NOTES FOR THE COURSE ANALISI DI FOURIER

MARCO M. PELOSO

Contents
1. Fourier series 1
1.1. The torus 1
1.2. Norm convergence of Fourier series 5
2. The Fourier transform 7
2.1. Convolutions 7
2.2. The Fourier transform 11
2.3. The Schwartz space 15
2.4. The space of tempered distributions 19
3. The Marcinkiewicz interpolation theorem and the Calderón–Zygmund decomposition 28
3.1. The Marcinkiewicz interpolation theorem 28
3.2. The Calderón–Zygmund decomposition of an L1 -function 31
4. The Hilbert transform 33
p
4.1. The L -boundedness of the Hilbert transform 36
4.2. Further properties of the Hilbert transform 39
5. Singular integrals 41
5.1. The Calderón–Zygmund theorem 41
5.2. Homogeneous distributions 43
5.3. The Riesz transforms 46
5.4. Solution of the Laplace equation 47
5.5. The heat operator 51
5.6. The Riesz potentials 52
5.7. †Construction of Calderón–Zygmund kernels 56
5.8. †Vector-valued singular integrals 62
6. Fourier multipliers 67
p
6.1. The space Mp of L -bounded Fourier multipliers 67
6.2. Multiple Fourier series and convergence in the Lp -norm 71
6.3. The Sobolev spaces H s 74
6.4. †The Mihlin–Hörmander multiplier theorem 78
6.5. †Proof of Thm. 6.24 80
7. Partial Differential Operators 82
7.1. The Sobolev spaces H s 82
7.2. Partial differential operators 85
8. †Littlewood–Paley theory 90
8.1. An application 90
8.2. The Littlewood–Paley theorem 93
8.3. The Marcinkiewicz multiplier theorem 96
9. Symbols and pseudodifferential operators 98
9.1. The kernel of a pseudodifferential operator 101
9.2. Sobolev continuity of a pseudodifferential operator 104

Appunti per il corso Analisi di Fourier per i Corsi di Laurea in Matematica dell’Università di Milano, a.a.
2017/18.
A † indicates the parts that are not required for the exam.
– June 25, 2018.
9.3. Asymptotic expansion of symbols 105
9.4. Adjoints and products of pseudodifferential operators 108
9.5. Local regularity of elliptic operators 112
10. Sum of squares of vector fields 115
Appendix A. Some results on Lp spaces 121
Appendix B. Some exercises 125
References 127
 HARMONIC ANALYSIS 1

1. Fourier series
In this brief first section we recall the main facts about classical Fourier series on the torus,
in part as a historical introduction, in part to motivate some of the results and techniques that
we will devolop later in the notes.
1.1. The torus. We will denote the additive group of real numbers by R and by Z the subgroup
consisting of the integers.
We define the torus T to be the group which is the quotient T = R/2πZ. There is an obvious
identification between functions on T and 2π-periodic functions on R.
This also allows an implicit introduction of notions such as continuity, differentiability, etc.
for functions on T. The Lebesgue measure on T, also, can be defined by means of the above
identification: a function f is integrable on T if the corresponding 2π-periodic function, which
we denote again by f , is integrable on [0, 2π) and we set
Z Z 2π
f (t) dt = f (x) dx .
T 0
A fundamental property of the measure dt on T is that it is translation invariant, that is, for
all t0 ∈ R Z Z
f (t − t0 ) dt = f (t) dt ,
T T
as it is easily checked, using the periodicity of f in R. (Here, of course, we have defined the
function t 7→ f (t − t0 ) via the translated of the periodic function x 7→ f (x) on the real line.)
We recall that the unit circle ∂D = {ζ ∈ C : ζ = eiθ } inherits the multiplicative structure
from C and it is homeomorphic (as a topological group) to the torus T via the mapping t 7→ eit ,
thus defining on T the structure of a topological group.
With an abuse of notation, we are going to use the same letter to denote functions on ∂D
and T, that is, we write
f (t) = f (eit )
for t ∈ R and f 2π-periodic.
We denote by Lp (T), 1 ≤ p < ∞ the Lebesgue space of measure (equivalence classes of)
functions such that  1 Z 1/p
kf kLp (T) := |f (t)|p dt
2π T
∞
is finite, and by L (T) the space of essentially bounded functions on T. Notice that, since the
total mass of T is finite, for 1 ≤ p ≤ q ≤ ∞ we have the (continuous) embedding
Lq (T) ⊂ Lp (T) .
We have the following fundamental definition of Fourier coefficient.
Definition 1.1. Let f ∈ L1 (T) (or, as an integrable function on the unit circle ∂D). We define
the n-th Fourier coefficient of the restriction of f
Z 2π
1
ˆ
f (n) = f (eiθ )e−inθ dθ .
2π 0

We turn to some elementary properties of Fourier coefficients.

2 M. M. PELOSO

Proposition 1.2. Let f, g ∈ L1 (T), then

(i) (f + g) b (n) = fˆ(n) + ĝ(n).
(ii) For any complex number α, (αf ) b (n) = αfˆ(n).
(iii) If f¯ is the complex conjugate of f , then (f¯) b (n) = fˆ(−n).
(iv) Writing fτ (t) = f (t − τ ), for t, τ ∈ T, we have (fτ ) b (n) = fˆ(n)e−inτ .
(v) |fˆ(n)| ≤ kf kL1 (T) .
Proof. The proof is elementary and left as an exercise. 2
Given f, g ∈ L1 (T), the convolution on T is defined as
Z 2π
1
f ∗ g(t) = f (t − τ )g(τ ) dτ . (1.1)
2π 0
Proposition 1.3. For f, g ∈ L1 (T), f ∗ g ∈ L1 (T) and
(i) kf ∗ gkL1 ≤ kf kL1 kgkL1 ;
(ii) (f ∗ g) b (n) = fˆ(n)ĝ(n).
Proof. Exercise. 2

Definition 1.4. A function of the form

N
X
p(t) = ak eikt ,
k=−N

for a positive integer N , is called a trigonometric polynomial.

It is immediate to observe that a trigonometric polynomial p is smooth, hence continuous and
in L1 and that, if p is as above, p̂(k) = ak , for all k (observing that ak = 0 for |k| > N ).
Therefore, if f ∈ L1 and p is as above,
(f ∗ p) b (k) = ak fˆ(k) . (1.2)

Definition 1.5. A family of 2π-periodic functions {ΦN }N =1,2,... is called a summability kernel
if they satisfy the following properties:
Z 2π
1
(1) ΦN (eiθ ) dθ = 1 for N = 1, 2, . . . ;
2π 0
(2) kΦN kL1 ≤ C with C independent of N ;
Z
(3) for every δ > 0, lim |ΦN (eiθ )| dθ = 0.
N →+∞ δ≤|θ|≤π

We can give an analogous definition of summability kernel using a countinous parameter t > 0
(and typically letting t → 0+ ).
We recall (see [Ka] e.g.) that if {ΦN }N >0 is a summability kernel on T and g ∈ Lp (T),
1 ≤ p < ∞, then ΦN ∗ g → g in the Lp -norm, and that if g ∈ C(T), then ΦN ∗ g → g in the
sup-norm.
 HARMONIC ANALYSIS 3

Two of the most useful summability kernels, and also best known, are Fejér’s kernel (which
we denote by {KN }) defined by
N 
X |j|  ijt
KN (t) = 1− e , (1.3)
N +1
j=−N

and the Poisson kernel, given by

+∞ +∞
X X 1 − r2
Pr (t) = r|n| eijt = 1 + 2 rj cos(jt) = . (1.4)
1 − 2r cos t + r2
j=−∞ j=1

The fact that KN satisfies condition (1) in Def. 1.5 above is obvious; that KN satisfies (2)
and (3) is clear from the following result.
Lemma 1.6. We have that
sin N 2+1 t 2


1
KN (t) = .
N +1 sin 2t
Proof. Recall that
t
1 1 1 1
sin2= (1 − cos t) = − e−it + − eit .
2 2 4 2 4
A direct computation of the coefficients in the product shows that
N
 1 −it 1 1 it  X  |j|  ijt 1  1 −i(N +1)t 1 1 i(N +1)t 
− e + − e 1− e = − e + − e .
4 2 4 N +1 N +1 4 2 4
j=−N

This proves the lemma. 2

We now have some consequences of the fact that {KN } is a summability kernel.
Corollary 1.7. The trigonometric polynomials are dense in Lp (T), 1 ≤ p < ∞ and in C(T) in
the uniform norm.
Moreover, if f ∈ L1 (T) and fˆ(n) = 0 for all n, then f = 0.
Proof. If we set σN (f ) = KN ∗ f , by (1.2) it follows that σN (f ) is a trigonometric polynomial.
The two statements now follow from the properties of summability kernels. 2
The terms σN (f ) are classically called the Cesaro means, that the corollary says that they
converge in norm to f .
Corollary 1.8. (The Riemann-Lebesgue Lemma) Let f ∈ L1 (T), then
lim fˆ(n) = 0 .
n→+∞

Proof. Let ε > 0 and let P be a trigonometric polynomial on T such that kf − P kL1 < ε. If |n|
is greater than degree of P , then
|fˆ(n)| = |(f − P )b(n)| ≤ kf − P kL1 < ε .
The conclusion follows. 2
4 M. M. PELOSO

Definition 1.9. Given an L1 function f , the formal series

+∞
X
fˆ(k)eikt
k=−∞
is called the Fourier series of f .

Perhaps one of the main questions in harmonic analysis is to establish under which condition
the Fourier series of f converges back to f . We observe that the partial sum
n
X n
 X 
Σn (f )(t) = fˆ(k)eikt = eikt ∗ f (t) ,
k=−n k=−n
is then given by the convolution with the kernel Dn
n
sin (n + 12 )t


X
ikt
Dn (t) = e = , (1.5)
k=−n
sin 2t
called the Dirichlet kernel. The last equality follows since
 t n
 X  n
−i 2t 1 1 1 1
X
i2
e −e eikt
= ei(k+ 2 )t − e−i(t− 2 )t = ei(n+ 2 )t − e−i(n+ 2 )t .
k=−n k=−n
It is important to notice that the Dirichlet kernel is not a summability kernel, since it satisfies
(1), but not (2-3) in Def. 1.5. For, in particular
sin (n + 12 )t sin (n + 12 )t
Z π
Z δ

1 1
kDn kL1 (T) = dt ≥ dt ≥ Cn . (1.6)
2π −π sin 2t 2π −δ sin 2t
Hence, the question of the norm convergence of the Fourier series of f to f iteself in the
Lp -norm, is a much harder question.
In some respects the greatest success in representing functions by means of their Fourier series
happens for square summable functions. The reason is that L2 (T) is a Hilbert space1, its inner
product being defined by Z
1
hf, gi = f (t)g(t) dt , (1.7)
2π T
Lemma 1.10. The set of exponentials {eint }n∈Z form a complete orthonormal system for L2 (T).
Proof. The orthogonality and the fact that are of norm 1 is a straightforward calculation, the
completeness follows Corollary 1.7. The details are left as an exercise. 2

We denote by `2 the Hilbert space of square summable sequences {ak }, that is,
+∞
X
`2 = {ak } : ak ∈ C, k ∈ Z and |ak |2 < ∞ .

(1.8)
k=−∞
From the abstract theory of Hilbert spaces and the lemma it now follows the following theorem.
Theorem 1.11. Let f ∈ L2 (T). Then
1For a quick tailored introduction to the theory of Hilbert spaces and convergence of Fourier series see, e.g.
[Ka].
 HARMONIC ANALYSIS 5

+∞ Z 2π
X 1
(i) |fˆ(k)|2 = |f (t)|2 dt;
2π 0
k=−∞
N
X
(ii) f = lim fˆ(k)eikt in the L2 (T) norm;
N →+∞
k=−N
(iii) for any {ak } ∈ `2 , there exists a unique f ∈ L2 (T) such that ak = fˆ(k);
(iv) let f, g ∈ L2 (T), then
Z 2π +∞
1 X
f (t)g(t) dt = fˆ(k)ĝ(k) .
2π 0
k=−∞

1.2. Norm convergence of Fourier series. We have the following fundamental definition.
On the other hand, we have
Definition 1.12. Given f ∈ Lp (T), 1 ≤ p < ∞, or f ∈ C(T), we say that the Fourier series of
f , j∈Z fˆ(j)eijt converges in norm to f if we set Σn (f ) = nj=−n fˆ(j)eijt we have
P P

kΣn (f ) − f k → 0 as n → +∞ ,
where k · k denotes the appropriate norm. In this case, we say that the given space admits
convergence of the Fourier series.
The question of the norm convergence of Fourier series while is tightly connected to the
boundedness of the mapping of real harmonic function to its harmonic conjugate, as we will
soon see. Suppose u is a real harmonic function on the unit disk. Then we know that u admits
a harmonic conjugate v, that is, a real harmonic function such that u + iv is holomorphic in D.
The function v is uniquely determined modulo constants, so we require v to vanish at the origin
z = 0, i.e. we set v(0) = 0.
We observe that if g is trigonometric polynomial of degree m, then Σn (g) = g for n ≥ m;
hence we trivially have norm convergence in any norm for all trigonometric polynomials.
We reduce the problem of the convergence of Fourier series to a single estimate.
Theorem 1.13. With the notation above, let B be any of the Banach spaces Lp (T), 1 ≤ p < ∞,
or C(T). Then B admits norm convergence of Fourier series if and only if there exists a constant
M > 0 such that for any positive integer n
kΣn k ≤ M ,
where kΣn k denotes the operator norm of Σn on B.
Proof. If Σn (f ) converges in norm for any f ∈ B, then kΣn (f )k is a bounded sequence. By the
uniform boundedness principle ([Ru]), the operator norms kΣn k are uniformly bounded.
Conversely, suppose that kΣn k ≤ M , for some M > 0 and all n. Given f ∈ B, fix ε > 0. Let
g be a trigonometric polynomial such that kf − gkB < ε. Then,
kΣn (f ) − f kB ≤ kΣn (f ) − Σn (g)kB + kΣn (g) − gkB + kg − f kB
≤ Mε + ε ,
for n sufficiently large. This proves the result. 
We begin by observing that Fourier series do not converge in L1 (T)-norm, nor in uniform
norm.
6 M. M. PELOSO

Theorem 1.14. Let B be either L1 (T) or C(T). Then B does not admit norm convergence of
Fourier series.
Proof. By the previous theorem, it suffices to estimates the operators norms kΣn k as n →
+∞, when B = L1 (T) and B = C(T). Since Σn (f ) = Dn ∗ f , we have that kΣn (f )kB ≤
kDn kL1 (T) kf kB , so that kΣn kB ≤ kDn kL1 (T) . On the other hand, let B = L1 (T) first. Recall
that Fejér’s kernel (1.3) KN is such that kKN kL1 (bT ) = 1, and KN ∗ f → f in the norm of B.
Then,
kΣn k ≥ kΣn (KN )kL1 (T) = kKN ∗ Dn kL1 (T) → kDn kL1 (T) as N → +∞.
Therefore, when B = L1 (T), kΣn k = kDn kL1 (T) ≥ Cn, see (1.6). Now let B = C(T). We wish
to show also in this case kΣn k → +∞ as n → +∞. Let gn ∈ C(T) be given by sgn(Dn ) where
sgn(Dn ) is continuous, and we may assume that the measure of the set where gn 6= sgn(Dn ) is
less than ε > 0. Then,
Z 2π
1
Σ(gn )(0) = gn (t)Dn (t) dt ≥ kDn kL1 (T) − ε ≥ Cn − ε → +∞ .
2π 0
This completes the proof. 
We now relate the problem of finding a uniform estimate for the norms kΣn k to the (harmonic)
conjugation.
Definition 1.15. Given f ∈ C ∞ (T), we define the conjugate Fourier series the series
X
−i sgn(j)fˆ(j)eijt = Q(f )(eit ) .
j∈Z
We say that the Banach space B admits conjugation if Q extends to a bounded operator on B.
We also define the Szegö projection. Given f ∈ L2 (T), we define
+∞
X
Sf (t) = fˆ(n)eint . (1.9)
n=0
Theorem 1.16. Let B any of the Banach spaces Lp (T), 1 ≤ p < ∞, or C(T). Then the
following are equivalent:
(i) B admits convergence of Fourier series;
(ii) Q is bounded on B;
(iii) the Szegö projection S is bounded on B.
Proof. Since 21 (I + iQ) = S − 12 M , where M g = ĝ(0), (ii) and (iii) are clearly equivalent.
Next, we notice that
2n
Σ]2n (f )
X
fˆ(j)eijt = eint Σn e−int f .


:=
j=0

Thus, the operator norms of Σ]2n are uniformly bounded if and only if the operator norms of Σn
are uniformly bounded. The equivalence of (i) and (iii) now follows from Thm. 1.13. 
Thus, we are left with determining whether, say, the operator Q is bounded on Lp (T), 1 <
p < ∞. To study this boundedness, we need the theory of singular integrals, and in particular
of the analysis of the so-called Hilbert transform. We will address these and further issues in all
the remaining parts of these notes.
 HARMONIC ANALYSIS 7

2. The Fourier transform

We begin by introducing some notation and basic definitions and recalling properties of a few
function spaces.
We denote partial derivatives by ∂xj . We denote by N the set of P non-negative integers. If
n
α is a multi-index, i.e. α ∈ Nn , we write ∂xα = ∂xα11 · · · ∂xαnn , |α| = j=1 α1 + · · · + αn , and
Qn
α! = j=1 α1 · · · αn .
Given a function f on Rn we denote by τy the translation by y, that is,
τy f (x) = f (x − y) . (2.1)
We recall that when g is a continuous function, its support supp
g is the smallest closed set
outside of which it vanishes, that is, it is the closure of g −1 c {0} . We denote by Cc = Cc (Rn )
the space of continuous function with compact support and by C0 = C0 (Rn ) the continuous
functions that vanish at infinity, that is, lim|x|→+∞ g(x) = 0.
The following result is elementary and its proof is left as an exercise (see Exercise 2 in
Appendix B).
Lemma 2.1. Endowed with the uniform norm, the space C0 (Rn ) is a Banach space and Cc (Rn )
is a dense subspace. Functions in C0 are uniformly continuous.
2.1. Convolutions. If f and g are measurable functions on Rn , their convolution f ∗ g is the
function defined as Z
f ∗ g(x) = f (x − y)g(y) dy
for all x for which the integral exists.
The question of measurability of the convolution deserves a discussion. For f Lebesgue
measurable, the function F (x, y) = f (x − y) turns out to be measurable on Rn × Rn . One can
see this fact by writing F = f ◦ g where g(x, y) = x − y and show that g −1 (E) of a Lebesgue
measurable set is still a Lebesgue measurable set.2
We now set some notation that will be use sistematically throughout these notes.
Given f on Rn and t > 0 we define
ft (x) = t−n f (x/t) and f t (x) = f (tx) . (2.2)
Notice that, if f ∈ L1 then
Z Z Z
−n
ft (x) dx = t f (x/t) dx = f (y) dy

and that, if the integral is well defined (for instance if f, g ∈ L2 )

Z Z
ft (x)g(x) dx = f (x)g t (x) dx ,

as a simple change of variables shows.

2In general, if f and g are Lebesgue measurable functions, say f, g : R → R, then the composition f ◦ g, even
when defined, may not be measurable. One way to remedy this, is to redefine f and g on sets of measure 0 and
take them to be Borel measurable. A function f is said to be Borel measurable if f −1 (U ) of any open set U is a
Borel set, that is countable union and/or intersection of open and closed sets.
8 M. M. PELOSO

The next result is elementary, however we need to define the support of a function f that is
locally integrable, but defined only a.e. When f is only a measurable function we define
supp f = c ∪ {A : A open and f = 0 a.e. on A } .



(2.3)
Thus, supp f is a closed set and outside of which f = 0 a.e. For consistency, we need to check
that if f is also continuous, the above definition coincides with the one given for continuous
functions. To do so, we only need to check that if E is a closed set outside of which f vanishes,
then E contains supp f as defined (2.3). But this is clear passing to the complementary sets
(and reversing the inclusion).

Proposition 2.2. Assume that the integrals below exist. Then

(i) f ∗ g = g ∗ f ;
(ii) f ∗ (g ∗ h) = (f ∗ g) ∗ h;
(iii) τy (f ∗ g) = (τy f ) ∗ g = f ∗ (τy g);
(iv) supp(f ∗ g) ⊆ supp f + supp g = {x ∈ Rn : x = y + z, y ∈ supp f, z ∈ supp g}.
From Fubini’s theorem it is clear that the convolution of two L1 -functions is still in L1 . But
more can be said. Now we prove the following.
Theorem 2.3. The following properties hold.
(i) (Young’s inequality) If f ∈ L1 and g ∈ Lp , with 1 ≤ p ≤ ∞ then f ∗ g exists for almost
every x, f ∗ g ∈ Lp and
kf ∗ gkp ≤ kf k1 kgkp .
(ii) (Generalized Young’s inequality) Let 1 ≤ p, q, r ≤ ∞ be such that 1/p + 1/q = 1 + 1/r.
Let f ∈ Lp , q ∈ Lq , then f ∗ g ∈ Lr and
kf ∗ gkr ≤ kf kp kgkq .
0 1 1
(iii) If f ∈ Lp and g ∈ Lp , where p + p0 = 1, 1 ≤ p < ∞, then f ∗ g ∈ C0 and
kf ∗ gk∞ ≤ kf kp kgkp0 .
(iv) If f ∈ L1 and g ∈ C k and ∂xα g ∈ L∞ for |α| ≤ k, then, for |α| ≤ k, ∂xα (f ∗ g) ∈ C0 and
∂xα (f ∗ g) = f ∗ ∂xα g .
Proof. Notice that (i) follows from (ii) by taking p = 1 and q = r in (ii). Now we prove (ii).
Notice that 1/p0 + 1/q 0 + 1/r = 1. Then, by the generalized Hölder’s inequality3
Z
|f ∗ g(x)| ≤ |f (x − y)| |g(y)| dy
Z
= |f (x − y)|p/r |f (x − y)|1−p/r |g(y)|q/r |g(y)|1−q/r dy
Z 1/r Z 1/q0 Z 1/p0
0 0
≤ |f (x − y)|p |g(y)|q dy |f (x − y)|q (1−p/r) dy |g(y)|p (1−q/r) dy
Z 1/r
p q
= |f (x − y)| |g(y)| dy kf k1−p/r
p kgkq1−q/r ,

3See Exercise 4.
 HARMONIC ANALYSIS 9

where we have used the fact that 1/q 0 = 1/p − 1/r so that q 0 (1 − p/r) = p and analogously
1/p0 = 1/q − 1/r and p0 (1 − q/r) = q. Then,
Z Z Z 
r
|f ∗ g(x)| dx ≤ |f (x − y)| |g(y)| dy kf kr−p
p q
p kgkq
r−q
dx
Z  Z 
= kf kr−p
p kgk r−q
q |g(y)| q
|f (x − y)| p
dx dy

= kf krp kgkrq
from which the conclusion follows.
For part (iii) see we use differentiation under the integral sign, see Exercise 5. Let |α| = 1, that
is, α = ej . Then, ∂xj g is bounded and |f (y)∂xj g(x − y)| ≤ C|f (y)|, where f ∈ L1 , indepently of
x. Then we can pass the derivative under the integral sign
Z  Z
∂ xj f (y)g(x − y) dy = f (y)∂xj g(x − y) dy ,

and finally proceed by induction. 2

Theorem 2.4. Let ϕ ∈ L1 and let ϕ(x) dx = a. Then,

R

(i) If f ∈ Lp , 1 ≤ p < ∞, then f ∗ ϕt → af in the Lp -norm, as t → 0.

(ii) If f is bounded and uniformly continuous, then ϕt ∗ f → af uniformly as t → 0.
Proof. We have
Z


f ∗ ϕt (x) − af (x) = f (x − y) − f (x) ϕt (y) dy
Z


= f (x − tz) − f (x) ϕ(z) dz
Z


= (τtz f )(x) − f (x) ϕ(z) dz .

Now we apply Minkowski’s integral inequality4

Z Z


kf ∗ ϕt − af kp = (τtz f ) − f (·)ϕ(z) dz p
≤ (τtz f ) − f p
|ϕ(z)| dz .

But k(τtz f )−f kp ≤ 2kf kp and tends to 0 as t → 0+ for each z (see Exercise I.6). Thus, assertion
(i) follows from the dominated convergence theorem.
If f is bounded and uniformly continuous, then it is still true that k(τtz f )−f k∞ ≤ 2kf k∞ and
tends to 0 as t → 0+ for each z. Thus, assertion (ii) also follows from the dominated convergence
theorem. 2

Definition 2.5. We denote by D (or Cc∞ ) the space of C ∞ -functions with compact support,
and more generally by Cc∞ (Ω) the space of C ∞ -functions having compact support contained in
Ω.
The next result is elementary, however, we invite the readers to convince themselves of its
validity.
4See Appendix, Theorem A.8.
10 M. M. PELOSO

Lemma 2.6. The space D is non-empty; for instance the function

− 1
(
e 1−|x|2 if |x| < 1
ϕ(x) =
0 if |x| ≥ 1
is an element of D.
By translation and dilation, it is easy to produce an element ψ of D with supp ψ ⊆ B(x0 , δ);
it suffices to take ϕ((x − x0 )/δ).
Proposition 2.7. (C ∞ -Urysohn’s Lemma) Let K be a compact set in Rn , U an open set that
contains K. Then there exists f ∈ D such that f ≥ 0, f = 1 on K and f vanishes outside U .
Proof. Let δ = dist(K, c U ) and let V = {y : dist(y, K) < δ/3}. Let ψ ∈ D, ψ = 1,
R
supp ψ ⊆ {x : |x| ≤ δ/3}. Notice that ψ(x) = cϕ(3x/δ), with ϕ as in the previous lemma,
satisfies the requirements for a suitable constant c. Then we set f = χV ∗ ψ. We have that,
supp f ⊆ V + supp ψ ⊆ {dist(x, K) < 2δ/3} ⊆ U ,
Moreover, 0 ≤ f ≤ 1, and notice that for x ∈ K, y ∈ V , |x − y| < δ/3 and
Z Z
f (x) = χV (y)ψ(x − y) dy = ψ(x − y) dy
V ∩{y: |x−y|<δ/3}
Z
= ψ(x − y) dy = 1 .
{y: |x−y|<δ/3}

This proves the lemma. 2

Corollary 2.8. The space D is dense in Lp , 1 ≤ p < ∞.

Proof. We have shown that, if f ∈ Lp , 1 ≤ p < ∞, and ϕ ∈ D, ϕ = 1, then f ∗ ϕt → f in
R
the Lp -norm. Let f ∈ Lp and let ε > 0 be fixed. Let K be a compact set chosen in such a way
that kf χ c K kp < ε/2. Next, let t0 be such that then kf χK − (f χK ) ∗ ϕt kp < ε/2 for 0 < t < t0 .
Notice that (f χK ) ∗ ϕt ∈ D for all t > 0. Then
kf − (f χK ) ∗ ϕt kp < kf − f χK kp + kf χK − (f χK ) ∗ ϕt kp < ε . 2
We conclude this section with a result (Prop. 2.10) that we will not use, but that is quite
useful in a variety of problems.
Lemma 2.9. Let E be a compact set in Rn and let {Uj }, j = 1 .P . . , N be a cover of E. Then,
there exist ϕ1 , . . . , ϕN ∈ D such that supp ϕj ⊆ Uj for all j, j ϕj (x) ≤ 1 for all x, and
P
j ϕ j (x) = 1 for all x ∈ E.

Proof. We first observe that there exist compact sets Ej such that Ej ⊂ Uj and E = ∪N j=1 Kj .
Indeed, for each x ∈ E there exists a ball B(x, rx ) such that B(x, rx ) ⊆ Uj for some j = 1, . . . , N .
Since E is compact, we can select B(x1 , rx1 ), . . . , B(xM , rxM ) so that E ⊆ ∪M
k=1 B(xk , rxk ). Now,
we let

Ej = ∪ B(xk , rxk ) : B(xk , rxk ) ⊆ Uj .
P the existence of ψj ∈ D with support in Uj and equal to 1 on Ej .
Next, Prop. 2.7 guarantees
Then we define ϕj = ψj /( j ψj ).
 HARMONIC ANALYSIS 11

It is easy to see that ϕj ∈ D, 0 ≤ ϕj ≤ 1 and that if x ∈ E, then x ∈ Ej for at least one j, so

that ϕj (x) 6= 0 (in fact ϕj (x) = 1) for at least one j. Therefore, if x ∈ E,
N N
X X ψj (x)
ϕ(x) = ϕj (x) = PN = 1. 
j=1 j=1 k=1 ψk (x)

Proposition 2.10. (Partition of unity) Let {Uk } be a locally finite open cover5 of Rn . Then,
there exists {ϕk } ⊆ D such that
• supp ϕk ⊆ Uk for all k;

• 0 ≤ ϕk ≤ 1 for all k;
n
P
• k ϕk (x) = 1 for all x ∈ R .
Proof. Let Q be the closed unit cube Q = {x : 0 ≤ xj ≤ 1, j = 1, . . . , n} and B the ball centered
at the center of Q and radius 1, so that Q ⊂ B. For κ ∈ Zn we set Qκ = Q + κ and Vκ = B + κ.
Then, Rn = ∪κ∈Z Qκ , Vκ is open containing Qκ and each x ∈ Rn belongs to at most Cn of the
Vκ ’s, with Cn independent of κ.
(κ)
For each κ ∈ Zn , we can find a finite sub-collection {Uj }j=1,...,Nκ of {Uk } that covers Qκ .
For each κ ∈ Zn fixed, we now apply Lemma 2.9 in the case E = Qκ and finite open cover
(κ) (κ)
{Uj }, j = 1, . . . , Nκ and obtain {ψκ,j }, j = 1, . . . , Nκ , such that ψκ,j ∈ D, supp ψκ,j ⊂ Uj ,
and 0 ≤ ψκ,j ≤ 1.
Finally we set
ψκ,j (x)
ϕκ,j (x) = P PNκ .
κ∈Zn j=1 ψ κ,j (x)
It is easy now to convince oneself that the functions ϕκ,j , κ ∈ Zn , j = 1, . . . , Nκ satisfy the
conclusions in the statement. 

2.2. The Fourier transform. We now come to the main object of our study. For f ∈ L1 (Rn )
we define the Fourier transform of f as
Z
f (ξ) = f (x)e−2πixξ dx .
ˆ

We will also denote it by Ff , so that to emphasyse the action of the Fourier transform as a
mapping.
It is clear that fˆ is well defined for f ∈ L1 and that kfˆk∞ ≤ kf k1 . Moreover, fˆ is continuous,
as an immediate application of the dominated convergence theorem. Thus,
F : L1 (Rn ) → C(Rn ) ∩ L∞ (Rn ) (2.4)
is bounded.
We now see the first elementary properties of the Fourier transform.
Proposition 2.11. Let f, g ∈ L1 . Then, the following hold.
(i) For any y, η ∈ Rn , F(τy f )(ξ) = e−2πiyξ fˆ(ξ) and τη fˆ(ξ) = F(e2πixη f )(ξ).
(ii) F(f ∗ g) = fˆĝ.
(iii) If xα f ∈ L1 for |α| ≤ k, then fˆ ∈ C k and ∂ξα fˆ(ξ) = F (−2πix)α f (ξ).

5A collection {U } of sets is called a locally finite cover of X if (i) ∪ U ⊇ X, and (ii) for each x ∈ X there
k k k
exist only finitely many Uk1 , . . . , Ukm that contains x.
12 M. M. PELOSO

(iv) Suppose f ∈ C k , ∂xα f ∈ L1 for |α| ≤ k, then F(∂xα f )(ξ) = (2πiξ)α fˆ(ξ).
Proof. (i) is elementary:
Z Z
(τy f )ˆ(ξ) = f (x − y)e−2πixξ dx = f (x)e−2πi(x+y)ξ dx

= e−2πiyξ fˆ(ξ) ,
and analogously for the other part of the statement.
(ii) is also elementary. By Fubini’s theorem,
ZZ
(f ∗ g)ˆ(ξ) = f (x − y)g(y) dye−2πixξ dx
ZZ
= f (x − y)e−2πi(x−y)ξ dxg(y)e−2πiyξ dy

= fˆ(ξ)ĝ(ξ) .
(iii) We argue as in the proof of Thm. 2.2 (iii). We begin with the case |α| = 1 and observe
that
|∂ξj f (x)e−2πixξ | ≤ C|xj f (x)| ∈ L1 (Rn ) .

Thus, we can pass the differentiation under the integral sign and then proceed by induction.
In order to prove (iv), we first assume that f is such that ∂xα f ∈ C0 for |α| ≤ k (in particular
if f also has compact support). Under this assumption, assume that |α| = 1 first. Then, since
∂xj f vanishes at ∞, by integration by parts we have
Z Z
−2πixξ
∂xj f (x)e dx = − f (x)(−2πiξj )e−2πixξ dx

= 2πiξj fˆ(ξ) .
The case when |α| > 1 follows by induction. The general case now follows by approximating
f ∈ L1 such that ∂xα f ∈ L1 when |α| ≤ k, as follows. Let η ∈ Cc∞ , 0 ≤ η ≤ 1, η = 1 when
|x| ≤ 1, and η = 0 when |x| ≥ 2. Clearly η ε f ∈ L1 having compact support, so the first part of
the argument applies. Then, it is easy to see that ∂xα η ε f → ∂xα f in L1 as ε → 0+ , for every

|α| ≤ k.6 Since F : L1 → C ∩ L∞ is bounded (recall (2.4)) we have that

(2πiξ)α (η ε f ) b (ξ) → (2πiξ)α fˆ(ξ) ,

but also
(2πiξ)α (η ε f ) b (ξ) = F ∂xα (η ε f ) (ξ) → F(∂xα f )(ξ)

uniformly. The conclusion thus follows. 2

Corollary 2.12. (Riemann–Lebesgue Lemma) The Fourier transform maps L1 continuously

into C0 (Rn ):
F : L1 (Rn ) → C0 (Rn ) .

6See also Exercise 8.

 HARMONIC ANALYSIS 13

Proof. We only need to show that for f ∈ L1 , fˆ vanishes at ∞ (that is lim|x|→+∞ ∂xα f (x) = 0).
For f ∈ L1 ∩ Cc1 we have that
n
X Z
ˆ
|ξ||f (ξ)| ≤ ξj f (x)e−2πixξ dx
j=1
≤ c0 ,

i.e. |ξ|fˆ is bounded, that is fˆ ∈ C0 . Finally, the result follows by the density of L1 ∩ Cc1 in L1 .
For, given g ∈ L1 and ε > 0, let fn ∈ L1 ∩ Cc1 such that kfˆn − ĝk∞ ≤ kfn − gk1 → 0. Thus,
fˆn ∈ C0 and converges uniformly to ĝ. The conlusion now follows. 2
We now compute a fundamental Fourier transform.
2
Lemma 2.13. Let a > 0 and f (x) = e−aπ|x| . Then
1 −π|ξ|2 /a
fˆ(ξ) = e .
an/2
Proof. We begin with the case n = 1. By the previous proposition, we can differentiate under
the integral sign and get
  
ˆ 0 −aπx2 ˆ

i −aπx2 0 ˆ


(f ) (ξ) = −2πixe (ξ) = e (ξ)
a
 
i
= (2πiξ)fˆ(ξ)
a
 
2π
= − ξ fˆ(ξ) .
a
2 /a
Therefore, the function eπξ fˆ(ξ) satisfies the differential equation
d πξ2 /a ˆ
e f (ξ) = 0 ,
dξ
2
e−t dt =
R
i.e. it is a constant. In order to compute the constant, we select ξ = 0 and recall that
√
π, so that
Z
2 1
fˆ(0) = e−aπx dx = √ .
a
The n-dimensional case follows from an application of Fubini’s theorem:
Z Z
2
f (ξ) = . . . e−2πixξ e−aπ|x| dx1 · · · dxn
ˆ
n Z
2
Y
= e−2πixj ξj e−aπ|xj | dxj
j=1
n  
1 2
√ e−πξj /a
Y
=
a
j=1
1 2 /a
= e−π|ξ| . 2
an/2
14 M. M. PELOSO

Remark 2.14. When we consider the two types of dilations defined in 2.2 we easily see have
that
t
fbt (ξ) = fˆ(tξ) = (fˆ) (ξ) and fbt (ξ) = t−n fˆ(ξ/t) = (fˆ)t (ξ) , (2.5)
(Exercise).

We now prove
Theorem 2.15. (Inversion Theorem) Let f ∈ L1 be such that fˆ ∈ L1 . Then
Z
f (x) = e2πiξx fˆ(ξ) dξ .

Proof. Let t > 0 and set

2
ψ(ξ) = e2πixξ−πt|ξ| .
Then,
2 |ξ|2 2 /t2
(Fψ)(y) = τx (Fe−πt )(y) = t−n τx e−π|y|


= ϕt (x − y) ,
where ϕ(x) = e −π|x|2
.
Next notice that, if f, ϕ ∈ L1 , then
Z ZZ Z
−2πixξ
f (x)ϕ̂(x) dx = f (x)ϕ(ξ)e dxdξ = fˆ(ξ)ϕ(ξ) dξ . (2.6)

Therefore,
Z Z Z
2πixξ−πt|ξ|2
e fˆ(ξ) dξ = ψ(ξ)fˆ(ξ) dξ = φ̂(y)f (y) dy
Z
= ϕt (x − y)f (y) dy = f ∗ ϕt (x) .
2
Since ϕ(x) dx = e−π|x| dx = 1, we have that f ∗ ϕt → f in the L1 norm as t → 0. On the
R R
left hand side of the equalities
R above we can use the dominated convergence theorem to show
2
that e2πixξ−πt|ξ| fˆ(ξ) dξ → e2πixξ fˆ(ξ) dξ as t → 0. 2
R

As an immediate consequence of the Inversion Theorem we have the following.

Corollary 2.16. If f ∈ L1 and fˆ = 0, then f = 0.
We now extend the definition of the Fourier transform to functions in L2 .
Theorem 2.17. (Plancherel Theorem) Let f ∈ L1 ∩ L2 . Then fˆ ∈ L2 and F|L1 ∩L2 extends
uniquely to a unitary isomorphism of L2 .
Proof. Consider the space H = {f ∈ L1 : fˆ ∈ L1 }. We use the facts that f ∈ L1 implies fˆ ∈ L∞
and that L1 ∩ L∞ ⊂ L2 to see that S ⊂ H ⊂ L2 . Since S is dense in L2 , also H is. Notice that,
for g ∈ H, by the inversion theorem, setting h̄ = ĝ
Z Z
−2πixξ
ĥ(ξ) = e ĝ(x) dx = e2πixξ ĝ(x) dx = g(ξ) .

