GAL Daniel Eceizabarrena - Distribution Theory and Fundamental Solutions of Differential Operators
GAL Daniel Eceizabarrena - Distribution Theory and Fundamental Solutions of Differential Operators
GAL Daniel Eceizabarrena - Distribution Theory and Fundamental Solutions of Differential Operators
Degree in Mathematics
Author:
Daniel Eceizabarrena Pérez
Supervisor:
Javier Duoandikoetxea Zuazo
Contents i
Introduction i
1 Distributions 1
1.1 The space D of testing functions . . . . . . . . . . . . . . . . . . . . 1
1.2 The space D0 of distributions . . . . . . . . . . . . . . . . . . . . . . 3
1.2.1 Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2.2 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2.3 Convergence . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.2.4 Operations on distributions . . . . . . . . . . . . . . . . . . . 8
1.2.5 Differentiation of distributions . . . . . . . . . . . . . . . . . 10
2 Tempered Distributions 15
2.1 The Schwartz space S . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.2 The space S 0 of tempered distributions . . . . . . . . . . . . . . . . . 20
2.2.1 Definition and some basic properties . . . . . . . . . . . . . . 20
2.2.2 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.2.3 Basic operations on tempered distributions . . . . . . . . . . 24
2.2.4 Convolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.2.5 Direct product . . . . . . . . . . . . . . . . . . . . . . . . . . 29
i
ii CONTENTS
Bibliography 67
Introduction
The objective of this dissertation is to study the theory of distributions and some
of its applications. Even if the vast majority of the theoretical results developed
throughout the Degree correspond to the time before 20th century, we have to focus
on the last 100 years to understand the matter of this document.
Certain concepts which we would include in the theory of distributions nowadays
have been widely used in several fields of mathematics and physics. For example,
many approaches to the very famous delta function were known to be used by
mathematicians such as Cauchy and Poisson, and some more recent as Kirchoff and
Hermite. In their works, they make use of some functions which they call impulse
functions. Nevertheless, it was Dirac who first introduced the delta function as we
know it, in an attempt to keep a convenient notation in his works in quantum mech-
anics. Their work contributed to open a new path in mathematics, as new objects,
similar to functions but not of their same nature, were being used systematically.
Distributions are believed to have been first formally introduced by the Soviet
mathematician Sergei Sobolev1 , in an attempt to find weak solutions to partial
differential equations. Nevertheless, it is the name of Laurent Schwartz2 which is
most deeply connected to distribution theory. His book Théorie des distributions,
published in 1950, is the source of the first systematic development of distributions,
and it highlighted their utility.
The aim of this project is to show how distribution theory can be used to obtain
what we call fundamental solutions of partial differential equations. The connection
between these two areas is given by the Fourier transform, probably one of the most
useful inventions of mathematics.
In short, distributions are linear and continuous functionals that assign a complex
number to every function of a certain space. The definition of the space of testing
functions, D, and of the space of distributions, D0 , will be given first. Neverthe-
less, the Fourier transform does not fit with the first and more basic definition of
distributions, so it will be necessary to introduce a more general space of functions,
the Schwartz space, S, in which the Fourier transform behaves properly. This fact
will naturally imply the existence of a new space of distributions, S 0 , which will be
known as the space of tempered distributions.
1
Sergei Lvovich Sobolev (1908, Saint Petersburg - 1989, Moscow) worked in analysis and partial
differential equations. The very famous Sobolev spaces are named after him.
2
See Appendix A.
i
ii INTRODUCTION
Distributions
2. ϕ is C ∞ .
Remark 1.2. Recall that the support of a continuous function is the closure of the
set of points in which the value of the function is not zero. In other words, supp ϕ =
Cl{x ∈ Ω | ϕ(x) 6= 0}. Because of this, it is enough to ask the support to be
bounded, as it will always be closed.
The first immediate property that can be obtained from the definition is given in
the following lemma. From now on, most of the times we will consider t = (t1 , · · · , tn )
to be the general variable in Rn , even if sometimes we will use x = (x1 , · · · , xn ) too.
∂ϕ
Lemma 1.3. If ϕ ∈ D(Ω), then ∂ti ∈ D(Ω), ∀i = 1, · · · , n.
∂ϕ
Proof. It is immediate, as ∂ti
is C ∞ because of the definition of ϕ. Moreover, if
t ∈
/ supp ϕ, as the complementary of the support is an open set, there exists a
∂ϕ
neighbourhood of t in which ϕ = 0. Thus, in that neighbourhood, ∂t i
= 0 and
∂ϕ ∂ϕ ∂ϕ
= 0, which implies that t ∈
∂ti (t) / supp ∂t i
. In other words, supp ∂t i
⊆ supp ϕ,
which completes the proof.
We will be able to give the space D an algebraic structure. The following propos-
ition shows why:
1
2 CHAPTER 1. DISTRIBUTIONS
Once we have defined the concept of testing function, let us show an example.
Example 1.5. Let ζ(t) be defined by parts:
( 1
e t2 −1 |t| < 1,
ζ(t) = (1.1)
0 |t| ≥ 1.
Clearly supp ζ = [−1, 1], and ζ is infinitely smooth whenever |t| 6= 1, so we only
have to check the situation at points |t| = 1. Anyway, as the function is even, it is
enough to analyse the case t = 1. We need to work out the derivatives in (−1, 1).
First,
2t 1
ζ 0 (t) = − 2 e t2 −1 , |t| < 1,
(t − 1)2
and it is fairly easy to observe that every derivative at |t| < 1 will be of the form
Pk (t) 1
ζ (k) = e t2 −1 , |t| < 1,
(t2− 1) 2k
Pk (t) 1
t2 −1 =
Pk (1) e1/2t
lim e lim = 0.
t→1− (t2 − 1)2k 4k t→0− t2k
This limit shows that every derivative of ζ is continuous, as ζ (k) (t) = 0 in |t| > 1.
As a consequence, ζ(t) ∈ D(R).
Remark 1.6. Most of the times, we will consider Ω = Rn , so from now on, we will use
the notation D(Rn ) (or simply D if it is clear the space we are working in) instead
of the more general D(Ω).
An interesting fact about testing functions is that one can produce new ones fairly
easy. The following proposition advocates so.
2. g(t)ϕ(t) ∈ D(Rn ).
1.2. THE SPACE D0 OF DISTRIBUTIONS 3
Proof. Every new function is infinitely smooth, as they are nothing but composi-
tions and products of C ∞ functions. Also, their supports are compact for being
translations, transpositions or dilations of that of ϕ, which is compact.
We will now define the concept of convergence in D, for it is key to understand
distributions, as we will later see.
Definition 1.8. Let {ϕm (t)}m∈N be a sequence of testing functions. We say it
converges to zero in D(Rn ) if
1. there exists K ⊆ Rn a bounded domain such that supp ϕm ⊆ K, ∀m ∈ N, and
2. each sequence {Dk ϕm (t)}m∈N converges uniformly to zero as m → ∞.
More generally, we say that the sequence {ϕm (t)}m∈N converges to ϕ(t) in D(Rn )
if every supp ϕm is contained in a single bounded domain K ⊂ Rn and the sequence
{ϕm (t) − ϕ(t)}m∈N converges to zero in D. In this case, we will write {ϕm } → ϕ, or
simply ϕm → ϕ.
Notation. We will use the notation Dk ϕ(t) with a multi-index k = (k1 , · · · , kn ) of
non-negative integers to represent the partial derivatives of the function ϕ. In other
words,
∂ k1 +···+kn
Dk ϕ(t) = ϕ(t1 , · · · , tn ).
∂tk11 · · · ∂tknn
We will also write |k| = k1 + · · · + kn .
Once we have defined the basics of testing functions, we are ready to introduce
the space of distributions.
1.2.1 Definition
Definition 1.9. A mapping T : D(Rn ) → C is called a distribution if:
1. it is linear, in the sense that if ϕ, φ ∈ D(Rn ) and a, b ∈ C, then
T (aϕ + bφ) = aT (ϕ) + bT (φ).
The space of distributions is denoted by D0 (Rn ). Some authors also call distri-
butions generalised functions.
Remark 1.10. In the same way we could define the space of testing functions over
an open set Ω ⊂ Rn , we can also define distributions over Ω. Anyway, as stated in
Remark 1.6, we will consider the whole space Rn .
A distribution can be represented by capital letters such as T , but it is usual to
use the same letters as for functions, such as f and g. Also, if ϕ ∈ D(Rn ), we will
write its image with f ∈ D0 (Rn ) as hf, ϕi.
