Mathématiques pour les glomorphes à rayures

Fourier Transform

The Fourier transform is a fundamental tool, in particular for the study of partial
derivative equations, because it “diagonalizes” (in a sense to be made precise) the dif-
ferential operators.

Let us begin by reminding the motivation for diagonalization in finite dimension.

Let E be a C-vector space of finite dimension n P N. Let A be an endomorphism of E
and u0 P E. We consider the following problem on E
#
u1 ptq “ Auptq, @t P R,
(1)
up0q “ u0 ,

where the unknown u is a function (say of class C 1 ) from R to E. The favorable case
is the case where u0 is an eigenvector of A, associated to an eigenvalue λ P C. In that
case, It is easy to see that we get a solution by setting, for any t P R,

uptq “ etλ u0 .

Indeed, the function u thus defined is of class C 8 from R to E, we have up0q “ u0 and,
for any t P R,
u1 ptq “ etλ λu0 “ etλ Au0 “ Auptq.
Since the equation u1 “ Au is linear with respect to u, we get that if u0 is a linear
combination of eigenvectors, that is if
k
ÿ
u0 “ αj e j ,
j“1

where k ď n, α1 , . . . , αk P R and for any j P J1, kK the vector ej is an eigenvector A

associated to an eigenvalue λj P C, then a solution of (1) is given by
k
ÿ
uptq “ etλj αj ej .
j“1

The interest of results about diagonalization is to give criteria ensuring that any vector
u0 of E can be written as a linear combination of eigenvectors of A. Thus, if A is diag-
onalizable and if we can determine its eigenvalues and eigenvectors, we can easily solve
Julien Royer
Université Toulouse III - Paul Sabatier
February 11, 2025

1
(1) for any initial data u0 .

It turns out that many models from concrete problems can be written in the form
(1), but in a functionnal space E of infinite dmension (typically L2 pRd q), and a linear
map A that is a differential operator. An important example (for example for the heat
equation, but also for Schrödinger equation or the wave equation) is the Laplace operator
d
ÿ B2
A“∆“ .
k“1
Bx2k

For example, the heat equation on Rd can be written as

Bt upt, xq “ ∆upt, xq, @t ě 0, @x P Rd .

The analysis is obviously more complicated in infinite dimension, so it is even more

important to use an appropriate setting.

To simplify the rest of the discussion, we assume that d “ 1. The simplest differential
operator on R is
Avpxq “ v 1 pxq.
We look for eigenvectors of A, that is functions v such that v 1 “ λv for some λ P C. The
candidates are obviously the exponential functions x ÞÑ eλx . These functions appears in
the Fourier series theory, that have precisely been introduced to solve the heat equation,
but on a bounded interval. According to the Fourier series theory, if we set
2iπnx
en : x ÞÑ e T , for n P Z,

then the family pen qnPZ is a Hilbert basis of the Hilbert space L2T pRq of locally integrable
and T -periodic functions from R to C, endowed with the norm defined by
żT
2 1 2
}f }L2 pRq “ |f pxq|2 dx.
T T ´ T
2

Thus for any f P L2T pRq we have

ż T
ÿ 1 2 2iπny
f“ cn pf qen where cn pf q “ ⟨f, en ⟩ “ e´ T f pyq dy.
nPZ
T ´ T2

In addition, each en is an eigenvector for the derivative operator. If f is regular we have

ÿ ÿ 2iπn
f1 “ cn pf qe1n “ cn pf qen .
nPZ nPZ
T

This holds for T -periodic functions, but we can do the same for the functions defined
on an interval of length T .
Here, we are interested in functions that are not periodic and are defined on all R.
Formally, we want to see what happens if we let T go to `8 in the previous expressions.
Observing that for ϕ P C08 pRq we have
ˆ ˙ ż
2π ÿ 2πn
ϕ ÝÝÝÝÝÑ ϕpξq dξ (2)
T nPZ T T Ñ`8 R

2
(this is essentially a Riemann sum), we want to write f pxq as
ż ˆż ˙
1
e ´iyξ
f pyq dy eixξ dξ. (3)
2π R R
If this has a meaning, and if it really gives an expression for f pxq, then we have
written f as an integral (that can be seen as the limit of a series, see (2), or as a “con-
tinuous sum”) of functions of the form cξ pf qeixξ , that are “eigenvectors” (it would be
necessary to specify the space in which one works) associated to the derivative operator.
In addition the corresponding coefficient cξ pf q can be written as a scalar product in
L2 pRq of f with the conjugate function x ÞÑ eixξ “ e´ixξ . Except that the function
x ÞÑ eixξ is not in L2 pRq. . .

