Section 7.7 · PDF pages 388–411
Rank selection, tangent envelopes, and Hecke descent
The notation transcription is provisional. Check it against the printed mathematics when citing or reusing a formula.
7.7 Rank selection, tangent envelopes, and Hecke descent
Together with Subsection 7.5 (which establishes the positive-line and adelic lattice coordinates and the base Hecke point and tangent dictionary) and Subsection 7.6 (which constructs the admissible nonzero trace tests Spec/SpecTop, the integral congruence plans PlanEx, the selected projective Kuga–Sato tower KSAll, the Hodge-trace identity (174), and the cohomological realization and comparison package Steps), the geometric realization GeoOne is now established unconditionally (Corollary 7.44), modulo only the standard PEL moduli, definite-unitary vector realization, and étale, crystalline, and Hodge–Tate comparison theorems cited at their exact interfaces above (all verified axiom-free in the Lean source). We now pass from GeoOne to a single rank-a odd seed.
Proposition 7.45 (The source’s rank-selection transition). Assume the divisor theorem DivAll and the conju-gation, weight-gap and generic-multiplicity assertions used in seedFinOdd_of, in addition to GeoOne. Then the source obtains a single rank-a odd seed, rather than making an independent choice of rank-a represen-tation at each place.
Proof. For the fixed automorphic U, the weight-gap assertion first turns its prescribed sufficiently separated weight into the lacunary-weight hypothesis required by the geometric construction. To apply that construc-tion at a chosen place v, replace each of the two auxiliary places, when necessary, by its complex conjugate. The SameOr alternative makes the new place have the required restriction to ℚ(x); CplMem keeps it over ℓ. The conjugation assertions preserve respectively the elliptic local type and the Satake/lacunary-weight conditions. Their separation equivalences preserve the exclusions in the original sphericity hypothesis. This produces the placewise geometric witnesses of Proposition 7.4 for each desired v.
Apply DivAll once, using the geometric base witness. It produces one rank-a representation r and a comparison assertion universally quantified over every geometric witness satisfying VBase. This universal comparison, denoted hr in the proof, is the crucial order of quantifiers. For a place away from ℓ, choose its geometric Eichler–Shimura witness and apply the first part of hr. The generic-multiplicity assertion converts the seed’s generic Satake hypothesis into the multiplicity hypothesis used by that comparison. At a place above ℓ, choose the crystalline/jump witness and apply the second part of hr. Choose, separately, the Frobenius-annihilation witness and apply the last part of hr. These applications all concern the same r already selected by the divisor theorem. The FieldOK hypothesis also makes every place above ℓ split when that is required by the last comparison. Thus the use of separate geometric witnesses is logically valid; it is the universal comparison in the divisor theorem that puts their conclusions on one representation.
Source: Foundations/GaloisRepresentations/IntegralStructures/ ZetaCharacterDivisibility/SeedFinOddOf.lean:44--83; ZetaCharacterDivisibility/MainFin.lean:13--15.
Here is the mathematical argument behind that universal comparison. For representations r of dimension a and ρ of dimension m, the source writes r ⊂ρ when an injective matrix intertwines their actions. It writes ρ ⊂r^{⊕q} when an injective intertwiner into some number of copies of r exists. Its names for these conditions are Inside and Covers. The latter condition is stronger than saying that r is one constituent of ρ.
Lemma 7.46 (Polynomial pinning). If an a-by-a matrix M has separable characteristic polynomial and a monic polynomial Q of degree a annihilates M, then Q = det(T −M).
Proof. Extend scalars to an algebraic closure. Every characteristic root α has a nonzero eigenvector v. Eval-uating Q(M)v gives Q(α)v = 0, hence Q(α) = 0. Separability gives a distinct characteristic roots. Their product ∏_{α}(T −α) therefore divides Q. Both polynomials are monic of degree a, so they are equal. Injec-tivity of extension of coefficients gives the equality over the original field.
Source: IntegralStructures/ZetaCharacterDivisibility/PSepPin.lean:8--54.
Proposition 7.47 (Uniformity of the divisor construction). Suppose the divisor-extraction theorem gives, for each geometric ρ attached to U, a continuous semisimple rank-a representation r_{ρ} with r_{ρ} ⊂ρ ⊂r_{ρ}^{⊕q}, and with the prescribed Satake characteristic polynomials on an open dense set of good Frobenius elements. Suppose also the determinant and local subrepresentation/jump comparisons used in the source hold. Then one selected r works in the comparisons for every geometric ρ.
Proof. First fix the determinant character ψ attached to the central character of U. For a Satake multiset S_{v} of cardinality a, set ∏
Selectable PDF notation • transcription unchecked
Q_{v}(T) = (T −ι^{−1}s).
s∈SvIt is monic of degree a, and Q_{v}(0) = (−1)^{a} ∏_{s∈Sv}ι^{−1}s. The geometric Frobenius relation says that Q_{v} an-nihilates the relevant inverse Frobenius on ρ. The central-character comparison identifies (−1)^{a}Q_{v}(0) with ψ of that inverse Frobenius. Good split Frobenius elements outside a finite set are dense; inversion preserves density. Thus the divisor-extraction theorem applies to this dense family of annihilating polynomials with their fixed constant term.
The extracted r_{ρ} has an open dense locus where its characteristic polynomial is separable. Its intertwiner into ρ shows that Q_{v} also annihilates r_{ρ} at each relevant Frobenius. On that open dense locus, Lemma 7.46 identifies its characteristic polynomial with Q_{v}. Trivial inertia on ρ restricts to trivial inertia on r_{ρ}. This proves the defining good-place properties of the extracted divisor representation.
If r_{ρ} and r_{ρ}′ are two such divisor representations, exclude the union of their finite bad-place sets and intersect their two open dense separability loci. Good split Frobenius elements in this intersection are still dense. Both characteristic polynomials there equal Q_{v}, including the inverse-Frobenius convention just used. Continuity extends their traces to equal traces on all of G_{E}. The semisimple character criterion in characteristic zero then makes the two representations isomorphic. To recall its algebraic content, form the finite-dimensional algebra generated by their direct-sum image. Its semisimple quotient is a product of matrix algebras; the characters of its distinct simple modules are linearly independent, since evaluation on a matrix-unit diagonal in one factor reads the multiplicity of that simple module. Equality of characters therefore gives equal multiplicities and an intertwining isomorphism.
Choose a divisor r from one initial geometric witness. For any other witness ρ, apply extraction to obtain r_{ρ}, and use the isomorphism just proved to transport both embeddings to
Selectable PDF notation • transcription unchecked
r ⊂ρ ⊂r^{⊕q}.The determinant comparison gives det r = ψ. The source’s generic-multiplicity comparison now recovers the desired Frobenius eigenvalues of r from an Eichler–Shimura annihilating polynomial for ρ. Its local subrepresentation comparison passes crystallinity from ρ to r. The embedding into copies of r, together with the jump comparison, recovers all prescribed jumps, rather than only an inclusion of some jumps. Finally the determinant and Frobenius-annihilation comparison gives the crystalline Frobenius polynomial of r. These are precisely the three comparisons in DivAll, now universally quantified over the arbitrary ρ. This proves the asserted uniformity.
Source: IntegralStructures/ZetaCharacterDivisibility/PDivOne.lean:16--88; PDivUniq.lean:18--37; PhiOfDivOf.lean:166--182.
Lemma 7.48 (Spectral projectors and lacunary off-diagonal blocks). Let X ∈ Mat_{N}(K) have a distinct eigenvalues e_{1}, . . . , e_{a}, and suppose a nonzero polynomial of degree at most a annihilates X. There are nonzero orthogonal idempotents p_{i}, polynomial in X, with
Selectable PDF notation • transcription unchecked
∑p_{i} = 1, X =∑e_{i}p_{i}.
i iSuppose K has characteristic zero and the e_{i}’s satisfy the unequal-average lacunarity condition. If a sub-space 𝔤 of matrices is stable under Y →→[X, Y ], then p_{i}Y p_{j} ∈𝔤 for Y ∈𝔤 and i ̸= j.
Proof. Each eigenvalue is a root of the minimal polynomial: apply an annihilating polynomial to a nonzero eigenvector. The distinct linear factors T −e_{i} are pairwise coprime, so their product divides the minimal polynomial. Its degree is at most∏ a, since it divides the given nonzero annihilator. Therefore it is exactly _{i}(T −e_{i}). Put
Selectable PDF notation • transcription unchecked
b_{i}(T) =∏^{T −ej}, p_{i} = b_{i}(X).
j̸=i^{ei −ej}A polynomial vanishing at every ∑ e_{i} is divisible by that minimal polynomial and vanishes at X. Evaluating b^{2}_{i}−b_{i}, b_{i}b_{j}, _{i}b_{i}−1 and T −∑_{i}e_{i}b_{i} therefore gives all the stated identities. A p_{i} cannot vanish: otherwise its polynomial b_{i}, of degree a−1, would be divisible by the degree-a minimal polynomial, although b_{i}(e_{i}) =
Selectable PDF notation • transcription unchecked
1.
Write D(Y ) = [X, Y ]. The block p_{k}Y p_{l} is a D-eigenvector of eigenvalue e_{k} −e_{l}, and Y =∑_{k,l}p_{k}Y p_{l}. For i ̸= j, lacunarity makes e_{i} −e_{j} different from every other ordered pair difference. Indeed equality with e_{k} −e_{l} would give e_{i} + e_{l} = e_{k} + e_{j}; cancelling common indices gives equal averages of two disjoint nonempty sets. If one side repeats an index, it gives instead the forbidden equality of that singleton’s average with the average of the two distinct indices. The remaining case is the identical ordered pair.
Define q(T) =∏_{(k,l)̸=(i,j)}(T −(e_{k} −e_{l})). Applying q(D) to the block decomposition kills all blocks except (i, j), and yields
Selectable PDF notation • transcription unchecked
q(D)Y = q(e_{i} −e_{j})p_{i}Y p_{j}, q(e_{i} −e_{j}) ̸= 0.Stability under D, addition and scalar multiplication gives stability under every polynomial in D. Divide by that nonzero scalar to conclude p_{i}Y p_{j} ∈𝔤.
Source: EtaleCohomology/DivisorCycles/ProjPf.lean:9--103; GradePf.lean:14--137.
Lemma 7.49 (Counting eigenvalues along idempotent cycles). Let K be algebraically closed of charac-teristic zero, and let p_{1}, . . . , p_{a} be nonzero orthogonal idempotents summing to 1. Suppose y is invertible and yp_{i} = p_{π(i)}y for a permutation π. Suppose also no eigenvalue of y occurs on nonzero eigenvectors supported in two different π-cycles. Then y has at least a distinct eigenvalues. If a nonzero polynomial of degree at most a annihilates y, its action on any fixed block p_{i}V , π(i) = i, is scalar.
Proof. For a cycle of length k, choose its initial index j. The nonzero space p_{j}V is stable under y^{k}. Choose an eigenvector v for that restriction, with y^{k}v = βv. Invertibility gives β ̸= 0. The polynomial T^{k}−β has k distinct roots in characteristic zero. For each such root λ, put
Selectable PDF notation • transcription unchecked
∑k−1
w_{λ} = λ^{−m}y^{m}v.
m=0Its p_{j}-component is v, since the intermediate cycle points are distinct. Thus w_{λ} ̸= 0. Shifting the sum, and using λ^{k} = β, gives yw_{λ} = λw_{λ}. It is supported on this cycle. Different cycles have disjoint sets of these eigenvalues by hypothesis. Adding their lengths therefore gives a distinct eigenvalues, with one chosen eigenvector indexed by every block.
Let q ̸= 0 be a degree-at-most-a annihilator. All those a eigenvalues are roots of q, so its roots are exactly these and are simple. On a fixed block, any eigenvector of the restriction must have the one eigenvalue λ_{i} assigned to that block: another assigned eigenvalue would contradict the cycle-disjointness hypothesis. Write q = (T −λ_{i})q^{′}, with q^{′}(λ_{i}) ̸= 0. The operator q^{′}(y) is injective on p_{i}V . Otherwise its kernel there is a nonzero y-stable space and has an eigenvector; its eigenvalue must be λ_{i}, which would make q^{′}(λ_{i}) = 0. Since q^{′}(y)(y −λ_{i}) = 0, injectivity on that block gives yp_{i} = λ_{i}p_{i}.
