An illustrated atlas · Lean blueprint

How the announced proof of the Riemann hypothesis is put together

Levent Alpöge's four-page announcement (20 November 2026) claims proofs of RH and of GRH for Dirichlet characters, formalized in a Lean archive of 175,187 modules. This atlas maps the argument component by component, links every step to pinned Lean source, and marks exactly what has been machine-checked.

✓✓One-sided prime windows ⇒ RH: checked by two kernels✓TopG ∧ VWinAt ⇒ RH ∧ Dirichlet GRH: kernel-checked✓✓Five endpoints, complete proof (175,187 modules): accepted by two independent kernels

Status recorded by the reading project at 2026-11-20 23:51 UTC: Two independent kernels accepted the complete proof, and all five theorems matched the canonical Mathlib challenge.

Background: the centered prime sum $(G(y)-\log 2\pi)/\sqrt y$, computed from the primes up to $10^7$ (see “About the images”).
01 · The gist

The proof in three sentences

The argument has one hard estimate and two well-understood transformations of it. Read the three cards from the arithmetic to the destination; every term in them is a component of the map below.

Arithmetic supplies a one-sided bound

For every primitive character $\psi$ and every small $\varepsilon$, build families on which a positive combination of centered prime sums satisfies $G_P(x)\le x^{1/2+9\varepsilon}$ at every point of arbitrarily late windows $[Y,Y^{3/2}]$. The tools: chains of about $\sqrt x$ theta lifts, Galois representations attached to automorphic forms, Iwasawa-type bounds for Galois cohomology.

Analysis turns the bound into a zero-free half-plane

The Mellin transform turns a zero $\rho$ into a pole with residue $-S(\rho)/(\rho(\rho-1))$, $S(\rho)>0$. High derivatives aimed at the windows would see that pole. Choosing $\theta$ after the zero and $\varepsilon$ after $\theta$, no zero survives to the right of $\tfrac12$.

Reflection delivers the critical line

Euler factors pass from primitive to all characters, and the functional equations reflect the half-plane onto the line: Mathlib's RiemannHypothesis and Dirichlet GRH for every positive modulus.

Where the difficulty lives. The analytic engine and the closing deductions are machine-checked with the window estimate as an explicit hypothesis (two verified cuts). The estimate itself comes from the arithmetic supply. Its terms belong to the complete proof, which two independent kernels accepted and several independent builders replayed (see Verification); in this atlas they are read at their pinned source lines, and only some of them were also checked one by one.
02 · The map

The proof in one picture

Fourteen components in four strata. Information flows down the arithmetic column, across the estimate, down the analytic engine and up to the endpoints. Arrows carry the Lean propositions through which the components talk. Click a card for its statement, idea, Lean declarations and status; toggle a verified cut to see which part has been machine-checked with the rest as a hypothesis.

Verified cuts: ✓✓Checked by two kernels✓Kernel-checked at the archive pin◐Partly checked⊂Inside a checked closure◇Source-read↗Full replay reported○Motivation

Arithmetic supply

Galois representations, theta lifts and Iwasawa-type cohomology bounds that are meant to produce the families on which a prime sum is small.

The estimate

A finite character comparison turns geometric control of a chain of theta lifts of length about √x into an upper bound for a weighted prime sum.

The analytic engine

A localized Landau oscillation argument converts one-sided bounds on whole power windows into a zero-free half-plane.

The destination

Euler factors, conductors and functional equations turn the half-plane theorem into the five exported statements, including Mathlib's RiemannHypothesis.

1

The five endpoints

Main.lean
Ordinary RH in Mathlib's own formulation, and Dirichlet GRH for every character of every positive modulus.
✓✓Checked by two kernels
2

Reflection

functional equations
A zero left of the critical line reflects to a zero right of it, so the right half-plane is all one has to clear.
✓Kernel-checked at the archive pin
3

Conductor descent

Euler factors
An imprimitive character differs from its primitive inducing character by finitely many Euler factors, none of which vanishes for $\operatorname{Re}s>0$.
✓Kernel-checked at the archive pin
4

The centered prime sum

what must be bounded
A ramp-weighted von Mangoldt sum, centered so that the main terms cancel exactly, and its positive combinations over a cyclic family of characters.
✓Kernel-checked at the archive pin
5

Mellin transform and positive residues

poleData
The Mellin transform of a ramp is $n^{-s}/(s(s-1))$; a zero of any family member becomes a pole whose residue cannot be cancelled, because the weights are positive.
✓Kernel-checked at the archive pin
6

The window criterion

localized Landau
If a positive family's prime sum stays below $Cy^\theta$ throughout arbitrarily late windows $[Y,Y^{1+\delta}]$, none of its $L$-functions vanishes to the right of $\theta$.
✓✓Checked by two kernels
7

The order of choices

θ after the zero, ε after θ
The loss exponents are fixed first; given a zero with $\beta>\tfrac12$ choose $\theta\in(\tfrac12,\beta)$, then $\varepsilon$ with $\tfrac12+9\varepsilon\le\theta$. No bound at $\varepsilon=0$ is ever needed.
✓Kernel-checked at the archive pin
8

The finite character comparison

core_gm · ch_core · ch_top
Finite Fourier orthogonality on a subgroup and an exact main-term identity convert a geometric discrepancy bound into the prime-sum bound, with exponent $\tfrac12+(e_D+e_L+7)\varepsilon$.
✓Kernel-checked at the archive pin
9

The geometric family and the theta chain

topDesWin · RankVal
A chain of about $\sqrt x$ theta lifts, with divisibility, slope and covolume control, is chosen so that both loss exponents are fixed numbers: $e_D=2$, $e_L=0$.
◐Partly checked
10

Hosts for every character

supplyChi
For each primitive character $\psi$, the cohomological window produces arithmetic hosts of width $3/2$ at all large scales of a system of scales.
✓Kernel-checked at the archive pin
11

