Adjoint methods for the innity Laplacian PDE

Lawrence C. Evans

and Charles K. Smart

Department of Mathematics
University of California, Berkeley
Abstract
To study ne properties of certain smooth approximations u

to a viscosity solution
u of the innity Laplacian PDE, we introduce Greens function

for the linearization.

We can then integrate by parts with respect to

and derive various useful integral

estimates.
We are in particular able to use these estimates (i) to prove the everywhere dier-
entiability of u and (ii) to rigorously justify interpreting the innity Laplacian equation
as a parabolic PDE.
1 Introduction
1.1 Basic equations. In this paper we consider the boundary value problem
(1.1)
_

u = 0 in U
u = g on U,
where U R
n
is an open set, g : U R is Lipschitz continuous, and we write

u := u
x
i
u
x
j
u
x
i
x
j
for the degenerate nonlinear innity-Laplacian operator. Since the unique viscosity solution
of (1.1) need not be smooth, we will study also the regularization:
(1.2)
_

= 0 in U
u

= g on U.

Supported in part by NSF grants DMS-0500452 and DMS-1001724

Supported in part by NSF grant DMS-1004595

1
Assume now that V U is a compactly contained open subset, with smooth boundary.
For each point x
0
V we introduce also this linear problem:
(1.3)
_
(u

x
i
u

x
j

)
x
i
x
j
+ 2(u

x
i
u

x
i
x
j

)
x
j

=
x
0 in V

= 0 on V .
Here
x
0 denotes the Dirac measure at x
0
.
Notation. We will write
(1.4) L

v := u

x
i
u

x
j
v
x
i
x
j
2u

x
i
u

x
j
x
j
v
x
j
v
for the linearization of (1.2), and
(1.5) L

w := (u

x
i
u

x
j
w)
x
i
x
j
+ 2(u

x
i
u

x
j
x
j
w)
x
i
w
for its adjoint. Thus (1.3) says
(1.6) L

=
x
0 in V ,
and so

is Greens function for the linear elliptic operator L

.
We will employ the functions

to extract information about the limiting behavior of u

as 0 and thus about the solution u of (1.1). The main new advances are a proof that u
is everywhere dierentiable and a rigorous interpretation of the innity Laplace equation as
a parabolic PDE, at least generically. Our companion paper [E-S] provides a simpler proof
of the everywhere dierentiability, employing only the maximum principle. This alternative
proof was inspired by the adjoint methods set forth here, which however provide much more
detailed information, as we will see.
Introducing the adjoint PDE (1.3) is inspired by the rst authors recent paper [E1] on
nonconvex Hamilton-Jacobi equations and also by various techniques for the PDE approach
to weak KAM theory (see [E2]). Savin [S] proved for n = 2 dimensions that the viscosity
solution u of (1.1) is in fact C
1
.
2 Solving the approximating PDE
2.1 Estimates for u

. We record some rst bounds, uniform in , proved in our other

paper [E-S]:
THEOREM 2.1 (i) There exists a unique solution u

of the (1.2), smooth on

U. Further-
more, we have the estimates
(2.1) max

U
|u

| C,
2
and for each open set V U
(2.2) max

V
|Du

| C.
Both constants are independent of and the constant in (2.3) depends upon dist(V, U) > 0.
(ii) We have
(2.3) u

u locally uniformly on

U,
where u is the unique viscosity solution of the boundary value problem (1.1).
2.2 The adjoint problem.
THEOREM 2.2 There exists a unique solution

of the linear adjoint problem (1.3),

smooth on

V {x
0
}. Furthermore,
(2.4)

0 in V .
Proof. 1. According to the maximum principle, the only solution of
_
L

v = 0 in V
v = 0 on V
is v 0. Thus 0 is not an eigenvalue of the operator L

and is consequently not an eigenvalue

of L

. The existence of Greens function

solving (1.3) follows from standard linear elliptic

PDE theory, and

is smooth away from the singularity at x

0
.
2. Given a smooth, nonnegative function f, we introduce the solution w

of the linear
boundary value problem
(2.5)
_
L

= 0 in V
w

= 0 on V .
Owing to the maximum principle, w

0. We multiply the PDE in (2.5) by

and integrate
by parts:
_
V
f

dx = w

(x
0
) 0.
This inequality is valid for all smooth f 0 and consequently

0.
3
3 Integral identities, more estimates
3.1 A rst integral identity. The following integral estimate will be useful later:
THEOREM 3.1 (i) For each smooth function : R
n
R we have the identity
(3.1) (Du

(x
0
)) +
_
V

p
k
p
l
(Du

)(u

x
i
u

x
i
x
k
u

x
j
u

x
j
x
l
+ u

x
i
x
l
u

x
i
x
k
)

dx =
_
V
(Du

dS
for
(3.2)

:=
_
_
u

_
2
+
_
|D

|.
(ii) In particular,
(3.3)
_
V

dS = 1
and
(3.4) Du

(x
0
) =
_
V
Du

dS.
(iii) We have this estimate for the second derivatives of u

:
(3.5)
_
V
(|D
2
u

Du

|
2
+ |D
2
u

|
2
)

dx C,
the constant C independent of .
Observe that the density

x
0
depends upon

and thus upon our choice of the point

x
0
V . Also, take note that although (3.4) resembles a linear representation formula for
Du

