OctoInformBeta

OpenAI preprint, September 2026 · §5. Uniform Gaussian survival · §5

Lemma 5.2 [One energy interval at one spectral scale]: ∥FiXt∥\|F_iX_t\| stays below a multiple of ki\sqrt{k_i} with probability at least 1/21/2 at one step per index and e−(H+1)(θn+1)e^{-(H+1)(\theta n+1)} at all steps

Survival bounds for a Gaussian chain seen through contractions.

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

Signs εi∈{−1,1}\varepsilon_i\in\{-1,1\} that keep every prefix sum ∑i<kεivi\sum_{i<k}\varepsilon_iv_i of vectors v0,…,vN−1v_0,\dots,v_{N-1} in ℝd\mathbb R^d short can be read off a Gaussian process, which must then stay in prescribed balls at many indices and times at once. Lemma 5.2 of the paper on the Euclidean Steinitz-Bergström theorem, One energy interval at one spectral scale, bounds the probability that it does.

Here the process is a stationary Gaussian chain X0,…,XnX_0,\dots,X_n in ℝm\mathbb R^m, read through real d×md\times m matrices FiF_i of operator norm at most 11, each with a size kik_i. How likely is it that every ∥FiXt∥\|F_iX_t\| stays below a fixed multiple of ki\sqrt{k_i}, when the matrices vary slowly along a parameter in [0,1][0,1] and the steps lie in a window set by a spectral scale θ\theta, and how fast does that probability decay over a long horizon? The library also answers the second question for the matrices that a filter builds at one scale.

The Gaussian Chain

Fix coefficients r∈[0,1]mr\in[0,1]^m and a horizon nn, and let the noise ξ=(ξ0,…,ξn)\xi=(\xi_0,\dots,\xi_n), with ξt∈ℝm\xi_t\in\mathbb R^m, follow the standard Gaussian measure on ℝ(n+1)×m\mathbb R^{(n+1)\times m}; every probability ℙ\mathbb P below is under this measure. The stationary Gaussian autoregressive chain X0,…,XnX_0,\dots,X_n in ℝm\mathbb R^m is given in each coordinate jj by X0,j=ξ0,j,Xt,j=rjXt−1,j+1−rj2ξt,j(1≤t≤n).X_{0,j}=\xi_{0,j},\qquad X_{t,j}=r_j\,X_{t-1,j}+\sqrt{1-r_j^2}\;\xi_{t,j}\quad(1\le t\le n).(1) The autoregressive row of rjr_j is the family of weights of the noises in Xt,jX_{t,j}, and the autoregressive evaluation at step tt is the matrix sending ξ\xi to XtX_t.

A real d×md\times m matrix FF is a contraction when ∥Fx∥≤∥x∥\|Fx\|\le\|x\| for every x∈ℝmx\in\mathbb R^m, and the squared Hilbert-Schmidt norm of a real matrix BB is the sum of the squares of its entries, ∥B∥HS2=∑i,jBij2.\|B\|_{HS}^2=\sum_{i,j}B_{ij}^2.(2)✦

The Filter at One Scale

A filter step on real 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 column jj; it acts by R↦CR+bejTR\mapsto CR+be_j^{T}, with eje_j the jjth unit vector of ℝm\mathbb R^m.✦ A filter step (C,b,j)(C,b,j) is good when CC is invertible and CCT+bbT=1.CC^{T}+bb^{T}=1.(3) A finite sequence LL of filter steps has fresh columns for a real d×md\times m matrix RR when every column of RR into which a step of LL writes is zero✦, and the run of LL on RR is the matrix obtained by applying the steps of LL to RR one after another, first to last.✦ For weights qq on the columns, the state of RR is the d×dd\times d matrix Rdiag⁡(q)RTR\,\operatorname{diag}(q)\,R^{T}.✦

A move on real d×dd\times d matrices is a pair π=(C,G)\pi=(C,G) of such matrices, its contraction CC and its gain GG; it carries a matrix DD to a matrix π(D)\pi(D) at a cost cπ(D)c_\pi(D).✦ The move of a filter step (C,b,j)(C,b,j) at the weights qq is (C,qjbbT)(C,\,q_j\,bb^{T})✦, and the energy cost of moves π0,…,πℓ−1\pi_0,\dots,\pi_{\ell-1} from DD is ℰ(D;π0,…,πℓ−1)=∑ν<ℓcπν(Dν),D0=D,Dν+1=πν(Dν).\mathcal E(D;\pi_0,\dots,\pi_{\ell-1})=\sum_{\nu<\ell}c_{\pi_\nu}(D_\nu),\qquad D_0=D,\quad D_{\nu+1}=\pi_\nu(D_\nu).(4)✦

The lift of a real d×dd\times d matrix SS along RR is the m×mm\times m matrix RTSRR^{T}SR.✦

The cutoff f0f_0 is the continuous piecewise linear function on ℝ\mathbb R with f0(x)={0,x≤1/2,2x−1,1/2<x<1,1,1≤x≤2,2−x/2,2<x<4,0,x≥4,f_0(x)=\begin{cases}0,&x\le 1/2,\\ 2x-1,&1/2<x<1,\\ 1,&1\le x\le 2,\\ 2-x/2,&2<x<4,\\ 0,&x\ge 4,\end{cases}(5)✦ and the size weight is the function φ(x)=min⁡{xa−1,x−1}\varphi(x)=\min\{x^{a-1},x^{-1}\} for the exponent a=3/2a=3/2.✦ The constants are the gain constant γ=a2a=32\gamma=a\,2^{a}=3\sqrt2✦, the increment constant I=230400(1+16γ1/a)I=230400\,(1+16\gamma^{1/a})✦, the moment constant M=I+20M=I+20✦, the epoch factor E=1+16γE=1+16\gamma✦ and the survival radius ρ=800(640M+2).\rho=800\,(640M+2).(6)✦

