Lecture notes on viscosity solutions

Jeff Calder

University of Minnesota
School of Mathematics
[email protected]

July 30, 2024

2
Contents

1 Introduction 5
1.1 An example . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.2 Motivation via dynamic programming . . . . . . . . . . . . . . 9
1.3 Motivation via vanishing viscosity . . . . . . . . . . . . . . . . 10
1.4 Motivation via the maximum principle . . . . . . . . . . . . . 11

2 Definitions 15

3 A comparison principle 23

4 The Hamilton-Jacobi-Bellman equation 31

5 Convergence of vanishing viscosity 39

5.1 Weak upper and lower limits . . . . . . . . . . . . . . . . . . . 40
√
5.2 The O( ε) rate . . . . . . . . . . . . . . . . . . . . . . . . . . 43
5.3 Semiconcavity and an O(ε) one-sided rate . . . . . . . . . . . 45

6 Boundary conditions 53
6.1 Time-dependent Hamilton-Jacobi equations . . . . . . . . . . 58
6.2 The Hopf-Lax Formula . . . . . . . . . . . . . . . . . . . . . . 60

7 The Perron method 63

8 Smoothing viscosity solutions 67

9 Finite difference schemes 77

9.1 Monotone schemes . . . . . . . . . . . . . . . . . . . . . . . . 78
9.2 Convergence of monotone schemes . . . . . . . . . . . . . . . . 81
9.3 Local truncation
√ error . . . . . . . . . . . . . . . . . . . . . . 85
9.4 The O( h) rate . . . . . . . . . . . . . . . . . . . . . . . . . . 87
9.5 One-sided O(h) rate . . . . . . . . . . . . . . . . . . . . . . . 93

3
4 CONTENTS

10 Homogenization 97

11 Discontinuous coefficients 103

12 Second order equations 111

12.1 Jensen’s Lemma . . . . . . . . . . . . . . . . . . . . . . . . . . 112
12.2 Comparison for continuous functions . . . . . . . . . . . . . . 115
12.3 Superjets and subjets . . . . . . . . . . . . . . . . . . . . . . . 117
12.4 Semicontinuous comparison . . . . . . . . . . . . . . . . . . . 120
12.5 A problem on an unbounded domain . . . . . . . . . . . . . . 124
Chapter 1

Introduction

These notes are concerned with viscosity solutions for fully nonlinear equa-
tions. A majority of the notes are concerned with Hamilton-Jacobi equations
of the form
H(Du, u, x) = 0.
First order equations generally do not admit classical solutions, due to the
possibility of crossing characteristics. On the other hand, there are infinitely
many Lipschitz continuous functions that satisfy the equation almost every-
where. Since the equation is nonlinear, we cannot define weak solutions via
integration by parts. In this setting, the correct notion of weak solution is the
viscosity solution, discovered by by Crandall, Evans and Lions [5,7]. At a high
level, the notion of viscosity solution selects, from among the infinitely many
Lipschitz continuous solutions, the one that is ‘physically correct’ for a very
wide range of applications.
Viscosity solutions have proven to be extremely useful, and this is largely
because very strong comparison and stability results are available via the max-
imum (or comparison) principle. As we shall see, these results come almost
directly from the definitions. As such, viscosity solutions could easily have
been called “comparison solutions” or “L∞ -stable solutions”. The term “vis-
cosity” comes from the original motivation for the definitions via the method
of vanishing viscosity (see Section 1.3 and Chapter 5). Viscosity solutions
have a wide range of applications, including problems in optimal control the-
ory. A good reference for the first order theory is the book by Bardi and
Capuzzo-Dolcetta [1], and Evans [11, Chapter 10].
Since viscosity solutions are defined by, and based upon, the maximum
principle, it is natural that they extend to fully nonlinear second order equa-
tions of the form
F (D2 u, Du, u, x) = 0,

5
6 CHAPTER 1. INTRODUCTION

provided F satisfies some form of ellipticity. However, in the early days of the
theory, it was not clear that uniqueness would hold for second order equations,
since the standard proof of uniqueness for first order equations does not directly
extend. The first uniqueness result for second order equations is due to Jensen
[13], and his role in the theory is immortalized in Jensen’s Lemma (see Lemma
12.1), which is a crucial technical tool in the second order theory. Good
references for second order theory include the User’s Guide [6], Crandall’s
introductory paper [4], and the book by Katzourakis [14].
These notes were designed to illustrate the theory and applications of vis-
cosity solutions. They are written in a lecture style and are not meant to be a
thorough reference. We do prove the comparison principle for first and second
order equations in full generality for semi-continuous sub- and supersolutions.
When considering applications, we take simple settings where the main ideas
are present, but the proofs are particularly simple. Almost all of the applica-
tions (e.g., convergence rates, homogenization, etc.) can be stated and proved
in far more generality. However, the ideas in these notes contain the essence
of the key tools for many of these problems.
The organization of these notes is as follows. In Sections 1.1, 1.2, 1.3, and
1.4 we give several different motivational examples leading to the definition of
viscosity solution. In Chapter 2 we give the main definitions of viscosity solu-
tions, and provide a number of interesting exercises. In Chapter 3 we prove the
comparison principle for viscosity solutions of first order equations. In Chap-
ter 4 we discuss the Hamilton-Jacobi-Bellman equation from optimal control
theory in the special case of shortest path problems (i.e., distance functions).
Chapter 5 treats the method of vanishing √viscosity, proving convergence via
the weak upper and lower limits, the O( ε) convergence rate, and a one-
sided O(ε) rate when the solution is semiconcave. In Chapter 6 we briefly
discuss boundary conditions in the viscosity sense. Chapter 7 covers the Per-
ron method for establishing existence of viscosity solutions. In Chapter 8 we
discuss the inf- and sup-convolutions and their role in constructing semiconvex
and semiconcave approximate viscosity sub- and supersolutions. In Chapter 9
we construct√convergent finite difference schemes for viscosity solutions, and
we prove O( h) and one-sided O(h) convergence rates. In Chapter 10 we
give a brief introduction to homogenization, and illustrate the perturbed test
function method. Chapter 11 establishes comparison principles for first or-
der equations with discontinuous coefficients. Finally, in Chapter 12 we prove
the comparison principle for viscosity solutions of second order equations, and
discuss some applications.
While most of the notes address first order Hamilton-Jacobi equations, I
have extended results to second order equations when the proofs are simi-
1.1. AN EXAMPLE 7

lar. In particular, Chapter 7 (the Perron method), Chapter 8 (inf- and sup-
convolutions), and Chapter 12 address general second order equations. Let me
also mention that the references are lacking; in future versions of these notes
I plan to extend the bibliography considerably.

1.1 An example
We begin with a simple example. Let Γ be a closed subset of Rn and let
u : Rn → [0, ∞) be the distance function to Γ, defined by

u(x) = dist(x, Γ) := min |x − y|. (1.1)

y∈Γ

Exercise 1.1. Verify that u is 1-Lipschitz, that is, |u(x) − u(y)| ≤ |x − y| for
all x, y ∈ Rn .

Let U = Rn \ Γ and fix any ball B(x, r) ⊂ U . We claim that

u(x) = r + min u(z). (1.2)

z∈∂B(x,r)

To see this, fix z ∈ ∂B(x, r) minimizing the right hand side of (1.2). Select
y ∈ Γ such that u(z) = |z − y| and compute

u(x) ≤ |x − y| ≤ |x − z| + |z − y| = r + u(z).

For the other direction, fix y ∈ Γ such that u(x) = |x − y|. Let z ∈ ∂B(x, r)
lie on the line segment between x and y. Then

u(x) = |x − y| = |x − z| + |z − y| ≥ r + u(z).

Equation (1.2) is called a dynamic programming principle (DPP), and it

gives a local characterization of the distance function u as the solution of a
Hamilton-Jacobi equation. Indeed, suppose for the moment that u ∈ C 1 (U ).
Notice we can rewrite the DPP as

max {u(x) − u(z) − r} = 0. (1.3)

z∈∂B(x,r)

Since u ∈ C 1 (U ) the Taylor expansion

u(x) − u(z) = Du(x) · (x − z) + o(|x − z|)

8 CHAPTER 1. INTRODUCTION

holds as z → x. Dividing both sides of (1.3) by r and using the Taylor

expansion yields
   
x−z
max Du(x) · − 1 = o (1) .
z∈∂B(x,r) r

Setting a = x−z
r
and sending r → 0+ we deduce

max {Du(x) · a − 1} = 0. (1.4)

|a|=1

The partial differential differential equation (1.4) is called a Hamilton-Jacobi-

Bellman equation, and is a direct consequence of the DPP (1.2). Notice that
(1.4) implies that Du(x) 6= 0, and the maximum occurs at a = Du(x)/|Du(x)|.
We therefore find that )
|Du| = 1 in U
(1.5)
u = 0 on Γ.
Equation (1.5) is a special case of the eikonal equation, which has applications
in geometric optics, wave propagation, level set methods for partial differential
equations, and computer vision.
In general u 6∈ C 1 (U ), so this argument is only a heuristic.
Exercise 1.2. Compute the distance function u to Γ = {0}, Γ = {xn = 0},
Γ = {x1 · · · xn = 0} and Γ = ∂B(0, 1). In which cases does it hold that
u ∈ C 1 (U )?
Exercise 1.3. Show that when U ⊂ Rn is open and bounded, and Γ = ∂U ,
there does not exist a classical solution u ∈ C 1 (U ) ∩ C(U ) of (1.5).
In light of these facts, the natural question is how to rescue this argument
so that it holds when u 6∈ C 1 (U )? Since the argument above relies only on a
first order Taylor expansion of u around x, the argument is valid at all points
of differentiability of u. As u is Lipschitz continuous, u is differentiable almost
everywhere. Thus u satisfies (1.5) almost everywhere, and is called a Lips-
chitz almost everywhere solution. Unfortunately, Lipschitz almost everywhere
solutions are in general not unique.
Exercise 1.4. Consider the following one dimensional version of (1.5):

|u′ (x)| = 1 for x ∈ (0, 1), and u(0) = u(1) = 0. (1.6)

Show that there are infinitely many Lipschitz almost everywhere solutions u
of (1.6).
1.2. MOTIVATION VIA DYNAMIC PROGRAMMING 9

Since our goal is to uniquely characterize u as a solution of (1.5), we need

a notion of solution of Hamilton-Jacobi equations that is weaker than classi-
cal solutions, yet more restrictive than Lipschitz almost everywhere solutions.
Equivalently, we need to discover some additional condition that selects among
the infinitely many Lipschitz almost everywhere solutions the one that is ‘phys-
ically correct’.

1.2 Motivation via dynamic programming

Before giving the definitions, let us proceed further with the distance function
example. As with all notions of weak solutions to partial differential equations,
we will push derivatives off of u and onto a class of smooth test functions.
Since (1.5) is not in divergence form, the classical trick of integration by parts
does not work. Instead, consider the following: Fix a point x ∈ U and let
φ ∈ C ∞ (Rn ) such that u − φ has a local maximum at x. Then we have

u(x) − u(z) ≥ φ(x) − φ(z) for all z near x.

Substituting this into the dynamic programming principle (1.3) we deduce

max {φ(x) − φ(z) − r} ≤ 0,

z∈∂B(x,r)

for r > 0 sufficiently small. Since φ is smooth, the argument in Section 1.1
can be used to conclude that

|Dφ(x)| − 1 = max {Dφ(x) · a − 1} ≤ 0. (1.7)

|a|=1

A similar argument can be used to show that for every x ∈ U and every
φ ∈ C ∞ (Rn ) such that u − φ has a local minimum at x

|Dφ(x)| − 1 = max {Dφ(x) · a − 1} ≥ 0. (1.8)

|a|=1

It is worthwhile taking a moment to observe what is going on here. If u − φ

has a local maximum at x, we can replace φ by φ + C so that u(x) = φ(x)
and u(y) ≤ φ(y) for all y near x. Thus, the graph of φ touches the graph
of u from above at the point x. We have shown above that any such test
function φ must be a subsolution of the PDE at x. Similarly, whenever the
graph of φ touches u from below at x, φ must be a supersolution of the PDE.
See Figure 1.1 for a visual illustration. These ideas should remind you of
maximum principle arguments. Indeed, in the case that u is differentiable at
10 CHAPTER 1. INTRODUCTION

Figure 1.1: An illustration of test functions touching a nonsmooth function u

from above and below. The functions φ1 and φ2 , drawn in red, touch u from
above, while ψ1 and ψ2 , drawn in blue, touch u from below.

x, we have Du(x) = Dφ(x) whenever φ touches u from above or below at x.

These observations say that the function u is not just any Lipschitz solution
of (1.5); it is a Lipschitz solution that respects the maximum principle in a
certain way. We will see shortly that u is the only Lipschitz solution of (1.5)
with these properties.

1.3 Motivation via vanishing viscosity

A general principle of PDEs is that equations are governed primarily by
their highest order terms. With this in mind, consider the semilinear viscous
Hamilton-Jacobi equation

H(Duε , uε , x) − ε∆uε = 0 in U. (1.9)

Since the highest order term in (1.9) is −ε∆uε , which is uniformly elliptic, we
can in very general settings prove existence and uniqueness of smooth solutions
uε of (1.9) subject to, say, Dirichlet boundary conditions u = g on ∂U . In fact,
we did this for a special case of (1.9) using Schaefer’s Fixed Point Theorem
earlier in the course. As a remark, the additional second order term ε∆uε
is called a viscosity term, since for the Navier-Stokes equations such a term
models the viscosity of the fluid.
1.4. MOTIVATION VIA THE MAXIMUM PRINCIPLE 11

Now suppose that uε converges uniformly to a continuous function u as

ε → 0+ . Let x ∈ U and let φ ∈ C ∞ (Rn ) such that u − φ has a local maximum
at x. By replacing φ(y) with φ(y) + |x − y|2 , we may assume that u − φ has
a strict local maximum at x. It follows that there exists a sequence εk → 0+
and xk → x such that uεk − φ has a local maximum at xk .
Exercise 1.5. Prove the preceding statement.
Therefore Duεk (xk ) = Dφ(xk ), ∆uεk (xk ) ≤ ∆φ(xk ), and

H(Dφ(xk ), uεk (xk ), xk ) − εk ∆φ(xk )

≤ H(Duεk (xk ), uεk (xk ), xk ) − εk ∆uεk (xk ) ≤ 0.

Sending εk → 0+ we find that

H(Dφ(x), u(x), x) ≤ 0.

We can similarly argue that whenever u − φ has a local minimum at x

H(Dφ(x), u(x), x) ≥ 0.

Notice we have recovered the same conditions on u as in Section 1.2. The

technique of sending ε → 0 in (1.9) is called the method of vanishing viscosity,
and served as the original motivation for the definition of viscosity solution.

1.4 Motivation via the maximum principle

A less common motivation for the definition of viscosity solution comes from
the maximum principle. Since the well-posedness theory for viscosity solutions
is based on the maximum principle, it is arguably a more important way
of thinking about viscosity solutions compared to the method of vanishing
viscosity.
Suppose that u ∈ C 1 (U ) ∩ C(U ) is a solution of

H(Du, x) = 0 in U. (1.10)

If φ ∈ C ∞ (Rn ) is any function satisfying

H(Dφ, x) > 0 in U,

then we immediately have

max(u − φ) = max(u − φ),

U ∂U
12 CHAPTER 1. INTRODUCTION

that is, the maximum principle holds when comparing u against strict super
solutions. In fact, we can say a bit more. Since we know that Du(x) 6= Dφ(x)
for all x ∈ U , the maximum of u − φ cannot be attained in U . This implies
that
u ≤ φ on ∂U =⇒ u < φ in U.
The observations above hold equally well for any V ⊂⊂ U . That is, if φ ∈
C ∞ (Rn ) satisfies
H(Dφ, x) > 0 in V, (1.11)
then we have
u ≤ φ on ∂V =⇒ u < φ in V. (1.12)
Similarly, if φ ∈ C ∞ (Rn ) satisfies

H(Dφ, x) < 0 in V, (1.13)

then we have
u ≥ φ on ∂V =⇒ u > φ in V. (1.14)
Now suppose we have a continuous function u ∈ C(U ) that satisfies the
maximum (or rather, comparison) principle against smooth strict super and
subsolutions, as above. What can we say about u? Does u solve (1.10) in
any reasonable sense? To answer these questions, we need to formulate what
it means for a continuous function to satisfy the maximum principles stated
above.
For every V ⊂⊂ U we define

S + (V ) = {φ ∈ C ∞ (Rn ) : H(Dφ(x), x) > 0 for all x ∈ V } , (1.15)

and

S − (V ) = {φ ∈ C ∞ (Rn ) : H(Dφ(x), x) < 0 for all x ∈ V } . (1.16)

Let u ∈ C(U ). Suppose that for every V ⊂⊂ U , u satisfies (1.12) for all
φ ∈ S + (V ) and (1.14) for all φ ∈ S − (V ). Such a function u could be called
a comparison solution of (1.10), since it is defined precisely to satisfy the
comparison or maximum principle.
We now derive a much simpler property that is satisfied by u. Let ψ ∈
C ∞ (Rn ) and x ∈ U such that u − ψ has a local maximum at x. This means
that for some r > 0

u(y) − ψ(y) ≤ u(x) − ψ(x) for all y ∈ B(x, r).

1.4. MOTIVATION VIA THE MAXIMUM PRINCIPLE 13

Define
φ(y) := ψ(y) + u(x) − ψ(x).
Then u ≤ φ on the ball B(x, r) and u(x) = φ(x). Therefore, u − φ attains
its maximum over the ball B(x, r) at the interior point x. It follows from our
definition of u that φ 6∈ S + (B 0 (x, r)), and hence

H(Dφ(xr ), xr ) ≤ 0 for some xr ∈ B 0 (x, r).

Sending r → 0 we have xr → x which yields

H(Dψ(x), x) ≤ 0. (1.17)

That is, for any x ∈ U and ψ ∈ C ∞ (Rn ) such that u − ψ has a local maximum
at x we deduce (1.17). It is left as an exercise to the reader to show that
whenever u − ψ has a local minimum at x we have

H(Dψ(x), x) ≥ 0.
14 CHAPTER 1. INTRODUCTION
Chapter 2

Definitions

Let us now consider a general second order nonlinear partial differential equa-
tion
H(D2 u, Du, u, x) = 0 in O, (2.1)
where H is continuous and O ⊂ Rn . We recall that a function u : O ⊂ Rn → R
is upper (resp. lower) semicontinuous at x ∈ O provided

lim sup u(y) ≤ u(x) (resp. lim inf u(y) ≥ u(x)).

O∋y→x O∋y→x

To be precise, we recall the definitions

lim sup u(y) := inf sup{u(y) : y ∈ O ∩ B(x, r)},

O∋y→x r>0

and
lim inf u(y) := sup inf{u(y) : y ∈ O ∩ B(x, r)}.
O∋y→x r>0

Let USC(O) (resp. LSC(O)) denote the collection of functions that are up-
per (resp. lower) semicontinuous at all points in O. We make the following
definitions.

Definition 2.1 (Viscosity solution). We say that u ∈ USC(O) is a viscosity

subsolution of (2.1) if for every x ∈ O and every φ ∈ C ∞ (Rn ) such that u − φ
has a local maximum at x with respect to O

H(D2 φ(x), Dφ(x), u(x), x) ≤ 0.

We will often say that u ∈ USC(O) is a viscosity solution of H ≤ 0 in O when

u is a viscosity subsolution of (2.1).

15
16 CHAPTER 2. DEFINITIONS

Similarly, we say that u ∈ LSC(O) is a viscosity supersolution of (2.1) if

for every x ∈ O and every φ ∈ C ∞ (Rn ) such that u − φ has a local minimum
at x with respect to O

H(D2 φ(x), Dφ(x), u(x), x) ≥ 0.

We also say that u ∈ LSC(O) is a viscosity solution of H ≥ 0 in O when u is

a viscosity supersolution of (2.1).
Finally, we say u is viscosity solution of (2.1) if u is both a viscosity sub-
solution and a viscosity supersolution.
We immediately have comparison against strict super and subsolutions.
Theorem 2.2. Let U ⊂ Rn be open and bounded. Suppose that φ ∈ C ∞ (Rn )
satisfies
H(D2 φ, Dφ, x) > 0 in U (2.2)
and let u ∈ USC(U ) be a viscosity solution of H ≤ 0 in U . Then we have that

max(u − φ) = max(u − φ). (2.3)

U ∂U

Similarly, if
H(D2 φ, Dφ, x) < 0 in U (2.4)
and u ∈ LSC(U ) is a viscosity solution of H ≥ 0 in U then

min(u − φ) = min(u − φ). (2.5)

U ∂U

Proof. Since u − φ is upper semicontinuous, its maximum is attained at some

point x ∈ U . If x ∈ U then by the definition of viscosity subsolution we would
have
H(D2 φ(x), Dφ(x), x) ≤ 0,
which contradicts (2.2) and completes the proof of (2.3). The proof of (2.5) is
similar.
Although we made the definitions for second order equations, we will
mostly study first order Hamilton-Jacobi equations of the form

H(Du, u, x) = 0 in O. (2.6)

Some remarks are in order.

Remark 2.3. The argument given in Section 1.2 shows that the distance
function u defined by (1.1) is a viscosity solution of (1.5).
 17

Remark 2.4. In light of Figure 1.1, when u − φ has a local maximum at x,

we will say that φ touches u from above at x. Likewise, when u − φ has a local
minimum at x, we will say that φ touches u from below at x.

Remark 2.5. It is possible that for some x ∈ O, there are no admissible test
functions φ in the definition of viscosity sub- or supersolution. For example, if
n = 1 and u(x) := |x|, there does not exist φ ∈ C ∞ (R) touching u from above
at x = 0 (why?). Of course, it is possible to touch u(x) = |x| from below at
x = 0 (e.g., take φ ≡ 0). A more intricate example is the function
(
x sin(log(|x|)), if x 6= 0
v(x) =
0, if x = 0.

Even though v is Lipschitz continuous, there are no smooth functions φ touch-

ing from above or below at x = 0.
It turns out the set of points at which there are admissible test functions
is dense (see Exercise 2.18).

Remark 2.6. A viscosity solution u is necessarily continuous, being both

upper and lower semicontinuous. It is very useful to allow viscosity sub- and
supersolutions to be merely semicontinuous. We will see applications of this
later.

Remark 2.7. The condition that φ ∈ C ∞ (Rn ) can be replaced by φ ∈ C(Rn )

and φ is differentiable at x (see [11, Section 10.1.2]). It follows that if u is a
viscosity solution of (2.6), O is open, and u is differentiable at x ∈ U , then

H(Du(x), u(x), x) = 0.

Indeed, we can simply take φ = u in the sub and supersolution properties.

Therefore, a Lipschitz viscosity solution is also an almost everywhere solution.
The converse is not true. A similar remark holds for second order equations,
provided u is twice differentiable at x.

Remark 2.8. The set O ⊂ Rn need not be open. In some settings, we may
take O = U ∪ Γ, where U ⊂ Rn is open and Γ ⊂ ∂U . The reader should note
that u − φ is assumed to have a local max or min at x ∈ O with respect to
the set O. This allows a very wide class of test function when x ∈ ∂O, and
as a consequence, classical solutions on non-open sets need not be viscosity
solutions at boundary points (see Exercise 2.12).
18 CHAPTER 2. DEFINITIONS

Remark 2.9. If u is a viscosity solution of (2.6), then u is generally not a

viscosity solution of
−H(Du, u, x) = 0 in O.
Although this is counterintuitive, it is a necessary peculiarity in the theory of
viscosity solutions, and is important to keep in mind. See Exercise 2.13.

Exercise 2.10. Let U ⊂ Rn be open, and suppose u ∈ C 1 (U ) is a classical

solution of H(Du, u, x) = 0 in U . Show that u is a viscosity solution of
H(Du, u, x) = 0 in U .

Exercise 2.11. Show that the distance function u defined by (1.1) is the
unique viscosity solution of (1.5) in the setting where Γ = ∂U and U is an
open and bounded set in Rn . [Hint: Use Theorem 2.2 and compare u against
a suitable family of strict super and subsolutions of (1.1).]

Exercise 2.12. Show that u(x) := x is a viscosity solution of u′ − 1 = 0 on

(0, 1], but is not a viscosity solution on [0, 1). [Hint: Examine the subsolution
condition at x = 0. This exercise shows that smooth solutions need not be
viscosity solutions at boundary points.]

Exercise 2.13. Verify that u(x) = −|x| is a viscosity solution of |u′ (x)|−1 = 0
on R, but is not a viscosity solution of −|u′ (x)| + 1 = 0 on R. What is the
viscosity solution of the second PDE?

Exercise 2.13 shows that, roughly speaking, viscosity solutions allow ‘cor-
ners’ or ‘kinks’ in only one direction. Changing the sign of the equation reverses
the orientation of the allowable corners.

Exercise 2.14. Let u : (0, 1) → R be continuous. Show that the following are
equivalent.

(i) u is nondecreasing.

(ii) u is a viscosity solution of u′ ≥ 0 on (0, 1).

(iii) u is a viscosity solution of −u′ ≤ 0 on (0, 1).

[Hint: For the hard direction, suppose that u′ ≥ 0 in the viscosity sense on
(0, 1), but u is not nondecreasing on (0, 1). Show that there exists 0 < x1 <
x2 < x3 < 1 such that u(x3 ) < u(x2 ) < u(x1 ). Construct a test function
φ ∈ C ∞ (R) with φ′ < 0 such that φ touches u from below somewhere in the
interval (x1 , x3 ). Drawing a picture might help.]
 19

Exercise 2.15. Let u : (0, 1) → R be continuous. Show that u is convex on

(0, 1) if and only if u is a viscosity solution of −u′′ ≤ 0 on (0, 1). Show that in
general, convex functions are not viscosity solutions of u′′ ≥ 0 on (0, 1). Note:
The PDE u′′ ≥ 0 is not even degenerate elliptic in the sense of (3.4).
[Hint: The hint for the hard direction is similar to Exercise 2.14. Suppose
that −u′′ ≤ 0 on (0, 1) but u is not convex on (0, 1). Then there exists
0 < x1 < x2 < 1 and λ ∈ (0, 1) such that

u(λx1 + (1 − λ)x2 ) > λu(x1 ) + (1 − λ)u(x2 ).

Construct a test function φ ∈ C ∞ (R) with φ′′ < 0 such that φ touches u from
above somewhere in the interval (x1 , x2 ).]
Exercise 2.16. Let u : Rn → R be Lipschitz continuous. Show that u is a
viscosity solution of |Du| ≤ Lip(u) and −|Du| ≥ −Lip(u) on Rn .
Exercise 2.17. We define the superdifferential of u at x to be
n o
D u(x) := p ∈ R : u(y) ≤ u(x) + p · (y − x) + o(|y − x|) as y → x .
+ n

Similarly, we define the subdifferential

n o
D− u(x) := p ∈ Rn : u(y) ≥ u(x) + p · (y − x) + o(|y − x|) as y → x .

Suppose U ⊂ Rn is open. Show that u ∈ USC(U ) is a viscosity subsolution of

H(Du, u, x) = 0 on U if and only if

H(p, u(x), x) ≤ 0 for all x ∈ U and p ∈ D+ u(x).

Similarly, show that u ∈ LSC(U ) is a viscosity supersolution of H(Du, u, x) =

0 in U if and only if

H(p, u(x), x) ≥ 0 for all x ∈ U and p ∈ D− u(x).