(x
0
), in fact

depends in a highly nonlinear and nonlocal way upon Du

and D
2
u

.
Proof. 1. Dierentiate the PDE (1.2) with respect to x
k
:
L

x
k
= u

x
i
u

x
j
u

x
k
x
i
x
j
2u

x
i
u

x
i
x
j
u

x
k
x
j
u

x
k
= 0.
Multiply by
p
k
(Du

), sum on k and rewrite, to discover that

L

= u

x
i
u

x
j

x
i
x
j
2u

x
i
u

x
i
x
j

x
j
=
p
k
p
l
(u

x
i
u

x
i
x
k
u

x
j
u

x
j
x
l
+ u

x
i
x
l
u

x
i
x
k
),
where = (Du

). We next multiply by

and integrate by parts twice. Recalling that

= 0 on V and remembering the PDE (1.3), we discover that

(Du

(x
0
)) +
_
V

p
k
p
l
(Du

)(u

x
i
u

x
i
x
k
u

x
j
u

x
j
x
l
+ u

x
i
x
l
u

x
i
x
k
)

dx
=
_
V
(Du

)((u

x
i
u

x
j

)
x
i
+

x
j
)
j
dS,
4
where = (
1
, . . . ,
n
) denotes the outward pointing unit normal along V . Again noting

= 0 on V , we observe that
_
V
((u

x
i
u

x
j

)
x
i
+

x
j
)
j
dS =
_
V
(u

x
i
u

x
j

x
i
+

x
j
)
j
dS =
_
V

dS,
since =
D

|D

|
. This proves (3.1).
2. The formulas (3.3) and (3.4) are special cases of (3.1), corresponding to 1 and
= p
k
(k = 1, . . . , n). The estimate (3.5) follows from the choice = |p|
2
and from (3.3).

3.2 A rst estimate on the L

1
norm of

. As an application of (3.3) and (3.5), we

derive a rough estimate on the integral of

:
THEOREM 3.2 There exists a constant C such that
(3.6)
_
V

dx
C

2
.
In general we do not have an L
1
bound for

that is independent of . For example,

if u = u

0, then

is 1/ times Greens function for the Laplacian, in which case

||

||
L
1 = O(1/). See the later Theorem 3.5 for a more rened estimate.
Proof. Let v := |x|
2
. Then according to (1.4),
L

v := u

x
i
u

x
j
v
x
i
x
j
2u

x
i
u

x
i
x
j
v
x
j
v = 2(|Du

|
2
+ n) 4u

x
i
u

x
i
x
j
x
j
.
Therefore
|x
0
|
2
=
_
V
vL

dx
=
_
V
L

dx +
_
V
|x|
2

dS
= 2
_
V
(|Du

|
2
+ n)

dx 4
_
V
u

x
i
u

x
i
x
j
x
j

dx +
_
V
|x|
2

dS.
Rearranging, we deduce that
_
V
(2|Du

|
2
+ 2n)

dx C + C
_
U
|D
2
u

Du

dx
C +
C

_
V
|D
2
u

Du

|
2

dx + n
_
V

dx

+ n
_
V

dx,
5
according to (3.3) and (3.5). This gives (3.6).
3.3 An exponential estimate. It is clear that when is convex, the second term on the
left hand side of (3.1) is nonnegative. One of our main observations is that this identity can
provide useful information for certain nonconvex functions , namely those of the form
(3.7) (p) = (|p|
2
)
where : R R is possibly decreasing. We write = (q).
THEOREM 3.3 (i) For each smooth we have the identity
(3.8)
(|Du

(x
0
)|
2
) + 2
_
V

(|Du

|
2
)(|D
2
u

Du

|
2
+ |D
2
u

|
2
)

dx
=
_
V
(|Du

|
2
)

dS 4
_
V

(|Du

|
2
)((u

)
2
+ |D
2
u

Du

|
2
)

dx.
(ii) There exists a constant > 0 for which
(3.9)
_
V
e
(
2

|Du

|
2
)

dS +
_
V
e
(
2

|Du

|
2
)

(|D
2
u

Du

|
2
+ |D
2
u

|
2
)

dx C,
where
(3.10)

:= |Du

(x
0
)|.
Notice that

occurs only within the last term in (3.8), and that this expression is O(),
according to (3.5).
We will see later that if liminf
0
|Du

(x
0
)| > 0, the exponential bound (3.9) implies
that u

(x
0
) and Du

(x
0
) are determined up to small errors by the boundary data on V
only at points where |Du

| |Du

(x
0
)|. That this is should be so is suggested by our
heuristic interpretation in 7 of the innity Laplacian PDE as a parabolic equation, with
time-like direction D
2
uDu = 1/2D(|Du|
2
). Therefore the values of u(x
0
) and Du(x
0
)
should be determined only by boundary data earlier in time, that is, at points on V
where |Du| |Du(x
0
)|.
Proof. 1. Plug the expression (3.7) into the identity (3.1), to nd
(|Du

(x
0
)|
2
)+2
_
V

(|Du

|
2
)(|D
2
u

Du

|
2
+ |D
2
u

|
2
)

dx
=
_
V
(|Du

|
2
)

dS 4
_
V

(|Du

|
2
)((

)
2
+ |D
2
u

Du

|
2
)

dx
=
_
V
(|Du

|
2
)

dS 4
_
V

(|Du

|
2
)((u

)
2
+ |D
2
u

Du

|
2
)

dx,
6
according to the PDE (1.2).
2. To establish the exponential estimate (3.9), we take
(3.11) (q) = e
(
2

q)

,
> 0 to be selected. Then (|Du

(x
0
)|
2
) = according to (3.10). Combining the two terms
in (3.8) involving integration over V , we compute
_
V
2

(|D
2
u

Du

|
2
+ |D
2
u

|
2
)

((u

)
2
+ |D
2
u

Du

|
2
)

dx
=
_
V
_
2(|D
2
u

Du

|
2
+ |D
2
u

|
2
) 4
2
((u

)
2
+|D
2
u

Du

|
2
)
_
e
(
2

|Du

|
2
)

dx
=
_
V
_
(2 4
2
)|D
2
u

Du

|
2
+ (2|D
2
u

|
2
4(u

)
2
)
_
e
(
2

|Du

|
2
)

dx

_
V
_
|D
2
u

Du

|
2
+ |D
2
u

|
2
_
e
(
2

|Du

|
2
)

dx
for some positive constant , provided we x > 0 suciently small.
3.4 A second integral identity. The identity (3.4) represents Du

