Category theory — functors & natural transformations

Functors are structure-preserving maps between categories; natural transformations are structure-preserving maps between functors. One-line: a functor sends objects to objects and arrows to arrows compatibly with composition and identities, while a natural transformation provides a coherent family of arrows relating two parallel functors. Key result: functors and natural transformations between two categories themselves form a category (the functor category), and naturality is exactly the condition that makes "morphisms of morphisms" compose correctly. Applies throughout algebra, topology, logic, and theoretical computer science. Introduced by Eilenberg and Mac Lane (1942–1945).

A functor $F : C \to D$ assigns to each object $X$ of $C$ an object $F (X)$ of $D$ , and to each morphism $f : X \to Y$ a morphism $F (f) : F (X) \to F (Y)$ , subject to functoriality: it preserves identities and composition. A natural transformation $η : F \Rightarrow G$ between two functors $F, G : C \to D$ is a family of $D$ -morphisms $η_{X} : F (X) \to G (X)$ , one per object, such that every morphism's image is respected — captured by a commuting square. These notions matter because they let mathematics treat constructions (homology, free groups, tensor products) and comparisons between constructions as first-class objects. The surprising fact, emphasized historically by Eilenberg and Mac Lane, is that category theory was invented in order to define natural transformations — the genuinely novel concept — and functors and categories were the necessary scaffolding to make "natural" precise.

The Core Insight

Functoriality turns the slogan "compatible with composition" into the only data you ever need to carry structure between mathematical worlds. Naturality then makes precise the elusive intuition that an isomorphism or transformation is "canonical" — independent of arbitrary choices — by demanding that it commute uniformly with all morphisms at once. The naturality square is the single diagram from which almost everything in higher category theory grows.

Formal Statement

Functor

Let $C$ and $D$ be categories. A (covariant) functor $F : C \to D$ consists of:

an object function $X \mapsto F (X)$ , $Ob (C) \to Ob (D)$ ;
for each pair of objects $X, Y$ , a morphism function $Hom_{C} (X, Y) \to Hom_{D} (F (X), F (Y))$ , written $f \mapsto F (f)$ ;

satisfying the two functoriality axioms:

F (id_{X}) = id_{F (X)} for all X \in Ob (C) .

⟳identities are uniquely characterized by

g \circ id = g

and

id \circ h = h

; applying

F

to these and using

[2]

forces

F (id_{X})

to be the identity on

F (X)

— so

[1]

is the minimal axiom making identities track.

F (g \circ f) = F (g) \circ F (f) for all composable f : X \to Y, g : Y \to Z .

⟳read both sides as morphisms

F (X) \to F (Z)

; the axiom states that the object function on hom-sets is a homomorphism of the composition operation, the categorical analogue of

ϕ (ab) = ϕ (a) ϕ (b)

Here $F (X), F (Y), F (Z)$ are objects of $D$ and all symbols $\circ$ denote composition in the respective category. A contravariant functor reverses arrows: $F (f) : F (Y) \to F (X)$ and $F (g \circ f) = F (f) \circ F (g)$ ; equivalently, it is a covariant functor $C^{op} \to D$ .

Natural Transformation

Let $F, G : C \to D$ be functors. A natural transformation $η : F \Rightarrow G$ assigns to each object $X \in Ob (C)$ a morphism $η_{X} : F (X) \to G (X)$ in $D$ , called the component at $X$ , such that for every morphism $f : X \to Y$ in $C$ the naturality square commutes:

G (f) \circ η_{X} = η_{Y} \circ F (f) .

⟳trace an element conceptually around the square

F (X) \to F (Y) \to G (Y)

versus

F (X) \to G (X) \to G (Y)

; both routes must land at the same morphism

F (X) \to G (Y)

, which is precisely the equality

[3]

If every component $η_{X}$ is an isomorphism in $D$ , then $η$ is a natural isomorphism, written $F ≅ G$ , and the componentwise inverses $η_{X}^{- 1}$ assemble into a natural transformation $η^{- 1} : G \Rightarrow F$ .

