Section 8.5 · PDF pages 475–479
Summing the character estimates
The notation transcription is provisional. Check it against the printed mathematics when citing or reusing a formula.
8.5 Summing the character estimates
The source fixes
B_{ψ} = {3} ∪{odd prime divisors of o m p_{K} h_{K} p_{1}}.
Here p_{K} ≥7 is prime, congruent to 3 (mod 4) and prime to m; h_{K} is the class number of ℚ(^{√}−p_{K}); and p_{1} ∤ 2mp_{K} is prime with p_{1} ≡ 1 (mod o_{odd}). All are fixed once ψ is fixed. Assume q, d exceed these fixed integers and 2t + 1. Let X_{e} consist of characters ϕ modulo N, trivial on the host subgroup, with ϕ(−1) = (−1)^{e}. There are at most N such characters, and at most b characters whose conductors divide b.
For the first bound, distinct multiplicative characters are linearly independent as functions over the fraction field of O, a space of dimension N. For the second they are induced from modulus gcd(N, b).
The required character reduction, with n = [E : ℚ], is
Selectable PDF notation • transcription unchecked
s^{+}_{ℓ}(E, e) ≤ ∑ v_{ℓ}(e, ϕ) + Nℒ_{ℓ}(N, n, e), (204)
ϕ∈Xe
ℒ_{ℓ}(N, n, e) = (n + n^{2})ν_{ℓ}(N, e) + ω(N)(n + 1)v_{ℓ}(n).If ℓ∤n, the entire loss term may be omitted. This is an inequality of extended lengths before finiteness. For
Selectable PDF notation • transcription unchecked
N ≥2, n ≤N, e ≥1,
ℒ_{ℓ}(N, n, e) ≤(Ne)^{6}. (205)Indeed, with M = Ne ≥ 2, the left side is at most 3M^{2}+ 3M^{3} ≤ 6M^{3} ≤ M^{6}, using ν_{ℓ} ≤ 2M and
Selectable PDF notation • transcription unchecked
n, ω(N), v_{ℓ}(n) ≤M.Lemma 8.10 (Finite abelian obstruction). Let Q be a finite abelian group acting by scalars on an O-module M with finite-length cyclic submodules. Then
length_{O} H^{2}(Q, M) ≤|Q|^{2} length_{O} M^{Q}.
If U ⊴G is open in a profinite group and G/U is abelian, the cokernel of
Selectable PDF notation • transcription unchecked
H^{1}(G, M) −→H^{1}(U, M)^{G/U}has length at most [G : U]^{2} length_{O} M^{G}.
Proof. Write c(g) for the scalar action and let a generate the ideal of all c(g) −1. For a 2-cocycle F, define β_{g}(x) = F(g, x) −F(x, g). The three cocycle identities at (g, x, y), (x, g, y), (x, y, g), and commutativity of Q, give
Selectable PDF notation • transcription unchecked
dβ_{g}(x, y) = (c(g) −1)F(x, y).Thus aZ^{2}⊆B^{2}, and H^{2}(Q, M) is a quotient of Z^{2}/aZ^{2}. Every cyclic submodule of Z^{2} has finite length, because evaluation embeds it into a finite product of cyclic submodules of M. The cokernel-versus-kernel argument in Lemma 8.7 applies to multiplication by a on Z^{2}. Moreover evaluation embeds Z^{2}[a] into (M[a])^{Q×Q}= (M^{Q})^{Q×Q}. This proves the first assertion.
For completeness the transgression used in the second assertion is the obstruction to extending an invari-ant cocycle on U. Choose a section s : G/U →G, which is continuous because the quotient is finite and discrete. Invariance of the class allows the conjugated cocycle at each s(q) to be corrected by a cobound-ary. Extend these choices to a 1-cochain on G using g = us(q). Its failure to satisfy the cocycle identity depends only on the two quotient coordinates and takes values in M^{U}; the identity d^{2} = 0 makes it a 2-cocycle. Changing the corrections changes it by a 2-coboundary. If the resulting class is zero, subtracting a lifted 1-cochain on the quotient removes the failure, so the original class is the restriction of a cocycle on G. This gives an injection of the restriction cokernel into H^{2}(G/U, M^{U}). Apply the first assertion to M^{U}; its quotient invariants are M^{G}, and its cyclic submodules retain finite length.
