Michael Shulman
exponentials in a 2-category

Cat is not locally cartesian closed

The 2-category Cat is cartesian closed, in an appropriate 2-categorical sense; see also the discussion here. However, it is not locally cartesian closed. This failure is fundamental and has nothing to do with strictness or size issues; pullbacks just don’t preserve colimits. For example, let X=(012) be the ordinal 3; then there is a pushout

(1) (12) (01) (012)\array{(1) & \to & (1\to 2)\\ \downarrow && \downarrow\\ (0\to 1) & \to & (0\to 1\to 2)}

in Cat/X which pulls back along the inclusion (02)X to

(2) (0) (02)\array{\emptyset & \to & (2)\\ \downarrow && \downarrow\\ (0) & \to & (0\to 2)}

which is certainly not a pushout. Note that the same counterexample applies in Pos.

It is similarly easy to write down examples of coinserters, coinverters, and coequifiers that are not preserved by pullbacks. Coproducts are preserved by pullback (in fact, Cat is extensive), as are quotients of 2-congruences (since Cat is regular), but these seem to be about it for pullback-stable colimits in Cat.

2-categories with exponentials

However, there are a number of other useful exponentiability properties that do hold when K=Cat.

  1. Fibrations and opfibrations are exponentiable. That is, if f:AX is an (op)fibration, then exponentials (BX) (AX) exist in the slice 2-category K/X. Equivalently, f *:K/XK/A has a right adjoint Π f. (Fibrations and opfibrations are not the only exponentiable morphisms in Cat, but they are certainly the most important and most commonly encountered ones.)

  2. If AX is an opfibration and BX is a fibration, then the exponential (BX) (AX) in K/X is a fibration, and dually.

  3. For any f:YX, the functor f *:Fib(X)Fib(Y) has a right adjoint Ran f, and likewise for f *:Opf(X)Opf(Y).

  4. For any f:YX, the functors f *:DFib(X)DFib(Y) and f *:DOpf(X)DOpf(Y) have right adjoints Ran f.

  5. The 2-categories Fib(X) and Opf(X) are all cartesian closed.

  6. K itself is cartesian closed.

  7. The 2-categories DFib(X) and DOpf(X) are all locally cartesian closed.

Note that Ran f and Π f are not the same even when both exist, and likewise the exponentials in Fib(X) are not the same as the exponentials in K/X even when both exist. The latter are better-behaved in some ways, for instance they are stable under pullback (because they are “fiberwise”).

Perhaps surprisingly, it turns out that the first of these properties is sufficient to imply all the others. The goal of the rest of this page is to prove this claim. Therefore, we define:

Definition

A 2-category K with finite limits is said to have exponentials if all fibrations and opfibrations in K are exponentiable.

Note that a 1-category, and in fact any (2,1)-category, has exponentials if and only if it is locally cartesian closed, since every morphism is a fibration and opfibration.

Basic observations

We begin with a couple of easy observations.

Proposition

If either fibrations or opfibrations are exponentiable in K, then K is cartesian closed.

Proof

Of course, K is cartesian closed just when every morphism A1 is exponentiable; but every such morphism is both a fibration and an opfibration.

Proposition

If AX is an opfibration and BX is a fibration, then the exponential (BX) (AX) in K/X (if it exists) is a fibration, and dually.

Proof

We say briefly how to construct an action morphism

X 2× X(BX) (AX)(BX) (AX).X ^{\mathbf{2}} \times_X(B\to X)^{(A\to X)} \to (B\to X)^{(A\to X)}.

Of course, by adjointness it suffices to construct a morphism

A× XX 2× X(BX) (AX)BA\times_X X ^{\mathbf{2}} \times_X (B\to X)^{(A\to X)} \to B

over X. Now the left-hand side is a limit of the diagram

(BX) (AX) X 2 X A X.\array{ &&&& (B\to X)^{(A\to X)}\\ &&&& \downarrow\\ && X ^{\mathbf{2}} &\to & X\\ && \downarrow\\ A & \to & X.}

We first map it via a diagonal to the limit of

(BX) (AX) X 2 X A X X 2\array{ &&&& (B\to X)^{(A\to X)}\\ &&&& \downarrow\\ && X ^{\mathbf{2}} &\to & X\\ && \downarrow && \uparrow\\ A & \to & X & \leftarrow & X ^{\mathbf{2}}}

then use the covariant action of X on A to map to the limit of

(BX) (AX) A X X X 2\array{ && (B\to X)^{(A\to X)}\\ && \downarrow\\ A &\to & X\\ && \uparrow\\ X & \leftarrow & X ^{\mathbf{2}}}

then the evaluation A× X(BX) (AX)B to get to the limit of

B X X X 2\array{ && B\\ && \downarrow\\ && X\\ && \uparrow\\ X & \leftarrow & X ^{\mathbf{2}}}

and finally the contravariant action of X on B to get to

B X.\array{ B\\ \downarrow \\ X.}

It is then straightforward to check that this action makes (BX) (AX) into a fibration.

