Section 6.1 · PDF pages 242–254
The comparison estimate and its numerical errors
The notation transcription is provisional. Check it against the printed mathematics when citing or reusing a formula.
6.1 The comparison estimate and its numerical errors
The local data choose disjoint finite sets R_{a}, C_{a} of places over ℓ, called rows and columns, with the rows the complex conjugates of the columns. Put d_{w} = [E_{w} : ℚ_{ℓ}]. Then
Selectable PDF notation • transcription unchecked
∑ d_{w} ≤2f, ∑ d_{u} ≤2f, |R_{a}|, |C_{a}| ≤2f. (135)
w∈Ra u∈CaIndeed ∑_{w|ℓ}[E_{w} : ℚ_{ℓ}] = [E : ℚ] = 2f, and every summand is at least one. Thus the inequalities hold for each subset; they do not require all places over ℓ to occur in the comparison.
The local level is s_{0}, and the discriminant error used in the source is
Selectable PDF notation • transcription unchecked
δ_{a} = ∑ (s_{0} + v_{ℓ}(j)).
1≤j<tFor a number L_{0} ≥s_{0} + 1 we have
Selectable PDF notation • transcription unchecked
0 ≤δ_{a} ≤tL_{0}. (136)To see this, the valuation sum is∑ v_{ℓ}((t −1)!). Counting the multiples of ℓ, ℓ^{2}, . . . gives v_{ℓ}((t −1)!) =
_{k≥1}⌊(t −1)/ℓ^{k}⌋≤t. The other part is (t −1)s_{0} ≤ts_{0}.
Here is the precise comparison inequality required before the numerical simplification. We keep its status distinct from the elementary lemmas which follow it.
Input 6.1 (Integral congruence–Selmer comparison). For every compatible host and setup, local index a,
good and good π ∈Z_{a}, there are a row–column declaration as above, an integer e ≥0, integers c_{0}(u) ≥0 for u ∈C_{a}, and b_{0} ≥0 such that
Selectable PDF notation • transcription unchecked
c_{a,π} ≤s_{a,π} + W_{ℓ}+ H_{ℓ}+ t ∑ λ_{a}(O_{a}/(ζ_{a,v} −1))
∑ v∈Qa
+ [e + d_{u}m(e + δ_{a}) + 2δ_{a} + h_{u} + d_{u}me + rc_{0}(u)]
u∈∑Ca
+ t (d_{w} + 2)b_{0} + 4V (|C_{a}| + 1)δ_{a} + t|C_{a}|h_{0}. (137)
w∈RaHere h_{u} is the normalized length of the specified W^{′} invariants at u, h_{0} is the inertia invariant exponent for the columns, and ζ_{a,v} is the value of the first cyclotomic block character on the chosen tame inertia generator at v. All displayed lengths are finite. The finite local conditions used to define s_{a,π} are the ones paired by local Tate duality with the corresponding dual conditions in the comparison.
The assertion is project-specific. It includes a comparison between the Hecke congruence ideal and the selected Galois lattices. Local Tate duality by itself is not that comparison. The exact remaining verification of this assertion is described below; it is not being discharged merely by assigning it a name.
The following reduction proves the length argument inside that comparison. It also specifies what the integral realization has to produce. Work for the moment with one coefficient ring O, fraction field K, and D = K/O. Write v_{O} for the valuation with v_{O}(ϖ) = 1 and put e_{O} = v_{O}(ℓ).
Lemma 6.2 (Relaxing finitely many local conditions). Let Q ⊂H be O-modules, let res_{w} : H →H_{w} be linear maps for w in a finite set U, and let F_{w} ⊂B_{w} ⊂H_{w}. Set
Selectable PDF notation • transcription unchecked
∩ ∩
S_{F} = Q ∩ res^{−}_{w}^{1}(F_{w}), S_{B} = Q ∩ res^{−}_{w}^{1}(B_{w}).
Selectable PDF notation • transcription unchecked
w∈U w∈U
If S_{F} and all B_{w}/F_{w} have finite length, then
Selectable PDF notation • transcription unchecked
∑
length_{O} S_{B} ≤ length_{O} S_{F} + length_{O}(B_{w}/F_{w}).
Selectable PDF notation • transcription unchecked
w∈U
Proof. Send z ∈ S_{B} to (res_{w}(z) + F_{w})_{w∈U} in ⊕_{w∈U}B_{w}/F_{w}. Its kernel is exactly S_{F} : vanishing of each component is exactly the corresponding defining local condition. Consequently S_{B}/S_{F} injects into that finite direct sum. Additivity of length in the sequence 0 → S_{F} → S_{B} → S_{B}/S_{F} → 0 proves both finiteness and the claimed inequality.
Here is the column module used for that argument. Let A be an O-algebra, with a central commutative coefficient algebra T^{′}, a homomorphism r : G → A^{×}, and a character χ : G → (T^{′})^{×}. Put u_{g} = χ(g)^{−1}r(g). Choose E, e_{1} ∈ A and δ ∈ T^{′} with E^{2} = δE and e^{2}_{1} = e_{1}. Let Y ⊂ A be the left A-submodule generated by y_{g} = (u_{g} −1)E. Given subgroups G_{l} ⊂G for finitely many columns l, integers j_{l} ≥0, and elements K_{la}(g), k_{l}(g) ∈A, define 𝒩 as the quotient of
Selectable PDF notation • transcription unchecked
⊕ ⊕j_{l} )
Y ⊕ Aν_{l} ⊕ Aσ_{la} ⊕Aσ_{0l}
l a=1by the relations
Selectable PDF notation • transcription unchecked
∑ Eν_{l} = 0, (1 −e_{1})σ_{la} = 0, (δ −E)σ_{0l} = 0,
y_{g} + (u_{g} −1)ν_{l} −δ K_{la}(g)σ_{la} −δk_{l}(g)σ_{0l} = 0 (g ∈G_{l}). (138)
aAll occurrences of y_{g} in this formula denote its image from Y . For an A-module M set ℋ_{M} = Hom_{A}(𝒩, M) and
Selectable PDF notation • transcription unchecked
Ψ_{M}(ϕ)(g) = ϕ(y_{g}).The identity y_{gh} = y_{g} + u_{g}y_{h} shows that this is a cocycle for the action g · m = u_{g}m. Thus Ψ_{M} is an O-linear map to Z^{1}(G, M), and its kernel consists exactly of the maps vanishing on the whole summand Y . The last assertion uses the fact that the y_{g} generate Y as an A-module.
