OpenAI finds point sets that advance Erdos unit-distance bound

An internal general-purpose model developed by OpenAI produced families of points in the plane that guarantee at least n^(1+δ) pairs at unit distance, a polynomial improvement over earlier lower-bound constructions. Mathematicians at Princeton University reviewed the output and confirmed the constructions and the supporting arguments.

The unit-distance problem, posed by Paul Erdős in 1946, asks how many pairs of points can be exactly one unit apart among n points in the plane. For decades, the best known explicit constructions produced counts only marginally larger than linear in n, typically written as n^(1+o(1)).

The OpenAI model proposed a new family of point arrangements that achieve at least n^(1+δ) unit-distance pairs for some fixed δ>0 that does not shrink as n grows. The constructions combine geometric arrangements with methods drawn from algebraic number theory, according to researchers involved in the review.

Princeton mathematicians examined the constructions, checked the calculations, and verified the mathematical arguments that establish the n^(1+δ) lower bound. The verification focused on the explicit point sets the model produced and the proofs showing how the number of unit-distance pairs scales with n.

The model used for the search was a general inference system under evaluation, not a system trained exclusively on formal mathematical proofs. Researchers reported that the output required subsequent human checking and formalization before verification was complete.

Tim Gowers wrote that the advance could have implications beyond geometry, including for cryptography and proof techniques in other fields. Arul Shankar described the result as a significant advance for discrete geometry.

Researchers involved noted that the result is a lower-bound existence statement: it provides explicit point sets that reach the higher count, rather than giving a new upper bound that applies to all possible point arrangements. They also noted that some areas of combinatorics, coding theory and cryptography depend on specially constructed examples, and that tools that produce new constructions may alter how researchers search for examples.

Background: Erdős posed the unit-distance problem in 1946. Traditional lower-bound constructions placed points on lattices or grids and adjusted scales to increase unit separations. Prior constructions left the exponent in the counting function only marginally above 1; the OpenAI-derived families are reported to be the first verified constructions with a fixed exponent strictly greater than 1.