Thus, the definition of the Fourier transform on R (or Rd for any d P N˚ ) is motivated
and guided by good properties of the Fourier series on a bounded interval, but to get
similar results in this setting, a specific analysis is necessary. This is the purpose of this
chapter.

1 The Schwartz space

In this chapter we will rely on Schwartz functions.
Definition 1.1. We denote by SpRd q the Schwartz space of functions ϕ on Rd which
are smooth and such that for any α, β P Nd there exists Cα,β ą 0 for which
ˇ ˇ
d ˇ α β
@x P R , ˇx B ϕpxqˇ ď Cα,β .
ˇ

We say that every derivative of ϕ is a rapidly decreasing function.

Example 1.2. • We have C08 pRd q Ă SpRd q.
2
• The map x ÞÑ e´|x| is in SpRd q but not in C08 pRd q.
Definition 1.3. We say that f : Rd Ñ C is a slowly increasing function if there exist
N P N and C ą 0 such that
@x P Rd , |f pxq| ď Cp1 ` |x|qN .
Proposition 1.4. SpRd q is a vector space, stable under the usual product, under convo-
lution, under derivation and multiplication by a function of class C 8 whose derivatives
are slowly increasing.
Proof. Let us prove that for f, g P SpRd q then pf ˚ gq P SpRd q. Let k P N and x P Rd .
Let y P Rd . If |x| ě 2 |y| we have
|x|
|x ´ y| ě |x| ´ |y| ě ,
2
so
ż
|x|k |pf ˚ gqpxq| ď |x|k |f px ´ yq| |gpyq| dy
yPRd
ż ż
k
ď k
2 |x ´ y| |f px ´ yq| |gpyq| dy ` |f px ´ yq| 2k |y|k |gpyq| dy.
|y|ď |x|
2
|y|ě |x|
2

3
This proves that f and g are rapidly decreasing functions, then pf ˚ gq is also a rapidly
decreasing function.
By derivation under the integral sign, we see that for f, g P SpRd q the convolution
pf ˚ gq is of class C 8 on Rd and for β P Nd we have B β pf ˚ gq “ pB β f q ˚ g. The previous
result applied to B β f and g ensures that the derivatives of pf ˚ gq are rapidly decreasing,
and so pf ˚ gq P SpRd q.

Since we will also work in different Lebesgue spaces, we give some useful links between
these different spaces.
Proposition 1.5. (i) For any p P r1, `8s we have SpRd q Ă Lp pRd q.
(ii) For p P r1, `8r, SpRd q is dense in Lp pRd q. More precisely, for p, q P r1, `8r and
f P Lp pRd q X Lq pRd q there exists a sequence pfn qnPN of Schwartz functions that
converges to f in both Lp pRd q and Lq pRd q.
The first point is easy, for the second one we use the fact that these properties of
density are valid for C08 pRd q, which is included in SpRd q.

We finish with the property of integration by parts on Rd . The fact that Schwartz
functions ar rapidly decreasing at infinity ensures that there is no difficulty with the
boundary terms to show the following result.
Proposition 1.6. Let u P C 1 pRd q be a slowly increasing function whose first order
partial derivatives are also slowly increasing. Let v P SpRd q. Then for j P J1, dK the
functions pBj uqv and upBj vq are integrable on Rd and we have
ż ż
pBj uqv dx “ ´ upBj vq dx.
Rd Rd

2 Fourier transform in L1
We begin by defining what will play the role of “Fourier coefficient” for a non periodic
function. It is in L1 pRd q that the following definition has a natural meaning.
Definition 2.1. Let f P L1 pRd q. For ξ P Rd we set
ż
fˆpξq “ e´ix¨ξ f pxq dx. (4)
Rd

This defines a function fˆ on Rd which is called the Fourier transform of f . It is also

denoted by Ff .
This definition is licit since the function x ÞÑ e´ix¨ξ f pxq is integrable for any ξ P Rd .
Moreover, the definition of fˆpξq does not depend of the choice of the representative of
f (in other words, the definition does not change if we replace f by a function equal to
f almost everywhere). In the sequel we will no longer discuss the distinction between
functions of L1 pRd q or equivalence classes modulo almost everywhere equality in L1 pRd q.
Notice that even if f is only defined for an equivalence class of function, fˆpξq is well
defined for each fixed. Notice also that fˆp0q is just the integral of f :
ż
ˆ
f p0q “ f pxq dx.
Rd