Therefore, Z Z Z Z
f ḡ dx = f ĥ, dx = fˆh dξ = fˆĝ dξ .
 HARMONIC ANALYSIS 15

Thus, F restricted to H preserves the L2 -scalar product, so that, by taking f = g we see that
kf k2 = kfˆk2 . Since F(H) = H and H is dense in L2 , F extends by continuity to a unique
mapping that is still an isometry. Thus, its image is closed and contains the dense subspace H,
that is, F is a surjective isomorphism of L2 .
Finally, one has to show that such an extension coincides with F on L1 ∩ L2 . Suppose
2
f ∈ L1 ∩ L2 and let ϕ(x) = e−π|x| . Consider f ∗ ϕt . Then f ∗ ϕt ∈ L1 ∩ L2 , F(f ∗ ϕt ) =
2
fˆF(e−πt|·| ) ∈ L1 (since fˆ is bounded). Therefore, f ∗ ϕt ∈ H and f ∗ ϕt → f both in L1 and L2
norms, and F(f ∗ ϕt ) → fˆ both uniformly and in L2 norm. 2

2.3. The Schwartz space. We recall a few basic facts about seminormed linear spaces.7
Let V be a complex linear space. A function % : V → [0, +∞) is called a seminorm if
(a) %(λv) = |λ|v for all v ∈ V and λ ∈ C;
(b) %(v1 + v2 ) ≤ %(v1 ) + %(v2 ) for all v1 , v2 ∈ V.
Notice that if we also require that %(v) = 0 implies v = 0, then % would be a norm.
Let V be a linear space on which there exists a family {%α }α∈A of seminorms with the following
property: For each pair of points v1 , v2 in V there exists %α such that %α (v1 ) 6= %α (v2 ). In this
case we say that the family of seminorms separates the points in V.
A linear space V on which there exists a family {%α }α∈A of seminorms that separates the
points is called a seminormed space.
On a seminormed space V with seminorms {%α }α∈A we define a topology by defining a system
of open sets

Ux,α,ε = y ∈ V : %α (x − y) < ε .
Let V be a (complex) seminormed space and suppose that adimits a countable family P of
seminorms {%k }. Let τP be the topology defined above, which also is the coarsest topology that
make continuous the identity
i : (V, τP ) → (V, %k )
for all k ∈ N.
We say that P is separating if for each v ∈ V there exists k ∈ N such that %k (v) > 0. We will
always assume that P is separating.
Proposition 2.18. With the above notation, the topology τP enjoys the following properties.
(i) The space (V, τP ) is a locally convex Hausdorff space.
(ii) The finite intersections of the sets
Bk,n = {v : %k (v) < 1/n}
form a fundamental system of neighborhoods of 0.
(iii) The topology τP is induced by the metric distance
X 1 %k (v − w)
dP (v, w) = .
2k 1 + %k (v − w)
k

7For proofs and more on this topic, see G. B. Folland, Real Analysis, Modern Techniques and Their Applications,
2 Ed..
16 M. M. PELOSO

(iv) A sequence {vn } in V converges to v in the τP topology if and only if it converges to v

in all the seminorms %k , that is if
lim %k (v − vn ) = 0
n→+∞

for all k ∈ N.
(v) A linear functional T on V, that is, a linear operator from V to C, is continuous in the
τP topology if and only if there exist seminorms %k1 , . . . , %kN and constant C > 0 such
that
XN
|T (v)| ≤ C %kj (v)
j=1

for all v ∈ V.
The distance dP is invariant, that is
dP (v + z, w + z) = dP (v, w)
for all v, w, z ∈ V.
We also recall that if T a linear operator between two linear normed spaces X , Y, then T is
continuous if and only if it is bounded, that is there exists a constant C > 0 such that
kT xkY ≤ CkxkX ,
for all x ∈ X . The proof of this fact is simple. If T linear is bounded, then it is Lipschitz, with
Lipschitz constant less or equal to C:
kT x1 − T x2 kY = kT (x1 − x2 )kY ≤ Ckx1 − x2 kX ,
which of course implies continuity. Conversely, if it continuous, the inverse image of the ball of
radius r > 0 is contained in the ball of radius R > 0, so that, for some R0 < R, and all x with
kxkX ≤ R0 ,
kT xkY ≤ r .
But then, for all x ∈ X ,
kxkX  R0 x  kxkX
kT xkY = T ≤ r.
R0 kxkX Y R0
As a consequence of the results discussed so far, we have the following proposition involving
the continuity (or boundedness) of a linear operator acting between seminormed spaces.
Proposition 2.19. Let V, W be seminormed linear spaces admitting family of seminorms {%α },
{σβ }, resp. Let T : V → W be a linear mapping. Then T is continuous if and only if, for each
seminorm σβ̄ on W there exist seminorms %α1 , . . . , %αN and a constant C = Cβ̄,α1 ,...,αN > 0 such
that for all v ∈ V we have
XN
σβ̄ (T v) ≤ C %αj (v) .
j=1
 HARMONIC ANALYSIS 17

A space of function of fundamental importance is the Schwartz space S(Rn ) (often denoted
as S). We denote by %(α,β) the seminorm

%(α,β) (f ) = sup |xα ∂ β f (x)|

x∈Rn

and define
S = f ∈ C ∞ (Rn ) : %(α,β) (f ) < ∞, for all multi-indices α, β .


Notice that for every positive integer N there exists a constant C = CN > 0 such that for all
x ∈ Rn we have
1 X α X
|x | ≤ (1 + |x|)N ≤ C |xα | . (2.7)
C
|α|≤N |α|≤N

(We leave the simple proof as an exercise.)

Observe that if ϕ ∈ S, for every positive integer N
X
(1 + |x|)N |ϕ(x)| ≤ C %α,0 (ϕ) ≤ CN ,
|α|≤N

so that
CN
|ϕ(x)| ≤ .
(1 + |x|)N
Thus, Schwartz functions are in C0 , hence uniformly continuous. Moreover, they decay faster
than the reciprical of any polynomial. This fact easily implies that S is contained in Lp , for all
p, 1 ≤ p ≤ ∞, and in particular that for 1 ≤ p < ∞,
Z
kϕkpp = (1 + |x|)N p |ϕ(x)|p (1 + |x|)−N p dx
X
≤C %α,0 (ϕ)(1 + |x|)−N p dx ,
|α|≤[N p]+1

which is finite if N is chosen so that N p − (n − 1) > 1.

2
It is clear that S contains D, so it is not empty. However, such inclusion is proper since e−|x| ,
1/ cosh x are examples of funtions in S that do not have compact support.
Lemma 2.20. Let p be a polinomial of degree d and γ be a multi-index. Then, the mapping
T : S 3 ϕ → ∂xγ pϕ) ∈ S
is bounded (i.e. continuous).
Proof. For a given seminorm %(α,β) and any ϕ ∈ S we have
%(α,β) ∂xγ ϕ = %(α,β+γ) (ϕ) ,

so ϕ 7→ ∂xγ ϕ is a continuous mapping

P of S into0 itself.
On the other hand, if p(x) = |α0 |≤d aα0 xα , then
0 0
X
∂xβ (pϕ = bα0 xα ∂xβ (ϕ)

|α0 |≤d, |β 0 |≤|β|

18 M. M. PELOSO

for some coefficients bα0 , so that

X
%(α,β) (pϕ) ≤ C %(α+α0 ,ββ0 ) (pϕ) .
|α0 |≤d, |β 0 |≤|β|

Thus, also ϕ → pϕ is a continuous mapping of S into itself, and the conclusion follows as
composition of continuous functions. 
Proposition 2.21. The space S is a complete metric space (i.e. a Frèchet space) in the topology
defined by the seminorms %(α,β) . Moreover, C0∞ is dense in S.
Proof. Using (iii) of the previous Prop. 2.18, we only need to prove that S is complete in the
given topology.
Let {fk } be a Cauchy sequence in S, then %(α,β) (fj − fk ) → 0 as j, k → +∞ for all (α, β) . In
particular, {∂ β fk } converges uniformly to a function gβ for all β. Denote by ej the j-th element
of the canonical basis in Rn . Notice that
g0 (x + tej ) − g0 (x) = lim fk (x + tej ) − fk (x)
k→+∞
Z t
= lim ∂xj fk (x + sej ) ds
k→+∞ 0
Z t
= gej (x + sej ) ds .
0

Therefore, ∂xj g0 = gej , and by induction on |β| we obtain that gβ = ∂ β g0 . Finally,

|xα ||∂xβ fk (x) − ∂xβ g0 (x)| = lim |xα ||∂xβ fk (x) − ∂xβ fj (x)|
j→+∞
≤ lim %(α,β) (fk − fj )
j→+∞
≤ ε,
for j, k ≥ k0 . Then %(α,β) (fk − g0 ) → 0 as k → +∞.
In order to show that C0∞ is dense in S, observe that, if η ∈ Cc∞ with η(0) = 1 and ϕ ∈ S,
then, for every ε > 0, η ε ϕ is in Cc∞ , where η ε (x) = η(εx). Now it is easy to check that η ε ϕ → ϕ
in the S-topology, as ε → 0. For, η ε ϕ − ϕ is identically 0 for |ξ| ≤ 1/ε, so that
%α,β (η ε ϕ − ϕ) = sup xα ∂xβ (η ε ϕ − ϕ)(x)
|x|≥1/ε
 
≤ sup xα ∂xβ (η ε ϕ)(x) + xα ∂xβ ϕ(x)
|x|≥1/ε
X
≤ Cε sup xα ∂xγ ϕ(x) .
|γ|≤|β| |x|≥1/ε

Using the fact that xα ∂xγ ϕ ∈ C0 for all α and γ, for a fixed ε̃, we can find ε0 such that the right
hand side above is less or equal to ε̃. The conclusion now follows. 2

A consequence of Prop. 2.11 and the Inversion Theorem is the following.

Proposition 2.22. The Fourier transform F : S → S is a continuous bijection, with continuous
inverse.
 HARMONIC ANALYSIS 19

We remark that, as a consequence of these facts, we obtain a simple proof that the convolution
of two Schwartz functions is still a Schwartz function. For, ϕ ∗ ψ = F −1 (ϕ̂ψ̂), where ϕ̂ψ̂ ∈ S.

Proof. In order to prove the continuity of F, and therefore the one of F −1 , we need to show
that for every semi-norm %α,β , there exist multi-indices α1 , β1 , . . . , αN , βN and C > 0 such that
for every f ∈ S,
XN


%(α,β) F(f ) ≤ C %(αj ,βj ) (f ) .
j=1
But,
ξ α ∂ξβ (fˆ) = cβ ξ α F(xβ f ) = cα,β F ∂xα (xβ f ) .

%(α,β) F(f ) = sup |ξ α ∂ξβ (fˆ)(ξ)|

ξ∈Rn

= cα,β sup F ∂xα (xβ f ) (ξ)

ξ∈Rn

≤ Ck∂xα (xβ f )k1

X
%(α0 ,0) ∂xα (xβ f ) .


≤C
|α0 |≤N

It follows that F is bounded from S into itself. The Inversion theorem, Thm. 2.15, shows that
ϕ 7→ ϕ̂(−·) is its inverse. So, F is one to one, onto and its inverse is also continuous. 2
We conclude this part by observing that, using the bounds in (2.7), if we set
%e(N,β) (f ) = sup (1 + |x|)N |∂xβ f (x)| ,
x∈Rn

for all non-negative intergers N and multi-indices β, we obtain a countable family of seminorms
{e
%N,β } that defines a topology in S equivalent to the given one.

2.4. The space of tempered distributions. We define the space S 0 of tempered distributions
as the dual space of S. We endow S 0 with the weak-∗ topology, that is, the weakest topology
that makes the elements of S continuous on S 0 , that is, the functionals
S 0 3 u 7→ u(ϕ)
for ϕS fixed. A ngbh basis for the weak-∗ topology is given by the sets
Iϕ1 ,...,ϕN ;ε (u) = v ∈ S 0 : |u(ϕj ) − v(ϕj )| < ε , j = 1, . . . , N ,

(2.8)
where ϕ1 , . . . , ϕN are in S and ε > 0.
There are many noticeable examples of tempered distributions.
Examples 2.23. (i) Functions in S, in any Lp class, 1 ≤ p ≤ ∞ give rise to bounded linear
functionals on S, by setting:
Z
Lf (ϕ) = f ϕ dx ϕ ∈ S, f ∈ Lp (or S) .
20 M. M. PELOSO

For,
Z
|Lf (ϕ)| ≤ |f ϕ| dx ≤ kf kLp kϕkLp0
Z 0 0 0
1/p0
= kf kLp (1 + |x|)−N p (1 + |x|)N p |ϕ(x)|p dx
Rn
≤ Ckf kLp sup{(1 + |x|)N |ϕ(x)|}
x
X
≤C %(α,0) (ϕ) ,
|α|≤N

dove N è scelto ≥ n + 1.
(ii) A function f is called tempered if there exists N > 0 such that (1 + |x|)−N f ∈ L1 . Then,
the formula Z
Lf (ϕ) = ϕf dx

defines a tempered distribution, since

Z Z X
| ϕf dx| ≤ sup{(1 + |x|)N |ϕ(x)|} (1 + |x|)−N |f | dx ≤ C %(α,0) (ϕ) .
x
|α|≤N
R
Analogously, a Borel measure µ is called tempered if there exists N > 0 such that (1 +
|x|)−N d|µ| < ∞. Then, the pairing
Z
Lµ (ϕ) = ϕ d|µ|

defines a tempered distribution. Then, Dirac deltas are examples of tempered distributions.
Notice however, not every C ∞ function defines a tempered distribution; for instance one with
exponential growth.

Remark 2.24. On S 0 we can introduce a few operations, besides the ones that define the vector
space structure.
(I) Differentiation. For u ∈ S 0 and α a multi-index we set

∂ α u(ψ) = u (−1)|α| ∂ α ψ .

Notice that, if u ∈ S ⊂ S 0 , then ∂ α u ∈ S and therefore it defines again an element of S 0 by the

integral pairing, and integrating by parts the boundary terms equal 0 (since both u and ψ are
Schwartz functions)
Z Z
(∂ u)(ψ) = (∂ u)ψ dx = − (∂ α−e1 u)(∂x−1 ψ) dx
α α

Z
|α|
= (−1) u∂ α ψ dx .

Hence, the definition of derivative of a distribution coincides with the classical definition if the
distribution admits classical derivatives.
 HARMONIC ANALYSIS 21

(II) Multiplication by a smooth function of polinomial growth. Given f ∈ C ∞ , we say that f

is of moderate growth if for each multi-index α there exist an integer N and a positive constant
C such that
|∂xα f (x)| ≤ C(1 + |x|)N .
Clearly such a function defines an element of S 0 , and if ϕ ∈ S, then f ϕ ∈ S. Moreover, if u ∈ S 0 ,
we can define another element f u of S 0 by setting
(f u)(ϕ) = u(f ϕ) .

(III) Fourier transform. Given u ∈ S 0 we define the tempered distribution û by setting

û(ψ) = u(ψ̂) ,
for any ψ ∈ S. Observe that, using identity (2.6), this defintion extends the definition of the
Fourier transform given on S; that is, if u above is in fact an element of S, then the equality
above is justified by (2.6).
Proposition 2.25. The Fourier transform F is a surjective topological isomorphism of S 0 onto
itself.
Proof. Clearly F maps S 0 into itself, and using the fact that F is a bijection on S, it is easy to
see that it is a bijection on S 0 too. For, if û(ϕ) = 0 for all ϕ, then u(ϕ̂) = 0 for all ϕ ∈ S, that
is, u(ψ) = 0 for all ψ ∈ S, i.e. u = 0. In order to see that F is onto, given u ∈ S 0 , define v ∈ S 0
by setting v(ψ) = u(F −1 ψ). Then, v is well defined and
u(ϕ) = u F −1 ϕ̂ = v(ϕ̂) = v̂(ϕ) ,

for all ϕ ∈ S. Then, u = Fv, that is, F is onto. Notice also that we have shown that the inverse
Fourier transform on tempered distributions is given by
F −1 u = u ◦ F −1 .
The fact that F and F −1 are continuous on S 0 follows from the description of the weak-∗
topology: a ngbh basis of u ∈ S 0 is given by the sets
Iϕ1 ,...,ϕN ;ε (u) = v ∈ S 0 : |u(ϕj ) − v(ϕj )| < ε , j = 1, . . . , N ,


where ϕ1 , . . . , ϕN are in S and ε > 0. Let v ∈ Iϕ1 ,...,ϕN ;ε (u). Then,

û(ϕj ) − v̂(ϕj ) = u(ϕ̂j ) − v(ϕ̂j ) < ε ,
if v ∈ Iϕ̂1 ,...,ϕ̂N ;ε (u), where Iϕ̂1 ,...,ϕ̂N ;ε (u) has the obvious meaning. Therefore,


F Iϕ̂1 ,...,ϕ̂N ;ε (u) ⊂ û(ϕj ) − v̂(ϕj )
and this shows that F : S 0 → S 0 is continuous and we are done. 2

(IV) Convolution with a Schwartz function ϕ. We first define the operator ˇ on functions by
setting ϕ̌(x) = ϕ(−x). Then, for ϕ ∈ S and u ∈ S 0 we set8


(u ∗ ϕ)(x) = u τx ϕ̌ . (2.9)
Proposition 2.26. Let ϕ, ψ ∈ S and u ∈ S 0 . Then
8We adopt the convention that τ ϕ̌ = ϕ̌(· − x) = ϕ(x − ·).
x
22 M. M. PELOSO

(i) u ∗ ϕ ∈ C ∞ and for every multi-index α

∂ α (u ∗ ϕ) = (∂ α u) ∗ ϕ = u ∗ (∂ α ϕ);
(ii) u ∗ ϕ is of moderate growth, so that u ∗ ϕ ∈ S 0 ;
(iii) we have Z

Z

u τx ϕ̌ ψ(x) dx = u τx ϕ̌ (y)ψ(x) dx ;
Rn Rn
(iv) the action of u ∗ ϕ on S is given by the formula
(u ∗ ϕ)(ψ) = u(ϕ̌ ∗ ψ) ;
(v) F(u ∗ ϕ) = ûϕ̂ and F(uϕ) = û ∗ ϕ̂;
(vi) (u ∗ ϕ) ∗ ψ = u ∗ (ϕ ∗ ψ).

Notice that the pairing in (iv) extends the case when also u ∈ S that we obtain by switching
the integration order:
Z Z 
(u ∗ ϕ)(ψ) = u(y)ϕ(x − y) dy ψ(x) dx
Z Z 
= u(y) ϕ(x − y)ψ(x) dx dy
Z
= u(y)(ϕ̌ ∗ ψ)(y) dy .

The same comments applies to the identity in (iv): if u is also in S, then the convolutions are
given by absolutely convergent integrals and we have
Z Z  Z Z 
(u ∗ ϕ) ∗ ψ(x) = u(z)ϕ(y − z) dz ψ(x − y) dy = u(z) ϕ(y − z)ψ(x − y) dy dz
Z Z  Z
0 0 0
= u(z) ϕ(y )ψ(x − z − y ) dy dz = u(z)(ϕ ∗ ψ)(x − z) dz
= u ∗ (ϕ ∗ ψ)(x) .
Proof. (i) We first claim that
ϕ(· + te1 ) − ϕ 1

ϕ(t) := = τ−te1 ϕ − ϕ → ∂x1 ϕ (2.10)
t t
as t → 0 in S.
Assume the validity of the claim for the moment. Using the fact that the convolution com-
mutes with translations, it follows that
(u ∗ ϕ)(x + te1 ) − (u ∗ ϕ)(x) 1 
= u(τx+te1 ϕ̌) − u(τx ϕ̌)
t t
1 
=u τte1 (τx ϕ̌) − τx ϕ̌
t

→ u − ∂y1 (τx ϕ̌)



= −u τx ∂y1 (ϕ̌) = u τx (∂y1 ϕ)ˇ
= (u ∗ ∂y1 ϕ)(x) ,
as t → 0, by definition (2.9).
 HARMONIC ANALYSIS 23

Iterating this argument we obtain that u ∗ ϕ is C ∞ and

∂ α (u ∗ ϕ) = u ∗ (∂ α ϕ) ,
for all multi-indices α. Since
u ∗ (∂ α ϕ)(x) = u τx (∂ α ϕ)ˇ = (−1)|α| u ∂ α (τx ϕ̌)

= (∂ α u) τx ϕ̌

= (∂ α u) ∗ ϕ(x) ,
also the first equality in (i) follows, modulo equality (2.10).
In order to prove (2.10) it suffices to prove that
1

F τ−te1 ϕ − ϕ − ∂x1 ϕ → 0
t
in S, that is,
 
t−1 e2πitξ1 − 1 − 2πiξ1 ϕ̂ → 0

in S, as t → 0. It this is easy to check that the function m(ξ) = t−1 e2πitξ1 − 1 − 2πiξ1 is C ∞ ,


of moderate growth together with all its derivatives, uniformly so in t, and they all tend to 0
uniformly as t → 0+ ; hence the conclusion. This proves (i).
(ii) In ordet to show that u∗ϕ is of moderate growth we need to show that for each multi-index
α there exist an integer N and a positive constant C such that
|∂xα (u ∗ ϕ)(x)| ≤ C(1 + |x|)N .
Since ∂xα (u ∗ ϕ) = u ∗ (∂xα ϕ), and ∂xα ϕ is again a Schwartz function, it suffices to prove the case
α = 0. Since u ∈ S 0 , given ϕ ∈ S there exist semi-norms %(α1 β1 ) , . . . , %(αN ,βN ) and C > 0 such
that
XN
|u(τx ϕ̌)| ≤ C %(αj βj ) (τx ϕ̌) .
j=1
Now, for any semi-norm %(α,β) ,

%(α,β) τx ϕ̌ = sup y α ∂yβ τx ϕ̌(y) = sup y α ∂yβ ϕ(x − y)

y∈Rn y∈Rn

= sup (x + y)α ∂yβ ϕ(y) ,

y∈Rn

which is clearly bounded by a polynomial in x.9

(iii) We have the following
Claim. Let η ∈ C0∞ . Then, the Riemann sums of the integral
R
Rn (τx ϕ̌)(y)η(x) dx converge in
the topology of S to ϕ̌ ∗ η(y).
For simplicity of notation assume n = 1, the general case is similar. If η ∈ C0∞ , supp η ⊆ I,
denoting by {x0 < x1 < · · · < xN } a partition of the interval I, ∆j = xj − xj−1 , x̃j an arbitrary

9The reader should check that expicitly this polynomial is a constant times (1 + |x|)k P
|α0 |≤|α| %(α ,β) (ϕ).
0
24 M. M. PELOSO

point in Ij = (xj−1 , xj ), we have

Z N
X
ϕ(x − y)η(x) dx = lim ∆j ϕ(x̃j − y)η(x̃j )
I N →+∞
j=1

and we want to show that the sequence

N
X Z
fN (y) := ∆j ϕ(x̃j − y)η(x̃j ) → ϕ(x − y)η(x) dx (2.11)
j=1 I

in S as N → +∞. In fact,
Z N Z 
X 
ϕ(x − y)η(x) dx − fN (y) = ϕ(x − y)η(x) − ϕ(x̃j − y)η(x̃j ) dx
I j=1 Ij

By the uniform continuity of the integrand, as a function of x, uniformly in y, given ε > 0 there
exists δ > 0 such that
ϕ(x − y)η(x) − ϕ(x̃j − y)η(x̃j ) < ε
when ∆j < δ, for all j = 1, . . . , N .
Therefore, choosing a partition of I such that ∆j < δ for all j, that is N ≥ |I|/δ, we have
Z XN Z
ϕ(x − y)η(x) dx − fN (y) ≤ ε = ε|I| .
I j=1 Ij

This shows that

Z 
%(0,0) ϕ(x − ·)η(x) dx − fN → 0
I
as N → +∞. The cases of the seminorms %(α,β) for all other α and β follow from this one.
In the case |β| > 0 we have
Z  Z 
β |β|
∂y ϕ(x − ·)η(x) dx − fN = (−1) ∂yβ ϕ(x − ·)η(x) dx − fβ;N ,
I I

where fβ;N is defined as fN in (2.11) with ϕ replaced by ∂yβ ϕ. Hence,

Z  Z 
%(0,β) ϕ(x − ·)η(x) dx − fN ≤ C%(0,0) ∂yβ ϕ(x − ·)η(x) dx − fβ;N → 0
I I
as N → +∞,
In the case |α| > 0,
Z 
α
y ϕ(x − ·)η(x) dx − fN (y)
I
Z N
X
α
= (y − x + x) ϕ(x − y)η(x) dx − ∆j (y − x̃j + x̃j )α ϕ(x̃j − y)η(x̃j )
I j=1
Z N
0

α0 β0 0
∆j (y − x̃j )α ϕ(x̃j − y)x̃βj η(x̃j ) .
X X
= c
α0 ,β 0 (x − y) ϕ(x − y)x η(x) dx −
α0 +β 0 =α I j=1
 HARMONIC ANALYSIS 25

Therefore,
Z  X Z 
%(α,0) ϕ(x − ·)η(x) dx − fN ≤ C %(0,0) ϕα0 (x − ·)ηβ 0 (x) dx − fα0 ,β 0 ;N → 0
I α0 +β 0 =α I

0 0
as N → +∞, where ϕα0 = xα ϕ, ηβ 0 = xβ η and fα0 ,β 0 ;N is defined as fN , with ϕ and η replaced
by ϕα0 , ηβ 0 , resp.
The claim, at least in the case n = 1, now follows by combining the above cases. We leave
the simple details to the reader.
Therefore, when η ∈ C0∞ (I),
Z
 Z 
u τx ϕ̌ (y)η(x) dx = u ϕ(x − y)η(x) dx
R I
 N
X 
= u lim ∆j ϕ(x̃j − y)η(x̃j )
N →+∞
j=1
N
X

= lim ∆j u ϕ(x̃j − ·) η(x̃j )
N →+∞
j=1
Z


= u ϕ(x − ·) η(x) dx
ZI


= u τx ϕ̌ η(x) dx .
R
Thus,
Z
 Z


u τx ϕ̌ (·)η(x) dx = u τx ϕ̌ η(x) dx (2.12)

when η ∈ C0∞ (I).

The case n > 1 can be proved with an analogous argmunent, which is just more involved
notationally (because of the expression of the Riemann sums in several variables and the decom-
position of the expressions of the form (y − x + x)α ). Again, we leave the details to the reader,
and therefore we consider that (2.12) (has been proved and) it is valid in the case of Rn .
Next, for a generic ψ ∈ S, let ηk ∈ C0∞ , ηk → ψ in S, then, as k → +∞,
Z Z



◦ τx ϕ̌ (y)ηk (x) dx → τx ϕ̌ (y)ψ(x) dx in S;
Z Z



◦ u τx ϕ̌ ηk (x) dx → u τx ϕ̌ ψ(x) dx .
These conditions imply that
Z
 Z

u τx ϕ̌ (·)ψ(x) dx = lim u τx ϕ̌ (·)ηk (x) dx
k→+∞
Z


= lim u τx ϕ̌ ηk (x) dx
k→+∞
Z


= u τx ϕ̌ ψ(x) dx .

This proves (iii).

26 M. M. PELOSO

In order to prove the identity (iv), notice that, by (iii)

Z



u ∗ ϕ (ψ) = u τx ϕ̌ ψ(x) dx
Rn
Z

=u τx ϕ̌ (y)ψ(x) dx
Rn
Z 
=u ϕ(x − y)ψ(x) dx
Rn
= u ϕ̌ ∗ ψ) .

(v) For all ψ ∈ S, since u ∗ ϕ ∈ S 0 , using (iv) we have

F(u ∗ ϕ)(ψ) = (u ∗ ϕ)(ψ̂) = u ϕ̌ ∗ ψ̂) = u F(ϕ̂ψ)
= û(ϕ̂ψ) = (ûϕ̂)(ψ) .
This shows that F(u ∗ ϕ) = ûϕ̂ Next, we wish to show that F(uϕ) = û ∗ ϕ̂. For we have ψ ∈ S,
F(uϕ)(ψ) = (uϕ)(ψ̂) = u(ϕψ̂) = u F(F −1 ϕ ∗ ψ)

= û F −1 ϕ ∗ ψ = û ∗ (F −1 ϕ)ˇ (ψ) = û ∗ ϕ̂) (ψ) .

This implies F(uϕ) = û ∗ ϕ̂, as we wished to show.

Finally (vi) follows at one from (v) by checking the equality of their Fourier transforms:




F (u ∗ ϕ) ∗ ψ = F (u ∗ ϕ) (ψ̂) = (ûϕ̂)(ψ̂) = F u ∗ (ϕ ∗ ψ) . 

(V) We conclude this part with a simple observation. It follows from (i) and (iv) that if u ∈ S 0
(∂ α u)ˆ(ϕ) = (∂ α u)(ϕ̂) = (−1)|α| u ∂ α ϕ̂ = u F (2πiξ)α ϕ̂ = (2πiξ)α u




b(ϕ) .

A consequence of these facts is that S is dense in S 0 .

Proposition 2.27. If ϕ ∈ S, ϕ = 1 and u ∈ S 0 , then u ∗ ϕt → u in S 0 , as t → 0. Moreover,
R
C0∞ , hence S, is dense in S 0 in its topology.
Proof. Given ψ ∈ S, we have
(u ∗ ϕt )(ψ) = u ϕ̌t ∗ ψ) → u(ψ)
R
as t → 0, since ϕt ∗ ψ → ψ in S and ϕ̌ = 1.
Next, arguing as in the proof of Prop. 2.21, let η ∈ C0∞ , 0 ≤ η ≤ 1, η(x) = 1 if |x| ≤ 1 and
0
η(x) = 0 if |x| ≥ 2. Since u ∗ ϕt ∈ C ∞ (and it is of moderate growth), η t (u ∗ ϕt ) ∈ C0∞ . For
ψ ∈ S we have,
0 0 0
η t (u ∗ ϕt ) (ψ) = (u ∗ ϕt )(η t ψ) = (u ∗ ϕt )(ψ) + (u ∗ ϕt )(η t ψ − ψ)


0 0
= u(ψ) + (u ∗ ϕt − u)(ψ) + u(η t ψ − ψ) + (u ∗ ϕt − u)(η t ψ − ψ) .
 HARMONIC ANALYSIS 27

0
Given ε > 0 we can choose t0 > 0 small enough so that |u(η t ψ − ψ)| < ε, and then t > 0 small
0
enough so that |(u ∗ ϕt − u)(η t ψ − ψ)| < ε and also |(u ∗ ϕt − u)(ψ)| < ε. Hence,
0
η t (u ∗ ϕt ) − u (ψ) < 3ε .

The conclusion now follows. 2

The next result, that we state without proof (for which we refer to [StWe, Thm. 3.16]) is an
example of the significance of the space of tempered distributions.
Theorem 2.28. Let T : Lp (Rn ) → Lq (Rn ) be a bounded linear operator, 1 ≤ p, q ≤ ∞, that
commutes with translations (that is, T (τx f ) = τx (T f )).
Then, there exists a tempered distribution K such that, for f ∈ S, T f = f ∗ K.
28 M. M. PELOSO

3. The Marcinkiewicz interpolation theorem and the Calderón–Zygmund

decomposition

3.1. The Marcinkiewicz interpolation theorem. We recall that, given a measurable func-
tion f on Rn , its distribution function αf is the function, defined for λ ≥ 0
αf (λ) = {x ∈ Rn : |f (x)| > λ} . (3.1)
Using Fubini’s theorem, it is easy to see that for p > 0
Z +∞
p
kf kLp = p λp−1 αf (λ) dλ . (3.2)
0

We also observed that

αf1 +f2 (λ) ≤ αf1 (λ/2) + αf2 (λ/2) ,
since
{x ∈ R : |(f1 + f2 )(x)| > λ} ⊆ {x ∈ R : |f1 (x)| > λ/2} ∪ {x ∈ R : |f2 (x)| > λ/2} .

An operator T defined on measurable functions on Rn is said to be sublinear if

- |T (f0 + f1 )(x)| ≤ |T (f0 )(x)| + |T (f1 )(x)|;
- |T (λf )(x)| = |λ||T f (x)|;
for all x ∈ Rn , λ ∈ C. We remark that, if T is a linear operator, |T | is sublinear. Moreover,
a typical example of an operator thatR is genuinely sublinear is the Hardy–Littlewood maximal
1
function M f (x) = supB(x,r),r>0 |B| B |f (x)|dx.

Definition 3.1. Given a sublinear operator T defined on measurable functions on Rn is of

weak-type (p, q) if there exists a constant C > 0 such that
C q
αT f (λ) = {x ∈ Rn : |T f (x)| > λ} ≤ kf kLp ,
λ
while it is said to be strong-type (p, q) if
T : Lp → Lq
is bounded.
It is easy to see that if T is of strong-type (p, q), then it is also of weak-type (p, q). For,
Z
αT f (λ) = dx
{x∈Rn : |T f (x)|>λ}
 q
|T f (x)|
Z
≤ dx
{x∈Rn : |T f (x)|>λ} λ
1
≤ kT f kqLq
λq
C q
≤ kf kLp .
λ
 HARMONIC ANALYSIS 29

Theorem 3.2. (Marcinkiewicz interpolation theorem) Let T be a sublinear operator de-

fined on Lp0 (Rn ) + Lp1 (Rn ), 1 ≤ pj , qj ≤ +∞, j = 0, 1. Suppose that T is of weak-type (p0 , q0 )
and of weak-type (p1 , q1 ). Then, T is of strong-type (p, q), where p and q are given by the
relations
1 θ 1−θ 1 θ 1−θ
= + , and = + ,
p p0 p1 q q0 q1
and 0 < θ < 1.
Proof. We will prove the theorem only in the case q0 = p0 and q1 = p1 . Then it suffices to
assume that p0 < p < p1 , without any reference to θ. For the proof in the general case (that we
need in Subsection 5.6) we refer the reader to [St1].
Let f ∈ Lp and λ > 0 be given. For a constant c > 0 to be selected later, we decompose f as
f = f0 + f1 , where
f0 = f χ{x:|f (x)|>cλ} , and f1 = f χ{x:|f (x)|≤cλ} .
Notice that f0 ∈ L and f1 ∈ Lp1 , since p0 < p < p1 . For,
p 0

|f0 (x)|p0 = |f (x)|p0 −p |f (x)|p χ{x:|f (x)|>cλ} (x)

< (cλ)p0 −p |f (x)|p χ{x:|f (x)|>cλ} (x) , (3.3)
so that f0 ∈ Lp0 . Analogous reasoning shows that f1 ∈ Lp1 :
|f1 (x)|p1 ≤ (cλ)p1 −p |f (x)|p χ{x:|f (x)|≤cλ} (x) . (3.4)
Then,
|T (f0 + f1 )(x)| ≤ |T (f0 )(x)| + |T (f1 )(x)|
and
αT (f0 +f1 ) (λ) ≤ αT f0 (λ/2) + αT f1 (λ/2) .
We distinguish two different cases. First assume that p1 < +∞. Then,
Z +∞
p
kT f kLp = p λp−1 αT f (λ) dλ
0
Z +∞ Z +∞
≤p λp−1 αT f0 (λ/2) dλ + p λp−1 αT f1 (λ/2) dλ
0 0
Z +∞  2A p 0 Z +∞  2A p 1
p−1 0 1
≤p λ kf0 kLp0 dλ + p λp−1 kf1 kLp1 dλ
0 λ 0 λ
Z +∞ Z
p0 p−p0 −1
≤ p(2A0 ) λ |f (x)|p0 dx dλ
0 {x:|f (x)|>cλ}
Z +∞ Z
p1 p−p1 −1
+ p(2A1 ) λ |f (x)|p1 dx dλ
0 {x:|f (x)|≤cλ}
Z Z |f (x)|/c
≤ p(2A0 )p0 |f (x)|p0 λp−p0 −1 dλ dx
R n 0
Z Z +∞
+ p(2A1 )p1
|f (x)|p1
λp−p1 −1 dλ dx
Rn |f (x)|/c
)p0 (2A1 )p0
 
(2A0
=p + kf kpLp .
(p − p0 )(cp−p0 ) (p1 − p)(cp−p1 )
30 M. M. PELOSO

In the second case, that is, p1 = +∞, then T : L∞ → L∞ is bounded. We may choose
c = 1/(2A1 ), where A1 is the (L∞ , L∞ )-norm of T , so that αT f1 (λ/2) = 0.10 Then, we have that
Z +∞  2A p 0
0
kT f kpLp ≤ p λp−1 kf0 kLp0 dλ
0 λ
Z Z |f (x)|/c
= p(2A0 )p0 |f (x)|p0 λp−p0 −1 dλ dx
Rn 0
p
= (2A0 )p0 (2A1 )p−p1 kf kpLp .
p − p0
This proves the theorem. 2
As an application, we show that the Hardy–Littlewood maximal operator M is bounded on
Lp , 1 < p ≤ ∞ and weak-type (1, 1). We recall that, for f ∈ L1loc ,
Z
1
Mf (x) = sup |f (y)| dy .
r>0 |B(x, r)| B(x,r)

We need a lemma. This type of result is called a covering lemma.

Lemma 3.3. Let {B1 , B2 , . . . , Bk } be a finite collection of open balls in Rn . Then there exists
a sub-collection {Bj1 , . . . , Bj` } of pairwise disjoint balls such that
`
X k
[
−n
|Bji | ≥ 3 Bj .
i=1 j=1

Proof. We let Bj1 to be a ball of largest volume among the balls in the collection {B1 , B2 , . . . , Bk }.
Next, we select Bj2 among the remaining balls that are disjoint from Bj1 to be of largest volume,
and we proceed in the same fashion. Since we have a finite number of balls, this process will
terminate after ` steps.
Thus, we have constructed a sub-collection {Bj1 , . . . , Bj` } of pairwise disjoint balls. It Bm
was not selected, that is, m 6∈ {j1 , . . . , j` }, this means that Bm intersects one of the balls
{Bj1 , . . . , Bj` }, say Bji , with |Bji | ≥ |Bm |. Then, 3Bji ⊇ Bm , where 3B denotes the ball with
the same center as B, with radius 3 times as large. The conclusion now follows. 
Theorem 3.4. The operator M is weak-type (1, 1) and bounded on Lp , 1 < p ≤ ∞.
Proof. The operator M is trivially bounded on L∞ . If we show that M is weak-type (1, 1), the
conclusion follows at once from Thm. 3.2.
Let λ > 0, then we wish to show that there exists C > 0 such that
C
{x ∈ Rn : |Mf (x)| > λ} ≤ kf kL1 .
λ
Let E be a compact subset of Ωλ := {x ∈ Rn : |Mf (x)| > λ}. For any x ∈ Ωλ there is an open
ball Bx = B(x, rx ) such that x ∈ Bx and
Z
1
|f (y)| dy > λ .
|Bx | Bx

10For, |T f (x)| ≤ A kf k ∞ ≤ A cλ ≤ λ/2, so that {x : |T f (x)| > λ/2} = ∅.

