Wolfram's 2,3 Turing Machine Not Universal

Fishbat writes "In a cutting message to the Foundations of Mathematics mailing list, Stanford's Vaughan Pratt has pointed out an elementary mistake in the recently announced proof that Wolfram's (2,3) machine is universal." Update: 10/30 04:18 GMT by KD : Ed Pegg Jr. from Wolfram Research points to this response to Dr. Pratt's note, which has been submitted to the FoM mailing list but has not yet appeared there due to moderation.

Ouch

[JoeShmoe950]

I'd hate to be involved in either the submission or the review of that proof. I was rather intriuged when the proof was first posted, but I must say, this is something of an embarrassment

Bad Headline

[EvanED]

Wolfram's 2,3 Turing Machine Not Universal

That's not, from my reading, what is true. What is true is that the proof is wrong, which means that it may not be universal, but reverts back to the unknown state.

Re:The Filter

[Anonymous Coward]

You misread the post. He said that if x + y = infinity and y is finite, then x must be infinity. This is TRUE for numbers. You cannot apply this by analogy to automata and think it is still true. It is not.

Title Proven To Be Misleading

[yerdaddie]

The Wolfram's 2,3 Turing Machine proof of universality was found to be flawed. This does *not mean* the 2,3 Turing Machine isn't universal. It just means it has not been proven to be universal. That would require another proof. Subtle distinction, I know; but in any case, the title of this article is fallacious.

No, just not proven

[billstewart]

As others [slashdot.org] have pointed out [slashdot.org], no, it doesn't mean anybody's proven that it's not universal, it just means that this alleged proof of universality was incorrect.

Back to the drawing board

[tepples]

Does this mean that it hasn't been proven to be universal (which is the case if there was just a bug in the proof)

It appears to mean the proof has a defect in the same way that message passing in Windows has a defect [wikipedia.org]. As far as I can tell based on the article, Vaughan Pratt alleges that so much of the proof relies on the defect that the proof would need a complete rewrite.

Re:duh

[ispeters]

You sound like a troll since you're so belligerent, but, in case anyone else here is legitimately wondering what it means for a Turing machine to be universal, I'll try to answer.

Basically, a Turing Machine [wikipedia.org] is an abstract "computer"--it's a tape (a skinny piece of paper) that has a start but no end (it's infinitely long, but it has a start), and a read/write head that can zip up and down the tape writing, reading, and erasing symbols on the tape. The Church-Turing Thesis [wikipedia.org] postulates that a computable algorithm is any algorithm that can be computed in a finite number of steps by a Turing Machine. There are some things that look like algorithms and seem like they should be computable but are in fact impossible. The classic example is the Halting Problem [wikipedia.org].

Anyway, a regular Turing Machine only computes one function--it's a single-purpose machine. A Universal Turing Machine [wikipedia.org] is a Turing Machine that can simulate any other Turing Machine by interpreting a codified description of the other machine. Since every computable function is isomorphic to some Turing Machine and every Turing Machine can be simulated by a Universal Turing Machine, every computable function can be computed by a Universal Turing Machine. The computer you're using to read this is an approximation to a Universal Turing Machine (the RAM would have to be infinite in size to be a proper Turing Machine), and the codified descriptions that it interprets are the binary executables that you run on it.

Hope that helps,

Ian

Re:The Filter

[jnana]

Indeed. A prior email in that thread -- by the same author, Pratt -- makes it very clear by giving the example of 2 pushdown automata [wikipedia.org] (PDA). A single PDA by itself is not universal, but the system comprised of 2 PDAs is universal, since each stack can represent one side of the Turing machine tape.

As Pratt states, the fallacy is of the following form: a system comprised of 2 PDAs, PDA A and PDA B, is universal. PDA A alone is not universal. Therefore, PDA B must be universal (because the system as a whole is universal). QED.

Of course, in the actual proof, it was not 2 PDAs, but a 2,3 machine and an encoder (i.e.,"PDA A" == "encoder" and "PDA B" == "2,3 machine").

Wolfram chimes in

[snark23]

Wolfram's lengthy response:

http://cs.nyu.edu/pipermail/fom/2007-October/012149.html [nyu.edu]

Re:The Filter

[Anonymous Coward]

Incidentally, for anyone who wants to learn something about automata and theory of computation and doesn't know where to start, I highly recommend the following book by Michael Sipser: Introduction to the Theory of Computation [amazon.com].

It's quite pricey for such a small book, but it's worth its weight in gold (i.e., the time you save by reading this little masterpiece instead of something else that's less well written). You can find the 1st edition used for much cheaper than the 2nd edition, and the differences between the two editions are pretty minor.

p.s. I have no connection to the book or the author. I'm just a very happy customer.