The cohomological window

VWinAt ψ 401
On a whole interval $[Y,Y^2]$, weighted lengths of locally trivial Galois cohomology of twisted modules are at most $\sqrt{\exp(x^{\varepsilon})}$, uniformly over the relevant fields.
◐Partly checked
12

Auxiliary primes and Behrend towers

QH.pairEx · hostThickExists
Two auxiliary primes with prescribed congruences and square-root prime-sum bounds for their quadratic twists, and an additive-combinatorial (Behrend) choice of the tower.
✓Kernel-checked at the archive pin
13

The Galois row

Row0b
For regular algebraic polarized cuspidal $\pi$ on $\mathrm{GL}_a$ over a CM field, a continuous semisimple $\ell$-adic representation whose Frobenius polynomials match the Satake parameters of $\pi$.
◇Source-read
14

Seven inputs at the decisive call

every_character_of_literature_of_vWin
Unitary-group and local theta packages, agreement of theta integrals, local matching, a covolume bound, local Langlands parameters and the window: the seven arguments of the arithmetic theorem.
◇Source-read
03 · The blueprint

107 declarations, 113 dependencies, all pinned

The reading project's declaration map (107 Lean declarations and 2 ideas quoted from the announcement), laid out by dependency depth: suppliers on the left, the five exported theorems on the right. Bar colour = stratum, dot = verification tier. Click a declaration for its statement in standard mathematics, its Lean head and source excerpt, and links to the pinned lines; drag to pan, scroll to zoom. The three tours follow the reading routes of the reading project.

Tours:
← dependencies run to the left · drag · scroll · click a box · minimap ↘
04 · Illustrated ideas

Eleven pictures of the argument

Amber tags mark figures computed from definitions (every number asserted by the build); grey tags mark schematic diagrams; green marks facts also proved in this repository's Lean library.

Prime sum versus the sum over zerosComputed: normalized G(y) from primes (amber) and the explicit formula with the first 100 zeta zeros (indigo), 100 <= y <= 2e6; dashed: the term a hypothetical zero at 0.62+40i would add (it grows like y^0.12).-0.04-0.020+0.02+0.04102103104105106yfrom primes: (G(y) − log 2π)/√yfrom the first 100 zeros (explicit formula)one hypothetical zero at 0.62 + 40i would add this
computed

The quantity the proof must control

The archive's statistic is $G(y)=\sum_{n\le y}\Lambda(n)(y/n-1)-y\log y+(1+\gamma)y$. By the explicit formula, $G(y)-\log 2\pi=-\sum_\rho y^\rho/(\rho(\rho-1))$ up to $O(y^{-2})$. The amber curve is computed from the primes, the indigo one from the first 100 zeros (computed here by Euler–Maclaurin summation); on $[10^2, 2\cdot10^6]$ they differ by at most $0.00064\sqrt y$. A zero at $\beta+i\gamma$ off the line would add oscillations of size $y^{\beta-1/2}$ (dashed: a hypothetical zero at $0.62+40i$). The proof needs only an upper bound for $G$, and only on windows $[Y,Y^{3/2}]$.

Critical strip, reflection and the half-plane theoremSchematic plane with the computed first zeta zeros on the critical line; a hypothetical off-line zero rho and its mirror image 1 - conj(rho) under the functional equation.0½1Re s10i20i30i40i50ireflection: L(s, χ) = 0 ⇒ L(1 − s, χ−1) = 0ρ ?1 − ρRe s ≥ 1: classicalnonvanishing (Mathlib)Re s > ½:half-planetheorem,every χfirst 10 zeros, computedon Re s = ½hypothetical zero
schematiccomputed

Only the right half-plane matters

If $L(\rho,\chi)=0$ in the strip, the functional equation gives $L(1-\rho,\chi^{-1})=0$; with complex conjugation this is the mirror image of $\rho$ in the critical line. So it suffices to clear $\operatorname{Re}s>\tfrac12$ for every character. The dots are the first 10 zeta zeros, computed.

Ramps and the centered sumComputed. Left: the weighted ramps Lambda(n)(y/n-1)_+ for prime powers n <= 13, each switching on at y = n with slope Lambda(n)/n. Right: their full sum minus y log y - (1+gamma) y, i.e. G(y), for 1 <= y <= 120; it hovers around log 2pi (dashed), the residue at s = 0.weighted ramps Λ(n)(y/n − 1)₊012342468101214y23457891113Σ Λ(n)(y/n − 1) − y log y + (1+γ)y = G(y)-20246020406080100120ylog 2π
computed

Ramps make the transform exact

Each prime power switches on a ramp $\Lambda(n)(y/n-1)_+$ at $y=n$. The Mellin transform of one ramp is $n^{-s}/(s(s-1))$ (checked numerically at 9 points, error at most 3.0e-10). Summed, the ramps produce $-L'/L$; the centring terms remove the pole at $s=1$, and what remains hovers around $\log 2\pi$, the residue at $s=0$.

Positive weights keep the residue aliveSchematic, exact in direction: at a common zero rho every member of a positive family contributes a positive multiple of the same complex number -1/(rho(rho-1)), so the contributions add; with signed weights two contributions could cancel.positive family: contributions add½ a₁ (m(χ₁) + m(χ₁⁻¹))½ a₂ (m(χ₂) + m(χ₂⁻¹))½ a₃ (m(χ₃) + m(χ₃⁻¹))total: S(ρ) · (−1/(ρ(ρ − 1)))S(ρ) = ½ Σᵢ aᵢ (ord_ρ L(·, χᵢ) + ord_ρ L(·, χᵢ⁻¹)) > 0 whenever some member vanishes at ρsigned weights: could cancel+a−athe source requires aᵢ > 0 and aprincipal member only of modulus 1
schematic

Zeros cannot hide in a positive family

At a common zero $\rho$ each member of a positive family contributes a positive multiple of $-1/(\rho(\rho-1))$ to the residue, so the contributions point the same way and add up to $S(\rho)>0$ times that number. With signed weights two zeros could cancel; the archive requires every weight $a_i>0$.

