Math 247A Lecture Notes

Classical Fourier Analysis

Professor: Monica Visan
Scribe: Daniel Raban

Contents
1 Review: The Fourier Transform 4
1.1 Properties of the Fourier transform . . . . . . . . . . . . . . . . . . . . . . . 4
1.2 The Riemann-Lebesgue lemma . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.3 Fourier transform of Gaussians . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.4 Fourier inversion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

2 Fourier Inversion and Plancherel’s Theorem 9

2.1 Fourier inversion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2 Plancherel’s theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.3 The Hausdorff-Young inequality . . . . . . . . . . . . . . . . . . . . . . . . . 11

3 The Littlewood Principle and Lorentz Spaces 13

3.1 The Littlewood principle and optimality of the Hausdorff-Young inequality 13
3.2 Weak Lp and Lorentz spaces . . . . . . . . . . . . . . . . . . . . . . . . . . 14

4 Relationships Between The Lorentz Quasinorms and Lp Norms 16

4.1 Order of growth of Lorentz quasinorms in terms of Lp and `q . . . . . . . . 16
4.2 Lorentz spaces are Banach spaces . . . . . . . . . . . . . . . . . . . . . . . . 17

5 Banach Space Properties of Lorentz Spaces 20

5.1 Proof of completeness, duality, and more . . . . . . . . . . . . . . . . . . . . 20

6 Hunt’s Interpolation Theorem 23

6.1 Strong type and weak type . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
6.2 Hunt’s interpolation theorem . . . . . . . . . . . . . . . . . . . . . . . . . . 24

7 Proofs of Interpolation Theorems 27

7.1 Proof of the Marcinkiewicz interpolation theorem . . . . . . . . . . . . . . . 27
7.2 Proof of Hunt’s interpolation theorem . . . . . . . . . . . . . . . . . . . . . 28

1
8 Interpolation and Maximal Function Estimates 31
8.1 Conclusion of proof of Hunt’s interpolation theorem . . . . . . . . . . . . . 31
8.2 Maximal and vector maximal functions . . . . . . . . . . . . . . . . . . . . . 32

9 Boundedness Properties of The Hardy-Littlewood Maximal Function and

Ap Weights 35
9.1 Boundedness properties of the Hardy-Littlewood maximal function . . . . . 35
9.2 Ap weights . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

10 Characterization of Ap Weights 39
10.1 Ap weights for p > 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
10.2 Proof sketch of characterization . . . . . . . . . . . . . . . . . . . . . . . . . 41

11 Ap Weights and The Vector-Valued Maximal Function 43

11.1 Use of reverse Hölder in the characterization of Ap weights . . . . . . . . . . 43
11.2 The vector-valued maximal function . . . . . . . . . . . . . . . . . . . . . . 45

12 Calderón-Zygmund Decomposition and Bounds for the Vector-Valued

Maximal Function 48
12.1 A Calderón-Zygmund decomposition . . . . . . . . . . . . . . . . . . . . . . 48
12.2 Weak-type bound for the vector-valued maximal function . . . . . . . . . . 49

13 The Hardy-Littlewood-Sobolev Inequality 52

13.1 Failure of bounds for L1 vector-valued maximal function . . . . . . . . . . . 52
13.2 The Hardy-Littlewood-Sobolev inequality . . . . . . . . . . . . . . . . . . . 52

14 The Sobolev Embedding Theorem 56

14.1 Fourier transforms of tempered distributions . . . . . . . . . . . . . . . . . . 56
14.2 Sobolev embedding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58

15 Riesz Tranforms and Calderón-Zygmund Convolution Kernels 60

15.1 Riesz tranforms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
15.2 L2 -boundedness of convolution with Calderón-Zygmund convolution kernels 62

16 Boundedness of Calderón-Zygmund Convolution Kernels 64

16.1 L2 -boundedness of convolution with Calderón-Zygmund kernels . . . . . . . 64
16.2 Lp bounds for Calderón-Zygmund convolution kernels . . . . . . . . . . . . 67

17 Lp Bounds for Calderón-Zygmund Convolution Kernels 68

17.1 Weak Lp bound for Calderón-Zygmund convolution kernels . . . . . . . . . 68
17.2 Application: The Hilbert transform . . . . . . . . . . . . . . . . . . . . . . . 70

2
18 The Mikhlin Multiplier Theorem 71
18.1 The Hilbert transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
18.2 Littlewood-Paley projections and the Mikhlin multiplier theorem . . . . . . 72

19 The Mikhlin Multiplier Theorem and Properties of Littlewood-Paley

Projections 75
19.1 The Mikhlin multiplier theorem . . . . . . . . . . . . . . . . . . . . . . . . . 75
19.2 Properties of Littlewood-Paley projections . . . . . . . . . . . . . . . . . . . 77

20 Littlewood-Paley Projections and Khinchine’s Inequality 80

20.1 Bernstein properties of Littlewood-Paley projections . . . . . . . . . . . . . 80
20.2 Khinchine’s inequality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
20.3 Littlewood-Paley square function estimate . . . . . . . . . . . . . . . . . . . 84

21 Estimates on the Littlewood-Paley Square Function and the Fractional

Product Rule 85
21.1 Estimates on the Littlewood-Paley square function . . . . . . . . . . . . . . 85
21.2 The fractional product rule . . . . . . . . . . . . . . . . . . . . . . . . . . . 88

22 The Fractional Chain Rule 90

22.1 Proof of the fractional chain rule . . . . . . . . . . . . . . . . . . . . . . . . 90

23 Introduction to Oscillatory Integrals 94

23.1 Decay of integrals with compactly supported integrand . . . . . . . . . . . . 94
23.2 The Van der Corput lemma . . . . . . . . . . . . . . . . . . . . . . . . . . . 95

24 Estimating Oscillatory Integrals With Stationary Phase 98

24.1 Estimation in the 1 dimensional case . . . . . . . . . . . . . . . . . . . . . . 98

25 Oscillatory Integrals in Higher Dimensions 102

25.1 Nonstationary phase . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
25.2 Stationary phase and Moore’s change of variables lemma . . . . . . . . . . . 102

3
1 Review: The Fourier Transform
1.1 Properties of the Fourier transform
This class is called “Classical Fourier Analysis,” but for the past 20 years, it has been
taught more like “Modern Harmonic Analysis.” Our treatment will be no different.
Definition 1.1. The Fourier transform of a function f ∈ L1 (Rd ) is given by
Z
(Ff )(ξ) = fb(ξ) = e−2πix·ξ f (x) dx.

Remark 1.1. By the triangle inequality,

kfbkL∞ ≤ kf kL1 .

We will prove quantitative results about nice sets of functions and extend these results
to more general functions via density arguments. What are our “nice” functions?
Definition 1.2. A C ∞ function f : Rd → C is called a Schwarz function if xα Dβ f ∈ L∞
for all multi-indices α, β ∈ Nd . The vector space of all such functions, S(Rd ), is called the
Schwarz space.
This says that all the derivatives of f decay faster than any polynomial. Recall that
for a multi-index α ∈ Nd , we denote

∂ |α|
|α| = α1 + · · · + αd , xα = xα1 1 xα2 2 · · · xαd d , Dα = .
∂xα11 · · · ∂xαdd
The Schwarz space is a Fréchet space with the topology generated by the countable
family of seminorms {ϕα,β }α,β∈Nd with ϕα,β (f ) = kxα Dβ f kL∞ .

Proposition 1.1 (properties of the Fourier transform). Fix f ∈ S(Rd ).

1. If g(x) = f (x − y) with y ∈ Rd fixed, then gb(ξ) = e−2πiy·ξ fb(ξ).

Proof.
Z Z
−2πix·ξ
gb(ξ) = e f (x − y) dx = e−2πi(x+y)·ξ f (x) dx = e−2πiy·ξ fb(ξ).

2. Let g(x) = e2πix·η f (x) for η ∈ Rd fixed. Then gb(ξ) = fb(ξ − η).

Proof. Z
gb(ξ) = e−2πix(ξ−η) f (x) dx = fb(ξ − η).

4
3. If f (x) = f (T x) for T ∈ GLd (R), then fb(ξ) = | det T |−1 fb((T −1 )> ξ).

Proof.
Z
fb(ξ) = e−2πix·ξ f (T x) dx
Z
y=T x −1 −1
= | det T | e−2πiT y·ξ f (y) dy
Z
−1 −1 >
= | det T | e−2πiy·(T ) ξ f (y) dy

= | det T |−1 fb((T −1 )> ξ).

4. If g = f , then gb(ξ) = fb(−ξ).

5. If g = Dα f with α ∈ Nd , then gb(ξ) = (2πiξ)α fb(ξ)

Proof. Using integration by parts,

Z
gb(ξ) = e−2πix·ξ Dα f (x) dx = (2πiξ)α fb(ξ).

6. If g(x) = xα f (x) for α ∈ Nd , then

1
gb(ξ) = Dα fb(ξ).
(−2πi)|α|

Proof.
Z
gb(ξ) = e−2πix·ξ xα f (x) dx
Z
1
= e−2πix·ξ (−2πix)α f (x) dx
(−2πi)|α|
1
= Dα fb(ξ).
(−2πi)|α|

7. Let g = k ∗ f for k ∈ L1 (Rd ). Then gb(ξ) = b

k(ξ)fb(ξ).

Proof.
Z
gb(ξ) = e−2πix·ξ (k ∗ f )(x) dx

5
 ZZ
= e−2πix·ξ k(x − y)f (y) dy dx
ZZ
z=x−y
= e−2πi(z+y)·ξ k(z)f (y) dz dy

=b
k(ξ)fb(ξ).

Remark 1.2. Properties 1, 2, 3, 4, and 7 extend to f ∈ L1 (Rd ).

Remark 1.3. Property 3 implies that any rotation and/or reflection symmetry T ∈ Od (R)
of f is inherited by fb. Indeed, if f (x) = f (T x), then

fb(ξ) = | det T |−1 fb((T −1 )> ξ) = fb(T ξ).

Exercise 1.1. Show that

1. If f ∈ S(Rd ), then fb ∈ S(Rd ).

d
S(R ) S(R )d
2. If fn −−−→ f , then fbn −−−→ fb.

These follow from properties 5 and 6.

1.2 The Riemann-Lebesgue lemma

Lemma 1.1 (Riemann-Lebesgue). If f ∈ L1 (Rd ), then fb ∈ C0 (Rd ) (continuous and van-
ishing at infinity). In particular, fb is uniformly continuous.
L1
Proof. Let fn ∈ S(Rd ) be such that fn −→ f . By the triangle inequality,
n→∞
kfbn − fbkL∞ ≤ kfn − f kL1 −−−→ 0.

Now fbn ∈ S(Rd ) ⊆ C0 (Rd ), and C0 (Rd ) is closed in L∞ . So fb ∈ C0 (Rd ).

1.3 Fourier transform of Gaussians

Lemma 1.2. Let A be a positive-definite, real-symmetric d × d matrix. Then
Z
2 −1
e−x·Ax e−2πix·ξ dx = (det A)−1/2 π d/2 e−π ξ·A ξ .

Proof. A real-symmetric, positive-definite matrix is diagonalizable, so there exists an or-

thogonal O ∈ Od (R) such that A = O> DO with D = diag(λ1 , . . . , λ` ) with λ1 , . . . , λd > 0.
Now
y=Ox X
x · Ax = x · O> DOx = Ox · DOx = y · Dy = λj yj2 .

6
We have
y=Ox η=Oξ
x · ξ = O−1 y · ξ = y · Oξ = y · η.
So Z Z
(λj yj2 −2πiyj ηj )
P
e−x·Ax e−2πix·ξ dx = e− dy.

This is a product of 1-dimensional integrals. Let’s look at the 1-dimensional integral

Z Z
πiη 2 π 2 η 2
−λy 2 −2πiyη
e dy = e−λ(y+ λ ) − λ dy
R ZR
2 2 2
= e−λy e−π η /λ dy
R
−1/2 1/2 −π 2 η 2 /λ
=λ π e .

So we get
Z d
−1/2 1/2 −π 2 ηj2 /λj
Y
−x·Ax −2πix·ξ
e e dx = (λj π e )
j=1
2 η·D −1 η
= (det A)−1/2 π d/2 e−π
2 ξ·O > D −1 Oξ
= (det A)−1/2 π d/2 e−π
2 ξ·A−1 ξ
= (det A)−1/2 π d/2 e−π .
2
Corollary 1.1. e−π|x| is n eigenvector of the Fourier transform with eigenvalue 1.

Proof.
2 A=πI 2
[F(e−π|x| )](ξ) = e−π|ξ| .

1.4 Fourier inversion

Theorem 1.1 (Fourier inversion). For f ∈ S(Rd ), we have

[(F ◦ F)f ](x) = f (−x),

or equivalently, Z
f (x) = e2πix·ξ fb(ξ) dξ.

We can think of this as decomposing f into a linear combination of characters with

Fourier coefficients.

7
Proof. We can’t use Fubini like we want to because the integrand is not necessarily abso-
lutely integrable. The (standard) trick is to force a Gaussian in there. For ε > 0, let
Z
2 2
Iε (x) = e−πε |ξ| e2πix·ξ fb(ξ) dξ.

e2πix·ξ fb(ξ) dξ as ε →
R
Then the dominated convergence theorem tells us that Iε (x) →
0.

Next time, we will complete the proof.

8
2 Fourier Inversion and Plancherel’s Theorem
2.1 Fourier inversion
Theorem 2.1 (Fourier inversion). For f ∈ S(Rd ), we have

[(F ◦ F)f ](−x) = f (x),

or equivalently, Z
f (x) = e2πix·ξ fb(ξ) dξ.

We can think of this as decomposing f into a linear combination of characters with

Fourier coefficients.

Proof. We can’t use Fubini like we want to because the integrand is not necessarily abso-
lutely integrable. The (standard) trick is to force a Gaussian in there. For ε > 0, let
Z
2 2
Iε (x) = e−πε |ξ| e2πix·ξ fb(ξ) dξ.

e2πix·ξ fb(ξ) dξ as ε → 0.
R
Then the dominated convergence theorem tells us that Iε (x) →
On the other hand,
ZZ
2 2
Iε (x) = e−πε |ξ| e2πix·ξ e−2πiy·ξ f (y) dy dξ
Z Z
2 2
= f (y) e−πε |ξ| e−2πi(y−x)·ξ dξ dy

Use our lemma from last time with the linear transformation A = πε2 I:
Z
2 1
= f (y)(πε2 )−d/2 π d/2 e−π (y−x) πε2 (y−x) dy
Z
π 2
= ε−d e− ε2 |x−y| f (y) dy.

π 2
ε−d e− ε2 |x| dx =
2
e−π|x| dx.
R R
Note that
ε→0
−−−→ f (x).
π 2 π 2
To show this convergence, we have ε−d e− ε2 |x−y| f (y) dy−f (x) = ε−d e− ε2 |x−y| dx[f (y)−
R R

f (x)] dy. For η > 0, there is a δ(η) > 0 such that |f (y) − f (x)| < η whenever |x − y| < δ.
Then Z Z
π 2 π 2
ε−d e ε2 |x−y| [f (y) − f (x)] dy ≤ η ε−d e ε2 |x−y| dy ≤ η,
|x−y|<δ |x−y|<δ

9
 Z Z
π π
|x−y|2 2
ε −d
e ε2 [f (y) − f (x)] dy ≤ 2kf kL∞ ε−d e− ε2 |y| dy
|x−y|>δ |y|>δ
Z
2
≤ 2kf kL∞ e−π|y| dy
|y|>δ
δ2
. kf kL∞ e−π 2ε2
ε→0
−−−→ 0.

First pick η  1. Then choose ε = ε(δ) = ε(η)  1.

Corollary 2.1. The Fourier transform is a homeomorphism on S(Rd ).

2.2 Plancherel’s theorem

Lemma 2.1. For f, g ∈ S(Rd ), we have
Z Z
f (ξ)b
b g (ξ) dξ = f (x)g(x) dx.

In particular,
kfbkL2 = kf kL2 .
so F is an isometry in L2 on S(Rd ).
Proof. For h ∈ S(Rd ),
Z ZZ
fb(ξ)h(ξ) dζ = e−2πix·ξ f (x)h(ξ) dx dξ
Z
= f (x)b
h(x) dx.

Now let h = ĝ. Then (Fh)(x) = F(b

g )(−x) = g(x).

Theorem 2.2 (Plancherel). The Fourier transform extends from S(Rd ) to a unitary map
on L2 (Rd ).
Proof. Fix f ∈ L2 (Rd ). To define the Fourier transform on F, let fn ∈ S(Rd ) be such that
L2 n,m→∞
fn −→ f . Since F is an isometry in L2 on S(Rd ), kfbn − fbm kL2 = kfn − fm kL2 −−−−−→ 0.
So {fbn }n≥1 is Cauchy and hence convergent in L2 (Rd ). Let fb be the L2 limit of the fbn .
We claim that fb does not depend on the sequence {fn }n≥1 . Let {gn }n≥1 ⊆ S(Rd ) be
L2
another sequence such that gn −→ f . Let
(
fk n = 2k − 1
hn =
gk n = 2k.

10
 L 2
We have that {hn } ⊆ S(Rd ), and hn −→ f . By the same argument as before, {b hn }n≥1
converges in L2 . This means that limn b
hn = limn fbn = limn gbn .
We now claim that kfbk2 = kf k2 for all f ∈ L2 (Rd ); i.e. F is an isometry on L2 . Indeed,

kfbk2 = lim kfbn k2 = lim kfn k2 = kf k2 .

n

Remark 2.1. This is not yet enough to show that F is unitary. In infinite dimensions,
isometries need not be unitary. For example, take T : `2 (N) → `2 (N) be T (a1 , a2 , . . . ) =
(0, a1 , a2 , . . . ). Then
X
hT (a1 , a2 , . . . ), (b1 , b2 , . . . )i = an bn+1 = h(a1 , a2 , . . . ), (b2 , b3 , . . . )i ,
n≥1

so T ∗ (a1 , a2 , . . . ) = (a2 , a3 , . . . ). So T ∗ T = id, but T T ∗ 6= id. What we need to get an

isometry is surjectivity.
We claim that F : L2 (Rd ) → L2 (Rd ) is onto. We will show that Ran(F) is closed in
L2 L2
L2 (Rd ). As Ran(F) ⊇ S(Rd ), this will give L2 (Rd ) = S(Rd ) ⊆ Ran(F) = Ran(F).
L2 L2
Let g ∈ Ran(F) . Then there exist fn ∈ L2 such that fbn −→ g. F is an isometry on
n,m→∞
L2 (Rd ), so kfn − fm k2 = kfbn − fbm k2 −−−−−→ 0. So {fn }n≥1 converges in L2 to some f .
Then g = fb because
n→∞
kfb − fbn k2 = kf − fn k2 −−−→ 0.
By the uniqueness of limits, we get g = fb. So we get g = fb ∈ Ran(F).

2.3 The Hausdorff-Young inequality

Theorem 2.3 (Hausdorff-Young). For f ∈ S(Rd ),

kfbkp0 ≤ kf kp , ∀1 ≤ p ≤ 2,

where 1/p + 1/p0 = 1.

Proof. This follows from interpolation, as we have F : L1 → L∞ with kfbkL∞ ≤ kf kL1 and
F : L2 → L2 with kfbkL2 = kf kL2 .

Remark 2.2. As in the proof of Plancherel’s theorem, we can use Hausdorff-Young to

extend the Fourier transform from S(Rd ) to Lp (Rd ) for any 1 ≤ p ≤ 2.
Note that the Riemann-Lebesgue lemma gives that for f ∈ L1 (Rd ), fb ∈ C0 (Rd ). So
we can think of evaluating the Fourier transform at a single point or on a measure 0 set,
such as a plane in R3 . The restriction problem asks: For which values of p can we make
sense of the Fourier transform on measure 0 sets, such as a parabaloid or a cone? This is
important in PDE, and it is very hard (still open!).
The next theorem says that the Hausdorff-Young inequality is the best we can do.

11
Theorem 2.4. If kfbkLq ≤ kf kLp for some 1 ≤ p, q ≤ ∞ and all f ∈ S(Rd ), then neces-
sarily, q = p0 and 1 ≤ p ≤ 2.

Proof. For f ∈ S(Rd ) with f 6≡ 0, define fλ (x) = f (x/λ) for λ > 0. Then kfλ kp = λd/p kf kp .
We also have Z
fbλ (ξ) = e−2πix·ξ f (x/λ) dx = λd fb(λξ),

so kfbλ kq = λd−d/q kfbkq . Then kfbλ kq ≤ kfλ kp if and only if λd−d/q kfbkq ≤ λd/p kf kp , so
λd(1−1/q−1/p) kfbkq ≤ kf kp . Letting λ → 0 or λ → ∞, we conclude that 1 − 1/q − 1/p = 1.
So we get q = p.

Next time, we will prove the remaining portion of this theorem, that 1 ≤ p ≤ 2.

12
3 The Littlewood Principle and Lorentz Spaces
3.1 The Littlewood principle and optimality of the Hausdorff-Young in-
equality
Last time we were proving the following theorem.

Theorem 3.1. If kfbkLq ≤ kf kLp for some 1 ≤ p, q ≤ ∞ and all f ∈ S(Rd ), then neces-
sarily, q = p0 and 1 ≤ p ≤ 2.

We have already proven the first statement. To prove the second we will use the
Littlewood principle: “A translation invariant operator does not improve decay.” So if
T : Lp → Lq , then q ≥ p. This is not a theorem but a general principle.
Say we have a bump function at 0 and we translate it far away. Keep doing this (N
times), and let f be the superposition of all the bump functions. If we apply T to f , since
T is translation invariant, we will get N translated copies of the modified bump. Then
kf kLp ∼ N 1/p , while kT f kLq ∼ N 1/q . Then we need N 1/q . N 1/p . Letting N → ∞, we
get 1/q ≤ 1/p, so p ≤ q.
The Fourier transform is not translation invariant, however. And the Fourier transform
of a compactly supported function no longer has compact support. However, we can use
the fast decay of the Gaussian to achieve the same effect.
2
Proof. Let ϕ(x) = e−π|x| . For 1 ≤ k ≤ N and α  1, define

ϕk (x) = e2πix·αke1 ϕ(x − αke1 ).

