How the argument works
The Lean proof is far too long to read. But a first guide to its strategy now exists, and the big idea fits on a page. This page explains it in three layers: a plain story, a student version, and an expert version.
TL;DR
- An old trick, going back to Chebyshev in the 1850s: if a whole number is divisible by many primes, and you know it isn't too big, then there can't be too many primes.
- The proof uses that trick with a far cleverer "number", built from symmetries of special curves, and gets a very precise upper limit on how many primes there are.
- An upper limit that is precise enough rules out zeros off the critical line. That is RH.
- The guide that explains this is a draft. Its hardest step is still taken from the computer-checked proof, not proved by hand.
Jump to: the plain story · student layer · expert layer · what is still unclear
This page follows "The Alpöge–Claude proof of the Riemann hypothesis explained for the Riemann zeta function" (working draft, 21 November 2026). It was written by three AI systems (Claude Fable 5.2, Codex Astra 6.2 and Gemini Carbon) at the direction of Michael O. Rubinstein. It is a work in progress. It proves that RH follows from one inequality, called (M), but it does not itself prove (M): it takes (M) from the Lean development. Section numbers like §I.2 below refer to that guide. What has been checked by computer is the Lean proof, not the guide. See also the Reddit discussion where the guide was shared.
Layer 1: the plain story
Divisible by lots of primes + not too big = not too many primes.
Chebyshev's trick
In the 1850s Chebyshev bounded the number of primes using the prime factors of factorials. The simplest version, due to Erdős (1932), goes like this (§I.0).
Take the number of ways to choose 10 things from 20. It is
$$\binom{20}{10}=184{,}756=2^2\times 11\times 13\times 17\times 19.$$Look at the primes between 10 and 20: they are 11, 13, 17 and 19. Every one of them divides 184,756. That is no accident. On top of the fraction \(\frac{20!}{10!\,10!}\), each of those primes appears once, and underneath it doesn't appear at all, so it can't cancel.
Now the second fact: this number isn't very big. It is at most \(4^{10}=1{,}048{,}576\), because choosing 10 from 20 is one of the \(2^{20}=4^{10}\) ways to choose any subset of 20 things.
Put the two facts together. The primes 11, 13, 17 and 19 all divide 184,756, so their product \(46{,}189\) is at most 184,756, which is at most \(4^{10}\). The same works for every \(n\): the primes between \(n\) and \(2n\) multiply to at most \(4^n\). So there can't be too many of them. That is a real theorem about primes, proved from two facts about one whole number: what divides it, and how big it is (§I.0, §I.2).
Try it: Chebyshev's number
The bars use a logarithmic scale (number of digits). The teal bar is always shorter than the coral one. That gap is how much the trick "wastes".
Why the trick isn't enough on its own
Chebyshev's trick tells you the right ballpark for how many primes there are, but no more. The waste comes from rounding: counting how often a prime divides a factorial involves rounding down, and those rounding errors are as big as the thing you're trying to measure (§I.2).
So the guide asks for a better "number" with three wishes (§I.2): no rounding, a size you know extremely precisely, and primes you can relate to all primes.
The new idea
- Zeros become primes. If a zero sat off the critical line, a certain prime count would sometimes be too big. So an upper limit on that prime count is enough to rule out such zeros (§I.1).
- Find a better divisible object. Special curves with extra symmetry (called "complex multiplication") have derivatives whose divisibility by primes follows a smooth pattern, with no rounding (§I.3–§I.4).
- Know its size very precisely. Instead of a single number, the proof uses maps between grids of points (lattices) built from theta functions. Their size can be computed almost exactly (§I.5, §I.7).
- Repeat it about \(\sqrt x\) times. Doing the step many times shrinks the errors. Balancing two kinds of error gives exactly the exponent \(\tfrac12\), the critical line (§I.8).
The guide's comparison is Stepanov's method for curves over finite fields, which proves the analogue of RH there by counting with multiplicity (§I.2, §I.7).
Layer 2: for students
If a weighted prime sum never goes far above its expected value on long stretches \([Y,Y^{3/2}]\), there are no zeros right of the line.
The prime sum
Let \(\Lambda(n)=\ln p\) if \(n\) is a power of a prime \(p\), and \(0\) otherwise. The guide uses a smoothed prime count, where each prime power \(n\le x\) is weighted by the "ramp" \(x/n-1\) (§I.1, (I.2)):
$$S(x)=\sum_{n\le x}\Lambda(n)\Big(\frac{x}{n}-1\Big),\qquad G(x)=S(x)-x\ln x+(1+\gamma)x,$$where \(\gamma\approx0.5772\) is Euler's constant. The subtracted terms are the expected size of \(S(x)\), so \(G(x)\) is the error.
