Kakeya maximal estimates and four-dimensional sets

Source and verification

This article explains family 074 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 074 contains two different Kakeya claims: a three-dimensional maximal-operator bound and a four-dimensional Hausdorff-dimension theorem. Their common object is a family of thin tubes with many directions, but the analytic and geometric conclusions have different strengths.

🏷️ Preliminaries: tubes and competing dimensions

A Kakeya set in ${\mathbb{R}}^n$ contains a unit line segment in every direction. Such a set need not have positive volume. The dimension question asks whether it must nevertheless have Hausdorff dimension $n$.

Let $T_{\delta }(a,e)$ be a tube of length one and radius $\delta$, located at $a$ and pointing in the unit direction $e$. Its volume is comparable to ${\delta }^{n-1}$. The tube maximal operator is

$$\begin{array}{c}{K}_{\delta }f(e)\\ =\underset{a}{\sup }\frac{1}{|T_{\delta }(a,e)|}{\int }_{T_{\delta }(a,e)}|f(x)|dx.\end{array}$$

The supremum chooses the tube’s position separately for each direction. A bound on this operator constrains how much the selected tubes can overlap.

Minkowski dimension uses coverings at one common scale. Hausdorff dimension allows many scales in the same cover, through sums ${\sum }_i(\mathrm{diam}{U}_i{)}^s$. A one-scale volume estimate therefore needs an additional multiscale argument before it yields a Hausdorff theorem.

🏷️ The two claims

Family 074, as stated in the manuscripts

In ${\mathbb{R}}^3$, for every $\epsilon >0$,

$$\Vert K_{\delta }f{\Vert }_{L^3(S^2)}\le C_{\epsilon }{\delta }^{-\epsilon }\Vert f{\Vert }_{L^3({\mathbb{R}}^3)}.$$

Separately, every Kakeya set in ${\mathbb{R}}^4$ has Hausdorff dimension four. The latter statement allows arbitrary sets and arbitrary choices of their witnessing segments. (OpenAI, 2026) (OpenAI, 2026a)

The first estimate permits every positive power loss, with a constant depending on that loss. It is not a uniform bound independent of $\delta$. The second manuscript does not assert the four-dimensional maximal conjecture.

🏷️ A consequence at one scale

For a bounded Kakeya set $E\subset {\mathbb{R}}^3$, let $E_{\delta }$ be its $\delta$-neighborhood. Suitable unit tubes around its witnessing segments lie in a constant multiple of this neighborhood. Thus $K_{\delta }{1}_{E_{c\delta }}(e)$ is bounded below uniformly in $e$. The claimed estimate implies

$$1{\lesssim }_{\epsilon }{\delta }^{-\epsilon }|E_{c\delta }{|}^{1/3},|E_{c\delta }|{\gtrsim }_{\epsilon }{\delta }^{3\epsilon }.$$

Since $\epsilon$ can be arbitrarily small, this forces full lower Minkowski dimension. The calculation shows precisely how a subpower operator loss still gives a sharp dimension conclusion. It does not by itself prove the corresponding Hausdorff statement for mixed-scale covers.

🏷️ Three dimensions: geometry and information

The maximal manuscript organizes a hypothetical failure into finite weighted tube configurations and then into nearly extremal configurations across scales. It records line positions, velocities, times, and transverse directions at compatible resolutions. The Shannon entropy of a finite random label, $H(X)=-{\sum }_x{p}_x\log p_x$, measures the effective number of choices left in that label.

Projection theorems require certain coordinate labels to gain entropy when observed in different frames. Excessive tube concentration demands less entropy growth. The proof aims to make these requirements contradictory.

A significant issue is dependence. Two observations can share a tube index, so their joint law need not be the product of its marginals. The source compares the actual joint distribution with a conditional product distribution and charges the discrepancy to mutual information. Its projection argument needs both the geometric estimate for the product distribution and control of the information cost of passing back to the actual one.

🏷️ Four dimensions: extending quadratic descriptions

The four-dimensional manuscript begins from a hypothetical Hausdorff deficit and produces a weighted line problem. It describes portions of that problem using quadratic tests on moving spatial boxes and time intervals. Each local description must capture substantial incidence mass; a good approximation of a negligible fragment would not advance the proof.

There is a tradeoff between the thickness of a quadratic approximation and the length of the interval on which it remains valid. The source selects extremal profiles for these quantities and separates three regimes according to how rapidly the configuration becomes narrow. Local comparison of scalar sheets, lower-dimensional projection estimates, and a correction restoring the required quadratic degree are used in the different regimes.

The final obligation is to assemble local improvements into a collection with enough total mass and recoverable labels in the original configuration. This assembly is essential: local fits alone do not contradict the original dimension deficit. The article records this architecture without treating the extensive comparison and assembly lemmas as independently verified.

