TRANSFORM: A REAL VARIABLE CHARACTERIZATION, I
MICHAEL T. LACEY, ERIC T. SAWYER,
CHUN-YEN SHEN, and IGNACIO URIARTE-TUERO
Abstract
Let andw be locally finite positive Borel measures on R which do not share a common point mass. Assume that the pair of weights satisfy a PoissonA2condition, and satisfy the testing conditions below, for the Hilbert transformH,
Z
I
H.1I/2dw .I /;
Z
I
H.w1I/2dw.I /;
with constants independent of the choice of intervalI. Then H./maps L2. /to L2.w/, verifying a conjecture of Nazarov, Treil, and Volberg. The proof has two com- ponents, aglobal-to-localreduction, carried out in this article, and an analysis of the localproblem, to be elaborated in a future Part II version of this article.
1. Introduction
Define a truncated Hilbert transform of a locally bounded signed measureby H;ı.x/WD
Z
<jyxj<ı
d.y/
yx; 0 < < ı:
Given weights (i.e., locally bounded positive Borel measures) andwon the real line R, we consider the followingtwo-weight norm inequalityfor the Hilbert transform,
sup
0<<ı
Z
R
ˇˇH;ı.f /ˇˇ2dwN2 Z
R
jfj2d; f 2L2. /; (1.1)
DUKE MATHEMATICAL JOURNAL
Vol. 163, No. 15, © 2014 DOI10.1215/00127094-2826690 Received 13 March 2012. Revision received 23 February 2014.
2010Mathematics Subject Classification. Primary 42A50; Secondary 47B38, 42B20.
Lacey’s work partially supported by National Science Foundation grant DMS-0968499, Simons Foundation grant 229596, and Australian Research Council grant ARC-DP120100399.
Sawyer’s work partially supported by the Natural Sciences and Engineering Research Council of Canada.
Uriarte-Tuero’s work partially supported by National Science Foundation grants DMS-0901524 and DMS- 1056965, Ministerio de Educación y Ciencias (Spain) grant MTM2010-16232 and MTM2009-14694-C02- 01, and by a Sloan Foundation Fellowship.
2795
where N is the best constant in the inequality, uniform over all truncations of the Hilbert transform kernel. Below, will write the inequality above askH.f /kL2.w/
NkfkL2.w/, that is, the uniformity over the truncation parameters is suppressed.
The primary question is to find a real variable characterization of this inequality, and the theorem below is an answer to the beautiful conjecture of Nazarov, Treil, and Volberg (see [29, p. 127]). Set
P .; I /WD Z
R
jIj
jIj2Cdist.x; I /2 .dx/;
which is approximately the Poisson extension of to the upper half-plane, evaluated at.xI;jIj/, wherexI is the center ofI.
THEOREM1.2
Let and w be locally finite positive Borel measures on the real line R with no common point masses. Then, the two-weight inequality (1.1) holds if and only if these three conditions hold uniformly over all intervalsI:
P .; I /P .w; I /A2; (1.3)
Z
I
ˇˇH.1I /ˇˇ2dwT2 .I /;
Z
I
ˇˇH.1Iw/ˇˇ2dT2w.I /: (1.4)
We have
N A1=22 CT DWH; (1.5)
whereA2andT are the best constants in the inequalities above.
It is well known (see [29]) that theA2condition is necessary for the norm inequal- ity, and the inequalities (1.4) are obviously necessary, thus the content of the theorem is the sufficiency of theA2and testing inequalities. In the present article we will carry out aglobal-to-localreduction in the proof of sufficiency, while the analysis of the localproblem will be made in a future article (Part II of this series; see [5]).
The Nazarov–Treil–Volberg conjecture has only been verified before under addi- tional hypotheses on the pair of weights, hypotheses which are not necessary for the two-weight inequality. The so-called pivotal condition of [29] is not necessary, as was proved in [9]. The pivotal condition is still an interesting condition: It is all that is needed to characterize the boundedness of the Hilbert transform, together with the maximal function in both directions. But, the boundedness of this triple of operators is decoupled in the two-weight setting (see [24]).
Our argument has these attributes. Certain degeneracies of the pair of weights must be addressed, such as the contribution made by Nazarov, Treil, and Volberg in their innovative 2004 paper [18] (see also [29]), which was further sharpened with the property ofenergyin [9], a crucial property of the Hilbert transform. This theme is fur- ther developed in the following pages, with the notion of functional energy explained in Section5.
The proof should proceed through the analysis of the bilinear formhH.f /; gwi, as one expects certain paraproducts to appear. Still, the paraproducts have no canoni- cal form, suggesting that the proof is highly nonlinear inf andg. The nonlinear point of view was initiated in [7] and is central to our work here. A particular feature of our arguments is a repeated appeal to certainquasiorthogonality arguments, providing (many) simplifications over prior arguments. For instance, we never find ourselves constructing auxiliary measures, and verifying that they are Carleson, a frequent step in many related arguments.
One can phrase a two-weight inequality question for any operator T, a ques- tion that became apparent with the foundational paper of Muckenhoupt [12] onAp
weights for the maximal function. Indeed, the case of Hardy’s inequality was quickly resolved by Muckenhoupt [11]. The maximal function was resolved by the second author in [26], as well as the fractional integrals, and, essential for this paper, Poisson integrals (see [27]). The latter paper established a result which closely paralleled the contemporaneousT 1theorem of David and Journé [1]. This connection, fundamental in nature, was not fully appreciated until the innovative work of Nazarov, Treil, and Volberg [14]–[16] in developing a nonhomogeneous theory of singular integrals. The two-weight problem for dyadic singular integrals was only resolved recently in [17].
