Skip to content

Section 8.14 · PDF pages 501–522

The cycles and the finite-difference construction

The notation transcription is provisional. Check it against the printed mathematics when citing or reusing a formula.

8.14 The cycles and the finite-difference construction

Lemma 8.43 (The Frobenius density consequence). Let L be a number field and S a finite set of primes. The conjugacy closure of chosen lifts of arithmetic Frobenius at primes outside S is dense in G_{L}. Consequently two continuous class functions from G_{L} to a Hausdorff space that agree at those Frobenius elements agree everywhere.

Proof. A nonempty open set contains a coset of an open normal subgroup. In the corresponding finite Galois quotient, apply the Chebotarev theorem cited in Lemma 8.32 to that coset’s conjugacy class, excluding S and the finitely many ramified primes. The chosen Frobenius lift has a conjugate in the prescribed coset. For the last assertion the equality locus is closed by the Hausdorff property, and contains this dense set.

We now give the chain operations in the arithmetic cycle construction. They explain why its eventual value has the factor p_{E} in (213). For a group G, write C_{n}(G, R) = R[G^{n}] for the inhomogeneous bar chains with trivial coefficients, with differential

Selectable PDF notation • transcription unchecked
n∑−1
∂[g_{1}, . . . , g_{n}] = [g_{2}, . . . , g_{n}] + (−1)^{j}[g_{1}, . . . , g_{j}g_{j+1}, . . . , g_{n}] + (−1)^{n}[g_{1}, . . . , g_{n−1}].
j=1

Pushing a tuple through a group homomorphism commutes with ∂, and ∂^{2} = 0: each pair of deleted or merged positions occurs in the two possible orders with opposite signs.

Lemma 8.44 (The cyclic bar relation). Let q = ℓ^{ν} > 1, R = ℤ/q, and let C_{q} be the cyclic group written additively. For E ≥1,

Selectable PDF notation • transcription unchecked
z = ∑ [j_{0}, 1, j_{1}, 1, . . . , j_{E−1}, 1] ∈C_{2E}(C_{q}, R)
j_{0},...,j_{E−1}∈R

is a cycle. For the 2E + 2 maps d_{t} : C_{q} →C_{q}^{2}, d_{t}(a) = (a, ta), there are n_{t}∈R, one of them a unit, and a chain b with

Selectable PDF notation • transcription unchecked
2∑E+1
n_{t}(d_{t})_{#}z = ∂b.
t=0

Proof. In the boundary of z, the first face is repeated q times as j_{0} varies, and is zero in R. Each subsequent adjacent pair of faces cancels after translating one of the summation variables by one; at the final pair this is the translation of j_{E−1}. The signs of each pair are opposite. Thus ∂z = 0.

The tensor product of the two standard periodic resolutions of a cyclic group computes the homology of C_{q}^{2}. In each periodic resolution the maps alternate between g −1 and 1 + g + · · · + g^{q−1}. Its exactness follows by comparing coefficients in the group ring: the kernel of g−1 consists of constant coefficients, and a coefficient vector of sum zero is a (g−1)-multiple. After trivial R-coefficients, both maps vanish. In degree 2E, their tensor resolution is consequently free of rank 2E + 1. The usual lifting of basis vectors along the surjections in a free resolution constructs chain comparison maps with the bar resolution, so H_{2E}(C_{q}^{2}, R) ≃

Selectable PDF notation • transcription unchecked
_{R}2E+1_{.}

The 2E + 2 cycle classes therefore have a relation with some unit coefficient. Explicitly lift their coor-dinate matrix to ℤ_{ℓ}, solve its underdetermined homogeneous linear system over ℚ_{ℓ}, and rescale a nonzero solution so its coordinates are integral and at least one is a unit. Reducing modulo q gives the required relation. A zero homology class is, by definition, a boundary, giving b.

Source: Development/AutomorphicArithmetic/AutomorphicVarieties/ModularityArgument/ClassCounting/FirstFace.lean:7--82; Development/EtaleCohomology/CyclotomicDiscriminants/SRelTPf.lean:270--314; Development/EtaleCohomology/CyclotomicDiscriminants/FinSRel.lean:4--6

Lemma 8.45 (Prisms and the integral stabiliser cycle). Define

Selectable PDF notation • transcription unchecked
∑n
P_{g}[h_{1}, . . . , h_{n}] = (−1)^{i}[h_{1}, . . . , h_{i}, g, g^{−1}h_{i+1}g, . . . , g^{−1}h_{n}g].
i=0

Then ∂P_{g} +P_{g}∂= (conj_{g}−1)_{#} −1. Suppose ρ : D →G, H ≤G, ρ(D) ⊂H, and a relation ∑_{t}n_{t}(d_{t})_{#}z = ∂b holds. Suppose g_{t} commutes with ρd_{t}(C_{q}), g_{t}^{−}_{0}^{1}ρ(D)g_{t0}⊂H, and g_{t}^{−1}g_{t0}∈H. There is a cycle y in H, one degree above z, whose image in G, modulo boundaries, is ∑_{t}n_{t}P_{gt}(ρd_{t})_{#}z.

Proof. Expanding the displayed prism differential cancels all inner faces in adjacent pairs. The two outer faces are the conjugated tuple and the original tuple, with the stated signs. For composition, the two succes-sive prisms and the prism of their product differ by a boundary on cycles. To check this without assuming functoriality of a chosen homotopy, use homogeneous bar tuples and the three vertex maps v →→v, v →→vg, v →→vgg^{′}. Sum the prisms with two switching positions 0 ≤i ≤j ≤n, with sign (−1)^{i+j}. Its boundary cancels all internal switches and leaves the three edge prisms, proving the claimed composition identity.

Let ρ_{H} be ρ with values in H, σ_{H}(x) = g_{t}^{−}_{0}^{1}ρ(x)g_{t0}, and k_{t} = g_{t}^{−}_{0}^{1}g_{t} ∈H. Use the explicit chain

Selectable PDF notation • transcription unchecked
y = (σ_{H})_{#}b −(ρ_{H})_{#}b +∑n_{t}P_{kt}(σ_{H}d_{t})_{#}z.
t

The commutation assumption gives k_{t}^{−1}σ_{H}(d_{t}(a))k_{t} = ρ_{H}(d_{t}(a)). The prism identity therefore makes its boundary zero: the two pushed-forward copies of ∂b cancel the two terms from the prism sum. By the composition identity, choose e_{t} with

∂e_{t} = P_{gt}v_{t} −P_{gt0}v_{t} −P_{kt}(conj_{g}−1)_{#}v_{t}, v_{t} = (ρd_{t})_{#}z.

Selectable PDF notation • transcription unchecked
t_{0}

The boundary of P_{gt0}ρ_{#}b−∑_{t}n_{t}e_{t} is exactly the image of y minus ∑_{t}n_{t}P_{gt}v_{t}. This supplies the boundary witness, as well as the cycle.

Source: Development/AutomorphicArithmetic/AutomorphicVarieties/ModularityArgument/ClassCounting/LatticeBlockClasses/CChain.lean:8--63; Development/AutomorphicArithmetic/AutomorphicVarieties/ModularityArgument/ClassCounting/ToInhPartialProd.lean:9--129; Development/AutomorphicArithmetic/AutomorphicVarieties/ModularityArgument/ClassCounting/SPrismMul.lean:194--205

Lemma 8.46 (The line-selecting difference). Let M < ℓ^{c}, c ≤ ν, κ = 1 + ℓ^{c}, R = ℤ/ℓ^{ν}, and t_{0} ∈ {0, . . . , M}. Put Q(T) =∏^{M}_{j=0}^{−1}(T −κ^{j}). There is a finitely supported integer function η on R^{2} whose pushforward by p_{t}(x, y) = x + ty is zero for t ̸= t_{0}, and whose pushforward at t_{0} is

Selectable PDF notation • transcription unchecked
∑M
[T^{i}]Q(T) [κ^{i}].
i=0

Proof. Work in the additive group rings, with multiplication [a][b] = [a + b]. Let S = {0, . . . , M} \ {t_{0}} and ∏

Selectable PDF notation • transcription unchecked
D_{0} = ([t_{0} −t] −1), F_{0} =∑[T^{i}]Q(T)[κ^{i}].
t∈S i

For 0 < |t_{0} −t| < ℓ^{c}, the subgroup generated by its residue contains ℓ^{c}; the geometric sum identity in the group ring shows that [t_{0} −t] −1 divides [ℓ^{c}] −1. Thus D_{0} | ([ℓ^{c}] −1)^{M}. Also F_{0} is divisible by this last power. For the latter assertion, write κ^{i}= 1 + ℓ(^{c}a)_{i} and expand [κ^{i}] = [1](1 + X)^{ai}, X = [ℓ^{c}] −1, modulo X^{M}. For j < M, the coefficient ∑_{i}[T^{ i}]Q(T)^{a}_{j}^{i} is zero: the binomial is a rational polynomial of degree

j in κ^{i}, and ∑_{i}[T^{ i}]Q(T)κ^{ih} = Q(κ^{h}) = 0 for h < M. These are integral coefficients, so the equality over ℚ is the required integer equality. Hence F_{0} = D_{0}A in ℤ[R].

Let ι([a]) = [(a, 0)]. The coefficients of

Selectable PDF notation • transcription unchecked
∏([(0, 1)] −[(t, 0)]) ι [−∑t]A)
t∈S t∈S

give η. Under p_{t}, t ̸= t_{0}, one factor is [t]−[t] = 0. Under p_{t0}, that product becomes [∑_{t∈S}t]D_{0}, cancelling the inserted inverse basis element and leaving F_{0}.

Source: Development/AutomorphicArithmetic/AutomorphicVarieties/ModularityArgument/ClassCounting/ExpStep.lean:111--182

Lemma 8.47 (The finite-field functional). Let k be a finite field containing a primitive q-th root ζ, q = ℓ^{ν}> 1, and let v : k^{×}→ℤ/q be additive on products. For E ≥1 there are boundary-vanishing functionals θ_{V} in degree 2E + 1, for every finite-dimensional k-space V , natural under isomorphisms and additive under direct sums, such that ( )

θ_{V} P_{g}(scal_{ζ}x)_{#}z = x^{E}v(det g).

Their pullbacks by reduction of integral matrices form the primitive family required in Proposition 8.33.

