Sparse Domination Principles Overview

• Sparse domination principles are techniques that provide pointwise or bilinear control of nonlocal operators using sparse families of cubes, balls, or stopping regions.
• They utilize recursive stopping-time arguments and local oscillation estimates to derive sharp weighted inequalities, endpoint bounds, and vector-valued extensions.
• Adaptations in product geometry, nondoubling measures, and martingale frameworks reveal both the robustness and limitations of sparse domination approaches.

Sparse domination principles are a family of estimates that majorize nonlocal operators by positive, local expressions indexed by sparse collections of cubes, balls, rectangles, or stopping regions. In their most standard form, they control either an operator pointwise by a sparse operator or its bilinear pairing by a sparse form; in more specialized settings they involve $\ell^r$-aggregation, square-function averages, cancellative percentiles, stopping times, or modified forms adapted to complexity and nondoubling geometry. Across one-parameter Calderón–Zygmund theory, metric spaces of homogeneous type, multiscale operators, vector-valued extensions, and martingale frameworks, sparse domination has become a unifying mechanism for deriving quantitative weighted inequalities, endpoint bounds, and structural necessary conditions. At the same time, the theory has sharp limitations: in product geometry, classical $(1,1)$-type sparse domination can fail outright for rectangles (Culiuc et al., 2016, Alonso et al., 2020, Barron et al., 2018).

1. Basic objects and formulations

A sparse family of cubes $S$ in $\mathbb{R}^d$ is typically defined by the existence, for each $Q\in S$, of a measurable set $E_Q\subset Q$ with $|E_Q|\ge \eta |Q|$ such that the sets $\{E_Q\}_{Q\in S}$ are pairwise disjoint. The associated positive sparse bilinear form is

$$\Lambda_S(f,g):=\sum_{Q\in S}|Q|\,\langle f\rangle_Q\,\langle g\rangle_Q, \qquad \langle f\rangle_Q:=\frac{1}{|Q|}\int_Q |f(x)|\,dx,$$

and the corresponding sparse operator is

$$\mathcal A_S f(x):=\sum_{Q\in S}\langle |f|\rangle_Q\,\mathbf 1_Q(x).$$

More generally, one uses $(1,1)$0-averages,

$(1,1)$1

and forms $(1,1)$2 (Culiuc et al., 2016).

In spaces of homogeneous type $(1,1)$3, the same paradigm is stated for balls: $(1,1)$4 The sparse family is now a collection of measurable sets with pairwise disjoint major subsets of comparable measure (Alonso et al., 2020).

Several variants arise when positivity alone is insufficient. In the $(1,1)$5 framework on spaces of homogeneous type, the sparse operator becomes

$(1,1)$6

with a dilation parameter $(1,1)$7 (Lorist, 2019). In cancellative martingale settings one replaces local averages by conditional percentiles and uses

$(1,1)$8

for an $(1,1)$9-sparse adapted sequence $S$0 (Alonso et al., 9 Mar 2026). In balanced nondoubling filtrations, classical sparse forms are supplemented by a modified term

$S$1

that records interactions between distinct sparse atoms at dyadic distance $S$2 (Conde-Alonso et al., 2023). In bi-parameter dyadic analysis, a natural form is

$S$3

with $S$4 dyadic rectangles and $S$5 the bi-parameter square function (Barron et al., 2017).

These formulations separate two basic notions. Pointwise sparse domination asserts an inequality such as

$S$6

while bilinear sparse domination asserts

$S$7

Much of the modern theory moves between these two forms, but the distinction becomes decisive in product and nondoubling settings.

2. Principal proof architectures

A central proof scheme is the Calderón–Zygmund decomposition combined with a recursive stopping-time argument. For dyadic shifts, one selects maximal subcubes $S$8 satisfying

$S$9

for at least one input, obtaining the packing estimate

$\mathbb{R}^d$0

The key iterative lemma then bounds the localized form on $\mathbb{R}^d$1 by a main sparse term plus recursively localized descendants; the dependence on the shift complexity $\mathbb{R}^d$2 is linear (Culiuc et al., 2016).