Partial information about the two-weight problem for singular integrals in [21] was basic to the resolution of theA2conjecture in [3], and several related results (see [4], [6], [21], [22]). Our result is the first real variable characterization of a two-weight inequality for a continuous singular integral.
Interest in the two-weight problem for the Hilbert transform arises from its natu- ral occurrence in questions related to operator theory (see [20], [25]), spectral theory (see [20]), model spaces (see [23]), and analytic function spaces (see [10]). In the context of operator theory, Sarason posed the conjecture (see [2]) that the Hilbert transform would be bounded if the pair of weights satisfied the (full) PoissonA2con- dition. This was disproved by Nazarov [13]. Advances on these questions have been linked to a finer understanding of the two-weight question (see, e.g., [19], [20]) built upon Nazarov’s counterexample.
2. Dyadic grids and Haar functions 2.1. Choice of truncation
We have stated the main theorem withhard cutoffs in the truncation of the Hilbert transform. There are many possible variants in the choice of truncation; moreover, the proof of sufficiency requires a different choice of truncation.
Consider a truncation given by HQ˛;ˇ.f /.x/WD
Z
f .y/K˛;ˇ.yx/ .dy/;
whereK˛;ˇ.y/is chosen to minimize the technicalities associated with off-diagonal considerations. Specifically, setK˛;ˇ.0/D0, and otherwiseK˛;ˇ.y/is odd and for y > 0,
K˛;ˇ.y/WD 8ˆ ˆˆ ˆ<
ˆˆ ˆˆ :
˛y2C2˛ 0 < y < ˛;
1
y ˛yˇ;
y
ˇ2Cˇ2 ˇ < y < 2ˇ;
0 2ˇy:
This is aC1function on.0; 2ˇ/, and is Lipschitz, convex, and monotone on.0;1/.
We now argue that we can use these truncations in the proof of the sufficiency bound of our main theorem.
PROPOSITION2.1
If the pair of weights; wsatisfy theA2bound (1.3), then one has the uniform norm estimate with thehardtruncations (1.1) if and only if one has uniform norm estimate for thesmoothtruncations,
sup
0<˛<ˇ
HQ˛;ˇ.f /
wNkfk:
Indeed,jH˛;ˇ.f / QH˛;ˇ.f /jA˛.jfj/CAˇ.jfj/, where these last two operators aresingle-scaleaverages, namely,
A˛. /.x/D˛1 Z
.x3˛;xC3˛/
.y/ .dy/:
But, the (simple)A2 bound is all that is needed to provide a uniform bound on the operatorsA˛. /. So the proposition follows.
Henceforth we use the truncationsHQ˛;ˇ, and we suppress the tilde in the notation.
The particular choice of truncation is motivated by this off-diagonal estimate on the kernels.
PROPOSITION2.2
Suppose that2jxx0j<jxyj; then
K˛;ˇ.yx0/K˛;ˇ.yx/DCx;x0;y
x0x .yx/.yx0/; whereCx;x0;yD1; 2˛ <jxyj<1
2ˇ; (2.3)
and is otherwise positive and never more than4.
Proof
The assumptions imply thatyx0 andyx have the same sign. Assume without loss of generality that 0 < yx0< yx. If2˛ <jxyj< ˇ=2, it follows that
˛ <jx0yj< ˇ, and so by the definition, K˛;ˇ.yx0/K˛;ˇ.yx/D 1
yx0 1
yx D x0x .yx/.yx0/: And, in the general case, we havejdtdK˛;ˇ.t /j 4t2, so that
0K˛;ˇ.yx0/K˛;ˇ.yx/
Z yx yx0
4
t2dtD4 x0x .yx/.yx0/: 2.2. Dyadic grids
A collection of intervalsG is agridif for allG; G02G, we haveG\G02 ¹;; G; G0º.
By a dyadic grid we mean a gridD of intervals of R such that for each interval I 2D, the subcollection¹I02DW jI0j D jIjºpartitionsR, aside from endpoints of the intervals. In addition, the left and right halves ofI, denoted byI˙, are also inD. ForI2D, the left and right halvesI˙are referred to as thechildrenofI. We denote byDI the unique interval inD havingI as a child, and we refer toDI as theD-parentofI.
We will work with subsets F D. We say that I has F-parentFI DF if F 2F is the minimal element ofF that containsI.
2.3. Haar functions
Let be a weight onR, one that does not assign positive mass to any endpoint of a dyadic gridD. IfI 2D is such that assigns nonzero weight to both children ofI, the associated Haar function is
hIWD s
.I/ .IC/ .I /
I
.I/C IC
.IC/ :
In this definition, we are identifying an interval with its indicator function, and we will do so throughout the remainder of the paper. This is anL2. /-normalized function, and has-integral zero. For any dyadic interval I0, we have that ¹ .I0/1=2I0º [
¹hIWI 2D; II0ºis an orthogonal basis forL2.I0; /.
We will use the notationf .I /O D hf; hIi, as well as
If D hf; hIihIDICEICf CIEIf IEIf:
The second equality is the familiar martingale difference equality, and so we will refer toIf as amartingale difference. It implies the familiar telescoping identity EJf DP
IWIJEJIf.