1 1 1 L 1 1
 HARMONIC ANALYSIS 31

Since E is compact, finitely many such balls {B1 , B2 , . . . , Bk } cover E. Then, we can apply
Lemma 3.3 to obtain a sub-collection {Bj1 , . . . , Bj` } of disjoint balls such that | ∪kj=1 Bj | ≤
3n `i=1 |Bji |. Therefore, for every compact E ⊆ Ωλ we have
P

k
[ `
X
|E| ≤ Bj ≤ 3n |Bji |
j=1 i=1
` Z
X 1
n
≤3 |f (y)| dy
λ Bj
i=1 i

C
≤ kf kL1 .
λ
This proves the theorem. 

3.2. The Calderón–Zygmund decomposition of an L1 -function. We now introduce a

fundamental decomposition of the whole space Rn , the dyadic decomposition. We define the
unit cube to be the set [0, 1)n and the set Q0 to be the family of sets obtained by translating
by mej , where m ∈ Z, and for j = 1, . . . , n, ej is an element of the canonical basis, and all their
possible combinations, that is,
Xn
Q0 = Q = [0, 1)n +

mj ej , mj ∈ Z .
j=1

Analogously, for k ∈ Z we define the family Qk as

Xn
Qk = Q = [0, 2−k )n + 2−k mj ej , mj ∈ Z .


j=1

The union of the famlies Qk , k ∈ Z, is called the family of dyadic cubes in Rn .

It is immediate to see that these sets are cubes with sides parallel to the axes, that a cube in
Qk have vertices at adjacent points in the lattices (2−k Z)n . Moreover, the following properties
are also easily checked.
(1) For each k fixed, the cubes in Qk are (mutually) disjoint and their union is all of Rn .
(2) Given any two dyadic cubes, they are either disjoint, or one is contained in the other
one.
(3) Given j, k ∈ Z, with j < k, then each cube in Qk is contained in a unique cube in Qj
and contains 2k+1 cubes in Qk+1 .
The proof of Thm. 4.6 uses an important and far-reaching result, called the Calderón–
Zygmund decomposition, that here we state and prove in the setting of Rn .
Theorem 3.5. Let f ∈ L1 (Rn ) and non-negative. Given λ > 0, there exists a sequence {Qj }of
disjoint dyadic cubes such that
(i) f (x) ≤ λ for almost all x 63 ∪j Qj ;
1
(ii) ∪j Qj ≤ kf kL1 ;
Zλ
1
(iii) λ < f (x) dx ≤ 2n λ.
|Qj | Qj
32 M. M. PELOSO

Proof. Let λ > 0 be fixed. Since f ∈ L1 , there exists k0 ∈ Z so that

Z
1
f dx ≤ λ
|Q| Q
for all Q ∈ Qk0 .
Now we consider the cubes of the next generation, that are obtained from the cubes in Qk0
by bisecting each cube of the collection Qk0 . We say that a cube Q ∈ Qk0 +1 belongs to the
sequence we week if Z
1
f dx > λ
|Q| Q
while (by construction),
Z
1
f dx ≤ λ for (the unique) Q0 ∈ Qk0 such that Q0 ⊇ Qk0 .
|Q0 | Q0
1
R
If a cube is not chosen, that is, |Q| Q f dx ≤ λ, we iterate this process and we bisect it by
considering the dydiac cubes of the next generation that are contained in it. Then, we say that
a cube Qj ∈ Qj , j > k0 is in the sequence if
Z
1
f dx > λ ,
|Qj | Qj
while, Z
1
f dx ≤ λ for Q0 ∈ Qj−1 such that Q0 ⊇ Qj .
|Q0 | Q0
We need to check that the sequence {Qj } so constructed satisfies the required conditions. By
definition, if Qj ∈ Qj is a cube in the sequence and Q ∈ Qj−1 contains Qj then
Z Z
1 1
λ< f (x) dx ≤ −n f (x) dx ≤ 2n λ ,
|Qj | Qj 2 |Q| Q
so that (iii) is satisfied.
Next,
Z
X 1X
| ∪j Qj | = |Qj | < f (x) dx
λ Qj
j j
1
≤ kf kL1 ,
λ
so that (ii) also holds.
Finally, (i) follows from the Lebesgue differentiation theorem. If x ∈ c (∪j Qj ), then for all
dyadic cubes Q containing x then Z
1
f dx ≤ λ .
|Q| Q
Letting the side length of the cube containing x tend to 0, by the Lebesgue differentiation
theorem we obtain Z
1
f (x) = lim f dx ≤ λ ,
`(Q)→0 |Q| Q
for a.a. x ∈ c (∪j Qj ). This proves the theorem. 
 HARMONIC ANALYSIS 33

4. The Hilbert transform

Throughout this section we assume that n = 1 and consider the real line R. We are going
to introduce and study perhaps the most important example of an integral operator that is
Lp -bounded. By this we mean that there exists a (measurable) kernel K(x, y) defined a.e. on
R × R and such that for the integral operator
Z
T f (x) = f (y)K(x, y) dy
R
defined on a suitable dense class of functions of Lp (R) there exists a constant A = Ap depending
on p, but not on the given function f , such that
kT f kLp (R) ≤ Akf kLp (R) .

Next we introduce a temperated distribution of fundamental importance.

Definition 4.1. We define the principal value of 1/x as the tempered distribution
Z
1 ϕ(x)
p.v. (ϕ) = lim dx .
x ε→0 |x|≥ε x

Definition 4.2. We define the Poisson kernel Py (x), defined for x ∈ R and y > 0 as the kernel
1 y
Py (x) = .
π y 2 + x2
1
We remark that the kernel Py (x) can be viewed as the “dilated” of the L1 function π1 1+x 2 , so

that Py (x) is a so-called summability kernel.

For a given function f ∈ Lp (R), 1 ≤ p < ∞, we define the Poisson integral of f the function
Z Z
1 1 y
Py ∗ f (x) = f (x − t)Py (t) dt = f (x − t) 2 dt .
π R π R y + t2
Definition 4.3. We are also goint to consider the family of tempered distributions Qy , with
y > 0, given by
1 x
Qy (x) = .
π y 2 + x2
The kernel Qy (x) is called the conjugate Poisson kernel.
For a function f ∈ Lp (R), 1 ≤ p < ∞, we define the conjugate Poisson integral of f to be the
function Z Z
1 1 t
Qy ∗ f (x) = f (x − t)Qy (t) dt = f (x − t) 2 dt .
π R π R y + t2
We remark that Qy is not an L1 function, so that Qy (t) is not a summability kernel (and
also that here we are using an abuse of notation, since the subindex t does not represent the
dilation defined in (2.2)). Nonetheless, the convolution Qy ∗ f is well defined and continuous
(since instead Qy ∈ Lq , for 1 < q ≤ ∞).
Remark 4.4. The three objects above arose naturally when trying to study the regularity of
the mapping that associates to a given real harmonic function f on the upper half-plane (or on
the unit disk) its harmonic conjugate.
34 M. M. PELOSO

To briefly illustrate this circle of ideas, suppose that f is a real function that is in Lp (T),
where T is the unit circle, that is, the boundary of the unit disk in the complex plane (and T
is identified with the one-dimensional torus, see [Ka]).
Consider its Poisson integral P(f ), a real harmonic function on the unit disk, that we still
denote by f . Then f admits a harmonic conjugate function f˜. It can be proved that f˜ defines
a boundary function (still denoted by f˜).
It is important to notice that the Cauchy integral of f
Z
1 f (ζ)
C(f )(z) = dζ
2πi γ ζ − z
where the integral is the complex line integral over the boundary of the unit disk, produces a
holomorphic function F . Recall that the Poisson kernel (on the unit disk) satisfies the identity
P = C + C̄ − 1 ,
that is, P = Re (2C −1). Thus, the real part of 2C(F )−f (0) coincides with f , and its imaginary
part with f˜.
Hence, the function f˜ can be obtained as the integral of f against the imaginary part of
the Cauchy kernel, kernel that is called the conjugate Poisson kernel Q. Moreover, it can be
proved that, as the variable z approaces the boundary, f˜(z) = Q(f )(z) tends to a function on
the boundary, that we still denote by f˜.
The correspondence of the boundary functions f 7→ f˜ is the “Hilbert transform”.

We collect in the next statement the definition and main properties of the Hilbert transform.
Theorem 4.5. The following properties hold:
1
(i) p.v. (ϕ) ≤ C kϕ0 k∞ + kxϕk∞ = C %(0,1) (ϕ0 ) + %(1,0) (ϕ) ;



x
0 1 1
(ii) in S , lim Qt = p.v. ;
t→0 π x
1 1
(iii) F p.v. (ξ) = −i sgn(ξ).
π x
The following expressions are equivalent and define the Hilbert transform Hf of a function
f ∈ S:
1 1
(i’) Hf (x) = p.v. ∗ f (x);
π x
(ii’) Hf (x) = lim Qt ∗ f (x);
t→0
(iii’) F Hf (ξ) = −i sgn(ξ)fˆ(ξ).


R
Proof. Using the fact that ε≤|x|≤1 1/x dx = 0 we write
Z Z
1 ϕ(x) ϕ(x)
p.v. (ϕ) = lim dx + dx
x ε→0 ε≤|x|≤1 x 1≤|x| x
ϕ(x) − ϕ(0)
Z Z
ϕ(x)
= lim dx + dx
ε→0 ε≤|x|≤1 x 1≤|x| x
ϕ(x) − ϕ(0)
Z Z
ϕ(x)
= dx + dx .
|x|≤1 x 1≤|x| x
 HARMONIC ANALYSIS 35

Now (i) follows using the mean value theorem.

In order to prove (ii) consider the functions u(ε) (x) = x1 χ{|x|≥ε} . These are locally integrable
and bounded, hence they define tempered distributions. By definition of the principal value
distribution we have that
1
lim u(ε) = p.v. .
ε→0 x
Therefore, we wish to prove that
 1 
lim Qt − u(t) = 0
t→0 π
in S 0 .
Let ϕ ∈ S be fixed. We have
Z Z

xϕ(x) ϕ(x)
πQt − u(t) (ϕ) = dx − dx
R t 2 + x2 |x|≥t x
Z Z
xϕ(x)  x 1
= 2 2
dx − 2 2
− ϕ(x) dx
|x|<t t + x |x|≥t t + x x
Z Z
xϕ(tx) ϕ(tx)
= 2
dx − dx .
|x|<1 1 + x |x|≥1 x(1 + x2 )
Now we can pass to the limit as t → 0, by applying the dominated convergence theorem, by
majorizing the integrands using the fact that |vp(tx)| ≤ kϕkL∞ . The limit equals ϕ(0) times
Z Z
x 1
2
dx − 2
dx = 0 ,
|x|<1 1 + x |x|≥1 x(1 + x )

since the functions under the integral signs are odd and the region of integration symmetric with
respect to the origin.
Next we compute the Fourier transform of π1 p.v. x1 , and we do this using (ii) and the continuity
of F on S 0 ; that is, we use the equality
1 1
F p.v. = lim FQt .
π x t→0

We observe that11
  Z
−1 −2πt|ξ|
F − i sgn(ξ)e (x) = −i sgn(ξ)e−2πt|ξ| e2πixξ dξ
R
Z 0 Z +∞
2π(t+ix)ξ
=i e dξ − i e2π(−t+ix)ξ dξ
−∞ 0
i i
= +
2π(t + ix) 2π(−t + ix)
1 x
= = Qt (x) .
π t 2 + x2
11The Fourier transform of Q can be computed directly using the calculus of indefinite integrals in complex
t
analysis that relies on the residue theorem. Here we use the inverse Fourier transform to avoid using complex
analysis.
36 M. M. PELOSO

Therefore, we obtain that

1 1
F p.v. = lim FQt = lim −i sgn(ξ)e−2πt|ξ|
π x t→0 t→0
= −i sgn(ξ) ,

as we wish to show.12

We define the Hilbert transform Hf of a Schwartz function f as

1 1
Hf (x) = p.v. ∗ f (x) ,
π x
and observe that
Z
1 1
Hf (x) = lim u(ε) ∗ f (x) = lim f (x − y) χ{|y|≥ε (y) dy
ε→0 ε→0 π R y
f (x − y)
Z
1
= lim dy . (4.1)
ε→0 π |y|≥ε y

Finally, (i’)-(iii’) follow from the first part and Remark 2.24. 2

4.1. The Lp -boundedness of the Hilbert transform. The regularity result for the Hilbert
transform is the following.

Theorem 4.6. The Hilbert transform H, initially defined on Schwartz functions, can be extented
to an operator, still denoted by H, which is weak-type (1, 1) and strong-type (p, p) when 1 < p <
∞. More precisely, there exists C > 0 such that
 C
x ∈ R : |Hf (x)| > λ ≤ kf kL1 (R) ,
λ
and
kHf kLp (R) ≤ Ckf kLp (R)
when 1 < p < ∞.

Proof. We first observe that H is bounded on L2 , since by Plancherel’s theorem

kHf kL2 = k(Hf )ˆkL2 = k sgn(ξ)fˆkL2 = kf kL2 .

12Notice however that the limits above are in the S 0 -sense, that is, for every ϕ ∈ S,
Z Z
lim − i sgn(ξ)e−2πt|ξ| (ϕ) = lim −i sgn(ξ)e−2πt|ξ| ϕ(ξ) dξ =



−i sgn(ξ)ϕ(ξ) dξ = − i sgn(ξ) (ϕ).
t→0 t→0 R R
 HARMONIC ANALYSIS 37

Next, suppose that we have shown that H is weak-type (1, 1). Then, by Marcinkiewicz it
follows that H is strong -type (p, p), for 1 < p ≤ 2. If 2 < p < ∞ we use duality:
Z
kHf kLp = sup Hf (x)g(x) dx
kgk 0 ≤1 R
Lp
Z
≤ sup f (x)Hg(x) dx
kgk 0 ≤1 R
Lp

≤ sup kf (x)kLp kHgkLp0

kgk 0 ≤1
Lp

≤ Ckf kLp ,

that is, H is bounded on Lp also when 2 < p < ∞.

Hence, we have reduced ourselves to show the weak-type (1, 1) inequality. Fix λ > 0 and
we form the Calderón–Zygmund decomposition at height λ. We obtain a sequence of disjoint
interval {Ij } such that:
(i) f (x) ≤ λ for almost all x 63 ∪j Ij ;
1
(ii) ∪j Ij ≤ kf kL1 ;
λ
Z
1
(iii) λ < f (x) dx ≤ 2λ.
|Ij | Ij

Now we decompose f as sum f = g + b, where


f (x)Z if x 6∈ ∪j Ij
g(x) = 1
f (x) dx if x ∈ Ij ,
|Ij | Ij


and
Z
X X 1 
b(x) = bj (x) = f (x) − f χIj (x) .
|Ij | Ij
j j

(The decomposition f = g + b is what is called the Calderón–Zygmund decomposition of the

non-negative L1 function f .)
We observe that 0 ≤ g(x) ≤ λ, if x 6∈ ∪j Ij , while 0 ≤ g(x) = |I1j | Ij f (x) ≤ 2λ if x ∈ Ij .
R

Hence, g(x) ≤ 2λ, Rand also g ∈R L1 since it is obviously integrable on c (∪Rj Ij ) and
R on each Ij
(that are disjoint) Ij g(x)dx = Ij f (x)dx < ∞. Notice in particular that R g = R f . On the
other hand, b has mean equal to 0, since each bj does.
We now proceed. Since |Hf (x)| ≤ |Hg(x)| + |Hb(x)|, we have

{x : |Hf (x)| > λ} ≤ {x : |Hg(x)| > λ/2} + {x : |Hb(x)| > λ/2} .

38 M. M. PELOSO

Now, on the “good function” g we use the L2 -boundedness of the Hilbert transform to obtain
|Hg(x)|2
Z Z
{x : |Hg(x)| > λ/2} = dx ≤ 2
dx
{x:|Hg(x)|>λ/2} {x:|Hg(x)|>λ/2} (λ/2)
Z Z
4 4
≤ 2 |Hg(x)|2 dx = 2 g(x)2 dx
λ R λ R
Z Z
8 8
≤ g(x) dx = f (x) dx .
λ R λ R
Next, let Ij∗ denote the interval with the same center cj as Ij with twice the length. Then we
estimate,
{x : |Hb(x)| > λ/2} = {x ∈ (∪j Ij∗ ) : |Hb(x)| > λ/2} + {x ∈ c (∪j Ij∗ ) : |Hb(x)| > λ/2}
≤ | ∪j Ij∗ | + {x ∈ c (∪j Ij∗ ) : |Hb(x)| > λ/2}
Z
2 2
≤ kf kL1 + |Hb(x)| dx .
λ λ R\(∪j Ij∗ )
P
Since |Hb(x)| ≤ j |Hbj (x)| a.e., it will suffice to prove that
XZ
|Hbj (x)| dx ≤ Ckf kL1 . (4.2)
j R\Ij∗

Now, bj 6∈ S, nonetheless, for x 6∈ Ij∗ we have the formula

Z
bj (y)
Hbj (x) = dy .
Ij x −y
Recall that bj has mean 0 and that we denote by cj the center of the interval Ij . Notice that
y ∈ Ij implies that |y − cj | ≤ |Ij |/2 and that x 6∈ Ij∗ and y ∈ Ij imply that |x − y| > |x − cj |/2.
Then,
Z Z Z
bj (y)
|Hbj (x)| dx = dy dx
R\Ij∗ R\Ij∗ Ij x − y
Z Z  
1 1
= bj (y) − dy dx
R\Ij∗ Ij x − y x − cj
|y − cj |
Z Z
≤ |bj (y)| dx dy
Ij R\Ij ∗ |x − y| |x − cj |
|Ij |
Z Z
≤ |bj (y)| 2
dx dy
Ij R\Ij∗ |x − cj |
Z
=2 |bj (y)| dy .
Ij

Therefore,
XZ XZ
|Hbj (x)| dx ≤ 2 |bj (y)| dy ≤ 4kf kL1 ,
j R\Ij∗ j Ij

which gives (5.33). This concludes the proof. 

 HARMONIC ANALYSIS 39

4.2. Further properties of the Hilbert transform. We now make a few observations.
(1) We have shown that, when f ∈ S, then Hf satisfies a weak-type (1, 1) bound and a
strong-type (p, p) bound. We now extend the definition of Hf to all of Lp and all of L1 , with
the same bounds.
If f ∈ Lp , 1 < p < ∞, then there exists a sequence {fn } of Schwartz functions converging in
p
L to f :
lim kfn − f kLp = 0
n→+∞
Then, in order to define Hf , notice that kHfn − Hfm kLp ≤ Ckfn − fm kLp , so that {Hfn } is a
Cauchy sequence in Lp and converges to g ∈ Lp . We set then Hf = g. It is immediate to see
that this definition is well-given, in the sense that does not depend on the choice of the sequence
{fn } and that H then satisfies the bound kHf kLp ≤ Ckf kLp , with the same constant C as in
Thm. 4.6.
(2) If f ∈ L1 , then there exists a sequence {fn } of Schwartz functions converging in L1 to f .
The weak (1, 1) inequality gives that, for all ε > 0 fixed,
C
lim {x ∈ R : (Hfn − Hfm )(x) > ε} ≤ kfn − fm kL1 = 0 .
n,m,→+∞ ε
Then, {Hfn } is a Cauchy sequence in measure, so it converges to a measurable function g a.e..
We set, Hf = g. It is easy to chech that H satisfies the same weak-type (1, 1) bound on all of
L1 .
(3) When p = 1, the strong-type inequality fails. For instance, if we take f = χ[0,1] , then we
have13
1 1 1
Z Z
1 f (y)
Hf (x) = lim dy = lim dy
ε→0 π |y|≥ε x − y ε→0 π ε x − y

1 x−ε 1
Z
= lim dt
ε→0 π x−1 t

1 |x|
= log .
π |x − 1|
It is clear that Hf 6∈ L1 , while f ∈ L1 .
(4) Having defined the Hilbert transform of a function f ∈ Lp , for 1 ≤ p < ∞ as limit in norm
(when 1 < p < ∞) or in measure (when p = 1), we now wish to defined it pointwise as well.
In this part we will omit the proofs and we refer the reader to [Du]. Consider the truncated
integrals
f (x − y)
Z
1
Hε f (x) = dy .
π |y|≥ε y
Proposition 4.7. Let 1 < p < ∞. Then, if f ∈ Lp , Hf = limε→0 Hε f in the Lp -norm.
When p = 1 and f ∈ L1 , then Hf = limε→0 Hε f in measure.
Proof. Notice that the function y1 χ{|y|>ε} is in Lq for all 1 < q ≤ ∞. Then, the function
Z
1 1
χ (y)f (x − y) dy = Hε f (x) ,
π y {|y|>ε}
13Here we use the equatity at the end of the proof of Thm. 4.5.
40 M. M. PELOSO

is well defined for all f ∈ Lp , 1 ≤ p < ∞.

Now,
e−2πiyξ
1  Z
F χ{|y|>ε} (ξ) = lim dy
y N →+∞ ε<|y|<N y
−i sin(2πyξ)
Z
= lim dy
N →+∞ ε<|y|<N y
Z 2πN |ξ|
sin x
= −2i sgn(ξ) lim dx .
N →+∞ −2πε|ξ| x
Therefore, F y1 χ{|y|>ε} (ξ) ≤ 2π, for all ε > 0. This implies that Hε is of strong-type (2, 2),

with constants uniform in ε.

The weak-type (1, 1) boundedness also follows with uniform constant, so the strong-type (p, p)
follows, with uniform constant, by interpolation and duality.
We consider now the case 1 < p < ∞; the case p = 1 begin analogous. If f ∈ Lp , then let
{fn } be in S converging to f in Lp . Then, using the uniform bound for the Lp -boundedness of
Hε (and the fact that fn ∈ S), we have
Hf = lim Hfn = lim lim Hε fn = lim lim Hε fn = lim Hε f .
n→+∞ n→+∞ ε→0 ε→0 n→+∞ ε→0
This proves the lemma. 
We conclude this section by stating the following result concerning the pointwise definition
of the Hilbert transform.
Theorem 4.8. Let 1 ≤ p < ∞, and f ∈ Lp . Then
Hf (x) = lim Hε f (x) a.e.x ∈ R .
ε→0

We only mention that, similarly to the case of the Hardy–Littlewood maximal function and
the Lebesgue differentiation theorem, the proof relies on the boundedness of a maximal function.
Set
H ∗ f (x) = sup |Hε f (x)| .
ε>0
Then the following result holds true.
Theorem 4.9. The maximal Hilbert transform H ∗ is weak-type (1, 1) and strong-type (p, p) for
1 < p < ∞.
 HARMONIC ANALYSIS 41

5. Singular integrals
5.1. The Calderón–Zygmund theorem. Main result of this section is the following theorem.
Given a function K that is locally integrable in Rn \ {0}, and it is also a tempered distribution,
it makes sense to consider the convolution K ∗ f with a Schwartz function f . Therefore, we
define the operator
T f (x) = K ∗ f (x) ,
initially defined on Schwartz functions.
The next result generalizes the theorem on the bounededness of the Hilbert transform.
Theorem 5.1. Let K be a tempered distribution that coincides with a locally integrable function
on Rn \ {0}. Suppose that
|K̂(ξ)| ≤ A , (5.1)
and furthermore
Z
|K(x − y) − K(x)| dx ≤ B . (5.2)
|x|>2|y|

Then, the operator T is bounded from Lp into itself, i.e., it is of strong-type (p, p) for 1 < p < ∞
and it is of weak-type (1, 1), that is,
 C
x : |T f (x)| > λ ≤ kf kL1 .
λ
Definition 5.2. A tempered distribution K that satisfies the conditions in the theorem, that
is, that coincides with a locally integrable function on Rn \ {0} and satisfies (5.1) and (5.2) is
called a Calderón–Zygmund convolution kernel.
Corollary 5.3. Let K ∈ C 1 (Rn \ {0}) be a tempered distribution that coincides with a locally
integrable function on Rn \ {0}. Suppose that K sastisfies (5.1) and also
B0
|∇K(x)| ≤ . (5.3)
|x|n+1
Then, for the operator T the same conclusions hold as in the previous theorem.
Proof. In order to prove the corollary, assuming the validity of the theorem, it suffices to show
that (5.3) implies (5.2). But this follows from the mean value theorem. Let σ(x, y) denote the
segment having endpoints in x − y and x. Then, using the fact that |x| > 2|y|, it holds that
|x − y| ≥ |x| − |x| |x|
2 = 2 , so that, if z ∈ σ(x, y) we have

|x|
|z| ≥ min{|x − y|, |x|} ≥ .
2
Therefore,
1
|K(x − y) − K(x)| ≤ |y| sup |∇K(z)| ≤ B 0 |y| sup n+1
z∈σ(x,y) z∈σ(x,y) |z|
|y|
≤ 2n+1 B 0 .
|x|n+1
42 M. M. PELOSO

Hence,14
Z Z
1
|K(x − y) − K(x)| dx ≤ 2n+1 B 0 |y| dx
|x|>2|y| |x|>2|y| |x|n+1
Z +∞
= 2n+1 B 0 ωn |y| r−2 dr
2|y|
=B.
Thus, (5.3) implies (5.2) and the conclusion follows from Thm. 5.1. 

Proof of Thm. 5.1. The proof goes along the same lines of the proof of Thm. 4.6.
By the assumption (5.1) and the Plancherel theorem it immediately follows that T is bounded
on L2 : Z Z
kT f k2L2 = |F(K ∗ f )(ξ)|2 dξ = |K̂(ξ)fˆ(ξ)|2 dξ ≤ A2 kf k2L2 .
Rn Rn
We now observe that the adjoint operator T ∗ has kernel K ∗ (x) = K(−x), so that it also
satisfies hypotheses (5.1) and (5.2). Thus, if we prove the weak-(1, 1) inequality for T , by
interpolation it would follow that T is Lp -bounded for 1 < p ≤ 2, and then, by duality is
Lp -bounded also for 2 ≤ p < ∞.15
Hence, it suffices to prove the weak-(1, 1) inequality. To do this, we may assume that f ≥ 0
and, let λ > 0 be fixed. We decompose f according to Thm. 3.5, as in the proof of Thm. 4.6.
We repeat it for simplicity: we a sequence of disjoint cubes {Qj } such that:
(i) f (x) ≤ λ for almost all x 63 ∪j Qj ;
1
(ii) ∪j Qj ≤ kf kL1 ;
Zλ
1
(iii) λ < f (x) dx ≤ 2n λ.
|Qj | Qj
Now we decompose f as sum f = g + b, where

f (x) Z if x 6∈ ∪j Qj
g(x) = 1
f (x) dx if x ∈ Qj ,
|Qj | Qj


and Z
X X 1 
b(x) = bj (x) = f (x) − f χQj (x) .
|Qj | Qj
j j
Arguing as in the proof of Thm. 4.6, we have
{x : |T f (x)| > λ} ≤ {x : |T g(x)| > λ/2} + {x : |T b(x)| > λ/2} ,
and then Z
C
{x : |T g(x)| > λ/2} ≤ f (x) dx .
λ R

14Here, and in the remaining of the notes, we denote by ω the volume of the unit sphere S n−1 in Rn .
n
15This boundedness is equivalent to the Lp -boundedness of T ∗ for 1 < p ≤ 2.
 HARMONIC ANALYSIS 43

Next,
Z
C 2
{x : |T b(x)| > λ/2} ≤ kf kL1 + |T b(x)| dx .
λ λ R\(∪j Q∗j )
√
where Q∗j denote the cube with the same center cj as Qj with side length 2 n-times longer.
Then we reduce ourselves to show that
XZ
|T bj (x)| dx ≤ Ckf kL1 ,
j R\Q∗j

which, in turn, is implied by the inequality

Z
|T bj (x)| dx ≤ Ckbj kL1 . (5.4)
R\Q∗j

We are going to use the Hörmander condition (5.2) and the fact that bj has integral equal to
0. For x 6∈ Q∗j ,
Z Z
 
T bj (x) = K(x − y)bj (y) dy = K(x − y) − K(x − cj ) bj (y) dy .
Qj Qj

Therefore,
Z Z Z
|T bj (x)| dx ≤ |bj (y)| K(x − y) − K(x − cj ) dx dy
Rn \Q∗j Qj Rn \Q∗j
Z
≤B |bj (y)| dy ,
Qj

since
Rn \ Q∗j ⊆ x ∈ Rn : |x − cj | > 2|y − cj | ,


so that
Z Z
K(x − y) − K(x − cj ) dx ≤ K(x − y) − K(x − cj ) dx
Rn \Q∗j |x−cj |>2|y−cj |
Z
= K(x0 − (y − cj )) − K(x0 ) dx0
|x0 |>2|y−c j|

≤B.
Thus, (5.4) follows and we are done. 

5.2. Homogeneous distributions. In order to describe some classical and fundamental exam-
ples of singular integrals in Rn , examples that generalize the case of the Hilbert transform in the
case of the real line, we need a preliminary discussion of the so-called homogenous distributions.

We recall that a function f is said to be homogenous of degree a if for all λ > 0 and all x ∈ Rn
f (λx) = λa f (x) .
44 M. M. PELOSO

Given another function ϕ, we see that

Z Z Z
f (x)ϕλ (x) dx = λ−n f (x)ϕ(x/λ) dx = f (λx0 )ϕ(x0 ) dx0
Rn n n
Z R R
a 0 0 0
=λ f (x )ϕ(x ) dx .
Rn
Therefore, we may extend the definition of homogeneity to tempered distributions as follows.
Definition 5.4. Given a tempered distribution u, we say that u is homogenous of degree a > 0
if for all ϕ ∈ S and λ > 0 we have
u(ϕλ ) = λa u(ϕ) .
Example 5.5. It is immediate to see that the distribution π1 p.v. x1 on S(R) is homogeous of
degree −1, while m(ξ) = −i sgn(ξ) is a homogeneous (function) of degree 0.
We now wish to generalize the tempered distribution π1 p.v. x1 and the Hilbert transform to
higher dimensions. A typical situation will be the following.
Let Ω(x) be a function in Rn \ {0} homogeneous of degree 0, that is, such that Ω(λx) = Ω(x)
for all λ > 0 and x ∈ Rn \ {0}. Notice that Ω, as all homogeneous functions, is uniquely
determined by its values on the unit sphere S n−1 . Suppose that Ω ∈ L1 (S n−1 ) and
Z
Ω(x0 ) dσ(x0 ) = 0 ,
S n−1
where dσ denotes the Lebesgue surface measure on S n−1 . We define the tempered distribution
Z
Ω(x)
u(ϕ) = p.v. n
ϕ(x) dx
R |x|
n
Z +∞ Z
= p.v. Ω(x0 )ϕ(rx0 ) dσ(x0 )r−1 dr
0 S n−1
Z +∞ Z
Ω(x0 ) ϕ(rx0 ) − ϕ(0) dσ(x0 )r−1 dr ,


=
0 S n−1
where we have used the fact that Ω has zero integral on the unit sphere. Now, using the mean
value theorem, as in the case of π1 p.v. x1 , it is easy to see that the last integral above converges
absolutely, for ϕ ∈ S(Rn ).
Lemma 5.6. If u ∈ S 0 is a tempered distribution homogeneous of degree a, then its Fourier
transform is a tempered distribution homogeneous of degree −n − a.
Proof. Recall the identities (2.5). For ϕ ∈ S and λ > 0 we have

λ

= λ−n u ϕ̂ λ−1




û(ϕλ ) = u ϕ
cλ = u ϕ̂
= λ−n−a u ϕ̂ = λ−n−a û(ϕ) ,

as we wished to show. 2

Lemma 5.7. Let 0 < a < n. Then the function |x|−a is locally integrable and defines a tempered
distribution. The Fourier transform, as an element of S 0 satisfies the equality
|x|−a b(ξ) = cn,a |ξ|a−n ,


 HARMONIC ANALYSIS 45

where
π a−n/2 Γ n−a


cn,a =
2 . (5.5)
Γ a2

Proof. We recall that the Fourier transform of a radial function is also radial.16 Suppose first
that n/2 < a < n, so that

|x|−a = |x|−a χ{|x|≤1} + |x|−a χ{|x|>1} = f1 (x) + f2 (x) ,

where f1 ∈ L1 and f2 ∈ L2 . Then we can compute the Fourier transform of |x|−a as a function.
Its Fourier transform is a radial function, and by the previous lemma, homogeneous of degree
−n + a. Hence, it is a constant multiple of |ξ|a−n , i.e.,

|x|−a b(ξ) = cn,a |ξ|a−n .

We now compute the constant cn,a . Using the Parseval formula (2.6) and the identity
2
2
e−π|x| ˆ(ξ) = e−π|ξ| of Lemma 2.13 we have
Z Z
−π|x|2 −a 2
e |x| dx = cn,a e−π|ξ| |ξ|a−n dξ . (5.6)
Rn Rn

Now, setting πr2 = s we see that

Z +∞ Z +∞
−πr2 b 1  s b/2 1
e r dr = √ e−s √ ds
0 2 π 0 π s
Z +∞
1
= (1+b)/2 e−s s(b−1)/2 ds
2π 0
1

= (1+b)/2 Γ (1 + b)/2 .
2π
From (5.6) it then follows
1
h 1
i−1
cn,a = Γ (n − a)/2 Γ a/2
2π (n−a)/2 2π a/2
π a−n/2 Γ n−a


2
= .
Γ a2

This proves the proposition in the case n/2 < a < n.

In the case 0 < a < n/2 we use the first part, since in this case n/2 < n − a < n, and the
inversion formula, valid also in S 0 . The case a = n/2 follows by passing to the limit, and using
the continuity of the Fourier transform in S 0 . 2

16This is immediate to check, using the fact that a funciton f is radial if and only if f (Ox) = f (x) for all
x ∈ Rn and all orthogonal transformations O.
46 M. M. PELOSO

5.3. The Riesz transforms. We now define the main generalizations of the Hilbert transform,
the Riesz transforms Rj , j = 1, . . . , n.
Given f ∈ S, we set
Z
yj
Rj f (x) = cn p.v. n+1
f (x − y) dy , (5.7)
Rn |y|

where the constant cn is given by17

Γ n+1


2
cn = (n+1)/2 .
π
Notice that for n = 1 we have c1 = 1/π and we recover the definition of the Hilbert transform.
Lemma 5.8. For f ∈ L2 (Rn ) and j = 1, . . . , n we have that
ξj ˆ
(Rj f )ˆ(ξ) = −i f (ξ) .
|ξ|
Hence,
n
X
Rj2 = −I .
j=1

Proof. We wish to show that

 yj  ξj
F cn p.v. n+1 (ξ) = −i ,
|y| |ξ|
as tempered distributions.
We claim that, again as tempered distributions,
xj
∂xj |x|−n+1 = (1 − n) p.v. , (5.8)
|x|n+1
Assuming the claim we finish the proof. Using Lemma 5.7 and Remark 2.24 (V) we have
 xj  1  
F p.v. n+1 (ξ) = F ∂xj |x|−n+1 (ξ)
|x| 1−n
2πiξj
F |x|−n+1 (ξ)


=
1−n
n
2πiξj π 2 −1 Γ( 21 ) −1
= |ξ|
1 − n Γ( n−1 2 )
π (n+1)/2 ξj
= −i

Γ (n + 1)/2 |ξ|
i ξj
=− ,
cn |ξ|
as we wish to prove.

17We recall that the volume of the unit ball B and of the unit sphere S n−1 in Rn are, resp., |B | =
n n
n/2
/Γ (n/2) + 1 and σ(S n−1 ) = 2π n/2 /Γ(n/2), resp.


π
 HARMONIC ANALYSIS 47

It only remains to prove the claim. For ϕ ∈ S we have

Z
xj xj
(1 − n) p.v. n+1 (ϕ) = lim (1 − n) 2 2 (n+1)/2
ϕ(x) dx
|x| ε→0 +
Rn (ε + |x| )
Z
1
= lim ∂xj 2 ϕ(x) dx
ε→0+ Rn (ε + |x|2 )(n−1)/2
Z
1
= − lim 2
∂ ϕ(x) dx
2 (n−1)/2 xj
Rn (ε + |x| )
ε→0 +
Z
1
=− ∂ ϕ(x) dx
n−1 xj
Rn |x|
= ∂xj |x|1−n (ϕ) .

This proves the claims and therefore the lemma. 2

Theorem 5.9. The Riesz transforms Rj , j = 1, . . . , n, are of weak-type (1, 1) and of strong-type
(p, p), for 1 < p < ∞.
Proof. It suffices to show that the the tempered distributions Kj (x) = cn p.v. xj /|x|n+1 satisfy


the hypotheses of Thm. 5.1, for j = 1, . . . , n, resp.
Clearly Kj coincides with a locally integrable function on Rn \ {0}. Moreover, by the lemma,
ξj
K̂j (ξ) = −i ,
|ξ|
which is clearly bounded.
Finally, on Rn \ {0}
1 n + 1 |xj ||x|
|∇Kj (x)| ≤ cn n+1
+ cn
|x| 2 |x|n+3
1
≤ C n+1 .
|x|
Thus, (5.2) is satisfied and we are done. 2

5.4. Solution of the Laplace equation. In this section we consider the Laplace operator
Xn
∆= ∂x2j , (5.9)
j=1

and prove a regularity result for the solution of the equation

∆g = f .
Given a partial differential operator P with constant coefficients, we say that a tempered
distribution E is a fundamental solution for P if P E = δ0 , the Dirac delta at the origin, that is,
if
P (f ∗ E) = f ∗ P E = f ∗ δ0 = f ,
for all f ∈ S. Notice in particular that, for a given f ∈ S, the tempered distribution f ∗ E solves
the equation P g = f .
We now compute the fundamental solution for ∆.
48 M. M. PELOSO

Proposition 5.10. Let ωn = 2π n/2 /Γ n/2 denote the volume of the unit sphere in Rn and


define
1 1
E(x) = for n > 2 ,
(2 − n)ωn |x|n−2
and
1
E(x) = log |x| for n = 2 .
2π
Then E is a fundamental solution for ∆.
Proof. We consider a family of regularized tempered distributions, namely E ε , given by
1 1
E ε (x) = for n > 2 ,
(2 − n)ωn (ε2 + |x|2 )(n−2)/2
and
1
E ε (x) = log(ε2 + |x|2 ) for n = 2 .
4π
It is easy to see that E ε → E pointwise as ε → 0, and that there exists g ∈ L1loc such that
ε
|E | ≤ g for all ε ≤ 1, in both cases n = 2 and n > 2. (It suffices to take |E| when n > 2 and
log |x| + 1 when n = 2.) Hence, E ε → E in S 0 and therefore, also ∆E ε → ∆E in S 0 .
Let ϕ ∈ S, we wish to show that ∆E ε (ϕ) → ϕ(0) as ε → 0. We provide the details for n > 2.
We have
1 1
∂xj E ε (x) = · (2 − n)xj
(2 − n)ωn (|x| + ε2 )n/2
2

nx2j
 
2 ε 1 1
∂xj E (x) = − ,
ωn (|x|2 + ε2 )n/2 (|x|2 + ε2 )(n+2)/2
so that
n 
nx2j

ε 1 X 1
∆E (x) = −
ωn
j=1
(|x|2 + ε2 )n/2 (|x|2 + ε2 )(n+2)/2
n ε2
= .
ωn (|x|2 + ε2 )(n+2)/2
Therefore,
∆E ε (x) = ψε (x) ,
where, as usual, ψε (x) = ε−n ψ(x/ε) and
n 1
ψ = ∆E 1 = .
ωn (|x| + 1)(n+2)/2
2

Next we use the fact that ψ(x) = ψ(−x) to see that

Z
∆E ε (ϕ) = ∆E ε (−x)ϕ(x) dx = ψε ∗ ϕ(0) → ( ψ)ϕ(0) ,
R

by the properities of the summability R kernels.