[Hint: Show that p ∈ D+ u(x) if and only if there exists φ ∈ C 1 (Rn ) such that
Dφ(x) = p and u − φ has a local maximum at x. A similar statement holds
for the subdifferential.]
Exercise 2.18. Let u ∈ USC(Rn ). Show that the set

A := {x ∈ Rn : D+ u(x) 6= ∅}

is dense in Rn . [Hint: Let x0 ∈ Rn and ε > 0, and consider the maximum of

u − φε over B(x0 , 1), where φε (x) := |x−xε 0 | . Send ε → 0.]
2
20 CHAPTER 2. DEFINITIONS

Exercise 2.19. Let U ⊂ Rn be open. Suppose that u ∈ C(U ) satisfies

Z
u(x) = − u dy + o(ε2 ) as ε → 0+
B(x,ε)

for every x ∈ U . Note this is an asymptotic version of the mean value property.
Show that u is a viscosity solution of

−∆u = 0 in U.

[Hint: Show that for every φ ∈ C ∞ (Rn )

Z
φ(x) − φ(y)
−∆φ(x) = 2(n + 2)− 2
dy + o(1) as ε → 0+ .
B(x,ε) ε

Then verify the viscosity sub- and supersolution properties directly from the
definitions.]

Exercise 2.20.

(a) Let u, v ∈ USC(U ). Suppose that w := u and w := v are viscosity

solutions of
H(D2 w, Dw, w, x) ≤ 0 in U. (2.7)
Show that w(x) := max{u(x), v(x)} is a viscosity solution of (2.7) (i.e.,
the pointwise maximum of two subsolutions is again a subsolution).

(b) Let u, v ∈ LSC(U ). Suppose that w := u and w := v are viscosity solutions

of
H(D2 w, Dw, w, x) ≥ 0 in U. (2.8)
Show that w(x) := min{u(x), v(x)} is a viscosity solution of (2.8).

Exercise 2.21. For each k ∈ N, let uk ∈ C(U ) be a viscosity solution of

H(D2 uk , Duk , uk , x) = 0 in U.

Suppose that uk → u locally uniformly on U (this means uk → u uniformly

on every V ⊂⊂ U ). Show that u is a viscosity solution of

H(D2 u, Du, u, x) = 0 in U.

Thus, viscosity solutions are stable under uniform convergence.

 21

Exercise 2.22. Suppose that p 7→ H(p, x) is convex for any fixed x. Let
u ∈ Cloc
0,1
(U ) satisfy

λu(x) + H(Du(x), x) ≤ 0 for a.e. x ∈ U,

where λ ≥ 0. Show that u is a viscosity solution of

λu + H(Du, x) ≤ 0 in U.

Give an example to show that the same result does not hold for supersolutions.
[Hint: Mollify u: uε := ηε ∗ u. For V ⊂⊂ U , use Jensen’s inequality to show
that
λuε (x) + H(Duε (x), x) ≤ hε (x) for all x ∈ V
and ε > 0 sufficiently small, where hε → 0 uniformly on V . Then apply an
argument similar to Exercise 2.21.]
22 CHAPTER 2. DEFINITIONS
Chapter 3

A comparison principle

The utility of viscosity solutions comes from the fact that we can prove exis-
tence and uniqueness under very broad assumptions on the Hamiltonian H.
Uniqueness of viscosity solutions is based on the maximum principle. In this
setting, the maximum principle gives a comparison principle, which states that
subsolutions must lie below supersolutions, provided their boundary conditions
do as well.
As motivation, let us give the formal comparison principle argument for
smooth sub- and super solutions. Let u, v ∈ C 2 (U ) ∩ C(U ) such that
)
H(D2 u, Du, u, x) < H(D2 v, Dv, v, x) in U
(3.1)
u≤v on ∂U.

We would like to find conditions on H for which (3.1) implies that u ≤ v in U

as well. The classical maximum principle argument examines the maximum
of u − v over U . We may as well assume maxU (u − v) > 0, or else we are done.
Therefore u − v attains a positive maximum at some x ∈ U . At this maximum
we have

u(x) > v(x), Du(x) = Dv(x), and D2 u(x) ≤ D2 v(x). (3.2)

Here, the notation X ≤ Y for symmetric matrices X and Y means that Y − X

is non-negative definite. Writing p = Du(x) = Dv(x) and recalling (3.1) we
deduce
0 < H(D2 v(x), p, v(x), x) − H(D2 u(x), p, u(x), x).
We obtain a contradiction provided

H(X, p, r, x) ≤ H(X, p, s, x) whenever r ≤ s, (3.3)

23
24 CHAPTER 3. A COMPARISON PRINCIPLE

and
H(X, p, z, x) ≥ H(Y, p, z, x) whenever X ≤ Y. (3.4)
The condition (3.4) is called ellipticity, or sometimes degenerate ellipticity.
The condition (3.3) is the familiar monotonicity we encountered when studying
linear elliptic equations
X
n X
n
Lu = − ij
a u xi xj + bi uxi + cu, (3.5)
i,j=1 i=1

where c ≥ 0 was necessary for maximum principle arguments to hold.

When u and v are semicontinuous viscosity solutions, it is possible that
at a maximum of u − v there may be no admissible test functions for u or v
(or both) in the definition of viscosity solution. It is quite extraordinary that
the basic structure of the argument can be rescued in this general setting.
The proof is based on the unusual (and clever) idea of doubling the variables.
We present the argument here in the simplest setting of a bounded domain
U ⊂ Rn in order to convey the main ideas, and present more general cases
later. We also restrict our study to first order equations, where the proofs
are considerably simpler. We refer the reader to the user’s guide [6] for more
details on the comparison principle for second order equations.
We require the additional regularity hypothesis

H(p, z, y) − H(p, z, x) ≤ ω(|x − y|(1 + |p|)) (3.6)

for all x, y ∈ U , z ∈ R, and p ∈ Rn , where ω is a modulus of continuity (i.e.,

ω is nonnegative, ω(0) = 0 and ω is continuous at 0).
Theorem 3.1 (Comparison with strict subsolution). Let U ⊂ Rn be open and
bounded, suppose H = H(p, z, x) satisfies (3.3) and (3.6), and fix ε > 0. Let
u ∈ USC(U ) be a viscosity solution of H ≤ −ε in U and let v ∈ LSC(U ) be a
viscosity solution of H ≥ 0 in U . If u ≤ v on ∂U then u ≤ v on U .
Proof. Assume to the contrary that supU (u − v) > 0. For α > 0 consider the
auxiliary function
α
Φ(x, y) = u(x) − v(y) − |x − y|2 (x, y ∈ U ). (3.7)
2
Since u ∈ USC(U ) and v ∈ LSC(U ), Φ ∈ USC(U × U ). As U is bounded, Φ
assumes its maximum value at some (xα , yα ) ∈ U × U . Note that
α
u(xα ) − v(yα ) − |xα − yα |2 = Φ(xα , yα ) ≥ sup(u − v) > 0.
2 U
 25

As u and −v are bounded above on U we have

C
|xα − yα |2 ≤ .
α
By passing to a subsequence, if necessary, we may assume that xα → x0 and
yα → x0 as α → ∞ for some x0 ∈ U . By the upper semicontinuity of Φ

lim sup Φ(xα , yα ) ≤ u(x0 ) − v(x0 ) ≤ lim inf Φ(xα , yα ).

α→∞ α→∞

Therefore limα→∞ Φ(xα , yα ) = u(x0 ) − v(x0 ) and it follows that

lim α|xα − yα |2 = 0. (3.8)

α→∞

Furthermore, since u ≤ v on ∂U and u(x0 ) > v(x0 ), we must have x0 ∈ U .

Therefore (xα , yα ) ∈ U × U for large enough α.
Consider φ(x) = α2 |x − yα |2 . By the definition of (xα , yα ), u − φ has a local
maximum at the point xα ∈ U . Therefore

H(α(xα − yα ), u(xα ), xα ) ≤ −ε. (3.9)

Likewise, v − ψ has a local minimum at yα , where ψ(y) = − α2 |xα − y|2 . There-

fore
H(α(xα − yα ), v(yα ), yα ) ≥ 0. (3.10)
Since Φ(xα , yα ) ≥ supU (u − v) > 0 we must have u(xα ) > v(yα ). Combining
the monotonicity of H (3.3) with (3.10) we have

H(α(xα − yα ), u(xα ), yα ) ≥ 0. (3.11)

Subtract (3.9) from (3.11) to find

ε ≤ H(α(xα − yα ), u(xα ), yα ) − H(α(xα − yα ), u(xα ), xα )

(3.6)
≤ ω(α|xα − yα |2 + |xα − yα |).

Sending α → ∞ and recalling (3.8) contradicts the positivity of ε.

It is a good idea to master the proof of Theorem 3.1. The main idea is
doubling the number of variables and defining the auxiliary function Φ (3.7).
The rest of the proof boils down to using the boundary conditions to show
that Φ assumes its maximum value at an interior point, and then obtaining
a contradiction from the viscosity sub- and supersolution conditions. The
26 CHAPTER 3. A COMPARISON PRINCIPLE

proof can be extended to unbounded domains, other boundary conditions, and

discontinuous Hamiltonians H by adding appropriate terms to the auxiliary
function.
The comparison result in Theorem 3.1 requires the subsolution to be strict,
i.e. H(Du, u, x) ≤ −ε < 0. This condition can be relaxed provided we can
perturb a viscosity solution u of H ≤ 0 so that it is a strict subsolution. This
requires some sort of strict monotonicity in the Hamiltonian H. We first record
a general result, and then discuss the type of monotonicity conditions one can
place on H.
Corollary 3.2. Let U ⊂ Rn be open and bounded and suppose H satisfies
(3.3) and (3.6). Let u ∈ USC(U ) be a viscosity solution of H ≤ 0 in U and
let v ∈ LSC(U ) be a viscosity solution of H ≥ 0 in U . Suppose there exists a
sequence uk ∈ USC(U ) such that uk → u pointwise on U , uk ≤ v on ∂U , and
each uk is a viscosity solution of H ≤ − k1 in U . Then u ≤ v on U .
Proof. By comparison with strict subsolution (Theorem 3.1), we have uk ≤ v
on U for all k. Since uk → u pointwise on U , u ≤ v on U .
The hypotheses of Corollary 3.2 do not hold in general, and require some
further conditions on H. We record some important cases here.
1. Suppose there exists γ > 0 such that

H(p, z + h, x) − H(p, z, x) ≥ γh (h > 0). (3.12)

This is strict form of the monotonicity condition (3.3). Then the hy-
potheses of Corollary 3.2 hold with uk = u − γk
1
. Notice that the mono-
tonicity condition (3.3) allows H to have no dependence on u, whereas
the strict monotonicity condition (3.12) requires such a dependence. A
special case of (3.12) is a Hamilton-Jacobi equation with zeroth order
term
u + H(Du, x) = 0 in U.

2. Suppose there exists γ > 0 and i ∈ {1, . . . , n} such that

H(p + hei , z, x) − H(p, z, x) ≥ γh (h > 0). (3.13)

Here, e1 , . . . , en are the standard basis vectors in Rn . The hypotheses

−a
of Corollary 3.2 hold with uk = u − xiγk , where a = minU xi . A special
case of (3.13) is the time-dependent Hamilton-Jacobi equation

ut + H(Du, x) = 0 in U × (0, T ).
 27

3. Consider the case where H has no dependence on u, i.e., H(p, z, x) =

H(p, x). Suppose p 7→ H(p, x) is convex, and there exists φ ∈ C ∞ (U )
such that
H(Dφ(x), x) + γ ≤ 0 in U
where γ > 0. Then the hypotheses of Corollary 3.2 hold with

uk = εk φ + (1 − εk )u,

where εk = γk 1
. Note that we can assume that φ ≤ 0, due to the fact
that H has no dependence on u. A special case is the eikonal equation
(1.5), in which case we can take φ ≡ 0.
Exercise 3.3. For each of the cases listed above, verify that uk is a viscosity
solution of H ≤ − k1 in U , uk ≤ u for all k, and uk → u uniformly on U .
The comparison principle from Corollary 3.2 shows that if H satisfies (3.3)
and (3.6), and any one of the conditions listed above holds, then there exists
at most one viscosity solution u ∈ C(U ) of the Dirichlet problem
)
H(Du, u, x) = 0 in U
(3.14)
u = g on ∂U

where g : ∂U → R is continuous and U ⊂ Rn is open and bounded. Indeed,

suppose u, v ∈ C(U ) are viscosity solutions of H = 0 in U . Then by (3.3) u−C
is a viscosity subsolution of H = 0 in U for all C > 0. Setting C = max∂U |u−v|
we have
u − C = u − max |u − v| ≤ v on ∂U.
∂U

By Corollary (3.2), u − C ≤ v in U . Swapping the roles of u and v we have

v − C ≤ u in U . Therefore

max |u − v| ≤ max |u − v|.

U ∂U

Hence, if u = v = g on ∂U , then u = v in U .
Exercise 3.4. Show that the following PDEs are degenerate elliptic.
P
(i) The linear elliptic operator (3.5), provided ni,j=1 aij ηi ηj ≥ 0.
(ii) The Monge-Ampère equation

− det(D2 u) + f = 0,

provided u is convex.
28 CHAPTER 3. A COMPARISON PRINCIPLE

(iii) The ∞-Laplace equation −∆∞ u + f = 0, where

1 X
n
∆∞ u := ux x u x ux .
|Du|2 i,j=1 i j i j

(iv) The level-set equation for motion by mean curvature

1 X
n
ut − ∆u + ux x ux ux = 0.
|Du|2 i,j=1 i j i j

[Hint: For (ii), use the linear algebra identity

Y
n
det(A) = min viT Avi ,
v1 ,...,vn
i=1

where A ≥ 0 is symmetric, and the minimum is over all orthonormal

bases of Rn . This is closely related to Hadamard’s determinant inequality.
For (iv), write the equation as
X
n
ut − aij (Du)uxi xj = 0,
i,j=1

where u xi u xj
aij (Du) = δij − .
|Du|2
Verify the ellipticity when Du 6= 0. (Remark: To handle Du = 0, we
redefine viscosity solutions by taking the upper and lower semicontinuous
envelopes of
Xn
F (X, p) = aij (p)Xij ,
i,j=1

in the sub- and supersolution properties, respectively.)]

Exercise 3.5. Consider the Hamilton-Jacobi equation
( √
|Du| − u = 0 in B 0 (0, 1)
(H)
u = 0 on ∂B(0, 1).
(a) Show that there are infinitely many nonnegative viscosity solutions u ∈
C(B(0, 1)) of (H). [Hint: For every 0 ≤ λ ≤ 1, show that
1
uλ (x) := max{0, λ − |x|}2
4
is a viscosity solution of (H).]
 29

(b) Explain why uniqueness fails for (H), i.e., which hypothesis from Theorem
3.1 is not satisfied.

(c) Show that u1 (i.e., uλ with λ := 1) is the unique viscosity solution of (H)
that is positive on B 0 (0, 1). [Hint: Show that if u e is√another viscosity
√
solution of (H) that is positive on B 0 (0, 1), then w := 2 u and w e := 2 u e
0
are both viscosity solutions of the eikonal equation in B (0, 1).]
30 CHAPTER 3. A COMPARISON PRINCIPLE
Chapter 4

The Hamilton-Jacobi-Bellman
equation

We now aim to generalize the distance function example in Section 1.1. Con-
sider the following calculus of variations problem:
n o
T (x, y) = inf I[w] : w ∈ C 1 ([0, 1]; U ), w(0) = x, and w(1) = y , (4.1)

where Z 1
I[w] := L(w′ (t), w(t)) dt. (4.2)
0
Here, U ⊂ Rn is open, bounded, and path connected with a Lipschitz bound-
ary, and x, y ∈ U . We assume that L : Rn × U → R is continuous,
p 7→ L(p, x) is positively 1-homogeneous, (4.3)
and
L(p, x) > 0 for all p 6= 0, x ∈ U . (4.4)
Recall that positively 1-homogeneous means means that L(αp, x) = αL(p, x)
for all α > 0 and x ∈ U .
Let us note a few consequences of these assumptions. First, the 1-homogeneity
requires that L(0, x) = 0 for all x. Since L is continuous on the compact set
{|p| = 1} × U , and L(p, x) > 0 for p 6= 0, we have
γ := inf {L(p, x)} > 0.
|p|=1
x∈U

Since L is 1-homogeneous, we conclude that

L(p, x) ≥ γ|p| for all p ∈ Rn , x ∈ U . (4.5)

31
32 CHAPTER 4. THE HAMILTON-JACOBI-BELLMAN EQUATION

Therefore, for any w ∈ C 1 ([0, 1]; U )

Z 1 Z 1
′
γ ℓ(w) = γ |w (t)| dt ≤ L(w′ (t), w(t)) dt = I[w], (4.6)
0 0

where ℓ(w) denotes the length of w. Hence, curves that minimize, or nearly
minimize I must have bounded length.
Instead of looking for minimizing curves w via the Euler-Lagrange equa-
tions, we consider the value function

u(x) = inf {g(y) + T (x, y)} , (4.7)

y∈∂U

where g : ∂U → R. In the case where U is convex, L(p, x) = |p|, and g ≡ 0,

u is the distance function to the boundary ∂U . We assume throughout this
section that the compatibility condition

g(x) − g(y) ≤ T (x, y) for all x, y ∈ ∂U (4.8)

holds. The compatibility condition ensures that u assumes its boundary values
u = g on ∂U continuously.

Proposition 4.1. For any x, y ∈ U such that the line segment between x and
y belongs to U we have
T (x, y) ≤ K|x − y|, (4.9)
where K = supx∈U ,|p|=1 L(p, x).

Proof. Take w(t) = x + t(y − x). Then

Z 1 Z 1
T (x, y) ≤ L(y − x, x + t(y − x)) dt = |x − y| L(p, x + t(y − x)) dt.
0 0

where p = (y − x)/|y − x| and we used the fact that L is positively 1-

homogeneous. The result immediately follows.

Lemma 4.2. For all x, y, z ∈ U we have

T (x, z) ≤ T (x, y) + T (y, z). (4.10)

Proof. Let ε > 0. For i = 1, 2, let wi ∈ C 1 ([0, 1]; U ) such that w1 (0) = x,
w1 (1) = y, w2 (0) = y, w2 (1) = z and

T (x, y) + T (y, z) + ε ≥ I[w1 ] + I[w2 ].

 33

Define (
w1 (2t), if 0 ≤ t ≤ 21
w(t) =
w2 (2t − 1), if 12 < t ≤ 1.
Note we can reparameterize w, if necessary, so that w ∈ C 1 ([0, 1]; U ), and we
easily compute that I[w1 ] + I[w2 ] = I[w]. Since w(0) = x and w(1) = z we
have
T (x, z) ≤ I[w] ≤ T (x, y) + T (y, z) + ε.
Sending ε → 0 completes the proof.
We now establish the important dynamic programming principle for the
value function u.
Lemma 4.3. For every B(x, r) ⊂ U we have
u(x) = inf {u(y) + T (x, y)} . (4.11)
y∈∂B(x,r)

Proof. Fix x ∈ U with B(x, r) ⊂ U , and let v(x) denote the right hand side
of (4.11).
We first show that u(x) ≥ v(x). Let ε > 0. Then there exists z ∈ ∂U and
w ∈ C 1 ([0, 1]; U ) such that w(0) = x, w(1) = z and
g(z) + I[w] ≤ u(x) + ε. (4.12)
Let y ∈ ∂B(x, r) and s ∈ (0, 1) such that w(s) = y. Define
w1 (t) = w(st) and w2 (t) = w(s + t(1 − s)).
Then we have I[w1 ]+I[w2 ] = I[w]. Furthermore, w1 (0) = x, w1 (1) = w2 (0) =
y and w2 (1) = z. Combining these observations with (4.12) and (4.1) we have
u(x) + ε ≥ g(z) + I[w1 ] + I[w2 ] ≥ u(y) + T (x, y) ≥ v(x).
Since ε > 0 is arbitrary, u(x) ≥ v(x).
To show that u(x) ≤ v(x), note that by Lemma 4.2 we have
g(z) + T (x, z) ≤ g(z) + T (y, z) + T (x, y),
for any y ∈ U and z ∈ ∂U . Therefore
u(x) = inf {g(z) + T (x, z)}
z∈∂U
≤ inf {g(z) + T (y, z)} + T (x, y)
z∈∂U
= u(y) + T (x, y), (4.13)
for any y ∈ U , and the result easily follows.
34 CHAPTER 4. THE HAMILTON-JACOBI-BELLMAN EQUATION

We now establish regularity of the value function u.

Lemma 4.4. The value function u is locally Lipschitz continuous in U and
assumes the boundary values u = g on ∂U , in the sense that for all x ∈ ∂U
lim u(y) = g(x).
y→x
y∈U

Proof. Let x, y ∈ U such that the line segment between x and y is contained
in U . By (4.13) and Proposition 4.1 we have
u(x) ≤ u(y) + T (x, y) ≤ u(y) + K|x − y|.
Therefore u is Lipschitz on any convex subset of U , and hence u is locally
Lipschitz.
To show that u assumes the boundary values g, we need to use the compat-
ibility condition (4.8) and the Lipschitzness of the boundary ∂U . Fix x0 ∈ ∂U .
Up to orthogonal transformation, we may assume that x0 = 0 and
U ∩ B(0, r) = {x ∈ B(0, r) : xn ≥ h(e
x)} ,
x, xn ) ∈ Rn , h : Rn−1 → R is Lipschitz
for r > 0 sufficiently small, where x = (e
continuous, and h(0) = 0. Let x ∈ U ∩ B(0, r) and define
x|).
y = (x1 , . . . , xn−1 , Lip(h)|e
Then |x − y| ≤ C|x| and yn = Lip(h)|e y |. It follows that the line segment
from y to 0 is contained in U , as well as the segment from x to y. In light of
Proposition 4.1 and Lemma 4.2 we have
u(x) = inf {g(z) + T (x, z)}
z∈∂U
≤ g(0) + T (x, 0)
≤ g(0) + T (x, y) + T (y, 0)
≤ g(0) + C|x − y| + C|y|
≤ g(0) + C|x|.
Now let ε > 0 and z ∈ ∂U such that
u(x) + ε ≥ g(z) + T (x, z).
Invoking the compatibility condition (4.8) we have
u(x) + ε ≥ g(0) − T (0, z) + T (x, z)
≥ g(0) − T (0, x) ≥ g(0) − C|x|.
Therefore |u(x) − g(0)| ≤ C|x| for all x ∈ U ∩ B(0, r), and the result immedi-
ately follows.
 35

We can now characterize u as the viscosity solution of a Hamilton-Jacobi

equation. We define

H(p, x) = sup {−p · a − L(a, x)} . (4.14)

|a|=1

Lemma 4.5. H is convex in p and satisfies (3.6).

Proof. Let p, q ∈ Rn , x ∈ U , and λ ∈ (0, 1). Then there exists a ∈ Rn with
|a| = 1 such that

H(λp + (1 − λ)q, x) = −(λp + (1 − λ)q) · a − L(a, x).

We compute

H(λp + (1 − λ)q, x) = λ (−p · a − L(a, x)) + (1 − λ) (−q · a − L(a, x))

≤ λH(p, x) + (1 − λ)H(q, x).

Therefore p 7→ H(p, x) is convex.

Now fix p ∈ Rn and x, y ∈ U . There exists a ∈ Rn with |a| = 1 such that

H(p, x) = −p · a − L(a, x).

Therefore we have

H(p, x) − H(p, y) ≤ −p · a − L(a, x) − (−p · a − L(a, y))

= L(a, y) − L(a, x) ≤ ω(|x − y|),

where ω is the modulus of continuity of L.

Theorem 4.6. The value function u is the unique viscosity solution of the
Hamilton-Jacobi-Bellman equation
)
H(Du, x) = 0 in U
(4.15)
u = g on ∂U.

Proof. We first verify that u is a viscosity subsolution of (4.15). Let x ∈ U

and let φ ∈ C ∞ (Rn ) such that u − φ has a local maximum at x. Choose r > 0
sufficiently small so that

u(y) − u(x) ≤ φ(y) − φ(x) for all y ∈ B(x, r) ⊂ U.

By the dynamic programming principle (4.11)

0= inf {u(y) − u(x) + T (x, y)} ≤ inf {φ(y) − φ(x) + T (x, y)}.
y∈∂B(x,r) y∈∂B(x,r)
36 CHAPTER 4. THE HAMILTON-JACOBI-BELLMAN EQUATION

Since φ ∈ C ∞ (Rn ) there exists C > 0 such that

φ(y) − φ(x) ≤ Dφ(x) · (y − x) + C|x − y|2 for all y ∈ B(x, r).
Therefore we have
0≤ inf {Dφ(x) · (y − x) + T (x, y)} + Cr2 . (4.16)
y∈∂B(x,r)

Notice now that

Z 1

T (x, y) ≤ r L y−x
r
,x + t(y − x) dt
Z0 1


≤r L y−x
r
,x + o(1) dt
0
y−x


= rL r
,x + o(r)
as r → 0+ , due to the continuity of L. Inserting this into (4.16) and dividing
by r we have


0 ≤ inf Dφ(x) · y−x
r
+ L y−x
r
, x + o(1),
y∈∂B(x,r)

as r → 0+ . Setting a = (y − x)/r and sending r → 0+ we find that

H(Dφ(x), x) = − inf {Dφ(x) · a + L(a, x)} ≤ 0.
|a|=1

Therefore u is a viscosity subsolution of (4.15).

We now show that u is a viscosity supersolution of (4.15). Let x ∈ U and
let φ ∈ C ∞ (Rn ) such that u − φ has a local minimum at x. We must show
that
H(Dφ(x), x) = − inf {Dφ(x) · a + L(a, x)} ≥ 0.
|a|=1

Suppose to the contrary that there exists θ > 0 such that

inf {Dφ(x) · a + L(a, x)} ≥ θ.
|a|=1

Then there exists r0 > 0 such that B(x, r0 ) ⊂ U and

θ
Dφ(y) · a + L(a, y) ≥ |a| (4.17)
2
for all a ∈ Rn and y ∈ B(x, r0 ). Notice we used the 1-homogeneity of L above.
Since u − φ has a local minimum at x, we may as well also assume that
u(y) − u(x) ≥ φ(y) − φ(x) for all y ∈ B(x, r0 ).
 37

Let 0 < r < r0 . By the dynamic programming principle (4.11) there exist
y ∈ ∂B(x, r) and w ∈ C 1 ([0, 1]; U ) with w(0) = x and w(1) = y such that
θ
u(x) ≥ u(y) + I[w] − r. (4.18)
4
By (4.6) and Lemma 4.4 we have
θ
γ ℓ(w) ≤ I[w] ≤ u(x) − u(y) + r ≤ Cr.
4
Fix 0 < r < r0 small enough so that ℓ(w) < r0 . Then w(t) ∈ B 0 (x, r0 ) for all
t ∈ [0, 1]. We can now invoke (4.17) to find that
u(y) − u(x) ≥ φ(y) − φ(x)
Z 1
d
= φ(w(t)) dt
0 dt
Z 1
= Dφ(w(t)) · w′ (t) dt
0
Z Z 1
θ 1 ′
by (4.17) ≥ |w (t)| dt − L(w′ (t), w(t)) dt
2 0 0
Z 1
θ
≥ r− L(w′ (t), w(t)) dt.
2 0

Combining this with (4.18) we have

Z 1
θ θ
r ≥ u(y) − u(x) + L(w′ (t), w(t)) dt ≥ r,
4 0 2
which is a contradiction.
Note that φ = 0 is smooth strict subsolution, since
H(Dφ(x), x) = − inf {L(a, x)} ≤ − inf {L(a, y)} = −γ < 0.
|a|=1 |a|=1
y∈U

Therefore H(Dφ(x), x) + γ ≤ 0 in U . Since p 7→ H(p, x) is convex, Corollary

3.2 and the remarks thereafter guarantee that u is the unique viscosity solution
of (4.15).
Remark 4.7. We remark that Theorem 4.6 establishes existence of a viscosity
solution of (4.15) when H is given by (4.14). In Chapter 5 we show how to
establish existence with the method of vanishing viscosity. Other techniques
for establishing existence include the Perron method [6] and convergence of
finite difference approximations.
38 CHAPTER 4. THE HAMILTON-JACOBI-BELLMAN EQUATION

Exercise 4.8. Let 1 < p < ∞ and define

! p1
X
n
|x|p := |xi |p .
i=1

Assume U ⊂ Rn is open, bounded, and path connected with Lipschitz bound-

ary ∂U , and let f : U → R be continuous and positive. Show that there exists
a unique viscosity solution u ∈ C(U ) of the p-eikonal equation
(
|Du|p = f in U
(P)
u = 0 on ∂U.