For any function theHaar supportoff is the collection¹I2DW Of .I /¤0º.
2.4. Good-bad decomposition
With a choice of dyadic gridD understood, we say thatJ2D is.; r/-goodif and only if for all intervalsI2DwithjIj 2r1jJj, the distance fromJto the boundary ofeither childofI is at leastjJjjIj1.
For f 2L2. /, we set Pgood f DP
I2D Iis.; r/-good
If. The projection Pgoodw g is defined similarly. To make the two reductions below, one must make a random selection of grids, as is detailed in [9] and [29]. The use of random dyadic grids has been a basic tool since the foundational work of [14]–[16]. Important elements of the suppressed construction of random grids are the following.
(1) It suffices to consider a single dyadic gridD.
(2) For any fixed0 < < 1=2, we can choose an integerr sufficiently large so that it suffices to consider f such that f DPgood f, and likewise for g2 L2.w/. Namely, it suffices to estimate the constant below, for an arbitrary dyadic gridD,
ˇˇhHf; giw
ˇˇNgoodkfkkgkw;
where it is required thatf DPgood 2L2. /andgDPgoodw 2L2.w/.
That the functions are good is, at some moments, an essential property. We sup- press it in notation, however, taking care to emphasize in the text those places in which we appeal to the property of being good.
A reduction, using randomized dyadic grids, allows one the extraordinarily useful reduction in the next lemma. This is a well-known reduction, due to Nazarov, Treil, and Volberg, explained in full detail in the current setting in [18, Section 4]. Below, H is as in (1.5), the normalized sum of theA2and testing constants.
LEMMA2.4
For all sufficiently small, and sufficiently larger, this holds. Suppose that for any
dyadic gridD, such that no endpoint of an intervalI 2D is a point mass for or w,we have
ˇˇhHPgood f; Pgoodw giw
ˇˇHkfkkgkw: (2.5)
Then, the same inequality holds without the projectionsPgood , andPgoodw .
Inequality (2.5) should be understood as an inequality, uniform over the class of smooth truncations of the Hilbert transform. But, we can suppress this in the notation without causing confusion. The bilinear form only needs to be controlled for.; r/- good functionsf andg, goodness being defined with respect to a fixed dyadic grid.
Suppressing the notation, we writegoodfor.; r/-good, and it is always assumed that the dyadic gridD is fixed, and only good intervals are in the Haar support off and g, though is also suppressed in the notation.
3. The global-to-local reduction
The goal of this section is to reduce the analysis of the bilinear form in (2.5) to the local estimate, (3.4). It is sufficient to assume thatf andgare supported on an interval I0; by trivial use of the interval testing condition, we can further assume thatf andg are of integral zero in their respective spaces. Thus,f is in the linear span of (good) Haar functionshI forI I0, and similarly forg, and
hHf; giwD X
I;JWI;JI0
hHIf; wJgiw:
The argument is independent of the choice of truncation that implicitly appears in the inner product above.
The double sum is broken into different summands. Many of the resulting cases are elementary, and we summarize these estimates as follows. Define the bilinear form
Babove.f; g/WD X
IWII0
X
JWJIJ
EIJIf hHIJ; wJgiw;
where here and throughout, J I means J I and2rjJj jIj. In addition, the argument of the Hilbert transform,IJ, is the child ofI that containsJ, so thatIf is constant onIJ. DefineBbelow.f; g/in the dual fashion.
This set of dyadic grids that fail this condition have probability zero in standard constructions of the random dyadic grids.
LEMMA3.1
We have, with the notation of (1.5),
ˇˇhHf; giwBabove.f; g/Bbelow.f; g/ˇ
ˇHkfkkgkw:
This is a common reduction in a proof of aT 1theorem, and in the current context, it only requires goodness of intervals and the A2 condition. For a proof, one can consult [29] and [18]. The lemma is specifically phrased and proved in this way in [7, Section 8].
These definitions are needed to phrase the global-to-local reduction. The follow- ing definition depends upon the essential energy inequality (4.7) in Section 4.
Definition 3.2
Given any intervalF0, defineFenergy.F0/to be the maximal subintervalsFF0such that
P .F0; F /2E.w; F /2w.F / > 10C0H2 .F /;
whereE.w; F / is defined in (4.6) and whereC0 is the constant in Proposition4.8.
We have .S
¹F WF 2F.F0/º/.=10/.F0/.
Definition 3.3
LetI0 be an interval, and letS be a collection of disjoint intervals contained inI0. A functionf 2L20.I0; /is said to beuniform (with respect toS)if the following conditions are met.
(1) Each energy stopping intervalF 2Fenergy.I0/is contained in someS2S.
(2) The functionf is constant on each intervalS2S.
(3) For any intervalII0which is not contained in anyS2S,EIjfj 1.
We will say that gisweakly adapted to a functionf uniform with respect toS if JS for some intervalS 2S implies thathg; hwJiwD0. We will also say thatgis weakly adapted toS.
The constantLis defined as the best constant in thelocal estimate ˇˇBabove.f; g/ˇˇL®
.I0/1=2C kfk¯
kgkw; (3.4)
wheref; g are of mean zero on their respective spaces, supported on an intervalI0. Moreover,f is uniform andgis weakly adapted tof. The inequality above is homo- geneous ing, but notf, since the term .I0/1=2is motivated by the bounded averages property off.
