Carleson–Hunt endpoint maximal estimates for the continuous AKNS-ZS/Dirac scattering transform with square-integrable potential

The scattering transform of the AKNS-ZS/Dirac system on the real line is a nonlinear analogue of the Fourier transform: the transfer matrix $G(x,k)$, normalized at the left endpoint of the truncation, takes values in $SU(1,1)$, and its upper-left entry $a(x,k)$ (the transmission coefficient of the potential truncated at $x$) satisfies $|a|\ge1$, so $\sqrt{\log|a|}$ is the natural size of the nonlinear Fourier transform, with the exact nonlinear Parseval identity $\int\log|a(\infty,k)|\,dk=\frac{\pi}{2}\|F\|_2^2$ at the un-maximized endpoint. Muscalu, Tao and Thiele (Nonlinearity 2003) proposed the Carleson-theorem analogue for this transform and proved the weak-type layer only in a dyadic Cantor-group model, stating they could not decide the truth of the continuous problem. The problem is the pair of endpoint maximal estimates they conjectured — the Carleson–Hunt strong $L^2$ estimate $\|\sup_x\sqrt{\log|a(x,\cdot)|}\|_{L^2(k)}\le C\|F\|_2$ and the weak-type estimate $|\{k:\sup_x\sqrt{\log|a(x,k)|}>\lambda\}|\le C\lambda^{-2}\|F\|_2^2$ — together with almost-everywhere finiteness of the maximal function for complex $L^2$ potentials. Later work reshaped the landscape without closing the estimates: Poltoratski (Annals of Mathematics 2024) proved a.e. convergence of the nonlinear Fourier transform and a.e. boundedness of generalized eigenfunctions for real half-line $L^2$ potentials; Poltoratski (2026) obtained the first quantitative weak-type maximal estimate, under dyadic $L^1$-mass and spectral-regularity hypotheses; Denisov (2026) refuted the strong-form nonlinear Carleson conjecture (divergence of the full scattering coefficient pair for some square-summable potentials) while leaving the modulus layer untouched; Denisov (2025) isolated quantitative Carleson–Hunt-type conjectures as the recognized open core. The strong and weak continuous maximal estimates in their stated generality remain open.

Background

The nonlinear Fourier transform of one-dimensional Dirac systems. The Ablowitz–Kaup–Newell–Segur / Zakharov–Shabat (AKNS-ZS) hierarchy contains the simplest nontrivial scattering problem, the 'special AKNS-ZS system', which is the eigenfunction equation of a Dirac-type operator on the real line: $\frac{df}{dx}=kJf+q(x)f$ with spectral parameter $k\in\mathbb{R}$, $J=\begin{pmatrix}-i&0\\ 0&i\end{pmatrix}$, and $q(x)=\begin{pmatrix}0&F(x)\\ \overline{F(x)}&0\end{pmatrix}$ for a potential $F\in L^2(\mathbb{R})$ (complex-valued unless stated otherwise). Writing the two components of a solution as $a(x)e^{-ikx}$ and $b(x)e^{ikx}$, the $2\times2$ transfer matrix $G$ solves $G'=WG$ with $W(x,k)=\begin{pmatrix}0&F(x)e^{2ikx}\\ \overline{F(x)}\,e^{-2ikx}&0\end{pmatrix}$ and initial condition $G(-\infty)=\mathrm{id}$ (for compactly supported $F$; for general $F\in L^2(\mathbb{R})$ one instead normalizes $G(0)=\mathrm{id}$). The matrix $G$ then remains in the Lie group $SU(1,1)$, consisting of the matrices $\begin{pmatrix}\alpha&\beta\\ \overline{\beta}&\overline{\alpha}\end{pmatrix}$ with $|\alpha|^2-|\beta|^2=1$, so in particular $|a|\ge1$ and $\sqrt{\log|a|}$ is a nonnegative measure of the transform's size. As $x\to\infty$ (for nice $F$), $G(x,k)$ tends to the scattering matrix at parameter $k$: $a(\infty,k)$ is the transmission coefficient, $b(\infty,k)$ the reflection coefficient, and the map $F\mapsto$ scattering data is the nonlinear Fourier transform. The analogy with the linear Fourier transform is quantitative: there is an exact nonlinear Parseval identity $\int_{\mathbb{R}}\log|a(\infty,k)|\,dk=\frac{\pi}{2}\|F\|_{L^2}^2$, and Christ and Kiselev proved Hausdorff–Young and maximal Hausdorff–Young inequalities for $1<p<2$, of the form $\|\sup_x\sqrt{\log|a(x,\cdot)|}\|_{L^{p'}(\mathbb R_k)}\le C_p\|F\|_{L^p}$, where $p'$ is the dual exponent. The maximal (sup-over-truncation) versions control, uniformly in the truncation point, the generalized eigenfunctions — the solutions of the system at spectral parameter $k$ — of Dirac-type operators, and thereby, through the AKNS-ZS framework that also encodes Schrödinger operators, the generalized eigenfunctions of one-dimensional Schrödinger operators with slowly decaying potentials.