Lemma 6.3 (The coboundary error). Suppose M = D^{a}, EM = 0, and δ acts on M as a nonzero δ_{O} ∈O. Suppose also that ∑
Selectable PDF notation • transcription unchecked
∑
E = t_{g}r(g), t_{g}χ(g) acts on M as δ_{O},
Selectable PDF notation • transcription unchecked
g g
where the sums are finite and t_{g} ∈T^{′}. If Ψ_{M}(ϕ)(g) = u_{g}v−v for some v ∈M, then δ_{O}^{2}v = 0. In particular the submodule im Ψ_{M} ∩B^{1}(G, M) has length at most 2a v_{O}(δ_{O}).
Proof. Write τ =∑_{g}t_{g}χ(g) and z = ∑_{g}t_{g}χ(g)y_{g} ∈Y . In A this is z = (E −τ)E. Applying ϕ and using the proposed coboundary gives
Selectable PDF notation • transcription unchecked
ϕ(z) =∑t_{g}χ(g)(u_{g}v −v) = Ev −τv = −δ_{O}v.
gSince τ is central and E^{2} = δE, we also have Ez = δz. The action of E on M is zero, so 0 = ϕ(Ez) = δ_{O}ϕ(z) = −δ_{O}^{2}v. Thus every coboundary in the indicated intersection is the image of an element of M[δ_{O}^{2}] under v →→(g →→u_{g}v −v). For a ̸= 0 in O the map O/aO →D[a], z →→z/a + O, is an isomorphism. It follows that length_{O} M[δ_{O}^{2}] = 2a v_{O}(δ_{O}). Length cannot increase on taking an image, which proves the last assertion.
Lemma 6.4 (Evaluation bounds for the column kernel). Retain the preceding assumptions and suppose each k_{l}(g) is a left multiple of E. Suppose e_{1}M ≃D. For each column let e_{l} ≥0 be such that
Selectable PDF notation • transcription unchecked
[ ∑ ]
g →→ K_{la}(g)s_{a} ∈B^{1}(G_{l}, M), s_{a} ∈e_{1}M =⇒ ℓ^{el}s_{a} = 0 for every a.
Selectable PDF notation • transcription unchecked
a
Put h_{l} = length_{O} H^{0}(G_{l}, M) and ∆= v_{O}(δ_{O}). If these lengths are finite, then
Selectable PDF notation • transcription unchecked
∑[ ]
length_{O} ker Ψ_{M} ≤ j_{l}(e_{O}e_{l} + ∆) + h_{l} + a |C|∆.
Selectable PDF notation • transcription unchecked
l
Proof. Take ϕ ∈ ker Ψ_{M}. Relation (138) gives ϕ(σ_{la}) ∈e_{1}M, δ_{O}ϕ(σ_{0l}) = 0, and, because k_{l}(g)M = 0,
Selectable PDF notation • transcription unchecked
(u_{g} −1)ϕ(ν_{l}) =∑K_{la}(g)δ_{O}ϕ(σ_{la}).
aThe assumed local lifting bound therefore kills δ_{O}ϕ(σ_{la}) by ℓ^{el}. Record all values ϕ(σ_{la}) and ϕ(σ_{0l}). This defines a map on ker Ψ_{M} with target
Selectable PDF notation • transcription unchecked
⊕(D[ℓ^{el}δ_{O}]^{jl}⊕M[δ_{O}]),
lusing the chosen isomorphism e_{1}M ≃D. Its target has length ∑_{l}j_{l}(e_{O}e_{l} + ∆) + a|C|∆. On the kernel of this recording map, the displayed cocycle relation says ϕ(ν_{l}) ∈H^{0}(G_{l}, M). Recording the ν_{l} values is injective there: if they too vanish, the map vanishes on Y and on every generator of the remaining finite direct sum, hence on 𝒩. Thus the first kernel has length at most ∑_{l}h_{l}. Add the lengths of the kernel and image of the recording map to obtain the result.
Combining these two lemmas with the quotient Z^{1}(G, M) →H^{1}(G, M) gives the main-block bound used by the source. Indeed, the length of ℋ_{M} is the sum of the length of ker Ψ_{M}, the length of the cobound-aries in its image, and the length of its image in H^{1}. The last image lies in the Selmer group with enlarged row conditions and the original column conditions. Lemma 6.2 bounds it by the selected Selmer length plus the sum of the row-condition quotients. Therefore
Selectable PDF notation • transcription unchecked
∑ ∑
length_{O} ℋ_{M} ≤ L_{w} + [j_{l}(e_{O}e_{l} + ∆) + h_{l}] + s + 4a(|C| + 1)∆. (139)
Selectable PDF notation • transcription unchecked
w l
Here L_{w} bounds the row quotient length and s bounds the strict Selmer length. The direct argument just given has the smaller discriminant error a(|C| + 2)∆; the source enlarges it to the displayed 4a(|C| + 1)∆. This harmless enlargement is part of the constants used in the subsequent estimate.