THEOREM3.5 (Global-to-local reduction) We have
ˇˇBabove.f; g/ˇ
ˇ¹HCLºkfkkgkw: The same inequality holds for the dual formBbelow.f; g/.
A reduction of this type is a familiar aspect of many proofs of aT 1 theorem, proved by exploiting standard off-diagonal estimates for Calderón–Zygmund ker- nels, but in the current setting, it is a much deeper fact, a consequence of thefunc- tional energy inequality of Section 5. We make the following construction for an f 2L2.I0; /, of-integral zero. AddI0 toF, and set˛f.I0/WDEI0jfj. In the inductive stage, ifF 2F is minimal, add toF those maximal descendantsF0ofF such thatF02Fenergy.F /orEF0jfj 10˛f.F /. Then define
˛f.F0/WD
´˛f.F / EF0jfj< 2˛f.F /;
EF0jfj otherwise:
If there are no such intervalsF0, the construction stops. We refer toF and˛f./as Calderón–Zygmund stopping data forf, following the terminology of [7, Definition 3.5]. Their key properties are collected here.
LEMMA3.6
ForF and˛f./as defined above, the following hold.
(1) The dyadic intervalI0is the maximal element ofF. (2) For allI 2D,II0, we haveEIjfj 10˛f.FI /.
(3) The datum˛f is monotonic: IfF; F02F andFF0, then˛f.F /˛f.F0/.
(4) The collectionF is-Carleson in that X
F2FWFS
.F /2 .S /; S2D: (3.7) (5) We have the inequality
X
F2F
˛f.F /F
kfk: (3.8)
Proof
The first three properties are immediate from the construction. The fourth, the -Carleson property is seen this way. It suffices to check the property forS 2F.
Now, theF-children can be inFenergy.S /, which satisfy X
F02Fenergy.S /
.F0/ 1 10 .S /:
Otherwise, note that by choice of˛f./, we haveESjfj 2˛f.S /. These intervals F0 satisfyEF0jfj 10˛f.S /5ESjfj. These intervals satisfy the display above with1=10replaced by1=5. Hence, (3.7) holds.
For the final property, let G F be the subset at which the stopping values change. IfF 2F G, andGis theG-parent ofF, then˛f.F /D˛f.G/. Set
ˆGWD X
F2FWGFDG
F:
DefineGkWD ¹ˆG2kº, forkD0; 1; : : :. The-Carleson property implies integra- bility of all orders in-measure ofˆG. Using the third moment, we have .Gk/ 23k .G/. Then, estimate
X
F2F
˛f.F /F
2 D
X
G2G
˛f.G/ˆG
2
X1 kD0
.kC1/C11X
G2G
˛f.G/2k1Gk
2
X1 kD0
.kC1/2X
G2G
˛f.G/2k1Gk.x/
2
X1 kD0
.kC1/2 X
G2G
˛f.G/222k .Gk/
X
G2G
˛f.G/2 .G/kMfk2kfk2:
Note that we have used Cauchy–Schwarz in k at the step marked by an . In the step marked with, for each pointx, the nonzero summands are a (super)geometric sequence of scalars, so the square can be moved inside the sum. Finally, we use the estimate on the-measure ofGk, and compare to the maximal functionMf to com- plete the estimate.
We will use the notation
PFf WD X
I2DWFIDF
If; F2F;
and similarly forQwF, but rather than useFJin the definition, we usePFJ, defined to be the minimalF 2F withJF. Without this alternate definition, some delicate case analysis would be forced upon us. The inequality (3.8) allows us to estimate
X
F2F
®˛f.F / .F /1=2C kPFfk
¯kQwFgkw
h X
F2F
®˛f.F /2 .F /C kPFfk2¯ X
F2F
kQFwgk2wi1=2
kfkkgkw: (3.9)
We will refer to this as thequasiorthogonalityargument, and we remark that it only requires orthogonality of the projectionsQwFg. It is very useful.
LEMMA3.10 We have
ˇˇBabove.f; g/BFabove.f; g/ˇ
ˇHkfkkgkw; whereBFabove.f; g/WD X
F2F
Babove.PFf; QFwg/:
Proof
We apply the functional energy inequality of Section5. Observe thatf DP
F2FPFf and that
X
JWJI0
wJgD X
F2F
QwFg:
From the definition of Babove.f; g/, we can assume that g equals the sum above.
Therefore,
Babove.f; g/D X
F02F
X
F2F
Babove.PF0f; QwFg/:
In the sum above, we can also add the restriction that F0\F ¤ ;, for otherwise Babove.PF0f; QwFg/D0. For a pair of intervalsJ IJ, note that this implies that JFI; that is,PFJFI. Therefore, we can add the restrictionF F0. The case ofF0DF is the definition ofBFabove.f; g/, so that it suffices to estimate
X
F ;F02F F0F
Babove.PF0f; QwFg/: (3.11)
Observe that the functionsgF WDQwFg areF-adapted in the sense of Defini- tion5.1, and by constructionF satisfies the Carleson measure condition (3.7). We take these steps to apply the functional energy inequality. The argument of the Hilbert transform isIF, the child of I that contains F. WriteIF DF C.IF F /, and use linearity ofH. Note that by the standard martingale difference identity and the construction of stopping data,
ˇˇ ˇ X
IWIF
EIFIfˇˇˇ˛f.F /; F 2F:
Hence, invoking interval testing, ˇˇ
ˇX
F2F
X
IWIF
EIFIf hHF; gFiw
ˇˇ ˇ X
F2F
˛f.F /ˇˇhHF; gFiw
ˇˇ
H X
F2F
˛f.F / .F /1=2kgFkw: Quasiorthogonality bounds this last expression.
