The trilinear Hilbert transform at the symmetric exponent

Source and verification

This article explains family 086 of the October 6, 2026 OpenAI collection. The research conclusions are attributed to its manuscripts. The source review covers statements, introductions, and the proof mechanisms described below; it is not an independent verification of the complete proofs.

Overview

Family 086 concerns the trilinear Hilbert transform on the line with slopes $1,2,3$. It claims the symmetric $L^3\times L^3\times L^3\to L^1$ bound. Quadratic phase patterns explain why the problem needs more than estimates for a fixed scale.

🏷️ Preliminaries: the operator and its scaling

For Schwartz functions $f_1,f_2,f_3:\mathbb{R}\to \mathbb{C}$, define

$$\begin{array}{c}T(f_1,f_2,f_3)(x)\\ =p.v.{\int }_{\mathbb{R}}\prod _{j=1}^3{f}_j(x-jt)\frac{dt}{t}.\end{array}$$

The principal value removes a symmetric interval around zero. For smooth inputs, subtracting the value at $t=0$ exposes cancellation there; rapid decay handles infinity.

The exponent relation $1/3+1/3+1/3=1$ is consistent with Hölder’s inequality and scaling. If $f_j^{\lambda }(x)=f_j(\lambda x)$, then a change of variable in $t$ gives $T(f_1^{\lambda },f_2^{\lambda },f_3^{\lambda })(x)=T(f_1,f_2,f_3)(\lambda x)$. Both the output $L^1$ norm and the product of input $L^3$ norms scale by ${\lambda }^{-1}$.

🏷️ The manuscript’s claim and its scope

Family 086, as stated in the manuscript

There is a finite constant $C$ such that

$$\Vert T(f_1,f_2,f_3){\Vert }_1\le C\prod _{j=1}^3\Vert f_j{\Vert }_3$$

for Schwartz inputs. Hence the operator has a unique bounded trilinear extension from $L^3(\mathbb{R}{)}^3$ to $L^1(\mathbb{R})$. (OpenAI, 2026)

The statement concerns the fixed slopes and displayed exponents. It does not state the full conjectured exponent range, a maximal-truncation theorem, or pointwise principal-value convergence for every triple of $L^3$ representatives. A bounded extension by density is a norm-level conclusion.

This operator is also distinct from the triangular Hilbert transform in family 082, whose inputs depend on two spatial variables.

🏷️ Quadratic modulation survives a four-term progression

Dualizing the $L^1$ norm introduces a bounded test function $f_0(x)$. After reversing the sign of $t$ if necessary, the four arguments lie at $x,x+t,x+2t,x+3t$.

Let $(c_0,c_1,c_2,c_3)=(-1,3,-3,1)$. Direct expansion shows

$$\begin{array}{c}\sum _{j=0}^3{c}_j(x+jt{)}^a=0,\\ a=0,1,2.\end{array}$$

Indeed, the sums of $c_j$, $j{c}_j$, and $j^2{c}_j$ all vanish. Thus for every polynomial $Q$ of degree at most two,

$$\prod _{j=0}^3{e}^{2\pi i{c}_{j}Q(x+jt)}=1.$$

Multiplying the four inputs by these matching phases leaves their product unchanged. This exact identity explains the persistence of quadratic modulation: large individual oscillations need not produce cancellation in the combined form.

🏷️ Why an inverse theorem is only an intermediate tool

A higher-order inverse theorem detects quadratic structure when a function has a large suitable uniformity norm. The manuscript uses a degree-two quantitative inverse theorem as an external input. That input can identify structured pieces at a given stage, but it does not automatically control how many pieces or scales a singular-integral proof will revisit.

The source therefore develops localized continuous descriptions of the quadratic structure, together with estimates for their behavior under repeated projections. Selected coefficients must remain recoverable after earlier selections. Its compression lemma controls products of these projections using a common budget for coefficient sizes and the number of labels along a spatial path.

The second ingredient localizes curvature in all four arguments at once. Adjoints of later projections may be evaluated on earlier cells, so the necessary separation has to persist through the construction. Comparing coupled narrow frequency bands introduces a block-length loss whose exponent can be made small by an auxiliary parameter.

The third ingredient separates errors with square-summable energy from coherent components that retain consistent representatives. A sparse counting estimate gives a positive power saving in a density parameter for the remaining progression tests. The final proof chooses the parameters so that this gain absorbs the accumulated losses.

🏷️ The significance of scale uniformity

On a single annulus $\epsilon <|t|<R$, taking absolute values and using Hölder yields a bound proportional to $\log (R/\epsilon )$. This follows by integrating $dt/|t|$ over the annulus. Such an estimate becomes useless as the truncation range grows.

The claimed theorem replaces this accumulating cost by a constant independent of the number of scales, using cancellation together with the structural and counting estimates above. The elementary modulation identity identifies an obstruction to naive cancellation; it is not itself a proof of the uniform bound.

