Sunday, December 05, 2021

Yes Virginia, there is a Santa Clause for Complexity Theorists, If you Only Believe

(Guest Post by Hunter Monroe) In this  guest post and discussion paper, I present a remarkable set of structurally similar conjectures which, if you only believe them, conjure up a dream world for theorists by asserting a new form of diagonalization based on naturally nonrelativizing facts invoking a deep linkage to underlying noncomputable languages. These conjectures, all stronger than the things to be proved, imply that the polynomial hierarchy does not collapse because the arithmetic hierarchy does not collapse, and P≠NP≠coNP. The diagonalizations imply the existence of hard instances, with the result that many complexity classes have speedup, including the Π side of PH, and proof speedup for tautologies stems from proof speedup for arithmetic. These conjectures do two things: (1) let us explore a hypothetical world where many open problems about uniform complexity classes are resolved and consider steps beyond e.g. to circuit complexity, and (2) reduce numerous open questions to a single plausible claim about how Turing machines have limited information about noncomputable languages. This would potentially allow a slew of open questions to be resolved at once with a skeleton key.

The following conjecture implying $\textbf{P!=NP}$ is remarkable: it hints at a deeper, unnoticed relationship between complexity and noncomputability; it is equivalent to speedup for all paddable $\textbf{coNP}$-complete languages and in proof length for tautologies; tweaked versions would separate other complexity classes; and if true it is a nonrelativizing fact.

Conjecture: (*) For any deterministic TM $M$ accepting the $\textbf{coNP}$-complete language ``nondeterministic TM $N$ on input $x$ does not halt within $t$ steps'' ($\texttt{coBHP}$), there exists a $\langle N',x'\rangle\in\texttt{coHP}$ ($N'$ does not halt on $x'$, ever) with $M$'s running time $f(t)=T_M(N',x',1^t)$ not bounded by any polynomial.

If true, (*) is a nonrelativizing fact; there is no hard $\langle N, x\rangle$ for $M$ with an exponential time oracle. The noncomputable language $\texttt{coHP}$ potentially explains why (*) is true, by analogy with this trivial theorem:

Theorem. For any $M$ accepting $\texttt{coBHP}$, there exists some non-halting $\langle N',x'\rangle\in\texttt{coHP}$ with $f(t)=T_M(N',x',1^t)$ not bounded by a constant.

Otherwise, $M$ would accept $\texttt{coHP}$ and have too much information about a non-c.e. language; (*) is just a stronger version. In the extreme, any $M$ is completely ignorant about some $\langle N',x'\rangle$ and requires on the order of $2^t$ steps to rule out every potentially halting branch. Tweaking (*) yields conjectures implying that $\textbf{PH}$ does not collapse:

Conjectures: For any $M^{\Pi^p_i}$ that accepts $\Pi^p_{i+1}=\{\langle N^{\Pi^p_i},x,1^t\rangle|$ $N^{\Pi^p_i}$ does not halt on input $x$ in $t$ steps$\}$, there is a non-halting $\langle N'^{\Pi^p_i},x'\rangle\in \Pi_i$ with $M^{\Pi^p_i}$'s running time not bounded by any polynomial.

By invoking every level $\Pi_i$ of the arithmetic hierarchy ($\textbf{AH}$), these conjectures state that the noncollapse $\textbf{PH}$ is due to the noncollapse of $\textbf{AH}$. The conjecture (*) can be calibrated depending on the desired separation to equip $M$ with an oracle or nondeterminism or constrain its resources, to choose a resource-bounded complete problem and underlying non-c.e. language, and to fine tune how hard a hard instance needs to be.

Proof speedup for tautologies (equivalent to (*)) may stem from the proof speedup for arithmetic that occurs when adding undecidable statements as new axioms, allowing new theorems to be proved and shortening the proof of existing theorems. This literature translates any arithmetic theorem free of existential quantifiers into a tautology by replacing $k$-bit numbers with $k$ Boolean variables. The analogy with (*) suggests this stronger conjecture may in fact also be equivalent:

Conjecture: The following two statements are equivalent: (1) there is no optimal propositional proof system; and (2) Any propositional proof system $P$ is outperformed by a sufficiently powerful conservative extension $T$ of the Peano arithmetic, and $T$ can be improved further by adding any undecidable statement in $T$ as a new axiom.

So (*) is a Swiss army knife for generating conjectures that give us a vision of a world in which answering essentially one question would serve as a skeleton key that unlocks many open problems.

5 comments:

  1. In words, it says that for any Turing machine M accepting the TMs that do not halt within t steps, there is one family of inputs for M indexed by t which is hard for M, that is the running time of M on that family is superpolynomial in t. Have you looked at the discussion paper https://speedupblogger.wordpress.com/2021/12/05/unifying-conjectures-on-complexity-and-noncomputability/ and the Theoretical Computer Science paper https://arxiv.org/abs/0906.3765? Are you saying the statement is imprecise or hard to interpret?

    ReplyDelete
    Replies
    1. Now I did, and I think I can understand (*). What I don't understand is, why it doesn't contradict the existence of a universal TM. That could simply run N' on x' for t steps and the simulation only takes a polynomial amount of time in t.

      ps. Instead of leaving a new comment, please reply to mine, and then I'll get notified. Now a week passed before I've again accidentally stumbled into this...

      Delete
    2. Because N' is nondeterministic, the brute force simulation would take time 2^t.

      Delete
  2. Thanks for the "Swiss army knife" ping; I enjoyed the conjecture spin of the software toolkit version analogy ... will take a look at things.

    ReplyDelete