The theory of Ap weights: a modern introduction1

CARLOS PEREZ
University of Seville

1 The Hardy-Littlewood maximal function.

The Hardy-Littlewood maximal function is the (non-linear) operator defined
by Z
1
M f (x) = sup |f (y)| dy
x∈Q |Q| Q

where the supremum is taken over all the cubes containing x. By cube we
always mean a cube with sides parallel to the axes. There are variations
of this definition such as replacing cubes by balls or just considering cubes
or balls centered at x but all of them are equivalent except for dimensional
constants. (See [Jo] for more information about all these).
Maximal functions arise very naturally in analysis, for proving theorems
about the existence almost everywhere of limits, for controlling pointwise
important objects such as the Poisson Integrals or for controlling, not point-
wise but at least in average, other basic operators such as singular integral
operators.
The model example of existence almost everywhere of limits is the Lebesgue
differentiation theorem:
Z
1
f (x) = lim f (y) dy (1)
r→0 |B(x, r)| B(x,r)

and this property is intimately related to the study of certain properties of

the Hardy-Littlewood maximal function. There are other almost everywhere
convergence problems in mathematics: Fourier series, Dirichlet problem, the
heat equation, the Schrodinger equation etc. All of them have the same
pattern, study first the associated maximal operator.
The key property to understand the Lebesgue differentiation theorem is
the so called “weak type” estimate or property of M . This is related to a
1
Notes for the CIMPA ABIDJAN 2014 school.

1
very special an fundamental space, the weak spaces or Marcinkiewicz spaces:

kf kLp,∞ (Rn ) = sup t |{x ∈ Rn : |f (x)| > t}|1/p p > 0.

t≥0

It is easy to see that Lp,∞ (Rn ) ⊂ Lp (Rn ) and that the content is strict.
In applications it is specially important the case p = 1.

Theorem 1.1 (Hardy–Littlewood–Wiener) M is a weak type (1, 1) op-

erator, namely
M : L1,∞ (Rn ) → L1 (Rn ),
which means that there exists a finite (dimensional) constant c such that the
following inequality holds,

kM f kL1,∞ (Rn ) ≤ c kf kL1 (Rn ) (2)

Corollary 1.2 M is of strong type (p, p), 1 < p < ∞ operator, namely,
there exists a constant c such that

kM f k ≤ ckf k (3)
Lp (Rn ) Lp (Rn )

Sometimes we will be using the notation

kT f kLp,∞ (X)
kT kLp,∞ (X) = sup
f 6=0 kf kLp (X)

or the standard one

kT f kLp (X)
kT kLp (X) = sup
f 6=0 kf kLp (X)
As a consequence of (2) we can derive the Lebesgue differentiation theorem
(1). To be more precise we will show that there is a set Lf the Lebesgue
set whose complement Nf is a ”null” set, namely |Nf | = 0 for which
Z
1
f (x) = lim f (y) dy x ∈ Lf (4)
r→0 |B(x, r)| B(x,r)

Observe that this is equivalent to saying

Z
1
lim |f (y) − f (x)| dy = 0 x ∈ Lf
r→0 |B(x, r)| B(x,r)

2
To do this we define
Z
1
Aλ = {x : lim sup |f (y) − f (x)| dy > λ} λ≥0
r→0 |B(x, r)| B(x,r)

It is enough to prove that |Aλ | = 0 for each λ > 0 since A0 = ∪j A1/j and
since x ∈
/ A0 , implies
Z
1
0 = lim sup |f (y) − f (x)| dy
r→0 |B(x, r)| B(x,r)
but then we can replace the lim sup by lim, namely,
Z
1
lim |f (y) − f (x)| dy = 0
r→0 |B(x, r)| B(x,r)

which is exactly (4) if we choose Nf = A0 .

We now prove then that |Aλ | = 0 for each λ > 0. A first important
remark is that (4) holds a priori for a dense class of functions, for instance
for class of continuous functions in Rn :
Z
1
g(x) = lim g(y) dy x ∈ Rn
r→0 |B(x, r)| B(x,r)

Now, if f ∈ L1 (Rn ) and for given  > 0 there exists a continuous function g
such that kf − gkL( Rn ) < ε. Then g satisfies
Z
1
lim |g(y) − g(x)| dy = 0 x ∈ Rn .
r→0 |B(x, r)| B(x,r)

So, if x ∈ Aλ and write f = g + h we immediately have

Z
1
λ< |f (y) − f (x)| dy ≤ M h(x) + |h(x)|
|B(x, r)| B(x,r)
and then
|Aλ | ≤ |{x ∈ Rn : M h(x) > λ/2}| + |{x ∈ Rn : h(x) > λ/2}|
and here comes the main point to apply (2):
Z
cn cn cn
|Aλ | ≤ |h(x)| dx = kf − gkL( Rn ) < ε
λ Rn λ λ
But ε is free so we can let ε → 0 obtaining |Aλ | = 0.
The moral: a crude estimate such as a weak type implies a very saddle
result, the Lebesgue differentiation theorem which involves cancellation of
the function f around x because the function change sign.

3
2 A first but remarkable weighted estimate
In this section we will prove Theorem 1.1 and its Corollary 1.2, actually we
will prove a weighted version for the same effort. This weighted estimates
are very important due to its many applications and because they open the
theory of weights.
We say that w is a weight if w is a a.e. positive locally
R integrable function
n
in R . If E is any measurable set we denote w(E) = E w. Sometimes we
can beyond locally integrable functions as is the case of the main theorem
below.

Theorem 2.1 a) There exists a constant cn such that for all f, w

Z
kM f k 1,∞ ≤ cn |f (x)| M w(x)dx. (5)
L (w)
Rn

b) As a consequence if 1 < p < ∞ there exists a constant C such that for

all f
kM f k p ≤ cn (p0 )1/p kf k p (6)
L (w) L (M w)

We may think of (5) inequality as a kind of duality for the (nonlinear)

operator. It can also be seen as an antecedent of the Ap theory weights that
we will be developing.
One important tool is the so called “the layer cake formula” formula will
plays a central role:
Z Z ∞
ϕ(f ) dµ = ϕ0 (t) µ({x ∈ X : f (x) > t}) dt f ≥0 (7)
X 0

Proof of Theorem 2.1:

We will beging with the proof of (6). We will use part a), namely the
weak type estimate(5). We may assume that f is nonnegative. Write f as
f = f t + ft , where f 1 (x) = f (x) if f (x) > t, and f t (x) = 0 otherwise. Then

M f (x) ≤ M f t (x) + M ft (x) ≤ M f t (x) + t.

