Download as pdf or txt
Download as pdf or txt
You are on page 1of 31

Balanced measures, sparse domination and complexity-dependent weight

classes

José M. Conde Alonso, Jill Pipher, and Nathan A. Wagner


arXiv:2309.13943v1 [math.CA] 25 Sep 2023

Abstract. We study sparse domination for operators defined with respect to an atomic filtration on a
space equipped with a general measure µ. In the case of Haar shifts, Lp -boundedness is known to require
a weak regularity condition, which we prove to be sufficient to have a sparse domination-like theorem. Our
result allows us to characterize the class of weights where Haar shifts are bounded. A surprising novelty is
that said class depends on the complexity of the Haar shift operator under consideration. Our results are
qualitatively sharp.

Introduction
Probabilistic techniques and methods are ubiquitous in harmonic analysis and especially so in Calde-
rón-Zygmund theory. Often these methods make use of the filtration of σ-algebras generated by the dyadic
system of cubes
[ 
D := Dk = 2−k [j1 , j1 + 1) × [j2 , j2 + 1) × . . . [jn , jn + 1) : j ∈ Zn .
k∈Z

Probabilistic stopping-time arguments have a long history as a critical tool in analysis; for example, in the
Calderón-Zygmund decomposition, which is used to prove a weak-type estimate for singular integrals. Some
fundamental insights into the close connection between certain function spaces and operators in analysis and
their probabilistic counterparts were established in: [12], which explored the relationship between BMO and
dyadic BMO, and later in the dyadic representation theorems of [14, 24]. In particular, the main result in
[24] recovers the Hilbert transform as an average of dyadic operators:
ˆ
(HRep) Hf (x) = c0 XD ω f (x)dP (ω),

where H is the Hilbert transform, (Ω, P ) is a probability space, D ω is a different dyadic system for each
value of ω and the dyadic Hilbert transform is the operator
X
(HDyad) XD ω f (x) = hf, hI i(hI− (x) − hI+ (x)),
I∈D ω

where hI = 2−1 (|I− |−1/2 1I− − |I+ |−1/2 1I+ ) denotes the Haar function associated with the interval I, and
I− (resp. I+ ) is the left (resp. right) dyadic child of I. The main result in [14] expands on (HRep) by
showing that any Calderón-Zygmund operator on Rn can be written as an average of higher-dimensional
generalizations of X.
Once it has been established a continuous operator can be recovered from probabilistic counterparts,
one can often more readily derive quantitative consequences, like weighted estimates. One efficient way
of doing so is the sparse domination technique, initiated in [16] and applied to get pointwise estimates
for Calderón-Zygmund operators via estimates for the dyadic objects that represent that operator [8, 17].
Sparse domination is not the only possible path: direct estimates are possible, like those in [15] or using the
one-third trick as in [19].
The setting of the results discussed above is Rn , equipped with the Lebesgue measure, which is a natural
one for the problems under study. However, other measures µ on Rn also appear naturally, arising from
1
2 JOSÉ M. CONDE ALONSO, JILL PIPHER, AND NATHAN A. WAGNER

questions in geometry or partial differential equations. For example, the class of representing measures
associated to elliptic or parabolic divergence form operators are all “doubling". A measure µ is said to be
doubling if there exists a constant cµ such that for all balls B centered at its support, there holds
(Db) µ(2B) ≤ cµ µ(B).
Another important class of measures is the class of measures of polynomial growth. These may fail (Db).
Despite the lack of the doubling property, using mainly geometric techniques, a very complete Calderón-
Zygmund theory has been developed for measures satisfying the polynomial growth condition in [22, 23,
26, 27]. These developments led to profound applications, among which are: the solution of the Painlevé
problem [28], the n − 1 case of the David-Semmes problem [21], and the characterization of the rectifiability
of harmonic measure [1].
In martingale language, (Db) is closely related to the regularity of the filtration. Regularity of the
b for all
filtration is equivalent, in the case of the dyadic filtration on (Rn , µ), to the condition µ(Q) ∼ µ(Q)
cubes in D, where Q̂ is the “parent" dyadic cube of Q. Probabilistic tools are usually versatile enough to be
applicable to σ-finite filtered spaces with highly non-regular structure, where the most important martingale
inequalities hold true [3, 10]. Therefore, one might be tempted to think that the techniques and results
described above are still powerful even when the geometric and measure theoretic properties of the ambient
space are not tightly coupled, even if (Db) fails. However, when the classical doubling condition of the
underlying measure fails, the situation becomes significantly more complicated. The one-third trick is no
longer useful in the usual way, and there are limited alternatives available [5]. Moreover, direct arguments of
sparse domination for Calderón-Zygmund operators [7, 30] yield quantitative estimates that seem far from
tight. To start, just as in the doubling setting, they do imply weighted inequalities for Calderón-Zygmund
operators. Recall that a weight is an a.e. positive function w ∈ L1loc (dµ) that we identify with the measure
dν = wdµ. The results in [7, 30] imply that a Calderón-Zygmund operator is bounded on L2 (wdµ) whenever
w ∈ A2 (µ), that is, the following holds:
w(B)w−1 (B)
(A2) [w]A2 (µ) := sup < ∞.
B ball µ(B)2
However, in [29], the necessary and sufficient condition that weights must satisfy so that Calderón-Zygmund
operators are bounded on L2 (wdµ), termed Z2 (µ), was shown to be strictly weaker than the classical A2 (µ)
condition. Moreover, the issue of a representation theorem along the lines of (HRep) or [14] remained
wide open. In that direction, the study of the operators that arise in the representation formulas is an
interesting problem that can inform what, if any, representation theorems one can expect when the measure
under consideration is not Lebesgue. Exploring the feasibility of extending this program to the non-doubling
setting was one of the motivations of this work. We begin this study on R, and point out that there will be
additional considerations needed to extend the forthcoming results to higher dimensions.

Sparse domination for Haar shifts with respect to µ. Let µ be a Radon, atomless measure on
R. We use the standard notation for dyadic ancestors: I (j) is the interval from D that contains I and has
sidelength equal to 2j ℓ(I). We denote
j
[
(j)
Dj (I) = {J ∈ D : J = I}, and also D≤j (I) = Dk (I).
k=0

Given an I ∈ D, the Haar function associated to I is constant on the dyadic children of I, normalized in
L2 (µ), and has mean value zero:
s  
µ(I− )µ(I+ ) 1I− (x) 1I+ (x)
hI (x) = − .
µ(I) µ(I− ) µ(I+ )

Given nonnegative integers s and t, a Haar shift of complexity (s, t) is an operator of the form
X X X
T s,t f (x) = αIJ,K hf, hJ ihK ,
I∈D J∈Dr (I) K∈Ds (I)
SPARSE DOMINATION FOR BALANCED MEASURES 3

where the pairing h·, ·i (and integral averages h·iI ) are taken with respect to µ:
1
ˆ ˆ
hf, gi = f (x)g(x)dµ(x), hf iQ = f (x)dµ(x).
Rn µ(Q) Q
In this language, X is an operator of complexity (0, 1). If α ∈ ℓ∞ , then T s,t is bounded on L2 (µ) regardless
of the properties of µ. We will always assume that kαkℓ∞ ≤ 1. In [18], a condition is identified that
characterizes Lp (µ) and weak type (1, 1) boundedness of T s,t when r and s are nonzero: if
µ(I− )µ(I+ )
m(I) = ,
µ(I)
then T s,t is of weak type (1, 1) if and only if, for all I ∈ D,
(Balance) m(I) ∼ m(I),b
where Ib denotes the dyadic parent of I. When (Balance) holds and µ is free of atoms, we will say that µ
is balanced 1. The first question that we address in this paper is whether the operators T s,t satisfy sparse
bounds for balanced measures µ. In the case of the Lebesgue measure, this is contained in [8, 14]. We first
consider estimates in the dual sense: we look for an inequality like
X
(ClasSp) hT s,t f1 , f2 i . hf1 iQ hf2 iQ µ(Q) =: AS (f1 , f2 ),
I∈S

for nice enough f1 and f2 and a family S = S(f1 , f2 ) ⊂ D which is sparse in the usual sense: for each I ∈ S,
there exists EI ⊂ I such that µ(EI ) ∼ µ(I) and the family {EI }I∈S is pairwise disjoint. The question is
natural, since sparse domination methods are usually well adapted to dyadic operators. Notably, the simple
argument in [9] might be expected to go through in the balanced case once the classical Calderón-Zygmund
decomposition employed there is replaced by the one in [18], (or the streamlined version from [4]). In this
direction, our first result is a negative one: we shall show that the classical sparse domination inequality,
(ClasSp), must fail, by constructing a balanced measure µ on the interval [0, 1] so that X fails sparse
domination. Of course, similar examples can be constructed to disprove sparse domination in general for
higher complexity dyadic shifts.
Modified sparse forms. The above discussion shows that sparse domination, in the classical formu-
lation, must fail for some balanced measures µ. To remedy that, we propose a variant of the sparse form in
the following way: X X
CS (f1 , f2 ) := [hf1 iI hf2 iJ + hf2 iI hf1 iJ ] m(I).
I∈S J∈S∩D≤2 (I)
b

The role of the second sum in the definition of CS takes into account the interaction of intervals I and its
dyadic neighbors, which is necessary for the sparse domination argument to work when the complexity is
nonzero. On the other hand, replacing µ(I) by m(I) on the right hand side helps one to obtain quantitative
bounds for CS , since we have m(I) ≪ µ(I) in general. With this definition, the result that we obtain reads
as follows:
Theorem A. Let µ be balanced, and let T be a Haar shift of complexity (s, t) with s + t ≤ 1. Then for
each pair of compactly supported, bounded nonnegative functions f1 , f2 there exists a sparse collection S ⊂ D
such that
|hT f1 , f2 i| . AS (f1 , f2 ) + CS (f1 , f2 ).
We postpone the full statement of Theorem A for Haar shifts of higher complexity to Section 2. As is
standard, the proof of Theorem A uses the weak (1, 1) bound for the maximal operator M associated to
the dyadic grid D. On the other hand, both forms AS and CS satisfy the expected weak-type and Lp (µ)
estimates, which is evidence that this definition of the sparse forms is the right one in the context of balanced
measures. The broad scheme of the proof of theorem A is similar to arguments that were used in the context
of doubling measures like [2, 6], and is most similar to [9]. However, we need to use a Calderón-Zygmund
decomposition adapted to general measures, originally inspired by Gundy’s decomposition for martingales
[4, 18]. An additional feature that we need to use is a BMO estimate for the good part of the decomposition,
which is a technical novelty key to our approach.
1In [18], the authors term these measures m-equilibrated, but we prefer the shorter term balanced.
4 JOSÉ M. CONDE ALONSO, JILL PIPHER, AND NATHAN A. WAGNER

A weighted theory for balanced measures. The main applications of sparse bounds are weighted
estimates. The classical sparse domination yields quantitatively sharp estimates for doubling measures. In
the case of the Lebesgue measure, both Calderón-Zygmund operators and Haar shifts are bounded on L2 (w)
if and only if w ∈ A2 . The corresponding condition with respect to a general measure µ is just (A2),
where we should take the sup with respect to all Q ∈ D in the Haar shift case (instead of over balls).
As we mentioned before, in Calderón-Zygmund case the A2 (µ) condition is not necessary in general, but
it is certainly sufficient [29]. In the dyadic setting, the situation is different: we can show that the A2 (µ)
condition is not sufficient for the boundedness of X, by constructing a pair (µ, w) where µ is balanced,
w ∈ A2 (µ) and X is not bounded on L2 (wdµ). Our main weighted result is the identification of the right
class of weights that governs the boundedness of Haar shifts. A notable feature of this identification is the
fact that it takes into account the higher complexity of the operators that we consider. Moreover, unlike the
sparse form, the weight class is the same for all operators of complexity greater than or equal to 1.
Theorem B. Let µ be balanced. There exists a collection of weights Ab2 (balanced A2 ) with the following
properties:
• Ab2 ⊆ A2 (µ).
• If (s, t) 6= (0, 0), all operators T s,t are bounded on L2 (wdµ) if and only if w ∈ Ab2 .
Moreover, there exists a balanced, nondoubling measure µ so that Ab2 ( A2 (µ).
We postpone, until section 3, the precise definition of the class Ab2 (µ), which can be described as a variant
of (A2) involving averages over dyadic intervals I, J where J ∈ D≤2 (I). b As one would expect, in case µ is
doubling, we have that Ab2 (µ) = A2 (µ). Our weighted results can be readily generalized to Lp (wdµ), with
a straightforward modification of our A2 -type class. Finally, we have a maximal function characterization.
We formulate it by identifying a family of variants of the usual dyadic maximal function, that we denote
by Mj . Mj is Lp (µ)-bounded and weak-type (1, 1), and satisfies the same sparse domination as the Haar
shifts T s,t whenever r + s = j. We then use each Mj to define a family of Ap -type classes of weights: we
say that w ∈ Aj1 (µ) if Mj w(x) . w(x), while we say that w ∈ Ajp (µ) if Mj : Lp (wdµ) → Lp (wdµ). Even
though the operators Mj are not equivalent, the weight classes that they determine are the same; that is,
Ajp (µ) = Abp (µ) for all j ≥ 1.
The rest of the paper is organized as follows: Section 1 contains an example that shows that sparse
domination in the usual sense must fail for Haar shifts and some balanced measures, and another one of
w ∈ A2 (µ) |with µ balanced| so that X is not bounded on L2 (wdµ). Section 2 contains the proof of
Theorem A and our sparse domination result for general complexity shifts on R. Section 3 contains the
weighted theory, including the proof of Theorem B.

