OctoInformBeta

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

Lemma 5.1 [Stationary survival and Dirichlet energy]: A survival rate bb in DD gives it norm at least e−b∏irie^{-b}\prod_i\sqrt{r_i}; rates B∑iqiB\sum_i q_i give sDsD the heat bound with constant (2B+1)/s2(2B+1)/s^2

Stationary survival bounds the compressed Gaussian kernel from below.

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

Let ι\iota be a finite set of coordinates and D⊂ℝιD\subset\mathbb{R}^\iota a measurable set of finite volume. In every coordinate ii run a stationary Gaussian autoregressive chain with correlation ri∈(0,1)r_i\in(0,1) between consecutive steps, the coordinates independent, and ask for the probability that the chain stays in DD at each of the steps 0,1,…,n0,1,\dots,n. Time is discrete. A survival rate bb of DD is a real number bb for which some real aa gives the lower bound e−a−bne^{-a-bn}, for every nn, on the probability that the chains stay in DD at each of the steps 0,1,…,n0,1,\dots,n. The question is what such a rate says about the Gaussian heat flow on DD: how much operator norm a Gaussian heat kernel keeps once it is cut down to DD, and what becomes of this when DD is dilated.

The Gaussian Chain

The autoregressive row of a real number rr is the array of reals ρr(t,k)\rho_r(t,k), indexed by a step t∈ℕt\in\mathbb{N} and a noise index k∈ℕk\in\mathbb{N}, defined by recursion on the step, with δ\delta the Kronecker delta: ρr(0,k)=δk0,ρr(t+1,k)=rρr(t,k)+1−r2δk,t+1.\rho_r(0,k)=\delta_{k0},\qquad \rho_r(t+1,k)=r\,\rho_r(t,k)+\sqrt{1-r^2}\,\delta_{k,t+1}.(1)✦ For −1<r<1-1<r<1, ρr(t,k)\rho_r(t,k) is the weight of the noise ξk\xi_k in the state XtX_t of the stationary chain X0=ξ0X_0=\xi_0, Xt=rXt−1+1−r2ξtX_t=rX_{t-1}+\sqrt{1-r^2}\,\xi_t, driven by independent standard Gaussians ξ0,ξ1,…\xi_0,\xi_1,\dots.

For correlations r∈ℝιr\in\mathbb{R}^\iota, a horizon n∈ℕn\in\mathbb{N} and a step t∈{0,…,n}t\in\{0,\dots,n\}, the autoregressive evaluation En,trE^r_{n,t} is the linear map from ℝ(n+1)×ι\mathbb{R}^{(n+1)\times\iota} to ℝι\mathbb{R}^\iota defined by (En,trx)i=∑k=0nρri(t,k)xk,i(i∈ι).(E^r_{n,t}x)_i=\sum_{k=0}^{n}\rho_{r_i}(t,k)\,x_{k,i}\qquad(i\in\iota).(2)✦ Under the standard Gaussian measure γ(n+1)×ι\gamma_{(n+1)\times\iota} on noise arrays xx, the vectors En,0rx,…,En,nrxE^r_{n,0}x,\dots,E^r_{n,n}x are the steps 00 to nn of the chains with correlations rir_i, one in each coordinate, and the survival probability of DD up to step nn is Pnr(D)=γ(n+1)×ι{x∈ℝ(n+1)×ι:En,trx∈D for t=0,1,…,n}.P^r_n(D)=\gamma_{(n+1)\times\iota}\bigl\{x\in\mathbb{R}^{(n+1)\times\iota} : E^r_{n,t}x\in D \text{ for } t=0,1,\dots,n\bigr\}.(3)

Kernels and Heat Operators

The product Gaussian density gvg_v of nonnegative weights v∈[0,∞)ιv\in[0,\infty)^\iota is the function on ℝι\mathbb{R}^\iota that multiplies the centered Gaussian densities with variances viv_i: gv(x)=∏i∈ι12πviexp⁡(−xi22vi).g_v(x)=\prod_{i\in\iota}\frac{1}{\sqrt{2\pi v_i}}\exp\Bigl(-\frac{x_i^2}{2v_i}\Bigr).(4)✦

