nLab adjoint functor

Redirected from "universal morphisms".
Contents

(The special case in Cat of the general notion of adjunction.)

Contents

Idea

The concept of adjoint functors [Kan (1958)] is a key concept in category theory β€” if not the key concept β€” and it is in large part through the manifold identification of examples of adjoint functors appearing ubiquituously in the practice of mathematics that category theoretic tools are brought to use in general mathematics.

Abstractly, the notion of adjoint functors embodies the concept of representable functors and has as special cases the fundamental universal constructions of category theory such as notably Kan extensions and hence (co-)limits and (co-)ends, while being itself the archetypical special case of a natural notion of adjunction which in 2-category theory embodies a general principle of duality.

Concretely, the concept of adjoint functors L⊣R:π’Ÿβ‡†π’žL \dashv R \,\colon\, \mathcal{D} \leftrightarrows \mathcal{C} is immediately transparent and compelling in its incarnation as a natural isomorphism on hom-sets (see below) and more generally on hom-objects (see at enriched adjoint functor), where it just says that adjoint functors are those that may coherently be switched left↔\leftrightarrowright in a hom-set

L⊣R:π’Ÿβ‡†RLπ’žmeansπ’Ÿ(L(βˆ’),βˆ’)β‰ƒπ’ž(βˆ’,R(βˆ’)). L \dashv R \;\colon\; \mathcal{D} \underoverset{R}{L}{\leftrightarrows} \mathcal{C} \;\;\;\;\;\;\;\;\;\;\;\;\; \text{means} \;\;\;\;\;\;\;\;\;\;\;\;\; \mathcal{D}\bigl(L(-),\,-\bigr) \;\simeq\; \mathcal{C}\bigl(-,\,R(-)\bigr) \,.

The striking analogy (in fact a kind of categorification) of this defining relation to the older notion of adjoint linear operators between Hermitian vector spaces/Hilbert spaces β„‹ i\mathscr{H}_i with inner products βŸ¨βˆ’,βˆ’βŸ© i\langle -,-\rangle_i

L=R †:β„‹ 2⇆ℋ 1means⟨L(βˆ’),βˆ’βŸ© 2=βŸ¨βˆ’,R(βˆ’)⟩ 1 L = R^\dagger \;\colon\; \mathscr{H}_2 \leftrightarrows \mathscr{H}_1 \;\;\;\;\;\;\;\;\;\;\; \text{means} \;\;\;\;\;\;\;\;\;\;\; \bigl\langle L(-),\, - \bigr\rangle_2 \;=\; \bigl\langle -,\, R(-) \bigr\rangle_1

is what gives the notion of adjoint functors its name (cf. further discussion at adjoint operator – history).

β€œthe universality of the concept of adjointness, which was first isolated and named in the conceptual sphere of category theory” [Lawvere (1969)]

β€œThe multiple examples, here and elsewhere, of adjoint functors tend to show that adjoints occur almost everywhere in many branches of Mathematics. It is the thesis of this book that a systematic use of all these adjunctions illuminates and clarifies these subjects.” [MacLane (1971), p. 103]

β€œIn all those areas where category theory is actively used the categorical concept of adjoint functor has come to play a key role.” [first line from An interview with William Lawvere, paraphrasing the first paragraph of Taking categories seriously]

Definition

There are various different but equivalent characterizations of adjoint functors, some of which are discussed below.

In terms of Hom isomorphism

We discuss here the definition of adjointness of functors L⊣RL \dashv R in terms of a natural bijection between hom-sets (Def. below):

{L(c)β†’d}≃{cβ†’R(d)} \{L(c) \to d\} \;\simeq\; \{ c \to R(d) \}

We show that this is equivalent to the abstract definition, in terms of an adjunction in the 2-category Cat, in Prop. below.

\,

Definition

(adjoint functors in terms of natural bijections of hom-sets)

Let π’ž\mathcal{C} and π’Ÿ\mathcal{D} be two categories, and let

π’ŸβŸΆR⟡Lπ’ž \mathcal{D} \underoverset {\underset{R}{\longrightarrow}}{\overset{L}{\longleftarrow}}{} \mathcal{C}

be a pair of functors between them, as shown. Then this is called a pair of adjoint functors (or an adjoint pair of functors) with LL left adjoint and RR right adjoint, denoted

if there exists a natural isomorphism between the hom-functors of the following form:

(1)Hom π’Ÿ(L(βˆ’),βˆ’)≃Hom π’ž(βˆ’,R(βˆ’)). Hom_{\mathcal{D}}\big(L(-),\,-\big) \;\simeq\; Hom_{\mathcal{C}}\big(-,\,R(-)\big) \,.

This means that for all objects cβˆˆπ’žc \in \mathcal{C} and dβˆˆπ’Ÿd \in \mathcal{D} there is a bijection of hom-sets

Hom π’Ÿ(L(c),d) βŸΆβ‰ƒ Hom π’ž(c,R(d)) (L(c)β†’fd) ↦ (cβ†’f˜R(d)) \array{ Hom_{\mathcal{D}}(L(c),d) &\overset{\simeq}{\longrightarrow}& Hom_{\mathcal{C}}(c,R(d)) \\ ( L(c) \overset{f}{\to} d ) &\mapsto& (c \overset{\widetilde f}{\to} R(d)) }

which is natural in cc and dd. This isomorphism is the adjunction isomorphism and the image f˜\widetilde f of a morphism ff under this bijections is called the adjunct of ff. Conversely, ff is called the adjunct of f˜\widetilde f.

Naturality here means that for every morphism g:c 2β†’c 1g \colon c_2 \to c_1 in π’ž\mathcal{C} and for every morphism h:d 1β†’d 2h\colon d_1\to d_2 in π’Ÿ\mathcal{D}, the resulting square

(2)Hom π’Ÿ(L(c 1),d 1) βŸΆβ‰ƒ(βˆ’)˜ Hom π’ž(c 1,R(d 1)) Hom π’Ÿ(L(g),h)↓ ↓ Hom π’ž(g,R(h)) Hom π’Ÿ(L(c 2),d 2) βŸΆβ‰ƒ(βˆ’)˜ Hom π’ž(c 2,R(d 2)) \array{ Hom_{\mathcal{D}}(L(c_1), d_1) &\underoverset{\simeq}{\widetilde{(-)}}{\longrightarrow}& Hom_{\mathcal{C}}(c_1, R(d_1)) \\ {}^{\mathllap{Hom_{\mathcal{D}}(L(g), h)}}\big\downarrow && \big\downarrow^{\mathrlap{Hom_{\mathcal{C}}(g, R(h))}} \\ Hom_{\mathcal{D}}(L(c_2),d_2) &\underoverset{\simeq}{\widetilde{(-)}}{\longrightarrow}& Hom_{\mathcal{C}}(c_2,R(d_2)) }

commutes (see also at hom-functor for the definition of the vertical maps here).

Explicitly, this commutativity, in turn, means that for every morphism f:L(c 1)β†’d 1f \;\colon\; L(c_1) \to d_1 with adjunct f˜:c 1β†’R(d 1)\widetilde f \;\colon\; c_1 \to R(d_1), the adjunct of the composition is

L(c 1) ⟢f d 1 L(g)↑ ↓ h L(c 2) d 2=c 1 ⟢f˜ R(d 1) g↑ ↓ R(h) c 2 R(d 2) \array{ L(c_1) & \overset{f}{\longrightarrow} & d_1 \\ {}^{\mathllap{L(g)}}\big\uparrow && \big\downarrow^{\mathrlap{h}} \\ L(c_2) && d_2 } \;\;\;=\;\;\; \array{ c_1 &\overset{\widetilde f}{\longrightarrow}& R(d_1) \\ {}^{\mathllap{g}}\big\uparrow && \big\downarrow^{\mathrlap{R(h)}} \\ c_2 && R(d_2) }
Definition

(adjunction unit and counit in terms of hom-isomorphism)

Given a pair of adjoint functors

π’ŸβŠ₯⟢R⟡Lπ’ž \mathcal{D} \underoverset {\underset{R}{\longrightarrow}}{\overset{L}{\longleftarrow}}{\bot} \mathcal{C}

according to Def. one says that

  1. for any cβˆˆπ’žc \in \mathcal{C} the adjunct of the identity morphism on L(c)L(c) is the unit morphism of the adjunction at that object, denoted

    Ξ· c≔id L(c)˜:c⟢R(L(c)) \eta_c \coloneqq \widetilde{id_{L(c)}} \;\colon\; c \longrightarrow R(L(c))
  2. for any dβˆˆπ’Ÿd \in \mathcal{D} the adjunct of the identity morphism on R(d)R(d) is the counit morphism of the adjunction at that object, denoted

    ϡ d:L(R(d))⟢d \epsilon_d \;\colon\; L(R(d)) \longrightarrow d
Proposition

(general adjuncts in terms of unit/counit)

Consider a pair of adjoint functors

π’ŸβŠ₯⟢R⟡Lπ’ž \mathcal{D} \underoverset {\underset{R}{\longrightarrow}}{\overset{L}{\longleftarrow}}{\bot} \mathcal{C}

according to Def. , with adjunction units Ξ· c\eta_c and adjunction counits Ο΅ d\epsilon_d according to Def. .

