The Alpöge–Claude proof of the Riemann hypothesis explained for the Riemann zeta function Claude Fable 5.2, Codex Astra 6.2, and Gemini Carbon at the behest of Michael O. Rubinstein Working draft, 21 November 2026 Abstract. We explain the argument of Alpöge and Claude for the Riemann hypothesis in the case of the Riemann zeta function. The analytic step shows that a one-sided bound for a smoothed sum over prime powers, valid throughout intervals [Y, Y 3/2 ] with Y arbitrarily large, excludes zeros to the right of the critical line. We derive this bound from one inequality between weighted sums over primes, the main inequality (M), and two estimates for Dirichlet characters. We prove the two estimates from theorems in the literature, and explain how the divisibility and the size of theta lifts enter (M). The proof of (M) is taken from the Lean development. This is a work in progress: Part III writes up parts of that proof and states exactly the inputs it takes from the Lean development; (M) itself is not proved here. Note. This is a working guide to Alpöge’s Lean-verified proof of the Riemann hypothesis, specialised to the Riemann zeta function. Its goal is a human-readable exposition of the proof in the style of classical analytic number theory. It is a work in progress, written as an aid to studying the proof. Written up so far: the deduction of the Riemann hypothesis from one inequality, called (M) throughout, and two estimates for Dirichlet characters (Parts I and II); proofs of those two estimates and the construction of two auxiliary sets (§III.1–III.3); the analytic part of the bound for the size of the theta kernels (§III.4; four passages there are conditional on inputs from the Lean development); a local divisibility estimate for power series at inert primes, under one hypothesis from the Lean development, proved for orbits of length two and with a further duality input for longer orbits (§III.5); the passage from that estimate to lattices: the lemmas on lattices are proved, the prime sum is split and several of its terms are bounded (the bounds for the remaining loss sums are quoted), and the theorem on the glued lattice is conditional on two statements of the Lean development (§III.6); the deduction of the size inequality from eleven named inputs, and the value of an index-corrected mean of periods under stated hypotheses (§III.7); the deduction of the dimension bound that is used from three statements of the Lean development (on traces, ranks and pairings) and stated conditions on sizes and subspaces, with a finite model of the trace estimates (§III.8); and the step from the bound for Galois cohomology to a bound for denominators, with an account of how the Lean development proves that bound (§III.9). Not proved here, all of it inside (M), and taken from the Lean development: (1) the integrality of the theta kernels in the coordinates in which the divisibility is proved, and the statement on valuations built on it; (2) the arithmetic inputs of the size bound: the inner-product formula for the theta lifts and its local factors, the covolumes of the lattices, the local index bounds and the passage between levels; (3) the trace-formula estimates themselves; (4) the bound for Galois cohomology itself; (5) the construction of the maps and their weights, and the assembly of (M), which in the Lean development is a mean over a family of maps (§III.6 treats a model case only). The table at the start of Part III indicates which statements are proved in these pages and which are still taken without proof from the Lean development. Contents (entries are links) Part I. How the proof works 4. The main inequality (M) . . . . . . . . . . 30 I.0 The strategy in one page . . . . . . . . . 3 5. From the table to a single prime sum . . . 33 I.1 From zeros to a sum over primes . . . . 4 6. Where the main inequality (M) comes from 37 I.2 Chebyshev’s method, and where it stalls 5 Sources . . . . . . . . . . . . . . . . . . . . . 