Then
bk (ξ) = e−2πiαkξ1 ϕ(ξ
ϕ b − αke1 ).
PN SN
Let f = k=1 ϕk and S = j=1 {x : |x − αj e1 | ≤ α/10}. Then

kf kLp = kf kLp (S) + kf kLp (Rd \S) .

We can bound each of these by

N
X
kf kLp (Rd \S) ≤ kϕk kLp (Rd \S) . N α−100 ,
k=1

N N
kf kpLp (S) ∼
X X
ϕk ∼ N (1 + O(α−100 ))
j=1 k=1 Lp (|x−αje1 |≤α/10)
| {z }
∼1+O( k6=j (α|k−j|)−100 )
P

because |x − αke1 | ≥ |α(j − k)e1 | − |x − αje1 | ≥ α|j − k| − α/10 ≥ (α/2)|j − k|.

13
 Taking α  1, we get kf kLp ∼ N . Similarly,
0
kfbkLp0 ∼ N 1/p
0
We need N 1/p ≤ N 1/p for all N ≥ 1. This means that 1/p0 ≤ 1/p, so p ≤ p0 . So
1 ≤ p ≤ 2.

3.2 Weak Lp and Lorentz spaces

Definition 3.1. For 1 ≤ p < ∞ and f : Rd → C, define

kf k∗Lp = sup λ|{x : |g(x)| > λ}|1/p .

weak λ>0

The weak Lp space is the set of measurable functions f : Rd → C for which kf k∗Lp < ∞.
weak
We denote it by Lpweak (Rd ).

Example 3.1. f (x) = |x|d/p is in Lpweak \ Lp . We have

kf k∗Lweak = sup λ|{x : |x|−d/p > λ}|1/p ∼ sup λ(λ−p )1/p ∼ 1.

λ>0 λ>0

Remark 3.1. We will show that the weak Lp “norm” is a quasinorm (not a norm) and
that is why we append ∗ to the usual norm notation.

By comparison, for 1 ≤ p < ∞,

Z
p
kf kLp = |f (x)|p dx
Z Z |f (x)|
= pλp−1 dλ dx
0
Z ∞
= pλp−1 |{x : |f (x)| > λ}| dλ
0
Z ∞
1
=p λp |{x : |f (x)| > λ}| dλ.
0 λ

So we can write
kf kLp = p1/p kλ|{x : |f (x)| > λ}|1/p kLp ((0,∞), dλ ) .
λ

With the convention that p1/∞ = 1, we also have

kf k∗Lp = p1/∞ kλ|{x : |f (x)| > λ}|1/p kL∞ ((0,∞), dλ ) .

weak λ

Can we do this to Lp spaces for other exponents?

14
Definition 3.2. For 1 ≤ p < ∞ and 1 ≤ q ≤ ∞, the Lorentz space Lp,q (Rd ) is the set
of measurable functions f : Rd → C for which

kf k∗Lp,q (Rd ) = p1/q kλ|{x : |f (x)| > λ}|1/p kLq ((0,∞), dλ ) < ∞.
λ

Note that Lp,p = Lp and Lp,∞ = Lpweak .

Lemma 3.1. kf k∗Lp,q (Rd ) is a quasinorm.

Proof. If kf k∗Lp,q = 0, then f = 0 a.e. For a 6= 0,

kaf k∗Lp,q = p1/q kλ|{x : |af (x)| > λ}|1/p kLq ( dλ )

λ

λ
= p1/q |a| |{x : |f (x)| > λ/|a|}|1/p
|a| Lq ( dλ )
λ

= |a|kf k∗Lp,q .

For the “triangle inequality,” we have

kf + gk∗Lp,q = p1/q kλ|{x : |f (x) + g(x)| > λ}|1/p kLq ( dλ )

λ
1/q
≤p kλ[|{x : |f (x)| > λ/2}| + |{x : |f (x)| > λ/2}|]1/p kLq ( dλ )
λ

By the concavity of fractional powers, we get

" #
1/q λ λ
≤p |{x : |f (x)| > λ/2}|1/p + |{x : |f (x)| > λ/2}|1/p
2 dλ
Lq ( ) 2 Lq ( dλ )
λ λ

≤ 2[kf k∗Lp,q + kgk∗Lp,q ].

Remark 3.2. We will show that for 1 < p < ∞ and 1 ≤ q ≤ ∞, there exists a norm
equivalent to this quasinorm. For p = 1 and q 6= 1, no such norm exists. Nonetheless, in
this latter case, there is a metric that generates the same topology. In all cases, Lp,q (Rd )
is complete.

Proposition 3.1. For f ∈ Lp,q (Rd ), decompose f =

P
m∈Z fm by defining fm (x) =
f (x)1{2m ≤|f (x)|<2m+1 } (x). Then

kf k∗Lp,q ∼ kfm kLp (Rd ) .

`qm (Z)

In particular, Lp,q1 ⊆ Lp,q2 whenever q1 ≤ q2 .

15
4 Relationships Between The Lorentz Quasinorms and Lp
Norms
4.1 Order of growth of Lorentz quasinorms in terms of Lp and `q
Last time, we had the quasinorm

kf k∗Lp,q (Rd ) = p1/q kλ|{x : |f (x)| > λ}|1/p kLq ((0,∞), dλ )

λ

Remark 4.1. If |g| ≤ |f |, then kgk∗Lp,q ≤ kf k∗Lp,q .

Proposition 4.1. If f ∈ Lp,q (Rd ) for 1 ≤ p < ∞ and 1 ≤ q ≤ ∞, write f =

P
m∈Z fm ,
where fm (x) = f (x)1{x:2m ≤|f (x)|<2m+1 } (x). Then

kf k∗Lp,q ∼ kfm kLp (Rd ) .

`qm (Z)

Proof. Both sides only concern |f |, so it suffices to prove this for f ≥ 0. Then

2m 1{2m ≤f (x)<2m+1 } ≤ fm < 2m+1 1{2m ≤f (x)<2m+1 } .

m∈Z 2 1Fm , where Fm are

m
P
Thus, by our previous remark, we may assume that f =
measurable, pairwise disjoint sets.
Z ∞
dλ
2n 1Fn > λ}|q/p
X
(kf k∗Lp,q )q = p λq |{x :
0 n
λ
m
X 2 Z
dλ
2n 1Fn > λ}|q/p
X
=p λq |{x :
2m−1 n
λ
m∈Z

For 2m−1 ≤ λ < 2m , {x : 2n 1Fn (x) > λ} =

P S
n≥m Fn .

 q/p
XZ 2m X dλ
∼ λq  |Fn |
2m−1 λ
m∈Z n≥m
 q/p
X X
∼ 2mq  |Fn |
m∈Z n≥m
 1/p q
X
∼ 2m  |Fn | .
n≥m
`qm

16
 We wanted to show that kf k∗Lp,q ∼ k2m |Fm |1/p k`qm . So we just need to show that
P 1/p
2m n≥m |Fn | ∼ k2m |Fm |1/p k`qm . We have the ≥ direction, so we just need the
`qm
other inequality:
 1/p
X X
2m  |Fn | ≤ 2m |Fn |1/p
n≥m n≥m
`qm `qm
X
. 2−k k2m+k |Fm+k |1/p k`qm
k≥0

Now reindex the `q sum by n = m + k.

X
. 2−k k2n |Fn |1/p k`qn
k≥0

. k2n |Fn |1/p k`qn .

4.2 Lorentz spaces are Banach spaces

Lemma 4.1. Let 1 ≤ q < ∞, and let S ⊆ 2Z , the dyadic integers. Then
!q  q
X X X
q
N ≤ N ≤ 2 sup N ≤ 2q N q.
N ∈S N ∈S N ∈S N ∈S

In other words, if we’re summing dyadic series, when we take the Lq norm, it doesn’t
really matter whether we have the q inside or outside the sum.

Theorem 4.1. For 1 < p < ∞ and 1 ≤ q ≤ ∞,

Z 
∗ ∗
kf kLp,q ∼ sup f (x)g(x) dx : kgkLp0 ,q0 ≤ 1 .

Thus, k · k∗Lp,q is equivalent to a norm, with respect to which Lp,q (Rd ) is a Banach space.
0 0
Moreover, for q 6= ∞, the dual of Lp,q is Lp ,q , under the natural pairing.

Remark 4.2. For p = 1, q 6= 1, there cannot be a norm equivalent to k · k∗L1,q . Let’s see
this for q = ∞ and d = 1. Assume, towards a contradiction, that k · k∗L1,∞ ∼ |||·|||. Let
f (x) = N 1
P
n=1 |x−n| for N  1. Then

∗  
1 1
= sup λ x: > λ = 2,
|x − n| L1,∞ λ>0 |x − n|

17
so
N N ∗
X 1 X 1
∼ = 2N.
|x − n| |x − n| L1,∞
n=1 n=1
Then we have
N
( )
X 1
kf k∗L1,∞ = sup λ x: >λ .
λ>0 |x − n|
n=1
PN 1 1 P
We claim that {x : n=1 |x−n| > log N } ⊇ [0, N ]. If x = 0, then
10 1/n > log(N + 1) ≥
1
10 log N . Now do the same with x = 1, x = 2, . . . . The worst case scenario is when
x ≈ N/2, but the inequality holds in this case, too. So we have
N
( )
1 X 1 1 N log N
kf k∗L1,∞ ≥ log N x : > log N ≥ .
10 |x − n| 10 10
n=1

So we have shown that

N log N
|||f ||| ∼ kf k∗L1,∞ ≥ .
10
This gives
N
X 1
N log N . |||f ||| ≤ ∼ N.
|x − n|
n=1
Let N → ∞ to get a contradiction.
Now let’s prove the theorem.

Proof. We may assume f ≥ 0, g ≥ 0. As both sides P are positive homogeneous, we may

assume that kf k∗Lp.q = 1. We may assume f = 2n 1Fn and g = 2 1Em with Fn
P m
measurable, pairwise disjoint and En measurable, pairwise disjoint. Then
1 = (kf k∗Lp,q )q
∼ k2n |Fn |1/p kq`q
X
∼ 2nq |Fn |q/p
n∈Z
X X
∼ 2nq |Fn |q/p
N ∈2Z n:N ≤|Fn |<2N
X X
∼ N q/p 2nq
N ∈2Z n:|Fn |∼N

By the lemma,
 q
X X
∼ N q/p  2n 
N ∈2Z n:|Fn |∼N

18
  q
X X
∼  2n |Fn |1/p  .
N ∈2Z n:|Fn |∼N

Similarly,
 q 0

q 0 X X
1/p0
1 ≥ kgkLp0 ,q0 ∼  2m |Em |  .
M ∈2Z m:|Em |∼M

Now
Z X
f (x)g(x) dx = 2n 2m |Fn ∩ Em |
n,m
X X X 0 min{N, M }
. 2n |Fn |1/p 2m |Em |1/p
N 1/p M 1/p0
N,M ∈2Z n:|Fn |∼N m:|Em |∼M
X  min{N, M } 1/q+1/q0 X X 0
. 2n |Fn |1/p 2m |Em |1/p
N 1/p M 1/p0
N,M ∈2
Z n:|Fn |∼N m:|Em |∼M

By Hölder’s inequality,
  q 1/q
X min{N, M }  X
. 2n |Fn |1/p  
N 1/p M 1/p0
N,M ∈2Z n:|Fn |∼N
  q0 1/q0
 X min{N, M }  X 0
· 2m |En |1/p   .

N 1/p M 1/p0
N,M ∈2Z m:|Fm |∼M

P min{N,M }
Now we just need M ∈2Z N 1/p M 1/p0 . 1. This comes from
( 1/p0  1/p ) X  M 1/p X  N 1/p0
X N M
min , . + . 1,
M N N M
M M ≤N M >N

as we get a geometric series.1

1
Instead of using Hölder’s inequality and the subsequent steps, we could alternatively use Schur’s test
for convergence of series. This kind of argument will be common in this course.

19
5 Banach Space Properties of Lorentz Spaces
5.1 Proof of completeness, duality, and more
Theorem 5.1. For 1 < p < ∞ and 1 ≤ q ≤ ∞,
Z 
∗ ∗
kf kLp,q ∼p,q sup f (x)g(x) dx : kgkLp0 ,q0 ≤ 1 .

Thus, k · k∗Lp,q is equivalent to a norm, with respect to which Lp,q (Rd ) is a Banach space.
0 0
Moreover, for q 6= ∞, the dual of Lp,q is Lp ,q , under the natural pairing.

Proof. Last time, we saw that it suffices to prove the equivalence for functions of the form

2n 1Fn
X
f=
n∈Z

with Fn measurable, pairwise disjoint, and kf k∗p,q

P ∼ k2n |Fn |1/p k`qn ∼ 1. Last time, we
showed that RHS . LHS by testing it on g = 2m 1Em with Em measurable, pairwise
∗ m 1/p 0
disjoint, and kgkLp0 q0 ∼ k2 |Em | k`q0 . 1. Let’s show that LHS . RHS.
m
|f |q−1 sgn f
Compare: in the case of Lq (Rd ), we take g = . Here, we take
kf kq−1
q

X q−1
|Fn |−1/p 1Fn
0
g= 2n |Fn |1/p

Check:
Z  q−1 0
X
f (x)g(x) dx = 2n 2n |Fn |1/p |Fn |−1/p |Fn |
n
X
= 2nq |Fn |q/p
n
= k2n |Fn |1/p kq`q
∼ (kf k∗Lp,q )q
∼ 1.

It remains to show that kgk∗Lp0 ,q0 . 1. By the proposition which evaluates the norm as a
dyadic sum,
 q 0 X 0 0 0
kgk∗Lp0 ,q0 ∼ N q |{x : N ≤ g(x) < 2N }|q /p
N ∈2Z

20
Note that {x : g ∼ N } = n∈SN Fn , where SN = {n ∈ Z : 2n(q−1) |Fn |q/p−1 ∼ N } = {n ∈
S

Z : |Fn | ∼ [N 2−n(q−1) ]p/(q−p) }.

 q0 /p0
0
X X
∼ Nq  |Fn |
N ∈2Z n∈SN

This is a dyadic sum, so we can pull the exponent inside (by our lemma).
0 0 0
X X
∼ Nq |Fn |q /p
N ∈2Z n∈SN
X q 0 0 0
∼ 2n(q−1) |Fn |q/p−1 |Fn |q /p
n∈Z

Use q 0 = q/(q − 1).

0
X
∼ 2nq |Fn |q (q/p−1/p)
n∈Z
X
∼ 2nq |Fn |q/p
n∈Z
∼ (kf k∗Lp,q )q
∼ 1.

The RHS defines a norm |||·|||. To see that Lp,q equiped with this norm is a Banach
space, one uses the usual Riesz-Fischer argument.
p,q are such that |||·||| < ∞, then there exists a function f ∈ Lp,q
P
Step 1: If kf
P n k ∈ L
such that f = fn in |||·|||.
Step 2: For a Cauchy sequence {fn }n≥1 ⊆ Lp,q , we pass to a subsequence so that
fkn+1 − fkn < 21n . So by Step 1,
n
X |||·|||
fkn = fk1 + fkj − fkj−1 −−→ f.
j=2

0 0
For 1 ≤ q < ∞, we want to show that the dual of Lp,q is Lp ,q . Let ` : Lp,q → R be a
linear functional, so |||`(f )||| . kf k∗Lp,q . For f = 1E with E of finite measure,

`(1E ) . |E|1/p .

So the measure E 7→ `(1E ) is absolutely continuous with respect to Lebesgue measure. So

there exists a g ∈ L1loc such that
Z
`(1E ) = g(x)1E (x) dx.

21
As ` is linear, we get Z
`(f ) = f (x)g(x) dx

for all simple functions f ∈ Lp,q .

0 0
• Claim 1: Boundedness of ` on simple functions yields g ∈ Lp ,q .

• Claim 2: Simple functions are dense in Lp,q if q 6= ∞.

Given these two claims, we get `(f ) = f (x)g(x) dx for all f ∈ Lp,q . Thus, the dual of
R
0 0
Lp,q is Lp ,q .
2 1Em with Em measurable,
P m
Proof of Claim 1: It is enough to show that if g =
pairwise disjoint, we have
0
k2m |Em |1/p k`q0 . 1.
 0
0 q −1
Choose f = |m|≤M 2m |Em |1/p |Em |−1/p 1Em . Then
P

Z
0 0 0 q 0 /q
`(f ) = f (x)g(x) dx ∼ k2m |Em |1/p kqq0 , kf kLp,q ∼ k2m |Em |1/p k 0 .
`|m|≤M `q|m|≤M

We have `(f ) . kf k∗Lp.q , so

0 0 1/(q−1)
k2m |Em |1/p k`q0 . k2m |Em |1/p k 0 .
|m|≤M `q|m|≤M

This gives
0
k2m |Em |1/p k`q0 .1
|m|≤M

0 0
uniformly in M . So g ∈ Lp ,q .
Proof of Claim 2: Consider f ≥ 0 and look at fm = f 1{2m ≤f <2m+1 } . For 2m+n ≤ k ≤
2 m+1+n − 1, let
  2m+n+1
X −1
k k+1 k
k
Em,n =f −1
, , ϕm,n = 1E .
2n 2n 2n m,nk
k=2

Then 0 ≤ fm −ϕm,n ≤ 21n . First choose ε > 0. If we look at kf k∗Lp,q kkfm kLp k`q , only finitely
many terms matter, so we can truncate the series. This lets us estimate kfm − ϕm,n k∗p0 ,q0 ,
as any large numbers we get will be multiplied by our small step size, 21n .

22
6 Hunt’s Interpolation Theorem
6.1 Strong type and weak type
Definition 6.1. We say that a map T on some measurable class of functions is sublinear
if

1. |T (cf )| ≤ |c||T f |,

2. |T (f + g)| ≤ |T (f )| + |T (g)|

for all constant c ∈ C and f, g in the domain of T .

Example 6.1. If T is linear, it is sublinear.

Example 6.2. If {Tt }t∈S is a family of linear maps, then

(T f )(x) = k(Tt f )(x)kL2t

is a sublinear map.

Definition 6.2. Let 1 ≤ p, q ≤ ∞, and let T be a sublinear map.

1. We say that T is of (strong) type (p, q) if there exists a constant C > 0 such that

kT f kLq (Rd ) ≤ Ckf kLp , ∀f ∈ Lp (Rd ).

2. If q < ∞, we say that T is of weak-type (p, q) if there exists a constant C > 0 such
that
kT f k∗Lq,∞ ≤ Ckf kLp (Rd ) ∀f ∈ Lp (Rd ).
If q = ∞, we say that T is of weak-type (p, q) if it is of strong type (p, q).

3. If p, q < ∞, we saw that T is of restricted weak-type (p, q) if there exists a constant

C > 0 such that

kT 1F k∗Lq,∞ ≤ C|F |1/p (≤ Ck1F k∗Lp,1 ) ∀F ⊆ Rd , |F | < ∞.

Remark 6.1.

Strong type (p, q) =⇒ weak-type (p, q) =⇒ restricted weak-type (p, q).

For the first implication, we have kT f k∗Lq,∞ . kT f k∗Lq,q . kf kLp . For the second implica-
tion,
kT 1F k∗Lq,∞ . k1F kLp = k1F k∗Lp,p . k1F k∗Lp,1 . |F |1/p .

23
Exercise 6.1. For 1 < p, q < ∞, let T be defined on functions on (0, ∞) via
Z ∞
−1/q 0
(T f )(x) = |x| |y|−1/p f (y) dy.
0

Then T is of restricted weak-type (p, q) but not of weak type (p, q).
Remark 6.2. Fix 1 < p, q < ∞. If T is of restricted weak-type (p, q), then for any
finite-measure sets E, F ⊆ Rd ,
Z
|(T 1F )(x)| · |1E (x)| dx . kT 1F k∗Lq,∞ k1E k∗Lq0 ,1 . |F |1/p |E|1/q .
0

Conversely, if this condition holds for all finite measure sets E, F ⊆ Rd , then T is of
restricted weak-type (p, q). Indeed,
Z
kT 1F kLq,∞ ∼ sup
∗
T 1F (x)g(x) dx .
kgk∗ q0 ,1 ≤1
L

2m 1Em with Em measurable and pairwise disjoint. Then

P
Take g =
Z Z
T 1F (x)g(x) dx ≤ |T 1F (x)| · |1Em (x)| dx
X
m
2
0
X
. 2m |F |1/p |Em |1/q
. |F |1/p kgk∗Lq0 ,1
. |F |1/p .

Remark 6.3. If 1 < p, q < ∞, then T is of restricted weak-type (p, q) if and only if there
is a constant C > 0 such that

kT f k∗Lq,∞ ≤ Ckf k∗Lp,1 ∀f ∈ Lp,1 (Rd ).

6.2 Hunt’s interpolation theorem

Theorem 6.1 (Hunt’s interpolation theorem). Let 1 ≤ p1 , p2 , q1 , q2 ≤ ∞ with p1 < p2 and
q1 6= q2 . Assume that T is a sublinear map satisfying kT f kLqj ,∞ . kf k∗ pj ,1 for j = 1, 2.
L
Then, for any 1 ≤ r ≤ ∞ and θ ∈ (0, 1), we have
1 θ 1−θ 1 θ 1−θ
kT f k∗Lqθ ,r . kf k∗Lpθ ,r , = + , = + .
pθ p1 p2 qθ q1 q2
Remark 6.4. 1. If pθ ≤ qθ , then T is of strong type (pθ , qθ ). Indeed, taking r = qθ , we
get
kT f kLqθ . kf k∗Lpθ ,qθ . kf kLpθ .

