4

Riesz Kernels.

A natural generalization of the Hilbert transform to higher dimension is

mutiplication of the Fourier Transform by homogeneous functions of degree
0, the simplest ones being
Ri f () =

i
f ()
||

(4.1)

i
Since the functions ki () = || are bounded functions it is clear that Ri
are bounded operators from L2 (Rd ) into L2 (Rd ). On the other hand ki
are not continuous at = 0, and therefore the formal kernel Ki with the
representation

Ri f (x) =

Rd

Ki (x y)f (y)dy

(4.2)

can not be in L1 (Rd ).

Lemma 1. The kernels Ki () are given by
Ki (x) = cd

xi
|x|d+1

(4.3)

where cd is a constant depending on the dimension.

Proof. We will begin with the following calculation. For any > 0, d > > 0
i<x,>

e
Rd

If we let

Rd

ei<x,>t

Rd 0

||2
4(t+ )

d
1
2

d
1
2

e(t+ )|x| dxdt

(t + ) 2 dt
d

(4.4)

0 in equation (??)
ei<x,>

lim
0

1
1
2
e |x| dx = d
d
|x|
( 2 )
cd
= d
( 2 )

1
cd
2
e |x| dx = d
d
|x|
( 2 )

||2
4t

t 2 1 dt =

cd ( 2 )
||
( d )
2

(d)x

( )

j
2
If we let f , (x) = |x|d+2 e |x| , lim 0 f , (x) = icd ( d ) j || . Finally we
2
let > 1 1.
It is not dicult to see that for any smooth function f (x) with compact
support

(Rj f )(x) =

|y|

yj
[f (x + y) f (x)]dy +
|y|d+1

|y|

yj
f (x + y)dy
|y|d+1

j
is independent of because S |y|d+1 dy = 0 for any shell S = { 1 |y| 2 }.
It is a smooth function of x. For large x, the rst term is 0, and the second
integral can be estimated by,

Rd

C
xj yj
xj
d+1 f (y) dy
d+1
|y x|
|x|
|x|d+1

if we use that f has compact support and satises

easy to compute
Rj f () =

Rd

f (y)dy = 0. It is now

1 j
f ()
cd ||

The next step is to show that the kernels Ki satisfy condition of equation (??).
y
Let us take x, y Rd and write y = r where r = |y| and = |y| S d1 .

|xy|C|y|

(x y) (x)

|dx
d
|x|d
|xy|C|y| |x y|
(x r) (x r)

|dx
d
|x|d
|xr|Cr| |x r|
|(x r) (x)|
+
dx
|x|d
|xr|Cr
1
1
C1
|
d |dx
d
|x|
|xr|Cr |x r|
|(x r) (x)|
+
dx
|x|d
|xr|Cr

xi yi
xi
d+1 |dx =
d+1
|x y|
|x|

where (x) =

|xy|C|y|

xi
.
|x|

If we make the substitution x = rx we get

xi yi
xi
d+1 |dx C1
d+1
|x y|
|x|

1
1

|dx
d
|x |d
|x |C |x |
|(x ) (x )|
+
dx
|x |d
|x |C
|

The estimate is clearly uniform in r. If C is large enough 0 and are excluded

from the domain of integrartion. For large x we get an extra cancellation in
both the integrals to make them converge with a bound that is uniform in
. For the second integral we need only that satises a Hlder condition
o
on S d1 .
We have therefore proved the following theorem.
Theorem 4.1. If the kernel K(x) is given by
x
( |x| )

K(x) =

|x|d

(4.5)

and () satises a Hlder condition on S d1 and has mean 0 on S d1 , then

o
convolution by K denes a bounded operator from Lp (Rd ) into Lp (Rd ) for all
p in the range 1 < p < . In particular the Riesz transforms 4.1 given by
4.2 with kernels 4.3 are bounded operators in every Lp in the same range.

Sobolev Spaces.