Lemma 6.5 (The character blocks). Suppose a one-dimensional block M = D has e_{1}M = 0, EM = 0, and the same scalar and coboundary identities as in Lemma 6.3. Suppose its strict global cohomology group has length at most S, its enlarged row quotients have lengths at most e_{O}b_{0}(d_{w} + 2), its auxiliary quotients have lengths at most v_{O}(ζ_{v} −1), and length_{O} H^{0}(G_{l}, M) ≤H_{0}. Then
Selectable PDF notation • transcription unchecked
∑
length_{O} ℋ_{M} ≤S + e_{O}b_{0} (d_{w} + 2) + 4(|C| + 1)∆+ |C|H_{0}
Selectable PDF notation • transcription unchecked
∑ w∈R
+ v_{O}(ζ_{v} −1).
v∈QFor the first character block instead suppose u_{g} acts trivially, e_{1}M = 0, and E, δ both act by δ_{O}. Suppose at each column some g has k_{l}(g) acting by ℓ^{el}δ_{O} times a unit. With strict global bound H and the same row and auxiliary quotient bounds, one has
Selectable PDF notation • transcription unchecked
∑ ∑
length_{O} ℋ_{M} ≤H + (e_{O}e_{l} + 2∆) + e_{O}b_{0} (d_{w} + 2)
Selectable PDF notation • transcription unchecked
l ∑ w∈R
+ 4(|C| + 1)∆+ v_{O}(ζ_{v} −1).
v∈QProof. In the higher character block the σ_{la} values vanish, since they belong to e_{1}M. For a map in ker Ψ_{M}, the σ_{0l} values are killed by δ_{O}, and after these values have been recorded, the ν_{l} values belong to H^{0}(G_{l}, M). The same generator argument as above gives length_{O} ker Ψ_{M} ≤|C|(∆+H_{0}). The coboundary intersection has length at most 2∆. Relaxing the row and auxiliary conditions gives the remaining summands. Enlarge (|C| + 2)∆ to 4(|C| + 1)∆.
For the first character block, Eν_{l} = 0 gives δ_{O}ϕ(ν_{l}) = 0. For ϕ ∈ ker Ψ_{M} the last relation in (138), evaluated at the specified g, gives ℓ^{el}δ_{O}^{2}ϕ(σ_{0l}) = 0; the unit can be cancelled. The σ_{la} values are again zero. Recording ν_{l} and σ_{0l} is injective and bounds the kernel length by ∑_{l}(e_{O}e_{l}+2∆)+|C|∆. Every coboundary is zero for a trivial action. Thus there is no middle coboundary length in this case. The strict global bound and local relaxation give the asserted inequality after enlarging |C|∆ to 4(|C| + 1)∆.
There is one main block of dimension r and t character blocks. For character block c ≥ 2 use S = length_{O} Sha_{ℓ}(c−1); for the first use H = e_{O}v_{ℓ}(h_{E}). Summing (139) and the two character-block estimates gives
Selectable PDF notation • transcription unchecked
∑t ∑ ∑ ∑
length_{O} ℋ_{Mc} ≤ L_{w} + [e_{O}e_{l} + j_{l}(e_{O}e_{l} + ∆) + 2∆+ h_{l}] + s + length_{O} Sha_{ℓ}(j)
Selectable PDF notation • transcription unchecked
c=0 w∈R l∈C ∑ 1≤j<t
+ e_{O}v_{ℓ}(h_{E}) + te_{O}b_{0} (d_{w} + 2)
w∈R ∑
+ 4(r + t)(|C| + 1)∆+ t|C|H_{0} + t v_{O}(ζ_{v} −1).
v∈QThe term |C|H_{0} occurs only t −1 times, so replacing that coefficient by t preserves the inequality. Every other character error occurs t times. This explains in particular both factors of t in the last line of the global comparison.
To obtain Input 6.1 from this calculation, two additional identifications remain essential. First, the inte-gral matrix construction must give
Selectable PDF notation • transcription unchecked
∑t
length_{O}(O/I_{a,π}) ≤ length_{O} ℋ_{Mc}. (140)
Selectable PDF notation • transcription unchecked
c=0
Second, the realized local conditions must give exactly the strict Selmer and Sha groups just used, and the local row quotients must match the claimed lifting bounds. Under these identifications divide by e_{O}, use ∆/e_{O} = δ_{a}, j_{l} ≤ d_{l}m, and pair each row with its conjugate column. The row bound is L_{w}/e_{O} ≤ d_{l}me+ rc_{0}(l); the column invariant term is divided by e_{O} to give the normalized h_{l} in (137). The displayed sum then becomes exactly (137), with V = r + t.
The determinant argument proving (140) is as follows. Its hypotheses are properties of a simultaneous integral matrix model, so we specify them rather than identifying them with local duality.
Lemma 6.6 (A nonzero column map away from the selected spectrum). Let V be a finite-dimensional K-representation of A on which δ acts by δ_{x} ̸= 0. Suppose EV and e_{1}V both have dimension one, and every nonzero η ∈EV is moved by some u_{g}. For each column l write d_{l}(v)(g) = u_{g}v −v. Suppose a subspace L_{l} ⊂ Map(G_{l}, V ) contains im d_{l} and all functions g →→K_{la}(g)s for s ∈e_{1}V , as well as g →→k_{l}(g)η for η ∈EV . Suppose
dim_{K}(L_{l}/ im d_{l}) ≤j_{l} + dim_{K} ker d_{l}.
Then Hom_{A}(𝒩, V ) ̸= 0.
Proof. Choose nonzero η ∈EV and s_{1} ∈e_{1}V . Then Eη = δ_{x}η and e_{1}s_{1} = s_{1}. For each column consider the linear map
Selectable PDF notation • transcription unchecked
V ⊕K^{jl}⊕K −→L_{l}, (v, (c_{a}), c_{0}) −−→d_{l}(v) −∑c_{a}K_{la}(·)s_{1} −c_{0}k_{l}(·)η.
aThe rank assumption and rank–nullity give dim L_{l} ≤ dim V + j_{l}, whereas this map has domain dimension dim V + j_{l} + 1. It therefore has a nonzero kernel vector (v_{l}, (c_{la}), c_{0l}). Since EV = Kη, either Ev_{l} = 0 or rescaling the vector makes Ev_{l} = η.
If some column l_{0} has Ev_{l0} = 0, define a map on the generators of the column module by sending Y to zero,
Selectable PDF notation • transcription unchecked
ν_{l0} →→δ_{x}v_{l0}, σ_{l0a} →→c_{l0a}s_{1}, σ_{0l0} →→c_{0l0}η,and sending all other column generators to zero. These values satisfy the first three relations in (138). The fourth is δ_{x} times the kernel equation just constructed. Thus the map descends to an A-linear map 𝒩 →V . It is nonzero because the chosen kernel vector is nonzero and δ_{x}, s_{1}, η are nonzero.
