M− (M, λ, Λ) = −M+ (−M, λ, Λ).

4. M± (αM ) = αM± (M ) for any α ≥ 0.

5. M+ (M ) + M− (N ) ≤ M+ (M + N ) ≤ M+ (M ) + M+ (N ).

6. M− (M ) + M− (N ) ≤ M− (M + N ) ≤ M− (M ) + M+ (N ).

80
 7. N ≥ 0 =⇒ λ|N | ≤ M− (N ) ≤ M+ (N ) ≤ nΛ|N |.

8. M− and M+ are uniformly elliptic with ellipticity constants λ, nΛ.

The proof of this Lemma is elementary and we leave it to the student.

Next, we define the class S of solutions of all elliptic equations.

Definition 9.14. Let f be a continuous function in Ω and 0 < λ ≤ Λ. We denote by S(λ, Λ, f )

the space of continuous functions u in Ω which are subsolutions M+ (D2 u, λ, Λ) ≥ f (x) in

the viscosity sense in Ω. Similarly S̄(λ, Λ, f ) are the set of viscosity subsolutions u satisfying

M− (D2 u, λ, Λ) ≤ f (x). Define

S(λ, Λ, f ) = S(λ, Λ, f ) ∩ S̄(λ, Λ, f ),

S ∗ (λ, Λ, f ) = S(λ, Λ, −|f |) ∩ S̄(λ, Λ, |f |).

We have the following simple Lemma.

Lemma 9.15. Assume that u ∈ S(λ, Λ, f ) ∩ C 2 (Ω). Then, for each x ∈ Ω, there is a uniformly

elliptic matrix λ ≤ aij (x) ≤ Λ so that

aij (x)Dij u ≥ f (x).

Proof. Since u ∈ C 2 we know that M+ (D2 u(x), λ, Λ) ≥ f (x). Thus,

max tr(AD2 u(x)) ≥ f (x).

A∈Aλ,Λ

We just take A = A(x) to be the maximizer.

Remark 9.16. A consequence of this fact is that once we prove some fact about all elements

of the class S (which one could view as essentially a “linear space”), they will automatically be

inherited by solutions to uniformly elliptic equations, without having to linearize F or anything

like that.

81
 Now we give a few basic properties of S, S̄, S, S ∗ (when their arguments are clear from

context, we will omit writing them).

Lemma 9.17. 1. λ0 ≤ λ ≤ Λ ≤ Λ0 =⇒ S(λ, Λ, f ) ⊂ S(λ0 , Λ0 , f ) and the same holds for

S̄, S, S ∗ . This follows from the second property of the Pucci operators above.

2. u ∈ S(λ, Λ, f ) =⇒ −u ∈ S̄(λ, Λ, −f ). This follows from the third property of the Pucci

operators.

3. If α, r > 0, u ∈ S(λ, Λ, f ), v(·) = αu( r· ), on rΩ, then

α ·
v ∈ S(λ, Λ, f ( )).
r2 r

This is a fundamental scaling law and it follows from the definition of the Pucci operators.

4. If u ∈ S(λ, Λ, f ), φ ∈ C 2 (Ω) and M+ (D2 φ(x)) ≤ g(x) on Ω, we have that

u − φ ∈ S(λ, Λ, f − g).

This is essentially a linearity property and it is what we alluded to in the previous remark.

Proof. We have already mentioned the proofs of (1)-(3) in the statement. To prove (4), let

ψ ∈ C 2 touch u − φ from above at x0 . Then, ψ + φ touches u from above at x0 . It follows that

M+ (D2 φ(x0 ) + D2 ψ(x0 )) ≥ f (x0 ).

However, we know that M+ is subadditive from the fifth property of the Pucci operators.

Therefore,

M+ (D2 ψ(x0 )) ≥ f (x0 ) − M+ (D2 φ(x0 )) ≥ f (x0 ) − g(x0 ),

by the assumption on φ.

Remark 9.18. Since M− is super-additive, a similar result can be established for supersolu-

tions.

82
 Combining some of our previous results, we get the following.

Lemma 9.19. • If u, v ∈ S(f ), then sup(u, v) ∈ S(f ).

• u ∈ S(0) =⇒ u+ ∈ S(0).

• S(f ), S̄(f ), S(f ) are all closed under uniform limits on compact sets.

Finally, let us mention the following proposition.

Proposition 9.20. Let u satisfy F (D2 u, x) ≥ f (x) in the viscosity sense. Then,

λ
u ∈ S( , Λ, f (x) − F (0, x)).
n

More generally, if φ ∈ C 2 (Ω), then

λ
u − φ ∈ S( , Λ, f (x) − F (D2 φ(x), x)).
n

For the supersolutions we just replace S by S̄.

Proof. We prove the second statement since the first is just the case φ = 0. Suppose that ψ

touches u − φ from above at x0 . Then, ψ + φ touches u from above at x0 . Thus, by definition

f (x0 ) ≤ F (D2 ψ(x0 ) + D2 φ(x0 ), x0 )

≤ F (D2 φ(x0 ), x0 ) + Λ|D2 ψ(x0 )+ | − λ|D2 ψ(x0 )− |

X λ X
≤ F (D2 φ(x0 ), x0 ) + Λ ei (ψ) + ei (ψ)
n
ei (ψ)>0 ei (ψ)<0

λ
= F (D2 φ(x0 ), x0 ) + M+ (D2 φ(x0 ), , Λ).
n

Therefore,
λ
M+ (D2 φ(x0 ), , Λ) ≥ f (x0 ) − F (D2 φ(x0 ), x0 )
n

83
10 Alexandroff Maximum Principle and Applications

We start by establishing the Alexandroff maximum principle for classical solutions and then

prove the viscosity case. For this chapter we will mainly follow the exposition of the book of

Lin and Han.

10.1 The case of classical solutions

We define the operator:

L = aij (x)Dij ,

with aij ∈ C(Ω). As usual, we assume that the matrix A is positive definite. Define

D∗ := det(A)1/n .

Note that D∗ is just the geometric mean of the eigenvalues of A. Note that since the arithmetic

mean is greater than the geometric mean, we have

1
tr(A) ≥ D∗ .
n

Let λ, Λ denote the largest and smallest eigenvalues of A respectively. If A is strictly positive

definite we have

0 < λ ≤ D∗ ≤ Λ.

10.1.1 Contact Set

We now define the concept of contact set. For u ∈ C 1 (Ω), we define

Γ+
u = {y ∈ Ω : u(x) ≤ u(y) + Du(y) · (x − y), for any x ∈ Ω}.

84
The set Γ+ is called the upper contact set of u and

D2 u ≤ 0 on Γ+ .

Note that the upper contact set can be defined for continuous functions as:

Γ+ n
u = {y ∈ Ω : u(x) ≤ u(y) + p(y) · (x − y), for any x ∈ Ω and some y ∈ R }.

u is concave if and only if Γ+ = Ω. If u ∈ C 1 , then p(y) = Du(y). We can now state the

Alexandroff Maximum Principle.

Theorem 20. Suppose u ∈ C(Ω̄) ∩ C 2 (Ω) satisfies aij ∂ij u ≥ f in Ω with

f
∈ Ln (Ω).
D∗

Then,
f−
sup u ≤ sup u+ + C| | n +,
Ω ∂Ω D∗ L (Γ )

where Γ+ is the upper contact set of u and C = C(n, diam(Ω)).

