· Interactive explainer

Waves and Theta Lifts

The figures are computed live in your browser from real data: the first 300 zeros of the zeta function, the primes up to a million, and actual lattices. Two diagrams in section 6 are schematics and say so. Section 6 also shows the shape of the new claim. Its Lean code has passed independent machine checks, but no human-readable proof has been published.

A companion to Levent Alpöge's note of November 20, 2026, which claims a proof, found with Claude, of the Riemann Hypothesis and of its generalization to Dirichlet L-functions, cousins of ζ that track primes in arithmetic progressions. Sections 1 to 5 explain established mathematics. Section 6 marks the new claim separately.

The short version

Give each prime p a weight, its logarithm ln p, and add up the weights of the primes (and prime powers) up to some number x. The total comes out almost exactly x. The Riemann Hypothesis (RH) is about the size of the error: it says the error is about √x in size, never bigger than √x times a slowly growing factor. For every x up to a million it stays under a thousand.

A formula going back to Riemann (1859) writes that error as a sum of waves, one for each zero of the zeta function ζ. The "trivial" zeros at −2, −4, −6, … give waves that die away, so set them aside. Every other zero gives a wave that grows like √x exactly when the zero sits on one particular line. A zero off the line always comes paired with a mirror-image zero, and one of the two makes a wave that grows faster than √x. So RH says every one of these zeros is on that line. Sections 1 to 3 let you watch the waves rebuild the primes.

RH has a sibling that is proved. Count the solutions of a curve's equation in clock arithmetic (arithmetic mod a prime), and the error is again about a square root: Hasse proved it in 1933 for the equations in section 6, and Weil in 1948 for all curves. One proof, which Stepanov began in 1969 and Bombieri extended to all curves in 1973, builds a polynomial that is zero many times over at every solution. The new note claims to carry that idea over to the whole numbers, using tools called theta lifts (sections 4 and 5). Section 6 puts the two side by side.

The claimed proof has been published so far only as code in Lean, a language whose proofs a computer can check. Independent checkers accept it, but a written account that people can read and check is still to come.

The prime staircase, rebuilt from waves

Walk along the number line and step up by ln p each time you reach a prime p or a power of a prime (4, 8, 9, 16, …; at 8 = 2³ the step is ln 2). The result is a staircase called ψ(x). It climbs at slope about 1, so ψ(x) ≈ x. That is the prime number theorem.

Von Mangoldt's explicit formula says that for x > 1 the staircase equals the straight line x, plus one wave for each nontrivial zero ρ of the zeta function, plus two small corrections:

ψ(x) = x − Σρ xρ/ρ − ln 2π − ½ ln(1 − x−2)

The zeros come in conjugate pairs, and every pair checked so far has the form ½ ± iγ. Each pair adds one real wave. At a jump, the formula gives the midpoint of the step. Add the waves one pair at a time, in order of height, and the straight line bends into the steps. The order matters, because the sum converges only conditionally: the formula adds each zero together with its conjugate, lowest pairs first.

25
ψ(x), the prime staircase x, the two corrections, and the waves from the first N zeros the line y = x
Move across the chart to compare values.
The zeros on the line of real part ½, by height γ (used ones in blue)
Each zero, with its conjugate, adds one wave. With no zeros you get only the corrections: a curve about 1.84 (that is, ln 2π) below the line y = x. A few dozen waves already trace the big steps. With all 300, the curve stays within about 0.4 of the staircase for x up to 50, except within a quarter unit of each jump, where it ripples (the Gibbs effect of Fourier series). At each jump it passes close to the midpoint of the step.

Linear-algebra lens. Think of a Fourier series, which writes a vector as a combination of waves. Here each zero supplies one frequency, its height γ, and the same zero also fixes that wave's coefficient, −1/ρ. Nothing else enters. So the list of zeros plays the role of a spectrum: the frequencies the primes are made of.

What "real part ½" controls

Take one zero ρ = σ + iγ. Its wave is xρ/ρ = xσ · eiγ ln x / ρ, and two separate things happen in it.

The height γ sets the frequency. The wave oscillates like cos(γ ln x), so its crests are evenly spaced on a logarithmic axis. The real part σ sets the growth: the wave's size grows like xσ.

RH says σ = ½ for every nontrivial zero, so every wave grows like √x. (The trivial zeros at −2, −4, −6, … only produce the small ½ ln(1 − x−2) term.) Drag σ to the right to see what a zero off the line would do.

