Fourier blocks classify equivariant linear layers and their dimension
OpenQAlgorithms.Papers.Independent.Intakenext100.P94.fourier_block_classificationFix a symmetric group and finite-dimensional real activation spaces , with linear representations . Choose a finite family of pairwise inequivalent nonzero real irreducible models of dimensions . Let be nonnegative multiplicities, and let be actual invertible linear changes of coordinates that intertwine with left action by on every column of the respective and blocks. Then a real linear map is equivariant if and only if there is a unique family of real matrices such that, for every input ,
Moreover, the vector space of these equivariant linear maps has dimension
This classifies the learnable linear operation before the fixed nonlinearity. The dimension counts linear maps, independently of any loss of information caused by that nonlinearity.
import Definitions.Def_QAlgorithms_Papers_Independent_Intakenext100_P94
noncomputable section
namespace QAlgorithms.Papers.Independent.Intakenext100.P94
open scoped BigOperators
attribute [local instance] Classical.propDecidable
/-- Theorem 3, source PDF p.4, with exact parameter count on p.4–5.
The actual invertible intertwining Fourier maps are inputs, not the classification
being asserted. Irreducible labels may be reindexed without changing the claim.
No unitary diagonalization of arbitrary nonorthogonal representations is asserted. -/
theorem fourier_block_classification {n : ℕ} {I E H : Type*} [Fintype I]
[AddCommGroup E] [Module ℝ E] [FiniteDimensional ℝ E]
[AddCommGroup H] [Module ℝ H] [FiniteDimensional ℝ H]
(A : Perm n →* Module.End ℝ E) (B : Perm n →* Module.End ℝ H)
(d tau upsilon : I → ℕ)
(rho : (a : I) → Perm n →* Module.End ℝ (Fin (d a) → ℝ))
(hirr : ∀ a, IsRealIrreducible (rho a))
(hneq : ∀ a b, a ≠ b → ModelsInequivalent (rho a) (rho b))
(U : E ≃ₗ[ℝ] FourierBlocks d tau) (V : H ≃ₗ[ℝ] FourierBlocks d upsilon)
(hU : ∀ s x, U (A s x) = fourierColumnAction rho s (U x))
(hV : ∀ s x, V (B s x) = fourierColumnAction rho s (V x)) :
(∀ L : E →ₗ[ℝ] H,
L ∈ equivariantLinearMaps (fun s => A s) (fun s => B s) ↔
∃! w : FourierWeights tau upsilon, ∀ x, V (L x) = applyBlocks w (U x)) ∧
Module.finrank ℝ (equivariantLinearMaps (fun s => A s) (fun s => B s)) =
∑ a, tau a * upsilon a := by sorry
end QAlgorithms.Papers.Independent.Intakenext100.P94
end
Read-back
What the Lean code literally says, in plain math · gpt-6.1-sol
Let be any natural number, let be a finite index type, and let be finite-dimensional real vector spaces, each equipped with an additive commutative group structure and a real module structure. Let and be representations of the permutation group on and , respectively, preserving identity and multiplication as maps into real-linear endomorphisms. Let , and for each let be a representation of on . Assume that each and every real-linear subspace of invariant under all is either zero or the whole space; assume also that whenever , there is no real-linear isomorphism satisfying for every . Suppose there are real-linear isomorphisms
such that, for every , , and , each column of is obtained by applying to the corresponding column of , and each column of is obtained by applying to the corresponding column of . Then both of the following hold: for every real-linear map ,
and the real dimension of the subspace of all such equivariant linear maps is
The matrix product in the first conclusion has entries . The theorem allows , empty , and zero values of or ; the positivity assumption excludes zero for any existing index .
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.