From zeros to the size of \(G\)
The explicit formula (§I.1, (I.4)) writes \(G\) as a sum over the zeros \(\rho\) of \(\zeta\):
$$G(x)=\ln 2\pi-\sum_{\rho}\frac{x^{\rho}}{\rho(\rho-1)}+O\!\left(x^{-2}\right).$$A zero \(\rho=\beta+i t\) contributes a term of size about \(x^{\beta}\). If every zero has \(\beta=\tfrac12\), the guide shows \(|G(x)-\ln2\pi|\le0.0462\sqrt x\) for large \(x\). Real values computed in the guide, which we rechecked for the first two: \(G(10^4)=1.749\), \(G(10^5)=3.671\), \(G(10^6)=11.040\). Compare \(0.0462\sqrt{10^6}=46.2\).
Why one side is enough
A zero \(\rho=\beta+i\gamma\) comes with its complex conjugate \(\beta-i\gamma\): same real part, because \(\zeta(\bar s)=\overline{\zeta(s)}\). Their two terms add up to a real wave, roughly \(x^{\beta}\cos(\gamma\ln x+c)\), that swings both up and down with size about \(x^{\beta}\). (Reflection across the critical line, \(\rho\mapsto1-\bar\rho\), is a different symmetry, coming from the functional equation; it is used below.) So if \(G(x)\) never goes far above \(x^{\theta}\), there can't be a zero with \(\beta\gt\theta\). This is made rigorous by a theorem of Landau on Mellin transforms of functions of one sign (§I.1):
One-sided criterion (§I.1). Let \(0\lt\theta\lt1\). If \(G(x)\le x^{\theta}\) for all large \(x\), then \(\zeta(s)\ne0\) for \(\operatorname{Re}s\gt\theta\).
On intervals. It is enough that there are arbitrarily large \(Y\) with \(G(x)\le x^{\theta}\) for every \(x\) in \([Y,Y^{3/2}]\). (Proof in Part II §1 of the guide.)
So to prove RH, it is enough to show: for every \(\varepsilon\gt0\) there are arbitrarily large \(Y\) with \(G(x)\le x^{1/2+\varepsilon}\) throughout \([Y,Y^{3/2}]\). Non-trivial zeros come in pairs \(\rho\), \(1-\rho\) (functional equation), so none to the right of the line means none to the left (§I.1).
And this is exactly the direction the divisibility trick gives. A nonzero whole number divisible by \(p^{e_p}\) for each \(p\) in some set satisfies (§I.2, (I.6))
$$\sum_p e_p\ln p\le\ln|\Delta|.$$Divisibility is about primes one at a time; size is about real numbers. This inequality always bounds primes from above, which is what the criterion needs.
Where the exponent \(\tfrac12\) comes from
The proof composes \(m\) steps. Errors made once get divided by \(m\), giving \(x/m\); errors that build up each step give \(m\). The total \(x/m+m\) is smallest when \(m\approx\sqrt x\), and then both are about \(\sqrt x\): exponent \(\tfrac12\). If the build-up were \(m^2\) instead, the balance \(x/m+m^2\) would give only \(\tfrac23\) (§I.8).
This balance only motivates the scale. The bound actually proved carries an extra factor \(x^{C\varepsilon}\) (see (M) below), so for each \(\varepsilon\) it gives \(G(x)\le x^{1/2+C''\varepsilon}\) on windows. Because \(\varepsilon\) can be taken as small as we like, that rules out zeros with any fixed real part \(\beta\gt\tfrac12\).
See the related prime-counting activity and the smoothed prime count on the Students page.
Layer 3 (experts): what follows from (M)
(M), plus Estimates B and C for auxiliary characters, gives Estimate A; A and B bound \(G\) on windows; the window form of Landau's lemma gives RH.
Skeleton (§I.0): (M) is quoted from Lean, not proved in the guide
$$\begin{gathered}\text{(M)},\ \text{B},\ \text{C}\\ \Downarrow\\ \text{Estimate A}\\ \Downarrow\\ G(x)\le x^{1/2+\varepsilon}\ \text{on } [Y,Y^{3/2}]\\ \Downarrow\\ \text{RH}\end{gathered}$$The data and (M) (§I.5)
Fix small \(\varepsilon\gt0\) and \(x\in[Y,Y^{3/2}]\). Choose primes \(q\equiv3\bmod 8\), \(d\equiv7\bmod16\) with \(q\mid d+1\), very large: \(\ln q\asymp Y^{\varepsilon/5}\). Let \(n\) be a power of 2 with \(x^{2+8\varepsilon}\le n\le 2x^{2+8\varepsilon}\), \(N=4n\), \(H\) the odd classes mod \(N\) up to sign, and \(b=e^{x^{\varepsilon}}\). Prime powers go into a \(2\times n\) table (row \(\tfrac12(1+\chi_q(r))\), column \([r]\)); cells carry weights \(w_i\), with distinguished cells (density \(\delta\)) of weight about \(b\) and all others tiny. \(P\) is a weighted sum of ramps \((a/w-1)^{+}\) with coefficients \(1-\chi_d\), and \(M\) its expected value. Averaging over translates \(u\in H\), labels \(t\in T\) and \(m\) repetitions,
$$\frac1n\sum_{u\in H}\sum_{t\in T}\sum_{j=1}^{m}\big(P_{u,t,j}-M_{u,t,j}\big)\le E,$$where
$$\frac{E}{B}\le x^{C\varepsilon}\Big(\frac{x}{m}+\frac{xm}{k}+m+\frac{k}{n^{2}}\Big),$$with \(B=|T|\,m\,b\,\delta/2\), \(m=x^{1/2+O(\varepsilon)}\), \(k=x^{9/2+O(\varepsilon)}\). The guide marks (M) as [quoted] from Lean.
