Section 6.2 · PDF pages 254–275
The reference-family construction
The notation transcription is provisional. Check it against the printed mathematics when citing or reusing a formula.
6.2 The reference-family construction
A local index is called live when some selected eigencharacter has a compatible Hecke–Galois representation whose specified rank-r complement is residually absolutely irreducible. This is an existence condition on a local index; it does not specify one reference representation in advance. The assertion needed in the comparison is that every live index has a good reference family with adapted inertia and corner data, and with all required corner roots in its coefficient ring. We now describe the finite enlargement which the source uses to arrange the last condition simultaneously.
Lemma 6.17 (Residual absolute irreducibility and matrix span). If a representation of a group on a nonzero finite-dimensional k-space V is absolutely irreducible, the k-span of its group matrices is End_{k}(V ).
Proof. First let k be algebraically closed and let A be that span, which is a unital algebra. An endomorphism commuting with A is scalar: it has an eigenvalue λ, and its nonzero λ-eigenspace is A-stable, hence all of V . We show by induction that for linearly independent v_{1}, . . . , v_{s}, the set U = {(av_{1}, . . . , av_{s}) : a ∈A} is all of V^{s}. For s = 1 it is a nonzero invariant subspace. For the induction step, projection onto the first s −1 coordinates is surjective by induction. Its kernel in U is either zero or the entire last copy of V , by irreducibility. In the latter case U = V^{s}. In the former case U is the graph of an A-linear map V^{s−1}→V . Each of its coordinate maps commutes with A and is therefore scalar. The identity element of A would then give v_{s} =∑_{i<s}λ_{i}v_{i}, a contradiction. Taking the v_{i} to be a basis shows that A realizes every possible list of images of a basis, so A = End_{k}(V ).
For arbitrary k, extend scalars to an algebraic closure. Absolute irreducibility allows the preceding argument there. The extended matrix span is the extension of the original span, so its dimension is both dim_{k} A and (dim_{k} V )^{2}. Since A ⊂ End_{k}(V ) already has this full dimension, equality follows over k.
Lemma 6.18 (Transfer of the full matrix span by traces). Let σ, ρ : G → GL_{r}(k) have equal traces on every group element. If the matrices of ρ span M_{r}(k), the matrices of σ span it too. This holds in positive characteristic as well.
Proof. For a finite relation ∑_{g}a_{g}σ(g) = 0 and every h ∈G, the trace of ∑_{g}a_{g}ρ(g)ρ(h) is
Selectable PDF notation • transcription unchecked
∑a_{g} tr ρ(gh) =∑a_{g} tr σ(gh) = 0.
g gAs the ρ(h) span all matrices, the trace of (∑_{g}a_{g}ρ(g))B is zero for every B ∈M_{r}(k). Take B to be each matrix unit in turn. The trace pairing extracts every entry, giving∑ ∑_{g}a_{g}ρ(g) = 0. Consequently sending
_{g}a_{g}σ(g) to ∑_{g}a_{g}ρ(g) defines a well-defined surjective linear map from the span of σ(G) to M_{r}(k). The former is already a subspace of the r^{2}-dimensional space M_{r}(k), so it must have dimension r^{2} and equal that space.
Lemma 6.19 (Equality from the common Frobenius polynomials). Two continuous integral Galois represen-tations with the same characteristic polynomials on unramified Frobenius elements at primes of the CM field E split from their complex conjugates, outside a fixed finite set, have the same characteristic polynomials at every element of G_{E}.
Proof. We check the extra split-prime restriction in the density argument of Lemma 8.43. Choose a ∈𝒪_{E} with a ̸= a, and exclude the finitely many prime divisors of a −a. Such an a exists because complex conjugation on a CM field is nontrivial. Prime ideals of residue degree at least two over ℚ have Dirichlet density zero: for real s > 1 their norm series is bounded by [E : ℚ]∑_{p}p^{−2s}, which stays bounded as s decreases to one. Chebotarev gives each conjugacy class in each finite Galois quotient a positive Dirichlet density. Consequently the class still occurs at a residue-degree-one prime outside the stated finite exclusions. This uses precisely the positive-density form of the Chebotarev theorem cited in that lemma.
At such a prime w the residue field is 𝔽_{p}, so an integer n has the same residue as a. Thus a −n ∈w, whereas a −n /∈w: otherwise a −a ∈w. This is the split-place witness required by the source. Therefore every finite quotient conjugacy class is met by the allowed split Frobenius elements. Their conjugacy closure is dense, by the same open-normal-subgroup argument as in Lemma 8.43. Each characteristic coefficient is a continuous class function, since it is a polynomial in the matrix entries. Equality on that dense set therefore implies equality everywhere in the Hausdorff coefficient field.
The local Hecke algebra has a single residue evaluation. To verify this assertion, let T ⊂O^{X} be local. Each evaluation T → O → k is surjective, since T contains the constants, and its kernel is therefore the unique maximal ideal of T. For f ∈T, the element f −f(x) is a nonunit because its value at x is zero. Its image under evaluation at any other y is therefore zero in k. Thus f(x) and f(y) have equal residues.
Apply this to the characteristic-polynomial coefficient functions of a good family. After reduction, the characteristic polynomial of r_{0}(g, x) is independent of x. At each selected x its ordered decomposition is ρ_{x} ⊕⊕_{c}χ_{c}, with the same character summands for every x. The characteristic polynomials therefore factor as det(z −ρ_{x}(g))∏_{c}(z −χ_{c}(g)). The common character product is monic and nonzero, so it can be cancelled in k[z]. The residual main representations have equal characteristic polynomials, and hence equal traces, at every g. At a live point one main representation is absolutely irreducible. It can be compared with the chosen good family at that same point by Lemma 6.19: their good Hecke–Galois condition specifies the same Frobenius polynomials. Cancellation of the common monic character product then identifies the residual main characteristic polynomials as well. Lemmas 6.17 and 6.18 show that the residual main matrix span is full at every selected x. Full matrix span remains full after scalar extension and forces every invariant subspace to be zero or the whole space. It therefore also proves absolute irreducibility for all these residual main representations. This is the residual-irreducibility part of the reference-setting construction, and it justifies the residue-span hypothesis used in Lemma 6.10.
Proposition 6.20 (Liveness descends to the reference coefficients). Let O_{0}, O_{1} be the coefficient DVRs at corresponding local indices, with residue fields k_{0}, k_{1}. Suppose the correspondence consists of an injective coefficient map ϕ : O_{0} → O_{1}, a continuous group isomorphism γ : G_{1} ≃ G_{0} with continuous inverse, and a bijection e : X_{1} ≃ X_{0}. Assume that ϕ respects the ℤ_{ℓ} structures, the complex embeddings, the Hecke eigenvalues, the specified characters, the selected sets, and the split Frobenius conditions. If a good reference family exists at index zero, every selected member of that family admits the prescribed main-plus-characters decomposition, and index one is live, then index zero is live.
Proof. The inverse image under ϕ of the maximal ideal of O_{1} is a prime ideal containing the nonzero element ℓ. A DVR has only one nonzero prime ideal, its maximal ideal. Thus ϕ is local and induces an injection ϕ : k_{0} ↪→k_{1}. It is continuous for the maximal-ideal-adic topologies: the image of the nth power of the first maximal ideal is contained in the nth power of the second. These are the specified coefficient topologies.
Choose a live point π ∈ Z_{1}, its good representation r_{1}, and its residual absolutely irreducible main summand ρ_{1}. At eπ ∈ Z_{0}, choose the good reference representation r_{0} and its prescribed integral main representation ρ_{0}. The representations ϕ(r_{0} ∘γ) and r_{1} are continuous. Their characteristic polynomials agree at the allowed split Frobenius elements by the Hecke-eigenvalue compatibility. Lemma 6.19, applied coefficientwise, therefore gives equality of their traces at every group element. The two decompositions express each trace as the main trace plus the sum of the specified character values. The character sums agree under ϕ, so cancellation gives
Selectable PDF notation • transcription unchecked
ϕ(tr ρ_{0}(γg))= tr ρ_{1}(g) (g ∈G_{1}).This equality survives reduction to k_{1}.
Lemma 6.17 makes the k_{1}-span of ρ_{1}(G_{1}) the full matrix algebra. Lemma 6.18 gives the same full span for ϕ(ρ_{0}(γG_{1})). Since γ is surjective, this is the scalar extension of the k_{0}-span of ρ_{0}(G_{0}). That k_{0}-span must itself be full. Indeed, a proper subspace of the finite-dimensional matrix space has a nonzero annihilating linear functional; extending its coefficients to k_{1} leaves it nonzero and annihilates the extended span, a contradiction. Full matrix span remains full over an algebraic closure, so ρ_{0} is absolutely irreducible. Together with the chosen good representation and decomposition at eπ, this is exactly a live point at index zero.
Lemma 6.21 (A wild cyclotomic choice). For the cyclotomic hosts under consideration and each local index, there is g_{0} ∈G_{E} in inertia at a place over ℓ such that
Selectable PDF notation • transcription unchecked
ε(g_{0}) = 1 + ℓ^{s0}v, v ∈ℤ^{×}_{ℓ}.The choice can be made with pro-ℓ cyclic closure. Consequently, at every place over ℓ whose inertia contains g_{0}, the element g_{0} is in wild inertia. This in particular gives the wildness required at the two auxiliary level primes.
Proof. We first verify the local field on which the choice must act trivially. The host field is generated inside the cyclotomic field of conductor N = Q_{S}qd. As q, d are distinct primes coprime to Q_{S}, the exponent v_{ℓ}(N) is at most s_{0} = max(1, v_{ℓ}(Q_{S})). If ℓ = q or d its exponent is one; otherwise it is v_{ℓ}(Q_{S}). At ℓ = 2 the specified s_{0} = v_{2}(Q_{S}) is at least two and the same bound holds. Write N = ℓ^{a}b with (b, ℓ) = 1. A root of unity of order b belongs to the maximal unramified local extension: choose f with b | ℓ^{f}−1 and lift the distinct roots in 𝔽_{ℓ}f by Hensel’s lemma. A root of unity of order N is a product of a bth root and an ℓ^{a}th root, by Bezout for b, ℓ^{a}. Thus the image of E at every ℓ-adic place lies in
Selectable PDF notation • transcription unchecked
ℚ^{nr}_{ℓ}(µ_{ℓ}s0).The cyclotomic character maps local inertia over ℚ_{ℓ} onto ℤ^{×}_{ℓ}. To recall the needed local argument, Φ_{ℓ}n(1 + z) is Eisenstein: modulo ℓ it is z^{ℓn−1(ℓ−1)} and its constant term is ℓ. Hence the ℓ^{n}-cyclotomic extension has degree ℓ^{n−1}(ℓ−1), is totally ramified, and has the full group (ℤ/ℓ^{n})^{×} acting on the roots. It is disjoint from every unramified extension, since an extension both unramified and totally ramified has degree one. The compatible finite cyclotomic actions can therefore be extended while fixing ℚ^{nr}_{ℓ} ; compactness of the inverse system of finite Galois extension maps gives an inertia element for every unit in ℤ^{×}_{ℓ}. Choose one, σ, with value 1 + ℓ^{s0}. It fixes µ_{ℓ}s0 and hence the displayed image of E. The usual decomposition-group identification at a chosen place over ℓ makes σ an element of its global inertia. The local and global cyclotomic characters agree, because both specify the action on every ℓ^{n}th root of unity.