24
 2. The condition pθ ≤ qθ is needed to obtain the strong-type conclusion. For example,
let (T f )(x) = f (x)|x|−1/2 . Then T : Lp (0, ∞) → L2p/(p+2),∞ (0, ∞) boundedly for
any 2 ≤ p < ∞. But T is not bounded from Lp to L2p/(p+2) for all 2 < p < ∞. To
see that T : Lp → L2p/(p+2),∞ is bounded, we use the Hölder inequality in Lorentz
spaces (which we will prove later): If 1 ≤ p1 , p2 , p < ∞ and 1 ≤ q1 , q2 , q ≤ ∞, then
1 1 1 1 1 1
kf1 f2 k∗Lp,q . kf1 k∗Lp1 ,q1 kf1 k∗Lp2 ,q2 , = + , = + .
p p1 p2 q q1 q2
Then
kT f k∗L2p/(p+2),∞ . k|x|−1/2 k∗L2,∞ kf k∗Lp,∞ . kf kLp .
Take
f (x) = |x|−1/p | log(x + 1/x)|−(p+2)/(2p) .
We get
Z ∞
dx
kf kpLp = | log(x + 1/x)|−(p+2)/(2p)
0 x
XZ 2n+1
dx
= | log(x + 1/x)|−1−p/2
n x
n∈Z 2
X
∼ |n|−1−p/2
n∈Z
<∞

On the other hand,

Z ∞
2p/(p+2) dx
kT f kL2p/(p+2) = | log(x + 1/x)|−1
0 x
X
∼ |n|−1
n∈Z
=∞

We know that T : Lp1 ,p1 → L2p1 /(p1 +2),∞ and T : Lp2 ,p2 → L2p2 /(p2 +2),∞ for 2 ≤ p1 <
p2 < ∞. Hunt’s interpolation theorem gives that for all 1 ≤ r ≤ ∞, T : Lpθ ,r →
L2pθ /(pθ +2),r . Note that p2p θ
θ +2
< pθ .

Next time, we will prove the following as a consequence of Hunt’s interpolation theorem.

Corollary 6.1 (Marcinkiewicz interpolation theorem). Let 1 ≤ p1 ≤ q1 ≤ ∞ and 1 ≤

p2 ≤ q2 ≤ ∞ with p1 ≤ p2 and q1 6= q2 . Let T be a sublinear map that satisfies

kT f k∗Lqj ,∞ . kf kLpj , j = 1, 2.

25
Then for any θ ∈ (0, 1), T is of strong type (pθ , qθ ), where

1 θ 1−θ 1 θ 1−θ
= + , = + .
pθ p1 p2 qθ q1 q2
We will also prove Hunt’s theorem next time.

26
7 Proofs of Interpolation Theorems
7.1 Proof of the Marcinkiewicz interpolation theorem
Last time, we introduced Hunt’s interpolation theorem.

Theorem 7.1 (Hunt’s interpolation theorem). Let 1 ≤ p1 , p2 , q1 , q2 ≤ ∞ with p1 < p2 and

q1 6= q2 . Assume that T is a sublinear map satisfying kT f kLqj ,∞ . kf k∗ pj ,1 for j = 1, 2.
L
Then, for any 1 ≤ r ≤ ∞ and θ ∈ (0, 1), we have

1 θ 1−θ 1 θ 1−θ
kT f k∗Lqθ ,r . kf k∗Lpθ ,r , = + , = + .
pθ p1 p2 qθ q1 q2
Before proving this, we will prove the Marcinkiewicz interpolation theorem as a corol-
lary.

Corollary 7.1 (Marcinkiewicz interpolation theorem). Let 1 ≤ p1 ≤ q1 ≤ ∞ and 1 ≤

p2 ≤ q2 ≤ ∞ with p1 ≤ p2 and q1 6= q2 . Let T be a sublinear map that satisfies

kT f k∗Lqj ,∞ . kf kLpj , j = 1, 2.

Then for any θ ∈ (0, 1), T is of strong type (pθ , qθ ), where

1 θ 1−θ 1 θ 1−θ
= + , = + .
pθ p1 p2 qθ q1 q2

Proof. As p1 ≤ q1 and p2 ≤ q2 , we get pθ ≤ qθ for all θ ∈ (0, 1). If p1 < p2 , Hunt’s theorem
yields
kT f k∗Lqθ ,r . kf k∗Lpθ ,r ∀1 ≤ r ≤ ∞.
Taking r = qθ , we get
kT f kLqθ . kf k∗Lpθ ,qθ . kf kLpθ .
Assume now that p1 = p2 =: p. Then

kT f k∗Lq1 ,∞ . kf kLp ⇐⇒ sup λ|{x : |T f (x)| > λ}|1/q1 . kf kp

λ>0
kf kp q1
 
=⇒ |{x : |T f (x)| > λ}| . ∀λ > 0.
λ

Similarly,

kT f k∗Lq1 ,∞ . kf kLp ⇐⇒ sup λ|{x : |T f (x)| > λ}|1/q2 . kf kp

λ>0
kf kp q2
 
=⇒ |{x : |T f (x)| > λ}| . ∀λ > 0.
λ

27
We now have
∞ Z
dλ
kT f kqLθqθ = qθ λqθ |{x : |(T f )(x)| > λ}|
λ
Z ∞0 q1 
kf kp q2 dλ
  
qθ kf kp
. λ min ,
0 λ λ λ
Say q1 < q2 .
kf kp  q1 ∞  q2
kf kp kf kp
Z Z
qθ dλ dλ
. λ +
0 λ λ kf kp λ λ
. kf kqp1 kf kqpθ −q1 + q2 qθ −q2
kf kp kf kp
. kf kqpθ .

7.2 Proof of Hunt’s interpolation theorem

Now let’s prove Hunt’s interpolation theorem. Recall that if 1 < p, q < ∞, T is of restricted
weak type (p, q) if
kT 1F k∗Lq,∞ . |F |1/p
for every finite measure set F . We saw that this is equivalent to
Z
|T 1F (x)||1E (x)| dx . |F |1/p |E|1/q ∀E, F ⇐⇒ ||T f k∗Lqθ ,∞ . kf k∗Lp,1
0
∀f ∈ Lp,1 .

Proof. Claim: It suffices to prove Hunt for 1 < p1 , p2 , q1 , q2 < ∞. Indeed, for every
θ ∈ (0, 1),
kT f k∗Lqθ ,∞ . kf k∗Lpθ ,1
Indeed, for any θ ∈ (0, 1), even if p1 = 1 and q1 = ∞, pθ ∈ (1, ∞). So we can use an
interpolation argument with a slightly modified p1 and p2 : It suffices to see that

kT 1F k∗qθ ,∞ . |F |1/pθ

for all finite measure sets F . We have

kT 1F k∗Lqθ ,∞ = sup λθ+(1−θ) |{x : |T f (x)| > λ}|θ/q1 +(1−θ)/q2

λ>0
 θ  1−θ
≤ sup |{x : |T 1F (x)| > λ} 1/q1
sup |{x : |T 1F (x)| > λ}1/q2
λ>0 λ>0

1 1

θ
= kT F k∗Lq1 ,θ (kT F k∗Lq2 ,∞ )1−θ
1/p1 ·θ 1/p2 ·(1−θ)
. |F | |F |
. |F |1/pθ .

28
 Henceforth, we assume 1 < p1 , p2 , q1 , q2 < ∞. We can write
Z
kT f k∗Lqθ ,r ∼ sup T f (x) · g(x) dx ,
kgk∗ ≤1
q 0 ,r 0
L θ

so it’s enough to show that

Z
T f (x)g(x) dx . 1 ∀kf k∗Lpθ ,1 = 1, kgk∗ q0 ,r0 . 1.
L θ

By splitting into real and imaginary parts (andPthen positive and negative parts), we may
assume f, g ≥ 0. We may also assume g = 2m 1Em , where Em are measurable and
pairwise disjoint. Caution: As T need not have monotonicity properties, we may not
assume f is a simple function.
Using the binary expansion, we write
X
f (x) = 2n an (x), an (x) ∈ {0, 1}.
n∈Z

Note that there exists a largest n(x) such that an(x) = 1 and an (x) = 0 for all n > n(x).
Also, we don’t allow recurrent 1s. Let {nk (x)}k≥1 be a decreasing sequence such that
ank (x) (x) = 1 and all other an (x) = 0. Then
X
f (x) = 2nk (x) .
k≥1

For ` ≥ 1, let f` (x) = 2n` (x) . We can write

2n 1Fn` ,
X
f` (x) = Fn` = {x : n` (x) = n}.
n∈Z

Then

f (x) = 2n` (x) + 2n`−1 (x) + · · · + 2n1 (x)

≥ 2n` (x) + 2n` (x)+1 + · · · + 2n` (x)+`−1
= 2n` (x) · (2` − 1)
≥ f` (x) · 2`−1 .

So we get
1
f` (x) ≤ f (x).
2`−1
As Lpθ ,r is a Banach space, then `≥1 f` = f in Lpθ ,r .
P

29
 Now we can tackle the bound:
Z " #
X Z
2m 1Em (x) dx
X
T f (x)g(x) dx ≤ T f` (x)
`≥1 m
Z
|T 1Fn` (x)||1Em (x)| dx
X X
≤ 2n 2m
`≥1 n,m∈Z
0 0
XX
. 2n 2m min{|Fn` |1/p1 |Em |1/q1 , |Fn` |1/p2 |Em |1/q2 }.
`≥1 n,m

We will show that this is . 1 next time.

30
8 Interpolation and Maximal Function Estimates
8.1 Conclusion of proof of Hunt’s interpolation theorem
Theorem 8.1 (Hunt’s interpolation theorem). Let 1 ≤ p1 , p2 , q1 , q2 ≤ ∞ with p1 < p2 and
q1 6= q2 . Assume that T is a sublinear map satisfying kT f kLqj ,∞ . kf k∗ pj ,1 for j = 1, 2.
L
Then, for any 1 ≤ r ≤ ∞ and θ ∈ (0, 1), we have
1 θ 1−θ 1 θ 1−θ
kT f k∗Lqθ ,r . kf k∗Lpθ ,r , = + , = + .
pθ p1 p2 qθ q1 q2
Proof. We may assume 1 < p1 , p2 , q1 , q2 < ∞ with p1 < p2 and q1 6= q2 . We know
Z
|T 1F (x)||1E (x)| dx . min{|Fn` |1/p1 |Em |1/q1 , |Fn` |1/p2 |Em |1/q2 }
0 0

Fix θ ∈ (0, 1) and 1 ≤ e ≤ ∞. We want to show that

kT f k∗Lqθ ,r . kf k∗Lpθ ,r

uniformly for f ∈ Lpθ ,r . It suffices to show that

Z
T f (x)g(x) dx . 1,

where kf k∗Lpθ ,r ∼ 1 and g = m 1Em , where Em are measurable, pairwise disjoint,

P
m∈Z 2
and
0
kgk∗ q0 ,r0 ∼ k2m |Em |1/qθ k`r0 . 1.
L θ
We write f = `≥1 f` , where f` = n∈Z 2n 1Fn` . We have
P P

Z
0 0
XX
T f (x)g(x) dx . 2n 2m min{|Fn` |1/p1 |Em |1/q1 , |Fn` |1/p2 |Em |1/q2 }
`≥1 n,m
0
X X
. 2n |Fn` |1/pθ 2m |Em |1/qθ
`≥1 n,m∈Z
0 0
· min{|Fn` |(1−θ)(1/p1 −1/p2 ) |Em |(1−θ)(1/q1 −1/q2 )
0 0
|Fn` |−θ(1/p1 −1/p2 ) |Em |−θ(1/q1 −1/q2 ) }

Using the same trick we have used before, we write this as a geometric series.
0
X X X X
. 2n N 1/pθ 2m M 1/qθ A(N, M )
` N,M ∈2Z n:|Fn` |∼N ` |∼M
m:|Em

where

31
 0 0 0 0
A(N, M ) = min{|Fn` |(1−θ)(1/p1 −1/p2 ) |Em |(1−θ)(1/q1 −1/q2 ) , |Fn` |−θ(1/p1 −1/p2 ) |Em |−θ(1/q1 −1/q2 ) }.

  r 1/r
X X X
.  A(N, M )  2n N 1/pθ  
`≥1 N,M ∈2Z n:|Fn` |∼N
  r 1/r
1/qθ0
X X
· A(N, M )  2m M   .
N,M ∈2Z m:|Em |∼M
P P
Note that supN M ∈2Z A(N, M ) . 1 and supM ∈2Z N ∈2Z A(N, M ) . 1. Fix M ∈ 2Z .
1/p −1/p2 −(1/q 0 0
Let n0 1 ∼ M 1 −1/q2 ) . Then

0 0 0 0
X X X
A(N, M ) = N (1−θ)(1/p1 −1/2 ) M (1−θ)(1/q1 −1/q2 ) + N −θ(1/p1 −1/p2 ) M −θ(1/q1 −1/q2 ) .
N ∈2Z N ≤N0 N >N0

Thus,
  r 1/r   r 1/r
Z X X  X 
1/qθ0
X X
T f (x)g(x) dx .  2n N 1/pθ   2m M 
   
``≥1 N n:|Fn |∼N M m:|Em |∼M
!1/r !1/r0
0 0 0
X X X
. 2nr |Fn` |r/pθ 2mr |Em |r /qθ
`≥1 n m
| {z }| {z }
kf` k∗ pθ ,r kgk∗
L q 0 ,r 0
L θ
X
. kf` k∗Lpθ ,r
`≥1

1
Since |f` | ≤ 2`−1
|f |,

. kf k∗Lpθ ,r
∼ 1.

Remark 8.1. We did not use anything specific about Lebesgue measure in our proof. So
these theorems hold for arbitrary measures µ.

8.2 Maximal and vector maximal functions

Recall the Hardy-Littlewood maximal function
Z
1
M f (x) = sup |f (y)| dy.
r>0 |B(x, r)| B(x,r)

32
Theorem 8.2.

1. If f ∈ Lp (Rd ) for some 1 ≤ p ≤ ∞, then M f is finite almost everywhere.

2. M is of weak-type (1, 1) and strong-type (p, p) for 1 < p ≤ ∞.

Remark 8.2.

1. M is not of strong-type (1, 1). Let ϕ ∈ Cc∞ (B(0, 1/2)). For |x| ≤ 1, M ϕ(x) ∼ 1.
If |x| > 1, then M ϕ(x) ∼ |x|1 d . So M ϕ(x) ∼ hxi−d , where this notation means
hxi := (1 + |x|2 )1/2 .2 So M ϕ ∈
/ L1 .

2. M is of weak-type (1, 1) means

kf k1
|{x : M f (x) > λ}| .
λ
uniformly in λ > 0 and f ∈ L1 . The decay in λ on the right hand side cannot be
improved. To see this, consider ϕ as above. Then M ϕ ∈ L∞ , so only the small λ are
relevant. For small λ,

|{x : M f (x) > λ}| = |{x : hxi−d & λ}|

= |{x : hxi . λ−1/d }|
. λ−1 .

/ L1,q (Rd ) for any q < ∞ because

Also, M ϕ ∈
Z ∞
dλ
kM ϕk∗L1,q ∼ λq |{x : M ϕ(x) > λ}|q = ∞.
0 | {z } λ
.λ−q

Theorem 8.3. Let ω : Rd → [0, ∞) be a locally integrable function (a weight), to which

we associate a measure via Z
ω(E) = ω(x) dx.
E
Then

1. M : L1 (M ω dx) → L1,∞ (ω dx) maps boundedly; that is,

Z
1
ω({x : M f (x) > λ}) . |f (y)|(M ω)(y) dy.
λ
2
This is known as “Japanese bracket notation” everywhere except Japan, where they just call it “bracket
notation.”

33
 2. M : Lp (M ω dx) → Lp (ω dx) boundedly for all 1 < p ≤ ∞; that is,
Z Z
|M f (x)|p ω(x) dx . |f (y)|p (M ω)(y) dy.

Remark 8.3.

1. If ω ≡ 1, then M ω ≡ 1, so we recover the previous theorem.

2. In order for the statement to

R be non-vacuous, we need M ω is finite somewhere. This
happens precisely when r1d |x|≤r ω(x) dx . 1 uniformly for sufficiently large r.
( =⇒ ): If x = 0, we are done, so assume x 6= 0. For r > 2d(x, 0),
Z Z
1 1
M ω(x) ≥ ω(y) dy & d ω(y) dy.
|B(x, r)| B(x,r) r |x|≤r/2

( ⇐= ): Choose x to be a Lebesgue point. The Lebesgue differentiation theorem

controls the maximal function at small scales, and the same argument controls the
maximal function at large scales.

34
9 Boundedness Properties of The Hardy-Littlewood Maxi-
mal Function and Ap Weights
9.1 Boundedness properties of the Hardy-Littlewood maximal function
The Hardy-Littlewood maximal function is given by
Z
1
M f (x) = sup |f (y| dy.
r>0 |B(x, r)| B(x,r)

Theorem 9.1. Let ω : Rd → [0, ∞) be a locally integrable function (a weight), to which

we associate a measure via Z
ω(E) = ω(x) dx.
E
Then

1. M : L1 (M ω dx) → L1,∞ (ω dx) maps boundedly; that is,

Z
1
ω({x : M f (x) > λ}) . |f (y)|(M ω)(y) dy
λ

uniformly in λ > 0 for all f ∈ L1 (M ω dx).

2. M : Lp (M ω dx) → Lp (ω dx) boundedly for all 1 < p ≤ ∞; that is,

Z Z
|M f (x)| ω(x) dx . |f (y)|p (M ω)(y) dy
p

uniformly for f ∈ Lp (M ω dx).

Just like the proof of the maximal inequality, we will start with a covering lemma.

Lemma 9.1 (Vitali). Given a finite collection of balls {B(xj , rj )}j∈J , there exists a sub-
collection S such that

1. Distinct balls are disjoint.

S S
2. j∈I B(xj , rj ) ⊆ j∈S B(xj , 3rj ).

Proof. We run the following algorithm. Set S = ∅.

1. Choose a ball of largest radius and add it to S.

2. Discard any balls that intersect balls in S.

3. If no balls remain, stop. Otherwise, return to step 1.

35
 Now let’s prove the theorem.

Proof. First note that M : L∞ (M ω dx) → L∞ (ω dx) boundedly:

kM f kL∞ (ω dx) = inf sup M f (x)

E:ω(E)=0 x∈E c

Since ω is locally integrable, it takes Lebesgue-null sets to ω-null sets.

≤ inf sup M f (x)

E:|E|=0 x∈E c

≤ kf kL∞ (dx)
= inf sup |f (x)|
E:|E|=0 x∈E c

M ω > 0 unless ω ≡ 0, so

= inf sup |f (x)|

E:(M ω)(E)=0 x∈E c

= kf kL∞ (M ω dx) .

So by the Marcinkiewicz interpolation theorem, it suffices to prove M : L1 (M ω dx)øL1,∞ (ω dx).

Fix λ > 0. Let K be a compact subset of {x : M f (x) > λ} (this suffices by regularity).
For x ∈ K, there is some r(x) > 0 such that
Z
1
|f (y)| dy > λ.
|B(x, r(x))| B(x,r(x))
S
Now
S K ⊆ x∈K B(x, r(x)), and by compactness, there exists a finite subcover such that
B(xj , rj ). By Vitali, there exists
j∈J S P a subcollection S of pairwise disjoint balls such that
K ⊆ j∈S B(xj , 3rj ). So ω(K) ≤ j∈S ω(B(xj , 3rj )).
For Lebesgue measure, we would just pull out the constant 3 and add the measures.
But here, we don’t have that property, so we will relate it to the maximal function. For
x ∈ B(xj , rj ),
Z
ω(B(xj , 3rj )) = ω(y) dy
B(xj ,3rj )
|B(x, 4rj )|
Z
≤ ω(y) dy
|B(x, 4rj )| B(x,4rj )

≤ 4d |B(x, 4rj )|M ω(x).

Now integrate this against f :

Z Z
1 d
ω(B(xj , 3rj )) |f (y)| dy ≤ 4 M ω(x)|f (y)| dy.
|B(xj , rj )| B(xj ,rj ) B(xj ,rj )

36
Remark 9.1. Rather than placing the weights outside the maximal function, one could
place them inside: Define
Z
1
Mµ f (x) = sup |f (y)| dµ(y),
r>0 µ(B(x, r)) B(x,r)

where µ is a nonnegative measure. If µ is a doubling measure (i.e. if B1 = B(x, r) and

B2 = B(x, 2r), then µ(B2 ) . µ(B1 ) uniformly for x ∈ Rd and r > 0), then with small
modifications, the proof of this theorem yields:

Mµ : L1 (dµ) → L1,∞ (dµ), Mµ : Lp (dµ) → Lp (dµ), ∀1 < p ≤ ∞

boundedly.

9.2 Ap weights
Can one characterize the nonnegative measure µ for which

M : Lp (dµ) → Lp (dµ), 1<p<∞

boundedly? Yes, these are the Ap weights.

