7 Regular Categories

Definition 7.1 (Image, cover). We say a category C has images if, for every AfB in C, there exists a least m:BB in Sub(B) through which f factors. We call m the image of f, and we say f is a cover if its image is 1B.

We write AfB to indicate that f is a cover.

Lemma 7.2. Any strong epimorphism is a cover. The converse holds if C has equalisers and pullbacks.

Proof. If f is strong epic, applying the definition to commutative squares of the form

           ′
 A       B

gfm1BB       B
shows that f is a cover.

For the converse, a cover AfB is epic since it can’t factor through the equaliser of any BhgC with gh. To verify the other condition, suppose given

 A       C

gfmhB       D
then the pullback of m along h is monic by Lemma 4.15, and f factors through it, so it’s an isomorphism. So we get BC by composing with the top edge of the pullback square.

Here, if C has images, image facorisation defines a functor [2,C][3,C]: given

  A      B


fghk C      D
if we form the image factorisations
A      I       B

C      J       D
we get a unique IJ making both squares commute.

Definition 7.3 (Regular category). We say C is regular if it has finite limits and images, and image factorisations are stable under pullback, i.e. if the left hand square above is a pullback then so are both right hand squares. (This is equivalent to saying that covers are stable under pullback).

Example 7.4.

Proposition 7.5. Assuming that:

Then covers coincide with regular epimorphisms.

Proof.

By a relation AB in a category C with finite products, we mean an isomorphism class of subobjects RA×B.

If C has images, we define the composite of ARBSC by forming the pullback

  P      S       C

  R      B


qpdcba A
forming the image of (ap,dq):PA×C.

This is well-defined up to isomorphism and has the A(1A,1A)A×A as 2-sided identities.

Lemma 7.6. Composition of relations in C is associative if and only if C is regular.

Proof.

We write Rel(C) for the category whose objects are those of C and whose morphisms are relations. Note that Rel(Set) is just Rel as defined in Example 1.3(e).

We have a faithful functor CRel(C) which is the identity on objects and sends AfB to A(1,f)A×B (for faithfulness, see Exercise 4.22(i)). We write f for (1A,f).

Note that there’s an isomporphism Rel(C)Rel(Cop) which is the identity on objects and sends R(a,b)A×B to R(b,a)B×A; we denote this by R, and write f for (f).

Also, Rel(C) is enriched over Poset (provided Rel(C) is locally small, i.e. C is well-powered), i.e. each Rel(C)(A,B) has a partial order which is preserved by composition.

We say ARB is left adjoint to BSA if 1ASR and RS1B.

Proposition 7.7. ARB is a left adjoint in Rel(C) if and only if it is of the form f.

Proof.

P. Freyd developed a theory of allegories which have the structure of categories of relations and axiomatised those allegories A for which the subcategory Ala is regular.

In a regular category C, we say a relation R:AA is reflexive if 𝟙AR, symmetric if R=R, and transitive if RRR. R is an equivalence relation if it has all three properties. For any AfB in C, the kernel-pair R(a,b)A×A of f is an equivalence relation. We say an equivalence relation R is effective if it occurs as a kernel-pair, and C is effective regular if all equivalence relations are effective.

tfAbGp is regular but not effective regular: {(m,n)×|mn(mod2)} is a non-effective equivalence relation on .

Note that an equivalence relation is idempotent in Rel(C), and if A is an allegory and E is a class of symmetric idempotents in A then A[Eˇ] (as defined in Exercise 1.18) is an allegory; and if A is Rel(C) for a regular category C, then:

Proposition 7.8. Assuming that:

Then Ceff=(Rel(C)[Eˇ])la is effective regular, and the embedding Rel(C)Rel(C)[Eˇ] restricts to a full and faithful regular functor CCeff which is universal among regular functors CD where D is effective regular.

Note that if C is effective regular, its equivalence relations are split idempotents in Rel(C): if ARA is the kernel-pair of AfB then it splits as ff (as we saw for C=Set in Exercise 1.19).

Definition 7.9 (Topos). A topos is a regular category E for which the embedding ERel(E) sending f to f has a right adjoint. We write the effect of the right adjoint on objects by APA, and the unit APA as {}A, and the counit PAA as APA×A.

In Set, PA is the power-set of A, the unit is the mapping a{a} of Example 1.7(c), and A={(A,a)|aA}PA×A.

Note that (isomorphism classes of) subobjects of A are in bijection with morphisms 1PA. C. J. Mikkelses showed that any topos has finite colimits; we’ll give Bob Paré’s proof, which is much simpler.

Proposition 7.10. Assuming that:

Then there exists a monadic functor EopE. In particular, Eop has finite colimits and if E has limits of shape J then it also has colimits of shape Jop.

Proof. We make the assignment APA into a functor P:EE and a functor P:EopE: given f:AB, Pf:PAPB corresponds to the image of APA×A1×fPA×B, and Pf corresponds to the pullback of

                 ∃B

1×f P B × A      P B × B
Given CgPA corresponding to RC×A, (Pf)g corresponds to the image of RC×A1×fC×B and similarly given SD×B, composing with Pf corresponds to pulling back along D×A1×fD×B.

Given a pullback square

  D      A


hkfg B      C
in E,
  P D      P A


PPPPk∗f∗Phg B      P C
commutes, since both ways correspond to the image of the left vertical composite in
E            ∃A


P A × D      P A × A


P A × B      P A × C
where both squares are pullbacks.

Now, as in Example 5.14(d), we have that if EeAgfB is a coreflexive in E, then

PPPPP∗∗∗gePfge B       PA       PE
is a split coequaliser coequaliser in E. Also, P is self-adjoint on the right, and it reflects isomorphisms by Exercise 7.17(v). The second assertion follows from Proposition 5.8(i).

