Universal Properties Instead of Elements: What the Categorical Method Buys and What It Costs — Epoche C2
The product of two sets admits two definitions that look unrelated. One says it is the set of ordered pairs; the other says it is an object equipped with two projections through which every pair of maps factors in exactly one way. This essay treats that pair of definitions as a controlled experiment on the method that produced the second, since the categorical method is a single self-imposed restriction — never look inside an object — and the product is the smallest example on which the restriction can be seen doing work. A category consists of objects, arrows between them, an associative composition of arrows and an identity arrow on each object, and every categorical statement must be expressible in those terms alone. That is the rule. What follows is an attempt to say precisely what obeying it proves, and precisely what it makes unsayable. The two styles on one example The element-based definition says that $X \times Y$ is the set of ordered pairs $(x,y)$ with $x$ in $X$ and $y$ in $Y$. The arrow-based definition never mentions a pair. It says that a product of $X$ and $Y$ is an object $P$ together with two arrows $\pi_1 : P \to X$ and $\pi_2 : P \to Y$, called projections, such that for every object $Z$ and every pair of arrows $f : Z \to X$ and $g : Z \to Y$ there is exactly one arrow $h : Z \to P$ with $\pi_1 \circ h = f$ and $\pi_2 \circ h = g$. That "exactly one" is the universal property, and it is the whole content. Written in hom-sets — $\mathrm{Hom}(A,B)$ denotes the collection of arrows from $A$ to $B$ — the definition is the bijection $$\mathrm{Hom}(Z,\, X \times Y) \;\cong\; \mathrm{Hom}(Z,X) \times \mathrm{Hom}(Z,Y),$$ required to hold naturally in $Z$, meaning compatibly with pre-composition by any arrow $Z' \to Z$. This is not a theorem but the definition rewritten: to give an arrow into the product is to give its two components, which is exactly what the universal property asserts. In the category of finite sets the bijection can be checked by counting. With $|X| = 2$, $|Y| = 3$ and $|Z| = 4$, the left side has $6^4 = 1296$ elements and the right side has $2^4 \times 3^4 = 16 \times 81 = 1296$. The same counting exposes the related property of exponentials: $|\mathrm{Hom}(X\times Y, Z)| = 4^6 = 4096$, while $Z^Y$, the set of functions $Y \to Z$, has $4^3 = 64$ elements, so $|\mathrm{Hom}(X, Z^Y)| = 64^2 = 4096$. Both identities are shadows of adjunctions — an adjunction being exactly a natural bijection of hom-sets of this shape: the product functor is right adjoint to the diagonal $Z \mapsto (Z,Z)$, and the exponential $(-)^Y$ is right adjoint to $- \times Y$, which is what "currying" names in the second case. What "naturally" adds, and why counting cannot check it The counting above verifies a bijection for one triple of sets, and it is worth pausing on the gap between that and what the definition demands, because the gap is where the subject began. Two functors can agree in cardinality at every object and still fail to be isomorphic as functors. The standard example is the one Eilenberg and Mac Lane wrote their 1945 paper to formalise: for a finite-dimensional vector space $V$ over $\mathbb{R}$, the dual $V^{*}$ has the same dimension as $V$, so $V \cong V^{*}$ for every $V$; yet no choice of such isomorphisms is natural, whereas the map $V \to V^{**}$ sending $v$ to evaluation at $v$ is. The failure can be proved in three lines rather than gestured at. Restrict to the groupoid of finite-dimensional real vector spaces and linear isomorphisms, on which $V \mapsto V^{*}$ becomes covariant via $f \mapsto (f^{-1})^{*}$. Take $V = \mathbb{R}$, so that $V^{*} \cong \mathbb{R}$ as well, and let $f$ be multiplication by $\lambda \neq 0$. Then $f^{*}$ is multiplication by $\lambda$, so $(f^{-1})^{*}$ is multiplication by $\lambda^{-1}$. A natural family $\eta$ would have to satisfy $\eta \circ f = (f^{-1})^{*} \circ \eta$; writing $\eta$ as multiplication by a scalar $a$, this reads $a\lambda = a\lambda^{-1}$, forcing $a = 0$ whenever $\lambda^{2} \neq 1$. So the only natural family is the zero map, which is not an isomorphism. Naturality is therefore a strictly stronger requirement than pointwise bijection, and it is the requirement that makes the arrow-only definitions determinate. A universal property stated without it would be satisfied by objects one does not want. What the restriction buys Having fixed what the definition demands, the payoff can be stated: the definition never mentioned what the objects were made of, so it applies wherever the axioms do, and one proof covers cases that look unrelated. The first thing it proves is that "the" product is legitimate. Suppose $(P,\pi_1,\pi_2)$ and $(P',\pi_1',\pi_2')$ are both products of $X$ and $Y$. Applying $P'$'s universal property with $Z = P$, $f = \pi_1$, $g = \pi_2$ yields a unique $u : P \to P'$ with $\pi_i' \circ u = \pi_i$; symmetrically there is a unique $v : P' \to P$ with $\pi_i \circ v = \pi_i'$. Then $\pi_i \circ (v \circ u) = \pi_i' \circ u = \pi_i$, and $\pi_i \circ \mathrm{id}_P = \pi_i$ as well; but $P$'s own universal property, applied with $Z = P$, says there is exactly one arrow with that property, so $v \circ u = \mathrm{id}_P$. The same argument gives $u \circ v = \mathrm{id}_{P'}$. The two products are isomorphic, by an isomorphism that is itself unique among those commuting with the projections. Nothing about sets, groups or spaces entered. Category What the product turns out to be What the universal property is doing Sets and functions Cartesian product pairing two functions into one Groups and homomorphisms direct product with componentwise operation forcing the pairing map to be a homomorphism Topological spaces and continuous maps product topology (for infinite products; for finite ones the box topology agrees) selecting the coarsest topology making both projections continuous A partially ordered set, viewed as a category greatest lower bound of two elements "below both, and above anything bel