5
$\begingroup$

$\DeclareMathOperator\Psh{Psh}\DeclareMathOperator\Sh{Sh}\newcommand\copower{\mathrm{copower}}$I was looking through Bodil Biering's thesis On the Logic of Bunched Implications - and its relation to separation logic, and became confused by the following Day convolution argument in Lemma 4.3.1. The statement of the lemma includes the isomorphism in \eqref{462744_1} below.

Consider a site $C$ which is also monoidal, with tensor product $\cdot : C \times C \to C$. We have a geometric embedding $\mathbf{a} \dashv i : \Sh(C) \hookrightarrow \Psh(C)$, so $\mathbf{a}$ is the sheafification functor, and $i$ the inclusion of sheaves to presheaves. Also let $\mathbf{y}: C \to \Psh(C)$ be the Yoneda embedding. The proof of Lemma 4.3.1 has a chain of equations/isomorphisms: \begin{align} \mathbf{a}\left(i\mathbf{a} P \otimes Q\right) &= \mathbf{a} \left( \int^{n,n'}_{\text{Psh}(C)} i \mathbf{a} P(n) \times Q(n') \times \mathbf{y}(n \cdot n')\right) \\ &\cong \int^{n,n'}_{\text{Sh}(C)} \mathbf{a} i \mathbf{a} P(n) \times \mathbf{a}Q(n') \times \mathbf{a}\mathbf{y}(n \cdot n')\tag{1}\label{462744_1}\\ &\quad\vdots\\ &\cong\mathbf{a}(P \otimes Q). \end{align} Biering notes that this isomorphism is a consequence of the property that $\mathbf{a}$ preserves colimits and finite limits. It makes sense that the co-continuity of $\mathbf{a}$ allows it to be brought inside of the coend. I am struggling to understand where is the finite limit (which allows $\mathbf{a}$ to be applied to all factors?).

Here is my incomplete understanding of the coend. The Day convolution of two presheaves $F,G : C^\text{op} \to \mathbf{Set}$ is a coend $\int^{n,n' : C} H\left(\langle n,n' \rangle,\langle n,n' \rangle \right)$, where $H$ is the profunctor $(C \times C)^\text{op} \times (C \times C) \to \Psh(C)$ given by the composition $$ (C \times C)^\text{op} \times (C \times C) \xrightarrow{F \boxtimes G} \mathbf{Set} \times (C \times C) \xrightarrow{\mathbf{y} \circ \cdot} \mathbf{Set} \times \Psh(C) \xrightarrow{\copower} \Psh(C), $$ where $F \boxtimes G$ is the presheaf $(C \times C)^\text{op} \to \mathbf{Set}$ given by $(F \boxtimes G)\langle n , n' \rangle = F(n) \times G(n')$, and $\copower : \mathbf{Set} \times \Psh(C) \to \Psh(C)$ is the coproduct $$ \copower(A,F) = \coprod_{\_ \in A} F. $$ If we apply the functor $\mathbf{a}$ to the result, we are taking the coend of the profunctor: $$ (C \times C)^\text{op} \times (C \times C) \xrightarrow{F \boxtimes G} \mathbf{Set} \times (C \times C) \xrightarrow{\mathbf{y} \circ \cdot} \mathbf{Set} \times \Psh(C) \xrightarrow{\copower} \Psh(C) \xrightarrow{\mathbf{a}} \Sh(C), $$ which (I think), when setting $F = i \mathbf{a} P$ and $G = Q$, corresponds to the "integral" $$ \int^{n,n'}_{\Sh(C)} i\mathbf{a}P(n) \times Q(n') \times \mathbf{a}\mathbf{y}(n\cdot n').\tag{2}\label{462744_2} $$ It seems that the product in coend \eqref{462744_1} is not a product of presheaves (hence not a finite limit in $\Psh(C)$). Instead it is consequence of how the usual copowering over $\mathbf{Set}$ is a generalization of the Cartesian product.

Perhaps there is something wrong with the above paragraph, or perhaps there is an isomorphism from \eqref{462744_1} to \eqref{462744_2}?