Thus, we only need to compute ψ. But, by passing to polar coordinates and making the
change of variables s = r2 /(1 + r2 ), so that ds = 2(1 + r2 )−2 rdr, and
rn−1  r2 (n−2)/2 r
= 2
.
2
(1 + r )(n+2)/2 1+r (1 + r2 )2
 HARMONIC ANALYSIS 49

We then find
Z Z
n 1
ψ(x) dx = dx
ωn (|x|2 + 1)(n+2)/2
Z +∞
1
=n 2 (n+2)/2
rn−1 dr
0 (1 + r )
Z 1
n
= s(n−2)/2 ds = 1 .
2 0
This completes the proof (at least in the case n > 2, the case n = 2 being left to the reader). 2
We now have the regularity result for the solutions of the Laplace equation.
Theorem 5.11. Let f ∈ S be given. Define the operator
T f (x) = ∂x2j xk f ∗ E = f ∗ ∂x2j xk E ,
j, k = 1, . . . , n. Then, T extends to an operator that is of weak-type (1, 1) and of strong-type
(p, p), for 1 < p < ∞.
Remark 5.12. Notice that the result shows that, for a given f ∈ Lp , it is possible to define the
solution u = f ∗ E of the equation ∆u = f is such a way that the map T (f ) = ∂x2j xk (f ∗ E) is
bounded on Lp .
On the other hand, the map I defined as f 7→ u = f ∗ E is bounded, when n ≥ 3,
I : Lp → Lq
as a consequence of Thm. 5.19 of the next section, where
1 1 2
= − .
q p n
Proof of Thm. 5.11. It suffices to show that the kernel K = ∂x2j xk E satisfies the hypotheses of
Thm. 5.1. It is clear that, by Lemma 5.7,
F ∂x2j xk E (ξ) = −(2π)2 ξj ξk Ê(ξ)

= Cn ξj ξk |ξ|−2 ,
so that (5.1) is satisfied.
On the other hand, also the condition (5.3) is also easily seen to be satisfied: if x 6= 0,
1
∇ ∂x2j xk E (x) ≤ C n+1 ,


|x|
The result now follows from 5.1. 2

We now consider the Dirichlet problem for the Laplacian on the half-space
(
∂t2 f (x, t) + ∆u(x, t) = 0 for (x, t) ∈ Rn+1 , t > 0
(5.10)
u(x, 0) = f (x) for x ∈ Rn .
Here ∆ = ∆x = nj=1 ∂x2J , so that ∂t2 + ∆x is the Laplacian in Rn+1 . Moreover,
P

Rn+1 := (x, t) ∈ Rn × R : t > 0 .


+
50 M. M. PELOSO

We assume that the boundary data g ∈ Lp (Rn ) and we look for a solution f defined on Rn+1
such that f (·, t) ∈ S 0 (Rn ) for all fixed t > 0. In this case, we can compute the Fourier transform
w.r.t. the variable x ∈ Rn . We have that
F (∂t2 + ∆x )f (·, t) (ξ) = ∂t2 fˆ(ξ, t) + F ∆x )f (·, t) (ξ) = ∂t2 fb(ξ, t) − |2πξ|2 fb(ξ, t) .

For the time being we consider ξ as a fixed parameter and write ϕ = fb(t, ·). Then the Dirichlet
problem (5.10) has been transformed into the Cauchy initial value problem
(
ϕ00 (t) − a2 ϕ(t) = 0
ϕ(0) = b ,

where a = |2πξ| and b = gb(ξ). Since the characteristic equation has roots λ± = ±|2πξ|, the
solutions of the ODE are
ϕ(t) = c+ e|2πξ|t + c− e−|2πξ|t .
Then, the solution of (5.10) has to satisfy

F −1 c+ e|2πξ|t + c0 e−|2πξ|t (x)

and the only possibility for the (inverse) Fourier transform to make sense is that c+ e|2πξ|t +
c0 e−|2πξ|t ∈ S 0 (Rn ). This forces c+ = 0 so that the solution of the Cauchy problem satifsies the
initial condition
ϕ(0) = c− = b = gb(ξ) .
Therefore,

f (x, t) = F −1 gbe−|2π·|t (x)

= g ∗ F −1 (e−|2π·|t ) (x) .

Thus, we need to compute F −1 (e−|2π·|t ) (x). Observe that by Remark 2.14 we have that

F −1 (e−|2π·|t ) (x) = Pt (x), so it suffices to determine Ψ such that Pb(x) = e−|2πξ .

We define the Poisson kernel for the half-space
t
Pt (x) = cn
n+1 (5.11)
|x|2 + t2 2

Γ( n+1 )
where ωn = 2
n+1 . We have the following
π 2

1
Proposition 5.13. The function P (x) = ωn
n+1 ∈ L1 (Rn ) and
2
1+|x|2

Pb(ξ) = e−|2πξ| .
For the proof, see Exercise 9, Appendix B.18

18Check the constants c and ω .

n n
 HARMONIC ANALYSIS 51

5.5. The heat operator. We now turn our attention to the heat operator in Rn+1 ∂t − ∆. We
begin by considering the Cauchy problem in the upper-half space Rn × (0, ∞), where we denote
the variables (x, t) ∈ Rn × R, called the Cauchy problem for the heat equation,
(
∂t u(x, t) − ∆u(x, t) = 0 for (x, t) ∈ Rn+1 , t > 0
(5.12)
u(x, 0) = f (x) for x ∈ Rn .

We assume that f ∈ S 0 and that the solution u is a tempered distribution in x, for any t
fixed. Then, by taking the Fourier transform in the x-variable, we obtain the ordinary differential
equation Cauchy problem (for any ξ ∈ Rn fixed)
(
∂t û(ξ, t) + 4π 2 |ξ|2 û(ξ, t) = 0 for t > 0
û(ξ, 0) = fˆ(ξ) .

We can easily solve this Cauchy problem and find that the solution is
2 2
û(ξ, t) = fˆ(ξ)e−4π |ξ| t for t > 0 .
Therefore, by undoing the Fourier transform we obtain
u(x, t) = f ∗ K√t (x) ,
where
2 |ξ|2 t 1 2
K√t (x) = K(x, t) = F −1 e−4π e−|x| /4t ,


= n/2
(4πt)
by Prop. 2.13. Notice that
K√t (x) = t−n/2 K1 (x/t1/2 ) ,
so that K√t dx = K1 dx = 1, for all t > 0.19
R R

We now compute a fundamental solution for the heat operator. We define K on all of Rn+1
by setting
( 2
(4πt)−n/2 e−|x| /4t for t > 0
K(x, t) = (5.13)
0 for t ≤ 0 .
Since Rn K(x, t) dx = 1 for t > 0, K(x, t) is integrable on every region of the form Rn ×(−∞, T ].
R

Proposition 5.14. The function defined by (5.13) is a fundamental solution for the heat oper-
ator.
Proof. Let ε > 0 be given. Define K(ε) (x, t) = K(x, t) for t > ε and K(ε) (x, t) = 0 for t ≤ ε.
Then K(ε) → K in the sense of distributions in Rn+1 as ε → 0. Then, we wish to show that
h(∂t K(ε) − ∆x K(ε) ), ϕi = hK(ε) , (−∂t − ∆x )(ϕ)i → ϕ(0, 0)

as ε → 0, for every ϕ ∈ D(Rn+1 ).

19The notation K√ is not standard, as this kernel is classically defined as K . We have elected to use the
t t
former notation for consistency.
52 M. M. PELOSO

Now, by integrating by parts we find that

Z +∞ Z
hK(ε) , (−∂t − ∆x )(ϕ)i = K(x, t)(−∂t − ∆x )(ϕ)(x, t) dxdt
ε Rn
Z +∞ Z Z Z +∞
= −∆x K(x, t)ϕ(x, t) dxdt + ∂t K(x, t)ϕ)(x, t) dtdx
ε Rn Rn ε
Z
+ K(x, ε)ϕ(x, ε) dx
Rn
Z +∞ Z Z
= (∂t − ∆x )K(x, t)ϕ(x, t) dxdt + K(x, ε)ϕ(x, ε) dx
ε Rn Rn
Z
=0+ K(−x, ε)ϕ(x, ε) dx
Rn
since K is even in x and it is annihilated by the heat operator on Rn × (ε, +∞). But, the right
hand side above is (the convolution being in Rn )
Z
K(−x, ε)ϕ(x, ε) dx = K(ε) ∗ ϕ(·, ε)(0)
Rn
 
= K(ε) ∗ ϕ(·, 0)(0) + K(ε) ∗ ϕ(·, ε) − K(ε) ∗ ϕ(·, 0) (0) .
The former term tends to ϕ(0, 0) as ε → 0, while the latter one is bounded by
sup |ϕ(x, ε) − ϕ(x, 0)|kK(ε) k1 = sup |ϕ(x, ε) − ϕ(x, 0)|
x∈Rn x∈Rn

which tends to 0 as ε → 0. 2

We conclude this part by solving the problem (5.12).

Theorem 5.15. Let 1 ≤ p < +∞ and g ∈ Lp (Rn ) be given. Then the boundary value problem
(5.12) admits a solution f (x, t) such that f (·, t) → g in Lp (Rn ) as t → 0+ .
Proof. Setting f = g ∗ K√t we see that f solves the equation in Rn+1
+ and f (·, t) = g ∗ K√t → g
p n +
in L (R ) as t → 0 . 

5.6. The Riesz potentials. We now discuss some operators Iα , called the Riesz potentials, that
arise in a natural way in connection with the Laplacian. The operators Iα are integral operators
whose kernel is homogenous of degree −(n − α), then locally integrable when 0 < α < n, case
to which we restrict ourselves.
Definition 5.16. Let 0 < α < n and define the Riesz potential of order α the integral operator
Z
1 1 1
Iα (f )(x) = f (y) dy ,
(2π) cn,n−α Rn |x − y|n−α
α

where cn,n−α is the constant defined in (5.5). (It will be sometimes convenient to set a = n − α.
Notice that 0 < a < n exactly when 0 < α < n.)
We remark that the Laplacian satisfies the identity
(−∆f )b(ξ) = (2π|ξ|)2 fˆ(ξ) ,
 HARMONIC ANALYSIS 53

for f ∈ S. By Lemma 5.7, it follows that

1 1
|x|−(n−α) ∗ f b(ξ)


(Iα f )b(ξ) = α
(2π) cn,n−α
= (2π|ξ|)−α fˆ(ξ) .
Hence, it is natural to define the fractional powers of −∆ by setting
 
−α/2
(−∆) f b(ξ) = (2π|ξ|)−α fˆ(ξ) , (5.14)
that is, by setting
(−∆)−α/2 = Iα .
We now study the (Lp , Lq )-boundedness of the operators Iα . We begin with a simple result
that will give us a necessary condition on p, q and α. We set Dλ f (x) = f λ (x) = f (λx), for
λ > 0.
Lemma 5.17. For λ > 0 we have the follwing identities:
(i) Dλ−1 Iα Dλ = λ−α Iα ;
(ii) kDλ f kLp = λ−n/p kf kLp ;
(iii) kDλ−1 Iα (f )kLq = λn/q kIα (f )kLq .
Proof. These are elementary. For a given function f , sufficiently regular, we have
(2π)α
Z
1
Dλ−1 Iα Dλ (f )(x) = −1
f (λy) dy
cn,n−α Rn |λ x − y|n−α
(2π)α 1
Z
1
= −1
f (z) dz
cn,n−α λ Rn |λ (x − z)|n−α
n

= λ−α Iα (f )(x) ,
as we claimed.
(ii) and (iii) are simply a change of variables. 2

Lemma 5.18. Let 0 < α < n, 1 < p, q < ∞ be given. Then, if Iα : Lp (Rn ) → Lq (Rn ) is
bounded, then necessarily
1 1 α
= − .
q p n
Proof. If Iα is bounded, there exists a constant C > 0 such that
kIα (f )kLq ≤ Ckf kLp
for all f ∈ Lp . Then, for all λ > 0 it must hold that
kIα (Dλ f )kLq ≤ CkDλ f kLp . (5.15)
Now, by Lemma 5.17,
kIα (Dλ f )kLq = kDλ Dλ−1 Iα (Dλ f )kLq = λ−n/q λ−α kIα (f )kLq ,
Then, (5.15) is equivalent to
n
+α− n
kIα (f )kLq ≤ Cλ q p kf kLp .
54 M. M. PELOSO

Since this must hold for all λ > 0, if the exponent of λ in the inquality above is 6= 0, by letting
λ → 0, or λ → +∞ we get a contradiction. The conclusion now follows. 2
We now prove the positive result, valid in the case 1 < p, q < ∞.
Theorem 5.19. Let 0 < α < n, 1 ≤ p, q < ∞ satisfy
1 1 α
= − , (5.16)
q p n
and consider the fractional integral Iα . For f ∈ Lp , the integral defining Iα (f ) converges abso-
lutely. Moreover,
(i) if 1 < p, q < ∞
Iα : Lp (Rn ) → Lq (Rn )
is strong-type (p, q);
(ii) if p = 1, then Iα is weak-type (1, q).
Proof. Clearly, it suffices to consider the convolution operator T : f 7→ K ∗ f , where K(x) =
|x|−n+α and prove (i) and (ii) for this operator.
By the general form of the Marcinkiewicz interpolation theorem, if we show that T is weak-
type (p1 , q1 ) and also weak-type (1, q0 ), both pairs satisfying the condition (5.16), then it follows
that it is strong-type (p, q), where
1 θ 1 θ 1−θ
= +1−θ, and = + .
p p1 q q1 q0
Notice that
1 1 α  α
=θ − + (1 − θ) 1 −
q p1 n n
θ α 1 α
= +1−θ− = − .
p1 n p n
We write K = K1 + K∞ , where
K1 (x) = K(x)χ{|x|≤R} , and K∞ (x) = K(x)χ{|x|>R} ,
and where R > 0 is a positive constant to be selected later.
We begin by observing that K1 ∈ L1 (Rn ), since
Z Z Z R
|K1 (x)| dx = |x|−n+α dx = ωn rα−1 dr
Rn |x|≤R 0
ωn α
= R . (5.17)
α
0
On the other hand, K∞ ∈ Lp , where p0 is the exponent conjugate to p. For,
Z Z
0 0
|K∞ (x)|p dx = |x|(−n+α)p dx
Rn |x|>R
Z +∞
0 0
= ωn rn(1−p )+αp −1 dr
R
ωn 0 0
= Rn(1−p )+αp , (5.18)
n(p0 − 1) − αp0
 HARMONIC ANALYSIS 55

and the integral is convergent since n(1 − p0 ) + αp0 < 0, that is, − np + α < 0, which is equivalent
to q < ∞.
Now, let f ∈ Lp , 1 ≤ p < ∞. Then, the integral defining K ∗ f converges absolutely since
0
K1 ∈ L1 and K∞ ∈ Lp . (Make sure you agree with this assertion.) This proves the first part of
the statement.
In order to prove (i) and (ii), it suffices to show that, for 1 ≤ p < q < ∞ such that 1/q =
1/p − α/n, T is weak-type (p, q), that is,
 kf k p q
 L
x : |K ∗ f (x)| > λ ≤ C . (5.19)
λ
Notice that we may assume that kf kLp = 1 (as it is easy to check, and we invite you to do
so).
Then we have,
  
x : |K ∗ f (x)| > λ ≤ x : |K1 ∗ f (x)| > λ/2 + x : |K∞ ∗ f (x)| > λ/2 .

Now, using (5.17) and the assumption kf kLp = 1, we have

|K1 ∗ f (x)| p
Z Z 

x : |K1 ∗ f (x)| > λ/2 = dx ≤ dx
{x:|K1 ∗f (x)|>λ/2} λ/2
2p kK1 ∗ f kpLp 2p kK1 kpL1 kf kpLp
= ≤
λp λp
 R α p
≤C . (5.20)
λ
Next, using (5.18) we see that

kK∞ ∗ f kL∞ ≤ kK∞ kLp0 kf kLp = kK∞ kLp0 ≤ CRα−n/p . (5.21)

We now choose R such that

λ
CRα−n/p = , that is, R = C 0 λ1/(α−n/p) = C 0 λ−q/n .
2
With choice of R, by (5.21) it follows that

x : |K∞ ∗ f (x)| > λ/2 = 0.

Therefore, by (5.21) (recalling that kf kLp = 1),

 R α p
p
≤ C λ−q(α/n+1/q)

x : |K ∗ f (x)| > λ ≤C
λ
 kf k p q
L
= Cλ−q = C .
λ
This concludes the proof of the theorem. 2
56 M. M. PELOSO

5.7. †Construction of Calderón–Zygmund kernels. In order to prove the Mihlin-Hörmander

multiplier theorem, Theorem 6.24 that follows, we describe a method, that is quite standard, to
construct Calderón–Zygmund convolution kernels, as defined in Def. 5.2.

Definition 5.20. An Lp -function f is said to satisfy the Lp -Lipschitz condition of order α,

α ∈ (0, 1), if there exists a constant C > 0 such that
Z 1/p
|f (x − h) − f (x)|p dx ≤ C|h|α
Rn
for every h ∈ Rn . In this case we set
Z 1/p
−α
kf kΛpα := kf kLp (Rn ) + sup |h| |f (x − h) − f (x)|p dx .
h6=0 Rn

Remark 5.21. The following facts are easy to prove.20

(1) An Lp -function f is in Λpα if and only if for every λ > 0 we have
Z 1/p
−α
sup |h| |f (x − h) − f (x)|p dx < ∞;
0<|h|<λ Rn

(2) let ϕ ∈ C0∞ be identically 1 in a nghb of the origin and set f = |x|α−n/p ϕ, then f ∈ Λpα .

Lemma 5.22. Let f ∈ Λ1α . Then, there exists C > 0 such that
|fˆ(ξ)| ≤ C(1 + |ξ|)−α kf kΛ1 . α

Proof. Notice that

Z Z
−πi −2πixξ 2
ˆ
−f (ξ) = e f (x)e dx = f (x)e−2πi(x+ξ/(2|ξ| ))ξ dx
Rn Rn
Z
f x − ξ/(2|ξ|2 ) e−2πixξ dx .


=
Rn
Therefore,
Z h
i
|fˆ(ξ)| = 1
2 f (x) − f x − ξ/(2|ξ|2 ) e−2πixξ dx
Rn
Z
1
f (x) − f x − ξ/(2|ξ|2 ) dx


≤ 2
Rn
≤ C|ξ|−α kf kΛ1α .
On the other hand
|fˆ(ξ)| ≤ kf kL1 ≤ kf kΛ1α ,
so that
(1 + |ξ|)α |fˆ(ξ)| ≤ C(1 + |ξ|α )|fˆ(ξ)| ≤ Ckf kΛ1α .
This proves the lemma. 2
We will need a smooth decomposition of the function identically 1.
Lemma 5.23. There exists a function ϕ ∈ C0∞ such that
20Exercise.
 HARMONIC ANALYSIS 57
X
(i) ϕ(2−j ξ) = 1 for all ξ 6= 0 .
j∈Z
Moreover, there exists another function ϕ0 ∈ C0∞ such that
+∞
X
(ii) ϕ0 (ξ) + ϕ(2−j ξ) = 1 for all ξ .
j=1

Proof. Let ϕ0 ∈ C0∞ (Rn ) such that ϕ ≥ 0, ϕ0 (ξ) = 1 if |ξ| ≤ 1 and ϕ0 (ξ) = 0 if |ξ| ≥ 2. (Such
function exists by the C ∞ -Urysohn’s lemma, Lemma 2.7). We set
ϕ(ξ) = ϕ0 (ξ) − ϕ0 (2ξ) .
Notice that ϕ(ξ) = 0 if |ξ| ≤ 1/2 or |ξ| ≥ 2, that is,
supp ϕ ⊆ {ξ : 1/2 ≤ |ξ| ≤ 2} .
By induction, it is elementary to prove that, for ` ≥ 1 and all ξ,
`
X
ϕ0 (ξ) + ϕ(2−j ξ) = ϕ0 (2−` ξ) . (5.22)
j=1

Letting ` → +∞ we obtain (ii).

Next, again by induction on `0 , with `0 ≤ 0, it is easy to show that
`
0
X
ϕ(2−j ξ) = ϕ0 (2−` ξ) − ϕ0 (2−` +1 ξ) . (5.23)
j=`0
0
Letting ` → +∞ and `0 → −∞ we obtain (i). (Notice that ϕ0 (2−` ξ) − ϕ0 (2−` +1 ξ) → 1 only
when ξ 6= 0.) 2

Corollary 5.24. Let ϕ0 , ϕ be as in Lemma 5.23, and set

, ψ(j) = F ϕ(2−j ·) = Fϕ 2−j , for j = 1, 2, . . .



ψ(0) = Fϕ0 , (5.24)
where Fϕ 2−j (ξ) = 2jn (Fϕ)(2j ξ) denotes the dilation defined in (2.2). Then the following


properties hold:
PN 0
(i) j=0 ψ(j) → δ0 in S as N → +∞;
(ii) kψ(j) kL1 ≤ C, kψ(j) kL∞ ≤ C2nj , for all j;
R
(iii) for all j, ψ(j) = 0.
Proof. It is immediate to see that the sum in Lemma 5.23 converges in S 0 . Then
 N
X  XN
−j
F ϕ0 + ϕ(2 ·) = ψ(j) → δ0
j=1 j=0

as N → +∞ in S 0.
Next,
kψ(j) kL1 = k(Fϕ)2−j kL1 = kFϕkL1 ≤ kFϕ0 kL1 + kF(ϕ0 (2·)kL1 ≤ C .
Moreover,
kψ(j) kL∞ ≤ kFψ(j) kL1 = kϕ(2−j ·)kL1 = 2jn kϕkL1 = C2jn .
58 M. M. PELOSO

Finally, Z
−j
ψ(j) (x) dx = F −1 (ψ(j) )(0) = ϕ2 (0) = ϕ(0) = 0 .
This concludes the proof. 2

Lemma 5.25. Let α ∈ (0, 1) and let p < (n + α)/n. Then, if f ∈ Λ1α we have f ∈ Lp and
kf kLp ≤ Ckf kΛ1α .
Proof. Let f ∈ Λ1α , and let ψ(j) be as in the previous lemma.
Using (iii), observe that for j ≥ 1,
Z Z


kf ∗ ψ(j) kL1 = f (x − y) − f (x) ψ(j) (y) dy dx
n n
ZR Z R
≤ f (x − y) − f (x) dx|ψ(j) (y)| dy
Rn Rn
Z
≤ kf kΛ1α |y|α |ψ(j) (y)| dy
n
ZR
= kf kΛ1α |y|α |2jn Fϕ(2j y)| dy
R n
Z
−jα
≤2 kf kΛ1α |z|α |Fϕ(z)| dz
Rn
−jα
≤ C2 kf kΛ1α .
Moreover, by (i)
kf ∗ ψ(j) kL∞ ≤ kf kL1 kψ(j) kL∞ ≤ C2nj kf kL1 ≤ C2nj kf kΛ1α .
Hence,
Z
kf ∗ ψ(j) kpLp ≤ kf ∗ ψ(j) kp−1
L∞ |f ∗ ψ(j) (x)|p dx ≤ C2nj(p−1) 2−jα kf kpΛ1 .
α
Rn
Therefore,
+∞
X N
X
kf ∗ ψ(j) k
Lp = lim kf ∗ ψ(j) kLp
N →+∞
j=0 j=0
N
X
≤ Ckf kΛ1α lim 2j[n(p−1)−α]/p
N →+∞
j=0
≤ Ckf kΛ1α ,
if n(p − 1) − α < 0,
Pthat is, p < (n + α)/n.
+∞
Consequentely, j=0 f ∗ ψ(j) convereges in the Lp -norm, and since it converges in S 0 to f (at
least when f ∈ S), we obtain that
+∞
X
kf kLp ≤ kf ∗ ψ(j) kLp
j=0
≤ Ckf kΛ1α ,
 HARMONIC ANALYSIS 59

when p < (n + α)/n. 2

We now obtain the reward for our work in this setting with the following result.

Theorem 5.26. Let k(j) ∈ L1 (Rn ), j ∈ Z, be functions such that there exist constants ε > 0,
α ∈ (0, 1) and C > 0 such that for all j ∈ Z,
Z
(1) (1 + |x|)ε |k(j) (x)| dx ≤ C;
Rn
Z
(2) k(j) (x) dx = 0;
Rn

(3) kk(j) kΛ1α ≤ C.

Then the series j∈Z k(j) 2j converges in S 0 to a Calderón–Zygmund kernel.

P

Proof. We wish to show that j∈Z k(j) 2j converges in S 0 to a distribution K that coincides on
P

Rn \ {0} with a locally integrable function and that K satisfies the conditions (5.2) and (5.1).


Step 1. We begin by showing that the series of the Fourier transforms of the k(j) 2j
converges
absolutely, that is,
X

F (k(j) )2j (ξ) ≤ C . (5.25)
j∈Z

We split the sum in two, the first one being for j such that |2j ξ| ≥ 1. Using Lemma 5.22 and
(3) we have21
X X
F k(j) (2j ξ)




F k(j) 2j
(ξ) =
j: |2j ξ|≥1 j: |2j ξ|≥1
X
≤C (1 + |2j ξ|)−α kk(j) kΛ1α
j: |2j ξ|≥1
X
≤C (1 + |2j ξ|)−α
j: |2j ξ|≥1
X
≤C (2j |ξ|)−α
j: |2j ξ|≥1
X
= C|ξ|−α 2−jα
j: 2j |ξ|≥1
≤C. (5.26)

21 Here we also use the elementary estimates P j j0

if 0 < q < 1, and j≤j0 q j ' q j0 if q > 1.
P
j≥j0 q ' q
60 M. M. PELOSO

When |2j ξ| < 1 we use (1), (2) and the estimate |e−2πit − 1| ≤ C|t|ε for |t| < 1 (and footnote
21 again). We have
Z
j
X X
k(j) (x) e−2πi2 ξx − 1 dx




F k(j) 2j (ξ) =
Rn
j: |2j ξ|<1 j: |2j ξ|<1
X Z
≤C |k(j) (x)| |2j ξx|ε dx
Rn
j: |2j ξ|<1
X
≤ C|ξ|ε 2εj
j: |2j ξ|<1
≤C. (5.27)

Estimates (5.26) and (5.27) now give (5.25). We set

X
F k(j) 2j (ξ) ∈ L∞ (Rn ) .



u(ξ) =
j∈Z

Since, for each f ∈ S, by the dominated convergence theorem,

Z XZ


u(ξ)f (ξ) dξ = F k(j) 2j (ξ)f (ξ) dξ
Rn j∈Z Rn

converges in S 0 . By continuity of the (inverse) Fourier transform in S 0 ,

P

so that j∈Z F k(j) 2j
X

k(j) 2j
=K (5.28)
j∈Z

for some K ∈ S 0 . This completes Step 1.

Step 2. We wish to prove that the distribution K defined in (5.28) is a Calderón–Zygmund

kernel.
WePfirst show
that K coincides with a locally integrable function on Rn \ {0}, by showing
that j∈Z k(j) 2j converges in L1 (E), for every compact E ⊂ Rn .
It suffices to consider compact sets E of the form {x : 2m ≤ |x| ≤ 2m+1 }, for some m ∈ Z.22
Notice that, with an obvious change of variables,
Z Z


| k(j) 2j (x)| dx = |k(j) (x)| dx .
{x: 2m ≤|x|≤2m+1 } {x: 2m−j ≤|x|≤2m−j+1 }

Assume first that j ≤ m. Then using (1) we have,

Z Z
(j−m)ε
|k(j) (x)| dx ≤ 2 |x|ε |k(j) (x)| dx
{x: 2m−j ≤|x|≤2m−j+1 } {x: 2m−j ≤|x|≤2m−j+1 }

≤ C2(j−m)ε .

22Make sure you agree with this reduction.

 HARMONIC ANALYSIS 61

On the other hand, if j > m we use (3) and Lemma 5.25 we have
Z Z 1/p
1/p0
|k(j) (x)| dx ≤ |k(j) (x)|p dx {|x| ≤ 2m−j+1 }
{x: 2m−j ≤|x|≤2m−j+1 } {|x|≤2m−j+1 }
0
≤ kk(j) kLp 2n(m−j+1)/p
0
≤ C2n(m−j+1)/p ,

for a suitable p (and p0 ).

Therefore,
0
X X X
kk(j) kL1 (E) ≤ C2(j−m)ε + C2n(m−j+1)/p ≤ C ,
j∈Z j≤m j>m

since both series converge.

We now show that K satisfies the two conditions (5.1) and (5.2). But, by construction and
(5.25)
X

|K̂(ξ)| ≤ F (k(j) )2j (ξ) ≤ C ,
j∈Z

so that (5.1) holds.

Finally, to prove that K satisfies Hörmander’s condition, let h ∈ Rn \ {0} and assume that
2 < 2|h| ≤ 2m+1 . Then we have
m

Z XZ


|K(x − h) − K(x)| dx ≤ | k(j) 2j (x − h) − k(j) 2j (x)| dx
|x|>2|h| j∈Z |x|>2m
XZ
= |k(j) (x − 2−j h) − k(j) (x)| dx
j∈Z |x|>2m−j
XZ
= |k(j) (x − 2−j h) − k(j) (x)| dx
j<m |x|>2m−j
XZ
+ |k(j) (x − 2−j h) − k(j) (x)| dx
j≥m |x|>2m−j
XZ X
≤2 |k(j) (x)| dx + |2−j h|α kk(j) kΛ1α
m−j−1
j<m |x|>2 j≥m
X Z X
≤2 2(j−m)ε |x|ε |k(j) (x)| dx + C |2−j h|α
j<m |x|>2m−j−1 j≥m
X X
(j−m)ε −j α
≤C 2 +C |2 h|
j<m j≥m
≤C,

where we have used conditions (1) and (3) in the hypotheses. This proves Step 2 and hence the
theorem. 2
62 M. M. PELOSO

5.8. †Vector-valued singular integrals. We conclude this part by extending Thm. 5.1 to
the case of vector-valued functions. This extension will be used in Section 8, when developing
the Littlewood–Paley theory.
Although this part of the theory works in the case of functions taking values in a separable
Banach space, we restrict our attention to the case of functions with values in a separable Hilbert
spaceH.
We say that a function f from Rn to H is measurable if for every v ∈ H, the mapping
Rn 3 x 7→ hf (x), viH ∈ C
is measurable.
For 1 ≤ p ≤ ∞ we define the space Lp (Rn , H) to be the space of measurable functions
f : Rn → H such that
Z 1/p
kf (x)kpH dx < ∞.
Rn
Then, Lp (Rn )
= Lp (Rn , C),
but we will still denote it by Lp for simplicity.
A typical example of L (Rn , H) is a function of the form f (x) = fs (x)v, where fs is a scalar-
p

valued function in Lp (Rn ) and v is a fixed element of H. The subset

Lp ⊗ H = f : f = N p
 P
j=1 fj vj where fj ∈ L , vj ∈ H ,

of finite linear combination of functions of the form above, is dense in Lp (Rn , H). (The proof
of this fact is simple, for this and other facts about this topic, see [Ru].)
Given f = N 1
P
j=1 fj vj ∈ L ⊗ H, we define its integral to be the element of H given by
Z N Z
X 
f (x) dx = fj (x) dx vj .
Rn j=1 Rn

map f 7→ f (x)dx extends to all of L1 (Rn , H) by density. Then, for f ∈ L1 (Rn , H),
R
This
R
f (x)dx is the unique element of H such that
DZ E Z
f (x)dx, v = hf (x), viH dx (5.29)
H Rn
for all v ∈ H.
0
Similarly reasoning works also in the case of f ∈ Lp (Rn , H) and g ∈ Lp (Rn , H): notice that
by applying the standard Hölder’s inequality it follows that
Z Z 0
1/p0  Z 1/p
f (x), g(x) H dx ≤ kg(x)kpH dx kf (x)kpH dx
Rn Rn Rn
= kgkLp0 (Rn ,H) kf kLp (Rn ,H) . (5.30)
R
Therefore Rn hf (x), g(x) H dx converges and it turns out that
n Z o
kgkLp0 (Rn ,H) = sup hf (x), g(x)i dx , kf kLp (Rn ,H) = 1 .
Rn
p0

∗
This implies that L (Rn , H) ⊆ Lp (Rn , H) . In fact equality holds (in the Hilbert case), that
is,

∗ 0
Lp (Rn , H) ≡ Lp (Rn , H) .
 HARMONIC ANALYSIS 63

We conclude this preliminary discussion introducing the singular integral in the vector-value
case.
Let H1 , H2 be Hilbert spaces and denote by L(H1 , H2 ) the space of bounded linear operators
from H1 to H2 . Suppose f : Rn → H1 and K ~ : Rn \ {0} → L(H1 , H2 ) are measurable functions.
We may consider the integral operator
Z
T~ f (x) = K(y)f (x − y) dy (5.31)
Rn \{0}

~
Here clearly the expression K(y)f (x − y) can only be interpreted as the element K(y) ~ of
L(H1 , H2 ) acting on f (x − y) ∈ H1 ; hence, it T~ f (x) is well defined, it is an element of H2 .
Theorem 5.27. Suppose T~ is a bounded linear operator from Lr (Rn , H1 ) to Lr (Rn , H2 ) for
some r, 1 < r < ∞, defined by the integral operator with kernel K as in (5.31). Further, assume
that K satisfies the (vector-valued) Hörmander condition
Z
~ − y) − K(x)k
kK(x ~ L(H1 ,H2 ) dx ≤ B . (5.32)
|x−y|>2|y|

Then T~ is bounded from Lp (Rn , H1 ) to Lp (Rn , H2 ) when 1 < p < ∞ and it is weak-type (1, 1),
that is,
C
x ∈ Rn : kT~ f (x)kH2 > λ ≤ kf kL1 (Rn ,H1 ) .

λ
Proof. For simplicity, we are going to drop the “ ~ ”-notation.
The proof does not follow from the scalar case, but, it follows from the same proof. We begin
by observing that the adjoint of T has kernel K ∗ (−x), since
Z DZ E
hT f, gi = K(x − y)f (y)dy, g(x) dx
n n H2
ZR Z R
= hK(x − y)f (y), g(x)iH2 dydx
n n
ZR ZR
= hf (y), K ∗ (x − y)g(x)iH2 dxdy
Rn Rn
Z D Z E
= f (y), K ∗ (x − y)g(x)dx dy .
Rn Rn H1

If we show that T is weak-type (1, 1), by Marcinkiewicz interpolation theorem it would follow
that T is of strong-type (p, p) for 1 < p ≤ r. The operator T ∗ satisfies the same assumption as
0 0
T , with the assumption that T ∗ is a bounded linear operator from Lr (Rn , H2 ) to Lr (Rn , H1 ),
so that it is also of strong-type (p, p) for 1 < p ≤ r0 . Thus, by duality, T is of strong-type (p, p)
for r < p < ∞, so it follows that T is of strong-type (p, p) for 1 < p < ∞.
Thus, (as in the scalar case) it suffices to show that T is weak-type (1, 1).
We now proceed as in the proof of Thm. 5.1. Hence, we only sketch the argument, indicating
the main differences.
The main point is to generalize the Calderón–Zygmund decomposition of an L1 -function. In
the proof of Thm. 4.6, as well as in Thm. 3.5, we assumed that f is non-negative
 (since we

wrote an arbitrary complex-valued function as f = (Re f )+ − (Re f )− + i (Im f )+ − (Im f )− ).
We provide the details for sake of completeness.
64 M. M. PELOSO

Given λ > 0, by applying the Calderón–Zygmund decomposition to the function kf (x)kH1 we

find that there exists a sequence of disjoint cubes {Qj } such that:
(i) 0 ≤ kf (x)kH1 ≤ λ for almost all x 6∈ ∪j Qj ;
1
(ii) ∪j Qj ≤ kf kL1 (Rn ,H1 ) ;
λ
Z
1
(iii) λ < kf (x)kH1 dx ≤ 2n λ.
|Qj | Qj
Now we decompose f as the sum f = g + b, where
X 1 Z 
g(x) = f (x)χ{ c ∪j Qj } (x) + f (y) dy χQj (x) ,
|Qj | Qj
j

and
Z
X X 1 
b(x) = bj (x) = f (x) − f (y) dy χQj (x) .
|Qj | Qj
j j

1
Notice that, as f (x), g(x) and b(x) are vectors in H1 , for a.a. x ∈ Rn , and that
R
|Qj | Qj f (y)dy
is also a well-defined element of H1 , since f ∈ L1 (Rn , H
1 ) and by the identity (5.29).
Notice that, similarly to the proof of Thm.’s 4.6 and 5.1, we have that
X 1 Z
kg(x)kH1 ≤ kf (x)kH1 χ{ c ∪j Qj } (x) + f (y) dy χQj (x)
|Qj | Qj H1
j
X 1 Z 
≤ λχ{ c ∪j Qj } (x) + kf (y)kH1 dy χQj (x)
|Qj | Qj
j
n
≤ 2 λ;

and also that

Z Z
kg(x)kH1 dx ≤ kf (x)kH1 dx .
Rn Rn

(This last inequality follows since g(x) = f (x) on c (∪j Qj ), while

R R
Qj g(x)dx = Qj f (x)dx for
all Qj .)
R
On the other hand, Rn b(x)dx = 0H1 , the zero element in H1 .

We now proceed and in the proof of Thm. 5.1. Since

kT f (x)kH2 ≤ kT g(x)kH2 + kT b(x)kH2 ,

we have

{x : kT f (x)kH2 > λ} ≤ {x : kT g(x)kH2 > λ/2} + {x : kT b(x)kH2 > λ/2} .

