And now, one from the archives.
In 1989, then MIT grad student Lance Fortnow (he’s now chair of the computer science department at Georgia Tech) wrote a mathematical proof and published it as conference proceedings. He later went to publish the proof in a journal.
But he then discovered “unexpected technical challenges” and published a retraction in 1997. Both are still available on his personal website.
Not everyone would be that transparent. We reached out to ask why he left them up for people to see. He gave us his rationale:
Part of the paper still has correct results and even though we couldn’t get the proof to work, the techniques we develop could prove useful in the future.
More importantly, you don’t retract a paper, you retract the results within. I’m dedicated to put all my published papers online, warts and all.
The vast majority of my papers do have, to the best of my beliefs, correct proofs. But four mistakes is enough to gain a reputation and means you should not trust my proofs on face value. I have also learned this lesson and try to get reliable people to check over my proofs before I make them public.
Fortnow was a Fulbright Scholar, as well as the winner of the 2014 Nerode Prize, a prestigious computer science award, so he seems to be doing alright — which is consistent with research on what happens when scientists come forward about their errors.
This is the notice, which ran in the Proceedings of the twenty-ninth annual ACM symposium on Theory of computing:
In the paper “Probabilistic Computation and Linear Time” [FS89] we claimed a proof of the following statement as out main result.
Question 1 There exists an oracle A such that
However technical problems have developed in the details of the proof. After unsuccessful attempts to repair the construction, we must retract our claim to a proof of Question 1.
Question 1 remains open.
Even assuming that the BPP machine has only nonadaptive access to the oracle there are subtle problems in the way we combine machines. Berg and Hastad [BH97] give a detailed analysis of the problem.
Related Problems
The proofs of Theorems 1 and 2 in Section 6 of [FS89] remain correct.
The Fortnow-Sipser paper [FS89] was motivated by the problem of establishing a time hierarchy on interactive proof systems. Such a hierarchy follows immediately from the series of results showing IP = PSPACE [LFKN92, Sha92].
Fortnow and Lund [FL93] have shown a strong time and space hierarchy for public-coin interactive proof systems.
The hierarchy question still remains open for bounded-round interactive proof systems. Also, hierarchy theorums remain open for other classes without complete sets like the time-bounded versions of UP and
Hat tip: Rolf Degen
I’d tend to agree.
Research grows on what came before. What came before is often wrong, and/or needs tweaking or elucidation to improve it. It needs to be built on, and not removed from our archives of knowledge. If we were to go back and retract all articles that are now deemed incorrect because newer research found out something they didn’t know back then, we’d be deleting half of the history of science.
Dishonesty? Retract! Fake data? Retract!
Legitimate research is published, but we forgot to do a certain form of analysis that debunks our conclusions – and our publication is recent enough – retract or correct.
Find out 8 years later that it was wrong based on further analysis that (I assume) was done based on discussion it generated? The original article was not a fraud, it was just mistaken, and is a legitimate development along the road of science? No retraction needed, imho.
So publish something and discover it’s wrong based on further studies? I say we should not retract, but do as this guy says.
I agree but disagree in some circumstances. When science is growing, all researcher get idea how the experiment was done or falsified in earlier published manuscript. For that it will take even decades to give justice to the reader. Some manuscript has a crystal clear shape of lies how they are presented and need a lot of time to understand either the data is accurate or not. So bulking literature in web always confuse reader and has no sense, thus need to be taken out from public domain.