nLab
adjunction

Contents

Idea

Two morphisms Cβ†’LDC \stackrel{L}{\to} D and Dβ†’RCD \stackrel{R}{\to} C in a 2-category π’ž\mathcal{C} form an adjunction if they are dual to each other.

There are two archetypical examples:

  • If AA is a monoidal category and π’ž=BA\mathcal{C} = \mathbf{B}A is the one-object 2-category incarnation of AA (the delooping of AA), so that the morphisms in π’ž\mathcal{C} correspond to the objects of AA, then the notion of adjoint morphisms in π’ž\mathcal{C} coincides precisely with the notion of dual objects in a AA.

  • If π’ž\mathcal{C} is the 22-category Cat, so that the morphisms in π’ž\mathcal{C} are functors, then the notion of adjoint morphisms in π’ž\mathcal{C} coincides precisely with the notion of adjoint functors.

General

The notion of adjunction may usefully be thought of as a weakened version of the notion of equivalence in a 2-category: a morphism in an adjunction need not be invertible, but it has in some sense a left inverse from below and a right inverse from above. If the morphism in an adjunction does happen to be a genuine equivalence, then we speak of the adjunction being an adjoint equivalence.

Essentially everything that makes category theory nontrivial and interesting beyond groupoid theory can be derived from the concept of adjoint functors. In particular universal constructions such as limits and colimits are examples of certain adjunctions. Adjunctions are already interesting (but simpler) in 2-posets, such as the 22-poset Pos of posets.

From hom-functors to units and counits

At the cost of some repetition (compare adjoint functor), we outline how one gets from the hom-functor formulation of adjunction in Cat to the elementary definition in terms of units and counits. This will motivate the definition in the section that follows, which is elementary (definable in the first-order theory of categories) and portable to any 2-category.

We start from a familiar example. Let U:Grpβ†’SetU: Grp \to Set from groups to sets be the usual forgetful functor. When we say ”F(X)F(X) is the free group generated by a set XX”, we mean there is a function Ξ· X:Xβ†’U(F(X))\eta_X: X \to U(F(X)) which is universal among functions from XX to the underlying set of a group, which means in turn that given a function f:Xβ†’U(G)f: X \to U(G), there is a unique group homomorphism g:F(X)β†’Gg: F(X) \to G such that

f=(X→η XU(F(X))→U(g)U(G))f = (X \stackrel{\eta_X}{\to} U(F(X)) \stackrel{U(g)}{\to} U(G))

Here Ξ· X\eta_X is a component of what we call the unit of the adjunction F⊣UF \dashv U, and the equation above is a recipe for the relationship between the map g:F(X)β†’Gg: F(X) \to G and the map f:Xβ†’U(G)f: X \to U(G) in terms of the unit.

Now we work more generally. Suppose given functors L:C→DL: C \to D, R:D→CR: D \to C and the structure of an adjunction in the form of a natural isomorphism

Ξ¨ c,d:hom D(L(c),d)β‰…hom C(c,R(d))\Psi_{c, d}: \hom_D(L(c), d) \cong \hom_C(c, R(d))

Now the idea is that, a la the Yoneda lemma, Ξ¨\Psi should be completely describable in terms of what it does to identity maps. With that in mind, define the unit Ξ·:1 Cβ†’RL\eta : 1_C \to R L by the formula Ξ· c=Ξ¨ c,L(c)(1 L(c))\eta_c = \Psi_{c, L(c)}(1_{L(c)}). Dually, define the counit Ξ΅:LRβ†’1 D\varepsilon : L R \to 1_D by the formula Ξ΅ d=Ξ¨ R(d),d βˆ’1(1 R(d))\varepsilon_d = \Psi^{-1}_{R(d), d}(1_{R(d)}). Then given g:L(c)β†’dg: L(c) \to d, the claim is that

Ψ c,d(g)=(c→η cR(L(c))→R(g)R(d)).\Psi_{c, d}(g) = (c \stackrel{\eta_c}{\to} R(L(c)) \stackrel{R(g)}{\to} R(d)).