Otherwise choose every kernel vector normalized by Ev_{l} = η. Choose w_{0} with Ew_{0} = η. Send y ∈Y to yw_{0}, and set
Selectable PDF notation • transcription unchecked
ν_{l} →→δ_{x}v_{l} −η, σ_{la} →→c_{la}s_{1}, σ_{0l} →→c_{0l}η.Again the first three relations follow from the displayed actions of E and e_{1}. On y_{g} the map has value u_{g}η−η; this cancels the −η contribution from the value of ν_{l}, and the remaining fourth relation is the kernel equation multiplied by δ_{x}. Hence it too descends to an A-linear map. Some u_{g}η −η is nonzero by assumption, so the descended map is nonzero. This proof includes an empty column set: in that case the second construction on Y alone applies.
Proposition 6.7 (The congruence determinant argument). Let T ⊂ O^{X} and T^{′} ⊂ O^{X′} be commutative O-algebras, let p : X^{′}→X, and let Z ⊂X. Suppose T^{′} is spanned over the pullback of T by b_{1}, . . . , b_{q} ∈ O^{X′}. Suppose that for each∏ x /∈ Z the restrictions of the b_{i} to p^{−1}(x) are O-linearly independent. Let A ⊂_{x}′_{∈X}′ M_{d}(K) be a left Noetherian O-algebra with this determinant property: for each square matrix Θ over A, the function x^{′} →→ det Θ(x^{′}) belongs to T^{′}, where evaluation replaces each entry by its d by d matrix.
Fix π^{′}∈X^{′} above π ∈Z. Suppose its evaluation is integral and block triangular, with diagonal maps ρ_{c} : A ↠M_{dc}(O) for 0 ≤c ≤t. Let 𝒩 be a finitely presented left A-module with g generators. For every x^{′} above X \ Z, suppose
Hom_{A}(𝒩, K_{x}^{d}′) ̸= 0.
If the residue field of O has more than g∑_{c}d_{c} elements, and all Hom_{A}(𝒩, D_{ρ}^{d}_{c}^{c}) have finite length, then
Selectable PDF notation • transcription unchecked
∑t
length_{O}(O/I_{π}) ≤ length_{O} Hom_{A}(𝒩, D_{ρ}^{d}_{c}^{c}).
Selectable PDF notation • transcription unchecked
c=0
Here I_{π} is the congruence ideal defined from T, Z.
Proof. Choose a finite presentation of 𝒩. Write its relation matrix as P, with g columns and h rows, so that applying Hom_{A}(−, M) identifies Hom_{A}(𝒩, M) with the kernel of the relation map P : M^{g} →M^{h}. For block c, let P_{c} be the hd_{c} by gd_{c} matrix over O obtained by applying ρ_{c}. Smith reduction over the discrete valuation ring shows that P_{c} has full column rank: otherwise its kernel on D^{gdc} contains a copy of D and has infinite length. If s_{c} is the length of that kernel, Smith reduction also shows that s_{c} is the minimum valuation of the maximal minors of P_{c}.
We claim that one matrix C with g rows and h columns over A can be chosen so that
v_{O} det(ρ_{c}(C)P_{c}) = s_{c} for every c. (141)
For each c, choose a maximal minor of valuation s_{c} and its row-selection matrix. Surjectivity of ρ_{c} lifts this row-selection matrix, entry by entry in d_{c} by d_{c} blocks, to a matrix C^{(c)} over A. Form C(z) =∑_{c}z_{c}C^{(c)} with central scalar variables. By the Cauchy–Binet formula every coefficient of det(ρ_{c}(C(z))P_{c}) is divisible by ϖ^{sc}. Dividing by that power and reducing modulo ϖ gives a nonzero polynomial F_{c}: setting the cth variable to 1 and the others to 0 gives a unit. Its total degree is at most gd_{c}. The product ∏_{c}F_{c} is nonzero and has total degree at most g∑_{c}d_{c}.
A nonzero polynomial of total degree less than the cardinality of a field cannot vanish at every tuple in that field. This follows by induction on the number of variables, expanding in the last variable and using that
a nonzero univariate polynomial has at most its degree many roots. Choose a residue tuple on which ∏_{c}F_{c} is nonzero and lift it to O. The corresponding C satisfies (141).
The determinant property gives f ∈T^{′} with f(x^{′}) = det((CP)(x^{′})). For p(x^{′}) /∈Z, a nonzero homo-morphism from 𝒩 to K_{x}^{d}_{′} gives a nonzero vector in the kernel of P(x^{′}). It is also in the kernel of (CP)(x^{′}), so f(x^{′}) = 0. Write f =∑_{i}(a_{i} ∘p)b_{i}, a_{i} ∈T. For x /∈Z, vanishing on the whole fiber p^{−1}(x) and the assumed independence imply a_{i}(x) = 0 for every i. Thus a_{i}(π) ∈I_{π}, and f(π^{′}) =∑_{i}a_{i}(π)b_{i}(π^{′}) ∈I_{π}.
At π^{′}, block triangularity gives f(π^{′}) = ∏_{c} det(ρ_{c}(C)P_{c}). Every factor is nonzero by (141); conse-quently ∑
length_{O}(O/I_{π}) ≤v_{O}(f(π^{′})) = s_{c}.
Selectable PDF notation • transcription unchecked
c
The definition of s_{c} identifies the last summands with the Hom-module lengths in the proposition, which proves it.
If the residue field is too small, the source first makes a finite unramified extension of the coefficient ring with sufficiently large residue field. For a finitely presented module, flat base change commutes with its Hom kernel in the finite presentation. For a finite-length module over O, an unramified extension pre-serves its length when measured over the new coefficient ring: a composition series has factors the residue field, which become the new residue field. Thus both sides of the determinant inequality descend with the same normalization. This reduction requires that the realized matrix algebra, its determinant functions, the separated fibers and the column module are indeed the compatible base changes. That compatibility is part of the integral matrix realization; it cannot be inferred just from the existence of an unrelated representation over a larger field.
Thus the length filtration, the coboundary error, the three kinds of blocks, their summation, and the deter-minant step have proofs above. The simultaneous integral matrix realization with the properties in Proposi-tion 6.7 (including the stated base-change compatibility) together with the local-condition identifications and cyclotomic lifting data for the selected Galois lattices are proved in Subsection 6.7 (Propositions 6.97–6.100 and Theorem 6.101), universally over every compatible setup GlobalTheory.NLink.SetupAlt.
