Mathématiques pour les glomorphes à rayures

Convolution

We introduce in this chapter the notion of convolution, essentially in Rd endowed

with the Lebesgue measure. The convolution appears in a natural way in many con-
texts. Typically, and it is one of our main motivation, convolution is intimately linked
to the regular product via the Fourier transform, that we will introduce later. Another
important application, that we will devellop in the second part of this chapter, is the
regularization of functions.

Without necessarily having mentioned it, we have already met convolution and the
links it has with the usual product. Let us have a quick and non exhaustive overview
before we get to the heart of the matter.

First, let us consider two real or complex sequences paj qjPN and pbk qkPN that are
eventually vanishing (there exist N, M P N such that aj “ 0 for j ą N and bk “ 0 for
k ą M ). We note P and Q the corresponding polynomials:
N
ÿ M
ÿ
P “ aj X j and Q “ bk X k .
j“0 k“0

The usual product of these polynomials is

˜ ¸
Nÿ`M n
ÿ
PQ “ aj bn´j X n .
n“0 j“0

This polynomial is associated to the eventually vanishing sequence pcn qnPN defined by
n
ÿ
@n P N, cn “ aj bn´j . (0.1)
j“0

If we extend the sequences paj qjPN , pbk qkPN et pcn qnPN by 0 to see them as Z indexed
sequences, we get that the sequence pcn qnPZ is exactly what we will define as the con-
volution of pan qnPZ and pbn qnPZ . Thus, the usual product of polynomials is defined by
convolution.

Julien Royer
Université de Toulouse
March 5, 2025

1
 If we remove the hypothesis saying that the sequences paj qjPN and pbk qkPN are even-
tually vanishing, the polynomials
ř`8 become
ř`8 power series, but the discussion
ř`8 remains the
j k
same. the product of j“0 aj z and k“0 bk z is the power series n“0 cn where cn is
again defined by (0.1).

This observation has, for example, applications in probabilities. If X and Y are two
independant random variables with values in N, then for any n P N we have
n
ÿ n
` ˘ ÿ
P pX ` Y “ nq “ P X “ j and Y “ n ´ j “ P pX “ jqP pY “ n ´ jq.
j“0 j“0

Thus, the distribution of X ` Y is given by the convolution of the distributions of X

and Y . Thus it is relevant to associate to a random variable X : Ω Ñ N the power series
(called probability-generating function) defined by
`8
ÿ
GX pzq “ P pX “ jqz j .
j“0

We have GX`Y “ GX GY (usual power series product), and from GX`Y we can then
identify X ` Y .

In the same spirit, but closer to the issues of this course, we consider the case of
Fourier series. Let f and g be two locally integrable and 2π-perodic functions on R. We
note pαn qnPZ and pβn qnPZ their Fourier coefficients. In L2 pSq we have
ÿ ÿ
f pxq “ αn einx and gpxq “ βn einx .
nPZ nPZ

The Fourier coefficients of f g are given by the convolution of the sequences of coefficients
of f and g: ˜ ¸
ÿ ÿ
pf gqpxq “ αk βn´k einx .
nPZ kPZ
We observe the same phenomenon for the Fourier transform. More precisely, the
Fourier transform of a product of function will be the convolution of the Fourier trans-
forms of these functions. We will come back on this in the next chapter, but for this,
we will need convolution for functions on Rd . This is the object of this chapter.

When we will study the Fourier transform, we will see that its main interest, as
for Fourier series, is its nice behaviour in the framework of differential equations. This
suggests that the convolution will naturally appear when solving differential equtions.
It is indeed the case, and can already be observed for the simplest cases. Consider α P R
and f a continuous function from R` to R. Using variation of parameters, we know
that the unique solution on R` of the Cauchy problem
#
y 1 ` αy “ f ptq, t ě 0,
yp0q “ 0,

is the function żt
y : t ÞÑ e´αpt´sq f psq ds.
0

2
It is precisely the convolution of s ÞÑ e´αs , solution of
#
y 1 ` αy “ 0, t ě 0,
yp0q “ 1,

with the source term f (these two functions being extended by 0 on R˚´ ). This is not
a coincidence, and the generalization of this remark will be one of the issues in the
following chapters.

1 Convolution
Rd and C are endowed with their usual Borel algebras. On pRd , BpRd qq we consider the
Lebesgue measure.

1.1 Definitions and first properties

Let f and g be two measurable functions from Rd to C. We observe once and for all
that for any x P Rd the function y ÞÑ f px ´ yqgpyq is measurable from Rd to C. We will
no longer mention this point in the proofs of this chapter.

Definition 1.1. For x P Rd such that the function y ÞÑ f px ´ yqgpyq is integrable, we

set ż
pf ˚ gqpxq “ f px ´ yqgpyq dy.
Rd
The function pf ˚ gq is called the convolution of f and g.

Example 1.2. ‚ For any f measurable on Rd we have f ˚ 0 “ 0.

