A guest post by Eric Allender prompted by an (incorrect) P ≠ NP proof recently published in Springer Nature's Frontiers of Computer Science.
For a time, I served as Editor-in-Chief of ACM Transactions on Computation Theory, and in this role I had to deal regularly with submissions that claimed to resolve the P vs NP problem. Finding referees for these papers was sometimes challenging, so I frequently ended up reviewing them myself. Dealing with such submissions involves enough overhead that ToCT, J.ACM and ACM Transactions on Algorithms limit the frequency with which authors can submit work of this sort. But for every submission, on any topic, the overall process was the same: Find someone on the Editorial Board who has sufficient expertise to find referees and make knowledgeable use of their recommendations. If there was no such person on the Editorial Board, then it was always the case that the submission was out of scope for the journal.
These thoughts are brought to mind by a recent case, where it seems to me that the editorial process broke down.
Springer publishes several high-quality academic journals that deal with Theoretical Computer Science. One Springer journal, Frontiers of Computer Science, recently published an article entitled SAT Requires Exhaustive Search, where one of the authors is a Deputy Editor-in-Chief of the journal. The abstract of the article states that it proves a result that is stronger than P not equal to NP. The Editorial Board of the journal has some members who are expert in computational complexity theory. However, all the ones whom I know personally have asserted that they had no knowledge of this paper, and that they were not involved at all in handling the paper.
When Ryan Williams and I learned about the publication of this article, we drafted a comment, which we sent to the Editor-in-Chief. We recommended that the paper be retracted, in which case there would be no need to publish our comment. However, the Editor-in-Chief declined to retract the article, saying that he could find no evidence of misconduct, and thus we have been assured that an edited version of our comment will appear in the journal.
Our comment calls attention to some shortcomings in the proof of the main theorem (similar to shortcomings in several other failed attempts to separate P from NP). But we were able to say more. Often, when one is looking for bugs in a purported proof, one has to deal with the situation where the claimed theorem is probably true, and the only problem is that the proof is not convincing. However, the main theorem in their paper (Theorem 3.2) states that a particular constraint satisfaction problem requires time more than \(d^{cn}\) for any constant \(c<1\) (where \(d\) is the domain size, and \(n\) is the number of variables). In particular, their purported proof claims that this holds even when \(k=2\) (meaning that each constraint has at most 2 variables). However, Ryan Williams presented an algorithm more than two decades ago that runs in time \(O(d^{(0.8)n})\) in this special case, contradicting the lower bound claimed in the article.
The article contains an appendix, with glowing testimonies from various researchers; the lead-in to the appendix also contains enthusiastic comments from Gregory Chaitin. I contacted Gregory Chaitin, and he asserts that he did not read the paper, and that he was quoted out of context.
The edited version of the comment that Ryan Williams and I wrote, which will supposedly appear in Frontiers of Computer Science soon, differs from the version linked here (our original submission) primarily in one respect: Our closing paragraph was removed. Here is that paragraph:
Finally, it is our opinion that the publication of this article is a complete embarrassment to this journal and its publisher. We believe that, at the very least, the paper should be withdrawn, and Springer should conduct an investigation to understand how such a paper could have made it through the peer review process.
Update (Lance, 8/5/25): The authors of the paper in question have asked me to link to their reply to the Allender-Williams comment. I do so with no endorsement.
Eric Allender asked me to note that their claim about Ryan’s algorithm requiring exponential space is misleading; the amount of space is not more than the runtime of the algorithm, which is less than the lower bound that they claim. (Their theorem does not state that \(d^{cn}\) time is required under the additional assumption that the algorithm use only \(n^{O(1)}\) space.)
See also https://arxiv.org/abs/2312.02071 for another rebuttal of an earlier version of the same paper
ReplyDeleteAlso see: https://arxiv.org/pdf/2401.01193 and https://arxiv.org/pdf/2309.05689 which are even more hilarious.
DeleteThis seems to be a popular topic. https://eprint.iacr.org/2025/445 but of course they don't perform peer review - it's just a preprint server.
ReplyDeleteThis is either hilarious or sad. Or, both.
ReplyDeleteBoth
DeleteWith so many attempts, something will slip thru the cracks. Good job pointing this out!
ReplyDeleteIf a paper does not brag that it solved P NE NP and that result appears in the middle of the paper, YES I can imagine that slipping through. but when the claim that P NE NP is in the abstract, this is not a `just slipped thru' scenario.