In dealing with dierential equations we often come across solutions that do

not have the smoothness necessary to be a solution in the ordinary sense. To
illustrate it by an example, suppose we want to solve the equation
u =

u xi xi = f

(5.1)

on Rd . If d = 1 the equation reduces to uxx = f which is easy to solve. We

need only to integrate f twice, and if f has d continuous derivatives u will
have d+2 continuous derivatives. On Rd it is conceivable that each uxi xi may
be singular, but somehow the singularities cancel miraculously to produce a
much nicer f . Working formally with Fourier trnasforms
||2 u() = f()
3

and
uxixj () =

i j
f()
||2

In other words
uxi xj = Ri Rj f
It says that for 1 < p < , if f Lp we can expect u to have two derivatives
in Lp , but if f is bounded and continuous one should not expect u to have
two continuous derivatives. In fact on d = 2, one can construct a counter
example, i.e. a function f which is continuous such that the soulution u of
Poissons equation exhibits a singularity of the individual second derivatives
at 0, that of course cancel to produce a continuous f .
p
The Sobolev spaces Wk (Rd ) are dened as the space of functions u on
n1
nd
Rd such that u and all its partial derivatives Dx1 Dxd u of order n =

n1 + + nd are in Lp . We could start with C functions with compact

support on Rd and complete it in the norm
u

k,p

n1
nd
Dx 1 D x d u

(5.2)

n1 ,...nd
n=n1 ++nd k

If u Lp and Di u = Dxi u Lp u should be more regular than an Lp function.

Let us consider the operator
Au() =

1
1

(1 + ||2) 2

u()

and consider its represenation by the kernel

u(x + y)a(y)dy

(Au)(x) =
Rd

where
ei<x,>

Rd

= kd

1
2
ei<x,>et(1+|| ) dt
t
Rd 0

kd
et |x|2
1
dt
2
4t dt =
et |x| e 4t d+1
d+1 e
d1
|y|
t 2
t 2
0

cd
1 dx =

(1 + ||2) 2

a(x) = cd

decays very rapidly at , is smooth for x = 0 and has a singularity of |x|1d

near the origin for d 2 and a logarithmic singularity at 0 when d = 1.
d
In particular a() Lq for q < d1 . By Hlders inequality, A will map Lp
o
into L for p > d. If d = p > 1 the result is false. Let us take d = 2
and a nonnegative function f with compact support such that f L2 but
f (x)
dx = . We saw that Af has a singularity at 0. Let us consider
Rd |x|
u = D1 (Af ). Clearly
u

2
2

= u

2
2

=
R2

2
1

|f()|2 d f
1 + ||2

2
2

= f

2
2

1
By Youngs inequality any K Lq maps Lp Lp provided 1 p = 1 1 .
p
q
1
1
Therefore f W1,p implies f Lp so long as 1 p < d . By induction
p
1
f Wk,p implies that f W1,p implies f Lp so long as 1 p < k ).
p
d
Therefore on Rd , f Wk,p implies the continuity of f if k > d .
p
1
1
Actually one can prove a stronger result to the eect that if 1 p = d .
p
then W1,p Lp as long as 1 < p < . This requires the following theorem.

Theorem 5.1. Let Ta be the operator of convolution by the kernel |x|ad on

Rd .
(Ta f )(x) =

|y|ad f (x + y)dy

(5.3)

Rd

Then Ta is bounded from Lp to Lp provided 1 < p <

d
a

and

1
p

1
p

a.
d

Proof. First, we note that for a > 0, Ta is well dened on bounded functions
with compact support. We start by proving a weak type inequlity of the
form
[x : |(Ta f )(x)| ] C

q
p

f
q

For any choice of 1 < p < d let f Lp . We can assume without loss of
a
generality that f 0. We write
(Ta f )(x) =

|y|

|y|adf (x + y)dy +

u1 + u2
5

|y|

|y|adf (x + y)dy

and estimate u1 , u2 by
u1
u2

C1 a f

|y|

We can now pick = ( 2C2

|y|

p (ad)