Functor Category

For categories $C, D$ (with $C$ small), the functors $C \to D$ as objects and natural transformations as morphisms form a category, the functor category $D^{C} = [C, D] = Fun (C, D)$ . Composition is vertical composition of natural transformations: for $η : F \Rightarrow G$ and $θ : G \Rightarrow H$ ,

(θ \circ η)_{X} = θ_{X} \circ η_{X} .

⟳pasting two naturality squares along their shared middle column

G (f)

yields a commuting rectangle whose outer boundary is exactly the naturality square for

θ \circ η

; the component formula is read off the bottom row.

The identity morphism on $F$ is the natural transformation $id_{F}$ with components $(id_{F})_{X} = id_{F (X)}$ .

Interchange / Horizontal Composition

Given functors and natural transformations

C F, G D H, K E, η : F \Rightarrow G, ϵ : H \Rightarrow K,

there is a horizontal composite

ϵ * η : H F \Rightarrow K G

with components

(ϵ * η)_{X} = ϵ_{G (X)} \circ H (η_{X}) = K (η_{X}) \circ ϵ_{F (X)} .

⟳the two expressions are equal precisely because the naturality square of

ϵ

applied to the morphism

η_{X}

commutes; the equality

[5]

is naturality of

ϵ

instantiated at

η_{X}

Horizontal and vertical composition satisfy the interchange law:

(θ * δ) \circ (η * ϵ) = (θ \circ η) * (δ \circ ϵ),

⟳both sides are computed componentwise; expand each using

[4]

and

[5]

, then apply functoriality

[2]

H, K

and naturality

[3]

to rearrange the resulting four-term composite — every reordering is licensed by one square commuting.

making the totality of categories, functors, and natural transformations a 2-category $Cat$ .

Why It Holds / Derivation

Vertical composition is well-defined and associative

We verify the functor category axioms.

Composite is natural. Given $η : F \Rightarrow G$ and $θ : G \Rightarrow H$ , and a morphism $f : X \to Y$ , we must show $H (f) \circ (θ \circ η)_{X} = (θ \circ η)_{Y} \circ F (f)$ . Using $[4]$ , the left side is $H (f) \circ θ_{X} \circ η_{X}$ .

By naturality of $θ$ (equation $[3]$ for $θ$ ), $H (f) \circ θ_{X} = θ_{Y} \circ G (f)$ . Substituting gives $θ_{Y} \circ G (f) \circ η_{X}$ .

By naturality of $η$ , $G (f) \circ η_{X} = η_{Y} \circ F (f)$ . Substituting gives $θ_{Y} \circ η_{Y} \circ F (f) = (θ \circ η)_{Y} \circ F (f)$ , as required. The two naturality squares pasted side by side close into one rectangle.

Identity laws. $(id_{G} \circ η)_{X} = id_{G (X)} \circ η_{X} = η_{X} = η_{X} \circ id_{F (X)} = (η \circ id_{F})_{X}$ , so $id$ acts as a two-sided unit.

Associativity. $((ζ \circ θ) \circ η)_{X} = (ζ_{X} \circ θ_{X}) \circ η_{X} = ζ_{X} \circ (θ_{X} \circ η_{X}) = (ζ \circ (θ \circ η))_{X}$ , inheriting associativity from $D$ componentwise. $■$

:::proof[The two formulas in $[5]$ agree] Apply the naturality square $[3]$ of the natural transformation $ϵ : H \Rightarrow K$ to the specific $D$ -morphism $η_{X} : F (X) \to G (X)$ . Naturality gives

K (η_{X}) \circ ϵ_{F (X)} = ϵ_{G (X)} \circ H (η_{X}) .

