We all know that we quickly encounter issues of *size w*hen we start category theory. Categories, we are told, are collections of objects and morphisms satisfying certain axioms. But we very soon meet, among our very first examples of categories, categories like **Set** whose objects are too many for the collection of them to itself be a set. So what kind of ‘collection’ do we have here?

But here’s another issue about that initial example of the category **Set** which is ignored by many writers of introductory texts.

The issue can be posed, bluntly, like this: when the category **Set **is mentioned early on in an introductory text *what category is that*?

Typically no explanation is given at this stage. But of course which category we are dealing with depends on our set theory. For an NF-iste, the category of **NFsets** has very different properties from what is intended (for a start **NFsets** is not cartesian closed and later we learn that **Set** is). But fair enough, in an intro book you aren’t going to mention that in the opening pages! Rather the authors are, surely, intending to point to stuff that their beginning readers can be assumed to know about, and are saying, “Hey, you in fact already know about some categories, for example …” The charitable reading, then, is that authors are relying on their readers to think of **Set **as comprising the sets they already know and love.

But what sets are *those*?

Well, for many readers (if they have done any more than had a bit of set-notation waved at them) the sets they know and love are the pure sets of the cumulative hierarchy — pure in that there are no urlements, no memberless entities in the universe of sets other than the empty set. If set theories with urelements are mentioned in passing in an intro set-theory course, it is usually only to be dismissed and forgotten about. So: in the absence of special explicit signals to the contrary, we might well reasonably take the category **Set** mentioned in the very early pages to be a category of pure sets of the usual hierarchy (or at least the hierarchy up to some inaccessible, or whatever). Just sometimes this is made explicit. Thus, Horst Schubert in his terse but very good and clear *Categories *(§3.1) writes “One has to be aware that the set theory used here has no “primitive (ur-) elements”; elements of sets, or classes …, are always themselves sets.”

But then what are we to make, a bit later in our introductory books, of e.g. the usual presentation of the Yoneda embedding as $latex \mathcal{Y}\colon \mathscr{C} \to [\mathscr{C}^{op}, \mathbf{Set}]$. Putting it this way, if you look at the details, assumes that hom-collections $latex \mathscr{C}(A, B)$ for $latex A, B \in \mathscr{C}$ actually live in $latex \mathbf{Set}$. And since such a hom-collection is a set of $latex \mathscr{C}$-morphisms, that assumes that the $latex \mathscr{C}$-morphisms — irrespective of what the [small] category $\mathscr{C}$ is — must live in the world of pure sets too. [Sure, we may want the relevant hom-collections to be set-sized in the Yoneda embedding case — but being no bigger than set-sized is one thing, living in the universe of pure sets is something else!]

*But do we really want to assume that morphisms are always pure sets?*

Might we not be looking to category theory for a story about how different bits of the mathematical universe hang together which need not presuppose some over-arching, all-in, set-theoretic reductionism, and so in particular doesn’t presuppose from day one that all morphisms are pure sets?

Now, as I noted, the foundational sections you often meet early in category theory books worry away about questions of size (sets vs classes etc.). But the present worry is orthogonal to all that, and is in a way more basic. If we want to make no assumption that the denizens of different bits of the mathematical universe are all cut from the same cloth, we won’t want to slip into assuming that *sets* of these denizens are all pure sets. So in particular, do we really want to assume that a collection of morphisms (hom-set) must actually live in **Set**, if that’s the category mentioned back almost on p.1 of the book which is naturally read (and sometimes explicitly said to be) a category of pure sets? It may be that **Set** eventually gets to play a special role in the mathematical universe, with other most other categories being representable inside it: but surely this should be something to be shown later, not an assumption to be built in from the start.

