OctoInformBeta

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

Corollary 4.3: Subfamily synthesis bounds control masked block operators, with bound 100100 for the chosen frame

Sparse assignments control rows and columns.

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, let II and KK be finite sets, and choose vectors vi∈ℂKv_i\in\mathbb C^K and elements si∈Gs_i\in G for i∈Ii\in I. An endpoint assignment chooses, for each x∈Gx\in G and i∈Ii\in I, the row xx or the column xsixs_i of the labelled edge from xx to xsixs_i. These choices divide a finite sum of rank-one shifts into a masked operator and its complement. The labels in II remain distinct when their group elements coincide.

The fibre E=ℂKE=\mathbb C^K is the complex Euclidean space of dimension |K||K|, with inner product (1). ⟨u,w⟩E=∑k∈Ku(k)¯w(k).\langle u,w\rangle_E=\sum_{k\in K}\overline{u(k)}w(k).(1) The Hilbert sum ⨁z∈ZEz\bigoplus_{z\in Z}E_z of a family of complex Hilbert spaces EzE_z consists of the square-summable coordinate families, with norm given by (2); write H=ℓ2(G;E)=⨁x∈GEH=\ell^2(G;E)=\bigoplus_{x\in G}E. ∑z∈Z∥ξz∥2<∞,∥ξ∥2=∑z∈Z∥ξz∥2.\sum_{z\in Z}\|\xi_z\|^2<\infty, \qquad \|\xi\|^2=\sum_{z\in Z}\|\xi_z\|^2.(2)

For x∈Gx\in G, the coordinate inclusion ιx:E→H\iota_x:E\to H is the complex-linear map that places a vector at xx and zero elsewhere✦, and the coordinate evaluation qx:H→Eq_x:H\to E is the complex-linear map that sends ξ\xi to ξ(x)\xi(x).✦ For a bounded complex-linear operator F∈ℬ(H)F\in\mathcal B(H) and x,y∈Gx,y\in G, its operator row at xx and operator column at yy are the compositions in (3). row⁡x(F)=qxF:H→E,col⁡y(F)=Fιy:E→H.\operatorname{row}_x(F)=q_xF:H\to E, \qquad \operatorname{col}_y(F)=F\iota_y:E\to H.(3)

Synthesis and Masked Operators

The bounded complex-linear synthesis operator VJ:ℂJ→EV_J:\mathbb C^J\to E associated to J⊆IJ\subseteq I forms the linear combination in (4). VJz=∑i∈Jzivi.V_Jz=\sum_{i\in J}z_i v_i.(4)✦ For the family v=(vi)i∈Iv=(v_i)_{i\in I}, a subfamily synthesis bound (p,b)∈ℕ×ℝ(p,b)\in\mathbb N\times\mathbb R is a pair such that every subfamily with at most pp labels, including the empty subfamily, satisfies (5). ∥VJ∥≤b(J⊆I,|J|≤p).\|V_J\|\le b \qquad (J\subseteq I,\ |J|\le p).(5)✦

For a bijection e:G→Ge:G\to G, the coordinate permutation Pe:H→HP_e:H\to H is the complex-linear map in (6).✦ Right translation RtR_t for t∈Gt\in G is the coordinate permutation induced by right multiplication by tt. (Peξ)(x)=ξ(e(x)),(Rtξ)(x)=ξ(xt).(P_e\xi)(x)=\xi(e(x)), \qquad (R_t\xi)(x)=\xi(xt).(6)✦

For bounded operators Fz∈ℬ(Ez)F_z\in\mathcal B(E_z) with a common bound ∥Fz∥≤κ\|F_z\|\le\kappa, where κ≥0\kappa\ge0, the direct sum ⨁z∈ZFz\bigoplus_{z\in Z}F_z on ⨁z∈ZEz\bigoplus_{z\in Z}E_z is the bounded complex-linear operator in (7). (⨁z∈ZFz)ξ=(Fzξz)z∈Z,∥⨁z∈ZFz∥≤κ.\left(\bigoplus_{z\in Z}F_z\right)\xi =(F_z\xi_z)_{z\in Z}, \qquad \left\|\bigoplus_{z\in Z}F_z\right\|\le\kappa.(7)✦