Definition 9.1. We say that a locally integrable weight ω : Rd → [0, ∞) satisfies the
A1 condition (and we write ω ∈ A1 ) if there is a C > 0 such that M ω(x) ≤ Cω(x) for
almost every x.
Remark 9.2. If ω ∈ A1 , then the theorem yields

M : Lp (ω dx) → Lp (ω dx) M : L1 (ω, dx) → L1,∞ (ω dx) ∀1 < p ≤ ∞

boundedly.
Let’s characterize these weights.
Lemma 9.2. The following are equivalent:
1. ω ∈ A1

2. Z
1
ω(y) dy . ω(x)
|B| B
uniformly for a.e. x ∈ B and all balls B.

3. Z Z
1 1
f (y) dy . f (y)ω(y) dy
|B| B ω(B) B
for all balls B and all f ≥ 0.

37
Proof. (1) =⇒ (2): Fix x with M ω(x) ≤ Cω(x), and let B be a ball of radius r that
contains x. Then
2d
Z Z
1
ω(y) dy ≤ ω(y) dy
|B| B |B(x, 2r)| B(x,2r)
≤ 2d M ω(x)
≤ 2d Cω(x).
(2) =⇒ (3): ω is bounded below by its maximal function, so
Z Z  Z 
1 1 1
f (y)ω(y) dy ≥ f (y) ω(z) dz dy
ω(B) B ω(B) B |B| B
Z
1
≥ f (y) dy.
|B| B
(3) =⇒ (2): Let x be a Lebesgue point for ω, and let B 3 x. Let r  1 be such that
B(x, r) ⊆ B. Set f = 1B(x,r) . Then
Z
1 1
|B(x, r)| . ω(y) dy.
|B| ω(B) B(x,r)
Rearranging this, we get
Z
ω(B) 1
. ω(y) dy → ω(x).
|B| |B(x, r)| B(x,r)

Definition 9.2. We say that a weight ω : Rd → [0, ∞) satisfies the Ap condition for
1 < p < ∞ if there exists an A > 0 such that
Z  Z p/p0
1 1 −p0 /p
sup ω(y) dy · ω(y) dy ≤ A,
balls B |B| B |B| B
or equivalently,
sup |B|−p ω(B)kω −1/(p−1) kp−1
L1 (B)
≤ A.
balls B
Remark 9.3.
1. This condition is invariant under ω 7→ λω and ω(x) 7→ ω(λx).
0
2. ω ∈ Ap if and only if σ = ω −p /p ∈ Ap0 . Indeed, the condition reads:
Z Z p/p0
1 −p/p0
sup p
σ(y) dy σ(y) dy ≤ A.
balls B |B| B
If we raise everything to the power p0 /p,
Z Z p0 /p
1 −p/p0 0
sup p 0 σ(y) dy σ(y) dy ≤ Ap /p .
balls B |B| B B

38
10 Characterization of Ap Weights
10.1 Ap weights for p > 1
Last time, we began to tackle the problem of characterizing nonnegative measures µ for
which Z
|M f (y)|p dµ(y) . |f (y|p dµ(p).

uniformly for f ∈ Lp (dµ) and some 1 < p < ∞. We will not prove the full details, but we
will give a compelling intuition of the results.
Fix 1 < p < ∞. Recall that a locally integrable weight ω : Rd → [0, ∞) satisfies the Ap
condition if there is an A > 0 such that
Z  Z p/p0
1 1 −p0 /p
sup ω(y) dy ω (y) dy ≤ A.
balls B |B| B |B| B

This is equivalent to
sup |B|−p ω(B)kω −1/(p−1) kp−1
L1 (R)
≤ A.
balls B

Remark 10.1. If 1 < p < q < ∞, then Ap ⊆ Aq : Let ω ∈ Ap . Then by Hölder,

kω −1/(q−1) kL1 (B) ≤ kω −1/q−1 kL(q−1)/(p−1) |B|1−(p−1)/(q−1)

(p−1)/(q−1)
= kω −1/(p−1) kL1 (B) |B|(q−p)/(q−1) .

So we get
q−1 p−1
|B|−q ω(B)kω −1/(q−1) kL1 (B) ≤ |B|
−p
ω(B)kω −1/(p−1) kL1 (B) . 1

uniformly in B.

Lemma 10.1. Fix 1 ≤ p < ∞. Then ω ∈ Ap if and only if

 Z p Z
1 1
f (y) dy . |f (y)|p ω(y) dy,
|B| B ω(B) B

uniformly in f ≥ 0 and balls B.

We proved this last time for p = 1.

Proof. It remains to consider 1 < p < ∞.

( =⇒ ):
 Z p  Z p
1 1
f (y) dy = f (y)ω 1/p−1/p dy
|B| B |B| B

39
 Z Z h p/p0 i
−p0 /p
≤ |B|−p |f (y)|p ω(y) dy ω(y) dy
B | {z }
.|B|p ·1/ω(B)
Z
1
. |f (y)|p ω(y) dy.
ω(B) B
0
( ⇐= ): Fix ε > 0 and a ball B. Let f = (ω + ε)−p /p . Then
 Z p Z
1 −p0 /p 1 0
(ω + ε) (y) dy . (ω + ε)−p (y)ω(y) dy
|B| B ω(B) B
Z
1 0
. (ω + ε)−p +1 (y) dy
ω(B)

Note that p0 − 1 = −p0 /p.

Z
1 0
. (ω + ε)−p /p (y) dy.
ω(B) B

So Z p/p0
0
|B|−p ω(B) (ω + ε)−p /p (y) dy .1
B
uniformly in B and ε > 0. Let ε → 0 and use the monotone convergence theorem

Corollary 10.1. Fix 1 ≤ p < ∞. If ω ∈ Ap , then ω is a doubling measure.

Proof. Let B = B(x, 2r) and f = 1B(x,r) . Then

 p
|B(x, r)| ω(B(x, r))
.
|B(x, 2r)| ω(B(x, 2r))

uniformly in x ∈ Rd and r > 0.

Remark 10.2. In fact, Ap weights with 1 ≤ p < ∞ satisfy a “fairness condition”: If

F ⊆ B, then taking f = 1F , we get

|F | p ω(F )
 
. .
|B| ω(F )

So if F is a large chunk of the ball B, ω has to give a large proportion of the measure of
the ball to F ; it has to treat F fairly.

40
10.2 Proof sketch of characterization
Theorem 10.1. Fix 1 < p < ∞. Then ω ∈ Ap if and only if
Z Z
|M f (y)| ω(y) dx . |f (y)|p ω(y) dy
p

uniformly for f ∈ Lp (ω dx) (that is, M : Lp (ω dx) → Lp (ω dx) boundedly).

This answers the question we proposed but only in the case that ω is a weight; i.e. ω
is absolutely continuous with respect to Lebesgue measure.

Remark 10.3. ( ⇐= ): The necessity holds under even weaker assumptions. If M :

Lp (ω dx) → Lp,∞ (ω dx) boundedly, then ω ∈ Ap .

Proof. ( ⇐= ): Fix a ball B of radius r > 0. Fix ε > 0, and let f = (ω + ε)−p /p 1B . For
0

x ∈ B,
Z
1
(ω + ε)−p /p 1B (x) dx
0
M f (x) = sup
R>0 |B(x, R)| B(x,R)
Z
1 0
≥ (ω + ε)−p /p (y) dy
|B(x, 2r)| B
Z
1 0
= d (ω + ε)p /p (y) dy
2 |B| B

Let’s give this a name

=: 2λ.

Now

ω(B) ≤ ω({x : M f (x) > λ})

−p0 (y)ω(y) dy
R
B (ω + ε)
.
λp
0
(ω + ε)−p /p (y) dy
R
B
. R .
[ B (ω + ε)−p0 /p (y) dy]p

So we get that
Z p/p0
−p p0 /p
|B| ω(B) (ω + ε) (y) dy . 1.
B

Remark 10.4. ( =⇒ ): The sufficiency, that is, ω ∈ Ap =⇒ M is bounded on Lp (ω dx),

rests on three ingredients:

41
 1. If 1 ≤ q < ∞, then M : Lq (ω dx) → Lq,∞ (ω dx) boundedly (this is Homework exercise
10). Look at Z
1
Mω f (x) = sup |f (y)|ω(y) dy.
r>0 ω(B(x, r)) B(x,r)

Then Mω : L1 (ω dx) → L1,∞ (ω dx) boundedly. Then

Z p Z
1 1
f (y) dy . |f (y)|p ω(y) dy,
|B| B ω(B) B

which tells us
|M f |p . Mω (|f |p ).

2. (Appears in Ch5 3
S of Stein’s Harmonic Analysis textbook ) A reverse Hölder inequality:
If ω ∈ A∞ = 1≤p<∞ Ap , then there exist and r > 1 and c > 0 (both depending on
ω) such that
 Z 1/r Z
1 r c 0
ω (y) dy ≤ ω(y) dy ⇐⇒ |B|1/r kωkLr (B) ≤ ckωkL1 (B) .
|B| B |B| B

This implies that if ω ∈ Ap , then ω ∈ Aq for some q < p.

Ingredients 1 and 2 give M : Lq (ω dx) → Lq,∞ (ω dx) boundedly for some q < p.

3. The Marcinkiewicz interpolation theorem with M : L∞ (ω dx) → L∞ (ω dx) (use the

fact that |E| = 0 ⇐⇒ ω(E) = 0 since ω > 0 a.e. as ω ∈ Ap ).

Next time, we will discuss how this generalizes to arbitrary measures, not just ones
absolutely continuous with respect to Lebesgue measure.

3
This book is the bible of Harmonic Analysis.

42
11 Ap Weights and The Vector-Valued Maximal Function
11.1 Use of reverse Hölder in the characterization of Ap weights
Last time, we proved the following theorem:
Theorem 11.1. Fix 1 < p < ∞. Then ω ∈ Ap if and only if M : Lp (ω dx) → Lp (ω dx)
boundedly.
We showed that ( ⇐= ) holds if L : Lp (ω dx) → Lp,∞ (ω dx). For the ( =⇒ ) direction,
we had 3 ingredients:
1. M : L1 (ω dx) → Lq,∞ (ω, dx) for all 1 ≤ q < ∞.

2. A reverse Hölder inequality yields if ω ∈ Ap , then ω ∈ Aq for some q < p.

3. M : L∞ (ω, dx)L∞ (ω dx) boundedly.

The reverse holds inequality says
Lemma 11.1. If ω ∈ Ap , then there exist an r > 1 and c > 0 such that
 Z 1/r Z
1 r c
ω(y) dy ≤ ω(y) dy.
|B| B |B| B

We will not prove this. Here’s how we use it:

Proof. Apply this to σ(y) = ω(y). Recall that ω ∈ Ap ⇐⇒ σ ∈ Ap . Then there exist
r > 1 and c > 0 depending on σ (and hence on ω) so that
 Z 1/r Z
1 −rp0 /p C 0
ω(y) dy ≤ ω(y)−p /p dy.
|B| B |B| B
So we get
 Z p/p0
1 1 −p0 /p
ω ∈ Ap ⇐⇒ sup ω(B) ω(y) dy .1
B |B| |B| B
0
|B| p /p |B| 1/(p−1)
Z   
1 −p0 /p
=⇒ ω(y) dy . . .
|B| B ω(B) ω(B)
We get
Z 1/r  1/(p−1)
−1/r −rp0 /p |B|
|B| ω(y) dy . .
B ω(B)
Write
1 1
rp0 /p = q 0 /q ⇐⇒ r =
p−1 q−1

43
 p−1
⇐⇒ a − 1 = <p−1
p
p−1
⇐⇒ q = 1 + ∈ (1, p).
p
We get
Z q/q0 ·1/(p−1)  1/(p−1)
−q/q 0 ·1/(p−1) −q 0 /q |B|
|B| ω(y) dy .
B ω(B)
Z q/q0
−1−(q−1) −q 0 /q
|B| ω(B) ω(y) dy . 1.
B
So ω ∈ Aq .

Theorem 11.2. Fix 1 ≤ p < ∞. If dµ is a nonnegative Borel measure such that M :

Lp (dµ) → Lp,∞ (dµ) boundedly, then dµ = ω dx and ω ∈ Ap .

Proof. It suffices to show that dµ is absolutely continuous with respect to Lebesgue mea-
sure. Write dµ = ω dx + dν with dν singular with respect to Lebsegue measure. Let K be
a compact set such that |K| = 0 and T ν(K) > 0. For n ≥ 1, let Un = {: d(x, k) < 1/n}.
Note that Un \ K ⊇ Un+1 \ K and (Un \ K) = ∅. Let fn = 1Un \K , so fn+1 ≤ fn and
fn → 0.
We claim that dµ is finite on compact sets. Assuming the claim, by the monotone
convergence theorem, Z
n→∞
|fn |p dµ −−−→ 0.

For x ∈ K,
Z
1
M fn (x) = sup 1Un \K (y) dy
r>0 |B(x, r)| B(x,r)
Z
1
≥ 1 (y) dy
|B(x, 1/n)| B(x,1/n) Un \K
Z
1
= 1 (y) dy
|B(x, 1/n)| K c B(x,1/n)

As |K| = 0,
Z
1
= 1B(x,1/n) (y) dy
|B(x, 1/n)| Rd
=1

Then Z
n→∞
µ(K) ≤ µ({x : M fn (x) > 1/2}) . |fn |p dµ −−−→ 0,

44
so we get a contradiction.
Now we prove the claim. Let E be a compact set such that 0 < µ(E) < ∞.
Z
1
M 1E (x) = sup 1E (y) dy
r>0 |B(x, r)| B(x,r)
|E|
&
[d(x, E) + diam(E)]d

So M 1E is bounded from below uniformly on compact sets: if F is a compact set, then for
x ∈ F,
|E|
M 1E (x) . =: C(F, E).
[dist(F, E) + diam F + diam E]d
Then
  Z
1
µ(F ) ≤ µ x : M 1E (x) > C(F, E) .F,E |1E (y)|p dµ(y) .F,E µ(E) < ∞.
2

11.2 The vector-valued maximal function

Definition 11.1. Let F : Rd → `2 , f (x) = {fn (x)}n≥1 . We write
Z 1/p
p
|f (x)| = k{fn (x)}n≥1 k`2 , kf k Lp = |f (x)| dx .

The vector-valued maximal function is

M f (x) = k{M fn (x)}n≥1 k`2 .

Theorem 11.3.

1. M is of weak-type (1, 1).

2. For 1 < p < ∞, M is of strong type (p, p).

Remark 11.1. We no longer have a trivial L∞ bound. In fact, it fails. Take d = 1. For
n ≥ 1, take fn = 1[2n−1 ,2n ) .
sX
|f (x)| = |fn |2 (x) = 1[1,∞) (x) ∈ L∞
n≥1

For |x| ≤ 2n ,
Z x+r
1
M fn (x) = sup 1[2n−1 ,2n ) (y) dy
r>0 2r x−r

45
 Z x+2n+1
1
≥ 1 n−1 n (y) dy
2 · 2n+1 x−2n+1 [2 ,2 )
1
= 2n−1
2 · 2n+1
1
= .
8
Now
sX
M f (x) = |M fn (x)|2
n≥1
v
u X  2
u 1
≥t
n
8
n:2 ≥|x|

=∞

/ L∞ .
So M f ∈

Remark 11.2. Boundedness of M on L2 follows from the scalar case:

Z X X X
2
kM f kL2 = |M fn (x)|2 = kM fn k2L2 . kfn k2L2
n≥1 n≥1 n≥1
XZ Z
= |fn (x)|2 dx ≤ |f (x)|2 dx = kf k2L2 .
n≥1

Let’s prove boundedness of M on Lp for 2 < p < ∞.

Proof. If ω ≥ 0 with ω ∈ L1loc , then

Z Z
|M fn | ω dx . |fn |2 (M ω) dx
2

uniformly in n. Summing in n, we get

Z Z
|M f (x)| ω(x) dx . |f (x)|2 (M ω)(x) dx.
2

Then

kM f k2Lp = k|M f |2 kLp/2

Z
= sup |M f |2 (x)ω(x) dx
kωk 0 ≤1
L(p/2)

46
 Z
. sup |f (x)|2 (M ω)(x) dx
kωk (p/2)0 ≤1
| {z } | {z }
L 0
∈Lp/2 ∈L(p/2)
2
. k|f | kLp/2 sup kM ωkL(p/2)0
kωk 0 ≤1 | {z }
L(p/2)
.kωk 0
L(p/2)

. kf k2Lp .

To prove M is of strong-type (p, p) for 1 < p < ∞, it suffices (by Marcinkiewicz) to

show that M is of weak-type (1, 1).

We will use the following.

Lemma 11.2 (A Calderón-Zygmund decomposition). If f ∈ L1 (Rd ) and λ > 0, then we

can decompose f = g + b such that

1. |g(x)| ≤ λ for almost every x ∈ Rd .

2. supp b is a union of cubes whose interiors are pairwise disjoint and

Z
1
λ< |b(x)| dx ≤ 2d λ.
|Qk | Qk

3. g = f [1 − 1S Qk ].

Next time, we will prove this decomposition and use it to prove the weak (1, 1) bound.

47
12 Calderón-Zygmund Decomposition and Bounds for the
Vector-Valued Maximal Function
12.1 A Calderón-Zygmund decomposition
Lemma 12.1 (A Calderón-Zygmund decomposition). If f ∈ L1 (Rd ) and λ > 0, then we
can decompose f = g + b such that
1. |g(x)| ≤ λ for almost every x ∈ Rd .
2. supp b is a union of cubes whose interiors are pairwise disjoint and
Z
1
λ< |b(x)| dx ≤ 2d λ.
|Qk | Qk

3. g = f [1 − 1S Qk ].
Remark 12.1. 1. Interpolating between the first conclusion and g ∈ L1 , we get g ∈ Lp
for all 1 ≤ p ≤ ∞.
1 P 1
P R R P
2. k |Qk | ∼ λ k k Qk |b(y)| dy, so k |Qk | . λ kf kL1 .
R
Remark 12.2. Modifying g further, we can ensure that Qk b(y) dy = 0 for all k. Indeed,
let ( S
f (x) x∈ / k Qk
g(x) = 1
R
|Qk | Qk f (y) dy x ∈ Qk .
Then for x ∈ Qk ,
Z
1
b(x) = f (x) − f (y) dy,
|Qk | Qk
so Z Z Z
b(x) dx = f (x) dx − f (y) dy = 0.
Qk Qk Qk
We lose a factor of 2 for the constant:
Z Z
1 2
|b(x)| dx ≤ |f (x)| dx ≤ 2d+1 λ.
|Qk | Qk |Qk | Qk
The price we have to pay is that |g(x)| ≤ 2d λ instead of λ.
Proof. Decompose Rd into dyadic cubes Q = [2n k1 , 2n (k1 + 1)] × · · · × [2n kd , 2n (kd + 1)],
where n is sufficiently large so that
Z
1
|f (y)| dy ≤ λ
|Q| Q
Fix such a Q and subdivide it into 2d congruent cubes (cut each side in half). Let Q0
denote one of the resulting children.

48
 • If 1
|f (y)| dy > λ, stop and add Q0 to the collection Qk .
R
|Q0 | Q0

• If |Q10 | Q0 |f (y)| dy ≤ λ, then continue subdividing until (if ever) we are forced into
R

case 1.

If we are in case 1, then

2d
Z Z
1
λ< 0 |f (y)| dy ≤ |f (y)| dy ≤ 2d λ.
|Q | Q0 |Q| Q

S remains to show that g = f [1 − 1 Qk ] satisfies |g| ≤ λ a.e. Fix a Lebesgue point

It S

x∈
/ Qk for f . Then
Z Z
1 1
f (y) dy − f (x) ≤ |f (y) − f (x)| dx
|Q| Q |Q| Q

for any cube, we can inscribe a ball in side it and we can circumscribe a ball around it.
Letting r ∼ diam(Q),
Z
1
. |f (y) − f (x) dx
|B(x, r)| B(x,r)
r→∞
−−−→ 0

So Z
1
f (x) = lim f (y) dy,
x3Q |Q| Q
diam Q→0

and we get |g(x)| = |f (x)| ≤ λ.

12.2 Weak-type bound for the vector-valued maximal function

Recall that for f : Rd to`2 with f = {fn }n≥1 , the vector-valued maximal function is

M f (x) = k{M fn }n≥1 k`2 .

Theorem 12.1.

1. M is of weak-type (1, 1).

2. For 1 < p < ∞, M is of strong type (p, p).

Proof. Last time, we remarked that we need only prove part 1. Fix f ∈ L1 and λ > 0. We
want to show that
kf kL1
|{x : M f (x) > λ}| . .
λ

49
 1
S R
Decompose f = g + b with |g| ≤ λ a.e., supp b = k Qk , and |Qk | Qk f |b(y)| dy ∼ λ. Then

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

By Chebyshev,

kM gk22 kgk22 λkgkL1 kf kL1

|{x : M g(x) > λ/2}| . 2
. 2
. 2
. .
λ λ λ λ
It is left to show that
kf kL1
|{x : M b(x) > λ/2}| . .
λ
We have
X X X1Z kf kL1
|2Qk | ≤ 2d |Qk | ∼ |b(y)| dy . .
λ Qk λ
k k k

We have to show now that

[ kf kL1
|{x ∈ [ (2Qk )]c : M b(x) > λ/2}| . .
λ
S
For x ∈
/ (2Qk ),
Z
1
M bn (x) = sup |bn (y)| dy
r>0 |B(x, r)| B(x,r)
1 XZ
= sup |bn (y)| dy
r>0 |B(x, r)| B(x,r)∩Qk k
√ √
If B(x, r) ∩ Qk 6= ∅, then r > `(Qk )/2. So Qk ⊆ B(x, r + d`(Qk )) ⊆ B(x, r(1 + 2 d)).

1 XZ
≤d sup √ √ |bn (y)| dy
r>0 |B(x, r(1 + 2 d))| B(x,r(1+2 d))
k
Z  Z 
1 1
1Qk (z)
X
. sup √ √ |bn (y)| dy dz.
r>0 |B(x, r(1 + 2 d))| B(x,r(1+2 d)) |Qk | Qk
k

Let bavg 1Qk |Q1k |

P R
n = Qk |bn (y)| dy. Then we have

M bn (x) . M bavg
n (x).

Let bavg = {bavg

n }n≥1 . For x ∈ Qk ,
Z
avg 1
|b (x)| = k{bavg
n (x)}n k`2 ≤ |b(y)| dy . λ.
|Qk | Qk

50
We also have
XZ
avg
kb kL1 = |b(y)| dy . kf kL1 .
k Qk

By Chebyshev, since M b(x) . M bavg (x),

[ [
|{x ∈ [ (2Qk )]c : M b(x) > λ/2}| . |{x ∈ [ (2Qk )]c : M bavg &}|
1
. kM bavg k|2L2
λ2
kbavg k2L2
.
λ2
avg
kb kL1 kf kL1
. . .
λ λ
Remark 12.3. One can replace `2 by `q for 1 < q ≤ ∞ for f : Rd toellq . Define

M q f (x) = k{M fn (x)}n≥1 k`q .

Then

1. M q is of weak-type (1, 1).

2. M q is of strong type (p, p) for all 1 < p < ∞.

The proof is as in the case q = 2 if 1 < q < ∞. The trivial estimate becomes that
M q : Lq → Lq is bounded. If q = ∞,

M ∞ f ≤ M k{fn }n≥1 k`∞ .

The estimates follow from the scalar case.

If q = 1, then these estimates fail. We will see an example next time.

51
13 The Hardy-Littlewood-Sobolev Inequality
13.1 Failure of bounds for L1 vector-valued maximal function
Let f : Rd → `1 , f = {fn }n≥1 with |f (x)| = n≥1 |fn (x)|. We define the L1 vector-valued
P
P
maximal function as M 1 f (x) = n≥1 M fn (x). The the following claims fail:

1. |{x : M 1 f (x) > λ}| ≤ λ1 kf kL1 uniformly for λ > 0, f ∈ L1 .

2. For 1 < p < ∞, kM 1 f kLp . kf kLp uniformly for f ∈ Lp .

Fix d = 1. Take [0, 1] and subdibide it into intervals I1 , . . . , IN of equal length. Let

1In 1 ≤ n ≤ N
(
fn =
0 n > N.