[Hint: Construct u as the value function

u(x) = inf{T (x, y) : y ∈ ∂U },

where

T (x, y) = inf I[w] : w ∈ C 1 ([0, 1]; U ), w(0) = x, w(1) = y ,
Z 1
I[w] = f (w(t))|w′ (t)|q dt,
0
1 1
and q is the Hölder conjugate of p, i.e., p
+ q
= 1.]
Chapter 5

Convergence of vanishing
viscosity

Consider the viscous Hamilton-Jacobi equation

)
uε + H(Duε , x) − ε∆uε = 0 in U
(5.1)
uε = 0 on ∂U.

In this section we show that solutions uε of (5.1) converge as ε → 0 to the

unique viscosity solution of
)
u + H(Du, x) = 0 in U
(5.2)
u = 0 on ∂U.

The main structural assumptions we place on H are coercivity:

lim inf H(p, x) > 0 uniformly in x ∈ U, (5.3)

|p|→∞

and nonnegativity
−H(0, x) ≥ 0 for all x ∈ U. (5.4)
The reason we call (5.4) nonnegativity is that when H(p, x) = G(p) − f (x) and
G(0) ≥ 0, (5.4) implies that f ≥ 0.
The main structural condition we place on U is the following exterior sphere
condition: There exists r > 0 such that for every x0 ∈ ∂U there is a point
x∗0 ∈ Rn \ U for which
B(x∗0 , r) ∩ U = {x0 }. (5.5)
Throughout this section we assume that U ⊂ Rn is open, bounded, and
satisfies the exterior sphere condition, and H satisfies (3.6) and is continuous,

39
40 CHAPTER 5. CONVERGENCE OF VANISHING VISCOSITY

coercive, and nonnegative. As a result, Corollary 3.2 guarantees that a com-

parison principle holds for (5.2). We note the results in this section hold under
more general assumptions than these, but the proofs are particularly simple
and illustrative in this special case.

5.1 Weak upper and lower limits

Our first technique for proving convergence of the vanishing viscosity method
will be the method of weak upper and lower limits. These techniques are very
general and apply to a wide range of problems.
We first need some basic estimates on solutions of (5.1).
Lemma 5.1. Let ε > 0 and let uε ∈ C 2 (U ) ∩ C(U ) be a solution of (5.1).
Then
0 ≤ uε ≤ sup |H(0, x)| in U. (5.6)
x∈U

Proof. The argument is based on the maximum principle. Due to the com-
pactness of U , uε must attain its maximum value at some x0 ∈ U . If x0 ∈ ∂U
then u(x0 ) = 0. If x0 ∈ U then Duε (x0 ) = 0 and ∆uε (x0 ) ≤ 0. Therefore

uε (x0 ) = ε∆uε (x0 ) − H(Duε (x0 ), x0 ) ≤ sup |H(0, x)|.

x∈U

Likewise, uε attains its minimum value at some y0 ∈ U . If y0 ∈ ∂U then

uε (y0 ) = 0. If y0 ∈ U then Duε (x0 ) = 0 and ∆uε (x0 ) ≥ 0. Since H is
nonnegative (5.4)

uε (y0 ) = ε∆uε (y0 ) − H(0, y0 ) ≥ 0.

Therefore uε ≥ 0 throughout U .
Definition 5.2. Let {uε }ε>0 be a family of real-valued functions on U .
The upper weak limit u : U → R of the family {uε }ε>0 is defined by

u(x) = lim sup uε (y). (5.7)

(y,ε)→(x,0+ )

Similarly, the lower weak limit u : U → R is defined by

u(x) = lim inf uε (y). (5.8)

(y,ε)→(x,0+ )

The limits above are taken with y ∈ U .

5.1. WEAK UPPER AND LOWER LIMITS 41

The upper and lower weak limits are fundamental objects in the theory of
viscosity solutions and allow passage to the limit in a wide variety of applica-
tions.
Lemma 5.3. Suppose the family {uε }ε>0 is uniformly bounded. Then u ∈
USC(U ) and u ∈ LSC(U ).
Proof. By the uniform boundedness assumption, u and u are bounded real-
valued functions on U . We will show that u ∈ USC(U ); the proof that u ∈
LSC(U ) is very similar.
We assume by way of contradiction that xk → x and u(xk ) ≥ u(x) + δ for
some δ > 0 and all k large enough, where xk , x ∈ U . By the definition of u,
for each k there exists yk and εk such that |xk − yk | < 1/k, εk < 1/k and
δ δ
uεk (yk ) ≥ u(xk ) − ≥ u(x) +
2 2
for sufficiently large k. Therefore
lim inf uεk (yk ) > u(x),
k→∞

which is a contradiction to the definition of u, since yk → x and εk → 0+ .

Theorem 5.4. For each ε > 0 let uε ∈ C 2 (U ) ∩ C(U ) solve (5.1). Then
uε → u uniformly on U as ε → 0+ , where u is the unique viscosity solution of
(5.2).
The idea of the proof is to show that u is a viscosity subsolution of (5.2),
and u is a viscosity supersolution of (5.2). Then provided we can show that
u ≤ u on ∂U , the comparison principle will show that u ≡ u, and uniform
convergence follows.
Proof. By Lemma 5.1, the family {uε }ε>0 is uniformly bounded, hence by
Lemma 5.3, u ∈ USC(U ) and u ∈ LSC(U ).
We claim that u is a viscosity solution of u + H(Du, x) ≤ 0 in U . To
establish the claim, let x ∈ U and φ ∈ C ∞ (Rn ) such that u − φ has a local
maximum at x. By replacing φ(y) with φ(y) + |x − y|2 , there is no loss in
assuming that u − φ has a strict local maximum at x. It follows that there
exists xk → x and εk → 0+ such that uεk (xk ) → u(x) and uεk − φ has a
local maximum at xk 1 . Since uε is twice continuously differentiable and φ is
smooth, we have
Duεk (xk ) = Dφ(xk ) and ∆uεk (xk ) ≤ ∆φ(xk ).
1
See Exercise 5.5
42 CHAPTER 5. CONVERGENCE OF VANISHING VISCOSITY

It follows that
u(x) + H(Dφ(x), x) = lim uεk (xk ) + H(Dφ(xk ), xk ) − εk ∆φ(xk )
k→∞
≤ lim uεk (xk ) + H(Duεk (xk ), xk ) − εk ∆uεk (xk ) = 0.
k→∞

Therefore u is a viscosity solution of u + H(Du, x) ≤ 0 in U . A similar

argument shows that u is a viscosity solution of u + H(Du, x) ≥ 0 in U .
By Lemma 5.1, uε ≥ 0 for all ε > 0, thus u ≥ 0 on ∂U . We claim that
u ≤ 0 on ∂U . To see this, fix x0 ∈ ∂U . By coercivity of H (5.3), select C > 0
and δ > 0 so that
H(Cp, x) ≥ δ for all x ∈ U and |p| = 1.
By the exterior sphere condition there exists r > 0 and x∗0 ∈ Rn \ U such that
|x0 − x∗0 | = r and
ψ(x) := C(|x − x∗0 | − r) ≥ 0 for all x ∈ U .
We note that
C(n − 1) C(n − 1)
|Dψ(x)| = C and ∆ψ(x) = ∗
≤
|x − x0 | r
for all x ∈ U . It follows that
ψ(x) + H(Dψ(x), x) − ε∆ψ(x) ≥ δ − C(n − 1)εr−1
for all x ∈ U . Since ψ ≥ 0 on ∂U , we can for sufficiently small ε > 0 use a
maximum principle argument similar to Lemma 5.1 to show that uε ≤ ψ on
U . It follows that
u(x0 ) ≤ lim sup ψ(x) = 0.
x→x0

This establishes the claim.

We have shown that u + H(Du, x) ≤ 0 and u + H(Du, x) ≥ 0 in the
viscosity sense, and u ≤ u on ∂U . By the comparison principle from Corollary
3.2 and the remarks thereafter, u ≤ u on U . Since u ≤ u by definition, we
have u ≡ u. It follows that uε → u uniformly on u as ε → 0+2 .
Exercise 5.5. Suppose u − φ has a strict local maximum at x ∈ U . Show
that there exists xk → x and εk → 0 such that uεk (xk ) → u(x) and uεk − φ
has a local maximum at xk .
Exercise 5.6. Show that if u ≡ u then uε → u uniformly on U .
2
See Exercise 5.6
 √
5.2. THE O( ε) RATE 43
√
5.2 The O( ε) rate
With a bit more work, and a few more hypotheses on H, we can actually say
something about the rate of convergence uε → u. This also gives us a second
approach for proving convergence of the vanishing viscosity method.
The main ingredient is the Lipschitz continuity of the viscosity solution
of (5.2). The following lemma gives a standard argument for proving that
viscosity subsolutions are Lipschitz continuous.

Lemma 5.7. Let u ∈ USC(U ) be a nonnegative viscosity subsolution of (5.2).

Then there exists C depending only on H such that

|u(x) − u(y)| ≤ C|x − y| for all x, y ∈ U .

Proof. By coercivity of H (5.3) select C > 0 and δ > 0 so that

H(Cp, x) ≥ δ for all x ∈ U and |p| = 1. (5.9)

Fix x ∈ U and define

w(y) = u(y) − C|y − x|.
Then w attains its maximum at some y0 ∈ U . If y0 ∈ ∂U then since u ≤ 0 on
∂U we have

u(x) = w(x) ≤ w(y0 ) = u(y0 ) − C|y0 − x| < 0,

which contradicts the nonnegativity of u. If y0 ∈ U and y0 6= x, then the

viscosity subsolution property implies that
 
y0 − x
u(y0 ) + H C , y0 ≤ 0, (5.10)
|y0 − x|

which contradicts (5.9). Therefore w must attain its maximum at y0 = x and

we have
u(y) − C|x − y| = w(y) ≤ w(x) = u(x)
for all y ∈ U . Since x ∈ U was arbitrary, the result follows.

To prove a rate of convergence uε → u, we need to assume Lipschitz

regularity of H. In particular, instead of (3.6), we assume that for every
R > 0 there exists CR such that

H(p, y) − H(p, x) ≤ CR |x − y| for all x, y ∈ U and |p| ≤ R. (5.11)

44 CHAPTER 5. CONVERGENCE OF VANISHING VISCOSITY

Theorem 5.8. For each ε > 0, let uε ∈ C 2 (U ) ∩ C(U ) solve (5.1), and let u
be the unique viscosity solution of (5.2). Then there exists C depending only
on H such that
√
|u − uε | ≤ C ε.
√
Proof. We first show that u − uε ≤ C ε. Define
α
Φ(x, y) = u(x) − uε (y) − |x − y|2 ,
2

where α is to be determined. Let (xα , yα ) ∈ U × U such that

max Φ = Φ(xα , yα ).
U ×U

Since Φ(xα , yα ) ≥ Φ(yα , yα ) we have

α
|xα − yα |2 ≤ u(xα ) − u(yα ) ≤ C|xα − yα |,
2
owing to the Lipschitz estimate from Lemma 5.7. Therefore

C
|xα − yα | ≤ . (5.12)
α
We claim that

u(xα ) − uε (yα ) ≤ C 1
α
+ αε . (5.13)
To see this: If xα ∈ ∂U then

u(xα ) − uε (yα ) ≤ 0,

due to Lemma 5.1 and the boundary condition u = 0 on ∂U . If yα ∈ ∂U then

C
u(xα ) − uε (yα ) ≤ u(xα ) − u(yα ) ≤ C|xα − yα | ≤ .
α
If (xα , yα ) ∈ U × U then x 7→ u(x) − α2 |x − yα |2 has a maximum at xα and
hence
u(xα ) + H(pα , xα ) ≤ 0, (5.14)
where pα = α(xα − yα ). Similarly, y 7→ uε (y) + α2 |xα − y|2 has a minimum at
yα and hence Duε (yα ) = pα and −∆uε (yα ) ≤ αn. Therefore

0 = uε (yα ) + H(pα , yα ) − ε∆uε (yα ) ≤ uε (yα ) + H(pα , yα ) + αnε.

5.3. SEMICONCAVITY AND AN O(ε) ONE-SIDED RATE 45

Subtracting this from (5.14) we have

C
u(xα ) − uε (yα ) ≤ H(pα , yα ) − H(pα , xα ) + αnε ≤ + αnε,
α
due to (5.11), (5.12) and the inequality |pα | = α|xα −yα | ≤ C. This establishes
the claim.
By (5.13) and the definition of Φ


max(u − uε ) ≤ Φ(xα , yα ) ≤ u(xα ) − uε (yα ) ≤ C α1 + αε .
U
√
Selecting α = 1/ ε completes √
the proof.
The proof that uε − u ≤ C ε is similar, and is left to Exercise 5.9.

Exercise 5.9. Complete the proof of Theorem 5.8 by showing that

√
uε − u ≤ C ε.

[Hint: Define the auxilliary function

α
Φ(x, y) = uε (x) − u(y) − |x − y|2 .
2
Then proceed as in the proof of Theorem 5.8. You will need to use the exterior
sphere condition and the barrier function method from the proof of Theorem
5.4 to handle the case when yα ∈ ∂U .]

Exercise 5.10. Show that the solution uε of

|u′ (x)| − εu′′ (x) = 1 for x ∈ (−1, 1)

satisfying u(−1) = u(1) = 0 is

 1 
uε (x) = 1 − |x| − ε e− ε |x| − e− ε .
1

In this case, |u − uε | ≤ Cε, where u(x) = 1 − |x| is the viscosity solution of

|u′ (x)| = 1 on (−1, 1) with u(−1) = u(1) = 0.

5.3 Semiconcavity and an O(ε) one-sided rate

Exercise 5.10 suggests that in some situations we might expect to see the better
rate |u − uε | ≤ Cε. To see when this might hold, let us proceed formally with
46 CHAPTER 5. CONVERGENCE OF VANISHING VISCOSITY

maximum principle arguments. Let x ∈ U be a maximum of uε − u. Then

u − uε has a minimum at x, and the definition of viscosity solution yields

u(x) + H(Duε (x), x) ≥ 0,

since uε ∈ C 2 (U ). Since uε solves (5.1), we find that

uε (x) − u(x) ≤ ε∆uε (x).

In the case that u ∈ C 2 (U ), we have ∆uε (x) ≤ ∆u(x) and thus

max(uε − u) ≤ ε sup ∆u(y) = Cε.

U y∈U

A similar argument shows that

max(u − uε ) ≤ −ε inf ∆u(y) = Cε.

U y∈U

In general, we do not expect u ∈ C 2 (U ). However, there are situations

where the second derivatives of u are bounded from above or from below,
and we can obtain one-sided rates uε − u ≤ Cε or u − uε ≤ Cε. This is
possible, for example, when H(p, x) = G(p) − f (x), where G is convex and f
has bounded second derivatives. To see why, we again proceed formally, and
assume u ∈ Cc∞ (Rn ) is a solution of

u + G(Du) = f in Rn .

Differentiate the PDE above twice in an arbitrary direction ξ ∈ Rn with |ξ| = 1

to obtain
X
n X
n
uξξ + Gpi pj (Du)uxi ξ uxj ξ + Gpi (Du)uxi ξξ = fξξ .
i,j=1 i=1

Since G is convex, the second term above is nonnegative. Setting v = uξξ we

find that
Xn
v+ Gpi (Du)vxi ≤ fξξ in Rn .
i=1

Since v is compactly supported, v attains its maximum over Rn at some x ∈ Rn

and Dv(x) = 0. Therefore

sup uξξ = v(x) ≤ fξξ (x).

y∈Rn
5.3. SEMICONCAVITY AND AN O(ε) ONE-SIDED RATE 47

It follows that D2 u ≤ cI at all points in Rn , where

c := maxn fξξ (x) ≥ 0. (5.15)

x∈R
|ξ|=1

Since u is not generally smooth, these arguments are only a heuristic. The
following theorem makes the arguments rigorous in the viscosity sense.

Theorem 5.11. Assume p 7→ G(p) is convex, G(0) = 0, and f ∈ Cc2 (Rn ). Let
u ∈ C(Rn ) be a compactly supported viscosity solution of

u + G(Du) = f in Rn . (5.16)

Then u is a viscosity solution of

−D2 u ≥ −cI in Rn , (5.17)

where c is given by (5.15).

Remark 5.12. We say u is a viscosity solution of (5.17) provided D2 φ(x) ≤ cI

whenever φ ∈ C ∞ (Rn ) and u − φ has a local minimum at x. This is equivalent
to the condition that u is a viscosity solution of

− max uξξ ≥ −c in Rn .
|ξ|=1

A function u satisfying (5.17) is called semiconcave, with semiconcavity con-

stant c. Notice that v := u − 12 c|x|2 is a viscosity solution of −D2 v ≥ 0, hence
v is concave (due to a generalization of Exercise 2.15).
We also note that (5.17) is equivalent to

u(x + h) − 2u(x) + u(x − h) ≤ c|h|2 for all x, h ∈ Rn ,

which is often the definition of semiconcavity. A function u is called semiconvex

if −u is semiconcave.

Proof. Consider the auxiliary function

α c c
Φ(x, y, z) = u(x) − 2u(y) + u(z) − |x − 2y + z|2 − |x − y|2 − |z − y|2 . (5.18)
2 2 2
Let (xα , yα , zα ) ∈ Rn × Rn × Rn such that

max Φ = Φ(xα , yα , zα ).
Rn ×Rn ×Rn
48 CHAPTER 5. CONVERGENCE OF VANISHING VISCOSITY

By the definition of viscosity solution we have

u(xα ) + G(pα + c(xα − yα )) ≤ f (xα ),

u(zα ) + G(pα + c(zα − yα )) ≤ f (zα ),

and
u(yα ) + G(pα + 2c (xα − yα ) + 2c (zα − yα )) ≥ f (yα ),
where pα = α(xα − 2yα + zα ). Since G is convex

2G(pα + 2c (xα − yα ) + 2c (zα − yα ))

≤ G(pα + c(xα − yα )) + G(pα + c(zα − yα )).

It follows that

u(xα ) − 2u(yα ) + u(zα ) ≤ f (xα ) − 2f (yα ) + f (zα ). (5.19)

Since Φ(y + h, y, y − h) ≤ Φ(xα , yα , zα ) for any y, h ∈ Rn we find that

u(y + h) − 2u(y) + u(y − h) − c|h|2

α
≤ f (xα ) − 2f (yα ) + f (zα ) −|xα − 2yα + xα |2
2
c c
− |xα − yα |2 − |zα − yα |2 . (5.20)
2 2
We now aim to bound the terms on the right hand side. Since Φ(xα , yα , zα ) ≥
Φ(0, 0, 0) = 0, we have
α c c
|xα − 2yα + zα |2 + |xα − yα |2 + |zα − yα |2 ≤ C.
2 2 2
Therefore, by passing to a subsequence, there exists y0 , h0 ∈ Rn such that

yα → y0 , xα − yα → h0 , and yα − zα → h0 ,

as α → ∞. Therefore

lim sup Φ(xα , yα , zα ) ≤ u(y0 + h0 ) − 2u(y0 ) + u(y0 − h0 ) − c|h0 |2 .

α→∞

For each α we have

Φ(xα , yα , zα ) ≥ Φ(y0 +h0 , y0 , y0 −h0 ) = u(y0 +h0 )−2u(y0 )+u(y0 −h0 )−c|h0 |2 ,

and so we deduce

lim Φ(xα , yα , zα ) = u(y0 + h0 ) − 2u(y0 ) + u(y0 − h0 ) − c|h0 |2 .

α→∞
5.3. SEMICONCAVITY AND AN O(ε) ONE-SIDED RATE 49

Therefore
α|xα − 2yα + zα |2 → 0 as α → ∞.
Passing to limits in (5.20) we have

u(y + h) − 2u(y) + u(y − h) − c|h|2

≤ f (y0 + h0 ) − 2f (y0 ) + f (y0 − h0 ) − c|h0 |2
≤ c|h0 |2 − c|h0 |2 ≤ 0,

for all y, h ∈ Rn . Now let φ ∈ C ∞ (Rn ) such that u − φ has a local minimum
at y ∈ Rn . Then

u(y + h) − u(y) ≥ φ(y + h) − φ(y) for small |h|.

Therefore

φ(y + h) − 2φ(y) + φ(y − h) ≤ u(y + h) − 2u(y) + u(y − h) ≤ c|h|2

for small |h|. It follows that φξξ (y) ≤ c for all ξ ∈ Rn with |ξ| = 1, and so
D2 φ(y) ≤ cI.
The second derivative estimate from Theorem 5.11 allows us to prove a
better one-sided rate in the method of vanishing viscosity.

Theorem 5.13. Assume p 7→ G(p) is convex and nonnegative with G(0) = 0,

and f ∈ Cc2 (U ) is nonnegative. Let u ∈ C(U ) be the viscosity solution of
)
u + G(Du) = f in U
(5.21)
u = 0 on ∂U,

and let uε ∈ C 2 (U ) ∩ C(U ) solve

)
uε + G(Duε ) − ε∆uε = f in U
(5.22)
uε = 0 on ∂U,

Then there exists a constant C such that

uε − u ≤ Cε.

Proof. Define (
u(x), if x ∈ U
v(x) =
0, otherwise.
50 CHAPTER 5. CONVERGENCE OF VANISHING VISCOSITY

We claim that v ∈ C(Rn ) is a viscosity solution of

v + G(Dv) = f in Rn .

Since v(x) = f (x) = G(0) = 0 for x 6∈ U , we just need to check that v is

a viscosity solution at boundary points x ∈ ∂U . Furthermore, since G ≥ 0
and u(x) = f (x) = 0 for x ∈ ∂U , the viscosity supersolution property holds
trivially. For the subsolution property, let x ∈ ∂U and φ ∈ C ∞ (Rn ) such that
v − φ has a local maximum at x. Since v is nonnegative (Lemma 5.1) and
v(x) = 0 we have
φ(y) − φ(x) ≥ v(y) − v(x) ≥ 0
for all y near x. Therefore φ has a local minimum at x, and so Dφ(x) = 0. It
follows that
v(x) + G(Dφ(x)) = 0 + G(0) = 0 ≤ f (x).
This establishes the claim.
By Theorem 5.11, v is a viscosity solution of −D2 v ≥ −cI on Rn , where
c is given by (5.15). Since u = v on the open set U , it follows that u is a
viscosity solution of −D2 u ≥ −cI on U . Note that u = uε on ∂U . Suppose
that maxU (uε − u) > 0 and let x ∈ U be a point at which uε − u assumes its
positive maximum. Then u − uε has a minimum at x and hence

u(x) + G(Duε (x)) ≥ f (x),

and D2 uε (x) ≤ cI. It follows that ∆uε (x) ≤ nc and therefore

max(uε − u) = uε (x) − u(x) ≤ ε∆uε (x) ≤ cnε.

U

Exercise 5.14.

(a) Let u ∈ C(U ) be a viscosity solution of

H(Du, u, x) = 0 in U.

Let Ψ : R → R be continuously differentiable with Ψ′ > 0. Show that

v := Ψ ◦ u is a viscosity solution of

H((Φ′ ◦ v)Dv, Φ ◦ v, x) = 0 in U,

where Φ := Ψ−1 .
5.3. SEMICONCAVITY AND AN O(ε) ONE-SIDED RATE 51

(b) Let u ∈ C(U ) be a viscosity solution of

H(Du) = f in U,

and suppose that H is positively 1-homogeneous. Define the Kružkov

Transform of u by v := −e−u . Use part (a) to show that v is a viscosity
solution of
f v + H(Dv) = 0 in U. (5.23)
[Remark: The Kružkov Transform is a standard technique for introduc-
ing a zeroth order term. When f > 0, this term has the correct sign
for a comparison principle to hold for (5.23). This also shows that we
do not lose much in the way of generality by studying equations with
zeroth order terms.]
52 CHAPTER 5. CONVERGENCE OF VANISHING VISCOSITY
Chapter 6

Boundary conditions in the

viscosity sense

We say that u satisfies the boundary condition from (3.14) in the strong sense
provided u = g on ∂U . This is the usual sense, and is how we have been inter-
preting boundary conditions thus far. However, depending on the geometry
of the projected characteristics, the Dirichlet problem (3.14) with boundary
conditions in the strong sense is in general overdetermined. For example, the
solution u of
ux1 + ux2 = 0 in B(0, 1) ⊂ R2
is constant along the projected characteristics

x(s) = (x0 + s, s) (x0 ∈ R).

Since each projected characteristic intersects ∂B(0, 1) at two points, we cannot

specify arbitrary Dirichlet conditions on ∂B(0, 1). The solution u is in fact
uniquely determined by its values on ∂B(0, 1) ∩ {x1 + x2 ≤ 0}. Clearly we
need some weaker notion of boundary conditions if we expect to get existence.

Exercise 6.1. Consider the ordinary differential equation

u′ε (x) − εu′′ε (x) = 1, uε (0) = uε (1) = 0.

Find explicitly the solution uε and sketch its graph. Show that uε (x) → x
pointwise on [0, 1) as ε → 0.

The previous exercise suggests that u(x) = x should be the viscosity solu-
tion of
u′ (x) = 1, u(0) = u(1) = 0,

53
54 CHAPTER 6. BOUNDARY CONDITIONS

even though u(1) 6= 0. The issue is that the problem above is overdetermined,
so we lose one of the boundary conditions in the vanishing viscosity limit. The
same thing happens in a more complicated manner in higher dimensions.
In order to make sense of this, we should consider carefully how boundary
conditions behave in the vanishing viscosity limit. Let uε be a smooth solution
of
H(Duε , uε , x) − ε∆uε = 0 in U, (6.1)
and assume that uε ≤ g on ∂U , where g : ∂U → R is continuous. Exercise
(6.1) shows that we cannot expect uε to converge uniformly on U . Instead, let
us consider the weak upper limit

u(x) = lim sup uε (y),

(y,ε)→(x,0+ )

where we assume that {uε }ε>0 is uniformly bounded, and y ∈ U . Select a

point x ∈ ∂U and let φ ∈ C ∞ (Rn ) such that u − φ has a strict local maximum
at x over the set U . Then there exists εk → 0+ and xk → x such that xk ∈ U ,
uεk (xk ) → u(x) and uεk − φ has a local max at xk over U . By passing to a
further subsequence, if necessary, we may assume that either (A) xk ∈ U for
all k, or (B) xk ∈ ∂U for all k.
If (A) holds, then we conclude, as in Section 1.3, that

H(Dφ(x), u(x), x) ≤ 0.