Then

  1. The adjunct f˜\widetilde f of any morphism L(c)β†’fdL(c) \overset{f}{\to} d is obtained from RR and Ξ· c\eta_c as the composite

    (3)f˜:c⟢η cR(L(c))⟢R(f)R(d) \widetilde f \;\colon\; c \overset{\eta_c}{\longrightarrow} R(L(c)) \overset{R(f)}{\longrightarrow} R(d)

    Conversely, the adjunct ff of any morphism c⟢f˜R(d)c \overset{\widetilde f}{\longrightarrow} R(d) is obtained from LL and ϡ d\epsilon_d as

    (4)f:L(c)⟢L(f˜)L(R(d))⟢ϡ dd f \;\colon\; L(c) \overset{L(\widetilde f)}{\longrightarrow} L(R(d)) \overset{\epsilon_d}{\longrightarrow} d
  2. The adjunction units Ξ· c\eta_c and adjunction counits Ο΅ d\epsilon_d are components of natural transformations of the form

    Ξ·:Id π’žβ‡’R∘L \eta \;\colon\; Id_{\mathcal{C}} \Rightarrow R \circ L

    and

    Ο΅:L∘Rβ‡’Id π’Ÿ \epsilon \;\colon\; L \circ R \Rightarrow Id_{\mathcal{D}}
  3. The adjunction unit and adjunction counit satisfy the triangle identities, saying that

    id L(c):L(c)⟢L(η c)L(R(L(c)))⟢ϡ L(c)L(c) id_{L(c)} \;\colon\; L(c) \overset{L(\eta_c)}{\longrightarrow} L(R(L(c))) \overset{\epsilon_{L(c)}}{\longrightarrow} L(c)

    and

    id R(d):R(d)⟢η R(d)R(L(R(d)))⟢R(ϡ d)R(d) id_{R(d)} \;\colon\; R(d) \overset{\eta_{R(d)}}{\longrightarrow} R(L(R(d))) \overset{R(\epsilon_d)}{\longrightarrow} R(d)
Proof

For the first statement, consider the naturality square (2) in the form

id L(c)∈ Hom π’Ÿ(L(c),L(c)) βŸΆβ‰ƒ(βˆ’)˜ Hom π’ž(c,R(L(c))) Hom π’Ÿ(L(id),f)↓ ↓ Hom π’ž(id,R(f)) Hom π’Ÿ(L(c),d) βŸΆβ‰ƒ(βˆ’)˜ Hom π’ž(c,R(d)) \array{ id_{L(c)} \in & Hom_{\mathcal{D}}(L(c), L(c)) &\underoverset{\simeq}{\widetilde{(-)}}{\longrightarrow}& Hom_{\mathcal{C}}(c, R(L(c))) \\ & {}^{\mathllap{Hom_{\mathcal{D}}(L(id), f)}}\big\downarrow && \big\downarrow^{\mathrlap{Hom_{\mathcal{C}}(id, R(f))}} \\ & Hom_{\mathcal{D}}(L(c), d) &\underoverset{\simeq}{\widetilde{(-)}}{\longrightarrow}& Hom_{\mathcal{C}}( c, R(d) ) }

and consider the element id L(c 1)id_{L(c_1)} in the top left entry. Its image under going down and then right in the diagram is f˜\widetilde f, by Def. . On the other hand, its image under going right and then down is R(f)∘η c R(f)\circ \eta_{c}, by Def. . Commutativity of the diagram means that these two morphisms agree, which is the statement to be shown, for the adjunct of ff.

The converse formula follows analogously.

The third statement follows directly from this by applying these formulas for the adjuncts twice and using that the result must be the original morphism:

id L(c) =id L(c)˜˜ =cβ†’Ξ· cR(L(c))˜ =L(c)⟢L(Ξ· c)L(R(L(c)))⟢ϡ L(c)L(c) \begin{aligned} id_{L(c)} & = \widetilde \widetilde { id_{L(c)} } \\ & = \widetilde{ c \overset{\eta_c}{\to} R(L(c)) } \\ & = L(c) \overset{L(\eta_c)}{\longrightarrow} L(R(L(c))) \overset{\epsilon_{L(c)}}{\longrightarrow} L(c) \end{aligned}

For the second statement, we have to show that for every morphism f:c 1β†’c 2f \colon c_1 \to c_2 the following square commutes:

c 1 ⟢f c 2 Ξ· c 1↓ ↓ Ξ· c 2 R(L(c 1)) ⟢R(L(f)) R(L(c 2)) \array{ c_1 &\overset{f}{\longrightarrow}& c_2 \\ {}^{\mathllap{\eta_{c_1}}}\big\downarrow && \big\downarrow^{\mathrlap{\eta_{c_2}}} \\ R(L(c_1)) &\underset{ R(L(f)) }{\longrightarrow}& R(L(c_2)) }

To see this, consider the naturality square (2) in the form

id L(c 2)∈ Hom π’Ÿ(L(c 2),L(c 2)) βŸΆβ‰ƒ(βˆ’)˜ Hom π’ž(c 2,R(L(c 2))) Hom π’Ÿ(L(f),id L(c 2))↓ ↓ Hom π’ž(f,R(id L(c 2))) Hom π’Ÿ(L(c 1),L(c 2)) βŸΆβ‰ƒ(βˆ’)˜ Hom π’ž(c 1,R(L(c 2))) \array{ id_{L(c_2)} \in & Hom_{\mathcal{D}}(L(c_2), L(c_2)) &\underoverset{\simeq}{\widetilde{(-)}}{\longrightarrow}& Hom_{\mathcal{C}}(c_2, R(L(c_2))) \\ & {}^{\mathllap{Hom_{\mathcal{D}}(L(f),id_{L(c_2)})}}\big\downarrow && \big\downarrow^{\mathrlap{Hom_{\mathcal{C}}(f, R(id_{L(c_2)}))}} \\ & Hom_{\mathcal{D}}(L(c_1),L(c_2)) &\underoverset{\simeq}{\widetilde{(-)}}{\longrightarrow}& Hom_{\mathcal{C}}(c_1,R(L(c_2))) }

The image of the element id L(c 2)id_{L(c_2)} in the top left along the right and down is η c 2∘f \eta_{c_2} \circ f, by Def. , while its image down and then to the right is L(f)˜=R(L(f))∘η c 1\widetilde{L(f)} = R(L(f)) \circ \eta_{c_1}, by the previous statement. Commutativity of the diagram means that these two morphisms agree, which is the statement to be shown.

The argument for the naturality of Ο΅\epsilon is directly analogous.

Proposition

(adjointness in terms of hom-isomorphism equivalent to adjunction in CatCat)

Two functors

π’ŸβŸΆR⟡Lπ’ž \mathcal{D} \underoverset {\underset{R}{\longrightarrow}}{\overset{L}{\longleftarrow}}{} \mathcal{C}

are an adjoint pair in the sense that there is a natural isomorphism (1) according to Def. , precisely if they participate in an adjunction in the 2-category Cat, meaning that

  1. there exist natural transformations

    Ξ·:Id π’žβ‡’R∘L \eta \;\colon\; Id_{\mathcal{C}} \Rightarrow R \circ L

    and

    Ο΅:L∘Rβ‡’Id π’Ÿ \epsilon \;\colon\; L \circ R \Rightarrow Id_{\mathcal{D}}
  2. which satisfy the triangle identities

    id L(c):L(c)⟢L(η c)L(R(L(c)))⟢ϡ L(c)L(c) id_{L(c)} \;\colon\; L(c) \overset{L(\eta_c)}{\longrightarrow} L(R(L(c))) \overset{\epsilon_{L(c)}}{\longrightarrow} L(c)

    and

    id R(d):R(d)⟢η R(d)R(L(R(d)))⟢R(ϡ d)R(d) id_{R(d)} \;\colon\; R(d) \overset{\eta_{R(d)}}{\longrightarrow} R(L(R(d))) \overset{R(\epsilon_d)}{\longrightarrow} R(d)
Proof

That a hom-isomorphism (1) implies units/counits satisfying the triangle identities is the statement of the second two items of Prop. .

Hence it remains to show the converse. But the argument is along the same lines as the proof of Prop. : We now define forming of adjuncts by the formula (3). That the resulting assignment f↦f˜f \mapsto \widetilde f is an isomorphism follows from the computation

f˜˜ =cβ†’Ξ· cR(L(c))β†’R(f)R(d)˜ =L(c)β†’L(Ξ· c)L(R(L(c)))β†’L(R(f))L(R(d))β†’Ο΅ dd =L(c)β†’L(Ξ· c)L(R(L(c)))β†’Ο΅ L(c)L(c)⟢fd =L(c)⟢fd \begin{aligned} \widetilde {\widetilde f} & = \widetilde{ c \overset{\eta_c}{\to} R(L(c)) \overset{R(f)}{\to} R(d) } \\ & = L(c) \overset{L(\eta_c)}{\to} L(R(L(c))) \overset{L(R(f))}{\to} L(R(d)) \overset{\epsilon_d}{\to} d \\ & = L(c) \overset{L(\eta_c)}{\to} L(R(L(c))) \overset{ \epsilon_{L(c)} }{\to} L(c) \overset{f}{\longrightarrow} d \\ & = L(c) \overset{f}{\longrightarrow} d \end{aligned}

where, after expanding out the definition, we used naturality of Ο΅\epsilon and then the triangle identity.