Elmentary Mistake in Title

[Cyberllama]

This doesn't mean it's not universal, just that it's not PROVEN that it is. Not at all the same thing.

Wrong, that is NOT Wolfram's response

[Weyoun]

Incorrect - that is not Wolfram's response to Pratt's message, it is a response to an earlier message. Compare the dates.

Serious authority

[Emnar]

I don't understand the math behind this argument and counter-argument, but Vaughan Pratt is a CS legend [wikipedia.org] and one of the early cofounders of Sun, to boot. You also might have run across his name in a cite or two in The Art of Computer Programming series by Donald Knuth. And if you don't care who Knuth is, then you probably don't care about this post at all.

I knew Pratt's daughter in college -- nice woman. Wrote her term papers in LaTeX, on a Linux workstation, in 1996 :P

Re:The Filter

[Anonymous Coward]

Prior to being accepted for publication, a paper is sent to (presumably) experts in the field to review the paper to make sure the research is worthy of publication. That's what is meant by "peer-reviewed" journals.

Re:Vaughn Pratt is confused

[Ed Pegg]

I'm posting from Wolfram Research. Basically, a message from Vaughan Pratt was posted to the correct spot, the FOM list. Dr. Pratt likely didn't expect his message to get a late night SlashDot level exposure. A response to his message has already been sent to the FOM list, but it is a moderated list, and the response is not visible yet. Here is a copy of the FOM posting from Todd Rowland, from the Wolfram prize committee. http://forum.wolframscience.com/showthread.php?s=&threadid=1472 [wolframscience.com] This is how math is done ... trying to poke holes in proofs.

Re:The Filter

[ais523]

This argument and Pratt's argument that the universality proof are incorrect both share the same flaw; see the response linked in the update of the original story for more details. Yes, a system consisting of two push-down automata that can communicate back and forth is universal. However, the proof of the universality of the 2,3 Turing machine in question doesn't set up a system where two systems can communicate back and forth; instead, the output of one is used as the input of the other, and there is no communication in the other direction. If one PDA is used as an encoder for another, the resulting system is not universal. The definition of universality in the original prize that was agreed on by the judges basically said that in the case of obviously simple encoders, if an encoder+system together are universal than the system itself is. The encoder used in the proof is not obviously simple (as it takes up several pages in the proof) but is clearly not doing any sort of universal calculation itself or engaging in part of such a calculation.

The real problem here is that with other definitions of 'universal', the system might indeed not be universal; the definitions are important in such cases. (For instance, a definition that restricted the initial condition to be finite and surrounded by nothing but yellow on the tape would result in the system being non-universal, but this definition is too restrictive to be useful.) Changing the definition of 'universal' is likely to result in changes to what is universal and what isn't, although it seems likely that the 2,3 Turing machine in question is well on the universal side of that edge (although not as far as some other systems). Even systems such as rule 110 that are generally accepted to be universal would be non-universal if a non-repeating finite initial condition was mandated by a definition of universality.

So in short, nothing has been disproven, the machine is still universal, and the headline is wrong.

Re:The Filter

[jnana]

I didn't understand Pratt to be arguing that the 2-PDA system is directly analogous to the proof system in the way that you state. What I understood Pratt's point to be was that the proof does not prove that the 2,3 machine must be universal, since it is entirely possible that while the encoding mechanism itself is not universal, it nevertheless could be sufficiently complex that universality does not (and perhaps cannot) exist without it.

The author of the proof states on page 26 (counting the first PDF page as 0) the following:

Therefore, this algorithm for determining an initial condition, whilst somewhat complicated, is definitely not itself universal, and so the universality discovered above is a property of system 0, and not of the algorithm used to find its initial condition.

I took Pratt to be arguing that this is not a valid inference, that one can't just say that because the encoding and creation of initial condition mechanism is not universal, while the system as a whole is universal, that the 2,3 machine is therefore universal.

I'm looking forward to learning more over the next few days, but that is what I meant by earlier post. I may have completely misinterpreted Prof. Pratt's criticism, but that's my take on it.

P.s. Prof. Davis stated on the FOM list that the committee has not actually ruled that the paper is acceptable yet. Apparently, people within Wolfram's organization were the ones who made the determination. Given that, I'm not sure that it's correct for you to state that "the machine is still universal", since the paper hasn't been peer-reviewed by even the expert peers on the committee, let alone the wider community, and so its universality is still far from established.

Re:Peer Review Rules

[rbarreira]

Every now and then I take a crack at P=NP, and sometimes, I feel like I've really got a good proof - a program idea, that, when implemented, could FACTOR fairly quickly.