This is exactly the claimed equality of the two component expressions; the common value is

(ϵ * η)_{X}

. Functoriality of

H

and

K

ensures these are morphisms

H F (X) \to K G (X)

■

Samuel Eilenberg & Saunders Mac Lane | 1909–1998, 1909–2005

Eilenberg (a Polish-American topologist) and Mac Lane (an American algebraist) jointly created category theory in a 1945 paper, General Theory of Natural Equivalences. Their motivation was strikingly concrete: in algebraic topology and homological algebra they kept encountering isomorphisms that were "natural" versus those that depended on arbitrary choices (e.g. a finite-dimensional vector space is isomorphic to its dual, but naturally isomorphic only to its double dual). To make "natural" rigorous they needed a notion of map between functors, which required defining functors, which in turn required defining categories. Eilenberg's topological instincts and Mac Lane's algebraic precision combined to yield a framework that Mac Lane later codified in the standard graduate text Categories for the Working Mathematician (1971). The pair also founded much of homological algebra (Eilenberg–MacLane spaces, the Eilenberg–Steenrod axioms).

Building Intuition

Think of a category as a "world" of objects and the legal ways to transform them. A functor is a translation of one world into another that respects the grammar of composition: if you can do $f$ then $g$ in the source, the translation of " $g$ after $f$ " equals "translation of $g$ after translation of $f$ ." Crucially a functor never invents structure — it can only carry along what composition gives.

A natural transformation is a dialect adjustment between two translations $F$ and $G$ of the same world. At each object it offers a bridge $η_{X} : F (X) \to G (X)$ . Naturality says these bridges are not arbitrary: whichever object you transform first — apply $f$ then cross the bridge, or cross the bridge then apply $f$ — you reach the same place. This uniformity is the mathematical content of the word "natural" (or "canonical"): no arbitrary choices were smuggled in, because the bridge commutes with everything simultaneously.

A useful mental image: a functor is a movie of one space deforming into another; a natural transformation is a homotopy between two such movies — and indeed under the nerve construction this analogy becomes literally true.

The naturality square: the one diagram everything grows from

Worked Examples

The free-group / abelianization functors and a natural transformation between forgetful composites

Given. Let

Grp

be the category of groups,

Ab

the category of abelian groups. Consider the functor

G = Id_{Grp}

and the functor

F = ι \circ Ab

, where

Ab : Grp \to Ab

is abelianization

H \mapsto H / [H, H]

and

ι : Ab ↪ Grp

is inclusion. We compare

Id

with

ι \circ Ab

via the quotient. Steps.

Define the component at a group $H$ as the canonical projection $η_{H} : H ↠ H / [H, H]$ . (Direction: we take $η : Id \Rightarrow ι \circ Ab$ .)
Take any homomorphism $ϕ : H \to K$ . We must verify $[3]$ : $(ι Ab) (ϕ) \circ η_{H} = η_{K} \circ ϕ$ .
The induced map $Ab (ϕ) : H / [H, H] \to K / [K, K]$ sends $h [H, H] \mapsto ϕ (h) [K, K]$ ; this is well defined because $ϕ$ carries commutators to commutators.
Trace $h \in H$ : clockwise $ϕ$ then project gives $ϕ (h) [K, K]$ ; project then induced map gives $h [H, H] \mapsto ϕ (h) [K, K]$ . They agree.

Concrete numbers. Take

H = S_{3}

(order

6

K = S_{3}

ϕ = id

. Then

[S_{3}, S_{3}] = A_{3} ≅ Z /3

, so

η_{S_{3}} : S_{3} ↠ S_{3} / A_{3} ≅ Z /2

. The transposition

(12)

maps to the nontrivial element of

Z /2

; the

3

-cycle

(123)

maps to

0

η : Id_{Grp} \Rightarrow ι \circ Ab is natural; η_{S_{3}} : S_{3} ↠ Z /2, (12) \mapsto 1, (123) \mapsto 0

Double-dual: a natural transformation that is a natural isomorphism only on finite dimensions

Given. Let

Vect_{k}

be vector spaces over a field

k

. Define

D = (-)^{**}

, the double-dual functor (covariant), and compare with

Id

. Steps.

