Skip to content

Section 8.10 · PDF pages 486–491

The determinant and finite-orbit bounds

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

8.10 The determinant and finite-orbit bounds

We record the lattice normalisation needed for the integral homology estimate. This also separates its geo-metric content from the rank statement (208).

Lemma 8.22 (The convex-body argument). Let Λ be a full lattice in a real vector space of dimension d, of covolume v. A convex, symmetric measurable set of volume greater than 2^{d}v contains a nonzero lattice vector. For d ≥1, if a real symmetric matrix G satisfies z^{t}Gz ≥1 for every nonzero z ∈ℤ^{d}, then

G > 0, det G ≥d^{−d}.

Proof. If a measurable set B has volume greater than v, its translates by Λ cannot be disjoint: intersect them with a measurable fundamental parallelepiped P, and translate back. If disjoint, countable additivity would give ∑

vol(B) = vol((B + λ) ∩P) ≤ vol(P) = v.

Selectable PDF notation • transcription unchecked
λ∈Λ

Apply this to B =^{1}_{2}A. Two points of B whose difference is a nonzero lattice vector give a point of A, since b_{1} −b_{2} =^{1}_{2}(2b_{1}) +^{ 1}_{2}(−2b_{2}). This proves the convex-body assertion.

The quadratic form of G is nonnegative on rational vectors: multiply by a denominator to reduce to an integer vector. By continuity it is nonnegative on all real vectors. For δ > 0, G_{δ} = G + δI is positive definite and still has integral minimum at least one. The open ellipsoid x^{t}G_{δ}x < 1 contains no nonzero integer vector. Its volume is vol_{√}(B_{d})/^{√}det G_{δ}, so the first assertion bounds it by 2^{d}. The unit Euclidean ball contains the cube with side 2/ d; therefore det G_{δ} ≥d^{−d}. Let δ decrease to zero. Continuity of determinant gives the same positive lower bound for det G. A nonnegative symmetric matrix of positive determinant is positive definite, by diagonalising it.

Source: Mathlib/MeasureTheory/Group/GeometryOfNumbers.lean:52--83

Pinned revision: 0df444a360eaa60ab8c11dca51a86af692955474.

Source: Formalism/PeriodIntegrals/UnitaryGroupMassFormula/AdelicClassFiniteness/HmkFactor.lean

Lemma 8.23 (The Hermitian determinant in the source normalisation). Let L be totally complex, n = [L : ℚ], N ≥1, d = nN, D = |∆_{L}|. At each complex place choose one embedding ι_{w}, and for a Hermitian family h = (h_{w}) put

Selectable PDF notation • transcription unchecked
∑

q_{h}(y) = Re(y_{w}^{∗}h_{w}y_{w}), ℛ = {h : q_{h}(y) ≥1 for 0 ̸= y ∈O_{L}^{N}}.

Selectable PDF notation • transcription unchecked
w

Every h ∈ℛ is positive definite at every place. With P(h) =∏_{w} det h_{w} and I_{0} = D^{N}d^{d}, one has

Selectable PDF notation • transcription unchecked
1 ≤I_{0}P(h).

If an integral matrix A of nonzero determinant has all columns minimal for h, meaning q_{h}(A_{•j}) = 1, then

I := | Norm_{L/ℚ} det A| ≤I_{0}.

For H_{w} = ι_{w}(A)^{∗}h_{w}ι_{w}(A), its diagonal entries are at most one and det H_{w} ≥I_{0}^{−1}.

Proof. Choose an integral basis b_{1}, . . . , b_{n} and take the real Gram matrix of q_{h} in the basis indexed by (j, i). Its entries are ∑

Selectable PDF notation • transcription unchecked
( )

G(h)_{(a,i),(b,j)} = Re ι_{w}(b_{i})(h_{w})_{ab}ι_{w}(b_{j}) .

Selectable PDF notation • transcription unchecked
w

The mixed real embedding is an isomorphism, and z^{t}G(h)z = q_{h}(y) for the corresponding integral vector.

Lemma 8.22 shows positivity and det G(h) ≥d^{−d}. Positivity of this real form is precisely positivity of each Hermitian component.

The discriminant identity for the full embedding matrix, pairing each complex embedding with its con-jugate, gives

2^{d} det G(h) = D^{N}P(h)^{2}. (212)

One can see the factor 2^{d} by replacing each pair of complex rows with its real and imaginary rows: each pair contributes the squared modulus of a determinant 2i, and there are d/2 such pairs. The block diagonal Hermitian form has real determinant ∏_{w}(det h_{w})^{2}, which proves the displayed formula. It follows that P(h) ≥2^{d/2}d^{−d/2}D^{−N/2}, and hence I_{0}P(h) ≥1.

