Michael Shulman
duality involution

An important structure possessed by the 2-category Cat is the duality involution () op:Cat coCat. Here we consider what the important properties of this involution are that should be generalized to other 2-categories.


The most obvious property of () op is that it is coherently self-inverse. To state this formally, let J be the walking isomorphism (01), considered as a 3-category with only identity 2-cells and 3-cells. Thus, a (pseudo) functor from J to any 3-category T can be considered an “internal adjoint (bi)equivalence in T.” If T is a 3-category, let C 3op denote the 3-cell dual of T; note that J 3opJ and that () co is a functor 2Cat2Cat 3op. Finally, let τ:JJJ 3op be the evident involution that switches 0 and 1.


A 2-contravariant involution (hereafter merely an involution) on a 2-category K is functor W:J2Cat such that W(0)=K and the square

J τ J 3op W W 2Cat () co 2Cat 3op\array{J & \overset{\tau}{\to} & J^{3op}\\ ^W \downarrow && \downarrow^W\\ 2Cat & \underset{(-)^{co}}{\to} & 2Cat^{3op}}

commutes (on the nose).

We write () o:KK co (or equivalently K coK) for the functor W(01); the rest of the data of W simply says that (() o) oId in a coherent way.


Any (2,1)-category K admits a canonical involution that is the identity on objects and morphisms and sends each 2-cell to its inverse.


Let C be a 2-category equipped with an involution and let K=[C,Cat]. Then K has a canonical induced involution defined by

[C,Cat] co[C co,Cat co][C,Cat co]() op[C,Cat].[C,Cat]^{co} \overset{\simeq}{\to} [C^{co},Cat^{co}] \overset{\simeq}{\to} [C, Cat^{co}] \overset{(-)^{op}\circ -}{\to} [C,Cat].

Conversely, if C is Cauchy-complete, then C op can be identified with the full subcategory of indecomposable projectives in K, and thus an involution on K induces an involution on C.

Thus, in particular, any (2,1)-truncated 2-presheaf 2-topos admits a canonical involution. However, in general a 2-category that admits an involution will admit many involutions.


For any 2-category K, the 2-category Inv(K) of involutions on K, if nonempty, is a torsor for the 3-group Aut(K) of (covariant) automorphisms of K.



In particular, any automorphism of a (2,1)-category C induces an involution on [C,Cat].

Duality Involutions

As observed by WeberYS2T, experience suggests that the most important additional property of the involution () op on Cat is that fibrations over A can be identified with opfibrations over A op, since both are equivalent to functors A opCat. So it is reasonable to consider, together with an involution () o on a 2-category K, a family of equivalences Fib(X)Opf(X o).

The next question is what additional properties should be required of such a family. The following definition is provisional, but I believe whatever eventual definition we settle on should allow the construction of all the data given below.


A duality involution on a 2-category K is an involution () o together with the following data.

  1. Equivalences of 2-categories

    V X:Fib(X)Opf(X o)V_X:Fib(X) \simeq Opf(X^o)

    which are natural in X (that is, V is a natural equivalence between functors K coop2Cat),

  2. An equivalence between V 1:Fib(1)Opf(1 o) and the composite Fib(1)KOpf(1)Opf(1 o) (note 1 o1 since () o is an equivalence).

  3. For any pullback square

    P A B X\array{P & \to & A\\ \downarrow && \downarrow\\ B & \to & X}

    in which all the displayed maps are fibrations, an invertible modification between the two composites

    Fib(P)V POpf(P o)Opf Opf(A o)(P oA o)V A 1Opf Fib(A)(V AP oA) Opf Fib Fib(X)(AX)(V AP oA)Fib Fib(X)(V X 1B o,A)\begin{gathered} Fib(P) \overset{V_P}{\to} Opf(P^o) \simeq Opf_{Opf(A^o)}(P^o\to A^o) \overset{V_A^{-1}}{\to} Opf_{Fib(A)}(V_A P^o \to A)\\ \simeq Opf_{Fib_{Fib(X)}(A\to X)}(V_A P^o\to A) \simeq Fib_{Fib(X)}(V_X^{-1} B^o, A) \end{gathered}


    Fib(P)Fib Fib(B)(PB)V BFib Opf(B o)(V BPB o)Fib Opf Opf(X o)(B oX o)(V BPB o) V X 1Fib Opf Fib(X)(V X 1B oX)(V X 1V BPV X 1B o)Fib Fib(X)(V X 1B o,A).\begin{gathered} Fib(P) \simeq Fib_{Fib(B)}(P\to B) \overset{V_B}{\to} Fib_{Opf(B^o)}(V_B P\to B^o) \simeq Fib_{Opf_{Opf(X^o)}(B^o\to X^o)}(V_B P\to B^o)\\ \overset{V_X^{-1}}{\to} Fib_{Opf_{Fib(X)}(V_X^{-1} B^o\to X)}(V_X^{-1} V_B P\to V_X^{-1} B^o) \simeq Fib_{Fib(X)}(V_X^{-1} B^o, A). \end{gathered}
  4. An invertible modification filling the square

