On Two Views of Computation in Computer Science

> TM+Halting oracle > TM

Why is this a failure of the Church-Turing thesis? The Church-Turing thesis says, "Anything that can be computed in our reality (i.e. by a human, ignoring resource constraints) can be computed by a TM." TM+Halting oracle doesn't satisfy the hypothesis.

>In practice, people write their programs in a programming language. Some languages (e.g., C, Java, Lisp) only provide leaky abstractions, so programmers always have access to the internal representation of everything, to the detriment of modularity. But there are languages (e.g., Standard ML) that aren't hobbled in this way. ... writing a function that inspects the representation of an abstract type isn't “difficult” - it's literally impossible.

Alright, at that point it becomes an issue of "what kinds of things count as computations" for the purposes of the CTT. You're right that some languages simply lack certain capabilities, but that's an issue of access, not computation per se. You can get the same result in pure math:

"C is incapable of computing the zlobnorg of a matrix."

'Ridiculous! I'm sure I can write a program that does it! So, what exactly is a zlobnorg?'

"Sorry, only the matrix itself can answer that."

If you're going to go that route, then yeah the CTT is false -- likewise, TMs can't see inside black holes, so they can't "compute" that either.

But the "computations" referred to in the CTT are assumed to meet certain criteria, e.g. that there's some API exposed for accessing the relevant information, and the output is a "pure function" of that input (or at least makes any external state changes accessible, etc). But then once you've done so, all the Turing-complete languages can access said information information and compute the correct result.

Now, if your point is that the day-to-day work of a programmer is a lot more than writing pure functions: I whole-heartedly agree and have said the same thing myself. I cringe at the whole "Maxwell's equations of programming". Of course you have to care about external state and access privileges and race conditions and whether the API is usable, etc -- things that the CT notion of "computation" doesn't deal with. But, for the same reason, that's also orthogonal to the CTT.

It is not necessarily representable. By diagonalization, in fact, nat -> nat can be shown to be uncountably infinite. Intuitively, an example nat -> nat function that is uncomputable is ~Ld, the function that, given a natural index i into the diagonal of the table of Turing machines and their accepted strings, returns the opposite of whether machine i accepts string i. The output is then clearly equal to no Turing machine, since it differs in at least one position. In order to encode this function, an oracle is required. As such, nat -> nat uncountability can be shown to be central to proving Turing recognizability.

I think you've got an interesting point, and I'd like to add to it a bit and refine/correct it a bit.

First off, minor thing, but as Jeremy has pointed out, your terminology is wrong; "representation theory" means something else. You seem to mean something more like "model theory".

But anyway -- I'm a mathematician, and I've noticed something like this too when I've dealt with physicists or their writings. But before I discuss physicists, let's discuss mathematicians, because I think you've mischaracterized them a little.

(Note, all claims here are just based on my experience.)

You talk about considering syntax vs. considering the different models of the syntax. But most mathematicians, unless they are logicians or set theorists or something, don't take either of those views. They take the Platonic view -- they're discussing a world of mathematical objects; the statements just describe it. They're not considering the statements as primary and then considering multiple models of it, they're considering the one true world of mathematical objects and describing it with statements (which might have other models, but who cares). This point of view breaks down somewhat when you have to deal with things like the continuum hypothesis and such, but that sort of thing doesn't really come up in ordinary mathematics.

(Note though that taking the Platonic view does not necessarily mean taking the "theory of programming" perspective the article discusses. I mean, I wouldn't take Aaronson's view, that computation is prior to everything else; that's pretty incompatible with it. But to my mind, computation is fundamentally about finite strings over a finite alphabet. Talking about computation with higher types -- where the Church-Turing thesis can fail -- as though it were, well, computation, feels really weird to me.)

Also, another terminological correction: Most mathematicans don't care about type theory. They care about types, at least implicitly, though they might not use that term; but they are not going to consider types as an object of study in and of themself. But yes -- when you want to understand a mathematical object, one of the first questions is, "What sort of object is that?" or "What space does that live in?"

So anyway -- yeah, physicists. I've noticed this too, and I don't know what to make of it. Where physicists don't tend to fully specify their, let's call it an ontology, and have a hard time answering questions like "What sort of object is that?" or "So that's a function from what to what?" Where they don't seem to properly distinguish between a mathematical model or description, and a method of calculation that's useful when working with that model or description. (I have so often failed to get an answer to, "So are virtual particles an actual physical phenomenon predicted by the quantum field theory, or are they just part of a useful method of calculation for working with quantum field theory?") Where they seem to care primarily about methods for getting numbers out, rather than building full, coherent models.

It's all weirdly instrumentalist, which to me seems backwards from what you'd expect -- like, mathematical Platonism is a philosophical question, but the physical world is definitely real! And sure, maybe the bits we can't measure are more of a philosophical question too -- especially since one can quite possibly come up with multiple isomorphic usable models for the laws of physics, whatever they turn out to be -- but still, mathematicians have no trouble just treating their mathematical objects as real things to be reasoned about, so this kind of instrumentalist view among physicists seems weird to me.

I hadn't even considered the possibility that they were thinking "syntax-first", so to speak! But I guess that does kind of sum it up pretty nicely. Bugs the hell out of me, though. Makes it pretty difficult to talk about the math of their theories with them, too.

> Yes but the real number that cannot be pinned down using a formula can also not be pinned down using a computation,

Undefinable real numbers exist, whether you can compute them or not.

> and we are talking of models of computation.

A theory of computation that rejects the existence of things that can't be computed is like a logic that rejects the existence of FALSE because it can't be proved. Where does the busy beaver function sit in your conceptual model?