AI Solved a 79-Year-Old Math Problem No Human Could Crack: The Erdős Unit Distance Conjecture Is Dead
- David Borish
- 1 day ago
- 5 min read
A Conjecture That Lasted Nearly Eight Decades
In 1946, Paul Erdős asked a deceptively simple question: if you place n points in the plane, how many pairs of them can be exactly distance 1 apart? He conjectured that the count could never grow faster than roughly n raised to the power 1 plus a correction term that shrinks toward zero, specifically n^(1 + C/log log n) for some absolute constant C.
The question sat at the intersection of geometry and combinatorics, and progress came slowly. By 1984, Spencer, Szemerédi, and Trotter had proven that the count could not exceed n^(4/3), an upper bound that remained the state of the art for decades, with only improvements to the constant factor. Meanwhile, the best known constructions barely exceeded n^(1 + c/log log n), which seemed to support Erdős's prediction. As recently as 2025, new results on unit distances for generic and typical norms were interpreted as evidence the conjecture was correct.
It was not correct. The new proof establishes that for some fixed positive constant δ, there exist infinitely many values of n for which the maximum unit-distance count exceeds n^(1+δ). That is a qualitatively stronger growth rate than Erdős conjectured, and it holds not just occasionally but for an infinite sequence of point set sizes.
How the Proof Works
The construction is geometric in its conclusion but arithmetic in its mechanism. The core idea is an extension of a classical observation about Gaussian integers: in the ring Z[i], a product of many primes that are 1 mod 4 has a large number of factorizations of the form z times its conjugate, which correspond geometrically to lattice vectors of the same length. The new proof replaces the Gaussian integers with a more elaborate algebraic structure.
The argument begins by building an infinite tower of number fields. Starting from a cyclic cubic base field F, the proof constructs an ascending chain of totally real number fields F_0, F_1, F_2, ... where each extension is unramified and has a Galois group that is a 3-group.
The Golod-Shafarevich theorem, a classical result in group theory, guarantees that such a tower can be made infinite. The key constraint is that the root discriminant of the fields, a measure of their arithmetic complexity, stays bounded throughout the tower. Minkowski's theorem then ensures that the class numbers of these fields grow at most exponentially in the field degree, which turns out to be the critical parameter for controlling how many useful arithmetic objects the construction produces.
To each field in the tower, the proof adjoins i to form a CM field, a totally imaginary extension. Elements of norm 1 in this CM field correspond, under every complex embedding, to complex numbers of absolute value 1. These are the candidate unit-distance translations. The number of such elements that can be harvested from any given level of the tower is controlled by the class number and the number of rational primes that split completely in the field. By using Chebotarev's density theorem to choose a fixed set of split primes and ensuring the tower remains infinite after the relevant Frobenius conditions are imposed, the proof guarantees a supply of norm-one elements that grows exponentially in the field degree.
The geometric part then takes these norm-one elements and turns them into an actual point set in the plane. The elements are embedded in a high-dimensional Minkowski space via the complex embeddings of the field, cut by a product of discs, and projected down to a single complex coordinate. The injectivity of this projection follows from the fact that a nonzero algebraic number cannot have any complex embedding equal to zero. The result is a planar point set in which the number of unit-distance pairs is proportional to the size of the set raised to a power strictly greater than 1.
The Role of the AI System
The paper includes an explicit statement about how the proof was produced. The problem was given to an internal model as a formal prompt, and the model's output was evaluated by an AI grading pipeline before any human examined it. Only after the automated evaluation indicated high confidence in the result did human researchers and mathematicians review the work. External number theory experts subsequently confirmed the proof's correctness and have since simplified parts of the argument.
The paper reproduces both the original prompt given to the model and the model's raw output verbatim, before any automated grading or human rewriting. The raw output is recognizable as the proof, written in compressed mathematical notation. What appears in the published paper is described as a human-edited exposition of the autonomous solution, with references, reorganized proofs, and explanatory material added afterward.
This framing matters because it draws a clear line between where the mathematical work occurred and where the presentation work occurred. The structure of the proof, the key insight connecting number field towers to planar geometry, and the sequence of lemmas that make the argument go through all came from the model. The published version is a cleaned-up rendering.
What This Result Means for Discrete Geometry
The Erdős unit distance conjecture was not an isolated problem. It belonged to a cluster of questions about distances in point sets that drove a significant portion of 20th-century combinatorial geometry. The distinct distances problem, asking how few distinct distances n points must determine, was resolved up to logarithmic factors by Guth and Katz in 2015. The unit distance problem resisted resolution for thirty years after that.
The new result does not determine the exact growth rate of the unit distance function, only that it exceeds Erdős's bound. The Spencer-Szemerédi-Trotter upper bound of n^(4/3) still stands. The gap between n^(1+δ) for some small fixed δ and n^(4/3) is the new frontier.
The proof's reliance on number field towers also raises a practical question: the construction works for an infinite sequence of point set sizes, but those sizes are not specified and may be astronomically large. The result is existential rather than constructive in a practical sense. No one is going to draw one of these point sets. What the proof provides is a definitive answer to the asymptotic question, not a recipe for building the configurations.
The Broader Pattern
This result is not the first time AI systems have contributed to mathematics. Systems have verified existing proofs, assisted with formalization, and been used to search for counterexamples. What is less common is autonomous production of a full solution to a major open problem, from problem statement to complete proof, without the solution being guided by human mathematical intuition at each step.
The proof is technically dense. It calls on the Golod-Shafarevich theorem, Chebotarev's density theorem, the Hajir-Maire class field tower method, Shafarevich's relation-rank estimate, and Minkowski's ideal-class bound. These are not obscure tools, but assembling them in the right sequence to attack a geometric problem in combinatorics required crossing disciplinary lines in a way that human researchers had not managed.
Whether the model found the connection by analogy to known strategies, by search over possible proof structures, or through some other mechanism is not described in the paper. What is described is the output: a correct proof of the negative resolution of the Erdős unit distance conjecture, produced autonomously, that human experts have verified and are now working to build upon.
The 1946 conjecture is now a theorem, in the negative direction. It took 79 years and required, in the end, a machine.
Comments