Tel Aviv University, Fall 2004

Lattices in Computer Science

Lecture 9
Fourier Transform
Lecturer: Oded Regev
Scribe: Gillat Kol
In this lecture we describe some basic facts of Fourier analysis that will be needed later. The rst section
discusses the Fourier transform, and the second discusses the Fourier series. We start each section with the
more familiar case of one-dimensional functions and then extend it to the higher dimensional case. As a
general rule, we will not worry too much about issues of convergence, differentiability etc., as these will
always be satised in our applications.
1 Fourier Transform
1.1 The one-dimensional case
DEFINITION 1 We dene L
1
(R) as the set of functions f : R C satisfying
_

|f(x)|dx < .
DEFINITION 2 For a function f L
1
(R) dene its Fourier transform as the function

f : R C given by

f(y) =
_

f(x)e
2ixy
dx
For example, the Fourier transform at point 0 is

f(0) =
_

f(x)dx.
EXAMPLE 1 Dene
f(x) =
_
1 if |x| < a
0 otherwise
.
Then,

f(y) =
_
a
a
e
2ixy
dx =
e
2iay
e
2iay
2iy
=
sin(2ay)
y
.
a
1
a
Figure 1: f(x) and

f(y)
EXAMPLE 2 Let f(x) = e
(
x
s
)
2
for some s > 0. Then,

f(y) =
_

e
(
x
s
)
2
e
2ixy
dx =
_

e
(
x
2
s
2
+2ixy)
dx
=
_

e
(
x
s
+iys)
2
e
(ys)
2
dx = e
(ys)
2
_

e
(
x
s
+iys)
2
dx.
We now perform a (complex) change of variable z =
x
s
+iys (which is possible by Cauchys theorem), and
see that the above is equal to
s e
(ys)
2
_

e
z
2
dz = s e
(ys)
2
.
Notice that for s = 1 we get that f(x) = e
x
2
satises

f = f.
1
1
s
1
s
s
Figure 2: f(x) and

f(y)
The following theorem lists some of the most important properties of the Fourier transform. The rst
property shows that the Fourier transform is linear. The third and fourth properties show that under the
Fourier transform, translation becomes multiplication by phase and vice versa. The sixth property shows
that scaling a function by some > 0 scales its Fourier transform by 1/ (together with the appropriate
normalization). The seventh property shows that under the Fourier transform, convolution becomes multipli-
cations and vice versa, where we dene the convolution of two functions as f g(y) =
_
R
f(x)g(y x)dx.
The last property shows that the Fourier transform of the derivative of a function can be obtained by simply
multiplying the Fourier transform of the function by 2iy.
THEOREM 1 For all f, g L
1
(R), x, y, z R, the following holds:
1.

f + g =

f + g and for all C,

(f) =

f
2. if

f is the complex conjugate of f then

(

f)(y) =

f(y)
3. if h(x) := f(x + z) then

h(y) =

f(y) e
2izy
4. if h(x) := e
2izx
f(x) then

h(y) =

f(y z)
5. |

f(y)|
_

|f(x)|dx
6. > 0, dene h(x) := f(x) then

h(y) =

f(
y

)
7.

f g =

f g and

f g =

f g
8. if h(x) = f

(x) L
1
(R) then

h(y) = 2iy

f(y)
PROOF: Most items are easy to verify. We only include a proof of two of them.
3.

h(y) =
_

f(x + z) e
2ixy
dx =
_

f(x)e
2i(xz)y
dx = e
2izy

f(y)
8. Since f

L
1
(R), lim
x
f(x) = f(0) +
_

0
f

(x)dx exists, and since f L

1
(R) this limit must
be zero. Similarly lim
x
f(x) = 0. Hence, using integration by parts
_

(x)e
2ixy
dx =
_

(2iy)f(x)e
2ixy
dx = 2iy

f(y)
2
2
The following theorem, known as the inversion formula, shows that a function can be recovered from its
Fourier transform. The proof is omitted.
THEOREM 2 If both f,

f L
1
(R) and f is continuous then f(x) =
_

f(y)e
2ixy
dy
1.2 The n-dimensional case
We now extend the Fourier transform to functions on R
n
. Dene L
1
(R
n
) as the set of functions f : R
n
C
satisfying
_
R
n
|f(x)|dx < . We also dene, for x, y R
n
, x, y =

n
i=1
x
i
y
i
and x =
_
x, x. We
now dene the n-dimensional Fourier transform.
DEFINITION 3 For f L
1
(R
n
) dene

f : R
n
C by

f(y) =
_
R
n
f(x)e
2ix,y
dx
As shown below, all properties listed in Theorem 1 can be extended to the n-dimensional case. The
proof is essentially the same as that in the one-dimensional case. The only new property is the last one: it
says that if an n-dimensional function can be factored as the product of n one-dimensional functions, then
its Fourier transform is the product of the individual Fourier transforms. The proof of this is left to the
reader.
THEOREM 3 For all f, g L
1
(R
n
), x, y, z R
n
, the following holds:
1.