Acknowledgements. The authors would like to thank Sergei Treil for several valuable discussions. The
first named author has been supported by grants CNS2022-135431 and RYC2019-027910-I (Ministerio de
Ciencia, Spain). The third named author is supported by grant DMS-2203272 (National Science Foundation).

1. Failure of usual sparse domination


We shall work with Borel measures µ on R that we will always assume to be balanced. This implies for
us that µ has no atoms -which we shall need in order to use the Carleson packing characterization of the
sparse condition below-. Fix 0 < η < 1. A collection S ⊂ D is called η-sparse (or just sparse) if for each
I ∈ S there exists EI ⊂ I so that EI ∩ EI ′ = ∅ if I 6= I ′ and µ(EI ) ≥ ηµ(I). According to [13, Theorem
1.3], the family S is sparse if and only if it satisfies a the Carleson packing condition
X
(1.1) µ(J) ≤ Cµ µ(I)
J⊂I:
J∈S

for all I ∈ D. One important consequence of (1.1) is the fact that the union of two sparse families is again
sparse (with a possibly different constant), and this observation we shall use throughout. The sparse operator
associated with S is X
AS f (x) = hf iI 1I (x),
I∈S
SPARSE DOMINATION FOR BALANCED MEASURES 5

while the corresponding bilinear form already appeared in the Introduction:


X
AS (f, g) = hAS f, gi = hf iI hgiI µ(I).
I∈S

We start constructing such a measure µ so that X (we omit the dyadic grid dependence from the notation)
does not admit sparse domination in the usual dual sense.
Proposition 1.1. Fix 0 < η < 1. There exists a balanced measure µ and two sequences of compactly
supported functions {fj }, {gj } ⊂ L2 (µ) such that for all η-sparse families S ⊂ D
(1.2) |hXfj , gj i| & j|AS (fj , gj )|.

Proof. We take the measure µ to be the one constructed in [18, Section 4, (a)]. In particular, µ is
supported on the unit interval2 and is defined inductively as follows: let ak = 1 − 1/k and bk = 1/k for
k ≥ 2. Set Ik = [0, 2−k ] for k ≥ 0, and denote by Ikb = [2−k , 2−k+1 ] its dyadic sibling. Set µ(I0 ) = 1,
µ(I1 ) = µ(I1b ) = 12 and for k ≥ 2, µ(Ik ) = ak µ(Ik−1 ) = ak µ(Ibk ), so µ(Ikb ) = bk µ(Ik−1 ). The construction is
finished by declaring that the density of µ is taken to be uniform on Ikb . µ is non-dyadically doubling because

µ(Ibk ) 1
b
= = k → ∞ as k → ∞.
µ(Ik ) bk

Clearly, we have µ(Ik ) ∼ µ(Ibk ) since the sequence ak is bounded from above and below. Straightforward
computations lead to

1 µ(Ibk ) µ(Ibk )
m(Ikb ) = bk µ(Ibk ) ∼ , m(Ik ) = ak+1 bk+1 ak µ(Ibk ) ∼ ,
4 k k
which show that µ is balanced. We also have the estimates
1 1
µ(Ik ) ∼ , µ(Ikb ) ∼ .
k k2
One can also show that µ does not have any point masses. Indeed, the only possible point mass is δ0 , but
this does not occur since µ(Ik ) → 0 as k → ∞ and so µ({0}) = 0. To prove (1.2), we assume η ∈ (0, 1) has
been fixed and suppress dependence on η in the computations that follow. For j ≥ 2, we take fj = 1Ij−1 b and
gj = 1Ijb . As a preliminary, we notice fj and gj have disjoint supports and if I ∈ D intersects both supp(f1 )
and supp(f2 ) non-trivially, we must have I = Ik with 0 ≤ k ≤ j − 2. We also observe

b
µ(Ij−1 ) µ(Ijb )
(1.3) Mfj (x) = , Mgj (x) = , x ∈ Ikb for 0 < k ≤ j − 2;
µ(Ik−1 ) µ(Ik−1 )

µ(Ijb ) b
(1.4) Mfj (x) = 1, Mgj (x) = , x ∈ Ij−1 ;
µ(Ij−2 )

b
µ(Ij−1 )
(1.5) Mfj (x) = , Mgj (x) = 1, x ∈ Ijb ;
µ(Ij−2 )

b
µ(Ij−1 ) µ(Ijb )
(1.6) Mfj (x) = , Mgj (x) = , x ∈ Ikb for k > j
µ(Ij−2 ) µ(Ij−1 )
Using the sparsity of S and equations (1.3) through (1.6), we can estimate the sum directly:

2If we want to consider measures supported on all of R, we can just take µ to be uniform with density 1 outside [0, 1] and
the argument still works.
6 JOSÉ M. CONDE ALONSO, JILL PIPHER, AND NATHAN A. WAGNER

X X
hfj iI hgj iI µ(I) . hfj iI hgj iI µ(EI )
I∈S I∈S
ˆ
≤ Mfj · Mgj dµ
[0,1]


≤ Mfj · Mgj dµ
k=1 Ikb
j−2
! ! !
X b
µ(Ij−1 ) µ(Ijb ) µ(Ijb )
b b
= · · µ(Ik ) + · µ(Ij−1 )
µ(Ik−1 ) µ(Ik−1 ) µ(Ij−2 )
k=1
! ∞
! !
b
µ(Ij−1 ) X b
µ(Ij−1 ) µ(Ijb )
b
+ · µ(Ij ) + · · µ(Ikb )
µ(Ij−2 ) µ(Ij−2 ) µ(Ij−1 )
k=j+1
!  
j−2 ∞
1 X 1 1 1 X 1 . 1.
. 4
1 + 3 + 3 + 2 2
j j j j k j3
k=0 k=j+1

Now we turn to a lower bound for |hXfj , gj i|. Given I ∈ D, denote by SI the simple shift SI f =
hf, hI i(hI− − hI+ ). Then, by definition and using the support properties of fj and gj , we have hSI fj , gj i =
hfj , hI i · hhI− − hI+ , gj i 6= 0 if and only if I = Ik with 0 ≤ k ≤ j − 2. Therefore, we may write
j−2
X j−2
X
hXfj , gj i = hSIk fj , gj i = hfj , hIk i · hhIk+1 − hIk+1
b , gj i.
k=0 k=0

For the last term in the sum, we have


q b
µ(Ij−1 ) q
hfj , hIj−2 i = − m(Ij−2 ) b
= − m(Ij−2 ),
µ(Ij−1 )
while q
hhIj−1 − hIj−1
b , gj i = hhIj−1 , gj i = − m(Ij−1 ).
Therefore,
q
µ(Ij ) 1
hfj , hIj−2 i · hhIj−1 − hIj−1
b , gj i =
m(Ij−1 )m(Ij−2 ) ∼ ∼ 2.
j j
Notice that all the other terms in the summation will similarly be positive, because if k < j − 2, fj is
supported entirely on the left half of Ik and gj is supported on the left half of Ik+1 . Therefore, we need not
consider the other terms and (1.2) follows immediately. 

Remark 1.2. If one were instead to consider the measure µ′ defined by b′k = k12 with parallel construction
to Proposition 1.1, then µ′ would have a point mass at 0. In this Pcase one can actually prove the much stronger
result that the bilinear form |hXf1 , f2 i| cannot be controlled by I∈S hf1 iI hf2 iI µ(I) for any dyadic collection
S, let alone a sparse one.
Our next construction uses the same measure µ that we used in Proposition 1.1 to show that the class
A2 (µ) is not the right one to study L2 -boundedness for higher complexity shifts. Again, we exemplify using
X.
Proposition 1.3. There exists a balanced measure µ and a weight w ∈ A2 (µ) so that X is unbounded
on L2 (wdµ).
Proof. Define µ as in the proof of Proposition 1.1. Define the weight w as follows:
( k
2− 2 , x ∈ I2bk for k ≥ 1,
w(x) =
1, otherwise.
SPARSE DOMINATION FOR BALANCED MEASURES 7

−1
We claim w ∈ A2 (µ). To see this, first note that if I is contained in Ikb for some k, then
P∞hwiI hwb iI = 1.
The only other case to consider is I = Iℓ for some ℓ. We will use the simple facts that j>ℓ µ(Ij ) = µ(Iℓ )
and µ(Ij ) ≤ µ(Ik ) if j ≥ k. We estimate
P P k
j>ℓ: µ(Ijb ) + b
k: 2 2 µ(I2k )
j6=2k for any k 2k >ℓ
hw−1 iIℓ =
µ(Iℓ )
P −k
k: 2 2 µ(I2k −1 )
2k >ℓ
≤1+ . 1,
µ(Iℓ )
using ℓµ(Iℓ ) = µ(Iℓ+1 ). The implicit constant is independent of ℓ. The estimate hwiIℓ . 1 is even easier, and
so hwiIℓ hw−1 iIℓ . 1 in this case as well, yielding the claim. However, notice that
k
hw−1 iI bk hwiI bk = 2 2 → ∞ as k → ∞.
2 2 +1

This behavior of the weight will lead to a contradiction: consider test functions fk ≡ w−1 1I bk and gk ≡ 1I bk .
2 2 +1
We compute
ˆ ! 21
k 1
kfk kL2 (wdµ) = w−1 dµ = 2 4 µ(I2bk ) 2 ,
I bk
2

and similarly we have


1
kgk kL2 (wdµ) = µ(I2bk +1 ) 2 .
We now obtain a lower bound for |hXfk , wgk i| using similar considerations as in the proof of Proposition
1.1:

k
2X −1
|hXfk , w gk i| = hSIj fk , w gk i
j=0

≥ |hfk , hI2k −1 i| · |hhI2k − hI bk , w gk i|


2
q p
k
=2 2 m(I2k −1 ) · m(I2k )
k 1 1
∼ 2 2 µ(I2bk ) 2 µ(I2bk +1 ) 2
k
= 2 4 kfk kL2 (wdµ) kgk kL2 (wdµ) .

This shows that X is unbounded on L2 (wdµ). 

2. Modified Sparse Domination and Consequences


2.1. Modified Sparse forms and Calderón-Zygmund estimates. Proposition 1.1 shows that usual
sparse forms are not enough to achieve sparse domination for Haar shifts in general. To remedy this, we
introduce other forms that reflect the complexity of the operator that we want to dominate. The dyadic
distance between two intervals I, J ∈ D that share a common ancestor is defined as follows:
distD (I, J) := min (s + t).
(s,t):I (s) =J (t)

If I, J do not share a common ancestor, then by convention we let distD (I, J) = ∞. Notice that distD (I, J) =
0 if and only if I = J. For any I ∈ D, we let CN the number of J ∈ D such that distD (I, J) ≤ N + 2.
CN does not depend on I. For convenience, we enumerate the dyadic intervals J ∈ D that satisfy 2 <
distD (I, J) ≤ N + 2 and J ∩ I = ∅ as c1 (I), c2 (I), · · · , cN ′ (I), where N ′ is a constant depending only on N .
We will suppose that we have fixed an arbitrary enumeration function for each I ∈ D and abuse notation by
writing cj to represent all enumerations for all intervals. Note that given a fixed J ∈ D, there are at most
8 JOSÉ M. CONDE ALONSO, JILL PIPHER, AND NATHAN A. WAGNER

N ′ intervals I satisfying cj (I) = J for some j. This fact will be used in various estimates without further
comment. Given S ⊂ D, N ∈ N, and f1 , f2 ∈ L2 (µ), we define
X p p
CSN (f1 , f2 ) := hf1 iJ hf2 iK m(J) m(K).
J,K∈S:
2<distD (J,K)≤N +2,
J∩K=∅

We will later show that the forms CSN are Lp -bounded and of weak-type (1, 1). What we show next
is a technical lemma that justifies considering only intervals at distance greater than 2 and disjoint in the
definition.
Lemma 2.1. Let N ∈ N and S ⊂ D be η-sparse. There exist 0 < η ′ ≤ η and S ′ ⊇ S η ′ -sparse such that
X p p
hf1 iJ hf2 iK m(J) m(K) . AS ′ (f1 , f2 ) + CSN′ (f1 , f2 )
J,K∈S:
1≤distD (J,K)≤N +2

Proof. Since we are going to construct S ′ so that S ⊆ S ′ , we only need to bound terms arising from
pairs (J, K) such that either distD (J, K) = 2 and J ∩ K = ∅, or J ∩ K ∈ {J, K} and distD (J, K) ≤ N + 2.
In the latter case, if J = K (k) for some k ≤ N + 2 -say-, we estimate
p p
hf1 iJ hf2 iK m(J) m(K) .k hf1 iJ hf2 iJ µ(J),
because m(J) ∼ m(K) and m(J) ≤ µ(J). In the former case, we must have K = J b . If we suppose J and
J b both belong to S, then

p q
hf1 iJ hf2 iJ b b
m(J) m(J b ) . hf1 iJb hf2 iJb µ(J),
where the implicit constant is independent of J because each dyadic interval has exactly two children. We
define now S ′ as the union of the family S and the intervals Jb for each J ∈ S such that J b ∈ S. It can be
seen, using (1.1) that S ′ is η ′ -sparse for some η ′ = η ′ (η, N ), and the assertion follows. 