For the second expression, when the argument of the Hilbert transform isIFF, first note that
ˇˇ ˇ X
IWIF
EIFIf .IF F / ˇˇ
ˇˆWD X
F02F
˛f.F0/F0; F2F:
Therefore, by the definition ofF-adapted, the monotonicity property (4.3) applies, and yields
ˇˇ ˇ X
IWIF
EIFIf˝
H.IFF /; gF˛
w
ˇˇ
ˇ X
J2J.F /
P .ˆ; J /D x
jJj; J gFE
w; F 2F: Here,J.F /are the maximal good intervalsJF, andgF WDP
J2J.F /WJFjbg.J /j hwJ, so that every term has a positive inner product withx. The sum overF 2F of this last expression is controlled by functional energy, and by the property that kˆkkfk. This completes the bound for (3.11).
Proof of Theorem3.5
By Lemma3.10, it remains to controlBFabove.f; g/. Keeping the quasiorthogonality argument in mind, we see that appropriate control on the individual summands is enough to control it. For eachF 2F, letSF be theF-children ofF. Observe that the function
C ˛f.F /1
PFf (3.12)
is uniform onFwith respect toSF, for an appropriate absolute constantC. Moreover, the functionQFwgdoes not have any intervalJ in its Haar support strongly contained in an intervalS2SF. That is, it is weakly adapted to the function in (3.12). Therefore, by assumption,
ˇˇBabove.PFf; QwFg/ˇ ˇL®
˛F.F / .F /1=2C kPFfk¯
kQFwgkw:
The sum over F 2F of the right-hand side is bounded by the quasiorthogonality argument of (3.9).
4. Energy, monotonicity, and Poisson
Our theorem (Theorem 1.2) is particular to the Hilbert transform, and so depends upon special properties of it. They largely extend from the fact that the derivative of1=yis positive. The following monotonicity property for the Hilbert transform was observed in [7, Lemma 5.8], and is basic to the analysis of the functional energy inequality.
LEMMA4.1 (Monotonicity property)
LetKI be two intervals, and assume that does not have point masses at the end point ofI. Then, for any functiong2L2.I; w/, withw-integral zero, andˇ > 2jKj,
P
.KI /; ID x jIj; gE
wlim inf
˛#0
˝H˛;ˇ
.KI /
; g˛
w: (4.2)
Here,gDP
J0jbg.J0/jhwJ0 is a Haar multiplier applied tog. IfJ is a good interval, JI, then, for functiong2L2.J; w/, withw-integral zero, and signed measures andsupported onKI, withjj , we have
sup
0<˛<ˇ
ˇˇhH˛;ˇ; giwˇ
ˇP .; J /D x jJj; gE
w: (4.3)
The truncations enter into the formulation of the lemma, since they play a notable role here. We need this preparation.
LEMMA4.4
LetI andJ be two intervals which share an endpointa, at which neither norw have a point mass. Then,
sup
0<˛<ˇ
ˇˇhH˛;ˇI; Jiw
ˇˇA1=22 p
.I /w.J /: (4.5) Proof
IfjIj ' jJj, this inequality is the weak boundedness principle of [9, Section 2.2]. So, let us assume that10jIj<jJj. Then, it remains to bound
ˇˇ˝H˛;ˇI; .J n10I /˛
w
ˇˇ X1 nD11
.I /w.J\..nC1/InnI //
njIj .I /
jIj1=2P .w; I /1=2w.J /1=2A1=22 p
.I /w.J /:
This depends upon obvious kernel bounds, and an application of Cauchy–Schwarz to derive the Poisson term above.
Proof of Lemma4.1
By linearity, it suffices to prove (4.2) in the case ofgDhwI. The point is to separate the supports of the functions involved. SinceI does not have a point mass at the end point ofI, we have . I nI /#0as #1. It follows that we can fix a > 1sufficiently small so thatP . .KI /; I /'P . .K I /; I /, and one more condition that we will come back to. Then, for0 < ˛ < .1=2/. 1/jIj, we estimate as below, where xI is the center ofI,
˝H˛;ˇ
.K I /
; hwI˛
w
D Z
KI
Z
I
®K˛;ˇ.yx/K˛;ˇ.yxI/¯
hwI.x/w.dx/ .dy/
D Z
KI
Z
I
xxI
.yx/.yxJ/hwI.x/w.dx/ .dy/
P
.KI /; IDxxI
jIj ; hwIE
w:
We have subtracted the term, sincehwI has integral zero, then we applied (2.3) with Cx;xJ;yD1, as follows from our choices of˛andˇ. Then, note that.xxJ/hwI 0, so that we can pull out the Poisson term. The last line follows by our selection of sufficiently close to1. Then, the last condition needed is to select sufficiently close to1that, in view of (4.5),
sup
˛;ˇ
ˇˇ˝H˛;ˇ. InI /; hwI˛
w
ˇˇA1=22 p
. I nI / < cP
.KI /; IDxxI
jIj ; hwIE
w:
In the last line,c > 0is an absolute constant. This completes the proof of (4.2).
