OctoInformBeta

OpenAI preprint, September 2026 · §5. Excluding five or more active cells

§5: Proof of Theorem 1.1: Probability budgets and determinant bounds allow at most four cells in a normalized obtuse configuration

Scalar constraints bound the number of active cells.

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 an optimal Gaussian partition, the active cells have nonzero centroids. Let A1,…,AmA_1,\ldots,A_m be the active cells of an optimal Gaussian partition in Euclidean space ℝd\mathbb R^d, with its usual inner product; write γd\gamma_d for standard Gaussian measure and use the notation in (1): zi=z(Ai)=∫Aixdγd(x),Pi=γd(Ai),C=∑i=1m∥zi∥2>0,ui=ziC,ri=∥ui∥.\begin{aligned} z_i&=z(A_i)=\int_{A_i}x\,d\gamma_d(x),& P_i&=\gamma_d(A_i),\\ C&=\sum_{i=1}^m\|z_i\|^2>0,& u_i&=\frac{z_i}{\sqrt C},\qquad r_i=\|u_i\|. \end{aligned}(1)

Scalar Data

A scalar probability datum of size mm is a pair of real sequences (ri,Pi)i=1m(r_i,P_i)_{i=1}^m satisfying (2), where AmA_m is the coefficient specified there: r1≥⋯≥rm>0,ri≤23,∑i=1mri2=∑i=1mPi=1,Am={0.929,m=5,0.884,m≠5,Pi≥Amri2,Pi≥0.415ri+0.15ri2(1≤i≤m).\begin{gathered} r_1\ge\cdots\ge r_m>0,\qquad r_i\le\frac23,\qquad \sum_{i=1}^m r_i^2=\sum_{i=1}^mP_i=1,\\ A_m=\begin{cases}0.929,&m=5,\\0.884,&m\ne5,\end{cases}\qquad P_i\ge A_m r_i^2,\qquad P_i\ge0.415r_i+0.15r_i^2\quad(1\le i\le m). \end{gathered}(2) For m≥3m\ge3, put t=rmt=r_m and c=r32c=r_3^2. The scalar feasible set 𝒮m\mathcal S_m is the set of scalar probability data satisfying the two determinant constraints (3): (c−t2)(r12+r22−c−2t2)≤8(3π)2t2,32(c−t2)2≤10(3π)2t2.\begin{aligned} (c-t^2)(r_1^2+r_2^2-c-2t^2)&\le8\left(\frac3\pi\right)^2t^2,\\ \frac32(c-t^2)^2&\le10\left(\frac3\pi\right)^2t^2. \end{aligned}(3)

