Limits, Colimits, and Kan Extensions

Limits and colimits are the universal cones and cocones over diagrams in a category, unifying products, equalizers, pullbacks, coproducts, coequalizers, and pushouts. Kan extensions are the most general universal construction in category theory — left and right adjoints to restriction along a functor — and "all concepts are Kan extensions" (Mac Lane). Applies across algebra, topology, logic, and theoretical computer science. Formalized by Daniel Kan (1958), Grothendieck, and the Yoneda school in the mid-20th century.

A limit of a diagram $D : J \to C$ is an object $lim D$ equipped with a universal cone, representing the functor sending $X$ to the set of cones over $D$ with apex $X$ ; colimits are the dual notion via cocones. These constructions package "the best simultaneous solution" to a system of constraints encoded by a diagram, subsuming nearly every gluing, g;quotient, or product construction in mathematics. Kan extensions generalize this further: given $K : A \to B$ and $F : A \to C$ , they extend $F$ along $K$ in a universally optimal way, and limits/colimits are merely Kan extensions along the functor to the terminal category. The surprising fact: every adjunction, every limit, every colimit, every notion of "induced representation," "sheafification," and "free construction" is a Kan extension — they are the atoms from which categorical universal constructions are built.

The Core Insight

A universal property says an object is determined up to unique isomorphism by how it maps to or from everything else. Limits collect a diagram into a single object via a terminal cone; colimits via an initial cocone. Kan extensions are the universal way to "transport" a functor along another functor, and they reduce to limits/colimits computed pointwise via (co)comma categories. The slogan "all concepts are Kan extensions" reflects that representability, adjunctions, and (co)limits are all instances of one construction.

Formal Statement

Diagram, Cone, and Cone Category

Let $J$ (the index or shape category) be small and $C$ a category. A diagram of shape $J$ is a functor $D : J \to C$ . For an object $X \in C$ , write $Δ X : J \to C$ for the constant diagram sending every object to $X$ and every morphism to $id_{X}$ .

A cone over $D$ with apex $X$ is a natural transformation $λ : Δ X \Rightarrow D$ , i.e. a family of morphisms $λ_{j} : X \to D (j)$ such that for every $f : j \to j^{'}$ in $J$ ,

D (f) \circ λ_{j} = λ_{j^{'}} .

⟳Naturality of

λ

on the morphism

f

gives the square; the

Δ X

side is the identity, collapsing it to a triangle.

A cocone under $D$ with nadir $Y$ is dually a natural transformation $μ : D \Rightarrow Δ Y$ , with components $μ_{j} : D (j) \to Y$ satisfying $μ_{j^{'}} \circ D (f) = μ_{j}$ .

Limit and Colimit

A limit of $D$ is a cone $(L, π : Δ L \Rightarrow D)$ that is terminal among cones: for any cone $(X, λ)$ there is a unique $u : X \to L$ with $π_{j} \circ u = λ_{j}$ for all $j$ . We write $L = lim D$ or $lim D$ . Equivalently, the limit represents the cone functor:

C (X, lim D) ≅ Cone (X, D) = Nat (Δ X, D),

naturally in

X

⟳Yoneda — a representing object for the functor

X \mapsto Nat (Δ X, D)

is precisely a terminal cone; the universal element is the limiting cone

π

A colimit $colim D = lim D$ is the initial cocone, with the dual representation

C (colim D, Y) ≅ Nat (D, Δ Y) .

⟳Dualize [2] by reversing every arrow and passing to

C^{op}

; a colimit in

C

is a limit in

C^{op}

Kan Extension

Let $K : A \to B$ and $F : A \to C$ be functors. The left Kan extension of $F$ along $K$ , written $Lan_{K} F : B \to C$ , is a functor equipped with a natural transformation $η : F \Rightarrow (Lan_{K} F) \circ K$ that is initial: for any $(G, γ : F \Rightarrow G K)$ there is a unique $α : Lan_{K} F \Rightarrow G$ with $(α K) \cdot η = γ$ . Equivalently, $Lan_{K}$ is left adjoint to restriction $K^{*} = (-) \circ K$ :

Nat (Lan_{K} F, G) ≅ Nat (F, G \circ K) naturally in G \in [B, C] .

⟳The defining universal property is exactly the unit/counit description of the adjunction

Lan_{K} ⊣ K^{*}

;

η

is the unit.

The right Kan extension $Ran_{K} F$ is right adjoint to $K^{*}$ , with $ε : (Ran_{K} F) \circ K \Rightarrow F$ terminal and

Nat (G, Ran_{K} F) ≅ Nat (G \circ K, F) .