DeleteDon't let President Trump know. Fake Science + China!
ReplyDeleteThe comment: “ authors can submit work of this sort” seems very derogatory .”Work of this sort” what the hell is this about here. Either have the courage to be fully descriptive but not so obnoxious please
ReplyDelete"Work of this sort" is referring to "submissions that claimed to resolve the P vs NP problem" from earlier in the paragraph.
DeleteThis was a reference to the P/NP policy of these journals, which is described at length here: https://dl.acm.org/journal/toct/pnp-policy
DeleteTo save you from having to follow the link, here is what it says: "No author may submit more than one paper to J. ACM, ACM Trans. on Algorithms, or ACM Trans. on Computation Theory in any 24-month period, purporting to resolve the P versus NP question or related long-standing questions in complexity theory, except by invitation of the Editor-in-Chief. This applies to resubmissions of previously rejected manuscripts."
I am sorry that you found my language to be obnoxious.
Another thing: an editor in chief of the journal should be aware of how such a journal made it into its journal. This journal is not as wide ranging as TCS … also likely is that the co-author who is an editor made things happen where other papers would have gotten stuck
ReplyDeleteWhen I hear stuff like “ lGregory Chaitin. I contacted Gregory Chaitin, and he asserts that he did not read the paper, and that he was quoted out of context.” i cringe… what does this even mean? Did he say it or not?
ReplyDeleteIs Greg still active in research or not? If not, I can understand. If you don’t read a paper, how do you have an opinion about it? How can you say something favorable other than … nice format and nice typeface! So something feels eerie here.
The paper says 'As of March 2025, more than 30 experts have provided highly positive feedback, including remarks like “truly revolutionary” from Prof. Gregory Chaitin'. I can believe Chaitin said something like "This result would be truly revolutionary if it's correct, so check the proof extremely carefully" and they quote-mined him.
DeleteThe article says "As of March 2025, more than 30 experts have provided highly positive feedback, including remarks like “truly revolutionary” from Prof. Gregory Chaitin." It's possible that Greg said something like, "if true, the result is truly revolutionary."
DeleteThis comment has been removed by the author.
DeleteTechnically, you did not point out the error, but merely that Ryan's result and the paper are inconsistent and therefore at least one of them is wrong. :)
ReplyDeleteNo, our full comment points to a serious (and obvious) mistake in the proof of the main result. The authors flatly assume that any algorithm for SAT must proceed by reducing an instance of size n to instances of size n-1 in a particular way.
DeleteThe fact that the statement of the main result was also directly refuted 20 years ago just makes the whole thing extra silly.
nothing "eerie" here - the paper in question is how a blatant academic fraud looks like.
ReplyDelete"eerie" is an eerie word.
DeleteWe should open a journal on proofs of P=NP and P\neq NP and any publications go in there. Problem solved?
ReplyDeleteWorld Scientific in 2016 even published a whole book implying P=NP, via a purported polynomial-size LP for TSP (something that was already disproved in our STOC'12 paper).
ReplyDeletehttps://www.worldscientific.com/worldscibooks/10.1142/9725#t=aboutBook
To be more precise & avoid confusion: they gave a purported polynomial-size extended formulation for the TSP polytope (which would imply a polynomial-size LP)
DeleteI took a look at the book wondering if the P=NP claim was buried in the middle (so the editor may have missed it) or overt (so no excuse to miss it). The TITLE of the book is fine: Advances in Combinatorial Optimization: LP formulations of the TSP and other hard opt problems (SO I THOUGHT: They will either formulate approx versions, or sho whow SOME TSP problems can be formulated as an LP or stuff like that).
DeleteMoney quote from the abstract on the website: This work also represents a proof of the equality of the complexity class ''P'' (polynomial time) and "NP" (nondeterministic polynomial time), and makes a contribution to the theory and application of `extended formulatins' (EFs).
Nothing subtle about that. I wonder (a) why this was not caughy early in the procss, and (b) why this book from April 2016 was not noticed earlier.
Next step: Either this book or the recent paper try to claim the $1,000,000 prize.
April 1 book?
DeleteSadly no- the book is an actual book, not an april fools day joke. For one thing, if it was an April fools day joke it would be a lot funnier.
Delete@gasarch, wait I am confused now. The authors of the paper I thought are not claiming that they’re resolved PvsNP but instead resolved a question they thought was more relevant. Hence my belief is they are not after the 1 m usd prize money. Correct me if i am wrong.
