Masullo Proof Atlas

How Masullo's claimed proof of GRH is put together

On the same day as the Alpöge–Claude Lean release, Ravi Vakil's blog posted a 940-page paper by Alessandro Masullo claiming the Riemann Hypothesis for the Hecke L-functions of every number field. It is a human-written paper, twelve years in the making, released before it was finished.

Nothing here says the proof is correct. This map is based on the introduction (pages 11 to 33), the statement of the main theorem (pages 519 to 522), and a text scan of the whole document for unfinished parts. The other 900 pages have not been read.

What is claimed

For any number field F and any unitary Hecke character χ, every zero of the completed L-function L(s, χ) in the critical strip has real part ½. Taking F to be the rationals gives the Riemann zeta function and all Dirichlet L-functions, which is the case the Lean proof covers. General number fields go further.

The argument in seven steps

1
Tate's thesis as a stageL-functions come from averaging functions on the adeles. Set up the spaces of those averages, the “theta functions”.
written outch. 2, 3, 5, 6
2
A space whose points are the zerosDivide all test functions by the theta functions. What is left is a space with one slot for each zero of L.
written outch. 5 to 7
3
Listen only to a thin stripComplete everything in a new norm, chosen so that only zeros within distance a of the critical line survive. This gives the spectral side.
written outch. 9, 10
4
Cut the kernel into one piece per primeThe difference between the convolution kernel and the spectral kernel is split along the primes, using explicit bases.
written outch. 4, 7, 8, 11
5
Almost every piece has trace zeroFor each prime outside a finite set S, Poisson summation shows the piece contributes nothing.
written outch. 11
6
Compute what is leftThe remaining piece is computed with a trace formula of Connes from 1998. The result is the main theorem, a trace formula.
written outch. 12, p. 519
7
Read off the zerosFeed a family of test functions into the trace formula to show the logarithmic derivative of L has no poles to the right of ½.
not writtench. 13 is a stub
written out a full chapter exists; I have not checked itsketched described in the introduction onlynot written placeholder in this release

Where to go

The idea, without the machinery

The explicit formula says a sum over zeros equals a sum over primes. Masullo's plan is to realize both sides as the trace of one operator, then change the space the operator acts on until the zeros away from the critical line can no longer be heard.

A trace computed two ways

The trace of a matrix can be computed as the sum of its eigenvalues or as the sum of its diagonal entries. A trace formula is that equality for an operator on an infinite-dimensional space. One side, the spectral side, sums eigenvalues. The other, the geometric side, integrates the operator's kernel along the diagonal.

Since Weil and Connes the hope has been to find an operator whose eigenvalues are the zeros of an L-function and whose geometric side is a sum over primes. That would make the explicit formula a trace formula. Masullo's paper is in this tradition and builds on Tate's thesis, Connes's 1998 paper, and work of Ralf Meyer.

The toy example

The introduction gives a model of the whole argument on a circle. Take a number γ greater than 1 and wrap the positive reals around so that t and γt are the same point. Functions on this circle have a basis of pure tones t2πim/log γ, one for each whole number m.

Let a test function φ act by convolution. Each tone is an eigenvector, so the trace is the sum of the eigenvalues, which are values of the Mellin transform of φ. Poisson summation turns that sum into something else:

Σ_m  M(φ)(2πim / log γ)   =   log γ · Σ_k  φ(γ^k)
1/γ1γγ² test function φ t, log scale φ = 0φ = 0φ = 0φ = 0
If φ lives strictly between 1 and γ it is zero at every power of γ, so the right-hand side vanishes and the trace is 0. In the paper γ becomes the size of a prime, and this is why a prime larger than the reach of φ contributes nothing.

The real version: one circle per prime

In the real setting the space is built from the adeles, and the operator's kernel has a contribution from every prime. The paper constructs a basis, which it calls an orbital Schauder basis, on the piece belonging to each prime. On that piece the calculation is the toy example, with γ replaced by the norm of the prime.

A test function φ supported between 1 and some bound T then sees only the finitely many primes of norm below T. Everything else has trace zero. The finite remainder is computed by Connes's method and gives the prime side of the explicit formula.

The thin strip

The new move is on the other side. Previous work used spaces in which every zero of L contributes an eigenvalue, wherever it lies. Masullo completes the spaces in a norm he calls Hardy–Sobolev, with a parameter a. In that completion only the zeros within distance a of the critical line still give eigenvalues. Evaluation at a zero further out is no longer a continuous operation, so it drops out of the trace.

0½1 ½ − a½ + a a zero out here would besilent in this space zeros inside the bandare counted the prime side of the formula comes out the same for every a
The claim that carries the proof: the geometric side can be computed inside the thin-strip space and gives the same answer however small a is. The spectral side then only ever contains zeros within a of the line.

