OctoInformBeta

OpenAI preprint, September 2026 · Section 5. Preparation, whitening, and the smooth limit

Proposition 5.2: Arbitrarily small changes and prediction loss controlled by the old good-arc complement

Refining flat spectral states in physical coordinates.

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

A flat spectral packet sum is to be changed so that a prescribed real observable can be approximated by a finite linear combination of its Koopman iterates. Physical closeness concerns derivatives of chosen integer iterates on compact sets, including iterates of the inverse map.

Write 𝕋=ℝ/ℤ\mathbb T=\mathbb R/\mathbb Z and e(t)=exp⁡(2πit)e(t)=\exp(2\pi i t). Let m1m_1 be normalized circle Haar measure, also used as phase Haar measure, and put m2=m1⊗m1m_2=m_1\otimes m_1 and m3=m1⊗m2m_3=m_1\otimes m_2. We identify the coordinate torus with the product torus 𝕋3=𝕋×𝕋×𝕋\mathbb T^3=\mathbb T\times\mathbb T\times\mathbb T.✦ The universal cover is ℝ3\mathbb R^3, with covering projection π(x0,x1,x2)=(x0,x1,x2)(mod⁡ℤ3)\pi(x_0,x_1,x_2)=(x_0,x_1,x_2)\pmod{\mathbb Z^3}. The slow-coordinate Haar measure m2m_2 is normalized Haar measure on 𝕋2\mathbb T^2. The product Haar measure m3m_3 is normalized Haar measure on 𝕋3\mathbb T^3.

Packet Data

A measurable packet family 𝒫=(mi,Ai)i∈J\mathcal P=(m_i,A_i)_{i\in J} consists of a finite set JJ, pairwise distinct nonzero integers mim_i, and measurable amplitudes Ai∈L2(𝕋2,m2;ℂ)A_i\in L^2(\mathbb T^2,m_2;\mathbb C).✦ Its angular packets pip_i and finite packet sum F𝒫F_{\mathcal P} are the functions in (1). pi(θ,u,a)=Ai(u,a)e(miθ),F𝒫(θ,u,a)=∑i∈Jpi(θ,u,a).p_i(\theta,u,a)=A_i(u,a)e(m_i\theta), \qquad F_{\mathcal P}(\theta,u,a)=\sum_{i\in J}p_i(\theta,u,a).(1) A reflection τ\tau of a packet family 𝒫=(mi,Ai)i∈J\mathcal P=(m_i,A_i)_{i\in J} is an involution τ:J→J\tau:J\to J reversing frequencies and conjugating amplitudes, as in (2). τ2=id,mτ(i)=−mi,Aτ(i)(u,a)=Ai(u,a)¯(i∈J,(u,a)∈𝕋2).\tau^2=\mathrm{id},\qquad m_{\tau(i)}=-m_i,\qquad A_{\tau(i)}(u,a)=\overline{A_i(u,a)} \quad(i\in J,\ (u,a)\in\mathbb T^2).(2) A packet family 𝒫\mathcal P is supported in [l,u][l,u] when every amplitude vanishes outside q([l,u])×𝕋q([l,u])\times\mathbb T, where q:ℝ→𝕋q:\mathbb R\to\mathbb T is reduction modulo one.✦

For a circle phase φ:𝕋→𝕋\varphi:\mathbb T\to\mathbb T, the packet spectral measure σi,φ\sigma_{i,\varphi} is the weighted pushforward in (3); the packet family spectrum σ𝒫,φ\sigma_{\mathcal P,\varphi} is their sum. σi,φ=((u,a)↦miφ(u))*(|Ai(u,a)|2dm2(u,a)),σ𝒫,φ=∑i∈Jσi,φ.\sigma_{i,\varphi} =\bigl((u,a)\mapsto m_i\varphi(u)\bigr)_* \bigl(|A_i(u,a)|^2\,dm_2(u,a)\bigr), \qquad \sigma_{\mathcal P,\varphi}=\sum_{i\in J}\sigma_{i,\varphi}.(3) A smooth one-periodic real profile VV induces the circle speed φV(u)=V(ũ)(mod⁡1)\varphi_V(u)=V(\widetilde u)\pmod1, where ũ\widetilde u is any real representative of uu.✦ A flat packet datum 𝒟=(𝒫,V;δ,l,u)\mathcal D=(\mathcal P,V;\delta,l,u) consists of a reflected packet family 𝒫\mathcal P supported in [l,u][l,u], a smooth one-periodic real profile VV, and real parameters δ,l,u\delta,l,u satisfying (4). 0≤δ<12,δ<l,u<1−δ,dVdx(x)>0(δ<x<1−δ),σ𝒫,φV=m1.0\leq\delta<\tfrac12,\qquad \delta<l,\qquad u<1-\delta,\qquad \frac{dV}{dx}(x)>0\quad(\delta<x<1-\delta),\qquad \sigma_{\mathcal P,\varphi_V}=m_1.(4) The good arc IδI_\delta of a flat packet datum 𝒟=(𝒫,V;δ,l,u)\mathcal D=(\mathcal P,V;\delta,l,u) is the arc Iδ=q((δ,1−δ))I_\delta=q((\delta,1-\delta)), with m1(𝕋\Iδ)=2δm_1(\mathbb T\setminus I_\delta)=2\delta.

