OctoInformBeta

OpenAI preprint, September 2026 · Section 5. Proof of the main theorem

Section 5: dimension ratio, collision energy and rescaling: Dimension ratios and collision energies accompany a direct sum obstruction with a prescribed common bound

Energy growth and rescaling obstruct unitarizability.

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

The question is how kernel energy can grow relative to fibre dimension while a representation retains a prescribed common operator bound. The constructions involve finite families of rank-one weights, with their labels retained when the associated group elements coincide.

Let GG be a discrete group and KK a finite set. The fibre ℂK\mathbb C^K is the Euclidean space of complex functions on KK, of dimension |K||K|. The space HKH_K in (1) consists of square-summable fibre-valued functions, and ℬ(HK)\mathcal B(H_K) denotes its bounded complex-linear operators. HK=ℓ2(G;ℂK),∥ξ∥2=∑x∈G∥ξ(x)∥2.H_K=\ell^2(G;\mathbb C^K),\qquad \|\xi\|^2=\sum_{x\in G}\|\xi(x)\|^2.(1)

For x,y∈Gx,y\in G, coordinate evaluation ex:HK→ℂKe_x:H_K\to\mathbb C^K sends ξ\xi to ξ(x)\xi(x), and coordinate inclusion ιy:ℂK→HK\iota_y:\mathbb C^K\to H_K places a vector at yy and is zero elsewhere.✦ The row, column and block of F∈ℬ(HK)F\in\mathcal B(H_K) are the maps in (2). row⁡xF=exF,col⁡yF=Fιy,Fx,y=exFιy.\operatorname{row}_xF=e_xF,\qquad \operatorname{col}_yF=F\iota_y,\qquad F_{x,y}=e_xF\iota_y.(2)

For B∈ℬ(ℂK)B\in\mathcal B(\mathbb C^K), its coefficient matrix mat⁡(B)\operatorname{mat}(B) is the vector in ℂK×K\mathbb C^{K\times K} whose (a,b)(a,b) entry is (Beb)a(Be_b)_a, where ebe_b is the standard basis vector.✦ Its Euclidean norm is the unnormalized Hilbert-Schmidt norm. The kernel energy ℰ(F)\mathcal E(F) of F∈ℬ(HK)F\in\mathcal B(H_K) is the sum in (3). ∥mat⁡(B)∥2=∥B∥HS2=tr⁡(B*B),ℰ(F)=∑g∈G∥mat⁡(F1,g)∥2.\|\operatorname{mat}(B)\|^2=\|B\|_{\mathrm{HS}}^2 =\operatorname{tr}(B^*B),\qquad \mathcal E(F)=\sum_{g\in G}\|\operatorname{mat}(F_{1,g})\|^2.(3)✦

For a bijection e:G→Ge:G\to G, the coordinate permutation QeQ_e is the operator on HKH_K given in (4). The left regular representation UU of GG on HKH_K is the monoid homomorphism given in (4); for s∈Gs\in G, right translation RsR_s is the operator on HKH_K given in the same display. (Qeξ)(x)=ξ(e(x)),(U(g)ξ)(x)=ξ(g−1x),(Rsξ)(x)=ξ(xs).(Q_e\xi)(x)=\xi(e(x)),\qquad (U(g)\xi)(x)=\xi(g^{-1}x),\qquad (R_s\xi)(x)=\xi(xs).(4)

The translation commutant 𝒞K\mathcal C_K is the set of F∈ℬ(HK)F\in\mathcal B(H_K) that commute with every U(g)U(g).✦ Conjugation by U(g)U(g) and the coboundary of an operator T∈ℬ(HK)T\in\mathcal B(H_K) are defined in (5). Ad⁡U(g)T=U(g)TU(g−1),δT(g)=T−Ad⁡U(g)T.\operatorname{Ad}_{U(g)}T=U(g)TU(g^{-1}),\qquad \delta T(g)=T-\operatorname{Ad}_{U(g)}T.(5)

Let II be a finite label set, vi∈ℂKv_i\in\mathbb C^K and si∈Gs_i\in G for i∈Ii\in I. For i∈Ii\in I, the rank-one operator PiP_i on ℂK\mathbb C^K and its coefficient matrix mat⁡(Pi)\operatorname{mat}(P_i) are defined in (6). The collision kernel 𝖪\mathsf K is the ℂK×K\mathbb C^{K\times K}-valued function on GG whose value 𝖪(g)\mathsf K(g) is the sum of the coefficient matrices of the rank-one operators PiP_i over the labels with si=gs_i=g. Piu=⟨vi,u⟩vi,mat⁡(Pi)a,b=vi(a)vi(b)¯,𝖪(g)=∑i:si=gmat⁡(Pi).P_i u=\langle v_i,u\rangle v_i,\qquad \operatorname{mat}(P_i)_{a,b}=v_i(a)\overline{v_i(b)},\qquad \mathsf K(g)=\sum_{i:\,s_i=g}\operatorname{mat}(P_i).(6)