Reading off the zeros

The last step uses the test functions ts on t > 1, approximated by allowed ones. With these the prime side becomes the logarithmic derivative of L, and the trace formula expresses it through zeros inside the band alone. A logarithmic derivative has a pole at every zero. So L has no zeros with real part beyond ½ + a, and since a can be as small as you like, none beyond ½. The functional equation gives the other side.

The introduction notes two things about this route. It does not use Weil's positivity criterion, the usual target in this tradition. And it uses known facts about L as inputs: no zeros on the line Re s = 1, and growth bounds in vertical strips.

Where a reader should look hardest. Weil's explicit formula has every zero on its spectral side. Masullo's has only those in the band, with the same prime side. Subtract the two and the zeros outside the band must contribute nothing, which is the theorem. So the whole proof rests on the claim that the geometric side really can be computed inside the thin-strip space: steps 4 to 6. That is an observation about the structure of the argument, not a finding that anything is wrong.

The argument by chapter

The seven steps, with where each lives in the 940 pages. Page numbers are those printed in the release of 20 November 2026.

1 · theta functions and zeta integrals 2 · quotient space: one slot per zero 3 · thin-strip completion: spectral side 4 · split the kernel prime by prime 5 · pieces outside S have trace 0 6 · the S-piece, by Connes's formula trace formula, Theorem 12.5.1.1 7 · log-derivative has no poles past ½ GRH for Hecke L-functions spectral sidegeometric side
The left column builds the side that sums over zeros, the right column the side that sums over primes. They meet in the trace formula. The dashed box is the step this release does not contain.
1Tate's thesis as a stagewritten out

Take a rapidly decaying function f on the adeles of F and sum it over the multiplicative orbit of F: Φ(f)(g) = Σ f(ξg). The result is a function on the idele class group CF, which the paper calls an adelic theta function. Tate showed in 1950 that the Mellin transform of such a function is an L-function times something entire. The first chapters build these spaces carefully, both globally and “S-locally”, using only a finite set S of places.

Chapter 2, Adelic θ-functions (p. 57) · Chapter 3, Idelic Zeta Functions (p. 117) · Chapter 5, Spaces of adelic θ-functions (p. 221) · Chapter 6, Local-Global structures (p. 287) · Appendices A, C, D

2A space whose points are the zeroswritten out

The theta functions sit inside the space of all rapidly decaying functions on CF. Under the Mellin transform they become exactly the multiples of L(s + ½, χ). So the quotient, all functions modulo theta functions, is a space of functions on the zero set of L: one dimension for each zero, counted with multiplicity. The paper calls this the fundamental divisor exact sequence.

0 → (theta functions) → (all test functions) → H_F → 0
           ↓ Mellin transform
0 → L(s+½, χ)·(entire) → (entire) → ⊕ over zeros z of ℂ^N(z) → 0

Showing this is exact with the right topologies is what the five “Structure Theorems” are for. It uses the convexity bound and polynomial growth of 1/L along a sequence of horizontal lines.

§1.2.1 (p. 13) · Chapter 5 Structure Theorems I to III · Chapter 6 Structure Theorems IV, V · Chapter 7, Schauder Bases of Idelic Zeta functions (p. 311)

3Listen only to a thin stripwritten out

Fix an interval I = [−a, a]. The Hardy–Sobolev space Ws,2I is the completion of the test functions in a weighted L² norm, with weight max(|g|2a, |g|−2a). Complete all three terms of the exact sequence. The quotient is now a Hilbert space spanned by the zeros in the strip ½ + I, and convolution by a test function φ acts on it with trace

Σ over zeros ρ with ρ − ½ in the strip of   N(ρ) · M(φ)(χ|·|^(ρ−½))

That is the spectral side. The paper says these spaces were not considered by Connes or Meyer, and that the case a = 0 recovers Connes's 1998 spectral interpretation.

§1.1.3, §1.2.1, §1.4.8 (pp. 11 to 14, 31) · Chapter 9, Hardy-Sobolev topologies (p. 395) · Chapter 10, Trace formula: Spectral Side (p. 433)

4Cut the kernel into one piece per primewritten out

Convolution by φ has kernel φ(xy⁻¹). If φ is supported away from 1 this kernel is zero on the diagonal, so its trace is zero. Define the geometric kernel as φ(xy⁻¹) minus the spectral kernel. Its diagonal integral is then minus the spectral side, by definition. The work is to compute that integral a second way.

The kernel is lifted to the ideles and cut along a filtration indexed by the primes, giving kgeom = kS + Σp kp. Each kp acts on an “orbital subquotient” that carries an explicit basis, a copy of the pure tones in the toy example. The paper calls these orbital Schauder bases and says they are its main novelty.

