Erdős Unit Distance Problem: What AI Disproved

Quick answer: The Erdős unit distance problem asks for the maximum number of pairs exactly one unit apart among n distinct points in the plane. On May 20, 2026, OpenAI announced an AI-generated construction disproving Erdős’s prediction that this maximum grows like n to the power 1 + o(1). For infinitely many n, it gives at least n to the power 1 + δ pairs for a fixed δ > 0. The exact growth rate is still unknown.

Here is a question you can try with a pencil. Place \(n\) dots on a sheet of paper. How many pairs of dots can be exactly 1 centimeter apart? With 3 dots you can get 3 pairs (an equilateral triangle). With 4 dots, 5 pairs (two triangles sharing a side). As \(n\) grows, the best arrangements become complicated, and finding the true growth rate is one of the most famous problems in discrete geometry.

Small Cases You Can Draw

A pair counts once (AB and BA are the same pair), the points must be distinct, and the segments may cross.

Points \(n\)Maximum unit distancesExample
21A single segment
33Equilateral triangle
45Two triangles sharing a side (a rhombus with one diagonal)
57Three triangles in a row: \((0,0), (1,0), (2,0), (\tfrac12,\tfrac{\sqrt3}{2}), (\tfrac32,\tfrac{\sqrt3}{2})\)

In the five-point example, the unit pairs are two bottom edges, one top edge and four slanted edges: 7 in total. You can check each one with the distance formula.

What Did Erdős Conjecture?

Let \(u(n)\) be the maximum number of unit-distance pairs among \(n\) points. In 1946, Paul Erdős gave a grid-based lower bound of about \(n^{1 + c/\log\log n}\) and an upper bound of \(O(n^{3/2})\). In 1984, Spencer, Szemerédi and Trotter improved the upper bound to \(O(n^{4/3})\).

Erdős conjectured that grids are essentially the best possible: \(u(n) = n^{1 + o(1)}\). The “\(o(1)\)” is an extra exponent that shrinks toward 0 as \(n\) grows. So the conjecture says that for any fixed \(\delta > 0\), you eventually cannot reach \(n^{1+\delta}\) pairs.

A lower bound shows what some arrangement can achieve. An upper bound limits what every arrangement can achieve. The true answer lies between them.

Why Rescaling a Grid Helps: A Check You Can Do

Take 25 points in a 5×5 square grid, with neighbors one unit apart. There are 20 horizontal and 20 vertical unit segments: 40 unit-distance pairs.

Now count pairs that are one step across and two steps up (or two across and one up). Each such pair is \(\sqrt5\) apart, because \(1^2 + 2^2 = 5\). There are 48 of them. Shrink the whole grid by a factor of \(1/\sqrt5\), and those 48 pairs become exactly one unit apart, more than the 40 we started with. Erdős’s construction does this on a large scale: it picks a distance that occurs in an unusually large number of ways and rescales it to 1.

What Did the AI Find?

On May 20, 2026, OpenAI announced that one of its internal models had disproved the conjecture. The construction shows that for infinitely many \(n\), there are \(n\) points with at least \(n^{1+\delta}\) unit-distance pairs for a fixed \(\delta > 0\). A fixed positive \(\delta\) is exactly what the conjecture ruled out, so the conjecture is false, even though the exact maximum is still unknown.

The method combines geometry with algebraic number theory. Instead of square grids, the points come from number fields with special structure, built using infinite class field towers and Golod–Shafarevich theory, so that one distance repeats far more often than in any grid.

How Was It Checked?

Nine mathematicians (Alon, Bloom, Gowers, Litt, Sawin, Shankar, Tsimerman, Wang and Wood) wrote a human-verified account of the proof (arXiv 2605.20695). Tsimerman called it “a really impressive piece of work,” and Gowers said he would have recommended acceptance had a human submitted it to the Annals of Mathematics; that is praise, not an actual publication. OpenAI described the result as the first prominent open problem solved autonomously by AI, which is the company’s own characterization.

