Category Theory
Natural Transformations: Morphisms Between Functors
Category theory was, by the confession of its founders Samuel Eilenberg and Saunders Mac Lane in 1945, invented not to study categories, nor even functors, but to make precise the word natural — as in "there is a natural isomorphism between a finite-dimensional vector space and its double dual." A natural transformation is the exact mathematical content of that word: a rule that transforms one functor into another uniformly, commuting with every morphism in sight.
Precisely: given two functors F, G : 𝒞 → 𝒟, a natural transformation η : F ⇒ G assigns to each object X of 𝒞 a morphism η_X : F(X) → G(X) in 𝒟 (its component at X), such that for every morphism f : X → Y the square G(f) ∘ η_X = η_Y ∘ F(f) commutes. When each η_X is an isomorphism, η is a natural isomorphism, and F and G are "the same up to a coherent relabeling."
- FieldCategory theory
- Introduced byEilenberg & Mac Lane, 1945
- Defining conditionNaturality square commutes: G(f)∘η_X = η_Y∘F(f)
- DataA component η_X : F(X) → G(X) for every object X
- Special caseNatural isomorphism (every η_X invertible)
- UnlocksFunctor categories, the Yoneda lemma, adjunctions, (co)limits
Interactive visualization
Press play, or step through manually. The visualization is yours to drive — try it before reading on.
Watch the 60-second explainer
A condensed visual walkthrough — narrated, captioned, under a minute.
The precise definition and the naturality square
Fix categories 𝒞 and 𝒟 and two functors F, G : 𝒞 → 𝒟. A natural transformation η : F ⇒ G consists of:
- for each object X ∈ 𝒞, a morphism η_X : F(X) → G(X) in 𝒟, called the component of η at X;
- subject to the naturality condition: for every morphism f : X → Y in 𝒞,
G(f) ∘ η_X = η_Y ∘ F(f).
Pictorially this is the commuting square with top edge η_X : F(X) → G(X), bottom edge η_Y : F(Y) → G(Y), left edge F(f), and right edge G(f). Both paths F(X) → G(Y) agree. That is the entire definition — no equations on objects, only this one family of squares.
An isomorphism in the sense above is a natural transformation each of whose components η_X is invertible in 𝒟; one then checks the inverses (η_X)⁻¹ automatically assemble into a natural transformation G ⇒ F. Two functors related by a natural isomorphism are naturally isomorphic, written F ≅ G.
The intuition: uniformity without arbitrary choices
A functor F sends each object X to F(X); a natural transformation is a way to slide the entire image of F over to the image of G, one object at a time, in a manner that respects all the arrows connecting them. The naturality square is the demand that this sliding be coherent: whatever you do to X, then transform, equals transforming first, then doing the same thing to Y.
The word to keep in mind is uniform. Anyone can rig up an isomorphism V ≅ V* between a finite-dimensional space and its dual — but only by choosing a basis, and different bases give incompatible isomorphisms. There is no way to make the choice commute with all linear maps at once. Naturality is precisely the condition that no arbitrary choice was made: the family {η_X} works simultaneously and compatibly for every object, forced by the structure rather than picked by hand. This is why 'natural' is a technical theorem-grade word, not a vibe.
The mechanism: functor categories and horizontal/vertical composition
The key structural idea is that natural transformations are themselves the morphisms of a new category. Given 𝒞 and 𝒟, form the functor category 𝒟^𝒞 (also written Fun(𝒞,𝒟) or [𝒞,𝒟]): its objects are functors 𝒞 → 𝒟 and its morphisms are natural transformations between them.
Vertical composition: given η : F ⇒ G and θ : G ⇒ H, define (θ∘η)_X = θ_X ∘ η_X. Naturality of the composite follows by pasting the two commuting squares side by side — the shared middle edge cancels. The identity natural transformation has components id_{F(X)}.
There is also horizontal composition (for η : F⇒G with F,G : 𝒞→𝒟 and η' : F'⇒G' with F',G' : 𝒟→ℰ), and the two compositions satisfy the interchange law. Together these make Cat into a 2-category: categories, functors, and natural transformations are its 0-, 1-, and 2-cells. The whole edifice — Yoneda, adjunctions, limits — is expressed in the language of these functor categories.
Canonical example: the double dual V ≅ V**
Let 𝒞 = 𝒟 = FinVect_k, finite-dimensional vector spaces over a field k. Let F be the identity functor and G = (−)** the double-dual functor, V ↦ V** = (V*)*. Define the component at V by evaluation:
η_V : V → V**, η_V(v)(φ) = φ(v) for φ ∈ V*.
Each η_V is an isomorphism precisely when dim V < ∞. Naturality: for a linear map f : V → W we need (f**) ∘ η_V = η_W ∘ f. Chasing an element, both sides send v to the functional φ ↦ φ(f(v)) on W* — they agree, so the square commutes for every f. This is a genuine natural isomorphism id ≅ (−)**.
Contrast the single dual: V ≅ V* holds dimensionwise but there is no natural isomorphism id ≅ (−)*, because (−)* is contravariant and, more tellingly, any attempted iso needs a basis and fails the square for a generic f. The double dual is 'natural'; the single dual is merely 'isomorphic'.
