OctoInformBeta

OpenAI preprint, September 2026 · Section 4. Finite-dimensional weights and sparse block operators

Proposition 4.4: The coboundary bound depends on subfamily synthesis and the number of edges assigned to each endpoint

Sparse operators have bounded translation differences.

Hover over a blue or green term in the text, or a node in the graph, to find it in the other. Select it to read its declaration both in Lean and in words.

The Setting

Finite sums of right translations with rank-one fibre coefficients give operators on Hilbert spaces over a discrete group. An endpoint assignment allows these coefficients to depend on the row. Proposition 4.4 concerns the change in the resulting operator under left translation.

Let GG be a discrete group with identity ee, and KK a finite set. The fibre E=ℂKE=\mathbb C^K is the complex Euclidean space of dimension |K||K|, with inner product as in (1). The group Hilbert space H=ℓ2(G;E)H=\ell^2(G;E) consists of square-summable fibre-valued functions, with the norm in (1); ℬ(H)\mathcal B(H) denotes its bounded complex-linear operators. ⟨u,w⟩E=∑a∈Ku(a)¯w(a),∥ξ∥H2=∑x∈G∥ξ(x)∥E2.\langle u,w\rangle_E=\sum_{a\in K}\overline{u(a)}w(a), \qquad \|\xi\|_H^2=\sum_{x\in G}\|\xi(x)\|_E^2.(1)

For a family (Eα)α∈Λ(E_\alpha)_{\alpha\in\Lambda} of normed abelian groups, its Hilbert sum ⨁αEα\bigoplus_\alpha E_\alpha consists of families zz with ∑α∥zα∥2<∞\sum_\alpha\|z_\alpha\|^2<\infty, and has squared norm equal to this sum.✦ When the summands are complex Hilbert spaces, a family of bounded complex-linear operators Bα∈ℬ(Eα)B_\alpha\in\mathcal B(E_\alpha) with common bound ∥Bα∥≤κ\|B_\alpha\|\leq\kappa, where κ≥0\kappa\geq0, defines the direct sum operator in (2). ((⨁αBα)z)α=Bαzα,∥⨁αBα∥≤κ.\left(\left(\bigoplus_\alpha B_\alpha\right)z\right)_\alpha =B_\alpha z_\alpha, \qquad \left\|\bigoplus_\alpha B_\alpha\right\|\leq\kappa.(2)✦

For a bijection σ:G→G\sigma:G\to G, the coordinate permutation PσP_\sigma on HH is the complex-linear map specified in (3). The fibre-valued left regular representation U:G→ℬ(H)U:G\to\mathcal B(H) is a monoid homomorphism, and RtR_t denotes right translation by t∈Gt\in G; their actions are also given in (3). (Pσξ)(x)=ξ(σ(x)),(U(g)ξ)(x)=ξ(g−1x),(Rtξ)(x)=ξ(xt).(P_\sigma\xi)(x)=\xi(\sigma(x)), \qquad (U(g)\xi)(x)=\xi(g^{-1}x), \qquad (R_t\xi)(x)=\xi(xt).(3)

The stage HjH_j associated with a sequence (kj)j≥0(k_j)_{j\geq0} of nonnegative integers is the Hilbert space in (4). Write UjU_j for its left regular representation. Hj=ℓ2(G;ℂkj),j=0,1,2,….H_j=\ell^2(G;\mathbb C^{k_j}),\qquad j=0,1,2,\ldots.(4)✦

The translation commutant 𝒞(U)\mathcal C(U) is the set of bounded operators on HH commuting with every left translation, as in (5).✦ 𝒞(U)={F∈ℬ(H):FU(g)=U(g)F for every g∈G}.\mathcal C(U)=\{F\in\mathcal B(H):FU(g)=U(g)F\text{ for every }g\in G\}.(5)✦

Assignment Operators

Let II be a finite label set, with families vi∈Ev_i\in E and si∈Gs_i\in G. Labels remain distinct even when their group elements coincide. For a complex mask c:G×I→ℂc:G\times I\to \mathbb C with |c(x,i)|≤1|c(x,i)|\leq1, the weighted rank-one shift Qi,cQ_{i,c} and the masked operator McM_c are defined in (6). Here the direct sum acts over the rows x∈Gx\in G, and Pi∈ℬ(E)P_i\in\mathcal B(E) is the rank-one operator displayed there. Piu=⟨vi,u⟩vi,Qi,c=(⨁x∈Gc(x,i)Pi)Rsi,Mc=∑i∈IQi,c.P_i u=\langle v_i,u\rangle v_i,\qquad Q_{i,c}=\left(\bigoplus_{x\in G}c(x,i)P_i\right)R_{s_i},\qquad M_c=\sum_{i\in I}Q_{i,c}.(6) The rank-one factor PiP_i has norm ∥vi∥2\|v_i\|^2, which is also a common bound for the fibre maps c(x,i)Pic(x,i)P_i.