We modify σ to have pro-ℓ closure without losing the exact first cyclotomic term. For k ≥0, choose by the Chinese remainder theorem an integer e_{k} with
e_{k} ≡1 (mod ℓ^{k+1}), e_{k} ≡0 (mod (k!)_{ℓ}′),
where (k!)_{ℓ}′ is the prime-to-ℓ part of k!. Take a cluster point g_{0} of σ^{ek} in the compact absolute Galois group. The chosen inertia is closed, so it contains g_{0}. In any finite quotient, if the order of σ divides d, then for k ≥d these powers stabilize at the component of σ of ℓ-power order: the exponent is one modulo the ℓ-part of d and zero modulo its prime-to-ℓ part. Thus the image of g_{0} in every finite quotient has ℓ-power order. Moreover e_{k} ≡1 (mod ℓ) and s_{0} ≥1 imply
(1 + ℓ^{s0})^{ek}≡1 + ℓ^{s0} (mod ℓ^{s0+1}).
Continuity preserves this closed congruence at g_{0}. It gives ε(g_{0}) = 1+ℓ^{s0}(1+ℓc), with a unit in parentheses.
Finally suppose g_{0} lies in inertia at any place above ℓ. For a nonzero algebraic element y, its finite normal extension is fixed by some g_{0}^{ℓa}, by the finite-quotient property. Put x = g_{0}(y)/y, which has valuation one. Inertia acts trivially on its residue, and
Selectable PDF notation • transcription unchecked
ℓ∏^{a}−1
g_{0}^{j}(x) = 1.
j=0Its residue therefore satisfies x^{ℓa} = 1. In residue characteristic ℓ this says (x −1)^{ℓa} = 0, so x = 1. Thus g_{0}(y)/y−1 has strictly positive valuation for every y ̸= 0, the defining wild-inertia condition used here.
Lemma 6.22 (Adapted integral idempotents). Let O be a finite DVR algebra over ℤ_{ℓ}, let X be finite, and let 𝒜 be the T-span of the matrices of a family r_{x}. At a selected point π suppose r_{π} ≃ρ⊕⊕^{t}_{c=1}χ_{c} over the fraction field, where r ≥2 and the integral span of ρ(G) is M_{r}(O). Then 𝒜 contains orthogonal idempotents e_{B}, e_{1} which act at π as the projection onto the character summands and as a rank-one projection in the main summand, respectively.
Proof. First we can realize any pair (P, 0) as an O-linear combination of the pairs
Selectable PDF notation • transcription unchecked
(ρ(g), (χ_{c}(g))_{c}).Let J be the span of these pairs inside M_{r}(O) × O^{ℕ}. It is closed under multiplication, since the product of two generating pairs is the pair for the product of their group elements. Its first projection is surjective by the assumed matrix span. Choose distinct indices i_{0}, i_{1}. Lift the matrix units E_{i0i0} and E_{i0i1} to J. The commutator of their lifts is (E_{i0i1}, 0), because the second components commute. If (M, 0) ∈ J, lifting arbitrary matrices U, V to J shows (UMV, 0) ∈ J. Sandwiching (E_{i0i1}, 0) between matrix units thus gives every (aE_{ij}, 0). Summing matrix entries proves the assertion for every P. The corresponding linear combination of the family matrices belongs to 𝒜 and acts at π as (P, 0) in the given decomposition. Constants in O belong to T, so these are permitted order elements.
We next record the integral idempotent construction used twice. Put R = ∏_{x∈X}M_{V} (O) with its ℓ-adic module topology. It is compact and Hausdorff. Every ℤ_{ℓ}-submodule is closed: it is finitely generated, hence is the continuous image of a finite power of the compact ring ℤ_{ℓ}. For b ∈ R, the closure C of its positive powers is a nonempty compact commutative subsemigroup. It contains an idempotent. Here is the compactness argument. The nonempty closed subsemigroups of C have a minimal member K: intersections of chains are nonempty by compactness and the finite intersection property, so the usual minimal-element principle applies. For a ∈K, the set aK is a nonempty closed subsemigroup of K, using commutativity; hence aK = K. The set {x ∈K : ax = a} is then a nonempty closed subsemigroup and must also equal K. In particular a^{2}= a. Choose an idempotent E ∈C.
If every positive power of b lies in a closed additive coset c + M, then so does E. Apply this first with M = 𝒜 and c = 0 to see E ∈𝒜. If b(π) is already idempotent, every positive power has value b(π), so E(π) = b(π). If yb = 0 then every positive power lies in the kernel of multiplication by y on the left, a closed submodule, and yE = 0. Similarly by = 0 implies Ey = 0. Thus this construction preserves both specified values at π and left and right annihilators.
Choose a ∈𝒜 acting at π as (I_{r}, 0) by the first paragraph. Apply the idempotent construction to 1 −a; its value at π is the character projection. This gives e_{B}. Next choose a^{′} ∈𝒜 acting at π as (E_{00}, 0), and apply the construction to (1 −e_{B})a^{′}(1 −e_{B}). Its value at π is the desired rank-one projection. Both its products with e_{B} are zero; preservation of annihilators gives e_{B}e_{1} = e_{1}e_{B} = 0. All assertions follow.
Lemma 6.23 (Distinct corner eigenvalues at every point). In the reference setting let e_{B}, e_{1} be the idem-potents of Lemma 6.22. Suppose the family is a sum of pairwise nonisomorphic absolutely irreducible con-stituents at every point, T is local and separates X, and its functions contain the characteristic coefficients of the family matrices. If χ_{1}(g_{0}), . . . , χ_{t}(g_{0}) are distinct, then, for every L ≥0, one can choose a correction s in the corner congruence module at π such that
Selectable PDF notation • transcription unchecked
X_{0} = e_{B}r(g_{0})e_{B} + ϖ^{L}shas t distinct eigenvalues on the image of e_{B}(x) for every x ∈X. Its corner polynomial is monic of degree
Selectable PDF notation • transcription unchecked
t.
Proof. The ranks of each of the two idempotents are constant in x. Indeed the characteristic-coefficient functions of every order element belong to T, by the determinant-functions argument above. All residue evaluations of local T agree. The characteristic polynomial of an idempotent of rank d is z^{V −d}(z −1)^{d}, and the distinct multiplicities at zero and one determine d even in positive characteristic. The integral im-age and kernel of an idempotent are complementary free modules over the DVR. Consequently its rank is unchanged by reduction or extension to the fraction field. Equality of residual characteristic polynomials proves constancy. The two ranks at π are t and one, so they have these values everywhere.
We verify that a separable direction exists in each corner. For a representation which is a direct sum of pairwise nonisomorphic absolutely irreducible K-modules V_{j}, its matrix span contains every endomor-phism preserving the summands. The simultaneous-orbit proof of Lemma 6.17 applies to a basis of each V_{j}, in succession. In that proof an invariant graph would give a module homomorphism from the previously considered copies of the V_{j} to the next copy. A homomorphism between distinct simple summands is zero; an endomorphism of an absolutely irreducible summand is scalar by the eigenvalue proof of Schur’s lemma used there. Thus the graph would express the next basis vector as a linear combination of its predecessors in that same summand, a contradiction. This proves that all simultaneous basis images can be prescribed, which is the stated matrix-span assertion.
At x, the idempotent e_{B}(x) belongs to this span and preserves each V_{j}. Its image is the direct sum of its images on those summands. Choose a basis of this image in each summand and assign pairwise distinct eigenvalues in the infinite field K to all these basis vectors. Extend this to an operator preserving each V_{j}. The preceding span assertion realizes that operator by a rational linear combination of family matrices. Its corner therefore has separable characteristic polynomial. Clearing denominators gives an integral order element with the same property up to a nonzero scalar, which preserves distinctness.
For x ̸= π, separation supplies f ∈T with f(x) −f(π) ̸= 0. Subtract the constant f(π), and multiply the integral order element just obtained by this function and by e_{B} on both sides. This is an allowed corner correction: its scalar coefficient vanishes at π. Its corner at x still has separable characteristic polynomial, up to a nonzero scalar. Multiplication by the nonzero ϖ^{L} does not change this property. At π no correction is needed, since the corner already has the distinct eigenvalues χ_{c}(g_{0}).
It remains to choose one correction for the finite set X. We give the finite avoidance argument to ensure the choice is simultaneous. For matrices B, A over K, the discriminant of det(z −B −uA) is a polynomial in u. If B has distinct eigenvalues, this polynomial is nonzero at u = 0 and has only finitely many bad values. If a nonzero scalar multiple of A has distinct eigenvalues, its leading homogeneous discriminant coefficient is nonzero, so again there are only finitely many bad values. Equivalently one can apply the first assertion to A + u^{−1}B and exclude u = 0.
Induct on the already treated subset of X, starting with the zero correction. To add x, if its original corner is separable use the current correction itself as direction; otherwise use the separable direction at x constructed above. In the first case the affine line has a separable member at parameter −1; in the second case the direction is separable up to scalar. At each previously treated point the current corner is separable, so it excludes only finitely many parameters as well. Choose a parameter in the infinite ring O outside this finite union and add that multiple of the direction. The correction module is an O-module, so the result stays in it. Induction finishes the construction at every point.
The corner relations e_{B}X_{0} = X_{0} = X_{0}e_{B} imply, in a basis of im e_{B} ⊕ker e_{B}, that ze_{B} −X_{0} +1−e_{B} is block diagonal with blocks z −X_{0} and 1. Its determinant is thus the characteristic polynomial on the rank-t image: it is monic of degree t and has the distinct roots just arranged. At π the correction vanishes, so all the prescribed selected-point actions are retained.