0.50
wave of the first zero, γ = 14.13 (on the line) the same wave moved off the line (hypothetical) size allowed by RH, ∝ √x
The x-axis is logarithmic, so the crests are evenly spaced. On the line, the wave stays inside the shaded √x band forever. Off the line, it grows like xσ and leaves the band. A zero off the line would come with a mirror partner with real part 1 − σ (at 1 − σ + iγ), and the one with σ > ½ dominates.

Here is the real error, measured: ψ(x) − x divided by √x, for every x from 11 to a million. If RH holds, this ratio can grow at most like (ln x)², which is very slow. If any zero had σ > ½, the ratio would reach size at least about xσ − ½ again and again, a power of x.

(ψ(x) − x) / √x, the actual primes ±1 what moving the first zero to your σ would add
Sieving the primes up to 1,000,000…
Each vertical sliver spans the lowest and highest value of the ratio over that stretch of x. Up to a million it hovers in a band of roughly constant width. Littlewood proved that it does creep outward forever: it gets past ±c·ln ln ln x again and again, for some fixed c > 0. The true rate is believed to be about (ln ln ln x)². Both grow far too slowly to see here. The magenta line uses the σ slider above.

Zeros and primes are a Fourier pair

In Fourier analysis a signal and its spectrum determine each other. Primes and zeros behave the same way, as long as you measure primes on a log scale, through ln p.

Top chart: add up cos(γ ln x) over the zeros, then flip the sign and scale as in the legend. The waves cancel almost everywhere and pile up at the prime powers, where the spikes approach height ln p as you use more zeros. Bottom chart: add up cos(t ln n) over the primes and prime powers, each weighted by ln p/√n, and flip the sign. That sum peaks at the heights γ of the zeros. The weights and the sign matter: the plain sum of cosines does not show the zeros.

Zeros → primes
150
−(2π√x / T) Σ cos(γ ln x), built only from zeros ln p at each prime power pk
Primes → zeros
10,000
−Σ ln p · p−k/2 cos(t ln pk), built only from primes true heights γ of the zeros
Top: T is the height of the last zero used, and the factor 2π√x/T comes from Landau's formula, which predicts spikes of height ln p in the limit of many zeros. With finitely many they fall short, more so at larger x. Bottom: the sum uses a triangular taper (1 − ln n / ln C) in ln n, where C is the "Primes up to" setting, whose Fourier transform is never negative, and subtracts a bump that ζ's pole at s = 1 would add near t = 0. More primes give sharper peaks.

The precise statement. Weil's explicit formula has the same shape as the Poisson summation in section 4. Summing a test function over the heights γ equals summing its Fourier transform over the numbers ln pk, each weighted by ln p/√pk, plus a term from ζ's pole at s = 1 and one from its gamma factor. Knowing all the primes pins down all the zeros, and the reverse.

A theta series turns a matrix into a function with symmetries

Theta lifts are built from theta series, and a theta series starts from something you know: a positive-definite symmetric matrix A. It defines a quadratic form Q(v) = vᵀAv. The level sets Q = n are ellipses, and their axes point along the eigenvectors of A.

Count the integer points on each ellipse: r(n) is the number of integer vectors v with Q(v) = n. Package all the counts into one function, the theta series θA = Σ r(n) qⁿ, a power series in a variable q. Hover a bar, or focus the chart and use the arrow keys, to light up its ellipse.

1
0
1
Integer lattice and the ellipses Q(v) = n
Counts r(n): the coefficients of θA
Start with A = I: Q = x² + y², and r(n) counts the ways to write n as a sum of two squares (r(5) = 8, from (±1, ±2) and (±2, ±1)). The dashed gray lines are the eigenvector axes of A. (For A = I every direction is an eigenvector, and the figure draws the coordinate axes.)

Now the hidden symmetry. Put q = e−πt, so θA(t) = Σv e−πt·vᵀAv: the values of one Gaussian, added up over all the integer points v. The Fourier transform of a Gaussian is again a Gaussian, with A replaced by A−1, t by 1/t, and a factor 1/(t√det A) in front (in two dimensions). Poisson summation says that summing a function over a lattice equals summing its Fourier transform over the dual lattice. Together they give

θA(t) = θA−1(1/t) / (t · √det A)

The inverse matrix, the determinant, and the flip t ↦ 1/t all appear. The two sides below are computed independently, one from A at t and one from A−1 at 1/t.

