S2.6, p. 26, and eq. (4), p. 3 — the UNiTE network is equivariant under local unitary changes of the atomic bases
ProvedOrbNetEqui.Equivariance.unite_equivariantConsider a molecule with atoms , each with a finite atomic-orbital basis , and an N-body tensor of order over the concatenated basis. Let be a group with subgroups , let be the representations on the atomic bases and unitary representations on the feature spaces , with finitely many labels. Let be the UNiTE network (Methods 1.1): the diagonal reduction , then steps of block convolution, message passing and point-wise interaction, then point-wise interaction steps. Assume:
- (S35) for all and ;
- (S33) in intertwining form, , for every step , every and ;
- (S32) for the diagonal reduction: for all and ;
- invariant attention: for every step and every ;
- in every EvNorm, and .
Then for every with , every tensor and every atom ,
where is the local action (S26), on each sub-tensor, and acts on every channel of label by .
This is the equivariance claim (4), , for the local unitary basis rotations (part (a) of the group of S2.2), which the paper proves in Appendix S2.6.
Formalization Note. is an abstract group ; scalars are complex. The diagonal reduction is any linear map satisfying (S32), which includes the paper's tabulated one (Corollary S3). The multilayer perceptrons are arbitrary functions, the EvNorm statistics are fixed parameters, the attention weights at each step are an arbitrary function of the step's inputs satisfying the paper's invariance condition, and (S33) is used in intertwining form. Matching weights carry the multiplicity index of (16) and (28) and may depend on the atom and step. Permutations (b) and translations (c) of the symmetry group are not part of the statement.
import Mathlib import Definitions.Def_OrbNetEqui_Equivariance_Basic
namespace OrbNetEqui.Equivariance
open Matrix
theorem unite_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)
(θ : UniteParams d n ι Λ dim nmax Kmul I J)
(hD : ∀ u, ∀ k ∈ S u, D u * π₂mat U Kmul u k = (π₁ u k : Matrix (ι u) (ι u) ℂ) * D u)
(hC : ∀ (t : ℕ) (u : Fin d), ∀ k ∈ S u, ∀ (c : Fin nmax) (L₁ L₂ L : Λ),
(θ.P t).C c L₁ L₂ L * Matrix.kroneckerMap (· * ·)
(U L₁ k : Matrix (Fin (dim L₁)) (Fin (dim L₁)) ℂ)
(U L₂ k : Matrix (Fin (dim L₂)) (Fin (dim L₂)) ℂ) =
(U L k : Matrix (Fin (dim L)) (Fin (dim L)) ℂ) * (θ.P t).C c L₁ L₂ L)
(hψ : ∀ u, ∀ k ∈ S u, ∀ (T : (Fin (n + 1) → ι u) → ℂ) (L : Λ) (ν : Fin (θ.μd u L)),
θ.ψ u (diagAct (π₁ u k : Matrix (ι u) (ι u) ℂ) T) L ν =
(U L k : Matrix (Fin (dim L)) (Fin (dim L)) ℂ) *ᵥ θ.ψ u T L ν)
(hα : ∀ (t : ℕ) (g : Fin d → K), (∀ u, g u ∈ S u) →
∀ (T : (Fin (n + 1) → Orb ι) → ℂ) (h : Fin d → Feature Λ dim nmax),
θ.α t (tensorAct π₁ g T) (fun u => featAct U (g u) (h u)) = θ.α t T h)
(hβ : ∀ (t : ℕ) (L : Λ) (c : Fin nmax), 0 < (θ.P t).βh L c ∧ 0 < (θ.P t).βq L c)
(hε : 0 < θ.ε)
(g : Fin d → K) (hg : ∀ u, g u ∈ S u) (T : (Fin (n + 1) → Orb ι) → ℂ) (u : Fin d) :
unite θ D (tensorAct π₁ g T) u = featAct U (g u) (unite θ D T u) := by sorry
end OrbNetEqui.Equivariance
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.