‚ If f is integrable on Rd we have
ż
p1 ˚ f qpxq “ f pyq dy.
Rd

‚ For x P R we have
$
ż1 &x
’ if x P r0, 1s,
p1r0,1s ˚ 1r0,1s qpxq “ 1r0,1s px ´ yq dy “ 2 ´ x if x P r1, 2s,
0 ’
0 if x R r0, 2s.
%

“
‚ Let f be the characteristic function of the interval ´ 21 , 12 s. Let g P L1loc pRq. For
x P R we have ż x` 1
2
pf ˚ gqpxq “ gpyq dy.
x´ 12

Thus, considering the convolution of g with f , at the point x we replace gpxq (that has
no
“ meaning since we consider almost everywhere equal functions) by the mean of g on
1 1
´ 2 , 2 s (that always has a meaning, even for g in L1loc pRq). We observe that we obtain
a continuous function pf ˚ gq. Moreover, if g is of class C k for some k P N, then pf ˚ gq
is of class C k`1 .

3
‚ Let α P R. For n P N we now consider the function gn : x ÞÑ α ` cospnxq. Thus, for
x P R we have ` ˘ ` ˘
sin npx ` 21 q ´ sin npx ´ 12 q
pf ˚ gn qpxq “ α ` .
n
The function gn is of class C 8 for any n P N, but it oscillates with higher and higher
frequency when n grows. The convolution erases these high frequency oscillations and
pf ˚ gn q is closer and closer to its mean value α.
These first examples show the regularizing effect of the convolution mentioned in the
introduction. We will come back on this aspect in the second part of this chapter. For
now, we begin by describing some general properties of the convolution.

Proposition 1.3. Let x P Rd . The convolution pf ˚ gqpxq is well defined if and only if
pg ˚ f qpxq is, and in that case their values coincide.

Proof. Applying the change of variables η “ x ´ y we get

ż ż
|f px ´ yq| |gpyq| dy “ |f pηq| |gpx ´ ηq| dη.
Rd Rd

This proves that pf ˚ gqpxq is defined if and only if pg ˚ f qpxq is defined. in that case,
the same computation considering f and g instead of |f | and |g| shows that pf ˚ gqpxq “
pg ˚ f qpxq.

Proposition 1.4. Suppose there exist R1 , R2 ą 0 such that f vanishes outside the ball
BpR1 q and g is null outside the ball BpR2 q. Then for x P Rd such that |x| ą R1 ` R2
the convolution pf ˚ gq is well defined at x and pf ˚ gqpxq “ 0.

Proof. Let x P Rd such that |x| ą R1 ` R2 . Let y P Rd . If |y| ď R2 then |x ´ y| ą R1 so

f px´yq “ 0, while if |y| ą R2 we have gpyq “ 0. In both cases we have f px´yqgpyq “ 0,
so the function y ÞÑ f px ´ yqgpyq is integrable with integral 0.

1.2 Convolution in Lp pRd q spaces

The purpose of this section is, just like the Hölder inequality for the usual product, to
give necessary conditions which ensure that pf ˚ gq is in some Lebesgue space if f and
g are. As for the usual product, there are some obvious cases.

Proposition 1.5. (i) Assume that f P L1 pRd q and g P L8 pRd q. Then pf ˚ gqpxq is
well defined for any x P Rd and pf ˚ gq P L8 pRd q with

}f ˚ g}8 ď }f }1 }g}8 .

(ii) We suppose that f P L1loc pRd q and g P L8 pRd q vanishes outside a compact of Rd .
Then pf ˚ gqpxq is well defined for any x P Rd .

Of course, we have analoguous properties if we switch the roles of f and g.

Proof. For this first proof, we emphasize the distinction between L1 pRd q and L1 pRd q or
L8 pRd q and L8 pRd q. We will not do this in the rest of the chapter.

4
 So let f P L1 pRd q and g P L8 pRd q. Let x P Rd . For almost every y P Rd we have
|gpyq| ď }g}8 , so
ż ż
|f px ´ yq| |gpyq| dy ď }g}8 |f px ´ yq| dy.
yPRd yPRd

Using the affine change of variables η “ x ´ y, dη “ dy, we get

ż ż
|f px ´ yq| |gpyq| dy ď }g}8 |f pηq| dη “ }f }1 }g}8 .
yPRd ηPRd

This proves that pf ˚ gqpxq is well defined and

ż
|pf ˚ gqpxq| ď |f px ´ yq| |gpyq| dy ď }f }1 }g}8 .
yPRd

Hence }pf ˚ gq}8 ď }f }1 }g}8 .

We now consider f P L1 pRd q and g P L8 pRd q. Let f1 , f2 P L1 pRd q be two elements of
the equivalence class of f , and g1 , g2 P L8 pRd q be two elements of the equivalence class
of g. Let x P Rd . The functions y ÞÑ f1 px´yqg1 pyq and y ÞÑ f2 px´yqg2 pyq are integrable
on Rd . Since f1 px ´ yqg1 pyq coincide for almost every y P Rd with f2 px ´ yqg2 pyq, their
integrals also coincide, so pf1 ˚g1 qpxq “ pf2 ˚g2 qpxq. We denote by pf ˚gqpxq this common
value. Moreover we have

}f ˚ g}8 “ }f1 ˚ g1 }8 ď }f1 }1 }g1 }8 “ }f }1 }g}8 .

