S2.6, p. 27 — message passing (S37)–(S38) is equivariant: m̃_{u₁} ↦ π¹(g_{u₁}) m̃_{u₁}, and ρ†_{u₁}(m̃) ↦ U(g_{u₁}) ρ†_{u₁}(m̃)
ProvedOrbNetEqui.Equivariance.messagePassing_equivariantAssume the setting of the block-convolution milestone: subgroups of a group , unitary representations and , and matching matrices satisfying (S35) for . Fix scalar attention weights ; using the same before and after the transformation expresses the paper's invariance condition .
For with , any tensor , features and atom , the aggregated message (S37)
satisfies:
- for every ;
- the reverse matching layer (S45) turns it into an equivariant feature: .
Together these give the equivariance of the second argument of the update (S38).
Formalization Note. The page displays only the first conclusion. The second one, about the reverse matching layer (S45), is the step that (S38) also needs, and it is added to complete the page's claim for "(S37)-(S38)". The reverse matching weights have the main-text shape (28), , and may depend on the atom.
import Mathlib import Definitions.Def_OrbNetEqui_Equivariance_Basic
namespace OrbNetEqui.Equivariance
open Matrix
theorem messagePassing_equivariant {K : Type*} [Group K] {d n : ℕ} {ι : Fin d → Type*}
[∀ u, Fintype (ι u)] [∀ u, DecidableEq (ι u)] {Λ : Type*} [Fintype Λ] [DecidableEq Λ]
{dim : Λ → ℕ} {nmax I J : ℕ} {Kmul : Fin d → Λ → ℕ}
(S : Fin d → Subgroup K)
(π₁ : (u : Fin d) → K →* Matrix.unitaryGroup (ι u) ℂ)
(U : (L : Λ) → K →* Matrix.unitaryGroup (Fin (dim L)) ℂ)
(D : (u : Fin d) → Matrix (ι u) (MIdx dim Kmul u) ℂ)
(W : (u : Fin d) → Fin I → (L : Λ) → Matrix (Fin (Kmul u L)) (Fin nmax) ℂ)
(Wd : (u : Fin d) → (L : Λ) → Matrix (Fin nmax) (Fin (Kmul u L) × Fin I × Fin J) ℂ)
(hD : ∀ u, ∀ k ∈ S u, D u * π₂mat U Kmul u k = (π₁ u k : Matrix (ι u) (ι u) ℂ) * D u)
(g : Fin d → K) (hg : ∀ u, g u ∈ S u)
(T : (Fin (n + 1) → Orb ι) → ℂ) (h : Fin d → Feature Λ dim nmax)
(α : (Fin (n + 1) → Fin d) → Fin J → ℂ) (u : Fin d) :
aggregate D W (tensorAct π₁ g T) (fun w => featAct U (g w) (h w)) α u =
(fun ij => (π₁ u (g u) : Matrix (ι u) (ι u) ℂ) *ᵥ aggregate D W T h α u ij) ∧
revMatching D Wd u (aggregate D W (tensorAct π₁ g T) (fun w => featAct U (g w) (h w)) α u) =
featAct U (g u) (revMatching D Wd u (aggregate D W T h α u)) := by sorry
end OrbNetEqui.Equivariance
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.