DeleteIn partial answer to Anonyous_5:44: The paper states unambiguously that they are proving stronger than P not equal to NP. Quoting from the abstract: "we prove, by the diagonalization method, that constructed self-referential CSPs cannot be solved by nonbrute-force computation, which is stronger than P ≠ NP".
DeleteMy comments are based on the paper as it was published in the journal.
sweet jesus. then i misunderstood their response or in some way they failed to explain themselves in a way that a student a non-native english speaking student can understand what they are actually after. if they just had framed it differently, like we actually don't care about pvsnp but think what we are showing is interesting -- i am not saying "more" interest than proving PvsNP -- just interesting ... i would have let them off the hook. but now that Eric Allender put it into perspective, i cannot find a reason to be supportive in any way. @ericAllender:8:09am, thx for clarifying.
DeleteEric or Ryan- I did a blog post a while back about how `proofs' of P=NPor P NE NP have never (literally never) had an interesting idea in them. So NO ``They don't prove P=NP but they have an idea for a better SAT Solver' or `Theydon't prove P NE NP but the ideas here could lead to SAT not in n^2''
ReplyDeleteSO- is there ANYTHING of interest in the current paper?
There are sections of the paper that discuss the probability with which a random instance of CSP (in their model) is satisfiable. I didn't go through this part of their paper carefully, but the reference that David Eppstein provided in an earlier comment (https://arxiv.org/abs/2312.02071) indicates that people have gone through that argument carefully and found no real problems there. This is conceivably of interest, but I don't think that it is very helpful for understanding why P is not equal to NP (assuming that this is the case). The paper also contains philosophical discussions (with mentions of Wittgenstein and others), which would have been more interesting if their argument had been correct.
ReplyDeleteThe Moser-Tardos Algorithm does not appear to be based on divide and conquer. As such, Xu and Zhou's assertion in their article that 'As we know, at least until now, all exact algorithms for CSPs are based on divide and conquer style, such as backtracking method' is inaccurate.
ReplyDeleteThe Moser-Tardos Algorithm is not an exact algorithm which can ensure to give a solution or say there is no solution.
DeleteIIUC, the Moser-Tardos Algorithm is not an exact algorithm which can either find a solution or proof there is no solution.
DeleteFirst of all thanks to Eric and Ryan for pointing this out.
ReplyDeleteAs a Ph.D. student in a peripherally related field who happen to see a post about this on X and came here, I am surprised that Eric/Ryan even bothered to draft this much of a comment in response to this paper in the first place. This paper seems to come out of a Chinese university that writes papers for the sake of publication alone, without any regards for quality. This reminds me of a similar instance a few years ago where one of the professors had purported to have discovered sequences that both converge and diverge (or something of this ridiculous sort), with the exception that this time the paper was even accepted to publication!
Alas, it is highly possible that there are many such instances (e.g., at ICML/NeurIPS) whose errors are undetected...
Maybe the paper was refereed by ChatGPT?
DeleteYou given money.
DeleteOr someone equally dumb.
ReplyDeleteThe authors’ country of origin, unfortunately, is highly predictable. Rampant academic misconduct. Even if it’s a small percentage, the population is so large that it’s too much volume of misconduct (incorrect results, plagiarism, etc )
ReplyDeleteAt reply from authors of paper. What do people think about it? It's hard to understand what is being conveyed partially because of language barrier and partially because the way it has been put down on paper. Good thing. The authors are real people. they do exist, they have a voice, and they have responded the comment.
ReplyDeleteThey have voiced that they have been misunderstood. (ok fine this is not relevant)
They have addressed the point that Williams and Allender raised -- "we do not claim that such an algorithm is likely to suceed anyhow ..." so back to Wiliams and Allender:
So if you two claims that there might be other ways, but are somewhat confident that they might not work anyhow ... i am left in a bubble about what to think. You guys don't point out anything concrete but are on equal footing with the authors of the paper.
What they did not do ... they did or did not address the concrete example by Williams?
Can someone point out if they addressed this point?
"(and obvious) mistake in the proof of the main result. The authors flatly assume that any algorithm for SAT must proceed by reducing an instance of size n to instances of size n-1 in a particular way.
The fact that the statement of the main result was also directly refuted 20 years ago just makes the whole thing extra silly."
How was it directly refuted 20 years ago and what are we actually talking about?
