# Using algebraic geometry to understand class field theory

In Algebraic Number Theory, S. Lang says "[a geometrical approach] allows one to have a much clearer insight into the whole class field theory, since the existence theorem and the reciprocity law become obvious once the machinery of algebraic geometry is available."

Inspired by this, I wonder if there is some (preferably modern) reference for class field theory using algebraic geometry.

• I'm not sure I'm interpreting this correctly but the analogy seems questionable. Wouldn't it mean that Langlands for $\mathrm{GL}_2$ over $\mathbb{Q}$ becomes obvious once Langlands for $\mathrm{GL}_2$ over function fields is established? – Faris Mar 24 at 6:04
• This is kind of what I had in mind when asking this question, which didn't really receive a definitive answer. – Tim Campion Mar 24 at 6:13
• @Faris, I unfortunately don't know how to answer your question. – Gabriel Mar 24 at 9:59
• I'd suspect that Lang meant that a heuristic for the number field case "becomes obvious" after the (etale cohomology...!) of the function-field/geometric case is understood. Lang did use hyperbole now and then. :) – paul garrett Mar 24 at 21:59
• In context, Lang meant that there is an alternative geometric approach to abelian class field theory in the function field case, which is why he doesn't cover that case in his book. – anon Mar 25 at 4:41

The usual reference here is Serre's Algebraic Groups and Class Fields. But it still felt to me like there was a lot of magic left over after I read this.

We had an earlier question along similar lines with many good answers. I remember really liking David Ben Zvi's lecture, but wishing that some one would write up the "decategorification" to get from the categorical statements in Ben Zvi's talk to concrete reciprocity laws.

This is not going to be a reference, but I think the following is a great exercise (using étale cohomology) to prove the fundamental exact sequence for Brauer groups of global function fields.

Theorem: Define $$k := \mathbf{F}_q$$ and let $$X/k$$ be a smooth proper geometrically connected curve with function field $$K$$. Then there is an exact sequence $$0 \to \text{Br}(K) \to \bigoplus_{\nu} \text{Br}(K_\nu) \to \mathbf{Q}/\mathbf{Z} \to 0.$$ The sum in the middle is over all (non-archimedean) places of $$K$$.

Let us first prove the following.

Lemma: Let $$A/k$$ be an abelian variety. Then for all $$i \geq 1$$, $$H^i_{\text{ét}}(k,A) = 0.$$ When $$i = 1$$, this is a special case of Lang's theorem. Now for $$i > 1$$, let $$n \in \mathbf{N}$$ be any integer, and consider the exact sequence in the fppf topology $$0 \to A[n] \to A \stackrel{n\cdot }{\to} A \to 0.$$ Now I claim that $$H^i_{\text{fppf}}(k, A[n]) = 0$$ for all $$i > 1$$. Indeed, since $$k$$ is perfect, we have (see here) $$H^i_{\text{fppf}}(k, A[n]) = H^i_{\text{ét}}(k, A[n])$$ which vanishes for cohomological dimension reasons. It follows that multiplication by $$n$$ on $$H^i_{\text{ét}}(k, A)$$ is an isomorphism. On the other hand, $$H^i_{\text{ét}}(k,A)$$ is also torsion and therefore must be zero since $$n$$ was arbitrary.

We can now prove the theorem. From now on all cohomology considered will be étale cohomology. Consider the divisor exact sequence $$0 \to \mathbf{G}_{m,X} \to j_\ast \mathbf{G}_{m, \eta} \to \bigoplus_{x \in |X|} (i_x)_\ast \mathbf{Z} \to 0$$ where $$\eta$$ is the generic point of $$X$$, $$j: \eta \to X$$ the canonical inclusion, and the sum on the right is over all closed points of $$X$$. Taking cohomology, we get $$H^2(X,\mathbf{G}_m) \to H^2(X, j_\ast \mathbf{G}_m) \to \oplus_{x \in |X|} H^2(X, (i_x)_\ast \mathbf{Z})\to H^3(X, \mathbf{G}_m) \to H^3(X, j_\ast \mathbf{G}_m).$$

We must now show that: $$\begin{eqnarray} H^2(X,\mathbf{G}_m) &=& 0 \\ H^2(X, j_\ast\mathbf{G}_m) &=& H^2(K,\mathbf{G}_m) \stackrel{\text{def}}{\equiv} \text{Br}(K) \\ H^3(X,\mathbf{G}_m) &=& \mathbf{Q}/\mathbf{Z}\\ H^3(X, j_\ast \mathbf{G}_m) &=& 0.\end{eqnarray}$$