Why the hypotheses bite, and the Yoneda connection
Every clause of the definition earns its keep. Drop the naturality square and you have merely an object-indexed family of morphisms — an 'unnatural' family, exactly the basis-dependent V ≅ V* that Eilenberg–Mac Lane built the theory to outlaw. Require the components to be isomorphisms and you get natural isomorphism; without invertibility you keep genuinely directional transformations (e.g. the unit X → GF(X) of an adjunction, generally not invertible).
The deepest payoff is the Yoneda lemma: for a functor F : 𝒞 → Set and object A, natural transformations Hom(A,−) ⇒ F correspond bijectively to elements of F(A), the bijection being η ↦ η_A(id_A). Its corollary, the Yoneda embedding 𝒞 ↪ [𝒞ᵒᵖ, Set], is fully faithful: an object is completely determined by the natural transformations into it. Naturality is the exact hypothesis that makes this work — relax it and the correspondence collapses. It also underlies representability, adjunctions (hom-set isos natural in both variables), and the definition of (co)limits as universal cones, which are themselves natural transformations from a constant functor.
Why it matters: the reason category theory exists
Natural transformations are not one tool among many — they are the raison d'être of the subject. Eilenberg and Mac Lane (1945) introduced functors solely so they could define natural transformations, and introduced categories solely so they could define functors. Mac Lane later joked that categories were defined 'in order to define functors, and functors in order to define natural transformations.'
Concretely they let you state, as theorems with content, claims like:
- Adjunctions: Hom_𝒟(F X, Y) ≅ Hom_𝒞(X, G Y) naturally in X and Y — the naturality is what makes an adjunction rigid and unique.
- Limits and colimits: a limit is a universal cone, i.e. a terminal natural transformation from a constant functor to a diagram.
- Homology and cohomology: the connecting homomorphisms and the comparison of theories are natural transformations; naturality of the long exact sequence is what lets you glue local computations.
- Change-of-basis, Fourier duality, Stone duality: each 'canonical isomorphism' in mathematics is really a claim of naturality.
In short, whenever a mathematician says a construction is 'canonical' or 'coordinate-free', the precise content is: it is a natural transformation.
| Concept | Lives between | Data | Compatibility law |
|---|---|---|---|
| Morphism f : X → Y | Objects of a category 𝒞 | A single arrow | Composition & identity axioms |
| Functor F : 𝒞 → 𝒟 | Categories | Object map + arrow map | F(g∘f)=F(g)∘F(f), F(id)=id |
| Natural transformation η : F ⇒ G | Parallel functors 𝒞 → 𝒟 | One component η_X per object X | Naturality square G(f)∘η_X = η_Y∘F(f) |
| Natural isomorphism | Parallel functors 𝒞 → 𝒟 | Components η_X, all invertible | Naturality + each η_X an iso; inverse is also natural |
| Modification | Parallel natural transformations | Component per object (2-categorical) | Higher coherence squares |
Frequently asked questions
What exactly is the naturality condition, in one line?
For every morphism f : X → Y in the source category, the square commutes: G(f) ∘ η_X = η_Y ∘ F(f). Equivalently, transforming an object and then applying G(f) gives the same result as applying F(f) and then transforming. This single family of commuting squares — one per arrow — is the entire content beyond having a component η_X : F(X) → G(X) at each object.
Why is V ≅ V* not natural but V ≅ V** is?
The double-dual map η_V(v) = (φ ↦ φ(v)) is defined without choosing a basis and satisfies the naturality square for every linear map f, giving a natural isomorphism id ≅ (−)**. The single dual V ≅ V* requires choosing a basis (or inner product); different choices give incompatible isomorphisms, and no choice makes the naturality square commute for all f. So V* is isomorphic to V dimensionwise but not naturally isomorphic — that gap is exactly what 'natural' measures.
Do natural transformations require the components to be isomorphisms?
No. A natural transformation only needs a component morphism η_X : F(X) → G(X) at each object plus naturality; the components can be arbitrary arrows. When every component happens to be invertible you get a natural isomorphism. Many important natural transformations are not isomorphisms — for instance the unit η : id ⇒ GF and counit ε : FG ⇒ id of an adjunction are generally non-invertible.
How do natural transformations compose?
Vertically: given η : F ⇒ G and θ : G ⇒ H, set (θ∘η)_X = θ_X ∘ η_X; naturality follows by pasting the two squares. This makes functors 𝒞 → 𝒟 into a category, the functor category [𝒞, 𝒟]. There is also horizontal composition of transformations between composable functors, and the two obey the interchange law, making Cat a 2-category.
What is the Yoneda lemma's relationship to natural transformations?
The Yoneda lemma says that for F : 𝒞 → Set, the natural transformations Hom(A, −) ⇒ F are in bijection with the set F(A), via η ↦ η_A(id_A). This is astonishing: an entire coherent family of maps is pinned down by a single element. It yields the Yoneda embedding, which is fully faithful, so any object is determined up to isomorphism by the natural transformations into it — the cornerstone of representability and much of modern category theory.
Who introduced natural transformations and why?
Samuel Eilenberg and Saunders Mac Lane introduced them in their 1945 paper 'General Theory of Natural Equivalences.' Their explicit goal was to make the informal word 'natural' — as used in phrases like 'the natural isomorphism of a space with its double dual' — into a precise mathematical notion. Functors and categories were defined as the scaffolding needed to state the definition of a natural transformation, which was the actual target of the theory.