A closely related but more explicit kernel decomposition appears for Dini-smooth Calderón–Zygmund operators. One writes

$\mathbb{R}^d$3

decomposes $\mathbb{R}^d$4 into single-scale pieces, constructs an exceptional set by maximal functions, performs a Whitney decomposition, and estimates good–good, good–bad, and bad–bad interactions. The Dini integral

$\mathbb{R}^d$5

enters precisely in summing oscillations across scales (Ballesta-Yagüe et al., 9 Sep 2025).

An alternative architecture dispenses with dyadic cubes altogether. In spaces of homogeneous type, sparse domination can be derived from a single-scale $\mathbb{R}^d$6-improving hypothesis. The assumptions are localization

$\mathbb{R}^d$7

uniform truncation bounds

$\mathbb{R}^d$8

and the two single-scale estimates

$\mathbb{R}^d$9

$Q\in S$0

for $Q\in S$1-atoms $Q\in S$2. Whitney coverings of balls replace dyadic systems, and stopping forms are summed by Carleson packing (Alonso et al., 2020).

Other proof mechanisms localize oscillation rather than kernels. An improved pointwise principle assumes a localized $Q\in S$3 property for $Q\in S$4 and weak type for the oscillation operator

$Q\in S$5

which is weaker than assuming weak type for Lerner’s earlier grand maximal truncated operator $Q\in S$6 (Lerner et al., 2019). An even more abstract version is operator-free: one works with arbitrary families $Q\in S$7 satisfying an $Q\in S$8-condition and a quasi-locality estimate, and produces sparse domination directly for the family rather than for a specific operator (Lerner et al., 2021).

3. Canonical positive theorems

Uniform sparse domination for dyadic shifts provides one of the basic one-parameter theorems. For every dyadic lattice $Q\in S$9 and every compactly supported $E_Q\subset Q$0, there exists a sparse collection $E_Q\subset Q$1 such that for all complexities $E_Q\subset Q$2,

$E_Q\subset Q$3

Combined with Hytönen’s dyadic representation, this yields sparse domination for the class $E_Q\subset Q$4 of singular integrals satisfying the assumptions of the classical $E_Q\subset Q$5-theorem (Culiuc et al., 2016).

Rough singular integrals and Bochner–Riesz means admit bilinear sparse domination of $E_Q\subset Q$6 type. For the rough homogeneous singular integrals $E_Q\subset Q$7 with $E_Q\subset Q$8, $E_Q\subset Q$9, and mean zero, one has

$|E_Q|\ge \eta |Q|$0

with dependence on $|E_Q|\ge \eta |Q|$1 when $|E_Q|\ge \eta |Q|$2 and on $|E_Q|\ge \eta |Q|$3 when $|E_Q|\ge \eta |Q|$4. The same scheme covers Bochner–Riesz means at the critical index (Conde-Alonso et al., 2016).

Square-function operators also fit the paradigm. For the Marcinkiewicz integral $|E_Q|\ge \eta |Q|$5 with $|E_Q|\ge \eta |Q|$6, homogeneous of degree zero and mean value zero, there exists a $|E_Q|\ge \eta |Q|$7-sparse family $|E_Q|\ge \eta |Q|$8 such that for $|E_Q|\ge \eta |Q|$9, $\{E_Q\}_{Q\in S}$0,

$\{E_Q\}_{Q\in S}$1

Here the square-function nature of $\{E_Q\}_{Q\in S}$2 dictates the restriction $\{E_Q\}_{Q\in S}$3 (Tao et al., 2019).

The metric-space theorem based on single-scale improving recovers sparse domination for Dini-continuous Calderón–Zygmund kernels on spaces of homogeneous type and extends further to maximal functions associated to convolutions with measures exhibiting Fourier decay, as well as to Radon transforms along polynomial submanifolds of $\{E_Q\}_{Q\in S}$4 (Alonso et al., 2020). A complementary multiscale theorem covers sums $\{E_Q\}_{Q\in S}$5, maximal operators, square functions, and variation-norm operators under support near the diagonal, uniform weak-type $\{E_Q\}_{Q\in S}$6 and restricted strong-type $\{E_Q\}_{Q\in S}$7 bounds for partial sums, a single-scale $\{E_Q\}_{Q\in S}$8 improving estimate, and an $\{E_Q\}_{Q\in S}$9-regularity condition; these hypotheses are also shown to be necessary in the stated sense (Beltran et al., 2020).