This gives the first property.

For the second, we do not give so many details, we just identify f and g with two of
their equivalence class representatives in L1loc pRd q and L8 pRd q, with g vanishing almost
everywhere outside Bp0, Rq for some R ą 0. Let x P Rd . As above, we have
ż ż
|f px ´ yq| |gpyq| dy “ |f px ´ yq| |gpyq| dy
yPRd yPBp0,Rq
ż
ď }g}8 |f px ´ yq| dy
yPBp0,Rq
ż
ď }g}8 |f | dη.
ηPBpx,Rq

Since f is integrable on Bpx, Rq, this proves that pf ˚ gqpxq is well defined.

Now we turn to the first important result about convolution. We prove that the
convolution of two integrable functions is an integrable function, and furthermore the
norm of the convolution is not greater than the product of the norms of the two factors.
From this perspective, convolution is more convenient than the usual product.

Proposition 1.6. Let f, g P L1 pRd q. The function y ÞÑ f px ´ yqgpyq is integrable (and

thus pf ˚ gqpxq is well defined) for almost every x P Rd . This defines a function pf ˚ gq
which is integrable on Rd , and

}f ˚ g}1 ď }f }1 }g}1 .

5
Proof. We identify f and g to representatives of their equivalence class in L1 pRd q (and
we check that everything is independant of the choice of the representative). The map
px, yq ÞÑ |f px ´ yq| |gpyq| is measurable from Rd ˆ Rd to r0, `8s. By the Fubini-Tonelli
theorem, the function ż
x ÞÑ |f px ´ yq| |gpyq| dy
Rd
is well defined and measurable from Rd to r0, `8s. Moreover,
ż ˆż ˙ ż ˆż ˙
|f px ´ yq| |gpyq| dy dx “ |gpyq| |f px ´ yq| dx dy
Rd Rd Rd Rd
ż ˆż ˙
“ |gpyq| |f pηq| dη dy
Rd Rd (1.1)
ż
“ }f }1 |gpyq| dy
Rd
“ }f }1 }g}1 .
For the third equality we used the change of variables η “ x ´ y, dη “ dx, in the
x-variable integral. This proves that for almost every x P Rd we have
ż
|f px ´ yq| |gpyq| dy ă `8.
Rd

Thus, by the Fubini-Lebesgue theorem, pf ˚gqpxq is well defined for almost every x P Rd .
Moreover,
ż ż ˆż ˙
|pf ˚ gqpxq| dx ď |f px ´ yq| |gpyq| dy dx “ }f }1 }g}1 .
Rd Rd Rd
The proposition is proved.
The aim of this section is to give more results of this kind, that is, prove that if
f and g are in suitable Lebesgue spaces, their convolution is well defined and satisfies
some useful properties.
The following result generalizes the first property of Proposition 1.5 and Proposition
1.6. If a function is in Lp pRd q for some p, then so is its convolution with an integrable
function.
Proposition 1.7. Let p P r1, `8s, f P L1 pRd q and g P Lp pRd q. Then the convolution
pf ˚ gqpxq is well defined for almost every x P Rd , we have pf ˚ gq P Lp pRd q and
}f ˚ g}p ď }f }1 }g}p .

Using the commutativity of convolution, we have an analoguous result for f P Lp pRd q

and g P L1 pRd q.
Proof. We identify f et g to representatives in L1 pRd q and Lp pRd q. The result is already
p
known if p “ `8, so we can assume that p P r1, `8r. Let q “ p´1 be the conjugate
d
exponent of p. For x P R we have by the Hölder inequality
ż ż
1 1
|f px ´ yq| |gpyq| dy “ |f px ´ yq| q |f px ´ yq| p |gpyq| dy
Rd Rd
1
ˆż ˙1
p
q p
ď }f }1 |f px ´ yq| |gpyq| dy .
Rd

6
By the Fubini-Tonelli Theorem we get
ż ˇż ˇp
p
ˇ ˇ
}|f | ˚ |g|}p “ ˇ
ˇ |f px ´ yq| |gpyq| dy ˇˇ dx
Rd Rd
p
ż ˆż ˙
ď }f }1q |f px ´ yq| |gpyq|p dy dx
d Rd
p
żR ˆż ˙
q p
ď }f }1 |gpyq| |f px ´ yq| dx dy
Rd Rd
p
ď }f }1 }g}pp .
We deduce that p|f | ˚ |g|q P Lp pRd q with }|f | ˚ |g|}p ď }f }1 }g}p . In particular the
function y ÞÑ f px ´ yqgpyq is integrable for almost every x P Rd , and
ż ˇż ˇp ż ˇż ˇp
|f px ´ yq| |gpyq| dy ˇˇ dx ď }f }p1 }g}pp .
ˇ ˇ ˇ ˇ
ˇ
ˇ f px ´ yqgpyq dy ˇ dx ď
ˇ ˇ
ˇ
Rd Rd d d R R

Hence pf ˚ gq P Lp pRd q and }f ˚ g}p ď }f }1 }g}p .