Proof. Write k^{×}= C_{h}, and choose its generator so that ζ is its h/q-th power; a primitive root admits such a generator by the elementary unit-reduction map (ℤ/h)^{×}→(ℤ/q)^{×}. On C_{h} let

Selectable PDF notation • transcription unchecked
⌊̃a +̃b⌋

u(a, b) = (mod q),

Selectable PDF notation • transcription unchecked
h

where tildes are representatives between 0 and h −1. Associativity of carrying proves du = 0. The cup cochain t = u^{E}⌣v is a (2E + 1)-cocycle, hence kills boundaries. The prism at a central element is a slant

operator: expanding its insertions in a cup product gives the graded derivation rule on cycles. The slant of u is zero, because its two inserted terms u(g, a) and u(a, g) are equal; the slant of v is v(g). Thus its value on the prism of the scalar image of z is v(g) times the value of the pulled-back u^{E} on z. Pullback along a →→ζ^{xa} sends u to xu_{q} modulo a coboundary, by comparing carries before and after multiplication by x. Finally u^{E}_{q}(z) = 1: in each independent pair (j, 1), precisely j = q −1 contributes a carry. The scalar prism value is therefore x^{E}v(g).

To extend this scalar cocycle to vector spaces, use transfer from a line stabiliser:

[θ_{V} ] = cor^{GL}_{PV}^{(V )}λ^{∗}_{V}[t],

where P_{V} stabilises a line and λ_{V} is its character on that line. Transfer here is a finite coset sum: lift a bar tuple successively along the chosen coset representatives, evaluate the pulled scalar cocycle, and sum over the initial coset. The boundary terms telescope, so it is a cochain map; changing representatives adds a prism coboundary. For dimension zero use zero. In dimension one it is the original scalar class. This also proves isomorphism naturality.

We verify its direct-sum law, which is the reason for the coefficient condition q | |k|−1. Under the block action on V ⊕W, projective lines are those lying in V , those lying in W, and the mixed lines. The first two pieces give the transfers defining θ_{V} , θ_{W} . A mixed line projects to a line in V , with |k|^{dim W}−1 points in every fibre. Its stabiliser character on the mixed line is the pulled-back character on that projected line. Grouping the transfer’s finite coset sum by this projection multiplies its class by |k|^{dim W}−1, which is zero modulo q. Thus the mixed contribution is zero. This proves the direct-sum identity on cycles. Choosing cocycle representatives defines the functionals; changes by coboundaries have no effect on any cycle used here.

Fix x, and put F_{V} (g) = θ_{V} (P_{g}(scal_{ζ}x)_{#}z). Scalar matrices commute with g, so this prism is a cycle. The prism composition identity makes F_{V} a homomorphism to the additive target. The preceding properties make it invariant under isomorphisms and additive on blocks. It vanishes on elementary transvections: conjugating u(a) by diag(ζ, 1) gives u(ζa); the homomorphism identity implies F(u((ζ −1)a)) = 0, and ζ −1 ̸= 0. Elementary row reduction generates SL(V ) by these transvections. Therefore F_{V} (g) depends only on det g. Its value on diag(a, 1, . . . , 1) is the already computed scalar value x^{E}v(a). This proves the claimed formula. Finally reduction preserves boundaries, and adjoining a trivial block contributes zero in positive degree. Repeating a representation d times multiplies its value by d, by direct-sum additivity. These are exactly the three primitive-family properties.

Source: Development/AutomorphicArithmetic/AutomorphicVarieties/ModularityArgument/ClassCounting/SThetaOf.lean:192--275; Development/AutomorphicArithmetic/AutomorphicVarieties/ModularityArgument/ClassCounting/STrans.lean:303--413; Development/AutomorphicArithmetic/AutomorphicVarieties/ModularityArgument/ClassCounting/FProd.lean:101--150

Lemma 8.48 (Lattice redistribution with prescribed determinant). For a number field F, write

Λ(a) = {x ∈F^{J} : ord_{P} (x_{j}) ≥a_{j,P} for every j, P},

where J is finite and the integral divisors a_{j} have finite support. If

Selectable PDF notation • transcription unchecked
∑

div(β) = (b_{j} −a_{j}), β ∈F^{×},

Selectable PDF notation • transcription unchecked
j

and J ̸= ∅ or β = 1, there is an F-linear isomorphism carrying Λ(a) onto Λ(b), with determinant exactly

Selectable PDF notation • transcription unchecked
β.

Proof. We give the determinant-one elementary move used in the source. For two divisors d, e, put C = ∏ P^{ dP} , D = ∏ P^{ eP} . Choose 0 ̸= p ∈ C. The integral ideal (p)C^{−1} has finite prime support. Choose q ∈D whose image in each D/PD, for those primes P, is nonzero. Such images exist because D_{P} is a free rank-one module over the discrete valuation ring O_{F,P} ; the Chinese remainder theorem on D chooses them simultaneously. Therefore (p)C^{−1}+ (q)D^{−1}= O_{F} . Writing 1 = ps + qr^{′}, with s ∈C^{−1}, r^{′}∈D^{−1}, and putting r = −r^{′}, gives ps −qr = 1. Thus

Selectable PDF notation • transcription unchecked
( )

p q div(p) ≥d, div(q) ≥e,

Selectable PDF notation • transcription unchecked
A = ,

r s div(r) ≥−e, div(s) ≥−d.

The inequality for zero entries has its usual infinite order interpretation.

For distinct coordinates i, j, take e = a_{i}+d−a_{j}. Insert A at these coordinates and the identity elsewhere. The first coordinate px_{i} + qx_{j} has divisor at least( ) a_{i} + d; the second rx_{i} + sx_{j} has divisor at least a_{j} −d. Applying the same inequalities to A^{−1}=^{ s}_{−r}^{−}_{p}^{q} proves the converse inclusion. This gives a determinant-one isomorphism

Selectable PDF notation • transcription unchecked
Λ(a) −→Λ(a + 1_{i}d −1_{j}d).

If two divisor families have equal sum, fix j_{0}. For each i ̸= j_{0}, make this move with d = b_{i} −a_{i}. It corrects coordinate i, preserves all other corrected coordinates, and transfers the difference to j_{0}. Equality of sums corrects the last coordinate. The composite determinant is one. In general first multiply coordinate j_{0} by β. Its determinant is β, and its new divisor sum is ∑_{j}b_{j}; apply the preceding redistribution. For empty J, the stipulated β = 1 permits the identity on the zero-dimensional space. This also explains the empty-block convention needed below.

Source: Development/EtaleCohomology/CyclotomicDiscriminants/SSteinTPf.lean:17--228; Development/EtaleCohomology/CyclotomicDiscriminants/FinSStein.lean:4--6

Lemma 8.49 (The polynomial and semilinear carrying maps). Let F/L be cyclic Galois with generator γ. Suppose F contains a primitive q = ℓ^{ν}-th root ζ, γζ = ζ^{κ}, κ = 1 + ℓ^{c}, and ord(γ) = ℓ^{ν−c}. Assume c ≥1, and c ≥2 if ℓ= 2. Let E ≥1, M = 2E + 1 < ℓ^{c}, and suppose

ord_{P} (α) = D_{γP} −κ^{E}D_{P} .

Put ∏

Selectable PDF notation • transcription unchecked
p(T) = (T −κ^{j}), Q(T) = p(T)(T −κ^{E}), β =∏(γ^{−i}α)^{[T i]p(T)}.
0≤_{j̸=}j<M_{E} i

For the function η of Lemma 8.46 and χ_{t}(j_{1}, j_{2}) = j_{1} +tj_{2}, there are L-linear automorphisms g_{t} of F^{(ℤ/q)2},

all carrying O_{F}^{(ℤ/q)2} onto Λ((η_{j}D)_{j}), and commuting with the cyclic diagonal action ζ^{aχt(j)}. They have a twisted block form in which all block determinants are one for t ̸= t_{0}, and for t = t_{0} only block 1 has determinant β.

Proof. The inverse powers of γ in β are essential. Equivariance of valuations gives

ord_{P} (γ^{−i}α) = D_{γ}i+1_{P} −κ^{E}D_{γ}i_{P} .

Multiply by [T^{i}]p and sum. The coefficient identity [T^{i}](p(T)(T −a)) = [T^{i−1}]p −a[T^{i}]p therefore gives

Selectable PDF notation • transcription unchecked
∑M

ord_{P} (β) = [T^{i}]Q(T)D_{γ}i_{P} . (217)

Selectable PDF notation • transcription unchecked
i=0

For t ̸= t_{0}, the pushforward of η by χ_{t} is zero. On each fibre χ_{t} = x, the target divisors η_{j}D sum to zero.

Lemma 8.48 supplies an F-linear block of determinant one from the zero-divisor lattice to these divisors. Their direct sum commutes with the scalar character on each block.

For the exceptional line we must change the fibres, rather than ignore the Galois twists. The elementary lifting-the-exponent identity

Selectable PDF notation • transcription unchecked
v_{ℓ}((1 + ℓ^{c})^{n}−1) = c + v_{ℓ}(n)

shows that κ modulo q has order ℓ^{ν−c}. One proves this identity first for ℓ ∤ n by the binomial expansion, and then iterates the same expansion for n →→ℓn; the specified 2-adic condition ensures the first term has strictly smaller valuation than the others. Thus κ^{i}= κ^{i′} implies γ^{i}= γ^{i′}. Define, on ℤ/q,

Selectable PDF notation • transcription unchecked
x T(x) B(x)
κ^{i} γ^{i} 1
x /∈⟨κ⟩ 1 x.

The first row is independent of the chosen exponent. In both rows

Selectable PDF notation • transcription unchecked
(T(x)^{−1}ζ)^{x}= ζ^{B(x)}.

Set a_{j,P} = η_{j}D_{T(χt0(j))P} . Apply the linear operation [x] →→ 1_{B(x)=y}D_{T(x)P} to the exceptional pushforward identity for η. Together with (217), it gives

Selectable PDF notation • transcription unchecked
∑ {

ord_{P} (β) y = 1,

Selectable PDF notation • transcription unchecked
_{B(χt0(j))=y}^{aj,P =} 0 y ̸= 1.

On every fibre of Bχ_{t0}, apply Lemma 8.48 with determinant β at y = 1 and one otherwise. The first fibre is nonempty since χ_{t0}(1, 0) = 1; empty other fibres have determinant one.

Write these F-linear blocks as g_{y}^{′}, and put s_{j} = T(χ_{t0}(j))^{−1}. Define the L-linear map g by the block equations ( )

Selectable PDF notation • transcription unchecked
s_{j}((gw)_{j}) = g_{B}^{′}_{(χt0(j))}( s_{i}(w_{i}) )_{B(χt0(i))=B(χt0(j))j}.

