Section 9.1 · PDF pages 680–682
Lattices, heights and an integral orthogonal cut
The notation transcription is provisional. Check it against the printed mathematics when citing or reusing a formula.
9.1 Lattices, heights and an integral orthogonal cut
Let K be a totally complex number field, let n_{K} be the number of its infinite places, and let V be a finite-dimensional K-vector space. Choose one embedding σ : K → ℂ at each infinite place and a positive definite Hermitian norm at that place. A lattice means a finitely generated 𝒪_{K}-submodule of V . Its rank is the dimension of its K-span. Realizing the lattice in the direct sum of the complex spaces at these places gives a discrete integer lattice of real rank 2n_{K}r, where r is its K-rank. All covolumes are taken in its own real span.
The source’s height and normalized slope are
log h(L) h(L) = 2^{nKr}| disc K|^{−r/2} covol(L), rµ(L) = − . (335)
Selectable PDF notation • transcription unchecked
n_{K}At rank zero the lattice is zero, its height is one, and the second equation reads 0 = 0. These conventions are needed in the cut argument when the bad part vanishes.
Lemma 9.1. Let ϕ : V →W be K-linear, let L be a lattice, and set L_{0} = L∩ker ϕ, I = ϕ(L) and r = rk I. Suppose the operator norm at place v is at most c_{v} ≥0. For the quotient norm on L/L_{0},
Selectable PDF notation • transcription unchecked
h(I) ≤∏c^{2}_{v}^{r}h(L/L_{0}), h(L/L_{0}) = h(L)/h(L_{0}). (336)
vSource: Formalism/HarderNarasimhan/HarderNarasimhanProofs/HtImage.lean:289--476; GramBlock.lean:10--228; StatementPl.lean:108--119.
Proof. Write ρ for the real realization. The sublattice ρ(L_{0}) is saturated in ρ(L): if a nonzero integer multiple of x ∈ L lies in L_{0}, then that same multiple of ϕ(x) is zero, and the characteristic-zero vector space has no such torsion. Equivalently, the quotient is the torsion-free integer group I. A basis of L_{0} as an integer group can therefore be extended to an integer basis of L. Project the additional basis vectors orthogonally off ⟨ρ(L_{0})⟩_{ℝ}. They form a basis of the quotient lattice Q. In orthonormal coordinates adapted to this subspace, the basis matrix is block triangular. Taking its determinant gives
covol(L) = covol(L_{0}) covol(Q).
The factors in (335) multiply in the same way because the ranks add. This proves the equality in (336).
The realized map kills the real span of L_{0}, so it induces a linear map from the span of Q to the span of I and an integer-group isomorphism Q →I. Both spans split orthogonally by infinite places, with real dimension 2r in each part. To see the dimension assertion, the K-span of L/L_{0} has dimension r; at each chosen complex embedding it becomes a complex r-space, hence a real 2r-space. The realized map preserves these parts and has norm at most c_{v} on part v.
Choose an orthonormal basis within each part. If c_{v} = 0 on a nonzero part, the image determinant is zero and the inequality follows directly. Otherwise divide that part’s map by c_{v}. Each resulting part is a contraction. Its Gram matrix is positive semidefinite and its quadratic form is at most the identity quadratic form; the real Hermitian spectral theorem therefore puts all its eigenvalues in [0, 1]. Its determinant is at most one. Undoing the scaling and taking the square root of the Gram determinant bounds the volume multiplier by ∏_{v}c_{v}^{2r}. Changing from these orthonormal bases to an integer basis multiplies both volumes by the same absolute determinant. Thus covol(I) ≤∏_{v}c_{v}^{2r} covol(Q). The height normalizations for I and Q have the same rank r and the same field K, so their factors agree. This proves the inequality. This is the source’s placewise Gram-determinant argument, including its zero-scaling case.