Turn to (4.3). The estimate (2.3) applies:
ˇˇhH˛;ˇ; giw
ˇˇDˇˇˇ Z
KI
Z
J
®K˛;ˇ.yx/K˛;ˇ.yxJ/¯
hwJ.x/w.dx/.dy/ˇˇˇ D
ˇˇ ˇ Z
KI
Z
J
Cx;xJ;y
.xxJ/
.yx/.yxJ/hwJ.x/w.dx/.dy/
ˇˇ ˇ:
But recall that0Cx;xJ;y4and that it equals1for˛sufficiently small. Moreover, yxandyxJ have the same sign, and.xxJ/hwJ.x/0. So an upper bound is obtained by passing fromto:
ˇˇhH˛;ˇ; giw
ˇˇ Z
KI
Z
J
.xxJ/
.yx/.yxJ/hwJ.x/w.dx/.dy/
'P .; J /D x jJj; hwJE
w:
The concept ofenergyis fundamental to the subject. For intervalI, define E.w; I /2WDEw.dx/I Ew.dxI 0/
.xx0/2 jIj2 D 2
w.I / X
JI
D x jIj; hwJE2
w: (4.6)
Now, consider theenergy constant, the smallest constant E such that this condition holds, as presented or in its dual formulation. For all dyadic intervalsI0, all partitions P ofI0 into dyadic intervals, we have
X
I2P
P .I0; I /2E.w; I /2w.I /E2 .I0/: (4.7) This was shown in [9, Proposition 2.11].
PROPOSITION4.8 For a finite constantC0,
E2C0¹A1=22 CTº2DC0H2: We will always estimateE byH. The proof is recalled here.
Proof
It suffices to consider the case of finite partitionsP ofI. We first prove a version of the energy inequality withholesin the argument of the Poisson. It follows from (4.2) that we can fix0 < ˛ < ˇsuch that
P
.I0I /; I2
E.w; I /2w.I /H˛;ˇ
.I0I /2L2.I; /; I2P: Then, using linearity and interval testing, we have
X
I2P
H˛;ˇ.I0/2L2.I; /H˛;ˇ.I0/2L2.I; /H2 .I0/
and
X
I2P
H˛;ˇ.I /2L2.I; /H2X
I2P
.I /H2 .I0/:
Then, by theA2 bound, we haveP .I; I /2E.w; I /2w.I / .I /, which we can sum over the partition. This completes the proof.
One should keep in mind that the concept of energy is related to the tails of the Hilbert transform. The energy inequality, and its multiscale extension to the functional energy inequality, show that the control of the tails is very subtle in this problem.
We also need the following elementary Poisson estimate from [29]; used occa- sionally in this argument, it is crucial to the proof of Lemma3.1.
LEMMA4.9
Suppose thatJII0, and thatJ is good. Then jJj21P
.I0I /; J
jIj21P
.I0I /; I
: (4.10)
Proof
We have dist.J; I0I / jJjjIj1, so that for anyx2I0I, we have jJj2
.jJj Cdist.x; J //2 jIj2 .jIj Cdist.x; I //2: Integrating this last expression, it follows that
jJj21P
.I0I /; J
D jJj21 Z
I0I
jJj
.jJj Cdist.x; J //2d jIj2
Z
I0I
1
.jJj Cdist.x; J //2d:
And this proves the inequality.
5. The functional energy inequality
We state an important multiscale extension of the energy inequality (4.7).
Definition 5.1
LetF be a collection of dyadic intervals. A collection of (good) functions¹gFºF2F
inL2.w/is said to beF-adaptedif, for allF 2F, the Haar support of the function gF is contained in¹JW PFJDFº.
Definition 5.2
LetF be the smallest constant in the inequality below, or its dual form. The inequality holds for all nonnegativeh2L2. /, all-Carleson collectionsF, and allF-adapted collections¹gFºF2F:
X
F2F
X
J2J.F /
P .h; J/ ˇˇ ˇD x
jJj; gFJE
w
ˇˇ
ˇFkhk
h X
F2F
kgFk2wi1=2
:
HereJ.F /consists of themaximalgood intervalsJF. Note that the estimate is universal inhandF, separately.
This constant was identified in [7], and is herein shown to be necessary from the A2and interval testing inequalities. Recall the definition ofH in (1.5).
THEOREM5.3
Assume thatF satisfies (3.7); then,F H.
The first step in the proof is the domination of the constantF by the best con- stant in a certain two-weight inequality for the Poisson operator, with the weights being determined bywandin a particular way. This is the decisive step, since there is a two-weight inequality for the Poisson operator proved by one of us in [27, Theo- rem 2]. It reduces the full norm inequality to simpler testing conditions, which are in turn controlled by theA2and Hilbert transform testing conditions.
5.1. The two-weight Poisson inequality Consider the weight
X
F2F
X
J2J.F /
PF ;Jw x
jJj 2
wı.xJ;jJj/: Here,PF ;Jw WDP
J0WJ0J;PFJDFwJ0. We can replacexbyxcfor any choice ofc we wish; the projection is unchanged. Andıqdenotes a Dirac unit mass at a pointqin the upper half-planeR2C. We prove the two-weight inequality for the Poisson integral
P.h /L2.R2
C;/Hkhk;
for all nonnegativeh. Above,P./denotes the Poisson extension operator to the upper half-plane, so that in particular
P.h / 2L2.R2
C;/D X
F2F
X
J2J.F /
P.h /
xJ;jJj2PF ;Jw x jJj
2
w;
wherexJ is the center of the intervalJ. The proof of Theorem5.3follows by duality.