Applying s_{j} transports the target divisor η_{j}D to a_{j}, so these equations prove the lattice equivalence. The displayed root identity makes the diagonal action scalar on each transformed block. Hence g commutes with that action. This is the asserted twisted block form; the nonexceptional construction is the same formula with all s_{j} = 1 and B = χ_{t}.

Source: Development/AutomorphicArithmetic/AutomorphicVarieties/ModularityArgument/ClassCounting/LatticeBlockClasses/CBetaDiv.lean:6--62; Development/AutomorphicArithmetic/AutomorphicVarieties/ModularityArgument/ClassCounting/LatticeBlockClasses/CGtNe.lean:7--29; Development/AutomorphicArithmetic/AutomorphicVarieties/ModularityArgument/ClassCounting/LatticeBlockClasses/GtEqZeta.lean:11--106

Proposition 8.50 (Construction of the arithmetic cycle). Under the hypotheses of Lemma 8.49, there are an integer N, a cycle

Y_{α} ∈C_{2E+1}(GL_{N}(O_{L}), ℤ/q),

and 0 ̸= b_{α} ∈O_{F} , chosen before the residue prime, which satisfy (213) at every prime described there. Thus the cycle-production assumption in Proposition 8.33 is discharged.

Proof. Choose an integral primitive generator for F/L. After inverting one nonzero element b_{0} ∈O_{L}, its conjugates form the splitting of its monic minimal polynomial, with all pairwise differences invertible, and O_{F} [1/b_{0}] is generated by it. Indeed denominators of a finite integral basis and the nonzero discriminant give such b_{0}. Consequently, whenever r : O_{F} →k is a residue map with r(O_{L}) = k and r(b_{0}) ̸= 0, evaluation gives

Selectable PDF notation • transcription unchecked
O_{F} ⊗_{OL} k ≃ k^{Gal(F/L)}, a −−→(r(σa))_{σ}. (218)

The Vandermonde determinant proves this is an isomorphism, including the specified order of the conjugate coordinates.

Choose a nonzero multiple b ∈O_{L} of b_{0} whose prime divisors in F include the support of D. To do this multiply nonzero elements in the contractions to O_{L} of those finitely many primes. Let Γ be the L-linear automorphisms of F^{J}, J = (ℤ/q)^{2}, preserving O_{F} [1/b]^{J}; let H be its subgroup preserving O_{F}^{J}. All g_{t} of the preceding lemma belong to Γ: the source and target lattices become O_{F} [1/b]^{J} after this localization. Let

Selectable PDF notation • transcription unchecked
ρ : C_{q}^{2}−→Γ, (ρ(a_{1}, a_{2})w)_{j} = ζ^{a1j1+a2j2}w_{j}.

It preserves both the zero lattice and every coordinate-divisor lattice. Thus ρ(C_{q}^{2}) ⊂H, g_{t}^{−}_{0}^{1}ρ(C_{q}^{2})g_{t0}⊂H, and g_{t}^{−1}g_{t0} ∈ H: the last assertion follows because both maps carry the zero lattice onto the same target lattice. Each g_{t} commutes with ρd_{t}, by the preceding lemma.

Choose n_{t}, t_{0} from Lemma 8.44. Lemma 8.45 supplies an H-cycle y whose image in Γ, modulo a boundary, is ∑

Selectable PDF notation • transcription unchecked
n_{t}P_{gt}(ρd_{t})_{#}z.
t

The O_{L}-module M_{0} = O_{F}^{J} is finite and torsion free, hence projective over the Dedekind domain O_{L}. Here the projectivity can be seen by the same rank-one ideal splitting used in Lemma 8.48: a nonzero coordinate projection has a fractional ideal as image, which is invertible and projective, so it splits; induction on rank decomposes the module into such ideals. A finite projective module is a summand of a finite free module: choose generators, split the resulting surjection O_{L}^{N} →M_{0}, and take its kernel M_{1}. Extend the action of H trivially on M_{1}. This gives λ : H → GL_{N}(O_{L}). Set

Y_{α} = λ_{#}(n^{−}_{t0}^{1}y), b_{α} = b viewed in O_{F} .

It is a cycle, and these choices have made no use of r, k, or v_{P} .

We calculate its reduction. The isomorphism (218) gives a representation µ : Γ → GL(k^{G×J}), where G = Gal(F/L). It is obtained by extending to O_{F} [1/b], reducing, and taking the conjugate coordinates in that isomorphism. On H, reduction of λ is isomorphic to µ|_{H} plus a trivial summand. This follows by tensoring M_{0} ⊕M_{1} ≃O_{L}^{N} with k. The functional of Lemma 8.47 is unchanged by this extra summand on a positive-degree cycle.

Write σζ = ζ^{κσ}. Consider any carrying map with twists s_{j}, labels B(j), blocks g_{x}^{′}, and determinants δ_{x} as in the preceding lemma. The following coordinate change retains all the twists:

Selectable PDF notation • transcription unchecked
(Φw)_{σ,x,j} = w_{σsj,j}, B(j) = x.

Its inverse reads coordinate (τ, j) from (τs^{−}_{j}^{1}, B(j), j). The defining block equation shows that µ(g) on

block (σ, x) is the reduction of σg_{x}^{′} ; its determinant is r(σδ_{x}). The compatibility (s_{j}ζ)^{χt(j)} = ζ^{B(j)} shows that µ(ρd_{t}) is scalar ζ^{κσx} on that block. These identities can also be checked on integral vectors and extended to all residue vectors using the surjectivity in (218). There is no determinant sign from a reordered basis: the same isomorphism Φ conjugates both operators. The finite-field prism formula therefore gives

Selectable PDF notation • transcription unchecked
θ(P_{µ(g)}(µρd_{t})_{#}z)=∑(κ_{σ}x)^{E}v_{P} (r(σδ_{x})).
σ,x

Every term is zero for∑ t ̸= t_{0}, since then δ_{x} = 1. For t = t_{0}, only x = 1 survives, and the answer is

Selectable PDF notation • transcription unchecked
_{σ}κ^{E}_{σ}v_{P}(r(σβ)).

All elements in this expression are units outside b. In fact the divisor relation for α has support above b, stable under G, and the same is true of its conjugates and their products. Thus f(u) = v_{P} (r(u)) is additive on the G-stable group of these units. Put

Selectable PDF notation • transcription unchecked
S =∑κ^{E}_{σ}f(σα).
σ

As κ_{στ} = κ_{σ}κ_{τ} modulo q, reindexing σ = τγ^{i} shows

Selectable PDF notation • transcription unchecked
∑κ^{E}_{σ}f(σγ^{−i}α) = (κ^{E})^{i}S.
σ

Expanding the definition of β gives ∑

Selectable PDF notation • transcription unchecked
κ^{E}_{σ}f(σβ) = p(κ^{E})S.
σ

Since γ generates G, the weighted norm 𝒩_{a}(α) =∏_{j<|G|}(γ^{j}α)^{aj}, a = κ^{E}, satisfies f(𝒩_{a}(α)) = S. Finally p(κ^{E}) = p_{E}. The boundary in the formula for y is killed by θ, and its single surviving prism has coefficient n_{t0}. Multiplication by n^{−}_{t0}^{1} in the definition of Y_{α} cancels that coefficient. This is exactly (213).

Source: Development/AutomorphicArithmetic/AutomorphicVarieties/ModularityArgument/ClassCounting/LatticeBlockClasses/CJoin.lean:5--187; Development/AutomorphicArithmetic/AutomorphicVarieties/ModularityArgument/ClassCounting/LatticeBlockClasses/CSplit.lean; Development/AutomorphicArithmetic/AutomorphicVarieties/ModularityArgument/ClassCounting/LatticeBlockClasses/CLatLoc.lean; Development/AutomorphicArithmetic/AutomorphicVarieties/ModularityArgument/ClassCounting/LatticeBlockClasses/BlkEquiv.lean:9--75; Development/AutomorphicArithmetic/AutomorphicVarieties/ModularityArgument/ClassCounting/LatticeBlockClasses/CBlockVal.lean:12--107; Development/AutomorphicArithmetic/AutomorphicVarieties/ModularityArgument/ClassCounting/LatticeBlockClasses/CBetaVal.lean:5--122

Lemma 8.51 (Restriction to the cyclotomic kernel, with local conditions). Let L be a number field containing µ_{ℓ} if ℓ is odd, or µ_{4} if ℓ= 2. Fix an algebraic closure and write

G = G_{L}, χ = χ_{ℓ}, H = ker χ, A = ℚ_{ℓ}/ℤ_{ℓ}(−E), E ≥1.

Let 𝒟 be the continuous additive characters φ : H →ℚ_{ℓ}/ℤ_{ℓ} that vanish on H ∩I_{w} at every finite place and on H ∩D_{w} at every place above ℓ. Then restriction is an isomorphism

Selectable PDF notation • transcription unchecked
𝒮_{ℓ}(L, E) ≃ {φ ∈𝒟 : φ(g^{−1}hg) = χ(g)^{E}φ(h) (g ∈G, h ∈H)}. (219)

All maps are ℤ_{ℓ}-linear.

Proof. We supply the extension and local coboundary arguments, rather than discard the local conditions in a restriction sequence. The image of χ is an infinite closed subgroup of the principal units 1 + ℓℤ_{ℓ}, or 1 + 4ℤ_{2}. It is infinite because

Selectable PDF notation • transcription unchecked
_{φ(ℓ}r_{)}
[L(µ_{ℓ}r) : L] ≥_{[L : ℚ]} −→∞.

The degree inequality follows by adjoining the primitive root over ℚ and applying the tower formula. Every nontrivial closed subgroup of these principal units is topologically cyclic. Indeed choose an element with minimal positive valuation a of u −1. Lifting the exponent shows that its powers surject onto each finite

quotient of 1 + ℓ^{a}ℤ_{ℓ}; their closure is that group. The minimality of a gives the reverse containment for the subgroup. Choose γ ∈G whose image u = χ(γ) is such a topological generator. It has infinite order. Put C = ⟨γ⟩. Compactness shows χ(C) = χ(G), hence G = CH.

Suppose φ is on the right of (219). First φ(C ∩H) = 0. To prove this, choose an open normal subgroup N of G such that φ(H ∩N) = 0. Such N exists by continuity at one and the profinite neighbourhood basis. Let the order of γ in G/N be d = ℓ^{a}d^{′}, with ℓ∤d^{′}. Choose j so large that

u^{m}≡1 (mod ℓ^{j}) =⇒ ℓ^{a}| m.

