Introduction
The Hardy spaces on R n \mathbb{R}^n R n were introduced in [9 ] by C. Fefferman and E. Stein, and have been extensively studied ever since. A key feature of these spaces is that they provide an effective alternative to Lebesgue spaces in the range 0 < p < 1 0<p<1 0 < p < 1 . In particular, while Riesz transforms fail to be bounded on L p ( R n ) L^p(\mathbb{R}^n) L p ( R n ) for p ≤ 1 p\le1 p ≤ 1 , they are bounded on the corresponding Hardy spaces H p ( R n ) H^p(\mathbb{R}^n) H p ( R n ) . To investigate the boundedness of operators such as singular integrals or fractional-type operators on Hardy spaces, one typically relies on their atomic or molecular characterizations. The atomic decomposition was first established by Coifman in one dimension (see [3 ] ) and later extended to higher dimensions by Latter [18 ] .
These descriptions suggest that the boundedness of linear operators on H p H^p H p could, in principle, be reduced to analyzing their action on atoms. However, this approach has certain limitations. Bownik [2 ] , building on an example due to Meyer, constructed a linear functional defined on a dense subspace of H 1 ( R n ) H^1(\mathbb{R}^n) H 1 ( R n ) that maps all ( 1 , ∞ , 0 ) (1,\infty,0) ( 1 , ∞ , 0 ) -atoms into bounded scalars, but does not admit a bounded extension to H 1 ( R n ) H^1(\mathbb{R}^n) H 1 ( R n ) .
2020 Mathematics Subject Classification. 42B20, 42B25, 42B30.
Key words and phrases. Fractional operators, weighted Hardy spaces, Muckenhoupt weights, Hörmander conditions.
The authors are partially supported by CONICET and SECYT-UNC.
Nevertheless, for classical operators T T T , uniform boundedness on atoms implies boundedness from H p H^p H p into L p L^p L p . This follows from the fact that T T T is bounded on L s L^s L s for some 1 < s < ∞ 1<s<\infty 1 < s < ∞ , together with the possibility of choosing atomic decompositions that converge in L s L^s L s ; see [28 ] and [21 ] .
The weighted Hardy spaces H w p ( R n ) H_w^p(\mathbb{R}^n) H w p ( R n ) are defined by adapting the definition of H p ( R n ) H^p(\mathbb{R}^n) H p ( R n ) to the measure w ( x ) d x w(x)\,dx w ( x ) d x (see [24 ] ), and the theory is particularly meaningful for w ∈ A ∞ w\in A_\infty w ∈ A ∞ . The atomic decomposition of H w p ( R n ) H_w^p(\mathbb{R}^n) H w p ( R n ) was established in [10 , 24 ] . In [22 ] , the author obtained, for all w ∈ A ∞ w\in A_\infty w ∈ A ∞ , an atomic decomposition of H w p ( R n ) H_w^p(\mathbb{R}^n) H w p ( R n ) converging in L s L^s L s , 1 < s < ∞ 1<s<\infty 1 < s < ∞ , and a molecular decomposition. As applications, he obtained the boundedness of certain singular integrals on H w p H_w^p H w p and from H w p H_w^p H w p into L w p L_w^p L w p , for all w ∈ A ∞ w\in A_\infty w ∈ A ∞ and 0 < p ≤ 1 0<p\leq1 0 < p ≤ 1 , and the H w p − H w q H_w^p-H_w^q H w p − H w q and H w p − L w q H_w^p-L_w^q H w p − L w q boundedness of the Riesz potential I α I_\alpha I α , for 1 q = 1 p − α n \frac{1}{q}=\frac{1}{p}-\frac{\alpha}{n} q 1 = p 1 − n α and certain weights w w w .
Let Ω ∈ L r ( S n − 1 ) \Omega\in L^r(\mathbb{S}^{n-1}) Ω ∈ L r ( S n − 1 ) , r > 1 r>1 r > 1 , be homogeneous of degree zero, and consider
T Ω , α f ( x ) = ∫ Ω ( x − y ) ∣ x − y ∣ n − α f ( y ) d y . T_{\Omega,\alpha}f(x)=\int\frac{\Omega(x-y)}{|x-y|^{n-\alpha}}f(y)\,dy. T Ω , α f ( x ) = ∫ ∣ x − y ∣ n − α Ω ( x − y ) f ( y ) d y .
Weighted estimates for T Ω , 0 T_{\Omega,0} T Ω , 0 were obtained in [17 , 7 , 26 ] , and for α > 0 \alpha>0 α > 0 in [5 ] . In [1 ] , the authors obtained a Coifman type inequality for general fractional integral operators whose kernels satisfy a size condition and a Hörmander condition given by a Young function. In [6 ] , Ding and Lu proved, under L r L^r L r -Dini type conditions on Ω \Omega Ω , the H p − L q H^p-L^q H p − L q boundedness of T Ω , α T_{\Omega,\alpha} T Ω , α for p = 1 p=1 p = 1 and, when 0 < α < 1 0<\alpha<1 0 < α < 1 , for n n + α ≤ p < 1 \frac{n}{n+\alpha}\leq p<1 n + α n ≤ p < 1 , as well as its H p − H q H^p-H^q H p − H q boundedness for a range of p < 1 p<1 p < 1 when 0 < α < 1 / 2 0<\alpha<1/2 0 < α < 1/2 . Strömberg and Wheeden [25 ] gave the weighted H p − L q H^p-L^q H p − L q and H p − H q H^p-H^q H p − H q boundedness of I α I_\alpha I α , and Ding, Lee and Lin [4 ] , using the atom–molecule theory of [10 , 19 ] , obtained the weighted H p − L q H^p-L^q H p − L q and H p − H q H^p-H^q H p − H q boundedness of T Ω , α T_{\Omega,\alpha} T Ω , α . More recently, Han and Wu [15 ] considered fractional Hörmander conditions of order δ \delta δ , which strengthen those of [1 ] , and obtained weighted H p − L q H^p-L^q H p − L q estimates for fractional integrals ∫ K α ( x − y ) f ( y ) d y \int K_\alpha(x-y)f(y)\,dy ∫ K α ( x − y ) f ( y ) d y with a single kernel satisfying such a condition; see [15 Definition 5.6 and Theorem 5.10] . Estimates for I α I_\alpha I α and for T Ω , α T_{\Omega,\alpha} T Ω , α with Ω \Omega Ω Lipschitz on Musielak–Orlicz Hardy spaces were recently obtained in [14 ] .
Let 0 ≤ α < n 0\leq\alpha<n 0 ≤ α < n and m ∈ N m\in\mathbb{N} m ∈ N . For 1 ≤ i ≤ m 1\leq i\leq m 1 ≤ i ≤ m let 1 < q i < ∞ 1<q_i<\infty 1 < q i < ∞ be such that
n q 1 + ⋯ + n q m = n − α . (1) \frac{n}{q_1}+\cdots+\frac{n}{q_m}=n-\alpha. \tag*{(1)} q 1 n + ⋯ + q m n = n − α . (1)
We study the integral operator
T α f ( x ) = ∫ R n K ( x , y ) f ( y ) d y , K ( x , y ) = k 1 ( x − A 1 y ) ⋯ k m ( x − A m y ) , (2) T_\alpha f(x)=\int_{\mathbb{R}^n}K(x,y)f(y)\,dy,\qquad K(x,y)=k_1(x-A_1y)\cdots k_m(x-A_my), \tag*{(2)} T α f ( x ) = ∫ R n K ( x , y ) f ( y ) d y , K ( x , y ) = k 1 ( x − A 1 y ) ⋯ k m ( x − A m y ) , (2)
where A 1 , … , A m A_1,\ldots,A_m A 1 , … , A m are invertible matrices, k 1 , … , k m k_1,\ldots,k_m k 1 , … , k m are measurable functions on R n \mathbb{R}^n R n , and f ∈ L c ∞ ( R n ) f\in L_c^\infty(\mathbb{R}^n) f ∈ L c ∞ ( R n ) .
When A i = a i I A_i=a_iI A i = a i I and k i ( x ) = ∣ x ∣ − n / q i k_i(x)=|x|^{-n/q_i} k i ( x ) = ∣ x ∣ − n / q i , Godoy and Urciuolo [12 ] obtained the L p − L q L^p-L^q L p − L q boundedness of T α T_\alpha T α for 1 < p < n α 1<p<\frac{n}{\alpha} 1 < p < α n , 1 q = 1 p − α n \frac{1}{q}=\frac{1}{p}-\frac{\alpha}{n} q 1 = p 1 − n α . For smooth Ω i \Omega_i Ω i , Rocha and Urciuolo [23 ] considered matrices satisfying
(H) A i A_i A i is invertible and A i − A j A_i-A_j A i − A j is invertible for i ≠ j i\neq j i = j , 1 ≤ i , j ≤ m 1\leq i,j\leq m 1 ≤ i , j ≤ m ,
and proved that T α T_\alpha T α is bounded from H p ( R n ) H^p(\mathbb{R}^n) H p ( R n ) into L q ( R n ) L^q(\mathbb{R}^n) L q ( R n ) for 0 < p < n α 0<p<\frac{n}{\alpha} 0 < p < α n ; they also showed that H p − H q H^p-H^q H p − H q boundedness cannot be expected, an important difference with the case m = 1 m=1 m = 1 . The weighted L p − L q L^p-L^q L p − L q boundedness of T α T_\alpha T α with rough homogeneous kernels was obtained by Riveros and Urciuolo [20 ] , and Ibáñez Firnkorn and Riveros [16 ] treated kernels satisfying size and Hörmander conditions of Young type. More recently, Wu and Zhang [27 ] studied these operators, with rough homogeneous kernels, on Orlicz–Morrey spaces.
In this paper we do not assume that the kernels are homogeneous. Following [1 Definition 3.1] , we only require a size condition on dyadic annuli. For the regularity of the kernels we use the fractional Hörmander conditions of order δ \delta δ of [15 ] , which strengthen the fractional Hörmander condition of [1 Definition 3.2] and are adapted to the Hardy spaces H p H^p H p with p < 1 p < 1 p < 1 . We study the structure of this scale of conditions and its sharpness, and we use it to obtain weighted H p − L q H^p-L^q H p − L q estimates for operators with several kernels composed with linear maps. These are the main contributions of this paper.
Hypotheses and main result
We write ∣ x ∣ ∼ s |x| \sim s ∣ x ∣ ∼ s for the annulus { x ∈ R n : s < ∣ x ∣ ≤ 2 s } \{x \in\mathbb{R}^n : s < |x| \le2s\} { x ∈ R n : s < ∣ x ∣ ≤ 2 s } and, for 1 ≤ τ < ∞ 1 \le\tau< \infty 1 ≤ τ < ∞ and s > 0 s > 0 s > 0 , we consider the normalized annular norm
∥ f ∥ τ , ∣ x ∣ ∼ s = ( 1 ∣ B ( 0 , 2 s ) ∣ ∫ s < ∣ x ∣ ≤ 2 s ∣ f ( x ) ∣ τ d x ) 1 / τ . \|f\|_{\tau,|x|\sim s}=\left(\frac{1}{|B(0,2s)|}\int_{s<|x|\le2s}|f(x)|^\tau\,dx\right)^{1/\tau}. ∥ f ∥ τ , ∣ x ∣ ∼ s = ( ∣ B ( 0 , 2 s ) ∣ 1 ∫ s < ∣ x ∣ ≤ 2 s ∣ f ( x ) ∣ τ d x ) 1/ τ .
with ∥ f ∥ ∞ , ∣ x ∣ ∼ s = ess sup s < ∣ x ∣ ≤ 2 s ∣ f ( x ) ∣ \|f\|_{\infty,|x|\sim s}=\operatorname*{ess\,sup}_{s<|x|\le2s}|f(x)| ∥ f ∥ ∞ , ∣ x ∣ ∼ s = ess sup s < ∣ x ∣ ≤ 2 s ∣ f ( x ) ∣ ; this is the norm ∥ f ∥ A , ∣ x ∣ ∼ s \|f\|_{A,|x|\sim s} ∥ f ∥ A , ∣ x ∣ ∼ s of [1 ] with A ( t ) = t τ A(t)=t^\tau A ( t ) = t τ . Throughout the paper 0 < α < n 0 < \alpha< n 0 < α < n , 0 < p < 1 0 < p < 1 0 < p < 1 ,
δ = n ( 1 p − 1 ) , 1 q = 1 p − α n . \delta=n\left(\frac{1}{p}-1\right),\qquad\frac{1}{q}=\frac{1}{p}-\frac{\alpha}{n}. δ = n ( p 1 − 1 ) , q 1 = p 1 − n α .
Besides (H) (which, for m = 1 m=1 m = 1 , only requires A 1 A_1 A 1 to be invertible), we assume:
(H₁) For each 1 ≤ i ≤ m 1 \le i \le m 1 ≤ i ≤ m there is q i < p i ≤ ∞ q_i < p_i \le\infty q i < p i ≤ ∞ such that ∥ k i ∥ p i , ∣ x ∣ ∼ s ≤ C s − n / q i \|k_i\|_{p_i,|x|\sim s} \le Cs^{-n/q_i} ∥ k i ∥ p i , ∣ x ∣ ∼ s ≤ C s − n / q i for all s > 0 s > 0 s > 0 .
(H₂) For each 1 ≤ i ≤ m 1 \le i \le m 1 ≤ i ≤ m , with the same p i p_i p i as in (H₁), there exist c i ≥ 1 c_i \ge1 c i ≥ 1 and C i > 0 C_i > 0 C i > 0 such that
∑ j = 1 ∞ 2 j δ ( 2 j R ) n / q i ∥ k i ( ⋅ − h ) − k i ∥ p i , ∣ x ∣ ∼ 2 j R ≤ C i , h ∈ R n , R > c i ∣ h ∣ . \sum_{j=1}^{\infty}2^{j\delta}(2^jR)^{n/q_i}\|k_i(\cdot-h)-k_i\|_{p_i,|x|\sim2^jR}\le C_i,\qquad h\in\mathbb{R}^n,\ R>c_i|h|. j = 1 ∑ ∞ 2 j δ ( 2 j R ) n / q i ∥ k i ( ⋅ − h ) − k i ∥ p i , ∣ x ∣ ∼ 2 j R ≤ C i , h ∈ R n , R > c i ∣ h ∣.
Condition (H₁) is the size condition of [1 Definition 3.1] . Condition (H₂) means that k i ∈ H n − n / q i , p i ( δ ) k_i\in H^{(\delta)}_{n-n/q_i,p_i} k i ∈ H n − n / q i , p i ( δ ) , where, for 1 ≤ τ ≤ ∞ 1\le\tau\le\infty 1 ≤ τ ≤ ∞ , 0 ≤ β < n 0\le\beta<n 0 ≤ β < n and δ ≥ 0 \delta\ge0 δ ≥ 0 , H β , τ ( δ ) H^{(\delta)}_{\beta,\tau} H β , τ ( δ ) denotes the class of kernels satisfying the fractional Hörmander condition of order δ \delta δ introduced by Han and Wu [15 Definition 5.6] (see Definition 3.2 ). For δ = 0 \delta=0 δ = 0 we recover the fractional Hörmander condition of [1 Definition 3.2] (the case k = 0 k=0 k = 0 there), while for δ = n ( 1 p − 1 ) \delta=n\left(\frac{1}{p}-1\right) δ = n ( p 1 − 1 ) the factor 2 j δ 2^{j\delta} 2 j δ is exactly the one produced by the action of T α T_\alpha T α on the atoms of H w p H^p_w H w p (see Remark 4.3).
The classes H β , τ ( δ ) H^{(\delta)}_{\beta,\tau} H β , τ ( δ ) form a genuine scale. They decrease as δ \delta δ increases, and Proposition 3.7 shows that, for n ≥ 2 n\ge2 n ≥ 2 , all these inclusions are strict:
S β , τ ∩ H β , τ ( δ 2 ) ⊊ S β , τ ∩ H β , τ ( δ 1 ) for 0 ≤ δ 1 < δ 2 , δ 1 < 1 , 1 ≤ τ ≤ ∞ . S_{\beta,\tau}\cap H^{(\delta_2)}_{\beta,\tau}\subsetneq S_{\beta,\tau}\cap H^{(\delta_1)}_{\beta,\tau}\qquad\text{for }0\le\delta_1<\delta_2,\ \delta_1<1,\ 1\le\tau\le\infty. S β , τ ∩ H β , τ ( δ 2 ) ⊊ S β , τ ∩ H β , τ ( δ 1 ) for 0 ≤ δ 1 < δ 2 , δ 1 < 1 , 1 ≤ τ ≤ ∞.
Moreover, the position of a kernel in this scale can be computed explicitly in natural examples: for the homogeneous kernels k ε ( x ) = ∣ x n ∣ ε ∣ x ∣ β − n − ε k_\varepsilon(x)=|x_n|^\varepsilon|x|^{\beta-n-\varepsilon} k ε ( x ) = ∣ x n ∣ ε ∣ x ∣ β − n − ε , with − 1 τ < ε < 1 − 1 τ -\frac{1}{\tau}<\varepsilon<1-\frac{1}{\tau} − τ 1 < ε < 1 − τ 1 and ε ≠ 0 \varepsilon\ne0 ε = 0 , we have k ε ∈ H β , τ ( δ ) k_\varepsilon\in H^{(\delta)}_{\beta,\tau} k ε ∈ H β , τ ( δ ) if and only if δ < ε + 1 τ \delta<\varepsilon+\frac{1}{\tau} δ < ε + τ 1 . In particular, for every δ > 0 \delta>0 δ > 0 condition (H₂) is strictly stronger than the classical one, and it becomes more restrictive as p p p decreases. These classes are studied in § 3 .
Since p i > q i p_i>q_i p i > q i , (1) gives ∑ i 1 p i < ∑ i 1 q i = 1 − α n < 1 \sum_i\frac{1}{p_i}<\sum_i\frac{1}{q_i}=1-\frac{\alpha}{n}<1 ∑ i p i 1 < ∑ i q i 1 = 1 − n α < 1 . We define t ≥ 1 t\geq1 t ≥ 1 and a = a ( t , q ) a=a(t,q) a = a ( t , q ) by
1 p 1 + ⋯ + 1 p m + 1 t = 1 , 1 a = 1 q − ∑ i = 1 m 1 p i = 1 q + 1 t − 1 , (3) \frac{1}{p_1}+\cdots+\frac{1}{p_m}+\frac{1}{t}=1,\qquad\frac{1}{a}=\frac{1}{q}-\sum_{i=1}^{m}\frac{1}{p_i}=\frac{1}{q}+\frac{1}{t}-1, \tag*{(3)} p 1 1 + ⋯ + p m 1 + t 1 = 1 , a 1 = q 1 − i = 1 ∑ m p i 1 = q 1 + t 1 − 1 , (3)
that is, a ( t , q ) = t q t + q − t q a(t,q)=\frac{tq}{t+q-tq} a ( t , q ) = t + q − tq tq . Since p < 1 p<1 p < 1 , we have 1 q = 1 p − α n > 1 − α n > ∑ i 1 p i \frac{1}{q}=\frac{1}{p}-\frac{\alpha}{n}>1-\frac{\alpha}{n}>\sum_i\frac{1}{p_i} q 1 = p 1 − n α > 1 − n α > ∑ i p i 1 , so a a a is well defined and positive. Note also that t < n α t<\frac{n}{\alpha} t < α n and a ≥ q a\geq q a ≥ q , with equality only if p i = ∞ p_i=\infty p i = ∞ for every i i i .
The main result of this paper is the following.
Theorem 1.1. Let 0 < α < n 0<\alpha<n 0 < α < n and max { n n + 1 , n n + α } < p < 1 \max\left\{\frac{n}{n+1},\frac{n}{n+\alpha}\right\}<p<1 max { n + 1 n , n + α n } < p < 1 , δ = n ( 1 p − 1 ) \delta=n\left(\frac{1}{p}-1\right) δ = n ( p 1 − 1 ) , 1 q = 1 p − α n \frac{1}{q}=\frac{1}{p}-\frac{\alpha}{n} q 1 = p 1 − n α . Let T α T_\alpha T α be defined by (2) , and assume (H), (H1 _1 1 ) and (H2 _2 2 ). If w w w is a weight such that w a ( t , q ) ∈ A 1 w^{a(t,q)}\in A_1 w a ( t , q ) ∈ A 1 and w ( A i x ) ≤ C w ( x ) w(A_i x)\leq Cw(x) w ( A i x ) ≤ C w ( x ) a.e. x ∈ R n x\in\mathbb{R}^n x ∈ R n , 1 ≤ i ≤ m 1\leq i\leq m 1 ≤ i ≤ m , then T α T_\alpha T α extends to a bounded operator from H w p ( R n ) H^p_w(\mathbb{R}^n) H w p ( R n ) into L w q ( R n ) L^q_w(\mathbb{R}^n) L w q ( R n ) .
The restriction on p p p is equivalent to δ < 1 \delta<1 δ < 1 and q > 1 q>1 q > 1 . The condition δ < 1 \delta<1 δ < 1 allows us to use atoms with vanishing integral only (see the proof of Theorem 1.1 ), and q > 1 q>1 q > 1 is used in Minkowski’s integral inequality in Lemma 4.2 . Moreover, δ < 1 \delta<1 δ < 1 is natural for the hypotheses themselves: if δ > 1 \delta>1 δ > 1 , then (H1 _1 1 ) and (H2 _2 2 ) force k i ≡ 0 k_i\equiv0 k i ≡ 0 (Remark 3.6).
Let us comment on the hypotheses of Theorem 1.1 . Although (H2 _2 2 ) is strictly stronger than the classical fractional Hörmander condition, the classes H β , τ ( δ ) H^{(\delta)}_{\beta,\tau} H β , τ ( δ ) are large: conditions (H1 _1 1 ) and (H2 _2 2 ) are satisfied by many kernels, homogeneous or not, such as kernels with pointwise Hölder regularity of order ε > δ \varepsilon>\delta ε > δ , rough homogeneous kernels Ω i ( x ) ∣ x ∣ − n / q i \Omega_i(x)|x|^{-n/q_i} Ω i ( x ) ∣ x ∣ − n / q i satisfying an L p i L^{p_i} L p i -Dini condition of order δ \delta δ , and kernels obtained from Calderón–Zygmund kernels (see § 3 ). For homogeneous kernels, the extra decay required in (H2 _2 2 ) cannot be obtained by changing the index of a classical fractional Hörmander class (Theorem 3.14 ), and the factor 2 j δ 2^{j\delta} 2 j δ cannot be removed: in § 4.5 we give an operator whose kernels satisfy (H1 _1 1 ) and the classical fractional Hörmander condition, but for which the conclusion of Theorem 1.1 fails; in fact, in general δ \delta δ cannot be replaced in (H2 _2 2 ) by any δ ′ < δ − 1 q + 1 p 1 \delta'<\delta-\frac{1}{q}+\frac{1}{p_1} δ ′ < δ − q 1 + p 1 1 (Remark 4.7). Theorem 1.1 also covers the case m = 1 m=1 m = 1 of a single kernel. In that case, Han and Wu [15 Theorem 5.10] obtained a related weighted estimate, for 0 < α ≤ 1 0<\alpha\leq1 0 < α ≤ 1 , assuming the Hörmander condition of order α \alpha α ; Theorem 1.1 only requires the condition of order δ = n ( 1 p − 1 ) ≤ α \delta=n\left(\frac{1}{p}-1\right)\leq\alpha δ = n ( p 1 − 1 ) ≤ α , allows any 0 < α < n 0<\alpha<n 0 < α < n , and has a weaker condition on the weight (see Remark 4.6 for a detailed comparison). Finally, the hypotheses on w w w and on the matrices are satisfied, for instance, by the power weights w ( x ) = ∣ x ∣ γ w(x)=|x|^\gamma w ( x ) = ∣ x ∣ γ , − n / a ( t , q ) < γ ≤ 0 -n/a(t,q)<\gamma\leq0 − n / a ( t , q ) < γ ≤ 0 , and arbitrary matrices satisfying (H); more examples are given in § 5 .
The paper is organized as follows. In § 2 we recall some preliminaries on weights and weighted Hardy spaces. § 3 is devoted to the classes H β , τ ( δ ) H^{(\delta)}_{\beta,\tau} H β , τ ( δ ) : we recall their definition, prove that they form a strictly decreasing scale (Proposition 3.7 ), show that 0 ≤ δ < 1 0\leq\delta<1 0 ≤ δ < 1 is the only meaningful range (Remark 3.6), give sufficient conditions covering kernels with pointwise Hölder regularity, rough homogeneous kernels with Dini conditions of order δ \delta δ and kernels obtained from Calderón–Zygmund kernels, and study the homogeneous case (Theorem 3.14 ). § 4 contains the proof of Theorem 1.1 , which is built on these classes: an extension theorem with two weights, the technical lemma, where (H2 _2 2 ) is used precisely to absorb the factor 2 j δ 2^{j\delta} 2 j δ produced by the atoms, the unweighted estimate, the proof itself, a comparison with [15 ] , and the sharpness of (H2 _2 2 ). In Section 5 we give examples of matrices and weights satisfying the hypotheses of Theorem 1.1 .
Preliminaries
Weights
A weight is a function w ∈ L l o c 1 ( R n ) w \in L^1_{\mathrm{loc}}(\mathbb{R}^n) w ∈ L loc 1 ( R n ) with values in ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) almost everywhere. Given a weight w w w and 0 < p < ∞ 0 < p < \infty 0 < p < ∞ , L w p ( R n ) L^p_w(\mathbb{R}^n) L w p ( R n ) is the space of measurable functions with ∥ f ∥ L w p p = ∫ R n ∣ f ∣ p w < ∞ \|f\|_{L^p_w}^p = \int_{\mathbb{R}^n} |f|^p w < \infty ∥ f ∥ L w p p = ∫ R n ∣ f ∣ p w < ∞ . For a measurable set E E E and 0 < p < ∞ 0 < p < \infty 0 < p < ∞ we write w p ( E ) = ∫ E w p w^p(E) = \int_E w^p w p ( E ) = ∫ E w p . For p = ∞ p = \infty p = ∞ , ( ∫ E ∣ f ∣ p ) 1 / p (\int_E |f|^p)^{1/p} ( ∫ E ∣ f ∣ p ) 1/ p stands for ∥ f χ E ∥ ∞ \|f\chi_E\|_\infty ∥ f χ E ∥ ∞ .
The Hardy–Littlewood maximal operator is M f ( x ) = sup B ∋ x 1 ∣ B ∣ ∫ B ∣ f ∣ Mf(x) = \sup_{B\ni x} \frac{1}{|B|}\int_B |f| M f ( x ) = sup B ∋ x ∣ B ∣ 1 ∫ B ∣ f ∣ . For 1 < p < ∞ 1 < p < \infty 1 < p < ∞ , the Muckenhoupt class A p A_p A p consists of the weights w w w such that
[ w ] A p : = sup B ( 1 ∣ B ∣ ∫ B w ) ( 1 ∣ B ∣ ∫ B w − 1 p − 1 ) p − 1 < ∞ . [w]_{A_p}:=\sup_B \left(\frac{1}{|B|}\int_B w\right)\left(\frac{1}{|B|}\int_B w^{-\frac{1}{p-1}}\right)^{p-1}<\infty. [ w ] A p := B sup ( ∣ B ∣ 1 ∫ B w ) ( ∣ B ∣ 1 ∫ B w − p − 1 1 ) p − 1 < ∞.
A weight w w w belongs to A 1 A_1 A 1 if M w ≤ c w Mw \leq cw M w ≤ c w a.e.; equivalently, there exists C > 0 C > 0 C > 0 such that, for every ball B B B ,
1 ∣ B ∣ ∫ B w ≤ C e s s i n f B w . \frac{1}{|B|}\int_B w \leq C\mathop{\mathrm{ess\,inf}}_B w. ∣ B ∣ 1 ∫ B w ≤ C ess inf B w .
We set A ∞ = ⋃ p ≥ 1 A p A_\infty= \bigcup_{p\geq1} A_p A ∞ = ⋃ p ≥ 1 A p and denote by q ~ w = inf { q > 1 : w ∈ A q } \widetilde q_w = \inf\{q > 1 : w \in A_q\} q w = inf { q > 1 : w ∈ A q } the critical index of w w w .