Proof. Without loss of generality assume that u ≤ 0 on ∂Ω. Let us denote by Ω+ = {x :

u(x) > 0}. As we noted before, D2 u ≤ 0 on Γ+ . Thus, if we consider the function u = u − |x|2 ,

D2 u < −2Id is negative semi-definite on Γ+ . This implies that ∇u : Γ+ ∩Ω+ → ∇u (Γ+ ∩Ω+ )

is a diffeomorphism and by the change of variables formula, we have:

Z Z
1= |det(D2 u − 2Id)|.
∇u (Γ+ ∩Ω+ ) Γ+ ∩Ω+

Sending  → 0, we see that

Z Z
+ + 2
 aij Dij u n f n
|∇u(Γ ∩ Ω )| ≤ −det(D u) ≤ − ≤| | n +,
Γ+ ∩Ω+ nD∗ nD∗ L (Γ )

1
since n multiplied by the trace always controls the n-th root of the determinant.

85
 Now we wish to show that

BM (0) ⊂ ∇u(Γ+ ∩ Ω+ ) ⊂ Rn ,

where
supΩ u − inf Ω u
M= .
diam(Ω)

We will be done once this is established.

Assume without loss of generality that u attains its maximum m > 0 at 0 ∈ Ω. Consider

the affine function L(x) defined by:

L(x) = m + a · x,

m
where |a| < diam(Ω) . By definition, L(x) > 0 on Ω and L(0) = m. Since u is maximal at 0 we

have that ∇u(0) = 0. Thus, there exists x1 close to 0 with u(x1 ) > L(x1 ) > 0, while u ≤ 0 < L

on ∂Ω. Therefore, there exists x̃ so that ∇u(x̃) = ∇L(x̃). Now we translate the plane up until

we have the highest such position. This point of tangency will be in the contact set. This means

that a ∈ Du(Γ+ ∩ Ω+ ). Now we are done.

Remark 10.1. To see the last step, it is best to imagine we have a hyperplane with small

“slope” coming down from infinity. The first point it touches has to be inside Γ+ ∩ Ω+ and the

gradient at that point is simply the “slope” a.

10.2 Maximum Principle on Small Domains

Remark 10.2. The above proof works almost identically when we had a lower order term c(x)u

to the elliptic equation when c ≤ 0.

Theorem 21. Suppose u ∈ C(Ω̄) ∩ C 2 (Ω) satisfies aij ∂ij u + cu ≥ 0 in Ω with u ≤ 0 on ∂Ω.

Assume diam(Ω) ≤ d. Then, there is a constant δ = δ(n, λ, Λ, d, |c+ |L∞ ) > 0 so that if |Ω| < δ

then u ≤ 0 in Ω.

86
Proof. We have

aij ∂ij u − c− u ≥ −c+ u.

By the AMP, we have

1
sup u+ ≤ C(λ, Λ)|c+ u+ |Ln (Ω) ≤ |c+ |L∞ c(n, λ, d)|Ω|1/n sup u+ ≤ sup u+ .
Ω Ω 2 Ω

10.3 ABP Estimate for Viscosity Solutions

We now move to the ABP estimate for viscosity solutions, following the book of Caffarelli and

Cabré. Define:

Γu (x) = sup{L(x) : L ≤ u, on Ω, L is affine}.

L

Our goal will now be to establish the following theorem.

Theorem 22. Suppose u ∈ S̄(λ, Λ, f ) in B1 with u ≥ 0 on ∂B1 for some f ∈ C(B̄1 ). Then we

have
Z 1/n
sup u− ≤ c(n, λ, Λ) (f + )n ,
B1 B1 ∩{u=Γu }

where Γu is the convex envelope of −u− = min{u, 0}, defined above.

Proof. We will show that if x0 ∈ {Γu = u} then we have

f (x0 ) ≥ 0

and
1
L(x) ≤ Γu (x) ≤ L(x) + C(f (x0 ) + (x))|x − x0 |2 ,
2

where L is an affine function and x is close to x0 and (x) → 0 as x → x0 . The second inequality

will be established by observing that if a parabola touches Γu from below at a contact point,

then it touches u from below. The key point is that since f is bounded, the opening of the

87
paraboloids touching from below has to be bounded. For example, Γu could not have a corner

singularity at a contact point if u ∈ S + . Note that the convex envelope need only be Lipschitz

in general.

With some work (see below), the above will show that Γu is C 1,1 on the contact set. Off

the contact set, Γu should be affine and uniformly smooth up to the contact set. We will then

be able to conclude that Γu ∈ C 1,1 . Then, we will apply the classical Alexandroff Maximum

Principle we proved above to Γu and it will imply the estimate in the statement of the theorem.

What we mentioned will be established rigorously via series of Lemmas that we now give.

Lemma 10.3. Let u ∈ S̄(λ, Λ, f ) in Bδ = Bδ (0). Assume that f ∈ L∞ (Bδ ) and that φ is convex

on Bδ and that

0 = φ(0) = u(0), 0≤φ≤u

in Bδ . Then,

φ(x) ≤ C|f + |L∞ (Bδ ) |x|2 ,

for all x ∈ Bνδ , where ν < 1, C only depend on λ, Λ, n.

Remark 10.4. The purpose of this is to establish the two inequalities we mentioned above.

Note that here we are proving the second inequality we mentioned in the sketch above in the

case where L = 0 (which one could assume WLOG by subtracting from u an affine function).

Proof. Fix 0 < r ≤ 4δ . Define

1
Cr = sup φ.
r2 Br

Because φ is convex, it attains its maximum at some x0 ∈ ∂Br :

φ(x0 ) = Cr r2 .

Since φ is convex, {x ∈ Bδ : φ(x) ≤ Cr r2 } is convex and contains Br by definition of Cr .

Without loss of generality, assume that x0 = (0, 0, 0, ..., r). Therefore, by a simple argument

88
done first for smooth convex functions, we see that

φ(x0 , r) ≥ φ(x0 ) = Cr r2 ∀x0 ∈ Rn−1 .

Our goal is to now to bound C̄ in terms of f . As alluded to in the sketch above, we will show

that φ must lie underneath a certain paraboloid.

Consider the set

Rr = {(x0 , xn ) : |x0 | ≤ N r, |xn | ≤ r}.

N will be chosen later to be sufficiently large depending on λ, Λ, n. Consider the polynomial:

h(x) = (xn + r)2 − b|x0 |2 .

Observe that h satisfies:

• for xn = −r, we have that h ≤ 0;

• for |x0 | = N r, we have that h ≤ (4b − N 2 )r2 ≤ 0 when we choose b = 4

N2
;

• for xn = r, we have that h ≤ 4r2 . Now we define:

Cr
h̃ = h.
4

Observe that h̃ is a paraboloid. Since φ ≥ 0, we have that

h̃ ≤ φ ≤ u on ∂Rr .

Note that we showed that φ ≥ r2 Cr on the whole of the upper boundary xn = r. Moreover,

observe that
Cr
h̃(0) = > φ(0) = u(0) = 0.
4r2

Thus, u − h̃ attains a local minimum somewhere inside of Rr . Since u ∈ S̄(λ, Λ, f ), we

89
 then see that since the eigenvalues of D2 h̃ are just given by Cr /2, − 2C
N2
r
, ..., − 2C
N2
r
, we have

that
Cr Cr
λ − 2Λ(n − 1) 2 ≤ |f |L∞ (Rr ) .
2 N

Therefore, if we choose
8Λ(n − 1)
N2 = ,
λ

we see that
2
Cr ≤ |f | ∞ .
λ L (Rr )
1
This concludes the proof by taking ν = N.