θA(t), summed over the lattice θA−1(1/t) / (t √det A), summed over the dual lattice
Change a, b, c above and the two sides still agree, to about 14 digits. This flip relates θA to its partner θA−1. Together with a simpler shift symmetry that comes from the exponents being whole numbers, it makes θA a modular form (of weight 1, for a smaller group of substitutions that depends on det A, and up to a sign-like factor called a character), the simplest kind of automorphic form: a function that changes only by a simple factor under a large group of substitutions. Automorphic forms are central to the note's argument.

A theta lift is a matrix

A theta lift is a linear map. Its input is a function on one symmetry group, its output is a function on another, and its matrix entries Θ(g, h) come from a theta series:

(θf)(h) = ∫ Θ(g, h) f(g) dg   compare   (Mf)i = Σj Mij fj

In the simplest genuine example the input side is a finite set, so the integral is literally a matrix. The set is finite because the forms are positive definite: their symmetry group is compact, so its functions live on finitely many points, the classes of lattices. The three forms below are all the positive-definite forms ax² + bxy + cy² with b² − 4ac = −23, counted up to changes of basis with determinant +1. (Here b is the whole xy coefficient, and a form and its mirror image count separately.) A function on this set is three numbers f₁, f₂, f₃. The lift sends it to Σ fᵢ θᵢ / |Aut(Lᵢ)|, where |Aut(Lᵢ)| counts the determinant +1 symmetries of the i-th lattice Lᵢ. Each has exactly two, v ↦ ±v, so |Aut| = 2. The columns of the matrix are theta series.

Input f:
x² + xy + 6y²
1
2x² + xy + 3y²
−0.5
2x² − xy + 3y²
−0.5
Lift matrix (rows n = 0 … 24, entries ri(n)/2) × f = coefficients of the output form

The claimed proof's strategy, next to a sibling that is already proved

RH has a sibling that is a theorem. Swap the whole numbers for arithmetic mod a prime p, and swap "count the primes" for "count the points on a curve". The analogue of RH then says the count is off by at most about a square root, and it was proved: by Hasse for elliptic curves in 1933, and by Weil for all curves in 1948. Starting in 1969, Stepanov found an elementary proof for curves like these. The note calls its own argument "a descendant of Stepanov's", carried over to the whole numbers.

So the most honest picture of the claimed proof is a side-by-side. Each row below puts one step of the finite-field story (left, proved, computed live apart from the schematic in step 5) next to the same step for the whole numbers (right). Read across a row to compare, or down a column for one story. Each row has a "More precisely" note for readers who want the exact statement. Steps 4 and 5 on the right are the new part, and they are the claim.

Count something, and compare with what you expect

Take the curve y² = x³ + Ax + B and count its points with coordinates mod p (whole numbers 0 to p − 1, arithmetic mod p), plus one extra "point at infinity". Call the count Np. Averaged over x, the equation has one solution y: about half the values of x³ + Ax + B are squares mod p, and each square has two square roots. So expect about p + 1 points.

Count the primes and prime powers up to x, each pk weighted by ln p. That is ψ(x), the staircase of section 1, and you expect about x.

The error is at most about a square root

Hasse proved |Np − (p + 1)| ≤ 2√p for every prime at which the curve stays smooth. Each dot is one prime, with its error divided by √p, so the bound is the fixed band ±2. Change the curve: the dots never leave it.

1
1
(Np − (p + 1)) / √p, one dot per prime Hasse's bound ±2

RH is equivalent to a bound of the same shape, with a slowly growing factor: |ψ(x) − x| ≤ √x (ln x)² / 8π for every x from 74 on (Schoenfeld, 1976). Here is the error divided by √x, for every x from 74 to a million, inside that band. Up to a million this is a computation; for all x it is exactly as hard as RH.

x from 74 to 1,000,000 (log scale)
(ψ(x) − x) / √x, its full range RH's bound ±(ln x)² / 8π
Sieving the primes up to 1,000,000…
More precisely

Schoenfeld proved that RH implies |ψ(x) − x| < √x (ln x)² / 8π for all x ≥ 73.2. Conversely, any bound of the form |ψ(x) − x| ≤ C x½ (ln x)k forces every zero onto the line, so the bound and RH stand or fall together. The band is loose for large x: up to a million the error stays within ±0.81 √x, while the band's half-width grows to 7.6 √x.