The filter at scale θ>0\theta>0 of RR, LL, the weights qq and a family (Pi)i∈ℐ(P_i)_{i\in\mathcal I} of prefixes of LL, indexed by a set ℐ\mathcal I, consists of real matrices FiF_i acting on ℝm\mathbb R^m, sizes kik_i and a horizon HH. The matrices FiF_i, the sizes kik_i and the horizon HH are built from the run of each PiP_i on RR and its state at the weights qθq_\theta, the weights qq rescaled by the scale θ\theta, through the lift, the cutoff, the size weight, the energy cost of the moves of PiP_i and the epoch factor.

The Claim

Theorem 1. Let r∈[0,1]mr\in[0,1]^m, let nn be a natural number and let X0,…,XnX_0,\dots,X_n be the chain (1). In (a) and (b), let (Fi)i∈ℐ(F_i)_{i\in\mathcal I} be a finite family of real d×md\times m contractions with sizes ki≥1/4k_i\ge 1/4, and let K≥0K\ge 0 and θ\theta be real numbers with ∥Fi∥HS2≤Kki\|F_i\|_{HS}^2\le Kk_i and tr⁡(Fidiag⁡(1−r)FiT)≤Kkiθ\operatorname{tr}\bigl(F_i\operatorname{diag}(1-r)F_i^{T}\bigr)\le Kk_i\theta for every ii.

(a) If steps ti∈{0,…,n}t_i\in\{0,\dots,n\} and parameters ai,bi∈[0,1]a_i,b_i\in[0,1] satisfy ∥Fi−Fj∥HS2≤K(1+ki+kj)|ai−aj|2/3\|F_i-F_j\|_{HS}^2\le K(1+k_i+k_j)\,|a_i-a_j|^{2/3} and θ|ti−tj|≤|bi−bj|\theta\,|t_i-t_j|\le|b_i-b_j| for all i,j∈ℐi,j\in\mathcal I, then ℙ{∥FiXti∥≤800(640K+2)ki for every i∈ℐ}≥1/2.\mathbb P\bigl\{\|F_iX_{t_i}\|\le 800\,(640K+2)\sqrt{k_i}\text{ for every }i\in\mathcal I\bigr\} \ge 1/2.(7)✦

(b) If θ≥0\theta\ge 0, H≥0H\ge 0 and heights hi∈[0,H]h_i\in[0,H] satisfy ∥Fi−Fj∥HS2≤K(1+ki+kj)|hi−hj|2/3\|F_i-F_j\|_{HS}^2\le K(1+k_i+k_j)\,|h_i-h_j|^{2/3} whenever ⌊hi⌋=⌊hj⌋\lfloor h_i\rfloor=\lfloor h_j\rfloor, then ℙ{∥FiXt∥≤800(640K+2)ki for every i∈ℐ and 0≤t≤n}≥e−(H+1)(θn+1).\mathbb P\bigl\{\|F_iX_t\|\le 800\,(640K+2)\sqrt{k_i}\text{ for every }i\in\mathcal I \text{ and }0\le t\le n\bigr\}\ge e^{-(H+1)(\theta n+1)}.(8)✦

(c) Let RR be a real d×md\times m matrix with orthonormal rows, RRT=1RR^{T}=1, let LL be a finite sequence of good filter steps with fresh columns for RR whose steps write into pairwise distinct columns, and let (Pi)i∈ℐ(P_i)_{i\in\mathcal I} be prefixes of LL. Let qq be positive weights on the columns with 1−rj≤qj1-r_j\le q_j for every column jj, and let θ>0\theta>0. Then the matrices FiF_i, sizes kik_i and horizon HH of the filter at scale θ\theta satisfy ℙ{∥FiXt∥≤ρki for every i∈ℐ and 0≤t≤n}≥e−(H+1)(θn+1).\mathbb P\bigl\{\|F_iX_t\|\le\rho\sqrt{k_i}\text{ for every }i\in\mathcal I \text{ and }0\le t\le n\bigr\}\ge e^{-(H+1)(\theta n+1)}.(9)✦

In the paper, the result reads: with U(τ)U(\tau) the paper’s process at time τ\tau and ℰθ(0,i)\mathcal E_\theta(0,i) its energy cost at scale θ\theta up to index ii, there is an absolute constant c1c_1 such that, if the values hi=ℰθ(0,i)h_i=\mathcal E_\theta(0,i), i∈ℐi\in\mathcal I, lie in an interval of length at most e0e_0, the paper’s fixed bound on that length, then for every u≥0u\ge 0 the probability that ∥FiU(τ)∥≤c1ki\|F_iU(\tau)\|\le c_1\sqrt{k_i} for every i∈ℐi\in\mathcal I and every u≤τ≤u+θ−1u\le\tau\le u+\theta^{-1} is at least 1/21/2. Part (a) is this statement in discrete time, the steps tit_i standing for the times τ\tau and the parameters aia_i for the energies hih_i.

What Is Checked

The claims (7) and (8) are stated over the stationary Gaussian autoregressive chain (1), through its autoregressive rows and evaluations under the standard Gaussian measure, and over contractions and the Hilbert-Schmidt norm (2); the claim (9) is stated over the same chain and over good filter steps (3) with fresh columns, their runs and states, the moves of the steps and their energy cost (4), the lift, the cutoff (5), the size weight, the epoch factor and the survival radius (6).