The spotlight of high derivativesComputed: kernels u^k e^{-u}/k! (normalized to peak 1) with k = ceil(1.25 U) for U = 20, 40, 80, 160 and the windows [U, 1.5 U] in u = log y; labels give the fraction of each kernel's mass inside its window.04080120160200240u = log yk = 2568.0% insidek = 5084.0% insidek = 10095.2% insidek = 20099.5% inside
computed

High derivatives are a spotlight

The $k$-th derivative of the Laplace transform weights $g(u)=G_F(e^u)$ by $u^k e^{-\sigma u}/k!$, a bump at $u=k/\sigma$ whose relative width shrinks like $k^{-1/2}$. Aimed at the window $[U,1.5U]$ with $k=\lceil 1.25\,\sigma U\rceil$ it puts 68.0%, 84.0%, 95.2%, 99.5% of its mass inside for $U=20,40,80,160$. A pole at distance $d$ contributes about $d^{-k}$; the window bound allows only $e^{\eta U}r^{-k}$ with $r>d$, which is impossible once $k\ge U$ and $\eta$ is small.

The order of choices (schematic)Schematic: the loss exponents are fixed first; a candidate zero beta > 1/2 is given; theta is chosen between 1/2 and beta; epsilon is chosen so that 1/2 + (e_D+e_L+7) eps <= theta; the arithmetic then supplies windows for that epsilon; the analytic criterion excludes the zero.fix the lossese_D = 2, e_L = 01suppose a zerowith Re ρ = β > ½2choose θ½ < θ < min(β, 1)3choose ε½ + 9ε ≤ θ4arithmetic supplieswindows [Y, Y^(3/2)]5the criterionexcludes Re s > θ6Why a bound at ε = 0 is never needed: θ is chosen after the zero, and ε after θLean: Formalism.HC.level_of_tiedWin takes θ = ½ + ½·min(β − ½, ½); the window width is 3/2.
schematic

The order of choices

The losses $e_D=2$, $e_L=0$ are fixed first. Given a zero with $\operatorname{Re}\rho=\beta>\tfrac12$, take $\theta=\tfrac12+\tfrac12\min(\beta-\tfrac12,\tfrac12)$ and then $\varepsilon$ with $\tfrac12+9\varepsilon\le\theta$; the arithmetic supplies windows for that $\varepsilon$, and the criterion excludes the zero. A bound at $\varepsilon=0$ is never needed.

Where the square root and each epsilon are paidLeft: exponents 1-a of x/m and a of m when m = x^a; the larger is smallest at a = 1/2. Right: exact staircase of the losses in units of epsilon, from the finite comparison to 1/2 + (e_D+e_L+7) eps = 1/2 + 9 eps for the source's choice e_D = 2, e_L = 0.00½½11a (m = xᵃ)x/mmchain length m ≈ √xlosses above ½, in units of ε+13/8balance x/m vs m (worst budget term)+5geometric error E ≤ x^(C*·ε)·B with C* = e_D + e_L + 3+1density of the averaging set: 2/δ ≤ 2·x^ε+3/8absorbing the constant: 26 ≤ x^(3ε/8)+1restoring discarded characters: 1 + 4o ≤ x^εtotal 9 = e_D + e_L + 7
exact

Where the square root, and every ε, is paid

The geometric discrepancy has a cost $x/m$ and a cost $m$; with $m=x^a$ their larger exponent $\max(1-a,a)$ is smallest at $a=\tfrac12$ — the announcement's "chain of length $\sqrt X$". The staircase adds the exact losses of the finite comparison: $\tfrac{13}{8}+(e_D+e_L+3)+1+\tfrac38+1=e_D+e_L+7$, which is $9$ for the archive's $e_D=2$, $e_L=0$.

One interval survives all exceptional twistsComputed with the induction of UpperTheory.V.Win.window1 on the book's example: forbidden logarithmic intervals [a, B a) for a = 8, 90 with B = 2 inside [0, 128]; the recursion lowers the ceiling 128 -> 32 and returns [16, 32], i.e. [Y, Y^2] = [2^16, 2^32].u = log₂ x (forbidden: [a, 2a) for each exceptional twist h, a = log₂(h)/C_t)020406080100120140160180200[8, 16)[90, 180)[16, 32]→ [Y, Y²] = [2¹⁶, 2³²]ceiling U = 128: start 90 lies in (32, 128] → delete itceiling U′ = 128/(2B) = 32: no start in (8, 32] → y = U′/2 = 16Ultrametric separation (ℓ = 3, Z = 80, centre 1): |h − 1|₃ < 3⁻³ forces h ≡ 1 mod 81, so h = 1 only;with ≤ instead of < one would also admit 28 and 55. Each ℓ-adic centre produces at most one exceptional twist.
computedformal

One interval survives every exceptional twist

An $\ell$-adic determinant estimate leaves at most $R$ centres, and each centre accounts for at most one exceptional twist (ultrametric separation, proved in this repository's Lean library). Each exceptional twist forbids a logarithmic interval $[a,Ba)$, and the induction of window1 finds $[y,2y]$ avoiding all of them. The reading project's example — twists $2^{16}$ and $2^{180}$, $C_t=2$, $\eta=1$, $X=2^{128}$ — gives $[Y,Y^2]=[2^{16},2^{32}]$.