Finally, that this construction satisfies the naturality condition (2) follows from the functoriality of the functors involved, and the naturality of the unit/counit:

c 2 ⟢η c 2 R(L(c 2)) g↓ ↓ R(L(g)) β†˜ R(L(g)∘f) c 1 ⟢η c 1 R(L(c 1)) ⟢R(f) R(d 1) R(h∘f)β†˜ ↓ R(h) R(d 2) \array{ c_2 &\overset{ \eta_{c_2} }{\longrightarrow}& R(L(c_2)) \\ {}^{\mathllap{g}}\downarrow && \downarrow^{\mathrlap{R(L(g))}} & \searrow^{\mathrlap{ R( L(g) \circ f ) }} \\ c_1 &\overset{\eta_{c_1}}{\longrightarrow}& R(L(c_1)) &\overset{R(f)}{\longrightarrow}& R(d_1) \\ && & {}_{R( h\circ f)}\searrow & \downarrow^{\mathrlap{ R(h) }} \\ && && R(d_2) }

In terms of representable functors

The condition (1) on adjoint functors L⊣RL \dashv R in Def. implies in particular that for every object dβˆˆπ’Ÿd \in \mathcal{D} the functor Hom π’Ÿ(L(βˆ’),d)Hom_{\mathcal{D}}(L(-),d) is a representable functor with representing object R(d)R(d). The following Prop. observes that the existence of such representing objects for all dd is, in fact, already sufficient to imply that there is a right adjoint functor.

This equivalent perspective on adjoint functors makes manifest that:

  1. adjoint functors are, if they exist, unique up to natural isomorphism, this is Prop. below;

  2. the concept of adjoint functors makes sense also relative to a full subcategory on which representing objects exists, this is the content of Remark below.

Global definition

Proposition

(adjoint functor from objectwise representing object)

A functor L:π’žβŸΆπ’ŸL \;\colon\; \mathcal{C} \longrightarrow \mathcal{D} has a right adjoint R:π’Ÿβ†’π’žR \;\colon\; \mathcal{D} \to \mathcal{C}, according to Def. , already if for all objects dβˆˆπ’Ÿd \in \mathcal{D} there is an object R(d)βˆˆπ’žR(d) \in \mathcal{C} such that there is a natural isomorphism

Hom π’Ÿ(L(βˆ’),d)βŸΆβ‰ƒ(βˆ’)˜Hom π’ž(βˆ’,R(d)), Hom_{\mathcal{D}}(L(-),d) \underoverset{\simeq}{\widetilde{(-)}}{\longrightarrow} Hom_{\mathcal{C}}(-,R(d)) \,,

hence for each object cβˆˆπ’žc \in \mathcal{C} a bijection

Hom π’Ÿ(L(c),d)βŸΆβ‰ƒ(βˆ’)˜Hom π’ž(c,R(d)) Hom_{\mathcal{D}}(L(c),d) \underoverset{\simeq}{\widetilde{(-)}}{\longrightarrow} Hom_{\mathcal{C}}(c,R(d))

such that for each morphism g:c 2β†’c 1g \;\colon\; c_2 \to c_1, the following diagram commutes

(5)Hom π’Ÿ(L(c 1),d) βŸΆβ‰ƒ(βˆ’)˜ Hom π’ž(c 1,R(d)) Hom π’ž(L(g),id d)↓ ↓ Hom π’ž(f,id R(d)) Hom π’Ÿ(L(c 2),d) βŸΆβ‰ƒ(βˆ’)˜ Hom π’ž(c 2,R(d)) \array{ Hom_{\mathcal{D}}(L(c_1),d) &\underoverset{\simeq}{\widetilde{(-)}}{\longrightarrow}& Hom_{\mathcal{C}}(c_1,R(d)) \\ {}^{\mathllap{ Hom_{\mathcal{C}}(L(g),id_d) }} \big\downarrow && \big\downarrow^{\mathrlap{ Hom_{\mathcal{C}}( f, id_{R(d)} ) }} \\ Hom_{\mathcal{D}}(L(c_2),d) &\underoverset{\simeq}{\widetilde{(-)}}{\longrightarrow}& Hom_{\mathcal{C}}(c_2,R(d)) }

(This is as in (2), except that only naturality in the first variable is required.)

In this case there is a unique way to extend RR from a function on objects to a function on morphisms such as to make it a functor R:π’Ÿβ†’π’žR \colon \mathcal{D} \to \mathcal{C} which is right adjoint to LL. , and hence the statement is that with this, naturality in the second variable is already implied.

Proof

Notice that

  1. in the language of presheaves the assumption is that for each dβˆˆπ’Ÿd \in \mathcal{D} the presheaf

    Hom π’Ÿ(L(βˆ’),d)∈[π’ž op,Set] Hom_{\mathcal{D}}(L(-),d) \;\in\; [\mathcal{C}^{op}, Set]

    is represented by the object R(d)R(d), and naturally so.

  2. In terms of the Yoneda embedding

    y:π’žβ†ͺ[π’ž op,Set] y \;\colon\; \mathcal{C} \hookrightarrow [\mathcal{C}^{op}, Set]

    we have

    (6)Hom π’ž(βˆ’,R(d))=y(R(d)) Hom_{\mathcal{C}}(-,R(d)) = y(R(d))

The condition (2) says equivalently that RR has to be such that for all morphisms h:d 1β†’d 2h \;\colon\; d_1 \to d_2 the following diagram in the category of presheaves [π’ž op,Set][\mathcal{C}^{op}, Set] commutes

Hom π’Ÿ(L(βˆ’),d 1) βŸΆβ‰ƒ(βˆ’)˜ Hom π’ž(βˆ’,R(d 1)) Hom π’ž(L(βˆ’),h)↓ ↓ Hom π’ž(βˆ’,R(h)) Hom π’Ÿ(L(βˆ’),d 2) βŸΆβ‰ƒ(βˆ’)˜ Hom π’ž(βˆ’,R(d 2)) \array{ Hom_{\mathcal{D}}(L(-),d_1) &\underoverset{\simeq}{\widetilde{(-)}}{\longrightarrow}& Hom_{\mathcal{C}}(-,R(d_1)) \\ {}^{\mathllap{ Hom_{\mathcal{C}}( L(-) , h ) }} \big\downarrow && \big\downarrow^{\mathrlap{ Hom_{\mathcal{C}}( -, R(h) ) }} \\ Hom_{\mathcal{D}}(L(-),d_2) &\underoverset{\simeq}{\widetilde{(-)}}{\longrightarrow}& Hom_{\mathcal{C}}(-, R(d_2)) }

This manifestly has a unique solution

y(R(h))=Hom π’ž(βˆ’,R(h)) y(R(h)) \;=\; Hom_{\mathcal{C}}(-,R(h))

for every morphism h:d 1β†’d 2h \colon d_1 \to d_2 under y(R(βˆ’))y(R(-)) (6). But the Yoneda embedding yy is a fully faithful functor (this prop.), which means that thereby also R(h)R(h) is uniquely fixed.

Remark

In more fancy language, the statement of Prop. is the following:

By precomposition LL defines a functor of presheaf categories

L *:[π’Ÿ op,Set]β†’[π’ž op,Set]. L^* \;\colon\; [\mathcal{D}^{op}, Set] \to [\mathcal{C}^{op}, Set] \,.

By restriction along the Yoneda embedding y:π’Ÿβ†’[π’Ÿ op,Set]y \;\colon\; \mathcal{D} \to [\mathcal{D}^{op}, Set] this yields the functor

LΒ―:π’Ÿ ⟢y [π’Ÿ op,Set] ⟢L * [π’ž op,Set] d ↦ Hom π’Ÿ(βˆ’,d) ↦ Hom π’Ÿ(L(βˆ’),d). \bar L \;\colon\; \array{ \mathcal{D} &\overset{y}{\longrightarrow}& [\mathcal{D}^{op}, Set] &\overset{L^*}{\longrightarrow}& [\mathcal{C}^{op}, Set] \\ d &\mapsto& Hom_{\mathcal{D}}(-,d) &\mapsto& Hom_{\mathcal{D}}(L(-),d) } \,.

The statement is that for all d∈Dd \in D this presheaf LΒ―(d)\bar L(d) is representable, then it is functorially so in that there exists a functor R:π’Ÿβ†’π’žR \colon \mathcal{D} \to \mathcal{C} such that

L¯≃y∘R. \bar L \;\simeq\; y \circ R \,.

Local definition

Remark

(relative adjoint functors)

The perspective of Prop. has the advantage that it yields useful information even if the adjoint functor RR does not exist globally, i.e. as a functor on all of π’Ÿ\mathcal{D}:

It may happen that

LΒ―(d)≔Hom D(L(βˆ’),d)∈[C op,Set] \bar L(d) \coloneqq Hom_D(L(-),d) \in [C^{op}, Set]

is representable for some object dβˆˆπ’Ÿd \in \mathcal{D} but not for all dd. The representing object may still usefully be thought of as R(d)R(d), and in fact it may be viewed as a right adjoint to LL relative to the inclusion of the full subcategory determined by those dds for which LΒ―(d)\bar L(d) is representable; see reflection along a functor and relative adjoint functor for more.

This global versus local evaluation of adjoint functors induces the global/local pictures of the definitions

as discussed there.

In terms of universal factorization through a (co)unit

We have seen in Prop. that the unit of an adjunction and counit of an adjunction plays a special role. One may amplify this by characterizing these morphisms as universal arrows in the sense of the following Def. . In fact the existence of these is already equivalent to the existence of an adjoint functor, this is the statement of Prop. below.

