Introduction
The classical fractional integral operators (or Riesz potentials) Iα are defined by
Iα(f)(x)=γα∫Rn∣y∣n−αf(x−y)dy,
where 0<α<n and γα=2−απ−n/2Γ(2α)Γ(2n−α). This family of fractional potential operators was first systematically investigated by Riesz [32]. The (Lp,Lq)-boundedness of Riesz potentials was proved by Hardy and Littlewood [9] in the one-dimensional case, and by Sobolev [35] in higher dimensions. The endpoint weak-type (L1,Ln/(n−α),∞) boundedness was subsequently obtained by Zygmund [39]. The classical result is the following.
Theorem A. Let 0<α<n and let p,q satisfy 1≤p<q<∞ and p1−q1=nα. Then for p>1,
∥Iα(f)∥Lq(Rn)≤C(n,p,α)∥f∥Lp(Rn). When p=1,
∥Iα(f)∥Ln/(n−α),∞(Rn)=λ>0supλ∣{x∈Rn:∣Iαf∣>λ}∣(n−α)/n≤C(n,α)∥f∥L1(Rn).
Lieb [20] calculated the sharp constant in the conformally symmetric case p=n+α2n, q=n−α2n (see also [8]). More recently, Tang and Wang [37] determined sharp weak-type constants for classical fractional integral operators.
The study of reverse fractional integral operators was initiated by Dou and Zhu [6], who proved the existence of extremal functions and computed the best constant for p=q=n+s2n. Chen, Liu, Lu and Tao [3] established the reverse weighted Hardy–Littlewood–Sobolev inequality, namely the reverse Stein–Weiss inequality,
∫Rn∫Rn∣x∣α∣x−y∣s−nf(x)g(y)∣y∣βdxdy≥Cn,α,β,p,q′∥f∥Lq′∥g∥Lp
for any nonnegative functions f∈Lq′(Rn), g∈Lp(Rn), and p,q′∈(0,1), α,β,s−n>0 such that
p1+q′1−nα+β+s−n=2.
Kassymov, Ruzhansky and Suragan [16] subsequently extended the reverse weighted and unweighted Hardy–Littlewood–Sobolev inequality to homogeneous groups. For the endpoint case, the present authors jointly with Zhang [11] established the sharp reverse weak estimates and the extremal functions for the reverse fractional integral operators Is. For related reverse Hardy–Littlewood–Sobolev inequality, we refer the reader to [2, 23].
In the weighted setting, Stein and Weiss [36] established the following weighted Hardy–Littlewood–Sobolev inequality (also called the Stein–Weiss inequality):
∫Rn∫Rn∣x∣−α∣x−y∣−λf(x)g(y)∣y∣−βdxdy≤Cn,α,β,p,q∥f∥Lp(Rn)∥g∥Lq(Rn),(1)
where 1<p,q<∞, α,β,λ satisfy
p1+q1+nα+β+λ=2,p1+q1≥1,
α+β≥0,α<p′n,β<q′n,0<λ<n.
The Ap weight theory began with Muckenhoupt’s theorem [21]: the Hardy–Littlewood maximal operator M is bounded on Lp(w) for 1<p<∞ if and only if the weight w satisfies the Ap condition
[w]Ap:=Qsup(∣Q∣1∫Qw)(∣Q∣1∫Qw−1/(p−1))p−1<∞.(2)
Hunt, Muckenhoupt, and Wheeden [12] proved that the Ap condition also characterizes the boundedness of the Hilbert transform
Hf(x)=p.v.π1∫x−yf(y)dy
on Lp(w). Coifman and Fefferman [4] developed the general theory for Calderón–Zygmund operators and introduced the class A∞=⋃p>1Ap. Muckenhoupt and Wheeden [22] characterized the boundedness of the fractional integral operators as follows:
(∫Rn∣Iαf(x)w(x)∣qdx)1/q≤C(∫Rn∣f(x)w(x)∣pdx)1/pfor all f≥0.(3)
where 1/q=1/p−α/n, holds if and only if [w]p,q<∞, where the weight characteristic is defined as
[w]p,q=Qsup(∣Q∣1∫Qwqdx)(∣Q∣1∫Qw−p′dx)q/p′<∞.
