2 Singular Integrals

We start with a very useful covering lemma.

Lemma 1. Suppose K S is a compact subset and I

is a covering of K.
There is a nite subcollection {I
j
} such that
1. {I
j
} are disjoint.
2. The intervals {3I
j
} that have the same midpoints as {I
j
} but three
times the lenghth cover K.
Proof. We rst choose a nite subcover. From the nite subcover we pick
the largest interval. In case of a tie pick any of the competing ones. Then, at
any stage, of the remaining intervals from our nite subcollection we pick the
largest one that is disjoint from the ones already picked. We stop when we
cannot pick any more. The collection that we end up with is clearly disjoint
and nite. Let x K. This is covered by one of the intervals I from our
nite subcollection covering K. If I was picked there is nothing to prove.
If I is not picked it must intersect some I
j
already picked. Let us look at
the rst such interval and call it I
j
. I is disjoint from all the previously
picked ones and I was passed over when we picked I
j
. Therefore inaddition
to intersecting I
j
, I is not larger than I
j
. Therefore 3I
j
I x.
This lemma is used in proving maximal inequalities. For instance, for the
Hardy-Littlewood maximal function we have
Theorem 1. Let f L
1
(S). Dene
M
f
(x) = sup
0<r<

2
1
2r

|yx|<r
|f(y)|dy (1)
[x : M
f
(x) > ]
3

|f(y)|dy

(2)
Proof. Let us denote by E

the set
E

= {x : M
f
(x) > }
6
and let K E

be an arbitrary compact set. For each x K there is an

interval I
x
such that

Ix
|f(y)|dy (I
x
)
Clearly {I
x
} is a covering of K and by lemma we get a nite disjoint sub
collection {I
j
} such that {3I
j
} covers K. Adding them up

|f(y)|dy

j
(I
j
)
1
3

j
(3I
j
) (K)
Sine K E

is arbitrary we are done.

There is no problem in replacing {x : |M
f
(x)| > } by {x : |M
f
(x)| }.
Replace by and let 0.
This theorem can be used to prove the Labesgue direntiability theorem.
Theorem 2. For any f L
1
(S),
lim
h0
1
2h

|xy|h
|f(y) f(x)|dy = 0 a.e. x (3)
Proof. It is sucient to prove that for any > 0
[x : limsup
h0
1
2h

|xy|h
|f(y) f(x)|dy ] = 0
Given > 0 we can write f = f
1
+g with f
1
continuous and g
1
and
[x : limsup
h0
1
2h

|xy|h
|f(y) f(x)|dy ]
= [x : limsup
h0
1
2h

|xy|h
|g(y) g(x)|dy ]
[x : sup
h>0
1
2h

|xy|h
|g(y) g(x)|dy ]

3h
1

Since > 0 is arbitrary we are done.

7
In other words the maximal inequality is useful to prove almost sure
convergence. Typically almost sure convergence will be obvious for a dense
set and the maximal inequality will be used to interchange limits in the
approximation.
Another summability method, like the Fejer sum that is often considred
is the Poisson sum
S(, x) =

n
a
n

|n|
e
inx
and the kernel corresponding to it is the Poisson kernel
p(, z) =
1
2

|n|
e
inz
=
1
2
1
2
(1 2 cos z +
2
)
(4)
so that
P(, x) =

f(y)p(, x y)dy
It is left as an exercise to prove that for for 1 p < , every f L
p
P(, ) f() in L
p
as 1. We will prove a maximal inequality for the
Poisson sum, so that as a consequence we will get the almost sure convergence
of P(, x) to f for every f in L
1
.
Theorem 3. For every f in L
1
[x : sup
0<1
P(, x) ]
Cf
1

(5)
Proof. The proof consists of estimating the Poisson maximal function in-
terms of the Hardy-Littlewood maximal function M
f
(x). We begin with
some simple estimates for the Poisson kernel p(, z).
p(, z) =
1
2
1
2
(1 )
2
+ 2(1 cos z)

