The concept of Cauchy completeness, ordinarily thought of as applying to metric spaces, was vastly generalized by Bill Lawvere in his influential paper Metric spaces, generalized logic, and closed categories. It is now seen by category theorists as a concept of enriched category theory, with close ties to the concept of Morita equivalence in the theory of modules. In category theory one also speaks of idempotent completeness.
The basic idea is that the Cauchy completion of a category is the closure of a category under what are called “absolute limits”, i.e., those limits that are preserved by any functor whatsoever. Equivalently, the Cauchy completion is the closure with respect to absolute colimits. If is small, the Cauchy completion of lies between and its “free cocompletion”, aka presheaf category
preserves small coproducts and coequalizers. All of these concepts generalize straightforwardly to the context of general -enriched categories, where is a complete, cocomplete, symmetric monoidal closed category.
Lawvere defines a point of the Cauchy completion of a small -category to be a -enriched bimodule (in other words, a -functor ) for which there is a bimodule right adjoint to (in the bicategory of enriched bimodules, see profunctor), where is the unit -category. Thus points of the Cauchy completion are certain -enriched presheaves , and together form a -category called the Cauchy completion whose homs are the presheaf homs. It is denoted .
and we say the -category is Cauchy-complete if this embedding is an equivalence. We work through a few examples in the following section.
For a small category write
This is called the Cauchy completion of .
For instance (BorceuxDejean, below theorem 1).
The Cauchy completion satisfies the following properties
This appears for instance as (BorceuxDejean, theorem 1).
is small because is a well-powered category. It contains as a full subcategory because the Yoneda embedding is a full and faithful functor. Every idempotent splits in because it does so in and because the composite of two retractions is a retraction.
A retract of a representale induces an idempotent on and hence by the Yoneda lemma an idempotent on . If is already idempotent complete, this splits and produces a retraction of in and hence of in . Since this is necessarily isomorphic to the original retraction, we find that every retract of the representable is itself representable, therefore in this case.
This appears for instance as (BorceuxDejean, prop. 2).
The following conditions are equivalent for a small category .
is Cauchy complete;
has all small absolute colimits.
Write for the terminal category (single object, single morphism). Let be a small category
The category is equivalent to the functor category out of the point
Its category of presheaves is equivalent to the profunctor category
In these terms the Yoneda embedding is the canonical inclusion
This appears for instance as (BorceuxDejean, prop. 4).
For a small category , the following are equivalent
This appears for instance as (BorceuxDejean, theorem 2).
of its Cauchy completion, def. 1, with the category of essential points of .
We first exhibit a full inclusion .
is fixed by a choice of copresheaf
naturally in , and hence that
A morphism between geometric morphisms is a geometric transformation, which is a natural transformation , hence by the above a natural transformation . By the Yoneda lemma these are in bijection with morphisms in . This gives the full inclusion .
The converse inclusion is now immediate by the same arguments: since the objects in are precisely the tiny objects each of them corresponds to a functor that has a right adjoint. Since this generally also has a left adjoint, it is the inverse image of an essential geometric morphism .
exhibited by the 2-adjunction
We first claim that when working with all categories instead of just the Cauchy-complete categories there is a 2-adjunction
natural in and . Here
the first equivalence is by definition of essential geometric morphism;
the second equivalence follows by observing that limits and colimits in presheaf categories are computed objectwise;
the third equivalence is again the definition of essential geometric morphisms.
are equivalences precisely if is Cauchy-complete. This means that restricted along the adjunction exhibits a coreflective embedding.
We discuss Cauchy completion in -enriched category theory, for a closed symmetric monoidal category with all limits and colimits. The discussion in ordinary category theory above is the special case where Set.
The key to the enriched version is the reformulation of ordinary Cauchy completion in terms of profunctors as discussed above. These have an immediate generalization to enriched category theory, and so one takes this formulation as the definition.
As before, we have
Every small -category is equivalence to the -enriched functor category
where is the -category with a single object and , the tensor unit in .
and the canonical -enriched functor
is the enriched Yoneda embedding.
For a small -enriched category, the Cauchy completion of is the full -subcategory
We consider first the classical case of metric spaces, but as redefined by Lawvere to mean a category enriched in the poset , with tensor product given by addition. So, to say is a Lawvere metric space means that with the set there is a distance function
for all in . (The associativity and identity axioms are here superfluous since is a poset.) A -enriched functor here just means a function from to such that
for all in (again, preservation of composition and of identities is superfluous here), so that -functors are short maps between vector spaces (Lipschitz maps with constant at most ). Finally, a -enriched transformation in this case boils down to an instance of a property: that
for all in . If are valued in , this just means for all .