I will compute the first of these and leave the rest as (great!) exercises for you. Let $$f : X \to \operatorname{Spec} k$$ denote the structure map. Consider the Leray spectral sequence with second page $$E_2^{i,j} := H^i(k, R^jf_\ast \mathbf{G}_m)$$ converging to $$E_\infty^{i+j} := H^{i+j}(X, \mathbf{G}_m).$$ We are interested in computing the $$E_\infty^2$$ term, namely $$H^2(X,\mathbf{G}_m)$$. By construction of the Leray spectral sequence, this admits a filtration $$0 = F^{-1} \subseteq F^0 \subseteq F^1 \subseteq F^2 = E_\infty^2$$ with $$\begin{eqnarray} F^0/F^{-1} &\simeq& E_\infty^{2,0} \\ F^1/F^0 &\simeq& E_\infty^{1,1} \\ F^2/F^1 &\simeq& E_\infty^{0,2}.\end{eqnarray}$$ We now compute each of these. First, we deal with the $$E_\infty^{2,0}$$ term. From the 2nd page of the Leray spectral sequence, we see that $$E_3^{2,0} \simeq E_2^{2,0}/\text{im}(E_2^{0,1} \to E_2^{2,0}).$$ But $$E_2^{2,0} = H^2(k, \mathbf{G}_m)$$ since $$f_\ast \mathbf{G}_m = \mathbf{G}_m$$ (cohomology and base change!). This can be identified with the (Azumaya) Brauer group of $$k$$, which we know is zero by Wedderburn's little theorem. Hence $$E_3^{2,0} \implies E_\infty^{2,0}$$ and thus $$F^0/F^{-1} = 0$$.

Now we show that $$F^1/F^0 = 0$$. Again, it is sufficient to show that $$E_3^{1,1} = 0$$. We see that $$E_3^{1,1} \simeq \ker(E_2^{1,1} \to E_2^{3,0})$$. However, the higher Galois cohomology of a finite field with coefficients in $$\mathbf{G}_m$$ is zero (by an argument similar to the Lemma above using the Kummer sequence), so it is sufficient to show that $$E_2^{1,1} = H^1(k, R^1 f_\ast \mathbf{G}_m) = 0$$. To this end, identify $$R^1f_\ast \mathbf{G}_m$$ with the Picard functor of $$X/k$$, and consider the exact sequence $$0 \to \text{Pic}^0_{X/k} \to \text{Pic}_{X/k} \to \mathbf{Z} \to 0,$$ where the last map is $$\mathcal{L} \mapsto \text{deg}(\mathcal{L})$$. Taking cohomology, we have an exact sequence $$H^1(k, \text{Pic}^0_{X/k}) \to H^1(k, \text{Pic}_{X/k}) \to H^1(k, \mathbf{Z}).$$ The left term is zero by the lemma above. The right term is zero because it can be identified with continuous homomorphisms from $$\text{Gal}(\overline{k}/k)$$ to $$\mathbf{Z}$$, of which there are none since the former is profinite while the latter discrete. This shows the middle term is zero, completing the proof that $$E_\infty^{1,1} = 0$$.

Finally, to show that $$E_\infty^{0,2}$$ is zero, observe that $$R^2f_\ast \mathbf{G}_m$$ is zero. The reason is because its stalk at the geometric point $$\overline{x} : \operatorname{Spec} \overline{k} \to \operatorname{Spec}k$$ is the Brauer group of $$\overline{X} := X \times_{k} \overline{k}$$. But $$\overline{X}$$ is a regular integral scheme, and hence $$\text{Br}(\overline{X})$$ must inject into the Brauer group of its function field which is zero by Tsen's theorem.