Bounds and Developments (Status Checked November 21, 2026)

DateResult for \(u(n)\)Status
1946Erdős: at least \(n^{1 + c/\log\log n}\); conjecture \(n^{1+o(1)}\)Lower bound proved; conjecture now disproved
1984Spencer–Szemerédi–Trotter: at most \(O(n^{4/3})\)Proved; the classical upper bound
May 2026OpenAI model: at least \(n^{1+\delta}\) for infinitely many \(n\), fixed \(\delta > 0\)Human-verified companion paper
May 2026Will Sawin: more than \(n^{1.014}\) for arbitrarily large \(n\)Preprint (arXiv 2605.20579)
September 2026OpenAI: at most \(O(n^{4/3 – \varepsilon})\) for some \(\varepsilon > 0\)Preprint, the first improvement on the 1984 exponent
October 2026Eric Naslund: exponent about 1.0432Computer-assisted preprint, partly formalized in Lean; not yet peer reviewed

So the exponent of \(u(n)\) is now known to lie somewhere between about 1.04 and a little under \(4/3\). Nobody knows where.

Unit Distances vs. Distinct Distances

Erdős posed a sister problem in the same 1946 paper: how few different distances can \(n\) points determine? A unit square shows the difference: it has 4 unit-distance pairs but only 2 distinct distance values, 1 and \(\sqrt2\). The distinct distances problem was essentially solved by Larry Guth and Nets Katz in 2010; the unit distance problem is still open. For more famous problems like these, see our guide to the Erdős problems.

Watch: The Unit Distance Problem in Video

If you learn better by watching, these videos are worth your time.

The Erdős Breakthrough (OpenAI). OpenAI’s own explanation of the unit distance result.
BREAKING NEWS: OpenAI has disproved Erdős’ unit distance conjecture (Alvaro Lozano-Robledo). A mathematician explains the statement and why the result surprised experts.
Why AI Just Disproved Erdős: The Unit Distance Problem (Noesis9). A short walkthrough of the grid construction and why it was beaten.

Check Your Understanding

Five quick questions to see how much of this has stuck. Pick an answer and press Submit; you will see a short explanation either way.

Answer each question and press Submit to see your result right away.

1. What does u(n) count in the unit distance problem?
Found a bug? Let us know!
2. How many unit distances can 3 points have at most?
Found a bug? Let us know!
3. What did Erdős conjecture about u(n)?
Found a bug? Let us know!
4. What did the 2026 AI result show?
Found a bug? Let us know!
5. Did the 2026 construction find the exact maximum number of unit distances for every n?
Found a bug? Let us know!

Frequently Asked Questions

What is the Erdős unit distance problem?

It asks for the largest number of pairs of points at distance exactly 1 among n points in the plane, and how fast that number grows as n increases. Paul Erdős posed it in 1946.

Did AI solve the Erdős unit distance problem?

An internal OpenAI model disproved Erdős’s conjecture about it, announced on May 20, 2026, by showing arrangements with at least n to the power 1 + delta unit distances. The exact growth rate is still unknown.

What are the current bounds for the unit distance problem?

As of November 2026, the best lower-bound exponent reported is about 1.043 (Eric Naslund, a computer-assisted preprint), building on the AI construction and Will Sawin’s explicit 1.014. A September 2026 OpenAI preprint improves the 1984 upper bound to n to the 4/3 minus epsilon.

Is the unit distance problem completely solved?

No. Erdős’s conjecture about it was disproved, but the true growth rate of the maximum number of unit distances is still unknown.

Is this the same as the distinct distances problem?

No. The unit distance problem asks how many pairs can share one distance. The distinct distances problem asks how few different distances n points can have; Guth and Katz essentially solved it in 2010.

Who checked the unit distance proof?

Nine mathematicians, including Noga Alon, Thomas Bloom, Tim Gowers, Will Sawin, Jacob Tsimerman and Melanie Matchett Wood, wrote a human-verified companion paper (arXiv 2605.20695).

Keep Exploring

Sources and Further Reading

Related to This Article

What people say about "Erdős Unit Distance Problem: What AI Disproved - Effortless Math"?

No one replied yet.

Leave a Reply