(x
0
) as an integral of
Du

over V with respect to the density

. We will see next that provided |Du

(x
0
)| is
bounded away from zero, there is a corresponding, but approximate, formula for u

(x
0
)
THEOREM 3.4 (i) For each smooth function : R R we have the indentity
(3.12)
(u

(x
0
)) +
_
V

(u

)(|Du

|
4
+ |Du

|
2
)

dx
=
_
V
(u

dS + 2
_
V
u

(u

dx.
(ii) If
(3.13) liminf
0
|Du

(x
0
)| > 0,
then the last term on the right of (3.12) is O(
1
2
). In particular,
(3.14) u

(x
0
) =
_
V
u

dS + O(
1
2
).
(iii) Furthermore, (3.13) implies the estimate
(3.15)
_
V
(|Du

|
4
+ |Du

|
2
)

dx C.
7
The constant C in (3.15) depends upon a positive lower bound for

:= |Du

(x
0
)|.
Proof. 1. Multiply the PDE (1.2) by

(u

) amd rewrite, to discover that

u

x
i
u

x
j

x
i
x
j
=

(|Du

|
4
+ |Du

|
2
),
where = (u

). Next multiply by

and integrate by parts. Similarly to the previous

proof, we learn that
(u

(x
0
)) +
_
V

(u

)(|Du

|
4
+ |Du

|
2
)

dx =
_
V
(u

dS 2
_
V
u

x
i
u

x
i
x
j

x
j

dx.
The last integral term is
_
V
u

x
i
u

x
i
x
j

x
j

dx =
_
V
u

x
i
u

x
j
u

x
i
x
j

(u

dx =
_
V
u

(u

dx,
according to the PDE (1.2).
2. Our task now is to estimate the last term in (3.12), under the assumption (3.13). The
main issue is that we do not yet have an L
1
estimate for

that is independent of .
We rst consider the case that (z) =
z
2
2
. Take so small that

> 0
for some xed number . The identity (3.12) for (z) =
z
2
2
reads
(3.16)
1
2
(u

(x
0
))
2
+
_
V
(|Du

|
4
+ |Du

|
2
)

dx =
1
2
_
V
(u

)
2

dS + 2
_
V
u

dx.
We write the last integral as
(3.17)
2
_
V
u

dx = 2
_
V {|Du

2
}
u

dx 2
_
V {|Du

|<

2
}

dx
=: A + B.
We estimate
(3.18)
|A| C
_
V {|Du

2
}
|D
2
u

dx

1
2
C
__
V
|D
2
u

|
2

dx
_1
2
_
_
V {|Du

2
}

dx
_1
2

1
2
C
__
V
|Du

|
4

dx
_1
2

1
2
_
V
|Du

|
4

dx + C.
8
The third inequality above follows from (3.5). Furthermore,
(3.19)
|B| C
_
V {|Du

|<

2
}
|Du

||D
2
u

Du

dx
C
_

_
V
|Du

|
2

dx
_1
2
_
1

_
V {|Du

|<

2
}
|D
2
u

Du

|
2

dx
_1
2
C
_

_
V
|Du

|
2

dx
_1
2
e
3
2
8

2
_
V
|Du

|
2

dx + Ce

,
for some > 0. We used the exponential estimate (3.9) for the third inequality in this
calculation.
Employing the estimates (3.18) and (3.19) in (3.17) and (3.16), we derive the bound
(3.15). Returning again to (3.18) and (3.19), and now using (3.15) in the next-to-last lines,
we deduce that
|A| +|B| C
1
2
+ Ce

= O(
1
2
).
This proves assertion (ii) for (z) =
z
2
2
and the general case follows at once from the
foregoing estimates.
3.5 An improved L
1
estimate for

. We derive next a uniform L

1
estimate for

,
under the assumption that the terms |Du

(x
0
)| are bounded away from zero. This will be
much more useful than the crude bound (3.6).
THEOREM 3.5 (i) There exists a constant > 0 such that for each 0 < <

, we have
(3.20)
_
V {|Du

|}

dx
C

e
(
2

,
where

:= |Du

(x
0
)|.
(ii) If
(3.21) liminf
0
|Du

(x
0
)| > 0,
we have the uniform L
1
bound
(3.22)
_
V

dx C.
9
Proof. 1. Let (p) = (|p|
2
) for (q) = e
(
2

q)

and the constant from the estimate (3.9).

Then v

:= (Du

)|x|
2
satises
L

= L

()|x|
2
+ L

(|x|
2
) 4u

x
i
u

x
j

x
i
x
j
4
x
j
x
j
= (2

(|D
2
u

Du

|
2
+ |D
2
u

|
2
) + 4

((u

)
2
+ |D
2
u

Du

|
2
))|x|
2
2|Du

|
2
2n 4u

x
i
u

x
j

x
i
x
j
4u

x
i
u

x
j
x
j
x
j
4
x
j
x
j
where = (Du

).
Multiplying by

and integrating, we deduce using the bound (3.9) that

(3.23)
2
_
V
(n +|Du

|
2
)

dx
C +
_
V
(4u

x
i
u

x
j

x
i
x
j
4u

x
i
u

x
j
x
j
x
j
4
x
j
x
j
)

dx
=: C + A
1
+ A
2
+ A
3
.
2. We have
|A
1
| C
_
V
e
(
2

|Du

|
2
)

|Du

||

dx
= C
_
V
e
(
2

|Du

|
2
)

|Du

||u

dx
C
_
V
e
(
2

|Du

|
2
)

(u

)
2

dx +
_
V
|Du

|
2

dx
C +
_
V
|Du

|
2

dx,
according to (3.9). We also compute
|A
2
| C
_
V
e
(
2

|Du

|
2
)

|D
2
u

Du

dx
C
_
V
e
(
2

|Du

|
2
)

|D
2
u

Du

|
2

dx +
n
2
_
V

dx
C +
n
2
_
V

dx,
again according to (3.9). The estimate for A
3
is similar:
|A
3
| C
_
V
e
(
2

|Du

|
2
)

|D
2
u

Du

dx C +
n
2
_
V

dx.
10
We insert our estimates for A
1
, A
2
, A
3
into (3.23), to deduce
_
V
(n +|Du

|
2
)

dx C;
therefore
e
(
2

2
)

_
V {|Du

|}

dx C.
This proves (3.20).
3. Assuming now (3.21), we take so small that

> 0
for some positive constant . Then (3.20) implies for =

2
that
_
V {|Du

|}

dx Ce

where > 0. This and (3.15) prove (3.22).