Phrasing things in this way brings a significant advantage. The characterization of the two-weight inequality for the Poisson operator (see [27]) reduces the full norm inequality above to these testing inequalities. For any dyadic intervalI2D we have
Z
R2CP.I /2d.x; t /H2 .I /; (5.4) Z
RP.tI /b 2 .dx/A2 Z
b I
t2d.x; t /; (5.5) wherebI DI Œ0;jIjis the box overI in the upper half-plane and wherePis the dual Poisson operator
P.tbI /D Z
b I
t2
t2C jxyj2.dy; dt /:
One should keep in mind that the intervalsI are restricted to be in our fixed dyadic grid, a reduction allowed as the integrations on the left-hand side in (5.4) and (5.5) are done over the entire space, eitherR2CorR. (Goodness of the intervalsI above is not needed.) This reduction is critical to the analysis below.
Remark 5.6
A gap in the proof of the Poisson inequality in [27, p. 542] can be fixed as in [28]
or [8].
5.2. The Poisson testing inequality: The core
This section is concerned with a part of inequality (5.4). Restrict the integral on the left-hand side to the setbIR2C:
Z
b
IP.I /2d.x; t /H2 .I /:
Since.xJ;jJj/2Ibif and only ifJI, we have Z
b
IP.I /.x; t /2d.x; t /D X
F2F
X
J2J.F /WJI
P.I /
xJ;jJj2PF ;Jw x jJj
2w:
For eachJ,
PF ;Jw x jJj
2 w
Z
J
ˇˇ
ˇxEwJx jJj
ˇˇ ˇ
2
dw.x/D2E.w; J /2w.J /2w.J /: (5.7) LetF0 be the maximalF 2F which are strictly contained inI, and letJ]be those dyadic J such that.xJ;jJj/is in the support of, but has no parent inF0.
These intervals are necessarily disjoint. Observe that by (5.7) and the energy inequal- ity,
X
J2J]
P.F /
xJ;jJj2
xJ;jJj
X
J2J]
P .F; J /2E.w; J /2w.J /
H2 .F /: (5.8)
We claim that X
F2F0
Z
FOP
.InF /
.x; t /2d.x; t /H .I /: (5.9) This is sufficient, since
Z
b
IP.I /.x; t /2d.x; t /
LHS(5.8) C LHS(5.9) C X
F2F0
Z
FOP.F /.x; t /2d.x; t / H2 .I /C X
F2F0
Z
FOP.F /.x; t /2d.x; t /:
The individual terms in the last sum are set up for a recursive application of this inequality. Due to the Carleson condition (3.7), this recursion will finish the proof.
It remains to prove (5.9), which is another instance of the energy inequality. For an intervalF02F0, andF 2F strictly contained inF0, each intervalJ2J.F /is contained in someJ02J.F0/. Then, the intervalsF 2F are not good, butJ and J0are good, hence
P
.InF0/
xJ;jJj2
xJ;jJj DhZ
InF0
jJj jJj2C jxxJj2
i2PF ;Jw x jJj
2
w
DhZ
InF0
1 jJj2C jxxJj2
i2
kPF ;Jw xk2w
hZ
InF0
jJ0j jJ0j2C jxxJ0j2
i2 PF ;Jw x
jJ0j
2 w: This follows from goodness. Forx2InF0,
jJj2C jxxJj2 jxxJj2 jxxJ0j2 jJ0jjF0j1:
But then, we can add the projectionsPF ;Jw , due to orthogonality, and use (5.7) again to see that
X
F2F FF0
X
J2J.F / JJ0
P
.InF0/
xJ;jJj2
xJ;jJj
P.I /
xJ0;jJ0j2 X
F2F FF0
X
J2J.F / JJ0
PF ;Jw x
jJ0j
2 w
P.I /
xJ0;jJ0j2
E.w; J0/2w.J0/:
The sum overF02F0, andJ02J.F0/is controlled by the energy inequality. This completes the proof of (5.9).