If (B) holds, then

u(x) = lim uεk (xεk ) = g(x).
k→∞

Hence we find that either u(x) = g(x) or H(Dφ(x), u(x), x) ≤ 0. If we weaken

the boundary condition by assuming merely that uε ≤ g on ∂U , then we would
find that for each x ∈ ∂U , either u(x) ≤ g(x) or H(Dφ(x), u(x), x) ≤ 0. This
can be compactly written as

min {H(Dφ(x), u(x), x), u(x) − g(x)} ≤ 0.

We can make the same argument with the weak lower limit u to find that
when u − φ has a local minimum at x ∈ ∂U we have

max {H(Dφ(x), u(x), x), u(x) − g(x)} ≥ 0,

provided uε ≥ g on ∂U .
This motivates the following definitions.
 55

Definition 6.2. We say u ∈ USC(U ) is a viscosity subsolution of (3.14) if for

all x ∈ U and every φ ∈ C ∞ (Rn ) such that u − φ has a local maximum at x
with respect to U
(
H(Dφ(x), u(x), x) ≤ 0, if x ∈ U
min {H(Dφ(x), u(x), x), u(x) − g(x)} ≤ 0 if x ∈ ∂U.

Likewise, we say that u ∈ LSC(U ) is a viscosity supersolution of (3.14) if

for all x ∈ U and every φ ∈ C ∞ (Rn ) such that u − φ has a local minimum at
x with respect to U
(
H(Dφ(x), u(x), x) ≥ 0, if x ∈ U
max {H(Dφ(x), u(x), x), u(x) − g(x)} ≥ 0 if x ∈ ∂U.

Finally, we say that u is a viscosity solution of (3.14) if u is both a viscosity

sub- and supersolution. In this case, we say that the boundary conditions in
(3.14) hold in the viscosity sense

Exercise 6.3. Show that u(x) = x is a viscosity solution of

u′ (x) = 1, u(0) = u(1) = 0,

on the interval U = (0, 1) in the sense of Definition 6.2.

It is possible to prove a comparison principle for viscosity sub- and super-

solutions in the sense of Definition 6.2, provided the semicontinuous solutions
attain their boundary values continuously. We will not give the proof in full
generality (see [1]). In some special cases, it is also possible to recover strong
boundary conditions from boundary conditions in the viscosity sense. The
typical approach is to select a sequence of test functions at a boundary point
for which the sub- or supersolution property is violated. This is often possible
because the class of admissible test functions at boundary points is very large
(since the admissibility condition is “one-sided”).
With the exception of Theorem 9.8, we will generally not use Definition 6.2
in these notes. Hence, unless otherwise stated, all viscosity solutions should
be interpreted in the sense of the definitions in Chapter 2.
We give here a comparison principle in the special case where we have addi-
tional information concerning at which boundary points the Dirichlet condition
holds, and at which points the PDE should hold. We also relax the assumption
that U is bounded, and instead assume that the sub- and supersolutions are
bounded.
56 CHAPTER 6. BOUNDARY CONDITIONS

We assume the usual monotonicity (3.3) and regularity (3.6) conditions on

H hold. In addition we assume

|H(p, z, x) − H(q, z, x)| ≤ ω1 (|p − q|), (6.2)

where ω1 is a modulus of continuity. In this case, we can prove the following

comparison principle.

Theorem 6.4. Let U ⊂ Rn be open and suppose ∂U = Γ1 ∪ Γ2 where Γ1

is relatively open and Γ1 ∩ Γ2 = ∅. Let u ∈ USC(U ) be a bounded viscosity
solution of H ≤ −ε < 0 on U ∪ Γ1 , and let v ∈ LSC(U ) be a bounded viscosity
solution of H ≥ 0 on U ∪ Γ1 . If u ≤ v on Γ2 then u ≤ v on U .

The proof is very similar to that of Theorem 3.1, so we will briefly outline
the details. The main difficulty is to ensure that the auxiliary function assumes
its maximum on the unbounded domain U . We also remark that the theorem
holds when U = Rn and Γ1 = Γ2 = ∅.

Proof. Let λ > 0 and define

λ
uλ (x) = u(x) − log(1 + |x|2 ).
2
Let x ∈ U ∪ Γ1 and let φ ∈ C ∞ (Rn ) such that uλ − φ has a local maximum at
x. Then u − λ2 log(1 + |x|2 ) − φ has a local max at x, and therefore

H(λ(1 + |x|2 )−1 x + Dφ(x), u(x), x) ≤ −ε.

By (6.2) and (3.3) we have

H(Dφ(x), uλ (x), x) ≤ H(Dφ(x), u(x), x)

≤ H(λ(1 + |x|2 )−1 x + Dφ(x), u(x), x) + ω1 (λ)
≤ −ε + ω1 (λ).

Therefore, there exists Λ > 0 such that for all 0 < λ < Λ, uλ is a viscosity
solution of
ε
H(Duλ , uλ , x) ≤ − in U ∪ Γ1 . (6.3)
2
We will prove that uλ ≤ v on U for all 0 < λ < Λ. To see this, assume to
the contrary that supU (uλ − v) > 0. For α > 0 define the auxiliary function
α
Φ(x, y) = uλ (x) − v(y) − |x − y|2 . (6.4)
2
 57

Since u and v are bounded

α λ
0≤ |x − y|2 ≤ uλ (x) − v(y) ≤ C − log(1 + |x|2 ),
2 2
for any (x, y) ∈ U × U such that Φ(x, y) ≥ 0. Since supU ×U Φ > 0, we find
that Φ attains its maximum at some (xα , yα ) ∈ U × U satisfying
λ α
log(1 + |xα |2 ) + |xα − yα |2 ≤ C.
2 2
It follows that there exists x0 ∈ U such that, up to a subsequence, xα → x0
and yα → x0 . Therefore

lim sup Φ(xα , yα ) ≤ uλ (x0 ) − v(x0 ) ≤ lim inf Φ(xα , yα ),

α→∞ α→∞

due to the upper semicontinuity of Φ. Therefore

lim Φ(xα , yα ) = uλ (x0 ) − v(x0 ) and lim α|xα − yα |2 = 0. (6.5)

α→∞ α→∞

Since Φ(xα , yα ) ≥ supU (uλ − v) > 0, we have

uλ (x0 ) > v(x0 ) and uλ (xα ) > v(yα ).

Since uλ ≤ v on Γ2 , we must have x0 ∈ U ∪ Γ1 . Since Γ1 is relatively open,

xα , yα ∈ U ∪ Γ1 for sufficiently large α.
By the viscosity sub- and supersolution properties, (3.3) and (3.6) we have
ε
≤ H(pα , v(yα ), yα ) − H(pα , uλ (xα ), xα )
2
≤ H(pα , uλ (xα ), yα ) − H(pα , uλ (xα ), xα )
≤ ω((1 + |pα |)|xα − yα |),

where pα = α(xα − yα ). By (6.5),

(1 + |pα |)|xα − yα | → 0 as α → ∞.

Sending α → ∞ contradicts the positivity of ε.

Therefore uλ ≤ v for all 0 < λ < Λ. It follows that u ≤ v on U .
Exercise 6.5. Let Γ ⊂ Rn be closed and bounded. Consider eikonal equation
(
|Du| = 1 in Rn \ Γ
(H)
u = 0 on Γ.
58 CHAPTER 6. BOUNDARY CONDITIONS

(a) Show that there is at most one viscosity solution u ∈ C(Rn ) of (H) satis-
fying the boundary condition at infinity
lim u(x) = ∞. (6.6)
|x|→∞

[Hint: Theorem 6.4 does not apply, since u and v are unbounded. To prove
uniqueness, let u, v ∈ C(Rn ) be two viscosity solutions of (H) satisfying
(6.6). Let Ψ : R → R be a smooth function satisfying


Ψ(s) = s, if s ≤ 1,
Ψ(s) ≤ 2, for all s ∈ R,

 ′
0 < Ψ (s) ≤ 1, for all s ∈ R.
For R > 1 define
w(x) := (R − 1) Ψ(R−1 u(x)).
Show that w ≤ 2R is a viscosity solution of
1
|Dw| + ≤ 1 in Rn \ Γ.
R
Use the doubling of the variables argument to show that w ≤ v on Rn \ Γ.
Complete the argument from here.]
(b) Show that the solution is not unique without (6.6).

6.1 Time-dependent Hamilton-Jacobi equations

As an application of Theorem 6.4, we will prove a comparison principle for the
time-dependent Hamilton-Jacobi equation
)
ut + H(Du, x) = 0 in Rn × (0, T )
(6.7)
u = g on Rn × {t = 0}.

We assume, as before, that H is continuous and satisfies (3.6) and (6.2).

Theorem 6.6. Let u ∈ USC(Rn × [0, T ]) be a bounded viscosity subsolution
of (6.7), and let v ∈ LSC(Rn × [0, T ]) be a bounded viscosity supersolution of
(6.7). Then u ≤ v on Rn × [0, T ].
Perhaps there is a bit of abuse of notation here. When we say u is a
subsolution of (6.7), we mean that u is a solution of ut + H ≤ 0 in Rn × (0, T )
and u ≤ g at t = 0. Likewise, a supersolution is assumed to satisfy v ≥ g at
t = 0, hence u ≤ v at t = 0.
6.1. TIME-DEPENDENT HAMILTON-JACOBI EQUATIONS 59

Proof. Set U = Rn × (0, T ), Γ1 = Rn × {t = T }, and Γ2 = Rn × {t = 0}. Our

aim is to apply Theorem 6.4. For this, we only need to show that u and v are
viscosity sub- and supersolutions on the extended set

U ∪ Γ1 = Rn × (0, T ].

That is, we need to allow t = T in the sub- and supersolution properties.

Let (x0 , T ) ∈ Γ1 , and let φ ∈ C ∞ (Rn × R) such that u − φ has a local
maximum at (x0 , T ). As before, we may assume the local maximum is strict.
For x ∈ Rn and 0 < t < T , define
ε
φε (x, t) := φ(x, t) + .
T −t
Then there exist sequences εk → 0+ and (xk , tk ) → (x0 , T ) such that 0 < tk <
T and u − φεk has a local maximum at (xk , tk ). Therefore

φεt k (xk , tk ) + H(Dφεk (xk , tk ), xk ) ≤ 0,

and hence
εk
φt (xk , tk ) + + H(Dφ(xk , tk ), xk ) ≤ 0.
(T − tk )2
Letting k → ∞ we find that

φt (x0 , T ) + H(Dφ(x0 , T ), x0 ) ≤ 0.

We can similarly verify that v is a viscosity supersolution of ut + H = 0 on

Γ1 . We can therefore invoke Theorem 6.4, Corollary 3.2, and the remarks
thereafter to obtain that u ≤ v on Rn × [0, T ].
We can also easily prove continuous dependence on the initial data.

Corollary 6.7. Let u, v ∈ C(Rn × [0, T ]) be bounded, and assume that w := u

and w := v are viscosity solutions of

wt + H(Dw, x) = 0 in Rn × (0, T ).

Then
sup |u − v| ≤ sup |u(x, 0) − v(x, 0)|.
Rn ×[0,T ] x∈Rn

Proof. Let C := supx∈Rn |u(x, 0) − v(x, 0)|. Then u − C ≤ v at t = 0, and by

Theorem 6.6 we have u − v ≤ C on Rn × [0, T ]. The inequality v − u ≤ C
follows by swapping the roles of u and v.
60 CHAPTER 6. BOUNDARY CONDITIONS

6.2 The Hopf-Lax Formula

In the case that H = H(p) and H is convex and superlinear, i.e.,

H(p)
lim = ∞,
|p|→∞ |p|

we have the Hopf-Lax formula

   
x−y
u(x, t) = minn tL + g(y) ,
y∈R t

where
L(v) = sup {p · v − H(p)}
p∈Rn

is the Legendre transform of H. In this case, H and L are Legendre duals,

and we also have
H(p) = sup {p · v − L(v)}.
v∈Rn

Under the assumption that g is Lipschitz continuous and bounded, we showed

last semester that u is a Lipschitz continuous almost everywhere solution of
(6.7). It turns out that the Hopf-Lax formula gives the unique viscosity so-
lution of (6.7). We’ll sketch the argument here, for a complete proof see [11,
Chapter 10].
Let (x0 , t0 ) ∈ Rn × (0, T ) and φ ∈ C ∞ (Rn × R) such that u − φ has a local
maximum at (x0 , t0 ). We may assume that for some r > 0

u(x0 , t0 ) = φ(x0 , t0 ) and u(x, t) ≤ φ(x, t) for |x − x0 |2 + |t − t0 |2 < r2 .

Since g is bounded, u is bounded, and we can also assume that u ≤ φ on

Rn × (0, T ). To see why, we note first that we can multiply φ by a bump
function to ensure φ ∈ Cc∞ (Rn × R). Then we can set

C := sup u + sup |φ|,

Rn ×[0,T ] Rn ×R

and
C
φ(x, t) := φ(x, t) + (|x − x0 |2 + |t − t0 |2 ).
r2
Then φ(x0 , t0 ) = u(x0 , t0 ), φ ≥ u for (x, t) ∈ B((x0 , t0 ), r), and

φ(x, t) ≥ φ(x, t) + C ≥ sup u ≥ u,

Rn ×[0,T ]
6.2. THE HOPF-LAX FORMULA 61

for (x, t) 6∈ B((x0 , t0 ), r). Since φt (x0 , t0 ) = φt (x0 , t0 ) and Dφ(x0 , t0 ) =

Dφ(x0 , t0 ), we can without loss of generality replace φ by φ.
We recall that the function u defined by the Hopf-Lax formula satisfies the
property    
x0 − x
u(x0 , t0 ) = minn (t − t0 )L + u(x, t)
x∈R t0 − t
for all 0 < t < t0 . Since u(x0 , t0 ) = φ(x0 , t0 ) and u ≤ φ we have
   
x0 − x
φ(x0 , t0 ) ≤ minn (t − t0 )L + φ(x, t) ,
x∈R t0 − t

for all 0 < t < t0 . The reader should notice the similarity with the proof of
Theorem 4.6. Since φ is smooth, the same arguments that showed that u is a
Lipschitz almost everywhere solution (see [11, Section 3.3]) prove that

φt + H(Dφ) ≤ 0 at (x0 , t0 ).

The proof that u is a viscosity supersolution is similar.

62 CHAPTER 6. BOUNDARY CONDITIONS
Chapter 7

The Perron method

We define the upper semicontinuous envelope of a function u : O → R by

u∗ (x) := lim sup u(y).

O∋y→x

The function u∗ is the smallest upper semicontinuous function that is pointwise

greater than or equal to u. The lower semicontinuous envelope of u, defined by
u∗ := −(−u)∗ , is correspondingly the greatest lower semicontinuous function
that is less than u. Note that u∗ ≤ u ≤ u∗ , and u∗ = u∗ = u if and only if u
is continuous.
The Perron method is a powerful technique for proving existence of vis-
cosity solutions. The idea is to construct a solution as an upper envelope of
subsolutions. Consider the second order nonlinear equation

H(D2 u, Du, u, x) = 0 in U, (7.1)

where H is continuous and U ⊂ Rn is open. Let w ∈ LSC(U ) be a viscosity

supersolution of (7.1) and define
n o
F := v ∈ USC(U ) : v is a subsolution of (7.1) and v ≤ w in U ,

and
u(x) := sup{v(x) : v ∈ F}. (7.2)
The function u is presumably a prime candidate for a viscosity solution of
(7.1).
We now establish two lemmas that are fundamental to the Perron method.
Lemma 7.1. Suppose F is nonempty. Then the upper semicontinuous func-
tion u∗ is a viscosity subsolution of (7.1)

63
64 CHAPTER 7. THE PERRON METHOD

Proof. Let x0 ∈ U and φ ∈ C ∞ (Rn ) such that u∗ − φ has a local maximum at

x0 . Replacing φ by φ + u∗ (x0 ) − φ(x0 ) + |x − x0 |4 , there exists B(x0 , r) ⊂ U
such that

u∗ (x0 ) = φ(x0 ) and u∗ (x) − φ(x) ≤ −|x − x0 |4 for x ∈ B(x0 , r). (7.3)

By definition of u∗ there exists a sequence xk → x0 such that u(xk ) → u∗ (x0 ).

For each k there exists uk ∈ F such that uk (xk ) ≥ u(xk ) − k1 . Since uk ∈
USC(U ), uk − φ attains its maximum over B(x0 , r) at some yk ∈ B(x0 , r).
Furthermore, since uk (yk ) ≤ u(yk ) ≤ u∗ (yk ) we have by (7.3) that

|yk − x0 |4 ≤ φ(yk ) − u∗ (yk )

≤ φ(yk ) − uk (yk )
≤ φ(xk ) − uk (xk )
1
≤ φ(xk ) − u(xk ) + .
k
It follows that yk → x0 as k → ∞. Thus for large enough k, uk − φ has a
local maximum at an interior point yk of the ball B(x0 , r). By the viscosity
subsolution property

H(D2 φ(yk ), Dφ(yk ), uk (yk ), yk ) ≤ 0.

Since u(xk ) → u∗ (x0 ) and

1
uk (yk ) − φ(yk ) ≥ uk (xk ) − φ(xk ) ≥ u(xk ) − φ(xk ) − ,
k
we have lim inf k→∞ uk (yk ) ≥ u∗ (x0 ). Since uk (yk ) ≤ u∗ (yk ) and u∗ is upper
semicontinuous, we find that uk (yk ) → u∗ (x0 ) as k → ∞. Sending k → ∞ and
using the continuity of H we have

H(D2 φ(x0 ), Dφ(x0 ), u∗ (x0 ), x0 ) ≤ 0.

Lemma 7.2. Let u ∈ F . If u∗ is not a viscosity supersolution of (7.1), then

there exists v ∈ F such that v(x) > u(x) for some x ∈ U .
Proof. Let u ∈ F and assume u∗ is not a viscosity supersolution of (7.1). Then
there exists φ ∈ C ∞ (Rn ) such that u∗ − φ has a local minimum at x0 and

H(D2 φ(x0 ), Dφ(x0 ), u∗ (x0 ), x0 ) < 0. (7.4)

We may assume that φ(x0 ) = u∗ (x0 ). If φ(x0 ) = w(x0 ) then w − φ has a local
minimum at x0 , which contradicts (7.4), as w is a supersolution. Therefore
 65

φ(x0 ) < w(x0 ). Hence there exists ε > 0 and a ball B(x0 , r) ⊂ U such that
φ ≤ u∗ and φ + ε ≤ w on B(x0 , r) and
H(D2 φ(x), Dφ(x), φ(x), x) + ε ≤ 0 for x ∈ B(x0 , r). (7.5)
Set  
r4
ψ(x) := φ(x) + δ − |x − x0 | ,
4
24
and choose δ > 0 small enough so that ψ ≤ w on B(x0 , r) and
H(D2 ψ(x), Dψ(x), ψ(x), x) ≤ 0 for x ∈ B(x0 , r).
Define (
max{u(x), ψ(x)}, if x ∈ B(x0 , r)
v(x) :=
u(x), otherwise.
Since u and ψ are subsolutions of H = 0 in B(x0 , r), v is a subsolution in
B(x0 , r). Furthermore, since
ψ(x) ≤ φ(x) ≤ u(x) for x ∈ B(x0 , r) \ B(x0 , 2r ),
we have u = v on the annulus B(x0 , r) \ B(x0 , 2r ). Therefore v is a subsolution
of (7.1) and v ≤ w on U . Therefore v ∈ F .
By definition of the lower semicontinuous envelope u∗ , there exists a se-
quence xk → x0 such that u(xk ) → u∗ (x0 ). Since v ≥ ψ on B(x0 , r), we
have
lim inf v(xk ) ≥ lim ψ(xk ) = u∗ (x0 ) + δ ′ ,
k→∞ k→∞
′ 4 4
where δ = δr /2 . Therefore, for k large enough
δ′
v(xk ) ≥ u(xk ) + ,
2
or v(xk ) > u(xk ). This completes the proof.
The remaining ingredient for Perron’s method is a comparison principle for
(7.1). Let us illustrate the technique on the time-dependent Hamilton-Jacobi
equation (6.7). As usual, we assume H is continuous and satisfies (3.3), (3.6),
and (6.2).
Theorem 7.3. Let g : Rn → R be bounded and Lipschitz continuous, and
suppose that
n o
K := sup |H(p, x)| : |p| ≤ Lip(g) and x ∈ Rn < ∞.

Then for every T > 0 there exists a unique bounded viscosity solution u ∈
C(Rn × [0, T ]) of (6.7).
66 CHAPTER 7. THE PERRON METHOD

Proof. Define
w(x, t) := g(x) + Kt.
If φ ∈ C ∞ (Rn × R) and w − φ has a local minimum at (x0 , t0 ) ∈ Rn × (0, T ),
then |Dφ(x0 , t0 )| ≤ Lip(g) and φt (x0 , t0 ) = K. Therefore

φt (x0 , t0 ) + H(Dφ(x0 , t0 ), x0 ) ≥ K − K = 0.

Therefore w is a bounded supersolution of (6.7).

Define
n o
F := v ∈ USC(Rn × [0, T ]) : v is a subsolution of (6.7) and v ≤ w ,

and
u(x, t) := sup{v(x, t) : v ∈ F}.
We can verify, as before, that w(x,e t) := g(x) − Kt is a subsolution of (6.7).
Therefore F is nonempty. Since u ≤ w and w is continuous, u∗ ≤ w and so
u∗ (x, 0) ≤ w(x, 0) = g(x). By Lemma 7.1, u∗ is a viscosity subsolution of
(6.7). Therefore u∗ ∈ F , and so u = u∗ .
Since we ≤ w, w e ∈ F and hence u ≥ w. e Since w e is continuous, u∗ (x, 0) ≥
w(x, 0) = g(x). By Lemma 7.2, u∗ is a viscosity supersolution of (6.7). Since
u∗ (x, 0) = u∗ (x, 0), we can use the comparison principle (Theorem 6.6) to show
that u∗ ≤ u∗ on Rn × [0, T ]. Since the opposite inequality is true by definition,
we have u∗ = u∗ = u. Therefore u ∈ C(Rn × [0, T ]) is a bounded viscosity
solution of (6.7). Uniqueness follows from Theorem 6.6.

Exercise 7.4. Consider the Hamilton-Jacobi equation

u + H(Du, x) = 0 in Rn .

What (non-trivial) conditions can you place on H to guarantee the existence

of a bounded viscosity solution u ∈ C(Rn )? [Hint: Use the Perron method.]
Chapter 8

Smoothing viscosity solutions

Since viscosity solutions are in general only continuous functions, it is useful

to be able to construct smoother approximations of viscosity solutions. That
is, given a viscosity solution u of F (D2 u, Du, u, x) = 0, we would like to
construct a sequence of smooth (or just smoother) functions uk such that
F (D2 uk , Duk , uk , x) → F (D2 u, Du, u, x) and uk → u as k → ∞, in some
appropriate sense. For linear constant coefficient PDEs, we can construct
smooth approximate solutions by mollification, that is, uε = ηε ∗ u, where ηε is
the standard mollifier [11]. The mollified function uε is infinitely differentiable
and uε → u locally uniformly as ε → 0 provided u is continuous. Since
mollification is a linear operation, it commutes with linear constant coefficient
PDEs and so uε is also a solution of the linear equation.
Unfortunately mollification is not useful for viscosity solutions, as the fol-
lowing exercise illustrates.

Exercise 8.1. Recall from Exercise 2.13 that u(x) = 1 − |x| is a viscosity
solution of |u′ (x)| − 1 = 0. In fact, this is the unique solution with boundary
conditions u(−1) = 0 = u(1) on the interval (−1, 1) (why?). Show that there
does not exist a sequence uk ∈ C 1 ([−1, 1]) such that uk → u and |u′k | → 1
uniformly as k → ∞. This shows that it is impossible, in general, to uniformly
approximate a viscosity solution by a classical solution.

Since we cannot smoothly approximate viscosity solutions, we are left to

consider approximations that are smoother than the continuous or Lipschitz
viscosity solutions, but less regular than classical solutions. Such approxima-
tions are provided by the inf- and sup-convolutions, defined below.

Definition 8.2. Let U ⊂ Rn , u : U → Rn , and ε > 0. We define the sup-

67
68 CHAPTER 8. SMOOTHING VISCOSITY SOLUTIONS
8 8

6 6

4 4

2 2

0 0

-2 -2

-4 -4

-6 -6
-1 -0.5 0 0.5 1 -1 -0.5 0 0.5 1

Figure 8.1: Examples of inf- and sup-convolutions of a Brownian motion sam-

ple path for ε = 0.1 (left) and ε = 0.01 (right)

convolution of u, denoted uε , to be
 
1
u (x) = sup u(y) − |x − y| .
ε 2
(8.1)
y∈U 2ε

Similarly, the inf-convolution of u, denoted uε , is defined by

 
1
uε (x) = inf u(y) + |x − y| . 2
(8.2)
y∈U 2ε

We remark that whenever the set U is not specified, it is taken to be the

domain of u.

