How Alpöge describes the proposed proof of the Riemann Hypothesis — in plain words
ARG Explains · Alliance Research Group · 22 November 2026
This text is dedicated to the memory of Prof. Henryk Poździk, the author’s high school mathematics teacher and class mentor during the years 1994–1997, and Prof. Kazimierz Musiał, the author’s high school physics teacher in the same period. Their clarity of exposition, intellectual rigor, and respect for precise reasoning shaped the author’s early understanding of mathematics, physics, and scientific responsibility.
1. What the Riemann Hypothesis is about
Prime numbers — 2, 3, 5, 7, 11, 13… — are the atoms of arithmetic: every other whole number is built from them. Yet they appear seemingly at random. In 1859 Bernhard Riemann noticed that their distribution is encoded in the zeros of a certain function — the zeta function — and conjectured that all the "important" zeros lie on a single line, the line ½.
Why does it matter? Because where the zeros sit decides how far the count of primes can stray from its average. In practice the Riemann Hypothesis says that the error grows at most like the square root of the range you look at (up to logarithmic factors). Remember that square root. It comes back.
2. What happened on 20 November 2026
Levent Alpöge, a mathematician working with Claude — Anthropic's AI model — published a four-page note and a formalization of a proof in Lean, a language in which a computer checks every step of the reasoning¹. The author writes that the same method also covers the generalized Riemann Hypothesis for so-called Dirichlet L-functions. According to our reading of the repository, the formalization has about 175,000 files, while the statement of the main theorem takes a single line² — which makes it easy to check what exactly has been proved without reading all the code.
Status today: the proof is announced, not confirmed. Independent researchers are compiling the code and checking the definitions². This is no paradox: once the code compiles, the computer guarantees there is no gap in the reasoning — but a human must check that the right thing was proved, for instance that it is the real zeta function and not a simplified version of it. That is what the Comparator tool is for². If compilation or the check of definitions reveals a problem, the proof will need repair — in the history of mathematics this has happened to great proofs too.
When will we know more? Independent compilations of the code are a matter of days (we are building it ourselves too, in an isolated environment). Checking the definitions and understanding the argument will take mathematicians rather months. The Clay Institute awards its prize only after publication and at least two years of acceptance by the community.
3. The idea: carry a proof over from a "toy world"
Mathematicians have long had a toy version of the Riemann Hypothesis — for curves over so-called finite fields, i.e. an arithmetic in which there are only finitely many numbers. There the hypothesis has been proved: by André Weil for curves in the 1940s, and by Pierre Deligne in 1974 for varieties of higher dimension.
In 1969 Sergei Stepanov gave a surprisingly elementary proof for a certain class of curves (hyperelliptic ones). Its heart is multiplicity: one builds an auxiliary polynomial that vanishes at many points many times over — like a function that does not just touch zero but "sticks" to it — and the needed bound follows.
The problem: in an argument about ordinary whole numbers it is hard to find a counterpart of multiplicity. In the toy world it comes from something like derivatives; over the integers it would be something like "differentiating numbers". The author writes that his proof descends from Stepanov's, but that multiplicity is exactly what is hard to find over the integers³.
4. The missing ingredient: a chain of "theta lifts"
According to the author, ideas from Fourier analysis in the toy model led the way, and then — in his work with Claude — chains of theta lifts: objects from the theory of automorphic forms, i.e. highly symmetric functions. In our simplification: each link of the chain mimics one degree of multiplicity³.
How long should the chain be? The author quotes Claude's suggestion: for a sum over primes (the so-called von Mangoldt sum) of length X — a chain of length roughly √X. And he adds: "And yet, what else could it be?"³
That line recalls section 1: the Riemann Hypothesis is about errors of square-root size, and the square root comes back as the length of the chain. But it is delight in the length, not a derivation of the exponent. The author goes on to record that the computation of the "cost" of the lifts stops a priori at the exponent 2/3, and that getting down to ½ took further tricks³. The square-root bracket is thus an analogy, not a proof — and not an obvious choice, even in the toy world.
The code shows the same architecture: bounding the error of a certain sum over primes by order x^(½+ε) — a classical argument (von Koch, Landau) — rules out zeros to the right of the line ½ (to the left, the symmetry of the zeta function rules them out). The bound itself comes from deep machinery: Galois representations, built among other things from theta lifts and so-called trace formulas (the author mentions a relative trace formula; the code contains automorphic trace formulas and the Lefschetz formula)⁴.
5. Why it is so huge — and whether it will get simpler
The author himself writes that the argument is "incredibly long at the moment", that an astounding amount of mathematics is present in it, and that — as he senses — it will be massively simplified, likely quickly³. He admits he may have baked in suboptimal choices, and says of one of the tricks that intuitively it seems unnecessary³. The text version generated with Claude runs to more than 250,000 pages; the author chose not to publish it and presented only the Lean³.
Our analysis of the code (without running it) sees that about a quarter of the files are technical lemmas and intermediate layers — though only a few percent of the lines. By the same analysis, real mathematics makes up most of the proof. Whether its deepest part — Galois representations of higher rank — could be replaced by something simpler remains an open question; reading the code text rather argues against it, but only compilation and an expert will settle it⁴.
6. What this does not change
Internet encryption rests on the difficulty of factoring huge numbers. A proof of the Riemann Hypothesis gives no fast way to factor — passwords and bank transfers are exactly as safe as before. What changes is mathematics: hundreds of theorems that begin "assuming the Riemann Hypothesis" would stop hanging on that assumption.
7. Knowing that — and knowing why
The author has announced an argument whose correctness still needs independent confirmation. Even once compilation confirms each of millions of steps, we will for now have certainty without understanding — because mathematics has always wanted more than certainty. One sentence promises understanding: the missing multiplicity can be replaced by a chain of symmetries as long as a square root. The author says his task now is to understand this argument, simplify it and write it up properly³. If that succeeds, the Riemann Hypothesis will be not only proved but told.
About the method. The text distinguishes what the author writes (the note¹), what is a checkable fact, what belongs to the known history of mathematics, what follows from our analysis of the code, and what is our simplification — a working version with every sentence labelled is kept in the ARG repository. The text was written by a human together with an AI model (Claude) and reviewed by models of other families and by ARG fleet nodes.
Notes
- ¹ L. Alpöge, note of 20 Nov 2026: alpo.ge/remark.pdf
- ² Formalization repository with the Comparator tool: github.com/leanprover-community/riemann-hypothesis
- ³ The author's note, pp. 1–4: paraphrases and one short verbatim quotation.
- ⁴ ARG static analysis (no code executed), 21 Nov 2026 — formalization map and analyses of the core and templates; results are hypotheses pending compilation.