Abstract
We extend to the multilinear setting classical inequalities of Marcinkiewicz and Zygmund on \(\ell ^r\)-valued extensions of linear operators. We show that for certain \(1 \le p, q_1, \dots , q_m, r \le \infty \), there is a constant \(C\ge 0\) such that for every bounded multilinear operator \(T:L^{q_1}(\mu _1) \times \cdots \times L^{q_m}(\mu _m) \rightarrow L^p(\nu )\) and functions \(\{f_{k_1}^1\}_{k_1=1}^{n_1} \subset L^{q_1}(\mu _1), \dots , \{f_{k_m}^m\}_{k_m=1}^{n_m} \subset L^{q_m}(\mu _m)\), the following inequality holds
In some cases we also calculate the best constant \(C\ge 0\) satisfying the previous inequality. We apply these results to obtain weighted vector-valued inequalities for multilinear Calderón-Zygmund operators.
Similar content being viewed by others
Avoid common mistakes on your manuscript.
1 Introduction and Main Results
The study of vector-valued inequalities for linear operators has its origins in the thirties, with works of Bochner, Marcinkiewickz, Paley and Zygmund among others. In this context we find the so-called Marcinkiewicz-Zygmund inequalities for linear operators, regarding the \(\ell ^r\)-valued extensions of linear operators between real \(L^p\)-spaces. That is, given \(1\le p,q,r \le \infty \), the triple (p, q, r) is said to satisfy a Marcinkiewicz-Zygmund inequality if there is a constant C such that for each bounded operator \(T:L^q(\mu ) \rightarrow L^p(\nu )\), (\(\mu \) and \(\nu \) arbitrary \(\sigma \)-finite measures, \(L^q(\mu )\) and \(L^p(\nu )\) real spaces), each \(n \in {\mathbb {N}}\) and functions \(f_1,\dots ,f_n \in L^q(\mu )\),
The infimum of all the constants \(C\ge 1\) satisfying (2) is denoted by \(k_{q,p}(r)\) (setting \(k_{q,p}(r)=\infty \) if there is not such constant). Fixed \(n \in {\mathbb {N}}\), let \(k_{q,p}^{(n)}(r) \in [1, \infty )\) be the infimum of all the constants \(C\ge 0\) satisfying (2) for each T but for only n functions \(f_k\). Note that given \(T:L^q(\mu ) \rightarrow L^p(\nu )\) we can consider the natural \(\ell ^r_n\)-valued extension \(T^{\ell ^r_n} :L^q(\mu , \ell ^r_n) \rightarrow L^p(\nu , \ell ^r_n)\) defined by
It is clear that \(k_{q,p}^{(n)}(r)=\sup \Vert T^{\ell ^r_n}\Vert \), where the supremum is taken over all the operators \(T :L^q(\mu ) \rightarrow L^p(\nu )\) with norm \(\Vert T\Vert \le 1\). Analogously, the validity of (2) is equivalent to saying that each \(T :L^q(\mu ) \rightarrow L^p(\nu )\) has a natural \(\ell ^r\)-valued extension and, in that case, \( k_{q,p}(r) = \sup \Vert T^{\ell ^r} :L^q(\mu , \ell ^r) \rightarrow L^p(\nu , \ell ^r) \Vert , \) where the supremum is taken over all the operators \(T :L^q(\mu ) \rightarrow L^p(\nu )\) with norm \(\Vert T\Vert \le 1\). It is known (see [11, 29.12]) that this supremum does not change if it is taken over two fixed measures \(\mu \) and \(\nu \) such that \(L^q(\mu )\) and \(L^p(\nu )\) are infinite-dimensional. It is worth mentioning that the problem of determining the constants \(k_{q,p}^{(n)}(r)\) is a generalization of the complexification problem, that is, the computation of the so-called complexification constants (\(k_{q,p}^{(2)}(2)\) in our terminology) of operators in real \(L_p\)-spaces, which relate the norm of an operator and its complexification.
In [24], Marcinkiewicz and Zygmund proved that \(k_{q,p}(2)<\infty \) for \(1 \le p,q < \infty \) and \(k_{q,p}(r)<\infty \) whenever \(1 \le \max (p,q)<r<2\). In these cases, they also obtained estimates for the constants \(k_{q,p}(r)\) in terms of the q-th moment of r-stable Lévy measures. In the particular case \(q=p\), they obtained \(k_{p,p}(2)=1\). Herz extended this last equality in [21], showing that \(k_{p,p}(r)=1\) for \(1<p<\infty \) and \(\min (p,2)\le r \le \max (p,2)\), and Grothendieck established the important case \(K_{G,{\mathbb {R}}}:=k_{\infty ,1}(2)<\infty \) (\(K_{G,{\mathbb {R}}}\) stands for the so-called Grothendieck constant in the real case). A systematic study of the constants \(k_{q,p}(r)\) and the precise asymptotic growth of \(k_{q,p}^{(n)}(r)\) is addressed in [12, 17]. In the following theorem, we state some properties of the constants \(k_{q,p}(r)\) (including monotonicity and duality) that can be found in the mentioned papers.
Theorem 1.1
-
(i)
\( k_{q,p}(r) = \lim _{n \rightarrow \infty } k_{q,p}^{(n)}(r). \)
-
(ii)
\(k_{q_1,p_1}^{(n)}(r) \le k_{q_2,p_2}^{(n)}(r)\) whenever \(q_1 \le q_2\) and \(p_2 \le p_1\).
-
(iii)
As a function of r, \(k_{q,p}(r)\) is decreasing on [1, 2] and increasing on \([2, \infty ]\).
-
(iv)
\(k_{q,p}(r) = k_{p', q'}(r')\).
In [12], the set of all the triples \(1 \le p,q,r \le \infty \) satisfying \(k_{q,p}(r) < \infty \) is determined and also the exact value of this constant is obtained in almost all the cases. We state as a theorem these results, that can be found in Sects. 4, 5 and 6 of the aforementioned article.
Theorem 1.2
[12]
-
(i)
If \(q=1\) or \(p=\infty \), then \(k_{q,p}(r)=1\).
-
(ii)
Let \(1< q \le p < \infty \). Then \(k_{q,p}(r) < \infty \) if and only if \(\min \{q,2\} \le r \le \max \{p,2\}\). Moreover, \(k_{q,p}(r)=1\) in that case.
-
(iii)
Let \(1 \le p \le q \le \infty \) with \(q\ne 1\) and \(p\ne \infty \) (these cases are considered in (i)). Then \(k_{q,p}(r) < \infty \) if and only if one of the following cases holds:
-
\(p=q=r\), in which case \(k_{q,p}(r)=1\);
-
\(1 \le p \le q \le 2\) and \(q < r \le 2\), in which case \(k_{q,p}(r) = \frac{c_{r,q}}{c_{r,p}}\);
-
\(2 \le p \le q \le \infty \) and \(2 \le r <p\), in which case \(k_{q,p}(r) = \frac{c_{r',p'}}{c_{r',q'}}\);
-
\(1 \le p \le 2 \le q \le \infty \) and \(r=2\); if, moreover, \(p=2\) or \(q=2\) then \(k_{q,p}(2) = \frac{c_{2,q}}{c_{2,p}}\).
The constant \(c_{r,q}\) denotes the q-th moment of r-stable Lévy measure. The only case in which \(k_{q,p}(r)\) is not determined is when \(1 \le p< r=2 < q \le \infty \).
-
We are interested in the study of Marcinkiewickz-Zygmund inequalities in the context of multilinear operators. In what follows, given \(\{f_k\}_k \subset L^p(\nu )\), we denote
Let \(m \in {\mathbb {N}}\), \(1\le q_1, \dots , q_m, p, r \le \infty \) and consider \({\vec {q}}=(q_1, \dots , q_m)\). We say that the triple \((p;{\vec {q}};r)\) satisfies the Marcinkiewicz-Zygmund inequality if there is a constant C such that for all bounded multilinear operators \(T:L^{q_1}(\mu _1) \times \cdots \times L^{q_m}(\mu _m) \rightarrow L^p(\nu )\) and functions \(\{f_{k_1}^1\}_{k_1=1}^{n_1} \subset L^{q_1}(\mu _1), \dots , \{f_{k_m}^m\}_{k_m=1}^{n_m} \subset L^{q_m}(\mu _m)\), the following inequality holds (with the usual modification when \(r=\infty \))
As in the linear case, we denote by \(k_{{\vec {q}},p}(r)\) the infimum of all the \(C\ge 1\) satisfying (4) and we put \(k_{{\vec {q}},p}(r)=\infty \) when there is no such constant. We denote by \(k_{q,p}^{(n)}(r) \in [1, \infty )\) the infimum of all the constants \(C\ge 0\) satisfying (4) only for \(n_1=\cdots = n_m=n\). It is easy to see that \(\lim _{n \rightarrow \infty } k_{{\vec {q}},p}^{(n)}(r) = k_{{\vec {q}},p}(r)\). As observed in (3) for linear operators, the validity of (4) is equivalent to saying that each \(T:L^{q_1}(\mu _1) \times \cdots \times L^{q_m}(\mu _m) \rightarrow L^p(\nu )\) has the \(\ell ^r\)-valued extension \(T^{\ell ^r}:L^{q_1}(\mu _1, \ell ^r) \times \cdots \times L^{q_m}(\mu _m, \ell ^r) \rightarrow L^p(\nu , \ell ^r({\mathbb {N}}\times \cdots \times {\mathbb {N}}))\) defined by
In that case \(k_{{\vec {q}},p}(r) = \sup \Vert T^{\ell ^r}\Vert \), where the supremum is taken over all the multilinear operators \(T :L^{q_1}(\mu _1) \times \cdots \times L^{q_m}(\mu _m) \rightarrow L^p(\nu )\) (and over all measures \(\mu _i\) and \(\nu \)) with norm \(\Vert T\Vert \le 1\).
As we will point out in Sect. 5.1.1, it is worth mentioning that, when dealing with inequalities of the form
the relation between the powers \(s, r_1, \dots , r_m\) is optimal when \(s=r_1=\cdots = r_m\), as in (4). This establishes a difference with the inequalities of the form
where the sum runs over only one index k and the optimal relation between the powers is given by \(\frac{1}{r}=\frac{1}{r_1}+\cdots + \frac{1}{r_m}\).