Source: EtaleCohomology/DivisorCycles/CycleCount2.lean, proofs cyc, eigs, fixed_scalar and CycleCount_of.
Lemma 7.50 (Scalar diagonal corners). Let X =∑_{i}e_{i}p_{i}, with the e_{i}’s distinct, over an algebraically closed characteristic-zero field. Suppose Y commutes with every p_{i}, and every X + sY , s ∈ K, has a nonzero annihilator of degree at most a. Then Y p_{i} is a scalar multiple of p_{i} for each i.
Proof. Let R be the finite set of eigenvalues of Y . Choose s ̸= 0 outside the finite set
Selectable PDF notation • transcription unchecked
{^{e}_{µ}^{j}_{ −}^{ −}_{µ}^{e}_{′}^{i} : i ̸= j, µ, µ^{′} ∈R, µ ̸= µ^{′}}.Choose also c so that y = X + sY + c is invertible; only finitely many c’s are forbidden by its characteristic polynomial. It preserves every p_{i}V . An eigenvector of y in that block is an eigenvector of Y , and its eigenvalue is e_{i} + sµ + c. Two different blocks cannot share such an eigenvalue: if µ = µ^{′}, this would give e_{i} = e_{j}; otherwise it would be one of the excluded values of s. Thus Lemma 7.49 applies with π = 1. Adding c to a matrix preserves the bound on the degree of an annihilator, by replacing its polynomial q(T) with q(T −c). The lemma therefore makes yp_{i} scalar. Subtract (e_{i} +c)p_{i} and divide by s to make Y p_{i} scalar as asserted.
Source: EtaleCohomology/DivisorCycles/DiagScalar2.lean:62--151.
Lemma 7.51 (Passing the corner identities to the algebra). Let 𝔤 be a matrix Lie algebra, and let the p_{i}’s be as above. Suppose each off-diagonal block p_{i}Y p_{j}, i ̸= j, of an element of 𝔤 belongs to 𝔤, and an element of 𝔤 commuting with every p_{i} acts by a scalar on each p_{i}V . Then the algebra A generated by 𝔤 has scalar diagonal corners, and its off-diagonal blocks still belong to 𝔤.
Proof. Consider all matrices with these two properties. They are closed under addition and scalar multipli-cation, and contain scalar matrices. An element Y ∈ 𝔤 has these properties: subtract all its off-diagonal blocks. The result remains in 𝔤, is block diagonal, and so has scalar diagonal blocks by hypothesis.
For closure under products write (ab)_{ik} =∑_{j}a_{ij}b_{jk} in block notation. When i ̸= k, the terms j = i, k are scalar multiples of off-diagonal blocks. For the other terms, a_{ij}b_{jk} = [a_{ij}, b_{jk}], since the reverse product is zero by p_{k}p_{i} = 0. Thus every term lies in 𝔤. In a diagonal corner the term j = i is scalar. For j ̸= i, the bracket D = [a_{ij}, b_{ji}] belongs to 𝔤, is supported on the two diagonal corners i, j, and commutes with every p_{k}. Therefore Dp_{i} = a_{ij}b_{ji} is scalar on p_{i}V . This proves closure under multiplication. The resulting subalgebra contains 𝔤, hence contains its generated algebra A, proving both claims.
Source: EtaleCohomology/DivisorCycles/AlgStruct2.lean:43--134.
Lemma 7.52 (The tangent Lie algebra and its envelope). For a matrix submonoid S, define its polynomial closure Z(S) by all polynomials vanishing on S, and define 𝔱(S) by
Y ∈𝔱(S) ⇐⇒ df_{1}(Y ) = 0 for every polynomial f|_{S} = 0.
This is a Lie algebra. If S = ρ(Γ), its generated algebra A is normalized by ρ(Γ) and preserves every ρ-invariant subspace. In characteristic zero, every polynomial vanishing on S vanishes formally on exp(tY ) for Y ∈𝔱(S).
Proof. All polynomial substitutions used below are explicit substitutions of linear forms in the matrix coor-dinates. For a linear map L, the pullback f ∘L is a polynomial, and differentiation of constants, sums and products gives d(f ∘L)_{x}(Y ) = df_{Lx}(LY ). Define the polynomial derivation
Selectable PDF notation • transcription unchecked
(D_{Y} f)(x) = df_{x}(Y x).For Y ∈𝔱(S) and f|_{S} = 0, fix s ∈S and pull back f by x →→xs. The pullback vanishes on S. Its derivative at 1 shows (D_{Y} f)(s) = df_{s}(Y s) = 0. Conversely evaluating this assertion at s = 1 proves that preservation of the vanishing ideal by D_{Y} characterizes 𝔱(S). On a coordinate x_{ij}, D_{Y} x_{ij} =∑_{k}Y_{ik}x_{kj}. Applying two such derivations gives [D_{Y} , D_{Z}] = D_{ZY −Y Z}. Preservation of the vanishing ideal therefore proves closure of 𝔱(S) under Y Z −ZY , as well as its evident vector-space closure.
Conjugation by ρ(g) preserves S; pulling back a vanishing polynomial by this linear conjugation and differentiating at 1 proves that it preserves 𝔱(S). The conjugation is an algebra automorphism, so it also preserves A. If W is ρ-invariant, v ∈ W, and ℓ is a linear form annihilating W, the linear polynomial x →→ ℓ(xv) vanishes on S. Thus ℓ(Y v) = 0 for every such ℓ, and Y v ∈ W. Sums, products and scalar matrices preserve W, so all of A does too.
Finally set F_{Y} (f) = f(exp(tY )) ∈K[[t]]. The coordinate identity (exp(tY ))^{′}= Y exp(tY ), followed by the polynomial product rule, gives
Selectable PDF notation • transcription unchecked
F_{Y} (f)^{′}= F_{Y} (D_{Y} f), F_{Y} (f)(0) = f(1).If f vanishes on S, so does D_{Y} f. Induct on the coefficient index: the constant term is zero, and the dis-played derivative identity says that (k + 1) times the next coefficient equals the k-th coefficient for D_{Y} f. Characteristic zero allows division by k+1. All coefficients therefore vanish. No convergence of the formal exponential is needed.
Source: EtaleCohomology/DivisorCycles/ChainPf.lean; TangentFieldExtension.lean, TanField\_of; TanLiePf.lean, DX\_comm and TanLie\_of; CoeMulInv.lean, lie\_conj, lie\_keeps and EnvStab\_of; ExpVanPf.lean, deriv\_F and coeff\_F.
Lemma 7.53 (The annihilator degree bound passes to tangents). Suppose every x ∈S is annihilated by a nonzero polynomial of degree at most a. The same holds for every x ∈Z(S), and, in characteristic zero, for every Y ∈𝔱(S).
Proof. The degree condition is equivalent to linear dependence of the a + 1 matrices 1, x, . . . , x^{a}: their dependence coefficients form the required polynomial, and its coefficients give the converse. After viewing these matrices as coordinate vectors, this is equivalent to vanishing of every (a + 1)-row determinant of the matrix whose columns are those vectors. Indeed dependence kills every such determinant, while independent columns have a + 1 independent rows and hence a nonzero minor. Each minor is a polynomial in the entries of x, defined over K, and the equivalence remains valid over every extension field. These equations already prove the assertion on Z(S).
For the tangent assertion put
Selectable PDF notation • transcription unchecked
∑_{Y}k+1U(t) = exp(tY ) = 1 + tZ(t), Z(t) = t^{k}.
Selectable PDF notation • transcription unchecked
_{k≥0}(k + 1)!The preceding lemma makes all these minor equations vanish on U(t). In the fraction field of K[[t]], they give a nonzero polynomial q of degree at most a with q(U(t)) = 0. The substitution T →→1+tT is invertible over this fraction field, so q(1 + tT) is nonzero, has the same degree, and annihilates Z(t). The same minor equations consequently vanish on Z(t) in the fraction field and hence already in K[[t]], whose map to that field is injective. Taking constant terms makes them vanish on Z(0) = Y . Their field-independent degree criterion now gives the required annihilator of Y .
Source: EtaleCohomology/DivisorCycles/DegPoly2.lean, deg\_iff\_not\_li, not\_li\_iff and DegPoly\_of; Tan-gentDegree.lean, tan\_deg and TanDeg\_of.
Lemma 7.54 (Finite-index groups have the same tangent space). If U is normal of finite index in Γ, then 𝔱(ρ(Γ)) = 𝔱(ρ(U)). Moreover the polynomial closure of a matrix submonoid is closed under multiplication.
Proof. For the multiplication assertion, if f vanishes on S and s ∈S, the polynomial y →→f(sy) vanishes on S, hence on Z(S). For y ∈Z(S), the polynomial x →→f(xy) therefore vanishes on S, hence on Z(S). This proves the assertion.
Write H = Z(ρ(U)), and let f vanish there. For each coset gU there is a polynomial q_{g} with q_{g}(1) ̸= 0 such that fq_{g} vanishes on ρ(gU). To see this, either some polynomial vanishing on that coset is nonzero at 1, in which case take it for q_{g}, or every such polynomial vanishes at 1. In the latter case pull back each equation of H by x →→ρ(g)^{−1}x. It vanishes on ρ(gU), so evaluation at 1 gives ρ(g)^{−1}∈H. Some positive power g^{k} belongs to U, and the multiplication assertion gives ρ(g) = ρ(g^{k})ρ(g)^{−(k−1)} ∈H. Thus ρ(gU) ⊂H, and one can take q_{g} = 1.
There are finitely many cosets. Put Q = ∏_{gU}q_{g}. Then Q(1) ̸= 0 and fQ vanishes on ρ(Γ). For Y ∈𝔱(ρ(Γ)), differentiate:
Selectable PDF notation • transcription unchecked
0 = d(fQ)_{1}(Y ) = Q(1)df_{1}(Y ) + f(1)dQ_{1}(Y ).Since 1 ∈ ρ(U), f(1) = 0, and division by Q(1) proves Y ∈ 𝔱(ρ(U)). The reverse inclusion follows immediately from ρ(U) ⊂ρ(Γ).
Source: EtaleCohomology/DivisorCycles/ZclMulPf.lean, left\_mem and mul\_mem; TangentIndexDvd.lean, coset\_poly and TanIndex\_of.
Lemma 7.55 (Choosing a generic diagonal element in the closure). Let K be algebraically closed of char-acteristic zero, let X =∑_{i}e_{i}p_{i} ∈𝔱(S), and suppose (e_{i}) is lacunary in the additive sense used above. For every finite B ⊂K there are τ_{i} ̸= 0 such that t =∑_{i}τ_{i}p_{i} ∈Z(S) and
Selectable PDF notation • transcription unchecked
∏ )|O^{′}| ∏ )|O|
τ_{i} ̸= b τ_{i} (185)
i∈O i∈O^{′}for all disjoint nonempty sets O, O^{′} of indices and all b ∈B.
Proof. Consider the formal series E_{i} = exp(e_{i}t). Each is nonzero. For the finitely many choices of O, O^{′}, b, the polynomial
Selectable PDF notation • transcription unchecked
∏ )|O^{′}| ∏ )|O|
z_{i} −b z_{i}
i∈O i∈O^{′}is nonzero after substitution∑ z_{i} = E_{i}. Otherwise exp(At) −b exp(Ct) = 0, with A = |O^{′}|∑_{i∈O}e_{i}, C = |O|_{i∈O}′ e_{i}. Comparing constant and linear coefficients gives b = 1, A = C, contrary to additive lacunarity.