The [w]p,q condition was first introduced in [22]. Sawyer [29] first gave the two-weight characterization for the maximal operator, and then set up the two-weight bounds for fractional and Poisson integrals [30]. Sawyer and Wheeden [31] extended the two-weight characterization of fractional integrals to Euclidean and homogeneous spaces. Pérez established two-weight norm inequalities for Riesz potentials [27], and subsequently for potential operators and fractional maximal operators [28].
Once the weighted boundedness of the operator is obtained, it is natural to consider how the operator bound depends on the weight. Buckley [1] proved that for the maximal operator,
∥M∥Lp(w)→Lp(w)≤c[w]Ap1/(p−1).
and that the exponent 1/(p−1) is sharp. For singular integral operators, the central open problem was so-called A2 conjecture, which asserts that for every Calderón–Zygmund operator T,
∥T∥L2(w)→L2(w)≤CT[w]A2.
Petermichl [25, 26] showed that
∥T∥Lp(w)→Lp(w)≤c[w]Apmax{1,1/(p−1)},
where T is either the Hilbert transform or one of the Riesz transforms
Rjf(x)=cnp.v.∫Rn∣x−y∣n+1xj−yjf(y)dy.
The A2 conjecture was resolved for general Calderón–Zygmund operators by Hytönen [13]. This landmark proof was subsequently simplified by Lerner [19], who gave an elegant argument via sparse pointwise domination. The sharp extrapolation theorem of Dragičević, Grafakos, Pereyra, and Petermichl [7] extended the sharp p=2 result to all 1<p<∞, yielding
∥T∥Lp(w)→Lp(w)≲[w]Apmax(1,1/(p−1)).
Furthermore, the sharp quantitative reverse Hölder theorem of Hytönen, Pérez, and Rela [15] and the sharp Fujii–Wilson A∞ bounds of Hytönen and Pérez [14] have also been established. Recently, Lacey, Moen, Pérez, and Torres [17] obtained the corresponding sharp bounds for weighted fractional integral operators Iα:
Theorem B. Let 1<p<n/α and q be defined by the equation 1/q=1/p−α/n, and [w]p,q<∞. Then,
∥Iα∥Lp(wp)→Lq(wq)≤c[w]p,q(1−α/n)max{1,p′/q}. Furthermore this estimate is sharp.
Further one-weight and two-weight results for fractional integral operators were obtained by Cruz-Uribe and Moen [5], who extend Theorem B to
∥Iα∥Lp(wp)→Lq(wq)≲[w]Ap1/p([w]A∞1/p′+[w1−p′]A∞1/p).
These existing results largely focus on power-weight settings, leaving open the general Muckenhoupt type one-weight characterization and sharp quantitative bounds for the best constant in reverse fractional integral operators. In the present paper we address two main questions for the reverse fractional integral operators Is with s>n:
What weight condition characterizes the boundedness of the reverse fractional integral operator, and what is the sharp dependence of the best constant on that weight?
What is the corresponding weight condition for endpoint boundedness?
The reverse fractional integral operators Is are defined by
Is(f)(x)=∫Rn∣x−y∣n−sf(y)dy.
where s>n. Throughout the strong-type discussion, set p′=p/(p−1)<0 and q′=q/(q−1)∈(0,1). Define the scale-invariant weight characteristic
[w]p,q:=Bsup(∣B∣1∫Bw−p′dx)−1/p′(∣B∣1∫Bwqdx)−1/q.
Here the supremum runs over all Euclidean balls B⊂Rn. By direct computation,
[w−p′]A1+p′/q=[w]p,q−p′,(4)
so that [w]p,q<∞ is equivalent to w−p′∈A1+p′/q. Our first main theorem provides the qualitative one-weight characterization.
Theorem 1.1. Let n≥1, s>n, n/s<p<1, and let q<0 satisfy 1/q=1/p−s/n. Let w>0 almost everywhere. Then there exists a constant C>0 such that
∥Isf∥Lq(wq)≥C∥f∥Lp(wp)(5) for every nonnegative measurable f if and only if [w]p,q<∞.
Remark 1. Let w(x)=∣x∣γ. Then (1.4) holds if and only if
n−pn<γ<s−pn.(6)
In fact, by Theorem 1.1, the required condition is ∣x∣−p′γ∈A1+p′/q, then we have
−n<−p′γ<n(1+p′/q−1)=qnp′.
Thus n−pn=n/p′<γ<−n/q=s−pn. The range (1.5) agrees with the one-weight specialization of the reverse Stein-Weiss parameter range in [3]. This gives an independent consistency check on the Muckenhoupt characterization.
