Tuesday, June 26, 2007

Church’s Thesis 14: Open texture and computability

Back at last to my blogview of the papers in Church’s Thesis After 70 Years (new readers can start here!) -- and we've reached a very nice paper by Stewart Shapiro, “Computability, Proof, and Open-Texture”, written with his characteristic clarity and good sense. One of the few ‘must read’ papers in the collection.

But I suspect that Shapiro somewhat misdescribes the basic logical geography of the issues in this area: so while I like many of the points he makes in his paper, I don’t think they support quite the conclusion that he draws. Let me explain.

There are three concepts hereabouts that need to be considered. First, there is the inchoate notion of what is computable, pinned down -- in so far as it is pinned down -- by examples of paradigm computations. Second, there is the idealized though still informal notion of effective computability. Third, there is the notion of Turing computability (or alternatively, recursive computability).

Church’s Thesis is standardly taken -- and I’ve been taking it -- to be a claim about the relation between the second and third concepts: they are co-extensive. And the point to emphasize is that we do indeed need to do some significant pre-processing of our initial inchoate notion of computability before we arrive at a notion, effective computability, that can reasonably be asserted to be co-extensive with Turing computability. After all, ‘computable’ means, roughly, ‘can be computed’: but ‘can’ relative to what constraints? Is the Ackermann function computable (even though for small arguments its value has more digits than particles in the known universe)? Our agreed judgements about elementary examples of common-or-garden computation don’t settle the answer to exotic questions like that. And there is an element of decision -- guided of course by the desire for interesting, fruitful concepts -- in the way we refine the inchoate notion of computability to arrive at the idea of effective computability (e.g. we abstract entirely away from consideration of the number of steps needed to execute an effective step-by-step computation, while insisting that we keep a low bound on the intelligence required to execute each particular step). Shapiro writes well about this kind of exercise of reducing the amount of ‘open texture’ in an inchoate informal concept and arriving at something more sharply bounded.

However, the question that has lately been the subject of some debate in the literature -- the question whether we can give an informal proof of Church’s Thesis -- is a question that arises after an initial exercise of conceptual refinement has been done, and we have arrived at the idea of effective computability. Is the next move from the idea of effective computability to the idea of Turing computability (or some equivalent) another move like the initial move from the notion of computability to the idea of effective computability? In other words, does this just involve further reduction in open texture, guided by more considerations ultimately of the same kind as are involved in the initial reduction of open texture in the inchoate concept of computability (so the move is rendered attractive for certain purposes but is not uniquely compulsory). Or could it be that once we have got as far as the notion of effective computability -- informal though that notion is -- we have in fact imposed sufficient constraints to force the effectively computable functions to be none other than the Turing computable functions?

Sieg, for example, has explored the second line, and I offer arguments for it in my Gödel book. And of course the viability of this line is not in the slightest bit affected by agreeing that the move from the initial notion of computability to the notion of effective computability involves a number of non-compulsory decisions in reducing open texture. Shapiro segues rather too smoothly from discussion of the conceptual move from the inchoate notion of computability to the notion of effective computability to discussion of the move from effective computability to Turing computability. But supposing that these are moves of the same kind is in fact exactly the point at issue in some recent debates. And that point, to my mind, isn’t sufficiently directly addressed by Shapiro in his last couple of pages to make his discussion of these matters entirely convincing.

But read his paper and judge for yourself!

No comments: