The Brauer Group of The Moduli Stack of Elliptic Curves: Benjamin Antieau and Lennart Meier

The Brauer group of the moduli stack of elliptic curves
Benjamin Antieau∗ and Lennart Meier†

arXiv:1608.00851v3 [math.AG] 5 May 2020
Abstract
We compute the Brauer group of M1,1 , the moduli stack of elliptic curves, over Spec Z,
its localizations, finite fields of odd characteristic, and algebraically closed fields of
characteristic not 2. The methods involved include the use of the parameter space of
Legendre curves and the moduli stack M(2) of curves with full (naive) level 2 structure,
the study of the Leray–Serre spectral sequence in étale cohomology and the Leray
spectral sequence in fppf cohomology, the computation of the group cohomology of S3
in a certain integral representation, the classification of cubic Galois extensions of Q,
the computation of Hilbert symbols in the ramified case for the primes 2 and 3, and
finding p-adic elliptic curves with specified properties.
Key Words. Brauer groups, moduli of elliptic curves, level structures, Hilbert sym-
bols.
Mathematics Subject Classification 2010. Primary: 14F22, 14H52, 14K10.

Secondary: 11G05, 11G07.
Contents
1 Introduction 2
2 Brauer groups, cyclic algebras, and ramification 4
3 The low-dimensional Gm -cohomology of BCm 10
4 A presentation of the moduli stack of elliptic curves 13
5 Beginning of the computation 17
6 The p-primary torsion in Br(MZ[ 12 ] ) for primes p ≥ 5 21
7 The 3-primary torsion in Br(MZ[ 1 ] ) 22

6
8 The ramification of the 3-torsion 24
9 The 2-primary torsion in Br(MZ[ 12 ] ) 26
10 The Brauer group of M 30
11 The Brauer group of M over Fq with q odd 31

∗ Benjamin Antieau was supported by NSF Grants DMS-1461847 and DMS-1552766.
† Lennart Meier was supported by DFG SPP 1786.
1
1. Introduction 2
1 Introduction
Brauer groups of fields have been considered since the times of Brauer and Noether; later
Grothendieck generalized Brauer groups to the case of arbitrary schemes. Although both
the definition via Azumaya algebras and the cohomological definition generalize to arbitrary
Deligne–Mumford stacks, Brauer groups of stacks have so far mostly been neglected espe-
cially for stacks containing arithmetic information. Some exceptions are the use of Brauer
groups of root stacks, as in the work of Chan–Ingalls [CI05] on the minimal model program
for orders on surfaces, the work of Auel–Bernardara–Bolognesi [ABB14] on derived cate-
gories of families of quadrics, and the work of Lieblich [Lie11], who computed the Brauer
group Br(Bµn ) over a field and applied it to the period-index problem. In this paper, we
study the Brauer group Br(M) of the moduli stack of elliptic curves M = M1,1 .
The case of the Picard group has been considered before. Mumford showed in [Mum65]
that Pic(Mk ) = H1ét (Mk , Gm ) ∼
= Z/12 if k is a field of characteristic not dividing 6, and
that the Picard group is generated by the Hodge bundle λ. The bundle λ is characterized
by the property that u∗ λ ∼ = p∗ Ω1E/T when u : T → M classifies a family p : E → T
of elliptic curves. This calculation was extended by Fulton and Olsson who showed that
Pic(MS ) ∼ = Pic(A1S ) ⊕ Z/12 whenever S is a reduced scheme [FO10].
In contrast, an equally uniform description of Br(MS ) does not seem possible (even if
we assume that S is regular noetherian); both the result and the proofs depend much more
concretely on the arithmetic on S. The following is a sample of our results in ascending
order of difficulty. We view (5) as the main result of this paper.
Theorem 1.1. (1) Br(Mk ) = 0 if k is an algebraically closed field of characteristic not

2,1
(2) Br(Mk ) ∼
= Z/12 if k is a finite field of characteristic not 2,
L L
(3) Br(MQ ) ∼
= Br(Q) ⊕ p6≡3 mod 4 Z/4 ⊕ p≡3 mod 4 Z/2 ⊕ H1 (Q, C3 ),
where p runs over all primes and −1,
(4) Br(M 1 )
∼
= Br(Z[ 12 ]) ⊕ Z/2 ⊕ Z/4, and
Z[ 2 ]
(5) Br(M) = 0.
In all cases the non-trivial classes can be explicitly described via cyclic algebras (see
Lemma 7.2, Proposition 9.6 and Remark 9.8). In general, the p-primary torsion for p ≥ 5
is often easy to control via the following theorem.
Theorem 1.2. Let S be a regular noetherian scheme and p ≥ 5 prime. Assume that
1
S[ 2p ] = S 1 is dense in S and that MS → S has a section. Then, the natural map
Z[ 2p ]
′ ′
p Br (S) → p Br (MS ) is an isomorphism.
Here, Br′ (MS ) denotes the cohomological Brauer group, which agrees with Br(MS ) when-
ever S is affine and at least one prime is invertible on S.
Let us now explain how to compute the 2- and 3-primary torsion in the example of
Br(M 1 ). We use the S3 -Galois cover M(2) → M 1 , where M(2) is the moduli stack of
Z[ 2 ] Z[ 2 ]
elliptic curves with full (naive) level 2-structure. The Leray–Serre spectral sequence reduces
the problem of computing Br(M 1 ) to understanding the low-degree Gm -cohomology of
Z[ 2 ]
M(2) together with the action of S3 on these cohomology groups and to the computation of
differentials.
1 Minseon Shin has proved in [Shi19] that when k is algebraically closed of characteristic 2 one has
Br(Mk ) ∼
= Z/2.
3 1. Introduction
To understand the groups themselves, it is sufficient to use the fact that M(2) ∼ = BC2,X ,
the classifying stack of cyclic C2 -covers over X, where X is the parameter space of Legendre
curves. Explicitly, X is the arithmetic surface
X = P1 1 − {0, 1, ∞} = Spec Z[ 21 , t±1 , (t − 1)−1 ],

Z[ 2 ]
and the universal Legendre curve over X is defined by the equation
y 2 = x(x − 1)(x − t).
The Brauer group of BC2,X can be described using a Leray–Serre spectral sequence as well,
but this description is not S3 -equivariant, which causes some complications, and we have
to use the S3 -equivariant map from M(2) to X, its coarse moduli space, to get full control.
Knowledge about the Brauer group of X leads for the 3-primary torsion to the following
conclusion.
Theorem 1.3. Let S be a regular noetherian scheme. If 6 is a unit on S, then there is an
exact sequence
0 → 3 Br′ (S) → 3 Br′ (MS ) → H1 (S, C3 ) → 0,
which is non-canonically split.
There is a unique cubic Galois extension of Q which is ramified at most at (2) and
(3), namely Q(ζ9 + ζ 9 ). This shows that the cokernel of Br(Z[ 61 ]) ֒→ Br(M 1 ) is Z/3.
Z[ 6 ]
The proof that this extra class does not extend to MZ[ 1 ] , which is similar to the strategy
2
discussed below for the 2-torsion, uses the computation of cubic Hilbert symbols at the
prime 3. Putting these ingredients together, we conclude that Br(M 1 ) is a 2-group,
Z[ 2 ]
and further computations first over Z[ 12 , i] and then over Z[ 21 ] let us deduce the structure in
Theorem 1.1(4). The corresponding general results on 2-torsion are contained in Proposition
9.3 and Theorem 9.1. They are somewhat more complicated to state so we omit them from
this introduction.
To show that Br(M) = 0, we need an extra argument since all our arguments using the
Leray–Serre spectral sequence presuppose at least that 2 is inverted. Note first that the
map Br(M) → Br(MZ[ 1 ] ) is an injection. Thus, we have only to show that the non-zero
2
classes in Br(MZ[ 1 ] ) do not extend to M. Our general method is the following. For each
2
non-zero class α in the Brauer group of M 1 , we exhibit an elliptic curve over Spec Z2 such
Z[ 2 ]
that the restriction of α to Spec Q2 is non-zero. Such an elliptic curve defines a morphism
Spec Z2 → M and we obtain a commutative diagram
Br(Q2 ) o Br(M 1 ) (1.4)

O Z[ ]
O 2
Br(Z2 ) o Br(M)
Together with the fact that Br(Z2 ) = 0, this diagram implies that the class α cannot come
from Br(M). This argument requires us to understand explicit generators for Br(MZ[ 1 ] )
2
and the computation of Hilbert symbols again.
We remark that the computation of Br(M) – while important in algebraic geometry and
in the arithmetic of elliptic curves – was nevertheless originally motivated by considerations
in chromatic homotopy theory, especially in the possibility in constructing twisted forms
of the spectrum TMF of topological modular forms. The Picard group computations of
Mumford and Fulton–Olsson are primary inputs into the computation of the Picard group
of TMF due to Mathew and Stojanoska [MS16].
2. Brauer groups, cyclic algebras, and ramification 4
Conventions. We will have occasion to use Zariski, étale, and fppf cohomology of schemes
and Deligne-Mumford stacks. We will denote these by HiZar (X, F ), Hi (X, F ), and Hipl (X, F )
when F is an appropriate sheaf on X. Note in particular that without other adornment,
Hi (X, F ) or Hi (R, F ) (when X = Spec R) always denotes étale cohomology. If G is a group
and F is a G-module, we let Hi (G, F ) denote the group cohomology.
For all stacks X appearing in this paper, we will have Br′ (X) ∼= H2 (X, Gm ). Thus, we
will use the two groups interchangeably. When working over a general base S, we will
typically state our results in terms of Br′ (S) or Br′ (MS ). However, when working over an
affine scheme, such as S = Spec Z[ 12 ], we will write Br(S) or Br(MS ). There should be no
confusion as in all of these cases we will have Br(S) = Br′ (S) and Br(MS ) = Br′ (MS ) and
so on by [dJ].
For an abelian group A and an integer n, we will denote by n A the subgroup of n-primary
torsion elements: n A = {x ∈ A : nk x = 0 for some k ≥ 1}.
Acknowledgments. We thank the Hausdorff Research Institute for Mathematics, UIC,

and Universität Bonn for hosting one or the other author in 2015 and 2016 while this paper
was being written. Furthermore, we thank Asher Auel, Akhil Mathew, and Vesna Stojanoska
for their ideas at the beginning of this project and Peter Scholze and Yichao Tian for the
suggestion to use fppf-cohomology and the idea for the splitting in Proposition 2.9. Finally,
we thank the anonymous referees for their helpful suggestions.
2 Brauer groups, cyclic algebras, and ramification

We review here some basic facts about the Brauer group, with special attention to providing
references for those facts in the generality of Deligne-Mumford stacks. For more details about
the Brauer group in general, see [Gro68a].
Any Deligne-Mumford stack has an associated étale topos, and we can therefore consider
étale sheaves and étale cohomology [LMB00].
Definition 2.1. If X is a quasi-compact and quasi-separated Deligne-Mumford stack, the
cohomological Brauer group of X is defined to be Br′ (X) = H2 (X, Gm )tors , the torsion
subgroup of H2 (X, Gm ).
Because of its definition as the torsion in a cohomology group, the cohomological Brauer
group is amenable to computation via Leray–Serre spectral sequences, long exact sequences,
and so on, as we will see in the next sections. However, our main interest is in the Brauer
group of Deligne-Mumford stacks.
Definition 2.2. An Azumaya algebra over a Deligne-Mumford stack X is a sheaf of
quasi-coherent OX -algebras A such that A is étale-locally on X isomorphic to Mn (OX ), the
sheaf of n × n-matrices over OX , for some n ≥ 1.
In particular, an Azumaya algebra A is a locally free OX -module, and the degree
n appearing in the definition is a locally constant function. If the degree n is in fact
constant, then A corresponds to a unique PGLn -torsor on X because the group of k-
algebra automorphisms of Mn (k) is isomorphic to PGLn (k) for fields k. The exact sequence
1 → Gm → GLn → PGLn → 1 gives a boundary map
δ : H1 (X, PGLn ) → H2 (X, Gm ).
For an Azumaya algebra A of degree n, we write [A] for the class δ(A) in H2 (X, Gm ).
In general, when X has multiple connected components, its invariant [A] ∈ H2 (X, Gm ) is
computed on each component.
Example 2.3. 1. If E is a vector bundle on X of rank n > 0, then A = End(E), the
sheaf of endomorphisms of E, is an Azumaya algebra on X. Indeed, in this case, A
5 2. Brauer groups, cyclic algebras, and ramification
is even Zariski-locally equivalent to Mn (OX ). The class of A in H1 (X, PGLn ) is the

image of E via H1 (X, GLn ) → H1 (X, PGLn ), so the long exact sequence in nonabelian
cohomology implies that [A] = 0 in H2 (X, Gm ).
2. If A and B are Azumaya algebras on X, then A ⊗OX B is an Azumaya algebra.
Definition 2.4. Two Azumaya algebras A and B are Brauer equivalent if there are
vector bundles E and F on X such that A ⊗OX End(E) ∼ = B ⊗OX End(F ). The Brauer
group Br(X) of a Deligne-Mumford stack X is the multiplicative monoid of isomorphism
classes of Azumaya algebras under tensor product modulo Brauer equivalence.
In terms of Azumaya algebras, addition is given by the tensor product [A] + [B] =
[A ⊗OX B], and −[A] = [Aop ]. Here are the basic structural facts we will use about the
Brauer group.
Proposition 2.5. (i) The Brauer group of a quasi-compact and quasi-separated Deligne-
Mumford stack X is the subgroup of Br′ (X) generated by [A] for A an Azumaya algebra
on X.
(ii) The Brauer group Br(X) is a torsion group for any quasi-compact and quasi-separated
Deligne-Mumford stack X.
(iii) If X is a regular and noetherian Deligne-Mumford stack, then H2 (X, Gm ) is torsion,

so in particular Br′ (X) ∼
= H2 (X, Gm ).
(iv) If X is a regular and noetherian Deligne-Mumford stack, and if U ⊆ X is a dense
open subset, then the restriction map H2 (X, Gm ) → H2 (U, Gm ) is injective.
(v) If X is a scheme with an ample line bundle, then Br(X) = Br′ (X).
(vi) If X is a regular and noetherian scheme with p invertible on X, then the morphism
i i 1
p H (X, Gm ) → p H (AX , Gm ) on p-primary torsion is an isomorphism for all i ≥ 0.
Proof. See [Gro68a, Section 2] for points (i) and (ii). The proof of (iv) is analogous to that
of [Lie08, Lemma 3.1.3.3], using an analogue of [LMB00, Proposition 15.4] to generalize
[Lie08, Lemma 3.1.1.9] to algebraic stacks. See also [ABB14, Proposition 1.26]. For (v), see
de Jong [dJ].
For (iii), see for instance [Gro68b, Proposition 1.4] in the case of schemes. We must
generalize it to the case of a regular noetherian Deligne-Mumford stack X. We can assume
that X is connected and hence irreducible as X is normal. Pick U ⊆ X a dense open such
that U admits a finite étale map V → U of degree n where V is a scheme. The composition
H2 (U, Gm ) → H2 (V, Gm ) → H2 (U, Gm ) of restriction and transfer is multiplication by n.
Since H2 (V, Gm ) is torsion, this implies that H2 (U, Gm ) is torsion. By (iv), H2 (X, Gm ) is
torsion as well.
Finally, we have to prove (vi). For i ≤ 1, the A1 -invariance is even true before taking
p-power torsion (see e.g. [Har77, II.6.6 and II.6.15]). Because p is invertible on S, the maps
Hi (X, µpm ) → Hn (A1X , µpm ) are isomorphisms for all i and m. This is the A1 -invariance of
étale cohomology and a proof can be found in [Mil80, Corollary VI.4.20]. The short exact
sequence
pm
1 → µpm → Gm −−→ Gm → 1
induces short exact sequences
0 / Hi−1 (X, Gm )/pm / Hi (X, µpm ) / Hi (X, Gm )[pm ] /0
∼
=

