Controlling the triangular Hilbert transform across all scales

← All digests

Family 082 · Real and complex analysis
Importance 5/5 · Provisional editorial rating
Sources consulted 2026-10-14

1. The question

Singular integrals balance positive and negative contributions near a singularity. The ordinary Hilbert transform performs this cancellation along one variable. What happens when two functions move in different coordinate directions, so their interactions become entangled?

The triangular Hilbert transform is a basic example. One input is shifted horizontally, the other vertically, but both use the same shift parameter and the same singular kernel. Pairing the output with a third function produces a cycle of three overlapping pairs of variables.

Family 082 contains three manuscripts. One claims a maximal estimate over both truncation endpoints, a later paper strengthens this to annular variation for every exponent above two, and another treats a dyadic model with unrestricted real inputs. The continuous results allow complex inputs; the dyadic theorem has its own precise real-input formulation.

2. Background and essential definitions

A finite annular truncation

For functions $F,G$ on $\mathbb R^2$, define

$$B_{\varepsilon,R}(F,G)(x,y)= \int_{\varepsilon<|t|<R}F(x+t,y)G(x,y+t)\,\frac{dt}{t}, \qquad0<\varepsilon<R<\infty.$$

The lower cutoff avoids the singularity at zero; the upper cutoff avoids an uncontrolled infinite tail. The odd factor $1/t$ makes cancellation between positive and negative shifts essential.

The operator is bilinear because it depends linearly on each input separately. Hölder’s inequality suggests the exponents $L^3\times L^3\to L^{3/2}$, since $1/3+1/3=2/3$. Pairing the output with another $L^3$ function gives a scalar trilinear form.

A change of variables writes that form as

$$\Lambda_{\varepsilon,R}(G_0,G_1,G_2)= \iiint_{\varepsilon<|a+b+c|<R} \frac{G_0(a,b)G_1(b,c)G_2(c,a)}{a+b+c}\,da\,db\,dc.$$

Each factor sees two of the three variables, and the shared pairs form a triangle. The symmetric point means all three input exponents equal three. It does not mean the input functions themselves are symmetric.

Maximal size and variation

The maximal transform is

$$B_*(F,G)(x,y)=\sup_{0<\varepsilon<R<\infty} |B_{\varepsilon,R}(F,G)(x,y)|.$$

The endpoints may be chosen separately at each output point. This is stronger than bounding each fixed truncation by a common norm constant.

For $r>2$, the annular variation is

$$V_r(F,G)=\sup_{0<t_0<\cdots<t_J} \left(\sum_{j=1}^{J}|B_{t_{j-1},t_j}(F,G)|^r\right)^{1/r},$$

where the supremum includes all finite partitions. It measures accumulated changes across disjoint annuli, not merely the largest single value. Taking $J=1$ shows that variation controls the maximal transform. Rational endpoints give the same supremum while making measurability immediate.

3. Development history and previous partial results

The flat operator appears in Demeter and Thiele’s work on the two-dimensional bilinear Hilbert transform. Its cyclic dependence makes it a model problem for entangled singular integrals. Thiele’s contribution to Some problems in harmonic analysis records the symmetric triangular estimate as a central question.

Kovač, Thiele and Zorin-Kranich obtained bounds for a Walsh or dyadic model when one input has a specified structural form. Their work already captures important related operators, but does not allow all three inputs to vary freely in the claimed new dyadic range.

Another line of progress measured how slowly truncated norms grow with the number of scales. Tao established sublogarithmic cancellation for multilinear Hilbert transforms, and Zorin-Kranich extended this kind of cancellation to simplex transforms. Durcik, Kovač and Thiele obtained power-type cancellation. At the symmetric triangular tuple, the bound retained a square-root dependence on the logarithmic scale range.

Results with curved shifts, such as $(t,t^2)$, or with an average over direction parameters gain additional cancellation. They do not directly settle the flat shifts $(t,t)$ underlying this family.

The manuscripts develop a matrix-energy approach from earlier telescoping and Bellman-function methods. The historical difficulty is that a triangle is a cyclic, non-bipartite interaction, so methods adapted to simpler patterns of shared variables do not transfer automatically.

4. What the new manuscripts claim

Continuous maximal and variation estimates

For arbitrary complex $F,G\in L^3(\mathbb R^2)$, the maximal manuscript states

$$\|B_*(F,G)\|_{3/2}\le C\|F\|_3\|G\|_3.$$

The later annular manuscript states the stronger family of bounds

$$\|V_r(F,G)\|_{3/2}\le C_r\|F\|_3\|G\|_3 \qquad(r>2).$$

Partitions can depend on the output point, and there is no bound on how many endpoints fall within one dyadic interval of radii.

The papers obtain a joint principal value as $\varepsilon\downarrow0$ and $R\uparrow\infty$, both almost everywhere and in $L^{3/2}$. The variation paper also states norm convergence of the supremum of the remaining errors over sufficiently small lower and sufficiently large upper cutoffs. Pairing with a third input gives the symmetric scalar triangular estimate.

The dyadic absolute-sum theorem

The dyadic model uses intervals $[2^kn,2^k(n+1))$ on $[0,\infty)$. A triple at one scale is admissible when its three integer indices have bitwise XOR equal to zero. XOR is binary addition without carries. Each interval has a Haar sign function, equal to one on its left half and minus one on its right half.