As an immediate consequence of the definition, we can make the condition of
continuity less restrictive if we know a functional is linear.
1.2.2 Examples
After defining what a distribution is, it will be useful to present some examples of
distributions that will be useful later on.
Example 1.12. We say a function on Rn is locally integrable if it is integrable on
every compact subset of Rn . In this situation, every locally integrable function
can be treated as a distribution. In fact, let f be a locally integrable function
on Rn . Then, for any ϕ ∈ D(Rn ), we can define the integral
Z
hf, ϕi = f (x)ϕ(x)dx. (1.2)
Rn
R
Observe that if K = supp ϕ, then hf, ϕi = K f (x)ϕ(x)dx, and as K is compact
and ϕ is infinitely smooth, f ϕ is integrable on K, and therefore on Rn .
So once seen the functional is well-defined, let us prove it is a distribution. The
linearity of the integral makes it to be linear. Thus, we must check it is continuous.
By Proposition 1.11, we only have to consider the situation of sequences which
converge to zero. So let ϕn → 0 in D. We have to see that
Z
lim hf, ϕn i = lim f (x)ϕn (x)dx = 0.
n→∞ n→∞ Rn
The convergence in D implies that the sequence {ϕn (t)} converges to zero uniformly,
and also that there exists a compact K ⊂ Rn in which the support of every ϕn is
contained. In other words,
1.2. THE SPACE D0 OF DISTRIBUTIONS 5
So ∀n > N we get
Z Z Z
f (x)ϕn (x)dx ≤ |f (x)||ϕn (x)|dx ≤ |f (x)|dx
K K K
R R
and as K |f (x)|dx is finite, limn→∞ Rn f (x)ϕn (x)dx = 0.
Therefore, making the identification (1.2), we can say that f ∈ D0 (Rn ). Some
authors also write Tf to denote this functional. This kind of distributions are called
regular distributions.
The question that arises is whether all distributions are regular. Fortunately, there
are many more. This fact is the reason to call distributions generalised functions,
as they take in more objects than usual functions. The next two examples illustrate
this fact.
Example 1.13. Let ϕ be a testing function in Rn . The functional δ : D(Rn ) → C
given by
ζn (t) = ζ(nt), ∀n ∈ N.
6 CHAPTER 1. DISTRIBUTIONS
It holds for every n ∈ N that ζn (0) = ζ(0) = e−1 , and supp ζn = [− n1 , n1 ]. So we can
write Z Z 1/n
f (t)ζn (t)dt = f (t)ζn (t)dt = e−1 .
R −1/n
and |χ[− 1 , 1 ] (t)f (t)ζn (t)| ≤ e−1 χ[−1,1] (t)|f (t)|, so we can bound the integrand by an
n n
integrable function. Thus, we can take the limit in (1.4) inside by the dominated
convergence theorem, and
Z 1/n Z
lim f (t)ζn (t)dt = lim χ[− 1 , 1 ] (t)f (t)ζn (t)dt = 0.
n→∞ −1/n R n→∞ n n
This is a contradiction with (1.4), so the delta function is not a regular distribution.
Distributions which are not regular are called singular distributions.
Example 1.14. In general, we cannot define a distribution through the expression
(1.2) if the function we choose is not locally integrable. This is the case of f (t) = 1/t,
as it is not integrable around the origin in R. We can think of a similar definition
anyway. Let T : D(R) → R be defined the following way:
Z
ϕ(x)
hT, ϕi = lim dx. (1.5)
→0 |x|≥ x
Now, we know that ϕ has compact support, so we can find a positive constant k
such that supp ϕ ⊆ [−k, k], and thus the limit (1.5) turns into
k
ϕ(x) − ϕ(−x)
Z
lim dx.
→0 x
1.2. THE SPACE D0 OF DISTRIBUTIONS 7
1.2.3 Convergence
The space of distributions can be given a concept of convergence, which in some
cases will be useful.
Definition 1.15. Let {fn } be a sequence of distributions. We say that the sequence
converges to the distribution f if for every testing function ϕ, the sequence hfn , ϕi
converges to hf, ϕi.
We also say that the sequence {fn } simply converges if every sequence hfn , ϕi
converges.
It can be proven that if a sequence of distributions converges, then the functional
f which assigns to every testing function the limit value of the sequence defines a
distribution. A proof can be consulted in [10, p. 37-39].
Apart from seeing that these new functionals are well-defined, to check they define
distributions is an easy exercise. Linearity is granted by that of f . To see they are
continuous, let {ϕn } be convergent to zero in D. If we prove that the sequences
{ϕn (t + x)}, {ϕn (−t)}, {ϕn at } converge to zero in D, we will get the desired result
After being able to define a product of a distribution with a certain kind of func-
tions, we wonder whether a general product between distributions can p be defined.
Unfortunately, this is not the case. Think of the function f (t) = 1/( |t|). This is
a locally integrable function on R, thus defines a distribution. But f 2 (t) = 1/|t|,
which is known not to be integrable on neighbourhoods of the origin. Hence, we
cannot define a distribution through the expression
Z
1 ϕ(t)
h , ϕi = dt,
|t| R |t|
because the integral will not exist for every testing function ϕ. So in general, we
will not be able to define the product of distributions. This is one of the
major drawbacks, if not the principal, of the theory of distributions, although many
attempts to fix it have been carried.
Up to now, we have defined what we can call basic operations in the space D0 . But
it is possible to define much more powerful operations which will play an important
role in reaching our ultimate objective. Differentiation will be discussed next, and
some others as convolution and the direct product will be treated in Chapter 2.
As most of the properties for distributions, this definition has its origins in the
behaviour of regular distributions. To see this, let us consider a regular distribution
1.2. THE SPACE D0 OF DISTRIBUTIONS 11
f ∈ D0 (R), for which the function f is smooth. Then, its derivative f 0 is also smooth,
so it defines a distribution and the integration by parts yields
Z Z Z
0 0 0
hf , ϕi = f (t)ϕ(t)dt = (f ϕ) (t)dt − f (t)ϕ0 (t)dt.
R R R
But as ϕ is a testing function, its support is compact, and thus the fundamental
theorem of calculus forces the first integral to be zero. Because of this, we can write
hf 0 , ϕi = −hf, ϕ0 i, ∀ϕ ∈ D.
Proof. The functional is well-defined, Lemma 1.3 showing that every partial deriv-
ative of a testing function is also a testing function. Its linearity is provided by that
of the derivative. Now, take a sequence of testing functions {ϕn } which converges
to zero in D. Then, as seen in the proof of Lemma 1.3, the support of the derivative
of a testing function is contained in the support of the testing function itself, so
as there exists a subset K ⊂ Rn containing every supp ϕn , it also contains every
supp ∂ϕ α ∞
∂ti . Moreover, we know that every {D ϕn }n=1 converges to zero uniformly,
n
Once we have fixed the definition of the derivatives of distributions, we are ready to
analyse some properties, which in some cases will enhance the performance of usual
functions. For example, recall Schwarz’ theorem, which allows to commute partial
derivatives of a twice differentiable function. Not only is this true for distributions,
but it also holds with no restrictions at all.
∂2f ∂2f
=
∂ti ∂tj ∂tj ∂ti
Proof. It is an immediate consequence of the fact that testing functions are infinitely
smooth, thus Schwarz’ theorem holds for them.
It is also interesting to note that the product rule also holds for distributions.
12 CHAPTER 1. DISTRIBUTIONS
Proposition 1.19. Let f ∈ D0 (Rn ) and g ∈ C ∞ (Rn ). Then, for any i ∈ {1, · · · , n},
∂ ∂g ∂f
(gf ) = f +g .
∂ti ∂ti ∂ti
Proof. Let ϕ be a testing function in Rn . Then,
∂(gf ) ∂ϕ ∂ ∂g ∂f ∂g
h , ϕi = −hf, g i = −hf, (gϕ)i + hf, ϕi = hg , ϕi + h f, ϕi.
∂ti ∂ti ∂ti ∂ti ∂ti ∂ti
The derivative of distributions, by the property seen in Proposition 1.17, can be
seen as an operator in the space of distributions. In some moments, it will be useful
to know it is linear and continuous.
hδ, ϕi = ϕ(0),
∂δ ∂ϕ ∂ϕ
h , ϕi = −hδ, i = − (0).
∂ti ∂ti ∂ti
Now, as ϕ has compact support, we can find k > 0 such that the support is contained
in [−k, k], and
Z ∞ Z k
0
− ϕ (t)dt = − ϕ0 (t)dt = −ϕ(k) + ϕ(0) = ϕ(0) = hδ, ϕi.