0 / Hi−1 (A1 , Gm )/pm / Hi (A1 , µpm ) / Hi (A1 , Gm )[pm ] /0
X X X
Inductively, using the A1 -invariance of Hi−1 (X, Gm ) if i ≤ 2 or of p Hi−1 (X, Gm ) if i > 2

as well as the fact that Hi−1 (X, Gm ) is torsion for i ≥ 3 since X is regular and noetherian
([Gro68b, Proposition 1.4]), we see by the five lemma that Hi (X, Gm )[pm ] → Hi (A1X , Gm )[pm ]
is an isomorphism for all i and m and hence we also get an isomorphism p Hi (X, Gm ) →
i 1
p H (AX , Gm ).
Remark 2.6. At a couple points, we use another important fact, due to Gabber, which says
that if p : Y → X is a surjective finite locally free map, if α ∈ Br′ (X), and if p∗ α ∈ Br(Y ),
then α ∈ Br(X). This is already proved in Gabber [Gab81, Chapter II, Lemma 4] for locally
ringed topoi with strict hensel local rings, so we need to add nothing further in our setting.
By far the most important class of Azumaya algebras arising in arithmetic applications
is the class of cyclic algebras. For a treatment over fields, see [GS06, Section 2.5]. These
algebras give a concrete realization of the cup product in fppf cohomology
H1pl (X, Cn ) × H1pl (X, µn ) → H2pl (X, µn ) → H2pl (X, Gm )
for an algebraic stack X. Given that Cn and Gm are smooth and that the image of such a cup
product is torsion, we can re-write this as H1 (X, Cn ) × H1pl (X, µn ) → H2pl (X, µn ) → Br′ (X).
Given χ ∈ H1 (X, Cn ) and u ∈ H1pl (X, µn ), we write [(χ, u)n ] or [(χ, u)] for the image of the
cup product in Br′ (X).
Fix an algebraic stack X. Let p(χ) : Y → X be the cyclic Galois cover defined by χ ∈
H1 (X, Cn ). Then, p(χ)∗ OY is a locally free OX -algebra of finite rank which comes equipped
with a canonical Cn -action. There is a natural isomorphism SpecX (p(χ)∗ OY ) ∼ = Y → X.
The group H1pl (X, µn ) fits into a short exact sequence
0 → Gm (X)/n → H1pl (X, µn ) → Pic(X)[n] → 0. (2.7)
It is helpful to have a more concrete description of H1pl (X, µn ), which will also show that the
exact sequence (2.7) is non-canonically split. Let H(X, n) be the abelian group of equivalence
classes of pairs (L, s) where L ∈ Pic(X)[n] and s is a choice of trivialization s : OX → L⊗n .
Two pairs (L, s) and (M, t) are equivalent if there is an isomorphism g : L → M and a unit
v ∈ Gm (X) such that g(s) = v n t. The group structure is given by tensor product of line
bundles and of trivializations. The following construction is part of Kummer theory and is
well-known in the scheme case (see e.g. [Mil80, p. 125], which also shows part of Proposition
2.9 below).
Construction 2.8. Let X be an algebraic stack and fix a class [u] ∈ H(X, n) with √ u =
⊗n
(L,
L s) for a line bundle L on
√ X with a trivialization s : O X → L . Define
√ O X ( n
u) as
⊗i
i∈Z L /(s − 1) and
√ X( n
u) → X as the affine morphism Spec O
X X ( n
u) → X. It is
easy to see that X( n u) → X is an fppf µn -torsor and that this construction defines a
group homomorphism H(X, n) → H1pl (X, µn ). Indeed, if X = Spec A and L is trivial, then
√ √ √
X( n u) ∼= Spec A( s−1 ) ∼
n
= Spec A( n s).
Proposition 2.9. The map H(X, n) → H1pl (X, µn ) is an isomorphism. In particular, there
is a non-canonical splitting
H1pl (X, µn ) ∼
= Gm (X)/n ⊕ Pic(X)[n].
√
Proof. We claim that the line bundle associated with the µn -torsor X( n u) → X (via the
map H1pl (X, µn ) → H1 (X, Gm ) induced by the inclusion µn → Gm ) is exactly L. Indeed, the
√ L ⊗i

obvious map from X( n u) to SpecX i∈Z L is equivariant along the inclusion µn → Gm
and the target is the Gm -torsor associated with L. Thus, the composition H(X, n) →
H1pl (X, µn ) → Pic(X)[n] is surjective. By (2.7), to prove that H(X, n) → H1pl (X, µn ) is an
isomorphism, it suffices to prove that the induced map from the kernel of this map to
Gm (X)/n is an isomorphism. But, this follows immediately from the definition of H(X, n).
It remains to construct the splitting. By Prüfer’s theorem [Fuc70, Theorem 17.2],

Pic(X)[n] is a direct sum of cyclic groups. Thus, we have only to show that for every divisor k
of n and each k-torsion element [L] in Pic(X), there exists a pre-image in H1pl (X, µn ) that is k-
torsion. This pre-image can be constructed as follows: choose
√ a trivialization s : OX → L⊗k
and take the µn -torsor associated with the µk -torsor X( k v) → X with v = (L, s).
eχ,u be the coproduct

Now, given χ and u as above, we let A
 
a √
p(χ)∗ OY OX ( n u)
OX
in the category of sheaves of quasi-coherent (associative and unital) OX -algebras. Finally,

we let
Aχ,u = Aeχ,u /(ab − bag(χ) )
be the quotient of A eχ,u by the two-sided ideal generated by terms ab − bag(χ) where a is a
√
local section of χ∗ OY , b is a local section of OX ( n u), and ag(χ) denotes the action of g(χ)
on χ∗ OY for g(χ) a fixed generator of the action of Cn on p(χ)∗ OY .
Lemma 2.10. Given an algebraic stack X and classes χ ∈ H1 (X, Cn ) and u ∈ H1pl (X, µn ),
the algebra Aχ,u is an Azumaya algebra on X.
Proof. It suffices to check this fppf-locally, and in particular we can assume that in fact X
is a scheme X, that L ∼ = OX , and that χ classifies a Cn -Galois cover p(χ) : Y → X. In
this case, we can write the OX -algebra p(χ)∗ OY Zariski locally as a quotient OX [x]/f (x) for
some monic polynomial f (x) of degree n. Then, locally, we have that
Aχ,u ∼
= OX hx, yi/(f (x), y n − u, xy − yxg(χ) ),
a quotient of the free algebra over OX on generators x and y. Note that the sections xi y j
for 0 ≤ i, j ≤ n − 1 form a basis of Aχ,u as an OX -module and in particular that Aχ,u
is (locally) a free OX -module. Examining the fibers of Aχ,u over X, we obtain the usual
definition of a cyclic algebra given in [GS06, Proposition 2.5.2]. So, Aχ,u is locally free with
central simple fibers and [Gro68a, Théorème 5.1] implies that Aχ,u is Azumaya.
The following proposition is well-known, but we do not know an exact reference. However,
in the case of quaternion algebras, it is given in [PS92, Lemma 8].
Proposition 2.11. Let X be a regular noetherian scheme and suppose that χ ∈ H1 (X, Cn )
and u ∈ H1pl (X, µn ) are fixed classes. In the notation above, we have [(χ, u)n ] = [Aχ,u ] in
Br′ (X). In particular [(χ, u)n ] ∈ Br(X).
Proof. We can assume that X is connected. As Br′ (X) → Br′ (K) is injective (see [Mil80,
IV.2.6] or Proposition 2.5(iv)), it is enough to check this on a generic point Spec K of X.
By definition and the previous example, Aχ,u is a standard cyclic algebra over K as defined
in [GS06, Chapter 2]. They check in [GS06, Proposition 4.7.3] that Aχ,u does indeed have
Brauer class given by the cup product. See also the remark at the beginning of the proof
of [GS06, Proposition 4.7.1].
Remark 2.12. The reader may notice that Aχ,u is defined in complete generality, but that
we only prove the equality [(χ, u)n ] = [Aχ,u ] for regular noetherian schemes. In fact, this
equality extends to arbitrary algebraic stacks, but a different argument is necessary. It is
given at the end of Section 3.
We will abuse notation and write (χ, u)n or even just (χ, u) for Aχ,u . This is called a
cyclic algebra. If there is a primitive nth root of unity ω ∈ µn (X) and the cyclic Galois
cover Y → X is obtained by adjoining an nth root of an element a ∈ Gm (X), we write
(a, u) = (a, u)ω for the corresponding cyclic algebra, where we need the choice of ω to fix
an isomorphism H1 (X, Cn ) ∼ = H1 (X, µn ). For n = 2, we obtain the classical notion of a
quaternion algebra.
For us, the key point about the cyclic algebra is that it allows us to compute the ramifi-
cation of a Brauer class explicitly. Before explaining this, we mention that by the Gabber–
Česnavičius purity theorem the Brauer group of a regular noetherian scheme X is insensitive
to throwing away high codimension subschemes.
Proposition 2.13 (Purity [Gab81, Chapter I], [Ces17]). Let X be a regular noetherian
scheme. If U ⊆ X is a dense open subscheme with complement of codimension at least 2,
then the restriction map Br′ (X) → Br′ (U ) is an isomorphism.
Let X be a regular noetherian scheme and let η be the scheme of generic points in X.
The purity theorem reduces the problem of computing Br(X) from Br(η) to the problem of
extending Brauer classes α ∈ Br(η) over divisors in X. This is controlled by ramification
theory. The following proposition is basically well-known, but we include a proof for the
reader’s convenience.
Proposition 2.14. Let X be a regular noetherian scheme, i : D ⊆ X a Cartier divisor that
is regular with complement U , and n an integer. There is an exact sequence
ram
0 → n Br′ (X) → n Br′ (U ) −−−D
→ n H3D (X, Gm ) → n H3 (X, Gm ) → n H3 (U, Gm ). (2.15)
If n is prime to the residue characteristics of X, we have n H3D (X, Gm ) ∼
= H1 (D, n Q/Z).
Proof. By [Gro68c, 6.1] or [Mil80, III.1.25] there is a long exact sequence
H2 (X, Gm ) → H2 (U, Gm ) → H3D (X, Gm ) → H3 (X, Gm ) → H3 (U, Gm ).
By [Gro68b, Propsition 1.4] all occurring groups are torsion so that the sequence is still
exact after taking n-primary torsion. Furthermore, by Proposition 2.5 the first map is an
injection.
We may assume that n = p is a prime, in which case p Q/Z ∼ = Qp /Zp . We have to show
that p H3D (X, Gm ) ∼
= H1 (D, Qp /Zp ). By either the relative cohomological purity theorem
of Artin [SGA4.3, Théorème XVI.3.7 and 3.8] (when both X and D are smooth over some
common base scheme S) or the absolute cohomological purity theorem of Gabber [Fuj02,
t
Theorem 2.1], we have the following identifications of local cohomology sheaves: HD (µpν ) =
2 ∼ ν
0 for t 6= 2 and HD (µp ) = i∗ Z/p (−1). It follows from the long exact sequence of local
ν
pν
cohomology sheaves associated to the exact sequence 1 → µpν → Gm −→ Gm → 1 that p
t
acts invertibly on HD (Gm ) for t 6= 1, 2. Moreover, since X is regular and noetherian, for
every open V ⊂ X, the map Pic(V ) → Pic(U ∩ V ) is surjective with kernel (i∗ Z)(V ) by
[Har77, II.6.5] and Br′ (V ) → Br′ (U ∩V ) is injective; thus HD
2 1
(Gm ) = 0 and HD (Gm ) ∼
= i∗ Z.
3
Therefore, the only contribution to p-primary torsion in HD (X, Gm ) in the local to global
spectral sequence
Hs (X, HDt
(Gm )) ⇒ Hs+t
D (X, Gm )
is p H2 (X, i∗ Z). We obtain
3
p HD (X, Gm )
∼
= p H2 (X, HD
1
(Gm )) ∼
= p H2 (D, Z) ∼
= H1 (D, Qp /Zp )
as desired, where the last isomorphism holds because Hi (D, Q) = 0 for i > 0 since D is
normal (see for example [Den88, 2.1]).
Note that in all cases where we use Proposition 2.14, the easier relative cohomological
purity theorem of Artin is applicable, so that in the end our paper does not rely on the
more difficult results of Gabber and Česnavičius.
We will need to know a special case of the ramification map ramD : Br(U ) → H1 (D, Q/Z).
Proposition 2.16. Let R be a discrete valuation ring with fraction field K and residue field
k. Set X = Spec R, U = Spec K, and x = Spec k. Let (χ, π)n be a cyclic algebra over K,
where χ is a degree n cyclic character of K, π is a uniformizing parameter of R (viewed as
an element of Gm (K)/n), and n is prime to the characteristic of k. Finally, let L/K be
the cyclic Galois extension defined by χ. If the integral closure S of R in L is a discrete
valuation ring with uniformizing parameter πS , then ram(π) (χ, π) is the class of the cyclic
extension S/(πS ) over k.
Proof. See [Sal99, Lemma 10.2].
Finally, we discuss cyclic algebras over local fields and some implications for global
calculations. Let K be a local field containing a primitive nth root of unity ω. Then there
is a pairing
−, −
: Gm (K)/n × Gm (K)/n → µn (K),
p
called the Hilbert symbol (where p stands for the maximal ideal of the ring of integers of
K). Our standard reference for this
pairing is [Neu99, Section V.3]. If p is generated by an
element π, we will also write a,b
π . We will use Hilbert symbols to check whether explicitly
defined cyclic algebras are zero in the Brauer group.
Proposition
2.17. For a, b ∈ K × , the cyclic algebra (a, b)ω is trivial in Br(K) if and only
a,b
if p = 1.

Proof. By Proposition V.3.2 of [Neu99] the Hilbert symbol a,b equals 1 if and only if a is
√
n
p
a norm from the extension K( b)|K. By [GS06, Corollary 4.7.7], this happens if and only
if (a, b)ω splits, i.e. defines the trivial class in Br(K).
More generally, local class field theory calculates Br(η) when η = Spec K where K is
a (non-archimedean) local field. Let X = Spec R and x = Spec k, where R is the ring of
integers in K and k is the residue field of R. As H3 (X, Gm ) ∼
= H3 (x, Gm ) = 0 (for instance
by [Gro68c, Théorème 1.1]), we find from [Gro68c, Corollaire 2.2] that there is an exact
sequence
0 → Br(X) → Br(η) → H1 (x, Q/Z) → 0.
The idea is similar to that of Proposition 2.14, but here the proof is easier as 0 → Gm →
j∗ Gm → i∗ Z → 0 is exact where j : η → X and i : x → X. Since K is local, k is finite,
so that H1 (x, Q/Z) ∼ = Q/Z. However, since R is Henselian, Br(X) = Br(x) (see [Gro68a,
Corollaire 6.2]), and Br(x) = 0 by a theorem of Wedderburn (see [Gro68c, Proposition 1.5]).
Now, let K be a number field, and let R be a localization of the ring of integers of K.
Set η = Spec K and X = Spec R. In this case, by [Gro68c, Proposition 2.1], there is an
exact sequence M
0 → Br(X) → Br(η) → Br(Spec Kp ),
p∈X (1)
(1)
where X denotes the set of codimension 1 points of X. This exact sequence is compatible
with (2.15) and with the exact sequence
M
0 → Br(η) → Br(Spec Kp ) → Q/Z → 0 (2.18)
p
of class field theory (see [NSW00, Theorem 8.1.17]). The sum ranges over the finite and the
infinite places of K, and the map Br(Spec Kp ) → Q/Z is the isomorphism described above
when p is a finite place, the natural inclusion Z/2 → Q/Z when Kp ∼ = R, and the natural
map 0 → Q/Z when Kp ∼ = C. Using these sequences, we can compute the Brauer group of
X.
The two fundamental observations we need about (2.18) are that a class α ∈ Br(η)
is ramified at no fewer than 2 places and that if K is purely imaginary, then α ∈ Br(η) is
ramified at no fewer than 2 finite places. The reader can easily verify the following examples.
3. The low-dimensional Gm -cohomology of BCm 10
Example 2.19. (1) Br(Z) = 0.