For the pulled-back form A^{∗}hA, every diagonal term is nonnegative and their sum over the places, for a fixed column, is one. Each diagonal entry is therefore at most one. Hadamard’s determinant inequality gives det H_{w} ≤1. Moreover A^{∗}hA ∈ℛ, and

Selectable PDF notation • transcription unchecked
P(A^{∗}hA) = IP(h).

The product on the left is at most one, so I ≤P(h)^{−1}≤I_{0}. Applying the already proved minimum bound to A^{∗}hA gives ∏_{w} det H_{w} ≥I_{0}^{−1}. Since all factors are positive and at most one, every factor is at least this product.

Source: Development/AutomorphicArithmetic/AutomorphicVarieties/HermiteNormalForms/Gramdet.lean:8--100; De-velopment/AutomorphicArithmetic/AutomorphicVarieties/HermiteNormalForms/Detlow.lean:10--81;Development/AutomorphicArithmetic/AutomorphicVarieties/HermiteNormalForms/Hadam.lean:6--37;Development/AutomorphicArithmetic/AutomorphicVarieties/HermiteNormalForms/DetPull.lean:9--86

Lemma 8.24 (Integral matrices of bounded norm determinant). For every positive integer J, the integral N-by-N matrices A with nonzero determinant and norm determinant at most J have at most J^{1+nN2} left GL_{N}(O_{L})-orbits.

Proof. Fix I = | Norm det A|. The integer I belongs to the ideal (det A): the additive group O_{L}/(det A) has order I and is annihilated by its order. Thus B = (I/ det A) adj(A) is integral and AB = BA = I 1. Use B mod I as a key. Since O_{L} is free of rank n over ℤ, there are I^{nN2} possible keys. If B −B^{′}= IX, and A^{′}, B^{′} satisfy the same scaled-inverse identities, put g = 1+A^{′}X, h = 1−AX. Then Ig = A^{′}B, Ih = AB^{′}. Multiplication and cancellation of the nonzero scalar I^{2} give gh = hg = 1, and IgA = IA^{′} gives gA = A^{′}. Equal keys therefore give the same left orbit. Sum over 1 ≤I ≤J: ∑I^{nN2}≤J^{1+nN2}.

Source: Development/AutomorphicArithmetic/AutomorphicVarieties/HermiteNormalForms/Scinv.lean:6--19; De-velopment/AutomorphicArithmetic/AutomorphicVarieties/HermiteNormalForms/Hnfeq.lean:6--41;Development/AutomorphicArithmetic/AutomorphicVarieties/HermiteNormalForms/Hnfone.lean:8--66;Development/AutomorphicArithmetic/AutomorphicVarieties/HermiteNormalForms/Hnf.lean:7--40

Lemma 8.25 (Vertices near a minimal matrix). For a fixed A as in Lemma 8.23, put R = N^{N}I_{0}^{2}. There are at most (2R + 1)^{d} integral vectors minimal for some h ∈ℛ for which all columns of A are minimal. There are at most

Selectable PDF notation • transcription unchecked
((2R + 1)^{d}+ 1)^{dN}

vertices of ℛ with these same column minima.

Proof. Take the scaled inverse B of the preceding proof. For a minimal vector y, fix a place w and put x = ι_{w}(By)/I. For H = H_{w}, we have x^{∗}Hx ≤ 1, the diagonal entries of H are at most one, and det H ≥I_{0}^{−1}. Cauchy–Schwarz for the positive form shows that all entries of H and of Hx have modulus at most one. Cramer’s rule, expanded into its N! determinant terms, therefore gives | det H| |x_{j}| ≤N! ≤N^{N}. Hence every conjugate of every coordinate of By has modulus at most II_{0}N^{N} ≤ R. Reduction modulo 2R+1 is injective on algebraic integers with this property: if the difference is (2R+1)z ̸= 0, all conjugates of the algebraic integer z have modulus at most 2R/(2R + 1) < 1, contradicting its nonzero integral norm. There are (2R + 1)^{n} residue classes, and N coordinates. The map y →→By is injective, because ABy = Iy. This proves the vector bound.

At a vertex, the active linear functionals h →→q_{h}(y), for minimal y, span the dual of the Hermitian form space; otherwise a nonzero direction killed by all active functionals would contradict the vertex condition. Select a basis among them. Its size is at most the real dimension of that space, which is at most dN. Their common value one determines the vertex uniquely. Thus a vertex is encoded injectively by a subset of at( ) most dN of the nearby vectors. For a set of size a, there are at most ∑^{dN}_{j=0}^{a}_{j} ≤(a + 1)^{dN} such subsets. Insert the vector bound.