f + g =

f + g and for all C,

(f) =

f
2. if

f is the complex conjugate of f then

(

f)(y) =

f(y)
3. if h(x) := f(x + z),

h(y) = e
2iy,z

f(y)
4. if h(x) := e
2ix,z
f(x) then

h(y) =

f(y z)
5. |

f(y)|
_
R
n
|f(x)|dx
6. for > 0, h(x) :=
n
f(x) then

h(y) =

f(
y

)
7.

f g =

f g and

f g =

f g
8. if
f
x
j
exists then

_
f
x
j
_
(y) = 2iy
j

f(y)
9. if f(x) = f
1
(x
1
) f
n
(x
n
) then

f(y) =

f
1
(y
1
)

f
n
(y
n
)
The following example will be used in future lectures.
EXAMPLE 3 Consider (x) := e
x
2
. Then
(x) = e
(x
2
1
++x
2
n
)
= e
x
2
1
e
x
2
n
hence we obtain that (y) = (y). More generally, for
s
(x) := e

x
s

2
,
s
(y) = s
n
1
s
(y).
We also have the following extension of the inversion formula (we omit the exact smoothness conditions
required from f).
THEOREM 4 For f,

f L
1
(R
n
), f(x) =
_
R
n

f(y)e
2ix,y
dy
3
2 The Fourier Series
2.1 The one-dimensional case
In this subsection, we consider functions on R with period 1, i.e., functions f that satisfy that f(x + y) =
f(x) for any x R, y Z.
DEFINITION 4 For a function f : R C with period 1 we dene
1
its Fourier series as the function

f : Z C given by

f(k) =
_
1
0
f(x)e
2ikx
dx.
The value

f(k) is sometimes called the kth Fourier coefcient.
Notice that unlike the Fourier transform, the Fourier series is only dened on a discrete set of points,
namely Z. The intuitive reason for this is that in a 1-periodic function, only integer frequencies appear.
Moreover, it is interesting to note that the Fourier coefcients can be seen as the limit of the Fourier transform
in the following sense. Consider a periodic function on R and restrict it to [R, R]. Then, roughly speaking,
as R goes to , its Fourier transform converges to 0 on non-integer points and to the Fourier coefcients on
integer points.
Most of the properties of the Fourier transform given in Theorem 1 also hold for the Fourier series. We
mention some below.
THEOREM 5 For any two functions f, g with period 1 we have
1.

f + g =

f + g and for any C,

f =

f
2. if h(x) := f(x + r) for some r R, then

h(k) =

f(k) e
2ikr
3. if h(x) := e
2ijx
f(x) for some j Z then

h(k) =

f(k j)
The following is the inversion formula for the Fourier series (also known as the Fourier convergence
theorem). Notice that in the case that f is continuous, the right hand side is simply f(x).
THEOREM 6 For any piecewise smooth f with period 1 we have

k=

f(k)e
2ikx
=
1
2
(f(x+) + f(x)) .
The following theorem is known as the Poisson summation formula. Its proof is based on a connection
between the Fourier transform and the Fourier series.
THEOREM 7 (PSF
1
) For a nice enough f L
1
(R),

j=
f(j) =

j=

f(j).
Equivalently, f(Z) =

f(Z).
1
To be precise, we should assume that

1
0
|f(x)|dx exists. From now on, we ignore such issues of convergence.
4
PROOF: Given a function f, dene (t) =

j=
f(t + j). Notice that has period 1, and we can
therefore consider its Fourier series,
(k) =
_
1
0
(t)e
2ikt
dt =

j=
_
1
0
f(t + j)e
2ikt
dt
=

j=
_
1
0
f(t + j)e
2ik(t+j)
dt
=
_

f(t)e
2ikt
dt =

f(k).
So, we see that is the restriction of

f to the integers. Using the inversion formula we have
(0) =

j=
(j),
and the theorem follows. 2
In the rest of this subsection, we extend our denition of the Fourier series to functions whose period
is not necessarily 1. It should be noted that this extension is not strictly necessary in the sense that any
function f with period can be transformed into a function with period 1 by simply dening g(x) := f(x).
Nevertheless, it serves as a good introduction to Fourier series on lattices since what we are doing here is
essentially dening the Fourier series of functions that are periodic on an arbitrary one-dimensional lattice
Z (whereas so far we only dealt with the lattice Z).
For a function f : R C with some period > 0 we dene its Fourier series as