Physical Coordinates and Jets

A physical coordinate frame ℱ=(χ,H,(Kk)k∈ℤ)\mathcal F=(\chi,H,(K_k)_{k\in\mathbb Z}) consists of a bijection χ:𝕋3→𝕋3\chi:\mathbb T^3\to\mathbb T^3 and smooth maps H,Kk:ℝ3→ℝ3H,K_k:\mathbb R^3\to\mathbb R^3 satisfying the projection identities (5). π∘H=χ∘π,π∘Kk=χ−1∘π(k∈ℤ).\pi\circ H=\chi\circ\pi,\qquad \pi\circ K_k=\chi^{-1}\circ\pi\quad(k\in\mathbb Z).(5)✦ For a physical coordinate frame ℱ\mathcal F, a real profile VV, real rr, and integer kk, the covering shear flow CV,rC_{V,r} and physical iterate lift Lℱ,V,kL_{\mathcal F,V,k} are the maps from ℝ3\mathbb R^3 to ℝ3\mathbb R^3 in (6). CV,r(x0,x1,x2)=(x0+rV(x1),x1,x2),Lℱ,V,k=Kk∘CV,k∘H.C_{V,r}(x_0,x_1,x_2)=(x_0+rV(x_1),x_1,x_2),\qquad L_{\mathcal F,V,k}=K_k\circ C_{V,k}\circ H.(6) A finite physical jet list Q=(kj,Ej,rj)j=1NQ=(k_j,E_j,r_j)_{j=1}^N specifies integers kjk_j, compact sets Ej⊂ℝ3E_j\subset\mathbb R^3, and derivative orders rj≥0r_j\geq0.✦ For a finite physical jet list QQ and a real tolerance ζ\zeta, physical jet closeness ∼Q,ζ\sim_{Q,\zeta} is the relation between pairs of physical frames and real profiles defined by the bounds in (7); here DrD^r denotes the iterated Fréchet derivative. ∥DrjL𝒢,W,kj(x)−DrjLℱ,V,kj(x)∥≤ζ(1≤j≤N,x∈Ej).\left\|D^{r_j}L_{\mathcal G,W,k_j}(x) -D^{r_j}L_{\mathcal F,V,k_j}(x)\right\|\leq\zeta \quad(1\leq j\leq N,\ x\in E_j).(7)✦

An admissible physical state s=(ℱ,𝒫,V;δ,l,u)s=(\mathcal F,\mathcal P,V;\delta,l,u) is a flat packet datum with δ>0\delta>0 and l<ul<u, together with a physical frame whose chart χ\chi is continuous and preserves m3m_3.✦ The physical packet function fsf_s of an admissible physical state s=(ℱ,𝒫,V;δ,l,u)s=(\mathcal F,\mathcal P,V;\delta,l,u) is the function fs=F𝒫∘χf_s=F_{\mathcal P}\circ\chi on 𝕋3\mathbb T^3, where χ\chi is the chart of ℱ\mathcal F; the same symbol denotes its class in the complex space L2(𝕋3,m3)L^2(\mathbb T^3,m_3) of square-integrable functions modulo equality almost everywhere.✦ The torus shear Rφ=Rφ1R_\varphi=R_\varphi^1 and the physical integer iterates SskS_s^k are given by (8) for every k∈ℤk\in\mathbb Z. Rφk(θ,u,a)=(θ+kφ(u),u,a),Ssk=χ−1∘RφVk∘χ.R_\varphi^k(\theta,u,a)=(\theta+k\varphi(u),u,a),\qquad S_s^k=\chi^{-1}\circ R_{\varphi_V}^k\circ\chi.(8) With the Koopman convention USg=g∘SU_Sg=g\circ S, a finite Laurent polynomial Pc(z)=∑k∈ℤckzkP_c(z)=\sum_{k\in\mathbb Z}c_kz^k, with finitely supported complex coefficients, gives the packet and physical Laurent sums in (9). ℒ𝒫,φ,c=∑k∈ℤckF𝒫∘Rφk,Pc(USs)fs=∑k∈ℤckUSskfs.\mathscr L_{\mathcal P,\varphi,c} =\sum_{k\in\mathbb Z}c_k F_{\mathcal P}\circ R_\varphi^k, \qquad P_c(U_{S_s})f_s=\sum_{k\in\mathbb Z}c_k U_{S_s}^k f_s.(9) The integer twist iterates RφkR_\varphi^k, for k∈ℤk\in\mathbb Z, are the maps of 𝕋3\mathbb T^3 that add kφ(u)k\varphi(u) to the angular coordinate and fix both slow coordinates. The physical Koopman orbit of an admissible state ss is the family (USskfs)k∈ℤ(U_{S_s}^k f_s)_{k\in\mathbb Z} in L2(𝕋3,m3)L^2(\mathbb T^3,m_3), obtained by composing its physical packet function with each physical integer iterate, including negative iterates.