\,

Definition

(universal arrow)

Given a functor R:π’Ÿβ†’π’žR \;\colon\; \mathcal{D} \to \mathcal{C}, and an object cβˆˆπ’žc\in \mathcal{C}, a universal arrow from cc to RR (or a reflection of cc along RR) is an initial object of the comma category (c/R)(c/R). This means that it consists of

  1. an object L(c)βˆˆπ’ŸL(c)\in \mathcal{D}

  2. a morphism η c:c→R(L(c))\eta_c \;\colon\; c \to R\big(L(c)\big), to be called the unit,

such that for any dβˆˆπ’Ÿd\in \mathcal{D}, any morphism f:cβ†’R(d)f \colon c\to R(d) factors through this unit Ξ· c\eta_c as

(7) c Ξ· c↙ β†˜ f R(L(c)) ⟢R(f˜) R(d) L(c) ⟢f˜ d \array{ && c \\ & {}^{\mathllap{\eta_c}}\swarrow && \searrow^{\mathrlap{f}} \\ R\big(L(c)\big) && \underset {R (\widetilde f)} {\longrightarrow} && R(d) \\ \\ L(c) && \underset{ \widetilde f}{\longrightarrow} && d }

for a unique f˜:L(c)⟢d\widetilde f \;\colon\; L(c) \longrightarrow d, to be called the adjunct of ff.

(e.g. Borceux, Vol. 1, Definition 3.1.1)

Proposition

(universal morphisms are initial objects in the comma category)

Let R:π’Ÿβ†’π’žR: \mathcal{D} \to \mathcal{C} be a functor and cβˆˆπ’žc \in \mathcal{C} an object. Then the following are equivalent:

  1. c→η cR(L(c))c \overset{\eta_c}{\to} R(L(c)) is a universal morphism into R(L(c))R(L(c)) (Def. );

  2. (c,Ξ· c)(c, \eta_c) is the initial object in the comma category c/Rc/R.

Proposition

(collection of universal arrows equivalent to adjoint functor)

Let R:π’Ÿβ†’π’žR \;\colon\; \mathcal{D} \to \mathcal{C} be a functor. Then the following are equivalent:

  1. RR has a left adjoint functor L:π’žβ†’π’ŸL \colon \mathcal{C} \to \mathcal{D} according to Def. ,

  2. for every object cβˆˆπ’žc \in \mathcal{C} there is a universal arrow c⟢η cR(L(c))c \overset{\eta_c}{\longrightarrow} R(L(c)), according to Def. .

Proof

In one direction, assume a left adjoint LL is given. Define the would-be universal arrow at cβˆˆπ’žc \in \mathcal{C} to be the unit of the adjunction Ξ· c\eta_c via Def. . Then the statement that this really is a universal arrow is implied by Prop. .

In the other direction, assume that universal arrows Ξ· c\eta_c are given. The uniqueness clause in Def. immediately implies bijections

Hom π’Ÿ(L(c),d) βŸΆβ‰ƒ Hom π’ž(c,R(d)) (L(c)β†’f˜d) ↦ (cβ†’Ξ· cR(L(c))β†’R(f˜)R(d)) \array{ Hom_{\mathcal{D}}(L(c),d) &\overset{\simeq}{\longrightarrow}& Hom_{\mathcal{C}}(c,R(d)) \\ \left( L(c) \overset{\widetilde f}{\to} d \right) &\mapsto& \left( c \overset{\eta_c}{\to} R(L(c)) \overset{ R(\widetilde f) }{\to} R(d) \right) }

Hence to satisfy (1) it remains to show that these are natural in both variables. In fact, by Prop. it is sufficient to show naturality in the variable dd. But this is immediate from the functoriality of RR applied in (7): For h:d 1β†’d 2h \colon d_1 \to d_2 any morphism, we have

c Ξ· c↙ β†˜ f R(L(c)) ⟢R(f˜) R(d 1) R(h∘f˜)β†˜ ↓ R(h) R(d 2) \array{ && c \\ & {}^{\mathllap{\eta_c}}\swarrow && \searrow^{\mathrlap{f}} \\ R (L(c)) &&\underset{R (\widetilde f)}{\longrightarrow}&& R(d_1) \\ && {}_{\mathllap{ R( h\circ \widetilde f ) }}\searrow && \downarrow^{\mathrlap{R(h)}} \\ && && R(d_2) }
Example

(localization via universal arrows)

The characterization of adjoint functors in terms of universal factorizations through the unit and counit (Prop. ) is of particular interest in the case that RR is a full and faithful functor

R:π’Ÿβ†ͺπ’ž R \;\colon\; \mathcal{D} \hookrightarrow \mathcal{C}

exhibiting π’Ÿ\mathcal{D} as a reflective subcategory of π’ž\mathcal{C}. In this case we may think of LL as a localization and of objects in the essential image of LL as local objects. Then the above says that:

  • every morphism cβ†’Rdc \to R d from cc into a local object factors throught the localization of cc.

In terms of comma categories

A functor L:Cβ†’DL \colon C \to D is left adjoint to a functor R:Dβ†’CR \colon D \to C if and only if there is an isomorphism (not equivalence) of comma categories L↓Dβ‰…C↓RL \downarrow D \cong C \downarrow R and this isomorphism commutes with the forgetful functors to the product category CΓ—DC \times D. See Β§B.I.2 of Functorial Semantics of Algebraic Theories.

This characterisation generalises (in the unenriched setting) to relative adjunctions by replacing C↓RC \downarrow R by J↓RJ \downarrow R.

In terms of cographs/correspondences/heteromorphisms

Every profunctor

k:C op×D→S k : C^{op} \times D \to S

defines a category C* kDC *^k D with Obj(C* kD)=Obj(C)βŠ”Obj(D)Obj(C *^k D) = Obj(C) \sqcup Obj(D) and with hom set given by

Hom C* kD(X,Y)={Hom C(X,Y) ifX,Y∈C Hom D(X,Y) ifX,Y∈D {k(X,Y)} ifX∈CandY∈D βˆ… otherwise Hom_{C *^k D}(X,Y) = \left\{ \array{ Hom_C(X,Y) & if X, Y \in C \\ Hom_D(X,Y) & if X,Y \in D \\ \left\{k(X,Y)\right\} & if X \in C and Y \in D \\ \emptyset & otherwise } \right.

(k(X,Y)k(X,Y) is also called a heteromorphism).

This category naturally comes with a functor to the interval category

C* kD→Δ 1. C *^k D \to \Delta^1 \,.

Now, every functor L:C→DL : C \to D induces a profunctor

k L(X,Y)=Hom D(L(X),Y) k^L(X,Y) = Hom_D(L(X), Y)

and every functor R:D→CR : D \to C induces a profunctor

k R(X,Y)=Hom C(X,R(Y)). k_R(X,Y) = Hom_C(X, R(Y)) \,.

The functors LL and RR are adjoint precisely if the profunctors that they define in the above way are equivalent. This in turn is the case if C⋆ LD≃(D op⋆ R opC op) opC \star^L D \simeq (D^{op} \star^{R^{op}} C^{op})^{op}.

We say that C⋆ kDC \star^k D is the cograph of the functor kk. See there for more on this.

In terms of graphs/2-sided discrete fibrations

Functors L:C→DL \colon C \to D and R:D→CR \colon D \to C are adjoint precisely if we have a commutative diagram of the form

(L↓Id D) β†’β‰… (Id C↓R) β†˜ ↙ CΓ—D, \array{ (L \downarrow Id_D) &&\stackrel{\cong}{\to}&& (Id_C \downarrow R) \\ & \searrow && \swarrow \\ && C \times D \mathrlap{\,,} }

where the downwards arrows denote the maps induced by the canonical projections out of the comma categories. This definition of adjoint functors was introduced by Lawvere in Functorial Semantics of Algebraic Theories, and was the original motivation for comma categories.

The above diagram may be recovered directly from the image under the equivalence [C opΓ—D,Set]→≃DFib(D,C)[C^{op} \times D, Set] \stackrel{\simeq}{\to} DFib(D,C) described at 2-sided fibration of the isomorphism of induced profunctors C opΓ—Dβ†’SetC^{op} \times D \to Set (see above at β€œIn terms of Hom isomorphism”). Its relation to the hom-set definition of adjoint functors can thus be understood within the general paradigm of Grothendieck construction-like correspondences. Consequently, this description is not viable for enriched adjunctions.

This description generalises to relative adjunctions by replacing Id CId_C with JJ.

In terms of Kan extensions/liftings

Given L:C→DL \colon C \to D, we have that it has a right adjoint R:D→CR\colon D \to C precisely if the left Kan extension Lan L1 CLan_L 1_C of the identity along LL exists and is absolute, in which case

R≃Lan L1 C. R \simeq \mathop{Lan}_L 1_C \,.

In this case, the universal 2-cell 1 C→RL1_C \to R L corresponds to the unit of the adjunction; the counit and the verification of the triangular identities can all be obtained through properties of Kan extensions and absoluteness.

It is also possible to express this in terms of Kan liftings: LL has a right adjoint RR if and only if:

In this case, we get the counit as given by the universal cell LR→1 DL R \to 1_D, while the rest of the data and properties can be derived from it through the absolute Kan lifting assumption.