A good reference is Neukirch's Algebraic Number Theory. It takes the analogy between algebraic number theory and algebraic geometry seriously, and it includes coverage of class field theory. But you'd probably be wise to learn algebraic number theory well (say from Lang's book) before or while tackling Neukirch's text.

$$\def\cO{\mathcal{O}}\def\FF{\mathbb{F}}\def\fq{\mathfrak{q}}\def\fp{\mathfrak{p}}\def\Gal{\mathrm{Gal}}\def\Frob{\mathrm{Frob}}$$I'm going to try to write up two insights that we get from the geometric picture. A basic result of class field theory is that, for $$K$$ a number field, the Galois group of the maximal unramified abelian extension of $$K$$ is isomorphic to the class group of $$K$$. (And then there are generalizations of this to the ramified case, but the unramified case illustrates the points I want to make.)

More precisely, let $$L/K$$ be an unramified abelian extension. Let $$\fq$$ be a prime of $$\cO_L$$ and let $$\fp$$ be $$\cO_K \cap \fq$$. There is a unique element $$\Frob(\fq)$$ of $$\Gal(L/K)$$ such that $$\Frob(\fq)$$ fixes $$\fq$$ and the induced action on $$\cO_L/\fq$$ is the $$\#(\cO_K/\fp)$$-power Frobenius. Using that $$\Gal(L/K)$$ is abelian, $$\Frob(\fp)$$ depends only on $$\fp$$, not on $$\fq$$. (In an nonabelian extension, changing $$\fq$$ to $$\fq'$$ would conjugate the Frobenius element.) The Artin map, from the ideal group of $$K$$ to $$\Gal(L/K)$$, sends $$\prod \fp^{a_p}$$ to $$\prod \Frob(\fp)^{a_p}$$. Then class field theory says two things:

(1) The Artin map factors through the class group and

(2) There is an unramified abelian extension, the class field, for which the Artin map is an isomorphism.

Let's see what each of these mean geometrically.

Let $$k$$ be the finite field with $$q$$ elements, let $$X$$ be a smooth, projective, geometrically irreducible curve over $$k$$, and let $$K$$ be the field of meromorphic functions on $$X$$. Then unramified extensions of $$K$$ correspond to unbranched covers $$Y$$ of $$X$$; such extensions are linearly disjoint from $$k^{\text{alg}}$$ if and only if $$Y$$ is geometrically connected.

Understanding (1): Let $$\pi: Y \to X$$ be an abelian unbranched cover with covering group $$G$$. We can think of $$Y$$ to $$X$$ as a principal bundle with group $$G$$. Let $$\mathfrak{p}$$ be a closed point of $$X$$, with residue field of size $$q^f$$. Geometrically, we can think of this as $$f$$ points $$x_1$$, $$x_2$$, ..., $$x_f$$ of $$X(k^{\text{alg}})$$ which are permuted cyclically by the $$q$$-power Frobenius, say $$x_1 \mapsto x_2 \mapsto \dots \mapsto x_f \mapsto x_1$$. Then $$q$$-power Frobenius must also cycle the fibers $$\pi^{-1}(x_1) \to \pi^{-1}(x_2) \to \cdots \to \pi^{-1}(x_f) \to \pi^{-1}(x_1)$$. So $$q^f$$ power Frobenius maps each $$\pi^{-1}(x_i)$$ to itself. This map is multiplication by some element of $$G$$, and this element is $$\Frob(\fp)$$. (If $$G$$ were not abelian, then the action of $$\Frob(\fp)$$ on $$\pi^{-1}(x_1)$$ is a permutation which commutes with the $$G$$-action on $$\pi^{-1}(x_1)$$. The group of such permutations is isomorphic to $$G$$, but the isomorphism is only natural up to inner conjugacy, so we only get a conjugacy class in $$G$$. I learned about this idea from John Baez, search for "dual groups" on the linked page.)

This gives us a geometric way to understand the Frobenius element, but it doesn't help us prove (1) yet. We can do that using an idea I learned from Ben-Zvi's lecture. For any positive integer $$N$$, we can look at $$Y^N \to X^N$$ as a principal bundle with group $$G^N$$ and can then quotient by $$\{ (g_1, g_2, \ldots, g_N) : \prod g_N = 1 \}$$ to get a principal $$G$$-bundle over $$X^N$$. (Note that what we are quotienting by is a subgroup because $$G$$ is abelian.) Moreover, this quotient descends to a principle $$G$$-bundle $$\psi: Z^N \to \mathrm{Sym}^N(X)$$.

