02.11.2012 Views

Reduction and Elimination in Philosophy and the Sciences

Reduction and Elimination in Philosophy and the Sciences

Reduction and Elimination in Philosophy and the Sciences

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

objects. But <strong>the</strong>y also leave unanswered a variety of<br />

foundational questions concern<strong>in</strong>g what it means to regard<br />

algorithms <strong>in</strong> this manner. For note that if algorithms are<br />

<strong>in</strong>deed treated as <strong>in</strong>tensional entities with<strong>in</strong> computer<br />

science, <strong>the</strong>n we might fear that will be forced to posit a<br />

novel class of non-extensional (<strong>and</strong> perforce nonma<strong>the</strong>matical)<br />

abstract objects <strong>in</strong> order to account for <strong>the</strong><br />

truth conditions of statements of types II) <strong>and</strong> III). 4 This<br />

concern serves to illustrate <strong>the</strong> importance of establish<strong>in</strong>g<br />

a positive answer to (A).<br />

Some hope that such an answer may be given<br />

comes from reflect<strong>in</strong>g on <strong>the</strong> orig<strong>in</strong>s of computer science<br />

with<strong>in</strong> computability <strong>the</strong>ory. The orig<strong>in</strong>s of <strong>the</strong> latter subject<br />

can be traced to <strong>the</strong> call to provide a ma<strong>the</strong>matical<br />

def<strong>in</strong>ition of <strong>the</strong> class C of effectively computable functions<br />

as it arose with<strong>in</strong> <strong>the</strong> Hilbert programme. It this context,<br />

<strong>the</strong>re was general agreement that a function f: N k → N is<br />

effectively computable just <strong>in</strong> case <strong>the</strong>re exists an<br />

algorithm for comput<strong>in</strong>g its values. But <strong>in</strong> order to show<br />

that a given function is not effectively computable required<br />

that C be given a precise def<strong>in</strong>ition.<br />

The developments lead<strong>in</strong>g to <strong>the</strong> consensus that C<br />

should be identified with <strong>the</strong> class of partial recursive<br />

functions — i.e. to <strong>the</strong> claim now known as Church's<br />

Thesis [CT] — are sufficiently familiar that <strong>the</strong>y need not<br />

be repeated here. What is less well recognized is that <strong>the</strong><br />

orig<strong>in</strong>al arguments for CT did not proceed by first giv<strong>in</strong>g a<br />

ma<strong>the</strong>matical def<strong>in</strong>ition of a class A which could plausibly<br />

be taken to consist of objects correspond<strong>in</strong>g to algorithms<br />

<strong>and</strong> <strong>the</strong>n def<strong>in</strong><strong>in</strong>g<br />

(C) C =df {f : N k → N : f(x1, ..., xn) = A(x1, ..., xn) & A ∈ A}<br />

Ra<strong>the</strong>r, Church, Tur<strong>in</strong>g, Gödel, <strong>and</strong> Post all proceeded by<br />

def<strong>in</strong><strong>in</strong>g a class of formal models M (which I will refer to<br />

somewhat <strong>in</strong>accurately as mach<strong>in</strong>es) which formalize<br />

different notions of what it means to be a f<strong>in</strong>itary<br />

procedure. For <strong>in</strong>stance, for Gödel, M consisted of <strong>the</strong><br />

class of general recursive def<strong>in</strong>itions. Such def<strong>in</strong>itions can<br />

be taken to formalize a variety of ways <strong>in</strong> which functions<br />

can be <strong>in</strong>troduced so that <strong>the</strong>ir values can be explicitly<br />

computed (e.g. by course of values recursion). But if we<br />

def<strong>in</strong>e FM to be <strong>the</strong> class of functions computable by<br />

members of M, for each choice of M, <strong>the</strong> question rema<strong>in</strong>s<br />

as to whe<strong>the</strong>r <strong>the</strong> correspond<strong>in</strong>g class FM exhausts all<br />

effectively computable functions.<br />

In order to demonstrate that we ought to accept C =<br />

FM thus requires an additional argument that for any<br />

<strong>in</strong>formally characterized algorithm A, <strong>the</strong>re exists a<br />

mach<strong>in</strong>e M ∈ M such that A <strong>and</strong> M determ<strong>in</strong>e <strong>the</strong> same<br />

function. This appears to be an extensional claim about<br />

<strong>the</strong> relationship between two classes of functions. But note<br />

that any argument <strong>in</strong> its favor must apparently proceed by<br />

<strong>the</strong> follow<strong>in</strong>g <strong>in</strong>tensional route: i) given any algorithm A,<br />