A point of the Cauchy completion is an -module , i.e., an enriched functor or short map
for which there is an -module on the other side, an enriched functor
that is right adjoint to in the sense of modules. This means there is a unit of the adjunction in the bicategory of modules:
and a counit:
Recall now that in the bicategory of modules is the unit bimodule given by the enriched Yoneda embedding, or in different words, . Recall also that module composition is defined by a coend formula for a tensor product. If one now tracks through the definitions, keeping in mind that we are in the very simple case of enrichment in a poset, the unit of the adjunction boils down to having the property
and the counit boils down to having the property
To better appreciate what these conditions mean, we point out that should be thought of as the distance between and the “ideal point” in the Cauchy completion , and should be thought of as the companion distance . Thus the unit condition above would come down to saying that for every there exists such that
and the counit condition imposes a necessary triangle inequality constraint on the distance functions and , in order that we get an actual Lawvere metric space . If are two points of the Cauchy completion thus defined, then their distance is defined by the usual formula for enriched presheaves:
It should be noted that even under the classical definition (where we impose symmetry , separation for , and finiteness ), this provides an elegant alternative definition of Cauchy completion. In essence, all it is doing is taking the metric closure of the embedding of into the already complete space of short maps:
The presheaf-hom definition of the distance formula for , being manifestly non-symmetric, is not the usual definition of distance in the classical symmetric case. However, if we first symmetrize the distance in :
then we do retrieve the classical formula
In other words, the completion of a symmetric metric space as a general (Lawvere) metric space is not necessarily the same as its completion as a symmetric metric space, but is the symmetrisation of .
The analysis of Cauchy complete Lawvere metric spaces contains some of the seeds of what happens in other enriched category contexts; the case of ordinary small categories, where the enrichment is no longer in a mere poset but in Set, reflects still more of the phenomena generally associated with Cauchy completions.
and the structure of the adjunction is given by unit and counit maps:
As we said in the case of metric spaces, and measure “distances” = homs:
where is shorthand for the module ; this is just an instance of the Yoneda lemma:
Now using the adjunction , there are, for any set , natural bijections
and maps in the bottom line are in bijection with maps . Therefore we have a natural bijection
and this proves .
With these identifications of and , the unit of the adjunction takes the form
The coend above is a quotient of
and hence the unit element is represented by a pair of transformations
for some .
Given that, it is now not hard – in fact it is fairly tautological – to verify that on the basis of the triangular equation of the adjunction which says
Indeed, the claim is that modules in the Cauchy completion are precisely those presheaves on which arise as retracts of representables in , or in other words may be identified with objects of the idempotent-splitting completion of (aka the Karoubi envelope of ). Therefore, in the -enriched case, the Cauchy completion is the idempotent-splitting completion. In particular, representables themselves are points of the Cauchy completion.
Notice that in a finitely complete category (such as or a presheaf category), idempotents split automatically: just take the equalizer of the pair
For that matter, in any finitely cocomplete category, taking the coequalizer of the above pair would also split the idempotent. Indeed, we can say that idempotents split in a category iff all equalizers of such pairs exist, iff all coequalizers of such pairs exist.
Notice that if and are categories, then any functor preserves retracts and therefore splittings of idempotents. Thus, the equalizers above are the sort of limits which are preserved by any functor whatsoever. They are called absolute limits for that reason. For the same reason, the coequalizers above are absolute colimits: they are precisely the colimits preserved by any functor whatsoever.
Pursuing this a bit further: if is any functor, then (because idempotents split in ) there is a unique extension of . Therefore we have an equivalence
and we say that and are Morita equivalent.
This appears as (Rosolini, prop. 2.1).
This appears as (Rosolini, corollary. 2.3).
Moreover, the characterization of Cauchy completion by left adjoint profunctors requires the internal axiom of choice:
In a given ambient context, the following are equivalent:
For instance (BorceuxDejean, prop. 5).
David: Concerning the result that on Set the terminal F-coalgebra is the Cauchy completion of the initial F-algebra, for certain F, I wonder if we have to factor completions through the metric space completion, as Barr does in Terminal coalgebras for endofunctors on sets. Perhaps Adamek’s work on Final Algebras are Ideal Completions of Initial Algebras is more natural.
How many kinds of completion are there for an enriched category? I see some may coincide in certain cases.
If two categories can be Morita equivalent, should this be reflected in the page Morita equivalence?
The notion of Cauchy complete categories was introduced in
Bill Lawvere, Metric spaces, generalized logic and closed categories Rend. Sem. Mat. Fis. Milano, 43:135–166 (1973)
Reprints in Theory and Applications of Categories, No. 1 (2002) pp 1-37 (tac)
Surveys are in
A. Carboni and Ross Street, Order ideal in categories Pacific J. Math., 124:275–288, 1986.
Further references include for instance
R. Walters, Sheaves and Cauchy complete categories , Cahiers Top. Geom. Diff. Cat. 22 no. 3 (1981) 283-286 (numdam)
R. Walters, Sheaves on sites as Cauchy-complete categories, J. Pure Appl. Algebra 24 (1982) 95-102