4
 1
Example 2.2. Let a ą 0. We consider on R the function f : x ÞÑ 2a 1r´a,as . Then
f P L1 pRq, fˆp0q “ 1 and for ξ P R˚ we have
1 a ´ixξ 1 e´iaξ ´ eiaξ
ż
ˆ sinpaξq
f pξq “ e dx “ “ .
2a ´a 2a ´iξ aξ
Example 2.3. Let a ą 0. For x P R we set f pxq “ e´a|x| . Then f P L1 pRq and for ξ P R
we have (exercise)
2a
fˆpξq “ .
a2 ` ξ 2
Remark 2.4. Different conventions co-exist for the definition of fˆ. We can also set
ż
ˆ 1
f pξq “ d e´ix¨ξ f pxq dx.
p2πq 2 Rd

The advantage of this definition is that the expression of the Fourier transform and its
inverse transform (see Proposition 3.1) are symetrical, and there is no 2π factor in the
d
Parseval identity (see Proposition 5.1). However a p2πq 2 factor appears for example for
the convolution (see Proposition 2.7). Another convention is to set
ż
ˆ
f pξq “ e´2iπx¨ξ f pxq dx.
Rd
None of these choices is better than the others, if we remove the 2π factor somewhere it
will appear somewhere else in the theory. Thus, when we use the Fourier transform, it
is better to recall which definition is used and to be consistent afterwards (even if the
p2πqd factor is in general harmless for the studied properties).
We now give a number of basic properties for the Fourier transform.
Proposition 2.5. (i) For f P L1 pRd q, fˆ is bounded and }fˆ}8 ď }f }1 .

(ii) The map f ÞÑ fˆ is linear from L1 pRd q to L8 pRd q.

(iii) For f P L1 pRd q, fˆ is continuous on Rd and tends to 0 when |ξ| tends to `8.
Proof. The first two properties come from the triangular inequality and linearity of the
integral. The continuity of fˆ is given by continuity under the integral sign.
Let us prove that fˆ tends to 0 when |ξ| tends to `8. We first assume that f is in
SpRd q. Let j P J1, dK. Let ξ “ pξ1 , . . . , ξn q P Rd such that ξj ‰ 0. With an integration
by part we get
ż ż
ˆ 1 ´ix¨ξ 1 1 z
f pξq “ ´Bxj e f pxq dx “ e´ix¨ξ Bxj f pxq dx “ Bx f pξq.
iξj Rd iξj Rd iξj j

Since Bxj f P L1 pRd q we have Bz 8 d ˆ

xj f P L pR q, and so f pξq tends to 0 when |ξj | tends to
`8. This applies to every j, so we get the result for f P SpRd q.
Now let us consider the general case. Let f P L1 pRd q. Let ε ą 0. Since SpRd q is
dense in L1 pRd q, there exists g P SpRd q such that }f ´ g}L1 ď 2ε . Then there exists
R ą 0 such that for |ξ| ě R we have |ĝpξq| ď 2ε . For |ξ| ě R we have
ε
|fˆpξq| ď |fˆpξq ´ ĝpξq| ` |ĝpξq| ď }f ´ g}L1 ` ď ε.
2
This proves that fˆpξq tends to 0 when |ξ| tends to `8.

5
 The proof of the following properties are left as exercises for the reader:

Proposition 2.6. Let f P L1 pRd q.

(i) If f is an even (respectively odd) function, then fˆ is even (respectively odd). More
generally, if we note P the operator that associates Pf : x ÞÑ f p´xq to f , then for
f P L1 pRd q we have
PFf “ FPf. (5)

(ii) If f is a real valued function fˆp´ξq “ fˆpξq for any ξ P Rd .

(iii) Let x0 P Rd et g : x ÞÑ f px ´ x0 q. Then ĝpξq “ fˆpξqe´ix0 ¨ξ for any ξ P Rd .

Let ξ0 P Rd and g : x ÞÑ eix¨ξ0 f pxq. Then ĝpξq “ fˆpξ ´ ξ0 q for any ξ P Rd .