Hence, by the cake layer formula

Z Z ∞ Z ∞
p p n dt p d
M f (y) w(y)dy = p t w{y ∈ R : M f (y) > t} = p2 tp w{y ∈ Rn : M f (y) > 2t}
Rn 0 t 0

4
 Z ∞ Z ∞ Z
p p n t dt dt
≤ p2 t w{y ∈ R : M f (x) > t} ≤ cn p2p p−1
t f (y) M w(y)dy
0 t 0 {y∈Rn :f (y)>t} t
by the key weak type inequality (5). Now we continue with
Z Z f (y) Z
p p−1 dt 0 p
≤ cn p 2 t f (y) M w(y)dy = C p 2 f (y)p M w(y)dy
Rn 0 t Rn

and this concludes the proof of (3)

A first proof of (5): The proof of a) is based on the classical Vitali
covering Lemma.

Lemma 2.2 (Vitali) Let F = {Qi }i=1,··· ,N be a finite family of cubes (or
balls) in Rn . Then we can extract from F a sequence of pairwise disjoint
cubes F 0 = {Qj }j=1,··· ,M such that
[ m
[
Q⊂ 5Qj .
Q∈F j=1

For the proof we remit [Ma] p.23 but there are many others good references.
The key point is the a geometrical point: the so called ”engulfing” property
of balls.
If we let Ωλ = {x ∈ Rn : M f (x) > λ} for a given λ and let K to be
any compact subset contained in Ωλ . Let x ∈ K then by defintion of the
maximal function there is a cube Q = Qx containg x such that
Z
1
|f (y)| dy > λ. (8)
|Q| Q
S
Then K ⊂ x∈K Qx and by compactness
S we can extract a finite family of
cubes F = {Q} such that K ⊂ Q∈F Q and where each cube satisfies (8).
Then by the Vitali S
lemma weSM can extract a pairwise disjoint family of cubes
M
{Qj }j=1 such that Q∈F ⊂ j=1 5Qj . Then
M M Z
X 1X 1
w(K) ≤ w(5Qj ) ≤ w(5Qj ) |f (y)| dy
j=1
λ j=1 |Qj | Qj

M Z M Z
C X w(5Qj ) CX
≤ |f (y)| dy ≤ |f (y)| M w(y)dy
λ j=1 |5Qj | Qj λ j=1 Qj

5
 Z
C
≤ |f (y)| M w(y)dy.
λ Rn

A second proof of (5): This second proof is just an application of the

Besicovitch covering lemma which plays a central role in modern analysis.

Lemma 2.3 (The Besicovitch covering lemma) Let K de a bounded set

in Rn and suppose that for every x ∈ K there is an (open) cube Q(x) with
center at x. Then we can find a sequence (possible finite) of points {xj } in
K such that [
K⊂ Q(xj )
j

and X
χQ(xj ) ≤ cn
j

where cn is a finite dimensional constant. In fact we can write

B(n)
[ [ [
Q(xj ) = Qi , (9)
j k=1 i∈Fk

where each of the family {Qi }i∈Fk , k = 1, · · · , B(n), is formed by pairwise

disjoint cubes. B(n) > 1 is usually called the Besicovitch constant.

The proof of this result can be found in several places such as the classical
lecture notes by M. De Guzman [dGuz] [Ma], also in [GrMF].
Indeed, assuming that w is bounded as we may, the first observation is
that (5) is equivalent to
Z
n w
sup λ w({x ∈ R : M (f )(x) > λ}) ≤ C |f (x)| w(x)dx λ > 0.
λ>0 Mw Rn

The second is that we trivially have the pointwise inequality

w
M (f )(x) ≤ cn Mwc f (x),
Mw
where Mwc is the weighted centered maximal function
Z
c 1
Mw f (x) = sup |f (y)| w(y)dy. (10)
r>0 w(Qr (x)) Qr (x)

6
Therefore (5) follows from
Z
n
sup λ w({x ∈ R : Mwc f (x) > λ}) ≤ C |f (x)| w(x)dx
λ>0 Rn
which is a consequence of the Besicovitch covering lemma where c si a di-
mensional constant.

The proof of b) is a particular case of one of the most important theorems

in modern mathematics: the Marcinkiewicz interpolation theorem. An
extension of above argument making a special emphasis on the control on
the bound is the following estimate:

1/p
1/p0
kT kLp (X) ≤ 2 (p0 )1/p kT kL1,∞ (X) kT kL∞ (X) .

2.1 An application: a vector-valued inequality

Inequality (5) was proved by C. Fefferman and E. Stein in a celebrated pa-
per [FS] to derive a vector–valued extension of (3). Indeed, for a vector of
functions f = {fi }i and for q > 0 the operator M q is defined by
∞
!1/q
X
M q f (x) = (M fi (x))q .
i=1

This operator can also be seen as a generalization

P of the qclassical integral of
Marcinkiewicz. Then if we denote |f (x)|q = ( ∞
1/q
|f
i=1 i (x)| ) = kf (x)k`q we
have for 1 < p < ∞ and 1 < q ≤ ∞
Z Z
p
M q f (x) dx ≤ C |f (x)|pq dx. (11)
Rn Rn
A corresponding weak type (1, 1) also holds in the range 1 < p < ∞. These
estimates play an important role in Harmonic Analysis, specially in the the-
ory of Littlewood-Paley.
If p = q there is nothing to be proved, the case p > q and let r = pq > 1.
By the Fatout lemma we may assume that the series is finite from i = 1 to
i = N and then let N → ∞. Raise (11) to the power 1r we can write
P XN Z
N q
i=1 (M fi ) = M fi (y)q g(y)dy
Lr (Rn ) Rn
i=0

7
 0
for some g ∈ Lr (Rn ) with unit norm. But then we can apply (6) to continue
with
N Z
X Z N
X
0 q 0
≤ cn q fi (y) M g(y)dy, ≤ cn q fi (y)q M g(y)dy
i=0 Rn Rn i=0

N
X
0
≤ cn q k fi (y)q kLr (Rn ) kM gkLr0 (Rn )
i=0

and by (6) (with no weight) we continue with

Z 1/r Z 1/r
0 1/r0 0 1/r0
≤ cn q r |f (x)|pq dx kgkLr0 (Rn ) = cn q r |f (x)|pq dx
Rn Rn

concluding the proof.

