Just in: An 11-page map of the claimed Riemann Hypothesis proof

Table of contents
  1. A Readable Map of the Proof
  2. From Stepanov to the Integer Setting
  3. The Analytic Framework
  4. Defining G and the Explicit Formula
  5. Landau’s Theorem and the Role of Upper Bounds
  6. Localized Bounds on Windows
  7. The Main Estimate
  8. The Arithmetic Estimate
  9. The Factorial Analogy
  10. Theta Functions and Divisibility
  11. The Product Formula and the Main Term
  12. Multiplicity Through Theta Lifts
  13. Auxiliary Primes, Iwasawa Theory, and Galois Representations
  14. Verification and Interpretation
  15. Formal Validity of the Lean Proof
  16. Accuracy of the Guide
  17. The Classical Analysis
  18. The Arithmetic Account and Its Future Review

Just in: An 11-page map of the claimed Riemann Hypothesis proof

Until yesterday evening, the proof of the Riemann hypothesis announced by Levent Alpöge existed in two forms, a four-page note and seventeen million lines of Lean, and neither of them let a mathematician see how the argument goes. A third document has now appeared. It is an eleven-page guide to the architecture of the proof, published in a repository that belongs to the mathematician Bartosz Naskręcki, and it was produced with the help of an AI system by reading the Lean code and the note.

It is not the author’s own account, and its later sections report what the code says without deriving it again. With that caution in mind, it is the first readable map of the route.

The note describes the argument as a descendant of a proof by Stepanov. For curves over a finite field, the analogue of the Riemann hypothesis was proved by Weil in the nineteen forties, and in nineteen sixty-nine Stepanov found an elementary proof. He constructs an auxiliary function that vanishes to high order at the points of the curve, and since a function cannot have more zeros than its degree allows, this bounds the number of points from above. Upper bounds over all the finite extensions of the field then turn out to be enough.

The announced proof looks for a counterpart over the integers to each of these four ingredients, and the hardest one to find is the multiplicity.

The first step is a reduction. Because of the functional equation, a zero to the left of the critical line would produce one to the right of it, so it is enough to show that there are no zeros with real part greater than one half.

The proof then works with a single function of a real variable y, which I will call G. By definition, G of y adds up the logarithms of the primes over all the prime powers n up to y, each one weighted by y over n, minus one, and then subtracts the expected main term. The weight is a ramp that falls to zero at n equals y, and the reason for choosing a ramp over a sharp cut-off is that the smoothing makes the sum over the zeros converge absolutely.

That sum is the explicit formula, which is a property of G and not part of its definition. It says that G of y equals the logarithm of two pi, minus a sum over the zeros rho of y to the power rho, divided by rho times rho minus one. If every zero has real part one half, each term has the size of the square root of y, and the whole sum is at most zero point zero four six times the square root of y. A zero further to the right, with real part beta, would contribute a term of size y to the power beta. When the function is computed up to ten million, it stays well inside that band.

The next step explains why the proof only ever needs to bound G from above. A theorem of Landau says that when a function keeps a constant sign, its Mellin transform has a singularity at the real point on the edge of its region of convergence. Applied to a constant times y to the power theta, minus G, it shows that a bound of that kind for all large y would exclude every zero with real part greater than theta. The heuristic is that all the zeros appear in the transform of G as poles with the same sign, so they cannot cancel one another, and a single zero cannot hide.

The arithmetic in this proof delivers the bound only on certain intervals, which the guide calls windows, running from Y to Y to the power three halves. The theorem is therefore proved in a localized form, in which a zero to the right of theta forces G above y to the power theta somewhere inside every sufficiently late window.

Everything that remains is contained in one statement, which says that for arbitrarily large Y the function is at most the square root of x, times a small power of x, at every point x of the window.

The idea behind that estimate can be seen in a classical example. The factorial of a is divisible by each prime p to a power close to a over p, plus a over p squared, and so on, And taking logarithms turns this into a sum over prime powers with a weight very much like the ramp. The other side of the equation is the logarithm of a factorial, which Stirling’s formula computes.

The simplest case in the guide replaces the factorial with the derivatives at the origin of a theta function on an elliptic curve with complex multiplication, divided by powers of its period. According to the note, these numbers are algebraic integers, and they are divisible by high powers of the primes that stay inert in the quadratic field.