A Behrend digit sphereComputed: the vectors of {0,1,2}^3 with squared norm 5 are the six permutations of (0,1,2); they lie on the circle where the sphere meets the plane x+y+z = 3. Their base-24 encodings avoid every relation j(s2-s1) = i(s3-s1) with 1 <= |i|,|j| <= 2, i != j: 1,440 checks.(0,1,2) ↦ 26(0,2,1) ↦ 49(1,0,2) ↦ 578(1,2,0) ↦ 624(2,0,1) ↦ 1153(2,1,0) ↦ 1176sphere x² + y² + z² = 5 ∩ plane x + y + z = 3Encode (x, y, z) ↦ 24²x + 24y + z (base 4Ld = 24).The base is large enough that no carries occur, soj(s₂ − s₁) = i(s₃ − s₁) between encodings forces thesame relation between the vectors: three points onone line. A line meets a sphere at most twice.1,440 relation checks, none holds(120 ordered triples × 12 pairs (i, j), 1 ≤ |i|, |j| ≤ 2, i ≠ j)Lean: BehrendReading.sphere_rel (this repository)Archive: Formalism.FrobeniusOrbits.ThickExists.noRel_image
computedformal

No three in a line: Behrend spheres

Integer vectors of fixed norm, written in a base large enough to avoid carries, satisfy no short affine relation, because a line meets a sphere at most twice. The six permutations of (0,1,2) in base 24 pass all 1,440 relation checks; the general statement is BehrendReading.sphere_rel, compiled in this repository.

The chain of theta lifts (schematic)Schematic: unitary groups U(X0) -> U(X1) -> ... -> U(Xm) linked by theta lifts; each lift carries a denominator D_j; the archive's RankVal puts m between sqrt(x delta1) and 2 sqrt(x delta1), delta1 = x^(-eps/8).U(X0)U(X1)U(X2)U(Xm−1)U(Xm)θ, ×D0θ, ×D1θ, ×Dm−1· · ·π ↦ π ⊞ μⱼ[t] at each step: each lift is meant to mimic one order of multiplicity√(xδ1) ≤ m ≤ 2√(xδ1), δ1 = x−ε/8 ⇒ m ≈ x1/2 − ε/16 GlobalTheory.EX.RankVal(∏ Dj) · (Tm−1 ⋯ T0 f)(v) ∈ S when every Djwj, every kernel entry and f are S-integral Formalism.ThetaLiftIndependence.ChainInt
schematic

A chain of about √x theta lifts

Each theta lift $U(X_j)\to U(X_{j+1})$ is meant to mimic one order of multiplicity in Stepanov's method. The archive's RankVal fixes the length between $\sqrt{x\delta_1}$ and $2\sqrt{x\delta_1}$, $\delta_1=x^{-\varepsilon/8}$; denominators clear by $\prod_j D_j$ (ChainInt).

The Galois row by quadratic descent (schematic)Schematic: representations r_K over infinitely many quadratic CM extensions K = E(sqrt(-p)) are patched to one rho over E whose Frobenius polynomials match the Satake parameters of pi.E (CM field), π on GL_aone ρ : G_E → GL_a(Q̄_ℓ)K1 = E(√−p1)rK1 from the good caseK2 = E(√−p2)rK2 from the good caseK3 = E(√−p3)rK3 from the good caseK4 = E(√−p4)rK4 from the good case…pi ≡ −1 mod M outside a finite bad set: √−pi ∉ E, since inertia at pi moves itindex-two patching of conjugation-invariant trace identities ⇒ one ρ over Eι(charpoly ρ(σ−1)) = ∏α∈B_v(X − q(a−1)/2α) for all but finitely many places v
schematic

A Galois representation by quadratic descent

For a regular algebraic polarized cuspidal $\pi$ on $\mathrm{GL}_a$ over a CM field $E$, representations over infinitely many $K=E(\sqrt{-p})$ are patched, through index-two trace identities, into one $\rho$ over $E$ whose Frobenius polynomials match the Satake parameters of $\pi$ at all but finitely many places.

05 · Verification

What has been checked, and by whom

Status recorded by the reading project at 2026-11-20 23:51 UTC. A verified cut is an implication; its premise is still a hypothesis. A source-read component has been located and read at pinned lines, not replayed.

✓✓Checked by two kernels

One-sided prime windows imply RH

Reading.WindowBridge.riemannHypothesis_of_prime_windows
$$\Bigl[\,\forall\theta\in(\tfrac12,1)\ \exists C\ \forall Y_0\ \exists Y\ge Y_0\ \forall y\in[Y,Y^{3/2}]:\ G(y)\le C\,y^{\theta}\,\Bigr]$$

⟹ Mathlib's RiemannHypothesis

Fresh Lean replay over 10,530 modules (766,938 constants) and the independent Rust kernel NanoDA on 61,697 exported declarations; 31 original modules; axioms propext, Classical.choice, Quot.sound.

Source: references/archive-bridges/WindowBridge.lean. Components inside: Reflection, The centered prime sum, Mellin transform and positive residues, The window criterion.

✓Kernel-checked at the archive pin

TopG and the cohomological windows imply RH and Dirichlet GRH

Reading.DirichletBridge.endpoints_of_topG_vWin401
$$\Bigl[\,\mathrm{TopG}\ \wedge\ \forall\psi\ \text{primitive}:\ \mathrm{VWinAt}\ \psi\ 401\,\Bigr]$$

⟹ RH ∧ half-plane nonvanishing ∧ critical-strip zeros, for every positive modulus

1,421 original modules at Lean 4.33.1; seven type/axiom reports; six defining modules replayed (Helios job 26647685).

Source: references/archive-bridges/DirichletBridge.lean. Components inside: Reflection, Conductor descent, The centered prime sum, Mellin transform and positive residues, The window criterion, The order of choices, The finite character comparison, Hosts for every character.