Let f = {fn }n≥1 . Then

|fn (x)| = 1[0,1] (x) ∈ Lp

X
|f (x)| = ∀1 ≤ p ≤ ∞,
n≥1

kf kLp = 1 ∀1 ≤ p ≤ ∞.
On the other hand,
N
X
M 1 f (x) = M fn (x)
n=1
N Z x+r
1
1In (y) dy
X
= sup
n=1 r>0 2r x−r

PbN/2c 1 1
For x ∈ [0, 1], M f (x) ≥ n=1 2(n/N ) · N & log(N ). This tells us that

1
|{x : M 1 f (x) > log N }| ≥ 1,
10
kM 1 f kLp & log N.

13.2 The Hardy-Littlewood-Sobolev inequality

Theorem 13.1 (Hardy-Littlewood-Sobolev). Fix 1 < p < r < ∞ and 1 < q < ∞ such
that 1 + 1r = p1 + 1q . Then
kf ∗ gkLr . kf kp kgk∗Lq,∞ ,

52
uniformly for f ∈ Lp , g ∈ Lq,∞ . In particular, for 0 < α < d,
1
f∗ . kf kLp ,
|x|α Lr
1 1
provided 1 + r = p + αd .
Proof. Fix g ∈ Lq,∞ . We may assume that kgk∗Lq,∞ = 1. We want to show that the
T
sublinear operator f 7→ f ∗ g is of strong type (p, r). By the Marcinkiewicz interpolation
theorem, it suffices to show T is of weak type (p, r) for all 1 < p < r < ∞ such that
1 + 1/r = 1/p + 1/q. Say the target is strong-type (p0 , r0 ). Then choose 1 < p1 < p0 <
p2 < ∞ and write p10 = pθ1 + 1−θ p2 . By Marcinkiewicz, if T is of weak-type (p1 , r1 ) and
(p2 , r2 ), then T is of strong-type (p0 , re), where
   
1 θ 1−θ 1 1 1 1 1 1 1
= + =θ + + −1 + (1 − θ) + −1 = + +1= .
re r1 r2 p1 q p2 q p0 q r0
Let’s show T is of weak-type (p, r):
 r
kf kp
|{x : |(f ∗ g)(x)| > λ}| .
λ
We may rescale so that kf kp = 1. Write g = g1 + g2 = g 1{|g|≤R} + g 1{|g|>R} . Then

|{x : |(f ∗ g)(x)| > λ}| . |{x : |(f ∗ g1 )(x)| > λ/2}| + |{x : |(f ∗ g2 )(x)| > λ/2}|

By Chebyshev,
kf ∗ g1 kss
|{x : |(f ∗ g1 )(x)| > λ/2}| .
λs
kf ksp kg1 ksps/(ps+p−s)
.
λs
∞
dα (ps+p−s)/p
Z 
. λ−s αps/(ps+p−s) 1{|g1 |>α}
0 α
Z R (ps+p−s)/p
dα
=λ −s
α ps/(ps+p−s)
1{|g|>α}
0 α
Z R (ps+p−s)/p
−s ∗ q·(ps+p−s)/p ps/(ps+p−s)−q dα
= λ (kgkLq,∞ ) α
0 α
. λ−s Rs−q·(ps+p−s)/p ,
1 1
Provided q >1+ s − p1 . Since 1 + 1
r = 1
q + p1 , this means that s > r.

53
 On the other hand, by Chebyshev,

kf ∗ g2 kpp
|{x : |(f ∗ g2 )(x)| > λ/2}| .
λp
kf kpp kg2 kp1
.
λpZ p
 ∞
−p
.λ |{x : |g2 |(x) > α}| dα
0
Z ∞ p
−p ∗ pq −q
. λ (kgkLq,∞ ) α dx
R
−p (1−q)p
.λ R .

Optimize in R:
λ−s Rs−qs(1+1/s−1/p) = λ−p R(1−q)p
λp−s = R(1−q)(p−s) Rq/p·(p−s)
λ = R1−q+q/p = Rq(1/q+1/p−1) = Rq/r .
So we optimize at R = λr/q .
So

|{x : |f ∗ g|(x) > λ}| . λ−p λr/q·p(1−q)

. λ−p(1−r/q+r)
. λ−pr(1/r−1/q+1)
. λ−pr/p
. λ−r .
1
Although we have just proven this claim, here is Hedberg’s proof of kf ∗ |x|α kr . kf kp
whenever 0 < α < d and 1 + 1r = p1 + αd .

Proof. Fix x ∈ Rd . Then

  Z Z Z
1 f (y) f (y) f (y)
f ∗ α (x) = dy = dy + dy.
|x| |x − y|α |x−y|≤R |x − y|α
|x−y|>R |x − y|α

Z XZ
f (y) f (y)
α
dy ≤ dy
|x−y|≤R |x − y| Z R≤|x−y|≤2r
|x − y|α
r∈2
r≤R
X 1
. r−α rd int |f (y)| dy
|B(x, 2r)| B(x,2r)
r≤R

54
 . Rd−α M f (x).

On the other hand,

Z
f (y) 1{|x|>R}
dy = f ∗ (x)
|x−y|>R |x − y|α |x|α
1{|x|>R}
. kf kp
|x|α p0
∞
rd−1
Z
. kf kp αp0
dr
R r
0
d/p −α
. kf kp R
. kf kp Rd(1−1/p−α/d)
. kf kp R−d/r .

Optimize in R: choose
Rd−α M f (x) = kf kp R−d/r
kf kp
Rd/p = Rd(1−α/d+1/r) = .
M f (x)
So
kf kp −p/r
   
1
f ∗ α (x) . kf kp . M f (x)p/r kf k1−p/r
p .
|x| M f (x)
So
1
f∗ . kf k1−p/r
p k(M f )p/r kr
|x|α r
. kf k1−p/r
p kM f kp/r
p
. kf kp .

55
14 The Sobolev Embedding Theorem
14.1 Fourier transforms of tempered distributions
Fix 0 < α < d, and consider Z
2 dt
e−πt|x| t(d−α)/2 .
t
If we let u = π|x|2 t, then this equals
Z ∞  (d−α)/2 Z ∞
u du 1 du
e−u = π (d−α)/2 d−α e−u u(d−α)/2
0 π|x|2 u |x| u
 0 
d−α 1
= π −(d−α)/2 Γ .
2 |x|d−α

We regard π −(d−α)/2 Γ d−α 1

2 |x|d−α
as a tempered distribution, that is an element of
S 0 (Rd ).
These are linear functionals on S(Rd ).
For T ∈ S 0 (Rd )givenby a density ϕ,,
Z
T (f ) = f (x)ϕ(x) dx.

In our case,
     Z
−(d−α)/2 d−α 1 −(d−α)/2 d−α f (x)
π Γ d−α
(f ) = π Γ dx.
2 |x| 2 |x|d−α
Since f is a Schwarz function, this integrand has the right decay at ∞. We have
Z Z Z
f (x) f (x) f (x)
dx ≤ dx + dx
|x|d−α |x|≤R |x|d−α |x|>R |x|d−α
Z
1
. kf k∞ dx + k|x|d f kL∞ R−(d−α) .
|x|d−α

Definition 14.1. For T ∈ S 0 (Rd ), we define its Fourier transform by

Tb(f ) = T (fb), f ∈ S(Rd ).

Let’s compute the Fourier transform of π −(d−α)/2 Γ d−α 1

2 |x|d−α
:
   ∧  Z b
−(d−α)/2 d−α 1 −(d−α)/2 d−α f (x)
π Γ (f ) = π Γ dx
2 |x|d−α 2 |x|d−α
  Z Z −2πix·ξ
−(d−α)/2 d−α e
=π Γ f (ξ) dx dξ
2 |x|d−α

56
 Z ∞ ZZ
2 dt
= e−πt|x| t(d−α)/2 e−2πix·ξ f (ξ) dx dξ
0 t
We already know the Fourier transform of a Gaussian.
Z ∞Z
2 dt
= π d/2 (πt)−d/2 e−π/t·|ξ| t(d−α)/2 f (ξ) dξ
t
Z0 ∞ Z
2 dt
= e−π/t·|ξ| t−α/2 f (ξ) dξ
0 t
Make the change of variables u = π|ξ|2 /t.
Z Z ∞  α/2
−udu u
= f (ξ)e dξ
0 u π|ξ|2
Z Z ∞
−α/2 1 du
= π e−u uα/2 f (ξ) dξ
|ξ|α 0 u
 
1
= π −α/2 Γ(α/2) α (f )
|ξ|
Remark 14.1. Take d = 3 and α = 2:
1 ∧
 
−1/2 1
π Γ(1/2) = π −1 Γ(1) 2 .
|x| |ξ|
That is,  ∧
1 1
= .
2π|x| 4π 2 |ξ|2
This allows us to solve Poisson’s equation: −∆u = f . If we take the Fourier transform,
this is
4π 2 |ξ|2 u
b(ξ) = fb(ξ),
so
1
u
b(ξ) = fb(ξ).
4π 2 |ξ|2
Taking the inverse Fourier transform, we get
1
u= ∗ f.
4π|x|
To make everything rigorous, use
• If T ∈ S 0 (Rd ) and f ∈ S(Rd ), then T ∗ f ∈ S 0 (Rd ) is given by (T ∗ f )(g) = T (fR ∗ g),
where fR (x) = f (−x):
Z
(T ∗ f )(g) = (ϕ ∗ f )(x)g(x) dx

57
 ZZ
= ϕ(y)f (x − y)g(x) dx dy

= T (fR ∗ g).

• If T ∈ S 0 (Rd ) and f ∈ S(Rd ), then T[

∗ f = Tbfb.

14.2 Sobolev embedding

Definition 14.2. Fix s > −d and f ∈ S(Rd ). Then |∇|s f ∈ S 0 (Rd ) is defined by its action
on the Fourier side:
(|∇|s f )∧ (ξ) = (2π|ξ|)s fb(ξ).

Theorem 14.1 (Sobolev embedding). For f ∈ S(Rd ) and 0 < s < d, we have

kf kq . k|∇|s f kp
1 1
whenever p = q + ds . The implicit constant is independent of f .

What does this say? It says that if s derivatives of f live in Lp , then the function must
be more regular/smooth (it lives is a higher Lp space).

Proof. By duality, kf kq = supkgk 0 =1

hf, gi. The idea is that by Plancherel,
Lq

hf, gi = hfb, gbi = h(2π|ξ|)s fb, (2π|ξ|−s gb(ξ)i,

where the first argument is in S 0 .

0
We claim that F = {g ∈ S 0 (Rd ) : gb vanishes on a neighborhood of 0} is dense in Lq .
0
It suffices to show that F is dense in S(Rd ) in the topology of Lq . Fix g0 ∈ S(Rd ). Fix
ε > 0 and ϕ ∈ Cc∞ (B(0, 2)) with ϕ ≡ 1 on B(0, 1). Define g[ [
ε (ξ) = g0 (ξ)(1 − ϕ(ξ/ε)) ∈ S.
d ∨
Then gε ∈ F. Then gb0 − gbε = gb0 ϕ(·/ε), so g0 − gε = g0 ∗ ε ϕ (ε ·).

kg0 − gε kq0 . kg0 k1 kεd ϕ∨ (ε ·)kq0

0
. kg0 k1 εd−d/q
ε→0
−−−→ 0.

Then

kf kq = sup hf, gi
g∈F :kgkq0 =1

= sup h(2π|ξ)s fb, (2π|ξ|)−s gbi

g∈F :kgkq0 =1 | {z } | {z }
∈S 0 ∈S

58
 = sup h|∇|s f , |∇−s gi
g∈F :kgkq0 =1| {z } | {z }
∈S 0 ∈S
. sup k|∇| f kp · k|∇|−s gkp0
s
g∈F :kgkq0 =1

We have
1
g ]∨ ∼
|∇|−s g = [(2π|ξ|)−sb ∗ g,
|x|d−s
so
1
k|∇|−s gkp0 ∼ . kgkq0
|x|d−s ∗g p0
1 1 d−s
by Hardy-Littlewood-Sobolev, provided 1 + p0 = q0 + d . We can rewrite this condition
as p1 = 1q = ds .

59
15 Riesz Tranforms and Calderón-Zygmund Convolution Ker-
nels
15.1 Riesz tranforms
Last time, we proved the Sobolev embedding theorem:
Theorem 15.1 (Sobolev embedding). For f ∈ S(Rd ) and 0 < s < d, we have
kf kq . k|∇|s f kp
1 1
whenever p = q + ds . The implicit constant is independent of f .
In particular,
1 1 1
kf kq . k|∇|f kp ,
= + .
p q d
However, the Fourier transform is not a local operator; it looks at the whole function.
However, we can ask whether it is true that
1 1 1
kf kq . k∇f kp = +
p q d
with 1 < p < q < ∞. This would follow from boundedness of the Riesz transforms on Lp
for 1 < p < ∞.
Definition 15.1. For 1 ≤ j ≤ d and f ∈ S(Rd ), we define the Riesz transforms as
iξj b
Rj f (ξ) = mj (ξ)f (ξ) = − f (ξ).
d b
|ξ|
j ∂
In other words, Rj = − |∇| .
We write
d
X
2π|ξ| = mj (ξ) · 2πiξj .
j=1

So
d
X
|∇| = Rj ∂j .
j=1

Then
d
X d
X
kf kq . k|∇|kp ≤ kRj ∂j f k . k∂j f kp . k∇f kp ,
j=1 j=1

if we knew the Riesz transforms were bounded on Lp . (The last step comes from the fact
that all finite dimensional vector space norms are equivalent.)

60
Remark 15.1. If we knew that the Riesz transforms are bounded on Lp for 1 < p < ∞,
we could also conclude that the solution u to the Poisson equation −∆u = f satisfies
∂j ∂k u ∈ Lp whenever f ∈ Lp . Indeed,
ξj ξk b
(∂j ∂k u)∧ (ξ) = −4π 2 ξj ξk u
b(ξ) = − f (ξ) = mj (ξ)mk (ξ)fb(ξ).
|ξ|2
So ∂j ∂k u = Rj Rk f .
How do we prove boundedness of Riesz transforms?
Definition 15.2. A function K : Rd \ {0} → C is a Calderón-Zygmund convolution
kernel if it satisfies:
1. |K(x)| . |x|−d uniformly for |x| > 0.
R
2. R1 ≤|x|≤R2 L(x) dx = 0 for all 0 < R1 < R2 < ∞ (cancellation condition).

3. |x|>2|y| |K(x + y) − K(x)| dx . 1 uniformly for y ∈ Rd (regularity condition).

R

Example 15.1. The Riesz tranforms correspond to Calderón-Zygmund convolution ker-

nels.
" #(d−1/2)
iξj 1 π −(d−1)/2 Γ((d − 1)/2) xj
mj = − =⇒ kj (x) = m∨
j (x) = − ∂j −1/2
∼d .
|ξ| 2π π Γ(1/2) |x|d+1

We have
1. |kj (x)| . x−d uniformly in |x| > 0.
R
2. R1 ≤|x|≤R2 kj (x) dx = 0 for all 0 < R1 < R2 < ∞ because it is odd in xj .
3. By the fundamental theorem of calculus,
Z Z Z 1
|kj (x + y) − kj (x)| dx ≤ |y| |∇kj (x + θy)| dθ dx
|x|≥2|y| |x|>2|y| 0

For |x| > 2y| and θ ∈ (0, 1), |x|/2 ≤ |x| − |y| ≤ |x + θy| ≤ |x| + |y| ≤ 3|x|/2.
Z
1
. |y| d+1 dx
|x|>2|y| |x|
1
. |y| · . 1.
|y|

More generally, we have proved the following.

Lemma 15.1. If |∇K(x)| . |x|−(d+1) uniformly for |x| > 0, then K satisfies the regularity
condition.

61
15.2 L2 -boundedness of convolution with Calderón-Zygmund convolu-
tion kernels
Here is a lemma we need.
Lemma 15.2. Let K : Rd \ {0} → C be a Calderón-Zygmund convolution kernel. For
ε > 0, let Kε = K 1{ε≤|x|≤1/ε} . Then Kε is a Calderón-Zygmund convolution kernel.

Proof. |kε (x)| . |k(x)| . |x|−d uniformly for |x| > 0. For the second condition,
Z Z
Kε (x) dx = K(x) dx = 0, ∀0 < R1 < R2 < ∞.
R1 ≤|x|≤R2 max{R1 ,ε}≤|x|≤min{R2 ,1/ε}

For the third condition,

Z Z
|Kε (x + y) − Kε (y)| dx ≤ ε≤|x|≤1/ε |Kε (x + y) − Kε (y)| dx
|x|>2|y| ε≤|x+y|≤1/ε
|x|>2|y|
Z
+ ε≤|x|≤1/ε |Kε (x)| dx
ε>|x+y| or |x+y|>1/ε
|x|>2|y|
Z
+ ε>|x| or x>1/ε |Kε (x + y)| dx.
ε≤|x+y|≤1/ε
|x|>2|y|

Look at I: If |x + y| < ε, then |x| ≤ |x + y| + |y| ≤ |x + y| + |x|/2, so |x| ≤ 2|x + y| ≤ 2ε.

The contribution is at most
Z Z
|K(x)| dx . |x|−d dx . 1,
ε≤|x|≤2ε ε≤|x|≤2ε

uniformly in ε > 0.
If |x + y| > 1/ε, then |x| ≥ |x + y| − |y| ≥ |x + y| − |x|/2, so |x| ≥ 23 |x + y| ≥ 2
3ε . The
contribution is at most Z
|K(X)| dx . 1,
2
3ε
≤|x|≤ 1ε

uniformly in ε > 0. Similarly, II . 1 uniformly in ε > 0 and y ∈ Rd .

Theorem 15.2. Ket K : Rd \ {0} → C be a Calderón-Zygmund convolution kernel. For

ε > 0, let Kε = K 1{ε<|x|≤1/ε} . THen

kKε ∗ f k2 . kf k2

uniformly for ε > 0, f ∈ L2 . Consequently, f 7→ K ∗ f (which is the L2 limit as ε → 0 of

Kε ∗ f ) extends continuously from S(Rd ) to a bounded map on L2 (Rd ).

62
Proof.
kKε ∗ f k2 = kK
\ ε ∗ f k2

= kK
b ε fbk2
≤ kK
b ε k∞ kwhf k2
≤ kK
b ε k∞ kf k2 .

Fix ξ ∈ Rd . Then
Z
K
b ε (ξ) = e−2πix·ξ Kε (x) dx
Z Z
= e−2πix·ξ Kε (x) dx + e−2πix·ξ Kε (x) dx.
|x|≤1/|ξ| |x|>1/|ξ|

Now observe that by condition 2 of the definition of the Calderón-Zygmund convolution

kernel,
Z Z
e−2πix·ξ Kε (x) dx = (e−2πix·ξ − 1)Kε (x) dx
ε≤|x|≤1/|ξ| |x|≤1/|ξ|

By condition 1,
Z
. |x||ξ||x|−d dx
|x|≤1/|ξ|
1
. |ξ|
|ξ|
. 1.
On the other hand, we have
Z Z
1
e−2πix·ξ Kε (x) dx = (1 − eπi )e−2 piix·ξ Kε (x) dx
|x|>1/|ξ| |x|>1/|ξ| 2
Z
1 −2πix·ξ
= e Kε (x) dx
|x|>1/|ξ| 2
Z
1 2
− e−2πiξ(x−ξ/(2|ξ| ) Kε (x) dx
2 |x|>1/|ξ|
Z
1 −2πix·ξ
= e Kε (x) dx
|x|>1/|ξ| 2
 
|xi|
Z
1
− e−2πix·ξ Kε x + dx,
2 x+ ξ 2 > 1 2|ξ|2
2|ξ| |ξ|

which puts us into a position to make a change of variables and use condition 3. We will
finish the proof next time.

63
16 Boundedness of Calderón-Zygmund Convolution Kernels
16.1 L2 -boundedness of convolution with Calderón-Zygmund kernels
Last time, we were proving the following theorem.

Theorem 16.1. Let K : Rd \ {0} → C be a Calderón-Zygmund convolution kernel. For

ε > 0, let Kε = K 1{ε<|x|≤1/ε} . Then

kKε ∗ f k2 . kf k2

uniformly for ε > 0, f ∈ L2 . Consequently, f 7→ K ∗ f (which is the L2 limit as ε → 0 of

Kε ∗ f ) extends continuously from S(Rd ) to a bounded map on L2 (Rd ).

Proof. By Plancherel,
kKε ∗ f kL2 ≤ kK
b ε k∞ kf kL2 ,
b ε k∞ . 1 uniformly in ε > 0. Fix ξ ∈ Rd . Then
so it suffices to show that kK
Z
b ε (ξ) = e−2πix·ξ Kε (x) dx
K
Z Z
= +
|x|≤1/|ξ| |x|>1/|ξ|

Because of property (b) and (a),

Z Z
−2πix·ξ
e Kε (x) dx = [e−2πix·ξ − 1]Kε (x)
ε≤|x|≤1/|ξ| |x|≤1/|ξ||
Z
1
. |x| · |ξ| · d dx . 1.
|x|≤1/|ξ| |x|

We also have
Z Z
−2πix·ξ 1 −2πix·ξ 2
e Kε (x) dx = (e − e−2πiξ·(x−ξ/(2|ξ| )) )Kε (x) dx
|x|>1/|ξ| |x|>1/|ξ| 2
Z
1
= e−2πix·ξ Kε (x) dx
2 |x|>1/|ξ|
Z
1
− e−2πix·ξ Kε (x + ξ/(2kξ|2 )) dx
2 x+ξ/(2|ξ|2 )>1/|ξ|
Z
1
= e−2πix·ξ [Kε (x) − Kε (x + ξ/(2|ξ|2 ))] dx
2 |x|>1/|ξ|
Z
1
+ e−2πix·ξ Kε (x + ξ/(2|ξ|2 )) dx
2

64
 R R R R
We can split x+ξ/(2|ξ|2 )>1/|ξ| = |x|>1/|ξ| − A − B , where A and B are a partition of the
symmetric difference (like a Venn diagram). So A = {x : |x| ≤ 1/|ξ| ≤ |x + ξ/(2|ξ|2 )|} and
B = {x : |x + ξ/(2|ξ|2 )| ≤ 1/|ξ| ≤ |x|}.
Z
= e−2πix·ξ [Kε (x) − Kε (x + ξ/(2|ξ|2 ))] dx
|x|>1/|ξ|
| {z }
I
Z
1
− e−2πix·ξ Kε (x + ξ/(2|ξ|2 )) dx
2 A
| {z }
II
Z
1
− e−2πix·ξ Kε (x + ξ/(2|ξ|2 )) dx .
2 B
| {z }
III

Looking at these terms individually:

Z
1
|I| ≤ |Kε (x) − Kε (x + ξ/(2|ξ|2 ))| dx . 1
2 |x|≥1/|ξ|

uniformly in ξ and ε > 0 by condition (c).