The case p ≤ q is easy but follows from vector-valued interpolation and
can be found in [GCRdF]

2.2 The A1 condition

The question is a one weighted question, namely for which weights w we have
the boundedness
M : L1 (w) → L1,∞ (w)
In view of the Fefferman-Stein’s inequality (5) these authors defined the
following class of weights.

Definition 2.4 A weight w belongs to the class A1 if there is a constant K

such that Z
1
w(y) dy ≤ K inf w, (12)
|Q| Q Q

and we will denote the infimum of the constant K by [w]A1 .

Observe that K ≥ 1 by the Lebesgue differentiation theorem. And with this

notation we have
kM kL1,∞ (w) ≤ cn [w]A1 .

8
So this means that the A1 condition is a sufficient condition for the bound-
edness M : L1 (w) → L1,∞ (w) but this condition is also necessary. In fact
we will show that
kM kL1,∞ (w) ≥ [w]A1
in an easy way in next section (see (15)).
It is important to have large class of A1 weights to our disposal and there
are eentially two methods. The first method is due to Coifman-Rochberg
[CR].
Lemma 2.5 Let µ be a positive Borel, then for each 0 < δ < 1
(M µ)δ ∈ A1
and furthermore
cn
[(M f )δ ]A1 ≤ . (13)
1−δ
The second method is based in the so called Rubio de Francia’s algorithm
Lemma 2.6 For any measurable function f define If
∞
X 1 M k (h)
R(h) =
k=0
2k kM kkLp (w)

Then we have
(A) h ≤ D(h)
(B) kR(h)kLp (w) ≤ 2 khkLp (w)

(C) [R(h)]A1 ≤ 2 kM kLp (w)

More information about this can be found in [CMP3].

3 Muckenhoupt Ap condition and Mucken-

houpt Ap theorem
The main breakthrough in the theory arrived in the early seventies with the
discovery by B. Muckenhoupt in [Mu] of the right class of weights for which
the Hardy-Littlewood maximal function is bounded on Lp (w), p > 1. This
result opened up the possibility of studying operators with higher singularity.
We show here non standard proofs

9
Definition 3.1 A weight w belongs to the class Ap , 1 < p < ∞, if
 Z  Z p−1
1 1 1−p0
[w]Ap = sup w(y) dy w(y) dy <∞ (14)
Q |Q| Q |Q| Q

Check that by Jensen’s inequality

1≤p≤q =⇒ [w]Aq ≤ [w]Ap i.e. Ap ⊂ Aq

We know prove Muckenhoupt‘s theorem with the sharp constant due to

S. Buckley. It came as a surprised that this theorem can be improved as
shown in [HP] an further simplified in [HPR]. To state these new results
we introduce the concept of A∞ constant. This recent results are intimately
related to the so called A2 conjecture and its improvement.

Theorem 3.2 Let p > 1. The following conditions are equivalent:

a) w ∈ Ap ;
b) there exists a constant K such that for each cube Q and nonnegative func-
tion f  Z p Z
1
f (y) dy w(Q) ≤ [w]Ap f (y)p w(y)dy
|Q| Q Q

c) M : Lp (w) → Lp,∞ (w) and furthermore

1 1
[w] p ≤ kM k ≤ cn [w] p . (15)
Ap Lp,∞ (w) Ap

d) M : Lp (w) → Lp (w) and furthermore

1
kM kLp (w) ≤ cn p0 [w]Ap

p−1
, (16)

where 1 < p < ∞. Furthermore the exponent is sharp in the sense that: for
any θ > 0
1
sup 1
−θ
kM kLp (w) = ∞ (17)
w∈Ap
[w]Ap−1
p

10
 A new way to understand the sharpeness of the exponent in (16) can be
found in [LPR]. It turns out that it is due to the fact that the norm of M
1
as operator on Lp (Rn ) blows up as p−1 when p → 1.
As already mentioned estimate (16) has been improved quite substantially
in [HP] and will be derived in Theorem 5.7.
1
In fact, we cannot replace the function ψ(t) = t p−1 by a “smaller” function
ψ : [1, ∞) → (0, ∞) in the sense that

ψ(t)
inf 1 =0
t>1
t p−1
ψ(t) tβ tβ
(or limt→∞ tβ
= 0, or supt>1 ψ(t)
= ∞ or limt→∞ ψ(t)
= ∞) since then

1
sup kM k = ∞ (18)
w∈Ap ψ([w]Ap )

Conditions a), b) and c) are also equivalent when p = 1. These and d)

are essentially due to Muckenhoupt except for the sharp constant. We use a
different approach. The sharp constant in d) is due to Buckley [Bu] however
we follow here the approach by Lerner [L]. Later on we will see a further
improvement.
Proof:
Since w ∈ Ap for each cube Q and nonnegative function f , by Holders
inequality
 Z p  Z p Z
1 1 1/p −1/p
f dy w(Q) = fw w dy w(Q) ≤ [w]Ap f (y)p w(y)dy.
|Q| Q |Q| Q Q

Hence 1 1
M f (x) ≤ 2n M c f (x) ≤ 2n [w] p Mwc (f p )(x) p ,
Ap

and by the Besicovitchs lemma:

1 1 1 1
kM f k ≤ cn [w] p kMwc (f p ) p k = cn [w] p kMwc (f p )k p 1,∞
Lp,∞ (w) Ap Lp,∞ (w) Ap L (w)

1 Z 1/p
p p
≤ cn [w] f w dx
Ap
Rn

This yieds that a) implies b) implies c).