Lifting the exponent gives this implication also for negative integers m. For x ∈C ∩H, approximate x by a power γ^{m} both modulo N and in the indicated cyclotomic congruence. Then d | md^{′}, so (γ^{m})^{d′}∈N. The group C is abelian, being the closure of a cyclic subgroup, and consequently x^{d′}∈H ∩N. Thus d^{′}φ(x) = 0. As φ(x) has ℓ-power order and ℓ∤d^{′}, this implies φ(x) = 0.

Write ρ(g) = χ(g)^{−E} for the action on A. Define

Selectable PDF notation • transcription unchecked
z(ch) = ρ(c)φ(h) (c ∈C, h ∈H).

If ch = c^{′}h^{′}, then c^{′−1}c = h^{′}h^{−1}∈C ∩H. Its φ-value is zero, and H acts trivially on A; these two facts prove independence of the decomposition. The product identity

Selectable PDF notation • transcription unchecked
(ch)(dv) = cd(d^{−1}hd)v

and the stipulated equivariance of φ give z(gg^{′}) = z(g) + ρ(g)z(g^{′}). The map C × H →G, (c, h) →→ch, is a continuous surjection from a compact space to a Hausdorff space, hence a quotient map. The defining formula on C × H is continuous, so z is a continuous cocycle. Its restriction to H is φ.

Conversely restricting a cocycle to H gives an additive character, since H acts trivially on A. The cocycle identity gives its displayed equivariance; the terms involving z(g) cancel in conjugation. Coboundaries restrict to zero. If a cocycle has zero restriction, choose m ∈A with

Selectable PDF notation • transcription unchecked
z(γ) = (u^{−E}−1)m.

This is possible because u^{−E}−1 ̸= 0 and multiplication by any nonzero element of ℤ_{ℓ} on ℚ_{ℓ}/ℤ_{ℓ} is onto: divide a representative in ℚ_{ℓ}. Subtract its coboundary. The resulting cocycle is zero on H and on all integer powers of γ, then on C by continuity, and then on G = CH. This proves injectivity on cohomology.

It remains to prove that the extended class has exactly the required local behaviour. For w ∤ℓ, inertia fixes the ℓ-power roots of unity: their integral reductions have distinct roots, because the residue characteristic does not divide their order. Thus I_{w} ⊂H, and z(I_{w}) = 0 by the assumption on φ. The cocycle therefore factors through the Galois group unramified outside ℓ; it vanishes on the closed normal subgroup generated by these inertias, using its cocycle identity and their trivial scalar action.

For w | ℓ, the image χ(D_{w}) is again an infinite closed subgroup of the same principal units. Its infinitude follows from

Selectable PDF notation • transcription unchecked
_{φ(ℓ}r_{)}
[L_{w}(µ_{ℓ}r) : L_{w}] ≥_{[Lw : ℚℓ]} −→∞.

Here the ℓ-power cyclotomic polynomial over ℚ_{ℓ} has the indicated degree: after replacing X by X + 1 it is Eisenstein at ℓ. The same topological-generator argument applies. The restriction of z to D_{w} vanishes on D_{w} ∩H. Kill its value on a lift of a generator by the divisible scalar argument just used. The residual cocycle is zero on the kernel and on a dense cyclic image, hence on D_{w}. Thus z|_{Dw} is a coboundary. This proves membership in 𝒮_{ℓ}(L, E). In the other direction a locally zero class restricts to zero on these intersections, since coboundaries vanish on H. Both local implications and linearity are now proved.

Source: Development/AutomorphicArithmetic/AutomorphicVarieties/GaloisLibrary/KummerTheory/VanishCyc.lean:5--163; Development/AutomorphicArithmetic/AutomorphicVarieties/GaloisLibrary/KummerTheory/ExtCocycle.lean:10--155; Development/AutomorphicArithmetic/AutomorphicVarieties/GaloisLibrary/KummerTheory/ResInfInj.lean:58--171; Development/AutomorphicArithmetic/AutomorphicVarieties/GaloisLibrary/KummerTheory/ResInfOnto.lean:15--166

Proposition 8.52 (One common operator for every cyclotomic layer). Retain the preceding hypotheses and put χ(G_{L}) = 1 + ℓ^{c0}ℤ_{ℓ}, with c_{0} ≥ 2 when ℓ = 2. Let L_{m} = L(µ_{ℓ}c_{0+}m). There are a single compact ℤ_{ℓ}-module X and, for every E ≥1, a continuous automorphism T_{E}, independent of m, such that

s_{ℓ}(L_{m}, E) = length_{ℤℓ} coker(T_{E}^{ℓm}−1). (220)

Lengths may initially be infinite; the equality includes finiteness in either direction. The generator γ, X, and the untwisted action are fixed before E and m vary.

Proof. All L_{m} have the same infinite cyclotomic extension L_{∞}, and the same cyclotomic kernel H = G_{L∞}. Their group inside G_{L} is

Selectable PDF notation • transcription unchecked
G_{m} = χ^{−1}(1 + ℓ^{c0+m}ℤ_{ℓ}).

The lifting-the-exponent calculation shows [G_{L} : G_{m}] = ℓ^{m} and that γ^{ℓm} generates its cyclotomic image. It also shows uniqueness: a finite cyclotomic field of degree ℓ^{m} is this field, since its fixing group is the corresponding principal congruence preimage. In particular this identification is valid for any abstractly given layer after embedding it into the fixed closure.

The module 𝒟 in Lemma 8.51 is common to all these fields. Indeed the inertia group in G_{m} is I_{w} ∩G_{m}, and its intersection with H is I_{w} ∩H; the same statement holds for decomposition groups above ℓ. Every place of each layer is obtained this way. Thus no local vanishing condition is weakened. On 𝒟 define

Selectable PDF notation • transcription unchecked
(τ_{E}φ)(h) = χ(γ)^{−E}φ(γ^{−1}hγ).

Conjugation preserves the collection of local subgroups, so this is an automorphism of 𝒟. Its invariants under the group G_{m} are precisely ker(τ_{E}^{ℓm}−1). To verify this last assertion, the twisted action of H is trivial because additive characters kill inner conjugation. Invariance under a generator then implies invariance under its powers and their closure by continuity. Applying (219) over L_{m} gives

𝒮_{ℓ}(L_{m}, E) ≃ ker(τ_{E}^{ℓm}−1).

Give 𝒟 its discrete topology and set X = Hom(𝒟, ℚ_{ℓ}/ℤ_{ℓ}) with the compact topology of pointwise convergence. Every element of 𝒟 has finite ℓ-power order, since its continuous image from compact H to a discrete group is finite. Consequently X is the inverse limit of the duals of the finite subgroups of 𝒟. Set T_{E} to be the dual of τ_{E}^{−1}. Duality gives

Selectable PDF notation • transcription unchecked
(

coker(T_{E}^{ℓm}−1))^{∨}≃ ker(τ^{ −}_{E}^{ℓm}−1) = ker(τ_{E}^{ℓm}−1).

For completeness, exactness here does not assert that a nonclosed image can be quotiented without care: the image of a continuous endomorphism of compact X is compact and closed. A continuous finite-order character of a closed subgroup extends to X. Choose an open subgroup whose intersection lies in its kernel, extend the resulting character in the finite quotient using the divisibility of ℚ_{ℓ}/ℤ_{ℓ}, and pull back. Conversely a character vanishing on an image factors through the compact quotient. These observations prove the ex-actness asserted. Finite cyclic ℓ-groups and their duals have the same order; a composition series therefore has the same length on both sides. If either side has infinite length, arbitrarily long finite subquotients give the same conclusion. This proves (220).

Source: Development/AutomorphicArithmetic/AutomorphicVarieties/NumberFieldTowers/Imports.lean:10--147; Development/AutomorphicArithmetic/AutomorphicVarieties/NumberFieldTowers/LayerGen.lean:10--39; Develop-ment/AutomorphicArithmetic/AutomorphicVarieties/NumberFieldTowers/LayerUniq.lean:11--43; Development/AutomorphicArithmetic/AutomorphicVarieties/NumberFieldTowers/ExistsDescends.lean:26--57

Lemma 8.53 (A common finite level for finitely many characters). Let L have largest ℓ-power root level c, with the root hypotheses of Lemma 8.51. Suppose every cyclic cyclotomic layer

Selectable PDF notation • transcription unchecked
F = L(µ_{ℓ}ν), ν ≥c, γ(ζ) = ζ^{1+ℓc},

has at most B characters ϑ : Cl(F) →ℤ/ℓ^{ν} satisfying

Selectable PDF notation • transcription unchecked
ϑ(γ^{−1}a) = (1 + ℓ^{c})^{E}ϑ(a).

Then 𝒮_{ℓ}(L, E) is finite and

Selectable PDF notation • transcription unchecked
ℓ^{sℓ(L,E)}= |𝒮_{ℓ}(L, E)| ≤B.

Proof. Take a finite subset of 𝒮_{ℓ}(L, E). Represent each of its restrictions to H by the continuous cocycle z : G_{L} →A constructed in Lemma 8.51. Each has finite image, since G_{L} is compact and A is discrete. At each of the finitely many places w | ℓ choose m_{w} with z(g) = (χ(g)^{−E}−1)m_{w} on D_{w}. Choose a common K annihilating the images of all these cocycles and all their m_{w}’s by ℓ^{K}. The same K handles all conjugate decomposition groups. In fact the cocycle formula shows that a conjugate local coboundary has parameter ρ(g)m_{w} −z(g); scalar units preserve the annihilator, and both terms are killed by ℓ^{K}.

Choose ν ≥ c, K and put U_{ν} = χ^{−1}(1 + ℓ^{ν}ℤ_{ℓ}). Every ρ(u) −1, u ∈ U_{ν}, is divisible by ℓ^{ν}. Thus z|_{Uν} is an additive character: its cocycle multiplication formula has trivial action on every value of z. It is unramified at every finite place. Away from ℓ the original cocycle vanishes on inertia; above ℓ its local coboundary formula is zero on D_{w} ∩U_{ν}. It also remains equivariant for the whole G_{L}, because

Selectable PDF notation • transcription unchecked
z(gug^{−1}) = ρ(g)z(u) + (1 −ρ(u))z(g) = ρ(g)z(u).

All its values lie in A[ℓ^{ν}].