Lemma 6.24 (Only finitely many additional roots are required). Fix a live local index with finite eigenchar-acter set X and selected subset Z. Suppose a good reference family and its adapted data g_{0}, e_{B}, e_{1}, X_{0} can be chosen for each π ∈Z. For each pair (π, x) ∈Z × X, suppose the corner polynomial
det(ze_{B}(x) −X_{0}(x) + 1 −e_{B}(x))
is monic of degree t. There is a finite set S ⊂ℂ containing every root of every such polynomial under the fixed complex coefficient embedding. If a compatible coefficient extension contains S in its fraction field, all these corner polynomials split over its integer ring.
Proof. Choose the adapted data once for each member of the finite set Z. Let S be the union, over Z × X, of the root sets of the resulting complex polynomials. Each root set is finite, and the union is finite. The fundamental theorem of algebra, monicity and the degree condition give an ordered factorization of each polynomial into t factors z −α_{i}, with roots listed with their multiplicities. The assumption on the new coefficient field gives a preimage there for each α_{i}. Map the proposed factorization to ℂ. Compatibility of the coefficient embedding makes it the known complex factorization, and injectivity proves the factorization over the new fraction field. Each root is integral over the new DVR, since its polynomial is monic with coefficients in that DVR. A DVR is integrally closed, so each root belongs to the integer ring. The factorization therefore holds there. Distinctness, when required by the adapted data, is preserved by the same injection.
At a live index these adapted choices are now supplied by the preceding lemmas. The residual full-span argument and Nakayama give the integral full main matrix span. Use Lemma 6.21 for g_{0}. Since χ_{c} = χ_{1}ε^{1−c}, its character values are distinct: the lifting-the-exponent calculation in Lemma 6.8 shows that no nonzero power ε(g_{0})^{j} for 1 ≤ j < t equals one. Lemma 6.22 gives e_{B}, e_{1}, and Lemma 6.23 gives X_{0}, with any chosen L, for example L = 0. Thus no existence of an adapted idempotent is being inferred merely from the finite root union.
Proposition 6.25 (Transport of a chosen reference family). Use the coefficient, group and eigencharacter correspondence of Proposition 6.20. Suppose the Hecke algebras are generated by the matching permitted Hecke eigenvalue functions. At index zero suppose there is a reference family r_{0}, ρ_{0} with the reference setting and, at each π_{0} ∈Z_{0}, chosen wild-cyclotomic data g_{0}, L, e_{B}, e_{1}, X_{0}. Suppose all the complex roots of its corner polynomials belong to a set S contained in the new coefficient fraction field. Then index one has a
good reference family with those adapted data and split corner polynomials. The integral ring over which the reference objects are constructed can remain O_{0}.
Proof. Write T̃_{0} = {f ∘e : f ∈ T_{0}} ⊂ O_{0}^{X1}. Pullback along e is an algebra isomorphism, with inverse pullback along e^{−1}. Define the reference representations over O_{0} by
Selectable PDF notation • transcription unchecked
r̃_{x}(g) = r_{0,ex}(γg), ρ̃_{x}(g) = ρ_{0,ex}(γg), χ̃_{c}(g) = χ_{0,c}(γg).The actual representations over O_{1} are their coefficient maps under ϕ. The ring O_{0} is still a finite DVR alge-bra over ℤ_{ℓ}; compatibility with ϕ is exactly the scalar tower required for using it as the reference coefficient ring.
Every reference-setting assertion over O_{0} is preserved under these pullbacks. Locality and separation follow from the algebra isomorphism and the bijection e. Matrix characteristic coefficients and their member-ship in T̃_{0} follow by substitution. Invariant subspaces, complements and intertwiners are unchanged by the surjective group isomorphism γ. This preserves semisimplicity, the decomposition into distinct absolutely irreducible constituents, and the residual absolute irreducibility of the selected main blocks. The prescribed decompositions use the same linear maps, evaluated at γg. The residual characteristic polynomials are like-wise unchanged.
The scalar extension of T̃_{0} spans precisely T_{1}. Indeed their permitted Hecke generators correspond under ϕ. The O_{1}-span of the image of T̃_{0} is an algebra: products of linear combinations are linear combinations of products in T̃_{0}. It contains the generators of T_{1}, so it contains T_{1}; the reverse inclusion follows since the image of every polynomial in the old generators is the same polynomial in the new generators. This proves the enlargement condition for the Hecke algebra, not merely an equality of rational eigenvalues.
The new family is good. Its continuity follows from continuity of ϕ and γ. The original reference representations are sums of absolutely irreducible constituents, so extension to the new characteristic-zero fraction field preserves their semisimplicity. Their Frobenius polynomial identities follow by mapping the original identities and using the stipulated Frobenius and Hecke compatibility.
For π ∈ Z_{1}, take the old chosen data at eπ, replace g_{0} by γ^{−1}g_{0}, and pull back each matrix-valued function along e. The two idempotent equations, their orthogonality, ranks and actions in the selected de-composition are unchanged. For clarity, an expression in the old matrix order is a sum ∑_{j}f_{j}(x)r_{0,x}(h_{j}); after pullback it is∑ ∑_{j}f_{j}(ex)r̃_{x}(γ^{−1}h_{j}), an expression in the pulled-back order. A correction term is a sum
_{j}a_{j}(x)e_{B}(x)b_{j}(x)e_{B}(x) with a_{j}(π_{0}) = 0. Its pullback has coefficients vanishing at π, and hence belongs to the new congruence correction module. The equation
Selectable PDF notation • transcription unchecked
X_{0} = e_{B}r(g_{0})e_{B} + ϖ^{L}stherefore persists with the same reference uniformizer ϖ. Distinctness of the corner roots also persists, because these polynomials over O_{0} have only been relabelled. The stipulated inertia, wild-inertia and cyclo-tomic compatibility makes γ^{−1}g_{0} a wild-cyclotomic choice at the new index.
Finally choose in Frac(O_{1}) the preimages of the complex roots in S. The compatibility of the complex embeddings identifies the mapped corner polynomial with the product of their linear factors. Injectivity proves this identity over Frac(O_{1}); monicity puts every root in O_{1} as in Lemma 6.24. These are all the required root and reference-family properties.
Lemma 6.26 (Enlarging the coefficient fields simultaneously). Fix a reference model M_{0}. Its local index sets over each coefficient place are finite. One can impose a finite additional coefficient requirement at each place so that every subsequent built model M_{1} admits compatible integral embeddings O_{0,i0} ↪→O_{1,i1} whenever i_{0}, i_{1} are over that place. Here the embeddings of their fraction fields in ℚ_{ℓ} use the same prescribed complex identification.
Proof. Each fraction field K_{0,i0} is finite over ℚ_{ℓ}. Choose a finite ℚ_{ℓ}-spanning set in its image in ℚ_{ℓ}. At a fixed coefficient place w, take the union of the complex images of these generators over all levels 1 ≤j ≤m and all reference indices over (j, w). This is finite, and so is an admissible additional requirement. Every new coefficient field over w contains these generators in its image, by the definition of a built model. That image is a ℚ_{ℓ}-subfield, so it contains the whole old field. Inverting the new field embedding therefore defines an embedding K_{0,i0} ↪→K_{1,i1} with the required complex compatibility.
This embedding restricts to the integer rings. Both are the rings of integers of finite extensions of ℚ_{ℓ}. Equivalently they consist of elements of absolute value at most one in their specified embeddings in ℚ_{ℓ}. The inclusion of the images preserves this condition. It also fixes ℚ_{ℓ}, and hence ℤ_{ℓ}. Restriction gives the claimed integral embedding and scalar compatibility. No finite union over all coefficient places has been taken.
Lemma 6.27 (Matching the eigencharacter blocks). Fix a level j and coefficient place w. Suppose two built models enumerate the same complex Hecke eigenvalue systems, each once, and their local blocks are exactly their congruence classes modulo the valuation at w. Then each block of the second model has a unique corresponding block of the first, and equality of complex eigenvalue systems gives a bijection between their labels. It preserves every permitted Hecke eigenvalue.
Proof. Choose one label x_{1} in the given block of model one. Its system occurs at a unique label y_{0} in model zero; let i_{0} be that label’s block. For any other x in the block of x_{1}, the model-zero label y with the same system is congruent to y_{0}, since the systems of x and x_{1} are congruent. Thus y lies in i_{0}. This defines a map of the desired blocks, injective because distinct labels have distinct systems. If y is any label in i_{0}, choose the model-one label x with its system. It is congruent to x_{1}, and hence belongs to the original model-one block. Thus the map is also surjective. By construction it identifies all the eigenvalues. The chosen anchor y_{0} also proves uniqueness of the corresponding block.
In the source construction the two premises of this last lemma are concrete: the localized points enumer-ate the nonzero common eigenspaces of the one fixed complex automorphic space, and the maximal-ideal index is determined by congruence of its permitted Hecke generators. Completeness of that enumeration is required; equality of a few eigenvalues by itself would not suffice. For the selected subsets one also uses the same complex automorphic realization. Its nonzero isotypic subspaces and their local, unramified and infinity types are unchanged when only the coefficient localization changes. A constituent in one realization is represented in the other by completeness; the resulting isomorphism identifies its Hecke scalars and all these types. The defining selected-point conditions therefore hold on one side exactly when they hold on the other. Likewise the twisting character is the same prescribed character attached to (j, w). Its first Galois character agrees after coefficient mapping on split Frobenius elements by the common type formula, and hence everywhere by the density argument of Lemma 6.19. Each remaining character is this first character times its prescribed integral power of the cyclotomic character. Compatibility on ℤ_{ℓ} proves equality for these characters as well.
The finite excluded sets also match: the auxiliary set is fixed by w, and the ramification set is determined by the common residual characteristic. Thus the permitted split Frobenius conditions coincide. The two models in this construction use the same algebraic closure of E, so the group map γ is the identity, with identical inertia and wild-inertia subgroups. These checks, together with Lemma 6.26, give precisely the localization correspondence used in Propositions 6.20 and 6.25, provided the asserted common automorphic realization and its complete localization exist.
Lemma 6.28 (A finite requirement splits all constituents). Fix ℓ, a complex identification of ℚ_{ℓ}, a dense set of allowed Frobenius elements, and their characteristic polynomials. There is a finite set S ⊂ℂ with the following property. Every continuous semisimple integral representation of the given dimension having these Frobenius polynomials has absolutely irreducible simple constituents over its coefficient fraction field
K, provided the image of K contains S. The coefficient field is finite over ℚ_{ℓ} and its embedding respects that field.
Proof. If no such representation exists, the empty set suffices. Otherwise choose one and view it over ℚ_{ℓ}, denoting it by r_{0}. It is semisimple there. In this use of scalar extension the coefficient fields have charac-teristic zero: the finite-dimensional semisimple image algebra is separable, so its extension to an algebraic closure is a product of matrix algebras. Decompose the representation into irreducible subspaces V_{j}. Choose matrices E_{j} identifying a coordinate space with V_{j} and left inverses D_{j}, so D_{j}E_{j} = 1.