0 0
Tempered Distributions
In this chapter we will introduce a new space of distributions, the space of tempered
distributions. Its importance will be covered in the next chapter, as the main mo-
tivation to define it is to be able to define the Fourier transform for distributions.
The space of tempered distributions, which we call S 0 , is a proper subspace of the
space D0 we already know. It can be defined, as we are to show, through a wider
class of testing fuctions.
2. for every m ∈ Z+ , k ∈ (Z+ )n , there exists a constant Cm,k > 0 such that
The space of all Schwartz functions is called Schwartz space, and it is denoted by
S(Rn ).
Remark 2.2. The Schwartz functions are also called functions of rapid descent.
The reason for this is that Condition 2 in Definition 2.1 is equivalent to
Cm,k
|Dk φ(t)| ≤ .
|t|m
This way, every Schwartz function and its derivatives must decrease to zero as |t|
tends to infinity, not slower than any power of 1/|t|.
Once we have defined the Schwartz space, it is time to present some very basic
properties.
15
16 CHAPTER 2. TEMPERED DISTRIBUTIONS
2. Dk φ ∈ S(Rn ).
Proof. By the triangle inequality, it is trivial to see that the sum of two Schwartz
functions and the product of a Schwartz function with a complex constant are again
Schwartz functions. This way, as every operation is performed in a field, S(Rn ) is a
linear space. Finally, to see 2 holds it is enough to check that for every i ∈ {1, · · · , n},
∂φ n ∞ implies that the partial derivative is so.
∂ti ∈ S(R ). Again, the fact that φ is C
For the second condition, we must see that for every m ∈ Z+ and k 0 ∈ (Z+ )n , there
0
exists Cm,k > 0 such that
k0 ∂φ
m
|t| D ≤ C0 0, ∀t ∈ Rn .
m,k
∂ti
In the beginning of the section, we have claimed our intention to define a space
wider than the space of testing functions. Let us state this fact in the following
proposition.
Theorem 2.4. If D(Rn ) is the space of testing functions and S(Rn ) is the Schwartz
space, then
D(Rn ) ⊂ S(Rn ),
the inclusion being strict.
So we see that |t|m |Dk φ(t)| is bounded, what shows that φ ∈ S, and thus D ⊆ S.
We want to prove that indeed D ⊂ S. So let φ be
2
φ(t) = e−|t| . (2.1)
Thus, if we consider |t| ≤ N , the expression (2.2) is bounded for being a C ∞ function
on a compact domain, say by a constant M > 0. Therefore, consider Cm,k =
max{M, }, and so,
|t|m |Dk φ(t)| ≤ Cm,k ,
Definition 2.5. Let {φn (t)}∞ n=1 be a sequence in the Schwartz space S. We say that
the sequence converges to zero in S if for every m ∈ Z+ , k ∈ (Z+ )n , the sequence
{|t|m Dk φn (t)}∞
n=1 converges to zero uniformly.
Following this definition, we say that the sequence {φn (t)}∞n=1 converges to φ if
the sequence {φn (t) − φ(t)}∞n=1 converges to zero.
Remark 2.6. As in the case of D, for simplicity, the sequence {φn (t)}∞
n=1 will also
be denoted φn , and convergence to the function φ will be written as φn → φ.
We have to take special care with the relationship between convergences in D and
in S. Suppose we are working with a sequence in S. Then, this sequence may not
be in D, so it makes no sense to ask about convergence in D. On the other hand,
every sequence in D is a sequence in S, so in order to keep a coherent structure, we
will have to ensure that convergence in D implies convergence in S. That is what
we present in the following proposition.
as the derivatives are zero outside K. Now, the convergence in D implies that
||Dk φn ||∞ → 0,
so the right hand side of (2.3) decays to zero as n → ∞. As a consequence,
sup |t|m |Dk φn (t)| → 0, as n → ∞.
Rn
This means that the sequence {|t|m Dk φn (t)} converges uniformly to zero. The proof
is now complete.
Granted that the definition of the Schwartz space is consistent with everything we
have defined before, we will introduce some operations on S, in a similar way as in
Proposition 1.7. But before that, we will present a little lemma, which will be useful
both to prove the following proposition and to get some results in next sections.
Lemma 2.8. Let a, b > 0 and m > 0. Then,
(a + b)m ≤ 2m (am + bm ).
Proof. We can prove it directly if we observe that
(a + b)m ≤ (2 max{a, b})m = 2m (max{a, b})m ≤ 2m (am + bm ).
Proposition 2.9. Let φ ∈ S(Rn ), τ ∈ Rn , a ∈ R − {0} and p(t) a polynomial in
Rn . Then,
1. φ(t + τ ), φ(−t), φ(at) ∈ S(Rn ).
2. p(t)φ(t) ∈ S(Rn ).
Proof. All of the newly defined functions are infinitely smooth for φ and p being so.
Now, for each of the cases, and for every m ∈ Z+ , k ∈ (Z+ )n , if we consider the
constants Cm,k for φ,
• Using the triangle inequality and Lemma 2.8, for φ(t + τ ),
|t|m |Dk φ(t + τ )| ≤ (|t + τ | + |τ |)m |Dk φ(t + τ )|
≤ 2m (|t + τ |m |Dk φ(t + τ )| + |τ |m |Dk φ(t + τ )|)
≤ 2m (Cm,k + |τ |m C0,k )
2.1. THE SCHWARTZ SPACE S 19
|t|m |Dk ϕ(t)| = |t|m |a|k| Dk φ(at)| = |a||k|−m |at|m |Dk φ(at)|
≤ |a||k|−m Cm,k .
∂k ∂ k−1 ∂k
(ti φ(t)) = k φ(t) + ti φ(t),
∂tki ∂tk−1
i
∂tki
There is one more property that plays a very important role when working with
Schwartz functions.
Proof. Let φ be a Schwartz function. We know that it is infinitely smooth, and for
every m ∈ Z+ and k ∈ (Z+ )n , there exists a constant Cm,k > 0 such that
|φ(t)|p (1 + |t|N )p
Z Z
p
|φ(t)| dt = dt,
Rn Rn (1 + |t|N )p
for some N ∈ N. Notice that the numerator can be bounded as follows:
So we need to check that the integral in (2.4) is finite. By a change to polar co-
ordinates, we can write
∞
rn−1
Z Z
1
dt = σ(Sn−1 ) dr
Rn (1 + |t|N )p 0 (1 + rN )p
Z 1 Z ∞ (2.5)
n−1 n−1 dr
≤ σ(S ) r dr + ,
0 1 rN p−n+1
where σ is a measure over the sphere Sn−1 . The first integral in (2.5) is finite, and
the second will be so if N p − n + 1 > 1, or equivalently, N > n/p. Thus, choosing
N as asserted, φ ∈ Lp (Rn ).
{T (φn )} → T (φ)
in C.
The space of tempered distributions is denoted by S 0 (Rn ). Some authors also call
tempered distributions distributions of slow growth. In general, tempered dis-
tributions will be represented using letters such as f and g as usual functions, and
the image f (φ) will be written as hf, φi.
S 0 (Rn ) ⊂ D0 (Rn ),
But the function φ defined by (2.1) in Theorem 2.4 is a Schwartz function, for which
Z Z Z
|t|2 −|t|2
hf, φi = f (t)φ(t)dt = e e dt = dt = ∞.
Rn Rn Rn
2.2.2 Examples
In the same way we did in Section 1.2.2, it is interesting to analyse some examples
of tempered distributions, for they will be important in the upcoming sections.
Example 2.15 (Delta function). Let us start with the most representative distri-
bution, the delta function. We saw in Section 1.2.2 that it defines a distribution.
We want to check it also defines a tempered distribution. We know that for every
φ ∈ S, hδ, φi = φ(0), so clearly, it is well-defined. We also checked that it is linear.
Now, consider a convergent sequence φn → 0 in S. Then, we know that choosing
constants m = 0 and k = (0, · · · , 0), the sequence {φn (t)} converges uniformly to
zero. Because of this,
This property can be exploited if we split the integral (2.6) into two terms:
Z Z Z
f (t)φ(t)dt = f (t)φ(t)dt + f (t)φ(t)dt. (2.9)
Rn |t|≤M |t|>M
2.2. THE SPACE S 0 OF TEMPERED DISTRIBUTIONS 23
Observe that if |t| ≤ M , then as φ ∈ C∞ (B(M )), it is bounded, and thus, the local
integrability of f implies that the first term of the right-hand side of (2.9) is finite.
