OctoInformBeta

OpenAI preprint, September 2026 · §5. Uniform Gaussian survival · §5.3

§5.3 (proof of Proposition 2.2): combining the constraints: Filtered Gaussian autoregressive states stay in a ball of radius 4ρ|ι|4\rho\sqrt{|\iota|} for nn steps with probability at least e−a−βn∑jqje^{-a-\beta n\sum_j q_j}

Dyadic scales and the survival of filtered Gaussian states.

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 finite index sets ι\iota and κ\kappa and a real matrix RR with rows indexed by ι\iota and columns by κ\kappa whose rows are orthonormal, RRT=IdRR^{T}=\mathrm{Id}. Run an independent stationary Gaussian autoregressive chain with coefficient rjr_j in each column jj, and read it through RR: with Xu∈ℝκX_u\in\mathbb{R}^{\kappa} the chains at step uu, the filtered state at step uu is RXu∈ℝιRX_u\in\mathbb{R}^{\iota}. A filter changes RR one step at a time, each step writing into a fresh column. The question is how likely it is that every filtered state, at every stage of the filter and every step u≤nu\le n up to a horizon nn, stays in a ball of radius a constant times |ι|\sqrt{|\iota|}, and how fast this survival probability can decay in nn.

The survival rate of the filtered states comes from combining constraints at finitely many dyadic scales. The library states the survival bound and, before it, two results on these scales: their choice, and the reconstruction of a bound on ∥Rx∥\|Rx\|, for a vector x∈ℝκx\in\mathbb{R}^{\kappa}, from bounds at each scale.

Filters and Their States

For a real ι×κ\iota\times\kappa matrix RR and weights q∈ℝκq\in\mathbb{R}^{\kappa}, the state matrix of qq and RR is the ι×ι\iota\times\iota matrix Rdiag(q)RT,R\,\mathrm{diag}(q)\,R^{T},(1)✦ the paper’s LtL_t. The lifted matrix of a real ι×ι\iota\times\iota matrix AA along RR is the κ×κ\kappa\times\kappa matrix RTARR^{T}AR.✦

A filter step is a triple (C,b,j)(C,b,j) of a real ι×ι\iota\times\iota matrix CC, a vector b∈ℝιb\in\mathbb{R}^{\iota} and a column j∈κj\in\kappa; it acts on real ι×κ\iota\times\kappa matrices by R↦CR+bejT,R\longmapsto CR+b\,e_j^{T},(2)✦ where eje_j is the jjth standard basis vector of ℝκ\mathbb{R}^{\kappa}. A filter step (C,b,j)(C,b,j) is good when CC is invertible and CCT+bbT=IdCC^{T}+bb^{T}=\mathrm{Id}.✦ A finite sequence FF of filter steps has fresh columns for RR when, for every step of FF, the column jj it writes into is zero in RR.✦ The run RPR_P of a finite sequence PP of filter steps on RR is the matrix obtained by applying the steps of PP to RR in order, first to last; the empty sequence leaves RR unchanged.✦ An initial segment of a finite sequence of filter steps is a sequence formed by its first steps, in order, possibly none.

Dyadic Scales

The trapezoid cutoff f0f_0 and the size weight ww are the real functions f0(x)={0,x≤12,2x−1,12<x<1,1,1≤x≤2,2−x2,2<x<4,0,x≥4,w(x)=min⁡{x1/2,x−1}.f_0(x)=\begin{cases} 0, & x\le\tfrac12,\\ 2x-1, & \tfrac12<x<1,\\ 1, & 1\le x\le 2,\\ 2-\tfrac{x}{2}, & 2<x<4,\\ 0, & x\ge 4, \end{cases} \qquad w(x)=\min\bigl\{x^{1/2},\,x^{-1}\bigr\}.(3) For a real ι×κ\iota\times\kappa matrix RR, weights q∈ℝκq\in\mathbb{R}^{\kappa} and an integer kk, the state matrix at scale 2k2^{k} is Lk=Rdiag(2−kq)RTL_k=R\,\mathrm{diag}(2^{-k}q)\,R^{T}, and f0(Lk)f_0(L_k) and w(Lk)w(L_k) are the matrices given by the functional calculus.

The Autoregressive Chain

The autoregressive row of a real coefficient rr is the array Φr(t,s)\Phi_r(t,s), over steps tt and noise indices ss in ℕ\mathbb{N}, defined by Φr(0,s)=𝟏{s=0},Φr(t+1,s)=rΦr(t,s)+1−r2𝟏{s=t+1}.\Phi_r(0,s)=\mathbf{1}_{\{s=0\}},\qquad \Phi_r(t+1,s)=r\,\Phi_r(t,s)+\sqrt{1-r^{2}}\;\mathbf{1}_{\{s=t+1\}}.(4)✦ For −1≤r≤1-1\le r\le 1 and independent standard Gaussian noises ξs\xi_s, Φr(t,s)\Phi_r(t,s) is the weight of ξs\xi_s in the state XtX_t of the stationary Gaussian autoregressive chain X0=ξ0X_0=\xi_0, Xt=rXt−1+1−r2ξtX_t=rX_{t-1}+\sqrt{1-r^{2}}\,\xi_t.