Remark 1.8. For p P r1, `8s et f P L1 pRd q, we have proved that the map F : g ÞÑ f ˚ g
is a continuous linear map on Lp pRd q, with }F }LpLp pXqq ď }f }1 .
Let us discuss another generalization of Proposition 1.5. We still arrive in L8 pRd q,
but we now consider f and g in Lp pRd q and Lq pRd q for general conjugate exponents
p and q. It is interesting to compare this result with Hölder inequality for the usual
product of functions.
Proposition 1.9. Let p, q P r1, `8s be two conjugate exponents. Let f P Lp pRd q and
g P Lq pRd q. Then the convolution pf ˚ gqpxq is well defined for any x P Rd , pf ˚ gq is
bounded on Rd and
}f ˚ g}L8 ď }f }Lp }g}Lq .
Moreover, pf ˚ gq is uniformly continuous and, if p and q are finite, then pf ˚ gqpxq tends
to 0 when |x| tends to `8.
Proof. We identify f and g to representatives of their class in Lp pRd q and Lq pRd q.
‚ Let x P Rd . By the Hölder Inequality we have
ż ˆż ˙ 1 ˆż ˙1
p q
p q
|f px ´ yq| |gpyq| dy ď |f px ´ yq| dy |gpyq| dy “ }f }p }g}q .
Rd Rd Rd
Hence pf ˚ gqpxq is well defined and
|pf ˚ gqpxq| ď }f }p }g}q .
This proves that pf ˚ gq is bounded with
}f ˚ g}8 ď }f }p }g}q .
‚ Let h P Rd . We suppose that p ă `8. For x P Rd we have, still using the Hölder
Inequality,
ż
|pf ˚ gqpx ` hq ´ pf ˚ gqpxq| ď |f px ` h ´ yq ´ f px ´ yq| |gpyq| dy
Rd
ˆż ˙1
p
p
ď }g}q |f px ` h ´ yq ´ f px ´ yq| dy
Rd
ď }τ´h f ´ f }p }g}q

7
(recall that τ´h f is the function η ÞÑ f pη ` hq). By continuity of the translation in
Lp pRd q,
sup |pf ˚ gqpx ` hq ´ pf ˚ gqpxq| ď }τh f ´ f }p }g}q ÝÝÝÑ 0,
xPRd hÑ0

so pf ˚ gq is uniformly continuous. If p “ `8 then q “ 1 and we can similarly write, by

commutativity of convolution,
ż
|pf ˚ gqpx ` hq ´ pf ˚ gqpxq| ď |f pyq| |gpx ` h ´ yq ´ gpx ´ yq| dy
Rd
ď }f }p }τ´h g ´ g}q
ÝÝÝÑ 0.
hÑ0

‚ Now assume that p and q are finite. Then, there exist sequences pfn qnPN and pgn qnPN
of compactly supported and continuous functions that tends to f and g in Lp pRd q and
in Lq pRd q, respectively. For any n P N, the function pfn ˚ gn q has compact support (see
proposition 1.4). By linearity of the convolution with respect to each factor, we can
write
pf ˚ gq ´ pfn ˚ gn q “ pf ´ fn q ˚ g ` fn ˚ pg ´ gn q,
hence
}pf ˚ gq ´ pfn ˚ gn q}8 ď }f ´ fn }p }g}q ` }f }p }g ´ gn }q ÝÝÝÑ 0.
nÑ8

Thus pf ˚ gq is the uniform limit of a sequence of compactly supported functions. This

implies that it goes to 0 when x goes to infinity.

We can give a completely general result about the existence of convolution in Lebesgue
spaces.

Proposition 1.10. Let p, q, r P r1, 8s such that

1 1 1
` “1` .
p q r

Let f P Lp pRd q and g P Lq pRd q. Then pf ˚ gqpxq is well defined for almost every x P Rd ,
pf ˚ gq belongs to Lr pRd q and

}f ˚ g}r ď }f }p }g}q .

Proof. This result is a consequence of Proposition 1.7 if q “ 1 and r “ p (where p “ 1

and r “ q) and of Proposition 1.9 if r “ 8. Thus we can assume that r ă 8 and p, q P
s1, `8r. Notice that maxpp, qq ă r. As above, we identify f and g to representatives in
Lp pRd q and Lq pRd q. The function p|f | ˚ |g|q is measurable from Rd to r0, `8s and for
x P Rd we can write
ż
p q ˘ p q
|f px ´ yq| r |gpyq| r |f px ´ yq|1´ r |gpyq|1´ r dy.
`
p|f | ˚ |g|qpxq “
Rd
rp rq
By the generalized Hölder Inequality applied with exponents r, r´p and r´q we get

ˆż ˙ 1 ˆż ˙ r´p ˆż ˙ r´q
r rp rq
p q p q
p|f |˚|g|qpxq ď |f px ´ yq| |gpyq| dy |f px ´ yq| dy |gpyq| dy .
Rd Rd Rd

8
We get ż
|p|f | ˚ |g|qpxq|r ď }f }r´p
p }g}qr´q |f px ´ yq|p |gpyq|q dy.
Rd

Integrating with respect to x P Rd we obtain by the Fubini Theorem that

ż ˆż ˙
r r´p r´q p q
}|f | ˚ |g|}r ď }f }p }g}q |f px ´ yq| |gpyq| dy dx
d Rd
żR ˆż ˙
r´p r´q q p
ď }f }p }g}q |gpyq| |f px ´ yq| dx dy
Rd Rd
ż
ď }f }rp }g}r´q
q |gpyq|q dy
Rd
ď }f }rp }g}rq .