A square-root error is the same thing as zeros on a line

For each prime the counts build a zeta function for the curve. Its zeros come from two complex numbers α and β with α + β = p + 1 − Np and αβ = p. Hasse's bound says exactly that both have size √p, so α/√p lies on the unit circle. Each dot is α/√p or its conjugate β/√p for one prime.

α/√p and β/√p, primes up to 3000 a pushed pair (slider below)

For ζ, every zero found so far, apart from the trivial ones, has the form ½ + iγ, as in sections 1 and 2. RH says every nontrivial zero lies on that line, where the real part is ½. Here are the first ones, at their true heights.

zeros ½ + iγ a pushed pair (slider below)

The circle and the line are one condition in two coordinates. For the curve, the zeros in s are where ps equals α or β, and the map s ↦ ps − ½ wraps the line "real part ½" around the unit circle. A zero off the line by δ lands off the circle, at distance pδ from the center, and its mirror partner at p−δ. Push a zero off the line and watch both pictures break the same way.

0.00
More precisely

The parallel with section 1 is exact. Over the field with pn elements the curve has pn + 1 − αn − βn points: a main term minus one term per zero, the same shape as the explicit formula that writes ψ(x) − x as a sum over the zeros of ζ. Since |α| = |β| = √p, each term has size pn/2, the square root of the main term.

The curve's zeta function is Z(T) = (1 − apT + pT²) / ((1 − T)(1 − pT)) with T = p−s and ap = p + 1 − Np, so its zeros are where ps is a root α or β of z² − apz + p. Hasse's bound says ap² ≤ 4p, which makes the roots complex conjugates with |α|² = αβ = p. For an elliptic curve that is the whole story: conjugates always have equal size, so an off-circle pair would have to be a real pair r√p, √p/r, which happens exactly when |ap| > 2√p. A curve of genus g has 2g roots, and for g ≥ 2 a complex pair off the circle, as drawn here, is possible a priori, and only Weil's theorem rules it out. The symmetry s ↔ 1 − s of each zeta function, combined with complex conjugation, is what pairs a zero with its mirror partner.

The key move: force a polynomial to vanish to high order

Stepanov's trick: build a polynomial that vanishes to high order m at every point being counted, meaning it and its first m − 1 derivatives are zero there. A nonzero polynomial of degree d has at most d roots counted with multiplicity, so (number of points) × m ≤ d. Too many points would force too high a degree.

2
At m = 1 the graph crosses the axis. At higher m it flattens and lingers at each root: that is what vanishing to high order looks like.

The whole numbers have nothing like this: there is no variable to differentiate, so Stepanov's way of forcing high-order vanishing has no direct counterpart. In the note's words, multiplicity is "hard to find over the integers, corresponding to something like differentiating whole numbers".

The note's answer is "chaining theta lifts", the matrices of section 5, with "each mimicking multiplicity". A rough timeline in the note, which Claude compiled and which the author passes on with the warning that parts of it "might make zero sense", puts the chain between totally definite unitary groups U0 → U1 → … → Um, winning multiplicity m. This is the new step, and no written account of how it works has been published.

More precisely

Over a finite field the derivatives are Hasse derivatives, and the polynomial exists by rank–nullity: the vanishing conditions are linear equations in its coefficients, with fewer equations than unknowns, and a short lemma shows the result is not the zero polynomial. What makes it cheap is a quirk of arithmetic mod p. For a field of q elements, q is a power of p, so the derivative of xq is q xq−1 = 0. Differentiation treats xq like a constant, while on the field's own elements xq = x. Polynomials built from x and xq can therefore vanish to high order at all the points without their degree blowing up. A unitary group here is a group of matrices, with entries in a field such as ℚ(i), that preserve a Hermitian form: the complex cousin of the rotations that preserve the quadratic forms of section 4. "Totally definite" means the form stays positive-definite under every embedding of the base field into the real numbers, like the positive-definite forms of sections 4 and 5. That makes the group compact at infinity, so at each level its automorphic forms live on a finite set, which is why a theta lift between two such groups is a finite matrix, like the input side of section 5's matrix.

