OctoInformBeta

OpenAI preprint, September 2026 · §4. Competition with two other cells

Lemma 4.1: Normalized Gram determinants with constant 3/π3/\pi, including degenerate residual pairs

Competition bounds residual area at an 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

At a fixed cell of a Gaussian partition, competition with two other cells gives two residual centroid vectors. Their Gram determinant measures the squared area of the parallelogram they span.

The space ℝd\mathbb R^d has its usual Euclidean inner product. Its standard Gaussian measure γd\gamma_d is the probability measure with mean zero and identity covariance.✦ The Gaussian centroid z(K)z(K) of a set K⊆ℝdK\subseteq\mathbb R^d is the unnormalized vector in (1); the centroid energy F(A)F(A) of a family A=(A1,…,Ak)A=(A_1,\ldots,A_k) is the sum in the same display. z(K)=∫Kxdγd(x),zi=z(Ai),F(A)=∑i=1k∥zi∥2.z(K)=\int_K x\,d\gamma_d(x),\qquad z_i=z(A_i),\qquad F(A)=\sum_{i=1}^k\|z_i\|^2.(1)

A Gaussian measurable partition AA of ℝd\mathbb R^d is a finite family of Borel sets such that γd\gamma_d-almost every point belongs to exactly one cell.✦ Empty cells are allowed. For a finite family AA, its cell probabilities PiP_i, active label set I(A)I(A), and number m(A)m(A) of active cells are defined by (2). Pi=γd(Ai),I(A)={i:Pi>0},m(A)=|I(A)|.P_i=\gamma_d(A_i),\qquad I(A)=\{i:P_i>0\},\qquad m(A)=|I(A)|.(2)

An optimal partition AA is a Gaussian measurable partition whose centroid energy is at least that of every Gaussian measurable partition in the same dimension with the same number of labels.✦ A minimal optimal partition AA is an optimal partition for which m(A)m(A) is no larger than the number of active cells of any other optimal partition with those labels.✦ Write C=F(A)C=F(A) for its value.

Residual Vectors and Normalization

For vectors z,y∈ℝdz,y\in\mathbb R^d, the negative projection coefficient a(z,y)a(z,y) and the unnormalized orthogonal residual R(z,y)R(z,y) are defined by (3), with a(0,y)=0a(0,y)=0. a(z,y)=−⟨z,y⟩∥z∥2,R(z,y)=y+a(z,y)z.a(z,y)=-\frac{\langle z,y\rangle}{\|z\|^2},\qquad R(z,y)=y+a(z,y)z.(3)

The Gram determinant of vectors u,v∈ℝdu,v\in\mathbb R^d is the quantity Δ(u,v)\Delta(u,v) in (4). Δ(u,v)=∥u∥2∥v∥2−⟨u,v⟩2.\Delta(u,v)=\|u\|^2\|v\|^2-\langle u,v\rangle^2.(4)

For a partition of positive value CC, its normalized centroids zˆi\widehat z_i and their lengths rir_i are given by (5); the constant BB sets the energy threshold. zˆi=ziC,ri=∥zi∥C,∑iri2=∑iPi=1,B=322π,B2=98π.\widehat z_i=\frac{z_i}{\sqrt C},\qquad r_i=\frac{\|z_i\|}{\sqrt C},\qquad \sum_i r_i^2=\sum_i P_i=1,\qquad B=\frac{3}{2\sqrt{2\pi}},\qquad B^2=\frac{9}{8\pi}.(5)

Fix a reference index ii. The coefficients aja_j, residuals RjR_j, and normalized residuals wjw_j relative to ziz_i are defined in (6), together with their normalized pair area DjℓD_{j\ell} and the reference length tt. aj=a(zi,zj),Rj=R(zi,zj),wj=RjC,t=ri,Djℓ=Δ(wj,wℓ).\begin{gathered} a_j=a(z_i,z_j),\qquad R_j=R(z_i,z_j),\qquad w_j=\frac{R_j}{\sqrt C},\qquad t=r_i,\\ D_{j\ell}=\sqrt{\Delta(w_j,w_\ell)}. \end{gathered}(6)

The Claim

Theorem 1 (Residual pair bounds). Let K⊆ℝdK\subseteq\mathbb R^d be Borel measurable, with nonzero centroid z=z(K)z=z(K). Let r,s∈ℝdr,s\in\mathbb R^d be orthogonal to zz, and let L,M>0L,M>0. Suppose that every x∈Kx\in K satisfies ⟨r,x⟩≤L⟨z,x⟩\langle r,x\rangle\le L\langle z,x\rangle and ⟨s,x⟩≤M⟨z,x⟩\langle s,x\rangle\le M\langle z,x\rangle. If Δ(r,s)>0\Delta(r,s)>0, the Gaussian residual area bound is (7). Δ(r,s)≤9∥z∥LM2π2π.\sqrt{\Delta(r,s)} \le \frac{9\|z\|LM}{2\pi\sqrt{2\pi}}.(7)✦

Let AA be a minimal optimal partition of positive value CC. Let i,j,ℓi,j,\ell be active indices with j≠ij\ne i and ℓ≠i\ell\ne i. The unnormalized residual pair satisfies (8), including when its Gram determinant is zero. Δ(Rj,Rℓ)≤9∥zi∥(1+aj)(1+aℓ)2π2π.\sqrt{\Delta(R_j,R_\ell)} \le \frac{9\|z_i\|(1+a_j)(1+a_\ell)}{2\pi\sqrt{2\pi}}.(8)✦

If in addition C>B2C>B^2, the normalized residual area satisfies (9). Djℓ≤3πri(1+aj)(1+aℓ).D_{j\ell}\le\frac{3}{\pi}\,r_i(1+a_j)(1+a_\ell).(9)✦

For a minimal optimal partition with a nonempty label set, all cells active, and C>B2C>B^2, put u=R(zˆi,zˆj)u=R(\widehat z_i,\widehat z_j) and v=R(zˆi,zˆℓ)v=R(\widehat z_i,\widehat z_\ell). For j≠ij\ne i and ℓ≠i\ell\ne i, the squared normalized area bound is (10). Δ(u,v)≤(3π)2∥zˆi∥2(1+a(zˆi,zˆj))2(1+a(zˆi,zˆℓ))2.\Delta(u,v)\le \left(\frac{3}{\pi}\right)^2\|\widehat z_i\|^2 \bigl(1+a(\widehat z_i,\widehat z_j)\bigr)^2 \bigl(1+a(\widehat z_i,\widehat z_\ell)\bigr)^2.(10)✦

In the paper, the result reads: for a minimal optimal partition of value C>B2C>B^2 with m≥5m\ge5 active cells, numbered 1,…,m1,\ldots,m with reference index i=mi=m, every distinct pair j,ℓ<mj,\ell<m satisfies Djℓ≤(3/π)t(1+aj)(1+aℓ)D_{j\ell}\le(3/\pi)t(1+a_j)(1+a_\ell).

What Is Checked

The Gaussian cell claim is stated over a Borel set with its Gaussian centroid and pointwise linear score bounds; the partition claims are stated over minimal optimal Gaussian measurable partitions, with every cell active in the squared normalized form.