Last time a site was a category equipped with a notion of covering family. A map between sites is not extra structure — it is exactly what is needed to compare sheaves on two sites, and hence to say what pushforward and pullback of sheaves mean.
Let $\mathcal{C}$ and $\mathcal{D}$ be sites. A continuous morphism of sites $f : \mathcal{C} \to \mathcal{D}$ is a functor
going in the reverse direction — exactly as a continuous map of spaces pulls back open sets — such that:
The direction reversal is the whole content of the definition, and it is worth stating as a slogan.
A continuous morphism of sites is a functor that carries covers to covers and commutes with the fiber products in which gluing conditions live. Preserving covers is what lets a sheaf on the source be pulled back to a sheaf on the target; preserving fiber products is what keeps the gluing conditions themselves intact.
Let $f : X \to Y$ be a continuous map of topological spaces and put $u(V) = f^{-1}(V)$ for $V \subset Y$ open. Then $u$ preserves fiber products, because fiber products in $\mathrm{Open}(Y)$ are intersections and preimages commute with intersections:
and $u$ sends covers to covers, because preimages commute with unions. So $f$ induces a continuous morphism of sites $X_{\mathrm{Zar}} \to Y_{\mathrm{Zar}}$, and the abstract definition recovers the familiar pushforward/pullback formalism for sheaves.
For a scheme $X$, the identity functor on the underlying category gives continuous morphisms of sites
(reading the arrows as maps of sites). Each inclusion of topologies goes the other way on underlying categories, and every Zariski cover is an étale cover, every étale cover is an fppf cover. This nesting is the reason a sheaf for a finer topology is automatically a sheaf for a coarser one: a condition imposed on more covering families is a stronger condition.
A topos is a category of sheaves on a site, together with the extra structure — colimits, finite limits, and a way to glue — that makes it behave like a category of sheaves on a space. Different sites can present the same topos; in that case the associated categories of sheaves are equivalent, and the sites are two descriptions of one object. This is the sense in which Grothendieck topologies generalize topological spaces rather than merely decorate them.
In practice one rarely checks the sheaf axiom on a site directly for every covering family. Two reduction principles do almost all the work.
Let $X$ be a scheme and $Y$ an $X$-scheme. Then the functor
is a sheaf on the fppf site $X_{\mathrm{fppf}}$; hence also on $X_{\mathrm{\acute{e}t}}$ and on $X_{\mathrm{Zar}}$. Consequently $\mathbb{G}_m$, $\mu_n$, $\underline{\mathbb{Z}/n\mathbb{Z}}$, and $\mathbb{P}^1$ all give sheaves on every one of these sites.
Maps into $Y$ glue along a cover because they glue on the underlying topological spaces and the gluing is compatible with the structure sheaves; faithful flatness of finite presentation is what upgrades compatibility on overlaps to a genuine map. The point to retain is structural: a functor represented by a scheme is automatically a sheaf, so the supply of sheaves is not a problem. The interesting questions are about the ones that are not representable, such as the cokernels of natural maps (recall $\mathbb{G}_m \xrightarrow{n} \mathbb{G}_m$ from §6 of the previous meeting).
The next observation is the entry point to descent. A quasi-coherent sheaf on $X$ is, by definition, a sheaf for the Zariski topology. It turns out to satisfy the sheaf condition for far more covers than that.
Let $\mathcal{F}$ be a quasi-coherent sheaf on a scheme $X$. Then $\mathcal{F}$ is a sheaf on the fppf site $X_{\mathrm{fppf}}$, and hence also on $X_{\mathrm{\acute{e}t}}$ and $X_{\mathrm{Zar}}$.
Equivalently, the evident functor is fully faithful:
A quasi-coherent sheaf is, Zariski-locally on $X$, a module $M$ over a ring $R$, and its sections over an $X$-scheme $U$ are $M \otimes_R \mathcal{O}(U)$. The sheaf condition for a cover $U' \to U$ becomes the statement that
is an equalizer whenever $R \to S$ is faithfully flat of finite presentation. That is exactly the affine lemma of §6 — the algebraic heart of descent. So the theorem is not an accident of the étale topology: it is the same flatness statement that makes descent work at all.
This is what allows us to speak of “the same” quasi-coherent sheaf in four different topologies, and hence to compare Zariski cohomology with étale and fppf cohomology of $\mathcal{O}_X$-modules. It also means that the interesting part of descent is not that quasi-coherent sheaves are sheaves, but that they can be reconstructed from data downstairs — the subject of the rest of the lecture.