🏷️ Key steps and proof outlines

Discretization keeps the weights and the time variable

In the three-dimensional argument, a tube configuration is represented by indexed lines and fine space–time cells. An incidence carries the weight of a particular line visiting a particular cell. It is important to retain the line index: two observations at different times are then known to belong to the same trajectory. If $\mu (Q)$ is the incidence weight in a fine cell $Q$, the expression ${\sum }_Q\mu (Q{)}^2$ counts weighted collisions. Cauchy–Schwarz bounds an average concentration parameter by this collision count divided by total incidence mass.

The proof then groups collisions by relative slope. Two nonparallel lines remain close only for a time interval controlled by their separation in slope. This turns geometric transversality into a bound on the number of common fine cells. It supplies the initial critical inequality from which the manuscript defines its extremal concentration profile. The later argument concerns equality or near-equality in that inequality, not arbitrary tube configurations. (OpenAI, 2026b)

Extremal profiles constrain every local improvement

The stationary-model construction varies the exponents in the critical inequality and differentiates the optimized bound at a suitable point. The derivative records how much scale length or clustering a near-extremizer must exhibit. If a local region could systematically improve the critical estimate, inserting that improvement into the original configuration would contradict extremality. Hence the selected model must satisfy both concentration identities and quantitative nonconcentration estimates.

A representative rescaling subtracts a parent matrix and rescales space while leaving time unchanged. The source explicitly tracks how many old spatial cells and time bins fit into a new cell. This accounting matters because a gain in geometric shape can be cancelled by an overlooked multiplicity in the rescaling. Common finite labels are selected only after negligible conditional-probability tails have been removed, so the retained mass and conditional counts refer to the same restricted configuration.

Two frames produce incompatible entropy requirements

For finite labels $X,Y$ and a context $Z$, conditional mutual information is

$$I(X;Y\mid Z)=H(X\mid Z)+H(Y\mid Z)-H(X,Y\mid Z).$$

It vanishes when $X$ and $Y$ are conditionally independent. More precisely, it is the averaged relative entropy between the actual conditional joint law and the product of its conditional marginals. Thus it measures the cost of replacing dependent observations by a product experiment.

The source observes position and velocity coordinates in two moving frames. The exact linear change of frame relates widths of the observed coordinate cells. Projection estimates say that sufficiently nonconcentrated product laws must create a definite amount of new information. The stationary concentration identities impose upper bounds on the same entropy increments. The difficult comparison charges the discrepancy between the genuine same-line experiment and the product experiment to mutual information. One cannot simply call the two sampled events independent because their times were sampled independently: their shared line index still couples their positions.

The closure section combines these constraints and excludes a positive-deficit stationary profile. The original discrete inequality then yields the maximal estimate by approximating measurable input functions and summing the appropriate dyadic levels.

Four-dimensional quadratic fitting and preservation of incidence mass

The four-dimensional companion uses a different reduction. A scalar quadratic trajectory has the form $x(t)=x_0+t{x}_1+t^2{x}_2$. Introducing an auxiliary parameter $D$ gives a planar affine motion

$$z_D(t)=(x_0+t(x_1-D),D+t{x}_2),(1,t)\cdot z_D(t)=x(t).$$

The cancellation of $D$ is exact. It allows a scalar quadratic concentration problem to be treated using estimates for affine motions, with the extra parameter’s cell count explicitly included. This is the mechanism in the scalar-concentration section; replacing a quadratic by an arbitrary linear approximation would lose the required identity. (OpenAI, 2026a)

The proof separates isotropic, finitely narrow, and unboundedly narrow regimes. In the isotropic case, independent samples of directions either lie near a lower-dimensional plane or produce a determinant bounded away from zero at the relevant scale. The latter transversality forces an improvement. In the narrow cases, scalar projections and comparison of bounded-degree sheets extend the time interval over which a local description is valid.

A candidate extension becomes useful only after it is lifted back to original labels and original resolution. The source’s retained-edge and witness-budget arguments preserve enough of the initial incidence mass during repeated deletions. Each deleted label is charged once; thresholds are selected so the total charge remains below a fixed fraction of retained mass. This yields an actual improvement of the extremal time profile and closes the Hausdorff-deficit contradiction. The argument therefore includes both local geometric fitting and a global mass-preserving assembly.

See Also

• Family 073: Falconer distances uses related nonconcentration ideas while studying pairs of points rather than an arbitrary family of witnessing segments.
• Family 077: Fourier extension explains how thin tubes arise from frequency-localized oscillatory waves.

🏷️ Manuscripts and verification scope

• The Kakeya maximal conjecture in three dimensions — the pinned manuscript; its directory supplies the source and citation metadata.
• Every Four-Dimensional Kakeya Set Has Full Hausdorff Dimension — the pinned manuscript; its directory supplies the source and citation metadata.

No family-specific scope page was available at lean/docs/074.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 2026a. Every Four-Dimensional Kakeya Set Has Full Hausdorff Dimension.

🐻  OpenAI 2026b. The Kakeya maximal conjecture in three dimensions.