Motivation from spectral theory. For Schrödinger operators with square-summable ($L^2$) potentials, Deift and Killip proved that the absolutely continuous spectrum fills the expected half-line; by Schnol'-type arguments (Schnol's theorem, as extended by Simon: absolute continuity of the spectrum follows from almost-everywhere boundedness of the generalized eigenfunctions) this is a strictly weaker statement than almost-everywhere boundedness itself. Christ–Kiselev's theory reaches every $1<p<2$; the endpoint $p=2$ is precisely the missing layer, and the conjectures below are its formulation for the nonlinear transform.

The linear prototype. Carleson's theorem (1966) states that the partial Fourier integrals $\int_{|x|<T}F(x)e^{2ikx}\,dx$ of an $L^2$ function converge for almost every $k$; the quantitative engine is the Carleson–Hunt maximal $L^2$ estimate for the supremum over $T$. Muscalu, Tao and Thiele proposed the nonlinear counterpart: replace the partial Fourier integral of $F\cdot\chi_{(-\infty,x)}$ by its scattering transform, and the modulus of the partial integral by $\sqrt{\log|a(x,k)|}$; the supremum over the truncation point $x$ is the nonlinear Carleson maximal operator. They formulated a strong (Carleson–Hunt) conjecture, a weak-type conjecture, and an almost-everywhere finiteness endpoint, and stated explicitly that they could not decide whether any of them is true or false. A weak-type $(2,2)$ estimate for an operator $\mathcal M$ means $|\{k:\mathcal M F(k)>\lambda\}|\le C\lambda^{-2}\|F\|_2^2$ for all $\lambda>0$, where $|\cdot|$ is Lebesgue measure.

The Cantor-group ($d$-adic) model. As supporting evidence, the same authors replaced the characters $e^{2ikx}$ of $\mathbb{R}$ by the characters $w(k,x)$ of a Cantor group — an infinite product of cyclic groups $\mathbb{Z}_d$, with $x$ and $k$ parametrized by their base-$d$ expansions — and proved the analogue of the weak-type estimate in this model (their Theorem 1.4), together with the model's Plancherel-type $L^2$ bound. For $d=2$ and real-valued $F$ the model degenerates: the transfer matrices commute and the model transform becomes the Walsh–Fourier transform (the Fourier transform on the Cantor group), so the model theorem reduces to Carleson's theorem for the Walsh system; the model is genuinely nonlinear for complex $F$ or $d>2$. Methodologically decisive for what follows, they also proved (their Proposition 5.1, complementing their published counterexample to a multilinear endpoint question of Christ and Kiselev) that the natural multilinear (Picard) expansion of the transform into powers of $F$ cannot yield the endpoint estimates: the individual multilinear terms violate $L^2$ bounds. Their model proof rests on 'swapping inequalities', combinatorial estimates for which no continuous analogue is known; the only known proof of the continuous Plancherel identity (contour integration) does not admit the decompositions that make the model work.