In [18], in the context of vector-valued inequalities for multilinear Calderón-Zygmund operators, Grafakos and Martell addressed the multilinear Marcinkiewicz-Zygmund inequality in the particular case \(r=2\), showing that \(k_{{\vec {q}},p}(2)<\infty \) for every \(1\le q_1, \dots , q_m, p<\infty \) (moreover, for \(0<q_1, \dots , q_m, p<\infty \)). They also proved the analogous inequality for multilinear operators from \( L^{q_1}(\mu _1) \times \cdots \times L^{q_m}(\mu _m)\) into \( L^{p,\infty }(\nu )\). Independently, Bombal, Pérez-García and Villanueva proved in [5, Thm. 4.2] a multilinear Marcinkiewicz-Zygmund inequality in the case \(r=2\) in the more general context of multilinear operators defined on Banach lattices. Our goal is to determine conditions on \(p, {\vec {q}},r\) such that the triple \((p,{\vec {q}},r)\) satisfies the Marcinkiewicz-Zygmund inequality (4) and, if it is possible, to calculate the exact value of the constant \(k_{{\vec {q}},p}(r)\). In this sense, our main result is the following.
Theorem 1.3
Let \(1 \le q_1, \dots , q_m \le \infty \), \(1 \le p < \infty \) and \(\varvec{q}=\max \{q_1, \dots , q_m\}\).
-
(i)
Suppose \(\varvec{q}\le p\).
-
(ia)
If \(1 \le \varvec{q}\le p \le 2\), then \(k_{{\vec {q}},p}(r)< \infty \) iff \(\varvec{q}\le r \le 2\) or \(\varvec{q}=1\) and \(1 \le r\le \infty \).
-
(ib)
If \(2 \le \varvec{q}\le p\), then \(k_{{\vec {q}},p}(r)< \infty \) iff \(2 \le r \le p\).
-
(ic)
If \(1 \le \varvec{q}\le 2 \le p\), then \(k_{{\vec {q}},p}(r)< \infty \) iff \(\varvec{q}\le r \le p\) or \(\varvec{q}=1\) and \(1 \le r\le \infty \).
-
(ia)
-
(ii)
Suppose \(p < \varvec{q}\).
-
(iia)
If \(1 \le p < \varvec{q}\le 2\), then \(k_{{\vec {q}},p}(r)< \infty \) iff \(\varvec{q}< r \le 2\) or \(\varvec{q}=r=2\).
-
(iib)
If \(2< p < \varvec{q}\), then \(k_{{\vec {q}},p}(r)< \infty \) implies \(2 \le r < p\). If \(r=2\) then \(k_{{\vec {q}},p}(r)<\infty \).
-
(iic)
If \(1 \le p \le 2 \le \varvec{q}\), then \(k_{{\vec {q}},p}(r)< \infty \) iff \(r=2\).
-
(iia)
If (ia), (ic), (iia) holds or if \(\varvec{q}\le r \le p\), then \(k_{{\vec {q}},p}(r)= k_{q_1,p}(r)\cdots k_{q_m,p}(r)\).
Note that the only case in which we do not get an equivalence is in (iib), where we obtain a necessary condition for \(k_{{\vec {q}},p}(r)< \infty \) (and the, already known, sufficient condition \(r=2\)). In Sect. 5 we address the case \(0< p < 1\) and the case of weak type estimates, which will be relevant in obtaining vector-valued estimates for multilinear singular integrals.
We also obtain some properties of the constants \(k_{{\vec {q}},p}(r)\) that, besides being important in the proof of the previous theorem, are interesting on their own. First, we study the relation between \(k_{{\vec {q}},p}(r)\) and the linear constants \(k_{q_1,p}(r), \dots , k_{q_m,p}(r)\). We get the following result which, in particular, shows that if \(k_{q_i,p}(r)=\infty \) for some \(1\le i\le m\), then \(k_{{\vec {q}},p}(r)=\infty \).
Proposition 1.4
Let \(1 \le p,q_1, \dots , q_m, r \le \infty \). Then, for each \(n \in {\mathbb {N}}\) we have
Consequently, \(k_{q_1;p}(r) \cdots k_{q_m;p}(r) \le k_{{\vec {q}};p}(r)\). Moreover, if \(p=r\) then equality holds.
We also prove the following monotonicity properties of \(k_{{\vec {q}}, p}(r)\) as a function of p and r, partially extending to multilinear setting the properties (ii) and (iii) of Theorem 1.1.
Proposition 1.5
Let \(1\le q_1, \dots , q_m,p \le \infty \).
-
(i)
If \(1\le p_2<p_1 \le \infty \), then \(k_{{\vec {q}},p_1}^{(n)}(r) \le k_{{\vec {q}},p_2}^{(n)}(r)\) for all \(1\le r \le \infty \). Consequently, \(k_{{\vec {q}},p_1}(r) \le k_{{\vec {q}},p_2}(r)\).
-
(ii)
If \(1\le r<s \le 2\), then \(k_{{\vec {q}},p}(s) \le k_{{\vec {q}},p}(r)\).
Although Theorem 1.3 seems to suggest that \(k_{{\vec {q}},p}(r) < \infty \) if and only if \(k_{q_i,p}(r)<\infty \) for all \(1 \le i \le m\), we establish an important difference between the Marcinkiewicz-Zygmund inequalities in the linear and multilinear cases when, for instance, we put \(p=\infty \). First, as a simple consequence of the optimality in the well-known Littlewood’s 4 / 3 inequality, we see this different behavior between the linear and bilinear cases when we put \(q_1=q_2=p=\infty \).
Remark 1.6
Denote \(\mathbf {\infty }=(\infty ,\infty )\). Then, although \(k_{\infty ,\infty }(r)=1\) for all \(1\le r\le \infty \) (this was stated in Theorem 1.2), we have \(k_{\mathbf {\infty },\infty }(r)=\infty \) whenever \(1\le r < 4/3\). Indeed, suppose that \(k_{\mathbf {\infty },\infty }(r) < \infty \). Then, for any \(n \in {\mathbb {N}}\) and \(T:\ell ^\infty _n \times \ell ^\infty _n \rightarrow {\mathbb {R}}\) we have
Thus, we obtain \( \left( \sum _{k_1, k_2=1}^\infty |T(e_{k_1}, e_{k_2})|^r\right) ^{1/r} \le k_{\mathbf {\infty },\infty }(r) \Vert T\Vert \) for every bilinear form \(T:c_0\times c_0 \rightarrow {\mathbb {R}}\). The optimality in Littlewood’s 4 / 3 inequality implies that \(r\ge 4/3\).
The next proposition generalizes the previous remark showing that, in contrast to the linear case where \(k_{q, \infty }(r) =1\) for all \(1 \le q, r \le \infty \), in the m-linear case \(k_{{\vec {q}}, \infty }(r)=\infty \) for appropriate choices of \(m \in {\mathbb {N}}\) and \(1 \le q_1, \dots , q_m,r \le \infty \). In particular, we see that \(k_{q_i,p}(r)<\infty \) for all \(1 \le i\le m\) does not imply, in general, \(k_{{\vec {q}},p}(r)<\infty \). It should be noted that, just as in Remark 1.6 the optimality of Littlewood’s 4 / 3 inequality was used, Proposition 1.7 is related to the optimality of some of its multilinear versions [14], such as the Bohnenblust-Hille inequality [6] (see also [10, 22]). In fact, our proof follows the spirit of the new approaches to these inequalities by means of probabilistic method [27] (see Lemma 4.1).
Proposition 1.7
Let \(1 \le q_1, \dots , q_m \le \infty \) and \(\varvec{q}= \max \{q_1, \dots , q_m\}\). Then, for
we have \(k_{{\vec {q}}, \infty }(r) = \infty \), while \(k_{q_i, \infty }(r)=1\), \(1 \le i \le m\).
Remark 1.8
It is easy to check that the expression in the right hand side of (7) is smaller than \(\min \{\varvec{q}, 2\}\). On the other hand, since \(k_{{\vec {q}},\infty }(\min \{\varvec{q}, 2\})<\infty \) and \(k_{{\vec {q}},\infty }(\infty )=1\) (see Theorem 1.3(ib) and Remark 2.3), by Corollary 2.6 below we have \(k_{{\vec {q}},\infty }(r)< \infty \) for \(\min \{\varvec{q}, 2\} \le r \le \infty \). This brings up the question on the behaviour of \(k_{{\vec {q}},\infty }(r)\) for the remaining range of values of r. Although the multilinear versions of Littlewood inequalities are understood for the full range of exponents, we were not be able to fill the gap for r in our results.
As the nature of Marcinkiewicz-Zygmund inequalities is related to the study of \(\ell ^r\)-valued extensions of bounded multilinear operators on \(L^p\)-spaces, a brief review of some of the inequalities involving Theorem 1.3 allows us to obtain vector-valued estimates for multilinear singular integrals. For instance, we prove that if \(1< q_1, \dots , q_m<r < 2\) and \(p>0\) are such that \( \frac{1}{p}=\frac{1}{q_1}+\cdots +\frac{1}{q_m} \) and \(\mathbf {w}=(w_1,\dots ,w_m)\) satisfies the multilinear \(A_{{\vec {q}}}\) condition defined in [23], then there exists a constant \(C>0\) such that for every Calderón-Zygmund multilinear operator T we have
where \(\nu _{\mathbf {w}}=\prod _{i=1}^m w_i^{p/q_i}\). We address the expected weak type estimates when some of the exponents \(q_i\) are equal to one. We compare our results with some known vector-valued estimates obtained in [1, 2, 8, 18].
Finally, we present the multilinear version of [16, Chapter V, Thm. 1.12], which states that if \(T:L^q(\mu ) \rightarrow L^p(\nu )\) is a positive linear operator (that is, \(f\ge 0\) implies \(T(f)\ge 0\)), then the Marcinkiewicz-Zygmund inequality (2) holds for all \(1 \le r \le \infty \), with \(C=1\).
Proposition 1.9
Let \(0< p, q_1, \dots , q_m \le \infty \), \(1 \le r \le \infty \) and \(T:L^{q_1}(\mu _1)\times \cdots L^{q_m}(\mu _m) \rightarrow L^p(\nu )\) be a positive multilinear operator. Then
for any choice of functions \(\{f_{k_i}^i\}_{k_i=1}^{n_i} \subset L^{q_i}(\mu _i)\), \(1\le i \le m\).
As an immediate consequence, we obtain the estimate
for the bilinear convolution operator, whenever \(\frac{1}{q_1} + \frac{1}{q_2} = \frac{1}{p} + 1\).