To prove Theorem A we will perform Calderón-Zygmund decompositions at different heights for two
functions f1 and f2 simultaneously. In the Lebesgue measure case, the Calderón-Zygmund decomposition
yields L∞ estimates for the good part of the function, but that is unavailable in our case, so we must content
ourselves with a BMO estimate. We use the standard martingale BMO norm (associated with the filtration
generated by D), which is
1
ˆ
kf kBMO := sup |f − hf iIb| dµ.
I∈D µ(I) I

Lemma 2.2. Let 0 ≤ f1 , f2 ∈ L1 (µ) be supported on I ∈ D and λ1 , λ2 > 0. Then there exist functions
gj , bj such fj = gj + bj for j = 1, 2 and satisfying the following properties:
P∞
(1) There exist pairwise disjoint, dyadic intervals {Ik } ⊂ D(I) so that bj = k=1 bj,k where each bj,k
is supported on Ibk , Ibk bj,k = 0, and kbj,k kL1 (µ) . Ik |fj | dµ for j = 1, 2. In particular, for j = 1, 2
´ ´

we have
bj,k = fj 1Ik − hfj 1Ik iIbk 1Ibk .

(2) For j = 1, 2, the function gj ∈ Lp (µ) for all 1 ≤ p < ∞ and satisfies kgj kpLp (µ) ≤ Cp λp−1
j kfj kL1 (µ) .
(3) For j = 1, 2 the function gj ∈ BMO and kgj kBMO ≤ λj .
Proof. The statement of the decomposition is contained in [4, 18], except for (3), which is new and for
which we give full details below. The intervals Ik are selected using the usual stopping-time argument: let
{Ik } be the set of maximal dyadic intervals contained in I so that either hf1 iIk > λ1 or hf2 iIk > λ2 . Then
(1) follows. We now turn to (3), which by interpolation proves (2). For j = 1, 2, we have
X
gj = fj 1(∪k Ik )c + hfj 1Ik iIbk 1Ibk := gj,1 + gj,2 .
k
SPARSE DOMINATION FOR BALANCED MEASURES 9

c
By the Lebesgue Differentiation theorem, for almost every point x ∈ (∪j Ij ) there is a sequence of unselected
cubes shrinking down to x and we have |fj (x)| ≤ λj , so kgj,1 kL∞ ≤ λj . It easily follows that kgj,1 kBMO ≤ 2λj
for j = 1, 2. It therefore suffices to show kgj,2 kBMO . λj for j = 1, 2. If I = Ik is a selected interval, we have
 
X µ(Ibℓ ) X
hgj,2 iIbk 1Ik =  hfj 1Iℓ iIbℓ + hfj 1Iℓ iIbℓ  1Ik
b b
µ(Ibk ) b b
ℓ:Iℓ (Ik ℓ:Ik ⊂Iℓ
 
X
´
f j dµ
= Ek 1Ik +  hfj 1Iℓ iIbℓ  1Ik ,
µ(Ibk ) b b
ℓ:Ik ⊂Iℓ

where [
Ek = Iℓ .
ℓ:Ibℓ (Ibk
On
P the other hand,  notice
P that by definition  and the disjointness of the selected intervals, gj,2 1Ik =
ℓ hfj 1Iℓ iIbℓ 1Ibℓ 1Ik = ℓ:Ibk ⊂Ibℓ hfj 1Iℓ iIbℓ 1Ik and therefore
´
1 fj dµ
ˆ
|gj,2 − hgj,2 iIbk | dµ = Ek ≤ hfj iIbk ≤ λj .
µ(Ik ) Ik µ(Ibk )
If I is not a selected interval, we can similarly argue
 
X
´
f j dµ
hgj,2 iIb1I = EI 1I +  hfj 1Iℓ iIbℓ  1I ,
b
µ(I)
ℓ:Ib⊂ Ibℓ
S
where EI = ℓ:Ibℓ (Ib Iℓ . However, in this case it is possible that I = Ibℓ for some ℓ, so we have g2 1I =
P  P 
ℓ:Ib⊂ Ibℓ hfj 1Iℓ iIbℓ 1I + ℓ:I=Ibℓ hfj 1Iℓ iIbℓ 1I . However, there are at most two terms in the second summa-
tion, and each average in the sum is controlled by λj , so the BMO norm can be estimated as before. 
We next introduce notation about collections of intervals. Suppose functions f1 , f2 and an initial interval
I0 ∈ D have been fixed as in Lemma 2.2. For I ⊆ I0 , let B(I) denote the collection of selected intervals in
Lemma 2.2 applied to fj 1I at heights λj = 16hfj iI , j = 1, 2. We then inductively define, for k ∈ N:
[ [
(2.1) B0 (I) = B(I), B1 (I) = B(J), Bk (I) = B(J).
J∈B(I) J∈Bk−1 (I)

Moreover, given N ∈ N, we set


N
[
B N (I) := Bk (I).
k=0
Finally, we set
G(I) := {J ∈ D(I) : J 6⊂ K for any K ∈ B(I)}, O(I) = {J ∈ D : J ) I}.

2.2. Haar shifts. Let α = {αIJ,K }I,J,K∈D be a (triply indexed) sequence with kαk∞ ≤ 1. For any two
integers s, t ∈ N, we define a Haar shift of complexity (s, t) by
X X X
Tαs,t f := αIJ,K hf, hJ ihK , f ∈ L2 (µ)
I∈D J∈D s (I) K∈D t (I)

The above sum converges in L2 (µ) and µ-almost everywhere. The family of Haar shifts of complexity (s, t)
will be denoted by 
HS(s, t) = Tαs,t : kαk∞ ≤ 1 .
It is immediate from the fact that {hI }I∈D forms an orthonormal basis for L2 (µ) that for any complexity
(s, t) and any measure µ, any T ∈ HS(s, t) is bounded on L2 (µ) with operator norm only depending on the
complexity. If µ is balanced, then T is bounded on Lp (µ) and of weak-type (1, 1), as shown in [18]. We are
now ready to state our main theorem.
10 JOSÉ M. CONDE ALONSO, JILL PIPHER, AND NATHAN A. WAGNER

Theorem 2.3. Let µ be balanced and N ∈ N. There exists η ∈ (0, 1) such that if (s, t) satisfies s+t ≤ N ,
the following holds: for all T ∈ HS(s, t) and each pair of compactly supported, bounded nonnegative functions
f1 , f2 there exists an η-sparse collection S ⊂ D such that

|hT f1, f2 i| . AS (f1 , f2 ) + CSN (f1 , f2) .
For the remainder of this subsection we assume that the pair (s, t) has been fixed, and the two functions
f1 and f2 are supported on a fixed dyadic interval I0 . We also suppose that the sequence α that defines the
Haar shift T is nonzero except for finitely many indices, since the estimates we obtain will be uniform over
all finite sums, and that α has been fixed as well. It will be useful to introduce the following notation: let
I = [2−k p, (p + 1)2−k ) ∈ D, where k, p ∈ Z. Given fixed s, t ∈ N, let 0 ≤ m ≤ 2s − 1 and 0 ≤ n ≤ 2t − 1 be
integers. We index the intervals J ∈ D s (I) via
Ism = [2−k p + m2−k−s , 2−k p + (m + 1)2−k−s ),
and we define Itn in an analogous manner. With this notation, the two dyadic children of a given interval Ism
2m 2m+1
are Is+1 and Is+1 . The following lemma contains the key iteration step in the sparse domination argument.
Lemma 2.4. Let 0 ≤ m < 2s , 0 ≤ n < 2t be integers and set N = s + t. If β = {βI }I∈D is a sequence
with kβk∞ ≤ 1 which is nonzero for finitely many indices, then
X  X 
p p
βJ hf1 , hJsm ihhJtn , f2 i . hf1 iI hf2 iI µ(I) + hf1 iS hf2 iT m(S) m(T ) .
J∈G(I) S∈Bs (I), T ∈Bt (I):
distD (S,T )≤N +2

Proof. Fix I ⊆ I0 . We begin with the usual splitting with fj 1I = gj + bj with respect to the Calderón-
Zygmund decomposition applied on I:

X X X
βJ hf1 , hJsm ihhJtn , f2 i ≤ βJ hg1 , hJsm ihhJtn , g2 i + βJ hg1 , hJsm ihhJtn , b2 i
J∈G(I) J∈G(I) J∈G(I)

X X
+ βJ hb1 , hJsm ihhJtn , g2 i + βJ hb1 , hJsm ihhJtn , b2 i
J∈G(I) J∈G(I)

:= (I) + (II) + (III) + (IV).


We handle the four terms in order. We estimate, using the fact that the hJ are orthonormal, the L2
control on the good functions, and the fact that kβkℓ∞ ≤ 1:
X
(I) ≤ |hg1 , hJm
s i| · |hhJ t , g2 i|
n
J∈G(I)
!1/2 !1/2
X X
2 2
≤ |hg1 , hK i| |hhL , g2 i|
K∈D L∈D
≤ kg1 kL2 (µ) kg2 kL2 (µ)
ˆ 1/2 ˆ 1/2
1/2 1/2
. (hf1 iI ) f1 dµ (hf2 iI ) f2 dµ = hf1 iI hf2 iI µ(I).
I I
This proves the desired estimate for (I). For (II), we have
X X
(II) ≤ |hg1 , hJsm i||hhJtn , b2,k i|.
J∈G(I) k

However, many of the terms in this summation are zero due to the cancellation/support of the b2,k and the
(2)
constraint J ∈ G(I). Indeed, the cancellation on the b2,k implies Jtn ( Ik , while the condition J ∈ G(I)
b
together with the support of b2,k implies that either Ik ( J or J ⊂ Ik . But in the latter case, b2,k is a constant
function on the support of hJtn (which is contained in Ikb ), so these terms may be eliminated as well by the
SPARSE DOMINATION FOR BALANCED MEASURES 11

cancellation of hJtn . Note that given any fixed Ik , there are finitely many intervals J that can satisfy these
(2+t)
constraints, and the number of intervals only depends on t, not k. Indeed, we may write Ik ( J ( Ik .
Therefore, after eliminating these terms, the double summation is equal to:
X X
|hg1 , hJsm i||hhJtn , b2,k i|.
k J∈G(I):
(2+t)
Ik (J(Ik

Fix J ∈ G(I) satisfying the additional two conditions in the summation. We will obtain a uniform estimate
on the inner summation. On the one hand, we estimate
p
|hg1 , hJsm i| = m(Jsm ) hg1 iJs+1
2m − hg1 i 2m+1
Js+1
p  
≤ m(Jsm ) hg1 iJs+1
2m − hg1 iJ m + hg1 iJ m − hg1 i 2m+1
s s Js+1
p p
≤ 2 m(Jsm )kg1 kBMO . m(Jsm )hf1 iI ,
by Lemma 2.2, (3). On the other hand, using the L∞ estimate of the Haar functions, the L1 control on the
bad function, and the fact that Jsm and Jtn share the common ancestor J, we estimate:
1
|hhJtn , b2,k i| ≤ p kb2,k kL1 (µ)
m(Jtn )
1
ˆ
. p |f2 | dµ.
m(Jtn ) Ik
1
ˆ
. p |f2 | dµ,
m(Jsm ) Ik
where in the last line we used our assumption that µ is balanced applied at most N times. Thus, in particular,
for fixed k, we have a bound ˆ
|hg1 , hJsm i||hhJtn , b2,k i| . hf1 iI |f2 | dµ,
Ik
that is independent of J. Putting these two estimates together and summing, we have
X X
(II) ≤ |hg1 , hJsm i||hhJtn , b2,k i|
k J∈G(I):
(2+t)
Ik (J(Ik
X X ˆ
. hf1 iI |f2 | dµ
k J∈G(I): Ik
(2+t)
Ik (J(Ik
X ˆ
. hf1 iI |f2 | dµ . hf1 iI hf2 iI µ(I),
k Ik

where in the penultimate estimate the implicit constant depends only on t and in the last step we used the
fact that the Ik ∈ B(I) are disjoint and contained in I for all k. Term (III) is handled similarly and we omit
the details, so we are left with (IV). We estimate
X
(IV) ≤ |hb1 , hJsm i| · |hhJtn , b2 i|.
J∈G(I)

So, we need to handle terms of the form |hb1,j , hJsm i| · |hhJtn , b2,k i| where b1,j is supported on Ij , b2,k is
supported on Ik , and both Ij and Ik are selected intervals. Again, there are some constraints on the interval
(2+s) (2+t)
J. In particular, we must have Ij ( J ( Ij and Ik ( J ( Ik , as we determined previously. So for
fixed j, k, there are finitely many intervals J that contribute to the sum, where the number of terms only
depends on s, t. Therefore, similar to before, it suffices to consider the sum
XX X
|hb1,j , hJsm i| · |hhJtn , b2,k i|
j k J∈G(I):
(2+s) (2+t)
Ij (J(Ij and Ik (J(Ik
12 JOSÉ M. CONDE ALONSO, JILL PIPHER, AND NATHAN A. WAGNER

and show it is controlled by the right hand side in the statement. Let us fix J, j, k as in the summation for
the time being, and consider a term |hb1,j , hJsm i| · |hhJtn , b2,k i|. Note that either Jsm = Ibj or Jsm ⊂ Ij . In the
former case, we can control the inner product as follows:
q
|hb1,j , hJs i| = m(Ibj ) hb1,j iIj − hb1,j iIjb
m

q
= m(Ibj ) hf1 iIj − hf1 1Ij iIbj + hf1 1Ij iIbj − hf1 1Ij iIjb
q
∼ m(Ij )hf1 iIj .