We use the unramified reciprocity theorem in precisely its Hilbert-class-field form: for a totally complex number field F, Artin induces an isomorphism from Cl(F) to the Galois group of its maximal abelian ex-tension unramified at all finite places.^{15} Here F is totally complex because it contains the stipulated nonreal root of unity. The character z|_{Uν} therefore factors uniquely through Cl(F). No splitting condition at ℓ is needed for this factorization; its stronger vanishing has already been checked. Identify A[ℓ^{ν}] with ℤ/ℓ^{ν} by a →→a/ℓ^{ν} mod ℤ_{ℓ}. Conjugation naturality of Artin gives, for its class character ϑ,

Selectable PDF notation • transcription unchecked
ϑ(γ^{−1}a) = (1 + ℓ^{c})^{E}ϑ(a).

To check the normalization, conjugation by a lift of γ carries a prime Frobenius to that of γP, and the preceding equivariance uses the scalar χ(γ)^{−E}; inverting γ gives the stated scalar. Such a lift exists with cyclotomic value exactly 1 + ℓ^{c}, since the full image is the principal group at level c.

Different original classes give different finite class characters: equality at this common finite level im-plies equality of their restrictions to H, and restriction is injective by (219). Thus the arbitrary finite subset has at most B members. An infinite set would have a subset of size B + 1, so the entire module is finite. A finite ℤ_{ℓ}-module is a direct sum of cyclic ℓ-groups; successively dividing by a subgroup of order ℓ shows its order is ℓ to its module length. Applying the same finite-subset bound to the full module proves the assertion.

^{15}J. S. Milne, Class Field Theory, v4.03 (2020), Introduction, Theorem 0.4 and the following paragraph, printed p. 4; Chapter V, Theorems 3.5–3.6 and Example 3.9, printed pp. 158–159: https://www.jmilne.org/math/CourseNotes/CFT.pdf.

Source: Development/AutomorphicArithmetic/AutomorphicVarieties/GaloisLibrary/KummerTheory/LevCompact.lean:6--84; Development/AutomorphicArithmetic/AutomorphicVarieties/GaloisLibrary/KummerTheory/CycZConjGal.lean:20--177; Development/AutomorphicArithmetic/AutomorphicVarieties/GaloisLibrary/KummerTheory/PullEq.lean:7--32; Development/AutomorphicArithmetic/AutomorphicVarieties/ModularityArgument/ClassCounting/RingEquivalences/OrderOfOneAddPrimePowZmodPrimePow.lean:348--490

Lemma 8.54 (The controlled auxiliary layer). For L, ℓ, E as above, put n = [L : ℚ], D = |∆_{L}|. There is a totally complex cyclotomic extension L^{′}/L, of degree d, with n^{′}= dn, D^{′}= |∆_{L}′|, and largest root level c^{′}, such that

Selectable PDF notation • transcription unchecked
2E + 1 < ℓ^{c′}, 2E + 1 < r_{2}(L^{′}),
d ≤(n + 1)(4E + 4), ℓ^{c′}≤2n^{′},
D^{′}≤D^{d}(ℓ^{c′})^{n′}, s_{ℓ}(L, E) ≤s_{ℓ}(L^{′}, E).

Proof. Let m be the least integer with ℓ^{m} ≥4E + 4, and take L^{′} = L(µ_{ℓ}c+m), d = ℓ^{m}. Then 4E + 4 ≤ d < ℓ(4E + 4). For odd ℓ, the primitive ℓ-th root in L has rational degree ℓ−1, hence ℓ≤n + 1; for ℓ= 2 this follows from n ≥ 1. Thus the required bound for d holds. The maximal root level c^{′} of L^{′} satisfies c^{′}≥c + m, so ℓ^{c′}≥d > 2E + 1. The field L^{′} is totally complex, and

Selectable PDF notation • transcription unchecked
2r_{2}(L^{′}) = n^{′}= dn ≥4E + 4 > 2(2E + 1).

Its primitive root has rational degree φ(ℓ^{c′}) ≤n^{′}. Since ℓ^{c′}/φ(ℓ^{c′}) = ℓ/(ℓ−1) ≤2, this gives ℓ^{c′}≤2n^{′}.

For clarity the discriminant estimate uses the different, not an asserted discriminant identity for a poten-tially nonmaximal monogenic order. If z is the generating root of order N = ℓ^{c+m}, its integral minimal polynomial f divides X^{N}−1. Write X^{N}−1 = fh. The derivative f^{′}(z) belongs to the relative different. One sees this from the trace dual of the power basis: for a monic separable polynomial of degree d, the coefficient-of-X^{d−1} identity ( )

Tr g(z)/f^{′}(z) = [X^{d−1}](g mod f)

identifies the trace dual of the monogenic order; the maximal-order complementary module is contained in this dual, giving the stated inclusion. Differentiating X^{N}−1 = fh at z gives

Selectable PDF notation • transcription unchecked
N = f^{′}(z)h(z)z ∈ 𝔇_{L}′_{/L}.

Thus the absolute norm of the relative different is at most N^{n′}. The discriminant tower formula, obtained by the determinant of the trace pairing in composed bases, now gives

Selectable PDF notation • transcription unchecked
D^{′}= D^{d}N_{L}′_{/ℚ}(𝔇_{L}′_{/L}) ≤D^{d}N^{n′}≤D^{d}(ℓ^{c′})^{n′}.

Finally use the common operator of Proposition 8.52. The identity

Selectable PDF notation • transcription unchecked
_{m} ℓ∑^{m}−1
T_{E}^{ℓ} −1 = (T_{E} −1) T_{E}^{i}
i=0

puts its image inside that of T_{E} −1. The corresponding cokernel therefore surjects onto the base cokernel; length is monotone under surjections. This proves the last inequality, including the possibility of infinite lengths.

Source: Development/EtaleCohomology/CyclotomicDiscriminants/SLayerCut.lean:11--106

Lemma 8.55 (The class-number estimate used in the exponent). For a number field of degree n and absolute discriminant D ≥1,

h_{L} ≤(2D)^{2n}, ⌊log_{2} h_{L}⌋≤4n log(2D).

Proof. In its real and complex embedding space, an integral ideal I has lattice covolume 2^{−r2√}D N(I). To check the factor, the determinant of the matrix of all embeddings of an integral basis has squared absolute value D. Replacing each conjugate complex pair by its real and imaginary coordinates divides the absolute determinant by two. An ideal of index N(I) multiplies this covolume by that index.

Apply Lemma_{√} 8.22 to the product of real intervals of radius R and complex disks of radius R, with R^{n}= 2 D N(I). Its volume 2^{r1}π_{√}^{r2}R^{n} is strictly greater than 2^{n} times the ideal covolume. Thus it contains_{√} 0 ̸= α ∈ I with |N_{L/ℚ}α| ≤ 2 D N(I). The integral ideal (α)I^{−1} then has norm at most 2 D and

Selectable PDF notation • transcription unchecked
√

represents the inverse ideal class. Hence every ideal class has an integral ideal of norm at most B = 2 D.

For a rational prime p, there are at most n prime ideals above p, since their ramification indices times residue degrees sum to n. An ideal of norm p^{k} is determined by at most n exponents, each between zero and k. For k ≥1, their number is at most (k + 1)^{n}≤p^{kn}; for k = 0 it is one. Unique factorization of nonzero ideals therefore bounds the number of ideals of norm m by m^{n}. The number of ideals of norm at most B is at most B^{n+1}. Consequently _{√}

Selectable PDF notation • transcription unchecked
h_{L} ≤(2 D)^{n+1}≤(2D)^{2n}.

The last inequality follows from n ≥1, D ≥1. Taking logarithms and using log 2 ≥1/2 gives the second assertion.

Theorem 8.56 (The integral cohomology bound). The statement (210) holds with an absolute exponent C_{K}; the estimates just proved permit C_{K} = 150. This assertion uses the standard homological-stability, stable-cohomology, class-field and arithmetic-duality theorems in the precise forms and ranges checked above.

Proof. First work over a totally complex base L^{′} with largest root level c and 2E + 1 < ℓ^{c}, r_{2}(L^{′}). For each ν ≥c, the corresponding cyclic extension F = L^{′}(µ_{ℓ}ν) has a generator with action 1 + ℓ^{c} on the roots. The arithmetic cycles of Proposition 8.50, followed by Proposition 8.33, give the finite-value bound

Selectable PDF notation • transcription unchecked
B_{ν} = (ℓ^{ν})^{r2(L′)}T_{2E+1}(L^{′}, 4E + 6) T_{2E}(L^{′}, 4E + 6)

for weighted norms separated modulo ℓ^{ν}-th powers. Its rank input is Proposition 8.35, with stability as checked in Lemma 8.34. Proposition 8.31 then gives

Selectable PDF notation • transcription unchecked
|ℰ_{ν}| (ℓ^{ν})^{r2(L′)}≤B_{ν} ℓ^{X}.

The positive common power cancels. The resulting bound

Selectable PDF notation • transcription unchecked
|ℰ_{ν}| ≤T_{2E+1}(L^{′}, 4E + 6) T_{2E}(L^{′}, 4E + 6) ℓ^{X}

is independent of ν: here

Selectable PDF notation • transcription unchecked
X = (b + f)g + c + f, b = 2E + (2E + 1)(c + 2E + 1), f = v_{ℓ}(E),

and g = 2[L^{′}: ℚ] + ⌊log_{2} h_{L}′⌋+ 2. Lemma 8.53 now gives finiteness of 𝒮_{ℓ}(L^{′}, E) and exactly the class-counting inequality in Proposition 8.21.

For an arbitrary base L under (210), choose L^{′} using Lemma 8.54. All its degree, discriminant, root and Sha monotonicity conditions have been proved there. The class-number bound just proved gives the required bound for g. Proposition 8.28 gives C_{0} = 24 for both integral torsion factors. Thus Proposition 8.21 gives C = 5(24 + 2) + 11 = 141 for the Sha term. Finally Proposition 8.29 supplies the exact global/local length formula, with its explicitly bounded local terms. The final calculation in Proposition 8.21 increases the exponent by nine, giving C_{K} = 150.

Lemma 8.57 (The determinant centres and growth estimate). Let X be a compact ℤ_{ℓ}-module generated by at most R ≥1 elements, and let g be a continuous automorphism, unipotent on X/ℓX. Let u ∈ℤ^{×}_{ℓ} satisfy v_{ℓ}(u −1) = c ≥1, with c ≥2 at ℓ= 2. Set j_{0} = ⌊log_{ℓ}R⌋+ 3 and

a(e, m) = length coker((u^{e}g)^{ℓm}−1).

