OctoInformBeta

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

Proposition 2.2 [Energy of a filtered path body]: Filtered states of a stationary Gaussian chain stay bounded with exponential probability, which gives the heat bound; dyadic scales control the state norm

Dilated filtered path bodies satisfy the diagonal heat bound.

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

In the construction of the signed walk, a filter carries a state y∈ℝdy\in\mathbb R^d along a point x∈ℝmx\in\mathbb R^m, reading one coordinate at a time: a step, given by a d×dd\times d matrix CC and a vector b∈ℝdb\in\mathbb R^d with CC𝖳+bb𝖳=IdCC^{\mathsf T}+bb^{\mathsf T}=I_d, replaces yy by Cy+bxjCy+b\,x_j, where the coordinate xjx_j has not been read before. The filtered path body is the subset of ℝm\mathbb R^m consisting of the points xx along which every state of the filter has norm below a fixed multiple of d\sqrt d. The question is how fast the heat semigroup on L2(ℝm)L^2(\mathbb R^m), run for a separate time in each coordinate direction and compressed to this body, can lose norm.

Filter Steps and State Matrices

A filter step is a triple (C,b,j)(C,b,j) of a d×dd\times d real matrix CC, a vector b∈ℝdb\in\mathbb R^d and a coordinate jj of ℝm\mathbb R^m; with eje_j the jj-th standard basis vector of ℝm\mathbb R^m, it acts on d×md\times m matrices by R↦CR+bej𝖳.R\longmapsto CR+b\,e_j^{\mathsf T}.(1) A filter step (C,b,j)(C,b,j) is good when CC is invertible and CC𝖳+bb𝖳=IdCC^{\mathsf T}+bb^{\mathsf T}=I_d.✦ For a d×md\times m matrix RR and filter steps (C1,b1,j1),…,(CN,bN,jN)(C_1,b_1,j_1),\dots,(C_N,b_N,j_N), the run of the steps from RR is the d×md\times m matrix obtained from RR by applying (1) with the first step, then with the second, and so on up to the last.✦ Filter steps (C1,b1,j1),…,(CN,bN,jN)(C_1,b_1,j_1),\dots,(C_N,b_N,j_N) are fresh for a d×md\times m matrix RR when the columns j1,…,jNj_1,\dots,j_N of RR vanish.✦

The state matrix of a d×md\times m matrix RR at weights q∈ℝmq\in\mathbb R^m is the d×dd\times d matrix Rdiag(q)R𝖳R\,\mathrm{diag}(q)\,R^{\mathsf T}.✦ The lift through RR of a d×dd\times d matrix MM is the m×mm\times m matrix R𝖳MRR^{\mathsf T}MR.✦ With the exponent p=3/2p=3/2, the trapezoid cutoff φ\varphi and the size weight ψ\psi are the real functions φ(x)={0,x≤12,2x−1,12<x<1,1,1≤x≤2,2−x2,2<x<4,0,x≥4,ψ(x)=min⁡(xp−1,x−1),\varphi(x)=\begin{cases}0,&x\le\tfrac12,\\ 2x-1,&\tfrac12<x<1,\\ 1,&1\le x\le2,\\ 2-\tfrac x2,&2<x<4,\\ 0,&x\ge4,\end{cases} \qquad \psi(x)=\min\bigl(x^{p-1},x^{-1}\bigr),(2) and both act on symmetric matrices through the spectral calculus.

The Gaussian Chain

Fix rates r∈[0,1]mr\in[0,1]^m and n∈ℕn\in\mathbb N, let ξ=(ξ0,…,ξn)\xi=(\xi_0,\dots,\xi_n) have the standard Gaussian law on ℝ(n+1)×m\mathbb R^{(n+1)\times m}, and write ℙ\mathbb P for probability under it. The autoregressive chain X=(X0,…,Xn)X=(X_0,\dots,X_n) with rates rr is the sequence of random points of ℝm\mathbb R^m given, coordinate by coordinate, by X0=ξ0,Xu=rXu−1+1−r2ξu(1≤u≤n),X_0=\xi_0,\qquad X_u=r\,X_{u-1}+\sqrt{1-r^2}\,\xi_u\quad(1\le u\le n),(3) and each XuX_u is a linear function of ξ\xi.