Hence |f | ˚ |g| P Lr pRd q with }|f | ˚ |g|}r ď }f }p }g}q . We conclude as before.

1.3 Other convolutions

The aim of this chapter is to study convolution on Rd , but since it was discussed in the
introduction, we briefly mention the standard convolution products in other contexts.
Let u and v be two sequences in CZ . When it makes sense, we define the sequence
pu ˚ vq by ÿ
pu ˚ vqn “ un´k vk . (1.2)
kPZ

We can also consider convolution on the space of 2π-periodic functions from R to C. If

f and g are two 2π-periodic functions we set

1 π
ż
pf ˚ gqpxq “ f pyqgpx ´ yq dy.
2π ´π

We cannot define this convolution on any measured space. We see here that we
need to have an additive group (to give sense to f px ´ yq). Thus we cannot define a
convolution on N or R` . However, we can extend by zeros a sequence indexed by N to
a sequence indexed by Z. Thus, if u and v are sequences in CN that we extend in Z
indexed sequences setting u´m “ v´m “ 0 for m P N, then (1.2) gives
#
0 si n ă 0,
pu ˚ vqn “ řn
k“0 un´k vk if n ě 0.

We obtain again a sequence whose terms with negative indices vanish, and we can see
pu ˚ vq as a sequence indexed by N.

Likewise, if f and g are two functions on R` , we can extend them by 0 on R´ (we

still note f and g the extended functions), and the convolution of f and g is given by
#
0 si x ă 0,
pf ˚ gqpxq “ şx
0 f px ´ yqgpyq dy if x ě 0.

As for sequences, we can consider pf ˚ gq as a function on R` .

9
2 Regularization of functions
Since we introduced them, we have worked a lot with functions in Lebesgue spaces, and
these functions can be very singular. We have already seen that the set of compact
support continuous functions is dense in Lp pRd q for p P r1, `8r. The aim of this section
is to go further and to approximate integrable functions by smooth (and compactly
supported) functions.

2.1 Approximation to the identity

We have seen that convolution defines a product on L1 pRd q. A natural question is wether
this product has an identity element 1, such that for any f P L1 pRd q

1 ˚ f “ f ˚ 1 “ f.

A formal computation shows that the “Dirac function” (supported on t0u and of integral
1) could play this role, but we know that such a function cannot exist.
Notice that we could introduce the convolution of a function and a measure. When
it has a meaning, we could set for a function f and a measure µ
ż
pµ ˚ f qpxq “ pf ˚ µqpxq “ f px ´ yq dµpyq.
Rd

With the Dirac measure, we would get pδ ˚ f qpxq “ f pxq. This will be done in the more
general cases of distributions.

For now, we only work with (equivalence classes of) functions. So there is no identity
element for the convolution, but we can consider arbitrarily close functions, in the sense
that their integral is equal to 1 and that most of their mass is close to 0. More precisely,
we introduce the following notion.

Definition 2.1. We call approximation to the identity a sequence pρn qnPN of measurable
functions on Rd such that

(i) ρn has positive values 1 for any n P N,

ş
(ii) Rd ρn dλ “ 1 for any n P N,

(iii) for any ε ą 0 we have ż

ρn dλ ÝÝÝÝÑ 0.
|x|ěε nÑ`8

We could consider approximations to the identity parametrized by a real number

instead of natural numbers.

In general we use the following result to build the sequences of approximation of the
unit.
1
we could also consider approximation of the unit of variable signs, but we need to ensure that the
sequence pρn qnPN is bounded in L1 . In practice, we use positive unit approximation.

10
 ş
Proposition 2.2. Let ρ : Rd Ñ R` be a measurable function such that Rd ρ dλ “ 1.
For n P N˚ et x P Rd we set
ρn pxq “ nd ρpnxq.
Then the sequence pρn qnPN˚ is an approximation of the unit on Rd .
For example, we can consider the mollifiers defined by
|x|2
e´ 2
ρpxq “ d , (2.1)
p2πq 2
or, on R,
1
ρpxq “ . (2.2)
πp1 ` x2 q
The proof of Proposition 2.2 simply uses the affine change of variables y “ nx.

