Section 9.2 · PDF pages 682–685
The finite-group projector in the cutoff
The notation transcription is provisional. Check it against the printed mathematics when citing or reusing a formula.
9.2 The finite-group projector in the cutoff
Let G be a finite group of order q and let 𝒳 be a finite set of distinct absolutely irreducible characters with values in a number field K. The actual group-algebra idempotent used in the source has coefficients
Selectable PDF notation • transcription unchecked
a(g) = ^{1}∑χ(1)χ(g^{−1}). (339)
^{q}χ∈𝒳It is important here to retain the character projector; it is not in general the averaging projector onto invariants.
Lemma 9.4. The coefficients in (339) satisfy
Selectable PDF notation • transcription unchecked
∑a(g)a(g^{−1}k) = a(k), (340)
g
a(hgh^{−1}) = a(g), (341)
qa(g) ∈𝒪_{K}, (342)
σ(a(g^{−1})) = σ(a(g)) (σ : K ↪→ℂ). (343)Source: Development/AbelianVarietyTheory/Proofs/DivisorProofs/GroupIdempotentCoefficients.lean:12--38; Formalism/LocalToGlobal/Divisibility/ProjLemmas.lean, idemOf_mul_self, idemOf_central and idemOf_integral.
Proof. At a complex embedding, let ρ, τ be irreducible representations of dimensions d_{ρ}, d_{τ}. For a linear map A between their spaces, averaging
Selectable PDF notation • transcription unchecked
𝒫(A) = ^{1}∑ρ(g)Aτ(g)^{−1}
^{q}g∈Ggives an intertwiner, by replacing g with hg. If the representations are inequivalent it is zero: the kernel and image of a nonzero intertwiner would make it an isomorphism by irreducibility. For one representation every intertwiner is scalar, since over ℂ it has an eigenvalue and the corresponding nonzero eigenspace is invariant. Taking the trace of the average then gives 𝒫(A) = (tr A/d_{ρ})id. Applying these identities to the matrix units is Schur’s matrix-coefficient orthogonality.
For a single character put a_{ρ}(g) = d_{ρ} tr ρ(g^{−1})/q. Expanding the two traces and applying the preceding matrix-unit identity gives
Selectable PDF notation • transcription unchecked
∑a_{ρ}(g)a_{τ}(g^{−1}k) ={^{aρ(k)ρ ≃τ,}
g^{0ρ ≃≃τ.}Indeed the unnormalized sum of the two traces is zero in the second case and is (q/d_{ρ}) tr ρ(k^{−1}) in the first; multiplying by d_{ρ}d_{τ}/q^{2} gives the displayed formula. Summing over the distinct characters proves (340). The embedding K ↪→ ℂ is injective, so this proves the coefficient identity over K. Characters are class functions, proving (341).
Every matrix ρ(g) has finite order. Its eigenvalues are roots of unity, so its trace is an algebraic integer. The number χ(1) is the integer d_{ρ}. Thus the numerator qa(g) in (339) is integral over ℤ and belongs to 𝒪_{K}. Finally average any positive Hermitian inner product over G; it remains positive and makes ρ unitary. Consequently χ(g^{−1}) = χ(g) at every complex embedding, while χ(1) is real. Substitution in (339) proves (343).
Proposition 9.5. Let ρ be a K-linear representation of G whose realized action at every place preserves the Hermitian inner product. Let η : G →K^{×} be a character. Put
Selectable PDF notation • transcription unchecked
c(g) = a(g)η(g)^{−1}, e =∑c(g)ρ(g).
gThen e^{2}= e and e is orthogonal at every place. If an 𝒪_{K}-lattice Λ is G-stable, and L is its intersection with the subspace on which a subgroup J acts through ηλ for a character λ of J, then qe(L) ⊆L.
Source: Development/AbelianVarietyTheory/Proofs/DivisorProofs/TwistedProjectorIdempotentOfGroupAlgebraIdem.lean:7--40; TwistedProjectorOrthogonalOfUnitaryAction.lean:9--52; CardSmulTwistedProjectorMemIntegralForms.lean:9--61.
Proof. The character law gives η(g)^{−1}η(g^{−1}k)^{−1} = η(k)^{−1}. Hence (340) also holds for c. Expanding e^{2} and changing variables from (g, h) to (g, k = gh) gives
Selectable PDF notation • transcription unchecked
e^{2}=∑ ∑c(g)c(g^{−1}k))ρ(k) = e.
k gThis is exactly the convolution reindexing in the source.