It remains to deal with the second term. We can work with 1/|t|, because we are
far away from the origin. Thus, using (2.8),
Z Z Z
f (t)φ(t)dt ≤ |f (t)||φ(t)|dt ≤ |t|N |φ(t)|dt
|t|>M |t|>M |t|>M
(2.10)
Z
|t|N +n+1 |φ(t)|
= dt.
|t|>M |t|n+1
Now, as φ ∈ S(Rn ), there exists a constant CN +n+1,0 > 0 such that |t|N +n+1 |φ(t)| <
CN +n+1,0 , and thus, it follows from (2.10) that
Z Z
dt
f (t)φ(t)dt ≤ CN +n+1,0 .
n+1
|t|>M |t|
|t|>M
where σ represents a measure over the (n − 1)-dimensional sphere Sn−1 . This shows
that the operator is well defined.
Apart from that, it is clear that the functional (2.6) is linear. To see that it is
continuous, we must prove that if φm converges to zero in S, then hf, φm i does so
in C. The idea is the same as in proving that the functional is well defined. Indeed,
we split the integral into two parts as is (2.9) (considering φm instead of φ). The
first part is bounded by
Z
sup |φm (t)| |f (t)|dt,
|t|<M |t|<M
in which the integral is finite and the supremum tends to zero, because the conver-
gence of φm is uniform. On the other hand, for the second part we need (2.8) and
a procedure similar to (2.10) to get a bound given by
Z
dt
sup{|t|N +n+1 |φm (t)|} n+1
,
Rn |t|>M |t|
where the integral, the same as (2.11), is also finite and the supremum tends to zero
by the convergence of φm in S. So eventually we get the result and f defines a
tempered distribution.
Example 2.17. As a remark, it turns out that every bounded function is of slow
growth, as it can be easily checked from the definition given in (2.7). In particular,
every Schwartz function is of slow growth, as they are bounded (Theorem
24 CHAPTER 2. TEMPERED DISTRIBUTIONS
2.10). Moreover, as Schwartz functions are infinitely smooth, they are locally in-
tegrable, so each of them defines a tempered distribution. As a consequence, and
considering Schwartz functions and their corresponding tempered distributions to
be the same, this example and Theorems 2.4 and 2.13 show that we can display a
chain of relationships among every space discussed so far,
D ⊂ S ⊂ S 0 ⊂ D0 .
Example 2.18. Let f ∈ Lp (Rn ) for some p ≥ 1. Then, f defines a tempered distri-
bution. Indeed, for every φ ∈ S(Rn ), from the definition of being Schwartz we can
find a constant K such that
1 + |t|N
Z Z
|f (t)|
Z
f (t)φ(t)dt ≤ |f (t)||φ(t)| dt ≤ K dt. (2.12)
1 + |t|N N
Rn 1 + |t|
n
R Rn
Now, recall the computations in Theorem 2.10. We saw in the proof, in (2.4) and in
(2.5) that the function 1/(1 + |t|N ) is in Lp whenever N > n/p. Therefore, taking
any N > n, it is in every Lp space, for p ≥ 1. Therefore, considering q to be the
Hölder conjugate of p, by Hölder’s inequality we can write
|f (t)|
Z
1
N
dt ≤ ||f ||p
N
< ∞.
R n 1 + |t| 1 + |t|
q
|hαf, φn i| ≤ |α||hf, φn i| → 0.
Remark 2.19. Notice that in this case, we cannot define the product of a tempered
distribution with a smooth function. In fact, not every product of a smooth function
and a Schwartz function is a Schwartz function. It is enough to consider φ(t) =
2 2
e−|t| ∈ S as in 2.1 and g(t) = e|t| .
However, it can be proved that a tempered distribution can be multiplied by any
C ∞ function for which itself and all its derivatives are of slow growth. If ψ is of that
kind, then φψ is Schwartz for every Schwartz function φ, and the sequence ψφn will
converge to zero if φn does so. Thus, for every f ∈ S 0 , ψ · f will define a tempered
distribution by the usual definition
hψ · f, φi = hf, ψφi, ∀φ ∈ S 0 .
In particular, as every Schwartz function is of slow growth and as all derivatives are
Schwartz functions and thus of slow growth, multiplication between tempered
distributions and Schwartz functions is permitted.
We cannot forget about one of the main feature of distributions, the derivative.
∂f
6. The derivative of a tempered distribution, ∂ti . It is immediate from the
definition of convergence in S that the sequence { ∂φ
∂ti } converges to zero in S,
n
2.2.4 Convolution
The ones presented in 2.2.3 are not the only operations we can define in the space
of tempered distributions. Indeed, we will need to work with not so basic ones when
we deal with fundamental solutions of differential operators in Chapter 4. The first
new operation we will work with is convolution. It is possible, and useful, to define
the convolution between tempered distributions under certain conditions, and even
on the space of usual distributions. However, we will focus on the convolution
between a tempered distribution and a Schwartz function, as it will be
enough to achieve our goal.
Recall the definition of the convolution between two integrable functions
f and g, Z
(f ∗ g)(x) = f (y)g(x − y)dy.
Rn
∂g
Let k ∈ {1, · · · , n}, and suppose both g and ∂tk are continuous. Moreover, suppose
∂g
∂tk is bounded by an integrable function in the variable x. In this situation,
Z
∂f ∂g
(t) = (t, x)dx.
∂tk Rn ∂tk
Once we know how to manage this situation, we present the following auxiliar
proposition.
Proof. We know that both f and g are smooth. As we want to compute, for every
k ∈ {1, · · · , n}, the derivative
Z
∂
f (x − y)g(y)dy,
∂xk Rn
we can use Lemma 2.21 to take the derivative inside, and thus
Z Z
∂ ∂
f (x − y)g(y)dy = (f (x − y)g(y))dy
∂xk Rn Rn ∂xk
Z
∂f ∂f
= (x − y)g(y)dy = ∗ g (x),
Rn ∂xk ∂xk
Remark 2.23. If we read the previous proof carefully, we will realize that g need not
be so regular. Indeed, it would be enough for it to be continuous and integrable. In
any case, if g is also Schwartz, by commuting the definition of convolution, the same
proof shows that
∂f ∂ ∂g
∗ g (x) = (f ∗ g)(x) = f ∗ (x).
∂xk ∂xk ∂xk
Now, writing |x| = |x − y + y|, we can use Lemma 2.8 to assert that
Z
|x|m |Dk (f ∗ g)(x)| ≤ 2m |x − y|m |g(x − y)||Dk f (y)|dy
R n
Z (2.13)
+ |y|m |g(x − y)||Dk f (y)|dy .
Rn
Now, |x − y|m |g(x − y)| and |y|m |Dk f (y)| can be bounded for the definition of f
and g being Schwartz, and as every Schwartz function is integrable (Theorem 2.10),
both Dk f (y) and g(x − y) are so, thus the sum of integrals (2.13) is bounded by a
constant depending on m and k, but not on x.
28 CHAPTER 2. TEMPERED DISTRIBUTIONS
The two integrals on the right-hand side of (2.15) are finite. Indeed, the first one
is concerning a Schwartz, thus integrable function. The second integrand can be
bounded by a factor C/|y|n+1 , where C is a constant, for ψ̃ being Schwartz. There-
fore, the convergence of φl in S makes both suprema tend to zero, what implies the
convergence of ψ̃ ∗ φn . Finally, the continuity of f implies that of f ∗ ψ.
We have seen that the derivatives are fairly easy to handle when working with
Schwartz functions. The next result shows that the situation does not change when
considering tempered distributions.
∂ ∂f ∂ψ
(f ∗ ψ) = ∗ψ =f ∗ , k = 1, · · · , n.
∂xk ∂xk ∂xk
First of all, let us present what intuitively the direct product could be. Let f and
g be locally integrable functions on Rn . Then, the function given by f (x)g(y) on
Rm+n is also locally integrable, thus defines a distribution,
Z
hf (x)g(y), ϕ(x, y)i = f (x)g(y)ϕ(x, y)dxdy, ∀ϕ ∈ D(Rn+m ).
Rn+m
For ϕ being bounded, Fubini’s theorem can be applied to consider the most con-
venient order. In this case, let us write
Z Z Z
f (x)g(y)ϕ(x, y)dxdy = f (x) g(y)ϕ(x, y) dx,
Rn+m Rn Rm
what suggests a definition for the direct product of distributions. In fact, we can
assert that the definition is correct.
