Enough distances?

0:00How many distances?

Take a compact set E in d-dimensional space and collect all distances between its points. That is the distance set. A set with interior has every small distance, but fractal sets can be thin. In 1985, Falconer asked how large the Hausdorff dimension of E must be to force a distance set of positive length. He showed that dimension above (d + 1)/2, is enough, and conjectured that above d/2 suffices. The threshold d/2 cannot be lowered: Falconer’s lattice-type examples have dimension exactly d/2 and distance sets of length zero. The reason is visible in a grid. In an N × N grid of spacing 1/N, every squared distance is a whole number over N², so the distances crowd onto at most about 2N² values, and only about N²/√(log N) of them actually occur. Replace each point by a disc slightly smaller than 1/N², and repeat with ever finer grids. The limit set has dimension one, which is d/2, yet its distances fill a set of length zero. Above the threshold, the conjecture says no compact set can crowd its distances like that. Progress came through Fourier analysis. Mattila turned the problem into estimates for spherical averages. In the plane, Wolff reached 4/3, and Guth, Iosevich, Ou and Wang reached 5/4. In space, Erdoğan reached 11/6; then, using decoupling, Du, Guth and coauthors reached 9/5, and Du, Ou, Ren and Zhang 12/7. At the critical dimension itself, Shmerkin and Wang showed that sets with equal Hausdorff and packing dimension d/2 have distance sets of dimension one.

1:46The claim

Result 073 is one OpenAI preprint, 64 pages, dated September 23. It claims the conjecture in every dimension d of at least two: every compact set of Hausdorff dimension greater than d/2 determines distances of positive Lebesgue measure. No regularity is assumed, the inequality is strict, and the statement is about the full distance set, not distances from a single pin. The paper says it uses no decoupling theorem and no earlier distance theorem at an improved threshold. OpenAI also published a Lean proof of the theorem in every dimension, with the planar case as a separate statement: about 2,500 files and 190,000 lines. We have not run it.

2:33Angles and shells

The proof starts from two separated Frostman measures on E, of dimension s greater than d/2, and pushes their pairs forward by the distance map. The goal is to show that this distance measure has a density. Two issues stand in the way. First, directions: the directions from one point to the other piece of the set need cap bounds of exponent d/2, even when radial projections cannot have densities. The paper builds those bounds scale by scale, deleting small exceptional sets of pairs. Integer exponents come from projecting to smaller spaces, where radial estimates are available. In odd dimensions, a thin-tube improvement adds the missing half dimension, using the incidence theorem of Orponen and Shmerkin. A Fourier calculation then turns cap bounds into estimates for belts on the sphere. In each frequency annulus of radius B, the bound carries a factor B to the power d minus twice the cap exponent, which is exactly one at the critical value d/2. Second, these angular bounds must become a distance estimate, for a fractal that can be irregular across scales. At each frequency, the measures split into almost uniform classes with mass profiles. A recursion trades a count of pairs with nearly equal distances against spherical Fourier energies, through stationary phase and a graph estimate, and a potential built from two separated depths pays for every step. The result is a power-small bound on the Fourier energy, in each dyadic shell, of measures that approximate the distance measure. Finally, the pair restrictions shrink with the scale, without a quantitative rate. A monotone shell lemma handles this: if positive measures decrease to a limit, and the N-th one is close to a measure with small Fourier mass in the N-th dyadic shell, with both errors decaying geometrically, then the limit is absolutely continuous. The limit is nonzero and lives on the distance set, so that set has positive length.

4:39Where it stands

The preprint is not yet peer reviewed, and it has not been revised. OpenAI published a Lean proof; we have not run the check. It cites one other paper from the release, result 166 on distinct distances, only as background. Its one deep outside input is the incidence theorem of Orponen and Shmerkin. The rest is standard tools, such as Frostman’s lemma, stationary phase and the Loomis–Whitney inequality, plus methods of Orponen, Shmerkin and Ren that the paper re-proves in the form it needs. We found no expert review. One public GitHub repository reports that OpenAI’s Lean proof passes the comparator check; we have not verified that. If it holds, the Falconer distance conjecture is settled in every dimension: above half the ambient dimension, every compact set has distances filling a set of positive length. The paper makes no claim at exactly d/2, and none about a single pin at the threshold.