Finally i find it amusing that they keep quoting Chaitin. How does Chaitin feel about it?
So the authors succeeded in what they set out to do, engage. whether work is frivolous or not. they got their engagement. Congratulations you two.
Regarding the paragraph after "so back to Williams and Allender:"
DeleteThe point that Ryan and I were attempting to make in our comment is the following (which I hope is obvious):
The statement of Theorem 3.2 claims that any algorithm that solves (their formulation of) the Constraint Satisfaction Problem requires time essentially d^n. This means that they must rule out the existence of ANY algorithm that solves CSP and has a smaller run time. It does NOT mean that they only need to rule out algorithms that use only a small amount of space. (If they could prove that, it would of course also be a fantastic result. But that's not what they claim, and it's not what they do.) It also does NOT mean that they only need to rule out algorithms that look just like the algorithms that we already know about. It's the algorithms that we DON'T know about that really need to be excluded. In our comment, Ryan and I pointed out that the argument that purports to prove Theorem 3.2 focuses only on algorithms that operate in a particular way, and we gave an example of an entire class of possible algorithms that are not addressed by their argument. (Namely, take the CSP instance, construct a graph from it, and look to see if it contains a particular kind of subgraph.)
In order to show that an argument is insufficient, it is not necessary to present an algorithm that beats the claimed lower bound. It doesn't matter that we don't think that this is a particularly promising class of algorithms: The point is that there is (still) no proof that an algorithm of this sort will not work.
But in fact we WERE able to point to an algorithm that beats their lower bound! The authors explicitly claim that their lower bound holds even in the case where k=2, and Ryan Williams published an algorithm 20 years ago whose runtime beats their claimed lower bound. So how can there be any additional discussion about whether their proof is correct?
Regarding "Can someone point out if they addressed this point?" They attempted to address this point, by pointing out that Ryan's algorithm uses quite a bit of space, and saying "Researchers do not include this type of algorithm in their discussions of exact algorithms for CSPs". Since Ryan's paper on this topic has more than 500 citations, this strikes me as a very dubious assertion.
The argument that "other algorithms might exist" feels analogous to claiming "aliens could exist in the universe".
DeleteIn computational complexity, space efficiency is fundamentally more valuable than time efficiency—this is a well-established principle. Allender’s attempt to dismiss space-complexity concerns by citing paper references strikes me as equally unconvincing.
After reviewing all the comments, I'm left with the impression that certain participants seem resistant to the idea of P vs NP being resolved at all.
I in no way intended to dismiss space complexity concerns. You seem to have mis-read my statement. Instead, I was intending to dismiss the assertion that researchers don't consider the algorithm in [Williams] to be a CSP algorithm. Clearly, there are quite a few researchers who do consider it to be a CSP algorithm worthy of consideration. The citation count backs me up.
DeleteRegarding 'The argument that "other algorithms might exist" feels analogous to claiming "aliens could exist in the universe".':
Consider the linear programming problem. In the 1970's, it was not known whether there was a polynomial-time algorithm to solve linear programming. Even at that time, there had been decades of experience by a large community of practitioners and researchers, solving linear programming using variants of the simplex algorithm. One could imagine someone trying to show that there was NO polynomial-time algorithm, by observing that all known variants of the simplex algorithm had superpolynomial running time, and (with some justification, considering the community at that time) claiming that researchers did not consider non-simplex algorithms to be legitimate approaches to solving linear programming. But of course Khachiyan's breakthrough approach opened the eyes of the community to new approaches to linear programming. That is: an "other algorithm" really did exist.
It is an extremely important question, whether or not an "other" approach to CSP can exist. But it is not particularly interesting to simply show that "no CSP algorithm that looks exactly like the algorithms that we're currently using can run much faster".
hi @anon8:31PM: not sure this is a fair comparison. I think time efficiency is as core as space. So, while you might be pro paper, i feel you are missing out. Fair point goes to both though. If it took exponential space to solve, then fine, we might make a point like this ... at the same time then, would there be novel ways to reduce the space required (for Williams k=2 case). If so, I think any criticism from paper is not warranted. I will have to side pro Williams and Allender on this one.
DeleteI fully acknowledge the significance of Williams' algorithmic contribution. However, my primary concern lies in the fundamental dimensional incomparability between the two approaches. Xu’s framework—whether employing brute-force or non-brute-force strategies—operates strictly within polynomial space, whereas Williams’ non-brute-force algorithm inherently requires superpolynomial space. This divergence in space complexity places the two methods in fundamentally different computational regimes, rendering direct theoretical comparisons between them problematic.