 HARMONIC ANALYSIS 65

Now, using the boundedness of T : Lr (Rn , H1 ) → Lr (Rn , H2 ),

kT g(x)krH2
Z Z
{x : kT g(x)kH2 > λ/2} = dx ≤ dx
{x:kT g(x)kH >λ/2} {x:kT g(x)kH >λ/2} (λ/2)r
2 2
Z
C
≤ r kT g(x)krH2 dx
λ Rn
Z Z
C C
≤ r kg(x)krH1 dx ≤ kg(x)kH1 dx
λ Rn λ Rn
Z
C
= kf (x)kH1 dx .
λ Rn
√
Next, let Q∗j denote the cube with the same center cj as Qj with side length 2n times longer.
Then we estimate,
{x : kT b(x)kH2 > λ/2} = {x ∈ (∪j Q∗j ) : kT b(x)kH2 > λ/2}
+ {x ∈ c (∪j Q∗j ) : kT b(x)kH2 > λ/2}
≤ | ∪j Q∗j | + {x ∈ c (∪j Q∗j ) : kT b(x)kH2 > λ/2}
Z
2 2
≤ kf kL1 (Rn ,H1 ) + kT b(x)kH2 dx .
λ λ Rn \(∪j Q∗j )
P
Since kT b(x)kH2 ≤ j kT bj (x)kH2 a.e., it will suffice to prove that
XZ
kT bj (x)kH2 dx ≤ Ckf kL1 (Rn ,H1 ) . (5.33)
j R\Q∗j
R
Notice that, since Qj bj (y)dy = 0H1 , also
Z Z
K(x − cj )bj (y)dy = K(x − cj ) bj (y)dy = 0H2 ,
Qj Qj

so that
Z Z Z
kT bj (x)kH2 dx = K(x − y)bj (y) dy dx
R\Q∗j R\Q∗j Qj H2
Z Z  
= K(x − y)bj (y) − K(x − cj )bj (y) dy dx
R\Q∗j Qj H2
Z Z


≤ K(x − y) − K(x − cj ) bj (y) H2
dy dx
R\Q∗j Qj
Z Z
≤ K(x − y) − K(x − cj ) L(H1 ,H2 )
kbj (y)kH1 dy dx
R\Q∗j Qj
Z Z
≤ kbj (y)kH1 K(x − y) − K(x − cj ) L(H1 ,H2 )
dx dy
Qj R\Q∗j
Z
≤B kbj (y)kH1 dy .
Qj
66 M. M. PELOSO

Therefore,
XZ XZ
kT bj (x)kH2 dx ≤ B kbj (y)kH1 dy ≤ 2Bkf kL1 (Rn ,H1 ) ,
j R\Q∗j j Qj

estimates that completes the proof. 2

 HARMONIC ANALYSIS 67

6. Fourier multipliers
In this chapter we can to the analysis of the so-called Fourier multipliers, that is, of the
operators, initially defined on Schwartz functions, of the form
Tm (f )(x) = F −1 mfˆ)(x) ,
where m ∈ S 0 (Rn ). We will often identify the operator Tm with the distribution m and call
them indifferently a Fourier multiplier.
Observe that m ∈ S 0 implies that m = K̂ for some K ∈ S 0 , so that
Tm (f ) = K ∗ f .
Therefore, the singular integrals (of convolution type) that we have seen in the previous sections,
are particular cases of Fourier multipliers.
Notice that fˆ ∈ S, so that mfˆ ∈ S 0 and it makes sense to define its inverse Fourier transform.
We mention, without proving, the following result, due to Hörmander, that shows that every
bounded linear operator between Lp -spaces that commutes with translations is given by the
convolution with a tempered distribution K; hence it is a Fourier multiplier.
Recall that, if a ∈ Rn we define [τa f ](x) = f (x + a). An operator T is said to commute with
translations if
 

τa T (f ) (x) = T [τa f ] (x) ,
for all a ∈ Rn , and a.a. x.
Theorem 6.1. Let 1 ≤ p, q ≤ ∞. Suppose that T : Lp (Rn ) → Lq (Rn ) is a bounded operator
that commutes with translations. Then there exists a unique u ∈ S 0 such that
T (f ) = u ∗ f
for all f ∈ S.
For the proof we refer to [StWe], Thm. 3.16.
We will first show a few basic properties of these operators, and then describe some sufficient
conditions on the function m to ensure that Tm is a bounded operator on Lp .

6.1. The space Mp of Lp -bounded Fourier multipliers. We define the space

Mp = m ∈ S 0 : Tm : Lp → Lp is bounded ,


and we set
kmkMp = kTm k(Lp ,Lp ) ≡ sup kTm (f )kLp .
kf kLp =1

By Plancherel’s theorem we know that if m is bounded then Tm is bounded on L2 . The

converse of this assertion is also true, as consequence of the following result.
Lemma 6.2. If Tm is bounded on Lp then m is bounded. In particular, Tm is bounded on L2 is
and only if m ∈ L∞ and, in this case, kTm k(L2 ,L2 ) = kmkL∞ . Finally, Mp = Mp0 , if p and p0
are conjugate exponents.
68 M. M. PELOSO

Proof. If m in not bounded, then it is easy to see that Tm cannot be bounded on L2 (exercise).
Now, it is easy to see that kTm k(L2 ,L2 ) = kmkL∞ . For, by Plancherel theorem,

kTm f kL2 = kmfˆkL2 ≤ kmkL∞ kfˆkL2 = kmkL∞ kf kL2 ,

so that kTm k(L2 ,L2 ) ≤ kmkL∞ .
Next, if ε > 0 is fixed, let

Eε = x : |x| ≤ 1/ε and |m(x)| > kmkL∞ − ε .

Then Eε has positive and finite measure, and let f ∈ L2 be such that fˆ = χEε . Hence,
Z
kTm f kL2 = kmf kL2 = |m(ξ)fˆ(ξ)|2 dξ
2 ˆ 2

Z
> (kmkL∞ − ε) 2
|fˆ(ξ)|2 dξ

= (kmkL∞ − ε)2 kf k2L2 .

It follows that kTm k(L2 ,L2 ) ≥ kmkL∞ , and then equality holds.
Next we show that if Tm is bounded on Lp , 1 ≤ p < ∞, then m is bounded. For, notice
that the adjoint operator Tm ∗ is given by convolution with Ǩ. Thus, T ∗ is also bounded on Lp .
m
0
Since it is bounded on Lp , by Marcinkiewicz interpolation theorem23 is also bounded on Lr for
r between p and p0 ; hence in particular on L2 . This implies that m ∈ L∞ and we are done. 2

Proposition 6.3. For 1 < p < q < 2 we have the inclusion

M1 ⊂ Mp ⊂ Mq ⊂ M2 (6.1)
and moreover,
kmkL∞ ≤ kmkMq ≤ kmkMp .
Proof. This follows easily from the previous lemma and the Marcinkiewicz theorem, if m ∈ M1 ,
then m is also in M2 and therefore in Mr for all 1 < r < 2. The same argument proves the
other inclusions in (6.1).
Using again the Riesz–Thorin theorem, for 1 < p < 2 we have that
1/2 1/2
kTm k(L2 ,L2 ) ≤ kTm k(Lp ,Lp ) kTm k(Lp0 ,Lp0 ) = kTm k(Lp ,Lp ) ,

since
1 1/2 1/2
= + 0 .
2 p p
A similar argument shows that, for 1 < p < q < 2, kTm k(Lq ,Lq ) ≤ kTm k(Lp ,Lp ) and we are done.
2

Proposition 6.4. For 1 ≤ p < ∞ the space Mp is a Banach space with respect to the norm
k · kMp . Moreover, Mp is closed under pointwise multiplication and it is a Banach algebra.24
23Or, more simply by the Riesz-Thorin Theorem, [Ka] or [So].
24An algebra A which is a Banach space w.r.t. the norm k·k is called a Banach algebra if the product satisfies
A
the inequality kxykA ≤ kxkA kykA for all x, y ∈ A.
 HARMONIC ANALYSIS 69

Proof. It suffices to consider the case 1 ≤ p ≤ 2.

It is obvious that Mp is a linear space and that the equality kmkMp = kTm k(Lp ,Lp ) defines a
norm on Mp .
Next, if m1 , m2 ∈ Mp , then
(Tm1 m2 f )ˆ(ξ) = m1 (ξ)m2 (ξ)fˆ(ξ) = Tm1 (Tm2 )f ˆ(ξ) ,

so that Tm1 m2 = Tm1 Tm2 . Moreover,

kTm1 m2 f kLp ≤ kTm1 k(Lp ,Lp ) kTm2 f kLp
≤ kTm1 k(Lp ,Lp ) kTm2 k(Lp ,Lp ) kf kLp ,
that is,
km1 m2 kMp = kTm1 m2 k(Lp ,Lp ) ≤ kTm1 k(Lp ,Lp ) kTm2 k(Lp ,Lp ) = km1 kMp km2 kMp .
This proves that Mp is a Banach algebra w.r.t. the pointwise product.
We now show that Mp is complete. Let {mj } be a Cauchy sequence in Mp . By Prop. 6.3 it
follows that {mj } is a Cauchy sequence in L∞ , so it converges, in the L∞ -norm, hence a.e., to
a function m. We wish to show that mj → m in Mp .
Let f ∈ S be fixed. Then, by Lebesgue’s dominated convergence theorem we have that
Z Z
Tmj (f )(x) = ˆ
f (ξ)mj (ξ)e 2πixξ
dξ → fˆ(ξ)m(ξ)e2πixξ dξ = Tm (f )(x)
Rn Rn
a.e. By Fatou’s lemma it follows that
Z Z
|Tm (f )(x)|p dx ≤ lim inf |Tmj (f )(x)|p dx
Rn j→+∞ Rn
≤ lim inf kmj kMp kf kLp ,
j→+∞

that is,
kmkMp ≤ lim inf kmj kMp ≤ C p
j→+∞
since {mj } is a Cauchy sequence, hence the sequence of the norms is bounded.
From the last inequality it also follows that
km − mk kMp ≤ lim inf kmj − mk kMp < ε
j→+∞

if k ≥ k0 , since {mj } is a Cauchy sequence. This proves the proposition. 2

Examples 6.5.
(1) The first example we consider is m(ξ) = −i sgn(ξ). Then Tm is the Hilbert transform,
and we know that Tm is weak-type (1, 1) and bounded on Lp for 1 < p < ∞.
(2) Consider the interval (a, b), and assume for the moment that it is bounded. Notice that
1 
χ(a,b) (ξ) = sgn(ξ − a) − sgn(ξ − b) .
2
Let a ∈ R and observe that the operator (called the modulation operator)
Ma f (x) = e2πiax f (x)
70 M. M. PELOSO

is bounded on Lp for all p, and also that (Ma f )ˆ(ξ) = fˆ(ξ − a) = (τa fˆ)(ξ). Then notice that, for
f ∈ S(R),
1
χ(a,b) (ξ)fˆ(ξ) = sgn(ξ − a) − sgn(ξ − b) fˆ(ξ)

2
1h
i
= τa sgn(ξ)τ−a (fˆ) (ξ) − τb sgn(ξ)τ−b (fˆ) (ξ)


2
i h i 
= F Ma HM−a − Mb HM−b (f ) (ξ) , (6.2)
2
where H denotes the Hilbert transform. (This last equality can be easily be checked by com-
puting the Fourier transform on the right hand side, recalling that (Hf )ˆ(ξ) = −i sgn(ξ)fˆ(ξ).)
Therefore,
ih i
F −1 χ(a,b) ∗ f = Ma HM−a − Mb HM−b (f ) ,


2
so that
i h i
Tχ(a,b) (f ) = Ma HM−a − Mb HM−b (f ) . (6.3)
2
Next, suppose that the interval is unbounded. Let (a, b) = (a, +∞), the other case begin
analogous. Then
1 
χ(a,+∞) (ξ) = 1 + sgn(ξ − a) ,
2
so that, arguing as in (6.2),
i h i 
χ(a,+∞) (ξ)fˆ(ξ) = F − iI + Ma HM−a (f ) (ξ) ,
2
i.e.
ih i
Tχ(a,+∞) (f ) = Ma HM−a − iI (f ) . (6.4)
2
Therefore, it follows from (6.3) and (6.4) that, for all a, b ∈ R ∪ {±∞}, m = χ(a,b) is in
Mp (R) for 1 < p < ∞, with operator norms uniformly bounded in a and b.
(3) Let n > 1. The Hilbert transform can be extended to define a family of Lp -bounded Fourier
multipliers in Rn . For j = 1, . . . , n define mj (ξ) = −i sgn(ξj ) and hence Tmj (f ) = F −1 mj fˆ).
It is easy to see that
Z +∞
1 1
Tmj f (x) = p.v. f (x1 , . . . , xj − t, . . . , xn ) dt .
π −∞ t
Now, taking j = n for simplicity, using the boundedness of the Hilbert transfom in the last
variable we have
Z Z
p
kTmn f kLp (Rn ) = |Tmn f (x1 , . . . , xn−1 , xn )|p dxn dx1 · · · dxn−1
Rn−1 R
Z Z
≤C |f (x1 , . . . , xn−1 , xn )|p dxn dx1 · · · dxn−1
Rn−1 R
= Ckf kpLp (Rn ) .
It also follows that, if a ∈ Rn , b ∈ R and
E = ξ ∈ Rn : a · ξ > b ,


then m = χE ∈ Mp , 1 < p < ∞. (Exercise.)

 HARMONIC ANALYSIS 71

(4) Let m = χ{|ξ|<1} be the characteristic function of the unit ball in Rn . If n = 1 it follows
from the previous discussion that m ∈ Mp for all p, 1 < p < ∞. On the other hand, the next
result, the celebrated Fefferman’s theorem on the ball multiplier, Thm. 6.12 below, states that
if n > 1 that χ{|ξ|<1} ∈ Mp if and only if p = 2.

We conclude this section by stating the characterization of M1 , a significant result, for whose
proof we refer to [StWe] or [Gr]. We recall Rthat a finite complex-valued Borel measure µ defines
a tempered distribution by setting µ(f ) = f (x)dµ(x).
Theorem 6.6. A measurable function m ∈ M1 if and only if m is the Fourier transform of a
finite complex-valued Borel measure µ.

6.2. Multiple Fourier series and convergence in the Lp -norm. In this section we briefly
illustrate an application of the theory of Fourier multipliers.
We define the n-dimensional torus to be the quotient group Tn = Rn /Zn . It is easy to
convince oneself that Tn can be identified with the compact set [0, 2π] × · · · × [0, 2π] ≡ [0, 2π]n .
Given an integrable function f on Tn we define its (multiple) Fourier series to be the expression
X
fˆ(k)e2πik·t , (6.5)
k∈Zn
where t ∈ Tn and Z
fˆ(k) = f (t)e−2πik·t dt .
Tn
The main question is in what sense the expression in (6.5) represents f , that is, if it converges
and if it does converge, whether it converges to f .
By elementary theory of Hilbert spaces applied to the case of H = L2 (Tn ) it follows at once
that, if f ∈ L2 (Tn ) then the series in (6.5) converges (unconditionally) to f in the L2 (Tn )-norm.
If p 6= 2 and we are still interested in the norm convergence of the series in (6.5), when n > 1
we need to specify what we mean by the summation, since Zn lacks of a natural order. Therefore,
we proceed as follows.
Definition 6.7. We call convex body an open convex set C ⊂ Rn containing the origin. For
λ > 0 we denote by λC the dilated of C by λ, that is the set {y = λx : x ∈ C, λ > 0}, and define
X
SCλ (f ) = fˆ(k)e2πik·t . (6.6)
k∈Zn ∩λC
We say that the Fourier series converge in Lp (Tn ),
1 ≤ p < ∞ in the summation method
given by the convex body C, or more briefly, in the C-sense, if
lim SCλ (f ) − f = 0,
λ→+∞ Lp (Tn )

for every f ∈ Lp (Tn ) ∩ C(Tn ).

Our goal now is to relate the norm convergence of Fourier series to the boundedness of a
single operator, as we have done in the case of the Fourier series on the torus. Namely, we have
Theorem 6.8. Let 1 < p < +∞ and let a convex body C in Rn be given. Then the following
conditions are equivalent.
72 M. M. PELOSO

(i) The space Lp (Tn ), 1 ≤ p < ∞, admits norm convergence for the Fourier series in the
summation method given by the convex body C.
(ii) There exists C > 0 such that
sup kSCλ k(Lp (Tn ),Lp (Tn )) ≤ C .
λ>0

(where kSCλ k(Lp (Tn ),Lp (Tn )) denotes the norm of SCλ as an operatort on Lp (Tn )).
(iii) The function χC ∈ Mp (Rn ).

We introduce the notion of Fourier multipliers on Tn .

Definition 6.9. Let ω : Zn → C be a complex-valued function on Zn . We say that ω is
a bounded Fourier multiplier on Lp (Tn ) if the operator Tω initially defined on trigonometric
polynomials as
X
Tω (f )(t) = ω(k)fˆ(k)e2πik·t
k∈Zn
extends to a bounded operator Tω : Lp (Tn ) → Lp (Tn ).
The following result is not difficult to prove, but for lack of adequate time we refer to [So],
Ch. IV for its proof.
Theorem 6.10. The following implications hold.
(1) If m ∈ Mp (Rn ) and m is continuous on Zn , then m̃ := m|Zn ∈ Mp (Tn ), and km̃kMp (Tn ) ≤
kmkMp (Rn ) .
(2) If m is a function defined a.e. in Rn and everywhere on Zn and it is such that mε ∈
Mp (Tn ) for every ε > 0 and
sup kTmε kMp (Tn ) ≤ C ,
ε>0

for some C > 0, then m ∈ Mp (Rn ) and km ∈ Mp (Rn )kMp (Rn ) ≤ C.

We recall that mε (x) = m(εx). We are now ready to prove Thm. 6.8.
Proof of Thm. 6.8. We first observe that same proof of Thm. 1.13 shows that (i) is equivalent
to (ii).
Next, we notice that since C is convex and 0 ∈ C, λ1 < λ2 implies that λ1 C ( λ2 C. In
particular, the summation method given by C and any of its dilated coincides. Moreover, ∂(C) ⊆
(1 + ε)C \ (1 − ε)C so that
 
|∂(λC)| ≤ (1 + ε)n − (1 − ε)n |C| ≤ εC|C|

for every ε > 0. Thus, |∂(C)| = 0. Thus, χλC is a.e. continuous in Rn , and clearly defined
everywhere on Zn Moreover, for k ∈ Zn there exists just one value λ > 0 such that k ∈ ∂(λC);
call such value λk . Thus, we may select an increasing sequence λ1 < λ2 < · · · , λm → +∞, such
that no point of Zn lies on ∂(λm C) for every m, in such a way that so that χλm C is continuous
on Zn , for every m, and moreover, for each λ > 0, there exists m such that
SCλ = SCλm .
 HARMONIC ANALYSIS 73

1/λ
Now we observe that, χλC (k) = χC (k) so that for any trigonometric polynomial f ,
X X
SCλ f (t) = fˆ(k)e2πik·t = χλC (k)fˆ(k)e2πik·t
k∈Zn ∩λC k∈Zn
X 1/λ
= χC (k)fˆ(k)e2πik·t
k∈Zn
= Tχ1/λ f (t) .
C

Suppose then that (ii) holds. This gives that Tχ1/λ is a bounded operator on Lp (Tn ), with
C
1/λ
uniformly bounded norms, that is, χC ∈ Mp (Tn ) and
1/λ
kχC kMp (Tn ) ≤ C .
Thm. 6.10 (2) gives that χC ∈ Mp (Rn ).
Suppose that (i) holds. Having selected the increasing sequence {λm } as above, χλm C is
continuous on Zn , then χλm C |Zn ∈ Mp (Tn ), by Thm. 6.10 (1). Therefore,

sup kSCλ k(Lp (Tn ),Lp (Tn )) = sup kSCλm k(Lp (Tn ),Lp (Tn )) = sup kTχ1/λm k(Lp (Tn ),Lp (Tn ))
λ>0 m∈N m∈N C

1/λ
= sup kχC m | n kMp (Tn )
Z
m∈N
1/λm
≤ sup kχC kMp (Rn )
m∈N
= kχC kMp (Rn ) .
In order to prove the last equality, it suffices to notice the norm k · kMp (Rn ) is dilation invariant,
that is,
kmε kMp (Rn ) = kmkMp (Rn ) .
for every ε > 0. We leave the elementary verification to the reader. This completes the proof. 
A subset S of Rn is called a half-space if there exist a ∈ Rn and b ∈ R such that
S = x ∈ Rn : a · x > b .


A bounded subset P ⊂ Rn is called a convex polyhedron if it is convex, contains the orgin,

and it is the intersection of a finite number of half-spaces.
The most typical examples of convex bodies are the unit ball B, which has smooth boundary
and nowhere vanishing curvature, and a convex polyhedron, that instead has only Lipschitz
boundary a.e. flat.
Theorem 6.11. Let P be a convex polyhedron. Then χP ∈ Mp , 1 < p < ∞. Hence, Lp (Tn )
admits norm convergence in the P -summation method.
Proof. If n = 1, then P = (a, b) for some a < 0 < b and the result follows from Example 6.5 (2).
When n > 1 we easily reduce ourselves to theQboundedness of the Riesz transforms. For, if
N
P ∩N
j=1 Sj , where Sj are half-spaces, then χP = j=1 χSj . Hence,

TχP = TQN χS j = TχS1 · · · TχSN .

j=1
74 M. M. PELOSO

Thus, it suffices to show that each TχSj is bounded on Lp (Rn ), 1 < p < ∞. Composing with a
rotation and a translation, it suffices to show that χS ∈ Mp , when
S = ξ ∈ Rn : ξ1 > 0 ,


and this follows from Example 6.5 (3). 2

The next result came as a big surprise. In the early ’70 C. Fefferman proved that the charac-
teristic function of the unit ball in Rn with n > 1 is not a bounded Fourier multiplier in Lp (Rn ),
unless p = 2.
Theorem 6.12. Let m = χB be the characteristic function of the unit ball B in Rn , where
n > 1. Then m ∈ Mp if and only if p = 2.
The proof of this theorem is quite elaborated and for the time being we refer to [St2] Ch. X
Section 2.5, or [So] Ch. 4, Section 3. (The latter proof is self-contained and it is essentially the
original proof by C. Fefferman, while the latter one shows the deep connection of this problem
with other important concepts and open problems in modern harmonic analysis.)

6.3. The Sobolev spaces H s . We have remarked that the characteristic function of the unit
ball in Rn , with n > 1, defines a Fourier multiplier that is bounded on Lp only for p = 2. It is
clear that, besides the boundedness, the function m needs to posses some regularity. In order
to measure regularity we introduce a family of spaces, called the Sobolev spaces, whose elements
have derivatives (in some weak sense) in L2 .
We begin our brief introduction with the case of a non-negative integer k. We define H k =
H (Rn ) to be the space of L2 functions f such that the distributional derivatives ∂ α f ∈ L2 for
k

all α with |α| ≤ k. On H k we can define an inner product that makes H k into a Hilbert space:
X Z
hf |gik = (∂ α f )(∂ α g) .25
|α|≤k

We wish to extend this definition to non-integral values of k and also to have a more “efficient”
way to represent the inner product. We observe that, by Plancherel theorem, f ∈ H k if and
only if ξ α fˆ ∈ L2 for all α, |α| ≤ k. We claim that there exist positive constants c1 , c2 such that,
for all ξ ∈ Rn , X
c1 1 + |ξ|2 )k/2 ≤ |ξ α | ≤ c2 1 + |ξ|2 )k/2 .
|α|≤k
For, if α is a multi-index with |α| ≤ k, if |ξ| ≥ 1 then
|ξ α | ≤ |ξ|k ≤ (1 + |ξ|2 )k/2 ,
while, if |ξ| ≤ 1,
|ξ α | ≤ 1 ≤ (1 + |ξ|2 )k/2 .
On the other hand, since |ξ|k and kj=1 |ξj |k are both homogeneous of degree k non-vanishing
P
for ξ 6= 0, we have
 X k  X
2 k/2 k 0
(1 + |ξ| ) ≤ c0 (1 + |ξ| ) ≤ c0 1 + c0 |ξj |k ≤ C |ξ α | .
j=1 |α|≤k

25The fact that this inner product defines a Hilbert space, hence in particular complete, is left as an Exercise.
 HARMONIC ANALYSIS 75

By the claim, it follows that f ∈ H k if and only if (1 + |ξ|2 )k/2 fˆ ∈ L2 . Moreover,

Z X Z
c1 1 + |ξ|2 )k |fˆ(ξ)|2 dξ ≤ |ξ α |2 |fˆ(ξ)|2 dξ
|α|≤k
X Z
≤c |∂ α f (x)|2 dx
|α|≤k
Z
≤ c02 1 + |ξ|2 )k |fˆ(ξ)|2 dξ ,

so that the two norms

X Z 1/2 Z 1/2
α 2
|∂ f (x)| dx and 1 + |ξ| ) |fˆ(ξ)|2 dξ
2 k

|α|≤k

are equivalent, and the latter one is defined for all k real, not just non-negative integer.

Definition 6.13. We define the Bessel potential of order s ∈ R of a distribution f ∈ S 0 as

 
−1 2 s/2 ˆ


Λs f = F 1 + |ξ| f .

Notice that Λs u is well defined for any u ∈ S 0 , since, for any Schwartz function ψ, if ϕ = F −1 ψ,
(Λs u)(ψ) = (Λs u)(ϕ̂) = (1 + |ξ|2 )s/2 û (ϕ)


Z
= (1 + |ξ|2 )s/2 û(ξ)ϕ(ξ) dξ = û (1 + |ξ|2 )s/2 ϕ

and (1 + |ξ|2 )s/2 ϕ(ξ) is again a Schwartz function.

Definition 6.14. Let s ∈ R. We define the Sobolev space H s of order s ∈ R as the space of
distributions f ∈ S 0 such that Λs f ∈ L2 , that is
Z
Hs = f ∈ S0 : |fˆ(ξ)|2 (1 + |ξ|2 )s dξ < ∞ .


We endow H s with the inner product

Z
hf |gis = (Λs f )(Λs g) ,

which also gives the norm

Z 1/2
kf kH s = |fˆ(ξ)|2 (1 + |ξ|2 )s dξ .

Proposition 6.15. The following properties are elementary.

(i) The Fourier transform F is a unitary isomorphism from H s to L2 (1 + |ξ|2 )s dξ .

(ii) The Schwartz space S is dense in H s for all s ∈ R.

(iii) If s > t then k · kH t ≤ k · kH s so that H s is a subspace of H t and it is also a dense
subspace.
(iv) The Bessel potential Λs is a unitary isomorphism from H t to H t−s for all s, t ∈ R. Its
inverse is Λ−s .
76 M. M. PELOSO

(v) H 0 = L2 , so that H s ⊂ L2 for all s > 0. For s < 0 the elements of H s may not be
functions.
(vi) The operators ∂ α are continuous from H s to H s−|α| , for all s and α.

Proposition 6.16. For s ∈ R, the L2 inner product h·|·i induces a unitary isomorphism between
H −s and the dual space (H s )∗ of H s .

Proof. Let f, ϕ ∈ S. Then, by Plancherel’s theorem,

Z
hϕ|f i = hϕ̂|f i ≤ (1 + |ξ|2 )−s/2 |fˆ(ξ)|(1 + |ξ|2 )s/2 |ϕ̂(ξ)| dξ
ˆ
Z 1/2 Z 1/2
2 −s
≤ (1 + |ξ| ) |fˆ(ξ)|2 dξ 2 s 2
(1 + |ξ| ) |ϕ̂(ξ)| dξ

= kf kH −s kϕkH s .

Therefore, the linear functional ϕ 7→ hϕ|f i extends to all of H s with norm ≤ kf kH −s . But,
we actually have equality of norms, since if g = (Λ−2s f ),
Z
hf |gi = hfˆ|ĝi = (1 + |ξ|2 )−s |fˆ(ξ)|2 dξ = kf k2H −s
= kf kH −s kgkH s .

Notice that the pairing
hϕ̂|fˆi is well defined for all f ∈ H −s and ϕ ∈ H s .
∗
Finally, let L ∈ H s . Then L ◦ F
−1 is a bounded linear functional on L2 (1 + |ξ|2 )s dξ , so

that there exists g ∈ L2 (1 + |ξ|2 )s dξ such that

Z
−1
L(ϕ) = L ◦ F (ϕ̂) = ϕ̂(ξ)g(ξ)(1 + |ξ|2 )s dξ
Z


= ϕ̂(ξ)FΛ2s (F −1 g)(ξ) dξ

= hϕ̂|Λ2s (F −1 g)i
= hϕ|f i .

But, f = Λ2s (F −1 g) ∈ H −s (since F −1 g ∈ H s ). This proves the proposition. 2

We conclude this part with the Sobolev immersion theorem. We set26

C0k = {f ∈ C k (Rn ) : ∂ α f ∈ C0 , for |α| ≤ k} .

Lemma 6.17. Let f be such that fˆ ∈ H s , where s > n/2. Then f ∈ L1 and for all ε < s − n/2
we have
Z
|f (x)|(1 + |x|)ε dx ≤ Cε kfˆkH s .
Rn

26We recall that we denote by C the space of continuous functions that vanishes at infinity.
0
 HARMONIC ANALYSIS 77

Proof. Using the Cauchy-Schwarz inequality we have

Z Z
|f (x)|(1 + |x|)ε dx ≤ 2ε |f (x)|(1 + |x|2 )ε/2 dx
Z 1/2  Z 1/2
≤ 2ε |f (x)|2 (1 + |x|2 )s dx (1 + |x|2 )ε−s dx (6.7)

≤ Cε kfˆkH s ,

1/2
where Cε = 2ε (1 + |x|2 )ε−s dx
R
is finite since ε < s − n/2. 2

Theorem 6.18. (Sobolev Embedding Theorem) Let t > k + (n/2). Then, H t embedds continu-
ously in C0k .
Proof. Let g ∈ H t and |α| ≤ k. By the previous lemma, F −1 ∂ α g ∈ L1 , since ∂ α g ∈ H s , where

s = t − |α| > n/2, and

Z
|F −1 ∂xα g (x)|(1 + |x|)ε dx ≤ Cε k∂ α gkH s ≤ Cε kgkH t .


Rn
By the Riemann-Lebesgue lemma, ∂xα g ∈ C0 , for |α| ≤ k. 2
For sake of clarity, we isolate the result that we need in the remaining of this part.
Corollary 6.19. Let s > n/2. Then, if f ∈ H s it follows that f is continuous and
kf kL∞ ≤ Ckf kH s .
We conclude this part with a lemma that we are going to need later on.
Lemma 6.20. Let f ∈ H s and ϕ ∈ S. Then f ϕ ∈ H s and
kf ϕkH s ≤ Cϕ kf kH s .
Proof. We begin with the preliminary observation that
1 + |ξ + ξ 0 |2 ≤ 1 + 2(|ξ|2 + |ξ 0 |2 ) ≤ 2(1 + |ξ|2 )(1 + |ξ 0 |2 ) . (6.8)
[
Next, since (f ϕ) = fˆ ∗ ϕ̂, using the Cauchy-Schwarz intequality, we have
Z Z Z
2
[
|(f ϕ)(ξ)|2 (1 + |ξ|2 )s dξ = fˆ(ξ − ξ 0 )ϕ̂(ξ 0 ) dξ 0 (1 + |ξ|2 )s dξ
Rn Rn n
ZR Z
≤ kϕ̂kL1 |fˆ(ξ − ξ 0 )|2 |ϕ̂(ξ 0 )| dξ 0 (1 + |ξ|2 )s dξ
n n
ZR R Z
0
≤ kϕ̂kL1 |ϕ̂(ξ )| |fˆ(ξ)|2 (1 + |ξ + ξ 0 |2 )s dξ dξ 0
Rn Rn
Z Z
0 0 2 s
s
≤ 2 kϕ̂kL1 |ϕ̂(ξ )|(1 + |ξ | ) |fˆ(ξ)|2 (1 + |ξ|2 )s dξ dξ 0
Rn Rn
≤ Cϕ kf k2H s ,
where we have used the fact that ϕ ∈ S. This proves the lemma. 2
78 M. M. PELOSO

6.4. †The Mihlin–Hörmander multiplier theorem. In this section we prove the Mihlin–
Hörmander multiplier theorem. The Mihlin–Hörmander condition that guarantees that a func-
tion gives rise to a Fourier multiplier that is bounded on Lp (Rn ) has the following features:
(i) it is invariant under the dilations m 7→ m(r·) for r > 0, in the sense that m satisfies this
condition then also m(r·) does;
(ii) the operator Tm : f 7→ (F −1 m) ∗ f is an operator whose kernel K = (F −1 m) is a
Calderón–Zygmund convolution kernel, hence it is of weak-type (1, 1) and therefore
bounded on Lp (Rn ), 1 < p < ∞;– in order to show this we are going to appeal to
Thm. 5.26.
We are now ready to define the space of Mihlin-Hörmander multipliers.
Definition 6.21. Let 0 < a0 < a < b < b0 and let ψ ∈ C0∞ be such that

(i) supp ψ ⊆ ξ : a0 ≤ |ξ| ≤ b0 ;
(ii) ψ ≥ 0 and ψ(ξ) = 1 for a ≤ |ξ| ≤ b.
We call Mihlin-Hörmander multiplier a function m such that
kmkMH s = sup km(r·)ψkH s , (6.9)
r>0
for s > n/2. We denote the space of such multipliers by MH s .
Remark 6.22.
(1) Although we are not going to prove this assertion, and the one below, it is important to
notice that the above definition is independent on the choice of ψ and that, a different ψ just
gives rise to an equivalent norm.
(2) The norm k · kMH s is invariant under dilation, in the sense that if m ∈ MH s and mr is
given by mr (ξ) = m(rξ), where r > 0, then mr ∈ MH s and kmr kMH s = kmkMH s .27
Lemma 6.23. For s > n/2, if m ∈ MH s , then m is bounded and
kmkL∞ ≤ CkmkMH s .
Proof. By assumption, the functions m(r·)ψ are in H s with s > n/2 and norms uniformly
bounded. By Cor. 6.19, it follows that km(r·)ψkL∞ = kmψ(r−1 ·)kL∞ ≤ CkmkMH s , with C
independent of r > 0.
Hence,
sup |m(ξ)|2 = sup sup |m(ξ)|2
ξ j 2−j ≤|ξ|≤2−j+1
X
= sup sup |m(ξ)ψ(2−k ξ)|2
j 2−j ≤|ξ|≤2−j+1 k
X
= sup sup |m(ξ)ψ(2−k ξ)|2
j 2−j ≤|ξ|≤2−j+1 k∈{j−1,...,j+2}

≤ 4 sup sup |m(ξ)ψ(2−k ξ)|2

k ξ∈Rn
≤ CkmkMH s .
This proves the lemma. 2

27Exercise: Prove (1) and (2).

 HARMONIC ANALYSIS 79

Theorem 6.24. (Mihlin–Hörmander) Let m ∈ MH s , with s > n/2. Then, the multiplier
operator Tm is bounded on Lp , 1 < p < ∞ and it weak-type (1, 1).
Before proving the theorem, we see a corollary.
Corollary 6.25. Let m ∈ C k \ {0}, with k > bn/2c + 1, be such that
 Z 1/2
|α| −n α 2
sup r r |∂ m(ξ)| dξ < ∞, (6.10)
r>0 r/2<|ξ|<2r

for all |α| ≤ k. Then, m ∈ Mp , for 1 < p < ∞.

In particular, m ∈ Mp , 1 < p < ∞, if
|∂ α m(ξ)| ≤ C|ξ|−|α| , (6.11)
for all |α| ≤ k.
Proof. We make the change of variables ζ = ξ/r in (6.10) to obtain
 Z 1/2 Z 1/2
|α| −n α 2 2|α| α 2
sup r r |∂ m(ξ)| dξ = sup r |(∂ m)(rζ)| dζ
r>0 r/2<|ξ|<2r r>0 1/2<|ζ|<2
Z 1/2
α 2


= sup | ∂ m(r·) (ζ)| dζ .
r>0 1/2<|ζ|<2

This quantity is, by the assumption (6.10), finite. We prove that (6.9) is also satisfied. For,
if |α| ≤ k, by the Leibnitz rule, using the facts that supp ψ ⊆ {1/2 ≤ |ξ| ≤ 2} and that
|∂ β ψ(ξ)| ≤ C, we see that
X
∂ α m(r·)ψ L2 = cα,β ∂ β m(r·)∂ α−β ψ L2

β≤α
X Z 1/2
2
∂ξβ m(r·) (ξ)∂ α−β ψ(ξ) dξ


≤C
β≤α Rn

X Z 1/2
β 2


≤C | ∂ m(r·) (ξ)| dξ ,
|β|≤k 1/2<|ξ|<2

where the constant C is independent of r > 0.

Therefore, m(r·)ψ ∈ H k with norm uniformly bounded in r > 0, that is, m ∈ MH k , with
k > bn/2c + 1. Applying Thm. 6.24 we obtain that m ∈ Mp , for 1 < p < ∞.
Finally, if m satisfies (6.11), then
 Z 1/2  Z 1/2
|α| −n α 2 |α| −n −2|α|
r r |∂ m(ξ)| dξ ≤ Cr r |ξ| dξ
r/2<|ξ|<2r r/2<|ξ|<2r
≤C,
that is, (6.10) is also satisfied, and the corollary is proven. 2
80 M. M. PELOSO

6.5. †Proof of Thm. 6.24. Let m ∈ MH s with s > n/2 and let ψ be as in Lemma 5.23 (i),
that is, ψ ∈ C0∞ and such that j∈Z ψ(2j ξ) = 1 for all ξ 6= 0.
P

We set mj (ξ) = m(2−j ξ)ψ(ξ), so that

j
X X
(mj )2 (ξ) = mj (2j ξ) = m(ξ)
j∈Z j∈Z

for all ξ 6= 0.
Morevore, we define
K = F −1 m , and k(j) = F −1 mj . (6.12)

Notice that then we have

X
2j
X
2 j X 
 X

FK = m = (mj ) = Fk(j) = F k(j) 2j = F k(j) 2j ,
j∈Z j∈Z j∈Z j∈Z

that is,
X

K= k(j) 2j
,
j∈Z

where the convergence is in S 0 .

We wish to prove that the k(j) ’s satisfy the conditions (1)-(3) in Thm. 5.26.

We begin with (1). Using Lemma 6.17 we have

Z Z
ε
(1 + |x|) |k(j) (x)| dx = (1 + |x|)ε |F −1 mj (x)| dx

≤ Cε kmj kH s = Cε km(2−j ·)ψkH s

≤ Cε kmkMH s .

Next,
Z
k(j) (x) dx = Fk(j) (0) = mj (0) = 0 ,

since ψ(0) = 0; fact that proves (2).

