The De Giorgi-Nash-Moser Estimates

We are going to discuss the the equation

Lu D
i
(a
ij
(x)D
j
u) = 0 in B
4
R
n
. (1)
The a
ij
, with i, j {1, . . . , n}, are functions on the ball B
4
. Here and in the
following doubly occurring indices are always understood to indicate sum-
mation. We assume that the coecients a
ij
satisfy the following ellipticity
condition
||
2
a
ij
(x)
i

j
for all R
n
and all x B
4
. (2)
with a positive constant . The equation Lu = 0 is then a second-degree
elliptic equation. We also require the a
ij
to be bounded and measurable,
satisfying a
ij

(B
4
)
with another constant > 0. (In case you wonder,
the radius 4 of the ball is to avoid fractions. Most of our estimates will be of
the kind some expression on B
1
another expression on B
4
and it would
be unconvenient to have things like 1/8 as a radius.)
It is clear that under these assumptions the equation (1) does not make
sense, for the a
ij
need not be dierentiable. In fact, we shall use it as an
abbreviation and really talk about so called weak solutions. Let us introduce
these terms. A function u H
1
(B
4
) is a weak solution to Lu = 0 if for all
H
1
0
(B
4
),
_
a
ij
D
j
uD
i
= 0. (3)
Conventional or strong solutions are obviously weak solutions as well. The
notion of weak solutions has come up because it provides a good way of
attacking equations. The question of solvability splits in two partsrst,
show that a weak solution exists; second, nd out how nice (continuous,
dierentiable) solutions are. Besides, there are equations without strong
solutions.
A function u is called a subsolution if
_
a
ij
D
j
uD
i
0 (4)
holds for all H
1
0
(B
4
) with 0. If the inequality in (4) is reversed, u is
a supersolution.
1
Our purpose is to carry out part of the second step described above. We
will demonstrate that a weak solution of Lu = 0 is not just an element of
the Sobolev space H
1
(B
4
) but is in fact in C

(B
4
), i.e., Holder continuous
with some exponent . Recall what this means: On a domain, a function u
is Holder continuous with exponent [0, 1] if it is continuous and
sup
x,yK
|u(x) u(y)|
|x y|

<
for every compact subset K of the domain.
Results of this kind can of course be obtained under dierent assump-
tions on the a
ij
; the mildest ones are probably those stated above. That
solutions to more general second-order equations with bounded coecients
a
ij
are Holder continuous was rst proved by De Giorgi (1957, for the elliptic
case) and independently by Nash (1958, for the parabolic case). De Giorgis
arguments were then much simplied and extended by Moser (1960, 1961).
The methods used involve what are called a priori estimates. One assumes
that an equation has a weak solution, usually in some Sobolev space, and then
tries to obtain general estimates on the solution, as in the case of Theorem 1
below. That way, regularity is proved for whole classes of solutions, only from
the fact that they are solutions; one does not consider individual properties.
(Hence the word a priori.)
One more word about notation. As usual, there are lots of constants.
Rather than lump them all together under one letter C we label them
successively by C
1
, C
2
, . . . so that one can work out the individual values if
needed. It is easier to check the dependence on the parameters that way.
1 Local Boundedness of Solutions
The rst step in proving Holder continuity is to show that solutions u to
Lu = 0 are locally bounded. This involves estimating the supremum of a
solution in terms of its L
2
-norm. But how in the world does one do that?
Because of the Sobolev inequality, u is in L
q
for q = 2n/(n 2). It turns
out that, u being a solution, these gains can be amplied by an iterative
procedure. The iteration used in the proof is due to Moser (and is named
after him).
2
Theorem 1. Let u H
1
(B
4
) be a subsolution to the equation, i.e., assume
that u satises (3). Then there is a positive constant C
4
, depending only on
n and /, such that
sup
B
2
u
+
C
4
u
L
2
(B
4
)
.
Proof. Let us rst give the general idea of the proof. By inserting a suitable
test function in the equation and playing around a bit, we can bound the
L
p
1
-norm of u in a smaller ball B
r
1
by the L
p
2
-norm in a larger ball B
r
2
,
where p
1
> p
2
2. In (13) below, we will have an estimate of the form
u
+