Component at $V$ : the evaluation map $η_{V} : V \to V^{**}$ , $η_{V} (v) (φ) = φ (v)$ for $φ \in V^{*}$ .
Naturality: for linear $T : V \to W$ we need $T^{**} \circ η_{V} = η_{W} \circ T$ . Compute on $v \in V$ , evaluated at $ψ \in W^{*}$ :

(T^{**} η_{V} (v)) (ψ) = η_{V} (v) (T^{*} ψ) = (T^{*} ψ) (v) = ψ (T v) = η_{W} (T v) (ψ) .

Both sides equal $ψ (T v)$ , so $[3]$ holds — $η$ is natural.

Concrete numbers. Let

k = R

V = R^{3}

. Then

dim V = dim V^{*} = dim V^{**} = 3

, and

η_{R^{3}}

is an isomorphism; e.g.

e_{1} \mapsto ev_{e_{1}}

, the functional "extract first coordinate of a dual vector." For

V = R^{(\infty)}

(countable direct sum),

dim V^{*} = 2^{ℵ_{0}} > ℵ_{0} = dim V

, so

η_{V}

is injective but not surjective.

η_{V} : V \to V^{**} is a natural iso ⟺ dim_{k} V < \infty

Common Misconceptions

A functor is not "a map sending objects to objects." The morphism data is part of the definition and is not optional or recoverable. Two functors can agree on all objects yet differ on morphisms (e.g. the identity vs. an inner-automorphism-twisted version on a one-object groupoid). Always specify both the object and arrow assignments. "Natural" is a technical term, not an aesthetic one. A construction being "obvious" or "canonical-looking" does not make it natural; you must exhibit the functors and verify the square

[3]

. Conversely a transformation can be defined by choices yet still turn out natural if the choices cohere. Naturality is a property of the whole family, not of single components. Each

η_{X}

being an isomorphism is necessary but not sufficient for a natural isomorphism — the squares must commute for all

f

. (In fact: if

η

is natural and each component is iso, the inverses are automatically natural; but a family of componentwise isos that does not satisfy

[3]

is simply not a natural transformation at all.)

A non-natural "isomorphism": $V \cong V^*$

For finite-dimensional $V$ , choosing a basis gives an isomorphism $β_{V} : V \to V^{*}$ (send basis vectors to dual basis). Every $β_{V}$ is an isomorphism. Yet there is no natural isomorphism $Id \Rightarrow (-)^{*}$ on $Vect_{k}$ , because $(-)^{*}$ is contravariant while $Id$ is covariant — there is no naturality square to even write, and forcing one fails. Concretely, take $V = k$ and scalar multiplication $T = (\cdot c) : k \to k$ . Naturality would demand $T^{*} \circ β_{k} = β_{k} \circ T$ ; but $T^{*}$ is multiplication by $c$ on $k^{*} ≅ k$ while the required direction reverses, giving $c \cdot β_{k} = β_{k} \cdot c^{- 1}$ after careful bookkeeping — impossible for $c \neq = \pm 1$ . This is exactly the phenomenon Eilenberg–Mac Lane formalized: $V ≅ V^{*}$ is unnatural, $V ≅ V^{**}$ is natural.

Connections

Functors and natural transformations are the backbone connecting nearly every categorical concept. The Yoneda Lemma is a statement about natural transformations out of representable functors $Hom (X, -)$ , showing $Nat (Hom (X, -), F) ≅ F (X)$ — naturality is its entire engine. Adjoint Functors are pairs $F ⊣ G$ specified by a natural isomorphism $Hom_{D} (F -, -) ≅ Hom_{C} (-, G -)$ ; the unit and counit are natural transformations $Id \Rightarrow GF$ and $F G \Rightarrow Id$ . A Monad is an endofunctor $T$ equipped with natural transformations $η : Id \Rightarrow T$ and $μ : T^{2} \Rightarrow T$ satisfying coherence. Limits and Colimits are objects representing the functor of cones, where a cone is itself a natural transformation from a constant functor. The functor category $[C, Set]$ is the Presheaf Category, the central object of topos theory and the natural habitat of the Yoneda embedding. The 2-categorical structure $Cat$ from $[6]$ launches Higher Category Theory.