A mask cc on G×IG\times I is a function c:G×I→ℂc:G\times I\to\mathbb C with |c(x,i)|≤1|c(x,i)|\le1. For i∈Ii\in I, the weighted rank-one shift DicD_i^c, the masked operator FcF_c, and the unmasked operator FF are the bounded operators on HKH_K defined in (7). (Dicξ)(x)=c(x,i)Piξ(xsi),Fc=∑i∈IDic,F=F𝟏.(D_i^c\xi)(x)=c(x,i)P_i\xi(xs_i),\qquad F_c=\sum_{i\in I}D_i^c,\qquad F=F_{\boldsymbol 1}.(7)

For a nonnegative integer dd, a real ρ\rho and a nonnegative integer ℓ\ell, put nℓ=dℓn_\ell=d^\ell. Here ⌈a⌉+=max⁡{0,⌈a⌉}\lceil a\rceil_+=\max\{0,\lceil a\rceil\} and log\log is the natural logarithm.

Definition 1. The endpoint capacity rℓr_\ell is the nonnegative integer in (8). rℓ=⌈nℓρℓ⌉+.r_\ell=\left\lceil n_\ell\rho^\ell\right\rceil_+.(8)✦ The fibre dimension kℓk_\ell is the nonnegative integer in (9). kℓ=⌈2rℓlog(2nℓ)⌉+.k_\ell=\left\lceil2r_\ell\log(2n_\ell)\right\rceil_+.(9)✦

The Hilbert direct sum ⨁α∈JEα\bigoplus_{\alpha\in J}E_\alpha of a family of complex Hilbert spaces (Eα)α∈J(E_\alpha)_{\alpha\in J} is the Hilbert space of families η=(ηα)\eta=(\eta_\alpha) with ∑α∥ηα∥2<∞\sum_\alpha\|\eta_\alpha\|^2<\infty, with squared norm equal to that sum.✦ The direct sum operator ⨁αBα\bigoplus_\alpha B_\alpha of a uniformly bounded family Bα∈ℬ(Eα)B_\alpha\in\mathcal B(E_\alpha) is the bounded operator on ⨁α∈JEα\bigoplus_{\alpha\in J}E_\alpha given by ((⨁αBα)η)α=Bαηα((\bigoplus_\alpha B_\alpha)\eta)_\alpha=B_\alpha\eta_\alpha.✦

For a dimension sequence (mj)j≥0(m_j)_{j\ge0}, the stage spaces HjH_j, their stage sum HH, and the obstruction space H⊕HH\oplus H are specified in (10). Hj=ℓ2(G;ℂmj),H=⨁j≥0Hj,∥(ξ,η)∥H⊕H2=∥ξ∥H2+∥η∥H2.H_j=\ell^2(G;\mathbb C^{m_j}),\qquad H=\bigoplus_{j\ge0}H_j,\qquad \|(\xi,\eta)\|_{H\oplus H}^2=\|\xi\|_H^2+\|\eta\|_H^2.(10)

A representation π\pi on a complex Hilbert space EE is a monoid homomorphism G→ℬ(E)G\to\mathcal B(E). A representation π\pi of GG on a complex Hilbert space EE is similar to a unitary representation if there is a bounded invertible complex-linear map S:E→ES:E\to E satisfying (11), with no restriction on ∥S∥∥S−1∥\|S\|\|S^{-1}\|. ∥Sπ(g)S−1ξ∥=∥ξ∥(g∈G,ξ∈E).\|S\pi(g)S^{-1}\xi\|=\|\xi\| \qquad(g\in G,\ \xi\in E).(11)✦