<strong>the</strong>re is an M ∈ M such that each step <strong>in</strong> <strong>the</strong> <strong>in</strong>formally<br />

characterized operation of A can be correlated with one or<br />

more steps <strong>in</strong> <strong>the</strong> operation of M; ii) hence <strong>the</strong> function<br />

<strong>in</strong>duced by <strong>the</strong> complete operation of M co<strong>in</strong>cides with that<br />

<strong>in</strong>duced by that of A. Incipient arguments to this effect may<br />

be found <strong>in</strong> <strong>the</strong> orig<strong>in</strong>al papers of Church, Tur<strong>in</strong>g, <strong>and</strong> Post<br />

from 1936. Better fleshed out versions appear <strong>in</strong> <strong>the</strong><br />

4 The gravity of this concern will ultimately depend upon how tightly <strong>the</strong> practice<br />

of computer science p<strong>in</strong>s down <strong>the</strong> identity conditions which must be<br />

imposed on algorithms. It follows from <strong>the</strong> example of <strong>the</strong> previous note that<br />

extensionally equivalent algorithms cannot be identified when <strong>the</strong>y differ with<br />

respect to a def<strong>in</strong>ite computational property such as asymptotic runn<strong>in</strong>g-time<br />

complexity, But at <strong>the</strong> same time, <strong>the</strong>re do not appear to be cases <strong>in</strong> which<br />

statements of algorithmic non-identity -- i.e. of <strong>the</strong> form A1 ≠ A2 -- are accepted<br />

<strong>in</strong> computer science when no such property serves to dist<strong>in</strong>guish A1 <strong>and</strong><br />

A2.<br />

Algorithms <strong>and</strong> Ontology — Walter Dean<br />

writ<strong>in</strong>gs of more recent commentators such as (Rogers<br />

1967), (G<strong>and</strong>y 1980), <strong>and</strong> (Sieg & Brynes 2000).<br />

Inasmuch as any sound argument for CT must<br />

proceed <strong>in</strong> <strong>the</strong> manner just suggested, one might<br />

reasonably conclude that at least certa<strong>in</strong> choices for M will<br />

<strong>in</strong>clude a formal representation of every algorithm. 5 And on<br />

this basis, one might conclude that it is allowable to take<br />

A = M <strong>in</strong> (C). But <strong>the</strong> members of M will generally be f<strong>in</strong>ite<br />

comb<strong>in</strong>atorial objects, <strong>and</strong> thus ma<strong>the</strong>matical objects par<br />

excellence. Thus one also might conclude that not only<br />

should (A) be answered <strong>in</strong> <strong>the</strong> positive, but such an<br />

answer is already implicit <strong>in</strong> our acceptance of CT.<br />

The fact that such a conclusion is not warranted<br />

follows by reflect<strong>in</strong>g fur<strong>the</strong>r on some basic results which<br />

have emerged from algorithmic analysis. As noted above,<br />

for <strong>in</strong>stance, algorithms are <strong>in</strong>dividuated at least as f<strong>in</strong>ely<br />

as <strong>the</strong>ir big-O runn<strong>in</strong>g-times. But for certa<strong>in</strong> choices of M,<br />

we can f<strong>in</strong>d examples of algorithms A comput<strong>in</strong>g certa<strong>in</strong><br />

functions f to which we assign runn<strong>in</strong>g-time O(tA(|x|)) but<br />

for which it may be shown that <strong>the</strong>re is no M ∈ M<br />

comput<strong>in</strong>g f with runn<strong>in</strong>g-time O(tM(|x|)) ≤ O(tA(|x|)). 6 If we<br />

take <strong>the</strong> property of hav<strong>in</strong>g runn<strong>in</strong>g-time O(tA(|x|)) to be a<br />

property of A itself, <strong>the</strong>n results like this suggest that we<br />

cannot take algorithms to be identical to members of any<br />

specific class M. For if we were to do so, <strong>the</strong>re would be<br />

no guarantee that <strong>the</strong>re is a member of M which faithfully<br />

represents A’s computational properties.<br />

This situation highlights <strong>the</strong> k<strong>in</strong>d of conceptual <strong>and</strong><br />

technical problems which arise when we attempt to settle<br />

(A) directly by identify<strong>in</strong>g algorithms with mach<strong>in</strong>es. For on<br />

