6.29 Limits and colimits of sheaves
Let X be a topological space. Let \mathcal{I} \to \mathop{\mathit{Sh}}\nolimits (X), i \mapsto \mathcal{F}_ i be a diagram.
Both \mathop{\mathrm{lim}}\nolimits _ i \mathcal{F}_ i and \mathop{\mathrm{colim}}\nolimits _ i \mathcal{F}_ i exist.
The inclusion functor i : \mathop{\mathit{Sh}}\nolimits (X) \to \textit{PSh}(X) commutes with limits. In other words, we may compute the limit in the category of sheaves as the limit in the category of presheaves. In particular, for any open U \subset X we have
(\mathop{\mathrm{lim}}\nolimits _ i \mathcal{F}_ i)(U) = \mathop{\mathrm{lim}}\nolimits _ i \mathcal{F}_ i(U).
The inclusion functor i : \mathop{\mathit{Sh}}\nolimits (X) \to \textit{PSh}(X) does not commute with colimits in general (not even with finite colimits – think surjections). The colimit is computed as the sheafification of the colimit in the category of presheaves:
\mathop{\mathrm{colim}}\nolimits _ i \mathcal{F}_ i = \Big(U \mapsto \mathop{\mathrm{colim}}\nolimits _ i \mathcal{F}_ i(U)\Big)^\# .
Let x \in X be a point. In general the stalk of \mathop{\mathrm{lim}}\nolimits _ i \mathcal{F}_ i at x is not equal to the limit of the stalks. But if the index category is finite then it is the case. In other words, the stalk functor is left exact.
Let x \in X. We always have
(\mathop{\mathrm{colim}}\nolimits _ i \mathcal{F}_ i)_ x = \mathop{\mathrm{colim}}\nolimits _ i \mathcal{F}_{i, x}.
The sheafification functor {}^\# : \textit{PSh}(X) \to \mathop{\mathit{Sh}}\nolimits (X) commutes with all colimits, and with finite limits. But it does not commute with all limits.
The proofs are all easy. Here is an example of what is true for directed colimits of sheaves.
Lemma 6.29.1. Let X be a topological space. Let I be a directed set. Let (\mathcal{F}_ i, \varphi _{ii'}) be a system of sheaves of sets over I, see Categories, Section 4.21. Let U \subset X be an open subset. Consider the canonical map
\Psi : \mathop{\mathrm{colim}}\nolimits _ i \mathcal{F}_ i(U) \longrightarrow \left(\mathop{\mathrm{colim}}\nolimits _ i \mathcal{F}_ i\right)(U)
If all the transition maps are injective then \Psi is injective for any open U.
If U is quasi-compact, then \Psi is injective.
If U is quasi-compact and all the transition maps are injective then \Psi is an isomorphism.
If U has a cofinal system of open coverings \mathcal{U} : U = \bigcup _{j\in J} U_ j with J finite and U_ j \cap U_{j'} quasi-compact for all j, j' \in J, then \Psi is bijective.
Proof.
Assume all the transition maps are injective. In this case the presheaf \mathcal{F}' : V \mapsto \mathop{\mathrm{colim}}\nolimits _ i \mathcal{F}_ i(V) is separated (see Definition 6.11.2). By the discussion above we have (\mathcal{F}')^\# = \mathop{\mathrm{colim}}\nolimits _ i \mathcal{F}_ i. By Lemma 6.17.5 we see that \mathcal{F}' \to (\mathcal{F}')^\# is injective. This proves (1).
Assume U is quasi-compact. Suppose that s \in \mathcal{F}_ i(U) and s' \in \mathcal{F}_{i'}(U) give rise to elements on the left hand side which have the same image under \Psi . Since U is quasi-compact this means there exists a finite open covering U = \bigcup _{j = 1, \ldots , m} U_ j and for each j an index i_ j \in I, i_ j \geq i, i_ j \geq i' such that \varphi _{ii_ j}(s) and \varphi _{i'i_ j}(s') have the same restriction to U_ j. Let i''\in I be \geq than all of the i_ j. We conclude that \varphi _{ii''}(s) and \varphi _{i'i''}(s') agree on the opens U_ j for all j and hence that \varphi _{ii''}(s) = \varphi _{i'i''}(s'). This proves (2).
Assume U is quasi-compact and all transition maps injective. Let s be an element of the target of \Psi . Since U is quasi-compact there exists a finite open covering U = \bigcup _{j = 1, \ldots , m} U_ j, for each j an index i_ j \in I and s_ j \in \mathcal{F}_{i_ j}(U_ j) such that s|_{U_ j} comes from s_ j for all j. Pick i \in I which is \geq than all of the i_ j. By (1) the sections \varphi _{i_ ji}(s_ j) agree over the overlaps U_ j \cap U_{j'}. Hence they glue to a section s' \in \mathcal{F}_ i(U) which maps to s under \Psi . This proves (3).