11
 To prove d) Since this theorem is so central we give another proof. we set
w(Q)  σ(3Q) p−1
Ap (Q) =
|Q| |Q|
then, we have using that

Ap (Q) ≤ 3np [w]Ap

and that for any x ∈ Q then Q ⊂ Q(x, 2`(Q)) ⊂ 3Q

 1
1
Z
1 |Q|  1
Z p−1  p−1
|f | = Ap (Q) p−1 |f |
|Q| Q w(Q) σ(3Q) Q
 Z 1
 p−1
0
1
1 −1 p−1
≤ 3np [w]Ap−1 c
M (f σ ) dx .
p
w(Q) Q σ

where Mσc is the weighted centered maximal function. Using again that

M f (x) ≤ 2n M c f (x)

we get
1
1
M f (x) ≤ 2n 3np [w]Ap−1 Mw Mσc (f σ −1 )p−1 w−1 (x) p−1
 c
p

We conclude using that both

kMwc kLp0 and kMσc kLpσ

w

are finite with constants uniformly in w. This follows from the Besicovitch
covering Lemma.
Another way: by dyadic, passing from dyadic to non dyadic by ”shifting”.
For the sharpness we consider n = 1 and 0 < ε < 1. Let

w(x) = |x|(1−ε)(p−1) .

It is easy to check that

1
1
[w]Ap−1 ≈
p
ε
and hence as in Buckley’s paper

f (y) = y −1+ε χ(0,1) (y)

12
Observe that:
1
kf kpLp (w) ≈
ε
To estimate now kM f kLp (w) we pick 0 < x < 1, hence
1 x
Z
1
M f (x) ≥ f (y) dy = f (x)
x 0 ε
and hence
1
kM f kLp (w) ≥ kf kLp (w)
ε
From which the rest follows easily.

4 A factorization theorem with control on

the constants

Muckenhoupt already observed in [M] that it follows from the definition

of the A1 weights that if w1 , w2 ∈ A1 , then
w = w1 w21−p ∈ Ap ,
and furthermore
[w]Ap ≤ [w1 ]A1 [w2 ]p−1
A1 (19)
In view of this Muckenhoupt conjectured that any Ap weight can be written
in this way. This theorem was P. Jones who proved the necessity of the
factorization theorem with a very difficult proof. The modern approach to
this theorem is based on completely different arguments and it is due to J. L.
Rubio de Francia. Here we give a proof of this result using sharp constants.
We will use a variation of the method of Rubio de Francia which appears in
[H].
It is also well known that the modern approach to this question uses
completely different path and it is due to J. L. Rubio de Francia as can be
found in [GrMF] where we remit the reader for more information about the
Ap theory of weights. Here we present a variation which appears in [H]. Here
we give a proof of this result using sharp constants. To be more precise we
have the following result.

13
Lemma 4.1 Let 1 < p < ∞ and let w ∈ Ap , then there are A1 weights
u, v ∈ A1 , such that
w = u v 1−p
in such a way that
1
[u]A1 ≤ cn [w]Ap & [v]A1 ≤ cn [w]Ap−1
p
(20)

Proof :
Following [H] we will be using the Rubio de Francia’s iteration scheme or
algorithm to our situation. Define
0
0 |f |p
S1 (f )p ≡ w1/p M ( ),
w1/p
and
1
S2 (f )p ≡ M (|f |p w1/p ),
w1/p
0 0
Observe that Si : Lpp (Rn ) → Lpp (Rn ) with constant
1/p
kSi kLpp0 (Rn ) ≤ cn [w]Ap i = 1, 2

by Buckley’s theorem.
0
Now, the operator S = S1 + S2 is bounded on Lpp (Rn ) with
1/p
kSkLpp0 (Rn ) ≤ cn [w]Ap .

Define the Rubio de Francia algorithm R by

∞
X 1 S k (h)
R(h) ≡ .
k=0
2k (kSkLpp0 (Rn ) )k
0 0
Observe that R is also bounded on Lpp (Rn ) . Now, if h ∈ Lpp (Rn ) is
fixed, R(h) ∈ A1 (S). More precisely
1/p
S(R(h)) ≤ 2 kSkLpp0 (Rn ) ≤ cn [w]Ap .

In particular R(h) ∈ A1 (Si ) i = 1, 2, with

1/p
Si (R(h)) ≤ cn [w]Ap R(h) i = 1, 2

14
Hence
0 p0 /p 0
M (R(h)p w−1/p ) ≤ cn [w]Ap R(h)p w−1/p
and
M (R(h)p w1/p ) ≤ cn [w]Ap R(h)p w1/p .
Finally, letting
0
u ≡ R(h)p w1/p & v ≡ R(h)p w−1/p
we have u, v ∈ A1 w = uv 1−p with
1
[u]A1 ≤ cn [w]Ap & [v]A1 ≤ cn [w]Ap−1
p

5 Improving the Ap theorem via the A∞ con-

stant
We know that the class of weights is increasing and hence it is natural to
define the A∞ class of weights as follows.
Definition 5.1 The A∞ class of weights is defined in a natural way by
A∞ := ∪p>1 Ap .
The A∞ class of weights shares a lot of interesting properties.
It is known that an equivalent way of defining the class A∞ is by means
of a the following quantity
Z Z 
expL
[w]A∞ := sup − w exp − − log w .
Q Q Q

This constant can be found in for instance in [GCRdF] and was introduced
by Hruščev in [Hr]. This definition is natural because all you have to do is to
let p → ∞ in the deifniition of the Ap constant. In fact we have by Jensen’s
inequality:
[w]expL
A∞ ≤ [w]Ap .
On the other hand in [HP] the authors use a “new” A∞ constant (which
was essentially introduced by Fujii in [Fujii] and rediscovered later by Wilson
in [Wil87]) which is more suitable.

15
Definition 5.2
Z
1
[w]A∞ := sup M (wχQ ) dx.
Q w(Q) Q

Observe that [w]Ap ≥ 1 by the Lebesgue differentiation theorem.

It is shown in [HP] that
Z
1
[w]A∞ := sup M (wχQ ) dx. (21)
Q w(Q) Q

In fact, it is shown in [HP] that there are examples showing that [w]A∞ is
much smaller than [w]expLA∞ .
There is another possible way of defying the A∞ class of weights. Indeed,
define Z
LlogL 1 w(y)
[w]A∞ := sup w(y) log(e + ) dy
Q w(Q) Q wQ
since because it can be shown that

[w]LlogL
A∞ ≤ 2n+1 [w]A∞

which follows from the following lemma.