Dually, we have that for R:D→CR\colon D \to C, it has a left adjoint L:C→DL \colon C \to D precisely if

  • L≃Ran R1 DL \simeq \mathop{Ran}_R 1_D, and this Kan extension is absolute

or, in terms of left Kan liftings:

  • L≃Lift R1 CL \simeq \mathop{Lift}_R 1_C, and this Kan lifting is absolute

This follows from the fact that the adjunction L⊣RL \dashv R induces adjunctions βˆ’βˆ˜RβŠ£βˆ’βˆ˜L- \circ R \dashv - \circ L and Lβˆ˜βˆ’βŠ£Rβˆ˜βˆ’L \circ - \dashv R \circ -.

The formulations in terms of liftings generalize to (unenriched) relative adjoints by allowing an arbitrary functor JJ in place of the identity; see there for more.

Transformation of adjoints

There are several layers of generality at which one may consider a notion of homomorphism between adjoint functors.

Here is a basic but important notion:

Definition

(conjugate transformations of adjoints)
Given a pair of pairs of adjoint functors between the same categories

π’žβŠ₯⟡R 1⟢L 1π’Ÿ π’žβŠ₯⟡R 2⟢L 2π’Ÿ \array{ \mathcal{C} \underoverset {\underset{R_1}{\longleftarrow}} {\overset{L_1}{\longrightarrow}} {\;\;\; \bot \;\;\;} \mathcal{D} \\ \mathcal{C} \underoverset {\underset{R_2}{\longleftarrow}} {\overset{L_2}{\longrightarrow}} {\;\;\; \bot \;\;\;} \mathcal{D} }

then a pair of natural transformations between the adjoints of the same chirality, of this form

Ξ»:L 1β†’L 2ρ:R 2β†’R 1 \lambda \,\colon\, L_1 \to L_2 \;\;\;\;\;\; \rho \,\colon\, R_2 \to R_1

is called conjugate for [MacLane (1971), Β§IV.7 (5)] or a pseudo-transformation of [Harpaz & Prasma (2015), Sec. 2.2] the given adjunctions if they make the following diagram of natural transformations between hom-sets commute:

π’ž(L 2(βˆ’),(βˆ’)) ⟢∼ π’Ÿ((βˆ’),R 2(βˆ’)) π’ž(Ξ» (βˆ’),βˆ’)↓ ↓ π’ž(βˆ’,ρ (βˆ’)) π’ž(L 1(βˆ’),(βˆ’)) ⟢∼ π’Ÿ((βˆ’),R 1(βˆ’)), \array{ \mathcal{C}\big( L_2(-) ,\, (-) \big) &\overset{\sim}{\longrightarrow}& \mathcal{D}\big( (-) ,\, R_2(-) \big) \\ \mathllap{{}^{ \mathcal{C}\big(\lambda_{(-)},\,-\big) }} \Big\downarrow && \Big\downarrow \mathrlap{{}^{ \mathcal{C}\big(-,\,\rho_{(-)}\big) }} \\ \mathcal{C}\big( L_1(-) ,\, (-) \big) &\overset{\sim}{\longrightarrow}& \mathcal{D}\big( (-) ,\, R_1(-) \big) \mathrlap{\,,} }

where the horizontal maps are the given hom-isomorphisms (1).

This condition is compatible with horizontal and vertical composition of natural transformations as 2-morphisms in Cat and hence yields:

Definition

The (very large) wide and locally full sub-2-category Cat Adj Cat_{Adj} of Cat

(8)Cat Adj⟢Cat Cat_{Adj} \longrightarrow Cat

whose

Proposition

Under the Grothendieck construction, the Grothendieck fibrations which arise from pseudofunctors β„¬βŸΆCat\mathcal{B} \longrightarrow Cat that factor through Cat AdjCat_{Adj} (8) are equivalently the bifibrations.

This may be category theory folklore; a proof has been spelled out in Harpaz & Prasma (2015), Prop. 2.2.1.

Properties

Basic properties

Proposition

(adjoint functors are unique up to natural isomorphism)

The left adjoint or right adjoint to a functor (Def. ), if it exists, is unique up to natural isomorphism.

Proof

Suppose the functor L:π’Ÿβ†’π’žL \colon \mathcal{D} \to \mathcal{C} is given, and we are asking for uniqueness of its right adjoint, if it exists. The other case is directly analogous.

Suppose that R 1,R 2:π’žβ†’π’ŸR_1, R_2 \;\colon\; \mathcal{C} \to \mathcal{D} are two functors which are right adjoint to LL. Then for each dβˆˆπ’Ÿd \in \mathcal{D} the corresponding two hom-isomorphisms (1) combine to say that there is a natural isomorphism

Ξ¦ d:Hom π’ž(βˆ’,R 1(d))≃Hom π’ž(βˆ’,R 2(d)) \Phi_d \;\colon\; Hom_{\mathcal{C}}(-,R_1(d)) \;\simeq\; Hom_{\mathcal{C}}(-,R_2(d))

As in the proof of Prop. , the Yoneda lemma implies that

Ξ¦ d=y(Ο• d) \Phi_d \;=\; y( \phi_d )

for some isomorphism

Ο• d:R 1(d)→≃R 2(d). \phi_d \;\colon\; R_1(d) \overset{\simeq}{\to} R_2(d) \,.

But then the uniqueness statement of Prop. implies that the collection of these isomorphisms for each object constitues a natural isomorphism between the functors.

Proposition

(left adjoints preserve colimits and right adjoints preserve limits)

Let (L⊣R):π’Ÿβ†’π’ž(L \dashv R) \colon \mathcal{D} \to \mathcal{C} be a pair of adjoint functors (Def. ). Then

Proof

Let y:Iβ†’π’Ÿy : I \to \mathcal{D} be a diagram whose limit lim ← iy i\lim_{\leftarrow_i} y_i exists. Then we have a sequence of natural isomorphisms, natural in x∈Cx \in C

Hom π’ž(x,Rlim ← iy i) ≃Hom π’Ÿ(Lx,lim ← iy i) ≃lim ← iHom π’Ÿ(Lx,y i) ≃lim ← iHom π’ž(x,Ry i) ≃Hom π’ž(x,lim ← iRy i), \begin{aligned} Hom_{\mathcal{C}}(x, R {\lim_\leftarrow}_i y_i) & \simeq Hom_{\mathcal{D}}(L x, {\lim_\leftarrow}_i y_i) \\ & \simeq {\lim_\leftarrow}_i Hom_{\mathcal{D}}(L x, y_i) \\ & \simeq {\lim_\leftarrow}_i Hom_{\mathcal{C}}( x, R y_i) \\ & \simeq Hom_{\mathcal{C}}( x, {\lim_\leftarrow}_i R y_i) \,, \end{aligned}

where we used the hom-isomorphism (1) and the fact that any hom-functor preserves limits (see there). Because this is natural in xx the Yoneda lemma implies that we have an isomorphism

Rlim ← iy i≃lim ← iRy i. R {\lim_\leftarrow}_i y_i \simeq {\lim_\leftarrow}_i R y_i \,.

The argument that shows the preservation of colimits by LL is analogous.

Remark

A partial converse to Prop. is provided by the adjoint functor theorem. See also Pointwise Expression below.

Proposition

Let L⊣RL \dashv R be a pair of adjoint functors (Def. ). Then the following holds:

(Further properties are listed in this MathOverflow discussion.)

A\phantom{A}adjunctionA\phantom{A}A\phantom{A}unit is iso:A\phantom{A}
A\phantom{A}coreflectionA\phantom{A}
A\phantom{A}counit is iso:A\phantom{A}A\phantom{A}reflectionA\phantom{A}A\phantom{A}adjoint equivalenceA\phantom{A}
Proof

For the characterization of faithful RR by epi counit components, notice (as discussed at epimorphism ) that LRx→xL R x \to x being an epimorphism is equivalent to the induced function

Hom(x,a)β†’Hom(LRx,a) Hom(x, a) \to Hom(L R x, a)

being an injection for all objects aa. Then use that, by adjointness, we have an isomorphism

Hom(LRx,a)→≃Hom(Rx,Ra) Hom(L R x , a ) \stackrel{\simeq}{\to} Hom(R x, R a)

and that, by the formula for adjuncts and the zig-zag identity, this is such that the composite

R x,a:Hom(x,a)β†’Hom(LRx,a)→≃Hom(Rx,Ra) R_{x,a} : Hom(x,a) \to Hom(L R x, a) \stackrel{\simeq}{\to} Hom(R x, R a)

is the component map of the functor RR (this Prop.):

(xβ†’fa) ↦(LRxβ†’xβ†’fa) ↦(RLRxβ†’Rxβ†’RfRa) ↦(Rxβ†’RLRxβ†’Rxβ†’RfRa) =(Rxβ†’RfRa). \begin{aligned} (x \stackrel{f}{\to} a) & \mapsto (L R x \to x \stackrel{f}{\to} a) \\ & \mapsto (R L R x \to R x \stackrel{R f}{\to} R a) \\ & \mapsto (R x \to R L R x \to R x \stackrel{R f}{\to} R a) \\ & = (R x \stackrel{R f}{\to} R a) \end{aligned} \,.

Therefore R x,aR_{x,a} is injective for all x,ax,a, hence RR is faithful, precisely if LRx→xL R x \to x is an epimorphism for all xx. The characterization of RR full is just the same reasoning applied to the fact that ϡ x:LRx→x\epsilon_x \colon L R x \to x is a split monomorphism iff for all objects aa the induced function

Hom(x,a)β†’Hom(LRx,a) Hom(x, a) \to Hom(L R x, a)

is a surjection.