DeleteAnonymous_12:57: Are you saying that the proofs in Xu's paper are correct and that it proves P ne NP, or are you saying there are some useful ideas in the paper even though it does not prove P ne NP?
DeleteAnonymous_12:57: at no point in the paper it is mentioned that this theorem only holds in polynomial space. The theorem as stated contradicts Williams' earlier result. This reason alone should be enough reason for a retraction or erratum. Not even mentioning the silly p not equals np claim.
DeleteIn my opinion, the brute-force method for CSP is widely recognized to have polynomial space complexity. The authors are discussing whether there exists a superior non-brute-force alternative capable of replacing it—one that, crucially, must also exhibit polynomial space complexity to ensure comparability. Since polynomial space complexity is already an inherent property of brute-force methods, explicitly stating it here would be redundant.
DeleteI'm deeply intrigued by the controversial aspects of this paper—we may very well be witnessing a pivotal moment in academic history.
DeleteHi @Eric Allender 9:21 PM: Computational problems can be categorized into two main types: combinatorial/unstructured problems and algebraic/structured problems [1]. The Constraint Satisfaction Problem (CSP) falls into the former category, whereas the Linear Programming Problem (LPP) belongs to the latter. These two types of problems exhibit significant differences in terms of computational complexity and algorithm design.
Delete[1] Barak, Boaz. "Structure vs combinatorics in computational complexity." Bulletin of EATCS 1.112 (2014).
ReplyDeleteI have read the paper, "SAT requires exhaustive search." ( I also strongly believe anyone who wants to make meaningful comments and arguments here should also read the paper at least once).
The authors argue that the most fundamental problem in computer science is the distinction between non-brute-force computation and brute-force computation, rather than P versus NP. In my view, this is the paper’s most significant point because it serves as the core motivation for their work and pertains to the strategic direction of computational complexity research.
The second most important point is the construction of self-referential instances, which illustrates what hard instances look like and why they are difficult to solve.
The least important point is the proof of the main result, as it directly employs the standard diagonalization method.
Regarding "The authors argue that the most fundamental problem in computer science is the distinction between non-brute-force computation and brute-force computation, rather than P versus NP":
DeleteI agree completely that the question about whether brute-force computation is required is fundamental.
This question has, in fact, received a GREAT DEAL of attention in the field of computational complexity theory. There is a significant literature regarding the “perebor conjecture”, going all the way back to the seminal work of Leonid Levin, and before), dealing precisely with the question of whether exhaustive search is required. (Interestingly, there is some surprising work showing that some formulations of the perebor conjecture are false. See [Mazor & Pass, ITCS 2024] and [Hirahara, Ilango, Williams; STOC 2024].) Work on the "Strong Exponential Time Hypothesis" also falls under the topic of whether brutue-force computation is required. This is certainly not a new question.
The parts of the paper that deal with the construction of self-referential instances would be much more interesting if they actually contributed to a correct proof of their Theorem 3.2. Unfortunately (as I've tried to explain elsewhere in this series of comments) no convincing proof of their Theorem 3.2 is in hand.
I have read this paper and my minor comment is that the "proof" is not, as claimed by the authors, a standard diagonalization. Refer to the paper Diagonal Arguments and Cartesian Closed Categories for I've read this paper, and my small comment is that this "proof" isn't a standard diagonalization, as the authors claim. For a standard construction of diagonalization, see the paper "Diagonal Arguments and Cartesian Closed Categories." This paper doesn't construct most of the structures that appear in proofs by Russell, Gödel, Cantor, and Turing. standard structure of diagonalization. Most of the structures present in Russell, Gödel, Cantor and Turing's proof are not constructed in this paper.
DeleteYes, the issue of exhaustive search has been discussed before. However, by linking this problem to Gödel's work, the authors of the paper "SAT Requires Exhaustive Search"have endowed it with a brand-new significance. Furthermore, no one has explicitly pointed out that this problem is actually more fundamental than the better-known P vs. NP problem.
