OctoInformBeta

OpenAI preprint, September 2026 · §4. Recovering bounded operators from spectral vectors · §4

Randomization, truncation and selection: From a unit band vector with mean orbit defect at least 4γ4\gamma, an element vv of MM with ∥v∥≤C1K\|v\| \le C_1K and ∥Tvξ∥≥γ/2\|Tv\xi\| \ge \sqrt{\gamma}/2

Random phases and a norm cutoff give bounded recovery.

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 MM be a von Neumann algebra on a complex Hilbert space HH, with a cyclic and separating unit vector ξ\xi and modular operator Δ\Delta, and let TT be a bounded operator on HH. Suppose a unit vector hh lies in a narrow band of the spectrum of log⁡Δ\log\Delta and its modular orbit is, on average over time, not small under TT. The question is whether some v∈Mv \in M of controlled norm gives a vector vξv\xi in a slightly wider band that is still not small under TT. The item concerns the end of this construction, in which a weighted sum of modular translates of one element of MM is mixed with random phases and cut off in norm by functional calculus.

The inner product of HH is conjugate-linear in the first argument, and B(H)B(H) is the algebra of bounded operators on HH. A real spectral calculus DD on HH is a **-homomorphism f↦f(D)f \mapsto f(D) from the bounded continuous complex functions on ℝ\mathbb{R} into B(H)B(H) that is nondegenerate: the vectors f(D)kf(D)k, with ff compactly supported and k∈Hk \in H, are dense in HH.✦ The unitary group of a real spectral calculus DD on HH consists of the operators Ut=eitDU_t = e^{itD}, the images of the spectral characters et(u)=eitue_t(u) = e^{itu}.✦ A vector h∈Hh \in H has DD-spectral support in a set I⊆ℝI \subseteq \mathbb{R} when f(D)h=0f(D)h = 0 for every bounded continuous ff vanishing on II; this expresses ED(I)h=hE_D(I)h = h.✦ The graph of eD/2e^{D/2} is the set of pairs (h,k)(h, k) in H×HH \times H with fn(D)h→kf_n(D)h \to k as n→∞n \to \infty, where fn(u)=emin⁡(n,u)/2f_n(u) = e^{\min(n, u)/2} is the half exponential cutoff.✦

Standard modular data on HH consist of a von Neumann algebra MM on HH, a cyclic and separating unit vector ξ\xi, a real spectral calculus DD on HH standing for log⁡Δ\log\Delta, and an antiunitary involution JJ of HH, subject to the Tomita graph identity {(aξ,a*ξ):a∈M}¯={(h,Jk):(h,k) in the graph of eD/2},\overline{\{(a\xi, a^*\xi) : a \in M\}} = \{(h, Jk) : (h, k) \text{ in the graph of } e^{D/2}\},(1)✦ the statement S=JΔ1/2S = J\Delta^{1/2} for the Tomita operator SS. For such data the modular flow on B(H)B(H) is σt(a)=UtaU−t\sigma_t(a) = U_t a U_{-t}✦ and the vector state is φ(a)=⟨ξ,aξ⟩\varphi(a) = \langle \xi, a\xi \rangle✦. The data have trivial centralizer, Mφ=ℂ1M_\varphi = \mathbb{C}1, when every a∈Ma \in M with σt(a)=a\sigma_t(a) = a for all real tt is a scalar multiple of the identity.✦

The orbit defect of a vector hh under a bounded operator TT is the function t↦∥TUth∥2t \mapsto \|TU_th\|^2 on ℝ\mathbb{R}.✦ For a free ultrafilter ω\omega on ℕ\mathbb{N}, the mean of a real function ff on ℝ\mathbb{R} is the limit along ω\omega of its symmetric averages over [−k,k][-k, k]: mtf(t)=limk→ω12k∫−kkf(t)dt.m_t f(t) = \lim_{k \to \omega} \frac{1}{2k} \int_{-k}^{k} f(t)\,dt.(2)✦

Random Phases and the Norm Cutoff

A phase is one of the four complex numbers 1,−1,i,−i.1, \quad -1, \quad i, \quad -i.(3)✦ For elements a1,…,ana_1, \dots, a_n of a complex **-algebra and a choice ε=(ε1,…,εn)\varepsilon = (\varepsilon_1, \dots, \varepsilon_n) of phases, the phase sum is aε=∑jεjaja_\varepsilon = \sum_j \varepsilon_j a_j.✦ The phase average of a function FF on the 4n4^n choices of phases, with values in the algebra or in ℝ\mathbb{R}, is 𝔼εF(ε)=4−n∑ε∈{1,−1,i,−i}nF(ε).\mathbb{E}_\varepsilon F(\varepsilon) = 4^{-n} \sum_{\varepsilon \in \{1, -1, i, -i\}^n} F(\varepsilon).(4)

For weights p1,…,pnp_1, \dots, p_n and times t1,…,tnt_1, \dots, t_n in ℝ\mathbb{R}, the weighted modular average of an operator a∈B(H)a \in B(H) is the operator ∑jpjσtj(a)\sum_j p_j \sigma_{t_j}(a).✦

The cutoff function qq at a real number KK is the real function q(r)=Kmax⁡(K,r).q(r) = \frac{K}{\max(K, \sqrt{r})}.(5)✦ The polar cutoff of an element zz of a unital C*^*-algebra at KK is the element z[K]=zq(z*z),z^{[K]} = z\,q(z^*z),(6)✦ with qq as in (5) and q(z*z)q(z^*z) given by the continuous functional calculus of z*zz^*z.