A bounded kernel on a measurable space XX is a measurable function kk on X×XX\times X with 0≤k≤M0\le k\le M for a constant MM, its bound.✦ The difference kernel of a measurable function p≥0p\ge0 on a Euclidean space with p≤p(0)p\le p(0) is the bounded kernel kp(x,y)=p(y−x)k_p(x,y)=p(y-x) with bound p(0)p(0)✦, and the Gaussian product kernel of vv is the difference kernel of gvg_v✦.

The kernel operator KK of a bounded kernel kk on XX, for a finite measure μ\mu on XX, is the bounded linear operator on the complex space L2(μ)L^2(\mu) defined by (Kf)(x)=∫Xk(x,y)f(y)dμ(y).(Kf)(x)=\int_X k(x,y)\,f(y)\,d\mu(y).(5)✦ For a set D⊂ℝιD\subset\mathbb{R}^\iota of finite volume, KvDK^D_v is the kernel operator of the Gaussian product kernel of vv for Lebesgue measure on DD, acting on L2(D)L^2(D) by (KvDf)(x)=∫Dgv(y−x)f(y)dy(K^D_vf)(x)=\int_D g_v(y-x)\,f(y)\,dy. Norms of operators are operator norms on L2L^2.

For a measurable complex function mm on a measure space (X,μ)(X,\mu) with |m|≤C|m|\le C almost everywhere, the multiplication operator MmM_m is the bounded linear operator f↦mff\mapsto mf on L2(μ)L^2(\mu)✦; the indicator mask 𝟏D\mathbf{1}_D of a measurable set DD is the multiplication operator of its indicator function✦. The Fourier multiplier operator of a measurable real function mm on a Euclidean space EE with |m|≤1|m|\le1 is ℱ−1Mmℱ\mathcal{F}^{-1}M_m\mathcal{F} on the complex space L2(E)L^2(E), with ℱ\mathcal{F} the unitary Fourier transform.✦ The heat operator HwH_w of a measurable function w≥0w\ge0 on EE is the Fourier multiplier operator of e−we^{-w}✦, and the compressed heat operator of a measurable set D⊂ED\subset E is 𝟏DHw𝟏D\mathbf{1}_DH_w\mathbf{1}_D✦.

The Gaussian kernel symbol of weights v∈[0,∞)ιv\in[0,\infty)^\iota is the function of a frequency ξ∈ℝι\xi\in\mathbb{R}^\iota σv(ξ)=12∑i∈ιvi(2πξi)2.\sigma_v(\xi)=\frac12\sum_{i\in\iota}v_i\,(2\pi\xi_i)^2.(6)✦ A measurable set D⊂ℝιD\subset\mathbb{R}^\iota satisfies the diagonal heat bound with constant cc if, for every v∈[0,∞)ιv\in[0,\infty)^\iota, its compressed heat operator for σv\sigma_v satisfies ∥𝟏DHσv𝟏D∥≥exp⁡(−c2∑i∈ιvi).\|\mathbf{1}_DH_{\sigma_v}\mathbf{1}_D\|\ \ge\ \exp\Bigl(-\frac c2\sum_{i\in\iota}v_i\Bigr).(7)✦

