Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

S2.6, pp. 26–27 — the block convolution (S36) is equivariant: m_{u⃗}(g·T, g·h) = π¹(g_{u₁}) m_{u⃗}(T, h)

Proved
OrbNetEqui.Equivariance.blockConv_equivariant

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

equivariancemachine-learningp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1representation-theory

Let KKK be a group with subgroups GuG_uGu​, one per atom uuu. Let πu1:K→U(ιu)\pi^1_u:K\to\mathrm U(\iota_u)πu1​:K→U(ιu​) be unitary representations on the atomic bases, (UL)L(U^L)_L(UL)L​ unitary representations on the feature spaces, and let the matching matrices satisfy (S35):

Du πu2(k)=πu1(k) Dufor all u and k∈Gu,D_u\,\pi^2_u(k)=\pi^1_u(k)\,D_u\qquad\text{for all }u\text{ and }k\in G_u,Du​πu2​(k)=πu1​(k)Du​for all u and k∈Gu​,

where πu2(k)\pi^2_u(k)πu2​(k) acts as UkLU^L_kUkL​ on each copy of VL\mathbb V^LVL in VuV_uVu​.

Then for every g=(gu)ug=(g_u)_ug=(gu​)u​ with gu∈Gug_u\in G_ugu​∈Gu​, every N-body tensor TTT, all features (hu)u(h_u)_u(hu​)u​, every sub-tensor index u⃗=(u1,…,uN)\vec u=(u_1,\dots,u_N)u=(u1​,…,uN​) and every convolution channel iii, the block convolution (S36)

(mu⃗)v1i=∑v2,…,vNTu⃗(v1,…,vN)∏j=2Nρuj(huj)vji‾(\mathbf m_{\vec u})^i_{v_1}=\sum_{v_2,\dots,v_N}T_{\vec u}(v_1,\dots,v_N)\prod_{j=2}^N\overline{\rho_{u_j}(h_{u_j})^i_{v_j}}(mu​)v1​i​=v2​,…,vN​∑​Tu​(v1​,…,vN​)j=2∏N​ρuj​​(huj​​)vj​i​​

satisfies

mu⃗(g⋅T, g⋅h)i=πu11(gu1) mu⃗(T,h)i,\mathbf m_{\vec u}(g\cdot T,\ g\cdot h)^i=\pi^1_{u_1}(g_{u_1})\,\mathbf m_{\vec u}(T,h)^i,mu​(g⋅T, g⋅h)i=πu1​1​(gu1​​)mu​(T,h)i,

where ggg acts on TTT by the local action (S26) and on huh_uhu​ by U(gu)U(g_u)U(gu​).

This is the first of the four building-block equivariances that make up the paper's proof of G-equivariance (S2.6).

Formalization Note. The complex conjugate in (S36) is kept: it is what makes the unitary factors cancel. The matching layer (S44) uses the main-text weight shape (16), with a multiplicity index for the copies of each VL\mathbb V^LVL in VuV_uVu​, and its weights may depend on the atom. The tensor has order N=n+1N=n+1N=n+1, and the first index is v1v_1v1​.

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

open Matrix

theorem blockConv_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 : 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)
    (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)
    (uvec : Fin (n + 1) → Fin d) (i : Fin I) :
    blockConv D W (tensorAct π₁ g T) (fun u => featAct U (g u) (h u)) uvec i =
      (π₁ (uvec 0) (g (uvec 0)) : Matrix (ι (uvec 0)) (ι (uvec 0)) ℂ) *ᵥ
        blockConv D W T h uvec i := by sorry

end OrbNetEqui.Equivariance
Source
Qiao et al., Informing Geometric Deep Learning with Electronic Interactions to Accelerate Quantum Chemistry, arXiv:2105.14655v4, pp. 26–27, S2.6, "Proof of G-equivariance for the convolution block (S36)"; eqs. (S26), (S35), (S36), (S44)
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