(iv) Let α ą 0 and g : x ÞÑ f pαxq. Then ĝpξq “ α1d fˆ αξ for any ξ P Rd .

` ˘

A remarkable property (which will play an important role in applications) is that the
Fourier transform of a convolution product is the usual product of Fourier transforms.

Proposition 2.7. Let f, g P L1 pRd q. Then we have

˚ g “ fˆĝ.
fz

Proof. Let ξ P Rd . The map px, yq ÞÑ e´ix¨ξ f px ´ yqgpyq is measurable. By the Fubini-
Tonelli theorem and the properties of the convolution product in L1 pRd q we have
ż ˇ ˇ ż ˆż ˙
ˇ ´ix¨ξ
ˇe f px ´ yqgpyqˇ dx dy ď |f px ´ yq| |gpyq| dy dx ď }f }1 }g}1 ă `8.
ˇ
Rd ˆRd Rd Rd

The, by the Fubini-Lebesgue theorem,

ż
f ˚ gpξq “
z e´ix¨ξ pf ˚ gqpξq dx
Rd
ż ˆż ˙
´ix¨ξ
“ e f px ´ yqgpyq dy dx
Rd Rd
ż ˆż ˙
´iy¨ξ ´ipx´yq¨ξ
“ e gpyq e f px ´ yq dx dy
Rd Rd
“ fˆpξqĝpξq.

3 Inversion formula
The property which makes the Fourier transform useful in applications is that we can
recover a function if we know its Fourier transform.

Proposition 3.1. Let f P L1 pRd q such that fˆ P L1 pRd q. Then for almost every x P Rd
we have ż
1
f pxq “ eix¨ξ fˆpξq dξ.
p2πqd Rd

6
 We observe that the formula giving the inverse Fourier transform is almost the
same as the one giving the Fourier transform itself (only the sign in the exponential
changes, and with our choice for the definition of F there is an extra factor p2πq´d ).
For g P L1 pRd q and x P Rd we set
ż
1 1
Fgpxq “ d
eix¨ξ gpξq dξ “ pFgqp´xq.
p2πq Rd p2πqd
Using the notation P introduced in proposition 2.6 this can be rewritten as
1 1
F“ PF “ FP.
p2πqd p2πqd
Thus we can write
1 1
f “ FFf “ d
PF Ff “ FF Pf “ FFf. (6)
p2πq p2πqd

Proof. ‚ For ε ą 0 and x P Rd we set

ż ż
1 ix¨ξ ˆ 1
F pxq “ e f pξq dξ and Fε pxq “ eix¨ξ e´εp|ξ1 |`¨¨¨`|ξd |q fˆpξq dξ.
p2πqd Rd p2πqd Rd

By continuity under the integral sign, these two functions are continuous on Rd . More-
over, by the dominated convergence theorem we have, for all x P Rd ,

Fε pxq ÝÝÝÑ F pxq.

εÑ0

‚ Let ε ą 0 and x P Rd . The function

py, ξq ÞÑ eipx´yq¨ξ e´εp|ξ1 |`¨¨¨`|ξd |q f pyq

is integrable on Rd ˆ Rd , so by the Fubini-Lebesgue Theorem we have

ż ˆż ˙
1 ix¨ξ ´εp|ξ1 |`¨¨¨`|ξd |q ´iy¨ξ
Fε pxq “ e e e f pyq dy dξ
p2πqd Rd Rd
˜ ¸
ż d ż
1 ź
“ eipxj ´yj qξj e´ε|ξj | dξj f pyq dy.
p2πqd Rd j“1 R

Let j P J1, dK. The computation of Example 2.3 gives

ż
2ε
eipxj ´yj qξj e´ε|ξj | dξj “ 2 .
R ε ` pxj ´ yj q2

For s “ ps1 , . . . , sj q P Rd we set

d
ź 1 1 ´s¯
χpsq “ and χε psq “ d χ .
j“1
πp1 ` s2j q ε ε

Then
p2εqd f pyq
ż
1
Fε pxq “ dy “ pχε ˚ f qpxq.
p2πqd
śd 2
Rd j“1 pε ` pxj ´ yj q2 q

7
Since pχε qεPs0,1s defines an approximation of unity, we have Fε P L1 pRd q and

}Fε ´ f }L1 ÝÝÝÑ 0.