From the table to \(G\) (§I.6)
- Two scales. Distinguished cells test primes near \(a/b\approx x\); all other cells test primes beyond \(\tfrac13\exp(x^{\varepsilon/3})\), where the classical zero-free region gives a saving better than any power of \(x\) (Estimate C).
- Averaging over translates removes the residue class mod \(N\), leaving \(R_{q,d}(x)=G(x)+G_{\chi_q}(x)-G_{\chi_d}(x)-G_{\chi_q\chi_d}(x)\). Estimate A: \(R_{q,d}(x)\le x^{1/2+C'\varepsilon}\).
- Estimate B. For \(0\lt\varepsilon\lt\tfrac1{52}\) and all large \(Y\) there are such \(q,d\) with \(|G_\chi(x)|\le x^{1/2+\varepsilon}\) for \(\chi=\chi_q,\chi_d,\chi_q\chi_d\) on \([Y,Y^{3/2}]\). This is not circular: \(q,d\) are astronomically larger than \(x\). It uses a zero-density estimate near \(\operatorname{Re}s=1\) and mean squares over the modulus, and needs only one good pair (proved in the guide, §III.1).
- Conclusion. \(G(x)\le R_{q,d}(x)+3x^{1/2+\varepsilon}\le x^{1/2+C''\varepsilon}\) on the window. The deduction from (M), B and C is proved in full in Part II §4–5.
Where (M) comes from: theta lifts (§I.7)
A theta lift between definite unitary groups is a finite matrix of derivatives of theta series of pairs of lattices, over the integers of a CM field \(E\). For an injective map \(T:L\to M\) of lattices with \(T(L)\subset p^{e}M\), comparing volumes gives the slope inequality (I.32): \(e\ln p\le\ln\|T\|+\mu_{\max}(M)-\mu(L)\). Three things must be established:
- Divisibility. At inert primes the matrix is divisible by \(p^e\) with \(e=\sum_{h\ \text{odd}}(a/p^h-w)^{+}\): derivatives of order \(a\) gain \(a(1/p+1/p^3+\cdots)\), and each minor absorbs only one derivative per direction (the \(-w\)). Composing \(m\) lifts adds exponents: the analogue of vanishing to order \(m\) in Stepanov's method. Written up in §III.5, conditional on hypothesis (B) from Lean.
- Size. \(\ln\|T\|^2\le a(\ln a-1-\ln\Omega_E)+\cdots\), leading term the Gaussian integral \(\|z^a\|^2=a!\). The period comes from Kronecker's limit formula / Chowla–Selberg (I.33); the remaining linear terms from the Siegel mass formula and the Weyl dimension formula. The same constants \(\gamma\) and \(L'/L(1,\chi_d)\) appear on the prime side, which the guide calls "the reason the method can work at all". The size bound itself remains quoted.
- Denominators and rank. Denominators are controlled by a bound, by \(b\), on orders of certain Galois cohomology groups (all primes \(\ell\), twists up to \(x^{400}\)), proved in the Lean development with automorphic forms and Galois representations. Rank uses many maps and a second-moment bound from a trace formula (§III.8).
Matching the guide to our proof map
We checked the three hardest names, ChainTied, HostExW and TopG, term by term: their Lean definitions against the guide's statements. "Confirmed in content" means every condition matches a statement in the guide, although the guide never uses that Lean name. File paths are under lean/src/ in the archive.
