OctoInformBeta

OpenAI preprint, September 2026 · §2. Constructing the signed walk · §2

Lemma 2.3 [Adding slabs]: A slab ∥Ax∥≤s\|Ax\|\le s with ∥A∥HS≤1\|A\|_{\mathrm{HS}}\le 1 costs a factor e−1e^{-1} and rate tr⁡(AQA𝖳)\operatorname{tr}(AQA^{\mathsf T}) per step, for weights qi≥1−riq_i\ge 1-r_i

Slabs cost the Gaussian autoregressive chain little survival.

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

Take the stationary Gaussian autoregressive chain X0,…,XnX_0,\dots,X_n in ℝm\mathbb R^m, with coefficient ri∈[0,1]r_i\in[0,1] in coordinate ii, and a convex body around the origin. The probability that the chain stays inside a convex body KK around the origin for all nn steps is its survival probability. Lemma 2.3 asks how much of that probability is lost when the body is cut by slabs, sets {x:∥Ax∥≤s}\{x:\|Ax\|\le s\} for a real matrix AA whose Hilbert-Schmidt norm (the square root of the sum of the squares of its entries) is at most one and a fixed radius ss, and how the loss per step depends on weights qi≥1−riq_i\ge 1-r_i attached to the coordinates.

The Chain and Its Slabs

Write ξ=(ξ0,…,ξn)\xi=(\xi_0,\dots,\xi_n) for a point of ℝ(n+1)×m\mathbb R^{(n+1)\times m} and ℙ\mathbb P for the standard Gaussian measure on that space. The autoregressive chain with coefficients r∈[0,1]mr\in[0,1]^m up to step nn is the sequence of linear maps X0,…,XnX_0,\dots,X_n from ℝ(n+1)×m\mathbb R^{(n+1)\times m} to ℝm\mathbb R^m given coordinatewise by X0,i=ξ0,i,Xt,i=riXt−1,i+1−ri2ξt,i(1≤t≤n).X_{0,i}=\xi_{0,i},\qquad X_{t,i}=r_i\,X_{t-1,i}+\sqrt{1-r_i^2}\,\xi_{t,i}\quad(1\le t\le n).(1)✦

The Hilbert-Schmidt norm ∥A∥HS\|A\|_{\mathrm{HS}} of a real matrix AA with mm columns is the square root of the sum of the squares of its entries. The weighted square of AA by an m×mm\times m matrix WW is the number tr⁡(AWA𝖳)\operatorname{tr}(AWA^{\mathsf T}); for W=diag⁡(q)W=\operatorname{diag}(q) with q∈[0,∞)mq\in[0,\infty)^m it equals ∥Adiag⁡(q)1/2∥HS2\|A\operatorname{diag}(q)^{1/2}\|_{\mathrm{HS}}^2.✦ The slab radius is the constant s=800(640+2)=513600s=800\,(640+2)=513600.✦

Filters and the Application Domain

A filter step on d×md\times m matrices is a triple (C,b,j)(C,b,j) of a real d×dd\times d matrix CC, a vector b∈ℝdb\in\mathbb R^d and a coordinate jj of ℝm\mathbb R^m; it acts on a d×md\times m matrix RR by R↦CR+bej𝖳,R\longmapsto CR+b\,e_j^{\mathsf T},(2)✦ where eje_j is the jj-th coordinate vector of ℝm\mathbb R^m. A good step is a filter step (C,b,j)(C,b,j) with CC invertible and CC𝖳+bb𝖳=IddCC^{\mathsf T}+bb^{\mathsf T}=\mathrm{Id}_d.✦ The run of a list L=(σ1,…,σN)L=(\sigma_1,\dots,\sigma_N) of filter steps from RR is the matrix obtained from RR by applying σ1,…,σN\sigma_1,\dots,\sigma_N in this order.✦ A list LL of filter steps is fresh for a d×md\times m matrix RR when the column of RR at the coordinate jj of each step of LL is zero.✦

Two constants enter the survival estimate in the application domain Ω\Omega defined below. The body rate is β=2α+1,α=84E+2,E=1+16γ,γ=p2p=32,p=32,\beta=2\alpha+1,\qquad \alpha=84E+2,\qquad E=1+16\gamma,\qquad \gamma=p\,2^{p}=3\sqrt2,\qquad p=\tfrac32,(3) with state rate α\alpha, epoch factor EE, gain constant γ\gamma and exponent pp. The survival radius is ρ=800(640M+2),M=I+20,I=230400(1+16γ1/p),\rho=800\,(640M+2),\qquad M=I+20,\qquad I=230400\,\bigl(1+16\gamma^{1/p}\bigr),(4)✦ with moment constant MM and increment constant II, and the state radius is 4ρ4\rho✦.