A weight w w w satisfies the reverse Hölder inequality with exponent s > 1 s > 1 s > 1 , w ∈ R H s w \in RH_s w ∈ R H s , if there is C > 0 C > 0 C > 0 such that for every ball B B B
( 1 ∣ B ∣ ∫ B w s ) 1 / s ≤ C ∣ B ∣ ∫ B w . \left(\frac{1}{|B|}\int_B w^s\right)^{1/s} \leq\frac{C}{|B|}\int_B w. ( ∣ B ∣ 1 ∫ B w s ) 1/ s ≤ ∣ B ∣ C ∫ B w .
If w ∈ R H s w \in RH_s w ∈ R H s then w ∈ R H s ′ w \in RH_{s'} w ∈ R H s ′ for 1 < s ′ < s 1 < s' < s 1 < s ′ < s and w ∈ R H s + ϵ w \in RH_{s+\epsilon} w ∈ R H s + ϵ for some ϵ > 0 \epsilon> 0 ϵ > 0 . We denote by r w = sup { r > 1 : w ∈ R H r } r_w = \sup\{r > 1 : w \in RH_r\} r w = sup { r > 1 : w ∈ R H r } the critical index for the reverse Hölder condition. It is well known that w ∈ A ∞ w \in A_\infty w ∈ A ∞ if and only if w ∈ R H s w \in RH_s w ∈ R H s for some s > 1 s > 1 s > 1 ; thus 1 < r w ≤ ∞ 1 < r_w \leq\infty 1 < r w ≤ ∞ for all w ∈ A ∞ w \in A_\infty w ∈ A ∞ .
We shall use the following well-known properties of A 1 A_1 A 1 weights; see, for instance, [11 ] , [8 ] or [13 ] .
Lemma 2.1. Let v ∈ A 1 v \in A_1 v ∈ A 1 .
(i) v θ ∈ A 1 v^\theta\in A_1 v θ ∈ A 1 for every 0 < θ ≤ 1 0 < \theta\leq1 0 < θ ≤ 1 .
(ii) There exists ϵ > 0 \epsilon> 0 ϵ > 0 such that v 1 + ϵ ∈ A 1 v^{1+\epsilon} \in A_1 v 1 + ϵ ∈ A 1 .
(iii) v θ ∈ R H 1 / θ v^\theta\in RH_{1/\theta} v θ ∈ R H 1/ θ for every 0 < θ < 1 0 < \theta< 1 0 < θ < 1 .
(iv) v ( λ B ) ≤ C λ n v ( B ) v(\lambda B) \leq C\lambda^n v(B) v ( λ B ) ≤ C λ n v ( B ) for every ball B B B and λ ≥ 1 \lambda\geq1 λ ≥ 1 .
Property (i) is a consequence of Jensen’s inequality, (ii) follows from the reverse Hölder inequality satisfied by A ∞ A_\infty A ∞ weights, and (iii) and (iv) follow directly from the definition of A 1 A_1 A 1 .
Weighted Hardy spaces
Topologize S ( R n ) \mathcal{S}(\mathbb{R}^n) S ( R n ) by the semi-norms
∥ ϕ ∥ μ , κ = sup x ∣ x μ ∂ κ ϕ ( x ) ∣ . \|\phi\|_{\mu,\kappa}=\sup_x |x^\mu\partial^\kappa\phi(x)|. ∥ ϕ ∥ μ , κ = x sup ∣ x μ ∂ κ ϕ ( x ) ∣.
For N ∈ N N \in\mathbb{N} N ∈ N let
S N = { ϕ ∈ S ( R n ) : ∥ ϕ ∥ μ , κ ≤ 1 , ∣ μ ∣ , ∣ κ ∣ ≤ N } , \mathcal{S}_N=\{\phi\in\mathcal{S}(\mathbb{R}^n):\|\phi\|_{\mu,\kappa}\leq1,\ |\mu|,|\kappa|\leq N\}, S N = { ϕ ∈ S ( R n ) : ∥ ϕ ∥ μ , κ ≤ 1 , ∣ μ ∣ , ∣ κ ∣ ≤ N } ,
and for f ∈ S ′ ( R n ) f \in\mathcal{S}'(\mathbb{R}^n) f ∈ S ′ ( R n ) let
M N f ( x ) = sup t > 0 sup ϕ ∈ S N ∣ t − n ϕ ( t − 1 ⋅ ) ∗ f ( x ) ∣ . M_Nf(x)=\sup_{t>0}\sup_{\phi\in\mathcal{S}_N}\left|t^{-n}\phi(t^{-1}\cdot)*f(x)\right|. M N f ( x ) = t > 0 sup ϕ ∈ S N sup t − n ϕ ( t − 1 ⋅ ) ∗ f ( x ) .
Given w ∈ A ∞ w\in A_\infty w ∈ A ∞ and p > 0 p>0 p > 0 , the weighted Hardy space H w p ( R n ) H_w^p(\mathbb{R}^n) H w p ( R n ) consists of all f ∈ S ′ ( R n ) f\in\mathcal{S}'(\mathbb{R}^n) f ∈ S ′ ( R n ) such that ∥ f ∥ H w p = ∥ M N f ∥ L w p < ∞ \|f\|_{H_w^p}=\|M_Nf\|_{L_w^p}<\infty ∥ f ∥ H w p = ∥ M N f ∥ L w p < ∞ , for N N N sufficiently large (see [24 Chapter VI] ). Let
D 0 = { ϕ ∈ S ( R n ) : ϕ ^ ∈ C c ∞ ( R n ) , 0 ∉ supp ϕ ^ } \mathcal{D}_0=\{\phi\in\mathcal{S}(\mathbb{R}^n):\widehat{\phi}\in C_c^\infty(\mathbb{R}^n),\ 0\notin\operatorname{supp}\widehat{\phi}\} D 0 = { ϕ ∈ S ( R n ) : ϕ ∈ C c ∞ ( R n ) , 0 ∈ / supp ϕ }
(see [24 Chapter VI] ).
Theorem 2.2 ([24 Chapter VII, Theorem 1] ). Let w w w be a doubling weight on R n \mathbb{R}^n R n . Then D 0 \mathcal{D}_0 D 0 is dense in H w p ( R n ) H_w^p(\mathbb{R}^n) H w p ( R n ) , 0 < p < ∞ 0<p<\infty 0 < p < ∞ .
We recall the definition of w w w -( p , p 0 , d ) \left(p,p_0,d\right) ( p , p 0 , d ) atom given in [22 ] . Let w ∈ A ∞ w\in A_\infty w ∈ A ∞ , 0 < p ≤ 1 0<p\leq1 0 < p ≤ 1 , max { 1 , p r w r w − 1 } < p 0 ≤ ∞ \max\{1,p\frac{r_w}{r_w-1}\}<p_0\leq\infty max { 1 , p r w − 1 r w } < p 0 ≤ ∞ (where r w r w − 1 = 1 \frac{r_w}{r_w-1}=1 r w − 1 r w = 1 if r w = ∞ r_w=\infty r w = ∞ ), and d ∈ Z d\in\mathbb{Z} d ∈ Z with d ≥ ⌊ n ( q w p − 1 ) ⌋ d\geq\lfloor n(\frac{q_w}{p}-1)\rfloor d ≥ ⌊ n ( p q w − 1 )⌋ . A function a a a is a w w w -( p , p 0 , d ) \left(p,p_0,d\right) ( p , p 0 , d ) atom centered at x 0 x_0 x 0 if
(a1) a ∈ L p 0 ( R n ) a\in L^{p_0}(\mathbb{R}^n) a ∈ L p 0 ( R n ) is supported in a ball B = B ( x 0 , r ) B=B(x_0,r) B = B ( x 0 , r ) ;
(a2) ∥ a ∥ L p 0 ≤ ∣ B ∣ 1 / p 0 w ( B ) − 1 / p \|a\|_{L^{p_0}}\leq|B|^{1/p_0}w(B)^{-1/p} ∥ a ∥ L p 0 ≤ ∣ B ∣ 1/ p 0 w ( B ) − 1/ p ;
(a3) ∫ x μ a ( x ) d x = 0 \int x^\mu a(x)\,dx=0 ∫ x μ a ( x ) d x = 0 for every multi-index μ \mu μ with ∣ μ ∣ ≤ d |\mu|\leq d ∣ μ ∣ ≤ d .
Lemma 2.3 ([22 Lemma 2.8] ). There exists C > 0 C>0 C > 0 such that ∥ a ∥ H w p ≤ C \|a\|_{H_w^p}\leq C ∥ a ∥ H w p ≤ C for every w w w -( p , p 0 , d ) \left(p,p_0,d\right) ( p , p 0 , d ) atom a a a .
Theorem 2.4 ([22 Theorem 2.9] ). Let w ∈ A ∞ w\in A_\infty w ∈ A ∞ , 0 < p ≤ 1 0<p\leq1 0 < p ≤ 1 , let p 0 p_0 p 0 and d d d be as in the definition of w w w -( p , p 0 , d ) \left(p,p_0,d\right) ( p , p 0 , d ) atom, and let f ∈ D 0 f\in\mathcal{D}_0 f ∈ D 0 . There exist w w w -( p , p 0 , d ) \left(p,p_0,d\right) ( p , p 0 , d ) atoms a j a_j a j and scalars λ j \lambda_j λ j with ∑ j ∣ λ j ∣ p ≤ c ∥ f ∥ H w p p \sum_j|\lambda_j|^p\leq c\|f\|_{H_w^p}^p ∑ j ∣ λ j ∣ p ≤ c ∥ f ∥ H w p p such that f = ∑ j λ j a j f=\sum_j\lambda_ja_j f = ∑ j λ j a j , where the series converges in L s ( R n ) L^s(\mathbb{R}^n) L s ( R n ) for every 1 < s < ∞ 1<s<\infty 1 < s < ∞ and pointwise.
The atoms constructed in the proof of [22 Theorem 2.9] satisfy ∥ a ∥ L ∞ ≤ w ( B ) − 1 / p \|a\|_{L^\infty}\leq w(B)^{-1/p} ∥ a ∥ L ∞ ≤ w ( B ) − 1/ p , so they are w w w -( p , p 0 , d ) \left(p,p_0,d\right) ( p , p 0 , d ) atoms for every admissible p 0 p_0 p 0 ; moreover, the order d d d of their vanishing moments can be prescribed.
Size and strengthened Hörmander conditions
Definitions and basic properties
The normalized annular norms introduced in Section 1 are invariant under dilations: if f λ ( x ) = f ( x / λ ) f_\lambda(x)=f(x/\lambda) f λ ( x ) = f ( x / λ ) , then ∥ f λ ∥ τ , ∣ x ∣ ∼ λ s = ∥ f ∥ τ , ∣ x ∣ ∼ s \|f_\lambda\|_{\tau,|x|\sim\lambda s}=\|f\|_{\tau,|x|\sim s} ∥ f λ ∥ τ , ∣ x ∣ ∼ λ s = ∥ f ∥ τ , ∣ x ∣ ∼ s . By Hölder’s inequality, ∥ f ∥ τ 1 , ∣ x ∣ ∼ s ≤ ∥ f ∥ τ 2 , ∣ x ∣ ∼ s \|f\|_{\tau_1,|x|\sim s}\leq\|f\|_{\tau_2,|x|\sim s} ∥ f ∥ τ 1 , ∣ x ∣ ∼ s ≤ ∥ f ∥ τ 2 , ∣ x ∣ ∼ s whenever 1 ≤ τ 1 ≤ τ 2 ≤ ∞ 1\leq\tau_1\leq\tau_2\leq\infty 1 ≤ τ 1 ≤ τ 2 ≤ ∞ .
Definition 3.1 (Size condition, [1 Definition 3.1] ). Let 1 ≤ τ ≤ ∞ 1\leq\tau\leq\infty 1 ≤ τ ≤ ∞ and 0 ≤ β < n 0\leq\beta<n 0 ≤ β < n . A measurable function k k k belongs to S β , τ S_{\beta,\tau} S β , τ if there exists C > 0 C>0 C > 0 such that
∥ k ∥ τ , ∣ x ∣ ∼ s ≤ C s β − n for all s > 0. \|k\|_{\tau,|x|\sim s}\leq Cs^{\beta-n}\qquad\text{for all }s>0. ∥ k ∥ τ , ∣ x ∣ ∼ s ≤ C s β − n for all s > 0.
We now recall the strengthened fractional Hörmander condition considered in this paper, which was introduced by Han and Wu [15 Definition 5.6] under the name of fractional L τ L^\tau L τ -Hörmander condition of order δ \delta δ . It is obtained from the fractional Hörmander condition of [1 ] by adding the factor 2 j δ 2^{j\delta} 2 j δ , which is adapted to the H w p H_w^p H w p –L q L^q L q setting of Theorem 1.1 : for δ = n ( 1 p − 1 ) \delta=n(\frac{1}{p}-1) δ = n ( p 1 − 1 ) , it is exactly the factor that appears when T α T_\alpha T α is applied to the atoms of H w p p H_{w^p}^p H w p p (see Remark 4.3).
Definition 3.2 ([15 Definition 5.6] ). Let 1 ≤ τ ≤ ∞ 1 \le\tau\le\infty 1 ≤ τ ≤ ∞ , 0 ≤ β < n 0 \le\beta< n 0 ≤ β < n , and δ ≥ 0 \delta\ge0 δ ≥ 0 . A measurable function k k k belongs to H β , τ ( δ ) H_{\beta,\tau}^{(\delta)} H β , τ ( δ ) if there exist c ≥ 1 c \ge1 c ≥ 1 and C > 0 C > 0 C > 0 such that, for every h ∈ R n h \in\mathbb{R}^{n} h ∈ R n and R > c ∣ h ∣ R > c|h| R > c ∣ h ∣ ,
∑ j = 1 ∞ 2 j δ ( 2 j R ) n − β ∥ k ( ⋅ − h ) − k ∥ τ , ∣ x ∣ ∼ 2 j R ≤ C . \sum_{j=1}^{\infty} 2^{j\delta}(2^{j}R)^{n-\beta}\lVert k(\cdot-h)-k\rVert_{\tau,|x|\sim2^{j}R} \le C. j = 1 ∑ ∞ 2 j δ ( 2 j R ) n − β ∥ k ( ⋅ − h ) − k ∥ τ , ∣ x ∣ ∼ 2 j R ≤ C . For δ = 0 \delta=0 δ = 0 this is the fractional Hörmander condition H β , τ H_{\beta,\tau} H β , τ of [1 Definition 3.2] (the case k = 0 k=0 k = 0 there).
The formulation in [15 ] differs from Definition 3.2 only in the normalization of the averages and in the constant c c c , and there the order is restricted to 0 ≤ δ ≤ 1 0 \le\delta\le1 0 ≤ δ ≤ 1 .
With this notation, ( H 1 ) (\mathrm{H}_{1}) ( H 1 ) and ( H 2 ) (\mathrm{H}_{2}) ( H 2 ) say that k i ∈ S n − n / q i , p i k_i \in S_{n-n/q_i,p_i} k i ∈ S n − n / q i , p i and k i ∈ H n − n / q i , p i ( δ ) k_i \in H_{n-n/q_i,p_i}^{(\delta)} k i ∈ H n − n / q i , p i ( δ ) for 1 ≤ i ≤ m 1 \le i \le m 1 ≤ i ≤ m .
Remark 3.3. Since 2 j δ 1 ≤ 2 j δ 2 2^{j\delta_1} \le2^{j\delta_2} 2 j δ 1 ≤ 2 j δ 2 for j ≥ 1 j \ge1 j ≥ 1 and by the monotonicity of the normalized norms in τ \tau τ , we have
H β , τ 2 ( δ 2 ) ⊆ H β , τ 1 ( δ 1 ) ⊆ H β , τ 1 , S β , τ 2 ⊆ S β , τ 1 , H_{\beta,\tau_2}^{(\delta_2)} \subseteq H_{\beta,\tau_1}^{(\delta_1)} \subseteq H_{\beta,\tau_1}, \qquad S_{\beta,\tau_2} \subseteq S_{\beta,\tau_1}, H β , τ 2 ( δ 2 ) ⊆ H β , τ 1 ( δ 1 ) ⊆ H β , τ 1 , S β , τ 2 ⊆ S β , τ 1 ,
whenever 1 ≤ τ 1 ≤ τ 2 ≤ ∞ 1 \le\tau_1 \le\tau_2 \le\infty 1 ≤ τ 1 ≤ τ 2 ≤ ∞ and 0 ≤ δ 1 ≤ δ 2 0 \le\delta_1 \le\delta_2 0 ≤ δ 1 ≤ δ 2 .
Example 3.4 (Homogeneous kernels). Let Ω i ∈ L p i ( S n − 1 ) \Omega_i \in L^{p_i}(S^{n-1}) Ω i ∈ L p i ( S n − 1 ) , q i < p i ≤ ∞ q_i < p_i \le\infty q i < p i ≤ ∞ , be homogeneous of degree zero and k i ( x ) = Ω i ( x ) ∣ x ∣ − n / q i k_i(x)=\Omega_i(x)|x|^{-n/q_i} k i ( x ) = Ω i ( x ) ∣ x ∣ − n / q i . If p i < ∞ p_i<\infty p i < ∞ , polar coordinates give
∥ k i ∥ p i , ∣ x ∣ ∼ s p i = ∥ Ω i ∥ L p i ( S n − 1 ) p i ∣ B ( 0 , 2 s ) ∣ ∫ s 2 s r n − 1 − n p i / q i d r = C s − n p i / q i , \lVert k_i\rVert_{p_i,|x|\sim s}^{p_i}=\frac{\lVert\Omega_i\rVert_{L^{p_i}(S^{n-1})}^{p_i}}{|B(0,2s)|}\int_s^{2s}r^{n-1-np_i/q_i}\,dr=C\,s^{-np_i/q_i}, ∥ k i ∥ p i , ∣ x ∣ ∼ s p i = ∣ B ( 0 , 2 s ) ∣ ∥ Ω i ∥ L p i ( S n − 1 ) p i ∫ s 2 s r n − 1 − n p i / q i d r = C s − n p i / q i ,
and the case p i = ∞ p_i=\infty p i = ∞ is immediate. Hence k i ∈ S n − n / q i , p i k_i \in S_{n-n/q_i,p_i} k i ∈ S n − n / q i , p i and ( H 1 ) (\mathrm{H}_{1}) ( H 1 ) holds. This is the setting of [20 ] .
The following consequence of the size condition replaces the polar coordinates computation available for homogeneous kernels.
Lemma 3.5. Let 1 < q i < p i ≤ ∞ 1<q_i<p_i\le\infty 1 < q i < p i ≤ ∞ , γ i = n q i − n p i > 0 \gamma_i=\frac{n}{q_i}-\frac{n}{p_i}>0 γ i = q i n − p i n > 0 , and k i ∈ S n − n / q i , p i k_i\in S_{n-n/q_i,p_i} k i ∈ S n − n / q i , p i . Then there is C > 0 C>0 C > 0 such that
∥ k i χ { ∣ x ∣ > u } ∥ L p i ≤ C u − γ i for all u > 0. \lVert k_i\chi_{\{|x|>u\}}\rVert_{L^{p_i}} \le C u^{-\gamma_i}\quad\text{for all }u>0. ∥ k i χ { ∣ x ∣ > u } ∥ L p i ≤ C u − γ i for all u > 0. Proof. For l ≥ 0 l\ge0 l ≥ 0 , by the size condition k i ∈ S n − n / q i , p i k_i\in S_{n-n/q_i,p_i} k i ∈ S n − n / q i , p i ,
∥ k i χ { 2 l u < ∣ x ∣ ≤ 2 l + 1 u } ∥ L p i = ∣ B ( 0 , 2 l + 1 u ) ∣ 1 / p i ∥ k i ∥ p i , ∣ x ∣ ∼ 2 l u ≤ C ( 2 l u ) n p i − n q i = C ( 2 l u ) − γ i , \lVert k_i\chi_{\{2^lu<|x|\le2^{l+1}u\}}\rVert_{L^{p_i}}=|B(0,2^{l+1}u)|^{1/p_i}\lVert k_i\rVert_{p_i,|x|\sim2^lu}\le C(2^lu)^{\frac{n}{p_i}-\frac{n}{q_i}}=C(2^lu)^{-\gamma_i}, ∥ k i χ { 2 l u < ∣ x ∣ ≤ 2 l + 1 u } ∥ L p i = ∣ B ( 0 , 2 l + 1 u ) ∣ 1/ p i ∥ k i ∥ p i , ∣ x ∣ ∼ 2 l u ≤ C ( 2 l u ) p i n − q i n = C ( 2 l u ) − γ i , with the obvious interpretation if p i = ∞ p_i=\infty p i = ∞ . Summing over l ≥ 0 l\ge0 l ≥ 0 and using the triangle inequality in L p i L^{p_i} L p i gives the claim, since γ i > 0 \gamma_i>0 γ i > 0 . □ \square □
Remark 3.6. The restriction δ < 1 \delta<1 δ < 1 is also natural from the point of view of the hypotheses: if δ > 1 \delta>1 δ > 1 , then S β , τ ∩ H β , τ ( δ ) = { 0 } S_{\beta,\tau}\cap H_{\beta,\tau}^{(\delta)}=\{0\} S β , τ ∩ H β , τ ( δ ) = { 0 } for every 1 ≤ τ ≤ ∞ 1\le\tau\le\infty 1 ≤ τ ≤ ∞ and 0 ≤ β < n 0\le\beta<n 0 ≤ β < n , so that ( H 1 ) (\mathrm{H}_{1}) ( H 1 ) and ( H 2 ) (\mathrm{H}_{2}) ( H 2 ) would force T α ≡ 0 T_\alpha\equiv0 T α ≡ 0 . Indeed, let k ∈ S β , τ ∩ H β , τ ( δ ) k\in S_{\beta,\tau}\cap H_{\beta,\tau}^{(\delta)} k ∈ S β , τ ∩ H β , τ ( δ ) with constants c c c and C C C , fix s > 0 s>0 s > 0 and let 0 < ∣ h ∣ < s / ( 2 c ) 0<|h|<s/(2c) 0 < ∣ h ∣ < s / ( 2 c ) . Choose j ≥ 1 j\ge1 j ≥ 1 with s / ( c 2 j + 1 ) ≤ ∣ h ∣ < s / ( c 2 j ) s/(c2^{j+1})\le|h|<s/(c2^j) s / ( c 2 j + 1 ) ≤ ∣ h ∣ < s / ( c 2 j ) and apply Definition 3.2 with R = 2 − j s > c ∣ h ∣ R=2^{-j}s>c|h| R = 2 − j s > c ∣ h ∣ . Keeping only the j j j -th term of the sum, we get
∥ k ( ⋅ − h ) − k ∥ τ , ∣ x ∣ ∼ s ≤ C s β − n 2 − j δ ≤ C s β − n ( 2 c ∣ h ∣ s ) δ , \lVert k(\cdot-h)-k\rVert_{\tau,|x|\sim s}\le Cs^{\beta-n}2^{-j\delta}\le Cs^{\beta-n}\left(\frac{2c|h|}{s}\right)^\delta, ∥ k ( ⋅ − h ) − k ∥ τ , ∣ x ∣ ∼ s ≤ C s β − n 2 − j δ ≤ C s β − n ( s 2 c ∣ h ∣ ) δ ,
which is o ( ∣ h ∣ ) o(|h|) o ( ∣ h ∣ ) as h → 0 h \to0 h → 0 , since δ > 1 \delta> 1 δ > 1 . Hence, if φ ∈ C c ∞ ( { s < ∣ x ∣ < 2 s } ) \varphi\in C^\infty_c(\{s < |x| < 2s\}) φ ∈ C c ∞ ({ s < ∣ x ∣ < 2 s }) and 1 ≤ l ≤ n 1 \le l \le n 1 ≤ l ≤ n , Hölder’s inequality gives
∣ ∫ k ∂ l φ ∣ = lim t → 0 ∣ ∫ k ( x − t e l ) − k ( x ) t φ ( x ) d x ∣ ≤ C s , φ lim t → 0 ∥ k ( ⋅ − t e l ) − k ∥ τ , ∣ x ∣ ∼ s ∣ t ∣ = 0. \left|\int k \partial_l\varphi\right|=\lim_{t\to0}\left|\int\frac{k(x-te_l)-k(x)}{t}\varphi(x)\,dx\right|\le C_{s,\varphi}\lim_{t\to0}\frac{\|k(\cdot-te_l)-k\|_{\tau,|x|\sim s}}{|t|}=0. ∫ k ∂ l φ = t → 0 lim ∫ t k ( x − t e l ) − k ( x ) φ ( x ) d x ≤ C s , φ t → 0 lim ∣ t ∣ ∥ k ( ⋅ − t e l ) − k ∥ τ , ∣ x ∣ ∼ s = 0.
Since the open annuli { s < ∣ x ∣ < 2 s } \{s < |x| < 2s\} { s < ∣ x ∣ < 2 s } , s > 0 s > 0 s > 0 , cover R n ∖ { 0 } \mathbb{R}^n \setminus\{0\} R n ∖ { 0 } , we get ∇ k = 0 \nabla k = 0 ∇ k = 0 in D ′ ( R n ∖ { 0 } ) \mathcal{D}'(\mathbb{R}^n \setminus\{0\}) D ′ ( R n ∖ { 0 }) , so k k k coincides a.e. with a constant c V c_V c V on each connected component V V V of R n ∖ { 0 } \mathbb{R}^n\setminus\{0\} R n ∖ { 0 } . As ∣ V ∩ { ∣ x ∣ ∼ s } ∣ / ∣ B ( 0 , 2 s ) ∣ |V\cap\{|x|\sim s\}|/|B(0,2s)| ∣ V ∩ { ∣ x ∣ ∼ s } ∣/∣ B ( 0 , 2 s ) ∣ does not depend on s s s , the size condition gives ∣ c V ∣ ≤ C s β − n |c_V|\le Cs^{\beta-n} ∣ c V ∣ ≤ C s β − n for every s > 0 s>0 s > 0 , and letting s → ∞ s\to\infty s → ∞ we obtain c V = 0 c_V=0 c V = 0 . In particular, the restriction δ ≤ 1 \delta\le1 δ ≤ 1 in [15 Definition 5.6] entails no loss of generality.
By Remark 3.3, the classes H β , τ ( δ ) H_{\beta,\tau}^{(\delta)} H β , τ ( δ ) decrease as δ \delta δ increases. The next result shows that these inclusions are strict. In part (a) we consider the homogeneous kernels k ε ( x ) = Ω ε ( x ) ∣ x ∣ β − n k_\varepsilon(x)=\Omega_\varepsilon(x)|x|^{\beta-n} k ε ( x ) = Ω ε ( x ) ∣ x ∣ β − n , with Ω ε ( x ) = ∣ x n / ∣ x ∣ ∣ ε \Omega_\varepsilon(x)=|x_n/|x||^\varepsilon Ω ε ( x ) = ∣ x n /∣ x ∣ ∣ ε , and determine exactly the values of δ \delta δ for which k ε ∈ H β , τ ( δ ) k_\varepsilon\in H_{\beta,\tau}^{(\delta)} k ε ∈ H β , τ ( δ ) : these are the δ < ε + 1 τ \delta<\varepsilon+\frac{1}{\tau} δ < ε + τ 1 . For ε > 0 \varepsilon>0 ε > 0 , the function Ω ε \Omega_\varepsilon Ω ε is Hölder continuous of order ε \varepsilon ε and vanishes on the hyperplane { x n = 0 } \{x_n=0\} { x n = 0 } , while for ε < 0 \varepsilon<0 ε < 0 it is singular there; the term 1 τ \frac{1}{\tau} τ 1 appears because the irregularity of k ε k_\varepsilon k ε is concentrated near this hyperplane. Since the threshold ε + 1 τ \varepsilon+\frac{1}{\tau} ε + τ 1 takes every value in ( 0 , 1 ) (0,1) ( 0 , 1 ) except 1 τ \frac{1}{\tau} τ 1 , any two classes H β , τ ( δ 1 ) H_{\beta,\tau}^{(\delta_1)} H β , τ ( δ 1 ) and H β , τ ( δ 2 ) H_{\beta,\tau}^{(\delta_2)} H β , τ ( δ 2 ) with δ 1 < δ 2 \delta_1<\delta_2 δ 1 < δ 2 , δ 1 < 1 \delta_1<1 δ 1 < 1 , are separated by some k ε k_\varepsilon k ε ; this is part (b).