⟳Right adjoint to

K^{*}

; dual of [4]. Both adjoints exist when

C

has enough (co)limits, by the pointwise formula below.

Pointwise Kan Extension Formula

Suppose $A$ is small and $C$ is cocomplete (resp. complete). Then $Lan_{K} F$ (resp. $Ran_{K} F$ ) exists and is computed pointwise by a colimit (resp. limit) over a comma category:

(Lan_{K} F) (b) = colim ((K ↓ b) proj A F C),

⟳The comma category

K ↓ b

has objects

(a, K a \to b)

; the colimit glues the values

F a

weighted by all maps into

b

. Verify the universal property [4] by Fubini for colimits / coend manipulation.

(Ran_{K} F) (b) = lim ((b ↓ K) proj A F C) .

⟳Dual to [6] using the comma category

b ↓ K

of objects

(a, b \to K a)

; this is a limit in

C

These admit the elegant (co)end expressions, valid in $Set$ -enriched or general $V$ -enriched contexts:

(Lan_{K} F) (b) ≅ \int^{a \in A} B (K a, b) \cdot F a, (Ran_{K} F) (b) ≅ \int_{a \in A} F a^{B (b, K a)} .

⟳

S \cdot c

is the

S

-fold copower and

c^{S}

the

S

-fold power; the coend over

a

replaces the comma-category colimit by a tensor weighted with the hom-set

B (K a, b)

. The Fubini theorem for ends justifies rewriting [6] as a coend.

Limits and Colimits as Kan Extensions

Let $1$ be the terminal category and $t : J \to 1$ the unique functor. For a diagram $D : J \to C$ ,

lim D ≅ (Ran_{t} D) (*), colim D ≅ (Lan_{t} D) (*) .

⟳The comma category

* ↓ t ≅ J

(resp.

t ↓ * ≅ J

), so [6]–[7] collapse to a colimit/limit over all of

J

. This is the precise sense in which (co)limits

are Kan extensions.

Why It Holds / Derivation

Pointwise Left Kan Extension Satisfies the Adjunction

We verify [4] using the pointwise definition [6]. Fix $G : B \to C$ .

Step 1 — Components. Define

(Lan_{K} F) (b) = colim_{(a, ϕ : K a \to b)} F a

, with colimit injections

ι_{a, ϕ} : F a \to (Lan_{K} F) (b)

. Functoriality in

b

comes from postcomposition

ϕ \mapsto g ϕ

for

g : b \to b^{'}

, inducing

(Lan_{K} F) (g)

. Step 2 — The unit. For

a \in A

, the object

(a, id_{K a})

lives in

K ↓ K a

, giving

η_{a} := ι_{a, id} : F a \to (Lan_{K} F) (K a)

. Naturality of

η

follows from compatibility of injections. Step 3 — From $γ$ to $α$ . Given

γ : F \Rightarrow G K

, define

α_{b} : (Lan_{K} F) (b) \to G b

on the cocone by

α_{b} \circ ι_{a, ϕ} = G (ϕ) \circ γ_{a} : F a \to G (K a) \to G b .

⟳The right-hand side is a cocone under the diagram

a \mapsto F a

(check compatibility using naturality of

γ

and functoriality of

G

); the universal property of the colimit produces the unique

α_{b}

Step 4 — Triangle identity. Setting

ϕ = id_{K a}

in [10] gives

α_{K a} \circ η_{a} = G (id) \circ γ_{a} = γ_{a}

, i.e.

(α K) \cdot η = γ

. Step 5 — Uniqueness. Any

α^{'}

with

(α^{'} K) \cdot η = γ

must satisfy [10] on every injection (decompose

ι_{a, ϕ}

through

ι_{a, id}

and naturality), and a map out of a colimit is determined by its restriction to the cocone. Hence

α^{'} = α

. This establishes the bijection [4], natural in

G

■

Limits Are Right Kan Extensions Along $t : \mathsf{J}\to\mathsf{1}$

Apply [5] with $K = t$ , $B = 1$ , $b = *$ . A functor $G : 1 \to C$ is just an object $X = G (*)$ , and $G \circ t = Δ X$ . Thus

C (X, (Ran_{t} D) (*)) ≅ Nat (G, Ran_{t} D) ≅ Nat (Gt, D) = Nat (Δ X, D) .