§1.2.2, §1.2.9, §1.3 (pp. 15 to 27) · Chapter 4, Periodic Schauder Bases (p. 155) · Chapter 8, Periodic Orbits and Fixed Points (p. 381) · Chapter 11, The Shell Structure (p. 475)

5Almost every piece has trace zerowritten out

On each orbital piece the trace is a sum over the zeros of a local Euler factor, and Poisson summation rewrites it as a sum of values of φ at powers of the prime's norm, exactly as in the toy example. If the support of φ lies below the norm of the prime, every term is zero. So for a fixed φ only the primes in a finite set S contribute.

Carrying the bases from the S-local spaces to the global ones is where the non-vanishing of L on the line Re s = 1 is used.

§1.2.8, §1.2.9, §1.4.7 (pp. 16 to 19, 30) · Theorem 11.3.2.1 · Chapter 11.3, Orbital Kernels and Orbital Traces (p. 491)

6Compute what is leftwritten out

The remaining piece kS involves only the places in S. Its diagonal integral is computed with a version of the S-local trace formula from Connes's 1998 paper. The Fourier transform on the S-adeles turns a factor |x|a into the change from multiplicative to additive measure only when a = ½, which the paper presents as the reason ½ appears in the hypothesis.

The outcome is the theorem the paper calls its main result. For a test function φ vanishing on the norm-one classes:

Σ over all places v of   (local integral of φ(g)·|g|^½ / |1 − g| at v)
    =  M(φ)(|·|^½) + M(φ)(|·|^−½)  −  Σ over zeros c in the strip of  N(c)·M(φ)(c·|·|^−½)

The left side is the prime side of Weil's explicit formula. The right side has the two pole terms and the zeros in the strip ½ + I.

§1.2.3 (p. 16) · Chapter 12, Trace formula: Geometric Side (p. 497) · Theorem 12.4.1.1 · Theorem 12.5.1.1 (p. 519)

7Read off the zerosnot written

The introduction sketches this in two pages. Approximate the functions φz(t) = tz for t > 1 by allowed test functions. Montel's theorem gives uniform convergence of both sides as holomorphic functions of z. The prime side converges to the logarithmic derivative of the finite L-function. The conclusion is that L′/L(s + ½, χ) is holomorphic for Re s > a, apart from the pole of zeta, for every a > 0.

Chapter 13, which should contain this, consists of two empty section headings, a restatement of the trace formula marked “Correct statement”, and a theorem with no statement at all. See What is missing.

The restated theorem also contains a multiplicity cap: each zero is counted min(N, [F : ℚ]) times, where N is its multiplicity. For the rationals that would count every zero once.

§1.4.2 to §1.4.5 (pp. 28 to 30) · Chapter 13, Zeros of Hecke L-functions (pp. 521 to 522)

What is missing

The author says the mathematics is complete and the writing is not. This page lists what a text scan of all 940 pages found unfinished. It does not assess the parts that are written.

The step that matters

WhereWhat is there
§13.1 Recollections on the Explicit Formulas (p. 522)empty Heading only.
§13.2 Holomorphic families of test functions (p. 522)empty Heading only. This is the approximation argument of §1.4.2 to §1.4.3.
Theorem 13.3.1.1, Trace Formula restated (p. 522)flagged Footnote: “Correct statement.”
Theorem 13.3.1.2, Divisor of L(s, χ) (p. 522)empty No statement. Footnotes: “Insert statement.” and “State and prove the Main Theorems, also explaining how to deduce them by approximation”. It is derived from “Theorem ??”.

So the release proves a trace formula and stops. The deduction of GRH from it exists only as the sketch in the introduction.

The introduction

WhereWhat is there
§1.1.1 Tate's Thesis (p. 11)empty
§1.2.4 to §1.2.6: poles of S-idelic zeta functions, Sarnak's long horocycles, periodic orbits (p. 16)empty Three headings in a row.
§1.3.3 The Shell structure, §1.3.4 Paley-Wiener Levels (p. 27)flagged Two or three sentences each, footnoted “Finish!”
§1.6 How this work was conducted, and why (p. 33)empty Including §1.6.2, “The role of Brian Conrad”.
§1.7.1 (p. 33)flagged Refers to “Caveat ??”.

Elsewhere in the body

WhereWhat is there
Chapter 12 (p. 514)flagged Footnote: “Include section.”
§14.1.2 Expected Multiplicity Bound (p. 528)flagged Footnote: “Correct this subsection.” Pages 528 and 533 cite a “Positivity Criterion … Theorem ??” that is not in the document. The introduction says positivity is not used for the main result.
Chapter 5 (p. 230, 234)flagged “Include version with I-exponentially decaying target”; one undefined equation reference.
Chapters 2, 4 (pp. 69, 88, 178)flagged Three undefined references to lemmas and a definition.

