OctoInformBeta

OpenAI preprint, September 2026 · §5. Local joining for the linear reserve

§5 (setting): variance-weighted reserve and pooling: A convex perspective bound for two restrictions of an interior cube field

Pooling entropy and variance in a reserve.

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

Fixing one coordinate of a cube field leaves two fields on a smaller cube. The quantity to control is their average entropy dissipation, with a correction depending on their entropy and the complements of their squared means.

For a nonnegative integer dd, the discrete cube is Ωd={−1,1}d\Omega_d=\{-1,1\}^d.✦ A real field u:Ωd→ℝu:\Omega_d\to\mathbb R has interior range if −1<u(x)<1-1<u(x)<1 at every vertex xx.✦ The uniform expectation 𝔼u\mathbb E u of a real field uu on Ωd\Omega_d is the real number given in (1). 𝔼u=2−d∑x∈Ωdu(x).\mathbb E u=2^{-d}\sum_{x\in\Omega_d}u(x).(1)✦

The binary entropy deficit ψ\psi and binary entropy HH are the functions of a real argument vv in (2), with L=log⁡2L=\log 2 and all logarithms natural. 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) The mean binary entropy e(u)e(u) and entropy Jensen gap I(u)I(u) of a real cube field uu are the real numbers given in (3). e(u)=𝔼H(u),I(u)=H(𝔼u)−e(u).e(u)=\mathbb E H(u),\qquad I(u)=H(\mathbb E u)-e(u).(3) The variance coefficient A(u)A(u) of a real cube field uu is the complement of its squared mean, as in (4).✦ A(u)=1−(𝔼u)2.A(u)=1-(\mathbb E u)^2.(4)✦

For a vertex x∈Ωdx\in\Omega_d and a coordinate index ii, let x(i)x^{(i)} be the coordinate reflection that reverses the iith sign and leaves the other signs unchanged.✦ The discrete derivative DiuD_i u of a real field uu on Ωd\Omega_d is the field, and its entropy dissipation 𝒟(u)\mathcal D(u) is the real number, given in (5). Diu(x)=u(x)−u(x(i))2,𝒟(u)=∑i=1d𝔼[atanh⁡(u)Diu].D_i u(x)=\frac{u(x)-u(x^{(i)})}{2},\qquad \mathcal D(u)=\sum_{i=1}^d\mathbb E\bigl[\operatorname{atanh}(u)D_i u\bigr].(5)

For a nonnegative integer nn, the coordinate restrictions g+g_+ and g−g_- of a real field g:Ωn+1→ℝg:\Omega_{n+1}\to\mathbb R are the fields obtained by fixing its first coordinate, as in (6). The paired mean pp and paired half-difference qq of the coordinate restrictions g+g_+ and g−g_- are the fields on Ωn\Omega_n given in (6). g+(x)=g(1,x),g−(x)=g(−1,x),p(x)=g+(x)+g−(x)2,q(x)=g+(x)−g−(x)2.\begin{aligned} g_+(x)&=g(1,x),& g_-(x)&=g(-1,x),\\ p(x)&=\frac{g_+(x)+g_-(x)}2,& q(x)&=\frac{g_+(x)-g_-(x)}2. \end{aligned}(6) The means m=𝔼pm=\mathbb E p and δ=𝔼q\delta=\mathbb E q of the paired mean pp and paired half-difference qq of gg are real numbers. The two-point entropy gap Γ(m,δ)\Gamma(m,\delta) at center mm and half-difference δ\delta is the real number given in (7). 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.(7)✦

The Reserve Correction

The inverse entropy deficit ψ−1\psi^{-1} at s∈[0,L]s\in[0,L] is the unique bias in [0,1][0,1] whose entropy deficit is ss.✦ The scalar dissipation rate FF and its dissipation excess QQ are the real functions on 0≤s<L0\le s<L given in (8). F(s)=ψ−1(s)atanh⁡(ψ−1(s)),Q(s)=F(s)−2s.F(s)=\psi^{-1}(s)\operatorname{atanh}\bigl(\psi^{-1}(s)\bigr), \qquad Q(s)=F(s)-2s.(8) The entropy-coordinate correction LLeanL_{\mathrm{Lean}} is the real function in (9). LLean(h)={Q(L−h),0<h<L,0,otherwise.L_{\mathrm{Lean}}(h)= \begin{cases} Q(L-h),&0<h<L,\\ 0,&\text{otherwise}. \end{cases}(9)✦ For positive variance coefficient AA and entropy ee, the variance-weighted correction ALLean(e/A)A L_{\mathrm{Lean}}(e/A) is the real number obtained by multiplying AA by the entropy-coordinate correction at e/Ae/A.

The Claim

Theorem 1. Let Φ:ℝ→ℝ\Phi:\mathbb R\to\mathbb R be convex on (0,∞)(0,\infty), and let A0,A1,e0,e1A_0,A_1,e_0,e_1 be strictly positive real numbers. Set A‾=(A1+A0)/2\bar A=(A_1+A_0)/2 and e‾=(e1+e0)/2\bar e=(e_1+e_0)/2. Then A‾Φ(e‾A‾)≤A1Φ(e1/A1)+A0Φ(e0/A0)2.\bar A\,\Phi\!\left(\frac{\bar e}{\bar A}\right) \le \frac{A_1\Phi(e_1/A_1)+A_0\Phi(e_0/A_0)}2.(10)✦

Let nn be a nonnegative integer and g:Ωn+1→ℝg:\Omega_{n+1}\to\mathbb R have interior range. Suppose that LLeanL_{\mathrm{Lean}} is convex on (0,∞)(0,\infty) and that the two restrictions satisfy the child bounds (11) for σ∈{+,−}\sigma\in\{+,-\}. 2I(gσ)+A(gσ)LLean(e(gσ)A(gσ))≤𝒟(gσ).2I(g_\sigma)+A(g_\sigma) L_{\mathrm{Lean}}\!\left(\frac{e(g_\sigma)}{A(g_\sigma)}\right) \le \mathcal D(g_\sigma).(11) Then 2(I(g)−Γ(m,δ))+(A(g)−δ2)LLean(e(g)A(g)−δ2)≤𝒟(g+)+𝒟(g−)2.\begin{aligned} 2\bigl(I(g)-\Gamma(m,\delta)\bigr) &+\bigl(A(g)-\delta^2\bigr) L_{\mathrm{Lean}}\!\left(\frac{e(g)}{A(g)-\delta^2}\right)\\ &\le \frac{\mathcal D(g_+)+\mathcal D(g_-)}2. \end{aligned}(12)✦

In the paper, the result reads: for information II and scalar surplus profile SS, set α=1−m2\alpha=1-m^2, β=α−δ2\beta=\alpha-\delta^2, e=H(m)−Ie=H(m)-I, and R(α,e)=αS(L−e/α)R(\alpha,e)=\alpha S(L-e/\alpha); the reserve profile is Bm(I)=2I+R(α,e)B_m(I)=2I+R(\alpha,e), and convexity pools the two child surplus terms into R(β,e)R(\beta,e).

What Is Checked

The scalar claim (10) is stated over a convex function on positive arguments with positive variance and entropy inputs; the cube claim (12) is stated over the uniform expectations and coordinate restrictions of an interior cube field satisfying (11), under convexity of LLeanL_{\mathrm{Lean}} on (0,∞)(0,\infty).