Let c:G×I→ℂc:G\times I\to\mathbb C satisfy |c(x,i)|≤1|c(x,i)|\le1. The rank-one operator Θi:E→E\Theta_i:E\to E and its weighted rank-one shift Wi,c:H→HW_{i,c}:H\to H are defined by (8). Θiu=⟨vi,u⟩vi,Wi,c=(⨁x∈Gc(x,i)Θi)Rsi.\Theta_i u=\langle v_i,u\rangle v_i, \qquad W_{i,c}=\left(\bigoplus_{x\in G}c(x,i)\Theta_i\right)R_{s_i}.(8)✦ The masked operator Mc∈ℬ(H)M_c\in\mathcal B(H) is the finite sum in (9)✦; the unmasked operator AA has constant coefficient one.✦ Mc=∑i∈IWi,c,A=M𝟏.M_c=\sum_{i\in I}W_{i,c}, \qquad A=M_{\mathbf 1}.(9)

A Boolean endpoint assignment a:G×I→{0,1}a:G\times I\to\{0,1\} is a map that assigns each labelled edge to its row when its value is 11 and to its column when its value is 00. For a Boolean endpoint assignment a:G×I→{0,1}a:G\times I\to\{0,1\}, the Boolean indicator ε:{0,1}→ℂ\varepsilon:\{0,1\}\to\mathbb C and the oriented assignment operator T∈ℬ(H)T\in\mathcal B(H) are the function and bounded complex-linear operator given by (10). ε(1)=1,ε(0)=0,T=Mε∘a.\varepsilon(1)=1,\qquad \varepsilon(0)=0, \qquad T=M_{\varepsilon\circ a}.(10) The row-one set Ix+I_x^+ and column-zero set Iy−I_y^- are the sets of labels assigned to the indicated endpoints, as in (11). Ix+={i∈I:a(x,i)=1},Iy−={i∈I:a(ysi−1,i)=0}.I_x^+=\{i\in I:a(x,i)=1\}, \qquad I_y^-=\{i\in I:a(ys_i^{-1},i)=0\}.(11)

The Claim

Theorem 1. Let r∈ℕr\in\mathbb N and b≥0b\ge0. Suppose vv has the subfamily synthesis bound (5) with p=min⁡(2r,|I|)p=\min(2r,|I|).

Suppose every row-one set has cardinality at most rr. Then the row bound (12) holds for every x∈Gx\in G. ∥qxT∥≤b2.\|q_xT\|\le b^2.(12)✦ Suppose instead that every column-zero set has cardinality at most rr. Then the complementary column bound (13) holds for every y∈Gy\in G. ∥(A−T)ιy∥≤b2.\|(A-T)\iota_y\|\le b^2.(13)✦ For the chosen frame bound b=b0=10b=b_0=10, the upper bound in each estimate is 100100.

In the paper, the result reads: under the preceding hypotheses and endpoint-assignment bounds, with qx=ιx*q_x=\iota_x^*, one has ∥qxT∥≤b02\|q_xT\|\le b_0^2 for every xx and ∥(A−T)ιy∥≤b02\|(A-T)\iota_y\|\le b_0^2 for every yy.

What Is Checked

The row and complementary column estimates are stated over arbitrary discrete groups, finite label and fibre sets, the subfamily synthesis bound at min⁡(2r,|I|)\min(2r,|I|), and the respective endpoint cardinality bounds.

The coordinate permutation of (6) satisfies (14). ∥Peξ∥=∥ξ∥.\|P_e\xi\|=\|\xi\|.(14)✦ The Boolean indicator used in (10) satisfies the coefficient bound (15). |ε(a(x,i))|≤1(x∈G,i∈I).|\varepsilon(a(x,i))|\le1 \qquad (x\in G,\ i\in I).(15)✦