    Fib(X) V X Opf(X o) () o () o Opf(X o) co V X 1 Fib(X) co.\array{Fib(X) & \overset{V_X}{\to} & Opf(X^o)\\ (-)^o\downarrow && \downarrow (-)^o\\ Opf(X^o)^{co}& \underset{V_X^{-1}}{\to} & Fib(X)^{co}.}

    which commutes with the previous two equivalences.

Perhaps there should also be some coherence data and axioms relating these data to each other.

The first two data should be fairly self-explanatory, but the third datum may seem quite complicated. In fact, it requires some thought to see that the two displayed composites are even well-defined. Note first that since P=A× XB, by naturality of V we have V BPB o× X oV XA, and thus V X 1V BPV X 1B o× XA. Also, P oB o× X oA o, so again by naturality V A 1P oV X 1B o× XA. Thus V X 1V BPV A 1P o, and the final two equivalences in each displayed composite are instances of the theorem on iterated fibrations. One way to think about this datum is as sort of an “inclusion-exclusion” property for reversal of arrows.

We call the final datum commutation of opposites. To explain it, observe that because () o is a 2-contravariant involution, we automatically have equivalences Fib K(X)Fib K co(X o)=Opf K(X o) co (these are the vertical maps in the displayed square). In Cat, this equivalence corresponds to composing a functor A opCat with () op:CatCat as well as reinterpreting it as an opfibration. Commutation of opposites then says that this operation commutes with V.

One immediate application of commutation of opposites is to deduce a version of the third datum for a pullback square of opfibrations, by passage along the equivalences Fib(X)Opf(X o) co. We will see others below.


If K is a (2,1)-category, then its canonical involution extends to a canonical duality involution, since Fib(X)K/XOpf(X).


If C is a (2,1)-category, then the canonical involution on K=[C,Cat] extends in a natural way to a duality involution. This does not seem to be the case if C is merely a 2-category equipped with an involution.

The 3-group Aut(K) also acts on the 2-category DualInv(K) of duality involutions on K; the action is free but (seemingly) no longer necessarily transitive. However, I do not know an example of two inequivalent duality involutions having the same underlying involution (as would be necessary for non-transitivity).

The differences between Definition 2 and the definition of WeberYS2T are threefold.

  1. Our definition is less strict; () o is only required to be self-inverse in a 2-categorically non-evil way.
  2. Our definition refers to arbitrary fibrations and opfibrations, rather than merely discrete ones.
  3. Our definition refers only to one-sided fibrations, whereas WeberYS2T requires an equivalence DFib(A×B,C)DFib(A,B o×C).

Of course, since any equivalence of 2-categories preserves discrete objects, our definition implies an equivalence DFib(X)DOpf(X o). More surprisingly, it turns out that the seemingly stronger two-sided version is a consequence of our definition.


If K is equipped with a duality involution, then we have natural equivalences Fib(A×B,C)Fib(A,B o×C).


We first construct an equivalence Opf(A×B)Fib(A,B o). Note that since Fib(1)Opf(1 o) is the identity, naturality implies that V A(A×B)A o×B. Also, since () o is an equivalence, and products in K are the same as products in K co, we have (A×B) oA o×B o. Now we have the composite equivalence (using the theorem on iterated fibrations):

Opf(A×B)Opf Opf(B)(A×BB)V BOpf Fib(B o)(A×B oB o)Fib(A,B o).Opf(A\times B) \simeq Opf_{Opf(B)}(A\times B \to B) \overset{V_B}{\to} Opf_{Fib(B^o)}(A\times B^o \to B^o) \simeq Fib(A,B^o).

Note that there is also another equivalence

Opf(A×B)V A×BFib((A×B) o)Fib Fib(A o)(A o×B oA o)V AFib Opf(A)(A×B oA)Fib(A,B o),Opf(A\times B) \overset{V_{A\times B}}{\to} Fib((A\times B)^o) \simeq Fib_{Fib(A^o)}(A^o\times B^o\to A^o) \overset{V_A}{\to} Fib_{Opf(A)}(A\times B^o\to A) \simeq Fib(A,B^o),

but the second and third data in the definition of a duality involution supply a canonical equivalence between these composites, so it doesn’t matter which we use. Finally, for the general case, we simply apply this twice:

Fib(A×B,C)Opf(A×B×C o)Opf(A×(B o×C) o)Fib(A,B o×C).Fib(A\times B, C) \simeq Opf(A\times B\times C^o) \simeq Opf(A\times (B^o\times C)^o) \simeq Fib(A, B^o\times C).