18✓Kernel-checked at the archive pin
Named declarations passed type, transitive-axiom (only propext, Classical.choice, Quot.sound) and kernel-replay checks at Lean 4.33.1.
19⊂Inside a checked closure
Called (directly or transitively) by a kernel-checked declaration, so it was elaborated and kernel-checked as part of that closure and its axioms are within that declaration's report. For a definition this only means that it typechecks: a proposition it defines (for instance VWinAt) can still be a hypothesis.
65◇Source-read
Statement and proof term located at pinned source lines and read; not replayed as a named declaration by the reading project's audit. It is part of the complete proof, whose replays are listed in the ledger.
0↗Full replay reported
Checked in a full replay of the archive reported by an external builder; the reading project has not (yet) recorded a run of its own for it.

Tiers of the 109 blueprint nodes, computed from the receipts registered in the reading project (as of 2026-11-20 23:51 UTC); motivation: 2. A component's own tier can also rest on receipts for declarations outside the blueprint, named in its status (for example Gz_eq, QH.pairEx and the bridge theorems).

External report

Andrew Sutherland and Simon DeDeo each report independently building and replaying all 175,187 original modules with Lean 4.33.1. DeDeo reports zero errors and crashes. Sutherland additionally records a con-leche comparison of the five targets and a lean4lean check of every original module. Five modules required higher reduction limits. The individual logs identify the names, flags and successful reruns.

Sutherland’s five named con-leche results · Sutherland’s five raised-fuel lean4lean reruns · Full solution and dated operator report · Second complete original-toolchain replay · Five individual axiom lists · Rust process log · Download checksums · Earlier challenge comparison · Independent five-target comparison · Independent rebuild with a separately generated proof export

The reading project's own checks

At Lean 4.33.1 we checked 24 named original results and 21 reading deductions across 2,244 distinct original modules. We also checked five endpoint equivalences and an Euler-factor boundary example. The one-sided prime-window implication passed fresh Lean replay with Mathlib and a second kernel on a portable proof of 61,697 declarations. The two-input deduction of ordinary RH and Dirichlet GRH, four steps from finite character comparison to the prime-sum bound and the auxiliary-prime theorem for three changing quadratic twists all passed. LemmaK, FixedIn and TwoIn yield one record of cohomological window constants for every cutoff exponent Ct ≥ 1. From Poitou–Tate duality, local finiteness and local Tate duality, a tower gives its cohomological bound. Adding Ferrero–Washington and LemmaK yields the joint window under those same hypotheses. Eight further teaching declarations use Lean 4.35.0. At each fixed primitive character, TopG and just its own cohomological estimate yield a cyclic-family bound with coefficient one. For the principal character modulo one this is the exact centered ordinary prime sum. The archive supplies the ordinary first moment unconditionally. Fresh kernel replay establishes that Schoenfeld’s RH-to-Chebyshev estimate suffices for the converse from RH to the window bound.

Headline at 2026-11-20 23:51 UTC: Two independent kernels accepted the complete proof, and all five theorems matched the canonical Mathlib challenge. Full original build on Helios (job 26647685): Building the original library targets and Main with 48 Lake workers and persistent group objects.

What the native receipts cover

ScopeAuthorityHypothesesConclusionOriginal modulesLean
Abstract window and reflection deductionsIndependent reading proofThe stated window and reflection premises.Zero exclusion and the critical-line deduction for abstract predicates.04.35.0
Integer-sphere proofIndependent reading proofInteger vectors in one finite dimension have the same squared norm. Their relation is j(v−u)=i(w−u), with j≠0 and i≠j.Positive squared distance, the chord identity, and exclusion of three distinct collinear lattice points.04.35.0
Closing deductionsOriginal archiveRight-half-plane nonvanishing of zeta or of all Dirichlet L-functions, as specified.The corresponding zeta RH or all-character strip statement follows.44.33.1
Analytic and Behrend landmarksOriginal archiveSeparate hypotheses for each selected result, given in the theorem table.Pole data for finite character families and a zero-free criterion from a uniform window bound. Relation avoidance on digit spheres and a quantitative large-set construction.384.33.1
Zeta prime-window implicationReading proof using the original archiveFor every ½ < θ < 1, one constant C bounds the named statistic by C·y^θ at every y in [Y,Y^(3/2)] for arbitrarily large Y. The statistic is Deduction.G.The singleton family identifies the analytic prime sum, and its bound implies ordinary RH.314.33.1
Geometric prime-window implicationOriginal comparison and reading proofFor every ½ < θ < 1, one constant C bounds the named statistic by C·y^θ at every y in [Y,Y^(3/2)] for arbitrarily large Y. The statistic is Formalism.Gz.An original pointwise normalization identity transfers the estimate and yields ordinary RH.1304.33.1
Exceptional-twist windowsOriginal archiveFinite exceptional sets and the stated scale inequalities.A whole interval simultaneously controls every exceptional twist.324.33.1
Local minor divisibilityOriginal archiveA commutative ring, prime p, w²=p, two input rows supported at coordinate zero modulo w, and M^(p^e)≡N′ modulo p entrywise.The specified two-by-two minor of the transformed rows is divisible by w.5094.33.1
Fresh kernel replay of the zeta prime-window implicationReading proof using the original archiveFor every ½ < θ < 1, one constant C bounds the named statistic by C·y^θ at every y in [Y,Y^(3/2)] for arbitrarily large Y. The statistic is Deduction.G.The same three reading declarations were rechecked from an empty kernel environment, including the safe non-partial constants imported from Mathlib.314.33.1
Auxiliary primes and their twistsOriginal archiveA positive integer M, 0<ε<1/52 and A≥1.Every sufficiently large Y admits a pair of primes with the required congruences, smoothed prime-sum and derivative bounds on [Y,Y^A], and finite zero-free boxes for the specified nontrivial quadratic twists.2934.33.1
Conditional zeta and Dirichlet endpointsOriginal deductions and reading proofThe archive’s TopG family estimate and its VWinAt cohomological bound with cutoff 401 for each primitive character.The common host construction, conductor reduction and reflection imply ordinary RH and both all-character endpoints.1,4214.33.1
Finite character balance and prime windowsOriginal archiveThe source's character estimates, size and norm inequalities, and compatible hosts on the window.Four deductions transfer the theta discrepancy to a weighted prime-sum upper bound with an explicit exponent.3164.33.1
The mathematical language of the five endpointsReading proof using the original archiveDefinitions of zeta, Dirichlet L-functions, positive levels and open strips.Five statement equivalences and a concrete level-two boundary zero.14.33.1
Arithmetic window assemblyOriginal archiveAn exponent satisfying LemmaK, the fixed odd-prime tower bounds FixedIn, and the uniform two-adic tower bound TwoIn.One record of constants K gives the joint arithmetic window ThmBab K Ct for every Ct≥1.6594.33.1
Inputs to the arithmetic tower windowOriginal archiveThe exact duality, finiteness, tower, Ferrero–Washington and growth premises specified for each theorem.A cohomological tower bound and a simultaneous choice of arithmetic windows.1,1114.33.1
The arithmetic source of the windowsReading proof using the original archiveAssume TopG and VWinAt ψ 401 for the prescribed primitive character.The cyclic family's sum is at most y^θ on whole late windows. For the character modulo one this sum equals G(y).1,4184.33.1
The converse for the ordinary zeta windowOriginal moment theorem and reading deductionsThe reverse implication uses Schoenfeld's RH-to-Chebyshev estimate.The archive supplies the first moment. Finite Abel summation and integration on a half-line connect its actual prime sum to the window bound.1484.33.1