Z
|II| . |Kε (x + ξ/(2|ξ|2 ))| dx
A

Note that A ⊆ x : 1/|ξ| ≤ |x + ξ|/(2|ξ|2 )| ≤ |x| + 1/(2|xi|) ≤ 3/(2|ξ|).

Z
. |Kε (y)| dy
1/|ξ|≤|y|≤3/(2|ξ|)
Z
. .
1/|ξ|≤|y|≤3/(2|ξ|)

Z
|III| . |Kε (x + ξ/(2|ξ|2 ))| dx
B

Note that B ⊆ x : 1/(2|ξ|) ≤ |x| − 1/(2|xi|) ≤ |x + ξ|/(2|ξ|2 )| ≤ 1/|ξ|.

Z
. |Kε (y)| dy . 1.
1/(2|ξ|)≤|y|≤1/|ξ|

So kK
b ε k∞ . 1, uniformly in ε > 0.

65
 We claim that for f ∈ S(R), {Kε ∗ f }ε is Cauchy ni L2 . Assuming the claim, for
f ∈ S(Rd ), let K ∗ f be the L2 limit of Kε ∗ f . Then

kK ∗ f k2 ≤ kKε ∗ f k2 + kK ∗ f − Kε ∗ f k2 .
| {z } | {z }
.kf k2 ε→0
−−−→0
So we have that
kK ∗ f k2 . kf k2 + o(1)
L2
as ε → 0. Let ε → 0 to get kK ∗ f k2 . kf k2 . For f ∈ L2 , let fn ∈ S be such that fn −→ f .
Then {fn }n is Cauchy in L2 , so {K ∗ fn }n≥1 is Cauchy in L2 . Let K ∗ f be the L2 -limit of
K ∗ fn . Now
kK ∗ f k2 = lim kK ∗ f k2 . lim kfn k2 = kf k2 .
n n

Now let’s prove the claim: Fix f ∈ S(Rd ) and 0 < ε1 < ε2 < 1. Then
Z Z
(Kε1 ∗ f − Kε2 ∗ f )(x) = K(y)f (x − y) dy − K(y)f (x − y) dy
ε1 ≤|y|≤1/ε1 ε2 ≤|y|≤1/ε2
Z Z
= K(y)f (x − y) dy + K(y)f (x − y) dy
ε1 ≤|y|≤ε2 1/ε2 ≤|y|≤1/ε1

Using property (b),

Z Z
K(y)f (x − y) dy = K(y)[f (x) − f (y)]) dy
ε1 ≤|y|≤ε2 ε1 ≤|y|≤ε2
Z Z 1
≤ |K(y)||y| |∇f (x − θy)| dθ dy
ε1 ≤|y|≤ε2 0

Using property (a),

Z Z
. |y|1−d |∇f (x − θy)| dθ dy
ε1 ≤|y|≤ε2 | {z }
.1/hx−θyid .1/hxid
1
. (ε2 − ε1 ) .
hxid

Alternatively, we could say

Z Z Z Z
1−d 1−d
|y| |∇f (x − θy)| dθ dy . |y| k∇f (x − θy)kL2x dθ dy
ε1 ≤|y|≤ε2 ε1 ≤|y|≤ε2
L2x
. k∇f kL2 (ε2 − ε1 )

66
 ε2 ,ε1 →0
−−−−−→ 0.

For the other term, using Young’s inequality, we have

Z Z
K(y)f (x − y) dy . kK 1{1/ε2 ≤|y|≤1/ε1 } k2 · kf k1
1/ε2 ≤|y|≤1/ε1
L2x
Z !1/2
−2d
. kf kL1 |y| dy
|y|≥1/ε2
d/2
. kf k1 ε2
ε →0
2
−− −→ 0.

Remark 16.1. The same argument show that for f ∈ S(Rd ), {Kε ∗ f }ε>0 is Cauchy in Lp
for 1 < p < ∞. It uses conditions (a), (b).

16.2 Lp bounds for Calderón-Zygmund convolution kernels

Theorem 16.2. Let K : Rd \ {0} → C be a Calderón-Zygmund convolution kernel. For
ε > 0, let Kε = K 1{ε≤|x|≤1/ε} . Then

1. |{x : |Kε ∗ f |(x) > λ}| . λ1 kf k1 uniformly in λ > 0, f ∈ L1 , ε > 0.

2. For any 1 < p < ∞, kKε ∗ f kp . kf kp uniformly for f ∈ Lp , ε > 0.

Consequently, f 7→ K ∗ f (the Lp -limit of Kε ∗ f ) extends continuous ly from S(Rd ) to a

bounded map on Lp when 1 < p < ∞.

Proof. First, assume that we have proven the first claim. By the Marcinkiewicz inter-
polation theorem, we get the second claim for 1 < p < 2. Now fix 2 < p < ∞. By
duality,

kKε ∗ f kp = sup hKε ∗ f, gi

kgkp0 =1

= sup hf, KεR ∗ gi

kgkp0 =1

. kf kp sup kKεR ∗ gkp0

kgkp0 =1

. kf kp .

We will prove the first claim last time.

67
17 Lp Bounds for Calderón-Zygmund Convolution Kernels
17.1 Weak Lp bound for Calderón-Zygmund convolution kernels
Theorem 17.1. Let K : Rd \ {0} → C be a Calderón-Zygmund convolution kernel. For
ε > 0, let Kε = K 1{ε≤|x|≤1/ε} . Then
1. |{x : |Kε ∗ f |(x) > λ}| . λ1 kf k1 uniformly in λ > 0, f ∈ L1 , ε > 0.
2. For any 1 < p < ∞, kKε ∗ f kp . kf kp uniformly for f ∈ Lp , ε > 0.
Consequently, f 7→ K ∗ f (the Lp -limit of Kε ∗ f ) extends continuoussly from S(Rd ) to a
bounded map on Lp when 1 < p < ∞.
Proof. Assuming that (1) holds, we proved (2) using interpolation and duality. To show the
last claim, it suffices to prove that {Kε ∗ f }ε>0 forms a Cauchy sequence in Lp (1 < p < ∞)
whenever f ∈ S(Rd ). We want to prove this using the L2 result and condition (c) of the
Calderón-Zygmund kernel; this will let our theory have more adaptability.
For 1 < p < 2, let 1 < q < p. Write p1 = θq + 1−θ 2 for some θ ∈ (0, 1). Then

kKε1 ∗ f − Kε2 ∗ f kp . kKε1 ∗ f + Kε2 f k21−θ kKε1 ∗ f + Kε2 f kθq

| {z } | {z }
ε1 ,ε2 →0 ≤(kKε1 ∗f kq +kKε2 ∗f kq )θ .kf kθq
−−−−−→0
ε1 ,ε2 →0
−−−−−→ 0.
1 1−θ
For 2 < p < ∞¡ let p < r < ∞ and write p = 2 + θr . Then

kKε1 ∗ f − Kε2 ∗ f kp ≤ kKε1 ∗ f − Kε2 ∗ f k21−θ kKε1 ∗ f − Kε2 ∗ f kθr

| {z }| {z }
ε1 ,ε2 →0 .kf kθr
−−−−−→0
Let’s show (1). For λ > 0, f ∈ L1 , and ε > 0, perform a Calderón-Zygmund de-
composition for f at level λ: f = g + b with supp b = Qk , Qok pairwise disjoint, and
S
P
k |Qk | ≤ kf k1 /λ. We can take
( S
f (x) x∈/ Qk
g(x) = 1
R o
|Qk | Qk f (y) dy x ∈ Qk .

Then |g| . λ, and b(x) = f (x) − |Q1k | Qk f (y) dy for x ∈ Qk , so

R

Z Z
1
b(x) dx = 0, |b(y)| . λ.
Qk |Qk | Qk
Then

|{x : |Kε ∗ f |(x) > λ}| ≤ |{x : |Kε ∗ g|(x) > λ/2}| + |{x : |Kε ∗ b|(x) > λ/2}|

68
 1 2
[ h[ ic
. kK ε ∗ gk2 + αQk + {x ∈ αQk : |Kε ∗ b|(x) > λ/2}
λ2
k

We have
1 2 kgk22 λkgk1 kf k1
2
kKε ∗ gk 2 . 2
. 2
.
λ λ λ λ
and
[ X X kf k1
αQk ≤ |αQk | ≤ αd |Qk | . αd .
λ
We are left with E := |{x ∈ [ αQk ]c : |Kε ∗ b|(x) > λ/2}|. Let x ∈
S S
/ αQk . Then
Z
Kε ∗ b(x) = Kε (x − y)b(y) dy
XZ
= Kε (x − y)b(y) dy
k Qk

Here, we only have a convolution, not an average. But a convolution is only as smooth as
its smoothest term. So we have to use the regularity of Kε (condition (c)).
XZ
= [Kε (x − y) − Kε (x − xk )]b(y) dy.
k Qk

Using Chebyshev,
Z
1
E. (Kε ∗ b)(x)
λ x∈/ S αQk
Z Z
1X
. |Kε (x − y) − Kε (x − xk )||b(y)| dy dx
λ x∈(αQk )c Qk
k

Change variables.
Z Z !
1X
. |b(y)| |Kε (x + xk − y) − Kε (x)| dx dy.
λ Qk (αQk )c −{xk }
k
√ √ √
For y ∈ Qk , |xk − y| ≤ 12 `(Qk ) d. So we need α`(Qk )/2 ≥ 2 12 `(Qk ) d. So take α ≥ 2 d.
Then using the regularity condition (c) of the convolution kernel, we get
Z
1X
E. |b(y)| · 1 dy
λ Qk k
kf k1
. .
λ
Remark 17.1. Once we have boundedness in L2 , the only condition we need to deduce
boundedness in Lp for 1 < p < ∞ is the regularity condition (c).

69
17.2 Application: The Hilbert transform
Here is an application.
1
Example 17.1. Let K : R \ {0} → R be K(x) = πx . This is a Calderón-Zygmund
convolution kernel. So the Hilbert transform,

f (x − y) f (x − y)
Z Z
1
Hf (x) = dy = lim dy.,
π y ε→0 |y|>ε y

is bounded on Lp for 1 < p < ∞.

Remark 17.2. Boundedness on L1 and L∞ may fail. Consider the Hilbert transofrm, and
take f = 1[a,b] ∈ L1 ∩ L∞ ; we will show that Hf ∈
/ L1 ∪ L∞ . For ε > 0,

1
Z
1[a,b] (x − y)
Hε f (x) := dy
π ε≤|y|≤1/ε y
Z
1 1
= dy
π ε≤|y|≤1/ε y
x−b≤y≤x−a
1 x−a
= log
π x−b

/ L1 ∪ L∞ .
almost everywhere. But Hf ∈
d (ξ) = −i sgn(ξ) · fb(ξ). For a > 0, let
Remark 17.3. We have Hf
( (
e−aξ ξ ≥ 0 0 ξ>0
fa (ξ) =
b gba (ξ) = qξ
0 ξ < 0, e ξ ≤ 0.

Then  
−aξ
e
 ξ>0
S 0 (R), a→0
1
 ξ>0
(fba − gba )(ξ) = 0 ξ = 0 −−−−−−−→ 0 ξ = 0 = sgn(ξ).

 aξ 
−e ξ<0 −1 ξ < 0


So we get
S 0 (R), a→0
fa − ga −−−−−−−→ sgn∨ .
Next time, we will complete this computation.

70
18 The Mikhlin Multiplier Theorem
18.1 The Hilbert transform
Recall the Hilbert transform
f (x − y)
Z
Hf (x) = P V dy.
πy

d (ξ) = −i sgn(ξ)fb(ξ). Let

We claimed that Hf
( (
e−aξ ξ > 0 0 ξ>0
fba (ξ) = gba (ξ) =
0 ξ ≤ 0, eaξ ξ≤0

Then  
−aξ
e
 ξ>0
S 0 (R), a→0
1
 ξ>0
(fba − gba )(ξ) = 0 ξ = 0 −−−−−−−→ 0 ξ = 0 = sgn(ξ).

 aξ 
−e ξ<0 −1 ξ<0


So we get
S 0 (R), a→0
fa (x) − ga (x) −−−−−−−→ sgn∨ .
Now compute
Z ∞
1
fa (x) = e2πixξ e−aξ dξ = ,
0 a − 2πix
Z 0
1
ga (x) = e2πixξ eaξ dξ = ,
−∞ a + 2πix
so
4πix
fa (x) − ga (x) = .
a2 + 4π 2 x2
Let ϕ ∈ S(Rd ), and compute
Z
4πix
lim (fa − ga )(ϕ) = lim ϕ(x) dx
a→0 a→0 a + 4π 2 x2
2

We can’t pull in the limit as is. We need the integrand to vanish near 0.
Z
4πix
= lim [ϕ(x) − ϕ(0)1[−ε,ε] (x)] dx
a→0 a2 + 4π 2 x2
Z ε Z
4πix 4πix
= lim [ϕ(x) − ϕ(0)] + lim ϕ(x) dx
a→0 −ε a2 + 4π 2 x2 a→0 |x|>ε a2 + 4π 2 x2

71
 Z ε Z
i i
= [ϕ(x) − ϕ(0)] + ϕ(x) dx .
−ε πx |x|>ε πx
| {z }
ε→0
−−−→i PV( πx )(ϕ)
1

On the other hand,

Z ε
i ε→0
(ϕ(x) − ϕ(0)) dx . εkϕkL∞ −−−→ 0.
−ε πx

So  
S 0 (R),a→0 1
fa − ga −−−−−−→ i PV (ϕ) = iH,
πx
where iH
b = sgn.

18.2 Littlewood-Paley projections and the Mikhlin multiplier theorem

Let’s construct a dyadic partition of unity. Let ϕ : Rd → [0, 1], ϕ ∈ Cc∞ with
(
1 |x| ≤ 1.4
ϕ(x) =
0 |x| > 1.42.

Let ψ(x) = ϕ(x) − ϕ(2x); if we graph ψ, it is 0 before 0.7, increases quickly to 1 between
0.7 and 0.71, plateaus on 0.71 to 1.4, and goes down to 0 by 1.42.
For N ∈ 2Z , let ψN 9x) = ψ(x/N ). Note that
X
ψN (x) = 1
N ∈2Z

a.e. (in fact for all x 6= 0.

Definition 18.1. The Littlewood-Paley projection to frequencies |ξ| ∼ N is given by

Pd
N f (ξ) = f (ξ)ψN (ξ),
b i.e. PN f = f ∗ [N d ψ ∨ (N ·)].

We also define
\
P ≤N f (ξ) = f (ξ)ϕ(ξ/n),
b i.e. P≤N f = [N d ϕ∨ (N ·)] ∗ f

Remark 18.1. Caution: PN is not a true projection since PN2 = PN .

We can also define

X
P>n = Id −P≤N , PM ≤·≤N = PK .
M ≤K≤N

72
Theorem 18.1 (Mikhlin multiplier theorem). Let m : Rd \{0} → C be such that |Dξα m(ξ)| .
|ξ|−|α| uniformly for |ξ| =
6 0 and 0 ≤ |α| ≤ d d+1
2 e. Then

f 7→ [m(ξ)fb(ξ)]∨ = m∨ ∗ f

is bounded on Lp for all 1 < p < ∞.

Proof. Taking α = 0, we get M ∈ L∞ . By Plancherel,

km∨ ∗ f k2 = kmfbk2 ≤ kmkL∞ kfbk2 . kf k2 .

It suffices to check the regularity condition (c) is satisfied by the kernel m∨ . We’ll first
prove this assuming |Dξα m(ξ)| . |ξ|−|α| for 0 ≤ |α| ≤ d + 2. In this case, we will show that
|∇m∨ (x)| . |x|−(d+1) uniformly for |x| =
6 0. This yields (c).
We have
k|xα |∇m∨ (x)kL∞x
. Dξα [ξm(ξ)] L1 ,
ξ
| {z }
O(|ξ|1−|α| )

But this is not integrable! However, we can integrate it on dyadic annuli. Write
X
m(ξ) = mN (ξ), mN (ξ) = m(ξ)ψN (ξ).
N ∈2Z

Then the chain rule gives

X
Dξα [ξmN (ξ)] = Dξα1 [ξm(ξ)]Dξα2 [ψN (ξ)],
α1 +α2 =α

so
X
|Dξα [ξmN (ξ)]| .α |ξ|1−|α1 | N −|α2 | |Dξα2 ψ|(ξ/N ).
α1 +α2 =α

Then

k|xα |∇m∨
N (x)kL∞
x
. kDξα [ξmn (ξ)]kL1
ξ
X Z
.α |ξ|1−|α1 | N −|α2 | dξ
α1 +α2 =α |ξ|∼N

.α N 1−|α|+d .

So we get
|∇m∨
N (x)| . min{N
d+1
, (N |x|d+2 )−1 }.

73
By the triangle inequality,
X
|∇m∨ (x)| ≤ |∇m∨
N (x)|
N ∈2Z
X X 1
. N d+1 +
N |x|d+2
N ≤|x|−1 N >|x|−1

. |x|−(d+1) ,

uniform in |x| =
6 0.
Now let’s prove condition (c) assuming this hypothesis holds for only 0 ≤ |α| ≤ d d+1
2 e.
Look at
Z X Z
|m∨ (x + y) − m∨ (x)| dx = |m∨ ∨
N (x + y) − mN (x)| dx
|x|≥2|y| |x|≥2|y|
N ∈2Z

If we have fbN (ξ) = fb(ξ)ψ( ξ), then fN (x) = (f ∗ N d ψ(N ·))(x) = f (x − y)N d ψ ∨ (N y) dy,
R

so |fN (x)| . |f (x − y)|N d hN 1yim dy.

R

X Z
. |m∨ ∨
N (x + y) − mN (x)| dx
|x|≥2|y|
N ≤|y|−1
X Z
+2 |Mn∨ (x)| dx
|x|≥|y|
N >|y|−1

Using the fundamental theorem of calculus,

X Z Z 1
. |y| · |∇m∨
N (x + θy)| dθ dx
|x|≥2|y| 0
N ≤|y|−1
X Z
+2 |m∨
N (x)| dx.
|x|≥|y|
N >|y|−1

We will complete the proof next time.

74
19 The Mikhlin Multiplier Theorem and Properties of Littlewood-
Paley Projections
19.1 The Mikhlin multiplier theorem
Theorem 19.1 (Mikhlin multiplier theorem). Let m : Rd \{0} → C be such that |Dξα m(ξ)| .
|ξ|−|α| uniformly for |ξ| =
6 0 and 0 ≤ |α| ≤ d d+1
2 e. Then

f 7→ [m(ξ)fb(ξ)]∨ = m∨ ∗ f

is bounded on Lp for all 1 < p < ∞.

Proof. By Plancherel and m ∈ L∞ , we get boundedness on L2 . So it suffices to check

regularity condition (c):
Z
|m∨ (x + y) − m∨ (x)| dx . 1
|x|≥2|y|

uniformly in y. We have
Z X Z
∨ ∨
|m (x + y) − m (x)| dx . |m∨ ∨
N (x + y) − mN (x)| dx
|x|≥2|y| |x|≥2|y|
N ∈2Z

where MN = mψN = mψ( · /N ).

X Z
≤ |m∨ ∨
N (x − y) − mN (x)| dx
|x|>2|y|
N ≤|y|−1
X Z
+2 |m∨
N (x)| dx
|x|≥|y|
N >|y|−1
X Z Z 1
≤ |y| |∇m∨
N (x + θy)| dθ dx
|x|≥2|y| 0
N ≤|y|−1
X Z
+2 |m∨
N (x)| dx.
|x|≥|y|
N >|y|−1

Last time, we had pointwise bound on derivatives by assuming more conditions for more
values of α. Here, instead, we will use Plancherel. By Plancherel,

k(2πix)α m∨ α
N (x)kL2x = kDξ mN kL2ξ
X 1
= cα1 ,α2 kDξα1 m(ξ) · (Dξα2 (ξ/N )k2
α1 +α2 =α
N |α2 |

75
 X Z 1/2
−2|α1 | −2|α2 |
.α |ξ| N dξ
α1 +α2

. N d/2−|α|

for all 0 ≤ α ≤ d d+1

2 e. By Cauchy-Schwarz,
Z
|m∨ ∨
N (x)| dx ≤ kmN k2 A
d/2
. (AN )d/2 .
|x|≤A

Similarly,
Z Z !1/2
|m∨
N (x)| dx ≤ kx α
m∨
N k2 |x| −2|α|
dx
|x|>A |x|>A

. N d/2−|α| Ad/2−|α| ,

provided |α| > d/2. So for α = d d+1

2 e > d/2, we get
Z
|m∨ N (x)| dx . (N A)
d/2−d(d+1)/2e
.
|x|>A

Then
X Z X
|m∨
N (x)| dx . (N |y|)d/2−d(d+1)/2e
|x|≥|y|
|x|>|y|−1 N >|y|−1

This is a geometric series, so it is smaller than a constant times its largest term.

. 1,

uniformly in y ∈ Rd . Taking A = N −1 in our relations, we get

Z
|m∨N (x)| . 1,

uniformly in N .
The same arguments would give
Z
|∇m∨
N (x)| dx . N,

uniformly in N . Indeed,

k(2πix)α ∇m∨ α
N k2 = kD (ξmN )k2

76
 Z !1/2
X
.α |ξ|2−2|α1 | N −2|α2 | dξ
α1 +α2 =α |ξ|∼N

. N 1+d/2−|α| ,

so we get Z
|∇|m∨
N| . N
1+d/2 d/2
A = n(N A)d/2 ,
|x|≤A
Z
|∇m∨
N | . N (N A)
d/2−d(d+1)/2e
.
|x|≥A

We can now estimate

X Z Z 1 X
|y| |∇m∨
N (x + θy)| dθ dx . |y| · N . 1,
|x|≥2|y| 0
N ≤|y|−1 N ≤|y|−1

uniformly in y.