Assume the hypothesis of (4). In particular we see that U is quasi-compact and hence by (2) we have injectivity of \Psi . Let s be an element of the target of \Psi . By assumption there exists a finite open covering U = \bigcup _{j = 1, \ldots , m} U_ j, with U_ j \cap U_{j'} quasi-compact for all j, j' \in J and for each j an index i_ j \in I and s_ j \in \mathcal{F}_{i_ j}(U_ j) such that s|_{U_ j} is the image of s_ j for all j. Since U_ j \cap U_{j'} is quasi-compact we can apply (2) and we see that there exists an i_{jj'} \in I, i_{jj'} \geq i_ j, i_{jj'} \geq i_{j'} such that \varphi _{i_ ji_{jj'}}(s_ j) and \varphi _{i_{j'}i_{jj'}}(s_{j'}) agree over U_ j \cap U_{j'}. Choose an index i \in I which is bigger or equal than all the i_{jj'}. Then we see that the sections \varphi _{i_ ji}(s_ j) of \mathcal{F}_ i glue to a section of \mathcal{F}_ i over U. This section is mapped to the element s as desired.
\square
Example 6.29.2. Let X = \{ s_1, s_2, \xi _1, \xi _2, \xi _3, \ldots \} as a set. Declare a subset U \subset X to be open if s_1 \in U or s_2 \in U implies U contains all of the \xi _ i. Let U_ n = \{ \xi _ n, \xi _{n + 1}, \ldots \} , and let j_ n : U_ n \to X be the inclusion map. Set \mathcal{F}_ n = j_{n, *}\underline{\mathbf{Z}}. There are transition maps \mathcal{F}_ n \to \mathcal{F}_{n + 1}. Let \mathcal{F} = \mathop{\mathrm{colim}}\nolimits \mathcal{F}_ n. Note that \mathcal{F}_{n, \xi _ m} = 0 if m < n because \{ \xi _ m\} is an open subset of X which misses U_ n. Hence we see that \mathcal{F}_{\xi _ n} = 0 for all n. On the other hand the stalk \mathcal{F}_{s_ i}, i = 1, 2 is the colimit
M = \mathop{\mathrm{colim}}\nolimits _ n \prod \nolimits _{m \geq n} \mathbf{Z}
which is not zero. We conclude that the sheaf \mathcal{F} is the direct sum of the skyscraper sheaves with value M at the closed points s_1 and s_2. Hence \Gamma (X, \mathcal{F}) = M \oplus M. On the other hand, the reader can verify that \mathop{\mathrm{colim}}\nolimits _ n \Gamma (X, \mathcal{F}_ n) = M. Hence some condition is necessary in part (4) of Lemma 6.29.1 above.
There is a version of the previous lemma dealing with sheaves on a diagram of spectral spaces. To state it we introduce some notation. Let \mathcal{I} be a cofiltered index category. Let i \mapsto X_ i be a diagram of spectral spaces over \mathcal{I} such that for a : j \to i in \mathcal{I} the corresponding map f_ a : X_ j \to X_ i is spectral. Set X = \mathop{\mathrm{lim}}\nolimits X_ i and denote p_ i : X \to X_ i the projection.
Lemma 6.29.3. In the situation described above, let i \in \mathop{\mathrm{Ob}}\nolimits (\mathcal{I}) and let \mathcal{G} be a sheaf on X_ i. For U_ i \subset X_ i quasi-compact open we have
p_ i^{-1}\mathcal{G}(p_ i^{-1}(U_ i)) = \mathop{\mathrm{colim}}\nolimits _{a : j \to i} f_ a^{-1}\mathcal{G}(f_ a^{-1}(U_ i))
Proof.
Let us prove the canonical map \mathop{\mathrm{colim}}\nolimits _{a : j \to i} f_ a^{-1}\mathcal{G}(f_ a^{-1}(U_ i)) \to p_ i^{-1}\mathcal{G}(p_ i^{-1}(U_ i)) is injective. Let s, s' be sections of f_ a^{-1}\mathcal{G} over f_ a^{-1}(U_ i) for some a : j \to i. For b : k \to j let Z_ k \subset f_{a \circ b}^{-1}(U_ i) be the closed subset of points x such that the image of s and s' in the stalk (f_{a \circ b}^{-1}\mathcal{G})_ x are different. If Z_ k is nonempty for all b : k \to j, then by Topology, Lemma 5.24.2 we see that \mathop{\mathrm{lim}}\nolimits _{b : k \to j} Z_ k is nonempty too. Then for x \in \mathop{\mathrm{lim}}\nolimits _{b : k \to j} Z_ k \subset X (observe that \mathcal{I}/j \to \mathcal{I} is initial) we see that the image of s and s' in the stalk of p_ i^{-1}\mathcal{G} at x are different too since (p_ i^{-1}\mathcal{G})_ x = (f_{b \circ a}^{-1}\mathcal{G})_{p_ k(x)} for all b : k \to j as above. Thus if the images of s and s' in p_ i^{-1}\mathcal{G}(p_ i^{-1}(U_ i)) are the same, then Z_ k is empty for some b : k \to j. This proves injectivity.
