Theorem S11 (only the reduced denominator matters).
ProvedQuantizedWeightLattices.DenominatorSharp.denominator_constant_reducedaether-catalogbridges
Theorem S11 (only the reduced denominator matters). For an arbitrary
mixing weight k/q with 0 < k < q, the exact maximal convexity defect is
(1 − g/q)·δ where g = gcd(k, q): the sharp constant depends on the reduced
denominator q/g alone. Two mixing weights of the same value have the same
convexity cost, and a weight with a small reduced denominator (e.g. 1/2) is
strictly gentler than a nearby weight with a large one.
theorem QuantizedWeightLattices.DenominatorSharp.denominator_constant_reduced{δ : ℝ} (hδ : 0 < δ) {k q : ℕ} (hk0 : 0 < k) (hkq : k < q) :
IsGreatest (defectSet δ k q) (δ * (1 - (Nat.gcd k q : ℝ) / (q : ℝ))) := by sorry
Formalization Note Transplanted verbatim from the Aether Catalog source Bridges/QuantizedWeightLatticesDenominatorSharp.lean; the statement is byte-identical to the source declaration, elaborated with autoImplicit disabled in the platform environment.
Preamble
-- Thm stub generated from Bridges/QuantizedWeightLatticesDenominatorSharp.lean
import Mathlib
import Definitions.Def_Bridges_QuantizedWeightLatticesDenominatorSharp
import Definitions.Def_Bridges_QuantizedWeightLatticesSharpConstant
/-
Copyright (c) 2026. Phase A Research Mission: Bridge NumberTheory ↔ Machine Learning.
# Arithmetic Geometry of Transformer Weight Lattices, VI:
# the denominator law is exact — the defect spectrum at a rational mixing weight
`Bridges.QuantizedWeightLatticesSharpConstant` proved the **denominator law**: for a
convex `L`-Lipschitz loss and the `δ`-grid quantizer, the convexity defect at the
interpolation weight `a = k/q` is at most `(1 − 1/q)·L·δ`, and it exhibited a
witness attaining the bound only for `k = q − 1`. Conjecture 1 of
`FUTURE_DIRECTIONS.md` asked whether the bound is attained for *every* `k`
coprime to `q`.
This file settles that conjecture **affirmatively, and in a stronger form**:
* `defect_spectrum` — for `gcd(k, q) = 1` and *every* `j` with `0 ≤ j < q` there is
an explicit convex `1`-Lipschitz loss (a "distance to a target weight" loss) and
an explicit pair of weights whose quantized convexity defect at mixing weight
`k/q` equals **exactly** `δ·j/q`. The whole arithmetic progression
`(δ/q)·{0, 1, …, q−1}` is realised: the defect spectrum is the full rank-one
lattice slice predicted by the arithmetic half of the denominator law.
* `denominator_constant_exact` — consequently the supremum of the defect over all
convex `1`-Lipschitz losses at mixing weight `k/q` is attained and equals
`(1 − 1/q)·δ`; formally an `IsGreatest` statement, so the constant of the
denominator law is optimal for every residue `k` coprime to `q`.
* `denominator_constant_reduced` — for a general (not necessarily reduced) weight
`k/q` the sharp constant is `(1 − gcd(k,q)/q)·δ`, i.e. only the *reduced*
denominator matters. Arithmetic, not size, of the mixing weight controls the
loss of convexity.
* `quantizeTensor_defect_denominator` — the denominator law itself lifts from `ℝ`
to whole weight tensors `ι → ℝ` with the sup norm: entrywise `δ`-grid
quantization of a transformer's weights loses at most `(1 − 1/q)·L·δ` of
convexity at any mixing weight of denominator `q`.
* `mesh_determined_by_defectSet`, `reducedDenominator_determined_by_defectSet` — a
spectral converse: the achievable convexity defects are an arithmetic
fingerprint. They determine the mesh `δ` of the quantizer and the *reduced*
denominator of the mixing weight, both recoverable from landscape measurements
alone.
The mechanism: all three roundings are lattice points, so the discrepancy
`a·Qx + (1−a)·Qy − Q(a x + (1−a) y)` lies in `(δ/q)·ℤ`; its numerator is
`k·(round X − round Y) mod q`, and as `round X − round Y` ranges over `ℤ` this
covers all of `ℤ/qℤ` exactly when `gcd(k, q) = 1`. Attainment is therefore a
covering statement for the cyclic group `ℤ/qℤ`, and the covering witnesses are
produced here by Bézout.
-/
open QuantizedWeightLattices.DenominatorSharp
open Set QuantizedWeightLattices QuantizedWeightLattices.SharpConstant
/-! ## Section 1: the covering lemma for `ℤ/qℤ` -/
/-! ## Section 2: exact roundings of the witness points -/
/-! ## Section 3: the witness family realising an arbitrary lattice defect -/
/-! ## Section 4: the defect spectrum at a coprime mixing weight -/
/-! ## Section 5: general (non-reduced) mixing weights -/Formal statement
theorem QuantizedWeightLattices.DenominatorSharp.denominator_constant_reduced{δ : ℝ} (hδ : 0 < δ) {k q : ℕ} (hk0 : 0 < k) (hkq : k < q) :
IsGreatest (defectSet δ k q) (δ * (1 - (Nat.gcd k q : ℝ) / (q : ℝ))) := by sorrySource