Comonadicity

The key observation for many of the proofs below is the following.

Lemma

If K has exponentials, then for any X, Fib(X) and Opf(X) are comonadic, as well as monadic, over K/X.

Proof

The monad on K/X whose category of algebras is Fib(X) takes AX to X 2× XA, or equivalently Σ st *A where (s,t):X 2X are the two projections. But t is an opfibration, so Σ st * has a right adjoint Π ts *. And it is a standard result, valid for 2-categories as for 1-categories, that when the underlying functor of a monad has a right adjoint, its right adjoint becomes a comonad whose category of coalgebras is equivalent to the category of algebras for the original monad.

Corollary

If K has exponentials, then Fib(X) and Opf(X) inherit any colimits possessed by K.

Proof

The category of coalgebras for any comonad inherits colimits from the base category, and slice (2-)categories always inherit colimits.

For our main applications of comonadicity, we require the following observation.

Proposition

Given a commutative square

D 1 G¯ D 2 U 1 U 2 C 1 G C 2\array{D_1 & \overset{\overline{G}}{\to} & D_2\\ ^{U_1}\downarrow && \downarrow ^{U_2}\\ C_1 & \underset{G}{\to} & C_2}

of functors between 2-categories, if U 1 and U 2 are monadic, D 1 has reflexive codescent objects, and G has a left adjoint, then G¯ also has a left adjoint. Dually, if U 1 and U 2 are comonadic, D 1 has reflexive descent objects, and G has a right adjoint, then G¯ also has a right adjoint.

Proof

The 1-categorical version of this, referring to reflexive (co)equalizers, is well-known; see for instance

  • P. T. Johnstone, Adjoint lifting theorems for categories of algebras.

The idea is the same as that in Beck’s (co)monadicity theorem: we express any object of D 2 as a reflexive coequalizer of free algebras, then apply the left adjoint of G to obtain a reflexive pair of free algebras in D 1 and take its coequalizer. The 2-categorical version is analogous, using the ideas of a 2-categorical monadicity theorem as found, for example, in

  • Claudio Hermida, Descent on 2-fibrations and strongly 2-regular 2-categories.

The dual is, of course, obvious.

Slicing and cartesian closure

Using comonadicity, we can show that certain exponentials are stable under slicing. First we observe:

Lemma

If YX is a fibration, then Fib(X)/Y is monadic over K/Y. If additionally K has exponentials, then Fib(X)/Y is also comonadic over K/Y.

Proof

The first statement is an instance of a general fact: if T is a monad on a (2-)category C and Y is a T-algebra, then there is an induced monad T Y on C/Y whose (2-)category of algebras is TAlg/Y, defined by taking AY to the composite TATYY.

The second statement is also an instance of a general fact: if such a T has a right adjoint G, then T Y also has a right adjoint G Y defined to take AY to the pullback

G YA GA Y GY.\array{G_Y A & \to & G A\\ \downarrow && \downarrow\\ Y & \to & G Y.}

Here the lower map YGY is the adjunct of the algebra structure map TYY.

Proposition

If K has exponentials, then a morphism in Fib(X) or Opf(X) is exponentiable if its underlying morphism in K is so. In particular, if K has exponentials, then for any X, fibrations are exponentiable in Fib(X) and opfibrations are exponentiable in Opf(X).

Proof

Suppose that f:ZY is a morphism in Fib(X) and that f is exponentiable in K. Then we have a commutative square

Fib(X)/Y f * Fib(X)/Z K/Y f * K/Z\array{Fib(X)/Y & \overset{f^*}{\to} & Fib(X)/Z\\ \downarrow && \downarrow\\ K/Y& \underset{f^*}{\to} & K/Z}

in which the vertical functors are comonadic by Lemma 2, and the bottom functor f * has a right adjoint since f is exponentiable in K. Therefore, by Proposition 3, the top functor f * has a right adjoint as well. The second statement follows because the underlying morphism in K of any fibration in Fib(X) is a fibration in K, and dually (see the theorems on iterated fibrations).