Proposition 3.7. Let n ≥ 2 n\ge2 n ≥ 2 , 0 ≤ β < n 0\le\beta<n 0 ≤ β < n and 1 ≤ τ ≤ ∞ 1\le\tau\le\infty 1 ≤ τ ≤ ∞ , where 1 τ = 0 \frac{1}{\tau}=0 τ 1 = 0 if τ = ∞ \tau=\infty τ = ∞ .
(a) Let − 1 τ < ε < 1 − 1 τ -\frac{1}{\tau}<\varepsilon<1-\frac{1}{\tau} − τ 1 < ε < 1 − τ 1 , ε ≠ 0 \varepsilon\ne0 ε = 0 , and k ε ( x ) = ∣ x n ∣ ε ∣ x ∣ β − n − ε k_\varepsilon(x)=|x_n|^\varepsilon|x|^{\beta-n-\varepsilon} k ε ( x ) = ∣ x n ∣ ε ∣ x ∣ β − n − ε . Then k ε ∈ S β , τ k_\varepsilon\in S_{\beta,\tau} k ε ∈ S β , τ , and k ε ∈ H β , τ ( δ ) k_\varepsilon\in H_{\beta,\tau}^{(\delta)} k ε ∈ H β , τ ( δ ) if and only if δ < ε + 1 τ \delta<\varepsilon+\frac{1}{\tau} δ < ε + τ 1 .
(b) If 0 ≤ δ 1 < δ 2 0\le\delta_1<\delta_2 0 ≤ δ 1 < δ 2 and δ 1 < 1 \delta_1<1 δ 1 < 1 , then
S β , τ ∩ H β , τ ( δ 2 ) ⊊ S β , τ ∩ H β , τ ( δ 1 ) . S_{\beta,\tau}\cap H_{\beta,\tau}^{(\delta_2)}\subsetneq S_{\beta,\tau}\cap H_{\beta,\tau}^{(\delta_1)}. S β , τ ∩ H β , τ ( δ 2 ) ⊊ S β , τ ∩ H β , τ ( δ 1 ) .
Proof. (a) We write x = ( x ~ , x n ) ∈ R n − 1 × R x=(\widetilde{x},x_n)\in\mathbb{R}^{n-1}\times\mathbb{R} x = ( x , x n ) ∈ R n − 1 × R . Let θ = ε + 1 τ \theta=\varepsilon+\frac{1}{\tau} θ = ε + τ 1 , so that 0 < θ < 1 0<\theta<1 0 < θ < 1 , and write k ε ( x ) = g ( x n ) ϱ ( x ) k_\varepsilon(x)=g(x_n)\varrho(x) k ε ( x ) = g ( x n ) ϱ ( x ) , with g ( u ) = ∣ u ∣ ε g(u)=|u|^\varepsilon g ( u ) = ∣ u ∣ ε and ϱ ( x ) = ∣ x ∣ β − n − ε \varrho(x)=|x|^{\beta-n-\varepsilon} ϱ ( x ) = ∣ x ∣ β − n − ε . Since ε τ > − 1 \varepsilon\tau>-1 ε τ > − 1 , the function ∣ x n / ∣ x ∣ ∣ ε |x_n/|x||^\varepsilon ∣ x n /∣ x ∣ ∣ ε belongs to L τ ( S n − 1 ) L^\tau(S^{n-1}) L τ ( S n − 1 ) , and k ε ∈ S β , τ k_\varepsilon\in S_{\beta,\tau} k ε ∈ S β , τ as in Example 3.4.
We use two elementary facts. First, for b ∈ R b\in\mathbb{R} b ∈ R ,
( ∫ R ∣ ∣ u − b ∣ ε − ∣ u ∣ ε ∣ τ d u ) 1 / τ = C ε , τ ∣ b ∣ ε + 1 τ , C ε , τ < ∞ . (4) \left(\int_{\mathbb{R}}\left||u-b|^\varepsilon-|u|^\varepsilon\right|^\tau\,du\right)^{1/\tau}=C_{\varepsilon,\tau}|b|^{\varepsilon+\frac{1}{\tau}},\qquad C_{\varepsilon,\tau}<\infty. \tag*{(4)} ( ∫ R ∣ ∣ u − b ∣ ε − ∣ u ∣ ε ∣ τ d u ) 1/ τ = C ε , τ ∣ b ∣ ε + τ 1 , C ε , τ < ∞. (4)
Indeed, the change of variables u = ∣ b ∣ v u=|b|v u = ∣ b ∣ v reduces (4) to b = ± 1 b=\pm1 b = ± 1 , and the integral converges because ε τ > − 1 \varepsilon\tau>-1 ε τ > − 1 (near v = 0 v=0 v = 0 and v = ± 1 v=\pm1 v = ± 1 ) and ( ε − 1 ) τ < − 1 (\varepsilon-1)\tau<-1 ( ε − 1 ) τ < − 1 (at infinity, where the integrand is O ( ∣ v ∣ ( ε − 1 ) τ ) O(|v|^{(\varepsilon-1)\tau}) O ( ∣ v ∣ ( ε − 1 ) τ ) ); if τ = ∞ \tau=\infty τ = ∞ , then 0 < ε < 1 0<\varepsilon<1 0 < ε < 1 and ∣ ∣ u − b ∣ ε − ∣ u ∣ ε ∣ ≤ ∣ b ∣ ε \left||u-b|^\varepsilon-|u|^\varepsilon\right|\le|b|^\varepsilon ∣ ∣ u − b ∣ ε − ∣ u ∣ ε ∣ ≤ ∣ b ∣ ε .
Second, since { ∣ x ∣ ≤ 2 s } ⊂ { ∣ x ~ ∣ ≤ 2 s , ∣ x n ∣ ≤ 2 s } \{|x|\le2s\}\subset\{|\widetilde{x}|\le2s,\ |x_n|\le2s\} { ∣ x ∣ ≤ 2 s } ⊂ { ∣ x ∣ ≤ 2 s , ∣ x n ∣ ≤ 2 s } , for every function F F F on R \mathbb{R} R we have
∥ F ( x n ) ∥ τ , ∣ x ∣ ∼ s ≤ C ( 1 s ∫ − 2 s 2 s ∣ F ( u ) ∣ τ d u ) 1 / τ . (5) \|F(x_n)\|_{\tau,|x|\sim s}\le C\left(\frac{1}{s}\int_{-2s}^{2s}|F(u)|^\tau\,du\right)^{1/\tau}. \tag*{(5)} ∥ F ( x n ) ∥ τ , ∣ x ∣ ∼ s ≤ C ( s 1 ∫ − 2 s 2 s ∣ F ( u ) ∣ τ d u ) 1/ τ . (5)
We first show that k ε ∈ H β , τ ( δ ) k_\varepsilon\in H_{\beta,\tau}^{(\delta)} k ε ∈ H β , τ ( δ ) if δ < θ \delta<\theta δ < θ . Let h ∈ R n h\in\mathbb{R}^n h ∈ R n , t = ∣ h ∣ t=|h| t = ∣ h ∣ , R > 4 t R>4t R > 4 t , j ≥ 1 j\ge1 j ≥ 1 and s = 2 j R s=2^jR s = 2 j R , so that t < s / 8 t<s/8 t < s /8 . If ∣ x ∣ ∼ s |x|\sim s ∣ x ∣ ∼ s , every point of the segment joining x − h x-h x − h and x x x has modulus between 7 s / 8 7s/8 7 s /8 and 3 s 3s 3 s , so ∣ ϱ ( x − h ) ∣ ≤ C s β − n − ε |\varrho(x-h)|\le Cs^{\beta-n-\varepsilon} ∣ ϱ ( x − h ) ∣ ≤ C s β − n − ε and, by the mean value theorem, ∣ ϱ ( x − h ) − ϱ ( x ) ∣ ≤ C t s s β − n − ε |\varrho(x-h)-\varrho(x)|\le C\frac{t}{s}s^{\beta-n-\varepsilon} ∣ ϱ ( x − h ) − ϱ ( x ) ∣ ≤ C s t s β − n − ε . Writing
k ε ( x − h ) − k ε ( x ) = [ g ( x n − h n ) − g ( x n ) ] ϱ ( x − h ) + g ( x n ) [ ϱ ( x − h ) − ϱ ( x ) ] , k_\varepsilon(x-h)-k_\varepsilon(x)=[g(x_n-h_n)-g(x_n)]\varrho(x-h)+g(x_n)[\varrho(x-h)-\varrho(x)], k ε ( x − h ) − k ε ( x ) = [ g ( x n − h n ) − g ( x n )] ϱ ( x − h ) + g ( x n ) [ ϱ ( x − h ) − ϱ ( x )] ,
and using (5) , (4) with b = h n b=h_n b = h n , and ∫ − 2 s 2 s ∣ u ∣ ε d u = C s ε + 1 \int_{-2s}^{2s}|u|^\varepsilon\,du=C s^{\varepsilon+1} ∫ − 2 s 2 s ∣ u ∣ ε d u = C s ε + 1 , we get
s n − β ∥ k ε ( ⋅ − h ) − k ε ∥ τ , ∣ x ∣ ∼ s ≤ C ( t s ) θ + C t s ≤ C ( t s ) θ , s^{n-\beta}\|k_\varepsilon(\cdot-h)-k_\varepsilon\|_{\tau,|x|\sim s}\leq C\left(\frac{t}{s}\right)^\theta+C\frac{t}{s}\leq C\left(\frac{t}{s}\right)^\theta, s n − β ∥ k ε ( ⋅ − h ) − k ε ∥ τ , ∣ x ∣ ∼ s ≤ C ( s t ) θ + C s t ≤ C ( s t ) θ ,
since θ < 1 \theta<1 θ < 1 and t < s t<s t < s . Hence
∑ j = 1 ∞ 2 j δ ( 2 j R ) n − β ∥ k ε ( ⋅ − h ) − k ε ∥ τ , ∣ x ∣ ∼ 2 j R ≤ C ( t R ) θ ∑ j = 1 ∞ 2 j ( δ − θ ) ≤ C , \sum_{j=1}^{\infty}2^{j\delta}(2^jR)^{n-\beta}\|k_\varepsilon(\cdot-h)-k_\varepsilon\|_{\tau,|x|\sim2^jR}\leq C\left(\frac{t}{R}\right)^\theta\sum_{j=1}^{\infty}2^{j(\delta-\theta)}\leq C, j = 1 ∑ ∞ 2 j δ ( 2 j R ) n − β ∥ k ε ( ⋅ − h ) − k ε ∥ τ , ∣ x ∣ ∼ 2 j R ≤ C ( R t ) θ j = 1 ∑ ∞ 2 j ( δ − θ ) ≤ C ,
so k ε ∈ H β , τ ( δ ) k_\varepsilon\in H_{\beta,\tau}^{(\delta)} k ε ∈ H β , τ ( δ ) , with c = 4 c=4 c = 4 .
We now show that k ε ∉ H β , τ ( δ ) k_\varepsilon\notin H_{\beta,\tau}^{(\delta)} k ε ∈ / H β , τ ( δ ) if δ ≥ θ \delta\geq\theta δ ≥ θ . Suppose that k ε ∈ H β , τ ( δ ) k_\varepsilon\in H_{\beta,\tau}^{(\delta)} k ε ∈ H β , τ ( δ ) with constants c c c and C C C . Let t > 0 t>0 t > 0 , h = t e n h=te_n h = t e n , R > max { c , 4 } t R>\max\{c,4\}t R > max { c , 4 } t , j ≥ 1 j\geq1 j ≥ 1 and s = 2 j R s=2^jR s = 2 j R . The set
E s = { x : 6 s / 5 < ∣ x ~ ∣ < 9 s / 5 , ∣ x n ∣ < t / 8 } E_s=\{x:6s/5<|\widetilde{x}|<9s/5,\ |x_n|<t/8\} E s = { x : 6 s /5 < ∣ x ∣ < 9 s /5 , ∣ x n ∣ < t /8 }
is contained in { s < ∣ x ∣ ≤ 2 s } \{s<|x|\leq2s\} { s < ∣ x ∣ ≤ 2 s } and, since n ≥ 2 n\geq2 n ≥ 2 , ∣ E s ∣ ≥ c n s n − 1 t |E_s|\geq c_ns^{n-1}t ∣ E s ∣ ≥ c n s n − 1 t . For x ∈ E s x\in E_s x ∈ E s we have ∣ x n ∣ < t / 8 |x_n|<t/8 ∣ x n ∣ < t /8 and 7 t / 8 < ∣ x n − t ∣ < 9 t / 8 7t/8<|x_n-t|<9t/8 7 t /8 < ∣ x n − t ∣ < 9 t /8 , hence, as ε ≠ 0 \varepsilon\neq0 ε = 0 and ∣ ε ∣ < 1 |\varepsilon|<1 ∣ ε ∣ < 1 ,
∣ g ( x n − t ) − g ( x n ) ∣ ≥ c ε t ε , g ( x n − t ) ≤ 2 t ε , c ε = ∣ ( 7 8 ) ε − ( 1 8 ) ε ∣ > 0. |g(x_n-t)-g(x_n)|\geq c_\varepsilon t^\varepsilon,\qquad g(x_n-t)\leq2t^\varepsilon,\qquad c_\varepsilon=\left|\left(\frac{7}{8}\right)^\varepsilon-\left(\frac{1}{8}\right)^\varepsilon\right|>0. ∣ g ( x n − t ) − g ( x n ) ∣ ≥ c ε t ε , g ( x n − t ) ≤ 2 t ε , c ε = ( 8 7 ) ε − ( 8 1 ) ε > 0.
Together with ∣ ϱ ( x − h ) − ϱ ( x ) ∣ ≤ C t s ϱ ( x ) |\varrho(x-h)-\varrho(x)|\leq C\frac{t}{s}\varrho(x) ∣ ϱ ( x − h ) − ϱ ( x ) ∣ ≤ C s t ϱ ( x ) , this gives, for x ∈ E s x\in E_s x ∈ E s ,
∣ k ε ( x − h ) − k ε ( x ) ∣ ≥ t ε ϱ ( x ) ( c ε − C t s ) ≥ c ε 2 t ε ϱ ( x ) ≥ c t ε s β − n − ε |k_\varepsilon(x-h)-k_\varepsilon(x)|\geq t^\varepsilon\varrho(x)\left(c_\varepsilon-C\frac{t}{s}\right)\geq\frac{c_\varepsilon}{2}t^\varepsilon\varrho(x)\geq ct^\varepsilon s^{\beta-n-\varepsilon} ∣ k ε ( x − h ) − k ε ( x ) ∣ ≥ t ε ϱ ( x ) ( c ε − C s t ) ≥ 2 c ε t ε ϱ ( x ) ≥ c t ε s β − n − ε
for every j ≥ j 0 j\geq j_0 j ≥ j 0 , where j 0 j_0 j 0 depends only on n n n , β \beta β and ε \varepsilon ε , since t / s < 2 − j t/s<2^{-j} t / s < 2 − j . As ∣ E s ∣ / ∣ B ( 0 , 2 s ) ∣ ≥ c t / s |E_s|/|B(0,2s)|\geq ct/s ∣ E s ∣/∣ B ( 0 , 2 s ) ∣ ≥ c t / s , we obtain, for j ≥ j 0 j\geq j_0 j ≥ j 0 ,
2 j δ ( 2 j R ) n − β ∥ k ε ( ⋅ − h ) − k ε ∥ τ , ∣ x ∣ ∼ 2 j R ≥ c 2 j δ ( t 2 j R ) θ = c ( t R ) θ 2 j ( δ − θ ) ≥ c ( t R ) θ , 2^{j\delta}(2^jR)^{n-\beta}\|k_\varepsilon(\cdot-h)-k_\varepsilon\|_{\tau,|x|\sim2^jR}\geq c2^{j\delta}\left(\frac{t}{2^jR}\right)^\theta=c\left(\frac{t}{R}\right)^\theta2^{j(\delta-\theta)}\geq c\left(\frac{t}{R}\right)^\theta, 2 j δ ( 2 j R ) n − β ∥ k ε ( ⋅ − h ) − k ε ∥ τ , ∣ x ∣ ∼ 2 j R ≥ c 2 j δ ( 2 j R t ) θ = c ( R t ) θ 2 j ( δ − θ ) ≥ c ( R t ) θ ,
so the series in Definition 3.2 diverges, a contradiction. This proves (a).
(b) Choose θ \theta θ with δ 1 < θ < min { δ 2 , 1 } \delta_1<\theta<\min\{\delta_2,1\} δ 1 < θ < min { δ 2 , 1 } and θ ≠ 1 τ \theta\neq\frac{1}{\tau} θ = τ 1 , and let ε = θ − 1 τ \varepsilon=\theta-\frac{1}{\tau} ε = θ − τ 1 . Then − 1 τ < ε < 1 − 1 τ -\frac{1}{\tau}<\varepsilon<1-\frac{1}{\tau} − τ 1 < ε < 1 − τ 1 and ε ≠ 0 \varepsilon\neq0 ε = 0 , so, by (a), k ε k_\varepsilon k ε belongs to S β , τ ∩ H β , τ ( δ 1 ) S_{\beta,\tau}\cap H_{\beta,\tau}^{(\delta_1)} S β , τ ∩ H β , τ ( δ 1 ) but not to H β , τ ( δ 2 ) H_{\beta,\tau}^{(\delta_2)} H β , τ ( δ 2 ) , since δ 2 > θ \delta_2>\theta δ 2 > θ . □
Sufficient conditions for (H2)
The following results show that (H2) is satisfied by large classes of kernels, homogeneous or not. In all of them the restriction δ < 1 \delta<1 δ < 1 of Theorem 1.1 plays a role.
Proposition 3.8 (Pointwise smoothness). Let 0 ≤ β < n 0\leq\beta<n 0 ≤ β < n , 0 < ε ≤ 1 0<\varepsilon\leq1 0 < ε ≤ 1 , and let k k k satisfy, for some c ≥ 1 c\geq1 c ≥ 1 ,
∣ k ( x − h ) − k ( x ) ∣ ≤ C ∣ h ∣ ε ∣ x ∣ n − β + ε , ∣ x ∣ > c ∣ h ∣ . |k(x-h)-k(x)|\leq C\frac{|h|^\varepsilon}{|x|^{n-\beta+\varepsilon}},\qquad|x|>c|h|. ∣ k ( x − h ) − k ( x ) ∣ ≤ C ∣ x ∣ n − β + ε ∣ h ∣ ε , ∣ x ∣ > c ∣ h ∣.
Then k ∈ H β , ∞ ( δ ) k\in H_{\beta,\infty}^{(\delta)} k ∈ H β , ∞ ( δ ) for every 0 ≤ δ < ε 0\leq\delta<\varepsilon 0 ≤ δ < ε . In particular, if ε = 1 \varepsilon=1 ε = 1 (the condition H β , ∞ ∗ H_{\beta,\infty}^* H β , ∞ ∗ of [1 Definition 3.3] ), then k ∈ H β , ∞ ( δ ) k\in H_{\beta,\infty}^{(\delta)} k ∈ H β , ∞ ( δ ) for every 0 ≤ δ < 1 0\leq\delta<1 0 ≤ δ < 1 .
Proof. Let R > c ∣ h ∣ R>c|h| R > c ∣ h ∣ and j ≥ 1 j\geq1 j ≥ 1 . If ∣ x ∣ > 2 j R |x|>2^jR ∣ x ∣ > 2 j R , then ∣ x ∣ > c ∣ h ∣ |x|>c|h| ∣ x ∣ > c ∣ h ∣ , and hence ∥ k ( ⋅ − h ) − k ∥ ∞ , ∣ x ∣ ∼ 2 j R ≤ C ∣ h ∣ ε ( 2 j R ) β − n − ε \|k(\cdot-h)-k\|_{\infty,|x|\sim2^jR}\leq C|h|^\varepsilon(2^jR)^{\beta-n-\varepsilon} ∥ k ( ⋅ − h ) − k ∥ ∞ , ∣ x ∣ ∼ 2 j R ≤ C ∣ h ∣ ε ( 2 j R ) β − n − ε . Therefore
∑ j = 1 ∞ 2 j δ ( 2 j R ) n − β ∥ k ( ⋅ − h ) − k ∥ ∞ , ∣ x ∣ ∼ 2 j R ≤ C ( ∣ h ∣ R ) ε ∑ j = 1 ∞ 2 j ( δ − ε ) ≤ C , \sum_{j=1}^{\infty}2^{j\delta}(2^jR)^{n-\beta}\|k(\cdot-h)-k\|_{\infty,|x|\sim2^jR}\leq C\left(\frac{|h|}{R}\right)^\varepsilon\sum_{j=1}^{\infty}2^{j(\delta-\varepsilon)}\leq C, j = 1 ∑ ∞ 2 j δ ( 2 j R ) n − β ∥ k ( ⋅ − h ) − k ∥ ∞ , ∣ x ∣ ∼ 2 j R ≤ C ( R ∣ h ∣ ) ε j = 1 ∑ ∞ 2 j ( δ − ε ) ≤ C ,
since ∣ h ∣ < R |h| < R ∣ h ∣ < R and δ < ε \delta< \varepsilon δ < ε .
The next result is the analogue of [1 Proposition 4.1] for the classes H β , τ ( δ ) H_{\beta,\tau}^{(\delta)} H β , τ ( δ ) . It allows us to produce fractional kernels from kernels of singular integral type.
Proposition 3.9. Let 1 ≤ τ ≤ ∞ 1 \le\tau\le\infty 1 ≤ τ ≤ ∞ , 0 < β < n 0 < \beta< n 0 < β < n and 0 ≤ δ < 1 0 \le\delta< 1 0 ≤ δ < 1 . If K ∈ S 0 , τ ∩ H 0 , τ ( δ ) K \in S_{0,\tau} \cap H_{0,\tau}^{(\delta)} K ∈ S 0 , τ ∩ H 0 , τ ( δ ) , then k ( x ) = ∣ x ∣ β K ( x ) k(x) = |x|^\beta K(x) k ( x ) = ∣ x ∣ β K ( x ) belongs to S β , τ ∩ H β , τ ( δ ) S_{\beta,\tau} \cap H_{\beta,\tau}^{(\delta)} S β , τ ∩ H β , τ ( δ ) .
Proof. Since ∣ x ∣ β ≤ ( 2 s ) β |x|^\beta\le(2s)^\beta ∣ x ∣ β ≤ ( 2 s ) β on { ∣ x ∣ ∼ s } \{|x| \sim s\} { ∣ x ∣ ∼ s } , we have ∥ k ∥ τ , ∣ x ∣ ∼ s ≤ C s β ∥ K ∥ τ , ∣ x ∣ ∼ s ≤ C s β − n \|k\|_{\tau,|x|\sim s} \le Cs^\beta\|K\|_{\tau,|x|\sim s} \le Cs^{\beta-n} ∥ k ∥ τ , ∣ x ∣ ∼ s ≤ C s β ∥ K ∥ τ , ∣ x ∣ ∼ s ≤ C s β − n , so k ∈ S β , τ k \in S_{\beta,\tau} k ∈ S β , τ .
Let c ≥ 1 c \ge1 c ≥ 1 be the constant of K K K , R > c ∣ h ∣ R > c|h| R > c ∣ h ∣ , j ≥ 1 j \ge1 j ≥ 1 and s = 2 j R s = 2^jR s = 2 j R . If ∣ x ∣ ∼ s |x| \sim s ∣ x ∣ ∼ s , then ∣ h ∣ < R ≤ s / 2 |h| < R \le s/2 ∣ h ∣ < R ≤ s /2 , so s / 2 < ∣ x − h ∣ < 5 s / 2 s/2 < |x-h| < 5s/2 s /2 < ∣ x − h ∣ < 5 s /2 and, by the mean value theorem,
∣ ∣ x − h ∣ β − ∣ x ∣ β ∣ ≤ C s β − 1 ∣ h ∣ . \left||x-h|^\beta-|x|^\beta\right| \le Cs^{\beta-1}|h|. ∣ x − h ∣ β − ∣ x ∣ β ≤ C s β − 1 ∣ h ∣.
Hence
∣ k ( x − h ) − k ( x ) ∣ ≤ ∣ x − h ∣ β ∣ K ( x − h ) − K ( x ) ∣ + ∣ K ( x ) ∣ ∣ ∣ x − h ∣ β − ∣ x ∣ β ∣ ≤ C s β ∣ K ( x − h ) − K ( x ) ∣ + C s β ∣ h ∣ s ∣ K ( x ) ∣ , \begin{aligned}
|k(x-h)-k(x)| \le|x-h|^\beta|K(x-h)-K(x)|+|K(x)|\left||x-h|^\beta-|x|^\beta\right| \\
&\le Cs^\beta|K(x-h)-K(x)|+Cs^\beta\frac{|h|}{s}|K(x)|,
\end{aligned} ∣ k ( x − h ) − k ( x ) ∣ ≤ ∣ x − h ∣ β ∣ K ( x − h ) − K ( x ) ∣ + ∣ K ( x ) ∣ ∣ x − h ∣ β − ∣ x ∣ β ≤ C s β ∣ K ( x − h ) − K ( x ) ∣ + C s β s ∣ h ∣ ∣ K ( x ) ∣ ,
and, using K ∈ S 0 , τ K \in S_{0,\tau} K ∈ S 0 , τ ,
∥ k ( ⋅ − h ) − k ∥ τ , ∣ x ∣ ∼ s ≤ C s β ∥ K ( ⋅ − h ) − K ∥ τ , ∣ x ∣ ∼ s + C s β − n ∣ h ∣ s . \|k(\cdot-h)-k\|_{\tau,|x|\sim s}\le Cs^\beta\|K(\cdot-h)-K\|_{\tau,|x|\sim s}+Cs^{\beta-n}\frac{|h|}{s}. ∥ k ( ⋅ − h ) − k ∥ τ , ∣ x ∣ ∼ s ≤ C s β ∥ K ( ⋅ − h ) − K ∥ τ , ∣ x ∣ ∼ s + C s β − n s ∣ h ∣ .
Multiplying by 2 j δ s n − β 2^{j\delta}s^{n-\beta} 2 j δ s n − β and summing over j ≥ 1 j \ge1 j ≥ 1 ,
∑ j = 1 ∞ 2 j δ ( 2 j R ) n − β ∥ k ( ⋅ − h ) − k ∥ τ , ∣ x ∣ ∼ 2 j R ≤ C ∑ j = 1 ∞ 2 j δ ( 2 j R ) n ∥ K ( ⋅ − h ) − K ∥ τ , ∣ x ∣ ∼ 2 j R + C ∣ h ∣ R ∑ j = 1 ∞ 2 j ( δ − 1 ) ≤ C , \begin{aligned}
\sum_{j=1}^{\infty}2^{j\delta}(2^jR)^{n-\beta}\|k(\cdot-h)-k\|_{\tau,|x|\sim2^jR} \\
&\le C\sum_{j=1}^{\infty}2^{j\delta}(2^jR)^n\|K(\cdot-h)-K\|_{\tau,|x|\sim2^jR}
+C\frac{|h|}{R}\sum_{j=1}^{\infty}2^{j(\delta-1)}\le C,
\end{aligned} j = 1 ∑ ∞ 2 j δ ( 2 j R ) n − β ∥ k ( ⋅ − h ) − k ∥ τ , ∣ x ∣ ∼ 2 j R ≤ C j = 1 ∑ ∞ 2 j δ ( 2 j R ) n ∥ K ( ⋅ − h ) − K ∥ τ , ∣ x ∣ ∼ 2 j R + C R ∣ h ∣ j = 1 ∑ ∞ 2 j ( δ − 1 ) ≤ C ,
since K ∈ H 0 , τ ( δ ) K \in H_{0,\tau}^{(\delta)} K ∈ H 0 , τ ( δ ) , ∣ h ∣ < R |h| < R ∣ h ∣ < R and δ < 1 \delta< 1 δ < 1 .