If Jtn = Ibk as well, then we obviously have the analogous estimate and therefore:
q p
(2.2) |hb1,j , hJsm i| · |hhJtn , b2,k i| . hf1 iIj hf2 iIk m(Ij ) m(Ik )
Moreover, notice that distD (Ij , Ik ) ≤ N + 2 and by definition, Ij , Ik ∈ B(I) and are disjoint unless Ij = Ik .
In the latter case when Jsm ⊂ Ij , we see that
p
|hb1,j , hJsm i| = m(Jsm ) hb1,j iJs−m − hb1,j iJ m
s+
p
m
= m(Js ) hf1 iJs− m − hf1 1I i b + hf1 1I i b − hf1 iJ m
j Ij j Ij s+
p p
(2.3) m
≤ m(Js )hf1 iJs− m + m
m(Js )hf1 iJs+ m .

p m
It suffices to control m(Jsm )hf1 iJs− m ; the other term is obviously handled in the same way. If J
s− ∈ B(Ij ),
m
we leave the estimate as is. So suppose that Js− ∈/ B(Ij ). Then either hf1 iJs−
m ≤ 16hf1 iI
j or there exists
m
K ( Ij with K ∈ B(Ij ) and Js ( K. In the former case, we have the estimate:
p q
(2.4) m(Jsm )hf1 iJs−
m . m(Ij )hf1 iIj ,
using the balanced property of µ. In the latter case, we apply the same procedure, letting K assume the role
m
of Ij . That is, either Js− ∈ B(K), in which case we stop and leave the original estimate as is, or we have

the control hf1 iJs−
m ≤ 16hf1 iK , or there exists an interval K ∈ B(K) so that J
m−
s ( K ′ , and we apply the
procedure again to K ′ . Notice that distD (Js−m
, Ij ) ≤ s. Iterating this procedure at most s times and using
the balanced condition, we can find an interval S ∈ B s (I) (more precisely, B s−1 (Ij )) containing Js− m
and
satisfying
p p
(2.5) m(Jsm )hf1 iJs−m . m(S)hf1 iS ,
where the implicit constant depends only on µ and s. Moreover, the same reasoning applies to Jtn , giving us
an estimate
p p
(2.6) m(Jtn )hf1 iJt−
n . m(T )hf1 iT .
for some T ∈ B t−1 (Ik ) (or T = Ik ) with Jt−
n
⊂ T. Let S, T be as above. We observe that distD (S, T ) ≤ N + 2,
since both have common ancestor J. Examining estimates (2.2) through (2.6) and using the fact that the
number of intervals J satisfying these constraints only depends on s, t, we deduce that for fixed j, k, we have
the bound:
X
|hb1,j , hJsm i| · |hhJtn , b2,k i|
J∈G(I):
(2+s) (2+t)
Ij (J(Ij and Ik (J(Ik
 q X 
p p p
. hf1 iIj hf2 iIk m(Ij ) m(Ik ) + hf1 iS hf2 iT m(S) m(T ) ,
S∈Bs (Ij ), T ∈Bt (Ik ):
distD (S,T )≤N +2

where the implicit constant depends on s, t. Summing over j, k (which correspond to disjoint intervals Ij , Ik )
then completes the proof. 

We can now turn to the proof of Theorem 2.3.


SPARSE DOMINATION FOR BALANCED MEASURES 13

Proof of Theorem 2.3. We first describe the construction of the sparse collection S and then show
how Lemma 2.4 leads to the desired domination. With the notation introduced in (2.1), we assume that
max{s,t}+1
f1 , f2 are supported on an interval I˜0 ∈ D, we set I0 = I˜0 and


[
S := I0 ∪ Bk (I0 ).
k=0

S is independent of s and t and only depends on f1 and f2 . We need to show that it is sparse, and we do
so showing that it satisfies (1.1). By an inductive argument, it suffices to show that for a given I ∈ Bj (I0 ),
there holds
X µ(I)
µ(J) ≤ .
2
J∈Bj+1 (I0 ):
J⊂I

For each such interval J, select s(J) ∈ {1, 2} such that hfs(J) iJ ≥ 16hfs(J) iI . We estimate, using the fact
that the intervals in the sum are pairwise disjoint:

X X
´
J
fs(J) dµ
µ(J) ≤
16hfs(J) iI
J∈Bj+1 (I0 ): J∈Bj+1 (I0 ):
J⊂I J⊂I
X ´ ´ 
µ(I) f dµ f2 dµ
≤ ´J 1 + ´J
16 f dµ
I 1
f dµ
I 2
J∈Bj+1 (I0 ):
J⊂I
µ(I)
≤ .
8

The claim then follows. We now turn to actually proving the domination. Since f1 and f2 are supported on
I0 , we may assume that αIJ,K = 0 if I ∩ I0 = ∅. Therefore, we may decompose
X X X
hT f1 , f2 i = αIJ,K hf1 , hJ ihf2 , hK i
I∈O(I0 ) J∈Ds (I) K∈Dt (I)
s t
2X −1 2X −1 X
+ αIIm
s ,I t hf1 , hI s ihf2 , hI t i
n m n
m=0 n=0 I∈D(I0 )
s t
2X −1 2X −1
=: hT0 f1 , f2 i + hTm,n f1 , f2 i.
m=0 n=0

Since we have the corona-type decomposition


[
D(I0 ) = G(J),
J∈S

we may apply Lemma 2.4 to each I ∈ S and sum to get


X p p
|hTm,n f1 , f2 i| . AS (f1 , f2 ) + hf1 iJ hf2 iK m(J) m(K).
J,K∈S:
1≤distD (J,K)≤N +2

after summing over I ∈ S. Lemma 2.1 then completes the estimate for the operators Tm,n after passing to a
second sparse collection S ′ . Finally, because of our choice of I0 , if I0 ( I then hJ is constant on any interval
14 JOSÉ M. CONDE ALONSO, JILL PIPHER, AND NATHAN A. WAGNER

contained in I˜0 if J ∈ Ds (I) ∪ Dt (I). Therefore,


X X X
|hT0 f1 , f2 i| ≤ |hf1 , hJ i||hf2 , hK i|
I∈O(I0 ) J∈Ds (I) K∈Dt (I)
X X X
= |hhf1 iI˜0 1I˜0 , hJ i||hhf2 iI˜0 1I˜0 , hK i|
I∈O(I0 ) J∈Ds (I) K∈Dt (I)
! 12 ! 12
X X
.s,t |hhf1 iI˜0 1I˜0 , hJ i|2 |hhf2 iI˜0 1I˜0 , hK i|2
J∈D K∈D

= hf1 iI˜0 1I˜0 2


hf2 iI˜0 1I˜0 2
= hf1 iI˜0 hf2 iI˜0 µ(I˜0 ),

and the result follows. 


Remark 2.5. Notice that CS0 (f1 , f2 )
≡ 0 for any pair f1 , f2 . Therefore, Theorem 2.3 with (s, t) = (0, 0)
recovers the known result for Haar multipliers, which admit the usual sparse domination by AS (f1 , f2 ) (see
[15]).
Remark 2.6. Theorem A is an immediate consequence of Theorem 2.3 when we take N = 1. Note that
this provides a sparse domination for the dyadic Hilbert transform X and -the same one- for its adjoint X∗ .
In addition, by following the proof of Theorem 2.3, one can obtain a refined sparse domination for X alone
of the form
 
X X X 
|hXf1 , f2 i| . C 
 hf1 iI hf2 iI µ(I) + hf1 iI hf2 iI−b m(I) + hf1 iI hf2 iI+b m(I)

I∈S I∈S: I∈S:
b b
I− ∈S I+ ∈S

We leave the details to the interested reader.


2.3. A maximal function. In what follows, given a pair of dyadic intervals I, J and p ∈ [1, ∞), let
(
1 I=J
(2.7) cp (I, J) =  m(I)p/2 m(J)p/2 
b

µ(J)µ(I)p−1 otherwise.
The following modified maximal function that depends on N ∈ N is intimately related to sparse domination
for T ∈ HS(s, t) for s + t = N .
Definition 2.7. Given N ∈ N, define the following maximal dyadic operator for f ∈ L1 (µ):

MN f (x) := sup cb1 (I, J)h|f |iI 1J (x)


I,J∈D:
distD (I,J)≤N +2

Under the assumption that µ is balanced, the operator MN enjoys similar boundedness properties as
the classical Hardy-Littlewood maximal function.
Lemma 2.8. MN is bounded on Lp (µ) for 1 < p ≤ ∞ and is of weak-type (1, 1).
Proof. MN is bounded on L∞ (µ) with constant depending on the parameter N and the balancing
constants of µ, so it suffices to show MN is weak-type (1, 1) and then interpolate. Take 0 ≤ f ∈ L1 (µ), and
for λ > 0 let
Eλ = {x : MN D f (x) > λ}.
The following countable collection of intervals then clearly covers Eλ :

Cλ = J ∈ D : ∃ I ∈ D so distD (I, J) ≤ N + 2 and cb1 (I, J)hf iI > λ .
By selecting the maximal dyadic intervals in Cλ , we obtain a countable, pairwise disjoint collection of intervals
{Jj }∞ ∞
j=1 that covers Eλ , together with an associated collection of dyadic intervals {Ij }j=1 satisfying

(2.8) cb1 (Ij , Jj )hf iIj > λ


SPARSE DOMINATION FOR BALANCED MEASURES 15

for all j. The intervals {Ij }∞


j=1 need not be pairwise disjoint. However, we can select a subsequence of
maximal ones that are pairwise disjoint, which we denote by {Ijk }∞
k=1 . For fixed k ∈ N, define the following
index collections:
Ak = {j ∈ N : Ij ⊂ Ijk }, Bk = {j ∈ Ak : Jj ∩ Ijk 6= ∅}.
{Ak }∞
k=1 partitions N. First, we claim that
(2.9) |Ak \ Bk | ≤ CN .
To see this, notice that if j ∈ Ak \ Bk , then Ij ⊂ Ijk and Jj ∩ Ijk = ∅. Notice that if K is a common dyadic
ancestor to Ij and Jj , it must contain Ijk , for otherwise it is contained in Ijk , and thus Ijk contains Jj ,
a contradiction. Thus K is also a common dyadic ancestor of Ijk and Jj . But then, by the definition of
dyadic distance, we have distD (Jj , Ijk ) ≤ distD (Jj , Ij ) ≤ CN , which implies (2.9). Also, notice that for any
j ∈ Ak \ Bk , by (2.8) and the balanced hypothesis we have the uniform control

  p !ˆ
1 µ(Jj ) m(Ij )m(Jj )
µ(Jj ) ≤ f dµ
λ µ(Ij ) µ(Jj ) Ij

1
ˆ
(2.10) . f dµ.
λ Ijk

On the other hand, if there exists j ∈ Bk that satisfies Jj ⊃ Ijk , then Jj is the only element of Bk , since the
Jj are pairwise disjoint, and moreover µ(Jj ) is controlled by (2.10) by the same argument as above. In this
case, |Ak | ≤ CN + 1 and we can redefine Bk = ∅ in the argument that follows. In the second case, we have
Jj ( Ijk for all j ∈ Bk , but this time Bk can be infinite. We have
[
(2.11) Jj ⊂ Ijk ,
j∈Bk

and moreover, by the definition of Ijk and using the fact that m(Ijk ) ∼ m(Jjk ) by the balanced condition
with implicit constant depending only on N :
p
1 m(Ijk )m(Jjk )
ˆ
µ(Ijk ) ≤ f dµ
λ µ(Jjk ) Ijk
1
ˆ
(2.12) . f dµ.
λ Ij
k

Collecting these facts together, we then have



X
µ(Eλ ) ≤ µ(Jj )
j=1
X∞ X ∞ X
X
= µ(Jj ) + µ(Jj )
k=1 j∈Ak \Bk k=1 j∈Bk

:= (I) + (II).

Note that for (I), using (2.9) and (2.10) and recalling that the Ijk are pairwise disjoint, we have the control
X∞ X ∞
1X X
ˆ
µ(Jj ) . f dµ
λ Ij
k=1 j∈Ak \Bk k=1 j∈Ak \Bk k

∞ ˆ
(CN + 1) X
≤ f dµ
λ Ij k=1 k

kf kL1 (µ)
. ,
λ
16 JOSÉ M. CONDE ALONSO, JILL PIPHER, AND NATHAN A. WAGNER

which is the bound we want. We bound (II) using the pairwise disjointess of the Jj , (2.11), and (2.12):

∞ X
X ∞
X
µ(Jj ) ≤ µ(Ijk )
k=1 j∈Bk k=1
X∞ ˆ
1
. f dµ
λ Ijk
k=1
kf kL1 (µ)
. ,
λ

which completes the proof. 

MN admits the same sparse domination as any T ∈ HS(s, t) for N = s + t. As before, we prove the
sparse domination for the bi-sublinear form defined by it.