Definition 7.11 (Support, totally supported, capital).

  • (a)
    By the support of an object A in a regular category, we mean the image of A1. We say A is well-supported if A1 is a cover.
  • (b)
    We say a regular category C is totally supported if every object is well-supported. We say C is almost totally supported if every object is either well-supported or a strict initial object, where we cann an object 0 strict if every A0 is an isomorphism. (Given finite limits, a strict object is initial since for any A there exists 0π10×Aπ2A, and the equaliser of any pair 0A is a).
  • (c)
    We say a regular category C is capital if its terminal object 1 is a detector, i.e. C(1,) reflects isomorphisms.

Example. Gp and AbGp are totally-supported since their terminal objects are initial. Set is almost totally-supported and capital. Note that capital implies almost totally-supported since if A isn’t well-supported there are no morphisms 1A.

A representable functor C(A,) always preserves limits, so it’s a regular functor if and only if A is cover-projective (c.f. Definition 2.10).

Lemma 7.12. Assuming that:

Then 1 is cover-projective.

Proof. Since covers are stable under pullback, we need to show that every A1 is split epic. If A1, nothing to prove. If not, the projections A×AA aren’t equal (since their coequaliser is A1, by Proposition 7.5). So there exists 1A×A not factoring through their equaliser, so there exists 1A×AA.

If C is regular, the full subcategory Cws of well-supported objects is closed under finite products since

A × B     A


B         1
is a pullback, and under pullbacks of covers since if AB then A and B have the same support.

We write Ctv for the category obtained from Cws by adjoining a strict initial object 0: this is regular and almost totally-supported and the functor CCtv sending all non-well-supported objects to 0 is regular (c.f. Exercise 5.19).

Lemma 7.13. Assuming that:

Then there exists an isomorphism-reflecting regular functor I:CC, where C is also small and almost totally-supported, such that for every well-supported AobC there exists a morphism 1IA in C not factoring through I(m) for any proper subobject m:AA in C.

Proof. Recall from Exercise 7.17: C regular implies CA regular for any A, and for any f:AB in C pullback along f defines a regular functor f:CBCA, which has a left adjoint Σf:CACB sending g:CA to fg. And f reflects isomorphisms if and only if f is a cover.

We’ll define C as (C^)tv where C^ is easier to describe.

To satisfy the desired conclusion for a single well-supported object A, enough to take (!)A:CC1CA, since (!A)A=(A×Aπ2A) acquires a point Δ:(A1A)(A×AA) not factoring through (A×AA) for any proper AA.

More generally, for any finite list A1,,An of well-supported objects, we can take Ci=1nAi.

We define a base to be a finite list A=(A1,,An) of distinct well-supported objecs of C. We preorder the set B of bases by AB if B contains all the members of A. We write A for the product i=1nAi and if AB we write πB,A for the product projection BA. This makes AA into a functor BopC.

Hence the assignment ACA, πB,AπB,A is ‘almost’ a functor BCat.

We now define C^: its objects are pairs (B,f) where B is a base and f:AB is an object of CB. Morphisms (B,f)(B,f) are represented by pairs (C,g) where C is a base containing B and B and g:πfπf in CC, subject to the relation which identifies (C,g) with (C,g) if CC and the pullback of g to CC is isomorphic to g.

Clearly, each CB sits inside C^ as a non-full subcategory; so in particular CC[] is a subcategory of C^, C^ is regular, and the inclusions CBC^ are isomorphism-reflecting regular functors.

Given a finite diagram in C^, we can choose B such that all edges of the diagram appear as morphisms in CB, and take the limit there, and this is a limit in C^. Similarly for images.

Also, if a morphism f becomes an isomorphism in C^, its inverse must live CB for some B, hence f is an isomorphism CB.

We define C=(C^)tv: the induced functor CC^C is still isomorphism reflecting since C is almost totally-supported.

Lemma 7.14. Assuming that:

Then there exists an isomorphism reflecting regular functor CC^ where C is capital. Hence in particular, there is an isomorphism-reflecting regular functor CSet.

Proof. Consider the sequence

C=C0C1C2,

where each Cn+1 is obtained from Cn by the construction of Lemma 7.13.

We define C^ to be the pseudo-colimit of this sequence: objects are pairs (n,A) where AobCn, and morphisms (n,A)(m,B) are represented by pairs (p,f) where pmax{m,n} and F:IAIB in Cp, modulo the identification of (p,f) with (p,f) if pp and f=If.

The proof that C is regular, and that the embeddings CnC^ are isomorphism-reflecting regular functors, is as in Lemma 7.13.

Given any non-invertible monomorphism AA in C^, it lives in Cn for some n, so there exists 1A in Cn+1 not factoring through AA.

But if AfB isn’t monic in C^, the legs RbaA of its kernel-pair aren’t equal, so there exists 1rR not factoring through their equation, so 1brarA are distinct but have the same composite with f.

So C^(1,) reflects monomorphisms and hence reflects isomorphisms.

Theorem 7.15. Assuming that:

Then there exists a set I and an isomorphism-reflecting regular functor CSetI.

Proof. Let I be a representative set of subobjects of 1 in C, and for each UI consider the composite

C(!U)CU(CU)tv(CU)tv^Set,

where the third factor is the functor of Lemma 7.14 and the fourth is represented by 1.

Given any non-invertible morphism AfB in C, if U is the support of B then (!)Hf remains non-invertible in CU and its codomain is well-supported there, so it remains non-invertible in (CU)tv and hence in Set.

So these functors collectively reflect isomorphisms.

Remark 7.16.

˙