$\endgroup$
1
  • 1
    $\begingroup$ TeX note: technically, {align} is meant for a group of equations. If you want analogous functionality for a single equation, then the AMSmath documentation charges you to use {split}, which is meant to be included in an ambient {equation} environment. But MathJax seems incorrectly to ignore \tag in that case, so we seem stuck with an {align}-based solution. Nonetheless, you did have a period after your \tag{1} that seemed to be accidental, so I removed it. (I also used \label+\eqref, as you might as well when tagging!) $\endgroup$
    – LSpice
    Commented Jan 23 at 22:09

1 Answer 1

1
$\begingroup$

$\require{AMScd}$I would say that the fact that $\bf a$ commutes with products is what allows you to apply it to all factors of the coend, but a more precise way to use its properties would be that $${\bf a}((Pn\times Qm)\cdot \hom(-,n\otimes m))\cong ({\bf a}Pn\times {\bf a}Qm)\cdot \hom(-,n\otimes m),$$ even if the copower operation in this specific case coincides with products.

Would you be interested in a different proof?

I'll give you two alternative proofs. Somehow I can't find totally satisfying the argument you're supposed to follow (but the statement is true, no doubt about it!)

Alternative proof 1 is based on cocontinuity of $\bf a$, plus the fact that ${\bf a}(P\times Q)\cong {\bf a}P \otimes Q$ where $P,Q$ are presheaves, $\bf a$ is sheafification, $\otimes$ is Day convolution (induced by a monoidal structure $\_\cdot\_$ on the domain of presheaves). Note that $P\otimes Q$ is a left Kan extension, precisely the left Kan extension of $P\times Q : C\times C \to Set \times Set \overset\times\to Set$ along $\_\cdot\_ : C\times C\to C$. But then, $${\bf a}(P \otimes Q) = {\bf a}(\text{Lan}_{(\_\cdot\_)}(P\times Q))\cong \text{Lan}_{(\_\cdot\_)}{\bf a}(P\times Q)\cong {\bf a}P\otimes Q$$ and now, clearly, also ${\bf a}(i{\bf a}P\otimes Q)$ is isomorphic to ${\bf a}P\otimes Q$ as well.

An alternative proof 2 relies, instead, on recognizing that the isomorphism you want is obtained applying sheafification $\bf a$ to a canonical morphism (this is mentioned in the thesis as well!). But given this, you can characterize completely the morphisms that sheafification sends to isomorphisms: given $f : A\to B$, the arrow ${\bf a}(f)$ is an iso if and only if, for every sheaf $R$, in the diagram $$\begin{CD} A @>q>> R \\ @VfVV @.\\ B \end{CD}$$ there is a unique $B\to R$ closing the triangle. Now, in your case $f = \eta_P\otimes Q : P\otimes Q \to i{\bf a}P\otimes Q$. You can easily prove that given a sheaf $R$, the condition above is satisfied.

$\endgroup$
4
  • $\begingroup$ Thank you for your answer! The part that trips me is the isomorphism in your answer. Especially since, e.g., $P n$ and $\mathbf{a} P n$ can be very different sets, even for subcanonical topologies. In the subcanonical case, the left side and the right side are coproducts of the same (pre)sheaf, but the cardinality of the indexing sets can change. $\endgroup$ Commented Jan 24 at 21:03
  • $\begingroup$ hmm, now I'm doubting my argument. I think I can prove the statement, but in a different way. $\endgroup$
    – fosco
    Commented Jan 25 at 8:36
  • 1
    $\begingroup$ I edited my first answer! $\endgroup$
    – fosco
    Commented Jan 26 at 7:43
  • 1
    $\begingroup$ I am running through the details of proof 2. It seems that the morphism $i\mathbf{a} P \otimes Q \to R$ is obtained by proving the following: if a pair $\alpha,\alpha' \in P n$ satisfies $\eta_P(\alpha) = \eta_P(\alpha')$, then for all $\beta \in Q n'$, $q(\alpha,\beta,\text{id}_{n \cdot n'}) = q(\alpha',\beta,\text{id}_{n\cdot n'})$. Does this not require some condition on the Grothendieck topology, e.g., if $f_i : X_i \to X$ is a cover, then $f_i \otimes 1_Y : X_i \otimes Y \to X \otimes Y$ is also a cover? $\endgroup$ Commented Jan 28 at 22:03