This gives our desired equivalence.

Note, though, that this proof crucially requires that the duality involution come with an equivalence Fib(X)Opf(X o), and not merely DFib(X)DOpf(X o).

Commutation of opposites, which we have not yet used, is necessary for the following important, and perhaps also surprising, result: duality involutions are preserved by (fibrational) slicing.


Any duality involution on K induces a duality involution on each fibrational slice Fib K(X) and Opf K(X).


We define () o X:Fib(X) coFib(X) to be the composite

Fib(X)() oOpf(X o) coV X 1Fib(X) co.Fib(X) \overset{(-)^o}{\to} Opf(X^o)^{co} \overset{V_{X}^{-1}}{\to} Fib(X)^{co}.

To show that () o X is an involution, we verify

(A o X) o X=V X 1((V X 1(A o)) o)V X 1(V X((A o) o))(A o) oA.(A^{o_X})^{o_X} = V_X^{-1}((V_X^{-1}(A^o))^o) \simeq V_X^{-1}(V_X((A^o)^o)) \simeq (A^o)^o \simeq A.

Here we use commutation of opposites in K to commute () o past V X. We now define

V A X:Fib Fib(X)(A)Opf Fib(X)(A o X)V^X_A: Fib_{Fib(X)}(A) \simeq Opf_{Fib(X)}(A^{o_X})

to be the composite

Fib Fib(X)(A)Fib(A)V AOpf(A o)Opf Opf(X o)(A o)V X 1Opf Fib(X)(V X(A o))=Opf Fib(X)(A o X).Fib_{Fib(X)}(A) \simeq Fib(A) \overset{V_A}{\to} Opf(A^o) \simeq Opf_{Opf(X^o)}(A^o) \overset{V_X^{-1}}{\to} Opf_{Fib(X)}(V_X(A^o)) = Opf_{Fib(X)}(A^{o_X}).

It is easy to see that V 1 X XV X 1V X1, and to construct the pullback-commutation equivalence. Finally, we construct commutation of opposites in Fib(X) as the composite

(V A X(B)) o XV X 1((V X 1(V A(B))) o)V X 1(V X((V A(B)) o))(V A(B)) oV A 1(B o)V A 1(V X(V X 1(B o)))=(V A X) 1(B o X).(V^X_A(B))^{o_X} \simeq V_X^{-1}((V_X^{-1}(V_A(B)))^o) \simeq V_X^{-1}(V_X((V_A(B))^o)) \simeq (V_A(B))^o\simeq V_A^{-1}(B^o)\simeq V_A^{-1}(V_X(V_X^{-1}(B^o))) = (V^X_A)^{-1}(B^{o_X}).

The case of Opf(X) is dual.

Fixing groupoids

We saw above that in general, a 2-category admitting an involution or duality involution will admit many. One way to get rid of some of these spurious involutions is to require that () o fix groupoidal objects, as is clearly the case for the involution of Cat and the “canonical” involutions of (2,1)-truncated 2-presheaf 2-toposes.

Note that any involution takes groupoidal objects to groupoidal objects.


An involution () o on K fixes groupoids it it restricts to the canonical involution on gpd(K), i.e. if X oX coherently whenever X is groupoidal.


If K is a regular n-category having enough groupoids, then it admits, up to equivalence, at most one involution that fixes groupoids. If K is also n-exact, then it admits exactly one such involution.


Suppose that () o is an involution on K that fixes groupoids. Then for any A and any groupoidal X, we have

K(X,A o)K co(X o,A)K co(X,A)K(X,A) op.K(X,A^o) \cong K^{co}(X^o,A) \cong K^{co}(X,A) \cong K(X,A)^{op}.

Now given B as well, with an eso p:XB where X is groupoidal, any morphism BA o is determined by a map XA o with an action by ker(p). But ker(p)=(p/p) is also groupoidal, so morphisms BA o are completely determined by morphisms to A out of groupoidal objects. Thus, by the Yoneda lemma, any two values of A o must be canonically isomorphic.

Now suppose K is exact, and let p:XA be an eso where X is groupoidal. Then its kernel (p/p) is also groupoidal, and so ker(p) is a homwise-discrete category in gpd(K). Thus we can consider its opposite, which is also a homwise-discrete category in gpd(K). In general, the opposite of a 2-congruence in a 2-category will not be a 2-congruence, since the condition of being a two-sided fibration is not preserved. But in a (2,1)-category such as gpd(K), this extra condition is automatic, so this opposite of is also a 2-congruence in gpd(K) (and, in fact, an n-congruence when K is an n-category). Thus, since K is n-exact, this opposite has a quotient, which we call A o. It is straightforward to verify that this defines an involution on K that fixes groupoids.