The inf- and sup-convolutions are tools that originally appeared in con-
vex analysis—the inf-convolution is called the Moreau envelop in optimiza-
tion [3]—and have been appropriated in the viscosity solution literature due
to their useful approximation properties. As we show below, the inf- and
sup-convolutions of a viscosity solution u are nearly C 2 functions, and are
approximate viscosity super- and subsolutions, respectively.
We first establish some basic properties of inf- and sup-convolutions

Proposition 8.3. Suppose u : U → R is open and let u : U → R. Then

(i) we have uε ≤ u ≤ uε ,

(ii) the function uε + 1

2ε
|x|2 is convex, and uε − 1
2ε
|x|2 is concave,

(iii) if y ε ∈ arg maxy∈U u(y) − 1
2ε
|x − y|2 then |x − y ε |2 ≤ 4kukL∞ (U ) ε,
 69

(iv) if yε ∈ arg miny∈U u(y) + 1
2ε
|x − y|2 then |x − yε |2 ≤ 4kukL∞ (U ) ε,
(v) both uε and uε are twice differentiable almost everywhere in U and
1  1/2

|uε (x) − uε (y)|, |uε (x) − uε (y)| ≤ |x − y| + 4kukL∞ (U ) ε1/2 |x − y|.
2ε
(8.3)
Proof. (i) is obvious. For (ii) Note that
 
1 2 1 1 2
u (x) + |x| = sup u(y) + x · y − |y| ,
ε
2ε y∈U ε 2ε

and recall that the supremum of a family of affine functions is convex. The
proof that uε − 2ε1
|x|2 is concave is similar.
For (iii), since
1
uε (x) = u(y ε ) − |x − y ε |2
2ε
and u (x) ≥ u(x) we have
ε

1
|x − y ε |2 = u(y ε ) − uε (x) ≤ u(y ε ) − u(x) ≤ 2kukL∞ (U ) .
2ε
The proof of (iv) is similar.
For (v), by the Alexandrov Theorem any convex function is twice differ-
entiable almost everywhere. Thus, it follows from (ii) that uε + 2ε 1
|x|2 and
uε − 2ε |x| are twice differentiable almost everywhere in U , and hence so are
1 2

uε and uε .
To prove the Lipschitz estimate (8.3), let x, y ∈ U and δ > 0. Let y ε ∈ U
such that
1
uε (x) ≤ u(y ε ) − |x − y ε |2 + δ.
2ε
Then we have
|x − y ε |2 ≤ 2(2kukL∞ (U ) + δ)ε. (8.4)
Since uε (y) ≥ u(y ε ) − 1
2ε
|y − y ε |2 we have
1

uε (x) − uε (y) ≤ |y − y ε |2 − |x − y ε |2 + δ
2ε
1

≤ (|x − y| + |x − y ε |)2 − |x − y ε |2 + δ
2ε
1

= |x − y|2 + 2|x − y||x − y ε | + δ
2ε
1
= (|x − y| + 2|x − y ε |) |x − y| + δ.
2ε
70 CHAPTER 8. SMOOTHING VISCOSITY SOLUTIONS

Recalling (8.4) and sending δ → 0 we have

1
uε (x) − uε (y) ≤ (|x − y| + 4kuk∞ ε )|x − y|.
1/2 1/2
2ε
Reversing the roles of x and y completes the proof. The Lipschitz estimate for
uε is similar.
Remark 8.4. Recalling Remark 5.12, it follows from Proposition 8.3(i) that
uε is semiconvex and uε is semiconcave. In particular, −D2 uε ≥ − 1ε I and
−D2 uε ≤ 1ε I in the viscosity sense. Roughly speaking, this means the second
derivatives of uε are bounded above by 1ε , while the second derivatives of uε
are bounded below by − 1ε .
To establish further properties of the sup- and inf-convolutions, we need
to assume u has additional regularity.
Lemma 8.5. Suppose U ⊂ Rn is open and bounded, and let u ∈ C(U ). Then
uε , uε → u uniformly on U . Furthermore, if u ∈ C 0,α (U ) for 0 < α ≤ 1 then

(i) |x − y ε |2−α ≤ 2ε[u]0,α;U for any y ε ∈ arg maxy∈U u(y) − 2ε
1
|x − y|2 ,

(ii) |x − yε |2−α ≤ 2ε[u]0,α;U for any yε ∈ arg miny∈U u(y) + 2ε1
|x − y|2 ,

(iii) kuε − ukL∞ (U ) , kuε − ukL∞ (U ) ≤ (2[u]0,α;U )2/(2−α) εα/(2−α) , and

(iv) uε , uε ∈ C 0,α (U ) and [uε ]0,α;U , [uε ]0,α;U ≤ C[u]0,α;U , with C independent
of ε > 0.
Proof. We first prove uniform convergence. Let
 
1
y ∈ arg max u(y) − |x − y|
ε 2

y∈U 2ε

and note that

1
|x − y ε |2 = u(y ε ) − uε (x) ≤ u(y ε ) − u(x). (8.5)
2ε
Therefore

|uε (x) − u(x)| = u(y ε ) − u(x) − 1

2ε
|x − y ε |2 ≤ 2|u(y ε ) − u(x)|. (8.6)
1/2
By Proposition 8.3(i), |x − y ε | ≤ 2kukLε (U ) ε1/2 , and so it follows from uniform
continuity of u that uε → u uniformly on U as ε → 0. The proof for uε is
similar.
 71

For (i), we can use (8.5) to deduce

1
|x − y ε |2 ≤ [u]0,α;U |x − y ε |α ,
2ε
and so
|x − y ε |2−α ≤ 2ε[u]0,α;U .
The proof of (ii) is similar.
For (iii), we use (8.6) and (i) to obtain
|uε (x) − u(x)| ≤ 2[u]0,α;U |x − y ε |α ≤ 2[u]0,α;U (2ε[u]0,α;U )α/(2−α) .

Finally, for (iv) let x, y ∈ U and let xε such that

1
uε (x) = u(xε ) − |x − xε |2 .
2ε
Using (iii) we obtain
|uε (x) − uε (y)| ≤ |uε (x) − u(x)| + |u(x) − u(y)| + |uε (y) − u(y)|
≤ 2(2[u]0,α;U )2/(2−α) εα/(2−α) + [u]0,α;U |x − y|α .
If |x − y|2−α ≥ ε[u]0,α;U then
|uε (x) − uε (y)| ≤ C[u]0,α;U |x − y|α .
On the other hand, if |x − y|2−α < ε[u]0,α;U then
1

uε (x) − uε (y) ≤ |y − xε |2 − |x − xε |2
2ε
1

≤ |x − y|2 + 2|y − x||x − xε |
2ε
1
= (|x − y|2−α + 2|x − y|1−α |x − xε |)|x − y|α
2ε
≤ C[u]0,α;U |x − y|α ,
which completes the proof.
Having established that the sup- and inf-convolutions are smooth(er) ap-
proximations of u, we turn to the problem of showing that sup- and inf-
convolutions preserve the viscosity sub- and super-solution properties, respec-
tively. For this, we introduce additional notation. Given u ∈ USC(U ) we
define
( )

M ε (u) = x ∈ U : arg max u(y) − 2ε 1
|x − y|2 ∩ U 6= ∅ , (8.7)
y∈U
72 CHAPTER 8. SMOOTHING VISCOSITY SOLUTIONS

and for u ∈ LSC(U ) we define

( )

Mε (u) = x ∈ U : arg min u(y) + 1
2ε
|x − y|2 ∩ U 6= ∅ . (8.8)
y∈U

For the moment, we consider equations of the form

F (D2 u, Du) = 0 in U. (8.9)

The following proposition is useful to state independently.

Proposition 8.6. Let ε > 0, u ∈ USC(Rn ), and x0 ∈ Rn . Let φ ∈ C ∞ (Rn )
such that uε − φ has a local max at x0 , and let xε ∈ Rn such that
1
uε (x0 ) = u(xε ) − |x0 − xε |2 . (8.10)
2ε
Define ψ(x) = φ(x + x0 − xε ). Then u − ψ has a local max at xε and
1
Dψ(xε ) = Dφ(x0 ) = (xε − x0 ). (8.11)
ε
Proof. Let r > 0 such that

uε (x0 ) − φ(x0 ) ≥ uε (x) − φ(x) (8.12)

whenever |x − x0 | < r. It follows from (8.10) and (8.12) that

1
u(xε ) − |x0 − xε |2 − φ(x0 ) ≥ uε (x) − φ(x)
2ε
1
≥ u(y) − |x − y|2 − φ(x), (8.13)
2ε
for any y ∈ Rn and |x − x0 | < r. Set y = xε to obtain
1 1
|x − xε |2 + φ(x) ≥ |x0 − xε |2 + φ(x0 )
2ε 2ε
for |x − x0 | < r. Therefore
1
x 7→ |x − xε |2 + φ(x)
2ε
has a local minimum at x = x0 and so
1
Dφ(x0 ) = (xε − x0 ),
ε
 73

which verifies (8.11).

Now in (8.13) set x0 − xε = x − y, that is, choose y = x − x0 + xε . Then

u(xε ) − φ(x0 ) ≥ u(y) − φ(y + x0 − xε ),

for |y − xε | = |x − x0 | < r. Since ψ(y) = φ(y + x0 − xε ) we find that

u(xε ) − ψ(xε ) ≥ u(y) − ψ(y)

whenever |y − xε | < r; that is, u − ψ has a local maximum at xε .

We now show that sup- and inf-convolutions preserve viscosity sub- and
super-solution properties.

Theorem 8.7. Let U ⊂ Rn be open and bounded. If u ∈ USC(U ) is a viscosity

subsolution of (8.9) then the sup-convolution uε is a viscosity solution of

F (D2 uε , Duε ) ≤ 0 in M ε (u) ⊂ U. (8.14)

Similarly, if u ∈ LSC(U ) is viscosity supersolution of (8.9) then the inf-

convolution uε is a viscosity solution of

F (D2 uε , Duε ) ≥ 0 in Mε (u) ⊂ U. (8.15)

Proof. Let u ∈ USC(U ) be a viscosity subsolution of (8.9) and define the

sup-convolution uε . Let x0 ∈ M ε (u). Then there exists y0 ∈ U such that
1
uε (x0 ) = u(y0 ) − |x0 − y0 |2 .
2ε
By Proposition 8.6 u − ψ has a local maximum at y0 , where ψ(x) = φ(x +
x0 − y0 ). Since u is a viscosity subsolution of (8.9) we have

F (D2 φ(x0 ), Dφ(x0 )) = F (D2 ψ(y0 ), Dψ(y0 )) ≤ 0.

The proof that uε is a viscosity solution of (8.15) is similar.

Remark 8.8. It can be the case that M ε (u) = U or Mε (u) = U . Indeed, if

u = 0 on ∂U and u > 0 in U , then we immediately have M ε (u) = U , and if
u < 0 in U then Mε (u) = U . In general, it follows from Lemma 8.5 that

M ε (u) ∩ Mε (u) ⊃ x ∈ U : dist(x, ∂U ) ≥ Cε1/(2−α) ,

where α = 0 when u is bounded, and α > 0 if u ∈ C 0,α (U ).

74 CHAPTER 8. SMOOTHING VISCOSITY SOLUTIONS

We can immediately state a corollary for equations of the form

F (D2 u, Du) = f in U, (8.16)

where f : U → R.

Corollary 8.9. Let U ⊂ Rn be open and bounded and suppose f ∈ C(U ) with
modulus of continuity ω. If u ∈ USC(U ) is a bounded viscosity subsolution of
(8.16) then the sup-convolution uε is a viscosity solution of

F (D2 uε , Duε ) ≤ f + ω(Cε1/2 ) in M ε (u) ⊂ U (8.17)

Similarly, if u ∈ LSC(U ) is bounded viscosity supersolution of (8.16) then the

inf-convolution uε is a viscosity solution of

F (D2 uε , Duε ) ≥ f − ω(Cε1/2 ) in Mε (u) ⊂ U. (8.18)

Proof. Let x0 ∈ M ε (u) and let φ ∈ C ∞ (Rn ) such that uε − φ has a local
maximum at x0 . As in the proof of Theorem 8.7 we define ψ(y) := φ(y+x0 −y0 )
where y0 ∈ U is a point for which uε (x0 ) = u(y0 ) − 2ε
1
|x0 − y0 |2 . Then u − ψ
has a local maximum at y0 and so

F (D2 φ(x0 ), Dφ(x0 )) = F (D2 ψ(y0 ), Dψ(y0 )) ≤ f (y0 ).

By Proposition 8.3 we have

f (y0 ) ≤ f (x0 ) + ω(|x0 − y0 |) ≤ f (x0 ) + ω(Cε1/2 ).

The proof of (8.17) is similar.

Remark 8.10. In Corollary 8.9, if u ∈ C 0,α (U ) for 0 < α ≤ 1, then we can

use Lemma 8.5 to obtain


F (D2 uε , Duε ) ≤ f + ω Cε1/(2−α) in M ε (u).

In particular, if u and f are Lipschitz continuous then

F (D2 uε , Duε ) ≤ f + Cε in M ε (u).

A similar remark holds for uε .

Our final approximation result is for general first order equations.

H(Du, u, x) = 0 in U. (8.19)
 75

Theorem 8.11. Let U ⊂ Rn be open and bounded, and assume H ∈ Cloc 0,1
(Rn ×
R × U ). If u ∈ C 0,1 (U ) is a viscosity subsolution of (8.19) then the sup-
convolution uε is a viscosity solution of

H(Duε , uε , x) ≤ Cε in M ε (u) ⊂ U. (8.20)

Similarly, if u ∈ C 0,1 (U ) is a viscosity supersolution of (8.19) then the inf-

convolution uε is a viscosity solution of

H(Duε , uε , x) ≥ −Cε in Mε (u) ⊂ U. (8.21)

In both cases, the constant C depends only on H and kukC 0,1 (U ) .

Proof. Since u ∈ C 0,1 (U ), there exists K > 0 such that kukL∞ (U ) ≤ K, and
|Du| ≤ K in U in the viscosity sense (see Exercise 2.16). Since H is locally
Lipschitz, there exists C > 0 such that

|H(p, z, x) − H(p, r, y)| ≤ C(|z − r| + |x − y|) (8.22)

for |z|, |r|, |p| ≤ K and x, y ∈ U .

Let x0 ∈ M ε (u) and let φ ∈ C ∞ (Rn ) such that uε −φ has a local maximum
at x0 . As in the proof of Theorem 8.7 we define ψ(y) := φ(y + x0 − y0 ) where
y0 ∈ U is a point for which uε (x0 ) = u(y0 ) − 2ε
1
|x0 − y0 |2 . Then u − ψ has a
local maximum at y0 and so

H(Dψ(y0 ), u(y0 ), y0 )) ≤ 0.

Since |Du| ≤ K in the viscosity sense, we also have |Dψ(y0 )| ≤ K. Since

Dψ(y0 ) = Dφ(x0 ), we have by (8.22) and Lemma 8.5 that

|H(Dψ(y0 ), u(y0 ), y0 )) − H(Dφ(x0 ), u(x0 ), x0 )|

≤ C(|u(x0 ) − u(y0 )| + |x0 − y0 |)
≤ C|x0 − y0 |
≤ Cε,

which completes the proof of (8.20). The proof of (8.21) is similar.

To conclude this section, we use the inf- and sup-convolution

√ tools devel-
oped here to give an alternative proof of the O( ε) rate in the method of
vanishing viscosity.
76 CHAPTER 8. SMOOTHING VISCOSITY SOLUTIONS

Alternative proof of Theorem 5.8. Recall that uε ∈ C 2 (U ) ∩ C(U ) is the clas-

sical solution of
uε + H(Duε , x) − ε∆uε = 0 in U (8.23)
with boundary condition uε = 0 on ∂U (and not the inf-convolution), and
u ∈ C 0,1 (U ) is the unique viscosity solution of
u + H(Du, x) = 0 in U
with Dirichlet condition u = 0 on ∂U . Our goal is to show that
√
|u − uε | ≤ C ε.
For δ > 0 let uδ be the sup-convolution
 
1
u (x) = sup u(y) − |x − y| .
δ 2

y∈U 2δ
Then uδ is semiconvex with constant − 1δ , i.e., −D2 uδ ≤ 1δ I on U in the viscosity
sense, and by Theorem 8.11 there exists C > 0 such that uδ is a viscosity
solution of
uδ + H(Duδ , x) ≤ Cδ in UCδ ,
where UCδ = {x ∈ U : dist(x, ∂U ) ≥ Cδ}. Let x0 ∈ U such that
max(uδ − uε ) = uδ (x0 ) − uε (x0 ).
U

If x0 6∈ UCδ then dist(x0 , ∂U ) < Cδ and hence by Lemma 8.5 and the nonneg-
ativity of uε we have
max(uδ − uε ) ≤ uδ (x0 ) − uε (x0 ) ≤ Cδ.
U

If x0 ∈ UCδ then uδ − uε has a local maximum at x0 and so −D2 uε (x0 ) ≤ 1δ I

and
uδ (x0 ) + H(Duε (x0 ), x0 ) ≤ Cδ.
Subtracting (8.23) we have


uδ (x0 ) − uε (x0 ) ≤ −ε∆uε (x0 ) + Cδ ≤ C ε
δ
+δ .
√
Optimizing over δ yields δ = ε and hence
√
u − uε ≤ uδ − uε ≤ C ε.
We leave the proof of the other direction to Exercise 8.12.
Exercise 8.12. √Complete the alternative proof of Theorem 5.8 by showing
the uε − u ≤ C ε. Use the inf-convolution instead of the sup-convolution,
and recall the barrier function argument from Theorem 5.4.
Chapter 9

Finite difference schemes

We briefly consider here the problem of approximating viscosity solutions with

finite difference schemes. For simplicity, we restrict our attention to the unit
box [0, 1]n . Thus we consider the Hamilton-Jacobi equation
)
H(Du, u, x) = 0 in (0, 1)n
(9.1)
u = g on ∂(0, 1)n .
Our goal is to design finite difference schemes for (9.1) that converge to the
viscosity solution of (9.1) as the grid resolution tends to zero.
We first introduce some notation. For h > 0 let Zh = {hz : z ∈ Z} and
Zh = (Zh )n . For a set O ⊂ Rn we define Oh := O ∩ Znh , and ∂Oh := (∂O) ∩ Znh .
n

We will always assume that 1/h is an integer. Given a function u : [0, 1]hn → R,
we define the forward and backward difference quotients by
u(x ± hei ) − u(x)
∇±
i u(x) := ± , (9.2)
h
and we set
∇± u(x) = (∇± ±
1 u(x), . . . , ∇n u(x)).

When u is a smooth function restricted to the grid, the forward and backward
difference quotients (9.2) offer O(h) (or first order) accurate approximations
of uxi . This can be immediately verified by expanding u via its Taylor series.
The idea is to restrict (9.1) to the grid [0, 1]nh , and replace each partial
derivative by a corresponding finite difference. However, some care must be
taken in how this is done.
Exercise 9.1. Consider the following finite difference scheme for the one di-
mensional eikonal equation (1.6) from Exercise 1.4:
|∇+
1 uh (x)| = 1 for x ∈ [0, 1)h , and uh (0) = uh (1) = 0. (9.3)

77
78 CHAPTER 9. FINITE DIFFERENCE SCHEMES

Show that the scheme is not well-posed, that is, depending on whether 1/h is
even or odd, there is either no solution, or there is more than one solution.

9.1 Monotone schemes

In light Exercise 9.1, we cannot arbitrarily select forward or backward dif-
ferences and hope to get a convergent scheme. To see what we should do,
consider the special case that (9.1) is the Hamilton-Jacobi-Bellman equation
(4.15), and H is given by (4.14), which we recall here:

H(p, x) = sup {−p · a − L(a, x)} .

|a|=1

In this case, the solution u satisfies the dynamic programming principle (4.11)

u(x) = inf {u(y) + T (x, y)}.

y∈∂B(x,r)

The infimum on the right is attained at some y ∈ ∂B(x, r) so we have

u(x) = u(y) + T (x, y).

Since T (x, y) > 0, this expresses two things. First, there must exist y ∈
∂B(x, r) such that u(y) < u(x), and second, u(x) depends only on the neigh-
boring values u(y) that are smaller than u(x). Keeping these ideas in mind,
we define the monotone finite differences
−
∇m
i u = m(∇i u, ∇i u),
+
(9.4)

where 

a, if a + b < 0 and a ≤ 0
m(a, b) = b, if a + b ≥ 0 and b ≥ 0


0, otherwise.
We also define the monotone gradient by

∇m u = (∇m
1 u, . . . , ∇n u).
m

The monotone finite difference ∇m

i u selects the forward difference when

u(x + hei ) = min{u(x + hei ), u(x − hei )} ≤ u(x),

the backward difference when

u(x − hei ) = min{u(x + hei ), u(x − hei )} ≤ u(x),

9.1. MONOTONE SCHEMES 79

and returns zero when

u(x) < min{u(x + hei ), u(x − hei )}.
This is consistent with our observations above that there must exist a neigh-
boring grid point where u is smaller, and u locally depends only on such a grid
point.
Exercise 9.2. Consider the following monotone finite difference scheme for
the one dimensional eikonal equation:

1 uh (x)| = 1 for x ∈ (0, 1)h , and uh (0) = uh (1) = 0.

|∇m
Find the solution uh explicitly, and show that uh → 1
2
− |x − 21 | as h → 0+ .
An important property of the monotone difference is the following anti-
monotonicity.
Proposition 9.3. If u(x) = v(x) and u ≤ v then
|∇m
i u(x)| ≥ |∇i v(x)|
m
for all i.
Proof. The proof follows immediately from the observation that
1
|∇m
i u(x)| = max {(u(x) − u(x + hei ))+ , (u(x) − u(x − hei ))+ } ,
h
where t+ := max{0, t}.
We should note that the condition u ≤ v from Proposition 9.3 only needs
to hold at neighboring grid points to x, i.e., x − hei and x + hei .
Lemma 9.4. Suppose H is given by (4.14) and L satisfies
L(a1 , . . . , an , x) = L(|a1 |, . . . , |an |, x) for all x. (9.5)
If u(x) = v(x) and u ≤ v then
H(∇m u(x), x) ≥ H(∇m v(x), x). (9.6)
Proof. By (9.5) we have
H(p1 , . . . , pn , x) = sup {−p · a − L(a, x)}
|a|=1

= sup {a1 |p1 | + · · · + an |pn | − L(a, x)} .

|a|=1

Therefore H(p, x) ≤ H(q, x) whenever |pi | ≤ |qi | for all i. Combining this with
Proposition 9.3 completes the proof.
80 CHAPTER 9. FINITE DIFFERENCE SCHEMES

Let us suppose now that H satisfies the hypotheses of Lemma 9.4 and
suppose uh : [0, 1]nh → R is a solution of the numerical scheme

H(∇m uh (x), x) = 0 for x ∈ (0, 1)nh .

Suppose we can also show that uh → u ∈ C(U ) uniformly on [0, 1]n as h → 0+ .

Let x ∈ (0, 1)n and φ ∈ C ∞ (Rn ) such that u − φ has a strict maximum at x.
Then there exists hk → 0+ and xk → x such that xk ∈ (0, 1)nh and uhk − φ has
a maximum at xk over the grid [0, 1]nhk . By shifting φ, if necessary, we may
assume that uhk (xk ) = φ(xk ) and uhk ≤ φ. By Lemma 9.4 we have

0 = H(∇m uhk (xk ), xk ) ≥ H(∇m φ(xk ), xk ) −→ H(Dφ(x), x)

as hk → 0+ . Therefore u is a viscosity subsolution of H = 0. We can argue

that u is a supersolution similarly.
We note that the key part of the argument above was using (9.6) to re-
place the numerical solution uhk by a smooth test function φ. In more general
situations, when H may not be given by (4.14), we can still look for a scheme
satisfying a condition like (9.6). Such schemes are called monotone, and some-
times upwind.
A general finite difference scheme for (9.1) is of the form
)
Sh (uh , uh (x), x) = 0 in (0, 1)nh
(9.7)
uh = g on ∂(0, 1)nh ,

where
Sh : Xh × R × Rn → R,
and Xh denotes the collection of real-valued functions on [0, 1]nh . We remark
that the first argument of Sh represents the dependence of Sh on neighboring
grid points, while the second argument represents the dependence of Sh on the
grid point x.
Definition 9.5. We say the scheme Sh is monotone if

u ≤ v =⇒ Sh (u, t, x) ≥ Sh (v, t, x) (9.8)

for all u, v ∈ Xh , t ∈ R and x ∈ Rn .

When H is given by (4.14) and L satisfies (9.5), the scheme

Sh (u, u(x), x) := H(∇m u(x), x)

is monotone by Lemma 9.4.

9.2. CONVERGENCE OF MONOTONE SCHEMES 81

9.2 Convergence of monotone schemes

We show here that monotone schemes are convergent provided they are stable
and consistent, and the limit PDE is well-posed. These results can be found
in [2] in the more general context of degenerate elliptic equations.
We now give the definitions of consistency and stability.

Definition 9.6. We say the scheme Sh is consistent if

lim Sh (φ + γ, φ(y) + γ, y) = H(Dφ(x), φ(x), x)

y→x
(9.9)
h→0+
γ→0

for all φ ∈ C ∞ (Rn ).

We note that even though viscosity solutions are not in general smooth,
consistency need only be verified for smooth test functions φ ∈ C ∞ (Rn ).

Definition 9.7. We say the scheme Sh is stable if the solutions uh are uni-
formly bounded as h → 0+ , that is, there exists C > 0 such that

sup sup |uh (x)| ≤ C.

h>0 x∈[0,1]n
h

Finally, we need to assume that a comparison principle holds for (9.1). In

particular we assume that (9.1) enjoys strong uniqueness. This means that
whenever u ∈ USC([0, 1]n ) and v ∈ LSC([0, 1]n ) are viscosity sub- and super-
solutions of (9.1) in the sense of Definition 6.2, we have u ≤ v on [0, 1]n .

Theorem 9.8. Suppose (9.1) enjoys strong uniqueness, and Sh is monotone,

consistent, and stable. Then uh → u uniformly on [0, 1]n as h → 0+ , where u
is the unique viscosity solution of (9.1).

The proof is similar to Theorem 5.4. We sketch the details below. We note
that in the context of the following proof, all viscosity solutions are interpreted
in the sense of Definition 6.2.

Proof. We define the upper and lower weak limits by

u(x) = lim sup uh (y) and u(x) = lim inf uh (y).

(y,h)→(x,0+ ) (y,h)→(x,0+ )

The limits above are taken with y ∈ [0, 1]nh . Since the scheme is stable, both
u and u are bounded real valued functions.
82 CHAPTER 9. FINITE DIFFERENCE SCHEMES

We claim that u is a viscosity subsolution of (9.1) in the sense of Definition

6.2. To see this, let x0 ∈ (0, 1)n and let φ ∈ C ∞ (Rn ) such that u − φ has a
strict maximum at x0 . As in the proof of Theorem 5.4, there exists hk → 0
and xk → x0 such that uhk (xk ) → u(x0 ) and uhk − φ has a maximum at xk ,
relative to the grid [0, 1]nh . Let us write

φk (x) = φ(x) + γk ,

where γk = uhk (xk ) − φ(xk ). Then φk (xk ) = uhk (xk ) and uhk ≤ φk . Since the
scheme Sh is monotone we have

0 = Sh (uhk , uhk (xk ), xk ) ≥ Sh (φ + γk , φ(xk ) + γk , xk ).

Since Sh is consistent, we can send k → ∞ to find that

H(Dφ(x0 ), φ(x0 ), x0 ) ≤ 0. (9.10)

If x0 ∈ ∂[0, 1]n , then we can arrange it so that xk ∈ ∂[0, 1]nhk for all k, or
xk ∈ (0, 1)nhk for all k. In the first case, we have

u(x0 ) = lim+ uhk (xk ) ≤ g(x0 ),

hk →0

due to the continuity of g. The second case proceeds as above and we find
that (9.10) holds. Therefore u is a viscosity subsolution of (9.1).
That u is a viscosity supersolution of (9.1) is verified similarly. By strong
uniqueness we have u = u. Therefore uh → u uniformly, where u is the unique
viscosity solution of (9.1).

Exercise 9.9. Suppose the numerical solutions uh are uniformly Lipschitz

continuous, i.e., there exists C > 0 such that

|uh (x) − uh (y)| ≤ C|x − y| for all x, y ∈ [0, 1]nh and h > 0.

This is a stronger form of stability. Prove Theorem 9.1 without the strong
uniqueness hypothesis. You can assume that ordinary uniqueness holds, that
is, there is at most one viscosity solution of (9.1) satisfying the boundary con-
ditions in the usual sense. [Hint: Use the Arzelà-Ascoli Theorem to extract a
subsequence uhk converging uniformly to a continuous function u ∈ C([0, 1]n ).
Show that u is the unique viscosity solution of (9.1), and conclude that the
entire sequence must converge uniformly to u.]
9.2. CONVERGENCE OF MONOTONE SCHEMES 83

Exercise 9.10. Suppose that Sh depends only on the forward and backward
neighboring grid points in each direction, so that we can write

Sh (u, u(x), x) = F (∇− −

1 u(x), −∇1 u(x), . . . , ∇n u(x), −∇n u(x), u(x), x).
+ +

Let us set F = F (a1 , . . . , a2n , z, x). You may assume that H and F are smooth.