A common misconception is that sparse domination is tied to smooth kernels. The rough-kernel results above, together with sparse bounds under $$\Lambda_S(f,g):=\sum_{Q\in S}|Q|\,\langle f\rangle_Q\,\langle g\rangle_Q, \qquad \langle f\rangle_Q:=\frac{1}{|Q|}\int_Q |f(x)|\,dx,$$0-Hörmander regularity and weak endpoint assumptions, show that the method survives substantial loss of smoothness, provided one has the correct local oscillation or improving estimates.

4. Vector-valued, multilinear, and operator-free extensions

For multilinear singular integrals, sparse domination can be proved under the multilinear $$\Lambda_S(f,g):=\sum_{Q\in S}|Q|\,\langle f\rangle_Q\,\langle g\rangle_Q, \qquad \langle f\rangle_Q:=\frac{1}{|Q|}\int_Q |f(x)|\,dx,$$1-Hörmander condition. If $$\Lambda_S(f,g):=\sum_{Q\in S}|Q|\,\langle f\rangle_Q\,\langle g\rangle_Q, \qquad \langle f\rangle_Q:=\frac{1}{|Q|}\int_Q |f(x)|\,dx,$$2 is bounded from $$\Lambda_S(f,g):=\sum_{Q\in S}|Q|\,\langle f\rangle_Q\,\langle g\rangle_Q, \qquad \langle f\rangle_Q:=\frac{1}{|Q|}\int_Q |f(x)|\,dx,$$3 to weak $$\Lambda_S(f,g):=\sum_{Q\in S}|Q|\,\langle f\rangle_Q\,\langle g\rangle_Q, \qquad \langle f\rangle_Q:=\frac{1}{|Q|}\int_Q |f(x)|\,dx,$$4 and its kernel satisfies the $$\Lambda_S(f,g):=\sum_{Q\in S}|Q|\,\langle f\rangle_Q\,\langle g\rangle_Q, \qquad \langle f\rangle_Q:=\frac{1}{|Q|}\int_Q |f(x)|\,dx,$$5-linear $$\Lambda_S(f,g):=\sum_{Q\in S}|Q|\,\langle f\rangle_Q\,\langle g\rangle_Q, \qquad \langle f\rangle_Q:=\frac{1}{|Q|}\int_Q |f(x)|\,dx,$$6-Hörmander condition, then for compactly supported $$\Lambda_S(f,g):=\sum_{Q\in S}|Q|\,\langle f\rangle_Q\,\langle g\rangle_Q, \qquad \langle f\rangle_Q:=\frac{1}{|Q|}\int_Q |f(x)|\,dx,$$7 there exists a sparse family $$\Lambda_S(f,g):=\sum_{Q\in S}|Q|\,\langle f\rangle_Q\,\langle g\rangle_Q, \qquad \langle f\rangle_Q:=\frac{1}{|Q|}\int_Q |f(x)|\,dx,$$8 such that

$$\Lambda_S(f,g):=\sum_{Q\in S}|Q|\,\langle f\rangle_Q\,\langle g\rangle_Q, \qquad \langle f\rangle_Q:=\frac{1}{|Q|}\int_Q |f(x)|\,dx,$$9

almost everywhere (Li, 2016).

Scalar sparse domination can also be lifted systematically to vector-valued sparse domination. For tuples of quasi-Banach function spaces one introduces the multilinear analogue of the UMD condition, denoted $$\mathcal A_S f(x):=\sum_{Q\in S}\langle |f|\rangle_Q\,\mathbf 1_Q(x).$$0, characterized by the boundedness of the multisublinear Hardy–Littlewood maximal operator. Under the assumption $$\mathcal A_S f(x):=\sum_{Q\in S}\langle |f|\rangle_Q\,\mathbf 1_Q(x).$$1, a scalar sparse bound of the form

$$\mathcal A_S f(x):=\sum_{Q\in S}\langle |f|\rangle_Q\,\mathbf 1_Q(x).$$2

implies

$$\mathcal A_S f(x):=\sum_{Q\in S}\langle |f|\rangle_Q\,\mathbf 1_Q(x).$$3

The framework covers iterated Lebesgue, Lorentz, and Orlicz spaces and yields sharp vector-valued weighted bounds directly from scalar sparse domination, without Rubio de Francia extrapolation (Lorist et al., 2020).