A real centered physical test hh is a continuous complex function on 𝕋3\mathbb T^3 satisfying h(x)¯=h(x)\overline{h(x)}=h(x), ∫hdm3=0\int h\,dm_3=0, and |h(x)|≤1|h(x)|\leq1 at every point✦; [h][h] denotes its L2L^2 class.✦ For complex functions f,gf,g on 𝕋3\mathbb T^3, the squared physical error E(f,g)E(f,g) and physical distance D(f,g)D(f,g) are the nonnegative quantities in (10). E(f,g)=∫𝕋3|f(x)−g(x)|2dm3(x),D(f,g)=E(f,g).E(f,g)=\int_{\mathbb T^3}|f(x)-g(x)|^2\,dm_3(x),\qquad D(f,g)=\sqrt{E(f,g)}.(10)

The Claim

Theorem 1. Let ε,η,κ,ζ>0\varepsilon,\eta,\kappa,\zeta>0 and let QQ be a finite physical jet list. Let 𝒟=(𝒫,V;δ,l,u)\mathcal D=(\mathcal P,V;\delta,l,u) be a flat packet datum, ℱ\mathcal F a physical frame, and hh a real centered physical test. There are a flat packet datum (𝒲,V′;δ′,l′,u′)(\mathcal W,V';\delta',l',u'), a physical frame ℱ′\mathcal F' with chart χ′\chi', a continuous m3m_3-preserving bijection α:𝕋3→𝕋3\alpha:\mathbb T^3\to\mathbb T^3, and finitely supported coefficients cc such that 0<δ′<min⁡{18,κ},l′<u′,χ′=α∘χ,(ℱ,V)∼Q,ζ(ℱ′,V′),D(F𝒫,F𝒲∘α)<ε,D(h,ℒ𝒲,φV′,c∘α)<2δ+η.\begin{gathered} 0<\delta'<\min\{\tfrac18,\kappa\},\qquad l'<u',\qquad \chi'=\alpha\circ\chi,\qquad (\mathcal F,V)\sim_{Q,\zeta}(\mathcal F',V'),\\ D(F_{\mathcal P},F_{\mathcal W}\circ\alpha)<\varepsilon,\qquad D\bigl(h,\mathscr L_{\mathcal W,\varphi_{V'},c}\circ\alpha\bigr) <\sqrt{2\delta}+\eta. \end{gathered}(11)✦ For every admissible physical state ss, with physical frame ℱs\mathcal F_s, profile VsV_s, and parameter δs\delta_s, and every real centered physical test hh, there are an admissible physical state tt, with corresponding components ℱt,Vt,δt\mathcal F_t,V_t,\delta_t, and finitely supported coefficients cc satisfying δt<κ,(ℱs,Vs)∼Q,ζ(ℱt,Vt),∥fs−ft∥2<ε,∥[h]−Pc(USt)ft∥2<2δs+η.\begin{gathered} \delta_t<\kappa,\qquad (\mathcal F_s,V_s)\sim_{Q,\zeta}(\mathcal F_t,V_t),\qquad \|f_s-f_t\|_2<\varepsilon,\\ \|[h]-P_c(U_{S_t})f_t\|_2<\sqrt{2\delta_s}+\eta. \end{gathered}(12)✦

In the paper, the result reads: given an admissible (S,f)(S,f) with good arc II, a real smooth physical function hh with ∫hdm3=0\int h\,dm_3=0 and ∥h∥∞≤1\|h\|_\infty\leq1, and η,ρ,τ>0\eta,\rho,\tau>0, in every prescribed neighborhood of (S,S−1)(S,S^{-1}) there is an admissible (S′,f′)(S',f') with (S′,S′−1)(S',S'^{-1}) in that neighborhood, a good arc I′I', and a finite Laurent polynomial PP such that ∥f′−f∥2<τ\|f'-f\|_2<\tau, m1(𝕋\I′)<ρm_1(\mathbb T\setminus I')<\rho, and ∥h−P(US′)f′∥2<m1(𝕋\I)1/2+η\|h-P(U_{S'})f'\|_2<m_1(\mathbb T\setminus I)^{1/2}+\eta.

What Is Checked

The statements (11) and (12) concern flat packet data with an arbitrary physical frame and admissible physical states, respectively, with finite jet restrictions and continuous physical tests. Every physical integer iterate SskS_s^k preserves m3m_3.✦