| Guide | Proof map step | Confidence |
|---|---|---|
| \(G(x)\), (I.2) | Step 22 G / Gm | Same formula |
| Landau's lemma on \([Y,Y^{3/2}]\), §I.1 | Step 21 FormBAt, used with \(\lambda=3/2\) in step 16 | Same shape of statement; the guide cites its own Lean file for the lemma |
| \(G(x)\le x^{1/2+C\varepsilon}\) on windows | Step 23 ChebyshevInequalityOnWindows | Same shape |
| Bound for Galois cohomology ("Theorem G"), §I.7 item 3, §III.9 | Step 28 VWin, proved by vwin_of_row from Row0b (steps 29–31) | The guide names these declarations |
| Rank count with labels, §III.8 | Step 35 Green | The guide names it as a hypothesis the final theorem depends on |
| The data and side conditions of (M) at one scale (§I.5; Part II §4; §III.5; §III.8) | Step 25 ChainTied (Formalism/SharpRestriction/GlobalSchurBounds/TieGlobal.lean:97) | Confirmed in content. RestSize (Formalism/SizeWindowBrackets/SChain.lean:53) asks \(|T|/n\ge\tfrac12x^{-\varepsilon/8}\), \(|T_2|/n\ge x^{-\varepsilon}\), \(m\ge x^{1/2-\varepsilon}\) and \(a\le 2xb\), the sizes in §I.5 and Part II §4. GrowthInequalitySizes (line 38) asks \(|a_t/b-x|\le1\). TieGlobal fixes \(\gamma\) and \(L'/L(1,\chi_d)\), the constants in the main term \(M\) (I.26). The weights w u0 j are indexed by translate and repetition, as in \(P_{u,t,j}\), and §III.8 describes the same per-translate records. ChainTied holds the hypotheses, not (M) itself. |
| Auxiliary primes and the two scales: Estimate B, Theorem III.1.1, Estimate C, Theorem III.2.1 | Step 27 HostExW (Development/AbelianVarietyTheory/Definitions/GlobalTwisting/ThinScalesIffHostBound.lean:120) | Confirmed in content. It quantifies over \(0\lt\varepsilon\lt\tfrac1{52}\). PairRow (PairRowIffQdRows.lean:68) asks \(q\equiv3\bmod8\) and \(d\equiv7\bmod16\) prime, \(\ln q\) near \(Y^{\varepsilon/5}\), and every prime up to \(Y^{\varepsilon/5}\) split: Theorem III.1.1(a)–(c). QDq is Estimate B and BoxH excludes real zeros above \(1-\tfrac1{400}\) (III.1.1(e)). ThinOK (line 61) is the conclusion of Estimate C almost word for word: \(|G_\chi(y)|\le y\,e^{-c x^{\varepsilon/8}}\) for \(y\ge\tfrac13e^{x^{\varepsilon/3}}\). |
| The construction step itself (§I.7, "not begun") | Step 18 TopG (Development/AutomorphicArithmetic/CharacterBoxes/ZeroFreeAllOfTopG.lean:12) | Partial match: the interface only. Its inputs are HostExW and HypV; by §III.9, HypV is the form in which a family receives the Galois-cohomology bound (DivinNOfHypV.lean:10). Its output is families satisfying ChainTied at every large scale. So it states the construction of (M)'s data, which the guide quotes and does not explain. |
| (M), B, C \(\Rightarrow\) A \(\Rightarrow\) bound (Part II §4–5) | Step 24 level_gm_le_on_family_scales | Same shape: it turns ChainTied into the windowed bound (LevelGmLeOnFamilyScales.lean:28). We did not compare the proofs line by line. |
| Theorem G at a shifted scale, twists below \(x^{401}\), window \([Y,Y^2]\to[Z,Z^{3/2}]\) (§III.9) | Step 26 supplyChi: VWinAt ψ 401 → HostExW (3/2) … | Matches the guide's numbers 401 and 3/2 |
What is still unclear
- (M) is not proved by hand. The guide proves RH from (M), B and C, proves B and C, and proves some steps toward (M). The construction of the maps and weights and the assembly of (M) "are not begun" (§I.7).
- The \(\sqrt X\) iteration: now located. The guide explains why \(m\approx\sqrt x\) (§I.8). In Lean it is fixed by
RankVal(PairRowIffQdRows.lean:28–32, a condition insideHostExW, and cited by the guide in §III.9): \(\sqrt{x\delta_1}\le m\le2\sqrt{x\delta_1}\) with \(\delta_1=x^{-\varepsilon/8}\) (DataParams₂.lean:45).RestSizeinChainTiedalso asks \(m\ge x^{1/2-\varepsilon}\). What we still can't explain is how the \(m\) lifts are built (insideTopG). - The sixfold cover of \(\mathrm{GL}_5\). The guide found "no trace" of it, or of the identity \(qN=Mk+R\), in the Lean development; it calls this an earlier formulation (§I.9).
TopG(step 18) matches the guide only at its interface: it states that the construction exists, and that construction is exactly the part of (M) the guide quotes without proof.- Declaration counts differ by source: the guide says the Lean theorem
riemannHypothesisdepends on 1,478,198 declarations (Part III); the independent con-leche check reported 1,468,418 declarations. They measure different things, and we don't reconcile them.