There is a finite set of coefficient numbers containing all the entries of r_{0}(g) in its generated ℤ_{ℓ}-algebra, uniformly in g. Indeed the original coefficient integer ring is a finite ℤ_{ℓ}-module; the images of a finite module-generating set suffice for every entry. Add the finitely many entries of every E_{j}, D_{j} and take their complex images as S. If K contains S, it contains every entry of r_{0}(g) and of E_{j}, D_{j}. The restrictions
Selectable PDF notation • transcription unchecked
D_{j}r_{0}(g)E_{j}therefore define representations over K. The same observation for g^{−1} gives their inverses over K. Their extension to ℚ_{ℓ} is the original irreducible representation on V_{j}, so each is absolutely irreducible.
Let r be any other representation in the statement. Continuity and density identify all its characteristic polynomials with those of r_{0} after mapping into ℚ_{ℓ}. Semisimple representations over this characteristic-zero algebraically closed field with these polynomials are isomorphic. One can see the character criterion directly on their finite-dimensional image algebras. Each is a product of full matrix algebras acting with positive integer multiplicities on the representation. Its trace pairing is nondegenerate, since on each factor it is that nonzero multiplicity times the matrix trace pairing. Equality of traces on group products therefore identifies the two image algebras: any linear relation on one side is a relation on the other. Under this identification the traces of the central simple-factor idempotents determine each multiplicity, by division by the dimension of its simple module in characteristic zero. The two modules are thus isomorphic.
For any irreducible K-submodule W of r, some V_{j} maps nontrivially into W ⊗_{K} ℚ_{ℓ} under that iso-morphism. The equations defining an intertwiner from the K-model of V_{j} to W are homogeneous linear equations over K. Although indexed by group elements, their row span has finite dimension, so finitely many of the equations suffice. A nonzero solution after field extension implies a nonzero solution over K, by equality of matrix ranks under field extension. Such a nonzero intertwiner is an isomorphism, because its source and target are irreducible. Hence W is absolutely irreducible, as required.
For use simultaneously over a coefficient place w, there are only finitely many Frobenius systems to consider. They come from the fixed finite-dimensional complex automorphic spaces at the finitely many levels 1 ≤ j ≤ m, after excluding the one fixed finite set of bad primes. Distinct common eigenvalue systems give linearly independent eigenvectors: from a shortest linear relation, apply a Hecke operator minus its eigenvalue on one vector; an operator distinguishing a second system gives a shorter nonzero relation. Thus their number is bounded by the dimension of the space. Taking the finite union of the sets S in the lemma over these systems imposes one finite requirement at w. Every compatible good representation in every later model then has absolutely irreducible constituents. This is the splitting requirement used before choosing the idempotents; its uniformity is over all later models and all their good representations, not only the initially chosen family.
Background fact 6.29 (Galois representations attached to polarized cusp forms). Let E be a CM num-ber field and (π, χ) a regular algebraic polarized cuspidal automorphic representation of GL_{a}(𝔸_{E}), with χ_{v}(−1) = (−1)^{a} at infinity. For every ℓ and ι : ℚ_{ℓ} ≃ ℂ there is a continuous semisimple polarized Ga-lois representation with the normalized local Langlands polynomials away from ℓ and pure Weil–Deligne parameters. It is de Rham, with Hodge–Tate weights a_{ιτ,i} + a −i. At a place above ℓ with Iwahori-fixed
vectors it is semistable and satisfies normalized local compatibility; if the local representation is unramified, it is crystalline.
This is Barnet-Lamb–Gee–Geraghty–Taylor, Potential automorphy and change of weight, Annals of Mathematics 179 (2014), 501–609, Theorem 2.1.1, page 538. Its CM, cuspidal, regular-algebraic and po-larization hypotheses are exactly the ordinary hypotheses stated here; the source polarization includes the displayed sign. The normalization is | det |^{(1−a)/2}. The source’s weight formula agrees because subtracting the infinitesimal-character entry −a_{ιτ,i}−(a+1−2i)/2 from (a−1)/2 gives a_{ιτ,i}+a−i. Identification of the abstract local period-model interfaces with these ordinary representations remains an additional application requirement.
Proposition 6.30 (The selected main complement exists integrally). Let r_{0} : G_{E} → GL_{V} (O) be continuous and semisimple over K. At a selected point suppose its automorphic parameter is the prescribed isobaric list of cuspidal strings, containing the distinguished string (1, µ, t), with regular algebraic total infinity type and the stipulated local matching. Suppose each cuspidal constituent is conjugate self-dual. Then the fixed integral characters χ_{1}, . . . , χ_{t} have a continuous integral main complement ρ : G_{E} → GL_{r}(O), where
Selectable PDF notation • transcription unchecked
r = V −t.
Proof. We first construct the character of the remaining strings over ℚ_{ℓ}. Write V = t + 2u + 1, as required by the rank parity. For a remaining string (a, π, d) its infinity entries have the form k +(V −d)/2, and those of µ have the form k_{0} +(V −t)/2. This follows by taking the corresponding submultisets of the regular total infinity parameter. They are distinct because that total parameter has no repetitions. If a + d −V = 2e + 1 is odd, the entries of π are (k −e) + (a −1)/2, so π is regular algebraic. Use its Galois representation from
Background fact 6.29 and, for 0 ≤b < d, twist it by ε^{e−b}. Here arithmetic Frobenius has cyclotomic value q_{w}. Its trace is the required string contribution because
Selectable PDF notation • transcription unchecked
^{1 −}_{2}^{a}+ e −b =^{d −}_{2}^{V} −b.If a + d −V = 2e is even, π ⊗µ is regular algebraic: its infinity entries minus (a −1)/2 are k + k_{0} + u + 1 −e ∈ℤ. It is still conjugate self-dual. Use its Galois representation, twist by χ^{−}_{1}^{1} and by ε^{e−u−1−b}
for 0 ≤b < d. The local type identity is χ_{1}(Frob_{w}) = µ_{w}q_{w}^{(t−V )/2}. It cancels the extra µ_{w} factor, and the remaining exponent is
Selectable PDF notation • transcription unchecked
^{1 −}_{2}^{a}−^{t −}_{2}^{V} + e −u −1 −b =^{d −}_{2}^{V} −b.Thus in both cases these are actual characters of continuous finite-dimensional representations, with the required normalized string traces. Sums over the finite string list remain such characters; denote the resulting trace by T(g).
At every allowed Frobenius outside finitely many bad primes, local matching and the sum of the string parameters give
Selectable PDF notation • transcription unchecked
∑t
tr r_{0}(g) = T(g) + χ_{c}(g).
∑c=1Indeed the distinguished string contributes µ_{w}q_{w}^{(t−V )/2t}_{b}^{−}_{=0}^{1}q_{w}^{−b}, which is exactly ∑_{c}χ_{c}(g) because χ_{c} = χ_{1}ε^{1−c}. Continuity and split Frobenius density extend the trace identity to every g. The characteristic-zero semisimple character criterion proved in Lemma 6.28 shows that the direct sum of these t characters occurs in r_{0} over ℚ_{ℓ}: its multiplicities are present in the displayed genuine character decomposition.
This character subrepresentation descends to K. Each character is already defined over K, and its si-multaneous eigenspace is the solution space of linear equations over K. Their finite row span shows that
dimensions are unchanged by field extension. Thus enough independent character eigenvectors can be cho-sen over K, including their prescribed multiplicities. As r_{0} is semisimple over K, their direct sum has an invariant complement W of dimension r. The intersection W ∩O^{V} is an invariant O-lattice: clearing denom-inators in a K-basis proves it spans W, and it is a finite torsion-free submodule of O^{V}. It is therefore free of rank r. Both g and g^{−1} preserve it, giving an integral invertible matrix representation ρ after choosing a basis. Its continuity follows by restricting the continuous action of r_{0} to this finite free lattice. The resulting rational decomposition is r_{0} ≃ρ ⊕⊕_{c}χ_{c}, as claimed.
Here a coefficient requirement is a function B assigning a finite subset B(w) ⊂ℂ to each coefficient place w. A model built for B is required to contain these numbers in its localized coefficient fields under the specified complex embeddings. There may be infinitely many places w; finiteness is required separately at each place. No union over all places is claimed to be finite.
Proposition 6.31 (A common enlargement for all live indices). Suppose the family construction has the following properties.
(i) For every coefficient requirement B there is a built model. For some initial B_{0}, all models built for B_{0} have good Hecke–Galois families, decompositions at selected eigencharacters, and the residual and integral order data needed to choose the adapted reference objects in Lemma 6.24.
(ii) Fixing one such initial model M_{0}, there is a requirement B_{1} ⊃B_{0} such that every model M_{1} built for
B_{1} has the following property. Each of its local indices i_{1} corresponds to an index i_{0} of M_{0} with the same coefficient place and a compatible localization and coefficient extension. It identifies eigenchar-acter labels, selected sets, Hecke eigenvalues, inertia, wild inertia and cyclotomic characters.
(iii) If i_{1} is live, its corresponding i_{0} is live. The good reference representations and their adapted data can
be transported through this correspondence; if their finitely many corner roots lie in the new fraction field, the transported data have all the required integral and root properties.
(iv) In M_{0} the indices over a place w are the union of finite sets locs_{0}(j, w) for 1 ≤j ≤m. A model built
for B contains B(w) in its coefficient fraction field at every index over w. Enlarging B preserves all previously imposed built-model properties.
Then there is a coefficient requirement B_{∗} such that every model built for B_{∗} has a good reference family with the required adapted wild-inertia, idempotent and split-corner data at every live local index.
Proof. Choose a model M_{0} built for B_{0} and then choose B_{1} from (ii). For each live index i_{0} of M_{0}, choose the finite root set S_{i0} and the corresponding adapted reference family of Lemma 6.24. Set S_{i0} = ∅ at a nonlive index. Define
Selectable PDF notation • transcription unchecked
∪m ∪
B_{∗}(w) = B_{1}(w) ∪ S_{i0}.
j=1 i_{0}∈locs_{0}(j,w)For each w the two unions have finite index sets and finite summands. Thus B_{∗} is a permitted coefficient requirement.
Let M_{1} be any model built for B_{∗} and let i_{1} be any live index in it. Since B_{1} ⊂B_{∗}, condition (ii) gives its corresponding i_{0} in M_{0}. It is live by (iii). Its level index lies between 1 and m, and it is over the same place w as i_{1}. Hence S_{i0} ⊂B_{∗}(w). Condition (iv) puts every number in S_{i0} in the coefficient fraction field at i_{1}. Transport the chosen reference family and its adapted data using (iii), and apply the finite-root lemma to obtain all the corner factorizations in that ring.