Heat Bounds and Good Bodies

Write ℱ\mathcal F for the Fourier transform on L2(ℝm)L^2(\mathbb R^m) and MgM_g for the operator of multiplication by a bounded measurable function gg.✦ For a measurable function gg with |g|≤1|g|\le1, the Fourier multiplier with symbol gg is the operator ℱ−1Mgℱ\mathcal F^{-1}M_g\mathcal F on L2(ℝm)L^2(\mathbb R^m).✦ For a weight v∈[0,∞)mv\in[0,\infty)^m, the heat operator PvP_v is the Fourier multiplier with symbol e−wve^{-w_v}✦, where wvw_v is the Gaussian symbol wv(ω)=12∑ivi(2πωi)2.w_v(\omega)=\frac12\sum_i v_i\,(2\pi\omega_i)^2 .(4)✦

For a measurable set D⊂ℝmD\subset\mathbb R^m, write 1D1_D for multiplication by its indicator✦; the compressed heat operator of DD is 1DPv1D1_DP_v1_D.✦ A measurable set D⊂ℝmD\subset\mathbb R^m satisfies the diagonal heat bound with constant cc when ∥1DPv1D∥L2→L2≥e−(c/2)∑ivifor every v∈[0,∞)m.\bigl\|1_DP_v1_D\bigr\|_{L^2\to L^2}\ge e^{-(c/2)\sum_iv_i}\quad\text{for every } v\in[0,\infty)^m.(5)✦ A set U⊂ℝmU\subset\mathbb R^m is a good body when it is nonempty, bounded, open, convex and symmetric about the origin and satisfies the diagonal heat bound with constant 10−410^{-4}.✦

The Application Body

The absolute constants are the gain constant, the epoch factor, the state rate, the body rate and the body scale, γ=p2p=32,E=1+16γ,α=84E+2,β=2α+1,Λ=1000(2β+1),\gamma=p\,2^{p}=3\sqrt2,\quad E=1+16\gamma,\quad \alpha=84E+2,\quad \beta=2\alpha+1,\quad \Lambda=1000\,(2\beta+1),(6) and the increment constant, the moment constant, the survival radius, the state radius and the slab radius, ℐ=230400(1+16γ1/p),M=ℐ+20,ρ=800(640M+2),A=4ρ,σ=800⋅642.\mathcal I=230400\,\bigl(1+16\gamma^{1/p}\bigr),\quad M=\mathcal I+20,\quad \rho=800\,(640M+2),\quad A=4\rho,\quad \sigma=800\cdot642.(7)

For d×md\times m matrices RtR_t, t∈Tt\in T, and rows hℓ∈ℝmh_\ell\in\mathbb R^m, ℓ∈J\ell\in J, the application body K⊂ℝmK\subset\mathbb R^m is the intersection of open norm constraints on the states RtxR_tx, an open ball and open slabs: K={x∈ℝm:∥Rtx∥<2Adfor t∈T,∥x∥<2Am,|⟨hℓ,x⟩|<2σfor ℓ∈J}.K=\bigl\{x\in\mathbb R^m:\ \|R_tx\|<2A\sqrt d\ \text{for } t\in T,\ \ \|x\|<2A\sqrt m,\ \ |\langle h_\ell,x\rangle|<2\sigma\ \text{for } \ell\in J\bigr\}.(8)✦

The Claim

Theorem 1.

  1. Let RR be a d×md\times m matrix with RR𝖳=IdRR^{\mathsf T}=I_d, let (C1,b1,j1),…,(CN,bN,jN)(C_1,b_1,j_1),\dots,(C_N,b_N,j_N) be good filter steps with distinct coordinates, fresh for RR, and for tt in an index set TT let Nt≤NN_t\le N be a natural number and RtR_t the run from RR of the first NtN_t of them. For weights q∈(0,∞)mq\in(0,\infty)^m and rates r∈[0,1]mr\in[0,1]^m with 1−rj≤qj1-r_j\le q_j for every jj, there is a real number aa such that, for every n∈ℕn\in\mathbb N, the chain (3) with rates rr satisfies ℙ(∥RtXu∥≤Adfor all t∈T and 0≤u≤n)≥e−a−αn∑jqj.\mathbb P\bigl(\|R_tX_u\|\le A\sqrt d\ \text{for all } t\in T \text{ and } 0\le u\le n\bigr) \ge e^{-a-\alpha n\sum_jq_j}.(9)✦

  2. Let s>0s>0, let D⊂ℝmD\subset\mathbb R^m be a measurable set of finite volume, and let BB be a real number. Suppose that for every q∈(0,∞)mq\in(0,\infty)^m there is a real number aa such that, for every n∈ℕn\in\mathbb N, the chain (3) with rates ri=1/(1+qi)r_i=1/(1+q_i) satisfies ℙ(Xu∈D for 0≤u≤n)≥e−a−Bn∑iqi\mathbb P(X_u\in D \text{ for } 0\le u\le n)\ge e^{-a-Bn\sum_iq_i}. Then the dilate sDsD satisfies the diagonal heat bound with constant (2B+1)/s2(2B+1)/s^2: ∥1sDPv1sD∥≥exp⁡(−2B+12s2∑ivi)for every v∈[0,∞)m.\bigl\|1_{sD}P_v1_{sD}\bigr\|\ge\exp\Bigl(-\frac{2B+1}{2s^2}\sum_iv_i\Bigr)\quad \text{for every } v\in[0,\infty)^m.(10)✦