Theorem 2.9. Let N ∈ N. There exists η ∈ (0, 1) so that for each pair of compactly supported, bounded,
nonnegative functions f1 , f2 ∈ L1 (µ), there exists an η-sparse collection S ⊂ D so that

(2.13) | MN f1 , f2 | . (AS (f1 , f2 ) + CSN (f1 , f2)).


Proof. Fix f1 , f2 as in the statement. By Lemma 2.1, it suffices to show there exists η ∈ (0, 1) and an
η-sparse collection S so that

X p p
(2.14) | MN f1 , f2 | . AS (f1 , f2) + hf1 iJ hf2 iK m(J) m(K).
J,K∈S:
1≤distD (J,K)≤N +2

We will use a level set argument to prove (2.14). For n ∈ Z set

En = {x ∈ R : MN f1 (x) > 2n }.

Cover En \ En+1 by maximal dyadic intervals {Jnj }∞ j


j=1 so that for each Jn , there exists a corresponding
j
In ∈ D satisfying

2n < cb1 (Inj , Jnj )hf1 iInj ≤ 2n+1 , distD (Inj , Jnj ) ≤ N + 2.

For fixed n, {Jnj } is a pairwise disjoint collection of intervals, although this is no longer necessarily the case
when one varies the parameter n and moves up and down level sets. However, even for fixed n, the collection
{Inj } need not be pairwise disjoint. Even so, by using an argument similar to that in Lemma 2.8, in the
bound we obtain we can replace {Inj } with a subcollection of intervals which satisfy a finite overlap condition.
We next describe the details of passing to this subcollection and obtaining the sparse bound. First, for fixed
n, select a maximal subsequence {Injk }∞ j
k=1 among the In . This collection will be pairwise disjoint. Then, for
k k
each k, define index collections An and Bn as in the proof of Lemma 2.8. Recall we argued that for each
k, we have |Akn \ Bnk | ≤ CN , and this bound is independent of n as well. If |Bnk | = 1, let ℓ(k) ∈ Bnk denote
ℓ(k) ℓ(k)
the unique index so that In ⊂ Injk and Jn ∩ Injk 6= ∅. Let Cn denote the collection of indices k so that
ℓ(k)
|Bnk | = 1 and Jn ⊃ Injk . Set


[
Sn1 := {Jnj }∞
j=1 , Sn2 := {Injk }∞
k=1 , Sn3 := {Inj }j∈Akn \Bnk , Sn4 := {Inℓ(k) }k∈Cn .
k=1
SPARSE DOMINATION FOR BALANCED MEASURES 17

The collections Sn1 , Sn1 , and Sn4 are pairwise disjoint by construction, while the collection Sn3 has finite overlap.
Next, we estimate:
ˆ
| MN f1 , f2 | = MN f1 · f2 dµ
R

X ˆ
n
∼ 2 f2 dµ
n=−∞ En \En+1

X∞ ∞ ˆ
X
≤ 2n f2 dµ
j
n=−∞ j=1 Jn

X X∞ X ˆ ∞
X ∞ X ˆ
X
n n
= 2 f2 dµ + 2 f2 dµ.
j j
n=−∞ k=1 j∈Ak k Jn n=−∞ k=1 j∈Bn
k Jn
n \Bn

For the first sum, we have:



X ∞
X X ˆ X ∞
∞ X X
2n f2 dµ . cb1 (Inj , Jnj )hf1 iInj hf2 iJnj µ(Jnj )
j
n=−∞ k=1 j∈Ak k Jn n=−∞ k=1 j∈Ak k
n \Bn n \Bn

X X p
. hf1 iI hf2 iJ m(I)m(J).
n=−∞ I∈S 3 , J∈S 1 :
n n
distD(I,J)≤N +2

For the second sum, we first consider the case when k ∈ Cn . Then it is easy to see we arrive at the estimate:
X∞ X ˆ X∞ X p
n
2 f2 dµ . hf1 iI hf2 iJ m(I)m(J).
ℓ(k)
n=−∞ k∈Cn Jn n=−∞ I∈S 4 , J∈S 1 :
n n
distD(I,J)≤N +2
′ ′
If k 6∈ Cn , then any j ′ ∈ Bnk must satisfy Jnj ⊂ Injk . Then we have, using the pairwise disjointness of the Jnj :
X∞ X X ˆ ∞
X X ˆ
n b jk jk
2 ′
f2 dµ . c1 (In , Jn )hf1 iI jk j f2 dµ
j n
n=−∞ k6∈Cn j ′ ∈Bn
k Jn n=−∞ k6∈Cn Ink


X X
. hf1 iI hf2 iI µ(I).
n=−∞ I∈Sn
2

Put

[ 4
[
Sj = Snj , and S= Sj .
n=−∞ j=1
Altogether, we have proven the desired estimate (2.14) for the collection S. We claim that the collection S
is sparse. For readability, we separate out this technical section of the proof into a dedicated lemma. The
proof of Theorem 2.9 is complete, up to the verification of the sparsity of S. 
Lemma 2.10. There exists η ∈ (0, 1) so that the collections Sj , constructed in the proof of Theorem 2.9,
are individually η-sparse, for j ∈ {1, 2, 3, 4}. Therefore, S is η ′ -sparse for some 0 < η ′ < η.
Proof. We first argue for the sparsity of S 1 . It is enough to show that S 1 is a union of finitely many

sparse collections. To begin with, note that if Jnj ′ ⊂ Jnj for some indices n, j, n′ , j ′ , we necessarily have

n′ ≥ n, since Jnj ⊂ En and Jnj ′ ∩ Enc ′ +1 6= ∅ by construction. By a standard argument, it suffices to show
that there exists a sufficiently large positive integer m (depending only on N ) so that for n′ = n + m, we
have, for each interval Jnj ∈ Sn1 ,
X ′ 1
(2.15) µ(Jnj ′ ) ≤ µ(Jnj ).
j′ j
2
j ′ : Jn′ ⊂Jn

First, we split the left hand side of (2.15) as follows:


18 JOSÉ M. CONDE ALONSO, JILL PIPHER, AND NATHAN A. WAGNER

X ′ X ′
µ(Jnj ′ ) + µ(Jnj ′ )
′j j j′ j
j ′ : Jn ′ ⊂Jn j ′ : Jn ′ ⊂Jn
j′ j j′ j
In ′ 6⊂Jn In ′ ⊂Jn

:= (I) + (II).
′ ′ ′ ′
To estimate (I), notice that if Inj ′ 6⊂ Jnj , we have distD (Jnj ′ , Jnj ) ≤ distD (Jnj ′ , Inj ′ ) ≤ N + 2, so the summation

actually has at most CN terms. Moreover, we also have distD (Inj ′ , Jnj ) ≤ N + 2. Then we estimate, using
these facts, the balanced hypothesis, and the fact that Jnj intersects En+1 c
:
X q
1 ′ ′
(I) ≤ n′ m(Inj ′ )m(Jnj ′ )hf1 iI j′
2 ′
n′
j j
j ′ : Jn ′ ⊂Jn
j ′ j
In ′ 6⊂Jn

X q
1 ′
. m(Inj ′ )m(Jnj )hf1 iI j′
2n′ ′
n′
j j
j ′ : Jn ′ ⊂Jn
j ′ j
In ′ 6⊂Jn

1 X CN
≤ µ(Jnj ) = µ(Jnj ),
2m−1 ′
2m−1
j j
j ′ : Jn ′ ⊂Jn
j′ j
In ′ 6⊂Jn

so we clearly have the desired control on (I) assuming a sufficiently large choice of m. To control (II), we
begin by observing that by definition,
[ ′
n ′
o
hf1 iJnj = cb1 (Jnj , Jnj )hf1 iJnj ≤ 2n+1 , and Jnj ′ ⊂ x ∈ Jnj : MN (1Jnj f1 )(x) > 2n .
j ′ j
j ′ : Jn ′ ⊂Jn
j′ j
In ′ ⊂Jn


Then we may estimate, using the pairwise disjointness of Jnj ′ and the weak-type estimate in the proof of
Lemma 2.8:
n ′
o
(II) ≤ µ x ∈ Jnj : MN (1Jnj f1 )(x) > 2n
kMN kL1 (µ)→L1,∞ (µ)
ˆ
≤ f1 dµ
2n′ Jn j

kMN kL1 (µ)→L1,∞ (µ)


≤ µ(Jnj ),
2m−1
which also establishes the necessary control for (II) as long as m is chosen large enough. Thus, we conclude
the collection S 1 is sparse.
Next, is not difficult to show that if we prove the collection S 2 is sparse, then this will also imply
that collections S 3 and S 4 are sparse, so we focus on checking the Carleson packing condition for each
I = Injk ∈ S 2 . To begin with, choose a to be the unique positive integer satisfying
2a < hf1 iI jk = cb1 (Injk , Injk )hf1 iI jk ≤ 2a+1 .
n n

This implies that Injk ⊂ Ea . Now, we split


X X X X X
(2.16) µ(K) = µ(K) + µ(K).
K∈S 2 : n′ <a K∈S 2 ′ : n′ ≥a K∈S 2 ′ :
j n n
K(Ink K(Ink
j
K(Ink
j

j′ j′
Take K = Inℓ′ as in the first sum. Note we must have Jnℓ′ ∩ Injk = ∅. Indeed, Injk is contained in Ea ,
j′ j′
while Jnℓ′ nontrivially intersects Enc ′ +1 , which is disjoint from Ea when n′ < a. Therefore, distD (Inℓ′ , Injk ) ≤
j′ j′
distD (Jnℓ′ , Inℓ′ ) ≤ N + 2, so the first sum in (2.16) is finite and in fact is bounded by CN µ(Injk ). The second
SPARSE DOMINATION FOR BALANCED MEASURES 19

summation may be estimated as follows, using the pairwise disjointness of each collection Sn2′ and our choice
of a:

X X j′
X 1 X j′ j′
ˆ
µ(Inℓ′ ) ≤ C1 (Inℓ′ , Jnℓ′ ) f1 dµ
2n′ Inℓ′
j′
n′ ≥a j′ 2
′ n ≥a j′ 2
Inℓ′ ∈Sn ′: Inℓ′ ∈Sn ′:
jℓ′ jℓ′
j j
In′ (Ink In′ (Ink
X 1 ˆ
. f1 dµ

2n′ Injk
n ≥a
X 1
≤ n′
2a+1 µ(Injk )

2
n ≥a

. µ(Injk ).

This bound proves the Carleson packing condition for I and completes the proof. 

Remark 2.11. The sparse domination given in Theorem 2.3 depends on the complexity in an essential
way. To prove this, we next show that a complexity 2 Haar shift cannot be dominated by a complexity 1
sparse form by a straightforward adaptation of Proposition 1.1. Define
X
T f := hf, hI ihI−− , f ∈ L2 (µ).
I∈D

Fix 0 < η < 1 and let µ be as in Proposition 1.1. We will show there exist two sequences of compactly
supported functions {fj }, {gj } ⊂ L2 (µ) such that for all η-sparse families S ⊂ D

(2.17) |hT fj , gj i| & j(|AS (fj , gj )| + |CS1 (fj , gj )|).

For j ≥ 2, we take fj = 1Ij−1


b and gj = 1Ij+1
b . As a preliminary, we notice fj and gj have disjoint supports
and if I ∈ D intersects both supp(fj ) and supp(gj ) non-trivially, we must have I = Ik with 0 ≤ k ≤ j − 2.
We also observe

b b
µ(Ij−1 ) µ(Ij+1 )
(2.18) M1 (fj )(x) . , M(gj )(x) = , x ∈ Ikb for 0 < k ≤ j − 3;
µ(Ik−1 ) µ(Ik−1 )

b
µ(Ij+1 )
(2.19) M1 (fj )(x) . 1, M(gj )(x) = , x ∈ Ikb for j − 2 ≤ k ≤ j;
µ(Ik−1 )

b
µ(Ij−1 )
(2.20) M1 (fj )(x) . b
, M(gj )(x) = 1, x ∈ Ij+1 ;
µ(Ij−2 )

b b
µ(Ij−1 ) µ(Ij+1 )
(2.21) M1 (fj )(x) . , M(gj )(x) = , x ∈ Ikb for k > j + 1
µ(Ij−2 ) µ(Ij )

Using the sparsity of S and equations (2.18) through (2.21), we can estimate the sums on the right hand
side of (2.17) directly:
20 JOSÉ M. CONDE ALONSO, JILL PIPHER, AND NATHAN A. WAGNER

X X p p
hfj iI hgj iI µ(I) + hfj iJ hgj iK m(J) m(K)
I∈S J,K∈S:
distD (J,K)=3,
J∩K=∅
p p
X X m(J) m(K)
. hfj iI hgj iI µ(EI ) + hfj iJ hgj iK µ(EK )
µ(K)
I∈S J,K∈S:
distD (J,K)=3,
J∩K=∅
ˆ
. M1 fj · Mgj dµ
[0,1]
∞ ˆ
X
≤ M1 fj · Mgj dµ
k=1 Ikb
j−3
! ! j
!
X b
µ(Ij−1 ) b
µ(Ij+1 ) X b
µ(Ij+1 )
b
. · · µ(Ik ) + · µ(Ikb )
µ(Ik−1 ) µ(Ik−1 ) µ(Ik−1 )
k=1 k=j−2
! ∞
! !
b
µ(Ij−1 ) X b
µ(Ij−1 ) b
µ(Ij+1 )
b
+ · µ(Ij+1 ) + · · µ(Ikb )
µ(Ij−2 ) µ(Ij−2 ) µ(Ij )
k=j+2
!  
j−3 ∞
1 X 1 1 1 X 1 . 1.
. 4
1 + 3 + 3 + 2 2
j j j j k j3
k=1 k=j+2

Now we turn to a lower bound for |hT fj , gj i|. Given I ∈ D, denote by SI the simple shift SI f = hf, hI ihI−− .
Then, by definition and using the support properties of fj and gj , we have hSI fj , gj i = hfj , hI i·hhI−− , gj i 6= 0
if and only if I = Ik with 0 ≤ k ≤ j − 2. Therefore, we may write
j−2
X j−2
X
hT fj , gj i = hSIk fj , gj i = hfj , hIk i · hhIk+2 , gj i.
k=0 k=0

For the last term in the sum, computations similar to Proposition 1.1 give
q
1
hfj , hIj−2 i · hhIj , gj i = m(Ij )m(Ij−2 ) ∼ 2 .
j
As before, all the other terms in the summation will similarly be positive, so (2.17) follows and the proof is
complete. Of course, one can adapt this argument to higher complexities in an obvious way. The key point
is that the higher complexity maximal function MN only needs to be applied to one of the functions fj , gj ,
which leads to a tighter upper bound than if it were applied to both.