Next, we make a general observation.

Lemma 10.5. Assume Γ ∈ C(B1 ) and that there exists a closed set A on which Γ can be

approximated in C 1,1 by paraboloids. Assume also that for the open set B = B1 \ A we have that

to each x ∈ B there is a unique connected maximal open set Bx on which Γ is affine with the

property that ∂Bx ⊂ A. Then, Γ ∈ C 1,1 (B1 ).

Proof. Let x, y ∈ B1 . We want to compute the second order difference quotient with x and

y. Let us denote it by Dx,y (Γ). If y ∈ Bx , then Dx,y (Γ) = 0. If not, then there is a x∗ ∈ A

with |x − x∗ |, |y − x∗ | ≤ |x − y|. Since Γ can be approximated by two paraboloids at x∗ , we are

done.

Finally, we observe that because of Lemma (10.3), Γu satisfies the assumptions of the pre-

ceding Lemma with the closed set A = {x : Γu = u}. Note that if x ∈ Ac , we must have that Γ

is affine in a neighborhood of x. This completes the proof of the Alexandroff-Baklemann-Pucci

estimate.

90
11 Harnack Inequality and Oscillation Lemma

We now move to establish the Harnack inequality for members of S ∗ . As we know, one of the

important consequences of the Harnack inequality is the oscillation lemma, which we use to

establish Hölder regularity. We will start by introducing two tools. The first is the construction

of a “barrier function” and the second is a decomposition of cubes similar to the Calderón-

Zygmund decomposition. Toward this end, we give some notation. Ql (x0 ) is the cube of side

length l centered at x0 . When we write Ql it means Ql (0).

11.1 Barrier and Calderón-Zygmund Decomposition

We start by constructing a “barrier function.”

Lemma 11.1. Given constants 0 < λ ≤ Λ, there exists φ ∈ C ∞ (Rn ) and universal constants

C, M > 0 such that

φ ≥ 0 in Rn \ B2√n (11.1)

φ ≤ −2 in Q3 (11.2)

M+ (D2 φ, λ, Λ) ≤ Cξ, (11.3)

where ξ ∈ Cc∞ (Q1 ) and 0 ≤ ξ ≤ 1. Moreover, φ ≥ −M on Rn .

This is a function which looks like a bowl with a dip in the center and a rather slow conver-

gence to a constant as |x| → ∞.

c :
Proof. We search for a radial φ which in fact takes the following form in B1/4

φ(x) = M1 − M2 |x|−α ,

where the constants M1 , M2 ≥ 0 are chosen

φ|B 3 √n = −2, φ|B2√n = 0.

2

91
Outside of B2√n , φ > 0 and is uniformly smooth. We can take φ to be radial and connecting
√ √
0 and −2 for the radial variable r ∈ [ 23 n, 2 n]. It does not really matter how we connect

φ. The key thing we need to check now is that φ is superharmonic outside B1/4 in the sense
1
that M+ (D2 φ, λ, Λ) ≤ 0 in |x| ≥ 4. Let us now compute the eigenvalues of D2 φ in B1/4
c .

Note that they must be a function of |x|. Thus, it suffices to compute the eigenvalues at

(r, ..., 0, 0). At such a point, it is not difficult to check that D2 φ is diagonal and the eigenvalues

are M2 r−α−2 (−α(1 + α), α, α, ..., α). Thus,

M+ (D2 φ, λ, Λ) ≤ M2 |x|−α−2 (−λα(1 + α) + (n − 1)Λα) ≤ 0

if α is sufficiently large. In fact, we can take α = max{1, (n − 1) Λλ − 1}.

We now recall the classical Calderón-Zygmund decomposition.

Lemma 11.2 (Calderón-Zygmund Decomposition). Suppose f ∈ L1 (Q1 ) is non-negative and

R
α > −Q1 f is a fixed constant. Then, there exists a sequence of non-overlapping dyadic cubes

{Qj } in Q1 such that

f (x) ≤ α for almost every x ∈ Q0 \ ∪j Qj

and
Z
α ≤ − f < 2n α.
Qj

Remark 11.3. The proof is elementary. The use of this result is that we can decompose Q1

into two sets, a good set on which f is uniformly bounded and a “bad” set on which f may not

be uniformly bounded but for which we control the average of f .

Proof. We just cut Q1 into 2n dyadic cubes. We keep a cube Q at any generation if −Q f ≥ α.
R

It must of course be that −Q f ≤ 2n α. We keep dissecting the ones we don’t keep. The result
R

follows by the Lebesgue density theorem.

Next, we give a slight modification. If Q is a daydic cube, we say that Q̃ is its predecessor

if Q is one of the 2n dyadic cubes obtained from cutting Q̃.

92
Lemma 11.4. Let A ⊂ B ⊂ Q1 be measurable sets and 0 < δ < 1 such that

• |A| ≤ δ, and

• If Q is a dyadic cube such that |A ∩ Q| > δ|Q|, then Q̃ ⊂ B.

Then, |A| ≤ δ|B|.

Proof. The proof is an application of the Calderón-Zygmund decomposition Lemma to χA .

Indeed, in that case, we find that

A ⊂ ∪∞
j=1 Qj ,

with Qj essentially disjoint. Moreover, we find that

|A ∩ Qj | |A ∩ Q̃j |
δ< ≤ 2n δ < δ,
|Qj | |Q̃j |

for all j. Relabel the Q̃j so that they are pairwise disjoint. Using the second assumption in the

theorem, we see that

Q̃j ⊂ B.

Thus,
X X
|A| = |A ∩ Q̃j | ≤ δ |Q̃j | ≤ δ|B|.

11.2 The Harnack Inequality

We now proceed to establish the Harnack inequality, which is the content of the following

Theorem.

Theorem 23. Let u ∈ S ∗ (λ, Λ, f ) in Q1 satisfy u ≥ 0 in Q1 , where f ∈ BC(Q1 ). Then,

sup u ≤ C( inf u + |f |Ln (Q1 ) ).

Q1/2 Q1/2

93
 In fact, it is not difficult to check that, by a scaling and covering argument it suffices to

establish the following Lemma.

Lemma 11.5. There exist universal constants 0 , C > 0 so that if u ∈ S ∗ (λ, Λ, f ) in Q4√n and

u ∈ C(Q̄4√n ) satisfy u ≥ 0 in Q4√n where f ∈ BC(Q4√n and inf Q1/4 u ≤ 1, then

sup u ≤ C.
Q1/4

Let us start by defining the distribution function.

Definition 11.6. Assume that g : Ω → R is measurable. Then, λg (t) = |{x ∈ Ω : u(x) > t}|, is

called the distribution function of g.

Remark 11.7. Note that λg is a decreasing function of t and that rate of decay of λg as t grows

essentially determines what Lp λg belongs to. sup g ≤ C is equivalent to λg (t) ≡ 0 for t > C.

Our first task will be to show that λu (t) ≤ Ct− where C,  are universal constants. For

this, we will use the Barrier argument, the cube decomposition lemma, and the ABP estimate.

A key feature of the ABP estimate we will use is that we integrate over the contact set on the

right hand side of the ABP estimate. The upper bound on the distribution function λu already

implies that

|u|Lq (Q1 ) ≤ C inf u,

Q3

for some small q and C that are universal. After showing this, we will show the same decay of

λu at “all scales,” by which we just mean that we can rescale u and the domain and get the

same type of bound on λu . Doing so will allow us to bound sup u.