For a family (Rτ)τ∈T(R_\tau)_{\tau\in T} of d×md\times m matrices and a finite family (hℓ)ℓ∈J(h_\ell)_{\ell\in J} of rows in ℝm\mathbb R^m, the application domain Ω\Omega of the two families is the set of points x∈ℝmx\in\mathbb R^m that satisfy the open norm constraints ∥Rτx∥<8ρd(τ∈T),∥x∥<8ρm,|hℓx|<2s(ℓ∈J).\|R_\tau x\|<8\rho\sqrt d\ \ (\tau\in T),\qquad \|x\|<8\rho\sqrt m,\qquad |h_\ell x|<2s\ \ (\ell\in J).(5)✦

The Claim

Theorem 1 (Adding slabs). Let q∈[0,∞)mq\in[0,\infty)^m and Q=diag⁡(q)Q=\operatorname{diag}(q).

(a) Let r∈[0,1]mr\in[0,1]^m satisfy 1−ri≤qi1-r_i\le q_i for every coordinate ii, let n∈ℕn\in\mathbb N, and let X0,…,XnX_0,\dots,X_n be the autoregressive chain with coefficients rr. If AA is a real matrix with mm columns and ∥A∥HS≤1\|A\|_{\mathrm{HS}}\le 1, then ℙ(∥AXt∥≤sfort=0,…,n)≥e−1−ntr⁡(AQA𝖳).\mathbb P\bigl(\|AX_t\|\le s\ \text{for}\ t=0,\dots,n\bigr)\ \ge\ e^{-1-n\operatorname{tr}(AQA^{\mathsf T})}.(6)✦ If (Aℓ)ℓ∈J(A_\ell)_{\ell\in J} is a finite family of real matrices with mm columns and ∥Aℓ∥HS≤1\|A_\ell\|_{\mathrm{HS}}\le 1 for every ℓ\ell, then ℙ(∥AℓXt∥≤sfor allℓ∈Jandt=0,…,n)≥e−|J|−n∑ℓ∈Jtr⁡(AℓQAℓ𝖳).\mathbb P\bigl(\|A_\ell X_t\|\le s\ \text{for all}\ \ell\in J\ \text{and}\ t=0,\dots,n\bigr) \ \ge\ e^{-|J|-n\sum_{\ell\in J}\operatorname{tr}(A_\ell QA_\ell^{\mathsf T})}.(7)✦

(b) If (hℓ)ℓ∈J(h_\ell)_{\ell\in J} is a finite family of rows in ℝm\mathbb R^m with ∑ℓ∈Jhℓ,i2≤1\sum_{\ell\in J}h_{\ell,i}^2\le 1 for every coordinate ii, then ∑ℓ∈JhℓQhℓ𝖳≤tr⁡Q.\sum_{\ell\in J}h_\ell Qh_\ell^{\mathsf T}\ \le\ \operatorname{tr}Q.(8)✦

(c) Suppose qi>0q_i>0 for every ii. Let RR be a d×md\times m matrix with orthonormal rows, let LL be a list of good filter steps that is fresh for RR and whose steps have distinct coordinates jj, and let each matrix of a family (Rτ)τ∈T(R_\tau)_{\tau\in T} be the run from RR of an initial segment of LL. Let (hℓ)ℓ∈J(h_\ell)_{\ell\in J} be a finite family of rows in ℝm\mathbb R^m with ∥hℓ∥≤1\|h_\ell\|\le 1 for every ℓ\ell and ∑ℓ∈Jhℓ,i2≤1\sum_{\ell\in J}h_{\ell,i}^2\le 1 for every coordinate ii, and let Ω\Omega be the application domain of the two families. Then there is a real number aa such that, for every n∈ℕn\in\mathbb N, the autoregressive chain X0,…,XnX_0,\dots,X_n with coefficients ri=1/(1+qi)r_i=1/(1+q_i) satisfies ℙ(Xt∈Ωfort=0,…,n)≥e−a−βntr⁡Q.\mathbb P\bigl(X_t\in\Omega\ \text{for}\ t=0,\dots,n\bigr)\ \ge\ e^{-a-\beta n\operatorname{tr}Q}.(9)✦

In the paper, the result reads: if D⊂ℝmD\subset\mathbb R^m is bounded, nonempty, open, convex and centrally symmetric, D*>0D_*>0, δ>0\delta>0, (hℓ)ℓ∈J(h_\ell)_{\ell\in J} are real rows and K=(2D*D)∩⋂ℓ∈J{x:|hℓx|<δ}K=(2D_*D)\cap\bigcap_{\ell\in J}\{x:|h_\ell x|<\delta\}, then λQ(K)≤D*−2λQ(D)+π2δ−2∑ℓ∈JhℓQhℓ𝖳\lambda_Q(K)\le D_*^{-2}\lambda_Q(D)+\pi^2\delta^{-2}\sum_{\ell\in J}h_\ell Qh_\ell^{\mathsf T} for every positive definite diagonal QQ, where λQ\lambda_Q is the paper’s diagonal Dirichlet energy. The library states the estimate for survival of the autoregressive chain, with the rate in (7) standing where the paper has its slab term.

What Is Checked

The theorem is stated over the autoregressive chain (1) under the standard Gaussian measure, the Hilbert-Schmidt norm and the weighted square, the slab radius ss, good filter steps with their runs and freshness, the application domain (5), and the constants β\beta and ρ\rho of (3) and (4).