This construction works for every live i_{1} in every model built for the one requirement B_{∗}. In particu-lar B_{∗} is chosen before M_{1}, before its live index, and before any subsequent compatible representation or
good eigencharacter. Condition (i) also gives an actual model built for B_{∗}, so the assertion is not only an implication on an uninhabited collection of models.
Lemma 6.32 (The definite unitary finite model). Let E/F be a CM extension and let G = U(1^{a}) be the unitary group of the standard positive definite Hermitian form. Fix an algebraic weight W, a compact open level J ⊂G(𝔸_{F,f}), and a character χ : J →ℂ^{×} whose kernel contains a principal congruence subgroup. Choose a number-field model of the weight and a field containing the finite values of χ. At each Hecke place below, assume that the level has its maximal integral local factor and that χ is trivial on that factor. Then the space
Selectable PDF notation • transcription unchecked
ℳ_{J,χ}(W) = {f : G(𝔸_{F,f}) →W : f(γgj) = ρ(γ)χ(j)f(g)}has a finite-dimensional model over that field. At each coefficient prime above ℓ it has a full integral lattice preserved by all split Hecke double-coset operators supported away from ℓ and from the nonmaximal levels. These operators are simultaneously semisimple and commuting. The fixed-level space has a complete finite decomposition into the automorphic isotypic constituents which occur in it.
Proof. At each infinite place the group is the compact unitary group of a positive definite complex Hermitian form. The group over F is connected reductive, and F is totally real. Thus the finite-class-set assertion applies: for a compact open K the set G(F)\G(𝔸_{F,f})/K is finite, and the arithmetic stabilizers are finite. We use Greenberg–Voight, Lattice methods for algebraic modular forms on classical groups, Contributions in Mathematical and Computational Sciences 6 (2014), 147–179, Section 2, Proposition 2.3 and Lemma 2.4; their Proposition 2.5 identifies the full Hecke constituents with automorphic representations of the specified weight and level.^{6} All its group and archimedean hypotheses have just been checked.
Choose representatives x_{1}, . . . , x_{h} for the class set at level J, and write ∆_{i} = G(F) ∩x_{i}Jx^{−}_{i}^{1}. A form is determined by f(x_{i}). The permitted values form the subspace of W^{h} cut out by
Selectable PDF notation • transcription unchecked
ρ(γ)χ(x^{−}_{i}^{1}γ^{−1}x_{i})v_{i} = v_{i} (γ ∈∆_{i}).The precise inverse follows by writing γx_{i}(x^{−}_{i}^{1}γ^{−1}x_{i}) = x_{i}. Conversely these equations make the prescrip-tion f(γx_{i}j) = ρ(γ)χ(j)v_{i} independent of the decomposition. They are finitely many linear equations over the chosen field. Kernels of such a finite matrix commute with field extension, proving the model assertion. The character values are finite: a principal congruence subgroup in its kernel has finite index in compact J. They are roots of unity, hence integral units at every coefficient prime.
Here is a stable lattice construction which also treats nontrivial weights. Fix a finite local coefficient extension K/ℚ_{ℓ} containing all the required embeddings of E and the character values, with integers O. The algebraic weight extends to a continuous representation ρ_{ℓ} of G(F ⊗ℚ_{ℓ}) over K. The compact image of the projection of J has a stable full O-lattice L_{W} : start with a basis lattice; its compact orbit is bounded, so its O-span is contained in a fixed uniformizer denominator of the basis lattice. That span is finite by Noetherianity, contains the basis lattice, and is stable under the compact group and its inverse. The weight models and integral substitution action used here are the ones of Lemma 10.55. Set L_{i} = ρ_{ℓ}(x_{i,ℓ})L_{W} , and let ⊕
Selectable PDF notation • transcription unchecked
L = ℳ_{J,χ}(W_{K}) ∩ L_{i}.
iThis intersection is finite and torsion-free over the DVR O, so free. It is full: take a K-basis of the finite kernel space and multiply its finitely many coordinates by one sufficiently large uniformizer power to put them in the L_{i}.
^{6}Author text: https://jvoight.github.io/articles/orthohecke-240123.pdf, pp. 4–7. The finite-class assertion is stated there as Gross’s Proposition 4.3.
Let JyJ = ∐_{b}y_{b}J be a Hecke double coset supported away from ℓ. Its representatives may all be chosen with ℓ-component equal to 1. Indeed the local factor of this double coset at ℓ is just J_{ℓ}; absorb it in the right J-factor. For each i, b write x_{i}y_{b} = γx_{j}k, with γ ∈G(F) and k ∈J. Then
Selectable PDF notation • transcription unchecked
γ_{ℓ}x_{j,ℓ}= x_{i,ℓ}k_{ℓ}^{−1}.Since ρ_{ℓ}(k_{ℓ}^{−1}) preserves L_{W} , ρ(γ) carries L_{j} into L_{i}. The scalar χ(k) is a unit. Consequently every term in
Selectable PDF notation • transcription unchecked
(T_{y}f)(x_{i}) =∑f(x_{i}y_{b})
blies in L_{i} when f ∈L. This proves integral stability. Any normalization by a power of the residue cardinality at the Hecke place preserves this conclusion, since that cardinality is an ℓ-adic unit.
For semisimplicity, put a positive invariant Hermitian metric on the complex weight by averaging over the compact infinite group. Integration on the compact automorphic quotient gives a positive Hermitian metric on the fixed-level space. Right translation is unitary. Thus convolution by a Hecke function has adjoint convolution by h^{∗}(g) = h(g^{−1}). At a split maximal place the spherical double-coset algebra is commutative, and contains this adjoint. Operators at distinct places commute as they act on different factors. Every operator under consideration therefore commutes with its adjoint and is normal. The finite-dimensional spectral theorem diagonalizes it; its eigenspaces are preserved by every commuting operator. Repeated refinement gives a common eigenbasis after at most the dimension many nontrivial refinements. The matrices are defined over the model field, so their minimal polynomials are square-free after algebraic closure at the local coefficient prime as well.
For completeness of the automorphic constituents, first use a principal congruence level K_{0} ⊂ ker χ. The full Hecke action on the finite space of K_{0}-fixed forms is closed under adjoints. Every invariant sub-space therefore has invariant orthogonal complement; induction on dimension decomposes it completely into irreducible modules for this full algebra. Proposition 2.5 identifies these with precisely the automorphic representations having the stated weight and nonzero K_{0}-fixed vectors. Taking the (J, χ)-type part keeps exactly their corresponding type subspaces. It creates no further constituents, and gives all of the specified fixed-level space. There are finitely many because that space is finite-dimensional.
The compactness of the selected level is a check on its actual local conditions. They put it inside the maximal integral level: use the product-level condition to isolate each local component; at a hyperspecial place its membership is maximal integral membership, and at the designated parabolic or unipotent places the defining reduction conditions also include integrality of the matrix and its inverse. Containing a principal congruence subgroup then makes it a finite union of its cosets in that compact maximal level. It is therefore compact open. This is why the definite model above applies to these levels.
The lemma constructs a stable integral model. The arithmetic input, however, uses a specified geometric lattice and specified complex realization maps. To identify this construction with that input one still needs their rational comparison and the assertion that the prescribed finite geometric generators span and are stable with integral coefficients. The existence of some stable lattice proves neither assertion about that particular lattice. Once those comparisons are supplied, the following proposition constructs the required local blocks.
Lemma 6.33 (Prime-to-ℓ isogenies of integral abelian models). Let R be a complete DVR of residue char-acteristic ℓ, with fraction field K, and let A/R be a polarized abelian scheme. Suppose f : A_{K} →B_{K} and f^{′}: B_{K} →A_{K} are isogenies with
Selectable PDF notation • transcription unchecked
f^{′}f = [p^{2}], ff^{′}= [p^{2}], p ̸= ℓ.Then B_{K} has an abelian scheme model B/R, both maps extend, and df : Lie A → Lie B is an isomorphism. Any commuting integral endomorphism action preserved by the isogenies is preserved in this construction. In particular good reduction and the ranks of its integral Lie eigenspaces are preserved.
Proof. The kernel H_{K} of f is a subgroup of A_{K}[p^{2}]. The latter extends to the finite etale group A[p^{2}] over R, since p is a unit. Its subgroup H_{K} extends to a finite etale closed subgroup H ⊂A[p^{2}]. For an explicit descent argument, pass to a finite unramified extension of the complete DVR which splits A[p^{2}]. There it is a constant finite group, and H_{K} is the union of the constant components belonging to one subgroup. That union extends over the DVR and is stable under the descent action because its generic subgroup was defined over K. Descent returns H over R.
Translation by H acts freely on the projective R-scheme A. The fppf quotient B = A/H is a scheme and A → B is an H-torsor. This is precisely the finite locally free, free-action, projective case of Stacks Project, Lemma 66.14.2, Tag 07S7.^{7} Every hypothesis holds here. Since H is etale, this is a finite etale surjection. Smoothness and properness descend along it. The group law descends because H is a subgroup of a commutative group, and each geometric fiber is connected as the image of a connected abelian variety. Thus B/R is an abelian scheme. Its generic fiber is the quotient A_{K}/H_{K}, canonically the given B_{K}.
Multiplication by p^{2} on A kills H, so the quotient property gives a map f^{′} : B → A. It extends the given generic map and satisfies f^{′}f = [p^{2}] on A. Surjectivity of f then gives ff^{′} = [p^{2}] on B. Taking differentials gives
Selectable PDF notation • transcription unchecked
df^{′}df = p^{2}1, df df^{′}= p^{2}1.Hence p^{−2}df^{′} is the integral inverse of df. If an integral endomorphism preserves H_{K}, it preserves H by the constant-group descent description, so descends uniquely to B. Commutation relations hold by descent, or by equality on the dense generic fiber of a flat scheme. The differential is equivariant, giving isomorphisms on each integral idempotent eigenspace. It carries an integral basis of each to an integral basis, proving the rank assertion.
Lemma 6.34 (Integral values after a unit change of frame). Let R^{′}/R be a finite extension of DVRs with fraction fields K^{′}/K, and let j : K^{′} ↪→ Ω extend the specified embedding of K. Suppose an algebraic section F has value b ∈R^{′} at an integral framed point y^{′}. Suppose its value at a Hecke-related geometric point z is
Selectable PDF notation • transcription unchecked
F(z) = j(ω(u)b),
where u is a tuple of integral units and ω is the integral-weight character of change of frame. Then F(z) is integral over the image of R in Ω. The same holds when integrality of b is specified by the nonnegative normalized valuation tests at y^{′}.