4 Flatness estimates
In this section we assume that u is a bounded viscosity solution of the innity Laplacian
equation
(4.1)

u = 0 in B(0, 3).
We as before introduce the regularization
(4.2)
_

= 0 in B(0, 3)
u

= u in B(0, 3).
According to Theorem 2.1,
max
B(0,2)
|u

|, |Du

| C.
We consider also the adjoint problem on the ball B(0, 2):
(4.3)
_
(u

x
i
u

x
j

)
x
i
x
j
+ 2(u

x
i
u

x
i
x
j

)
x
j

=
x
0 in B(0, 2)

= 0 on B(0, 2),
11
for a given point x
0
B(0, 1). As in the previous section
(4.4)
_
B(0,2)

dx = 1
for

=
_
_
u

_
2
+
_
|D

|. Furthermore, if
(4.5) liminf
0
|Du

(x
0
)| > 0,
we know from Section 3 that
(4.6)
_
B(0,2)
(1 +|D
2
u

Du

|
2
+ |D
2
u

|
2
)

dx C,
and
(4.7)
_
B(0,2){|Du

|}

dx
C

e
(
2

,
for some > 0, where

:= |Du

(x
0
)| and 0 < <

. The constants C in (4.6) and (4.7)

depend upon a positive lower bound for the

.
In this section we make the additional atness assumption that the function u

is
uniformly close to an ane function in B(0, 2), which without loss we take to be the linear
function x
n
:
(4.8) max
B(0,2)
|u

x
n
| =: ,
where is small.
The ideal result would be that (4.8) forces the gradient Du

to be close to the unit vector

e
n
= (0, . . . , 0, 1) everywhere within the ball B(0, 1). This however is very subtle, and we are
not able to prove this. We can however show that Du

(x
0
) is close e
n
, provided x
0
B(0, 1),
is small, and |Du

(x
0
)| is close to one.
THEOREM 4.1 Assume the condition (4.5) that the gradient Du

(x
0
) is bounded away
from zero and also the atness condition (4.8). Select x
0
B(0, 1).
(i) We then have the estimate
(4.9)
_
B(0,2)
(|Du

|
2
u

xn
)
2

ds C.
(ii) Furthermore
(4.10)
_
B(0,2){|Du

|1+}

dx
C

2
for each > 0.
12
Proof. 1. Put v

:= (u

x
n
)
2
; then
(4.11)
L

= u

x
i
u

x
j
v

x
i
x
j
2u

x
i
u

x
j
x
j
v

x
j
v

= 2(u

x
n
)(u

x
i
u

x
j
u

x
i
x
j
2u

x
i
u

x
j
x
j
(u

x
j

jn
) u

)
2(|Du

|
2
u

xn
)
2
2|Du

e
n
|
2
.
Multiply the

and integrate over B(0, 2):

(u

(x
0
) x
0
n
)
2
=
_
B(0,2)
v

dx =
_
B(0,2)
L

dx +
_
B(0,2)
v

dS.
Using (4.4), (4.8) and (4.11), we deduce that
_
B(0,2)
(|Du

|
2
u

xn
)
2

dx C
2
+ C
_
B(0,2)
(|D
2
u

Du

| + |D
2
u

|)

dx
C
2
+ C,
the last inequality a consequence of (4.6). This proves (4.9).
2. On the set {|Du

| 1 + } we have
|Du

|
2
u

xn
|Du

|(|Du

| 1) ,
and so (4.10) follows from (4.9).
Next we strengthen (4.5), now to require that |Du

(x
0
)| be close to one, and then estimate
by how much Du

(x
0
) diers from e
n
:
THEOREM 4.2 Select any point x
0
B(0, 1). Suppose that
(4.12) 1 |Du

(x
0
)|
2
1 +
for a small constant > 0 and that the atness condition (4.8) holds.
Then
(4.13) |Du

(x
0
) e
n
|
2
C
_
e

1
2
+

1
2

+
1
4
+
1
2
+
1
2
_
.
The conclusion (4.13) is a strong consequence of the atness condition (4.8), since we
will later be able to adjust the various parameters to make the right hand side small. But
notice that we can only deduce this if we assume (4.12), that the length of the gradient is
close to one.
13
Proof. 1. Select a smooth function such that
(4.14) 1 on B(0, 1), = 0 on B(0, 2).
Then
(4.15)
_
B(0,2)
L