Current landscape. The qualitative a.e.-convergence layer of the continuous problem was later settled for real potentials: Poltoratski (published in the Annals of Mathematics, 2024) proved that the nonlinear Fourier transforms of the truncations $f\cdot\chi_{(0,T)}$ of every real $f\in L^2(\mathbb{R}^+)$ converge almost everywhere, which implies almost-everywhere boundedness of the generalized eigenfunctions of half-line Dirac systems with real $L^2$ potentials; the proof is non-perturbative, based on the motion of resonances (zeros of the Hermite–Biehler functions, the entire functions generating the de Branges spaces — the spectral-theoretic analogues of Paley–Wiener spaces — associated with the Dirac system). Denisov subsequently showed (2026) that the strong form of the nonlinear Carleson conjecture, almost-everywhere convergence of the full pair of scattering coefficients, is false: there exist square-summable potentials (discrete $\ell^2$ data, and $L^2(\mathbb R)$ potentials) for which those limits fail to exist at every spectral point, a divergence living in the phase layer, while the modulus layer $\log|a|$ is not addressed by his construction. The modulus layer is exactly what the source conjectures quantify, and it remains open at every quantitative level: beyond the $p<2$ maximal Hausdorff–Young bounds and the un-maximized Parseval identity, the only general partial result is a 2026 conditional weak-type maximal estimate of Poltoratski for the real half-line system under dyadic $L^1$-mass and spectral-regularity hypotheses.

Problem Statement

Fix the special AKNS-ZS/Dirac system on the real line, i.e. the transfer-matrix initial value problem $G'(x,k)=W(x,k)\,G(x,k),\qquad W(x,k)=\begin{pmatrix}0&F(x)e^{2ikx}\\ \overline{F(x)}\,e^{-2ikx}&0\end{pmatrix},\qquad G(0,k)=\mathrm{id},$ for a potential $F\in L^2(\mathbb R)$ (complex-valued; for compactly supported $F$, after translating the support of $F$ into $(0,\infty)$ — a harmless operation, since a simultaneous translation of $F$ and of the truncation point leaves $\|F\|_{L^2}$ and the maximal function $\sup_x\sqrt{\log|a(x,k)|}$ unchanged — one may equivalently normalize $G(-\infty)=\mathrm{id}$, in which case $x$ ranges over the truncation endpoints of $F\cdot\chi_{(-\infty,x)}$). Let $a(x,k)$ denote the upper-left entry of the $SU(1,1)$-valued solution $G(x,k)$, so that $|a|\ge1$ and $\log|a|\ge0$. Decide whether the following Carleson-Hunt-type endpoint estimates for the continuous (non-Cantor) scattering transform hold, with $C$ absolute, independent of $F$ and of $\lambda$:

Conjecture 1.2 (Carleson-Hunt estimate, strong endpoint). $\Big\|\sup_{x\in\mathbb R}\sqrt{\log|a(x,\cdot)|}\Big\|_{L^2(\mathbb R_k)}\le C\,\|F\|_{L^2(\mathbb R)}.$

Conjecture 1.3 (weak type Carleson estimate). $\big|\{k\in\mathbb R:\ \sup_{x\in\mathbb R}\sqrt{\log|a(x,k)|}>\lambda\}\big|\le C\,\lambda^{-2}\,\|F\|_{L^2(\mathbb R)}^2\qquad\text{for all }\lambda>0.$

Qualitative endpoint. For every $F\in L^2(\mathbb R)$ (complex-valued), $\sup_{x\in\mathbb R}\sqrt{\log|a(x,k)|}<\infty$ for almost every $k\in\mathbb R$.

These three statements form one coherent objective, the endpoint Carleson theory of the nonlinear Fourier transform, since Conjecture 1.2 implies Conjecture 1.3 by Chebyshev's inequality and Conjecture 1.3 implies the qualitative endpoint by letting $\lambda\to\infty$. The problem is to prove or refute them at the stated generality. An answer restricted to real-valued $F$, one-sided (half-line) potentials, dyadic $L^1$-mass conditions, regular spectral sets, or lacunary maximal functions is partial progress (such restrictions currently exist in the literature), not a resolution. The almost-everywhere finiteness layer is already resolved for real half-line potentials (Poltoratski, Annals of Mathematics 2024), and the phase layer can diverge for square-summable potentials (Denisov 2026), so the conjectures concern exactly the modulus layer $\log|a|$.