Surjectivity. Let s be a section of p_ i^{-1}\mathcal{G} over p_ i^{-1}(U_ i). By Topology, Lemma 5.24.5 the set p_ i^{-1}(U_ i) is a quasi-compact open of the spectral space X. By construction of the pullback sheaf, we can find an open covering p_ i^{-1}(U_ i) = \bigcup _{l \in L} W_ l, opens V_{l, i} \subset X_ i, sections s_{l, i} \in \mathcal{G}(V_{l, i}) such that p_ i(W_ l) \subset V_{l, i} and p_ i^{-1}s_{l, i}|_{W_ l} = s|_{W_ l}. Because X and X_ i are spectral and p_ i^{-1}(U_ i) is quasi-compact open, we may assume L is finite and W_ l and V_{l, i} quasi-compact open for all l. Then we can apply Topology, Lemma 5.24.6 to find a : j \to i and open covering f_ a^{-1}(U_ i) = \bigcup _{l \in L} W_{l, j} by quasi-compact opens whose pullback to X is the covering p_ i^{-1}(U_ i) = \bigcup _{l \in L} W_ l and such that moreover W_{l, j} \subset f_ a^{-1}(V_{l, i}). Write s_{l, j} the restriction of the pullback of s_{l, i} by f_ a to W_{l, j}. Then we see that s_{l, j} and s_{l', j} restrict to elements of (f_ a^{-1}\mathcal{G})(W_{l, j} \cap W_{l', j}) which pullback to the same element (p_ i^{-1}\mathcal{G})(W_ l \cap W_{l'}), namely, the restriction of s. Hence by injectivity, we can find b : k \to j such that the sections f_ b^{-1}s_{l, j} glue to a section over f_{a \circ b}^{-1}(U_ i) as desired.
\square
Next, in addition to the cofiltered system X_ i of spectral spaces, assume given
a sheaf \mathcal{F}_ i on X_ i for all i \in \mathop{\mathrm{Ob}}\nolimits (\mathcal{I}),
for a : j \to i an f_ a-map \varphi _ a : \mathcal{F}_ i \to \mathcal{F}_ j
such that \varphi _ c = \varphi _ b \circ \varphi _ a whenever c = a \circ b. Set \mathcal{F} = \mathop{\mathrm{colim}}\nolimits p_ i^{-1}\mathcal{F}_ i on X.
Lemma 6.29.4. In the situation described above, let i \in \mathop{\mathrm{Ob}}\nolimits (\mathcal{I}) and let U_ i \subset X_ i be a quasi-compact open. Then
\mathop{\mathrm{colim}}\nolimits _{a : j \to i} \mathcal{F}_ j(f_ a^{-1}(U_ i)) = \mathcal{F}(p_ i^{-1}(U_ i))
Proof.
Recall that p_ i^{-1}(U_ i) is a quasi-compact open of the spectral space X, see Topology, Lemma 5.24.5. Hence Lemma 6.29.1 applies and we have
\mathcal{F}(p_ i^{-1}(U_ i)) = \mathop{\mathrm{colim}}\nolimits _{a : j \to i} p_ j^{-1}\mathcal{F}_ j(p_ i^{-1}(U_ i)).
A formal argument shows that
\mathop{\mathrm{colim}}\nolimits _{a : j \to i} \mathcal{F}_ j(f_ a^{-1}(U_ i)) = \mathop{\mathrm{colim}}\nolimits _{a : j \to i} \mathop{\mathrm{colim}}\nolimits _{b : k \to j} f_ b^{-1}\mathcal{F}_ j(f_{a \circ b}^{-1}(U_ i))
Thus it suffices to show that
p_ j^{-1}\mathcal{F}_ j(p_ i^{-1}(U_ i)) = \mathop{\mathrm{colim}}\nolimits _{b : k \to j} f_ b^{-1}\mathcal{F}_ j(f_{a \circ b}^{-1}(U_ i))
This is Lemma 6.29.3 applied to \mathcal{F}_ j and the quasi-compact open f_ a^{-1}(U_ i).
\square
Comments (2)
Comment #8679 by Peter Fleischmann on
Comment #9387 by Stacks project on