19.2 Properties of Littlewood-Paley projections

Recall the Littlewood-Paley projections:
(
1 |x| ≤ 1.4
ϕ(x) = ψ(x) = ϕ(x) − ϕ(2x).
0 |x| > 1.42,

Then we had
fN = PN f = f ∗ N d ψ ∨ (N ·),
f≤N = P≤N f = f ∗ N d ϕ∨ (N ·).
Here are the basic properties of Littlewood-Paley projections.

Theorem 19.2.

1. kfn kp + kf≤N kp . kf kp uniformly in N and for 1 ≤ p ≤ ∞.

2. |fN (x)| + |f≤N (x)| . M f (x).

Lp
3. For f ∈ Lp with 1 < p < ∞, we have f =
P
N ∈2Z fN .

4. (Bernstein’s inequality) For 1 ≤ p ≤ q ≤ ∞,

kfN kq . N d/p−d/q kfN kp

kf≤N kq . N d/p−d/q kf≤N kp .

77
 5. (Bernstein) For 1 ≤ p ≤ ∞ and s ∈ R,

k|∇|s fN kp ∼ N s kfN kp .

In particular, for s > 0 and 1 ≤ p ≤ ∞,

k|∇|s f≤N kp . N s kf≤N kp .

k∇|−s f>N kp . N −s kf>N kp .

Proof.

1. By Young’s inequality,

kfN kp = kf ∗ N d ψ ∨ (N ·)kp
. kf kp kN d ψ ∨ (N ·)k1 = kψ ∨ k1
| {z }
. kf kp ,

kf≤N kp = kf ∗ N d ϕ∨ (N ·)kp
. kf kp kϕ∨
1 k1
. kf kp .

2.
Z
|fN (x)| ≤ |f (y)N d ψ ∨ (N (x − y))| dy
Z
1
. N d |f (y)| dy
hN (x − y)i2d
|f (y)|
Z X Z
d d
.N |f (y)| dy + N dy
|x−y|≤1/N R/N ≤|x−y|≤2R/N R2d
R∈2Z

How do we make this look like the maximal function?

Z
1
. |f (y)| dy
|B(x, 1/N )| B(x,1/N )
Z
X
−d 1
+ R |f (y)| dy
|B(x, 2R/N )| B(x,2R/N )
R∈2Z
 
X
. M f (x) 1 + R−d 
R∈2Z

. M f (x).

78
 3. First assume f ∈ S(Rd ). By Plancherel and dominated convergence,
N →0
kf − PN ≤·≤1/N f k2 −−−→ 0.
1 1−θ 1+θ
For 1 < p < 2, write p =θ+ 2 = 2 .

kf − PN ≤·≤1/N f kp ≤ kf − PN ≤·≤1/N f kθ1 kf − PN ≤·≤1/N f k1−θ

2
≤ (kf k1 + kPN ≤·≤1/N f k1 )θ · kf − PN ≤·≤1/N f k1−θ
2
N →0
−−−→ 0

by property (1). For 2 < p < ∞,

2/p
kf − PN ≤·≤1/N f kp ≤ k k2 k k1−2/p
∞
| {z } | {z }
N →0 .kf k∞
−−−→0

If f ∈ Lp , let g ∈ S(Rd ) such that kf − gkp ≤ δ. Then

kf − PN ≤·≤1/N f kp . kg − PN ≤·≤1/N gkp + kf − gk|p + kPN ≤·≤1/N (f − g)kp

. o(1) + δ

as N → 0.

We will prove (4) and (5) next time.

R
Remark 19.1. (3) fails for p = 1 and p = ∞. For p = 1, PN f = Pd
n f (0) = 0, so pick
some function with mean 0.

79
20 Littlewood-Paley Projections and Khinchine’s Inequality
20.1 Bernstein properties of Littlewood-Paley projections
Last time, we were proving properties of Littlewood-Paley projections.
Theorem 20.1.

1. kfn kp + kf≤N kp . kf kp uniformly in N and for 1 ≤ p ≤ ∞.

2. |fN (x)| + |f≤N (x)| . M f (x).

Lp
3. For f ∈ Lp with 1 < p < ∞, we have f =
P
N ∈2Z fN .

4. (Bernstein’s inequality) For 1 ≤ p ≤ q ≤ ∞,

kfN kq . N d/p−d/q kfN kp

kf≤N kq . N d/p−d/q kf≤N kp .

5. (Bernstein) For 1 ≤ p ≤ ∞ and s ∈ R,

k|∇|s fN kp ∼ N s kfN kp .

In particular, for s > 0 and 1 ≤ p ≤ ∞,

k|∇|s f≤N kp . N s kf≤N kp .

kf>N kp . N −s k|∇|s f>N kp .

We proved properties (1) to (3) last time.

Proof. Here is 4: We have fN = f ∗ N d ψ ∨ (N ·), so by Young’s inequality,

kfn kq . kf kp · kN d ψ ∨ (N ·)kqp/(qp+p−q)
. kf kN d−d(1+1/q−1/p)
. N d/p−d/q kf kp

To recover fN on the RHS, we use a common trick. Let ψ(ξ) e = ψ(2ξ) + ψ(ξ) + ψ(ξ/2),
ψN (ξ) = ψ(ξ/N ), and define the fattened LIttlewood-Paley projection
e e

PeN f (ξ) = fb(ξ) · ψeN (ξ).

d

Note that PeN Pn = Pn since ψe ≡ 1 on supp ψ. Write

fN = PeN f = fn ∗ [N d ψe∨ (N ·)]

80
and argue as before. The same argument gives kf≤N kq ≤ N d/p−d/q kf≤N kp . (We use
P≤4N P≤N = P≤N .)
Here is 5: Note that

|∇|s fN = [(2π|ξ|)s ψN (ξ)]∨ ∗ f

∨
2π|ξ| s
 
s
=N ψ(ξ/N )) ∗ f.
N

Let
χ(ξ) = (2π|ξ|)s ψ(ξ) ∈ Cc∞ (Rd \ {0}), χN (ξ) = χ(ξ/N ).
Then |∇|s fN = N s [N d χ∨ (N ·)] ∗ f . So

k|∇|s fN kp . N s kf kp kN d χ∨ (N ·)k
| {z }
=kχ∨ k1

. N s kf kp .

Using the fattened Littlewood-Paley projection PN , we get

k|∇|s fN kp . N s kfn kp .

On the other hand,

kfn kp = k|∇|−s |∇|s fn kp . N −s k|∇|s fN kp .

Finally, for s > 0,

X
k|∇|s f≤N kp . k|∇|s fM kp
M ≤N
X
. M s kfM kp
M ≤N
| {z }
.kf kp
s
. N kf kp .

For high frequencies,

X
kf>N kp . kfM kp
M >N
X
. M −s k|∇|s fM kp
| {z }
M >N
.k|∇|s f kp

. N −s k|∇|s f kp .

81
20.2 Khinchine’s inequality
Lemma 20.1 (Khinchine’s inequality). Let {Xn }n≥1 be independent, identically distributed
random variables with Xn = ±1 with equal probability. Let {cn }n≥1 ⊆ C and 0 < p < ∞.
Then  p 1/p
X sX
E  cn Xn  ∼p |cn |2 .
n≥1 n≥1

One way to think about this is that a random variable’s “size” is given by its variance.
For p = 2,
 2
X X  X 
E cn Xn  = E[ cn Xn cm Xm ]
n≥1
X X :0
|cn |2 E[Xn2 ] +


= cn cm
E[Xn Xm ]
 
n6=m
X
= |cn |2 .

So this basically says that this orthogonality persists, even in an Lp sense.

Proof. Without loss of generality, we may assume cn ∈ R.

 p
X Z ∞ X  dλ
E cn Xn  = p λp P cn Xn > λ
0 λ
n≥1

By Chebyshev,
X  P
P cn Xn > λ ≤ e−λt E[et cn Xn ]
" #
Y
= e−λt E etcn Xn
n
Y
−λt
=e E[ctcn Xn ]
n
Y etcn + e−tcn
= e−λt
n
2
Y
= e−λt cosh(tcn )
n
2 /2
Use that cosh x ≤ ex .
t 2 c2 /2
Y
= e−λ et n

82
 2 2
P
= e−λt+t /2( cn ) .

Choose t such that λt = t2 c2n ; so t = λ/ c2n . We get

P P

X  2
P 2
P cn Xn > λ ≤ e−λ /(2 cn ) .

The same argument gives

X  P 2
P 2
P cn Xn < −λ ≤ e−λt E[e−t cn Xn ] ≤ e−λ /(2 cn ) .

So we have
 p
X Z ∞ X  dλ
E cn Xn =p λp P cn Xn > λ
0 λ
n≥1
Z ∞
2 /(2 c2n ) dλ
P
.p λp e−λ
0 λ
pP
Make the change of variables β = λ/ c2n .

p/2 Z ∞
X 2 /2 dβ
.p c2n β p e−β .
β
|0 {z }
.p 1

For the other inequality, for 1 < p < ∞,

X 
X 2
2
|cn | = E cn Xn
 0
p0 1/p
X
hX p i1/p
.E cn Xn E cn Xn ,
| √ {z }
|cn |2
P
.

which gives us
qX hX p i1/p
c2n . E cn Xn .

For 0 < p ≤ 1, we use Cauchy-Schwarz instead:

X 
X 2
2
|cn | = E cn Xn
X 
p/2 X 2−p/2
=E cn Xn cn Xn

83
 4−p 1/2
X 
hX p i1/2
.E cn Xn E cn Xn .
| {z }
|cn |2 )1/2·1/2·(4−p)
P
.(

So we get that
X p/4 hX p i1/2
|cn |2 .E cn Xn .

Now raise both sides to the power 2/p.

20.3 Littlewood-Paley square function estimate

Theorem 20.2 (Littlewood-Paley square function estimate). Let f ∈ S(Rd ) and define
the square function qX
S(f ) = |fN |2 .
Then
kS(f )kp ∼p kf kp ∀1 < p < ∞.

Proof. Let’s prove kSf kp .p kf kp . Let {XN }n∈2Z be iid random variables with Xn = ±1
with equal probability. Let X
mX (ξ) = XN ψN (ξ).
N ∈2Z

Note that X
m∨
X ∗f = XN fN .
N ∈2Z

We claim that mX is a Mikhlin multiplier uniformly in the choice of XN .

X
|Dξα mX (ξ)| . N |α| |Dξα ψ|(ξ/N )
N ∈2Z

Since ψ has compact support on R \ {0}, only finitely many N contribute to the sum.

. |ξ|−α .

We will finish the proof next time.

Remark 20.1. We could replace ψ by any Cc∞ (Rd \ {0}) and still get a Mikhlin multiplier.

84
21 Estimates on the Littlewood-Paley Square Function and
the Fractional Product Rule
21.1 Estimates on the Littlewood-Paley square function
Theorem 21.1 (Littlewood-Paley square function estimate). Let f ∈ S(Rd ) and define
the square function qX
S(f ) = |fN |2 .
Then
kS(f )kp ∼p kf kp ∀1 < p < ∞.

Proof. Let {XN }N ∈2Z be iid random variables

P with Xn = ±1 with equal probability, and
define the random variable mX (ξ) = Xn ψN (ξ). Last time, we showed that mX is a
Mikhlin multiplier, uniformly in the choice of XN . This holds even if we replace ψ be
another Cc∞ (Rd \ {0}) function. Now
X
m∨
X ∗f = XN fN .
N ∈2Z

By Kinchine’s inequality,
qX
E[|m∨ 2 1/p
X ∗ f| ] ∼ |fN |2 ∼p S(f ).

Now
Z
kS(f )kpp ∼ E[|m∨ p
X ∗ f | (x)] dx
Z 
∨ p
∼E |mX ∗ f | (x) dx
| {z }
p
km∨
X ∗f kp

. E[kf kpp ]
. kf kpp .

Again, note that this holds for any Cc∞ (Rd \ {0}) function in place of ψ.
To prove the reverse inequality, we argue by duality and use the generality under which
we proved the first inequality. We say

kf kp = sup hf, gi
kgkp0 =1
DX E
= sup PN f, g
kgkp0 =1

85
Since PeN PN = PN and Pen is self-adjoint,
X
= sup hPN f, PeN gi
kgkp0 =1
N ∈2Z
Z sX sX
≤ sup |PN f |2 |PeN g|2 dx
kgkp0 =1 N N
Using Hölder,
sX
≤ kS(f )|p sup |PeN g|2 .
kgkp0 =1 N
p0

Replacing ψ by ψ(ξ)
e = ψ(2ξ) + ψ(ξ) + ψ(ξ/2) ∈ Cc∞ (Rd \ {0}) in the previous argument,
we get
sX
|PeN g|2 . kgkp0 . 1.
N
p0

Corollary 21.1. Fix 1 < p < ∞. Then

1. Whenever s > −d and f ∈ S(Rd ) (or s ∈ R and fb ∈ Cc∞ (Rd \ {0})),

qX
k|∇|s f kp ∼p N 2s |fN |2 .
p

2. For s > 0 and f ∈ S(Rd ),

qX
k|∇|s f kp ∼p N 2s |f≥N |2 .
p

Proof.
pP
1. Let’s show that N 2s |fN |2 . k|∇|s f kp . We have
p
X X
N 2s |fN |2 = N 2s ||∇|−s |∇|s fN |2
X
= |N s |∇|−s PN (|∇|s f )|2
 s
1 ∞ d N
Replacing ψ by χ(ξ) = (2π|ξ|) s ψ(ξ) ∈ Cc (R \{0} and ψN by χN (ξ) = 2π|ξ| ψN (ξ),
we get
qX
|N s |∇|−s PN (|∇|s f )|2 .p k|∇|s f kp .
p

86
 To prove the reverse inequality, we argue by duality:

k|∇|s f kp = sup h|∇|s f, gi

kgkp0 =1
XD E
= sup |∇|s fN , PeN g
kgkp0 =1 N
XD E
s −s se
= sup N fN , N |∇| PN g
kgkp0 =1 N

0
Recall that F = {h ∈ S(Rd ) : bh vanishes in a nbhd of 0} is dense in Lp . So we can
always take g to be in this family. So
Z qX qX
s
k|∇| f kp ≤ sup 2s
N |fN | 2 N −2s ||∇|s Pen g|2 dx
g∈F
kgkp0 =1
qX qX
≤ N 2s |fN |2 sup N −2s ||∇|s PeN g|2 .
p kgkp0 =1 p

Replacing ψ by
 s
s 2π|ξ|
χ(ξ) = (2π|ξ|) ψ(ξ),
e χn (ξ) = ψeN (ξ),
N
we get
qX
N −2s ||∇|s PeN g|2 . kgkp0 . 1.
p0

N 2s |f≥N |2 ∼
N 2s |fN |2 . We have
P P
2. We claim that
  
X X X X
N 2s |f≥N |2 = N 2s  fN1   fN2 
N N N1 ≥N N2 ≥N

By paying a factor of 2, we can assume N1 ≤ N2 .

X
≤2 N 2s |fN1 | · |fN2 |
N ≤N1 ≤N2
X
. N12s |fN1 ||fN2 |
N1 ≤N2
X  N1 s
. (N1s |fN1 |) (N2s |fN2 )
N2
N1 ≤N2

87
 By Cauchy-Schwarz,
X
. N 2s |fN |2 .
N

On the other hand,

|fN | = |f≥N − f≥2N | ≤ |f≥N f≥2N |.

So
X X X
N 2s |fN |2 . N 2s |f≥n |2 + 2−2s (2N )2s |f≥2N |2
N N N
X
2s 2
. N |f≥N | .
N

21.2 The fractional product rule

Theorem 21.2 (Fractional product rule, Christ-Weinstein, 1991). Fix s > 0 and 1 <
p, p1 , p2 , q1 , q2 < ∞,. Then

k|∇|s (f g)kp . k|∇|s f kp1 kgkp2 + kf kq1 + k|∇|s gkq2 .

1 1 1 1 1
whenever p = p1 + p2 = q1 + q2 .

Remark 21.1. p2 and q1 are allowed to be ∞.

We really should only be proving this for 0 < s < 1, since for integers, we can just use
the regular product rule and then look at the fractional part.

Proof. We have
sX
s
k|∇| (f g)kp ∼ N 2s |PN (f g)|2 .
N
p

We write f g = f≥N/4 g + f>N/4 g≥N/4 + f<N/4 g<N/4 , so


: 0
PN (f g) = PN (f≥N/4 g) + PN (f>N/4 g≥N/4 ) + PN (f g<N/4 ).
<N/4


This gives

|PN (f g)| . M (f≥N/4 g) + M (f≤N/4 g≥N/4 )

. M (f≥N/4 g) + M ((M f )g≥N/4 ).

88
So we get
X X X
N 2s |PN (f g)|2 . |M ((N s f≥N/4 )g)|2 + |((M f ) · N s g≥N/4 )|2 ,

which gives
qX qX qX
N 2s |PN (f g)|2 . |M ((N s f≥N/4 )g)|2 + |((M f ) · N s g≥N/4 )|2 .

So we get
q q
k|∇|s (f g)kp . N 2s |f≥ N/4|2 g + Mf N 2s |g≥ N/4|2
p p

By the corollary,

. k∇|s f kp1 kgkp2 + kf kq1 k|∇|s gkq2 .

89
22 The Fractional Chain Rule
22.1 Proof of the fractional chain rule
Theorem 22.1 (Fractional chain rule, Christ-Weinstein, 1991). Let F : C → C be such
that
|F (u) − F (v)| ≤ |u − v|[G(u) + G(v)],
where G : C → [0, ∞). Then for 0 < s < 1, 1 < p, p1 < ∞, 1 < p2 ≤ ∞ such that
1 1 1
p = p1 + p2 ,
k|∇|s (F ◦ u)kp . k|∇|s ukp1 · kG ◦ ukp2 .
Example 22.1. Consider some nonlinear interaction: Let F (u) = |u|p u, where p > 0.
Then
|F (u) − F (v)| . |u − v|[|u|p + |v|p ],
so we get a bound.
Proof. Last time, we showed that
qX
k|∇|s (F ◦ u)kp ∼ N 2s |PN (F ◦ u)|2 .
p

Let’s calculate
Z
[PN (f ◦ u)](x) = N d ψ ∨ (N u)(F ◦ u)(x − y) dy

We
R ∨want to isolate u in this expression. We will use the locally Lipschitz condition. Since
ψ dy = ψ(0) = 0,
Z
= N d ψ∨ (N y)[(F ◦ u)(x − y) − (F ◦ u)(x)] dy.

So we have
Z
|PN (F ◦ u)|(x) ≤ N d |ψ ∨ (N y)| · |u(x − y) − u(x)|[(G ◦ u)(x − y) + (G ◦ u)(x)] dy

We expect cancellation in the u terms at low frequencies. So we decompose

X
|u(x − y) − u(x)| ≤ |u>N (x − y)| + |u>N (x)| + |uk (x − y) − uk (x)| .
| {z } | {z }
I II k≤N
| {z }
III

Let’s consider the contribution of I:

Z
N d |ψ ∨ (N y)|u>N (x − y)|[(G ◦ u)(x − y) + (G ◦ u)(x)] dy

90
 R d ∨
We can bound this using the maximal function. We have N |ψ (N y)||g(x
R − y)| dy .
d 1 |g(x − y)| dy . 1
R P R
|y|≤1/N |g(x − y)| dy + R∈2N R/N ≤|y|≤2R/N N R2d |B(0,1/N )| B(0,1/N ) |g(x −
y)| dy + · · · .

. M (u≥N (G ◦ u))(x) + M (u>N )(x)(G ◦ u)(x)

. M (u>N (G ◦ u))(x) + M (u>N )(x)M (G ◦ u)(x).

This contributes the following to the original estimate:

qX qX
N 2s |M (u>N (G ◦ u))|2 + N 2s |M (u>N )M (G ◦ u)|2
p p
qX qX
. |M (N s u>N )(G ◦ u)|2 + M (G ◦ u) |M (N s u>N )|2
p p

Using our bounds for the vector-valued maximal function and Hölder,
sX qX
. |N s u>N |2 (G ◦ u) + |N s u>N |2 kM (G ◦ u)kp2
N p1

. k|∇|s ukp1 · kG ◦ ukp2 .

This is an acceptable contribution for what we want to prove.

Let’s look at what II. To PN (F ◦ u), this contributes
Z
N d |ψ ∨ (N y)|u>N (x)|[(G ◦ u)(x − y) + (G ◦ u)(x)] dy

. |u>N (x)|M (G ◦ u)(x) + |u>N (x)|(G ◦ u)(x)

. M (u>n )(x)M (G ◦ u)(x).

As before, the contribution of II to the right hand side of the original estimate is acceptable.
We turn to III. We claim that

|uk (x − y) − uk (x)| . k|y| · [M (uk )(x − y) + M uk (x)]

We split into cases:

1. k|y| > 1: Then

|uk (x − y) − uk (x)| ≤ |Pek uk |(x − y) + |Pek uk |(x)

. (M uk )(x − y) + (M uk )(x).

2. k|y| ≤ 1:
Z
|uk (x − y) − uk (x)| = k d ψe∨ (kz)[uk (x − y − z) − uk (x − z)] dz

91
 Z
= k d [ψe∨ (k(z − y)) − ψe∨ (kz)]uk (x − z) dz

Using the fundamental theorem of calculus,

Z Z 1
d
= k k|y| |∇ψe∨ |(kz − θky) |uk (x − z)| dθ dz
0 | {z }
.1/hkz−θkyi2d .1/hkzi2d

. k|y| · (M uk )(x),

proving the claim.