Next, we obtain the sharp weighted bounds for the reverse fractional integral operators:
Theorem 1.2. Let n≥1, s>n, n/s<p<1, and let q<0 satisfy 1/q=1/p−s/n and [w]p,q<∞. Then,
∥Is∥Lp(wp)→Lq(wq)≥cn,s,p[w]p,qmin{n(1−p)p(n−s),sp−np(n−s)}. Furthermore this estimate is sharp.
Remark 2. Through the proof of Theorem 1.2, we have the following better bounds,
C∗(w)≥cn,s,p[w]p,q−1min{[w−p′]A∞1/q,[wq]A∞1/p′}.(7)
Combining (7), (1.3), the inequalities
[w−p′]A∞≲n,p,q[w−p′]A1+p′/qand[wq]A1+q/p′=[w−p′]A1+p′/qq/p′,
we obtain the desired lower bound in Theorem 1.2.
The critical point is pc=n+s2n, where the two symmetric terms coincide, giving p′=q. Thus
Γ∗={sp−np(s−n),n(1−p)p(s−n),sn<p≤n+s2n,n+s2n≤p<1.
At the conformally symmetric reverse Hardy-Littlewood-Sobolev exponent p=2n/(n+s), the sharp exponent is
C∗(w)≳[w]p,q−2.
At the endpoint p=1, we obtain a lower bound estimate of the reverse fractional integral operators.
Theorem 1.3. Let n≥1, s>n. The inequality
∥Isf∥Ln−sn,∞(wn−sn)≥snvnnn−s[wn−sn]A1nn−s∥f∥L1(w) holds for all f≥0 with a positive constant if and only if wn−sn∈A1.
For two quantities X, Y depending on parameters p, s, n, . . . , the notation X≲Y (resp. X≳Y) means X≤CY (resp. X≥cY) for constants c, C>0 that depend only on the indicated subscripts; X≈Y means both X≲Y and X≳Y.
The paper is organised as follows. Section 2 collects facts on Ap and A∞ weights, the bilinear reduction, and some lemmas. Section 3 proves necessity of the Muckenhoupt condition. In Section 4, a negative-index duality argument reduces the lower bound to a bilinear case, and we prove Theorems 1.1 and 1.2. We use power-weight to show that the exponent is sharp. Section 5 treats the endpoint p=1 and proves Theorem 1.3 by the layer-cake identity, the A1 condition, and Tonelli’s theorem, while necessity follows by testing on approximate point masses and applying the Lebesgue differentiation.
Preliminaries
In this section, we recall facts on Ar and A∞ weights and establish the bilinear and geometric lemmas used below.
In the main theorems, w is positive and finite almost everywhere; local integrability of w itself is not required. We use w(E)=∫Ew(x)dx when this integral is defined. For 1<r<∞, a positive, locally integrable function u belongs to Ar if
[w]Ar:=Bsup(∣B∣1∫Bw)(∣B∣1∫Bw−1/(r−1))r−1<∞.
Ball and cube definitions yield the same Ar class, with comparison constants depending on the dimension and the exponent, see [21].
The dual weight is
w∗=w−1/(r−1),
and
[w∗]Ar′=[w]Ar1/(r−1).(8)
Introducing the scale-invariant characteristic
[w]p,q=Bsup(∣B∣1∫Bw−p′dx)−1/p(∣B∣1∫Bwqdx)−1/q.
Equivalently, [w]p,q<∞ means
w−p′=wp′/(1−p)∈A1+p′/q=Ap(s−n)/(n(1−p)).
It is well known that A∞=⋃r>1Ar. Equivalently, we use the Fujii-Wilson characteristic
[w]A∞:=Bsupw(B)1∫BM(wχB)(x)dx,(9)
where M is the uncentered Hardy–Littlewood maximal operator. Equivalent A∞ characteristics are discussed in, for example, [38, 14]. The crucial quantitative fact is the sharp reverse-Hölder theorem of Hytönen–Pérez–Rela [15]. In the Euclidean setting it implies that there are dimensional constants cn,Cn>0 such that, for every w∈A∞, every ball B, and every measurable E⊂B,
w(B)w(E)≤Cn(∣B∣∣E∣)δw,δw≥[w]A∞cn.(10)
For q<0 and a positive measurable weight w, define
∥f∥Lq(w):=(∫Rnf(x)qw(x)dx)1/q,
with the standard extended-value convention. The Lp norm for p<1 is not a norm for a vector space. For q<0, and a weight w, the reverse weak-type quasi-norm is
∥f∥Lq,∞(w):=λ>0infλw({x:f(x)<λ})1/q,
where the sublevel set uses f<λ.