11.2.1 Bound on the distribution function

We will first show a base Lemma and then iterate it along with the cube decomposition Lemma

to get bounds on the distribution function.

94
Lemma 11.8. There exist universal constants 0 > 0, 0 < µ < 1 and M > 1 so that if u ∈ S̄(|f |)

in Q4√n and u ∈ C(Q4¯√n ) and f satisfy

u ≥ 0 in Q4√n

inf u ≤ 1
Q3

|f |Ln (Q4√n ) ≤ 0 ,

then we have that

|{u ≤ M } ∩ Q1 | > µ,

where M = −2.

Remark 11.9. We will apply the ABP estimate to w = u + φ where φ is the barrier function

of Lemma (11.1).

Proof. Define w = u + φ. Note that B2√n ⊂ Q4√n and that M+ (D2 φ) ≤ Cξ (pointless). Thus,

we have that w ∈ S̄(|f | + Cξ) in B2√n . Now, w ≥ 0 on ∂B2√n since φ = 0 there and u ≥ 0. We

also have that inf Q3 (w) ≤ −1 since φ ≤ −2 there. Thus, applying the ABP estimate to w, we

see that:

1 ≤ C|f + Cξ|Ln ({w=Γw }∩B2√n ) ≤ C0 + C|{w = Γw } ∩ Q1 |1/n .

Taking 0 sufficiently small then tells us that

1
≤ |{w = Γw } ∩ Q1 |.
(2C)n

Observe, however, that since w = u + φ, when w(x) = Γw (x) we have w(x) ≤ 0 (recall that Γw

is the convex envelope of −w− ). Thus, w = Γw implies that u ≤ M . Thus,

1
≤ |{u ≤ M } ∩ Q1 |.
(2C)n

95
 Now we establish the bounds on the distribution function using the cube decomposition

lemma.

Lemma 11.10. Let u be as in the previous Lemma. Then,

|{u > M k } ∩ Q1 | ≤ (1 − µ)k .

Proof. Observe that we have the bound for k = 1. Suppose that the bound holds for k − 1 and

define

A = {u > M k } ∩ Q1 , B = {u > M k−1 } ∩ Q1 .

The result will be established once we show that

|A| ≤ (1 − µ)|B|.

The key is to use the cube lemma. The first assumptions of the lemma are satisfied. Indeed,

we have A ⊂ B ⊂ Q1 and |A| ≤ (1 − µ). We must now show that if Q is a dyadic cube,

Q = Q1/2i (x0 ) then we have that

|A ∩ Q| > (1 − µ)|Q| =⇒ Q̃ ⊂ B.

Let us suppose that Q̃ 6⊂ B. That is, take x̃ ∈ Q̃ such that u(x̃) ≤ M k−1 . Then, let us transform:

1
x = x0 + y,
2i

so that y ∈ Q1 for all x ∈ Q = Q1/2i (x0 ). Consider the function:

1
ũ(y) = u(x).
M k−1

96
We claim that ũ satisfies the assumptions of the previous lemma. If so, then:

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

Thus, |Q \ A| > µ|Q|, which contradicts our assumption about Q. To complete the proof, we

just need to show that ũ satisfies the assumptions of the previous Lemma. Observe that by a

simple scaling argument,

f (x1 + 21i y)
ũ(y) ∈ S̄( ) = S̄(f˜(y))
22i M k−1
1
in Q4√n . Moreover, x ∈ Q̃ =⇒ y ∈ Q3 . Thus, ũ ≥ 0 and inf Q3 ≤ M k−1
u(x̃) ≤ 1 by assumption.

Moreover, it is easy to see that |f˜|Ln (Q4√n ≤ |f |Ln (Q4√n ) ≤ 0 . Now we are done.

11.2.2 Bounds on all scales

Note that to establish the above bounds, we only used that u ∈ S̄. We will now observe that

since u ∈ S ∗ (f ), we also have that −u ∈ S̄(|f |). This means that the above considerations also

apply to −u. We also have the following Lemma.

Lemma 11.11. Let u ∈ S(−|f |) in Q4√n . Assume that |f |Ln ≤ 0 and that

|{u ≥ t} ∩ Q1 | ≤ dt− ,

with d a universal constant.

M0
Then, there exist universal constants M0 > 1, σ > 0 so that for ν = M0 − 12
> 1, the following

holds: if j ≥ 1 is an integer and x0 is such that

1
|x0 |∞ ≤
4

and

u(x0 ) ≥ ν j−1 M0 ,

97
then

Qj := Qlj (x0 ) ⊂ Q1 sup u ≥ ν j M0 ,

Qj

−/n −j/n
where lj = σM0 ν .

Remark 11.12. What we are saying now is that if there is an x0 with u(x0 ) ≥ M0 , then there is

x1 so that u(x1 ) ≥ νM0 , and an x2 so that u(x2 ) ≥ ν 2 M0 and so on. Furthermore, these points

get closer and closer to each other. We will later show that there is then a sequence of points

making u unbounded, which will contradict the continuity of u. It will follow that u < M0 ,

which will conclude the proof of the Harnack inequality.

Proof. Take σ > 0 and M0 > 1 so that

1 n √
σ > d2 (4 n)n ,
2

−/n 1
σM0 + dM0− ≤ .
2

It follows that Qlj (x0 ) ⊂ Q1 . Now suppose that supQj u < ν j M0 . We will derive a contradiction

using our previous Lemma. Observe that

 M −
0
|{u ≥ ν j M0 /2} ∩ Qlj /4√n (x0 )| ≤ |{u ≥ ν j M0 /2} ∩ Q1 | ≤ dν −j .
2

We will now use that supQl √

n (x0 )
u < ν j M0 to show that the complimentary set must also be
j /4

small. This will be the contradiction. Toward this end, consider the transformation

lj
x = x0 + √ y, y ∈ Q4√n , x ∈ Qj = Qlj (x0 )
4 n

and the function

u
νM0 − ν j−1 (x)
v(y) = .
(ν − 1)M0

Note that v > 0 on Q4√n by assumption. Let us claim that v satisfies the assumptions of the

98
first Lemma 11.8. We will show this is the case later. Thus we have that

|{v(y) > M0 } ∩ Q1 | ≤ dM0− .

Since u(x) < ν j M20 ⇐⇒ v(y) > M0 we have that

M0  l n
j
|{u(x) < ν j } ∩ Qlj /4√n (x0 )| ≤ √ dM0− .
2 4 n

Thus using the bound on the set of {u ≥ ν j M0 /2} ∩ Qlj /4√n (x0 ), we have that

 l n  l j n  M −
j 0
√ ≤ √ dM0− + dν −j .
4 n 4 n 2

But remember that

−/n −j/n 1 n √ −/n 1

lj = σM0 ν , σ > d2 (4 n)n , σM0 ≤
2 2

This is a contradiction. To complete the proof, we just have to show that v satisfies the hy-

potheses of Lemma 11.8. Indeed,

v ∈ S̄(f˜),

l l
where f˜(y) = ( 4√j n )2 (ν j−1 (ν − 1)M0 )−1 |f (x)|, where x = x0 + √j y.
4 n
Since u(x0 ) ≥ ν j−1 M0 ,

inf Q3 v ≤ 1. Moreover,

 l 2 1 lj  l  1
j j
|f˜|Ln (Q4√n ) = √ |f (x 0 + √ y)| n √
L (Q4 n ) = √ |f | n
j−1
4 n ν (ν − 1)M0 4 n 4 n ν (ν − 1)M0 L (Qlj (x0 ))
j−1

≤ 0 ,

ν
where we used that (ν − 1)M0 = 2 and ν > 1.

