OctoInformBeta

OpenAI preprint, September 2026 · §8. Sharp entropy interpolation under noise

§8 (proof setting): what the supplied flow proof establishes: Sharp dissipation comparison for nonconstant increasing Boolean functions at correlations 0<ρ<10<\rho<1

Noise contraction and monotone rearrangement.

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

For a nonnegative integer nn, the discrete cube is Ωn={−1,1}n\Omega_n=\{-1,1\}^n.✦ A sign-valued function f:Ωn→ℝf:\Omega_n\to\mathbb R takes only the values −1-1 and 11.✦ The question concerns the information retained by such a function after independent noise in each coordinate.

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 for every coordinate ii.✦ An interior field g:Ωn→ℝg:\Omega_n\to\mathbb R takes its values in (−1,1)(-1,1).✦ The uniform cube expectation 𝔼g\mathbb E g of a real field g:Ωn→ℝg:\Omega_n\to\mathbb R is the normalized sum in (1). 𝔼g=2−n∑x∈Ωng(x).\mathbb E g=2^{-n}\sum_{x\in\Omega_n}g(x).(1)✦

For real bias vv, the binary entropy deficit ψ(v)\psi(v) and binary entropy H(v)H(v) in natural logarithmic units are the functions in (2), where L=log⁡2L=\log 2. 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).(2) For a real field gg on Ωn\Omega_n, its mean binary entropy H¯(g)\overline H(g) is the expectation of H(g)H(g)✦, and its information deficit I(g)I(g) is the entropy Jensen gap defined in (3). H¯(g)=𝔼[H(g)],I(g)=H(𝔼g)−H¯(g).\overline H(g)=\mathbb E[H(g)], \qquad I(g)=H(\mathbb E g)-\overline H(g).(3)✦

The noise kernel Kρ(x,y)K_\rho(x,y) at real correlation ρ\rho and the noise operator TρT_\rho on real fields are given by (4). For 0≤ρ≤10\leq\rho\leq1, the coordinate flip probability is ε=(1−ρ)/2\varepsilon=(1-\rho)/2. Kρ(x,y)=∏i=1n{(1+ρ)/2,xi=yi,(1−ρ)/2,xi≠yi,Tρg(x)=∑y∈ΩnKρ(x,y)g(y),Pt=Te−t(t≥0).\begin{aligned} 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}\\ T_\rho g(x)&=\sum_{y\in\Omega_n}K_\rho(x,y)g(y), \qquad P_t=T_{e^{-t}}\quad(t\geq0). \end{aligned}(4) The notation PtP_t in (4) parametrizes noise by time tt.

Dissipation and Its Scalar Profile

For x∈Ωnx\in\Omega_n and a coordinate ii, the reflected point x(i)∈Ωnx^{(i)}\in\Omega_n is obtained by reversing xix_i and leaving every other coordinate unchanged.✦ For a real field g:Ωn→ℝg:\Omega_n\to\mathbb R and a coordinate ii, the discrete derivative DigD_i g is the field given by the half-difference across reflection in coordinate ii, and the entropy dissipation D(g)D(g) is the sum of expectations in (5). Dig(x)=g(x)−g(x(i))2,D(g)=∑i=1n𝔼[atanh(g)Dig].D_i g(x)=\frac{g(x)-g(x^{(i)})}{2}, \qquad D(g)=\sum_{i=1}^n\mathbb E\!\left[ \operatorname{atanh}(g)D_i g\right].(5) The mean variance V(g)V(g) of a real field gg is the complement of its squared mean, as in (6). V(g)=1−(𝔼g)2.V(g)=1-(\mathbb E g)^2.(6)✦

For 0≤s≤L0\leq s\leq L, the inverse entropy deficit ψ−1(s)\psi^{-1}(s) is the unique a∈[0,1]a\in[0,1] with ψ(a)=s\psi(a)=s.✦ The scalar dissipation rate FF and its dissipation excess RR are the functions of the deficit coordinate ss in (7). F(s)=ψ−1(s)atanh⁡(ψ−1(s)),R(s)=F(s)−2s.F(s)=\psi^{-1}(s)\operatorname{atanh}(\psi^{-1}(s)), \qquad R(s)=F(s)-2s.(7) The monotone correction LLeanL_{\mathrm{Lean}} is the function of a real entropy coordinate hh defined in (8). The small deficit threshold s0s_0 is specified in the same display. LLean(h)={R(L−h),0<h<L,0,otherwise,s0=ψ(1/5).L_{\mathrm{Lean}}(h)= \begin{cases} R(L-h),&0<h<L,\\ 0,&\text{otherwise}, \end{cases} \qquad s_0=\psi(1/5).(8)✦

The Claim

Theorem 1. Fix the dimension nn.

  1. Suppose every increasing interior field gg on Ωn\Omega_n satisfies (9), and every interior field gg on Ωn\Omega_n with I(g)≤s0I(g)\leq s_0 satisfies (10): 2I(g)+V(g)LLean(H¯(g)V(g))≤D(g),2I(g)+V(g)L_{\mathrm{Lean}}\!\left( \frac{\overline H(g)}{V(g)}\right)\leq D(g),(9) F(I(g))≤D(g).F(I(g))\leq D(g).(10) For every nonconstant increasing sign-valued function ff on Ωn\Omega_n and every 0<ρ<10<\rho<1, the conditional noise bound is (11): I(Tρf)≤ψ(ρ).I(T_\rho f)\leq\psi(\rho).(11)✦

  2. For every nonconstant increasing sign-valued function ff on Ωn\Omega_n and every 0<ρ<10<\rho<1, the noise bound (12) holds without the dissipation assumptions of the first assertion: I(Tρf)≤ψ(ρ).I(T_\rho f)\leq\psi(\rho).(12)✦

  3. Let ff be any sign-valued function on Ωn\Omega_n and let 0≤ρ≤10\leq\rho\leq1. There exists an increasing sign-valued function gg on Ωn\Omega_n with the mean and noisy information specified in (13): 𝔼g=𝔼f,I(Tρf)≤I(Tρg).\mathbb E g=\mathbb E f, \qquad I(T_\rho f)\leq I(T_\rho g).(13)✦

In the paper, the result reads: integrating D(g)≥F(I(g))D(g)\geq F(I(g)) gives ψ−1(I(Ptg))≤e−tψ−1(I(g))\psi^{-1}(I(P_tg))\leq e^{-t}\psi^{-1}(I(g)), and continuity at zero time and reflection for negative correlation give Theorem 1.1.

What Is Checked

The conditional assertion quantifies its dissipation assumptions over interior fields on the finite uniform cube, the two noise bounds concern nonconstant increasing sign-valued inputs at 0<ρ<10<\rho<1, and rearrangement concerns arbitrary sign-valued inputs at 0≤ρ≤10\leq\rho\leq1.