This may be left as an exercise in the yoga of the Yoneda lemma, applied to hom D(L(c),βˆ’)β†’hom C(c,R(βˆ’))\hom_D(L(c), -) \to \hom_C(c, R(-)). By duality, given f:cβ†’R(d)f: c \to R(d),

Ξ¨ c,d βˆ’1(f)=(L(c)β†’L(f)L(R(d))β†’Ξ΅ dd).\Psi^{-1}_{c, d}(f) = (L(c) \stackrel{L(f)}{\to} L(R(d)) \stackrel{\varepsilon_d}{\to} d).

(In fact, we spell out the Yoneda-lemma proof of this dual form below.)

Finally, these operations should obviously be mutually inverse, but that can again be entirely encapsulated Yoneda-wise in terms of the effect on identity maps. Thus, if Ξ· c≔Ψ c,L(c)(1 L(c))\eta_c \coloneqq \Psi_{c, L(c)}(1_{L(c)}), the recipe just given for Ξ¨ βˆ’1\Psi^{-1} yields back

1 L(c)=(L(c)β†’L(Ξ· c)LRL(c)β†’Ξ΅ L(c)L(c))1_{L(c)} = (L(c) \stackrel{L(\eta_c)}{\to} L R L(c) \stackrel{\varepsilon_{L(c)}}{\to} L(c))

and this is one of the famous triangular equations: 1 L=(Lβ†’LΞ·LRLβ†’Ξ΅LL)1_L = (L \stackrel{L \eta}{\to} L R L \stackrel{\varepsilon L}{\to} L). Note that juxtaposition in the diagram above is neither functor application, nor vertical composition, nor horizontal composition, but is actually whiskering. By duality, we have the other triangular equation 1 R=(Rβ†’Ξ·RRLRβ†’RΞ΅R)1_R = (R \stackrel{\eta R}{\to} R L R \stackrel{R \varepsilon}{\to} R). These two triangular equations are enough to guarantee that the recipes for Ξ¨\Psi and Ξ¨ βˆ’1\Psi^{-1} are indeed mutually inverse.

Thus, it is perfectly sufficient to define an adjoint pair of functors in CatCat as given by unit and counit transformations η:1 C→RL\eta: 1_C \to R L, Ρ:LR→1 D\varepsilon: L R \to 1_D, satisfying triangular equations as above.

One thing often heard is that the definition of adjunctions via units and counits is an β€œelementary” definition (so that by implication, the formulation in terms of hom-functors is not elementary). This means that whereas the hom-functor formulation relies on a background category of sets, the formulation in terms of units and counits is purely in the first-order language of categories and makes no reference to a background model of set theory. It is therefore a perfectly serviceable definition of adjunction without assumptions of local smallness.

Yoneda-lemma argument

We claim that Ξ¨ c,d βˆ’1:hom C(c,R(d))β†’hom D(L(c),d)\Psi^{-1}_{c, d}: \hom_C(c, R(d)) \to \hom_D(L(c), d) can be defined by the formula

Ξ¨ c,d βˆ’1(f:cβ†’R(d))=(L(c)β†’L(f)L(R(d))β†’Ξ΅ dd)\Psi^{-1}_{c, d}(f: c \to R(d)) = (L(c) \stackrel{L(f)}{\to} L(R(d)) \stackrel{\varepsilon_d}{\to} d)

where Ξ΅ d≔Ψ R(d),d βˆ’1(1 R(d))\varepsilon_d \coloneqq \Psi^{-1}_{R(d), d}(1_{R(d)}). This is by appeal to the proof of the Yoneda lemma applied to the transformation

Ξ¨ βˆ’,d βˆ’1:hom C(βˆ’,R(d))β†’hom D(L(βˆ’),d)\Psi^{-1}_{-, d}: \hom_C(-, R(d)) \to \hom_D(L(-), d)