(2) Br(Z[ p1 ]) ∼
= Z/2.
1
(3) Br(Z[ pq ]) ∼
= Z/2 ⊕ Q/Z.
(4) Br(Z[ p1 , ζp ]) = 0.
We will use these computations and those like them throughout the paper, often without
comment.
3 The low-dimensional Gm -cohomology of BCm

Let S be a scheme. Write Cn,S for the constant étale group scheme on the cyclic group Cn
of order n ≥ 2 over S. We will often suppress the base in the notation and simply write
Cn when the base is clear from context. The purpose of this section is to make a basic
computation of the Gm -cohomology of the Deligne-Mumford stack BCn = BCn,S . In fact,
we are only interested in the cases n = 2 and n = 4, but the general case is no more difficult.
The first tool for our computations of the étale cohomology of an étale sheaf F is the
convergent Leray–Serre spectral sequence. If π : Y → X is a G-Galois cover where X and
Y are Deligne-Mumford stacks, then this spectral sequence has the form
Ep,q p q ∗
2 = H (G, H (Y, π F)) ⇒ H
p+q
(X, F),
with differentials dr of bidegree (r, 1 − r).

We will use the spectral sequence in this section for the Cn -Galois cover π : S → BCn ,
where Cn acts trivially on S. In this case it is of the form
Ep,q p q
2 = H (Cn , H (S, Gm )) ⇒ H
p+q
(BCn,S , Gm )
and Figure 1 displays the low-degree part of its E2 -page. The fact that Cn acts trivially on
the cohomology of S implies that the left-most column is simply the Gm -cohomology of S.
For F a constant Cn -module, we use the standard isomorphisms Hi (Cn , F ) ∼ = F [n] when
i > 0 odd, and Hi (Cn , F ) ∼
= F/n when i > 0 is even. We are only interested in Hi (BCn , Gm )
for 0 ≤ i ≤ 2.
H2 (S, Gm )
Pic(S) Pic(S)[n] ❩❩❩❩❩❩Pic(S)/n
❩❩❩❩❩❩❩❩
❩❩❩❩❩❩-
Gm (S) µn (S) Gm (S)/n µn (S)
Figure 1: The E2 -page of the Leray–Serre spectral sequence computing Hi (BCn , Gm ).
Recall Grothendieck’s theorem that the natural morphism Hi (X, G) → Hipl (X, G) is an
isomorphism for i ≥ 0 when G is a smooth group scheme on X (such as Cn or Gm ). See
[Gro68c, Section 5, Thèoréme 11.7]. Note that this implies the agreement of étale and fppf
cohomology on a Deligne–Mumford stack. For example, Grothendieck’s theorem implies
that the morphism from the Leray–Serre spectral sequence above to the analogous Leray–
Serre spectral sequence
Ep,q p q p+q
2 = H (Cn , Hpl (S, Gm )) ⇒ Hpl (BCn , Gm )
for the fppf cohomology is an isomorphism; thus the comparison map
Hi (BCn , Gm ) → Hipl (BCn , Gm )

11 3. The low-dimensional Gm -cohomology of BCm
is also an isomorphism. For Gm -coefficients, we will thus not distinguish between étale and
fppf cohomology in what follows.
We use these observations to compute the Picard and Brauer groups of BCn,S via a
Leray spectral sequence. The idea is borrowed from [Lie11, Section 4.1]. Consider the map
to the coarse moduli space c : BCn,S → S. We claim that
R0pl c∗ Gm = Gm
R1pl c∗ Gm = µn
R2pl c∗ Gm = 0.
Indeed, Ripl c∗ Gm is the fppf-sheafification of U 7→ Hi (BCn,U , Gm ) and in the Leray–Serre

spectral sequences all classes in Hp (Cn , Hq (U, Gm )) for q > 0 are killed by some fppf cover
of U . Furthermore every unit has an n-th root fppf-locally so that the fppf-sheafification of
the presheaf Gm /n vanishes. This implies the claim.
It follows that the fppf-Leray spectral sequence
Ep,q p q p+q
2 = Hpl (S, Rpl c∗ Gm ) ⇒ Hpl (BCn,S , Gm ) (3.1)
for c takes the form given in Figure 2 in low degrees. As π ∗ c∗ = id, we see that the edge
0
µn (S) ❩❩❩❩❩❩❩H❩1pl (S, µn ) ❬❬❬❬❬❬❬
❩❩❩❩❩❩❩❩ ❬
❩❩❩❩❩- ❬❬❬❬❬❬❬❬❬❬❬❬❬-
Gm (S) Pic(S) H2 (S, Gm ) H3 (S, Gm )
Figure 2: The E2 -page of the Leray spectral sequence (3.1) computing Hi (BCn , Gm ).
homomorphisms Hi (S, Gm ) → Hi (BCn,S , Gm ) from the bottom line in the Leray spectral
sequence are all split injections. In particular, the displayed differentials d0,1
2 and d1,1
2 are
zero.
Proposition 3.2. There is an isomorphism
Gm (BCn,S ) ∼
= Gm (S)
and short exact sequences

c∗
0 → Pic(S) −→ Pic(BCn,S ) → µn (S) → 0
and
c∗ r
0 → H2 (S, Gm ) −→ H2 (BCn,S , Gm ) −
→ H1pl (S, µn ) → 0,
c∗ r
0 → Br′ (S) −→ Br′ (BCn,S ) −
→ H1pl (S, µn ) → 0,
c∗ r
→ H1pl (S, µn ) → 0,
0 → Br(S) −→ Br(BCn,S ) −
which are split. The isomorphism and the short exact sequences are functorial in BCn,S
(i.e. endomorphisms of BCn,S induces endomorphisms of exact sequences in a functorial
manner), but the splittings are only functorial in S.
Proof. By the discussion above, the Leray spectral sequence proves everything except for
the split exactness of the last two sequences. For the sequence involving the cohomological
Brauer group, we just apply the torsion subgroup functor to the split exact sequence involv-
ing H2 (−, Gm ). By Remark 2.6 we furthermore see that a = π ∗ c∗ a ∈ H2 (S, Gm ) is in Br(S)
if and only if c∗ a ∈ Br(BCn,S ), implying split exactness for the last exact sequence.
3. The low-dimensional Gm -cohomology of BCm 12
Later on we will need not only the computation of the Brauer group of BCn , but also a
description of the classes coming from the inclusion Gm (S)/n ֒→ Br(BCn ), which is either
defined via the Leray–Serre spectral sequence or using the splitting in Proposition 3.2 (the
proof of the following lemma will, in particular, show that these two maps differ at most by
a unit). These classes are described via the classical cyclic algebra construction from the
previous section.
Lemma 3.3. Let X be an algebraic stack and n a positive integer. Let σ ∈ H1 (BCn,X , Cn )
be the class of the universal Cn -torsor X → BCn,X . Then there is an integer k prime to n
(that only depends on n) such that the map
s : H1pl (X, µn ) → H2 (BCn,X , Gm )
defined by s(u) = k[(σ, u)n ] is a section to the map r from Proposition 3.2.
Proof. It suffices to consider the universal case of X = Bpl µn over Spec Z, the stack classify-
ing fppf µn torsors. Note that Bpl µn is indeed a stack by [TS, Tag 04UR] and is an algebraic
stack by [TS, Tag 06DC] with fppf atlas Spec Z → Bpl µn . Let d : Bpl µn → Spec Z denote
the structure map, and let Rqpl d∗ µn denote the derived functors of the push-forward in the
= µn and we claim that R1pl d∗ µn ∼
fppf topos. Then, it is easy to see that R0pl d∗ µn ∼ = Cn . To
see the latter isomorphism, consider the natural transformations
Homgp 1 1
Spec R (µn,R , Gm,R ) → H (Bpl µn,R , Gm )[n] ← Hpl (Bpl µn,R , µn ) (3.4)
of presheaves on affine schemes over Z; the first map sends a homomorphism f : µn,R →
Gm,R to the image of the canonical class H1 (Bpl µn,R , µn ) under f∗ and the second map
is part of the Kummer sequence. The leftmost term is a sheaf and it is a standard fact
that it is represented by the constant étale group scheme Cn . See [CSS97, Section V.2.10]
for example. The fppf-sheafification of the rightmost term is R1pl d∗ µn . To see that the
induced map of sheaves are isomorphisms, it is sufficient to check on stalks in the fppf
topology [GK15, Remark 1.8, Theorem 2.3] and in particular if R is a Henselian local ring
with algebraically closed residue field [GK15, Lemma 3.3]. If R is such a local ring, then
Gm (R)/n = 0 so that H1pl (Bpl µn,R , µn ) ∼
= H1pl (Bpl µn,R , Gm )[n]. Using that Pic(R) = 0, the
Leray–Serre spectral sequence for the cover Spec R → Bpl µn,R shows that
H1pl (Bpl µn,R , Gm ) ∼

= H1group (µn,R , Gm,R ) ∼
gp
= HomSpec R (µn,R , Gm,R ),
where H1group (µn,R , Gm,R ) is the first cohomology of the cobar complex
Gm (S) → Gm (µn,R ) → Gm (µn,R ×Spec R µn,R ) → · · ·
with differentials as in the usual definition of group cohomology. This shows that the
morphisms in (3.4) are isomorphisms on fppf-stalks and thus that R1pl d∗ µn ∼
= Cn .
Now, the fppf-Leray spectral sequence for d : Bpl µn → Spec Z yields an exact sequence
0 → H1pl (Spec Z, µn ) → H1pl (Bpl µn , µn ) → H0 (Spec Z, Cn ) → 0. (3.5)
The right hand term is isomorphic to Z/n, and the sequence is split by applying the pullback
map along Spec Z → Bpl µn . We denote by τ ∈ H1pl (Bpl µn , µn ) the class of the universal
µn -torsor over Bpl µn . This is (exactly) of order n and pulls back to zero on Spec Z.
Consider c : BCn,Bpl µn → Bpl µn and the class α = [(σ, c∗ τ )n ]. The class of α has
order (exactly) n as there are cyclic algebras of order n over fields, for example by [GS06,
Lemma 5.5.3]. As π ∗ α = 0 for π : Bpl µn → BCn,Bpl µn the projection, it follows from
Proposition 3.2 that r(α) in H1pl (Bpl µn , µn ) has order n as well. On the other hand, r(α)
pulls back to zero over Spec Z so it is a non-zero multiple of τ (using the split-exact sequence
(3.5)). Thus, r(α) = mτ for some m prime to n. This completes the proof if we set k to be
a number such that km ≡ 1 mod n.
13 4. A presentation of the moduli stack of elliptic curves
Corollary 3.6. Suppose that χ : X → Y is a Cn -torsor for some positive integer n. Let
u ∈ Gm (Y )/n be the class of a unit, and write αu for the corresponding class in Br′ (Y )
(defined via the Leray–Serre spectral sequence). Then we have αu = k[(χ, u)] in Br′ (Y ),
where k is some number prime to n which only depends on n.
We do not know the value of k in the corollary. Perhaps it is always ±1, as is the case
in similar computations, such as the result of Lichtenbaum (see [GS06, Theorem 5.4.10]),
which computes the exact value of the map Pic(Xk )G ∼ = Z → Br(k) when X is a Severi–
Brauer variety of a field with Galois group G, or the computation of [GQM11] of the sign
of the Rost invariant.
Proposition 3.7. Let X be an algebraic stack and suppose that χ ∈ H1 (X, Cn ) and u ∈
H1pl (X, µn ) are fixed classes. In the notation above, we have [(χ, u)n ] = [Aχ,u ] in Br′ (X).
Proof. Both [Aχ,u ] and [(χ, u)n ] define classes in H2pl (BCn × Bpl µn , Gm ). As at the end of
the proof of Proposition 3.3, we see that [Aχ,u ] = k[(χ, u)n ] for some k prime to n. We saw
in Proposition 2.11 that they agree when pulled back to regular noetherian schemes. The
result follows.
4 A presentation of the moduli stack of elliptic curves

We will compute Br(M) using that it injects into Br(MZ[ 1 ] ) by Proposition 2.5(iv) and using
2
a specific presentation of M 1 , which we now describe. This presentation is standard and
Z[ 2 ]
we claim no originality in our presentation of it. For references, see [DR73] or [KM85],
Definition 4.1. A full level 2 structure on an elliptic curve E over a base scheme S is
a fixed isomorphism (Z/2)2S → E[2], where (Z/2)2S denotes the constant group scheme on
(Z/2)2 over S and E[2] is the subgroupscheme of order 2 points in E. If there exists an
isomorphism (Z/2)2S ∼ = E[2], an equivalent way of specifying a level 2 structure is to order
the points of exact order 2 in E(S) (over each connected component of S).
Remark 4.2. These full level 2 structures are sometimes called naive to distinguish them
from the level structures considered by Drinfeld, which allow one to extend M(2) to a stack
supported over all of Spec Z. We will not need this generalization in this paper. It is the
subject of [KM85].
The moduli stack M(2) of elliptic curves with fixed level 2 structures is a regular noethe-
rian Deligne-Mumford stack. Moreover, since the existence of a full level 2 structure implies
that 2 is invertible in S (by [KM85, Corollary 2.3.2] for example), the functor M(2) → M
which forgets the level structure factors through M 1 . This map is clearly equivariant
Z[ 2 ]
for the right S3 -action on M(2) that permutes the non-zero 2-torsion points and the trivial
S3 -action on M. Note that in general, being G-equivariant for a map f : X → Y of stacks
with G-action is extra structure: for every g ∈ G one has to provide compatible 2-morphism
σg : gf → f g (see [Rom05] for details). In the case of M(2) → MZ[ 21 ] though the equivariance
is strict in the sense that all σg are the identity 2-morphisms of M(2) → M 1 .
Z[ 2 ]
We have the following well-known statement; to fix ideas, we will provide a proof.
Lemma 4.3. The map M(2) → M 1 is an S3 -Galois cover.