1.1 Structure of the Article
The article is organized as follows. In Sect. 2 we prove the mentioned properties of the constants \(k_{{\vec {q}},p}(r)\). The inequality \(k_{q_1;p}(r) \cdots k_{q_m;p}(r) \le k_{{\vec {q}};p}(r)\) is almost immediate, while the equality in the case \(p=r\) follows by induction. For the monotonicity of \(k_{{\vec {q}},p}(r)\) as a function of p we use a duality argument. The monotonicity in r makes use of r-stable Lévy measures and a generalization of some arguments in [17]. We include in this section a kind of interpolation property of \(k_{{\vec {q}},p}(r)\) as a function of r (see Corollary 2.6) and we also prove that \(k_{{\vec {q}},p}(r)=1\) when \({\vec {q}}= (1, \dots , 1)\). Our main results are given in Sect. 3, where we determine conditions on \({\vec {q}},p\) and r so that \(k_{{\vec {q}},p}(r) < \infty \) and we obtain (in many cases) the exact value of these constants. In Theorem 3.1 we focus on the case \(1\le q_1, \dots , q_m \le 2\), \(1 \le p < \infty \). The proof of this theorem combines some classical arguments with properties from Sect. 2 and the values of the constants \(k_{q_i,p}(r)\) obtained in [12]. In Sect. 3.2 we include the proof of Theorem 1.3, which at that point is just a combination of the results previously obtained. In Sect. 4 we treat the case \(p=\infty \), showing (see Proposition 1.7) that the behavior of \(k_{{\vec {q}},\infty }(r)\) is very different from that of \(k_{q_i, \infty }(r)\). The proof relies on a multilinear version of a Kahane-Salem-Zygmund inequality. In Sect. 5 we discuss some applications of the Marcinkiewicz-Zygmund inequalities. We deal with weighted vector-valued inequalities for multilinear Calderón-Zygmund operators in Sect. 5.1 and, in Sect. 5.2, we prove Proposition 1.9, which gives vector-valued inequalities for positive multilinear operators (such as the convolution).
1.2 Notation
All the Banach spaces considered are real and all the measure spaces \((\Omega , \mu )\) are \(\sigma \)-finite. As usual, given a measure space \((\Omega , \mu )\) the space of measurable functions \(f:\Omega \rightarrow {\mathbb {R}}\) such that \(\int _{\Omega }|f(\omega )|^p\, d\mu (\omega ) < \infty \) is denoted by \(L^p(\mu )=L^p(\Omega ; d\mu )\) (we omit the set \(\Omega \) which will be clear by context). Given a Banach space X, we denote by \(L^p(\mu , X)\) the vector space of X-valued \(\mu \)-measurable functions \(f :\Omega \rightarrow X\) such that \(\Vert f(\cdot )\Vert _{X}^p\) is \(\mu \)-integrable. When \(1 \le p \le \infty \), this is a Banach space with the natural norm \( \Vert f\Vert _{L^p(\mu , X)} = \left( \int \Vert f(\omega )\Vert _{X}^p d\mu (\omega )\right) ^{1/p} \) (the case \(p=\infty \) is defined analogously). Following the standard notation, we denote by \(\ell ^p\) the space of sequences \((a_k)_{k\in {\mathbb {N}}}\) in \({\mathbb {R}}\) such that \(\Vert (a_k)_k\Vert _{\ell ^p} = \left( \sum _k |a_k|^p\right) ^{1/p} < \infty \) and \(\ell ^p_n = ({\mathbb {R}}^n, \Vert \cdot \Vert _{\ell ^p})\), with the usual modification when \(p=\infty \).
2 Basic Properties of \(k_{{\vec {q}},p}(r)\)
2.1 Relation Between the Linear and Multilinear Constants
Before proving Proposition 1.4 we state the following easy remark, whose proof is omitted. Recall that if \((\Omega _1, \nu _1),\dots , (\Omega _m, \nu _m)\) are measure spaces, then \(\nu _1 \otimes \cdots \otimes \nu _m\) denotes the product measure on \(\Omega _1 \times \cdots \times \Omega _m\).
Remark 2.1
Let \(T_i :L^{q_i}(\mu _i) \rightarrow L^p(\nu _i)\) be bounded operators (\(i=1,\dots , m\)) and let \(T :L^{q_1}(\mu _1) \times \cdots \times L^{q_m}(\mu _m) \rightarrow L^p(\nu _1 \otimes \cdots \otimes \nu _m)\) be the m-linear operator defined as \( T(f^1, \dots , f^m)= T_1(f^1) \otimes \cdots \otimes T_m(f^m), \) , i.e.,
Then \(\Vert T\Vert =\Vert T_1\Vert \cdots \Vert T_m\Vert \) and
for each \(n \in {\mathbb {N}}\) and \(\{f_{k_i}^i\}_{k_i=1}^n \subset L^{q_i}(\mu _i)\).
Proof of Proposition 1.4
It is clear that \(k_{q_1;p}^{(n)}(r) \cdots k_{q_m;p}^{(n)}(r)\) is the infimum of all the constants \(C\ge 0\) satisfying
for every \(T_i :L^{q_i}(\mu _i) \rightarrow L^p(\nu _i)\) and all functions \(\{f_{k_i}^i\}_{k_i=1}^n \subset L^{q_i}(\mu _i)\) (\(1 \le i \le m\)). By the previous remark, we see that \(k_{q_1;p}^{(n)}(r) \cdots k_{q_m;p}^{(n)}(r)\) is the infimum of all the constants \(C\ge 0\) satisfying
for every m-linear operator \(T:L^{q_1}(\mu _1) \times \cdots \times L^{q_m}(\mu _m) \rightarrow L^p(\nu _1 \otimes \cdots \otimes \nu _m)\) of the form \(T(f^1, \cdots , f^m)=T_1(f^1)\otimes \cdots \otimes T_m(f^m)\) and functions \(\{f_{k_i}^i\}_{k_i=1}^n \subset L_{q_i}(\mu _i)\). This shows that the constant \(k_{{\vec {q}},p}^{(n)}(r)\), which is the infimum of all the constants satisfying (10) for every m-linear operator from \(L^{q_1}(\mu _1) \times \cdots \times L^{q_m}(\mu _m)\) to \(L^p(\nu )\), is greater than or equal to \(k_{q_1;p}^{(n)}(r) \cdots k_{q_m;p}^{(n)}(r)\).
To prove equality when \(p=r\), we reason by induction on m (see, for instance, [5, Thm. 3.1] for a similar argument). For \(m=1\) the result is trivial, then let \(m \ge 2\) and suppose that the result holds for \(m-1\). We write the proof for \(1 \le r<\infty \), the case \(r=\infty \) being analogous. Fix \(T:L^{q_1}(\mu _1) \times \cdots \times L^{q_m}(\mu _m) \rightarrow L^r(\nu )\), \(n \in {\mathbb {N}}\) and \(\{f_{k_i}^i\}_{k_i=1}^{n} \subset L^{q_i}(\mu _i)\). Our goal is to prove
which would give the remaining inequality \( k_{{\vec {q}},r}^{(n)}(r) \le k_{q_1,r}^{(n)}(r) \cdots k_{q_m,r}^{(n)}(r)\). For each \(2\le i \le m\) let \(\nu _i\) be the counting measure on \({\mathbb {N}}\) and define \(S:L^{q_1}(\mu _1) \rightarrow L^r(\nu \otimes \nu _2 \otimes \cdots \otimes \nu _m)\) by
It is easy to check that
Now, on the one hand we have
and on the other hand, if we call \(T_{f^1}(f^2, \dots , f^m) = T(f^1,\dots , f^m)\) and \({\vec {q}}_{2, \dots , m}=(q_2, \dots , q_m)\), then
from where we deduce
Putting (11), (12) and (13) together we obtain \(k_{{\vec {q}},r}^{(n)}(r) \le k_{q_1,r}^{(n)}(r) k_{{\vec {q}}_{2, \dots , m}, r}^{(n)}(r)\). By the induction hypothesis we have \(k_{{\vec {q}}_{2, \dots , m}, r}^{(n)}(r) \le k_{q_2,r}^{(n)}(r) \cdots k_{q_m,r}^{(n)}(r)\) and this proves the statement. \(\square \)
If \(T_1:L^{q_1}(\mu _1) \rightarrow L^p(\nu _1)\) and \(T_2:L^{q_2}(\mu _2) \times \cdots \times L^{q_m}(\mu _m) \rightarrow L^p(\nu _2)\), then
is an m-linear operator satisfying \(\Vert T\Vert =\Vert T_1\Vert \Vert T_2\Vert \) and an equality analogous to (9). Then, reasoning as in the proof of Proposition 1.4, we see that if \(({\vec {q}},p,r)\) satisfies the (m-linear) Marcinkiewicz-Zygmund inequality, then \(((q_2,\dots ,q_m),p,r)\) satisfies the (\((m-1)\)-linear) Marcinkiewicz-Zygmund inequality. Clearly, the same reasoning applies with any \(q_i\) instead of \(q_1\). Then, we have the following.
Remark 2.2
Let \(1 \le p,q_1, \dots , q_m, r \le \infty \). Then,
for \(j=1,\dots ,m\).
Applying Theorem 1.2 and Proposition 1.4 for the case \(p=r\), we see that \(k_{{\vec {q}},r}(r) < \infty \) if and only if \(r=2\) or \(q_i \le r\) for all \(1 \le i \le m\). Moreover, if \(q_i \le r\) for all \(1 \le i \le m\) then \(k_{{\vec {q}},r}(r)=1\). This, together with the monotonicity in p (see Proposition 1.5), gives the following result.
Remark 2.3
If \(1 \le q_1, \dots , q_m \le r \le p \le \infty \) then \(k_{{\vec {q}},p}(r) =1\).
2.2 Monotonicity in p
Next, we prove the property of monotonicity of \(k_{{\vec {q}},p}(r)\) as a function of p stated in Proposition 1.5(i). The proof follows a duality argument analogous to that of [18, Thm. 6].
Proof of monotonicity in p
Let \(T :L^{q_1}(\mu _1) \times \cdots \times L^{q_m}(\mu _m) \rightarrow L^{p_1}(\nu )\) be an m-linear operator and \(\{f_{k_i}^i\}_{k_i=1}^n \subset L^{q_i}(\mu _i)\). Assume \(1 \le r < \infty \), the case \(r=\infty \) being completely analogous. By duality we have
Now, fixed \(g \in L^{(p_1/p_2)'}(\nu )\) with norm at most 1, consider \(T_g :L^{q_1}(\mu _1) \times \cdots \times L^{q_m}(\mu _m) \rightarrow L^{p_2}(\nu )\) defined by
and note that (14) gives
Since \(\Vert T_g\Vert \le \Vert T\Vert \) for every \(g \in L^{(p_1/p_2)'}(\nu )\) of norm at most 1 (just apply Hölder’s inequality), we obtain
which gives \(k_{{\vec {q}},p_1}^{(n)}(r) \le k_{{\vec {q}},p_2}^{(n)}(r)\). \(\square \)
2.3 Monotonicity and Interpolation Property in r
For the proof of the monotonicity of \(k_{{\vec {q}},p}(r)\) as a function of r (when \(1\le r \le 2\)), we make use of r-stable variables (or r-stable Lévy measures). A random variable is said to be r-stable (\(0< r\le 2\)) if its Fourier transform on \({\mathbb {R}}\) is equal to \(e^{-|x|^r}\). The s-th moment of the r-stable variable w is given by \(c_{r,s}=\left( \int _{{\mathbb {R}}} |t|^s w(t) dt \right) ^{1/s}\). Note that 2-stable variables are just Gaussian variables. One importance of r-stable variables relies on the following well known property, whose proof can be found in [26, 21.1.3]: For \(0<s<r<2\) or \(r=2\) and \(0<s<\infty \), if \(\{w_k\}_{k}\) is a sequence of independent r-stable random variables defined on [0, 1], then
for every sequence \(\{a_k\}_k\subset \mathbb R\) and any \(n \in {\mathbb {N}}\).