For the naturality of Ξ¨ βˆ’1\Psi^{-1} in the argument (βˆ’)(-) would imply that given f:cβ†’R(d)f: c \to R(d), we have a commutative square

hom C(R(d),R(d)) β†’Ξ¨ R(d),d βˆ’1 hom D(L(R(d)),d) hom C(f,R(d))↓ ↓hom D(L(f),d) hom C(c,R(d)) β†’Ξ¨ c,d βˆ’1 hom D(L(c),d)\array{ \hom_C(R(d), R(d)) & \stackrel{\Psi^{-1}_{R(d), d}}{\to} & \hom_D(L(R(d)), d) \\ \hom_C(f, R(d)) \downarrow & & \downarrow \hom_D(L(f), d) \\ \hom_C(c, R(d)) & \underset{\Psi^{-1}_{c, d}}{\to} & \hom_D(L(c), d) }

Chasing the element 1 R(d)1_{R(d)} down and then across, we get f:cβ†’R(d)f: c \to R(d) and then Ξ¨ c,d βˆ’1(f)\Psi^{-1}_{c, d}(f). Chasing across and then down, we get Ξ΅ d\varepsilon_d and then Ξ΅ d∘L(f)\varepsilon_d \circ L(f). This completes the verification of the claim.

Definition

Direct definition

An adjunction in a 2-category is a pair of objects C,DC,D together with morphisms L:Cβ†’DL: C \to D, R:Dβ†’CR : D \to C and 2-morphisms Ξ·:1 Cβ†’R∘L\eta: 1_C \to R \circ L, Ο΅:L∘Rβ†’1 D\epsilon: L \circ R \to 1_D satisfying the equations

(RΟ΅)β‹…(Ξ·R)=1 Ri.e.Rβ†’Ξ·βˆ˜1 RR∘L∘Rβ†’1 R∘ϡR=Rβ†’1 RRi.e. 1 C ⇓η D β†’R C β†’L D β†’R C ⇓ϡ 1 D =Dβ†’RC (R \epsilon) \cdot (\eta R) = 1_R \qquad \text{i.e.} \qquad R \stackrel{\eta \circ 1_R}{\to} R \circ L \circ R \stackrel{1_R \circ \epsilon}{\to} R = R \stackrel{1_R}{\to} R \qquad \text{i.e.} \qquad \array{\arrayopts{ \padding{0} } &&&&1_C& \\ &&\cellopts{\colspan{5}}\begin{svg} <svg xmlns="http://www.w3.org/2000/svg" width="8.5em" height="2em" viewBox="0 0 85 20"> <defs> <marker id='svg195arrowhead' markerHeight='5' markerUnits='strokeWidth' markerWidth='8' orient='auto' refX='0' refY='5' viewBox='0 0 10 10'> <path d='M 0 0 L 10 5 L 0 10 z'/> </marker> </defs> <path marker-end='url(#svg195arrowhead)' stroke-width="1" stroke="#000" fill="none" d="M5 15q40-28 75 0"/> <foreignObject height='20' width='20' x='40' y='3' font-size='10'><math xmlns="http://www.w3.org/1998/Math/MathML" display='inline'><msup><mo>&#8659;</mo><mi>&#951;</mi></msup></math></foreignObject> </svg> \end{svg}\\ D & \stackrel{R}{\to}& C & \stackrel{L}{\to}& D & \stackrel{R}{\to}& C \\ \cellopts{\colspan{4}}\begin{svg} <svg xmlns="http://www.w3.org/2000/svg" width="8.5em" height="2em" viewBox="0 0 85 20"> <path marker-end='url(#svg195arrowhead)' stroke-width="1" stroke="#000" fill="none" d="M5 5q40 28 75 0"/> <foreignObject height='20' width='20' x='40' y='0' font-size='10'><math xmlns="http://www.w3.org/1998/Math/MathML" display='inline'><msup><mo>&#8659;</mo><mi>&#1013;</mi></msup></math></foreignObject> </svg> \end{svg} \\ &&1_D& } \quad = \quad D \stackrel{R}{\to} C