Z[ 2 ]
Proof. It is enough to show that for every affine scheme Spec R over Spec Z[ 12 ] and every
elliptic curve E over Spec R, we can find a full level 2 structure étale locally. Indeed, if there
is one full level 2 structure on E, the map
(S3 )Spec R = S3 × Spec R → Spec R ×M 1

M(2)
Z[
2]
4. A presentation of the moduli stack of elliptic curves 14
from the constant group scheme on S3 is an isomorphism since we get every other full level
2 structure on E by permuting the non-zero 2-torsion points.
The elliptic curve E defines an R-point E : Spec R → MZ[ 1 ] of the moduli stack of elliptic
2
curves. Zariski locally we can assume the pullback E ∗ λ of the Hodge bundle to be trivial, in
which case there exists a nowhere vanishing invariant differential ω. By [KM85, Section 2.2],
we can then write E in Weierstrass form over Spec R, which after a coordinate change takes
the form
y 2 = x3 + b2 x2 + b4 x + b6 .
As a point (x, y) on E is 2-torsion if and only if y = 0, we have a full level 2 structure
after adjoining the three roots e1 , e2 and e3 of x3 + b2 x2 + b4 x + b6 to R. This defines an
étale extension as the discriminant of this cubic polynomial does not vanish (because E is
smooth).
Definition 4.4. A Legendre curve with parameter t over S is an elliptic curve Et with
Weierstrass equation
y 2 = x(x − 1)(x − t).
As the discriminant of this equation is 16t2 (t − 1)2 , such an equation defines an elliptic (and
hence Legendre) curve if and only if 2, t and t − 1 are invertible on S.
The points (0, 0), (1, 0), and (t, 0) define three non-zero 2-torsion points on Et . Taking
them in this order fixes a full level 2 structure on E. This defines a morphism
π : X → M(2),
where X is the parameter space of Legendre curves. In fact, X is an affine scheme, given as

X = Spec Z 12 , t±1 , (t − 1)−1 = A1Z[ 1 ] − {0, 1}.
2
We will use X throughout this paper to refer specifically to this moduli space of Legendre
curves. In particular, X is naturally defined over Z[ 21 ]. In general, given a scheme S, we let
XS = A1S − {0, 1}. Note that this is a slight abuse of notation as we do not assume that 2
is invertible on S.
We equip the map π : X → M(2) with the structure of a C2 -equivariant map with the
trivial C2 -action on X and M(2) by choosing σg : gπ → πg to be [−1] (i.e. multiplication by
−1 on the universal elliptic curve) for g ∈ C2 the non-trivial element. Note that [−1] fixes
the level 2 structure and so indeed defines a natural automorphism of idM(2) . The structure
of a C2 -equivariant map on π induces a map [X/C2 ] = BC2,X → M(2).
Proposition 4.5. The C2 -equivariant map π : X → M(2) is a C2 -torsor. Thus, the map
BC2,X → M(2) is an equivalence.
Proof. First we will show that an elliptic curve E → Spec R with full level 2 structure
can étale locally be brought into Legendre form. Our proof will be along the lines of
[Sil09, Proposition III.1.7], but we have to take a little bit more care.
As in the proof of Lemma 4.3, Zariski locally over Spec R, we can write E in the form
y 2 = x2 + b2 x2 + b4 x + b6
and the full level 2 structure allows us to factor the right hand side as
(x − e1 )(x − e2 )(x − e3 ),
where (e1 , 0), (e2 , 0) and (e3 , 0) are the nonzero 2-torsion points. We set p = e2 − e1 and
q = e3 − e1 . By a linear coordinate change, we get y 2 = x(x − p)(x − q).
Since the equation y 2 = x(x−p)(x−q) defines an elliptic curve, p, q and p−q are nowhere
√
vanishing. Thus, the extension R → R[ p] is étale so that we can (and will) assume étale
15 4. A presentation of the moduli stack of elliptic curves
√
locally to have a (chosen) square root p. Now, E is isomorphic to y 2 = x(x − 1)(x − t)
q
for t = p , where the isomorphism is given by x = px′ and y = p3/2 y ′ . Thus our original
E is indeed étale locally (on the base) isomorphic to a Legendre curve as an elliptic curve
with level 2 structure. It is moreover an elementary check with coordinate transformations
that there is at most one choice of t ∈ R such that the Legendre curve Et with paramter t
is isomorphic to E in M(2).
Now assume that our elliptic curve E over R is in Legendre form and assume further
that Spec R is connected. By definition, for a commutative R-algebra R′ , an element of
(X ×M(2) Spec R)(R′ ) consists of a Legendre curve Et together with an isomorphism of Et
to ER′ in M(2). By assumption this set is non-empty and it is indeed a torsor under the
group of automorphisms of ER′ in M(2). By [KM85, Corollary 2.7.2], the only non-trivial
automorphism of ER′ with level 2 structure is [−1]. Thus, the C2 -action exactly interchanges
the two elements of (X ×M(2) Spec R)(R′ ) and we obtain a C2 -equivariant equivalence
C2 × Spec R ≃ X ×M(2) Spec R.
As every E with full level 2 structure satisfies étale locally our assumptions, this implies
that X → M(2) is a C2 -torsor.
By the general fact that for a G-torsor X → Y, the induced map [X/G] → Y is an
equivalence, we obtain in our case the equivalence BC2,X ≃ M(2).
Corollary 4.6. The map c : M(2) → X sending y 2 = (x − e1 )(x − e2 )(x − e3 ) to y 2 =
x(x − 1)(x − ee23 −e
−e1
1
) exhibits X as the coarse moduli space of M(2).
Proof. The set of maps from M(2) ≃ BC2,X to X is in bijection with C2 -equivariant maps
X → X. Thus, a map M(2) → X exhibits X as the coarse moduli space if and only if the
precomposition with π is the identity. This is clearly the case for c.
It follows that the right S3 -action on M(2) induces a right S3 action on X. We can
describe this explicitly as follows. Consider the generators σ = (1 3 2) and τ = (2 3) of
GL2 (Z/2) ∼= S3 , of orders 3 and 2, respectively. Then,
t−1
σ(t) = ,
t
1
τ (t) = .
t
By a simple computation, the map c : M(2) → X defined above is strictly S3 -equivariant.
In contrast, the map π : X → M(2) described above is not S3 -equivariant, as one notes
for example by checking that the elliptic curves y 2 = x(x − 1)(x − t) and y 2 = x(x − 1)(x − 1t )
are generally not isomorphic. To actually explain the correct S3 -action on BC2,X , we have
to fix some notation.
Consider again an elliptic curve E given by y 2 = (x − e1 )(x − e2 )(x − e3 ). Set again
p = e2 − e1 and q = e3 − e1 so that we can write E as
y 2 = x(x − p)(x − q).
The only possible coordinate changes fixing the form of this equation are the transformations
y 7→ u3 y and x 7→ u2 x; such a coordinate change results in multiplying the standard invariant
differential ω = −dx
2y by u
−1
and sending p to u2 p and q to u2 q. Thus, pω ⊗2 and qω ⊗2 define
⊗2
canonical sections of λ on M(2), not dependent on any choice of Weierstrass form. Note
√
that these sections are nowhere vanishing. We can consider the C2 -torsor M(2)( p) → M(2)
L ⊗i

defined as the cyclic cover SpecM(2) i∈Z λ /(1 − p) . Étale locally on some Spec√ R, we
can trivialize λ so that p becomes an element of R and the C2 -torsor becomes Spec R[ p] →
√
Spec R. The C2 -torsor M(2)( p) → M(2) is equivalent to X → M(2). Indeed, we have
shown in the proof of Proposition 4.5 that the latter has a section as soon as we have a
chosen square root of p.
4. A presentation of the moduli stack of elliptic curves 16
As g ∗ λ for g ∈ S3 on M(2) is canonically isomorphic to λ (as this is pulled back from

M), we have an action of S3 on H0 (M(2), λ⊗∗ ). Consider the section
g(p)
∈ H0 (M(2), OM(2) ) ∼
= H0 (X, OX ),
p
e −e
which can for E as above be written as g(2)e2 −e1
g(1)
. For example, we have g(p) q
p = p for g = τ ,
which equals t on X. For a scheme S with a map f : S → X or f : S → M(2), we denote
the torsor adjoining the square root of f ∗ g(p)
p by Tf,g → S.
For the next lemma, we recall that an object in BC2,X (S) corresponds to a C2 -torsor
T → S and a C2 -equivariant map T → X, where X has the trivial C2 -action. Equivalently,
an object can be described as a C2 -torsor T → S with a map f : S → X. Let S3 act on
BC2,X in the following way: g ∈ S3 acts (from the right) on (T, f ) ∈ BC2,X (S) by setting
f g
g(T ) to be (T ×S Tf,g )/C2 and the map g(f ) to be the composition S −
→X−
→ X.
Lemma 4.7. The natural map M(2) → BC2,X induces an S3 -equivariant equivalence
BC2,X ≃ M(2).
√
Proof. As noted above, the map M(2) → BC2,X classifying the torsor M(2)( p) → M(2) is
an equivalence by Proposition 4.5 (as this torsor is equivalent to X → M(2)). We have only
to check the S3 -equivariance of this map. √
Given an f : S → M(2), the corresponding object in BC2,X is the torsor S( f ∗ p) → S
together with S → M(2) → X. The composition gf for g ∈ S3 corresponds to the torsor
p g
S( (gf )∗ p) → S together with S → M(2) →X − → X as M(2) → X is S3 -equivariant. As
(gf )∗ p = f ∗ (g(p)), we have (gf )∗ p = f ∗ p · f ∗ g(p)
p . Thus, we have a natural isomorphism
√ ∼
= p
(S( p)×S Tf,g )/C2 − → S( (gf )∗ p). One can check that these isomorphisms are compatible
(similarly to [Rom05, Definition 2.1], although we do not have a strict S3 -action on BC2,X )
so that one actually gets the structure of an S3 -equivariant map.
Of particular import will be the action of S3 on the units of X. Let ρ be the tautological
permutation representation of S3 on Z⊕3 and let ρ̃ be the kernel of the morphism
ρ∼
= indSC32 Z → Z
to the trivial representation, the adjoint to the identity.
Lemma 4.8. For any connected normal noetherian scheme S over Z[ 12 ], there is an S3 -
equivariant exact sequence
0 → Gm (S) → Gm (XS ) → ρ̃ → 0,
where S3 acts on Gm (S) trivially and ρ̃ is additively generated by the images of t and t − 1.
This exact sequence is non-equivariantly split.
Proof. Denote by π : XS → S the structure map. We have a map f : Z2 ⊕ Gm → π∗ Gm,XS

of sheaves on S, where f takes the two Z-summands to t and (t − 1), respectively. We claim
that this map is an isomorphism. It is enough to check this on affine connected opens Spec R,
where it follows from R being an integral domain (as it is normal). The non-equivariant
statement follows.
Moreover, the action of S3 onGm (S)
 is trivial
 by  definition. Set σ = (1 3 2) and τ = (2 3).
0 1
If we choose the basis vectors  1  and  0  for ρ̃, we obtain exactly the same S3 -
−1 −1
representation as on Gm (XS )/Gm (S) ∼ = Z{t, t − 1}, where the latter denotes the free Z-
module on t and t − 1 with elements thought of as tk (t − 1)l .
17 5. Beginning of the computation
5 Beginning of the computation

Let S be a connected regular noetherian scheme over Z[ 12 ], let MS be the moduli stack of
elliptic curves over S, and let M(2)S be the moduli stack of elliptic curves with full level 2
structure over S. The Leray–Serre spectral sequence for M(2)S → MS takes the form
Ep,q p q
2 : H (S3 , H (M(2)S , Gm )) ⇒ H
p+q
(MS , Gm ), (5.1)
with differentials dr of bidegree (r, 1 − r). In this section, we will collect the basic tools to
compute the E2 -term. We start with two brief remarks about the cohomology of S3 .
Lemma 5.2. Let M be a trivial S3 -module. Then,
H1 (S3 , M ) ∼
= M [2],
2
H (S3 , M ) =∼ M/2,
H3 (S3 , M ) ∼
= M [6],
4
H (S3 , M ) =∼ M/6.
Proof. We use the Lyndon–Hochschild–Serre spectral sequence for
1 → Z/3 → S3 → Z/2 → 1.
A reference is [Wei94, Example 6.7.10]. On the E2 -page, Epq 2 = 0 whenever p > 0 and
q > 0 because the cohomology of Z/3 is 3-torsion. Moreover, Z/2 acts on Hq (Z/3, M ) by
multiplication by −1 for q ≡ 1, 2 mod 4 and by 1 for q ≡ 0, 3 mod 4.
The next lemma is about the cohomology of the reduced regular representation ρ̃ of S3
introduced in Section 4.
Lemma 5.3. Let M be an abelian group, and let ρ̃ ⊗ M be an S3 -module through the action
on ρ̃. Then,
H0 (S3 , ρ̃ ⊗ M ) ∼
= M [3],
H (S3 , ρ̃ ⊗ M ) ∼
1
= M/3,
H2 (S3 , ρ̃ ⊗ M ) ∼
=0
H (S3 , ρ̃ ⊗ M ) ∼
3
= 0.
Proof. There is a short exact sequence of S3 -modules
0 → ρ̃ ⊗ M → ρ ⊗ M → M → 0.
In the associated long exact sequence in cohomology, note that Hi (S3 , ρ ⊗ M ) ∼

= Hi (C2 , M )
∼ S3
by Shapiro’s lemma, as ρ ⊗ M = indC2 M . The map
Hi (S3 , ρ ⊗ M ) ∼
= Hi (C2 , M ) → Hi (S3 , M )
is the transfer. We obtain short exact sequences
0 → coker trSC32 (Hi−1 ) → Hi (S3 , ρ̃ ⊗ M ) → ker trSC32 (Hi ) → 0.
Because C2 → S3 has a retraction, the restriction map Hi (S3 , M ) → Hi (C2 , M ) is the

projection to a direct summand. The transfer equals 3 times the inclusion of this summand as
can easily be deduced from the equation trSC32 resSC32 = 3. Thus, the transfer is multiplication
by 3 on H0 , an isomorphism on H1 and H2 and the inclusion M [2] → M [6] on H3 . The
lemma follows.
5. Beginning of the computation 18
These computations allow us to compute the E2 -term of the Leray–Serre spectral se-
quence
Ep,q p q
2 : H (S3 , H (M(2)S , Gm )) ⇒ H
p+q
(MS , Gm )
in a range. Using the results of the last two sections, we can analyze Hq (M(2)S , Gm ) in
terms of Hq (XS , Gm ). Especially Proposition 3.2 turns out to be useful as the short exact
sequences in it are S3 -equivariant by naturality. Using additionally Lemma 4.8 for the first
one, we obtain the S3 -equivariant extensions
0 → Gm (S) → Gm (M(2)S ) ∼
= Gm (XS ) → ρ̃ → 0, (5.4)
0 → Pic(XS ) ∼
= Pic(S) → Pic(M(2)S ) → µ2 (S) → 0, (5.5)
and
0 → Br′ (XS ) → Br′ (M(2)S ) → H1 (XS , µ2 ) → 0. (5.6)
The only point needing justification is that the pullback map Pic(S) → Pic(XS ) is an
isomorphism. It is injective because XS has an S-point. It is surjective as it factors through
the isomorphism Pic(S) → Pic(A1S ) and since j ∗ : Pic(A1S ) → Pic(XS ) is surjective, where
j denotes the inclusion XS ⊆ A1S . Indeed, given a line bundle L on XS , we take a coherent
subsheaf F of j∗ L with j ∗ F ∼
= L. The double dual of F is a reflexive sheaf L′ with j ∗ L′ still
isomorphic to L. By [Har80, Prop 1.9], L′ is a line bundle.
The sequence (5.5) is S3 -equivariantly split and thus consists only of S3 -modules with
the trivial action. Indeed, the morphism S → Spec Z[ 12 ] induces by pullback a morphism
from the exact sequence
0 → 0 → Pic(M(2)) → µ2 (Z[ 12 ]) → 0,
where the splitting is clearly S3 -equivariant. As µ2 (Z[ 12 ]) → µ2 (S) is an isomorphism for S

connected, the result follows. These observations allow us to compute the q = 0, 1 lines of
the Leray–Serre spectral sequence (5.1).
Lemma 5.7. If S is a connected regular noetherian scheme over Z[ 21 ], then there are natural
extensions
0 → Hp (S3 , Gm (S)) → Hp (S3 , Gm (M(2)S )) → Hp (S3 , ρ̃) → 0
for 0 ≤ p ≤ 3, and natural isomorphisms
Hp (S3 , H1 (M(2)S , Gm )) ∼
= Hp (S3 , Pic(S)) ⊕ Hp (S3 , µ2 (S))
for all p ≥ 0.
Proof. The first exact sequence follows from Lemma 5.2 and Lemma 5.3 using that H1 (S3 , ρ̃)
is 3-torsion and H2 (S3 , Gm (S)) is 2-torsion. The direct sum decomposition follows from the
fact that (5.5) is S3 -equivariantly split.
The only necessary remaining group we need to understand for our computations is
H2 (M(2)S , Gm )S3 , which we analyze using the short exact sequence (5.6).
Lemma 5.8. If S is a regular noetherian scheme over Z[ 12 ], then there is a canonical
isomorphism H1 (S, µ2 ) ∼
= H1 (XS , µ2 )S3 .
Proof. Using the S3 -equivariant short exact sequence
0 → Gm (XS )/2 → H1 (XS , µ2 ) → Pic(XS )[2] → 0,
we get a long exact sequence

S3
0 → (Gm (XS )/2) → H1 (XS , µ2 )S3 → Pic(XS )[2]S3 → H1 (S3 , Gm (XS )/2) → · · ·
19 5. Beginning of the computation
As the canonical map XS → S is S3 -equivariant, we obtain a map into this from the exact
sequence
0 → Gm (S)/2 → H1 (S, µ2 ) → Pic(S)[2] → 0.
S
As the maps Gm (S)/2 → (Gm (XS )/2) 3 and Pic(S)[2] → Pic(XS )[2] are isomorphisms
(using the exact sequence (5.4) and Lemma 5.3), the five lemma implies that H1 (S, µ2 ) →
H1 (XS , µ2 )S3 is an isomorphism as well.
From (5.6), we obtain a long exact sequence
0 → Br′ (XS )S3 → Br′ (M(2))S3 → H1 (S, µ2 ) → H1 (S3 , Br′ (XS )) → · · · (5.9)
Lemma 5.10. Let S be a regular noetherian scheme over Spec Z[ p1 ] for some prime p, and
let XS = A1S − {0, 1} as before. There is a non-canonically split exact sequence
0 → p Br′ (S) → p Br′ (XS ) → p H3{0,1} (A1S , Gm ) → 0.
Proof. By Proposition 2.14 we have an exact sequence
0 → p Br′ (A1S ) → p Br′ (XS ) → p H3{0,1} (A1S , Gm ) → p H3 (A1S , Gm ) → p H3 (XS , Gm ).
Because p is invertible on S, Proposition 2.5 implies that p Hi (S, Gm ) ∼

