OctoInformBeta

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

Section 4.2: operators associated to an endpoint assignment: Constant and Boolean masks on square-summable functions with finite Euclidean fibres

Endpoint assignments determine masked operators.

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

Let GG be a discrete group, and let II and KK be finite sets. Fix vectors vi∈ℂKv_i\in\mathbb C^K and group elements si∈Gs_i\in G, indexed by i∈Ii\in I. The list positions remain distinct even when some of the sis_i coincide. The operators considered here couple the fibre at xsixs_i to the fibre at xx, with a coefficient chosen separately for each pair (x,i)(x,i).

The Hilbert sum ⨁j∈JEj\bigoplus_{j\in J}E_j of a family (Ej)j∈J(E_j)_{j\in J} of normed abelian groups is the square-summable direct sum in (1), with the indicated norm: ⨁j∈JEj={ξ=(ξj)j∈J:∑j∈J∥ξj∥2<∞},∥ξ∥2=∑j∈J∥ξj∥2.\bigoplus_{j\in J}E_j =\left\{\xi=(\xi_j)_{j\in J}:\sum_{j\in J}\|\xi_j\|^2<\infty\right\}, \qquad \|\xi\|^2=\sum_{j\in J}\|\xi_j\|^2.(1)✦ The fibre ℂK\mathbb C^K is the complex Euclidean space of functions on KK, with its coordinate inner product, and the common translation space ℋ\mathcal H is its Hilbert sum over GG, as in (2): ⟨z,w⟩=∑k∈Kz(k)¯w(k),ℋ=ℓ2(G;ℂK)=⨁x∈GℂK.\langle z,w\rangle=\sum_{k\in K}\overline{z(k)}w(k), \qquad \mathcal H=\ell^2(G;\mathbb C^K)=\bigoplus_{x\in G}\mathbb C^K.(2) The fibre has dimension |K||K|.

For a bijection e:G→Ge:G\to G, the coordinate permutation QeQ_e on ℋ\mathcal H is the complex-linear map in (3).✦ The right translation RsR_s on ℋ\mathcal H, for s∈Gs\in G, is the coordinate permutation associated with x↦xsx\mapsto xs: (Qeξ)(x)=ξ(e(x)),(Rsξ)(x)=ξ(xs).(Q_e\xi)(x)=\xi(e(x)), \qquad (R_s\xi)(x)=\xi(xs).(3)

Let EjE_j be complex Hilbert spaces and let Bj:Ej→EjB_j:E_j\to E_j be bounded complex-linear operators with ∥Bj∥≤κ\|B_j\|\leq\kappa for a common nonnegative real number κ\kappa. The direct sum operator ⨁j∈JBj\bigoplus_{j\in J}B_j is the bounded complex-linear map on ⨁jEj\bigoplus_jE_j given by (4): ((⨁j∈JBj)ξ)j=Bjξj.\left(\left(\bigoplus_{j\in J}B_j\right)\xi\right)_j=B_j\xi_j.(4)✦

A mask c:G×I→ℂc:G\times I\to\mathbb C is a function satisfying |c(x,i)|≤1|c(x,i)|\leq1 for every x∈Gx\in G and i∈Ii\in I. For each ii, let PiP_i be the rank-one operator on ℂK\mathbb C^K and Bc,iB_{c,i} the weighted rank-one shift on ℋ\mathcal H defined in (5): Piη=⟨vi,η⟩vi,Bc,i=(⨁x∈Gc(x,i)Pi)Rsi.P_i\eta=\langle v_i,\eta\rangle v_i, \qquad B_{c,i}=\left(\bigoplus_{x\in G}c(x,i)P_i\right)R_{s_i}.(5)✦ The masked operator TcT_c on ℋ\mathcal H is the bounded complex-linear map in (6), given by the finite sum of these shifts: Tc=∑i∈IBc,i,(Tcξ)(x)=∑i∈Ic(x,i)⟨vi,ξ(xsi)⟩vi.T_c=\sum_{i\in I}B_{c,i}, \qquad (T_c\xi)(x)=\sum_{i\in I}c(x,i)\langle v_i,\xi(xs_i)\rangle v_i.(6)✦

An endpoint assignment a:G×I→{0,1}a:G\times I\to\{0,1\} is a function that chooses the row endpoint at xx when a(x,i)=1a(x,i)=1 and the column endpoint at xsixs_i when a(x,i)=0a(x,i)=0. The Boolean indicator 𝟏a:G×I→ℂ\mathbf1_a:G\times I\to\mathbb C of an endpoint assignment a:G×I→{0,1}a:G\times I\to\{0,1\} is the complex-valued function in (7): 𝟏a(x,i)={1,a(x,i)=1,0,a(x,i)=0.\mathbf1_a(x,i)= \begin{cases} 1,&a(x,i)=1,\\ 0,&a(x,i)=0. \end{cases}(7)✦

The Definition

Definition 1. Let GG be a discrete group, let II and KK be finite sets, and fix vi∈ℂKv_i\in\mathbb C^K and si∈Gs_i\in G for i∈Ii\in I. For the masked operators of (6), the constant assignment operator AA on ℋ\mathcal H is the bounded complex-linear map given by (8). A:=T1,1(x,i)=1.A:=T_1,\qquad 1(x,i)=1.(8)✦ For an endpoint assignment a:G×I→{0,1}a:G\times I\to\{0,1\}, the oriented assignment operator TT on ℋ\mathcal H is the bounded complex-linear map given by (9). T:=Ta:=T𝟏a.T:=T_a:=T_{\mathbf1_a}.(9)✦

In the paper, the result reads: fix n≥2n\geq2 and r≥1r\geq1, the frame parameters and unit vectors from Lemma 4.1, a countable group GG with a product list (si)i=1n(s_i)_{i=1}^n, and an assignment a:G×{1,…,n}→{0,1}a:G\times\{1,\ldots,n\}\to\{0,1\} with at most rr row-assigned edges at each row and at most rr column-assigned edges at each column, and set A=T1A=T_1, T=TaT=T_a and (U(g)ξ)(x)=ξ(g−1x)(U(g)\xi)(x)=\xi(g^{-1}x).

What Is Checked

The definitions (8) and (9) are stated over an arbitrary discrete group, finite list and fibre index sets, arbitrary fibre vectors and group elements, and an arbitrary endpoint assignment. The coordinate permutations in (3) are isometries✦, and the direct sum operator in (4) satisfies the uniform norm bound in (10): ∥Qeξ∥=∥ξ∥,∥⨁j∈JBj∥≤κ.\|Q_e\xi\|=\|\xi\|, \qquad \left\|\bigoplus_{j\in J}B_j\right\|\leq\kappa.(10) For every endpoint assignment, the Boolean indicator (7) satisfies the estimate (11): |𝟏a(x,i)|≤1(x∈G,i∈I).|\mathbf1_a(x,i)|\leq1\qquad(x\in G,\ i\in I).(11)✦