Finally we prove (3). It suffices to show that28

Z
sup |h|−α |k(j) (x − h) − k(j) (x)| dx < ∞ .
0<|h|<1

28See footnote 20.

 HARMONIC ANALYSIS 81

Since s > n/2, using the Cauchy-Schwarz inequality we have that for any α ∈ (0, 1) and
0 < |h| < 1
Z
|h|−α |k(j) (x − h) − k(j) (x)| dx
Rn
Z 1/2
−α
≤ |h| |k(j) (x − h) − k(j) (x)|2 (1 + |x|2 )s dx
Rn
Z Z 1
2 1/2
= |h|−α h · ∇k(j) (x − th) dt (1 + |x|2 )s dx
Rn 0
Z Z 1 1/2
2
≤ |h|1−α ∇k(j) (x − th) dt(1 + |x|2 )s dx
Rn 0
Z Z 1 1/2
2
≤C ∇k(j) (x) (1 + |x + th|2 )s dt dx
Rn 0
Z 1/2
2
≤C ∇k(j) (x) (1 + |x|2 )s dx
Rn
n
X
≤C kF −1 (∂ξi k(j) )kH s ,
i=1
where we have used an estimate as in (6.8).
Thus, we only need to prove that
Xn
kF −1 (∂ξi k(j) )kH s ≤ CkmkMH s . (6.13)
i=1
Notice that
F −1 (∂ξi k(j) ) = 2πiξi F(k(j) )(ξ) = 2πiξi mj (ξ) = 2πiξi m(2j ξ)ψ(ξ)
= 2πiξi η(ξ) m(2j ξ)ψ(ξ) ,

where η ∈ C0∞ and it is identically 1 on the support of ψ. Applying Lemma 6.20 we have then
Xn Xn
−1
k ξi η m(2j ·)ψkH s


kF (∂ξi k(j) )kH s ≤ C
i=1 i=1
≤ Ckm(2j ·)ψkH s
≤ CkmkMH s ,
thus proving (10.4) and therefore the theorem. 2
82 M. M. PELOSO

7. Partial Differential Operators

In this section we apply the Fourier analysis techniques developed so far, to the study of
partial differential operators and equations. We begin by introducing a scale of spaces, called
Sobolev spaces and denoted by H s , that measure smoothness of functions, and that are suitably
defined using the Fourier transform.

7.1. The Sobolev spaces H s . Let k be a non-negative integer. We define H k = H k (Rn ) to

be the space of L2 functions f such that the distributional derivatives ∂ α f ∈ L2 for all α with
|α| ≤ k. On H k we can define an inner product that makes H k into a Hilbert space:
X Z
hf |gik = (∂ α f )(∂ α g) .29
|α|≤k

We wish to extend this definition to non-integral values of k and also to have a more “efficient”
way to represent the inner product. We observe that, by Plancherel theorem, f ∈ H k if and
only if ξ α fˆ ∈ L2 for all α, |α| ≤ k. We claim that there exist positive constants c1 , c2 such that,
for all ξ ∈ Rn , X
c1 1 + |ξ|2 )k/2 ≤ |ξ α | ≤ c2 1 + |ξ|2 )k/2 .
|α|≤k
For, if α is a multi-index with |α| ≤ k, if |ξ| ≥ 1 then
|ξ α | ≤ |ξ|k ≤ (1 + |ξ|2 )k/2 ,
while, if |ξ| ≤ 1,
|ξ α | ≤ 1 ≤ (1 + |ξ|2 )k/2 .
On the other hand, since |ξ|k and kj=1 |ξj |k are both homogeneous of degree k non-vanishing
P
for ξ 6= 0, we have
 X k  X
2 k/2 k 0
(1 + |ξ| ) ≤ c0 (1 + |ξ| ) ≤ c0 1 + c0 |ξj |k ≤ C |ξ α | .
j=1 |α|≤k

By the claim, it follows that f ∈ H k if and only if (1 + |ξ|2 )k/2 fˆ ∈ L2 . Moreover,

Z X Z
2 k ˆ
c1 2
1 + |ξ| ) |f (ξ)| dξ ≤ |ξ α |2 |fˆ(ξ)|2 dξ
|α|≤k
X Z
≤c |∂ α f (x)|2 dx
|α|≤k
Z
≤ c02 1 + |ξ|2 )k |fˆ(ξ)|2 dξ ,

so that the two norms

X Z 1/2 Z 1/2
α 2
|∂ f (x)| dx and 1 + |ξ| ) |fˆ(ξ)|2 dξ
2 k

|α|≤k

29We will keep the notation h·, ·i to indicate the bilinear pairing between a test function and a distribution,
while h·|·i will denote a hermitian inner product.
 HARMONIC ANALYSIS 83

are equivalent, and the latter one is defined for all k real, not just non-negative integer.

Definition 7.1. We define the Bessel potential of order s ∈ R of a distribution f ∈ S 0 as

 
−1 2 s/2 ˆ


Λs f = F 1 + |ξ| f .

Notice that Λs f is well defined for any f ∈ S 0 , since, for any Schwartz function ψ, if ϕ = F −1 ψ
hΛs f, ψi = hΛs f, ϕ̂i = h(1 + |ξ|2 )s/2 fˆ, ϕi
Z
= (1 + |ξ|2 )s/2 fˆ(ξ)ϕ(ξ) dξ = hfˆ, (1 + |ξ|2 )s/2 ϕi

and (1 + |ξ|2 )s/2 ϕ(ξ) is again a Schwartz function.

Definition 7.2. We define the Sobolev space H s of order s ∈ R as the space of distributions
f ∈ S 0 such that Λs f ∈ L2 , that is
Z
Hs = f ∈ S0 : |fˆ(ξ)|2 (1 + |ξ|2 )s dξ < ∞ .


we endow H s with the inner product

Z
hf |gis = (Λs f )(Λs g)

which also gives the norm

Z 1/2
kf kH s = |fˆ(ξ)|2 (1 + |ξ|2 )s dξ .

Proposition 7.3. The following properties are elementary.

(i) The Fourier transform F is a unitary isomorphism from H s to L2 (1 + |ξ|2 )s dξ .

(ii) The Schwartz space S is dense in H s for all s ∈ R.

(iii) If s > t then k · kH t ≤ k · kH s so that H s is a subspace of H t and it is also a dense
subspace.
(iv) The Bessel potential Λs is a unitary isomorphism from H t to H t−s for all s, t ∈ R. Its
inverse is Λ−s .
(v) H 0 = L2 , so that H s ⊂ L2 for all s > 0. For s < 0 the elements of H s may not be
functions.
(vi) The operators ∂ α are continuous from H s to H s−|α| , for all s and α.
Proposition 7.4. For s ∈ R, the L2 inner product h·|·i induces a unitary isomorphism between
H −s and the dual space (H s )∗ of H s .
Proof. Let f, ϕ ∈ S. Then, by Plancherel,
Z
hϕ|f i = hϕ̂|f i ≤ (1 + |ξ|2 )−s/2 |fˆ(ξ)|(1 + |ξ|2 )s/2 |ϕ̂(ξ)| dξ
ˆ
Z 1/2 Z 1/2
2 −s
≤ (1 + |ξ| ) |fˆ(ξ)|2 dξ 2 s 2
(1 + |ξ| ) |ϕ̂(ξ)| dξ

= kf kH −s kϕkH s .
84 M. M. PELOSO

Therefore, the linear functional ϕ 7→ hϕ|f i extends to all of H s with norm ≤ kf kH −s . But,
we actually have equality of norms, since if g = (Λ−2s f ),
Z
hf |gi = hf |ĝi = (1 + |ξ|2 )−s |fˆ(ξ)|2 dξ = kf k2H −s
ˆ

= kf kH −s kgkH s .

Notice that the pairing
hϕ̂|fˆi is well defined for all f ∈ H −s and ϕ ∈ H s .
∗
Finally, let L ∈ H s . Then L ◦ F
−1 is a bounded linear functional on L2 (1 + |ξ|2 )s dξ , so

that there exists g ∈ L2 (1 + |ξ|2 )s dξ such that

Z
L(ϕ) = L ◦ F −1 (ϕ̂) = ϕ̂(ξ)g(ξ)(1 + |ξ|2 )s dξ
Z


= ϕ̂(ξ)FΛ2s (F −1 g)(ξ) dξ

= hϕ̂|Λ2s (F −1 g)i
= hϕ|f i .
But, f = Λ2s (F −1 g) ∈ H −s (since F −1 g ∈ H s ). This proves the proposition. 2
We conclude this part with the Sobolev immersion theorems. We set
k
C(0) = {f ∈ C k (Rn ) : ∂ α f ∈ C(0) , for |α| ≤ k} .

Theorem 7.5. (Sobolev Embedding Theorem) Let s > k + (n/2). Then, H s embedds continu-
k .
ously in C(0)

Proof. Let f ∈ H s . Then,

Z Z Z
α |α| αˆ
|(∂ f ) b (ξ)| dξ = (2π) |ξ f (ξ)| dξ ≤ (1 + |ξ|2 )k/2 dξ
Z 1/2 Z 1/2
|α|
≤ (2π) |fˆ(ξ)|2 (1 + |ξ|2 )s dξ 2 k−s
(1 + |ξ| ) dξ
Z 1/2
|α| 2 k−s
= (2π) (1 + |ξ| ) dξ kf kH s .

Now, (1 + |ξ|2 )k−s dξ is finite exactly when 2(k − s) < −n, that is s > k + (n/2).
R

This argument shows that, when s > k + (n/2), (∂ α f ) b ∈ L1 . By the Fourier inversion
theorem, and the Reimann-Lebesgue theorem, it follows that (∂ α f ) ∈ C(0) . 2
This result can be made more precise, taking into account also fractional regularity. For
0 < α < 1 we define the Lipschitz space
Λα (Rn ) = f ∈ C ∩ L∞ (Rn ) : sup |f (x) − f (y)|/|x − y|α < ∞ .

x6=y

Then we define
k,α
(Rn ) = f ∈ C(0)
k
(Rn ) : ∂ β f ∈ Λα

C(0) for all β, |β| = k .
Then we have
 HARMONIC ANALYSIS 85

Theorem 7.6. (Sobolev Embedding Theorem, second version) Let 0 < α < 1 and s = (n/2) + α.
Then, H s embedds continuously in Λα . If s = (n/2) + k + α. Then, H s embedds continuously
k,α
in C(0) (Rn ).

Proof. Exercise (or see [H], Theorem 4.5.13). 2

7.2. Partial differential operators. A linear partial differential operator with smooth coeffi-
cients30 is an operator of the form
X
P f (x) = aα (x)∂xα f (x) .
|α|≤m

The smooth functions aα are called the coefficients of the operator P and m is called the order
of the operator. If the coefficients aα are all constants, then
X
(P f ) b (ξ) = aα (2πiξ)α fˆ(ξ) .
|α|≤m

Thus, in order to obtain a cleaner expression, we introduce the differential operator

D = (2πi)−1 ∂ so that Dα = (2πi)−|α| ∂ α . (7.1)
α,
P
In this way, if P = |α|≤m aα D then
X
(P f ) b (ξ) = aα (ξ)α fˆ(ξ) .
|α|≤m
α
P
We will often denote an operator P = |α|≤m aα D as p(D), where p is the polynomial
of degree m |α|≤m aα ξ α . The principal sympbol of P is defined to the top-order part of the
P

polynomial p: X
pm (ξ) = aα ξ α .
|α|=m
We now give two important definitions.
Definition 7.7. Let P = |α|≤m aα (x)Dα be a PDO with smooth coefficients on an open set
P

Ω. We say that P is hypoellptic on Ω if for every open set Ω0 ⊆ Ω and u ∈ D0 such that P u is
C ∞ on Ω0 , then u is also C ∞ on Ω0 .
Definition 7.8. Let P = |α|≤m aα (x)Dα be a PDO with smooth coefficients on an open set
P

Ω. We say that P is locally solvable in a point x0 ∈ Ω if there exists a neighborhood of x0 such

that for every f ∈ D there exists u ∈ D0 such that
Pu = f on U .
Given a PDO P with constant coefficients, we say that a distribution E is a fundamental
solution for P if P E = δ0 , the Dirac delta at the origin, that is if
hf, P Ei = h t P f, Ei = f (0)
for all f ∈ D.
Lemma 7.9. Let P be a PDO with constant coefficients. Then, the following are equivalents.
30Sometime it will be written PDO for short.
86 M. M. PELOSO

(i) P is locally solvable at a point x0 ∈ Rn ;

(ii) P is locally solvable at any point x0 ∈ Rn ;
(iii) P admits a fundamental solution.
Proof. It is a simple matter to see that a PDO P = |α|≤m aα (x)Dα has constant coefficients
P

if and only if it is invariant under translations. For, let y ∈ Rn , then

X
P (τy f )(x) = aα (x)Dα (τy f )(x)
|α|≤m
X
= aα (x)(Dα f )(x − y)
|α|≤m
X 
α
= τy (τ−y aα )(D f ) (x)
|α|≤m

and the above equals τy (P f ) if and only if, τ−y aα = aα . Now, it is easy to see that (i) and (ii)
are equivalent.
Next, if (iii) holds, let E be a fundamental solution for P . Given x0 ∈ Rn fix a neighborhood
U of x0 and a distribution u that we assume to have compact support (otherwise, replace u by
u0 = uϕ, where ϕ ∈ D, ϕ = 1 on U ). Let f = ψ(u ∗ E), where ψ ∈ D, ψ = 1 on U . Then


P f = P ψ(u ∗ E) = ψP (u ∗ E) + v
= ψ(u ∗ δ) + v
= ψu + v .
Notice that the distribution v is obtained when some of the derivatives falls on ψ. Since ψ = 1
on U , the support of v does not intersect U . Therefore, on U , P f = ψu = u; that is P is locally
solvable at x0 .
Finally, the proof that (ii) implies (iii) (i.e. the existence of a fundamental solution) requires
further notion of functional analysis and we refer to [H] for a proof. 2
We now show that all PDO’s with constant coefficients admits a fundamental solution, i.e.
are locally solvable.
Theorem 7.10. (Malgrange-Ehrenpreis) Let P = |α|≤m aα Dα be a constant coefficient PDO.
P

Then P admits a fundmental solution.

Let f ∈ D be given. Let P = p(D). If a tempered distribution u solves the equation P u = f ,
then
p(ξ)û(ξ) = fˆ(ξ) ,
We would like to define
fˆ(ξ)
Z
u(x) = e2πixξ dξ ,
p̂(ξ)
the problem being the zeros of p on the real space Rn . The proof is based on a deformation of
the integral to a path in Cn in order to avoid the zeros of p. We begin with a reduction and a
couple of lemmas.
By a rotation we may assume that the polynomial p has the leading term that does not vanish
at the point (0, . . . , 0, 1). By multiplying by a non-zero constant, we may assume that
p(ξ) = ξnm + lower order terms in ξn .
 HARMONIC ANALYSIS 87

Therefore, writing ξ = (ξ 0 , ξn ) with ξ 0 ∈ Rn−1 , we can find λj (ξ 0 ), j = 1, 2, . . . , m such that

m
Y
0
ξn − λj (ξ 0 ) .


p(ξ , ξn ) =
j=1

Lemma 7.11. There exists a measurable function φ : Rn−1 → [−m, m] such that
|φ(ξ 0 ) − λj (ξ 0 )| ≥ 1 .

inf min
ξ 0 ∈Rn−1 j=1,2,...,m

Proof. Let ξ 0 be fixed. The interval [−m − 1, m + 1] can be divided into the m + 1 intervals
[−m − 1, −m + 1], . . . , [m − 1, m + 1]. At least one of them does contain any of the values
Im λj (ξ), j = 1, 2, . . . , m. Define φ(ξ 0 ) to be the middle point of that interval. More precisely,
define
Uk = ξ 0 ∈ Rn−1 : Im λj (ξ 0 ) 6∈ [−m − 1 + 2k, −m + 1 + 2k), j = 1, 2, . . . , m .


The sets Uk cover Rn−1 and they are Borel sets since the functions Im λj are continuous. Thus,
set
φ(ξ 0 ) = −m + 2k for Uk \ ∪k−1
k=0 Ul
for 0 ≤ k ≤ m. 2

Lemma 7.12. Let h be a polynomial of one complex variable of degree m, with leading coefficient
1. Suppose h(0) 6= 0 and let λ1 , . . . , λm be its zeros. Then
|h(0)| ≥ (d/2)m
where d = minj |λj |.
Qm
Proof. We can write h(z) = j=1 (z − λj ) so that
m
h(z) Y z
= 1− ≤ 2m
h(0) λj
j=1

when |z| ≤ d. Notice that h(m) (z) is constant and equals m!. By Cauchy integral formula for
the derivatives
Z Z
m! h(z)
m! = h(m) (0) = dz
2πi |z|=d z m+1
Z 2π
m 1 1
≤ m!2 |h(0)| ddθ
2π 0 dm+1
m!2m
= m |h(0)| .
d
This proves the desired result. 2

Proof of Theorem 7.10. With φ as in the previous lemma set

fˆ(ξ)
Z Z
u(x) = e2πixξ dξn dξ .
Rn−1 Im ξn =φ(ξ 0 ) p(ξ)
88 M. M. PELOSO

We apply Lemma 7.12 to the polynomial h(z) = p(ξ 0 , ξn + z). Notice that h(z) = 0 when
z = λj (ξ 0 ) − ξn , j = 1, . . . , m. Then

|λj (ξ 0 ) − ξn | m
 
|p(ξ)| = |h(0)| ≥
2
0

m
Im λj (ξ ) − ξn

≥
2
λj (ξ ) − φ(ξ 0 ) m
0
 
=
2
≥ 2−m

when Im ξn = φ(ξ 0 ). By the Paley-Wiener Theorem we know that fˆ is an entire function that
decays rapidly when |Re ξ| → +∞, as long as Im ξ remains bounded. This shows that the
integral converges absolutely.
The same argument shows that we can differentiate under the integral sign as many times as
we like. Therefore, u ∈ C ∞ .
Now,
Z Z
P u(x) = e2πixξ fˆ(ξ) dξn dξ 0 .
Rn−1 Im ξn =φ(ξ 0 )

We use Cauchy integral theorem to deform the contour of integration to bring back the inte-
gration in ξn on the real axis. For, let ΓR be the contour in the complex ξn -plane given by the
rectangle of vertices = R, ξn = R + iIm φ(ξ 0 ), ξn = −R + iIm φ(ξ 0 ), and ξn = −R, with this
R ξn 2πixξ
orientation. Now, ΓR e fˆ(ξ) dξn = 0, by Cauchy theorem. If we show that the integrals over
the two vertical sides tend to 0 as R → +∞, we would obtain that
Z Z Z Z +∞
e2πixξ fˆ(ξ) dξn dξ 0 = e2πixξ fˆ(ξ) dξn dξ 0 .
Rn−1 Im ξn =φ(ξ 0 ) Rn−1 −∞

By Fourier inversion theorem the result now would follow. Thus, consider the integral over the
right vertical segment
Z Im φ(ξ0 ) Z Im φ(ξ0 )
0 0
e2πi(x ξ +xn (R+is) fˆ(ξ 0 , R + is)ds ≤ e−2πxn s |fˆ(ξ 0 , R + is)|ds
0 0
≤ me2π|xn |m cN (1 + R)−N

which tends to 0 as R → +∞.

Thus, we have shown that given f ∈ D there exists u ∈ C ∞ such that P u = f . We finally
construct a fundamental solution for P .
Define E by setting
fˆ(−ξ)
Z Z
hE, f i = dξn dξ 0 .
R n−1 0
Im ξn =φ(ξ ) p(ξ)
As before, we see that the integral converges absolutely. Moreover,

hP E, f i = hE, t P f i
 HARMONIC ANALYSIS 89

and notice that t P is the operator with symbol p(−ξ). Therefore, deforming the integral in ξn
as before,
( t P f ) b (−ξ)
Z Z
hP E, f i = hE, t P f i = dξn dξ 0
R n−1 0
Im ξn =φ(ξ ) p(ξ)
Z Z
= fˆ(−ξ) dξn dξ 0
Rn−1 0
Im ξn =φ(ξ )
Z
= fˆ(ξ)dξ
Rn
= f (0) = hδ0 , f i .
This finishes the proof of the theorem. 2
We conclude this section by characterizing the hypoelliptic PDO with constant coefficients
by the regularity of their fundamental solutions.
Proposition 7.13. Let P be a PDO with constant coefficients. Then, the following are equiva-
lent.
(i) P admits a fundamental solution that is in C ∞ (Rn \ {0});
(ii) every fundamental solution of P is in C ∞ (Rn \ {0});
(iii) P is hypoelliptic.
Proof. Suppose (iii) holds and let E be any fundamental solution. Then, from the equation
P E = δ0 we deduce that E ∈ C ∞ (Rn \ {0}) since δ0 is. Thus, (ii) holds. (ii) implies (i) trivially.
In order to show that (i) implies (iii) we need the following fact: If u and v are distributions (i.e.
elements of D0 ) with f ∈ C ∞ \ {0} and v with compact support, then sing supp (u ∗ v) ⊆ supp v.
Assuming the claim for now, we finish the proof. Let K be a fundamental solution of P which
is C ∞ (Rn \ {0}). Let g be a distribution such that P g is C ∞ on an open set Ω. Let x ∈ Ω and
let ϕ ∈ C0∞ (B(x, ε)) be ≡ 1 on B(x, ε/2). Then, P (ϕg) = ϕP g + v, where v ≡ 0 on B(x, ε/2)
and outside B(x, ε). We write
K ∗ P (ϕg) = K ∗ ϕP g + K ∗ v .
Now, ϕP g is a C0∞ function, so that K ∗ ϕP g is C ∞ . Also, K ∗ v is C ∞ on B(x, ε/2) by the claim.
Therefore, K ∗ P (ϕg) is C ∞ on B(x, ε/2). But,
K ∗ P (ϕg) = P K ∗ (ϕg) = ϕg .
Therefore ϕg is C∞ on B(x, ε/2), but ϕ ≡ 1 on B(x, ε/2), so that ϕg = g on that ball. Therefore,
g ∈ C ∞ on B(x, ε/2)
Finally, we prove the claim. Let x 6∈ supp v. Let ε > 0 small so that B(x, ε) ∩ supp v = ∅. Let
ψ ∈ C0∞ (B(0, ε/2)), ψ = 1 on B(0, ε/4). Then
u ∗ v = (ψu) ∗ v + [(1 − ψ)u] ∗ v .
But, (1 − ψ)u ∈ C ∞ , so that [(1 − ψ)u] ∗ v is C ∞ in Rn . Moreover,


supp (ψu) ∗ v ⊆ supp ψ + supp g
which does not intersect B(x, ε/2).
Hence, on B(x, ε/2) u ∗ v = [(1 − ψ)u] ∗ v and we are done. 2
90 M. M. PELOSO

8. †Littlewood–Paley theory
In order to present the Littlewood–Paley theory we need the extension of the singular integrals
to the case of vector-valued functions developed in Subsection 5.8.

8.1. An application. A simple application of the theory of vector-valued singular integrals is

~ is a constant sequence. In this
the following result. In this theorem the vector-valued kernel K
case it is easy to prove the boundedness of the corresponding vector-valued singular integral T~ .
Theorem 8.1. Let T be a convolution operator which is bounded L2 (Rn ) and whose integral
kernel satisfy the Hörmander condition (5.2). Let 1 < p, r < ∞. Then we have the strong
(p, p)-bound
X 1/r X 1/r
|T fj |r ≤ Cp,r |fj |r ,
p p
j j
and the weak (1, 1)-bound
n X 1/r o Cr  X 1/r
x ∈ Rn : |T fj |r >λ ≤ |fj |r .
λ L1
j j

Proof. We have stated the theorem in the case of the `r -norm, 1 < r < ∞. Since we stated and
proved Thm. 5.27 in the case of functions taking values in a Hilbert space (rather than in a
separable, reflexive, Banach space), here we consider only the case r = 2.
We only have to check that the hypotheses of Thm. 5.27 are satisfied.
In the current situation we consider functions f : Rn → `2 , that is, f (x) = {fj (x)}, where fj
are scalar valued functions. Then, we have that
T~ (f ) = T~ {fj } = K ∗ fj .



Then T~ (f ) = K
~ ∗ f , where
~
K(x) : `2 → `2
{aj } 7→ {K(x)aj } .
Hence, T~ : L2 (Rn , `2 ) → L2 (Rn , `2 ) is clearly bounded since, (as in the discussion prior to the
theorem) if f ∈ L2 (Rn , `2 ),
Z X Z X
~ 2
kT (f )kL2 (Rn ,`2 ) = 2
|K ∗ fj (x)| dx = |K̂(ξ)|2 |fˆj (ξ)|2 dξ
Rn j Rn j

≤ Akf k2L2 (Rn ,`2 ) .

Next,
Z Z
kK(x − y) − K(x)kL(`2 ,`2 ) dx = |K(x − y) − K(x)| dx ≤ B .
|x|>2|y| |x|>2|y|
The result now follows from Thm. 5.27. 
In order to illustrate the Littlewood–Paley theorem we consider a collection {Ej } of mutually
disjoint measurable sets in Rn and the operator S initially defined on Schwartz functions,
X 
Sf = F −1 (χEj fˆ) . (8.1)
j
 HARMONIC ANALYSIS 91

Then, it is immediate to see that S is bounded on L2 , since, by Plancherel’s theorem

Z XZ
2
X X
2
kSf kL2 = k ˆ 2
(χEj f )kL2 = ˆ
(χEj f )(ξ) dξ = |fˆ(ξ)|2 dξ = kf k2L2 . (8.2)
j Rn j j Ej

Starting from this simple observation we now state the following result. Notice that in this
theorem, we restrict ourselves to the 1-dimensional case.
Theorem 8.2. Let Ej = (−2j+1 , −2j ] ∪ [2j , 2j+1 ) ⊂ R and let Sj be defined as
Sj f ˆ(ξ) = χEj (ξ)fˆ(ξ) .

Then, for 1 < p < ∞ there exists C > 0 such that, for all f ∈ Lp (R)
1 X 1/2
kf kLp ≤ |Sj f |2 ≤ Ckf kLp .
C Lp
j

We will prove this theorem as a consequence of the Littlewood–Paley theorem, Thm. 8.5 in
the next section, that holds true also in Rn . In that theorem we will deal with an operator as
the one in (8.1). Notice that
X X
χEj (ξ) = χE0 (2−j ξ) = 1
j j

for all ξ ∈ R \ {0}. We will need a smooth decomposition of the function identically 1, and we
then recall Lemma 5.23: There exists a function ϕ ∈ C0∞ such that
X
(i) ϕ(2−j ξ) = 1 for all ξ 6= 0 .
j∈Z
Moreover, there exists another function ϕ0 ∈ C0∞ such that
+∞
X
(ii) ϕ0 (ξ) + ϕ(2−j ξ) = 1 for all ξ .
j=1
A consequence of the lemma is the following.
Corollary 8.3. There exists ψ ∈ S(Rn ) such that supp ψ ⊆ {ξ : 1/2 ≤ |ξ| ≤ 2} and
X
|ψ(2−j ξ)|2 = 1 for ξ 6= 0 . (8.3)
j∈Z

Proof. With ϕ as in the Lemma, we define Φ by setting (Φ̂)2 = ϕ so that

2
2
F(Φ2−j ) (ξ) = Φ̂(2−j ξ) = ϕ(2−j ξ) ,
and therefore, X X
|F(Φ2−j )(ξ)|2 = ϕ(2−j ξ) = 1
j∈Z j∈Z
for ξ 6= 0.
b 2
Then, it suffices to set ψ = Φ.
The next result is the first part of the Littlewood–Paley Thm. In a sense, it is the simplest,
and least interesting, part. It shows that a certain decomposition on the Fourier transform side
of a function gives rise to an operator that is bounded in Lp . It is more interesting to show that,
92 M. M. PELOSO

under an additional hypothesis, such an operator is also bounded from below– as we shall see
in Thm. 8.5.

Theorem 8.4. (Littlewood–Paley Thm., part 1) Let Φ̃ ∈ S(Rn ) be such that

X b̃
|Φ(2−j ξ)|2 ≤ C
j∈Z

˜ j f = f ∗ Φ̃2−j . Then, for 1 < p < ∞, there exists C > 0 such that
for all ξ 6= 0, and define ∆

X 1/2
˜ j f |2
|∆ ≤ Ckf kLp . (8.4)
Lp
j∈Z

Proof. If we set

T~ (f ) = {∆
˜ j f }j∈Z

if suffices to show that T~ satisfies the hypotheses of Thm. 5.27, when

T~ : Lp (Rn ) → Lp (Rn , `2 ) .

As we have seen already, T~ is bounded when p = 2. For,

Z Z X
kT~ f k2L2 (Rn ,`2 ) = kT~ (f )(x)k2`2 dx = |∆j (f )(x)|2 dx
Rn Rn j∈Z
XZ b̃ −j ξ)|2 |fˆ(ξ)|2 dξ
= |Φ(2
j∈Z Rn
Z
≤C |fˆ(ξ)|2 dξ
Rn
= Ckf k2L2 .

Thus, it will suffice to show that

~ C
k∇K(x)kL(C,`2 ) ≤ .
|x|n+1

Now,

~ = ∇Φ̃2−j (x) = ∇ 2nj Φ̃(2j x)

 

∇K(x)
= 2(n+1)j ∇Φ̃ (2j x) .


 HARMONIC ANALYSIS 93

Then, using the fact that Φ̃ is a Schwartz function, for any integer N > 0 there exists a constant
CN > 0 such that
2 1/2
X 
~ 2(n+1)j ∇Φ̃ (2j x)


k∇K(x)k `2 =
j∈Z
X
2(n+1)j ∇Φ̃ (2j x)


≤
j∈Z
X
2(n+1)j min 1, |2j x|−N


≤ CN
j∈Z
X X 
≤ CN 2(n+1)j + 2(n+1)j |2j x|−N
j≤j0 j>j0
X 1 X (n+1−N )j 
= CN 2(n+1)j + 2 ,
|x|N
j≤j0 j>j0

with j0 to be selected. Choosing j0 so that 2j0 ' |x|−1 , and N > n + 1 we have31,

~

(n+1)j0 2(n+1−N )j0  C
k∇K(x)k `2 ≤ CN 2 + N
≤ ,
|x| |x|n+1
as we wished to show. 2

8.2. The Littlewood–Paley theorem.

Theorem 8.5. (Littlewood–Paley theorem) Let ψ ∈ S(Rn ) be as in (8.3). Let Φ = F −1 ψ
and set
∆j f = f ∗ Φ2−j .
Then, for 1 < p < ∞, there exists C > 0 such that
1 X 1/2
kf kLp ≤ |∆j f |2 ≤ Ckf kLp .
C Lp
j

Proof. The bound from above follows directly from Thm. 8.4, where we use only the estimate
−j ξ)|2 ≤ C.
P
j∈Z |ψ(2
For the bound from below, we notice that
Z X 1/2
kf kL2 = |fˆ(ξ)|2 |ψ(2−j ξ)|2 dξ
Rn j∈Z
Z X 1/2
= |fˆ(ξ)ψ(2−j ξ)|2 dξ
Rn j∈Z
X 1/2
= |∆j f |2 .
L2
j

31Recall footnote 21.

94 M. M. PELOSO

Having the identity kf kL2 = kT~ (f )kL2 (Rn ,`2 ) , we can polarized it32, to obtain that for all f, g ∈
L2 (Rn ) Z Z X
f ḡ = ∆j f ∆j g .
j
Therefore,
Z Z X
kf kLp = sup f (x)g(x) dx = sup ∆j f (x)∆j g(x) dx
kgk 0 =1 kgk 0 =1 j
Lp Lp
Z X 1/2  X 1/2
≤ sup |∆j f (x)|2 |∆j g(ξ)|2 dx
kgk 0 =1 j j
Lp
Z X p/2 1/p  Z  X p0 /2 1/p0
2
≤ sup |∆j f (x)| dx |∆j g(ξ)|2 dx
kgk 0 =1 j j
Lp
Z X p/2 1/p
≤C |∆j f (x)|2 dx , (8.5)
j

where we have used the estimate from above. This proves the theorem. 2
We are now ready to prove Thm. 8.2. We recall that in this situation the space dimension is
n = 1.
Proof of Thm. 8.2. Let ψ ∈ C0∞ (R), supp ψ ⊆ {ξ : 1/2 ≤ |ξ| ≤ 4} and such that ψ(ξ) = 1 if
1 ≤ |ξ| ≤ 2, that is, if ξ ∈ E0 . Therefore,
−j
ψ 2 (ξ)χEj (ξ) = ψ(2−j ξ)χE0 (2−j ξ) = χE0 (2−j ξ) = χEj (ξ)
for all ξ. This implies that, if we define ∆˜ j f by setting
˜ j f ˆ(ξ) = ψ(2−j ξ)fˆ(ξ) ,


∆
then
˜ j Sj f = Sj f .
∆
Of course we can write ∆ ˜ j f = Φ̃2−j ∗ f , where Φ b̃ = ψ.
Notice that for any given ξ, since supp ψ ⊆ {ξ : 1/4 ≤ |ξ| ≤ 4}, there at most 3 indices
j1 , j2 , , j3 such that ψj (ξ) 6= 0. Therefore,
X
|ψ(2−j ξ)|2 ≤ C ,
j

for all ξ 6= 0.
We now make the following claim: For 1 < p < ∞, there exists C > 0 such that for all
g = {gj } ∈ Lp (Rn , `2 )
X 1/2 X 1/2
|Sj gj |2 p
≤ C |gj |2
p
= CkgkLp (Rn ,`2 ) . (8.6)
L L
j j

32Recall the identity 4hf, gi = kf + gk2 − kf − gk2 + ikf + igk2 − ikf − igk2 valid on any Hilbert space,
H H H H H
RP
and notice that in our case hf, giH = j ∆j f ∆j g.
 HARMONIC ANALYSIS 95

Assuming the claim we obtain the estimate from above, since

X 1/2 X 1/2
|Sj f |2 = |Sj
˜ j f |2
∆
pL Lp
j j
X 1/2
≤C ˜ j f |2
|∆
Lp
j
≤ Ckf kLp ,
by Thm. 8.4.
The estimate from below follows by polarizing the identity (8.2) and using duality, as in the
proof of Thm. 8.5, inequality (8.5), by replacing ∆j with Sj .
Thus, we only need to prove the claim.
~ : Lp (Rn , `2 ) → Lp (Rn , `2 ) defined as
We now consider the vector-valued operator S
~
 
S(g) = Sj (gj ) ≡ Kj ∗ gj

where Kj = F −1 (χEj ). For this operator we wish to prove the bound (8.6).
~ and its kernel satisfy the hypotheses of Thm. 5.27, we
Instead of proving that the operator S
proceed in a direct way.
Recall Example (6.5) (2). By (6.3) above we have that
Sj (gj )(x) = F −1 χ(−2j+1 ,−2j ] + χ[2j ,2j+1 ) ∗ gj (x)

ih i
= M−2j+1 HM2j+1 − M−2j HM2j + M2j HM−2j − M2j+1 HM−2j+1 (gj ) ,
2
so that





Sj (gj )(x) ≤ H M2j+1 gj (x) + H M2j gj (x) + H M−2j gj (x) + H M−2j+1 gj (x) .
Therefore, in order to prove the claim it suffices to show that each of the four operators on the
right hand side above satisfies (8.6), that is,
X
1/2 X 1/2
|H Maj gj |2 p
≤ C |gj |2
p
,
L L
j j

where {aj } is a sequence of real numbers.

Applying Thm. 8.1 we have that
X
1/2 X 1/2
|H Maj gj |2 ≤C |Maj gj |2
Lp Lp
j j
X 1/2
=C |gj |2 .
Lp
j

This proves the claim and therefore the theorem. 2

We conclude this part with an extension of Thm. 8.2 to n-dimensions.

96 M. M. PELOSO

The dyadic cubes in Rn can be seen as the cartesian product of dyadic intervals of the same
size; e.g. χ[0,1)n (x) = χ[0,1) (x1 ) · · · χ[0,1) (xn ). A dyadic rectangle R in Rn is the cartesian product
of dyadic intervals in Ij1 , . . . , Ijn of possibly different scales:
χR (x) = χIj1 (x1 ) · · · χIjn (xn ) .
We consider the operator
Sj11 f ˆ(ξ) = χIj1 (ξ1 )fˆ(ξ) ,

and more generally,

S~j f ˆ(ξ) = Sj11 · · · Sjnn f ˆ(ξ) = χIj1 (ξ1 ) · · · χIjn (ξn )fˆ(ξ) .

Corollary 8.6. Let f ∈ Lp (Rn ), 1 < p < ∞. Then there exists a positive constant Cp > 0 such
that
1  X 1/2
kf kLp ≤ |S~j f |2 ≤ Cp kf kLp ,
Cp Lp
~j ∈Zn

where {Ij1 }, . . . , {Ijn } are decompositions of the real line R into dyadic intervals of (possibly
different) scales 2k1 , . . . , 2kn , resp.
Proof. Notice that, if we set R~j , the collection {R~j }~j∈bZ n is a decomposition of Rn into mutually
disjoint dyadic rectangles.
The estimate from above follows at once from Thm. 8.4. The estimate from below follows
instead by applying Thm. 8.2 to the operators Sjkk , k = 1, 2, . . . , n.

8.3. The Marcinkiewicz multiplier theorem. We conclude with another multiplier theorem,
the Marcinkiewicz multiplier theorem, which is tightly connected with the theory of product
spaces.
The simplest form of the Mihlin-Hörmander condition is given by the inequality (6.11) for
k ≥ bn/2c + 1. We now consider the condition
∂ α m(ξ) ≤ C|ξ1 |−α1 · · · |ξn |−αn . (8.7)
Notice that when n = 1 it coincides with the Mihlin-Hörmander condition. But, when n > 1
the two conditions are different. In the theorem we give a more general condition for the Lp -
boundedness of the multiplier operator. It is easy to see that the condition is satisfied if (8.7)
holds, for k ≥ n.

Definition 8.7. Let s = (s1 , . . . , sn ) ∈ Rn . We define the product Sobolev space H s as the space
of tempered distributions f such that
Z
(1 + |ξ1 |2 )s1 · · · (1 + |ξn |2 )sn |fˆ(ξ)|2 dξ1 · · · dξn < ∞ .
Rn

Similarly to Lemma 6.17 one can easily prove the following.

