---
title: Carleson Operators in Harmonic Analysis
url: https://www.emergentmind.com/topics/carleson-operators
type: topic
---

# Carleson Operators in Harmonic Analysis

Carleson operators are maximally modulated singular integral operators whose prototype is the classical operator
\[
Cf(x)=\sup_{N\in\mathbb R}\left|\int_{\mathbb R} f(x-t)e^{iNt}\,\frac{dt}{t}\right|,
\]
a central object in harmonic analysis because its boundedness is equivalent to almost-everywhere convergence phenomena for Fourier expansions and partial Fourier integrals [2012.08913]. In contemporary usage, the term also encompasses polynomial, oscillatory, variational, anisotropic, multilinear, and abstract metric-space analogues, all of which preserve the characteristic feature of taking a supremum over a modulation family while retaining a singular kernel. The subject now spans classical time-frequency analysis, oscillatory integral theory, weighted inequalities, vector-valued extensions, and non-Euclidean generalizations [1105.4504].

## 1. Classical operator and the convergence problem

The classical Carleson operator can be written either as the maximally modulated Hilbert transform
\[
\mathcal C f(x)=\sup_{\xi\in\mathbb R}\bigl|H(M_\xi f)(x)\bigr|,
\qquad
M_\xi f(x)=e^{2\pi i \xi x}f(x),
\]
or in the oscillatory singular-integral form displayed above [1312.0833]. In one dimension, this operator governs maximal partial Fourier integrals and is the analytic core of Carleson’s theorem on almost-everywhere convergence of Fourier series.

A detailed modern metric-space presentation states the classical theorem in quantitative Egorov form: for a \(2\pi\)-periodic uniformly continuous \(f\) with \(|f(x)|\le 1\), and every \(0<\epsilon<1\), there is a Borel set \(E\subset[0,2\pi]\) with \(|E|\le\epsilon\) and an integer \(N_0\) such that
\[
|f(x)-S_Nf(x)|\le \epsilon
\]
for all \(x\in[0,2\pi]\setminus E\) and all \(N>N_0\) [2405.06423]. That work also makes explicit the reduction from Fourier partial sums to a real-line Carleson maximal operator built from the localized kernel
\[
T f(x)=\sup_{n\in \mathbb{Z}} \sup_{r>0}\left|\int_{r<|x-y|<1} f(y)\kappa(x-y) e^{iny}\, dy\right|,
\]
thereby exhibiting the classical theorem as a special case of a broader Carleson framework on doubling metric measure spaces [2405.06423].

Classical boundedness is now understood in several complementary forms. Carleson’s original theorem gives the \(L^2\) result, while later developments yield \(L^p\)-boundedness for every \(1<p<\infty\) [2012.08913]. In the weighted setting, Di Plinio and Lerner proved that the classical operator satisfies the same linear \(A_q\)-dependence as the Hilbert transform for \(1\le q<p\):
\[
\|\mathcal C\|_{L^p(w)\to L^p(w)} \le c(q,p)\,[w]_{A_q},
\]
and also obtained the \(A_p\)-bound
\[
\|\mathcal C\|_{L^p(w)\to L^p(w)} \le c(p)\,[w]_{A_p}^{\max\left(1,\frac1{p-1}\right)} \log\log\!\bigl(e^e+[w]_{A_p}\bigr),
\]
showing that weighted Carleson theory reflects the near-\(L^1\) endpoint behavior of Fourier convergence [1312.0833].

## 2. Polynomial and oscillatory generalizations

A fundamental nonlinear extension is the polynomial Carleson operator
\[
C_d f(x)=\sup_{\deg(P)\le d} \left| \int_{\mathbb R} f(x-t)e^{iP(t)}\,\frac{dt}{t} \right|,
\]
whose one-dimensional \(L^p\)-boundedness for \(1<p<\infty\) was proved by Victor Lie, resolving the one-dimensional case of Stein’s conjecture [2012.08913]. Lie’s formulation emphasizes the generalized modulation symmetries
\[
M_{j,a}f(x)=e^{iax^j}f(x),
\]
and recasts the proof in a higher-order wave-packet framework adapted to polynomial phases rather than linear frequencies [1105.4504].

