We have an fppf cover $f : U \to X$ and the pullback functor
Last time we established full faithfulness by reducing to the affine case $X = \operatorname{Spec} R$, $U = \operatorname{Spec} S$ with $R \to S$ faithfully flat of finite presentation, where the assertion becomes the exactness of
The plan for today is:
Let $f : U \to X$ be an fppf cover. Then $f^* : \mathbf{QCoh}(X) \to \mathrm{Desc}(U/X)$ is an equivalence of categories.
Only essential surjectivity remains: given a descent datum $(\mathcal{F}, \varphi)$ on $U$, we must produce $\mathcal{G} \in \mathbf{QCoh}(X)$ with $f^*\mathcal{G} \cong \mathcal{F}$ compatible with $\varphi$.
Suppose $U \to X$ is affine. Then $X = \operatorname{Spec} R$, $U = \operatorname{Spec} S$ with $R \to S$ faithfully flat of finite presentation, and the descent datum is an $S$-module $M$ together with a gluing isomorphism
satisfying the cocycle condition on $S \otimes_R S \otimes_R S$. (The reduction from a general fppf cover to the affine case is Exercise 1 in §7: one first reduces to $U \to X$ affine, using that descent data for quasi-coherent sheaves is local on $X$, and then to $X$ affine.)
Define
the elements of $M$ fixed by the gluing isomorphism. Then $N$ is an $R$-module and there is a natural $S$-linear comparison map
$\kappa$ is an isomorphism. Since $R \to S$ is faithfully flat, it suffices to check this after base changing along $S \to S \otimes_R S$, i.e. to show that $\kappa \otimes_S (S \otimes_R S)$ is an isomorphism. After this base change the sequence of the affine lemma acquires a section, hence splits: base changing $M$ along $S \to S \otimes_R S$ exhibits $M \otimes_S (S \otimes_R S)$ as a direct summand carrying a copy of $M$ on which the two face maps agree. The fixed locus of $\varphi$ is then exactly the image of $M$ under this splitting, and $\kappa$ identifies the two. This is the step where the cocycle condition is used: it is what makes $\varphi$ an action of a groupoid rather than an arbitrary isomorphism, and hence what makes “fixed elements” well behaved. Independence of $N$ from the choices made is a formal consequence of the same splitting.
□
For an fppf cover $U \to X$, a quasi-coherent sheaf on $X$ is the same thing as a quasi-coherent sheaf on $U$ equipped with a compatible gluing isomorphism on the double overlap. Descent is not an extra theorem about gluing — it is the statement that Zariski gluing is a special case of a mechanism that works for every fppf cover.
Now let $\mathcal{C}$ be a site, $U \to X$ a cover in $\mathcal{C}$, and consider sheaves of $\mathcal{O}$-modules rather than quasi-coherent sheaves on a scheme. The same two questions arise, and the answers are no longer symmetric.
Let $f : U \to X$ be an étale (or fppf) cover of a scheme $X$ and let $\mathcal{F}, \mathcal{G}$ be sheaves of $\mathcal{O}_X$-modules on the corresponding site. Then
is an equalizer. Equivalently, the functor from sheaves on $X$ to descent data on $U/X$ is fully faithful.
Reduce to the case where $f : U \to X$ is affine and $X$ is affine, so that the situation is $R \to S$ faithfully flat with $M, N$ modules. Then use the affine lemma of the previous meeting to exhibit the diagram of $\operatorname{Hom}$-sets as an equalizer. The reduction uses that a map of sheaves is determined locally and that the sheaf condition itself is an equalizer condition — so the argument is the same shape as the quasi-coherent one, with “module” replaced throughout by “sheaf of modules”.
For quasi-coherent sheaves on a scheme, descent data along an fppf cover is always effective: essential surjectivity holds. For general sheaves this fails. Two things then need separating:
Let $U \to X$ be a cover and let $\mathcal{A}$ be a sheaf of $\mathcal{O}_U$-algebras equipped with descent data. Then $\mathcal{A}$ descends to a sheaf of $\mathcal{O}_X$-algebras $\mathcal{B}$ (algebras are modules, so the mechanism of §2 applies, together with compatibility of the multiplication with $\varphi$). Applying the relative spectrum construction,
so the descended scheme is recovered from descended algebra data. This is the bridge from descent of sheaves to descent of schemes, and it is the reason relative spectra appear naturally in this lecture. The condition that $\mathcal{A}$ be quasi-coherent as a sheaf of algebras is the analogue of the effectiveness hypothesis above.