For q<0<p≤1, we say an operator T is of weak-type (p,q), if there exists a constant C(p,q,n)>0 such that ∥f∥Lq,∞≥C(p,q,n)∥f∥Lp for all f∈Lp(Rn).
Lemma 2.1 ([6]). For 0<r<1, put r′=r/(r−1)<0. For nonnegative measurable functions f,g, the reverse Hölder inequality holds whenever the quantities are defined:
∫f(x)g(x)dx≥∥f∥Lr∥g∥Lr′.
We also need the exact dual formula for a negative exponent.
Lemma 2.2. Let q<0 and q′=q/(q−1)∈(0,1). For every positive measurable f,
∥f∥Lq=∥g∥Lq′=1, g≥0inf∫Rnf(x)g(x)dx.
Proof. Reverse Hölder’s inequality with exponents q′ and q gives the inequality for every admissible g. Put A=∫Rnfq(x)dx. If 0<A<∞, then g=A−1/q′fq−1 has ∥g∥q′=1 and
∫Rnf(x)g(x)dx=A1/q=∥f∥q.
If A=∞, choose increasing measurable sets Ej for which 0<∫Ejfq(x)dx<∞ and ∫Ejfq(x)dx→∞. The same construction with fq−1χEj gives integrals (∫Ejfq(x)dx)1/q→0=∥f∥q. □
The strong (p,q) inequality for the weighted reverse fractional integral operators can then be equivalently reformulated as a bilinear form controlled from below by a product of norms. Given F(y)=f(y)w(y), define
SwF(x)=w(x)Is(F/w)(x).
Then
∥Isf∥Lq(wq)=∥SwF∥Lq,∥f∥Lp(wp)=∥F∥Lp.
By Lemma 2.2, the best strong constant equals the best constant in
Bw(F,G):=∬Rn×Rn∣x−y∣s−nF(y)G(x)w(y)w(x)dydx≥C∥F∥p∥G∥q′.(11)
This formulation exhibits the symmetry
Cp,q′(w)=Cq′,p(w−1).(12)
It will be useful in the sharpness argument.
Next, we introduce the following lemma which play an important roles in the proof of the main theorems.
Lemma 2.3. Let ∥F∥Lp=∥G∥Lq′=1 for q<0. Assume [w]p,q<∞ and let B=B(z,r). Fix η∈(0,1). If
∫BF(y)pdy≥η,∫(2B)cG(x)q′dx≥η, then
Bw(F,G)≥cn,s,p,η[w]p,q−1[w−p′]A∞1/q.(13)
If instead
∫BG(x)q′dx≥η,∫(2B)cF(y)pdy≥η,
then
Bw(F,G)≥cn,s,p,η[w]p,q−1[wq]A∞1/p′.
Proof. We prove the first assertion. Decompose
(2B)c=k≥1⋃Ak,Ak=2k+1B∖2kB,
and set qk=∫AkG(x)q′dx. Then
k≥1∑qk≥η.
If y∈B and x∈Ak, then
∣x−y∣≥c2kr.
Hence
Bw(F,G)≥ck≥1∑(2kr)s−n(∫Bw(y)F(y)dy)(∫AkG(x)w(x)dx).
Reverse Hölder gives
∫Bw(y)F(y)dy≥(∫BF(y)pdy)1/p(∫Bw−p′(y)dy)1/p′≥η1/p(∫Bw−p′(y)dy)1/p′,
and
∫AkG(x)w(x)dx≥qk1/q′(∫Akwq(y)dy)1/q≥qk1/q′(∫2k+1Bwq(y)dy)1/q.
Applying the definition of the weight condition on 2k+1B yields
Bw(F,G)≥cη1/pk≥1∑qk1/q′1(2kr)s−n(∫Bw−p′(y)dy)1/p′(∫2k+1Bwq(y)dy)1/q≥cn,s,p,η[w]p,q−1k≥1∑(w−p′(B)w−p′(2k+1B))−1/p′qk1/q′1.(14)
By the A∞ estimate (10), we have
w−p′(2k+1B)w−p′(B)≤Cn2−knδu,δu≥[w−p′]A∞cn.