That higher-order framework replaces ordinary time-frequency rectangles by curved polynomial tiles. In Lie’s one-dimensional construction, a tile is a \((d+1)\)-tuple
\[
P=[\alpha^1,\alpha^2,\dots,\alpha^d,I], \qquad |\alpha^j|=|I|^{-1},
\]
encoding a localized neighborhood of a polynomial graph over \(I\). The operator is decomposed as
\[
T=\sum_{P\in\mathbb P}T_P,
\]
and the proof proceeds through trees, forests, density and geometric factors, and a mass decomposition that yields BMO control of counting functions and eliminates exceptional sets from the analysis [1105.4504]. One consequence highlighted in that paper is a direct strong \(L^2\) argument for generalized Carleson operators, without passing through interpolation from weak-\(L^2\) [1105.4504].

A more restricted polynomial family fixes the nonlinear phase and maximizes only over a linear modulation:
\[
C_P f(x)=\sup_{N\in\mathbb R} \left| \int_{\mathbb R} f(x-t)e^{iNt+iP(t)}\,\frac{dt}{t} \right|.
\]
Ramos proved that \(C_P\) is bounded on \(L^p\), \(1<p<\infty\), with bounds depending only on \(p\) and \(\deg(P)\), and did so by reducing the problem to the classical linear Carleson operator together with a quadratic analysis, rather than using the full polynomial Carleson machinery [2012.08913]. The result is conceptually narrower than the full theorem of Lie and Zorin-Kranich, but it isolates a mechanism by which linear Carleson control forces boundedness for a nonlinear modulation family [2012.08913].

Oscillatory Carleson operators with nonlinear monomial phase form another major branch:
\[
C_\alpha f(x)=\sup_{N\in\mathbb R}\left|\int_{\mathbb R} f(x-t)e^{iNt+i|t|^\alpha}\,\frac{dt}{t}\right|.
\]
For \(\alpha>0\), \(\alpha\notin\{0,1\}\), Ramos proved \(L^p\)-boundedness for \(1<p<\infty\) together with the uniform estimate
\[
\sup_{\alpha\in(2-\beta,\,2+\beta)} \|C_\alpha\|_{L^p\to L^p} <\infty
\qquad\text{whenever } 0<\beta<1,
\]
thereby controlling the constants in a neighborhood of the delicate quadratic regime \(\alpha=2\) [2012.08913]. Earlier proofs treated \(\alpha=2\) and \(\alpha\neq2\) by sharply different methods and lost control as \(\alpha\to2\); the novelty is the uniformity in \(\alpha\) [2012.08913]. The same paper notes that the analogous odd-phase family with \(e^{i\operatorname{sgn}(t)|t|^\alpha}\) admits the same local uniformity near \(2\), while the regimes \(\alpha\to0\) and especially \(\alpha\to1\) remain more delicate [2012.08913].

## 3. Weighted, variational, and vector-valued forms

Weighted Carleson theory has developed into a precise extension of sharp Calderón–Zygmund inequalities. Di Plinio and Lerner treated the classical Carleson operator \(\mathcal C\), its lacunary version \(\mathcal C_{\mathrm{lac}}\), and the Walsh-Carleson operator \(\mathcal W\) using local mean oscillation and sparse domination rather than the original time-frequency decomposition [1312.0833]. Their abstract framework applies to maximally modulated Calderón–Zygmund operators
\[
T^{\mathcal F}f(x)=\sup_{\alpha\in A}\bigl|T(M^{\phi_\alpha}f)(x)\bigr|,
\qquad
M^{\phi_\alpha}f(x)=e^{2\pi i \phi_\alpha(x)}f(x),
\]
under weak-\(L^p\) input near \(p=1\) [1312.0833].

For the classical operator they proved the weak estimate
\[
\|\mathcal C f\|_{L^{p,\infty}} \lesssim p'\log\log\!\bigl(e^e+p'\bigr)\,\|f\|_{L^p},
\qquad 1<p\le 2,
\]
while for the lacunary operator
\[
\|\mathcal C_{\mathrm{lac}} f\|_{L^{p,\infty}} \lesssim \log(e+p')\,\|f\|_{L^p},
\qquad 1<p\le 2,
\]
and for the Walsh-Carleson operator
\[
\|\mathcal W f\|_{L^{p,\infty}(\mathbb T)} \lesssim p'\|f\|_{L^p(\mathbb T)},
\qquad 1<p\le2
\]
[1312.0833]. The resulting \(A_p\)-bounds exhibit a hierarchy: a doubly logarithmic loss for \(\mathcal C\), a single logarithm for \(\mathcal C_{\mathrm{lac}}\), and no logarithmic loss for \(\mathcal W\), mirroring the corresponding endpoint theories [1312.0833].