For rough homogeneous kernels, ( H 2 ) (H_2) ( H 2 ) follows from a Dini type condition of order δ \delta δ . For Ω ∈ L τ ( S n − 1 ) \Omega\in L^\tau(S^{n-1}) Ω ∈ L τ ( S n − 1 ) let
ϖ τ ( u ) = sup ∣ ϱ ∣ ≤ u ∥ Ω ( ϱ ⋅ ) − Ω ∥ L τ ( S n − 1 ) , \varpi_\tau(u)=\sup_{|\varrho|\le u}\|\Omega(\varrho\cdot)-\Omega\|_{L^\tau(S^{n-1})}, ϖ τ ( u ) = ∣ ϱ ∣ ≤ u sup ∥Ω ( ϱ ⋅ ) − Ω ∥ L τ ( S n − 1 ) ,
where the supremum is taken over rotations ϱ \varrho ϱ with ∣ ϱ ∣ = ∥ ϱ − I ∥ ≤ u |\varrho|=\|\varrho-I\|\le u ∣ ϱ ∣ = ∥ ϱ − I ∥ ≤ u .
Proposition 3.10. Let 1 < τ < ∞ 1 < \tau< \infty 1 < τ < ∞ , 0 < β < n 0 < \beta< n 0 < β < n , 0 ≤ δ < 1 0 \le\delta< 1 0 ≤ δ < 1 , and let Ω ∈ L τ ( S n − 1 ) \Omega\in L^\tau(S^{n-1}) Ω ∈ L τ ( S n − 1 ) be homogeneous of degree zero with
∫ 0 1 ϖ τ ( u ) u 1 + δ d u < ∞ . (6) \int_0^1\frac{\varpi_\tau(u)}{u^{1+\delta}}\,du<\infty. \tag*{(6)} ∫ 0 1 u 1 + δ ϖ τ ( u ) d u < ∞. (6) Then k ( x ) = Ω ( x ) ∣ x ∣ β − n k(x)=\Omega(x)|x|^{\beta-n} k ( x ) = Ω ( x ) ∣ x ∣ β − n belongs to S β , τ ∩ H β , τ ( δ ) S_{\beta,\tau}\cap H_{\beta,\tau}^{(\delta)} S β , τ ∩ H β , τ ( δ ) .
Proof. That k ∈ S β , τ k \in S_{\beta,\tau} k ∈ S β , τ follows as in Example 3.4. For the smoothness we use [6 Lemma 4] (whose proof follows [17 Lemma 5] ): since (3.3) implies the L τ L^\tau L τ -Dini condition
∫ 0 1 ϖ τ ( u ) u d u < ∞ , \int_0^1\frac{\varpi_\tau(u)}{u}\,du<\infty, ∫ 0 1 u ϖ τ ( u ) d u < ∞ ,
and after dividing by ∣ B ( 0 , 2 s ) ∣ 1 / τ |B(0,2s)|^{1/\tau} ∣ B ( 0 , 2 s ) ∣ 1/ τ , there is C > 0 C>0 C > 0 such that, for s > 4 ∣ h ∣ s>4|h| s > 4∣ h ∣ ,
∥ k ( ⋅ − h ) − k ∥ τ , ∣ x ∣ ∼ s ≤ C s β − n ( ∣ h ∣ s + ∫ ∣ h ∣ / 2 s ∣ h ∣ / s ϖ τ ( u ) d u u ) . (7) \|k(\cdot-h)-k\|_{\tau,|x|\sim s}\le Cs^{\beta-n}\left(\frac{|h|}{s}+\int_{|h|/2s}^{|h|/s}\varpi_\tau(u)\frac{du}{u}\right). \tag*{(7)} ∥ k ( ⋅ − h ) − k ∥ τ , ∣ x ∣ ∼ s ≤ C s β − n ( s ∣ h ∣ + ∫ ∣ h ∣/2 s ∣ h ∣/ s ϖ τ ( u ) u d u ) . (7)
Let R > 2 ∣ h ∣ R > 2|h| R > 2∣ h ∣ , j ≥ 1 j \ge1 j ≥ 1 and s = 2 j R s = 2^jR s = 2 j R , so that s > 4 ∣ h ∣ s > 4|h| s > 4∣ h ∣ . If u ∈ [ ∣ h ∣ / 2 j + 1 R , ∣ h ∣ / 2 j R ] u \in[|h|/2^{j+1}R, |h|/2^jR] u ∈ [ ∣ h ∣/ 2 j + 1 R , ∣ h ∣/ 2 j R ] , then 2 j ≤ ∣ h ∣ / ( R u ) ≤ u − 1 2^j \le|h|/(Ru) \le u^{-1} 2 j ≤ ∣ h ∣/ ( R u ) ≤ u − 1 , so 2 j δ ≤ u − δ 2^{j\delta} \le u^{-\delta} 2 j δ ≤ u − δ . Hence, by (7) ,
∑ j = 1 ∞ 2 j δ ( 2 j R ) n − β ∥ k ( ⋅ − h ) − k ∥ τ , ∣ x ∣ ∼ 2 j R ≤ C ∣ h ∣ R ∑ j = 1 ∞ 2 j ( δ − 1 ) + C ∫ 0 1 ϖ τ ( u ) u 1 + δ d u ≤ C . \sum_{j=1}^{\infty} 2^{j\delta}(2^jR)^{n-\beta}\lVert k(\cdot-h)-k\rVert_{\tau,|x|\sim2^jR}
\le C\frac{|h|}{R}\sum_{j=1}^{\infty}2^{j(\delta-1)}+C\int_0^1\frac{\varpi_\tau(u)}{u^{1+\delta}}\,du\le C. j = 1 ∑ ∞ 2 j δ ( 2 j R ) n − β ∥ k ( ⋅ − h ) − k ∥ τ , ∣ x ∣ ∼ 2 j R ≤ C R ∣ h ∣ j = 1 ∑ ∞ 2 j ( δ − 1 ) + C ∫ 0 1 u 1 + δ ϖ τ ( u ) d u ≤ C .
For δ = 0 \delta=0 δ = 0 , Proposition 3.10 is essentially [1 Proposition 4.2] , with k = 0 k=0 k = 0 there and the Young function t ↦ t t t\mapsto t^t t ↦ t t . For 0 ≤ δ < 1 0\le\delta<1 0 ≤ δ < 1 , the conclusion of Proposition 3.10 is stated without proof in [15 Remark 5.7] .
Remark 3.11. For m = 1 m=1 m = 1 , Ding and Lu [6, Theorem 2] proved the (unweighted) H p − L q H^p-L^q H p − L q boundedness of T Ω , α T_{\Omega,\alpha} T Ω , α , for 0 < α < 1 0<\alpha<1 0 < α < 1 and n n + α ≤ p < 1 \frac{n}{n+\alpha}\le p<1 n + α n ≤ p < 1 , for Ω ∈ L r ( S n − 1 ) \Omega\in L^r(S^{n-1}) Ω ∈ L r ( S n − 1 ) with r > n n − α r>\frac{n}{n-\alpha} r > n − α n satisfying the Dini condition ∫ 0 1 ϖ r ( u ) u − 1 − α d u < ∞ \int_0^1\varpi_r(u)u^{-1-\alpha}\,du<\infty ∫ 0 1 ϖ r ( u ) u − 1 − α d u < ∞ . Since δ = n ( 1 p − 1 ) ≤ α \delta=n(\frac{1}{p}-1)\le\alpha δ = n ( p 1 − 1 ) ≤ α in that range, condition (6) , of order δ \delta δ , is weaker than theirs; in fact, their argument (see [6 (3.6)] ) only uses the Dini condition of order δ \delta δ . Thus, for m = 1 m=1 m = 1 , Theorem 1.1 and Proposition 3.10 give a weighted version of [6 Theorem 2] for n n + α < p < 1 \frac{n}{n+\alpha}<p<1 n + α n < p < 1 , under the Dini condition of order δ \delta δ .
Remark 3.12. If Ω \Omega Ω is Hölder continuous of order ε \varepsilon ε on S n − 1 S^{n-1} S n − 1 , then ϖ τ ( u ) ≤ C u ε \varpi_\tau(u)\le Cu^\varepsilon ϖ τ ( u ) ≤ C u ε for every τ \tau τ , and (6) holds whenever ε > δ \varepsilon>\delta ε > δ ; in this case Proposition 3.8 also gives k ∈ H β , ∞ ( δ ) k\in H^{(\delta)}_{\beta,\infty} k ∈ H β , ∞ ( δ ) . Compare with Remark 4.7, where Ω 1 \Omega_1 Ω 1 is Hölder of order ε ≤ δ − 1 / q \varepsilon\le\delta-1/q ε ≤ δ − 1/ q and the conclusion of Theorem 1.1 fails.
Example 3.13 (A non-homogeneous kernel). Let Ω ∈ C 1 ( S n − 1 ) \Omega\in C^1(S^{n-1}) Ω ∈ C 1 ( S n − 1 ) , extended homogeneously of degree zero, and
k i ( x ) = Ω ( x ) cos ( log ∣ x ∣ ) ∣ x ∣ n / q i . k_i(x)=\frac{\Omega(x)\cos(\log|x|)}{|x|^{n/q_i}}. k i ( x ) = ∣ x ∣ n / q i Ω ( x ) cos ( log ∣ x ∣ ) .
Then ∣ k i ( x ) ∣ ≤ C ∣ x ∣ − n / q i |k_i(x)|\le C|x|^{-n/q_i} ∣ k i ( x ) ∣ ≤ C ∣ x ∣ − n / q i and ∣ ∇ k i ( x ) ∣ ≤ C ∣ x ∣ − n / q i − 1 |\nabla k_i(x)|\le C|x|^{-n/q_i-1} ∣∇ k i ( x ) ∣ ≤ C ∣ x ∣ − n / q i − 1 for x ≠ 0 x\ne0 x = 0 . If ∣ x ∣ > 2 ∣ h ∣ |x|>2|h| ∣ x ∣ > 2∣ h ∣ , the segment from x − h x-h x − h to x x x stays in { ∣ z ∣ > ∣ x ∣ / 2 } \{|z|>|x|/2\} { ∣ z ∣ > ∣ x ∣/2 } , so ∣ k i ( x − h ) − k i ( x ) ∣ ≤ C ∣ h ∣ ∣ x ∣ − n / q i − 1 |k_i(x-h)-k_i(x)|\le C|h||x|^{-n/q_i-1} ∣ k i ( x − h ) − k i ( x ) ∣ ≤ C ∣ h ∣∣ x ∣ − n / q i − 1 . Hence k i ∈ S n − n / q i , ∞ k_i\in S_{n-n/q_i,\infty} k i ∈ S n − n / q i , ∞ and, by Proposition 3.8, k i ∈ H n − n / q i , ∞ ( δ ) k_i\in H^{(\delta)}_{n-n/q_i,\infty} k i ∈ H n − n / q i , ∞ ( δ ) for every 0 ≤ δ < 1 0\le\delta<1 0 ≤ δ < 1 . Thus (H1) and (H2) hold with p i = ∞ p_i=\infty p i = ∞ , although k i k_i k i is not homogeneous. More generally, by Proposition 3.9, one may take k i ( x ) = ∣ x ∣ n − n / q i K i ( x ) k_i(x)=|x|^{n-n/q_i}K_i(x) k i ( x ) = ∣ x ∣ n − n / q i K i ( x ) with K i ∈ S 0 , ∞ ∩ H 0 , ∞ ( δ ) K_i\in S_{0,\infty}\cap H^{(\delta)}_{0,\infty} K i ∈ S 0 , ∞ ∩ H 0 , ∞ ( δ ) , for instance a Calderón–Zygmund kernel with Hölder regularity of order ε > δ \varepsilon>\delta ε > δ . Other non-homogeneous kernels satisfying size and Hörmander conditions of Young type are given, in dimension one, in [16 Section 4] .
The homogeneous case
In this subsection k i ( x ) = Ω i ( x ) ∣ x ∣ − n / q i k_i(x)=\Omega_i(x)|x|^{-n/q_i} k i ( x ) = Ω i ( x ) ∣ x ∣ − n / q i as in Example 3.4. Definition 3.2 makes sense for every β ∈ R \beta\in\mathbb{R} β ∈ R , and here we allow any real β \beta β . For β , δ ∈ R \beta,\delta\in\mathbb{R} β , δ ∈ R , h ∈ R n h\in\mathbb{R}^n h ∈ R n and R > 0 R>0 R > 0 set
Λ β ( δ ) ( h , R ) = ∑ j = 1 ∞ 2 j δ ( 2 j R ) n − β ∥ k i ( ⋅ − h ) − k i ∥ p i , ∣ x ∣ ∼ 2 j R , Λ β = Λ β ( 0 ) . \Lambda_\beta^{(\delta)}(h,R)=\sum_{j=1}^{\infty}2^{j\delta}(2^jR)^{n-\beta}\lVert k_i(\cdot-h)-k_i\rVert_{p_i,|x|\sim2^jR},\qquad\Lambda_\beta=\Lambda_\beta^{(0)}. Λ β ( δ ) ( h , R ) = j = 1 ∑ ∞ 2 j δ ( 2 j R ) n − β ∥ k i ( ⋅ − h ) − k i ∥ p i , ∣ x ∣ ∼ 2 j R , Λ β = Λ β ( 0 ) .
The next result shows that, among classical fractional Hörmander classes, the only one compatible with the homogeneity of k i k_i k i is H n − n / q i , p i H_{n-n/q_i,p_i} H n − n / q i , p i . Consequently, the extra decay required in (H2) cannot be obtained by changing the index β \beta β , and has to be imposed through the factor 2 j δ 2^{j\delta} 2 j δ .
Theorem 3.14. Let 1 ≤ i ≤ m 1 \le i \le m 1 ≤ i ≤ m , 1 ≤ p i ≤ ∞ 1 \le p_i \le\infty 1 ≤ p i ≤ ∞ , and k i ( x ) = Ω i ( x ) ∣ x ∣ − n / q i k_i(x) = \Omega_i(x)|x|^{-n/q_i} k i ( x ) = Ω i ( x ) ∣ x ∣ − n / q i with Ω i ∈ L 1 ( S n − 1 ) \Omega_i \in L^1(S^{n-1}) Ω i ∈ L 1 ( S n − 1 ) homogeneous of degree zero and k i ≢ 0 k_i \not\equiv0 k i ≡ 0 .
(a) If k i ∈ H β , p i k_i \in H_{\beta,p_i} k i ∈ H β , p i for some β ∈ R \beta\in\mathbb{R} β ∈ R , then β = n − n / q i \beta= n - n/q_i β = n − n / q i .
(b) In particular, k i ∉ H n − n / q i − δ , p i k_i \notin H_{n-n/q_i-\delta,p_i} k i ∈ / H n − n / q i − δ , p i for every δ ≠ 0 \delta\ne0 δ = 0 .
(c) For every δ ∈ R \delta\in\mathbb{R} δ ∈ R , Λ n − n / q i ( δ ) ( λ h , λ R ) = Λ n − n / q i ( δ ) ( h , R ) \Lambda_{n-n/q_i}^{(\delta)}(\lambda h,\lambda R) = \Lambda_{n-n/q_i}^{(\delta)}(h,R) Λ n − n / q i ( δ ) ( λh , λ R ) = Λ n − n / q i ( δ ) ( h , R ) for all λ > 0 \lambda> 0 λ > 0 .
Proof. The proof relies on the behavior of the sums Λ β ( δ ) \Lambda_\beta^{(\delta)} Λ β ( δ ) under dilations, which is a consequence of the homogeneity of k i k_i k i . For h ∈ R n h \in\mathbb{R}^n h ∈ R n , R > 0 R > 0 R > 0 and j ≥ 1 j \ge1 j ≥ 1 , let N j ( h , R ) = ∥ k i ( ⋅ − h ) − k i ∥ p i , ∣ x ∣ ∼ 2 j R N_j(h,R) = \|k_i(\cdot-h)-k_i\|_{p_i,|x|\sim2^jR} N j ( h , R ) = ∥ k i ( ⋅ − h ) − k i ∥ p i , ∣ x ∣ ∼ 2 j R . Since k i k_i k i is homogeneous of degree − n / q i -n/q_i − n / q i , we have k i ( λ x − λ h ) − k i ( λ x ) = λ − n / q i ( k i ( x − h ) − k i ( x ) ) k_i(\lambda x-\lambda h)-k_i(\lambda x)=\lambda^{-n/q_i}(k_i(x-h)-k_i(x)) k i ( λ x − λh ) − k i ( λ x ) = λ − n / q i ( k i ( x − h ) − k i ( x )) for a.e. x x x and every λ > 0 \lambda> 0 λ > 0 . If p i < ∞ p_i < \infty p i < ∞ , the change of variables x = λ y x=\lambda y x = λ y and the identity ∣ B ( 0 , 2 j + 1 λ R ) ∣ = λ n ∣ B ( 0 , 2 j + 1 R ) ∣ |B(0,2^{j+1}\lambda R)|=\lambda^n|B(0,2^{j+1}R)| ∣ B ( 0 , 2 j + 1 λ R ) ∣ = λ n ∣ B ( 0 , 2 j + 1 R ) ∣ give
N j ( λ h , λ R ) = ( λ n λ − n p i / q i λ n ∣ B ( 0 , 2 j + 1 R ) ∣ ∫ 2 j R < ∣ y ∣ ≤ 2 j + 1 R ∣ k i ( y − h ) − k i ( y ) ∣ p i d y ) 1 / p i = λ − n / q i N j ( h , R ) , \begin{aligned}
N_j(\lambda h,\lambda R)
&=\left(\frac{\lambda^n\lambda^{-np_i/q_i}}{\lambda^n|B(0,2^{j+1}R)|}\int_{2^jR<|y|\le2^{j+1}R}|k_i(y-h)-k_i(y)|^{p_i}\,dy\right)^{1/p_i}\\
&=\lambda^{-n/q_i}N_j(h,R),
\end{aligned} N j ( λh , λ R ) = ( λ n ∣ B ( 0 , 2 j + 1 R ) ∣ λ n λ − n p i / q i ∫ 2 j R < ∣ y ∣ ≤ 2 j + 1 R ∣ k i ( y − h ) − k i ( y ) ∣ p i d y ) 1/ p i = λ − n / q i N j ( h , R ) ,
and the same identity holds for p i = ∞ p_i=\infty p i = ∞ , taking essential suprema. Multiplying by 2 j δ ( 2 j λ R ) n − β 2^{j\delta}(2^j\lambda R)^{n-\beta} 2 j δ ( 2 j λ R ) n − β and summing in j j j , we obtain
Λ β ( δ ) ( λ h , λ R ) = λ θ Λ β ( δ ) ( h , R ) , θ = n − β − n q i . (8) \Lambda_\beta^{(\delta)}(\lambda h,\lambda R)=\lambda^\theta\Lambda_\beta^{(\delta)}(h,R),\qquad\theta=n-\beta-\frac{n}{q_i}. \tag*{(8)} Λ β ( δ ) ( λh , λ R ) = λ θ Λ β ( δ ) ( h , R ) , θ = n − β − q i n . (8)
When β = n − n / q i \beta=n-n/q_i β = n − n / q i we have θ = 0 \theta=0 θ = 0 , and (8) is precisely the statement of (c).
We now prove (a). Assume that k i ∈ H β , p i k_i\in H_{\beta,p_i} k i ∈ H β , p i with constants c c c and C C C and suppose, by contradiction, that β ≠ n − n / q i \beta\ne n-n/q_i β = n − n / q i , that is, θ ≠ 0 \theta\ne0 θ = 0 . Fix h ∈ R n h\in\mathbb{R}^n h ∈ R n and R > c ∣ h ∣ R>c|h| R > c ∣ h ∣ . Since λ R > c ∣ λ h ∣ \lambda R>c|\lambda h| λ R > c ∣ λh ∣ for every λ > 0 \lambda>0 λ > 0 , the Hörmander condition can be applied to the pair ( λ h , λ R ) (\lambda h,\lambda R) ( λh , λ R ) , and (8) gives λ θ Λ β ( h , R ) ≤ C \lambda^\theta\Lambda_\beta(h,R)\le C λ θ Λ β ( h , R ) ≤ C for all λ > 0 \lambda>0 λ > 0 . Letting λ → ∞ \lambda\to\infty λ → ∞ if θ > 0 \theta>0 θ > 0 , or λ → 0 \lambda\to0 λ → 0 if θ < 0 \theta<0 θ < 0 , we conclude that Λ β ( h , R ) = 0 \Lambda_\beta(h,R)=0 Λ β ( h , R ) = 0 . Consequently N j ( h , R ) = 0 N_j(h,R)=0 N j ( h , R ) = 0 for every j ≥ 1 j\ge1 j ≥ 1 , which means that k i ( x − h ) = k i ( x ) k_i(x-h)=k_i(x) k i ( x − h ) = k i ( x ) for a.e. x x x with ∣ x ∣ > 2 R |x|>2R ∣ x ∣ > 2 R .
This holds for every R > c ∣ h ∣ R>c|h| R > c ∣ h ∣ , and we can let R R R decrease to c ∣ h ∣ c|h| c ∣ h ∣ . Indeed, let h ≠ 0 h\ne0 h = 0 (the case h = 0 h=0 h = 0 is trivial) and R l = c ∣ h ∣ ( 1 + 1 l ) R_l=c|h|(1+\frac{1}{l}) R l = c ∣ h ∣ ( 1 + l 1 ) . Since { ∣ x ∣ > 2 c ∣ h ∣ } = ⋃ l { ∣ x ∣ > 2 R l } \{|x|>2c|h|\}=\bigcup_l\{|x|>2R_l\} { ∣ x ∣ > 2 c ∣ h ∣ } = ⋃ l { ∣ x ∣ > 2 R l } and a countable union of null sets is null, we obtain k i ( x − h ) = k i ( x ) k_i(x-h)=k_i(x) k i ( x − h ) = k i ( x ) for a.e. x x x with ∣ x ∣ > 2 c ∣ h ∣ |x|>2c|h| ∣ x ∣ > 2 c ∣ h ∣ . In other words, for every r > 0 r>0 r > 0 the function k i k_i k i is invariant under all sufficiently small translations on the exterior region U r = { x : ∣ x ∣ > r } U_r=\{x:|x|>r\} U r = { x : ∣ x ∣ > r } :
k i ( ⋅ − h ) = k i a.e. in U r , whenever ∣ h ∣ < r 2 c . (9) k_i(\cdot-h)=k_i\quad\text{a.e. in }U_r,\qquad\text{whenever }|h|<\frac{r}{2c}. \tag*{(9)} k i ( ⋅ − h ) = k i a.e. in U r , whenever ∣ h ∣ < 2 c r . (9)
We next show that this invariance forces k i k_i k i to be locally constant on U r U_r U r . Since Ω i ∈ L 1 ( S n − 1 ) \Omega_i\in L^1(S^{n-1}) Ω i ∈ L 1 ( S n − 1 ) , we have k i ∈ L l o c 1 ( R n ∖ { 0 } ) k_i\in L^1_{\mathrm{loc}}(\mathbb{R}^n\setminus\{0\}) k i ∈ L loc 1 ( R n ∖ { 0 }) . Let φ ∈ C c ∞ ( U r ) \varphi\in C_c^\infty(U_r) φ ∈ C c ∞ ( U r ) and 1 ≤ l ≤ n 1\le l\le n 1 ≤ l ≤ n . By (9) , for ∣ t ∣ |t| ∣ t ∣ small enough,
∫ k i ( x ) φ ( x + t e l ) − φ ( x ) t d x = ∫ k i ( x − t e l ) − k i ( x ) t φ ( x ) d x = 0 , \int k_i(x)\frac{\varphi(x+t e_l)-\varphi(x)}{t}\,dx
=
\int\frac{k_i(x-t e_l)-k_i(x)}{t}\varphi(x)\,dx=0, ∫ k i ( x ) t φ ( x + t e l ) − φ ( x ) d x = ∫ t k i ( x − t e l ) − k i ( x ) φ ( x ) d x = 0 ,
and letting t → 0 t\to0 t → 0 , by dominated convergence, we get ∫ k i ∂ l φ = 0 \int k_i\,\partial_l\varphi=0 ∫ k i ∂ l φ = 0 . Thus ∇ k i = 0 \nabla k_i=0 ∇ k i = 0 in D ′ ( U r ) D'(U_r) D ′ ( U r ) , and therefore k i k_i k i coincides a.e. with a constant c V c_V c V on each connected component V V V of U r U_r U r .
To identify these components, observe that the map Φ ( x ) = ( x / ∣ x ∣ , ∣ x ∣ ) \Phi(x)=(x/|x|,|x|) Φ ( x ) = ( x /∣ x ∣ , ∣ x ∣ ) is a homeomorphism from U r U_r U r onto S n − 1 × ( r , ∞ ) S^{n-1}\times(r,\infty) S n − 1 × ( r , ∞ ) . Hence, if n ≥ 2 n\ge2 n ≥ 2 , U r U_r U r is connected, while if n = 1 n=1 n = 1 its connected components are ( r , ∞ ) (r,\infty) ( r , ∞ ) and ( − ∞ , − r ) (-\infty,-r) ( − ∞ , − r ) . In both cases, since the dilation x ↦ 2 x x \mapsto2x x ↦ 2 x corresponds under Φ \Phi Φ to ( θ , ρ ) ↦ ( θ , 2 ρ ) (\theta,\rho) \mapsto(\theta,2\rho) ( θ , ρ ) ↦ ( θ , 2 ρ ) , every connected component V V V of U r U_r U r satisfies 2 V ⊂ V 2V \subset V 2 V ⊂ V . Using the homogeneity of k i k_i k i once more, we obtain
c V = k i ( 2 x ) = 2 − n / q i k i ( x ) = 2 − n / q i c V c_V=k_i(2x)=2^{-n/q_i}k_i(x)=2^{-n/q_i}c_V c V = k i ( 2 x ) = 2 − n / q i k i ( x ) = 2 − n / q i c V
for a.e. x ∈ V x \in V x ∈ V , and since 2 − n / q i ≠ 1 2^{-n/q_i} \ne1 2 − n / q i = 1 , it follows that c V = 0 c_V=0 c V = 0 . Hence k i = 0 k_i=0 k i = 0 a.e. on U r U_r U r for every r > 0 r>0 r > 0 , that is, k i = 0 k_i=0 k i = 0 a.e. on R n \mathbb{R}^n R n , which contradicts k i ≠ 0 k_i \ne0 k i = 0 . This proves (a). Finally, if δ ≠ 0 \delta\ne0 δ = 0 , then n − n / q i − δ ≠ n − n / q i n-n/q_i-\delta\ne n-n/q_i n − n / q i − δ = n − n / q i , so (b) is an immediate consequence of (a). □
Remark 3.15. By Theorem 3.14 (c), ( H 2 ) (\mathrm{H}_2) ( H 2 ) is compatible with the homogeneity of k i k_i k i : since 2 j δ = ( 2 j R / R ) δ 2^{j\delta}=(2^jR/R)^\delta 2 j δ = ( 2 j R / R ) δ , the factor does not depend on the scale. Writing 2 j δ = ( 2 j R ) δ R − δ 2^{j\delta}=(2^jR)^\delta R^{-\delta} 2 j δ = ( 2 j R ) δ R − δ , condition ( H 2 ) (\mathrm{H}_2) ( H 2 ) reads
∑ j = 1 ∞ ( 2 j R ) n / q i + δ ∥ k i ( ⋅ − h ) − k i ∥ p i , ∣ x ∣ ∼ 2 j R ≤ C i R δ , R > c i ∣ h ∣ . \sum_{j=1}^{\infty}(2^jR)^{n/q_i+\delta}\left\|k_i(\cdot-h)-k_i\right\|_{p_i,\lvert x\rvert\sim2^jR}\le C_iR^\delta,\qquad R>c_i\lvert h\rvert. j = 1 ∑ ∞ ( 2 j R ) n / q i + δ ∥ k i ( ⋅ − h ) − k i ∥ p i , ∣ x ∣ ∼ 2 j R ≤ C i R δ , R > c i ∣ h ∣ .