DeleteA key aspect of Gödel's work lies in the construction of self-referential propositions. To demonstrate that such propositions are unprovable within finite formal systems, Gödel put forward assumptions regarding the nature of proof: a proof is a finite sequence of symbols, and the correctness of a proof can be verified mechanically. On this basis, Gödel's incompleteness theorem was proven using the diagonalization method. Similarly, in the paper "SAT Requires Exhaustive Search", the authors constructed self-referential CSP instances and proposed assumptions about how CSP instances are detremined to be satisfiable or unsatisfiable. They then employed the standard diagonalization method to prove their main result. Thus, the problem that requires discussion is not whether their simple and straightforward proof is correct, but whether we can replace the current assumption with a better one. In other words, there ought to be a bridge connecting self-referential instances to the diagonalization method. The question then becomes: can we construct a bridge that is better than the current one?
It seems that the paper's framework is based on Gödel's framework. If this framework is valid, then the paper should be valid. Many scholars have utlized Turing's model and Shannon's circuit model to prove lower bounds. My question is why Gödel's framework has never been considered until now.
DeleteAnonymous_5:31: A "framework" is not a proof.
DeleteIn my perspective, the primary challenge in applying Gödel's method lies in the construction of self-referential instances. Some scholars have put forward analogous ideas, although they have not explicitly mentioned Gödel's method itself. For instance, Mulmuley [1] noted that to resolve the self-referential paradox in the P vs. NP question, a method should exist that can efficiently find counterexamples for any algorithm claiming to solve the hard problem.
Delete[1] K Mulmuley. Explicit proofs and the flip. arXiv preprint arXiv:1009.0246, 2010
I don't know what to say. Are the authors saying we don't claim to have resolved PvsNP? but happy we resolved what we think matters most. So these folks are very practical computer science students. that is ok, then. they should go publish under that thesis.
ReplyDeleteAssuming we're addressing two distinct problems:
ReplyDelete1. Is there a better algorithm than brute-force search to solve all Constraint Satisfaction Problems (CSPs)?
2. Is P equal to NP (P vs. NP)?
I definitely feel that problem (1) is more intuitive than problem (2). A potential solution to (1) would be to construct specific CPS instances that cannot be solved more efficiently than with a brute-force search. If such instances exist, the answer to problem (1) would be "no."
If the answer to problem (1) is no, this directly implies that P != NP. The existence of a CSP that requires brute-force search (which is exponential in time) means there's a problem in NP that can't be solved in polynomial time.
However, I also feel that problem (1) might be harder to formalize precisely than problem (2). One challenge, as other comments have pointed out, is the assumption that algorithms for CPS must be based on a "divide-and-conquer" approach. I assume this is the status quo for many existing exact algorithms for CPS.
Isn't the whole point that #1 is an open problem which this paper incorrectly claims to close for SAT, thereby closing #2 as well?
DeleteConsider how each of these three could "look" like exhaustive search could be required:
(a) determining whether a planar graph is 3-colorable is NPC
(b) determining whether a planar graph is 4-colorable is "yes, it is"
(c) determining whether a planar graph is 2-colorable is trivial
A proof of the #1 you state is not given in the paper even though the title makes the claim about SAT, with the specifics of this pointed out several times here already.
Frontiers of Computer Science should be about science, not sophistry.
Nature should have learned from earlier mistakes they have made.
I noticed that many comments focus on space complexity, so I compared Williams's and Xu's approaches. Williams's algorithm achieves a time complexity of $O(d^{0.8n}$ by partitioning the n variables into k parts, each requiring $d^{n/k}$ assignments. However, once edge counting is taken into account, the space complexity becomes $O(d^{2n/k}$. In the RB model, where $d=Poly(n)$ , this leads to an impractically large space requirement.
ReplyDeleteIn fact, Williams explicitly raises the question in his conclusion of whether faster algorithms exist for 2-CSP optimization using only polynomial space. Given this, I find the claim that Xu's work contradicts Williams's to be unconvincing.
The discussion of space complexity arose only because of Theorem 3.2 (Model RB cannot be solved in O(d^{cn}) time for any constant 0 < c < 1). Perhaps the intent of the authors was to condition the validity of Theorem 3.2 under their central assumption (this task is finished by dividing the original problem into subproblems) but they stated it without such a condition. As a result, the Williams algorithm provides a counterexample to Theorem 3.2. Presumably, the central assumption is so strong that it would have (implicitly) excluded all algorithms of this form.
DeleteIt is well-known that a Turing machine operating in t steps cannot visit more than O(t) cells. Thus you appear to be claiming that the algorithm in [Williams] cannot be implemented in the Turing machine model within the given time bound. However, I believe that the analysis in [Williams] holds in the Turing machine model, which means that I reject your assertion about the space usage.