Here is the algebraic specialization step, including its field hypothesis. Let I be the kernel of K[z_{1}, . . . , z_{a}] → K[[t]], z_{i} →→ E_{i}, and let P be the product of the coordinates and all the displayed binomials. The quotient by I is a domain and P /∈I. Its localization at P is a nonzero finitely generated K-algebra. It has a maximal ideal, whose residue field is K by the following finite-generation argument. A field finitely generated as a K-algebra has a maximal algebraically independent subset t_{1}, . . . , t_{r} of its generators. Its remaining generators are algebraic, and, after clearing finitely many denominators, are integral over K[t_{1}, . . . , t_{r}, 1/f] for some nonzero f. The field is therefore integral over this ring. A subring over which a field is integral is itself a field: the integral equation for b^{−1}, multiplied by a suitable power of b ̸= 0, expresses b^{−1} in the subring. If r > 0, choose c ∈K so that t_{1} −c does not divide f; there are only
finitely many excluded c’s. Then t_{1} −c is a nonunit in that localized polynomial ring, a contradiction. Thus r = 0, and the residue field is a finite algebraic extension of the algebraically closed field K, hence equals
Selectable PDF notation • transcription unchecked
K.
The residue homomorphism gives τ ∈ K^{a} annihilating I with P(τ) ̸= 0. It gives all the desired nonvanishings. If f vanishes on S, the formal-exponential lemma shows that f(∑_{i}E_{i}p_{i}) = 0, since the powers of the orthogonal idempotents give exp(tX) = ∑_{i}E_{i}p_{i}. Therefore the polynomial f(∑_{i}z_{i}p_{i}) belongs to I, and f(t) = 0. This proves t ∈Z(S).
Source: EtaleCohomology/DivisorCycles/SpecPtPf.lean, SpecPt\_of; TorusPtPf.lean, E\_ne, entry\_exp andTorusPt\_of. The source implements specialization through the vanishing-ideal form of the Nullstellensatz.
Lemma 7.56 (Separating the cycle spectra). Suppose x is invertible and xp_{i} = p_{π(i)}x. There is a finite set B such that every diagonal t = ∑_{i}τ_{i}p_{i} satisfying (185) has the following property: xt has no common eigenvalue on the subspaces supported on two distinct cycles of π.
Proof. For a cycle O of size r, put T_{O} =∏_{i∈O}τ_{i}. Iterating xtp_{i} = τ_{i}p_{π(i)}x gives (xt)^{r}p_{i} = T_{O}x^{r}p_{i} for i ∈O. If a nonzero vector supported on this cycle has eigenvalue λ for xt, it consequently has eigenvalue β = λ^{r}/T_{O} ̸= 0 for x^{r}. For a second cycle O^{′} of size s with the same λ, write similarly λ^{s} = T_{O}′β^{′}. Eliminating λ yields
Selectable PDF notation • transcription unchecked
T_{O}^{s} = (β^{′r}/β^{s})T^{ r}_{O}′.Take B to contain these ratios for 1 ≤r, s ≤a and all nonzero characteristic roots β of x^{r}, β^{′} of x^{s}. This is finite. The last equality contradicts (185).
Source: EtaleCohomology/DivisorCycles/GenSpecProof.lean, ypow, eig and GenSpec\_of.
Lemma 7.57 (Enlarging to an algebra containing the group). Suppose A has scalar corners at the p_{i}’s and is normalized by ρ(Γ). Suppose
Selectable PDF notation • transcription unchecked
∃a ∈A : p_{i}aρ(g)p_{i} ̸= 0
=⇒ ∃w ∈A : p_{i}ρ(g) = p_{i}w (g ∈Γ, i).^{(186)}Then R = span_{K}{aρ(g) : a ∈ A, g ∈ Γ} is a subalgebra containing A and ρ(Γ), and still has scalar corners. If A preserves every ρ-invariant subspace, so does R.
Proof. The generator product is
Selectable PDF notation • transcription unchecked
(aρ(g))(bρ(h)) = a(ρ(g)bρ(g)^{−1})ρ(gh).Normalization puts the middle conjugate in A. Thus the span is closed under multiplication, and it contains the identity and both indicated sets.
For a generator aρ(g), its i-th diagonal corner is scalar if zero. Otherwise (186) supplies w ∈ A. Move a past ρ(g) by writing aρ(g) = ρ(g)(ρ(g)^{−1}aρ(g)). Its corner becomes a corner of the product w(ρ(g)^{−1}aρ(g)) ∈A, hence is scalar. Linearity gives scalar corners on the entire span. Normalization by the group is preserved on these generators and thus on the span. Finally, a ρ-invariant subspace is preserved first by ρ(g) and then by a; it is therefore preserved by every generator and its linear span.
Source: EtaleCohomology/DivisorCycles/BigCorner2.lean:11--133.
Lemma 7.58 (Semisimplicity excludes the nilpotent obstruction). Let ρ : Γ → GL(V ) be a semisimple representation on a finite-dimensional vector space over a field K. Let A ⊂ End_{K}(V ) be a subalgebra normalized by ρ(Γ), and suppose that A preserves every ρ-invariant subspace of V . If Y ∈A and every aY , a ∈A, is nilpotent, then Y = 0.
Proof. Put
S = {Z ∈A : aZ is nilpotent for every a ∈A}, U = span_{K}{Zv : Z ∈S, v ∈V }.
Conjugation by ρ(g) preserves S: after conjugation the product with a is conjugate to the product of Z with ρ(g)^{−1}aρ(g) ∈ A. Consequently U is ρ-invariant. Choose a ρ-invariant complement W, using semisim-plicity. It is preserved by A, by hypothesis. For Z ∈S and w ∈W, the vector Zw belongs both to U by definition and to W by preservation. Thus SW = 0.
Let J be the Jacobson radical of A. Every element of S lies in J: for every a ∈ A, nilpotence of aZ makes 1 + aZ invertible by the finite geometric sum. This is the radical criterion used by the source. Here the radical is nilpotent because A is finite-dimensional. For completeness, the descending sequence J^{r} stabilizes as vector spaces. For a stabilized term M = J^{r} = JM, choose a minimal finite set of left A-module generators. If it is nonempty, write its last generator as m_{s} =∑_{i}a_{i}m_{i} with a_{i} ∈J. Since 1 −a_{s} is a unit, this expresses m_{s} using the other generators, a contradiction. Hence J^{r}= 0.
Now U ⊂JV . Inductively, if U ⊂J^{k}V , decompose any v ∈V as u + w with u ∈U, w ∈W. For Z ∈ S, Zv = Zu ∈ J^{k+1}V . Taking spans gives U ⊂ J^{k+1}V for every k ≥ 1. Nilpotence of J forces U = 0. In particular Y V = 0, so Y = 0. The source implements the same induction with the subspace spanned by the images of each ideal power.
Source: EtaleCohomology/DivisorCycles/SemisPf.lean, img, img\_step, nil\_conj and Semis\_of.
Lemma 7.59 (Constructing the corner matrix units). Let p_{1}, . . . , p_{a} be nonzero pairwise orthogonal idem-potents summing to 1, in a subalgebra A ⊂ End_{K}(V ). Suppose p_{i}Ap_{i} = Kp_{i} and suppose A has the conclusion of Lemma 7.58. Then there are elements E_{ij} satisfying all the matrix-unit hypotheses of Proposi-tion 7.62 with R = A.
Proof. First take a nonzero Y ∈p_{i}Ap_{j}. The nilpotent obstruction property supplies b ∈A for which bY is not nilpotent. Set w = p_{j}bp_{i}. The two scalar-corner hypotheses give
Selectable PDF notation • transcription unchecked
Y w = cp_{i}, wY = dp_{j}.Associating Y wY in the two ways gives cY = dY , hence c = d, since Y ̸= 0. Moreover c ̸= 0: otherwise Y bY = Y wY = 0, so (bY )^{2}= 0, contrary to the choice of b. Therefore x = Y has a partner y = c^{−1}w ∈ p_{j}Ap_{i} with xy = p_{i}, yx = p_{j}.
Write i ∼j if such a pair (x, y) exists. The pair (p_{i}, p_{i}) proves reflexivity, interchange proves symmetry, and the products (xx^{′}, y^{′}y) prove transitivity. Choose one representative r(i) of each equivalence class, constant on that class. Choose partners
Selectable PDF notation • transcription unchecked
X_{i} ∈p_{i}Ap_{r(i)}, Y_{i} ∈p_{r(i)}Ap_{i}, X_{i}Y_{i} = p_{i}, Y_{i}X_{i} = p_{r(i)}.Define E_{ij} = X_{i}Y_{j} when i ∼ j, and E_{ij} = 0 otherwise. In the first case the inverse partner is E_{ji}, so E_{ij} ̸= 0, and E_{ii} = p_{i}. If i ∼j ∼k, then
Selectable PDF notation • transcription unchecked
E_{ij}E_{jk} = X_{i}(Y_{j}X_{j})Y_{k} = X_{i}p_{r(j)}Y_{k} = E_{ik} ̸= 0.The support identities p_{i}E_{ij}p_{j} = E_{ij} follow directly from the supports of X_{i}, Y_{j}.
It remains to prove that these elements span the whole corners, rather than merely give some matrix units. For a partner pair (x, y) between i, j and a ∈A, the scalar-corner hypothesis applied to ap_{j}y gives p_{i}ap_{j}y = cp_{i}. Multiplying on the right by x, and using yx = p_{j}, gives p_{i}ap_{j} = cx. Thus, if i ∼j, take x = E_{ij} to obtain p_{i}Ap_{j} = KE_{ij}. If i ∼∼j, any nonzero member of that corner would have a partner by the first paragraph and would contradict i ∼∼j. That corner is therefore zero, as required. This also proves symmetry of the nonzero pattern and every assertion in the proposition’s matrix-unit contract.
Source: EtaleCohomology/DivisorCycles/MatUnitsOf.lean, partner, oneDim, rel\_equiv and MatUnits\_of.
Lemma 7.60 (Matching two systems of scalar corners). Work in characteristic zero. Suppose A has the nilpotent-obstruction property, and both (p_{i})^{a}_{i=1} and (q_{i})^{a}_{i=1} are complete nonzero orthogonal idempotent systems with scalar corners in A. There are a unit v ∈A and a permutation π such that
Selectable PDF notation • transcription unchecked
vp_{i} = q_{π(i)}v.They can be chosen so that π(i) = i whenever q_{i}Ap_{i} ̸= 0.
Proof. Say q_{j} and p_{k} are isomorphic if there are α ∈ q_{j}Ap_{k}, β ∈ p_{k}Aq_{j} with αβ = q_{j}, βα = p_{k}. Any nonzero z ∈q_{j}Ap_{k} produces such a pair. Indeed the partner argument of Lemma 7.59 uses only the two sup-port identities and scalar corners, and applies equally when these idempotents come from different systems. Thus nonisomorphic cross corners are zero.
Let Q(k) be the set of j’s isomorphic to p_{k}. It is nonempty: otherwise every q_{j}p_{k} would vanish, contra-dicting ∑_{j}q_{j} = 1 and p_{k} ̸= 0. Two of the sets Q(k) are equal or disjoint, since composing the displayed partial inverses transports an isomorphism through a common q_{j}. These sets also cover the j’s, by the same argument with the two systems interchanged.
For a set S = Q(k_{0}) put P_{S} = {k : Q(k) = S}. Completeness of the systems and vanishing of the other cross corners give ∑
Selectable PDF notation • transcription unchecked
p_{k} =∑q_{j}.
k∈P_{S} j∈SEvery idempotent in these sums has the same trace t, since tr(αβ) = tr(βα) for each isomorphic pair. This trace is nonzero: a nonzero idempotent is the identity on its nonzero image, so its trace is its positive integer rank, nonzero in characteristic zero. Taking traces in the last equality yields |P_{S}| = |S|. Choose a bijection P_{S} →S fixing their intersection; such a bijection exists by matching their equal-size set differences. Combining these bijections gives π, and i ∈Q(i) ensures that i is fixed. ∑
Choose the partial inverse pair (α_{i}, β_{i}) between q_{π(i)} and p_{i}, and set v =∑_{i}α_{i}, v^{′} =_{i}β_{i}. Orthogonal supports kill every off-diagonal term in vv^{′} and v^{′}v; the diagonal terms sum to 1. Thus v^{′} = v^{−1}, and the support identities give vp_{i} = α_{i} = q_{π(i)}v. Finally, q_{i}Ap_{i} ̸= 0 means i ∈Q(i), giving the required fixed-index property.