Open verification questions

  1. Exact endpoint replay. Compile the five unmodified exports at the archive pins, verify their exact types, inspect their transitive axioms, and replay the selected kernel checker.
  2. Separate the analytic implication. Trace the finite positive-weight character family through poleData, landau_window and the uniform window bound. Check the pole treatment and the order of quantifiers.
  3. Nonempty compatible data. Follow the witnesses for the same geometric family through the test-vector grids, linked bundles and nonempty definite-side hypotheses.
  4. Source comparison of the Galois input. Unfold the row, the local-normalization convention and the theta-lift producer together. Check all quantified places, twists and weight restrictions.
  5. Uniform cohomological bound. Track finiteness and the absolute constants from the four counting cases to the finite prime sums on the claimed windows.
06 · The architecture, compiled

A Lean skeleton of the whole argument

This repository's Lean library (Lean 4.35.0, no imports) states every component as an explicit hypothesis over abstract types and proves only how they compose — the shape of the proof, checked. A toy model with zeros on the line shows the hypotheses are jointly satisfiable. It is a map, not a proof of RH.

The arithmetic supply, as the single hypothesis

/-- The arithmetic supply, in the form consumed by `zeroFreeAll_of_topG`. For each primitive `ψ`
and each admissible small `ε` there is a family containing `ψ` whose prime sum satisfies the
window bound with exponent `loss ε = 1/2 + (e_D + e_L + 7)·ε`, where `e_D = 2`, `e_L = 0` were
fixed **before** `ε`. In the archive this is produced from `TopG` and `VWinAt ψ 401` (through
`supplyChi`, `core_gm` and `topDesWin`). It is the one non-classical input. -/
structure Supply (P : Plane) (E : Engine P) where
  Eps : Type
  Small : Eps → Prop
  loss : Eps → E.Exp
  family : P.Char → Eps → E.Family
  mem_family : ∀ ψ ε, E.Mem ψ (family ψ ε)
  windows : ∀ ψ ε, P.Primitive ψ → Small ε → E.Windows (family ψ ε) (loss ε)
  /-- θ after the zero, ε after θ (`level_of_tiedWin`): for `Re s > 1/2` take
  `θ = 1/2 + min(Re s - 1/2, 1/2)/2`, so that `Re s > θ`, then a small `ε` with `loss ε ≤ θ`.
  This is possible because `loss ε → 1/2` as `ε → 0` with the losses fixed. -/
  choose : ∀ s, P.Right s → ∃ θ, E.RightOf s θ ∧ ∃ ε, Small ε ∧ E.Le (loss ε) θ

Quoted from lean/Atlas/Architecture.lean.

The second verified cut, in the skeleton

/-- **The second verified cut** (mirrors `Reading.DirichletBridge.endpoints_of_topG_vWin401`,
kernel-checked against the archive in a 1,421-module closure): the arithmetic supply implies
RH, the half-plane theorem for every character and Dirichlet GRH in the strip. -/
theorem endpoints_of_supply (R : Reflection P) (D : Descent P) (E : Engine P) (S : Supply P E) :
    RH P ∧ HalfPlane P ∧ StripGRH P :=
  have hhalf := halfPlane_of_primitive D (primitiveHalfPlane_of_supply E S)
  ⟨rh_of_halfPlane R D hhalf, hhalf, strip_of_halfPlane R D hhalf⟩

Quoted from lean/Atlas/Architecture.lean.

The whole architecture in one record

/-- The four layers of the argument bundled together. The first three are classical or
machine-checked; the fourth, `supply`, is where the arithmetic lives. -/
structure Architecture where
  plane : Plane
  reflection : Reflection plane
  descent : Descent plane
  engine : Engine plane
  supply : Supply plane engine

/-- The five endpoints, as conjunctions on the abstract plane. -/
theorem Architecture.endpoints (A : Architecture) :
    RH A.plane ∧ HalfPlane A.plane ∧ StripGRH A.plane :=
  endpoints_of_supply A.reflection A.descent A.engine A.supply

Quoted from lean/Atlas/Architecture.lean.