Appendices

WhereWhat is there
Appendix E, Topological stacks (p. 823)flagged Footnote on the title: “Complete this Appendix, correcting it and shortening it.” Also “Check this definition is correct” (p. 862), “This proof needs correction” (p. 850), and eight undefined references.
Appendix B, analytic number theory (pp. 591, 603, 626, 648)flagged Four reviewer-style queries, including one asking whether a bound is meant to be uniform, and a reference to “Chapter ??”.
Appendices C, D (pp. 667 to 809)flagged About eight queries asking for references, expansions or clarifications.

Totals from the scan

  • 16 undefined cross-references (printed as “??”), most of them in Appendix E.
  • 33 footnotes that are editing notes or questions to the author.
  • 8 section headings with no text under them, all in Chapter 1 and Chapter 13.

The scan works on extracted text and matches patterns, so it can miss gaps that are not marked, such as a proof that is present but incomplete. The script is tools/gapscan.py in the repository and the page-by-page results are in notes/masullo-gaps.md.

Two proofs compared

Both appeared on 20 November 2026. They share a conclusion and almost nothing else.

Alpöge with ClaudeMasullo
ClaimRH and GRH for Dirichlet L-functionsGRH for Hecke L-functions of every number field, which includes the Dirichlet case
FormLean 4 source: 175,187 files, about 1 GB, no comments, no paperA 940-page paper, no formalization
Made byOne mathematician working with an AI system over about six weeksOne mathematician over about twelve years, advised by Brian Conrad
TraditionStepanov's elementary proof for curves over finite fields, carried to number fieldsTate's thesis, Weil's explicit formulas, Connes's trace formula
Key deviceA chain of about √x theta lifts between unitary groups, standing in for “vanishing to high order”A function space that only registers zeros in a thin strip, plus explicit bases that split a kernel prime by prime
How zeros are excludedLandau: a one-sided bound on a smoothed prime sumThe logarithmic derivative of L is shown to have no poles right of ½
Heavy inputsGalois representations, base change, about 195 theorems from the literature re-proved in LeanFunctional analysis of nuclear spaces, non-vanishing on Re s = 1, growth bounds for L
Function fieldsThe method descends from the function-field proofThe author explains why the method does not work for function fields (§1.7.1, §14.2)
What is checkable todayEverything, by machine, once the kernel replay finishes. Nobody yet understands it.The trace formula can be read by experts. The final deduction is not written.
Statuskernel check in progressincomplete manuscript

If both were right they would be independent proofs by unrelated methods, and Masullo's would be the more general theorem. Neither has been verified.

Lean notes

Small pieces of the argument stated, and where possible proved, in Lean 4 with Mathlib. The aim is to pin down exactly what each step says. It is not an attempt to formalize the paper.

The files are in the repository under lean/RH/, built against Lean 4.29.0 and Mathlib v4.29.0. A statement followed by sorry has been checked to make sense to Lean but has no proof yet.
FileWhat it saysState
Targets.leanThe target for the rationals: DirichletGRH, every zero of a Dirichlet L-function in the open strip has real part ½. Also NoZerosRightOfHalf, what step 7 delivers before reflection; it is the same statement the Alpöge proof calls EveryCharacter.compiles definitions only
Toy.leantrace_identity: the toy trace computed two ways, Σm M(φ)(2πim/log γ) = log γ · Σk φ(γk).stated proof is sorry
Toy.leantrace_eq_zero: if φ is supported strictly between 1 and γ, the trace is 0.proved from trace_identity
LogDeriv.leanne_zero_of_logDeriv_analyticAt: if the logarithmic derivative of an analytic function extends analytically across a point, the function is not zero there. This is how step 7 turns “no poles” into “no zeros”.stated proof is sorry

What Mathlib has and lacks for this paper

  • Has: the Riemann zeta function and Dirichlet L-functions with their functional equations, the statement RiemannHypothesis, the Mellin transform, Fourier analysis and Poisson summation on the real line, adele rings of number fields.
  • Lacks, as far as I can tell: Hecke characters with their L-functions for general number fields, Tate's thesis, Schwartz–Bruhat spaces on adeles, nuclear spaces and trace-class operators at the generality the paper needs.

So the target theorem can be stated today only for the rationals. The toy example and the final complex-analysis step need nothing beyond what Mathlib has.

Candidates for next

  • Prove the toy trace identity from Mathlib's Poisson summation.
  • State the Hardy–Sobolev norm on the positive reals and prove that evaluation of the Mellin transform at a point outside the strip is not continuous for it. This is step 3 in miniature.
  • State the trace formula for F = ℚ and the trivial character as a hypothesis, and derive RH from it formally. That would write down the missing Chapter 13 in the simplest case and show exactly what it needs.
← Quiz: Part I10 A trace computed two ways →