1. in a sentence like “a group is a set equipped with the following operations (…), such that the following axioms are satisfied (…)”, the word “set” denotes a base domain for mathematical structures, equipped with an equality relation (needed to express the axioms of group);
2. in the phrase “the set of even numbers”, the word “set” denotes a subdomain of the domain of natural numbers (which is a set in the first meaning).

One dogma of 20th century mathematics was that one cannot define sets. But this applies only to sets in the sense of objects of a so-called set theory, like ZF, not to sets in the two mathematical meanings, which can be defined.

We will work in a sufficiently rich typed language. A general type won’t be assumed to have any structure, in particular we don’t preequip types with an equality relation. We denote by the syntactical assertion that the term is of type (for example “” means that the type of the term is , and “let be ” is used to declare that the variable is of type ). It is important not to use for this relation between terms and types in the metalanguage, to avoid confusion with the membership relation between elements and sets² in the language of mathematics (see below). Remark: some people in type theory call types “sets”, which add to the confusion.

1. Meaning 1

We call set¹ a type equipped with an equivalence relation (or equality), allowing to express that two objects of this type are equal or not. This is essentially the definition by Errett Bishop in Foundations of constructive analysis; it is usually used in formalisations of mathematics in proof assistants (in this context, it is sometimes called a setoid). As in Bishop’s book, an operation from a type to a type is given, for each , by an object . A function between two sets¹ is an operation preserving equality.

The only structure on the domain of a group needed to express the axioms of group is an equality, so the more general definition of a group one can give is a set¹ equipped with the usual operations satisfying the axioms of group. More generally, a model of a (let’s say first-order) theory with equality is a set¹ equipped with some operations, relations, and satisfying some equations.

Some examples of sets¹ are:

• the set¹ of natural numbers, whose objects are the natural numbers with the ordinary equality of natural numbers;
• the set¹ of natural numbers modulo , whose objects are the natural numbers and equality is equality modulo ;
• the set¹ of rational numbers, whose objects are the quotients , where and are two integers such that , with the usual equality of rationals ( iff );
• the set¹ of functions between two sets¹, with pointwise equality: by definition, two functions are equal if, for every , ;
• the empty set¹, which has no objects, the equality is trivial;
• the one-element set¹ which has only one object, equal to itself.

Let us recall that a category is a type equipped, for each pair of objects , with a set¹ (whose objects are called “morphisms from to ”), with an associative composition and, for each object , an identity , which is a neutral element for the composition.

The right identity criterion for sets¹ and the constructions one can use to create new sets¹ follow from the following assertion:

1. the identity criterion for objects in a category being isomorphism, two sets¹ are considered to be “the same” if there exists a bijection (an isomorphism in the category of sets¹) between them;
2. the constructions one can do with sets¹ are the limits, colimits, etc. in the category of sets¹: cartesian product, coproduct (“disjoint union”), quotient of a set¹ by an equivalence relation, exponentiation, etc.

The first point (as well as the cardinality functor described below) suggests that sets¹ are the right formalisation of Cantor’s notion of cardinal number (as Lawvere has noticed). This identity criterion is the right one for sets as bases of mathematical structures: if a set¹ is equipped with a certain structure, and if it is in bijection with , then we can transfer the structure from to . Two sets¹ equipped with the same kind of structure (e.g. two groups, two rings, etc.) are “the same” if there is a bijection between them preserving and reflecting the structure.

On the other hand, we cannot define two sets¹ to be equal if they have the same elements, because there is no way to compare the objects of two (syntactically) different sets.

2. Meaning 2

The second meaning of “set” in mathematics is probably the first, historically: this is a collection of objects of a same type, like sets of oranges, sets of natural numbers, sets of functions from to , etc. In a second-order theory, the “sets” over which one quantify are interpreted in a model as being the sets² of the domain of the model, which is a set¹.

Let us fix a set¹ . We define a set² of type as an injection . A common abuse of notation consists in writing just instead of , which contributes to the confusion of the two meanings of “set”. It is for this notion of set that a membership relation is available, between objects and sets² of type : we say that is an element of (and we write ) if is in the range of (there exists such that ). Here are a few examples:

• if we are given different objects of , there is a set² where is the set¹ with elements (let us denote its elements by ) and where the injection maps to ; its elements are exactly the objects equal to one of ;
• if we are given a property , with a free variable of type , we can construct the set² of the objects of satisfying : it is where the objects of the set¹ are the objects of satisfying , the equality of is defined as in , and the inclusion maps such an object to itself.

We can define an order relation on sets² of type : we say that if every element of is an element of . So

sets² of type live in an order

As a consequence,

1. the criterion of identity for objects of an order being given by iff and , two sets² and are “the same” if they have the same elements (so, for sets² of we have extensionality);
2. it is with this notion of set that the usual constructions of “naïve set theory” are available: these are the typical constructions in an order, i.e. intersection, union, implication, the smallest set² (the empty set² of type ), the greatest set² (the identity on the whole set¹ ).

Every set² of type can be equivalently described as a selection of some of the objects of : the set² is equal to the set² .

