S2.6, p. 27 — EvNorm (S39): h̄ is invariant and ĥ is equivariant under the unitary representations U^L
ProvedOrbNetEqui.Equivariance.evNorm_equivariantequivariancemachine-learningnormalizationp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1
Let be unitary representations of a group on , and let be a feature, i.e. a vector for each label and channel . EvNorm (S39) with statistics , positive scales and stability factor is
with the regularized norm of (23).
Then for every , writing ,
The invariant part feeds the multilayer perceptrons of the point-wise interaction block, and the equivariant part carries the directions.
Formalization Note. Scalars are complex, so the norm uses . The statistics are fixed parameters, so the statement holds for every value of them (batch or layer statistics alike). The page's positivity of and is kept as hypotheses. No subgroup restriction is needed, because every is unitary.
Preamble
import Mathlib import Definitions.Def_OrbNetEqui_Equivariance_Basic
Formal statement
namespace OrbNetEqui.Equivariance
open Matrix
theorem evNorm_equivariant {K : Type*} [Group K] {Λ : Type*} {dim : Λ → ℕ} {nmax : ℕ}
(U : (L : Λ) → K →* Matrix.unitaryGroup (Fin (dim L)) ℂ)
(μ σ β : Λ → Fin nmax → ℝ) (ε : ℝ) (hβ : ∀ L c, 0 < β L c) (hε : 0 < ε)
(k : K) (x : Feature Λ dim nmax) :
evNormBar μ σ ε (featAct U k x) = evNormBar μ σ ε x ∧
evNormHat β ε (featAct U k x) = featAct U k (evNormHat β ε x) := by sorry
end OrbNetEqui.Equivariance
Source
Qiao et al., Informing Geometric Deep Learning with Electronic Interactions to Accelerate Quantum Chemistry, arXiv:2105.14655v4, p. 27, S2.6, "Proof of G-equivariance for EvNorm (S39)"; eqs. (22)–(23) p. 9, (S39) p. 26
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.