Therefore
(w−p′(B)w−p′(2k+1B))−1/p′≥cn2−p′nδuk,where 2p′nδu>1.
Set ak=2−nδuk/p′. Since 1/q+1/q′=1 and q<0, discrete reverse Hölder gives
k≥1∑(w−p′(B)w−p′(2k+1B))−1/p′qk1/q′≳nk≥1∑akqk1/q′≥(k≥1∑akq)1/q(k≥1∑qk)1/q′≥η1/q′2−nδu/p′(1−2−nδuq/p′)−1/q.
We may choose 0<δu≤1 with δu≳n[w−p′]A∞−1. Since q/p′>0, the last expression is bounded below by cn,s,p,ηδu−1/q≳n,s,p,η[w−p′]A∞1/q. Together with (2.7), this proves (2.6). The second assertion follows by exchanging (F,p,w) with (G,q′,w−1); the weight characteristic is unchanged under this exchange.
The following lemma guarantees that such separated functions in Lemma 2.3 always hold.
Lemma 2.4. Let 0<p<1, 0<q′<∞, and let F,G≥0 be measurable functions on Rn with
∫RnF(y)pdy=1,∫RnG(x)q′dx=1. And there is a dimensional constant ηn>0 such that at least one of the following alternatives holds for some ball B:
∫BFp(y)dy≥ηn,∫(2B)cGq′(x)dx≥21;(15) ∫BGq′(x)dx≥ηn,∫(2B)cFp(y)dy≥21.(16)
Proof. For a nonnegative integrable function H with ∫RnH(y)dy=1, define
cH(r)=z∈Rnsup∫B(z,r)H(y)dy,rH=inf{r>0:cH(r)≥1/2}.
Since H∈L1, absolute continuity of the integral gives cH(r)→0 as r↓0, uniformly in the center. Since ∫RnH(y)dy=1, there holds cH(r)→1 as r→∞. Hence 0<rH<∞, and by definition every ball B of radius strictly smaller than rH satisfies ∫BH(y)dy<1/2.
We claim there is a ball BH of radius 2rH with ∫BHH≥1/3. Indeed, choose rj↓rH with rj<2rH and cH(rj)≥1/2. For each j there is a center zj with ∫B(zj,rj)H>1/3, and B(zj,2rH)⊃B(zj,rj).
Apply this to H=Fp and H=Gq′, obtaining balls BF,BG of radii 2rF,2rG respectively, each carrying mass ≥1/3. Choose L=16.
Case 1: rF≤rG/L. Then 2BF has radius 4rF<rG, so ∫2BFGq′<1/2 while ∫BFFp≥1/3. Thus (2.8) holds with B=BF and ηn=1/3.
Case 2: rG≤rF/L. Symmetrically, (2.9) holds with B=BG. Case 3: L−1rG<rF<LrG. Cover BF, whose radius is 2rF, by at most Nn balls of radius rG/4, where Nn depends only on n and L. Such a bound follows by taking a maximal rG/4-separated set of centers in BF: the radius-rG/8 balls are disjoint and lie in a ball of controlled radius, so their number is bounded by a volume ratio. Since ∫BFFp(y)dy≥1/3, one of these covering balls, say B∗, satisfies
∫B∗Fp≥3Nn1.
The radius of 2B∗ is rG/2<rG, hence ∫2B∗Gq′<1/2. This proves (2.8) with ηn=(3Nn)−1. Interchanging Fp and Gq′ proves (2.9) as well. □
Necessity of the Muckenhoupt condition
Proposition 1. If (1.4) holds with constant C>0, then
[w]p,q≤Cn,sC−1.(17)
In particular, w−p′∈A1+qp′.
Proof. Fix a ball B=B(x0,R). We first show wq(B)<∞. For sufficiently large N, the set DN=B∩{N−1≤w≤N} has positive measure. For f=χDN, both sides of (1.4) are positive and finite on B. If wq(B)=∞, the negative q-norm of Isf is zero, contradicting (1.4). Now let f≥0 be supported in B. For x,y∈B,
∣x−y∣s−n≤(2R)s−n,
so
Isf(x)≤(2R)s−n∫Bf(y)dy.
Restricting the integral to B gives an upper bound for the negative norm. Hence
∥Isf∥Lq(wq)≤(2R)s−n(∫Bf(y)dy)(∫Bw(y)qdy)1/q.
Combining this with the assumed lower bound yields
C≤(2R)s−n(∫Bf(y)pw(y)pdy)1/p∫Bf(y)dy(∫Bw(y)qdy)1/q.