The aim of an approximation of the unit is to . . . approach the unit. In particular,

we expect that pρn ˚ f q is close to f for n large. We begin with the most convenient
case.
Proposition 2.3. Let pρn qnPN be an approximation to the unit. Let f be a bounded and
continuous function on Rd . Then f ˚ ρn converges uniformly to f as n goes to `8.
ş
Proof. Since Rd ρn dλ “ 1 we have for all x P Rd
ż
pf ˚ ρn qpxq ´ f pxq “ pf px ´ yq ´ f pxqq ρn pyq dy.
Rd

Let ε ą 0. There exists δ ą 0 such that for x, y P Rd with |x ´ y| ď δ we have

|f pxq ´ f pyq| ď 2ε . Let M ą 0 be a bound for |f |. There exists N P N such that for all
n ě N we have ż
ε
ρn pyq dy ď .
|y|ěη 4M
Then for x P Rd and n ě N we have

|pf ˚ ρn qpxq ´ f pxq|

ż ż
ď |f px ´ yq ´ f pxq| ρn pyq dy ` |f px ´ yq ´ f pxq| ρn pyq dy
yPBpx,δq yPRd zBpx,δq
ż ż
ε
ď ρn pyq dy ` 2M ρn pyq dy
2 yPBpx,δq yPRd zBpx,δq
ε ε
ď ` “ ε.
2 2
The result follows.

Now we show that ρn ˚ f is close to f for any f P L1 pRd q, and even for f P Lp pRd q
for p P r1, `8r. Taking for example ρn “ 2n1 1 1 ‰ for n P N˚ and f “ 1r´1,1s we see
r´ n , n
that the following result cannot hold for p “ `8.
Proposition 2.4. Let pρn qnPN be an approximation of the unit. Let p P r1, `8r and
f P Lp pRd q. We have
}pf ˚ ρn q ´ f }p ÝÝÝÑ 0.
nÑ8

11
 ş
Proof. We identify f a representative in Lp pRd q. Let n P N˚ . Since Rd ρn dλ “ 1 we
have for any x P Rd
ż
pf ˚ ρn qpxq ´ f pxq “ pf px ´ yq ´ f pxqq ρn pyq dy
Rd
ż
1 1
“ pf px ´ yq ´ f pxqq ρn pyq p ρn pyq q dy,
Rd

where q the conjugate exponent of p. By the Hölder Inequality we get

ˆż ˙ ˆż ˙p
q
p p
|pf ˚ ρn qpxq ´ f pxq| ď |f px ´ yq ´ f pxq| ρn pyq dy ρn pyq dy .
Rd Rd

The last integral is equal to 1. Integrating with respect to x P Rd an using the Fubini-
Tonelli Theorem we get
ż ˆż ˙ ż
p p
}f ˚ ρn ´ f }p ď |f px ´ yq ´ f pxq| dx ρn pyq dy ď }τy f ´ f }pp ρn pyq dy.
Rd Rd Rd

Let ε ą 0. By continuity of the translation there exists η ą 0 such that for any y P Rd
with |y| ď η we have
ε
}τy f ´ f }pp ď .
2
On the other hand there exists N P N such that for any n ě N we have
ż
p p ε
2 }f }p ρn pyq dy ď .
|y|ěη 2

Then for n ě N we have

ż ż
p
}f ˚ ρn ´ f }p ď }τy f ´ f }pp ρn pyq dy ` }τy f ´ f }pp ρn pyq dy ď ε.
|y|ďη |y|ěη

This proves that

}f ˚ ρn ´ f }pp ÝÝÝÝÑ 0.
nÑ`8

At the beginning of the chapter, we observed that convolution has a regularizing

effect. So we can expect ρn ˚ f to be more regular than f . All the more so if ρn is itself
regular. Hence, Proposition 2.4 will allow us to approach Lp by a sequence of regular
functions. To do that, we will use a sequence of approximation of the unit made of
regular functions for example by building it from regular kernels (2.1) or (2.2). If we
want to approach f by regular functions that have compact support, we will need to use
compact support regularizing functions. This is not the case for (2.1) or (2.2).

2.2 Localized and regular functions

Let Ω be an open set of Rd . We begin by giving basic properties of the space of smooth
and compactly supported functions on Ω. In particular, the fact that this set is not
reduced to the zero function.

Definition 2.5. We denote by C08 pΩq, Cc8 pΩq or DpΩq the set of smooth and compactly
supported functions on Ω.

12
 Before going any further, we check that C08 pRd q contains non trivial functions.

Proposition 2.6. The function f defined on Rd by

# ´ ¯
1
exp ´ 1´|x| 2 if |x| ă 1,
f pxq “
0 if |x| ě 1,

belongs in C08 pRd q. Its support is the closed ball Bp0, 1q.

Proof. ‚ The function f is positive on the unit ball Bp0, 1q and vanishes on RzBp0, 1q,
so its support is Bp0, 1q. Moreover, f is of class C 8 on Bp0, 1q and RzBp0, 1q.
‚ For r ą 0 we set # ´ ¯
1
exp ´ 1´r if r ă 1,
gprq “
0 if r ě 1.
g is of class C 8 on s0, 1r. We can check by induction on n P N that there exists a
polynomial Pn P RrXs such that for any r Ps0, 1r we have
ˆ ˙ ˆ ˙
pnq 1 1
g prq “ Pn exp ´ .
1´r 1´r

Let n P N. The function s ÞÑ Pn psqe´s tends to 0 when s tends to `8, so

g pnq prq ÝÝÝÑ 0.

rÑ1

In particular, g is continuous at x “ 1. In addition, the derivatives of g vanish on

s1, `8r. By the theorem of C 1 extension (based on the mean value theorem) we get by
induction that g is n times differentiable at x “ 1 with g pnq p1q “ 0. Finaly g is of class
C 8 on s0, `8r.
‚ By composition of g with the smooth function x ÞÑ |x|2 “ x21 ` ¨ ¨ ¨ ` x2d , we get that
f is of class C 8 on Rd z t0u. Since it is also C 8 on Bp0, 1q, it is of class C 8 on Rd .