(a) Show that Sh is monotone if and only if Fai ≥ 0 for all i.

(b) Show that Sh is consistent if and only if

F (p1 , −p2 , . . . , pn , −pn , z, x) = H(p, z, x)

for all p ∈ Rn , z ∈ R and x ∈ [0, 1]nh .

(c) Find a monotone and consistent scheme for the linear PDE

a1 ux1 + · · · + an uxn = f (x),

where a1 , . . . , an are real numbers. Compare your scheme with the direc-
tion of the projected characteristics. [Hint: Your solution should depend
on the signs of the ai .]

(d) Suppose that H is Lipschitz continuous and define

a := sup {|Dp H(p, z, x)| : p ∈ Rn , z ∈ R, x ∈ [0, 1]n } .

The Lax-Friedrichs scheme is defined by

ah
Sh (u, u(x), x) := H (∇h u(x), u(x), x) − ∆h u(x),
2
where
 
u(x + he1 ) − u(x − he1 ) u(x + hen ) − u(x − hen )
∇h u(x) := ,..., ,
2h 2h

and
X
n
u(x + hei ) − 2u(x) + u(x − hei )
∆h u(x) := .
i=1
h2
Show that the Lax-Friedrichs scheme is monotone and consistent. [Hint:
Rewrite the scheme as a function of the forward and backward differences
∇±i u(x), as above.]
84 CHAPTER 9. FINITE DIFFERENCE SCHEMES

Exercise 9.11. Let U := B 0 (0, 1) and ε > 0. Consider the nonlocal integral
equation
 Z
(1 + cε2 )uε (x) − − uε dy = cε2 f (x) if x ∈ U
(Iε ) B(x,ε)

uε (x) = 0 if x ∈ Γε ,

where c = 1
2(n+2)
, uε : Γε ∪ U → R, f ∈ C(U ), and

Γε = {x ∈ Rn \ U : dist(x, ∂U ) ≤ ε}.

Follow the steps below to show that as ε → 0+ , uε converges uniformly to the

viscosity solution u of
(
u − ∆u = f in U
(P)
u = 0 on ∂U.

The proof is based on recognizing (Iε ) as a monotone approximation scheme for

(P). Unless otherwise specified, any function u : U → R is implicitly extended
to be identically zero on Γε .
(a) Show that there exists a unique function uε ∈ C(U ) solving (Iε ). [Hint:
Show that the mapping T : C(U ) → C(U ) defined by
Z
1 cε2
T [u](x) := − u dy + f (x)
1 + cε2 B(x,ε) 1 + cε2

is a contraction mapping. Use the usual norm kuk := maxU |u| on C(U ).
Then appeal to Banach’s fixed point theorem.]

(b) Define Sε : L∞ (U ∪ Γε ) × R × U → R by
Z
Sε (u, t, x) := (1 + cε )t − −
2
u dy.
B(x,ε)

Show that Sε is monotone, i.e., for all t ∈ R, x ∈ U , and u, v ∈ L∞ (U ∪ Γε )

u ≤ v on B(x, ε) =⇒ Sε (u, t, x) ≥ Sε (v, t, x).

(c) Show that the following comparison principle holds: Let u, v ∈ L∞ (U ∪

Γε ) such that u|U , v|U ∈ C(U ). If u ≤ v on Γε and Sε (u, u(x), x) ≤
Sε (v, v(x), x) at all x ∈ U , then u ≤ v on U .
9.3. LOCAL TRUNCATION ERROR 85

(d) Use the comparison principle to show that there exists C > 0 such that
|uε (x)| ≤ C(1 + 3ε − |x|2 ),
for all x ∈ U and 0 < ε ≤ 1, where C depends only on kf k = maxU |f |.
[Hint: Compare against v(x) := C(1 + 3ε − |x|2 ) and −v, and adjust the
constant C appropriately.]
(e) Use the method of weak upper and lower limits to show that uε → u
uniformly on U , where u is the viscosity solution of (P). You may assume
a comparison principle holds for (P) for semicontinuous viscosity solutions.
That is, if u ∈ USC(U ) is a viscosity subsolution of (P) and v ∈ LSC(U )
is a viscosity supersolution, and u ≤ v on ∂U , then u ≤ v in U . [Hint:
You will find the identity in the hint from Exercise 2.19 useful.]

9.3 Local truncation error

The monotonicity condition (Definition 9.5) places severe restrictions on the
types of schemes available. It turns out that all monotone schemes for first
order equations are at best O(h) accurate. By this, we mean that the local
truncation error (obtained by substituting smooth functions into the scheme)
is no better than O(h). In other words, any scheme with a local truncation
error of O(h2 ) or better cannot be monotone.
There are many higher order schemes for Hamilton-Jacobi equations. For
example, essentially non-oscillatory (ENO) schemes [15], which were originally
proposed for hyperbolic conservation laws and have been adapted to Hamilton-
Jacobi equations, are widely used and quite successful. Since the ENO schemes
are non-monotone, there is as of yet no rigorous theory guaranteeing conver-
gence to the viscosity solution.
As in Exercise 9.10, we write our monotone scheme as
F [u](x) = F (∇− −
1 u(x), −∇1 u(x), . . . , ∇n u(x), −∇n u(x), u(x), x).
+ +

Thus, we are assuming that the neighborhood N (x) of x contains just the
forward and backward neighbors in each coordinate direction. Let us write
F = F (a1 , . . . , a2n , z, x)
for notational simplicity. Recall from Exercise (9.8) that F is monotone if
and only if F is nondecreasing in each ai , i.e., Fai ≥ 0 for all i. In this case,
consistency of the scheme states that
F (p1 , −p1 , . . . , pn , −pn , z, x) = H(p, z, x). (9.11)
86 CHAPTER 9. FINITE DIFFERENCE SCHEMES

Let M > 0 and define

n o
∞
SM := φ ∈ C (R ) : kφkC 3 (Rn ) ≤ M .
n

We define the local truncation error by

err(M, h) := sup |F [φ](x) − H(Dφ(x), φ(x), x)|.

φ∈SM
x∈[0,1]n

Theorem 9.12. Let F be monotone and smooth, and assume H is smooth.

Suppose that for some p ∈ Rn , z ∈ R, x ∈ [0, 1]n , and i ∈ {1, . . . , n}

Hpi (p, z, x) 6= 0. (9.12)

Then there exists M > 0, C > 0, c > 0 and h > 0 such that for all 0 < h < h

ch ≤ err(M, h) ≤ Ch. (9.13)

Remark 9.13. The condition (9.12) says that H is not a trivial zeroth order
PDE, such as H(p, z, x) = z.
Proof. The basic idea of the proof is that the monotonicity of F ensures that
the second order terms in the Taylor expansion for u(x) − u(x ± hei ) cannot
be cancelled out to improve accuracy.
Without loss of generality, let us assume that i = 1, z = 0 and x = 0 in
(9.12). If p1 = 0, then we can find a nearby point p where p1 6= 0 and (9.12)
holds, by smoothness of H. Hence we may assume that p1 > 0. By consistency
(Eq. (9.11)) we have

Fa1 (a0 , 0, 0) − Fa2 (a0 , 0, 0) 6= 0,

where a0 = (p1 , −p1 , . . . , pn , −pn ). Since F is monotone, Fai ≥ 0 for all i. We

may, without loss of generality, assume that Fa1 (a0 , 0, 0) > 0. Hence, there
exists r > 0 and θ > 0 such that

Fa1 (a, 0, 0) ≥ θ whenever |a − a0 |2 ≤ r2 . (9.14)

Define
1 1
φ(x) = p1 (x1 + 1)2 − p1 + p2 x2 + · · · + pn xn .
2 2
Then Dφ(0) = p and φ(0) = 0. We also note that
1 − 1 ±
1 φ(0) = p1 + p1 h, ∇1 φ(0) = p1 − p1 h, and ∇i φ(0) = pi
∇+
2 2
 √
9.4. THE O( h) RATE 87

for i = 2, . . . , n. Due to the monotonicity of F and (9.14) we have

1 1
F [φ](0) = F (p1 − p1 h, −p1 − p1 h, p2 , −p2 , . . . , pn , −pn , 0, 0)
2 2
1
≤ F (p1 , −p1 , . . . , pn , −pn , 0, 0) − p1 θh
2
1
= H(p, 0, 0) − p1 θh,
2
provided 12 p1 h < r. Thus, there exists h > 0 such that
1
H(Dφ(0), φ(0), 0) − F [φ](0) ≥ p1 θh =: ch
2
for all 0 < h < h. We can multiply φ by a bump function to ensure that
φ ∈ Cc∞ (Rn ), hence φ ∈ SM for M large enough.
The upper bound in (9.13) follows from the fact that F is smooth, hence
Lipschitz on compact sets, and

|∇± φ(x) − Dφ(x)| ≤ Ch,

for all φ ∈ SM , where C depends on M .

√
9.4 The O( h) rate
Even though monotone schemes have O(h) local truncation errors, it turns
out that the√best global errors that can be established√rigorously are worse;
they are O( h). This should be compared with the O( ε) convergence rates
established in Chapter 5 for the method of vanishing viscosity. Intuitively, the
reason for this is that local truncation errors consider how the scheme acts on
smooth functions, and viscosity solutions are in general not smooth. Thus,
the usual trick of substituting the solution of the PDE into the scheme does
not convert local errors into global errors for viscosity solutions. However,
see Section 5.3 for situations where the viscosity solution satisfies a one-sided
second derivative bound. In this situation, we would expect a one-sided O(h)
rate.
Nevertheless, it is commonplace in practice to observe global errors on
the order of O(h) in numerical experiments, even when the solutions are not
smooth. There is currently no theory that fully explains this difference between
the experimental and theoretical convergence rates.
We assume that L is Lipschitz continuous, and satisfies (9.5) as well as all
of the assumptions of Chapter 4, and we take H to be given by (4.14).
88 CHAPTER 9. FINITE DIFFERENCE SCHEMES

Proposition 9.14. The Hamiltonian H is Lipschitz continuous.

Proof. It follows from Lemma 4.5 that H is Lipschitz in x. Let x, p, q ∈ Rn
and choose a ∈ Rn with |a| = 1 such that

H(p, x) = −p · a − L(a, x).

Then we have
H(q, x) ≥ −q · a − L(a, x),
and so
H(p, x) − H(q, x) ≤ (q − p) · a ≤ |q − p|.
Therefore H is Lipschitz continuous.
Let u ∈ C 0,1 ([0, 1]n ) be the unique viscosity solution of
)
H(Du, x) = 0 in (0, 1)n
(9.15)
u = 0 on ∂(0, 1)n ,

and consider the monotone finite difference scheme

)
H(∇m uh (x), x) = 0 in (0, 1)nh
(9.16)
uh = 0 on ∂(0, 1)nh .

We first aim to establish existence of a unique solution of (9.16). For this,

we need a discrete comparison principle. We say that uh : [0, 1]nh → R is a
subsolution of (9.16) if H(∇m uh , x) ≤ 0 in (0, 1)nh and uh ≤ 0 on ∂(0, 1)nh . We
define supersolutions analogously.
The comparison principle for (9.16) is based on the maximum principle.
When u and v are smooth functions, the maximum principle is based on the
fact that when u − v has a maximum at x0 , we have Du(x0 ) = Dv(x0 ) and
hence
H(Du(x0 ), x0 ) = H(Dv(x0 ), x0 ). (9.17)
If u and v are functions on the grid [0, 1]nh , then at a max of u − v we have

∇− −
i u(x0 ) ≥ ∇i v(x0 ) and ∇i u(x0 ) ≤ ∇i v(x0 ).
+ +

Hence we cannot expect equality like (9.17) at the discrete level. Monotone
schemes are designed precisely to give the correct inequality so that the max-
imum principle holds.
To see how this works, recall from Lemma 9.4 that if u(x0 ) = v(x0 ) and
u ≤ v, then H(∇m u(x0 ), x0 ) ≥ H(∇m v(x0 ), x0 ). We can rephrase this in
 √
9.4. THE O( h) RATE 89

the language used above. Suppose that u − v has a maximum at x0 . Then

e(x0 ) = v(x0 ) and u
u e ≤ v, where u
e(x) := u(x) + v(x0 ) − u(x0 ). Since u
e and u
differ by a constant, we have

e(x0 ), x0 ) ≥ H(∇m v(x0 ), x0 ).

H(∇m u(x0 ), x0 ) = H(∇m u

This is important, so we we repeat

If u − v has a max at x then H(∇m u(x), x) ≥ H(∇m v(x), x). (9.18)

This is the discrete analogue of (9.17) and is exactly what allows maximum
principle arguments to hold for monotone finite difference schemes.

Lemma 9.15. If u and v are sub- and supersolutions of (9.16), respectively,

then u ≤ v on [0, 1]nh .

Proof. We will show that for every θ ∈ (0, 1), θu ≤ v. Fix θ ∈ (0, 1) and
assume to the contrary that max[0,1]nh (θu − v) > 0. Let x ∈ [0, 1]nh be a point
at which θu − v attains its positive maximum. Then by (9.18) we have

H(∇m v(x), x) ≤ H(θ∇m u(x), x).

Since u ≤ 0 ≤ v on ∂(0, 1)nh , we must have x ∈ (0, 1)nh . Due to the convexity
of H we have

H(θ∇m u(x), x) = H (θ∇m u(x) + (1 − θ) · 0, x)

≤ θH (∇m u(x), x) + (1 − θ)H(0, x) ≤ −(1 − θ)γ,

where γ is a positive constant depending on L. We therefore deduce

H(∇m v(x), x) ≤ −(1 − θ)γ,

which is a contradiction. Therefore θu ≤ v for all θ ∈ (0, 1), and hence

u ≤ v.

Lemma 9.16. There exists a unique grid function uh : [0, 1]nh → R satisfying
the monotone scheme (9.16). Furthermore, the sequence uh is nonnegative and
uniformly bounded.

The proof of Lemma 9.16 is based on the Perron method, but is consider-
ably simpler due to the discrete setting.
90 CHAPTER 9. FINITE DIFFERENCE SCHEMES

Proof. Define
n o
F = u : [0, 1]nh → R : u is a nonnegative subsolution of (9.16) .

Since H(0, x) ≤ 0, u ≡ 0 is a subsolution, so F is nonempty. Furthermore, we

claim that there exists a constant C > 0 such that for every u ∈ F
u ≤ C.
To see this, note that there exists C > 0 so that L(a, x) ≤ C for all x ∈ [0, 1]n
and |a| = 1. Therefore
H(p, x) ≥ sup {−p · a − C} ≥ |p| − C,
|a|=1

and so v(x) := Cx1 is a supersolution of (9.16). By Lemma 9.15, u ≤ v ≤ C,

which establishes the claim.
We now define
uh (x) := sup{u(x) : u ∈ F}.
We first show that uh is a subsolution of (9.16). Fix x0 ∈ (0, 1)nh and let
v k ∈ F such that v k (x0 ) → uh (x0 ) as k → ∞. By passing to a subsequence,
if necessary, we may assume that v k → v on [0, 1]nh as k → ∞ for some
v : [0, 1]nh → R. Then clearly uh (x0 ) = v(x0 ), uh ≥ v and by continuity of H,
v is a subsolution of (9.16). It follows from monotonicity that
H(∇m uh (x0 ), x0 ) ≤ H(∇m v(x0 ), x0 ) ≤ 0.
This establishes that uh is a subsolution of (9.16).
We now show that uh is a supersolution of (9.16). Assume to the contrary
that there exists x0 ∈ (0, 1)nh such that
H(∇m uh (x0 ), x0 ) < 0.
For ε > 0 define (
uh (x) + ε, if x = x0
v(x) :=
uh (x), otherwise.
By continuity, we can choose ε > 0 small enough so that
H(∇m uh (x0 ), x0 ) ≤ 0.
By monotonicity, uh remains a subsolution of (9.16) at other grid points.
Therefore v ∈ F and v(x0 ) > uh (x0 ), which is a contradiction. Therefore uh
is a solution of (9.16).
Uniqueness follows from Lemma 9.15.
 √
9.4. THE O( h) RATE 91
√
We are now ready to prove the O( h) convergence rate.

Theorem 9.17. There exists a constant C > 0 such that

√
|u − uh | ≤ C h.

Proof. For θ ∈ (0, 1), to be selected later, define the auxiliary function

1
Φ(x, y) = θu(x) − uh (y) − √ |x − y|2 ,
h
for x ∈ [0, 1]n and y ∈ [0, 1]nh . Let (xh , yh ) ∈ [0, 1]n × [0, 1]nh such that

Φ(xh , yh ) = max Φ.
[0,1]n ×[0,1]n
h

Since Φ(xh , yh ) ≥ Φ(yh , yh ) we have

1
θu(xh ) − uh (yh ) − √ |xh − yh |2 ≥ θu(yh ) − uh (yh ).
h
Therefore
1
√ |xh − yh |2 ≤ θ(u(xh ) − u(yh )) ≤ C|xh − yh |
h
due to the Lipschitzness of u. Therefore
√
|xh − yh | ≤ C h.

The proof of this is split into three cases now.

Case 1. If xh ∈ ∂(0, 1)n then

θu(xh ) − uh (yh ) ≤ 0

since uh is nonnegative and u(xh ) = 0.

Case 2. If yh ∈ ∂(0, 1)nh then
√
θu(xh ) − uh (yh ) = θu(xh ) − θu(yh ) ≤ C|xh − yh | ≤ C h,

since uh (yh ) = u(yh ) = 0 and θ ∈ (0, 1).

Case 3. Suppose xh ∈ (0, 1)n and yh ∈ (0, 1)nh . Then

1
x 7→ u(x) − √ |x − yh |2
θ h
92 CHAPTER 9. FINITE DIFFERENCE SCHEMES


has a maximum at xh . Letting p = √2h (xh − yh ) we have H p
θ
, xh ≤ 0.
Therefore
 p 
H(p, xh ) = H θ + (1 − θ) · 0, xh
 θp 
≤ θH , xh + (1 − θ)H(0, xh ) ≤ −(1 − θ)γ, (9.19)
θ
for some γ > 0 depending only on L. Note we used the convexity of H with
respect to p above.
Notice that
1
y 7→ uh (y) + √ |xh − y|2
h
has a maximum at y = yh . By (9.18) we have

0 = H(∇m uh (yh ), yh ) ≤ H(∇m ψ(yh ), yh ),

where ψ(y) := − √1h |xh − y|2 . Since |D2 ψ| ≤ √1 ,

h
a Taylor expansion gives
√
|∇m ψ(yh ) − Dψ(yh )| ≤ C h.

Since H is Lipschitz in all variables and Dψ(yh ) = p we have

√
H(p, yh ) + C h ≥ 0.

Combining this with (9.19) yields

√ √
γ(1 − θ) ≤ H(p, yh ) − H(p, xh ) + C h ≤ C h

for all θ ∈ (0, 1), due to the Lipschitzness of H. For h sufficiently small, we
set
(C + 1) √
θ =1− h,
γ
to obtain √ √
(C + 1) h ≤ C h.
Since this is a contradiction, case 3 is impossible. √
We have shown that there exists K > 0 such that when θ := 1 − K h we
have √
θu(xh ) − uh (yh ) ≤ C h
for h > 0 sufficiently small. Therefore, there exists h > 0 such that
√
maxn (θu − uh ) ≤ θu(xh ) − uh (yh ) ≤ C h,
[0,1]h
9.5. ONE-SIDED O(h) RATE 93

for 0 < h < h. Therefore

√ √ √
u − uh = θu − uh + (1 − θ)u ≤ C h + K h maxn u = C h.
[0,1]

For h ≥ h we have
√
e h,
maxn (u − uh ) ≤ maxn |u| + maxn |uh | ≤ C ≤ C
[0,1]h [0,1] [0,1]h

√
e := C/ h. This completes the proof.
due to Lemma 9.16, where C

Exercise
√ 9.18. Complete the proof of Theorem 9.17 by showing that uh −u ≤
C h.

9.5 One-sided O(h) rate

When the solution is semiconcave, we can prove an O(h) one-sided rate. This
result is a direct analog of the one-sided rate we obtained for the method of
vanishing viscosity in Section 5.3 via semiconcavity.
As in Section 9.4, we assume that L is Lipschitz continuous, and satisfies
(9.5) as well as all of the assumptions of Chapter 4, and we take H to be given
by (4.14).

Proposition 9.19. Suppose u : Rn → R is semiconcave with semiconcavity

constant c. Then for almost every x0 ∈ Rn

c
i u(x0 )| ≥ |uxi (x0 )| − h.
|∇m (9.20)
2

Proof. Since u is semiconcave with constant c, u − 2c |x − x0 |2 is concave. It is a

general fact that convex or concave functions are locally Lipschitz continuous,
hence differentiable almost everywhere. Hence, for almost every x0 we have

c
u(x) − |x − x0 |2 ≤ u(x0 ) + Du(x0 ) · (x − x0 ) for all x ∈ Rn .
2

Therefore
u(x0 ± hei ) − u(x0 ) c
≤ ±uxi (x0 ) + h,
h 2
94 CHAPTER 9. FINITE DIFFERENCE SCHEMES

for all h > 0 and almost every x0 ∈ Rn . It follows that

|∇m
i u(x0 )| = max{(u(x0 ) − u(x0 − hei ))+ , (u(x0 ) − u(x0 + hei ))+ }
 
c   c 
≥ max uxi (x0 ) − h , −uxi (x0 ) − h
2 + 2 +
c
≥ max {(uxi (x0 ))+ , (−uxi (x0 ))+ } − h
2
c
= |uxi (x0 )| − h
2
for almost every x0 .

Theorem 9.20. Suppose the viscosity solution u ∈ C([0, 1]n ) of (9.15) is

semiconcave, and let uh : [0, 1]nh → R be the solution of the monotone finite
difference scheme (9.16). Then there exists a constant C > 0 such that

uh − u ≤ Ch. (9.21)

Proof. Let x0 ∈ (0, 1)n such that u is differentiable at x0 . By Proposition 9.19

we have
c
|∇mi u(x0 )| ≥ |uxi (x0 )| − h
2
for all i and h > 0. As in the proof of Lemma 9.4
 c 
H(∇ u(x0 ), x0 ) ≥ H Du(x0 ) − h1, x0 ≥ −Ch,
m
2
due to the Lipschitzness of H, and the fact that the Lipschitz viscosity solution
u satisfies the PDE (9.15) at each point of differentiability; in particular, at
x0 . In the above, 1 = (1, . . . , 1) ∈ Rn . By continuity of H and u, we conclude
that
H(∇m u(x), x) + Ch ≥ 0 for all x ∈ (0, 1)nh . (9.22)
Now set vh (x) = θuh (x) for θ > 0 to be determined. Then we have

H(∇m vh (x), x) = H(θ∇m uh (x) + (1 − θ) · 0, x)

≤ θH(∇m uh (x), x) + (1 − θ)H(0, x)
≤ −(1 − θ)γ,

for a constant γ > 0 depending only on L. Note we used the convexity of H

above. Set θ = 1 − Cγ h to find that

H(∇m vh (x), x) + Ch ≤ 0 for all x ∈ (0, 1)nh ,

9.5. ONE-SIDED O(h) RATE 95

for h > 0 small enough that θ > 0. By the discrete comparison principle
(Lemma 9.15) we have that vh ≤ u on [0, 1]nh . Therefore (1 − Ch)uh ≤ u.
Since the sequence uh is uniformly bounded, we conclude that

uh − u ≤ Ch

for h > 0 sufficiently small.

Notice in the proof of Theorem 9.20, we only used the fact that u is a Lip-
schitz almost everywhere solution of (9.15). Therefore, we have the following
result.

Corollary 9.21. If w ∈ C 0,1 ([0, 1]n ) is a semiconcave Lipschitz almost every-

where solution of (9.15), then w is the unique viscosity solution of (9.15).

Proof. Let u denote the unique viscosity solution of (9.15). By Exercise 2.22, w
is a viscosity subsolution of (9.15), and so by comparison, w ≤ u. By Theorem
9.20, there exists a constant C such that w ≥ uh − Ch for all h > 0, where
uh is the solution of the monotone scheme (9.16). Since uh → u uniformly as
h → 0, we have w ≥ u, hence w = u.
96 CHAPTER 9. FINITE DIFFERENCE SCHEMES
Chapter 10

Homogenization

Let uε ∈ C(U ) be a viscosity solution of

 
x
uε + H Duε , = 0 in U 
ε (10.1)
uε = 0 on ∂U,

where H : Rn × Rn → R and U ⊂ Rn is open and bounded. Since (10.1)

is highly oscillatory when ε > 0 is small, we expect uε to have an oscillatory
component. The goal of homogenization theory is to describe these oscillations
and understand the behavior of the sequence uε as ε → 0+ .
Our primary assumption is

(Periodicity) y 7→ H(p, y) is Zn -periodic for all p ∈ Rn . (10.2)

This means that H(p, y + z) = H(p, y) for all z ∈ Zn and y ∈ Rn . We also

assume that H satisfies (3.6), (6.2), and is

(Coercive) lim inf H(p, y) > 0 uniformly in y ∈ Rn , (10.3)

|p|→∞

and
(Nonnegative) − H(0, y) ≥ 0 for all y ∈ Rn . (10.4)
We first record a Lipschitz estimate on the solution uε .

Lemma 10.1. There exists a constant C such that for all ε > 0

kuε kC 0,1 (U ) ≤ C. (10.5)

97
98 CHAPTER 10. HOMOGENIZATION

The proof of Lemma 10.1 is very similar to Lemmas 5.1 and 5.7, so we
omit it.
By the Arzelà-Ascoli Theorem, we can pass to a subsequence uεj so that
uεj → u ∈ C 0,1 (U ) uniformly on U . The goal is to identify a PDE that is
satisfied by u. To do this, we need to understand locally the structure of
uε − u. Let us suppose that near a point x0 , uε has the form

uε (x) = u(x) + εv( xε ) + O(ε2 ) as ε → 0+ ,

where the function v is Zn -periodic and may depend on the choice of x0 .

Substituting this into (10.1) we formally have

u(x) + εv(y) + H(Du(x) + Dv(y), y) + o(1) = 0 as ε → 0+

for x near x0 , where y := xε . Setting p = Du(x0 ) and formally sending ε → 0+

we find that
H(p + Dv(y), y) = λ in Rn (10.6)
for some λ ∈ R (here, λ = −u(x0 )). Equation (10.6) is called a cell problem,
and its solution v is called a corrector function. The corrector describes the
high frequency oscillations of uε about the limit u near the point x0 .
While the above argument is only a heuristic, it is important because it
allows us to identify the cell problem (10.6), which we can study rigorously.
Lemma 10.2. For each p ∈ Rn , there exists a unique real number λ such that
(10.6) has a Zn -periodic viscosity solution v ∈ C 0,1 (Rn ).
In light of the lemma, we write

H(p) := λ, (10.7)

and the heuristics above suggest that u should be the viscosity solution of

u + H(Du) = 0 in U,

satisfying u = 0 on ∂U . The function H is called the effective Hamiltonian.