Proof. The character ω(u) is a product of integer powers of units; negative powers remain integral units. Thus ω(u)b ∈R^{′}. The ring R^{′} is finite and torsion-free over the DVR R, hence finite free. Multiplication by this element on R^{′} has a matrix over R. Its monic characteristic polynomial annihilates the element, by applying the matrix Cayley–Hamilton identity to 1 ∈ R^{′}. Applying j to that equality proves the required integrality. For the valuation version, a finite extension multiplies the valuation by a positive ramification index. Its normalizing denominator is positive too. Nonnegativity therefore says exactly that the value lies in R^{′}, reducing to the first case.
These lemmas establish the good-reduction model and value arguments used at prime-to-ℓ Hecke points. Application to the prescribed geometric lattice also needs the actual framed theta-line comparison to be by integral units and its valuation tests to be the tests defining that lattice. For the degenerate integral models used elsewhere in that definition, the corresponding model and frame-extension assertion must be checked there as well.
Input 6.35 (The prescribed automorphic rational and integral realization). For every selected host and each of its weight and twist indices, the chosen complex automorphic space ℳ has the following simultaneous
^{7}Primary statement and proof: https://stacks.math.columbia.edu/tag/07S7.
data. There is a number-field space M_{K} and a specified isomorphism ℂ ⊗_{K,σ} M_{K} ≃ℳ agreeing with its geometric realization map. Every normalized split Hecke generator used in the construction preserves the image of M_{K}. At each twisted coefficient place w, the prescribed finite set of geometric lattice generators belongs to that image and spans ℳ; the image of each generator under every such Hecke operator is a combination of the same prescribed generators whose coefficients in K have nonnegative w-valuation. All identifications use the selected weight, level, character and theta realization.
Status. Input 6.35 is discharged in Subsection 6.6: Proposition 6.90 proves all eight clauses of GlobalTheory.HeightCounting.MtInput unconditionally for the exact prescribed geometric data (j_{Mt}, MtLat) of the selected family, while Propositions 6.92–6.95 establish the full six-field series and Hecke package SerN (including the theta-line unit pin and degenerate-model extension in Proposition 6.94).
Lemma 6.36 (The Hecke action on the prescribed lattice). Hypothesis 6.35 gives commuting rational Hecke endomorphisms of M_{K} and a full stable lattice at each twisted coefficient place. The normalized operator is independent of the uniformizer used at its split Hecke place.
Proof. For a Hecke generator T and v ∈M_{K} there is a unique preimage T_{K}v of T(jv) under the injective realization j. Complex linearity and injectivity show T_{K}(v + v^{′}) = T_{K}v + T_{K}v^{′} and T_{K}(av) = aT_{K}v for a ∈K. Thus T_{K} is K-linear. The same injectivity transports commutation of two complex Hecke operators to their rational preimages.
Pull the prescribed finite generating set back to M_{K}. It spans M_{K}: if its K-span were proper, a nonzero K-linear functional vanishing on that span would extend to a nonzero complex functional vanishing on the specified complex spanning set, a contradiction. Each displayed integral Hecke combination pulls back to the identical combination in M_{K}, again by injectivity. Its O_{K,w}-span is consequently stable. After extension to the complete local coefficient ring it is finite, torsion-free and full, and therefore finite free over the DVR.
Finally two uniformizers at the same split place differ by a unit. The diagonal matrices with the first k entries equal to these uniformizers differ by an integral invertible diagonal matrix. They define the same maximal-compact double coset. The chosen pure local representatives differ only by elements on which the level character is trivial. Their Hecke sums are therefore the same; the residue cardinality used for normalization is unchanged. The lifted rational endomorphisms are the same by uniqueness.
Thus the remaining initial assertion is the prescribed rational and integral comparison in the hypothesis, rather than finiteness of an unspecified automorphic space or the existence of an unspecified invariant lattice. The phase, theta-line, valuation and lattice identifications which prove it must be made in the same selected family.
Proposition 6.37 (The localized Hecke models from an integral action). Fix a finite-dimensional automor-phic coefficient space over a finite extension K_{0}/ℚ_{ℓ} and a full lattice L_{0} preserved by its commuting Hecke operators. Suppose those operators are semisimple after algebraic closure and the fixed complex realiza-tion identifies them with the required automorphic double-coset operators. Then, after a finite coefficient extension, the Hecke action has finite complete eigencharacter sets and canonical local blocks. One may simultaneously require any specified finite collection of complex numbers in the coefficient field. Every resulting local lattice is finite free and the local Hecke algebra acts faithfully on it.
Proof. Commuting semisimple operators can be diagonalized simultaneously over an algebraic closure. To see this for a possibly infinite operator family, decompose into eigenspaces of one nonscalar operator; all other operators preserve them. Repeat on any summand where an operator is nonscalar. Each such refinement strictly increases the number of nonzero summands, so after at most the dimension many refinements every operator acts by scalars on every summand. Thus only finitely many operators, eigenvalues and change-of-basis coefficients are needed to realize the simultaneous decomposition. Adjoin them to K_{0}. The finitely
many specified complex numbers also have algebraic preimages under the fixed identification ℚ_{ℓ} ≃ ℂ; adjoining them is another finite extension. Any additional finite coefficient requirements for the prescribed local bridges or finite-order twisting characters can be included in the same compositum.
Write O for its integer ring and L = L_{0}⊗O. The Hecke operators preserve L, and their generated algebra T ⊂ End_{O}(L) is a finite torsion-free O-module: it is an O-submodule of the finite free endomorphism module over the Noetherian DVR. All its eigenvalues are integral, since an endomorphism of a finite free lattice satisfies its monic characteristic polynomial. They consequently lie in O. Let X be the distinct simultaneous eigencharacters which actually occur. Evaluation embeds T in O^{X}: an element acting as zero on every eigenspace is the zero operator. It separates distinct x ∈ X by their definition. Moreover T ⊗K = K^{X}. For each fixed x, choose for each y ̸= x a generator or an element f_{y} ∈T with f_{y}(x) ̸= f_{y}(y). The product of (f_{y}−f_{y}(y))/(f_{y}(x)−f_{y}(y)) has value one at x and zero at every other point. These coordinate functions span K^{X}.
The finite O-algebra T is complete. Its quotient modulo the uniformizer is a finite-dimensional com-mutative algebra over the residue field, hence is the product of its local Artinian factors. Their mutually orthogonal idempotents lift uniquely to T: successive solution of z^{2}−z = 0 is possible because 2z −1 is a unit at an idempotent modulo the uniformizer. The lifted idempotents remain orthogonal and sum to one by uniqueness. Thus T =∏_{i}T_{i}, with each T_{i} local. For explicit locality, every maximal ideal of T_{i} contains the uniformizer, by integrality over O; its reduction is the unique maximal ideal of the corresponding Artinian factor.
Put L_{i} = e_{i}L. It is a finite free O-module, being a direct summand of the free lattice. The action of T_{i} on L_{i} is faithful: an element annihilating it annihilates all other summands automatically, and therefore annihilates L. Its eigencharacter set X_{i} consists of those x where e_{i}(x) = 1. This set is nonempty when T_{i} ̸= 0, since the rational evaluation algebra is faithful. All these eigenspaces are nonzero by the definition of X; no artificial eigencharacters have been introduced. The blocks partition X according to the maximal ideal obtained by residue evaluation, exactly the congruence partition used in Lemma 6.27.
The construction is compatible with a larger finite coefficient extension. The lattice and its operator action extend by tensor product, all the same scalar eigencharacters still occur, and the local factors are again determined by their residue congruence classes. Completeness of the eigencharacter enumeration and faithfulness have just been proved for every such extension. Applying this construction at each of the finitely many levels over each coefficient place gives all the localized Hecke and lattice objects required for the reference enlargement. The choices are made separately at each place, so they also accommodate a finite requirement at every place.
The proposition is the coefficient-enlargement argument used for the reference-family assertion. Its liveness and transport hypotheses are proved in Propositions 6.20 and 6.25; the adapted choices and finite root sets are proved above. Its localization correspondence follows from the finite coefficient enlargement and matching-block lemmas once the common complex automorphic realization is supplied with its complete localization. Its initial good-family and splitting requirements must likewise come from that realization and the attached Galois representations, using Background fact 6.29, Lemma 6.28, and Proposition 6.30. The remaining existence assertion is therefore precise: for every permitted coefficient requirement, construct the finite Hecke/localization models of the stated common automorphic space, with its selected isobaric parameters, and identify their local Galois and period data with those used in the congruence construction. It is not the assertion that an arbitrary family of matrices has adapted idempotents or that an unspecified root set can be adjoined.
We next prove the algebraic lifting mechanism behind the local crystalline exponent. This separates its actual calculation from the assertion that the required period modules realize the Galois representations in the application.
Let A be a commutative ℤ_{ℓ}-algebra with commuting endomorphisms φ, γ and elements X, Q satisfying
φ(X) = XQ, γ(X) = X(u + Xt), γ(Q) = vQ, v ∈A^{×}, v ≡1 (mod X),
where u ∈ ℤ^{×}_{ℓ}. Let D be an A-module with commuting semilinear operators φ, γ and an A-submodule N. Multiplication by X is injective on D, and multiplication by Q is bijective. Fix M ≥ 0 and suppose Q^{M}φ(N) ⊂N and (γ −1)N ⊂XN. Write
Selectable PDF notation • transcription unchecked
L = Q^{−M}N, C^{1}= L × XN, C^{2}= XL, d(y, z) = (γ −1)y −(φ −1)z.All these are ℤ_{ℓ}-modules. The operators γ preserve N, L, XL, and d(C^{1}) ⊂XL. For example, applying γ to Q^{M}y ∈N, using γ(Q) = vQ, and dividing by the unit v^{M} shows γy −y ∈XL, since v^{M}−1 ∈XA. Also (φ −1)(Xn) = X(Qφn −n) with Qφn −n ∈L. These give the asserted differential and stability properties explicitly.
Lemma 6.38 (Frobenius coordinates and the annihilator bound). Let O be a complete DVR of residue charac-teristic ℓ, R = O[[x]], and φ(x) = (1+x)^{ℓ}−1, with φ fixing O. Put X_{s} = φ^{s−1}(x) and Q_{s} = φ^{s−1}(φ(x)/x) for s ≥1. Let (𝔞_{i}) be a cyclic family of ideals of R. If, for some c, H ≥0,
x^{c}∈𝔞_{i}, Q^{H}_{s}𝔞_{i}⊂Rφ(𝔞_{i−1}) for every i,
then X_{s}^{H}∈𝔞_{i} for every i.