The following lemma will be a main tool to prove the desired monotonicity and will also be useful in Sect. 3.1.
Lemma 2.4
Let \(0<r\le 2\) and \(0<p<\infty \). If \(\{w_{k_1}\}_{k_1}, \dots , \{w_{k_m}\}_{k_m}\) are sequences of mutually independent r-stable random variables defined on [0, 1], then
for every sequence \(\{a_{k_1, \dots , k_m}\}_{k_1, \dots , k_m}\subset \mathbb R\) and any \(n \in {\mathbb {N}}\), where \(C=c_{r,p}^{m}\) if \(0<p< r <2\) or \(0<p\le r=2\), \(C=c_{2,2}^{m}\) if \(r=2< p < \infty \) and \(C=c_{r, s}^m\) if \(r<2\) and \(r\le p\), for any \(0<s<r\).
Proof
We begin with the cases \(0<p< r <2\) and \(0<p\le r=2\). Note that, by (15), we have
Then, applying the continuous Minkowski inequality with \(r/p\ge 1\) we obtain
which together with (17) gives
Using (15) again, we see that
and, in virtue of (18),
Repeating this argument, in \((m-1)\)-steps we obtain
and applying Minkowski’s inequality (with \(r/p\ge 1\)) and (15) we get
which proves the statement.
The remaining cases, where p is greater than or equal to r, follow choosing \(s=r=2\) or \(s<r<2\) and applying the previous cases to get
\(\square \)
Another useful tool in the proof of the monotonicity in r, is the following generalization to the multilinear setting of a result that can be found inside the proof of [17, Thm. 1].
Lemma 2.5
Let \(1 \le q_1, \dots , q_m, p\le \infty \) and \(1 \le r < \infty \). Then
for each bounded multilinear operator \(T:L^{q_1}(\mu _1) \times \cdots \times L^{q_m}(\mu _m) \rightarrow L^p(\nu )\), functions \(\{f_{k_i}^i\}_{k_i=1}^{n} \subset L^{q_i}(\mu _i)\), \(\{g_{k_i}^i\}_{k_i=1}^n \subset L^r[0,1]\) (\(i=1, \dots , m\)) and \(n \in {\mathbb {N}}\).
Proof
Fix \(T:L^{q_1}(\mu _1) \times \cdots \times L^{q_m}(\mu _m) \rightarrow L^p(\nu )\), \(n \in {\mathbb {N}}\) and \(\{f_{k_i}^i\}_{k_i=1}^{n} \subset L^{q_i}(\mu _i)\), \(\{g_{k_i}^i\}_{k_i=1}^n \subset L^r[0,1]\). Along the proof, we will write \(\Vert \cdot \Vert _r\) instead of \(\Vert \cdot \Vert _{L^r[0,1]}\) or \(\Vert \cdot \Vert _{L^r[0,1]^m}\), depending on the context. Let \(\varepsilon >0\) and, for each \(1 \le i \le m\), consider a sequence of simple functions \(\{s_{k_i}^i\}_{k_i=1}^n\) such that \(\Vert g_{k_i}^i - s_{k_i}^i\Vert _{r} < \varepsilon \), \(k_i=1, \dots , n\). On the one hand, reasoning as in [17, Thm. 1], we obtain
On the other hand, \(\Vert g_{k_1}^1 \cdots g_{k_m}^m - s_{k_1}^1 \cdots s_{k_m}^m\Vert _{r} < \gamma (\varepsilon )\) for some \(\gamma (\varepsilon ) \xrightarrow [\varepsilon \rightarrow 0]{}\, 0\) (just add and subtract the terms \(s_{k_1}^1 g_{k_2}^2 \cdots g_{k_m}^m, s_{k_1}^1 s_{k_2}^2 \cdots g_{k_m}^m, \dots , s_{k_1}^1 \cdots s_{k_{m-1}}^{m-1} g_{k_m}^m\) and apply triangle inequality). Then, we have
Note that \(\gamma (\varepsilon )\) is independent of the \(k_i\)’s; it depends on \(\max _{i, k_i} \Vert g_{k_i}^i\Vert _{r}\) but this is not a problem since the functions \(\{g_{k_i}^i\}_{k_i}\) are fixed. Now, the previous inequality together with the monotonicity of the norm \(\Vert \cdot \Vert _p\), gives
Then, if we prove that
we would have
and since \(\varepsilon >0\) was chosen arbitrarily, letting \(\varepsilon \rightarrow 0\),
which is the desired statement. Then, it only remains to prove (22). Since each \(s_{k_j}^j\) is a simple function, there exist \(c_{i_j, k_j}^j \in {\mathbb {R}}\) and \(A_{i_j}\) measurable (disjoint) subsets of [0, 1] such that \( s_{k_j}^j(\cdot ) = \sum _{i_j} c_{i_j, k_j}^j \chi _{A_{i_j}}(\cdot ). \) Now, if we denote by \(\lambda (A_{i_j})\) the measure of \(A_{i_j}\),
which proves (22) (once again, we are assuming \(1 \le p < \infty \), the case \(p=\infty \) being completely analogous). \(\square \)
Finally, we prove the monotonicity in r stated in Proposition 1.5(ii).
Proof of monotonicity in r
Let \(1 \le r<s \le 2\). The proof follows by a simple application of Lemma 2.4 and Lemma 2.5 with the sequences \(\{g_{k_i}^i\}_{k_i}=\{w_{k_i}\}_{k_i}\) of s-stable random variables and functions \(\{f_{k_i}^i\}_{k_i=1}^{n} \subset L^{q_i}(\mu _i)\). Indeed, we have
and this proves that \(k_{{\vec {q}},p}^{(n)}(s) \le k_{{\vec {q}},p}(r)\). Consequently, \(k_{{\vec {q}},p}(s) \le k_{{\vec {q}},p}(r)\). \(\square \)
It is worth mentioning that the above proof does not assure \(k_{{\vec {q}},p}^{(n)}(s) \le k_{{\vec {q}},p}^{(n)}(r)\) for each \(n \in {\mathbb {N}}\). The problem arises in Lemma 2.5, where we cannot put \(k_{{\vec {q}},p}^{(n)}(r)\) instead of \(k_{{\vec {q}},p}(r)\). This problem was stated in [17, Problem 1] in the linear case.
We will see now a kind of interpolation behavior of \(k_{{\vec {q}},p}(r)\) as a function of r. First, we need the following known result (see, for instance, [3, Sections 5.6, 5.7 and 5.8]).
Let\(1 \le r_1,r_2 \le \infty \)andTbe a multilinear operator which is bounded from \(L^{q_1}(\mu _1, \ell ^{r_i}) \times \cdots \times L^{q_m}(\mu _m, \ell ^{r_i})\)into\(L^p(\nu , \ell ^{r_i}({\mathbb {N}}\times \cdots \times {\mathbb {N}}))\)with norm\(M_i\) (\(i=1,2\)). If\(\frac{1}{r}=\frac{1-\theta }{r_1} + \frac{\theta }{r_2}\)with\(0<\theta <1\), then\(T :L^{q_1}(\mu _1, \ell ^{r}) \times \cdots \times L^{q_m}(\mu _m, \ell ^{r}) \rightarrow L^p(\nu , \ell ^{r}({\mathbb {N}}\times \cdots \times {\mathbb {N}}))\)with norm less than or equal to\(M_1^{1 - \theta }M_2^{\theta }\). The same is true if we put \(\ell ^r_n\)instead of \(\ell ^r\), where \(\ell ^r_n({\mathbb {N}}\times \cdots \times {\mathbb {N}})\)is the space of sequences\((a_{k_1, \dots , k_m})_{k_1, \dots , k_m=1}^n\)with the normr.
As a consequence, we obtain the mentioned property of \(k_{{\vec {q}},p}(r)\).
Corollary 2.6
Let \(1 \le p, q_1, \dots , q_m, r_1, r_2 \le \infty \). If \(\frac{1}{r}=\frac{1-\theta }{r_1} + \frac{\theta }{r_2}\) with \(0<\theta <1\), then \(k_{{\vec {q}},p}^{(n)}(r) \le k_{{\vec {q}},p}^{(n)}(r_1)^{1-\theta } k_{{\vec {q}},p}^{(n)}(r_2)^{\theta }\). Consequently, if \(k_{{\vec {q}},p}(r_i)< \infty \) for \(i=1,2\), then \(k_{{\vec {q}},p}(r) \le k_{{\vec {q}},p}(r_1)^{1-\theta } k_{{\vec {q}},p}(r_2)^{\theta } < \infty \).
Proof
It is enough to prove that, given \(T :L^{q_1}(\mu _1) \times \cdots \times L^{q_m}(\mu _m) \rightarrow L^p(\nu )\) with norm \(\Vert T\Vert \le 1\), the \(\ell ^r_n\)-valued extension \(T^{\ell ^r_n} :L^{q_1}(\mu _1, \ell ^{r}_n) \times \cdots \times L^{q_m}(\mu _m, \ell ^{r}_n) \rightarrow L^p(\nu , \ell ^{r}_n({\mathbb {N}}\times \cdots \times {\mathbb {N}}))\) defined as in (5) has norm less than or equal to \(k_{{\vec {q}},p}^{(n)}(r_1)^{1-\theta } k_{{\vec {q}},p}^{(n)}(r_2)^{\theta }\). For this, simply note that each \(\ell ^{r_i}_n\)-valued extension \(T^{\ell ^{r_i}_n}\) (\(i=1,2\)) has norm less than or equal to \(k_{{\vec {q}},p}^{(n)}(r_i)\) and apply the previous interpolation theorem. \(\square \)
2.4 The Case \({\vec {q}}=(1, \dots , 1)\)
To finish this section we prove a generalization to the multilinear setting of a result stated in Theorem 1.2(i), which asserts that \(k_{1,p}(r) =1\) for all \(1 \le p,r \le \infty \). Recall that \(L^p(\mu , X)\) is the space of p-integrable X-valued functions. When \(p=1\), there is a natural (isometric) identification between the projective tensor product \(L^1(\mu ) \tilde{\otimes }_{\pi } X\) and the space \(L^1(\mu , X)\). Having in mind this isometry, we adopt the tensor notation \(g\otimes x(\omega ) = g(\omega ) x\). We refer readers to [11, 3.3] for the definition of the projective tensor product and a detailed exposition of this topics. For our purposes, we will need the following well-known fact: given \(f \in L^1(\mu ,X)\) (or \(L^1(\mu ) \tilde{\otimes }_{\pi } X\), via the identification) and \(\varepsilon >0\) there exist bounded sequences \((g_k)_k \subset L^1(\mu )\) and \((x_k)_k \subset X\) such that the series \(\sum _{k=1}^\infty g_k \otimes x_k\) converges to f (in the projective norm) and \( \sum _{k=1}^\infty \Vert g_k\Vert _{L^1(\mu )} \Vert x_k\Vert _X < \Vert f\Vert _{L^1(\mu , X)} + \varepsilon . \) Then, it is clear that
where the infimum is taken over all the representations \(f=\sum _{k=1}^\infty g_k \otimes x_k\) with \(g_k \in L^1(\mu )\) and \(x_k \in X\).