A related Banach-space-valued development replaces the usual $$\mathcal A_S f(x):=\sum_{Q\in S}\langle |f|\rangle_Q\,\mathbf 1_Q(x).$$4 sparse operator by an $$\mathcal A_S f(x):=\sum_{Q\in S}\langle |f|\rangle_Q\,\mathbf 1_Q(x).$$5 sparse operator dictated by the operator’s structure and the geometry of the underlying Banach spaces. On spaces of homogeneous type, this produces pointwise $$\mathcal A_S f(x):=\sum_{Q\in S}\langle |f|\rangle_Q\,\mathbf 1_Q(x).$$6-sparse domination for vector-valued operators, an $$\mathcal A_S f(x):=\sum_{Q\in S}\langle |f|\rangle_Q\,\mathbf 1_Q(x).$$7-theorem for vector-valued Calderón–Zygmund operators, an anisotropic mixed norm Mihlin multiplier theorem, and quantitative weighted inequalities for the Rademacher maximal operator (Lorist, 2019).

Sparse domination via the helicoidal method extends the theory to multilinear Fourier multipliers whose symbols are singular along a $$\mathcal A_S f(x):=\sum_{Q\in S}\langle |f|\rangle_Q\,\mathbf 1_Q(x).$$8-dimensional subspace of

$$\mathcal A_S f(x):=\sum_{Q\in S}\langle |f|\rangle_Q\,\mathbf 1_Q(x).$$9

with $(1,1)$00, and to the variational Carleson operator. The method yields sparse estimates for multiple vector-valued extensions and Fefferman–Stein type inequalities in $(1,1)$01 for all $(1,1)$02 (Benea et al., 2017).

The operator-free framework goes further still. For arbitrary families $(1,1)$03 satisfying an $(1,1)$04-condition and quasi-locality, one obtains sparse domination results that apply to generalized Poincaré–Sobolev inequalities, tent spaces, general dyadic sums, and vector-valued square functions that are not localizable in the sense of earlier operator-based schemes (Lerner et al., 2021). This suggests that sparse domination is not merely an operator theorem but a structural principle for local oscillation and aggregation.

5. Cancellative, martingale, and product-geometry variants

In continuous time, sparse domination can be formulated through stopping times rather than cubes. If $(1,1)$05 and $(1,1)$06 are uniformly integrable càdlàg Hilbert-space-valued martingales and $(1,1)$07 is differentially subordinate to $(1,1)$08, then there exists a sparse increasing sequence of stopping times $(1,1)$09 such that

$(1,1)$10

This pathwise domination leads to weighted martingale bounds in terms of the martingale $(1,1)$11 characteristic $(1,1)$12, and in the special case $(1,1)$13 one gets the optimal exponent $(1,1)$14 (Domelevo et al., 2016).

A distinct development is cancellative sparse domination. Instead of dominating by noncancellative averages, one precomposes sparse operators with percentile maximal functions. In the martingale setting,

$(1,1)$15

and the paper proves the sparse characterization

$(1,1)$16

It also yields cancellative sparse domination for martingale transforms, square functions, dyadic Haar shifts, and smooth Calderón–Zygmund operators, and gives the bi-parameter bound

$(1,1)$17

for the dyadic strong maximal function (Alonso et al., 9 Mar 2026).

Product geometry requires further modifications. For bi-parameter martingale transforms, cancellative dyadic shifts, and paraproduct-free Journé operators, one has sparse domination not by averages of $(1,1)$18 but by averages of square functions: $(1,1)$19 This reflects the Chang–Fefferman product structure and the role of Córdoba–Fefferman selection in place of one-parameter stopping-time trees (Barron et al., 2017).

In general filtered measure spaces with balanced measures, classical sparse domination may fail and must be replaced by a sparse domination-like inequality

$(1,1)$20

where $(1,1)$21 records interactions up to dyadic distance $(1,1)$22. The associated admissible weight classes depend on the complexity $(1,1)$23 (Conde-Alonso et al., 2023). This is a sharp indication that cancellation, filtration regularity, and product geometry can force sparse principles away from the classical one-parameter model.

6. Weighted inequalities and quantitative sharpness

Sparse domination is especially effective because positive sparse forms admit precise weighted estimates. In the dyadic-shift/T(1) framework, sparse domination recovers Hytönen’s sharp $(1,1)$24 bounds in the scalar setting and yields, for matrix weights,

$(1,1)$25

uniformly over $(1,1)$26 (Culiuc et al., 2016).