Remark 2.12. We can obtain a similar result as Remark 2.11 for the sparse domination of the maximal
function M N in Theorem 2.9. In particular, we can show that it is impossible to dominate hM N f, gi by a
complexity N ′ sparse form where N ′ < N . The proof is very similar to the Haar shift case.

2.4. Lp and weak-type estimates. We can now turn to estimates for modified sparse forms. We
start with an Lp one.

Lemma 2.13. Let S be sparse and N ∈ N. For all nonnegative f1 ∈ Lp (µ) and f2 ∈ Lp (µ), there holds

CSN (f1 , f2 ) .p kf1 kLp (µ) kf2 kLp′ (µ) .

Proof. Fix an integer j satisfying 1 ≤ j ≤ N ′ . It clearly suffices to prove


X q
hf1 iI hf2 icj (I) m(I)m(cj (I)) . kf1 kLp (µ) kf2 kLp′ (µ) ,
I∈S:
cj (I)∈S
SPARSE DOMINATION FOR BALANCED MEASURES 21

which we do as follows:

X q X ′
hf1 iI hf2 icj (I) m(I)m(cj (I)) . hf1 iI hf2 icj (I) m(I)1/p m(cj (I))1/p
I∈S: I∈S:
cj (I)∈S cj (I)∈S
!1/p !1/p′
X X ′
. (hf1 iI )p µ(I) (hf2 iJ )p µ(J)
I∈S J∈S
!1/p !1/p′
X X
p p′
. (hf1 iI ) µ(EI ) (hf2 iJ ) µ(EJ )
I∈S J∈S
≤ kMf1 kLp (µ) kMf2 kLp′ (µ)
≤ kf1 kLp(µ) kf2 kLp′ (µ) .

where we have used the fact that µ is balanced in the first step. This finishes the proof. 

Proposition 2.14. Let S be sparse and N ∈ N. Then

sup sup inf′ sup CSN (|f1 |, |f2 |) < ∞.


f1 : G⊂R G: f2 :|f2 |≤1G′
kf1 kL1 (µ) ≤1 µ(G)≤2µ(G′ )

Proof. Let G be an arbitrary Borel set, and f1 with kf1 kL1 (µ) ≤ 1. We define the following set:

 
C0
H= x∈R: MN
D f1 (x) >
µ(G)

where we will choose the value of C0 momentarily. By Lemma 2.8, we get

µ(G) µ(G)
µ(H) ≤ kMN
D kL1 (µ)→L1,∞ (µ) ≤ ,
C0 2

′ ′
after choosing C0 = 2kMN D kL1 (µ)→L1,∞ (µ) . Take G = G \ H. The above estimate shows µ(G) ≤ 2µ(G ), so

G is a contender in the infimum. For any f2 with |f2 | ≤ 1G′ ., it is enough to prove

X p q
hf1 iI hf2 icj (I) m(I) m(cj (I)) . 1
I∈S:
cj (I)∈S

for each j ∈ {1, · · · , N ′ }. Let I ∈ S be such that cj (I) ∈ S. If we have

p
m(I)m(cj (I)) C0
hf1 iI > ,
µ(cj (I)) µ(G)

then by definition cj (I) ⊂ H. But since f2 is supported in H c , in this case we have hf2 icj (I) = 0. Therefore,
we may assume without loss of generality that if I ∈ S is such that cj (I) ∈ S, then we have the estimate

µ(cj (I)) C0
(2.22) hf1 iI ≤ p .
m(I)m(cj (I)) µ(G)
22 JOSÉ M. CONDE ALONSO, JILL PIPHER, AND NATHAN A. WAGNER

We then estimate, using Cauchy-Schwarz and the L2 bound for the ordinary maximal function:
 1/2
  !1/2
X q  X X
 2 m(cj (I)) 
 2
hf1 iI hf2 icj (I) m(I)m(cj (I)) .  (hf1 iI ) m(I) (hf2 iJ ) µ(J)
µ(cj (I)) 
I∈S: I∈S: J∈S
cj (I)∈S cj (I)∈S
 1/2
  !1/2
 X m(I)m(cj (I))  X
.
 (hf1 iI )2 µ(cj (I))

2
(hf2 iJ ) µ(EJ )
µ(cj (I))2
I∈S: J∈S
cj (I)∈S
 1/2
 
 X m(I)m(cj (I)) 
.
 (hf1 iI )2 µ(Ecj (I) )
 µ(G)1/2 .
µ(cj (I))2
I∈S:
cj (I)∈S

We will now show


 1/2
 
 X m(I)m(cj (I)) 
 (hf1 iI )2 µ(Ecj (I) ) . µ(G)−1/2 ,
 µ(cj (I))2 
I∈S:
cj (I)∈S
which is enough to conclude. Let
X
fe = cb1 (I, cj (I)) hf1 iI 1Ecj (I) ,
I∈S:
cj (I)∈S

and notice that we have


 1/2
 
 X m(I)m(cj (I)) 
kfekL2 (µ) ∼ 
 (hf1 iI )2 µ(Ecj (I) )
 .
µ(cj (I))2
I∈S:
cj (I)∈S

Moreover, we have the L∞ estimate, using (2.22):


C0
kfekL∞ (µ) . .
µ(G)
Also notice that for almost every x ∈ R, we have
|fe(x)| . MN f1 (x).
Therefore, we can write using the distribution function, the weak-type estimate for MN again, and the L1
normalization of f1 :
ˆ ∞
kfek2L2 (µ) = 2 λµ({|fe| > λ}) dλ
0
ˆ C
µ(G)
. λµ({x : MN
D f1 (x) > λ}) dλ.
0
ˆ C′
µ(G)
. kf1 kL1 (µ) 1 dλ
0
kf1 kL1 (µ) 1
. ≤ ,
µ(G) µ(G)
proving the required estimate. 
Remark 2.15. As a consequence of Lemma 2.13, Proposition 2.14 and the corresponding ones for AS ,
we recover the known results for the Lp bounds and weak-type estimates for X and other Haar shifts (see
[18]).
SPARSE DOMINATION FOR BALANCED MEASURES 23

Remark 2.16. We leave as an interesting open question whether pointwise sparse estimates can be
obtained, instead of the bilinear ones proved in this section.

3. Weighted Estimates
We now consider weights w on R with respect to a fixed balanced measure µ. In what follows, we write
1
ˆ
hf iw
I = f (x)w(x) dµ(x),
w(I) I
and Mw to indicate the dyadic Hardy-Littlewood maximal operator with respect to the measure wdµ. We
now turn to natural question of weighted Lp estimates for operators T ∈ HS(s, t). We first define the
appropriate weight classes and quantities which reflect the complexity of the dyadic shift.

3.1. A new class of weights. For p ∈ (1, ∞), we say w ∈ Abp (µ) if

[w]Abp (µ) := sup cbp (I, J)hwiI hw1−p ip−1
J < ∞.
I,J∈D:
b b
J∈{I,I− ,I+ } or
b b
I∈{J− ,J+ }

If p = 1, we say w ∈ Ab1 (µ) if


[w]Ab1 (µ) := sup cb1 (I, J)hwiI kw−1 kL∞ (J) ≤ C.
I,J∈D:
b b
J∈{I,I− ,I+ } or
b b
I∈{J− ,J+ }

Remark 3.1. If w ∈ Abp (µ), [w]Ap (µ) ≤ [w]Abp (µ) . If µ is dyadically doubling, Ap (µ) = Abp (µ). Proposition
1.3 shows that, for general balanced measures, the inclusion Abp (µ) ⊂ Ap (µ) is proper. As one would expect,

we have the duality relationship w ∈ Abp (µ) if and only if w1−p ∈ Abp′ (µ).

We are going to use a different quantitative characterization of the classes Abp (µ) that will be useful to
study the boundedness of MN . If p ∈ (1, ∞) and N ∈ N, we set
 
b 1−p′ p−1
[w]ANp (µ)
:= sup c p (I, J) hwiI hw iJ .
I,J∈D:
0≤distD (J,K)≤N +2

For p = 1, we set
[w]AN
1 (µ)
:= sup cb1 (I, J)hwiI kw−1 kL∞ (J) .
I,J∈D:
distD (I,J)≤N +2

Remark 3.2. [w]AN


1 (µ)
< ∞ if and only if there exists C such that

MN w(x) ≤ Cw(x).
Moreover, [w]AN
1 (µ)
is equal to the infimum of all such C.
Lemma 3.3. Let µ be balanced. For N ∈ N and 1 ≤ p < ∞,
 2N −1
[w]Abp (µ) . [w]AN
p (µ)
. N [w] b
Ap (µ) .

In particular, [w]AN
p (µ)
< ∞ exactly when [w]Abp (µ) < ∞.

Proof. We use induction on N . For the base case N = 1, it is obvious that [w]Abp . [w]A1p (µ) . Conversely,
if w ∈ Abp (µ) and I ⊂ J, for p > 1 we have
   
m(I)p/2 m(J)p/2 1−p′ p−1 m(I)p/2 m(J)p/2 ′

p−1
hwiI hw iJ ≤ p
hwiJ hw1−p ip−1
J
µ(J)µ(I) µ(I)
. [w]Abp ,
24 JOSÉ M. CONDE ALONSO, JILL PIPHER, AND NATHAN A. WAGNER

An entirely analogous argument works when J ⊂ I. The only other case to consider is when J = I b , in
which case we have the control:
 
m(I)p/2 m(I b )p/2 ′ ′
hwiI hw1−p ip−1
Ib
. hwiIbhw1−p ip−1 .
µ(I b )µ(I)p−1 Ib

The case p = 1 can be dealt with with straightforward modifications. This completes the proof of the base
case.
Now, assume that the statement holds for some positive integer j. The inequality
[w]Ajp (µ) ≤ [w]Aj+1
p (µ)

is immediate. Additionally, it is enough to consider disjoint intervals I, J when bounding [w]Aj+1


p (µ) , so
1 b and µ(J) < µ(Jb). For
suppose I, J are disjoint with distD (I, J) = j + 3. We may assume µ(I) < 2 µ(I) 1
2
1 b b
example, if µ(I) ≥ µ(I), then in fact µ(I) ∼ µ(I) and we would have the control
2

  !
m(I)p/2 m(J)p/2 ′ b p/2 m(J)p/2
m(I) ′
(3.1) hwiI hw1−p ip−1
J . hwiIbhw1−p ip−1
J ≤ [w]Ajp (µ) ,
µ(J)µ(I)p−1 µ(J)µ(I)b p−1

b J) = j + 2. Therefore, under this assumption we have m(I) ∼ µ(I) and m(J) ∼ µ(J), since
since distD (I,
b ∼ min{µ(I), µ(I b )} for any dyadic interval I and µ is balanced. Let K denote the minimal common
m(I)
ancestor of I and J. It is not difficult to see that either I (2) ⊂ K or J (2) ⊂ K. Assume without loss of
generality that I (2) ⊂ K. We consider two cases. If µ(Ibb ) ≥ 21 µ(I (2) ), then we have
b ∼ m(I (2) ) ∼ m(I) ∼ µ(I),
µ(I)
and we can then use the doubling argument as in (3.1). Otherwise, µ(Ibb ) ∼ m(Ibb ) and we may estimate by
Hölder’s inequality:

   
m(I)p/2 m(J)p/2 ′ m(I)p/2 m(J)p/2 ′ ′
hwiI hw1−p ip−1
J ≤ hwiI hw1−p ip−1 hwiIbb hw1−p ip−1
J
µ(J)µ(I)p−1 µ(J)µ(I)p−1 Ibb
" ! #
m(I)p/2 m(Ibb )p/2 ′
∼ hwiI hw1−p iIp−1bb
µ(Ibb )µ(I)p−1
" ! #
m(Ibb )p/2 m(J)p/2 ′ p−1
× hwiIbb hw1−p iJ
µ(J)µ(Ibb )p−1
≤ [w]2Aj (µ) ,
p

since distD (I, Ibb ) = 3 ≤ j + 2 and distD (Ibb , J) = j + 2. This completes the induction. 
Remark 3.4. Regardless of the sharpness of the inequalities in the statement of Lemma 3.3, it is probably
not the case that [w]Ajp (µ) ∼ [w]Akp (µ) with implicit constants independent of w, j, k. More compelling
evidence to support the fact that the characteristics are not quantitatively equivalent comes from the sparse
domination in Theorems 2.3 and 2.9. The sparse domination depends in an essential way on the complexity.
Remark 2.11 shows that, in general, a complexity one sparse form CS1 cannot possibly dominate a dyadic
shift with complexity 2, and so on for higher complexities. Therefore, obtaining sharp quantitative weighted
estimates for dyadic shift operators defined with respect to balanced measures remains an open problem,
but one would expect the quantitative bounds to depend on the complexity of the operator.
The proof of our next result justifies the introduction of the quantities [w]AN
p (µ)
.