Source: EtaleCohomology/DivisorCycles/ConjugateIdempotents.lean, key, compose, count and IdemConj\_of.
Lemma 7.61 (Factoring a corner-algebra unit). Under the matrix-unit hypotheses above, every invertible matrix v ∈A has a factorization
Selectable PDF notation • transcription unchecked
v = uD, D =∑d_{i}p_{i}, d_{i} ̸= 0,
iwhere u is a product of elements 1 + p_{i}ap_{j}, a ∈A, i ̸= j.
Proof. Partition the indices by E_{ij} ̸= 0. For each class T, its sum z_{T} =∑_{i∈T}p_{i} is a central idempotent in A: decompose each a into the corners p_{i}ap_{j}, and note that a nonzero corner has both indices in the same class. On this class, write z_{T} v =∑_{i,j∈T}M_{ij}E_{ij}. The matrix M is invertible. Otherwise choose a nonzero vector w with Mw = 0, let M^{′′} have every column equal to w, and put X =∑_{i,j∈T}M_{ij}^{′′}E_{ij}. Matrix-unit multiplication gives vX = 0, while a nonzero coordinate w_{i} gives p_{i}Xp_{i} = w_{i}p_{i} ̸= 0. This contradicts invertibility of v.
An invertible finite matrix is a product of transvections and an invertible diagonal matrix. The elimination proof used here is finite induction: choose a nonzero entry in the first column, move it to the first row, and use row additions to clear that column below it. The lower right block is invertible, so repeat there and clear the remaining entries in the first row. A row interchange is also allowed, since its two-coordinate matrix is
Selectable PDF notation • transcription unchecked
(0 1)
= T_{12}(1)T_{21}(−1)T_{12}(1) diag(−1, 1).
Selectable PDF notation • transcription unchecked
1 0
The empty-matrix case starts the induction.∑ Transport these factors into A using M →→ 1 − z_{T} +
_{i,j∈T}M_{ij}E_{ij}. The map preserves products. A transvection maps to 1 + cE_{ij} = 1 + p_{i}(cE_{ij})p_{j}; a diagonal matrix maps to ∑_{i}d_{i}p_{i}, with entries 1 outside T.
Finally these factorizations can be multiplied without losing the stated order. Diagonal factors normalize the elementary subgroup, because
Selectable PDF notation • transcription unchecked
D(1 + p_{i}ap_{j})D^{−1}= 1 + (d_{i}/d_{j})p_{i}ap_{j}.Their products remain invertible diagonal factors. The central block factors 1 −z_{T} + z_{T} v multiply to v, since their supports are disjoint and sum to 1. Moving all their diagonal factors to the right gives the asserted
Selectable PDF notation • transcription unchecked
uD.
Source: EtaleCohomology/DivisorCycles/UnitFactorizationTheorem.lean, S\_mul, block, main and UnitFactor\_of; the block proof uses diagonal\_transvection\_induction\_of\_det\_ne\_zero.
Proposition 7.62 (The finite matrix-unit part of extraction). Let V = K^{N}, let ρ : Γ → GL(V ), and let R ⊂ End_{K}(V ) contain its image. Suppose there are nonzero orthogonal idempotents p_{1}, . . . , p_{a} summing to 1, and elements E_{ij} ∈p_{i}Rp_{j} such that E_{ii} = p_{i}, nonvanishing of E_{ij} is symmetric in i, j, E_{ij}E_{jk} = E_{ik} ̸= 0 whenever the two factors are nonzero, and p_{i}Rp_{j} = KE_{ij}. Then there is a rank-a representation r embedded in ρ. The images of all intertwining maps r → ρ span V . If ρ is semisimple, then r is semisimple and ρ embeds in finitely many copies of r. With continuous matrix entries for ρ, the constructed r is continuous.
Proof. For x ∈R, define x_{ij} by p_{i}xp_{j} = x_{ij}E_{ij}, taking zero if E_{ij} = 0. The nonzero cases have unique coefficients. Inserting 1 =∑_{j}p_{j} between two factors gives
Selectable PDF notation • transcription unchecked
(xy)_{ik} =∑x_{ij}y_{jk}, (1)_{ij} = δ_{ij}.
jIf E_{ik} = 0, every term in the first sum is zero, because two nonzero consecutive matrix units would force E_{ik} ̸= 0. Otherwise the usual matrix-unit product gives the equality by comparing its nonzero E_{ik} coefficient. Thus r(g) = (ρ(g)_{ij}) is a representation; its inverse is r(g^{−1}).
The relation E_{ij} ̸= 0 is an equivalence relation on the indices. Choose one index b in each class. Choose v_{i} ̸= 0 in p_{i}V , set w_{0} =∑_{i}v_{i}, and choose a functional u_{i} with u_{i}(v_{i}) = 1. Define an N-by-a matrix A with column A_{i} = E_{i,b(i)}w_{0}, and define row i of B as u_{b(i)}E_{b(i),i}. Matrix-unit multiplication and orthogonality give BA = 1: its diagonal entry is u_{b(i)}p_{b(i)}w_{0} = 1, and the off-diagonal products vanish. Expanding xE_{j,b(j)} by the p_{i}’s shows xA = A(x_{ij}), hence A intertwines r with ρ and embeds it.
For any w ∈ V , the columns E_{i,b(i)}w give another intertwiner T_{w} : r → ρ. If an endomorphism Z annihilates every intertwiner, then ZT_{w} = 0 for every w. Thus ZE_{i,b(i)} = 0. Multiplying by E_{b(i),i} gives Zp_{i} = 0, so Z = 0. This is the spanning property of the intertwiners used by the source.
If ρ is semisimple, its embedded subrepresentation r is semisimple. We give the source’s finite-cover argument rather than infer a multiplicity bound. Let ℐ be the span of composites T^{′}T, where T : V → r, T^{′} : r → V intertwine. Start with its idempotent E = 0. If ker E ̸= 0, put P = 1 −E. The spanning property gives an intertwiner T^{′} with PT^{′}̸= 0. Split its kernel in r and split its image in V , using semisimplicity. Inverting PT^{′} on these complements gives an intertwiner T : V → r whose composite PT^{′}T is the projection onto that nonzero image. Consequently H = PT^{′}TP is an idempotent in ℐ, with EH = HE = 0, and E + H has strictly smaller kernel. Repeating decreases a nonnegative dimension, so eventually 1 ∈ℐ. Express it as a finite sum 1 = ∑_{j}T^{ ′}_{j}T_{j}, absorbing coefficients into T^{ ′}_{j}. The map v →→(T_{j}v)_{j} embeds V into r^{⊕q}, with left inverse (v_{j})_{j} →→∑_{j}T^{ ′}_{j}v_{j}. Finally r(g) = Bρ(g)A, so its matrix entries are continuous when those of ρ are continuous; inversion follows by g →→g^{−1}.
Source: EtaleCohomology/DivisorCycles/MinCongr.lean, coefficient multiplication, explicit A,B and Rep_of; FinEmbedPf.lean, projection step and finite-kernel descent.
Proposition 7.63 (The key corner implication and a simple-spectrum point). Let K be algebraically closed of characteristic zero, let ρ be semisimple, and suppose every ρ(g) has annihilator degree at most a. Suppose the tangent Lie algebra contains a matrix with a lacunary eigenvalues. Let (p_{i}) be its spectral projectors and A its generated Lie envelope. Then (186) holds. Moreover for every normal finite-index U ⊂ Γ and every g ∈Γ there is m ∈Z(ρ(U)) such that ρ(g)m has at least a distinct eigenvalues.
Proof. The tangent-degree lemma supplies the degree bound on the whole Lie algebra. The spectral, grading and diagonal-corner lemmas therefore apply, giving scalar corners in A, with every off-diagonal corner of A lying in the tangent Lie algebra. The envelope lemma gives normalization and preservation of invariant subspaces; semisimplicity gives the nilpotent-obstruction property. Thus the matching and unit-factorization lemmas apply to A.
Fix g, U, put R = ρ(g), and compare p_{i} with q_{i} = R^{−1}p_{i}R. Conjugation preserves the envelope and scalar corners. Matching gives v ∈A, invertible, and π with vp_{j} = q_{π(j)}v; it fixes i whenever q_{i}Ap_{i} ̸= 0. Factor v = uD as above. Since D commutes with all p_{j},
Selectable PDF notation • transcription unchecked
Rup_{j} = p_{π(j)}Ru.Each elementary factor 1 + E of u has E = p_{j}ap_{k}, j ̸= k, in the tangent Lie algebra, and E^{2}= 0. Finite-index tangent equality puts E in 𝔱(ρ(U)). Its formal exponential is exactly 1+tE; formal vanishing of every equation is then polynomial vanishing, so evaluation at t = 1 proves 1 + E ∈ Z(ρ(U)). Multiplicative closure gives u ∈Z(ρ(U)).
Apply the cycle-separation lemma to Ru, obtaining its finite exceptional set B. The diagonal special-ization lemma, again using finite-index tangent equality, supplies t =∑_{i}τ_{i}p_{i} in Z(ρ(U)) avoiding B. Put y = Rut. It is invertible and permutes the blocks by π. It belongs to Z(ρ(Γ)), so its annihilator degree is at most a. Distinct cycles have disjoint spectra. Lemma 7.49 therefore gives a distinct eigenvalues of y and, for every fixed index of π, yp_{i} = c_{i}p_{i}. The desired first conclusion uses m = ut ∈Z(ρ(U)).
For the corner implication, suppose p_{i}aRp_{i} ̸= 0 for some a ∈A. Conjugating a gives a^{′}= R^{−1}aR ∈A, and
Selectable PDF notation • transcription unchecked
q_{i}a^{′}p_{i} = R^{−1}p_{i}aRp_{i} ̸= 0.Thus π(i) = i. Since y preserves this block, the preceding scalar identity also reads p_{i}y = c_{i}p_{i}. The inverse of v belongs to A: its characteristic polynomial has nonzero constant coefficient, and Cayley–Hamilton
expresses v^{−1} as a polynomial in v. Consequently
Selectable PDF notation • transcription unchecked
(ut)^{−1}= t^{−1}Dv^{−1}∈A, t^{−1}=∑τ_{j}^{−1}p_{j}.
jIt follows that p_{i}R = p_{i}y(ut)^{−1}= p_{i}w, where w = c_{i}(ut)^{−1}∈A. This is precisely (186).
Source: EtaleCohomology/DivisorCycles/SumMulSum.lean, sys\_conj, tv\_zcl and Key\_of.
Lemma 7.64 (Descending the lacunary matrix to the coefficient field). If the lacunary tangent matrix and its a eigenvalues are initially defined over an extension of an algebraically closed coefficient field K, they can be replaced by such a matrix and eigenvalues over K itself.
Proof. Use variables for its N^{2} matrix entries and its a eigenvalues, and take the kernel of their evaluation in the extension field. Every equation df_{1}(X) = 0 defining the original tangent condition belongs to this kernel; it is linear in the matrix variables. So does each equation det(e_{i}1 −X) = 0. There are only finitely many nonvanishing conditions expressing lacunarity, namely
Selectable PDF notation • transcription unchecked
|B|∑e_{i} −|A|∑e_{i} ̸= 0
i∈A i∈Bfor disjoint nonempty subsets A, B of the index set. Localize the kernel quotient at their product and take a K-valued point, by the specialization argument in Lemma 7.55. All kernel equations, including the possibly infinite family of tangent equations, are preserved by this point; the finite product remains nonzero. The resulting matrix has exactly the required tangent, root and lacunarity properties over K.
Source: EtaleCohomology/DivisorCycles/SpecLiePf.lean, SpecLie\_of.