The factor R δ R^\delta R δ on the right-hand side is exactly what a classical Hörmander condition does not allow and, by Theorem 3.14 (b), it cannot be removed by changing the index. On the other hand, Remark 4.7 shows that the classical condition H n − n / q i , p i H_{n-n/q_i,p_i} H n − n / q i , p i is not sufficient for our results, for any 1 ≤ p i ≤ ∞ 1\le p_i\le\infty 1 ≤ p i ≤ ∞ .
Proof of the main result
In this section we prove Theorem 1.1 . We first state an extension theorem and prove the technical lemma, which gives the uniform boundedness of T α T_\alpha T α on atoms; then we obtain the unweighted estimate needed to apply them and prove the theorem. Finally, we show that ( H 2 ) (\mathrm{H}_2) ( H 2 ) cannot be replaced by the classical Hörmander condition.
An extension theorem
We state the extension result with two weights, which is the form needed for Theorem 1.1 . It extends [22 Theorem 2.10] , which corresponds to q 0 = p 0 q_0=p_0 q 0 = p 0 , u = v u=v u = v and q = p q=p q = p .
Theorem 4.1. Let T T T be a bounded linear operator from L p 0 ( R n ) L^{p_0}(\mathbb{R}^n) L p 0 ( R n ) into L q 0 ( R n ) L^{q_0}(\mathbb{R}^n) L q 0 ( R n ) for some 1 < p 0 ≤ q 0 < ∞ 1<p_0\le q_0<\infty 1 < p 0 ≤ q 0 < ∞ . Let v ∈ A ∞ v\in A_\infty v ∈ A ∞ , 0 < p ≤ 1 0<p\le1 0 < p ≤ 1 with r v r v − 1 p < p 0 \frac{r_v}{r_v-1}p<p_0 r v − 1 r v p < p 0 , d ≥ ⌊ n ( q 0 p − 1 ) ⌋ d\ge\lfloor n(\frac{q_0}{p}-1)\rfloor d ≥ ⌊ n ( p q 0 − 1 )⌋ , u u u a weight, and p ≤ q < ∞ p\le q<\infty p ≤ q < ∞ . Then T T T extends to a bounded operator from H v p ( R n ) H_v^p(\mathbb{R}^n) H v p ( R n ) into L u q ( R n ) L_u^q(\mathbb{R}^n) L u q ( R n ) , that is, there is C > 0 C>0 C > 0 such that ∥ T f ∥ L u q ≤ C ∥ f ∥ H v p \|Tf\|_{L_u^q}\le C\|f\|_{H_v^p} ∥ T f ∥ L u q ≤ C ∥ f ∥ H v p for every f ∈ D 0 f\in D_0 f ∈ D 0 , if and only if
sup { ∥ T a ∥ L u q : a is a v - ( p , p 0 , d ) atom } < ∞ . \sup\{\|Ta\|_{L_u^q}:a\text{ is a }v\text{-}(p,p_0,d)\text{ atom}\}<\infty. sup { ∥ T a ∥ L u q : a is a v - ( p , p 0 , d ) atom } < ∞.
Proof. Assume first that ∥ T f ∥ L u q ≤ C 1 ∥ f ∥ H v p \|Tf\|_{L_u^q}\le C_1\|f\|_{H_v^p} ∥ T f ∥ L u q ≤ C 1 ∥ f ∥ H v p for every f ∈ D 0 f\in D_0 f ∈ D 0 , and let a a a be a v v v -( p , p 0 , d ) (p,p_0,d) ( p , p 0 , d ) atom. Since a a a need not belong to D 0 D_0 D 0 , we approximate it. Fix ζ ∈ S ( R n ) \zeta\in\mathcal{S}(\mathbb{R}^n) ζ ∈ S ( R n ) such that ζ ^ ∈ C c ∞ ( R n ) \widehat{\zeta}\in C_c^\infty(\mathbb{R}^n) ζ ∈ C c ∞ ( R n ) and ζ ^ ≡ 1 \widehat{\zeta}\equiv1 ζ ≡ 1 in a neighborhood of the origin, let ζ λ ( x ) = λ − n ζ ( x / λ ) \zeta_\lambda(x)=\lambda^{-n}\zeta(x/\lambda) ζ λ ( x ) = λ − n ζ ( x / λ ) for λ > 0 \lambda>0 λ > 0 , and set
f k = a ∗ ζ 1 / k − a ∗ ζ k , k ∈ N . f_k=a*\zeta_{1/k}-a*\zeta_k,\qquad k\in\mathbb{N}. f k = a ∗ ζ 1/ k − a ∗ ζ k , k ∈ N .
Since a a a has compact support, a ^ ∈ C ∞ ( R n ) \widehat{a}\in C^\infty(\mathbb{R}^n) a ∈ C ∞ ( R n ) , and f k ^ ( ξ ) = a ^ ( ξ ) ( ζ ^ ( ξ / k ) − ζ ^ ( k ξ ) ) \widehat{f_k}(\xi)=\widehat{a}(\xi)\bigl(\widehat{\zeta}(\xi/k)-\widehat{\zeta}(k\xi)\bigr) f k ( ξ ) = a ( ξ ) ( ζ ( ξ / k ) − ζ ( k ξ ) ) is smooth, has compact support and vanishes in a neighborhood of the origin; hence f k ∈ D 0 f_k\in D_0 f k ∈ D 0 . As k → ∞ k\to\infty k → ∞ , a ∗ ζ 1 / k → a a*\zeta_{1/k}\to a a ∗ ζ 1/ k → a in L p 0 L^{p_0} L p 0 , because ∫ ζ = ζ ^ ( 0 ) = 1 \int\zeta=\widehat{\zeta}(0)=1 ∫ ζ = ζ ( 0 ) = 1 , while
∥ a ∗ ζ k ∥ p 0 ≤ ∥ a ∥ 1 ∥ ζ k ∥ p 0 = k − n ( 1 − 1 / p 0 ) ∥ a ∥ 1 ∥ ζ ∥ p 0 ⟶ 0 ; \|a*\zeta_k\|_{p_0}\le\|a\|_1\|\zeta_k\|_{p_0}=k^{-n(1-1/p_0)}\|a\|_1\|\zeta\|_{p_0}\longrightarrow0; ∥ a ∗ ζ k ∥ p 0 ≤ ∥ a ∥ 1 ∥ ζ k ∥ p 0 = k − n ( 1 − 1/ p 0 ) ∥ a ∥ 1 ∥ ζ ∥ p 0 ⟶ 0 ;
thus f k → a f_k\to a f k → a in L p 0 L^{p_0} L p 0 . The norms ∥ f k ∥ H v p \|f_k\|_{H_v^p} ∥ f k ∥ H v p are uniformly bounded. Indeed, let ψ ∈ S N \psi\in S_N ψ ∈ S N , t , λ > 0 t,\lambda> 0 t , λ > 0 and σ = max { t , λ } \sigma=\max\{t,\lambda\} σ = max { t , λ } . Then ψ t ∗ ζ λ = ( ψ t / σ ∗ ζ λ / σ ) σ \psi_t * \zeta_\lambda=(\psi_{t/\sigma} * \zeta_{\lambda/\sigma})_\sigma ψ t ∗ ζ λ = ( ψ t / σ ∗ ζ λ / σ ) σ , and either t / σ = 1 t/\sigma=1 t / σ = 1 or λ / σ = 1 \lambda/\sigma=1 λ / σ = 1 . A direct computation shows that the functions ψ ∗ ζ s \psi* \zeta_s ψ ∗ ζ s and ψ s ∗ ζ \psi_s * \zeta ψ s ∗ ζ , with 0 < s ≤ 1 0<s\le1 0 < s ≤ 1 and ψ ∈ S N \psi\in S_N ψ ∈ S N , are uniformly bounded in the semi-norms ∥ ⋅ ∥ μ , κ \|\cdot\|_{\mu,\kappa} ∥ ⋅ ∥ μ , κ with ∣ μ ∣ , ∣ κ ∣ ≤ N − n − 1 |\mu|,|\kappa|\le N-n-1 ∣ μ ∣ , ∣ κ ∣ ≤ N − n − 1 . Hence M N ( g ∗ ζ λ ) ≤ C M N − n − 1 g M_N(g*\zeta_\lambda)\le C M_{N-n-1}g M N ( g ∗ ζ λ ) ≤ C M N − n − 1 g with C C C independent of λ \lambda λ and, since the space H v p H_v^p H v p does not depend on N N N for N N N large (see [24 Chapter VI] ), ∥ g ∗ ζ λ ∥ H v p ≤ C ∥ g ∥ H v p \|g*\zeta_\lambda\|_{H_v^p}\le C\|g\|_{H_v^p} ∥ g ∗ ζ λ ∥ H v p ≤ C ∥ g ∥ H v p for every g ∈ H v p ( R n ) g\in H_v^p(\mathbb{R}^n) g ∈ H v p ( R n ) and λ > 0 \lambda>0 λ > 0 . By Lemma 2.3, sup k ∥ f k ∥ H v p ≤ C ∥ a ∥ H v p ≤ C \sup_k\|f_k\|_{H_v^p}\le C\|a\|_{H_v^p}\le C sup k ∥ f k ∥ H v p ≤ C ∥ a ∥ H v p ≤ C .
Finally, since T T T is bounded from L p 0 L^{p_0} L p 0 into L q 0 L^{q_0} L q 0 , T f k → T a Tf_k\to Ta T f k → T a in L q 0 L^{q_0} L q 0 , so a subsequence converges a.e., and Fatou’s lemma gives
∥ T a ∥ L u q ≤ lim inf k → ∞ ∥ T f k ∥ L u q ≤ C 1 sup k ∥ f k ∥ H v p ≤ C , \|Ta\|_{L_u^q}\le\liminf_{k\to\infty}\|Tf_k\|_{L_u^q}\le C_1\sup_k\|f_k\|_{H_v^p}\le C, ∥ T a ∥ L u q ≤ k → ∞ lim inf ∥ T f k ∥ L u q ≤ C 1 k sup ∥ f k ∥ H v p ≤ C ,
with C C C independent of a a a .
Conversely, assume ∥ T a ∥ L u q ≤ C 0 \|Ta\|_{L_u^q}\le C_0 ∥ T a ∥ L u q ≤ C 0 for all v v v -( p , p 0 , d ) (p,p_0,d) ( p , p 0 , d ) atoms. Let f ∈ D 0 f\in D_0 f ∈ D 0 and let f = ∑ j λ j a j f=\sum_j\lambda_j a_j f = ∑ j λ j a j be the decomposition of Theorem 2.4 , which converges in L p 0 L^{p_0} L p 0 . Since T T T is bounded from L p 0 L^{p_0} L p 0 into L q 0 L^{q_0} L q 0 , ∑ j λ j T a j \sum_j\lambda_jTa_j ∑ j λ j T a j converges to T f Tf T f in L q 0 L^{q_0} L q 0 , so a subsequence of the partial sums converges a.e. and
∣ T f ( x ) ∣ ≤ ∑ j ∣ λ j ∣ ∣ T a j ( x ) ∣ a.e. x ∈ R n . (10) |Tf(x)|\le\sum_j|\lambda_j||Ta_j(x)|\qquad\text{a.e. }x\in\mathbb{R}^n. \tag*{(10)} ∣ T f ( x ) ∣ ≤ j ∑ ∣ λ j ∣∣ T a j ( x ) ∣ a.e. x ∈ R n . (10)
If p ≤ q < 1 p\le q<1 p ≤ q < 1 , then (10) and the inequality ( ∑ j b j ) q ≤ ∑ j b j q (\sum_j b_j)^q\le\sum_j b_j^q ( ∑ j b j ) q ≤ ∑ j b j q give
∥ T f ∥ L u q q ≤ ∑ j ∣ λ j ∣ q ∥ T a j ∥ L u q q ≤ C 0 q ∑ j ∣ λ j ∣ q , \|Tf\|_{L_u^q}^q\le\sum_j|\lambda_j|^q\|Ta_j\|_{L_u^q}^q\le C_0^q\sum_j|\lambda_j|^q, ∥ T f ∥ L u q q ≤ j ∑ ∣ λ j ∣ q ∥ T a j ∥ L u q q ≤ C 0 q j ∑ ∣ λ j ∣ q ,
and ( ∑ j ∣ λ j ∣ q ) 1 / q ≤ ( ∑ j ∣ λ j ∣ p ) 1 / p (\sum_j|\lambda_j|^q)^{1/q}\le(\sum_j|\lambda_j|^p)^{1/p} ( ∑ j ∣ λ j ∣ q ) 1/ q ≤ ( ∑ j ∣ λ j ∣ p ) 1/ p since p ≤ q p\le q p ≤ q . If q ≥ 1 q\ge1 q ≥ 1 , Minkowski’s inequality and monotone convergence give
∥ T f ∥ L u q ≤ ∑ j ∣ λ j ∣ ∥ T a j ∥ L u q ≤ C 0 ∑ j ∣ λ j ∣ ≤ C 0 ( ∑ j ∣ λ j ∣ p ) 1 / p , \|Tf\|_{L_u^q}\le\sum_j|\lambda_j|\|Ta_j\|_{L_u^q}\le C_0\sum_j|\lambda_j|\le C_0\left(\sum_j|\lambda_j|^p\right)^{1/p}, ∥ T f ∥ L u q ≤ j ∑ ∣ λ j ∣∥ T a j ∥ L u q ≤ C 0 j ∑ ∣ λ j ∣ ≤ C 0 ( j ∑ ∣ λ j ∣ p ) 1/ p ,
since p ≤ 1 p\le1 p ≤ 1 . In both cases ∥ T f ∥ L u q ≤ C ∥ f ∥ H v p \|Tf\|_{L_u^q}\le C\|f\|_{H_v^p} ∥ T f ∥ L u q ≤ C ∥ f ∥ H v p for all f ∈ D 0 f\in D_0 f ∈ D 0 , and the result follows from Theorem 2.2 , since v v v is doubling. □
The technical lemma
We now prove that T α T_\alpha T α maps w p − w_p^- w p − -( p , p 0 , 0 ) (p,p_0,0) ( p , p 0 , 0 ) atoms uniformly into L w q q L_{w^q}^q L w q q . For a single kernel, an argument of this type is given in the proof of [15 Theorem 5.10] , following [4 ] ; the main new difficulty here is the presence of several kernels composed with the matrices A i A_i A i , whose singularities are located at the different points A i x 0 A_i x_0 A i x 0 .
Lemma 4.2. Let A 1 , … , A m A_1,\ldots,A_m A 1 , … , A m be invertible matrices, and let T α T_\alpha T α satisfy ( H 1 ) (H_1) ( H 1 ) and ( H 2 ) (H_2) ( H 2 ) , where 0 < α < n 0<\alpha<n 0 < α < n and p , δ , q p,\delta,q p , δ , q are as in Theorem 1.1 . Let s 0 = a ( t , q ) s_0=a(t,q) s 0 = a ( t , q ) be as in (1.3), let η > q \eta>q η > q , and assume
w max { s 0 , η } ∈ A 1 and w ( A i x ) ≤ C w ( x ) a.e. , 1 ≤ i ≤ m . w^{\max\{s_0,\eta\}}\in A_1\qquad\text{and}\qquad w(A_i x)\le Cw(x)\ \text{a.e.},\quad1\le i\le m. w m a x { s 0 , η } ∈ A 1 and w ( A i x ) ≤ C w ( x ) a.e. , 1 ≤ i ≤ m . Let 1 < p 0 < n α 1<p_0<\frac{n}{\alpha} 1 < p 0 < α n , 1 q 0 = 1 p 0 − α n \frac{1}{q_0}=\frac{1}{p_0}-\frac{\alpha}{n} q 0 1 = p 0 1 − n α , be such that T α : L p 0 → L q 0 T_\alpha:L^{p_0}\to L^{q_0} T α : L p 0 → L q 0 is bounded and
q 0 > q , q ( q 0 q ) ′ ≤ η . q_0>q,\qquad q\left(\frac{q_0}{q}\right)'\le\eta. q 0 > q , q ( q q 0 ) ′ ≤ η .
Then there is C > 0 C>0 C > 0 such that ∥ T α a ∥ L w q q ≤ C \|T_\alpha a\|_{L^q_{w^q}} \le C ∥ T α a ∥ L w q q ≤ C for every function a a a supported in a ball B B B with ∥ a ∥ p 0 ≤ ∣ B ∣ 1 / p 0 [ w p ( B ) ] − 1 / p \|a\|_{p_0} \le|B|^{1/p_0}[w^p(B)]^{-1/p} ∥ a ∥ p 0 ≤ ∣ B ∣ 1/ p 0 [ w p ( B ) ] − 1/ p and ∫ a = 0 \int a=0 ∫ a = 0 ; in particular, for every w p w^p w p -( p , p 0 , d ) (p,p_0,d) ( p , p 0 , d ) atom.
Proof. Let B = B ( x 0 , r ) B=B(x_0,r) B = B ( x 0 , r ) be the ball associated with a a a . The idea of the proof is to split R n \mathbb{R}^n R n into a region near the images A i ( B ) A_i(B) A i ( B ) of the support of a a a , where we only use the size of a a a and the L p 0 − L q 0 L^{p_0}-L^{q_0} L p 0 − L q 0 boundedness of T α T_\alpha T α , and a far region, where we use the cancellation of a a a together with the regularity of the kernels given by ( H 2 ) (\mathrm{H}_2) ( H 2 ) . Before doing so, we collect the estimates on the weight and on the kernels that will be needed.
We first note that, by Hölder’s inequality and the size condition on a a a ,
∥ a ∥ 1 ≤ ∣ B ∣ 1 − 1 / p 0 ∥ a ∥ p 0 ≤ ∣ B ∣ [ w p ( B ) ] − 1 / p . (11) \|a\|_1 \le|B|^{1-1/p_0}\|a\|_{p_0} \le|B|[w^p(B)]^{-1/p}. \tag*{(11)} ∥ a ∥ 1 ≤ ∣ B ∣ 1 − 1/ p 0 ∥ a ∥ p 0 ≤ ∣ B ∣ [ w p ( B ) ] − 1/ p . (11)
Let
ρ = ( q 0 q ) ′ , γ i = n q i − n p i > 0 , 1 ≤ i ≤ m . \rho=\left(\frac{q_0}{q}\right)',\qquad\gamma_i=\frac{n}{q_i}-\frac{n}{p_i}>0,\qquad1\le i\le m. ρ = ( q q 0 ) ′ , γ i = q i n − p i n > 0 , 1 ≤ i ≤ m .
Since p < q ≤ s 0 p<q\le s_0 p < q ≤ s 0 and q < q ρ ≤ η q<q\rho\le\eta q < q ρ ≤ η , all the exponents s 0 s_0 s 0 , q ρ q\rho q ρ , p p p and q q q are at most max { s 0 , η } \max\{s_0,\eta\} max { s 0 , η } , so Lemma 2.1 (i), applied to v = w max { s 0 , η } v=w^{\max\{s_0,\eta\}} v = w m a x { s 0 , η } , shows that w s 0 w^{s_0} w s 0 , w q ρ w^{q\rho} w q ρ , w p w^p w p , w q ∈ A 1 w^q\in A_1 w q ∈ A 1 . In particular, Lemma 2.1 (iii), applied to v = w q ρ v=w^{q\rho} v = w q ρ with θ = 1 / ρ \theta=1/\rho θ = 1/ ρ , gives w q ∈ R H ρ w^q\in RH_\rho w q ∈ R H ρ , and, by Lemma 2.1 (iv), the measures w p ( x ) d x w^p(x)\,dx w p ( x ) d x and w q ( x ) d x w^q(x)\,dx w q ( x ) d x satisfy w p ( λ Q ) ≤ C λ n w p ( Q ) w^p(\lambda Q)\le C\lambda^n w^p(Q) w p ( λ Q ) ≤ C λ n w p ( Q ) and w q ( λ Q ) ≤ C λ n w q ( Q ) w^q(\lambda Q)\le C\lambda^n w^q(Q) w q ( λ Q ) ≤ C λ n w q ( Q ) for every ball Q Q Q and λ ≥ 1 \lambda\ge1 λ ≥ 1 .
The weights w p w^p w p and w q w^q w q can be compared on balls. Indeed, since w q ∈ A 1 w^q\in A_1 w q ∈ A 1 and, trivially, ess inf Q w ≤ ( w p ( Q ) / ∣ Q ∣ ) 1 / p \operatorname*{ess\,inf}_Q w\le(w^p(Q)/|Q|)^{1/p} ess inf Q w ≤ ( w p ( Q ) /∣ Q ∣ ) 1/ p , we have
( w q ( Q ) ∣ Q ∣ ) 1 / q ≤ C ess inf Q w ≤ C ( w p ( Q ) ∣ Q ∣ ) 1 / p , \left(\frac{w^q(Q)}{|Q|}\right)^{1/q}\le C\operatorname*{ess\,inf}_Q w\le C\left(\frac{w^p(Q)}{|Q|}\right)^{1/p}, ( ∣ Q ∣ w q ( Q ) ) 1/ q ≤ C Q ess inf w ≤ C ( ∣ Q ∣ w p ( Q ) ) 1/ p ,
and, as 1 q − 1 p = − α n \frac{1}{q}-\frac{1}{p}=-\frac{\alpha}{n} q 1 − p 1 = − n α , this gives
[ w q ( Q ) ] 1 / q ≤ C ∣ Q ∣ − α / n [ w p ( Q ) ] 1 / p for every ball Q . (12) [w^q(Q)]^{1/q}\le C|Q|^{-\alpha/n}[w^p(Q)]^{1/p}\qquad\text{for every ball }Q. \tag*{(12)} [ w q ( Q ) ] 1/ q ≤ C ∣ Q ∣ − α / n [ w p ( Q ) ] 1/ p for every ball Q . (12)
In the same way, the condition w s 0 ∈ A 1 w^{s_0}\in A_1 w s 0 ∈ A 1 allows us to control the L s 0 L^{s_0} L s 0 norm of w w w on subsets of a ball by its w q w^q w q measure: for every measurable E ⊂ Q E\subset Q E ⊂ Q ,
∥ w χ E ∥ s 0 ≤ ( C ∣ Q ∣ ess inf Q w s 0 ) 1 / s 0 ≤ C ∣ Q ∣ 1 / s 0 − 1 / q [ w q ( Q ) ] 1 / q . (13) \|w\chi_E\|_{s_0}\le(C|Q|\operatorname*{ess\,inf}_Q w^{s_0})^{1/s_0}\le C|Q|^{1/s_0-1/q}[w^q(Q)]^{1/q}. \tag*{(13)} ∥ w χ E ∥ s 0 ≤ ( C ∣ Q ∣ Q ess inf w s 0 ) 1/ s 0 ≤ C ∣ Q ∣ 1/ s 0 − 1/ q [ w q ( Q ) ] 1/ q . (13)
As for the kernels, ( H 1 ) (\mathrm{H}_1) ( H 1 ) and Lemma 3.5 give the tail estimate
∥ k i χ { ∣ x ∣ > u } ∥ p i ≤ C u − γ i , u > 0 , 1 ≤ i ≤ m . (14) \|k_i\chi_{\{|x|>u\}}\|_{p_i}\le Cu^{-\gamma_i},\qquad u>0,\quad1\le i\le m. \tag*{(14)} ∥ k i χ { ∣ x ∣ > u } ∥ p i ≤ C u − γ i , u > 0 , 1 ≤ i ≤ m . (14)
We now define the two regions. Let c i c_i c i be the constants in ( H 2 ) (\mathrm{H}_2) ( H 2 ) , and fix
M > max { 1 , max 1 ≤ i ≤ m ∥ A i ∥ , max 1 ≤ i ≤ m c i ∥ A i ∥ } , L > max { 1 , 4 M max 1 ≤ i ≤ m ∥ A i − 1 ∥ } . M>\max\left\{1,\max_{1\le i\le m}\|A_i\|,\max_{1\le i\le m}c_i\|A_i\|\right\},\qquad L>\max\left\{1,4M\max_{1\le i\le m}\|A_i^{-1}\|\right\}. M > max { 1 , 1 ≤ i ≤ m max ∥ A i ∥ , 1 ≤ i ≤ m max c i ∥ A i ∥ } , L > max { 1 , 4 M 1 ≤ i ≤ m max ∥ A i − 1 ∥ } .
Let B ∗ = B ( x 0 , L r ) B^*=B(x_0,Lr) B ∗ = B ( x 0 , L r ) , U = ⋃ i = 1 m A i ( B ∗ ) U=\bigcup_{i=1}^m A_i(B^*) U = ⋃ i = 1 m A i ( B ∗ ) and D = R n ∖ U D=\mathbb{R}^n\setminus U D = R n ∖ U . Then
∥ T α a ∥ L w q q ≤ ∑ i = 1 m I i + I m + 1 , I i = ∥ T α a χ A i ( B ∗ ) ∥ L w q q , I m + 1 = ∥ T α a χ D ∥ L w q q , \|T_\alpha a\|_{L^q_{w^q}}\le\sum_{i=1}^m I_i+I_{m+1},\qquad I_i=\|T_\alpha a\chi_{A_i(B^*)}\|_{L^q_{w^q}},\qquad I_{m+1}=\|T_\alpha a\chi_D\|_{L^q_{w^q}}, ∥ T α a ∥ L w q q ≤ i = 1 ∑ m I i + I m + 1 , I i = ∥ T α a χ A i ( B ∗ ) ∥ L w q q , I m + 1 = ∥ T α a χ D ∥ L w q q ,
and it suffices to show that each of these terms is bounded by a constant independent of a a a .
We begin with the terms I i I_i I i , 1 ≤ i ≤ m 1 \le i \le m 1 ≤ i ≤ m , which correspond to the near region. Hölder’s inequality with exponents q 0 / q q_0/q q 0 / q and ρ \rho ρ gives
I i q ≤ ∥ T α a ∥ q 0 q ( ∫ A i ( B ∗ ) w q ρ ) 1 / ρ . I_i^q \le\|T_\alpha a\|_{q_0}^q \left(\int_{A_i(B^*)} w^{q\rho}\right)^{1/\rho}. I i q ≤ ∥ T α a ∥ q 0 q ( ∫ A i ( B ∗ ) w q ρ ) 1/ ρ .
To estimate the integral of the weight, we change variables y = A i x y=A_i x y = A i x and use the hypothesis w ( A i x ) ≤ C w ( x ) w(A_i x) \le Cw(x) w ( A i x ) ≤ C w ( x ) , which reduces the integral over A i ( B ∗ ) A_i(B^*) A i ( B ∗ ) to an integral over B ∗ B^* B ∗ ; then the reverse Hölder property w q ∈ R H ρ w^q \in RH_\rho w q ∈ R H ρ yields
( ∫ A i ( B ∗ ) w q ρ ) 1 / ρ = ∣ det A i ∣ 1 / ρ ( ∫ B ∗ w ( A i x ) q ρ d x ) 1 / ρ ≤ C ( ∫ B ∗ w q ρ ) 1 / ρ ≤ C ∣ B ∗ ∣ 1 ρ − 1 w q ( B ∗ ) . \left(\int_{A_i(B^*)} w^{q\rho}\right)^{1/\rho}
=|\det A_i|^{1/\rho}\left(\int_{B^*}w(A_i x)^{q\rho}\,dx\right)^{1/\rho}
\le C\left(\int_{B^*}w^{q\rho}\right)^{1/\rho}
\le C|B^*|^{\frac{1}{\rho}-1}w^q(B^*). ( ∫ A i ( B ∗ ) w q ρ ) 1/ ρ = ∣ det A i ∣ 1/ ρ ( ∫ B ∗ w ( A i x ) q ρ d x ) 1/ ρ ≤ C ( ∫ B ∗ w q ρ ) 1/ ρ ≤ C ∣ B ∗ ∣ ρ 1 − 1 w q ( B ∗ ) .