We can now finish the proof of the Harnack inequality by proving 11.5. Note that if M0 is

99
sufficiently large, we have that
∞
X 1
lj ≤ .
4
j=1

Note that taking M0 larger does only strengthens the results of the preceding Lemma. We will
1
now show that supQ1/4 u ≤ M0 . Suppose not. Then, there exists |x1 |∞ ≤ 8 so that u(x1 ) ≥ M0 .

This gives x2 with

l1
|x2 − x1 | ≤ l1 u(x2 ) ≥ νM0 .
2
1
Note that the factors of 2 are coming from the definition of Qlj . Repeating this process gives a

sequence of points xj with

1
|xj |∞ ≤ u(xj+1 ) ≥ ν j M0 .
4

This contradicts the continuity of u on Q1 .

Now the arguments we gave at the end of the proof of the De Giorgi Theorem allow us to

conclude the C α regularity of elements of S̄(f ).

12 Uniqueness and C 1,α Regularity

In the first part of this section, we will establish the following important Theorem.

Theorem 24. Let u be a viscosity subsolution of F (D2 w) = 0 in Ω and let v be a viscosity

supersolution. Then,
λ
u − v ∈ S( , Λ)
n

in Ω.

Remark 12.1. Note that we already established this if one of u or v is C 2 .

Corollary 12.2. The Dirichlet problem F (D2 u) = 0 in Ω, u = φ on ∂Ω has at most one

viscosity solution u ∈ C(Ω̄).

The result follows directly from the preceding theorem and the maximum principle for vis-

cosity solutions, which relies on the Alexandroff Maximum Principe. The proof of the theorem

100
will rely on approximate solutions introduced by Jensen.

12.1 The upper  envelope and its properties

Definition 12.3. Let u ∈ C(Ω) and assume that H ⊂⊂ Ω is an open set. For  > 0 we defined

the upper -envelope of u with respect to H:

1
u (x0 ) = sup u(x) +  − |x − x0 |2 ,
x∈H̄ 

for x0 ∈ H.

Observe that u (x0 ) ≥ u(x0 ) +  for all x0 ∈ H. We will show the following Theorem.

Example 12.4. It is instructive to consider the following two examples. First, when u1 (x) = |x|,

it is easy to compute u1 as follows:

5
u1 (x0 ) = |x0 | + ,
4

 
since it is easy to check that the maximum occurs at x0 + 2 when x0 ≥ 0 and x0 − 2 when

x0 < 0. Notice that u1 is an example of a corner that is opening up, so it is already “regular”

from below. Let us now consider the function u2 (x) = −|x|. The situation is now a bit more

interesting. We observe that when x0 > 2 , the maximal x = x0 − 

2 while when x0 < − 2 the

maximal x = x0 + 2 . Thus, we again have:

5 
u2 (x0 ) = −|x0 | + , for |x0 | >
4 2

In the middle region there is a change and the optimal x is actually x = 0. In that case, we see

that
1 
u2 (x0 ) =  − x0 2 , for |x0 | ≤ .
 2

Observe that this function is C 1,1 everywhere.

101
Theorem 25 (Aleksandrov). If f is a convex function on Ω ⊂ Rd , then f is twice differentiable

almost everywhere.

Remark 12.5. A proof of this result can be found in a paper of Crandall, Ishii, and Lions on

viscosity solutions. Perhaps we will look at it at a later time.

Corollary 12.6. If f is C 1,1 from below, then f is twice differentiable almost everywhere.

Proof. C 1,1 from below means, in the notation of the first section on viscosity solutions, that

Θ̄(f, ) ≥ −M for some , M > 0. That is, the second order difference quotients are bounded

from below, or there are parabloids of fixed opening M touching f from below at any point

in the domain. Therefore, f + M |x|2 is convex. The result now follows from the Aleksandrov

theorem.

We will use this to say that the upper envelope u is in fact twice differentiable almost

everywhere.

Theorem 26. The upper -envelope u enjoys the following properties.

1. u ∈ C(H) and u decreases to u uniformly in H as  → 0.

2. For any x0 ∈ H, there is a concave paraboloid of opening 2/ that touches u from below

at x0 in H. Hence, u is C 1,1 from below.

3. Suppose that u is a viscosity subsolution of F (D2 u) = 0 in Ω and that H1 ⊂⊂ H is

open. Then there exists 0 depending on u, H, H0 so that for all  ≤ 0 , u is a viscosity

subsolution of F (D2 u) = 0 in H1 . In particular, F (D2 u ) ≥ 0 almost everywhere in H1 .

Toward establishing the theorem, let us mention a few useful properties of the upper envelope.

Lemma 12.7. Assume x0 , x1 ∈ H. We have that

1. There exists x∗0 ∈ H̄ so that u (x0 ) = u(x∗0 ) +  − 1 |x∗0 − x0 |2 .

2. u (x0 ) ≥ u(x0 ) + .

102
 3. |u (x0 ) − u (x1 )| ≤ 3 diam(H)|x0 − x1 |.

0
4. If 0 <  < 0 u (x0 ) ≤ u (x0 ).

5. |x∗0 − x0 |2 ≤ oscH (u).

6. 0 < u (x0 ) − u(x0 ) ≤ u(x∗0 ) − u(x0 ) + .

Proof of Lemma 12.7. Points 1,2,4,6 are obvious from the definition. Now let us move to estab-

lishing 3. We have that

1 1 1 2
u (x0 ) ≥ u(x) +  − |x − x0 |2 ≥ u(x) +  − |x − x1 |2 − |x0 − x1 |2 − |x0 − x1 ||x − x1 |
   

1 3
≥ u(x) +  − |x − x1 |2 − diam(H)|x1 − x0 |
 

Therefore,
3
u (x0 ) ≥ u (x1 ) − diam(H)|x1 − x0 |.


Finally, to prove (5), observe that

1 ∗
|x − x0 |2 = u(x∗0 ) +  − u (x0 ) ≤ u(x0 ∗) − u(x0 ) ≤ osc(u).
 0

Next, we move to prove the Theorem.

Proof of the Theorem 26. The Lipschitz continuity of u follows from the third property in

Lemma (12.7). The other properties imply the uniform convergence u → u. Now for fixed

x0 ∈ H, we define
1
P0 (x) = u(x∗0 ) +  − |x − x∗0 | ≤ u (x),


for any x while P0 (x0 ) = u (x0 ). It follows that u is C 1,1 from below and thus almost every-

where twice differentiable by the Aleksandrov Theorem. Now we prove that u is a subsolution

103
whenever u is. Assume that F (D2 u) ≥ 0 in the viscosity sense on H. Fix x0 ∈ H1 ⊂⊂ H and

suppose that P is a paraboloid touching u from above at x0 . Now consider

1
Q(x) = P (x + x0 − x∗0 ) + |x0 − x∗0 |2 − .


Note that Q is just a horizontal and vertical translation of P . Now, we can ensure that if

x0 ∈ H1 , then x∗0 ∈ H for  ≤ 0 . Now if x is very close to x∗0 we have that

1
u (x + x0 − x∗0 ≥ u(x) +  − |x0 − x∗0 |2 ,


which implies that

1
u(x) ≤ u (x + x0 − x∗0 ) + |x0 − x∗0 |2 − .