Definition 2.29. Let f ∈ D0 (Rn ) and g ∈ D0 (Rm ) be any two distributions. Their
direct product is defined as
We have seen that the definition works fine for locally integrable functions, but
when it comes to distributions, we have to ensure that this definition makes sense.
For instance, we should check that the expression hg(y), ϕ(x, y)i defines a testing
function on Rn whenever g is a distribution on Rm and ϕ is a testing function on
Rn+m . The following lemma asserts so, although we will not prove it. A proof can
be found in [8, p.96-98].
Lemma 2.30. For every distribution g ∈ D0 (Rm ) and every testing function ϕ ∈
D(Rn+m ), the function
ψ(x) = hg(y), ϕ(x, y)i
is a testing function on Rn . Moreover, its derivatives are given by
for every multi-index α, where the notation Dxα stands for the partial derivatives
with respect to x and with indices α. Also, if ϕn → ϕ in D(Rn+m ), then
in D(Rn ).
This lemma allows us to prove that the direct product is indeed a distribution.
2.2. THE SPACE S 0 OF TEMPERED DISTRIBUTIONS 31
It can also be proved that the direct product is a commutative operation. This
means that for every distributions f and g, it holds that
hδ(x) · δ(y), ϕ(x, y)i = hδ(x), ϕ(x, 0)i = ϕ(0, 0) = hδ(x, y), ϕ(x, y)i,
In this chapter, we will study what the effect of applying the Fourier transform to
functions in S is, in order to define the Fourier transform of distributions. Several
problems appear when the Fourier transform is defined for the space D, but it works
properly when we extend the space to S. We will discuss this along next pages.
33
34 CHAPTER 3. THE FOURIER TRANSFORM
The integrand is bounded by 2|f (x)| and the limit can be taken inside. It is clear
that it tends to zero.
Definition 3.3. The Fourier transform on the Schwartz space S(Rn ) is an operator
F : S(Rn ) → L∞ (Rn ) which asigns to every φ ∈ S(Rn ) the function
Z
Fφ(ξ) = φ̂(ξ) = φ(x)e−i(ξ·x) dx.
Rn
Our next objective is to see that the Fourier transform is a bijective endomorphism
(thus an automorphism). In other words, we want to prove:
1. that the Fourier transform of every Schwartz function is also a Schwartz func-
tion, and
To get these two results, it is vital to know how to manage the derivatives of
the Fourier transform of a function, as the definition of S is heavily dependent on
differentiation. For that reason, we will fist develop some properties concerning F.
properties which are probably the major advantage of working in the Fourier space.
These properties are the ones which make a direct relation between differentiation
and multiplication. Let us analyse them in detail.
Proof. 1. We want to make the derivative of the Fourier transform φ̂(ξ). Let
i ∈ {1, · · · , n}, and
Z
∂ φ̂ ∂
(ξ) = φ(x)e−i(ξ·x) dx.
∂ξi ∂ξi Rn
Observe that
∂ −i(ξ·x)
φ(x)e = φ(x)(−ixi )e−i(ξ·x) = |xi φ(x)|.
∂ξi
2. Choose again i ∈ {1, · · · , n}. We will compute the Fourier transform of the
derivative. By definition,
Z
∂φ ∂φ
F( )(ξ) = (x)e−i(ξ·x) dx.
∂xi Rn ∂x i
∂φ −i(ξ·x) =
Observe that the function we are integrating is in L1 , because ∂x i
(x)e
∂φ ∂φ
∂xi (x), and ∂xk ∈ S ⊆ L1 , as we saw in Proposition 2.3 and in Theorem 2.10.
So the order of the integral can be changed by Fubini’s theorem, and
Z Z Z
∂φ ∂φ
(x)e−i(ξ·x) dx = (x)e−i(ξ·x) dx. (3.2)
Rn ∂xi R n−1 R ∂x i
36 CHAPTER 3. THE FOURIER TRANSFORM
Now, integrating by parts, the boundary term vanishes for φ being Schwartz,
and Z Z
∂φ −i(ξ·x)
(x)e dxi = iξk φ(x)e−i(ξ·x) dxi .
R ∂xi R
So rewriting (3.2),
Z Z
∂φ
(x)e−i(ξ·x) dx = iξk φ(x)e−i(ξ·x) dx.
Rn ∂xi Rn
∂φ
In other words, F( ∂xi
)(ξ) = iξk F(φ). Applying this last expression |k| times,
Proposition 3.4 conveys how derivatives work in interaction with the Fourier trans-
form. This is key to prove our first objective.
1. φ̂ ∈ C ∞ , and
2. for every m ∈ Z+ , k ∈ (Z+ )n , ∃Cm,k > 0 | |ξ|m |Dk φ̂(ξ)| < Cm,k .
1. By Proposition 3.4,
Observe now that Dim (xk φ) is a sum of several products between polynomials
and derivatives of φ. For that, Dim (xk φ) ∈ S holds. This means that it is
integrable, and thus, Mi = Rn |Dim (xk φ(x))|dx is finite.
R
√ m
Also observe that |ξ|m ≤ n maxi=1,··· ,n |ξi |m , so we can write
√ m √ m
|ξ|m |Dk φ̂(ξ)| ≤ n max |ξim Dk φ̂(ξ)| ≤ n max Mi < ∞. (3.4)
i=1,··· ,n i=1,··· ,n
3.2. THE FOURIER TRANSFORM IN S 37
Some ideas of this proof yield to the fact that the Fourier transform F is a linear
continuous operator from S to S.
Proof. It is trivial to see that the operator F is linear. To see that it is continuous,
for the same reason as in Proposition 1.11, it is enough to check the case {φl }∞
l=1 → 0.
m k
Supposing so, we have to see that every sequence |ξ| D φ̂l (ξ) → 0 uniformly, being
m ∈ Z+ , k ∈ (Z+ )n . But if we consider (3.4) from the proof of Theorem 3.5, we
know that
√ m
|ξ|m |Dk φ̂l (ξ)| ≤ n max Mi,l ,
i=1,··· ,n
where Z
Mi,l = |Dim (xk φl (x))|dx. (3.5)
Rn
Therefore, it is enough to check that maxi=1,··· ,n Mi,l tends to zero as l → ∞.
As the result of expanding the derivative of (3.5) will be a finite sum of products of
polynomials and derivatives of φl , the convergence of φl implies that the derivatives
will converge uniformly to zero.
Moreover, if we choose q ∈ N, the expression |x|q |Dim (xk φl )(x)| will be bounded by
the finite sum mentioned above times |x|q , which, by the convergence of φl in S, will
be bounded by a constant Cq,m (it will be the sum of the constants corresponding
to each summand). This way,
Now, as we can have a bound with a integrable function, we can apply the dominated
convergence theorem to assert that
lim Mi,l = 0, ∀i = 1, · · · , n.
l→∞
Instead of using the definition, we will think in an alternative way. First of all,
observe that, if we allow a little abuse of notation when considering g in one variable,
2 2 2 2 2
g(x) = e−k|x| = e−k(x1 +...+xn ) = e−kx1 · . . . · e−kxn = g(x1 ) · . . . · g(xn ). (3.7)
Also observe that as g ∈ S (recall Theorem 2.4, equation (2.1)), it is integrable, and
Fubini’s theorem allows us to split the integral using (3.7), and get
so (3.9) is equivalent to
iξFg(ξ) + 2kiF(g)0 (ξ) = 0,
which in turn is
ξ
F(g)0 (ξ) +
Fg(ξ) = 0.
2k
A solution can be easily found; it must be of the form
ξ2
Fg(ξ) = Ce− 4k , C ∈ R.
By repeating every calculation for the case where k > 0, we can assert that
2 /4z
ĝ(ξ) = Ke−ξ ,
This integral cannot be computed as we did in the real case. It needs an special
treatment. So, for a > 0, consider
Z
2
Ia (t) = e−(a+ti)x dx.
R
Anyway, the integral in 3.14 is not absolutely convergent, and thus, Fubini’s theorem
cannot be applied. To solve this problem, we will need some properties of the Fourier
transform, concerning its behaviour with respect to certain operations.
1. If τ ∈ Rn , then
F(φ(x − τ ))(ξ) = e−iξ·τ Fφ(ξ). (3.15)
2. If λ ∈ R, then
1 ξ
F(φ(λx))(ξ) = Fφ . (3.16)
|λ|n λ
3. If φ, ϕ ∈ S(Rn ), then
Z Z
Fφ(x)ϕ(x)dx = φ(x)Fϕ(x)dx. (3.17)
Rn Rn
Proof. Properties 1 and 2 are a direct result of simply applying a change of variables
in the integral of the definition of the Fourier transform. Finally, property 3 is a
direct consequence of Fubini’s theorem. Observe that Fubini’s results can be applied
as every Schwartz function is integrable (see Theorem 2.10).