To see that the sheaves appearing in practice really are sheaves for both topologies, note that
are all representable (§3 of the previous meeting), hence sheaves on $X_{\mathrm{fppf}}$ and therefore on $X_{\mathrm{\acute{e}t}}$ and $X_{\mathrm{Zar}}$. The Kummer sequence
is exact as a sequence of sheaves on $X_{\mathrm{\acute{e}t}}$ when $n$ is invertible on $X$, and on $X_{\mathrm{fppf}}$ for every $n$ — the fact established at the end of the previous meeting, and the reason the fppf topology is worth the trouble.
We now change subject, from descent to the structure of the category of sheaves itself. This is the beginning of the machinery that will produce étale cohomology.
Let $\mathcal{C}$ be a site. Then the category
of abelian-group-valued sheaves on $\mathcal{C}$ is an abelian category, and it has enough injectives. Consequently every object admits an injective resolution, and right derived functors of left exact functors out of $\mathrm{Sh}(\mathcal{C},\mathbf{Ab})$ — in particular $H^i(\mathcal{C}, -) = R^i\Gamma(\mathcal{C}, -)$ — are defined.
Categories of presheaves with values in $\mathbf{Ab}$ are very well behaved: limits and colimits are computed objectwise, so kernels and cokernels exist and the abelian-category axioms are inherited from $\mathbf{Ab}$. Sheaves are a full subcategory cut out by an equalizer condition, and the subtlety is that this subcategory is not closed under cokernels computed in presheaves. The missing ingredient is a left adjoint to the inclusion — sheafification — which both repairs the cokernels and supplies enough injectives. This is the content of the next section.
Injective objects in a category of presheaves are easy to manufacture: for a presheaf category, the functors $\prod_{U} \operatorname{Hom}_{\mathbf{Ab}}(\mathbb{Z}[h_U], -)$-shaped constructions give injectives, using the injective objects of $\mathbf{Ab}$ — that is, divisible abelian groups. Because the forgetful functor $\mathrm{Sh} \to \mathrm{Pre}$ has a left adjoint, it is right adjoint to something, so it preserves injectives up to the appropriate statement: pushing a presheaf injective into sheaves via the right adjoint produces injectives in $\mathrm{Sh}$. Every sheaf embeds into such an object, which is enough injectives. (This is the standard two-step argument; the adjunction of §5 is the only non-formal input.)
Everything up to now has been setting the stage. Étale cohomology is the right derived functor of global sections on this category. Without abelian-ness and enough injectives there is no long exact sequence, no derived functor, and no cohomology at all — so this unglamorous structural theorem is the gate through which the whole subject passes.
Let $\mathcal{C}$ be a site. The forgetful functor
admits a left adjoint, called sheafification (or the associated sheaf functor), written $F \mapsto F^{+}$. That is, for a presheaf $F$ and a sheaf $G$ there is a natural bijection
and the unit $F \to F^{+}$ is the universal map from $F$ to a sheaf.
The associated sheaf is built in two steps:
Finally, we make the definition of the previous meeting do work. Let
be a morphism of schemes, and consider the induced continuous morphism of sites obtained from the functor $U \mapsto U \times_Y X$ — the algebraic analogue of $V \mapsto f^{-1}(V)$. Then $u$ preserves fiber products (fiber products commute with base change) and sends covers to covers (base change of an étale or fppf cover is again one).
For a sheaf $\mathcal{G}$ on $X$, define the pushforward (direct image) $f_*\mathcal{G}$ to be the sheaf on $Y$ given on an object $V \to Y$ by
Equivalently, $f_*\mathcal{G} = \mathcal{G} \circ u$, the composite of $\mathcal{G}$ with the functor $u : Y \to X$ of the continuous morphism. One checks directly that $f_*\mathcal{G}$ satisfies the sheaf condition, precisely because $u$ sends covers to covers and preserves the fiber products along which the gluing conditions are imposed.