1
2
1
2
(1 )
2
=
1
2
1 +
1

1

1
1
The problem therefore is only as 1. Lets us assume that
1
2
.
8
For any symmetric function (z) the intgral
|

f(z)(z)dz|
= |

0
[f(z) + f(z)](z)dz|
= |

0
(z)[
d
dz

z
z
f(y)dy]dz|
|

(z)[

z
z
f(y)dy]dz| +|()|

f(z)dz|

0
2|z

(z)|[

|f(y)|
z
(dy)]dz +()|

|f(z)|dz
2M
f
(0)

0
|z

(z)|dz +()||M
f
(0)|
For the Poisson kernel
|z
d
dz
p(, z)| =
1
2
1
2
(1 2 cos z +
2
)
2
2|z sin z|

(1 )z
2
(1 )
4
+ (1 cos z)
2
C
(1 )z
2
(1 )
4
+z
4
and

|z
d
dz
p(, z)|dz C

(1 )z
2
(1 )
4
+z
4
dz
=

1

1
z
2
1 +z
4
dz

z
2
1 +z
4
dz C
1
uniformly in .
Interpolation theorems play a very important role in Harmonic Analysis.
An example is the following
9
Theorem 4 (Marcinkiewicz). Let T be a sublinear map deened on L
p

L
q
that satises weak type inequlities
[|x| : |(Tf)(x)| ]
C
i
f
p
i
p
i

p
i
(6)
for i = 1, 2 where 1 p
1
< p
2
< . Then for p
1
< p < p
2
, there are
constants C
p
such that
Tf
p
C
p
f
p
(7)
Note that T need not be linear. It need only satisfy
|T(f +g)|(x) |Tf|(x) +|Tg|(x) (8)
Proof. Let p (p
1
, p
2
) be xed. For any function f L
p
and for any positive
number a we deine f
a
= f
{|f|a}
and f
a
=
{|f|>a}
. Clearly f
a
L
p
2
and
f
a
L
p
1
[x : |Tf(x)|| 2] [x : |Tf
a
(x)|| ] +[x : |Tf
a
(x)|| ]

C
2

p
2

|f(x)|a
|f(x)|
p
2
d +
C
1

p
1

|f(x)|>a
|f(x)|
p
1
d
Take a = , multiply by
p1
and integrate with respect to from 0 to .
Use Fubinis theorem. We get

p1
[x : |Tf(x)|| 2]d [
C
2
p
2
p
+
C
1
p p
1
]

|f(x)|
p
d (9)
Since the left hand side is
Tf
p
p
p
we are done.
There is a slight variation of the argument that allows p
2
to be innite pro-
vided T is bounded on L

. If we denote the norm by C

2
we use
[x : |Tf(x)|| (C + 1)] [x : |Tf
a
(x)|| ]
and proceed as before.
A dierent interpolation theorem for linear maps T is the following
10
Theorem 5 (Riesz-Thorin). If a linear map T is bounded from L
p
i
into
L
p
i
with a bound C
i
for i = 1, 2 then for p
1
p p
2
it is bounded from L
p
into L
p
with a bound C
p
that can be taken to be
C
p
= C
t
1
C
1t
2
(10)
where t is determined by
1
p
= t
1
p
1
+ (1 t)
1
p
2
(11)
Proof. The proof uses methods from the theory of functions of a complex
variable. The starting point is the maximum modulus principle. Let us
assume that u(z) is analytic in the open strip a < Rez < b and bounded
and continuous in the closed strip a Rez b. Let M(x) be the maximum
modulus of the function on the line Rez = x. Then log M(x) is a convex
function of x.This is not hard to see. Clearly the maximum principle dictates
that
M(x) max[M(a), M(b)]
If one is worried about the maximum being attained, one can always mutiply
by e
z
2
and let go to 0. Replacing u(z) by u(z)e
tz
yields the inequality
M(x) max[M(a)e
t(ax)
, M(b)e
t(bx)
]
optimizing with respect to t we get,
M(x) max[[M(a)]
bx
ba
, [M(b)]
xa
ba
]
which is the required convexity.
We note that the maximum of any collection of convex functions is again
convex. The proof is completed by representing log F(p), where F(p) is the
norm of T from L
p
to L
p
, as the supremum of a bunch of functions that are
11
convex in x =
1
p
.
T
p,p
= sup
fp1
gq1
|