There are s_{1}, . . . , s_{R} ∈ℤ_{ℓ}, chosen independently of e, m, such that

Selectable PDF notation • transcription unchecked
∑R

a(e, m) ≤ {max(m, j_{0}) + c + v_{ℓ}(s_{i} −e)},

Selectable PDF notation • transcription unchecked
i=1
a(e, m) ≤a(e, j_{0}) + R(m −j_{0})_{+}.

The inequalities use extended lengths and v_{ℓ}(0) = +∞.

Proof. We first prove growth, allowing torsion in X. For an endomorphism w, write w ≡1 (mod ℓ^{a}) to mean (w −1)X ⊆ ℓ^{a}X; this notation does not require dividing an endomorphism on a torsion module. Nilpotence on the vector space X/ℓX has index at most its dimension, hence at most R. Thus (g −1)^{R}X ⊆ ℓX. The same holds for u^{e}g, since its reduction is that of g. Put k = ⌊log_{ℓ}R⌋+ 1. Then ℓ^{k} ≥R, and the binomial coefficients in the Frobenius identity modulo ℓ show

Selectable PDF notation • transcription unchecked
k

(u^{e}g)^{ℓ} ≡1 (mod ℓ).

If w ≡1 (mod ℓ), then w^{ℓ} ≡1 (mod ℓ^{2}). Indeed for x, write (w −1)x = ℓy and (w −1)y = ℓy^{′}. The geometric-sum identity gives

Selectable PDF notation • transcription unchecked
(w^{ℓ}−1)x = ℓ∑w^{i}y = ℓ^{2}((y +∑∑w^{j}y^{′})).
i<ℓ i<ℓ j<i

Further positive powers preserve this divisibility. Consequently w = (u^{e}g)^{ℓn} satisfies w ≡1 (mod ℓ^{2}) for

Selectable PDF notation • transcription unchecked
n ≥k + 1.

For such w, let S = 1 + w + · · · + w^{ℓ−1}. Every Sx is ℓx modulo ℓ^{2}X; therefore

Selectable PDF notation • transcription unchecked
SX ⊆ℓX, ℓX ⊆SX + ℓ(ℓX).

The finite ℤ_{ℓ}-module ℓX/SX equals its multiple by ℓ, so Nakayama’s lemma makes it zero. A direct proof of this instance is the determinant trick: choose generators q_{i}, express q_{i} = ℓ∑a_{ij}q_{j}, and multiply by the adjugate of I −ℓA; its determinant is a unit, so every q_{i} = 0. Thus SX = ℓX, and length(X/SX) ≤R. For two endomorphisms f, h, the sequence

Selectable PDF notation • transcription unchecked
X/hX−→^{f} X/fhX −→X/fX −→0

is right exact. It bounds the middle length by the sum of the other two. Apply it to w^{ℓ}−1 = S(w −1), giving

length coker(w^{ℓ}−1) ≤ length coker(w −1) + R.

Iteration from j_{0} = k + 2 proves growth for m ≥j_{0}. For m < j_{0}, the polynomial factorization of w^{ℓj0}−1 by w^{ℓm}−1 gives a surjection of cokernels in the required direction. This proves the second bound.

Choose a surjection π : ℤ^{R}_{ℓ} →X. Lift each image g^{ℓj0}π(e_{i}) to get a matrix A with πA = g^{ℓj0}π. For every scalar b, π induces a surjection

coker(bA −I) −→ coker(bg^{ℓj0}−1).

The length of the free-module cokernel is v_{ℓ}(det(bA −I)). One verifies this by choosing an entry with minimal valuation, using invertible row and column operations to clear its row and column, and inducting. This produces the diagonal form over the discrete valuation ring. If the determinant is zero there is a nonzero free quotient and both the determinant valuation and the relevant upper bound are infinite.

In a normed algebraic closure of ℚ_{ℓ}, let z_{1}, . . . , z_{R} be the roots of the characteristic polynomial of A, with multiplicity. Factoring that polynomial gives

Selectable PDF notation • transcription unchecked
∏

det(bA −I) = (bz_{i} −1).

Selectable PDF notation • transcription unchecked
i

This includes zero roots and does not assume that A is invertible. We now prove the required estimate for one factor, with j = j_{0} and arbitrary z. Write b_{e} = (u^{ℓj})^{e}. Lifting the exponent gives

Selectable PDF notation • transcription unchecked
|b_{a} −b_{b}|_{ℓ}= ℓ^{−(j+c)}|a −b|_{ℓ}.

Since |b_{e}|_{ℓ}= 1, the ultrametric inequality gives

|b_{a} −b_{b}|_{ℓ}≤ max{|b_{a}z −1|_{ℓ}, |b_{b}z −1|_{ℓ}}.

For |z| ≥1, subtract the two factors and divide by |z|; for |z| < 1, each factor has norm one. These cover all z. Thus the closed balls in ℤ_{ℓ} with centres e ∈ ℕ and radii ℓ^{j+c}|b_{e}z −1|_{ℓ} have the finite intersection property. For a finite set of them choose the centre having least radius; the displayed pairwise inequality puts that centre in every ball. Compactness of ℤ_{ℓ} gives a common point s, satisfying

Selectable PDF notation • transcription unchecked
ℓ^{−(j+c)}|s −e|_{ℓ}≤|b_{e}z −1|_{ℓ} (e ∈ℕ).

Choose s_{i} in this way for each z_{i}. Multiplication of the inequalities and the determinant factorization give

Selectable PDF notation • transcription unchecked
a(e, j_{0}) ≤∑{j_{0} + c + v_{ℓ}(s_{i} −e)}.
i

Combine this with the growth bound; add m −j_{0} once for each of the R terms. This proves the first inequality.

Source: Development/AutomorphicArithmetic/AutomorphicVarieties/AdditiveValuationNorms/UnipPow.lean:8--46; Development/AutomorphicArithmetic/AutomorphicVarieties/AdditiveValuationNorms/PowIn.lean:6--29; Development/AutomorphicArithmetic/AutomorphicVarieties/AdditiveValuationNorms/ImportsTer.lean:7--46; Development/AutomorphicArithmetic/AutomorphicVarieties/AdditiveValuationNorms/ImportsMore.lean:9--22; Development/AutomorphicArithmetic/AutomorphicVarieties/AdditiveValuationNorms/Growth.lean:11--56; Devel-opment/AutomorphicArithmetic/AutomorphicVarieties/AdditiveValuationNorms/ImportsAlt.lean:7--44; Devel-opment/AutomorphicArithmetic/AutomorphicVarieties/AdditiveValuationNorms/Helly.lean:6--25; Development/AutomorphicArithmetic/AutomorphicVarieties/AdditiveValuationNorms/DetFactor.lean:6--49; Development/AutomorphicArithmetic/AutomorphicVarieties/AdditiveValuationNorms/PadicUnitsDistanceLeMax.lean:8--63; Development/AutomorphicArithmetic/AutomorphicVarieties/AdditiveValuationNorms/ExistsCentresLevel.lean:17--69

Proposition 8.58 (The quantitative tower bounds from finite generation). If the common cyclotomic module X over L_{0} in Proposition 8.52 is generated by at most R ≥1 elements, then both (215) and (216) hold, with at most R centres in ℚ_{ℓ} and c the largest root level of L_{0}.

Proof. The untwisted conjugation action on X factors through the principal cyclotomic group, which is pro- ℓ. Its action on the finite vector space X/ℓX therefore has ℓ-power order. In characteristic ℓ, g^{ℓa}= 1 implies (g−1)^{ℓa}= 0, proving the unipotence needed in the preceding lemma. The dual twisted operator is T_{e} = u^{e}g, where u = χ(γ) has level c. The common-operator formula and the first inequality of that lemma give (215) as soon as |s_{i} −e|_{ℓ}≥ℓ^{−B}. Its second inequality gives

Selectable PDF notation • transcription unchecked
s_{ℓ}(L_{m}, e) ≤s_{ℓ}(L_{j0}, e) + R(m −j_{0})_{+}.

By Proposition 8.29, s_{ℓ}(L_{j0}, e) is at most the integral degree-two length there: the global invariant group injects into the product of the local invariant groups, so the extra local length minus global length is nonneg-ative. These groups are finite by Theorem 8.56, which applies to every layer of L_{0}. This proves (216).

Source: Development/AutomorphicArithmetic/AutomorphicVarieties/AdditiveValuationNorms/ABoundOfDescends.lean:36--55; Development/AutomorphicArithmetic/AutomorphicVarieties/AdditiveValuationNorms/ABoundOfIsTower.lean:7--11

Lemma 8.59 (Characters across an index-two subgroup). Let U be an open subgroup of index two in a profinite group V , and let S ⊂ U be stable under conjugation by V . Write 𝒞(V, S) for the continuous homomorphisms V → F_{2} vanishing on S. If |𝒞(V, S)| ≤n, then

Selectable PDF notation • transcription unchecked
|𝒞(U, S)| ≤n^{2}.

The same upper bound holds when the index divides two. Finiteness of the left side is part of the conclusion.

Proof. Choose σ ∈ V \ U. Conjugation by σ induces an involution c on 𝒞(U, S), since σ^{2} ∈ U and characters kill inner conjugation. For P = ker(1 + c), define t(φ) = φ(σ^{2}). If t is nonzero choose φ_{0} ∈P with t(φ_{0}) = 1; otherwise put φ_{0} = 0. For φ ∈P, the character

Selectable PDF notation • transcription unchecked
φ̃ = φ + t(φ)φ_{0}

is invariant under c and kills σ^{2}. It extends to a character ψ_{φ} of V by assigning value zero to σ: write an element in the two open cosets as uσ^{a}, a = 0, 1, and assign ̃φ(u). In multiplying two such forms, conjugating u does not change its value, and the sole carry σ^{2} has value zero. This proves multiplicativity; openness of U proves continuity. The extension still kills S. Let ϵ be the quotient character V /U → F_{2}, which also kills S. The map

Selectable PDF notation • transcription unchecked
P −→𝒞(V, S), φ −−→ψ_{φ} + t(φ)ϵ

is injective: its value at σ records t, and its restriction to U then recovers ̃φ and φ. Hence |P| ≤n.

On 𝒞(U, S), N = 1 + c satisfies N^{2}= 0 in characteristic two. Its kernel is P, and its image lies in P. Every nonempty fibre of N is a coset of its kernel. Therefore the whole space is finite of size at most n^{2}. For index one the spaces are equal and the bound is immediate because n ≥1.