If $$x_1$$, $$x_2$$, ..., $$x_f$$ are as two paragraphs ago, then $$(x_1, x_2, \ldots, x_f)$$ is a fixed point of the $$q$$-power Frobenius action on $$\mathrm{Sym}^f(X)$$ and the action of $$q$$-power Frobenius on $$\psi^{-1} (x_1, x_2, \ldots, x_f)$$ is $$\Frob(\fp)$$. More generally, if $$D$$ is an effective element of the divisor class group, with degree $$N$$, then $$D$$ gives a Frobenius-fixed point $$[D]$$ of $$\mathrm{Sym}^N(X)$$, and the Artin map describes the action of $$q$$-power Frobenius action on $$\psi^{-1}([D])$$.

This gives us a geometrically natural way to think of the Artin map, but can it prove (1)? Let $$D$$ and $$E$$ be effective divisors both of degree $$N$$ which represent the same class in $$\mathrm{Pic}^N(X)$$. So $$[D]$$ and $$[E]$$ are in the same fiber of $$\mathrm{Sym}^N(X) \to \mathrm{Pic}^N(X)$$. How can we see that the Frobenius actions on $$\psi^{-1}([D])$$ and $$\psi^{-1}([E])$$ coincide?

For $$N$$ large enough, all the fibers of $$\mathrm{Sym}^N(X) \to \mathrm{Pic}^N(X)$$ are projective spaces! (And, if $$N$$ isn't large enough, we can add the same high degree divisor to $$D$$ and $$E$$ to make it large enough.) And projective space is geometrically simply connected! So the $$G$$-principal bundle $$\psi : Z^N \to \mathrm{Sym}^N(X)$$ must be geometrically trivial when restricted to any fiber of $$\mathrm{Sym}^N(X) \to \mathrm{Pic}^N(X)$$. ("Geometrically" meaning after we base change to $$k^{\text{alg}}$$.) So, restricting to the fiber through $$[D]$$ and $$[E]$$, the principal bundle $$\psi$$ is geometrically trivial, and Frobenius must act by permuting the geometric components. In particular, Frobenius acts the same way on $$\psi^{-1}([D])$$ and $$\psi^{-1}([E])$$.

Understanding (2). In number fields, the class group is finite. In function fields, the analogue of the class group is $$\mathrm{Pic}(X)(k)$$, which is not quite finite, it sits in a short exact sequence $$0 \to \mathrm{Pic}^0(X)(k) \to \mathrm{Pic}(X)(k) \to \mathbb{Z} \to 0. \quad (\ast)$$ So we are not going to ask for the Artin map to be an isomorphism, but we can still take finite quotients of $$\mathrm{Pic}(X)(k)$$ and ask what covers they correspond to.

The quotient $$\mathrm{Pic}(X)(k) \to \mathbb{Z} \to \mathbb{Z}/d \mathbb{Z}$$ corresponds to extending the ground field to have $$q^d$$ elements; it is a good exercise to check that the Artin map does fit into this story.

The more interesting thing is covers that come from $$\mathrm{Pic}^0(X)(k)$$. However, these are harder to talk about because the extension $$(\ast)$$ doesn't come with a natural splitting. (It is splittable, since $$\mathbb{Z}$$ is projective in the category of abelian groups.) However, if $$X$$ has a $$k$$-point $$x_{\infty}$$, then we can make a splitting $$\mathrm{Pic}(X)(k) \longrightarrow \mathrm{Pic}^0(X)(k)$$ by $$D \mapsto D - \mathrm{deg}(D) x_{\infty}$$, and we can ask what covering of $$X$$ corresponds to this quotient. (If you want to see the theory without the assumption of a $$k$$-point on $$X$$, I recall that Serre has a rather difficult chapter on this point.)

Let's abbreviate the Jacobian $$\mathrm{Pic}^0(X)$$ to $$J$$, and lets abbreviate the finite group $$J(k) = \mathrm{Pic}^0(X)(k)$$ to $$G$$. Frobenius is an addditive map $$F:J \to J$$ and, using the group structure on $$J$$, we can talk about the isogeny $$F-\mathrm{Id} : J \to J$$. The kernel of $$F - \mathrm{Id}$$ is $$G$$, so $$F-\mathrm{Id} : J \to J$$ is a principal $$G$$-bundle.

Using the point $$x_{\infty}$$, we can embed $$X$$ into $$J$$ by $$x \mapsto x-x_{\infty}$$, so we can pull the $$G$$-bundle $$F-\mathrm{Id}: J \to J$$ back to $$X$$; let's call this $$\pi : Y \to X$$.