Proposition 3.5. Let p ∈ (1, ∞) and N ∈ N. The operator MN is bounded on Lp (wdµ) if and only if
w ∈ Abp (µ). MN is bounded from L1 (wdµ) to L1,∞ (wdµ) if and only if w ∈ Ab1 (µ).
Proof. We apply Lemma 3.3 to use the quantity [w]AN
p (µ)
instead of [w]Abp (µ) . To begin with, fix

1 < p < ∞ and suppose that kMN kLp (wdµ)→Lp (wdµ) = C. Let I ∈ D be arbitrary. Suppose that w1−p
SPARSE DOMINATION FOR BALANCED MEASURES 25


is locally integrable (otherwise, consider (w + ε)1−p and use a limiting argument). Let J be any interval

satisfying 0 ≤ dist(I, J) ≤ N + 2 and take f = w1−p 1J . A computation shows that
ˆ 1/p

kf kLp(wdµ) = w1−p dµ .
J

Using the above and our hypothesis we estimate:


ˆ 

Cp w1−p dµ ≥ kMN f kpLp (wdµ)
J
ˆ  p

≥ [cb1 (J, I)]p hw1−p iJ wdµ
I
ˆ  p
b µ(J) ′
= cp (I, J) hw1−p iJ wdµ.
µ(I) I
Simple rearrangement then yields

cbp (I, J)hwiI hw1−p ip−1
J ≤ Cp,
which means [w]AN p (µ)
< ∞, since I, J were arbitrary. We prove the necessity when p = 1 with an argument
similar to [20]. Put now
kMN kL1 (wdµ)→L1,∞ (wdµ) = C,
and let I, J ∈ D satisfy distD (I, J) ≤ N + 2. Let ε > 0 be arbitrary and let

A = {x ∈ J : w(x) < ess inf J w + ε}.

Note that µ(A) > 0. Take f = 1A . Then, by definition we have


 
µ(A) b
I ⊂ x : MN f (x) ≥ c1 (J, I) .
µ(J)
It then follows from the supposed weak-type estimate that
µ(J)
ˆ
w(I) ≤ C b w(x)dµ(x)
c1 (J, I)µ(A) A
µ(J)
≤C b (ess inf J w + ε).
c1 (J, I)
Since ε was arbitrary and C is independent of ε, simple rearrangment yields

cb1 (I, J)hwiI ≤ C(ess inf J w),

which is precisely what we wanted to show.


We next turn to the sufficiency of the condition [w]AN p (µ)
< ∞. For p > 1, it is immediately given by
a weighted bound for the sparse form, which we postpone to the proof of Corollary 3.9, which is obtained
independently from the maximal function, together with Theorem 2.9. When p = 1, we argue like in Lemma
2.8. Suppose w ∈ Ab1 (µ) and take f ∈ L1 (wdµ). Put Eλ = {x : MN f (x) > λ}, cover Eλ by disjoint maximal
intervals {Jj } with corresponding averaging intervals {Ij }. Select the maximal dyadic intervals from {Ij },
label them {Ijk }, and define the index sets Ak , Bk as in the proof of Lemma 2.8. We may write

X
w(Eλ ) ≤ hwiJj µ(Jj )
j=1
X∞ X ∞ X
X
= hwiJj µ(Jj ) + hwiJj µ(Jj )
k=1 j∈Ak \Bk k=1 j∈Bk

:= (I) + (II).
26 JOSÉ M. CONDE ALONSO, JILL PIPHER, AND NATHAN A. WAGNER

We control (I) as follows, noticing that cb1 (Jj , Ij )hwiJj ≤ MN w(x)1Ij (x) if x ∈ Ij :

X X ∞
1X X ˆ
hwiJj µ(Jj ) . hwiJj cb1 (Jj , Ij ) f dµ
λ Ij
k=1 j∈Ak \Bk k=1 j∈Ak \Bk

1X X ˆ
≤ f · MN w dµ
λ Ij
k=1 j∈Ak \Bk
∞ ˆ
X
(CN + 1)[w]AN
1 (µ)
≤ f wdµ
λ Ijk
k=1
(CN + 1)[w]AN
1 (µ)
≤ kf kL1 (wdµ) .
λ
For (II), we can assume without loss of generality Jj ( Ijk for all j ∈ Bk . We then have
∞ X
X ∞
X
w(Jj ) ≤ hwiIjk µ(Ijk )
k=1 j∈Bk k=1
X∞
1
ˆ
. hwiIjk f dµ
λ Ijk
k=1
X∞ ˆ
1
≤ f (Mw) dµ
λ Ijk
k=1
∞ ˆ
X
[w]AN
1 (µ)
≤ f wdµ
λ Ijk
k=1
[w]AN
1 (µ)
. kf kL1 (wdµ) .
λ


3.2. Haar Shifts. We now turn to weighted estimates for Haar shift operators. For convenience in
what follows, we prove all weighted estimates for p = 2. However, all of the proofs carry over to the other
values of p with minor modifications. We first record the following standard lemma, which is an analog of
an A∞ property we will need in the sequel.

Lemma 3.6. Let w ∈ Ap (µ). Then for any I ∈ D and EI ⊂ I with µ(EI ) ≥ η(µ(I)) for some η ∈ (0, 1),

ηp
w(EI ) ≥ w(I).
[w]Ap (µ)

Proposition 3.7. Let S be sparse and N ∈ N. Then


(p−1)2 +1
1/p
kCSN kLp (wdµ)→Lp (wdµ) .N,p [w]Ap(p−1)
p (µ)
[w]AN (µ) .
p

Proof. We only prove the result for p = 2, and we proceed by duality: we need to show that for all
bounded, compactly supported nonnegative f1 , f2 ∈ L2 (wdµ),
1/2
|CSN f1 , wf2 i| . [w]A2 (µ) [w]AN (µ) kf1 kL2 (wdµ) kf2 kL2 (wdµ) .
2

As in the proof of Theorem 2.13, we only need to show the following estimate:
X q
1/2
hf1 iI hwf2 icj (I) m(I)m(cj (I)) ≤ C[w]A2 (µ) [w]AN (µ) kf1 kL2 (w) kf2 kL2 (w) .
2
I∈S:
cj (I)∈S
SPARSE DOMINATION FOR BALANCED MEASURES 27

We estimate as follows:
X q X −1
q
−1
hf1 iI hwf2 icj (I) m(I)m(cj (I)) = hwf1 iw
I hf2 iw
cj (I) hw iI hwicj (I) m(I)m(cj (I))
I∈S: I∈S:
cj (I)∈S cj (I)∈S
s
X −1 m(I)m(cj (I)) 1/2 −1
∼ hwf1 iw
I hf2 iw
cj (I) hw−1 iI hwicj (I) (w (I))1/2 w(cj (I))1/2
µ(I)µ(cj (I))
I∈S:
cj (I)∈S
1/2
X −1
−1
. [w]AN (µ) hwf1 iw
I hf2 iw
cj (I) (w (I))1/2 w(cj (I))1/2
2
I∈S:
cj (I)∈S
!1/2 !1/2
1/2
X −1 X
. [w]AN (µ) (hwf1 iw
I )2 w−1 (I) (hf2 iw 2
J ) w(J)
2
I∈S J∈S
!1/2 !1/2
1/2
X −1 X
. [w]A2 (µ) [w]AN (µ) (hwf1 iw
I )2 w−1 (EI ) (hf2 iw 2
J ) w(EJ )
2
I∈S J∈S
1/2 −1
≤ [w]A2 (µ) [w]AN (µ) kMw (f1 w)kL2 (w−1 dµ) kMw (f2 )kL2 (wdµ)
2
1/2
. [w]A2 (µ) [w]AN (µ) kf1 wkL2 (w−1 dµ) kf2 kL2 (wdµ)
2
1/2
= [w]A2 (µ) [w]AN (µ) kf1 kL2 (wdµ) kf2 kL2 (wdµ) ,
2

where we used Lemma 3.6 in the fifth line. This completes the proof. 

We can also obtain a weighted weak-type estimate for the sparse form when p = 1.

Proposition 3.8. Let S be sparse and N ∈ N. Then

kCSN kL1 (wdµ)→L1,∞ (wdµ) .N [w]2AN (µ) .


1

Proof. It is enough to show

sup sup inf′ sup CSN (|f1 |, w|f2 |) . [w]2AN (µ) .


G: 1
f1 : G⊂R f2 :|f2 |≤1G′
kf1 kL1 (wdµ) ≤1 w(G)≤2w(G′ )

We follow the general scheme of the proof of Theorem 2.14. Let G be an arbitrary Borel set, and f1 with
kf1 kL1 (wdµ) ≤ 1. Define
 
N C0
H = x ∈ R : M f1 (x) > ,
w(G)
where C0 is some sufficiently large positive constant. By Proposition 3.5, w(H) ≤ w(G)/2 as long as
C0 ≥ C0′ [w]AN1 (µ)
, where C0′ depends only on µ and N . As before, we take G′ = G \ H. Let f2 be arbitrary
with |f2 | ≤ 1G′ . It is enough to show, for 1 ≤ j ≤ CN , that we have
X p q
hf1 iI hwf2 icj (I) m(I) m(cj (I)) . [w]2AN (µ)
1
I∈S:
cj (I)∈S

By the same argument as the proof of Proposition 3.5, we may assume without loss of generality that if
I ∈ S is such that cj (I) ∈ S, then we have the estimate

µ(cj (I)) C0′ [w]AN


1 (µ)
hf1 iI ≤ p .
m(I)m(cj (I)) w(G)
28 JOSÉ M. CONDE ALONSO, JILL PIPHER, AND NATHAN A. WAGNER

We then estimate, using Cauchy-Schwartz inequality, Lemma 3.6, the fact that [w]A2 (µ) ≤ [w]AN
1 (µ)
, and the
weighted L2 bound for Mw :
X q
hf1 iI hwf2 icj (I) m(I)m(cj (I))
I∈S:
cj (I)∈S
 1/2
  !1/2
 X m(I)m(cj (I))  X
2 µ(J)
2
.
 (hf1 iI )2 w(cj (I))
 (hwf2 iJ )
µ(cj (I))2 w(J)
I∈S: J∈S
cj (I)∈S
 1/2
  !1/2
 X m(I)m(cj (I))  X
. [w]AN  (hf1 iI )2 w(Ecj (I) ) (hf2 iw 2
1 (µ)   J ) w(EJ )
µ(cj (I))2
I∈S: J∈S
cj (I)∈S
 1/2
 
 X m(I)m(cj (I)) 
. [w]AN  (hf1 iI )2 w(Ecj (I) ) w(G)1/2 .
1 (µ)  µ(cj (I))2 
I∈S:
cj (I)∈S

We will now show


 1/2
 
 X m(I)m(cj (I)) 
(∗) := 
 (hf1 iI )2 w(Ecj (I) )
 . [w]AN
1 (µ)
w(G)−1/2 .
µ(cj (I))2
I∈S:
cj (I)∈S

To see this, let


X
fe = cb1 (I, cj (I)) hf1 iI 1Ecj (I) ,
I∈S:
cj (I)∈S

which satisfies
C0
kfekL∞ (µ) . , |fe| . MN f1 , and kfekL2 (wdµ) ∼ (∗).
w(G)
This, combined with the weighted weak-type estimate for MN , and the L1 (wdµ) normalization of f1 , yields
ˆ ∞
kfek2L2 (wdµ) = 2 λw({|fe| > λ}) dλ
0
C[w]
AN (µ)
ˆ 1
w(G)
. λw({x : MN f1 (x) > λ}) dλ.
0
C[w]
AN (µ)
ˆ 1
w(G)
. [w]AN
1 (µ)
kf1 kL1 (wdµ) 1 dλ
0
[w]2AN (µ)
1
. ,
w(G)
proving the required estimate. 