Lemma 7.65 (Simple spectrum occurs in every finite-index coset). For the representation r constructed above, each coset gU of a normal finite-index subgroup contains an element h for which charpol(r(h)) is separable.
Proof. Keep A, B from the finite matrix realization, so BA = 1 and r(h) = Bρ(h)A. Put ϕ(x) = Bρ(g)xA. By Proposition 7.63 choose m ∈ Z(ρ(U)) such that y = ρ(g)m has a distinct eigenvalues. For every intertwiner T : r →ρ, the equation
Selectable PDF notation • transcription unchecked
ρ(g)xT = Tϕ(x)
holds at x = ρ(u), u ∈U. Its entries are linear polynomials, so it holds at m as well. Let q = charpol(ϕ(m)). Multiplying the equation repeatedly gives q(y)T = Tq(ϕ(m)) = 0. The intertwiner-spanning property forces q(y) = 0. Each of the a distinct eigenvalues of y is then a root of the monic degree-a polynomial q: apply q(y) to a corresponding nonzero eigenvector. Hence q is separable.
The resultant of a matrix’s characteristic polynomial and its derivative is a polynomial in its entries, nonzero exactly for separable characteristic polynomial. If every r(gu) had inseparable characteristic poly-nomial, this polynomial composed with ϕ would vanish on ρ(U), hence at m, contradicting separability of q. This proves the claim. The source’s extra multiplication by B when encoding the intertwiner identity is harmless, because multiplying back by A uses BA = 1.
Source: EtaleCohomology/DivisorCycles/DenseAlg2.lean, aeval\_intertwine, sep\_of\_roots and DenseAlg\_of; SepPolyProof.lean, SepPoly\_of.
Lemma 7.66 (From cosets to an open dense separability locus). Suppose Γ is a compact topological group, K a Hausdorff topological field of characteristic zero, and r continuous. If the conclusion of the preceding lemma holds, its separability locus is open and dense in Γ.
Proof. Openness follows by continuity of the resultant polynomial. To prove density, suppose a nonempty open set O lies in its zero set. Compactness gives finitely many left translates of O covering the group, or equivalently finitely many a’s such that each x has ax ∈O.
Equip Γ temporarily with the topology induced by the evaluation-kernel map into the prime spectrum of the polynomial ring in the entries of r. This topology is Noetherian, since that polynomial ring is Noetherian. Polynomial zero sets are closed in it. Left translations are homeomorphisms: pulling back a polynomial uses the linear substitution X →→r(a)X. Let X be an irreducible component containing 1. Its translates are irreducible components, of which there are finitely many. The stabilizer of X therefore has finite index; let U be its normal core, still of finite index.
Let Z be the polynomial zero set of the separability resultant. The finite translates a^{−1}Z cover Γ, because they already cover it with O in place of Z. Irreducibility gives X ⊂a^{−1}Z for one of them. For u ∈U, the equality uX = X and 1 ∈X imply u ∈X, hence au ∈Z. Thus an entire coset of a normal finite-index subgroup lies in Z, contrary to the preceding lemma. The original separability locus is consequently dense. Only the finite-cover step used the original compact topology.
Source: EtaleCohomology/DivisorCycles/FinDensePf.lean, FinDense\_of; AutomorphicArithmetic/IntegralCasts/CapelliIdentities/CosetOfCover.lean, private finite\_cover, coset\_of\_cover and main.
Proposition 7.67 (The algebraic divisor extraction). Let Γ be a compact topological group, K an alge-braically closed Hausdorff topological field of characteristic zero, and ρ : Γ → GL_{N}(K) continuous and semisimple. Suppose S ⊂Γ is dense and each s ∈S has a monic degree-a polynomial P_{s} annihilating ρ(s). Suppose a lacunary a-eigenvalue tangent matrix exists, possibly over an extension of K. Finally suppose ψ : Γ →K^{×} is continuous and ψ(s) = (−1)^{a}P_{s}(0) for s ∈S. Then there is one continuous semisimple rank-a representation r, with
r ⊂ρ ⊂r^{⊕q}, det r = ψ,
whose separability locus is open dense. Its characteristic polynomial annihilates ρ(g) for every g. For s ∈S, charpol(r(s)) = P_{s} whenever either that characteristic polynomial is separable or the root multiset of P_{s} has the source’s multiplicative-genericity property.
Proof. The minor equations for annihilator degree are continuous functions of g; their vanishing on S ex-tends to all Γ. Descend the lacunary matrix to K by tangent specialization. The tangent, spectral and corner lemmas construct the envelope with scalar corners. The key proposition and the enlargement lemma give R containing ρ(Γ) with the same scalar corners, normalization and invariant-space preservation. Semisimplic-ity again supplies its nilpotent-obstruction property. Matrix-unit construction and finite matrix realization now give r ⊂ ρ ⊂ r^{⊕q}, semisimplicity and continuity. The two preceding lemmas give its open dense separability locus D.
For any polynomial q, intertwining gives q(ρ(g))A = Aq(r(g)). Since BA = 1, q(ρ(g)) = 0 implies q(r(g)) = 0. Conversely, if q(r(g)) = 0, then q(ρ(g))T = 0 for every intertwiner T, so the spanning property gives q(ρ(g)) = 0. In particular Cayley–Hamilton for r(g) proves the asserted annihilator, and each P_{s} annihilates r(s). Polynomial pinning gives charpol(r(s)) = P_{s} on S ∩D. This intersection is dense because S is dense and D is open dense. There det r(s) = (−1)^{a}P_{s}(0) = ψ(s). Continuity and Hausdorffness extend this equality to every g.
For the other pinning case, let M be the characteristic-root multiset of r(s). Every member is a root of P_{s}, because P_{s} annihilates its eigenvectors. Both multisets have cardinality a, and their products agree by the
determinant equality. The precise genericity condition is that no different cardinality-a multiset supported on the roots of P_{s} has that same product. It forces M = P_{s}.roots. Factoring both monic polynomials into their linear factors proves the last assertion.
Source: EtaleCohomology/DivisorCycles/EntriesCont.lean, DegAll\_of; FinSubOf.lean, FinSub\_of; FinPolyPf.lean, FinPoly\_of; DivPlus.lean, plus\_of; Imports.lean, goal\_of.
Proposition 7.68 (The required Tate-twist invariant vanishing). Let K/ℚ_{ℓ} be finite and let C_{ℓ} be the comple-tion of an algebraic closure, with its usual Galois action and cyclotomic character χ_{cyc}. For every integer
Selectable PDF notation • transcription unchecked
m ̸= 0,
g(x) = χ_{cyc}(g)^{m}x (g ∈G_{K}) =⇒ x = 0.This also applies to the source’s P.χ_{Q} on every finite local extension.
Proof. Tate’s original Theorem 2 in Section 3, printed page 177, proves the vanishing of invariants of C(η) when the extension cut out by a continuous character η : G_{K} →K^{×} is, after a finite intermediate extension, totally ramified with group ℤ_{ℓ} [8]. Apply it to η = χ^{−}_{cyc}^{m}, so that the displayed eigenvector is an invariant of
Selectable PDF notation • transcription unchecked
C(η).
We check that ramification hypothesis. Over ℚ_{ℓ}, the cyclotomic extensions are totally ramified, as fol-lows from the Eisenstein polynomial of a primitive ℓ^{r}-th root minus 1. The inertia subgroup of K has finite index in the inertia subgroup of ℚ_{ℓ}; therefore its cyclotomic image is open in ℤ^{×}_{ℓ}. After passing to a finite intermediate extension of the field cut out by η, its Galois image is a torsion-free principal-unit subgroup contained in the inertia image. It is ℤ_{ℓ}, and the remaining extension is totally ramified. A nonzero integer power has zero kernel on that torsion-free group, so this works for every m ̸= 0, including negative m and ℓ= 2 (where one uses 1 + 4ℤ_{2}).
For the source normalization, P.χ_{Q}(g) has norm one and hence comes from a unique unit of ℤ_{ℓ}, denoted chiZ(P)(g). Its residue modulo ℓ^{r} describes the action of g on every ℓ^{r}-th root of unity: the defining ap-proximation in P.χ_{Q} makes the exponents congruent modulo ℓ^{r}. A primitive root identifies this residue with that of the ordinary cyclotomic character. Equality at every r gives equality of the two ℤ_{ℓ} units, hence of their ℚ_{ℓ} images. The action on C_{ℓ} is the continuous extension of the action on its dense algebraic subfield. Thus both the character and action agree with Tate’s statement; arbitrary choices of the tied period data do not change this invariant calculation.
Source: AbelianVarietyTheory/Proofs/InertiaEigenvalues/ TateNoTwistedInvariantsComplexPadic.lean, chiZ, chiZ\_coe, chiZ\_cyc and tate\_chiQ.
Lemma 7.69 (Taking symbols of a filtered de Rham frame). Let B be a filtered field with Fil^{j}B = t^{j}Fil^{0}B, and let θ : Fil^{0}B →C be a surjective ring map with kernel Fil^{1}B. Let a common coefficient field F embed in Fil^{0}B and in C, compatibly with θ. Suppose G fixes F, preserves the filtration, and acts equivariantly on θ. Let χ : G →F^{×} satisfy g(t) = χ(g)t, and let ρ : G → GL_{N}(F). Suppose a basis u_{i} of B^{N} satisfies
u_{i} ∈(Fil^{ki}B)^{N}, ρ(g)g(u_{i}) = u_{i},
and is a filtered basis: for every y ∈(Fil^{j}B)^{N}, its coordinates y =∑_{i}f_{i}u_{i} satisfy f_{i} ∈ Fil^{j−ki}B. Then the columns θ(t^{−ki}u_{i}) form an invertible matrix P over C and
ρ(g)g(P) = P diag(χ(g)^{−ki}).
Proof. All entries of t^{−ki}u_{i} are in Fil^{0}B, so the columns are defined. If Pc = 0, lift each c_{i} to γ_{i} ∈ Fil^{0}B. Then z = ∑_{i}γ_{i}t^{−ki}u_{i} has all its symbols zero, hence lies in (Fil^{1}B)^{N}. Its filtered-basis coordinates f_{i}
belong to Fil^{1−ki}B. Uniqueness of basis coordinates gives f_{i} = γ_{i}t^{−ki}. Multiplying by t^{ki} gives γ_{i} ∈ Fil^{1}B, so c_{i} = 0. Thus P has zero kernel and is invertible.
Finally ρ(g)g(t^{−ki}u_{i}) = χ(g)^{−ki}t^{−ki}u_{i}. Apply θ, which respects the indicated Galois action and coeffi-cient embedding. This gives the formula column by column. Taking χ^{−1} for the character in Proposition 7.71 retains the integers k_{i} themselves as the cocharacter exponents.
Source: AbelianVarietyTheory/Proofs/InertiaEigenvalues/ HodgeTateSymbolMatrixIsUnitDet.lean, colHT\_isUnit\_det; HodgeTateSymbolMatrixEquivariant.lean, colHT\_equivariant.
For the source’s de Rham comparison, the filtered basis is obtained by choosing a basis adapted to the finite filtration on the specified embedding component of D_{dR}. Lift bases of its successive graded quotients, starting from the deepest nonzero filtration step. Their union is a basis, and strictness of the filtered compari-son isomorphism gives exactly the coordinate filtration in the lemma. The number of basis vectors of degree j is the dimension of that graded quotient, hence the required multiplicity of j in the jump multiset. These are the finite-dimensional steps of the source’s exists_dR_frame construction. The existence and strictness of the comparison isomorphism for every tied model and bridge remain part of the period comparison input isolated below.
Source: AbelianVarietyTheory/Proofs/InertiaEigenvalues/ DeRhamExistsHodgeTateEigenbasis.lean, exists\_dR\_frame and htc\_basis.