⟳First iso is Yoneda in

[1, C] ≃ C

; second is the

Ran

adjunction [5]; the last equality unfolds

Gt = Δ X

Comparing with the limit representation [2], $(Ran_{t} D) (*)$ represents the same functor $X \mapsto Nat (Δ X, D)$ , hence is canonically $lim D$ . The colimit case is dual via [4]. $■$

Daniel M. Kan | 1927–2013

Born Daniel Kan in Amsterdam, he survived WWII and emigrated to Israel, completing his doctorate under Samuel Eilenberg's influence. In his 1958 paper "Adjoint Functors" (Trans. AMS) he simultaneously introduced adjoint functors and the extension functors now bearing his name. Kan's instinct was profoundly homotopical: he developed simplicial methods, the Kan condition on simplicial sets (fibrant objects), Kan loop groups, and was a central architect of modern homotopy theory at MIT. The Kan extension was his tool for transporting homotopy-theoretic structure between categories; Mac Lane later elevated the slogan "all concepts are Kan extensions" to a guiding principle of Categories for the Working Mathematician (1971). Kan's combination of abstract categorical machinery with concrete homotopy computation set the template for the field.

Building Intuition

Think of a diagram $D$ as a system of compatible constraints. A cone over $D$ is a candidate "solution" mapping consistently into every constraint; the limit is the most efficient solution — every other solution factors uniquely through it. For products this is "carry all coordinates at once"; for equalizers it is "the largest subobject on which two maps agree"; for pullbacks it is "matched pairs." Dually, the colimit is the most economical gluing: a coproduct is disjoint union, a coequalizer identifies as little as forced, a pushout glues two objects along a shared piece.

For Kan extensions, picture $K : A \to B$ as a "change of base." You know $F$ on $A$ and want a functor on the larger $B$ . The left Kan extension is the freest extension — it builds $(Lan_{K} F) (b)$ by gluing all the data $F a$ that maps into $b$ via $K$ (a colimit). The right Kan extension is the cofreest — it assembles $F a$ for all maps out of $b$ (a limit). The comma categories $K ↓ b$ and $b ↓ K$ are precisely "everything in $A$ that sees $b$ from below / from above."

Reading for the working scientist

If you think of $J$ as a spacetime lattice and $D$ as a field configuration assigning data to each site, the limit is a global section consistent with all local restriction maps — like a solution to gauge constraints. Colimits are coarse-grainings: gluing local field data along overlaps (a pushout) is precisely how one builds a manifold from charts or a fiber bundle by clutching. Kan extension is the categorical version of induced and coinduced representations in physics: extend a symmetry from a subgroup $H$ to $G$ — induction is a left Kan extension, coinduction a right Kan extension along $B H \to B G$ .

Limit as terminal cone and colimit as initial cocone over a diagram $D$

<svg viewBox="0 0 760 360" xmlns="http://www.w3.org/2000/svg">
  <rect width="760" height="360" fill="#0a0e17"/>
  <text x="30" y="30" fill="#e2e8f0" font-size="16">Limit (terminal cone)</text>
  <!-- diagram objects -->
  <circle cx="180" cy="180" r="6" fill="#94a3b8"/>
  <text x="190" y="184" fill="#e2e8f0" font-size="14">D(j)</text>
  <circle cx="300" cy="120" r="6" fill="#94a3b8"/>
  <text x="310" y="118" fill="#e2e8f0" font-size="14">D(j')</text>
  <circle cx="300" cy="250" r="6" fill="#94a3b8"/>
  <text x="310" y="262" fill="#e2e8f0" font-size="14">D(j'')</text>
  <line x1="186" y1="176" x2="294" y2="124" stroke="#94a3b8" stroke-width="1.5"/>
  <line x1="186" y1="184" x2="294" y2="244" stroke="#94a3b8" stroke-width="1.5"/>
  <!-- apex limit -->
  <circle cx="70" cy="180" r="7" fill="#d4a853"/>
  <text x="20" y="160" fill="#d4a853" font-size="15">L = lim D</text>
  <line x1="77" y1="178" x2="174" y2="180" stroke="#d4a853" stroke-width="2"/>
  <line x1="76" y1="174" x2="294" y2="122" stroke="#d4a853" stroke-width="2"/>
  <line x1="76" y1="186" x2="294" y2="248" stroke="#d4a853" stroke-width="2"/>
  <text x="120" y="172" fill="#d4a853" font-size="12">π_j</text>
  <!-- competing cone X -->
  <circle cx="60" cy="300" r="6" fill="#94a3b8"/>
  <text x="20" y="320" fill="#e2e8f0" font-size="13">X (any cone)</text>
  <path d="M64 295 Q70 240 70 187" stroke="#94a3b8" stroke-width="1.5" fill="none" stroke-dasharray="4 3"/>
  <text x="74" y="250" fill="#94a3b8" font-size="12">∃! u</text>