A Boolean mask a:G×I→{false,true}a:G\times I\to\{\mathrm{false},\mathrm{true}\} assigns the labelled edge from row xx to column xsixs_i to its row when true and to its column when false. The complex indicator εa\varepsilon_a of a Boolean mask aa on G×IG\times I is the complex-valued function on G×IG\times I with the values in (7). The constant assignment operator AA and the oriented assignment operator TT associated with aa are the masked operators on HH specified in (7). εa(x,i)={1,a(x,i)=true,0,a(x,i)=false,A=M1,T=Mεa.\varepsilon_a(x,i)= \begin{cases}1,&a(x,i)=\mathrm{true},\\0,&a(x,i)=\mathrm{false},\end{cases} \qquad A=M_1,\qquad T=M_{\varepsilon_a}.(7)

The row-one set Ra(x)R_a(x) and the column-zero set Ca(y)C_a(y) are the sets of labels assigned to those respective endpoints, given by (8). Ra(x)={i∈I:a(x,i)=true},Ca(y)={i∈I:a(ysi−1,i)=false}.R_a(x)=\{i\in I:a(x,i)=\mathrm{true}\},\qquad C_a(y)=\{i\in I:a(ys_i^{-1},i)=\mathrm{false}\}.(8)

For any family (vi)i∈I(v_i)_{i\in I} in EE, with II an arbitrary index set, and a finite subset J⊆IJ\subseteq I, the synthesis operator SJ:ℂJ→ES_J:\mathbb C^J\to E forms linear combinations of the vectors with labels in JJ, as in (9).✦ SJz=∑i∈Jzivi.S_J z=\sum_{i\in J}z_i v_i.(9)✦ For a family v=(vi)i∈Iv=(v_i)_{i\in I} in EE, a subfamily synthesis bound (p,b)(p,b) is a pair consisting of a nonnegative integer pp and a real number bb for which (10) holds, including the empty subfamily. ∥SJ∥≤bfor every finite J⊆I with |J|≤p.\|S_J\|\leq b \qquad\text{for every finite }J\subseteq I\text{ with }|J|\leq p.(10)✦

For a monoid homomorphism V:G→ℛV:G\to\mathcal R into a unital ring and Q∈ℛQ\in\mathcal R, conjugation Ad⁡V(g)\operatorname{Ad}_V(g) is a map from ℛ\mathcal R to itself, and the coboundary δVQ\delta_VQ is a function from GG to ℛ\mathcal R, with values specified in (11). The group inverse supplies the inverse of V(g)V(g). Ad⁡V(g)(Q)=V(g)QV(g−1),(δVQ)(g)=Q−Ad⁡V(g)(Q).\operatorname{Ad}_V(g)(Q)=V(g)QV(g^{-1}),\qquad (\delta_VQ)(g)=Q-\operatorname{Ad}_V(g)(Q).(11) For the oriented assignment operator, write D=δUTD=\delta_U T.

The Claim

Theorem 1. Let GG be a discrete group. On the stages associated with any dimension sequence, left translation satisfies (12). ∥Uj(g)ξ∥=∥ξ∥(j≥0,g∈G,ξ∈Hj).\|U_j(g)\xi\|=\|\xi\| \qquad(j\geq0,\ g\in G,\ \xi\in H_j).(12)✦ For arbitrary finite label and fibre sets and arbitrary families v,sv,s, the constant assignment operator satisfies (13). A∈𝒞(U).A\in\mathcal C(U).(13)✦ Let rr be a nonnegative integer and b≥0b\geq0. Suppose vv has subfamily synthesis bound (min⁡(2r,|I|),b)(\min(2r,|I|),b) and every row-one set and column-zero set in (8) has cardinality at most rr. The oriented assignment coboundary satisfies (14). ∥D(g)∥≤b2(g∈G).\|D(g)\|\leq b^2\qquad(g\in G).(14)✦ For any monoid homomorphism V:G→ℛV:G\to\mathcal R into a unital ring and Q∈ℛQ\in\mathcal R, the coboundary satisfies (15) and (16). (δVQ)(e)=0.(\delta_VQ)(e)=0.(15)✦ (δVQ)(gh)=(δVQ)(g)+Ad⁡V(g)((δVQ)(h))(g,h∈G).(\delta_VQ)(gh)=(\delta_VQ)(g) +\operatorname{Ad}_V(g)\bigl((\delta_VQ)(h)\bigr) \qquad(g,h\in G).(16)✦

In the paper, the result reads: under the preceding hypotheses and endpoint-assignment bounds, UU is a unitary representation and AA commutes with every U(g)U(g), while D(g)=T−U(g)TU(g)−1D(g)=T-U(g)TU(g)^{-1} satisfies sup⁡g∈G∥D(g)∥≤b02=100\sup_{g\in G}\|D(g)\|\leq b_0^2=100 for the chosen frame bound b0=10b_0=10, D(e)=0D(e)=0, and D(gh)=D(g)+U(g)D(h)U(g)−1D(gh)=D(g)+U(g)D(h)U(g)^{-1} for all g,h∈Gg,h\in G.

What Is Checked

The stage assertion is stated over arbitrary discrete groups and dimension sequences starting at zero. The masked operator assertions use finite label and fibre sets, with the synthesis and endpoint bounds required for (14); the identities (15) and (16) are stated in arbitrary unital rings.

The coordinate permutation isometry PσP_\sigma has the norm identity (17) for every bijection σ\sigma of GG. ∥Pσξ∥=∥ξ∥(ξ∈H).\|P_\sigma\xi\|=\|\xi\|\qquad(\xi\in H).(17)✦ For the Boolean coefficients in (7), the modulus bound is (18). |εa(x,i)|≤1(x∈G,i∈I).|\varepsilon_a(x,i)|\leq1\qquad(x\in G,\ i\in I).(18)✦