  3. Let RR, the steps and the RtR_t be as in (i), and let hℓ∈ℝmh_\ell\in\mathbb R^m, ℓ∈J\ell\in J, be finitely many rows with ∥hℓ∥≤1\|h_\ell\|\le1 for every ℓ\ell and ∑ℓhℓi2≤1\sum_\ell h_{\ell i}^2\le1 for every coordinate ii. Then the dilate ΛK\Lambda K of the application body (8) is a good body, so that ∥1ΛKPv1ΛK∥≥e−∑ivi/20000for every v∈[0,∞)m.\bigl\|1_{\Lambda K}P_v1_{\Lambda K}\bigr\|\ge e^{-\sum_iv_i/20000}\quad \text{for every } v\in[0,\infty)^m.(11)✦

  4. Let RR be a d×md\times m matrix with RR𝖳=IdRR^{\mathsf T}=I_d, let q∈(0,∞)mq\in(0,\infty)^m and L=Rdiag(q)R𝖳L=R\,\mathrm{diag}(q)\,R^{\mathsf T}, and let S⊂ℤS\subset\mathbb Z be a finite set such that every eigenvalue of LL lies in [2k,2k+1][2^k,2^{k+1}] for some k∈Sk\in S. Every x∈ℝmx\in\mathbb R^m with ∥R𝖳φ(2−kL)Rx∥≤ρ(tr⁡ψ(2−kL))1/2\|R^{\mathsf T}\varphi(2^{-k}L)R\,x\|\le\rho\,\bigl(\operatorname{tr}\psi(2^{-k}L)\bigr)^{1/2} for all k∈Sk\in S satisfies ∥Rx∥≤Ad.\|Rx\|\le A\sqrt d.(12)✦

  5. For every finite family (Lt)t∈T(L_t)_{t\in T} of positive definite d×dd\times d matrices there is a finite set S⊂ℤS\subset\mathbb Z with spec⁡Lt⊂⋃k∈S[2k,2k+1](t∈T),2k≤λfor some t∈T,λ∈spec⁡Lt(k∈S).\operatorname{spec}L_t\subset\bigcup_{k\in S}\,[2^k,2^{k+1}]\ \ (t\in T),\qquad 2^k\le\lambda\ \text{for some } t\in T,\ \lambda\in\operatorname{spec}L_t\ \ (k\in S).(13)✦

In the paper, the result reads: there are absolute constants A0,B0,H0>0A_0,B_0,H_0>0 such that for d≥1d\ge1, N≥0N\ge0 and m=d+Nm=d+N, invertible CtC_t and vectors bt∈ℝdb_t\in\mathbb R^d with CtCt𝖳+btbt𝖳=IdC_tC_t^{\mathsf T}+b_tb_t^{\mathsf T}=I_d for 1≤t≤N1\le t\le N, and the maps R0x=(x1,…,xd)R_0x=(x_1,\dots,x_d) and Rtx=CtRt−1x+btxd+tR_tx=C_tR_{t-1}x+b_tx_{d+t}, the body 𝒟={x∈ℝm:|xj|<B0 for all j,∥Rtx∥<A0d for 0≤t≤N}\mathcal D=\{x\in\mathbb R^m:|x_j|<B_0\text{ for all }j,\ \|R_tx\|<A_0\sqrt d\text{ for } 0\le t\le N\} satisfies λQ(𝒟)≤H0tr⁡Q\lambda_Q(\mathcal D)\le H_0\operatorname{tr}Q for every positive definite diagonal QQ, where λQ\lambda_Q is the diagonal Dirichlet energy.

In the library, the diagonal heat bound (5) with constant cc stands for λQ(D)≤ctr⁡Q\lambda_Q(D)\le c\operatorname{tr}Q over all positive definite diagonal QQ, and (11) is drawn for the application body, which has a Euclidean ball and slabs in place of the coordinate box.

What Is Checked

The claim is stated over d×md\times m matrices RR with RR𝖳=IdRR^{\mathsf T}=I_d; good filter steps, fresh for RR, and their runs; the autoregressive chain (3) under the standard Gaussian measure; the compressed heat operators 1DPv1D1_DP_v1_D, the diagonal heat bound (5) and good bodies; the application body (8) with the constants (6) and (7); and the state matrices, their lifts through RR, the cutoff φ\varphi and the size weight ψ\psi.

The library also proves that the state matrix Rdiag(q)R𝖳R\,\mathrm{diag}(q)\,R^{\mathsf T} is positive definite whenever RR𝖳=IdRR^{\mathsf T}=I_d and q∈(0,∞)mq\in(0,\infty)^m✦, and that the Gaussian symbol wvw_v is nonnegative.✦ On L2L^2, it proves that multiplication by a measurable function that is bounded almost everywhere is an operator whose norm is at most that bound.✦