Source: Development/AutomorphicArithmetic/AutomorphicVarieties/QuadraticExtensions/InvTwo.lean:7--117; De-velopment/AutomorphicArithmetic/AutomorphicVarieties/QuadraticExtensions/IdxTwo.lean:8--102

Lemma 8.60 (The cost of allowing one tame prime). Let V be a closed subgroup of an absolute Galois group. Suppose full inertia at a prime p ̸= ℓ lies in V , and the places above p have at most k orbits under V . If the space of F_{ℓ}-valued characters with a specified set of allowed ramified primes has size at most ℓ^{a}, then allowing p in addition raises this bound to at most ℓ^{a+k}.

Proof. Choose one place in each orbit. Wild inertia there is pro-p, hence has no nonzero continuous additive character to F_{ℓ}. The ℓ-primary tame inertia quotient is procyclic, so choose an element that generates it. Restriction of a character is determined by its value at that element. Record these at the at most k chosen places. If two newly allowed characters have the same record, their difference vanishes on full inertia at each chosen place, and then at every place above p, by conjugation invariance of additive characters. It belongs to the old space. Each record fibre consequently injects into that old space by subtracting a chosen member. There are at most ℓ^{k} records. This proves both finiteness and the asserted cardinality bound.

Source: Development/AutomorphicArithmetic/AutomorphicVarieties/QuadraticExtensions/TameGen.lean:7--60; De-velopment/AutomorphicArithmetic/AutomorphicVarieties/QuadraticExtensions/CharStep.lean:29--108

Proposition 8.61 (A uniform rank for the varying two-prime bases). Fix the abelian field K belonging to M, of degree n_{M}, and put K_{∞} = K(µ_{2}∞). There is an integer g, depending only on M, such that every admissible pair q, d has

Selectable PDF notation • transcription unchecked
( √^{√} )

X K(i, −q, −d) generated by at most R = 4(g + 6n_{M}) elements over ℤ_{2}.

Here X is the locally split cyclotomic module of Proposition 8.52. In particular R is chosen before q, d.

Proof. We first state exactly the established Iwasawa input used for this fixed abelian base. For the cyclo-tomic ℤ_{p}-extension of a finite abelian number field, the unramified Iwasawa module is finitely generated over ℤ_{p}. This follows from its finite torsion ℤ_{p}[[T]]-module structure and Ferrero–Washington’s µ = 0 theorem. A precise account is R. Sharifi, Iwasawa Theory, Proposition 3.2.11, Theorem 6.2.1 and the mod-ule structure Theorem 2.1.22; the locally split module is identified in Proposition 3.2.20.^{16} The hypotheses include p = 2; no odd-prime eigenspace assertion is being used here. To spell out the algebraic conse-quence, the structure theorem gives a pseudo-isomorphism to a sum of modules ℤ_{p}[[T]]/(f(T)^{a}), with f distinguished, and no summand ℤ_{p}[[T]]/(p^{b}), since µ = 0. The displayed summands are finite free over ℤ_{p} by division by the monic distinguished polynomial. The kernel and cokernel of the pseudo-isomorphism are finite. Noetherianity of ℤ_{p} therefore proves finite generation over ℤ_{p} itself.

Apply this at p = 2 to the fixed field K(i). Its unramified Iwasawa module has finite F_{2}-dimension g modulo two. By unramified Artin reciprocity, the continuous characters of G_{K∞} unramified at every finite place thus have at most 2^{g} elements.

There are at most 2n_{M} places of K_{∞} above q, and at most 4n_{M} above d. Indeed in ℚ(µ_{2}∞)/ℚ, a decomposition group at an odd prime p has image the closure of the powers of p in ℤ^{×}_{2}. If q ≡3 (mod 8), then v_{2}(q^{2}−1) = 3; lifting the exponent makes the even powers dense in 1+8ℤ_{2}, so that the full closure has index two. If d ≡7 (mod 16), then v_{2}(d^{2}−1) = 4, and the analogous index is four. Adjoining K increases the number of places by at most [K : ℚ] = n_{M}, proving the claimed orbit bounds. Since q, d are prime to 2m, where m is the fixed conductor, these full inertia groups lie in G_{K∞}. The cyclotomic extensions at conductor prime to the residue characteristic are unramified, as follows by distinct reduction of their roots.

Apply Lemma 8.60 twice, first at q and then at d. Characters of G_{√K∞} unramified outside q, d have at most 2^{g+6nM} elements. Now pass to the kernels adjoining ^{√}−q and −d. Each has index dividing two. Inertia outside q, d is contained in these kernels: the quadratic polynomials define extensions unramified away from 2q and 2d, respectively, and the infinite cyclotomic base already contains all 2-power roots. For the local prime two, adjoining ^{√}−q, with q ≡3 (mod 8), is unramified over ℚ_{2}, while −d ≡1 (mod 8) is a square in ℚ_{2}. Thus no new inertia at two is introduced. The quadratic unramified assertion follows from the integral quadratic discriminant −q ≡5 (mod 8); the square assertion follows by successively lifting an

^{16}https://www.math.ucla.edu/~sharifi/notes/iwasawa.pdf, printed pp. 43, 80, 83, 154. The original vanishing the-orem is B. Ferrero and L. C. Washington, Annals of Mathematics 109 (1979), 377–395, https://doi.org/10.2307/1971116.

odd square root modulo 2^{a}. Consequently the hypotheses concerning S in Lemma 8.59 hold at each step. Two squarings give

#{characters of G_{K∞(}^{√}_{−q,}^{√}_{−d)} unramified outside q, d} ≤2^{4(g+6nM)}.

Characters unramified everywhere and split above two form a subspace, so the same bound applies to them.

By the duality defining X, this is a bound of R for dim_{F2} X/2X. Lift a basis to X. The image Y of the resulting map ℤ^{R}_{2} →X is compact and closed, and X/Y = 2(X/Y ). Every finite quotient of the pro-2 module X/Y is then zero, by the elementary finite Nakayama argument; these quotients separate points. Thus X = Y , proving the generator bound uniformly in q, d.

Source: Development/AutomorphicArithmetic/AutomorphicVarieties/QuadraticExtensions/OrbTwo.lean:11--42; Development/AutomorphicArithmetic/AutomorphicVarieties/QuadraticExtensions/IneTwo.lean:29--44; Develop-ment/AutomorphicArithmetic/AutomorphicVarieties/QuadraticExtensions/QuadStep.lean:11--50; Development/AutomorphicArithmetic/AutomorphicVarieties/QuadraticExtensions/IsAlgClosureRat.lean:37--110

Corollary 8.62 (The fixed and two-prime tower inputs). The tower estimates needed in Proposition 8.41 hold for every fixed admissible abelian base and for the two-prime bases of Lemma 8.42. In the latter case their rank R is uniform before the auxiliary primes vary.

Proof. For a fixed abelian base containing the required roots, the Iwasawa and Ferrero–Washington state-ments used in Proposition 8.61 give finite generation over ℤ_{ℓ} of the maximal unramified module. Its quotient imposing complete splitting at ℓ is finitely generated too. Choose a generator bound R ≥1, and apply Propo-sition 8.58. This is uniform for all layers and twists of that fixed base. For the varying two-prime bases use the single R = 4(g+6n_{M}) just proved, then apply the same proposition to every base. Lemma 8.42 supplies their degree, conductor, root and relative-layer comparison. Thus all hypotheses used by the exceptional-set and window repair arguments, apart from their already explicit size conditions on the host parameters, have now been checked.

Lemma 8.63 (The primitive Bernoulli value and its nonvanishing). For a primitive Dirichlet character η of conductor f, and k ≥2,

Selectable PDF notation • transcription unchecked
L(1 −k, η) = −^{B}_{k }^{k,η}.

If η(−1) = (−1)^{k}, this value and B_{k,η} are nonzero, including f = 1.

Proof. For real t > 0, put

Selectable PDF notation • transcription unchecked
∑^{f}_{η(a)e}−at
F_{η}(t) =^{a}_{1}^{=1}_{ −e−ft} .

For Re s > 1, termwise integration of its absolutely convergent exponential expansion gives

Selectable PDF notation • transcription unchecked
∫ ∞
Γ(s)L(s, η) =_{0} t^{s−1}F_{η}(t) dt.

The Bernoulli generating function and reflection B_{j}(1 −x) = (−1)^{j}B_{j}(x) give, near zero,

Selectable PDF notation • transcription unchecked
∑_{(−1)}j_{Bj,η}
F_{η}(t) = t^{j−1}.
j≥0^{j!}

Subtract the terms through j = k on (0, 1). The remaining Mellin integral is holomorphic near s = 1 −k: the remainder is O(t^{k}), and the tail at infinity decays exponentially. The subtracted term with j = k has the

sole residue there, (−1)^{k}B_{k,η}/k!. The residue of Γ(s) at s = 1 −k is (−1)^{k−1}/(k −1)!; it follows by the recurrence Γ(s + 1) = sΓ(s) from its residue one at zero. Dividing the two Laurent expansions proves the asserted special value. This also treats the trivial conductor-one character: its possible pole at s = 1 is away from the point under consideration.

The primitive functional equation, with τ(η) =∑^{f}_{a=1}η(a)e^{2πia/f}, is

Selectable PDF notation • transcription unchecked
_{f}s−1_{Γ(s)}
L(1 −s, η) =_{(2π)s} {e^{−πis/2}+ η(−1)e^{πis/2}}τ(η)L(s, ̄η).

This is the convention in the source’s finite Fourier functional equation; an exact standard reference is NIST DLMF, (25.15.5), with Gauss sum convention (25.15.6).^{17} At s = k and matching parity, the braces are 2(−i)^{k}. The gamma factor is (k −1)!, and the other real factors are positive. The primitive Gauss sum has squared absolute value f. One checks this by expanding its square and using the orthogonality of the finite additive characters; primitivity makes its Fourier transform zero at nonunits and ̄η(b)τ(η) at a unit b. Plancherel then gives φ(f)|τ(η)|^{2} = fφ(f). For f = 1 the sum and its norm are one. Finally the Euler product at k > 1 is nonzero: its factors are nonzero and the sum of their deviations from one converges absolutely. Thus L(1 −k, η) ̸= 0, and the special value identity proves the claim for B_{k,η}.