L
p
1(Br
1
)
Cu
+

L
p
2(Br
2
)
,
some kind of reversed Holder inequality. We will iterate this, choosing r
i
and
p
i
carefully, to get our result.
To begin with, introduce the following two functions. For positive num-
bers k and m, set u = k + u
+
and
u
m
=
_
u if u < m
k + m if u m
The point is that u
m
is still an element of H
1
(B
4
), but bounded from below
by k and from above by (k +m). Also note that D u
m
= 0 whenever u < 0 or
u > m and that u
m
= u at all other points. The function u is always positive
and D u = Du if u 0. Both k and m are needed to make the argument
work; in the end, we will let k 0+ and m , so that both u and u
m
converge to u
+
.
We start from the fact that u is a subsolution. In the inequality (4),
we use a test function of the form =
2
( u

m
u k
+1
), where 0 is
an arbitray real number and C
1
0
(B
4
) a nonnegative cut-o function to
be chosen later on. This function is an element of H
1
0
(B
4
) because u
m
is
bounded; it is also nonnegative and can therefore be used as a test function.
An explicit calculation with the weak derivatives gives
D =
2
u
1
m
D u
m
u + D u
2
u

m
+ 2D( u

m
u k
+1
)
=
2
u

m
(D u
m
+ D u) + 2D( u

m
u k
+1
),
(5)
where we used positivity of u
m
and the fact that u = u
m
whenever D u
m
= 0.
3
Now insert this expression into the inequality (4) to obtain
0
_
a
ij
D
j
uD
i

=
_
a
ij
D
j
u
2
u

m
(D
i
u
m
+ D
i
u) + 2
_
a
ij
D
j
uD
i
( u

m
u k
+1
)

_

2
u

m
_
|D u
m
|
2
+ |D u|
2
_
2
_

a
ij
D
j
uD
i

m
u k
+1

.
(6)
Notice that D = 0 whenever u < 0, so that all integrals are eectively over
the set {u 0} only. This allowed us to replace D
j
u by D
j
u, since the two
are equal if u 0. We also used the ellipticity condition (2) and the fact
that D
i
u = D
i
u
m
whenever the latter is nonzero.
Let us estimate the second integral. From Cauchys inequality,
2
_

a
ij
D
j
uD
i

m
u k
+1

2
_
n|D u||D| | u

m
u k
+1
|,
and since u

m
u k
+1
0,
2n
_
|D u||D| u

m
u =
_
_
|D u| u
/2
m
__
2n|D| u u
/2
m
_
.
Apply Youngs inequality
_
ab

2
a
2
+
1
2
b
2
_
to the two bracketed factors to
get

2
_
|D u|
2

2
u

m
+
2n
2

_
|D|
2
u
2
u

m
.
Combining this result with (6) and simplifying slightly, we obtain

_

2
u

m
|D u
m
|
2
+
_

2
u

m
|D u|
2

4n
2

2
_
|D|
2
u
2
u

m
= C
1
_
|D|
2
u
2
u

m
(7)
where C
1
= 4n
2
(/)
2
.
We dene an additional function w H
1
(B
4
) by w = u
/2
m
u. As before,
one calculates that Dw = u
/2
m
(/2 D u
m
+ D u); therefore
|Dw|
2
= u

2
D u
m
+ D u

2
= u

m
_

2
4
|D u
m
|
2
+ D u
m
D u +|D u|
2
_
= u

m
_
(/4 + 1)|D u
m
|
2
+|D u|
2
_
u

m
( + 1)
_
|D u
m
|
2
+|D u|
2
_
.
(8)
4
In combination with (7), we have
_
|Dw|
2

2
( + 1)
_

_

2
u

m
|D u
m
|
2
+
_

2
u

m
|D u|
2
_
( + 1)C
1
_
|D|
2
w
2
and so
_
|D(w)|
2
2
_
|D|
2
w
2
+|Dw|
2

2
2
_
1 + C
1
( + 1)
_
_
|D|
2
w
2
4C
1
( + 1)
_
|D|
2
w
2
.
(9)
Remember that we want to estimate stronger L
p
-norms by weaker ones.
Here is how we do it. From the Sobolev inequality, with = n/(n 2) > 1
for n > 2 and any xed > 2 for n = 2, we get
__
(w)
2
_1