Corollary

If K has exponentials and a duality involution, then Fib(X) and Opf(X) also have exponentials for any X.

Proof

After Proposition 4, it remains to show that opfibrations are exponentiable in Fib(X) (and dually, fibrations are exponentiable in Opf(X)). Note that an opfibration in Fib(X) will not, in general, be an opfibration in K. But with a duality involution we have Fib(X)Opf(X o), and opfibrations are exponentiable in Opf(X o), hence also in Fib(X).

We say that a 2-category K has local exponentials if K has exponentials and each 2-category Fib(X) and Opf(X) also has local exponentials. Of course, the recursion in this definition is not well-founded, but we can reformulate it in explicit terms to say that

  1. K has exponentials,
  2. Each Opf(X) and Fib(X) has exponentials,
  3. Each Opf Fib(X)(AX) and Fib Opf(X)(AX) has exponentials,
  4. and so on.

Since duality involutions are stable under fibrational slicing, Corollary 2 implies that if K has exponentials and a duality involution, then it has local exponentials. It would be nice to have a finite list of axioms that implies local exponentials without invoking a duality involution (since not all Grothendieck 2-toposes have dualities).

Corollary

If K has exponentials, then the 2-categories Fib(X) and Opf(X) are all cartesian closed.

Proof

Combine Propositions 4 and 1.

Corollary

If K has exponentials, then the categories DFib(X) and DOpf(X) are all locally cartesian closed.

Proof

Since right adjoints preserve discrete objects, Corollary 3 implies that DFib(X) and DOpf(X) are cartesian closed for any X. Now, given a discrete fibration AX, we have Fib K(A)Fib Fib(X)(AX) by the theorem on iterated fibrations, and so DFib K(A)DFib Fib(X)(AX)DFib K(X)/(AX). Thus, since DFib K(A) is cartesian closed, so is DFib K(X)/(AX), and thus DFib K(X) is locally cartesian closed.

Left Kan extensions

We now turn to the existence of left and right adjoints to pullback functors. So far what we know is

  • f *:K/YK/X always has a left adjoint Σ f given by composition with f.
  • If K has exponentials and f is a fibration or opfibration, then (by definition) f *:K/YK/X has a right adjoint Π f.
  • If K has a comprehensive factorization (such as when it is countably-coherent), then f *:DFib(Y)DFib(X) and f *:DOpf(Y)DOpf(X) have left adjoints Lan f, which satisfy the Beck-Chevalley condition for comma squares.

Note that even if f is a fibration, so that Σ f maps Fib(Y) to Fib(X), it will not in general be a left adjoint to f *:Fib(X)Fib(Y). However, if f is a discrete fibration, then Σ f:DFib(Y)DFib(X) is left adjoint to f *:DFib(X)DFib(Y), since DFib(X) is a full sub-2-category of K/X.

We now consider how to construct left adjoints for non-discrete fibrations.

Proposition

If K has exponentials and reflexive codescent objects, then each pullback functor f *:Fib(Y)Fib(X) and f *:Opf(Y)Opf(X) has a left adjoint Lan f.

Proof

Because of the existence of Σ f, by Proposition 3 it suffices to show that Fib(X) and Opf(X) always have reflexive codescent objects; but this follows from Corollary 1.

The basic first-order structure we are interested in doesn’t imply the existence of such codescent objects, and a Heyting 2-pretopos, such as FinCat, need not have them. But they can be constructed with some infinitary structure; see colimits in an n-pretopos.

I do not know whether these adjunctions always satisfy the Beck-Chevalley condition for comma squares (although they do in one important case; see below).

Right Kan extensions

We now consider the existence of right adjoints to pullback functors between fibrational slices.

Proposition

If K has exponentials, then f *:Opf(X)Opf(Y) and f *:Fib(X)Fib(Y) have right adjoints Ran f for any f.

Proof

If f is exponentiable in K, this follows directly from Lemma 1 and Proposition 3. In particular, this applies when f is a fibration or opfibration.

Now let f:YX be any morphism, and consider the comma square:

Comma Square (1/f) X Y X t s 1 f

Suppose first that all the left adjoints Lan g exist for opfibrational slices and satisfy the Beck-Chevalley condition for comma squares. Then this comma square gives us an equivalence f *Lan ts *, so that f *:Opf(X)Opf(Y) has the right adjoint Ran f:=Ran st *.

We now show that regardless of the overall existence of left adjoints, the particular value Lan ts *(A) exists for any AOpf(X) and is given by f *A. This clearly suffices to prove the result.

