How does a simple counting trick, run inside a special kind of number system, land points exactly one unit from the center?

This explores the step in the recent unit-distance constructions where a counting argument inside a CM field produces points of absolute value exactly 1. Points at distance 1 from the origin are what turn into many pairs of points exactly one unit apart. The collection does not spell this step out, so the explanation below comes from the standard number theory these papers build on rather than from the notes themselves. What the collection does have is the context: OpenAI's counterexample to Erdős's conjecture uses classical tools (CM fields and Golod–Shafarevich towers), and its new ingredient was letting the field's degree grow without limit What made OpenAI's unit distance counterexample succeed?. A follow-up made the gain explicit, giving more than n^1.014 unit-distance pairs among n points How many unit distances can points in a plane have?.

The intuition is this. A CM field has a built-in 'complex conjugation' that behaves the same way under every way of placing the field inside the complex numbers. So if an element α times its conjugate ᾱ equals a positive real number, then the ratio α/ᾱ has absolute value exactly 1, and that holds simultaneously in every complex embedding. That uniformity is what makes CM fields the natural place to look. The hard part is finding many such elements. Pigeonhole does that job: build lots of ideals out of a split prime, and since there are only as many 'ideal classes' as the class number, many of those ideals must land in the same class. The quotient of two same-class ideals is principal, so it has a generator, and dividing that generator by its conjugate gives a point on the unit circle. More collisions mean more unit-modulus points, and differences between them give unit distances.

This also explains why growing the degree mattered. Pigeonhole pays off most when the number of classes is small compared with the number of objects you are sorting. Letting [K : Q] → ∞ while a fixed split prime keeps the class number and discriminant under control is what tips the count from linear to superlinear What made OpenAI's unit distance counterexample succeed?. The method is old; the new part is choosing a setting where the old counting argument wins by a margin.

There's a wider lesson in how AI-assisted mathematics is being judged. Tao's view is that an opaque tool's output is fine to trust when it can be checked independently Can opaque machine learning models help prove new mathematics?. The Leiden Declaration says human authors still carry responsibility for correctness and for explaining the result Can AI-generated proofs ever replace human mathematical understanding?. The question asked here is exactly that kind of explanation: knowing why the pigeonhole step lands on the unit circle is what turns a machine-found construction into mathematics people understand. Read the two unit-distance notes for the construction itself. For a full proof of the pigeonhole step, you'll need the original papers.