🏷️ Key steps and proof outlines

Local quadratic detection must be quantitative and stable

The dual four-input form is localized to spatial intervals and finitely many singular-integral scales. If an input has large higher-order uniformity on such an interval, the inverse theorem identifies a quadratic structured component. The source represents that component by a localized quadratic chart, including smooth cutoffs and finite horizontal coordinates. These charts are closed under conjugation, multiplication, translation, and the fixed affine changes among the four progression arguments.

Closure under these operations matters because later projections act on previously selected components. A representation useful only at its first detection would be lost as soon as an adjoint projection was moved to an earlier cell. The chart-expansion and Weyl-alternative lemmas provide quantitative complexity and tail bounds after each allowed operation. Their output is therefore both a structured approximation and a certificate that later manipulations remain in a controlled class. (OpenAI, 2026)

Compression controls repeated use of structured directions

Successive inverse-theorem applications could produce an ever-growing list of charts. The compression argument keeps coefficient budgets and private coordinates that distinguish selected columns. It estimates products of inherited projections, not merely each individual projection’s norm. This is the mechanism that limits the accumulated complexity along a spatial path.

The stopping construction is organized into attempts. If the exceptional child roots in each attempt have total length at most $\rho$ times the parent length, with $\rho <1$, then the total length of all generation-$n$ roots is at most ${\rho }^n|J|$. Hence

$$\sum _{n\ge 0}\sum _{I\text{ in generation }n}|I|\le \frac{|J|}{1-\rho }.$$

Several stopping conditions can be imposed by taking their maximal exceptional intervals, provided the sum of their fractional costs remains below one. The source uses this restart budget to convert an estimate for successful attempts into a bound over the entire interval. It prevents unsuccessful detection or localization stages from being ignored.

Curvature separation has to hold in all four slots

For a quadratic phase, curvature is its quadratic coefficient after normalization to the current interval. Products of charts carry differences of these coefficients. If two localized columns have a large inner product, a smoothed interval test detects a large Fourier coefficient of their product. The quantitative Weyl alternative then forces their curvature data to be compatible at the prescribed accuracy.

This implication builds an ambient graph of compatible tags and short paths. Colors with controlled changes organize regions where one coherent representative can be retained. The construction uses two accuracy levels: one controls structural approximation and another controls the separation required after later adjoints are pulled backward. Identifying these tolerances would leave no room for accumulated localization errors.

The four arguments $x+jt$ are treated together. The exact quadratic modulation identity above shows why detecting curvature in only one input is inadequate: its phase can be cancelled by matching phases in the other three inputs.

Persistent components and square-summable errors are separated

The linearization section expands a coherent component at its first compatible match, fixes the rational splitting there, and controls the resulting tails. Persistent components retain representatives through later scales. The remaining errors are placed in a square-summable energy budget. This distinction determines which summation tool is available: coherent terms need a counting argument, while energy errors can be handled by Cauchy–Schwarz and orthogonality-type estimates.

On short depth blocks, side terms are compared using mixed-norm maximal estimates and weighted extrapolation. Their losses depend on block length with an exponent that can be made small by an auxiliary parameter. The source does not claim these terms vanish; it chooses the parameters so that their cost is dominated by the later positive gain.

A common successful height makes the final count meaningful

The counting argument puts the finitely many labels, shifts, and scales on one sufficiently fine grid. Approximations preserve the required separation, including the divisions by the progression coefficients. Selecting one common successful height ensures that the counted frequency tests belong to compatible data rather than to unrelated favorable slices.

The resulting sparse count gives a positive power saving in density. In the completion, first fix that saving, then choose the comparison parameters and block lengths so all accumulated losses have smaller exponents. Sum the retained-cell estimates and the restart errors. This supplies a constant independent of the number of singular-integral scales. Density then extends the Schwartz-input operator in norm, with the scope limitations stated earlier.

See Also

• Family 082: triangular Hilbert transforms uses matrix-energy telescoping for a different entangled singular integral.
• Family 075: Fourier convergence also needs to organize many possible phases without paying separately for every frequency or scale.

Pointwise multiple ergodic averages explains how companion papers strengthen the local fixed-input estimates for depth-dependent oscillation tests and transfer them to orbit averages.

🏷️ Manuscripts and verification scope

• An $L^3$ bound for the trilinear Hilbert transform — the pinned manuscript; its directory supplies the source and citation metadata.

No family-specific scope page was available at lean/docs/086.md in this snapshot. This records the catalogue lookup, not a claim that no related Lean code exists. No Lean build or axiom audit was performed.

OpenAI collection gives the reading sequence and source policy; coverage register records which families have articles and which remain to be written.

🐻  OpenAI 2026. An L³ bound for the trilinear Hilbert transform.