The Claim

Theorem 1. Let d≥2d\ge2 be an integer and ρ≥0\rho\ge0 a real number. The dimensions satisfy (12) for every ℓ≥0\ell\ge0. kℓnℓ≤2(ρℓ+1nℓ)log⁡(2nℓ)+1nℓ.\frac{k_\ell}{n_\ell} \le 2\left(\rho^\ell+\frac1{n_\ell}\right)\log(2n_\ell) +\frac1{n_\ell}.(12)✦ If ρ<1\rho<1, the dimension ratio has the limit in (13). kℓnℓ→0(ℓ→∞).\frac{k_\ell}{n_\ell}\longrightarrow0 \qquad(\ell\to\infty).(13)✦ For 0<ρ<10<\rho<1 and a real sequence (Eℓ)(E_\ell) satisfying Eℓ≥nℓE_\ell\ge n_\ell for every ℓ\ell, the energy ratio has the limit in (14). Eℓkℓ→+∞(ℓ→∞).\frac{E_\ell}{k_\ell}\longrightarrow+\infty \qquad(\ell\to\infty).(14)✦

Let GG be a discrete group, I,KI,K finite sets, and let vi∈ℂKv_i\in\mathbb C^K and si∈Gs_i\in G. The unmasked operator has the coefficient kernel in (15) for every g∈Gg\in G. mat⁡(F1,g)=𝖪(g).\operatorname{mat}(F_{1,g})=\mathsf K(g).(15)✦ When every viv_i is a unit vector, the collision kernel and the unmasked operator satisfy (16) and (17), respectively. ∑g∈G∥𝖪(g)∥2≥|I|.\sum_{g\in G}\|\mathsf K(g)\|^2\ge |I|.(16)✦ ℰ(F)≥|I|.\mathcal E(F)\ge |I|.(17)✦

Let (mj)j≥0(m_j)_{j\ge0} be positive integers and let Aj,Tj∈ℬ(Hj)A_j,T_j\in\mathcal B(H_j), with AjA_j in the translation commutant on HjH_j. Suppose that a common real κ≥0\kappa\ge0 satisfies the bounds in (18), together with the energy growth specified there. ∥row⁡1Tj∥≤κ,∥col⁡1(Aj−Tj)∥≤κ,∥δTj(g)∥≤κ(j≥0,g∈G),ℰ(Aj)mj→+∞(j→∞).\begin{gathered} \|\operatorname{row}_1T_j\|\le\kappa,\qquad \|\operatorname{col}_1(A_j-T_j)\|\le\kappa,\qquad \|\delta T_j(g)\|\le\kappa \quad(j\ge0,\ g\in G),\\ \frac{\mathcal E(A_j)}{m_j}\longrightarrow+\infty \qquad(j\to\infty). \end{gathered}(18) For every real t>0t>0, there exists a representation π\pi on H⊕HH\oplus H with the properties in (19). ∥π(g)∥≤1+tκ(g∈G),π is not similar to a unitary representation.\|\pi(g)\|\le1+t\kappa\quad(g\in G),\qquad \pi\text{ is not similar to a unitary representation}.(19)✦

In the paper, the result reads: for a countable nonamenable group GG and the parameters S,d,ρS,d,\rho of Lemma 3.1, the quantities nℓ=dℓn_\ell=d^\ell, rℓ=⌈nℓρℓ⌉r_\ell=\lceil n_\ell\rho^\ell\rceil and kℓ=⌈2rℓlog⁡(2nℓ)⌉k_\ell=\lceil2r_\ell\log(2n_\ell)\rceil satisfy kℓ/nℓ≤2(ρℓ+1/nℓ)log⁡(2nℓ)+1/nℓ→0k_\ell/n_\ell\le2(\rho^\ell+1/n_\ell)\log(2n_\ell)+1/n_\ell\to0; for unit rank-one weights and arbitrary product collisions, Wℓ(s)=∑i:si=sPiW_\ell(s)=\sum_{i:\,s_i=s}P_i has ∑s∥Wℓ(s)∥HS2≥nℓ\sum_s\|W_\ell(s)\|_{\mathrm{HS}}^2\ge n_\ell; rescaling Aℓ,TℓA_\ell,T_\ell by t=ε/100t=\varepsilon/100, for ε>0\varepsilon>0, preserves divergence of energy divided by fibre dimension and gives common bound ε\varepsilon in Proposition 2.2, hence a nonunitarizable representation with bound 1+ε1+\varepsilon.

What Is Checked

The statements concern arbitrary discrete groups, finite label and fibre sets, and Hilbert direct sums indexed by natural numbers starting at zero; the rescaling statement assumes positive stage dimensions and the commutant condition, bounds and energy growth in (18).

The coordinate permutations in (4) are complex-linear isometries. A family of coordinate operators with a common nonnegative norm bound κ\kappa has a direct sum operator of norm at most κ\kappa. The rank-one operators in (6) have norm ∥vi∥2\|v_i\|^2, and the fibre maps u↦c(x,i)Piuu\mapsto c(x,i)P_i u in (7) have this same common upper bound.