= p Hi (A1S , Gm ) for
all i ≥ 0. But, since XS has an S-point, it follows that p H (AS , Gm ) ∼
i 1
= p Hi (S, Gm ) →
i
p H (XS , Gm ) is split injective for all i.
Lemma 5.11. For any prime p and any regular noetherian scheme over Spec Z[ p1 ], there
is a canonical isomorphism
Hq{0,1} (A1S , Gm ) ∼ q q
= H{0} (A1S , Gm ) ⊕ H{1} (A1S , Gm ).
The action of S3 on p Br′ (XS )/p Br′ (S) is isomorphic to
ρ̃ ⊗ p H3{0} (A1S , Gm ) ∼
= ρ̃ ⊗ H1 (S, Qp /Zp ).
Proof. Given any étale sheaf F, there is a canonical isomorphism
H0{0,1} (A1S , F) ∼
= H0{0} (A1S , F) ⊕ H0{1} (A1S , F),
as one sees by an easy diagram chase. By deriving this isomorphism, the first part of the
lemma follows.
To prove the second statement, we compare the sequence of Lemma 5.10 with the long
exact sequence for étale cohomology with supports coming from the open inclusion XS ⊆ P1S .
Using the natural map of long exact sequences, we obtain a commutative diagram
0 / p Br′ (P1 ) / p Br′ (XS ) / p H3 1

S {0,1,∞} (PS , Gm )
∼
= =

0 / p Br′ (A1 ) / p Br′ (XS ) / p H3 (A 1
S , Gm )
/0
S {0,1}
with exact rows, where the left-hand vertical map is an isomorphism because it is injective
(by Proposition 2.5(iv)), p Br′ (S) → p Br′ (A1S ) is an isomorphism, and there is an S-point
of A1S ⊆ P1S .
Now, by Proposition 2.14,
M
3 1 ∼
p H{0,1,∞} (PS , Gm ) = H1 (S, Qp /Zp )
{0,1,∞}
5. Beginning of the computation 20
and M
3 1
p H{0,1} (AS , Gm )
∼
= H1 (S, Qp /Zp ).
{0,1}
With this description, the right-hand vertical map above is the natural projection away from
the factor of H1 (S, Qp /Zp ) corresponding to ∞. Let χ0 and χ1 be p-primary characters of S,
i.e., elements of H1 (S, Qp /Zp ). Then, as χ0 , χ1 vary, the Azumaya algebras (χ0 , t)⊗(χ1 , t−1)
give elements of Br(XS ) whose ramification classes (χ0 , χ1 ) span p H3{0,1} (A1S , Gm ). The
ramification of such a class computed in H3{0,1,∞} (P1S , Gm ) is (χ0 , χ1 , −χ0 − χ1 ). This
follows from Proposition 2.16, the fact that ram(π) (χ, π −1 ) = −ram(π) (χ, π) in the notation
of that proposition, and the fact that both t−1 and (t− 1)−1 are uniformizing parameters for
the divisor at ∞ of P1S . It follows that the image of p Br′ (XS ) inside p H3{0,1,∞} (P1S , Gm ) ∼
=
L 1 1
{0,1,∞} H (S, Qp /Zp ) can be identified with ρ̃ ⊗ H (S, Qp /Zp ).
We will analyze the implications for p-primary torsion for p > 2 in the next three sections.
For the rest of this section, we will begin the study of the 2-primary torsion of Br′ (MS ) where
S is a Z[ 12 ]-scheme. By Lemmas 5.11 and 5.3, we know that 2 H3{0,1} (A1S , Gm )S3 = 0. Thus
from Lemma 5.10, we see that 2 Br′ (S) → 2 Br′ (XS )S3 is an isomorphism. If we tensor the
sequence (5.9) with Z(2) , we obtain (using Lemmas 5.2 and 5.3) the exact sequence
∂
→ H1 (S3 , 2 Br′ (XS )) ∼
0 → 2 Br′ (S) → 2 Br′ (M(2))S3 → H1 (S, µ2 ) − = Br′ (S)[2] → · · · (5.12)
Here we recall that we denote for an abelian group A by A[2] its 2-torsion, while 2 A denotes
its 2-primary torsion. We want to analyze the boundary map ∂.
Lemma 5.13. If u ∈ H1 (S, µ2 ), then ∂(u) equals the Brauer class of the cyclic (quaternion)
algebra (−1, u).
Proof. We assume that S is connected. Denote by π : XS → BC2,XS the projection and by
c : BC2,XS → XS the canonical map to the coarse moduli space. Denote by
r : Br′ (BC2,XS ) → H1 (XS , µ2 )
the map obtained from the Leray spectral sequence. Finally, let
s : H1 (XS , µ2 ) → Br′ (BC2,XS )
be given by s(u) = [(χ, u)] = [(χ, c∗ u)], where χ ∈ H1 (BC2,XS , C2 ) classifies π. We have
r(s(u)) = u by Lemma 3.3.
Using [Ser97, Section 5.4], we can compute a crossed homomorphism representing ∂(u)
thus as
g 7→ π ∗ (g(s(u)) − s(u)) ∈ Br′ (XS ).
Consider the subgroup C2 = h(2 3)i ⊂ S3 and the C2 -equivariant morphism z : S → XS
classifying y 2 = x(x − 1)(x + 1) (i.e., t = −1). It follows from Lemmas 5.2 and 5.3 that
the morphism z ∗ resSC32 induces an isomorphism H1 (S3 , 2 Br′ (XS )) → H1 (C2 , 2 Br′ (S)). The
isomorphism H1 (C2 , Br′ (S)) → Br′ (S)[2] is given by evaluating the crossed homomorphism
at the non-trivial element (2 3) ∈ C2 . Thus, the coboundary map ∂ : H1 (S, µ2 ) → Br′ (S)[2]
sends u to
z ∗ π ∗ ((2 3)(s(u)) − s(u)).
As the pullback of XS → BC2,XS along π ◦ z : S → XS → BC2,XS is the trivial C2 -
torsor, z ∗ π ∗ s(u) = (πz)∗ (χ, u) defines the trivial Brauer class. By Lemma
√ 4.7, the action
of (2 3) multiplies the torsor XS → BC2,XS with the torsor BC2,XS ( t) → BC2,XS . Thus,
z ∗ π ∗ (2 3)(s(u)) = (z ∗ t, u) = (−1, u).
Summarizing, we obtain the following result.

21 6. The p-primary torsion in Br(MZ[ 1 ] ) for primes p ≥ 5
2
Proposition 5.14. Let S be a regular noetherian scheme over Z[ 12 ]. We have an exact

sequence
S ∂
→ Br′ (S)[2]
0 → 2 Br′ (S) → 2 Br′ (M(2)S ) 3 → H1 (S, µ2 ) −
S3
with ∂(u) = [(−1, u)]. The map 2 Br′ (S) → 2 Br′ (M(2)S ) is non-canonically split.
Proof. The exact sequence is exactly (5.12). The identification of ∂(u) follows from the
previous lemma. For the splitting, choose an S-point S → M(2)S . Then the composition
Br′ (M(2)S )S3 → Br′ (M(2)S ) → Br′ (S) provides the splitting.
6 The p-primary torsion in Br(MZ[ 21 ] ) for primes p ≥ 5

Before we proceed to study the 3-primary and 2-primary torsion, we will show in this section
that for a large class of S there is no p-primary torsion for p ≥ 5 in the Brauer group of
MS . Lemma 5.3 implies the crucial fact that there are no S3 -invariant classes in p Br′ (XS )
ramified at {0, 1} when p 6= 3 is invertible on S. The main point of the following theorem is
that this is true for p ≥ 5 even for certain regular noetherian schemes where p is not a unit.
Theorem 6.1. Let S be a regular noetherian scheme over Z and p ≥ 5 prime. Assume
1
that S[ 2p ] = SZ[ 1 ] is dense in S and that MS → S has a section. Then the natural map
2p
p Br′ (S) → p Br′ (MS ) is an isomorphism.
Proof. Assume first that 2 is invertible on S. The only contribution to p Br′ (MS ) in the
Leray–Serre spectral sequence (5.1) occurs as
S3
p H2 (BC2,XS , Gm )S3 = 2
p H (BC2,XS , Gm ) (6.2)
because Hi of S3 for i ≥ 1 can never have p-primary torsion for p ≥ 5.

We will argue that the p-group (6.2) is isomorphic to p Br′ (S) for all primes p ≥ 5 if
additionally p is invertible on S. To do so, note first that
2
p H (BC2,XS , Gm )
∼
= p Br′ (XS )
for p 6= 2 by Proposition 3.2. By Lemmas 5.10, 5.11 and 5.3, we see that p Br′ (S) →
′ S3
p Br (XS ) is an isomorphism.
This shows the theorem if 2p is invertible on S. Let now S be arbitrary regular noetherian
1
such that S[ 2p ] ⊂ S is dense and MS has an S-point. Consider the commutative diagram
p Br′ (MS ) / p Br′ (MS[ 1 ] )

2p
∼
=

p Br′ (S) / p Br′ (S[ 1 ]).
2p
induced by the choice of an S-point of MS . As MS has a cover by a scheme that is fppf

over S and fppf morphisms are open [Gro65, Thm 2.4.6], MS → S is open as well. Thus,
′ ′
MS[ 2p
1
] ⊂ MS is dense and hence Br (MS ) → Br (MS[ 2p
1 ) is injective by Proposition 2.5.
]
This implies that Br′ (MS ) → Br′ (S) is injective as well. As it is also split surjective, we see
that it is an isomorphism.
Remark 6.3. In general, it is a subtle question to decide whether MS has an S-point. For
example for S = Z[ 12 ] or S = Z[ 13 ], there is such an S-point, but for S = Z[ 15 ] or S = Z[ 29
1
]
there is none (for this and other examples see [EdGT90, Cor 1]). Nevertheless, sometimes
one can still control the p-power torsion for p ≥ 5 if there is no S-point, as the following
corollary shows.
7. The 3-primary torsion in Br(M 1 ) 22
Z[ ]
6
Corollary 6.4. The Brauer groups Br(M) ⊆ Br(MZ[ 1 ] ) have only 2 and 3-primary torsion.
2
Proof. Indeed, Br(M) ⊆ Br(MZ[ 1 ] ), and there is no p-torsion in Br(Z[ 12 ]) ∼

= Z/2 for p 6=
2
2.
7 The 3-primary torsion in Br(MZ[ 1 ] )

6
The next theorem describes the 3-primary torsion in Br′ (MS ) in many cases.
Theorem 7.1. Let S be a regular noetherian scheme. If 6 is a unit on S, then there is an

exact sequence
0 → 3 Br′ (S) → 3 Br′ (MS ) → H1 (S, C3 ) → 0,
which is non-canonically split. The map 3 Br′ (MS ) → H1 (S, C3 ) can be described as the
composition of pullback to XS and taking the ramification at the divisor {0} in A1S defined
by t (using Proposition 2.14).
Proof. The 3-primary torsion in Br′ (M(2)S ) ∼ = Br′ (BC2,XS ) is just the 3-primary torsion in
′ 1
Br (XS ) by Proposition 3.2 as Hpl (XS ; µ2 )(3) = 0. Similarly, since 2 is invertible in S, the
Leray–Serre spectral sequence (5.1) together with the group cohomology computations of
Lemma 5.2 and Lemma 5.3 and the exact sequences (5.4)-(5.6) say that
S3
3 Br′ (MS ) ∼
= 3 Br′ (XS ) .
Since 3 is invertible in S, we have a short exact sequence
0 → 3 Br′ (S) → 3 Br′ (XS ) → 3 H3{0,1} (A1S , Gm ) → 0
by Lemma 5.10.
The 0 and 1 sections are disjoint, so that there is an isomorphism of S3 -modules
M
3 1
3 H{0,1} (AS , Gm )
∼
= H1 (S, Q3 /Z3 ) ∼
= ρ̃ ⊗ H1 (S, Q3 /Z3 )
i=0,1
by Proposition 2.14 and Lemma 5.11. Thus, Lemma 5.3 implies that the long exact sequence
in S3 -cohomology takes the following form:
0 → 3 Br′ (S) → 3 Br′ (XS )S3 → H1 (Spec S, C3 ) → H1 (S3 , 3 Br′ (S)).
However, the action of S3 on 3 Br′ (S) is trivial, so the group on the right vanishes by
Lemma 5.2. Since MS has an S-point (because 2 is inverted), the splitting follows.
L
The map 3 Br′ (XS ) → i=0,1 H1 (S, Q3 /Z3 ) takes the ramification at the divisors {0}
S3
and {1}. At this point, we need to make the isomorphism ρ̃ ⊗ H1 (S, Q3 /Z3 ) → H1 (S, C3 )
∼
more explicit. By choosing the ordered basis t, t − 1 of Gm (XS )/Gm (S) = ρ̃, we see from
the description of the action that Z/3 ∼ = (ρ̃/3)S3 ⊆ ρ̃/3 is generated by t(t − 1). Indeed,
σ(t(t − 1)) = −t−2 (t − 1) and τ (t(t − 1)) = −t−2 (t − 1) for σ = (1 3 2) and τ = (2 3) and
thus
σ(t(t − 1)) ≡ t(t − 1) ≡ τ (t(t − 1)) ∈ Gm (XS )/3.
S3
This implies that ρ̃ ⊗ H1 (S, Q3 /Z3 ) → H1 (S, C3 ) can be identified with projection onto
′
the first coordinate. Thus, 3 Br (SR ) → H1 (S, C3 ) takes the ramification at the divisor
S3
{0}.
23 7. The 3-primary torsion in Br(M 1 )
Z[ ]
6
We want to be more specific about the Azumaya algebras arising from H1 (S, C3 ). For
that purpose consider the section ∆ ∈ H0 (M, λ⊗12 ), which is defined as follows. Given an
elliptic curve E over S, we can write it Zariski locally in Weierstrass form. Consider its
discriminant ∆E ∈ O(S) and its invariant differential ω ∈ Ω1E/R (E) ∼= λ(R). It is easy to
⊗12 ⊗12
see by [Sil09, Table 3.1] that ∆ = ∆E ω is a section of λ , which is invariant under
coordinate changes. Thus, ∆ defines a section of λ⊗12 on M.
Lemma 7.2. By Construction 2.8, we can associate √ with the line bundle L = λ⊗4 and
the trivialization ∆ : OM → L the µ3 -torsor M( ∆) → M whose class in H1 (M, µ3 ) we
⊗3 3
denote by [∆]3 . If S is a regular noetherian scheme and 6 is a unit on S, then the composite
∪(−[∆]3 )
H1 (S, C3 ) → H1 (MS , C3 ) −−−−−−→ H2 (MS , µ3 ) → 3 Br′ (MS )
is a section of the map 3 Br′ (MS ) → H1 (S, C3 ) of Theorem 7.1.
Remark 7.3. Informally, this section associates with χ ∈ H1 (S, C3 ) the symbol algebra
[(χ, ∆−1 )3 ].
Proof. The pullback of ∆ to X is the discriminant of the universal Legendre curve, which
is 16t2 (t − 1)2 (using the standard trivialization of λ on X given by dx 1
2y ). For χ ∈ H (S, C3 )
−1 ′
the pullback of [(χ, ∆ )3 ] to Br (XS ) is thus [(χ, 4t(t − 1))3 ]. As 4t(t − 1) is a uniformizer
for the local ring of A1S at t = 0, Proposition 2.16 implies the result.
Remark 7.4. While this map H1 (S, C3 ) → Br′ (MS ) is defined whether or not 6 is a unit on
S, without this assumption we do not know that H1 (S, C3 ) is the cokernel of 3 Br′ (S) →
′
3 Br (MS ).
Corollary 7.5. When S = Spec Z[ 16 ], there is an isomorphism 3 Br(M 1 )

