On the afternoon of August 27, 2026, somebody began deleting Nicholas Polson’s academic career.
Andrew Gelman (b. 1965) watched it happen from his blog at Columbia. At 9:05 that morning he had posted about the Chicago Booth statistician’s SSRN author page, which carried 258 working papers posted in 2026 alone. By 3:32 PM the count had fallen to 99. Then 44. Then one, a paper called “Sequential Bayesian Pricing of AI Infrastructure” from May 29. Then the page was gone.
Ten hours, from a blog post to an erasure.
The story reached Gelman through the economist Jeremy Horpedahl, who had counted the papers and noticed that six of them went up on August 26. Horpedahl’s point was that these were not notes. They ran twenty-seven pages, thirty-two pages, fifty-eight pages. Read any one of them, he wrote, and it feels like a normal academic paper.
Gelman picked one to test that. “Theories of Human Connection”, eighty pages, written February 20 and posted March 11, by Polson (b. 1963) and Vadim Sokolov of George Mason University. Its preface assembles sixteen thinkers who each identified a dimension of human connection the others left in shadow: Gary Becker, Gregory Bateson, John von Neumann, Thomas Schelling, John Maynard Keynes, Desmond Morris, Diane Vaughan, Viktor Frankl, Abraham Maslow, John Gottman (b. 1942), Paramahansa Yogananda, Swami Vivekananda, Nisargadatta Maharaj, Polson, Katherine Woodward Thomas, and Gwyneth Paltrow (b. 1972).
The chapter that stopped Gelman is Chapter 10, titled “The Mathematical Proof: Gottman’s Model of Relational Dynamics.” It reports that Gottman’s models predict dissolution with over ninety percent accuracy and calls the result extraordinary. The abstract carries the same claim into the indexed record. Gelman has been correcting that number since 2010. It comes from fitting a model to the sample it was built on. Two Bayesian statisticians published it as evidence.
That much has been reported. What follows has not.
Go to Polson’s research page at Chicago Booth. Near the top, above the year-by-year listings, sits a line reading “RH: Hilbert 8.” It is a hyperlink. It goes to arXiv 1806.07964, a manuscript called “Riemann Hypothesis: a GGC factorisation.”
That document was submitted June 18, 2018. It went through seven versions between June and October of that year, each four or five kilobytes, which is to say two or three pages. The last revision landed October 15, 2018. Nothing since. It concludes that the denominator of a certain representation cannot vanish and therefore that all nontrivial zeros of the zeta function lie on the critical line.
It claims a proof of the Riemann hypothesis.
Eight years later his employer’s website still directs readers to it. The page carries no note about what happened to the companion paper.
Here is what happened to the companion paper. In August 2018 the Brazilian Journal of Probability and Statistics published Polson’s “On Hilbert’s 8th problem” at volume 32, number 3, pages 670 to 678. Received September 1, 2017. Accepted January 1, 2018. The journal then published a retraction. The notice says the paper was withdrawn by the author over an erratum in the application of Grosswald’s result on the existence of a certain function, and that the error invalidates the proof of Theorem 2. Emil Grosswald (1912-1989) was a number theorist at Temple. His theorem had been asked to establish something over a range where it did not apply.
Polson retracted it himself. Nobody forced him. That is a fact in his favor and it belongs in any honest account of this man.
The arXiv version of that paper tells the rest of the story. Fifteen revisions between August 2017 and July 2018, all five to seven kilobytes, a short note hammered at again and again. Then the retraction. Then version sixteen in April 2019 drops to three kilobytes, the sound of an argument being cut back to what survived. A slow rebuild through 2019 and 2020. Version twenty-one lands February 9, 2021.
Then silence for four years and four months.
Version twenty-two appears June 1, 2026. Version twenty-three appears July 4, 2026, at eleven kilobytes, more than half again the size of anything before it.
Both documents sit under Mathematics, General Mathematics. That is arXiv’s category for work its moderators decline to file as number theory. In the August 2017 listing Polson’s paper appears four entries above “Collatz Conjecture: Is It False?” and a few below “Magic Tours of Knight in 4x4x4 Cube.”
The current abstract of version twenty-three is careful in a way the 2018 note was not. It says the positive measure it constructs is the input the Thorin condition requires. An input. Not a conclusion.