For coefficients r∈ℝκr\in\mathbb{R}^{\kappa}, a horizon nn and a step u∈{0,…,n}u\in\{0,\dots,n\}, the autoregressive evaluation Λu\Lambda_u is the real matrix that sends a noise array x=(xs,j)x=(x_{s,j}) in ℝ{0,…,n}×κ\mathbb{R}^{\{0,\dots,n\}\times\kappa} to the states at step uu of these chains, one with coefficient rjr_j in each coordinate jj: (Λux)j=∑s=0nΦrj(u,s)xs,j(j∈κ).(\Lambda_u x)_j=\sum_{s=0}^{n}\Phi_{r_j}(u,s)\,x_{s,j}\qquad(j\in\kappa).(5)✦

The Constants

The bounds use absolute constants built from the exponent 3/23/2: the gain constant γ\gamma, the epoch factor EE, the increment constant II, the moment constant MM, the survival radius ρ\rho and the state rate β\beta, given by γ=32⋅23/2=32,E=1+16γ,I=230400(1+16γ2/3),M=I+20,\gamma=\tfrac32\cdot 2^{3/2}=3\sqrt{2},\qquad E=1+16\gamma,\qquad I=230400\,\bigl(1+16\gamma^{2/3}\bigr),\qquad M=I+20,(6) ρ=800(640M+2),β=84E+2,\rho=800\,(640M+2),\qquad \beta=84E+2,(7) and the state radius is 4ρ4\rho✦.

The Claim

Theorem 1 (Finite set of dyadic scales). Let TT be a finite set and let AtA_t, for t∈Tt\in T, be positive definite real ι×ι\iota\times\iota matrices with eigenvalues λi(At)\lambda_i(A_t), i∈ιi\in\iota. Then there is a finite set SS of integers such that ⋃t∈T{λi(At):i∈ι}⊆⋃k∈S[2k,2k+1],\bigcup_{t\in T}\{\lambda_i(A_t):i\in\iota\}\subseteq\bigcup_{k\in S}\,[2^{k},2^{k+1}],(8)✦ and such that every k∈Sk\in S has 2k≤λi(At)2^{k}\le\lambda_i(A_t) for some t∈Tt\in T and i∈ιi\in\iota.

Theorem 2 (Reconstruction bound). Let RR be a real ι×κ\iota\times\kappa matrix with RRT=IdRR^{T}=\mathrm{Id}, let q∈ℝκq\in\mathbb{R}^{\kappa} have qj>0q_j>0 for every jj, and let SS be a finite set of integers such that every eigenvalue of the state matrix Rdiag(q)RTR\,\mathrm{diag}(q)\,R^{T} lies in [2k,2k+1][2^{k},2^{k+1}] for some k∈Sk\in S. If a vector x∈ℝκx\in\mathbb{R}^{\kappa} satisfies ∥RTf0(Lk)Rx∥≤ρtr⁡w(Lk)for every k∈S,\bigl\|R^{T}f_0(L_k)R\,x\bigr\|\le\rho\sqrt{\operatorname{tr}w(L_k)}\qquad\text{for every }k\in S,(9) then ∥Rx∥≤4ρ|ι|.\|Rx\|\le 4\rho\sqrt{|\iota|}.(10)✦

Theorem 3 (Survival of the filtered states). Let RR be a real ι×κ\iota\times\kappa matrix with RRT=IdRR^{T}=\mathrm{Id}, let FF be a finite sequence of good filter steps with fresh columns for RR and with pairwise distinct columns, let PtP_t, for tt in an index set TT, be initial segments of FF, and let q,r∈ℝκq,r\in\mathbb{R}^{\kappa} satisfy qj>0q_j>0, 0≤rj≤10\le r_j\le 1 and 1−rj≤qj1-r_j\le q_j for every j∈κj\in\kappa. Then there is a real number aa such that, for every n∈ℕn\in\mathbb{N}, a standard Gaussian array xx in ℝ{0,…,n}×κ\mathbb{R}^{\{0,\dots,n\}\times\kappa} and the autoregressive evaluations Λu\Lambda_u with coefficients rr satisfy ℙ(∥RPtΛux∥≤4ρ|ι| for all t∈T and 0≤u≤n)≥e−a−βn∑j∈κqj.\mathbb{P}\Bigl(\bigl\|R_{P_t}\Lambda_u x\bigr\|\le 4\rho\sqrt{|\iota|} \ \text{ for all } t\in T \text{ and } 0\le u\le n\Bigr) \ \ge\ e^{-a-\beta n\sum_{j\in\kappa}q_j}.(11)✦

What Is Checked

The statements (8), (10) and (11) are stated over finite families of positive definite matrices and their eigenvalues; real matrices RR with RRT=IdRR^{T}=\mathrm{Id}, their state matrices (1), the matrices LkL_k under the functional calculus of the trapezoid cutoff and the size weight (3), and the lifted matrices RTf0(Lk)RR^{T}f_0(L_k)R; good filter steps (2) with fresh, pairwise distinct columns and their runs along initial segments; the autoregressive evaluations (5) built from the rows (4), under the standard Gaussian measure; and the constants (6) and (7).

About these constructions the library proves that the state matrix Rdiag(q)RTR\,\mathrm{diag}(q)\,R^{T} of a real matrix RR with RRT=IdRR^{T}=\mathrm{Id} and weights qj>0q_j>0 is positive definite.✦