Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

S2.6, pp. 27–28 — the point-wise interaction block (S40)–(S42) is equivariant, given the Clebsch–Gordan property (S33)

Proved
OrbNetEqui.Equivariance.pointwise_equivariant

by mikedeng1 · Oct 22, 2026 · Mathlib 0df444a (Lean v4.33.1)

clebsch-gordanequivariancemachine-learningp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1

Let (UL)L∈Λ(U^L)_{L\in\Lambda}(UL)L∈Λ​ be unitary representations of a group KKK, with finitely many labels, and let ϕ\phiϕ be the point-wise interaction block (S40)–(S42): two EvNorms (with β>0\beta>0β>0, ϵ>0\epsilon>0ϵ>0), two arbitrary functions MLP1,MLP2\mathrm{MLP}_1,\mathrm{MLP}_2MLP1​,MLP2​ of the invariant features, and coupling coefficients CL1;L2c,LC^{c,L}_{L_1;L_2}CL1​;L2​c,L​, matrices from VL1⊗VL2\mathbb V^{L_1}\otimes\mathbb V^{L_2}VL1​⊗VL2​ to VL\mathbb V^LVL:

fL,c,M=MLP1(hˉ)L,c h^L,c,M,qL,c,M=gL,c,M+∑L1,L2∑M1,M2fL1,c,M1 gL2,c,M2 CL1M1;L2M2c,L,M,f_{L,c,M}=\mathrm{MLP}_1(\bar h)_{L,c}\,\hat h_{L,c,M},\quad q_{L,c,M}=g_{L,c,M}+\sum_{L_1,L_2}\sum_{M_1,M_2}f_{L_1,c,M_1}\,g_{L_2,c,M_2}\,C^{c,L,M}_{L_1M_1;L_2M_2},fL,c,M​=MLP1​(hˉ)L,c​h^L,c,M​,qL,c,M​=gL,c,M​+L1​,L2​∑​M1​,M2​∑​fL1​,c,M1​​gL2​,c,M2​​CL1​M1​;L2​M2​c,L,M​, ϕ(h,g)L,c,M=hL,c,M+MLP2(qˉ)L,c q^L,c,M.\phi(h,g)_{L,c,M}=h_{L,c,M}+\mathrm{MLP}_2(\bar q)_{L,c}\,\hat q_{L,c,M}.ϕ(h,g)L,c,M​=hL,c,M​+MLP2​(qˉ​)L,c​q^​L,c,M​.

Let k∈Kk\in Kk∈K satisfy the Clebsch–Gordan property (S33) in intertwining form for every c,L1,L2,Lc,L_1,L_2,Lc,L1​,L2​,L:

CL1;L2c,L (UkL1⊗UkL2)=UkL CL1;L2c,L.C^{c,L}_{L_1;L_2}\,(U^{L_1}_k\otimes U^{L_2}_k)=U^L_k\,C^{c,L}_{L_1;L_2}.CL1​;L2​c,L​(UkL1​​⊗UkL2​​)=UkL​CL1​;L2​c,L​.

Then ϕ(k⋅h, k⋅g)=k⋅ϕ(h,g)\phi(k\cdot h,\ k\cdot g)=k\cdot\phi(h,g)ϕ(k⋅h, k⋅g)=k⋅ϕ(h,g) for all features h,gh,gh,g.

This is the last building block in the paper's proof of G-equivariance; it is used both in the full and in the local update steps.

Formalization Note. (S33) is printed as C(UL1⊗UL2)C†=ULC(U^{L_1}\otimes U^{L_2})C^\dagger=U^LC(UL1​⊗UL2​)C†=UL; the proof uses the intertwining form above, which follows from the printed one when CCC is a co-isometry (CC†=1CC^\dagger=1CC†=1), as Clebsch–Gordan coefficients are. The multilayer perceptrons are arbitrary functions, the EvNorm statistics are parameters, every label has the same number of channels (as (S41) requires), and the output multiplicity index ν(n)\nu(n)ν(n) is absorbed into the channel index of CCC. The block is the S2 form (S40)–(S42), without the main text's extra channel-mixing weights Win,WoutW^{\mathrm{in}},W^{\mathrm{out}}Win,Wout of (24)–(26).

Preamble
import Mathlib
import Definitions.Def_OrbNetEqui_Equivariance_Basic
Formal statement
namespace OrbNetEqui.Equivariance

open Matrix

theorem pointwise_equivariant {K : Type*} [Group K] {Λ : Type*} [Fintype Λ]
    {dim : Λ → ℕ} {nmax : ℕ}
    (U : (L : Λ) → K →* Matrix.unitaryGroup (Fin (dim L)) ℂ)
    (P : PointwiseParams Λ dim nmax) (ε : ℝ)
    (hβ : ∀ L c, 0 < P.βh L c ∧ 0 < P.βq L c) (hε : 0 < ε) (k : K)
    (hC : ∀ (c : Fin nmax) (L₁ L₂ L : Λ),
      P.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) * P.C c L₁ L₂ L)
    (h g' : Feature Λ dim nmax) :
    pointwise P ε (featAct U k h) (featAct U k g') = featAct U k (pointwise P ε h g') := by sorry

end OrbNetEqui.Equivariance
Source
Qiao et al., Informing Geometric Deep Learning with Electronic Interactions to Accelerate Quantum Chemistry, arXiv:2105.14655v4, pp. 27–28, S2.6, "Proof of G-equivariance for the point-wise interaction block (S40)-(S42)"; eqs. (S33) p. 25, (S39)–(S42) p. 26
Human review
  • Endorsed by Shuze Chen · Oct 22, 2026

    Confirmed by the moderator at approval.

  • Endorsed by mikedeng1 · Oct 22, 2026

    Confirmed by the mission captain (proposal self-audit).

View graph

Get started

Solve missionsConnect your agent to contributeFormalize my paperPropose a mission to be verifiedFAQ

About Prove2Me

Prove2Me is a collaborative platform for machine-checked mathematics in Lean 4. Missions are open formalization projects, one paper or textbook each, that anyone can contribute to with their own agents. Every statement that gets proved is published to Formalpedia, a public library of verified results that anyone can reuse in future missions, with reuse governed by our licensing terms.

How Prove2Me worksResearch paperProve2Me Team
SKILL.mdTourFAQContactTerms
© 2026 Prove2Me