Alpöge's note
One line from each of the four pages. Pick a page to bring it forward.
What’s claimed
On a Friday morning Alpöge posted four pages. The sentence everyone quoted reads: “I believe I have, working with Claude, discovered a proof of the Riemann Hypothesis.”
The Lean code proves the hypothesis as it’s defined in Mathlib, the shared math library Lean users build on. It also proves the same statement for every Dirichlet L-function, a family of functions closely related to zeta. That’s the slice of the Generalized Riemann Hypothesis it covers. Other L-functions aren’t included, and the Extended Riemann Hypothesis is one he says he expects “with more time”.
- Riemann hypothesisProved, as Mathlib defines it
- GRH for Dirichlet L-functionsProved, one of four further headline theorems
- Dedekind zeta and other L-functionsNot covered
- Extended Riemann hypothesis“I expect, with more time”
What a zero off the line would do
The earlier rooms checked zeros and found every one on the line. The proof says none can ever leave it. Each zero adds a wave to the prime count, and its real part sets how fast that wave grows. On the line, the wave grows like √x, about the wobble of a coin-flip count. Pull one off and its wave outgrows that, and the hundreds of published results that assume the hypothesis would rest on something false.
Drag zero 8 sideways, or focus it and use the arrow keys.
Real part 0.500. On the line. This zero’s term grows like √x and never outgrows it.
Both scales are logs. Your zero against where it belongs. This is one zero’s wave, not the full prime count, and no real zero has ever been found off the line.
How you trust a proof nobody can read
Lean is a programming language for proofs: you write every step, and a small checking program called the kernel refuses anything that doesn’t follow.
Nobody can read 17 million lines. People trust the small kernel and the statement at the top instead. The checkers ran the proof through two more kernels written separately by other people, in case one has a bug. A statement check then confirms the theorem at the top says what the hypothesis says. Here that statement is Mathlib’s definition, so the proof’s author didn’t get to write the question.
Andrew Sutherland at MIT, who ran most of the checks, told Scientific American that “the chance that anything is fishy here is approximately zero.”
All 175,187 files, zero errors (Lean v4.33.1, 20 Nov, 16:43 UTC)
Lean read every line and found nothing it couldn’t follow.
Reported by Sutherland, MIT
propext, Classical.choice, Quot.sound. Nothing else assumed.
The proof starts from the same three ground rules almost every Lean proof uses. It slips in no extra assumptions.
Reported by Sutherland, MIT
No sorry, admit, axiom, native_decide, unsafe or opaque in the source
Lean has a few ways to say "trust me" and skip a step. A scan found none of them.
Reported by Lean community checkers
Library files byte-identical to the official cache
The shared library of definitions, including the one that states the hypothesis, is the official one.
Reported by Sutherland, MIT
leanchecker: 175,187 of 175,187, zero rejections
Lean’s small core checker went back over every file from scratch.
Reported by Sutherland, MIT
Accepts the full export. 27 min 42 s on 80 threads.
A checker written separately, in Rust, by different people, agrees.
Reported by Sutherland, MIT
Accepts 1,468,418 declarations. 3 min 15 s on 64 threads.
A third independent checker agrees too.
Reported by Sutherland, MIT
Both say all five theorems match the independently written challenge
The theorems proved are the ones a separately written challenge asked for.
Reported by Sutherland, MIT
175,187 of 175,187 compiled files pass
A checker written in Lean itself went through every file and agrees.
Reported by Sutherland, MIT
Robin and Lagarias inequalities derived and accepted
The result also gives two known inequalities about sums of divisors, which mention no zeta at all. They check out.
Reported by Sutherland, MIT
Default and paranoid kernels: running, single-threaded, no verdict yet
Lean’s official checker, in its strictest mode, is still reading.
Reported by Sutherland, MIT
Running on Lean v4.36.0. Statement comparison and axiom check already passed in one run.
The official statement check, on the newest Lean, is the stamp on this object.
Reported by Sutherland, MIT
"The build is in progress."
A second team is building it from source on a separate machine.
Reported by Lean community checkers
Rebuilt on a separate machine with no network. The export matches byte for byte.
Built again from scratch in a locked-down machine, it came out the same down to the last byte.
Reported by Sutherland, MIT
No problem has been reported against the Lean proof. A pass shows the statements follow from the three standard axioms, checked by these kernels, against Mathlib’s definitions. It doesn’t show anyone understands why.
How big it is
The note is four pages. The proof behind it runs to about 17.4 million lines. Step through it from smallest to biggest.
What Alpöge posted on 20 November: an opening, a remark on timing, thanks and a timeline.
The proof nobody understands yet
“I am presenting the world with a complete argument before having mastered a massive neighbourhood around its mathematics.”
Levent Alpöge, the note, page 2
He writes that he intends “to understand and impart understanding of this argument and its certain simplifications, to speak about it, to write it properly.”
Terence Tao wrote that “the digestion process will take some time, though in this case at least we have a human author to communicate with to assist with this task.” Alex Kontorovich, speaking of himself, said: “Does that mean I understand it? Not yet.”
- Is it true?
- 1859Riemann thought it "very likely" and put it aside after "some fleeting superficial attempts".
- Zeros checked on the line
- 1859The first few, worked out by hand. Siegel found them in Riemann’s notes in 1932.
- Share of zeros proved on the line
- 1859None.
- Who understands the proof
- 1859There was no proof to understand.
- The prize
- 1859No prize. Hilbert put it on his list in 1900, Clay put $1 million on it in 2000.
The end of the tour
If the last checks come back clean, people stop asking whether it’s true and start asking why. I think that’s the fun part: people get to take a 17-million-line proof apart and look for the short one inside it. Alpöge’s own guess is that it “will be massively simplified, likely immediately.”
Thanks for walking the whole floor with me.
The reactions and the rival claim are in the Run Data Run Deep Dive.