Do you mean factor numbers? Even though it would be impressive to have an algorithm which factors numbers quickly, it wouldn't prove anything about the P=NP? problem. Factoring numbers is not known to be an NP-complete problem, so solving it in polynomial time doesn't automatically imply that P=NP.

Re:The Filter

[ais523]

Whether this is a valid inference depends on the definition of 'universal'; the problem here is that Pratt appears to be using a different definition from the one that the prize committee and the proof adopt. The inference may be incorrect with certain definitions of 'universal'; however, it is correct with other definitions. My point is that the counterexamples that Pratt gives to the inference are incorrect; that does not, of course, mean that the inference itself is correct, but it means that looking into where the problems with the counterexample came from is worth doing, and it appears to be a misconception between systems interacting backwards and forwards, and preprocessors. It seems to me that most general definitions of 'universal' will cause the inference in question to be correct (there is at least one definition which includes the inference as part of the definition itself, that is 'A system is universal if it can emulate any Turing machine if given an appropriate initial condition encoded using a non-universal system', although there is some circular reasoning involved in that definition).

Re:Vaughn Pratt is confused

[evgalois]

Well, that's a good start: you accepting you are not an expert in the field but pretending to judge something that of course experts reviewed. It is nice that people like you think on this problem and possible flaws, though. Of course nobody is changing the well stablished definition of a Turing machine. But it is accepted even by the experts that there is no clear definition on universality. You should follow all the FOM posts and not only those that you think are better to critic others. The definition of universality has been modify several times, the first to make a generalization exactly as the used by Alex Smith was Watanabe as early as in the 60's. Alex Smith contribution is a generalization of the same sort. If you want some references feel free to check these and then reply again: [1] Manfred Kudlek. Small deterministic Turing machines. Theoretical Computer Science, 168(2):241-255, November 1996. [2] Manfred Kudlek and Yurii Rogozhin. A universal Turing machine with 3 states and 9 symbols. In Werner Kuich, Grzegorz Rozenberg, and Arto Salomaa, editors, Developments in Language Theory (DLT) 2001, vol. 2295 of LNCS, pp. 311-318, Vienna, May 2002. Springer. [3] Maurice Margenstern and Liudmila Pavlotskaya. On the optimal number of instructions for universality of Turing machines connected with a finite automaton. International Journal of Algebra and Computation, 13(2):133-202, April 2003. [4] Claudio Baiocchi. Three small universal Turing machines. In Maurice Margenstern and Yurii Rogozhin, editors, Machines, Computations, and Universality (MCU), volume 2055 of LNCS, pp. 1-10, Chisinau Moldavia, May 2001. Springer. [5] Turlough Neary and Damien Woods. Four small universal Turing machines. Machines, Computations, and Universality (MCU), volume 4664 of LNCS, pp. 242-254, Orleans, France, September 2007. Springer. [6] Yurii Rogozhin. Small universal Turing machines. Theoretical Computer Science, 168(2):215-240, November 1996. [7] Shigeru Watanabe. 5-symbol 8-state and 5-symbol 6-state universal Turing machines. Journal of the ACM, 8(4):476-483, October 1961. [8] Shigeru Watanabe. 4-symbol 5-state universal Turing machines. Journal of Information Processing Society of Japan, 13(9):588-592, 1972. That is why many serious authors as Vaughn Pratt can be confused. Because they are not updated on the subtles of the field. We are glad that this is making a contribution to the discussion on universality. We expect that others will further clarify what Alex Smith has done.

Re:Vaughn Pratt is confused

[evgalois]