Fix a cover $f : U \to X$. Given a sheaf on $X$, restricting along $f$ gives a sheaf on $U$; but restricting alone forgets information, because a cover has overlaps on which the restrictions must be compatible.
Let $f : U \to X$ be a cover and write $U^{\times_X 2} = U \times_X U$, $U^{\times_X 3} = U \times_X U \times_X U$, with the projections $p_1, p_2 : U \times_X U \to U$ and $p_{ij} : U^{\times_X 3} \to U \times_X U$. Descent data for a quasi-coherent sheaf along $f$ is a quasi-coherent sheaf $\mathcal{F}$ on $U$, together with an isomorphism
on $U \times_X U$ — the gluing isomorphism — satisfying the cocycle condition on $U^{\times_X 3}$:
The three projections and their two ways around the triple overlap are worth reading as a picture: the cocycle condition says that the three identifications of $\mathcal{F}$ pulled back from the three copies of $U$ are consistent on the triple intersection. This is literally the same condition as the usual cocycle condition for transition functions of a vector bundle.
Descent data is “a sheaf on the cover plus consistent identifications on overlaps” — the same shape as gluing local data in the Zariski topology, but with the open cover replaced by an arbitrary cover in a Grothendieck topology.
A morphism of descent data $(\mathcal{F}, \varphi) \to (\mathcal{G}, \psi)$ is a morphism $h : \mathcal{F} \to \mathcal{G}$ of quasi-coherent sheaves on $U$ compatible with the gluing isomorphisms, i.e.
Descent data on $U/X$ therefore forms a category, denoted $\mathrm{Desc}(U/X)$ in the statement below.
For a Zariski cover $U = \bigsqcup_i U_i \to X$ of a scheme $X$, descent data along $f$ is exactly gluing data: quasi-coherent sheaves $\mathcal{F}_i$ on the pieces together with isomorphisms $\varphi_{ij} : \mathcal{F}_i|_{U_{ij}} \to \mathcal{F}_j|_{U_{ij}}$ on the overlaps satisfying $\varphi_{jk}\varphi_{ij} = \varphi_{ik}$ on the triple overlaps. The conclusion is the familiar statement that a quasi-coherent sheaf on $X$ is the same thing as compatible quasi-coherent sheaves on the pieces — which is why Zariski descent is unremarkable and the theorem below is not.
Let $L/k$ be a finite Galois extension with group $G$ and take $U = \operatorname{Spec} L \to \operatorname{Spec} k = X$. A quasi-coherent sheaf on $\operatorname{Spec} L$ is a vector space $V$ over $L$, and descent data is an action of $G$ on $V$ by semilinear maps: a collection of isomorphisms $g : V \to V$ satisfying the cocycle condition $gh = g \circ h$. The descent theorem below then asserts that
recovering classical Galois descent. This is Exercise 3 in §8; it is the single best example to carry in mind for everything that follows.
Let $f : U \to X$ be an fppf cover (faithfully flat and locally of finite presentation). Then the pullback functor
sending a quasi-coherent sheaf $\mathcal{F}$ to $(f^*\mathcal{F}, \text{the canonical gluing isomorphism})$ is an equivalence of categories.
An equivalence is two statements, and they are proved by different means:
Full faithfulness is a statement about morphisms and reduces to an equalizer diagram; essential surjectivity is a statement about objects and requires a construction. We take them in turn.
We must show that
is an equalizer: a morphism downstairs is exactly a morphism on the cover whose two pullbacks to the overlap agree. Since $f$ is fppf, we may work Zariski-locally on $X$ and $U$ and reduce to the affine case $X = \operatorname{Spec} R$, $U = \operatorname{Spec} S$, where $\mathcal{F}, \mathcal{G}$ correspond to $R$-modules $M, N$ and $f^*$ corresponds to $- \otimes_R S$. The assertion becomes: for a faithfully flat $R$-algebra $S$ of finite presentation and $R$-modules $M, N$, the diagram
is an equalizer. This follows from the affine lemma of §6 applied to $M$ and to $N$, together with the exactness of $-\otimes_R S$: a map is determined by its base change because the first map is injective, and a compatible map on the cover descends because the middle exactness of the lemma supplies the needed element downstairs. This is where faithful flatness is used, and only there.
The whole of fppf descent rests on one algebraic fact, which is worth isolating because it is also the engine of Galois descent and of flat descent for modules.
Let $R \to S$ be a faithfully flat ring homomorphism and let $M$ be an $R$-module. Then the sequence
is exact, where $d_0(m) = m \otimes 1$, $d_1(m \otimes s) = m \otimes s \otimes 1$, and $d_2(m \otimes s) = m \otimes 1 \otimes s$. That is, $M$ is the equalizer of $d_1$ and $d_2$.
