3 Multidimensional Versions

The problem of convergence of Fourier Series in several dimensions is more

complicated because there is no natural truncation. If n = {n
1
, . . . , n
d
} is a
multi-index, then the sum

n
a
n
e
in.x
is natuarally computed by summing over nite stes D
N
which are allowed to
increase to Z
d
. One tries to recover the function f by
f = lim
N

nD
N
a
n
e
in.x
(1)
For smooth functions there is no problem because a
n
decays fast. The
degree of smoothness needed gets worse as dimension goes up. In d dimn-
sions we need |a
n
| to decay like |n|
d+
for some > 0 to be sure of uniform
convergence of the Fourier Series. On the other hand the orthogonality rela-
tions imply that in f L
2
, the series converges in L
2
and again D
N
can be
arbitrary. However for 1 < p < but dierent from 2 the situation is far
from clear.
If we take D
N
= {n : |n
j
| N, j = 1, . . . , d} the partial sum operator we
need to look at is convolution by

1
2

|n
j
|N
j=1,... ,d
e
i<n,x>
=
d
j=1
sin(N +
1
2
)x
j
2 sin
x
j
2
=
d
j=1
t
N
(x
j
)
The partial sum operator S
N
is therefore the product
T
N
=
d
j=1
T
N
j
where T
N
j
is the convolution in the variable x
j
by the kernel t
N
(x
j
). It is easy
to see that as operators T
N
j
have a bound that is uniform in N. The bound in
the context of a single variable extends to d variables because t
N
j
acts only on
the single variable x
j
. Therefore T
N
have a uniform bound as well. Therefore
we have with the choice of the cube D
N
= {n : |n
j
| N, j + 1, . . . d}, we
18
have convergence in L
p
of the partial sums to f, for every f L
p
provided
1 < p < .
It is known that the result is false for any p = 2 if we choose D
N
= {n :
n
2
1
+ +n
2
d
N
2
.
We now look at Fourier Transforms on R
d
. If f(x) is a function in L
1
(R
d
)
its Fourier transform

f(y) is dened by

f(y) =

R
d
e
i <x,y>
f(x)dx (2)
We denote by S the class of all functions f on R
d
that are innitely dieren-
tiable such that the function and its derivitives of all orders decay faster than
any power, i.e. for every n
1
, n
2
, . . . , n
d
0 and k 0 there are constants
C
n
1
,n
2
,... ,n
d
,k
such that
|[(
d
dx
1
)
n
1
(
d
dx
1
)
n
2
. . . (
d
dx
d
)
n
d
f](x)| C
n
1
,n
2
, ,n
d
,k
(1 +x)
k
It is easy to show by repeated integration by parts that if f S so does

f.
Theorem 1. The Fourier transform has the inverse
f(x) =

R
d
e
i <x,y>

f(y)dy (3)
proving that the Fourier transform is a one to one mapping of S onto itself.
In addition the Fourier transform extends as a unitary map from L
2
(R
d
)
onto L
2
(R
d
).
Proof. Clearly
g(x) =

R
d
e
i <x,y>

f(y)dy
is well dened as a function in S. We only have to identify it. We compute
19
g as
g(x) =

R
d
e
i <x,y>

f(y)dy
= lim
0

R
d
e
i <x,y>

f(y)e

y
2
2
dy
= lim
0

R
d

R
d
e
i <z,y>
f(z)dz

e
i <x,y>
e

y
2
2
dy
= lim
0

1
2

R
d

R
d
e
i <zx,y>
f(z)e

y
2
2
dydz
= lim
0

1
2

R
d
f(z)

R
d
e
i <zx,y>
e

y
2
2
dy

dz
= lim
0

R
d
f(z)e

zx
2
2
dz
= f(x)
Here we have used the identity
1

R
e
i xy
e

x
2
2
dx = e

y
2
2
We now turn to the computation of L
2
norm of

f. We calculate it as

f
2
2
= lim
0

R
d
|

f(y)|
2
e

y
2
2
dy
= lim
0

R
d

R
d

R
d
f(x)

f(z)e
i<xz,y>
e

y
2
2
dydxdz
= lim
0

R
d

R
d
f(x)

f(z)e

xz
2
2
dxdz
= lim
0

R
d
f(x)[K

f ](x)dx
=

R
d
|f(x)|
2
dx
20
We see that the Fourier transform is a bounded linear map from L
1
to L

as well as L
2
to L
2
with corresponding bounds C = (
1

2
)
d
and 1. By the
Riesz-Thorin interpolation theorem the Fourier transform is bounded from
L
p
into L
p
p1
for 1 p 2. If
1
p
= 1.t +
1
2
(1 t) then
1
2
(1 t) = 1
1
p
=
p1
p
.
See exercise to show that for f L
p
with p > 2 the Fourier Transform need
not exist.
For convolution operators of the form
(Tf)(x) = (k f)(x) =