Proposition 2.7
If \({\vec {q}}=(1, \dots , 1)\), then \(k_{{\vec {q}},p}(r)=1\) for all \(1 \le p,r \le \infty \).
Proof
We suppose \(1 \le p, r < \infty \) (the proof is the same in the remaining cases). Let \(T:L^1(\mu _1) \times \cdots \times L^1(\mu _m) \rightarrow L^p(\nu )\) and functions \(\{f_{k_i}^i\}_{k_i=1}^n \in L^1(\mu _i)\), \(i=1, \dots , m\). Consider \(h_i \in L^1(\mu _i, \ell ^r)\) defined by \(h_i=\sum _{k_i=1}^n f_{k_i}^i \otimes e_{k_i}\) and note that
By (23), given \(\varepsilon >0\) we can take (for each \(1\le i \le m\)) a representation \(h_i= \sum _{l_i} g_{l_i}^i \otimes x_{l_i}^i\) such that
Now, consider \(T^{\ell ^r} :L^1(\mu _1, \ell ^r) \times \cdots \times L^1(\mu _m, \ell ^r) \rightarrow L^p(\nu , \ell ^r({\mathbb {N}}\times \cdots \times {\mathbb {N}}))\) the m-linear operator defined by \(T^{\ell ^r}(f^1 \otimes x^1, \dots , f^m \otimes x^m)(\cdot ) = T(f^1, \dots , f^m)(\cdot ) x^1\otimes \cdots \otimes x^m\), where \({x^1\otimes \cdots \otimes x^m}\) stands for \((x^1(i_1) \cdots x^m(i_m))_{i_1, \dots , i_m=1}^\infty \in \ell ^r({\mathbb {N}}\times \cdots \times {\mathbb {N}})\). On the one hand we have that \(T^{\ell ^r}(h_1, \dots , h_m)(\cdot )=\sum _{k_1, \dots ,k_m=1}^n T(f_{k_1}^1, \dots , f_{k_m}^m)(\cdot ) e_{k_1}\otimes \cdots \otimes e_{k_m}\) and, hence,
On the other hand \( T^{\ell ^r}(h_1, \dots , h_m)(\cdot )=\sum _{l_1, \dots , l_m} T(g_{l_1}^1, \dots , g_{l_m}^m)(\cdot ) x_{l_1}^1 \otimes \cdots \otimes x_{l_m}^m, \) from where
the last inequality due to (24) and (25). Putting together (26) and the inequality above we obtain
and since \(\varepsilon >0\) was arbitrary we deduce that \(k_{(1, \dots ,1), p}^{(n)}(r) =1\) for all \(n \in {\mathbb {N}}\). \(\square \)
3 The Triples \(({\vec {q}},p,r)\) Satisfying \(k_{{\vec {q}},p}(r) < \infty \)
3.1 The Case \(1 \le q_1,\dots , q_m \le 2\), \(p<\infty \)
In this section we see that, in the particular case \(1 \le q_1,\dots , q_m \le 2\) and \(1 \le p<\infty \), we can determine the triples \((p,{\vec {q}},r)\) satisfying the Marcinkiewicz-Zygmund inequality and the exact values of the constants \(k_{{\vec {q}},p}(r)\). We follow ideas of [17] (see Theorems 2 and 3 in there) to obtain upper estimates for the constants \(k_{{\vec {q}},p}(r)\) and then use Proposition 1.4 for the lower bounds. The exact value of \(k_{{\vec {q}},p}(r)\) is then determined by the exact values of the constants \(k_{q_i,p}(r)\) obtained in [12].
Theorem 3.1
Let \(1\le q_1, \dots , q_m \le 2\), \(1 \le p < \infty \) and denote \(\varvec{q}= \max \{q_1, \dots , q_m\}\). Then \(k_{{\vec {q}},p}(r) < \infty \) if and only if one of the following conditions is satisfied:
-
(i)
\(r=2\);
-
(ii)
\(\varvec{q}=1\) and \(1 \le r \le \infty \);
-
(iii)
\(\varvec{q}< r \le \max \{p,2\}\);
-
(iv)
\(\varvec{q}=r \le p\).
Moreover, in all these cases we have \(k_{{\vec {q}},p}(r) = k_{q_1,p}(r) \cdots k_{q_m,p}(r)\).
It can be easily deduced from the preceding theorem and Theorem 1.2 that, in the case \(1 \le q_1, \dots , q_m \le 2\) and \(1 \le p < \infty \), we have \(k_{{\vec {q}},p}(r)<\infty \) if and only if \(k_{q_i,p}(r) < \infty \) for all \(1 \le i\le m\). Then, in virtue of the monotonicity of \(k_{q_i,p}(r)\) as a function of \(q_i\) (see Theorem 1.1), it follows that \(k_{{\vec {q}},p}(r) < \infty \) if and only if \(k_{\varvec{q},p}(r) < \infty \). Note that, as pointed out in the Introduction (see Proposition 1.7 and the paragraph above), this equivalence does not hold in general, as we will see in Sect. 4.
Proof of Theorem 3.1
Suppose first that \(k_{{\vec {q}},p}(r) < \infty \) and (ii) is not satisfied, and let us see that either (i), (iii) or (iv) is satisfied. By Proposition 1.4 we have \(k_{q_i,p}(r) < \infty \) for all \(1 \le i \le m\) and, in particular, \(k_{\varvec{q},p}(r) < \infty \). If \(\varvec{q}\le p\), we have \(1 \le \varvec{q}\le p < \infty \) and \(k_{\varvec{q},p}(r)<\infty \) and, by Theorem 1.2, it follows that \(\varvec{q}\le r \le \max \{p,2\}\). If \(\varvec{q}> p\), then we have \(1\le p < \varvec{q}\le 2\) and \(k_{\varvec{q},p}(r)< \infty \) and by Theorem 1.2 it follows that \(r=2\) or \(\varvec{q}<r\le 2 = \max \{p,2\}\). We deduce from the cases above that either \(r=2\) or \(\varvec{q}< r \le \max \{p,2\}\) or \(\varvec{q}=r\) and \(\varvec{q}\le p\). For the converse, assume the following facts.
-
(a)
If \(1 \le p, \varvec{q}<r<2\) or \(1 \le p, \varvec{q}\le r=2\), then \(k_{{\vec {q}},p}(r) = k_{q_1,p}(r)\cdots k_{q_m,p}(r) < \infty \).
-
(b)
If \(\varvec{q}\le r \le p\), then \(k_{{\vec {q}},p}(r) = 1 = k_{q_1,p}(r)\cdots k_{q_m,p}(r)\).
If (i) holds then, whether \(p<2\) or \(p\ge 2\), \(k_{{\vec {q}},p}(r) = k_{q_1,p}(r)\cdots k_{q_m,p}(r)<\infty \) in virtue of (a) and (b). If (ii) holds then by Proposition 2.7 we have \(k_{{\vec {q}},p}(r) = 1 = k_{q_1,p}(r)\cdots k_{q_m,p}(r)\). Suppose we are under the hypothesis of (iii). If \(p<r\) then \(\varvec{q}< r \le \max \{p,2\}= 2\) and hence \(k_{{\vec {q}},p}(r) = k_{q_1,p}(r)\cdots k_{q_m,p}(r) < \infty \) by (a), while if \(p \ge r\) then \(\varvec{q}< r \le p\) and, consequently, \(k_{{\vec {q}},p}(r) = 1 = k_{q_1,p}(r)\cdots k_{q_m,p}(r)\) by (b). Finally, if (iv) holds then \(k_{{\vec {q}},p}(r) = 1 = k_{q_1,p}(r)\cdots k_{q_m,p}(r)\) by (b). Then, it only remains to prove (a) since the assertion in (b) was already stated in Remark 2.3. Fix \(T:L^{q_1}(\mu _1)\times \cdots \times L^{q_m}(\mu _m) \rightarrow L^p(\nu )\) and functions \(\{f_{k_i}^i\}_{k_i=1}^n \subset L^{q_i}(\mu _i)\), \(1\le i \le m\). By Lemma 2.4 and the m-linearity of T we have
Applying Fubini and the boundedness of T we obtain,
Then, it will be suffice to obtain upper bounds for each \(\int _0^1 \left\| \sum _{k_i=1}^n f_{k_i}^i w_{k_i}(t_i) \right\| ^p_{L^{q_i}(\mu _i)} dt_i\). On the one hand, for those \(q_i<p\) we have
On the other hand, if \(p \le q_i\)
These inequalities together with (27) give
where \(q_{j_1}, \dots , q_{j_{m_2}}\) are those \(q_i\ge p\). Noting that \(k_{q_i,p}(r) =1\) if \(q_i<p\) and \(k_{q_i,p}(r)=\frac{c_{r,q_i}}{c_{r,p}}\) if \(q_i \ge p\) (see Theorem 1.2), the last inequality gives \(k_{{\vec {q}},p}(r) \le k_{q_1,p}(r) \cdots k_{q_m,p}(r) < \infty \). The equality holds as a consequence of Proposition 1.4 and this proves (a). \(\square \)
Note that in the proof of the statement (a) when \(r<2\), the hypothesis \(p, \varvec{q}< r\) is necessary in order to apply (15). When \(r=2\), there is no need of this hypothesis to see, with the same proof, that \(k_{{\vec {q}},p}(2)< \infty \) (we will point out this fact in Proposition 5.3 below). In this case, we need the assumption \(p, \varvec{q}\le 2\) to ensure \(k_{{\vec {q}},p}(2)=k_{q_1,p}(2) \cdots k_{q_m,p}(2)\).