Using Theorem 2.3 and the estimates for AS (which are analogous to those in the case of the Lebesgue
measure), we get the following consequence:
Corollary 3.9. Let N ∈ N. If T ∈ HS(s, t) and s + t ≤ N or T = MN , then
• If p > 1,
(p−1)2 +1
max{1, 1 } p(p−1) 1/p
kT kLp(wdµ)→Lp (wdµ) . [w]Ap (µ) p−1 + [w]Ap (µ) [w]AN (µ) .
p
SPARSE DOMINATION FOR BALANCED MEASURES 29

• The following weak-type estimate holds:


kT kL1(wdµ)→L1,∞ (wdµ) . [w]2AN (µ) .
1

Remark 3.10. One can follow similar arguments to obtain “one sided” weight conditions that, for
example, are sufficient for the boundedness on L2 (wdµ) (or Lp (wdµ)) of X or other particular operators
T ∈ HS(s, t). Indeed, one can check that the condition
(3.2) [w]A0,1 (µ) := sup cb2 (K, J)hwiK hw−1 iJ < ∞
2
J∈D:
b b
K∈{J,J− ,J+ }

is sufficient for the boundedness of X. However, such conditions are less natural in the sense that they
are asymmetric with respect to w and w−1 and are specialized to the form of the particular operator under
consideration.
We finally consider the necessity of the Abp (µ) condition.
Lemma 3.11. Let N ∈ N and p ∈ (1, ∞). Suppose there exists a constant C so that
max sup kT kLp(wdµ)→Lp (wdµ) ≤ C.
s,t∈N: T ∈HS(s,t)
s+t≤N

Let J, K ∈ D be such that 2 < distD (J, K) ≤ N + 2 and J ∩ K = ∅. Then


 ′

cbp (J, K)) hwiJ hw1−p ip−1
K . Cp,
where the implicit constant is independent of the particular intervals J, K.
Proof. By definition, J, K have a unique (minimal) common dyadic ancestor interval L, and there exist
nonnegative integers s, t satisfying s + t ≤ N so that distD (K, L) = s + 1 and distD (J, L) = t + 1. Without
loss of generality, we may suppose K = Lm n
s− and J = Lt− . Again, for simplicity we only prove the result for
−1
p = 2. We consider test functions f1 = w 1Ls− and f2 = 1Lnt− . Straightforward computations yield
m

kf1 k2L2 (wdµ) = w−1 (Lm−


s ), kf2 k2L2 (wdµ) = w(Ln−
t ).

Choose the sequence α to be the unique sequence of signs satisfying αIIsm ,Itn hf1 , hIsm ihhItn , wf2 i ≥ 0 for all
I ∈ D (and αIJ,K otherwise), and consider the operator T ∈ HS(s, t) defined by α. Then we estimate, using
the hypothesis :
Ckf1 kL2 (wdµ) kf2 kL2 (wdµ) ≥ h|T f1 , wf2 i|
X
= αIIsm ,Itn hf1 , hIsm ihhItn , wf2 i
I∈D
X
= |hf1 , hIsm i||hhItn , wf2 i|
I∈D
≥ |hf1 , hLms
i||hhLnt , wf2 i|
 −1 m   
p w (Ls− ) w(Lnt− )
= m(Lm )m(L n)
s t
µ(Lm s− ) µ(Lnt− )
s
m(J)m(K) 1/2
∼ hwiJ hw−1 iK |f1 kL2 (w) kf2 kL2 (w) .
µ(J)µ(K)
Squaring both sides and cancelling the L2 (wdµ) norms then produces the desired estimate. 
The above result is the remaining bit that allows us to characterize the class of weights for which Haar
shifts are Lp -bounded.
Corollary 3.12. Let p ∈ (1, ∞) and N ∈ N. There exists a constant C such that all operators
[
T ∈ HS(s, t)
s,t∈N:
s+t≤N
30 JOSÉ M. CONDE ALONSO, JILL PIPHER, AND NATHAN A. WAGNER

satisfy
kT kLp(wdµ)→Lp (wdµ) ≤ C
if and only if w ∈ Abp (µ).
Proof. One direction immediately follows from Corollary 3.9. For the other, we need to recall a relevant
fact concerning dyadic operators of total complexity 0. In particular, given a sequence α = {αJ }J∈D ∈ ℓ∞ ,
define the Haar multiplier Πα
D as follows:
X
ΠαD (f ) := αJ hf, hJ ihJ
J∈D

It is immediate that Πα 2
D is bounded on L (µ). It is also well-known that for any Borel measure µ, such
an operator is bounded on Lp (µ) for 1 < p < ∞. In fact, these operators admit sparse domination in the
usual sense, see [15]. Moreover, the collection of such operators with kαkℓ∞ ≤ 1 is uniformly bounded on
Lp (wdµ) if and only if w ∈ Ap (µ), see [25] for p = 2. For p 6= 2, the sufficiency of the Ap (µ) condition may
be obtained via the sparse domination, while the necessity may be obtained in the same way as p = 2, see
the proof of [25, Proposition 2.3]. This fact together with Lemma 3.11 completes the proof. 
We end this subsection observing that Theorem B follows from Proposition 1.3 and Corollary 3.12.
3.3. Higher-Dimensional Haar Shifts. It is a natural question whether all results in this paper can
be generalized to Haar shift operators defined on Rn , n > 1. A central point of departure from the one-
dimensional case occurs in constructing the Haar system (say according to the abstract definition in [18]),
as there are more degrees of freedom. We make a couple of remarks concerning the higher dimensional case.
Remark 3.13. The higher dimensional dyadic case for non-doubling measures was partially addressed
in [18], where sufficient conditions were given for the weak (1, 1) bound of dyadic shifts defined with respect
to two generalized Haar systems. In such a general context, we do not know if there are appropriate
generalizations of the modified sparse forms and weight classes so that our sparse domination or weighted
estimates remain true. It would be very interesting to investigate this question, as the higher dimensional
theory could be significantly more complicated. The general martingale setting, with possibly nonatomic
filtrations replacing the one generated by D, is also open.
Remark 3.14. In the special case that the Haar system {hQ }Q∈D consists of functions taking only two
values, we expect there to be a parallel theory based on the condition
sup khQ kL1 (µ) khQ′ kL∞ (µ) . 1,
Q

for Q, Q′ with prescribed dyadic distance, which is the right higher dimensional analog of balanced. This case
had already been identified in [18]. Although we have not checked the details, the two-value Haar functions
can be essentially regarded as one-dimensional objects, which is why we are confident this generalization will
work without much modification to the arguments.

References
[1] Jonas Azzam, Steve Hofmann, José María Martell, Svitlana Mayboroda, Mihalis Mourgoglou, Xavier Tolsa, and Alexander
Volberg, Rectifiability of harmonic measure, Geom. Funct. Anal. 26 (2016), no. 3, 703–728, DOI 10.1007/s00039-016-0371-x.
MR3540451
[2] Frédéric Bernicot, Dorothee Frey, and Stefanie Petermichl, Sharp weighted norm estimates beyond Calderón-Zygmund
theory, Anal. PDE 9 (2016), no. 5, 1079–1113, DOI 10.2140/apde.2016.9.1079. MR3531367
[3] D. L. Burkholder and R. F. Gundy, Extrapolation and interpolation of quasi-linear operators on martingales, Acta Math.
124 (1970), 249–304, DOI 10.1007/BF02394573. MR440695
[4] Léonard Cadilhac, José M. Conde-Alonso, and Javier Parcet, Spectral multipliers in group algebras and noncommutative
Calderón-Zygmund theory, J. Math. Pures Appl. (9) 163 (2022), 450–472, DOI 10.1016/j.matpur.2022.05.011 (English,
with English and French summaries). MR4438906
[5] José M. Conde-Alonso, BMO from dyadic BMO for nonhomogeneous measures, Pub. Mat (2020).
[6] José M. Conde-Alonso, Amalia Culiuc, Francesco Di Plinio, and Yumeng Ou, A sparse domination principle for rough
singular integrals, Anal. PDE 10 (2017), no. 5, 1255–1284, DOI 10.2140/apde.2017.10.1255. MR3668591
[7] José M. Conde-Alonso and Javier Parcet, Nondoubling Calderón-Zygmund theory: a dyadic approach, J. Fourier Anal.
Appl. 25 (2019), no. 4, 1267–1292, DOI 10.1007/s00041-018-9624-4. MR3977117
[8] José M. Conde-Alonso and Guillermo Rey, A pointwise estimate for positive dyadic shifts and some applications, Math.
Ann. (2016).
SPARSE DOMINATION FOR BALANCED MEASURES 31

[9] Amalia Culiuc, Francesco Di Plinio, and Yumeng Ou, Uniform sparse domination of singular integrals via dyadic shifts,
Math. Res. Lett. 25 (2018), no. 1, 21–42.
[10] Burgess Davis, On the integrability of the martingale square function, Israel J. Math. 8 (1970), 187–190, DOI
10.1007/BF02771313. MR268966
[11] Javier Duoandikoetxea, Francisco J. Martín-Reyes, and Sheldy Ombrosi, On the A∞ conditions for general bases, Math.
Z. 282 (2016), no. 3-4, 955–972.
[12] John B. Garnett and Peter W. Jones, BMO from dyadic BMO, Pacific J. Math. 99 (1982), no. 2, 351–371. MR658065
[13] Timo S. Hänninen, Equivalence of sparse and Carleson coefficients for general sets, Ark. Mat. 56 (2018), no. 2, 333–339,
DOI 10.4310/ARKIV.2018.v56.n2.a8. MR3893778
[14] Tuomas P. Hytönen, The sharp weighted bound for general Calderón-Zygmund operators, Ann. of Math. (2) 175 (2012),
no. 3, 1473–1506, DOI 10.4007/annals.2012.175.3.9. MR2912709
[15] Michael T. Lacey, An elementary proof of the A2 bound, Israel J. Math. 217 (2017), no. 1, 181–195, DOI 10.1007/s11856-
017-1442-x. MR3625108
[16] Andrei K. Lerner, On an estimate of Calderón-Zygmund operators by dyadic positive operators, J. Anal. Math. 121 (2013),
141–161, DOI 10.1007/s11854-013-0030-1. MR3127380
[17] Andrei K. Lerner and Fedor Nazarov, Intuitive dyadic calculus: the basics, Expo. Math. 37 (2019), no. 3, 225–265, DOI
10.1016/j.exmath.2018.01.001. MR4007575
[18] Luis Daniel López-Sánchez, José María Martell, and Javier Parcet, Dyadic harmonic analysis beyond doubling measures,
Adv. Math. 267 (2014), 44–93.
[19] Tao Mei, BMO is the intersection of two translates of dyadic BMO, C. R. Math. Acad. Sci. Paris 336 (2003), no. 12,
1003–1006, DOI 10.1016/S1631-073X(03)00234-6 (English, with English and French summaries). MR1993970
[20] Benjamin Muckenhoupt, Weighted norm inequalities for the Hardy maximal function, Trans. Amer. Math. Soc. 165 (1972),
207–226.
[21] Fedor Nazarov, Xavier Tolsa, and Alexander Volberg, On the uniform rectifiability of AD-regular measures with bounded
Riesz transform operator: the case of codimension 1, Acta Math. 213 (2014), no. 2, 237–321, DOI 10.1007/s11511-014-
0120-7. MR3286036
[22] F. Nazarov, S. Treil, and A. Volberg, Weak type estimates and Cotlar inequalities for Calderón-Zygmund operators on
nonhomogeneous spaces, Internat. Math. Res. Notices 9 (1998), 463–487, DOI 10.1155/S1073792898000312. MR1626935
[23] , The T b-theorem on non-homogeneous spaces, Acta Math. 190 (2003), no. 2, 151–239, DOI 10.1007/BF02392690.
MR1998349
[24] Stefanie Petermichl, Dyadic shifts and a logarithmic estimate for Hankel operators with matrix symbol, C. R. Acad. Sci.
Paris Sér. I Math. 330 (2000), no. 6, 455–460, DOI 10.1016/S0764-4442(00)00162-2.
[25] Christoph Thiele, Sergei Treil, and Alexander Volberg, Weighted martingale multipliers in the non-homogeneous setting
and outer measure spaces, Adv. Math. 285 (2015), 1155–1188.
[26] Xavier Tolsa, BMO, H 1 , and Calderón-Zygmund operators for non doubling measures, Math. Ann. 319 (2001), no. 1,
89–149, DOI 10.1007/PL00004432. MR1812821
[27] , A proof of the weak (1, 1) inequality for singular integrals with non doubling measures based on a Calderón-
Zygmund decomposition, Publ. Mat. 45 (2001), no. 1, 163–174. MR1829582
[28] , Painlevé’s problem and the semiadditivity of analytic capacity, Acta Math. 190 (2003), no. 1, 105–149, DOI
10.1007/BF02393237. MR1982794
[29] , Weighted norm inequalities for Calderón-Zygmund operators without doubling conditions, Publ. Mat. 51 (2007),
no. 2, 397–456. MR2334796
[30] Alexander Volberg and Pavel Zorin-Kranich, Sparse domination on non-homogeneous spaces with an application to Ap
weights, Rev. Mat. Iberoam. 34 (2018), no. 3, 1401–1414, DOI 10.4171/RMI/1029. MR3850292

José M. Conde Alonso


Departamento de Matemáticas
Universidad Autónoma de Madrid
C/ Francisco Tomás y Valiente sn
28049 Madrid, Spain
Email address: jose.conde@uam.es

Jill Pipher
Department of Mathematics
Brown University
151 Thayer Street
Providence, RI, 02912 USA
Email address: jill_pipher@brown.edu

Nathan A. Wagner
Department of Mathematics
Brown University
151 Thayer Street
Providence, RI, 02912 USA
Email address: nathan_wagner@brown.edu

You might also like