Let us understand how the Artin map works for $$\pi : Y \to X$$. Again, let $$\fp$$ be a closed point of $$X$$ corresponding to a Frobenius orbit of geometric points $$x_1 \mapsto x_2 \mapsto \cdots \mapsto x_f \mapsto x_1$$. Choose a geometric point $$y_1$$ over $$x_1$$ in $$Y$$, and let $$y_j = F^{j-1}(y_1)$$. So $$\pi(y_j) = x_j$$. Writing out the defnition of $$\pi$$, we have $$F(y_j) - y_j = x_j$$ or, in other words, $$F^{j+1}(y_1) - F^j(y_1) = x_j$$. This telescopes to give $$F^f(y_1) - y_1 = \sum x_i$$. But $$\Frob(\fp)$$ is the element of $$G$$ by which $$F^f$$ acts on $$\pi^{-1}(x_1)$$. So we have just computed that $$\Frob(\fp)=\sum x_i$$.

More generally, if $$D$$ is any effective divisor on $$X$$, we have just computed that the Artin map takes $$D$$ to the sum of the geometric points of $$D$$, computed in $$G = J(k)$$. So the Artin map is the isomorphism to the class group that we want.

• Excellent explanation! A brief note: (2) works, without any real additional difficulty, for ramified class field theory - just use the "generalized Jacobian" parameterizing line bundles with a trivialization over a fixed divisor $N$. For (1), proving the ramified analogue is much messier. – Will Sawin Apr 6 at 15:40
• Thanks! A computation that I remember doing when I read Serre was working out the ray class fields of $\mathbb{P}^1$ with respect to the conductors $(0)+(\infty)$ and $2 (\infty)$ from the generalized Jacobian definition: It was really fun to see the equations $y^{p-1}=x$ and $z^p-z = x$ pop out of the abstraction. – David E Speyer Apr 6 at 16:36
• I never thought before about what happens to (1) in the ramified case. Thinking now: It looks like the fibers of Sym(X) to GeneralizedJacobian(X) are not projective spaces, but projective spaces with a locus deleted which, after extending the base field, looks like a bunch of hyperplanes. I imagine you have to do some sort of computation to show that, if the conductor is chosen large enough, the monodromy around those hyperplane vanishes? – David E Speyer Apr 6 at 17:18
• Yes, this is what's needed. I don't actually know anywhere that the computation is done in that language, but I believe it is possible – Will Sawin Apr 6 at 18:33

Ben Lim's answer reminded me of Mazur's article on the etale cohomology of number fields.

From the introduction:

"In the setting of etale cohomology, M. Artin and J.-L. Verdier have proved a duality theorem for constructible abelian sheaves over the scheme Spec D, where D is the ring of integers in a number field {see [AV]). This duality theorem contains within it Tate duality for finite Galois modules over local and global fields, and is indeed the natural extension of Tate duality to the context of such schemes. In what follows, I would like to poke at this result from various angles, and make what I hope to be enlightening remarks and computations, so as to convey in a concrete way a sense of the information contained in this theorem."

• There is a beautiful writeup by Brian Conrad and Alessndro Maria Masullo here: math.stanford.edu/~conrad/BSDseminar/Notes/L4.pdf – Squid with Black Bean Sauce Mar 28 at 16:24
• Dear Matt, I may be too naïve (and I surely don't yet understand even a small part of this) but isn't this (Artin-Verdier duality) precisely what I wanted? If I recall correctly, we can deduce all of class field theory from Tate's duality. Then if we can prove (a generalization) of Tate's duality using étale cohomology, this would be precisely what I want. – Gabriel Mar 28 at 17:17
• Dear Gabriel, Maybe ... . But I would guess Lang is more referring to geometric proofs of Class Field Theory in the function field case which rely on simple geometric properties of $\mathbb P^1$, which is something a bit different and more concrete than the sheaf theory of Artin--Verdier. – Matt E Mar 29 at 1:22
• @MattE indeed. I think I have a satisfactory answer for the function field case, but I am also interested in the number field case. What I am not yet sure is if we don't have to use class field theory to prove Artin-Verdier. – Gabriel Mar 29 at 8:15

I would like to point out a letter from A Grothendieck to L Breen

[pdf]

In which he explains some details about it. See for example page 45