But arXiv 1804.10043, “Riemann, Thorin, van Dantzig Pairs, Wald Couples and Hadamard Factorisation,” revised April 29, 2026, does claim the conclusion. Its Theorem 25 ends with the sentence that every nontrivial zero of zeta satisfies the real part equal to one half. Section 7 extends the machinery to Dirichlet and modular L-functions and to Birch and Swinnerton-Dyer. Section 9 lists the applications: the Riemann hypothesis, the generalized Riemann hypothesis, Birch and Swinnerton-Dyer.
Four months later comes “Thorin Quantisation”, written August 13 and posted August 15, eighteen pages, Polson alone, acknowledging conversations with Lennart Bondesson and Pierre Patie.
Read its Lemma 5.3. It considers what happens when you factor the completed zeta function through the pole of zeta, and shows that the linear factor contributes a term that looks like a Thorin atom at three quarters. Then it proves the atom cannot exist. The function is finite and nonzero at the point where an atom would force a pole. The apparent atom is an artifact of the factorization, cancelled by the pole of zeta at one.
Now read Remark 6.6 of the same paper. It observes that an exponential tilt at a certain parameter fails if one carries the spurious atom, because the component in question has no exponential moment there.
Then go back to arXiv 1804.10043 and read the proof of Theorem 23. It constructs precisely that atom. And the proof of Theorem 25, which is the theorem that concludes the Riemann hypothesis, performs precisely that tilt.
The August paper identifies both halves of the failure in the April paper. It does not cite Theorem 25. It does not use the word erratum. The two documents sit on two different servers under the same name, and nothing connects them.
This is the same species of error as 2018. An existence condition applied where it does not hold, invalidating the theorem that rests on it.
I am not a number theorist and I do not offer a verdict on the mathematics. I am reporting what two documents by the same author say about each other, and noting that both are public, both are checkable in an afternoon by anyone who works in this area, and nobody has checked.
“Thorin Quantisation” does serious work. It engages a 2022 paper by Konstantopoulos, Patie and Sarkar on the van Dantzig problem. It produces what it says are the first explicit hitting times of self-similar Markov processes that are self-decomposable and provably not generalized gamma convolutions. It reduces the Riemann hypothesis to positive semidefiniteness of Hankel matrices built from a computable moment sequence. Its Table 2 reports numbers computed from the Taylor coefficients of the completed zeta function at a single point, with no input from the zeros, whose ratios converge to 14.1349. The first zero ordinate is 14.134725.
Whoever produced that document was doing mathematics.
Which is why the other artifacts are hard to place.
Open “The Riemann Ξ-Function as a Characteristic Function”, dated August 13, 2026, thirty-one pages. Turn to page 30. In the bibliography, after an entry for a 2024 arXiv paper on primes in arithmetic progressions, sits a parenthesis reading: Author list to be checked.
The authors are Kaisa Matomäki, Jori Merikoski and Joni Teräväinen, and the arXiv record names all three. The reference also sorts wrong, sitting between Heath-Brown and Montgomery, which is where an entry with no author key lands.
Turn to page 31. Another entry cites a Polson companion manuscript and notes that its title and numbering are still to be finalized. A twenty-eight page paper answering that description, “Thorin Measures and the Reciprocal Xi Function”, went up eight days later.
Now go back to “Theories of Human Connection” and check its references. It cites Buunk, Dijkstra, Fetchenhauer and Kenrick, “Age and Gender Differences in Mate Selection Criteria for Various Involvement Levels,” Personal Relationships 15 (2), 2008, pages 271 to 278, first author given as Abraham P. Buunk.
The article exists. It ran in 2002, in volume 9, issue 3, at pages 271 to 278. The first author publishes as Bram P. Buunk.
Year wrong. Volume wrong. Issue wrong. Pages correct. Name expanded to a form that appears nowhere on the article. And the year and volume moved together, since volume 15 of that journal is 2008, so the false pair is internally consistent. A person miscopying a citation gets one field wrong. This reference was reconstructed.
The same eighty pages say fourteen thinkers in Chapter 1 and again in the conclusion, list sixteen in the preface, describe thirteen dimensions four times, and print a table of twelve. Polson appears in the document exactly twice: on the title page and in the preface roster. He is named as one of the minds who identified a dimension of human connection, and then no chapter discusses his contribution and no work of his appears in the bibliography.
None of the eight papers I examined carries a statement disclosing the use of generative AI. SSRN has required one since March 7, 2023, and tightened the language on August 3, 2026 to specify that material use must be disclosed in the PDF along with the tool and the extent of its contribution. I make no claim that any particular paper crossed that threshold. I report that no disclosure appears.
Two hundred fifty-eight papers in eight months is not a rate a person achieves. Something in this workflow was generating manuscripts. That much needs no forensics.