Proof. The ring R is free over φ(R) with basis 1, (1 + x), . . . , (1 + x)^{ℓ−1}. We recall a direct integral proof of the coordinate assertion. Modulo the uniformizer ϖ it is the familiar unique grouping of a power series according to its exponent modulo ℓ, since φ(x) ≡x^{ℓ}. Replacing x^{j} by (1 + x)^{j} is a triangular change of basis with diagonal one. Lift these residue coordinates arbitrarily to R, subtract their sum from the given series, and divide the error by ϖ. Repeat. Completeness of O, coefficient by coefficient, sums the resulting ϖ-adic series to coordinates in R. The Frobenius substitution and finite sum commute with these limits. For uniqueness, a zero relation has all its coordinates divisible by ϖ after reduction; divide and repeat. Separatedness gives zero coordinates.
Write the basis as e_{j} = (1+x)^{j}. An element of Rφ(𝔟) has unique Frobenius coordinates all in 𝔟: expand each scalar multiplying a generator φ(b) in the basis, and absorb b into its coordinates. Consequently, if φ(z) = zq and q𝔞 ⊂Rφ(𝔟), then zy ∈𝔞 implies that the coordinates y_{j} in y =∑_{j}e_{j}φ(y_{j}) satisfy zy_{j} ∈𝔟. Indeed multiply this coordinate expansion by φ(z) and use uniqueness of coordinates. We apply this with z = X_{s}^{H} and q = Q^{H}_{s}.
Suppose for contradiction that X_{s}^{H}∈/𝔞_{i}. Inductively, for every m ≥0,
Selectable PDF notation • transcription unchecked
X_{s}^{H}y ∈𝔞_{i+m} =⇒ y ∈(ϖ, x^{m+1}).For m = 0, y cannot be a unit, since otherwise X_{s}^{H} ∈ 𝔞_{i}. A nonunit of O[[x]] has constant coefficient divisible by ϖ, giving membership in (ϖ, x). At the next stage the coordinate observation gives X_{s}^{H}y_{j} ∈ 𝔞_{i+m} for all coordinates y_{j}. The inductive hypothesis gives y_{j} ∈(ϖ, x^{m+1}). Its Frobenius image belongs to (ϖ, x^{m+2}), because φ(x) ≡x^{ℓ} (mod ϖ) and ℓ(m + 1) ≥m + 2. The coordinate sum for y belongs to this same ideal, finishing the induction. Take m = c and y = x^{c}, which is permitted because x^{c} ∈𝔞_{i+c}. This yields x^{c}∈(ϖ, x^{c+1}), impossible by its coefficient of degree c. The contradiction proves the lemma.
Lemma 6.39 (Root expansions detect integral divisibility). Let O be a complete DVR of mixed characteristic (0, ℓ), enlarged if necessary to contain the ℓ^{s}th roots of unity. Put Q_{s}(x) = Φ_{ℓ}s(1 + x). For f ∈ O[[x]], divisibility by Q^{k}_{s} is equivalent to vanishing to order at least k in the expansion at every x = ζ −1, with ζ primitive of order ℓ^{s}. The same test holds for any local expansion parameter with nonzero linear coefficient.
Proof. The polynomial Q_{s} is monic and reduces modulo the uniformizer to x^{d}, where d = ℓ^{s−1}(ℓ−1). Every f ∈O[[x]] has a unique division f = Q^{k}_{s}g + r, with g ∈O[[x]] and r ∈O[x] of degree less than dk. A direct division proof is to perform this first modulo the uniformizer, where the divisor is x^{dk}, lift quotient and remainder, divide the error by the uniformizer, and repeat. Coefficientwise completeness gives g and the bounded-degree polynomial r. Reduction and repeated division of a zero relation proves uniqueness.
Each ζ−1 lies in the maximal ideal: in the residue field of characteristic ℓ, its residue satisfies (ζ−1)^{ℓs}= 0 and hence is zero. The expansions therefore converge. The d primitive roots are distinct in characteristic zero. If the stated order test holds, r has multiplicity at least k at each of these d roots, since Q^{k}_{s}g already does. But deg r < dk, so r = 0. This proves divisibility. The converse is immediate from the simple roots of Q_{s}. An invertible formal change of parameter preserves the order of vanishing, giving the last assertion.
Lemma 6.40 (Hodge jumps give the two height inclusions). Suppose a realized period lattice has the follow-ing expansion description at each primitive cyclotomic point. Its completed expansion lattice is the lattice of bounded-below Laurent series ∑_{k}d_{k}t^{k} with d_{k} ∈D_{cris} and d_{k} ∈ Fil^{−k}D_{dR} under the comparison map. It is full and bounded below. At the lower cyclotomic levels the expansion lattice is D_{cris}[[t]]. Expansion is compatible with scalars and Frobenius, Q expands as t times a unit, and the Frobenius expansion is obtained from the preceding level by applying crystalline Frobenius and multiplying its coefficient of degree k by ℓ^{k}. Assume also that the expansions of an integral basis are bases of these completed lattices, with the scalar expansions of Lemma 6.39. If the Hodge jumps lie in [−M, h], then
Selectable PDF notation • transcription unchecked
Q^{M}φ(N_{i−1}) ⊂N_{i}, Q^{h}N_{i} ⊂Rφ(N_{i−1}).Here the usual Frobenius commensurability after inverting Q and equality of the lattice rank with the crys-talline dimension are part of the realized expansion data.
Proof. The absence of filtration jumps below −M implies that every series with coefficients in D_{cris} van-ishing in degrees below M belongs to the expansion lattice. To check this using only fullness, place each coefficient in an arbitrarily high single degree, where fullness puts it in the required filtration. Constancy of the filtration below −M then puts it in the filtration for its original degree. The stated coefficient description proves membership of the whole series. Conversely every member of the lattice vanishes below −h. If a nonzero coefficient occurred in a lower degree, constancy of the filtration above h would allow its single monomial to be moved to arbitrarily negative degrees while remaining in the lattice. This contradicts its uniform lower bound.
For a basis vector at the preceding level the expansion is nonnegative, since that lower-level lattice is standard. The Frobenius expansion is therefore nonnegative too. The constant terms of these Frobenius expansions are linearly independent. Indeed the old basis expansions generate the standard lattice, so their constant terms span its crystalline space. Crystalline Frobenius maps that space onto the new one. Equality of rank and dimension turns their spanning family into a basis. Multiplying a Frobenius-expanded basis by Q^{m} consequently gives expansions beginning in degree m with linearly independent leading coefficients.
Let x ∈ N_{i−1}. Choose k with Q^{k}φ(x) ∈ N_{i} by commensurability. Put v = Q^{k+M}φ(x) ∈ N_{i}. Its expansion starts in degree k + M. After shifting down by k degrees it starts in degree M and therefore belongs to the expansion lattice by the first paragraph. Thus every coefficient of v in the integral basis has expansion divisible by t^{k} at every root. The root-divisibility criterion gives divisibility by Q^{k} in the coefficient ring. Divide those basis coefficients by Q^{k}, and cancel the invertible Q^{k} in the ambient period module. The result is Q^{M}φ(x) ∈N_{i}.
For the other inclusion let y ∈N_{i} and choose k with Q^{k}y =∑_{j}a_{j}φ(b_{j}), where the b_{j} are a basis of N_{i−1}. Choose one m with Q^{m}φ(b_{j}) ∈N_{i} for all j. Then
Selectable PDF notation • transcription unchecked
Q^{h+m+k}y =∑(Q^{h}a_{j}) Q^{m}φ(b_{j}).
jThe left side expands only in degrees at least m+k, because y has no terms below −h. On the right the basis-vector expansions begin in degree m with independent leading coefficients. Comparing the coefficients of degrees m, m + 1, . . . , m + k −1 successively shows that each scalar Q^{h}a_{j} has expansion divisible by t^{k}. At the first step independence kills their constant terms; after those vanish it kills their linear terms, and so on. Divisibility at all roots gives Q^{k} | Q^{h}a_{j} in the coefficient ring. Divide by Q^{k} in the displayed identity and cancel it in the ambient period module to obtain Q^{h}y ∈Rφ(N_{i−1}). All steps apply separately to each cyclic coefficient component, giving the two claimed inclusions.
Lemma 6.41 (The double-saturated image). Let A =∏_{i}O[[u]], let E = ∏_{i}(O[[u]][1/u])_{ℓ}^{∧}, and let N_{1}, N_{2} be finite free A-lattices in finite free E-modules D_{1}, D_{2}. Suppose E ⊗_{A} N_{j} = D_{j}. Give them commuting semilinear Frobenius and γ actions, with
Selectable PDF notation • transcription unchecked
φ(u) = (1 + u)^{ℓ}−1 = uq_{1}, γ(u) = uv, v ∈A^{×}.Frobenius may cyclically permute the coefficient components. Suppose its linearization on each D_{j} is an isomorphism, γN_{j} = N_{j}, (γ −1)N_{j} ⊂ uN_{j}, and that N_{j} and Aφ(N_{j}) contain one another after multi-plication by powers of Q_{s}. Let f : D_{1} → D_{2} be a surjection commuting with both actions, and suppose f(N_{1}) ⊂N_{2}. Write I = f(N_{1}) and
S = {z ∈N_{2} : for some b, ℓ^{b}z ∈I and u^{b}z ∈I}.
Then S is finite free over A, has E-span D_{2}, is γ-stable, and satisfies (γ −1)S ⊂ uS. Moreover S and Aφ(S) contain one another after multiplication by powers of Q_{s}.
Proof. The set S is an A-submodule: for a sum, increase the two exponents to their maximum; scalar multi-plication preserves both tests. It contains I and is contained in N_{2}, so it is finite over the Noetherian ring A. We will repeatedly use the following consequence of its definition:
Selectable PDF notation • transcription unchecked
z ∈N_{2}, ℓz ∈S, uz ∈S =⇒ z ∈S.Indeed the first test supplies an ℓ-power multiple of z in I, and the second supplies a u-power multiple; increase their exponents to one common exponent.
Work in one factor O[[u]]. Multiplication by u on S is injective. If ℓx = uy with x, y ∈S, the constant coefficient of x in an N_{2}-basis is killed by ℓ, so x = uz for z ∈ N_{2}. Cancellation gives y = ℓz. The preceding test puts z in S. Thus S/uS has no ℓ-torsion, hence no uniformizer torsion over O; it is finite free over O. Lift a basis to S. Nakayama gives a surjection O[[u]]^{h}→S. Its kernel J satisfies J/uJ = 0, because multiplication by u on S is injective and the reduced map is an isomorphism. Since J is finite, Nakayama gives J = 0. This proves freeness.