Proof. The proof is split into several steps.
1. Let δ > 0 and consider the approximating PDE

δwδ + H(p + Dwδ , y) = 0 in Rn . (10.8)

The addition of the zeroth order term guarantees that a comparison principle
holds for (10.8) (see Theorem 6.4 and Corollary 3.2). We can prove existence
 99

of a viscosity solution of (10.8) via the Perron method. Indeed, let C > 0 be
large enough so that

δC + H(p, y) ≥ 0 for all y ∈ Rn .

Then the constant C is a viscosity supersolution of (10.8). Set

n o
F := u ∈ USC(Rn ) : v is a subsolution of (10.8) and v ≤ C .

e is a viscosity subsolution for large enough C

Since u = −C e > 0, the set F is
nonempty. Define
wδ (x) = sup{u(x) : u ∈ F}.
By Lemmas 7.1 and 7.2, wδ = wδ∗ is a bounded viscosity subsolution of (10.8)
and wδ,∗ is a bounded viscosity supersolution of (10.8). By comparison for
(10.8) we have wδ ≤ wδ,∗ , therefore wδ ∈ C(Rn ) is the unique bounded viscosity
solution of (10.8). By comparison for (10.8) we have wδ ≤ wδ,∗ , therefore
wδ ∈ C(Rn ) is the unique bounded viscosity solution of (10.8).
2. We now claim that wδ is Zn -periodic. Suppose to the contrary that there
exists y0 ∈ Rn and z ∈ Zn such that wδ (y0 + z) > wδ (y0 ). By the periodicity
of H, u(y) := wδ (y + z) is a viscosity solution of (10.8), and u ≤ C. Therefore
u ∈ F and by the definition of wδ

wδ (y0 ) ≥ u(y0 ) = wδ (y0 + z) > wδ (y0 ),

which is a contradiction.
3. Since wδ is Zn -periodic, similar arguments to the proof of Lemma 10.1
show that there exists C > 0 such that

kδwδ kC(Rn ) ≤ C

and
|wδ (x) − wδ (y)| ≤ C|x − y| for all x, y ∈ Rn ,
where C is independent of δ. We now define

vδ := wδ − min wδ .
nR

Then each vδ is a viscosity solution of

δvδ + H(p + Dvδ , y) = − min δwδ ,

n R
100 CHAPTER 10. HOMOGENIZATION

and the sequence vδ satisfies

sup kvδ kC 0,1 (Rn ) < ∞.

δ>0

Utilizing the above information and the Arzelà-Ascoli Theorem, we can extract
a subsequence δj → 0 such that

v δj → v and δj wδj → −λ uniformly on Rn ,

where v ∈ C 0,1 (Rn ) is Zn -periodic and λ ∈ R. By the stability of viscosity

solutions under uniform convergence we find that v is a viscosity solution of

H(p + Dv, y) = λ in Rn .

4. We now show that λ is unique. Suppose, by way of contradiction, that

v̂ ∈ C 0,1 (Rn ) is a Zn -periodic viscosity solution of

H(p + Dv̂, y) = λ̂ in Rn , (10.9)

and, say, λ̂ > λ. By the comparison principle from Theorem 6.4, we have
v ≤ v̂ in Rn . This contradicts the fact that we can add an arbitrary constant
to v̂ without changing (10.9).

Theorem 10.3. The sequence uε converges uniformly on U to the unique

viscosity solution u ∈ C 0,1 (U ) of
)
u + H(Du) = 0 in U
(10.10)
u = 0 on ∂U.

The proof of Theorem 10.3 is based on the “perturbed test function” tech-
nique, which was pioneered in [9, 10].
Proof. By Lemma 10.1 and the Arzelà-Ascoli Theorem, there exists a function
u ∈ C 0,1 (U ) and a subsequence εj → 0 such that uεj → u uniformly on U . We
claim that u is the unique viscosity solution of (10.10). Once this is established,
it immediately follows that uε → u uniformly on U .
We first verify that u is a viscosity subsolution of (10.10). The proof is
split into three steps.
1. Let x0 ∈ U and φ ∈ C ∞ (Rn ) such that u− φ has a strict local maximum
at x0 and u(x0 ) = φ(x0 ). We must show that

φ(x0 ) + H(Dφ(x0 )) ≤ 0.
 101

Assume, by way of contradiction, that

φ(x0 ) + H(Dφ(x0 )) =: θ > 0. (10.11)

Set p = Dφ(x0 ) and let v ∈ C 0,1 (U ) be a Zn -periodic viscosity solution of

H(p + Dv, y) = H(p) in Rn . (10.12)

By adding a constant to v, we may assume v ≥ 0. We now define the perturbed

test function x
φε (x) := φ(x) + εv .
ε
We note that φε is Lipschitz continuous, but not C 1 in general. Since v ≥ 0,
φε ≥ φ.
2. We now claim that
 x θ
φε + H Dφε , ≥ in B 0 (x0 , r) (10.13)
ε 2
in the viscosity sense for small enough r > 0 (to be selected later). To see this,
let x1 ∈ B 0 (x0 , r) and ψ ∈ C ∞ (Rn ) such that φε − ψ has a local minimum at
x1 and ψ(x1 ) = φε (x1 ). Then the mapping
x
x 7→ εv − (ψ(x) − φ(x)) has a minimum at x = x1 ,
ε
and hence
x1
y 7→ v(y) − η(y) has a minimum at y1 := ,
ε
where
1
η(y) := (ψ(εy) − φ(εy)).
ε
Since v is a viscosity solution of (10.12) we deduce

H(p + Dη(y1 ), y1 ) ≥ H(p).

Since φε (x0 ) ≥ φ(x0 ) we have by (10.11) that

 x1 
φε (x0 ) + H Dφ(x0 ) − Dφ(x1 ) + Dψ(x1 ), ≥ θ.
ε
By (6.2), there exists a sufficiently small radius r > 0 such that
 x1  θ
φε (x1 ) + H Dψ(x1 ), ≥ ,
ε 2
102 CHAPTER 10. HOMOGENIZATION

which establishes the claim.

3. We can select r > 0 smaller, if necessary, so that u+2δ ≤ φ on ∂B(x0 , r)
for some 0 < δ < 2θ . Then for εj sufficiently small, we have

uεj + δ ≤ φ ≤ φεj on ∂B(x0 , r).

We note that u := uεj + δ is a viscosity solution of

 x θ
u + H Du, =δ< in B 0 (x0 , r).
ε 2
By the comparison principle we have uεj + δ ≤ φεj throughout the ball
B 0 (x0 , r). Sending εj → 0, we arrive at the contradiction u(x0 ) + δ ≤ φ(x0 ).
Therefore u is a viscosity subsolution of (10.10).
The proof that u is a viscosity supersolution of (10.10) is similar.
Chapter 11

Discontinuous coefficients

Here, we briefly consider Hamilton-Jacobi equations with discontinuous coeffi-

cients, and illustrate how to define viscosity solutions and extend the compar-
ison principle to this setting. For simplicity, we consider the Hamilton-Jacobi
equation )
H(Du) = f in U
(11.1)
u = g on ∂U.
The comparison principles we have established so far, Theorems 3.1 and 6.4,
require f ∈ C(U ). There are many applications where f may be discontinuous,
as is illustrated by the example below.
Example 11.1 (Shape from shading). Let n = 2. Suppose we are photograph-
ing an object given by the graph of a function u : U → R with a camera
and light source positioned at (0, 0, T ) and pointing in the downward direction
(0, 0, −1). The camera’s light source illuminates the object, and the camera
captures a grayscale image I : U → [0, 1] proportional to the amount of re-
flected light returning to the camera. The shape from shading problem is to
reconstruct the object u from the image I.
We will show that the shape from shading problem reduces to solving a
Hamilton-Jacobi equation. Let
(ux , ux , 1)
n(x) = p 1 2
|Du|2 + 1
be the upward unit normal to the object. If the light source and camera are
far from the object (so T  1) then the intensity I(x) of the image can be
approximated by
1
I(x) = n(x) · (0, 0, 1) = p . (11.2)
|Du|2 + 1

103
104 CHAPTER 11. DISCONTINUOUS COEFFICIENTS

Therefore, we find that

p
|Du| = I(x)−2 − 1 =: f (x). (11.3)

Hence, given appropriate boundary conditions, the shape from shading problem
reduces to solving the eikonal equation. If the object u is not a smooth graph–
it may have corners–then I, and hence f , may be discontinuous. For more
details on shape from shading and connections to viscosity solutions, we refer
the reader to [16].

Example 11.1 motivates the need for a theory of viscosity solutions with
discontinuous coefficients. Since our definition of viscosity solution (Definition
2.1) assumed continuity, we first need to revisit definitions.
As motivation, we consider the method of vanishing viscosity for f possibly
discontinuous. In the viscous regularization, we replace f with the mollification
fε := η ε ∗ f , where η ε is the standard mollifier, yielding
)
H(Duε ) − ε∆uε = fε in U
(11.4)
uε = gε on ∂U.

Suppose, as before, that uε → u uniformly as ε → 0. Let x0 ∈ U and

φ ∈ C ∞ (Rn ) such that u − φ has a strict local max at x0 . Then there exists
xk → x0 and εk → 0 such that uεk − φ has a local max at xk . Therefore
Duεk (xk ) = Dφ(xk ) and ∆uεk (xk ) ≤ ∆φ(xk ). This yields

H(Dφ(xk )) − εk ∆φ(xk ) ≤ H(Duεk (xk )) − εk ∆uεk (xk ) = fεk (xk ).

Sending εk → 0 we have

H(Dφ(x0 )) ≤ lim inf fεk (xk ).

k→∞

Noting that Z
fε (x) = η ε (x − y)f (y) dy ≤ sup f
B(x,ε) B(x,ε)

we find that
H(Dφ(x0 )) ≤ lim inf sup f ≤ f ∗ (x0 ),
k→∞ B(x,ε )
k

where f ∗ is the upper semicontinuous envelope of f , defined in Chapter 7.

The discussion above motivates the following definitions.
 105

Definition 11.1 (Viscosity solution). Let f : U → R. We say that u ∈

USC(U ) is a viscosity subsolution of (11.1) if for every x ∈ U and every
φ ∈ C ∞ (Rn ) such that u − φ has a local maximum at x we have

H(Dφ(x)) ≤ f ∗ (x).

Similarly, we say that u ∈ LSC(U ) is a viscosity supersolution of (11.1) if for

every x ∈ U and every φ ∈ C ∞ (Rn ) such that u − φ has a local minimum at
x we have
H(Dφ(x)) ≥ f∗ (x).

In general, the standard doubling variables arguments used to prove com-

parison in Theorems 3.1 and 6.4 does not extend directly to Definition 11.1,
since f is not continuous. However, there are special cases where the auxiliary
function can be modified to compensate for discontinuous f . We first illustrate
this first in a special case.

Theorem 11.2. Let U = B 0 (0, 1) and set B + = U ∩{xn > 0}, B − = U ∩{xn <
0}, and Γ = U ∩ {xn = 0}. Assume that f |B + ∈ C(B + ), f |B − ∈ C(B − ) and
for all x ∈ Γ
−
lim f (y) ≤ +lim f (y). (11.5)
B ∋y→x B ∋y→x

Let ε > 0 and let u, v ∈ C 0,1 (U ) such that H(Du) ≤ f and H(Dv) ≥ f + ε in
U in the viscosity sense of Definition 11.1. Then

max(u − v) = max(u − v). (11.6)

U ∂U

The proof of Theorem 11.2 uses a modified doubling the variables argu-
ment. The proof given below is borrowed in part from [8].

Proof. First, we claim that if x, y ∈ U with yn > xn then

f ∗ (x) − f∗ (y) ≤ 2ω(|x − y|), (11.7)

where ω is the modulus of continuity of f |B + and f |B − . To see this, note that

if xn = 0 then by (11.5) f ∗ (x) = limB + ∋y→x f (y) and f is continuous at y ∈ B +
yielding
f ∗ (x) − f∗ (y) = f ∗ (x) − f (y) ≤ ω(|x − y|).
If yn = 0 then f∗ (y) = limB − ∋y→x f (y) and f is continuous at x ∈ B − . Hence

f ∗ (x) − f∗ (y) = f (x) − f∗ (y) ≤ ω(|x − y|).

106 CHAPTER 11. DISCONTINUOUS COEFFICIENTS

Finally, assume xn 6= 0 and yn 6= 0. If x, y ∈ B − or x, y ∈ B + then (11.7) is

trivially true, so we can assume x ∈ B − and y ∈ B + . Let z ∈ Γ such that z
lies on the line between x and y, that is z = θx + (1 − θ)y for some θ ∈ (0, 1)
and zn = 0. Then we have

f ∗ (x) − f∗ (y) = f (x) − f∗ (z) + f∗ (z) − f ∗ (z) + f ∗ (z) − f (y)

≤ f (x) − −lim f (w) + +lim f (w) − f (y)
B ∋w→z B ∋w→z
≤ ω(|x − z|) + ω(|y − z|)
≤ 2ω(|x − y|),

which establishes the claim.

We now prove (11.6). We can assume that u ≤ v on ∂U , and assume to
the contrary that δ := maxU (u − v) > 0. Then there exists x0 ∈ U such that
u(x0 ) − v(x0 ) = δ. We define the auxiliary function
2
α 1
Φ(x, y) = u(x) − v(y) − x − y + √ en (11.8)
2 α

where en = (0, 0, . . . , 0, 1). Let (xα , yα ) ∈ U × U such that

Φ(xα , yα ) = max Φ.
U ×U

We claim that
lim Φ(xα , yα ) = δ. (11.9)
α→∞

To see this, note that

Φ(xα , yα ) ≥ Φ(x0 , x0 + √1 en )
α
= u(x0 ) − v(x0 + √1 en )
α
→δ

as α → ∞, and so
lim inf Φ(xα , yα ) ≥ δ > 0.
α→∞

So for α > 0 large enough Φ(xα , yα ) > 0 and so

2
α 1
xα − yα + √ e n ≤ u(xα ) − v(yα ) ≤ C.
2 α

Therefore
1 C
xα − yα + √ en , |xα − yα | ≤ √ . (11.10)
α α
 107

It follows that
C
Φ(xα , yα ) ≤ u(xα ) − v(yα ) = u(xα ) − u(yα ) + u(yα ) − v(yα ) ≤ √ + δ
α
and so lim supα→∞ Φ(xα , yα ) ≥ δ, which establishes the claim.
For large enough α, u(xα ) − v(yα ) ≥ 2δ , and so xα , yα ∈ U . By the viscosity
sub- and super-solution properties we have
H(pα ) ≤ f ∗ (xα ) and H(pα ) ≥ f∗ (yα ) + ε,
where  
1
pα = α xα − yα + √ e n .
α
Therefore
ε ≤ f ∗ (xα ) − f∗ (yα ). (11.11)
√  
Setting wα = α xα − yα + √1α en we have

1
yα = xα + √ (en − wα ) . (11.12)
α
Notice that
1
|wα |2 = u(xα ) − v(yα ) − Φ(xα , yα )
2
≤ u(xα ) − u(yα ) + u(yα ) − v(yα ) − Φ(xα , yα )
C
≤ √ + δ − Φ(xα , yα ),
α
and so wα → 0 as α → ∞. It follows from (11.12) that yα,n > xα,n for α
sufficiently large. Thus by (11.7) and (11.11) we have
ε ≤ f ∗ (xα ) − f∗ (yα ) ≤ ω(|xα − yα |) ≤ ω(Cα−1/2 ).
Sending α → ∞ yields a contradiction.
We can generalize the argument in some ways. We follow [8] and make the
assumption that
(D) For all x0 ∈ U there exists εx0 > 0 and ηx0 ∈ Sn−1 such that
f ∗ (x) − f∗ (x + rd) ≤ ω(|x − x0 | + r), (11.13)
for all x ∈ U , r > 0 and d ∈ Sn−1 such that |d−ηx0 | < εx0 and x+rd ∈ U ,
where ω is a modulus of continuity.
108 CHAPTER 11. DISCONTINUOUS COEFFICIENTS

This models the situation where the domain can be decomposed as the disjoint
union U = U1 ∪ U2 ∪ Γ where U1 , U2 are open and Γ = ∂U1 ∩ ∂U2 ∩ U is the
boundary between U1 and U2 . Then (D) is satisfied provided Γ is a Lipschitz
hypersurface, f |U1 ∈ C(U1 ), f |U2 ∈ C(U2 ), and

lim f (y) ≤ lim f (y)

U1 ∋y→x U2 ∋y→x

for all x ∈ Γ.

Exercise 11.3. Prove the assertion above.

We now give a more general comparison principle assuming (D) holds, and
that H is continuous.

Theorem 11.4. Let U ⊂ Rn be open and bounded, assume f : U → R satisfies

(D) and H ∈ C(Rn ). Let ε > 0 and let u, v ∈ C 0,1 (U ) such that H(Du) ≤ f
and H(Dv) ≥ f + ε in U in the viscosity sense of Definition 11.1. Then

max(u − v) = max(u − v). (11.14)

U ∂U

Proof. We sketch the proof, as it is similar to Theorem 11.2. We may assume

that u ≤ v on ∂U , and assume to the contrary that δ := maxU (u − v) > 0.
Let x0 ∈ U such that u(x0 ) − v(x0 ) = δ. We define the auxiliary function
2
α 1
Φ(x, y) = u(x) − v(y) − x − y + √ ηx 0 − |x − x0 |2 . (11.15)
2 α

Let (xα , yα ) ∈ U × U such that

Φ(xα , yα ) = max Φ.
U ×U

As usual, we have |xα − yα | ≤ C/α, Φ(xα , yα ) → δ, and

2
α 1
x − y + √ η x0 + |xα − x0 |2 → 0 as α → ∞.
2 α

Therefore xα , yα → x0 as α → ∞ and so xα , yα ∈ U for α sufficiently large.

Write pα = α(xα − yα + √1α ηx0 ). By the viscosity sub- and supersolution
properties

H(pα + 2(x − x0 )) ≤ f ∗ (xα ) and H(pα ) ≥ f∗ (yα ) + ε.

 109

Therefore

0 < ε ≤ H(pα ) − H(pα + xα − x0 ) + f ∗ (xα ) − f∗ (yα ). (11.16)

Since u, v ∈ C 0,1 (U ), there exists C > 0 such that (see Exercise 2.16)

|pα |, |pα + 2(xα − x0 )| ≤ C

for all α. Since H is uniformly continuous on compact sets we have

lim H(pα ) − H(pα + xα − x0 ) = 0.

α→∞

√  
Setting wα = α xα − yα + √1 ηx
α 0
we have

1
yα = xα + √ (ηx0 − wα ) . (11.17)
α

Since wα → 0 as α → ∞ we can invoke (D) to find that

 
f ∗ (xα ) − f∗ (yα ) ≤ ω |xα − x0 | + √1α

for α sufficiently large. Inserting this into (11.16) and taking α → ∞ yields a
contradiction.
110 CHAPTER 11. DISCONTINUOUS COEFFICIENTS
Chapter 12

Second order equations

We consider in this section the comparison principle for viscosity solutions of

second order equations

F (D2 u, Du, u, x) = 0 in U, (12.1)

where U ⊂ Rn , and F is degenerate elliptic (see (3.4)) and satisfies the usual
monotonicity in u (see (3.3)). Our treatment will loosely follow [4], though
we prefer to avoid the super/sub-jet terminology. A comprehensive reference
on the theory of second order equations with the sharpest results is given the
User’s Guide [6].
We first examine why the method of proof we used for first order equations
(see Theorem 3.1) does not work here. The comparison principle for first order
equations is based on doubling the variables and examining the maximum of
α
Φ(x, y) = u(x) − v(y) − |x − y|2
2
as α → ∞. The key step was identifying that at a maximum (xα , yα ) of Φ, the
smooth function φ(x) := α2 |x − yα |2 touches u from above at xα , and ψ(y) =
− α2 |xα − y|2 touches v from below at yα . Furthermore, we have the magic
property Dφ(xα ) = Dψ(yα ), which replaces the classical identity Du(x) =
Dv(x) at a maximum of u − v when u, v are differentiable. For second order
equations, we also need the identity D2 u(x) ≤ D2 v(x) at the max of u − v.
However, D2 φ(xα ) = αI  −αI = D2 ψ(yα ). So we appear to be at an
impasse.
However, we have not used one important piece of information; namely
that (x, y) 7→ Φ(x, y) is jointly maximal at (xα , yα ). If we, for the moment,
assume u, v ∈ C 2 , then the condition that (xα , yα ) maximize Φ can be written

111
112 CHAPTER 12. SECOND ORDER EQUATIONS

as    
D2 u(xα ) 0 I −I
≤α . (12.2)
0 −D2 v(yα ) −I I
Since the right hand side annihilates vectors of the form (η, η) for η ∈ Rn ,
we find that η T D2 u(xα )η ≤ η T D2 v(yα )η for all η ∈ Rn —that is D2 u(xα ) ≤
D2 v(yα ). So when u, v are sufficiently smooth, the doubling variables argument
contains enough information to utilize the maximum principle for second order
equations.
This suggests performing some regularization of u and v, and then ap-
plying the doubling variables argument to the regularizations. The standard
regularizers in viscosity solutions are the inf- and sup-convolutions, defined in
Chapter 8. We replace the subsolution u with the sup-convolution uε , and the
supersolution v with the inf-convolution vε . Thus, we consider the doubling
variables argument in the form
α
Φε (x, y) := uε (x) − vε (y) − |x − y|2 .
2
ε
The key is that (see Chapter 8) u remains a subsolution (approximately)
and vε remains a supersolution, so we have not lost much by making this
substitution, and we have gained a great deal of regularity. However, to use this
additional regularity, we require a more refined understanding of semiconvex
functions.

12.1 Jensen’s Lemma

In this section we prove Jensen’s Lemma, which gives necessary conditions for
a semiconvex function to attain its maximum value.
Lemma 12.1 (Jensen’s Lemma). Let φ : Rn → R be semiconvex and let x0
be a strict local maximum of φ. For p ∈ Rn set φp (x) = φ(x) − p · (x − x0 ).
Then for r > 0 sufficiently small and all δ > 0 the set
K = {y ∈ B(x0 , r) : ∃p ∈ B(0, δ) such that φp (x) ≤ φp (y) for x ∈ B(x0 , r)}
has positive measure.
Remark 12.2. Note that the condition φp (x) ≤ φp (y) for x ∈ B(x0 , r) is
simply stating that x 7→ φ(x) − p · (x − x0 ) has a local maximum at y, and
so the linear function L(x) = p · (x − x0 ) touches φ from above at y. In other
words, near a strict maximum of a semiconvex function there are a lot of points
with very small gradients, so the function looks in some sense “round” near
its maximum.
12.1. JENSEN’S LEMMA 113

We now turn to the proof of Jensen’s Lemma. The proof requires the
area formula, which is a generalization of the change of variables formula in
Lebesgue integration. Note we write #A to denote the number of points in
A ⊂ Rn and |A| to denote the Lebesgue measure.

Theorem 12.3 (Area formula). Let f : Rn → Rn be a Lipschitz map and set

Jf (x) = |det(Df (x))|. Then for every Lebesgue measurable A ⊂ Rn
Z Z
−1
#(A ∩ f ({x})) dx = Jf (x) dx. (12.3)
f (A) A
R
In particular, if f is injective then |f (A)| = A
Jf (x) dx.

We refer the reader to [12] for a proof of the area formula.

Remark 12.4. Since #(A ∩ f −1 ({x})) ≥ 1 for x ∈ f (A), it follows from the
area formula that Z Z
|f (A)| = dx ≤ Jf (x) dx. (12.4)
f (A) A

This form of the area formula is used in the proof of Jensen’s Lemma.

Proof of Jensen’s Lemma. Let r > 0 be small enough so that φ(x0 ) > φ(x)
for all x ∈ B(x0 , r) with x 6= x0 , and let a > 0 such that φ(x) + a ≤ φ(x0 ) for
all x ∈ ∂B(x0 , r). Let ε > 0 and define the mollification φε = φ ∗ η ε . Then
φε → φ uniformly on B(x0 , r) as ε → 0. Define the corresponding sets

K ε = y ∈ B(x0 , r) : ∃p ∈ B(0, δ) such that φεp (x) ≤ φεp (y) for x ∈ B(x0 , r) ,

where φεp (x) = φε (x) − p · (x − x0 ). Notice that for y ∈ ∂B(x0 , r) we have

φεp (y) − φεp (x0 ) = φεp (y) − φ(y) + φ(y) − φ(x0 ) + φ(x0 ) − φεp (x0 )
≤ 2kφ − φε kL∞ (B(x0 ,r)) + 2|p|r − a.

Therefore, for ε and δ sufficiently small, φεp (y) < φεp (x0 ) for all p with |p| ≤ δ
and all y ∈ ∂B(x0 , r). Thus, every maximum of φεp with respect to B(x0 , r)
lies in the interior B 0 (x0 , r) when |p| ≤ δ. At a maximum y ∈ B 0 (x0 , r) of φεp
we have Dφε (y) = p, and so Dφε (K ε ) ⊃ B(0, δ). For the rest of the proof we
fix δ > 0 sufficiently small, as above.
Now, let λ > 0 such that φ(x) + λ2 |x|2 is convex. This yields

−λI ≤ D2 φε (x) = D2 φεp (x) ≤ 0 for x ∈ K ε .

