OctoInformBeta

OpenAI preprint, September 2026 · §2. Extremal partitions and their centroids

Lemma 2.2: Centroid rank, score cones and separate bounds for small configurations

The geometry of a sparsest optimum.

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

A Gaussian partition assigns a centroid to each cell. At an energy maximum with as few positive-measure cells as possible, the geometry concerns the angles between these centroids and the regions on which each centroid has maximal linear score.

The space ℝd\mathbb R^d has its usual Euclidean inner product, and γd\gamma_d is standard Gaussian measure, with mean zero and identity covariance. A Gaussian measurable partition A=(A1,…,Ak)A=(A_1,\ldots,A_k) of ℝd\mathbb R^d is a family of Borel sets for which γd\gamma_d-almost every point belongs to exactly one cell.✦ Empty cells are allowed. The unnormalized Gaussian centroid z(E)z(E) of a set E⊆ℝdE\subseteq\mathbb R^d is the vector in ℝd\mathbb R^d, and the centroid energy F(A)F(A) of a finite labelled family A=(A1,…,Ak)A=(A_1,\ldots,A_k) of subsets of ℝd\mathbb R^d is the real number, given in (1): z(E)=∫Exdγd(x),zi=z(Ai),F(A)=∑i=1k‖zi‖2.z(E)=\int_E x\,d\gamma_d(x),\qquad z_i=z(A_i),\qquad F(A)=\sum_{i=1}^k\lVert z_i\rVert^2.(1)

An optimal partition AA is a Gaussian measurable partition that maximizes FF among Gaussian measurable partitions in the same dimension with the same number of labels; write C=F(A)C=F(A) for its value.✦ For a finite labelled family A=(A1,…,Ak)A=(A_1,\ldots,A_k) of subsets of ℝd\mathbb R^d, the active labels form the set I(A)I(A) of labels of positive Gaussian measure, their number is the active count mm, and their centroids span the subspace VV of ℝd\mathbb R^d given in (2): I(A)={i:γd(Ai)>0},m=|I(A)|,V=span⁡{zi:i∈I(A)}.I(A)=\{i:\gamma_d(A_i)>0\},\qquad m=|I(A)|,\qquad V=\operatorname{span}\{z_i:i\in I(A)\}.(2) A minimal optimal partition AA is an optimal partition whose active count is no larger than that of any optimal partition with the same dimension and labels.✦

For a finite vector family v=(vj)j∈Jv=(v_j)_{j\in J} in ℝd\mathbb R^d, the closed maximizing cone Ki(v)K_i(v) is the set on which viv_i has maximal score, as defined in (3): Ki(v)={x∈ℝd:⟨vj,x⟩≤⟨vi,x⟩ for every j∈J}.K_i(v)=\{x\in\mathbb R^d: \langle v_j,x\rangle\leq\langle v_i,x\rangle \text{ for every }j\in J\}.(3)✦

Normalized Configurations

The constant BB is defined in (4): B=322π,B2=98π.B=\frac{3}{2\sqrt{2\pi}},\qquad B^2=\frac{9}{8\pi}.(4)

For vectors z,y∈ℝdz,y\in\mathbb R^d, the negative projection coefficient α(z,y)\alpha(z,y) and orthogonal residual R(z,y)R(z,y) are the scalar and vector in (5): α(z,y)={−⟨z,y⟩/‖z‖2,z≠0,0,z=0,R(z,y)=y+α(z,y)z.\alpha(z,y)= \begin{cases} -\langle z,y\rangle/\lVert z\rVert^2,&z\ne0,\\ 0,&z=0, \end{cases} \qquad R(z,y)=y+\alpha(z,y)z.(5)

Write φ\varphi and Φ\Phi for the standard Gaussian density and distribution function on ℝ\mathbb R. The unnormalized density ρ\rho, its upper tail TT and the Gaussian normalizing constant aa are defined in (6): ρ(x)=e−x2/2,T(x)=∫x∞ρ(t)dt,a=12π,φ(x)=aρ(x).\rho(x)=e^{-x^2/2},\qquad T(x)=\int_x^\infty\rho(t)\,dt,\qquad a=\frac1{\sqrt{2\pi}},\qquad \varphi(x)=a\rho(x).(6) The inverse Mills ratio λ\lambda and the inverse Mills derivative GG, written with the paper’s argument convention, are the functions on ℝ\mathbb R given in (7); the reflected argument is x=−qx=-q: λ(x)=φ(x)Φ(−x),G(q)=φ(q)(qΦ(q)+φ(q))Φ(q)2,G(−x)=λ(x)(λ(x)−x).\lambda(x)=\frac{\varphi(x)}{\Phi(-x)},\qquad G(q)=\frac{\varphi(q)\bigl(q\Phi(q)+\varphi(q)\bigr)}{\Phi(q)^2}, \qquad G(-x)=\lambda(x)\bigl(\lambda(x)-x\bigr).(7) The Gaussian upper-tail probability pp and Gaussian loss loss\operatorname{loss} are the functions defined by (8): p(x)=aT(x)=Φ(−x),loss⁡(x)=p(x)2G(−x).p(x)=aT(x)=\Phi(-x),\qquad \operatorname{loss}(x)=p(x)^2G(-x).(8)