g(Tf)d|
= sup
fp1,f0,||=1
gq1,g0,||=1
|

(g)(T(f))d|
= sup
f
1
1,f>0,||=1
g
1
1,g>0,||=1
|

(g
x
)(T(f
1x
))d|
= sup
f
1
1,f>0,||=1
g
1
1,g>0,||=1
Rez=x
|

(g
z
)(T(f
1z
))d|
= sup
f
1
1,f>0,||=1
g
1
1,g>0,||=1
sup
Rez=x
|u(f, g, , , z)|
In particular for the Hardy-Littlewood or Poisson maximal function the
L

bound is trivial and we now have a bound for the L

p
norm of the maximal
function in terms of the L
p
norm of the original function provided p > 1.
For a convolution operator of the form
(Tf)(x) =

f(y)k(x y)dy (12)

we saw that for it to be bounded as an operator from L
1
into itself we need
k to be in L
1
. However for 1 < p < the operator can some times be
bounded even if k is not in L
1
. This is proved by establishing a bound from
L
2
to L
2
and a weak type inequality in L
1
. We can then use Marcinkiewicz
interpolation, followed by Riesz-Thorin interpolation.
Theorem 6. If

k(n) =

e
inz
k(z)dz
is bounded in absolute value by C, then the convolution operator given by
equation (12) is bounded by C as an operator from L
2
to L
2
.
12
Proof. Use the the orthonormal basis e
inx
to diagonalize T
Te
inx
=

k(n)e
inx
(13)
We now proceed to establish weak type (1, 1) estimate. We shall assune that
we have a kernel k in L
1
that satises
1.
sup
n
|

k(y)e
iny
dy| = C
1
< (14)
2.
sup
y

x:|xy|>2|y|
|k(x y) k(x)|dx = C
2
< (15)
Although we have assumed that k is in L
1
we will prove a weak type (1, 1)
bound.
Theorem 7. The operator of convolution by k
(T
k
f)(x) =

k(x y)f(y)dy (16)

satises the weak type inequality (1,1)
[|x| : |(Tf)(x)| ]
C

f
1
(17)
with a constant C that depends only on C
1
and C
2
.
Proof. Proof involves several steps.
First we observe that the Hardy-Littlewood maximal function given
by (1) satises equation 2). The set G = [x : M
f
(x) ] is an open
set in [, ] and has Lebsgue measure atmost
3f
1

. We assume that
>
3f
1
2
so that B = G
c
is nonempty. We write the open set G as
a possible countable union of disjoint open intervals I
j
of length r
j
centered at x
j
. Note that the end points x
j

1
2
r
j
necessarily belong to
B. The maximal inequality assures us that

j
r
j

3f
1

13
Let us dene the averages
m
j
=
1
r
j

I
j
f(y)dy
and write f in the form
f(x) = [f(x)1
B
(x) +

j
m
j
1
I
j
(x)] +

j
[f(x) m
j
]1
I
j
(x)
= g(x) +

j
h
j
(x)
We have the bounds
|m
j
|
1
r
j

I
j
|f(y)|dy
1
r
j

I
j
|f(y)|dy
2
1
2r
j

I
j
|f(y)|dy 2M
f
(x
j
r
j
) 2
Here

I
j
is the interval centered around x
j

r
j
2
of length 2r
j
. In particular
g

2. On the other hand

j
h
j

1
=

I
j
|f(y) m
j
|dy 2

I
j
|f(y)|dy 2f
1
We therefore have
g
1
3f
1
Note that the decomposition depends on . Let us write the corre-
sponding sum
u = T
k
f = T
k
g +

j
T
k
h
j
= v +

j
w
j
= v +w
We estimate the L
2
norm of v and the L
1
norm of w on large enough
set. Then use Tchebychevs inequality.
[x : |v(x)|