DeleteNobody is denying that it is an interesting question, whether CSP (even 2-CSP) can be solved more efficiently than exhaustive search using only polynomial space. However, this question is still open; the paper in question discusses only a small class of possible algorithms.
Xu's Theorem 3.2 doesn't mention space. If you want to rewrite the paper so that it is correct, we can then see what you have proved.
DeleteTo the best of my knowledge, in existing studies investigating exact algorithms for SAT and CSP problems (e.g., [1-4]), researchers seldom explicitly address space complexity, although polynomial space complexity appears to be an implicit requirement in these algorithmic frameworks.
Delete[1] Scheder , D., Steinberger, J. PPSZ for General k-SAT and CSP—Making Hertli’s Analysis Simpler and 3-SAT Faster. comput. complex. 33, 13 (2024). https://doi.org/10.1007/s00037-024-00259-y
[2] Shibo Li & Dominik Scheder (2021). Impatient PPSZ - a faster algorithm for CSP. In 32nd International Symposium on Algorithms and Computation, ISAAC 2021. https://doi.org/10.4230/LIPIcs.ISAAC.2021.33.
[3] Dominik Scheder (2021). PPSZ is better than you think. In 62nd IEEE Annual Symposium on Foundations of Computer Science, FOCS 2021. https://doi.org/10.1109/FOCS52979.2021.00028.
[4] Dominik Scheder (2019). PPSZ for k ≥ 5: more is better. TOCT 11(4), 25:1–25:22. https://doi.org/10.1145/3349613.
Side note: Here is a paper that says that MAX-2-SAT is in P and as a result, concludes that P=NP. The paper is on arXiv, and not published to my knowledge, so not fraud, but I think maybe Gemini and ChatGPT ingested this. Both told me that MAX-2-SAT is in P. The paper was discussed on Hacker News, etc, and thus seems to have high PageRank. https://arxiv.org/abs/2304.12517
ReplyDeleteInteresting that this one is on its 22nd revision. There are atleast fifty wrong proofs of P vs NP a year, but this is also not the only wrong one about MAX2SAT specifically.
DeleteI am less concerned with the details of the proof itself. What truly captures my interest is the comparison between non-brute-force computation vs brute-force computation and P vs NP.
ReplyDeleteA widespread consensus exists among researchers that P ≠ NP, yet proving this conjecture remains an extraordinarily daunting challenge. One might even propose accepting it as an axiom outright. However, even if we were to adopt this assumption, it would still fall short of meeting the demands of contemporary computational complexity research. Increasingly, scholars are basing their work on stronger conjectures. A notable example is explored in the paper: Vassilevska Williams, Virginia. "Hardness of easy problems: Basing hardness on popular conjectures such as the strong exponential time hypothesis (invited talk)." 10th International Symposium on Parameterized and Exact Computation (IPEC 2015).
In light of this, the field may ultimately shift its focus from the classic P vs NP problem toward embracing stronger conjectures as foundational pillars.
The authors of the paper "SAT Requires Exhaustive Search" also note in their abstract that stronger conjectures may prove easier to tackle: "Specifically, proving lower bounds for many problems, such as 3-SAT, can be challenging because these problems have various effective strategies available to avoid exhaustive search. However, for self-referential examples that are extremely hard, exhaustive search becomes unavoidable, making its necessity easier to prove. Consequently, it renders the separation between non-brute-force computation and brute-force computation much simpler than that between P and NP."
DeleteInterestingly, I have observed that other researchers hold a similar perspective on this point. For instance, on page 29 of Scott Aaronson’s excellent survey paper "P =? NP" (included in Open Problems in Mathematics, published by Springer in 2016, with the URL: https://www.scottaaronson.com/papers/pnp.pdf), the author notes: "In my view, the central reason why proving P ≠ NP is hard is simply that, in case after case, there are amazingly clever ways to avoid brute-force search, and the diversity of those ways rivals the diversity of mathematics itself."
In mathematics, it is indeed true that sometimes a stronger conjecture is actually easier to prove. This is because stronger conjectures often reveal more universal laws and deeper structural connections, providing more systematic tools and clearer ideas for proofs. In contrast, weaker conjectures may, due to their confinement to specific cases, obscure the essential logic instead. Examples include Fermat's Last Theorem and the Taniyama-Shimura Conjecture, as well as the problem of primes in arithmetic progressions and Dirichlet's Theorem.