A different refinement replaces maximal control by variation control. For Banach-space-valued functions \(f\in S(\mathbb R;X)\), partial Fourier integrals are
\[
C_{\xi}f(x) := \int_{-\infty}^{\xi} \hat{f}(\eta) e^{2\pi i \eta x} \, d\eta,
\]
and the \(r\)-variational Carleson operator is
\[
V^{r}_{*} f(x):= \|\xi \mapsto C_{\xi} f(x)\|_{V^r(\mathbb{R};X)}.
\]
Hytönen, Lorist, and Veraar proved that if \(X\) is an \(r_0\)-intermediate UMD space, then
\[
\| V_{*}^{r} f\|_{L^p(\mathbb R)} \lesssim_{p,r,X} \|f\|_{L^p(\mathbb R;X)}
\]
for all \(r_0<r<\infty\) and \((r/(r_0-1))'<p<\infty\) [2003.02742]. This strengthens vector-valued Carleson–Hunt theory by controlling the oscillation of the truncation path \(\xi\mapsto C_\xi f(x)\), not merely its supremum [2003.02742].

Their method combines time-frequency analysis with outer Lebesgue spaces on the time-frequency-scale domain \(\mathbb R^3_+\), using a wave-packet embedding
\[
E_{\phi}[f](\eta,y,t)= \langle f , \Lambda_{(\eta,y,t)}\phi\rangle
\]
and a truncated embedding \(A^\pm_{\mathfrak c}[g]\) adapted to the linearized variation increments [2003.02742]. The result recovers the maximal vector-valued Carleson theorem for intermediate UMD spaces as the limiting case \(r\to\infty\), but its actual content is stricter: finite \(r\)-variation yields almost-everywhere convergence with quantitative oscillation control [2003.02742].

## 4. Bilinear, anisotropic, and geometric variants

Carleson-type maximal modulation also appears in bilinear settings. The non-resonant bilinear Hilbert–Carleson operators are
\[
BC^{a}(f,g)(x):= \sup_{\lambda\in {\mathbb R}}
\left|\int f(x-t)\, g(x+t)\, e^{i\lambda t^a} \, \frac{dt}{t} \right|,
\qquad a>0,
\]
with \(t^a\) interpreted as \(|t|^a\) or \(\operatorname{sgn}(t)|t|^a\) for noninteger \(a\) [2106.09697]. These operators interpolate between the bilinear Hilbert transform, recovered at \(\lambda=0\), and nonlinear-phase Carleson operators, recovered when one input is frozen to \(1\) [2106.09697].

For every non-resonant exponent
\[
a\in (0,\infty)\setminus\{1,2\},
\]
Guo, Roos, and Yung proved that \(BC^a\) maps \(L^p(\mathbb R)\times L^q(\mathbb R)\) to \(L^r(\mathbb R)\) whenever
\[
\frac1p+\frac1q=\frac1r,\qquad 1<p,q\le \infty,\qquad \frac23<r<\infty,
\]
and they established decay for high-oscillation pieces:
\[
\|T_m(f,g)\|_{L^r}\lesssim_{a,p,q}2^{-\delta a m}\|f\|_{L^p}\|g\|_{L^q}
\]
for some \(\delta=\delta(p,q)>0\) [2106.09697]. The excluded exponents \(a=1,2\) are resonant because additional modulation symmetries destroy the scale-decay mechanism [2106.09697]. The paper describes the resulting operators as genuinely hybrid, combining zero-curvature features inherited from bilinear Hilbert transform modulation symmetry with nonzero-curvature oscillatory cancellation [2106.09697].