Lemma 8.8. Let g ∈ S 0 be such that ĝ ∈ H s with si > 1/2 for i = 1, . . . , n. Then g ∈ L1 (Rn )
and for 0 < ε < si − 1/2 for all i
Z
(1 + |x1 |2 )ε · · · (1 + |xn |2 )ε |g(x)| dx ≤ Cε kĝkH s .
Rn
 HARMONIC ANALYSIS 97

Definition 8.9. For i = 1, . . . , n, let ηi ∈ C ∞ (R) be a function as in the hypothesis of Lemma

5.23 and set η(ξ) = η1 (ξ1 ) · · · ηn (ξn ). For r = (r1 , . . . , rn ) ∈ (R+ )n and ξ = (ξ1 , . . . , ξn ) ∈ Rn we
also set rξ = (r1 ξ1 , . . . , rn ξn ).
A tempered distribution m is said to be a Marcinkiewicz multiplier if
sup km(r·)ηkH s =: kmkMs < ∞ .
r∈(R+ )n

Theorem 8.10. (Marcinkiewicz multiplier theorem) Let m be a Marcinkiewicz multiplier

in H s , with si > 1/2 for i = 1, . . . , n. Then m is a bounded Fourier multiplier on Lp for
1 < p < ∞, that is, m ∈ Mp for 1 < p < ∞.
The next result follows easily from the theorem.
Corollary 8.11. Let m ∈ L∞ (Rn ) be such that there exists a constant B > 0 such that for all
0<k≤n Z
sup ∂ξkj ···ξjk m(ξ) dξj1 · · · dξjk ≤ B , (8.8)
1
Ij1 ,...,Ijk Ij1 ×···×Ijk
where Iji , i = 1, . . . , k, are dyadic intervals in R.
98 M. M. PELOSO

9. Symbols and pseudodifferential operators

In this section we introduce the notion of pseudodifferential operator. This turns out to be
of fundamental importance in the analysis of differential operators and their inverses. We begin
with the definition of symbol.
Definition 9.1. Let X, Y be open sets in Rn . For m a real number, we denote by S̃ m (X × Y )
the set of all a ∈ C ∞ (X × Y × Rn ) such that, for all KX , KY compact sets contained in X and
Y respectively, and all multi-indices α, β there exists a constant C = CKX ,KY ,α,β such that
β
∂x,y ∂ξα a(x, y, ξ) ≤ C(1 + |ξ|)m−|α| for (x, y, ξ) ∈ KX × KY × Rn .
m (X × Y ) the set of all
More generally, for 0 ≤ %, δ ≤ 1, one defines the class of symbols S%,δ
∞ n
a ∈ C (X × Y × R ) such that, for all KX , KY compact sets contained in X and Y respectively,
and all multi-indices α, β there exists a constant C > 0 such that
β
∂x,y ∂ξα a(x, y, ξ) ≤ C(1 + |ξ|)m−%|α|+δ|β| for (x, y, ξ) ∈ KX × KY × Rn .
But, in these notes we restrict ourselves to the classical, simpler case % = 1 and δ = 0.
It is immediate to see that,Pif m is a positive integer and aα (x, y) are smooth functions
α
of x and y, and a(x, y, ξ) := |α|≤m aα (x, y)ξ (that is, a is a polynomial in ξ), then a ∈
S̃ m (Rn × Rn × Rn ).
We now consider symbols of (only apparent) simpler nature.
Definition 9.2. For m a real number, we denote by S m the set of all a ∈ C ∞ (Rn × Rn ) such
that for all K compact in Rn and all multi-indices α, β there exists a constant C = Cα,β,K > 0
such that
∂xβ ∂ξα a(x, ξ) ≤ C(1 + |ξ|)m−|α| for (x, ξ) ∈ K × Rn .
We now define a pseudodiffential operator P with symbol a.
Definition 9.3. Let a ∈ S m . We define the pseudodifferential operator Pa with symbol a the
operator, initially defined on Schwartz functions, as
Z
Pa u(x) = a(x, ξ)e2πix·ξ û(ξ) dξ .

Notice that, since u ∈ S and for fixed x, |a(x, ξ)| ≤ C(1 + |ξ|)m , the integral above converges
absolutely.
Moreover, if, for ã ∈ S̃ m , we consider the application defined on D
x
P̃ã : u 7→ ã(x, y, ξ)e2πi(x−y)·ξ u(y) dydξ
the integral converges absolutely, for any fixed x, if m < −n, since then |ã(x, y, ξ)u(y)| ≤
C(1 + |ξ|)−n−ε |u(y)|, which is in L1 (Rn × Rn ). For generic m ∈ R, the double integral
x
e2πi(x−y)·ξ ã(x, y, ξ)u(y) dydξ
does not converge absolutely and it must be interpreted as an iterated integral in the following
way. Observe that, for all positive integers N ,
(I − ∆y )N
e2πi(x−y)ξ = e2πi(x−y)ξ ,


2 2
(1 + 4π |ξ| ) N
 HARMONIC ANALYSIS 99

so that, integrating by parts in the inner integral in the y-variable we obtain that
x
e2πi(x−y)·ξ ã(x, y, ξ)u(y) dydξ
x (I − ∆ )N
y 2πi(x−y)ξ


= e ã(x, y, ξ)u(y) dydξ
(1 + 4π 2 |ξ|2 )N
x e2πi(x−y)ξ N


= (I − ∆ y ) ã(x, y, ξ)u(y) dydξ . (9.1)
(1 + 4π 2 |ξ|2 )N
For N > 0 large enough, the integral above now converges absolutely.
Definition 9.4. Let ã ∈ S̃ m . We define the pseudodiffential operator Pã with symbol ã the
operator, initially defined on test functions (i.e. smooth functions with compact support), by
equation (9.1) above.

Notice that if the symbol ã ≡ a does not depend on y, then

x Z
Pã u(x) = ã(x, ξ)e2πi(x−y)·ξ u(y) dydξ = a(x, ξ)e2πix·ξ û(ξ) dξ = Pa u(x) ,

that is, the two definitions coincide in this case.

A few examples are now in order. Let P P be the partial differential operator with smooth
coefficients of order m, given by P f (x) = α
|α|≤m aα (x)Dx f (x). By the Fourier inversion
theorem
X Z
P f (x) = aα (x) e2πix·ξ ξ α fˆ(ξ) dξ
|α|≤m
Z X 
= e2πix·ξ aα (x)ξ α fˆ(ξ) dξ
|α|≤m
Z
= e2πix·ξ a(x, ξ)fˆ(ξ) dξ

where a(x, ξ) := |α|≤m aα (x)ξ α . Thus, P = Pa is a pseudo differential operator of order m

P

(that is the definitions of orders coincide).

P If the operator P has constant coefficients aα , then P = Pa where the symbol a(x, ξ) :=
α = a(ξ) is just a polynomial in ξ.
|α|≤m aα ξ

Definition 9.5. For m ∈ R we denote by Ψm the set of pseudodifferential operators with

symbol a ∈ S m .
Example 9.6. For each m ∈ R, the function σm (ξ) = (1 + |ξ|2 )m/2 defines and element of
S m (Rn ), so that Λm ∈ Ψm . Moreover, if P is a PDO (with C ∞ coefficients) of order k, then
P ∈ Ψk .
If a ∈ S m , then ∂xα a ∈ S m while ∂ξα a ∈ S m−|α| for any multiindex α.
The next result is rather obvious.
Lemma 9.7. If m1 < m2 are real numbers, then S m1 ⊆ S m2 . If a1 ∈ S m1 and a2 ∈ S m2 , then
a1 a2 ∈ S m1 +m2 and a1 + a2 ∈ S max(m1 ,m2 ) .
100 M. M. PELOSO

It is natural to then set

S ∞ = ∪m∈R S m , Ψ∞ = ∪m∈R Ψm and S −∞ = ∩m∈R S m , Ψ−∞ = ∩m∈R Ψm .
We mention that in general, if b is a measurable function, then the Fourier multiplier operator
Mb is defined as
Mb f (x) = F −1 b(ξ)fˆ(ξ) (x) .

Before presenting the next example, we prove the following simple fact.
Lemma 9.8. Let Mb be a Fourier multiplier operator as above. Then, the following conditions
are equivalent:
(i) b is bounded, i.e. b ∈ L∞ (Rn );
(ii) Mb : L2 (Rn ) → L2 (Rn ) is bounded;
(iii) for all s ∈ R, Mb : H s (Rn ) → H s (Rn ) is bounded.
Proof. By Plancherel’s theorem it is immediate to see that
Z Z
|Mb (f )(x)| dx = |b(ξ)fˆ(ξ)|2 dξ ≤ kbk2L∞ kf k2L2 ,
2

that is (i) implies (ii). It is also easy to see that if g 7→ bg is bounded on L2 , then b must be
bounded (Exercise V.2). By Plancherel again it follows then that (ii) implies (i).
Finally, if (i) holds, then
Z
kMb (f )ks = |b(ξ)fˆ(ξ)|2 (1 + |ξ|2 )s dξ
2

Z
≤ kbk2L∞ |fˆ(ξ)|2 (1 + |ξ|2 )s dξ

= Ckf k2s .
Hence, (i) implies (iii). And, if (iii) holds, also (ii) holds, so does (i) 2
Notice that this result holds true even without the assuption that b ∈ C ∞ , and thus without
any assumption on the behaviour of the derivatives of b.
Corollary 9.9. If m ≤ 0 and b = b(ξ) ∈ S m then Pb : H s → H s for all s ∈ R. If ψ ∈ S then
the mappings u 7→ ψu and u 7→ ψ ∗ u, initially defined on S, are in Ψ0 .
We wish to check that a pseudodifferential operator Pa : S → D0 boundedly. Let then ϕ ∈ S
and ψ ∈ D and consider hPa ϕ, ψi. Let K = supp ψ, then we have
x
|hPa ϕ, ψi| = e2πixξ a(x, ξ)ϕ̂(ξ) dξ ψ(x) dx
Z Z
≤ |ϕ̂(ξ)| |a(x, ξ)| |ψ(x)| dx dξ
Z Z
≤ CK |ϕ̂(ξ)|(1 + |ξ|)m |ψ(x)| dx dξ
X
≤C kϕk(α,n+1) kψk(0,n+1) , (9.2)
|α|≤N

for some N large enough.

 HARMONIC ANALYSIS 101

More is true. Let E 0 denote the space of distributions having compact support. That is,
u ∈ D0 (or in S 0 ) is in E 0 if there exists a compact set K such that hψ, ui = 0 for all ψ ∈ D such
that supp ψ ∩ K = ∅.
Then, if a ∈ S m , then Pa : E 0 → D0 continuously. Now we have a definition that will be used
in what follows.
Definition 9.10. A linear operator T : E 0 → C ∞ is called a smoothing operator.

9.1. The kernel of a pseudodifferential operator. If T is an integral operator initially

defined on S having values in S 0 , (or initially defined on D having values in D0 )
Z
T ϕ(x) = K(x, y)ϕ(y) dy ,
Rn

then, for ψ also in S we have that, if the double integral makes sense,
x
hT ϕ, ψi = K(x, y)ϕ(y)ψ(x) dydx ,

or, in general,
hT ϕ, ψi = hK, ϕ ⊗ ψi .
The distribution K on Rn × Rn is called the distributional kernel of the operator T . In this
setting, recall also the so-called Schwartz kernel theorem, Theorem ??
We wish to calculate the distributional kernel of a pseudodifferential operator. Let a ∈ S m ,
ϕ ∈ S and ψ ∈ D, then, using (9.2), we have
x
hPa ϕ, ψi = e2πixξ a(x, ξ)ϕ̂(ξ)ψ(x) dξdx .

Now, let χ ∈ C0∞ , χ = 1 in a neighborhood of the origin. Then set aε (x, ξ) = χε (ξ)a(x, ξ), so
that aε has compact support in ξ, while tends to a as ε → 0. Then,
y
hPa ϕ, ψi = lim e−2πi(y−x)ξ aε (x, ξ)ϕ(y)ψ(x) dydξdx
ε→0
y
= lim e−2πi(y−x)ξ χ(εξ)a(x, ξ) dξ ϕ(y)ψ(x) dydx
ε→0
x 
= lim χ̂ε ∗ F2 a(x, ·) (y − x) ϕ(y)ψ(x) dydx
ε→0
x
= F2 a(x, y − x)ϕ(y)ψ(x) dydx ,

where F2 a(x, ·) denotes the Fourier transform of a(x, ξ) in the second variable. (Notice that this
is well defined, since if a ∈ S m , a(x, ·) is C ∞ and of moderate growth.)
Theorem 9.11. Let a ∈ S m and let K(x, y) = F2 a(x, y − x) be its distributional kernel. Then,
(i) if |α| > m+n+k, the distribution uα = z α F2 a(x, z) is in fact a function in C k (Rn ×Rn )
and its derivatives of order ≤ k are bounded in K × Rn , for all compact K ⊆ Rn ;
(ii) if |α| > m + n + k, the distribution (x − y)α F2 a(x, y − x) is in fact a function in
C k (Rn ×Rn ), in particular K ∈ C ∞ (Rn ×Rn \4R2n ), where 4R2n = {(x, y) ∈ Rn ×Rn :
x = y}.
102 M. M. PELOSO

Proof. By the discussion above, it suffices to prove (i). Now, the distribution uα is the Fourier
transform in the second variable of Dξα a(x, ξ). If |α| > m + n + k, for x varying in a compact
set K, |Dξα a(x, ξ)| ≤ C(1 + |ξ|)m−|α| , which is integrable. Thus we can apply Proposition 2.11
and the result follows. 2

Corollary 9.12. If P ∈ Ψ−∞ , then P is a smoothing operator.

Proof. By the previous theorem it follows that the distributional kernel of P is in fact C ∞ (Rn ×
Rn ). Then, for u a distribution with compact support we have
P u(x) = hK(x, ·), ui .
By (i) in the Theorem, we can differentiate in x under the pairing of duality (which can be
thought to be an integral, by approximating u with elements of D) to obtain that
Dxα P u(x) = hDxα K(x, ·), ui .
This proves the result. 2

Definition 9.13. We say that a subset Ω of Rn × Rn is proper if for every compact set K ⊂ Rn
πx−1 (K) ∩ Ω and πy−1 (K) ∩ Ω are compact sets. Here we denote by πx and πy the projections of
Rn × Rn onto its first and second component, respectively.
A pseudodifferential operator Pa is said to be properly supported if its distributional kernel
K(x, y) is properly supported.
An important fact, that we leave without proof (see [Fo2]) is that the composition of two
properly supported operators is still properly supported.
Proposition 9.14. Let ã = ã(x, y, ξ) ∈ S̃ m . Then, the pseudodifferential operator Pã can be
written as sum of a properly supported and of a smoothing pseudodifferential operators.
Proof. Let χ be a test function, supported on the unit ball and identically 1 on the ball of radius
1/2. Write the symbol pseudodifferential operator as sum


ã(x, y, ξ) = χ(x − y)ã(x, y, ξ) + 1 − χ(x − y) ã(x, y, ξ) ≡ ã1 (x, yξ) + ã∞ (x, y, ξ) .
This gives the desired decomposition. 2
Notice that if ã(x, y, ξ) is supported in a neigborhood of the diagonal 4R2n , then, for x and
ξ fixed the function y 7→ ã(x, y, ξ) is C ∞ with compact support. Then, Pã f is well defined for
any C ∞ function f . We will make use of this remark in the next result.
The notion of properly supported pseudodifferential operators allows us to make a reduction.
This is a somewhat technical result, which is however of some significance.
Theorem 9.15. Let ã ∈ S̃ m and Pã properly supported. Set
b(x, ξ) = e−2πixξ Pã e2πi(·)ξ (x) .

Then b ∈ S m and Pã = Pb modulo Ψ−∞ .

Proof. We begin by arguing as in the discussion preceeding Theorem 9.11. Let χ ∈ C0∞ , χ = 1
in a neighborhood of the origin. Then set ãε (x, y, η) = χε (η)a(x, y, η), so that ãε has compact
support in η, while tends to ã as ε → 0 (as a tempered distribution in η, depending smoothly
 HARMONIC ANALYSIS 103

in x, y). Recall also that, for fixed x, ã(x, y, η), and so also ãε (x, y, η), has compact support in
y. Let Pãε be the corresponding operator. Then, for a given ϕ ∈ S,
x
Pãε ϕ(x) = e2πi(x−y)η ãε (x, y, η)ϕ(y) dydη
x Z
2πi(x−y)η
= e ãε (x, y, η) e2πiyξ ϕ̂(ξ) dξ dydη
Z  x 
2πixξ −2πixξ 2πi(x−y)η 2πiyξ
= e ϕ̂(ξ) e e ãε (x, y, η)e dydη dξ
Z  
2πixξ −2πixξ 2πi(·)ξ


= e ϕ̂(ξ) e Pãε e (x) dξ .

Passing to the limit as ε → 0+ , we obtain that

Pã ϕ(x) = lim Pãε ϕ(x)
ε→0+
Z  
2πixξ −2πixξ 2πi(·)ξ


= lim e ϕ̂(ξ) e Pãε e (x) dξ
ε→0+
Z  
= e2πixξ ϕ̂(ξ) e−2πixξ Pã e2πi(·)ξ (x) dξ

Z
= e2πixξ ϕ̂(ξ)b(x, ξ) dξ ,

where Pã u must be interpreted as in Definition 9.4.

Thus, we only need to show that
b(x, ξ) = e−2πixξ Pã e2πi(·)ξ (x)


x
= e−2πixξ e2πi(x−y)η ã(x, y, η)e2πiyξ dydη
is an element of S m (where, again, the last double integral above is an iterated integral that
must be must be interpreted in the sense of Definition 9.4).
Now we set σ(x, y, η) = ã(x, x + y, η), so that
x
b(x, ξ) = e2πi(x−y)(η−ξ) ã(x, y, η) dydη
x
= e2πi(x−y)(η−ξ) σ(x, y − x, η) dydη
x
= e−2πiz(η−ξ) σ(x, z, η) dydη
Z
= F2 σ(x, η − ξ, η) dη .

Hence, Z
b(x, ξ) = F2 σ(x, η, η + ξ) dη . (9.3)
For x varying in a fixed compact set K, y 7→ σ(x, ·, η) is C ∞ with compact support so that
for any N > 0,
∂xβ ∂ζα F2 σ(x, η, ζ) ≤ CK,α,β,N (1 + |ζ|)m−|α| (1 + |η|)−N . (9.4)
Next we apply the simple fact that
(1 + |x|2 )s (1 + |y|2 )−s ≤ 2|s| (1 + |x − y|2 )|s| (9.5)
104 M. M. PELOSO

for all x, y ∈ Rn and s ∈ R, that follows from the triangle inequality

(1 + |x|2 ) ≤ 2(1 + |x − y|2 )(1 + |y|2 ) .
Then, by applying (9.5) with x = ξ + η and y = ξ to (9.4), we obtain
∂xβ ∂ξα F2 σ(x, η, η + ξ) ≤ CK,α,β,N (1 + |ξ + η|)m−|α| (1 + |η|)−N
m−|α| −N
≤ CK,α,β,N (1 + |ξ|)m−|α| (1 + |η|) .
Inserting this estimate in (9.3) we obtain
Z
m−|α| −N
∂xβ ∂ξα b(x, ξ) ≤ CK,α,β,N (1 + |ξ|)m−|α| (1 + |η|) dη

≤ CK,α,β (1 + |ξ|)m−|α| ,
for all x ∈ K and ξ ∈ Rn . This shows that b ∈ S m . 2

9.2. Sobolev continuity of a pseudodifferential operator. We now wish to study the

mapping properties of pseudodifferential operators. The nature of the definition of symbol
offers no control in the behaviour of the symbol itself, and consequentely of the action of a
pseudodifferential operator on a test function (say) Pa ϕ(x) when the output variable x tends to
infinity.
For this reason we need to restrict our attention to the local behaviour of Pa ϕ. We then have
the following definition.
Definition 9.16. Let s ∈ R. We set
s
= f ∈ S 0 : ϕf ∈ H s for all ϕ ∈ C0∞ .

Hloc
s−m
The next result shows that a pseudodifferential operator of order m maps H s into Hloc
continuously, for all s, m ∈ R.
Theorem 9.17. Let P be a pseudodifferential operator of order m. Then, for all s ∈ R,
s−m
Pa : H s → Hloc is continuous, that is, for each χ ∈ C0∞ there exists a constant C = Cχ,s,m > 0
such that
kχPa ϕkH s−m ≤ CkϕkH s .
s , we may just assume that a(x, ξ) has compact
Proof. Since we wish to show that Pa (ϕ) is in Hloc
support in the x-variable– it suffices to multiply a by the cut-off function χ– and prove that
Pa : H s → H s−m
for all s ∈ R. Thus, we assume that a has compact support in x throughout the rest of the
proof.
We begin by making a reduction. We may assume that Pa is of order 0. In fact, suppose
a ∈ S m and define b(x, ξ) = (1 + |ξ|2 )−m/2 a(x, ξ). Then b ∈ S 0 as it is immediate to check (and
still has compact support in x). Moreover,
Z
Pa (ϕ)(x) = e2πix·ξ b(x, ξ)(1 + |ξ|2 )m/2 ϕ̂(ξ) dξ = Pb (Λm ϕ)(x) .

If we had shown the H s -boundedness of operators of order 0, then it would follow that
kPa (ϕ)kH s−m = kPb (Λm ϕ)kH s−m ≤ ckΛm ϕkH s−m ≤ ckϕkH s .
 HARMONIC ANALYSIS 105

Thus, we assume that a ∈ S 0 . Since a has compact support in x, we may compute its Fourier
transform in x. We set Z
F1 a(t, ξ) = e−2πit·x a(x, ξ) dx ,

so that Z
a(x, ξ) = e2πit·x F1 a(t, ξ) dt .

An integration by parts now shows that for all multi-indices α

Z
(2πit)α F1 a(t, ξ) = ∂x a(x, ξ) e−2πit·x dx .
 α 

Therefore,
sup |F1 a(t, ξ)| ≤ CN (1 + |t|)−N (9.6)
ξ

for arbitrary N ≥ 0.
Notice that for ϕ ∈ S
Z Z
−2πixξ
F(Pa ϕ)(ξ) = e e2πixη a(x, η)ϕ̂(η) dηdx
Z
= F1 a(ξ − η, η)ϕ̂(η) dη .

Finally, let s ∈ R. Then, using the estimate (9.6) above, Hölder inequality, (9.4) and choosing
N large enough, we have
Z Z
2 2
kPa (ϕ)kH s = F1 a(ξ − η, η)ϕ̂(η) dη (1 + |ξ|2 )s dξ
Z Z 2
≤ CN (1 + |ξ − η|2 )−N |ϕ̂(η)| dη (1 + |ξ|2 )s dξ
Z Z
≤ CN |ϕ̂(η)| 2
(1 + |ξ − η|2 )−N (1 + |ξ|2 )s dξdη
Z Z
≤ CN |ϕ̂(η)| (1 + |η| ) dη (1 + |ξ − η|2 )−N +|s| dξdη
2 2 s

≤ Ckϕk2H s ,
by choosing N > |s| + n, as we may. Here we have used (9.5) with x = ξ and y = η.
This concludes the proof. 2

9.3. Asymptotic expansion of symbols. We now present the notion of symptotic expansion
of our symbols.
Definition 9.18. LetPaj ∈ S mj , mj ∈ R, j = 0, 1, 2, . . . . We say that a symbol a has the
asymptotic expansion j aj , and we write
X
a∼ aj ,
j
106 M. M. PELOSO

if for every M > 0 there exists NM > 0 such that

NM
X
a− aj ∈ S −M .
j=0

Proposition 9.19. Let aj ∈ S mj , j = 0, 1, 2, . . . , where mj & −∞. Then there exists a symbol
a ∈ S m0 , unique modulo symbols in S −∞ (and we will usually say, with an abuse of language,
modulo smoothing operators), such that
+∞
X
a∼ aj .
j=0

Proof. Let θ be a C ∞ function that is identically 0 when |ξ| ≤ 1 and identically 1 when |ξ| ≥ 2.
Let Ωj be an increasing sequence of bounded open set tending to the whole of Rn , say Ωj =
B(0, Rj ), with Rj % +∞.
Let t0 = 1 < t1 < t2 < · · · be an increasing sequence of positive numbers tending to +∞, to
be selected later. We set
+∞
X
a(x, ξ) = θ(ξ/tj )aj (x, ξ) . (9.7)
j=0

Notice that θ(ξ/tj ) = 0 when |ξ| ≤ tj , so that, for each (x, ξ) fixed, only finitely many terms in
the series are non-zero. Therefore, a(x, ξ) is a well-defined C ∞ function.
Now we claim that we can find a sequence {tj } in such a way that
∂xβ ∂ξα θ(ξ/tj )aj (x, ξ) ≤ 2−j (1 + |ξ|)mj−1 −|α|


(9.8)
for |x| ≤ Rj and |α| + |β| ≤ j.
Assuming the claim for the moment, we show that the function a defined in (9.7) has the
required properties. Having fixed a compact set K ⊂ Rn , let k be such that K ⊂ B(0, Rk ) and
let |α| + |β| ≤ k. We use the facts that, for j ≤ k aj ∈ S mj , and for j > k the claim above, to
obtain that
k
X +∞
X
|∂xβ ∂ξα a(x, ξ)| ∂xβ ∂ξα ∂xβ ∂ξα θ(ξ/tj )aj (x, ξ)



≤ θ(ξ/tj )aj (x, ξ) +
j=0 j=k+1
k
X +∞
X
≤ Cj (1 + |ξ|)mj −|α| + 2−j (1 + |ξ|)mj−1 −|α|
j=0 j=k
m0 −|α|
≤ Ck (1 + |ξ|) + C(1 + |ξ|)mk−1 −|α| (9.9)
m0 −|α|
≤ C(1 + |ξ|) .
Therefore, a ∈ S m0 . We now need to show that a admits the right asymptotic expansion. We
can write
k−1
X  k−1
X k−1
 X 


a(x, ξ) − aj (x, ξ) = a(x, ξ) − θ(ξ/tj )aj (x, ξ) + θ(ξ/tj ) − 1 aj (x, ξ)
j=0 j=0 j=0
=: b(x, ξ) + r(x, ξ) .
 HARMONIC ANALYSIS 107

By the previous argument (9.9) we see that

k−1
X +∞
X
b(x, ξ) = a(x, ξ) − θ(ξ/tj )aj (x, ξ) = θ(ξ/tj )aj (x, ξ)
j=0 j=k

is a well-defined C ∞ function and it satisfied the differential inequalities

|∂xβ ∂ξα b(x, ξ)| ≤ C(1 + |ξ|)mk −|α| ,

that is, b ∈ S mk . On the other hand, notice that

supp(1 − θ(ξ/tj )) ⊆ {ξ : |ξ| ≤ 2tj } ⊆ {ξ : |ξ| ≤ 2tk } (9.10)

for j = 0, . . . , k − 1. Then r(x, ξ) has compact support in ξ, that is, r ∈ S −∞ . These two facts
imply that
k−1
X
a(x, ξ) − aj (x, ξ) ∈ S mk ,
j=0

as we wished to show.
In order to show that this asymptotic expansion is unique (that is, thatPa is unique modulo
smoothing terms), let b(x, ξ) be another symbol in S m such that b ∼ j aj . Then, given
0

M > 0, let k be such that mk < −M so that

 k−1
X   k−1
X 
a(x, ξ) − b(x, ξ) = a(x, ξ) − aj (x, ξ) − b(x, ξ) − aj (x, ξ) ∈ S mk .
j=0 j=0

Hence, a − b ∈ S mk for all k, that is, a − b ∈ S −∞ , i.e. a = b modulo S −∞ , as we wished to

show.
It only remains to prove the claim (9.8). We begin by observing that (9.10) also gives that
θ(ξ/tj ) = 1 for all ξ with |ξ| ≥ 2tj and all j. Then, if |α| > 0, ∂ξα θ(ξ/tj ) has compact support
for all j, and, in fact, |ξ| ≤ 2tj for ξ ∈ supp ∂ξα θ(ξ/tj ), for all j. Hence,

1
|α|
≤ C(1 + |ξ|)−|α|
tj

for ξ ∈ supp ∂ξα θ(ξ/tj ) and for all j, where the constant C is independent of j. Thus,

1
|∂ξα θ(ξ/tj )| = |α|
|(∂ξα θ)(ξ/tj )|
tj
≤ C(1 + |ξ|)−|α| sup |(∂ξα θ)(ξ)|
ξ
−|α|
≤ Cα (1 + |ξ|) ,

where the constant Cα is again independent of j. Notice that the above estimate holds true also
when α = 0.
108 M. M. PELOSO

Therefore33, for |α| + |β| ≤ j and x ∈ B(0, Rj ),

X α γ
β α
∂ξ θ(ξ/tj )∂xβ ∂ξα−γ aj (x, ξ)


∂x ∂ξ θ(ξ/tj )aj (x, ξ) =
γ
γ≤α
X α 
≤ Cγ (1 + |ξ|)−|γ| Cα,γ (1 + |ξ|)mj −(|α|−|γ|)
γ
γ≤α

≤ Cj (1 + |ξ|)mj −|α| ,
for some constant Cj = max{Cα,β : |α| + |β| ≤ j}, |x| ≤ Rj , |α| + |β| ≤ j. Recall that θ(ξ/tj )
is supported in {|ξ| ≥ tj } and that mj & −∞. Inductively, having chosen tj−1 we select tj in
such a way that
|ξ| ≥ tj implies Cj (1 + |ξ|)mj ≤ 2−j (1 + |ξ|)mj−1 .
This establish the claim and prove the proposition. 2

9.4. Adjoints and products of pseudodifferential operators. In order to compute the

adjoint of a pseudodifferential operator we need to rely on the more general definition of symbol
ã(x, y, ξ) given in Definition 9.1.
Theorem 9.20. Let ã ∈ S̃ m and Pã properly supported. Let
b(x, ξ) = e−2πixξ Pã e2πi(·)ξ (x)

be as in Theorem 9.15. Then b admits asymptotic expansion

∞
X
b(x, ξ) ∼ bj (x, ξ) ,
j=0

where
X 1
bj (x, ξ) = ∂ α Dα ã(x, y, ξ) .
α! ξ y y=x
|α|=j

Proof. From Theorem 9.15 we know that b is well defined and is in S m . We need to show that
b has the prescribed asymptotic expansion.
We use the notation from the proof of Theorem 9.15. From equation (9.3) we know that
Z
b(x, ξ) = F2 σ(x, η, η + ξ) dη ,

and from (9.4) that

∂xβ ∂ξα F2 σ(x, η, ξ) ≤ CK,α,β,N (1 + |ξ|)m−|α| (1 + |η|)−N ,
for x ∈ K, K compact, and any N > 0.
Observe that this inequality implies that, for |α| = j,
Z Z
β α0 m−j−|α0 |
∂x ∂ξ α α
∂ξ F2 σ(x, η, ξ)η dη ≤ CK,α,α0 ,β,N (1 + |ξ|) (1 + |η|)−N |η|j dη
0
≤ C(1 + |ξ|)m−j−|α | , (9.11)
33We use the product formula for higher order partial derivatives.
 HARMONIC ANALYSIS 109

that is, Z
∂ξα F2 σ(x, η, ξ)η α dη ∈ S m−j for |α| = j .

Next, we expand the function ξ 7→ F2 σ(x, η, ξ):

X ∂ξα F2 σ(x, η, ξ)
F2 σ(x, η, ξ + η) − ηα
α!
|α|<k

≤ ck sup |η|k ∂ξα F2 σ(x, η, ξ + tη)

|α|=k, 0≤t≤1

≤ CK,k,N sup |η|k (1 + |ξ + tη|)m−k (1 + |η|)−N .

0≤t≤1

If |η| < 21 |ξ| (taking N = k) we have

X ∂ξα F2 σ(x, η, ξ)
F2 σ(x, η, ξ + η) − η α ≤ C 0 (1 + |ξ|)m−k ,
α!
|α|<k

and if |η| ≥ 12 |ξ| (taking N large) we have

X ∂ξα F2 σ(x, η, ξ)
F2 σ(x, η, ξ + η) − η α ≤ C”(1 + |η|)−n−k .
α!
|α|<k

Integrating, we have
Z X 1 Z
F2 σ(x, η, ξ + η) dη − ∂ξα F2 σ(x, η, ξ)η α dη
α!
|α|<k
X 1 Z
= b(x, ξ) − ∂ξα F2 σ(x, η, ξ)η α dη
α!
|α|<k
Z Z
≤ C0 (1 + |ξ|)m−k dη + C” (1 + |η|)−n−k dη
|η|< 21 |ξ| |η|≥ 12 |ξ|
 
≤ Ck (1 + |ξ|)m+n−k + (1 + |ξ|)−k

≤ Ck (1 + |ξ|)mk .
Notice that mk & −∞, as k → +∞.
A repeatition of the same argument shows that, for all multi-indices α, β and K compact,
there exists C = Cα,β,K such that, for all x ∈ K, ξ ∈ Rn ,
X 1 Z
β α
∂x ∂ξ b(x, ξ) − β α
∂ ∂ ∂ξα F2 σ(x, η, ξ)η α dη ≤ C(1 + |ξ|)mk −|α| .
α! x ξ
|α|<k

Therefore, since mk & −∞, given any M > 0, we can find k = k(M ) in such a way that
mk < −M and
k−1 X Z
X 1
b(x, ξ) − ∂ξα F2 σ(x, η, ξ)η α dη ∈ S −M .
α!
j=0 |α|=j
110 M. M. PELOSO

Thus,
+∞ X Z
X 1
b(x, ξ) ∼ ∂ξα F2 σ(x, η, ξ)η α dη .
α!
j=0 |α|=j
Finally, notice that
Z  Z 
α α α 2πiyη α
∂ξ F2 σ(x, η, ξ)η dη = Dy e ∂ξ F2 σ(x, η, ξ) dη
y=0
= Dyα ∂ξα σ(x, y, ξ) y=0
 

= Dyα ∂ξα a(x, y, ξ) y=x .

 

This proves the theorem. 2

Definition 9.21. Let a ∈ S m , Pa be the corresponding pseudodifferential operator. Assume

that Pa is properly supported. Then, we define the transpose t Pa and the adjoint Pa∗ resp., as
hPa ϕ, ψi = hϕ, t Pa ψi , and hPa ϕ|ψi = hϕ|Pa∗ ψi , resp. .
Lemma 9.22. Let a ∈ S m . Then the transpose t Pa of Pa is the pseudodifferential operator
Pσ̃ where σ̃(x, y, ξ) = a(y, −ξ), while the adjoint Pa∗ is the pseudodifferential operator Pτ̃ where
τ̃ (x, y, ξ) = a(y, ξ).
Proof. We only present the case of t Pa since the other case is essentially identical.
For ϕ, ψ ∈ D (recall that Pa ϕ is only C ∞ , so we need at least ψ to be smooth and with
compact support to define the pairing and have convergent integrals), we write
x
hPa ϕ, ψi = e2πixξ a(x, ξ)ϕ̂(ξ)ψ(x) dξdx
Z
= ϕ̂(ξ)g(ξ) dξ
Z
= ϕ(y)ĝ(y) dy = hϕ, ĝi ,

where g is given by Z
g(ξ) = e2πixξ a(x, ξ)ψ(x) dx
and so g, and therefore ĝ too, is a Schwartz function. For, g is clearly C ∞ . Moreover, for every
α,
Z  
α
ξ g(ξ) = Dxα e2πixξ a(x, ξ)ψ(x) dx
Z  
= (−1)|α| e2πixξ Dxα a(x, ξ)ψ(x) dx ,

and this last integral is bounded since ψ has compact support. Similar reasoning shows that all
the derivatives of g decay rapidly.
Therefore,
x
t
Pa ψ(y) = ĝ(y) = e2πi(x−y)ξ a(x, ξ)ψ(x) dxdξ
x
= e2πi(y−x)ξ a(x, −ξ)ψ(x) dxdξ ,
 HARMONIC ANALYSIS 111

and the lemma is proved. 2

Theorem 9.23. Let a ∈ S m and suppose Pa is properly supported. Denote by aP 0 and aP ∗ the
symbols of t Pa and Pa∗ resp. Then, aP 0 , aP ∗ ∈ S m and we have
X (−1)|α|
aP 0 (x, ξ) ∼ Dxα ∂ξα a(x, −ξ)
α!
|α|≥0
X 1
aP ∗ (x, ξ) ∼ Dα ∂ α a(x, ξ) .
α! x ξ
|α|≥0

Proof. Now this follows at once. Let’s consider the case of a t P . By the lemma we know that
t P = P , where σ̃(x, y, ξ) = a(y, −ξ). Since P is properly supported, so is P . By Theorems
a σ̃ a σ̃
9.20 and 9.15 we then have that, modulo smoothing operators, Pσ̃ = Pa t P , where
X (−1)|α|
aP 0 (x, ξ) ∼ Dxα ∂ξα a(x, −ξ) .
α!
|α|≥0

This proves the theorem. 2

0 0
Theorem 9.24. Let a ∈ S m , b ∈ S m . Then Pa Pb ∈ Ψm+m and its symbol σ = σPa Pb has
asymptotic expansion
X 1
σ(x, ξ) ∼ ∂ α a(x, ξ)Dxα b(x, ξ) .
α! ξ
|α|≥0

Proof. Since P = (P 0 )0 , we can write

x
P ϕ(x) = e2πi(x−y)ξ aP 0 (y, −ξ)ϕ(y) dydξ ,
so that Z
Pcϕ(ξ) = e−2πiyξ aP 0 (y, −ξ)ϕ(y) dyd .
Therefore,
Z x
Pb Pa ϕ(x) = e2πixξ b(x, ξ)Pcϕ(ξ) dξ = e2πi(x−y)ξ b(x, ξ)aP 0 (y, −ξ)ϕ(y) dydξ .

This means that Pb Pa = Pτ̃ , where

τ̃ (x, y, ξ) = b(x, ξ)aP 0 (y, −ξ) .
Clearly τ̃ ∈ S̃ m+m0
. We obtain the asymptotic expansion by applying Theorem 9.15. Modulo
smoothing operators, Pτ̃ = Pσ , where σ has asymptotic expansion
X 1
∂ξα Dxα b(x, ξ)aP 0 (y, −ξ)
 
σ(x, ξ) ∼ .
α! x=y
|α|≥0

The fact that

X 1 X 1
∂ξα Dxα b(x, ξ)aP 0 (y, −ξ) ∂ α a(x, ξ)Dxα b(x, ξ)
 
∼
α! x=y α! ξ
|α|≥0 |α|≥0

is a (fairly) elementary verification that is left to the reader. 2

112 M. M. PELOSO

0
Corollary 9.25. Let Pa , Pb be properly supported, with a ∈ S m , b ∈ S m . Then, if σ denotes
the symbol of Pa Pb we have that
0
σ(x, ξ) = a(x, ξ)b(x, ξ) mod S m+m −1 .