∼
= Q3 /Z3 ⊕Z/3.
Z[ 6 ]
The 3-torsion subgroup is generated by classes σ and θ, which can be described as follows.
Let χ ∈ H1 (Spec Z[ 61 ], C3 ) be the character of the Galois extension Q(ζ9 + ζ 9 ) of Q. Then
σ = [(χ, 6)3 ] and θ = [(χ, 16∆−1 )3 ], which pulls back to [(χ, t(t − 1))3 ] on X 1 .
Z[ 6 ]
Proof. We claim first that Q(ζ9 + ζ 9 ) is the only cyclic cubic extension L of Q that ramifies
at most at 2 and 3. This can either be deduced from [Has48, I.§1.2] or shown as follows.
By the Kronecker–Weber theorem, any cyclic cubic extension L of Q has to embed into a
cyclotomic extension Q(ζn ) and is more precisely its fixed field under a normal subgroup
H ⊂ (Z/n)× of index 3. As H contains all elements of 2-power order, we can assume that
n is odd and thus L does not ramify at 2 but only at 3. Proposition 3.1 of [Lem05] shows
that L is unique and must be Q(ζ9 + ζ 9 ). This implies that H1 (Spec Z[ 16 ], C3 ) ∼
= Z/3.
Using that 3 Br(Z[ 61 ]) ∼
= Q3 /Z3 , the structure of the Brauer group Br(M 1 )[3] follows
Z[ 6 ]
from Theorem 7.1. The description of θ follows directly from the last lemma (where we have
modified the section by an element of 3 Br(Z[ 16 ]) for convenience).
Last we need to show that [(χ, 6)3 ] is non-zero in Br(Z[ 16 ])[3] ∼
= Z/3. It suffices to check
that (χ, 6) is ramified at the prime (2). Note that the minimal polynomial of ζ9 + ζ 9 is
w3 + w + 1. By Proposition 2.16, the ramification at (2) in H1 (F2 , C3 ) ∼= Z/3 is the class of
the extension w3 + w + 1 over F2 . Since this polynomial is irreducible (it has no solutions
in F2 and it has degree 3), it follows that the ramification is non-zero.
Corollary 7.6. Let R = Z[ 21 ] or R = Z. Then the order of 3 Br(MR ) is either 1 or 3.
Proof. Using the injectivity of Br(M) → Br(M 1 ), it suffices to prove this when R =
Z[ 2 ]
Z[ 21 ]. Now, we claim that no non-zero class α ∈ 3 Br(Z[ 16 ]) ⊆ 3 Br(MZ[ 1 ] ) extends to
6
8. The ramification of the 3-torsion 24
MZ[ 1 ] . Indeed, we can take the Legendre curve y 2 = x(x − 1)(x − 2), which defines a point
2
Spec Z[ 21 ] → M. Using the commutative diagram
Spec Z[ 61 ] /M 1
Z[ ]
/ Spec Z[ 1 ]
6
6

Spec Z[ 12 ] /M 1 ,
Z[ 2 ]
we see that if α ∈ 3 Br(Z[ 16 ]) did extend to M 1 , then it would be zero in the Brauer group
Z[ 2 ]
of Spec Z[ 61 ], as it would extend to 3 Br(Z[ 12 ]) = 0. However, these classes are all non-zero
in Br(Z[ 16 ]) since the composition at the top of the commutative diagram is the identity for
any Z[ 16 ]-point of the moduli stack.
Now, if α = β + mθ extends, where β ∈ 3 Br(Z[ 61 ]), then 3α = 3β also extends. Hence, it
must be that α has order at most 3. In particular, this means that every class of 3 Br(M 1 )
Z[ 2 ]
is actually 3-torsion and hence this group is a subgroup of (Z/3)2 . But, we have already
seen that σ does not extend. So, it is a proper subgroup, and hence it has order at most
3.
Proposition 7.7. Suppose there are Legendre curves Ei : y 2 = x(x − 1)(x − ti ) over
Spec Z3 [ζ3 ] for i = 1, 2 such that
[(χ, t1 (t1 − 1))] 6= 0 and [(χ, t2 (t2 − 1))] = 0
in Br(Q3 (ζ3 ))[3] = Z/3, where χ denotes the pullback of the Galois extension Q(ζ9 + ζ 9 ) of
Q to Q3 (ζ3 ). Then 3 Br(M[ 21 ]) = 0.
Proof. Suppose that α = aσ + bθ is a linear combination of the classes found in Corollary 7.5
where we can assume that b ∈ {1, 2} since σ does not extend. Suppose that α extends to
Br(MZ[ 1 ] ). We can pull back α along the two Q3 (ζ3 )-points of M defined by Ei and compute
2
the ramifications in Br(Q3 (ζ3 )). Let k = [(χ, t1 (t1 − 1))]. For i = 1, we get a + bk and for
i = 2 we get a. Since k is non-zero, these cannot be simultaneously zero modulo 3. But
the two maps Br(MZ[ 1 ] ) → Br(Q3 (ζ3 )) factor over Br(Z3 [ζ3 ]) = 0, which is a contradiction.
2
Thus, α cannot extend to Br(M 1 ).
Z[ 2 ]
8 The ramification of the 3-torsion

Our aim in this section is to show that 3 Br(M[ 21 ]) = 0 using Proposition 7.7.
Lemma 8.1. The natural inclusion Q(ζ9 + ζ 9 , ζ3 ) → Q(ζ9 ) is an isomorphism.
Proof. The left hand side is a subfield of Q(ζ9 ) that is strictly larger than Q(ζ9 + ζ 9 ) and
thus is equal to Q(ζ9 ).
We will need the following lemma to aid our Hilbert symbol calculations below.
Lemma 8.2. Consider the cyclotomic field Q(ζ) with ζ = ζ3 and π = 1 − ζ and denote by
Tr the trace for Q(ζ) over Q. Then


 (−1)3k · 33k · 2 if l = 0



 (−1)3k
· 3 3k+1
if l = 1


(−1)3k · 33k+1 if l = 2
Tr(π 6k+l ) =

 0 if l = 3



 (−1)3k+1
· 3 3k+2
if l = 4


 3k+1 3k+3
(−1) ·3 if l = 5
25 8. The ramification of the 3-torsion
Proof. We have π 2 = (1 − ζ)2 = −3ζ and thus π 6k = (−3)3k . Therefore, we have just to
compute Tr(π l ) for l = 0, . . . , 5, which is easily done.
We come to a key arithmetic point in our proof, where we compute the Hilbert symbol
at the prime 3 of certain degree 3 cyclic algebras. By Proposition 2.17, this will allow us to
check whether certain cyclic algebras are zero in the Brauer group.
Lemma 8.3. Consider the cyclotomic field Q3 (ζ) with ζ = ζ3 and π = 1−ζ the uniformizer.
Then we have
ζ, t(t − 1) 2
= ζ 1−b
π
in µ3 (Q3 (ζ)), where t = 2 + bπ with b ∈ Z3 .
Proof. We use the formula of Artin–Hasse (see [Neu99, Theorem V.3.8]) to compute this
Hilbert symbol. By this formula, we have

ζ, a
= ζ Tr(log a)/3 ,
π
where a ∈ 1 + p (for p ⊂ Z3 [ζ3 ] the maximal ideal) and Tr the trace for Q3 (ζ) over Q3 .
This formula directly applies to t − 1 = 1 + bπ. We have
∞
X (bπ)i
log(t − 1) = (−1)i+1 .
i=1
i
Tr(π i )
Again, it follows easily from Lemma 8.2 that i is divisible by 9 for i ≥ 3. Thus,
Tr(log(t − 1)) Tr(bπ) Tr(b2 π 2 ) b2

≡ − ≡b− mod 3
3 3 6 2
ζ,t−1
b2
and π = ζ b− 2 .
ζ,t

ζ,−1 2
To compute note that ζ,t
π π =
ζ,−t ζ,−1
π π = ζ,−t
π . Indeed, π = 1 and hence

also ζ,−1
π = 1 (in µ3 (Q3 (ζ3 ))). We have −t = 1 + (−3 − bπ). Thus,
∞
X (−3 − bπ)i
log(−t) = (−1)i+1 .
i=1
i
Tr((−3−bπ)i ))
It follows easily from Lemma 8.2 that i is divisible by 9 for i ≥ 3. Thus,
Tr(log(−t)) Tr(−3 − bπ) Tr((−3 − bπ)2 ) b2

≡ − ≡ −2 − b − mod 3.
3 3 6 2
b2
Thus, ζ,t
π =
ζ,−t
π = ζ −2−b− 2 . It follows that

ζ, t(t − 1) ζ, t ζ, t − 1 2 2
= = ζ −2−b = ζ 1−b ,
π π π
as desired.
Theorem 8.4. We have
3 Br(M) =3 Br(MZ[ 21 ] ) = 0.
Proof. By Proposition 7.7 and Proposition 2.17, it suffices to find two Legendre curves E1
and E2 over Z3 [ζ3 ] with corresponding classes [(χ, t1 (t1 − 1))] 6= 0 and [(χ, t2 (t2 − 1))] = 0.

The associated condition on the Hilbert symbols is ζ,t1 (tπ1 −1) 6= 1 and ζ,t2 (tπ2 −1) = 1.
(Recall here that χ is the character associated with adjoining ζ9 + ζ 9 which over Q3 (ζ3 ) is
9. The 2-primary torsion in Br(MZ[ 1 ] ) 26
2
isomorphic to Q3 (ζ9 ) by Lemma 8.1.) Take ti = 2 + bi π, where b1 = 0 and b2 = 1. Consider

the two elliptic curves
E1 : y 2 = x(x − 1)(x − 2)
and
E2 : y 2 = x(x − 1)(x − (2 + π)).
The previous lemma says that we have

ζ, t1 (t1 − 1) ζ, 2(2 − 1)
= = ζ 6= 1
π π
and
ζ, t2 (t2 − 1) ζ, (2 + π)(1 + π)
= = ζ 0 = 1.
π π
This completes the proof.
9 The 2-primary torsion in Br(MZ[ 21 ] )

Throughout this section, let S denote a connected regular noetherian scheme over Spec Z[ 21 ].
′
Given a stack X over S, let Br (X) = coker(Br′ (S) → Br′ (X)).
Theorem 9.1. Let S be a regular noetherian scheme over Z[ 21 ] with Pic(S) = 0. There is
a natural exact sequence
′
0 → Gm (S)/2 → 2 Br (MS ) → G → 0,
where G ⊂ Gm (S)/2 is the subgroup of all those u with [(−1, u)] = 0 ∈ Br(S).
We will prove the theorem after several preliminaries. Figure 3 shows a small part of
the 2-local Leray–Serre spectral sequence (5.1) for the S3 -Galois cover M(2)S → MS . The
description follows from Lemmas 5.2, 5.3, 5.7 and Proposition 5.14.
2 Br′ (S) ⊕ Pic(S)[2] ⊕ G ❭❭❭❭❭❭❭

❭❭❭❭❭❭❭❭❭❭❭❭
❭❭❭❭❭❭❭❭.
Pic(S)(2) ⊕ µ2 (S) ❭❭❭❭❭❭❭❭❭Pic(S)[2] ⊕ µ2 (S) ❭❭❭❭❭❭Pic(S)/2 ⊕ µ2 (S)
❭❭❭❭❭❭❭❭❭❭❭❭ ❭
❭❭❭❭❭❭❭❭❭❭❭❭❭❭❭❭❭❭❭❭❭❭❭❭❭❭❭❭❭❭❭
Gm (S)(2) µ2 (S) . G (S)/2
m
❭- µ (S)
2
Figure 3: The E2 -page of the Leray–Serre spectral sequence computing Hi (MS , Gm )(2) for
i ≤ 2.
From now on, we will localize everything in this section implicitly at 2.

Proposition 9.2. Let S be a connected regular noetherian scheme over Z[ 12 ]. The differen-
tial d20,1 in the Leray–Serre spectral sequence of Figure 3 always vanishes and d0,2
2 and d0,2
3
vanish if Pic(S) = 0.
Proof. The map
Pic(MS ) → E0,1 ∼
2 = Pic(S)(2) ⊕ µ2 (S)
is surjective as −1 ∈ µ2 (S) can be realized as λ⊗6 . This implies that there can be no differ-
ential originating from E0,1 ′ ′
2 . Moreover, 2 Br (S) splits off from Br (MS ), so the differentials
0,2 0,3 ′
d2 and d3 vanish on 2 Br (S).
Now assume Pic(S) = 0. Then also Pic(S)[2] = 0 and Pic(S)/2 = 0. So, we are
concerned with the vanishing of d0,22 : G → µ2 (S) and d0,2 3 : G → µ2 (S). However, by
pulling back the spectral sequence to a geometric point x of S, we find G = Gm (x)/2 = 0,
while µ2 (x) ∼ 0,2 0,2
= Z/2. This implies that d2 and d3 vanish.
27 9. The 2-primary torsion in Br(MZ[ 1 ] )
2
To resolve the differential d1,1

2 and solve possible extension issues we will first consider
schemes S over Z[ 12 , i]. In this case we can compare the Leray–Serre spectral sequence
considered above with the Leray–Serre spectral sequence for the C2 -Galois cover BC2,S →
BC4,S .
Proposition 9.3. If S is a regular noetherian Z[ 12 , i]-scheme, then Br′ (MS ) ∼

= Br′ (BC4,S ).
Proof. Consider the elliptic curve E : y 2 = x(x − 1)(x + 1) over Z[ 12 , i] with discriminant 64.
It has an automorphism η of order 4 given by y 7→ iy and x 7→ −x, which defines a map
BC4,Z[ 21 ,i] → M 1 . The 2-torsion points of E are (0, 0), (1, 0) and (−1, 0); taking them in
Z[ 2 ]
this order defines a full level 2 structure. We can base change this elliptic curve together
with its level structure to an arbitrary Z[ 12 , i]-scheme S. This results in pullback squares
` `
S3 S /
S3 /C2 BC2,S / M(2)S (9.4)

S / BC4,S / MS .
`
Here we use that η acts on the scheme S3 S of level structures on ES by multiplication with
`
S3 and in particular η 2 acts trivially. Thus, the stack
the cycle (2 3) ∈` `quotient S3 S /C4
is equivalent to S3 /C2 BC2,S . More precisely, the S3 -Galois cover S3 /C2 BC2,S → BC4,S
is induced along an inclusion C2 → S3 from the C2 -Galois cover BC2,S → BC4,S . The
right square is indeed cartesian as can be checked after base change along the étale cover
S → BC4,S .
In the Leray–Serre spectral sequence
a
Ep,q p
2 = H (S3 , H (
q
BC2,S , Gm )) ⇒ Hp+q (BC4,S , Gm ),
S3 /C2
`
the S3 -modules Hq ( S3 /C2 BC2,S ) are all induced up from C2 . Thus, the spectral sequence
is isomorphic to the Leray–Serre spectral sequence for the C2 -Galois cover BC2,S → BC4,S .
The Leray–Serre spectral sequence computing Hp+q (BC4,S , Gm ) from the Gm -cohomology
of BC2,S is displayed in Figure 4. The computation follows from Proposition 3.2 together
with the fact that C2 acts trivially on the cohomology of BC2,S (as indeed the morphism
t : BC2,S → BC2,S for t ∈ C2 the generator is the identity; only the natural transformation
id → t2 is not the identity).
Br′ (S) ⊕ Pic(S)[2] ⊕ Gm (S)/2 ❭❭❭❭❭❭