Having presented these properties, we are ready to face our main objective.
This is the time when the Gaussian comes into play. If we consider g to be the
Gaussian function (3.6) with k = 1/2,
2 /2
g(x) = e−|x| ,
then (3.11) and (3.19) convey that
Z √ nZ
ˆ
f (x)dx = f (0)( 2π) g(x)dx = (2π)n f (0).
Rn Rn
Reordering, Z
1
f (0) = fˆ(x)dx, (3.20)
(2π)n Rn
which is the result we seek at ξ = 0, valid for every Schwartz function. The last step
is to use (3.15). Let ξ ∈ Rn . Considering f (x + ξ) ∈ S, then by (3.20) and (3.15),
Z Z
1 1
f (ξ) = F(f (x + ξ))(y)dy = e−i(y·(−ξ)) fˆ(y)dy
(2π)n Rn (2π)n Rn
Z
1
= fˆ(x)ei(ξ·x) dx,
(2π)n Rn
which is the definitive result. This shows that F −1 Ff = f for every Schwartz
function f . The symmetrical result is achieved in a similar way.
With this result, we are allowed to say that the inverse Fourier transform of every
Schwartz function is again a Schwartz function, and that the operator F −1 is also
linear and continuous, by the same arguments used for F.
Theorem 3.10. The inverse Fourier transform of every Schwartz function is a
Schwartz function, and the operator F −1 : S → S is a linear and continuous oper-
ator.
Proof. That F −1 is an endomorphism in S is a direct consequence of Theorem 3.9.
The linearity and continuity of the inverse operator can be obtained by a similar
procedure as in Theorem 3.6.
42 CHAPTER 3. THE FOURIER TRANSFORM
1. F −1 φ(ξ) = 1
(2π)n F φ̃(ξ).
2. F φ̃(ξ) = Fφ(−ξ).
As we said in Section 2.2.4, convolution will play an important role. For this
reason, we need to know how to manage its Fourier transform.
1.
F(φ ∗ ϕ)(ξ) = Fφ(ξ)Fϕ(ξ). (3.21)
2.
1
F(φϕ)(ξ) = (Fφ ∗ Fϕ)(ξ). (3.22)
(2π)n
F(φ ∗ ϕ) = Fφ · Fϕ.
The equality ||Fφ||2 = (2π)n/2 ||φ||2 for φ ∈ S and the fact that S is dense in L2
can be used to extend the Fourier transform to L2 . Nevertheless, for our purposes,
we need to extend it to the space of tempered distributions, where we know L2 is
contained.
In the same way every operation analysed before has a justification in terms of
regular distributions, Definition 3.14 can also be justified by considering Schwartz
functions as tempered distributions. In fact, if φ is a Schwartz function and we treat
it as a distribution, then for every ϕ ∈ S,
Z
hFφ, ϕi = Fφ(x)ϕ(x)dx,
Rn
44 CHAPTER 3. THE FOURIER TRANSFORM
thus the definition. It is important to check that we are defining a tempered distri-
bution.
Proof. The hardest part of this proof is already done in previous sections. Definition
3.14 and Proposition 3.5 show the new functional Ff is well-defined. The linearity
of both the Fourier transform in S and the tempered distribution f show that Ff is
also linear. Last, we saw in Theorem 3.6 that the Fourier transform is a continuous
operator in S, showing that if a sequence φn converges to zero in S, then so does the
sequence φ̂n . Thus, because of the continuity of f , the sequence hFf, φn i = hf, φ̂n i
converges to zero in C, showing that Ff is continuous.
We can also define the inverse Fourier transform in S 0 in the same way we did it
for the direct one.
Theorem 3.18. The Fourier inversion formula still works in S 0 . More precisely,
for every tempered distribution f ,
FF −1 f = f and F −1 Ff = f.
2. xk Ff = (−i)|k| F(Dk f ).
3. F 2 f = (2π)n f˜.
3.3. THE FOURIER TRANSFORM IN S 0 45
4. F 4 f = (4π 2 )n f.
Proof. The properties are direct consequence of those which hold for Schwartz func-
tions, given in Propositions 3.4 and 3.11.
It is also easy to show that the Fourier transform is a linear and continuous
operator in the space of tempered distributions. For this, we must recall the concept
of convergence of tempered distributions, which is absolutely identical to that of
usual distributions defined in Section 1.2.3.
Theorem 3.20. The Fourier transform operator F : S 0 (Rn ) → S 0 (Rn ) and the in-
verse Fourier transform operator F −1 : S 0 (Rn ) → S 0 (Rn ) are linear and continuous
operators.
F(f ∗ ψ) = Ff · Fψ.
to assert that
hF(f ∗ ψ), φi = hFf, Fψ · φi = hFf · Fψ, φi.
Finally, we want to see what is the effect of applying the Fourier transform to a
direct product of tempered distributions. The result is surprisingly simple.
3.3.2 Examples
The Fourier transform plays an important role in the pursuit of fundamental solu-
tions of differential equation, as we will analyse in the following chapters. For this
reason, it is interesting to work out the Fourier transform of some distributions. In
this section, we will work with the delta function and polynomials.
Example 3.23. We want to check that for every multi-index k ∈ (Z+ )n and τ ∈ Rn ,
F(Dk [δ(x − τ )])(ξ) = i|k| ξ k e−iξ·τ . (3.24)
For that, we only need to use properties we presented in Section 3.3.1. Indeed,
choosing any Schwartz function φ,
hF(Dk [δ(x − τ )]), φi = (−1)|k| hδ(x − τ ), Dk φ̂(x)i
= (−1)|k| (−i)|k| hδ(x − τ ), F(ξ k φ)i
(3.25)
Z
|k| |k|
k
= i F(ξ φ)(τ ) = i ξ k φ(ξ)e−i(τ ·ξ) dξ
Rn
|k| k −i(ξ·τ )
= hi ξ e , φi.
In particular, for k = 0 and τ = 0,
Fδ(ξ) = 1(ξ). (3.26)
3.3. THE FOURIER TRANSFORM IN S 0 47
i|k| xk e−ix·τ .
and we will use the mentioned (3.24) to prove so. Indeed, applying F and recovering
results from Proposition 3.19, we get
h(Dk [δ(ξ − τ )])(−ξ), φ(ξ)i = (−1)2|k| Dk φ(−τ ) = (−1)|k| hDk δ(ξ + τ ), φi,
For that reason, the convergence does not change even if we apply the Fourier
transform, so by (3.13),
r n
π 2 2
F(g )(ξ) = e−|ξ| /4(+bi) → F(e−bi|x| )
+ bi
48 CHAPTER 3. THE FOURIER TRANSFORM
what shows that the formula shown in (3.13) is also valid for an imaginary number.
Chapter 4
Fundamental Solutions of
Differential Operators
In this last chapter, we will develop an efficient and general theory to reach what
we will call fundamental solutions of some differential operators. Eventually, we will
apply this theory to analyse some representative cases such as the heat equation,
the Schrödinger equation and the Laplace equation.
In general, we are looking for a distribution u ∈ D0 (Rn ) which satisfies (4.3). Notice
that this solution may not be a usual solution we are used to, for it need not be a
function, and even if it is a function, its derivatives need not be functions. This is
where the concept of generalised solution comes from (also remember that another
name for distributions is generalised functions).
49
CHAPTER 4. FUNDAMENTAL SOLUTIONS OF DIFFERENTIAL
50 OPERATORS
In this situation, we will be able to obtain some particular solutions which will be
of extreme importance.
We will work with these fundamental solutions from now on, but of course, it is
important to know the reason for this. Notice that the connection between recently
defined generalised solutions and fundamental solutions remains unknown. The
following theorem will help to enlighten the situation.
Theorem 4.3. Let L(D) be a differential operator with constant coefficients and E
a tempered fundamental solution of it. Let also f be a Schwartz function on Rn .
Then, a solution to the equation L(D)u = f is given by
u = E ∗ f.
Recall the rules of differentiation of the convolution given in (2.23); we can write
m
X m
X
aα Dα (E ∗ f ) = aα (Dα E ∗ f ).
|α|=0 |α|=0
4.1. GENERALISED AND FUNDAMENTAL SOLUTIONS 51
as we saw in (2.16) that the delta function is the identity element with respect to
convolution. As a consequence, the distribution E ∗ f is a solution to the equation
in the statement.
