OctoInformBeta

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

§2.2 (setting): symmetric contractions (symmetric-filter) and predictor rows (predictor-rows): Symmetric steps R↦CR+bejTR\mapsto CR+be_j^{T} with C=I−ηbbTC=I-\eta bb^{T}, one per vector bib_i, and the predictor rows ηbTR\eta b^{T}R read before each step

Symmetric filter steps and their predictor rows.

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

The signed walk is built by a linear filter that takes vectors b0,…,bn−1b_0,\dots,b_{n-1} in ℝι\mathbb{R}^{\iota} one at a time, ι\iota a finite set of coordinates. The filter keeps a real matrix of coefficients: its rows are the coordinates, and its columns are a state block, a copy of ι\iota, followed by one noise coordinate for each vector. Each vector moves the matrix by one step, and just before that step the filter reads off a row of coefficients, the predictor row of the vector.

The Constants

The scale that normalizes the predictor rows is built from a chain of numerical constants. The gain constant is γ=3223/2=32\gamma=\tfrac32\,2^{3/2}=3\sqrt{2}; the epoch factor EE, the state rate α\alpha, the body rate β\beta and the body scale BB are E=1+16γ,α=84E+2,β=2α+1,B=1000(2β+1),E=1+16\gamma,\qquad \alpha=84E+2,\qquad \beta=2\alpha+1,\qquad B=1000\,(2\beta+1),(1) and the number s=800(640+2)=513600s=800\,(640+2)=513600 is the slab radius.

The Definition

Definition 1. Let ι\iota be a finite set and κ\kappa a set, and let RR be a real matrix with rows indexed by ι\iota and columns by κ\kappa. A filter step on such matrices is a triple p=(C,b,j),C∈ℝι×ι,b∈ℝι,j∈κ.p=(C,b,j),\qquad C\in\mathbb{R}^{\iota\times\iota},\quad b\in\mathbb{R}^{\iota},\quad j\in\kappa.(2)✦ A filter step p=(C,b,j)p=(C,b,j) acts on a real matrix RR with rows indexed by ι\iota and columns by κ\kappa by pR=CR+bejT,pR=CR+b\,e_j^{T},(3)✦ where bejTb\,e_j^{T}, with eje_j the jj-th standard basis vector, is the matrix whose column jj is bb and whose other columns vanish. A filter step p=(C,b,j)p=(C,b,j) is good if CC is invertible and CCT+bbT=I.CC^{T}+bb^{T}=I.(4)✦ The run of RR through a list P=(p0,…,pk−1)P=(p_0,\dots,p_{k-1}) of filter steps applies the steps in the order of the list, PR=pk−1(⋯p1(p0R)⋯),PR=p_{k-1}\bigl(\cdots p_1(p_0R)\cdots\bigr),(5)✦ and is RR itself for the empty list. The matrix RR has fresh columns for PP if every column into which a step of PP writes is zero in RR: Rej=0for every step (C,b,j) of P.Re_j=0\quad\text{for every step }(C,b,j)\text{ of }P.(6)✦ When κ\kappa is finite, the state matrix of RR for weights q∈ℝκq\in\mathbb{R}^{\kappa} on its columns is the ι×ι\iota\times\iota matrix Lq(R)=Rdiag(q)RT.L_q(R)=R\,\mathrm{diag}(q)\,R^{T}.(7)✦

The mass of a vector b∈ℝιb\in\mathbb{R}^{\iota} is its squared Euclidean norm ∥b∥2=∑a∈ιba2,\|b\|^2=\sum_{a\in\iota}b_a^2,(8)✦ its normalizer is the real number η(b)=11+1−∥b∥2,\eta(b)=\frac{1}{1+\sqrt{1-\|b\|^2}},(9)✦ and its symmetric contraction is the ι×ι\iota\times\iota matrix C(b)=I−η(b)bbT.C(b)=I-\eta(b)\,bb^{T}.(10)✦ The symmetric step of bb into a column j∈κj\in\kappa is the filter step S(b,j)=(C(b),b,j).S(b,j)=\bigl(C(b),\,b,\,j\bigr).(11)✦

The raw predictor row of bb at RR is the row vector indexed by κ\kappa m(R,b)=η(b)bTR.m(R,b)=\eta(b)\,b^{T}R.(12)✦ The predictor rows of RR along a list P=(p0,…,pk−1)P=(p_0,\dots,p_{k-1}) of filter steps, with pt=(Ct,bt,jt)p_t=(C_t,b_t,j_t), are the raw predictor rows of the vectors btb_t at the matrices their steps meet: with P<t=(p0,…,pt−1)P_{<t}=(p_0,\dots,p_{t-1}), mt(R,P)=m(P<tR,bt),t=0,…,k−1.m_t(R,P)=m\bigl(P_{<t}R,\,b_t\bigr),\qquad t=0,\dots,k-1.(13)✦

For a set ν\nu of noise coordinates, the coefficients ι⊔ν\iota\sqcup\nu are the state block ι\iota followed by the noise block ν\nu. The initial matrix on the coefficients ι⊔ν\iota\sqcup\nu is the ι×(ι⊔ν)\iota\times(\iota\sqcup\nu) matrix R0=(I0),R_0=\begin{pmatrix} I & 0 \end{pmatrix},(14)✦ The noise part y0y^{0} of a vector y∈ℝι⊔νy\in\mathbb{R}^{\iota\sqcup\nu} is the vector that agrees with yy on the noise block and vanishes on the state block.✦ For vectors b0,…,bn−1b_0,\dots,b_{n-1} in ℝι\mathbb{R}^{\iota} and ν={0,…,n−1}\nu=\{0,\dots,n-1\}, the family of symmetric steps writes each bib_i into its noise coordinate ii: Pb=(S(b0,0),S(b1,1),…,S(bn−1,n−1)).P_b=\bigl(S(b_0,0),\,S(b_1,1),\dots,S(b_{n-1},n-1)\bigr).(15)✦ The predictor row of bib_i is the noise part of the ii-th predictor row of R0R_0 along PbP_b, mˆi=(mi(R0,Pb))0.\hat m_i=\bigl(m_i(R_0,P_b)\bigr)^{0}.(16)✦ The signing scale is the real number σ=18Bs,\sigma=\frac{1}{8Bs},(17)✦ and the normalized predictor row of bib_i is the row vector m̃i=mˆi/σ.\tilde m_i=\hat m_i/\sigma.(18)✦

In the paper, the result reads: bi=σvib_i=\sigma v_i, ηi=1/(1+1−∥bi∥2)\eta_i=1/(1+\sqrt{1-\|b_i\|^2}), Ci=I−ηibibiTC_i=I-\eta_ib_ib_i^{T} and mix=ηibiTRi−1x0m_ix=\eta_ib_i^{T}R_{i-1}x^{0}, for the paper’s vectors viv_i, with Ri−1R_{i-1} the matrix that the step of bib_i meets and xx a vector of coefficients.

What Is Checked

The definitions are stated over real matrices with rows indexed by a finite set ι\iota and columns by a set κ\kappa, and, for the signed walk, over vectors b0,…,bn−1b_0,\dots,b_{n-1} in ℝι\mathbb{R}^{\iota} and the numerical constants BB and ss.