Lectures on

FOURIER TRANSFORM

BY

S. THANGAVELU

1. Fourier transform: Basic properties

In this section we show that there exists a remarkable unitary oper-

ator F on the Hilbert space L2 (R) which we call the Fourier transform
and study some of the basic properties of that operator.

1.1. Unitary operators: some examples. We begin with some def-

initions. Given two Hilbert spaces H1 and H2 consider a bounded linear
operator T : H1 → H2 . We define its adjoint, denoted by T ∗ the unique
operator from H2 into H1 determined by the condition

(T u, v)2 = (u, T ∗ v)1 , u ∈ H1 , v ∈ H2

where (·, ·)j stand for the inner product in Hj . Note that T ∗ is bounded.
We say that T is unitary if T T ∗ = I2 , T ∗ T = I1 where Ij is the identity
operator on Hj . If T is unitary then we have (u, v)1 = (T u, T v)2 for all
u, v ∈ H1 . In particular kT uk2 = kuk1 , u ∈ H1 .
We give some examples of unitary operators. Let H1 = L2 (S 1 ) and
H2 = L2 (Z). Take T to be the operator T f (k) = fˆ(k) where
Z 1
fˆ(k) = f (t)e−2πikt dt.
0
∗
Then it can be checked that T is given by
∞
X
∗
T ϕ(t) = ϕ(k)e−2πikt .
−∞

The Plancherel theorem for the Fourier series shows that T is unitary.
Another simple example is provided by the translation τa f (x) = f (x −
a) defined from L2 (R) into itself. We give some more examples below.
1
2 THANGAVELU

Let us take the nonabelian group H1 which is R × R × R with the

group law

(x, y, t)(x0 , y 0 , t0 ) = (x + x0 , y + y 0 , t + t0 + xy 0 ).

Then it is clear that H1 is nonabelian and the Lebesgue measure dxdydt

is both left and right invariant Haar measure on H1 . With this measure
we can form the Hilbert space L2 (H1 ). Let Γ = Z × Z × Z. Then it
is easy to check that Γ is a subgroup of H1 so that we can form the
quotient M = Γ/H1 consisting of all right cosets of Γ. Functions on
M are naturally identified with left Γ−invariant functions on H1 . As
the Lebesgue measure dxdydt is left Γ− invariant we can form L2 (M )
using the Lebesgue measure restricted to M. As a set we can identify
M with [0, 1)3 and we just think of L2 (M ) as L2 ([0, 1)3 ).
Fourier expansion in the last variable allows us to decompose L2 (M )
into a direct sum of orthogonal subspaces. Simply define Hk to be the
set of all f ∈ L2 (M ) which satisfy the condition

f (x, y, t + s) = e2πiks f (x, y, t).

Then Hk is orthogonal to Hj whenever k 6= j and any f ∈ L2 (M ) has

the unique expansion
∞
X
f= fk , f k ∈ Hk .
k=−∞

We are mainly interested in H1 which is a Hilbert space in its own

right.(why?) It is interesting to note that functions in H1 are also
invariant under the left action of Γ.
Our next example of a unitary operator is the following. Consider
the map J : H1 → H1 given by J(x, y, t) = (y, −x, t − xy). Then J is
an automorphism of the group H1 which satisfies (i)J 4 = I,(ii) J(Γ) =
Γ(i.e. J leaves Γ invariant) and (iii) J restricted to the center of H1 is
just the identity; i.e. J(0, 0, t) = (0, 0, t). Using this automorphism we
define an operator, denoted by the same symbol, on H1 by

Jf (x, y, t) = f (J(x, y, t)) = f (y, −x, t − xy).

 FOURIER TRANSFORM 3

It is clear that J ∗ f (x, y, t) = f (−y, x, t − xy) so that J is unitary. We

also observe that J 2 f (x, y, t) = f (−x, −y, t).
We now define another very important unitary operator which takes
2
L (R) onto H1 . This operator used by Weil and Brezin is called the
Weil- Brezin transform and is defined as follows. For f ∈ L2 (R),
∞
X
2πit
V f (x, y, t) = e f (x + n)e2πiny .
n=−∞

As f ∈ L2 (R) we know that f (x + n) is finite for almost every x ∈ R.

The above series converges in L2 ([0, 1)) as a function of y and we have
Z 1 X∞ Z 1
2
|V f (x, y, t)| dy = |f (x + n)|2 .
0 n=−∞ 0

Thus it follows that V f ∈ H1 and

Z Z ∞
2
|V f (x, y, t)| dxdydt = |f (x)|2 dx.
[0,1)3 −∞

Proposition 1.1. V is a unitary operator from L2 (R) onto H1 .

To prove this proposition we need to calculate V ∗ . It is clear that V

is one to one but is also onto. To see this, given F ∈ H1 consider f
defined as follows. For x ∈ [m, m + 1) define
Z 1
f (x) = F (x − m, y, 0)e−2πimy dy.
0

Then it is clear that f ∈ L2 (R) and

∞
X Z 1 
2πit −2πimu
V f (x, y, t) = e F (x, u, 0))e du e2πimy = F (x, y, t).
m=−∞ 0

Moreover, if f, g ∈ L2 (R) then

X∞ Z 1
(f, g) = f (x + m)g(x + m)dx.
m=−∞ 0

The sum is nothing but

Z 1
V f (x, y, t)V g(x, y, t)dy
0
4 THANGAVELU

and hence we have (f, g) = (V f, V g). This shows that V ∗ = V −1 and

hence V is unitary.

1.2. Fourier transform: Plancherel and inversion theorems.

Definition 1.2. The unitary operator V ∗ JV from L2 (R) onto itself is

called the Fourier transform and is denoted by F .

We record some important properties of the Fourier transform in the

following theorem.

Theorem 1.3. The Fourier transform F satisfies: (i) F 4 f = f, for

every f ∈ L2 (R) (ii) F 2 f (x) = f (−x) for almost every x ∈ R and (iii)
kF f k2 = kf k2 .

We only need to check (ii) as (i) follows immediately since J 4 = I.