1
p

dy

= C2 ad+ p f

) dap and estimate

sup u2 (x)
x

= C3
= C3
pd
dap

or

1
q

1
p

p
p

p
[x : u1 (x) ] 2p C1 a p
2

where q =

ap2
+p
dap

a.
d

Now, an application of Marcinkiewicz interpolation gives boundedness from

Lp to Lq in the same range and with the same relation between p and q.
We can also dene the fractional derivative operarors
(|D|a f )(x) =
Rd

f (x + y) f (x)
dy
|y|d+a

(5.4)

for 0 < a < 2. A calculation shows that in terms of Foirier transforms it is

multiplication by

Rd

ei< , y> 1
dy = cd,a ||a
|y|d+a

Therefore |D|a and Ta are essentially (upto a constant) inverses of each other.
If r > 0 is written as k + a, where k is a nonnegative integer and 0 a < 1,
then one denes the norm corresponding to r th derivative by
u

r,p

n k
i

n
n
D1 1 D d d u

n =k
i

n
n
D1 1 D d d u

a,p

(5.5)

This way the Sobolev spaces Wr,p are dened for fractional derivatives as
well.
6

Theorem 5.2. The inclusion map is well dened and bounded from Wr,p
1
into Ws,q provided s < r, 1 < p < q < , and 1 p rs . The extreme
q
d
1
value of q = is allowed if 1 > p rs .
q
d
Proof. We can assume without loss of generality that 0 < r s < 1. We can
go from Wr,p to Ws,q in a nite number of steps, with 0 < r s < 1 at each
step. We write I = cd,a Ta |D|a where a = r s. By denition |D|a maps
Wr,p boundedly into Ws,p . By the earlier theorem Ta maps Ws,p boundedly
into Ws,q . Although we proved it for s = 0, it is true for every s because Ta
commutes with |D|a . The cae q = is covered as well by this argument.

Generalized Functions.

Let us begin with the space W1,2 . This is a Hilbert Space with the inner
product
d

< f, g >1 =

[f g +

Rd

fxi gxi ]dx =

f hdx
Rd

where h = g d gxixi . Since g W1,2 , gxi L2 and gxixi is the derivative

1
of an L2 function. In fact since we can write f gxi dx as fxi gdx, Any
derivative of an L2 function can be thought of as a bounded linear functional
on the space W1,2 . A simlar reasoning applies to all the spaces Wr,p . The
dual space of Wr,p is Wr,q where 1 + 1 = 1.
p
q
For a function to be in Lp its singularities as well as decay at must be
controlled. We can get rid of the condition at by demaniding that f be
in Lp (K) for every bounded set K or equivalently by insisting that f Lp
for every C function with compact support. This denition makes sense
loc
for Wr,p as well. We say that f Wr,p if f Wr,p for every C function
with compact support. One needs to check that on Wr,p mutiplication by a
smooth function is a bounded linear map. One can use Leibnitzs rule if r is
an integer. For 0 < r < 1 we need the following lemma.
Lemma 2. If f Wr,p and C r with r < r 1 i.e. is a bounded
function satisfying |(x) (y)| C|x y|r , for all x, y, then f Wr,p .
Proof. We need to prove
g(x) =
Rd

(y)f (y) (x)f (x)

dy
|y x|d+r
7

is in Lp . We can write
(y)f (y) (x)f (x) = (x)[f (y) f (x)] + [(y) (x)]f (y).
The contribution of rst term is easy to control. To control the second term
it is sucient to show that
sup
x

Rd

|(y) (x)|
dy <
|y x|d+r

which is not hard. We split the integral into two regions |x y| 1 and
|xy| > 1, use the Hlder property of to obtain an estimate on the integral
o
over |x y| 1 and the boundedness of to get an estimate over |x y| > 1,
both of which are uniform in x.