← Dean Bisogno

RH Verification

checking the Lean formalization of Alpöge's announced proof of RH/GRH · last updated 2026-11-21

Levent Alpöge has announced a proof of RH and of GRH for Dirichlet L-functions, given as a Lean 4 formalization distributed as alpo.ge/lean.tar.gz.part01–part14.

Andrew Sutherland has independently checked the same archive and reports that it passes. Separately, Alessandro Maria Masullo has circulated a human-written draft proof of GRH. Both are linked under Elsewhere.

This page contains some of my own thoughts and an independent check of Alpöge's formalization on my personal hardware: what has been checked, how, and what remains.


My own evolving thoughts

The past year has seen rapid AI adoption and considerable progress throughout the sciences, perhaps nowhere more so than in mathematics research. A confluence of modern factors now challenges practicing mathematicians: constrained department budgets, large-scale reductions in science funding, and agentic systems adept at navigating the intricate machinery of mathematical proof. For a subject that demands such high investment, the vision of a new style of mathematics is discomforting. Others have written and spoken better about the human side of mathematics (see Quanta and The Daily).

Navier–Stokes was a bellwether. But to me, an arithmetic geometer by training, a proof of GRH is akin to the moon landing. A Lean-verified proof without complete human interpretation is something else: a turn in history we watched approach for miles that nevertheless arrived sooner than some of us expected.

Evolution in our standards of proof is the history of mathematics. Euclid built geometry on axioms, while Diophantus proved by construction. In the Renaissance and early modern period, proof became laissez-faire. Tartaglia guarded the solution of the cubic as a trade secret until Cardano published it, algebra left the geometric standard of rigor behind, and results were stated without proof (Fermat's margin famously so). Algebraists rebuilt their subject on Cantor's set theory (Katz and Parshall, Taming the Unknown, 2014), and from 1900 through the 1940s Hilbert and then Bourbaki called for mathematics to return to simple axiomatic systems. An established, machine-checked proof of GRH would turn the next page. If this is the new contract of proof, then I wonder what responsibility the individual holds for verifying theorems they wish to accept.

I have been heartened to read Masullo's draft manuscript and I am excited to follow its peer review. Unpacking Alpöge's result will undoubtedly yield novel insights and techniques as well.

I'm running my own verification with an AI agent's assistance, as seems fitting for this proof. I priced out rented cloud compute, but the payoff wasn't there, so I am running it on my personal machine: AMD Ryzen 7 5700X3D (8C/16T), 32 GB RAM (WSL2 limited to 26 GB plus 48 GB swap), Ubuntu 24.04.5 on WSL2 (kernel 6.18.33.2), Windows 11, build on a PNY CS2130 500 GB SSD. Completion ETA is sometime on Monday unless I hit a memory issue.

— Dean Bisogno, 2026-11-21


Elsewhere


Status

[in progress] Steps 1–4 done. Full build (step 5) started 2026-11-21 09:55 MST. As of 11:46 MST, 6,332 of 175,187 modules are built (3.6%), with no errors. At the current rate the build would finish around 2026-11-23, though per-module cost varies.


Steps


Scope

A successful check shows the stated theorems follow from Lean's logic, Mathlib's definitions and Lean's standard axioms, assuming the kernel is correct. It depends on step 3: the theorems must say what they are claimed to say.

It does not give a human-readable proof or check the informal account in the announcement.


Checksums (SHA-256)

86974b1c242ca7b6748f3299df4fc64ee0f6e06c776ac964fb64888101a3de77  part01
1c2a2477354a9dfa97ff1ce54861e96ed885bd456fffbaaabc2028314a229c42  part02
1465a23b1b579a881e24e5ffaf7c4c3727af28ea24741f1ec80e1d67947530a3  part03
3929e327eecb4e2c5c7cae30e104e6a4a52fa6b82d28325bcda8e9da1c86893b  part04
f25d7e80f7c4ad6d674aa18b323ca760f848857c8755501c149317c38a66c75d  part05
a8c651a68327811b37dce3cca8eda5e0b2102fa227c0d395cbb7553e4dd69939  part06
65c7e502a9b236a0f68d9e0aa6e089d4c09b9fc7a0960a149384063ccbe5e869  part07
71b027ad6e4179c3a50ce72921e3fec0a2dd751ec7472ac2c644ad145a50ec7e  part08
47718977ebaa8613f1647e93134479780db9320c5fa911463fd312a6257b796f  part09
1d85ece78e0779cec26b0cce29c6462d9f90180e56d78120b152df101715240b  part10
cf83a159d1a9540012229c376562587dd966da10cd657cbcef53acfd03232448  part11
1aa5415bd2db83b2689afeeb946419303b3c4f599ae790ad74d38465f768b502  part12
2de4edbce8ebdbfd0b94bc89bab518c7294209f408ec5b14a7be9e4832d9da81  part13
53f8ea43a9b5dcab97b84a52f3ac2848711c844a5420a4c56418093bacb3c6f6  part14
9631b4f020199e580ba86178cb56c6e9b09f690c2369a80dc4334aa93f49c7fc  joined

lean.tar.gz.partNN, fetched 2026-11-21; joined is their concatenation.


Updates

Updated at each milestone.