The two cuts against the real archive

/-- The constant may depend on the exponent. A chosen window must work at every real point. -/
def OneSidedWindows (g : ℝ → ℝ) : Prop :=
  ∀ θ : ℝ, 1 / 2 < θ → θ < 1 → ∃ C : ℝ,
    ∀ Y₀ : ℝ, ∃ Y : ℝ, Y₀ ≤ Y ∧
      ∀ y ∈ Set.Icc Y (Y ^ (3 / 2 : ℝ)), g y ≤ C * y ^ θ

/-- A one-sided estimate on arbitrarily late three-halves windows implies Mathlib's RH. -/
theorem riemannHypothesis_of_prime_windows (h : OneSidedWindows Deduction.G) : RiemannHypothesis := by

/-- Pass to arbitrary positive moduli, then use the archived critical-strip
and zeta reductions. The right-half-plane predicate retains its pole exception. -/
theorem endpoints_of_topG_vWin401 (hG : UpperTheory.Top.TopG)
    (hV : ∀ ψ : PChar, VWinAt ψ 401) :
    RiemannHypothesis ∧ DirichletLNonvanishingRightOfHalf ∧ DirichletLZerosOnLine := by

Quoted from references/archive-bridges/; compiled against the archive at Lean 4.33.1 by the reading project (receipts in its evidence/).

The destination layer, with Mathlib's ζ

/-- **The destination layer, checked against Mathlib**: right-half-plane nonvanishing of `ζ`
implies Mathlib's `RiemannHypothesis`, by the skeleton's `rh_of_halfPlane` instantiated with `ζ`. -/
theorem riemannHypothesis_of_right
    (h : ∀ s : ℂ, 1 / 2 < s.re → s ≠ 1 → riemannZeta s ≠ 0) : RiemannHypothesis :=
  rh_iff.mp (Atlas.rh_of_halfPlane reflection descent (halfPlane_iff.mpr h))

Quoted from lean-mathlib/AtlasZeta.lean: the skeleton's reflection step instantiated with riemannZeta and Mathlib's functional equation. Compiled against Mathlib at the archive's pins (Lean 4.33.1); axioms of riemannHypothesis_of_right: propext, Classical.choice, Quot.sound.

Axiom report of this repository's library

Lean (version 4.35.0, arm64-apple-darwin24.6.0, commit eb6cafb7290e8bd5f2937522d8d804f58f02bafc, Release) · build passed · problems: 0 · generated 2026-11-21T00:21:04+00:00

DeclarationAxioms
Atlas.halfPlane_of_primitive∅none
Atlas.strip_of_halfPlane∅none
Atlas.rh_of_halfPlane∅none
Atlas.primitiveHalfPlane_of_supply∅none
Atlas.endpoints_of_supply∅none
Atlas.rh_of_prime_windows∅none
Atlas.Architecture.endpoints∅none
Atlas.toyArchitecture∅none
Atlas.Separation.eq_of_dvd_subpropext, Quot.sound
Atlas.Separation.at_most_one_per_centrepropext, Quot.sound
Reading.excludeZeroByWindows∅none
Reading.criticalLineByReflection∅none
BehrendReading.normSq_nonnegpropext
BehrendReading.coordinates_zero_of_normSq_zeropropext, Quot.sound
BehrendReading.distance_pos_of_nepropext, Quot.sound
BehrendReading.chord_identitypropext, Classical.choice, Quot.sound
BehrendReading.sphere_relpropext, Classical.choice, Quot.sound
BehrendReading.sphere_midpointpropext, Classical.choice, Quot.sound
07 · From announcement to archive

The twenty timeline entries

Dates as written on pages 3–4 of the announcement (some mean “from that day on”), coloured by the component each entry corresponds to; grey entries are motivation without a matching theorem. Click a dot.