For a real parameter AA, the scalar charge qAq_A is the quadratic polynomial in (4), and its charge threshold τA\tau_A is the scalar given there: qA(s)=0.415s−(A−0.15)s2,τA=0.415A−0.15.q_A(s)=0.415s-(A-0.15)s^2,\qquad \tau_A=\frac{0.415}{A-0.15}.(4) Budgeted ordered radii with parameter AA are real sequences (ri,δi)i=1m(r_i,\delta_i)_{i=1}^m whose radii satisfy the ordering, positivity, cap and squared normalization in (2), and whose excesses satisfy (5): δi≥0,qA(ri)≤δi(1≤i≤m),∑i=1mδi=1−A.\delta_i\ge0,\qquad q_A(r_i)\le\delta_i\quad(1\le i\le m),\qquad \sum_{i=1}^m\delta_i=1-A.(5)✦ For real parameters AA and kk, the clipping coefficient βA\beta_A and smaller budget root tA,kt_{A,k} are real numbers, and the terminal charge budget HA,kH_{A,k} is a real function, defined in (6). βA=(A−0.15)0.42τA+0.42,tA,k=0.415−0.4152−4(A−0.15)(1−A)/k2(A−0.15),HA,k(s)=qA(2.6s)+kqA(s).\begin{aligned} \beta_A&=\frac{(A-0.15)\,0.42}{\tau_A+0.42},\\ t_{A,k}&=\frac{0.415- \sqrt{0.415^2-4(A-0.15)(1-A)/k}}{2(A-0.15)},\\ H_{A,k}(s)&=q_A\!\left(\sqrt{2.6s}\right)+kq_A(s). \end{aligned}(6) An admissible scalar regime (A,k)(A,k) is a pair consisting of a real parameter AA and an integer kk satisfying (7); write a=tA,ka=t_{A,k} within that display: 0.884≤A<1,k≥2,0.066<a<0.12,kqA(a)=1−A,2.6a<τA2,(A−0.15)(0.42+a)<0.415,1−A<kqA(0.42),qA(0.066)≤qA(0.45),0<qA(0.066),1−A<βA(4τA2−1+(k−1)0.0662)+(k−1)qA(0.066),1−A<qA(0.23)+kqA(0.066),1−A<HA,k(0.066).\begin{gathered} 0.884\le A<1,\qquad k\ge2,\qquad 0.066<a<0.12,\\ kq_A(a)=1-A,\qquad 2.6a<\tau_A^2,\\ (A-0.15)(0.42+a)<0.415,\qquad 1-A<kq_A(0.42),\\ q_A(0.066)\le q_A(0.45),\qquad 0<q_A(0.066),\\ 1-A<\beta_A\bigl(4\tau_A^2-1+(k-1)0.066^2\bigr) +(k-1)q_A(0.066),\\ 1-A<q_A(0.23)+kq_A(0.066),\qquad 1-A<H_{A,k}(0.066). \end{gathered}(7)

Normalized Configurations

The unnormalized Gaussian density ρ\rho, Gaussian normalizing constant κ\kappa and unnormalized upper tail TT are the functions and constant in (8): ρ(x)=e−x2/2,κ=12π,T(x)=∫x∞ρ(y)dy.\rho(x)=e^{-x^2/2},\qquad \kappa=\frac1{\sqrt{2\pi}},\qquad T(x)=\int_x^\infty\rho(y)\,dy.(8) Let φ\varphi and Φ\Phi denote the standard Gaussian density and distribution function. The Gaussian upper tail pp, inverse Mills ratio λ\lambda, inverse Mills derivative GG in the paper’s argument, and Gaussian loss loss\operatorname{loss} are the real functions defined by (9); the reflected argument is x=−qx=-q: p(x)=κT(x)=Φ(−x),λ(x)=ρ(x)T(x)=φ(x)Φ(−x),G(q)=φ(q)(qΦ(q)+φ(q))Φ(q)2,G(−x)=λ(x)(λ(x)−x),loss⁡(x)=p(x)2G(−x).\begin{aligned} p(x)&=\kappa T(x)=\Phi(-x),& \lambda(x)&=\frac{\rho(x)}{T(x)}=\frac{\varphi(x)}{\Phi(-x)},\\ G(q)&=\frac{\varphi(q)\bigl(q\Phi(q)+\varphi(q)\bigr)}{\Phi(q)^2},& G(-x)&=\lambda(x)\bigl(\lambda(x)-x\bigr),\\ \operatorname{loss}(x)&=p(x)^2G(-x). \end{aligned}(9)

For a nonzero vector z∈ℝdz\in\mathbb R^d and y∈ℝdy\in\mathbb R^d, the negative projection coefficient a(z,y)a(z,y) is the real number and the orthogonal residual v(z,y)v(z,y) is the vector in ℝd\mathbb R^d defined by (10). For normalized vectors, use the abbreviations in the same display: a(z,y)=−⟨z,y⟩∥z∥2,v(z,y)=y+a(z,y)z,aij=a(ui,uj),wij=v(ui,uj).\begin{aligned} a(z,y)&=-\frac{\langle z,y\rangle}{\|z\|^2},& v(z,y)&=y+a(z,y)z,\\ a_{ij}&=a(u_i,u_j),&w_{ij}&=v(u_i,u_j). \end{aligned}(10) For centroids normalized as in (1), wijw_{ij} is the unnormalized residual divided by C\sqrt C.

