OctoInformBeta

OpenAI preprint, September 2026 · §2. Constructing the signed walk · §2.2

§2.2 (displayed): absolute parameters and the application body (application-body) to (absolute-parameters): Under conditions on filter runs and rows, state constraints, a ball and slabs |hjx|<2s|h_jx|<2s, dilated by B=1000(2β+1)B=1000(2\beta+1), form a good convex body

The body scale that makes the application body good.

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

Section 2 of the paper builds the signed walk, and its Proposition 2.1 is applied there to one convex body of coefficient vectors x∈ℝκx\in\mathbb{R}^\kappa, where κ\kappa is a finite index set of columns. State constraints cut that body out together with a coefficient ball and predictor slabs: the state constraints bound the norms ∥Rtx∥\|R_tx\| for matrices RtR_t that a filter, a sequence of steps defined below, produces from an initial matrix RR one step at a time, the ball bounds ∥x∥\|x\|, and the slabs bound |hjx||h_jx| for rows hjh_j of norm at most one.

The question is how far this body has to be dilated to become a good convex body, a notion built on heat operators compressed to the body. The item fixes that dilation, the body scale BB, along with the chain of constants behind it.

Filter Steps and Their Runs

Fix finite index sets ι\iota for rows and κ\kappa for columns. A filter step p=(C,b,j)p=(C,b,j) on real ι×κ\iota\times\kappa matrices is a triple 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, acting on a matrix RR by p(R)=CR+bej𝖳,p(R)=CR+b\,e_j^{\mathsf T},(1)✦ with eje_j the jj-th standard basis vector of ℝκ\mathbb{R}^\kappa. A filter step p=(C,b,j)p=(C,b,j) is good if CC is invertible and CC𝖳+bb𝖳=I.CC^{\mathsf T}+bb^{\mathsf T}=I.(2)✦

The run of a list L=(p1,…,pk)L=(p_1,\dots,p_k) of filter steps on a matrix RR is the matrix pk(⋯p1(R)⋯)p_k(\cdots p_1(R)\cdots); the run of the empty list is RR.✦ A list LL of filter steps has fresh columns for a matrix RR if each step (C,b,j)(C,b,j) of LL writes into a zero column of RR: Rij=0R_{ij}=0 for every row i∈ιi\in\iota.✦

The Application Body

The constants of the body are real numbers defined in a chain from the exponent p=3/2p=3/2. The gain constant γ\gamma, the epoch factor EE, the state rate α\alpha and the body rate β\beta are γ=p2p=32,E=1+16γ,α=84E+2,β=2α+1;\gamma=p\,2^{p}=3\sqrt2,\qquad E=1+16\gamma,\qquad \alpha=84E+2,\qquad \beta=2\alpha+1;(3) the increment constant II, the moment constant MM, the survival radius ρ\rho and the slab radius ss are I=230400(1+16γ1/p),M=I+20,ρ=800(640M+2),s=800(640+2)=513600;I=230400\,\bigl(1+16\,\gamma^{1/p}\bigr),\qquad M=I+20,\qquad \rho=800\,(640M+2),\qquad s=800\,(640+2)=513600;(4) the state radius is r=4ρr=4\rho✦, and the body scale is B=1000(2β+1).B=1000\,(2\beta+1).(5)✦

For matrices Rt∈ℝι×κR_t\in\mathbb{R}^{\iota\times\kappa} indexed by t∈Tt\in T and rows hj∈ℝ1×κh_j\in\mathbb{R}^{1\times\kappa} indexed by j∈Jj\in J, the application domain DD is the set of vectors xx in the coefficient space ℝκ\mathbb{R}^\kappa with ∥Rtx∥<2r|ι| for t∈T,∥x∥<2r|κ|,|hjx|<2s for j∈J.\|R_tx\|<2r\sqrt{|\iota|}\ \text{ for } t\in T,\qquad \|x\|<2r\sqrt{|\kappa|},\qquad |h_jx|<2s\ \text{ for } j\in J.(6)✦

The squared Hilbert-Schmidt norm of a real matrix AA is the sum ∥A∥HS2=∑i,jAij2\|A\|_{\mathrm{HS}}^2=\sum_{i,j}A_{ij}^2 of the squares of its entries.✦