As J 2 f (x, y, t) = f (−x, −y, t) we have
Z 1
2
F f (x) = V f (−x − m, −y, t)e−2πimy dy
0

whenever x ∈ [m, m + 1). If we recall the definition of V f the above is

simply f (−x).
The property (iii), namely kF f k2 = kf k2 is called the Plancherel
theorem for the Fourier transform.
Before proceeding further let us calculate the Fourier transforms of
some well known functions. As our first example let us take the Gauss-
2
ian ϕ(x) = e−πx .

Proposition 1.4. The Fourier transform of ϕ is itself: F ϕ = ϕ.

Proof. By definition, when x ∈ [0, 1),

Z 1 X∞
F ϕ(x + m) = ϕ(y + n)e−2πi(y+n)(x+m) dy
0 n=−∞

which can be rewritten as

Z 1 ∞
X
−π(x+m)2 2
e e−π(y+n+i(x+m)) dy.
0 n=−∞
 FOURIER TRANSFORM 5

We claim that the integral is a constant. To see this, note that

Z 1 X∞ Z ∞
−π(y+n+iw)2 2
G(w) = e dy = e−π(y+iw) dy
0 n=−∞ −∞

is an entire function of w = u + iv and G(iv) = G(v). Hence G(x + m)

is a constant and we get F ϕ = cϕ. But G(0) = 1 and so c = 1 proving
the proposition. 

The above proposition shows that the Gaussian ϕ is an eigenfunc-

tion of the Fourier transform. We will say more about the spectral
decomposition of F in the next section.
We introduced the Fourier transform as a unitary operator on L2 (R).
Now we extend the definition to L1 (R) and prove a useful inversion
formula.

Theorem 1.5. For f ∈ L1 (R) ∩ L2 (R) the Fourier transform is given

by Z
F f (ξ) = f (x)e−2πixξ dx.
R
1
If we further assume that F f ∈ L (R) then for almost every x we have
Z
f (x) = F f (ξ)e2πixξ dξ.
R

Proof. If ξ = x + m, x ∈ [0, 1) it follows from the definition that

Z 1 X ∞
F f (ξ) = f (y + n)e−2πi(y+n)(x+m) dy.
0 n=−∞

As f is integrable we can interchange the order of summation and

integration to arrive at the formula
Z
F f (ξ) = f (x)e−2πixξ dx.
R

Under the assumption that F f is also integrable the inversion formula

F 2 f (x) = f (−x) leads to
Z
f (x) = F f (ξ)e2πixξ dξ.
R

This completes the proof of the theorem. 

6 THANGAVELU

It is customary to denote the Fourier transform F f of integrable

functions by fˆ. Thus
Z
ˆ
f (ξ) = f (x)e−2πixξ dx.
R

It is clear from this that the Fourier transform can be defined on all of
L1 (R). Note that fˆ for f ∈ L1 (R) is a bounded function and
Z
|fˆ(ξ)| ≤ |f (x)|dx = kf k1 .
R

It can be easily checked, by an application of the Lebesgue dominated

convergence theorem, that fˆ is in fact continuous. But something more
is true. The following result is known as the Riemann-Lebesgue lemma
in the literature.

Theorem 1.6. For all f ∈ L1 (R), fˆ vanishes at infinity; i.e.,fˆ(ξ) → 0

as |ξ| → ∞.

Proof. As L1 (R) ∩ L2 (R) is dense in L1 (R) it is enough to prove the

result for f ∈ L1 (R) ∩ L2 (R). Recall that fˆ(x + m), x ∈ [0, 1) is the
m−th Fourier coefficient of the integrable periodic function
∞
X
F (y) = f (y + n)e−2πix(y+n)
n=−∞

and hence it is enough to show that the Fourier coefficients of an in-

tegrable function vanish at infinity. It is clearly true of trigonometric
polynomials and as they are dense in L1 ([0, 1)) the same true for all
integrable functions. 

Another immediate consequence of our definition of the Fourier

transform is the so called Poisson summation formula. If the inte-
grable function f satisfies the estimate |f (y)| ≤ C(1 + y 2 )−1 then the
series defining V f (x, y, t) converges uniformly. The same is true of V fˆ
if fˆ also satisfies such an estimate. For such functions we have the
following result.
 FOURIER TRANSFORM 7

Theorem 1.7. Assume that f is measurable and satisfies |f (y)| ≤

C(1 + y 2 )−1 and |fˆ(ξ)| ≤ C(1 + ξ 2 )−1 . Then
∞ ∞
fˆ(n).
X X
f (n) =
n=−∞ n=−∞

Proof. Since fˆ = V ∗ JV f we have JV f (x, y, t) = V fˆ(x, y, t). As both

series defining JV f and V fˆ converge uniformly we can evaluate them
at (0, 0, 0) which gives the desired result. 
1 1
When we take f (x) = t− 2 ϕ(t− 2 x) for t > 0, it follows that
fˆ(ξ) = ϕ(t 2 ξ) and hence Poisson summation formula gives the in-
1

teresting identity
∞ ∞
− 21
X X
−πtn2 −1 2
e =t e−πt n .
n=−∞ n=−∞

We can obtain several identities of this kind by considering eigenfunc-

tions of the Fourier transform.

1.3. Spectral decomposition of F . The spectrum of the Fourier

transform F is contained in the unit circle as F is unitary. Moreover,
as F 4 = I any λ in its spectrum σ(F ) satisfies λ4 = 1. Hence, σ(F ) =
{1, −1, i, −i}. In this subsection we describe explicitly the orthogonal
projections associated to each point of the spectrum.
We have identified at least one eigenfunction, namely the Gaussian.
2
Let us search for eigenfunctions of the form f (x) = p(x)e−πx where p
is a real valued polynomial. The reason is the following: the Fourier
transform of such a function is given by,
Z 1 X∞
−π(x+m)2 2
F f (x + m) = e p(y + n)e−π(y+n+i(x+m)) dy
0 n=−∞

for x ∈ [0, 1). As before we are led to consider the function

Z ∞
2
G(w) = p(y)e−π(y+iw) dy
−∞

which is entire. For v ∈ R,

Z ∞
2
G(iv) = p(y + v)e−πy dy
−∞
8 THANGAVELU
2
is a polynomial and hence F f (x) = q(x)e−πx for another polynomial
q. So it is reasonable to expect eigenfunctions among this class of func-
tions. Let us record this in the following.