C(n)
_
|D(w)|
2
4C
1
C(n)( + 1)
_
|D|
2
w
2
. (10)
Now choose a suitable cut-o function. For 0 < r < R 4, take C
1
0
(B
4
)
with 1 in B
r
and |D| 2/(R r). Then
__
Br
w
2
_1

__
(w)
2
_1

4C
1
C(n)( + 1)
_
|D|
2
w
2
16C
1
C(n)
( + 1)
(R r)
2
_
B
R
w
2
= C
2
( + 1)
(R r)
2
_
B
R
w
2
,
(11)
and, if we let = + 2 2, recall the denition of w and use that u
m
u,
__
Br
u

m
_1

=
__
Br
u
2
u

m
_1

=
__
Br
w
2
_1

C
2
( + 1)
(R r)
2
_
B
R
w
2
C
2
( 1)
(R r)
2
_
B
R
u
2
u

m
C
2
( 1)
(R r)
2
_
B
R
u

.
(12)
Finally, let m and k 0+ (use Fatous Lemma) to obtain the crucial
estimate
u
+

(Br)

_
C
2
( 1)
(R r)
2
_1

u
+

(B
R
)
. (13)
5
Observe how the stronger L

-norm is estimated by the weaker L

-norm. As
a trade-o, we have to increase the ball from B
r
to the larger B
R
. As we
said above, the two ingredients were the Sobolev inequality and the equation
itself.
The key observation is that (13) is valid for all 0 < r < R 4 and
for all 2. This suggests an iteration, taking successively the values
= 2, 2, 2
2
, . . . . Dene, for all i = 0, 1, 2, . . . ,

i
= 2
i
and r
i
= 2 +
1
2
i1
.
For any i 0, insert r = r
i+1
, R = r
i
and =
i
into (13),
u
+

i+1
(Br
i+1
)

_
C
2
(
i
1)
(1/2
i
)
2
_ 1

i
u
+

i (Br
i
)
C
i

i
3
u
+

i (Br
i
)
.
C
3
depends only on n and / because is a function of n alone. By
iteration,
u
+

i (Br
i
)
C
P
i

i
3
u
+

L
2
(B
4
)
= C
4
u
L
2
(B
4
)
,
valid for all i 1; since all r
i
are greater than 2, one arrives at
u
+

i (B
2
)
= C
4
u
L
2
(B
4
)
.
Now let i ; this entails
i
and gives us
sup
B
2
u
+
C
4
u
L
2
(B
4
)
.
We have proved the estimate.
There are technical reasons for stating the theorem in terms of subsolu-
tions: this helps in the proofs of the next section. For solutions u of (3), we
can immediately derive a stronger result, namely boundedness on compact
subsets of B
4
.
Corollary 1. Let u H
1
(B
4
) be a weak solution to Lu = 0 in the ball B
4
.
Then u satises
u
L

(B
2
)
C
4
u
L
2
(B
4
)
with the same constant C
4
as in Theorem 1. Moreover, u is bounded on each
compact subset K of B
4
.
6
Proof. If u is a solution to (3), then both u and (u) are subsolutions. In
addition to the estimate in the theorem, we therefore have
sup
B
2
u

C
4
u
L
2
(B
4
)
.
If we combine the two, we get
u
L

(B
2
)
C
4
u
L
2
(B
4
)
. (14)
Now one quickly sees that u is bounded on each compact subset K. Con-
sider rst the case of a ball B
s
(a) contained in B
4
, with s > 0. If we dene
v(x) = u(a + sx), v is an element of H
1
(B
4
), and by taking test functions
with support in B
s
(a) in the original equation (3) and changing coordinates,
we see that v satises _
a
ij
D
j
vD
i
= 0
for all H
1
0
(B
4
), where a
ij
(x) = a
ij
(a + sx). Inequality (14) above, when
applied to v, gives us
u
L

(B
s/2
(a))

C
4

s
u
L
2
(Bs(a))

C
4

s
u
L
2
(B
4
)
.
An arbitrary compact subset K can be covered by nitely many open
balls B(a
i
, s
i
/2) of positive radius s
i
, such that B(a
i
, s
i
) B
4
. Then
u
L