Lemma 7.70 (Separating the semilinear weights). Let a group G act on a field C, and let χ : G →C^{×} be a character. Suppose
Selectable PDF notation • transcription unchecked
m ̸= 0, g(c) = χ(g)^{m}c (g ∈G) =⇒ c = 0. (187)Give the coordinates of a finite-support vector space integer weights w(j). If a C-subspace J is preserved by the semilinear maps (T_{g}ϕ)_{j} = χ(g)^{w(j)}g(ϕ_{j}), then the part of any ϕ ∈J of each individual weight also belongs to J.
Proof. Induct on support size. Normalize a nonzero coefficient ϕ_{j0} to 1, and put δ_{g} = χ(g)^{−w(j0)}T_{g}ϕ−ϕ ∈ J. Each δ_{g} has smaller support: its j_{0}-coordinate is zero and no new coordinate appears. If all δ_{g} = 0, then g(ϕ_{j}) = χ(g)^{w(j0)−w(j)}ϕ_{j}. Condition (187) forces every coefficient of a different weight to vanish, proving this case. Otherwise choose g, j_{1} with (δ_{g})_{j1} ̸= 0. The vector
Selectable PDF notation • transcription unchecked
ϕ^{′}= ϕ −(ϕ_{j1}/(δ_{g})_{j1})δ_{g}also has smaller support. Induction applies to both ϕ^{′} and δ_{g}, so it applies to their linear combination ϕ. Undo the initial scalar normalization to finish the proof.
Source: AbelianVarietyTheory/Proofs/InertiaEigenvalues/ TateWeightGraded.lean, tate\_weight\_graded.
Proposition 7.71 (The cocharacter lies in the polynomial closure). Assume (187). Suppose every entry of ρ(h) is fixed by G and an invertible matrix P satisfies
ρ(g)g(P) = P diag(χ(g)^{ki}) (g ∈G) (188)
for integers k_{i}. Every polynomial vanishing on ρ(G) vanishes on P diag(t^{ki})P^{−1}, for all t ∈C^{×}. If C is infinite, then P diag(k_{i})P^{−1} is a tangent matrix, with characteristic roots k_{i}.
Proof. For a vanishing polynomial f, put p(Y ) = f(Y P^{−1}). Let J be the C-space of polynomials vanishing at every ρ(h)P. It contains p. Give a coordinate Y_{ij} weight k_{j}, so a monomial Y^{m} has weight ∑_{ij}m_{ij}k_{j}.
The coefficient operation c_{m} →→χ(g)∑^{ mijkj}g(c_{m}) preserves J. Indeed
ρ(h)P diag(χ(g)^{ki}) = g(ρ(hg)P),
using fixed matrix entries and (188); evaluation of the transformed polynomial gives the Galois transform of a value zero on the right. The preceding lemma therefore puts every weight part p_{d} in J. Evaluation at P = ρ(1)P gives p_{d}(P) = 0, hence
Selectable PDF notation • transcription unchecked
∑
f(P diag(t^{ki})P^{−1}) = t^{d}p_{d}(P) = 0.
Selectable PDF notation • transcription unchecked
d
This is a Laurent polynomial in t. Over an infinite field it is the zero Laurent polynomial: multiply by a high power of t and use the finite root bound. Evaluate it at the dual-number unit t = 1 + ϵ, ϵ^{2} = 0. For all integers k, including negative ones, (1 + ϵ)^{k} = 1 + kϵ, by multiplication and inversion. The matrix becomes 1 + ϵP diag(k_{i})P^{−1}. The coefficient of ϵ in f is precisely its derivative at 1 in that direction, by the polynomial product rule, and is zero. Conjugating the diagonal matrix gives its roots.
Source: AbelianVarietyTheory/Proofs/InertiaEigenvalues/ VanishAtHodgeTateCocharacter.lean, vanJ\_stableand vanish\_at\_cocharacter; WeilBound/CoreStatements/ PadicHodgeComparison/CocharacterLieDivisors/CocharacterLieAlgebra/PCocharToLie.lean, inLie\_of\_cochar and cocharToLie\_holds.
Proposition 7.72 (The period comparison supplies the lacunary tangent matrix). Let ρ : G_{E} → GL_{N}(ℚ_{ℓ}) be continuous. Suppose that for one place above ℓ, one permitted tied period model and bridge, and one induced embedding, its restriction is de Rham and its jump multiset contains a distinct lacunary integers. Then, over an algebraically closed extension of ℚ_{ℓ}, the polynomial tangent space of the global image contains a matrix having these integers as characteristic roots.
Proof. We first verify the established comparison being used. Fontaine’s Theorem 3.8(i), printed page 142, states that a de Rham representation is Hodge–Tate, that its canonical de Rham comparison is strictly com-patible with filtrations, and that its Hodge–Tate module is the associated graded de Rham module [41]. Its base field is complete discretely valued of characteristic zero with perfect residue field; the present base is finite over ℚ_{ℓ}. Dimension equality implies admissibility by Proposition 1.4.2(ii), since the de Rham period ring is Galois-regular by Proposition 3.6. We now check the source’s coefficients, model and embedding against this statement.
The source’s concrete period ring is the fraction field of the completion of W(𝒪_{C}^{♭}_{ℓ})[1/ℓ] at the kernel of θ, with its valuation filtration. This is the usual B_{dR} construction. The tied-model equivalence e_{D} respects the Galois action and every filtration step by its actD and fil fields. Its zD field fixes ℤ_{ℓ}, hence ℚ_{ℓ} after inversion. It also respects the algebraic coefficient embeddings: its om field gives a lift with the prescribed θ-residue, and that lift is unique for an algebraic element. Indeed its separable minimal polynomial has nonzero derivative at the residue; Hensel uniqueness in the complete discrete valuation ring identifies its lift. Thus 1 ⊗e_{D} identifies the two invariant spaces, coefficient actions and filtrations. The dimension condition called IsDR is consequently the ordinary one.
Explicitly, a model has coefficients in a finite extension L/ℚ_{ℓ}, and its dimension condition is
dim_{L} D = [K : ℚ_{ℓ}]N.
On forgetting L, this says dim_{K} D = [L : ℚ_{ℓ}]N, the dimension of the underlying ℚ_{ℓ}-representation. Hence the quoted admissibility criterion applies. Its continuity hypothesis also holds. The inverse of the bridge’s Galois identification is continuous: fixing a finite set of global algebraic elements pulls back to fixing their finite set of local algebraic images, an open subgroup. Compose with the given continuous ρ. Passing to its finite coefficient model preserves continuity, since L ↪→ℚ_{ℓ} has the induced topology.
Enlarge L if needed so that it contains all embeddings of K. This does not change jump multiplicities: ex-pand in an L-basis of the enlarged coefficient field; invariance and membership in each period filtration hold coordinate by coordinate. Let τ : K →L be the embedding in the jump datum. Extend it to an automorphism
s of the local algebraic closure, replace K by sK, and replace the local representation by h →→ρ(s^{−1}hs), retaining its coefficient matrices. On the period tensor, 1 ⊗s intertwines these actions and preserves the filtration; the τ-eigencondition becomes the inclusion sK ⊂ L. Invariant dimensions and each filtered dimension are unchanged. This is exactly the conjugation used in the source’s mem_closure_of_jumps.
In this inclusion component the strict comparison has rank∏ N. One can see the component directly from L⊗_{ℚℓ}K ≃_{σ:K→L}L: its idempotents project the invariant module to the embedding components, and eval-uation in the chosen embedding L ↪→B_{dR} projects the period tensor to B_{dR}^{N}. The inclusion component sur-vives and the other K-embedding components vanish. Restricting the comparison gives B_{dR} ⊗_{L} D_{τ} ≃B_{dR}^{N}, strictly filtered. Choose an adapted basis. Its degrees k_{i} have exactly the prescribed graded multiplicities, so their multiset is the given jump multiset; in particular that multiset has cardinality N. Lemma 7.69 now supplies the invertible symbol matrix P. Restrict the local Galois group to a finite extension containing the coefficient field, so its entries are fixed scalars, as required by the polynomial cocharacter proposition.
Proposition 7.68 supplies its remaining hypothesis, with the inverse cyclotomic character. Thus P diag(t^{ki})P^{−1} annihilates every polynomial equation of that restricted local image. Every equation of the global image is such an equation after restriction, so its derivative at 1 vanishes on X = P diag(k_{i})P^{−1}. This proves the required global tangent condition. Select the a distinct lacunary integers from the jump multiset; each is one of the k_{i}, hence a characteristic root of X. The ambient coefficient field can be taken to be C_{ℓ}, an algebraically closed extension of ℚ_{ℓ}.
Source: Formalism/LocalRingCardinality/AdicCompletionLocal/ LocTied.lean, fields eD, zD, om, actD, fil and t; InstIsAdicCompleteOCSpanSingletonSetCast.lean:34--61; AbelianVarietyTheory/Proofs/IwasawaDecomposition/ LocalIndepBridge.lean, conjugation and multV\_conj; Proofs/InertiaEigenvalues/MemClosureOfJumps.lean; Foundations/GaloisRepresentations/IntegralStructures/ ZetaCharacterDivisibility/HtCocharOf.lean:20--47.
The arithmetic extraction now has the following scope. Let ρ be continuous semisimple and have the sufficiently rich, lacunary Hodge–Tate data specified by HTRichEx. Suppose that on a dense set every ρ(g) is annihilated by a monic polynomial of degree a, whose constant term is (−1)^{a}ψ(g) for one continuous character ψ. Proposition 7.72 supplies the lacunary tangent matrix from the selected jumps. Therefore
Proposition 7.67 now constructs r ⊂ ρ ⊂ r^{⊕q}, its prescribed determinant and its open dense separability locus. The universal-comparison argument above then retains one such divisor across the separate geometric witnesses. The remaining local-jump and geometric realization inputs are separate obligations.
Proposition 7.73 (Local comparisons of the selected divisor). Suppose r ⊂ρ ⊂r^{⊕q}, both representations are continuous and semisimple, and r has rank a. Fix one of the permitted local period models and bridges. If ρ is crystalline there, then so is r, and its local Weil–Deligne characteristic polynomial on each Weil element divides that of ρ. If the jump multiset of ρ contains a specified set J of a distinct integers, the jump multiset of r at the same embedding is exactly J.
Proof. Choose a complement to the image of the embedding r →ρ, using semisimplicity. A basis formed by the image of the basis of r and a basis of its complement identifies ρ with r ⊕s. This is the source’s subSplit construction. The change of basis and its inverse have only finitely many algebraic entries. Adjoin them to a coefficient field for a local model of ρ. The conjugated matrices still have entries in that finite extension, and their two diagonal blocks give finite coefficient models for r, s. Their group laws and inverses follow by taking blocks in the group laws for ρ.
For either de Rham or crystalline periods, tensoring r⊕s with the period ring and taking invariants is the direct sum of the two invariant spaces: a vector is fixed precisely when its two coordinates are fixed. The filtration and Frobenius also act componentwise. Conjugation of a model induces an isomorphism on these spaces, and extending the coefficient field preserves them by expansion in a basis of that extension. Thus
all these calculations descend to the source’s existential finite models. Crystalline representations are closed under subobjects and are de Rham, by Fontaine’s theorem immediately following 5.1.7, printed page 158
[41]; the tied e_{B}, e_{D} and their action and Frobenius identities give the same coefficient-model transport as above. Hence r, s are crystalline and de Rham whenever ρ is crystalline. The direct-sum matrices for the induced Weil–Deligne action give
charpol(WD(ρ)(g)) = charpol(WD(r)(g)) charpol(WD(s)(g)),
which proves divisibility. Frobenius semisimplification preserves characteristic polynomials and this direct-sum decomposition.
Similarly the dimensions of each filtered embedding component add, so their successive differences add: J_{ρ} = J_{r} + J_{s} as multisets. The embedding ρ → r^{⊕q} also splits, since a finite sum of semisimple representations is semisimple. Applying the same calculation gives J_{ρ} ≤qJ_{r}. Therefore every member of the specified J occurs in J_{r}. Distinctness of J turns this support inclusion into multiset inclusion J ≤J_{r}. The filtered comparison proved above gives |J_{r}| = a = |J|, so equality holds. Every step was made for the fixed arbitrary model, bridge and induced embedding; this proves the same universally quantified local assertion used by the divisor comparison.