Proposition 1.8. Let p be a polynomial with real coefficients. Then

2
f (x) = p(x)e−πx is an eigenfunction of F with eigenvalue λ if and
only if Z ∞
2
p(x − iy)e−πx dx = λp(y).
−∞

From the above equation we can infer several things. Calculating

the the derivatives at the origin we have
Z ∞
k 2
(−i) p(k) (x)e−πx dx = λp(k) (0).
−∞

Since p is real valued and λ = (−i)n for n = 0, 1, 2, 3 the degree

m of p should satisfy the condition (−i)m+n is real. This means m
should be odd (even) whenever n is odd (resp. even). Moreover, since
F 2 f (x) = f (−x) we infer that the polynomials p corresponding to real
(imaginary) eigenvalues are even (resp. odd) functions. If we assume
that each p is monic then we also get that p should be of degree 4k + n
if λ = (−i)n , n = 0, 1, 2, 3. With these preparations we can easily show
the existence of eigenfunctions of the Fourier transform.

Theorem 1.9. There exist monic polynomials pk of degree k ∈ N such

2
that pk (x)e−πx is an eigenfunction of F with eigenvalue (−i)k .

Proof. We consider only the case of λ = 1. The other cases can be

treated similarly. In this case we have to find polynomials p of degree
4k such that Z ∞
2
p(x − iy)e−πx dx = p(y).
−∞
This leads to the equations
4k 4k
X 1 j
X 1 (j)
Cj (−iy) = p (0)y j
j=0
j! j=0
j!
where Z ∞
2
Cj = p(j) (x)e−πx dx.
−∞
 FOURIER TRANSFORM 9

As p is a function of x2 it follows that Cj = 0 unless j is even. Thus

we are led to the equations
Z ∞
(2j) j 2
p (0) = (−1) p(2j) (x)e−πx dx.
−∞

These equations can be solved recursively starting with p(4k) (0) = (4k)!.
The details are left to the reader. 

The polynomials whose existence is guaranteed by the above theorem

are called the Hermite polynomials and denoted by Hk (x). We define
2
the Hermite functions ϕk (x) = ck Hk (x)e−πx with suitably chosen ck
so as to make kϕk k2 = 1. The importance of the Hermite functions lie
in the following theorem.
Theorem 1.10. The Hermite functions ϕk , k ∈ N form an orthonor-
mal basis for L2 (R).
Proof. Here we only prove that they form an orthonormal system. The
completeness will be proved later. Since (f, g) = (fˆ, ĝ) for all f, g ∈
L2 (R) it follows that
(ϕ4k+n , ϕ4j+m ) = (−i)n−m (ϕ4k+n , ϕ4j+m )
for any n, m ∈ {0, 1, 2, 3}. Thus (ϕ4k+n , ϕ4j+m ) = 0 whenever n 6=
m. This argument does not prove the orthogonality within the same
eigenspace.
d2
2 2
Consider the operator H = − dx 2 + 4π x . By integration by parts

we can easily verify that F (Hf ) = H(F f ) for all functions of the form
2
f (x) = p(x)e−πx with p polynomial. The above shows that Hϕk is
an eigenfunction of F with the same eigenvalue and hence there are
constants λk such that Hϕk = λk ϕk . If we can show that λk 6= λj for
k 6= j then we are done. To check this the equation Hf = λk f for
2
f (x) = p(x)e−πx reduces to
−p(2) (x) + 4πxp(1) (x) + 2πp(x) = λk p(x).
By direct calculation one can show that this equation has a degree k
polynomial solution only for certain discrete values of λk and the λk
are all distinct. The details are left to the reader. 
10 THANGAVELU

We are now ready to state the explicit spectral decomposition of

F . For j = 0, 1, 2, 3 define L2j (R) be the subspace of L2 (R) for which
{ϕ4k+j : k ∈ N} is an orthonormal basis. Let Pj stand for the orthog-
onal projection of L2 (R) onto L2j (R).

Theorem 1.11. We have L2 (R) = ⊕3j=0 L2j (R). Every f ∈ L2 (R) can
P3
be uniquely written as f = j=0 Pj f. The projections are explicitly
given by the Hermite expansion
∞
X
Pj f = (f, ϕ4k+j )ϕ4k+j .
k=0

We conclude this section with the following generalisation of the

Poisson summation formula.

Theorem 1.12. Let f = ϕ4k+j be any of the Hermite functions. Then

we have
∞
X ∞
X
−2πix(y+n) j
f (y + n)e = (−i) f (x + n)e2πiny .
n=−∞ n=−∞

Proof. As F = V ∗ JV the equation F f = (−i)j f translates into

JV f (x, y, t) = (−i)j V f (x, y, t). This proves the theorem. 

2. Invariant subspaces of L2 (Rn )

In this section we define the Fourier transform for functions on Rn

and study some subspaces of L2 (Rn ) that are left invariant by the
Fourier transform.

2.1. Fourier transform on Rn . In the previous section we defined

the Fourier transform on functions defined on R. We can easily extend
the definition to functions on Rn . Instead of H1 we consider the (2n+1)
dimensional group Hn which is Rn × Rn × R with the group law
(x, y, t)(x0 , y 0 , t0 ) = (x + x0 , y + y 0 , t + t0 + x · y 0 ).
Let Γ = Zn × Zn × Z. Then Γ is a subgroup of Hn so that we can form
the quotient M = Γ/Hn consisting of all right cosets of Γ. Functions on
M are naturally identified with left Γ−invariant functions on Hn . As
 FOURIER TRANSFORM 11

the Lebesgue measure dxdydt is left Γ− invariant we can form L2 (M )

using the Lebesgue measure restricted to M. As a set we can identify
M with [0, 1)2n+1 and we just think of L2 (M ) as L2 ([0, 1)2n+1 ).
As before Fourier expansion in the last variable allows us to decom-
pose L2 (M ) into a direct sum of orthogonal subspaces. Simply define
Hk to be the set of all f ∈ L2 (M ) which satisfy the condition

f (x, y, t + s) = e2πiks f (x, y, t).