R
d
k(x y)f(y)dy (4)
we want to estimate T
p
, the operator norm from L
p
to L
p
for 1 p .
As before for p = 1, ,
T
p
=

R
d
|k(y)|dy.
Let us suppose that for some constant C,
1. The Fourier transform

k(y) of k() satises
sup
yR
d
|

k(y)| C < (5)

2. In addition,
sup
xR
d

{y:xyCx}
|k(y x) k(y)|dy C < (6)
We will estimate T
p
in terms of C. The main step is to establish a
weak type (1, 1) inequality. Then we will use the interpolation theorems to
get boundedness in the range 1 < p 2 and duality to reach the interval
2 p < .
Theorem 2. The function g(x) = (Tf)(x) = (k f)(x) satises a weak type
(1, 1) inequality
{x : |g(x)| } C
0
f
1

(7)
with a constant C
0
that depends only on C.
21
We rst prove a decomposition lemma that we will need for the proof of
the theorem.
Lemma 1. Given any open set G R
d
of nite Lebesgue measure we can
nd a countable set of balls {S(x
j
, r
j
)} with the following properties. The
balls are all disjoint. G =
j
S(x
j
, 2r
j
) is the countable union of balls with
the same centers but twice the radius. More over each point of G is covered
at most 9
d
times by the covering G =
j
S(x
j
, 2r
j
). Finally each of the balls
S(x
j
, 8r
j
) has a nonempty intersection with G
c
.
Basically, the lemma says that it is possible to write G as a nearly disjoint
countable union of balls each having a radius that is comparable to the
distance of the center from the boundary.
Proof. Suppose G is an open set in the plane of nite volume. Let d(x) =
d(x, G
c
) be the distance from x to G
c
or the boundary of G. Let d
0
=
sup
xG
d(x). Since the volume of G is nite, G cannot contain any large balls
and consequently d
0
cannot be innite. We consider balls S(x, r(x)) around x
of radius r(x) =
d(x)
4
. They are contained in G and provide a covering of G as
x varies over G. All these balls have the property that S(x, 5r(x))intersects
G
c
. We select a countable subcover from this covering
xG
S(x, r(x)). We
choose x
1
such that d(x
1
) >
d
0
2
. Having chosen x
1
, . . . , x
k
the choice of x
k+1
is made as follows. We consider the balls S(x
i
, r(x
i
)) for i = 1, 2, . . . , k.
Look at the set G
k
= {x : S(x, r(x)) S(x
i
, r(x
i
)) = for 1 i k} and
dene d
k
= sup
xG
k
d(x). We pick x
k+1
G
k
such that d(x
k+1
) >
d
k
2
. We
proceed in this fashion to get a countable collection of balls {S(x
j
, r(x
j
))}.
By construction, they are disjoint balls contained in the set G of nite volume
and therefore r(x
j
) 0 as j . Since, d
j
2d(x
j+1
) 8r(x
j+1
) it must
also necessarily go to 0 as j . Every S(x
j
, 5r(x
j
)) intersects G
c
. We
now worry about how much of G they cover. First we note that G
0
G
1

