MathematicsPreprintTheory3 min read

IF AB = 1, IS BA = 1? NOT ALWAYS

With ordinary numbers, if a × b = 1, then b × a = 1: the order of multiplication does not matter. In many algebraic worlds it does. A ring — a set of objects that can be added and multiplied, like whole numbers — is called directly finite when the shortcut still holds: whenever ab = 1, then also ba = 1.

The question concerns a particular kind of ring, the group algebra. Take a group G — a set of elements that can be combined and undone, like the rotations of an object — and a field k of coefficients, a number system in which one can add, subtract, multiply and divide, like the fractions; the elements of the group algebra k[G] are finite sums of group elements, each multiplied by a coefficient. In the book Fields and Rings, Kaplansky conjectured that every such group algebra is directly finite.

Proved again and again

The conjecture was confirmed in many cases, according to the paper:

  • Kaplansky proved it in characteristic zero — for coefficients such as fractions or real numbers.
  • Ara, O’Meara and Perera proved it for a large family called amenable groups, whatever the field.
  • Elek and Szabó extended it to sofic groups, those that can be approximated by actions on finite sets.

These proofs share a tool: a way of assigning a “rank”, a size, to every matrix over the algebra. If such a rank never gives zero to a nonzero element, the algebra is directly finite. The open question was whether some group escapes every such rank.

A counterexample with 1 + 1 = 0

In a manuscript dated 23 September 2026, OpenAI constructed a group algebra that is not directly finite, with coefficients in characteristic two — a number system in which 1 + 1 = 0. Andrei Jaikin-Zapirain now gives what the note calls a “human-readable exposition” of that construction, then pushes it further: for every prime number p, there is a group whose algebra with coefficients modulo p is not directly finite. And these groups are hyperbolic: they act on a space with negative curvature.

The construction is a striking mix of fields:

  • Finite geometry. Take the points and straight lines of a space whose coordinates belong to a finite number system with q elements, remove some points called “holes”, and keep only lines that avoid them, so that every remaining point lies on exactly the same number of lines.
  • A graph without short loops. Points and lines form a network. If its shortest loop has at least 14 steps, gluing small finite groups along it produces a hyperbolic group.
  • Polynomials that cancel on every line. A low-degree polynomial always sums to zero along a full line of this space. Such polynomials, built with tools from error-correcting codes, supply matrices that no rank can treat consistently.
  • Randomness. The lines are partly chosen at random, and a probabilistic tool, the Lovász local lemma, guarantees that some choice avoids every short loop.

The scale is extreme: the space has 1,680 dimensions, and each point lies on more than q¹⁰⁰ lines. Finding a suitable prime in that range calls on recent results about prime numbers in arithmetic progressions.

Non-sofic groups, almost by accident

Because sofic groups always satisfy the conjecture, any counterexample must come from a group that is not sofic. According to the note, the very first examples of non-sofic groups were constructed only a few months earlier, also by OpenAI. The new construction immediately gives non-sofic hyperbolic groups. Sandhu and Thom have built one too, with related ideas.

A crowded month

The note lists further counterexamples announced around the same time: one in a specific odd characteristic, one with a finitely presented torsion-free group, one by Bunina, Lee, Miasnikov and Muliarchyk for every odd prime, and one by Mghirbi and Douglas for every prime with torsion-free hyperbolic groups, which also disproves three other conjectures of Kaplansky — on zero divisors, idempotents and units — for that class.

The author’s AI-use statement is explicit. The original construction is due to OpenAI; AI assistance also contributed to developing the ideas for the extensions, and ChatGPT was used to review arguments and calculations. The author takes full responsibility, including for verifying every statement and proof: “AI systems are not authors of this paper.”