The Claim

Theorem 1. Let r∈(0,1)ιr\in(0,1)^\iota, put wi=(1−ri2)/riw_i=(1-r_i^2)/r_i, and let D⊂ℝιD\subset\mathbb{R}^\iota be a measurable set of finite volume whose survival probabilities satisfy Pnr(D)≥e−a−bnP^r_n(D)\ge e^{-a-bn} for every n∈ℕn\in\mathbb{N}, for some reals aa and bb. Then ∥KwD∥≥e−b∏i∈ιri,\|K^D_w\|\ \ge\ e^{-b}\prod_{i\in\iota}\sqrt{r_i},(8)✦ and for every s>0s>0 such that the dilation sD={sx:x∈D}sD=\{sx : x\in D\} has finite volume, ∥Ks2wsD∥≥e−b∏i∈ιri.\|K^{sD}_{s^2w}\|\ \ge\ e^{-b}\prod_{i\in\iota}\sqrt{r_i}.(9)✦ Second, let s>0s>0, let BB be real, and let D⊂ℝιD\subset\mathbb{R}^\iota be a measurable set of finite volume whose dilation sDsD is measurable of finite volume. Suppose that for every q∈(0,∞)ιq\in(0,\infty)^\iota there is a real aa with Pnr(D)≥e−a−Bn∑iqiP^r_n(D)\ge e^{-a-Bn\sum_iq_i} for every n∈ℕn\in\mathbb{N}, where ri=1/(1+qi)r_i=1/(1+q_i). Then sDsD satisfies the diagonal heat bound with constant (2B+1)/s2(2B+1)/s^2: for every v∈[0,∞)ιv\in[0,\infty)^\iota, ∥𝟏sDHσv𝟏sD∥≥exp⁡(−2B+12s2∑i∈ιvi).\|\mathbf{1}_{sD}H_{\sigma_v}\mathbf{1}_{sD}\|\ \ge\ \exp\Bigl(-\frac{2B+1}{2s^2}\sum_{i\in\iota}v_i\Bigr).(10)✦

In the paper, the result reads: for a bounded nonempty open convex set DD, a positive definite diagonal matrix QQ and every T>0T>0, the stationary Ornstein-Uhlenbeck process UU satisfies ℙ{U([0,T])⊂D}≤e−TλQOU(D)\mathbb{P}\{U([0,T])\subset D\}\le e^{-T\lambda^{\mathrm{OU}}_Q(D)}, and λQ(D)≤λQOU(D)+12tr⁡Q\lambda_Q(D)\le\lambda^{\mathrm{OU}}_Q(D)+\tfrac12\operatorname{tr}Q, where λQ(D)\lambda_Q(D) is the diagonal Dirichlet energy of DD and λQOU(D)\lambda^{\mathrm{OU}}_Q(D) its Ornstein-Uhlenbeck energy.

The box answers it in discrete time, with the chain of correlations ri=1/(1+qi)r_i=1/(1+q_i) in place of UU and the diagonal heat bound in place of a bound on λQ\lambda_Q, the summand 11 in 2B+12B+1 standing for 12tr⁡Q\tfrac12\operatorname{tr}Q.

What Is Checked

The claim is stated over the autoregressive rows (1) and evaluations (2), the survival probabilities (3) under the standard Gaussian measure, the kernel operators of the Gaussian product kernel on L2(D)L^2(D) and L2(sD)L^2(sD), and the diagonal heat bound (7), built from the indicator mask, the heat operator and the symbol σv\sigma_v.

About these constructions the library proves 0≤gv(x)≤gv(0)0\le g_v(x)\le g_v(0) for the product Gaussian density and σv≥0\sigma_v\ge0 for the symbol✦. For a bounded kernel with bound MM and a finite measure μ\mu, it shows that the integral (5) of every f∈L2(μ)f\in L^2(\mu) is again in L2(μ)L^2(\mu)✦, additive in ff✦, with ∥Kf∥≤M∥1∥L2(μ)2∥f∥\|Kf\|\le M\,\|1\|_{L^2(\mu)}^2\,\|f\| for the constant function 11✦. For a measurable mm with |m|≤C|m|\le C almost everywhere, the product mfmf of every f∈L2(μ)f\in L^2(\mu) lies in L2(μ)L^2(\mu)✦, with ∥mf∥≤C∥f∥\|mf\|\le C\,\|f\|✦.