Source: IntegralStructures/ZetaCharacterDivisibility/ PSubSplit.lean, subSplit\_of; PLocConj.lean, ex-ists\_finExt\_gl and locConj\_of; PLocSum.lean, model\_split and locSum\_of; PSubLoc.lean, subLoc\_of; PCoverIn.lean, coverIn\_of; PJumpSeed.lean, jumpSeed\_of.
The other crystalline polynomial comparison is obtained by the same root-multiset argument already used in extraction. Suppose the geometric assertion gives charpol(WD(ρ)(F)) | P^{m} for a Frobenius element of degree one and the normalized Satake polynomial P. The proposition gives charpol(WD(r)(F)) | P^{m}. All roots of its degree-a polynomial therefore belong to the roots of P. The determinant comparison with the rank-one central-character representation makes their products equal. Under the stated multiplicative-genericity hypothesis, the two root multisets, and hence the two monic polynomials, agree. Both local pa-rameters are unramified: the first because r is crystalline, the second by its spherical Satake construction. Thus every Weil element acts, up to inertia, by a power of degree-one Frobenius, and its characteristic poly-nomial is obtained by taking that power of the same root multiset. Equality follows on every Weil element. This proves the polynomial part of phiSeed, subject precisely to the geometric divisibility, the rank-one determinant comparison and the stated genericity input.
Source: IntegralStructures/ZetaCharacterDivisibility/ PPhiSeed.lean, phiSeed\_of.
Lemma 7.74 (The usable Frobenius elements are dense). Suppose d > 0 and E contains an element x with x^{2}= −d. Fix a finite set X of finite places and an automorphic unitary representation U. Frobenius elements at places v satisfying all the following requirements are dense in G_{E}: they avoid X and ℓ, lie over different places of E^{+} from every member of X, are split in E/E^{+}, are spherical for U, and −d is a square in the q-adic field at their underlying rational prime q. Their inverses are dense as well.
Proof. There are only finitely many split places at which U is not spherical. To check this, choose a nonzero smooth automorphic vector and its open finite-adelic stabilizer. The restricted product topology makes this stabilizer contain each maximal compact local subgroup away from finitely many places. At a split such place the vector has a nonzero spherical local component and hence a Satake multiset. Add the finite exceptional set to X.
Set ∏ M = ℓ(2d) Norm(u).
Selectable PDF notation • transcription unchecked
u∈X
It is nonzero. Exclude its finitely many prime ideal factors and the finitely many primes ramified over ℚ. The degree-one Chebotarev density argument of Proposition 6.19, using Lemma 8.43, still gives a dense set of Frobenius elements outside these exclusions. A degree-one place v has residue field 𝔽_{q} and norm q. Since it is unramified, it must split in the quadratic extension E/E^{+}: an inert prime would have relative residue degree two, contrary to absolute residue degree one. Avoiding M excludes ℓ, all members of X, and every place over the same place of E^{+} as a member of X, since its rational prime would then divide that member’s norm.
The element x is integral because it satisfies T^{2}+ d. Its residue in 𝔽_{q} is an integer residue a with a^{2}+ d ≡0 (mod q). Since q ∤2d, one has 2a ≡≡0 (mod q). Hensel’s lemma lifts this simple root to ℤ_{q}, giving a square root of −d in ℚ_{q}. These are exactly the extra conditions in the source’s UsPl(d, v). Finally inversion is a homeomorphism of G_{E}, so it preserves density.
Source: AbelianVarietyTheory/Proofs/SatakeParameters/ AutomorphicLevelAwayNonsplit.lean:31--91; AbsNormMemOfUnderEq.lean, isSquare\_neg\_padic and usPl\_of\_absNorm\_prime; IntegralStructures/ZetaCharacterDivisibility/ PDenseUs.lean, denseUs\_of.
This density lemma is the precise adapter needed at the arithmetic extraction boundary. At each usable inverse Frobenius choose its geometric degree-a annihilating polynomial with the central-character constant term. The set on which at least one such polynomial exists is dense by the lemma. Choose one polynomial there, and assign an arbitrary value elsewhere; no continuity of that choice is required by Proposition 7.67. Global semisimplicity and Proposition 7.72 supply its other representation-theoretic hypotheses. The result-ing representation is continuous as a group representation because its matrix entries are continuous and its inverse entries are obtained by composition with g →→g^{−1}.
Source: IntegralStructures/ZetaCharacterDivisibility/ PDivSet.lean, divSet\_of.
Proposition 7.75 (The algebraic central character of the unitary representation). Let U ̸= 0 be an irreducible automorphic representation of the definite unitary group of rank a ≥1. The central map
Selectable PDF notation • transcription unchecked
ζ : 𝔸^{×}_{E}−→U(a)(𝔸_{E}+), z −−→(z/z^{c})1_{a}acts on U through an algebraic Hecke character χ, trivial on conjugation-fixed ideles. At a split spherical place with Satake multiset B_{v} = {α_{1}, . . . , α_{a}}, it is unramified and χ_{v}(ϖ_{v}) =∏_{i}α_{i}.
Proof. The scalar z/z^{c} has norm one because (z/z^{c})^{c}(z/z^{c}) = 1. Hence ζ lands in the unitary group. It is continuous, central, takes principal ideles to rational unitary matrices, and is the identity on conjugation-fixed ideles.
We justify scalar action using the source’s finite-dimensional argument. For a nonzero automorphic vector φ, its orbit under the compact infinite group spans a finite-dimensional space S. An open subgroup K of the finite adelic group fixes φ, hence all of S, since the finite and infinite factors commute. The space M of linear maps S →U intertwining the infinite action and having K-fixed images is finite-dimensional. Indeed the finite class set from Proposition 7.98 gives representatives g_{1}, . . . , g_{h}; evaluation T →→(v →→T(v)(g_{j}))_{j} embeds M into (S^{∗})^{⊕h}. Rational left invariance, infinite equivariance and right K-invariance prove this injectivity. The inclusion S ↪→U makes M ̸= 0.
A central element z acts on M by T →→ R(z)T. Over ℂ this endomorphism has an eigenvalue and a nonzero eigenvector. Some value of that map is a nonzero vector in the corresponding z-eigenspace of U. Since z is central, that eigenspace is invariant under the entire group; irreducibility makes it all of U. This proves scalar action, without applying a finite-dimensional Schur lemma to the whole automorphic representation.
Write R(ζ(z))φ = χ(z)φ. Choose g_{0} with φ(g_{0}) ̸= 0. Then
Selectable PDF notation • transcription unchecked
χ(z) = φ(g_{0}ζ(z))/φ(g_{0}).This proves continuity. Composition of right translations and φ ̸= 0 proves multiplicativity and χ(1) = 1; evaluating at z^{−1} proves nonvanishing. Rational left invariance and centrality make it trivial on principal ideles, so it is a Hecke character. The last assertion about conjugation-fixed ideles follows from ζ(z) = 1.
For algebraicity, write the character on an infinite exponential path at τ as exp(pz + qz̄), with p, q its two logarithmic exponents. Real z = t gives a conjugation-fixed idele, so this expression is constantly 1. Differentiation at t = 0 gives p + q = 0. For z = πi, its exponential is −1, again fixed by conjugation, so exp(2πip) = 1. Thus p ∈ℤ and q = −p. At every pair of embeddings the infinity character is consequently an integral monomial (z/̄z)^{p}, precisely the source’s algebraicity condition.
Finally choose a nonzero spherical Satake vector in the local component, and map it into U by the nonzero local intertwiner. Its image is nonzero, since the local representation is irreducible. A scalar local unit lies in GL_{a}(𝒪_{v}) and fixes the vector, proving χ_{v} unramified. The Hecke operator associated with the scalar ϖ_{v}1_{a} has one coset, because this matrix is central. Its Satake eigenvalue is the top elementary symmetric function,∏ _{i}α_{i}. Comparison with the central scalar action on that same nonzero vector gives the claimed value of
Selectable PDF notation • transcription unchecked
χ_{v}.Source: IntegralStructures/ZetaCharacterDivisibility/ PZetaEx.lean, zetaEx\_of; PSchurU.lean, schurU\_of; PCharOfZeta.lean, charOfZeta\_of; PAlgOfFix.lean, algOfFix\_of; PSatCent.lean, satCent\_of; PSmall.lean, centU\_of.
Selectable PDF notation • transcription unchecked
( )
Lemma 7.76 (The determinant normalization). Put c_{a} =^{a}_{2} and λ = χ| · |^{−ca}. The normalized Satake
roots at v are β_{i} = q_{v}^{(a−1)/2}α_{i}, and
Selectable PDF notation • transcription unchecked
λ_{v}(ϖ_{v}) = q_{v}^{ca}∏α_{i} =∏β_{i}.
i iThe character λ is algebraic.
Proof. The local absolute value of ϖ_{v} is q_{v}^{−1}, so the first equality follows from the proposition. The product of a scaling factors is q_{v}^{a(a−1)/2}= q_{v}^{ca}, giving the second equality. At infinity multiplication by this integral absolute-value power shifts both logarithmic exponents by the integer −c_{a}, so algebraicity is preserved.
Source: IntegralStructures/ZetaCharacterDivisibility/ PDetEx.lean, algSat\_prod\_eq and detEx\_of.
Lemma 7.77 (Global reciprocity in the row’s convention). For a number field E, the global reciprocity map with geometric Frobenius convention has all the quotient, kernel and local compatibility properties assumed in Proposition 7.78. Its local image has closure equal to the image of the decomposition group in G^{ab}_{E}.
Proof. We use the finite reciprocity and existence theorems in Milne, Class Field Theory, Chapter V, Propo-sition 5.2 and Theorems 5.3 and 5.5, pp. 178–179; local normalization is Chapter I, Theorem 1.1 and para-graphs 1.8–1.10, pp. 20–23. These apply to every number field and to every finite extension of ℚ_{ℓ}, respec-tively. They give compatible finite Artin maps, kill principal ideles, identify each finite abelian quotient with its norm quotient, and realize every open finite-index idele-class subgroup. The local convention there is arithmetic Frobenius. We invert the resulting maps to obtain the convention used here.[23]
Here are the inverse-limit and topology details. Write I = 𝔸^{×}_{E} and H = E^{×}I_{∞}^{0}. The product of the local Artin maps into a finite abelian extension has only finitely many nontrivial factors: outside the ramified places an idele component is a unit except at finitely many places. These products commute with restriction of finite extensions. They therefore define a continuous map into Gal(E^{ab}/E). Continuity follows by checking its finite coordinates. Each coordinate is onto, so the image is dense. Principal ideles are killed by reciprocity. The image of I_{∞}^{0} is trivial because a connected group has trivial image in a finite discrete group. Thus H is contained in the kernel.
Conversely, let x /∈ H. Take an identity neighborhood V with x /∈ HV . For some nonzero finite modulus 𝔪, the subgroup ∏
Selectable PDF notation • transcription unchecked
U_{𝔪} = I_{∞}^{0} × (1 + 𝔭^{m}_{v}^{v}) ×∏𝒪_{v}^{×}
v|𝔪 v∤𝔪satisfies HU_{𝔪} ⊂ HV ; one first chooses the finite compact-open part inside V , then absorbs the whole infinite identity component into H. The quotient I/(HU_{𝔪}) is the narrow ray class group for this modulus. Indeed weak approximation multiplies an idele by a principal element so that its components at v | 𝔪 lie in the specified congruence groups and all its real components are positive. The remaining valuations give a fractional ideal prime to 𝔪. Changing this principal adjustment changes that ideal by the ideal of a totally positive element congruent to one at the modulus, and these are exactly the changes killed in the ray class group. The converse uses the same adjustment to an idele representing such a principal ideal. This proves both the quotient identification and its finite index; it is the same narrow quotient as in Proposition 10.28.