Thus, if x is close to x∗0 ,

1
u(x) ≤ P (x + x0 − x∗0 ) + |x0 − x∗0 |2 −  = Q(x),


while u(x∗0 ) = Q(x∗0 ), since P (x0 ) = u (x0 ). Thus, Q touches u from above at x∗0 . Thus,

0 ≤ F (D2 Q) = F (D2 P ).

Thus, F (D2 u ) ≥ 0 in the viscosity sense.

Now we move to prove the following result:

Theorem 27. Let u be a viscosity subsolution of F (D2 w) = 0 in Ω and let v be a viscosity

supersolution. Then,
λ
u − v ∈ S( , Λ)
n

in Ω.

Proof. By approximation, it suffices to show that u − v ∈ S( nλ , Λ) in H1 ⊂⊂ H ⊂⊂ Ω, where

104
u is the upper  envelope of u and v is the lower  envelope of v. This is because uniform limits

of viscosity subsolutions on compact sets are viscosity subsolutions, as we showed in the past.

Toward showing that u − v ∈ S( nλ , Λ), Let P be a paraboloid touching u − v from above

at x0 ∈ H1 . We need to show that M+ (D2 P, nλ , Λ) ≥ 0. Let us assume that Br (x0 ) ⊂ H1 ,

B2r (x0 ) ⊂ H. Let δ > 0 and define:

w(x) = v (x) − u (x) + P (x) + δ|x − x0 |2 − δr2 .

Observe that w ≥ 0 on ∂Br (x0 ) and w(x0 ) < 0, since P touches u − v from above. Since

w ∈ C 1,1 almost everywhere, we can apply one step of the Alexandrov Maximum Principle to

w. Namely, we have that

Z
−
0 < sup w ≤ det(D2 Γw ).
Br (x0 )∩{w=Γw }

Now, we have that

F (D2 v ) ≤ 0 F (D2 u (x)) ≥ 0,

for almost every x ∈ Br (x0 ). Since Γw is convex and Γw ≤ w, we have that

D2 w ≥ 0

for almost every x ∈ {w = Γw }. From the lower bound of the integral, we see that

|{w = Γw }| > 0,

so that there exists x1 ∈ {w = Γw } where w is twice differentiable. Thus we have that

0 ≤ F (D2 u (x1 )) = F (D2 v (x1 ) − D2 w(x1 ) + D2 P + 2δI)

≤ F (D2 v (x1 ) + D2 P + 2δI) ≤ F (D2 v (x1 ) + D2 P ) + 2Λδ

105
 ≤ F (D2 v (x1 )) + Λ||(D2 P )+ || − λ||(D2 P )− || + 2Λδ

≤ M+ (D2 P, λ/n, Λ) + 2Λδ.

Sending δ → 0 completes the proof.

12.2 C 1,α regularity of solutions to F (D2 u)

A direct corollary of the preceding theorem is

Corollary 12.8. Let u be a viscosity solution of F (D2 u) = 0 in Ω. Assume that h ∈ R\{0}, e ∈

1
Rn is a fixed unit vector, and that 0 ≤ β ≤ 1. Define uh,e,β u by uh,e,β (x) = |h|β
(u(x+he)−u(x)).

Then,
λ
uh,e,β ∈ S( , Λ).
n

Corollary 12.9. We have that

|uh,e,β |C α (Br ) ≤ Cr,s (|uh,e,β |L∞ (Bs ) ,

where α depends only on n, λ, Λ and C depends only on n, λ, Λ, s, r.

Lemma 12.10. Let 0 < α < 1, 0 < β ≤ 1 and K > 0 be constants. Let u ∈ L∞ ([−1, 1] and

assume that |u|L∞ ≤ K. Consider uh,β := uh,e1 ,β . Assume that

|uh,β |C α ≤ K,

independent of h.

• If α + β < 1, then u ∈ C α+β ([−1, 1]) and |u|C α+β ≤ CK.

• If α + β > 1, then u ∈ C 0,1 ([−1, 1]) and |u|C 0,1 ≤ CK,

where C depends only on α + β.

Proof. Exercise.

106
 Noting that uh,e,α ∈ L∞ (B1 ) by the Harnack inequality (+oscillation Lemma etc), we can

apply the above corollary finitely many times (a universal number) to deduce that uh,e,1 ∈ L∞

on B3/4 from which we see:

|uh,e,1 |C α (B1/2 ) ≤ |u|C 0,1 (B3/4 ) .

Now we arrive at the following result.

Corollary 12.11. Let u be a viscosity solution of F (D2 u) = 0 in B1 . Then, we have that

u ∈ C 1,α (B1 ) and

|u|C 1,α (B1/2 ) ≤ C(|u|L∞ (B1 ) + |F (0)|),

where C, α > 0 are universal constants.

12.3 The case of Concave Equations

Definition 12.12. We say that the equation F (D2 u) = 0 is concave if

M +N F (M ) + F (N )
F( )≥ .
2 2

An important consequence of concavity is that ∂ee u is a subsolution whenever u is a solution.

Assuming that the solution is sufficiently smooth, this is just the following computation first in

one dimension

(F (u00 ))00 = F 0 (u)(u00 )00 + F 00 (u00 )(u0 )2 = 0,

from which we see that

F 0 (u00 )(u00 )00 ≥ 0,

using that F is concave. In the multidimensional case, we observe that

∂ee (F (D2 u)) = ∂e (Fij (D2 u)(∂e ∂ij u)) = (Fij )(D2 u)∂ij (∂ee u) + Fij,kl (D2 u)∂ij ∂e u ∂kl ∂e u = 0.

Note that Fij is the derivative of F with respect to the ij variable of F . Now concavity implies

107
that the second quantity is non-positive. To see this in the general viscosity case, we will rely

on the following Theorem applied in a suitable way to the second order difference quotients of

u.

Theorem 28. Let F be concave and let u and v be subsolutions of F (D2 w) = 0 in Ω. Then,
1
2 (u + v) is a viscosity suboslution of F (D2 w) = 0 in Ω.

Corollary 12.13. Assume that F is concave and that F (D2 u) = 0 in the viscosity sense. If e

is a fixed unit vector, h > 0, then

u(x + he) + u(x − he) − 2u(x) λ

2
∈ S( , Λ).
h n

Corollary 12.14. If u ∈ C 2 (Ω) is a solution of F (D2 u) = 0 in Ω, then

λ
∂ee u ∈ S( , Λ)
n

in Ω.

The first corollary follows since 12 (u(x+eh)+u(x−eh)−2u(x)) can be written as the difference

of a viscosity subsolution and supersolution (because of the above theorem) of F (D2 w) = 0.

The second corollary follows from the first since the class S is closed under uniform limits. Now

we turn to the proof of the Theorem, which again uses the  envelope.

Proof of the Theorem. We need to show that 21 (u +v  ) is a viscosity subsolution of F (D2 w) = 0.
1 
Let P be a paraboloid that touches 2 (u + v  ) from above at x0 . We need to show that

F (D2 P ≥ 0). Define

1
w(x) = P (x) − (u + v  ) + δ|x − x0 |2 − δr2 .
2

Applying the exact same strategy as we did before, we find an x1 where u , v  , w are all second

108
order differentiable at x1 and for which

1
D2 (P + δ|x − x0 |2 − (u + v  ))(x1 ) = D2 w(x1 ) ≥ 0,
2

while F (D2 u (x1 )) ≥ 0 and F (D2 v  (x1 )) ≥ 0. Since F is concave, we have that

1
F (D2 (u + v  )(x1 )) ≥ 0.
2

Thus,

F (D2 P + 2δI) ≥ 0.