Then Hk is orthogonal to Hj whenever k 6= j and any f ∈ L2 (M ) has

the unique expansion
∞
X
f= fk , f k ∈ Hk .
k=−∞

We are mainly interested in H1 which is a Hilbert space in its own

right.
Define J as in the one dimesional case and let V : L2 (Rn ) → H1 be
defined by
X
V f (x, y, t) = e2πit f (x + m)e2πim·y .
m∈Zn
Then V is unitary and we simply define F = V ∗ JV. It is then clear that
F is a unitary operator on L2 (Rn ). All the results proved for functions
on R remain true now. For example, for f ∈ L1 (Rn ) we have
Z
ˆ
F f (ξ) = f (ξ) = f (x)e−2πix·ξ .
Rn

The eigenfunctions of F are obtained by taking tensor products of the

one dimensional Hermite functions.
For each α ∈ Nn we define

Φα (x) = ϕα1 (x1 )ϕα2 (x2 )...ϕα1 (xn ).

Then it follows that F Φα = (−i)|α| Φα where |α| = nj=1 αj . Moreover,

P

{Φα : α ∈ Nn } is an orthonormal basis for L2 (Rn ). Thus every f ∈

L2 (Rn ) has an expansion
X
f= (f, Φα )Φα
α∈Nn
12 THANGAVELU

the series being convergent in the L2 sense. Defining Pk f =

P
|α|=k (f, Φα )Φα we can write the above in the compact form f =
P∞ k
k=0 Pk f. Note that F (Pk f ) = (−i) Pk f.

2.2. The Schwartz space. Our first example of an invariant subspace

of L2 (Rn ) is provided by the the space of Schwartz class functions. First
we make

Definition 2.1. A subspace W of L2 (Rn ) is said to be invariant under

F if F f ∈ W whenever f ∈ L2 (Rn ).

As {Φα : α ∈ Nn } is an orthonormal basis for L2 (Rn ) it follows

that f ∈ L2 (Rn ) if and only if α∈Nn |(f, Φα )|2 < ∞. Since (F f, Φα ) =
P

(−i)|α| (f, Φα ) any subspace of L2 (Rn ) defined in terms of the behaviour

of |(f, Φα )| will be invariant under the Fourier transform. We can define
a whole family of invariant subspaces. Indeed, for each s > 0 define
WHs (Rn ) to be the subspace of L2 (Rn ) consisting of functions f for
which
X
kf k22,s = (2|α| + n)s |(f, Φα )|2 < ∞.
α∈Nn

We let S(Rn ) = ∩s>0 WHs (Rn ). This is called the Schwartz space, mem-
bers of which are called Schwartz functions.

Theorem 2.2. The Schwartz space has the following properties. (i)
S(Rn ) is a dense subspace of L2 (Rn ); (ii) S(Rn ) is invariant under F
and (iii) F : S(Rn ) → S(Rn ) is one to one and onto.

The density follows from the fact that finite linear combinations of
Hermite functions form a subspace of S(Rn ) which is dense in L2 (Rn ).
The invariance follows from that of each of WHs (Rn ). As F (F ∗ f ) = f
the surjectivity is proved.
The Schwartz space can be made into a locally convex topological
vector space such that S(Rn ) is continuously embedded in WHs (Rn ) for
every s > 0. The dual of S(Rn ) denoted by S 0 (Rn ) is called the space
of tempered distributions. It can be shown that u ∈ S 0 (Rn ) if and only
 FOURIER TRANSFORM 13

if
X
(2|α| + n)−s |(f, Φα )|2 < ∞
α∈Nn
for some s > 0. The Fourier transform has a natural extension to
S 0 (Rn ) given by (F u, f ) = (u, F f ) where the brackets here stand for
the action of a tempered distribution on a Schwartz function.

2.3. Weighted Bergman spaces. From the definition it follows that

f ∈ S(Rn ) if and only if for every s > 0, there are constants Cs such
that
|(f, Φα )| ≤ Cs (2|α| + n)−s , α ∈ Nn .
It is natural to consider conditions of the form

|(f, Φα )| ≤ Ce−(2|α|+n)s , α ∈ Nn

for some s > 0. This will lead to another family of invariant subspaces
which can be identified with certain Hilbert spaces of entire functions.
For each t > 0 let us consider the weight function
n 2 −coth(2t)|y|2 )
Ut (x, y) = 2n (sinh(4t))− 2 e2π(tanh(2t)|x| .

We define Ht (Rn ) to be the subspace of L2 (Rn ) consisting all f which

extend to Cn as entire functions and satisfy
Z Z
2
kf kHt = |f (x + iy)|2 Ut (x, y)dxdy < ∞.
Rn Rn

These are examples of weighted Bergman spaces. We call them

Hermite-Bergman spaces for reasons which will become clear soon.

Theorem 2.3. For each t > 0 the space Ht (Rn ) is invariant under the
Fourier transform.
2
Since Φα (x) = Hα (x)e−π|x| where Hα is a polynomial we can extend
2
Φα to Cn as an entire function simply by setting Φα (z) = Hα (z)e−πz
where z 2 = nj=1 zj2 , z = (z1 , z2 , ..., zn ) ∈ Cn . The reader can verify by
P

direct calulation that Φα ∈ Ht (Rn ) for any t > 0. Moreover, it can be

shown that the functions Φ̃α (x) = e−(2|α|+n)t Φα (x) form an orthonormal
system in Ht (Rn ). More is true.
14 THANGAVELU

Theorem 2.4. The family {Φ̃α : α ∈ Nn } is an orthonormal basis for

Ht (Rn ).

We will not attempt a proof of this theorem but only indicate a major
step involved in the proof. Before that let us see how this theorem can
be used to prove the invariance of Ht (Rn ). Any f ∈ Ht (Rn ) has an
expansion
X
f= cα Φ̃α
α∈Nn
where the sequence cα is square summable. It is now obvious that
X
Ff = (−i)|α| cα Φ̃α
α∈Nn

also belongs to Ht (Rn ).

Let Bt (Rn ) be the Bergman space consisting of all f ∈ L2 (Rn ) which
extend to Cn as entire functions that are square integrable with respect
to the weight function
n π 2
pt/2 (y) = t− 2 e− t |y| .

That is, f ∈ Bt (Rn ) if and only if f (x + iy) is entire and

Z Z
2
kf kBt (Rn ) = |f (x + iy)|2 pt/2 (y)dxdy < ∞.
Rn Rn

We have the following characterisation of these Bergman spaces.