Lemma 5.3 For any cube Q and any measurable function w,

Z Z
w n+1
w log(e + ) dx ≤ 2 M (wχQ ) dx, (22)
Q wQ Q

and hence if w ∈ A∞
[w]LlogL
A∞ ≤ 2n+1 [w]A∞

Inequality (22) plays an important role in many situations in particular

when dealing with local estimates. The essential idea behind the proof can
be traced back to the well known L log L estimate for M in [St] (see also
[Wil-LNM, p. 17, inequality (2.15)] for a different proof). A proof is supplied
in [HP].

16
5.1 The Reverse Hölder property
In the classical situation, if w ∈ A∞ , then there is a constant r > 1 such that
 Z 1/r Z
1 r c
w ≤ w
|Q| Q |Q| Q

Observe that the other inequality is immediate by Jensen’s inequality. This

property has played a central role in the area. The standard proofs show
that there is a bad dependence on the constant c = c(r, [w]A1 ). To improve
our results we need a more precise estimate.
The main result is the following

Theorem 5.4 (Sharp Reverse Hölder Inequality)

a) Let w ∈ A∞ and let Q0 be a cube. Then
Z 1/r(w) Z Z
r(w)
− w ≤ 2− w ≤ 2− w dx,
Q Q Q0

where r(w) = cd [w]1 A .

∞
b) Furthermore, the result is optimal up to a dimensional factor: If a
weight w satisfies the reverse Hölder inequality
Z 1/r Z
r
− w ≤ K− w,
Q Q

then [w]A∞ ≤ cn · K · r0 .

This result was obtained in [HP] (using Sawyer’e discretization method)

and has been improved in [HPR] with a much simpler proof. A previous
reverse Holder’s inequality in this spirit was obtained in [LOP1] but for A1
weights.
Part b) follows from the boundedness of the maximal function in Lr with
constant cn r0 :
Z Z 1/r
r
− M (χQ w) ≤ − M (χQ w)
Q Q
Z 1/r Z
0 r 0
≤ cn · r − w ≤ cn · r · K− w.
Q Q

17
 In this section we give a different proof of Theorem 5.4. The constant is
a slightly worst but the method is more direct and may be of interest. The
main idea is to linearize the maximal function following Sawyer’s well know
method but instead at global level at local one.
The main use of the Lemma is the following key observation that we
borrow from [Wil-LNM], p. 45:

Lemma 5.5 Let S ⊂ Q and let λ > 0, then

|S| w(S) 2d+2 [w]A∞

< e−λ implies < + e−λ/2 (23)
|Q| w(Q) λ

proof: Indeed, if Eλ = {x ∈ Q : w(x) > eλ hwiQ } then w(Eλ ) ≤

2d+1
λ
[w]A∞ w(Q) by (22). Therefore:

2d+2 [w]A∞
w(S) ≤ w(S ∩ Eλ/2 ) + w(S \ Eλ/2 ) ≤ w(Q) + eλ/2 hwiQ |S|
λ
2d+2 [w]A∞
≤ w(Q) + eλ/2 e−λ w(Q) by the hypothesis in (23)
λ
2d+2 [w]A∞
= w(Q) + e−λ/2 w(Q)
λ
and this proves the claim (23).
2

proof [of Theorem 5.4] Recall that we have to prove that

Z 1/r(w) Z
r(w)
− w ≤ 2− w.
Q Q

where
1
r(w) := 1 + ,
τd [w]0A∞
and where τd is a large dimensional constant. R
Observe that by homogeneity we can assume that −Q w = 1. We use the

18
dyadic maximal function on the dyadic subcubes of a given Q:
Z Z Z ∞
1+ 
w ≤ Md (wχQ ) w = t−1 w({x ∈ Q : Md (wχQ ) > t}) dt.
Q Q 0
Z1 Z ∞
−1 dt
≤ t w(Q) dt +  t w({x ∈ Q : Md (wχQ ) > t})
0 1 t
XZ ak+1
dt
≤ |Q| +  t w({x ∈ Q : Md (wχQ ) > t})
k≥0 ak t
Z ak+1

X
k dt
≤ |Q| + a a w({x ∈ Q : Md (wχQ ) > ak }) , for a  1,
k≥0 ak t
X
= |Q| + a log a ak w(Ωk )
k≥0

where
Ωk = {x ∈ Q : Md (wχQ (x) > ak }.
R
Since ak ≥ 1 = −Q w we can consider the Calderón–Zygmund decomposi-
tion w adapted to Q. There is a family of maximal S non-overlapping dyadic
cubes {Qk,j } strictly contained in Q for which Ωk = j Qk,j and
Z
a < − w ≤ 2d ak .
k
(24)
Qk,j

Now,
Z !
X X X 1
ak w(Ωk ) = ak w(Qk,j ) ≤ w(y) dy w(Qk,j )
k≥0 k,j k,j
|Qk,j | Qk,j

We now estimate w(Qk,j ), for each (k, j) we set Ek,j = Qk,j \ Ωk+1 . Ob-
serve that the sets of the family Ek,j are pairwise disjoint. We claim that for
a > 2n and for each k, j:
a
|Qk,j | < |Ek,j |. (25)
a − 2n
We now apply (23) with Q = Qk,j and S = Qk,j ∩ Ωk+1 . Choose λ such
n
that e−λ = 2a , namely λ = log 2an . Then applying (23) we have that
w(Qk,j ∩ Ωk+1 ) 2d+2 [w]A∞ 2n 1/2
< + ( ) .
w(Qk,j ) log 2an a

19
Since a > 2n is available we choose a = 2d eL [w]A∞ , with L a large dimensional
constant to be chosen. If in particular L ≥ 2d+4 we have

w(Qk,j ∩ Ωk+1 ) 2d+2 1 1 1

< + e− [w]A∞ L/2 < + =
w(Qk,j ) L 4 4 2

This yields that w(Qk,j ) ≤ 2w(Ek,j ) and we can continue with the sum
estimate:
Z !
X X 1
ak w(Ωk ) ≤ 2 w(y) dy w(Ek,j )
k≥0 k,j
|Qk,j | Qk,j

XZ Z

≤2 Md (wχQ ) wdx ≤ 2 Md (wχQ ) wdx
k,j Ek,j Q

Combining estimates we end up with

Z Z
− Md (wχQ ) w ≤ 1 + 2 a log a − Md (wχQ ) wdx
 
Q Q

for any  > 0. Recall that a = 2d eL [w]A∞ . Hence if we choose

1 1
L = 2d+4 , = = ,
27 L [w] A∞ 211+d [w]A∞
we can compute
Z
 1
2 a log a < , − Md (wχQ ) w ≤ 2,
2 Q

concluding the proof of the theorem.