Source: Development/AutomorphicArithmetic/AutomorphicVarieties/GaloisLibrary/SelmerTransport/LengthSpanCocycleFinite.lean:28--118; Development/AutomorphicArithmetic/AutomorphicVarieties/GaloisLibrary/SelmerTransport/TransgOpen.lean:282--362;Development/AutomorphicArithmetic/AutomorphicVarieties/GaloisLibrary/SelmerTransport/CokerAb.lean:7--29
Lemma 8.11 (Adding local splitting conditions). Let ϕ^{n}_{G}= 1, and e ≥1. Adding the splitting conditions at primes dividing N, to the unramified condition and the splitting condition at ℓ, decreases the O-length by at most ω(N)v_{ℓ}(n)e_{O}.
Proof. At p ̸= ℓ, the kernel of restriction from a decomposition group to inertia is the unramified H^{1}:
Selectable PDF notation • transcription unchecked
M_{e,ϕ}^{Ip}/(F_{p} −1)M_{e,ϕ}^{Ip},where F_{p} is arithmetic Frobenius. This formula follows directly by evaluating a cocycle of the procyclic unramified quotient at F_{p}; a continuous cocycle is determined by that value, and coboundaries change it by (F_{p} −1)m. If ϕ_{G} is nontrivial on inertia, select an inertia value ζ ̸= 1. It satisfies ζ^{n}= 1; the invariants are annihilated by ζ −1. The identity
Selectable PDF notation • transcription unchecked
n∑−1
n = (1 −ζ^{i})
i=0shows that ζ −1 divides n, so these invariants have length at most v_{O}(n) = e_{O}v_{ℓ}(n). Their quotient has the same upper bound. If the finite character is trivial on inertia, M_{e,ϕ}^{Ip} = D_{O}. Frobenius acts by α = ϕ_{G}(F_{p})^{−1}p^{−e} ̸= 1: equality would imply p^{en} = 1, which is impossible for e, n ≥1. Multiplication by the nonzero element α −1 is surjective on D_{O}, so the unramified H^{1} vanishes. Finally map a global unramified class to these local unramified groups, one for each p | N, p ̸= ℓ. The kernel imposes exactly the extra splitting conditions. Additivity of length gives the asserted sum of local upper bounds. Conjugate prolongations give equivalent local tests by the cocycle identity, so one representative at each prime suffices.
Source: Development/AutomorphicArithmetic/AutomorphicVarieties/CharacterLevelBounds/LocUnr.lean:38--141; Development/AutomorphicArithmetic/AutomorphicVarieties/CharacterLevelBounds/VLe.lean:9--60
Proposition 8.12 (Character reduction with its degree loss). The inequalities (204) hold. Their loss term can be omitted when ℓ∤n.
Proof. Let G = (ℤ/N)^{×}/H = Gal(E/ℚ), and enlarge O to contain all |G|-th roots of unity. The nor-malised lengths do not change: a further finite coefficient extension multiplies both the numerator and e_{O} by its ramification index. Conjugation makes 𝒮_{ℓ}(E, e) ⊗_{ℤℓ}O a G-module. Twisting that action by ϕ^{−}_{G}^{1} identifies its ϕ-eigenspace with the G-invariants of the Selmer group on G_{E} with coefficients M_{e,ϕ}. This follows at cochain level from
Selectable PDF notation • transcription unchecked
(g · z)(u) = g z(g^{−1}ug);the extra scalar in the twisted action is precisely ϕ_{G}(g)^{−1}. Coefficient extension multiplies its length by e_{O}, and the positive eigenspace consists of the characters with ϕ(−1) = (−1)^{e}.