2
]
v
2
2

2

C
1
g
2
2

2

2C
1
g
1

2
=
6C
1
f
1

14
Let us denote by

I
j
the interval centered around x
j
of length 3r
j
and
by U =
j

I
j
. We begin by estimating w.1
U
c
1
.
w.1
U
c
1

U
c

j
|

I
j
k(x y)[f(y) m
j
]dy|dx
=

U
c

j
|

I
j
[k(x y) k(x x
j
)][f(y) m
j
]dy|dx

U
c

I
j
|k(x y) k(x x
j
)||f(y) m
j
|dydx
=

I
j
|f(y) m
j
|dy

U
c
|k(x y) k(x x
j
)|dx

I
j
|f(y) m
j
|dy

I
c
j
|k(x y) k(x x
j
)|dx

I
j
|f(y) m
j
|dy

x:|xy|2|yx
j
|
|k(x y) k(x x
j
)|dx
C
2

I
j
|f(y) m
j
|dy
2C
2
f
1
We have used here two facts. f(y) m
j
has mean zero on I
j
. If y I
j
and x

I
c
j
, then |y x| r
j
2|y x
j
|. On the other hand
(U)

I
j
) 3

(I
j
) = 3

j
r
j

9f
1

Finally we can put the pieces together.

(x : |u(x)| 2) (x : |v(x)| ) + (x : |w(x)| )

6C
1
f
1

+
9f
1

+
2C
2
f
1

or
(x : |u(x)| )
(12C
1
+ 18 + 4C
2
)f
1

=
Cf
1

15
There is one point that we should note. For the interval doubling construction
on the circle we should be sure that we do not see for instance any interval
of lenghth larger than

2
in G. This can be ensured if we take >
6f
1

. The
inequality is however satised for all because C 12.
We want to look at the special kernel k(y) =
1
y
which is not in L
1
. We
consider its truncation
k

(y) =
1
y
1
{|y|}
(y)
First we estimate the Fourier transform
|

|y|
e
iny
y
dy| = 2|

sin ny
y
dy|
= 2|

n
n
sin y
y
dy| 4 sup
0<a<
|

a
0
sin y
y
dy| C
1
Next in order to verify the condition (15) we need to estimate the following
quantity uniformly in y and .

x:|xy|>2|y|
|k

(x y) k

(x)|dx
There are three sets over which the integral does not vanish.
F
1
= {x : |x y| > 2|y|, |x y| , |x| }
F
2
= {x : |x y| > 2|y|, |x y| , |x| }
F
3
= {x : |x y| > 2|y|, |x y| , |x| }
We consider

F
1
|
1
x y

1
x
|dx

x:|xy|2|y|
|
1
x y

1
x
|dx

|z1|2
|
1
z 1

1
z
|dz
= C
3
16
It is clear that F
2
[2, 2]. Therefore

F
2
1
|x|
dx 2

dx
x
= C
4
Finally F
3
[x : |x y| 2] and works similarly. With C
2
= C
3
+ 2C
4
we
are done.
We are now ready to prove
Theorem 8. For any f L
p
the partial sums s
N
(f, x) converge to f in L
p
provided 1 < p < .
Proof. We need only prove, for 1 < p < , a bound from L
p
to L
p
, for the
partial sum operators
(T
N
f)(x) =

f(x y)k
N
(y)dy
with
k
N
(z) =
1
2
sin(N +
1
2
)z
sin
z
2
that is uniform in N. In terms of multipliers we are looking at a uniform L
p
bound for the operators dened by

k
N
(n) = 1
{|n|N}
(n)
Let us dene the operators M
k
as multiplication by e
ikx
which are isometries
in every L
p
. P
0
is the operator of projection to constants, i.e. the operator
with multiplier 1
{0}
(n) which is clearly bounded in every L
p
. Finally the
Hilbert transform S is the one with multiplier signum n. It is easy to verify
that
T
N
= M
N
1
2
[(S +I) + P
0
]M
N
M
(N+1)
]
1
2
[(S +I) + P
0
]M
(N+1)
]
This reduces the problem to proving that a single operator S is bounded on
L
p
. The kernel is calculated to be
s(z) =
1
2
cot
z
2
This can be replaced by the modied kernel
k(z) =
1
z
and we are done.
17