Theorem 2.5. A function f ∈ Bt (Rn ) if and only if f = g ∗ pt for

some g ∈ L2 (Rn ). Moreover,
Z Z Z
2
|f (x + iy)| pt/2 (y)dxdy = c |g(x)|2 dx.
Rn Rn Rn
2
Proof. When f = g ∗ pt so that F f (ξ) = F g(ξ)e−2πt|ξ| the inversion
formula shows that the prescription
Z
2
f (x + iy) = F g(ξ)e−2πt|ξ| e2πi(x+iy)·ξ dξ
Rn

gives an entire extension of f. Moreover, by Plancherel theorem

Z Z
2 2
|f (x + iy)| dx = |F g(ξ)|2e−4πt|ξ| e−4πy·ξ dξ.
Rn Rn
 FOURIER TRANSFORM 15

Integrating both sides against pt/2 (y) and simplifying we get

Z Z Z
2
|f (x + iy)| pt/2 (y)dxdy = c |g(x)|2 dx.
Rn Rn Rn

As all the steps are reversible we get the theorem. 

2.4. Spherical harmonics. In this subsection we look for some more

eigenfunctions of the Fourier transform which have some symmetry. As
in the one dimensional case we consider functions of the form f (x) =
2
p(x)e−π|x| . This will be an eigenfunction of F if and only p satisfies
Z
2
p(x − iy)e−π|x| dx = λp(y).
Rn

If this is true for all y ∈ Rn then we should also have

Z
2
p(x + y)e−π|x| dx = λp(iy).
Rn

Integrating in polar coordinates the integral on the left is

Z ∞ Z 
n−1 2
|S | p(y + rω)dσ(ω) e−πr r n−1 dr
0 S n−1

where dσ is the normalised surface measure on the unit sphere S n−1 . If

p is homogeneous of degree m then p(iy) = im p(y) and hence for such
polynomilas the equation
Z
2
p(x + y)e−π|x| dx = λim p(y)
Rn

will be satisfied for λ = (−i)m if p has the mean value property

Z
p(y + rω)dσ(ω) = p(y).
S n−1

Such functions are precisely the harmonic functions satisfying ∆u = 0.

Thus we have proved
2
Theorem 2.6. Let f (x) = p(x)e−π|x| where p is homogeneous of degree
m and harmonic. Then F f = (−i)m f.

Let Pm stand for the finite dimensional space of homogeneous har-

monic polynomilas of degree m. The above theorem says that the fi-
nite dimensional subspace of L2 (Rn ) consisting of functions of the form
16 THANGAVELU
2
p(x)e−π|x| , p ∈ Pm is invariant under the Fourier transform. We claim
that the following extension is true.

Theorem 2.7. Let f ∈ L2 (Rn ) be of the form f (x) = p(x)g(|x|) where

p ∈ Pm . Then F f (ξ) = p(ξ)G(|ξ|). Thus the subspace of functions
of the form f (x) = p(x)g(|x|), p ∈ Pm is invariant under the Fourier
transform.

Proof. When f (x) = p(x)g(|x|), p ∈ Pm is from L2 the function g

satisfies Z ∞
|g(r)|2r n+2m−1 dr < ∞.
0
We let Dn+2m to stand for the space of all such functions with the
obvious norm. We claim that the subspace W consisting of finite linear
2
combinations of e−πt|x| as t runs through positive reals is dense in
Dn+2m . To see this suppose g ∈ Dn+2m satisfies
Z ∞
2
e−πtr g(r)r n+2m−1 dr = 0
0

for all t > 0. Differentiating the integral k times at t = 1 we get

Z ∞
2
e−πr r 2k g(r)r n+2m−1 dr = 0
0
1 2
for all k ∈ N. Thus the function g(r)r n+2m−1 e− 2 πr is orthogonal to all
1 2
functions of the form P (r 2 )e− 2 πr where P runs through all polynomi-
las. As this is a dense class in L2 ((0, ∞), dr) we get g = 0.
In view of this density, it is enough to prove that W is invariant
under Fourier transform. But this is easy to check. For t > 0 we let
2 2
ft (x) = tn f (tx) so that F (ft )(ξ) = F f (t−1 ξ). If f (x) = p(x)e−πt |x|
2
take g(x) = p(x)e−π|x| and consider

F (f )(ξ) = t−n−m F (gt )(ξ) = t−n−m F (g)(t−1 ξ).

Since F g(ξ) = (−i)m g(ξ) we get

−2 |x|2
F (f )(ξ) = t−n−2m (−i)m p(x)e−πt .

This proves that W is invariant and hence the theorem follows. 

 FOURIER TRANSFORM 17

The above theorem gives rise to an operator Tmn on the space Dn+2m
defined as follows. If g ∈ Dn+2m then for p ∈ Pm the function
p(x)g(|x|) ∈ L2 (Rn ) whose Fourier transform is of the form p(x)G(|x|).
As F is unitary it follows that G ∈ Dn+2m . We define Tmn g = G. Note
that kTmn gk = kgk where k · k is the norm in Dn+2m . We can think of
g(|x|) as a radial function on Rn+2m whose n + 2m dimensional Fourier
transform will be a radial function, say G0 (|x|). We define another op-
erator T0n+2m on Dn+2m by letting T0n+2m g = G0 It is also clear that
kT0n+2m gk = kgk. If we denote the Fourier transform on Rn by Fn then
T0n+2m = Fn+2m . Calculations done in the proof of the above theorem
shows that Tmn g = (−i)m T0n+2m g whenever g ∈ W. The density of this
subspace gives
Theorem 2.8. Let f ∈ L2 (Rn ) be of the form f (x) = p(x)g(|x|), p ∈
Pm . Then Fn (f ) = (−i)m pFn+2m g.