We may bound a finite abelian module by the sum of its character eigenspace lengths even when ℓ | |G|. For an operator T with T^{r} = 1, ∏_{ζ}r_{=1}(T −ζ) = 0. The elementary inequality length ker(AB) ≤ length ker A+length ker B therefore gives length M ≤∑_{ζ}r_{=1} length ker(T −ζ). Decompose G as a product of finite cyclic groups and apply this inequality successively; the operators commute, so each successive eigenspace is stable under the others. Applied to the positive eigenspace, only the characters in X_{e} survive. Opposite signs at −1 have zero intersection, because their difference is 2, a unit of O.
We bound the eigenspace for one ϕ. By Lemma 8.10, the obstruction to descending its invariant coho-mology class from G_{E} to G_{ℚ} costs at most n^{2}e_{O}ν_{ℓ}(N, e). For classes which descend, the local conditions can fail over ℚ only at ℓ and the primes dividing N: at every other prime, inertia already lies in G_{E}. At ℓ, restriction from a decomposition group to its intersection with G_{E} has kernel of length at most ne_{O}ν_{ℓ}(N, e), by Lemma 8.7 and the local version of Lemma 8.8; the proof of that lemma used inertia and hence gives that
local version unchanged. At p | N, p ̸= ℓ, the cyclotomic character is trivial on inertia. The finite quotient of inertia has order dividing n. If its scalar action is nontrivial, its invariant module has length at most e_{O}v_{ℓ}(n), as in Lemma 8.11, and the scalar restriction lemma bounds its kernel by n times this. If its action is trivial, a cocycle in the kernel is a homomorphism of that finite quotient and is determined by at most n values in D_{O}[n], giving the same bound. The sum of these local losses is at most ω(N)ne_{O}v_{ℓ}(n).
Thus the eigenspace length is at most the length of 𝒮_{ℓ}(M_{e,ϕ}) plus
Selectable PDF notation • transcription unchecked
e_{O}{(n + n^{2})ν_{ℓ}(N, e) + ω(N)nv_{ℓ}(n)}.Lemma 8.11 changes this group to 𝒮_{ℓN}(M_{e,ϕ}) at an additional cost e_{O}ω(N)v_{ℓ}(n). Sum over the at most N characters in X_{e} and divide by e_{O}. This is (204).
When ℓ ∤ n, all finite-quotient cohomology in positive degree vanishes by averaging over the finite quotient, because n is a unit in O. Restriction consequently identifies the global group with the invariant subgroup over E, and the same averaging establishes this identification for the local conditions. The extra-splitting loss also vanishes since v_{ℓ}(n) = 0. The preceding character sum therefore has no loss term.
Source: Development/AutomorphicArithmetic/AutomorphicVarieties/CharacterLevelBounds/ReductionToDirichletCharAnyLevelOfLe.lean:30--106; Development/AutomorphicArithmetic/AutomorphicVarieties/HeightAndVanishing/RedAll.lean:6--21; Development/AutomorphicArithmetic/AutomorphicVarieties/HeightAndVanishing/RedPrime.lean:11--57
Proposition 8.13 (Primes in the exceptional set). Assume (204). Fix ℓ∈B_{ψ}, a layer height j, and the layer bound B ≥0 of Proposition 8.9. Suppose the other characters have finite lengths with
log ℓv_{ℓ}(e, ϕ) ≤(ℓNe)^{b} (1 ≤e < t, ϕ ∈X_{e})
for a fixed b. There is then C, depending only on ψ, ℓ, j, b, such that
Selectable PDF notation • transcription unchecked
∑
log ℓ s^{+}_{ℓ}(E, e) ≤(Nt)^{C} + CB,
Selectable PDF notation • transcription unchecked
1≤e<t
and all lengths on the left are finite.