From this first non trivial example of C08 pRd q function, we can build “peak func-
tions”, that we will use as approximations of the unit.

Proposition
ż 2.7. Let ε ą 0. There exists ρε P C08 pRd , R` q such that supppρε q Ă Bp0, εq
and ρε “ 1.
Rd

Proof. Let f be the function given by proposition 2.6. It is continuous and has compact
support thus it is integrable on Rd . Moreover, it is non-negative and non identically
zero, so its integral is positive. Thus for x P Rd we can set

f pxq
ρpxq “ ş .
Rd f ptq dt

This defines a smooth function ρ : Rd Ñ R` whose support is Bp0, 1q and whose integral
is equal to 1. We conclude by setting, for any x P R,
1 ´x¯
ρε pxq “ d ρ .
ε ε

13
2.3 Regularization of functions
In this paragraph we prove that the convolution of a general function with a C08 pRd q
function is well defined and regular. We will then deduce that C08 pRd q is dense in Lp pRd q
for any p P r1, `8r.

We first recall some notation for derivatives in any dimension. For α “ pα1 , . . . , αn q P
Nd we set
|α| “ α1 ` ¨ ¨ ¨ ` αd ,
and then, for f P C 8 pRd q,
ˆ ˙α1 ˆ ˙ αd
α B |α| f B B
B f “ α1 “ ¨¨¨ f.
Bx1 . . . Bxαd d Bx1 Bxd
Recall that by the Schwarz Theorem, the order of differentiation is not important.

Proposition 2.8. Let ρ P C08 pRd q. For f P L1loc pRd q we have pρ ˚ f q P C 8 pRd q and for
any α P Nd we have
Bα
ˆ α ˙
B ρ
α
pρ ˚ f q “ ˚ f. (2.3)
Bx Bxα
Proof. The convolution is well defined by Proposition 1.5. Let j P J1, dK. There exists
R ą 0 and M ą 0 such that for any x P Rd
|Bj ρpxq| ď M 1Bp0,Rq pxq.

Let r ą 0. For any y P Rd the map x ÞÑ ρpx ´ yqf pyq is of class C 1 on Bp0, rq and
ˇ ˇ ˇ ˇ
ˇ B ` ˘ˇ ˇ B ˇ
ˇ
ˇ Bxj ρpx ´ yqf pyq ˇ “ ˇ
ˇ ˇ Bxj ρpx ´ yqˇ |f pyq| ď M 1Bp0,R`rq pyq |f pyq| .
ˇ

By differentiation under the integral sign, the function pρ ˚ f q is differentiable with

respect to xj on Bp0, rq and its derivative is
B
pρ ˚ f q “ pBj ρq ˚ f.
Bxj
Since this holds for any r ą 0, the proposition is proved for |α| “ 1. We conclude by
induction, replacing ρ by its successive derivatives.

Theorem 2.9. Let p P r1, `8r. Then C08 pRd q is dense in Lp pRd q.
Proof. Let f P Lp pRd q and ε ą 0. There exists R ą 0 such that g “ 1BpRq f satisfies
}f ´ g}p ď 2ε . Let pρn qnPN be a sequence of approximation to the identity such that
ρn P C08 pRd q for all n P N. By Proposition 2.8, pρn ˚ gq is of class C 8 for all n P N˚ .
It also has compact support by Proposition 1.4. Finally, by Proposition 2.4, we have
}g ´ pg ˚ ρn q}p ď 2ε for n large enough, so }f ´ pg ˚ ρn q}p ď ε.

Remark 2.10. Let p, q P r1, `8r. Let f P Lp pRd q X Lq pRd q. Let ε ą 0. In the previous
proof we can chose R such that }f ´ g}p ď 2ε and }f ´ g}q ď 2ε . Since pg ˚ ρn q tends to
g in Lp pRd q and Lq pRd q, there exists n P N such that }g ´ gn }p ď 2ε and }g ´ gn }q ď 2ε .
This proves that one can construct a sequence pfm qmPN of functions in C08 pRd q such
that fm tends to f both in Lp pRd q and in Lq pRd q.