dx =
_
B(0,2)
L

dx = (x
0
) = 1.
We further compute that
L

(u

xn
) = L

xn
+ u

xn
L

2u

x
i
u

x
j
u

xnx
i

x
j
2u

xnx
i

x
i
= u

xn
L

2u

x
i
u

x
j
u

xnx
i

x
j
2u

xnx
i

x
i
,
since our dierentiating the PDE (4.2) shows L

xn
= 0.
Thus (4.14) and (4.15) imply
(4.16)
u

xn
(x
0
) 1 =
_
B(0,2)
(u

xn
1)L

dx 2
_
B(0,2)
u

x
i
u

x
j
u

xnx
i

x
j
+ u

xnx
i

x
i

dx
=: A + B.
2. Estimate of A. We recall (4.6), to compute
|A|
_
B(0,2)
|u

xn
1||L

dx
C
_
B(0,2)
|u

xn
1|
_
1 +|D
2
u

Du

| + |D
2
u

|
_

dx
C
__
B(0,2)
(u

xn
1)
2

dx
_1
2
C
__
B(0,2){|Du

|
2
12}
(u

xn
1)
2

dx
_1
2
+ C
__
B(0,2){|Du

|
2
1+2}
(u

xn
1)
2

dx
_1
2
+ C
__
B(0,2){1+2|Du

|
2
12}
(u

xn
1)
2

dx
_1
2
=: A
1
+ A
2
+ A
3
.
Owing to (4.7),
|A
1
|
C

1
2
e

2
.
14
Furthermore, (4.10) lets us estimate
|A
2
| C
__
B(0,2){|Du

|
2
1+2}

dx
_1
2

C
1
2

.
Finally, on the set {1 + 2 |Du

|
2
1 2} we have
(4.17) (u

xn
1)
2
C(|Du

|
2
u

xn
)
2
+ C
2
.
Consequently,
|A
3
| C
__
B(0,2)
(|Du

|
2
u

xn
)
2

dx + C
2
_1
2
C(
1
2
+ ).
in view of the estimate (4.9).
Collecting the foregoing bounds, we conclude that
(4.18) |A|
C

1
2
e

2
+
C
1
2

+ C(
1
2
+ ).
3. Estimate of B. To control this term, we rst observe that
u

x
i
u

x
j
u

xnx
i

x
j
=
1
2
n1

j=1
u

x
j
(|Du

|
2
)
xn

x
j
+
1
2
u

xn
(|Du

|
2
)
xn

xn
=
1
2
n1

j=1
u

x
j
(|Du

|
2
)
xn

x
j

_
1
2
n1

j=1
u

x
j
(|Du

|
2
)
x
j
+ u

xn
according to the PDE (4.2). Consequently,
(4.19) |u

x
i
u

x
j
u

xnx
i

x
j
| C|D

||D
2
u

Du

| + C|D
2
u

|,
where D

:= (u

x
1
, . . . , u

x
n1
, 0).
And so
|B|
_
B(0,2)
(|D

||D
2
u

Du

| + C|D
2
u

|)

dx
C
__
B(0,2)
|D

|
2

dx
_1
2
+ C
1
2
C
__
B(0,2){|Du

|
2
12}
|D

|
2

dx
_1
2
+ C
__
B(0,2){|Du

|
2
1+2}
|D

|
2

dx
_1
2
+ C
__
B(0,2){1+2|Du

|
2
12}
|D

|
4

dx
_1
4
+ C
1
2
=: B
1
+ B
2
+ B
3
+ C
1
2
.
15
As above,
|B
1
|
C

1
2
e

2
, |B
2
|
C
1
2

.
In addition, on the set {1 + 2 |Du

|
2
1 2}, we have
|D

|
2
1 (u
xn
)
2
+ 2 C|1 u

xn
| + C.
We therefore have from (4.17) that
|D

|
4
C(|Du

|
2
u

xn
)
2
+ C
2
.
Consequently, estimate (4.9) lets us conclude that
|B
3
| C(
1
4
+
1
2
).
Combining all these estimates, we see that we conclude that
(4.20) |B|
C

1
2
e

2
+
C
1
2

+ C(
1
4
+
1
4
) + C
1
2
.
4 The last inequality and the similar bound (4.18) for the term A prove that |u

xn
1|,
and therefore |u

xn
1|
2
, are less than or equal to the right hand side of (4.13).
To estimate the other derivatives, we see from (4.12) that at the point x
0
|D

|
2
1 (u

xn
)
2
+ C|1 u

xn
| + .
This and the foregoing estimate for |u

xn
1| complete the proof of (4.13).
5 Everywhere dierentiability
5.1 Blow up limits. If

u = 0 in the viscosity sense in some open subset U R

n
and
B(x, r) U, we dene
L
+
r
(x) :=
max
B(x,r)
u u(x)
r
, L

r
(x) :=
u(x) min
B(x,r)
u
r
As proved for example in [C-E-G], the limits
L(x) := lim
r0
L
+
r
(x) = lim
r0
L

r
(x)
exist and are equal for each point x U. (We will see later that in fact L(x) = |Du(x)|).
The paper [C-E-G] proves the following theorem, asserting that any blow-up limit around
any point x U must be a linear function. See [C-E] for a simplifed proof.
16
THEOREM 5.1 Let u be a viscosity solution of

u = 0 in U
and select any point x U.
For each sequence {r
j
}

j=1
converging to zero, there exists a subsequence {r
j
k
}

k=1
such
that
(5.1)
u(r
j
k
y + x) u(x)
r
j
k
a y locally uniformly,
for some a R
n
such that
(5.2) |a| = L(x).
Since solutions of

u = 0 are locally Lipschitz continuous, the rescaled functions

u
r
(y) :=
u(ry+x)u(x)
r
are locally bounded and Lipschitz continuous and consequently contain a
locally uniformly convergent subsequence. Theorem 5.1 asserts that each such limit is linear,
but does not prove that various blow-up limits, corresponding to dierent subsequences of
radii going to zero, are the same (unless L(x) = 0).
5.2 Dierentiability. This section resolves this uncertainty by proving the uniqueness of
the blow-up limits (5.1).
LEMMA 5.2 Assume b R
n
, |b| = 1. Let v be a smooth function satisfying
max
B(0,1)
|v b x|
for some small constant . Then there exists a point x
0
B(0, 1) at which
|Dv(x
0
) b| 6
Proof. Dene
w := b x 3|x|
2
+ .
We select the constant so that v w in B(0, 1), but v(x
0
) = w(x
0
) at some interior point
x
0
. Then Dv(x
0
) = Dw(x
0
) = b 6x
0
.
THEOREM 5.3 Let u be the unique viscosity solution of
(5.3)
_