Source: Development/AutomorphicArithmetic/AutomorphicVarieties/GaloisLibrary/BernoulliNumbers/BLval.lean:8--40; Development/AutomorphicArithmetic/AutomorphicVarieties/GaloisLibrary/BernoulliNumbers/BLprim.lean:7--13; Development/AutomorphicArithmetic/AutomorphicVarieties/GaloisLibrary/BernoulliNumbers/FeLFunctionConstMul.lean:13--81; Development/AutomorphicArithmetic/AutomorphicVarieties/GaloisLibrary/BernoulliNumbers/BNzC.lean:10--46

Proposition 8.64 (The integral nonzero Bernoulli product). For E/ℚ abelian, of conductor F and degree n, and k ≥2, the factors α_{k}(η) defined above are algebraic integers. On the parity set η(−1) = (−1)^{k}, their product is a nonzero rational integer M_{k} and satisfies

log |M_{k}| ≤n log{16F^{k+1}k!}.

Thus these hypotheses in Proposition 8.15 are established.

Proof. Put P_{k} = ∏_{p≤k+1}p. The von Staudt–Clausen denominator theorem says that for positive even j, the denominator of B_{j} is the product of the primes with p −1 | j. Thus P_{k}B_{j} ∈ ℤ for j ≤ k. For the remaining indices use B_{0} = 1, B_{1} = −1/2, and the vanishing of odd Bernoulli numbers above one.^{18} For f | F, expand a summand:

Selectable PDF notation • transcription unchecked
∑k (_{k})
f^{k−1}B_{k}(a/f) = B_{j}f^{j−1}a^{k−j}.
j=0^{j}

Multiplication by FP_{k} makes every rational coefficient integral. For j = 0 this uses F/f ∈ℤ; for j ≥1 it uses the integer P_{k}B_{j}. Every η_{0}(a) is zero or a root of unity, hence an algebraic integer. Their finite sum α_{k}(η) is therefore integral.

The Fourier series for B_{k}(x), valid on [0, 1] for k ≥2, gives

Selectable PDF notation • transcription unchecked
|B_{k}(x)| ≤^{2k!}_{k} ∑m^{−k}≤^{4k!}_{k} .
^{(2π)} m≥1^{(2π)}

^{17}https://dlmf.nist.gov/25.15.E5 and https://dlmf.nist.gov/25.15.E6. ^{18}NIST DLMF, (24.10.1) with its denominator statement, https://dlmf.nist.gov/24.10.E1; the convention B1 = −1/2 is fixed by the generating function (24.2.1).

Selectable PDF notation • transcription unchecked
∫ ∞

Here ∑_{m≥1}m^{−k} ≤2 follows by comparison with 1 +_{1} t^{−k} dt. The Fourier formula is DLMF (24.8.3), with precisely this endpoint range.^{19} Since |η_{0}(a)| ≤ 1, summing the f terms in the primitive definition yields

Selectable PDF notation • transcription unchecked
_{4f}k_{k!}
|B_{k,η}| ≤_{(2π)k} .

The elementary primorial bound( ) P_{k} ≤4^{k+1} has the following useful proof. For even N = 2m, the primes between m and 2m divide ^{2}_{m}^{m} ≤4^{m}; combine this with the induction bound for the product up to m. For

Selectable PDF notation • transcription unchecked
( )

odd N = 2m + 1, the primes strictly above m + 1 divide ^{2m}_{m}^{+1} ≤ 4^{m}, because the two equal largest binomial coefficients sum to at most 2^{2m+1}. Combine with the product up to m + 1. The cases N ≤2 start the induction, proving the product up to N is at most 4^{N}. Consequently

Selectable PDF notation • transcription unchecked
(_{ 4} )k
|α_{k}(η)| ≤16F^{k+1}k!_{2π} ≤16F^{k+1}k!,

using f ≤F and π > 2. This bound holds after every complex embedding, since embedding the root-of-unity values just replaces the Dirichlet character.

Take a cyclotomic field containing all the character values. Each of its Galois automorphisms permutes the characters of Gal(E/ℚ) and preserves their parity. It also preserves primitive conductors: the kernel and every factorization through a smaller modulus are preserved by an injective coefficient automorphism. The primitive Bernoulli finite sum therefore commutes with that automorphism. Thus the parity product is fixed by the whole Galois group, and is rational. It is integral by the first part, hence is an integer. Every factor is nonzero by Lemma 8.63. The product is consequently nonzero. There are at most n factors, each bounded as above, and the bound is at least one. Taking absolute values and logarithms gives the claimed estimate, also when the parity set is empty and the product is one.

Source: Development/AutomorphicArithmetic/AutomorphicVarieties/GaloisLibrary/BernoulliNumbers/Primden.lean:6--38; Development/AutomorphicArithmetic/AutomorphicVarieties/GaloisLibrary/BernoulliNumbers/Denint.lean:8--26; Development/AutomorphicArithmetic/AutomorphicVarieties/GaloisLibrary/BernoulliNumbers/Polybound.lean:6--68; Development/AutomorphicArithmetic/AutomorphicVarieties/GaloisLibrary/BernoulliNumbers/BUpper.lean:7--39; Development/AutomorphicArithmetic/AutomorphicVarieties/GaloisLibrary/BernoulliNumbers/Denupper.lean:7--38; Development/AutomorphicArithmetic/AutomorphicVarieties/GaloisLibrary/BernoulliNumbers/Ratprod.lean:7--49; Development/AutomorphicArithmetic/AutomorphicVarieties/GaloisLibrary/BernoulliNumbers/Prodint.lean:12--50

Lemma 8.65 (The residual generator bound for the minus group). Let E be a CM subfield of the F-th cyclotomic field, n = [E : ℚ], and let ℓ be odd. For every integral twist e, there is an integer g with

dim_{Fℓ}H^{1}(G_{E,Sℓ}, ℚ_{ℓ}/ℤ_{ℓ}(−e))[ℓ] ≤g ≤8ℓn log(ℓF).

The same bound applies to any subgroup, including the locally split minus subgroup.

Proof. Put K = E(µ_{ℓ}), m = [K : ℚ], M = ℓF. Then m ≤(ℓ−1)n, m ≥2, and K embeds in the M-th cyclotomic field. Restriction on the ℓ-torsion cohomology is injective: corestriction followed by restriction in the other order gives multiplication by [K : E], a unit modulo ℓ. The extension is unramified away from ℓ, so these are maps of the indicated S_{ℓ}-ramified cohomology groups. The long exact sequence for

Selectable PDF notation • transcription unchecked
0 −→A[ℓ] −→A→→^{ℓ} A −→0

^{19}https://dlmf.nist.gov/24.8.E3.

shows that H^{1}(G_{K,Sℓ}, A)[ℓ] is a quotient of H^{1}(G_{K,Sℓ}, A[ℓ]). Over K, the latter coefficient module is trivial and is isomorphic to µ_{ℓ}. Kummer theory identifies its cohomology with

{x ∈K^{×}/(K^{×})^{ℓ}: ord_{P} (x) ≡0 (mod ℓ) (P ∤ℓ)}.

One obtains the identification by sending x to g →→g(^{√ℓ}x)/^{√ℓ}x. Its kernel is the ℓ-th powers, and Hilbert 90 proves surjectivity; at a prime away from ℓ, the unramified condition is exactly the displayed valuation congruence.

Taking the fractional ideal root outside the primes above ℓ gives an exact sequence from the S_{ℓ}-units modulo powers into this group and then into the ℓ-torsion of the S_{ℓ}-class group. The unit rank is at most (m−1)+#S_{ℓ}≤2m−1; its cyclic roots-of-unity contribution has at most ℓ classes. The class group quotient has order at most h_{K}. Thus this Kummer group has at most ℓ^{2m}h_{K} elements.

We need a conductor-sensitive class bound, slightly better than Lemma 8.55. Its lattice argument gives an integral ideal of norm at most 2^{√}D_{K} in every class. For B ≥1, the number of ideals with norm at most B is at most ∑

Selectable PDF notation • transcription unchecked
B^{2} N(I)^{−2}= B^{2}ζ_{K}(2) ≤B^{2}ζ(2)^{m}< B^{2}2^{m}.
I

The Euler-factor inequality follows because there are at most m primes above each rational prime and their residue degrees are positive. Absolute convergence at two justifies multiplication of these positive Euler factors. The bound ζ(2) < 2 follows by the integral comparison. Hence h_{K} ≤4D_{K} 2^{m}. If C = ℚ(µ_{M}), the derivative argument of Lemma 8.54 gives D_{C} ≤M^{[C:ℚ]}. The discriminant tower identity gives D_{K}^{[C:K]}≤ D_{C}, and therefore D_{K} ≤M^{m}. Since m ≥2, M ≥2,

log h_{K} ≤ log 4 + m log 2 + m log M ≤3m log M.

Let g be the dimension of the original finite ℓ-torsion group. The preceding injection and quotient bounds give

g log ℓ≤2m log ℓ+ log h_{K} ≤5m log M.

Using log ℓ≥ log 2 > 5/8 yields g ≤8m log M ≤8ℓn log(ℓF). Restriction to a subgroup can only lower its ℓ-torsion dimension.

Source: Development/AutomorphicArithmetic/AutomorphicVarieties/OddResidueBounds/Imports.lean:11--46; Development/AutomorphicArithmetic/AutomorphicVarieties/OddResidueBounds/UnrInj.lean; Development/AutomorphicArithmetic/AutomorphicVarieties/OddResidueBounds/UnrSel.lean; Development/AutomorphicArithmetic/AutomorphicVarieties/OddResidueBounds/SelCard.lean; Development/AutomorphicArithmetic/AutomorphicVarieties/OddResidueBounds/LogH.lean

Lemma 8.66 (From the Stickelberger exponent to the constant 240). Suppose the locally split minus module is annihilated by ℓ^{r}, where

r log ℓ≤30n(e + 1) log(eℓF).

Then it is finite and satisfies (207).

Proof. For any ℓ^{r}-annihilated module M whose ℓ-torsion has dimension at most g, one has length M ≤rg. Induct on r. For r = 0, M = 0. The subgroup ℓM is annihilated by ℓ^{r−1}, and its ℓ-torsion is a subspace of M[ℓ]. The sequence

Selectable PDF notation • transcription unchecked
0 −→M[ℓ] −→M →→^{ℓ} ℓM −→0

therefore proves finiteness by induction and adds at most g to the length. Apply this with g ≤8ℓn log(ℓF) from Lemma 8.65. Since e ≥1, its product with the assumed exponent estimate is at most

30n(e + 1) log(eℓF) 8ℓn log(ℓF) ≤240ℓn^{2}(e + 1) log^{2}(eℓF).

Source: Development/AutomorphicArithmetic/AutomorphicVarieties/OddResidueBounds/LenPow.lean; Development/AutomorphicArithmetic/AutomorphicVarieties/OddResidueBounds/ThmBcOfPsiBound.lean