1

Are there any known or established or researched logics (not restricted to sequent logic) where the proof object

  • Is not necessarily verifiable in every case whether the proof object is valid
  • If it is verifiable, then it can be verified in finite time

The intention is that this logic would be more general than those restricted to primitive recursive computing models; for example, the proof of Godel's incompleteness theorem might not apply as usually presented.

DanielV
  • 23,556
  • I may also ask this question on the cs stackexchange because they also have a very dedicated logic section. – DanielV Sep 30 '17 at 18:16
  • Have you looked into infinitary logic – Rob Arthan Sep 30 '17 at 22:58
  • There is also work on how the incompleteness theorem breaks down if you admit non-recursive axiom systems for number theory. – Rob Arthan Sep 30 '17 at 23:22
  • @RobArthan I haven't heard the term "infinitary logic" although I've seen the general concept discussed. The first sentence from wikipedia makes me hesitant though, "allows infinitely long statements and/or infinitely long proofs" would suggest a required computing model strictly stronger than Turing Machines. I would expect the proof object itself to be of finite size. Can you clarify what a "non-recursive axiom system" is? Does that mean "non primitive recursive" or a computing model strictly stronger than Turing Machines ? – DanielV Oct 01 '17 at 00:18
  • 1
    Every first-order theory comes with an axiom set. If that set is recursive (its members are decided by some program) then the incompleteness theorems apply to it. But even if the set is not recursive, as long as it is semi-recursive (there is a program that accepts exactly its members) or equivalently recursively enumerable (it is the range of a program taken as a partial function), then as described in my post the incompleteness theorems still apply to it! @RobArthan, you may be interested in it too. Another natural question is how much non-computability is needed to obtain a complete theory. – user21820 Oct 04 '17 at 05:08
  • 1
    It turns out that all you need is the first Turing jump (the halting oracle) and you can decide the axiom set of some complete theory extending PA. The trick is to impose some computable enumeration of sentences over PA, such as length-lexicographic ordering. Then we construct a program $P$ using the halting oracle that on input $x$ will initialize a list $L$ to be empty, and then go through every sentence $y$ over PA up to $x$ in order, and if $y,\neg y$ are both not in $L$ then add $y$ to $L$ if $\text{PA}+L+y$ is consistent and add $\neg y$ otherwise. Finally $P$ accepts iff $x$ is in $L$. – user21820 Oct 04 '17 at 05:15
  • 1
    At each step, $L$ is finite and so consistency can be decided using the halting oracle, as it is a simple question about whether the first-order deductive system proves a contradiction from $PA+L+y$ or not. Thus $P$ decides some set of sentences $S$ over PA. Note that $S$ is complete because for every sentence either it or its negation is accepted by $P$. Also note that $S$ extends PA (assuming PA is consistent), since every axiom of PA must be accepted by $P$. Now one might ask what kind of theory $S$ is. It is in fact not $\text{Th}(\mathbb{N})$ because deciding that needs all finite jumps. – user21820 Oct 04 '17 at 05:17

1 Answers1

2

From the comments it seems that you want proof objects that are finite, even if they are potentially unverifiable. In the generalized incompleteness theorem a general formal system is simply a proof verifier program $V$ that halts on every input and accepts (the code of) $(x,y)$ iff $x$ is a (valid) proof of $y$. Your proposal then naturally corresponds to the idea of relaxing the requirement on $V$ and allowing it to not halt on inputs that it does not accept. It turns out that your proposal will still fall to the generalized incompleteness theorems, and we will have incompleteness for "verifiable" instead of "provable". Namely, we have the following results for your system $S$ where we say that $S \vdash P$ iff there is a verifiable proof of $P$ over $S$:

  1. You will never be able to use such a system to construct a program that decides the truth value (according to the meta-system) of every arithmetical sentence (or equivalently every sentence about finite binary strings) correctly. This follows easily from the unsolvability of the halting problem.

  2. If $S$ interprets PA$^-$ (equivalently TC) via a computable translation $ι$ of arithmetical sentences into $S$, and if $S \nvdash ι(0=1)$, then there is some arithmetical sentence $P$ such that $S \nvdash ι(P)$ and $S \nvdash ι(\neg P)$. The proof is the same as for the generalized incompleteness theorem linked above, except that $G$ performs the executions of $V$ in parallel, and so it does not matter that $V$ does not halt when it is unable to verify a proof. Indeed the stipulation that $V$ does halt when the proof can be verified is the crucial reason the proof works. Compared to (1), this requires $S$ to interpret PA$^-$, but does not care whether $S$ proves (the translation of) false arithmetical sentences. In other words, having mere arithmetical consistency is enough to force $S$ to be arithmetically incomplete, even if $S$ is arithmetically unsound.

  3. As mentioned in this more recent post explaining the generalized incompleteness theorem, the computability-based proof relativizes easily. So even if the proof verifier uses an uncomputable oracle, as long as the system can reason about programs using that same oracle, it will suffer incompleteness.

By the way, see this post for further discussion of reasons for or against wanting a foundational system to interpret $PA^-$, and a self-verifying theory of arithmetic that escapes incompleteness.

Anyway your questions are very interesting! I enjoyed thinking about them!

user21820
  • 57,693
  • 9
  • 98
  • 256
  • My internet was down for a while. I'm going to take some time to read through this. – DanielV Oct 04 '17 at 06:39
  • @DanielV: Does my answer make sense? Feel free to ask to clarify any point. =) – user21820 Oct 10 '17 at 16:20
  • @DanielV: I recently posted this, which is a relatively rigorous statement and proof of Godel's and Rosser's incompleteness theorems in computability terms, which should hence be easy for you to understand. – user21820 Nov 13 '17 at 08:02
  • 1
    @DanielV: Hi! I just came across this post again, and decided to update my answer. – user21820 Nov 27 '18 at 16:38