In spaces of homogeneous type, if $(1,1)$27 satisfies the single-scale improving hypotheses with exponents $(1,1)$28, then for weights

$(1,1)$29

one has

$(1,1)$30

The same machinery yields vector-valued inequalities and endpoint consequences where appropriate (Alonso et al., 2020).

For rough homogeneous singular integrals $(1,1)$31, sparse domination gives quantitative weighted estimates. If $(1,1)$32 and $(1,1)$33, then for $(1,1)$34,

$(1,1)$35

with

$(1,1)$36

When $(1,1)$37 one recovers the exponent

$(1,1)$38

for both $(1,1)$39 and critical Bochner–Riesz means (Conde-Alonso et al., 2016).

For the Marcinkiewicz integral $(1,1)$40 with rough $(1,1)$41 kernel, the bilinear sparse domination implies

$(1,1)$42

as well as the endpoint estimate

$(1,1)$43

A Coifman–Fefferman type inequality

$(1,1)$44

also follows (Tao et al., 2019).

For variational Carleson, the sparse bound yields

$(1,1)$45

for $(1,1)$46 and $(1,1)$47 (Benea et al., 2017). In the cancellative Hardy-space setting, smooth Calderón–Zygmund operators satisfy

$(1,1)$48

and if $(1,1)$49 then $(1,1)$50 is bounded $(1,1)$51 (Alonso et al., 9 Mar 2026).

7. Obstructions, modified forms, and open directions

Sparse domination is not universal in its classical $(1,1)$52 form. For the dyadic strong maximal function $(1,1)$53 on axis-parallel rectangles in $(1,1)$54, the main theorem states that for every $(1,1)$55 and every $(1,1)$56 there exist compactly supported integrable $(1,1)$57 such that for every $(1,1)$58-sparse collection $(1,1)$59 of dyadic rectangles,

$(1,1)$60

Quantitatively, the construction produces discrete measures $(1,1)$61 with

$(1,1)$62

and the choice $(1,1)$63 forces any Carleson constant $(1,1)$64 to blow up (Barron et al., 2018).

The obstruction is geometric. The proof uses an anisotropic “distance”

$(1,1)$65

a maximally separated set $(1,1)$66, and a near-far cloud $(1,1)$67 with unique associations and controlled clustering. The same mechanism rules out $(1,1)$68-type sparse domination for bi-parameter martingale transforms, excludes $(1,1)$69 sparse bounds for

$(1,1)$70

and propagates the failure to all dimensions $(1,1)$71 by tensor reduction (Barron et al., 2018).

A second obstruction occurs in balanced nondoubling filtrations. There, classical sparse domination fails even for positive complexity Haar shifts, and the correct weighted classes depend on the complexity $(1,1)$72 through characteristics such as

$(1,1)$73

The modified sparse form $(1,1)$74 and the maximal operator $(1,1)$75 are not artefacts of proof but necessary features of the theory (Conde-Alonso et al., 2023).

In bi-parameter harmonic analysis this does not mean that all sparse control disappears. Instead, one often has to replace averages of $(1,1)$76 by averages of square functions, or replace classical sparse forms by cancellative or complexity-sensitive versions (Barron et al., 2017, Alonso et al., 9 Mar 2026). A plausible implication is that “sparse domination” in product and nondoubling settings is better understood as a family of domination mechanisms rather than a single canonical formula.

Several open directions are explicit in the literature. For the strong maximal function and related bi-parameter operators, it remains open whether $(1,1)$77-type sparse bounds hold for $(1,1)$78 with $(1,1)$79, whether mixed sparse forms or restricted classes of rectangles can recover useful domination, and which other multiparameter operators admit salvageable sparse frameworks (Barron et al., 2018). In the metric Whitney-covering approach, extending beyond Dini moduli and broadening maximal Radon applications remain open (Alonso et al., 2020). In the Calderón–Zygmund decomposition approach to Dini kernels, transferring the scheme to nondoubling measures is obstructed by the lack of classical Whitney coverings and good dyadic geometry for arbitrary measures (Ballesta-Yagüe et al., 9 Sep 2025).