Zulip Chat Archive

Stream: PrimeNumberTheorem+

Topic: Fixing a sign error in Erdos 392


Terence Tao (Jan 16 2026 at 21:00):

When reviewing the blueprint graph I noticed an oddity in the Erdos 392 project: the final theorems (erdos-sol-1, erdos-sol-2) were mysteriously disconnected from the rest of the Erdos 392 lemmas; and the (AI-provided) proofs were suspiciously short. After some inspection I realized the problem: I had asked to prove the (trivial) statements that n!n! can be factored into at least nn/logn+o(n/logn)n-n/\log n + o(n/\log n) factors of size n\leq n, and at least n/2n/2logn+o(n/logn)n/2-n/2\log n + o(n/\log n) factors of size n2\leq n^2, when in fact the Erdos problem asks to factor into at most nn/logn+o(n/logn)n-n/\log n + o(n/\log n) factors of size n\leq n, and at most n/2n/2logn+o(n/logn)n/2-n/2\log n + o(n/\log n) factors of size n2\leq n^2. (The trivial factorization n!=1××nn! = 1 \times \dots \times n would satisfy the previous versions of the theorems, which is what the AI-generated proof found.) So I am reopening PNT#648 PNT#649 to actually finish the proof of Erdos 392 (which is otherwise almost done outside of two small lemmas, including PNT#515 which is currently unclaimed).

Another cautionary tale not to blindly trust AI auto-formalization, even when it typechecks...

Eric Vergo (Jan 16 2026 at 21:23):

It sounds like you were able to catch this, in part, because the dependency graph is generated using the LeanArchitect system. Maybe checking for a disconnected dependency graph can be automated and happen as the dependency graph is being generated. It would be especially useful when working towards a specific result, such as this project.

Thomas Zhu (Jan 16 2026 at 21:43):

I'm happy our LeanArchitect tool helped in this case! Indeed, this level of dependency debugging is also one of its purposes.

Eric Vergo (Jan 16 2026 at 21:47):

Awesome, I made a feature request for this.

Pietro Monticone (Jan 16 2026 at 21:48):

Cool! Thanks @Terence Tao and @Eric Vergo.

Terence Tao (Feb 23 2026 at 01:07):

Happy to announce that the formalization of the solution to Erdos 392 is now complete. Many thanks to everyone who contributed to this solution, and particularly to @Alex Kontorovich for his heroic efforts in formalizing the final step (PNT#784) (which eluded two AI tools due to some non-trivial typos and refactoring required).


Last updated: Feb 28 2026 at 14:05 UTC