Lemma 6.8 (The cyclotomic discriminant error). Suppose the chosen cyclotomic element has value x = 1 + ℓ^{s0}u in ℤ^{×}_{ℓ}, where u is a unit, s_{0} ≥1, and s_{0} ≥2 if ℓ= 2. For a unit a ∈O^{×} put
Selectable PDF notation • transcription unchecked
δ_{O} = ∏ (a −ax^{−j}).
1≤j<tThen δ_{O} ̸= 0 and
Selectable PDF notation • transcription unchecked
^{vO(δO)} = ∑ (s_{0} + v_{ℓ}(j)) = δ_{a}.
^{eO} 1≤j<tProof. For z ∈ℓℤ_{ℓ} with the stated stronger divisibility at 2, expand (1 +( ) z)^{ℓ}−1. Its first term is ℓz. Every intermediate term has valuation at least 1 + 2v_{ℓ}(z), since ℓ divides _{k}^{ℓ} for 1 ≤ k < ℓ. Its last term has valuation ℓv_{ℓ}(z) > 1 + v_{ℓ}(z) under exactly the stated divisibility conditions. The first term is consequently the unique term of least valuation, so v_{ℓ}((1 + z)^{ℓ}−1) = v_{ℓ}(z) + 1. If b is positive and prime to ℓ, the same expansion of (1 + z)^{b}−1 has first term bz of valuation v_{ℓ}(z) and all other terms of strictly larger valuation. Writing j = ℓ^{k}b therefore gives v_{ℓ}(x^{j}−1) = s_{0} + k by iteration.
Now a −ax^{−j} = ax^{−j}(x^{j}−1). Its first factor is a unit, so its O-valuation is e_{O}(s_{0} + v_{ℓ}(j)) and the factor is nonzero. Sum these valuations over the finite product. For t = 1 the product is 1 and both sums are zero, giving the same conclusion.
For completeness, the published local statements relevant to those identifications can be specified pre-cisely. They do not involve the Hecke congruence ideal.
Lemma 6.9 (The integral algebra of ordered distinct roots). Let X be finite and nonempty, T ⊂O^{X}, and let P_{x} ∈ O[z] be monic of degree t. Suppose all its coefficient functions belong to T and every P_{x} splits into distinct roots in O. Let X^{′} consist of all pairs (x, (α_{0}, . . . , α_{t−1})) with P_{x} = ∏_{i}(z −α_{i}), and let p : X^{′}→X. The algebra T^{′}⊂O^{X′} generated by the pullback of T and the ordered root functions is finite free of rank t! over T. There are t! generators b_{i} which form a basis and whose restrictions to every fiber of p are O-linearly independent.
Proof. The coefficient functions define a monic polynomial P ∈T[z] of degree t. Construct its universal ordered splitting algebra recursively. First adjoin a root by T_{1} = T[z_{0}]/(P(z_{0})). Division by a monic polynomial gives the basis 1, z_{0}, . . . , z_{0}^{t−1} over T. In T_{1}[z] divide P by z −z_{0}; the remainder is zero, so the quotient is monic of degree t −1. Adjoin its first root and continue. At the end one obtains an algebra S generated by t roots, free over T of rank t(t −1) · · · 1 = t!. It has the required universal property: an ordered factorization in any commutative T-algebra first gives a map out of T_{1}, and then a map out of each successive root quotient. These maps are unique since the roots generate the algebra. The construction for t = 0 is S = T.
Every point of X^{′} therefore gives an evaluation S →O. Together they define S →O^{X′}, with image exactly T^{′}: the image contains the pullback coefficients and roots, and every element of S is an expression in those generators. Choose a T-basis β_{1}, . . . , β_{t!} of S, and let b_{i} be its evaluations.
Fix x. Over K, the t! permutations of the distinct roots give t! distinct algebra homomorphisms S →K extending evaluation at x. They are linearly independent as functions. Indeed, a nonzero relation with the fewest terms, evaluated at a product ab and compared with its value at b multiplied by one of the homomor-phisms’ values at a, gives a relation with fewer terms. Choose a separating two of the homomorphisms; the new relation is nonzero, a contradiction.
Each of these homomorphisms is determined linearly by its values on the β_{i}. The square matrix of those values consequently has independent rows and is invertible over K. If ∑_{i}c_{i}b_{i} vanishes on p^{−1}(x), with ∑c_{i} ∈O, this matrix sends (c_{i}) to zero, so all c_{i} = 0. This proves fiber independence. Finally, if an element
_{i}a_{i}β_{i}, a_{i} ∈T, has zero evaluation everywhere, fiber independence gives a_{i}(x) = 0 for every x. As T is an algebra of functions on X, this means a_{i} = 0. Hence S →T^{′} is injective as well as surjective, proving the basis assertion.
Lemma 6.10 (The canonical integral matrix order). In the preceding construction let r_{0}(g, x) ∈ GL_{d}(O) be a finite family of representations. Let A be the T^{′}-span of the matrices r(g, x^{′}) = r_{0}(g, p(x^{′})). It is an O-finite algebra; every evaluation is integral. Suppose evaluation at π is block triangular with a main block of dimension r and t one-dimensional character blocks. If the reduction of the main block’s image spans M_{r}(O/ϖO), all its block evaluation maps A →M_{dc}(O) are surjective.
Suppose in addition that e_{B}, e_{1} and a perturbation s belong to the original T-span of r_{0}(G), with e^{2}_{B}= e_{B}, e_{1}^{2}= e_{1}, and e_{1} of rank one. Let X_{0} = e_{B}r_{0}(g_{0})e_{B} + s preserve the t-dimensional space e_{B}K^{d}, have zero action on its complementary summand, and have the ordered distinct eigenvalues α_{0}, . . . , α_{t−1} there. Then
Selectable PDF notation • transcription unchecked
∏t−1 ∏t−1
E = e_{B} (X_{0} −α_{i}e_{B}), δ = (α_{0} −α_{i})
i=1 i=1belong to A and T^{′} respectively, satisfy E^{2}= δE, and have rank one and nonzero δ at every point. If e_{B}(π) projects onto the character blocks in that order, then E acts by δ(π) on the first character and by zero on every other block. If e_{1}(π) is supported on the main block, then e_{1}D^{r}≃D and e_{1} kills the character blocks.