Real-World Applications

Algebraic topology — homology and cohomology. Singular homology $H_{n} : Top \to Ab$ is a functor; the long exact sequence and connecting homomorphisms are natural transformations, and the Eilenberg–Steenrod axioms are phrased entirely functorially.
Functional programming (Haskell, Scala). The Functor typeclass with fmap is the categorical functor on the category of types; Applicative and Monad use natural transformations (pure, join) and the interchange law $[6]$ governs how effects compose.
Database theory. Spivak's functorial data model treats a schema as a category $C$ and a database instance as a functor $C \to Set$ ; data migration between schemas is composition with a functor, and consistency constraints are natural transformations.
Representation theory. Induction and restriction functors $Ind, Res$ between $Rep (H)$ and $Rep (G)$ form adjoint pairs; Frobenius reciprocity is a natural isomorphism of hom-sets.
Quantum computing / TQFT. A topological quantum field theory is a (symmetric monoidal) functor from a cobordism category to $Vect$ ; physical equivalences of theories are natural transformations.
Logic & type theory. The Curry–Howard–Lambek correspondence makes proofs natural transformations and cut-elimination compatible with naturality; parametric polymorphism corresponds (Reynolds, Wadler's "Theorems for free") to naturality conditions.
Machine learning — categorical/structured deep learning. Recent work models neural network architectures as functors and weight-tying / equivariance constraints as natural transformations (e.g. group-equivariant CNNs).

Historical Development

1942 — Eilenberg & Mac Lane publish Group Extensions and Homology, where the need to express "natural" isomorphisms first crystallizes in homological computations.
1945 — General Theory of Natural Equivalences (Trans. AMS) formally introduces categories, functors, and natural transformations; the trio is born explicitly to define "natural."
1950s — Functorial methods spread through algebraic topology; the Eilenberg–Steenrod axioms (1952) recast homology theories as functors with natural transformations.
1957 — Grothendieck's Tôhoku paper develops abelian categories and derived functors, making functorial language indispensable in homological algebra.
1958 — Daniel Kan introduces adjoint functors and Kan extensions, defined via universal natural transformations; "all concepts are Kan extensions."
1958–1963 — Lawvere's program reformulates set theory and algebraic theories functorially; functorial semantics treats models as functors.
1964 — Yoneda's lemma (named by Mac Lane after Nobuo Yoneda's 1954 observations) becomes the cornerstone result about natural transformations.
1965 — Eilenberg & Kelly formalize enriched and 2-categorical structure; the interchange law $[6]$ is recognized as defining 2-categories.
1971 — Mac Lane's Categories for the Working Mathematician canonizes the definitions and proofs used to this day.

Subtleties & Limits

Size and foundations. The functor category

[C, D]

is well-behaved when

C

is small; for large

C

Nat (F, G)

may be a proper class, and one must invoke Grothendieck universes or a careful set/class distinction. The "category of all categories" is not naively a category for this reason. Strictness vs. coherence. Equation

[2]

demands functoriality on the nose. In higher category theory one weakens this to functoriality up to coherent natural isomorphism (pseudofunctors, lax functors), and the management of these coherence isomorphisms (Mac Lane's coherence theorem, and its open higher analogues) becomes deeply nontrivial. Whether every weak

\infty

-functor can be strictified is delicate and false in general beyond low dimensions. Naturality vs. dinaturality and parametricity. For functors of mixed variance

T (X^{op}, X)

, ordinary naturality fails and one needs dinatural transformations, whose composites do not always compose — a known pathology with no fully satisfactory general fix. Relating syntactic parametricity (Reynolds) to genuine (di)naturality in all type-theoretic models remains an active research interface between logic and category theory.