Of course, the only interesting values of n in the above lemma are 2 and (1,2), since any groupoid-fixing involution on a (2,1)-category is equivalent to the canonical one. Of particular note is that any (2,1)-truncated Grothendieck 2-topos admits a unique groupoid-fixing involution.

In fact, this groupoid-fixing involution should also extend to a duality involution. For when K is n-exact, the functor Opf:K op2Cat ought to be a 3-sheaf? (a 2-stack). This means that if p:XA is an eso with X groupoidal, then Opf(A) can be reconstructed from Opf(X) and Opf(p/p) with descent data. Similarly, Fib:K coop2Cat should also be a 3-sheaf, and so Fib(A) can be recovered from Fib(X) and Fib(p/p) with “op-descent data.” But since X is groupoidal, Opf(X)Fib(X), and likewise for (p/p). Therefore, descent data for ker(p) over Opf(X), which determines an object of Opf(A), is the same as op-descent data for the opposite of ker(p) over Fib(X), which determines an object of Fib(A o). This gives the equivalence Opf(A)Fib(A o).

Alternately, there is also a stronger theorem that if K is 2-exact with enough groupoids, then it is equivalent to a particular 2-category of internal categories and anafunctors in the (2,1)-category gpd(K) (more precisely, it is the 2-exact completion of gpd(K)). And it is fairly easy to see, by its construction, that this 2-category has a duality involution, just as the 2-category of internal categories in any 1-category does.

Modulo the action of automorphisms, these are the only examples of duality involutions that I know. I do not know an example of a duality involution on a 2-category that does not have enough groupoids.

Fixing groupoids locally

Now, since a duality involution on K extends to a duality involution on Fib K(X) and Opf K(X), it is natural to strengthen the notion of groupoid-fixing so that it carries over to fibrational slices. We write GFib(X)=gpd(Fib(X)) and similarly GOpf(X).


A duality involution () o on K fixes groupoids locally if we have a equivalence modification between V X:GFib(X)GOpf(X o) and the composite GFib(X)() oGOpf(X o) coGOpf(X o).

The latter equivalence GOpf(X o) coGOpf(X o) is the canonical involution on the (2,1)-category GOpf(X o). Taking X=1, together with the second condition in the definition of a duality involution, this implies that () o fixes groupoids in the previous sense. Moreover, we have:


If () o is a duality involution that fixes groupoids locally, then the induced duality involutions on Fib(X) and Opf(X) also fix groupoids locally.


For a fibration AX, on Fib Fib(X)(A) we have () o X=V X 1() o and V A X=V X 1V A, so the equivalence between () o and V A in K induces one between () o X and V A X in Fib(X).

Now, suppose that () o is a duality involution that fixes groupoids locally. Then for any A and B, the pullback-commutation datum shows that the composite equivalence

Fib(A×B o)Fib(B,A)Opf(A o×B)Opf((A×B o) o)Fib(A\times B^o) \simeq Fib(B,A) \simeq Opf(A^o\times B)\simeq Opf((A\times B^o)^o)

derived from Theorem 1 is equivalent to V A×B o. But if we restrict to groupoidal fibrations, then

V A×B o:GFib(A×B o)GOpf(A o×B)V_{A\times B^o}:GFib(A\times B^o) \simeq GOpf(A^o\times B)

is equivalent to () o. Together with commutation of opposites, this implies that the canonical equivalence

GFib(B,A)GFib(A o,B o)GFib(B,A)\simeq GFib(A^o,B^o)

is, up to equivalence, simply given by () o. In particular, since () o preserves powers, the canonical equivalence

DFib(A,A)DFib(A o,A o)DFib(A,A)\simeq DFib(A^o,A^o)

takes A 2 to (A o) 2. In the 2-internal logic, this corresponds to the statement that ”hom A(x,y)hom A o(y,x),” which is certainly an expected part of the behavior of opposite categories. Of course, for this it suffices that () o fix discretes locally, which has the evident definition.

Further axioms

Another natural requirement is that when X is groupoidal, V X should be equivalent to the composite

Fib(X)K/XOpf(X)Opf(X o).Fib(X)\simeq K/X \simeq Opf(X)\simeq Opf(X^o).

Note that this includes the second datum in the definition of a duality involution as a special case, since 1 is groupoidal. There seems no reason for this condition to be stable under slicing, but it could be stabilized.

Revised on June 12, 2012 11:10:00 by Andrew Stacey? (