5.2 The open property

As a corollary we deduce the following useful result.

20
Corollary 5.6 (The Precise Open property) Let 1 < p < ∞ and let
w ∈ Ap . Then w ∈ Ap− where
p−1 p−1
= 0
=
rσ 1 + τn [σ]A∞
0
where as usual σ = w1−p . Furthermore

[w]Ap−ε ≤ 2p−1 [w]Ap

Proof
Since w ∈ Ap σ ∈ Ap0 ⊂ A∞ , and hence
Z Z  p−1 Z  Z p−1
rσ rσ
− w − σ ≤ − w 2− σ .
Q Q Q Q
p−1 p−1
Choose ε so that rσ
= p − ε − 1, namely  = 0 .
rσ
Observe that ε > 0 and
p − ε > 1.

5.3 Improving Muckenhoupt-Buckley’s theorem

Now we prove one of the main result.

Theorem 5.7 Let p > 1.

1/p
kM kLp (w) ≤ cn p0 [w]Ap [σ]A∞ , (26)
1
where 1 < p < ∞, σ = w 1−p .

Proof: We will use the weak type estimate (15) and

{x ∈ Rn : M f (x) > 2t} ⊂ {x ∈ Rn : M ft (x) > t}

where ft = f χf >t . Now, since w ∈ Ap by ”The Precise Open property”

p−1
Corollary 5.6 if w ∈ Ap− , where  = r(σ)0 and [w]Ap−ε ≤ 2p−1 [w]Ap .
Then
Z Z ∞ Z ∞
p p n dt p dt
(M f ) wdx = p t w{y ∈ R : M f (x) > t} = p2 tp w{y ∈ Rn : M f (x) > 2t}
Rn 0 t 0 t

21
 ∞ Z ∞ Z
ft (y)p−ε
Z
p p n dt p−ε p dt
≤ p2 t w{y ∈ R : M ft (x) > t} ≤ p cn [w]Ap−ε 2 tp p−ε
w(y)dy
0 t 0 Rn t t
Z Z f (y) Z
dt 1
≤ p cp−ε
n 2p−1 [w]Ap 2p tε f (y)p−ε w(y)dy = 22p−1 p cp−ε
n [w]Ap f (y)p w(y)dy
Rn 0 t ε Rn

r(σ)0
Z Z
1 + τn [σ]A∞ 2p−1 p−ε
= p 22p−1 cp−ε
n [w]p p
f (y) w(y)dy = p2 cn [w]p f (y)p w(y)dy
p−1 Rn p−1 Rn

and this yields (26) .

Of course we recover here part d) of Theorem (3.2):

1
kM kLp (w) ≤ cp0 [w] p−1
Ap

since 1
[σ]A∞ ≤ [σ]Ap0 = [w]Ap−1
p

6 The sharp extrapolation theorem

This result was proved in [DGPP] but it is found [CMP3] a much simpler
proof. (see also [D]).

Theorem 6.1 Let 1 < p0 < ∞, and α > 0 and let T be any operator such
that for some c > 0
kT k p0 ≤ c [w]αA . (27)
L (w) p0

Then, for 1 < p < p0 there is universal c > 0 such that

p0 −1
α max{1, p−1
}
kT k ≤ c [w]Ap . (28)
Lp (w)

Proof. Let 1 < p < p0 . For such a p and w ∈ Ap perform the iteration
algorithm R as follows:
∞
X 1 M k (h)
R(h) =
k=0
2k kM kkLp (w)

22
Then we have
(A) h ≤ D(h)
(B) kR(h)kLp (w) ≤ 2 khkLp (w)
1
(C) [R(h)]A1 ≤ 2 kM kLp (w) ≤ cn p0 [w]Ap−1 p
by Theorem 3.2, d).
Then
Z 1/p
p −(p0 −p) pp (p0 −p) pp
kT (f )kLp (w) = |T f | (Rf ) 0 (Rf ) 0 wdx
Rn

Z 1/p0  Z 0 −p
 ppp
p0 −(p0 −p) p 0
≤ |T f | (Rf ) wdx (Rf ) wdx
Rn Rn
Z 1/p0 p0 −p

≤ c [R(f )−(p0 −p) w]αAp0 |f |p0 (Rf )−(p0 −p) wdx kf kLpp0(w)
Rn
Z 1/p0 p0 −p
−(p0 −p)
≤ c [R(f ) w]αAp0 |f |p wdx kf kLpp0(w)
Rn

= c [R(f )−(p0 −p) w]αAp0 kf kLp (w)

To conclude we claim the following:
p0 −1

[R(f )−(p0 −p) w]Ap0 ≤ [w]Ap−1

p

Indeed for any cube and the definition of A1

Z  1 Z −(p0 −p) 1 Z
1 −(p0 −p) p0 −p
R(f ) wdx ≤ [Rf ]A1 R(f ) dx wdx
|Q| Q |Q| Q |Q| Q
−1 0 −1
and also if we set q = pp00 −p > 1 and observe that q 0 = pp−1 . Then, Holder’s
gives

1
Z  1−p00 1
Z p0 −p 0
−(p0 −p)
R(f ) w dx = R(f ) p0 −1 w1−p0 dx
|Q| Q |Q| Q
 1 Z  pp0 −p  1 Z 0
 pp−1
0 −1 −1
≤ R(f ) dx w1−p dx 0
|Q| Q |Q| Q
Hence, combining we have