Injectivity of $d_0$. If $m \neq 0$ in $M$, choose a linear functional $\lambda : M \to R$ with $\lambda(m) \neq 0$; then $m \otimes 1 = 0$ in $M \otimes_R S$ would force $\lambda(m) \otimes 1 = 0$ in $S$, contradicting faithful flatness.
Exactness in the middle. First suppose the map $R \to S$ splits, i.e. there is an $R$-module map $S \to R$ with $1 \mapsto 1$. Then the sequence is split exact, which is a formal computation with the two face maps. In general one reduces to the split case: $S \to S \otimes_R S$, $s \mapsto s \otimes 1$, is a faithfully flat base change that admits a section, namely multiplication $S \otimes_R S \to S$. Exactness can be checked after this base change because $-\otimes_R S$ is faithfully flat, and it holds there by the split case. (This is Exercise 1 in §8.)
□
Everything in the theorem is a consequence of this one sequence. “Descent” is the statement that a module over $R$ is exactly a module over $S$ with a compatible isomorphism between its two pullbacks to $S \otimes_R S$ — the cocycle condition of §4 is the triple-overlap shadow of the same exactness.
It remains to build the descended object. Fix an fppf cover $U \to X$ and a descent datum $(\mathcal{F}, \varphi)$ on $U$.
First reduce to the case where $U \to X$ is an affine morphism, so that $X = \operatorname{Spec} R$, $U = \operatorname{Spec} S$ with $R \to S$ faithfully flat of finite presentation. The descent datum is an $S$-module $M$ together with an isomorphism of $S \otimes_R S$-modules
satisfying the cocycle condition on $S \otimes_R S \otimes_R S$. Set
the descent of $M$: the elements of $M$ that are fixed by the gluing isomorphism. Then $N$ is an $R$-module, and one checks that the natural map $N \otimes_R S \to M$ is an isomorphism. The verification that $N \otimes_R S \to M$ is an isomorphism uses the affine lemma of §6 applied after the faithfully flat base change to $S$: the sequence acquires a section, hence splits, and the comparison map can be read off from the resulting decomposition. Independence of the construction from the choice of presentation is where the cocycle condition is used.
□
The reduction of the general case to the affine case, and the verification that the construction is independent of choices, are the technical parts. The same argument, run for sheaves of modules rather than modules, gives descent for sheaves I: for a cover $U \to X$, descent data on $U/X$ is fully faithful for the category of sheaves, and the construction above supplies the descended object when a presentation exists. The subtlety — in fact the whole reason fppf descent is a theorem rather than a tautology — is that for arbitrary sheaves the functor need not be essentially surjective; it is on quasi-coherent sheaves that the argument above applies.
Next time: the second part of fppf descent, descent for sheaves relative to an étale cover, and the beginning of the study of the category of sheaves. Following Litt’s Lecture 5.
These are the exercises marked in Litt’s Lecture 4. They follow the order in which the ideas appeared above: the equalizer lemma first, then the reduction steps that make the descent theorem available, and finally Galois descent as the guiding example.
Let $R \to S$ be a faithfully flat ring homomorphism and $M$ an $R$-module. Prove that
is an exact sequence. Hint: handle the case in which $R \to S$ splits, then reduce to it by base changing along $S \to S \otimes_R S$ and using that exactness can be checked after a faithfully flat base change.
Let $U \to X$ be an fppf cover. Show that in order to prove full faithfulness of $f^* : \mathbf{QCoh}(X) \to \mathrm{Desc}(U/X)$, it suffices to treat the case in which $X = \operatorname{Spec} R$ and $U = \operatorname{Spec} S$ are affine with $R \to S$ faithfully flat of finite presentation. What to check: that the statement is Zariski-local on $X$ and on $U$, and that restriction of quasi-coherent sheaves to an affine open commutes with the morphism sets involved.
Let $L/k$ be a finite Galois extension with Galois group $G$, and take the fppf cover $\operatorname{Spec} L \to \operatorname{Spec} k$.
Hint: the cocycle condition on the triple overlap is exactly the relation $(gh)(v) = g(h(v))$; average over $G$ to identify the descended subspace with the invariants $V^G$ when $L/k$ is finite.
Let $U = \bigsqcup_i U_i \to X$ be a Zariski cover. Check that descent data for quasi-coherent sheaves along this cover agrees with the usual gluing data
so that the fppf descent theorem specializes to the statement that quasi-coherent sheaves glue in the Zariski topology — a statement we already know. Moral: a theorem of descent should be checked against the cases where there is nothing to prove.