Todd Wilson asks:

As someone who is "striving to realize the dream of a theoretical computer science", I would better like to understand the point that Lawvere is making here. Am I right in assuming that, in using terms such as "spurious ingredients" and "rigidification", Lawvere is referring to the fact that (to use some computer science terminology) set theory is too much implementation and not enough specification? That the rigid epsilon-structure of set theory cannot represent abstract mathematical structure faithfully, without introducing unwanted details?

Executive summary of a long reply ------------------------------------------------- * Spaces - for instance topological spaces - cannot always be analysed as set + structure. * Category theory enables structure to be found amongst the morphisms.(generalized elements) instead of within the objects (amongst concrete elements in a set-theoretic analysis). * Variation of kind of space bears some relation to variation of logic. * The logical questions are important in computer science, even if topology and geometry at first appear not to be. ------------------------------------------------------------------ The discussion of Cartesian product does illustrate set theory vs. categories as implementation vs. specification, and, sure, there's a lot to be said for specifying categorically before you implement set-theoretically. However, Bill's point is much deeper than that. He is saying that set theory is actually inadequate as an implementation language in many contexts. I'll set out my own understanding of this, but I hope I'm not misrepresenting Bill if I draw on his note. I think within his note you already see two criticisms of set theory. The less deep of the two is of the "rigidity" of the epsilon-structure. Most mathematicians would probably agree that it doesn't make much sense to investigate the elements of pi (say), yet still view set theory as a remarkably successful encoding of mathematics into an epsilon-structure, successful enough to be used as foundations. The deeper criticism is that even for things that are unambiguously collections of some kind, so they really do have elements, the set-theoretic approach is inadequate. Bill refers to "cohesive and variable sets (or `spaces')", and I take it that an example of "cohesive" set is a topological space. There it is not enough to know merely what the elements are. They also "cohere", sit in rather subtle relationship with each other. Maps should respect the coherence, i.e. be continuous. One way to view this is that in order to know the coherence in a space X you need to know not just its "point-like" elements but also its line-like, plane-like or any-shape-you-choose elements - as Bill said, "Volterra already recognized that spaces have elements other than points". These are the "generalized elements", the maps from other spaces to X. Hence in the category of spaces the structure of X is seen not within the object but amongst the morphisms. Correspondingly, we get two approaches to topological spaces. 1) Set theory - structure within objects. Space = set + topology. The topology provides criteria so that out of all the functions, mostly discontinuous, between the sets of points, some can be identified as continuous. 2) Category theory - structure amongst morphisms. We are now free to describe the objects in point-free ways, for instance the frames of locale theory, the schemes of algebraic geometry, or Grothendieck toposes as generalized spaces, and still hope to get (via the generalized elements) a good spatial feel for the objects. Given a choice, why use the point-free ways at all? But in fact there's less of a choice than one might think, as the geometers discovered. Approach (1) does not generalize well to situations where you want to vary the set theory. This want arises naturally, for concrete mathematical reasons, when you start working with sheaves, since the class of sheaves over a space provides a non-standard set theory, usually non-classical, and one in which rigid "element of" or "subset of" are not useful. As you vary the set theory there are ways of transforming sets in one to sets in another, but it turns out that sets of points of spaces cannot be transformed in the same way: the space and its set of points part company. [See Appendix for a technical discussion.] Examples thus show that the decomposition "space = set + topology" is not stable under change of set theory. The inadequacies of point-set topology are well known in constructive mathematics, where the point-set statements of important results such as the Heine-Borel theorem turn out to be false while point-free versions still hold. Note that category theory in itself doesn't tell you what category to use (sheaves? locales? toposes? schemes?). But it does help you to see what you are looking for in a category of spaces that goes beyond space = set+structure. Computer science --------------------------- That was a long preamble to say Beware! Spaces might not be sets! But why should workers on Knowledge Interchange Formats worry about this? Why can't they just use first-order logic, assume a classical set-theoretic semantics and keep well away from sheaves? One reason lies in the classical first-order logic itself. It works in a blandly uniform way on its formulae that ignores any difference in status as knowledge. For instance, maybe a knowledge database should distinguish between formulae that represent observed facts and those that represent background assumptions or scientific hypotheses. It could make a difference to what negation means or how negation behaves. But such nuances are not present in classical first-order logic. These logical questions could be a natural part of the theory of knowledge representation. The interesting thing is that the variability of logic can be connected to the possible un-set-likeness of spaces. When you replace set theory's elements by category theory's generalized elements, you sometimes get a calculus that's almost like set theory but with different logic. For instance, intuitionistic logic (internal logic of toposes) for sheaves, geometric logic for spaces. The effect is of enabling the point-free spaces and maps to be handled just like sets and functions - describe a space by its points, define a map without needing to prove continuity - but with a funny logic. Hence although Knowledge Interchange Formats might seem a million miles away from topology or algebraic geometry, subtle logical questions might create direct links. (My particular favourite is geometric logic, whose distinction between formulae and sentences actually looks quite like that between observed facts and background assumptions.) Steve Vickers. Appendix ------------- Let f: X -> Y be a map between spaces and let f^* be the inverse image functor on sheaves. It is the way of transforming "sets over Y" to "sets over X". If p: Z -> Y is a local homeomorphism - one way of describing a sheaf - then its inverse image f^*(p): f^*(Z) -> X is just the pullback of p along f. Just from the categorical property of pullback one can say that the (generalized) points of f^*(Z) are the pairs (x,z), x a point of X and z a point of Z, such that f(x) = p(z). Now it makes good sense to say that a space over Y is just an arbitrary map (not necessarily a local homeomorphism) g: W -> Y and again it is natural to use pullback to transform it to a space g': Z' -> X over X. Any such space over Y has a corresponding sheaf over Y, its "set of points". However, f^* of the set of points is not in general the set of points of the pullback space.