Skip to content

Section 9.7 · PDF pages 694–695

The transference bound used to find a CM frame

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

9.7 The transference bound used to find a CM frame

Proposition 9.20 (The source’s covering radius, with its discriminant). Let E be a CM field of degree 2f, choose exactly one complex embedding σ_{i} from each conjugate pair, and let V = E^{r} have a nondegenerate hermitian form S. Suppose that invertible complex matrices Q_{i} identify it with the signature-(p, r −p) form at each chosen embedding. Let L ⊆V be a full finitely generated 𝒪_{E}-lattice, self-dual for S:

Selectable PDF notation • transcription unchecked
L = {y ∈V : S(y, L) ⊆𝒪_{E}}.

Assume λ_{∗}> 0 and that every nonzero v ∈L satisfies ‖Q_{i}σ_{i}(v)‖≥λ_{∗} for at least one i. Put D = | disc E| and

Selectable PDF notation • transcription unchecked
µ = 4(fr)^{3}D/λ_{∗}.

Then every w ∈ ∏_{i}ℂ^{r} has some l ∈ L with ∑_{i}‖Q_{i}σ_{i}(l) −w_{i}‖^{2} ≤ µ^{2}. There are also r vectors of L, linearly independent over E, each with ∑_{i}‖Q_{i}σ_{i}(v)‖^{2} ≤µ^{2}.

Source: Formalism/LocalToGlobal/ProjectiveLattices/ArchimedeanQuadraticData/FundamentalDomainD5.lean:17--333; ReciprocityTheory/ReductiveGroups/ArithmeticGeometry/FundamentalDomains/CoveringRadiusStatement.lean:7--21.

Proof. If r = 0 the ambient space is zero and both conclusions hold, with the empty basis. Suppose r > 0 and put N = 2fr ≥2. Embed V into ℝ^{N} by the real and imaginary coordinates of Φ(v) = (Q_{i}σ_{i}(v))_{i}. A ℤ-basis of L gives a real basis under this map. One can verify its independence directly using the nondegenerate trace pairing, as follows. Let Θ change the sign of the last r −p coordinates in every complex block. The hermitian form convention gives

Selectable PDF notation • transcription unchecked
⟨ΘΦ(y), Φ(v)⟩_{ℝ}N =^{1}_{2}Tr_{E/ℚ} S(y, v). (C.23)

The factor 1/2 occurs because only one embedding in each conjugate pair is included. The trace form is nondegenerate: if its pairing with every v vanishes, pairing with av for all a ∈ E and the nondegenerate field trace forces S(y, v) = 0 for all v, hence y = 0. Thus its Gram matrix in a ℚ-basis is invertible. Equation (C.23) then makes the real embedding matrix invertible as well. Consequently Γ = Φ(L) is a full Euclidean lattice.

We claim that every nonzero vector of the Euclidean dual lattice Γ^{∗} has length at least λ_{∗}/D. Indeed its pairings with the real images of the fixed ℤ-basis are integers. Inverting the rational trace Gram matrix in (C.23) produces y ∈V such that this dual vector is ΘΦ(y). It satisfies Tr S(y, v) ∈2ℤ ⊆ℤ for every v ∈L. Since av ∈L for every a ∈𝒪_{E}, it follows that S(y, v) lies in the inverse different, that is, the trace-dual of 𝒪_{E}. Multiplication by the integer D sends that trace-dual into 𝒪_{E}: take an integral basis of 𝒪_{E} and its integral trace Gram matrix; its determinant is the signed discriminant, and its inverse is its integral adjugate divided by that determinant. Hence S(Dy, L) ⊆𝒪_{E}, and self-duality gives Dy ∈L. The vector is nonzero, so Dy ̸= 0. The assumed lower bound gives

Selectable PDF notation • transcription unchecked
λ^{2}_{∗}≤∑‖Q_{i}σ_{i}(Dy)‖^{2}= D^{2}‖ΘΦ(y)‖^{2}.
i

This proves the claim, including the precise discriminant factor.

Use Banaszczyk’s Euclidean transference theorem, Theorem 2.1, p. 631 [59], in the form

Selectable PDF notation • transcription unchecked
λ_{N}(Γ)λ_{1}(Γ^{∗}) ≤N.

The dual here is for the ordinary real inner product, exactly as in (C.23); the theorem requires only a full Euclidean lattice, already checked. Thus λ_{N}(Γ) ≤ ND/λ_{∗}. Taking a positive slack ND/λ_{∗} as in the source’s shortDual_holds, choose N real-linearly independent lattice vectors of length at most 2ND/λ_{∗}. Their preimages span V over ℚ, and therefore over E. Extract r of them forming an E-basis. Each has the asserted bound, since

Selectable PDF notation • transcription unchecked
2ND/λ_{∗}≤N^{3}D/(2λ_{∗}) = µ (N ≥2).

For the covering assertion retain all N real independent vectors. They need not be a ℤ-basis of the full lattice. They are a real basis, so write any w as their real linear combination and round each coefficient to its nearest integer. The error has length at most one half the sum of their lengths, hence at most N^{2}D/λ_{∗}≤ N^{3}D/(2λ_{∗}) = µ. The rounded combination belongs to Γ even if the vectors span a proper sublattice. Squaring this norm proves the required sum-of-squares inequality. This is the source’s successive-minima and rounding argument; no claim that short independent vectors form an integral basis is used.