Maximal Functions in Analysis

Robert Feerman
June 20, 2005
The University of Chicago REU
Scribe: Philip Ascher
Abstract
This will be a self-contained introduction to the theory of maximal
functions, which are some of the most important objects in modern
harmonic analysis and partial dierential equations. We shall consider
various generalizations of the Fundamental Theorem of Calculus, and
wind up with an elementary introduction to Calderon-Zygmund The-
ory. One of the basic ingredients of the study of maximal functions
is a deep understanding of the geometry of collections of simple sets
such as balls or rectangles in Euclidean spaces. We shall investigate
this geometry in some detail.
Contents
1 The Fundamental Theorem of Calculus 2
2 Lebesgue Measure 2
3 The Generalized FTC 4
4 The Hardy-Littlewood Maximal Theorem 4
5 The Vitali Covering Lemma 5
6 Proof of The Hardy-Littlewood Theorem 5
7 The Generalized FTC Revisited 6
8 Proof of Hardy-Littlewood when p > 1 8
9 Some Closing Questions 10
1
1 The Fundamental Theorem of Calculus
Theorem 1.1 (The Fundamental Theorem of Calculus). If f is con-
tinuous then
__
x
a
f dx
_
= f and
_
b
a
F

dx = F(b) F(a).
Proof.
lim
h0
_
x+h
a
f dx
_
x
a
f dx
h
= lim
h0
_
x+h
x
f dx
h
. .
small average of f
= f
Later, well generalize the FTC to R
n
(n > 1), and to the Lebesgue Integral.
Denition 1.2. The Dirichlet Funtion is dened as:
D(x) =
_
1 x Q
0 x / Q
Since it is measurable, the Dirichlet funtion has the property that
_
1
0
Ddx = 1 ({D = 1}) + 0 ({D = 0})
where is Lebesgue Measure.
2 Lebesgue Measure
Lets recall some basic facts about Lebesgue Measure on R. If we have

_
k=1
I
k
S then (S)

k=1
length(I
k
)
Denition 2.1. The Lebesgue Measure of a set S is dened as:
(S) = inf
all covers

k=1
length(I
k
)
Remark 2.2.
(Q [0, 1]) <

k=1

2
k
= 0 > 0
2
Remark 2.3.

_
Q [0, 1]
_
= 1
Remark 2.4. For disjoint measurable sets

_

_
k=1
E
k
_
=

k=1
(E
k
)
Denition 2.5. A Lebesgue Simple Function S is a measurable function
with nitely many values v
1
, . . . , v
N
.
Remark 2.6. For a simple function S
_
1
0
S dx =
N

k=1
v
k
({S = v
k
})
Denition 2.7 (Lebesgue Integral). Let f(x) 0 be a measurable func-
tion on [0,1], then
_
1
0
f(x) dx
def
= sup
S simple
0Sf
_
1
0
S dx
Denition 2.8 (L
p
spaces). for 1 p < dene:
L
1
([0, 1])
def
=
_
f measurable |
_
1
0
|f| dx <
_
L
p
([0, 1])
def
=
_
f measurable |
_
1
0
|f|
p
dx <
_
Recall the Dirichlet Function
D(x) =
_
1 x Q
0 x / Q
We have that
__
x
0
Ddx
_
= D almost everywhere (that is, except on a set of
measure 0). We can verify this: if x / Q then
__
x
0
D(x) dx
_
= 0 = D(x) = 0,
so its false only when x Q (which has measure 0).
3
3 The Generalized FTC
Theorem 3.1 (The Generalized FTC for the Lebesgue Integral). Let
f L
1
([0, 1]). Then
__
x
0
f dx
_
= f almost everywhere
Denition 3.2. If f L
1
([0, 1] [0, 1]) then
sup
h>0
1
2h
_
x+h
xh
f dx = sup
xI
1
|I|
_
I
f dx = Mf(x)
Denition 3.3. If f L
1
(R
n
) then the Maximal Function of f is
M(f)(x)
def
= sup
xB
1
(B)
_
B
|f| dx
4 The Hardy-Littlewood Maximal Theorem
Theorem 4.1 (Hardy-Littlewood Maximal Theorem). If f L
p
(R
n
)
a) p = 1 {M(f)(x) > }
C

f
1
(a weak-type on L
1
)
b) p > 1 M(f)
p
C
p,n
f
p
(bounded on L
p
)
The rst part of this theorem is related to Chebychevs Inequality.
Theorem 4.2 (Chebychevs Inequality). If Q 0 then
({Q > })
1

