Endpoint uniform estimates ($p_i=\infty$ and BMO) for the Muscalu-Tao-Thiele paraproduct family

The uniform paraproduct estimates of Muscalu, Tao and Thiele (2002) bound a general family of multilinear paraproduct forms and their maximal variants, with constants independent of all integer frequency-localization parameters (polytope offsets and relative scale shifts), but only for finite Lebesgue exponents $1<p_i<\infty$; the paper explicitly declines to treat the endpoints $p_i=\infty$ and BMO. Bernicot (2009) partially resolved this endpoint question: he proved endpoint bounds with $L^\infty_c$ and Hardy-space slots and a BMO target, and established parameter-uniform bounds under an explicit homogeneity condition ($\gamma\ge 1/2$, or all exponents finite), conditions that cannot hold in the all-infinite configuration in which the target is BMO, whose parameter-uniformity therefore remains open; he also showed by an explicit bilinear counterexample that uniform $L^\infty$-slot bounds are impossible in the complementary extreme configuration. The problem that remains open is to complete the characterization: determine whether the known sufficient conditions are necessary (in particular the regime $0<\gamma<1/2$), whether uniform bounds hold with full $L^\infty$ and BMO input slots, whether the BMO-target bound is uniform in the parameters, and the endpoint behaviour of the maximal-operator versions.

Background

A paraproduct is a multilinear operator built from a frequency decomposition: if $Q_{j,i}$ denotes a Fourier multiplier (an operator acting on a function by multiplying its Fourier transform by a fixed function called the symbol) whose symbol is a smooth bump function adapted to the frequency ball $\{\xi:\ |\xi|\le 2^j\}$, a degree-$n$ paraproduct form on the real line is a sum over scales of integrals of products $\prod_{i=1}^n Q_{k,i}f_i(x)$, together with a cancellation assumption saying that at least one participating symbol vanishes at the origin. Classical Calderon-Zygmund theory, developed by Coifman and Meyer, shows that such forms are bounded on products of Lebesgue spaces $L^{p_i}(\mathbb R)$ whenever $1<p_i<\infty$ and the homogeneity condition $\sum_i 1/p_i=1$ holds. Here $L^p$ is the usual Lebesgue space and BMO denotes the space of locally integrable functions of bounded mean oscillation, the standard substitute for $L^\infty$ in endpoint estimates, since many operators of this type (for instance the Hilbert transform $Hf(x)=\mathrm{p.v.}\int f(x-y)\,dy/y$) are not bounded on $L^\infty$ but do map $L^\infty$ into BMO.

Muscalu, Tao and Thiele studied how the constants in these estimates depend on the frequency-localization data. Their Theorem 1.1 considers multiplier families $Q_{j,i}$ whose bump symbols are adapted to $\{|\xi|\le 2^j\}$ and vanish at the origin, summed over integer multi-indices $(j_1,\dots,j_n)$ ranging over a set $\Omega\subset\mathbb Z^n$ defined by finitely many constraints $j_{i_\alpha}\ge j_{i'_\alpha}+A_\alpha$ with integer offsets $A_\alpha$ (a convex polytope invariant under translation in the direction $(1,\dots,1)$), and proves the bounds

$\Big|\int_{\mathbb R}\sum_{(j_1,\dots,j_n)\in\Omega}\prod_{i=1}^n Q_{j_i,i}f_i\,dx\Big|\le C_{p,n,K}\prod_{i=1}^n\|f_i\|_{L^{p_i}},$

together with a maximal version in which a supremum over the last scale $j_n$ is taken and the output is measured in $L^{p_n'}$, for all $1<p_i<\infty$ with $\sum_i 1/p_i=1$, with a constant $C_{p,n,K}$ depending only on the exponents, the degree $n$ and the number $K$ of constraints, but not on the integer offsets $A_\alpha$. A companion statement (Corollary 1.2) gives the same uniformity in integer shifts $M_i$ for multipliers adapted to $\{|\xi|\le 2^{j+M_i}\}$ with, at each scale, at least one symbol vanishing at the origin. This uniformity is essential for the authors' subsequent analysis of multilinear singular integrals with a one-dimensional modulation symmetry, a class containing rescaled versions of the bilinear Hilbert transform $B(f,g)(x)=\mathrm{p.v.}\int f(x-y)g(x+y)\,dy/y$, whose $L^p$ theory requires time-frequency analysis (the study of operators by simultaneous decomposition of physical space and frequency space into localized wave packets); there, auxiliary paraproduct-like operators arise with a priori uncontrolled relative frequency scales. The uniformity is genuinely subtle: the multipliers in question satisfy only product-type Marcinkiewicz conditions (symbol estimates bounding each derivative by a product of powers of the single frequencies) with constants independent of the parameters, and Grafakos and Kalton showed that such conditions do not imply boundedness in general, so the polytope structure must be used; the source paper also notes that the estimate fails if the absolute values are moved inside the integral.