Note that the cut-off function fN=w−p′1B∩{N−1≤w≤N}, whose integral is always finite. If ∫Bw−p′(y)dy=∞, the resulting ratio tends to zero, contradicting the assumed lower bound. Hence w−p′ is locally integrable, and using reverse Hölder’s inequality, we have
finf(∫Bf(y)pw(y)pdy)1/p∫Bf(y)dy=(∫Bw(y)−p′dy)1/p′.
Thus
(∫Bw−p′)−1/p′(∫Bwq)−1/q≤C−1(2R)s−n.
Since ∣B∣−1/p′−1/q=∣B∣(s−n)/n≃nRs−n, division by ∣B∣(s−n)/n proves (3.1). □
Strong-type bounds and sharpness
Proof of Theorem 1.2. By homogeneity, normalize ∥F∥p=∥G∥q′=1 and apply Lemma 2.4. If (2.8) holds, then Lemma 2.3 gives
Bw(F,G)≥cn,s,p[w]p,q−1[w−p′]A∞1/q.
If (2.9) holds, then
Bw(F,G)≥cn,s,p[w]p,q−1[wq]A∞1/p′.
Thus, we obtain
Bw(F,G)≥cn,s,p[w]p,q−1min{[w−p′]A∞1/q,[wq]A∞1/p′}.
Taking the infimum over G with ∥G∥q′=1 and using Lemma 2.2 proves (1.6).
Proof of Theorem 1.1. The implication (i)⇒(ii) is Proposition 1. If (ii) holds, then (1.3) gives w−p′∈A1+qp′ and its dual weight wq∈A1+p′q, so both are in A∞. The mixed lower bound of Theorem 1.2 is therefore strictly positive and yields (i).
In the following subsections, we first prove the lower bound, then show that the exponent cannot be improved.
Proof. Since
[w−p′]A∞≲n,p,q[w−p′]A1+qp′=[w]p,qp′,
and, using (2.1),
[wq]A∞≲n,p,q[wq]A1+p′q=[w]p,q−q,
Theorem 1.2 gives the lower bound estimate
C∗(w)≳[w]p,q−1min{[w]p,q−p′/q,[w]p,q−q/p′}=[w]p,q−(1+max{p′/q,q/p′})=[w]p,q−Γ∗.
Next, we show the sharpness of the exponent. Note that
∣x∣δ∈Ar⟺−n<δ<n(r−1).(18)
and, as δ→−n and δ→n(r−1),
[∣x∣−n+ε]Ar≈n,rε−1,(19)
[∣x∣n(r−1)−ε]Ar≈n,rε−(r−1).(20)
Let
wε−(x)=∣x∣(n−ε)/p′,0<ε<1.
By (4.2),
[wε−]p,q≈ε1/p′.(21)
Choose
fε(x)=∣x∣−n+εχB(0,1)(x).
Since −p/p′=1−p,
∥fε∥Lp((wε−)p)p=∫B(0,1)∣x∣−n+εdx=ε∣Sn−1∣.(22)
For ∣x∣≥2 and ∣y∣<1, ∣x−y∣≤23∣x∣, hence
Isfε(x)≤(23)s−n∣x∣s−nε∣Sn−1∣.
And
∫Rn(Isfε(x))q(wε−(x))qdx≥cn,sε−q∫∣x∣>2∣x∣(s−n)q+(n−ε)q/p′dx=cn,sε−q∣Sn−1∣∫2∞r−1−εq/p′dr≥cn,s,pε−q−1.
Therefore
∥Isfε∥Lq((wε−)q)≤Cn,s,pε−1/q−1.(23)
Combining (4.5) and (4.6),
C∗(wε−)≤Cn,s,pε−1/p′−1/q=Cn,s,pε(s−n)/n.(24)
(4.4) implies
C∗(wε−)≲[wε−]p,q−(1+p′/q).(25)
Hence, we obtain C∗(w)≳[w]p,q−Γ with Γ≥1+p′/q.
On the other hand, let
wε+(x)=∣x∣−n/q+ε/p′,0<ε<1.
By (4.3),
[wε+]p,q≈ε1/q.(26)
The bilinear symmetry (2.5) exchanges
(p,−p′,w−p′)⟷(q′,−q,wq),w⟷w−1.
