OctoInformBeta

OpenAI preprint, September 2026 · §1. The contraction theorem and its mechanism

§1 (setting): entropy production and the noise flow: Mean conservation and coordinate splitting for the noise flow

Entropy production under coordinate noise.

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 Ωn={−1,1}n\Omega_n=\{-1,1\}^n be the discrete cube with the uniform probability measure.✦ For a real field g:Ωn→ℝg\colon\Omega_n\to\mathbb R, the quantities of interest are its entropy Jensen gap and the rate at which that gap changes under coordinate noise. The uniform expectation of gg is the average in (1): 𝔼ng=2−n∑x∈Ωng(x).\mathbb E_n g=2^{-n}\sum_{x\in\Omega_n}g(x).(1)✦

All logarithms are natural, and L=log⁡2L=\log 2.✦ The binary entropy deficit ψ\psi is the scalar function of a real variable vv given in (2), with the convention 0log⁡0=00\log 0=0: ψ(v)=1+v2log⁡(1+v)+1−v2log⁡(1−v).\psi(v)=\frac{1+v}{2}\log(1+v)+\frac{1-v}{2}\log(1-v).(2) The binary entropy HH is the scalar function defined by (3): H(v)=L−ψ(v).H(v)=L-\psi(v).(3)✦ The mean binary entropy 𝔼nH(g)\mathbb E_n H(g) of a real field g:Ωn→ℝg\colon\Omega_n\to\mathbb R is the uniform average of its binary entropy. The information deficit I(g)I(g) of a real field gg on the discrete cube is the entropy Jensen gap in (4): I(g)=H(𝔼ng)−𝔼nH(g).I(g)=H(\mathbb E_n g)-\mathbb E_n H(g).(4)✦

For a real parameter ρ\rho and points x,y∈Ωnx,y\in\Omega_n, the noise kernel KρK_\rho on the cube is the product kernel in (5): Kρ(x,y)=∏i=1n{(1+ρ)/2,xi=yi,(1−ρ)/2,xi≠yi.K_\rho(x,y)=\prod_{i=1}^n \begin{cases} (1+\rho)/2,&x_i=y_i,\\ (1-\rho)/2,&x_i\ne y_i. \end{cases}(5) The noise operator TρT_\rho is the linear operator on real fields defined by (6): (Tρg)(x)=∑y∈ΩnKρ(x,y)g(y).(T_\rho g)(x)=\sum_{y\in\Omega_n}K_\rho(x,y)g(y).(6)✦

To split a field g:Ωn+1→ℝg\colon\Omega_{n+1}\to\mathbb R at its first coordinate, its coordinate restrictions are the fields g+,g−:Ωn→ℝg_+,g_-\colon\Omega_n\to\mathbb R given by g+(x)=g(1,x)g_+(x)=g(1,x) and g−(x)=g(−1,x)g_-(x)=g(-1,x).✦ The paired mean mm and paired half-difference δ\delta of a real field gg on Ωn+1\Omega_{n+1} are the real fields on Ωn\Omega_n defined in (7) and (8), respectively: m(x)=g+(x)+g−(x)2.m(x)=\frac{g_+(x)+g_-(x)}2.(7)✦ δ(x)=g+(x)−g−(x)2.\delta(x)=\frac{g_+(x)-g_-(x)}2.(8)✦ For a two-point field with values m+δm+\delta and m−δm-\delta, its two-point entropy gap Γ(m,δ)\Gamma(m,\delta) is the scalar quantity in (9): Γ(m,δ)=H(m)−H(m+δ)+H(m−δ)2.\Gamma(m,\delta)=H(m)-\frac{H(m+\delta)+H(m-\delta)}2.(9)✦ Writing atanh\operatorname{atanh} for the inverse hyperbolic tangent, the two-point dissipation d(m,δ)d(m,\delta) of a field with values m+δm+\delta and m−δm-\delta is the scalar quantity in (10): d(m,δ)=δ2(atanh⁡(m+δ)−atanh⁡(m−δ)).d(m,\delta)=\frac{\delta}{2} \bigl(\operatorname{atanh}(m+\delta)-\operatorname{atanh}(m-\delta)\bigr).(10)

The Claim

Theorem 1 (Noise flow and coordinate identities). For x∈Ωnx\in\Omega_n and a coordinate ii, the coordinate flip x(i)x^{(i)} is the point specified by (11): (x(i))j={−xi,j=i,xj,j≠i.(x^{(i)})_j= \begin{cases} -x_i,&j=i,\\ x_j,&j\ne i. \end{cases}(11)✦ The discrete derivative DigD_i g of a real field g:Ωn→ℝg\colon\Omega_n\to\mathbb R in coordinate ii is the field on Ωn\Omega_n given in (12): Dig(x)=g(x)−g(x(i))2.D_i g(x)=\frac{g(x)-g(x^{(i)})}{2}.(12)✦ An increasing field g:Ωn→ℝg\colon\Omega_n\to\mathbb R is a function satisfying (13) for every x,y∈Ωnx,y\in\Omega_n: g(x)≤g(y)whenever xj≤yj for every coordinate j.g(x)\le g(y)\qquad\text{whenever }x_j\le y_j\text{ for every coordinate }j.(13)✦ An interior field g:Ωn→ℝg\colon\Omega_n\to\mathbb R is a function whose values all belong to (−1,1)(-1,1).✦

The noise flow PtP_t on real fields on Ωn\Omega_n is the family of operators defined for real tt by (14): Ptg=Te−tg.P_t g=T_{e^{-t}}g.(14) The entropy dissipation D(g)D(g) of a real field g:Ωn→ℝg\colon\Omega_n\to\mathbb R is the scalar quantity defined in (15): D(g)=∑i=1n𝔼n[atanh⁡(g)Dig].D(g)=\sum_{i=1}^n\mathbb E_n\bigl[\operatorname{atanh}(g)D_i g\bigr].(15)

For every real field g:Ωn→ℝg\colon\Omega_n\to\mathbb R and every real time tt, mean conservation takes the form (16): 𝔼n(Ptg)=𝔼ng.\mathbb E_n(P_tg)=\mathbb E_n g.(16)✦ If PtgP_tg is interior at a real time tt, the information deficit along the noise flow is differentiable there, with derivative (17): ddsI(Psg)|s=t=−D(Ptg).\left.\frac{d}{ds}I(P_sg)\right|_{s=t}=-D(P_tg).(17)✦

For a real field g:Ωn+1→ℝg\colon\Omega_{n+1}\to\mathbb R, use its coordinate restrictions g+,g−g_+,g_- and paired fields m,δm,\delta. The dissipation decomposition is (18): D(g)=D(g+)+D(g−)2+𝔼nd(m,δ).D(g)=\frac{D(g_+)+D(g_-)}2+\mathbb E_n d(m,\delta).(18)✦ The information deficit chain rule is (19): I(g+)+I(g−)2=I(g)−Γ(𝔼nm,𝔼nδ).\frac{I(g_+)+I(g_-)}2 =I(g)-\Gamma(\mathbb E_n m,\mathbb E_n\delta).(19)✦

What Is Checked

The statements concern real fields on finite uniform cubes and restrictions obtained by fixing one coordinate; the derivative identity assumes an interior value of the noise flow at the specified real time.