For the characterization of faithful LL by monic units notice that analogously (as discussed at monomorphism) x→RLxx \to R L x is a monomorphism if for all objects aa the function

Hom(a,x)β†’Hom(a,RLx) Hom(a,x ) \to Hom(a, R L x)

is an injection. Analogously to the previous argument we find that this is equivalent to

L a,x:Hom(a,x)β†’Hom(a,RLx)→≃Hom(La,Lx) L_{a,x} : Hom(a,x ) \to Hom(a, R L x) \stackrel{\simeq}{\to} Hom(L a, L x)

being an injection. So LL is faithful precisely if all xβ†’RLxx \to R L x are monos. For LL full, it’s just the same applied to xβ†’RLxx \to R L x split epimorphism iff the induced function

Hom(a,x)β†’Hom(a,RLx) Hom(a,x ) \to Hom(a, R L x)

is a surjection, for all objects aa.

The proof of the other statements proceeds analogously.

Parts of this statement can be strengthened:

Proposition

Let (L⊣R):Dβ†’C(L \dashv R) : D \to C be a pair of adjoint functors such that there is any natural isomorphism

LR≃Id, L R \simeq Id \,,

then also the counit ϡ:LR→Id\epsilon : L R \to Id is an isomorphism.

This appears as (Johnstone, lemma 1.1.1).

Proof

Using the given isomorphism, we may transfer the comonad structure on LRL R to a comonad structure on Id DId_D. By the Eckmann-Hilton argument the endomorphism monoid of Id DId_D is commutative. Therefore, since the comultiplication of the comonad Id DId_D is a left inverse to the counit (by the co-unitality property applied to this degenerate situation), it is in fact a two-sided inverse and hence the Id DId_D-counit is an isomorphism. Transferring this back one finds that also the counit of the comand LRL R, hence of the adjunction (L⊣R)(L \dashv R) is an isomorphism.

Pointwise expression

Proposition

(pointwise expression of left adjoints in terms of limits over comma categories)

A functor R:π’žβŸΆπ’ŸR \;\colon\; \mathcal{C} \longrightarrow \mathcal{D} has a left adjoint L:π’ŸβŸΆπ’žL \;\colon\; \mathcal{D} \longrightarrow \mathcal{C} precisely if

  1. RR preserves all limits that exist in π’ž\mathcal{C};

  2. for each object dβˆˆπ’Ÿd \in \mathcal{D}, the limit of the canonical functor out of the comma category of RR under dd

    d/RβŸΆπ’ž d/R \longrightarrow \mathcal{C}

    exists.

In this case the value of the left adjoint LL on dd is given by that limit:

(9)L(d)≃lim⟡(c,d ↓ f R(c))∈d/Rc L(d) \;\simeq\; \underset{\underset{ \left( c, \array{ d \\ \downarrow^{\mathrlap{f}} \\ R(c) } \right) \in d/R }{\longleftarrow}}{\lim} c

(e.g. MacLane, chapter X, theorem 2)

Proof

First assume that the left adjoint exist. Then

  1. RR is a right adjoint and hence preserves limits since all right adjoints preserve limits;

  2. by Prop. the adjunction unit provides a universal morphism Ξ· d\eta_d into L(d)L(d), and hence, by Prop. , exhibits (L(d),Ξ· d)(L(d), \eta_d) as the initial object of the comma category d/Rd/R. The limit over any category with an initial object exists, as it is given by that initial object.

Conversely, assume that the two conditions are satisfied and let L(d)L(d) be given by (9). We need to show that this yields a left adjoint.

By the assumption that RR preserves all limits that exist, we have

(10)R(L(d)) =R(lim⟡(c,d ↓ f R(c))∈d/Rc) ≃lim⟡(c,d ↓ f R(c))∈d/RR(c) \array{ R(L(d)) & = R\left( \underset{\underset{ \left( c, \array{ d \\ \downarrow^{\mathrlap{f}} \\ R(c) } \right) \in d/R }{\longleftarrow}}{\lim} c \right) \\ & \simeq \underset{\underset{ \left( c, \array{ d \\ \downarrow^{\mathrlap{f}} \\ R(c) } \right) \in d/R }{\longleftarrow}}{\lim} R(c) }

Since the d→fR(d)d \overset{f}{\to} R(d) constitute a cone over the diagram of the R(d)R(d), there is universal morphism

d⟢AAη dAAR(L(d)). d \overset{\phantom{AA} \eta_d \phantom{AA}}{\longrightarrow} R(L(d)) \,.

By Prop. it is now sufficient to show that Ξ· d\eta_d is a universal morphism into L(d)L(d), hence that for all cβˆˆπ’žc \in \mathcal{C} and d⟢gR(c)d \overset{g}{\longrightarrow} R(c) there is a unique morphism L(d)⟢f˜cL(d) \overset{\widetilde f}{\longrightarrow} c such that

d Ξ· d↙ β†˜ f R(L(d)) ⟢AAR(f˜)AA R(c) L(d) ⟢AAf˜AA c \array{ && d \\ & {}^{\mathllap{ \eta_d }}\swarrow && \searrow^{\mathrlap{f}} \\ R(L(d)) && \underset{\phantom{AA}R(\widetilde f)\phantom{AA}}{\longrightarrow} && R(c) \\ L(d) &&\underset{\phantom{AA}\widetilde f\phantom{AA}}{\longrightarrow}&& c }

By Prop. , this is equivalent to (L(d),Ξ· d)(L(d), \eta_d) being the initial object in the comma category c/Rc/R, which in turn is equivalent to it being the limit of the identity functor on c/Rc/R (this prop.). But this follows directly from the limit formulas (9) and (10).

See at adjoint functor theorem for more.

Relation between adjunctions and monads

There is a close relation between adjunctions (adjoint functors) and monads:

Monad induced by an adjunction

Every adjunction (L⊣R)(L \dashv R) induces a monad R∘LR \circ L and a comonad L∘RL \circ R.

(Huber 1961, Β§4; see eg. MacLane 1971, Β§VI.1 (p 134); Borceux 1994, vol. 2, prop. 4.2.1).

In detail:

Proposition

Let (π’ž,π’Ÿ,F,U,Ξ·,Ο΅)(\mathcal{C},\mathcal{D},F,U,\eta,\epsilon) be a pair of adjoint functors ie F⊣UF \dashv U are adjoint functors where F:π’žβ†’π’ŸF \colon \mathcal{C} \rightarrow \mathcal{D}, U:π’Ÿβ†’π’žU \colon\mathcal{D} \rightarrow \mathcal{C}, Ξ· A:Aβ†’U(F(A))\eta_{A}:A \rightarrow U(F(A)) is the unit and Ο΅ B:F(U(B))β†’B\epsilon_{B} \colon F(U(B)) \rightarrow B is the counit. Then:

  • U∘FU \circ F is a monad on π’ž\mathcal{C}, with unit Ξ·\eta and multiplication U(Ο΅ F(A)):U(F(U(F(A))))β†’U(F(A))U(\epsilon_{F(A)}):U(F(U(F(A)))) \rightarrow U(F(A)).

  • F∘UF \circ U is a comonad on π’Ÿ\mathcal{D}, with counit Ο΅\epsilon and comultiplication F(Ξ· U(B)):F(U(B))β†’F(U(F(U(B))))F(\eta_{U(B)}):F(U(B)) \rightarrow F(U(F(U(B)))).

Proof

We verify that we obtain a monad, the argument for the comonad is formally dual.

(1) We know that this diagram commutes:

By applying UU, we obtain the first part of the unitaly of the monad:

(2) We know that this diagram commutes: By putting B=F(A)B=F(A), we obtain the second part of the unitality of the monad:

(3) The naturality of ϡ\epsilon is written, for every f:B→B′f:B \rightarrow B':

We apply it to f=Ο΅ F(A)f=\epsilon_{F(A)} and it gives:

Finally apply UU to this diagram to obtain:

which is exactly the associativity of the multiplication of the monad.

(4) The naturality of the multiplication U(Ο΅ F(A)):U(F(U(F(A))))β†’U(F(A))U(\epsilon_{F(A)}):U(F(U(F(A)))) \rightarrow U(F(A)) is obtained by two whiskerings of the counit Ο΅ B:U(F(B))β†’B\epsilon_{B}:U(F(B)) \rightarrow B.

Category of adjunction-resolutions of a monad

An adjunction inducing a monad TT (as above) is also called a resolution of TT.

There is in general more than one such resolution, in fact there is a category of adjunctions for a given monad whose morphisms are β€œcomparison functors” (eg. MacLane 1971, Β§VI.3).

In this category:

(e.g. Borceux 1994, vol. 2, prop. 4.2.2)

Semantics-structure adjunction

The above passage from adjunctions to monads and back to their monadic adjunctions constitutes itself an adjunction, sometimes called the semantics-structure adjunction.

Opposite adjoint functors

Given a pair of adjoint functors

π’ŸβŠ₯⟢R⟡Lπ’ž \mathcal{D} \underoverset {\underset{R}{\longrightarrow}} {\overset{L}{\longleftarrow}} {\;\;\;\;\bot\;\;\;\;} \mathcal{C}

there is an induced opposite adjunction of opposite functors between their opposite categories of the form

π’Ÿ opβŠ₯⟡L op⟢R opπ’ž op. \mathcal{D}^{op} \underoverset {\underset{L^{op}}{\longleftarrow}} {\overset{R^{op}}{\longrightarrow}} {\;\;\;\;\bot\;\;\;\;} \mathcal{C}^{op} \,.