Proof. The span is closed under multiplication because r(g)r(h) = r(gh) and the coefficients in T^{′} are central. It contains the identity, so it is an algebra. The set X^{′} is finite: each fiber has t! ordered root tuples. All evaluations lie in M_{d}(O), so A is an O-submodule of the finite free module M_{d}(O)^{X′}. Since a discrete
valuation ring is Noetherian, A is finite over O. Every ascending chain of left ideals of A is an ascending chain of O-submodules, so A is left Noetherian too.
For the main block let B be the image in M_{r}(O). The assumed residue span gives B + ϖM_{r}(O) = M_{r}(O). Its finite cokernel M_{r}(O)/B therefore equals its multiple by ϖ and is zero by Nakayama’s lemma. This proves main-block surjectivity. Each one-dimensional block map contains all scalars from O, so is surjective directly.
Pullback of the original span lies in A; therefore e_{B}, e_{1}, X_{0} belong to A, and the ordered roots are in T^{′}. The polynomial formula for E belongs to A. In an eigenbasis on e_{B}K^{d} it has value δ on the α_{0}-line and zero on all other eigenlines; it also vanishes on ker e_{B}. Distinctness makes δ ̸= 0, and this diagonal description proves both rank one and E^{2}= δE. Evaluation on the specified blocks at π gives the asserted scalar actions.
Finally an integral rank-one idempotent splits O^{r} = e_{1}O^{r}⊕(1 −e_{1})O^{r}. Its first summand is free of rank one over the discrete valuation ring. Passing to K/O in this split decomposition gives e_{1}D^{r} = e_{1}K^{r}/e_{1}O^{r}≃K/O. Its stipulated support makes its other blocks zero.
Lemma 6.11 (Determinants in the character-coefficient algebra). Let A be the T^{′}-span of r(G) as above. Suppose every coefficient function of det(1−zr(g)) belongs to T^{′} for every g ∈G. For every square matrix Θ over A, its evaluated block determinant is a function in T^{′}.
Proof. First consider a d by d formal matrix series U(z) with U(0) = 1 and every coefficient in A. Fix a degree N. We may factor U modulo z^{N+1} into finitely many factors
Selectable PDF notation • transcription unchecked
1 −cz^{j}r(g), c ∈T^{′}, j ≥1.To construct the factorization, suppose a current residual series is 1 + z^{j}A_{j} modulo z^{j+1}. Express A_{j} as a finite T^{′}-linear combination of r(g) and multiply by the inverses of the corresponding factors. Each inverse exists modulo z^{N+1} by the finite geometric series. The coefficient at degree j is cancelled; all cross terms have degree at least 2j ≥ j + 1. The new coefficients remain in A, since products of the r(g) are again group matrices. Repeating through N proves the factorization.
The determinant of each factor has coefficients in T^{′} by the assumed characteristic-polynomial condition, after substituting cz^{j} for z. Multiplicativity of determinant shows that the coefficients of det U through degree N belong to T^{′}. As N was arbitrary, this holds for every coefficient.
Now apply this result to the diagonal pivot blocks in a formal block elimination of 1 −zΘ. Its first d by d block has constant term 1, so its inverse is a formal series with coefficients in A. Multiplying on the left by the block lower-unitriangular matrix which eliminates the first block column does not change the determinant. The remaining block matrix is the Schur complement; it again has constant term 1 and all coefficients in A. Induct on the number of block rows. Each pivot determinant has coefficients in T^{′}, and therefore so does det(1 −zΘ).
If Θ has g block rows, the coefficient of z^{gd} in this last polynomial is (−1)^{gd} det Θ. Since T^{′} is closed under signs, the required determinant function belongs to T^{′}. This argument uses only finite congruences at each coefficient; it involves no convergence assertion or division by a factorial in O.
In the intended model, the characteristic-polynomial condition in Lemma 6.11 is the source’s condition that the coefficients of the characteristic polynomial of every Galois matrix lie in the Hecke algebra; pull-back puts them in T^{′}. Thus the determinant property needed for Proposition 6.7 follows from that condition. Lemmas 6.9 and 6.10 likewise prove the root-order, finiteness and block-surjectivity portions of the realiza-tion from their displayed matrix hypotheses. The remaining matrix task is to obtain the adapted idempotents, separable corner perturbation and residual full-matrix span for the selected Galois family, with the specified coefficient-change compatibility. Those are existence and compatibility assertions about that family.
Lemma 6.12 (Integral normalization at the selected character). Suppose the rational representation r_{0}(·, π) is isomorphic to
Selectable PDF notation • transcription unchecked
ρ ⊕χ_{1} ⊕· · · ⊕χ_{t},where ρ and all the characters have the specified integral lattices. Replace only r_{0}(·, π) by this block di-agonal integral representation, leaving the family unchanged at every x ̸= π. The old and new T-span algebras are isomorphic as O-algebras by rational conjugation at π and identity elsewhere. This isomor-phism preserves all polynomial identities between their elements, pointwise ranks, and corner characteristic polynomials. It also preserves the condition that the characteristic-polynomial coefficient functions belong to T.
Assume the main summand is irreducible of dimension greater than one and that the characters are pair-wise distinct. Adapted idempotent and corner data expressed using any such ordered rational decomposition give the same specified main and character actions after this normalization. In particular a rank-t idempo-tent acting by one on every character becomes the standard character projection; an orthogonal rank-one idempotent is supported on the main block.
Proof.∑ Let P ∈ GL_{d}(K) conjugate the old representation at π to the chosen direct sum. For a(x) =
_{g}f_{g}(x)r_{0}(g, x), with finitely many f_{g} ∈ T, define a^{n}(x) = ∑_{g}f_{g}(x)r_{0}^{n}(g, x). At x ̸= π it equals a(x), while a^{n}(π) = Pa(π)P^{−1} in M_{d}(K). Thus the definition is independent of the expression for a. It is integral because the latter expression is a sum of integral matrices with coefficients in O. Conjugating back defines its inverse by the same argument. Pointwise addition, multiplication and scalar multiplication are preserved, proving the algebra isomorphism even though P itself need not be integral.