A normalized obtuse configuration of size kk in ℝd\mathbb R^d is a pair of families (uj)j=1k(u_j)_{j=1}^k and (Pj)j=1k(P_j)_{j=1}^k of vectors and real weights satisfying (9), (10) and (11), together with (12) when k=4k=4; write rj=‖uj‖r_j=\lVert u_j\rVert and 𝒞d,k\mathcal C_{d,k} for the set of these configurations. ∑juj=0,∑jrj2=∑jPj=1,uj≠0,ri≥rj(i≤j),⟨ui,uj⟩<0(i≠j),rj≤23,Pj≤12.\begin{gathered} \sum_j u_j=0,\qquad \sum_j r_j^2=\sum_jP_j=1,\qquad u_j\ne0,\\ r_i\geq r_j\quad(i\leq j),\qquad \langle u_i,u_j\rangle<0\quad(i\ne j),\qquad r_j\leq\frac23,\qquad P_j\leq\frac12. \end{gathered}(9) ckrj2≤Pj,0.415rj+0.15rj2≤Pj,ck={0.929,k=5,0.884,k≠5.\begin{gathered} c_k r_j^2\leq P_j,\qquad 0.415r_j+0.15r_j^2\leq P_j,\\ c_k=\begin{cases}0.929,&k=5,\\0.884,&k\ne5.\end{cases} \end{gathered}(10) ‖R(ui,uj)‖2‖R(ui,uℓ)‖2−⟨R(ui,uj),R(ui,uℓ)⟩2≤(3π)2ri2(1+α(ui,uj))2(1+α(ui,uℓ))2(i≠j,i≠ℓ).\begin{aligned} &\lVert R(u_i,u_j)\rVert^2\lVert R(u_i,u_\ell)\rVert^2 -\langle R(u_i,u_j),R(u_i,u_\ell)\rangle^2\\ &\qquad\leq \left(\frac3\pi\right)^2r_i^2 \bigl(1+\alpha(u_i,u_j)\bigr)^2 \bigl(1+\alpha(u_i,u_\ell)\bigr)^2 \qquad(i\ne j,\ i\ne\ell). \end{aligned}(11) rj≤83Pj(1−Pj),for each j there is xj≥0 such that p(xj)=Pj and B2rj21+1−43rj2≤loss⁡(xj).\begin{gathered} r_j\leq\frac83P_j(1-P_j),\\ \text{for each }j\text{ there is }x_j\geq0\text{ such that }p(x_j)=P_j \text{ and } \frac{B^2r_j^2}{1+\sqrt{1-\frac43r_j^2}} \leq\operatorname{loss}(x_j). \end{gathered}(12)✦

The Claim

Theorem 1. Let AA be a minimal optimal Gaussian partition, with active labels I(A)I(A), active count mm and active centroid span VV. For distinct i,j∈I(A)i,j\in I(A), ⟨zi,zj⟩<0.\langle z_i,z_j\rangle<0.(13)✦ If I(A)I(A) is nonempty, then dim⁡V=m−1.\dim V=m-1.(14)✦

Let (vj)j∈J(v_j)_{j\in J} be a finite family in a real inner product space, with ⟨vj,vℓ⟩<0\langle v_j,v_\ell\rangle<0 for distinct j,ℓ∈Jj,\ell\in J, and fix i∈Ji\in J. Then (vj)j∈J\{i} is linearly independent.(v_j)_{j\in J\setminus\{i\}}\text{ is linearly independent}.(15)✦ If also ∑j∈Jvj=0\sum_{j\in J}v_j=0, then (vj−vi)j∈J\{i} is linearly independent.(v_j-v_i)_{j\in J\setminus\{i\}}\text{ is linearly independent}.(16)✦ For any finite family (vj)j∈J(v_j)_{j\in J} in a real vector space whose sum is zero, and any i∈Ji\in J, span⁡{vj−vi:j∈J\{i}}=span⁡{vj:j∈J}.\operatorname{span}\{v_j-v_i:j\in J\setminus\{i\}\} =\operatorname{span}\{v_j:j\in J\}.(17)✦

If the minimal optimal partition AA has C>0C>0, then for each i∈I(A)i\in I(A), Ai=Ki((zj)j=1k)γd-almost everywhere.A_i=K_i((z_j)_{j=1}^k)\qquad\gamma_d\text{-almost everywhere}.(18)✦

For every Gaussian measurable partition QQ with k≤3k\leq3 labels, F(Q)≤B2.F(Q)\leq B^2.(19)✦

For normalized obtuse configurations in ℝd\mathbb R^d, 𝒞d,4=⌀.\mathcal C_{d,4}=\varnothing.(20)✦

In the paper, the result reads: if n≥4n\geq4, N=n+1N=n+1 and an optimal partition of ℝn\mathbb R^n with NN labels has value C>B2C>B^2 and is chosen to minimize the number of positive-measure cells, then, after its null cells are discarded and the remaining centroids are denoted z1,…,zmz_1,\ldots,z_m, distinct centroids have negative inner product, their span VV has dimension m−1m-1, every proper subfamily is linearly independent, for each ii the vectors zi−zjz_i-z_j with j≠ij\ne i form a basis of VV, its cells agree γn\gamma_n-almost everywhere with Ki((zj)j=1m)K_i((z_j)_{j=1}^m), and m≥5m\geq5.

What Is Checked

The partition assertions (13), (14), (18) and (19) are stated over Borel Gaussian partitions with fixed dimension and labels; (15) and (16) concern finite families in real inner product spaces under their stated hypotheses, (17) concerns finite zero-sum families in real vector spaces, and (20) concerns the normalized obtuse configurations defined above.