9.5. Local regularity of elliptic operators. We now introduce the classical notion of an
elliptic partial differential operator.
Definition 9.26. Let P = |α|≤m aα (x)Dα be a PDO with smooth coefficients. We say that
P

P is elliptic at x0 if pm (x0 , ξ) 6= 0 for all ξ ∈ Rn \ {0}, that is if

X
aα (x0 )ξ α 6= 0, for all ξ ∈ Rn \ {0} .
|α|=m

Notice that, in this case there exists a constant C0 > 0 such that
X
aα (x0 )ξ α ≥ C0 |ξ|m ,
|α|=m

(where C0 can be taken to be the minimum of the left hand side on the sphere |ξ| = 1).
The operator P is said to be elliptic on an open set Ω if it is elliptic at every point x ∈ Ω.
In analogy with the classical definition, we introduce the notion of elliptic symbols.
Definition 9.27. Let a ∈ S m , where m is a real number, and let Pa be the corresponding
pseudodifferential operator. We say that P is elliptic if, for every compact set K ⊂ Rn there
exist positive constants C1 , C2 such that
|a(x, ξ)| ≥ C1 |ξ|m ,
for all x ∈ K and ξ ∈ Rn with |ξ| ≥ C2 .
Now a technical lemma.
Lemma 9.28. Let a ∈ S m be elliptic. Then, there exists χ ∈ C ∞ (Rn × Rn ) such that, for all
compact set K there exist constant c, C > 0 such that
(i) χ(x, ξ) = 1 for |ξ| ≥ C;
(ii) |a(x, ξ)| ≥ c|ξ|m when χ(x, ξ) 6= 0.
Proof. Exercise V.6. 2

Definition 9.29. Let P ∈ Ψ∞ be a pseudodifferential operator. We say that a properly

supported pseudodifferential operator Q is a parametrix for P if,
P Q − I, QP − I ∈ Ψ−∞ .
Theorem 9.30. If P ∈ Ψm is elliptic, then P has a parametrix Q ∈ Ψ−m .
s , then u ∈ H m+s . In particular, if
Moreover, if P is properly supported, u ∈ E 0 and P u ∈ Hloc loc
P u ∈ C ∞ (Ω) for some open set Ω, then u ∈ C ∞ (Ω).
 HARMONIC ANALYSIS 113

Proof. Let P = Pa where a ∈ S m is elliptic, and let χ be as in Lemma 9.28. We set

(
χ(x,ξ)
if χ(x, ξ) 6= 0
b0 (x, ξ) = a(x,ξ)
0 if χ(x, ξ) = 0 .
It is clear that b0 is well defined. Indeed, having fixed a compact set K ⊂ Rn , letting x vary
−1
in K we have that |a(x, ξ)| ≥ cK |ξ|m when χ(x, ξ) 6= 0. Moreover, b0 (x, ξ) = a(x, ξ)

when
|ξ| ≥ CK , since χ(x, ξ) = 1 when |ξ| ≥ CK .
For the time being, we work on a fixed compact set K ⊂ Rn and let x ∈ K. All the constants
that appear will depend on K and the equality holds for x ∈ K, even if not explicitely indicated.
To complete the proof it will then suffice to repeat the argument for compact sets Kj % Rn .
Now we wish to estimate ∂xβ ∂ξα b0 (x, ξ) as |ξ| → +∞. Thus, it suffices to compute the above
derivatives for |ξ| ≥ CK .
Notice that, avoiding to write the arguments of the various functions, ∂ξj a−1 = −a−2 ∂ξj a,
∂ξj ξk a−1 = a−3 2∂ξj a∂ξk a + a∂ξ2j ξk a .
2
 

Inductively, one can easily show that

−1  −1−|α|
∂ξα a(x, ξ)

= a(x, ξ) A(x, ξ) ,
where A is a finite sum of the form
∂ξγ1 a(x, ξ) · · · ∂ξγk a(x, ξ) ,
where |γ1 | + · · · + |γk | = |α|. Notice that necessarly k ≤ |α|. Analogous formula holds for the
partial derivatives in x.
We now restrict ourselves to the case m ≥ 0. The case m < 0 can be obtained along the same
lines. We leave the easy checking to the reader.
Then, using the fact that [a(x, ξ)]−1 ≤ C(1 + |ξ|)−m (or ≤ C(1 + |ξ|)−|m| in the general case)
for |ξ| ≥ CK and the above remark, we have
−1
∂xβ ∂ξα b0 (x, ξ) = ∂xβ ∂ξα a(x, ξ)

X
≤ C(1 + |ξ|)−m(1+|α|) (1 + |ξ|)m−|γ1 | · · · (1 + |ξ|)m−|γk |
|γ1 |+···+|γk |=|α|
−m(1+|α|)
≤ C(1 + |ξ|) (1 + |ξ|)m|α|−|α|
≤ C(1 + |ξ|)−m−|α| .
This shows that b0 ∈ S −m .
Moreover,
(
χ(x, ξ) if χ(x, ξ) 6= 0
b0 (x, ξ)a(x, ξ) =
0 if χ(x, ξ) = 0
= 1 − r0 (x, ξ) ,
where r0 ∈ S −∞ as it is easy to check, using the fact that supp r0 ⊂ ∪K%Rn K × {|ξ| ≤ CK }.
Now, it follows from Theorem 9.24 that the symbol of the operator Pb0 Pa , denote it by σPb0 Pa ,
satisfies
σPb0 Pa (x, ξ) = b0 (x, ξ)a(x, ξ) − r10 (x, ξ) ,
where r10 ∈ S −1 , that is, σPb0 Pa = 1 − r1 with r1 ∈ S −1 .
114 M. M. PELOSO

Next we set (
χ(x,ξ)r1 (x,ξ)
a(x,ξ) if χ(x, ξ) 6= 0
b1 (x, ξ) =
0 if χ(x, ξ) = 0 .
Then an argument similar as the one above shows that b1 (x, ξ) ∈ S −m−1 and that
b1 (x, ξ)a(x, ξ) = (1 − χ(x, ξ))r1 (x, ξ) = r1 (x, ξ) − r20 (x, ξ) ,
where r20 ∈ S −2 .
Again, denote by σ(Pb0 +Pb1 )Pa the symbol of (Pb0 + Pb1 )Pa . By Theorem 9.24 we have that



σ(Pb0 +Pb1 )Pa (x, ξ) = 1 − r1 (x, ξ) + r1 (, ξ) − r2 (x, ξ)
= 1 − r2 (x, ξ) ,
where r2 ∈ S −2 .
Inductively, having picked k, one constructs bj with j ≥ 2 with bj ∈ S m−j for j < k in such a
way that
σ(Pb0 +Pb1 +···+Pb )Pa (x, ξ) = 1 − rj+1
j

where rj+1 ∈ S −j−1 ,

for all j < k.
Then, by the asymptoticP expansion of symbols Theorem 9.15 we obtain there exists a properly
supported b such that b ∼ j bj , b ∈ S −m and such that, for all k,
σPb Pa = σ(Pb0 +Pb1 +···+Pb )Pa (x, ξ) = 1 − rk+1 ∈ S −k−1 .
k

This proves that Pb Pa − I ∈ Ψ−∞ ,

With the same procedure, we can construct b0 ∈ S m such that Pa Pb0 = I modulo Ψ−∞ .
But then also,
Pa Pb − I = (Pa Pb0 − I) − Pa Pb0 + Pa Pb Pa Pb0 + Pa Pb − Pa Pb Pa Pb0
= (Pa Pb0 − I) + Pa (Pb Pa − I)Pb0 − Pa Pb (Pa Pb0 − I)
∈ Ψ−∞ .
Then Pb is a parametrix for Pa . This proves the first part of the theorem.
The second and third part follow immediately from Theorem 9.17 (Exercise V.7). 2
As a corollary we also obtain the following classical result.
The elliptic partial differential operators enjoy of the so-called local regularity property. More
precisely, they satisfy the following apriori local estimate, which follows from Theorem 9.30.
Theorem 9.31. Let Ω be a bounded open set and let P = |α|≤m aα (x)Dα be a PDO with
P

smooth coefficients, elliptic in a neighborhood of Ω. Then, for every s ∈ R there exists a

constant C = Cs > 0 such that for all u ∈ Cc∞ (Ω) we have


kukH s ≤ Cs kP ukH s−m + kukH s−1 .
Proof. Exercise V.8. 2
 HARMONIC ANALYSIS 115

10. Sum of squares of vector fields

We now turn our attention to operators with variable coefficients that are not elliptic, but
share some regularity properties with them.
Finally, we study the famous Hörmander’s theorem on the hypoellipticity of sums of squares
of vector fields.
We fix an open set Ω in Rn . A vector field (with smooth coefficients) is a first order PDO
with no constant term
Xn
X= aj (x)∂xj .
j=1

We will (typically) assume that the coefficients aj are real-valued. (In case we are interested in
complex-valued coefficients, we will refer to them as complex-vector fields.)
We consider the PDO P defined as
k
X
P = Xj2 + X0 + b(x) , (10.1)
j=1

where X0 , . . . , Xk are (real-valued) vector fields and b is a smooth real-valued function.

Definition 10.1. The commutator of two vector fields X = nj=1 aj (x)∂xj and Y = nj=1 bj (x)∂xj
P P

is the vector field, denoted by [X, Y ], defined by the action on a smooth function f
[X, Y ](f ) = (XY − Y X)(f ) .

It is easy to check that indeed (XY − Y X) is a PDO of first order, with no constant term
and that
n
X  Xn 
[X, Y ](f )(x) = (XY − Y X)(f )(x) = X bj (x)∂xj f (x) − Y aj (x)∂xj f (x)
j=1 j=1
n
X n
X



= ak (x) ∂xk bj (x) ∂xj f (x) − bk (x) ∂xk aj (x) ∂xj f (x)
j,k=1 j,k=1
X n Xn 
= ak (x)∂xk bj (x) − bk (x)∂xk aj (x) ∂xj f (x) ,
j=1 k=1

since the second order terms cancel out.

Definition 10.2. We denote by X1 the collection of the Xj ’s, j = 0, 1, . . . , k and inductively, by

Xm the set of Xm−1 and the collection of the commutators [X, Y ], where X ∈ X1 and Y ∈ Xm−1 .
We say that the operator P as in (10.1) is of finite type of order m in Ω if Xm spans the whole
tangent space of Rn at every point in Ω.

Theorem 10.3. (Hörmander) Let P be the operator in (10.1). Suppose that P is of finite
type at every point in Ω. Then P is hypoelliptic in Ω.

This result follows by the next, more precise, theorem.

116 M. M. PELOSO

Theorem 10.4. (Hörmander) Let P be the operator in (10.1). Suppose that Xm spans the
tangent space at every point in Ω. Then for every s ∈ R and N > 0 there exists a constant Cs,N
such that


kukH s+ε ≤ Cs,N kP ukH s + kukH−N ,
for every u ∈ Cc∞ (Ω), where ε = 21−m .

PnTypical2
examples of operators to which these theorems apply are the heat operator P =
2 2 2 2
j=1 ∂xj − ∂t and the Grushin operator in R , P = ∂x + x ∂y . Another classical example is
the so-called sub-Laplacian L on the Heisenberg group. The HeisenbergP group H is identified
with Rn × Rn × R, with coordinates (x, y, t), and the operator L = − 12 nj=1 Xj2 + Yj2 , where
Xj = ∂xj − y2 ∂t , Xj = ∂yj + x2 ∂t , and T = ∂t . Then, it is easy to see that [Xj , Yj ] = −T , so that
the Hörmander condition is satisfied with m = 2.
Pn
We write Xj = `=1 aj` ∂x` , and by shrinking the domain Ω, we may assume that b and
aj,` ∈ C ∞ (Ω) for all j, `.

Lemma 10.5. Let P be defined as in (10.1). Then there exists C > 0 such that
k
X
kXj uk2 ≤ C |(P u|u)| + kuk2 ,

j=1

for all u ∈ Cc∞ (Ω).

Proof. We denote by Xj∗ the adjoint of Xj (w.r.t. the L2 -inner product). It is easy to check that
(recall that the Xj are real-valued) Xj∗ = −Xj + hj , where hj = − n`=1 ∂x` aj` . Integrating by
P
parts we see that

−(Xj2 u|u) = −(Xj u|Xj∗ u) = (Xj u|Xj u) − (Xj u|hj u)

≤ kXj uk2 + CkXj uk kuk
= kXj uk2 + O kXj uk kuk ,

while
(X0 u|u) = −(u|X0 u) + O(kuk2 ) ,
so that
Re (X0 u|u) = O(kuk2 ) .
Therefore,
k
X k
X
2
(Xj2 u|u) + O kXj uk kuk


kXj uk = −
j=1 j=1
k
X 
2
= −Re (P u|u) + O kXj uk kuk + kuk . (10.2)
j=1
 HARMONIC ANALYSIS 117

2
Using the estimate ab ≤ ε2 a2 + 2ε12 b2 , valid for all a, b ∈ R, ε > 0, we see that
X k  Xk
O kXj uk kuk ≤ C kXj uk kuk
j=1 j=1
 k 
C 2X 2 1 2
≤ ε kXj uk + 2 kuk
2 ε
j=1
k
Cε2 X
= kXj uk2 + C 0 kuk2 .
2
j=1

By choosing ε > 0 suitably small, we can take part of the right hand side of (10.2) on the left
and obtain the result. 2

Lemma 10.6. Let L be a PDO operator with smooth coefficients of order `, s ∈ R. Then the
operators LΛs and Λs L are psuedodifferential operators of order ` + s, while the commutator
[Λs , L] is a psuedodifferential operator of order ` + s − 1, in the sense that
t t−(`+s)
LΛs , Λs L : Hloc → Hloc ,
while
t t−(`+s−1)
[Λs , L] : Hloc → Hloc
for all t ∈ R.
Proof. This follows at once from the results in Section 5. 2

Theorem 10.7. If Xm spans the tangent space at every point in Ω, then there exist ε > 0 and
C > 0 such that
X k 
2 2 2
kukH ε ≤ C kXj uk + kuk ,
j=0
for all u ∈ Cc∞ (Ω). Here we can take ε = 21−m .
Proof. We denote by Zm a generic element of Xm . Then,
k
X   X 
2 2 2 2 2
kukH ε ≤ C kDj ukH ε−1 + kuk ≤ C kZm ukH ε−1 + kuk ,
j=0 Zm ∈Xm

where the last sum is over finitely many terms. Therefore, it suffices to show that, for each
Zm ∈ Xm
Xk 
2 2 2
kZm ukH ε−1 ≤ C kXj uk + kuk ,
j=0
for some ε > 0.
If m = 1 we can take ε = 1. Let then m ≥ 2 and assume the estimate valid for m − 1 and
ε = 21−(m−1) . We may assume that Zm = XZm−1 − Zm−1 X, with X ∈ X1 . Then,
kZm uk2H ε−1 = (Zm u|Λ2ε−2 Zm u)
= (XZm−1 u|Λ2ε−2 Zm u) − (Zm−1 Xu|Λ2ε−2 Zm u) , (10.3)
118 M. M. PELOSO

and we estimate these last two terms separately.

By Lemma 10.6, using the fact that 2ε − 1 ≤ 0, we have


|(XZm−1 u|Λ2ε−2 Zm u)| = |(Zm−1 u|XΛ2ε−2 Zm u)| + O |(Zm−1 u|Λ2ε−2 Zm u)|


= |(Zm−1 u|XΛ2ε−2 Zm u)| + O |(Λ2ε−1 Zm−1 u|Λ−1 Zm u)|


= |(Zm−1 u|XΛ2ε−2 Zm u)| + O kZm−1 ukH 2ε−1 kuk


= |(Zm−1 u|Λ2ε−2 Zm Xu)| + O kZm−1 ukH 2ε−1 kuk
≤ C kXuk2 + kZm−1 uk2H 2ε−1 + kuk2 .

Next,


|(Zm−1 Xu|Λ2ε−2 Zm u)| = |(Xu|Zm−1 Λ2ε−2 Zm u)| + O kukkXuk
≤ C kXuk2 + kZm−1 uk2H 2ε−1 + kuk2 .

Substituting into (10.3) and using the induction hypothesis, by taking ε = 21−m , we obtain
X k 
2 2 2 2
kZm ukH ε−1 ≤ C kXj uk + kZm−1 ukH 2ε−1 + kuk
j=0
k
X 
≤C kXj uk2 + kZm−1 uk2 1−(m−1) −1 + kuk2
H2
j=0
k
X 
≤C kXj uk2 + kuk2 .
j=0

This proves the theorem. 2

We now prove
Theorem 10.8. If Xm spans the tangent space at every point in Ω, then there exist ε > 0 and
C > 0 such that
kuk2H ε ≤ C kP uk2 + kuk2 ,

for all u ∈ Cc∞ (Ω). Here we can take ε = 21−m .

Proof. The proof is quite simple in the case that P = kj=1 Xj2 , i.e. there is no first order term
P
X0 and no zero order term b in P .
Under these assumptions it suffices to combine Lemma 10.5 and Theorem 10.7:
X k 
2 2 2
kukH ε ≤ C kXj uk + kuk
j=0

≤ C |(P u|u)| + kuk2

≤ C kP uk2 + kuk2 .

This proves the theorem in the special case P = kj=1 Xj2 .

P

For the general case we refer to [?] Thm. 8.2.6. 2

We now “boot-strap” the argument of the previous theorem to prove
 HARMONIC ANALYSIS 119

Theorem 10.9. Let Xm span the tangent space at every point in Ω. Then, for every s ∈ R and
N > 0 there exists C = Cs,m,N > 0 such that

kuk2H s+ε ≤ C kP uk2H s + kuk2H −N ,

for all u ∈ Cc∞ (Ω). Here, again, ε = 21−m .

Proof. We wish to apply Thm. 10.8 to Λs u. We assume that the estimates in Thm. 10.8 holds
for functions u in Cc∞ (Ω0 ), with Ω0 open, Ω0 ⊆ Ω.
Let ϕ, ϕ1 be smooth cut-off functions supported in Ω, ϕ = 1 on Ω0 and ϕ1 = 1 on supp ϕ.
We now claim that for each s ∈ R and N > 0 there exists C > 0 such that for all v ∈ S

k(1 − ϕ1 )Λs (ϕv)k ≤ CkvkH −N .

This depends on the fact that the distribuitions 1 − ϕ1 and ϕv have disjoint supports, so that
(1 − ϕ1 )Λs (ϕv) ∈ Ψ−∞ .
Next, we assume the claim. Notice that u = ϕu (since (1 − ϕ)u = 0). We then have

kukH s+ε = kΛs ukH ε

≤ kϕ1 Λs ukH ε + k(1 − ϕ1 )Λs (ϕu)kH ε


≤ C kP (ϕ1 Λs u)k + kϕ1 Λs uk + kukH −N


≤ C k[P, ϕ1 Λs ]u)k + kP ukH s + kukH s + kukH −N . (10.4)

Notice that [P, ϕ1 Λs ] is a pseudodifferential operator that can be written as

k
X
[P, ϕ1 Λs ] = Qsj Xj + Qsk+1 (10.5)
j=1

where we denote by Qs , Qsj etc. a generic element in Ψs . For, by Thm. 9.24 we have
k
hX i
[P, ϕ1 Λs ] = Xj2 + X0 + b, ϕ1 Λs
j=1
k
X  2
Xj , ϕ1 Λs + Qs ,

=
j=1

where Qs ∈ Ψs with the convention fixed above. Now,

 2
Xj , ϕ1 Λs = Xj2 ϕ1 Λs − ϕ1 Λs Xj2 = Xj2 ϕ1 Λs − Xj ϕ1 Λs Xj + Xj ϕ1 Λs Xj − ϕ1 Λs Xj2

 
= Xj [Xj , ϕ1 Λs ] + [Xj , ϕ1 Λs ]Xj = [Xj , ϕ1 Λs ]Xj + Xj , [Xj , ϕ1 Λs ] + [Xj , ϕ1 Λs ]Xj
s
= Qsj Xj + Q0 j .

This proves (10.5). As a consequence,

k
X 
k[P, ϕ1 Λs ]uk ≤ C kXj ukH s + kukH s . (10.6)
j=1
120 M. M. PELOSO

Next, applying Lemma 10.5 and using (10.6) we have,

kXj uk2H s = kΛs Xj uk2 ≤ C kϕΛs Xj uk2 + k(1 − ϕ)Λs ϕ1 Xj uk2

≤ C kXj ϕΛs uk2 + k[Xj , ϕΛs u]k2 + kuk2H −N

 
≤ C P (ϕΛs u)|ϕΛs u + kuk2H s + kuk2H −N


 
≤ C ϕΛs P (u)|ϕΛs u + [P, ϕΛs ](u)|ϕΛs u + kuk2H s + kuk2H −N

 k
X 
kXj kH s + kukH s kukH s + kuk2H s + kuk2H −N


≤ C kP ukH s kukH s +
j=1
k
X 
kXj kH s kukH s + C kP uk2H s + kuk2H s + kuk2H −N .


≤C (10.7)
j=1
Therefore, using the “small constant-big constant” argument we obtain,
X k Xk 
2
kXj kH s kukH s + C kP uk2H s + kuk2H s + kuk2H −N


kXj ukH s ≤ C
j=1 j=1
k
X
kXj k2H s + C kP uk2H s + kuk2H s + kuk2H −N ,


≤ Cε
j=1
so that,
k
X

kXj ukH s ≤ C kP ukH s + kukH s + kukH −N ,
j=1
for some suitable large constant C > 0. Inserting the above estimate and the one from (10.6)
into (10.4) we see that


kukH s+ε ≤ C kP ukH s + kukH s + kukH −N . (10.8)
In order to conclude the proof it suffices to show that for any δ > 0 there exists a constant
C = Cε,s,N,δ such that
kukH s ≤ δkukH s+ε + CkukH −N ,
and then choose δ > 0 sufficiently small in order to bring the term δkukH s on the left hand side
of (10.8). This concludes the proof of the theorem. 2
 HARMONIC ANALYSIS 121

Appendix A. Some results on Lp spaces

We begin by recalling a few basic facts about the Lp spaces in Rn .

For proofs and further details of the facts and ideas presented here, we refer to [WZ], or also
[Fo1] and [Ru], for example.
We denote by Lp = Lp (Rn ) the space of Lebesgue measurable functions, which are p-integrable
functions, if 0 ≤ p < ∞, and for p = ∞ essentially bounded. These norms are
Z 1/p
p
kf kp = |f (x)| dx
Rn
kf k∞ = inf {x ∈ Rn : |f (x)| > α} = 0 .
α≥0
It is clear that Lp are vector spaces, since
|f + g|p ≤ 2p |f |p + |g|p .
 

In order to prove that k · kp is indeed a norm we need to prove the triangle inequality. We
begin by showing that when 0 < p < 1 instead the triangle inequality fails. For, if A and B are
disjoint sets of positive, finite measure, then, since 0 < p < 1,

1/p
kχA + χB kp = |A| + |B| > |A|1/p + |B|1/p = kχA kp + kχB kp .
We now prove some basic inequalities concerning the Lp -norms. For p ≥ 1 we denote by p0 its
conjugate exponent (i.e. 1/p + 1/p0 = 1, which gives p0 = p/(p − 1) if p > 1, p = 1 and p0 = ∞
being conjugate exponents).
Theorem A.1. (Hölder’s inequatlity) Let 1 < p < ∞, and let p0 be its conjugate exponent. If f
and g are measurable functions, we have
kf gk1 ≤ kf kp kgkp0 .
0
Equality holds if and only if α|f |p = β|g|p for some constants α and β.
Proof. The result is obvious if either kf kp = 0 or kgkp0 = 0. In fact, in this case either f = 0
a.e. or g = 0 a.e., which gives f g = 0 a.e.
The result is also clear if either kf kp = ∞ or kgkp0 = ∞.
Next we claim that for a, b ≥ 0 and 0 < λ < 1 then
aλ b1−λ ≤ λa + (1 − λ)b . (A.1)
For, suppose b = 0, then the inequality holds. If b 6= 0, setting t = a/b, inequality (A.1) becomes
tλ ≤ λt + (1 − λ) .
Let ϕ(t) = tλ − λt, we obtain ϕ0 (t) = λ t(λ−1) − 1 , which is > 0 for 0 < t < 1 and < 0 when

t > 1. Thus, the maximum is attained when t = 1 with value 1 − λ. This proves (A.1) and
equality holds only if t = 1, i.e. a = b.
Now suppose that both kf kp and kgkp0 are different from 0. We use the claim above with

p
p 0
a = |f (x)|/kf kp , b = |g(x)|/kgkp0 and λ = 1/p .
We obtain
|f (x)||g(x)| 1 |f (x)| 1 |g(x)|
≤ + 0 .
kf kp kgkp0 p kf kp p kgkp0
122 M. M. PELOSO

Integrating both sides of the inequality gives

Z
1 1 1
|f (x)||g(x)| dx ≤ + 0 = 1 .
kf kp kgkp0 p p
This proves the inequality. Finally equality holds if and only if a = b, i.e.

p
p 0
|f (x)|/kf kp = |g(x)|/kgkp0 ,
0
that is α|f |p = β|g|p . 2

Theorem A.2. (Minkowski’s inequatlity) Let 1 ≤ p ≤ ∞, and let p0 be its conjugate exponent.
If f and g are measurable functions, we have
kf + gkp ≤ kf kp + kgkp0 .
Proof. The result is clear if p = 1 or p = +∞, and also if f + g = 0.
Suppose that 1 < p < ∞, and f + g 6= 0 a.e. Then
|f + g|p ≤ |f | + |g| |f + g|p−1

and using Hölder’s inequality with conjugate exponents p and p0 we obtain

Z
|f + g|p dx ≤ kf kp |f + g|p−1 p0 + kgkp |f + g|p−1 p0

= kf kp + kgkp |f + g|p−1 p0

Z 1/p0
p


= kf kp + kgkp |f + g| dx .

Since 1/p0 = 1 − 1/p it follows that

kf + gkp ≤ kf kp + kgkp0 . 2
Theorem A.3. For 1 ≤ p ≤ +∞, Lp is a Banach space.
Proof. We need to prove the completeness. We begin with the case p = +∞. Let {fk } be a
Cauchy sequence in L∞ . Then
|fk (x) − fm (x)| ≤ kfk − fm k∞
for all x outside a set Ek,m of measure 0. If E = ∪∞
k,m=1 Ek,m , then |E| = 0 and
|fk (x) − fm (x)| ≤ kfk − fm k∞
for all x 6∈ E.
Thus, {fk } converges uniformly to a bounded limit f outside a set of measure 0, i.e. kfk −
f k∞ → 0.
Now let 1 ≤ p < ∞. We use the following characterization of Banach spaces: A normed space
X is complete if and only if every series converging absolutely it converges in norm (that is, if
P PN
k kxk kX < ∞ then there exists x ∈ X such that limP N →+∞ k k=1 xk − xkX = 0). P
p N
Let {fPk } be a sequence of functions in L such that k kfk kLp = C < ∞. Let gN = k=1 |fk |
∞
and g = k=1 |fk |. By the monotone convergence theorem
Z
kgkpp = lim |gN (x)|p dx ≤ C p .
N →+∞
 HARMONIC ANALYSIS 123

Then g ∈ Lp , and in particular g(x) < ∞ a.e. This implies that the series
P
k fk converges a.e.
to a function f . For such a function, |f | ≤ g, so that f ∈ Lp . Finally,
N
X
|f − fk |p ≤ (g + g)p = 2p g p ,
k=1
so that by the dominated convergence theorem
N Z N
p p
X X
lim f− fk p = lim f (x) − fk (x) dx = 0 ,
N →+∞ N →+∞
k=1 k=1
Lp -norm.
P
that is k fk converges to f in the 2
PN
Proposition A.4. For 1 ≤ p < ∞, the set of simple functions {f : f = k=1 αk χEk , with |Ek |
< ∞, for all k} is dense in Lp .
Proof. We first show that if f is measurable, then there exists a sequence of simple functions
sm such that |s1 | ≤ |s2 | ≤ · · · ≤ |f |, sm → f pointwise, and sm → f uniformly on any set where
f is bounded. By writing f = f+ − f− we may assume that f ≥ 0.
Let m be a positive integer and let 0 ≤ k ≤ 22m − 1. It suffices to define
k
= f −1 (k2−m , (k + 1)2−m ] , and Fm = f −1 (2m , +∞] ,



Em
and set
22m
X −1
sm = k2−m χEm m
k + 2 χFm .

k=0
It is not difficult to check that 0 ≤ sm ≤ sm+1 ≤ f for all m = 1, 2, . . . and that 0 ≤ f −sm ≤ 2−m
on the set where f ≤ 2m . The claim now follows.
Now, let f ∈ Lp , and let {sm } be a sequence of simple measurable functions converging
pointwise a.e. to f , |sm | ≤ f . Then, sm ∈ Lp for all m and |f − sm |p ≤ 2p |f |p , which is a
function in L1 . Then, by the dominated converge theorem, kf − sm kp → 0 as m → +∞. 2

Proposition A.5. For 1 ≤ p < ∞ the set of continuous functions with compact support Cc is
dense in Lp .
Proof. Since the simple functions are dense in Lp , it suffices to show that for any measurable set
E, with |E| < ∞, we can approximate χE in the Lp -norm with compactly suported, continuous
functions. Given any measurable set E, |E| < ∞, we can find a (so-called Borel set) E 0 , with
E 0 ⊆ E, |E \ E 0 | = 0, such that for any ε > 0 there exist an open set U and a compact set K
such that
K ⊆ E0 ⊆ U , |U \ K| < ε .
Now, by Urysohn’s lemma, we can find f ∈ Cc such that χK ≤ f ≤ χU . Therefore,
kχE − f kp = kχE 0 − f kp ≤ |U \ K|1/p ≤ ε1/p . 2
Next we wish to describe the dual space of Lp , when 1 ≤ p < ∞.
Theorem A.6. (Reverse Hölder’s inequality) Let g be a measurable function, 1 ≤ p < ∞. Then
R
kgkp0 = sup f g dx : kf kp = 1 .
124 M. M. PELOSO
 R
Proof. Set M (g) = sup f g dx : kf kp = 1 . From Hölder’s inequality we know that, if
0
g ∈ Lp , Z
f g dx ≤ kf kp kgkp0
that is M (g) ≤ kgkp0 . The statement holds true if g = 0 a.e. If g 6= 0 and p0 < ∞, select
0
f = sgn g(|g|/kgkp0 )p −1


where sgn g (x) = g(x)/|g(x)| if g(x) 6= 0 and equals 0 if g(x) = 0. Then,
Z
1 0
p
kf kp = p(p0 −1)
|g(x)|p(p −1) dx
kgkp0
=1
while
Z Z
1 0
M (g) ≥ f g dx = p(p0 −1)
|g(x)|p dx
kgkp0
kgkp0 .
Finally, if p0 = ∞, given any ε > 0, let E be a set such that
E ⊆ {x : |g(x)| > kgk∞ − ε}, 0 < |E| < ∞ .
Select
1
sgn gχE .
f=
|E|
Then Z Z
1
M (g) ≥ f g dx = |g(x)| dx ≥ kgk∞ − ε .
|E| E
This proves the theorem. 2
0
Theorem A.7. (Duality of the Lp spaces) For 1 ≤ p < ∞ the dual of Lp is Lp , with equatity
of norms.
0
Proof. By Hölder’s inequality is clear that any f ∈ Lp gives rise to a bounded linear functional
Lf on Lp , 1 ≤ p < ∞. Conversely, let L be a bounded linear functional on Lp , 1 ≤ p < ∞. We
0
wish to find an f ∈ Lp such that Z
L(g) = g f¯ dx ,
that is, L = Lf . The construction of such an f goes a little beyond the scope of these notes and
lectures. Therefore, we refer to [Fo1], or [WZ] for a proof. 2
We conclude this section with a few further inequalities.
Theorem A.8. (Minkowski’s integral inequality) Let 1 ≤ p ≤ ∞, f a measurable function
defined on the product space Rn1 × Rn2 . Suppose that f (·, y) ∈ Lp for a.e. y and that the
function y 7→ kf (·, y)kp ∈ L1 . Then the function f (x, ·) ∈ L1 for a.e. x, the function x 7→
f (x, y) dy ∈ Lp , and
R
Z Z
f (·, y) dy p
≤ kf (·, y)kp dy .
 HARMONIC ANALYSIS 125

Proof. When p = 1 this is just Fubini’s theorem on iterated integrals. Let 1 < p < ∞, then we
0
use the duality between Lp spaces. Let p0 be the conjugate exponent of p, and g ∈ Lp . We have
then
Z Z  x
f (x, y) dy g(x) dx ≤ |f (x, y)||g(x)| dxdy
Z Z 1/p
p
≤ |f (x, y)| dx kgkp0 dy
Z
= kgkp0 kf (·, y)kp dy .
0
By taking the supremum over g ∈ Lp , kgkp0 = 1, the statement follows. 2

Theorem A.9. (Chebyshev’s inequality) Let 1 ≤ p < ∞. Then, for any λ > 0
kf kp p
 
{x : |f (x)| > λ} ≤ .
λ
Proof. Define Eλ = {x : |f (x)| > λ}. We have
Z Z Z
kf kpp = |f |p dx ≥ |f |p dx ≥ λp dx = λp |Eλ | .
Eλ Eλ

This gives the desired inequality. 2

Appendix B. Some exercises

1. Evaluate the following limits, justifying your answers:
x k x/2
Z k  Z k
x
k −2x
lim 1− e dx lim 1+ e dx .
k→+∞ 0 k k→+∞ 0 k
2. Prove Lemma 2.1.
3. (Measure of the ball and sphere in Rn .) Recall that for a, b > 0,
Z Z 1
−a|x|2

n/2 Γ(a)Γ(b)
e dx = π/a , and (1 − s)a−1 sb−1 ds = ,
Rn 0 Γ(a + b)
R∞
(where the Euler’s Gamma function is defined as Γ(x) = 0 tx−1 e−t dt).
Use these facts and integration in polar coordinates to compute the surface measure of the
unit sphere S n−1 in Rn and the volume of the ball of radius r, Br :
2π n/2 π n/2
σ(S n−1 ) = , |Br | = .
Γ( n2 ) Γ( n2 + 1)
Pm
4. (Generalized Hölder’s inequality.) Show that if fj ∈ Lpj , j = 1, . . . , m and j=1 1/pj =1
then m 1
Q
j=1 fj ∈ L and
Z Y m m
Y
| fj (x)| dx ≤ kfj kpj .
j=1 j=1
126 M. M. PELOSO

5. (Differentiation under the integral sign.) Let f : Rn × [a, b] be measurable and denote by
(x, y) the variables in Rn × R. Suppose ∂y f exists and suppose that there exists g ∈ L1 (Rn )
such that |∂y f (x, y)| ≤ |g(x)| for all (x, y) ∈ Rn × [a, b]. Show that
Z
F (y) = f (x, y) dx
Rn
is differentiable and that F 0 (y) = Rn ∂y f (x, y) dx. Use this result to prove Theorem 2.3 (iii).
R

6. Show that the translation is a continuous operation in Lp when 1 ≤ p < ∞, but it is not
continuous for p = ∞.
7. Let f, g ∈ L2 . Show that both sides of the equality F(f ∗ g) = fˆĝ are well defined, and prove
such an equality.
8. Let 1 ≤ p < ∞, f ∈ Lp such that ∂xα f ∈ Lp for |α| ≤ N . Show that there exists a sequence
of functions {ϕk } in C0∞ such that ∂xα ϕk → ∂xα f in Lp , for all |α| ≤ N .
9. Prove that F(ϕ)(x) = e−2π|ξ| , where ϕ(x) = ωn (1+|x|2 )−(n+1)/2 and ωn = Γ (n+1)/2 π −(n+1)/2 ,

by proving: (i) for β ≥ 0,

1 +∞ e−iβt
Z
−β
e = dt ;
π −∞ 1 + t2
(ii) for β ≥ 0,
Z +∞ −s −β 2 /4s
−β e e
e = √ ds ;
0 πs
(iii) set β = 2π|ξ| in (ii) and use Lemma 2.13.
10. Let G be an invertible linear transfomation of Rn and let g = f ◦ G, for f ∈ L1 . Express ĝ
in terms of fˆ (that is, prove that
1
fˆ t G−1 (ξ) . )


ĝ(ξ) =
| det G|
11. Let δ denote the Dirac delta at the origin, that is the distribution hδ, ϕi = ϕ(0). Prove that
∂ α δ(ξ) = (2πiξ)α and that (xα ) b = (−2πi)−|α| ∂ α δ.
12. On R1 , let Φ be the distribution defined as
Z
1
hΦ, ϕi = lim φ(x) dx .
ε→0 |x|≥ε πx
Show that the equality above defines indeed a distribution and show that the limit equals
ϕ(x) − ϕ(0)
Z Z
ϕ(x)
dx + dx .
|x|≤1 πx |x|>1 πx

Moreover, prove that Φ̂ = −isgn ξ. Call g = Φ̂. Prove that f 7→ f ∗ g extends to a unitary
operator on L2 (called the Hilbert transform).
 HARMONIC ANALYSIS 127

References
[Du] J. Duoandikoetxea, Fourier Analysis, Graduate Studies in Mathematics 29, American Mathematical
Society, Rhode Island 2001.
[Fo1] G. B. Folland, Real Analysis, Modern Techniques and Their Applications. Second Edition, Wyley & Sons,
New York 1999.
[Fo2] G. B. Folland, Introduction to Partial Differential Equations. Second Edition, Princeton University Press,
Princeton 1995.
[Gr] L. Grafakos, Classical Fourier Analysis, GTM 249, Springer-Verlag Ed., New York 2008.
[H] L. Hörmander, Anaysis of Partial Diffrential Operators, vol 1, Distribution theory and Fourier analysis,
Springer-Verlag, Berlino 2003.
[Ka] Y. Katznelson, An Introduction to Harmonic Analysis, Dover, New York 1968.
[Ru] W. Rudin, Functional Analysis, 2nd Ed., Mc Graw Hill, Toronto 1991.
[So] P.M. Soardi, Serie di Fourier in più variabili, Quaderni dell’Unione Matematica Italiana 26 , Pitagora
Ed., Bologna 1984.
[St1] E. M. Stein, Singular Integrals and Differentiability Properties of Functions, Princeton Univ. Press, Prince-
ton 1970.
[St2] E. M. Stein, Harmonic Analysis, Real-variable Methods, Orthogonality, and Oscillatory Integrals, Prince-
ton Univ. Press, Princeton 1993.
[StWe] E. M. Stein, G. Weiss, Introduction to Fourier Analysis on Euclidean Spaces, Princeton Univ. Press,
Princeton 1971.
[WZ] R. Wheeden, A. Zygmund, Measure and Integral, An Introduction to Real Analysis, Marcel Dekker, New
York 1977.

Dipartimento di Matematica, Università degli Studi di Milano, Via C. Saldini 50, 20133 Milano,
Italy
E-mail address: marco.peloso@umimi.it
URL: http://wwww.mat.unimi.it/~peloso/