The Constants

Let ψ\psi be a fixed smooth function on ℝ\mathbb{R} with values in [0,1][0, 1], equal to 11 on [−1,1][-1, 1] and vanishing outside (−2,2)(-2, 2).✦ The base kernel g0g_0 is the Fourier transform of the Schwartz function x↦ψ(2πx)x \mapsto \psi(2\pi x)✦, and C1C_1 is its L1L^1 norm: g0(t)=∫ℝe−2πixtψ(2πx)dx,C1=∥g0∥1=∫ℝ|g0(t)|dt.g_0(t) = \int_{\mathbb{R}} e^{-2\pi i x t}\, \psi(2\pi x)\,dx, \qquad C_1 = \|g_0\|_1 = \int_{\mathbb{R}} |g_0(t)|\,dt.(7)✦ The recovery moment bound c(s)c(s) at s∈ℝs \in \mathbb{R} is the real number c(s)=26+16es+1.c(s) = 26 + 16e^{s+1}.(8)

The Claim

Theorem 1. For elements a1,…,ana_1, \dots, a_n of a complex **-algebra, the phase sums aε=∑jεjaja_\varepsilon = \sum_j \varepsilon_j a_j satisfy 𝔼ε[(aε*aε)2]=(∑jaj*aj)2+∑jaj*(∑lalal*)aj−∑j(aj*aj)2.\mathbb{E}_\varepsilon\big[(a_\varepsilon^* a_\varepsilon)^2\big] = \Big(\sum_j a_j^* a_j\Big)^2 + \sum_j a_j^* \Big(\sum_l a_l a_l^*\Big) a_j - \sum_j (a_j^* a_j)^2.(9)✦

Given standard modular data on HH, a bounded operator TT on HH, weights p1,…,pn≥0p_1, \dots, p_n \ge 0 and bounded operators z1,…,znz_1, \dots, z_n on HH, put zε=∑jpjεjzjz_\varepsilon = \sum_j \sqrt{p_j}\,\varepsilon_j z_j, A=∑jpjzj*zjA = \sum_j p_j z_j^* z_j and B=∑jpjzjzj*B = \sum_j p_j z_j z_j^*. Then 𝔼ε∥Tzεξ∥2=∑jpj∥Tzjξ∥2\mathbb{E}_\varepsilon \|T z_\varepsilon \xi\|^2 = \sum_j p_j \|T z_j \xi\|^2(10)✦ and 𝔼εRe⁡φ((zε*zε)2)≤Re⁡φ(A2)+∑jpjRe⁡φ(zj*Bzj).\mathbb{E}_\varepsilon \operatorname{Re}\varphi\big((z_\varepsilon^* z_\varepsilon)^2\big) \le \operatorname{Re}\varphi(A^2) + \sum_j p_j \operatorname{Re}\varphi(z_j^* B z_j).(11)✦

For every real function FF on the 4n4^n choices of phases, some choice ε\varepsilon satisfies 𝔼ηF(η)≤F(ε).\mathbb{E}_\eta F(\eta) \le F(\varepsilon).(12)✦

If zz is an element of a unital C*^*-algebra and K>0K > 0, then ∥z[K]∥≤K\|z^{[K]}\| \le K(13)✦ and (z−z[K])*(z−z[K])≤K−2(z*z)2.(z - z^{[K]})^*(z - z^{[K]}) \le K^{-2} (z^*z)^2.(14)✦ If zz is a bounded operator on HH, xx a vector of HH and K>0K > 0, then ∥(z−z[K])x∥2≤K−2Re⁡⟨x,(z*z)2x⟩.\|(z - z^{[K]})x\|^2 \le K^{-2} \operatorname{Re}\langle x, (z^*z)^2 x\rangle.(15)✦ If zz lies in a von Neumann algebra MM on HH, then for every real KK z[K]∈M.z^{[K]} \in M.(16)✦

Given a bounded operator TT on HH, a real γ>0\gamma > 0 and a real ss, some K>0K > 0 satisfies ∥T∥2K−2c(s)≤γ/4.\|T\|^2 K^{-2} c(s) \le \gamma/4.(17)✦