A normalized obtuse configuration with mm labels in ℝd\mathbb R^d is a family of vectors uiu_i and real cell probabilities PiP_i whose lengths ri=∥ui∥r_i=\|u_i\| form a scalar probability datum, and which satisfy (11) and (12): ∑i=1mui=0,ui≠0,⟨ui,uj⟩<0(i≠j),Pi≤12.\sum_{i=1}^m u_i=0,\qquad u_i\ne0,\qquad \langle u_i,u_j\rangle<0\quad(i\ne j),\qquad P_i\le\frac12.(11) ∥wij∥2∥wiℓ∥2−⟨wij,wiℓ⟩2≤(3π)2ri2(1+aij)2(1+aiℓ)2(j≠i,ℓ≠i).\|w_{ij}\|^2\|w_{i\ell}\|^2-\langle w_{ij},w_{i\ell}\rangle^2 \le\left(\frac3\pi\right)^2r_i^2(1+a_{ij})^2(1+a_{i\ell})^2 \quad(j\ne i,\ \ell\ne i).(12) When m=4m=4, a normalized obtuse configuration also satisfies the four-cell cap and loss conditions (13), for every ii: ri≤83Pi(1−Pi),there is xi≥0 with p(xi)=Pi and 98πri21+1−43ri2≤loss⁡(xi).\begin{gathered} r_i\le\frac83P_i(1-P_i),\\ \text{there is }x_i\ge0\text{ with }p(x_i)=P_i\text{ and }\quad \frac{9}{8\pi}\frac{r_i^2}{1+\sqrt{1-\frac43r_i^2}} \le\operatorname{loss}(x_i). \end{gathered}(13)

The Claim

Theorem 1. Let m≥5m\ge5. For budgeted ordered radii with parameter A≥0.884A\ge0.884, the first constraint in (3) implies 0.066<t.0.066<t.(14)✦ For budgeted ordered radii in an admissible scalar regime (A,k)(A,k), suppose that 3+k≤m3+k\le m and 0.066<t0.066<t. The tail cap, minimum bound and third-radius estimate are, respectively, ri≤0.42(4≤i≤m),r_i\le0.42\qquad(4\le i\le m),(15)✦ t≤tA,k,t\le t_{A,k},(16)✦ 2.6t<r32.2.6t<r_3^2.(17)✦ For real numbers c,tc,t with 0<t<0.120<t<0.12 and 2.6t<c2.6t<c, one has 32(c−t2)2>10(3π)2t2.\frac32(c-t^2)^2>10\left(\frac3\pi\right)^2t^2.(18)✦ The scalar feasible set satisfies 𝒮m=⌀(m≥5).\mathcal S_m=\varnothing\qquad(m\ge5).(19)✦ Every normalized obtuse configuration in ℝd\mathbb R^d has m≤4.m\le4.(20)✦

In the paper, the result reads: for m≥5m\ge5, positive ordered radii r1≥⋯≥rm=tr_1\ge\cdots\ge r_m=t with ∑iri2=1\sum_i r_i^2=1 and ri≤2/3r_i\le2/3, and probabilities summing to one, cannot satisfy the three bounds of Lemma 3.1 and the two determinant constraints of Corollary 4.2, with a scalar threshold t0t_0 and the estimates t>0.066t>0.066, rj≤0.42r_j\le0.42 for j≥4j\ge4, t≤t0<0.12t\le t_0<0.12 and c>2.6tc>2.6t preceding the contradiction to the second determinant constraint.

What Is Checked

The scalar assertion (19) is stated over the finite real sequences in (2) and (3); the radius estimates (14), (15), (16) and (17) concern budgeted ordered radii and admissible scalar regimes. The determinant assertion (18) concerns two real scalars, and the cell-count assertion (20) concerns normalized obtuse configurations with the residual-pair and four-cell conditions above.