14
Theorem 2.11. Let Ω be an open subset of Rd and p P r1, `8r. Then C08 pRd q is dense
in Lp pRd q.

Proof. ‚ Let f P Lp pΩq. Assume that there exists a compact K of Ω such that f “ 0
outside K. We extend f by 0 on Rd . This defines a function f˜ P Lp pRd q. There exists
δ ą 0 such thatş Bpx, 2δq Ă Ω for all x P K. Let ρ P C08 pRd , R` q be supported in Bp0, δq
and such that ρ dx “ 1. For n P N˚ and x P Rd we set ρn pxq “ n´d ρpnxq. Then we set
f˜n “ f˜ ˚ ρn . Then we have
}f˜n ´ f˜}Lp pRd q ÝÝÝÑ 0.
nÑ8

Moreover, f˜n is compactly supported in Ω, so its restriction fn to Ω belongs to C08 pΩq

and satisfies
}fn ´ f }Lp pΩq ÝÝÝÑ 0.
nÑ8

This proves the result in this case.

‚ For m P N˚ we set
" *
1
Km “ x P Ω | |x| ď n and Bpx, q Ă Ω .
m

Then Km is compact for all m P N˚ and

ď
Ω“ Km .
mPN˚

Let f P Lp pΩq and ε ą 0. By the dominated convergence theorem we have

}f ´ f 1Kn }Lp pΩq ÝÝÝÑ 0,

nÑ8

so there exists n P N such that

ε
}f ´ f 1Kn }Lp pΩq ď .
2
By the first step there exists g P C08 pΩq such that
ε
}f 1Kn ´ g}Lp pΩq ď .
2
Then }f ´ g}Lp pΩq ď ε, and the conclusion follows.

Exercise 1. Let Ω be an open subset of Rd . For m P N˚ we set

" *
d 1
Km “ x P Ω | |x| ď m and distpx, R zΩq ě .
m

Let f P L1 pΩq and ε ą 0. The purpose of the exercise is to construct g P C08 pΩq such
that }f ´ g}L1 pΩq ď ε.
1. We first consider a particular case. Assume that there exists a compact K Ă Ω such
that f “ 0 outside K. For x P Rd we set
#
f pxq if x P Ω,
f˜pxq “
0 otherwise.

15
 ş
Let ρ P C08 pRd , R` q supported in Bp0, 1q and such that Rd ρ dx “ 1. For n P N˚ and
x P Rd we set ρn pxq d
› “ n ρpnxq.›
a. Prove that ›pf˜ ˚ ρn q ´ f˜› 1 d Ñ 0 as n Ñ 8.
› ›
L pR q
b. Prove that f˜ ˚ ρn is supported in Ω for n large enough.
c. Conclude in this case.
2. Prove the general case.

2.4 Partitions of unity

We end this chapter with partitions of unity. The purpose is to be able to localize a
problem, considering a function f defined on an open set Ω as a sum of functions that
are localized on smaller and simpler open subsets. We could do it by multiplying f by
characteristic functions of a partition of Ω, but we often need to localize with regular
functions.
Proposition 2.12. Let K and O be compact and open subsets of Rd such that K Ă O.
Then there exists χ P C08 pRd , r0, 1sq such that χ “ 1 in a neighbourhood of K and χ “ 0
in a neighbourhood of Rd zO.
Proof. There exists ε ą 0 such that for any x P K we have Bpx, 4εq Ă O. We consider
a peak function ρε as given by proposition 2.7 and we set
ď
Kε “ Bpx, 2εq,
xPK

so that K Ă K̊ε Ă Kε Ă O. It only remains to regularize the characteristic function of

Kε . For this we set, for x P R,
ż
χpxq “ 1Kε px ´ yqρε pyq dy.
R

We check that χ verifies the required properties.

Proposition 2.13. Let K be a compact set of Rd . Let n P N˚ and ω1 , . . . , ωn be open

sets of Rd such that
n
ď
KĂ ωj .
j“1

Then there exist

řnfunctions χ1 , . . . , χn P C08 pRd , r0, 1sq such that supppχj q Ă ωj for any
j P J1, N K and j“1 χj is equal to 1 on K.

Proof. ‚ Let x P K. There exists jpxq P J1, nK and rpxq ą 0 such that Bpx, rpxqq Ă
ωjpxq . The family of balls Bpx, rpxqq for x P K is an open cover of K, so there exist
k P N˚ and x1 , . . . , xk P K such that
k
ď
KĂO“ Bpxl , rpxl qq.
l“1

For j P J1, nK we set ď

Kj “ Bpxl , rpxl qq.
1ďlďk
jpxl q“j

16
Then Kj is a compact included in ωj . In addition we have
n
ď
KĂ Kj .
j“1

‚ We consider ϕ0 P C 8 pRd , r0, 1sq equal to 1 outside O and equal to 0 on K. For

instance, we can set ϕ0 “ 1 ´ χ, where χ P C08 pRd , r0, 1sq is given by Proposition
2.12. For j P J1, nK we also consider ϕj P C08 pRd , r0, 1sq such that ϕj “ 1 on Kj and
supppϕj q Ă ωj . For any x P Rd we have
n
ÿ
ϕi ą 0.
i“0

For j P J1, nK we set

ϕj
χj “ řn .
i“0 ϕi
The χj , 0 ď j ď n, are of class C 8 and they take values in r0, 1s, their sum is 1
everywhere
ř and χj has compact support for j ‰ 0. Moreover, since χ0 vanishes on K,
we have nj“1 χj “ 1 on K.

17