❭❭❭❭❭❭❭❭❭❭❭❭❭
❭❭❭❭❭❭❭❭.
Pic(S) ⊕ µ2 (S) Pic(S)[2] ⊕ µ2 (S) ❭❭❭❭❭❭Pic(S)/2
❭❭❭❭❭❭❭❭❭⊕❭❭µ2 (S)
❭❭❭❭❭❭❭❭❭❭-
Gm (S) µ2 (S) Gm (S)/2 µ2 (S)
Figure 4: Part of the Leray–Serre spectral sequence for BC2,S → BC4,S .
By the considerations above, the pullback square (9.4) induces a map
Hp (S3 , Hq (M(2)S , Gm )) → Hp (C2 , Hq (BC2,S , Gm )).
Note first that G = Gm (S)/2 in our case as −1 is a square. If we identify M(2)S with
BC2,XS this map on cohomology groups is induced by the maps S → XS (classifying the
Legendre curve ES ) and C2 → S3 . This induces an isomorphism of spectral sequences for
p + q ≤ 3 and q ≤ 1 for p + q = 3 by Figure 3.
9. The 2-primary torsion in Br(MZ[ 1 ] ) 28
2
Corollary 9.5. Let S be a regular noetherian scheme over Z[ 21 ]. The restriction of the
differential d1,1
2 to µ2 (S) in the Leray–Serre spectral sequence for M(2)S → MS defines an
∼
=
isomorphism µ2 (S) − → µ2 (S), while d0,2
3 = 0.
Proof. Consider S ′ = Spec Z[ 21 , i]. By Proposition 3.2, we see that
Br′ (BC4,S ′ ) ∼
= Br′ (S ′ ) ⊕ Pic(S ′ )[4] ⊕ Gm (S ′ )/4 ∼
= Gm (S ′ )/4,
since the Brauer and Picard groups of Z[ 12 , i] are zero. Hence, Br′ (BC4,S ′ ) ∼= Z/4 ⊕ Z/4,
with generators given as i and 1 + i. In Figure 4, we see that the only way to have a group of
order 16 in the abutment Br′ (BC4,S ′ ) is that d1,1 ′ ′
2 : µ2 (S ) → µ2 (S ) is an isomorphism, both
in the Leray–Serre spectral sequence for H (BC4,S ′ , Gm ) and for H∗ (MS ′ , Gm ). It follows
∗
that this differential is already an isomorphism in the Leray–Serre spectral sequence for
H∗ (MZ[ 1 ] , Gm ) by naturality. This in turn implies by naturality that d21,1 |µ2 (S) : µ2 (S) →
2
µ2 (S) in the Leray–Serre spectral sequence for H∗ (MS , Gm ) is an isomorphism for any
regular noetherian Z[ 12 ]-scheme S. As the target of d0,2
3 is already zero on E3 , the differential
d0,2
3 must vanish.
Finally, we prove the theorem from the beginning of the section.
Proof of Theorem 9.1. The claim follows from the determination of the differentials in the
range pictured in Figure 3.
We want to be more specific about the Brauer group classes coming from Gm (S)/2.
Recall the section√∆ ∈ H0 (M, λ⊗12 ) from
L Section 7. As in
Construction 2.8, we can define
⊗6i
the C2 -torsor M( ∆) = SpecM 1 i∈Z λ /(∆ − 1) → MZ[ 1 ] that adjoins a square
Z[
2] 2
root of ∆ to M 1 . For a unit u ∈ Gm (Z[ 21 ]), we denote by (∆, u) the symbol (quaternion)
Z[ 2 ]
algebra associated with this torsor.
Proposition 9.6. Let S denote a connected regular noetherian scheme over Spec Z[ 12 ].
Then the map
Gm (S)/2 → Br′ (MS )
from the Leray–Serre spectral sequence sends u to [(u, ∆)2 ].
Proof. We consider the Leray–Serre spectral sequence
Hp (C3 , Hq (M(2)S , Gm ))(2) ⇒ Hp+q (M(2)S /C3 , Gm )(2) ,
where M(2)/C3 denotes the stack quotient by the subgroup C3 ⊂ S3 . Its E2 -term is clearly
concentrated in the column p = 0. From Lemma 5.7 and Proposition 5.14, it is easy to see
that H0 (C3 , Hq (M(2), Gm )(2) ∼
= H0 (S3 , Hq (M(2), Gm )(2) ). We can now consider the further
Leray–Serre spectral sequence
Hp (C2 , Hq (M(2)S /C3 , Gm ))(2) ⇒ Hp+q (MS , Gm )(2) ,
and we see that it has the same E2 -term as the Leray–Serre spectral sequence for the S3 -cover
M(2)S → MS in the range depicted in Figure 9. √
We claim that the C2 -torsor M(2)S /C3 → MS agrees with MS ( ∆) → MS . For this
it suffices to show that ∆ becomes a square on M(2)S /C3 . With p, q ∈ H0 (M(2), λ⊗2 ) as
in the discussion after Corollary 4.6, we have ∆ = 16p2 q 2 (p − q)2 . The C3 -action permutes
p, (−q) and (q − p) cyclically so that 4pq(p − q) is a C3 -invariant section of λ⊗6 whose square
is indeed ∆.
Now the statement follows from Corollary 3.6.
The following is one of our main results. We recall the convention that everything is
implicitly 2-local so that Br(MR ) for a ring R denotes really Br(MR )(2) .
29 9. The 2-primary torsion in Br(MZ[ 1 ] )
2
Proposition 9.7. Let P be a set of prime numbers including 2 and denote by ZP ⊂ Q the
subset of all fractions where the denominator is only divisible by primes in P . Then
M M
Br(MZP ) ∼
= Br(ZP ) ⊕ Z/2 ⊕ Z/4.
p∈P ∪{−1}, p∈P,
p≡3 mod 4 p6≡3 mod 4
Proof. First we have to compute the subgroup

M
G ⊂ Gm (ZP )/2 ∼
= F2 .
P ∪{−1}
By Proposition
2.17, a quaternion algebra (a, b) ramifies at
p if and only if the Hilbert
symbol a,b equals −1. By [Neu99, Theorem V.3.6], −1,−1
= −1 if and only if p = 2, ∞
−1,q
p p
and p = −1 if and only if q ≡ 3 mod 4 and p = 2, q (for q a prime number). We see
that G has an F2 -basis given by the primes not congruent to 3 mod 4.
We obtain a diagram
0 / Gm (ZP )/2 / Br(MZP ) /G /0

0 / Gm (ZP [i])/2 / Br(MZ [i] )
P
/ Gm (ZP [i])/2 / 0.
By Proposition 9.3, Br(M 1 )

∼
= Gm (ZP [i])/4. As the map G → Gm (ZP [i])/2 is injective,
Z[ 2 ,i]
we see that none of the nonzero lifts of elements of G to Br(MZP ) are 2-torsion. The
proposition follows.
This shows the 2-local part of the computation of Br(MQ ) and Br(MZ[ 1 ] ) in Theorem
2
1.1, while the 3-local part was already contained in Theorems 7.1 and 8.4 and the p-local
part for p > 3 in Theorem 6.1.
Remark 9.8. We can describe all the Brauer classes in Br(MZP ) explicitly when P is again
a set of prime numbers including 2. We already saw in the last two propositions that
Br(MZP )[2] has an F2 -basis given by [(p, ∆)2 ], where p ∈ P ∪ {−1}. When p ∈ S and either
p = 2 or p ≡ 1 mod 4, we will give explicit elements of order 4 in the Brauer group of MZP
generating the Z/4-subgroups of the proposition.
To describe the 4-torsion we start with a small observation. Given a cyclic algebra (χ, υ),
where χ is a C4 -torsor and υ a µ4 -torsor, 2[(χ, υ)] is represented by (χ′ , υ ′ ), where χ′ and
υ ′ are obtained from χ and υ via the morphisms C4 → C2 and µ4 → µ2 . Concretely, this
means that χ′ are the C2 -fixed points of χ and that if υ is given by adjoining the 4-th root
of a section u of L⊗4 , then υ ′ is given by adjoining a square root of u, a section of (L⊗2 )⊗2 .
For primes p ≡ 1 mod 4, we construct P a C4 -Galois
extension L of Q whose C2 -fixed
√ √ √
points are Q( p). As p is the Gauß sum p−1 a=1 p
a a
ζp , we see that Q( p) ⊂ Q(ζp ). The
Galois group of the Q-extension Q(ζp ) is cyclic of order p − 1, which is divisible by 4. Thus,
√
it has a unique cyclic subextension L of degree 4 whose C2 -fixed points are Q( p). Note
Pp−1
that L is only ramified at p. Explicitly, L is generated by the Gauß sum a=1 ϕ(a)ζpa , where
ϕ : (Z/p)× → µ4 (C) is a surjective character.
For p = 2, we take L = Q(ζ16 + ζ 16 ) instead, which is the unique C4 -Galois subextension
of Q(ζ16 ) over Q. If we denote for p = 2 or p ≡ 1 mod 4 the character of L/Q by χ, these
define C4 -Galois covers of MS by pullback and we abuse notation and write χ also for these
covers. The cup product [(χ, ∆)4 ] in Br′ (MZP ) is a class such that 2[(χ, ∆)4 ] = [(p, ∆)2 ]
and thus has exact order 4. It follows that the classes [(χ, ∆)4 ] give a basis of the 4-torsion
of Br(MZP ).
10. The Brauer group of M 30
10 The Brauer group of M

In this section, we will complete the computation of Br(M). In the last section, we saw that
Br(M[ 21 ]) ∼
= Z/2 ⊕ Z/2 ⊕ Z/4
with generators α = [(−1, −1)2 ], β = [(−1, ∆)2 ] and 12 γ for γ = [(2, ∆)2 ]. We will study
the ramification of these classes for certain elliptic curves and the reader can find more
information about these curves in The L-functions and Modular Forms Database [TL].
As above, we will write a,b 2 ∈ µ2 (Q2 ) for the Hilbert symbol in Q2 at the prime 2.

Hence, a,b 2 = ±1. Recall from Proposition 2.17 that if χ ∈ H1 (Q2 , C2 ) ∼
= H1 (Q2 , µ2 )
corresponds
to a unit v ∈ Gm (Q2 )/2, then the degree 2 cyclic algebra (χ, u) has class
u,v
in Br(Q ∼ ∼
2 )[2] = Z/2 = µ2 (Q2 ). Since Br(Z2 ) = 0, the Hilbert symbol measures the
2
ramification along (2) in Spec Z2 .
Proposition 10.1. Every non-zero linear combination of α, β, 12 γ is ramified along (2), so

these linear combinations are not in the image of Br(M) → Br(M[ 21 ]).
Proof. It suffices to prove this for all seven nonzero linear combinations of α, β, γ. Indeed,
if all these linear combinations are ramified, then any linear combination rα + sβ + 12 γ is
ramified as well. As explained in the introduction around the diagram (1.4), it suffices to
construct for each non-zero linear combination ρ an elliptic curve Spec Z2 → M such that
the pullback of ρ to Spec Q2 is non-zero in Br(Q2 ). Indeed, Br(Z2 ) = 0.
Let
E1 : y 2 + y = x3 − x2 ,
the elliptic curve with Cremona label 11a3. This curve has discriminant −11, which is a
unit, so we get an elliptic curve over Z2 , and two associated Brauer classes, (2, −11) and
(−1, −11) over Q2 . We can ask what the ramification is. The Hilbert symbol in this case
is computed as follows [Ser73, Chapter III]. Given a = 2α u and b = 2β v ∈ Gm (Q2 ), where
u, v ∈ Gm (Z2 ), we have
a, b
= (−1)ǫ(u)ǫ(v)+αω(v)+βω(u) ,
2
u−1 u2 −1
where ǫ(u) ≡ 2 and ω(u) ≡ 8 .
Hence,
2, −11
= (−1)ω(−11) = (−1)15 = −1.
2
Hence, (∆, 2) is ramified at 2. Similarly,

−1, −11
= (−1)ǫ(−1)ǫ(−11) = (−1)6 = 1.
2
The curve
E2 : y 2 + xy = x3 − 2x2 + x
has Cremona label 15a8 and discriminant −15. This time, the Hilbert symbols are

2, −15
=1
2
and
−1, −15
= 1.
2
The curve
E3 : y 2 + xy + y = x3 − x2
31 11. The Brauer group of M over Fq with q odd
has Cremona label 53a1 and discriminant −53. In this case, the Hilbert symbols are

2, −53
= −1
2
and
−1, −53
= −1.
2
Now, let ρ = α + kβ + mγ with k and m integers. The elliptic curve E2 gives a point
x : Spec Q2 → M where v2 (x∗ ρ) = −1. It follows that all classes of the form ρ are ramified
along (2). Now, E3 proves that β and γ are ramified along (2). Finally, E1 proves that β + γ
is ramified along (2).
Theorem 10.2. Br(M) = 0.
Proof. This follows from Corollary 6.4, Theorem 8.4, and Propositions 9.7 and 10.1.
11 The Brauer group of M over Fq with q odd

As another application of our methods, we compute the Brauer group of MFq when q =
pn and p is an odd prime. There is remarkable regularity in this case, which is possibly
surprising based on what happens for number fields.
Theorem 11.1. Let q = pn where p is an odd prime. Then, Br(MFq ) ∼
= Z/12.
Proof. Recall that Br(Fq ) = 0. Thus, by Theorem 9.1, there is an extension 0 → F× q /2 →
× × ∼
2 Br(MFq ) → Fq /2 → 0. Since q is odd, Fq /2 = Z/2. The remainder of Section 9, especially
Remark 9.8, implies that the extension is non-split so that in fact 2 Br(MFq ) ∼= Z/4.
Now, let ℓ ≥ 3 be a prime, which we do not assume is different from p. The possible
terms contributing to ℓ Br(MFq ) in the Leray–Serre spectral sequence for M(2)Fq → MFq in
Gm cohomology, besides ℓ Br(XFq )S3 , are H1 (S3 ,ℓ Pic(M(2)Fq ) and H2 (S3 ,ℓ Gm (M(2)Fq )) ∼
=
H2 (S3 ,ℓ Gm (XFq )). The first of these is zero since ℓ Pic(M(2)Fq ) ∼
= ℓ Pic(XFq ) = 0 for
ℓ odd. The second has no odd primary torsion. This follows from the exact sequence
0 → Gm (Fq ) → Gm (XFq ) → ρ̃ ⊗ Z → 0 together with Lemmas 5.2 and 5.3.
S
Thus, we see that ℓ (Br(MFq )) ∼= ℓ Br(XFq ) 3 for ℓ odd. So, it suffices to compute the
Brauer group of XFq as an S3 -module. In general, our argument in the rest of the paper relies
fundamentally on Lemma 5.11, which requires ℓ to be invertible to analyze the ramification
map as in Proposition 2.14. However, in this case, XFq is a curve over a finite field, so
the ramification theory simplifies drastically. Consider the commutative diagram of exact
ramification sequences due to [Gro68c, Proposition 2.1]
L
0 / Br(P1 ) / Br(η) / Q/Z / Q/Z /0
Fq x∈(P1Fq )(1)
=
L
0 / Br(XFq ) / Br(η) /
x∈(XFq )(1) Q/Z
where η is the generic point of XFq . Note that the exactness

L at the right in the top sequence
is due to the fact (see [GS06, Corollary 6.5.4]) that x∈(P1 )(1) Q/Z → Q/Z is given
Fq
by summation in Q/Z. Since Br(P1Fq ) = 0 by [Gro68c, Remarques 2.5.b], we see that

Br(XFq ) ⊆ Br(η) is the subgroup consisting of classes ramified only at 0, 1, ∞. Using the
top row of the diagram, it follows that Br(XFq ) fits into an exact sequence
M
0 → Br(XFq ) → Q/Z → Q/Z → 0,
0,1,∞
REFERENCES 32
from which it follows that