53 I.3 A divisible number with smooth exponents 6 Part III. Proofs of the quoted statements I.4 Repairing the three defects . . . . . . . 7 III.1 The auxiliary primes: proof of Estimate B 55 I.5 The inequality that comes out . . . . . 13 III.2 The large scale: proof of Estimate C . 61 I.6 From the table to G(x) . . . . . . . . . 14 III.3 The sets of columns T2 and T . . . . . 64 I.7 Where (M) comes from . . . . . . . . . 16 III.4 The size of the theta kernels (partly I.8 Why the exponent is one half . . . . . 18 written) . . . . . . . . . . . . . . . . . 66 I.9 Dictionary with Alpöge’s announcement 19 III.5 Divisibility at inert primes . . . . . . . 80 I.10 The same proof for Dirichlet L-functions 21 III.6 Divisibility: from derivatives to lattices 91 I.11 Notes on the literature . . . . . . . . . 22 III.7 The size bound: the arithmetic inputs . 111 I.12 Summary: what is proved where . . . . 25 III.8 The images in general position: the Part II. Details and proofs trace-formula estimates . . . . . . . . . 129 1. What a zero forces . . . . . . . . . . . . . 26 III.9 The bound for Galois cohomology . . . 144 2. How divisibility counts primes . . . . . . . 28 Computations . . . . . . . . . . . . . . . . . . 161 3. The reduction to one prime sum . . . . . . 29 References . . . . . . . . . . . . . . . . . . . . 161 What this is. In November 2026 Levent Alpöge announced a proof, found with the AI system Claude, of the Riemann hypothesis. It exists as a formal proof in the Lean proof assistant, about 17 million lines long. This paper explains how the argument works for the Riemann zeta function, for a reader who knows a first course in analytic number theory: the prime number theorem, Dirichlet characters, and Mellin transforms. How it is organised. Part I is an exposition: the strategy first, then each step with its motivation and with examples computed on real primes. Part II gives the details and proofs to which Part I refers. A reader can stop after Part I and know how the proof goes; Part II is there to check each step. Part III, which is being written, is to prove the statements that Parts I and II quote. It proves Estimates B and C and constructs the two auxiliary sets of residue classes (the “sets of columns” of §I.5). For the main inequality (M) it proves some steps, proves others conditionally on statements of the Lean development which it names, and takes the rest from that development; the table at the start of Part III says which. What is proved here and what is not. The argument rests on three statements. One of them, the main inequality (M), is the arithmetic heart of the matter. Two further estimates (B and C) concern auxiliary Dirichlet characters. We state B and C exactly, and (M) in the form in which it is used, and explain why one should expect them. That the three together imply the Riemann hypothesis is proved in Part II. Estimates B and C are proved in Part III, from theorems on Dirichlet L-functions that are cited there. But (M) is not yet proved here: we are working through its proof in the Lean development and adding the details to Part III section by section. Of what it needs, Part III so far constructs the two sets of columns (§III.3) and proves: the analytic part of the bound for the size of the theta kernels, with four conditional passages (§III.4); a local divisibility estimate for power series at inert primes, under the integrality of the theta kernels in the coordinates of the height-two logarithm (hypothesis (B)), for orbits of length two and with a further 2 duality input for longer orbits (§III.5); the lemmas on lattices that carry this estimate to lattices and part of the sum over primes, the main theorem there being conditional on two statements of the Lean development (§III.6); that the size inequality follows from eleven named inputs (§III.7); that the dimension bound follows from three statements of the Lean development on traces, ranks and pairings (§III.8); and the step from the bound for Galois cohomology to a bound for denominators (§III.9). Part II proves the elementary lemmas. The