The reason for we bring fundamental solutions into the limelight is clear now. If
we manage to obtain one for an operator, much of the work will be done. In any
case, it is important to remark that, in general, they are not unique. Suppose we
have been able to get a solution to the homogeneous equation L(D)u = 0. If we call
it E0 , then
L(D)(E + E0 ) = L(D)E + L(D)E0 = δ + 0 = δ,
showing that E + E0 is also a fundamental solution.
It would be interesting to get, if possible, a procedure to work out fundamental
solutions. Here is where the Fourier transform comes into play.
Theorem 4.4. Let L(D) be a differential operator with constant coefficients and
E ∈ S 0 (Rn ). Then, E is a fundamental solution of L(D) if and only if
L(iξ)F(E)(ξ) = 1,
Once we know this, consider L(D)E = δ. Then, applying the Fourier transform and
by (4.6),
F(E)(ξ)L(iξ) = Fδ = 1.
Conversely, suppose F(E)(ξ)L(iξ) = 1. Then, by the inverse Fourier transform and
(4.6),
F(L(D)E) = 1 ⇒ L(D)E = F −1 Fδ = δ.
The question is how we could exploit this interesting result. It conveys that it is
enough to solve
L(iξ)F(E)(ξ) = 1,
or what we could expect to be the same,
1
F(E)(ξ) = . (4.7)
L(iξ)
CHAPTER 4. FUNDAMENTAL SOLUTIONS OF DIFFERENTIAL
52 OPERATORS
But we must be careful at this point. To consider (4.7), we need to be sure that the
expression on the right hand side is a distribution. Indeed, it is a usual function.
If it were locally integrable and of slow growth, we could treat it as a tempered
distribution and work with expression (4.7), being our solution
−1 1
E=F .
L(iξ)
But matters are not so simple if the function turns out not to be locally integrable.
In that case, we will need to manage to obtain a tempered distribution F solving
the equation
L(iξ)F = 1,
to finally obtain the result by the inverse Fourier transform, E = F −1 (F ).
where Dxi represents the partial derivative with respect to the variable xi . We know
that if we want to obtain a fundamental solution of LH (D), we need to solve the
equation
LH (D)E = Dt E(x, t) − a2 ∆x E(x, t) = δ(x, t).
Instead of applying the general Fourier transform as we did in Section 4.1, we will
only consider the Fourier transform on the variable x. For that,
∂E
Fx − a2 Fx (∆x E) = Fx (δ). (4.10)
∂t
4.2. THREE FUNDAMENTAL SOLUTIONS 53
An application of Lemma 2.21 yields to say that the Fourier transform in x and the
partial derivative in t commute, as they are concerning different variables, so
∂E ∂
Fx = Fx (E)(ξ, t). (4.11)
∂t ∂t
Also observe that
n
X
Fx (∆x E) = Fx (Dx2i E),
i=1
and by properties seen in Proposition 3.19,
It remains to handle the right hand side of (4.10). In this case, by properties (2.18)
and (3.22),
Fx (δ(x, t)) = Fx (δ(x) · δ(t)) = Fx (δ(x)) · δ(t),
as can be easily checked. We also know that Fx (δ(x)) = 1(ξ), so
where the only variation is that Dt (1(ξ) · H(t)) = 1(ξ) · Dt H(t). But we can simplify
that expression. In fact,
2 |ξ|2 t 2 2
h(1(ξ) · H(t))e−a , ϕ(ξ, t)i = hH(t), h1(ξ), e−a |ξ| t ϕ(ξ, t)ii
Z Z
2 2
= H(t)e−a |ξ| t ϕ(ξ, t)dξdt
R Rn
2
= hH(t)e−a|ξ| t , ϕ(ξ, t)i.
For this,
2 |ξ|2 t
F (ξ, t) = H(t)e−a .
In short, if the fundamental solution we are looking for is E, then by the inverse
Fourier transform,
2 2
E = Fξ−1 (H(t)e−a |ξ| t ).
Clearly, the transform has no effect on H, and as properties 1 and 2 from Proposition
3.11 can be generalised to distributions,
H(t) 2 2
E= n
Fξ (e−a |ξ| t ).
(2π)
But we know how to work out the Fourier transform of the Gaussian; we analysed
it in section 3.2.2. Thus,
r n
H(t) π 2 2 H(t) 2 2
E(x, t) = n 2
e−|x| /4a t = √
n
e−|x| /4a t . (4.17)
(2π) a t (2a πt)
We will make use of the Fourier transform of the Gaussian to get the result.
Indeed, recall from Section 3.2.2 that
r n
−k|x|2 π 2
F(e )(ξ) = e−|ξ| /4k .
k
Also recall that hFf, φi = hf, Fφi for every Schwartz function φ, from the
definition of the Fourier transform in S 0 . So taking f to be the Gaussian, we
can write the following equality.
r n Z Z
π −|ξ|2 /4k 2
φ(ξ)e dξ = φ̂(x)e−k|x| dx. (4.20)
k Rn Rn
and the value is finite for 1/|x|2 being locally integrable and of slow
growth.
– The left hand side is a bit more tricky. Again by Fubini’s theorem, we
can write equivalently
Z Z ∞
π n/2 −|ξ|2 /4k
φ(ξ) e dk dξ.
Rn 0 k
CHAPTER 4. FUNDAMENTAL SOLUTIONS OF DIFFERENTIAL
56 OPERATORS
we can write
√ √
(2 π)n ∞ n −2 −r (2 π)n n
Z
r 2 e dr = Γ − 1 .
4|ξ|n−2 0 4|ξ|n−2 2
Remember that we seek F −1 (1/|ξ|2 ). We can use equation (4.22) for that,
because we know that
Γ n2 − 1
−1 1 1 1 1
F = F = √ .
|ξ|2 (2π)n |ξ|2 4( π)n |x|n−2
Γ n2 − 1
1
E(x) = − √ n .
4( π) |x|n−2
and observe that every En defines a regular distribution for being continuous
and thus locally integrable, being also of slow growth. We want to see that
En → E in S 0 (R2 ) if n → ∞. Let φ ∈ S(R2 ). Then,
Z
1 1
hEn , φi = log |x|2 + 2 φ(x)dx.
2 R2 n
Observe that | log(|x|2 + 1/n2 )φ(x)| ≤ (| log(|x|2 + 1)| + | log(|x|2 )|)|φ(x)|, and
the right-hand side function is integrable, because logarithms are locally integ-
rable functions of slow growth and φ is Schwartz, so the sum of the products is
integrable. For this, we can use the dominated convergence theorem and say
that
Z Z
1 2
lim hEn , φi = log |x| φ(x)dx = log |x|φ(x)dx = hE, φi.
n→∞ 2 R2 R2
2/n2
Z Z
2 x
2 2 2
φ(x)dx = 2 2
φ dx.
R2 (|x| + 1/n ) R2 (|x| + 1) n
The question now is if we can commute the limit with the integral. The
answer is positive, as for being Schwartz, φ is bounded, and by a change to
polar coordinates,
Z Z ∞
2 rdr
2 2
dx = 2π = 2π.
R2 (|x| + 1) 0 (1 + r2 )2
So taking every step done together,
Z
2
h∆E(x), φi = 2 2
φ(0)dx = 2πφ(0) = 2πhδ, φi.
R2 (|x| + 1)
CHAPTER 4. FUNDAMENTAL SOLUTIONS OF DIFFERENTIAL
58 OPERATORS
ux = vy and uy = −vx .
To obtain an expression for the operator, let us define the following operators in R2 :
1 1
Dz = (Dx + iDy ) and Dz = (Dx − iDy ) .
2 2
One can prove that the condition Dz f = 0 is the same as saying that u and v satisfy
the Cauchy-Riemann conditions. It is also easy to check that
1
Dz Dz f = Dz Dz f = ∆f.
4
1
We have seen in Section 4.2.2 that the function 4π log |x2 + y 2 | is the fundamental
solution of the Laplace operator in the plane. This can be used to obtain the
1
fundamental solutions to both Dz and Dz . Indeed, calling E = 4π log |x2 + y 2 |,
1 1
Dz Dz E = Dz Dz E = ∆E = δ(x, y).
4 4
Therefore, it is clear that the fundamental solutions of the operators Dz and Dz are,
respectively, 4Dz E and 4Dz E, or explicitly,
1 1 1 1
Dz log |x2 + y 2 | = and Dz log |x2 + y 2 | = .