The E-span of I is D_{2}: apply f to the E-span of N_{1} and use surjectivity. Hence S also spans D_{2}. In a component, inclusion S ⊂N_{2} bounds its free rank above by the rank of N_{2}, whereas the surjection from its E-span onto D_{2} bounds it below by that same rank. Consequently the span map is represented by a surjective square matrix over the corresponding component of E. Its determinant is a unit, as follows by taking determinants of a right inverse, so this map is an isomorphism. In particular an A-basis of S is an E-basis of D_{2}.
The image I is γ-stable. If z ∈S has exponent b, then ℓ^{b}γz ∈I and u^{b}γz ∈I, the latter by applying γ to u^{b}z and dividing by v^{b}. Thus γS = S. Write (γ −1)z = uy with y ∈ N_{2}. Choose w ∈ N_{1} with f(w) = ℓ^{b}z and write (γ −1)w = uw^{′}, w^{′}∈N_{1}. Commutation and cancellation of u give ℓ^{b}y = f(w^{′}) ∈I. Also u^{b+1}y = u^{b}(γz −z) ∈ I by the two tests just proved. Increasing the exponent gives y ∈ S. This proves the asserted triviality modulo uS.
For the forward Frobenius inclusion, fix z ∈S and form the ideal
J_{z} = {a ∈A : for some c, Q^{c}_{s}aφz ∈I}.
Common denominators show that this is an ideal. Frobenius commensurability for N_{1}, transported by f, shows that ℓ^{b}∈J_{z} and (uq_{1})^{b}∈J_{z}. Since uq_{1} ≡u^{ℓ} (mod ℓ), the radical of J_{z} contains both ℓ and u. Some power u^{d} therefore belongs to J_{z}. Choose one Q_{s}-power which makes both ℓ^{b}φz and u^{d}φz lie in I, and also makes φz lie in N_{2}. Increasing the two testing exponents puts this Q_{s}-multiple of φz in S. Finitely many basis vectors give one uniform power.
For the reverse inclusion, first note that Q_{s} is a unit in E: modulo ℓ it is a power of the unit u, and an inverse lifts by an ℓ-adic geometric series. If b_{1}, . . . , b_{a} is an A-basis of S, then φb_{1}, . . . , φb_{a} is an E-basis of D_{2}, by invertibility of the linearized Frobenius. Put P =∑_{j}Aφb_{j} = Aφ(S). We need its elementary double-saturation property. The inclusion A/ℓ^{c}A →E/ℓ^{c}E is injective: the latter is obtained by inverting u in the former, where u is a non-zero-divisor. If y =∑_{j}e_{j}φb_{j} and both ℓ^{c}y, u^{c}y lie in P, write their integral coordinates as a_{j}, b_{j}. Uniqueness gives a_{j} = ℓ^{c}e_{j} and b_{j} = u^{c}e_{j} in E. The injectivity just noted implies a_{j} = ℓ^{c}a_{j}^{′} in A. The equality ℓ^{c}b_{j} = u^{c}a_{j}, followed by cancellation in A, gives b_{j} = u^{c}a^{′}_{j}. Since u is a unit in E, e_{j} = a^{′}_{j}∈A. Thus y ∈P.
Finally, for z ∈S, both ℓ^{b}z and u^{b}z belong to I. The reverse Frobenius inclusion for N_{1}, followed by f, puts a Q_{s}-power multiple of each in P. Choose a common power and use the double-saturation of P to put that multiple of z in P. Once again finitely many basis vectors give a common exponent, completing the proof.
Lemma 6.42 (The matrix divisibility in Wach uniqueness). In the notation of the preceding lemma, set X_{s} = φ^{s−1}(u). Suppose γ(u) = (1+u)^{η}−1, where η ∈1+ℓ^{s}ℤ_{ℓ} has infinite order. Let B, G, G_{1}, P, P_{1}, P_{1}^{′} be square matrices over A with
G ≡G_{1} ≡1 (mod u),
Selectable PDF notation • transcription unchecked
X_{s}^{a}Gγ(B) = γ(X_{s})^{a}BG_{1},
P_{1}P_{1}^{′}= Q^{m}_{s}^{′}1, Pφ(B) = Q^{m}_{s}BP_{1}.Then every entry of B is divisible by X_{s}^{a} in A.
Proof. Pass to the coefficient field and expand first at u = 0 using t = log(1 + u). In this expansion X_{s} = tw for a unit w, and γ acts by t →→ ηt. Suppose inductively that B = t^{b}C with b < a. Substitute in the second matrix identity and cancel t^{a+b}. Taking constant coefficients, using G(0) = G_{1}(0) = 1 and w(0) ̸= 0, gives
Selectable PDF notation • transcription unchecked
(η^{b}−η^{a})C(0) = 0.Infinite order of η makes the scalar nonzero. Thus C(0) = 0, increasing the divisibility exponent by one. Starting with b = 0 proves divisibility by t^{a} at u = 0.
Multiplying the last identity on the right by P_{1}^{′} gives
Selectable PDF notation • transcription unchecked
Q^{m}_{s}^{+m′}B = Pφ(B)P_{1}^{′}.Now expand at u = ζ −1 for ζ of order ℓ^{j}, 1 ≤ j ≤ s −1, with local parameter t = log((1 + u)/ζ). Frobenius carries this expansion to the one at ζ^{ℓ}−1, with parameter ℓt and the correspondingly permuted coefficient component. Induction on j therefore makes every entry of φ(B) divisible by t^{a} at this point. Here Q_{s}(ζ −1) ̸= 0, so Q_{s} is a unit in the local formal-series ring. The displayed matrix identity gives the same divisibility for B.
After a finite coefficient extension containing all these roots, X_{s} = (1 + u)^{ℓs−1}−1 is a monic distin-guished polynomial with precisely these distinct roots, including u = 0. Division by X_{s}^{a}, by the integral division argument of Lemma 6.39, leaves a polynomial remainder of degree less than a deg X_{s}. Its vanish-ing to order a at all the roots forces it to be zero. The quotient already has coefficients in the original O, by uniqueness of monic integral division. This proves the assertion in every coefficient component.
Lemma 6.43 (Uniqueness and the initial cokernel powers). Suppose two free integral lattices in the same D satisfy the Frobenius and γ conditions of Lemma 6.41 and are commensurable by powers of X_{s}. Then they are equal. Consequently, in that lemma, if its S and N_{2} are commensurable by powers of X_{s}, there is c such that both ℓ^{c}N_{2} ⊂f(N_{1}) and u^{c}N_{2} ⊂f(N_{1}).
Proof. For lattices M, M^{′} choose a with X_{s}^{a}M^{′} ⊂M, and write, in column bases b, b^{′}, X_{s}^{a}b^{′} = bB. The action matrices G, G_{1} of γ are the identity modulo u. Applying γ to this equality gives exactly X_{s}^{a}Gγ(B) = γ(X_{s})^{a}BG_{1}. Choose Frobenius denominators
Selectable PDF notation • transcription unchecked
Q^{h}_{s}φ(b) = bF, Q^{h}_{s}^{1}φ(b^{′}) = b^{′}F_{1}, φ(b^{′})P^{′}= Q^{e}_{s}b^{′}.All matrices are integral, by the two commensurability inclusions. Coefficient comparison gives
Selectable PDF notation • transcription unchecked
F_{1}P^{′}= Q^{h}_{s}^{1+e}1, (Q^{h}_{s}^{1}F)φ(B) = Q^{a}_{s}^{+h}BF_{1},where φ(X_{s}) = X_{s}Q_{s} was used in the second equality. Lemma 6.42 gives B = X_{s}^{a}B_{0} with B_{0} integral. Cancel X_{s}^{a} in the ambient period module to obtain b^{′} = bB_{0}, hence M^{′} ⊂ M. Interchanging the lattices proves equality.
Apply this result to S and N_{2}. The preceding lemma supplies every action and freeness condition, so the additional power-commensurability hypothesis gives S = N_{2}. Choose finitely many A-generators of N_{2}. Each satisfies the two defining tests for S with some exponent; their maximum works for every generator and hence every element. This is the asserted common c.
Lemma 6.44 (Frobenius equations force integral period coefficients). Let R be a perfect valuation ring of characteristic ℓ with fraction field F. Write φ for Witt Frobenius. If matrices over the indicated rings satisfy
Z ∈ Mat_{a×b}(W(F)), S, S^{′}, Y over W(R), φ(Z) = SZS^{′}+ Y,
then every entry of Z belongs to W(R). The same assertion holds after tensoring both Witt rings with any finite free ℤ_{ℓ}-algebra O, with Frobenius acting on the Witt factor.
Proof. Reduce the equation modulo ℓ, so Frobenius becomes the ℓth power on F. Among the finitely many entries of the reduced Z, choose one of largest multiplicative valuation c. If c > 1, every entry of SZS^{′}+Y has valuation at most c, since the coefficients of S, S^{′}, Y have valuation at most 1 and the valuation is ultrametric. But the selected entry on the left has valuation c^{ℓ} > c, a contradiction. Hence all reduced entries are in R.
Lift these entries by Teichmuller representatives to a matrix Z_{0} over W(R) and write Z = Z_{0} + ℓZ_{1} in W(F). Substitution gives
Selectable PDF notation • transcription unchecked
ℓ(φZ_{1} −SZ_{1}S^{′}) = SZ_{0}S^{′}+ Y −φZ_{0}.The right side lies in W(R) and has zero constant Witt coefficient, because it maps to an ℓ-multiple in W(F) and R →F is injective. Since R is perfect, the kernel of the constant Witt coefficient is ℓW(R). Thus the right side is ℓY_{1} with Y_{1} integral. Cancellation gives φZ_{1} = SZ_{1}S^{′}+ Y_{1}, the same type of equation. Iterate. For every n, each entry of Z agrees modulo ℓ^{n} with an element of W(R). Consequently each of its first n Witt coefficients lies in R. Taking all n proves that all its Witt coefficients lie in R, which is precisely membership in W(R).
For coefficients choose a ℤ_{ℓ}-basis o_{1}, . . . , o_{d} of O and use the regular representation of O by its left-multiplication matrices. It sends O⊗W(F) to d-by-d matrices over W(F), respects products and commutes with Frobenius. Replace each entry in the equation by this block matrix and apply the proved assertion. To recover the original entry, write 1 =∑_{j}c_{j}o_{j}. If ρ(z)_{ij} are its regular matrix entries, then
Selectable PDF notation • transcription unchecked
z =∑c_{j} o_{i} ⊗ρ(z)_{ij}.
i,jAll terms on the right are integral, proving the coefficient version without any assumption that O is unrami-fied.