On the other hand, ∥ T α a ∥ q 0 ≤ C ∥ a ∥ p 0 \|T_\alpha a\|_{q_0}\le C\|a\|_{p_0} ∥ T α a ∥ q 0 ≤ C ∥ a ∥ p 0 , and the size condition on a a a together with the doubling property w p ( B ∗ ) ≤ C w p ( B ) w^p(B^*)\le Cw^p(B) w p ( B ∗ ) ≤ C w p ( B ) and ∣ B ∣ ≤ ∣ B ∗ ∣ |B|\le|B^*| ∣ B ∣ ≤ ∣ B ∗ ∣ lead to
I i q ≤ C ∣ B ∣ q p 0 [ w p ( B ) ] − q p ∣ B ∗ ∣ 1 ρ − 1 w q ( B ∗ ) ≤ C ∣ B ∗ ∣ q p 0 + 1 ρ − 1 [ w p ( B ∗ ) ] − q p w q ( B ∗ ) . I_i^q\le C|B|^{\frac{q}{p_0}}[w^p(B)]^{-\frac{q}{p}}|B^*|^{\frac{1}{\rho}-1}w^q(B^*)\le C|B^*|^{\frac{q}{p_0}+\frac{1}{\rho}-1}[w^p(B^*)]^{-\frac{q}{p}}w^q(B^*). I i q ≤ C ∣ B ∣ p 0 q [ w p ( B ) ] − p q ∣ B ∗ ∣ ρ 1 − 1 w q ( B ∗ ) ≤ C ∣ B ∗ ∣ p 0 q + ρ 1 − 1 [ w p ( B ∗ ) ] − p q w q ( B ∗ ) .
Since 1 ρ = 1 − q p 0 \frac{1}{\rho}=1-\frac{q}{p_0} ρ 1 = 1 − p 0 q , the exponent of ∣ B ∗ ∣ |B^*| ∣ B ∗ ∣ is q p 0 + 1 ρ − 1 = q ( 1 p 0 − 1 q 0 ) = q α n \frac{q}{p_0}+\frac{1}{\rho}-1=q\left(\frac{1}{p_0}-\frac{1}{q_0}\right)=\frac{q\alpha}{n} p 0 q + ρ 1 − 1 = q ( p 0 1 − q 0 1 ) = n q α , which is exactly compensated by (12) applied with Q = B ∗ Q=B^* Q = B ∗ :
I i q ≤ C ∣ B ∗ ∣ q α n [ w p ( B ∗ ) ] − q p ∣ B ∗ ∣ − q α n [ w p ( B ∗ ) ] q p = C . I_i^q\le C|B^*|^{\frac{q\alpha}{n}}[w^p(B^*)]^{-\frac{q}{p}}|B^*|^{-\frac{q\alpha}{n}}[w^p(B^*)]^{\frac{q}{p}}=C. I i q ≤ C ∣ B ∗ ∣ n q α [ w p ( B ∗ ) ] − p q ∣ B ∗ ∣ − n q α [ w p ( B ∗ ) ] p q = C .
We now turn to I m + 1 I_{m+1} I m + 1 . Here y y y is far from all the points A i x 0 A_i x_0 A i x 0 , and we organize D D D according to the nearest of these points and to the dyadic distance to it. If y ∈ D y\in D y ∈ D , then ∣ A ℓ − 1 y − x 0 ∣ ≥ L r |A_\ell^{-1}y-x_0|\ge Lr ∣ A ℓ − 1 y − x 0 ∣ ≥ L r , hence
∣ y − A ℓ x 0 ∣ ≥ L r ∥ A ℓ − 1 ∥ > 4 M r |y-A_\ell x_0|\ge\frac{Lr}{\|A_\ell^{-1}\|}>4Mr ∣ y − A ℓ x 0 ∣ ≥ ∥ A ℓ − 1 ∥ L r > 4 M r
for every 1 ≤ ℓ ≤ m 1\le\ell\le m 1 ≤ ℓ ≤ m .
For 1 ≤ ℓ ≤ m 1\le\ell\le m 1 ≤ ℓ ≤ m let D ℓ D_\ell D ℓ be the set of points y ∈ D y\in D y ∈ D such that ∣ y − A ℓ x 0 ∣ < ∣ y − A i x 0 ∣ |y-A_\ell x_0|<|y-A_i x_0| ∣ y − A ℓ x 0 ∣ < ∣ y − A i x 0 ∣ for i < ℓ i<\ell i < ℓ and ∣ y − A ℓ x 0 ∣ ≤ ∣ y − A i x 0 ∣ |y-A_\ell x_0|\le|y-A_i x_0| ∣ y − A ℓ x 0 ∣ ≤ ∣ y − A i x 0 ∣ for i > ℓ i>\ell i > ℓ ; that is, each point of D D D is assigned to a nearest center A ℓ x 0 A_\ell x_0 A ℓ x 0 , ties being assigned to the smallest index. The sets D ℓ D_\ell D ℓ are measurable and form a partition of D D D , and so do the sets
D ℓ , j = D ℓ ∩ { y : 2 j M r < ∣ y − A ℓ x 0 ∣ ≤ 2 j + 1 M r } , 1 ≤ ℓ ≤ m , j ≥ 1. D_{\ell,j}=D_\ell\cap\{y:2^jMr<|y-A_\ell x_0|\le2^{j+1}Mr\},\qquad1\le\ell\le m,\ j\ge1. D ℓ , j = D ℓ ∩ { y : 2 j M r < ∣ y − A ℓ x 0 ∣ ≤ 2 j + 1 M r } , 1 ≤ ℓ ≤ m , j ≥ 1.
If y ∈ D ℓ , j y\in D_{\ell,j} y ∈ D ℓ , j , 1 ≤ i ≤ m 1\le i\le m 1 ≤ i ≤ m and z ∈ B z\in B z ∈ B , then ∣ y − A i x 0 ∣ ≥ ∣ y − A ℓ x 0 ∣ > 2 j M r |y-A_i x_0|\ge|y-A_\ell x_0|>2^jMr ∣ y − A i x 0 ∣ ≥ ∣ y − A ℓ x 0 ∣ > 2 j M r , since A ℓ x 0 A_\ell x_0 A ℓ x 0 is a nearest center, and therefore, as M > ∥ A i ∥ M>\|A_i\| M > ∥ A i ∥ ,
∣ y − A i z ∣ ≥ ∣ y − A i x 0 ∣ − ∥ A i ∥ ∣ z − x 0 ∣ > ( 2 j − 1 ) M r ≥ 2 j − 1 M r . (15) |y-A_i z|\ge|y-A_i x_0|-\|A_i\||z-x_0|>(2^j-1)Mr\ge2^{j-1}Mr. \tag*{(15)} ∣ y − A i z ∣ ≥ ∣ y − A i x 0 ∣ − ∥ A i ∥∣ z − x 0 ∣ > ( 2 j − 1 ) M r ≥ 2 j − 1 M r . (15)
In other words, on D ℓ , j D_{\ell,j} D ℓ , j all the singularities of the kernel are at distance comparable to 2 j r 2^jr 2 j r or larger.
Since ∫ a = 0 \int a=0 ∫ a = 0 , we can write
T α a ( y ) = ∫ B [ K ( y , z ) − K ( y , x 0 ) ] a ( z ) d z , T_\alpha a(y)=\int_B[K(y,z)-K(y,x_0)]a(z)\,dz, T α a ( y ) = ∫ B [ K ( y , z ) − K ( y , x 0 )] a ( z ) d z ,
and, writing the difference of the products as a telescoping sum, ∣ K ( y , z ) − K ( y , x 0 ) ∣ ≤ ∑ i 0 = 1 m F i 0 , z ( y ) |K(y,z)-K(y,x_0)|\le\sum_{i_0=1}^m F_{i_0,z}(y) ∣ K ( y , z ) − K ( y , x 0 ) ∣ ≤ ∑ i 0 = 1 m F i 0 , z ( y ) , where
F i 0 , z ( y ) = ∏ i < i 0 ∣ k i ( y − A i x 0 ) ∣ ∣ k i 0 ( y − A i 0 z ) − k i 0 ( y − A i 0 x 0 ) ∣ ∏ i > i 0 ∣ k i ( y − A i z ) ∣ . F_{i_0,z}(y)=\prod_{i<i_0}|k_i(y-A_i x_0)|\,|k_{i_0}(y-A_{i_0}z)-k_{i_0}(y-A_{i_0}x_0)|\prod_{i>i_0}|k_i(y-A_i z)|. F i 0 , z ( y ) = i < i 0 ∏ ∣ k i ( y − A i x 0 ) ∣ ∣ k i 0 ( y − A i 0 z ) − k i 0 ( y − A i 0 x 0 ) ∣ i > i 0 ∏ ∣ k i ( y − A i z ) ∣.
In each F i 0 , z F_{i_0,z} F i 0 , z only the i 0 i_0 i 0 -th factor is a difference, and it will be controlled by ( H 2 ) (\mathrm{H}_2) ( H 2 ) , while the remaining factors will be controlled by the tail estimate (14) . To make this precise, let h i 0 ( z ) = A i 0 ( z − x 0 ) h_{i_0}(z)=A_{i_0}(z-x_0) h i 0 ( z ) = A i 0 ( z − x 0 ) and
N i 0 , k ( z ) = ∥ [ k i 0 ( ⋅ − h i 0 ( z ) ) − k i 0 ] χ { 2 k M r < ∣ ⋅ ∣ ≤ 2 k + 1 M r } ∥ p i 0 , k ≥ 1. N_{i_0,k}(z)=\left\|\left[k_{i_0}(\cdot-h_{i_0}(z))-k_{i_0}\right]\chi_{\{2^kMr<|\cdot|\leq2^{k+1}Mr\}}\right\|_{p_{i_0}},\qquad k\geq1. N i 0 , k ( z ) = [ k i 0 ( ⋅ − h i 0 ( z )) − k i 0 ] χ { 2 k M r < ∣ ⋅ ∣ ≤ 2 k + 1 M r } p i 0 , k ≥ 1.
Since ∣ h i 0 ( z ) ∣ < ∥ A i 0 ∥ r |h_{i_0}(z)|<\|A_{i_0}\|r ∣ h i 0 ( z ) ∣ < ∥ A i 0 ∥ r and M > c i 0 ∥ A i 0 ∥ M>c_{i_0}\|A_{i_0}\| M > c i 0 ∥ A i 0 ∥ , we have M r > c i 0 ∣ h i 0 ( z ) ∣ Mr>c_{i_0}|h_{i_0}(z)| M r > c i 0 ∣ h i 0 ( z ) ∣ , so ( H 2 ) (\mathrm{H}_2) ( H 2 ) can be applied with R = M r R=Mr R = M r and h = h i 0 ( z ) h=h_{i_0}(z) h = h i 0 ( z ) . Taking into account that ∥ f χ { ∣ x ∣ ∼ s } ∥ p i 0 = ∣ B ( 0 , 2 s ) ∣ 1 / p i 0 ∥ f ∥ p i 0 , ∣ x ∣ ∼ s \|f\chi_{\{|x|\sim s\}}\|_{p_{i_0}}=|B(0,2s)|^{1/p_{i_0}}\|f\|_{p_{i_0},|x|\sim s} ∥ f χ { ∣ x ∣ ∼ s } ∥ p i 0 = ∣ B ( 0 , 2 s ) ∣ 1/ p i 0 ∥ f ∥ p i 0 , ∣ x ∣ ∼ s , it gives
∑ k = 1 ∞ 2 k δ ( 2 k M r ) γ i 0 N i 0 , k ( z ) ≤ C , uniformly in z ∈ B . (16) \sum_{k=1}^{\infty}2^{k\delta}(2^kMr)^{\gamma_{i_0}}N_{i_0,k}(z)\leq C,\qquad\text{uniformly in }z\in B. \tag*{(16)} k = 1 ∑ ∞ 2 k δ ( 2 k M r ) γ i 0 N i 0 , k ( z ) ≤ C , uniformly in z ∈ B . (16)
We now estimate F i 0 , z F_{i_0,z} F i 0 , z on each set D ℓ , j D_{\ell,j} D ℓ , j . By (15) and (14) , for i ≠ i 0 i\ne i_0 i = i 0 the L p i L^{p_i} L p i norm on D ℓ , j D_{\ell,j} D ℓ , j of the factor k i ( ⋅ − A i x 0 ) k_i(\cdot-A_ix_0) k i ( ⋅ − A i x 0 ) or k i ( ⋅ − A i z ) k_i(\cdot-A_iz) k i ( ⋅ − A i z ) is bounded by C ( 2 j r ) − γ i C(2^jr)^{-\gamma_i} C ( 2 j r ) − γ i . For the difference factor, the change of variables x = y − A i 0 x 0 x=y-A_{i_0}x_0 x = y − A i 0 x 0 and the inclusion D ℓ , j − A i 0 x 0 ⊂ { ∣ x ∣ > 2 j M r } D_{\ell,j}-A_{i_0}x_0\subset\{|x|>2^jMr\} D ℓ , j − A i 0 x 0 ⊂ { ∣ x ∣ > 2 j M r } show that
∥ [ k i 0 ( ⋅ − A i 0 z ) − k i 0 ( ⋅ − A i 0 x 0 ) ] χ D ℓ , j ∥ p i 0 ≤ ∑ k = j ∞ N i 0 , k ( z ) . \left\|\left[k_{i_0}(\cdot-A_{i_0}z)-k_{i_0}(\cdot-A_{i_0}x_0)\right]\chi_{D_{\ell,j}}\right\|_{p_{i_0}}\leq\sum_{k=j}^{\infty}N_{i_0,k}(z). [ k i 0 ( ⋅ − A i 0 z ) − k i 0 ( ⋅ − A i 0 x 0 ) ] χ D ℓ , j p i 0 ≤ k = j ∑ ∞ N i 0 , k ( z ) .
For the weight, let Q ℓ = B ( A ℓ x 0 , 2 M r ) Q_\ell=B(A_\ell x_0,2Mr) Q ℓ = B ( A ℓ x 0 , 2 M r ) . Since D ℓ , j ⊂ 2 j Q ℓ D_{\ell,j}\subset2^jQ_\ell D ℓ , j ⊂ 2 j Q ℓ , (13) and the doubling property of w q w^q w q yield
∥ w χ D ℓ , j ∥ s 0 ≤ C ( 2 j r ) n s 0 − n q [ w q ( 2 j Q ℓ ) ] 1 / q ≤ C ( 2 j r ) n s 0 − n q 2 j n q [ w q ( Q ℓ ) ] 1 / q . \|w\chi_{D_{\ell,j}}\|_{s_0}\leq C(2^jr)^{\frac{n}{s_0}-\frac{n}{q}}[w^q(2^jQ_\ell)]^{1/q}\leq C(2^jr)^{\frac{n}{s_0}-\frac{n}{q}}2^{j\frac{n}{q}}[w^q(Q_\ell)]^{1/q}. ∥ w χ D ℓ , j ∥ s 0 ≤ C ( 2 j r ) s 0 n − q n [ w q ( 2 j Q ℓ ) ] 1/ q ≤ C ( 2 j r ) s 0 n − q n 2 j q n [ w q ( Q ℓ ) ] 1/ q .
We combine these three estimates by means of the generalized Hölder inequality with 1 q = ∑ i = 1 m 1 p i + 1 s 0 \frac{1}{q}=\sum_{i=1}^m\frac{1}{p_i}+\frac{1}{s_0} q 1 = ∑ i = 1 m p i 1 + s 0 1 . Since, by (1) and (3) ,
− ∑ i ≠ i 0 γ i + n s 0 − n q = α − n + γ i 0 and α − n + n q = δ , -\sum_{i\ne i_0}\gamma_i+\frac{n}{s_0}-\frac{n}{q}=\alpha-n+\gamma_{i_0}\qquad\text{and}\qquad\alpha-n+\frac{n}{q}=\delta, − i = i 0 ∑ γ i + s 0 n − q n = α − n + γ i 0 and α − n + q n = δ ,
the powers of 2 j r 2^jr 2 j r combine into
∥ F i 0 , z χ D ℓ , j ∥ L w q ≤ C r α − n [ w q ( Q ℓ ) ] 1 / q 2 j δ ( 2 j r ) γ i 0 ∑ k = j ∞ N i 0 , k ( z ) . (17) \|F_{i_0,z}\chi_{D_{\ell,j}}\|_{L^q_w}\leq Cr^{\alpha-n}[w^q(Q_\ell)]^{1/q}2^{j\delta}(2^jr)^{\gamma_{i_0}}\sum_{k=j}^{\infty}N_{i_0,k}(z). \tag*{(17)} ∥ F i 0 , z χ D ℓ , j ∥ L w q ≤ C r α − n [ w q ( Q ℓ ) ] 1/ q 2 j δ ( 2 j r ) γ i 0 k = j ∑ ∞ N i 0 , k ( z ) . (17)
It is at this point that the factor 2 j δ 2^{j\delta} 2 j δ appears, and it is absorbed by ( H 2 ) (\mathrm{H}_2) ( H 2 ) when we sum in j j j . Indeed, since γ i 0 + δ > 0 \gamma_{i_0}+\delta>0 γ i 0 + δ > 0 , interchanging the order of summation and using (16) we obtain
∑ j = 1 ∞ 2 j δ ( 2 j r ) γ i 0 ∑ k = j ∞ N i 0 , k ( z ) = r γ i 0 ∑ k = 1 ∞ N i 0 , k ( z ) ∑ j = 1 k 2 j ( γ i 0 + δ ) ≤ C ∑ k = 1 ∞ 2 k δ ( 2 k r ) γ i 0 N i 0 , k ( z ) ≤ C , \begin{aligned}
\sum_{j=1}^{\infty}2^{j\delta}(2^jr)^{\gamma_{i_0}}\sum_{k=j}^{\infty}N_{i_0,k}(z)
=r^{\gamma_{i_0}}\sum_{k=1}^{\infty}N_{i_0,k}(z)\sum_{j=1}^{k}2^{j(\gamma_{i_0}+\delta)} \\
&\leq C\sum_{k=1}^{\infty}2^{k\delta}(2^kr)^{\gamma_{i_0}}N_{i_0,k}(z)\leq C,
\end{aligned} j = 1 ∑ ∞ 2 j δ ( 2 j r ) γ i 0 k = j ∑ ∞ N i 0 , k ( z ) = r γ i 0 k = 1 ∑ ∞ N i 0 , k ( z ) j = 1 ∑ k 2 j ( γ i 0 + δ ) ≤ C k = 1 ∑ ∞ 2 k δ ( 2 k r ) γ i 0 N i 0 , k ( z ) ≤ C ,
where the powers of M M M have been absorbed into C C C . Therefore, since q > 1 q>1 q > 1 , Minkowski’s integral inequality and (17) give
I m + 1 ≤ ∑ ℓ = 1 m ∑ j = 1 ∞ ∑ i 0 = 1 m ∫ B ∣ a ( z ) ∣ ∥ F i 0 , z χ D ℓ , j ∥ L w q d z ≤ C r α − n ∥ a ∥ 1 ∑ ℓ = 1 m [ w q ( Q ℓ ) ] 1 / q . I_{m+1}\leq\sum_{\ell=1}^{m}\sum_{j=1}^{\infty}\sum_{i_0=1}^{m}\int_B|a(z)|\|F_{i_0,z}\chi_{D_{\ell,j}}\|_{L^q_w}\,dz\leq Cr^{\alpha-n}\|a\|_1\sum_{\ell=1}^{m}[w^q(Q_\ell)]^{1/q}. I m + 1 ≤ ℓ = 1 ∑ m j = 1 ∑ ∞ i 0 = 1 ∑ m ∫ B ∣ a ( z ) ∣∥ F i 0 , z χ D ℓ , j ∥ L w q d z ≤ C r α − n ∥ a ∥ 1 ℓ = 1 ∑ m [ w q ( Q ℓ ) ] 1/ q .
It remains to compare w q ( Q ℓ ) w^{q}(Q_\ell) w q ( Q ℓ ) with w q ( B ) w^{q}(B) w q ( B ) . Since Q ℓ ⊂ A ℓ ( B ( x 0 , 2 M ∥ A ℓ − 1 ∥ r ) ) Q_\ell\subset A_\ell(B(x_0,2M\|A_\ell^{-1}\|r)) Q ℓ ⊂ A ℓ ( B ( x 0 , 2 M ∥ A ℓ − 1 ∥ r )) , the change of variables y = A ℓ x y=A_\ell x y = A ℓ x , the hypothesis w ( A ℓ x ) ≤ C w ( x ) w(A_\ell x)\leq Cw(x) w ( A ℓ x ) ≤ C w ( x ) , and the doubling property of w q w^{q} w q (note that 2 M ∥ A ℓ − 1 ∥ > 2 ∥ A ℓ ∥ ∥ A ℓ − 1 ∥ ≥ 2 2M\|A_\ell^{-1}\|>2\|A_\ell\|\|A_\ell^{-1}\|\geq2 2 M ∥ A ℓ − 1 ∥ > 2∥ A ℓ ∥∥ A ℓ − 1 ∥ ≥ 2 ) give w q ( Q ℓ ) ≤ C w q ( B ) w^{q}(Q_\ell)\leq Cw^{q}(B) w q ( Q ℓ ) ≤ C w q ( B ) . Finally, by (11) and (12) with Q = B Q=B Q = B ,
I m + 1 ≤ C r α − n ∣ B ∣ [ w p ( B ) ] − 1 / p [ w q ( B ) ] 1 / q ≤ C r α − n ∣ B ∣ 1 − α / n ≤ C . I_{m+1}\leq Cr^{\alpha-n}|B|[w^{p}(B)]^{-1/p}[w^{q}(B)]^{1/q}\leq Cr^{\alpha-n}|B|^{1-\alpha/n}\leq C. I m + 1 ≤ C r α − n ∣ B ∣ [ w p ( B ) ] − 1/ p [ w q ( B ) ] 1/ q ≤ C r α − n ∣ B ∣ 1 − α / n ≤ C .
This, together with the bound for I 1 , … , I m I_1,\ldots,I_m I 1 , … , I m , proves the lemma. □
Remark 4.3. In the proof of Lemma 4.2, the factor 2 j δ 2^{j\delta} 2 j δ in (17) comes exactly from the doubling estimate w q ( 2 j Q ℓ ) ≤ C 2 j n w q ( Q ℓ ) w^{q}(2^{j}Q_\ell)\leq C2^{jn}w^{q}(Q_\ell) w q ( 2 j Q ℓ ) ≤ C 2 j n w q ( Q ℓ ) , combined with the factor r α − n r^{\alpha-n} r α − n . Indeed, after the generalized Hölder inequality, the powers of 2 j 2^{j} 2 j that are not absorbed into ( 2 j r ) γ i 0 (2^{j}r)^{\gamma_{i_0}} ( 2 j r ) γ i 0 are 2 j ( α − n ) 2^{j(\alpha-n)} 2 j ( α − n ) and 2 j n / q 2^{jn/q} 2 j n / q , and α − n + n / q = δ \alpha-n+n/q=\delta α − n + n / q = δ . For w ≡ 1 w\equiv1 w ≡ 1 this doubling estimate is an identity, ∣ 2 j Q ℓ ∣ = 2 j n ∣ Q ℓ ∣ |2^{j}Q_\ell|=2^{jn}|Q_\ell| ∣ 2 j Q ℓ ∣ = 2 j n ∣ Q ℓ ∣ , so no loss occurs at this step. This explains why the counterexample in Remark 4.7 already works without weights.
The unweighted estimate
The following lemma is a consequence of the results in [16 ] .
Lemma 4.4. Assume (H) and (H1 _1 1 ), and that k i ∈ H n − n / q i , p i k_i\in H_{n-n/q_i,p_i} k i ∈ H n − n / q i , p i for 1 ≤ i ≤ m 1\leq i\leq m 1 ≤ i ≤ m (this holds, for instance, under (H2 _2 2 ), by Remark 3.3). Let t t t be as in (3) . If t < p 0 < n α t<p_0<\frac{n}{\alpha} t < p 0 < α n and 1 q 0 = 1 p 0 − α n \frac{1}{q_0}=\frac{1}{p_0}-\frac{\alpha}{n} q 0 1 = p 0 1 − n α , then T α T_\alpha T α is bounded from L p 0 ( R n ) L^{p_0}(\mathbb{R}^n) L p 0 ( R n ) into L q 0 ( R n ) L^{q_0}(\mathbb{R}^n) L q 0 ( R n ) .
Proof. Set α i = n / q i \alpha_i=n/q_i α i = n / q i , so that 0 < α i < n 0<\alpha_i<n 0 < α i < n and α 1 + ⋯ + α m = n − α \alpha_1+\cdots+\alpha_m=n-\alpha α 1 + ⋯ + α m = n − α by (1) . We first reduce the problem to finite exponents, since the results of [16 ] are stated for Young functions. Let r i = p i r_i=p_i r i = p i if p i < ∞ p_i<\infty p i < ∞ , and let r i < ∞ r_i<\infty r i < ∞ be large if p i = ∞ p_i=\infty p i = ∞ . By Remark 3.3, k i ∈ S n − n / q i , r i ∩ H n − n / q i , r i k_i\in S_{n-n/q_i,r_i}\cap H_{n-n/q_i,r_i} k i ∈ S n − n / q i , r i ∩ H n − n / q i , r i for 1 ≤ i ≤ m 1\leq i\leq m 1 ≤ i ≤ m . Let t ~ \widetilde{t} t be defined by 1 r 1 + ⋯ + 1 r m + 1 t ~ = 1 \frac{1}{r_1}+\cdots+\frac{1}{r_m}+\frac{1}{\widetilde{t}}=1 r 1 1 + ⋯ + r m 1 + t 1 = 1 . Then t ≤ t ~ t\leq\widetilde{t} t ≤ t , and t ~ → t \widetilde{t}\to t t → t as the exponents r i r_i r i corresponding to p i = ∞ p_i=\infty p i = ∞ tend to infinity; since t < p 0 t<p_0 t < p 0 , we may choose them so that t ~ < p 0 \widetilde{t}<p_0 t < p 0 .
Let Ψ i ( u ) = u r i \Psi_i(u)=u^{r_i} Ψ i ( u ) = u r i and φ ( u ) = u t ~ \varphi(u)=u^{\widetilde{t}} φ ( u ) = u t , which are Young functions. The normalized annular norms of [16 ] coincide with ours (the average is taken over B ( 0 , 2 s ) B(0,2s) B ( 0 , 2 s ) ), so the conditions k i ∈ S n − n / q i , r i ∩ H n − n / q i , r i k_i\in S_{n-n/q_i,r_i}\cap H_{n-n/q_i,r_i} k i ∈ S n − n / q i , r i ∩ H n − n / q i , r i mean that k i ∈ S n − α i , Ψ i ∩ H n − α i , Ψ i k_i\in S_{n-\alpha_i,\Psi_i}\cap H_{n-\alpha_i,\Psi_i} k i ∈ S n − α i , Ψ i ∩ H n − α i , Ψ i in the notation of [16 ] , Section 2. Since the constant c c c in the Hörmander condition may be enlarged, we may assume c > 1 c>1 c > 1 , as required in [16 ] . Moreover, Ψ 1 − 1 ( u ) ⋯ Ψ m − 1 ( u ) φ − 1 ( u ) = u \Psi_1^{-1}(u)\cdots\Psi_m^{-1}(u)\varphi^{-1}(u)=u Ψ 1 − 1 ( u ) ⋯ Ψ m − 1 ( u ) φ − 1 ( u ) = u , so φ \varphi φ is the complementary function of Ψ 1 , … , Ψ m \Psi_1,\ldots,\Psi_m Ψ 1 , … , Ψ m , and M α , φ = M α , t ~ M_{\alpha,\varphi}=M_{\alpha,\widetilde{t}} M α , φ = M α , t .