_
R
n
Qdx
Proof. Starting out with a trivial inequality,

{Q>}
(x) Q(x)

_
R
n

{Q>}
dx
_
R
n
Qdx
({Q > })
_
R
n
Qdx
({Q > })
1

_
R
n
Qdx
4
5 The Vitali Covering Lemma
Question: can you get a subset of an arbitrary collection of nitely many
balls in R
n
that is disjoint and counts a xed fraction of their total area?
Lemma 5.1 (The Vitali Covering Lemma). Let B
1
, B
2
, . . . , B
N
be balls
in R
n
. Then a sub-collection of balls,

B
1
,

B
2
, . . . ,

B
M
, such that:
a) The

B
k
are pairwise disjoint
b)
_
M
_
k=1

B
k
_

1
3
n

_
N
_
k=1
B
k
_
Proof. First, arrange the balls by decreasing radius. Choose B
1
, so that

B
1
= B
1
, for B
2
choose

B
2
= B
2
if and only if it is disjoint from

B
1
= B
1
.
Continue this - so given a ball B
k
, choose it for the sub-collection if and only
if it is disjoint from all previously chosen balls.
Claim:
_
B
k

_
_

B
j
_
3
Where the notation (

B
j
)
3
means a concentric enlargement of the ball by a
factor of 3.
Pf of Claim: Take one B
k
and ask: why is B
k


_

B
j
_
3
? This is easy
to see by drawing a picture, and keeping in mind the way we chose the {

B
j
}
above. So:

_
_
B
k
_

_
_
_

B
j
_
3
_

__

B
j
_
3
_
=

3
n

_

B
j
_
= 3
n

_
_

B
j
_
since the {

B
j
} were chosen to be disjoint.
6 Proof of The Hardy-Littlewood Theorem
Theorem 6.1 (Restating Thm 4.1). If f L
p
(R
n
)
a) p = 1 {M(f)(x) > }
C

f
1
(a weak-type on L
1
)
b) p > 1 M(f)
p
C
p,n
f
p
(bounded on L
p
)
5
Proof. For p = , its easy to check that M(f)

1 f

. Now, well
use the Vitali Covering Lemma: x E

= {Mf > }, choose B

x
such that
1
(B
x
)
_
Bx
|f| dx >
and {B
x
}
xE
covers E

. Without loss of generality, B

1
, B
2
, . . . , B
k
, a
sequence of disjoint open balls from the B
x
s such that
_
k
B
k

_
xE
B
x
E

So were going to show

({M(f)(x) > })
_
N
_
k=1
B
k
_

C

f
1
N (C independent of N)
The proof is now fairly straightforward:

_
N
_
k=1
B
k
_
=
N

k=1
(B
k
)
<
N

k=1
1

_
B
k
|f| dx =
1

_
S
B
k
|f| dx
1

f
1
The second step in this string of inequalities comes from the following obser-
vation: look where
1
(B)
_
B
|f| dx > (B
k
) <
1