εÑ0

In particular, there exists a sequence pεn qnPN that tends to 0 and such that Fεn tends
to f almost everywhere. This proves that F “ f almost everywhere.

Thus we have proved that under the hypotheses of the proposition 3.1, the expres-
sion obtained formally in (3) holds for almost any x P Rd .

We observe that Proposition 3.1 gives injectivity for the Fourier transform in L1 pRd q.

Corollary 3.2. If f P L1 pRd q is such that fˆ “ 0 almost everywhere, then f “ 0.

Proof. If fˆ “ 0 then in particular fˆ P L1 pRd q. So we can apply Proposition 3.1, which

shows that f pxq “ 0 for almost every x P Rd .

4 Derivation and multiplication by a polynomial - Fourier

transform in S
We now discuss the property for which the Fourier transform has been introduced,
namely the good behavior with respect to differential operators. We work with functions
x ÞÑ eix¨ξ since their derivatives are given by multication by a scalar. More precisely, if
ż
1
f pxq “ eix¨ξ fˆpξq dξ,
p2πqd Rd
we expect ż
1
Bxj f pxq “ iξj eix¨ξ fˆpξq dξ.
p2πqd Rd
The computations of the proof of proposition 2.5 gives the following property.

Proposition 4.1. Let f P SpRd q. Let j P J1, dK. Then for ξ P R we have

ˆ
xj f pξq “ iξj f pξq
Bz

In particular, the map ξ ÞÑ ξj fˆpξq is in L8 pRd q.

This property can in fact be extended to any function f P L1 pRd q X C 1 pRd q such
that the partial derivative Bxj f is also in L1 pRd q.

Since the inverse Fourier transform has an expression analogous to the Fourier trans-
form itself, it is not surprising to find that we actually have an analogous property in
the other direction. That is, the Fourier transform changes the multiplication by the
variable xj into a derivative with respect to ξj :

Proposition 4.2. Let f P SpRd q. Let j P J1, dK. Then fˆ is derivable with respect to ξj
and for any ξ P Rd we have
x
y ˆ
j f pξq “ iBξj f pξq.

8
Proof. For px, ξq P Rd ˆ Rd we set φpx, ξq “ e´ix¨ξ f pxq. Then φ is differentiable with
respect to ξj and for any px, ξq P pRd q2 we have
ˇ ˇ ˇ
ˇ Bφ ˇ ˇ ˇ
ˇ px, ξq ˇ “ ˇ´ixj e´ixξ f pxqˇˇ “ |xj f pxq| .
ˇ Bξj ˇ

Since x ÞÑ |xj f pxq| is integrable on Rd , we get by differentiation under the integral sign
that fˆ is differentiable with respect to ξj in Rd and, for ξ P Rd ,
ż
ˆ
iBξj f pξq “ e´ix¨ξ xj f pxq dx “ x
y j f pξq.
Rd

Notice that this proof holds for any function f P L1 pRd q such that x ÞÑ xj f pxq is
integrable.

The Schwartz space is the space of functions for which one can derive and multiply by
the variable as many times as one wants. Moreover, these two operations play symmetric
roles with respect to the Fourier transform. The two previous propositions suggest that
it is a space in which the Fourier transform is particularly convenient.

Proposition 4.3. The Fourier transform F defines a bijection from SpRd q to itself,
and its inverse is the restiction of F to SpRd q. Moreover, for f P SpRd q, α, β P Nd and
ξ P Rd we have
ξ α Bξβ Ff pξq “ p´iq|α|`|β| FpBxα xβ f qpξq. (7)

Proof. Let f P SpRd q and α, β P Nd . Applying |β| times Proposition 4.2, we get that
the partial derivative Bξβ fˆ exists and

β f pξq “ i|β| B β fˆpξq.

xy ξ

Now applying Proposition 4.1 to the successive derivatives of xβ f we get

α β |α|`|β| α β ˆ
x x f pξq “ i
B{ ξ Bξ f pξq.

In particular, the function ξ ÞÑ ξ α B β fˆpξq is bounded. This ensures that fˆ P SpRd q and
proves (7). The restriction FS of F to SpRd q is then injective by Corollary 3.2. Finally,
1
according to (6) we have f “ p2πq 2d FF Pf , so FS is surjective. Thus FS is a bijection

in SpRd q, and its inverse function is the restriction of F to SpRd q.