Source: Development/AutomorphicArithmetic/AutomorphicVarieties/HermiteNormalForms/Cramer.lean:6--68; Development/AutomorphicArithmetic/AutomorphicVarieties/HermiteNormalForms/Nearcoord.lean:10--64;Development/AutomorphicArithmetic/AutomorphicVarieties/HermiteNormalForms/Nearvec.lean:9--38; Development/AutomorphicArithmetic/AutomorphicVarieties/HermiteNormalForms/Nearvx.lean:9--26

Proposition 8.26 (The normal-cone complex and its orbit bound). Let X_{p} consist of (p+1)-tuples of vertices of ℛ whose normal cones have a common point in the open cone of linear forms strictly positive on nonzero positive semidefinite forms. Face maps delete entries. Its augmented integer chain complex is exact. The GL_{N}(O_{L})-stabilisers have order at most s = 3^{dN}, and

Selectable PDF notation • transcription unchecked
( )dN(p+1)

|X_{p}/ GL_{N}(O_{L})| ≤A_{p} := I_{0}^{1+dN} (2R + 1)^{d}+ 1 .

Proof. We spell out the geometric properties of the complex. Near a positive form, only finitely many of the inequalities q_{h}(y) ≥ 1 can be active: compare with a fixed positive form and use discreteness of the embedded integral lattice. Thus ℛ is locally polyhedral. It lies in the positive cone by Lemma 8.23. A linear form strictly positive on the semidefinite cone has bounded sublevel sets there: on the compact unit section of that cone it has a positive minimum. Its minimum on the nonempty closed set ℛ is attained and its minimising face is compact. That face has a vertex, since a nonempty compact polyhedron has one. Consequently the normal cones of the vertices cover the open dual cone. They are locally finite there; a compact dual subregion bounds the associated minimising primal forms, where local polyhedrality gives only finitely many vertices. Every finite intersection with the open dual cone is convex.

The augmented complex of tuples in this cover is acyclic. One direct verification is the usual convex-cover double complex: take singular chains in the finite intersections in one direction and alternating deletion in the other. In one direction, convex contraction gives homology only in degree zero, namely the tuple complex. In the other, subdivide a singular simplex until its pieces lie in members of the cover; its augmented finite cover complex is a simplex and is exact. The resulting homology is that of the convex open dual cone, which is ℤ in degree zero. For closed polyhedral members, carry out this argument with small open thickenings on the compact support of any given finite chain. Local finiteness makes the relevant collection finite; empty intersections remain empty for sufficiently small thickening on that compact set. Thus the tuple nerve and its finite-chain filling are unchanged. This proves the claimed augmented exactness.

For any allowed tuple, all its vertices have common minimal vectors spanning L^{N}. To see this, average its vertices and examine the vectors minimal at the average. They are minimal at each vertex: each cor-responding value is at least one and their average is one. If their L-span were proper, at each embedding

choose a nonzero semidefinite Hermitian direction annihilating that span. Decreasing the averaged form slightly in that direction preserves the active inequalities and, by local polyhedrality and positivity, the inac-tive inequalities. It decreases the common strictly positive linear functional, contradicting the normal-cone minimum condition. Choose N independent common minimal vectors as columns of A. Lemma 8.23 gives | Norm det A| ≤I_{0}. Move A to one of the at most I_{0}^{1+dN} representatives of Lemma 8.24. Every vertex of the moved tuple lies among the nearby vertices counted in Lemma 8.25. Choosing such an encoding for each orbit is injective: equal encoded matrices and vertex tuples give the same orbit. This proves the formula for

Selectable PDF notation • transcription unchecked
A_{p}.

A vertex stabiliser permutes its finite set of minimal vectors, which spans the full vector space; hence it is finite. Reduction modulo 3 is injective on any finite subgroup of GL_{N}(O_{L}). Indeed, in the integer representation on O_{L}^{N}, a nonidentity finite-order matrix congruent to one modulo 3 would give one of prime order p. Write it 1 + 3^{a}B, a ≥1, with B not identically divisible by 3. The binomial expansion of its p-th power has leading term p3^{a}B of strictly smaller 3-adic valuation than all other terms, for p ̸= 3 and also for p = 3; it cannot equal one. Since O_{L}/3O_{L} has 3^{n} elements, its matrix ring has 3^{nN2}= 3^{dN} elements. This bounds every tuple stabiliser, which is a subgroup of a vertex stabiliser.