rest of (M) is for now quoted: hypothesis (B) and the statement on valuations built on it, the arithmetic inputs of the size estimate, the trace-formula estimates themselves, the bound for Galois cohomology itself, the construction of the maps and their weights, and the assembly of (M) from these. Statements marked [quoted] are taken from the Lean development without proof. The full proof of (M) is where almost all of the formal development goes, and writing it up is the work in progress. Every numerical example, except where stated, can be reproduced with the scripts listed under Computations, after Part III. Part I. How the proof works I.0 The strategy in one page Chebyshev’s idea about counting primes. Chebyshev [Che52] bounded the number of primes using the prime factorization of factorials. In the simplest form of the argument, due to Erdős [Erd32], one considers the middle binomial coefficient 2n . It is divisible by every prime between n  n and 2n, and it is at most 4 . So the product of those primes is at most 4n . That is an upper bound n for a count of primes, obtained from two facts about one integer: what divides it, and how big it is. One piece of notation. For a real number v we write v+ for its positive part: v+ = v if v > 0, and v+ = 0 otherwise. For example 5+ = 5 and (−3)+ = 0. A sum of terms of the form (· · · )+ is automatically cut off where the terms would become negative. The plan. The proof of the Riemann hypothesis described here is Chebyshev’s argument, run with a far better “integer”. There are five steps. 1. Zeros to primes. The Riemann hypothesis is equivalent to a bound |G(x)| ≤ Cx1/2+ε for a certain smoothed count of primes G(x), for every ε > 0, with C = C(ε). In fact an upper bound G(x) ≤ x1/2+ε is enough, and it is enough to have it on a sparse sequence of long intervals. (§I.1.) 2. Why Chebyshev’s method stalls. For n! the divisibility is known exactly: the exponent of p is h ⌊n/ph ⌋. The floor functions are the obstacle; they are why the classical method gives P constants but not asymptotics. (§I.2.) 3. A divisible object with no floor functions. Derivatives of functions on an elliptic curve with complex multiplication are divisible by primes to exponents of the smooth form a/p + a/p3 + · · · , but only at the primes that stay prime in the associated imaginary quadratic field. Subtracting a large weight w turns each term a/ph into (a/ph − w)+ = w (x/ph − 1)+ with x = a/w: exactly the smooth kernel that G uses, magnified by w. (§I.3 and §I.4.) 4. Size, to the precision needed. “Divisibility ≤ size” now gives an upper bound for a sum over primes. For it to be useful the size must be known not just to leading order a log a but through the term linear in a. This forces one to replace a single number by a linear map between lattices, built from theta series, whose size is governed by classical formulas 3 (Chowla–Selberg, the mass formula). The resulting comparison of divisibility with size is the main inequality (M). (§I.5 and §I.7.) 5. Clean up and balance. The sum in (M) runs over primes in certain residue classes. Averaging over the classes leaves G(x) plus three sums twisted by quadratic characters, which are small for well-chosen auxiliary primes. Repeating the construction m times leaves two competing √ errors, x/m and m. Taking m ≈ x gives G(x) ≤ x1/2+ε . (§I.6 and §I.8.) Logical skeleton. §I.6 §I.6 §I.1 (M) , B, C −−−→ Estimate A −−−→ G(x) ≤ x1/2+ε on long intervals −−−→ RH. (I.1) |{z} | {z } [quoted] Part III Three statements enter: the main inequality (M), which is quoted, and two estimates for auxiliary Dirichlet characters, one at the scale x (Estimate B) and one at a much larger scale (Estimate C), which are proved in Part III from cited theorems. The order of choices is this. For each small ε there are arbitrarily large Y for which a single pair of auxiliary primes q, d serves for every x in [Y, Y 3/2 ]; the rest of the construction is then made for each x separately. I.1 From zeros to a sum over primes The statistic. Let