Example 4.4. Let f P SpRd q. Then the Fourier transform of ´∆f is ξ ÞÑ |ξ|2 fˆpξq.
Example 4.5. We want to compute on R the Fourier transform of the Gaussian f : x ÞÑ
x2
e´ 2 . Notice that f P SpRq. In particular fˆ P SpRq. Since f 1 pxq “ ´xf pxq for any
x P R, we have that for ξ P R

fˆ1 pξq “ ´ixf

x pξq “ ifp1 pξq “ ´ξ fˆpξq.

This proves that

ξ 2
fˆpξq “ e´ 2 fˆp0q.

9
But we know that ż
x2 ?
fˆp0q “ e´ 2 dx “ 2π,
R
So for any ξ P R we have
? ξ2
fˆpξq “ 2πe´ 2 .
x2
More generally, for σ ą 0, if we consider fσ : x ÞÑ e´ 2σ2 then for ξ P R we have
? 2 2
´ σ 2ξ
σ pξq “ σ 2πe
fx .

5 Fourier transform in L2
So far we have discussed the properties of the Fourier transform in L1 pRd q (because it
is the space in which the definition naturally makes sense) or of its restriction to SpRd q
(where it is the most comfortable). But our favorite space in the analysis of partial
differential equations will be L2 pRd q, because it is a Hilbert space. The problem is that
a function in L2 pRd q is not necessarily integrable, and the Fourier transform as defined
in (4) does not even make sense in this case.

We start by going a little further with the analysis of the Fourier transform for
Schwartz functions. Since SpRd q is included in L2 pRd q, it can be endowed with the
corresponding norm. The Fourier transform of SpRd q is then a quasi-isomery for this
norm:

Proposition 5.1. Let f, g P SpRd q. Then we have

ż ż
fˆpξqĝpξq dξ “ p2πqd f pxqgpxq dx.
Rd Rd

In particular,
d
}fˆ}L2 “ p2πq 2 }f }L2 .

Proof. The map px, ξq ÞÑ e´ix¨ξ f pxqĝpξq is integrable on Rd ˆ Rd , so by the Fubini-

Lebesgue Theorem we have
ż ż ˆż ˙
ˆ
f pξqĝpξq dξ “ e ´ix¨ξ
f pxq dx ĝpξq dξ
Rd Rd Rd
ż ˆż ˙
´ix¨ξ
“ f pxq e ĝpξq dξ dx
Rd Rd
ż ˆż ˙
“ f pxq ix¨ξ
e ĝpξq dξ dx
Rd Rd
ż
“ p2πqd f pxqgpxq dx.
Rd

The second propery follows by taking g “ f .

Considering g “ ϕ̂ we can re-write the previous equality as follows.

10
Corollary 5.2. For f, ϕ P SpRd q we have
ż ż
fˆϕ dy “ f ϕ̂ dy.
Rd Rd

Now we use the fact that SpRd q is dense in L2 pRd q to extend the Fourier transform
on SpRd q into an ismoetry of L2 pRd q.

Theorem 5.3. There exists a unique isomorphism F of L2 pRd q which coincides with
the Fourier transform on SpRd q and such that
d
@f P L2 pRd q, }Ff }L2 pRd q “ p2πq 2 }f }L2 pRd q . (8)

Proof. ‚ Assume that two isomorphisms of L2 pRd q satisfy the conditions of the theo-
rem. They are in particular continuous applications on L2 pRd q which coincide on the
dense subset SpRd q. Then they coincide everywhere, which gives uniqueness.
‚ Let f P L2 pRd q. There exists a sequence pfn qnPN of Schwartz functions that converge
to f in L2 pRd q. By Proposition 5.1, the sequence pfx 2 d
n qnPN is a Cauchy sequence in L pR q.
2 d 2 d
Since L pR q is complete, fx n has a limit in L pR q, which we denote by Ff . We can
check that this definition of Ff does not depend on the choice of the sequence pfn qnPN
and that for f P SpRd q the limit obtained this way coincides with the Fourier transform
of f . By linearity of the Fourier transform on SpRd q and by Proposition 5.1, the map
F is linear and satisfies (8). By continuity, the equality (6) still holds in L2 pRd q, which
proves that F is surjective.

Remark 5.4. For f, g P L2 pRd q we have

ż ż
Ff Fg dξ “ p2πqd f g dx.
Rd Rd

and ż ż
pFf qg dy “ f pFgq dy.
Rd Rd

Proposition 5.5. The isomorphism of Theorem 5.3 coincides with the Fourier trans-
form already defined on L1 pRd q X L2 pRd q.