π πz π πz
As stated before, we will work with f ∈ S(Rn ). In the following lines, we will try to
obtain a solution through a fundamental solution. In the case of the heat operator,
we need to solve
(
Et (x, t) − a2 ∆x E(x, t) = 0, for t > 0, x ∈ Rn ,
(4.25)
E(x, 0) = δ(x), for x ∈ Rn .
The method for solving this problem is to use the Fourier transform in variable x in
a similar way we did in 4.2.1. So applying F to the first equation in (4.25),
for some function c(ξ). This function can be determined by using the initial condi-
tion. In fact,
Fx (E)(ξ, 0) = Fx (E(x, 0))(ξ) = F(δ)(ξ) = 1(ξ),
and we deduce that,
2 |ξ|2 t
Fx (E)(ξ, t) = e−a . (4.27)
CHAPTER 4. FUNDAMENTAL SOLUTIONS OF DIFFERENTIAL
60 OPERATORS
So we have to compute its inverse Fourier transform in order to get the fundamental
solution function E. Observe that the function of which inverse Fourier transform
we have to compute is the same we computed to obtain (4.17), so
2 |ξ|2 t 1 2 2
E(x, t) = Fξ−1 (e−a )= √
n
e−|x| /4a t
(2a πt)
is the fundamental solution to the Cauchy problem of the heat operator. Now, as
we said in the beginning of the section, the solution to the Cauchy problem is given
by a convolution in the variable x,
Z
1 −|x|2 /4a2 t 1 2 2
u(x, t) = f ∗x √
n
e = √
n
f (y)e−|x−y| /4a t dy. (4.28)
(2a πt) (2a πt) Rn
u(x, 0) = f (x).
In this section, we will try to obtain a solution for the Cauchy problem corresponding
to the Schrödinger equation, so we will try to obtain a fundamental solution E which
fulfils (4.29) and also
E(x, 0) = δ(x).
Once we get that, the solution will be of the form E ∗x f .
In the same way we did for the heat equation, we will make use of the Fourier
transform. The similarity of the Schrödinger operator to the heat operator is obvi-
ous; the only important change is the complex number i. Nevertheless, this little
variation generates a need of caution in every computation, although the result will
be analogous to (4.28). The procedure is the same, so as the Fourier transform does
not affect constants, every step repeats until (4.27). Hence,
2
Fx (E)(ξ, t) = e−ik|ξ| t ,
so we need to compute the inverse Fourier transform of a function which looks like
a Gaussian. In this case, the exponent is a complex number whose real part is zero.
4.4. THE MALGRANGE-EHRENPREIS THEOREM 61
In this moment, we must recall Example 3.25, and more precisely, (3.32). With that
result, and also with Property 1 in Proposition 3.11, we can write
r n
1 π 2
E(x, t) = n
e−|x| /4kti ,
(2π) ikt
which is the fundamental solution to the Cauchy problem for the Schrödinger oper-
ator. Finally, the solution to the original problem is given by
Z
1 2
u(x, t) = E ∗x f = √ f (y)ei|x−y| /4kt dy.
n
(2 πkti) Rn
Several different proofs have been published so far, also using completely inde-
pendent arguments. The former proofs were non-constructive, and they made use of
the Hahn-Banach theorem. Other proofs were based on the L2 -theory. Some of them
also assert that the fundamental solutions can be taken from the space of tempered
distributions. In the last decades, constructive proofs have been developed, with
explicit expressions for a solution which have been turning more and more compact.
A recent short proof using Fourier transforms is in [9], where references to other
proofs can be found.
1
Bernard Malgrange (born in 1928, Paris) was a student of Laurent Schwartz. He has been
professor at Strasbourg, Orsay and Grenoble, and he works on differential equations.
2
Leon Ehrenpreis (1930-2010, Brooklyn, NY) was professor at Temple University, Philadelphia,
PA.
Appendix A
Laurent Schwartz was a brilliant mathematician who not only focused on scientific
activities during his life. He was also very committed to French and international
social and political problems.
He was born in Paris in 1915 to a non-practicing Jewish family. His father was
the first Jew surgeon in a hospital in Paris, in a time when anti-Semitism was on the
rise in France. Mathematics were present in his childhood; his uncle was a teacher,
and the famous Jacques Hadamard was the uncle of his mother. Even if he was
good at them, he showed better abilities in literature and languages, but advised
by some of his teachers, and after a hard work, he enrolled in the École Normale
Supérieure (ENS), where he met the daughter of the also mathematician Paul Lévy,
Marie-Hélène. They married some years later.
The environment at ENS made him get close to communist, indeed Trotskyist
ideas, an inclination he would never leave behind, and which would cause him some
problems. The young Laurent finished his studies in 1937, and after two years of
military service, he moved, together with his wife, to Toulouse, where both of them
started working in a research institute. Nevertheless, their condition of Jews made
their situation extremely delicate after France fell under Nazi control in 1940 and
the pro-axis Vichy government led by Marshal Philippe Pétain imposed restrictive
laws in the area. They decided to move to the city of Clermont-Ferrand for academic
reasons, where the University of Strasbourg had been transferred during the war.
It was there where Schwartz made contact with the Bourbaki group, a collective
which had a decisive influence in his academic development. After finding himself
forced to move to the Italian border for political reasons, he stayed in Grenoble for
some months. In this time, and during the following years, he started to investigate
in several fields, such as Fréchet spaces and generalised derivatives, after reading
articles about harmonic and polyharmonic functions. One night, he came up with
the idea of generalised functions in the most beautiful night of his life.
Once the war was over, he became professor of the University of Grenoble for a
little time before moving to Nancy, city where he worked in his ideas, and published
63
64 APPENDIX A. LAURENT SCHWARTZ: LIFE AND WORK
several articles in which he presented many concepts and results concerning distri-
butions. These innovative contributions made him well-known in the whole country
and even internationally. In fact, he received a Fields medal for his work in 1950
in Harvard. In the following year he published his two-volume treatise Théorie des
distributions, where he put all his discoveries together.
In 1952, he became professor of the University of Paris. Some years later, in 1958,
his father-in-law Paul Lévy retired as professor in the prestigious, though lately old-
fashioned École Polytechnique. Even if he was not intending to make a request for
the position at first, he was persuaded by representatives of the École, in an attempt
to modernise their institution. He accepted, and from that time on, until he died, he
carried an important work in the modernisation not only of the École, but also of the
whole French university system. He rapidly revitalised the mathematical activity in
a centre which had fallen in a situation of academic paralysis.
He stayed at the École Polytechnique until 1980, but it was not his last work at all.
In 1985, he became chairman of the National Evaluation Committee, an organisation
created by the French government after his advise. He also was member of the French
Academy of Science since 1975. He died in Paris in 2002 at the age of 87.
Distributions were not his unique area of interest, and he worked in several fields
during his life. Before developing the theory which made him famous, he worked on
polynomials, semi-periodic functions, harmonic analysis and exponential sums. His
thesis in 1942 was indeed involving this last matter. Once he published his main
work Théorie des distributions, a second edition with some corrections and more
information was reprinted in 1966, and extended the field with vectorial distributions
and the kernel theorem. Afterwards, he succeeded in discovering applications of his
work to theoretical physics. He also made important contributions to the field of
probability and integration, especially in his latter works.
As stated in the beginning, Laurent Schwartz was a politically active man. He
enrolled in the Trotskyst party in France during his stay in ENS, and excepting the
lapse of the war, he was an active militant. He even presented himself for candidate
in the legislative elections in 1945 and in 1946, but he quitted in 1947. He no longer
took part in political parties, but anyway he had serious trouble when entering the
USA in 1950 to receive his Fields medal. His colleague Marshall Stone discovered
he was classified as a dangerous communist for the federal government.
Schwartz also played an active part in the decolonisation processes of the French
Indochina in 1954 and Morocco and Tunisia in 1956, and in 1960, together with
many intellectuals as Jean Paul Sartre and Simone de Beauvoir, he made a public
manifesto for the right of the French youth not to take part in the Algerian War.
This last action had a professional cost for Schwartz; indeed, he was working at
the Ministry of Defence-run École Polytechnique, from which he was fired in 1961.
Nevertheless, he was readmitted in 1963.
In the next decade, he made a strong campaign against the Vietnam War. He
even travelled to that country and met the communist leader Ho Chi Minh, being
part of the Russell Tribunal to find evidence of war crimes. In the context of the
Afghanistan War after the Soviet invasion, he was president of the International
65
[6] Karen Saxe. Beginning Functional Analysis. Springer-Verlag New York, Inc.,
2002.
67