OctoInformBeta

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

Theorem 1.3: Interior fields have a tangent-truncated correction, with an additional variance reserve under coordinatewise increase

Information deficit imposes nonlinear dissipation bounds.

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 n≥0n\geq0 be an integer and equip the discrete cube Ωn={−1,1}n\Omega_n=\{-1,1\}^n with the uniform probability measure. A real field g:Ωn→ℝg:\Omega_n\to\mathbb R is interior-valued if −1<g(x)<1-1<g(x)<1 at every vertex.✦ A real field g:Ωn→ℝg:\Omega_n\to\mathbb R is increasing if g(x)≤g(y)g(x)\leq g(y) whenever xi≤yix_i\leq y_i in every coordinate.✦ The question is how much entropy dissipation a prescribed information deficit forces, and how coordinatewise increase changes that constraint.

The uniform cube expectation 𝔼g\mathbb E g of a real field gg is its average over all vertices; write mm for the mean in (1). 𝔼g=12n∑x∈Ωng(x),m=𝔼g.\mathbb E g=\frac{1}{2^n}\sum_{x\in\Omega_n}g(x), \qquad m=\mathbb E g.(1)✦ The mean variance reserve V(g)V(g) of a real field gg on Ωn\Omega_n is the quantity in (2). V(g)=1−m2.V(g)=1-m^2.(2)✦

All logarithms are natural, and L=log⁡2L=\log2.✦ The binary entropy deficit ψ\psi and binary entropy HH at a bias v∈[−1,1]v\in[-1,1] are the functions in (3), with 0log⁡0=00\log0=0. ψ(v)=1+v2log⁡(1+v)+1−v2log⁡(1−v),H(v)=L−ψ(v).\psi(v)=\frac{1+v}{2}\log(1+v)+\frac{1-v}{2}\log(1-v), \qquad H(v)=L-\psi(v).(3) The mean binary entropy H¯(g)\overline H(g) of a field g:Ωn→[−1,1]g:\Omega_n\to[-1,1] is the uniform expectation of H(g)H(g). The information deficit I(g)I(g) of a field g:Ωn→[−1,1]g:\Omega_n\to[-1,1] is the entropy Jensen gap in (4). H¯(g)=𝔼[H(g)],I(g)=H(m)−H¯(g).\overline H(g)=\mathbb E[H(g)], \qquad I(g)=H(m)-\overline H(g).(4)

For a vertex x∈Ωnx\in\Omega_n and a coordinate i∈{1,…,n}i\in\{1,\ldots,n\}, let x(i)x^{(i)} be the vertex obtained by reversing the sign of xx in coordinate ii and leaving every other coordinate unchanged.✦ The coordinate derivative DigD_i g of a real field g:Ωn→ℝg:\Omega_n\to\mathbb R in coordinate i∈{1,…,n}i\in\{1,\ldots,n\} is the field given by half the difference across coordinate ii, as in (5). Dig(x)=g(x)−g(x(i))2.D_i g(x)=\frac{g(x)-g(x^{(i)})}{2}.(5) The entropy dissipation D(g)D(g) of an interior field gg is the sum of the coordinate averages in (6). D(g)=∑i=1n𝔼[atanh⁡(g)Dig].D(g)=\sum_{i=1}^{n}\mathbb E\bigl[\operatorname{atanh}(g)D_i g\bigr].(6)✦

The Scalar Corrections

The inverse entropy deficit ψ−1:[0,L]→[0,1]\psi^{-1}:[0,L]\to[0,1] is the inverse of ψ\psi restricted to [0,1][0,1].✦ Write a(s)=ψ−1(s)a(s)=\psi^{-1}(s). The scalar dissipation profile FF and its surplus GG are the functions on 0≤s<L0\leq s<L given by (7). F(s)=a(s)atanh⁡(a(s)),G(s)=F(s)−2s.F(s)=a(s)\operatorname{atanh}(a(s)), \qquad G(s)=F(s)-2s.(7) The profile slope on 0<s<L0<s<L has the expression in (8); its assigned value for s≤0s\leq0 is 22. F′(s)=1+a(s)(1−a(s)2)atanh⁡(a(s)),0<s<L.F'(s)=1+\frac{a(s)}{(1-a(s)^2)\operatorname{atanh}(a(s))}, \qquad 0<s<L.(8)✦ The truncation threshold s0s_0 is the entropy deficit at bias 1/51/5, as in (9). s0=ψ(1/5).s_0=\psi(1/5).(9)✦

The entropy-coordinate correction LLean:ℝ→ℝL_{\mathrm{Lean}}:\mathbb R\to\mathbb R is the surplus expressed in the entropy variable on (0,L)(0,L) and set to zero elsewhere, as in (10). LLean(h)={G(L−h),0<h<L,0,otherwise.L_{\mathrm{Lean}}(h)= \begin{cases} G(L-h),&0<h<L,\\ 0,&\text{otherwise}. \end{cases}(10)✦ The tangent-truncated dissipation correction K:ℝ→ℝK:\mathbb R\to\mathbb R is the function that vanishes for s≤0s\leq0 and has the positive branches specified in (11) and (12).✦

The Claim

Theorem 1. For 0<s≤s00<s\leq s_0, the correction satisfies K(s)=G(s).K(s)=G(s).(11)✦ For s>0s>0 with s0≤ss_0\leq s, it satisfies K(s)=G(s0)+(F′(s0)−2)(s−s0).K(s)=G(s_0)+\bigl(F'(s_0)-2\bigr)(s-s_0).(12)✦ Let n≥0n\geq0 be an integer and let g:Ωn→(−1,1)g:\Omega_n\to(-1,1). If gg is increasing, then D(g)≥2I(g)+V(g)LLean(H¯(g)V(g)).D(g)\geq2I(g)+V(g)L_{\mathrm{Lean}} \!\left(\frac{\overline H(g)}{V(g)}\right).(13)✦ For every such interior field gg, D(g)≥2I(g)+K(I(g)).D(g)\geq2I(g)+K(I(g)).(14)✦ If I(g)≤s0I(g)\leq s_0, then D(g)≥F(I(g)).D(g)\geq F(I(g)).(15)✦

In the paper, the result reads (16): for every finite n≥0n\geq0 and every g:Ωn→(−1,1)g:\Omega_n\to(-1,1), with m=𝔼gm=\mathbb E g, D(g)≥Bˆm(I(g))=max⁡{Bm(I(g)),F(I(g))}≥F(I(g)),Bm(I)=2I+(1−m2)S(L−H(m)−I1−m2),\begin{aligned} D(g)&\geq\widehat B_m(I(g)) =\max\{B_m(I(g)),F(I(g))\}\geq F(I(g)),\\ B_m(I)&=2I+(1-m^2)S\!\left(L-\frac{H(m)-I}{1-m^2}\right), \end{aligned}(16) where SS is the zero-extended surplus F−2IdF-2\operatorname{Id} and Id\operatorname{Id} denotes the identity function.

What Is Checked

The dissipation bound (14) is stated over interior real fields on every finite uniform cube, with coordinatewise increase required in (13) and the information threshold required in (15); the scalar branch identities (11) and (12) concern positive real arguments.