A non-zero algebraic integer that is divisible by many primes cannot be small, by the product formula, and so the divisibility, which is a prime sum with a ramp, is bounded above by the logarithm of the size of the number, which can be computed. The computation produces a main term that begins with a times the logarithm of a minus one minus Euler’s constant, and this is the same centring that appears in the definition of G.

A single derivative gives error terms that are far too large, and this is where multiplicity comes in. The derivative is replaced by a chain of m theta lifts, which are linear maps between spaces of automorphic forms on unitary groups of growing dimension, defined over a field that climbs a cyclotomic tower as x grows.

The product formula is replaced by an inequality of Bost on the slopes of lattices, which says that whatever the chain gains in divisibility at the primes, it has to pay for in its norm at the infinite places. According to the guide, the errors come out as x over m, plus m, plus the square root of x. The first term is the error of the whole chain shared among its lifts, and the second is a cost that each lift incurs as the chain grows longer, so the best choice is m of the order of the square root of x. That choice is the step that the note attributes to Claude.

Three further pieces complete the route, and I will only name them here. The bound arrives for a combination of all the characters of the Galois group of the field, so an average with positive weights is used to keep the characters that are wanted, and two auxiliary primes are chosen to remove the others.

The divisibility holds only up to a defect measured by Galois cohomology, which is controlled through Iwasawa theory and a theorem of Ferrero and Washington. That control is available only at certain scales, which is the reason the analytic theorem had to work on windows. Both of these rest on the construction of Galois representations attached to automorphic forms over these fields, which the archive proves formally for itself.

Two different questions should be kept apart, and the first is whether the formal proof is valid. The complete Lean proof has now been accepted by two independently written kernels, and its five final theorems, the Riemann hypothesis among them, match the statements in Mathlib. The guide also reports a separate machine check of the last step, which leads from the final estimate to the Riemann hypothesis.

The second question is whether the guide’s account of how the proof works is accurate, and the answer depends on the part. The classical part, which covers the function G, the explicit formula and Landau’s theorem, is textbook analysis, and its formulas and the computation up to ten million have been rechecked independently of the guide.

The arithmetic part, with the theta lifts, the slopes and the Iwasawa theory, was read from the Lean code with the help of an AI system, and no specialist has refereed that reading yet. Alpöge has said that he intends to write the argument up properly, and that account, when it comes, is the one this map should be tested against.

Questions and answers
What does the newly published eleven-page guide contribute, and what limitation should readers keep in mind?
It offers the first readable map of the proof’s route, based on the Lean code and the short note. However, it is not Alpöge’s own account, and some later sections report what the code says rather than independently deriving those arguments.
Why is the function G smoothed with a ramp instead of using a sharp cut-off?
The ramp makes the sum over the zeros in the explicit formula converge absolutely, which is why the proof uses this smoothed function.
Why does the argument need to bound G only from above to rule out zeros to the right of the critical line?
Landau’s theorem implies that an upper bound of the relevant form for all sufficiently large y would exclude zeros with real part greater than the bound’s exponent. The heuristic is that the zeros contribute poles with the same sign in the transform, preventing one zero from being hidden by cancellation.
Why does the proof establish its arithmetic bound on windows rather than at every sufficiently large point?
The divisibility control supplied by Iwasawa theory is available only at certain scales. Consequently, the analytic theorem is formulated locally on windows, from Y to Y to the power three halves, rather than as a bound at every large point.
Why does the proof replace a single theta-function derivative with a chain of lifts, and how is the chain length chosen?
A single derivative gives errors that are too large, so the proof uses a chain of m theta lifts to obtain multiplicity. The reported errors are x/m, m, and the square root of x; balancing the first two terms gives m of order the square root of x.
Does the formal acceptance of the Lean proof also establish that the guide accurately explains the proof’s arithmetic ideas?
No. The formal proof’s validity and the accuracy of the guide’s explanation are separate questions. The arithmetic account was reconstructed from the Lean code with AI assistance and has not yet been refereed by a specialist.
Is the computation of G up to ten million enough to prove the Riemann hypothesis, given that it stays within the predicted band?
No. The computation is supporting evidence, but the proof requires an argument that rules out zeros beyond the critical line; the text presents the computation separately from the analytic and arithmetic bounds that are meant to establish that result.
How exactly do the two auxiliary primes remove the unwanted characters, and why are those particular primes chosen?
The text says only that two auxiliary primes are chosen to remove the unwanted characters; it does not explain how the procedure works or how the primes are selected. Those details remain unclear from this account.