The specified contraction construction spans third-order linear layers
OpenQAlgorithms.Papers.Independent.Intakenext100.P94.third_order_contraction_constructionLet and consider real linear maps on full order-three tensors, equivariant for simultaneous inverse permutation of all indices. Form contractions by identifying any subset of the three input indices into blocks, retaining any subset of those blocks as free indices, and summing each remaining block once. Blocks are canonically named by their least input slot. There are respectively contractions of ranks .
For generic output entries, inject the retained indices into the three output slots in every possible way. Add the analogous terms on each of the three equal-pair output supports, using two output slots, and on the all-equal support, using one slot. These are additions to the generic formula, so they also operate on intersections of the supports. If is the resulting explicitly defined equality-constraint kernel, the weighted map is
There are generic, terms for each of the three pair supports, and all-diagonal terms. Every equivariant linear map has a unique such weight vector:
This states the paper’s contraction construction itself as a complete linear parametrization.
import Definitions.Def_QAlgorithms_Papers_Independent_Intakenext100_P94
noncomputable section
namespace QAlgorithms.Papers.Independent.Intakenext100.P94
open scoped BigOperators
attribute [local instance] Classical.propDecidable
/-- Appendix C §10.2, source PDF pp.15–16: the specified contraction algorithm
spans the equivariant linear layers and has 203 independent weights for n≥6.
Corrects “three” all-diagonal scalar contractions to one (p.16, item 4c) and
uses the stated two-index sum in item 3a. No closed formula is invented. -/
theorem third_order_contraction_construction (n : ℕ) (hn : 6 ≤ n) :
(∀ r : Fin 4,
Fintype.card {p : Recipe // p.val.2.card = r.val} = (![5,10,6,1] : Fin 4 → ℕ) r) ∧
Fintype.card (PlacedRecipe 3) = 77 ∧
Fintype.card (PlacedRecipe 2) = 37 ∧
Fintype.card (PlacedRecipe 1) = 15 ∧
Fintype.card ThirdIndex = 203 ∧
(∀ L : Tensor n 3 →ₗ[ℝ] Tensor n 3,
L ∈ equivariantLinearMaps (permutationTensorAction (n := n) (k := 3)) permutationTensorAction ↔
∃! w : ThirdIndex → ℝ, ∀ F, L F = thirdFormula w F) := by sorry
end QAlgorithms.Papers.Independent.Intakenext100.P94
end
Read-back
What the Lean code literally says, in plain math · gpt-6.1-sol
For every natural number satisfying , write and let be the real vector space of all functions . A recipe is a pair , where satisfies and for every , and consists of fixed points of ; a placed recipe with positions additionally specifies an injection . The theorem asserts jointly that, for each , the number of recipes with is respectively for ; the numbers of placed recipes with three, two, and one positions are respectively ; and the disjoint union of the placed recipes with three positions, the pairs of an element of and a placed recipe with two positions, and the placed recipes with one position has cardinality . It further asserts that every real linear map satisfies the following equivalence: commutes with every permutation action , meaning for every permutation of and every tensor , if and only if there exists exactly one weight function such that, for every and ,
Here, for a placed recipe , is precisely when for all and for every , and is otherwise, with no inequality required between coordinates belonging to different values of . For an index in the three-position component, ; for an index in the two-position component, equals when , and equals otherwise; for an index in the one-position component, equals when all coordinates of are equal, and equals otherwise. The additions are modulo .
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.