Proof. Let f P L1 pRd q X L2 pRd q. For this proof, we denote by fˆ the Fourier transform
of f in the sense (4), and by Ff the function of L2 pRd q given by Theorem 5.3. Let
pfn qnPN be a sequence of Schwartz functions which converges to f both in L1 pRd q and
L2 pRd q. With the same convention as for f , we have fˆn “ Ff for any n P N. Since F
is continuous on L2 pRd q we have

}Ffn ´ Ff }2 ÝÝÝÝÑ 0.
nÑ`8

In particular, possibly after extracting a subsequence, we can assume that Ffn tends to
Ff for almost every x P Rd . On the other hand we have

}fˆn ´ fˆ}8 ď }f ´ fn }1 ÝÝÝÝÑ 0.

nÑ`8

This implies that Ff “ fˆ almost everywhere.

11
 Recall that for a non integrable function, we cannot use the expression of Ff given
by (4). We at least have the following property, where f˜n is a kind of Fourier transform
for locally integrable functions.
Proposition 5.6. Let f P L2 pRd q. For n P N and ξ P Rd we set
ż
f˜n pξq “ e´ix¨ξ f pxq dx,
Bpnq

where Bpnq denotes the ball of radius n centered at 0. Then we have

›f˜n ´ Ff › ÝÝÝÝÑ 0.
› ›
2 nÑ`8

Proof. By the dominated convergence theorem we have

› ›
›1Bpnq f ´ f › ÝÝÝÝÑ 0.
2 nÑ`8

Since 1Bpnq f P L1 X L2 for any n P N, we get using proposition 5.5 that

}f˜n ´ Ff }2 “ ›Fp1Bpnq f q ´ Ff ›2 ÝÝÝÝÑ 0.

› ›
nÑ`8

6 Example of application
For f P SpRd q, we consider the equation

´∆u ` u “ f, (9)

of unknown u P SpRd q. Taking the Fourier transform, we see that u is a solution if and
only if for any ξ P Rd we have

p|ξ|2 ` 1qûpξq “ fˆpξq,

or equivalently
fˆpξq
ûpξq “ .
1 ` |ξ|2
fˆpξq
We consider the function f˜ which maps ξ P Rd to 1`|ξ|2
. This defines a function
f˜ P SpRd q and we have
d
}f˜}L2 ď }fˆ}L2 “ p2πq 2 }f }L2 .
We then set Rf “ F ´1 f˜. This also defines a function of SpRd q and we have
d
}Rf }L2 “ p2πq´ 2 }f˜}L2 ď }f }L2 .

It is not very difficult to check that the map f ÞÑ Rf is linear. Thus R defines a linear
map from SpRd q to itself, continuous for the L2 norm. It is therefore extended by con-
tinuity into a continuous linear map on L2 pRd q.

For f P SpRd q we thus have

p´∆ ` 1qRf “ f.

12
We can then check that for any u P SpRd q we have

Rp´∆ ` 1qu “ u.

Thus R is on SpRd q the inverse function of p´∆ ` 1q (however, be careful, R is not a

bijection on L2 pRd q).

A final question about this example. Given f P SpRd q, can we give a slightly more
explicit expression for the solution Rf of the problem (9)? We have to compute the
inverse Fourier transform of the product of two functions. This is where the convolution
product appears (see Proposition 2.7). Thus, if G is a function which Fourier transform
is given by
1
Ĝpξq “ ,
1 ` |ξ|2
then we will have
x pξq “ Ĝpξqfˆpξq,
Rf
and hence
Rf pxq “ pG ˚ f qpxq.
It remains to determine G. At least for the dimension d “ 1, we have already done the
computations at example 2.3 and we get

e´|x|
Gpxq “ .
2
This gives
e´|x´y| e´|y|
ż ż
Rf pxq “ f pyq dy “ f px ´ yq dy.
R 2 R 2
We can now check with an explicit computation (left as an exercise) that if f P SpRq
then for any x P R we have

´pRf q2 pxq ` Rf pxq “ f pxq.

In dimension 1, the result is not very spectacular, because it is only an ordinary differ-
ential equation with constant coefficients, and we could have got the same result using
only a double variation of constants, but the method described on this simple example
will be useful to understand more subtle problems. . .

13