Conjugation preserves matrix rank and characteristic polynomial. For a corner pair (e, A) its polynomial is det(ze −A + 1 −e); conjugating both entries conjugates the matrix inside this determinant, so preserves this polynomial too. All other asserted identities follow from the algebra isomorphism. The characteristic-polynomial functions do not change at any point, so retain their membership in T.
To check independence of the chosen decomposition, compare its intertwiner with the fixed ordered direct sum. There is no nonzero intertwining map between the irreducible main summand and a one-dimensional summand, and there is none between two distinct characters: the kernel and image of an intertwiner are invariant, and irreducibility would otherwise make the summands isomorphic. The change between decom-positions is therefore block diagonal, with a scalar on each character. It preserves the asserted character actions of every element of the span algebra. A rank-t idempotent already having t one-dimensional identity blocks has rank zero on the main block; that block is zero. Orthogonality to this projection forces the rank-one idempotent to have zero character blocks. The same block diagonal comparison preserves the prescribed scalar actions of the corner operator.
Lemma 6.13 (Cancellation of the fixed character summands)⊕. For representations V, W over a field K, an equivariant isomorphism V ⊕⊕^{t}_{i=1}K(χ_{i}) ≃W ⊕_{i}^{t}_{=1}K(χ_{i}) implies an equivariant isomorphism V ≃W. The characters need not be distinct for this cancellation assertion.
Proof. It suffices to cancel one character and then repeat. Let e : V ⊕K(χ) →W ⊕K(χ) be the isomor-phism. Choose a χ-eigenvector v_{0} ∈ V such that the second coordinate δ of e(v_{0}, 1) is nonzero. Such a choice always exists. If the second coordinate of e(0, 1) is nonzero, take v_{0} = 0. Otherwise write e^{−1}(0, 1) = (v_{1}, d^{′}). Equivariance shows that v_{1} is a χ-eigenvector; moreover e(v_{1}, 1) = e(v_{1}, d^{′}) + (1 −d^{′})e(0, 1) has second coordinate 1. Take v_{0} = v_{1} in this case.
Write e(v_{0}, 1) = (w_{0}, δ). The vector w_{0} is also a χ-eigenvector. If e(v, 0) = (a(v), b(v)), define
Selectable PDF notation • transcription unchecked
f(v) = a(v) −^{b(}_{δ }^{v)}w_{0}.If e^{−1}(w, 0) = (a^{′}(w), b^{′}(w)), define g(w) = a^{′}(w) −b^{′}(w)v_{0}. Both maps are linear and equivariant; for instance a intertwines the representations and b transforms by χ, exactly as w_{0} does. The identities
Selectable PDF notation • transcription unchecked
e^{−1}(f(v), 0) = (v, 0) −^{b(}_{δ}^{v)}(v_{0}, 1),
e(g(w), 0) = (w, 0) −b^{′}(w)(w_{0}, δ)give g(f(v)) = v and f(g(w)) = w by substituting the two coordinate definitions. Thus f is the required equivariant isomorphism. Applying this step successively to the t ordered character summands proves the assertion, including t = 0.
In the source the reference Galois family and the selected family have the same Hecke characteristic poly-nomials at the relevant Frobenius elements. Their asserted rational conjugacy must use continuity, semisim-plicity and density of those Frobenius classes. Once that conjugacy and the ordered decompositions exist,
Lemma 6.13 identifies their main summands, and Lemma 6.12 explains why conjugating the integral span data does not require an integral conjugating matrix. The reference family’s existence, including its adapted wild-inertia element and idempotent data, is a separate existence requirement; a conjugacy theorem alone does not supply it.
Background fact 6.14 (The finite local conditions are dual). Let F/ℚ_{ℓ} be finite and let L be a finite free ℤ_{ℓ}-module with continuous G_{F} -action such that L⊗ℚ_{ℓ} is de Rham. Under the local Tate pairing, H_{f}^{1}(F, L) and H_{f}^{1}(F, L^{∗}(1) ⊗ℚ_{ℓ}/ℤ_{ℓ}) are exact annihilators. The rational finite conditions are exact annihilators as well. Here the lattice condition is the inverse image of the rational condition, and the divisible condition is its propagated condition on the lattice quotient.
This is Proposition 3.8 of S. Bloch and K. Kato, L-functions and Tamagawa numbers of motives, The Grothendieck Festschrift I, Progress in Mathematics 86 (1990), 333–400, at page 354; the finite condition is defined in (3.7.2)–(3.7.3).^{4} For the application the fields are completions E_{w}, the lattices are finite free, and continuity is part of the compatible setup. Crystalline local representations satisfy the de Rham hypothesis. Application to the abstract local conditions still requires their identification with these propagated finite conditions.
Background fact 6.15 (Integral prismatic realization). For a complete discretely valued mixed-characteristic field with perfect residue field, the category of prismatic F-crystals on its ring of integers is equivalent, by etale realization, to finite free ℤ_{ℓ}-lattices in crystalline Galois representations. Every equivariant map of such lattices comes from a unique map of their prismatic F-crystals.
This is Theorem 1.2, restated as Theorem 5.6, of B. Bhatt and P. Scholze, Prismatic F-crystals and crystalline Galois representations, Cambridge Journal of Mathematics 11 (2023), no. 2, 507–562.^{5} The field hypothesis holds for each E_{w}/ℚ_{ℓ}. The statement concerns integral lattices whose rational representations are crystalline; it does not assert full faithfulness for arbitrary torsion representations. Identifying the abstract period and crystal models in the setup with this realization remains necessary before using this citation for them.