Let f ∈ L c ( R n ) f\in L^c(\mathbb{R}^n) f ∈ L c ( R n ) . We first check that T α f ∈ L q 0 T_\alpha f\in L^{q_0} T α f ∈ L q 0 . If m ≥ 2 m\geq2 m ≥ 2 , since p 0 > 1 p_0>1 p 0 > 1 we have q 0 > n n − α > n c − α q_0>\frac{n}{n-\alpha}>\frac{n}{c-\alpha} q 0 > n − α n > c − α n , so [16 ] , Lemma 5.1 with ν ≡ 1 \nu\equiv1 ν ≡ 1 gives T α f ∈ L q 0 T_\alpha f\in L^{q_0} T α f ∈ L q 0 . (Its proof shows that ∣ T α f ( x ) ∣ ≤ C ∣ x ∣ α − n |T_\alpha f(x)|\leq C|x|^{\alpha-n} ∣ T α f ( x ) ∣ ≤ C ∣ x ∣ α − n for ∣ x ∣ |x| ∣ x ∣ large, and this function is in L q 0 L^{q_0} L q 0 near infinity precisely because q 0 > n n − α q_0>\frac{n}{n-\alpha} q 0 > n − α n .) If m = 1 m=1 m = 1 , then T α f T_\alpha f T α f is bounded, since k 1 k_1 k 1 is integrable near the origin by the size condition. Moreover, if supp f ⊂ B ( 0 , ρ ) \operatorname{supp} f\subset B(0,\rho) supp f ⊂ B ( 0 , ρ ) and ∣ x ∣ > 2 ∥ A 1 ∥ ρ |x|>2\|A_1\|\rho ∣ x ∣ > 2∥ A 1 ∥ ρ , Hölder’s inequality and k 1 ∈ S n − n / q 1 , r 1 k_1\in S_{n-n/q_1,r_1} k 1 ∈ S n − n / q 1 , r 1 give ∣ T α f ( x ) ∣ ≤ C ∣ x ∣ α − n + n / r 1 |T_\alpha f(x)|\leq C|x|^{\alpha-n+n/r_1} ∣ T α f ( x ) ∣ ≤ C ∣ x ∣ α − n + n / r 1 , and this function is in L q 0 L^{q_0} L q 0 near infinity because 1 q 0 = 1 p 0 − α n < 1 t ~ − α n = 1 − α n − 1 r 1 \frac{1}{q_0}=\frac{1}{p_0}-\frac{\alpha}{n}<\frac{1}{\widetilde{t}}-\frac{\alpha}{n}=1-\frac{\alpha}{n}-\frac{1}{r_1} q 0 1 = p 0 1 − n α < t 1 − n α = 1 − n α − r 1 1 , since p 0 > t ~ = r 1 p_0>\widetilde{t}=r_1 p 0 > t = r 1 . Hence [16 ] , Theorem 3.3 with w ≡ 1 w\equiv1 w ≡ 1 and exponent q 0 q_0 q 0 yields
∥ T α f ∥ q 0 ≤ C ∥ M α , t ~ f ∥ q 0 = C ∥ M α t ~ ( ∣ f ∣ t ~ ) ∥ q 0 / t ~ 1 / t ~ ≤ C ∥ ∣ f ∣ t ~ ∥ p 0 / t ~ 1 / t ~ = C ∥ f ∥ p 0 , \|T_\alpha f\|_{q_0}\leq C\|M_{\alpha,\widetilde{t}}f\|_{q_0}=C\|M_{\alpha\widetilde{t}}(|f|^{\widetilde{t}})\|_{q_0/\widetilde{t}}^{1/\widetilde{t}}\leq C\||f|^{\widetilde{t}}\|_{p_0/\widetilde{t}}^{1/\widetilde{t}}=C\|f\|_{p_0}, ∥ T α f ∥ q 0 ≤ C ∥ M α , t f ∥ q 0 = C ∥ M α t ( ∣ f ∣ t ) ∥ q 0 / t 1/ t ≤ C ∥∣ f ∣ t ∥ p 0 / t 1/ t = C ∥ f ∥ p 0 ,
where we used [16 ] , Remark 2.2 and the boundedness of the fractional maximal operator M α t ~ M_{\alpha\widetilde{t}} M α t from L p 0 / t ~ L^{p_0/\widetilde{t}} L p 0 / t into L q 0 / t ~ L^{q_0/\widetilde{t}} L q 0 / t , which holds because 1 < p 0 t ~ < n α t ~ 1<\frac{p_0}{\widetilde{t}}<\frac{n}{\alpha\widetilde{t}} 1 < t p 0 < α t n and t ~ q 0 = t ~ p 0 − α t ~ n \frac{\widetilde{t}}{q_0}=\frac{\widetilde{t}}{p_0}-\frac{\alpha\widetilde{t}}{n} q 0 t = p 0 t − n α t . The result follows by density of L c ∞ L^\infty_c L c ∞ in L p 0 L^{p_0} L p 0 . □
Remark 4.5. Hypothesis (H) is not used in the proof of Lemma 4.2 , which holds for arbitrary invertible matrices A 1 , … , A m A_1,\ldots,A_m A 1 , … , A m ; in the proof of Theorem 1.1 it enters only through Lemma 4.4 .
Proof of Theorem 1.1
Let s 0 = a ( t , q ) ≥ q s_0=a(t,q)\ge q s 0 = a ( t , q ) ≥ q . By Lemma 2.1 (ii) applied to w s 0 w^{s_0} w s 0 , there is ε > 0 \varepsilon>0 ε > 0 such that w η ∈ A 1 w^\eta\in A_1 w η ∈ A 1 with η = s 0 ( 1 + ε ) > q \eta=s_0(1+\varepsilon)>q η = s 0 ( 1 + ε ) > q . In particular w max { s 0 , η } = w η ∈ A 1 w^{\max\{s_0,\eta\}}=w^\eta\in A_1 w m a x { s 0 , η } = w η ∈ A 1 .
Since η > q > p \eta>q>p η > q > p and t < n α t<\frac{n}{\alpha} t < α n , we can choose p 0 p_0 p 0 with
max { 1 , t , p η η − p , ( α n + 1 q − 1 η ) − 1 } < p 0 < n α . \max\left\{1,t,\frac{p\eta}{\eta-p},\left(\frac{\alpha}{n}+\frac{1}{q}-\frac{1}{\eta}\right)^{-1}\right\}<p_0<\frac{n}{\alpha}. max { 1 , t , η − p p η , ( n α + q 1 − η 1 ) − 1 } < p 0 < α n .
The interval is nonempty: 1 p − 1 η > 1 p − 1 q = α n \frac{1}{p}-\frac{1}{\eta}>\frac{1}{p}-\frac{1}{q}=\frac{\alpha}{n} p 1 − η 1 > p 1 − q 1 = n α and α n + 1 q − 1 η > α n \frac{\alpha}{n}+\frac{1}{q}-\frac{1}{\eta}>\frac{\alpha}{n} n α + q 1 − η 1 > n α . With 1 q 0 = 1 p 0 − α n \frac{1}{q_0}=\frac{1}{p_0}-\frac{\alpha}{n} q 0 1 = p 0 1 − n α , the last lower bound gives 1 q − 1 q 0 > 1 η \frac{1}{q}-\frac{1}{q_0}>\frac{1}{\eta} q 1 − q 0 1 > η 1 , that is, q 0 > q q_0>q q 0 > q and q ( q 0 q ) ′ < η q\left(\frac{q_0}{q}\right)^{\prime}<\eta q ( q q 0 ) ′ < η . Since t < p 0 < n α t<p_0<\frac{n}{\alpha} t < p 0 < α n , Lemma 4.4 gives the boundedness of T α : L p 0 → L q 0 T_\alpha:L^{p_0}\to L^{q_0} T α : L p 0 → L q 0 .
By Lemma 2.1 (iii), applied to v = w η v=w^\eta v = w η with θ = p / η \theta=p/\eta θ = p / η , w p ∈ R H η / p w^p\in RH_{\eta/p} w p ∈ R H η / p , so r w p ≥ η / p r_{w^p}\ge\eta/p r w p ≥ η / p and
η r w p r w p − 1 ≤ p η η − p < p 0 . \frac{\eta r_{w^p}}{r_{w^p}-1}\le\frac{p\eta}{\eta-p}<p_0. r w p − 1 η r w p ≤ η − p p η < p 0 .
Since w p ∈ A 1 w^p\in A_1 w p ∈ A 1 , we have q ~ w p = 1 \widetilde q_{w^p}=1 q w p = 1 , and δ < 1 \delta<1 δ < 1 gives ⌊ n ( 1 p − 1 ) ⌋ = 0 \lfloor n(\frac{1}{p}-1)\rfloor=0 ⌊ n ( p 1 − 1 )⌋ = 0 ; hence we may take d = 0 d=0 d = 0 . By Lemma 4.2 , ∥ T α a ∥ L w q q ≤ C \|T_\alpha a\|_{L^{q}_{w^q}}\le C ∥ T α a ∥ L w q q ≤ C for every w p w^p w p -( p , p 0 , 0 ) (p,p_0,0) ( p , p 0 , 0 ) atom a a a . Since q > 1 ≥ p q>1\ge p q > 1 ≥ p , Theorem 4.1 with v = w p v=w^p v = w p and u = w q u=w^q u = w q shows that T α T_\alpha T α extends to a bounded operator from H w p p ( R n ) H^p_{w^p}(\mathbb{R}^n) H w p p ( R n ) into L w q q ( R n ) L^q_{w^q}(\mathbb{R}^n) L w q q ( R n ) .
Remark 4.6. For a single kernel, the result closest to Theorem 1.1 is [15 Theorem 5.10] , which states the following: if $0<\alpha\le1$, n n + α ≤ p < 1 \frac{n}{n+\alpha}\le p<1 n + α n ≤ p < 1 , 1 q = 1 p − α n \frac{1}{q}=\frac{1}{p}-\frac{\alpha}{n} q 1 = p 1 − n α , s > n n − α s>\frac{n}{n-\alpha} s > n − α n , the kernel K α K_\alpha K α satisfies the L 1 L^1 L 1 size condition ∫ r < ∣ x ∣ ≤ 2 r ∣ K α ( x ) ∣ d x ≤ C r α \int_{r<|x|\le2r}|K_\alpha(x)|\,dx\le Cr^\alpha ∫ r < ∣ x ∣ ≤ 2 r ∣ K α ( x ) ∣ d x ≤ C r α for $r>0 and the fractional L s L^s L s -Hörmander condition of order α \alpha α , that is, K α ∈ H α , s ( α ) K_\alpha\in H^{(\alpha)}_{\alpha,s} K α ∈ H α , s ( α ) , and ω n s / ( n s − n − s α ) ∈ A 1 \omega^{ns/(ns-n-s\alpha)}\in A_1 ω n s / ( n s − n − s α ) ∈ A 1 , then f ↦ ∫ K α ( x − y ) f ( y ) d y f\mapsto\int K_\alpha(x-y)f(y)\,dy f ↦ ∫ K α ( x − y ) f ( y ) d y is bounded from H ω p p ( R n ) H^p_{\omega^p}(\mathbb{R}^n) H ω p p ( R n ) into L ω q q ( R n ) L^q_{\omega^q}(\mathbb{R}^n) L ω q q ( R n ) .
Theorem 1.1 with m = 1 m=1 m = 1 , A 1 = I A_1=I A 1 = I , k 1 = K α k_1=K_\alpha k 1 = K α , q 1 = n n − α q_1=\frac{n}{n-\alpha} q 1 = n − α n and p 1 = s p_1=s p 1 = s applies to the same operators. In this case, (H1 _1 1 ) and (H2 _2 2 ) mean that K α ∈ S α , s ∩ H α , s ( δ ) K_\alpha\in S_{\alpha,s}\cap H^{(\delta)}_{\alpha,s} K α ∈ S α , s ∩ H α , s ( δ ) , and (1.3) gives a ( t , q ) = s q s − q a(t,q)=\frac{sq}{s-q} a ( t , q ) = s − q s q . The main differences between both results are the following.
Operators. Theorem 1.1 covers m ≥ 2 m\ge2 m ≥ 2 kernels composed with matrices A i A_i A i , which are not considered in [15 ] . Even for m = 1 m=1 m = 1 , it allows a matrix A 1 ≠ I A_1\ne I A 1 = I and the case s = ∞ s=\infty s = ∞ .
Smoothness. Theorem 1.1 requires the Hörmander condition of order δ = n ( 1 p − 1 ) \delta=n(\frac{1}{p}-1) δ = n ( p 1 − 1 ) instead of α \alpha α . Since δ < α \delta<\alpha δ < α when p > n n + α p>\frac{n}{n+\alpha} p > n + α n , this is a weaker requirement and, for n ≥ 2 n\ge2 n ≥ 2 , it is strictly weaker by Proposition 3.7 .
Range of α \alpha α . In Theorem 1.1, α \alpha α may be any number in ( 0 , n ) (0,n) ( 0 , n ) , since only δ < 1 \delta<1 δ < 1 is needed, while [15 Theorem 5.10] requires 0 < α ≤ 1 0<\alpha\le1 0 < α ≤ 1 .
Weights. The condition on the weight in Theorem 1.1 is weaker. Indeed, with a = s q s − q a=\frac{sq}{s-q} a = s − q s q ,
1 a = 1 p − α n − 1 s > 1 − α n − 1 s = n s − n − s α n s . \frac{1}{a}=\frac{1}{p}-\frac{\alpha}{n}-\frac{1}{s}>1-\frac{\alpha}{n}-\frac{1}{s}=\frac{ns-n-s\alpha}{ns}. a 1 = p 1 − n α − s 1 > 1 − n α − s 1 = n s n s − n − s α .
so that a < n s n s − n − s α a < \frac{ns}{ns-n-s\alpha} a < n s − n − s α n s and, by Lemma 2.1 (i), ω n s / ( n s − n − s α ) ∈ A 1 \omega^{ns/(ns-n-s\alpha)} \in\mathcal{A}_1 ω n s / ( n s − n − s α ) ∈ A 1 implies ω a ∈ A 1 \omega^a \in\mathcal{A}_1 ω a ∈ A 1 . This improvement comes from the self-improvement property of A 1 \mathcal{A}_1 A 1 weights (Lemma 2.1 (ii)) used in the proof of Theorem 1.1 .
Range of p p p and size condition. On the other hand, [15 Theorem 5.10] includes the endpoint p = n n + α p = \frac{n}{n+\alpha} p = n + α n and only requires an L 1 L^1 L 1 size condition, while Theorem 1.1 requires p > n n + α p > \frac{n}{n+\alpha} p > n + α n and the L s L^s L s size condition K α ∈ S α , s K_\alpha\in S_{\alpha,s} K α ∈ S α , s .
Sharpness of (H2)
We now show that (H2 _2 2 ) cannot be replaced by the classical fractional Hörmander condition.
Remark 4.7 (The factor 2 j δ 2^{j\delta} 2 j δ in (H2 _2 2 ) cannot be removed). Let n ≥ 2 n \ge2 n ≥ 2 , 0 < α < n 0 < \alpha< n 0 < α < n , p p p as in Theorem 1.1 , and assume δ > 1 / q \delta> 1/q δ > 1/ q . For instance, n = 2 n = 2 n = 2 , α = 9 / 5 \alpha= 9/5 α = 9/5 and p = 10 / 13 p = 10/13 p = 10/13 give q = 5 / 2 q = 5/2 q = 5/2 , δ = 3 / 5 \delta= 3/5 δ = 3/5 and δ − 1 / q = 1 / 5 \delta- 1/q = 1/5 δ − 1/ q = 1/5 . Fix 0 < ε ≤ δ − 1 / q 0 < \varepsilon\le\delta- 1/q 0 < ε ≤ δ − 1/ q , let m = 2 m = 2 m = 2 , q 1 = q 2 = 2 n / ( n − α ) q_1 = q_2 = 2n/(n-\alpha) q 1 = q 2 = 2 n / ( n − α ) , so that (1.1) holds, A 1 = I A_1 = I A 1 = I , A 2 = − I A_2 = -I A 2 = − I , w ≡ 1 w \equiv1 w ≡ 1 , and
k 1 ( x ) = ∣ x n ∣ ε ∣ x ∣ n / q 1 + ε = Ω 1 ( x ) ∣ x ∣ n / q 1 , Ω 1 ( x ) = ∣ x n ∣ x ∣ ∣ ε , k 2 ( x ) = 1 ∣ x ∣ n / q 2 . k_1(x) = \frac{|x_n|^\varepsilon}{|x|^{n/q_1+\varepsilon}} = \frac{\Omega_1(x)}{|x|^{n/q_1}}, \qquad\Omega_1(x) = \left|\frac{x_n}{|x|}\right|^\varepsilon, \qquad k_2(x) = \frac{1}{|x|^{n/q_2}}. k 1 ( x ) = ∣ x ∣ n / q 1 + ε ∣ x n ∣ ε = ∣ x ∣ n / q 1 Ω 1 ( x ) , Ω 1 ( x ) = ∣ x ∣ x n ε , k 2 ( x ) = ∣ x ∣ n / q 2 1 .
The hypotheses hold. (H) holds, since A 1 A_1 A 1 , A 2 A_2 A 2 and A 1 − A 2 = 2 I A_1-A_2 = 2I A 1 − A 2 = 2 I are invertible. Since Ω 1 , Ω 2 ∈ L ∞ ( S n − 1 ) \Omega_1,\Omega_2 \in L^\infty(\mathbb{S}^{n-1}) Ω 1 , Ω 2 ∈ L ∞ ( S n − 1 ) , Example 3.4 gives (H1 _1 1 ) with p 1 = p 2 = ∞ p_1 = p_2 = \infty p 1 = p 2 = ∞ ; then t = 1 t = 1 t = 1 and a ( t , q ) = q a(t,q) = q a ( t , q ) = q . Clearly w q ≡ 1 ∈ A 1 w^q \equiv1 \in\mathcal{A}_1 w q ≡ 1 ∈ A 1 and w ( A i x ) = w ( x ) w(A_i x) = w(x) w ( A i x ) = w ( x ) . Since ∣ ∣ s ∣ ε − ∣ s ′ ∣ ε ∣ ≤ ∣ s − s ′ ∣ ε \left||s|^\varepsilon- |s'|^\varepsilon\right| \le|s-s'|^\varepsilon ∣ ∣ s ∣ ε − ∣ s ′ ∣ ε ∣ ≤ ∣ s − s ′ ∣ ε , Ω 1 \Omega_1 Ω 1 is Hölder continuous of order ε \varepsilon ε on S n − 1 \mathbb{S}^{n-1} S n − 1 , and a standard computation gives, for ∣ x ∣ ≥ 2 ∣ h ∣ |x| \ge2|h| ∣ x ∣ ≥ 2∣ h ∣ ,
∣ k 1 ( x − h ) − k 1 ( x ) ∣ ≤ C ( ∣ h ∣ ∣ x ∣ ) ε ∣ x ∣ − n / q 1 , ∣ k 2 ( x − h ) − k 2 ( x ) ∣ ≤ C ∣ h ∣ ∣ x ∣ ∣ x ∣ − n / q 2 . |k_1(x-h)-k_1(x)| \le C\left(\frac{|h|}{|x|}\right)^\varepsilon|x|^{-n/q_1}, \qquad|k_2(x-h)-k_2(x)| \le C\frac{|h|}{|x|}|x|^{-n/q_2}. ∣ k 1 ( x − h ) − k 1 ( x ) ∣ ≤ C ( ∣ x ∣ ∣ h ∣ ) ε ∣ x ∣ − n / q 1 , ∣ k 2 ( x − h ) − k 2 ( x ) ∣ ≤ C ∣ x ∣ ∣ h ∣ ∣ x ∣ − n / q 2 .
Consequently, for 0 ≤ δ ′ < ε 0 \le\delta' < \varepsilon 0 ≤ δ ′ < ε , c i = 2 c_i = 2 c i = 2 and R > 2 ∣ h ∣ R > 2|h| R > 2∣ h ∣ ,
∑ j = 1 ∞ 2 j δ ′ ( 2 j R ) n / q i ∥ k i ( ⋅ − h ) − k i ∥ ∞ , ∣ x ∣ ∼ 2 j R ≤ C ( ∣ h ∣ R ) ε ∑ j = 1 ∞ 2 j ( δ ′ − ε ) ≤ C , i = 1 , 2. \sum_{j=1}^{\infty} 2^{j\delta'}(2^jR)^{n/q_i}\|k_i(\cdot-h)-k_i\|_{\infty,|x|\sim2^jR} \le C\left(\frac{|h|}{R}\right)^\varepsilon\sum_{j=1}^{\infty}2^{j(\delta'-\varepsilon)} \le C, \qquad i=1,2. j = 1 ∑ ∞ 2 j δ ′ ( 2 j R ) n / q i ∥ k i ( ⋅ − h ) − k i ∥ ∞ , ∣ x ∣ ∼ 2 j R ≤ C ( R ∣ h ∣ ) ε j = 1 ∑ ∞ 2 j ( δ ′ − ε ) ≤ C , i = 1 , 2.
In particular k i ∈ H n − n / q i , ∞ k_i \in H_{n-n/q_i,\infty} k i ∈ H n − n / q i , ∞ , and in fact k i ∈ H n − n / q i , ∞ ( δ ′ ) k_i \in H^{(\delta')}_{n-n/q_i,\infty} k i ∈ H n − n / q i , ∞ ( δ ′ ) for every δ ′ < ε \delta' < \varepsilon δ ′ < ε . Moreover T α T_\alpha T α is bounded from L p 0 L^{p_0} L p 0 into L q 0 L^{q_0} L q 0 for every 1 < p 0 < n / α 1 < p_0 < n/\alpha 1 < p 0 < n / α , 1 q 0 = 1 p 0 − α n \frac{1}{q_0} = \frac{1}{p_0} - \frac{\alpha}{n} q 0 1 = p 0 1 − n α . This follows from Lemma 4.4 , since t = 1 t = 1 t = 1 , and also from [12 ] , since ∣ k 1 ( x ) ∣ ≤ ∣ x ∣ − n / q 1 |k_1(x)| \le|x|^{-n/q_1} ∣ k 1 ( x ) ∣ ≤ ∣ x ∣ − n / q 1 gives
∣ T α f ( x ) ∣ ≤ ∫ ∣ x − y ∣ − n / q 1 ∣ x + y ∣ − n / q 2 ∣ f ( y ) ∣ d y . |T_\alpha f(x)| \le\int|x-y|^{-n/q_1}|x+y|^{-n/q_2}|f(y)|\,dy. ∣ T α f ( x ) ∣ ≤ ∫ ∣ x − y ∣ − n / q 1 ∣ x + y ∣ − n / q 2 ∣ f ( y ) ∣ d y .
An atom whose image is not in L q L^q L q . Write z = ( z ~ , z n ) ∈ R n − 1 × R z = (\widetilde{z},z_n) \in\mathbb{R}^{n-1} \times\mathbb{R} z = ( z , z n ) ∈ R n − 1 × R . Let φ ∈ C ∞ ( R n − 1 ) \varphi\in C^\infty(\mathbb{R}^{n-1}) φ ∈ C ∞ ( R n − 1 ) with φ ≥ 0 \varphi\ge0 φ ≥ 0 , supp φ ⊂ { ∣ z ~ ∣ < 1 / 2 } \operatorname{supp}\varphi\subset\{|\widetilde{z}| < 1/2\} supp φ ⊂ { ∣ z ∣ < 1/2 } and ∫ φ = 1 \int\varphi= 1 ∫ φ = 1 , let ψ = χ [ 0 , 1 / 2 ] − χ [ − 1 / 2 , 0 ) \psi= \chi_{[0,1/2]} - \chi_{[-1/2,0)} ψ = χ [ 0 , 1/2 ] − χ [ − 1/2 , 0 ) , and put a ( z ) = φ ( z ~ ) ψ ( z n ) a(z) = \varphi(\widetilde{z})\psi(z_n) a ( z ) = φ ( z ) ψ ( z n ) . Then a ∈ L ∞ a \in L^\infty a ∈ L ∞ , supp a ⊂ B ( 0 , 1 ) \operatorname{supp}a \subset B(0,1) supp a ⊂ B ( 0 , 1 ) and ∫ a = 0 \int a = 0 ∫ a = 0 . Since w ≡ 1 w \equiv1 w ≡ 1 and δ < 1 \delta< 1 δ < 1 , a suitable multiple of a a a is a ( p , p 0 , 0 ) (p,p_0,0) ( p , p 0 , 0 ) -atom. Define
G ( u ) = ∫ R ∣ u − s ∣ ε ψ ( s ) d s , u ∈ R . G(u) = \int_{\mathbb{R}} |u-s|^\varepsilon\psi(s)\,ds, \qquad u \in\mathbb{R}. G ( u ) = ∫ R ∣ u − s ∣ ε ψ ( s ) d s , u ∈ R .
The function G G G is continuous and
G ( 1 / 2 ) = ∫ 0 1 / 2 τ ε d τ − ∫ 1 / 2 1 τ ε d τ < 0 , G(1/2) = \int_0^{1/2}\tau^\varepsilon\,d\tau- \int_{1/2}^1\tau^\varepsilon\,d\tau< 0, G ( 1/2 ) = ∫ 0 1/2 τ ε d τ − ∫ 1/2 1 τ ε d τ < 0 ,
so there exist c > 0 c > 0 c > 0 and an interval I ⊂ ( 0 , 1 ) I \subset(0,1) I ⊂ ( 0 , 1 ) such that ∣ G ∣ ≥ c |G| \ge c ∣ G ∣ ≥ c on I I I . Let y = ( y ~ , u ) y = (\widetilde{y},u) y = ( y , u ) with u ∈ I u \in I u ∈ I and ∣ y ~ ∣ = ρ |\widetilde{y}| = \rho ∣ y ∣ = ρ . For z ∈ supp a z \in\operatorname{supp}a z ∈ supp a , uniformly,
∣ y − z ∣ − n / q 1 − ε ∣ y + z ∣ − n / q 2 = ρ α − n − ε ( 1 + O ( ρ − 1 ) ) , |y-z|^{-n/q_1-\varepsilon}|y+z|^{-n/q_2} = \rho^{\alpha-n-\varepsilon}\left(1+O(\rho^{-1})\right), ∣ y − z ∣ − n / q 1 − ε ∣ y + z ∣ − n / q 2 = ρ α − n − ε ( 1 + O ( ρ − 1 ) ) ,
by (1) . Since ∫ φ = 1 \int\varphi=1 ∫ φ = 1 ,
T α a ( y ) = ∫ ∣ u − z n ∣ ε ∣ y − z ∣ − n / q 1 − ε ∣ y + z ∣ − n / q 2 a ( z ) d z = ρ α − n − ε ( G ( u ) + O ( ρ − 1 ) ) . T_{\alpha a}(y)=\int|u-z_n|^{\varepsilon}|y-z|^{-n/q_1-\varepsilon}|y+z|^{-n/q_2}a(z)\,dz=\rho^{\alpha-n-\varepsilon}\left(G(u)+O(\rho^{-1})\right). T α a ( y ) = ∫ ∣ u − z n ∣ ε ∣ y − z ∣ − n / q 1 − ε ∣ y + z ∣ − n / q 2 a ( z ) d z = ρ α − n − ε ( G ( u ) + O ( ρ − 1 ) ) .
Hence there is ρ 1 > 0 \rho_1>0 ρ 1 > 0 such that ∣ T α a ( y ) ∣ ≥ ε 2 ρ α − n − ε |T_{\alpha a}(y)|\geq\frac{\varepsilon}{2}\rho^{\alpha-n-\varepsilon} ∣ T α a ( y ) ∣ ≥ 2 ε ρ α − n − ε whenever ∣ y ~ ∣ = ρ ≥ ρ 1 |\widetilde y|=\rho\geq\rho_1 ∣ y ∣ = ρ ≥ ρ 1 and u ∈ I u\in I u ∈ I . Integrating in polar coordinates in y ~ \widetilde y y (with counting measure on S 0 S^0 S 0 when n = 2 ) n=2) n = 2 ) ,
∫ R n ∣ T α a ∣ q ≥ ( ε 2 ) q ∣ I ∣ ∣ S n − 2 ∣ ∫ ρ 1 ∞ ρ ( α − n − ε ) q + n − 2 d ρ = ∞ , \int_{\mathbb{R}^n}|T_{\alpha a}|^q\geq\left(\frac{\varepsilon}{2}\right)^q|I||S^{n-2}|\int_{\rho_1}^{\infty}\rho^{(\alpha-n-\varepsilon)q+n-2}\,d\rho=\infty, ∫ R n ∣ T α a ∣ q ≥ ( 2 ε ) q ∣ I ∣∣ S n − 2 ∣ ∫ ρ 1 ∞ ρ ( α − n − ε ) q + n − 2 d ρ = ∞ ,
because ( α − n − ε ) q + n − 2 ≥ − 1 (\alpha-n-\varepsilon)q+n-2\geq-1 ( α − n − ε ) q + n − 2 ≥ − 1 is equivalent to ε ≤ α − n + n − 1 q = δ − 1 q \varepsilon\leq\alpha-n+\frac{n-1}{q}=\delta-\frac{1}{q} ε ≤ α − n + q n − 1 = δ − q 1 .