Lemma 9.2. Let L ⊆M be lattices and e : V →V an idempotent which is orthogonal for every chosen norm. Let g ≥1 be an integer such that ge(L) ⊆L. Set L_{g} = L ∩ im e, L_{b} = L ∩ ker e, and d_{g} = rk L_{g}, d_{b} = rk L_{b}. Then
d_{g} + d_{b} = rk L, (337)
(d_{g} + d_{b})µ(L) ≤d_{g}µ(L_{g}) + d_{b}µ(L_{b}) + 2d_{b} log g. (338)
If the maximal slope of M is at most Q, then d_{b}µ(L_{b}) ≤d_{b}Q.
Source: Formalism/MeanHeightDescent/AverageDegreeBounds/GoodBadCut.lean:19--479; Development/AbelianVarietyTheory/Proofs/DivisorProofs/MajAtAddRight.lean:18--109,136--181.
Proof. Put S = KL. For x ∈L the vector gex belongs to L, hence ex ∈S; thus eS ⊆S. The spans of L_{g}, L_{b} are respectively eS, (1 −e)S: for example gex ∈L_{g} supplies the reverse inclusion for the first, and g(x −ex) ∈L_{b} does so for the second. These spaces are complementary because e^{2}= e. Their dimensions give (337).
Consider ϕ = g(1−e). Its kernel is im e. Its image lattice I = ϕ(L) lies in L_{b}, since both gx and gex lie in L and e(x −ex) = 0. It contains gL_{b}, so I and L_{b} have the same K-span and rank d_{b}. As integer lattices they have finite index; hence h(I) = [L_{b} : I]h(L_{b}) ≥h(L_{b}). At each infinite place 1 −e is an orthogonal projection, so has norm at most one. The norm of g(1 −e) is at most g. Applying Lemma 9.1 gives
Selectable PDF notation • transcription unchecked
h(L_{b}) ≤h(I) ≤g^{2nKdb}h(L)/h(L_{g}).All heights are positive. Taking logarithms, moving the two lattice terms to the other side and dividing by n_{K} > 0 yields (338) through (335).
The source initially states orthogonality as a Pythagorean identity for ex and y −ey. Expanding their Hermitian norm square makes the real part of their pairing zero. At a complex place choose a ∈ K with ℑσ(a) ̸= 0; such an a exists since the embedding is not real. Apply the same identity to ax to make the real part of σ(a)⟨ex, y −ey⟩ zero. The two equations make the imaginary part zero as well. Thus the subspaces really are orthogonal, verifying the projection-norm premise just used; no positivity has been deduced from an arbitrary idempotent.
Finally L_{b} ⊆ M. The defining characterization of maximal slope bounds the slope of every nonzero sublattice of M, so µ(L_{b}) ≤Q when d_{b} > 0. When d_{b} = 0, both products in the asserted bound are zero. This proves the remaining statement.
Lemma 9.3. Suppose d_{g}, d_{b} ≥0, d_{g} + d_{b} > 0, 0 ≤δ ≤1/12, and
Selectable PDF notation • transcription unchecked
(d_{g} + d_{b})µ ≤d_{g}µ_{g} + d_{b}(m + ℓ), d_{b} ≤3δ(d_{g} + d_{b}), m ≤Q, −Q ≤µ, 0 ≤ℓ≤Q.Then µ −12δQ ≤µ_{g}.
Source: Development/AbelianVarietyTheory/Proofs/DivisorProofs/MajAtAddRight.lean:112--134; GlobalTheory.CutNum.numeric.
Proof. We have Q ≥0 from 0 ≤ℓ≤Q. If d_{g} = 0, the rank bound gives d_{b} ≤d_{b}/4, contradicting d_{b} > 0. Thus d_{g} > 0. Rearranging the same bound and using 3δ ≤1/4 gives d_{b} ≤4δd_{g}. Subtract d_{b}µ from the first inequality. Then
Selectable PDF notation • transcription unchecked
d_{g}µ ≤d_{g}µ_{g} + d_{b}(m + ℓ−µ) ≤d_{g}µ_{g} + 3d_{b}Q ≤d_{g}µ_{g} + 12δd_{g}Q.Every multiplication preserves its inequality because d_{b}, Q, δ are nonnegative. Divide by d_{g} > 0.