The Substrate

The setting must do two things at once. It must let a map be applied to an argument — the bare content of computation — and it must let a map itself become an object, a thing that can be named, passed, and fed its own index. The first is products; the second is exponentials; and the precise statement that binds them is an adjunction. We take the three universal properties in turn, then collapse them into the one that matters.

Two examples fix the boundary of the definition — and the second is more instructive than the first, because it shows that closure is a real demand, easily failed.

There are two ways to speak of an "element" of an object, and confusing them is the most common error in reading a categorical theorem. A global point is a map 1 → A. A generalized element is a map X → A from any object — an "element of A at stage X." Yoneda's lemma says the second notion is complete: the assignment A ↦ Hom(−, A) is a full and faithful embedding, so an object is determined, to isomorphism, by the totality of its generalized elements and how they restrict along maps. Global points are just the slice at X = 1 — and that slice can be empty, or too coarse to tell two maps apart.

The cleanest failure is small and complete. Let G = ℤ/2 and work in G-Set, sets with an involution, equivariant maps between them — a topos, hence cartesian closed. Its terminal object is the one-point set with trivial action, so a point 1 → A is precisely a fixed point of the involution on A. Now take A = {x, σx} with σ swapping the two — the free G-set on one generator.

One last object earns its place here, because the "limit face" of the construction will need it. In a topos the subobject classifier Ω represents subobjects — Sub(A) ≅ Hom(A, Ω) — and plays the role of "the object of truth values." In Set, Ω = {⊤, ⊥}: two values, Boolean, with negation a fixed-point-free involution. In a general topos Ω is an internal Heyting algebra, possibly with many values and a negation that need not be an involution at all. The diagonal's limitative face — the one that yields Cantor and Gödel — turns on the codomain B carrying a fixed-point-free twist; over Ω that is the law of the excluded middle made geometric. Where Ω is non-Boolean, the swap that drives those theorems must be supplied by hand, not assumed. The setting fixes not only what can be named, but what can be negated.

Two surjectivity notions must be kept apart, because the theorem consumes the weaker one and the difference is the whole reason it applies in arithmetic and computation, where the stronger one is false.

Now the necessity. Weak point-surjectivity is not merely enough; it is the most that the corpus's instances ever provide, which is why the theorem must ask for no more. The diagonal lemma of arithmetic gives, for each formula ψ(x), a sentence G with ⊢ G ↔ ψ(⌜G⌝) — agreement at the single point ⌜G⌝, a self-reference holding at one code, not a surjection of codes onto formulas as objects. The s-m-n theorem of computability gives a total s with φ_{s(e,x)}(y) = φ_e(x, y) — enough to internalize a parameter and so to manufacture, for any given function, one index that behaves correctly, again on points. Neither delivers a point-surjection in the strong sense; both deliver exactly weak point-surjectivity. The theorem is calibrated to its instances, not the instances stretched to fit the theorem — and the binary form makes plain that what they share is the cheapest structure imaginable: a way to feed something its own index, inside a category that can merely form pairs.

So a reflexive object cannot be a set — at least, not a set with two distinguishable points and the classical logic that lets them be swapped. It must live where the order of approximation replaces the algebra of membership: in domain theory, the category DCPO of directed-complete partial orders with least element and Scott-continuous maps. That category is cartesian closed — [D → E] is the continuous maps under the pointwise order, itself a dcpo — and in it Dana Scott, in 1969–72, constructed a non-trivial D ≅ Dᴰ: the first mathematical model of the untyped λ-calculus, and the proof that self-application need not be paradoxical, only partial in the right way.

A map f has no fixed point exactly when it displaces every element. The three classical theorems each rule that out by a different surplus of structure: self-reference makes the space name its own maps, so one of them must be f itself caught at its index; order-completeness makes every ascending approximation have a limit the map cannot overshoot; geometry — connectedness, or shrinking distance — leaves the map nowhere to send a boundary, or forces its iterates to converge. The hypotheses are incomparable, but where two coincide the theorems can be made to speak about the same object, and then the discipline is to ask whether they also agree. We take each genus to its constructive core, then map the overlaps.

The third genus splits into two theorems that share a habitat but not a method, and the contrast is worth drawing precisely, because one is mere existence and the other is everything.