Now we send δ → 0.

13 The Evans-Krylov Theorem, C 2,α regularity for solutions to

Concave equations

The goal of this section is to establish C 2,α regularity of viscosity solutions to F (D2 u) = 0 when

F is concave. We start by showing that C 2 solutions are actually C 2,α .

13.1 Evans-Krylov Theorem (C 2 to C 2,α )

In this subsection we will establish the following Theorem.

Theorem 29. Let F be concave and let u ∈ C 2 satisfy

F (D2 u) = 0

in B1 . Then, u ∈ C 2,α (B1 ) and

|u|C 2,α (B1/2 ) ≤ C|u|C 1,1 (B3/4 ).

The proof will follow once we establish the following Lemma.

109
Lemma 13.1. Under the hypotheses of the Theorem, there exists a universal constant 0 < δ0 < 1

so that osc(D2 u(B1 )) = 2 implies osc(D2 u(Bδ0 )) ≤ 1.

Note that the assumption that the oscillation of D2 u being bounded by 2 can be easily
1
ensured by dividing u by a constant. Note that t F (t·) has the same ellipticity constants as

F (·). The main tool we will use to establish the Theorem is that we already know that

λ
∂ee u ∈ S( , Λ).
n

We first “recall” the following result that we established in the course of proving the Harnack

inequality:

Lemma 13.2. Let v ∈ S̄(λ, Λ, 0) in B1 and assume that v ≥ 0 in B1 . Then,

inf v ≥ C|{v ≥ 1} ∩ B1/4 |δ ,

B1/2

with C and δ universal.

Let us now observe the following simple consequence of ellipticity.

Lemma 13.3. Assume that F is uniformly elliptic. If F (M1 ) = F (M2 ) = 0, then

λ
kM2 − M1 k ≤ k(M2 − M1 )+ k = sup ((M2 − M1 )e, e)+ .
λ+Λ |v|=1

Remark 13.4. The second equality follows from diagonalizing the matrix M2 − M1 . The first

inequality is the key and it is intuitively clear. Indeed, recall that F (·) is “like” a trace operator

and so F (M1 ) = F (M2 ) = 0 implies that the positive and negative eigenvalues of M1 and M2

are controlling each other in a sense.

Proof. Observe that

0 = F (M2 ) ≤ F (M1 )+Λk(M2 −M1 )+ k−λk(M2 −M1 )− k ≤ (Λ+λ)k(M2 −M1 )+ k−λkM2 −M1 k,

110
where we used the triangle inequality in the second inequality.

Now we come to the key Lemma.

Lemma 13.5 (Key Lemma). Under the hypotheses of the Theorem, assume also that

1 < diam(D2 u(B1 )) ≤ 2

and that D2 u(B1 ) is covered by N balls {B i }N

i=1 of maximal radius not exceeding 0 , where 0 is

universal. Then, D2 u(B1/2 ) is covered by N − 1 balls among the B i .

Assuming this Lemma holds, let us now prove the main oscillation lemma.

Oscillation Lemma. Since diam(D2 u(B1 )) = 2, we can cover D2 u(B1 ) by N balls of radius 0 ,

where 0 is as in the last Lemma (N is universal since 0 is). By the last Lemma, we know that

D2 u(B1/2 ) is covered by N − 1 balls of radius 0 . If diamD2 u(B1/2 ) ≤ 1 we are done. If not,

let us define

w(x) = 4u(x/2),

for x ∈ B1 . Then, F (D2 w) = 0 in B1 and diam(D2 w(B1 )) ≤ 2. It follows that D2 u(B1/4 ) =

D2 w(B1/2 ) can be covered by N − 2 of the original balls. Thus, there exists k ≤ N so that

diam(D2 u(B1/2k )) ≤ 1.

Thus we are done once we establish the key lemma.

Now we turn to prove the Key Lemma.

λ c0
Proof of the Key Lemma. Let c0 = Λ+λ (as in Lemma 13.3). Take 0 ≤ 32 . By assumption, we

thus have points xi so that if

Mi = D2 u(xi ),

D2 u(B1 ) ⊂ ∪N
i=1 Bc0 /16 (Mi ).

111
In fact, since the radii are now fixed at c0 /16, we have that

0
D2 u(B1 ) ⊂ ∪N
i=1 Bc0 /8 (Mi ),

where N 0 ≤ N is universal. It follows that there exists one of the Mi , let us say it is M1 , for

which

|(D2 u)−1 (Bc0 /8 (M1 )) ∩ B1/4 | ≥ η > 0,

where η is universal. This is just because they are finitely many and {(D2 u)−1 (Bc0 /8 (Mi ))}
1
cover B1 . Since the diameter of D2 u(B1 ) > 1 and since 0 ≤ 32 , there must exist M2 so that

|M2 − M1 | ≥ 41 . Thus, by the Ellipticity Lemma 13.3, there exists a unit vector e ∈ Rn with

c0
uee (x2 ) ≥ uee (x1 ) + .
4

Now define

K = sup uee , v = K − uee .

B1

Using the concavity of F , we have that 0 ≤ v ∈ S̄(λ/n, Λ, 0) in B1 . Since the measure of points

for which D2 u(x) is within c0 /8 of D2 u(x1 ) is larger than η and since uee (x2 ) ≥ uee (x1 ) + c0 /4,

we have that

|{v ≥ c0 /8} ∩ B1/4 | ≥ η.

It follows now that there is a universal constant C1 so that

K − sup uee = inf (K − uee ) = inf v ≥ C1 > 0.

B1/2 B1/2 B1/2

Now since B2 (Mi ) cover D2 u(B1 ), we can find a single j so that

K − uee (xj ) < 3.

Observe now that for all x ∈ B1/2 we have that (D2 u(x)e, e) ≤ K − C1 while all matrices in

112
M ∈ B2 (Mj ) satisfy (M e, e) > K − 5. Now we take 50 ≤ C1 and it follows that B2 (Mj ) ∩

D2 (B1/2 ) = φ. This completes the proof.

13.2 L∞ to C 2

We will now prove a stronger version of the preceding Theorem: that viscosity solutions are C 2,α

and the C 2,α norm of solutions to concave equations can be controlled just by its L∞ norm.

Theorem 30. Let F be concave and u be a viscosity solution of F (D2 u) = 0 in B1 . Then

2,α
u ∈ Cloc (B1 ) and

|u|C 2,α (B1/2 ≤ C(|u|L∞ (B1 ) + |F (0)|),

where 0 < α < 1 and C > 0 are universal constants.

Remark 13.6. The existence of a classical solution, uniqueness of the viscosity solution, and

the Evans-Krylov Theorem imply the above Theorem. The proof given by Caffarelli and Cabre

of the above theorem, however, is quite soft and independently interesting (it also seems to

apply to a wider class of F than what was previously obtained using existence, uniqueness,

and Evans-Krylov). Why does the a-priori bound follow from Evans-Krylov? It is similar to a

computation we did previously in the proof of the De Giorgi theorem.

Remark 13.7. We may assume that F (0) = 0. Indeed, for a general elliptic F , there exists
|F (0)|
a unique t ∈ R so that F (tI) = 0 while |t| ≤ λ . This just follows from the definition of

ellipticity. Now if we define the paraboloid P (x) = 2t |x − x0 |2 , we see that we can define

F (D2 u) = F (D2 (u − P ) + tI) := G(D2 (u − P )),

where now we have G(0) = 0 and G is uniformly elliptic.