The paper proves its estimates only for $1<p_i<\infty$ and states, in the introduction: 'We shall not persue here results for endpoints $p_i=\infty$ or $p_i=$ BMO.' The obstruction is methodological: the proof proceeds by induction over a graph-theoretic decomposition of the polytope, using Holder's inequality and the boundedness of Littlewood-Paley square functions (square sums of frequency-localized pieces of a function) and of the Hardy-Littlewood maximal operator (the map sending a function to the supremum of its averages over intervals containing a given point) on $L^p$; these tools are available exactly for $1<p<\infty$ and fail, or must be replaced by Carleson-measure estimates (estimates for measures on the upper half-space whose mass on each box is controlled by the size of its base, the standard device for controlling BMO) and John-Nirenberg-type arguments (the fundamental exponential-integrability property of BMO functions), at the endpoints. Endpoint results for classical (non-uniform) paraproducts were known from multilinear Calderon-Zygmund theory: continuities with Hardy-space slots $H^{p_i}$ (the substitute for $L^{p_i}$ when $p_i\le 1$), weak-type targets $L^{p,\infty}$, and BMO at the infinite end; the open issue is whether the constants can be made independent of the integer parameters.

Bernicot took up exactly this question. Writing the relative frequency scales as parameters $\lambda_i$ (equivalent to the shifts $2^{-M_i}$) and introducing the anisotropic distance $d_\lambda(\xi,0)=\sum_i|\lambda_i\xi_i|$ on frequency space, he proved endpoint continuities for a symbol class containing the Muscalu-Tao-Thiele paraproducts: slots with $p_i=\infty$ may be taken in $L^\infty_c$ (bounded compactly supported functions), slots with $p_i\le 1$ in Hardy spaces $H^{p_i}$ or $L^1$ with a weak-type target, and when all exponents are infinite the output may be taken in BMO. He proved these bounds are uniform in the parameters under either of two conditions: (a) all $p_i<\infty$, or (b) a homogeneity condition $\gamma\ge 1/2$, where $\gamma$ is the sum of $1/p_j$ over those indices $j$ whose parameter $|\lambda_j|$ is comparable to the maximum of the $|\lambda_l|$; without such a condition his bound may depend on the ratio of the extreme parameters. In the BMO-target configuration all exponents are infinite, so condition (a) fails and $\gamma=0<1/2$ makes (b) fail; the parameter-uniformity of the BMO-target bound is therefore not established by these results, and his negative result is stated only for finite target exponents. Conversely, he gave a bilinear counterexample: admissible operators in the class for which no $L^p\times L^\infty\to L^p$ bound ($1<p<\infty$) with constant uniform in the parameters can hold, because in the degenerating limit the operators converge to the pointwise product $f\cdot H(g)$, whose boundedness would imply the false statement that $H$ maps $L^\infty$ to $L^\infty$. He also states that it is not known whether his uniformity conditions are necessary, and that working with the full space $L^\infty$ rather than $L^\infty_c$ is possible only in particular cases. Later work has extended the uniform technology in other directions (Bellman-function proofs, arguments through a specially constructed convex function that governs extremal dyadic estimates, for generalized and twisted paraproducts, endpoint $p_i=1$ weak-type estimates for classical paraproducts, local BMO bounds for bilinear paraproducts, uniform bounds for convex-set-type multipliers in local $L^2$ ranges), but none of it addresses the remaining endpoint-uniformity questions.

Problem Statement

Characterize completely the endpoint configurations in which the uniform paraproduct estimates of Muscalu, Tao and Thiele extend, with constants still independent of the integer parameters.

The base estimates to be extended are those of Theorem 1.1 and Corollary 1.2 of the source paper: for multipliers whose symbols are bump functions adapted to $\{|\xi|\le 2^{j+M_i}\}$, with at least one symbol vanishing at the origin at each scale, summed over translation-invariant polytopes $\Omega=\{(j_1,\dots,j_n):\ j_{i_\alpha}\ge j_{i'_\alpha}+A_\alpha\}\subset\mathbb Z^n$ with integer offsets $A_\alpha$, the bounds

$\Big|\int_{\mathbb R}\sum_{(j_1,\dots,j_n)\in\Omega}\prod_{i=1}^n Q_{j_i,i}f_i\,dx\Big|\le C\prod_{i=1}^n\|f_i\|_{X_i},\qquad \sum_{i=1}^n\frac{1}{p_i}=1,$

together with the corresponding maximal-operator version (a supremum over one of the scales, output measured in the dual-exponent space), hold for $1<p_i<\infty$ (with $X_i=L^{p_i}$ in the base estimates) with $C$ independent of the offsets $A_\alpha$ and the shifts $M_i$. The problem is to determine, for this family, the complete range of endpoint configurations in which estimates of this type remain true with a uniform constant, where the endpoints in question are: (i) some exponent $p_i=\infty$, i.e., an input slot measured in $L^\infty(\mathbb R)$, or in BMO (the standard endpoint replacement for $L^\infty$), rather than in $L^{p_i}$ with $p_i$ finite; (ii) an $L^{p_i}$ input slot replaced by BMO; (iii) the target of the maximal estimate taken in BMO (formally $p_n=1$, so that the dual exponent is infinite). The uniformity question spans these configurations through shared parameter conditions: whether Bernicot's sufficient conditions (all exponents finite, or $\gamma\ge 1/2$, where $\gamma$ is the sum of $1/p_j$ over those indices $j$ whose frequency parameter $|\lambda_j|$ is comparable to the largest of the $|\lambda_l|$) are necessary has not been determined in any endpoint regime, including the intermediate range $0<\gamma<1/2$.