In higher-dimensional anisotropic Euclidean settings, Roos studied
\[
\mathscr C_m f(x) = \sup_{N\in\mathbb{R}^n} \left| \int_{\mathbb{R}^n} \widehat f(\xi)\, e^{ix\cdot \xi}\, m(\xi-N)\,d\xi \right|,
\]
where \(m\) is a bounded \(C^\nu\) multiplier homogeneous of degree \(0\) under anisotropic dilations
\[
\delta_\lambda(x)=\bigl(\lambda^{\alpha_1}x_1,\dots,\lambda^{\alpha_n}x_n\bigr)
\]
[1710.10962]. If \(\nu_0\ge 3|\alpha|+2\), then
\[
\|\mathscr C_m f\|_{L^{2,\infty}(\mathbb{R}^n)} \le C\,\|m\|_{\mathscr M^{\nu_0}}\,\|f\|_{L^2(\mathbb{R}^n)},
\]
recovering the classical weak \((2,2)\) Carleson bound when \(n=1\), \(\alpha_1=1\) [1710.10962]. The proof extends the Lacey–Thiele tree method to anisotropic tiles, anisotropic wave packets, and an anisotropic cone decomposition [1710.10962].

A geometric Radon-transform analogue appears in polynomial Carleson operators along the monomial curve \((t,t^m)\subset\mathbb R^2\):
\[
H_N^{m,d}f(x,y)=p.v.\int_{\mathbb{R}} f(x-t,y-t^m)e^{iNt^d}\frac{dt}{t}.
\]
Guo and Pierce studied partial linearizations in which the stopping-time function depends on only one ambient variable:
\[
A_N^{m,d}f(x,y)=H_{N(x)}^{m,d}f(x,y), \qquad
B_N^{m,d}f(x,y)=H_{N(y)}^{m,d}f(x,y),
\]
and proved \(L^p\)-boundedness for all \(1<p<\infty\) when \(m,d>1\) and \(m\neq d\), as well as for the special modulation-invariant cases \(A_N^{m,1}\) with \(m\ge3\) and \(B_N^{m,m}\) with \(m\ge2\) [1605.05812]. Their analysis combines Stein–Wainger-style \(TT^*\) estimates, including phases with fractional monomials, with vector-valued Carleson–Hunt bounds in the symmetric cases [1605.05812]. The same paper observes that \(L^2\) bounds for certain partial curved operators imply the usual Carleson theorem or even the quadratic Carleson theorem, so these are not merely auxiliary models [1605.05812].

A related higher-dimensional geometric variant acts along the paraboloid in \(\mathbb R^{n+1}\):
\[
T_\lambda f(x,t)=\int_{\mathbb R^n} f(x-y,t-|y|^2)e^{iP_\lambda(y)}K(y)\,dy,
\]
where the phase has no linear term and its quadratic homogeneous part is not a nonzero multiple of \(|y|^2\) [1505.03882]. For \(1<p<\infty\), Pierce proved
\[
\big\|\sup_{\lambda} | T_\lambda f| \big\|_{L^p(\mathbb{R}^{n+1})} \leq A_p \|f\|_{L^p(\mathbb{R}^{n+1})},
\]
using a Littlewood–Paley decomposition and square-function comparison with a smoother auxiliary operator, precisely because the usual stopping-time \(TT^*\) method fails in the Radon setting [1505.03882].

## 5. Abstract metric-space formulations

Recent work has detached Carleson theory from Euclidean algebra and rebuilt it on doubling metric measure spaces. In that formulation, the ambient singular kernel is a one-sided Calderón–Zygmund kernel \(K:X\times X\to\mathbb C\) satisfying
\[
|K(x,y)| \leq \frac{2^{a^3}}{V(x,y)}, \qquad V(x,y):=\mu(B(x,\rho(x,y))),
\]
together with Hölder regularity in the second variable [2405.06423]. The generalized Carleson operator is then
\[
Tf(x):=\sup_{\theta\in\Theta} \sup_{0 < R_1 < R_2}
\left| \int_{R_1 <  \rho(x,y) < R_2}  K(x,y) f(y) e(\theta(y)) \, \mathrm{d}\mu(y) \right|,
\]
where \(\Theta\) is a compatible cancellative collection of phase functions equipped, for each ball \(B\subset X\), with a metric \(d_B\) measuring relative oscillation [2405.06423].

The metric-space theorem of Buss, Cladek, and others proves a restricted weak-type bound: if the associated non-tangential maximal truncation \(T_*\) satisfies
\[
\|T_{*}g\|_{2} \leq 2^{a^3} \|g\|_2,
\]
then for \(1<q\le2\), Borel sets \(F,G\subset X\), and \(|f|\le \mathbf 1_F\),
\[
\left|\int_{G} T f \, \mathrm{d}\mu\right|
\le \frac{2^{450a^3}}{(q-1)^5}\,
\mu(G)^{1-\frac{1}{q}} \mu(F)^{\frac{1}{q}}
\]
[2405.06423]. The 2025 sequel sharpens and streamlines the abstract theory, formulating main results as Theorem 1.1 and Theorem 1.2, and states that these results have been computer verified in Lean/mathlib, with the 2024 paper serving as the detailed formalization-oriented blueprint [2508.05563].