Proof. Put M = Nt, L = log ℓ, a_{4} = ℓmℓ^{j}+3. Let 𝒪 be the union of characters covered by Proposition 8.9. Its proof gives cond ϕ | m, hence |𝒪| ≤m; each such character contributes at most B +M^{a4} when summed over e. For the other characters (ℓNe)^{b}≤M^{2b}. Splitting the double sum and enlarging nonnegative sums gives ∑ ∑
Selectable PDF notation • transcription unchecked
L v_{ℓ}(e, ϕ) ≤Lm(B + M^{a4}) + tNM^{2b}.
e ϕ∈XeCharacter reduction adds at most LNtM^{6}. The total is at most LmB + LmM^{a4}+ M^{2b+1}+ LM^{7}. Here M ≥3, L ≤ℓ≤M, m ≤M. Choose C = a_{4} +2b+9+ℓm. The first term is at most CB; each remaining term is at most M^{C−1}, and 3M^{C−1}≤M^{C}. Finiteness follows from the character bounds and the extended length inequality.
Source: Development/AutomorphicArithmetic/AutomorphicVarieties/HeightAndVanishing/FinalLe.lean:86--263
Proposition 8.14 (Primes outside the exceptional set). Assume the loss-free character reduction when ℓ∤n. Suppose for each 1 ≤e < t the primes outside B_{ψ} with ℓ−1 ∤2e satisfy the collective bound
Selectable PDF notation • transcription unchecked
∑ ∑
log ℓ v_{ℓ}(e, ϕ) ≤(Nt)^{C1}
Selectable PDF notation • transcription unchecked
ℓ∈T ϕ∈Xe
for every finite set T of those primes. For the remaining primes suppose the individual characters have finite lengths with log ℓv_{ℓ}(e, ϕ) ≤(ℓNe)^{K}. Then every finite set T of odd primes outside B_{ψ} satisfies
Selectable PDF notation • transcription unchecked
∑ ∑
log ℓ s^{+}_{ℓ}(E, e) ≤(Nt)^{C1+3K+4}.
Selectable PDF notation • transcription unchecked
ℓ∈T 1≤e<t
Proof. The host subgroup has index dividing 2^{a+2}o: its four component indices divide 2^{a}, o, 2, 2, and the intersection index divides their product in an abelian group. Thus every ℓ∈/B_{ψ} is prime to n = [E : ℚ], so character reduction has no loss. Put M = Nt. For a fixed e, a prime with ℓ−1 | 2e has ℓ≤2e + 1 < 2t. There are at most 2t such primes, at most N characters at each, and ℓNe ≤ 2Mt ≤ M^{3}, since M ≥ 2. Combining the collective and individual estimates gives
Selectable PDF notation • transcription unchecked
∑ ∑
log ℓ v_{ℓ}(e, ϕ) ≤M^{C1}+ 2tNM^{3K}= M^{C1}+ 2M^{3K+1}.
Selectable PDF notation • transcription unchecked
ℓ∈T ϕ∈Xe
Sum over at most t ≤M values of e. The result is at most M^{C1+1}+2M^{3K+2}≤M^{C1+3K+4}. The loss-free reduction gives the asserted bound and finiteness.
Source: Development/AutomorphicArithmetic/AutomorphicVarieties/HeightAndVanishing/SumOut.lean:50--231
Suppose now that the inside estimate holds for each ℓ∈B_{ψ}, with the same layer bound B, and that the outside estimate holds. Let D dominate the finitely many inside exponents, k = |B_{ψ}|, and C_{0} the outside exponent. Splitting any finite prime sum at B_{ψ} bounds it by k(M^{D}+ DB) + M^{C0}. For M ≥ 2, set C = C_{0} + D + k + kD. Bernoulli’s inequality gives M^{k} ≥k + 1, whence M^{C} ≥kM^{D}+ M^{C0}; also C ≥kD. Thus the sum is at most M^{C}+ CB, the exact plus estimate in Proposition 8.4.
Source: Development/AutomorphicArithmetic/AutomorphicVarieties/HeightAndVanishing/PointArith.lean:4--92