The question asks, for each endpoint configuration, whether such a parameter-uniform bound holds or fails.

The question retains the full generality of the source remark: input spaces are to be the full $L^\infty$ and BMO (not only the dense subclass of bounded compactly supported functions used in the known partial results), the maximal-operator versions are included, and the constants are to be independent of all integer data (the shifts $M_i$ and the polytope offsets $A_\alpha$). The known partial resolution (uniform bounds for $L^\infty_c$ slots and a BMO target under the conditions 'all exponents finite' or $\gamma\ge 1/2$, and the failure of uniformity when $\gamma=0$ with finite target exponent) is part of the boundary that the complete answer must subsume and delimit, not a restriction of the question.

References

1. Camil Muscalu, Terence Tao, Christoph Thiele, 'Uniform estimates on paraproducts', Journal d'Analyse Mathématique 87 (2002), 369-384. DOI: 10.1007/BF02868481. URL: https://doi.org/10.1007/BF02868481. arXiv:math/0106092.
2. Frédéric Bernicot, 'Uniform estimates for paraproducts and related multilinear multipliers', Revista Matemática Iberoamericana 25 (2009), no. 3, 1055-1088. DOI: 10.4171/RMI/589. URL: https://doi.org/10.4171/RMI/589. arXiv:0806.0459.
3. Camil Muscalu, Terence Tao, Christoph Thiele, 'Uniform estimates on multi-linear operators with modulation symmetry', Journal d'Analyse Mathématique 88 (2002), 255-309. DOI: 10.1007/BF02786579. URL: https://doi.org/10.1007/BF02786579. arXiv:math/0107059.
4. Loukas Grafakos, Nigel J. Kalton, 'The Marcinkiewicz multiplier condition for bilinear operators', Studia Mathematica 146 (2001), no. 2, 115-156. DOI: 10.4064/sm146-2-2. URL: https://doi.org/10.4064/sm146-2-2.
5. Xiaochun Li, 'Uniform estimates for some paraproducts', preprint (2007, revised 2008). arXiv:0709.2906. URL: https://arxiv.org/abs/0709.2906.
6. Vjekoslav Kovač, 'Bellman function technique for multilinear estimates and an application to generalized paraproducts', Indiana University Mathematics Journal 60 (2011), no. 3, 813-846. DOI: 10.1512/iumj.2011.60.4784. URL: https://doi.org/10.1512/iumj.2011.60.4784. arXiv:1108.0917.
7. Loukas Grafakos, Danqing He, Nigel Kalton, Mieczysław Mastyło, 'Multilinear paraproducts revisited', Revista Matemática Iberoamericana 31 (2015), no. 2, 609-616. DOI: 10.4171/RMI/847. URL: https://doi.org/10.4171/RMI/847.
8. Salvador Rodríguez-López, Wolfgang Staubach, 'Some endpoint estimates for bilinear paraproducts and applications', Journal of Mathematical Analysis and Applications 421 (2015), no. 2, 1021-1041. DOI: 10.1016/j.jmaa.2014.07.057. URL: https://doi.org/10.1016/j.jmaa.2014.07.057. arXiv:1406.1663.
9. Olli Saari, Christoph Thiele, 'Paraproducts for bilinear multipliers associated with convex sets', Mathematische Annalen 385 (2023), no. 3-4, 2013-2036. DOI: 10.1007/s00208-022-02397-6. URL: https://doi.org/10.1007/s00208-022-02397-6. arXiv:2102.10305.
10. Marco Fraccaroli, Olli Saari, Christoph Thiele, 'Uniform bounds for bilinear symbols with linear K-quasiconformally embedded singularity', Analysis & PDE 18 (2025), no. 9, 2293-2323. DOI: 10.2140/apde.2025.18.2293. URL: https://doi.org/10.2140/apde.2025.18.2293. arXiv:2402.11661.
11. Dong Dong, Xiao Chun Li, 'On a hybrid of bilinear Hilbert transform and paraproduct', Acta Mathematica Sinica, English Series 34 (2018), no. 1, 29-41. DOI: 10.1007/s10114-017-6415-9. URL: https://doi.org/10.1007/s10114-017-6415-9.
12. Frédéric Bernicot, Polona Durcik, 'Boundedness of some multi-parameter fiber-wise multiplier operators', preprint (2020). arXiv:2007.02211. URL: https://arxiv.org/abs/2007.02211.
13. Peter Friz, Pavel Zorin-Kranich, 'Rough semimartingales and p-variation estimates for martingale transforms', Annals of Probability 51 (2023), no. 2, 397-441. DOI: 10.1214/22-AOP1598. URL: https://doi.org/10.1214/22-AOP1598. arXiv:2008.08897.