u = 0 in U
u = g on U.
Then u is dierentiable at each point in U.
17
Proof. 1. Select any point within U, which without loss we may assume is 0. Suppose that
the blow up discussed in 5.1 does not produce a unique tangent plane at 0. This means
there exist two sequences {r
j
}

j=1
, {s
k
}

k=1
, each converging to zero, for which
(5.4) max
B(0,r
j
)
1
r
j
|u(x) u(0) a x| 0
and
(5.5) max
B(0,s
k
)
1
s
k
|u(x) u(0) b x| 0,
for distinct vectors a, b R
n
, with |a| = |b| > 0. We may assume without loss that
a = e
n
, |b| = 1, b = e
n
.
Dene
(5.6) := |b e
n
| > 0.
2. Hereafter C denotes the constant on the right hand side of estimate (4.13). We now
adjust various parameters to make the right hand side of this inequality small as compared
with
2
.
First select > 0 so small that
(5.7) C
1
2

2
24
.
Now x > 0 so that
(5.8) C
_

1
2

+
1
4
_

2
24
.
Next select
1
> 0 so that
(5.9) C
_
e

1
2
+
1
2
_

2
24
.
for all 0 <
1
.
We then use (5.4) (with a = e
n
) to select a radius r
j
> 0 for which
max
B(0,r
j
)
1
r
j
|u(x) u(0) x
n
|

2
.
18
We may without loss assume that r
j
= 2 and that u(0) = 0, as we can otherwise rescale and
consider the function
u(r
j
x)u(0)
r
j
. Hence
(5.10) max
B(0,2)
|u x
n
|

2
.
Now pick
2
> 0 so that
(5.11) max
B(0,2)
|u

x
n
| .
for all 0 <
2
.
3. We introduce yet another constant > 0, picked so that
(5.12) 12 + 36
2
, 6 ,
and
(5.13) 72
2

2
4
.
In view of (5.5), we can nd a (possibly very small) radius 0 < s < 1 for which
max
B(0,s)
1
s
|u b x|

2
.
We select
3
> 0 so that
(5.14) max
B(0,s)
1
s
|u

b x| ,
for all 0 <
3
, and hereafter take
(5.15) := min{
1
,
2
,
3
}.
Rescaling (5.14) to the unit ball and applying the Lemma, we secure a point x
0

B(0, s) B(0, 1) at which

(5.16) |Du

(x
0
) b| 6.
Then since |b| = 1, we have
(5.17) |Du

(x
0
)|
2
(1 + 6)
2
1 +
19
according to (5.12). Furthermore, |Du

(x
0
)| 1 6 and therefore
(5.18) |Du

(x
0
)|
2
1 ,
again owing to (5.12).
4. Now (5.17) and (5.18) allow us to invoke the key estimate (4.13):
|Du

(x
0
) e
n
|
2
C
_
e

1
2
+

1
2

+
1
4
+
1
2
+
1
2
_
.
In view of our choices (5.7), (5.8) and (5.9), it follows that
(5.19) |Du

(x
0
) e
n
|
2

2
8
.
Using (5.6), (5.13), (5.16) and (5.19), we at last reach the contradiction that

2
= |b e
n
|
2
2|Du

(x
0
) b|
2
+ 2|Du

(x
0
) e
n
|
2
72
2
+

2
4


2
2
.

Our paper [E-S] presents a simpler proof of the everywhere dierentiability.

6 The innity Laplacian PDE as a parabolic equation
This section provides heuristics to justify our claim that the innity Laplacian PDE
(6.1)

u = 0
should be regarded as a parabolic, and not an elliptic, equation. (G. Aronsson has made a
similar observation in his old paper [A], although for dierent reasons.)
6.1 Linearization. The only reasonable way to assert that a given nonlinear PDE is elliptic
or parabolic or hyperbolic at a particular solution is to classify, if possible, the type of its
linearization around this solution. We therefore consider the formal linearization of (6.1),
which is the PDE
(6.2) Lv := u
x
i
u
x
j
v
x
i
x
j
2u
x
i
u
x
i
x
j
v
x
j
= 0
for the variation v.
We contend that L is a parabolic equation, at least generically. Indeed the the second-
order term u
x
i
u
x
j
v
x
i
x
j
corresponds to diusion along the line parallel to the gradient Du,
20
whereas the rst-order term2u
x
i
u
x
i
x
j
v
x
j
corresponds to transport in the direction D
2
uDu.
According to the innity Laplacian equation (6.1), the direction of diusion Du is orthog-
onal to the direction of transport.
The linearized PDE (6.2) is therefore analogous to the one-dimensional linear heat equa-
tion
v
t
= v
xx
,
except that (6.2) has variable coecients, depending upon u, and holds in many variables.
We may think of the direction of D
2
uDu as the time-like direction and the perpendicular
directions, including that of Du, as space-like. In particular a critical point x
0
, where
|Du(x
0
)| = 0, is at time-like innity. Several of our rigorous assertions are consistent with
this interpretation, most notably the exponential estimate (3.9) which asserts that the value
of

is negligible at points y U where |Du

(y)| < |Du

(x
0
)|. Such points are forwards
in time for x
0
and so should not aect the solution at that point.
(If our smooth solution u of (6.1) happens also to be a solution of the eikonal equation
|Du| for some constant , the time-like term does not appear and the linearization is a
degenerate elliptic equation.)
6.2 Finite dierence approximation. Our revisiting a standard nite dierence approx-
imation for the innity Laplacian also reveals the parabolic structure.
Fix a step size h > 0 and dene the nonlinear nite dierence operator operator
(6.3)
h