3.2 Proof of the Main Theorem
We prove now Theorem 1.3, which determines the set of those triples \((p,{\vec {q}},r)\) satisfying \(k_{{\vec {q}},p}(r)<\infty \), with the exception of the cases \(p=\infty \) (partially discussed in Sect. 4) and \(2 \le p < \max \{q_1, \dots , q_m\}\). At this point, the results stated in the theorem are simple consequences of those obtained in the previous sections.
Proof of Theorem 1.3
We begin by proving (i). Items (ia) and (ic) are consequences of Theorem 3.1, then we only need to prove (ib). If \(k_{{\vec {q}},p}(r)< \infty \) then \(k_{\varvec{q},p}(r)< \infty \) (by Proposition 1.4) and by Theorem 1.2(ii) it follows that \(2\le r \le p\). For the converse, take \(2 \le r \le p\). If \(\varvec{q}\le r \le p\) then \(k_{{\vec {q}},p}(r) = 1\) by Remark 2.3, while if \(2 \le r \le \varvec{q}\) then \(k_{{\vec {q}},p}(r) < \infty \) by Corollary 2.6, since \(k_{{\vec {q}},p}(2)<\infty \) (this is the particular case of the Marcinkiewicz-Zygmund inequalities proved in [5, 18]) and \(k_{{\vec {q}},p}(\varvec{q}) = 1 < \infty \).
Now we prove (ii). Item (iia) follows from Theorem 3.1. For (iib) (resp. (iic)) note that \(k_{{\vec {q}},p}(r)< \infty \) implies \(k_{\varvec{q},p}(r)< \infty \) and, by Theorem 1.2(iii), this gives \(2 \le r <p\) (resp. \(r=2\)). The converse in (iic) and the sufficient condition in (iib) are again consequences of the multilinear Marcinkiewicz-Zygmund inequality for \(r=2\) obtained in [5, 18]. \(\square \)
4 Proof of Proposition 1.7
The key tool in this section is a variant of a Kahane-Salem-Zygmund inequality. The idea is to construct a multilinear operator from \(\ell ^{q_1}_n \times \cdots \times \ell ^{q_m}_n\) into \(\ell ^{\infty }_n\) with relatively small supremum norm but for which \(\left( \sum _{i_1, \dots , i_m=1}^n |T(e_{i_1}, \dots , e_{i_m})|^r \right) ^{1/r}\) has a relatively large \(\ell ^\infty \)-norm. This, together with an appropriate choice of the indices \(q_1, \dots , q_m, r\), will force \(\lim _n k_{{\vec {q}}, \infty }^{(n)}(r) = \infty \) as desired. In [4, Thm. 4], Boas showed that there exists an \((m+1)\)-linear map \(T :\ell ^{2}_n \times \cdots \times \ell ^{2}_n \rightarrow {\mathbb {R}}\) of the form
where \(\varepsilon _{j_1, \dots , j_{m+1}}=\pm 1\), such that \(\Vert T\Vert _{\ell ^{2}_n \times \cdots \times \ell ^{2}_n \rightarrow {\mathbb {R}}}\prec n^{1/2}\) (here, \(\prec \) means that there exist a constant \(C_m>0\), depending only on m, such that \(\Vert T\Vert _{\ell ^{2}_n \times \cdots \times \ell ^{2}_n \rightarrow {\mathbb {R}}}\le C_m n^{1/2}\)). The restriction of this operator to \(\ell ^{1}_n \times \cdots \times \ell ^{1}_n\) has norm one and hence, by an interpolation argument, if \(1<q<2\) (note that \(\frac{1}{q}= \frac{1 - \theta }{1} + \frac{\theta }{2}\) with \(\theta =\frac{2}{q'}\)) then \(T :\ell ^{q}_n \times \cdots \times \ell ^{q}_n \rightarrow {\mathbb {R}}\) has norm \(\Vert T\Vert _{ \ell ^{q}_n \times \cdots \times \ell ^{q}_n \rightarrow {\mathbb {R}}}\prec 1^{1-\theta } n^{\frac{\theta }{2}} = n^{\frac{1}{q'}}\). It is clear then that, if \(1 \le q_1, \dots , q_{m+1} \le 2\) and \(\varvec{q}=\max \{q_1, \dots , q_{m+1}\}\), \(T:\ell ^{q_1}_n \times \cdots \times \ell ^{q_{m+1}}_n \rightarrow {\mathbb {R}}\) has norm \(\Vert T\Vert _{ \ell ^{q_1}_n \times \cdots \times \ell ^{q_{m+1}}_n \rightarrow {\mathbb {R}}} \le \Vert T\Vert _{ \ell ^{\varvec{q}}_n \times \cdots \times \ell ^{\varvec{q}}_n \rightarrow {\mathbb {R}}} \prec n^{\frac{1}{\varvec{q}'}}\). If, instead, \(1\le q_1, \dots , q_{m+1} \le \infty \) with \(q_i>2\) for at least one \(1 \le i \le m+1\), we can take \(T :\ell ^{2}_n \times \cdots \times \ell ^{2}_n \rightarrow {\mathbb {R}}\) as in (28) and compose it with the identities \(id^n_{q_i,2} :\ell ^{q_i}_n \rightarrow \ell ^2_n\).
Note that \(id^n_{q_i,2}\) has norm one if \(q_i \le 2\) and norm \(n^{\frac{1}{2}-\frac{1}{q_i}}\) if \(q_i>2\). As a consequence of the previous observations and the isometric correspondence between m-linear maps \( \ell ^{q_1}_n \times \cdots \times \ell ^{q_{m}}_n \rightarrow \ell ^{q_{m+1}'}_n\) and \((m+1)\)-linear maps \( \ell ^{q_1}_n \times \cdots \times \ell ^{q_{m+1}}_n \rightarrow {\mathbb {R}}\), we obtain the following lemma.
Lemma 4.1
Let \(m,n \in {\mathbb {N}}\), \(1 \le q_1, \dots , q_m, p \le \infty \) and \(\varvec{q}= \max \{q_1, \dots , q_m\}\). Then there exists an m-linear map \(T :\ell ^{q_1}_n \times \cdots \times \ell ^{q_m}_n \rightarrow \ell ^{p}_n\) of the form
where \(\varepsilon _{j_1, \dots , j_{m+1}}=\pm 1\), such that
Proof of Proposition 1.7
Our goal is to show that \(k_{{\vec {q}}, \infty }^{(n)}(r) \succ n^{\gamma }\) for some \(\gamma >0\) independent of n, in which case \(k_{{\vec {q}}, \infty }(r)=\lim _n k_{{\vec {q}}, \infty }^{(n)}(r) = \infty \). In virtue of Lemma 4.1, there is an m-linear operator \(T:\ell ^{q_1}_{n} \times \cdots \times \ell ^{q_m}_n \rightarrow \ell ^\infty _n\) as in (29) such that
where \(\sum _{q_i>2} \left( \frac{1}{2}-\frac{1}{q_i} \right) =0\) if \(\varvec{q}\le 2\). Now, since \(|T(e_{i_1}, \dots , e_{i_m})| = (1, \dots , 1)\) for all \(1 \le i_1, \dots , i_m \le n\) we have
Then,
The statement follows from this inequality. \(\square \)
5 Applications
5.1 Weighted Vector-Valued Estimates for Multilinear Calderón-Zygmund Operators
The study of multilinear Calderón-Zygmund theory finds its origins in the seventies with works of Coifman and Meyer, but a systematic treatment of this topic appears later with works of Grafakos and Torres [19, 20]. Recall the definition of a multilinear Calderón-Zygmund operator: let \(T:S({\mathbb R}^n)\times \dots \times S({\mathbb R}^n)\rightarrow S'({\mathbb R}^n)\) be a multilinear operator initially defined on the m-fold product of Schwartz spaces and taking values into the space of tempered distributions; we say that T is an m-linear Calderón-Zygmund operator if, for some \(1\le q_1, \dots , q_m<\infty \) and \(\frac{1}{m} \le p < \infty \) satisfying \(\frac{1}{p}=\frac{1}{q_1}+\dots +\frac{1}{q_m}\), it extends to a bounded multilinear operator from \(L^{q_1}\times \dots \times L^{q_m}\) to \(L^p\), and if there exists a function K defined off the diagonal \(x=y_1=\dots =y_m\) in \(({\mathbb R}^n)^{m+1}\) satisfying the appropriate decay and smoothness conditions (see [19, 20]) and such that
for all \(x \notin \cap _{i=1}^{m} \mathrm{supp}f_i\). In [19] it was shown that, if \(\frac{1}{p}=\frac{1}{q_1}+\dots +\frac{1}{q_m}\) with \(1< q_1, \dots , q_m < \infty \), then an m-linear Calderón-Zygmund operator T maps \(L^{q_1}({\mathbb {R}}^n)\times \dots \times L^{q_m}({\mathbb {R}}^n)\) into \(L^p({\mathbb {R}}^n)\). If \(1 \le q_1, \dots , q_m < \infty \) and at least one \(q_i=1\), then T maps \(L^{q_1}({\mathbb {R}}^n)\times \dots \times L^{q_m}({\mathbb {R}}^n)\) into \(L^{p, \infty }({\mathbb {R}}^n)\). Regarding the weighted norm inequalities, the first result was obtained in [20] (see also [25]) where the authors proved that, if \(1< q_1, \dots , q_m < \infty \) and w is a weight in the Muckenhoupt \(A_{q_0}\) class for \(q_0=\min \{q_1, \dots , q_m\}\), an m-linear Calderón-Zygmund operator T maps \(L^{q_1}(w)\times \dots \times L^{q_m}(w)\) into \(L^p(w)\). The same approach of [20] shows that T maps \(L^{q_1}(w_1)\times \dots \times L^{q_m}(w_m) \) into \( L^{p}(\nu _{\mathbf {w}})\), where \(\nu _{\mathbf {w}}=\prod _{i=1}^m w_i^{p/q_i}\) and \(w_i\) is in \(A_{q_i}\). As expected, if at least one \(q_i=1\), then the weak endpoint estimate holds. In [23], the authors developed the right class of multiple weights for m-linear Calderón-Zygmund operators. We briefly review the weighted estimate proved in there for multilinear Calderón-Zygmund operators. As usual, let \(1\le q_1, \dots , q_{m}<\infty \) and \(\frac{1}{m} \le p < \infty \) be such that \(\frac{1}{p}=\frac{1}{q_1}+\dots +\frac{1}{q_m}\). We say that \(\mathbf {w}=(w_1,\dots ,w_m)\) satisfies the multilinear\(A_{{\vec {q}}}\)condition if
where the supremum is taken over all cubes Q (when \(q_i=1\), \(\Big (\frac{1}{|Q|}\int _Q w_i^{1-q'_i}\Big )^{1/q'_i}\) is understood as \(\displaystyle (\inf _Q w_i)^{-1}\)). Now, if \(\mathbf {w}\) satisfies the \(A_{{\vec {q}}}\) condition and \(1<q_1,\dots ,q_m<\infty \), then an m-linear Calderón-Zygmund operator T maps \(L^{q_1}(w_1)\times \dots \times L^{q_m}(w_m) \) into \( L^{p}(\nu _{\mathbf {w}})\). If at least one \(q_i=1\), then T maps \(L^{q_1}(w_1)\times \dots \times L^{q_m}(w_m) \) into \( L^{p,\infty }(\nu _{\mathbf {w}})\). It is shown that \(\prod _{i=1}^m A_{q_i} \subseteq A_{{\vec {q}}}\) and that this inclusion is strict; then, the above weighted estimates improve those obtained in [20]. In fact, if T is the m-linear Riesz Transform, it was proved in [23] that \(A_{{\vec {q}}}\) is a necessary condition for such weighted estimate of T.