Good Convex Bodies

For a measurable function mm on ℝκ\mathbb{R}^\kappa with |m|≤C|m|\le C almost everywhere, the multiplication operator MmM_m is the bounded operator f↦mff\mapsto mf on the complex space L2(ℝκ)L^2(\mathbb{R}^\kappa).✦ For a measurable real function mm with |m|≤1|m|\le1, the Fourier multiplier operator with multiplier mm is ℱ−1Mmℱ\mathcal{F}^{-1}M_m\mathcal{F}, with ℱ\mathcal{F} the unitary Fourier transform.✦ For a measurable function w≥0w\ge0, the heat operator ℋw\mathcal{H}_w is the Fourier multiplier operator with multiplier e−we^{-w}✦, and for a measurable set K⊆ℝκK\subseteq\mathbb{R}^\kappa the compressed heat operator is M𝟏KℋwM𝟏KM_{\mathbf 1_K}\mathcal{H}_wM_{\mathbf 1_K}.

For a weight v∈[0,∞)κv\in[0,\infty)^\kappa, the Gaussian kernel symbol is the function φv(ξ)=12∑i∈κvi(2πξi)2,ξ∈ℝκ.\varphi_v(\xi)=\tfrac12\sum_{i\in\kappa}v_i\,(2\pi\xi_i)^2,\qquad \xi\in\mathbb{R}^\kappa.(7)✦ A measurable set K⊆ℝκK\subseteq\mathbb{R}^\kappa satisfies the diagonal heat bound with constant cc if, for every weight v∈[0,∞)κv\in[0,\infty)^\kappa, the operator norm on L2L^2 obeys ∥M𝟏KℋφvM𝟏K∥≥e−(c/2)∑i∈κvi.\bigl\|M_{\mathbf 1_K}\mathcal{H}_{\varphi_v}M_{\mathbf 1_K}\bigr\|\ \ge\ e^{-(c/2)\sum_{i\in\kappa}v_i}.(8)✦ A good convex body is a set KK in the coefficient space ℝκ\mathbb{R}^\kappa that is nonempty, open, bounded and convex, with −x∈K-x\in K for every x∈Kx\in K, and that satisfies the diagonal heat bound with constant c=1/10000c=1/10000.

The Claim

Theorem 1. Let RR be a real ι×κ\iota\times\kappa matrix with RR𝖳=IRR^{\mathsf T}=I, and let LL be a list of good filter steps that has fresh columns for RR and in which no two steps write into the same column. For each tt in an index set TT, let RtR_t be the run on RR of a prefix PtP_t of LL. Let hj∈ℝ1×κh_j\in\mathbb{R}^{1\times\kappa}, for jj in a finite set JJ, be rows with ∥hj∥HS2≤1\|h_j\|_{\mathrm{HS}}^2\le1 and with ∑j∈J(hj)i2≤1\sum_{j\in J}(h_j)_i^2\le1 for every column i∈κi\in\kappa, and let DD be the application domain of the matrices RtR_t and the rows hjh_j. Then the dilation of DD by the body scale, BD={Bx:x∈D},B\,D=\{\,Bx : x\in D\,\},(9)✦ is a good convex body, and the body scale satisfies 2β+1B2≤110000.\frac{2\beta+1}{B^2}\le\frac{1}{10000}.(10)✦

In the paper, the result reads: the body K=(2D*𝒟)∩{x:|mix|<δ}K=(2D_*\mathcal{D})\cap\{x:|m_ix|<\delta\}, built from a domain 𝒟\mathcal{D}, rows mim_i and a width δ\delta, satisfies the all-diagonal energy bound λQ(K)≤tr⁡Q/10000\lambda_Q(K)\le\operatorname{tr}Q/10000 for the diagonal Dirichlet energy λQ\lambda_Q and every positive definite diagonal QQ, the constants D*D_* and σ\sigma being fixed so that it does.

What Is Checked

The claim is stated over filter steps, their goodness (2), their runs and fresh columns; over the application domain (6), with its open norm constraints, the state radius rr, the slab radius ss and the Hilbert-Schmidt norm of its rows; over the body scale (5) and the constants (3) and (4) behind it; and over good convex bodies, defined through the diagonal heat bound (8) on compressed heat operators.

About these constructions the library also proves that the Gaussian kernel symbol φv\varphi_v is nonnegative at every frequency✦, and that for a measurable function mm with |m|≤C|m|\le C almost everywhere and every f∈L2f\in L^2, the product mfmf lies in L2L^2✦ and ∥mf∥≤C∥f∥\|mf\|\le C\|f\|.✦