u(x) :=
1
h
2
( max
B(x,h)
u + min
B(x,h)
u 2u(x)).
Then
(6.4)
h

u(x) =
1
h
2
(u(x
+
) + u(x

) 2u(x)),
the points x

are selected so that

u(x
+
) = max
B(x,h)
u, u(x

) = min
B(x,h)
u.
If u is smooth and Du = 0, we have
(6.5) lim
h0

u =

u
|Du|
2
:
see for instance Armstrong-Smart [A-S].
21
LEMMA 6.1 If u is smooth and Du(x) = 0, then
(6.6) x
+
= x + h
Du
|Du|
+ h
2
_
D
2
uDu
|Du|
2

uDu
|Du|
4
_
+ O(h
3
)
and
(6.7) x

= x h
Du
|Du|
+ h
2
_
D
2
uDu
|Du|
2

uDu
|Du|
4
_
+ O(h
3
),
Du and D
2
u evaluated at the point x.
Proof. Without loss, we may assume x = 0. Then
x
+
= h
Du(x
+
)
|Du(x
+
)|
= h
_
Du(0) + D
2
u(0)x
+
+ O(h
2
)
_
_
1
|Du(0)|

Du(0) D
2
u(0)x
+
|Du(0)|
3
+ O(h
2
)
_
,
and so
x
+
= h
Du(0)
|Du(0)|
+ O(h
2
).
Plugging this into the previous expansion, we deduce (6.6). The derivation of (6.7) is similar.

We observe that in view of (6.6) and (6.7) the dierence scheme (6.4) is that for a
parabolic PDE, involving O(h) steps in the space-like directions Du and an O(h
2
) step
in the time-like direction D
2
uDu. It is straightforward to check the consistency condition
that (6.5) follows from (6.6), (6.7).
6.3 Stochastic dierential equations. We introduce next a stochastic dierential equa-
tion, which provides a probabilistic interpretation of

and

:
(6.8)
_
dX

= Du

(X

)dW + D
2
u

(X

)Du

(X

)dt + (2)
1
2
dW
X

(0) = x
0
,
where W is a one-dimensional Brownian motion and W = (W
1
, . . . , W
n
) is an independent
n-dimensional Brownian motion.
Then

is the density of the distribution of X

(), where =

x
0
is the rst hitting time
for V . Furthermore if E V is a Borel set, then
_
E

dx = E
__

0

{X

(t)E}
dt
_
22
records the amount of time that the process X

spends within E before exiting V .

We can check using Itos calculus that Du

(X

) is a martingale, although in general

u

(X

) is not. This is why the formula (3.4) for the gradient Du

(x
0
) is exact, whereas the
expression (3.14) for u

(x
0
) has an error term (which is however small as 0).
7 Some numerical experiments
In a series of experiments we have studied numerically the limiting behavior of

and

as
0. We employed both a monotone and a second-order nite dierence scheme, and only
report computations for which both methods gave nearly identical results.
7.1 A monotone scheme. A. Obermans monotone nite dierence scheme [O] for the
normalized innity Laplacian PDE is easily adapted to our case: we need only multiply his
nite dierence operator by a suitable approximation of |Du|
2
. Given a step size h > 0, an
integer d > 0 and a function u : hZ
2
R, we therefore dene

h,d

u(x) :=
1
12(hd)
4
_
max
zN(x)
(u(z) u(x))
3
+ min
zN(x)
(u(z) u(x))
3
_
,
where
N(x) := {z hZ
2
| h(d 1/2) |x z| h(d + 1/2)}.
It is easy to see that
h,d

is monotone. Furthermore, for any smooth function , we have

h,d

locally uniformly as d and hd 0. Combining

h,d

with the standard

5-point Laplacian
h
, we obtain a monotone nite dierence scheme of the form

h,d

u
h,d

h
u
h,d
= 0.
A theorem of Barles and Souganidis [B-S] immediately implies the convergence of this scheme.
7.2 A second-order scheme. To obtain a higher-order scheme, we exploit the variational
structure of the regularized PDE (1.2). If we multiply by the integration factor exp(
1
2
|Du

|
2
),
we obtain
(7.1) div
_
e
1
2
|Du

|
2
Du

_
= 0;
and this is the Euler-Lagrange equation for minimizers of the functional
(7.2)
_
U
e
1
2
|Dv|
2
|Dv|
2
dx.
We can now construct a second-order convergent nite dierence approximation for (7.1)
using standard techniques (see for example Hackbusch [H] or LeVeque [L]). We in particular
23
selected a second-order accurate discretization of (7.1) and then solved the Euler-Lagrange
equations for the discrete variational problem.
We must however be very careful when implementing such a scheme, as the fast growth
of exp
_
1
2
|Dv|
2
_
increases the condition number of the linearization. Preconditioning is
required to obtain an accurate solution when is small. Even with this adjustment, numer-
ical instability manifests itself as a failure of the maximum principle for the adjoint of the
linearization when the step size h is insuciently small relative to .
7.3 Experimental results. For each trial we took several small values of and approxi-
mated u

and

for xed boundary data on the square

Q := {x R
2
| |x
1
|, |x
2
| < 1}.
Computation 1. In our rst experiment, we set x
0
= 0 and used boundary data given by
the argument function
u(x) := arctan
_
x
2
2 + x
1
_
.
Since u solves the regularized PDE (1.2) in R
2
{(2, 0)} for all > 0, we expect that

converges as 0 to the solution of

(7.3)
_
(u
x
i
u
x
j
)
x
i
x
j
2(u
x
i
u
x
i
x
j
)
x
j
=
x
0 in Q,
= 0 on Q.
This is exactly what appears to happen in Figure 1 below.
Computation 2. As a second numerical experiment, we put x
0
= (1/10, 1/2) and used
boundary data given by
(7.4) u(x) := x
4/3
1
x
4/3
2
,
an innity harmonic function discovered by Aronsson that is nonsmooth along the coordinate
axes {x
1
x
2
= 0} (which we regard as weak shocks).
We argue heuristically for this example that

and

csnnot concentrate solely within

the rst quadrant Q {x
1
, x
2
> 0} as 0, and therefore trajectories of the stochastic
dierential equation (6.8) with positive probability diuse across the forming weak shocks
along the coordinate axes. To see this, remember from (3.4) that
(7.5) Du