4 Oct8 Oct15 Oct22 Oct1 Nov8 Nov15 Nov1. Stepanov determinant (10/6/26) — Historical motivation14. Determinantal positivity with multiplicity one (10/6/26+) — Historical motivation411. The tower of fields (10/6/26) — Supporting comparison1113. The slope inequality (10/6/26) — Mathematical statement matched136. Two-support positivity test (10/8/26) — Historical calculation62. Invariant differentials on a CM curve (10/9/26) — Historical calculation27. Interlacing weights and periods (10/14/26) — Supporting comparison716. Period pairings and approximate orthogonality (10/14/26) — Supporting comparison163. The finite-field sign argument (10/19/26+) — Historical analogy319. A Behrend set in the tower construction (10/19/26) — Supporting comparison195. The smoothed von Mangoldt sum (10/20/26) — Formula matched58. A chain with square-root length (11/15/26) — Quantitative comparison89. Norms, levels and arithmetic periods (11/15/26) — Supporting comparison910. Divisibility of repeated lifts (11/15/26) — Supporting comparison1012. Seesaw and Bessel periods (11/15/26) — Supporting comparison1214. The provisional exponent two-thirds (11/15/26) — Intermediate strategy1415. A common integral structure across characters (11/15/26) — Supporting comparison1517. Cohomology and a weakening of Vandiver (11/15/26) — Quantitative comparison1718. Covolumes, lattice indices and congruences (11/15/26) — Supporting comparison1820. Dirichlet GRH (11/16/26) — Endpoint statement matched20
#DateEntryCorrespondence with the archiveComponent
110/6/26Stepanov determinant
The single-line entry records an origin of the argument. Its underlying function-field determinant is unspecified.
Historical motivation
210/9/26Invariant differentials on a CM curve
The entry describes inert-prime divisibility of iterated derivatives. The argument for this displayed exponent was not identified in the selected source path.
Historical calculation
310/19/26+The finite-field sign argument
The footnote compares a zero mode with an error bounded by iterated Parseval. Its objects and hypotheses differ from the Dirichlet L-function endpoint.
Historical analogy—
410/6/26+Determinantal positivity with multiplicity one
The short entry leaves W and S implicit. An exact theorem-to-type identification would require definitions of both.
Historical motivation—
510/20/26The smoothed von Mangoldt sum
The displayed zeta error has the exact ramp and the correction (1 + Euler’s constant)x. Character twists enter the level average. The sentence continues onto page 4.
Formula matched
610/8/26Two-support positivity test
The two indicator functions are a proposed model for a zero-free exponent. The final power-window criterion has a different statement.
Historical calculation
710/14/26Interlacing weights and periods
A local Bessel model is constructed from two interlacing conditions. The announcement’s claim about determinant values 0 and 1 has further content.
Supporting comparison
811/15/26A chain with square-root length
The arithmetic rank specification puts m between one and two times x raised to 1/2 minus epsilon/16. The finite kernel recursion explains iteration.
Quantitative comparison
911/15/26Norms, levels and arithmetic periods
One result relates a CM period to a logarithmic derivative. Another cancels a level volume against an index. The complete order-m cancellation requires more data.
Supporting comparison
1011/15/26Divisibility of repeated lifts
A specific minor vanishes modulo a divisor and a lifted column has a higher-order remainder. Turning these local facts into the required global gain is an additional step.
Supporting comparison
1110/6/26The tower of fields
The formal window ranges over fields specified by a fixing subgroup in a cyclotomic extension and by layer parameter e(x).
Supporting comparison
1211/15/26Seesaw and Bessel periods
Conditional unfolding identifies a theta-lift period with an integral. A middle-step volume identity is independent of the admissible character. A full iterated comparison is a stronger assertion.
Supporting comparison
1310/6/26The slope inequality
The Euclidean lattice model bounds the slope of an injectively mapped source by the maximal target slope plus the logarithm of the map bound. Identification of the particular period lattice is separate.
Mathematical statement matched
1411/15/26The provisional exponent two-thirds
The author describes a barrier for an earlier estimate. The endpoint reads a one-half-plus-loss exponent. The earlier barrier is not derived in this comparison.
Intermediate strategy
1511/15/26A common integral structure across characters
The final deduction imposes local and global controls on the same family. A precise comparison with the gcd in the historical description remains open.
Supporting comparison
1610/14/26Period pairings and approximate orthogonality
Unfolding relates a Poincare-series pairing to an integral of periods. The later host comparison turns suitable bounds into an estimate for the weighted prime sum.
Supporting comparison
1711/15/26Cohomology and a weakening of Vandiver
The window requires finite lengths and an aggregate weighted bound. Further implications for individual K-groups are conjectural comments in the announcement.
Quantitative comparison
1811/15/26Covolumes, lattice indices and congruences
A covolume bound is an explicit premise of the automorphic assembly. The local congruence lemma gives a representative higher-order equality.
Supporting comparison
1910/19/26A Behrend set in the tower construction
Encoding a populated integer sphere in a large base excludes short affine relations and controls intersections with short orbits.
Supporting comparison
2011/16/26Dirichlet GRH
The source declares that zeros of every Dirichlet L-function in the open critical strip are on the critical line. This entry records source-statement correspondence.
Endpoint statement matched
08 · Glossary

Words that carry weight here

Centered prime sum
$G(y)=\sum_{n\le y}\Lambda(n)(y/n-1)-y\log y+(1+\gamma)y$; its twisted versions $g_\chi$ add $y\,(L'/L)(1,\chi)$ instead of the main terms.
Power window
All $y$ with $Y\le y\le Y^{\lambda}$; the archive uses $\lambda=3/2$ for the prime sums and $[Y,Y^2]$ for the cohomology.
One-sided bound
An upper bound $G_F(y)\le Cy^\theta$ at every point of a window; no lower bound is required.
Positive family
A finite family of characters with weights $a_i>0$ (principal member only of modulus one); residues at common zeros add up.
Primitive character
A Dirichlet character at its conductor; every character is induced by a unique primitive one.
Theta lift
Transfer of automorphic forms between groups by a theta kernel; the archive chains about $\sqrt x$ of them.
Galois row
Row0b: the local–global compatibility package attaching $\ell$-adic Galois representations to regular algebraic polarized cuspidal automorphic representations over CM fields.
Cohomological length
The $\mathbb Z_\ell$-length of the locally trivial part of $H^1$ of a twisted module; VWinAt bounds weighted sums of these lengths.
Verified cut
An implication machine-checked against the archive with the remaining inputs as explicit hypotheses.
Axiom audit
Lean's list of axioms a declaration depends on; the audits permit only propext, Classical.choice and Quot.sound.
Source-read
A declaration located at pinned source lines and read, but not separately replayed by this audit.
Pinned link
A URL into the public mirror at commit 40e56bab, whose tree is byte-identical to the archive.
09 · Sources, pins and images

Where everything comes from

Pins

  • Archive SHA-256 9631b4f020199e580ba86178cb56c6e9b09f690c2369a80dc4334aa93f49c7fc (14 parts, 175,187 modules), Lean leanprover/lean4:v4.33.1, Mathlib 0df444a360ea.
  • Public mirror Deicyde/riemann-hypothesis at 40e56bab0265; every Lean link points there.
  • Announcement PDF SHA-256 3fc97784b71ddd44ff8aa0a177e99622a1c61b13cfa843a0989ee4a7a0cfeb8d — alpo.ge/remark.pdf.
  • Reading project (the data of this atlas): the reading project (private working copy) at commit 348ec48748e0, imported 2026-11-21T00:17:58+00:00.

Reproduce

make data      # import + model
make figures   # compute figures, assert facts
make lean      # build the Lean library, axiom audit
make atlas     # this page
make pdf guide book
make check

Built 2026-11-21T01:19:17+00:00. Fonts and MathJax load from a CDN; without a network the page still works with system fonts and TeX source for formulas.

About the images