5.1.1 Some Known Estimates and Their Comparison with Marcinkiewicz-Zygmund Inequalities
In the context of vector-valued inequalities for multilinear Calderón-Zygmund operators, the following result was proved in [8, Corollary 3.3] by means of extrapolation techniques. See also [9, Section 6.4] where this type of inequalities are obtained in the more general context of rearrangement invariant quasi-Banach function spaces.
Theorem 5.1
[8] Let T be a multilinear Calderón-Zygmund operator, \(1 \le q_1, \dots , q_m < \infty \), \(1< r_1, \dots , r_m < \infty \) and \(0<p,r < \infty \) such that
If \(1<q_1, \dots , q_m < \infty \) and \(w \in A_{q_0}\) for \(q_0=\min \{q_1, \dots , q_m\}\), then
If at least one \(q_i=1\) and \(w \in A_1\), then
The estimate (32) was obtained independently in [18] as a particular case of the inequality (34) for m-tuples of weights. Also a weaker version of (33) was obtained as a consequence of a multilinear Marcinkiewicz-Zygmund inequality for \(r=2\).
Theorem 5.2
[18] Let T be a multilinear Calderón-Zygmund operator, \(1 \le q_1, \dots , q_m < \infty \), \(1< r_1, \dots , r_m < \infty \) and \(0<p,r < \infty \) such that
If \(1< q_1, \dots , q_m < \infty \) and \((w_1^{q_1}, \dots , w_{m}^{q_m}) \in A_{q_1} \times \cdots \times A_{q_m}\), then
If at least one \(q_i=1\) and \(w \in A_1\), then
It is worth mentioning that, in the above weighted vector-valued estimates, the multilinear \(A_{{\vec {q}}}\) condition is not considered. We will consider this appropriate class of multiple weights in Corollary 5.4, where vector-valued inequalities for Calderón-Zygmund operators are obtained. Also, as we emphasize below, the vector-valued inequalities that we obtain are significantly different from those stated in (32), (33), (34). In our case, the sum on the left-hand side is replaced by a multi-indexed sum over \(k_1, \dots , k_m\) and we consider only one power r instead of the powers \(1<r_1, \dots , r_m < \infty \).
At this point we would like to stress the relation \( \frac{1}{r}=\frac{1}{r_1}+\cdots +\frac{1}{r_m}\) that appears in the hypotheses of the previous theorems which, by the way, shows why the estimate (35) is weaker than (33). When proving estimates like (34), one is interested in the study of the following inequalities:
If in the previous inequality we put \(f^i_k=f^i \in L^{q_i}(\mu _i)\) for \(1 \le k \le n\), \(1\le i \le m\) and \(f^i_k=0\) otherwise, we obtain
and, since C is independent of n, this forces \( \frac{1}{s} \le \frac{1}{r_1}+\cdots +\frac{1}{r_m}. \) Thus, the estimate (36) is optimal when \(s=r\), where r satisfies \(\frac{1}{r} = \frac{1}{r_1}+\cdots +\frac{1}{r_m}\). In the case of (35) the relation between the powers is not optimal as in (33). In fact, if (33) holds for \(r=2\) then \(2 \le r_i\) for all \(1 \le i \le m\) and, since \(\ell ^2 \hookrightarrow \ell ^{r_i}\) with \(\Vert \cdot \Vert _{\ell ^{r_i}} \le \Vert \cdot \Vert _{\ell ^2}\), then
which yields the (hence, weaker) estimate in (35).
Now, the Marcinkiewicz-Zygmund inequalities we were studying in the previous sections have a significant difference with those in (36). In the former ones the sum
ranges over the indices \(k_1, \dots , k_m\), while in (36) we sum over the diagonal, hence only one index k. This produces another optimal relation between the powers \(r, r_1, \dots , r_m\). Indeed, suppose we are interested in estimates of the form
If we choose \(f^i_{k_i} = f^i \in L^{q_i}(\mu _i)\) for \(1 \le k_i \le n\), \(1 \le i \le m\) and \(f^i_{k_i}=0\) otherwise, we have
from where we get \(\frac{m}{s} \le \frac{1}{r_1}+\cdots +\frac{1}{r_m}\). Also, if in (38) we take any sequence \(\{f^1_{k_1}\}_{k_1} \subset L^{q_1}(\mu _1)\) and, for each \(2 \le i \le m\), we put \(f^i_1=f^i \in L^{q_i}(\mu _i)\) and \(f^i_{k_i}=0\) for \(k_i\ge 2\), then we obtain
where \(\tilde{T}:L^{q_1}(\mu _1) \rightarrow L^p(\nu )\) is the linear operator \(\tilde{T}(\cdot ) = T(\cdot , f^2, \dots , f^m)\). In virtue of (37) (when \(m=1\)) this last inequality gives \(\frac{1}{s}\le \frac{1}{r_1}\). Analogously, \(\frac{1}{s}\le \frac{1}{r_i}\) for all \(1 \le i \le m\). In sum, when we are dealing with inequality (38) we have the following conditions over the powers \(s, r_1, \dots , r_m\),
This shows that the inequality (38) is optimal when \(s=r_1=\cdots = r_m=r\), which is just the case treated in the Marcinkiewicz-Zygmund inequalities.
Camil Muscalu pointed out to us that, in the unweighted case, Marcinkiewicz-Zygmund estimates for multilinear Calderón-Zygmund operators can be deduced from more general multiple vector-valued inequalities obtained in his recent works with Cristina Benea. In [1, 2], the authors developed a powerful method, which they called the helicoidal method, that allows to obtain vector-valued inequalities in harmonic analysis. They apply this method to obtain vector-valued inequalities for paraproducts (which can be regarded as bilinear multiplier operators) and the bilinear Hilbert transform. We also refer to [7, 15] where, with different methods, the authors obtain vector-valued inequalities for multilinear multiplier operators. As far as we know, Marcinkiewicz-Zygmund inequalities for multilinear Calderón-Zygmund operators were not addressed in the weighted case. Also, although it might be possible to derive similar estimates via the helicoidal method, our approach is completely different.
5.1.2 Weighted Vector-Valued Estimates
We state now as a proposition a result that is partially demonstrated in the proof of Theorem 3.1.
Proposition 5.3
Let \(0<p, q_1, \dots , q_m< r<2\) or \(r=2\) and \(0<p, q_1, \dots , q_m < \infty \) and, for each \(1 \le i \le m\), consider \(\{f^i_{k_i}\}_{k_i} \subset L^{q_i}(\mu _i)\). Then, there exists a constant \(C>0\) such that the following estimates hold.
-
(i)
If \(T:L^{q_1}(\mu _1) \times \cdots \times L^{q_m}(\mu _m) \rightarrow L^p(\nu )\) then
$$\begin{aligned} \left\| \left( \sum _{k_1, \dots , k_m} |T(f^1_{k_1}, \dots , f^m_{k_m})|^r \right) ^{\frac{1}{r}} \right\| _{L^p(\nu )} \le C \Vert T\Vert \prod _{i=1}^m \left\| \left( \sum _{k_i} |f^i_{k_i}|^{r} \right) ^{\frac{1}{r}} \right\| _{L^{q_i}(\mu _i)}. \end{aligned}$$(39) -
(ii)
If \(T:L^{q_1}(\mu _1) \times \cdots \times L^{q_m}(\mu _m) \rightarrow L^{p,\infty }(\nu )\) then
$$\begin{aligned}&\left\| \left( \sum _{k_1, \dots , k_m} |T(f^1_{k_1}, \dots , f^m_{k_m})|^r \right) ^{\frac{1}{r}} \right\| _{L^{p, \infty }(\nu )} \nonumber \\&\quad \le C e^{\frac{1}{p}} \Vert T\Vert _{weak} \prod _{i=1}^m \left\| \left( \sum _{k_i} |f^i_{k_i}|^{r} \right) ^{\frac{1}{r}} \right\| _{L^{q_i}(\mu _i)}. \end{aligned}$$(40)
Proof of Theorem 1.3
The inequality (39) was proved in Theorem 3.1 for \(1 \le p, q_1, \dots , q_m< r<2\) (see the assertion (a) inside the proof). It is easy to check that, with exactly the same proof, the statement remains valid for \(0<p, q_1, \dots , q_m< r<2\). The case \(r=2\) also follows the same proof, with slight modifications in the involved constants. See the case \(r=2<p\) in Lemma 2.4 (see also [18, Thm. 6 (a)]). Then, we only need to prove (ii) and, for this purpose, we follow ideas of [18, Thm. 6 (b)] where the particular case \(r=2\) is addressed. Recall that, for \(0<s<p<\infty \),
Then,
Now for each measurable set E of positive and finite \(\nu \)-measure, consider \(T_E:L^{q_1}(\mu _1) \times \cdots \times L^{q_m}(\mu _m) \rightarrow L^s(\nu )\) defined by
(the fact that \(T_E\) takes values in \(L^s(\nu )\) follows from (41)). In virtue of the second inequality in (41) we have
and, consequently, \(\nu (E)^{\frac{1}{p}-\frac{1}{s}} \Vert T_E\Vert \le \left( \frac{p}{p-s}\right) ^{\frac{1}{s}} \Vert T\Vert _{weak}\). Going back to (42), we obtain
and since \(0<s<p\) was arbitrary, letting \(s \rightarrow 0\) we obtain the desired estimate. \(\square \)
As a consequence of the previous proposition, we obtain the following vector-valued estimates for multilinear Calderón-Zygmund operators which should be compared with those of Theorems 5.1 and 5.2.