23
 p0 −p p0 −1

[R(f )−(p0 −p) w]Ap0 ≤ c [Rf ]pA01−p [w]Ap ≤ c kM kpL0p−p p−1

(w) [w]Ap ≤ [w]Ap [w]Ap = [w]Ap−1
p

Let p > p0 . For such a p and w ∈ Ap perform the iteration algorithm R0

as follows: ∞
0
X 1 (M 0 )k (h)
R (h) =
k=0
2k kM 0 kkLp (w)
M 0 (f w)
where M 0 is the operator defined by w
. Observe that by Theorem 3.2
1
p0 −1
part d) kM 0 kLp (w) = kM kLp (σ) ≤ cn p0 [σ]Ap0 = cn p0 [w]Ap . Then it is easy
to verify that
(A’) h ≤ R0 (h)
(B’) kR0 (h)kLp0 (w) ≤ 2 khkLp0 (w)

(C’) [R0 (h)w]A1 ≤ 2 kM 0 kLp0 (w) ≤ cn p [w]Ap Now by duality,

Z
kT (f )kLp (w) = sup |T f |hwdx
h
0
for any non-negative h ∈ Lp (w) such that kf kLp (w) = 1. Now we set W =
p−p0
(R0 (h)) p−1 w and claim that W ∈ Ap0 . Assuming the claim for the time
being we use Holder’s inequality together with the extrapolation hypothesis
to derive Z p−p0 (p0 −1)p
|T f |p (R0 h) p0 (p−1) (R0 h) p0 (p−1) wdx
Rn
Z 1/p0  Z  p1
p0 0 p0
≤ |T f | W dx (R h) wdx 0
Rn Rn
Z p−p0
1/p0
≤ cp0 ,T, [W ]αAp0 |f |p0 (R0 h) p−1 dx
Rn
Z 1/p p−p0
0
≤ cp0 ,T, [W ]αAp0 |f | wdxp p0 (p−1)
|R (h)kLp0 (w)
Rn
Z 1/p
α
≤ cp0 ,T, [W ]Ap0 |f |p wdx .
Rn
To conclude we claim the following:

[W ]Ap0 ≤ cn,p,p0 [w]Ap

24
 We now estimate [W ]Ap0 : Indeed by Holder’s inequality with exponent
p−1 (p/p0 )0
q = p−p0
= p0
> 1 and q 0 = pp−1
0 −1
.
0 −1
 pp−1
Z Z  1 Z 1/q  1 Z
1 1 0
p−p0
0
W dx = (R (h)) p−1 wdx ≤ R (h) wdx wdx
|Q| Q |Q| Q |Q| Q |Q| Q

but on the other hand by the A1 property from (C’)

 1 Z p0 −1  1 Z p−p p0 −1
1−p00 0 − (p−1)(p0 −1) 1−p00
W dx = (R (h)) 0 w
|Q| Q |Q| Q
p−p0
 1 Z − p−p 0
1
Z p−p0 p0 −1
0 p−1 0 p−1 1−p00
≤ [R (h)w]A1 R (h)w w (p−1)(p0 −1)
w dx
|Q| Q |Q| Q
p−p0 
1
Z − p−p
p−1
0
1
Z
0
p0 −1
0 0
p−1
= [R (h)w]A1 R (h)w w1−p dx
|Q| Q |Q| Q
Combining estimates we have and using again condition (C’)
p−p0  0 −1 
 pp−1
Z  1 Z p0 −1 Z Z p0 −1
1 1−p00 0 1 1 0
W dx W dx p−1
≤ [R (h)w]A1 wdx w1−p dx
|Q| Q |Q| Q |Q| Q |Q| Q
p−p0 p0 −1

≤ [R0 (h)w]Ap−1
1
[w]Ap−1
p
≤ c [w]Ap .

Applications to Singular Integrals, commutators, square functions, vector-

valued maximal function and other operators can be found in [Hy], [HL],
[HLP], [HP], [CMP1], [CMP2], [ChPP].

7 Two weight problem: sharp Sawyer’s the-

orem
Recall Sawyer’s theorem [S]:

Let 1 < p < ∞, then there is a finite C

kM (f σ)kLp (u) ≤ C kf kLp (σ) (29)

25
 if and only if there is a finite constant K such that for any cube
Q
Z 1/p
p
|M (σ χQ )| u dx ≤ K σ(Q)1/p
Q

7.1 Moen version

Kabe Moen version of Sawyer’s theorem: let
R 1/p
p
Q
|M (σ χQ )| u dx
[u, σ]Sp = sup
Q σ(Q)1/p

Theorem 7.1 [M] Let 1 < p < ∞ and let kM k the smallest of the constants
C in (29). Then
[u, σ]Sp ≤ kM k ≤ cn p0 [u, σ]Sp

Very recently, with E. Rela we have found an application of this result

and as a particular case we obtain the following ([PR]).

Corollary 7.2
kM k ≤ cn p0 ([u, σ]Ap [σ]A∞ )1/p (30)

Proof (skecth): Let a > 1 a universal parameter

R to be chosen. Fix a
k0 k0 +1
cube Q and let k0 an integer such that a ≤ Q σ < a
− . Then
Z Z ∞ Z
X
p p
|M (σ χQ )| u dx = |M (σ χQ )| u dx+ |M (σ χQ )|p u dx
Q M (σ χQ )≤ak0 k=k0 ak <M (σ χ Q )≤ak+1

Z ∞
X
≤ (− σ)p u(Q) + ap akp u({x ∈ Q : M (σ χQ ) > ak })
Q k=k0
k
Rbut by the standard local Calderón–Zygmund covering lemma since a >
− σ, k ≥ k0 we can find dyadic relative with respect to Q subcubes {Qk,j }
Q
such that for each of these k

{x ∈ Q : M (σ χQ ) > ak } = ∪k,j Qkj

26
and Z
a < − σ ≤ 2n ak
k
j∈Z
Qkj

Furthermore the family of cubes satisfies the sparness property. Then using
the Ap condition we continue with t
!p !p−1 Z
X Z X Z
≤ [u, σ]Ap σ(Q)+ap − σ dx u(Qkj ) ≤ [u, σ]Ap σ(Q)+ap − σ dx − u dx σ(Qkj )
k,j Qkj k,j Qkj Qkj

X
≤ [u, σ]Ap σ(Q) + cn [u, σ]Ap σ(Qkj )
j,k

Now by the sparness property of the cubes we can continue last sum with
XZ XZ Z
k
≤ cn − σ dx |Ej | ≤ cn M (χQ σ) dx ≤ cn M (χQ σ) dx ≤ cn [σ]A∞ σ(Q)
j,k Qkj j,k Ejk Q

since the family Ejk is pairwise disjoint and all of them contained in Q.
Putting together we are done.
2

References
[Bu] S.M. Buckley, Estimates for operator norms on weighted spaces and
reverse Jensen inequalities, Trans. Amer. Math. Soc., 340 (1993),
no. 1, 253–272.