(K)
max
i
u
L

(B
s
i
/2
(a
i
))
max
i
C
4

s
i
u
L
2
(B
4
)
which is nite. Therefore u is bounded on K.
7
2 Holder Continuity of Solutions
In this section, we show that solutions to Lu = 0 are Holder continuous. On
the way we meet with the very pretty Theorem 2 which gives a lower bound
on positive solutions. The following lemma on subsolutions and supersolu-
tions is helpful.
Lemma 1. Let be a convex and locally Lipschitz continuous function on
some interval I.
1. If u is a subsolution with values in I and

0, then v = (u) is also

a subsolution, provided it is in H
1
loc
(B
4
).
2. If u is a supersolution with values in I and

0, then v = (u) is a
subsolution, provided it is in H
1
loc
(B
4
).
Proof. Let us prove the second statement only, since it is the one used in
Theorem 2 below (the rst one is dealt with in an analogous manner). Since
C
1
0
is dense in H
1
0
, it is enough to consider test functions C
1
0
. If one as-
sumes that C
2
loc
(I), then

(t) 0 and

(t) 0. Take any nonnegative

C
1
0
(B
4
). A direct calculation gives
_
a
ij
D
j
vD
i
=
_
a
ij

(u)D
j
uD
i
=
=
_
a
ij
D
j
uD
i
(

(u))
_
(a
ij
D
j
uD
i
u)

(u)

_
a
ij
D
j
uD
i
(

(u))
_
|Du|
2

(u) 0
because

(u) H
1
0
(B
4
) is nonnegative. Therefore (u) is a subsolution.
In general, let

be the standard mollier and set

(t) =

(t). Then

(t) =

(t) 0 and

(t) =

(t) 0. By what we have just

proved,

(t) is a subsolution. Because

(t)

(t) a.e. as 0+ and

because has compact support, the dominated convergence theorem implies
that
0
_
a
ij

(u)D
j
uD
i

_
a
ij

(u)D
j
uD
i
=
_
a
ij
D
j
vD
i
,
which gives the result.
8
We use this lemma in the following way. The function (t) = (log t)

is
convex and satises local Lipschitz conditions on (0, ). If u H
1
(B
4
) is a
positive supersolution to the equation, then (u) = (log u)

is a subsolution,
provided it is still in H
1
loc
B
4
, which is the case if, say, u is bounded from
below by a positive number.
The next two theorems show that solutions to the equation (3) cannot
oscillate too much. This is reminiscent of the behavior of harmonic functions,
for example of Harnacks inequality for positive solutions to u = 0 on a
domain . It states that for any compact subset K of , there is an absolute
constant, depending only on K and , such that
sup
K
u C inf
K
u
holds for any positive harmonic function u on .
But maybe these connections come as no surprise, for the Laplace equa-
tion is a special case of (1). In fact, Moser was able to prove an analogue
of Harnacks inequality for weak solutions to (3). For more details, see [1],
Chapter 4.4.
Theorem 2 (Density Theorem). Suppose that u H
1
(B
4
) is a positive
supersolution with
m
_
{x B
2
: u 1}
_
m
_
B
2
_
for some > 0. Then there exists a constant C
7
= C
7
(, n, /) (0, 1)
such that
inf
B
1
u C
7
.
Proof. Since one can always add a small constant to u, we can assume that
u > 0 (let 0+ in the end). By Lemma 1, the function v = (log u)

is a subsolution, bounded by log 1/. After a dilatation, Theorem 1 tells us

that
sup
B
1
v
C
4

2
v
L
2
(B
2
)
.
Since m
_
{x B
2
: v = 0}
_
= m
_
{x B
2
: u 1}
_
m
_
B
2
_
, one of the
versions of the Poincare inequality (as discussed in class) implies
sup
B
1
v
C
4
C(, n)

2
Dv
L
2
(B
2
)
. (15)
9
To show that the right-hand side is bounded, use a test function =
2
/u
(with C
1
0
(B
2
)),
0
_
a
ij
D
j
uD
i
_

2
u
_
=
_

2
a
ij
D
j
uD
i
u
u
2
+ 2
_
a
ij
D
j
uD
i

u
.
From the ellipticity condition and Holder inequality, one gets

_

2
|Dlog u|
2
2n
__

2
|Dlog u|
2
_1
2
__
|D|
2
_1
2
,
which implies
_

2
|Dlog u|
2

4
2
n
2

2
_
|D|
2
.
If we take a xed C
1
0
(B
4
) with 1 in B
2
, we have
_
|Dlog u|
2

4
2
n
2

2
_
|D|
2
= C
5
. (16)
where C
5
is a constant depending on n and /.
Along with (15) we obtain
sup
B
1
v = sup
B
1
(log u)