Therefore T α a ∉ L q T_{\alpha a}\notin L^q T α a ∈ / L q . Hence the conclusion of Lemma 4.2 fails and, by Theorem 4.1 with v = u ≡ 1 v=u\equiv1 v = u ≡ 1 , T α T_\alpha T α does not extend to a bounded operator from H p ( R n ) H^p(\mathbb{R}^n) H p ( R n ) into L q ( R n ) L^q(\mathbb{R}^n) L q ( R n ) , so the conclusion of Theorem 1.1 fails as well, although the kernels satisfy ( H 1 ) (\mathrm{H}_1) ( H 1 ) , the classical Hörmander condition, and even k i ∈ H n − n / q i , ∞ ( δ ′ ) k_i\in H^{(\delta')}_{n-n/q_i,\infty} k i ∈ H n − n / q i , ∞ ( δ ′ ) for every δ ′ < ε \delta'<\varepsilon δ ′ < ε . In particular, δ \delta δ cannot be replaced in ( H 2 ) (\mathrm{H}_2) ( H 2 ) by any δ ′ < δ − 1 / q \delta'<\delta-1/q δ ′ < δ − 1/ q .
The same example gives more information when ( H 1 ) (\mathrm{H}_1) ( H 1 ) and ( H 2 ) (\mathrm{H}_2) ( H 2 ) are considered with a finite exponent for k 1 k_1 k 1 . Let q 1 < p 1 < ∞ q_1<p_1<\infty q 1 < p 1 < ∞ and p 2 = ∞ p_2=\infty p 2 = ∞ . Since Ω 1 \Omega_1 Ω 1 is bounded, ( H 1 ) (\mathrm{H}_1) ( H 1 ) holds with these exponents. Moreover, since 0 < ε ≤ δ − 1 q < 1 − 1 p 1 0<\varepsilon\leq\delta-\frac{1}{q}<1-\frac{1}{p_1} 0 < ε ≤ δ − q 1 < 1 − p 1 1 , Proposition 3.7 (a), with β = n − n / q 1 \beta=n-n/q_1 β = n − n / q 1 and τ = p 1 \tau=p_1 τ = p 1 , gives k 1 ∈ H n − n / q 1 , p 1 ( δ ′ ) k_1\in H^{(\delta')}_{n-n/q_1,p_1} k 1 ∈ H n − n / q 1 , p 1 ( δ ′ ) for every δ ′ < ε + 1 p 1 \delta'<\varepsilon+\frac{1}{p_1} δ ′ < ε + p 1 1 , while k 2 ∈ H n − n / q 2 , ∞ ( δ ′ ) k_2\in H^{(\delta')}_{n-n/q_2,\infty} k 2 ∈ H n − n / q 2 , ∞ ( δ ′ ) for every δ ′ < 1 \delta'<1 δ ′ < 1 . Taking ε = δ − 1 q \varepsilon=\delta-\frac{1}{q} ε = δ − q 1 , we conclude that, for this choice of exponents, δ \delta δ cannot be replaced in ( H 2 ) (\mathrm{H}_2) ( H 2 ) by any δ ′ < δ − 1 q + 1 p 1 \delta'<\delta-\frac{1}{q}+\frac{1}{p_1} δ ′ < δ − q 1 + p 1 1 . This is consistent with Theorem 1.1 , since 1 p 1 < 1 q 1 < 1 − α n < 1 q \frac{1}{p_1}<\frac{1}{q_1}<1-\frac{\alpha}{n}<\frac{1}{q} p 1 1 < q 1 1 < 1 − n α < q 1 , so that δ − 1 q + 1 p 1 < δ \delta-\frac{1}{q}+\frac{1}{p_1}<\delta δ − q 1 + p 1 1 < δ .
Examples of matrices and weights
In this section we give examples of matrices and weights satisfying the hypotheses of Theorem 1.1 , that is, ( H ) (\mathrm{H}) ( H ) for the matrices and
w a ( t , q ) ∈ A 1 , w ( A i x ) ≤ C w ( x ) a.e. x ∈ R n , 1 ≤ i ≤ m , (18) w^{a(t,q)}\in A_1,\qquad w(A_i x)\leq Cw(x)\quad\text{a.e. }x\in\mathbb{R}^n,\quad1\leq i\leq m, \tag*{(18)} w a ( t , q ) ∈ A 1 , w ( A i x ) ≤ C w ( x ) a.e. x ∈ R n , 1 ≤ i ≤ m , (18)
for the weight. We also discuss the choice of the exponents p i p_i p i and give a concrete family of operators.
Matrices
Example 5.1. The following families of matrices satisfy (H).
(a) A i = a i A A_i=a_iA A i = a i A , where A A A is an invertible matrix and a 1 , … , a m a_1,\ldots,a_m a 1 , … , a m are distinct nonzero real numbers, since A i − A j = ( a i − a j ) A A_i-A_j=(a_i-a_j)A A i − A j = ( a i − a j ) A . In particular, A i = a i I A_i=a_iI A i = a i I , which is the setting of [12 ] .
(b) Diagonal matrices A i = diag ( d i , 1 , … , d i , n ) A_i=\operatorname{diag}(d_{i,1},\ldots,d_{i,n}) A i = diag ( d i , 1 , … , d i , n ) with d i , l ≠ 0 d_{i,l}\ne0 d i , l = 0 and d i , l ≠ d j , l d_{i,l}\ne d_{j,l} d i , l = d j , l for i ≠ j i\ne j i = j and 1 ≤ l ≤ n 1\leq l\leq n 1 ≤ l ≤ n .
(c) In R 2 \mathbb{R}^2 R 2 , identified with C \mathbb{C} C , the matrices of the maps z ↦ c i z z\mapsto c_i z z ↦ c i z , where c 1 , … , c m ∈ C c_1,\ldots,c_m\in\mathbb{C} c 1 , … , c m ∈ C are distinct and nonzero. Indeed, A i − A j A_i-A_j A i − A j is the matrix of z ↦ ( c i − c j ) z z\mapsto(c_i-c_j)z z ↦ ( c i − c j ) z , whose determinant is ∣ c i − c j ∣ 2 |c_i-c_j|^2 ∣ c i − c j ∣ 2 . For instance, the rotations A i = R θ i A_i=R_{\theta_i} A i = R θ i with angles θ 1 , … , θ m \theta_1,\ldots,\theta_m θ 1 , … , θ m distinct modulo 2 π 2\pi 2 π .
Remark 5.2. In odd dimensions rotations behave differently: if n n n is odd and Q , Q ′ ∈ SO ( n ) Q,Q' \in\operatorname{SO}(n) Q , Q ′ ∈ SO ( n ) , then Q − Q ′ Q-Q' Q − Q ′ is not invertible, so two rotations never satisfy (H). Indeed, P = Q ′ − 1 Q ∈ SO ( n ) P=Q'^{-1}Q \in\operatorname{SO}(n) P = Q ′ − 1 Q ∈ SO ( n ) and, since P − I = P ( I − P T ) P-I=P(I-P^{\mathsf{T}}) P − I = P ( I − P T ) ,
det ( P − I ) = det ( I − P T ) = det ( I − P ) = ( − 1 ) n det ( P − I ) = − det ( P − I ) , \det(P-I)=\det(I-P^{\mathsf{T}})=\det(I-P)=(-1)^n\det(P-I)=-\det(P-I), det ( P − I ) = det ( I − P T ) = det ( I − P ) = ( − 1 ) n det ( P − I ) = − det ( P − I ) ,
so det ( P − I ) = 0 \det(P-I)=0 det ( P − I ) = 0 and Q − Q ′ = Q ′ ( P − I ) Q-Q'=Q'(P-I) Q − Q ′ = Q ′ ( P − I ) is singular. On the other hand, A 1 = I A_1=I A 1 = I and A 2 = − I A_2=-I A 2 = − I satisfy (H) in every dimension.
Weights
We first observe that A 1 A_1 A 1 is invariant under linear changes of variables.
Lemma 5.3. Let v ∈ A 1 v \in A_1 v ∈ A 1 and let A A A be an invertible matrix. Then v ∘ A ∈ A 1 v \circ A \in A_1 v ∘ A ∈ A 1 .
Proof. Let B = B ( x B , r ) B=B(x_B,r) B = B ( x B , r ) be a ball. The ellipsoid A ( B ) A(B) A ( B ) is contained in the ball B ′ = B ( A x B , ∥ A ∥ r ) B'=B(Ax_B,\lVert A\rVert r) B ′ = B ( A x B , ∥ A ∥ r ) , and ∣ B ′ ∣ / ∣ A ( B ) ∣ = ∥ A ∥ n / ∣ det A ∣ ≤ ( ∥ A ∥ ∥ A − 1 ∥ ) n |B'|/|A(B)|=\lVert A\rVert^n/|\det A|\leq(\lVert A\rVert\lVert A^{-1}\rVert)^n ∣ B ′ ∣/∣ A ( B ) ∣ = ∥ A ∥ n /∣ det A ∣ ≤ (∥ A ∥ ∥ A − 1 ∥ ) n . Hence
1 ∣ B ∣ ∫ B v ( A x ) d x = 1 ∣ A ( B ) ∣ ∫ A ( B ) v ≤ ( ∥ A ∥ ∥ A − 1 ∥ ) n ∣ B ′ ∣ ∫ B ′ v ≤ C ess inf B ′ v ≤ C ess inf A ( B ) v = C ess inf B ( v ∘ A ) . \frac{1}{|B|}\int_B v(Ax)\,dx=\frac{1}{|A(B)|}\int_{A(B)}v\leq\frac{(\lVert A\rVert\lVert A^{-1}\rVert)^n}{|B'|}\int_{B'}v
\leq C\mathop{\operatorname{ess\,inf}}_{B'}v\leq C\mathop{\operatorname{ess\,inf}}_{A(B)}v=C\mathop{\operatorname{ess\,inf}}_B(v\circ A). ∣ B ∣ 1 ∫ B v ( A x ) d x = ∣ A ( B ) ∣ 1 ∫ A ( B ) v ≤ ∣ B ′ ∣ (∥ A ∥ ∥ A − 1 ∥ ) n ∫ B ′ v ≤ C ess inf B ′ v ≤ C ess inf A ( B ) v = C ess inf B ( v ∘ A ) .
:::
Proposition 5.4. Let φ : ( 0 , ∞ ) → ( 0 , ∞ ) \varphi:(0,\infty)\to(0,\infty) φ : ( 0 , ∞ ) → ( 0 , ∞ ) be nonincreasing and such that φ ( r / 2 ) ≤ C φ ( r ) \varphi(r/2)\leq C\varphi(r) φ ( r /2 ) ≤ C φ ( r ) for every r > 0 r>0 r > 0 , and let w ( x ) = φ ( ∣ x ∣ ) w(x)=\varphi(|x|) w ( x ) = φ ( ∣ x ∣ ) . Then, for every invertible matrix A A A , there is C A > 0 C_A>0 C A > 0 such that w ( A x ) ≤ C A w ( x ) w(Ax)\leq C_Aw(x) w ( A x ) ≤ C A w ( x ) for all x ≠ 0 x\ne0 x = 0 .
Proof. Since ∣ A x ∣ ≥ ∣ x ∣ / ∥ A − 1 ∥ |Ax|\geq|x|/\lVert A^{-1}\rVert ∣ A x ∣ ≥ ∣ x ∣/ ∥ A − 1 ∥ and φ \varphi φ is nonincreasing, w ( A x ) ≤ φ ( ∣ x ∣ / ∥ A − 1 ∥ ) w(Ax)\leq\varphi(|x|/\lVert A^{-1}\rVert) w ( A x ) ≤ φ ( ∣ x ∣/ ∥ A − 1 ∥) . If ∥ A − 1 ∥ ≤ 1 \lVert A^{-1}\rVert\leq1 ∥ A − 1 ∥ ≤ 1 , this is at most φ ( ∣ x ∣ ) = w ( x ) \varphi(|x|)=w(x) φ ( ∣ x ∣ ) = w ( x ) . Otherwise, let k ∈ N k\in\mathbb{N} k ∈ N with 2 k ≥ ∥ A − 1 ∥ 2^k\geq\lVert A^{-1}\rVert 2 k ≥ ∥ A − 1 ∥ ; then w ( A x ) ≤ φ ( 2 − k ∣ x ∣ ) ≤ C k w ( x ) w(Ax)\leq\varphi(2^{-k}|x|)\leq C^kw(x) w ( A x ) ≤ φ ( 2 − k ∣ x ∣ ) ≤ C k w ( x ) . □
Example 5.5. Let a = a ( t , q ) a=a(t,q) a = a ( t , q ) and − n < γ ≤ 0 -n<\gamma\leq0 − n < γ ≤ 0 . The weights
w 1 ( x ) = ∣ x ∣ γ / a , w 2 ( x ) = ( 1 + ∣ x ∣ ) γ / a w_1(x)=|x|^{\gamma/a},\qquad w_2(x)=(1+|x|)^{\gamma/a} w 1 ( x ) = ∣ x ∣ γ / a , w 2 ( x ) = ( 1 + ∣ x ∣ ) γ / a satisfy (18) for arbitrary invertible matrices A 1 , … , A m A_1,\ldots,A_m A 1 , … , A m . Indeed, w 1 a = ∣ x ∣ γ ∈ A 1 w_1^a=|x|^\gamma\in A_1 w 1 a = ∣ x ∣ γ ∈ A 1 , and w 2 a = ( 1 + ∣ x ∣ ) γ w_2^a=(1+|x|)^\gamma w 2 a = ( 1 + ∣ x ∣ ) γ is comparable to min { 1 , ∣ x ∣ γ } \min\{1,|x|^\gamma\} min { 1 , ∣ x ∣ γ } , which belongs to A 1 A_1 A 1 because M ( min { u , v } ) ≤ min { M u , M v } M(\min\{u,v\})\leq\min\{Mu,Mv\} M ( min { u , v }) ≤ min { M u , M v } . Moreover, both weights are of the form φ ( ∣ x ∣ ) \varphi(|x|) φ ( ∣ x ∣ ) with φ \varphi φ nonincreasing and φ ( r / 2 ) ≤ 2 − γ / a φ ( r ) \varphi(r/2)\leq2^{-\gamma/a}\varphi(r) φ ( r /2 ) ≤ 2 − γ / a φ ( r ) , so the second condition in (18) follows from Proposition 5.4. If the matrices A i A_i A i are orthogonal, every radial weight w w w with w a ∈ A 1 w^a\in A_1 w a ∈ A 1 satisfies (18) , with w ( A i x ) = w ( x ) w(A_ix)=w(x) w ( A i x ) = w ( x ) .
Non-radial examples can be obtained by symmetrization when the matrices generate a finite group.
Proposition 5.6. Let G G G be a finite group of invertible matrices containing A 1 , … , A m A_1,\ldots,A_m A 1 , … , A m , let v ∈ A 1 v\in A_1 v ∈ A 1 , and set
w = ( ∑ g ∈ G v ∘ g ) 1 / a ( t , q ) . w=\left(\sum_{g\in G}v\circ g\right)^{1/a(t,q)}. w = g ∈ G ∑ v ∘ g 1/ a ( t , q ) . Then w w w satisfies (18) , and in fact w ( A i x ) = w ( x ) w(A_ix)=w(x) w ( A i x ) = w ( x ) for a.e. x x x and $1\leq i\leq m.
Proof. By Lemma 5.3, v ∘ g ∈ A 1 v\circ g\in A_1 v ∘ g ∈ A 1 for every g ∈ G g\in G g ∈ G , and a finite sum of A 1 A_1 A 1 weights is an A 1 A_1 A 1 weight, since M ( u 1 + u 2 ) ≤ M u 1 + M u 2 M(u_1+u_2)\leq Mu_1+Mu_2 M ( u 1 + u 2 ) ≤ M u 1 + M u 2 . Hence w a ( t , q ) ∈ A 1 w^{a(t,q)}\in A_1 w a ( t , q ) ∈ A 1 ; since a ( t , q ) ≥ 1 a(t,q)\geq1 a ( t , q ) ≥ 1 , w w w is locally integrable. Moreover, for each i i i the map g ↦ g A i g\mapsto gA_i g ↦ g A i is a bijection of G G G , so
∑ g ∈ G v ( g A i x ) = ∑ g ∈ G v ( g x ) , \sum_{g\in G}v(gA_ix)=\sum_{g\in G}v(gx), g ∈ G ∑ v ( g A i x ) = g ∈ G ∑ v ( g x ) ,
that is, w ( A i x ) = w ( x ) w(A_ix)=w(x) w ( A i x ) = w ( x ) . □
Example 5.7. Let A 1 = I A_1=I A 1 = I and A 2 = − I A_2=-I A 2 = − I , so that G = { I , − I } G=\{I,-I\} G = { I , − I } , and let x 0 ∈ R n ∖ { 0 } x_0\in\mathbb{R}^n\setminus\{0\} x 0 ∈ R n ∖ { 0 } and − n < γ < 0 -n<\gamma<0 − n < γ < 0 . By Proposition 5.6 with v ( x ) = ∣ x − x 0 ∣ γ v(x)=|x-x_0|^\gamma v ( x ) = ∣ x − x 0 ∣ γ , the weight
w ( x ) = ( ∣ x − x 0 ∣ γ + ∣ x + x 0 ∣ γ ) 1 / a ( t , q ) , w(x)=\left(|x-x_0|^\gamma+|x+x_0|^\gamma\right)^{1/a(t,q)}, w ( x ) = ( ∣ x − x 0 ∣ γ + ∣ x + x 0 ∣ γ ) 1/ a ( t , q ) ,
which has two singularities, satisfies (18) . Similarly, in R 2 ≃ C \mathbb{R}^2\simeq\mathbb{C} R 2 ≃ C , if A i z = c i z A_i z=c_i z A i z = c i z with c 1 , … , c m c_1,\ldots,c_m c 1 , … , c m distinct k k k -th roots of unity and G G G is the group of the k k k -th roots of unity, one obtains weights with k k k singularities at the vertices of a regular polygon.
Remark 5.8. For power weights with a singularity away from the origin the symmetrization cannot be avoided. Indeed, let m ≥ 2 m\geq2 m ≥ 2 , assume (H), and let w ( x ) = ∣ x − x 0 ∣ γ w(x)=|x-x_0|^\gamma w ( x ) = ∣ x − x 0 ∣ γ with γ < 0 \gamma<0 γ < 0 . If w ( A i x ) ≤ C w ( x ) w(A_i x)\leq Cw(x) w ( A i x ) ≤ C w ( x ) a.e., then A i x 0 = x 0 A_i x_0=x_0 A i x 0 = x 0 : otherwise y 0 = A i − 1 x 0 ≠ x 0 y_0=A_i^{-1}x_0\neq x_0 y 0 = A i − 1 x 0 = x 0 , and for x x x in a small ball centered at y 0 y_0 y 0 the quotient w ( A i x ) / w ( x ) w(A_i x)/w(x) w ( A i x ) / w ( x ) is arbitrarily large. Thus A 1 x 0 = A 2 x 0 = x 0 A_1x_0=A_2x_0=x_0 A 1 x 0 = A 2 x 0 = x 0 , so ( A 1 − A 2 ) x 0 = 0 (A_1-A_2)x_0=0 ( A 1 − A 2 ) x 0 = 0 and x 0 = 0 x_0=0 x 0 = 0 by (H).
The choice of the exponents p i p_i p i
Remark 5.9. If (H1 _1 1 ) and (H2 _2 2 ) hold for some exponents p i p_i p i , the condition on the weight depends on them through
1 a ( t , q ) = 1 q − ∑ i 1 p i . \frac{1}{a(t,q)}=\frac{1}{q}-\sum_i\frac{1}{p_i}. a ( t , q ) 1 = q 1 − i ∑ p i 1 .
Increasing the p i p_i p i decreases a ( t , q ) a(t,q) a ( t , q ) and, by Lemma 2.1(i), weakens the condition w a ( t , q ) ∈ A 1 w^{a(t,q)}\in A_1 w a ( t , q ) ∈ A 1 ; the weakest one, w q ∈ A 1 w^q\in A_1 w q ∈ A 1 , corresponds to p i = ∞ p_i=\infty p i = ∞ for every i i i . On the other hand, by Remark 3.3, (H1 _1 1 ) and (H2 _2 2 ) become stronger as the p i p_i p i increase. For instance, if Ω i \Omega_i Ω i is Hölder continuous of order ε > δ \varepsilon>\delta ε > δ on S n − 1 S^{n-1} S n − 1 , one may take p i = ∞ p_i=\infty p i = ∞ (Proposition 3.8), while for a rough Ω i ∈ L τ ( S n − 1 ) ∖ L ∞ ( S n − 1 ) \Omega_i\in L^\tau(S^{n-1})\setminus L^\infty(S^{n-1}) Ω i ∈ L τ ( S n − 1 ) ∖ L ∞ ( S n − 1 ) satisfying (6) one takes p i = τ p_i=\tau p i = τ (Proposition 3.10), at the price of a stronger condition on the weight.
A concrete family
Example 5.10. Let n = 2 n=2 n = 2 , α = 9 / 5 \alpha=9/5 α = 9/5 , m = 2 m=2 m = 2 , q 1 = q 2 = 20 q_1=q_2=20 q 1 = q 2 = 20 , A 1 = I A_1=I A 1 = I , A 2 = − I A_2=-I A 2 = − I and, for 0 < ε ≤ 1 0<\varepsilon\leq1 0 < ε ≤ 1 ,
k 1 ( x ) = ∣ x 2 ∣ x ∣ ∣ ε ∣ x ∣ − 1 / 10 , k 2 ( x ) = ∣ x ∣ − 1 / 10 , k_1(x)=\left|\frac{x_2}{|x|}\right|^\varepsilon|x|^{-1/10},\qquad k_2(x)=|x|^{-1/10}, k 1 ( x ) = ∣ x ∣ x 2 ε ∣ x ∣ − 1/10 , k 2 ( x ) = ∣ x ∣ − 1/10 ,
which is the setting of § 4.5 with n = 2 n=2 n = 2 . Let p = 10 / 13 p=10/13 p = 10/13 , so that q = 5 / 2 q=5/2 q = 5/2 and δ = 3 / 5 \delta=3/5 δ = 3/5 .
If 3 / 5 < ε ≤ 1 3/5<\varepsilon\leq1 3/5 < ε ≤ 1 , then T α T_\alpha T α is bounded from H w 10 / 13 10 / 13 ( R 2 ) H^{10/13}_{w^{10/13}}(\mathbb{R}^2) H w 10/13 10/13 ( R 2 ) into L w 5 / 2 5 / 2 ( R 2 ) L^{5/2}_{w^{5/2}}(\mathbb{R}^2) L w 5/2 5/2 ( R 2 ) for every weight w w w such that w 5 / 2 ∈ A 1 w^{5/2}\in A_1 w 5/2 ∈ A 1 and w ( − x ) ≤ C w ( x ) w(-x)\leq Cw(x) w ( − x ) ≤ C w ( x ) ; for instance, w ( x ) = ∣ x ∣ γ w(x)=|x|^\gamma w ( x ) = ∣ x ∣ γ or w ( x ) = ( 1 + ∣ x ∣ ) γ w(x)=(1+|x|)^\gamma w ( x ) = ( 1 + ∣ x ∣ ) γ with − 4 / 5 < γ ≤ 0 -4/5<\gamma\leq0 − 4/5 < γ ≤ 0 , or w ( x ) = ( ∣ x − x 0 ∣ γ + ∣ x + x 0 ∣ γ ) 2 / 5 w(x)=(|x-x_0|^\gamma+|x+x_0|^\gamma)^{2/5} w ( x ) = ( ∣ x − x 0 ∣ γ + ∣ x + x 0 ∣ γ ) 2/5 with x 0 ≠ 0 x_0\neq0 x 0 = 0 and − 2 < γ < 0 -2<\gamma<0 − 2 < γ < 0 .
If 0 < ε ≤ 1 / 5 0<\varepsilon\leq1/5 0 < ε ≤ 1/5 , this fails even for w ≡ 1 w\equiv1 w ≡ 1 .
Indeed, Ω 1 \Omega_1 Ω 1 and Ω 2 ≡ 1 \Omega_2\equiv1 Ω 2 ≡ 1 are bounded, so (H1 _1 1 ) holds with p 1 = p 2 = ∞ p_1=p_2=\infty p 1 = p 2 = ∞ ; then t = 1 t=1 t = 1 and a ( t , q ) = q = 5 / 2 a(t,q)=q=5/2 a ( t , q ) = q = 5/2 . As in § 4.5 , for ∣ x ∣ ≥ 2 ∣ h ∣ |x|\geq2|h| ∣ x ∣ ≥ 2∣ h ∣ ,
∣ k 1 ( x − h ) − k 1 ( x ) ∣ ≤ C ( ∣ h ∣ ∣ x ∣ ) ε ∣ x ∣ − 1 / 10 , ∣ k 2 ( x − h ) − k 2 ( x ) ∣ ≤ C ∣ h ∣ ∣ x ∣ ∣ x ∣ − 1 / 10 , |k_1(x-h)-k_1(x)|\leq C\left(\frac{|h|}{|x|}\right)^\varepsilon|x|^{-1/10},\qquad|k_2(x-h)-k_2(x)|\leq C\frac{|h|}{|x|}|x|^{-1/10}, ∣ k 1 ( x − h ) − k 1 ( x ) ∣ ≤ C ( ∣ x ∣ ∣ h ∣ ) ε ∣ x ∣ − 1/10 , ∣ k 2 ( x − h ) − k 2 ( x ) ∣ ≤ C ∣ x ∣ ∣ h ∣ ∣ x ∣ − 1/10 ,
so Proposition 3.8 gives (H2 _2 2 ) whenever δ = 3 / 5 < ε \delta=3/5<\varepsilon δ = 3/5 < ε . Since max { 2 3 , 10 19 } < 10 13 < 1 \max\left\{\frac{2}{3},\frac{10}{19}\right\}<\frac{10}{13}<1 max { 3 2 , 19 10 } < 13 10 < 1 and (H) holds, (i) follows from Theorem 1.1 and Examples 5.5 and 5.7, while (ii) is the case ε ≤ δ − 1 / q = 1 / 5 \varepsilon\leq\delta-1/q=1/5 ε ≤ δ − 1/ q = 1/5 of Remark 4.7.
Acknowledgement
We express our gratitude to Pablo Rocha (Universidad Nacional del Sur) and Marta S. Urciuolo (Universidad Nacional de Córdoba) for their helpful clarifications throughout the course of this work.
AI Disclosure
During the preparation of this work, the authors used Claude (Opus 5.5, Anthropic ) to assist in editing the manuscript and in writing parts of it, to check some technical steps, and to consult and check the bibliography. It was also used to construct some of the examples and auxiliary results, in particular those in § 5 .
The main ideas and results of this work were developed by the authors, who verified all the arguments and proofs and take full responsibility for the content of the publication.
A. L. Gallo, FaMAF, Universidad Nacional de Córdoba, CIEM (CONICET), 5000 Córdoba, Argentina
Email address: andrea.gallo@unc.edu.ar
M. S. Riveros, FaMAF, Universidad Nacional de Córdoba, CIEM (CONICET), 5000 Córdoba, Argentina
Email address: msriveros@unc.edu.ar
L. A. Vallejos, FaMAF, Universidad Nacional de Córdoba, CIEM (CONICET), 5000 Córdoba, Argentina
Email address: lucas.vallejos@unc.edu.ar