The above result is known as the Hecke-Bochner formula for the

Fourier transform. We conclude our discussion on invariant subspaces
with the following result which shows that the Fourier transform of a
radial function reduces to an integral transform whose kernel is a Bessel
function. Let Jα stand for the Bessel function of type α > −1.
Theorem 2.9. If f (x) = g(|x|) is radial and integrable then
Z ∞
J n −1 (2πr|ξ|) n−1
Fn (f )(ξ) = cn g(r) 2 n r dr.
0 (2πr|ξ|) 2 −1
Proof. As f is radial
Z ∞ Z 
n−1 −2πirω·ξ
Fn (f )(ξ) = |S | g(r) e dσ(ω) r n−1 dr.
0 S n−1

The inner integral is clearly a radial function as the measure dσ is

rotation invariant. It can be shown that the inner integral is a constant
J n −1 (2πr|ξ|)
multiple of 2 n −1 . This completes the proof. 
(2πr|ξ|) 2

3. Ultravariant subspaces of L2 (Rn )

In the previous section we studied several subspaces of L2 (Rn ) that

are invariant under F . But not all subspaces are invariant. For example,
18 THANGAVELU

L1 ∩ L2 (Rn ) is not invariant under F . In this section we are interested

in subspaces which are extremely sensitive to the action of the Fourier
transform. We make this precise in the following definition.

Definition 3.1. We say that a subspace W of L2 (Rn ) is ultravariant

if the conditions f ∈ W, F f ∈ W imply f = 0.

A priori it is not clear if there is any ultravariant subspace of L2 (Rn )

but in this section we show that there are many such subspaces.

3.1. Paley-Wiener spaces. Our first example of an ultravariant sub-

space is the Paley-Wiener space defined as follows. For each a > 0 let
P Wa (Rn ) stand for the subspace of L2 (Rn ) consisting of functions hav-
ing entire extensions to Cn and satisfying
Z
|f (x + iy)|2 dx ≤ Ce4πa|y|
Rn

for all y ∈ Rn . We define P W (Rn ) = ∪a>0 P Wa (Rn ) and call it the

Paley-Wiener space. The space P W (Rn ) is not empty since any f ∈
L2 (Rn ) whose Fourier tranform is compactly supported belongs to the
Paley-Wiener space. To see this, suppose F f vanishes for |ξ| > a and
consider the inversion formula
Z
f (x) = F f (ξ)e2πix·ξ dξ.
|ξ|≤a

It is clear then that the prescription

Z
f (x + iy) = F f (ξ)e2πi(x+iy)·ξ dξ
|ξ|≤a

defines and entire function and by Plancherel we also have

Z
|f (x + iy)|2 dx ≤ Ce4πa|y| .
Rn

We show below that the converse is also true.

Theorem 3.2. An L2 function f belongs to P Wa (Rn ) if and only if

F f is supported in {ξ : |ξ| ≤ a}.
 FOURIER TRANSFORM 19

Proof. It is enough to show that F f is compactly supported in {ξ :

|ξ| ≤ a} whenever f ∈ P Wa (Rn ) since the converse has been already
proved. First we claim that P W (Rn ) ⊂ ∩t>0 Bt (Rn ). To see this let
f ∈ P Wa (Rn ) and consider
Z Z Z
2 −n π 2
|f (x + iy)| pt/2 (y)dxdy ≤ Ct 2 e4πa|y| e− t |y| dy.
Rn Rn Rn

A simple calculation shows that

Z Z
n 2
|f (x + iy)|2 pt/2 (y)dxdy ≤ C(a2 t) 2 e4πa t .
Rn Rn

This proves our claim. In view of Theorem 2.5 we get gt ∈ L2 (Rn ) such
that f = gt ∗ pt and
Z Z Z
2
|gt (x)| dx = C |f (x + iy)|2 pt/2 (y)dxdy.
Rn Rn Rn

For each δ > 0 consider

Z Z
|fˆ(ξ)| dξ ≤ e |fˆ(ξ)|2 e4πt|ξ| dξ.
2 −4πt(a+δ) 2 2

|ξ|>a+δ Rn

As f = gt ∗ pt the last integral is equal to

Z Z Z
2
|gt (x)| dx = C |f (x + iy)|2 pt/2 (y)dxdy.
Rn Rn Rn

This along with our earlier estimate gives

Z
n
|fˆ(ξ)|2 dξ ≤ C(a2 t) 2 e−4πt(a+δ) e4πta .
2 2

|ξ|>a+δ

Letting t go to infinity we conclude that fˆ vanishes for |ξ| > a + δ.

Since δ is arbitrary we see that fˆ is supported in |ξ| ≤ a. 

Theorem 3.3. P W (Rn) is ultravariant.

The thorem follows immediately from the Paley-Wiener theorem. If

f ∈ P W (Rn ) then fˆ is compactly supported and hence cannot have
an entire extension unless f = 0.
The Paley-Wiener space is strictly contained in ∩t>0 Bt (Rn ). It turns
out that the bigger space ∩t>0 Bt (Rn ) is also ultravariant. Even more
20 THANGAVELU

surprising is the following result. Note that the heat kernel ps ∈ Bt (Rn )
only for t < s since
2 n
F ps (x) = e−2πs|x| = (2s)− 2 p 1 (x).
4s

1
It is also clear that F ps ∈ Bt (Rn ) only for t < 4s . Therefore, if 0 <
t < 1/2 then for any s satisfying t < s < 1/4t the function ps and F ps
both belong to Bt (Rn ). This means that for such t the space Bt (Rn ) is
not ultravariant. But the behaviour is different for other values of t.

Theorem 3.4. The Bergman space Bt (Rn ) is ultravariant for all t ≥

1/2.

This is a very special case of a theorem of Cowling and Price which is

viewed as an uncertainty principle for the Fourier transform. There is
a more general theorem due to Bonami et al from which Cowling-Price
theorem can be deduced. We just state the result without proof.

Theorem 3.5. The only function f ∈ L2 (Rn ) for which

Z Z
|f (x)||F f (y)|e2π|x·y|dxdy < ∞
Rn Rn

is the trivial function f = 0.

In the one dimensional case this result is due to Beurling. Let us use
this to prove the ultravariance of Bt (Rn ) for t > 1/2. If both f and F f
are in Bt (Rn ) we have
Z
2
|g(x)|2 e4πt|x| dx < ∞
Rn

for g = f as well as for g = F f. It is then easy to check that the

hypothesis of Beurling’s theorem is satisfied.