Lemma 6.16 (The integral cyclotomic prism lattice). Let K_{0}/ℚ_{ℓ} be finite unramified, with residue field k, let s ≥1, and put K = K_{0}(ζ_{ℓ}s). Let T be a finite free ℤ_{ℓ}-lattice with continuous G_{K}-action and crystalline rational representation. Put
Selectable PDF notation • transcription unchecked
S = W(k)[[u]], Q_{s}(u) = Φ_{ℓ}s(1 + u), φ(u) = (1 + u)^{ℓ}−1,^{4}Primary text: https://math.stanford.edu/~conrad/BSDseminar/refs/BKTamagawa.pdf. ^{5}Primary text: https://people.mpim-bonn.mpg.de/scholze/PrismaticCrystals.pdf.
with Witt Frobenius on W(k). There is a finite free S-module M(T), functorial in T, with the following properties. It embeds in W(C^{♭}) ⊗_{ℤℓ}T, where C is a completed algebraic closure of K. Its basis becomes a basis of that tensor product. Frobenius induces an isomorphism
Selectable PDF notation • transcription unchecked
(φ^{∗}M(T))[1/Q_{s}] ≃ M(T)[1/Q_{s}].The group G_{K} preserves M(T), acts semilinearly by g(u) = (1 + u)^{χℓ(g)} −1, and acts trivially on M(T)/uM(T). The subgroup ker(χ_{ℓ}|_{GK}) fixes M(T) pointwise in its ambient tensor product.
If T is free of rank a over the integers O of a finite coefficient extension containing K_{0}, then M(T) has the compatible O-action. It is free of rank a on each factor of
Selectable PDF notation • transcription unchecked
O ⊗_{ℤℓ}W(k)[[u]] ≃ ∏ O[[u]].
τ:k↪→O/𝔪_{O}Proof. The pair (S, (Q_{s})) is a bounded prism over O_{K}. Here are the ring checks in this particular construc-tion. The polynomial Q_{s} is monic and Q_{s}(u) mod ℓ = u^{ℓs−1(ℓ−1)}, while Q_{s}(0) = ℓ. It is Eisenstein, so evaluation at u = ζ_{ℓ}s −1 identifies S/(Q_{s}) with O_{K}. The topology defined by (ℓ, Q_{s}) is the (ℓ, u)-adic topology. The element Q_{s} is a non-zero-divisor, and S/(Q_{s}) has no ℓ-torsion. Writing δ(a) = (φ(a)−a^{ℓ})/ℓ, we have
δ(Q_{s}) mod (ℓ, u) = 1.
Indeed its constant coefficient is 1 −ℓ^{ℓ−1}. Thus δ(Q_{s}) is a unit, and ℓδ(Q_{s}) = φ(Q_{s}) −Q^{ℓ}_{s} gives the prism condition ℓ∈(Q_{s}, φ(Q_{s})).
Choose compatible roots of unity and write ϵ = (1, ζ_{ℓ}, ζ_{ℓ}2, . . .) in the tilt. The map
Selectable PDF notation • transcription unchecked
S −→A_{inf} = W(O_{C}^{♭}), u −−→[ϵ^{1/ℓs}] −1commutes with Frobenius and sends Q_{s} to
Selectable PDF notation • transcription unchecked
[ϵ] −1_{,}
[ϵ^{1/ℓ}] −1a generator of ker θ. Its reduction modulo ℓ is injective: the image of u has strictly positive valuation, and in a nonzero power series over k the lowest nonzero term has strictly smaller valuation than all later terms. Injectivity over ℤ_{ℓ} follows by successive division by ℓ and separatedness. This is a map of prisms over O_{K}.
Use Fact 6.15 to obtain the prismatic F-crystal ℰ_{T} realizing T, and set M(T) = ℰ_{T} (S, (Q_{s})). Evaluation is finite projective, hence free over the local ring S. The crystal property gives its base change to A_{inf}; the etale comparison identifies the further base change with W(C^{♭}) ⊗T. Precisely, this is Construction 4.8 and the comparison in the construction following Theorem 5.2 of Bhatt–Scholze, printed pp. 13 and 15–16. It is compatible with Frobenius, Galois action and maps of lattices. The basis is therefore independent in the ambient tensor product, giving the stated embedding. The Frobenius assertion is the defining F-crystal isomorphism evaluated at this prism. In particular, both inclusions between M(T) and the span of φM(T) hold after multiplication by a sufficiently large power of Q_{s}: take a common denominator for the finitely many matrix entries of the isomorphism and its inverse.
For g ∈G_{K} the indicated substitution is an automorphism of S, commutes with Frobenius and induces the identity on S/(Q_{s}) = O_{K}, because χ_{ℓ}(g) ≡1 (mod ℓ^{s}). Crystal transport gives the semilinear action on M(T). The map u →→0 is a map from this prism to (W(k), (ℓ)), considered over O_{K} by the residue map. Its composite with every such g is unchanged. Base change of the crystal therefore identifies the action on M(T)/uM(T) with the identity. If χ_{ℓ}(g) = 1, the automorphism of S itself is the identity. Naturality of the comparison then says that g fixes every image vector of M(T). This also shows directly that the construction lands in the indicated cyclotomic-invariant period module.
For the coefficient assertion, full faithfulness lifts every endomorphism given by an element of O uniquely to the crystal. Uniqueness preserves addition, multiplication and the unit, so these lifts define an O-action, commuting with all the preceding structures. Put A = O ⊗_{ℤℓ}S and decompose it by the displayed unramified embeddings. Each component M_{i} is a finite O[[u]]-module on which u is injective, since M(T) is S-free. Its quotient M_{i}/uM_{i} is O-torsion-free: it has no ℓ-torsion, and a power of a uniformizer of O equals ℓ times a unit. Being finite over the DVR O, it is free. Lift an O-basis to obtain a surjection from a finite free O[[u]]-module onto M_{i}, by Nakayama’s lemma. If J is its kernel, injectivity of u on M_{i} shows that reduction modulo u is still exact at J. Thus J/uJ = 0; a second application of Nakayama gives J = 0. Finally the comparison, with its O-action, identifies each component after scalar extension with the corresponding component of W(C^{♭}) ⊗_{ℤℓ}T. The latter has rank a. Each embedding component survives in this scalar extension, so every M_{i} has rank a as asserted.
This construction supplies the free integral cyclotomic lattice, its commuting actions and Frobenius com-mensurability, rather than just the existence of a rational crystalline representation. The height bounds in
Lemma 6.40 still require its stated filtered expansion identification. The integral Herr comparison and its finite-class assertion are also separate requirements; neither follows from freeness alone.