DeleteThe discussion here unfortunately got a bit derailed by the question of polynomial space in Ryan's algorithm. However, there is a much simpler polynomial-space algorithm that works for every fixed k. Consider a nontrivial constraint on k variables; by nontrivial, we mean that not all the d^k possible assignments are satisfying. So we can assume that there are at most d^k-1 possible assignments on these k variables that satisfy the constraint. We branch into at most d^k-1 directions by fixing the value of these k variables, thereby reducing the number of variables by k. This means that the depth of the search tree is at most n/k, resulting in a search tree whose number of leaves is at most (d^k-1)^{n/k}=(d^{k-1/k})^n =d^{cn} for c=(k-1)/k<1. I believe this is precisely the type of reduction algorithm that the authors try to rule out.
ReplyDeleteThis idea works for any fixed k, not just k=2. In general, there is nothing mysterious about the fact that certain problems described by bounded-size constraints can be solved faster than brute force.
The article is a bit vague if k is fixed or not: Theorem 3.2 does not state this, but the proof argues for the fixed k=2 case.
If k is unbounded, that the algorithm sketched above does not give an improvement over brute force. However, if k is unbounded, then there is the issue of how the constraints are represented in the input and how it contributes to the size of the input instance: a full truth table can be exponentially large. For example, one can consider a setting where every constraint is violated only by one assignment (as in the case of SETH) and a popular model in database theory assumes that all the satisfying tuples are listed in the input.
Hi @Dániel Marx
ReplyDeleteIf (d^k-1)^(n/k) = d^{cn}, then c = log(d^k-1)/log(d^k).
If d and k are fixed, then c<1. However, d grows polynomially with n in this paper.
You are right, my calculation is off and this gives c<1 only if d and k are both fixed.
DeleteThis paper appears to present several novel ideas. A straightforward method to assess the validity of a new idea is to examine its applicability to other problems. Specifically, this papers aims to demonstrate the necessity of brute-force computation by constructing self-referential instances, and confirms that such instances can indeed be constructed for Constraint Satisfaction Problems (CSPs) with growing domains. To apply this idea to other problems, we need to identify the properties under which self-referential instances can be constructed to establish the necessity of brute-force computation.
ReplyDeleteIn Appendix B of this paper, I note that reviewer Bin Wang highlighted a crucial yet easily understandable property: growing domains significantly weakens the correlation between assignments, making the solution space appear nearly independent. Formally speaking, two assignments I and J are independent if Pr(I, J are solutions)=Pr(I is a solution)Pr(J is a solution). The near-independence property ensures that the slight change made by the symmetry mapping will not affect the remaining solution space, thereby enabling the construction of self-referential instances.
This property reminds me the well-known hard problems like Independent Set (or Clique), where the task is determine whether a graph with n vertices contains a subset of k vertices such that no two are adjacent. Currently, no known algorithm can find an independent set of size k=n^epsilon more efficiently than brute-force search [1, 2]. For this problem, it is easy to observe that between two random candidate sets of size k=n^epsilon, with high probability these two sets have no overlap. Consequently, such a solution space exhibits near-independence. Given this characteristic, it should be feasible to construct self-referential instances using an approach analogous to that presented in this paper.
[1] R. G. Downey, M. R. Fellows. Fixed-parameter tractability and completeness II: On completeness for W [1]. Theoretical Computer Science, 1995, 141(1-2): 109-131.
[2] J. Chen, X. Huang, I. A. Kanj, G. Xia. Strong computational lower bounds via parameterized complexity. Journal of Computer and System Sciences, 2006, 72(8): 1346-1367.
The challenge of establishing computational lower bounds has long remained an enigma. As Prof. Avi Wigderson observed in his book Mathematics and Computation: "Concluding, I view the mystery of the difficulty of proving (even the slightest non-trivial) computational difficulty of natural problems to be one of the greatest mysteries of contemporary mathematics." In this light, I concur that constructing self-referential instances may offer a promising path toward resolving this mystery. This approach has been instrumental in proving several renowned mathematical impossibility results, yet intriguingly, it has not been applied to the domain of computational lower bounds. This unexplored potential makes it a compelling direction for future investigation.
Deletelog(d^k-1)/log(d^k) = ((k-1)*log d)/(k*log d) = (k-1)/k
ReplyDeleteWhat is inside the Log function is d^k-1, not d^(k-1).
Delete