G
k
G
k+1
. We claim that
k
G
k
= . If not let x G
k
for every
k. Then d
k
d(x) > 0 for every k contradicting the convergence of d
k
to 0.
Since x G
0
= G, we can nd k 1 be such that x / G
k
but x G
k1
.
Then S(x, r(x)) must intersect S(x
k
, r(x
k
)) giving us the inequality |xx
k
|
r(x) +r(x
k
)
d(x)
4
+r(x
k
)
d
k1
4
+r(x
k
)
d(x
k
)
2
+r(x
k
) =
3
2
r(x
k
). Clearly
S(x
k
, 2r(x
k
) will contain x. Since
3
2
r(x) < d(x) the enlarged ball is still within
G. This means G =
k
S(x
k
, 2r(x
k
)). Now we worry about how often a point
x can be covered by {S(x
k
, 2r(x
k
)}. Let for some k, |xx
k
| 2r(x
k
). Then
by the triangle inequality |d(x) d(x
k
)| 2r(x
k
) =
1
2
d(x
k
). This implies
22
that for the ratio
r(x)
r(x
k
)
=
d(x)
d(x
k
)
we have
1
2

r(x)
r(x
k
)

3
2
In particular any ball
S(x
j
, 2r(x
j
) that covers x, must have its center with in a distance of 4r(x)
and the corresponding r(x
j
) must be in the range
2
3
r(x) r(x
j
) 2r(x).
The balls S(x
j
, r(x
j
) are then contained in S(x, 6r(x)) are disjoint and have
a radius of atleast
2
3
r(x). There can be atmost 9
d
of them by considering the
total volume. We can choose our norm in R
d
to be max
i
|x
i
| and force the
spheres to be cubes.
Proof of theorem. The proof is similar to the one-dimensional case with some
modications.
1. We let G

be the open set where the maximal function M

f
(x) satises
|M
f
(x)| > . From the maximal inequality
[G

] C
f
1

(8)
2. We write G

=
j
B
j
=
j
S(x
j
, 2r
j
), a countable union of cubes ac-
cording to the lemma.
3. If we let
(x) =

j
1
B
j
(x)
then 1 (x) 9
d
on G

.
4. Let us dene a weighted average m
j
of f(y) on B
j
by

B
j
[f(y) m
j
]
dy
(y)
= 0 (9)
and write
f(x) = f(x)1
G
c

(x) +
1
(x)

j
f(x)1
B
j
(x)
= f(x)1
G
c

(x) +
1
(x)

j
m
j
1
B
j
(x) +
1
(x)

j
[f(x) m
j
]1
B
j
(x)
= h
0
(x) +

j
h
j
(x) (10)
23
5. For any cube B
j
with center x
j
there is a cube with 4 times its size and
with the same center that contains a point x

j
G
c

with |M
f
(x

j
)|
. The cube S(x

j
, 10r
j
) contains B
j
. Therefore with some constant
depending only on the dimension
|m
j
| C
d
(11)
Moreover on G
c

, |f(x)| M
f
(x) . Hence
h
0

+C
d
= (C
d
+ 1) (12)
On the other hand
h
0

1
f
1
+C
d

j
[B
j
]
f
1
+C
2
d
[G

]
(1 +CC
2
d
)f
1
(13)
and therefore
h
0

2
2
(C
d
+ 1)h
0

1
C
1
f
1
(14)
From the boundedness of T from L
2
to L
2
this gives
{x : |(Th
0
)(x)| } C
2
f
1

(15)
6. We now turn our attention to the functions {h
j
}
w = T[

j
h
j
] =

B
j
[f(y) m
j
]k(x y)
dy
(y)
=

B
j
[f(y) m
j
][k(x y) k(x x
j
)]
dy
(y)

B
j
|f(y) m
j
||k(x y) k(x x
j
)|dy (16)
We estimate |w(x)| for x /
j
U
j
where U
j
is the cube with the same
center x
j
as B
j
but enlarged by a factor C + 1. In particular if y B
j
24
and x U
c
j
, then |y x| |x x
j
| |y x
j
| C|y x
j
|.

j
U
c
j
|w(x)|dx

j
U
c
j
[

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

B
j
|f(y) m
j
|[

E
j
|k(x y) k(x x
j
)|dx]dy
(17)
where E
j
{x : |x y| C|y x
j
|}. Therefore,

E
j
|k(x y) k(x x
j
)|dx
sup
y,j

{x:|xy|C|yx
j
|}
|k(x y) k(x x
j
)|dx
sup
y

{x:|xy|C|y|}
|k(x y) k(x)|dx
C (18)
giving us the estimate

j
U
c
j
|w(x)|dx C

B
j
|f(y) m
j
|dy
C(f
1
+ [sup
j
m
j
]

j
[B
j
])
C
1
f
1
(19)
7. We put the pieces together and we are done.
25