(x
0
) =
_
Q
Du

dS.
We assume now that Du

is close to Du along Q. Since then u

x
1
> 0 > u

x
2
and since
|Du

| > |Du

(x
0
)| on Q {x
1
, x
2
> 0}, the identity (7.5) could not be true as 0 if

were to concentrate only on Q {x

1
, x
2
> 0}.
24
Observe also that if we set = 0 in the stochastic dierential equation (6.8), the transport
vector
D
2
uDu =
4
27
(x
1
1
, x
1
2
),
not integrable near the coordinate axes {x
1
x
2
= 0}; whereas the diusion matrix
Du Du =
16
9
_
x
2/3
1
0
0 x
2/3
2
_
,
is bounded. So presumably a competition occurs as 0 between the decay of the diusion
and the growth of the transport in (6.8); and in the limit some positive portion of the mass
of

must remain outside of the rst quadrant.

It appears from the numerical data that

converges as 0 to a function that

solves (7.3) in Q {x
1
x
2
= 0}, but is singular on {|Du| > |Du(x
0
)|} {x
1
x
2
= 0}. There
are corresponding singularities in the limit of the

at the four points where these weak

shocks hit the boundary. This is most apparent in the bottom image in Figure 2, in which
we see cusps forming in the graph the of

as 0.
Computation 3. In our nal experiment, we set x
0
= (0, 1/10) and used the boundary
data
u(x) :=
(1 + rx
1
)
4/3
(rx
2
)
4/3
1
r
,
for small r > 0. That is, we zoomed in to a small neighborhood of a point on the weak
shocks of the Aronsson function (7.4) and get a closer view of the apparent singularities in

and

forming as 0. See Figure 3.

References
[A-S] S. Armstrong and C. Smart, A nite dierence approach to the innity Laplace
equation and tug-of-war games, to appear in Trans. AMS.
[A] G. Aronsson, On the partial dierential equation u
x
2
u
xx
+ 2u
x
u
y
u
xy
+ u
y
2
u
yy
= 0,
Ark. Mat. 7 (1968), 395425.
[B-S] G. Barles and P. E. Souganidis, Convergence of approximation schemes for fully
nonlinear second order equations, Asymptotic Analysis 4 (1991), 271-283.
[C-E] M. Crandall and L. C. Evans, A remark on innity harmonic functions, Proceedings
of the USA-Chile Workshop on Nonlinear Analysis (Vina del Mar-Valparaiso, 2000),
123129, Electron. J. Dier. Eq. Conf. 6, 2001.
25
[C-E-G] M. Crandall, L. C. Evans and R. Gariepy, Optimal Lipschitz extensions and the
innity Laplacian, Calc. Var. Partial Dierential Equations 13 (2001), 123139.
[E1] L. C. Evans, A survey of partial dierential equations methods in weak KAM theory,
Comm. Pure Appl. Math. 57 (2004) 445480.
[E2] L. C. Evans, Adjoint and compensated compactness methods for HamiltonJacobi
PDE, Archive Rational Mech Analysis 197 (2010), 10531088.
[E-S] L. C. Evans and C. Smart, Everywhere dierentiability of innity harmonic func-
tions, to appear.
[H] W. Hackbusch, Elliptic Dierential Equations: Theory and Numerical Treatment,
Springer Series in Computational Mathematics 18, Springer, 1992.
[L] R. J. LeVeque, Finite Dierence Methods for Ordinary and Partial Dierential
Equations, SIAM, 2007.
[O] A. Oberman, A convergent dierence scheme for the innity Laplacian: construction
of absolutely minimizing Lipschitz extensions, Math. Comp. 74 (2005), 1217-1230.
[S] O. Savin, C
1
regularity for innity harmonic functions in two dimensions, Archive
Rational Mech Analysis 176 (2005), 351361
26
1 0.5 0 0.5 1
1
0.5
0
0.5
1
1 0.5 0 0.5 1
1
0.5
0
0.5
1
(1,1) (1,1) (1,1) (1,1) (1,1)
0
0.5
1
1.5
2
2.5
3

1/4

1/16

1/64
Student Version of MATLAB
Figure 1: Clockwise from the upper left are the level sets of u

for = 1/64, the level

sets of

for = 1/64 and

for = 1/4, 1/16, 1/64. The boundary data are g(x, y) :=

arctan(y/(x + 2)) and the initial point is x
0
= 0.
27
1 0.5 0 0.5 1
1
0.5
0
0.5
1
1 0.5 0 0.5 1
1
0.5
0
0.5
1
(1,1) (1,1) (1,1) (1,1) (1,1)
10
0
10
5

1/2

1/8

1/32
Student Version of MATLAB
Figure 2: Clockwise from the upper left are the level sets of u

for = 1/32, the level

sets of

for = 1/32, and

for = 1/2, 1/8, 1/32. The boundary data are given by

g(x, y) := x
4/3
y
4/3
and the initial point is x
0
= (1/10, 1/2).
28
1 0.5 0 0.5 1
1
0.5
0
0.5
1
1 0.5 0 0.5 1
1
0.5
0
0.5
1
(1,1) (1,1)
10
0
10
5

1/4

1/16

1/64
1 0.5 0 0.5 1
10
0
10
5

1/4

1/16

1/64
Student Version of MATLAB
Figure 3: Clockwise from the upper left are the level sets of u

for = 1/64, the level

sets of

for = 1/64, the cross sections of

on {x
1
= 8/10} (indicated by the dotted
line), and

along the right-most edge of the domain. The boundary data are g(x, y) :=
r
1
[(1 + rx)
4/3
(ry)
4/3
] for r = 1/10 and the initial point is x
0
= (0, 1/10).
29