C
4
C(, n)

2
C
5
= C
6
,
which gives
inf
B
1
u e
C
6
= C
7
> 0.
Theorem 3 (Oscillation Theorem). Suppose that u is a solution of Lu = 0
in B
4
. Then there exists a number = (n, /) (1/2, 1) such that
osc
B
1
2
u osc
B
2
u
Proof. In Corollary 1, it was shown that u is bounded on compact subsets
of B
4
. We may thus dene

1
= sup
B
2
u and
1
= inf
B
2
u
10
as well as

2
= sup
B
1
2
u and
2
= inf
B
1
2
u.
Excluding the trivial case of constant u, the two functions
u
1

1
and

1
u

1
are positive solutions to the equation on B
2
. Note the following two equiva-
lences:
u
1
2
(
1
+
1
)
u
1

1
2
u
1
2
(
1
+
1
)

1
u

1
2
Depending on whether u is generally big or not, there are two possibilities.
Case 1. Suppose that
m
_
_
x B
1
: 2
u
1

1
1
_
_

1
2
m
_
B
1
_
.
Apply the Density Theorem (with = 1/2) to the function 2
u
1

1
0, but
in B
2
instead of in B
4
. For some constant C
7
(0, 1), we have
inf
B
1
2
u
1

C
7
2
from which we obtain the estimate

2
= inf
B
1
2
u
1
+
C
7
2
(
1

1
).
Case 2. Now suppose that
m
_
_
x B
1
: 2

1
u

1
1
_
_

1
2
m
_
B
1
_
.
Here, the result is

2
= sup
B
1
2
u
1

C
7
2
(
1

1
).
11
with the same constant C
7
.
Since clearly
2

1
and
2

1
, we have in both cases

2

_
1
C
7
2
_
(
1

1
),
which is our inequality with = 1 C
7
/2.
After all that work, we are now able to come to the following, triumphant
conclusion (called De Giorgis theorem).
Theorem 4. Suppose that u H
1
(B
4
) is a weak solution of the equation
Lu = 0 in B
4
. Then there holds
sup
xB
2
|u(x)| + sup
x,yB
2
|u(x) u(y)|
|x y|

C
8
(n, /)u
L
2
(B
4
)
with some real number = (n, /) (0, 1). Moreover, u C

(B
4
).
Proof. One half of the estimate, namely
sup
xB
2
|u(x)| C
4
(n, /)u
L
2
(B
4
)
, (17)
is given by Theorem 1.
For the second half, x two arbitray distinct points x, y B
2
and set
r = |x y|. For some n 0, we have 4
n+1
> r 4
n
. Let us rst consider
the interesting case n > 0. By applying the oscillation theorem several times
on suitable dilates of u, we obtain
osc
Br(x)
u
n1
osc
B
4
n1
r
(x)
u
n1
2C
4
u
L
2
(B
4
)
,
invoking Theorem 1 in the last step. In particular,
|u(x) u(y)|
n1
2C
4
u
L
2
(B
4
)
.
Now r 4
n
and so r

4
n
for any (0, 1). We get
|u(x) u(y)|
|x y|

n1
4
n
2C
4
u
L
2
(B
4
)
8C
4
(4

)
n1
u
L
2
(B
4
)
.
Take such that 4

= 1; since (1/2, 1), actually falls in the range

(0, 1/2).
12
The case n = 0 is somewhat easier. Because r 1, we directly conclude
that
|u(x) u(y)|
|x y|

2C
4
u
L
2
(B
4
)
.
We see that the choice C
8
= 8C
4
gives the inequality in the theorem.
To prove that u is an element of C

(B
4
), that is to say, Holder continuous
on each compact subset of B
4
, one can proceed in exactly the same way as
in Corollary 1. We omit this argument.
13
References
[1] Q. Han and F. Lin, Elliptic partial dierential equations, New York Uni-
versity, Courant Institute for Mathematical Sciences, New York, 1997
[2] J. N. Mather et al., Notices Amer. Math. Soc. 47 (2000), no. 11, 1392
1405
14