The values of η are roots of unity, because η(g)^{ord(g)}= 1. They are algebraic integers of complex norm one. Thus qc(g) ∈𝒪_{K}, and σ(c(g^{−1})) = σ(c(g)). Unitarity of ρ gives ρ(g)^{∗}= ρ(g^{−1}). Taking the adjoint of the sum for e and replacing g with g^{−1} therefore gives e^{∗}= e. It follows that
Selectable PDF notation • transcription unchecked
⟨ex, y −ey⟩= ⟨x, e(y −ey)⟩= 0,
using e^{2}= e. Expanding the norm square proves the Pythagorean identity required by the cutoff.
Since each ρ(g) preserves Λ and each qc(g) is integral, qe(Λ) ⊆ Λ. The function c is central: this holds for a, and a one-dimensional character is invariant under conjugation. Reindexing the projector sum by g →→hgh^{−1} shows that e commutes with ρ(h) for every h. Thus for x in the specified J-eigenspace,
Selectable PDF notation • transcription unchecked
ρ(h)e(x) = e(ρ(h)x) = η(h)λ(h)e(x) (h ∈J).
Both factors defining L are therefore preserved after multiplication by q. This proves qe(L) ⊆L.
For the source’s cutoff take G = p.P.Gp, 𝒳 = p.P.A3.A.X, and the character η from its form-space rep-resentation. The point conditions state that every member of 𝒳 is absolutely irreducible. Its lattice-stability and inner-product-invariance assertions give the other two representation hypotheses above. Consequently the algebraic and orthogonality clauses of CutAt hold with g = |G|. The remaining size clause is the separate inequality 2 log |G| ≤b_{0}; it must still be checked from the host parameters, not inferred from idempotence.
Lemma 9.6. Suppose the source’s finite group is a product of 2f groups GL_{r}(ℤ/q_{v}ℤ), with q_{v} > 0, and let ℒ ≥10, f ≥8, 0 ≤r ≤n_{V} . If
log q_{v} ≤9ℒ + 1, frn_{V} ℒ ≤X, X > 0, X^{3}≤b_{0},
then 2 log |G| ≤b_{0}.
Source: Development/AutomorphicArithmetic/ThetaCorrespondence/ThetaPairLifts/ThirdPlaceSeries/TwoLogCardLevelsLe.lean:17--94; Development/AbelianVarietyTheory/Proofs/DivisorProofs/LogGroupOrderBound.lean:14--26.
Proof. An invertible r × r matrix is determined by its r^{2} entries, so |GL_{r}(ℤ/q_{v}ℤ)| ≤q_{v}^{r2}. Multiply these bounds and take logarithms. All cardinalities are positive, and each q_{v} > 0, so
log |G| ≤2fr^{2}(9ℒ + 1).
If r = 0, the group has one element and the assertion follows from b_{0} ≥X^{3}> 0. Otherwise r ≥1, since it is an integer rank. Put Y = fr^{2}ℒ. Then Y ≥80 and Y ≤frn_{V} ℒ ≤X. As ℒ ≥10, we have
2 log |G| ≤4fr^{2}(9ℒ + 1) ≤40Y ≤40X ≤X^{3}≤b_{0};
the penultimate inequality uses X ≥80, hence X^{2}≥40. This is the source’s separate zero-rank case and its polynomial room estimate.
In the actual application ℒ = x^{ϵ}, b_{0} = exp(x^{ϵ}) and X = x^{A}. The source’s host bound has A = 22+|C_{t}| for a fixed family dependence exponent C_{t}. Its size inequalities give fr ≤x^{20}, n_{V} ≤2t ≤2x^{Ct} and ℒ ≤x for all sufficiently large x. For x ≥2 these imply
Selectable PDF notation • transcription unchecked
frn_{V} ℒ ≤x^{20}x^{1+Ct}x = x^{22+Ct}≤x^{A}.The same family hypotheses give f ≥8 eventually and q_{v} ≤ exp(9ℒ + 1). Since ϵ > 0, x^{3A} ≤ exp(x^{ϵ}) eventually: taking logarithms reduces this to 3A log x ≤x^{ϵ}. Choose the maximum of these finitely many
thresholds and the threshold for ℒ ≥10. Lemma 9.6 now proves the cutoff size clause simultaneously at every row and point in the family. This is the part of HostNAlt 3 actually consumed by covering_cut_of_local_traces_alt; its other clauses are not being substituted for an unproved bound on |G|. The source’s family size and prime bounds used here are the corresponding components of SizesAt and FamHyp, not claims for arbitrary hosts.
Source: Development/AbelianVarietyTheory/Proofs/DivisorProofs/HostBoundHolds.lean:15--33,37--109; Develop-ment/AutomorphicArithmetic/ThetaCorrespondence/ThetaPairLifts/ThirdPlaceSeries/HostN.lean:14--44; Cover-ingCutOfLocalTraces.lean:19--23.