Corollary 5.4
Let T be a multilinear Calderón-Zygmund operator, \(1 \le q_1, \dots , q_m<r < 2\) (or \(r=2\) and \(1 \le q_1, \dots , q_m <\infty \)) and \(p>0\) such that \( \frac{1}{p}=\frac{1}{q_1}+\cdots +\frac{1}{q_m}. \) Suppose \(\mathbf {w}=(w_1,\dots ,w_m)\) satisfies the multilinear \(A_{{\vec {q}}}\) condition and consider \(\nu _{\mathbf {w}}=\prod _{i=1}^m w_i^{p/q_i}\). Then, there exists a constant \(C>0\) such that the following estimates hold.
-
(i)
If \(q_i>1\) for all \(1 \le i \le m\), then
$$\begin{aligned} \left\| \left( \sum _{k_1, \dots , k_m} |T(f^1_{k_1}, \dots , f^m_{k_m})|^r \right) ^{\frac{1}{r}} \right\| _{L^p(\nu _{\mathbf {w}})} \le C \Vert T\Vert \prod _{i=1}^m \left\| \left( \sum _{k_i} |f^i_{k_i}|^{r} \right) ^{\frac{1}{r}} \right\| _{L^{q_i}(w_i)}. \end{aligned}$$(43) -
(ii)
If at least one \(q_i=1\), then
$$\begin{aligned}&\left\| \left( \sum _{k_1, \dots , k_m} |T(f^1_{k_1}, \dots , f^m_{k_m})|^r \right) ^{\frac{1}{r}} \right\| _{L^{p,\infty }(\nu _{\mathbf {w}})} \nonumber \\&\quad \le C e^{\frac{1}{p}} \Vert T\Vert _{weak} \prod _{i=1}^m \left\| \left( \sum _{k_i} |f^i_{k_i}|^{r} \right) ^{\frac{1}{r}} \right\| _{L^{q_i}(w_i)}. \end{aligned}$$(44)
5.2 Marcinkiewicz-Zygmund Inequalities for Positive Multilinear Operators
Following the proof of the Marcinkiewicz-Zygmund inequality for positive linear operators stated in [16, Chapter V.1, Thm. 1.12], we prove Proposition 1.9 which extends this result to the multilinear setting. Recall that a multilinear operator \(T:L^{q_1}(\mu _1)\times \cdots L^{q_m}(\mu _m) \rightarrow L^p(\nu )\) is positive if \(f^1, \dots , f^m \ge 0\) implies \(T(f^1, \dots , f^m) \ge 0\). It can be seen that if T is positive, then \(|T(f^1, \dots , f^m)| \le T(g^1, \dots , g^m)\) whenever \(|f^1|\le g^1, \dots , |f^m|\le g^m\).
Proof of Proposition 1.9
It suffices to show that
for any choice of functions \(\{f_{k_i}^i\}_{k_i=1}^{n_i} \subset L^{q_i}(\mu _i)\), \(1\le i \le m\). The case \(r=\infty \) is immediate by the positivity of T. Then, we assume \(1 \le r < \infty \) and prove (45) by induction on m. The case \(m=1\) is proved in [16, Chapter V.1, Thm. 1.12] via a duality argument. Now, let \(m\ge 2\) and suppose (45) holds for \(m-1\). By duality we know that, given \(\mathbf {a}=(a_{k_1 \dots k_m})_{k_1, \dots , k_m=1}^\infty \),
where the supremum is taken over all \(\mathbf {b} = (b_{k_1 \dots k_m})_{k_1, \dots , k_m}\) such that \(\Vert \mathbf {b}\Vert _{\ell ^{r'}({\mathbb {N}}\times \cdots \times {\mathbb {N}})}\le 1\). Take any \(\mathbf {b} \in \ell ^{r'}({\mathbb {N}}\times \cdots \times {\mathbb {N}})\) and note that
where the last inequality follows from Hölder’s inequality and the positivity of T. A repeated application of H\(\ddot{\text {o}}\)lder’s inequality and the induction hypothesis shows that
Hence,
and taking supremum over all \(\Vert \mathbf {b}\Vert _{\ell ^{r'}({\mathbb {N}}\times \cdots \times {\mathbb {N}})} \le 1\) we get (45). \(\square \)
As an immediate consequence of Proposition 1.9 and Young’s inequality, we have the following vector-valued inequality for the convolution \(f*g(x) = \int _{\mathbb {R}}f(x-y) g(y) \, dy\).
Corollary 5.5
Let \(1 \le q_1,q_2,p \le \infty \) satisfying \(\frac{1}{q_1} + \frac{1}{q_2} = \frac{1}{p} + 1\) and let \(1 \le r \le \infty \). Then,
for any choice of functions \(\{f_{k_i}^i\}_{k_i=1}^{n_i} \subset L^{q_i}({\mathbb {R}})\), \(1\le i \le 2\).
The inequality (8) should be compared with [13, Thm. 6.2], where it is proved that, given a positive m-linear operator \(T:X_1 \times \cdots \times X_m \rightarrow Y\) between Banach lattices and \(1 \le r_1, \dots , r_m, r \le \infty \) such that \(\frac{1}{r}=\frac{1}{r_1} + \cdots + \frac{1}{r_m}\), we have
for any choice of sequences \(\{x_{k}^i\}_{k=1}^{n} \subset X_i\), \(1\le i \le m\). As a consequence, if \(1 \le q_1,q_2,p \le \infty \) satisfy \(\frac{1}{q_1} + \frac{1}{q_2} = \frac{1}{p} + 1\) and \(1 \le r_1, r_2, r \le \infty \) are such that \(\frac{1}{r}=\frac{1}{r_1} + \frac{1}{r_2}\), then
for any choice of functions \(\{f_{k}^i\}_{k=1}^{n} \subset L^{q_i}({\mathbb {R}})\), \(1\le i \le 2\).
References
Benea, C., Muscalu, C.: Multiple vector valued inequalities via the helicoidal method. Anal. PDE 9, 1931–1988 (2016)
Benea, C., Muscalu, C.: Quasi-Banach valued inequalities via the helicoidal method. Preprint, arXiv:1609.01090
Bergh, J., L\(\ddot{\text{o}}\)fstr\(\ddot{\text{ o }}\)m, J.: Interpolation Spaces. An Introduction. Grund- lehren der mathematischen Wissenschaften, vol. 223. Springer, Berlin (1976)
Boas, H.: Majorant series. J. Korean Math. Soc. 37(2), 321–337 (2000)
Bombal, F., Pérez-García, D., Villanueva, I.: Multilinear extensions of Grothendieck’s theorem. Quart. J. Math. 55(4), 441–450 (2004)
Bohnenblust, H.F., Hille, E.: On the absolute convergence of Dirichlet series. Ann. Math. 32(3), 600–622 (1931)
Culiuc, A., Di Plinio, F., Ou, Y.: Domination of multilinear singular integrals by positive sparse forms. Preprint, arXiv:1603.05317
Cruz-Uribe, D., Martell, J.M., Pérez, C.: Extrapolation from \(A_{\infty }\) weights and applications. J. Funct. Anal. 213(2), 412–439 (2004)
Curbera, G., García-Cuerva, J., Martell, J.M., Pérez, C.: Extrapolation with weights, rearrangement-invariant function spaces, modular inequalities and applications to singular integrals. Adv. Math. 203, 256–318 (2006)
Davie, A.M.: Quotient algebras of uniform algebras. J. London Math. Soc. 7(2), 31–40 (1973)
Defant, A., Floret, K.: Tensor Norms and Operator Ideals. North-Holland Mathematics Studies, vol. 176. North-Holland Publishing Co., Amsterdam (1993)
Defant, A., Junge, M.: Best constants and asymptotics of Marcinkiewicz-Zygmund inequalities. Studia Math. 125(3), 271–287 (1997)
Defant, A., Mastyło, M.: Interpolation of Fremlin tensor products and Schur factorization of matrices. J. Funct. Anal. 262, 3981–3999 (2012)
Defant, A., Sevilla-Peris, P.: A new multilinear insight on littlewood’s 4/3-inequality. J. Funct. Anal. 256(5), 1642–1664 (2009)
Di Plinio, F., Ou, Y.: Banach-valued multilinear singular integrals. Preprint, arXiv:1506.05827. To appear in Indiana Univ. Math. J
García-Cuerva, J., Rubio de Francia, J.L.: Weighted Norm Inequalities and Related Topics. North-Holland Mathematics Studies, vol. 116. North-Holland Publishing Co., Amsterdam (1985)
Gasch, J., Maligranda, L.: On vector-valued inequalities of Marcinkiewicz-Zygmund, Herz and Krivine type. Math. Nachr. 167, 95–129 (1994)
Grafakos, L., Martell, J.M.: Extrapolation of weighted norm inequalities for multivariable operators and applications. J. Geom. Anal. 14(1), 19–46 (2004)
Grafakos, L., Torres, R.H.: Multilinear Calderón-Zygmund theory. Adv. Math. 165(1), 124–164 (2002)
Grafakos, L., Torres, R.H.: Maximal operator and weighted norm inequalities for multilinear singular integrals. Indiana Univ. Math. J. 51(5), 1261–1276 (2002)
Herz, C.: Theory of \(p\)-spaces with an application to convolution operators. Trans. Am. Math. Soc. 154, 69–82 (1971)
Kaijser, S.: Some results in the metric theory of tensor products. Studia Math. 63(2), 157–170 (1978)
Lerner, A., Ombrosi, S., Pérez, C., Torres, R., Trujillo-González, R.: New maximal functions and multiple weights for the multilinear Calderón-Zygmund theory. Adv. Math. 220(4), 1222–1264 (2009)
Marcinkiewicz, J., Zygmund, A.: Quelques inégalités pour les opérations linéaires. Fund. Math. 32, 113–121 (1939)
Pérez, C., Torres, R.H.: Sharp maximal function estimates for multilinear singular integrals. Contemp. Math. 320, 323–333 (2003)
Pietsch, A.: Operator Ideals. North-Holland Publishing Co., Amsterdam (1980)
Quefflec, H.H.: Bohr’s vision of ordinary Dirichlet series; old and new results. J. Anal. 3, 43–60 (1995)
Acknowledgements
The authors wish to thank C. Muscalu for his valuable comments regarding this work. This project was supported in part by CONICET PIP 11220130100329CO, ANPCyT PICT 2015-2299 and UBACyT 20020130100474. The second author has a postdoctoral position from CONICET.
Author information
Authors and Affiliations
Corresponding author
Additional information
Communicated by Rodolfo H. Torres.
Rights and permissions
About this article
Cite this article
Carando, D., Mazzitelli, M. & Ombrosi, S. Multilinear Marcinkiewicz-Zygmund Inequalities. J Fourier Anal Appl 25, 51–85 (2019). https://doi.org/10.1007/s00041-017-9563-5
Received:
Revised:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s00041-017-9563-5