  <text x="430" y="30" fill="#e2e8f0" font-size="16">Colimit (initial cocone)</text>
  <circle cx="540" cy="180" r="6" fill="#94a3b8"/>
  <text x="490" y="184" fill="#e2e8f0" font-size="14">D(j)</text>
  <circle cx="540" cy="120" r="6" fill="#94a3b8"/>
  <text x="490" y="118" fill="#e2e8f0" font-size="14">D(j')</text>
  <circle cx="540" cy="250" r="6" fill="#94a3b8"/>
  <text x="490" y="262" fill="#e2e8f0" font-size="14">D(j'')</text>
  <circle cx="700" cy="180" r="7" fill="#d4a853"/>
  <text x="640" y="160" fill="#d4a853" font-size="15">colim D</text>
  <line x1="546" y1="180" x2="693" y2="180" stroke="#d4a853" stroke-width="2"/>
  <line x1="546" y1="122" x2="694" y2="176" stroke="#d4a853" stroke-width="2"/>
  <line x1="546" y1="248" x2="694" y2="184" stroke="#d4a853" stroke-width="2"/>
  <text x="610" y="150" fill="#d4a853" font-size="12">ι_j</text>
</svg>

Worked Examples

Pullback in $\mathbf{Set}$ — a concrete limit

Given. Let

A = {1, 2, 3}

B = {x, y}

C = {u, v}

, with maps

f : A \to C

f (1) = f (2) = u, f (3) = v

, and

g : B \to C

g (x) = u, g (y) = v

. The shape

J

is the cospan

∙ \to ∙ \leftarrow ∙

. Steps.

The pullback $A \times_{C} B$ is the limit, computed as ${(a, b) : f (a) = g (b)}$ — a set of matched pairs.
Enumerate: $(1, x)$ since $f (1) = u = g (x)$ ; $(2, x)$ likewise; $(3, y)$ since $f (3) = v = g (y)$ . No others: $f (a) = u$ forces $b = x$ , $f (a) = v$ forces $b = y$ .
Projections $π_{A} (a, b) = a$ , $π_{B} (a, b) = b$ form the limiting cone, and the square commutes: $f π_{A} = g π_{B}$ .

Answer.

A \times_{C} B = {(1, x), (2, x), (3, y)}, ∣ A \times_{C} B ∣ = 3.

Left Kan extension computes the free monoid / induced functor (numerical)

Given. Let

A

be the discrete category on one object

{a}

B

the two-object discrete category

{b_{1}, b_{2}}

K : A \to B

sending

a \mapsto b_{1}

, and

F : A \to Set

with

F (a) = S

∣ S ∣ = 4

. Target

Set

is cocomplete. Steps.

Use the pointwise formula [6]: $(Lan_{K} F) (b) = colim_{(a, ϕ : K a \to b)} F a$ .
At $b_{1}$ : comma category $K ↓ b_{1}$ has one object $(a, id_{b_{1}})$ since the only map $K a = b_{1} \to b_{1}$ is the identity (discrete). Colimit over a one-object discrete category is just $F a = S$ . So $(Lan_{K} F) (b_{1}) = S$ , size $4$ .
At $b_{2}$ : $K ↓ b_{2}$ has objects $(a, ϕ : b_{1} \to b_{2})$ , but $B$ discrete means no morphism $b_{1} \to b_{2}$ . The comma category is empty; colimit over the empty diagram is the initial object $\emptyset$ .

Answer.

(Lan_{K} F) (b_{1}) = S (∣ \cdot ∣ = 4), (Lan_{K} F) (b_{2}) = \emptyset (∣ \cdot ∣ = 0) .

This is the categorical "extension by zero" — left Kan extension along an inclusion fills the new spot with the initial object.

Common Misconceptions

Functors do not automatically preserve (co)limits. A functor

G

preserves limits iff it maps limiting cones to limiting cones; this is a theorem to verify, not a definition. Key fact: right adjoints preserve limits (RAPL) and left adjoints preserve colimits (LAPC), but the converses fail without further hypotheses (need a solution-set condition — Freyd's Adjoint Functor Theorem). The forgetful functor

Grp \to Set

preserves limits (it has a left adjoint, the free group functor) but does not preserve colimits: the coproduct of groups is the free product, whose underlying set is not the disjoint union. Kan extensions need not be "pointwise." The pointwise formulas [6]–[7] require

C

to have the relevant (co)limits. A Kan extension defined only by the universal property [4]/[5] may exist without being computable pointwise (in enriched or higher-categorical settings the distinction is genuine). In ordinary

Set

-categories with

C

(co)complete the two coincide, but never assume the comma-category formula holds by default.