<strong>the</strong> one h<strong>and</strong>, <strong>the</strong> argument for CT sketched above<br />

promises to show that for every <strong>in</strong>formally presented<br />

algorithm A, <strong>the</strong>re will exist a mach<strong>in</strong>e M A ∈ M which<br />

mimics its step-by-step operation. But on <strong>the</strong> o<strong>the</strong>r h<strong>and</strong>,<br />

<strong>the</strong> question of determ<strong>in</strong><strong>in</strong>g when <strong>the</strong> existence of a<br />

particular form of step-by-step correlation is sufficient to<br />

allow us to conclude that M A is identical to A appears to<br />

require that we have a prior characterization of <strong>the</strong><br />

properties of A itself. This is to say that before we can be<br />

<strong>in</strong> a position to assess whe<strong>the</strong>r a given argument for (A) of<br />

this form might be successful, we must first agree on how<br />

our computational practices fix <strong>the</strong> properties of <strong>in</strong>dividual<br />

algorithms.<br />

At this po<strong>in</strong>t, a number of analogies between (A)<br />

<strong>and</strong> various reductive proposals <strong>in</strong> <strong>the</strong> philosophy of<br />

ma<strong>the</strong>matics can be drawn. For note that if we agree that<br />

algorithms are regarded as <strong>in</strong>tensional objects <strong>in</strong> computer<br />

science, (A) amounts to <strong>the</strong> claim that reference to such<br />

entities can be elim<strong>in</strong>ated <strong>in</strong> favor of extensional<br />

ma<strong>the</strong>matical ones. The desire to demonstrate such a<br />

claim can thus be compared to <strong>the</strong> traditional nom<strong>in</strong>alist<br />

desire to show how reference to ma<strong>the</strong>matical entities can<br />

be elim<strong>in</strong>ated <strong>in</strong> favor of reference to concrete ones.<br />

This observation suggests that o<strong>the</strong>r strategies are<br />

available <strong>in</strong> attempt<strong>in</strong>g to demonstrate (A) than simply<br />

attempt<strong>in</strong>g to identify a ma<strong>the</strong>matical object to correlate<br />

with each <strong>in</strong>dividual algorithm — i.e. what (Burgess &<br />

5 This po<strong>in</strong>t is put by (Rogers 1967) (p. 19) as follows: “[T]here is a sense <strong>in</strong><br />

which each of <strong>the</strong> st<strong>and</strong>ard formal characterizations appears to <strong>in</strong>clude all<br />

possible algorithms ... For given a formal characterization..., <strong>the</strong>re is a uniform<br />

effective way to ‘translate’ any set of <strong>in</strong>structions (i.e. algorithm) of that characterization<br />

<strong>in</strong>to a set of <strong>in</strong>structions of one of <strong>the</strong> st<strong>and</strong>ard formal characterizations.”<br />

6 A number of examples of lower-bound results of this nature are known for<br />

<strong>the</strong> s<strong>in</strong>gle-tape, s<strong>in</strong>gle-head Tur<strong>in</strong>g mach<strong>in</strong>e model T which is most often<br />

referenced <strong>in</strong> Rogers-style translation arguments. For <strong>in</strong>stance, while <strong>the</strong>re is<br />

a trivial O(n) algorithm for determ<strong>in</strong><strong>in</strong>g whe<strong>the</strong>r a b<strong>in</strong>ary str<strong>in</strong>g is a pal<strong>in</strong>drome,<br />

no mach<strong>in</strong>e T ∈ T can solve this problem <strong>in</strong> time faster than O(n2) (cf. Hopcort<br />

<strong>and</strong> Ulman 1979).<br />

59

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!