Hence where LL was the left adjoint, its opposite becomes the right adjoint, and dually for RR.

This is immediate from the definition of opposite categories and the characterization of adjoint functors via the corresponding hom-isomorphism.

The adjunction unit of the opposite adjunction has as components the components of the original adjunction counit, regarded in the opposite category, and dually:

Ο΅ c R opL op:R op∘L op(c)β†’(Ξ· c RL) opc,AAAAAAΞ· d L opR op:dβ†’(Ο΅ d LR) opL op∘R op(d). \epsilon^{R^{op} L^{op}}_{c} \;\colon\; R^{op}\circ L^{op}(c) \xrightarrow{\;\; \big( \eta^{R L}_c \big) ^{op} \;\;} c \,, {\phantom{AAAAAA}} \eta^{L^{op} R^{op}}_{d} \;\colon\; d \xrightarrow{\;\; \big( \epsilon^{L R}_d \big) ^{op} \;\;} L^{op} \circ R^{op}(d) \,.

Composing adjunctions

Given two pairs of adjoint functors

β„°βŠ₯⟢Rβ€²βŸ΅Lβ€²π’ŸβŠ₯⟢R⟡Lπ’ž \mathcal{E} \underoverset {\underset{R'}{\longrightarrow}} {\overset{L'}{\longleftarrow}} {\;\;\;\;\bot\;\;\;\;} \mathcal{D} \underoverset {\underset{R}{\longrightarrow}} {\overset{L}{\longleftarrow}} {\;\;\;\;\bot\;\;\;\;} \mathcal{C}

there is an induced composite adjunction Lβ€²βˆ˜L⊣R∘Rβ€²L' \circ L \dashv R \circ R' between π’ž\mathcal{C} and β„°\mathcal{E}.

This is immediate from the characterization of adjoint functors in terms of hom-isomorphisms:

Hom β„°(Lβ€²(L(βˆ’)),βˆ’)≃Hom π’Ÿ(L(βˆ’),Rβ€²(βˆ’))≃Hom π’ž(βˆ’,R(Rβ€²(βˆ’))) Hom_{\mathcal{E}}(L'(L(-)),-) \;\simeq\; Hom_{\mathcal{D}}(L(-),R'(-)) \;\simeq\; Hom_{\mathcal{C}}(-,R(R'(-)))

Pre- and postcomposite adjoint functors

Given a pair of adjoint functors

π’ŸβŠ₯⟢R⟡Lπ’ž \mathcal{D} \underoverset {\underset{R}{\longrightarrow}} {\overset{L}{\longleftarrow}} {\;\;\;\;\bot\;\;\;\;} \mathcal{C}

and a category β„°\mathcal{E}, there is an induced adjunction of precomposition functors between the functor categories [π’ž,β„°][\mathcal{C}, \mathcal{E}] and [π’Ÿ,β„°][\mathcal{D}, \mathcal{E}] of the form

[π’Ÿ,β„°]βŠ₯βŸΆβˆ’βˆ˜LβŸ΅βˆ’βˆ˜R[π’ž,β„°]. [\mathcal{D}, \mathcal{E}] \underoverset {\underset{- \circ L}{\longrightarrow}} {\overset{- \circ R}{\longleftarrow}} {\;\;\;\;\bot\;\;\;\;} [\mathcal{C}, \mathcal{E}] \,.

Hence where LL was the left adjoint, its precomposition functor βˆ’βˆ˜L- \circ L becomes the right adjoint, and dually for RR.

The components Ξ· F:Fβ‡’F∘R∘L\eta_F : F \Rightarrow F \circ R \circ L of the unit and Ο΅ F:F∘L∘Rβ‡’F\epsilon_F : F \circ L \circ R \Rightarrow F of the counit are given by whiskering the original unit and counit with FF on the left.

By uniqueness of adjoints, this implies that left Kan extensions along LL are given by precomposition with RR, which is another way of saying that RR is the absolute left Kan extension of the identity functor along LL. Dually, right Kan extensions along RR are given by precomposition with LL.

There is also an induced adjunction of postcomposition functors between [β„°,π’ž][\mathcal{E}, \mathcal{C}] and [β„°,π’Ÿ][\mathcal{E}, \mathcal{D}] of the form

[β„°,π’Ÿ]βŠ₯⟢Rβˆ˜βˆ’βŸ΅Lβˆ˜βˆ’[β„°,π’ž]. [\mathcal{E}, \mathcal{D}] \underoverset {\underset{R \circ -}{\longrightarrow}} {\overset{L \circ -}{\longleftarrow}} {\;\;\;\;\bot\;\;\;\;} [\mathcal{E}, \mathcal{C}] \,.

The components Ξ· F:Fβ‡’R∘L∘F\eta_F : F \Rightarrow R \circ L \circ F of the unit and Ο΅ F:L∘R∘Fβ‡’F\epsilon_F : L \circ R \circ F \Rightarrow F of the counit are given by whiskering the original unit and counit with FF on the right.

By uniqueness of adjoints, this implies that left Kan lifts along RR are given by postcomposition with LL, which is another way of saying that LL is the absolute left Kan lift of the identity functor along RR. Dually, right Kan lifts along LL are given by postcomposition with RR.

Slicing of adjoint functors

Proposition

(sliced adjoints)
Let

π’ŸβŠ₯⟢R⟡Lπ’ž \mathcal{D} \underoverset {\underset{\;\;\;\;R\;\;\;\;}{\longrightarrow}} {\overset{\;\;\;\;L\;\;\;\;}{\longleftarrow}} {\bot} \mathcal{C}

be a pair of adjoint functors (adjoint ∞-functors), where the category (∞-category) π’ž\mathcal{C} has all pullbacks (homotopy pullbacks).

Then:

  1. For every object bβˆˆπ’žb \in \mathcal{C} there is induced a pair of adjoint functors between the slice categories (slice ∞-categories) of the form

    (11)π’Ÿ /L(b)βŠ₯⟢R /b⟡L /bπ’ž /b, \mathcal{D}_{/L(b)} \underoverset {\underset{\;\;\;\;R_{/b}\;\;\;\;}{\longrightarrow}} {\overset{\;\;\;\;L_{/b}\;\;\;\;}{\longleftarrow}} {\bot} \mathcal{C}_{/b} \mathrlap{\,,}

    where:

    • L /bL_{/b} is the evident induced functor (applying LL to the entire triangle diagrams in π’ž\mathcal{C} which represent the morphisms in π’ž /b\mathcal{C}_{/b});

    • R /bR_{/b} is the composite

      R /b:π’Ÿ /L(b)⟢Rπ’ž /(R∘L(b))⟢(Ξ· b) *π’ž /b R_{/b} \;\colon\; \mathcal{D}_{/{L(b)}} \overset{\;\;R\;\;}{\longrightarrow} \mathcal{C}_{/{(R \circ L(b))}} \overset{\;\;(\eta_{b})^*\;\;}{\longrightarrow} \mathcal{C}_{/b}

      of

      1. the evident functor induced by RR;

      2. the (homotopy) pullback along the (L⊣R)(L \dashv R)-unit at bb (i.e. the base change along η b\eta_b).

  2. For every object bβˆˆπ’Ÿb \in \mathcal{D} there is induced a pair of adjoint functors between the slice categories of the form

    (12)π’Ÿ /bβŠ₯⟢R /b⟡L /bπ’ž /R(b), \mathcal{D}_{/b} \underoverset {\underset{\;\;\;\;R_{/b}\;\;\;\;}{\longrightarrow}} {\overset{\;\;\;\;L_{/b}\;\;\;\;}{\longleftarrow}} {\bot} \mathcal{C}_{/R(b)} \mathrlap{\,,}

    where:

    • R /bR_{/b} is the evident induced functor (applying RR to the entire triangle diagrams in π’Ÿ\mathcal{D} which represent the morphisms in π’Ÿ /b\mathcal{D}_{/b});

    • L /bL_{/b} is the composite

      L /b:π’Ÿ /R(b)⟢Lπ’ž /(L∘R(b))⟢(Ο΅ b) !π’ž /b L_{/b} \;\colon\; \mathcal{D}_{/{R(b)}} \overset{\;\;L\;\;}{\longrightarrow} \mathcal{C}_{/{(L \circ R(b))}} \overset{\;\;(\epsilon_{b})_!\;\;}{\longrightarrow} \mathcal{C}_{/b}

      of

      1. the evident functor induced by LL;

      2. the composition with the (L⊣R)(L \dashv R)-counit at bb (i.e. the left base change along ϡ b\epsilon_b).

The first statement appears, in the generality of (∞,1)-category theory, as HTT, prop. 5.2.5.1. For discussion in model category theory see at sliced Quillen adjunctions.
Proof

(in 1-category theory)

Recall that (this Prop.) the hom-isomorphism that defines an adjunction of functors (this Def.) is equivalently given in terms of composition with

  • the adjunction unit Ξ· c:cβ†’R∘L(c)\;\;\eta_c \colon c \xrightarrow{\;} R \circ L(c)

  • the adjunction counit Ο΅ d:L∘R(d)β†’d\;\;\epsilon_d \colon L \circ R(d) \xrightarrow{\;} d

as follows:

Using this, consider the following transformations of morphisms in slice categories, for the first case:

(1a)

(2a)

(2b)

(1b)

Here:

  • (1a) and (1b) are equivalent expressions of the same morphism ff in π’Ÿ /L(b)\mathcal{D}_{/L(b)}, by (at the top of the diagrams) the above expression of adjuncts between π’ž\mathcal{C} and π’Ÿ\mathcal{D} and (at the bottom) by the triangle identity.