From the FOM list where most of you jumped to conclusions without even waiting to an official answer from Wolfram Research: 1. Concerning the prize committee, we did not expect them to go through the details of the proof but to support the reviewers on technical issues related to their fields of expertise. We would have been delighted to have had them spend more time on the proof but that wasn't our expectation. We did talk to people about how to set the prize up. Basically we needed judgement about their areas and we were very careful to cover those that could come up in a universality proof. We are grateful to the committee for covering those areas and we are happy that they were more active than many such committees. In the review process, committee members agreed with the Wolfram Science internal reviewers, while others asked for additional clarifications (such as fulfilling universality basic requirements for non-halting machines). All committee members had Alex Smith's proof in hand for a couple of months. But it was Wolfram staff who did the hard work to achieve a satisfactory level of verification (between revisions and revision requests), and we thought that by waiting any longer we would not have learned more. It helped that a lot of the proof was submitted in the form of programs that could be run and analyzed in a systematic way. Smith's proof is available to everybody, anyone can point out anything, and it is nice to get the sense that there are people out there who would be interested in verifying proofs. This is great news, and if we do other prizes we would definitely like to take advantage of that. We would be also delighted to hear opinions on how to set up next prizes. 2. Concerning the definition of universality, a halting state or halting instruction wasn't a requirement. This is a common usage nowadays in the field of small universal Turing machines (see detailed references below), which is a generalization of previous definitions. If there is no clear definition is because there is no clear-cut, established procedure to determine when an initial condition is computationally simple enough to be acceptable. Some would wish universal computation stick to a finite initial condition with an unbounded tape filled with "blanks", because that's the only case where the theory is entirely clear. However, others accept generalizations such as periodic "blank" words as long as they remain computationally simple enough (possibly generated in the same fashion as an unbounded "blank" tape). So Alex Smith's use of a non-periodic but still sufficiently computationally simple background is a natural generalization of this sort. The key point is that the background can be set up without doing universal computation, so the 2,3 machine itself actually carries out the computation. One could fairly argue that two non-universal coupled automata can reach universality, but it would require the couple to interact at every step, as in Margenstern's nice example [1], but that's very different from a non-universal second automaton intervening only once at the first step. We are glad that this is making a contribution to the discussion on universality. We expect that others will further clarify what Alex Smith has done. We particularly hope that his methods can be extended to other similar proofs. Accepting that you are not experts in the field but pretending to judge something that of course experts reviewed is a good start. It is nice that people like you think on this problem and possible flaws, though. Of course nobody is changing the well stablished definition of a Turing machine. But it is accepted even by the experts that there is no clear definition on universality. You should follow all the FOM posts and not only those that you think are better to criticise others. The definition of universality has been modify several times (if one can say so, since there is no original or first accepted definition), the first to make a generalization exactly as the used by Alex Smith was Watanabe as early as in the 60's. Alex Sm

Re:The Filter

[mightybaldking]

You are perfectly correct given the layman's imprecise use of mathematical language. That is, you understand the concept quite well, but are using the language imprecisely. x + y = infinity is not a valid statement. Nor is x = infinity. infinity is not a number. Infinite is an adjective that can be used to describe a number that you might think of a infinity. In other words, infinity has no value, and can't be "equaled" You would be more correct to say (X + Y) is infinite, or produces a non-finite result. Anybody doing number theory will use set theory and cardinality (look up a few posts) to be completely precise. Even in calculus, where infinity is used more frequently, we really mean "The value of X approaches the infinite"

Comment removed

[account_deleted]

Comment removed based on user account deletion

Re:wolfram inc. sinking deeper and deeper

[tlord]

Yes, exploited. It's a professional embarrasment at the start of his career. I'm not sure $25K is a fair price.

It's a basic logic problem. The original challenge problem could be restated: "Prove either that there exists a non-universal machine which emulates the 2-3 machine OR that the 2-3 machine can only be emulated by a universal machine." The proof does neither. Wolfram Inc. reps have come back with "Well, perhaps we should change the definition of universality!" Only, they aren't very concrete in offering an alternative with any rigor and the vague suggestions they are making don't add up (e.g., don't answer Vaughn Pratt's counter-example of paired push-down automata).

What the student proved here may turn out to be an interesting and useful result (not the universality of 2-3 but the universality of this interesting combination of machines). Students should be encouraged to work on such problems. Students should be encouraged to write as well as this student is learning to do -- it's a nicely presented paper (trivial formatting bugs aside). But students shouldn't be encouraged to go in those directions on false premises.

At least two interesting questions come from the students work. To Wolfram Inc.'s credit, they are pointing in the general direction of these new questions (even while not yet acknowledging their mistake). The new questions: Do there exist simple machines whose universality is undecidable (and might 2-3 be an example)? If 2-3's universality is either false or undecidable (and especially in the latter case) can we find any useful structure to what combinations of it with other machines clearly are universal?

I'll leave it as an exercise to figure out how raising those questions relate to the "Priciple of Computational Equivalence" in NKS but, meanwhile, leave the student out of it!

We'll see. With an additional step that "finitizes" the student's construction the proof is rescued and raised to the status of an important lemma -- but if that step isn't very quickly forthcoming, the prize -- in no small part an advertising vehicle -- was administered in a pedagogically misleading way.

-t

Re:Peer Review Rules

[hansraj]

A problem L is NP-hard if any polynomial algorithm for L can be used to solve all problems in *NP* with only a polynomial slowdown. That is the definition of NP-hard. NP-complete problems are NP-hard and have extra property that they are in NP as well.

So, solving any NP-hard (and hence any NP-complete problem) in polynomial time is as good as solving *all* problems in NP in polynomial time. P is just the class of problems solvable in polynomial time.

Hence, polynomial algorithm for *any* NP-complete problem => (P=NP)