114 CHAPTER 12. SECOND ORDER EQUATIONS

In particular, it follows that |det(D2 φε (x))| ≤ λn for x ∈ K ε and so by the

area formula (Theorem 12.3) we have
Z
|B(0, δ)| ≤ |Dφ(K )| ≤
ε
|detD2 φ(x)| dx ≤ |K ε |λn .
Kε

Therefore
α(n)δ n
|K ε | ≥ . (12.5)
λn
Since φε → φ uniformly, if x ∈ K εj for a sequence εj → 0 then x ∈ K.
Therefore
χK (x) ≥ lim sup χK m1 (x),
m→∞
where χA is the indicator function of the set A. By Fatou’s Lemma
Z Z
α(n)δ n
|K| = dx ≥ lim sup dx ≥ ,
K m→∞ Km
1
λn
which completes the proof.
The following proposition illustrates the usefulness of Jensen’s Lemma in
establishing the maximum principle for semiconvex functions.
Proposition 12.5. Let φ : Rn → R be semiconvex and let x0 be a local
maximum of φ. Then there exists xk → x0 such that φ is twice differentiable
at xk , Dφ(xk ) → 0 as k → ∞ and D2 φ(xk ) ≤ εk I for a sequence εk → 0.
Remark 12.6. Proposition 12.5, which is a restatement of Jensen’s Lemma,
is the semiconvex analog of the condition that Dφ = 0 and D2 φ ≤ 0 at the
maximum of a C 2 function.
Proof. Define ψ(x) = φ(x) − |x − x0 |4 . Then ψ has a strict local max at x0 ,
and ψ is semiconvex. Let rk > 0 be a decreasing sequence of real numbers con-
verging to zero. By Lemma 12.1 (Jensen’s Lemma), there is a corresponding
decreasing sequence δk > 0 such that δk → 0 and
{y ∈ B(x0 , rk ) : ∃p ∈ B(0, δk ) such that ψp (x) ≤ ψp (y) for x ∈ B(x0 , rk )}
has positive measure, where ψp (x) = ψ(x) + p · (x − x0 ). Since ψ is twice
differentiable almost everywhere, there exists xk ∈ B 0 (x0 , rk ) and pk ∈ B(0, δk )
such that ψpk has a local maximum at xk and ψ is twice differentiable at xk .
Hence xk → x0 , pk → 0, Dψ(xk ) = pk and D2 ψ(xk ) = D2 ψpk (xk ) ≤ 0.
The proof is completed by noting that φ is also twice differentiable along the
sequence xk , and that
|Dφ(xk ) − pk | ≤ 4|xk − x0 |3 and D2 φ(xk ) ≤ 12|xk − x0 |2 I.
12.2. COMPARISON FOR CONTINUOUS FUNCTIONS 115

12.2 Comparison for continuous functions

We now use Jensen’s lemma to prove comparison principles for second order
equations of the form (12.1) for continuous sub- and supersolutions.
We assume throughout this section that U ⊂ Rn is open and bounded,
and F is continuous, degenerate elliptic (i.e., satisfies (3.4)), and satisfies the
monotonicity condition (3.3). We also assume F satisfies
F (X, p, z, y) − F (X, p, z, x) ≤ ω(|x − y|(1 + |p|)) (12.6)
for all x, y ∈ U , z ∈ R, p ∈ Rn , and symmetric matrices X, where ω is a
modulus of continuity.
We first prove comparison when the subsolution is semiconvex, and the
supersolution is semiconcave.
Lemma 12.7 (Semiconvex comparison). Let u ∈ C(U ) be a semiconvex vis-
cosity subsolution of (12.1), and let v ∈ C(U ) be a semiconcave viscosity
solution of
F (D2 v, Dv, v, x) − δ ≥ 0 in U,
for some δ > 0. If u ≤ v on ∂U then u ≤ v in U .
Proof. We use doubling the variables and Jensen’s lemma. Assume to the
contrary that maxU (u − v) > 0. Define the auxiliary function
α
Φ(x, y) = u(x) − v(y) − |x − y|2 , (12.7)
2
and let xα , yα ∈ U such that
Φ(xα , yα ) = max Φ.
U ×U

As in the proof of Theorem 3.1 we have xα , yα ∈ U for α large enough and

α|xα − yα |2 −→ 0. (12.8)
Since u and −v are semiconvex, the auxiliary function Φ : U × U → R
is semiconvex. By Proposition 12.5 there exists a sequence (xkα , yαk ) ∈ U × U
such that xkα → xα and yαk → yα as k → ∞, Φ is twice differentiable at
(xkα , yαk ), Dxy Φ(xkα , yαk ) → 0 as k → ∞, and Dxy 2
Φ(xkα , yαk ) ≤ εk I for a sequence
εk → 0. Here, Dxy denotes the gradient jointly in (x, y). It follows that u is
twice differentiable at xkα and v is twice differentiable at yαk . Furthermore, for
pα = α(xα − yα ) we have Du(xkα ) → pα and Dv(xkα ) → pα as k → ∞, and
 2   
D u(xkα ) 0 I −I
−CI ≤ ≤α + εk I. (12.9)
0 −D2 v(yαk ) −I I
116 CHAPTER 12. SECOND ORDER EQUATIONS

The lower bound in (12.9) follows from semiconvexity of u and −v, while the
upper bound follows from Dxy 2
Φ(xkα , yαk ) ≤ εk I. By conjugating both sides with
vectors of the form (η, η) ∈ R2n we have
η T D2 u(xkα )η ≤ η T D2 v(xkα )η + 2εk |η|2 ,
for all η ∈ Rn , and hence D2 u(xkα ) ≤ D2 v(xkα ) + 2εk I. Using (12.9), we can,
upon passing to a subsequence, assume that D2 u(xkα ) → Xα and D2 v(yαk ) → Yα
as k → ∞, where Xα ≤ Yα .
By the viscosity sub- and supersolution properties, and Remark 2.7, we
have
F (D2 u(xkα ), Du(xkα ), u(xkα ), xkα ) ≤ 0 (12.10)
and
F (D2 v(yαk ), Dv(yαk ), v(yαk ), yαk ) ≥ δ. (12.11)
Taking k → ∞ and using continuity of F , u, and v we have
F (Xα , pα , u(xα ), xα ) ≤ 0
and
F (Yα , pα , v(yα ), yα ) ≥ a,
where Xα ≤ Yα . Since Φ(xα , yα ) ≥ maxU (u − v) > 0 we have u(xα ) > v(yα ),
and so by monotonicity and degenerate ellipticity, we have
δ ≤ F (Yα , pα , v(yα ), yα ) ≤ F (Xα , pα , u(xα ), yα ).
Applying (12.6) and (12.10) we find that
δ ≤ ω((1 + |pα |)|xα − yα |).
Sending α → ∞ and recalling (12.8) yields a contradiction.
Using the inf- and sup-convolutions, we can extend the semiconvex com-
parison principle to continuous functions. Here, we assume F has the form
F (X, p, z, x) = λz + H(X, p) − f (x), (12.12)
where λ ≥ 0. Then the regularity condition (12.6) is equivalent to the condi-
tion f ∈ C(U ).
Theorem 12.8 (Continuous comparison). Assume F has the form (12.12).
Let u ∈ C(U ) be a viscosity subsolution of (12.1), and let v ∈ C(U ) be a
viscosity solution of
F (D2 v, Dv, v, x) − δ ≥ 0 in U,
for some δ > 0. If u ≤ v on ∂U then u ≤ v in U .
12.3. SUPERJETS AND SUBJETS 117

Proof. For ε > 0, let uε be the sup-convolution of u, and let vε be the inf-
convolution of v, as defined in Chapter 8. By an argument similar to Corollary
8.9, we have that
λuε + H(D2 uε , Duε ) − f ≤ g(ε) in Uε
and
λvε + H(D2 vε , Dvε ) − f ≥ δ − h(ε) in Uε
hold in the viscosity sense, where
Uε = {x ∈ U : dist(x, ∂U ) ≥ Cε}
for a constant C > 0, and g, h are nonnegative continuous functions with
g(0) = h(0) = 0. Let mε = supU \Uε (uε − vε ). Since u, v ∈ C(U ), uε → u
and vε → v uniformly, and u ≤ v on ∂U , we have mε → 0 as ε → 0. Define
wε = uε − mε . Then wε satisfies
λwε + H(D2 wε , Dwε ) − f ≤ g(ε) − λmε in Uε
in the viscosity sense, and wε ≤ vε on ∂Uε . For ε > 0 sufficiently small, we
can apply Lemma 12.7 to show that wε ≤ vε on Uε , and so
uε ≤ v ε + m ε on Uε .
Sending ε → 0 we recover u ≤ v on U .

12.3 Superjets and subjets

There is a common alternative definition of viscosity solutions that is worth
discussing briefly. We first make a definition.
Definition 12.9. Let O ⊂ Rn , u : O → R, and x0 ∈ O. The superjet
JO2,+ u(x0 ) is defined as the set of all (p, X) ∈ Rn × S(n) for which
1
u(x) ≤ u(x0 ) + p · (x − x0 ) + (x − x0 )T X(x − x0 ) + o(|x − x0 |2 )
2
as O 3 x → x0 .
Similarly, the subjet JO2,− u(x0 ) is defined as the set of all (p, X) ∈ Rn ×S(n)
for which
1
u(x) ≥ u(x0 ) + p · (x − x0 ) + (x − x0 )T X(x − x0 ) + o(|x − x0 |2 )
2
as O 3 x → x0 .
118 CHAPTER 12. SECOND ORDER EQUATIONS

The following proposition is immediate.

Proposition 12.10. Let u : U → R where U ⊂ Rn is open. If u : O → R is
twice differentiable at x ∈ U then

J 2,+ u(x) ∩ J 2,− u(x) = (Du(x), D2 u(x)).

The relationship between superjets and subjets and viscosity solutions is

illuminated by the following result.
Proposition 12.11. Let u : U → R where U ⊂ Rn is open. We have

J 2,+ u(x0 ) = (Dφ(x0 ), D2 φ(x0 )) : φ ∈ C 2 (Rn ) and u − φ has a local max at x0 .

and

J 2,− u(x0 ) = (Dφ(x0 ), D2 φ(x0 )) : φ ∈ C 2 (Rn ) and u − φ has a local min at x0 .

Proof. If u − φ has a local maximum at x0 , then clearly (Dφ(x0 ), D2 φ(x0 )) ∈

J 2,+ u(x0 ).
We now prove the converse. We may take x0 = 0 and u(0) = 0 for simplic-
ity. Let (p, X) ∈ J 2,+ u(0). By definition, for some r0 > 0 we have
1
u(x) ≤ p · x + xT Xx + g(x)|x|2
2
for |x| ≤ r0 , where g : U → R is continuous and g(0) = 0. Define

ρ(r) = max |g(x)|.

|x|≤r

Then ρ : [0, r0 ] → [0, ∞) is continuous and nondecreasing with ρ(0) = 0 and

g(x) ≤ ρ(|x|). Therefore
1
u(x) ≤ p · x + xT Xx + ρ(|x|)|x|2
2
for |x| ≤ r0 . Define the C 2 function
Z rZ s
σ(r) = ρ(t) dt ds,
0 0

and note that

Z 3r Z s Z 3r
σ(3r) ≥ ρ(t) dt ds ≥ (s − r)ρ(r) ds ≥ 2r2 ρ(r).
r r r
12.3. SUPERJETS AND SUBJETS 119

Therefore, for |x| ≤ r0 we have

1
u(x) ≤ p · x + xT Xx + σ(3|x|) =: φ(x).
2
Notice that φ ∈ C 2 (B(0, r0 )), u − φ has a local max at x = 0, p = Dφ(0)
and X = D2 φ(0). We can extend φ to a function φ ∈ C 2 (Rn ) with a bump
function argument. The proof for the superjet is similar.
We now give an alternative characterization of viscosity solutions of (12.1)
in terms of superjets and subjets.
Theorem 12.12. Let U ⊂ Rn be open and assume F is continuous in all
variables. If u ∈ USC(U ) is a viscosity subsolution of (12.1) then

F (X, p, u(x), x) ≤ 0 for all x ∈ U and (p, X) ∈ J 2,+ u(x). (12.13)

Similarly, if v ∈ LSC(U ) is a viscosity supersolution (12.1) then

F (X, p, v(x), x) ≥ 0 for all x ∈ U and (p, X) ∈ J 2,− v(x). (12.14)

Remark 12.13. The conditions (12.13) and (12.14) are sometimes given as
the definitions of viscosity solutions (see, e.g., [6]). While this notation may
seem convenient and compact, nobody quite likes this “jet” business [4].
Proof. Let u ∈ USC(U ) be a viscosity subsolution of (12.1) and let x0 ∈ U
and (p, X) ∈ J 2,+ u(x). By Proposition 12.11 there exists φ ∈ C 2 (Rn ) such
that u − φ has a strict local maximum at x0 , p = Dφ(x0 ) and X = D2 φ(x0 ).
Define the standard mollification φε = η ε ∗ φ. Since φε → φ locally uniformly,
there exists εk → 0 and xk → x0 such that u(xk ) → u(x0 ) and u − φεk has a
local max at xk for each k ≥ 1. Since φ ∈ C ∞ (Rn ), the viscosity subsolution
property yields

F (D2 φεk (xk ), Dφεk (xk ), u(xk ), xk ) ≤ 0 (k ≥ 1).

Since φ ∈ C 2 (Rn ), we have Dφε → Dφ and D2 φε → D2 φ locally uniformly

as ε → 0. Since F is continuous and u(xk ) → u(x0 ), we can send k → ∞ to
obtain
F (D2 φ(x0 ), Dφ(x0 ), u(x0 ), x0 ) ≤ 0.
Since p = Dφ(x0 ) and X = D2 φ(x0 ) we have

F (X, p, u(x0 ), x0 ) ≤ 0,

which completes the proof. The proof for the superjet is similar.
120 CHAPTER 12. SECOND ORDER EQUATIONS

As a consequence of Theorem 12.12, we can prove consistency of viscosity

solutions with classical solutions under minimal regularity assumptions.
Corollary 12.14. Let U ⊂ Rn be open and assume F is continuous in all
variables. If u ∈ USC(U ) is a viscosity subsolution of (12.1) and u is twice
differentiable at some x ∈ U then
F (D2 u(x), Du(x), u(x), x) ≤ 0.
Similarly, if v ∈ LSC(U ) is a viscosity supersolution of (12.1) and v is
twice differentiable at some x ∈ U then
F (D2 v(x), Dv(x), v(x), x) ≥ 0.
Proof. The proof follows directly from Proposition 12.10 and Theorem 12.12,
since (Du(x), D2 u(x)) ∈ J 2,+ u(x) and (Dv(x), D2 v(x)) ∈ J 2,− v(x).

12.4 Semicontinuous comparison

In order to use the Perron method (Chapter 7), or the weak upper and lower
limits (Chapters 5 and 9), we require a comparison principle for semicontinuous
sub- and supersolutions, which is somewhat more involved than the continuous
comparison principle given in Section 12.2. Often, the comparison principle in
its full semicontinuous glory is proved using the superjet and subjet definitions
of viscosity solutions introduced in Section 12.3. Good references for this
include [4, 6].
Here, we give a proof that avoids the notion of superjets and subjets, yet
is sharp in its generality. The key technical point is that the inf- and sup-
convolutions must be done within the doubling variables argument, and in
particular must be applied to the penalty term α2 |x − y|2 as well, instead of
performing them separately as was done in Theorem 12.8. Furthermore, we do
not send ε → 0 in the sup-convolution, so we avoid the issue that uε does not
converge uniformly to u in the semi-continuous case. The proof of Theorem
12.15 is borrowed partly from [4], with appropriate translations to use test
functions instead of superjets and subjets.
Theorem 12.15. Assume U ⊂ Rn is open and bounded, and that F is con-
tinuous, and satisfies (3.3), (3.4), and (12.6). Let u ∈ USC(U ) be a viscosity
subsolution of (12.1), and let v ∈ LSC(U ) be a viscosity solution of
F (D2 v, Dv, v, x) − δ ≥ 0 in U,
for some δ > 0. If u ≤ v on ∂U then u ≤ v in U .
12.4. SEMICONTINUOUS COMPARISON 121

Proof. We assume, by way of contradiction, that maxU (u − v) > 0. For α > 0

define
α
Φ(x, y) = u(x) − v(y) − |x − y|2 , (12.15)
2
and let xα , yα ∈ U such that

Φ(xα , yα ) = max Φ.
U ×U

As in the proof of Theorem 3.1 we have xα , yα ∈ U for α large enough and

α|xα − yα |2 −→ 0. (12.16)

The proof is split into 3 steps.

1. We first make a reduction. Define

f (x) = u(x + xα ) − u(xα ) − αx · (xα − yα ),

and
g(y) = v(y + yα ) − v(yα ) − αy · (xα − yα ).
Then we have
α
f (x) − g(y) − |x − y|2
2
α
= u(x + xα ) − v(y + yα ) − |x − y|2 − α(x − y) · (xα − yα ) + v(yα ) − u(xα )
2
α α
= u(x + xα ) − v(y + yα ) − |x − y + xα − yα |2 + |xα − yα |2 + v(yα ) − u(xα )
2 2
α
= Φ(x + xα , y + yα ) + |xα − yα | + v(yα ) − u(xα ).
2
2
Therefore
α
f (x) − g(y) − |x − y|2
2
attains its maximum at (x, y) = (0, 0), and f (0) = g(0) = 0. Thus, we have
α
f (x) − g(y) ≤ |x − y|2 (12.17)
2
for x, y near 0.
2. We now take the sup-convolution on both sides of (12.17) jointly in
(x, y) to obtain (see Exercise 12.16)
α
f ε (x) − gε (y) ≤ (1 − 2αε)−1 |x − y|2 (12.18)
2
122 CHAPTER 12. SECOND ORDER EQUATIONS

for x, y near 0. Since f ε ≥ f and gε ≤ g we have

f ε (0) − gε (0) ≥ f (0) − g(0) = 0

and so by (12.18) we have f ε (0) = gε (0). Since f ε (0) ≥ f (0) = 0 and gε (0) ≤
g(0) = 0 we have f ε (0) = 0 = gε (0). Therefore the function
α
f ε (x) − gε (y) − (1 − 2αε)−1 |x − y|2 − |x|4 − |y|4
2
attains a strict local maximum at (x, y) = (0, 0). By Lemma 12.1 (Jensen’s
Lemma), there exists xk , y k → 0 and ξ k , ζ k → 0 such that
α
f ε (x) − gε (y) − (1 − 2αε)−1 |x − y|2 − |x|4 − |y|4 − ξ k · x − ζ k · y
2

has a local maximum at (xk , y k ), f ε is twice differentiable at xk , and gε is twice

differentiable at y k . Let rk = max{|xk |2 , |y k |2 }. The first order conditions for
a maximum yield

pk := Df ε (xk ) = (1 − 2αε)−1 α(xk − y k ) + 4|xk |2 xk + ξ k , (12.19)

and

q k := Dgε (y k ) = (1 − 2αε)−1 α(xk − y k ) − 4|y k |2 y k + ζ k . (12.20)

Note that pk , q k → 0 as k → ∞. The second order condition for a maximum,

and semiconvexity, yield
 k   
1 X 0 −1 I −I
− I≤ k ≤ (1 − 2αε) α + 12rk2 I, (12.21)
ε 0 −Y −I I

where X k = D2 f ε (xk ) and Y k = D2 gε (y k ). In particular, for all η ∈ Rn we

have
η T X k η ≤ η T Y k η + 12rk2 |η|2 ,
and so X k ≤ Y k + 12rk2 I. Passing to a subsequence, if necessary, there exists
Xα , Yα ∈ S(n) such that X k → Xα , Y k → Yα , and Xα ≤ Yα .
3. By Proposition 12.11, there exists φ ∈ C 2 (Rn ) such that f ε − φ has a
local max at xk and D2 φ(xk ) = X k . Let xkε ∈ U such that

1 k
f ε (xk ) = f (xkε ) − |x − xkε |2 . (12.22)
2ε
12.4. SEMICONTINUOUS COMPARISON 123

By Proposition 8.6 f (x) − φε (x) has a local maximum at xkε , where φε (x) :=
φ(x + xk − xkε ), and we have
1
Dφε (xkε ) = pk = (xkε − xk ),
ε
and D φε (xε ) = X . Since p → 0 as k → ∞ we have xkε → 0 as k → ∞. In
2 k k k

particular, from (12.22) and the continuity of f ε we have

f (xkε ) −→ f ε (0) = f (0) = 0 (12.23)
as k → ∞. Since f − φε has a local max at xkε we see that
x 7→ u(x) − α(x − xα ) · (xα − yα ) − φε (x − xα )
has a local max at xkε + xα . Therefore, setting pα = α(xα − yα ) we have
F (X k , pα + pk , u(xkε + xα ), xkε + xα ) ≤ 0.
Since f (xkε ) → 0 as k → ∞ we have u(xkε + xα ) → u(xα ). Thus, we can take
k → ∞ above and recall F is continuous to obtain
F (Xα , pα , u(xα ), xα ) ≤ 0. (12.24)
We can make a similar argument for gε to obtain
F (Yα , pα , v(yα ), yα ) ≥ δ. (12.25)
Since Φ(xα , yα ) > 0 we have u(xα ) > v(yα ). Subtracting (12.24) from (12.25)
and using (3.3), (3.4) and (12.6) we have
δ ≤ F (Yα , pα , v(yα ), yα ) − F (Xα , pα , u(xα ), xα )
≤ F (Xα , pα , u(xα ), yα ) − F (Xα , pα , u(xα ), xα )
≤ ω(|xα − yα |(1 + |pα |)).
Sending α → ∞ yields a contradiction, due to (12.16).
Exercise 12.16. Define
α
w(x, y) = |x − y|2 .
2
Show that the sup-convolution
 
α ′ ′ 2 1 ′ 2 1 ′ 2
ε
w (x, y) = sup |x − y | − |x − x | − |y − y |
(x′ ,y ′ )∈Rn ×Rn 2 2ε 2ε
is given by
α
wε (x, y) = (1 − 2αε)−1 |x − y|2 ,
2
provided 1 − 2αε 6= 0.
124 CHAPTER 12. SECOND ORDER EQUATIONS

12.5 A problem on an unbounded domain

In this section we give a brief application of the comparison principle from
Section 12.4. In particular, we study the equation

u + F (D2 u) = f on Rn . (12.26)

The well-posedness theory is far more general; we study this simple problem to
illustrate the main ideas, and to show how to handle the unbounded domain.
Throughout this section, we assume F is uniformly continuous and degenerate
elliptic, and f is continuous.
We first prove a comparison principle.

Lemma 12.17. Let u ∈ USC(Rn ) be a viscosity subsolution of (12.26) and

let v ∈ LSC(Rn ) be a viscosity supersolution of (12.26). If

u(x) − v(x)
lim =0 (12.27)
|x|→∞ |x|2

then u ≤ v on Rn .

Proof. For ε > 0 define

1
uε (x) = u(x) − ε|x|2 − cε ,
2
for cε > 0 to be determined. Then uε is a viscosity solution of
1
uε + F (D2 uε + εI) ≤ f − ε|x|2 − cε on Rn .
2
Since F is uniformly continuous, we can choose cε → 0 as ε → 0 so that
1
uε + F (D2 uε ) ≤ f − ε|x|2 − ε on Rn
2
in the viscosity sense. Note that

uε (x) − v(x) u(x) − v(x) 1

≤ − ε.
|x| 2 |x|2 2

Thus, by (12.27) there exists rε > 0 such that uε ≤ v for |x| ≥ rε . By Theorem
12.15 we have uε ≤ v on B(0, r) for all r > rε , and thus uε ≤ v on Rn . Sending
ε → 0 completes the proof.
12.5. A PROBLEM ON AN UNBOUNDED DOMAIN 125

By Lemma 12.17 there is at most one viscosity solution of (12.26) satisfying

the growth bound
u(x)
lim =0 (12.28)
|x|→∞ |x|2

We now turn to proving existence of a solution, via the Perron method.

Here, we look for a solution with at most linear growth at infinity.
Theorem 12.18. Suppose there exists Cf > 0 so that

|f (x)| ≤ Cf (1 + |x|). (12.29)

Then there exists a unique viscosity solution u ∈ C(Rn ) of (12.26) satisfying

lim|x|→∞ u(x)
|x|2
= 0. Furthermore, we have

|u(x)| ≤ C(1 + |x|) (12.30)

for some C > 0.

Proof. Let ψ ∈ C ∞ (Rn ) such that ψ(x) = |x| for |x| ≥ 1, and ψ ≥ 0. Define
w(x) = K1 + K2 ψ(x), for K1 , K2 to be determined. Note that
1 x⊗x
D2 ψ(x) = I− for |x| ≥ 1.
|x| |x|3

Therefore |D2 ψ| is bounded on Rn and

  
F (K2 D2 ψ) 1
w + F (D w) − f ≥ |x| K2 +
2
− Cf 1 +
|x| |x|
for |x| ≥ 1. Choose K2 = 2Cf . Then there exists r > 1 such that for |x| > r
we have
1
w + F (D2 w) − f ≥ Cf |x| > 0.
2
Now, for |x| ≤ r we have

w + F (D2 w) − f ≥ K1 + F (K2 D2 ψ) − f (x) > 0

for K1 > 0 sufficiently large. Therefore, for large enough K1 , K2 , w is a

supersolution of (12.26) on Rn and

|w(x)| ≤ C(1 + |x|). (12.31)

We also note that the same construction yields that v := −w is a subsolution

of (12.26).
126 CHAPTER 12. SECOND ORDER EQUATIONS

Now, define

F = {v ∈ USC(Rn ) : v is a subsolution of (12.26) and v ≤ w}, (12.32)

and
u(x) := sup{v(x) : v ∈ F}. (12.33)
Since v = −w ∈ F, we have −w ≤ u ≤ w, and so u satisfies (12.30). By
Lemma 7.1 u∗ is a viscosity subsolution of (12.26). Since u ≤ w and w is
continuous, we have u∗ ≤ w and so u∗ ∈ F and u = u∗ . By Lemma 7.2, u∗
is a viscosity supersolution of (12.26). Since u satisfies (12.30), we can invoke
Lemma 12.17 to obtain u∗ ≤ u∗ , and so u = u∗ = u∗ is the unique viscosity
solution of (12.26).

Exercise 12.19. Modify the proof of Theorem 12.18 to hold under the con-
dition
|f (x)| ≤ Cf (1 + |x|α ) (12.34)
for some Cf > 0 and α < 2.
Bibliography

[1] M. Bardi and I. Capuzzo-Dolcetta. Optimal control and viscosity solutions

of Hamilton-Jacobi-Bellman equations. Springer Science & Business Me-
dia, 2008.

[2] G. Barles and P. E. Souganidis. Convergence of approximation schemes

for fully nonlinear second order equations. Asymptotic analysis, 4(3):271–
283, 1991.

[3] A. Beck. First-Order Methods in Optimization, volume 25. SIAM, 2017.

[4] M. G. Crandall. Viscosity solutions: a primer. In Viscosity solutions and

applications, pages 1–43. Springer, 1997.

[5] M. G. Crandall, L. C. Evans, and P.-L. Lions. Some properties of viscosity

solutions of Hamilton-Jacobi equations. Transactions of the American
Mathematical Society, 282(2):487–502, 1984.

[6] M. G. Crandall, H. Ishii, and P.-L. Lions. User’s guide to viscosity solu-
tions of second order partial differential equations. Bulletin of the Amer-
ican Mathematical Society, 27(1):1–67, 1992.

[7] M. G. Crandall and P.-L. Lions. Viscosity solutions of Hamilton-Jacobi

equations. Transactions of the American Mathematical Society, 277(1):1–
42, 1983.

[8] K. Deckelnick and C. M. Elliott. Uniqueness and error analysis for

hamilton-jacobi equations with discontinuities. Interfaces and free bound-
aries, 6(3):329–349, 2004.

[9] L. C. Evans. The perturbed test function method for viscosity solutions
of nonlinear PDE. Proceedings of the Royal Society of Edinburgh: Section
A Mathematics, 111(3-4):359–375, 1989.

127
128 BIBLIOGRAPHY

[10] L. C. Evans. Periodic homogenisation of certain fully nonlinear partial

differential equations. Proceedings of the Royal Society of Edinburgh: Sec-
tion A Mathematics, 120(3-4):245–265, 1992.

[11] L. C. Evans. Partial Differential Equations, volume 19 of Graduate Studies

in Mathematics. AMS, Providence, Rhode Island, 1998.

[12] L. C. Evans and R. Gariety. Measure theory and fine properties of func-
tions. Studies in Advanced Mathematics. CRC Press. Boca Raton London
New York Washington, DC, 1992.

[13] R. Jensen. The maximum principle for viscosity solutions of fully non-
linear second order partial differential equations. Archive for Rational
Mechanics and Analysis, 101(1):1–27, 1988.

[14] N. Katzourakis. An introduction to Viscosity Solutions for fully nonlinear

PDE with applications to Calculus of Variations in Linfinity. Springer,
2014.

[15] S. Osher and C.-W. Shu. High-order essentially nonoscillatory schemes

for Hamilton-Jacobi equations. SIAM Journal on numerical analysis,
28(4):907–922, 1991.

[16] E. Rouy and A. Tourin. A viscosity solutions approach to shape-from-

shading. SIAM Journal on Numerical Analysis, 29(3):867–884, 1992.