  • (2a) and (2b) are equivalent expression of the same morphism f˜\tilde f in π’ž /b\mathcal{C}_{/b}, by the universal property of the pullback.

Hence:

  • starting with a morphism as in (1a) and transforming it to (2)(2) and then to (1b) is the identity operation;

  • starting with a morphism as in (2b) and transforming it to (1) and then to (2a) is the identity operation.

In conclusion, the transformations (1) ↔\leftrightarrow (2) consitute a hom-isomorphism that witnesses an adjunction of the first claimed form (11).


The second case follows analogously, but a little more directly since no pullback is involved:

(1a)

(2)

(1b)

In conclusion, the transformations (1) ↔\leftrightarrow (2) consitute a hom-isomorphism that witnesses an adjunction of the second claimed form (12).

Remark

(left adjoint of sliced adjunction forms adjuncts)
The sliced adjunction (Prop. ) in the second form (12) is such that the sliced left adjoint sends slicing morphism Ο„\tau to their adjuncts Ο„Λœ\widetilde{\tau}, in that (again by this Prop.):

L /d(c ↓ Ο„ R(b))=(L(c) ↓ Ο„Λœ b)βˆˆπ’Ÿ /b L_{/d} \, \left( \array{ c \\ \big\downarrow {}^{\mathrlap{\tau}} \\ R(b) } \right) \;\; = \;\; \left( \array{ L(c) \\ \big\downarrow {}^{\mathrlap{\widetilde{\tau}}} \\ b } \right) \;\;\; \in \; \mathcal{D}_{/b}

The two adjunctions in admit the following joint generalisation, which is proven HTT, lem. 5.2.5.2. (Note that the statement there is even more general and here we only use the case where K=Ξ” 0K = \Delta^0.)

Proposition

(sliced adjoints)
Let

π’žβŠ₯⟡R⟢Lπ’Ÿ \mathcal{C} \underoverset {\underset{\;\;\;\;R\;\;\;\;}{\longleftarrow}} {\overset{\;\;\;\;L\;\;\;\;}{\longrightarrow}} {\bot} \mathcal{D}

be a pair of adjoint ∞-functors, where the ∞-category π’ž\mathcal{C} has all homotopy pullbacks. Suppose further we are given objects cβˆˆπ’žc \in \mathcal{C} and dβˆˆπ’Ÿd \in \mathcal{D} together with a morphism Ξ±:cβ†’R(d)\alpha: c \to R(d) and its adjunct Ξ²:L(c)β†’d\beta:L(c) \to d.

Then there is an induced a pair of adjoint ∞-functors between the slice ∞-categories of the form

(13)π’ž /cβŠ₯⟡R /b⟢L /bπ’Ÿ /d, \mathcal{C}_{/c} \underoverset {\underset{\;\;\;\;R_{/b}\;\;\;\;}{\longleftarrow}} {\overset{\;\;\;\;L_{/b}\;\;\;\;}{\longrightarrow}} {\bot} \mathcal{D}_{/d} \mathrlap{\,,}

where:

  • L /cL_{/c} is the composite

    L /c:π’ž /c⟢Lπ’Ÿ /L(c)⟢β !π’Ÿ /d L_{/c} \;\colon\; \mathcal{C}_{/{c}} \overset{\;\;L\;\;}{\longrightarrow} \mathcal{D}_{/{L(c)}} \overset{\;\;\beta_!\;\;}{\longrightarrow} \mathcal{D}_{/d}

    of

    1. the evident functor induced by LL;

    2. the composition with Ξ²:L(c)β†’d\beta:L(c) \to d (i.e. the left base change along Ξ²\beta).

  • R /dR_{/d} is the composite

    R /d:π’Ÿ /d⟢Rπ’ž /R(d)⟢(Ξ± *π’ž /c R_{/d} \;\colon\; \mathcal{D}_{/{d}} \overset{\;\;R\;\;}{\longrightarrow} \mathcal{C}_{/{R(d)}} \overset{\;\;(\alpha^*\;\;}{\longrightarrow} \mathcal{C}_{/c}

    of

    1. the evident functor induced by RR;

    2. the homotopy along α:c→R(d)\alpha:c \to R(d) (i.e. the base change along α\alpha).

Examples

The central point about examples of adjoint functors is:

Adjoint functors are ubiquitous .

To a fair extent, category theory is all about adjoint functors and the other universal constructions: Kan extensions, limits, representable functors, which are all special cases of adjoint functors – and adjoint functors are special cases of these.

Listing examples of adjoint functors is much like listing examples of integrals in analysis: one can and does fill books with these. (In fact, that analogy has more to it than meets the casual eye: see coend for more).

Keeping that in mind, we do list some special cases and special classes of examples that are useful to know. But any list is necessarily wildly incomplete.

General

  • A pair of adjoint functors between posets is a Galois correspondence.

  • A pair of adjoint functors (L⊣R)(L \dashv R) where RR is a full and faithful functor exhibits a reflective subcategory.

    In this case LL may be regarded as a localization. The fact that the adjunction provides universal factorization through unit and counit in this case means that every morphism f:c→Rdf : c \to R d into a local object factors through the localization of cc.

  • A pair of adjoint functors that is also an equivalence of categories is called an adjoint equivalence.

  • A pair of adjoint functors where CC and DD have finite limits and LL preserves these finite limits is a geometric morphism. These are one kind of morphisms between toposes. If in addition RR is full and faithful, then this is a geometric embedding.

  • The left and right adjoint functors p !p_! and p *p_* (if they exist) to a functor p *:[Kβ€²,C]β†’[K,C]p^* : [K',C] \to [K,C] between functor categories obtained by precomposition with a functor p:Kβ†’Kβ€²p : K \to K' of diagram categories are called the left and right Kan extension functors along pp

    (Lan p⊣p *⊣Ran p):=(p !⊣p *⊣p *):[K,C]β†’p *←p *β†’p ![Kβ€²,C]. (Lan_p \dashv p^* \dashv Ran_p) := (p_! \dashv p^* \dashv p_*) : [K,C] \stackrel{\overset{p_!}{\to}}{\stackrel{\overset{p^*}{\leftarrow}}{\underset{p_*}{\to}}} [K',C] \,.

    If Kβ€²=*K' = {*} is the terminal category then this are the limit and colimit functors on [K,C][K,C].

    If C=C = Set then this is the direct image and inverse image operation on presheaves.

  • if RR is regarded as a forgetful functor then its left adjoint LL is a regarded as a free functor.

  • If CC is a category with small colimits and KK is a small category (a diagram category) and Q:Kβ†’CQ : K \to C is any functor, then this induces a nerve and realization pair of adjoint functors

    (|βˆ’| Q⊣N Q):Cβ†’N Q←|βˆ’| Q[K op,Set] (|-|_Q \dashv N_Q) : C \stackrel{\overset{|-|_Q}{\leftarrow}}{\underset{N_Q}{\to}} [K^{op}, Set]

    between CC and the category of presheaves on KK, where

    • the nerve functor is given by

      N Q(c):=Hom C(Q(βˆ’),c):k↦Hom C(Q(k),c) N_Q(c) := Hom_C(Q(-),c) : k \mapsto Hom_C(Q(k),c)
    • and the realization functor is given by the coend

      |F| Q:=∫ k∈KQ(k)β‹…F(k), |F|_Q := \int^{k \in K} Q(k)\cdot F(k) \,,

      where in the integrand we have the canonical tensoring of CC over Set (Q(k)β‹…F(k)=∐ s∈F(k)Q(k)Q(k) \cdot F(k) = \coprod_{s \in F(k)} Q(k)).

    A famous examples of this is obtained for C=C = Top, K=Ξ”K = \Delta the simplex category and Q:Ξ”β†’TopQ : \Delta \to Top the functor that sends [n][n] to the standard topological nn-simplex. In this case the nerve functor is the singular simplicial complex functor and the realization is ordinary geometric realization.

Related concepts

References

For the basics, see any text on category theory (and see the references at adjunction), for instance:

Though the definition of an adjoint equivalence appears in Grothendieck's Tohoku paper, the idea of adjoint functors in general goes back to

  • Daniel Kan, Adjoint functors, Transactions of the American Mathematical Society 87 2 (1958) 294-329 [jstor:1993102]

and in relation to co/monads to

and its fundamental relevance for category theory was highlighted in

  • Peter Freyd, Abelian categories – An introduction to the theory of functors, Harper’s Series in Modern Mathematics, Harper & Row, New York (1964) [pdf]

  • William Lawvere, Adjointness in Foundations, (TAC), Dialectica 23 (1969), 281-296

The history of the idea that adjoint functors formalize aspects of dialectics is recounted in

  • Joachim Lambek, The Influence of Heraclitus on Modern Mathematics, In Scientific Philosophy Today: Essays in Honor of Mario Bunge, edited by Joseph Agassi and Robert S Cohen, 111–21. Boston: D. Reidel Publishing Co. (1982) (doi:10.1007/978-94-009-8462-2_6)

    (more along these lines at adjoint modality)

See also:

More on the notion of transformation of adjoints:

Last revised on September 22, 2024 at 13:58:02. See the history of this page for a list of all contributions to it.