The class-field existence theorem applied to this open finite-index subgroup of I/E^{×} gives a finite abelian extension whose Artin kernel in I is HU_{𝔪}. Since x does not belong to that kernel, it is not in the kernel of the inverse-limit map. This proves that the full kernel is precisely H.
The quotient I/H is compact. Choose finitely many fractional-ideal representatives for the ordinary ideal class group and fixed finite ideles with those valuation ideals. Multiplication by a principal idele puts every finite valuation ideal into one of these representatives. The remaining finite components lie in the compact group ∏_{v}𝒪_{v}^{×}. At infinity, division by I_{∞}^{0} leaves only the finitely many real signs. Thus a finite union of compact sets maps onto I/H. Its induced map into the profinite Galois group has compact, hence closed, image. The density already proved therefore gives surjectivity. The induced bijection from the compact Hausdorff quotient is a homeomorphism. Since the quotient homomorphism by a subgroup is open, the full Artin map is an open quotient map.
The topological abelianization of G_{E} is this Galois group: its kernel under restriction consists of the closure of commutators. One inclusion follows because every finite abelian quotient kills commutators; for the other, a nontrivial element of the profinite abelian quotient by their closure is separated from one by a finite abelian quotient. This identifies exactly the source’s target, with its quotient topology.
For completeness, the local unit and decomposition conditions also survive the inverse limit. Local
Selectable PDF notation • transcription unchecked
̂
reciprocity completes E_{v}^{×} = 𝒪_{v}^{×}× ϖ_{v}^{Z} to 𝒪_{v}^{×}× ϖ_{v}^{Z}. The second factor is the unramified quotient, with ϖ_{v} mapping to arithmetic Frobenius before our inversion. The first factor is therefore exactly the image of inertia. Its compactness makes its image closed after mapping to G^{ab}_{E}. The image of E_{v}^{×} is dense in the local abelian Galois group, so its image closure globally is the image of the decomposition group. Replacing a uniformizer by a unit multiple ranges over the entire corresponding inertia coset. Consequently its Artin images are precisely the images of inverse arithmetic Frobenius lifts, as required by the source’s quantified Frobenius condition.
Finally the source’s units of its adele ring are the usual ideles: both an element and its inverse are inte-gral at almost all finite places exactly when their valuations are zero there. Their topology induced by the element-and-inverse embedding is the restricted product topology with respect to 𝒪_{v}^{×}. Each prime-adic com-pletion identifies with E_{v} by the unique continuous extension of the dense field E, preserving its normalized valuation. Under these identifications principal, local and infinite ideles are the same elements used above. Thus the proof supplies the source’s actual quotient and local maps, including their signs.
Proposition 7.78 (Descent of an algebraic Hecke character). Let E be a number field, let ι : ℚ_{ℓ}≃ℂ, and let λ : 𝔸^{×}_{E}/E^{×} → ℂ^{×} be a continuous Hecke character. Suppose its infinity character on the identity component is the monomial with integer exponents n_{τ}, indexed by τ : E ↪→ℂ. Assume global reciprocity in the following precise form: a continuous open quotient
Art_{E} : 𝔸^{×}_{E}−→G^{ab}_{E}
has kernel the closure of the subgroup generated by E^{×} and the connected component at infinity; local units map onto inertia, and local uniformizers map to geometric Frobenius modulo inertia. There is then a unique
Selectable PDF notation • transcription unchecked
×
continuous character ψ_{λ,n} : G_{E} →ℚ_{ℓ} satisfying the formula below. At every unramified place v ∤ℓ,
ψ_{λ,n}(Frob^{−}_{v}^{1}) = ι^{−1}λ_{v}(ϖ_{v}), ψ_{λ,n}(I_{v}) = 1,
where Frob_{v} denotes arithmetic Frobenius.
Proof. For an idele x, let p_{∞}(x) be the product of its infinite components under the embeddings, raised to the powers n_{τ}. At a complex place this product uses both embeddings, and at a real place its single embedding. Let ∏
Selectable PDF notation • transcription unchecked
p_{ℓ}(x) = ∏ σ(x_{v})^{nισ|E}.
^{v|ℓ} σ:Ev↪→ℚ_{ℓ}Both are continuous multiplicative functions into their respective fields: only finitely many local embeddings occur in the second product. The notation n_{ισ|E} includes restriction along E →E_{v}.
Put θ = λ/p_{∞}. The asserted equality of infinity characters makes θ trivial on the connected component at infinity. This follows even if algebraicity was initially stated only for totally positive principal elements: weak approximation makes those elements dense in that infinite component, and the two continuous charac-ters agree there by passage to the closure.
The kernel of θ is open. Here is the source’s topological reason, which is necessary because ι is not a continuous map of valued fields. The finite idele group has arbitrarily small compact open subgroups. Choose one whose image under θ lies in {z : |z−1| < 1/2}. A subgroup of ℂ^{×} contained in this set is trivial: if all positive powers of z satisfy that bound, then ∑_{k<N}z^{k} is within N/2 of N, while (z −1) ∑_{k<N}z^{k} = z^{N}−1; letting N grow forces z = 1. Together with the open identity component of the infinite idele group this gives an identity neighborhood in the kernel. Therefore ι^{−1}θ is locally constant, hence continuous. In particular no continuity of ι^{−1} has been used.
The function
Selectable PDF notation • transcription unchecked
Ψ(x) = ι^{−1}(^{ λ}_{p∞}^{(}_{(}^{x}_{x}^{)}_{)})p_{ℓ}(x) (189)is a continuous idele character. For k ∈ E^{×}, restriction gives a bijection between all continuous local ℓ-adic embeddings and all embeddings E ↪→ℚ_{ℓ}: an embedding defines its unique ℓ-adic place and extends uniquely to its completion. Thus
Selectable PDF notation • transcription unchecked
p_{ℓ}(k) = ι^{−1} ∏τ(k)^{nτ})= ι^{−1}p_{∞}(k).
τSince λ(k) = 1, equation (189) is one on principal ideles. It is also one on the connected component at infinity, since θ is one there and every finite component is one. Its kernel is closed, the target being Hausdorff, so it contains the entire reciprocity kernel. Consequently it descends uniquely through Art_{E}. Quotient topology proves continuity of the descended character; composing with G_{E} → G^{ab}_{E} gives ψ_{λ,n}. The quotient’s surjectivity proves uniqueness for formula (189).
For a local idele at v ∤ℓ, both power factors are one. Hence its value is ι^{−1}λ_{v}(x). If λ_{v} is unramified, this is one on local units, proving the inertia assertion. Evaluating on a uniformizer and using the stated geometric reciprocity normalization gives the displayed inverse-Frobenius value.
Source: AbelianVarietyTheory/Proofs/IwasawaDecomposition/ IsOpenKerOfInfConn.lean:18--208; PublishedTheorems/AutomorphicForms2/ExistsUniqueHcharOfFiniteOrder.lean, IsGlobalArtin, IsLAdicOf, loc and loc\_away.
Proposition 7.79 (The above-ℓ rank-one comparison). For the character of Proposition 7.78, let v | ℓ and suppose λ_{v} is unramified. Its restriction to G_{Ev} is crystalline. At the embedding σ : E_{v} ↪→ℚ_{ℓ} its filtered jump is −n_{ισ|E}. Its Weil–Deligne representation has zero monodromy and is unramified, with value
Selectable PDF notation • transcription unchecked
( )−deg g
ι^{−1}λ_{v}(ϖ_{v}) on a Weil element g.
These statements hold for each tied period-ring model and each induced bridge appearing in the row.
Proof. On local reciprocity the character has the explicit value
Selectable PDF notation • transcription unchecked
δ(x) = ι^{−1}λ_{v}(x)∏σ(x)^{nισ|}E.
σIt takes values in a finite extension L/ℚ_{ℓ}: adjoin the finitely many images σ(E_{v}) and ι^{−1}λ_{v}(ϖ_{v}). The unramified factor is generated by the latter value, and the unit factor lies in the compositum of the former fields. The image of local reciprocity is dense in G^{ab}_{Ev}; continuity and closedness of the complete field L therefore give the same coefficient field for the local Galois character.
Apply Kedlaya–Pottharst–Xiao, Construction 6.2.4 and Example 6.2.6(3), with their geometric reci-procity convention in Hypothesis 6.0.1. A character whose unit restriction is ∏_{σ}σ^{kσ} is crystalline and has jumps −k_{σ}; its crystalline φ^{fv} scalar is [35]
Selectable PDF notation • transcription unchecked
α = δ(ϖ_{v})∏σ(ϖ_{v})^{−kσ}.
σHere k_{σ} = n_{ισ|E}, so cancellation gives α = ι^{−1}λ_{v}(ϖ_{v}). Their Definition 6.2.5 uses the actual filtered jumps, which fixes the minus sign in the assertion.
For a crystalline character the potentially semistable period module has trivial inertial descent and zero monodromy. The source defines its Weil action by the descent action multiplied by φ^{−fv deg g}. Substituting the scalar α gives α^{−deg g}, as asserted; in particular degree −1 has value α. Frobenius semisimplification changes nothing in dimension one. This also explains why geometric Frobenius on the Galois side and the displayed Weil formula agree.
For an arbitrary tied model, its given equivalences preserve multiplication, filtration, Frobenius and the Galois action. Applying these equivalences to the period invariant equations identifies the invariant spaces and all their filtered pieces; applying them to the Frobenius equations preserves α. The tensor scalar maps commute with the induced embedding by the bridge hypothesis. Coefficient extension preserves these one-dimensional equations. The same transport used in Proposition 7.72 therefore proves the claims in every permitted model, rather than only in a preferred realization of the rings.
Source: ThetaParseval/SplitLineSubtraction/PCharAbove.lean; AbelianVarietyTheory/Definitions/AutomorphicMembership/ FiniteDimensionalLetter.lean:219--227; FiniteDimensionalWdSpace.lean:103--149.
Lemma 7.80 (Faithfulness of the Hecke-character assignment). If two algebraic Hecke characters, with any permitted integer exponent data, give the same ℓ-adic character in Proposition 7.78, the two Hecke characters are equal.
Proof. Outside their finite ramification sets and the places above ℓ, equality of the Galois characters and the inverse-Frobenius formula give equality of the two uniformizer values. For their quotient η, every such local character is unramified and has uniformizer value one. It is therefore identically one on that local multiplicative group, since every nonzero element is a unit times an integral power of a uniformizer.
Let S contain the finitely many omitted finite places. A continuous idele character trivial on all local groups outside S factors through the product of the infinite multiplicative groups and the groups E_{v}^{×}, v ∈S.
To check this, first truncate the finitely many nonunit valuation components of an idele; then the finite partial products of its remaining unit components converge in the restricted product topology. Every partial product has character value one, so continuity gives value one for their limit. The kernel of the stated projection is therefore killed.
The image of E^{×} is dense in this finite product by weak approximation, with nonzero coordinates an open condition. The factored character is one on that image, because η is a Hecke character. Its kernel is closed, so it is one on the product. Consequently η = 1, proving equality. This is the source’s cofinite Hecke multiplicity-one argument; it requires no automorphic multiplicity-one theorem.
Source: ThetaParseval/SplitLineSubtraction/PCharInj.lean; IwahoriSphericalTransfer/ImportsFurther.lean:14--29; IwasawaDecomposition/DenseRangeUnits.lean:137--209.
Lemma 7.77 and the preceding three results give the source’s CharEx simultaneously at all finite places. Choose its unique character for each (λ, n) satisfying algebraicity, and choose the trivial character otherwise; every use of CharEx supplies the algebraicity proof. The one-dimensional characteristic polynomial is T − ι^{−1}λ_{v}(ϖ_{v}) away from ℓ, while the above-ℓ assertion has jump multiset {−n_{ισ|E}}. The injectivity clause applies to any two valid pairs. Combining this construction with Lemma 7.76 provides the exact determinant character used in detEx and phiSeed.
This closes the rank-one realization and determinant comparison input. It does not construct the higher-rank geometric Hecke eigenspace required for the ambient divisibility argument.