References

1. Camil Muscalu, Terence Tao, and Christoph Thiele, 'A Carleson type theorem for a Cantor group model of the scattering transform', Nonlinearity 16 (2003), no. 1, 219-246. DOI: 10.1088/0951-7715/16/1/314; arXiv: math/0205139; https://arxiv.org/abs/math/0205139
2. Alexei Poltoratski, 'Pointwise convergence of the non-linear Fourier transform', Annals of Mathematics (2) 199 (2024), no. 2. DOI: 10.4007/annals.2024.199.2.4; arXiv: 2103.13349; https://arxiv.org/abs/2103.13349
3. Alexei Poltoratski, 'A maximal estimate for the non-linear Fourier transform', preprint (2026). arXiv: 2608.24606; https://arxiv.org/abs/2608.24606
4. Sergey A. Denisov, 'Pointwise behavior of SU(1,1) nonlinear Fourier transform', preprint (2026). arXiv: 2605.25108; https://arxiv.org/abs/2605.25108
5. Sergey A. Denisov, 'The strong version of nonlinear Carleson conjecture fails', preprint (2026). arXiv: 2605.01658; https://arxiv.org/abs/2605.01658
6. Sergey A. Denisov, 'Two quantitative versions of the Nonlinear Carleson Conjecture', preprint (2025). arXiv: 2505.06788; https://arxiv.org/abs/2505.06788
7. Sergey A. Denisov, 'Convergence of sparse square-summable NLFT', preprint (2026). arXiv: 2606.06661; https://arxiv.org/abs/2606.06661
8. Gevorg Mnatsakanyan, 'Weak $L^2$ bound of the lacunary Carleson operator for the non-linear Fourier transform', preprint (2022, revised 2025). arXiv: 2210.00309; https://arxiv.org/abs/2210.00309
9. Michel Alexis, Gevorg Mnatsakanyan, and Christoph Thiele, 'One sided orthogonal polynomials and a pointwise convergence result for $SU(2)$-valued nonlinear Fourier series', preprint (2025). arXiv: 2507.05124; https://arxiv.org/abs/2507.05124
10. Terence Tao and Christoph Thiele, 'Nonlinear Fourier Analysis', lecture notes from the Park City Mathematics Institute graduate summer school (lectures delivered 2003; arXiv posting of January 2012). arXiv: 1201.5129; https://arxiv.org/abs/1201.5129
11. Vjekoslav Kovač, 'Uniform constants in Hausdorff-Young inequalities for the Cantor group model of the scattering transform', Proceedings of the American Mathematical Society 140 (2012), no. 3, 915-926. DOI: 10.1090/S0002-9939-2011-11078-5; arXiv: 1012.3146; https://arxiv.org/abs/1012.3146
12. Michael Christ and Alexander Kiselev, 'Maximal functions associated to filtrations', Journal of Functional Analysis 179 (2001), no. 2, 409-425. DOI: 10.1006/jfan.2000.3687; https://doi.org/10.1006/jfan.2000.3687
13. Michael Christ and Alexander Kiselev, 'WKB asymptotic behavior of almost all generalized eigenfunctions for one-dimensional Schrödinger operators with slowly decaying potentials', Journal of Functional Analysis 179 (2001), no. 2, 426-447. DOI: 10.1006/jfan.2000.3688; https://doi.org/10.1006/jfan.2000.3688
14. P. Deift and R. Killip, 'On the absolutely continuous spectrum of one-dimensional Schrödinger operators with square summable potentials', Communications in Mathematical Physics 203 (1999), no. 2, 341-347. DOI: 10.1007/s002200050615; https://doi.org/10.1007/s002200050615
15. Camil Muscalu, Terence Tao, and Christoph Thiele, 'A counterexample to a multilinear endpoint question of Christ and Kiselev', Mathematical Research Letters 10 (2003), no. 2, 237-246. DOI: 10.4310/mrl.2003.v10.n2.a10; https://doi.org/10.4310/mrl.2003.v10.n2.a10
16. Lennart Carleson, 'On convergence and growth of partial sums of Fourier series', Acta Mathematica 116 (1966), 135-157. DOI: 10.1007/BF02392815; https://doi.org/10.1007/BF02392815
17. Pavel Gubkin, 'Mate-Nevai-Totik theorem for Krein systems', preprint (2020). arXiv: 2011.14374; https://arxiv.org/abs/2011.14374