For a sheaf $\mathcal{F}$ on $Y$, define $f^{-1}\mathcal{F}$ on $X$ by
the colimit over the category of factorizations $U \to V \to Y$ — and then sheafify:
The sheafification is essential: $f^{-1}\mathcal{F}$ is only a presheaf in general, and it is the sheafification that makes it a sheaf on a finer topology. For this reason pullback is right exact but not left exact; its derived functors — the higher direct images, on the other side of the adjunction — are exactly what Leray spectral sequences and later lectures will compute.
The functors $f^{-1}$ and $f_*$ are adjoint,
and consequently $f^{*}$ is left adjoint to $f_*$ on sheaves of modules. Taking global sections of the adjunction recovers the usual formalism: a map of sheaves downstairs is the same as a map downstairs after pulling back.
With sheafification, the abelian category of sheaves, and the pushforward/pullback adjunction in hand, the machinery of cohomology is complete in outline. The next block (Litt’s Lectures 6–10) develops stalks, enough injectives in more detail, derived-functor cohomology, Çech cohomology, the Artin–Schreier sequence, and the first computations of $H^i_{\mathrm{\acute{e}t}}(X, \mathbb{G}_m)$ and $H^i_{\mathrm{\acute{e}t}}(X, \mu_\ell)$ — the point at which the theory starts returning answers. Following the pilot, the group decides whether to continue.
These are the exercises marked in Litt’s Lecture 5. They follow the order in which the ideas appeared above: reducing to the affine case, completing the descent theorem, and then the pullback construction for maps of sites.
Let $f : U \to X$ be an fppf cover of schemes. Show that in order to prove essential surjectivity of $f^* : \mathbf{QCoh}(X) \to \mathrm{Desc}(U/X)$ — and in order to prove full faithfulness — it is enough to treat the case where $U \to X$ is an affine morphism, and then the case where $X$ itself is affine.
What to check: that descent data restricts along an open immersion $X' \hookrightarrow X$ to descent data for the restricted cover, that a quasi-coherent sheaf on $X$ can be produced by gluing its restrictions to a Zariski cover of $X$, and that these two constructions are inverse to one another. Moral: descent statements are Zariski-local on the target, which is why “affine” is always available as a hypothesis.
Keep the notation of §2: $R \to S$ is faithfully flat of finite presentation, $M$ an $S$-module with a gluing isomorphism $\varphi$ satisfying the cocycle condition, and
Show that the sequence of the affine lemma splits after the faithfully flat base change $S \to S \otimes_R S$, and deduce that $\kappa$ is an isomorphism. Hint: the multiplication map $S \otimes_R S \to S$, $s_1 \otimes s_2 \mapsto s_1 s_2$, is a section of $s \mapsto s \otimes 1$; use the resulting idempotent to exhibit $M$ as a direct summand carrying the fixed locus of $\varphi$.
Let $U \to X$ be a cover and let $\mathcal{A}$ be a sheaf of $\mathcal{O}_U$-algebras with descent data $(\mathcal{A}, \varphi)$ in which $\varphi$ is compatible with the multiplication. Show that $\mathcal{A}$ descends to a sheaf of $\mathcal{O}_X$-algebras $\mathcal{B}$, and that the relative spectrum construction descends:
What to check: that the descended multiplication is well defined, that $\varphi$ being multiplicative is exactly the compatibility needed, and that the universal property of the relative spectrum is preserved under descent.
Let $f : X \to Y$ be a morphism of schemes and $\mathcal{F}$ a sheaf on $Y_{\mathrm{\acute{e}t}}$. Consider the presheaf on $X$ defined by
Give an example showing that this presheaf need not be a sheaf on $X_{\mathrm{\acute{e}t}}$, so that the sheafification in the definition of $f^{*}$ is genuinely necessary rather than a formality. Hint: take $\mathcal{F}$ representable and ask whether the colimit description already satisfies the gluing condition; the failure is the same phenomenon as the failure of the presheaf cokernel to be a sheaf. Then verify the adjunction $\operatorname{Hom}_X(f^{-1}\mathcal{F}, \mathcal{G}) \cong \operatorname{Hom}_Y(\mathcal{F}, f_*\mathcal{G})$ directly from the definitions.