f :
1

Z C by

f(y) =
1

_

0
f(x)e
2ixy
dx.
The inversion formula becomes the following.
THEOREM 8
f(x) =

y
1

f(y)e
2ixy
We now obtain the following extension of the Poisson summation formula (we remark that this extension
can also be derived directly from Theorem 7).
LEMMA 9 (PSF
2
) For any > 0 and any nice enough function f,

xZ
f(x) =
1

yZ/

f(y).
Equivalently, f(Z) =
1

f(
1

Z).
5
PROOF: Dene (x) =

j=
f(x + j). Then has period and for y
1

Z,
(y) =
1

_

0
(x)e
2ixy
dx
=
1

j=
_

0
f(x + j)e
2ixy
dx
=
1

f(x)e
2ixy
dx
=
1

f(y).
By the inversion formula, we have
(0) =

y
1

Z
(y).
2
EXAMPLE 4 For f(x) = e
x
2
, we obtain that for any > 0,

j=
e
j
2
=
1

j=
e

2
.
2.2 The n-dimensional case
In this subsection, we extend the denition of the Fourier series to the n-dimensional case. We start by
considering the Fourier series of functions on R
n
that are Z
n
-periodic, that is, functions f : R
n
C such
that f(x + y) = f(x) for any x R
n
, y Z
n
.
DEFINITION 5 For a Z
n
-periodic function f dene its Fourier series

f : Z
n
C as

f(y) =
_
[0,1)
n
f(x)e
2ix,y
dx
THEOREM 10 For a nice enough f we have that for all x
f(x) =

yZ
n

f(y)e
2ix,y
LEMMA 11 (PSF
3
) For a nice enough f we have f(Z
n
) =

f(Z
n
)
We would now like to extend the above to functions that are -periodic for some full-rank lattice .
Notice that we already did a similar thing in the previous subsection for one-dimensional lattices. Indeed,
we started with Z-periodic functions and then extended our discussion to -periodic functions for any one-
dimensional lattice . The Fourier series of Z-periodic functions was dened as a function on the dual
lattice
1

Z. Moreover, in Lemma 9 we proved that f() = det(

)

f(

) for any one-dimensional lattice

.
Let B be a basis of some full-rank lattice and let f be a -periodic function, i.e., a function f : R
n

C such that f(x + y) = f(x) for any x R

n
, y . The Fourier series of f is the function

f :

C
given by

f(y) =
1
det()
_
P(B)
f(x)e
2ix,y
dx.
6
As the following exercise shows, this denition is independent of the choice of basis B, and is therefore
well-dened.
EXERCISE 1 Show that if g is a -periodic function for some lattice = L(B), then its integral on P(B)
is the same for any choice of basis B. (A possible approach is to show that the integral is invariant under the
basic operations and then use the fact that a basis can be transformed into any other basis using a sequence
of basic operations.) Deduce that

f is well-dened.
The inversion formula is now of the following form.
THEOREM 12 For a nice enough f we have that for all x
f(x) =

f(z)e
2ix,z
Finally, we have the following general formulation of the Poisson summation formula. This formulation
will be often used is future lectures.
LEMMA 13 (PSF
4
) For a nice enough f and any full-rank lattice , f() = det(

)

f(

).
PROOF: The function (x) =

z
f(x + z) is -periodic and hence we can consider its Fourier series.
For any y

we have
(y) =
1
det()
_
P(B)

z
f(x + z)e
2ix,y
dx
=
1
det()

z
_
P(B)
f(x + z)e
2ix,y
dx
=
1
det()

z
_
P(B)
f(x + z)e
2ix+z,y
dx
=
1
det()
_
R
n
f(x)e
2ix,y
dx
= det(

)

f(y),
where we used that z, y Z. By the inversion formula, (0) =

y
(y). 2
Let us remark that it is possible to derive Lemma 13 directly from Lemma 11 by using the fact that if
f : R
n
C is a -periodic function for some lattice with basis B, then the function g : R
n
C given
by g(x) = f(Bx) is Z
n
-periodic.
EXAMPLE 5 Applying this to the function dened in Example 3, we obtain that () = det(

)(

).
More generally, we obtain
s
() = s
n
det(

)
1/s
(

).
7