On November 20, 2026, Levent Alpöge released a proof of the Riemann Hypothesis, written with Claude, as a Lean library of about 17 million lines. A day later the Lean and Mathlib communities certified both its statement and its proof. Clay's prize still needs publication and time, but as a machine-checked result the question is answered: every nontrivial zero of ζ(s) lies on the line Re(s) = ½.
A proof is a finite object. Lean does not walk the zeros one by one. It checks that every step follows from the last. So this post is not a complaint about size. It asks a different question, one a proof does not answer by itself:
If a zero had left the line, how long would it take to notice?
The twin
Map each zero ρ through w = (ρ − 1)/ρ. On the critical line |w| = 1 exactly. The functional equation pairs a zero at σ + it with one at (1 − σ) + it, and their images obey
|w| · |w′| = 1
So if a zero sits at σ > ½, its image falls inside the unit circle and its twin falls outside. In Li's criterion, λₙ = Σ [1 − (1 − 1/ρ)ⁿ], the inside one fades and the outside one grows like |w′|ⁿ. For large t,
log|w′| ≈ Δ / t², with Δ = σ − ½.
The twin's growth eventually pushes some λₙ below zero. That negative value is the signature of an off-line zero.
The cost
Growth of Δ/t² per step is very slow, and the twin has to beat something: the on-line zeros, whose combined contribution itself grows as (n/2)(ln n − 1 + γ − ln 2π). Setting the twin equal to that background gives the threshold:
n_crit · Δ / t² ≈ ln n_crit + ln(ln n_crit + c) − ln 4
So the number of Li coefficients you need scales as t² log t / Δ. The t²/Δ part was known (Oesterlé 2000; Voros 2017). The log factor matters, because it means the "constant" in front is not constant. An earlier version of my own paper got this wrong: it claimed a universal value of 29.93. That value fails at every height I tested.
Testing it on real zeros
I built λₙ from the first 200 actual zeros of ζ plus the known density of zeros above them. I checked it against the exact λ₁…λ₅, then inserted one hypothetical off-line pair and scanned for the first negative λₙ:
| σ | t | first λₙ < 0 | n·Δ/t² (actual) | formula above |
|---|---|---|---|---|
| 0.60 | 10 | 9,809 | 9.81 | 9.76 |
| 0.60 | 30 | 112,538 | 12.50 | 12.49 |
| 0.55 | 30 | 239,788 | 13.32 | 13.32 |
| 0.60 | 100 | 1,534,362 | 15.34 | 15.34 |
At t = 10¹², the same formula needs about 6 × 10²⁶ coefficients.
The core of the model fits in a few lines:
import math G = 0.5772156649015329 def background(n): # growth of lambda_n from on-line zeros return n/2*(math.log(n) - 1 + G - math.log(2*math.pi)) def n_crit(t, D): # first n where the twin wins L = 0.5*math.log1p(2*D/(t*t + 0.25 - D + D*D)) # exact log|w'| n = 10.0 while 2*math.expm1(n*L) <= background(n): n *= 1.001 return n for t in (10, 30, 100, 1000): n = n_crit(t, 0.1) print(f"t={t:5} n_crit={n:14.0f} I={n*0.1/t**2:.2f}") What the cost is a property of
Here is the part I find interesting. Nothing in that threshold uses the primes. It depends on two things only: the pairing (|w|·|w′| = 1) and how fast the background grows. For L-functions of conductor q, the cost barely moves. On real zeros of L(s, χ₋₄) the shift in n·Δ/t² is about +0.15, not ln 4.
A function whose zeros really do leave the line
The off-line zeros above were inserted by hand. Some functions don't need that. Davenport and Heilbronn built Dirichlet series with the same kind of functional equation as ζ but no Euler product, and their zeros genuinely wander off the line. I took the partner function mod 5. Its strongest off-line zero sits at 2.3086 + 8.9184i.
Before computing anything exact, I wrote down the prediction: the background for conductor 5, plus the twins of its off-line zeros. Then I computed its Li coefficients exactly, from Taylor coefficients of log Λ around s = 1. That method uses no zeros at all, and two independent runs agree to 10⁻⁸.
| first λₙ < 0 | |
|---|---|
| exact | 285 |
| prediction | 285 |
| old "universal constant" formula | 1,316 |
No primes in the input, a real off-line zero, and the crossing lands exactly where the pairing says it should.
So the cost of seeing is geometry, and the reason nothing is there to see is arithmetic. The Davenport–Heilbronn function has the geometry without the arithmetic, and its zeros leave the line. The proof settles the second. The threshold describes the first: how far out a twin would have to be, at a given height, before a sign test could see it.
Both kinds of result belong in mathematics. A proof tells you that. A measured invariant tells you how hard it is to look.
Paper, code and data: 10.5281/zenodo.24447722
- bitsabhi •/>