Br(XFq ) ∼
= ρ̃ ⊗ Q/Z.
S3 S3 ∼
By Lemma 5.3, we find that ℓ Br(XFq ) = 0 if ℓ ≥ 5 and 3 Br(XFq ) = Z/3. This proves
the theorem.
Remark 11.2. We can again be more specific about the Azumaya algebras representing the
classes in Br(MFq ) with q odd. Let T be a Fq -scheme (or stack) with an elliptic curve E of
discriminant ∆. Let χm be the pullback of a character of the Galois extension Fqm of Fq
to T . We claim that there is a generator a ∈ Br(MFq ) such that the pullback of a to Br(T )
agrees with [(χ12 , ∆)]. Informally, a = [(χ12 , ∆)] in the universal case T = MFq , where
more precisely we should replace here ∆ by the µ12 -torsor corresponding to ∆ ∈ Γ(λ⊗12 )
via Construction 2.8.
First we consider the 3-torsion. The proof of Lemma 7.2 applies here to show that
[(χ3 , ∆)] is indeed the pullback of a generator of Br(MFq )[3]. Moreover, Proposition 9.6
implies that the√unique 2-torsion element 6a ∈ Br(MFq ) pulls back to [(χ2 , ∆)] as Fq2
agrees with Fq [ x] for an arbitrary non-square x in Fq . As in Remark 9.8 we see that
6[(χ12 , ∆)] = [(χ2 , ∆)] 6= 0 and 4[(χ12 , ∆)] = [(χ3 , ∆)] 6= 0. Thus, [(χ12 , ∆)] is indeed a
generator of Br(MFq ) ∼ = Z/12.
Finally, we also treat the easier case of an algebraically closed base.
Proposition 11.3. Let k be an algebraically closed field of characteristic not 2. Then
Br(Mk ) = 0.
Proof. By Theorem 9.1, 2 Br(Mk ) = 0. As in the last proof we see that ℓ (Br(Mk )) ∼ =
S3
ℓ Br(X k ) for ℓ an odd prime. By Tsen’s theorem, Br(η) vanishes for η the generic point
of Xk . By [Gro68c, Proposition 2.1] we obtain that Br(Xk ) = 0.
Remark 11.4. In an earlier version of this paper we suggested using the GL2 (Z/3)-cover
M(3) → M 1 to determine Br(Mk ) also for algebraically closed fields k of characteristic
Z[ 3 ]
2. Combining this approach with a new idea, Minseon Shin has proved in the meanwhile
in [Shi19] that Br(Mk ) ∼
= Z/2 for such fields k.
References
[ABB14] A. Auel, M. Bernardara, and M. Bolognesi, Fibrations in complete intersections of quadrics,
Clifford algebras, derived categories, and rationality problems, J. Math. Pures Appl. (9) 102
(2014), no. 1, 249–291. ↑1, 2
[Ces17] K. Cesnavicius, Purity for the Brauer group, ArXiv preprint arXiv:1711.06456 (2017). ↑2.13
[CI05] D. Chan and C. Ingalls, The minimal model program for orders over surfaces, Invent. Math. 161
(2005), no. 2, 427–452. ↑1
[CSS97] G. Cornell, J. H. Silverman, and G. Stevens (eds.), Modular forms and Fermat’s last theorem,
Springer-Verlag, New York, 1997. Papers from the Instructional Conference on Number Theory
and Arithmetic Geometry held at Boston University, Boston, MA, August 9–18, 1995. ↑3
[dJ] A. J. de Jong, A result of Gabber, available at http://www.math.columbia.edu/~ dejong/. ↑1, 2
[DR73] P. Deligne and M. Rapoport, Les schémas de modules de courbes elliptiques, Modular functions
of one variable, II (Proc. Internat. Summer School, Univ. Antwerp, Antwerp, 1972), Springer,
Berlin, 1973, pp. 143–316. Lecture Notes in Math., Vol. 349. ↑4
[Den88] Ch. Deninger, A proper base change theorem for nontorsion sheaves in étale cohomology, J. Pure
Appl. Algebra 50 (1988), no. 3, 231–235. ↑2
[EdGT90] B. Edixhoven, A. de Groot, and J. Top, Elliptic curves over the rationals with bad reduction at
only one prime, Math. Comp. 54 (1990), no. no. 189, 413–419. ↑6.3
[Fuc70] L. Fuchs, Infinite abelian groups. Vol. I, Pure and Applied Mathematics, Vol. 36, Academic Press,
New York-London, 1970. ↑2
[Fuj02] K. Fujiwara, A proof of the absolute purity conjecture (after Gabber), Algebraic geometry 2000,
Azumino (Hotaka), Adv. Stud. Pure Math., vol. 36, Math. Soc. Japan, Tokyo, 2002, pp. 153–183.
↑2
33 REFERENCES
[FO10] W. Fulton and M. Olsson, The Picard group of M1,1 , Algebra & Number Theory 4 (2010), no. 1,
87–104. ↑1
[Gab81] O. Gabber, Some theorems on Azumaya algebras, The Brauer group (Sem., Les Plans-sur-Bex,
1980), Lecture Notes in Math., vol. 844, Springer, Berlin, 1981, pp. 129–209. ↑2.6, 2.13
[GK15] O. Gabber and S. Kelly, Points in algebraic geometry, J. Pure Appl. Algebra 219 (2015), no. 10,
4667–4680. ↑3
[GS06] P. Gille and T. Szamuely, Central simple algebras and Galois cohomology, Cambridge Studies in
Advanced Mathematics, vol. 101, Cambridge University Press, Cambridge, 2006. ↑2, 2, 2, 2, 3, 3,
11
[GQM11] P. Gille and A. Quéguiner-Mathieu, Formules pour l’invariant de Rost, Algebra Number Theory
5 (2011), no. 1, 1–35. ↑3
[Gro65] A. Grothendieck, Éléments de géométrie algébrique. IV. Étude locale des schémas et des mor-
phismes de schémas. II, Inst. Hautes Études Sci. Publ. Math. 24 (1965), 231. ↑6
[Gro68a] A. Grothendieck, Le groupe de Brauer. I. Algèbres d’Azumaya et interprétations diverses, Dix
Exposés sur la Cohomologie des Schémas, North-Holland, Amsterdam, 1968, pp. 46–66. ↑2, 2, 2,
2
[Gro68b] , Le groupe de Brauer. II. Théorie cohomologique, Dix Exposés sur la Cohomologie des
Schémas, North-Holland, Amsterdam, 1968, pp. 67–87. ↑2, 2
[Gro68c] , Le groupe de Brauer. III. Exemples et compléments, Dix Exposés sur la Cohomologie
des Schémas, North-Holland, Amsterdam, 1968, pp. 88–188. ↑2, 2, 3, 11, 11
[Har77] R. Hartshorne, Algebraic geometry, Springer-Verlag, New York-Heidelberg, 1977. Graduate Texts
in Mathematics, No. 52. ↑2, 2
[Har80] , Stable reflexive sheaves, Math. Ann. 254 (1980), no. 2, 121–176. ↑5
[Has48] H. Hasse, Arithmetische Bestimmung von Grundeinheit und Klassenzahl in zyklischen kubischen
und biquadratischen Zahlkörpern, Abh. Deutsch. Akad. Wiss. Berlin. Math.-Nat. Kl. 1948 (1948),
no. 2, 95 pp. (1950). ↑7
[KM85] N. M. Katz and B. Mazur, Arithmetic moduli of elliptic curves, Annals of Mathematics Studies,
vol. 108, Princeton University Press, Princeton, NJ, 1985. ↑4, 4.2, 4, 4, 4
[LMB00] G. Laumon and L. Moret-Bailly, Champs algébriques, Ergebnisse der Mathematik und ihrer
Grenzgebiete. 3. Folge., vol. 39, Springer-Verlag, Berlin, 2000. ↑2, 2
[Lem05] F. Lemmermeyer, Kronecker-Weber via Stickelberger, J. Théor. Nombres Bordeaux 17 (2005),
no. 2, 555–558. ↑7
[Lie08] M. Lieblich, Twisted sheaves and the period-index problem, Compos. Math. 144 (2008), no. 1,
1–31. ↑2
[Lie11] , Period and index in the Brauer group of an arithmetic surface, J. Reine Angew. Math.
659 (2011), 1–41. With an appendix by Daniel Krashen. ↑1, 3
[Lie15] , The period-index problem for fields of transcendence degree 2, Ann. of Math. (2) 182
(2015), no. 2, 391–427. ↑
[MS16] A. Mathew and V. Stojanoska, The Picard group of topological modular forms via descent theory,
Geom. Topol. 20 (2016), no. 6, 3133–3217. ↑1
[Mil80] J. S. Milne, Étale cohomology, Princeton Mathematical Series, vol. 33, Princeton University Press,
Princeton, N.J., 1980. ↑2, 2, 2, 2
[Mum65] D. Mumford, Picard groups of moduli problems, Arithmetical Algebraic Geometry (Proc. Conf.
Purdue Univ., 1963), Harper & Row, New York, 1965, pp. 33–81. ↑1
[Neu99] J. Neukirch, Algebraic number theory, Grundlehren der Mathematischen Wissenschaften, vol. 322,
Springer-Verlag, Berlin, 1999. Translated from the 1992 German original and with a note by
Norbert Schappacher; With a foreword by G. Harder. ↑2, 2, 8, 9
[NSW00] J. Neukirch, A. Schmidt, and K. Wingberg, Cohomology of number fields, Grundlehren der Mathe-
matischen Wissenschaften [Fundamental Principles of Mathematical Sciences], vol. 323, Springer-
Verlag, Berlin, 2000. ↑2
[PS92] R. Parimala and V. Srinivas, Analogues of the Brauer group for algebras with involution, Duke
Math. J. 66 (1992), no. 2, 207–237. ↑2
[Rom05] M. Romagny, Group actions on stacks and applications, Michigan Math. J. 53 (2005), no. 1,
209–236. ↑4, 4
[Sal99] D. J. Saltman, Lectures on division algebras, CBMS Regional Conference Series in Mathematics,
vol. 94, Published by American Mathematical Society, Providence, RI, 1999. ↑2
[Ser73] J.-P. Serre, A course in arithmetic, Springer-Verlag, New York-Heidelberg, 1973. ↑10
REFERENCES 34
[Ser97] J.-P. Serre, Galois cohomology, Springer-Verlag, Berlin, 1997. Translated from the French by
Patrick Ion and revised by the author. MR1466966 ↑5
[SGA4.3] Théorie des topos et cohomologie étale des schémas. Tome 3, Lecture Notes in Mathematics, Vol.
305, Springer-Verlag, Berlin, 1973. Séminaire de Géométrie Algébrique du Bois-Marie 1963–1964
(SGA 4); Dirigé par M. Artin, A. Grothendieck et J. L. Verdier. Avec la collaboration de P.
Deligne et B. Saint-Donat. ↑2
[Shi19] M. Shin, The Brauer group of the moduli stack of elliptic curves over algebraically closed fields
of characteristic 2, J. Pure Appl. Algebra 223 (2019), no. 5, 1966–1999. ↑1, 11.4
[Sil09] J. H. Silverman, The arithmetic of elliptic curves, 2nd ed., Graduate Texts in Mathematics,
vol. 106, Springer, Dordrecht, 2009. ↑4, 7
[TL] The LMFDB Collaboration, The L-functions and Modular Forms Database, available at
http://www.lmfdb.org. Accessed: 26 May 2016. ↑10
[TS] The Stacks Project Authors, Stacks Project. Accessed: 17 October 2017. ↑3
[Wei94] C. A. Weibel, An introduction to homological algebra, Cambridge Studies in Advanced Mathe-
matics, vol. 38, Cambridge University Press, Cambridge, 1994. ↑5

The Brauer Group of The Moduli Stack of Elliptic Curves: Benjamin Antieau and Lennart Meier

Uploaded by

Copyright:

Available Formats

The Brauer Group of The Moduli Stack of Elliptic Curves: Benjamin Antieau and Lennart Meier

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

The Brauer Group of The Moduli Stack of Elliptic Curves: Benjamin Antieau and Lennart Meier

Uploaded by

Copyright:

Available Formats

The Brauer group of the moduli stack of elliptic curves

Benjamin Antieau∗ and Lennart Meier†

Mathematics Subject Classification 2010. Primary: 14F22, 14H52, 14K10.

2 Brauer groups, cyclic algebras, and ramification 4

3 The low-dimensional Gm -cohomology of BCm 10

4 A presentation of the moduli stack of elliptic curves 13

5 Beginning of the computation 17

6 The p-primary torsion in Br(MZ[ 12 ] ) for primes p ≥ 5 21

7 The 3-primary torsion in Br(MZ[ 1 ] ) 22

8 The ramification of the 3-torsion 24

9 The 2-primary torsion in Br(MZ[ 12 ] ) 26

10 The Brauer group of M 30

11 The Brauer group of M over Fq with q odd 31

Theorem 1.1. (1) Br(Mk ) = 0 if k is an algebraically closed field of characteristic not

X = P1 1 − {0, 1, ∞} = Spec Z[ 21 , t±1 , (t − 1)−1 ],

and the universal Legendre curve over X is defined by the equation

y 2 = x(x − 1)(x − t).

Br(Q2 ) o Br(M 1 ) (1.4)

Acknowledgments. We thank the Hausdorff Research Institute for Mathematics, UIC,

2 Brauer groups, cyclic algebras, and ramification

δ : H1 (X, PGLn ) → H2 (X, Gm ).

is even Zariski-locally equivalent to Mn (OX ). The class of A in H1 (X, PGLn ) is the

2. If A and B are Azumaya algebras on X, then A ⊗OX B is an Azumaya algebra.

(iii) If X is a regular and noetherian Deligne-Mumford stack, then H2 (X, Gm ) is torsion,

0 / Hi−1 (X, Gm )/pm / Hi (X, µpm ) / Hi (X, Gm )[pm ] /0

Inductively, using the A1 -invariance of Hi−1 (X, Gm ) if i ≤ 2 or of p Hi−1 (X, Gm ) if i > 2

H1pl (X, Cn ) × H1pl (X, µn ) → H2pl (X, µn ) → H2pl (X, Gm )

0 → Gm (X)/n → H1pl (X, µn ) → Pic(X)[n] → 0. (2.7)

It remains to construct the splitting. By Prüfer’s theorem [Fuc70, Theorem 17.2],

eχ,u be the coproduct

in the category of sheaves of quasi-coherent (associative and unital) OX -algebras. Finally,

Example 2.19. (1) Br(Z) = 0.

3 The low-dimensional Gm -cohomology of BCm

with differentials dr of bidegree (r, 1 − r).

Figure 1: The E2 -page of the Leray–Serre spectral sequence computing Hi (BCn , Gm ).

for the fppf cohomology is an isomorphism; thus the comparison map

Hi (BCn , Gm ) → Hipl (BCn , Gm )

Indeed, Ripl c∗ Gm is the fppf-sheafification of U 7→ Hi (BCn,U , Gm ) and in the Leray–Serre

and short exact sequences

s : H1pl (X, µn ) → H2 (BCn,X , Gm )

H1pl (Bpl µn,R , Gm ) ∼

Gm (S) → Gm (µn,R ) → Gm (µn,R ×Spec R µn,R ) → · · ·

0 → H1pl (Spec Z, µn ) → H1pl (Bpl µn , µn ) → H0 (Spec Z, Cn ) → 0. (3.5)

4 A presentation of the moduli stack of elliptic curves

Lemma 4.3. The map M(2) → M 1 is an S3 -Galois cover.

(S3 )Spec R = S3 × Spec R → Spec R ×M 1

C2 × Spec R ≃ X ×M(2) Spec R.

y 2 = x(x − p)(x − q).

As g ∗ λ for g ∈ S3 on M(2) is canonically isomorphic to λ (as this is pulled back from

to the trivial representation, the adjoint to the identity.

Proof. Denote by π : XS → S the structure map. We have a map f : Z2 ⊕ Gm → π∗ Gm,XS

5 Beginning of the computation

Lemma 5.2. Let M be a trivial S3 -module. Then,

Proof. We use the Lyndon–Hochschild–Serre spectral sequence for

Proof. There is a short exact sequence of S3 -modules

In the associated long exact sequence in cohomology, note that Hi (S3 , ρ ⊗ M ) ∼

is the transfer. We obtain short exact sequences

0 → coker trSC32 (Hi−1 ) → Hi (S3 , ρ̃ ⊗ M ) → ker trSC32 (Hi ) → 0.

Because C2 → S3 has a retraction, the restriction map Hi (S3 , M ) → Hi (C2 , M ) is the