To PN (f ◦ u), the term III contributes
Z X
N d |ψ ∨ (N y)| K|y|[(M uk )(x − y) − M (uk )(x)] · [(G ◦ u)(x − y) + (G ◦ u)(x)] dy
K≤N
X k Z N |y|
. Nd [(M uk )(x − y) + (M uk )(x)] · [(G ◦ u)(x − y) + (G ◦ u)(x)] dy
k≤N
N hN |y|i3d
X k
. · [M ((M uk ) · (G ◦ u))(x) + M (M uk )(x) · M (G ◦ u)(x)].
N
k≤N

The contribution of III to the right hand side of the original estimate is
v v
u 2 u 2
uX X k u X k
N 2s + tN 2s
u u
. t M ((M uk )(G ◦ u)) M (M uk ) · M (G ◦ u)
N N
N k≤N k≤N
p p

Both cases have terms like

2
X X k X kL
2s
N ck ≤2 N 2s |ck ||cL |
N NN
N k≤N k≤L≤N
X k
. L2s |ck ||cL |
L
k≤L
X  k 1−s
. k s |ck |Ls |cL |
L
k≤L

By Cauchy-Schwarz (or Schur’s test),

sX sX
. k 2s |ck |2 L2s |cL |2
k L

92
 X
. N 2s |cN |2 .
N

And we use our maximal function bounds to finish the proof.

93
23 Introduction to Oscillatory Integrals
23.1 Decay of integrals with compactly supported integrand
Oscillatory integrals are of two types
1. First kind: Z
I(λ) = eiλφ(x) ψ(x) dx,

where λ > 0, φ : Rd → R, and ψ : Rd → C. In this case, we are interested in

the asymptotic behavior of I(λ) as λ → ∞ (think of λ as time). This is covered in
Chapter 8 of Stein’s Harmonic Analysis textbook.

2. Second kind: Z
(Tλ f )(x) = eiλφ(x,y) K(x, y)f (y) dy,

where λ > 0, φ : Rd × Rd → R, K : Rd × Rd → C, and f : Rd → C. We are interested

in the asymptotic behavior of the norm of Tλ as λ → ∞. This is covered in Chapter
9 of Stein’s Harmonic Analysis textbook.
In this class, we’ll discuss oscillatory integrals of the first kind, first for d = 1, where
we will develop a more complete theory.
Proposition 23.1. Let φ : R → R, ψ : R → C be smooth functions. AssumeR supp ψ ⊆
b
(a, b) (nonempty interval), and suppose φ0 (x) 6= 0 for all x ∈ [a, b]. Then I(λ) = a eiλφ(x) ψ(x) dx
satisfies |I(λ)| .N λ −N for all N ≥ 0.
Proof. We use integration by parts. Write
1 d iλφ(x)
eiλφ(x) = 0
(e ).
iλφ (x) dx
Integrating by parts,
Z b   Z b  
d iλφ(x) ψ(x) iλφ(x) d ψ(x)
I(λ) = (e ) dx = e − dx.
a dx iλφ0 (x) a dx iλφ0 (x)
Let
1 df
(Df )(x) = (x).
iλφ0 (x) dx
The transpose of D is  
t d f (x)
Df (x) = − .
dx iλφ0 (x)
Note that
DN (e−λφ ) = eiλφ ∀N ≥ 1.

94
Now
Z b
I(λ) = DN (eiλφ(x) )ψ(x) dx
a
Z b
= eiλφ(x) (t D)N ψ(x) dx
a
Z b  N
iλφ(x) d 1
= e − .
a dx iλφ0 (x)

We get
N
X X ψ (β) φ(1+α1 ) · · · φ(1+αk )
|I(λ)| . λ−N .N λ−N .
(φ0 )N +k
k=0 β+α1 +···+αk =N L1 (a,b)
αj ≥1

Remark 23.1. If ψ is not compactly supported inside (a, b), then we don’t expect better
than λ−1 decay.
Z b
eiλb − eiλa
eiλx dx = .
a iλ

23.2 The Van der Corput lemma

Proposition 23.2 (Van der Corput lemma). Let φ : R → R be smooth. Fix k ≥ 1, and
assume |φ(k) (x)| ≥ 1 for all x ∈ [a, b]. If k = 1, assume also that φ0 is monotonic on [a, b].
Rb
Then I(λ) = a eiλφ(x) dx satisfies |I(λ)| ≤ ck λ−1/k , where ck is independent of φ, λ, a, b.

Remark 23.2. If k = 1, the assumption |φ0 (x)| ≥ 1 is not sufficient to get the claim. Set
λ = 1. Then Z b Z b
|I(λ)| = eiφ(x) dx ≥ cos(φ(x)) dx .
a a

If φ0 is large on {x : cos(φ(x)) < 0} and small on {x : cos(φ(x)) > 0}, then |{x : cos(φ(x)) <
b→∞
0}|  |{x : cos(φ(x)) > 0}|. In particular, |I(λ)| −−−→ ∞.

Let’s prove the Van der Corput lemma.

Proof. We argue by induction on k. First, let k = 1. Then

Z b
I(λ) = eiλφ(x) dx
a
Z b
1 d iλφ(x)
= 0
(e ) dx
a iλφ (x) dx

95
 b
eiλφ(b) eiλφ(a)
Z  
iλφ(x) d 1
= − − e dx.
iλφ0 (b) iλφ0 (a) a dx iλφ0 (x)

So
Z b
2 1 d 1
|I(λ)| ≤ + dx
λ λ a dx φ0 (x)
Since φ0 is monotonic,
Z b
2 1 d 1
= + dx
λ λ a dx φ0 (x)
2 1 1 1
= + −
λ λ φ0 (b) φ0 (a)
| {z }
≤1
3
≤ .
λ
So c1 = 3.
For the inductive step, assume the claim holds for some k ≥ 1. Assume |φ(k+1) (x)| ≥ 1
for all x ∈ [a, b]. Replacing φ by −φ if necessary, we may assume that φ(k+1) (x) ≥ 1 for all
x ∈ [a, b]. So φ(k) is increasing on [a, b]. Then there exists at most one point c ∈ [a, b] such
that φ(k) (c) = 0. We have two cases:

1. ∃c ∈ [a, b] such that φ(k) (c) = 0: Since φ(k) grows at least linearly, there is a δ such
that |φ(k) (x)| ≥ δ for all x ∈ [a, b] \ (c − δ, c + δ). Then

dk
φ(δ −1/k x) ≥ 1 ∀δ −1/k x ∈ [a, b] \ (c − δ, c + δ).
dxk

Then
Z c−δ Z c+δ Z b
iλφ(x) iλφ(x)
I(λ) = e dx + e dx + eiλφ(x) dx.
a c−δ c+δ

Using the change of variables x = δ −1/k y,

Z c−δ Z δ 1/k (c−δ)
−1/k y)
e iλφ(x)
dx = eiλφ(δ δ −1/k dy ≤ ck (δλ)−1/k
a δ 1/k a

by the inductive hypothesis. Similarly,

Z b
eiλφ(x) dx ≤ ck (λδ −1/k .
c+δ

96
 For the remaining term, we have
Z c+δ
eiλφ(x) dx ≤ 2δ.
c−δ

We get
I(λ)| ≤ 2ck (λδ)−1/k + 2δ.
Now choose δ such that (λδ)−1/k = δ =⇒ δ = λ−1/(k+1) . Then
|I(λ)| ≤ 2(ck + 1) λ−1/(k+1) .
| {z }
ck+1

2. φ(k) (x) 6= 0 for all x ∈ [a, b]: If φ(k) (a) > 0,

Z a+δ Z b
|I(λ)| ≤ eiλφ(x) dx + eiλφ(x) dx ≤ δ + ck (λδ)−1/k
a a+δ

as in the previous case. Setting δ = λ−1/(k+1) , we get

|I(λ)| ≤ (ck + 1)λ−1/(k+1) .
Similarly, if φ(k) (a) < 0, then φ(k) (b) < 0. So we split
Z b−δ Z b
−λφ(x)
|I(λ)| ≤ e dx + eiλφ(x) dx ≤ ck (δδ)−1/k + δ ≤ (ck + 1)δ −1/(k+1) .
a b−δ

Corollary 23.1. Let φ : R → R and ψ : R → C be smooth. Fix k ≥ 1, and assume

|φ(k) (x)| ≥ 1 for all x ∈ [a, b]. If k = 1, assume also that φ0 is monotonic. Then
Z b
I(λ) = eiλφ(x) ψ(x) dx
a

satisfies  Z b 
|I(λ)| ≤ ck λ−1/k |ψ(b)| + |ψ 0 (x)| dx .
a
Proof. Write
Z b  Z x 
d iλφ(y)
I(λ) = ψ(x) e dy dx
a dx a
Using integration by parts,
Z b Z b Z x 
iλφ(y) 0 iλφ(y)
= ψ(b) e dy − ψ (x) · e dy dx.
a a a

97
24 Estimating Oscillatory Integrals With Stationary Phase
24.1 Estimation in the 1 dimensional case
Proposition 24.1 (stationary phase). Assume φ : R → R is smooth and has a non-
degenerate critical points at x0 ; that is, φ0 (x0 ) = 0 and φ00 (x0 ) 6= 0. Assume ψ : R → C is
smooth and supported in a sufficiently small neighborhood of x0 . Then
Z
I(λ) = eiλφ(x) ψ(x) dx
√ 00
= eiλφ(x0 ) ψ(x0 ) 2πei(π/4) sgn(φ (x0 )) |φ00 (x0 )|−1/2 |−1/2 λ−1/2 + O(λ−3/2 )
as λ → ∞.
Remark 24.1. If we are not interested in the coefficient of the leading order term, then
we can argue as follows: Let a ∈ Cc∞ be such that
(
1 |x| ≤ 1
a(x) =
0 |x| > 2
and decompose
I(λ) = I1 (λ) + I2 (λ),
Z
I1 (λ) = eiλφ(x) ψ(a)a(λ1/2 (x − x0 )) dx.

Then
Z
|I1 (λ)| ≤ kψk∞ |a(λ1/2 (x − x0 ))| dx

≤ kψk∞ kak∞ · λ1/2 ,

Z
I2 (λ) = eiλφ(x) ψ(x)[1 − a(λ1/2 (x − x0 ))] dx.

Note that supp(ψ(x)[1 − a(λ1/2 (x − x0 ))]) ⊆ {λ−1/2 ≤ |x − x0 | .ψ 1}. If supp ψ is such

that φ0 (x) 6= 0 for x ∈ (supp ψ) \ {x0 }, then integration by parts gives
|I2 (λ)| .m λ−m ∀m ≥ 0.
Proof. Write

 : 0 φ00 (x0 )

φ(x) = φ(x0 ) + φ0
(x
0 )(x
 − x0 ) + (x − x0 )2 + O(|x − x0 |3 ).
 2
Rewrite this as
φ00 (x0 )
φ(x) − φ(x0 ) = (x − x0 )2 [1 + η(x)],
2
where η(x) = O(|x − x0 |). Let U be a small neighborhood of x0 such that

98
 1. |η(x)| < 1 for all x ∈ U
2. φ0 (x) 6= 0 for all x ∈ U \ {x0 }.
p
Assume supp ψ ⊆ U . Change variables to y(x) = (x − x0 ) 1 + η(x). This is a diffeomor-
phism from U to a neighborhood of y = 0. Then
Z
iλφ(x0 )
I(λ) = e eiλ[φ(x)−φ(x0 )] ψ(x) dx
Z
iλφ(x0 ) 00 2
=e eiλ(φ (x0 )/2)y ψ(y)
e dy,

where ψe ∈ Cc∞ is supported in a neighborhood of y = 0 and ψ(0) e = ψ(x0 ).

λφ 00 (x )
Let ψ ∈ Cc∞ be such that ψe = 1 on supp ψ.
e Let λe= 0
. Then
ee e
2
Z
e 2 2 2
I(λ) = eiλφ(x0 ) eiλy e−y [ey ψ(y)] ψ(y) dy
e ee

Using a Taylor expansion, we write

N 1
dN +1 |·|2 e
Z
y2
X
j N +1 1
e ψ(y)
e = aj y + y RN (y), RN (y) = (1 − t)N [e ψ](ty) dt.
N! 0 dy N +1
j=0

This leads us to consider 3 terms:

N Z
2 2
X
I=e iλφ(x0 )
aj eiλy e−y y j dy,
e

j=0
Z
e 2 2
II = eiλφ(x0 ) eiλy e−y PN (y)[ψ(y) − 1] dy,
ee
Z
e 2 2
III = eiλφ(x0 ) eiλy e−y y N +1 RN 9y)ψ(y) dy.
ee

d iλy 2
Since dy e , we can pull off a factor of y using integration by parts. By picking N to be
e

large enough, we can get as much decay in III as we want.

Let’s look at I. Note that the terms with j odd vanish. Consider j = 0 and note that
a0 = ψ(x0 ). The contribution is
√
Z
e 2 2
e iλφ(x0 )
ψ(x0 ) eiλy e−y dy = eiλφ(x0 ) ψ(x0 )(1 − iλ)
e 1/2 π.

p
To see what happens when λ → ∞, write 1−iλ e = reiσ , where r = e2 and tan σ = −λ.
1+λ e
e −1/2 = r−1/2 e−iσ/2 . Then
Then (1 − iλ)
 r −1/2
−1/2 1
r = |λ| 1 +
e
e2
λ

99
 1 −1/4
 
−1/2
= |λ|
e 1+
e2
λ
 00
−1/2
λ|φ (x0 )|
1 + O(λ−2 )


=
2
 1/2
2
= + O(λ−5/2 ),
λ|φ00 (x0 )|
λφ00 (x0 ) λ→∞
tan σ = −λ
e=− −−−→ − sgn(φ00 (x0 )) · ∞.
2
So σ → − sgn(φ00 (x0 )) · π2 , and we get
s
2π 00
eiλφ(x0 ) ψ(x0 ) 00
esgn(φ (x0 ))π/4 + O(λ−5/2 ).
λ|φ (x0 )|

For j ≥ 2 even,
Z Z
e 2 2
e −1/2−j/2 e−y2 y j dy . λ−(j+1)/2 . λ−3/2 .
eiλy e−y y j dy = (1 − iλ)

2
7 e−y PN (y)[ψ(y)
Now consider II. Note that y → − 1] is supported away from the origin.
ee
Integration by parts gives
| II | .m λ−m ∀m ≥ 0.
Consider III. Decompose III = III1 + III2 , where
Z
e 2 2
III1 = e iλφ(x0 )
eiλy e−y y N +1 RN (y)ψ(y)a(y/ε) dy.
ee

Then
Z
| III1 | . |y|N +1 dy . εN +2 .
|y|≤ε

The other term is Z

2
III2 = eiλφ(x0 ) eiλy y N +1 [1 − a(y/ε]b(y) dy,
e

2
where b(y) = e−y RN (y)ψ(y). Integration by parts gives
ee
Z  m
iλφ(x0 ) e 2
iλy d 1
III2 = e e · − [y N +1 (1 − ay/εb(y)] dy,
dy 2iλy
e
so
m
1 X X y N +1−α1 ε−α2 a(α2 ) (y/ε)b(α3 ) (y)
| III2 | .m
λm y m+k
k=0 α1 +α2 +α3 =m−k L1

100
 Z
1
.m |y|N +1−2m dy
λm |y|≥ε
ε N +2−2m
.
λm
N +2
if m > 2 . Now choose ε such that

εN +2−2m
εN +2 = ⇐⇒ ε = λ−1/2
λm

to get | III | . λ−(N +2)/2 . λ−3/2 if N ≥ 1.

101
25 Oscillatory Integrals in Higher Dimensions
25.1 Nonstationary phase
Here is the case of nonstationary phase.

Proposition 25.1. Let φ : Rd → R, ψ : Rd → CRbe smooth. Assume supp ψ is compact

6 0 for all x ∈ supp ψ. Then I(λ) = eiλφ(x) ψ(x) dx satisfies
and k∇φ(x)| =

|I(λ)| .m λ−m ∀m ≥ 0.

Proof. As in the 1 dimensional case, we use integration by parts. We write

∇φ(x)
eiλφ(x) = · ∇(eiλφ(x) ).
iλ|∇φ(x)|2

Then
 
∇φ(x)
Z
I(λ) = eiλφ(x) ∇ · ψ(x) dx,
iλ|∇φ(x)|2
so

|I(λ)| . λ−1 ,

where the implicit constant depends on the C 2 norm of φ and the C 1 norm of ψ. Now
iterate.

There is an equivalent of Van der Corput’s lemma.

Proposition 25.2. Let φ : Rd → R, ψ : Rd → C be smooth. Assume ψ is compactly

supported and |Dα φ(x)| ≥ 1 for al x ∈ supp ψ for some α ∈ Nd with |α| ≥ 1. Then
R iλφ(x)
I(λ) = e ψ(x) dx satisfies

|I(λ)| ≤ C(|α|, φ)λ−1/|α| [kψk∞ + k∇ψk1 ].

Remark 25.1. This is worse than the previous proposition when |α| = 1. We will also
beat it when |α| = 2, so we will not actually prove it.

25.2 Stationary phase and Moore’s change of variables lemma

Here is the case of stationary phase.

Proposition 25.3 (stationary phase). Let φ : Rd → R be smooth, h 2andi assume φ has a

∂ φ
nondegenerate critical point at x0 ; that is, ∇φ(x0 ) = 0, but det ∂xi xj
(x0 ) 6= 0.
1≤i,j≤d

102
Assume that ψ : Rd → C is smooth and supported in a sufficiently small neighborhood of
x0 . Then
Z
I(λ) = eiλφ(x) ψ(x) dx

= eiλφ(x0 ) ψ(x0 )(2πi)d/2 λ−d/2 (det[D2 φ(x0 )])−1/2 + O(λ−d/2−1 )

as λ → ∞.

Remark 25.2. If we just aim for the correct decay order (and not the precise coefficient),
we argue as follows: Let a : Rd → R be a cutoff with
(
1 |x| ≤ 1
a(x) =
0 |x| ≥ 2

and decompose I(λ) = I1 (λ) + I2 (λ), where

Z
I1 (λ) = eiλφ(x) ψ(x)a(λ1/2 (x − x0 )) dx.

Then

|I1 (λ)| . λ−d/2 .

Integration by parts gives

|I2 (λ)| .m λ−m ∀m ≥ 0.

Lemma 25.1 (Morse). If x0 is a nondegenerate critical point of a smooth function φ :

Rd → R, then there exists a smooth change of variables x 7→ y(x) such that y(x0 ) = 0,
∂y
∂x (x0 ) = Id, and
d
X 1
φ(x) − φ(x0 ) = λj yj2 ,
2
j=1

where λ1 , . . . , λd are the eigenvalues of D2 φ(x0 ).

Proof. Performing an orthogonal change of variables, we may assume that D2 φ(x0 ) =

diag(λ1 , . . . , λd ). By Taylor expansion,
Z 1
:0 d2
φ(x) = φ(x0 ) +  ∇φ(x0 ) · (x − x0 ) + (1 − t) 2 [φ(x0 + t(x − x0 ))] dt.
 


0 dt
So
Z 1
d
φ(x) − φ(x0 ) = (1 − t) [(x − x0 ) · ∇φ(x0 + t(x − x0 ))] dt
0 dt

103
 1
XZ ∂2φ
= (1 − t)(x − x0 )i (x − x0 )j (x0 + t(x − x0 )) dt
∂xi ∂xj
i,j≥1 0
X
= (x − x0 )i (x − x0 )j mi,j (x),
i,j≥1

R1 2
where mi,j (x) = 0 (1−t) ∂x∂i ∂x
φ
j
(x0 +t(x−x0 )) dt. Note that the mi,j are smooth, mi,j (x) =
1 ∂2φ
mj,i (x), and mi,j (x0 ) = 2 ∂xi ∂xj (x0 ). So

1
[mi,j (x0 )]1≤i,j≤d = diag(λ1 , . . . , λd ).
2
We argue inductively. Assume
1 1 X
φ(x) − φ(x0 ) = λ1 y12 + · · · + λr−1 yr−1
2
+ m
e i,j (y)yi yj
2 2
i,j≥r

∂y
for some 1 ≤ r ≤ d, where y(x0 ) =0, ∂x (x0 ) = Id, and m e i,j = me j,i . We know that
2
D [RHS(x)]x=x0 = diag(λ1 , . . . , λd ). Then

∂2 ∂ 2 yk
   
1 2 ∂yk ∂yk
λk yk = λk + λk yk
∂xi ∂xj 2 x=x0 ∂xi ∂xj ∂xi ∂xj x=x0
= λk δi,k δj,k .

So   
X
D2  m
e i,j (y)yi yj  (x0 ) = diag(0, . . . , 0, λr , . . . , λd ).
i,j≥r

We now have
 
∂2 X X
 m
e i,j (y)yi , yj  = m
e i,j (0) (δk,i δ`,j + δ`,i δk,j ) .
∂xk ∂x`
i,j≥r i,j≥r
x=x0

This tells us that

1
[m
e i,j (0)]r≤i,j≤d = diag(λr , . . . , λd )
2
Change variables as follows:

y 0 = yj j 6= r
j q  
m
e j,r (y)
yr0 = m e r,r (y) P
λr /2 yr + j≥r+1 e r,r (y) yj
m .

104
 ∂y 0
We need to show that this is a diffeomorphism with y 0 (x0 ) = 0, ∂x |x=x0 = Id, and

1 1 X
φ(x) = φ(x0 ) = λ1 y12 + · · · + λr (yr0 )2 + m
e i,j (y)yi yj .
g
2 2
i,j≥r+1

We have y 0 (x0 ) = 0 because each yi is 0 at x0 . For j 6= r,

∂yj0
= δi,j ,
∂xi x−x0

so  
s 0
∂yr0 m
e r,r (0)  X m
e j,r(0)
>

= δi,r + δj,i  = δi,r

∂xi x=x0 λr /2 m
e 
 r,r (0)
j≥r+1

Now we have
X 1 X
e i,j (y)yi yj − λr (yr0 )2 =
m m
e i,j (y)yi yj
2
i,j≥r i,j≥r

X m e j,r (y)
e r,r yr2 + 2
−m yi , yr
me r,r (y)
j≥r+1

X m e i,r (y)me j,r (y)
+ yi , yj 
m
e r,r (y)me r,r (y)
i,j≥r+1
X  m
e i,r (y)me j,r (y)

= e i,j (y) −
m yi , yj .
me r,r (y)
i,j≥r+1 | {z }
=m
e i,j (y)
e

This completes the proof.

105