Let us start with a sketch of the proof. The first step is to observe that if F (0) = 0, we can

find a supporting hyperplane for F at 0. That is, we can find a linear functional L on the space

of symmetric matrices for which:

L(M ) ≥ F (M ).

113
Since L is linear, we must have L(M ) = tr(AM ). Since F is elliptic, it is easy to check that A

is positive definite with eigenvalues in [λ, Λ] by a rotation, we can assume that L(M ) = tr(M ).

It follows that any solution to F (D2 u) = 0 satisfies ∆u ≥ 0 in the viscosity sense. We will then

be able to show that

Z
u(x0 ) ≤ − u,
∂Bh (x0 )

for all x0 ∈ B1/2 . From there we will observe that the functions u∗h (x) = 1 −
R
(
h2 ∂Bh (x0 )
u − u(x))

belong to the class S. A simple argument will show that u∗h ∈ L1 uniformly in h, from which

we will see that it is uniformly L∞ in h (Because it is an element of S). This will then allow us

to say that ∆u ∈ L∞ and eventually D2 u ∈ L∞ . From there, we will modify slightly the proof

of the Evans-Krylov theorem to allow for u ∈ C 1,1 rather than u ∈ C 2 .

We will now start the formal proof.

Proof. The proof will proceed in steps. First observe that by scaling we may as well assume that

|u|L∞ ≤ 1 and so our task will be to show first that |u|C 1,1 (B1/2 ≤ C and then |u|C 2,α (B1/2 ≤ C,

for some universal α, C.

Step 1: We may assume that ∆u ≥ F (D2 u) whenever u is C 2

As described above, there exists a linear functional L so that

L(M ) ≥ F (M ),

for all symmetric matrices M . Since L is linear we can write:

X
L(M ) = aij mij = tr(AM ).
i,j

Now we just observe that if ξ ∈ Rd , the positive definite matrix ξξ t satisfies:

aij ξi ξj = L(ξξ t ) ≥ F (ξξ t ) ≥ F (0) + λ|ξξ t |2 = λ|ξ|2 ,

114
by the definition of ellipticity. Similarly,

aij ξi ξj = −L(−ξξ t ) ≤ −F (−ξξ t ) ≤ −F (0) + Λkξξ t k = Λ|ξ|2 .

It follows that A is symmetric and positive definite. We can thus change coordinates so that A

is diagonal and further change coordinates (by scaling) so that A is the identity. Note that this

change of coordinates is universal since the eigenvalues of A are in [λ, Λ]. It follows that we may

assume that A = Id. Note that |u|C 2,α in this coordinate is equivalent to |u|C 2,α in the usual

coordinates.

Step 2: ∆u ≥ 0 and mean value property

We thus see that for any C 2 function, ∆φ ≥ F (D2 φ). Now, if u is a viscosity solution, we

have that for any φ touching u from above, ∆φ ≥ F (D2 φ) ≥ F (D2 u) = 0. Thus, u satisfies

∆u ≥ 0

1
in the viscosity sense. Now we claim that for any x0 ∈ B1/2 and 0 < h < 2 we have that

Z
u(x0 ) ≤ − u.
∂Bh (x0 )

Indeed, for fixed h, let ∆w = 0 in Bh (x0 ) and h = w on ∂Bh (x0 ). By the maximum principle

for viscosity subsolutions (from the ABP estimate), we have that u ≤ w. Thus, in particular,

Z
u(x0 ) ≤ w(x0 ) = − u,
∂Bh (x0 )

where we used the mean value property on the harmonic function w and w = u on ∂Bh (x0 ).
R 
Step 3: Properties of u∗h (x) = 1
h2
−
∂Bh (x) u − u(x)

Our goal will be to show that the function u∗h (x) is uniformly in L1 .

Note that the function u∗h is continuous and nonnegative on B1/2 for any 0 < h < 1/2.

115
Observe that if φ ∈ C 2 we have that φ∗h → 1
2n ∆φ on compact sets. This can be seen easily

using a Taylor expansion of u and can just be checked for quadratic polynomials. Note also that

|φ∗h |L∞ ≤ Cn |D2 φ|L∞ .

Note now that u∗h ≥ 0 so we can bound |u∗h |L1 as follows. Fix φ ≥ 0 be in Cc∞ (B1/2 ) with

φ ≡ 1 on B1/3 . Then, since |u|L∞ (B1 ) ≤ 1,

Z Z Z Z Z Z  Z
1
|u∗h | = u∗h ≤ u∗h φ = − u− φ−u(x)φ(x) = uφ∗h ≤ C(n)|D2 φ|L∞ ≤ C(n).
B1/3 B1/3 B1/2 h2 B1/2 Bh (x) Bh (x) B1

Step 4: u∗h ∈ S(λ/n, Λ) and application

Note that u∗h is an average of u minus u. We want to see that the average of u is a subsolution

of F (D2 u) = 0. Since F is concave, if u1 , ..., uk are viscosity subsolutions then ki=1 λi ui is also
P

P
a viscosity subsolution whenever λi ≥ 0 and λi = 1. Observe that

Z Z
− u=− u(y + x)dy.
Sh (x) Sh (0)

Now note that u(· + y) is a subsolution for each y ∈ Sh (0). Now by the closeness of of the class

of viscosity subsolutions of F (D2 u) = 0 under uniform limits, we see that

Z
h2 u∗h = (− u − u(x))

is a difference of a viscosity subsolution and supersolution of F (D2 u) = 0. It follows that

u∗h ∈ S(λ/n, Λ).

It thus follows from the proof of the Harnack inequality that

|u∗h |L∞ (B1/4 ) ≤ C|u∗h |L1 (B1/3 ) ≤ C,

for C independent of h.

116
 Step 5: u ∈ C 1,1

It follows from the preceding bound that ∆u ∈ L∞ (B1/4 ). Indeed, let φ ∈ Cc∞ (B1/4 ). We

know that 2nψh∗ → ∆φ uniformly. Thus we see that

Z Z Z
u∆ψ = 2n lim uψh∗ ≤ 2n sup | u∗h ψ| ≤ C|h|L1 .
h→0 h B1/4

It follows by duality that

|∆u|L∞ (B1/4 ) ≤ C.

Thus, by the standard elliptic regularity theory, u ∈ H 2 (B1/5 ) and

|D2 u|L2 (B1/5 ≤ C.

Now, we know that the second order difference quotients

1
∆2he u(x) = (u(x + he) + u(x − he) − 2u(x)]
h2

belong to S(λ/n, Λ), since F is concave. Thus, since u ∈ H 2 (B1/5 ), we have that ∆2he u ∈
1
L2 (B1/10 ) for any 0 < h < 10 . It thus follows from the Harnack inequality that

sup ∆2he u ≤ C.
B1/11

1
for all 0 < h < 10 . Thus, u − C2 |x|2 is concave and so u is twice differentiable almost everywhere

by the Alexandrov Theorem and

uee (x) ≤ C

almost everywhere for any |e| = 1. Since F (D2 u) = 0 in the viscosity sense and u is twice

differentiable almost everywhere, F (D2 u(x)) = 0 almost everywhere. Using the definition of

117
ellipticity we see that

|D2 u(x)| ≤ C sup uee (x)+ ≤ C.

|e|=1

Thus,

|u|C 1,1 (B1/11 ) ≤ C,

and that u is twice differentiable on a set A of full measure in B1/11 where F (D2 u(x)) = 0 and

uee (x) = limh→0 ∆2he u(x).

118