According to the proof of Proposition 5, Lan ts *(A) should be given by the codescent object of

Descent object R S T

where

T= Σ ts *A× YY 2 S= Σ t(s *A× (1/f)(1/f) 2)× YY 2 R= Σ t(s *A× (1/f)(1/f) 2× (1/f)(1/f) 2)× YY 2.\begin{aligned} T =& \Sigma_t s^* A \times_Y Y ^{\mathbf{2}}\\ S =& \Sigma_t \left(s^* A\times_{(1/f)} (1/f)^{\mathbf{2}}\right) \times_Y Y ^{\mathbf{2}}\\ R =& \Sigma_t \left(s^* A\times_{(1/f)} (1/f)^{\mathbf{2}} \times_{(1/f)} (1/f)^{\mathbf{2}}\right) \times_Y Y ^{\mathbf{2}}. \end{aligned}

Now we have a map

q:Σ ts *A× YY 2f *A× YY 2f *A;q:\Sigma_t s^* A \times_Y Y ^{\mathbf{2}} \to f^* A \times_Y Y ^{\mathbf{2}} \to f^*A;

the first map is a Beck-Chevalley morphism, and the second comes from the fact that f *A is an opfibration. We claim that q is actually a codescent object of the above diagram, and in fact it is a split codescent object. Recall from

  • Claudio Hermida, Descent on 2-fibrations and strongly 2-regular 2-categories.

that a split codescent object is a diagram

Descent object R π 01 π 12 π 02 μ S c d τ T q σ Q

with isomorphisms cτσq, π 12μτc, π 02μτd, 1π 01μ, 1dτ, and 1qσ satisfying certain axioms. In this situation, q is necessarily the codescent object of the diagram composed of d,c,π 01,π 02,π 12.

Tracing through the definitions above, we see that our T is the limit of the span AXY weighted by

(0)(012)(12).(0) \;\to\; (0\to 1\to 2) \;\leftarrow\; (1\to 2).

Likewise, S is the limit of AXY weighted by

(0)(0 1 2 3 4)(234)(0) \;\to\; \left(\array{0&\to&1\\ \downarrow && \downarrow \\ 2 &\to & 3& \to & 4}\right) \;\leftarrow\; (2 \to 3\to 4)

and R is the limit by an evident more complicated weight. Finally, f *A itself is the limit of the same span weighted by

()()().(\cdot) \;\to\; (\cdot) \;\leftarrow\; (\cdot).

The morphism q is induced by the action of X on the opfibration A, applied twice. The morphism d is induced by the inclusion of the second arrow in a composable pair and the action of X on A applied once. The morphism c is induced by the composition of a composable pair. We define the splitting σ,τ,μ in terms of the limit weights by adding in identity morphisms in appropriate places. It is then straightforward to check that the axioms for a split codescent object are satisfied.

Corollary

If K has exponentials, then f *:DOpf(X)DOpf(Y) and f *:DFib(X)DFib(Y) have right adjoints Ran f for any f.

Proof

Right adjoints preserve discrete objects.

This completes the proof that exponentiability of fibrations and opfibrations implies all the other notable exponentiability properties we might want to require of a 2-category.

Enrichment

One further observation:

Proposition

If K has exponentials, then each category DFib(X) is enriched over the cartesian closed category disc(K).

Proof

We define the hom-object DFib(X)(A,B) to be Ran XB A, where B A is the exponential in DFib(X) and

Ran X:DFib(X)DFib(1)disc(K)Ran_X:DFib(X) \to DFib(1)\simeq disc(K)

is the right adjoint to X *. A composition map is obtained by adjointness in the usual way. We also have

K(1,DFib(X)(A,B)) K(1,Ran XB A) DFib(X)(1 X,B A) DFib(X)(A,B)\begin{aligned} K(1,\mathbf{DFib(X)}(A,B)) \cong& K(1, Ran_X B^A)\\ \cong& DFib(X)(1_X, B^A)\\ \cong& DFib(X)(A,B) \end{aligned}

so the underlying ordinary category of this disc(K)-enriched category is the ordinary category DFib(X).

In fact, more is true: I believe DFib(X) can be made into a locally internal category, or equivalently a locally small fibration, over disc(K). Its fiber over a discrete object ZK is the category DFib(X×Z). This follows by localizing the previous lemma in the slices Fib(Z)K/Z; one also has to check that a suitable Beck-Chevalley condition is satisfied by the right adjoints Ran f.

Revised on February 27, 2009 14:59:16 by Jacques Distler? (128.83.114.63)