A non-existent limit

Limits need not exist. In the category $Field$ of fields, the product of $Q$ and $F_{2}$ does not exist: any field receiving maps to both would need characteristic simultaneously $0$ and $2$ , which is impossible since there is no field of mixed characteristic, and $Field$ has no terminal object. Thus the empty diagram (whose limit is the terminal object) already fails, and binary products fail too. Completeness is a strong condition, not automatic.

Connections

Limits and colimits are the load-bearing structure beneath nearly all of category theory. They are special weighted limits in the Enriched Category Theory sense, and the (co)end formulas [8] live there naturally. Via the Yoneda Lemma, every presheaf is canonically a colimit of representables (the density theorem), which is itself a Kan extension statement: the identity functor on presheaves is $Lan_{よ} よ$ . The relation $Lan_{K} ⊣ K^{*} ⊣ Ran_{K}$ exhibits restriction as having both adjoints, tying directly to Adjoint Functors. The pointwise formula is the categorical avatar of the Coend Calculus, and in homotopical settings these become homotopy (co)limits and derived Kan extensions, the foundation of Model Categories and $\infty$ -category theory.

Real-World Applications

Sheafification. The associated-sheaf functor on a site is a left Kan extension along the inclusion of a basis, built from colimits over covering sieves — central to algebraic geometry (Grothendieck topologies).
Induced representations. For groups $H \leq G$ , $Ind_{H}^{G}$ and $Coind_{H}^{G}$ are the left and right Kan extensions along $B H ↪ B G$ ; Frobenius reciprocity is the Kan adjunction [4]/[5].
Databases (functorial data migration). Spivak's model treats schemas as categories and instances as functors $S \to Set$ ; the migration operators $Σ$ (union/merge), $Π$ (join), $Δ$ (project) are exactly $Lan$ , $Ran$ , and restriction.
Type theory & programming. In Haskell, Codensity, free monads, and "right Kan extension" Ran are used to optimize and define monad transformers; limits/colimits model product/sum types and recursion schemes.
Topological data analysis. Persistent homology is a functor from a poset (filtration) to vector spaces; interleavings and the structure theorem rest on (co)limits in this functor category.
Quantum field theory. Functorial/TQFT axioms (Atiyah–Segal) glue spacetimes via colimits (pushouts of cobordisms), and extended TQFTs use left Kan extensions to descend from cobordisms.
Logic and semantics. Quantifiers $\exists$ and $\forall$ are left and right adjoints to substitution (Lawvere), hence Kan extensions along projection functors of indexed categories.

Historical Development

1942–45 — Eilenberg & Mac Lane introduce categories, functors, and natural transformations, providing the language in which (co)limits will be phrased.
1948 — Mac Lane formulates products and coproducts via universal properties, foreshadowing the general limit notion.
1957 — Grothendieck's Tôhoku paper systematizes abelian categories, exactness, and the existence of limits/colimits (his axioms AB3–AB6), making colimits a working tool in homological algebra.
1958 — Daniel Kan, "Adjoint Functors", introduces adjoints and the extension functors $Lan$ , $Ran$ , conceived for homotopy theory.
1960 — Freyd develops the Adjoint Functor Theorem, connecting completeness, (co)limits, and the existence of adjoints (hence Kan extensions).
1964 — Lawvere's thesis reframes algebraic theories and recognizes quantifiers as adjoints, an early Kan-extension viewpoint in logic.
1965 — Yoneda's lemma and the density theorem are crystallized, showing presheaves as colimits of representables.
1971 — Mac Lane's Categories for the Working Mathematician canonizes the slogan "all concepts are Kan extensions" and the comma/coend formulas.
1982 — Kelly's Basic Concepts of Enriched Category Theory gives the definitive enriched/weighted treatment of limits and Kan extensions via ends and coends.
2009 — Lurie's Higher Topos Theory develops $\infty$ -categorical limits, colimits, and Kan extensions, now foundational in derived geometry and stable homotopy.

Subtleties & Limits

The passage from ordinary to higher-categorical Kan extensions is subtle and still actively developed. Pointwise formulas in $(\infty, 1)$ -categories require homotopy colimits over $\infty$ -comma categories, and ensuring functoriality/coherence is technically demanding (handled via straightening–unstraightening or via $\infty$ -cosmoi). Beyond $(\infty, 1)$ , fully lax and oplax Kan extensions, and Kan extensions in double categories and bicategories, lack a uniformly clean theory; existence and pointwise-computability can fail or split into inequivalent notions. Open programmatic questions include the right framework for derived Kan extensions in motivic and condensed mathematics, and effective/computable presentations of Kan extensions for use in proof assistants and database theory where size and decidability obstructions (solution-set conditions, accessibility) become practically binding.