3.2. Theta transform and Hardy’s theorem. In this section we

return to the Hilbert space L2 (M ) introduced in section 1.1. We in-
troduce and study a transform called the theta transform. As applica-
tions we show that the Hermite functions form an orthonormal basis
for L2 (R) and prove a theorem of Hardy.
 FOURIER TRANSFORM 21
2
Let ϕiτ (x) = eπiτ x which belongs to L2 (R) even for complex τ pro-
1 ∂
vided =(τ ) > 0. Let ψiτ (x) = 2πiτ ϕ (x) = xϕiτ (x). Recall that for
∂x iτ
2
f ∈ L (R) the Weil-Brezin transform is given by
∞
X
2πit
V f (x, y, t) = e f (x + n)e2πiny .
n=−∞

Definition 3.6. The theta transform is defined on L2 (M ) by

Z
Θ(F, τ ) = (V ϕiτ , F ) = V ϕiτ (g)F̄ (g)dg.
M

We also define Θ∗ (F, τ ) = (V ψiτ , F ).

Since V is a unitary operator we get the formulas

Z ∞
¯
Θ(V f , τ ) =
2
f (x)eπiτ x dx
−∞

and Z ∞
Θ (V f¯, τ ) =
∗ 2
xf (x)eπiτ x dx
−∞
∗
Note that Θ(F, τ ) and Θ (F, τ ) are functions defined on the upper
half-plane R2+ = {τ ∈ C : =(τ ) > 0}.

Theorem 3.7. For F ∈ L2 (M ) both Θ(F, τ ) and Θ∗ (F, τ ) are holo-

morphic in the upper half-plane.

Proof. It is clear that Θ(F, τ ) is holomorphic in the upper half-plane

when F ∈ C ∞ (M ). If Fn ∈ C ∞ (M ) converges to F ∈ L2 (M ) then
Θ(Fn , τ ) converges to Θ(F, τ ) uniformly over compact subsets of the
upper half-plane. This follows from the fact that V (ϕiτ ) is bounded
when τ is restricted to compact subsets. This shows that Θ(F, τ ) is
holomorphic. The proof for Θ∗ (F, τ ) is the same. 

We can decompose H1 as H1o ⊕ H1e where H1o (resp. H1e ) is the

image under V of all odd (even) functions. Note that V ϕiτ ∈ H1e and
V ψiτ ∈ H1e . Moreover, Θ(F, τ ) = 0 if F ∈ H1o and Θ∗ (F, τ ) = 0 if
F ∈ H1e . We can now prove the following uniqueness theorem for the
theta transform.
22 THANGAVELU

Theorem 3.8. For F ∈ H1 , F = 0 if and only if Θ(F, τ ) = Θ∗ (F, τ ) =

0 for all τ ∈ R2+ . Consequently, the set of all functions {V ϕiτ , V ψiτ , τ ∈
R2+ } is dense in H1 .
Proof. If F = G + H where G ∈ H1e and H ∈ H1o then Θ(F, τ ) =
Θ(G, τ ). Therefore, Θ(F, τ ) = 0 for all τ gives, by taking τ = t + i,
Z ∞
2 2
g(x)e−πx eπitx dx = 0
0

where G = V (ḡ). By making a change of variables we get

Z ∞
1 1
g(s 2 )s− 2 e−πs eist ds = 0.
0

As the integrand belongs to L1 (R) by the uniqueness theorem for

the Fourier transform we get g = 0. Similarly, the other condition
Θ∗ (F, τ ) = 0 gives h = 0. Hence the theorem. 
Corollary 3.9. The Hermite functions {ϕk : k ∈ N} form an or-
thonormal basis for L2 (R).
2
Proof. It is enought to show that the set of all functions {tn e−πt : n ∈
N} is dense in L2 (R). Suppose f is orthogonal to all these functions.
Let F = V (f¯) and consider θ(τ ) = Θ(F, τ ). Evaluating the derivatives
of θ at τ = i we get
Z ∞
(n) 2
θ (i) = f (t)t2n e−πt dt = 0.
−∞

As θ is holomorphic we get Θ(F, τ ) = 0. As before, if F = G + H, G ∈

H1e , H ∈ H10 we have Θ(G, τ ) = Θ(F, τ ) = 0 and Θ∗ (G, τ ) = 0. Hence
G = 0. This means that f is odd. Working with Θ∗ V (f¯) we can also
conclude that f is even. Hence f = 0 proving the corollary. 

We now use properties of the theta transform to prove a result on

Fourier tranform pairs due to Hardy. This result will be used to con-
struct some more examples of invariant and ultravariant subspaces.
Theorem 3.10. Suppose f ∈ L2 (R) satisfies the growth conditions
2 π 2
|f (x)| ≤ Ce−πtx , |F f (y)| ≤ Ce− t y
2
for some t > 0. Then f (x) = Ce−πtx .
 FOURIER TRANSFORM 23

Proof. By dilating by t we can assume that t = 1. Recalling the defini-

tions of ϕiτ and ψiτ we can easily calculate that
1 3
Θ(V ϕ−1 , τ ) = (1 − iτ )− 2 , Θ∗ (V ψ−1 , τ ) = (1 − iτ )− 2 .

Given f as in the theorem we can write it as f = g + h, where g(x) =

1
2
(f (x) + f (−x)) and h(x) = 21 (f (x) − f (−x)). Observe that both g
and h satisfy the same growth conditions as f. We therefore, prove the
theorem for even and odd functions separately.
If f is even, consider the decomposition f = g + h where g = 21 (f +
F f ) and h = 21 (f − F f ). Then F g = g, F h = −h and both satisfy the
conditions of the theorem. If f is odd the decompostion g = 21 (f +iF f )
and h = 21 (f − iF f ) gives f = g + h with F g = −ig and F h =
ih. This shows that we can assume without loss of generality f is an
eigenfunction of F . We start with the even case, F f = cf where c = 1
or −1.
We consider the function θ(τ ) = Θ(V (f¯), τ ) which is given by the
integral
Z ∞
2
α(τ ) = f (x)eπiτ x dx.
−∞

The growth condition on f shows that α(τ ) is holomorphic in =(τ ) >

−1. Since
Θ(V (f¯), τ ) = (V ϕiτ , V f¯) = (ϕiτ , f¯)

using the result (f, g) = (F f, F g) we get

1 1 −1
Θ(V (f¯), τ ) = (−iτ )− 2 (ϕ −i , f¯) = c(−iτ )− 2 α( ).
τ τ
In the above calculation we have used the facts that F ϕiτ =
(−iτ )− 2 ϕ −i and F f¯ = cf¯. Therefore, α satisfies α(τ ) = c(−iτ )− 2 α( −1
1 1
τ
)
τ
−1
provided both =(τ ) > −1 and =( τ ) > −1.
1
Define a new function β(τ ) = (1 − iτ ) 2 α(τ ). If we can show that
β(τ ) is a constant which means that α(τ ) = CΘ(V ϕ−1 , τ ) then by the
uniqueness theorem for the theta transform we get f = ϕ−1 . This will
take care of the even case.
24 THANGAVELU