For the corresponding local trilinear contributions $L_I$, the dyadic manuscript claims

$$\sum_{k\in S}\sum_{I\text{ admissible at }k} |L_I(F_0,F_1,F_2)| \le40\prod_{v=0}^2\|F_v\|_{L^3(\mathbb R_+^2)}$$

for every finite scale set $S$ and arbitrary real $L^3$ inputs. Each $L_I$ is the integral of the cyclic product over the interval triple, multiplied by the three Haar signs and by $2^{-k}$.

Because the absolute values are taken separately, coefficients of modulus at most one may vary independently among admissible triples. This is a theorem for a distinct operator; it is not an automatic replacement for the continuous principal-value statement.

5. Why the advance matters

Importance: 5/5. If correct, the continuous estimates settle the symmetric triangular Hilbert-transform problem and give strong control of its truncation behavior. The dyadic theorem removes structural restrictions on the inputs and controls even the sum of absolute local contributions.

Uniformity in the number of scales is decisive. A bound growing like the square root of the scale count demonstrates cancellation, but still diverges as more and more scales are included. The new claims remove that dependence in the appropriate norm.

Variation is stronger still: it limits repeated oscillation as the cutoffs change. This provides a natural route to convergence of singular integrals rather than only boundedness of isolated truncations. The family would therefore advance both the boundedness and convergence aspects of an important entangled operator.

6. The main idea

Encode the cycle with matrices

Sample the three pair-dependent inputs on a finite grid. Each input becomes a matrix between the coordinate spaces of the two variables it uses. Their cyclic product is then represented by a matrix trace, which naturally retains the triangular arrangement.

At each vertex of the triangle, place the two incident matrices side by side in a row operator $R$. The energy is

$$\operatorname{tr}\bigl((R^*R)^{3/2}\bigr).$$

Here $R^*$ is the conjugate transpose, and the positive matrix power is defined using eigenvalues. The exponent gives the energy the same cubic scaling as a product of three inputs.

The manuscripts state matrix inequalities whose constants do not grow with the grid size. This dimension independence is necessary before passing from sampled functions to continuous ones.

Pay for truncations through changes in energy

In the continuous argument, Gaussian weights smooth the sampled inputs. As the Gaussian variance changes, a heat-flow identity makes the decrease of the nonnegative energy control the required cyclic products.

A pointwise choice of truncation means that different entries of the matrix representing the dual output test switch on and off at different scales. The energy can jump at these switches. The maximal paper’s extra estimate bounds the total cost of those jumps using the input norms, without charging separately for every scale.

The variation paper must handle many disjoint annuli. It first proves a count estimate: the sum of the absolute contributions from at most $m$ disjoint annuli has norm bounded by

$$C\sqrt m\log(2+m)\|F\|_3\|G\|_3.$$

Sorting the increments by size and grouping their ranks into dyadic blocks gives a convergent series when $r>2$. The square-root growth is precisely why that variation range appears.

Hard annular endpoints create additional errors beyond smooth Gaussian cutoffs. The paper groups endpoints by scale and uses a frequency-block estimate to control many jumps together, rather than paying a linear cost in their number.

A discrete telescoping counterpart

For the dyadic theorem, Gaussian flow is replaced by passage from an interval triple to its four XOR-admissible child triples. The local contribution is bounded by an increase in average child energy.

Summing these increases over the tree makes intermediate energies cancel. Only controlled boundary terms remain. This telescoping is what allows an estimate independent of the finite scale set, even with absolute local contributions.

7. Scope, limitations, and open questions

The continuous claims are at the symmetric $L^3\times L^3\to L^{3/2}$ point. The family does not state the full expected range of unequal exponents or all higher simplex transforms.

The variation theorem requires $r>2$. It makes no endpoint assertion at $r=2$, and the constant may depend on $r$. Its pointwise partition supremum is stronger than a variation quantity formed after taking output norms; these two formulations should not be confused.

The dyadic estimate uses the specified XOR model and real inputs. Its absolute sum is a sum of absolute local integrals, not the integral of the absolute value of the singular kernel times the inputs. Cancellation still occurs inside each local contribution.

The principal value is joint in the two cutoffs. This is stronger than proving convergence only along one predetermined sequence of annuli, and it remains distinct from absolute integrability of the untruncated kernel.

8. Sources and further reading

Thiele’s problem formulation, the structured-input dyadic theorem, and the power-type cancellation theorem provide the main historical comparisons. The three source manuscripts below state the continuous variation, dyadic absolute-sum and continuous maximal results separately.

Optional introductions to the Hilbert transform and Haar wavelets explain the continuous kernel and the discrete signs used in the models.

Source manuscripts

• M1. OpenAI, Annular variation of the triangular Hilbert transform at the symmetric point (October 5, 2026).
• M2. OpenAI, An L³ bound for the dyadic triangular Hilbert form (October 5, 2026).
• M3. OpenAI, The maximal triangular Hilbert transform at the symmetric point (September 24, 2026).

Sources consulted on 14 October 2026 at repository revision 1581d4c. Download the manuscript citations.