Take standard modular data on HH, a bounded operator TT on HH and reals ss, QQ, d>0d > 0, γ>0\gamma > 0 and K>0K > 0 with ∥T∥2K−2Q≤γ/4\|T\|^2 K^{-2} Q \le \gamma/4. Let y∈My \in M be such that yξy\xi has DD-spectral support in [s−d,s+d][s - d, s + d], and let weights p1,…,pn≥0p_1, \dots, p_n \ge 0 and times t1,…,tn∈ℝt_1, \dots, t_n \in \mathbb{R} satisfy γ≤∑jpj∥Tσtj(y)ξ∥2\gamma \le \sum_j p_j \|T\sigma_{t_j}(y)\xi\|^2 and Re⁡φ(A2)+∑jpjRe⁡φ(σtj(y)*Bσtj(y))≤Q\operatorname{Re}\varphi(A^2) + \sum_j p_j \operatorname{Re}\varphi\big(\sigma_{t_j}(y)^* B \sigma_{t_j}(y)\big) \le Q, where now AA and BB are the weighted modular averages of y*yy^*y and of yy*yy^*. Then some v∈Mv \in M has ∥v∥≤C1K,∥Tvξ∥≥γ/2,vξ has D-spectral support in [s−2d,s+2d].\|v\| \le C_1 K, \qquad \|Tv\xi\| \ge \sqrt{\gamma}/2, \qquad v\xi \text{ has $D$-spectral support in } [s - 2d, s + 2d].(18)✦

Take standard modular data on HH with Mφ=ℂ1M_\varphi = \mathbb{C}1, a free ultrafilter ω\omega on ℕ\mathbb{N}, a bounded operator TT on HH and reals ss, 0<ρ≤1/20 < \rho \le 1/2, γ>0\gamma > 0 and K>0K > 0 with ∥T∥2K−2c(s)≤γ/4\|T\|^2 K^{-2} c(s) \le \gamma/4. If a unit vector hh has DD-spectral support in [s−ρ,s+ρ][s - \rho, s + \rho] and its orbit defect has mean mt∥TUth∥2≥4γm_t\|TU_th\|^2 \ge 4\gamma, as in (2), then some v∈Mv \in M has ∥v∥≤C1K,∥Tvξ∥≥γ/2,vξ has D-spectral support in [s−4ρ,s+4ρ].\|v\| \le C_1 K, \qquad \|Tv\xi\| \ge \sqrt{\gamma}/2, \qquad v\xi \text{ has $D$-spectral support in } [s - 4\rho, s + 4\rho].(19)✦

In the paper, the result reads: for independent phases εj\varepsilon_j uniform on {1,−1,i,−i}\{1, -1, i, -i\} and z=∑jpjεjzjz = \sum_j \sqrt{p_j}\,\varepsilon_j z_j, one has 𝔼∥Tzξ∥2=∑jpj∥Tzjξ∥2\mathbb{E}\|Tz\xi\|^2 = \sum_j p_j \|Tz_j\xi\|^2 and 𝔼[(z*z)2]=A2+∑jpjzj*Bzj−∑jpj2(zj*zj)2\mathbb{E}[(z^*z)^2] = A^2 + \sum_j p_j z_j^* B z_j - \sum_j p_j^2 (z_j^* z_j)^2, the truncation z[K]=umin⁡(|z|,K)z^{[K]} = u\min(|z|, K), with z=u|z|z = u|z|, has ∥z[K]∥≤K\|z^{[K]}\| \le K and 𝔼∥(z−z[K])ξ∥2≤CK−2\mathbb{E}\|(z - z^{[K]})\xi\|^2 \le C K^{-2}, and for the paper’s filter FdnF_{d_n} and the filtered truncation v=Fdn(z[K])v = F_{d_n}(z^{[K]}) with K≥2∥T∥C/γK \ge 2\|T\|\sqrt{C/\gamma} some realization of the phases gives ∥Tvξ∥≥γ/2\|Tv\xi\| \ge \sqrt{\gamma}/2, the paper’s CC and dnd_n standing for the page’s QQ and dd.

What Is Checked

The claim is stated over standard modular data, built from a von Neumann algebra with a cyclic and separating unit vector, a real spectral calculus with its unitary group and the graph of eD/2e^{D/2}, and an antiunitary involution, and over DD-spectral support, the modular flow, the vector state, trivial centralizers, the orbit defect and its mean along a free ultrafilter, phases with their sums and averages, weighted modular averages, the cutoff function, the polar cutoff, and the constants C1C_1 and c(s)c(s).

About these constructions the library proves that the orbit defect t↦∥TUth∥2t \mapsto \|TU_th\|^2 is a bounded continuous function on ℝ\mathbb{R} with values at most ∥T∥2∥h∥2\|T\|^2\|h\|^2, and that the base kernel g0g_0 is a Schwartz function. It also proves that the half exponential cutoff fn(u)=emin⁡(n,u)/2f_n(u) = e^{\min(n,u)/2} is a bounded continuous function on ℝ\mathbb{R} with values at most en/2e^{n/2}, and that each spectral character et(u)=eitue_t(u) = e^{itu} is a bounded continuous function on ℝ\mathbb{R} with values of modulus one.