I was considering the following situation. I have a directed/filtering system of categories $C_i$. I understand how take its direct limit (aka colimit) $C$ in the category of categories. My question is:

If $C_i$ is cocomplete is $C$ also cocomplete?

I think in general the answer is no. For example, taking $C_i = C_0 \times C_0 \cdots \times C_0$ for some fixed category $C_0$ and taking $C_i \to C_{i+1}$ to be the obvious inclusion. I don't think the limiting category $C$ has coproducts.

Can I add (non trivial) assumptions so that $C$ suddenly has colimits? (if so, how do I compute them?)

In the situation I am considering I have a bunch more of assumptions, which might be handy. All the categories $C_i$ are Grothendieck abelian. The functors $C_i \to C_j$ are exact, essentially surjective and have right adjoints. The indexing system has an initial element $0$, so any object in $C_i$ comes from $C_0 \to C_i$. Also, all I really want in $C$ are direct limits (ie filtered colimits), not necessarily arbitrary ones.

  • 2
    $\begingroup$ If the diagram is $\kappa$-filtered, any structure of size less than $\kappa$ (like e.g. $\kappa$-colimits) that is shared by all categories and preserved by the transition functors is inherited by the $\kappa$-filtered colimit. $\endgroup$
    – godelian
    Jan 21, 2017 at 16:00
  • $\begingroup$ @godelian that's great! how do you prove it? I can't see what the recipe is to form colimits in the limiting category. $\endgroup$ Jan 21, 2017 at 20:50
  • 1
    $\begingroup$ It follows from the construction of $\kappa$-filtered colimits of categories. If you have a diagram in the colimit category, you can find representatives of each arrow in the categories involved in the $\kappa$-filtered colimit. Because of $\kappa$-filteredness, you can find a unique category where all representatives lie. Then you just take the colimit in this category and it will be a representative of the colimit in the colimiting category. $\endgroup$
    – godelian
    Jan 21, 2017 at 21:01
  • $\begingroup$ @godelian I'm probably not seeing something obvious. Take the example of a diagram $x_i$ which can be entirely represented in the first category $C_0$. I would like to conclude that $f(colim_i x_i) = colim_i f(x_i)$ where $f\colon C_0 \to C$ is the functor from $C_0$ to the colimit category $C$. But if I try and verify the universal property, ie $C(f(colim_i x_i), y) = lim_i C(x_i, y)$ I get stuck, essentially because limits and colimits do not commute in general. Am I missing something? If so, would you care to write an answer? $\endgroup$ Jan 21, 2017 at 21:17
  • $\begingroup$ More explicitly, say $x,y \in C$ can be represented in the initial category $C_0$, we have $C(x,y) = colim_j C_j(f_j(x), f_j(y))$ where $f_j \colon C_0 \to C_j$ is the transition functor. For the universal property to be satisfied I would have to have $colim_j C_j( f_j colim_i x_i, y) = lim_i colim_j C_j(f_j(x_i), f_j(y))$. Assuming $f_j$ commutes with colimits (which I'd be happy to), this amounts to $colim_j lim_i C_j(f_j(x_i), f_j(y)) = lim_i colim_j C_j(f_j(x_i), f_j(y))$ which doesn't hold in general. $\endgroup$ Jan 21, 2017 at 21:20

1 Answer 1


Let $C_i$ a diagram of categories on a filtred small category $I$, $C:=\varinjlim_{i} C_i$, let write $[A, B]$ the category of funtors (and natural transformation) between categories $A, B$, observe that a colimit of categories is also a 2-colimit (i.e. it is natural respect natural transformations).

If $J$ is a finite category then $\varinjlim_{i} [J, C_i]\cong [J, C]$ (a functor $J \to C $ as a lifting on some $C_i$, and two such lifting have another one with usual coherence). Consider the colimit adjunction $[J, C_i]\leftrightarrows [\ast, C_i]$ then getting the colimit we have the adjoint diagram $[J, C]\leftrightarrows [\ast, C]$.

THere is a easy generalization to more high cardinal..


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.