_
B
k
|f| dx
1

f
1
This completes the proof of the Hardy-Littlewood Theorem for p = 1 (see
section 8 for the other part).
7 The Generalized FTC Revisited
Theorem 7.1 (Generalized FTC in R
n
). If f L
1
(R
n
) then
lim
r0
1
(B
r
(x))
_
Br(x)
f dx = f outside a set of measure 0
6
Proof. Given any f L
1
(R
n
), a beautiful function C L
1
(R
n
) such that
_
R
n
|f C| dx is as small as you wish. Dene:
A
r
(f)(x) =
1
(B
r
(x))
_
Br(x)
f dx
We want to prove that
lim
r0
A
r
(f)(x) f(x) outside a set of measure 0
Note that A
r
is additive:
A
r
(f
1
+f
2
)(x) = A
r
(f
1
)(x) +A
r
(f
2
)(x)
We have f L
1
(R
n
). Let f = C + (C is the beautiful function above)
where = error = f C and
_
R
n
|| dx < . Since A
r
is additive:
A
r
(f)(x) = A
r
(C)(x) +A
r
()(x)
Subtracting equations we get:
A
r
(f)(x) f(x) = [A
r
(C)(x) C(x)]
. .
this0 x
+ [A
r
()(x) (x)]
Now we play some tricks with limsups:
limsup
r0
|A
r
(f)(x) f(x)|
= lim
k0
sup
0<r<k
|A
r
(f)(x) f(x)|
limsup
r0
|A
r
(C)(x) C(x)|
. .
this0 by usual FTC
+ limsup
r0
|A
r
()(x) (x)|
sup
r>0
|A
r
()(x)| +|(x)|
M()(x) +|(x)|
Remember that we want to prove that
lim
r0
A
r
(f)(x) f(x) outside a set of measure 0
7
so lets look at the following set:

_
limsup
r0
|A
r
(f)(x) f(x)| >
_

_
M() >

2
_
+
_
|(x)| >

2
_

C
_

2
_
1
+
_
|(x)| >

2
_
(by H-L)

2C

1
+
2

1
(by Chebychev)
8 Proof of Hardy-Littlewood when p > 1
Theorem 8.1 (Restating Thm 4.1). If f L
p
(R
n
)
a) p = 1 {M(f)(x) > }
C

f
1
(a weak-type on L
1
)
b) p > 1 M(f)
p
C
p,n
f
p
(bounded on L
p
)
Proof. We proved part a) in section 6, so now were going to prove part b),
and were going to show that a) implies b), that is: {M(f)(x) > }
C

f
1
implies that M(f)
p
C
p,n
f
p
for p > 1. Dene for > 0 the
distribution function of f:

f
() = {|f(x)| > }
Now were going to play with some integrals and use Fubinis Theorem:
_
R
n
|f(x)|
p
dx =(f
p
)
p
= p
_

0

p1

f
() d
= p
_

0

p1
{|f| > } d
= p
_

0

p1
_
R
n

{|f(x)|>}
dx d
=
_
R
n
_
|f(x)|
0
p
p1
d dx
=
_
R
n
|f(x)|
p
dx
8
Now f L
p
(R
n
), and dene f = f

+f

where
f

(x) =
_
f(x) where |f(x|

2
0 where |f(x| <

2
and
f

(x) =
_
0 where |f(x|

2
f(x) where |f(x| <

2
Since the maximum operator is a supremum of averages of values over balls,
it is sub-additive:
M(f) = M(f

+ f

) M(f

) +M(f

)
Now note that f

L
1
(R
n
) and f

(R
n
). Were going to make some
observations about the sets, and then look at their measures:
{M(f) > }
_
M(f

) >

2
_

_
M(f

) >

2
_
{M(f) > }
_
M(f

) >

2
_
+
_
M(f

) >

2
_

C
_

2
_f

1
(by part a) of H-L)
Now were going to use the distribution function of Mf:

Mf
()
2C

_
R
n
|f

(x)| dx =
2C

_
{|f(x)|>

2
}
|f(x)| dx
9
Using the same logic as before (and Fubini again):
_
R
n
|Mf(x)|
p
dx =(Mf
p
)
p
= p
_

0

p1

Mf
() d
p
_

0
2C

_
{|f(x)|>

2
}
|f(x)| dx
p1
d
= 2pC
_
R
n
_
2|f(x)|
0

p2
d |f(x)| dx
= 2pC
_
R
n
(2|f(x)|)
p1
p 1
|f(x)| dx
=
2
p
pC
p 1
_
R
n
|f(x)|
p
dx
= C
p
(f
p
)
p
9 Some Closing Questions
Do you think the generalized FTC is still true if we replace B
r
(x) with
rectangles, and state the theorem the same way? If not, when is it true (by
placing restrictions on f)? When is it true in R
2
? In R
n
?
10