In Chapter 3 of his fine *Basic Category Theory*, Tom Leinster makes the right moves here. Back track. Ask again: what are the sets that, before we start into category theory, we know and love and which are pointed to as constituting an exemplar of a category? Not the sets of pure ZFC (which mathematicians other than set theorists could even spell out the axioms which are supposed to tell you what these are?) but rather what we might call the sets of the working mathematician — sets of naturals, perhaps, or sets of reals, or sets of points in a space, or a set of continuous functions, or … And these are sets with ur-elements, elements that are not themselves presumed to be sets. (Or if we ascend to talk of sets of sets of naturals, for example, it remains the case that when we descend again through members of members … everything typically bottoms out with primitive non-sets.)

We assume various operations are permissible on these sets, and we can codify the permissible operations (for the purposes of the non-set-theorist, falling short of ZFC). Now Leinster associates this with Lawvere’s codification of the Elementary Theory of the Category of Sets (ETCS). But of course, Zermelo was in the same game, aiming to codify the principles of set-building that the working mathematician actually needs, aiming to provide enough to do the usual constructions but not enough to threaten paradox, and Zermelo’s original set theory too allowed for ur-elements. But then, if **Set** is to be a category of sets-with-urelements, built up using the principles working mathematicians actually need (whatever they are) — which indeed makes some of appeals to **Set** in category theory much more natural — then this really needs to be said (as Leinster does, and others don’t, and some actually deny).

Of course, we *could* go Schubert’s way and take hom-sets to be pure sets, so that morphisms and the other apparatus of category theory are all pure sets. But seeing category theory in this way as simply in the service of an across-the-board set-theoretic reductionism would surely be an unhappy way of looking at things: category theory seemed to promise more than that!

Paolo G. GiarrussoI think that Awodey has a counterpoint that you might find interesting, in relation to both this post and to the discussion in your category theory notes (http://philpapers.org/rec/AWOAAT, available https://www.andrew.cmu.edu/user/awodey/preprints/awodeyVhellman.pdf).

I’ll risk venturing my own (short) version, though my set-theory background is limited, so I’d be happy to learn if I’m saying nonsense. When you ask “are morphisms pure sets?” you’re asking a question on the specific model of the logical theory of categories, but the point of category theory is to abstract from specific models. And in fact, I believe caring for a specific model for Set is missing the point. (Unless ZFC with urelements is more abstract than I assume).

Most textbooks pick a specific model for Set, namely ZFC, but it seems they do so because they don’t care too much, and just want avoid the issue, and I will concede with you that’s questionable (though of course there are didactic reasons for that). But you *seem* to be looking for another specific model, even more so in your guide; and I’m not so happy with that.

I like Leinster’s answer (which I learned from Sec. 2 of the excellent http://arxiv.org/abs/1012.5647v3, and seems to agree with the Google Books preview of his book). But Leinster’s answer doesn’t require picking a specific model of ETCS (though it does require Set to be Cartesian closed, hence excluding NF). In particular, he laments that in ZFC everything’s a set. But, I believe, nothing he says requires that, say, 2 is *not* a set. I think it’s enough that isset(2) is independent from our axioms. (Though I’m unsure whether you could state anything like isset(2) in ETCS).

This will not satisfy some foundational and ontological questions (which you might have), because this way, we aren’t proving the existence of a model; to do that, we can resort again to the cumulative hierarchy*, but we can forget it again after we’ve used it to prove the existence of a model. Moreover, we are not explaining what are functions intrinsically — because we’re just saying what properties they must have, and we’re leaving open what they can be. To quote Dana Scott (“Relating theories of the lambda-calculus”, p. 408, where he’s looking for a theory of functions):

“[C]ategory theory has no existential import. (It was not meant to.) Set theory has “too much” existential import. (It was meant to.)”

Peter SmithMany thanks for this. There are a number of points here I’d like to respond to … when I get back from a holiday, so in a couple of weeks.

Paolo G. GiarrussoThanks for your answer — and enjoy your holiday then! I’ll take it into account when commenting on your CT notes.

David AuerbachWow, two consecutive posts about Schubert.