[ChPP] D. Chung, M.C. Pereyra and C. Pérez Sharp bounds for general
commutators on weighted Lebesgue spaces, Trans. Amer. Math. Soc.
364 (2012), 1163-1177.

[CR] R. Coifman and R. Rochberg, Another charaterzation of B.M.O.,

Proc. AMS. 79 (1980), 249–254.

[CMP1] D. Cruz-Uribe,SFO, J.M. Martell, C. Pérez, Sharp weighted esti-

mates for classical operators, Advances in Mathematics, 229, (2012),
408-441.

27
[CMP2] D. Cruz-Uribe,SFO, J.M. Martell, C. Pérez, Sharp weighted esti-
mates for approximating dyadic operators, Electronic Research An-
nouncements in the Mathematical Sciences, 17 (2010),12-19.
[CMP3] D. Cruz-Uribe, J.M. Martell and C. Pérez, Weights, Extrapola-
tion and the Theory of Rubio de Francia, book, Series: Op-
erator Theory: Advances and Applications, Vol. 215, Birkauser,
Basel, (http://www.springer.com/mathematics/analysis/book/978-
3-0348-0071-6).
[dGuz] M. de Guzmán, Differentiation of Integrals in Rn , Lect. Notes Math.
481, Springer Verlag, (1975).
[DGPP] O. Dragičević, L. Grafakos, C. Pereyra, S. Petermichl Extrapola-
tion and sharp norm estimates for classical operators on weighted
Lebesgue spaces, Publ. Mat. 49 (2005), no. 1, 73–91.
[D] J. Duoandikoetxea, Extrapolation of weights revisited: new proofs
and sharp bounds, J. Funct. Anal., 260 (2011), 1886-1901.
[FS] C. Fefferman and E.M. Stein, Some maximal inequalities, Amer. J.
Math. 93 (1971), 107-115.
[Fujii] Nobuhiko Fujii, Weighted bounded mean oscillation and singular in-
tegrals, Math. Japon. 22 (1977/78), no. 5, 529–534.
[GrMF] L. Grafakos, Modern Fourier Analysis, Springer-Verlag, Graduate
Texts in Mathematics 250, Second Edition, (2008).
[GCRdF] J. Garcı́a-Cuerva and J.L. Rubio de Francia, Weighted Norm In-
equalities and Related Topics, North Holland Math. Studies 116,
North Holland, Amsterdam, 1985.
[H] E. Hernández, Fatorization and extrapolization of pairs of weights,
Studia Math, 95 (1989), 179-193.
[Hr] Sergei V. Hruščev, A description of weights satisfying the A∞ con-
dition of Muckenhoupt, Proc. Amer. Math. Soc. 90 (1984), no. 2,
253–257.
[Hy] T. Hytönen, The sharp weighted bound for general Calderón-
Zygmund operators, Annals of Math. 175 (2012), no. 3, 1473–1506.

28
[HL] T. Hytönen and M. Lacey, The Ap − A∞ inequality for general
Calderón-Zygmund operators, Indiana Univ. Journal of Math. (to
appear).
[HLP] T. Hytönen, M. Lacey and C. Pérez, Sharp weighted bounds for the q-
variation of singular integrals, Bulletin London Math. Soc. 45 (2013)
529-540.
[HP] Tuomas Hytönen and Carlos Pérez, Sharp weighted bounds
involving A∞ , Analysis & PDE 6 (2013), 777–818. DOI
10.2140/apde.2013.6.777.
[HPR] T. Hytönen, C. Pérez and E. Rela, Sharp Reverse Hölder property
for A∞ weights on spaces of homogeneous type, Journal of Functional
Analysis 263, (2012) 3883–3899.
[Jo] J.-L. Journé, Calderón–Zygmund Operators, Pseudo-Differential
Operators and the Cauchy Integral of Calderón, Lecture Notes in
Mathematics, 994, Springer Verlag, New York, 1983.
[L] A.K. Lerner, An elementary approach to several results on the Hardy-
Littlewood maximal operator, Proc. Amer. Math. Soc. 136 (2008) no
8, 2829-2833.
[LOP1] A. Lerner, S. Ombrosi and C. Pérez, Sharp A1 bounds for Calderón-
Zygmund operators and the relationship with a problem of Mucken-
houpt and Wheeden, International Mathematics Research Notices,
2008, 6, Art. ID rnm161, 11 pp. 42B20.
[LPR] T. Luque, C. Pérez and Ezequiel Rela, Optimal exponents in weighted
estimates without examples, to appear Mathematical Research Let-
ters.
[Ma] P. Mattila, Geometry of sets and Measures in Euclidean spaces,
Cambridge University Press, Cambridge, (1995).
[M] K. Moen, Sharp one-weight and two-weight bounds for maximal op-
erators, Studia Math. 194 (2009), 163-180
[Mu] B. Muckenhoupt, Weighted norm inequalities for the Hardy–
Littlewood maximal function, Trans. Amer. Math. Soc. 165 (1972),
207–226.

29
[PR] C. Pérez and Ezequiel Rela, A new quantitative two weight theorem
for the Hardy-Littlewood maximal operator, to appear Proceedings
of the American Mathematical Society.

[S] E. T. Sawyer, A characterization of a two weight norm weight in-

equality for maximal operators, Studia Math. 75 (1982), 1–11.

[St] E. M. Stein, Note on the class L log L, Studia Math. 32 (1969),

305-310.

[Wil87] J. Michael Wilson, Weighted inequalities for the dyadic square func-
tion without dyadic A∞ , Duke Math. J. 55 (1987), no. 1, 19–50.

[Wil-LNM] Michael Wilson. Weighted Littlewood-Paley theory and

exponential-square integrability, volume 1924 of Lecture Notes in
Mathematics.

30