For w=wε+,
σ=(wε+)q=∣x∣−n+εq/p′,
which is just the lower-boundary family with the parameters exchanged. Applying (4.7) with −p′ and −q exchanged gives
C∗(wε+)≲ε−1/p′−1/q=ε(s−n)/n.
(4.9) gives
C∗(wε+)≲[wε+]p,q−(1+q/p′).
Hence Γ≥1+q/p′. Together with (4.8), this proves optimality of Γ∗=max{1+p′/q,1+q/p′} and completes the proof of sharp exponent of Theorem 1.2.
Endpoint theorem and explicit characteristic bounds
We now consider the endpoint case p=1. For a positive measure σ, we use
∥h∥Ln−sn,∞(σ)=λ>0infλσ({x:h(x)<λ})nn−s.
The A1 characteristic is
[σ]A1=BsupessinfBσ∣B∣1∫Bσ.
Proof. We first prove the sufficient. Assume wn−sn∈A1 and put A=[wn−sn]A1. Fix f≥0 and λ>0, and set
Eλ={x:Isf(x)<λ}.
If f is nonzero and Isf is finite at some point, then each superlevel set Eλ is bounded. Indeed, if ∫Bf(y)dy>0 for some ball B, then Isf(x) grows at least like ∣x∣s−n as ∣x∣→∞. Thus it suffices to consider 0<wn−sn(Eλ)<∞. For almost every y and every r>0, the A1 condition gives
wn−sn(Eλ∩B(y,r))≤wn−sn(B(y,r))≤Awn−sn(y)vnrn.
Set Ry=(Avnwn−sn(y)wn−sn(Eλ))1/n. The layer-cake identity and Tonelli’s theorem give
∫Eλ∣x−y∣s−nwn−sn(x)dx=(s−n)∫0∞rs−n−1wn−sn(Eλ∖B(y,r))dr≥(s−n)∫0Ryrs−n−1(wn−sn(Eλ)−Awn−sn(y)vnrn)dr=sn(Avn)nn−swn−sn(Eλ)s/nwn−sn(y)1/qe=sn(Avn)nn−swn−sn(Eλ)s/nw(y).
Consequently, Tonelli’s theorem yields
λwn−sn(Eλ)≥∫EλIsf(x)wn−sn(x)dx≥sn(Avn)nn−swn−sn(Eλ)s/n∫Rnf(y)w(y)dy.
Division by wn−sn(Eλ)s/n and then taking the infimum over λ, we obtain
∥Isf∥Ln−sn,∞(wn−sn)≥snvnnn−s[wn−sn]A1nn−s∥f∥L1(w).
We show necessity. First, wn−sn is locally integrable. Indeed, for any ball B, choose N so that B∩DN has positive measure, where DN={N−1≤w≤N}. With f=χB∩DN, ∥f∥L1(w)>0 and Isf is bounded on B. If wn−sn(B)=∞, which makes the reverse weak norm zero, contradicting
∥Isf∥Ln−sn,∞(wn−sn)≥snvnnn−s[wn−sn]A1nn−s∥f∥L1(w).
Fix N and choose a point y∈DN that is also a Lebesgue point of wχDN and wn−sn. Define
fε(z)=∫B(y,ε)∩DNw(x)dxχB(y,ε)∩DN(z),∫Rnfε(z)w(z)dz=1.
By Lebesgue differentiation, we obtain ∫fε(z)dz→w(y)−1. For x∈B(y,R),
Isfε(x)≤(R+ε)s−n∫Rnfε(z)dz.
Since n−sn<0, let ε→0, this yields
∣B∣1∫B(y,R)wn−sn(z)dz≤vn−1(snvnnn−s[wn−sn]A1nn−swn−sn(y))n−sn(27)
for almost every y and every R>0. For an arbitrary ball B(c,R) and almost every y∈B(c,R),
∣B∣1∫B(c,R)wn−sn(z)dz≤2n∣B∣1∫B(y,2R)wn−sn(z)dz
≤2nvn−1(snvnnn−s[wn−sn]A1nn−swn−sn(y))n−sn.
Take the essential infimum in y, we have wn−sn∈A1. This completes the proof of Theorem 1.3. □
School of Science, Xi’an University of Posts and Telecommunications, Xi’an, 710121, China, E-mail: [email protected]
Laboratory of Mathematics and Complex Systems (Ministry of Education), School of Mathematical Sciences, Beijing Normal University, Beijing, 100875, China, E-mail: [email protected]