Λ(n) = log p if n = ph is a prime power and 0 otherwise. Define x  S(x) = Λ(n) −1 , G(x) = S(x) − x log x + (1 + γ)x, (I.2) X n≤x n where γ is Euler’s constant. Equivalently S(x) = n Λ(n)(x/n − 1)+ , where v+ = max(v, 0): with P the positive part, the condition n ≤ x under the sum can be omitted. Compared with Chebyshev’s ψ(x) = n≤x Λ(n), the sum S weights n by the ramp (x/n − 1)+ : as a function of n it is large for P small n, decreases, reaches 0 at n = x, and is 0 for n > x. Also S(x) = x 1x ψ(t) dt/t2 . The two R subtracted terms are the main term of S, so G is a remainder. R∞ Why this statistic. The Mellin transform of the ramp is simple: 1 (x/n − 1)+ x−s−1 dx = n−s /(s(s − 1)) for Re s > 1. Summing over n with weights Λ(n), ζ ′ (s) 1 1 1+γ Z ∞ F (s) = G(x)x−s−1 dx = − · − + . (I.3) 1 ζ(s) s(s − 1) (s − 1) 2 s−1 The last two terms cancel the pole at s = 1 (both terms of its principal part). What remains are simple poles of F at the zeros ρ of ζ, with residue −1/(ρ(ρ − 1)) at a simple zero, and a pole at s = 0. (The trivial zeros −2, −4, . . . also give poles; they produce the term O(x−2 ) below.) Shifting the contour in the inverse transform gives the explicit formula (compare [Dav00, Ch. 17]) xρ G(x) = log 2π − + O(x−2 ), (I.4) X ρ ρ(ρ − 1) the sum being over the nontrivial zeros. It converges absolutely, because of the factor ρ(ρ − 1); that is the benefit of smoothing. One has ρ ρ(1−ρ)1 = 2 + γ − log 4π = 0.0461 . . . . If every zero P √ has real part 2 then |ρ(ρ − 1)| = ρ(1 − ρ), and therefore, for large x, |G(x) − log 2π| ≤ 0.0462 x. 1 A zero with real part β > 12 would contribute a term of size xβ . 4 A sanity check on real primes. Summing over the prime powers up to 2·106 gives G(104 ) = 1.749, G(105 )√= 3.671, G(106 ) = 11.040, G(2 · 106 ) = −12.268, to be compared with log 2π = 1.838 and 0.0462 106 = 46.2. One side is enough. The terms of a zero off the line and of its complex conjugate together form a real function of x that oscillates and takes both signs. So one expects a zero with real part β to make G(x) exceed +cxβ for some arbitrarily large x. Thus an upper bound for G should already exclude it. That an upper bound suffices follows from a theorem of Landau on Mellin transforms of functions of one sign [Lan05]; see [MV07, §15.1]. Lemma (one-sided criterion). Let 0 < θ < 1. If G(x) ≤ xθ for all large x, then ζ(s) ̸= 0 for Re s > θ. Proof. The integral H(s) = 1∞ xθ − G(x) x−s−1 dx equals 1/(s − θ) − F (s), and its integrand is R  nonnegative for large x. Landau’s theorem says that a Mellin transform of an eventually nonnegative function has a singularity at the real point of its abscissa of convergence. But by (I.3) the function H is regular at every real s > θ: the only real singularities of F in s > 0 would come from real zeros of ζ in (0, 1), and there are none. Hence the abscissa is at most θ, the integral converges for Re s > θ, and F is holomorphic there. A zero of ζ in that half-plane would be a pole of F . The arithmetic argument below produces the upper bound only for x in certain long intervals. The lemma survives this weakening. Lemma (the same, on intervals). Let 0 < θ < 1. If there are arbitrarily large Y such that G(x) ≤ xθ for all x ∈ [Y, Y 3/2 ], then ζ(s) ̸= 0 for Re s > θ. The proof replaces Landau’s theorem by an estimate of high derivatives of F , which see only one interval at a time; it is one page and is given in full in Part II, §1. Conclusion of step 1. To prove the Riemann hypothesis it suffices to show: for every ε > 0 there are arbitrarily large Y with G(x) ≤ x1/2+ε throughout [Y, Y 3/2 ]. (Zeros come in pairs ρ, 1 − ρ, so no zeros to the right of the line means none to the left.) From now on the goal is an upper bound for a sum over primes. I.2 Chebyshev’s method, and where it stalls The method. Let N = 2n n . Every prime p with n < p ≤ 2n divides (2n)! and not (n!) , so it 2  divides N . Also N ≤ (1 + 1)2n = 4n . A positive integer divisible by distinct primes is at least their product, so log p ≤ log N ≤ 2n log 2. (I.5) X n