So, for each set¹ , we get a model of a genuine set(²) theory: this is a theory with two sorts (let’s denote them by and ), with a binary relation symbol on (satisfying the axioms for equality), a binary relation symbol on (satisfying the axioms for order), and a binary relation symbol from to , satisfying two axioms:

1. extensionality: iff for all , implies
2. comprehension: for any property , where , there is such that .

3. Comparison; cardinality

It is important to realise that these two notions of set are very different: they live in different structures and have different criteria of identity. But there is a link between these two notions: if we fix a set¹ , there is a “forgetful” functor , which maps a set² to the set¹ . This is Cantor’s cardinality functor for sets² of type (but of course at Cantor’s time, the notion of functor didn’t exist): it maps a set² of type to the “number” of objects of this set², which is a set¹. For example, it maps the set² to the four-elements set¹.

Of course, the cardinality functor can make different sets² identical. For example, in the order , the set² of even numbers is strictly smaller than the set² of all natural numbers, but they have the same “number” of objects, i.e., in the category of sets¹, the set¹ whose objects are the even numbers is in bijection with , so they are the same set¹, when we forget the injection into (we can even represent the set² of even numbers as the injection , since its elements are exactly the even numbers). And the set² in and the set² in live in different orders, so we cannot even ask if they are the same, but if we forget the inclusions into or , we get just two different presentations of the three-element set¹.

4. Meaning 3?

There is a third meaning, which is not included in the title, since it is not in use in mathematics, but only in logic and philosophy of “mathematics”, and sometimes, I’m afraid, in category theory. Moreover, it seems abusive to use the word “set” for this third notion. Originally, it is a special case of the second meaning, the special case we get if, in the theory of sets² given at the end of Section 3, we identify the two sorts and , so that there is only one kind of objects, which play both roles: they are elements and sets² at the same time. Of course, by Cantor’s theorem this is inconsistent (this gives Russell’s paradox). People tried to make consistent this inconsistent theory by restricting in a more or less elegant way the properties to which applies the scheme of comprehension. This has given various so-called “set theories” such as Zermelo-Fränkel, Bernays-Gödel, New Foundations, etc., formalising neither the notion of set¹, nor the notion of set², but some hybrids between elements and sets². Let us call set³ an object of such a theory.

If we have a model of such a theory (if such a model exists…), with the base set¹ of all sets³, to each set³ corresponds a set² of type , the set² of all sets³ “belonging” to , and so corresponds a set¹, by applying to it the cardinality functor. In this way, when we pretend to do mathematics in such a theory, sets³ play the rôle both of sets¹ and sets² (this is probably the origin of the confusion between these two notions). Moreover, the “mathematics” one can develop in such a theory, once an encoding of the basic notions is given, not only admit meaningless assertions (for example, in ZF with the most used encoding of ordered pairs and natural numbers, we can express and even prove the proposition ), but moreover are strongly limited:

1. all sets¹ are canonically sets² of a big set¹ (which in some theories is not even represented by a set³ [e.g. in Z.F., because of Russell's Paradox]), hence they share a universal equality (the equality of ), and so:
2. one cannot anymore speak of types not equipped with an equality (for example, we cannot define general categories, without equality at the level of objects, such as the category of sets¹).

5. Conclusion

As far as I am concerned, I always distinguish sets¹ and sets². I reserve the word “set” for sets¹ and use “subset of A” for sets² of type (following the tradition in category theory). And, of course, I never use sets³, since I’ve never seen any use for them.

6. Links

• For more details about the definitions, see Bishop’s set theory by Erik Palmgren (and Bishop’s book).
• In topos theory, one studies categories whose objects behave like sets(¹). Then subobjects of a given object are usually defined, like sets², as monomorphisms (which generalise injections) with codomain . See Sets for mathematics by Lawvere and Rosebrugh, or Todd Trimble’s posts on his blog about the Elementary Theory of the Category of Sets for a semi-elementary approach to toposes. Note that, in a topos, for each object , there is an object playing the rôle of the “object of all subobjects of ”; but I think that we shouldn’t assume that all subsets of a set(¹) form themselves a set(¹), but rather an order, so I think that topos theory should be based on orders rather than on sets¹; I intend to speak about that later.
• The paper In the Search of a Naive Type Theory, Lecture Notes in Computer Science 4941 (2008), by Agnieszka Kozubek and Paweł Urzyczyn (here are slides with the same title) advocates the same distinction; here is an excerpt of the introduction:
  

  In fact, there are two very basic intuitions that are glued together into the notion of a “set”:

  • Set as a domain or universe;
  • Set as a result of selection.

  We used to treat this identification as natural and obvious. But perhaps only because we were taught to do so. These two ideas are in fact different, and this very confusion is responsible for Russel’s paradox. In addition, ordinary mathematical practice often makes an explicit difference between the two aspects.

• See also Towards a categorical foundation of mathematics by Michael Makkai.

(pdf file of this entry; for the comments, if you type “set<sup>1</sup>”, you get “set¹”)