Balancing two costs gives the square root

More multiplicity shrinks one error term but raises the degree, which grows the other. With constants dropped, for a field of q elements (q is p or a power of p, unrelated to the q of section 4) the error behaves like q/m + m. It is smallest at m = √q, where it has size about √q: that balance is where the square root comes from.

8
Schematic, constants dropped (q = 400)
error left by low multiplicity, q/m cost of the higher degree, m total

For a sum over the primes up to X, such as ψ(X) − X, the note says Claude proposed a chain of about √X lifts: the counterpart of m ≈ √q. The note's timeline lists a "precise computation of the 'cost' of each lift" among its steps but gives no formula. The details are in the Lean code and no written account has been published, so there is no honest curve to draw here. The diagram shows only the shape of the claim.

10⁶
Each arrow is one theta lift, claimed to add one order of multiplicity.
More precisely

Stepanov's method on its own bounds the count from above, by q + O(√q) with a constant larger than 2. A bound from below follows from the same argument applied to a twisted curve, and the sharp constant 2 follows by running the bound over every extension field and using the curve's zeta function. Bombieri (1973) extended the method to all curves.

Where things stand

Proved. Hasse (1933) for elliptic curves, Weil (1948) for all curves over finite fields; Stepanov (1969–72) and Bombieri (1973) gave the elementary proofs sketched here. It is textbook material.

Machine-checked, not yet understood. The note publishes only Lean code and says a readable write-up is still to come. On November 20, 2026, Andrew Sutherland (MIT) reported that the code compiles, that its five main theorems use only Lean's three standard axioms, and that several independently written proof checkers accept it. The final end-to-end check (Comparator) was still running. A machine check confirms that the formal statements follow from the axioms; it does not explain the argument.

What the Lean code states. The code (mirrored by the Lean community) is 175,187 files. Its main file states five theorems. The first, riemannHypothesis : RiemannHypothesis, uses Mathlib's own definition (Mathlib is Lean's shared math library): every zero s of ζ other than the trivial zeros −2, −4, … has real part ½. The other four say the same for every Dirichlet L-function: no zeros with real part above ½, and every zero in the strip 0 < Re s < 1 on the line Re s = ½, both for all characters and for primitive ones. Those four are defined in a 16-line file that imports only Mathlib, so none of the proof's own code can change what they mean.

A map of the claim

One picture for three questions. What does the claimed proof stand on? What would it give if it holds? And what would still be open? The map reads left to right, or top to bottom on a phone. Pick a box to see what it says, where it comes from, and, where the post covers it, which section explains it.

Two columns need care. The second lists tools named in the rough timeline that Claude provided and the note passes on, with its author's warning that parts of it paraphrase a summary that "might make zero sense". It shows what the argument is said to use, and little about how. The column just right of the claim is conditional: each result there is already proved assuming RH or GRH, and would become a theorem if the claim holds.

Lines show what builds on what, and which questions lie beyond the claim. Picking a box highlights everything it builds on, everything that builds on it, and what lies just beyond it. Every count and bound in the boxes is checked by a test against an independent computation.

Check your picture

Places where an intuitive picture often drifts from the real one:

  1. The waves live in ln x. Each zero gives cos(γ ln x), so on an ordinary axis the crests spread out as x grows. They are evenly spaced only on a log axis.
  2. Height and real part do different jobs. The height γ is a frequency. The real part σ is a growth rate. RH is a statement about growth rates only: all of them equal ½.
  3. The zeros are not evenly spaced. After rescaling for their growing density, they appear to be spread like the eigenvalues of large random Hermitian matrices. Montgomery conjectured this for pairs of zeros in 1973, Odlyzko's computations bear it out for pairs and for gaps, and it is the best-known of the links to physics the note alludes to.
  4. A theta lift does not act on ζ itself. It maps between spaces of functions with symmetries, and L-functions come back out of what it produces. They appear as the coefficient series of the output (the Eisenstein series from section 5's constant input carries ζ(s) · L(s, χ), with χ the character attached to discriminant −23) or as integrals (periods) of the lifted functions, which is one reason the note talks about periods.
  5. The quadratic form is the hinge. One symmetric matrix gives a geometry (ellipses, eigen-axes) and, through Fourier duality A ↔ A−1, a function with symmetries. Theta lifts generalize this.
  6. Multiplicity means flatness at a root. The function and many of its derivatives vanish there. The claim is that a chain of lifts can stand in for derivatives the integers do not have.