The central structural innovation is the axiomatization of the phase family. A compatible collection \(\Theta\) is required to satisfy oscillation control
\[
\sup_{x,y\in B}|\theta(x)-\theta'(x)- \theta(y)+\theta'(y)| \le d_{B}(\theta,\theta'),
\]
monotonicity and doubling-type properties for the local metrics \(d_B\), and a covering condition in phase space [2405.06423]. A cancellative collection also satisfies an oscillatory integral estimate against Lipschitz test functions on each ball, abstracting Euclidean van der Corput or integration-by-parts arguments [2405.06423]. This framework is explicitly designed to include the classical linear phases, polynomial phases, and related non-Euclidean modulation classes [2508.05563].

The proof architecture remains recognizably time-frequency-theoretic: Christ-type dyadic cubes, tiles \(\mathfrak p\), antichains, trees, forests, density functionals, and \(TT^*\)-style tile correlation estimates all reappear in metric form [2405.06423]. What changes is that Euclidean frequency intervals are replaced by modulation cells inside \(\Theta\), and derivative-based phase separation is replaced by the local metrics \(d_B\) [2508.05563].

## 6. Methods, structural themes, and terminological boundaries

Across the subject, three proof paradigms recur. The first is classical time-frequency analysis: tiles, trees, forests, and almost orthogonality in the style of Fefferman and Lacey–Thiele. Lie’s polynomial theorem extends this to higher-order wave packets and curved tiles adapted to polynomial phase graphs [1105.4504]. The second is oscillatory integral analysis via \(TT^*\), stationary phase, and van der Corput estimates, especially in Stein–Wainger-type settings, oscillatory Carleson families, and Radon-transform analogues [2012.08913]. The third is domination technology—sparse domination, local mean oscillation, and outer \(L^p\) embeddings—which has been particularly effective in weighted and vector-valued questions [1312.0833].

Several papers emphasize that the main technical difficulty is not merely singularity, but the interaction between maximal modulation and underlying symmetry. In quadratic or near-quadratic regimes, modulation symmetry can survive strongly enough to force a time-frequency treatment, even when oscillatory curvature is present [2012.08913]. In bilinear Hilbert–Carleson theory, this interaction is explicit: non-resonant exponents permit curvature-driven decay, whereas the resonant cases \(a=1\) and \(a=2\) produce additional modulation invariances that obstruct the available mechanism [2106.09697]. In the monomial-curve and paraboloid settings, curvature from the underlying geometry and oscillation from the phase are inseparable, and partial or smoothed models become necessary [1605.05812].

The available results also leave a clear map of unresolved regimes. Uniform bounds for oscillatory Carleson operators are established only locally near \(\alpha=2\), not globally in \(\alpha\) [2012.08913]. The resonant bilinear Hilbert–Carleson cases \(a=1,2\) remain open [2106.09697]. Several partial operators along monomial curves, including \(A_N^{2,1}\), \(A_N^{m,m}\), and \(B_N^{m,1}\), are not covered by full \(L^p\) theory [1605.05812]. In the variational vector-valued setting, no endpoint statements are proved at \(r=r_0\) or \(p=(r/(r_0-1))'\), and it remains open whether intermediate UMD can be replaced by general UMD [2003.02742]. Abstract metric-space theorems presently yield restricted weak type in the lower range \(1<q\le2\), with strong \(L^q\) bounds for \(1<q<2\) obtained by interpolation, while broader endpoint and full-range extensions remain natural targets [2405.06423].

The term “Carleson” also appears outside Fourier-analytic operator theory. In elliptic PDE, “Carleson perturbations” refers to Carleson-measure-controlled perturbations of elliptic coefficients, not to maximally modulated singular integrals [2011.06574]. That usage is historically connected through Carleson measure estimates but designates a different theory. Within harmonic analysis proper, however, Carleson operators remain the prototypical maximally modulated singular integrals, and their modern generalizations continue to organize a large part of time-frequency and oscillatory analysis.

Source: https://www.emergentmind.com/topics/carleson-operators