and

(Ο΅L)β‹…(LΞ·)=1 Li.e.Lβ†’1 L∘ηL∘R∘Lβ†’Ο΅βˆ˜1 LL=Lβ†’1 LLi.e. 1 C ⇓η C β†’L D β†’R C β†’L D ⇓ϡ 1 D =Cβ†’LD (\epsilon L) \cdot (L \eta) = 1_L \qquad \text{i.e.} \qquad L \stackrel{1_L \circ \eta}{\to} L \circ R \circ L \stackrel{\epsilon \circ 1_L}{\to} L = L \stackrel{1_L}{\to} L \qquad \text{i.e.} \qquad \array{\arrayopts{ \padding{0} } &&1_C& \\ \cellopts{\colspan{5}}\begin{svg} <svg xmlns="http://www.w3.org/2000/svg" width="8.5em" height="2em" viewBox="0 0 85 20"> <defs> <marker id='svg195arrowhead' markerHeight='5' markerUnits='strokeWidth' markerWidth='8' orient='auto' refX='0' refY='5' viewBox='0 0 10 10'> <path d='M 0 0 L 10 5 L 0 10 z'/> </marker> </defs> <path marker-end='url(#svg195arrowhead)' stroke-width="1" stroke="#000" fill="none" d="M5 15q40-28 75 0"/> <foreignObject height='20' width='20' x='40' y='3' font-size='10'><math xmlns="http://www.w3.org/1998/Math/MathML" display='inline'><msup><mo>&#8659;</mo><mi>&#951;</mi></msup></math></foreignObject> </svg> \end{svg}\\ C & \stackrel{L}{\to}& D & \stackrel{R}{\to}& C & \stackrel{L}{\to}& D \\ &&\cellopts{\colspan{4}}\begin{svg} <svg xmlns="http://www.w3.org/2000/svg" width="8.5em" height="2em" viewBox="0 0 85 20"> <path marker-end='url(#svg195arrowhead)' stroke-width="1" stroke="#000" fill="none" d="M5 5q40 28 75 0"/> <foreignObject height='20' width='20' x='40' y='0' font-size='10'><math xmlns="http://www.w3.org/1998/Math/MathML" display='inline'><msup><mo>&#8659;</mo><mi>&#1013;</mi></msup></math></foreignObject> </svg> \end{svg} \\ &&&&1_D& } \quad = \quad C \stackrel{L}{\to} D

variously called the triangle identities or the zig-zag identities. We call LL the left adjoint (of RR) and RR the right adjoint (of LL). We call Ξ·\eta the unit of the adjunction and Ο΅\epsilon the counit of the adjunction.

When interpreted in the prototypical 2-category Cat, CC and DD are categories, LL and RR are functors, and Ξ·\eta and Ο΅\epsilon are natural transformations. In this case (which was of course the first to be defined) there are a number of equivalent definitions of an adjunction, which can be found on the page adjoint functor. Conversely, the definition in any 2-category can be obtained by internalization from the definition in Cat\Cat.

In terms of string diagrams

The definition of an adjunction may be nicely expressed using string diagrams. The data L:Cβ†’DL: C \to D, R:Dβ†’CR : D \to C and 2-cells Ξ·:1 Cβ†’R∘L\eta: 1_C \to R \circ L, Ο΅:L∘Rβ†’1 D\epsilon: L \circ R \to 1_D are depicted as

String diagram for a left adjoint (for 'Adjunction') Β Β Β Β Β Β Β Β  String diagram of a right adjoint (for 'Adjunction') Β Β Β Β Β Β Β Β  String diagram of an adjunction unit (for 'Adjunction') Β Β Β Β Β Β Β Β  String diagram of an adjunction co-unit (for 'Adjunction')

(where 1-cells read from right to left and 2-cells from bottom to top), and the zigzag identities are expressed as β€œpulling zigzags straight” (hence the name):

String diagram of first zigzag identity (for 'Adjunction') Β Β Β Β Β Β Β Β 

Often, arrows on strings are used to distinguish LL and RR, and most or all other labels are left implicit; so the zigzag identities, for instance, become:

Minimal string diagram of first zigzag identity (for 'Adjunction') Β Β Β Β Β Β Β Β  Minimal string diagram of second zigzag identity (for 'Adjunction')

Properties

Relation to monads

See at monad – Relation to adjunctions.

Examples

References

Revised on February 27, 2014 08:32:43 by Urs Schreiber (82.113.98.143)