Source: Development/AutomorphicArithmetic/AutomorphicVarieties/HermiteNormalForms/Imports.lean:10--21; Development/AutomorphicArithmetic/AutomorphicVarieties/ConvexNerveComplexes/NerveFaceClosed.lean:33--57; Development/AutomorphicArithmetic/AutomorphicVarieties/HermiteNormalForms/ColsMinSmul.lean:34--75; Development/AutomorphicArithmetic/AutomorphicVarieties/HermiteNormalForms/Assembly.lean:19--30

Lemma 8.27 (Torsion from an exact permutation complex). Suppose an augmented complex · · · → ℤ[X_{1}] →ℤ[X_{0}] →M →0 of G-modules is exact through degree m −1. Suppose X_{p} has at most A orbits and stabilisers of order at most s ≥1, for p ≤m, and each boundary of a generator of X_{m+1} has ℓ^{1}-norm at most b ≥1. Then

| tors H_{m}(G, M)| ≤b^{A}s^{mAsm}.

Proof. For m = 0, coinvariants have at most A generators, and their relations have Euclidean norm at most their ℓ^{1}-norm, hence at most b. Choose linearly independent relation columns spanning their rational span. Hadamard’s inequality bounds their nonzero-volume determinant by b^{r}, r ≤A; the index of the full relation lattice in its saturation is no greater, since adding relations decreases that index. This proves the torsion bound b^{A}.

For the induction, put K = ker(ℤ[X_{0}] → M). The shifted augmented complex resolves K. The homology exact sequence of 0 → K → ℤ[X_{0}] → M → 0 compares the desired torsion with that of H_{m−1}(G, K) and H_{m}(G, ℤ[X_{0}]). The latter is the sum, over the at most A orbits, of the stabiliser homology, by the explicit induced-module bar resolution. For a finite group of order at most s, positive-degree homology is killed by its order (sum the translates in the bar resolution) and has at most s^{m} generators (the rank of the degree-m bar group). Its order is therefore at most s^{sm}. Thus |H_{m}(G, ℤ[X_{0}])| ≤ s^{Asm}. The shifted induction gives b^{A}s^{(m−1)Asm−1}. Their product is at most b^{A}s^{mAsm}. In this exact-sequence comparison the intervening positive stabiliser homology is finite, so taking its quotient cannot create torsion from an uncontrolled free quotient.

Source: Development/AutomorphicArithmetic/AutomorphicVarieties/TorsionBoundsChains/LogLeExp.lean:14--119; Development/AutomorphicArithmetic/AutomorphicVarieties/TorsionBoundsChains/GroupHomologyLongExact.lean:5--42

Proposition 8.28 (The uniform torsion estimate). For totally complex L, E ≥1, 1 ≤N ≤4E + 6, and m ≤2E + 1, the torsion subgroup of H_{m}(GL_{N}(O_{L}), ℤ) is finite, and

Selectable PDF notation • transcription unchecked
( )

log T_{m}(L, N) ≤ exp (2E[L : ℚ])^{24} log(2|∆_{L}|) .

Proof. Apply Lemma 8.27 to Proposition 8.26, with A = A_{m}, s = 3^{dN}, b = m + 2. Deleting an entry has exactly m + 2 terms in degree m + 1, so its boundary has the required norm bound. Writing D = |∆_{L}|, direct power estimates give

Selectable PDF notation • transcription unchecked
A_{m} ≤(2Dd)^{12(m+1)d3N}.

Here is the exponent check. Put B = 2Dd ≥2. Then I_{0} ≤B^{2d}, N^{N} ≤B^{d}, and (2N^{N}I_{0}^{2}+ 1)^{d} + 1 ≤ B^{(5d+2)d+1}. The total exponent in A_{m} is at most

Selectable PDF notation • transcription unchecked
2d + 2d^{2}N + 5d^{3}N(m + 1) + 2d^{2}N(m + 1) + dN(m + 1) ≤12(m + 1)d^{3}N.

Since s ≤e^{2dN}, the preceding torsion bound implies

Selectable PDF notation • transcription unchecked
( )

log T_{m}(L, N) ≤ exp 12(m + 1)d^{3}N log(2Dd) + 2(m + 1)dN + 2m .

This follows by taking its logarithm; use log(m + 2) ≤m + 1, and absorb the two resulting positive terms in the additional 2m in the exponent. For m = 0 the inequality follows directly from the base bound 2^{A}.

For completeness put a = E[L : ℚ] ≥1 and Y = log(2D) >0.69. Then N, d ≤10a, m + 1 ≤4a, and log(2Dd) = Y + log d ≤16aY , using log d ≤d −1. The first two terms inside the exponent are at most 7 681 200a^{6}Y , and 2m ≤12a^{6}Y . Their sum is at most (2a)^{24}Y , because 7 681 212 < 2^{24} and a^{6} ≤a^{24}. Exponentiate to obtain the stated bound.