An easy calculation shows that the function β satisfies β(τ ) = cβ( −1

τ
)
−1
whenever both =(τ ) > −1 and =( τ ) > −1. If τ = a + ib, b < −1, then
−b = |b| < b2 + a2 so that =( −1 τ
b
) = (a2 +b 2 ) > −1. Therefore, β can

be extended to the entire complex plane except possibly τ = −i. With

1
τ = a + ib we have the estimate |α(τ )| ≤ C(1 + b)− 2 , b > −1 which
follows from the integral defining α and the hypothesis on f. When
a2 + b2 ≤ 1 writing b = −1 + δ, δ > 0 we have a2 ≤ 1 − b2 = δ(2 − δ).
This gives the estimate
|a| 3
|(1 − iτ )β(τ )| ≤ C(1 + b)(1 + )2
1+b
1 3 1 1 3
≤ Cδ(1 + 2δ − 2 ) 2 ≤ Cδ 4 (2 + δ 2 ) 2 .
This together with the property β(τ ) = cβ( −1τ
) shows that (1 − iτ )β(τ )
tends to zero as τ goes to −i. Hence β is entire. It is also bounded
(since β(τ ) = cβ( −1
τ
) ) and hence reduces to a constant.
When f is an odd eigenfunction of the Fourier transform we work
with (1 − iτ ) 2 Θ∗ (V f¯, τ ) and show that f (x) = Cxe−πx . But now the
3 2

constant C has to be zero if the growth condition on f is satisfied.

Thus the odd componet of f is zero proving the theorem.


The above theorem can be generalised and extended to functions on

n
R . The following result is known as the Hardy’s uncertainty principle
for the Fourier transform.

Theorem 3.11. Let f ∈ L2 (Rn ) satisfy the estimates

2 2
|f (x)| ≤ Cp(x)e−πt|x| , |F f (y)| ≤ Cq(y)e−πs|x|

for some s, t > 0 and p, q polynomials. Then f = 0 whenever st > 1

2
and f (x) = Cp(x)e−πt|x| when st = 1.

Proof. First we consider the case st > 1 in one dimension. We can

choose , δ > 0 such that (s − δ)(t − ) = 1. Then we have the estimates
2 2
|f (x)| ≤ C e−π(t−)x , |F f (y)| ≤ Cδ e−π(s−δ)y .
 FOURIER TRANSFORM 25
2
By the previous theorem we conclude that f (x) = Ce−π(t−)x . But this
cannot satisfy the hypothesis unless C = 0 proving that f = 0.
To prove the theorem in n dimensions, we fix a vector ω ∈ S n−1 and
consider the function defined on R by
Z
fω (v) = f (u ⊕ vω)du.
Rn−1
Then it is easy to see that
Z
fω (v)e−2πivr dv = fˆ(rω).
R
Thus the function fω and its Fourier transform satisfy the hypothesis
with st > 1 and hence fˆ(rω) = 0. As this is true for all ω we conclude
that f = 0.
To prove the equality case st = 1 with polynomial factors, again
we need to consider the one dimensional case. The proof of Theorem
can be modified to take care of this case. The details are left to the
reader. 

Hardy’s theorem gives us one more example of ultravariant space.

For each t > 0 consider the Hardy class Ht (Rn ) consisting of functions
2
satifying |f (x)| ≤ Cp(x)e−πt|x| for some polynomial. Then we have

Theorem 3.12. The Hardy class Ht (Rn ) is ultravariant for all t > 1.

Since the Hermite functions Φα ∈ Ht (Rn ) for all t ≤ 1 it follows that

Ht (Rn ) is not ultravariant for t ≤ 1. We do not know if it is invariant
or not.
26 THANGAVELU

Notes and references

There are several excellent books on Fourier transform and related
topics. We recommend Dym and McKean [4], Stein and Weiss [5] and
Sugiura [6]. We have followed Auslander and Meyer [1] and Howe [2]
in introducing the Fourier transform via Heisenberg group. For more
about this group we refer to Folland [3] and Thangavelu [7]. For the
weighted Bergman spaces the reader may consult the CIMPA lecture
notes [9]. The proof of Hardy’s theorem presented here is due to Tolim-
ieri [10]. For other proofs and extensions of Hardy’s theorem we refer
to Thangavelu [8].

References
[1] L. Auslander and Y. Meyer, A generalised Poisson summation formula, Appl.
and Comp. Harmonic Analysis 3, 372-376 (1996).
[2] R. Howe, On the role of the Heisenberg group in Harmonic Analysis, Bull.
Amer. Math. Soc., 3, 821-843 (1980).
[3] G. B. Folland, Harmonic Analysis in Phase Space, Ann. Math. Stud. 122,
Princeton Univ. press, Princeton, N.J. (1989).
[4] H. Dym and H. P. McKean, Fourier series and integrals, Academic press, New
York (1972).
[5] E. M. Stein and G. Weiss, Introduction to Fourier Analysis on Euclidean
Spaces, Princeton Univ. press, Princeton, N.J. (1971).
[6] M. Sugiura, Unitary representations and Harmonic Analysis, Wiley, New York
(1975).
[7] S. Thangavelu, Harmonic Analysis on the Heisenberg group, Prog. in Math.
Vol. 159, Birkhäuser, Boston (1998).
[8] S. Thangavelu, An Introduction to the Uncertainty Principle, Prog. in Math.
Vol. 217, Birkhäuser, Boston (2004).
[9] S. Thangavelu, Hermite and Laguerre semigroups: some recent developments,
CIMPA lecture notes (2006).
[10] R. Tolimieri, The theta transform and the Heisenberg group, J. Funct. Anal.
24, 353-363 (1977).

Department of Mathematics,, Indian Institute of Science,, Banga-

lore 560 012, India E-mail : [email protected]