Everything past it does. Which model, operated by whom, with how much review, and whether every named coauthor agreed to appear are open questions, and the people who can answer them have not. Gelman offered four explanations, ranked by his own generosity: a joke, an impersonation, an attempt to stake priority claims cheaply, or an effort to teach. He landed on teaching. Commenters added delusion and an attempt to seed future training data. Kaiser Fung, who wrote it up separately, pointed at the ratio of citations to downloads.
I would add a fifth that requires no motive at all. The friction between having a thought and publishing it went to zero, and a man who already published outside the referee’s reach kept publishing. He has been posting Riemann work to General Mathematics since 2017 and to his own faculty page since before that. The disposition is old. What is new is the rate.
There is a version of this story that treats the corpus as slop and stops. That version cannot account for a paper that identifies a subtle error in the author’s own posted proof, or for a moment sequence that recovers the first zero of the zeta function from local data. Something here is doing competent work and shipping the error and its correction to different servers without ever noticing they are the same argument.
The institutional fact is simpler and it survives every possible answer to the rest.
Chicago Booth’s faculty page for the Robert Law, Jr. Professor of Econometrics and Statistics has listed nothing since 2018. Not one of the 258 appears on it. Its most recent entries are deep learning papers with Sokolov and a sparsity paper with Peter McCullagh. It links an SSRN profile holding sixteen papers, the most recent from 2018. And near the top, above all of it, sits a link labeled “RH: Hilbert 8,” pointing at a two-page note from 2018 that says the Riemann hypothesis is proved, through a retraction the university’s own professor requested, past a 2026 paper by the same man saying the condition remains open.
Two records, two servers, one name, and nobody minding either.
I wrote to Polson and to Sokolov on August 27.
Dear Professor Polson,
I write at lukeford.net. I have no credentials. I’m working on a piece about your SSRN output this year and about Theories of Human Connection in particular, coauthored with Vadim Sokolov, posted March 11, 2026. Andrew Gelman discussed it on his blog yesterday. I’m going further into the documents than he did, and I’d like your account before I publish.
Specific items:
Jeremy Horpedahl counts 258 papers posted to your SSRN author page in 2026 through August 26, 104 of them in August. Is that count accurate, and were all of them posted by you or with your authorization?
Was generative AI used to produce Theories of Human Connection or others among the 2026 papers? I find no AI disclosure in the PDF or on the abstract page, and SSRN requires disclosure of material use.
The abstract and Chapter 10 assert that Gottman’s models predict separation with over ninety percent accuracy. That number is a fit to the original sample, not a prediction, as you would know better than most. Do you stand behind it?
The preface lists sixteen thinkers, including yourself, between Nisargadatta Maharaj and Katherine Woodward Thomas. Chapter 1 says fourteen. Table 13.1 lists twelve. You do not appear in any chapter or in the references. Can you explain the inclusion and the discrepancies?
Two references appear corrupted. Buunk, Dijkstra, Fetchenhauer and Kenrick is given as 2008, Personal Relationships 15, pages 271 to 278; I believe the paper is 2002, volume 9, same pages. The Bhagavad Gita is listed as Anonymous, 1944, Penguin, London; the Penguin edition is Mascaró 1962 and the 1944 translation is Prabhavananda and Isherwood, Hollywood. Do you know how these arose?
Separately: arXiv 1804.10043, revised April 29, 2026, concludes at Theorem 25 that every non-trivial zero of zeta lies on the critical line, and the abstract says the Hardy and Ramanujan growth bound is verified using Hayman and Grosswald. You retracted On Hilbert’s 8th Problem from the Brazilian Journal of Probability and Statistics in 2018 over an error in applying Grosswald’s result. Has that error been repaired in the current version, and if so, where?
Dear Professor Sokolov,
I write at lukeford.net. I have no credentials. I’m working on a piece about the paper you coauthored with Nicholas Polson, Theories of Human Connection, SSRN 6278378, posted March 11, 2026.
Three questions:
Did you read the manuscript in full before it was posted?
Was generative AI used in its preparation? SSRN has required disclosure of material AI use since 2023, and I can’t find a disclosure statement in the PDF or on the abstract page. If one exists somewhere I’ve missed, I’d like to know where.
Chapter 10 and the abstract both say Gottman’s models predict separation with over ninety percent accuracy. That figure comes from fitting a model to the sample it was built on rather than from out-of-sample prediction, and it has been disputed for two decades. Do you stand behind the claim as published?
I’ve received no reply.