Source: Development/AutomorphicArithmetic/AutomorphicVarieties/HermiteNormalForms/Orbnum.lean:6--92; Development/AutomorphicArithmetic/AutomorphicVarieties/ModularityArgument/ClassCounting/ExponentLe.lean:6--125

Proposition 8.29 (The precise duality comparison). For totally complex L, E ≥1, and T = ℤ_{ℓ}(E + 1), A = ℚ_{ℓ}/ℤ_{ℓ}(−E), there is an exact sequence of compact ℤ_{ℓ}-modules

Selectable PDF notation • transcription unchecked
0 −→𝒮_{ℓ}(L, E)^{∨}−→H^{2}(G_{L,Sℓ}, T) −→∏H^{2}(G_{Lv}, T) −→H^{0}(G_{L,Sℓ}, A)^{∨}−→0.
v|ℓ

Moreover H^{2}(G_{Lv}, T) ≃ H^{0}(G_{Lv}, A)^{∨}. Consequently (211) holds and its local sum is at most n(2 + ⌊log_{ℓ}n⌋+ E).

Proof. We use finite-module Poitou–Tate duality, including the perfect pairing between the two localisation kernels, and finite local Tate duality. A precise reference is J. S. Milne, Arithmetic Duality Theorems, second edition (2006), I, Theorem 4.10(a)–(b), with local duality I, Corollary 2.3.^{10} Their hypotheses hold for T_{r} = T/ℓ^{r}T, its Cartier dual A_{r} = ℓ^{−r}ℤ_{ℓ}/ℤ_{ℓ}(−E), and the finite set S_{ℓ} together with the infinite places: the modules are finite and continuous, their orders are units in the ring of S_{ℓ}-integers, and L has no real places. Thus there are no real-place Tate corrections.

Take the inverse limits on the integral side and direct limits on the torsion-dual side of these compatible finite duality sequences. The finite H^{1} groups satisfy the Mittag–Leffler condition: for each fixed finite target the descending images eventually stabilise. Hence their derived inverse limit is zero. The continuous cochain descriptions then identify the inverse limit degree-two groups with H^{2}(T); the direct limit groups on the other side are the groups with coefficients A. Pontryagin duality interchanges these limits and reverses exact sequences. The finite perfect pairing between localisation kernels therefore identifies the integral kernel with 𝒮_{ℓ}(L, E)^{∨}; exactness at the last two terms gives the displayed sequence. The same limit argument applied to the finite local duality pairing identifies its local degree-two term as H^{0}(A)^{∨}. Equivalently, these statements can be checked at every matching ℓ^{r}-quotient: a nonzero element in a compact kernel has a

^{10}https://www.jmilne.org/math/Books/ADTnot.pdf, printed pp. 28, 56–58.

nonzero finite quotient, where the finite perfect pairing detects it. This checks both injectivity and that no extra limit cokernel has been inserted. Additivity of extended module length in the displayed sequence, and preservation of length by Pontryagin duality on finite modules, give (211) whenever its right side is finite, which is the case needed here.

We give its local upper bound explicitly. For v | ℓ, the image of G_{Lv} in ℤ^{×}_{ℓ} has index at most [L_{v} : ℚ_{ℓ}] ≤ n, by intersection with the local cyclotomic extension. The principal unit subgroup is procyclic 1 + ℓℤ_{ℓ} for odd ℓ, and 1+4ℤ_{2} for ℓ= 2. An index at most n subgroup therefore contains 1+ℓ^{b}ℤ_{ℓ}, with b = 2+⌊log_{ℓ}n⌋. Apply its generator 1 + ℓ^{b} to an invariant of A. Lifting the exponent gives

Selectable PDF notation • transcription unchecked
v_{ℓ}((1 + ℓ^{b})^{E}−1) = b + v_{ℓ}(E) ≤b + E.

Thus the local invariant group has length at most 2 + ⌊log_{ℓ}n⌋+ E. There are at most n places above ℓ, proving the sum bound. In particular these local groups, and the global invariant subgroup, are finite.

Source: Development/AbelianVarietyTheory/Proofs/FoundationProofs/LengthEqLengthInvariantsOfTorsionLevels.lean:295--314,393--528; Development/AutomorphicArithmetic/AutomorphicVarieties/GaloisLibrary/GaloisCohomologyLength/Formula.lean:15--32; Development/AutomorphicArithmetic/AutomorphicVarieties/ModularityArgument/ClassCounting/SFinal.lean:8--98