5.3. The Poisson testing inequality: The remainder
Now we turn to proving the following estimate for the global part of the first testing condition (5.4):
Z
R2CbIP.I /2dA2 .I /:
Decompose the integral on the left-hand side into four terms. WithFJ the unique F 2F withJ2J.F /, and using (5.7),
Z
R2CbIP.I /2d
D X
JW.xJ;jJj/2R2CbI
P.I /
xJ;jJj2PFw
J;J
x jJj
2 w
° X
JWJ\3ID;
jJjjIj
C X
JWJ3II
C X
JWJ\ID;
jJj>jIj
C X
JWJI
±P.I /
xJ;jJj2
w.J /
DACBCCCD:
Decompose termAaccording to the length ofJ and its distance fromI, to obtain A
X1 nD0
X1 kD1
X
JWJ3kC1I3kI jJjD2njIj
2njIj
dist.J; I /2 .I /2
w.J /
X1 nD0
22n X1 kD1
jIj2 .I /w.3kC1I3kI / j3kIj4 .I /
X1 nD0
22n X1 kD1
32k° .3kC1I /w.3kC1I / j3kIj2
± .I /A2 .I /:
Decompose termBaccording to the length ofJand then use the Poisson inequal- ity (4.10), available to use because of goodness of intervalsJ. We then obtain
B X1 nD0
X
JWJ3II jJjD2n
2n.24/ .I /2 jIj2 w.J /
X1 nD0
2n.24/ .3I /w.3I /
j3Ij2 .I /A2 .I /:
For termC, fornD1; 2; : : :, setJn to be those good dyadic intervalsJ with jJj>jIj,J\I D ;and
.n1/jJj dist.I; J / < njJj:
These intervals have bounded overlaps. Indeed, suppose that J1 Jr are all members forJ1. Then, by goodness,
dist.J1; I /dist.Jr; I /.n1/2rjJ1j Cdist.J1; @Jr/ ®
.n1/2rC2r.1/¯ jJ1j;
which is a contradiction to membership inJn. Restricting the sum to intervals inJn, we have
X
J2Jn
P.I /
xJ;jJj2
w.J / .I /2 X
J2Jn
w.J / n4jJj2 .I /2
jIj X
J2Jn
w.J / jIj n4jJj2 .I /
n2 .I /
jIj P .w; I /A2
.I / n2 : And this is summable inn2N.
In the last termD, all the intervalsJ containI. Note that X
JWJI
P.I /
xJ;jJj2
w.J / .I /2 X
JWJI
w.J / jJj2 .I / .I /
jIj X
JWJI
w.J / jIj jJj2 .I / .I /
jIj P .w; I /A2 .I /:
5.4. The dual Poisson testing inequality
We are considering (5.5). Note that there is a power oft on both sides and that the expressions on the two sides of this inequality are
Z
b I
t2.dx; dt /D X
F2F
X
J2J.F / JI
kPF ;Jw xk2w;
P.tbI /.x/D X
F2F
X
J2J.F / JI
kPF ;Jw xk2w jJj2C jxxJj2:
We are to dominatekP.tI /kb 2by the first expression above. The squared norm will be the sum over integerssofTs below, in which the relative lengths ofJ andJ0are fixed bys. Suppressing the requirement thatJ; J0I,
TsWD X
F2F
X
J2J.F /
X
F02F
X
J02J.F / jJ0jD2sjJj
Z kPF ;Jw xk2w
jJj2C jxxJj2 kPFw0;J0xk2w jJ0j2C jxxJ0j2d
Ms
X
F2F
X
J2J.F /
kPF ;Jw xk2w;
where Ms sup
F2F
sup
J2J.F /
X
F02F
X
J02J.F / jJ0jD2sjJj
Z 1
jJj2C jxxJj2 w.J0/ jJ0j2 jJ0j2C jxxJ0j2d:
The estimate (5.7) has been used in the definition ofMs. We claim the termMs is at most a constant timesA22s, and it is here that the full PoissonA2condition is used.
FixJ, and letn2Nbe the integer chosen so that.n1/jJj dist.J; J0/njJj.
Estimate the integral in the definition ofMs by w.J0/
jJ0j
Z jJ0j2
jJj2C jxxJj2 jJ0j
jJ0j2C jxxJ0j2dA222s: This estimate is adequate fornD0; 1; 2. Then estimate the sum overJ0as
X
F02F
X
J02J.F0/WjJ0jD2sjJj .n1/jJjdist.J;J0/njJj
22s2s;
because the relative lengths of J and J0 are fixed, and each J0 is in at most one J.F /.
For the case ofn3, restrictJ0to be to the right ofJ, and lettnDxJCxJ0=2, so thatjxJtnj;jxJ0tnj 'njJj. First, estimate the integral in the definition ofMs
on the intervalŒtn;1/:
w.J0/ jJ0j
Z 1 tn
jJ0j2
jJj2C jxxJj2 jJ0j
jJ0j2C jxxJ0j2dA2
22s n2 :
Then estimate the sum overJ0as follows:
X
F02F
X
J02J.F0/WjJ0jD2sjJj .n1/jJjdist.J;J0/njJj
22s n2 2s
n2 :
This is clearly summable inn4.
Now, estimate on the integral on the interval.1; tn/:
w.J0/ jJ0j
Z tn 1
jJ0j2
jJj2C jxxJj2 jJ0j
jJ0j2C jxxJ0j2d D22sw.J0/
jJj Z tn
1
jJj
jJj2C jxxJj2 jJj2
jJ0j2C jxxJ0j2d 22sw.J0/
n2jJjP .; J /:
Drop the term with the geometric decay ins, and sum overnandJ0to see that X1
nD4
X
F02F
X
J02J.F0/WjJ0jD2sjJj .n1/jJjdist.J;J0/njJj
w.J0/
n2jJjP .; I /P .w; J /P .; J /A2:
Here, we have appealed to the full PoissonA2condition. This completes the control of the dual Poisson testing condition.
Acknowledgments.The authors benefited from a stimulating conference on two-weight inequalities in October 2011 at the American Institute of Mathematics in Palo Alto, California, for which they express their appreciation. The reviewing process led to many improvements in this paper, including a streamlining of the main contribution, for which we thank the referees.
References
[1] G. DAVIDandJ.-L. JOURNÉ,A boundedness criterion for generalized Calderón-Zygmund operators, Ann. of Math. (2)120(1984), 371–397.
MR 0763911.DOI 10.2307/2006946. (2797)